1 //===-- AMDGPUTargetMachine.cpp - TargetMachine for hw codegen targets-----===// 2 // 3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4 // See https://llvm.org/LICENSE.txt for license information. 5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6 // 7 //===----------------------------------------------------------------------===// 8 // 9 /// \file 10 /// The AMDGPU target machine contains all of the hardware specific 11 /// information needed to emit code for R600 and SI GPUs. 12 // 13 //===----------------------------------------------------------------------===// 14 15 #include "AMDGPUTargetMachine.h" 16 #include "AMDGPU.h" 17 #include "AMDGPUAliasAnalysis.h" 18 #include "AMDGPUExportClustering.h" 19 #include "AMDGPUMacroFusion.h" 20 #include "AMDGPUTargetObjectFile.h" 21 #include "AMDGPUTargetTransformInfo.h" 22 #include "GCNIterativeScheduler.h" 23 #include "GCNSchedStrategy.h" 24 #include "R600MachineScheduler.h" 25 #include "SIMachineFunctionInfo.h" 26 #include "SIMachineScheduler.h" 27 #include "TargetInfo/AMDGPUTargetInfo.h" 28 #include "llvm/Analysis/CGSCCPassManager.h" 29 #include "llvm/CodeGen/GlobalISel/IRTranslator.h" 30 #include "llvm/CodeGen/GlobalISel/InstructionSelect.h" 31 #include "llvm/CodeGen/GlobalISel/Legalizer.h" 32 #include "llvm/CodeGen/GlobalISel/Localizer.h" 33 #include "llvm/CodeGen/GlobalISel/RegBankSelect.h" 34 #include "llvm/CodeGen/MIRParser/MIParser.h" 35 #include "llvm/CodeGen/Passes.h" 36 #include "llvm/CodeGen/RegAllocRegistry.h" 37 #include "llvm/CodeGen/TargetPassConfig.h" 38 #include "llvm/IR/LegacyPassManager.h" 39 #include "llvm/IR/PassManager.h" 40 #include "llvm/InitializePasses.h" 41 #include "llvm/Passes/PassBuilder.h" 42 #include "llvm/Support/TargetRegistry.h" 43 #include "llvm/Transforms/IPO.h" 44 #include "llvm/Transforms/IPO/AlwaysInliner.h" 45 #include "llvm/Transforms/IPO/GlobalDCE.h" 46 #include "llvm/Transforms/IPO/Internalize.h" 47 #include "llvm/Transforms/IPO/PassManagerBuilder.h" 48 #include "llvm/Transforms/Scalar.h" 49 #include "llvm/Transforms/Scalar/GVN.h" 50 #include "llvm/Transforms/Scalar/InferAddressSpaces.h" 51 #include "llvm/Transforms/Utils.h" 52 #include "llvm/Transforms/Utils/SimplifyLibCalls.h" 53 #include "llvm/Transforms/Vectorize.h" 54 55 using namespace llvm; 56 57 namespace { 58 class SGPRRegisterRegAlloc : public RegisterRegAllocBase<SGPRRegisterRegAlloc> { 59 public: 60 SGPRRegisterRegAlloc(const char *N, const char *D, FunctionPassCtor C) 61 : RegisterRegAllocBase(N, D, C) {} 62 }; 63 64 class VGPRRegisterRegAlloc : public RegisterRegAllocBase<VGPRRegisterRegAlloc> { 65 public: 66 VGPRRegisterRegAlloc(const char *N, const char *D, FunctionPassCtor C) 67 : RegisterRegAllocBase(N, D, C) {} 68 }; 69 70 static bool onlyAllocateSGPRs(const TargetRegisterInfo &TRI, 71 const TargetRegisterClass &RC) { 72 return static_cast<const SIRegisterInfo &>(TRI).isSGPRClass(&RC); 73 } 74 75 static bool onlyAllocateVGPRs(const TargetRegisterInfo &TRI, 76 const TargetRegisterClass &RC) { 77 return !static_cast<const SIRegisterInfo &>(TRI).isSGPRClass(&RC); 78 } 79 80 81 /// -{sgpr|vgpr}-regalloc=... command line option. 82 static FunctionPass *useDefaultRegisterAllocator() { return nullptr; } 83 84 /// A dummy default pass factory indicates whether the register allocator is 85 /// overridden on the command line. 86 static llvm::once_flag InitializeDefaultSGPRRegisterAllocatorFlag; 87 static llvm::once_flag InitializeDefaultVGPRRegisterAllocatorFlag; 88 89 static SGPRRegisterRegAlloc 90 defaultSGPRRegAlloc("default", 91 "pick SGPR register allocator based on -O option", 92 useDefaultRegisterAllocator); 93 94 static cl::opt<SGPRRegisterRegAlloc::FunctionPassCtor, false, 95 RegisterPassParser<SGPRRegisterRegAlloc>> 96 SGPRRegAlloc("sgpr-regalloc", cl::Hidden, cl::init(&useDefaultRegisterAllocator), 97 cl::desc("Register allocator to use for SGPRs")); 98 99 static cl::opt<VGPRRegisterRegAlloc::FunctionPassCtor, false, 100 RegisterPassParser<VGPRRegisterRegAlloc>> 101 VGPRRegAlloc("vgpr-regalloc", cl::Hidden, cl::init(&useDefaultRegisterAllocator), 102 cl::desc("Register allocator to use for VGPRs")); 103 104 105 static void initializeDefaultSGPRRegisterAllocatorOnce() { 106 RegisterRegAlloc::FunctionPassCtor Ctor = SGPRRegisterRegAlloc::getDefault(); 107 108 if (!Ctor) { 109 Ctor = SGPRRegAlloc; 110 SGPRRegisterRegAlloc::setDefault(SGPRRegAlloc); 111 } 112 } 113 114 static void initializeDefaultVGPRRegisterAllocatorOnce() { 115 RegisterRegAlloc::FunctionPassCtor Ctor = VGPRRegisterRegAlloc::getDefault(); 116 117 if (!Ctor) { 118 Ctor = VGPRRegAlloc; 119 VGPRRegisterRegAlloc::setDefault(VGPRRegAlloc); 120 } 121 } 122 123 static FunctionPass *createBasicSGPRRegisterAllocator() { 124 return createBasicRegisterAllocator(onlyAllocateSGPRs); 125 } 126 127 static FunctionPass *createGreedySGPRRegisterAllocator() { 128 return createGreedyRegisterAllocator(onlyAllocateSGPRs); 129 } 130 131 static FunctionPass *createFastSGPRRegisterAllocator() { 132 return createFastRegisterAllocator(onlyAllocateSGPRs, false); 133 } 134 135 static FunctionPass *createBasicVGPRRegisterAllocator() { 136 return createBasicRegisterAllocator(onlyAllocateVGPRs); 137 } 138 139 static FunctionPass *createGreedyVGPRRegisterAllocator() { 140 return createGreedyRegisterAllocator(onlyAllocateVGPRs); 141 } 142 143 static FunctionPass *createFastVGPRRegisterAllocator() { 144 return createFastRegisterAllocator(onlyAllocateVGPRs, true); 145 } 146 147 static SGPRRegisterRegAlloc basicRegAllocSGPR( 148 "basic", "basic register allocator", createBasicSGPRRegisterAllocator); 149 static SGPRRegisterRegAlloc greedyRegAllocSGPR( 150 "greedy", "greedy register allocator", createGreedySGPRRegisterAllocator); 151 152 static SGPRRegisterRegAlloc fastRegAllocSGPR( 153 "fast", "fast register allocator", createFastSGPRRegisterAllocator); 154 155 156 static VGPRRegisterRegAlloc basicRegAllocVGPR( 157 "basic", "basic register allocator", createBasicVGPRRegisterAllocator); 158 static VGPRRegisterRegAlloc greedyRegAllocVGPR( 159 "greedy", "greedy register allocator", createGreedyVGPRRegisterAllocator); 160 161 static VGPRRegisterRegAlloc fastRegAllocVGPR( 162 "fast", "fast register allocator", createFastVGPRRegisterAllocator); 163 } 164 165 166 static cl::opt<bool> EnableR600StructurizeCFG( 167 "r600-ir-structurize", 168 cl::desc("Use StructurizeCFG IR pass"), 169 cl::init(true)); 170 171 static cl::opt<bool> EnableSROA( 172 "amdgpu-sroa", 173 cl::desc("Run SROA after promote alloca pass"), 174 cl::ReallyHidden, 175 cl::init(true)); 176 177 static cl::opt<bool> 178 EnableEarlyIfConversion("amdgpu-early-ifcvt", cl::Hidden, 179 cl::desc("Run early if-conversion"), 180 cl::init(false)); 181 182 static cl::opt<bool> 183 OptExecMaskPreRA("amdgpu-opt-exec-mask-pre-ra", cl::Hidden, 184 cl::desc("Run pre-RA exec mask optimizations"), 185 cl::init(true)); 186 187 static cl::opt<bool> EnableR600IfConvert( 188 "r600-if-convert", 189 cl::desc("Use if conversion pass"), 190 cl::ReallyHidden, 191 cl::init(true)); 192 193 // Option to disable vectorizer for tests. 194 static cl::opt<bool> EnableLoadStoreVectorizer( 195 "amdgpu-load-store-vectorizer", 196 cl::desc("Enable load store vectorizer"), 197 cl::init(true), 198 cl::Hidden); 199 200 // Option to control global loads scalarization 201 static cl::opt<bool> ScalarizeGlobal( 202 "amdgpu-scalarize-global-loads", 203 cl::desc("Enable global load scalarization"), 204 cl::init(true), 205 cl::Hidden); 206 207 // Option to run internalize pass. 208 static cl::opt<bool> InternalizeSymbols( 209 "amdgpu-internalize-symbols", 210 cl::desc("Enable elimination of non-kernel functions and unused globals"), 211 cl::init(false), 212 cl::Hidden); 213 214 // Option to inline all early. 215 static cl::opt<bool> EarlyInlineAll( 216 "amdgpu-early-inline-all", 217 cl::desc("Inline all functions early"), 218 cl::init(false), 219 cl::Hidden); 220 221 static cl::opt<bool> EnableSDWAPeephole( 222 "amdgpu-sdwa-peephole", 223 cl::desc("Enable SDWA peepholer"), 224 cl::init(true)); 225 226 static cl::opt<bool> EnableDPPCombine( 227 "amdgpu-dpp-combine", 228 cl::desc("Enable DPP combiner"), 229 cl::init(true)); 230 231 // Enable address space based alias analysis 232 static cl::opt<bool> EnableAMDGPUAliasAnalysis("enable-amdgpu-aa", cl::Hidden, 233 cl::desc("Enable AMDGPU Alias Analysis"), 234 cl::init(true)); 235 236 // Option to run late CFG structurizer 237 static cl::opt<bool, true> LateCFGStructurize( 238 "amdgpu-late-structurize", 239 cl::desc("Enable late CFG structurization"), 240 cl::location(AMDGPUTargetMachine::EnableLateStructurizeCFG), 241 cl::Hidden); 242 243 static cl::opt<bool, true> EnableAMDGPUFunctionCallsOpt( 244 "amdgpu-function-calls", 245 cl::desc("Enable AMDGPU function call support"), 246 cl::location(AMDGPUTargetMachine::EnableFunctionCalls), 247 cl::init(true), 248 cl::Hidden); 249 250 static cl::opt<bool, true> EnableAMDGPUFixedFunctionABIOpt( 251 "amdgpu-fixed-function-abi", 252 cl::desc("Enable all implicit function arguments"), 253 cl::location(AMDGPUTargetMachine::EnableFixedFunctionABI), 254 cl::init(false), 255 cl::Hidden); 256 257 // Enable lib calls simplifications 258 static cl::opt<bool> EnableLibCallSimplify( 259 "amdgpu-simplify-libcall", 260 cl::desc("Enable amdgpu library simplifications"), 261 cl::init(true), 262 cl::Hidden); 263 264 static cl::opt<bool> EnableLowerKernelArguments( 265 "amdgpu-ir-lower-kernel-arguments", 266 cl::desc("Lower kernel argument loads in IR pass"), 267 cl::init(true), 268 cl::Hidden); 269 270 static cl::opt<bool> EnableRegReassign( 271 "amdgpu-reassign-regs", 272 cl::desc("Enable register reassign optimizations on gfx10+"), 273 cl::init(true), 274 cl::Hidden); 275 276 static cl::opt<bool> OptVGPRLiveRange( 277 "amdgpu-opt-vgpr-liverange", 278 cl::desc("Enable VGPR liverange optimizations for if-else structure"), 279 cl::init(true), cl::Hidden); 280 281 // Enable atomic optimization 282 static cl::opt<bool> EnableAtomicOptimizations( 283 "amdgpu-atomic-optimizations", 284 cl::desc("Enable atomic optimizations"), 285 cl::init(false), 286 cl::Hidden); 287 288 // Enable Mode register optimization 289 static cl::opt<bool> EnableSIModeRegisterPass( 290 "amdgpu-mode-register", 291 cl::desc("Enable mode register pass"), 292 cl::init(true), 293 cl::Hidden); 294 295 // Option is used in lit tests to prevent deadcoding of patterns inspected. 296 static cl::opt<bool> 297 EnableDCEInRA("amdgpu-dce-in-ra", 298 cl::init(true), cl::Hidden, 299 cl::desc("Enable machine DCE inside regalloc")); 300 301 static cl::opt<bool> EnableScalarIRPasses( 302 "amdgpu-scalar-ir-passes", 303 cl::desc("Enable scalar IR passes"), 304 cl::init(true), 305 cl::Hidden); 306 307 static cl::opt<bool> EnableStructurizerWorkarounds( 308 "amdgpu-enable-structurizer-workarounds", 309 cl::desc("Enable workarounds for the StructurizeCFG pass"), cl::init(true), 310 cl::Hidden); 311 312 static cl::opt<bool> EnableLDSReplaceWithPointer( 313 "amdgpu-enable-lds-replace-with-pointer", 314 cl::desc("Enable LDS replace with pointer pass"), cl::init(false), 315 cl::Hidden); 316 317 static cl::opt<bool, true> EnableLowerModuleLDS( 318 "amdgpu-enable-lower-module-lds", cl::desc("Enable lower module lds pass"), 319 cl::location(AMDGPUTargetMachine::EnableLowerModuleLDS), cl::init(true), 320 cl::Hidden); 321 322 static cl::opt<bool> EnablePreRAOptimizations( 323 "amdgpu-enable-pre-ra-optimizations", 324 cl::desc("Enable Pre-RA optimizations pass"), cl::init(true), 325 cl::Hidden); 326 327 extern "C" LLVM_EXTERNAL_VISIBILITY void LLVMInitializeAMDGPUTarget() { 328 // Register the target 329 RegisterTargetMachine<R600TargetMachine> X(getTheAMDGPUTarget()); 330 RegisterTargetMachine<GCNTargetMachine> Y(getTheGCNTarget()); 331 332 PassRegistry *PR = PassRegistry::getPassRegistry(); 333 initializeR600ClauseMergePassPass(*PR); 334 initializeR600ControlFlowFinalizerPass(*PR); 335 initializeR600PacketizerPass(*PR); 336 initializeR600ExpandSpecialInstrsPassPass(*PR); 337 initializeR600VectorRegMergerPass(*PR); 338 initializeGlobalISel(*PR); 339 initializeAMDGPUDAGToDAGISelPass(*PR); 340 initializeGCNDPPCombinePass(*PR); 341 initializeSILowerI1CopiesPass(*PR); 342 initializeSILowerSGPRSpillsPass(*PR); 343 initializeSIFixSGPRCopiesPass(*PR); 344 initializeSIFixVGPRCopiesPass(*PR); 345 initializeSIFoldOperandsPass(*PR); 346 initializeSIPeepholeSDWAPass(*PR); 347 initializeSIShrinkInstructionsPass(*PR); 348 initializeSIOptimizeExecMaskingPreRAPass(*PR); 349 initializeSIOptimizeVGPRLiveRangePass(*PR); 350 initializeSILoadStoreOptimizerPass(*PR); 351 initializeAMDGPUFixFunctionBitcastsPass(*PR); 352 initializeAMDGPUAlwaysInlinePass(*PR); 353 initializeAMDGPUAnnotateKernelFeaturesPass(*PR); 354 initializeAMDGPUAnnotateUniformValuesPass(*PR); 355 initializeAMDGPUArgumentUsageInfoPass(*PR); 356 initializeAMDGPUAtomicOptimizerPass(*PR); 357 initializeAMDGPULowerKernelArgumentsPass(*PR); 358 initializeAMDGPULowerKernelAttributesPass(*PR); 359 initializeAMDGPULowerIntrinsicsPass(*PR); 360 initializeAMDGPUOpenCLEnqueuedBlockLoweringPass(*PR); 361 initializeAMDGPUPostLegalizerCombinerPass(*PR); 362 initializeAMDGPUPreLegalizerCombinerPass(*PR); 363 initializeAMDGPURegBankCombinerPass(*PR); 364 initializeAMDGPUPromoteAllocaPass(*PR); 365 initializeAMDGPUPromoteAllocaToVectorPass(*PR); 366 initializeAMDGPUCodeGenPreparePass(*PR); 367 initializeAMDGPULateCodeGenPreparePass(*PR); 368 initializeAMDGPUPropagateAttributesEarlyPass(*PR); 369 initializeAMDGPUPropagateAttributesLatePass(*PR); 370 initializeAMDGPUReplaceLDSUseWithPointerPass(*PR); 371 initializeAMDGPULowerModuleLDSPass(*PR); 372 initializeAMDGPURewriteOutArgumentsPass(*PR); 373 initializeAMDGPUUnifyMetadataPass(*PR); 374 initializeSIAnnotateControlFlowPass(*PR); 375 initializeSIInsertHardClausesPass(*PR); 376 initializeSIInsertWaitcntsPass(*PR); 377 initializeSIModeRegisterPass(*PR); 378 initializeSIWholeQuadModePass(*PR); 379 initializeSILowerControlFlowPass(*PR); 380 initializeSIPreEmitPeepholePass(*PR); 381 initializeSILateBranchLoweringPass(*PR); 382 initializeSIMemoryLegalizerPass(*PR); 383 initializeSIOptimizeExecMaskingPass(*PR); 384 initializeSIPreAllocateWWMRegsPass(*PR); 385 initializeSIFormMemoryClausesPass(*PR); 386 initializeSIPostRABundlerPass(*PR); 387 initializeAMDGPUUnifyDivergentExitNodesPass(*PR); 388 initializeAMDGPUAAWrapperPassPass(*PR); 389 initializeAMDGPUExternalAAWrapperPass(*PR); 390 initializeAMDGPUUseNativeCallsPass(*PR); 391 initializeAMDGPUSimplifyLibCallsPass(*PR); 392 initializeAMDGPUPrintfRuntimeBindingPass(*PR); 393 initializeGCNNSAReassignPass(*PR); 394 initializeGCNPreRAOptimizationsPass(*PR); 395 } 396 397 static std::unique_ptr<TargetLoweringObjectFile> createTLOF(const Triple &TT) { 398 return std::make_unique<AMDGPUTargetObjectFile>(); 399 } 400 401 static ScheduleDAGInstrs *createR600MachineScheduler(MachineSchedContext *C) { 402 return new ScheduleDAGMILive(C, std::make_unique<R600SchedStrategy>()); 403 } 404 405 static ScheduleDAGInstrs *createSIMachineScheduler(MachineSchedContext *C) { 406 return new SIScheduleDAGMI(C); 407 } 408 409 static ScheduleDAGInstrs * 410 createGCNMaxOccupancyMachineScheduler(MachineSchedContext *C) { 411 ScheduleDAGMILive *DAG = 412 new GCNScheduleDAGMILive(C, std::make_unique<GCNMaxOccupancySchedStrategy>(C)); 413 DAG->addMutation(createLoadClusterDAGMutation(DAG->TII, DAG->TRI)); 414 DAG->addMutation(createAMDGPUMacroFusionDAGMutation()); 415 DAG->addMutation(createAMDGPUExportClusteringDAGMutation()); 416 return DAG; 417 } 418 419 static ScheduleDAGInstrs * 420 createIterativeGCNMaxOccupancyMachineScheduler(MachineSchedContext *C) { 421 auto DAG = new GCNIterativeScheduler(C, 422 GCNIterativeScheduler::SCHEDULE_LEGACYMAXOCCUPANCY); 423 DAG->addMutation(createLoadClusterDAGMutation(DAG->TII, DAG->TRI)); 424 return DAG; 425 } 426 427 static ScheduleDAGInstrs *createMinRegScheduler(MachineSchedContext *C) { 428 return new GCNIterativeScheduler(C, 429 GCNIterativeScheduler::SCHEDULE_MINREGFORCED); 430 } 431 432 static ScheduleDAGInstrs * 433 createIterativeILPMachineScheduler(MachineSchedContext *C) { 434 auto DAG = new GCNIterativeScheduler(C, 435 GCNIterativeScheduler::SCHEDULE_ILP); 436 DAG->addMutation(createLoadClusterDAGMutation(DAG->TII, DAG->TRI)); 437 DAG->addMutation(createAMDGPUMacroFusionDAGMutation()); 438 return DAG; 439 } 440 441 static MachineSchedRegistry 442 R600SchedRegistry("r600", "Run R600's custom scheduler", 443 createR600MachineScheduler); 444 445 static MachineSchedRegistry 446 SISchedRegistry("si", "Run SI's custom scheduler", 447 createSIMachineScheduler); 448 449 static MachineSchedRegistry 450 GCNMaxOccupancySchedRegistry("gcn-max-occupancy", 451 "Run GCN scheduler to maximize occupancy", 452 createGCNMaxOccupancyMachineScheduler); 453 454 static MachineSchedRegistry 455 IterativeGCNMaxOccupancySchedRegistry("gcn-max-occupancy-experimental", 456 "Run GCN scheduler to maximize occupancy (experimental)", 457 createIterativeGCNMaxOccupancyMachineScheduler); 458 459 static MachineSchedRegistry 460 GCNMinRegSchedRegistry("gcn-minreg", 461 "Run GCN iterative scheduler for minimal register usage (experimental)", 462 createMinRegScheduler); 463 464 static MachineSchedRegistry 465 GCNILPSchedRegistry("gcn-ilp", 466 "Run GCN iterative scheduler for ILP scheduling (experimental)", 467 createIterativeILPMachineScheduler); 468 469 static StringRef computeDataLayout(const Triple &TT) { 470 if (TT.getArch() == Triple::r600) { 471 // 32-bit pointers. 472 return "e-p:32:32-i64:64-v16:16-v24:32-v32:32-v48:64-v96:128" 473 "-v192:256-v256:256-v512:512-v1024:1024-v2048:2048-n32:64-S32-A5-G1"; 474 } 475 476 // 32-bit private, local, and region pointers. 64-bit global, constant and 477 // flat, non-integral buffer fat pointers. 478 return "e-p:64:64-p1:64:64-p2:32:32-p3:32:32-p4:64:64-p5:32:32-p6:32:32" 479 "-i64:64-v16:16-v24:32-v32:32-v48:64-v96:128" 480 "-v192:256-v256:256-v512:512-v1024:1024-v2048:2048-n32:64-S32-A5-G1" 481 "-ni:7"; 482 } 483 484 LLVM_READNONE 485 static StringRef getGPUOrDefault(const Triple &TT, StringRef GPU) { 486 if (!GPU.empty()) 487 return GPU; 488 489 // Need to default to a target with flat support for HSA. 490 if (TT.getArch() == Triple::amdgcn) 491 return TT.getOS() == Triple::AMDHSA ? "generic-hsa" : "generic"; 492 493 return "r600"; 494 } 495 496 static Reloc::Model getEffectiveRelocModel(Optional<Reloc::Model> RM) { 497 // The AMDGPU toolchain only supports generating shared objects, so we 498 // must always use PIC. 499 return Reloc::PIC_; 500 } 501 502 AMDGPUTargetMachine::AMDGPUTargetMachine(const Target &T, const Triple &TT, 503 StringRef CPU, StringRef FS, 504 TargetOptions Options, 505 Optional<Reloc::Model> RM, 506 Optional<CodeModel::Model> CM, 507 CodeGenOpt::Level OptLevel) 508 : LLVMTargetMachine(T, computeDataLayout(TT), TT, getGPUOrDefault(TT, CPU), 509 FS, Options, getEffectiveRelocModel(RM), 510 getEffectiveCodeModel(CM, CodeModel::Small), OptLevel), 511 TLOF(createTLOF(getTargetTriple())) { 512 initAsmInfo(); 513 if (TT.getArch() == Triple::amdgcn) { 514 if (getMCSubtargetInfo()->checkFeatures("+wavefrontsize64")) 515 MRI.reset(llvm::createGCNMCRegisterInfo(AMDGPUDwarfFlavour::Wave64)); 516 else if (getMCSubtargetInfo()->checkFeatures("+wavefrontsize32")) 517 MRI.reset(llvm::createGCNMCRegisterInfo(AMDGPUDwarfFlavour::Wave32)); 518 } 519 } 520 521 bool AMDGPUTargetMachine::EnableLateStructurizeCFG = false; 522 bool AMDGPUTargetMachine::EnableFunctionCalls = false; 523 bool AMDGPUTargetMachine::EnableFixedFunctionABI = false; 524 bool AMDGPUTargetMachine::EnableLowerModuleLDS = true; 525 526 AMDGPUTargetMachine::~AMDGPUTargetMachine() = default; 527 528 StringRef AMDGPUTargetMachine::getGPUName(const Function &F) const { 529 Attribute GPUAttr = F.getFnAttribute("target-cpu"); 530 return GPUAttr.isValid() ? GPUAttr.getValueAsString() : getTargetCPU(); 531 } 532 533 StringRef AMDGPUTargetMachine::getFeatureString(const Function &F) const { 534 Attribute FSAttr = F.getFnAttribute("target-features"); 535 536 return FSAttr.isValid() ? FSAttr.getValueAsString() 537 : getTargetFeatureString(); 538 } 539 540 /// Predicate for Internalize pass. 541 static bool mustPreserveGV(const GlobalValue &GV) { 542 if (const Function *F = dyn_cast<Function>(&GV)) 543 return F->isDeclaration() || AMDGPU::isEntryFunctionCC(F->getCallingConv()); 544 545 GV.removeDeadConstantUsers(); 546 return !GV.use_empty(); 547 } 548 549 void AMDGPUTargetMachine::adjustPassManager(PassManagerBuilder &Builder) { 550 Builder.DivergentTarget = true; 551 552 bool EnableOpt = getOptLevel() > CodeGenOpt::None; 553 bool Internalize = InternalizeSymbols; 554 bool EarlyInline = EarlyInlineAll && EnableOpt && !EnableFunctionCalls; 555 bool AMDGPUAA = EnableAMDGPUAliasAnalysis && EnableOpt; 556 bool LibCallSimplify = EnableLibCallSimplify && EnableOpt; 557 558 if (EnableFunctionCalls) { 559 delete Builder.Inliner; 560 Builder.Inliner = createFunctionInliningPass(); 561 } 562 563 Builder.addExtension( 564 PassManagerBuilder::EP_ModuleOptimizerEarly, 565 [Internalize, EarlyInline, AMDGPUAA, this](const PassManagerBuilder &, 566 legacy::PassManagerBase &PM) { 567 if (AMDGPUAA) { 568 PM.add(createAMDGPUAAWrapperPass()); 569 PM.add(createAMDGPUExternalAAWrapperPass()); 570 } 571 PM.add(createAMDGPUUnifyMetadataPass()); 572 PM.add(createAMDGPUPrintfRuntimeBinding()); 573 if (Internalize) 574 PM.add(createInternalizePass(mustPreserveGV)); 575 PM.add(createAMDGPUPropagateAttributesLatePass(this)); 576 if (Internalize) 577 PM.add(createGlobalDCEPass()); 578 if (EarlyInline) 579 PM.add(createAMDGPUAlwaysInlinePass(false)); 580 }); 581 582 Builder.addExtension( 583 PassManagerBuilder::EP_EarlyAsPossible, 584 [AMDGPUAA, LibCallSimplify, this](const PassManagerBuilder &, 585 legacy::PassManagerBase &PM) { 586 if (AMDGPUAA) { 587 PM.add(createAMDGPUAAWrapperPass()); 588 PM.add(createAMDGPUExternalAAWrapperPass()); 589 } 590 PM.add(llvm::createAMDGPUPropagateAttributesEarlyPass(this)); 591 PM.add(llvm::createAMDGPUUseNativeCallsPass()); 592 if (LibCallSimplify) 593 PM.add(llvm::createAMDGPUSimplifyLibCallsPass(this)); 594 }); 595 596 Builder.addExtension( 597 PassManagerBuilder::EP_CGSCCOptimizerLate, 598 [EnableOpt](const PassManagerBuilder &, legacy::PassManagerBase &PM) { 599 // Add infer address spaces pass to the opt pipeline after inlining 600 // but before SROA to increase SROA opportunities. 601 PM.add(createInferAddressSpacesPass()); 602 603 // This should run after inlining to have any chance of doing anything, 604 // and before other cleanup optimizations. 605 PM.add(createAMDGPULowerKernelAttributesPass()); 606 607 // Promote alloca to vector before SROA and loop unroll. If we manage 608 // to eliminate allocas before unroll we may choose to unroll less. 609 if (EnableOpt) 610 PM.add(createAMDGPUPromoteAllocaToVector()); 611 }); 612 } 613 614 void AMDGPUTargetMachine::registerDefaultAliasAnalyses(AAManager &AAM) { 615 AAM.registerFunctionAnalysis<AMDGPUAA>(); 616 } 617 618 void AMDGPUTargetMachine::registerPassBuilderCallbacks(PassBuilder &PB) { 619 PB.registerPipelineParsingCallback( 620 [this](StringRef PassName, ModulePassManager &PM, 621 ArrayRef<PassBuilder::PipelineElement>) { 622 if (PassName == "amdgpu-propagate-attributes-late") { 623 PM.addPass(AMDGPUPropagateAttributesLatePass(*this)); 624 return true; 625 } 626 if (PassName == "amdgpu-unify-metadata") { 627 PM.addPass(AMDGPUUnifyMetadataPass()); 628 return true; 629 } 630 if (PassName == "amdgpu-printf-runtime-binding") { 631 PM.addPass(AMDGPUPrintfRuntimeBindingPass()); 632 return true; 633 } 634 if (PassName == "amdgpu-always-inline") { 635 PM.addPass(AMDGPUAlwaysInlinePass()); 636 return true; 637 } 638 if (PassName == "amdgpu-replace-lds-use-with-pointer") { 639 PM.addPass(AMDGPUReplaceLDSUseWithPointerPass()); 640 return true; 641 } 642 if (PassName == "amdgpu-lower-module-lds") { 643 PM.addPass(AMDGPULowerModuleLDSPass()); 644 return true; 645 } 646 return false; 647 }); 648 PB.registerPipelineParsingCallback( 649 [this](StringRef PassName, FunctionPassManager &PM, 650 ArrayRef<PassBuilder::PipelineElement>) { 651 if (PassName == "amdgpu-simplifylib") { 652 PM.addPass(AMDGPUSimplifyLibCallsPass(*this)); 653 return true; 654 } 655 if (PassName == "amdgpu-usenative") { 656 PM.addPass(AMDGPUUseNativeCallsPass()); 657 return true; 658 } 659 if (PassName == "amdgpu-promote-alloca") { 660 PM.addPass(AMDGPUPromoteAllocaPass(*this)); 661 return true; 662 } 663 if (PassName == "amdgpu-promote-alloca-to-vector") { 664 PM.addPass(AMDGPUPromoteAllocaToVectorPass(*this)); 665 return true; 666 } 667 if (PassName == "amdgpu-lower-kernel-attributes") { 668 PM.addPass(AMDGPULowerKernelAttributesPass()); 669 return true; 670 } 671 if (PassName == "amdgpu-propagate-attributes-early") { 672 PM.addPass(AMDGPUPropagateAttributesEarlyPass(*this)); 673 return true; 674 } 675 return false; 676 }); 677 678 PB.registerAnalysisRegistrationCallback([](FunctionAnalysisManager &FAM) { 679 FAM.registerPass([&] { return AMDGPUAA(); }); 680 }); 681 682 PB.registerParseAACallback([](StringRef AAName, AAManager &AAM) { 683 if (AAName == "amdgpu-aa") { 684 AAM.registerFunctionAnalysis<AMDGPUAA>(); 685 return true; 686 } 687 return false; 688 }); 689 690 PB.registerPipelineStartEPCallback( 691 [this](ModulePassManager &PM, PassBuilder::OptimizationLevel Level) { 692 FunctionPassManager FPM; 693 FPM.addPass(AMDGPUPropagateAttributesEarlyPass(*this)); 694 FPM.addPass(AMDGPUUseNativeCallsPass()); 695 if (EnableLibCallSimplify && 696 Level != PassBuilder::OptimizationLevel::O0) 697 FPM.addPass(AMDGPUSimplifyLibCallsPass(*this)); 698 PM.addPass(createModuleToFunctionPassAdaptor(std::move(FPM))); 699 }); 700 701 PB.registerPipelineEarlySimplificationEPCallback( 702 [this](ModulePassManager &PM, PassBuilder::OptimizationLevel Level) { 703 if (Level == PassBuilder::OptimizationLevel::O0) 704 return; 705 706 PM.addPass(AMDGPUUnifyMetadataPass()); 707 PM.addPass(AMDGPUPrintfRuntimeBindingPass()); 708 709 if (InternalizeSymbols) { 710 PM.addPass(InternalizePass(mustPreserveGV)); 711 } 712 PM.addPass(AMDGPUPropagateAttributesLatePass(*this)); 713 if (InternalizeSymbols) { 714 PM.addPass(GlobalDCEPass()); 715 } 716 if (EarlyInlineAll && !EnableFunctionCalls) 717 PM.addPass(AMDGPUAlwaysInlinePass()); 718 }); 719 720 PB.registerCGSCCOptimizerLateEPCallback( 721 [this](CGSCCPassManager &PM, PassBuilder::OptimizationLevel Level) { 722 if (Level == PassBuilder::OptimizationLevel::O0) 723 return; 724 725 FunctionPassManager FPM; 726 727 // Add infer address spaces pass to the opt pipeline after inlining 728 // but before SROA to increase SROA opportunities. 729 FPM.addPass(InferAddressSpacesPass()); 730 731 // This should run after inlining to have any chance of doing 732 // anything, and before other cleanup optimizations. 733 FPM.addPass(AMDGPULowerKernelAttributesPass()); 734 735 if (Level != PassBuilder::OptimizationLevel::O0) { 736 // Promote alloca to vector before SROA and loop unroll. If we 737 // manage to eliminate allocas before unroll we may choose to unroll 738 // less. 739 FPM.addPass(AMDGPUPromoteAllocaToVectorPass(*this)); 740 } 741 742 PM.addPass(createCGSCCToFunctionPassAdaptor(std::move(FPM))); 743 }); 744 } 745 746 //===----------------------------------------------------------------------===// 747 // R600 Target Machine (R600 -> Cayman) 748 //===----------------------------------------------------------------------===// 749 750 R600TargetMachine::R600TargetMachine(const Target &T, const Triple &TT, 751 StringRef CPU, StringRef FS, 752 TargetOptions Options, 753 Optional<Reloc::Model> RM, 754 Optional<CodeModel::Model> CM, 755 CodeGenOpt::Level OL, bool JIT) 756 : AMDGPUTargetMachine(T, TT, CPU, FS, Options, RM, CM, OL) { 757 setRequiresStructuredCFG(true); 758 759 // Override the default since calls aren't supported for r600. 760 if (EnableFunctionCalls && 761 EnableAMDGPUFunctionCallsOpt.getNumOccurrences() == 0) 762 EnableFunctionCalls = false; 763 } 764 765 const R600Subtarget *R600TargetMachine::getSubtargetImpl( 766 const Function &F) const { 767 StringRef GPU = getGPUName(F); 768 StringRef FS = getFeatureString(F); 769 770 SmallString<128> SubtargetKey(GPU); 771 SubtargetKey.append(FS); 772 773 auto &I = SubtargetMap[SubtargetKey]; 774 if (!I) { 775 // This needs to be done before we create a new subtarget since any 776 // creation will depend on the TM and the code generation flags on the 777 // function that reside in TargetOptions. 778 resetTargetOptions(F); 779 I = std::make_unique<R600Subtarget>(TargetTriple, GPU, FS, *this); 780 } 781 782 return I.get(); 783 } 784 785 int64_t AMDGPUTargetMachine::getNullPointerValue(unsigned AddrSpace) { 786 return (AddrSpace == AMDGPUAS::LOCAL_ADDRESS || 787 AddrSpace == AMDGPUAS::PRIVATE_ADDRESS || 788 AddrSpace == AMDGPUAS::REGION_ADDRESS) 789 ? -1 790 : 0; 791 } 792 793 bool AMDGPUTargetMachine::isNoopAddrSpaceCast(unsigned SrcAS, 794 unsigned DestAS) const { 795 return AMDGPU::isFlatGlobalAddrSpace(SrcAS) && 796 AMDGPU::isFlatGlobalAddrSpace(DestAS); 797 } 798 799 unsigned AMDGPUTargetMachine::getAssumedAddrSpace(const Value *V) const { 800 const auto *LD = dyn_cast<LoadInst>(V); 801 if (!LD) 802 return AMDGPUAS::UNKNOWN_ADDRESS_SPACE; 803 804 // It must be a generic pointer loaded. 805 assert(V->getType()->isPointerTy() && 806 V->getType()->getPointerAddressSpace() == AMDGPUAS::FLAT_ADDRESS); 807 808 const auto *Ptr = LD->getPointerOperand(); 809 if (Ptr->getType()->getPointerAddressSpace() != AMDGPUAS::CONSTANT_ADDRESS) 810 return AMDGPUAS::UNKNOWN_ADDRESS_SPACE; 811 // For a generic pointer loaded from the constant memory, it could be assumed 812 // as a global pointer since the constant memory is only populated on the 813 // host side. As implied by the offload programming model, only global 814 // pointers could be referenced on the host side. 815 return AMDGPUAS::GLOBAL_ADDRESS; 816 } 817 818 TargetTransformInfo 819 R600TargetMachine::getTargetTransformInfo(const Function &F) { 820 return TargetTransformInfo(R600TTIImpl(this, F)); 821 } 822 823 //===----------------------------------------------------------------------===// 824 // GCN Target Machine (SI+) 825 //===----------------------------------------------------------------------===// 826 827 GCNTargetMachine::GCNTargetMachine(const Target &T, const Triple &TT, 828 StringRef CPU, StringRef FS, 829 TargetOptions Options, 830 Optional<Reloc::Model> RM, 831 Optional<CodeModel::Model> CM, 832 CodeGenOpt::Level OL, bool JIT) 833 : AMDGPUTargetMachine(T, TT, CPU, FS, Options, RM, CM, OL) {} 834 835 const GCNSubtarget *GCNTargetMachine::getSubtargetImpl(const Function &F) const { 836 StringRef GPU = getGPUName(F); 837 StringRef FS = getFeatureString(F); 838 839 SmallString<128> SubtargetKey(GPU); 840 SubtargetKey.append(FS); 841 842 auto &I = SubtargetMap[SubtargetKey]; 843 if (!I) { 844 // This needs to be done before we create a new subtarget since any 845 // creation will depend on the TM and the code generation flags on the 846 // function that reside in TargetOptions. 847 resetTargetOptions(F); 848 I = std::make_unique<GCNSubtarget>(TargetTriple, GPU, FS, *this); 849 } 850 851 I->setScalarizeGlobalBehavior(ScalarizeGlobal); 852 853 return I.get(); 854 } 855 856 TargetTransformInfo 857 GCNTargetMachine::getTargetTransformInfo(const Function &F) { 858 return TargetTransformInfo(GCNTTIImpl(this, F)); 859 } 860 861 //===----------------------------------------------------------------------===// 862 // AMDGPU Pass Setup 863 //===----------------------------------------------------------------------===// 864 865 namespace { 866 867 class AMDGPUPassConfig : public TargetPassConfig { 868 public: 869 AMDGPUPassConfig(LLVMTargetMachine &TM, PassManagerBase &PM) 870 : TargetPassConfig(TM, PM) { 871 // Exceptions and StackMaps are not supported, so these passes will never do 872 // anything. 873 disablePass(&StackMapLivenessID); 874 disablePass(&FuncletLayoutID); 875 // Garbage collection is not supported. 876 disablePass(&GCLoweringID); 877 disablePass(&ShadowStackGCLoweringID); 878 } 879 880 AMDGPUTargetMachine &getAMDGPUTargetMachine() const { 881 return getTM<AMDGPUTargetMachine>(); 882 } 883 884 ScheduleDAGInstrs * 885 createMachineScheduler(MachineSchedContext *C) const override { 886 ScheduleDAGMILive *DAG = createGenericSchedLive(C); 887 DAG->addMutation(createLoadClusterDAGMutation(DAG->TII, DAG->TRI)); 888 return DAG; 889 } 890 891 void addEarlyCSEOrGVNPass(); 892 void addStraightLineScalarOptimizationPasses(); 893 void addIRPasses() override; 894 void addCodeGenPrepare() override; 895 bool addPreISel() override; 896 bool addInstSelector() override; 897 bool addGCPasses() override; 898 899 std::unique_ptr<CSEConfigBase> getCSEConfig() const override; 900 901 /// Check if a pass is enabled given \p Opt option. The option always 902 /// overrides defaults if explicitely used. Otherwise its default will 903 /// be used given that a pass shall work at an optimization \p Level 904 /// minimum. 905 bool isPassEnabled(const cl::opt<bool> &Opt, 906 CodeGenOpt::Level Level = CodeGenOpt::Default) const { 907 if (Opt.getNumOccurrences()) 908 return Opt; 909 if (TM->getOptLevel() < Level) 910 return false; 911 return Opt; 912 } 913 }; 914 915 std::unique_ptr<CSEConfigBase> AMDGPUPassConfig::getCSEConfig() const { 916 return getStandardCSEConfigForOpt(TM->getOptLevel()); 917 } 918 919 class R600PassConfig final : public AMDGPUPassConfig { 920 public: 921 R600PassConfig(LLVMTargetMachine &TM, PassManagerBase &PM) 922 : AMDGPUPassConfig(TM, PM) {} 923 924 ScheduleDAGInstrs *createMachineScheduler( 925 MachineSchedContext *C) const override { 926 return createR600MachineScheduler(C); 927 } 928 929 bool addPreISel() override; 930 bool addInstSelector() override; 931 void addPreRegAlloc() override; 932 void addPreSched2() override; 933 void addPreEmitPass() override; 934 }; 935 936 class GCNPassConfig final : public AMDGPUPassConfig { 937 public: 938 GCNPassConfig(LLVMTargetMachine &TM, PassManagerBase &PM) 939 : AMDGPUPassConfig(TM, PM) { 940 // It is necessary to know the register usage of the entire call graph. We 941 // allow calls without EnableAMDGPUFunctionCalls if they are marked 942 // noinline, so this is always required. 943 setRequiresCodeGenSCCOrder(true); 944 } 945 946 GCNTargetMachine &getGCNTargetMachine() const { 947 return getTM<GCNTargetMachine>(); 948 } 949 950 ScheduleDAGInstrs * 951 createMachineScheduler(MachineSchedContext *C) const override; 952 953 bool addPreISel() override; 954 void addMachineSSAOptimization() override; 955 bool addILPOpts() override; 956 bool addInstSelector() override; 957 bool addIRTranslator() override; 958 void addPreLegalizeMachineIR() override; 959 bool addLegalizeMachineIR() override; 960 void addPreRegBankSelect() override; 961 bool addRegBankSelect() override; 962 void addPreGlobalInstructionSelect() override; 963 bool addGlobalInstructionSelect() override; 964 void addFastRegAlloc() override; 965 void addOptimizedRegAlloc() override; 966 967 FunctionPass *createSGPRAllocPass(bool Optimized); 968 FunctionPass *createVGPRAllocPass(bool Optimized); 969 FunctionPass *createRegAllocPass(bool Optimized) override; 970 971 bool addRegAssignAndRewriteFast() override; 972 bool addRegAssignAndRewriteOptimized() override; 973 974 void addPreRegAlloc() override; 975 bool addPreRewrite() override; 976 void addPostRegAlloc() override; 977 void addPreSched2() override; 978 void addPreEmitPass() override; 979 }; 980 981 } // end anonymous namespace 982 983 void AMDGPUPassConfig::addEarlyCSEOrGVNPass() { 984 if (getOptLevel() == CodeGenOpt::Aggressive) 985 addPass(createGVNPass()); 986 else 987 addPass(createEarlyCSEPass()); 988 } 989 990 void AMDGPUPassConfig::addStraightLineScalarOptimizationPasses() { 991 addPass(createLICMPass()); 992 addPass(createSeparateConstOffsetFromGEPPass()); 993 addPass(createSpeculativeExecutionPass()); 994 // ReassociateGEPs exposes more opportunites for SLSR. See 995 // the example in reassociate-geps-and-slsr.ll. 996 addPass(createStraightLineStrengthReducePass()); 997 // SeparateConstOffsetFromGEP and SLSR creates common expressions which GVN or 998 // EarlyCSE can reuse. 999 addEarlyCSEOrGVNPass(); 1000 // Run NaryReassociate after EarlyCSE/GVN to be more effective. 1001 addPass(createNaryReassociatePass()); 1002 // NaryReassociate on GEPs creates redundant common expressions, so run 1003 // EarlyCSE after it. 1004 addPass(createEarlyCSEPass()); 1005 } 1006 1007 void AMDGPUPassConfig::addIRPasses() { 1008 const AMDGPUTargetMachine &TM = getAMDGPUTargetMachine(); 1009 1010 // There is no reason to run these. 1011 disablePass(&StackMapLivenessID); 1012 disablePass(&FuncletLayoutID); 1013 disablePass(&PatchableFunctionID); 1014 1015 addPass(createAMDGPUPrintfRuntimeBinding()); 1016 1017 // This must occur before inlining, as the inliner will not look through 1018 // bitcast calls. 1019 addPass(createAMDGPUFixFunctionBitcastsPass()); 1020 1021 // A call to propagate attributes pass in the backend in case opt was not run. 1022 addPass(createAMDGPUPropagateAttributesEarlyPass(&TM)); 1023 1024 addPass(createAMDGPULowerIntrinsicsPass()); 1025 1026 // Function calls are not supported, so make sure we inline everything. 1027 addPass(createAMDGPUAlwaysInlinePass()); 1028 addPass(createAlwaysInlinerLegacyPass()); 1029 // We need to add the barrier noop pass, otherwise adding the function 1030 // inlining pass will cause all of the PassConfigs passes to be run 1031 // one function at a time, which means if we have a nodule with two 1032 // functions, then we will generate code for the first function 1033 // without ever running any passes on the second. 1034 addPass(createBarrierNoopPass()); 1035 1036 // Handle uses of OpenCL image2d_t, image3d_t and sampler_t arguments. 1037 if (TM.getTargetTriple().getArch() == Triple::r600) 1038 addPass(createR600OpenCLImageTypeLoweringPass()); 1039 1040 // Replace OpenCL enqueued block function pointers with global variables. 1041 addPass(createAMDGPUOpenCLEnqueuedBlockLoweringPass()); 1042 1043 // Can increase LDS used by kernel so runs before PromoteAlloca 1044 if (EnableLowerModuleLDS) { 1045 // The pass "amdgpu-replace-lds-use-with-pointer" need to be run before the 1046 // pass "amdgpu-lower-module-lds", and also it required to be run only if 1047 // "amdgpu-lower-module-lds" pass is enabled. 1048 if (EnableLDSReplaceWithPointer) 1049 addPass(createAMDGPUReplaceLDSUseWithPointerPass()); 1050 1051 addPass(createAMDGPULowerModuleLDSPass()); 1052 } 1053 1054 if (TM.getOptLevel() > CodeGenOpt::None) 1055 addPass(createInferAddressSpacesPass()); 1056 1057 addPass(createAtomicExpandPass()); 1058 1059 if (TM.getOptLevel() > CodeGenOpt::None) { 1060 addPass(createAMDGPUPromoteAlloca()); 1061 1062 if (EnableSROA) 1063 addPass(createSROAPass()); 1064 if (isPassEnabled(EnableScalarIRPasses)) 1065 addStraightLineScalarOptimizationPasses(); 1066 1067 if (EnableAMDGPUAliasAnalysis) { 1068 addPass(createAMDGPUAAWrapperPass()); 1069 addPass(createExternalAAWrapperPass([](Pass &P, Function &, 1070 AAResults &AAR) { 1071 if (auto *WrapperPass = P.getAnalysisIfAvailable<AMDGPUAAWrapperPass>()) 1072 AAR.addAAResult(WrapperPass->getResult()); 1073 })); 1074 } 1075 1076 if (TM.getTargetTriple().getArch() == Triple::amdgcn) { 1077 // TODO: May want to move later or split into an early and late one. 1078 addPass(createAMDGPUCodeGenPreparePass()); 1079 } 1080 } 1081 1082 TargetPassConfig::addIRPasses(); 1083 1084 // EarlyCSE is not always strong enough to clean up what LSR produces. For 1085 // example, GVN can combine 1086 // 1087 // %0 = add %a, %b 1088 // %1 = add %b, %a 1089 // 1090 // and 1091 // 1092 // %0 = shl nsw %a, 2 1093 // %1 = shl %a, 2 1094 // 1095 // but EarlyCSE can do neither of them. 1096 if (isPassEnabled(EnableScalarIRPasses)) 1097 addEarlyCSEOrGVNPass(); 1098 } 1099 1100 void AMDGPUPassConfig::addCodeGenPrepare() { 1101 if (TM->getTargetTriple().getArch() == Triple::amdgcn) 1102 addPass(createAMDGPUAnnotateKernelFeaturesPass()); 1103 1104 if (TM->getTargetTriple().getArch() == Triple::amdgcn && 1105 EnableLowerKernelArguments) 1106 addPass(createAMDGPULowerKernelArgumentsPass()); 1107 1108 if (TM->getOptLevel() > CodeGenOpt::Less) 1109 addPass(&AMDGPUPerfHintAnalysisID); 1110 1111 TargetPassConfig::addCodeGenPrepare(); 1112 1113 if (isPassEnabled(EnableLoadStoreVectorizer)) 1114 addPass(createLoadStoreVectorizerPass()); 1115 1116 // LowerSwitch pass may introduce unreachable blocks that can 1117 // cause unexpected behavior for subsequent passes. Placing it 1118 // here seems better that these blocks would get cleaned up by 1119 // UnreachableBlockElim inserted next in the pass flow. 1120 addPass(createLowerSwitchPass()); 1121 } 1122 1123 bool AMDGPUPassConfig::addPreISel() { 1124 if (TM->getOptLevel() > CodeGenOpt::None) 1125 addPass(createFlattenCFGPass()); 1126 return false; 1127 } 1128 1129 bool AMDGPUPassConfig::addInstSelector() { 1130 // Defer the verifier until FinalizeISel. 1131 addPass(createAMDGPUISelDag(&getAMDGPUTargetMachine(), getOptLevel()), false); 1132 return false; 1133 } 1134 1135 bool AMDGPUPassConfig::addGCPasses() { 1136 // Do nothing. GC is not supported. 1137 return false; 1138 } 1139 1140 //===----------------------------------------------------------------------===// 1141 // R600 Pass Setup 1142 //===----------------------------------------------------------------------===// 1143 1144 bool R600PassConfig::addPreISel() { 1145 AMDGPUPassConfig::addPreISel(); 1146 1147 if (EnableR600StructurizeCFG) 1148 addPass(createStructurizeCFGPass()); 1149 return false; 1150 } 1151 1152 bool R600PassConfig::addInstSelector() { 1153 addPass(createR600ISelDag(&getAMDGPUTargetMachine(), getOptLevel())); 1154 return false; 1155 } 1156 1157 void R600PassConfig::addPreRegAlloc() { 1158 addPass(createR600VectorRegMerger()); 1159 } 1160 1161 void R600PassConfig::addPreSched2() { 1162 addPass(createR600EmitClauseMarkers(), false); 1163 if (EnableR600IfConvert) 1164 addPass(&IfConverterID, false); 1165 addPass(createR600ClauseMergePass(), false); 1166 } 1167 1168 void R600PassConfig::addPreEmitPass() { 1169 addPass(createAMDGPUCFGStructurizerPass(), false); 1170 addPass(createR600ExpandSpecialInstrsPass(), false); 1171 addPass(&FinalizeMachineBundlesID, false); 1172 addPass(createR600Packetizer(), false); 1173 addPass(createR600ControlFlowFinalizer(), false); 1174 } 1175 1176 TargetPassConfig *R600TargetMachine::createPassConfig(PassManagerBase &PM) { 1177 return new R600PassConfig(*this, PM); 1178 } 1179 1180 //===----------------------------------------------------------------------===// 1181 // GCN Pass Setup 1182 //===----------------------------------------------------------------------===// 1183 1184 ScheduleDAGInstrs *GCNPassConfig::createMachineScheduler( 1185 MachineSchedContext *C) const { 1186 const GCNSubtarget &ST = C->MF->getSubtarget<GCNSubtarget>(); 1187 if (ST.enableSIScheduler()) 1188 return createSIMachineScheduler(C); 1189 return createGCNMaxOccupancyMachineScheduler(C); 1190 } 1191 1192 bool GCNPassConfig::addPreISel() { 1193 AMDGPUPassConfig::addPreISel(); 1194 1195 if (TM->getOptLevel() > CodeGenOpt::None) 1196 addPass(createAMDGPULateCodeGenPreparePass()); 1197 1198 if (isPassEnabled(EnableAtomicOptimizations, CodeGenOpt::Less)) { 1199 addPass(createAMDGPUAtomicOptimizerPass()); 1200 } 1201 1202 if (TM->getOptLevel() > CodeGenOpt::None) 1203 addPass(createSinkingPass()); 1204 1205 // Merge divergent exit nodes. StructurizeCFG won't recognize the multi-exit 1206 // regions formed by them. 1207 addPass(&AMDGPUUnifyDivergentExitNodesID); 1208 if (!LateCFGStructurize) { 1209 if (EnableStructurizerWorkarounds) { 1210 addPass(createFixIrreduciblePass()); 1211 addPass(createUnifyLoopExitsPass()); 1212 } 1213 addPass(createStructurizeCFGPass(false)); // true -> SkipUniformRegions 1214 } 1215 addPass(createAMDGPUAnnotateUniformValues()); 1216 if (!LateCFGStructurize) { 1217 addPass(createSIAnnotateControlFlowPass()); 1218 } 1219 addPass(createLCSSAPass()); 1220 1221 return false; 1222 } 1223 1224 void GCNPassConfig::addMachineSSAOptimization() { 1225 TargetPassConfig::addMachineSSAOptimization(); 1226 1227 // We want to fold operands after PeepholeOptimizer has run (or as part of 1228 // it), because it will eliminate extra copies making it easier to fold the 1229 // real source operand. We want to eliminate dead instructions after, so that 1230 // we see fewer uses of the copies. We then need to clean up the dead 1231 // instructions leftover after the operands are folded as well. 1232 // 1233 // XXX - Can we get away without running DeadMachineInstructionElim again? 1234 addPass(&SIFoldOperandsID); 1235 if (EnableDPPCombine) 1236 addPass(&GCNDPPCombineID); 1237 addPass(&SILoadStoreOptimizerID); 1238 if (isPassEnabled(EnableSDWAPeephole)) { 1239 addPass(&SIPeepholeSDWAID); 1240 addPass(&EarlyMachineLICMID); 1241 addPass(&MachineCSEID); 1242 addPass(&SIFoldOperandsID); 1243 } 1244 addPass(&DeadMachineInstructionElimID); 1245 addPass(createSIShrinkInstructionsPass()); 1246 } 1247 1248 bool GCNPassConfig::addILPOpts() { 1249 if (EnableEarlyIfConversion) 1250 addPass(&EarlyIfConverterID); 1251 1252 TargetPassConfig::addILPOpts(); 1253 return false; 1254 } 1255 1256 bool GCNPassConfig::addInstSelector() { 1257 AMDGPUPassConfig::addInstSelector(); 1258 addPass(&SIFixSGPRCopiesID); 1259 addPass(createSILowerI1CopiesPass()); 1260 return false; 1261 } 1262 1263 bool GCNPassConfig::addIRTranslator() { 1264 addPass(new IRTranslator(getOptLevel())); 1265 return false; 1266 } 1267 1268 void GCNPassConfig::addPreLegalizeMachineIR() { 1269 bool IsOptNone = getOptLevel() == CodeGenOpt::None; 1270 addPass(createAMDGPUPreLegalizeCombiner(IsOptNone)); 1271 addPass(new Localizer()); 1272 } 1273 1274 bool GCNPassConfig::addLegalizeMachineIR() { 1275 addPass(new Legalizer()); 1276 return false; 1277 } 1278 1279 void GCNPassConfig::addPreRegBankSelect() { 1280 bool IsOptNone = getOptLevel() == CodeGenOpt::None; 1281 addPass(createAMDGPUPostLegalizeCombiner(IsOptNone)); 1282 } 1283 1284 bool GCNPassConfig::addRegBankSelect() { 1285 addPass(new RegBankSelect()); 1286 return false; 1287 } 1288 1289 void GCNPassConfig::addPreGlobalInstructionSelect() { 1290 bool IsOptNone = getOptLevel() == CodeGenOpt::None; 1291 addPass(createAMDGPURegBankCombiner(IsOptNone)); 1292 } 1293 1294 bool GCNPassConfig::addGlobalInstructionSelect() { 1295 addPass(new InstructionSelect(getOptLevel())); 1296 return false; 1297 } 1298 1299 void GCNPassConfig::addPreRegAlloc() { 1300 if (LateCFGStructurize) { 1301 addPass(createAMDGPUMachineCFGStructurizerPass()); 1302 } 1303 } 1304 1305 void GCNPassConfig::addFastRegAlloc() { 1306 // FIXME: We have to disable the verifier here because of PHIElimination + 1307 // TwoAddressInstructions disabling it. 1308 1309 // This must be run immediately after phi elimination and before 1310 // TwoAddressInstructions, otherwise the processing of the tied operand of 1311 // SI_ELSE will introduce a copy of the tied operand source after the else. 1312 insertPass(&PHIEliminationID, &SILowerControlFlowID, false); 1313 1314 insertPass(&TwoAddressInstructionPassID, &SIWholeQuadModeID); 1315 insertPass(&TwoAddressInstructionPassID, &SIPreAllocateWWMRegsID); 1316 1317 TargetPassConfig::addFastRegAlloc(); 1318 } 1319 1320 void GCNPassConfig::addOptimizedRegAlloc() { 1321 // Allow the scheduler to run before SIWholeQuadMode inserts exec manipulation 1322 // instructions that cause scheduling barriers. 1323 insertPass(&MachineSchedulerID, &SIWholeQuadModeID); 1324 insertPass(&MachineSchedulerID, &SIPreAllocateWWMRegsID); 1325 1326 if (OptExecMaskPreRA) 1327 insertPass(&MachineSchedulerID, &SIOptimizeExecMaskingPreRAID); 1328 1329 if (isPassEnabled(EnablePreRAOptimizations)) 1330 insertPass(&RenameIndependentSubregsID, &GCNPreRAOptimizationsID); 1331 1332 // This is not an essential optimization and it has a noticeable impact on 1333 // compilation time, so we only enable it from O2. 1334 if (TM->getOptLevel() > CodeGenOpt::Less) 1335 insertPass(&MachineSchedulerID, &SIFormMemoryClausesID); 1336 1337 // FIXME: when an instruction has a Killed operand, and the instruction is 1338 // inside a bundle, seems only the BUNDLE instruction appears as the Kills of 1339 // the register in LiveVariables, this would trigger a failure in verifier, 1340 // we should fix it and enable the verifier. 1341 if (OptVGPRLiveRange) 1342 insertPass(&LiveVariablesID, &SIOptimizeVGPRLiveRangeID, false); 1343 // This must be run immediately after phi elimination and before 1344 // TwoAddressInstructions, otherwise the processing of the tied operand of 1345 // SI_ELSE will introduce a copy of the tied operand source after the else. 1346 insertPass(&PHIEliminationID, &SILowerControlFlowID, false); 1347 1348 if (EnableDCEInRA) 1349 insertPass(&DetectDeadLanesID, &DeadMachineInstructionElimID); 1350 1351 TargetPassConfig::addOptimizedRegAlloc(); 1352 } 1353 1354 bool GCNPassConfig::addPreRewrite() { 1355 if (EnableRegReassign) 1356 addPass(&GCNNSAReassignID); 1357 return true; 1358 } 1359 1360 FunctionPass *GCNPassConfig::createSGPRAllocPass(bool Optimized) { 1361 // Initialize the global default. 1362 llvm::call_once(InitializeDefaultSGPRRegisterAllocatorFlag, 1363 initializeDefaultSGPRRegisterAllocatorOnce); 1364 1365 RegisterRegAlloc::FunctionPassCtor Ctor = SGPRRegisterRegAlloc::getDefault(); 1366 if (Ctor != useDefaultRegisterAllocator) 1367 return Ctor(); 1368 1369 if (Optimized) 1370 return createGreedyRegisterAllocator(onlyAllocateSGPRs); 1371 1372 return createFastRegisterAllocator(onlyAllocateSGPRs, false); 1373 } 1374 1375 FunctionPass *GCNPassConfig::createVGPRAllocPass(bool Optimized) { 1376 // Initialize the global default. 1377 llvm::call_once(InitializeDefaultVGPRRegisterAllocatorFlag, 1378 initializeDefaultVGPRRegisterAllocatorOnce); 1379 1380 RegisterRegAlloc::FunctionPassCtor Ctor = VGPRRegisterRegAlloc::getDefault(); 1381 if (Ctor != useDefaultRegisterAllocator) 1382 return Ctor(); 1383 1384 if (Optimized) 1385 return createGreedyVGPRRegisterAllocator(); 1386 1387 return createFastVGPRRegisterAllocator(); 1388 } 1389 1390 FunctionPass *GCNPassConfig::createRegAllocPass(bool Optimized) { 1391 llvm_unreachable("should not be used"); 1392 } 1393 1394 static const char RegAllocOptNotSupportedMessage[] = 1395 "-regalloc not supported with amdgcn. Use -sgpr-regalloc and -vgpr-regalloc"; 1396 1397 bool GCNPassConfig::addRegAssignAndRewriteFast() { 1398 if (!usingDefaultRegAlloc()) 1399 report_fatal_error(RegAllocOptNotSupportedMessage); 1400 1401 addPass(createSGPRAllocPass(false)); 1402 1403 // Equivalent of PEI for SGPRs. 1404 addPass(&SILowerSGPRSpillsID); 1405 1406 addPass(createVGPRAllocPass(false)); 1407 return true; 1408 } 1409 1410 bool GCNPassConfig::addRegAssignAndRewriteOptimized() { 1411 if (!usingDefaultRegAlloc()) 1412 report_fatal_error(RegAllocOptNotSupportedMessage); 1413 1414 addPass(createSGPRAllocPass(true)); 1415 1416 // Commit allocated register changes. This is mostly necessary because too 1417 // many things rely on the use lists of the physical registers, such as the 1418 // verifier. This is only necessary with allocators which use LiveIntervals, 1419 // since FastRegAlloc does the replacments itself. 1420 addPass(createVirtRegRewriter(false)); 1421 1422 // Equivalent of PEI for SGPRs. 1423 addPass(&SILowerSGPRSpillsID); 1424 1425 addPass(createVGPRAllocPass(true)); 1426 1427 addPreRewrite(); 1428 addPass(&VirtRegRewriterID); 1429 1430 return true; 1431 } 1432 1433 void GCNPassConfig::addPostRegAlloc() { 1434 addPass(&SIFixVGPRCopiesID); 1435 if (getOptLevel() > CodeGenOpt::None) 1436 addPass(&SIOptimizeExecMaskingID); 1437 TargetPassConfig::addPostRegAlloc(); 1438 } 1439 1440 void GCNPassConfig::addPreSched2() { 1441 addPass(&SIPostRABundlerID); 1442 } 1443 1444 void GCNPassConfig::addPreEmitPass() { 1445 addPass(createSIMemoryLegalizerPass()); 1446 addPass(createSIInsertWaitcntsPass()); 1447 1448 if (TM->getOptLevel() > CodeGenOpt::None) 1449 addPass(createSIShrinkInstructionsPass()); 1450 1451 addPass(createSIModeRegisterPass()); 1452 1453 if (getOptLevel() > CodeGenOpt::None) 1454 addPass(&SIInsertHardClausesID); 1455 1456 addPass(&SILateBranchLoweringPassID); 1457 if (getOptLevel() > CodeGenOpt::None) 1458 addPass(&SIPreEmitPeepholeID); 1459 // The hazard recognizer that runs as part of the post-ra scheduler does not 1460 // guarantee to be able handle all hazards correctly. This is because if there 1461 // are multiple scheduling regions in a basic block, the regions are scheduled 1462 // bottom up, so when we begin to schedule a region we don't know what 1463 // instructions were emitted directly before it. 1464 // 1465 // Here we add a stand-alone hazard recognizer pass which can handle all 1466 // cases. 1467 addPass(&PostRAHazardRecognizerID); 1468 addPass(&BranchRelaxationPassID); 1469 } 1470 1471 TargetPassConfig *GCNTargetMachine::createPassConfig(PassManagerBase &PM) { 1472 return new GCNPassConfig(*this, PM); 1473 } 1474 1475 yaml::MachineFunctionInfo *GCNTargetMachine::createDefaultFuncInfoYAML() const { 1476 return new yaml::SIMachineFunctionInfo(); 1477 } 1478 1479 yaml::MachineFunctionInfo * 1480 GCNTargetMachine::convertFuncInfoToYAML(const MachineFunction &MF) const { 1481 const SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>(); 1482 return new yaml::SIMachineFunctionInfo( 1483 *MFI, *MF.getSubtarget().getRegisterInfo(), MF); 1484 } 1485 1486 bool GCNTargetMachine::parseMachineFunctionInfo( 1487 const yaml::MachineFunctionInfo &MFI_, PerFunctionMIParsingState &PFS, 1488 SMDiagnostic &Error, SMRange &SourceRange) const { 1489 const yaml::SIMachineFunctionInfo &YamlMFI = 1490 reinterpret_cast<const yaml::SIMachineFunctionInfo &>(MFI_); 1491 MachineFunction &MF = PFS.MF; 1492 SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>(); 1493 1494 if (MFI->initializeBaseYamlFields(YamlMFI, MF, PFS, Error, SourceRange)) 1495 return true; 1496 1497 if (MFI->Occupancy == 0) { 1498 // Fixup the subtarget dependent default value. 1499 const GCNSubtarget &ST = MF.getSubtarget<GCNSubtarget>(); 1500 MFI->Occupancy = ST.computeOccupancy(MF.getFunction(), MFI->getLDSSize()); 1501 } 1502 1503 auto parseRegister = [&](const yaml::StringValue &RegName, Register &RegVal) { 1504 Register TempReg; 1505 if (parseNamedRegisterReference(PFS, TempReg, RegName.Value, Error)) { 1506 SourceRange = RegName.SourceRange; 1507 return true; 1508 } 1509 RegVal = TempReg; 1510 1511 return false; 1512 }; 1513 1514 auto diagnoseRegisterClass = [&](const yaml::StringValue &RegName) { 1515 // Create a diagnostic for a the register string literal. 1516 const MemoryBuffer &Buffer = 1517 *PFS.SM->getMemoryBuffer(PFS.SM->getMainFileID()); 1518 Error = SMDiagnostic(*PFS.SM, SMLoc(), Buffer.getBufferIdentifier(), 1, 1519 RegName.Value.size(), SourceMgr::DK_Error, 1520 "incorrect register class for field", RegName.Value, 1521 None, None); 1522 SourceRange = RegName.SourceRange; 1523 return true; 1524 }; 1525 1526 if (parseRegister(YamlMFI.ScratchRSrcReg, MFI->ScratchRSrcReg) || 1527 parseRegister(YamlMFI.FrameOffsetReg, MFI->FrameOffsetReg) || 1528 parseRegister(YamlMFI.StackPtrOffsetReg, MFI->StackPtrOffsetReg)) 1529 return true; 1530 1531 if (MFI->ScratchRSrcReg != AMDGPU::PRIVATE_RSRC_REG && 1532 !AMDGPU::SGPR_128RegClass.contains(MFI->ScratchRSrcReg)) { 1533 return diagnoseRegisterClass(YamlMFI.ScratchRSrcReg); 1534 } 1535 1536 if (MFI->FrameOffsetReg != AMDGPU::FP_REG && 1537 !AMDGPU::SGPR_32RegClass.contains(MFI->FrameOffsetReg)) { 1538 return diagnoseRegisterClass(YamlMFI.FrameOffsetReg); 1539 } 1540 1541 if (MFI->StackPtrOffsetReg != AMDGPU::SP_REG && 1542 !AMDGPU::SGPR_32RegClass.contains(MFI->StackPtrOffsetReg)) { 1543 return diagnoseRegisterClass(YamlMFI.StackPtrOffsetReg); 1544 } 1545 1546 auto parseAndCheckArgument = [&](const Optional<yaml::SIArgument> &A, 1547 const TargetRegisterClass &RC, 1548 ArgDescriptor &Arg, unsigned UserSGPRs, 1549 unsigned SystemSGPRs) { 1550 // Skip parsing if it's not present. 1551 if (!A) 1552 return false; 1553 1554 if (A->IsRegister) { 1555 Register Reg; 1556 if (parseNamedRegisterReference(PFS, Reg, A->RegisterName.Value, Error)) { 1557 SourceRange = A->RegisterName.SourceRange; 1558 return true; 1559 } 1560 if (!RC.contains(Reg)) 1561 return diagnoseRegisterClass(A->RegisterName); 1562 Arg = ArgDescriptor::createRegister(Reg); 1563 } else 1564 Arg = ArgDescriptor::createStack(A->StackOffset); 1565 // Check and apply the optional mask. 1566 if (A->Mask) 1567 Arg = ArgDescriptor::createArg(Arg, A->Mask.getValue()); 1568 1569 MFI->NumUserSGPRs += UserSGPRs; 1570 MFI->NumSystemSGPRs += SystemSGPRs; 1571 return false; 1572 }; 1573 1574 if (YamlMFI.ArgInfo && 1575 (parseAndCheckArgument(YamlMFI.ArgInfo->PrivateSegmentBuffer, 1576 AMDGPU::SGPR_128RegClass, 1577 MFI->ArgInfo.PrivateSegmentBuffer, 4, 0) || 1578 parseAndCheckArgument(YamlMFI.ArgInfo->DispatchPtr, 1579 AMDGPU::SReg_64RegClass, MFI->ArgInfo.DispatchPtr, 1580 2, 0) || 1581 parseAndCheckArgument(YamlMFI.ArgInfo->QueuePtr, AMDGPU::SReg_64RegClass, 1582 MFI->ArgInfo.QueuePtr, 2, 0) || 1583 parseAndCheckArgument(YamlMFI.ArgInfo->KernargSegmentPtr, 1584 AMDGPU::SReg_64RegClass, 1585 MFI->ArgInfo.KernargSegmentPtr, 2, 0) || 1586 parseAndCheckArgument(YamlMFI.ArgInfo->DispatchID, 1587 AMDGPU::SReg_64RegClass, MFI->ArgInfo.DispatchID, 1588 2, 0) || 1589 parseAndCheckArgument(YamlMFI.ArgInfo->FlatScratchInit, 1590 AMDGPU::SReg_64RegClass, 1591 MFI->ArgInfo.FlatScratchInit, 2, 0) || 1592 parseAndCheckArgument(YamlMFI.ArgInfo->PrivateSegmentSize, 1593 AMDGPU::SGPR_32RegClass, 1594 MFI->ArgInfo.PrivateSegmentSize, 0, 0) || 1595 parseAndCheckArgument(YamlMFI.ArgInfo->WorkGroupIDX, 1596 AMDGPU::SGPR_32RegClass, MFI->ArgInfo.WorkGroupIDX, 1597 0, 1) || 1598 parseAndCheckArgument(YamlMFI.ArgInfo->WorkGroupIDY, 1599 AMDGPU::SGPR_32RegClass, MFI->ArgInfo.WorkGroupIDY, 1600 0, 1) || 1601 parseAndCheckArgument(YamlMFI.ArgInfo->WorkGroupIDZ, 1602 AMDGPU::SGPR_32RegClass, MFI->ArgInfo.WorkGroupIDZ, 1603 0, 1) || 1604 parseAndCheckArgument(YamlMFI.ArgInfo->WorkGroupInfo, 1605 AMDGPU::SGPR_32RegClass, 1606 MFI->ArgInfo.WorkGroupInfo, 0, 1) || 1607 parseAndCheckArgument(YamlMFI.ArgInfo->PrivateSegmentWaveByteOffset, 1608 AMDGPU::SGPR_32RegClass, 1609 MFI->ArgInfo.PrivateSegmentWaveByteOffset, 0, 1) || 1610 parseAndCheckArgument(YamlMFI.ArgInfo->ImplicitArgPtr, 1611 AMDGPU::SReg_64RegClass, 1612 MFI->ArgInfo.ImplicitArgPtr, 0, 0) || 1613 parseAndCheckArgument(YamlMFI.ArgInfo->ImplicitBufferPtr, 1614 AMDGPU::SReg_64RegClass, 1615 MFI->ArgInfo.ImplicitBufferPtr, 2, 0) || 1616 parseAndCheckArgument(YamlMFI.ArgInfo->WorkItemIDX, 1617 AMDGPU::VGPR_32RegClass, 1618 MFI->ArgInfo.WorkItemIDX, 0, 0) || 1619 parseAndCheckArgument(YamlMFI.ArgInfo->WorkItemIDY, 1620 AMDGPU::VGPR_32RegClass, 1621 MFI->ArgInfo.WorkItemIDY, 0, 0) || 1622 parseAndCheckArgument(YamlMFI.ArgInfo->WorkItemIDZ, 1623 AMDGPU::VGPR_32RegClass, 1624 MFI->ArgInfo.WorkItemIDZ, 0, 0))) 1625 return true; 1626 1627 MFI->Mode.IEEE = YamlMFI.Mode.IEEE; 1628 MFI->Mode.DX10Clamp = YamlMFI.Mode.DX10Clamp; 1629 MFI->Mode.FP32InputDenormals = YamlMFI.Mode.FP32InputDenormals; 1630 MFI->Mode.FP32OutputDenormals = YamlMFI.Mode.FP32OutputDenormals; 1631 MFI->Mode.FP64FP16InputDenormals = YamlMFI.Mode.FP64FP16InputDenormals; 1632 MFI->Mode.FP64FP16OutputDenormals = YamlMFI.Mode.FP64FP16OutputDenormals; 1633 1634 return false; 1635 } 1636