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