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 "llvm/CodeGen/GlobalISel/IRTranslator.h" 30 #include "llvm/CodeGen/GlobalISel/InstructionSelect.h" 31 #include "llvm/CodeGen/GlobalISel/Legalizer.h" 32 #include "llvm/CodeGen/GlobalISel/RegBankSelect.h" 33 #include "llvm/CodeGen/MIRParser/MIParser.h" 34 #include "llvm/CodeGen/Passes.h" 35 #include "llvm/CodeGen/TargetPassConfig.h" 36 #include "llvm/IR/Attributes.h" 37 #include "llvm/IR/Function.h" 38 #include "llvm/IR/LegacyPassManager.h" 39 #include "llvm/Pass.h" 40 #include "llvm/Support/CommandLine.h" 41 #include "llvm/Support/Compiler.h" 42 #include "llvm/Support/TargetRegistry.h" 43 #include "llvm/Target/TargetLoweringObjectFile.h" 44 #include "llvm/Transforms/IPO.h" 45 #include "llvm/Transforms/IPO/AlwaysInliner.h" 46 #include "llvm/Transforms/IPO/PassManagerBuilder.h" 47 #include "llvm/Transforms/Scalar.h" 48 #include "llvm/Transforms/Scalar/GVN.h" 49 #include "llvm/Transforms/Utils.h" 50 #include "llvm/Transforms/Vectorize.h" 51 #include <memory> 52 53 using namespace llvm; 54 55 static cl::opt<bool> EnableR600StructurizeCFG( 56 "r600-ir-structurize", 57 cl::desc("Use StructurizeCFG IR pass"), 58 cl::init(true)); 59 60 static cl::opt<bool> EnableSROA( 61 "amdgpu-sroa", 62 cl::desc("Run SROA after promote alloca pass"), 63 cl::ReallyHidden, 64 cl::init(true)); 65 66 static cl::opt<bool> 67 EnableEarlyIfConversion("amdgpu-early-ifcvt", cl::Hidden, 68 cl::desc("Run early if-conversion"), 69 cl::init(false)); 70 71 static cl::opt<bool> 72 OptExecMaskPreRA("amdgpu-opt-exec-mask-pre-ra", cl::Hidden, 73 cl::desc("Run pre-RA exec mask optimizations"), 74 cl::init(true)); 75 76 static cl::opt<bool> EnableR600IfConvert( 77 "r600-if-convert", 78 cl::desc("Use if conversion pass"), 79 cl::ReallyHidden, 80 cl::init(true)); 81 82 // Option to disable vectorizer for tests. 83 static cl::opt<bool> EnableLoadStoreVectorizer( 84 "amdgpu-load-store-vectorizer", 85 cl::desc("Enable load store vectorizer"), 86 cl::init(true), 87 cl::Hidden); 88 89 // Option to control global loads scalarization 90 static cl::opt<bool> ScalarizeGlobal( 91 "amdgpu-scalarize-global-loads", 92 cl::desc("Enable global load scalarization"), 93 cl::init(true), 94 cl::Hidden); 95 96 // Option to run internalize pass. 97 static cl::opt<bool> InternalizeSymbols( 98 "amdgpu-internalize-symbols", 99 cl::desc("Enable elimination of non-kernel functions and unused globals"), 100 cl::init(false), 101 cl::Hidden); 102 103 // Option to inline all early. 104 static cl::opt<bool> EarlyInlineAll( 105 "amdgpu-early-inline-all", 106 cl::desc("Inline all functions early"), 107 cl::init(false), 108 cl::Hidden); 109 110 static cl::opt<bool> EnableSDWAPeephole( 111 "amdgpu-sdwa-peephole", 112 cl::desc("Enable SDWA peepholer"), 113 cl::init(true)); 114 115 static cl::opt<bool> EnableDPPCombine( 116 "amdgpu-dpp-combine", 117 cl::desc("Enable DPP combiner"), 118 cl::init(true)); 119 120 // Enable address space based alias analysis 121 static cl::opt<bool> EnableAMDGPUAliasAnalysis("enable-amdgpu-aa", cl::Hidden, 122 cl::desc("Enable AMDGPU Alias Analysis"), 123 cl::init(true)); 124 125 // Option to run late CFG structurizer 126 static cl::opt<bool, true> LateCFGStructurize( 127 "amdgpu-late-structurize", 128 cl::desc("Enable late CFG structurization"), 129 cl::location(AMDGPUTargetMachine::EnableLateStructurizeCFG), 130 cl::Hidden); 131 132 static cl::opt<bool, true> EnableAMDGPUFunctionCallsOpt( 133 "amdgpu-function-calls", 134 cl::desc("Enable AMDGPU function call support"), 135 cl::location(AMDGPUTargetMachine::EnableFunctionCalls), 136 cl::init(true), 137 cl::Hidden); 138 139 // Enable lib calls simplifications 140 static cl::opt<bool> EnableLibCallSimplify( 141 "amdgpu-simplify-libcall", 142 cl::desc("Enable amdgpu library simplifications"), 143 cl::init(true), 144 cl::Hidden); 145 146 static cl::opt<bool> EnableLowerKernelArguments( 147 "amdgpu-ir-lower-kernel-arguments", 148 cl::desc("Lower kernel argument loads in IR pass"), 149 cl::init(true), 150 cl::Hidden); 151 152 // Enable atomic optimization 153 static cl::opt<bool> EnableAtomicOptimizations( 154 "amdgpu-atomic-optimizations", 155 cl::desc("Enable atomic optimizations"), 156 cl::init(false), 157 cl::Hidden); 158 159 // Enable Mode register optimization 160 static cl::opt<bool> EnableSIModeRegisterPass( 161 "amdgpu-mode-register", 162 cl::desc("Enable mode register pass"), 163 cl::init(true), 164 cl::Hidden); 165 166 // Option is used in lit tests to prevent deadcoding of patterns inspected. 167 static cl::opt<bool> 168 EnableDCEInRA("amdgpu-dce-in-ra", 169 cl::init(true), cl::Hidden, 170 cl::desc("Enable machine DCE inside regalloc")); 171 172 static cl::opt<bool> EnableScalarIRPasses( 173 "amdgpu-scalar-ir-passes", 174 cl::desc("Enable scalar IR passes"), 175 cl::init(true), 176 cl::Hidden); 177 178 extern "C" void LLVMInitializeAMDGPUTarget() { 179 // Register the target 180 RegisterTargetMachine<R600TargetMachine> X(getTheAMDGPUTarget()); 181 RegisterTargetMachine<GCNTargetMachine> Y(getTheGCNTarget()); 182 183 PassRegistry *PR = PassRegistry::getPassRegistry(); 184 initializeR600ClauseMergePassPass(*PR); 185 initializeR600ControlFlowFinalizerPass(*PR); 186 initializeR600PacketizerPass(*PR); 187 initializeR600ExpandSpecialInstrsPassPass(*PR); 188 initializeR600VectorRegMergerPass(*PR); 189 initializeGlobalISel(*PR); 190 initializeAMDGPUDAGToDAGISelPass(*PR); 191 initializeGCNDPPCombinePass(*PR); 192 initializeSILowerI1CopiesPass(*PR); 193 initializeSIFixSGPRCopiesPass(*PR); 194 initializeSIFixVGPRCopiesPass(*PR); 195 initializeSIFixupVectorISelPass(*PR); 196 initializeSIFoldOperandsPass(*PR); 197 initializeSIPeepholeSDWAPass(*PR); 198 initializeSIShrinkInstructionsPass(*PR); 199 initializeSIOptimizeExecMaskingPreRAPass(*PR); 200 initializeSILoadStoreOptimizerPass(*PR); 201 initializeAMDGPUFixFunctionBitcastsPass(*PR); 202 initializeAMDGPUAlwaysInlinePass(*PR); 203 initializeAMDGPUAnnotateKernelFeaturesPass(*PR); 204 initializeAMDGPUAnnotateUniformValuesPass(*PR); 205 initializeAMDGPUArgumentUsageInfoPass(*PR); 206 initializeAMDGPUAtomicOptimizerPass(*PR); 207 initializeAMDGPULowerKernelArgumentsPass(*PR); 208 initializeAMDGPULowerKernelAttributesPass(*PR); 209 initializeAMDGPULowerIntrinsicsPass(*PR); 210 initializeAMDGPUOpenCLEnqueuedBlockLoweringPass(*PR); 211 initializeAMDGPUPromoteAllocaPass(*PR); 212 initializeAMDGPUCodeGenPreparePass(*PR); 213 initializeAMDGPURewriteOutArgumentsPass(*PR); 214 initializeAMDGPUUnifyMetadataPass(*PR); 215 initializeSIAnnotateControlFlowPass(*PR); 216 initializeSIInsertWaitcntsPass(*PR); 217 initializeSIModeRegisterPass(*PR); 218 initializeSIWholeQuadModePass(*PR); 219 initializeSILowerControlFlowPass(*PR); 220 initializeSIInsertSkipsPass(*PR); 221 initializeSIMemoryLegalizerPass(*PR); 222 initializeSIOptimizeExecMaskingPass(*PR); 223 initializeSIPreAllocateWWMRegsPass(*PR); 224 initializeSIFormMemoryClausesPass(*PR); 225 initializeAMDGPUUnifyDivergentExitNodesPass(*PR); 226 initializeAMDGPUAAWrapperPassPass(*PR); 227 initializeAMDGPUExternalAAWrapperPass(*PR); 228 initializeAMDGPUUseNativeCallsPass(*PR); 229 initializeAMDGPUSimplifyLibCallsPass(*PR); 230 initializeAMDGPUInlinerPass(*PR); 231 } 232 233 static std::unique_ptr<TargetLoweringObjectFile> createTLOF(const Triple &TT) { 234 return llvm::make_unique<AMDGPUTargetObjectFile>(); 235 } 236 237 static ScheduleDAGInstrs *createR600MachineScheduler(MachineSchedContext *C) { 238 return new ScheduleDAGMILive(C, llvm::make_unique<R600SchedStrategy>()); 239 } 240 241 static ScheduleDAGInstrs *createSIMachineScheduler(MachineSchedContext *C) { 242 return new SIScheduleDAGMI(C); 243 } 244 245 static ScheduleDAGInstrs * 246 createGCNMaxOccupancyMachineScheduler(MachineSchedContext *C) { 247 ScheduleDAGMILive *DAG = 248 new GCNScheduleDAGMILive(C, make_unique<GCNMaxOccupancySchedStrategy>(C)); 249 DAG->addMutation(createLoadClusterDAGMutation(DAG->TII, DAG->TRI)); 250 DAG->addMutation(createStoreClusterDAGMutation(DAG->TII, DAG->TRI)); 251 DAG->addMutation(createAMDGPUMacroFusionDAGMutation()); 252 return DAG; 253 } 254 255 static ScheduleDAGInstrs * 256 createIterativeGCNMaxOccupancyMachineScheduler(MachineSchedContext *C) { 257 auto DAG = new GCNIterativeScheduler(C, 258 GCNIterativeScheduler::SCHEDULE_LEGACYMAXOCCUPANCY); 259 DAG->addMutation(createLoadClusterDAGMutation(DAG->TII, DAG->TRI)); 260 DAG->addMutation(createStoreClusterDAGMutation(DAG->TII, DAG->TRI)); 261 return DAG; 262 } 263 264 static ScheduleDAGInstrs *createMinRegScheduler(MachineSchedContext *C) { 265 return new GCNIterativeScheduler(C, 266 GCNIterativeScheduler::SCHEDULE_MINREGFORCED); 267 } 268 269 static ScheduleDAGInstrs * 270 createIterativeILPMachineScheduler(MachineSchedContext *C) { 271 auto DAG = new GCNIterativeScheduler(C, 272 GCNIterativeScheduler::SCHEDULE_ILP); 273 DAG->addMutation(createLoadClusterDAGMutation(DAG->TII, DAG->TRI)); 274 DAG->addMutation(createStoreClusterDAGMutation(DAG->TII, DAG->TRI)); 275 DAG->addMutation(createAMDGPUMacroFusionDAGMutation()); 276 return DAG; 277 } 278 279 static MachineSchedRegistry 280 R600SchedRegistry("r600", "Run R600's custom scheduler", 281 createR600MachineScheduler); 282 283 static MachineSchedRegistry 284 SISchedRegistry("si", "Run SI's custom scheduler", 285 createSIMachineScheduler); 286 287 static MachineSchedRegistry 288 GCNMaxOccupancySchedRegistry("gcn-max-occupancy", 289 "Run GCN scheduler to maximize occupancy", 290 createGCNMaxOccupancyMachineScheduler); 291 292 static MachineSchedRegistry 293 IterativeGCNMaxOccupancySchedRegistry("gcn-max-occupancy-experimental", 294 "Run GCN scheduler to maximize occupancy (experimental)", 295 createIterativeGCNMaxOccupancyMachineScheduler); 296 297 static MachineSchedRegistry 298 GCNMinRegSchedRegistry("gcn-minreg", 299 "Run GCN iterative scheduler for minimal register usage (experimental)", 300 createMinRegScheduler); 301 302 static MachineSchedRegistry 303 GCNILPSchedRegistry("gcn-ilp", 304 "Run GCN iterative scheduler for ILP scheduling (experimental)", 305 createIterativeILPMachineScheduler); 306 307 static StringRef computeDataLayout(const Triple &TT) { 308 if (TT.getArch() == Triple::r600) { 309 // 32-bit pointers. 310 return "e-p:32:32-i64:64-v16:16-v24:32-v32:32-v48:64-v96:128" 311 "-v192:256-v256:256-v512:512-v1024:1024-v2048:2048-n32:64-S32-A5"; 312 } 313 314 // 32-bit private, local, and region pointers. 64-bit global, constant and 315 // flat, non-integral buffer fat pointers. 316 return "e-p:64:64-p1:64:64-p2:32:32-p3:32:32-p4:64:64-p5:32:32-p6:32:32" 317 "-i64:64-v16:16-v24:32-v32:32-v48:64-v96:128" 318 "-v192:256-v256:256-v512:512-v1024:1024-v2048:2048-n32:64-S32-A5" 319 "-ni:7"; 320 } 321 322 LLVM_READNONE 323 static StringRef getGPUOrDefault(const Triple &TT, StringRef GPU) { 324 if (!GPU.empty()) 325 return GPU; 326 327 // Need to default to a target with flat support for HSA. 328 if (TT.getArch() == Triple::amdgcn) 329 return TT.getOS() == Triple::AMDHSA ? "generic-hsa" : "generic"; 330 331 return "r600"; 332 } 333 334 static Reloc::Model getEffectiveRelocModel(Optional<Reloc::Model> RM) { 335 // The AMDGPU toolchain only supports generating shared objects, so we 336 // must always use PIC. 337 return Reloc::PIC_; 338 } 339 340 AMDGPUTargetMachine::AMDGPUTargetMachine(const Target &T, const Triple &TT, 341 StringRef CPU, StringRef FS, 342 TargetOptions Options, 343 Optional<Reloc::Model> RM, 344 Optional<CodeModel::Model> CM, 345 CodeGenOpt::Level OptLevel) 346 : LLVMTargetMachine(T, computeDataLayout(TT), TT, getGPUOrDefault(TT, CPU), 347 FS, Options, getEffectiveRelocModel(RM), 348 getEffectiveCodeModel(CM, CodeModel::Small), OptLevel), 349 TLOF(createTLOF(getTargetTriple())) { 350 initAsmInfo(); 351 } 352 353 bool AMDGPUTargetMachine::EnableLateStructurizeCFG = false; 354 bool AMDGPUTargetMachine::EnableFunctionCalls = false; 355 356 AMDGPUTargetMachine::~AMDGPUTargetMachine() = default; 357 358 StringRef AMDGPUTargetMachine::getGPUName(const Function &F) const { 359 Attribute GPUAttr = F.getFnAttribute("target-cpu"); 360 return GPUAttr.hasAttribute(Attribute::None) ? 361 getTargetCPU() : GPUAttr.getValueAsString(); 362 } 363 364 StringRef AMDGPUTargetMachine::getFeatureString(const Function &F) const { 365 Attribute FSAttr = F.getFnAttribute("target-features"); 366 367 return FSAttr.hasAttribute(Attribute::None) ? 368 getTargetFeatureString() : 369 FSAttr.getValueAsString(); 370 } 371 372 /// Predicate for Internalize pass. 373 static bool mustPreserveGV(const GlobalValue &GV) { 374 if (const Function *F = dyn_cast<Function>(&GV)) 375 return F->isDeclaration() || AMDGPU::isEntryFunctionCC(F->getCallingConv()); 376 377 return !GV.use_empty(); 378 } 379 380 void AMDGPUTargetMachine::adjustPassManager(PassManagerBuilder &Builder) { 381 Builder.DivergentTarget = true; 382 383 bool EnableOpt = getOptLevel() > CodeGenOpt::None; 384 bool Internalize = InternalizeSymbols; 385 bool EarlyInline = EarlyInlineAll && EnableOpt && !EnableFunctionCalls; 386 bool AMDGPUAA = EnableAMDGPUAliasAnalysis && EnableOpt; 387 bool LibCallSimplify = EnableLibCallSimplify && EnableOpt; 388 389 if (EnableFunctionCalls) { 390 delete Builder.Inliner; 391 Builder.Inliner = createAMDGPUFunctionInliningPass(); 392 } 393 394 Builder.addExtension( 395 PassManagerBuilder::EP_ModuleOptimizerEarly, 396 [Internalize, EarlyInline, AMDGPUAA](const PassManagerBuilder &, 397 legacy::PassManagerBase &PM) { 398 if (AMDGPUAA) { 399 PM.add(createAMDGPUAAWrapperPass()); 400 PM.add(createAMDGPUExternalAAWrapperPass()); 401 } 402 PM.add(createAMDGPUUnifyMetadataPass()); 403 if (Internalize) { 404 PM.add(createInternalizePass(mustPreserveGV)); 405 PM.add(createGlobalDCEPass()); 406 } 407 if (EarlyInline) 408 PM.add(createAMDGPUAlwaysInlinePass(false)); 409 }); 410 411 const auto &Opt = Options; 412 Builder.addExtension( 413 PassManagerBuilder::EP_EarlyAsPossible, 414 [AMDGPUAA, LibCallSimplify, &Opt](const PassManagerBuilder &, 415 legacy::PassManagerBase &PM) { 416 if (AMDGPUAA) { 417 PM.add(createAMDGPUAAWrapperPass()); 418 PM.add(createAMDGPUExternalAAWrapperPass()); 419 } 420 PM.add(llvm::createAMDGPUUseNativeCallsPass()); 421 if (LibCallSimplify) 422 PM.add(llvm::createAMDGPUSimplifyLibCallsPass(Opt)); 423 }); 424 425 Builder.addExtension( 426 PassManagerBuilder::EP_CGSCCOptimizerLate, 427 [](const PassManagerBuilder &, legacy::PassManagerBase &PM) { 428 // Add infer address spaces pass to the opt pipeline after inlining 429 // but before SROA to increase SROA opportunities. 430 PM.add(createInferAddressSpacesPass()); 431 432 // This should run after inlining to have any chance of doing anything, 433 // and before other cleanup optimizations. 434 PM.add(createAMDGPULowerKernelAttributesPass()); 435 }); 436 } 437 438 //===----------------------------------------------------------------------===// 439 // R600 Target Machine (R600 -> Cayman) 440 //===----------------------------------------------------------------------===// 441 442 R600TargetMachine::R600TargetMachine(const Target &T, const Triple &TT, 443 StringRef CPU, StringRef FS, 444 TargetOptions Options, 445 Optional<Reloc::Model> RM, 446 Optional<CodeModel::Model> CM, 447 CodeGenOpt::Level OL, bool JIT) 448 : AMDGPUTargetMachine(T, TT, CPU, FS, Options, RM, CM, OL) { 449 setRequiresStructuredCFG(true); 450 451 // Override the default since calls aren't supported for r600. 452 if (EnableFunctionCalls && 453 EnableAMDGPUFunctionCallsOpt.getNumOccurrences() == 0) 454 EnableFunctionCalls = false; 455 } 456 457 const R600Subtarget *R600TargetMachine::getSubtargetImpl( 458 const Function &F) const { 459 StringRef GPU = getGPUName(F); 460 StringRef FS = getFeatureString(F); 461 462 SmallString<128> SubtargetKey(GPU); 463 SubtargetKey.append(FS); 464 465 auto &I = SubtargetMap[SubtargetKey]; 466 if (!I) { 467 // This needs to be done before we create a new subtarget since any 468 // creation will depend on the TM and the code generation flags on the 469 // function that reside in TargetOptions. 470 resetTargetOptions(F); 471 I = llvm::make_unique<R600Subtarget>(TargetTriple, GPU, FS, *this); 472 } 473 474 return I.get(); 475 } 476 477 TargetTransformInfo 478 R600TargetMachine::getTargetTransformInfo(const Function &F) { 479 return TargetTransformInfo(R600TTIImpl(this, F)); 480 } 481 482 //===----------------------------------------------------------------------===// 483 // GCN Target Machine (SI+) 484 //===----------------------------------------------------------------------===// 485 486 GCNTargetMachine::GCNTargetMachine(const Target &T, const Triple &TT, 487 StringRef CPU, StringRef FS, 488 TargetOptions Options, 489 Optional<Reloc::Model> RM, 490 Optional<CodeModel::Model> CM, 491 CodeGenOpt::Level OL, bool JIT) 492 : AMDGPUTargetMachine(T, TT, CPU, FS, Options, RM, CM, OL) {} 493 494 const GCNSubtarget *GCNTargetMachine::getSubtargetImpl(const Function &F) const { 495 StringRef GPU = getGPUName(F); 496 StringRef FS = getFeatureString(F); 497 498 SmallString<128> SubtargetKey(GPU); 499 SubtargetKey.append(FS); 500 501 auto &I = SubtargetMap[SubtargetKey]; 502 if (!I) { 503 // This needs to be done before we create a new subtarget since any 504 // creation will depend on the TM and the code generation flags on the 505 // function that reside in TargetOptions. 506 resetTargetOptions(F); 507 I = llvm::make_unique<GCNSubtarget>(TargetTriple, GPU, FS, *this); 508 } 509 510 I->setScalarizeGlobalBehavior(ScalarizeGlobal); 511 512 return I.get(); 513 } 514 515 TargetTransformInfo 516 GCNTargetMachine::getTargetTransformInfo(const Function &F) { 517 return TargetTransformInfo(GCNTTIImpl(this, F)); 518 } 519 520 //===----------------------------------------------------------------------===// 521 // AMDGPU Pass Setup 522 //===----------------------------------------------------------------------===// 523 524 namespace { 525 526 class AMDGPUPassConfig : public TargetPassConfig { 527 public: 528 AMDGPUPassConfig(LLVMTargetMachine &TM, PassManagerBase &PM) 529 : TargetPassConfig(TM, PM) { 530 // Exceptions and StackMaps are not supported, so these passes will never do 531 // anything. 532 disablePass(&StackMapLivenessID); 533 disablePass(&FuncletLayoutID); 534 } 535 536 AMDGPUTargetMachine &getAMDGPUTargetMachine() const { 537 return getTM<AMDGPUTargetMachine>(); 538 } 539 540 ScheduleDAGInstrs * 541 createMachineScheduler(MachineSchedContext *C) const override { 542 ScheduleDAGMILive *DAG = createGenericSchedLive(C); 543 DAG->addMutation(createLoadClusterDAGMutation(DAG->TII, DAG->TRI)); 544 DAG->addMutation(createStoreClusterDAGMutation(DAG->TII, DAG->TRI)); 545 return DAG; 546 } 547 548 void addEarlyCSEOrGVNPass(); 549 void addStraightLineScalarOptimizationPasses(); 550 void addIRPasses() override; 551 void addCodeGenPrepare() override; 552 bool addPreISel() override; 553 bool addInstSelector() override; 554 bool addGCPasses() override; 555 }; 556 557 class R600PassConfig final : public AMDGPUPassConfig { 558 public: 559 R600PassConfig(LLVMTargetMachine &TM, PassManagerBase &PM) 560 : AMDGPUPassConfig(TM, PM) {} 561 562 ScheduleDAGInstrs *createMachineScheduler( 563 MachineSchedContext *C) const override { 564 return createR600MachineScheduler(C); 565 } 566 567 bool addPreISel() override; 568 bool addInstSelector() override; 569 void addPreRegAlloc() override; 570 void addPreSched2() override; 571 void addPreEmitPass() override; 572 }; 573 574 class GCNPassConfig final : public AMDGPUPassConfig { 575 public: 576 GCNPassConfig(LLVMTargetMachine &TM, PassManagerBase &PM) 577 : AMDGPUPassConfig(TM, PM) { 578 // It is necessary to know the register usage of the entire call graph. We 579 // allow calls without EnableAMDGPUFunctionCalls if they are marked 580 // noinline, so this is always required. 581 setRequiresCodeGenSCCOrder(true); 582 } 583 584 GCNTargetMachine &getGCNTargetMachine() const { 585 return getTM<GCNTargetMachine>(); 586 } 587 588 ScheduleDAGInstrs * 589 createMachineScheduler(MachineSchedContext *C) const override; 590 591 bool addPreISel() override; 592 void addMachineSSAOptimization() override; 593 bool addILPOpts() override; 594 bool addInstSelector() override; 595 bool addIRTranslator() override; 596 bool addLegalizeMachineIR() override; 597 bool addRegBankSelect() override; 598 bool addGlobalInstructionSelect() override; 599 void addFastRegAlloc() override; 600 void addOptimizedRegAlloc() override; 601 void addPreRegAlloc() override; 602 void addPostRegAlloc() override; 603 void addPreSched2() override; 604 void addPreEmitPass() override; 605 }; 606 607 } // end anonymous namespace 608 609 void AMDGPUPassConfig::addEarlyCSEOrGVNPass() { 610 if (getOptLevel() == CodeGenOpt::Aggressive) 611 addPass(createGVNPass()); 612 else 613 addPass(createEarlyCSEPass()); 614 } 615 616 void AMDGPUPassConfig::addStraightLineScalarOptimizationPasses() { 617 addPass(createLICMPass()); 618 addPass(createSeparateConstOffsetFromGEPPass()); 619 addPass(createSpeculativeExecutionPass()); 620 // ReassociateGEPs exposes more opportunites for SLSR. See 621 // the example in reassociate-geps-and-slsr.ll. 622 addPass(createStraightLineStrengthReducePass()); 623 // SeparateConstOffsetFromGEP and SLSR creates common expressions which GVN or 624 // EarlyCSE can reuse. 625 addEarlyCSEOrGVNPass(); 626 // Run NaryReassociate after EarlyCSE/GVN to be more effective. 627 addPass(createNaryReassociatePass()); 628 // NaryReassociate on GEPs creates redundant common expressions, so run 629 // EarlyCSE after it. 630 addPass(createEarlyCSEPass()); 631 } 632 633 void AMDGPUPassConfig::addIRPasses() { 634 const AMDGPUTargetMachine &TM = getAMDGPUTargetMachine(); 635 636 // There is no reason to run these. 637 disablePass(&StackMapLivenessID); 638 disablePass(&FuncletLayoutID); 639 disablePass(&PatchableFunctionID); 640 641 addPass(createAtomicExpandPass()); 642 643 // This must occur before inlining, as the inliner will not look through 644 // bitcast calls. 645 addPass(createAMDGPUFixFunctionBitcastsPass()); 646 647 addPass(createAMDGPULowerIntrinsicsPass()); 648 649 // Function calls are not supported, so make sure we inline everything. 650 addPass(createAMDGPUAlwaysInlinePass()); 651 addPass(createAlwaysInlinerLegacyPass()); 652 // We need to add the barrier noop pass, otherwise adding the function 653 // inlining pass will cause all of the PassConfigs passes to be run 654 // one function at a time, which means if we have a nodule with two 655 // functions, then we will generate code for the first function 656 // without ever running any passes on the second. 657 addPass(createBarrierNoopPass()); 658 659 if (TM.getTargetTriple().getArch() == Triple::amdgcn) { 660 // TODO: May want to move later or split into an early and late one. 661 662 addPass(createAMDGPUCodeGenPreparePass()); 663 } 664 665 // Handle uses of OpenCL image2d_t, image3d_t and sampler_t arguments. 666 if (TM.getTargetTriple().getArch() == Triple::r600) 667 addPass(createR600OpenCLImageTypeLoweringPass()); 668 669 // Replace OpenCL enqueued block function pointers with global variables. 670 addPass(createAMDGPUOpenCLEnqueuedBlockLoweringPass()); 671 672 if (TM.getOptLevel() > CodeGenOpt::None) { 673 addPass(createInferAddressSpacesPass()); 674 addPass(createAMDGPUPromoteAlloca()); 675 676 if (EnableSROA) 677 addPass(createSROAPass()); 678 679 if (EnableScalarIRPasses) 680 addStraightLineScalarOptimizationPasses(); 681 682 if (EnableAMDGPUAliasAnalysis) { 683 addPass(createAMDGPUAAWrapperPass()); 684 addPass(createExternalAAWrapperPass([](Pass &P, Function &, 685 AAResults &AAR) { 686 if (auto *WrapperPass = P.getAnalysisIfAvailable<AMDGPUAAWrapperPass>()) 687 AAR.addAAResult(WrapperPass->getResult()); 688 })); 689 } 690 } 691 692 TargetPassConfig::addIRPasses(); 693 694 // EarlyCSE is not always strong enough to clean up what LSR produces. For 695 // example, GVN can combine 696 // 697 // %0 = add %a, %b 698 // %1 = add %b, %a 699 // 700 // and 701 // 702 // %0 = shl nsw %a, 2 703 // %1 = shl %a, 2 704 // 705 // but EarlyCSE can do neither of them. 706 if (getOptLevel() != CodeGenOpt::None && EnableScalarIRPasses) 707 addEarlyCSEOrGVNPass(); 708 } 709 710 void AMDGPUPassConfig::addCodeGenPrepare() { 711 if (TM->getTargetTriple().getArch() == Triple::amdgcn) 712 addPass(createAMDGPUAnnotateKernelFeaturesPass()); 713 714 if (TM->getTargetTriple().getArch() == Triple::amdgcn && 715 EnableLowerKernelArguments) 716 addPass(createAMDGPULowerKernelArgumentsPass()); 717 718 TargetPassConfig::addCodeGenPrepare(); 719 720 if (EnableLoadStoreVectorizer) 721 addPass(createLoadStoreVectorizerPass()); 722 } 723 724 bool AMDGPUPassConfig::addPreISel() { 725 addPass(createLowerSwitchPass()); 726 addPass(createFlattenCFGPass()); 727 return false; 728 } 729 730 bool AMDGPUPassConfig::addInstSelector() { 731 addPass(createAMDGPUISelDag(&getAMDGPUTargetMachine(), getOptLevel())); 732 return false; 733 } 734 735 bool AMDGPUPassConfig::addGCPasses() { 736 // Do nothing. GC is not supported. 737 return false; 738 } 739 740 //===----------------------------------------------------------------------===// 741 // R600 Pass Setup 742 //===----------------------------------------------------------------------===// 743 744 bool R600PassConfig::addPreISel() { 745 AMDGPUPassConfig::addPreISel(); 746 747 if (EnableR600StructurizeCFG) 748 addPass(createStructurizeCFGPass()); 749 return false; 750 } 751 752 bool R600PassConfig::addInstSelector() { 753 addPass(createR600ISelDag(&getAMDGPUTargetMachine(), getOptLevel())); 754 return false; 755 } 756 757 void R600PassConfig::addPreRegAlloc() { 758 addPass(createR600VectorRegMerger()); 759 } 760 761 void R600PassConfig::addPreSched2() { 762 addPass(createR600EmitClauseMarkers(), false); 763 if (EnableR600IfConvert) 764 addPass(&IfConverterID, false); 765 addPass(createR600ClauseMergePass(), false); 766 } 767 768 void R600PassConfig::addPreEmitPass() { 769 addPass(createAMDGPUCFGStructurizerPass(), false); 770 addPass(createR600ExpandSpecialInstrsPass(), false); 771 addPass(&FinalizeMachineBundlesID, false); 772 addPass(createR600Packetizer(), false); 773 addPass(createR600ControlFlowFinalizer(), false); 774 } 775 776 TargetPassConfig *R600TargetMachine::createPassConfig(PassManagerBase &PM) { 777 return new R600PassConfig(*this, PM); 778 } 779 780 //===----------------------------------------------------------------------===// 781 // GCN Pass Setup 782 //===----------------------------------------------------------------------===// 783 784 ScheduleDAGInstrs *GCNPassConfig::createMachineScheduler( 785 MachineSchedContext *C) const { 786 const GCNSubtarget &ST = C->MF->getSubtarget<GCNSubtarget>(); 787 if (ST.enableSIScheduler()) 788 return createSIMachineScheduler(C); 789 return createGCNMaxOccupancyMachineScheduler(C); 790 } 791 792 bool GCNPassConfig::addPreISel() { 793 AMDGPUPassConfig::addPreISel(); 794 795 if (EnableAtomicOptimizations) { 796 addPass(createAMDGPUAtomicOptimizerPass()); 797 } 798 799 // FIXME: We need to run a pass to propagate the attributes when calls are 800 // supported. 801 802 // Merge divergent exit nodes. StructurizeCFG won't recognize the multi-exit 803 // regions formed by them. 804 addPass(&AMDGPUUnifyDivergentExitNodesID); 805 if (!LateCFGStructurize) { 806 addPass(createStructurizeCFGPass(true)); // true -> SkipUniformRegions 807 } 808 addPass(createSinkingPass()); 809 addPass(createAMDGPUAnnotateUniformValues()); 810 if (!LateCFGStructurize) { 811 addPass(createSIAnnotateControlFlowPass()); 812 } 813 814 return false; 815 } 816 817 void GCNPassConfig::addMachineSSAOptimization() { 818 TargetPassConfig::addMachineSSAOptimization(); 819 820 // We want to fold operands after PeepholeOptimizer has run (or as part of 821 // it), because it will eliminate extra copies making it easier to fold the 822 // real source operand. We want to eliminate dead instructions after, so that 823 // we see fewer uses of the copies. We then need to clean up the dead 824 // instructions leftover after the operands are folded as well. 825 // 826 // XXX - Can we get away without running DeadMachineInstructionElim again? 827 addPass(&SIFoldOperandsID); 828 if (EnableDPPCombine) 829 addPass(&GCNDPPCombineID); 830 addPass(&DeadMachineInstructionElimID); 831 addPass(&SILoadStoreOptimizerID); 832 if (EnableSDWAPeephole) { 833 addPass(&SIPeepholeSDWAID); 834 addPass(&EarlyMachineLICMID); 835 addPass(&MachineCSEID); 836 addPass(&SIFoldOperandsID); 837 addPass(&DeadMachineInstructionElimID); 838 } 839 addPass(createSIShrinkInstructionsPass()); 840 } 841 842 bool GCNPassConfig::addILPOpts() { 843 if (EnableEarlyIfConversion) 844 addPass(&EarlyIfConverterID); 845 846 TargetPassConfig::addILPOpts(); 847 return false; 848 } 849 850 bool GCNPassConfig::addInstSelector() { 851 AMDGPUPassConfig::addInstSelector(); 852 addPass(&SIFixSGPRCopiesID); 853 addPass(createSILowerI1CopiesPass()); 854 addPass(createSIFixupVectorISelPass()); 855 addPass(createSIAddIMGInitPass()); 856 return false; 857 } 858 859 bool GCNPassConfig::addIRTranslator() { 860 addPass(new IRTranslator()); 861 return false; 862 } 863 864 bool GCNPassConfig::addLegalizeMachineIR() { 865 addPass(new Legalizer()); 866 return false; 867 } 868 869 bool GCNPassConfig::addRegBankSelect() { 870 addPass(new RegBankSelect()); 871 return false; 872 } 873 874 bool GCNPassConfig::addGlobalInstructionSelect() { 875 addPass(new InstructionSelect()); 876 return false; 877 } 878 879 void GCNPassConfig::addPreRegAlloc() { 880 if (LateCFGStructurize) { 881 addPass(createAMDGPUMachineCFGStructurizerPass()); 882 } 883 addPass(createSIWholeQuadModePass()); 884 } 885 886 void GCNPassConfig::addFastRegAlloc() { 887 // FIXME: We have to disable the verifier here because of PHIElimination + 888 // TwoAddressInstructions disabling it. 889 890 // This must be run immediately after phi elimination and before 891 // TwoAddressInstructions, otherwise the processing of the tied operand of 892 // SI_ELSE will introduce a copy of the tied operand source after the else. 893 insertPass(&PHIEliminationID, &SILowerControlFlowID, false); 894 895 // This must be run just after RegisterCoalescing. 896 insertPass(&RegisterCoalescerID, &SIPreAllocateWWMRegsID, false); 897 898 TargetPassConfig::addFastRegAlloc(); 899 } 900 901 void GCNPassConfig::addOptimizedRegAlloc() { 902 if (OptExecMaskPreRA) { 903 insertPass(&MachineSchedulerID, &SIOptimizeExecMaskingPreRAID); 904 insertPass(&SIOptimizeExecMaskingPreRAID, &SIFormMemoryClausesID); 905 } else { 906 insertPass(&MachineSchedulerID, &SIFormMemoryClausesID); 907 } 908 909 // This must be run immediately after phi elimination and before 910 // TwoAddressInstructions, otherwise the processing of the tied operand of 911 // SI_ELSE will introduce a copy of the tied operand source after the else. 912 insertPass(&PHIEliminationID, &SILowerControlFlowID, false); 913 914 // This must be run just after RegisterCoalescing. 915 insertPass(&RegisterCoalescerID, &SIPreAllocateWWMRegsID, false); 916 917 if (EnableDCEInRA) 918 insertPass(&RenameIndependentSubregsID, &DeadMachineInstructionElimID); 919 920 TargetPassConfig::addOptimizedRegAlloc(); 921 } 922 923 void GCNPassConfig::addPostRegAlloc() { 924 addPass(&SIFixVGPRCopiesID); 925 if (getOptLevel() > CodeGenOpt::None) 926 addPass(&SIOptimizeExecMaskingID); 927 TargetPassConfig::addPostRegAlloc(); 928 } 929 930 void GCNPassConfig::addPreSched2() { 931 } 932 933 void GCNPassConfig::addPreEmitPass() { 934 addPass(createSIMemoryLegalizerPass()); 935 addPass(createSIInsertWaitcntsPass()); 936 addPass(createSIShrinkInstructionsPass()); 937 addPass(createSIModeRegisterPass()); 938 939 // The hazard recognizer that runs as part of the post-ra scheduler does not 940 // guarantee to be able handle all hazards correctly. This is because if there 941 // are multiple scheduling regions in a basic block, the regions are scheduled 942 // bottom up, so when we begin to schedule a region we don't know what 943 // instructions were emitted directly before it. 944 // 945 // Here we add a stand-alone hazard recognizer pass which can handle all 946 // cases. 947 // 948 // FIXME: This stand-alone pass will emit indiv. S_NOP 0, as needed. It would 949 // be better for it to emit S_NOP <N> when possible. 950 addPass(&PostRAHazardRecognizerID); 951 952 addPass(&SIInsertSkipsPassID); 953 addPass(&BranchRelaxationPassID); 954 } 955 956 TargetPassConfig *GCNTargetMachine::createPassConfig(PassManagerBase &PM) { 957 return new GCNPassConfig(*this, PM); 958 } 959 960 yaml::MachineFunctionInfo *GCNTargetMachine::createDefaultFuncInfoYAML() const { 961 return new yaml::SIMachineFunctionInfo(); 962 } 963 964 yaml::MachineFunctionInfo * 965 GCNTargetMachine::convertFuncInfoToYAML(const MachineFunction &MF) const { 966 const SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>(); 967 return new yaml::SIMachineFunctionInfo(*MFI, 968 *MF.getSubtarget().getRegisterInfo()); 969 } 970 971 bool GCNTargetMachine::parseMachineFunctionInfo( 972 const yaml::MachineFunctionInfo &MFI_, PerFunctionMIParsingState &PFS, 973 SMDiagnostic &Error, SMRange &SourceRange) const { 974 const yaml::SIMachineFunctionInfo &YamlMFI = 975 reinterpret_cast<const yaml::SIMachineFunctionInfo &>(MFI_); 976 MachineFunction &MF = PFS.MF; 977 SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>(); 978 979 MFI->initializeBaseYamlFields(YamlMFI); 980 981 auto parseRegister = [&](const yaml::StringValue &RegName, unsigned &RegVal) { 982 if (parseNamedRegisterReference(PFS, RegVal, RegName.Value, Error)) { 983 SourceRange = RegName.SourceRange; 984 return true; 985 } 986 987 return false; 988 }; 989 990 auto diagnoseRegisterClass = [&](const yaml::StringValue &RegName) { 991 // Create a diagnostic for a the register string literal. 992 const MemoryBuffer &Buffer = 993 *PFS.SM->getMemoryBuffer(PFS.SM->getMainFileID()); 994 Error = SMDiagnostic(*PFS.SM, SMLoc(), Buffer.getBufferIdentifier(), 1, 995 RegName.Value.size(), SourceMgr::DK_Error, 996 "incorrect register class for field", RegName.Value, 997 None, None); 998 SourceRange = RegName.SourceRange; 999 return true; 1000 }; 1001 1002 if (parseRegister(YamlMFI.ScratchRSrcReg, MFI->ScratchRSrcReg) || 1003 parseRegister(YamlMFI.ScratchWaveOffsetReg, MFI->ScratchWaveOffsetReg) || 1004 parseRegister(YamlMFI.FrameOffsetReg, MFI->FrameOffsetReg) || 1005 parseRegister(YamlMFI.StackPtrOffsetReg, MFI->StackPtrOffsetReg)) 1006 return true; 1007 1008 if (MFI->ScratchRSrcReg != AMDGPU::PRIVATE_RSRC_REG && 1009 !AMDGPU::SReg_128RegClass.contains(MFI->ScratchRSrcReg)) { 1010 return diagnoseRegisterClass(YamlMFI.ScratchRSrcReg); 1011 } 1012 1013 if (MFI->ScratchWaveOffsetReg != AMDGPU::SCRATCH_WAVE_OFFSET_REG && 1014 !AMDGPU::SGPR_32RegClass.contains(MFI->ScratchWaveOffsetReg)) { 1015 return diagnoseRegisterClass(YamlMFI.ScratchWaveOffsetReg); 1016 } 1017 1018 if (MFI->FrameOffsetReg != AMDGPU::FP_REG && 1019 !AMDGPU::SGPR_32RegClass.contains(MFI->FrameOffsetReg)) { 1020 return diagnoseRegisterClass(YamlMFI.FrameOffsetReg); 1021 } 1022 1023 if (MFI->StackPtrOffsetReg != AMDGPU::SP_REG && 1024 !AMDGPU::SGPR_32RegClass.contains(MFI->StackPtrOffsetReg)) { 1025 return diagnoseRegisterClass(YamlMFI.StackPtrOffsetReg); 1026 } 1027 1028 return false; 1029 } 1030