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 "AMDGPUCallLowering.h" 19 #include "AMDGPUInstructionSelector.h" 20 #include "AMDGPULegalizerInfo.h" 21 #include "AMDGPUMacroFusion.h" 22 #include "AMDGPUTargetObjectFile.h" 23 #include "AMDGPUTargetTransformInfo.h" 24 #include "GCNIterativeScheduler.h" 25 #include "GCNSchedStrategy.h" 26 #include "R600MachineScheduler.h" 27 #include "SIMachineFunctionInfo.h" 28 #include "SIMachineScheduler.h" 29 #include "TargetInfo/AMDGPUTargetInfo.h" 30 #include "llvm/CodeGen/GlobalISel/IRTranslator.h" 31 #include "llvm/CodeGen/GlobalISel/InstructionSelect.h" 32 #include "llvm/CodeGen/GlobalISel/Legalizer.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/TargetPassConfig.h" 37 #include "llvm/IR/Attributes.h" 38 #include "llvm/IR/Function.h" 39 #include "llvm/IR/LegacyPassManager.h" 40 #include "llvm/Pass.h" 41 #include "llvm/Support/CommandLine.h" 42 #include "llvm/Support/Compiler.h" 43 #include "llvm/Support/TargetRegistry.h" 44 #include "llvm/Target/TargetLoweringObjectFile.h" 45 #include "llvm/Transforms/IPO.h" 46 #include "llvm/Transforms/IPO/AlwaysInliner.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/Utils.h" 51 #include "llvm/Transforms/Vectorize.h" 52 #include <memory> 53 54 using namespace llvm; 55 56 static cl::opt<bool> EnableR600StructurizeCFG( 57 "r600-ir-structurize", 58 cl::desc("Use StructurizeCFG IR pass"), 59 cl::init(true)); 60 61 static cl::opt<bool> EnableSROA( 62 "amdgpu-sroa", 63 cl::desc("Run SROA after promote alloca pass"), 64 cl::ReallyHidden, 65 cl::init(true)); 66 67 static cl::opt<bool> 68 EnableEarlyIfConversion("amdgpu-early-ifcvt", cl::Hidden, 69 cl::desc("Run early if-conversion"), 70 cl::init(false)); 71 72 static cl::opt<bool> 73 OptExecMaskPreRA("amdgpu-opt-exec-mask-pre-ra", cl::Hidden, 74 cl::desc("Run pre-RA exec mask optimizations"), 75 cl::init(true)); 76 77 static cl::opt<bool> EnableR600IfConvert( 78 "r600-if-convert", 79 cl::desc("Use if conversion pass"), 80 cl::ReallyHidden, 81 cl::init(true)); 82 83 // Option to disable vectorizer for tests. 84 static cl::opt<bool> EnableLoadStoreVectorizer( 85 "amdgpu-load-store-vectorizer", 86 cl::desc("Enable load store vectorizer"), 87 cl::init(true), 88 cl::Hidden); 89 90 // Option to control global loads scalarization 91 static cl::opt<bool> ScalarizeGlobal( 92 "amdgpu-scalarize-global-loads", 93 cl::desc("Enable global load scalarization"), 94 cl::init(true), 95 cl::Hidden); 96 97 // Option to run internalize pass. 98 static cl::opt<bool> InternalizeSymbols( 99 "amdgpu-internalize-symbols", 100 cl::desc("Enable elimination of non-kernel functions and unused globals"), 101 cl::init(false), 102 cl::Hidden); 103 104 // Option to inline all early. 105 static cl::opt<bool> EarlyInlineAll( 106 "amdgpu-early-inline-all", 107 cl::desc("Inline all functions early"), 108 cl::init(false), 109 cl::Hidden); 110 111 static cl::opt<bool> EnableSDWAPeephole( 112 "amdgpu-sdwa-peephole", 113 cl::desc("Enable SDWA peepholer"), 114 cl::init(true)); 115 116 static cl::opt<bool> EnableDPPCombine( 117 "amdgpu-dpp-combine", 118 cl::desc("Enable DPP combiner"), 119 cl::init(true)); 120 121 // Enable address space based alias analysis 122 static cl::opt<bool> EnableAMDGPUAliasAnalysis("enable-amdgpu-aa", cl::Hidden, 123 cl::desc("Enable AMDGPU Alias Analysis"), 124 cl::init(true)); 125 126 // Option to run late CFG structurizer 127 static cl::opt<bool, true> LateCFGStructurize( 128 "amdgpu-late-structurize", 129 cl::desc("Enable late CFG structurization"), 130 cl::location(AMDGPUTargetMachine::EnableLateStructurizeCFG), 131 cl::Hidden); 132 133 static cl::opt<bool, true> EnableAMDGPUFunctionCallsOpt( 134 "amdgpu-function-calls", 135 cl::desc("Enable AMDGPU function call support"), 136 cl::location(AMDGPUTargetMachine::EnableFunctionCalls), 137 cl::init(true), 138 cl::Hidden); 139 140 // Enable lib calls simplifications 141 static cl::opt<bool> EnableLibCallSimplify( 142 "amdgpu-simplify-libcall", 143 cl::desc("Enable amdgpu library simplifications"), 144 cl::init(true), 145 cl::Hidden); 146 147 static cl::opt<bool> EnableLowerKernelArguments( 148 "amdgpu-ir-lower-kernel-arguments", 149 cl::desc("Lower kernel argument loads in IR pass"), 150 cl::init(true), 151 cl::Hidden); 152 153 static cl::opt<bool> EnableRegReassign( 154 "amdgpu-reassign-regs", 155 cl::desc("Enable register reassign optimizations on gfx10+"), 156 cl::init(true), 157 cl::Hidden); 158 159 // Enable atomic optimization 160 static cl::opt<bool> EnableAtomicOptimizations( 161 "amdgpu-atomic-optimizations", 162 cl::desc("Enable atomic optimizations"), 163 cl::init(false), 164 cl::Hidden); 165 166 // Enable Mode register optimization 167 static cl::opt<bool> EnableSIModeRegisterPass( 168 "amdgpu-mode-register", 169 cl::desc("Enable mode register pass"), 170 cl::init(true), 171 cl::Hidden); 172 173 // Option is used in lit tests to prevent deadcoding of patterns inspected. 174 static cl::opt<bool> 175 EnableDCEInRA("amdgpu-dce-in-ra", 176 cl::init(true), cl::Hidden, 177 cl::desc("Enable machine DCE inside regalloc")); 178 179 static cl::opt<bool> EnableScalarIRPasses( 180 "amdgpu-scalar-ir-passes", 181 cl::desc("Enable scalar IR passes"), 182 cl::init(true), 183 cl::Hidden); 184 185 extern "C" void LLVMInitializeAMDGPUTarget() { 186 // Register the target 187 RegisterTargetMachine<R600TargetMachine> X(getTheAMDGPUTarget()); 188 RegisterTargetMachine<GCNTargetMachine> Y(getTheGCNTarget()); 189 190 PassRegistry *PR = PassRegistry::getPassRegistry(); 191 initializeR600ClauseMergePassPass(*PR); 192 initializeR600ControlFlowFinalizerPass(*PR); 193 initializeR600PacketizerPass(*PR); 194 initializeR600ExpandSpecialInstrsPassPass(*PR); 195 initializeR600VectorRegMergerPass(*PR); 196 initializeGlobalISel(*PR); 197 initializeAMDGPUDAGToDAGISelPass(*PR); 198 initializeGCNDPPCombinePass(*PR); 199 initializeSILowerI1CopiesPass(*PR); 200 initializeSIFixSGPRCopiesPass(*PR); 201 initializeSIFixVGPRCopiesPass(*PR); 202 initializeSIFixupVectorISelPass(*PR); 203 initializeSIFoldOperandsPass(*PR); 204 initializeSIPeepholeSDWAPass(*PR); 205 initializeSIShrinkInstructionsPass(*PR); 206 initializeSIOptimizeExecMaskingPreRAPass(*PR); 207 initializeSILoadStoreOptimizerPass(*PR); 208 initializeAMDGPUFixFunctionBitcastsPass(*PR); 209 initializeAMDGPUAlwaysInlinePass(*PR); 210 initializeAMDGPUAnnotateKernelFeaturesPass(*PR); 211 initializeAMDGPUAnnotateUniformValuesPass(*PR); 212 initializeAMDGPUArgumentUsageInfoPass(*PR); 213 initializeAMDGPUAtomicOptimizerPass(*PR); 214 initializeAMDGPULowerKernelArgumentsPass(*PR); 215 initializeAMDGPULowerKernelAttributesPass(*PR); 216 initializeAMDGPULowerIntrinsicsPass(*PR); 217 initializeAMDGPUOpenCLEnqueuedBlockLoweringPass(*PR); 218 initializeAMDGPUPromoteAllocaPass(*PR); 219 initializeAMDGPUCodeGenPreparePass(*PR); 220 initializeAMDGPUPropagateAttributesEarlyPass(*PR); 221 initializeAMDGPUPropagateAttributesLatePass(*PR); 222 initializeAMDGPURewriteOutArgumentsPass(*PR); 223 initializeAMDGPUUnifyMetadataPass(*PR); 224 initializeSIAnnotateControlFlowPass(*PR); 225 initializeSIInsertWaitcntsPass(*PR); 226 initializeSIModeRegisterPass(*PR); 227 initializeSIWholeQuadModePass(*PR); 228 initializeSILowerControlFlowPass(*PR); 229 initializeSIInsertSkipsPass(*PR); 230 initializeSIMemoryLegalizerPass(*PR); 231 initializeSIOptimizeExecMaskingPass(*PR); 232 initializeSIPreAllocateWWMRegsPass(*PR); 233 initializeSIFormMemoryClausesPass(*PR); 234 initializeAMDGPUUnifyDivergentExitNodesPass(*PR); 235 initializeAMDGPUAAWrapperPassPass(*PR); 236 initializeAMDGPUExternalAAWrapperPass(*PR); 237 initializeAMDGPUUseNativeCallsPass(*PR); 238 initializeAMDGPUSimplifyLibCallsPass(*PR); 239 initializeAMDGPUInlinerPass(*PR); 240 initializeGCNRegBankReassignPass(*PR); 241 initializeGCNNSAReassignPass(*PR); 242 } 243 244 static std::unique_ptr<TargetLoweringObjectFile> createTLOF(const Triple &TT) { 245 return llvm::make_unique<AMDGPUTargetObjectFile>(); 246 } 247 248 static ScheduleDAGInstrs *createR600MachineScheduler(MachineSchedContext *C) { 249 return new ScheduleDAGMILive(C, llvm::make_unique<R600SchedStrategy>()); 250 } 251 252 static ScheduleDAGInstrs *createSIMachineScheduler(MachineSchedContext *C) { 253 return new SIScheduleDAGMI(C); 254 } 255 256 static ScheduleDAGInstrs * 257 createGCNMaxOccupancyMachineScheduler(MachineSchedContext *C) { 258 ScheduleDAGMILive *DAG = 259 new GCNScheduleDAGMILive(C, make_unique<GCNMaxOccupancySchedStrategy>(C)); 260 DAG->addMutation(createLoadClusterDAGMutation(DAG->TII, DAG->TRI)); 261 DAG->addMutation(createStoreClusterDAGMutation(DAG->TII, DAG->TRI)); 262 DAG->addMutation(createAMDGPUMacroFusionDAGMutation()); 263 return DAG; 264 } 265 266 static ScheduleDAGInstrs * 267 createIterativeGCNMaxOccupancyMachineScheduler(MachineSchedContext *C) { 268 auto DAG = new GCNIterativeScheduler(C, 269 GCNIterativeScheduler::SCHEDULE_LEGACYMAXOCCUPANCY); 270 DAG->addMutation(createLoadClusterDAGMutation(DAG->TII, DAG->TRI)); 271 DAG->addMutation(createStoreClusterDAGMutation(DAG->TII, DAG->TRI)); 272 return DAG; 273 } 274 275 static ScheduleDAGInstrs *createMinRegScheduler(MachineSchedContext *C) { 276 return new GCNIterativeScheduler(C, 277 GCNIterativeScheduler::SCHEDULE_MINREGFORCED); 278 } 279 280 static ScheduleDAGInstrs * 281 createIterativeILPMachineScheduler(MachineSchedContext *C) { 282 auto DAG = new GCNIterativeScheduler(C, 283 GCNIterativeScheduler::SCHEDULE_ILP); 284 DAG->addMutation(createLoadClusterDAGMutation(DAG->TII, DAG->TRI)); 285 DAG->addMutation(createStoreClusterDAGMutation(DAG->TII, DAG->TRI)); 286 DAG->addMutation(createAMDGPUMacroFusionDAGMutation()); 287 return DAG; 288 } 289 290 static MachineSchedRegistry 291 R600SchedRegistry("r600", "Run R600's custom scheduler", 292 createR600MachineScheduler); 293 294 static MachineSchedRegistry 295 SISchedRegistry("si", "Run SI's custom scheduler", 296 createSIMachineScheduler); 297 298 static MachineSchedRegistry 299 GCNMaxOccupancySchedRegistry("gcn-max-occupancy", 300 "Run GCN scheduler to maximize occupancy", 301 createGCNMaxOccupancyMachineScheduler); 302 303 static MachineSchedRegistry 304 IterativeGCNMaxOccupancySchedRegistry("gcn-max-occupancy-experimental", 305 "Run GCN scheduler to maximize occupancy (experimental)", 306 createIterativeGCNMaxOccupancyMachineScheduler); 307 308 static MachineSchedRegistry 309 GCNMinRegSchedRegistry("gcn-minreg", 310 "Run GCN iterative scheduler for minimal register usage (experimental)", 311 createMinRegScheduler); 312 313 static MachineSchedRegistry 314 GCNILPSchedRegistry("gcn-ilp", 315 "Run GCN iterative scheduler for ILP scheduling (experimental)", 316 createIterativeILPMachineScheduler); 317 318 static StringRef computeDataLayout(const Triple &TT) { 319 if (TT.getArch() == Triple::r600) { 320 // 32-bit pointers. 321 return "e-p:32:32-i64:64-v16:16-v24:32-v32:32-v48:64-v96:128" 322 "-v192:256-v256:256-v512:512-v1024:1024-v2048:2048-n32:64-S32-A5"; 323 } 324 325 // 32-bit private, local, and region pointers. 64-bit global, constant and 326 // flat, non-integral buffer fat pointers. 327 return "e-p:64:64-p1:64:64-p2:32:32-p3:32:32-p4:64:64-p5:32:32-p6:32:32" 328 "-i64:64-v16:16-v24:32-v32:32-v48:64-v96:128" 329 "-v192:256-v256:256-v512:512-v1024:1024-v2048:2048-n32:64-S32-A5" 330 "-ni:7"; 331 } 332 333 LLVM_READNONE 334 static StringRef getGPUOrDefault(const Triple &TT, StringRef GPU) { 335 if (!GPU.empty()) 336 return GPU; 337 338 // Need to default to a target with flat support for HSA. 339 if (TT.getArch() == Triple::amdgcn) 340 return TT.getOS() == Triple::AMDHSA ? "generic-hsa" : "generic"; 341 342 return "r600"; 343 } 344 345 static Reloc::Model getEffectiveRelocModel(Optional<Reloc::Model> RM) { 346 // The AMDGPU toolchain only supports generating shared objects, so we 347 // must always use PIC. 348 return Reloc::PIC_; 349 } 350 351 AMDGPUTargetMachine::AMDGPUTargetMachine(const Target &T, const Triple &TT, 352 StringRef CPU, StringRef FS, 353 TargetOptions Options, 354 Optional<Reloc::Model> RM, 355 Optional<CodeModel::Model> CM, 356 CodeGenOpt::Level OptLevel) 357 : LLVMTargetMachine(T, computeDataLayout(TT), TT, getGPUOrDefault(TT, CPU), 358 FS, Options, getEffectiveRelocModel(RM), 359 getEffectiveCodeModel(CM, CodeModel::Small), OptLevel), 360 TLOF(createTLOF(getTargetTriple())) { 361 initAsmInfo(); 362 } 363 364 bool AMDGPUTargetMachine::EnableLateStructurizeCFG = false; 365 bool AMDGPUTargetMachine::EnableFunctionCalls = false; 366 367 AMDGPUTargetMachine::~AMDGPUTargetMachine() = default; 368 369 StringRef AMDGPUTargetMachine::getGPUName(const Function &F) const { 370 Attribute GPUAttr = F.getFnAttribute("target-cpu"); 371 return GPUAttr.hasAttribute(Attribute::None) ? 372 getTargetCPU() : GPUAttr.getValueAsString(); 373 } 374 375 StringRef AMDGPUTargetMachine::getFeatureString(const Function &F) const { 376 Attribute FSAttr = F.getFnAttribute("target-features"); 377 378 return FSAttr.hasAttribute(Attribute::None) ? 379 getTargetFeatureString() : 380 FSAttr.getValueAsString(); 381 } 382 383 /// Predicate for Internalize pass. 384 static bool mustPreserveGV(const GlobalValue &GV) { 385 if (const Function *F = dyn_cast<Function>(&GV)) 386 return F->isDeclaration() || AMDGPU::isEntryFunctionCC(F->getCallingConv()); 387 388 return !GV.use_empty(); 389 } 390 391 void AMDGPUTargetMachine::adjustPassManager(PassManagerBuilder &Builder) { 392 Builder.DivergentTarget = true; 393 394 bool EnableOpt = getOptLevel() > CodeGenOpt::None; 395 bool Internalize = InternalizeSymbols; 396 bool EarlyInline = EarlyInlineAll && EnableOpt && !EnableFunctionCalls; 397 bool AMDGPUAA = EnableAMDGPUAliasAnalysis && EnableOpt; 398 bool LibCallSimplify = EnableLibCallSimplify && EnableOpt; 399 400 if (EnableFunctionCalls) { 401 delete Builder.Inliner; 402 Builder.Inliner = createAMDGPUFunctionInliningPass(); 403 } 404 405 Builder.addExtension( 406 PassManagerBuilder::EP_ModuleOptimizerEarly, 407 [Internalize, EarlyInline, AMDGPUAA, this](const PassManagerBuilder &, 408 legacy::PassManagerBase &PM) { 409 if (AMDGPUAA) { 410 PM.add(createAMDGPUAAWrapperPass()); 411 PM.add(createAMDGPUExternalAAWrapperPass()); 412 } 413 PM.add(createAMDGPUUnifyMetadataPass()); 414 PM.add(createAMDGPUPropagateAttributesLatePass(this)); 415 if (Internalize) { 416 PM.add(createInternalizePass(mustPreserveGV)); 417 PM.add(createGlobalDCEPass()); 418 } 419 if (EarlyInline) 420 PM.add(createAMDGPUAlwaysInlinePass(false)); 421 }); 422 423 const auto &Opt = Options; 424 Builder.addExtension( 425 PassManagerBuilder::EP_EarlyAsPossible, 426 [AMDGPUAA, LibCallSimplify, &Opt, this](const PassManagerBuilder &, 427 legacy::PassManagerBase &PM) { 428 if (AMDGPUAA) { 429 PM.add(createAMDGPUAAWrapperPass()); 430 PM.add(createAMDGPUExternalAAWrapperPass()); 431 } 432 PM.add(llvm::createAMDGPUPropagateAttributesEarlyPass(this)); 433 PM.add(llvm::createAMDGPUUseNativeCallsPass()); 434 if (LibCallSimplify) 435 PM.add(llvm::createAMDGPUSimplifyLibCallsPass(Opt, this)); 436 }); 437 438 Builder.addExtension( 439 PassManagerBuilder::EP_CGSCCOptimizerLate, 440 [](const PassManagerBuilder &, legacy::PassManagerBase &PM) { 441 // Add infer address spaces pass to the opt pipeline after inlining 442 // but before SROA to increase SROA opportunities. 443 PM.add(createInferAddressSpacesPass()); 444 445 // This should run after inlining to have any chance of doing anything, 446 // and before other cleanup optimizations. 447 PM.add(createAMDGPULowerKernelAttributesPass()); 448 }); 449 } 450 451 //===----------------------------------------------------------------------===// 452 // R600 Target Machine (R600 -> Cayman) 453 //===----------------------------------------------------------------------===// 454 455 R600TargetMachine::R600TargetMachine(const Target &T, const Triple &TT, 456 StringRef CPU, StringRef FS, 457 TargetOptions Options, 458 Optional<Reloc::Model> RM, 459 Optional<CodeModel::Model> CM, 460 CodeGenOpt::Level OL, bool JIT) 461 : AMDGPUTargetMachine(T, TT, CPU, FS, Options, RM, CM, OL) { 462 setRequiresStructuredCFG(true); 463 464 // Override the default since calls aren't supported for r600. 465 if (EnableFunctionCalls && 466 EnableAMDGPUFunctionCallsOpt.getNumOccurrences() == 0) 467 EnableFunctionCalls = false; 468 } 469 470 const R600Subtarget *R600TargetMachine::getSubtargetImpl( 471 const Function &F) const { 472 StringRef GPU = getGPUName(F); 473 StringRef FS = getFeatureString(F); 474 475 SmallString<128> SubtargetKey(GPU); 476 SubtargetKey.append(FS); 477 478 auto &I = SubtargetMap[SubtargetKey]; 479 if (!I) { 480 // This needs to be done before we create a new subtarget since any 481 // creation will depend on the TM and the code generation flags on the 482 // function that reside in TargetOptions. 483 resetTargetOptions(F); 484 I = llvm::make_unique<R600Subtarget>(TargetTriple, GPU, FS, *this); 485 } 486 487 return I.get(); 488 } 489 490 TargetTransformInfo 491 R600TargetMachine::getTargetTransformInfo(const Function &F) { 492 return TargetTransformInfo(R600TTIImpl(this, F)); 493 } 494 495 //===----------------------------------------------------------------------===// 496 // GCN Target Machine (SI+) 497 //===----------------------------------------------------------------------===// 498 499 GCNTargetMachine::GCNTargetMachine(const Target &T, const Triple &TT, 500 StringRef CPU, StringRef FS, 501 TargetOptions Options, 502 Optional<Reloc::Model> RM, 503 Optional<CodeModel::Model> CM, 504 CodeGenOpt::Level OL, bool JIT) 505 : AMDGPUTargetMachine(T, TT, CPU, FS, Options, RM, CM, OL) {} 506 507 const GCNSubtarget *GCNTargetMachine::getSubtargetImpl(const Function &F) const { 508 StringRef GPU = getGPUName(F); 509 StringRef FS = getFeatureString(F); 510 511 SmallString<128> SubtargetKey(GPU); 512 SubtargetKey.append(FS); 513 514 auto &I = SubtargetMap[SubtargetKey]; 515 if (!I) { 516 // This needs to be done before we create a new subtarget since any 517 // creation will depend on the TM and the code generation flags on the 518 // function that reside in TargetOptions. 519 resetTargetOptions(F); 520 I = llvm::make_unique<GCNSubtarget>(TargetTriple, GPU, FS, *this); 521 } 522 523 I->setScalarizeGlobalBehavior(ScalarizeGlobal); 524 525 return I.get(); 526 } 527 528 TargetTransformInfo 529 GCNTargetMachine::getTargetTransformInfo(const Function &F) { 530 return TargetTransformInfo(GCNTTIImpl(this, F)); 531 } 532 533 //===----------------------------------------------------------------------===// 534 // AMDGPU Pass Setup 535 //===----------------------------------------------------------------------===// 536 537 namespace { 538 539 class AMDGPUPassConfig : public TargetPassConfig { 540 public: 541 AMDGPUPassConfig(LLVMTargetMachine &TM, PassManagerBase &PM) 542 : TargetPassConfig(TM, PM) { 543 // Exceptions and StackMaps are not supported, so these passes will never do 544 // anything. 545 disablePass(&StackMapLivenessID); 546 disablePass(&FuncletLayoutID); 547 } 548 549 AMDGPUTargetMachine &getAMDGPUTargetMachine() const { 550 return getTM<AMDGPUTargetMachine>(); 551 } 552 553 ScheduleDAGInstrs * 554 createMachineScheduler(MachineSchedContext *C) const override { 555 ScheduleDAGMILive *DAG = createGenericSchedLive(C); 556 DAG->addMutation(createLoadClusterDAGMutation(DAG->TII, DAG->TRI)); 557 DAG->addMutation(createStoreClusterDAGMutation(DAG->TII, DAG->TRI)); 558 return DAG; 559 } 560 561 void addEarlyCSEOrGVNPass(); 562 void addStraightLineScalarOptimizationPasses(); 563 void addIRPasses() override; 564 void addCodeGenPrepare() override; 565 bool addPreISel() override; 566 bool addInstSelector() override; 567 bool addGCPasses() override; 568 569 std::unique_ptr<CSEConfigBase> getCSEConfig() const override; 570 }; 571 572 std::unique_ptr<CSEConfigBase> AMDGPUPassConfig::getCSEConfig() const { 573 return getStandardCSEConfigForOpt(TM->getOptLevel()); 574 } 575 576 class R600PassConfig final : public AMDGPUPassConfig { 577 public: 578 R600PassConfig(LLVMTargetMachine &TM, PassManagerBase &PM) 579 : AMDGPUPassConfig(TM, PM) {} 580 581 ScheduleDAGInstrs *createMachineScheduler( 582 MachineSchedContext *C) const override { 583 return createR600MachineScheduler(C); 584 } 585 586 bool addPreISel() override; 587 bool addInstSelector() override; 588 void addPreRegAlloc() override; 589 void addPreSched2() override; 590 void addPreEmitPass() override; 591 }; 592 593 class GCNPassConfig final : public AMDGPUPassConfig { 594 public: 595 GCNPassConfig(LLVMTargetMachine &TM, PassManagerBase &PM) 596 : AMDGPUPassConfig(TM, PM) { 597 // It is necessary to know the register usage of the entire call graph. We 598 // allow calls without EnableAMDGPUFunctionCalls if they are marked 599 // noinline, so this is always required. 600 setRequiresCodeGenSCCOrder(true); 601 } 602 603 GCNTargetMachine &getGCNTargetMachine() const { 604 return getTM<GCNTargetMachine>(); 605 } 606 607 ScheduleDAGInstrs * 608 createMachineScheduler(MachineSchedContext *C) const override; 609 610 bool addPreISel() override; 611 void addMachineSSAOptimization() override; 612 bool addILPOpts() override; 613 bool addInstSelector() override; 614 bool addIRTranslator() override; 615 bool addLegalizeMachineIR() override; 616 bool addRegBankSelect() override; 617 bool addGlobalInstructionSelect() override; 618 void addFastRegAlloc() override; 619 void addOptimizedRegAlloc() override; 620 void addPreRegAlloc() override; 621 bool addPreRewrite() override; 622 void addPostRegAlloc() override; 623 void addPreSched2() override; 624 void addPreEmitPass() override; 625 }; 626 627 } // end anonymous namespace 628 629 void AMDGPUPassConfig::addEarlyCSEOrGVNPass() { 630 if (getOptLevel() == CodeGenOpt::Aggressive) 631 addPass(createGVNPass()); 632 else 633 addPass(createEarlyCSEPass()); 634 } 635 636 void AMDGPUPassConfig::addStraightLineScalarOptimizationPasses() { 637 addPass(createLICMPass()); 638 addPass(createSeparateConstOffsetFromGEPPass()); 639 addPass(createSpeculativeExecutionPass()); 640 // ReassociateGEPs exposes more opportunites for SLSR. See 641 // the example in reassociate-geps-and-slsr.ll. 642 addPass(createStraightLineStrengthReducePass()); 643 // SeparateConstOffsetFromGEP and SLSR creates common expressions which GVN or 644 // EarlyCSE can reuse. 645 addEarlyCSEOrGVNPass(); 646 // Run NaryReassociate after EarlyCSE/GVN to be more effective. 647 addPass(createNaryReassociatePass()); 648 // NaryReassociate on GEPs creates redundant common expressions, so run 649 // EarlyCSE after it. 650 addPass(createEarlyCSEPass()); 651 } 652 653 void AMDGPUPassConfig::addIRPasses() { 654 const AMDGPUTargetMachine &TM = getAMDGPUTargetMachine(); 655 656 // There is no reason to run these. 657 disablePass(&StackMapLivenessID); 658 disablePass(&FuncletLayoutID); 659 disablePass(&PatchableFunctionID); 660 661 // This must occur before inlining, as the inliner will not look through 662 // bitcast calls. 663 addPass(createAMDGPUFixFunctionBitcastsPass()); 664 665 // A call to propagate attributes pass in the backend in case opt was not run. 666 addPass(createAMDGPUPropagateAttributesEarlyPass(&TM)); 667 668 addPass(createAtomicExpandPass()); 669 670 671 addPass(createAMDGPULowerIntrinsicsPass()); 672 673 // Function calls are not supported, so make sure we inline everything. 674 addPass(createAMDGPUAlwaysInlinePass()); 675 addPass(createAlwaysInlinerLegacyPass()); 676 // We need to add the barrier noop pass, otherwise adding the function 677 // inlining pass will cause all of the PassConfigs passes to be run 678 // one function at a time, which means if we have a nodule with two 679 // functions, then we will generate code for the first function 680 // without ever running any passes on the second. 681 addPass(createBarrierNoopPass()); 682 683 if (TM.getTargetTriple().getArch() == Triple::amdgcn) { 684 // TODO: May want to move later or split into an early and late one. 685 686 addPass(createAMDGPUCodeGenPreparePass()); 687 } 688 689 // Handle uses of OpenCL image2d_t, image3d_t and sampler_t arguments. 690 if (TM.getTargetTriple().getArch() == Triple::r600) 691 addPass(createR600OpenCLImageTypeLoweringPass()); 692 693 // Replace OpenCL enqueued block function pointers with global variables. 694 addPass(createAMDGPUOpenCLEnqueuedBlockLoweringPass()); 695 696 if (TM.getOptLevel() > CodeGenOpt::None) { 697 addPass(createInferAddressSpacesPass()); 698 addPass(createAMDGPUPromoteAlloca()); 699 700 if (EnableSROA) 701 addPass(createSROAPass()); 702 703 if (EnableScalarIRPasses) 704 addStraightLineScalarOptimizationPasses(); 705 706 if (EnableAMDGPUAliasAnalysis) { 707 addPass(createAMDGPUAAWrapperPass()); 708 addPass(createExternalAAWrapperPass([](Pass &P, Function &, 709 AAResults &AAR) { 710 if (auto *WrapperPass = P.getAnalysisIfAvailable<AMDGPUAAWrapperPass>()) 711 AAR.addAAResult(WrapperPass->getResult()); 712 })); 713 } 714 } 715 716 TargetPassConfig::addIRPasses(); 717 718 // EarlyCSE is not always strong enough to clean up what LSR produces. For 719 // example, GVN can combine 720 // 721 // %0 = add %a, %b 722 // %1 = add %b, %a 723 // 724 // and 725 // 726 // %0 = shl nsw %a, 2 727 // %1 = shl %a, 2 728 // 729 // but EarlyCSE can do neither of them. 730 if (getOptLevel() != CodeGenOpt::None && EnableScalarIRPasses) 731 addEarlyCSEOrGVNPass(); 732 } 733 734 void AMDGPUPassConfig::addCodeGenPrepare() { 735 if (TM->getTargetTriple().getArch() == Triple::amdgcn) 736 addPass(createAMDGPUAnnotateKernelFeaturesPass()); 737 738 if (TM->getTargetTriple().getArch() == Triple::amdgcn && 739 EnableLowerKernelArguments) 740 addPass(createAMDGPULowerKernelArgumentsPass()); 741 742 TargetPassConfig::addCodeGenPrepare(); 743 744 if (EnableLoadStoreVectorizer) 745 addPass(createLoadStoreVectorizerPass()); 746 } 747 748 bool AMDGPUPassConfig::addPreISel() { 749 addPass(createLowerSwitchPass()); 750 addPass(createFlattenCFGPass()); 751 return false; 752 } 753 754 bool AMDGPUPassConfig::addInstSelector() { 755 // Defer the verifier until FinalizeISel. 756 addPass(createAMDGPUISelDag(&getAMDGPUTargetMachine(), getOptLevel()), false); 757 return false; 758 } 759 760 bool AMDGPUPassConfig::addGCPasses() { 761 // Do nothing. GC is not supported. 762 return false; 763 } 764 765 //===----------------------------------------------------------------------===// 766 // R600 Pass Setup 767 //===----------------------------------------------------------------------===// 768 769 bool R600PassConfig::addPreISel() { 770 AMDGPUPassConfig::addPreISel(); 771 772 if (EnableR600StructurizeCFG) 773 addPass(createStructurizeCFGPass()); 774 return false; 775 } 776 777 bool R600PassConfig::addInstSelector() { 778 addPass(createR600ISelDag(&getAMDGPUTargetMachine(), getOptLevel())); 779 return false; 780 } 781 782 void R600PassConfig::addPreRegAlloc() { 783 addPass(createR600VectorRegMerger()); 784 } 785 786 void R600PassConfig::addPreSched2() { 787 addPass(createR600EmitClauseMarkers(), false); 788 if (EnableR600IfConvert) 789 addPass(&IfConverterID, false); 790 addPass(createR600ClauseMergePass(), false); 791 } 792 793 void R600PassConfig::addPreEmitPass() { 794 addPass(createAMDGPUCFGStructurizerPass(), false); 795 addPass(createR600ExpandSpecialInstrsPass(), false); 796 addPass(&FinalizeMachineBundlesID, false); 797 addPass(createR600Packetizer(), false); 798 addPass(createR600ControlFlowFinalizer(), false); 799 } 800 801 TargetPassConfig *R600TargetMachine::createPassConfig(PassManagerBase &PM) { 802 return new R600PassConfig(*this, PM); 803 } 804 805 //===----------------------------------------------------------------------===// 806 // GCN Pass Setup 807 //===----------------------------------------------------------------------===// 808 809 ScheduleDAGInstrs *GCNPassConfig::createMachineScheduler( 810 MachineSchedContext *C) const { 811 const GCNSubtarget &ST = C->MF->getSubtarget<GCNSubtarget>(); 812 if (ST.enableSIScheduler()) 813 return createSIMachineScheduler(C); 814 return createGCNMaxOccupancyMachineScheduler(C); 815 } 816 817 bool GCNPassConfig::addPreISel() { 818 AMDGPUPassConfig::addPreISel(); 819 820 if (EnableAtomicOptimizations) { 821 addPass(createAMDGPUAtomicOptimizerPass()); 822 } 823 824 // FIXME: We need to run a pass to propagate the attributes when calls are 825 // supported. 826 827 // Merge divergent exit nodes. StructurizeCFG won't recognize the multi-exit 828 // regions formed by them. 829 addPass(&AMDGPUUnifyDivergentExitNodesID); 830 if (!LateCFGStructurize) { 831 addPass(createStructurizeCFGPass(true)); // true -> SkipUniformRegions 832 } 833 addPass(createSinkingPass()); 834 addPass(createAMDGPUAnnotateUniformValues()); 835 if (!LateCFGStructurize) { 836 addPass(createSIAnnotateControlFlowPass()); 837 } 838 839 return false; 840 } 841 842 void GCNPassConfig::addMachineSSAOptimization() { 843 TargetPassConfig::addMachineSSAOptimization(); 844 845 // We want to fold operands after PeepholeOptimizer has run (or as part of 846 // it), because it will eliminate extra copies making it easier to fold the 847 // real source operand. We want to eliminate dead instructions after, so that 848 // we see fewer uses of the copies. We then need to clean up the dead 849 // instructions leftover after the operands are folded as well. 850 // 851 // XXX - Can we get away without running DeadMachineInstructionElim again? 852 addPass(&SIFoldOperandsID); 853 if (EnableDPPCombine) 854 addPass(&GCNDPPCombineID); 855 addPass(&DeadMachineInstructionElimID); 856 addPass(&SILoadStoreOptimizerID); 857 if (EnableSDWAPeephole) { 858 addPass(&SIPeepholeSDWAID); 859 addPass(&EarlyMachineLICMID); 860 addPass(&MachineCSEID); 861 addPass(&SIFoldOperandsID); 862 addPass(&DeadMachineInstructionElimID); 863 } 864 addPass(createSIShrinkInstructionsPass()); 865 } 866 867 bool GCNPassConfig::addILPOpts() { 868 if (EnableEarlyIfConversion) 869 addPass(&EarlyIfConverterID); 870 871 TargetPassConfig::addILPOpts(); 872 return false; 873 } 874 875 bool GCNPassConfig::addInstSelector() { 876 AMDGPUPassConfig::addInstSelector(); 877 addPass(&SIFixSGPRCopiesID); 878 addPass(createSILowerI1CopiesPass()); 879 addPass(createSIFixupVectorISelPass()); 880 addPass(createSIAddIMGInitPass()); 881 return false; 882 } 883 884 bool GCNPassConfig::addIRTranslator() { 885 addPass(new IRTranslator()); 886 return false; 887 } 888 889 bool GCNPassConfig::addLegalizeMachineIR() { 890 addPass(new Legalizer()); 891 return false; 892 } 893 894 bool GCNPassConfig::addRegBankSelect() { 895 addPass(new RegBankSelect()); 896 return false; 897 } 898 899 bool GCNPassConfig::addGlobalInstructionSelect() { 900 addPass(new InstructionSelect()); 901 return false; 902 } 903 904 void GCNPassConfig::addPreRegAlloc() { 905 if (LateCFGStructurize) { 906 addPass(createAMDGPUMachineCFGStructurizerPass()); 907 } 908 addPass(createSIWholeQuadModePass()); 909 } 910 911 void GCNPassConfig::addFastRegAlloc() { 912 // FIXME: We have to disable the verifier here because of PHIElimination + 913 // TwoAddressInstructions disabling it. 914 915 // This must be run immediately after phi elimination and before 916 // TwoAddressInstructions, otherwise the processing of the tied operand of 917 // SI_ELSE will introduce a copy of the tied operand source after the else. 918 insertPass(&PHIEliminationID, &SILowerControlFlowID, false); 919 920 // This must be run just after RegisterCoalescing. 921 insertPass(&RegisterCoalescerID, &SIPreAllocateWWMRegsID, false); 922 923 TargetPassConfig::addFastRegAlloc(); 924 } 925 926 void GCNPassConfig::addOptimizedRegAlloc() { 927 if (OptExecMaskPreRA) { 928 insertPass(&MachineSchedulerID, &SIOptimizeExecMaskingPreRAID); 929 insertPass(&SIOptimizeExecMaskingPreRAID, &SIFormMemoryClausesID); 930 } else { 931 insertPass(&MachineSchedulerID, &SIFormMemoryClausesID); 932 } 933 934 // This must be run immediately after phi elimination and before 935 // TwoAddressInstructions, otherwise the processing of the tied operand of 936 // SI_ELSE will introduce a copy of the tied operand source after the else. 937 insertPass(&PHIEliminationID, &SILowerControlFlowID, false); 938 939 // This must be run just after RegisterCoalescing. 940 insertPass(&RegisterCoalescerID, &SIPreAllocateWWMRegsID, false); 941 942 if (EnableDCEInRA) 943 insertPass(&RenameIndependentSubregsID, &DeadMachineInstructionElimID); 944 945 TargetPassConfig::addOptimizedRegAlloc(); 946 } 947 948 bool GCNPassConfig::addPreRewrite() { 949 if (EnableRegReassign) { 950 addPass(&GCNNSAReassignID); 951 addPass(&GCNRegBankReassignID); 952 } 953 return true; 954 } 955 956 void GCNPassConfig::addPostRegAlloc() { 957 addPass(&SIFixVGPRCopiesID); 958 if (getOptLevel() > CodeGenOpt::None) 959 addPass(&SIOptimizeExecMaskingID); 960 TargetPassConfig::addPostRegAlloc(); 961 } 962 963 void GCNPassConfig::addPreSched2() { 964 } 965 966 void GCNPassConfig::addPreEmitPass() { 967 addPass(createSIMemoryLegalizerPass()); 968 addPass(createSIInsertWaitcntsPass()); 969 addPass(createSIShrinkInstructionsPass()); 970 addPass(createSIModeRegisterPass()); 971 972 // The hazard recognizer that runs as part of the post-ra scheduler does not 973 // guarantee to be able handle all hazards correctly. This is because if there 974 // are multiple scheduling regions in a basic block, the regions are scheduled 975 // bottom up, so when we begin to schedule a region we don't know what 976 // instructions were emitted directly before it. 977 // 978 // Here we add a stand-alone hazard recognizer pass which can handle all 979 // cases. 980 // 981 // FIXME: This stand-alone pass will emit indiv. S_NOP 0, as needed. It would 982 // be better for it to emit S_NOP <N> when possible. 983 addPass(&PostRAHazardRecognizerID); 984 985 addPass(&SIInsertSkipsPassID); 986 addPass(&BranchRelaxationPassID); 987 } 988 989 TargetPassConfig *GCNTargetMachine::createPassConfig(PassManagerBase &PM) { 990 return new GCNPassConfig(*this, PM); 991 } 992 993 yaml::MachineFunctionInfo *GCNTargetMachine::createDefaultFuncInfoYAML() const { 994 return new yaml::SIMachineFunctionInfo(); 995 } 996 997 yaml::MachineFunctionInfo * 998 GCNTargetMachine::convertFuncInfoToYAML(const MachineFunction &MF) const { 999 const SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>(); 1000 return new yaml::SIMachineFunctionInfo(*MFI, 1001 *MF.getSubtarget().getRegisterInfo()); 1002 } 1003 1004 bool GCNTargetMachine::parseMachineFunctionInfo( 1005 const yaml::MachineFunctionInfo &MFI_, PerFunctionMIParsingState &PFS, 1006 SMDiagnostic &Error, SMRange &SourceRange) const { 1007 const yaml::SIMachineFunctionInfo &YamlMFI = 1008 reinterpret_cast<const yaml::SIMachineFunctionInfo &>(MFI_); 1009 MachineFunction &MF = PFS.MF; 1010 SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>(); 1011 1012 MFI->initializeBaseYamlFields(YamlMFI); 1013 1014 auto parseRegister = [&](const yaml::StringValue &RegName, unsigned &RegVal) { 1015 if (parseNamedRegisterReference(PFS, RegVal, RegName.Value, Error)) { 1016 SourceRange = RegName.SourceRange; 1017 return true; 1018 } 1019 1020 return false; 1021 }; 1022 1023 auto diagnoseRegisterClass = [&](const yaml::StringValue &RegName) { 1024 // Create a diagnostic for a the register string literal. 1025 const MemoryBuffer &Buffer = 1026 *PFS.SM->getMemoryBuffer(PFS.SM->getMainFileID()); 1027 Error = SMDiagnostic(*PFS.SM, SMLoc(), Buffer.getBufferIdentifier(), 1, 1028 RegName.Value.size(), SourceMgr::DK_Error, 1029 "incorrect register class for field", RegName.Value, 1030 None, None); 1031 SourceRange = RegName.SourceRange; 1032 return true; 1033 }; 1034 1035 if (parseRegister(YamlMFI.ScratchRSrcReg, MFI->ScratchRSrcReg) || 1036 parseRegister(YamlMFI.ScratchWaveOffsetReg, MFI->ScratchWaveOffsetReg) || 1037 parseRegister(YamlMFI.FrameOffsetReg, MFI->FrameOffsetReg) || 1038 parseRegister(YamlMFI.StackPtrOffsetReg, MFI->StackPtrOffsetReg)) 1039 return true; 1040 1041 if (MFI->ScratchRSrcReg != AMDGPU::PRIVATE_RSRC_REG && 1042 !AMDGPU::SReg_128RegClass.contains(MFI->ScratchRSrcReg)) { 1043 return diagnoseRegisterClass(YamlMFI.ScratchRSrcReg); 1044 } 1045 1046 if (MFI->ScratchWaveOffsetReg != AMDGPU::SCRATCH_WAVE_OFFSET_REG && 1047 !AMDGPU::SGPR_32RegClass.contains(MFI->ScratchWaveOffsetReg)) { 1048 return diagnoseRegisterClass(YamlMFI.ScratchWaveOffsetReg); 1049 } 1050 1051 if (MFI->FrameOffsetReg != AMDGPU::FP_REG && 1052 !AMDGPU::SGPR_32RegClass.contains(MFI->FrameOffsetReg)) { 1053 return diagnoseRegisterClass(YamlMFI.FrameOffsetReg); 1054 } 1055 1056 if (MFI->StackPtrOffsetReg != AMDGPU::SP_REG && 1057 !AMDGPU::SGPR_32RegClass.contains(MFI->StackPtrOffsetReg)) { 1058 return diagnoseRegisterClass(YamlMFI.StackPtrOffsetReg); 1059 } 1060 1061 return false; 1062 } 1063