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/MC/TargetRegistry.h" 43 #include "llvm/Passes/PassBuilder.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 static cl::opt<bool> EnablePromoteKernelArguments( 310 "amdgpu-enable-promote-kernel-arguments", 311 cl::desc("Enable promotion of flat kernel pointer arguments to global"), 312 cl::Hidden, cl::init(true)); 313 314 extern "C" LLVM_EXTERNAL_VISIBILITY void LLVMInitializeAMDGPUTarget() { 315 // Register the target 316 RegisterTargetMachine<R600TargetMachine> X(getTheAMDGPUTarget()); 317 RegisterTargetMachine<GCNTargetMachine> Y(getTheGCNTarget()); 318 319 PassRegistry *PR = PassRegistry::getPassRegistry(); 320 initializeR600ClauseMergePassPass(*PR); 321 initializeR600ControlFlowFinalizerPass(*PR); 322 initializeR600PacketizerPass(*PR); 323 initializeR600ExpandSpecialInstrsPassPass(*PR); 324 initializeR600VectorRegMergerPass(*PR); 325 initializeGlobalISel(*PR); 326 initializeAMDGPUDAGToDAGISelPass(*PR); 327 initializeGCNDPPCombinePass(*PR); 328 initializeSILowerI1CopiesPass(*PR); 329 initializeSILowerSGPRSpillsPass(*PR); 330 initializeSIFixSGPRCopiesPass(*PR); 331 initializeSIFixVGPRCopiesPass(*PR); 332 initializeSIFoldOperandsPass(*PR); 333 initializeSIPeepholeSDWAPass(*PR); 334 initializeSIShrinkInstructionsPass(*PR); 335 initializeSIOptimizeExecMaskingPreRAPass(*PR); 336 initializeSIOptimizeVGPRLiveRangePass(*PR); 337 initializeSILoadStoreOptimizerPass(*PR); 338 initializeAMDGPUFixFunctionBitcastsPass(*PR); 339 initializeAMDGPUCtorDtorLoweringPass(*PR); 340 initializeAMDGPUAlwaysInlinePass(*PR); 341 initializeAMDGPUAttributorPass(*PR); 342 initializeAMDGPUAnnotateKernelFeaturesPass(*PR); 343 initializeAMDGPUAnnotateUniformValuesPass(*PR); 344 initializeAMDGPUArgumentUsageInfoPass(*PR); 345 initializeAMDGPUAtomicOptimizerPass(*PR); 346 initializeAMDGPULowerKernelArgumentsPass(*PR); 347 initializeAMDGPUPromoteKernelArgumentsPass(*PR); 348 initializeAMDGPULowerKernelAttributesPass(*PR); 349 initializeAMDGPULowerIntrinsicsPass(*PR); 350 initializeAMDGPUOpenCLEnqueuedBlockLoweringPass(*PR); 351 initializeAMDGPUPostLegalizerCombinerPass(*PR); 352 initializeAMDGPUPreLegalizerCombinerPass(*PR); 353 initializeAMDGPURegBankCombinerPass(*PR); 354 initializeAMDGPUPromoteAllocaPass(*PR); 355 initializeAMDGPUPromoteAllocaToVectorPass(*PR); 356 initializeAMDGPUCodeGenPreparePass(*PR); 357 initializeAMDGPULateCodeGenPreparePass(*PR); 358 initializeAMDGPUPropagateAttributesEarlyPass(*PR); 359 initializeAMDGPUPropagateAttributesLatePass(*PR); 360 initializeAMDGPUReplaceLDSUseWithPointerPass(*PR); 361 initializeAMDGPULowerModuleLDSPass(*PR); 362 initializeAMDGPURewriteOutArgumentsPass(*PR); 363 initializeAMDGPUUnifyMetadataPass(*PR); 364 initializeSIAnnotateControlFlowPass(*PR); 365 initializeSIInsertHardClausesPass(*PR); 366 initializeSIInsertWaitcntsPass(*PR); 367 initializeSIModeRegisterPass(*PR); 368 initializeSIWholeQuadModePass(*PR); 369 initializeSILowerControlFlowPass(*PR); 370 initializeSIPreEmitPeepholePass(*PR); 371 initializeSILateBranchLoweringPass(*PR); 372 initializeSIMemoryLegalizerPass(*PR); 373 initializeSIOptimizeExecMaskingPass(*PR); 374 initializeSIPreAllocateWWMRegsPass(*PR); 375 initializeSIFormMemoryClausesPass(*PR); 376 initializeSIPostRABundlerPass(*PR); 377 initializeAMDGPUUnifyDivergentExitNodesPass(*PR); 378 initializeAMDGPUAAWrapperPassPass(*PR); 379 initializeAMDGPUExternalAAWrapperPass(*PR); 380 initializeAMDGPUUseNativeCallsPass(*PR); 381 initializeAMDGPUSimplifyLibCallsPass(*PR); 382 initializeAMDGPUPrintfRuntimeBindingPass(*PR); 383 initializeAMDGPUResourceUsageAnalysisPass(*PR); 384 initializeGCNNSAReassignPass(*PR); 385 initializeGCNPreRAOptimizationsPass(*PR); 386 } 387 388 static std::unique_ptr<TargetLoweringObjectFile> createTLOF(const Triple &TT) { 389 return std::make_unique<AMDGPUTargetObjectFile>(); 390 } 391 392 static ScheduleDAGInstrs *createSIMachineScheduler(MachineSchedContext *C) { 393 return new SIScheduleDAGMI(C); 394 } 395 396 static ScheduleDAGInstrs * 397 createGCNMaxOccupancyMachineScheduler(MachineSchedContext *C) { 398 ScheduleDAGMILive *DAG = 399 new GCNScheduleDAGMILive(C, std::make_unique<GCNMaxOccupancySchedStrategy>(C)); 400 DAG->addMutation(createLoadClusterDAGMutation(DAG->TII, DAG->TRI)); 401 DAG->addMutation(createAMDGPUMacroFusionDAGMutation()); 402 DAG->addMutation(createAMDGPUExportClusteringDAGMutation()); 403 return DAG; 404 } 405 406 static ScheduleDAGInstrs * 407 createIterativeGCNMaxOccupancyMachineScheduler(MachineSchedContext *C) { 408 auto DAG = new GCNIterativeScheduler(C, 409 GCNIterativeScheduler::SCHEDULE_LEGACYMAXOCCUPANCY); 410 DAG->addMutation(createLoadClusterDAGMutation(DAG->TII, DAG->TRI)); 411 return DAG; 412 } 413 414 static ScheduleDAGInstrs *createMinRegScheduler(MachineSchedContext *C) { 415 return new GCNIterativeScheduler(C, 416 GCNIterativeScheduler::SCHEDULE_MINREGFORCED); 417 } 418 419 static ScheduleDAGInstrs * 420 createIterativeILPMachineScheduler(MachineSchedContext *C) { 421 auto DAG = new GCNIterativeScheduler(C, 422 GCNIterativeScheduler::SCHEDULE_ILP); 423 DAG->addMutation(createLoadClusterDAGMutation(DAG->TII, DAG->TRI)); 424 DAG->addMutation(createAMDGPUMacroFusionDAGMutation()); 425 return DAG; 426 } 427 428 static MachineSchedRegistry 429 SISchedRegistry("si", "Run SI's custom scheduler", 430 createSIMachineScheduler); 431 432 static MachineSchedRegistry 433 GCNMaxOccupancySchedRegistry("gcn-max-occupancy", 434 "Run GCN scheduler to maximize occupancy", 435 createGCNMaxOccupancyMachineScheduler); 436 437 static MachineSchedRegistry 438 IterativeGCNMaxOccupancySchedRegistry("gcn-max-occupancy-experimental", 439 "Run GCN scheduler to maximize occupancy (experimental)", 440 createIterativeGCNMaxOccupancyMachineScheduler); 441 442 static MachineSchedRegistry 443 GCNMinRegSchedRegistry("gcn-minreg", 444 "Run GCN iterative scheduler for minimal register usage (experimental)", 445 createMinRegScheduler); 446 447 static MachineSchedRegistry 448 GCNILPSchedRegistry("gcn-ilp", 449 "Run GCN iterative scheduler for ILP scheduling (experimental)", 450 createIterativeILPMachineScheduler); 451 452 static StringRef computeDataLayout(const Triple &TT) { 453 if (TT.getArch() == Triple::r600) { 454 // 32-bit pointers. 455 return "e-p:32:32-i64:64-v16:16-v24:32-v32:32-v48:64-v96:128" 456 "-v192:256-v256:256-v512:512-v1024:1024-v2048:2048-n32:64-S32-A5-G1"; 457 } 458 459 // 32-bit private, local, and region pointers. 64-bit global, constant and 460 // flat, non-integral buffer fat pointers. 461 return "e-p:64:64-p1:64:64-p2:32:32-p3:32:32-p4:64:64-p5:32:32-p6:32:32" 462 "-i64:64-v16:16-v24:32-v32:32-v48:64-v96:128" 463 "-v192:256-v256:256-v512:512-v1024:1024-v2048:2048-n32:64-S32-A5-G1" 464 "-ni:7"; 465 } 466 467 LLVM_READNONE 468 static StringRef getGPUOrDefault(const Triple &TT, StringRef GPU) { 469 if (!GPU.empty()) 470 return GPU; 471 472 // Need to default to a target with flat support for HSA. 473 if (TT.getArch() == Triple::amdgcn) 474 return TT.getOS() == Triple::AMDHSA ? "generic-hsa" : "generic"; 475 476 return "r600"; 477 } 478 479 static Reloc::Model getEffectiveRelocModel(Optional<Reloc::Model> RM) { 480 // The AMDGPU toolchain only supports generating shared objects, so we 481 // must always use PIC. 482 return Reloc::PIC_; 483 } 484 485 AMDGPUTargetMachine::AMDGPUTargetMachine(const Target &T, const Triple &TT, 486 StringRef CPU, StringRef FS, 487 TargetOptions Options, 488 Optional<Reloc::Model> RM, 489 Optional<CodeModel::Model> CM, 490 CodeGenOpt::Level OptLevel) 491 : LLVMTargetMachine(T, computeDataLayout(TT), TT, getGPUOrDefault(TT, CPU), 492 FS, Options, getEffectiveRelocModel(RM), 493 getEffectiveCodeModel(CM, CodeModel::Small), OptLevel), 494 TLOF(createTLOF(getTargetTriple())) { 495 initAsmInfo(); 496 if (TT.getArch() == Triple::amdgcn) { 497 if (getMCSubtargetInfo()->checkFeatures("+wavefrontsize64")) 498 MRI.reset(llvm::createGCNMCRegisterInfo(AMDGPUDwarfFlavour::Wave64)); 499 else if (getMCSubtargetInfo()->checkFeatures("+wavefrontsize32")) 500 MRI.reset(llvm::createGCNMCRegisterInfo(AMDGPUDwarfFlavour::Wave32)); 501 } 502 } 503 504 bool AMDGPUTargetMachine::EnableLateStructurizeCFG = false; 505 bool AMDGPUTargetMachine::EnableFunctionCalls = false; 506 bool AMDGPUTargetMachine::EnableFixedFunctionABI = false; 507 bool AMDGPUTargetMachine::EnableLowerModuleLDS = true; 508 509 AMDGPUTargetMachine::~AMDGPUTargetMachine() = default; 510 511 StringRef AMDGPUTargetMachine::getGPUName(const Function &F) const { 512 Attribute GPUAttr = F.getFnAttribute("target-cpu"); 513 return GPUAttr.isValid() ? GPUAttr.getValueAsString() : getTargetCPU(); 514 } 515 516 StringRef AMDGPUTargetMachine::getFeatureString(const Function &F) const { 517 Attribute FSAttr = F.getFnAttribute("target-features"); 518 519 return FSAttr.isValid() ? FSAttr.getValueAsString() 520 : getTargetFeatureString(); 521 } 522 523 /// Predicate for Internalize pass. 524 static bool mustPreserveGV(const GlobalValue &GV) { 525 if (const Function *F = dyn_cast<Function>(&GV)) 526 return F->isDeclaration() || F->getName().startswith("__asan_") || 527 F->getName().startswith("__sanitizer_") || 528 AMDGPU::isEntryFunctionCC(F->getCallingConv()); 529 530 GV.removeDeadConstantUsers(); 531 return !GV.use_empty(); 532 } 533 534 void AMDGPUTargetMachine::adjustPassManager(PassManagerBuilder &Builder) { 535 Builder.DivergentTarget = true; 536 537 bool EnableOpt = getOptLevel() > CodeGenOpt::None; 538 bool Internalize = InternalizeSymbols; 539 bool EarlyInline = EarlyInlineAll && EnableOpt && !EnableFunctionCalls; 540 bool AMDGPUAA = EnableAMDGPUAliasAnalysis && EnableOpt; 541 bool LibCallSimplify = EnableLibCallSimplify && EnableOpt; 542 bool PromoteKernelArguments = 543 EnablePromoteKernelArguments && getOptLevel() > CodeGenOpt::Less; 544 545 if (EnableFunctionCalls) { 546 delete Builder.Inliner; 547 Builder.Inliner = createFunctionInliningPass(); 548 } 549 550 Builder.addExtension( 551 PassManagerBuilder::EP_ModuleOptimizerEarly, 552 [Internalize, EarlyInline, AMDGPUAA, this](const PassManagerBuilder &, 553 legacy::PassManagerBase &PM) { 554 if (AMDGPUAA) { 555 PM.add(createAMDGPUAAWrapperPass()); 556 PM.add(createAMDGPUExternalAAWrapperPass()); 557 } 558 PM.add(createAMDGPUUnifyMetadataPass()); 559 PM.add(createAMDGPUPrintfRuntimeBinding()); 560 if (Internalize) 561 PM.add(createInternalizePass(mustPreserveGV)); 562 PM.add(createAMDGPUPropagateAttributesLatePass(this)); 563 if (Internalize) 564 PM.add(createGlobalDCEPass()); 565 if (EarlyInline) 566 PM.add(createAMDGPUAlwaysInlinePass(false)); 567 }); 568 569 Builder.addExtension( 570 PassManagerBuilder::EP_EarlyAsPossible, 571 [AMDGPUAA, LibCallSimplify, this](const PassManagerBuilder &, 572 legacy::PassManagerBase &PM) { 573 if (AMDGPUAA) { 574 PM.add(createAMDGPUAAWrapperPass()); 575 PM.add(createAMDGPUExternalAAWrapperPass()); 576 } 577 PM.add(llvm::createAMDGPUPropagateAttributesEarlyPass(this)); 578 PM.add(llvm::createAMDGPUUseNativeCallsPass()); 579 if (LibCallSimplify) 580 PM.add(llvm::createAMDGPUSimplifyLibCallsPass(this)); 581 }); 582 583 Builder.addExtension( 584 PassManagerBuilder::EP_CGSCCOptimizerLate, 585 [EnableOpt, PromoteKernelArguments](const PassManagerBuilder &, 586 legacy::PassManagerBase &PM) { 587 // Add promote kernel arguments pass to the opt pipeline right before 588 // infer address spaces which is needed to do actual address space 589 // rewriting. 590 if (PromoteKernelArguments) 591 PM.add(createAMDGPUPromoteKernelArgumentsPass()); 592 593 // Add infer address spaces pass to the opt pipeline after inlining 594 // but before SROA to increase SROA opportunities. 595 PM.add(createInferAddressSpacesPass()); 596 597 // This should run after inlining to have any chance of doing anything, 598 // and before other cleanup optimizations. 599 PM.add(createAMDGPULowerKernelAttributesPass()); 600 601 // Promote alloca to vector before SROA and loop unroll. If we manage 602 // to eliminate allocas before unroll we may choose to unroll less. 603 if (EnableOpt) 604 PM.add(createAMDGPUPromoteAllocaToVector()); 605 }); 606 } 607 608 void AMDGPUTargetMachine::registerDefaultAliasAnalyses(AAManager &AAM) { 609 AAM.registerFunctionAnalysis<AMDGPUAA>(); 610 } 611 612 void AMDGPUTargetMachine::registerPassBuilderCallbacks(PassBuilder &PB) { 613 PB.registerPipelineParsingCallback( 614 [this](StringRef PassName, ModulePassManager &PM, 615 ArrayRef<PassBuilder::PipelineElement>) { 616 if (PassName == "amdgpu-propagate-attributes-late") { 617 PM.addPass(AMDGPUPropagateAttributesLatePass(*this)); 618 return true; 619 } 620 if (PassName == "amdgpu-unify-metadata") { 621 PM.addPass(AMDGPUUnifyMetadataPass()); 622 return true; 623 } 624 if (PassName == "amdgpu-printf-runtime-binding") { 625 PM.addPass(AMDGPUPrintfRuntimeBindingPass()); 626 return true; 627 } 628 if (PassName == "amdgpu-always-inline") { 629 PM.addPass(AMDGPUAlwaysInlinePass()); 630 return true; 631 } 632 if (PassName == "amdgpu-replace-lds-use-with-pointer") { 633 PM.addPass(AMDGPUReplaceLDSUseWithPointerPass()); 634 return true; 635 } 636 if (PassName == "amdgpu-lower-module-lds") { 637 PM.addPass(AMDGPULowerModuleLDSPass()); 638 return true; 639 } 640 return false; 641 }); 642 PB.registerPipelineParsingCallback( 643 [this](StringRef PassName, FunctionPassManager &PM, 644 ArrayRef<PassBuilder::PipelineElement>) { 645 if (PassName == "amdgpu-simplifylib") { 646 PM.addPass(AMDGPUSimplifyLibCallsPass(*this)); 647 return true; 648 } 649 if (PassName == "amdgpu-usenative") { 650 PM.addPass(AMDGPUUseNativeCallsPass()); 651 return true; 652 } 653 if (PassName == "amdgpu-promote-alloca") { 654 PM.addPass(AMDGPUPromoteAllocaPass(*this)); 655 return true; 656 } 657 if (PassName == "amdgpu-promote-alloca-to-vector") { 658 PM.addPass(AMDGPUPromoteAllocaToVectorPass(*this)); 659 return true; 660 } 661 if (PassName == "amdgpu-lower-kernel-attributes") { 662 PM.addPass(AMDGPULowerKernelAttributesPass()); 663 return true; 664 } 665 if (PassName == "amdgpu-propagate-attributes-early") { 666 PM.addPass(AMDGPUPropagateAttributesEarlyPass(*this)); 667 return true; 668 } 669 if (PassName == "amdgpu-promote-kernel-arguments") { 670 PM.addPass(AMDGPUPromoteKernelArgumentsPass()); 671 return true; 672 } 673 return false; 674 }); 675 676 PB.registerAnalysisRegistrationCallback([](FunctionAnalysisManager &FAM) { 677 FAM.registerPass([&] { return AMDGPUAA(); }); 678 }); 679 680 PB.registerParseAACallback([](StringRef AAName, AAManager &AAM) { 681 if (AAName == "amdgpu-aa") { 682 AAM.registerFunctionAnalysis<AMDGPUAA>(); 683 return true; 684 } 685 return false; 686 }); 687 688 PB.registerPipelineStartEPCallback( 689 [this](ModulePassManager &PM, OptimizationLevel Level) { 690 FunctionPassManager FPM; 691 FPM.addPass(AMDGPUPropagateAttributesEarlyPass(*this)); 692 FPM.addPass(AMDGPUUseNativeCallsPass()); 693 if (EnableLibCallSimplify && Level != OptimizationLevel::O0) 694 FPM.addPass(AMDGPUSimplifyLibCallsPass(*this)); 695 PM.addPass(createModuleToFunctionPassAdaptor(std::move(FPM))); 696 }); 697 698 PB.registerPipelineEarlySimplificationEPCallback( 699 [this](ModulePassManager &PM, OptimizationLevel Level) { 700 if (Level == OptimizationLevel::O0) 701 return; 702 703 PM.addPass(AMDGPUUnifyMetadataPass()); 704 PM.addPass(AMDGPUPrintfRuntimeBindingPass()); 705 706 if (InternalizeSymbols) { 707 PM.addPass(InternalizePass(mustPreserveGV)); 708 } 709 PM.addPass(AMDGPUPropagateAttributesLatePass(*this)); 710 if (InternalizeSymbols) { 711 PM.addPass(GlobalDCEPass()); 712 } 713 if (EarlyInlineAll && !EnableFunctionCalls) 714 PM.addPass(AMDGPUAlwaysInlinePass()); 715 }); 716 717 PB.registerCGSCCOptimizerLateEPCallback( 718 [this](CGSCCPassManager &PM, OptimizationLevel Level) { 719 if (Level == OptimizationLevel::O0) 720 return; 721 722 FunctionPassManager FPM; 723 724 // Add promote kernel arguments pass to the opt pipeline right before 725 // infer address spaces which is needed to do actual address space 726 // rewriting. 727 if (Level.getSpeedupLevel() > OptimizationLevel::O1.getSpeedupLevel() && 728 EnablePromoteKernelArguments) 729 FPM.addPass(AMDGPUPromoteKernelArgumentsPass()); 730 731 // Add infer address spaces pass to the opt pipeline after inlining 732 // but before SROA to increase SROA opportunities. 733 FPM.addPass(InferAddressSpacesPass()); 734 735 // This should run after inlining to have any chance of doing 736 // anything, and before other cleanup optimizations. 737 FPM.addPass(AMDGPULowerKernelAttributesPass()); 738 739 if (Level != OptimizationLevel::O0) { 740 // Promote alloca to vector before SROA and loop unroll. If we 741 // manage to eliminate allocas before unroll we may choose to unroll 742 // less. 743 FPM.addPass(AMDGPUPromoteAllocaToVectorPass(*this)); 744 } 745 746 PM.addPass(createCGSCCToFunctionPassAdaptor(std::move(FPM))); 747 }); 748 } 749 750 int64_t AMDGPUTargetMachine::getNullPointerValue(unsigned AddrSpace) { 751 return (AddrSpace == AMDGPUAS::LOCAL_ADDRESS || 752 AddrSpace == AMDGPUAS::PRIVATE_ADDRESS || 753 AddrSpace == AMDGPUAS::REGION_ADDRESS) 754 ? -1 755 : 0; 756 } 757 758 bool AMDGPUTargetMachine::isNoopAddrSpaceCast(unsigned SrcAS, 759 unsigned DestAS) const { 760 return AMDGPU::isFlatGlobalAddrSpace(SrcAS) && 761 AMDGPU::isFlatGlobalAddrSpace(DestAS); 762 } 763 764 unsigned AMDGPUTargetMachine::getAssumedAddrSpace(const Value *V) const { 765 const auto *LD = dyn_cast<LoadInst>(V); 766 if (!LD) 767 return AMDGPUAS::UNKNOWN_ADDRESS_SPACE; 768 769 // It must be a generic pointer loaded. 770 assert(V->getType()->isPointerTy() && 771 V->getType()->getPointerAddressSpace() == AMDGPUAS::FLAT_ADDRESS); 772 773 const auto *Ptr = LD->getPointerOperand(); 774 if (Ptr->getType()->getPointerAddressSpace() != AMDGPUAS::CONSTANT_ADDRESS) 775 return AMDGPUAS::UNKNOWN_ADDRESS_SPACE; 776 // For a generic pointer loaded from the constant memory, it could be assumed 777 // as a global pointer since the constant memory is only populated on the 778 // host side. As implied by the offload programming model, only global 779 // pointers could be referenced on the host side. 780 return AMDGPUAS::GLOBAL_ADDRESS; 781 } 782 783 //===----------------------------------------------------------------------===// 784 // GCN Target Machine (SI+) 785 //===----------------------------------------------------------------------===// 786 787 GCNTargetMachine::GCNTargetMachine(const Target &T, const Triple &TT, 788 StringRef CPU, StringRef FS, 789 TargetOptions Options, 790 Optional<Reloc::Model> RM, 791 Optional<CodeModel::Model> CM, 792 CodeGenOpt::Level OL, bool JIT) 793 : AMDGPUTargetMachine(T, TT, CPU, FS, Options, RM, CM, OL) {} 794 795 const TargetSubtargetInfo * 796 GCNTargetMachine::getSubtargetImpl(const Function &F) const { 797 StringRef GPU = getGPUName(F); 798 StringRef FS = getFeatureString(F); 799 800 SmallString<128> SubtargetKey(GPU); 801 SubtargetKey.append(FS); 802 803 auto &I = SubtargetMap[SubtargetKey]; 804 if (!I) { 805 // This needs to be done before we create a new subtarget since any 806 // creation will depend on the TM and the code generation flags on the 807 // function that reside in TargetOptions. 808 resetTargetOptions(F); 809 I = std::make_unique<GCNSubtarget>(TargetTriple, GPU, FS, *this); 810 } 811 812 I->setScalarizeGlobalBehavior(ScalarizeGlobal); 813 814 return I.get(); 815 } 816 817 TargetTransformInfo 818 GCNTargetMachine::getTargetTransformInfo(const Function &F) { 819 return TargetTransformInfo(GCNTTIImpl(this, F)); 820 } 821 822 //===----------------------------------------------------------------------===// 823 // AMDGPU Pass Setup 824 //===----------------------------------------------------------------------===// 825 826 std::unique_ptr<CSEConfigBase> llvm::AMDGPUPassConfig::getCSEConfig() const { 827 return getStandardCSEConfigForOpt(TM->getOptLevel()); 828 } 829 830 namespace { 831 832 class GCNPassConfig final : public AMDGPUPassConfig { 833 public: 834 GCNPassConfig(LLVMTargetMachine &TM, PassManagerBase &PM) 835 : AMDGPUPassConfig(TM, PM) { 836 // It is necessary to know the register usage of the entire call graph. We 837 // allow calls without EnableAMDGPUFunctionCalls if they are marked 838 // noinline, so this is always required. 839 setRequiresCodeGenSCCOrder(true); 840 substitutePass(&PostRASchedulerID, &PostMachineSchedulerID); 841 } 842 843 GCNTargetMachine &getGCNTargetMachine() const { 844 return getTM<GCNTargetMachine>(); 845 } 846 847 ScheduleDAGInstrs * 848 createMachineScheduler(MachineSchedContext *C) const override; 849 850 ScheduleDAGInstrs * 851 createPostMachineScheduler(MachineSchedContext *C) const override { 852 ScheduleDAGMI *DAG = createGenericSchedPostRA(C); 853 const GCNSubtarget &ST = C->MF->getSubtarget<GCNSubtarget>(); 854 DAG->addMutation(ST.createFillMFMAShadowMutation(DAG->TII)); 855 return DAG; 856 } 857 858 bool addPreISel() override; 859 void addMachineSSAOptimization() override; 860 bool addILPOpts() override; 861 bool addInstSelector() override; 862 bool addIRTranslator() override; 863 void addPreLegalizeMachineIR() override; 864 bool addLegalizeMachineIR() override; 865 void addPreRegBankSelect() override; 866 bool addRegBankSelect() override; 867 void addPreGlobalInstructionSelect() override; 868 bool addGlobalInstructionSelect() override; 869 void addFastRegAlloc() override; 870 void addOptimizedRegAlloc() override; 871 872 FunctionPass *createSGPRAllocPass(bool Optimized); 873 FunctionPass *createVGPRAllocPass(bool Optimized); 874 FunctionPass *createRegAllocPass(bool Optimized) override; 875 876 bool addRegAssignAndRewriteFast() override; 877 bool addRegAssignAndRewriteOptimized() override; 878 879 void addPreRegAlloc() override; 880 bool addPreRewrite() override; 881 void addPostRegAlloc() override; 882 void addPreSched2() override; 883 void addPreEmitPass() override; 884 }; 885 886 } // end anonymous namespace 887 888 AMDGPUPassConfig::AMDGPUPassConfig(LLVMTargetMachine &TM, PassManagerBase &PM) 889 : TargetPassConfig(TM, PM) { 890 // Exceptions and StackMaps are not supported, so these passes will never do 891 // anything. 892 disablePass(&StackMapLivenessID); 893 disablePass(&FuncletLayoutID); 894 // Garbage collection is not supported. 895 disablePass(&GCLoweringID); 896 disablePass(&ShadowStackGCLoweringID); 897 } 898 899 void AMDGPUPassConfig::addEarlyCSEOrGVNPass() { 900 if (getOptLevel() == CodeGenOpt::Aggressive) 901 addPass(createGVNPass()); 902 else 903 addPass(createEarlyCSEPass()); 904 } 905 906 void AMDGPUPassConfig::addStraightLineScalarOptimizationPasses() { 907 addPass(createLICMPass()); 908 addPass(createSeparateConstOffsetFromGEPPass()); 909 addPass(createSpeculativeExecutionPass()); 910 // ReassociateGEPs exposes more opportunities for SLSR. See 911 // the example in reassociate-geps-and-slsr.ll. 912 addPass(createStraightLineStrengthReducePass()); 913 // SeparateConstOffsetFromGEP and SLSR creates common expressions which GVN or 914 // EarlyCSE can reuse. 915 addEarlyCSEOrGVNPass(); 916 // Run NaryReassociate after EarlyCSE/GVN to be more effective. 917 addPass(createNaryReassociatePass()); 918 // NaryReassociate on GEPs creates redundant common expressions, so run 919 // EarlyCSE after it. 920 addPass(createEarlyCSEPass()); 921 } 922 923 void AMDGPUPassConfig::addIRPasses() { 924 const AMDGPUTargetMachine &TM = getAMDGPUTargetMachine(); 925 926 // There is no reason to run these. 927 disablePass(&StackMapLivenessID); 928 disablePass(&FuncletLayoutID); 929 disablePass(&PatchableFunctionID); 930 931 addPass(createAMDGPUPrintfRuntimeBinding()); 932 addPass(createAMDGPUCtorDtorLoweringPass()); 933 934 // This must occur before inlining, as the inliner will not look through 935 // bitcast calls. 936 addPass(createAMDGPUFixFunctionBitcastsPass()); 937 938 // A call to propagate attributes pass in the backend in case opt was not run. 939 addPass(createAMDGPUPropagateAttributesEarlyPass(&TM)); 940 941 addPass(createAMDGPULowerIntrinsicsPass()); 942 943 // Function calls are not supported, so make sure we inline everything. 944 addPass(createAMDGPUAlwaysInlinePass()); 945 addPass(createAlwaysInlinerLegacyPass()); 946 // We need to add the barrier noop pass, otherwise adding the function 947 // inlining pass will cause all of the PassConfigs passes to be run 948 // one function at a time, which means if we have a nodule with two 949 // functions, then we will generate code for the first function 950 // without ever running any passes on the second. 951 addPass(createBarrierNoopPass()); 952 953 // Handle uses of OpenCL image2d_t, image3d_t and sampler_t arguments. 954 if (TM.getTargetTriple().getArch() == Triple::r600) 955 addPass(createR600OpenCLImageTypeLoweringPass()); 956 957 // Replace OpenCL enqueued block function pointers with global variables. 958 addPass(createAMDGPUOpenCLEnqueuedBlockLoweringPass()); 959 960 // Can increase LDS used by kernel so runs before PromoteAlloca 961 if (EnableLowerModuleLDS) { 962 // The pass "amdgpu-replace-lds-use-with-pointer" need to be run before the 963 // pass "amdgpu-lower-module-lds", and also it required to be run only if 964 // "amdgpu-lower-module-lds" pass is enabled. 965 if (EnableLDSReplaceWithPointer) 966 addPass(createAMDGPUReplaceLDSUseWithPointerPass()); 967 968 addPass(createAMDGPULowerModuleLDSPass()); 969 } 970 971 if (TM.getOptLevel() > CodeGenOpt::None) 972 addPass(createInferAddressSpacesPass()); 973 974 addPass(createAtomicExpandPass()); 975 976 if (TM.getOptLevel() > CodeGenOpt::None) { 977 addPass(createAMDGPUPromoteAlloca()); 978 979 if (EnableSROA) 980 addPass(createSROAPass()); 981 if (isPassEnabled(EnableScalarIRPasses)) 982 addStraightLineScalarOptimizationPasses(); 983 984 if (EnableAMDGPUAliasAnalysis) { 985 addPass(createAMDGPUAAWrapperPass()); 986 addPass(createExternalAAWrapperPass([](Pass &P, Function &, 987 AAResults &AAR) { 988 if (auto *WrapperPass = P.getAnalysisIfAvailable<AMDGPUAAWrapperPass>()) 989 AAR.addAAResult(WrapperPass->getResult()); 990 })); 991 } 992 993 if (TM.getTargetTriple().getArch() == Triple::amdgcn) { 994 // TODO: May want to move later or split into an early and late one. 995 addPass(createAMDGPUCodeGenPreparePass()); 996 } 997 } 998 999 TargetPassConfig::addIRPasses(); 1000 1001 // EarlyCSE is not always strong enough to clean up what LSR produces. For 1002 // example, GVN can combine 1003 // 1004 // %0 = add %a, %b 1005 // %1 = add %b, %a 1006 // 1007 // and 1008 // 1009 // %0 = shl nsw %a, 2 1010 // %1 = shl %a, 2 1011 // 1012 // but EarlyCSE can do neither of them. 1013 if (isPassEnabled(EnableScalarIRPasses)) 1014 addEarlyCSEOrGVNPass(); 1015 } 1016 1017 void AMDGPUPassConfig::addCodeGenPrepare() { 1018 if (TM->getTargetTriple().getArch() == Triple::amdgcn) { 1019 addPass(createAMDGPUAttributorPass()); 1020 1021 // FIXME: This pass adds 2 hacky attributes that can be replaced with an 1022 // analysis, and should be removed. 1023 addPass(createAMDGPUAnnotateKernelFeaturesPass()); 1024 } 1025 1026 if (TM->getTargetTriple().getArch() == Triple::amdgcn && 1027 EnableLowerKernelArguments) 1028 addPass(createAMDGPULowerKernelArgumentsPass()); 1029 1030 TargetPassConfig::addCodeGenPrepare(); 1031 1032 if (isPassEnabled(EnableLoadStoreVectorizer)) 1033 addPass(createLoadStoreVectorizerPass()); 1034 1035 // LowerSwitch pass may introduce unreachable blocks that can 1036 // cause unexpected behavior for subsequent passes. Placing it 1037 // here seems better that these blocks would get cleaned up by 1038 // UnreachableBlockElim inserted next in the pass flow. 1039 addPass(createLowerSwitchPass()); 1040 } 1041 1042 bool AMDGPUPassConfig::addPreISel() { 1043 if (TM->getOptLevel() > CodeGenOpt::None) 1044 addPass(createFlattenCFGPass()); 1045 return false; 1046 } 1047 1048 bool AMDGPUPassConfig::addInstSelector() { 1049 addPass(createAMDGPUISelDag(&getAMDGPUTargetMachine(), getOptLevel())); 1050 return false; 1051 } 1052 1053 bool AMDGPUPassConfig::addGCPasses() { 1054 // Do nothing. GC is not supported. 1055 return false; 1056 } 1057 1058 llvm::ScheduleDAGInstrs * 1059 AMDGPUPassConfig::createMachineScheduler(MachineSchedContext *C) const { 1060 ScheduleDAGMILive *DAG = createGenericSchedLive(C); 1061 DAG->addMutation(createLoadClusterDAGMutation(DAG->TII, DAG->TRI)); 1062 return DAG; 1063 } 1064 1065 //===----------------------------------------------------------------------===// 1066 // GCN Pass Setup 1067 //===----------------------------------------------------------------------===// 1068 1069 ScheduleDAGInstrs *GCNPassConfig::createMachineScheduler( 1070 MachineSchedContext *C) const { 1071 const GCNSubtarget &ST = C->MF->getSubtarget<GCNSubtarget>(); 1072 if (ST.enableSIScheduler()) 1073 return createSIMachineScheduler(C); 1074 return createGCNMaxOccupancyMachineScheduler(C); 1075 } 1076 1077 bool GCNPassConfig::addPreISel() { 1078 AMDGPUPassConfig::addPreISel(); 1079 1080 if (TM->getOptLevel() > CodeGenOpt::None) 1081 addPass(createAMDGPULateCodeGenPreparePass()); 1082 1083 if (isPassEnabled(EnableAtomicOptimizations, CodeGenOpt::Less)) { 1084 addPass(createAMDGPUAtomicOptimizerPass()); 1085 } 1086 1087 if (TM->getOptLevel() > CodeGenOpt::None) 1088 addPass(createSinkingPass()); 1089 1090 // Merge divergent exit nodes. StructurizeCFG won't recognize the multi-exit 1091 // regions formed by them. 1092 addPass(&AMDGPUUnifyDivergentExitNodesID); 1093 if (!LateCFGStructurize) { 1094 if (EnableStructurizerWorkarounds) { 1095 addPass(createFixIrreduciblePass()); 1096 addPass(createUnifyLoopExitsPass()); 1097 } 1098 addPass(createStructurizeCFGPass(false)); // true -> SkipUniformRegions 1099 } 1100 addPass(createAMDGPUAnnotateUniformValues()); 1101 if (!LateCFGStructurize) { 1102 addPass(createSIAnnotateControlFlowPass()); 1103 } 1104 addPass(createLCSSAPass()); 1105 1106 if (TM->getOptLevel() > CodeGenOpt::Less) 1107 addPass(&AMDGPUPerfHintAnalysisID); 1108 1109 return false; 1110 } 1111 1112 void GCNPassConfig::addMachineSSAOptimization() { 1113 TargetPassConfig::addMachineSSAOptimization(); 1114 1115 // We want to fold operands after PeepholeOptimizer has run (or as part of 1116 // it), because it will eliminate extra copies making it easier to fold the 1117 // real source operand. We want to eliminate dead instructions after, so that 1118 // we see fewer uses of the copies. We then need to clean up the dead 1119 // instructions leftover after the operands are folded as well. 1120 // 1121 // XXX - Can we get away without running DeadMachineInstructionElim again? 1122 addPass(&SIFoldOperandsID); 1123 if (EnableDPPCombine) 1124 addPass(&GCNDPPCombineID); 1125 addPass(&SILoadStoreOptimizerID); 1126 if (isPassEnabled(EnableSDWAPeephole)) { 1127 addPass(&SIPeepholeSDWAID); 1128 addPass(&EarlyMachineLICMID); 1129 addPass(&MachineCSEID); 1130 addPass(&SIFoldOperandsID); 1131 } 1132 addPass(&DeadMachineInstructionElimID); 1133 addPass(createSIShrinkInstructionsPass()); 1134 } 1135 1136 bool GCNPassConfig::addILPOpts() { 1137 if (EnableEarlyIfConversion) 1138 addPass(&EarlyIfConverterID); 1139 1140 TargetPassConfig::addILPOpts(); 1141 return false; 1142 } 1143 1144 bool GCNPassConfig::addInstSelector() { 1145 AMDGPUPassConfig::addInstSelector(); 1146 addPass(&SIFixSGPRCopiesID); 1147 addPass(createSILowerI1CopiesPass()); 1148 return false; 1149 } 1150 1151 bool GCNPassConfig::addIRTranslator() { 1152 addPass(new IRTranslator(getOptLevel())); 1153 return false; 1154 } 1155 1156 void GCNPassConfig::addPreLegalizeMachineIR() { 1157 bool IsOptNone = getOptLevel() == CodeGenOpt::None; 1158 addPass(createAMDGPUPreLegalizeCombiner(IsOptNone)); 1159 addPass(new Localizer()); 1160 } 1161 1162 bool GCNPassConfig::addLegalizeMachineIR() { 1163 addPass(new Legalizer()); 1164 return false; 1165 } 1166 1167 void GCNPassConfig::addPreRegBankSelect() { 1168 bool IsOptNone = getOptLevel() == CodeGenOpt::None; 1169 addPass(createAMDGPUPostLegalizeCombiner(IsOptNone)); 1170 } 1171 1172 bool GCNPassConfig::addRegBankSelect() { 1173 addPass(new RegBankSelect()); 1174 return false; 1175 } 1176 1177 void GCNPassConfig::addPreGlobalInstructionSelect() { 1178 bool IsOptNone = getOptLevel() == CodeGenOpt::None; 1179 addPass(createAMDGPURegBankCombiner(IsOptNone)); 1180 } 1181 1182 bool GCNPassConfig::addGlobalInstructionSelect() { 1183 addPass(new InstructionSelect(getOptLevel())); 1184 return false; 1185 } 1186 1187 void GCNPassConfig::addPreRegAlloc() { 1188 if (LateCFGStructurize) { 1189 addPass(createAMDGPUMachineCFGStructurizerPass()); 1190 } 1191 } 1192 1193 void GCNPassConfig::addFastRegAlloc() { 1194 // FIXME: We have to disable the verifier here because of PHIElimination + 1195 // TwoAddressInstructions disabling it. 1196 1197 // This must be run immediately after phi elimination and before 1198 // TwoAddressInstructions, otherwise the processing of the tied operand of 1199 // SI_ELSE will introduce a copy of the tied operand source after the else. 1200 insertPass(&PHIEliminationID, &SILowerControlFlowID, false); 1201 1202 insertPass(&TwoAddressInstructionPassID, &SIWholeQuadModeID); 1203 insertPass(&TwoAddressInstructionPassID, &SIPreAllocateWWMRegsID); 1204 1205 TargetPassConfig::addFastRegAlloc(); 1206 } 1207 1208 void GCNPassConfig::addOptimizedRegAlloc() { 1209 // Allow the scheduler to run before SIWholeQuadMode inserts exec manipulation 1210 // instructions that cause scheduling barriers. 1211 insertPass(&MachineSchedulerID, &SIWholeQuadModeID); 1212 insertPass(&MachineSchedulerID, &SIPreAllocateWWMRegsID); 1213 1214 if (OptExecMaskPreRA) 1215 insertPass(&MachineSchedulerID, &SIOptimizeExecMaskingPreRAID); 1216 1217 if (isPassEnabled(EnablePreRAOptimizations)) 1218 insertPass(&RenameIndependentSubregsID, &GCNPreRAOptimizationsID); 1219 1220 // This is not an essential optimization and it has a noticeable impact on 1221 // compilation time, so we only enable it from O2. 1222 if (TM->getOptLevel() > CodeGenOpt::Less) 1223 insertPass(&MachineSchedulerID, &SIFormMemoryClausesID); 1224 1225 // FIXME: when an instruction has a Killed operand, and the instruction is 1226 // inside a bundle, seems only the BUNDLE instruction appears as the Kills of 1227 // the register in LiveVariables, this would trigger a failure in verifier, 1228 // we should fix it and enable the verifier. 1229 if (OptVGPRLiveRange) 1230 insertPass(&LiveVariablesID, &SIOptimizeVGPRLiveRangeID, false); 1231 // This must be run immediately after phi elimination and before 1232 // TwoAddressInstructions, otherwise the processing of the tied operand of 1233 // SI_ELSE will introduce a copy of the tied operand source after the else. 1234 insertPass(&PHIEliminationID, &SILowerControlFlowID, false); 1235 1236 if (EnableDCEInRA) 1237 insertPass(&DetectDeadLanesID, &DeadMachineInstructionElimID); 1238 1239 TargetPassConfig::addOptimizedRegAlloc(); 1240 } 1241 1242 bool GCNPassConfig::addPreRewrite() { 1243 if (EnableRegReassign) 1244 addPass(&GCNNSAReassignID); 1245 return true; 1246 } 1247 1248 FunctionPass *GCNPassConfig::createSGPRAllocPass(bool Optimized) { 1249 // Initialize the global default. 1250 llvm::call_once(InitializeDefaultSGPRRegisterAllocatorFlag, 1251 initializeDefaultSGPRRegisterAllocatorOnce); 1252 1253 RegisterRegAlloc::FunctionPassCtor Ctor = SGPRRegisterRegAlloc::getDefault(); 1254 if (Ctor != useDefaultRegisterAllocator) 1255 return Ctor(); 1256 1257 if (Optimized) 1258 return createGreedyRegisterAllocator(onlyAllocateSGPRs); 1259 1260 return createFastRegisterAllocator(onlyAllocateSGPRs, false); 1261 } 1262 1263 FunctionPass *GCNPassConfig::createVGPRAllocPass(bool Optimized) { 1264 // Initialize the global default. 1265 llvm::call_once(InitializeDefaultVGPRRegisterAllocatorFlag, 1266 initializeDefaultVGPRRegisterAllocatorOnce); 1267 1268 RegisterRegAlloc::FunctionPassCtor Ctor = VGPRRegisterRegAlloc::getDefault(); 1269 if (Ctor != useDefaultRegisterAllocator) 1270 return Ctor(); 1271 1272 if (Optimized) 1273 return createGreedyVGPRRegisterAllocator(); 1274 1275 return createFastVGPRRegisterAllocator(); 1276 } 1277 1278 FunctionPass *GCNPassConfig::createRegAllocPass(bool Optimized) { 1279 llvm_unreachable("should not be used"); 1280 } 1281 1282 static const char RegAllocOptNotSupportedMessage[] = 1283 "-regalloc not supported with amdgcn. Use -sgpr-regalloc and -vgpr-regalloc"; 1284 1285 bool GCNPassConfig::addRegAssignAndRewriteFast() { 1286 if (!usingDefaultRegAlloc()) 1287 report_fatal_error(RegAllocOptNotSupportedMessage); 1288 1289 addPass(createSGPRAllocPass(false)); 1290 1291 // Equivalent of PEI for SGPRs. 1292 addPass(&SILowerSGPRSpillsID); 1293 1294 addPass(createVGPRAllocPass(false)); 1295 return true; 1296 } 1297 1298 bool GCNPassConfig::addRegAssignAndRewriteOptimized() { 1299 if (!usingDefaultRegAlloc()) 1300 report_fatal_error(RegAllocOptNotSupportedMessage); 1301 1302 addPass(createSGPRAllocPass(true)); 1303 1304 // Commit allocated register changes. This is mostly necessary because too 1305 // many things rely on the use lists of the physical registers, such as the 1306 // verifier. This is only necessary with allocators which use LiveIntervals, 1307 // since FastRegAlloc does the replacements itself. 1308 addPass(createVirtRegRewriter(false)); 1309 1310 // Equivalent of PEI for SGPRs. 1311 addPass(&SILowerSGPRSpillsID); 1312 1313 addPass(createVGPRAllocPass(true)); 1314 1315 addPreRewrite(); 1316 addPass(&VirtRegRewriterID); 1317 1318 return true; 1319 } 1320 1321 void GCNPassConfig::addPostRegAlloc() { 1322 addPass(&SIFixVGPRCopiesID); 1323 if (getOptLevel() > CodeGenOpt::None) 1324 addPass(&SIOptimizeExecMaskingID); 1325 TargetPassConfig::addPostRegAlloc(); 1326 } 1327 1328 void GCNPassConfig::addPreSched2() { 1329 addPass(&SIPostRABundlerID); 1330 } 1331 1332 void GCNPassConfig::addPreEmitPass() { 1333 addPass(createSIMemoryLegalizerPass()); 1334 addPass(createSIInsertWaitcntsPass()); 1335 1336 if (TM->getOptLevel() > CodeGenOpt::None) 1337 addPass(createSIShrinkInstructionsPass()); 1338 1339 addPass(createSIModeRegisterPass()); 1340 1341 if (getOptLevel() > CodeGenOpt::None) 1342 addPass(&SIInsertHardClausesID); 1343 1344 addPass(&SILateBranchLoweringPassID); 1345 if (getOptLevel() > CodeGenOpt::None) 1346 addPass(&SIPreEmitPeepholeID); 1347 // The hazard recognizer that runs as part of the post-ra scheduler does not 1348 // guarantee to be able handle all hazards correctly. This is because if there 1349 // are multiple scheduling regions in a basic block, the regions are scheduled 1350 // bottom up, so when we begin to schedule a region we don't know what 1351 // instructions were emitted directly before it. 1352 // 1353 // Here we add a stand-alone hazard recognizer pass which can handle all 1354 // cases. 1355 addPass(&PostRAHazardRecognizerID); 1356 addPass(&BranchRelaxationPassID); 1357 } 1358 1359 TargetPassConfig *GCNTargetMachine::createPassConfig(PassManagerBase &PM) { 1360 return new GCNPassConfig(*this, PM); 1361 } 1362 1363 yaml::MachineFunctionInfo *GCNTargetMachine::createDefaultFuncInfoYAML() const { 1364 return new yaml::SIMachineFunctionInfo(); 1365 } 1366 1367 yaml::MachineFunctionInfo * 1368 GCNTargetMachine::convertFuncInfoToYAML(const MachineFunction &MF) const { 1369 const SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>(); 1370 return new yaml::SIMachineFunctionInfo( 1371 *MFI, *MF.getSubtarget().getRegisterInfo(), MF); 1372 } 1373 1374 bool GCNTargetMachine::parseMachineFunctionInfo( 1375 const yaml::MachineFunctionInfo &MFI_, PerFunctionMIParsingState &PFS, 1376 SMDiagnostic &Error, SMRange &SourceRange) const { 1377 const yaml::SIMachineFunctionInfo &YamlMFI = 1378 reinterpret_cast<const yaml::SIMachineFunctionInfo &>(MFI_); 1379 MachineFunction &MF = PFS.MF; 1380 SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>(); 1381 1382 if (MFI->initializeBaseYamlFields(YamlMFI, MF, PFS, Error, SourceRange)) 1383 return true; 1384 1385 if (MFI->Occupancy == 0) { 1386 // Fixup the subtarget dependent default value. 1387 const GCNSubtarget &ST = MF.getSubtarget<GCNSubtarget>(); 1388 MFI->Occupancy = ST.computeOccupancy(MF.getFunction(), MFI->getLDSSize()); 1389 } 1390 1391 auto parseRegister = [&](const yaml::StringValue &RegName, Register &RegVal) { 1392 Register TempReg; 1393 if (parseNamedRegisterReference(PFS, TempReg, RegName.Value, Error)) { 1394 SourceRange = RegName.SourceRange; 1395 return true; 1396 } 1397 RegVal = TempReg; 1398 1399 return false; 1400 }; 1401 1402 auto diagnoseRegisterClass = [&](const yaml::StringValue &RegName) { 1403 // Create a diagnostic for a the register string literal. 1404 const MemoryBuffer &Buffer = 1405 *PFS.SM->getMemoryBuffer(PFS.SM->getMainFileID()); 1406 Error = SMDiagnostic(*PFS.SM, SMLoc(), Buffer.getBufferIdentifier(), 1, 1407 RegName.Value.size(), SourceMgr::DK_Error, 1408 "incorrect register class for field", RegName.Value, 1409 None, None); 1410 SourceRange = RegName.SourceRange; 1411 return true; 1412 }; 1413 1414 if (parseRegister(YamlMFI.ScratchRSrcReg, MFI->ScratchRSrcReg) || 1415 parseRegister(YamlMFI.FrameOffsetReg, MFI->FrameOffsetReg) || 1416 parseRegister(YamlMFI.StackPtrOffsetReg, MFI->StackPtrOffsetReg)) 1417 return true; 1418 1419 if (MFI->ScratchRSrcReg != AMDGPU::PRIVATE_RSRC_REG && 1420 !AMDGPU::SGPR_128RegClass.contains(MFI->ScratchRSrcReg)) { 1421 return diagnoseRegisterClass(YamlMFI.ScratchRSrcReg); 1422 } 1423 1424 if (MFI->FrameOffsetReg != AMDGPU::FP_REG && 1425 !AMDGPU::SGPR_32RegClass.contains(MFI->FrameOffsetReg)) { 1426 return diagnoseRegisterClass(YamlMFI.FrameOffsetReg); 1427 } 1428 1429 if (MFI->StackPtrOffsetReg != AMDGPU::SP_REG && 1430 !AMDGPU::SGPR_32RegClass.contains(MFI->StackPtrOffsetReg)) { 1431 return diagnoseRegisterClass(YamlMFI.StackPtrOffsetReg); 1432 } 1433 1434 auto parseAndCheckArgument = [&](const Optional<yaml::SIArgument> &A, 1435 const TargetRegisterClass &RC, 1436 ArgDescriptor &Arg, unsigned UserSGPRs, 1437 unsigned SystemSGPRs) { 1438 // Skip parsing if it's not present. 1439 if (!A) 1440 return false; 1441 1442 if (A->IsRegister) { 1443 Register Reg; 1444 if (parseNamedRegisterReference(PFS, Reg, A->RegisterName.Value, Error)) { 1445 SourceRange = A->RegisterName.SourceRange; 1446 return true; 1447 } 1448 if (!RC.contains(Reg)) 1449 return diagnoseRegisterClass(A->RegisterName); 1450 Arg = ArgDescriptor::createRegister(Reg); 1451 } else 1452 Arg = ArgDescriptor::createStack(A->StackOffset); 1453 // Check and apply the optional mask. 1454 if (A->Mask) 1455 Arg = ArgDescriptor::createArg(Arg, A->Mask.getValue()); 1456 1457 MFI->NumUserSGPRs += UserSGPRs; 1458 MFI->NumSystemSGPRs += SystemSGPRs; 1459 return false; 1460 }; 1461 1462 if (YamlMFI.ArgInfo && 1463 (parseAndCheckArgument(YamlMFI.ArgInfo->PrivateSegmentBuffer, 1464 AMDGPU::SGPR_128RegClass, 1465 MFI->ArgInfo.PrivateSegmentBuffer, 4, 0) || 1466 parseAndCheckArgument(YamlMFI.ArgInfo->DispatchPtr, 1467 AMDGPU::SReg_64RegClass, MFI->ArgInfo.DispatchPtr, 1468 2, 0) || 1469 parseAndCheckArgument(YamlMFI.ArgInfo->QueuePtr, AMDGPU::SReg_64RegClass, 1470 MFI->ArgInfo.QueuePtr, 2, 0) || 1471 parseAndCheckArgument(YamlMFI.ArgInfo->KernargSegmentPtr, 1472 AMDGPU::SReg_64RegClass, 1473 MFI->ArgInfo.KernargSegmentPtr, 2, 0) || 1474 parseAndCheckArgument(YamlMFI.ArgInfo->DispatchID, 1475 AMDGPU::SReg_64RegClass, MFI->ArgInfo.DispatchID, 1476 2, 0) || 1477 parseAndCheckArgument(YamlMFI.ArgInfo->FlatScratchInit, 1478 AMDGPU::SReg_64RegClass, 1479 MFI->ArgInfo.FlatScratchInit, 2, 0) || 1480 parseAndCheckArgument(YamlMFI.ArgInfo->PrivateSegmentSize, 1481 AMDGPU::SGPR_32RegClass, 1482 MFI->ArgInfo.PrivateSegmentSize, 0, 0) || 1483 parseAndCheckArgument(YamlMFI.ArgInfo->WorkGroupIDX, 1484 AMDGPU::SGPR_32RegClass, MFI->ArgInfo.WorkGroupIDX, 1485 0, 1) || 1486 parseAndCheckArgument(YamlMFI.ArgInfo->WorkGroupIDY, 1487 AMDGPU::SGPR_32RegClass, MFI->ArgInfo.WorkGroupIDY, 1488 0, 1) || 1489 parseAndCheckArgument(YamlMFI.ArgInfo->WorkGroupIDZ, 1490 AMDGPU::SGPR_32RegClass, MFI->ArgInfo.WorkGroupIDZ, 1491 0, 1) || 1492 parseAndCheckArgument(YamlMFI.ArgInfo->WorkGroupInfo, 1493 AMDGPU::SGPR_32RegClass, 1494 MFI->ArgInfo.WorkGroupInfo, 0, 1) || 1495 parseAndCheckArgument(YamlMFI.ArgInfo->PrivateSegmentWaveByteOffset, 1496 AMDGPU::SGPR_32RegClass, 1497 MFI->ArgInfo.PrivateSegmentWaveByteOffset, 0, 1) || 1498 parseAndCheckArgument(YamlMFI.ArgInfo->ImplicitArgPtr, 1499 AMDGPU::SReg_64RegClass, 1500 MFI->ArgInfo.ImplicitArgPtr, 0, 0) || 1501 parseAndCheckArgument(YamlMFI.ArgInfo->ImplicitBufferPtr, 1502 AMDGPU::SReg_64RegClass, 1503 MFI->ArgInfo.ImplicitBufferPtr, 2, 0) || 1504 parseAndCheckArgument(YamlMFI.ArgInfo->WorkItemIDX, 1505 AMDGPU::VGPR_32RegClass, 1506 MFI->ArgInfo.WorkItemIDX, 0, 0) || 1507 parseAndCheckArgument(YamlMFI.ArgInfo->WorkItemIDY, 1508 AMDGPU::VGPR_32RegClass, 1509 MFI->ArgInfo.WorkItemIDY, 0, 0) || 1510 parseAndCheckArgument(YamlMFI.ArgInfo->WorkItemIDZ, 1511 AMDGPU::VGPR_32RegClass, 1512 MFI->ArgInfo.WorkItemIDZ, 0, 0))) 1513 return true; 1514 1515 MFI->Mode.IEEE = YamlMFI.Mode.IEEE; 1516 MFI->Mode.DX10Clamp = YamlMFI.Mode.DX10Clamp; 1517 MFI->Mode.FP32InputDenormals = YamlMFI.Mode.FP32InputDenormals; 1518 MFI->Mode.FP32OutputDenormals = YamlMFI.Mode.FP32OutputDenormals; 1519 MFI->Mode.FP64FP16InputDenormals = YamlMFI.Mode.FP64FP16InputDenormals; 1520 MFI->Mode.FP64FP16OutputDenormals = YamlMFI.Mode.FP64FP16OutputDenormals; 1521 1522 return false; 1523 } 1524