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