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 substitutePass(&PostRASchedulerID, &PostMachineSchedulerID); 813 } 814 815 GCNTargetMachine &getGCNTargetMachine() const { 816 return getTM<GCNTargetMachine>(); 817 } 818 819 ScheduleDAGInstrs * 820 createMachineScheduler(MachineSchedContext *C) const override; 821 822 ScheduleDAGInstrs * 823 createPostMachineScheduler(MachineSchedContext *C) const override { 824 ScheduleDAGMI *DAG = createGenericSchedPostRA(C); 825 const GCNSubtarget &ST = C->MF->getSubtarget<GCNSubtarget>(); 826 DAG->addMutation(ST.createFillMFMAShadowMutation(DAG->TII)); 827 return DAG; 828 } 829 830 bool addPreISel() override; 831 void addMachineSSAOptimization() override; 832 bool addILPOpts() override; 833 bool addInstSelector() override; 834 bool addIRTranslator() override; 835 void addPreLegalizeMachineIR() override; 836 bool addLegalizeMachineIR() override; 837 void addPreRegBankSelect() override; 838 bool addRegBankSelect() override; 839 void addPreGlobalInstructionSelect() override; 840 bool addGlobalInstructionSelect() override; 841 void addFastRegAlloc() override; 842 void addOptimizedRegAlloc() override; 843 844 FunctionPass *createSGPRAllocPass(bool Optimized); 845 FunctionPass *createVGPRAllocPass(bool Optimized); 846 FunctionPass *createRegAllocPass(bool Optimized) override; 847 848 bool addRegAssignAndRewriteFast() override; 849 bool addRegAssignAndRewriteOptimized() override; 850 851 void addPreRegAlloc() override; 852 bool addPreRewrite() override; 853 void addPostRegAlloc() override; 854 void addPreSched2() override; 855 void addPreEmitPass() override; 856 }; 857 858 } // end anonymous namespace 859 860 AMDGPUPassConfig::AMDGPUPassConfig(LLVMTargetMachine &TM, PassManagerBase &PM) 861 : TargetPassConfig(TM, PM) { 862 // Exceptions and StackMaps are not supported, so these passes will never do 863 // anything. 864 disablePass(&StackMapLivenessID); 865 disablePass(&FuncletLayoutID); 866 // Garbage collection is not supported. 867 disablePass(&GCLoweringID); 868 disablePass(&ShadowStackGCLoweringID); 869 } 870 871 void AMDGPUPassConfig::addEarlyCSEOrGVNPass() { 872 if (getOptLevel() == CodeGenOpt::Aggressive) 873 addPass(createGVNPass()); 874 else 875 addPass(createEarlyCSEPass()); 876 } 877 878 void AMDGPUPassConfig::addStraightLineScalarOptimizationPasses() { 879 addPass(createLICMPass()); 880 addPass(createSeparateConstOffsetFromGEPPass()); 881 addPass(createSpeculativeExecutionPass()); 882 // ReassociateGEPs exposes more opportunities for SLSR. See 883 // the example in reassociate-geps-and-slsr.ll. 884 addPass(createStraightLineStrengthReducePass()); 885 // SeparateConstOffsetFromGEP and SLSR creates common expressions which GVN or 886 // EarlyCSE can reuse. 887 addEarlyCSEOrGVNPass(); 888 // Run NaryReassociate after EarlyCSE/GVN to be more effective. 889 addPass(createNaryReassociatePass()); 890 // NaryReassociate on GEPs creates redundant common expressions, so run 891 // EarlyCSE after it. 892 addPass(createEarlyCSEPass()); 893 } 894 895 void AMDGPUPassConfig::addIRPasses() { 896 const AMDGPUTargetMachine &TM = getAMDGPUTargetMachine(); 897 898 // There is no reason to run these. 899 disablePass(&StackMapLivenessID); 900 disablePass(&FuncletLayoutID); 901 disablePass(&PatchableFunctionID); 902 903 addPass(createAMDGPUPrintfRuntimeBinding()); 904 addPass(createAMDGPUCtorDtorLoweringPass()); 905 906 // This must occur before inlining, as the inliner will not look through 907 // bitcast calls. 908 addPass(createAMDGPUFixFunctionBitcastsPass()); 909 910 // A call to propagate attributes pass in the backend in case opt was not run. 911 addPass(createAMDGPUPropagateAttributesEarlyPass(&TM)); 912 913 addPass(createAMDGPULowerIntrinsicsPass()); 914 915 // Function calls are not supported, so make sure we inline everything. 916 addPass(createAMDGPUAlwaysInlinePass()); 917 addPass(createAlwaysInlinerLegacyPass()); 918 // We need to add the barrier noop pass, otherwise adding the function 919 // inlining pass will cause all of the PassConfigs passes to be run 920 // one function at a time, which means if we have a nodule with two 921 // functions, then we will generate code for the first function 922 // without ever running any passes on the second. 923 addPass(createBarrierNoopPass()); 924 925 // Handle uses of OpenCL image2d_t, image3d_t and sampler_t arguments. 926 if (TM.getTargetTriple().getArch() == Triple::r600) 927 addPass(createR600OpenCLImageTypeLoweringPass()); 928 929 // Replace OpenCL enqueued block function pointers with global variables. 930 addPass(createAMDGPUOpenCLEnqueuedBlockLoweringPass()); 931 932 // Can increase LDS used by kernel so runs before PromoteAlloca 933 if (EnableLowerModuleLDS) { 934 // The pass "amdgpu-replace-lds-use-with-pointer" need to be run before the 935 // pass "amdgpu-lower-module-lds", and also it required to be run only if 936 // "amdgpu-lower-module-lds" pass is enabled. 937 if (EnableLDSReplaceWithPointer) 938 addPass(createAMDGPUReplaceLDSUseWithPointerPass()); 939 940 addPass(createAMDGPULowerModuleLDSPass()); 941 } 942 943 if (TM.getOptLevel() > CodeGenOpt::None) 944 addPass(createInferAddressSpacesPass()); 945 946 addPass(createAtomicExpandPass()); 947 948 if (TM.getOptLevel() > CodeGenOpt::None) { 949 addPass(createAMDGPUPromoteAlloca()); 950 951 if (EnableSROA) 952 addPass(createSROAPass()); 953 if (isPassEnabled(EnableScalarIRPasses)) 954 addStraightLineScalarOptimizationPasses(); 955 956 if (EnableAMDGPUAliasAnalysis) { 957 addPass(createAMDGPUAAWrapperPass()); 958 addPass(createExternalAAWrapperPass([](Pass &P, Function &, 959 AAResults &AAR) { 960 if (auto *WrapperPass = P.getAnalysisIfAvailable<AMDGPUAAWrapperPass>()) 961 AAR.addAAResult(WrapperPass->getResult()); 962 })); 963 } 964 965 if (TM.getTargetTriple().getArch() == Triple::amdgcn) { 966 // TODO: May want to move later or split into an early and late one. 967 addPass(createAMDGPUCodeGenPreparePass()); 968 } 969 } 970 971 TargetPassConfig::addIRPasses(); 972 973 // EarlyCSE is not always strong enough to clean up what LSR produces. For 974 // example, GVN can combine 975 // 976 // %0 = add %a, %b 977 // %1 = add %b, %a 978 // 979 // and 980 // 981 // %0 = shl nsw %a, 2 982 // %1 = shl %a, 2 983 // 984 // but EarlyCSE can do neither of them. 985 if (isPassEnabled(EnableScalarIRPasses)) 986 addEarlyCSEOrGVNPass(); 987 } 988 989 void AMDGPUPassConfig::addCodeGenPrepare() { 990 if (TM->getTargetTriple().getArch() == Triple::amdgcn) { 991 addPass(createAMDGPUAttributorPass()); 992 993 // FIXME: This pass adds 2 hacky attributes that can be replaced with an 994 // analysis, and should be removed. 995 addPass(createAMDGPUAnnotateKernelFeaturesPass()); 996 } 997 998 if (TM->getTargetTriple().getArch() == Triple::amdgcn && 999 EnableLowerKernelArguments) 1000 addPass(createAMDGPULowerKernelArgumentsPass()); 1001 1002 TargetPassConfig::addCodeGenPrepare(); 1003 1004 if (isPassEnabled(EnableLoadStoreVectorizer)) 1005 addPass(createLoadStoreVectorizerPass()); 1006 1007 // LowerSwitch pass may introduce unreachable blocks that can 1008 // cause unexpected behavior for subsequent passes. Placing it 1009 // here seems better that these blocks would get cleaned up by 1010 // UnreachableBlockElim inserted next in the pass flow. 1011 addPass(createLowerSwitchPass()); 1012 } 1013 1014 bool AMDGPUPassConfig::addPreISel() { 1015 if (TM->getOptLevel() > CodeGenOpt::None) 1016 addPass(createFlattenCFGPass()); 1017 return false; 1018 } 1019 1020 bool AMDGPUPassConfig::addInstSelector() { 1021 // Defer the verifier until FinalizeISel. 1022 addPass(createAMDGPUISelDag(&getAMDGPUTargetMachine(), getOptLevel()), false); 1023 return false; 1024 } 1025 1026 bool AMDGPUPassConfig::addGCPasses() { 1027 // Do nothing. GC is not supported. 1028 return false; 1029 } 1030 1031 llvm::ScheduleDAGInstrs * 1032 AMDGPUPassConfig::createMachineScheduler(MachineSchedContext *C) const { 1033 ScheduleDAGMILive *DAG = createGenericSchedLive(C); 1034 DAG->addMutation(createLoadClusterDAGMutation(DAG->TII, DAG->TRI)); 1035 return DAG; 1036 } 1037 1038 //===----------------------------------------------------------------------===// 1039 // GCN Pass Setup 1040 //===----------------------------------------------------------------------===// 1041 1042 ScheduleDAGInstrs *GCNPassConfig::createMachineScheduler( 1043 MachineSchedContext *C) const { 1044 const GCNSubtarget &ST = C->MF->getSubtarget<GCNSubtarget>(); 1045 if (ST.enableSIScheduler()) 1046 return createSIMachineScheduler(C); 1047 return createGCNMaxOccupancyMachineScheduler(C); 1048 } 1049 1050 bool GCNPassConfig::addPreISel() { 1051 AMDGPUPassConfig::addPreISel(); 1052 1053 if (TM->getOptLevel() > CodeGenOpt::None) 1054 addPass(createAMDGPULateCodeGenPreparePass()); 1055 1056 if (isPassEnabled(EnableAtomicOptimizations, CodeGenOpt::Less)) { 1057 addPass(createAMDGPUAtomicOptimizerPass()); 1058 } 1059 1060 if (TM->getOptLevel() > CodeGenOpt::None) 1061 addPass(createSinkingPass()); 1062 1063 // Merge divergent exit nodes. StructurizeCFG won't recognize the multi-exit 1064 // regions formed by them. 1065 addPass(&AMDGPUUnifyDivergentExitNodesID); 1066 if (!LateCFGStructurize) { 1067 if (EnableStructurizerWorkarounds) { 1068 addPass(createFixIrreduciblePass()); 1069 addPass(createUnifyLoopExitsPass()); 1070 } 1071 addPass(createStructurizeCFGPass(false)); // true -> SkipUniformRegions 1072 } 1073 addPass(createAMDGPUAnnotateUniformValues()); 1074 if (!LateCFGStructurize) { 1075 addPass(createSIAnnotateControlFlowPass()); 1076 } 1077 addPass(createLCSSAPass()); 1078 1079 if (TM->getOptLevel() > CodeGenOpt::Less) 1080 addPass(&AMDGPUPerfHintAnalysisID); 1081 1082 return false; 1083 } 1084 1085 void GCNPassConfig::addMachineSSAOptimization() { 1086 TargetPassConfig::addMachineSSAOptimization(); 1087 1088 // We want to fold operands after PeepholeOptimizer has run (or as part of 1089 // it), because it will eliminate extra copies making it easier to fold the 1090 // real source operand. We want to eliminate dead instructions after, so that 1091 // we see fewer uses of the copies. We then need to clean up the dead 1092 // instructions leftover after the operands are folded as well. 1093 // 1094 // XXX - Can we get away without running DeadMachineInstructionElim again? 1095 addPass(&SIFoldOperandsID); 1096 if (EnableDPPCombine) 1097 addPass(&GCNDPPCombineID); 1098 addPass(&SILoadStoreOptimizerID); 1099 if (isPassEnabled(EnableSDWAPeephole)) { 1100 addPass(&SIPeepholeSDWAID); 1101 addPass(&EarlyMachineLICMID); 1102 addPass(&MachineCSEID); 1103 addPass(&SIFoldOperandsID); 1104 } 1105 addPass(&DeadMachineInstructionElimID); 1106 addPass(createSIShrinkInstructionsPass()); 1107 } 1108 1109 bool GCNPassConfig::addILPOpts() { 1110 if (EnableEarlyIfConversion) 1111 addPass(&EarlyIfConverterID); 1112 1113 TargetPassConfig::addILPOpts(); 1114 return false; 1115 } 1116 1117 bool GCNPassConfig::addInstSelector() { 1118 AMDGPUPassConfig::addInstSelector(); 1119 addPass(&SIFixSGPRCopiesID); 1120 addPass(createSILowerI1CopiesPass()); 1121 return false; 1122 } 1123 1124 bool GCNPassConfig::addIRTranslator() { 1125 addPass(new IRTranslator(getOptLevel())); 1126 return false; 1127 } 1128 1129 void GCNPassConfig::addPreLegalizeMachineIR() { 1130 bool IsOptNone = getOptLevel() == CodeGenOpt::None; 1131 addPass(createAMDGPUPreLegalizeCombiner(IsOptNone)); 1132 addPass(new Localizer()); 1133 } 1134 1135 bool GCNPassConfig::addLegalizeMachineIR() { 1136 addPass(new Legalizer()); 1137 return false; 1138 } 1139 1140 void GCNPassConfig::addPreRegBankSelect() { 1141 bool IsOptNone = getOptLevel() == CodeGenOpt::None; 1142 addPass(createAMDGPUPostLegalizeCombiner(IsOptNone)); 1143 } 1144 1145 bool GCNPassConfig::addRegBankSelect() { 1146 addPass(new RegBankSelect()); 1147 return false; 1148 } 1149 1150 void GCNPassConfig::addPreGlobalInstructionSelect() { 1151 bool IsOptNone = getOptLevel() == CodeGenOpt::None; 1152 addPass(createAMDGPURegBankCombiner(IsOptNone)); 1153 } 1154 1155 bool GCNPassConfig::addGlobalInstructionSelect() { 1156 addPass(new InstructionSelect(getOptLevel())); 1157 return false; 1158 } 1159 1160 void GCNPassConfig::addPreRegAlloc() { 1161 if (LateCFGStructurize) { 1162 addPass(createAMDGPUMachineCFGStructurizerPass()); 1163 } 1164 } 1165 1166 void GCNPassConfig::addFastRegAlloc() { 1167 // FIXME: We have to disable the verifier here because of PHIElimination + 1168 // TwoAddressInstructions disabling it. 1169 1170 // This must be run immediately after phi elimination and before 1171 // TwoAddressInstructions, otherwise the processing of the tied operand of 1172 // SI_ELSE will introduce a copy of the tied operand source after the else. 1173 insertPass(&PHIEliminationID, &SILowerControlFlowID, false); 1174 1175 insertPass(&TwoAddressInstructionPassID, &SIWholeQuadModeID); 1176 insertPass(&TwoAddressInstructionPassID, &SIPreAllocateWWMRegsID); 1177 1178 TargetPassConfig::addFastRegAlloc(); 1179 } 1180 1181 void GCNPassConfig::addOptimizedRegAlloc() { 1182 // Allow the scheduler to run before SIWholeQuadMode inserts exec manipulation 1183 // instructions that cause scheduling barriers. 1184 insertPass(&MachineSchedulerID, &SIWholeQuadModeID); 1185 insertPass(&MachineSchedulerID, &SIPreAllocateWWMRegsID); 1186 1187 if (OptExecMaskPreRA) 1188 insertPass(&MachineSchedulerID, &SIOptimizeExecMaskingPreRAID); 1189 1190 if (isPassEnabled(EnablePreRAOptimizations)) 1191 insertPass(&RenameIndependentSubregsID, &GCNPreRAOptimizationsID); 1192 1193 // This is not an essential optimization and it has a noticeable impact on 1194 // compilation time, so we only enable it from O2. 1195 if (TM->getOptLevel() > CodeGenOpt::Less) 1196 insertPass(&MachineSchedulerID, &SIFormMemoryClausesID); 1197 1198 // FIXME: when an instruction has a Killed operand, and the instruction is 1199 // inside a bundle, seems only the BUNDLE instruction appears as the Kills of 1200 // the register in LiveVariables, this would trigger a failure in verifier, 1201 // we should fix it and enable the verifier. 1202 if (OptVGPRLiveRange) 1203 insertPass(&LiveVariablesID, &SIOptimizeVGPRLiveRangeID, false); 1204 // This must be run immediately after phi elimination and before 1205 // TwoAddressInstructions, otherwise the processing of the tied operand of 1206 // SI_ELSE will introduce a copy of the tied operand source after the else. 1207 insertPass(&PHIEliminationID, &SILowerControlFlowID, false); 1208 1209 if (EnableDCEInRA) 1210 insertPass(&DetectDeadLanesID, &DeadMachineInstructionElimID); 1211 1212 TargetPassConfig::addOptimizedRegAlloc(); 1213 } 1214 1215 bool GCNPassConfig::addPreRewrite() { 1216 if (EnableRegReassign) 1217 addPass(&GCNNSAReassignID); 1218 return true; 1219 } 1220 1221 FunctionPass *GCNPassConfig::createSGPRAllocPass(bool Optimized) { 1222 // Initialize the global default. 1223 llvm::call_once(InitializeDefaultSGPRRegisterAllocatorFlag, 1224 initializeDefaultSGPRRegisterAllocatorOnce); 1225 1226 RegisterRegAlloc::FunctionPassCtor Ctor = SGPRRegisterRegAlloc::getDefault(); 1227 if (Ctor != useDefaultRegisterAllocator) 1228 return Ctor(); 1229 1230 if (Optimized) 1231 return createGreedyRegisterAllocator(onlyAllocateSGPRs); 1232 1233 return createFastRegisterAllocator(onlyAllocateSGPRs, false); 1234 } 1235 1236 FunctionPass *GCNPassConfig::createVGPRAllocPass(bool Optimized) { 1237 // Initialize the global default. 1238 llvm::call_once(InitializeDefaultVGPRRegisterAllocatorFlag, 1239 initializeDefaultVGPRRegisterAllocatorOnce); 1240 1241 RegisterRegAlloc::FunctionPassCtor Ctor = VGPRRegisterRegAlloc::getDefault(); 1242 if (Ctor != useDefaultRegisterAllocator) 1243 return Ctor(); 1244 1245 if (Optimized) 1246 return createGreedyVGPRRegisterAllocator(); 1247 1248 return createFastVGPRRegisterAllocator(); 1249 } 1250 1251 FunctionPass *GCNPassConfig::createRegAllocPass(bool Optimized) { 1252 llvm_unreachable("should not be used"); 1253 } 1254 1255 static const char RegAllocOptNotSupportedMessage[] = 1256 "-regalloc not supported with amdgcn. Use -sgpr-regalloc and -vgpr-regalloc"; 1257 1258 bool GCNPassConfig::addRegAssignAndRewriteFast() { 1259 if (!usingDefaultRegAlloc()) 1260 report_fatal_error(RegAllocOptNotSupportedMessage); 1261 1262 addPass(createSGPRAllocPass(false)); 1263 1264 // Equivalent of PEI for SGPRs. 1265 addPass(&SILowerSGPRSpillsID); 1266 1267 addPass(createVGPRAllocPass(false)); 1268 return true; 1269 } 1270 1271 bool GCNPassConfig::addRegAssignAndRewriteOptimized() { 1272 if (!usingDefaultRegAlloc()) 1273 report_fatal_error(RegAllocOptNotSupportedMessage); 1274 1275 addPass(createSGPRAllocPass(true)); 1276 1277 // Commit allocated register changes. This is mostly necessary because too 1278 // many things rely on the use lists of the physical registers, such as the 1279 // verifier. This is only necessary with allocators which use LiveIntervals, 1280 // since FastRegAlloc does the replacements itself. 1281 addPass(createVirtRegRewriter(false)); 1282 1283 // Equivalent of PEI for SGPRs. 1284 addPass(&SILowerSGPRSpillsID); 1285 1286 addPass(createVGPRAllocPass(true)); 1287 1288 addPreRewrite(); 1289 addPass(&VirtRegRewriterID); 1290 1291 return true; 1292 } 1293 1294 void GCNPassConfig::addPostRegAlloc() { 1295 addPass(&SIFixVGPRCopiesID); 1296 if (getOptLevel() > CodeGenOpt::None) 1297 addPass(&SIOptimizeExecMaskingID); 1298 TargetPassConfig::addPostRegAlloc(); 1299 } 1300 1301 void GCNPassConfig::addPreSched2() { 1302 addPass(&SIPostRABundlerID); 1303 } 1304 1305 void GCNPassConfig::addPreEmitPass() { 1306 addPass(createSIMemoryLegalizerPass()); 1307 addPass(createSIInsertWaitcntsPass()); 1308 1309 if (TM->getOptLevel() > CodeGenOpt::None) 1310 addPass(createSIShrinkInstructionsPass()); 1311 1312 addPass(createSIModeRegisterPass()); 1313 1314 if (getOptLevel() > CodeGenOpt::None) 1315 addPass(&SIInsertHardClausesID); 1316 1317 addPass(&SILateBranchLoweringPassID); 1318 if (getOptLevel() > CodeGenOpt::None) 1319 addPass(&SIPreEmitPeepholeID); 1320 // The hazard recognizer that runs as part of the post-ra scheduler does not 1321 // guarantee to be able handle all hazards correctly. This is because if there 1322 // are multiple scheduling regions in a basic block, the regions are scheduled 1323 // bottom up, so when we begin to schedule a region we don't know what 1324 // instructions were emitted directly before it. 1325 // 1326 // Here we add a stand-alone hazard recognizer pass which can handle all 1327 // cases. 1328 addPass(&PostRAHazardRecognizerID); 1329 addPass(&BranchRelaxationPassID); 1330 } 1331 1332 TargetPassConfig *GCNTargetMachine::createPassConfig(PassManagerBase &PM) { 1333 return new GCNPassConfig(*this, PM); 1334 } 1335 1336 yaml::MachineFunctionInfo *GCNTargetMachine::createDefaultFuncInfoYAML() const { 1337 return new yaml::SIMachineFunctionInfo(); 1338 } 1339 1340 yaml::MachineFunctionInfo * 1341 GCNTargetMachine::convertFuncInfoToYAML(const MachineFunction &MF) const { 1342 const SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>(); 1343 return new yaml::SIMachineFunctionInfo( 1344 *MFI, *MF.getSubtarget().getRegisterInfo(), MF); 1345 } 1346 1347 bool GCNTargetMachine::parseMachineFunctionInfo( 1348 const yaml::MachineFunctionInfo &MFI_, PerFunctionMIParsingState &PFS, 1349 SMDiagnostic &Error, SMRange &SourceRange) const { 1350 const yaml::SIMachineFunctionInfo &YamlMFI = 1351 reinterpret_cast<const yaml::SIMachineFunctionInfo &>(MFI_); 1352 MachineFunction &MF = PFS.MF; 1353 SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>(); 1354 1355 if (MFI->initializeBaseYamlFields(YamlMFI, MF, PFS, Error, SourceRange)) 1356 return true; 1357 1358 if (MFI->Occupancy == 0) { 1359 // Fixup the subtarget dependent default value. 1360 const GCNSubtarget &ST = MF.getSubtarget<GCNSubtarget>(); 1361 MFI->Occupancy = ST.computeOccupancy(MF.getFunction(), MFI->getLDSSize()); 1362 } 1363 1364 auto parseRegister = [&](const yaml::StringValue &RegName, Register &RegVal) { 1365 Register TempReg; 1366 if (parseNamedRegisterReference(PFS, TempReg, RegName.Value, Error)) { 1367 SourceRange = RegName.SourceRange; 1368 return true; 1369 } 1370 RegVal = TempReg; 1371 1372 return false; 1373 }; 1374 1375 auto diagnoseRegisterClass = [&](const yaml::StringValue &RegName) { 1376 // Create a diagnostic for a the register string literal. 1377 const MemoryBuffer &Buffer = 1378 *PFS.SM->getMemoryBuffer(PFS.SM->getMainFileID()); 1379 Error = SMDiagnostic(*PFS.SM, SMLoc(), Buffer.getBufferIdentifier(), 1, 1380 RegName.Value.size(), SourceMgr::DK_Error, 1381 "incorrect register class for field", RegName.Value, 1382 None, None); 1383 SourceRange = RegName.SourceRange; 1384 return true; 1385 }; 1386 1387 if (parseRegister(YamlMFI.ScratchRSrcReg, MFI->ScratchRSrcReg) || 1388 parseRegister(YamlMFI.FrameOffsetReg, MFI->FrameOffsetReg) || 1389 parseRegister(YamlMFI.StackPtrOffsetReg, MFI->StackPtrOffsetReg)) 1390 return true; 1391 1392 if (MFI->ScratchRSrcReg != AMDGPU::PRIVATE_RSRC_REG && 1393 !AMDGPU::SGPR_128RegClass.contains(MFI->ScratchRSrcReg)) { 1394 return diagnoseRegisterClass(YamlMFI.ScratchRSrcReg); 1395 } 1396 1397 if (MFI->FrameOffsetReg != AMDGPU::FP_REG && 1398 !AMDGPU::SGPR_32RegClass.contains(MFI->FrameOffsetReg)) { 1399 return diagnoseRegisterClass(YamlMFI.FrameOffsetReg); 1400 } 1401 1402 if (MFI->StackPtrOffsetReg != AMDGPU::SP_REG && 1403 !AMDGPU::SGPR_32RegClass.contains(MFI->StackPtrOffsetReg)) { 1404 return diagnoseRegisterClass(YamlMFI.StackPtrOffsetReg); 1405 } 1406 1407 auto parseAndCheckArgument = [&](const Optional<yaml::SIArgument> &A, 1408 const TargetRegisterClass &RC, 1409 ArgDescriptor &Arg, unsigned UserSGPRs, 1410 unsigned SystemSGPRs) { 1411 // Skip parsing if it's not present. 1412 if (!A) 1413 return false; 1414 1415 if (A->IsRegister) { 1416 Register Reg; 1417 if (parseNamedRegisterReference(PFS, Reg, A->RegisterName.Value, Error)) { 1418 SourceRange = A->RegisterName.SourceRange; 1419 return true; 1420 } 1421 if (!RC.contains(Reg)) 1422 return diagnoseRegisterClass(A->RegisterName); 1423 Arg = ArgDescriptor::createRegister(Reg); 1424 } else 1425 Arg = ArgDescriptor::createStack(A->StackOffset); 1426 // Check and apply the optional mask. 1427 if (A->Mask) 1428 Arg = ArgDescriptor::createArg(Arg, A->Mask.getValue()); 1429 1430 MFI->NumUserSGPRs += UserSGPRs; 1431 MFI->NumSystemSGPRs += SystemSGPRs; 1432 return false; 1433 }; 1434 1435 if (YamlMFI.ArgInfo && 1436 (parseAndCheckArgument(YamlMFI.ArgInfo->PrivateSegmentBuffer, 1437 AMDGPU::SGPR_128RegClass, 1438 MFI->ArgInfo.PrivateSegmentBuffer, 4, 0) || 1439 parseAndCheckArgument(YamlMFI.ArgInfo->DispatchPtr, 1440 AMDGPU::SReg_64RegClass, MFI->ArgInfo.DispatchPtr, 1441 2, 0) || 1442 parseAndCheckArgument(YamlMFI.ArgInfo->QueuePtr, AMDGPU::SReg_64RegClass, 1443 MFI->ArgInfo.QueuePtr, 2, 0) || 1444 parseAndCheckArgument(YamlMFI.ArgInfo->KernargSegmentPtr, 1445 AMDGPU::SReg_64RegClass, 1446 MFI->ArgInfo.KernargSegmentPtr, 2, 0) || 1447 parseAndCheckArgument(YamlMFI.ArgInfo->DispatchID, 1448 AMDGPU::SReg_64RegClass, MFI->ArgInfo.DispatchID, 1449 2, 0) || 1450 parseAndCheckArgument(YamlMFI.ArgInfo->FlatScratchInit, 1451 AMDGPU::SReg_64RegClass, 1452 MFI->ArgInfo.FlatScratchInit, 2, 0) || 1453 parseAndCheckArgument(YamlMFI.ArgInfo->PrivateSegmentSize, 1454 AMDGPU::SGPR_32RegClass, 1455 MFI->ArgInfo.PrivateSegmentSize, 0, 0) || 1456 parseAndCheckArgument(YamlMFI.ArgInfo->WorkGroupIDX, 1457 AMDGPU::SGPR_32RegClass, MFI->ArgInfo.WorkGroupIDX, 1458 0, 1) || 1459 parseAndCheckArgument(YamlMFI.ArgInfo->WorkGroupIDY, 1460 AMDGPU::SGPR_32RegClass, MFI->ArgInfo.WorkGroupIDY, 1461 0, 1) || 1462 parseAndCheckArgument(YamlMFI.ArgInfo->WorkGroupIDZ, 1463 AMDGPU::SGPR_32RegClass, MFI->ArgInfo.WorkGroupIDZ, 1464 0, 1) || 1465 parseAndCheckArgument(YamlMFI.ArgInfo->WorkGroupInfo, 1466 AMDGPU::SGPR_32RegClass, 1467 MFI->ArgInfo.WorkGroupInfo, 0, 1) || 1468 parseAndCheckArgument(YamlMFI.ArgInfo->PrivateSegmentWaveByteOffset, 1469 AMDGPU::SGPR_32RegClass, 1470 MFI->ArgInfo.PrivateSegmentWaveByteOffset, 0, 1) || 1471 parseAndCheckArgument(YamlMFI.ArgInfo->ImplicitArgPtr, 1472 AMDGPU::SReg_64RegClass, 1473 MFI->ArgInfo.ImplicitArgPtr, 0, 0) || 1474 parseAndCheckArgument(YamlMFI.ArgInfo->ImplicitBufferPtr, 1475 AMDGPU::SReg_64RegClass, 1476 MFI->ArgInfo.ImplicitBufferPtr, 2, 0) || 1477 parseAndCheckArgument(YamlMFI.ArgInfo->WorkItemIDX, 1478 AMDGPU::VGPR_32RegClass, 1479 MFI->ArgInfo.WorkItemIDX, 0, 0) || 1480 parseAndCheckArgument(YamlMFI.ArgInfo->WorkItemIDY, 1481 AMDGPU::VGPR_32RegClass, 1482 MFI->ArgInfo.WorkItemIDY, 0, 0) || 1483 parseAndCheckArgument(YamlMFI.ArgInfo->WorkItemIDZ, 1484 AMDGPU::VGPR_32RegClass, 1485 MFI->ArgInfo.WorkItemIDZ, 0, 0))) 1486 return true; 1487 1488 MFI->Mode.IEEE = YamlMFI.Mode.IEEE; 1489 MFI->Mode.DX10Clamp = YamlMFI.Mode.DX10Clamp; 1490 MFI->Mode.FP32InputDenormals = YamlMFI.Mode.FP32InputDenormals; 1491 MFI->Mode.FP32OutputDenormals = YamlMFI.Mode.FP32OutputDenormals; 1492 MFI->Mode.FP64FP16InputDenormals = YamlMFI.Mode.FP64FP16InputDenormals; 1493 MFI->Mode.FP64FP16OutputDenormals = YamlMFI.Mode.FP64FP16OutputDenormals; 1494 1495 return false; 1496 } 1497