1 //===- Parsing, selection, and construction of pass pipelines -------------===// 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 /// \file 9 /// 10 /// This file provides the implementation of the PassBuilder based on our 11 /// static pass registry as well as related functionality. It also provides 12 /// helpers to aid in analyzing, debugging, and testing passes and pass 13 /// pipelines. 14 /// 15 //===----------------------------------------------------------------------===// 16 17 #include "llvm/Passes/PassBuilder.h" 18 #include "llvm/ADT/StringSwitch.h" 19 #include "llvm/Analysis/AliasAnalysis.h" 20 #include "llvm/Analysis/AliasAnalysisEvaluator.h" 21 #include "llvm/Analysis/AssumptionCache.h" 22 #include "llvm/Analysis/BasicAliasAnalysis.h" 23 #include "llvm/Analysis/BlockFrequencyInfo.h" 24 #include "llvm/Analysis/BranchProbabilityInfo.h" 25 #include "llvm/Analysis/CFGPrinter.h" 26 #include "llvm/Analysis/CFLAndersAliasAnalysis.h" 27 #include "llvm/Analysis/CFLSteensAliasAnalysis.h" 28 #include "llvm/Analysis/CGSCCPassManager.h" 29 #include "llvm/Analysis/CallGraph.h" 30 #include "llvm/Analysis/DDG.h" 31 #include "llvm/Analysis/DemandedBits.h" 32 #include "llvm/Analysis/DependenceAnalysis.h" 33 #include "llvm/Analysis/DominanceFrontier.h" 34 #include "llvm/Analysis/GlobalsModRef.h" 35 #include "llvm/Analysis/IVUsers.h" 36 #include "llvm/Analysis/LazyCallGraph.h" 37 #include "llvm/Analysis/LazyValueInfo.h" 38 #include "llvm/Analysis/LoopAccessAnalysis.h" 39 #include "llvm/Analysis/LoopCacheAnalysis.h" 40 #include "llvm/Analysis/LoopInfo.h" 41 #include "llvm/Analysis/MemoryDependenceAnalysis.h" 42 #include "llvm/Analysis/MemorySSA.h" 43 #include "llvm/Analysis/ModuleSummaryAnalysis.h" 44 #include "llvm/Analysis/OptimizationRemarkEmitter.h" 45 #include "llvm/Analysis/PhiValues.h" 46 #include "llvm/Analysis/PostDominators.h" 47 #include "llvm/Analysis/ProfileSummaryInfo.h" 48 #include "llvm/Analysis/RegionInfo.h" 49 #include "llvm/Analysis/ScalarEvolution.h" 50 #include "llvm/Analysis/ScalarEvolutionAliasAnalysis.h" 51 #include "llvm/Analysis/ScopedNoAliasAA.h" 52 #include "llvm/Analysis/StackSafetyAnalysis.h" 53 #include "llvm/Analysis/TargetLibraryInfo.h" 54 #include "llvm/Analysis/TargetTransformInfo.h" 55 #include "llvm/Analysis/TypeBasedAliasAnalysis.h" 56 #include "llvm/CodeGen/MachineModuleInfo.h" 57 #include "llvm/CodeGen/PreISelIntrinsicLowering.h" 58 #include "llvm/CodeGen/UnreachableBlockElim.h" 59 #include "llvm/IR/Dominators.h" 60 #include "llvm/IR/IRPrintingPasses.h" 61 #include "llvm/IR/PassManager.h" 62 #include "llvm/IR/SafepointIRVerifier.h" 63 #include "llvm/IR/Verifier.h" 64 #include "llvm/Support/CommandLine.h" 65 #include "llvm/Support/Debug.h" 66 #include "llvm/Support/FormatVariadic.h" 67 #include "llvm/Support/Regex.h" 68 #include "llvm/Target/TargetMachine.h" 69 #include "llvm/Transforms/AggressiveInstCombine/AggressiveInstCombine.h" 70 #include "llvm/Transforms/IPO/AlwaysInliner.h" 71 #include "llvm/Transforms/IPO/ArgumentPromotion.h" 72 #include "llvm/Transforms/IPO/Attributor.h" 73 #include "llvm/Transforms/IPO/CalledValuePropagation.h" 74 #include "llvm/Transforms/IPO/ConstantMerge.h" 75 #include "llvm/Transforms/IPO/CrossDSOCFI.h" 76 #include "llvm/Transforms/IPO/DeadArgumentElimination.h" 77 #include "llvm/Transforms/IPO/ElimAvailExtern.h" 78 #include "llvm/Transforms/IPO/ForceFunctionAttrs.h" 79 #include "llvm/Transforms/IPO/FunctionAttrs.h" 80 #include "llvm/Transforms/IPO/FunctionImport.h" 81 #include "llvm/Transforms/IPO/GlobalDCE.h" 82 #include "llvm/Transforms/IPO/GlobalOpt.h" 83 #include "llvm/Transforms/IPO/GlobalSplit.h" 84 #include "llvm/Transforms/IPO/HotColdSplitting.h" 85 #include "llvm/Transforms/IPO/InferFunctionAttrs.h" 86 #include "llvm/Transforms/IPO/Inliner.h" 87 #include "llvm/Transforms/IPO/Internalize.h" 88 #include "llvm/Transforms/IPO/LowerTypeTests.h" 89 #include "llvm/Transforms/IPO/MergeFunctions.h" 90 #include "llvm/Transforms/IPO/PartialInlining.h" 91 #include "llvm/Transforms/IPO/SCCP.h" 92 #include "llvm/Transforms/IPO/SampleProfile.h" 93 #include "llvm/Transforms/IPO/StripDeadPrototypes.h" 94 #include "llvm/Transforms/IPO/SyntheticCountsPropagation.h" 95 #include "llvm/Transforms/IPO/WholeProgramDevirt.h" 96 #include "llvm/Transforms/InstCombine/InstCombine.h" 97 #include "llvm/Transforms/Instrumentation.h" 98 #include "llvm/Transforms/Instrumentation/AddressSanitizer.h" 99 #include "llvm/Transforms/Instrumentation/BoundsChecking.h" 100 #include "llvm/Transforms/Instrumentation/CGProfile.h" 101 #include "llvm/Transforms/Instrumentation/ControlHeightReduction.h" 102 #include "llvm/Transforms/Instrumentation/GCOVProfiler.h" 103 #include "llvm/Transforms/Instrumentation/HWAddressSanitizer.h" 104 #include "llvm/Transforms/Instrumentation/InstrOrderFile.h" 105 #include "llvm/Transforms/Instrumentation/InstrProfiling.h" 106 #include "llvm/Transforms/Instrumentation/MemorySanitizer.h" 107 #include "llvm/Transforms/Instrumentation/PGOInstrumentation.h" 108 #include "llvm/Transforms/Instrumentation/PoisonChecking.h" 109 #include "llvm/Transforms/Instrumentation/SanitizerCoverage.h" 110 #include "llvm/Transforms/Instrumentation/ThreadSanitizer.h" 111 #include "llvm/Transforms/Scalar/ADCE.h" 112 #include "llvm/Transforms/Scalar/AlignmentFromAssumptions.h" 113 #include "llvm/Transforms/Scalar/BDCE.h" 114 #include "llvm/Transforms/Scalar/CallSiteSplitting.h" 115 #include "llvm/Transforms/Scalar/ConstantHoisting.h" 116 #include "llvm/Transforms/Scalar/CorrelatedValuePropagation.h" 117 #include "llvm/Transforms/Scalar/DCE.h" 118 #include "llvm/Transforms/Scalar/DeadStoreElimination.h" 119 #include "llvm/Transforms/Scalar/DivRemPairs.h" 120 #include "llvm/Transforms/Scalar/EarlyCSE.h" 121 #include "llvm/Transforms/Scalar/Float2Int.h" 122 #include "llvm/Transforms/Scalar/GVN.h" 123 #include "llvm/Transforms/Scalar/GuardWidening.h" 124 #include "llvm/Transforms/Scalar/IVUsersPrinter.h" 125 #include "llvm/Transforms/Scalar/IndVarSimplify.h" 126 #include "llvm/Transforms/Scalar/InductiveRangeCheckElimination.h" 127 #include "llvm/Transforms/Scalar/InstSimplifyPass.h" 128 #include "llvm/Transforms/Scalar/JumpThreading.h" 129 #include "llvm/Transforms/Scalar/LICM.h" 130 #include "llvm/Transforms/Scalar/LoopAccessAnalysisPrinter.h" 131 #include "llvm/Transforms/Scalar/LoopDataPrefetch.h" 132 #include "llvm/Transforms/Scalar/LoopDeletion.h" 133 #include "llvm/Transforms/Scalar/LoopDistribute.h" 134 #include "llvm/Transforms/Scalar/LoopFuse.h" 135 #include "llvm/Transforms/Scalar/LoopIdiomRecognize.h" 136 #include "llvm/Transforms/Scalar/LoopInstSimplify.h" 137 #include "llvm/Transforms/Scalar/LoopLoadElimination.h" 138 #include "llvm/Transforms/Scalar/LoopPassManager.h" 139 #include "llvm/Transforms/Scalar/LoopPredication.h" 140 #include "llvm/Transforms/Scalar/LoopRotation.h" 141 #include "llvm/Transforms/Scalar/LoopSimplifyCFG.h" 142 #include "llvm/Transforms/Scalar/LoopSink.h" 143 #include "llvm/Transforms/Scalar/LoopStrengthReduce.h" 144 #include "llvm/Transforms/Scalar/LoopUnrollAndJamPass.h" 145 #include "llvm/Transforms/Scalar/LoopUnrollPass.h" 146 #include "llvm/Transforms/Scalar/LowerAtomic.h" 147 #include "llvm/Transforms/Scalar/LowerConstantIntrinsics.h" 148 #include "llvm/Transforms/Scalar/LowerExpectIntrinsic.h" 149 #include "llvm/Transforms/Scalar/LowerGuardIntrinsic.h" 150 #include "llvm/Transforms/Scalar/LowerMatrixIntrinsics.h" 151 #include "llvm/Transforms/Scalar/LowerWidenableCondition.h" 152 #include "llvm/Transforms/Scalar/MakeGuardsExplicit.h" 153 #include "llvm/Transforms/Scalar/MemCpyOptimizer.h" 154 #include "llvm/Transforms/Scalar/MergeICmps.h" 155 #include "llvm/Transforms/Scalar/MergedLoadStoreMotion.h" 156 #include "llvm/Transforms/Scalar/NaryReassociate.h" 157 #include "llvm/Transforms/Scalar/NewGVN.h" 158 #include "llvm/Transforms/Scalar/PartiallyInlineLibCalls.h" 159 #include "llvm/Transforms/Scalar/Reassociate.h" 160 #include "llvm/Transforms/Scalar/RewriteStatepointsForGC.h" 161 #include "llvm/Transforms/Scalar/SCCP.h" 162 #include "llvm/Transforms/Scalar/SROA.h" 163 #include "llvm/Transforms/Scalar/Scalarizer.h" 164 #include "llvm/Transforms/Scalar/SimpleLoopUnswitch.h" 165 #include "llvm/Transforms/Scalar/SimplifyCFG.h" 166 #include "llvm/Transforms/Scalar/Sink.h" 167 #include "llvm/Transforms/Scalar/SpeculateAroundPHIs.h" 168 #include "llvm/Transforms/Scalar/SpeculativeExecution.h" 169 #include "llvm/Transforms/Scalar/TailRecursionElimination.h" 170 #include "llvm/Transforms/Scalar/WarnMissedTransforms.h" 171 #include "llvm/Transforms/Utils/AddDiscriminators.h" 172 #include "llvm/Transforms/Utils/BreakCriticalEdges.h" 173 #include "llvm/Transforms/Utils/CanonicalizeAliases.h" 174 #include "llvm/Transforms/Utils/EntryExitInstrumenter.h" 175 #include "llvm/Transforms/Utils/InjectTLIMappings.h" 176 #include "llvm/Transforms/Utils/LCSSA.h" 177 #include "llvm/Transforms/Utils/LibCallsShrinkWrap.h" 178 #include "llvm/Transforms/Utils/LoopSimplify.h" 179 #include "llvm/Transforms/Utils/LowerInvoke.h" 180 #include "llvm/Transforms/Utils/Mem2Reg.h" 181 #include "llvm/Transforms/Utils/NameAnonGlobals.h" 182 #include "llvm/Transforms/Utils/SymbolRewriter.h" 183 #include "llvm/Transforms/Vectorize/LoadStoreVectorizer.h" 184 #include "llvm/Transforms/Vectorize/LoopVectorize.h" 185 #include "llvm/Transforms/Vectorize/SLPVectorizer.h" 186 187 using namespace llvm; 188 189 static cl::opt<unsigned> MaxDevirtIterations("pm-max-devirt-iterations", 190 cl::ReallyHidden, cl::init(4)); 191 static cl::opt<bool> 192 RunPartialInlining("enable-npm-partial-inlining", cl::init(false), 193 cl::Hidden, cl::ZeroOrMore, 194 cl::desc("Run Partial inlinining pass")); 195 196 static cl::opt<int> PreInlineThreshold( 197 "npm-preinline-threshold", cl::Hidden, cl::init(75), cl::ZeroOrMore, 198 cl::desc("Control the amount of inlining in pre-instrumentation inliner " 199 "(default = 75)")); 200 201 static cl::opt<bool> 202 RunNewGVN("enable-npm-newgvn", cl::init(false), 203 cl::Hidden, cl::ZeroOrMore, 204 cl::desc("Run NewGVN instead of GVN")); 205 206 static cl::opt<bool> EnableGVNHoist( 207 "enable-npm-gvn-hoist", cl::init(false), cl::Hidden, 208 cl::desc("Enable the GVN hoisting pass for the new PM (default = off)")); 209 210 static cl::opt<bool> EnableGVNSink( 211 "enable-npm-gvn-sink", cl::init(false), cl::Hidden, 212 cl::desc("Enable the GVN hoisting pass for the new PM (default = off)")); 213 214 static cl::opt<bool> EnableUnrollAndJam( 215 "enable-npm-unroll-and-jam", cl::init(false), cl::Hidden, 216 cl::desc("Enable the Unroll and Jam pass for the new PM (default = off)")); 217 218 static cl::opt<bool> EnableSyntheticCounts( 219 "enable-npm-synthetic-counts", cl::init(false), cl::Hidden, cl::ZeroOrMore, 220 cl::desc("Run synthetic function entry count generation " 221 "pass")); 222 223 static const Regex DefaultAliasRegex( 224 "^(default|thinlto-pre-link|thinlto|lto-pre-link|lto)<(O[0123sz])>$"); 225 226 // This option is used in simplifying testing SampleFDO optimizations for 227 // profile loading. 228 static cl::opt<bool> 229 EnableCHR("enable-chr-npm", cl::init(true), cl::Hidden, 230 cl::desc("Enable control height reduction optimization (CHR)")); 231 232 PipelineTuningOptions::PipelineTuningOptions() { 233 LoopInterleaving = EnableLoopInterleaving; 234 LoopVectorization = EnableLoopVectorization; 235 SLPVectorization = RunSLPVectorization; 236 LoopUnrolling = true; 237 ForgetAllSCEVInLoopUnroll = ForgetSCEVInLoopUnroll; 238 LicmMssaOptCap = SetLicmMssaOptCap; 239 LicmMssaNoAccForPromotionCap = SetLicmMssaNoAccForPromotionCap; 240 } 241 242 extern cl::opt<bool> EnableHotColdSplit; 243 extern cl::opt<bool> EnableOrderFileInstrumentation; 244 245 extern cl::opt<bool> FlattenedProfileUsed; 246 247 static bool isOptimizingForSize(PassBuilder::OptimizationLevel Level) { 248 switch (Level) { 249 case PassBuilder::O0: 250 case PassBuilder::O1: 251 case PassBuilder::O2: 252 case PassBuilder::O3: 253 return false; 254 255 case PassBuilder::Os: 256 case PassBuilder::Oz: 257 return true; 258 } 259 llvm_unreachable("Invalid optimization level!"); 260 } 261 262 namespace { 263 264 /// No-op module pass which does nothing. 265 struct NoOpModulePass { 266 PreservedAnalyses run(Module &M, ModuleAnalysisManager &) { 267 return PreservedAnalyses::all(); 268 } 269 static StringRef name() { return "NoOpModulePass"; } 270 }; 271 272 /// No-op module analysis. 273 class NoOpModuleAnalysis : public AnalysisInfoMixin<NoOpModuleAnalysis> { 274 friend AnalysisInfoMixin<NoOpModuleAnalysis>; 275 static AnalysisKey Key; 276 277 public: 278 struct Result {}; 279 Result run(Module &, ModuleAnalysisManager &) { return Result(); } 280 static StringRef name() { return "NoOpModuleAnalysis"; } 281 }; 282 283 /// No-op CGSCC pass which does nothing. 284 struct NoOpCGSCCPass { 285 PreservedAnalyses run(LazyCallGraph::SCC &C, CGSCCAnalysisManager &, 286 LazyCallGraph &, CGSCCUpdateResult &UR) { 287 return PreservedAnalyses::all(); 288 } 289 static StringRef name() { return "NoOpCGSCCPass"; } 290 }; 291 292 /// No-op CGSCC analysis. 293 class NoOpCGSCCAnalysis : public AnalysisInfoMixin<NoOpCGSCCAnalysis> { 294 friend AnalysisInfoMixin<NoOpCGSCCAnalysis>; 295 static AnalysisKey Key; 296 297 public: 298 struct Result {}; 299 Result run(LazyCallGraph::SCC &, CGSCCAnalysisManager &, LazyCallGraph &G) { 300 return Result(); 301 } 302 static StringRef name() { return "NoOpCGSCCAnalysis"; } 303 }; 304 305 /// No-op function pass which does nothing. 306 struct NoOpFunctionPass { 307 PreservedAnalyses run(Function &F, FunctionAnalysisManager &) { 308 return PreservedAnalyses::all(); 309 } 310 static StringRef name() { return "NoOpFunctionPass"; } 311 }; 312 313 /// No-op function analysis. 314 class NoOpFunctionAnalysis : public AnalysisInfoMixin<NoOpFunctionAnalysis> { 315 friend AnalysisInfoMixin<NoOpFunctionAnalysis>; 316 static AnalysisKey Key; 317 318 public: 319 struct Result {}; 320 Result run(Function &, FunctionAnalysisManager &) { return Result(); } 321 static StringRef name() { return "NoOpFunctionAnalysis"; } 322 }; 323 324 /// No-op loop pass which does nothing. 325 struct NoOpLoopPass { 326 PreservedAnalyses run(Loop &L, LoopAnalysisManager &, 327 LoopStandardAnalysisResults &, LPMUpdater &) { 328 return PreservedAnalyses::all(); 329 } 330 static StringRef name() { return "NoOpLoopPass"; } 331 }; 332 333 /// No-op loop analysis. 334 class NoOpLoopAnalysis : public AnalysisInfoMixin<NoOpLoopAnalysis> { 335 friend AnalysisInfoMixin<NoOpLoopAnalysis>; 336 static AnalysisKey Key; 337 338 public: 339 struct Result {}; 340 Result run(Loop &, LoopAnalysisManager &, LoopStandardAnalysisResults &) { 341 return Result(); 342 } 343 static StringRef name() { return "NoOpLoopAnalysis"; } 344 }; 345 346 AnalysisKey NoOpModuleAnalysis::Key; 347 AnalysisKey NoOpCGSCCAnalysis::Key; 348 AnalysisKey NoOpFunctionAnalysis::Key; 349 AnalysisKey NoOpLoopAnalysis::Key; 350 351 } // End anonymous namespace. 352 353 void PassBuilder::invokePeepholeEPCallbacks( 354 FunctionPassManager &FPM, PassBuilder::OptimizationLevel Level) { 355 for (auto &C : PeepholeEPCallbacks) 356 C(FPM, Level); 357 } 358 359 void PassBuilder::registerModuleAnalyses(ModuleAnalysisManager &MAM) { 360 #define MODULE_ANALYSIS(NAME, CREATE_PASS) \ 361 MAM.registerPass([&] { return CREATE_PASS; }); 362 #include "PassRegistry.def" 363 364 for (auto &C : ModuleAnalysisRegistrationCallbacks) 365 C(MAM); 366 } 367 368 void PassBuilder::registerCGSCCAnalyses(CGSCCAnalysisManager &CGAM) { 369 #define CGSCC_ANALYSIS(NAME, CREATE_PASS) \ 370 CGAM.registerPass([&] { return CREATE_PASS; }); 371 #include "PassRegistry.def" 372 373 for (auto &C : CGSCCAnalysisRegistrationCallbacks) 374 C(CGAM); 375 } 376 377 void PassBuilder::registerFunctionAnalyses(FunctionAnalysisManager &FAM) { 378 #define FUNCTION_ANALYSIS(NAME, CREATE_PASS) \ 379 FAM.registerPass([&] { return CREATE_PASS; }); 380 #include "PassRegistry.def" 381 382 for (auto &C : FunctionAnalysisRegistrationCallbacks) 383 C(FAM); 384 } 385 386 void PassBuilder::registerLoopAnalyses(LoopAnalysisManager &LAM) { 387 #define LOOP_ANALYSIS(NAME, CREATE_PASS) \ 388 LAM.registerPass([&] { return CREATE_PASS; }); 389 #include "PassRegistry.def" 390 391 for (auto &C : LoopAnalysisRegistrationCallbacks) 392 C(LAM); 393 } 394 395 FunctionPassManager 396 PassBuilder::buildFunctionSimplificationPipeline(OptimizationLevel Level, 397 ThinLTOPhase Phase, 398 bool DebugLogging) { 399 assert(Level != O0 && "Must request optimizations!"); 400 FunctionPassManager FPM(DebugLogging); 401 402 // Form SSA out of local memory accesses after breaking apart aggregates into 403 // scalars. 404 FPM.addPass(SROA()); 405 406 // Catch trivial redundancies 407 FPM.addPass(EarlyCSEPass(true /* Enable mem-ssa. */)); 408 409 // Hoisting of scalars and load expressions. 410 if (Level > O1) { 411 if (EnableGVNHoist) 412 FPM.addPass(GVNHoistPass()); 413 414 // Global value numbering based sinking. 415 if (EnableGVNSink) { 416 FPM.addPass(GVNSinkPass()); 417 FPM.addPass(SimplifyCFGPass()); 418 } 419 } 420 421 // Speculative execution if the target has divergent branches; otherwise nop. 422 if (Level > O1) { 423 FPM.addPass(SpeculativeExecutionPass()); 424 425 // Optimize based on known information about branches, and cleanup afterward. 426 FPM.addPass(JumpThreadingPass()); 427 FPM.addPass(CorrelatedValuePropagationPass()); 428 } 429 FPM.addPass(SimplifyCFGPass()); 430 if (Level == O3) 431 FPM.addPass(AggressiveInstCombinePass()); 432 FPM.addPass(InstCombinePass()); 433 434 if (!isOptimizingForSize(Level)) 435 FPM.addPass(LibCallsShrinkWrapPass()); 436 437 invokePeepholeEPCallbacks(FPM, Level); 438 439 // For PGO use pipeline, try to optimize memory intrinsics such as memcpy 440 // using the size value profile. Don't perform this when optimizing for size. 441 if (PGOOpt && PGOOpt->Action == PGOOptions::IRUse && 442 !isOptimizingForSize(Level) && Level > O1) 443 FPM.addPass(PGOMemOPSizeOpt()); 444 445 // TODO: Investigate the cost/benefit of tail call elimination on debugging. 446 if (Level > O1) 447 FPM.addPass(TailCallElimPass()); 448 FPM.addPass(SimplifyCFGPass()); 449 450 // Form canonically associated expression trees, and simplify the trees using 451 // basic mathematical properties. For example, this will form (nearly) 452 // minimal multiplication trees. 453 FPM.addPass(ReassociatePass()); 454 455 // Add the primary loop simplification pipeline. 456 // FIXME: Currently this is split into two loop pass pipelines because we run 457 // some function passes in between them. These can and should be removed 458 // and/or replaced by scheduling the loop pass equivalents in the correct 459 // positions. But those equivalent passes aren't powerful enough yet. 460 // Specifically, `SimplifyCFGPass` and `InstCombinePass` are currently still 461 // used. We have `LoopSimplifyCFGPass` which isn't yet powerful enough yet to 462 // fully replace `SimplifyCFGPass`, and the closest to the other we have is 463 // `LoopInstSimplify`. 464 LoopPassManager LPM1(DebugLogging), LPM2(DebugLogging); 465 466 // Simplify the loop body. We do this initially to clean up after other loop 467 // passes run, either when iterating on a loop or on inner loops with 468 // implications on the outer loop. 469 LPM1.addPass(LoopInstSimplifyPass()); 470 LPM1.addPass(LoopSimplifyCFGPass()); 471 472 // Rotate Loop - disable header duplication at -Oz 473 LPM1.addPass(LoopRotatePass(Level != Oz)); 474 // TODO: Investigate promotion cap for O1. 475 LPM1.addPass(LICMPass(PTO.LicmMssaOptCap, PTO.LicmMssaNoAccForPromotionCap)); 476 LPM1.addPass(SimpleLoopUnswitchPass()); 477 LPM2.addPass(IndVarSimplifyPass()); 478 LPM2.addPass(LoopIdiomRecognizePass()); 479 480 for (auto &C : LateLoopOptimizationsEPCallbacks) 481 C(LPM2, Level); 482 483 LPM2.addPass(LoopDeletionPass()); 484 // Do not enable unrolling in PreLinkThinLTO phase during sample PGO 485 // because it changes IR to makes profile annotation in back compile 486 // inaccurate. 487 if ((Phase != ThinLTOPhase::PreLink || !PGOOpt || 488 PGOOpt->Action != PGOOptions::SampleUse) && 489 PTO.LoopUnrolling) 490 LPM2.addPass(LoopFullUnrollPass(Level, /*OnlyWhenForced=*/false, 491 PTO.ForgetAllSCEVInLoopUnroll)); 492 493 for (auto &C : LoopOptimizerEndEPCallbacks) 494 C(LPM2, Level); 495 496 // We provide the opt remark emitter pass for LICM to use. We only need to do 497 // this once as it is immutable. 498 FPM.addPass(RequireAnalysisPass<OptimizationRemarkEmitterAnalysis, Function>()); 499 FPM.addPass(createFunctionToLoopPassAdaptor( 500 std::move(LPM1), EnableMSSALoopDependency, DebugLogging)); 501 FPM.addPass(SimplifyCFGPass()); 502 FPM.addPass(InstCombinePass()); 503 // The loop passes in LPM2 (IndVarSimplifyPass, LoopIdiomRecognizePass, 504 // LoopDeletionPass and LoopFullUnrollPass) do not preserve MemorySSA. 505 // *All* loop passes must preserve it, in order to be able to use it. 506 FPM.addPass(createFunctionToLoopPassAdaptor( 507 std::move(LPM2), /*UseMemorySSA=*/false, DebugLogging)); 508 509 // Delete small array after loop unroll. 510 FPM.addPass(SROA()); 511 512 // Eliminate redundancies. 513 if (Level != O1) { 514 // These passes add substantial compile time so skip them at O1. 515 FPM.addPass(MergedLoadStoreMotionPass()); 516 if (RunNewGVN) 517 FPM.addPass(NewGVNPass()); 518 else 519 FPM.addPass(GVN()); 520 } 521 522 // Specially optimize memory movement as it doesn't look like dataflow in SSA. 523 FPM.addPass(MemCpyOptPass()); 524 525 // Sparse conditional constant propagation. 526 // FIXME: It isn't clear why we do this *after* loop passes rather than 527 // before... 528 FPM.addPass(SCCPPass()); 529 530 // Delete dead bit computations (instcombine runs after to fold away the dead 531 // computations, and then ADCE will run later to exploit any new DCE 532 // opportunities that creates). 533 FPM.addPass(BDCEPass()); 534 535 // Run instcombine after redundancy and dead bit elimination to exploit 536 // opportunities opened up by them. 537 FPM.addPass(InstCombinePass()); 538 invokePeepholeEPCallbacks(FPM, Level); 539 540 // Re-consider control flow based optimizations after redundancy elimination, 541 // redo DCE, etc. 542 if (Level > O1) { 543 FPM.addPass(JumpThreadingPass()); 544 FPM.addPass(CorrelatedValuePropagationPass()); 545 FPM.addPass(DSEPass()); 546 FPM.addPass(createFunctionToLoopPassAdaptor( 547 LICMPass(PTO.LicmMssaOptCap, PTO.LicmMssaNoAccForPromotionCap), 548 EnableMSSALoopDependency, DebugLogging)); 549 } 550 551 for (auto &C : ScalarOptimizerLateEPCallbacks) 552 C(FPM, Level); 553 554 // Finally, do an expensive DCE pass to catch all the dead code exposed by 555 // the simplifications and basic cleanup after all the simplifications. 556 // TODO: Investigate if this is too expensive. 557 FPM.addPass(ADCEPass()); 558 FPM.addPass(SimplifyCFGPass()); 559 FPM.addPass(InstCombinePass()); 560 invokePeepholeEPCallbacks(FPM, Level); 561 562 if (EnableCHR && Level == O3 && PGOOpt && 563 (PGOOpt->Action == PGOOptions::IRUse || 564 PGOOpt->Action == PGOOptions::SampleUse)) 565 FPM.addPass(ControlHeightReductionPass()); 566 567 return FPM; 568 } 569 570 void PassBuilder::addPGOInstrPasses(ModulePassManager &MPM, bool DebugLogging, 571 PassBuilder::OptimizationLevel Level, 572 bool RunProfileGen, bool IsCS, 573 std::string ProfileFile, 574 std::string ProfileRemappingFile) { 575 assert(Level != O0 && "Not expecting O0 here!"); 576 // Generally running simplification passes and the inliner with an high 577 // threshold results in smaller executables, but there may be cases where 578 // the size grows, so let's be conservative here and skip this simplification 579 // at -Os/Oz. We will not do this inline for context sensistive PGO (when 580 // IsCS is true). 581 if (!isOptimizingForSize(Level) && !IsCS) { 582 InlineParams IP; 583 584 IP.DefaultThreshold = PreInlineThreshold; 585 586 // FIXME: The hint threshold has the same value used by the regular inliner. 587 // This should probably be lowered after performance testing. 588 // FIXME: this comment is cargo culted from the old pass manager, revisit). 589 IP.HintThreshold = 325; 590 591 CGSCCPassManager CGPipeline(DebugLogging); 592 593 CGPipeline.addPass(InlinerPass(IP)); 594 595 FunctionPassManager FPM; 596 FPM.addPass(SROA()); 597 FPM.addPass(EarlyCSEPass()); // Catch trivial redundancies. 598 FPM.addPass(SimplifyCFGPass()); // Merge & remove basic blocks. 599 FPM.addPass(InstCombinePass()); // Combine silly sequences. 600 invokePeepholeEPCallbacks(FPM, Level); 601 602 CGPipeline.addPass(createCGSCCToFunctionPassAdaptor(std::move(FPM))); 603 604 MPM.addPass(createModuleToPostOrderCGSCCPassAdaptor(std::move(CGPipeline))); 605 606 // Delete anything that is now dead to make sure that we don't instrument 607 // dead code. Instrumentation can end up keeping dead code around and 608 // dramatically increase code size. 609 MPM.addPass(GlobalDCEPass()); 610 } 611 612 if (!RunProfileGen) { 613 assert(!ProfileFile.empty() && "Profile use expecting a profile file!"); 614 MPM.addPass(PGOInstrumentationUse(ProfileFile, ProfileRemappingFile, IsCS)); 615 // Cache ProfileSummaryAnalysis once to avoid the potential need to insert 616 // RequireAnalysisPass for PSI before subsequent non-module passes. 617 MPM.addPass(RequireAnalysisPass<ProfileSummaryAnalysis, Module>()); 618 return; 619 } 620 621 // Perform PGO instrumentation. 622 MPM.addPass(PGOInstrumentationGen(IsCS)); 623 624 FunctionPassManager FPM; 625 FPM.addPass(createFunctionToLoopPassAdaptor( 626 LoopRotatePass(), EnableMSSALoopDependency, DebugLogging)); 627 MPM.addPass(createModuleToFunctionPassAdaptor(std::move(FPM))); 628 629 // Add the profile lowering pass. 630 InstrProfOptions Options; 631 if (!ProfileFile.empty()) 632 Options.InstrProfileOutput = ProfileFile; 633 // Do counter promotion at Level greater than O0. 634 Options.DoCounterPromotion = true; 635 Options.UseBFIInPromotion = IsCS; 636 MPM.addPass(InstrProfiling(Options, IsCS)); 637 } 638 639 void PassBuilder::addPGOInstrPassesForO0(ModulePassManager &MPM, 640 bool DebugLogging, bool RunProfileGen, 641 bool IsCS, std::string ProfileFile, 642 std::string ProfileRemappingFile) { 643 if (!RunProfileGen) { 644 assert(!ProfileFile.empty() && "Profile use expecting a profile file!"); 645 MPM.addPass(PGOInstrumentationUse(ProfileFile, ProfileRemappingFile, IsCS)); 646 // Cache ProfileSummaryAnalysis once to avoid the potential need to insert 647 // RequireAnalysisPass for PSI before subsequent non-module passes. 648 MPM.addPass(RequireAnalysisPass<ProfileSummaryAnalysis, Module>()); 649 return; 650 } 651 652 // Perform PGO instrumentation. 653 MPM.addPass(PGOInstrumentationGen(IsCS)); 654 // Add the profile lowering pass. 655 InstrProfOptions Options; 656 if (!ProfileFile.empty()) 657 Options.InstrProfileOutput = ProfileFile; 658 // Do not do counter promotion at O0. 659 Options.DoCounterPromotion = false; 660 Options.UseBFIInPromotion = IsCS; 661 MPM.addPass(InstrProfiling(Options, IsCS)); 662 } 663 664 static InlineParams 665 getInlineParamsFromOptLevel(PassBuilder::OptimizationLevel Level) { 666 auto O3 = PassBuilder::O3; 667 unsigned OptLevel = Level > O3 ? 2 : Level; 668 unsigned SizeLevel = Level > O3 ? Level - O3 : 0; 669 return getInlineParams(OptLevel, SizeLevel); 670 } 671 672 ModulePassManager 673 PassBuilder::buildModuleSimplificationPipeline(OptimizationLevel Level, 674 ThinLTOPhase Phase, 675 bool DebugLogging) { 676 ModulePassManager MPM(DebugLogging); 677 678 bool HasSampleProfile = PGOOpt && (PGOOpt->Action == PGOOptions::SampleUse); 679 680 // In ThinLTO mode, when flattened profile is used, all the available 681 // profile information will be annotated in PreLink phase so there is 682 // no need to load the profile again in PostLink. 683 bool LoadSampleProfile = 684 HasSampleProfile && 685 !(FlattenedProfileUsed && Phase == ThinLTOPhase::PostLink); 686 687 // During the ThinLTO backend phase we perform early indirect call promotion 688 // here, before globalopt. Otherwise imported available_externally functions 689 // look unreferenced and are removed. If we are going to load the sample 690 // profile then defer until later. 691 // TODO: See if we can move later and consolidate with the location where 692 // we perform ICP when we are loading a sample profile. 693 // TODO: We pass HasSampleProfile (whether there was a sample profile file 694 // passed to the compile) to the SamplePGO flag of ICP. This is used to 695 // determine whether the new direct calls are annotated with prof metadata. 696 // Ideally this should be determined from whether the IR is annotated with 697 // sample profile, and not whether the a sample profile was provided on the 698 // command line. E.g. for flattened profiles where we will not be reloading 699 // the sample profile in the ThinLTO backend, we ideally shouldn't have to 700 // provide the sample profile file. 701 if (Phase == ThinLTOPhase::PostLink && !LoadSampleProfile) 702 MPM.addPass(PGOIndirectCallPromotion(true /* InLTO */, HasSampleProfile)); 703 704 // Do basic inference of function attributes from known properties of system 705 // libraries and other oracles. 706 MPM.addPass(InferFunctionAttrsPass()); 707 708 // Create an early function pass manager to cleanup the output of the 709 // frontend. 710 FunctionPassManager EarlyFPM(DebugLogging); 711 EarlyFPM.addPass(SimplifyCFGPass()); 712 EarlyFPM.addPass(SROA()); 713 EarlyFPM.addPass(EarlyCSEPass()); 714 EarlyFPM.addPass(LowerExpectIntrinsicPass()); 715 if (Level == O3) 716 EarlyFPM.addPass(CallSiteSplittingPass()); 717 718 // In SamplePGO ThinLTO backend, we need instcombine before profile annotation 719 // to convert bitcast to direct calls so that they can be inlined during the 720 // profile annotation prepration step. 721 // More details about SamplePGO design can be found in: 722 // https://research.google.com/pubs/pub45290.html 723 // FIXME: revisit how SampleProfileLoad/Inliner/ICP is structured. 724 if (LoadSampleProfile) 725 EarlyFPM.addPass(InstCombinePass()); 726 MPM.addPass(createModuleToFunctionPassAdaptor(std::move(EarlyFPM))); 727 728 if (LoadSampleProfile) { 729 // Annotate sample profile right after early FPM to ensure freshness of 730 // the debug info. 731 MPM.addPass(SampleProfileLoaderPass(PGOOpt->ProfileFile, 732 PGOOpt->ProfileRemappingFile, 733 Phase == ThinLTOPhase::PreLink)); 734 // Cache ProfileSummaryAnalysis once to avoid the potential need to insert 735 // RequireAnalysisPass for PSI before subsequent non-module passes. 736 MPM.addPass(RequireAnalysisPass<ProfileSummaryAnalysis, Module>()); 737 // Do not invoke ICP in the ThinLTOPrelink phase as it makes it hard 738 // for the profile annotation to be accurate in the ThinLTO backend. 739 if (Phase != ThinLTOPhase::PreLink) 740 // We perform early indirect call promotion here, before globalopt. 741 // This is important for the ThinLTO backend phase because otherwise 742 // imported available_externally functions look unreferenced and are 743 // removed. 744 MPM.addPass(PGOIndirectCallPromotion(Phase == ThinLTOPhase::PostLink, 745 true /* SamplePGO */)); 746 } 747 748 // Interprocedural constant propagation now that basic cleanup has occurred 749 // and prior to optimizing globals. 750 // FIXME: This position in the pipeline hasn't been carefully considered in 751 // years, it should be re-analyzed. 752 MPM.addPass(IPSCCPPass()); 753 754 // Attach metadata to indirect call sites indicating the set of functions 755 // they may target at run-time. This should follow IPSCCP. 756 MPM.addPass(CalledValuePropagationPass()); 757 758 // Optimize globals to try and fold them into constants. 759 MPM.addPass(GlobalOptPass()); 760 761 // Promote any localized globals to SSA registers. 762 // FIXME: Should this instead by a run of SROA? 763 // FIXME: We should probably run instcombine and simplify-cfg afterward to 764 // delete control flows that are dead once globals have been folded to 765 // constants. 766 MPM.addPass(createModuleToFunctionPassAdaptor(PromotePass())); 767 768 // Remove any dead arguments exposed by cleanups and constand folding 769 // globals. 770 MPM.addPass(DeadArgumentEliminationPass()); 771 772 // Create a small function pass pipeline to cleanup after all the global 773 // optimizations. 774 FunctionPassManager GlobalCleanupPM(DebugLogging); 775 GlobalCleanupPM.addPass(InstCombinePass()); 776 invokePeepholeEPCallbacks(GlobalCleanupPM, Level); 777 778 GlobalCleanupPM.addPass(SimplifyCFGPass()); 779 MPM.addPass(createModuleToFunctionPassAdaptor(std::move(GlobalCleanupPM))); 780 781 // Add all the requested passes for instrumentation PGO, if requested. 782 if (PGOOpt && Phase != ThinLTOPhase::PostLink && 783 (PGOOpt->Action == PGOOptions::IRInstr || 784 PGOOpt->Action == PGOOptions::IRUse)) { 785 addPGOInstrPasses(MPM, DebugLogging, Level, 786 /* RunProfileGen */ PGOOpt->Action == PGOOptions::IRInstr, 787 /* IsCS */ false, PGOOpt->ProfileFile, 788 PGOOpt->ProfileRemappingFile); 789 MPM.addPass(PGOIndirectCallPromotion(false, false)); 790 } 791 if (PGOOpt && Phase != ThinLTOPhase::PostLink && 792 PGOOpt->CSAction == PGOOptions::CSIRInstr) 793 MPM.addPass(PGOInstrumentationGenCreateVar(PGOOpt->CSProfileGenFile)); 794 795 // Synthesize function entry counts for non-PGO compilation. 796 if (EnableSyntheticCounts && !PGOOpt) 797 MPM.addPass(SyntheticCountsPropagation()); 798 799 // Require the GlobalsAA analysis for the module so we can query it within 800 // the CGSCC pipeline. 801 MPM.addPass(RequireAnalysisPass<GlobalsAA, Module>()); 802 803 // Require the ProfileSummaryAnalysis for the module so we can query it within 804 // the inliner pass. 805 MPM.addPass(RequireAnalysisPass<ProfileSummaryAnalysis, Module>()); 806 807 // Now begin the main postorder CGSCC pipeline. 808 // FIXME: The current CGSCC pipeline has its origins in the legacy pass 809 // manager and trying to emulate its precise behavior. Much of this doesn't 810 // make a lot of sense and we should revisit the core CGSCC structure. 811 CGSCCPassManager MainCGPipeline(DebugLogging); 812 813 // Note: historically, the PruneEH pass was run first to deduce nounwind and 814 // generally clean up exception handling overhead. It isn't clear this is 815 // valuable as the inliner doesn't currently care whether it is inlining an 816 // invoke or a call. 817 818 // Run the inliner first. The theory is that we are walking bottom-up and so 819 // the callees have already been fully optimized, and we want to inline them 820 // into the callers so that our optimizations can reflect that. 821 // For PreLinkThinLTO pass, we disable hot-caller heuristic for sample PGO 822 // because it makes profile annotation in the backend inaccurate. 823 InlineParams IP = getInlineParamsFromOptLevel(Level); 824 if (Phase == ThinLTOPhase::PreLink && PGOOpt && 825 PGOOpt->Action == PGOOptions::SampleUse) 826 IP.HotCallSiteThreshold = 0; 827 MainCGPipeline.addPass(InlinerPass(IP)); 828 829 // Now deduce any function attributes based in the current code. 830 MainCGPipeline.addPass(PostOrderFunctionAttrsPass()); 831 832 // When at O3 add argument promotion to the pass pipeline. 833 // FIXME: It isn't at all clear why this should be limited to O3. 834 if (Level == O3) 835 MainCGPipeline.addPass(ArgumentPromotionPass()); 836 837 // Lastly, add the core function simplification pipeline nested inside the 838 // CGSCC walk. 839 MainCGPipeline.addPass(createCGSCCToFunctionPassAdaptor( 840 buildFunctionSimplificationPipeline(Level, Phase, DebugLogging))); 841 842 for (auto &C : CGSCCOptimizerLateEPCallbacks) 843 C(MainCGPipeline, Level); 844 845 // We wrap the CGSCC pipeline in a devirtualization repeater. This will try 846 // to detect when we devirtualize indirect calls and iterate the SCC passes 847 // in that case to try and catch knock-on inlining or function attrs 848 // opportunities. Then we add it to the module pipeline by walking the SCCs 849 // in postorder (or bottom-up). 850 MPM.addPass( 851 createModuleToPostOrderCGSCCPassAdaptor(createDevirtSCCRepeatedPass( 852 std::move(MainCGPipeline), MaxDevirtIterations))); 853 854 return MPM; 855 } 856 857 ModulePassManager PassBuilder::buildModuleOptimizationPipeline( 858 OptimizationLevel Level, bool DebugLogging, bool LTOPreLink) { 859 ModulePassManager MPM(DebugLogging); 860 861 // Optimize globals now that the module is fully simplified. 862 MPM.addPass(GlobalOptPass()); 863 MPM.addPass(GlobalDCEPass()); 864 865 // Run partial inlining pass to partially inline functions that have 866 // large bodies. 867 if (RunPartialInlining) 868 MPM.addPass(PartialInlinerPass()); 869 870 // Remove avail extern fns and globals definitions since we aren't compiling 871 // an object file for later LTO. For LTO we want to preserve these so they 872 // are eligible for inlining at link-time. Note if they are unreferenced they 873 // will be removed by GlobalDCE later, so this only impacts referenced 874 // available externally globals. Eventually they will be suppressed during 875 // codegen, but eliminating here enables more opportunity for GlobalDCE as it 876 // may make globals referenced by available external functions dead and saves 877 // running remaining passes on the eliminated functions. These should be 878 // preserved during prelinking for link-time inlining decisions. 879 if (!LTOPreLink) 880 MPM.addPass(EliminateAvailableExternallyPass()); 881 882 if (EnableOrderFileInstrumentation) 883 MPM.addPass(InstrOrderFilePass()); 884 885 // Do RPO function attribute inference across the module to forward-propagate 886 // attributes where applicable. 887 // FIXME: Is this really an optimization rather than a canonicalization? 888 MPM.addPass(ReversePostOrderFunctionAttrsPass()); 889 890 // Do a post inline PGO instrumentation and use pass. This is a context 891 // sensitive PGO pass. We don't want to do this in LTOPreLink phrase as 892 // cross-module inline has not been done yet. The context sensitive 893 // instrumentation is after all the inlines are done. 894 if (!LTOPreLink && PGOOpt) { 895 if (PGOOpt->CSAction == PGOOptions::CSIRInstr) 896 addPGOInstrPasses(MPM, DebugLogging, Level, /* RunProfileGen */ true, 897 /* IsCS */ true, PGOOpt->CSProfileGenFile, 898 PGOOpt->ProfileRemappingFile); 899 else if (PGOOpt->CSAction == PGOOptions::CSIRUse) 900 addPGOInstrPasses(MPM, DebugLogging, Level, /* RunProfileGen */ false, 901 /* IsCS */ true, PGOOpt->ProfileFile, 902 PGOOpt->ProfileRemappingFile); 903 } 904 905 // Re-require GloblasAA here prior to function passes. This is particularly 906 // useful as the above will have inlined, DCE'ed, and function-attr 907 // propagated everything. We should at this point have a reasonably minimal 908 // and richly annotated call graph. By computing aliasing and mod/ref 909 // information for all local globals here, the late loop passes and notably 910 // the vectorizer will be able to use them to help recognize vectorizable 911 // memory operations. 912 MPM.addPass(RequireAnalysisPass<GlobalsAA, Module>()); 913 914 FunctionPassManager OptimizePM(DebugLogging); 915 OptimizePM.addPass(Float2IntPass()); 916 OptimizePM.addPass(LowerConstantIntrinsicsPass()); 917 918 // FIXME: We need to run some loop optimizations to re-rotate loops after 919 // simplify-cfg and others undo their rotation. 920 921 // Optimize the loop execution. These passes operate on entire loop nests 922 // rather than on each loop in an inside-out manner, and so they are actually 923 // function passes. 924 925 for (auto &C : VectorizerStartEPCallbacks) 926 C(OptimizePM, Level); 927 928 // First rotate loops that may have been un-rotated by prior passes. 929 OptimizePM.addPass(createFunctionToLoopPassAdaptor( 930 LoopRotatePass(), EnableMSSALoopDependency, DebugLogging)); 931 932 // Distribute loops to allow partial vectorization. I.e. isolate dependences 933 // into separate loop that would otherwise inhibit vectorization. This is 934 // currently only performed for loops marked with the metadata 935 // llvm.loop.distribute=true or when -enable-loop-distribute is specified. 936 OptimizePM.addPass(LoopDistributePass()); 937 938 // Now run the core loop vectorizer. 939 OptimizePM.addPass(LoopVectorizePass( 940 LoopVectorizeOptions(!PTO.LoopInterleaving, !PTO.LoopVectorization))); 941 942 // Eliminate loads by forwarding stores from the previous iteration to loads 943 // of the current iteration. 944 OptimizePM.addPass(LoopLoadEliminationPass()); 945 946 // Cleanup after the loop optimization passes. 947 OptimizePM.addPass(InstCombinePass()); 948 949 // Now that we've formed fast to execute loop structures, we do further 950 // optimizations. These are run afterward as they might block doing complex 951 // analyses and transforms such as what are needed for loop vectorization. 952 953 // Cleanup after loop vectorization, etc. Simplification passes like CVP and 954 // GVN, loop transforms, and others have already run, so it's now better to 955 // convert to more optimized IR using more aggressive simplify CFG options. 956 // The extra sinking transform can create larger basic blocks, so do this 957 // before SLP vectorization. 958 OptimizePM.addPass(SimplifyCFGPass(SimplifyCFGOptions(). 959 forwardSwitchCondToPhi(true). 960 convertSwitchToLookupTable(true). 961 needCanonicalLoops(false). 962 sinkCommonInsts(true))); 963 964 // Optimize parallel scalar instruction chains into SIMD instructions. 965 if (PTO.SLPVectorization) 966 OptimizePM.addPass(SLPVectorizerPass()); 967 968 OptimizePM.addPass(InstCombinePass()); 969 970 // Unroll small loops to hide loop backedge latency and saturate any parallel 971 // execution resources of an out-of-order processor. We also then need to 972 // clean up redundancies and loop invariant code. 973 // FIXME: It would be really good to use a loop-integrated instruction 974 // combiner for cleanup here so that the unrolling and LICM can be pipelined 975 // across the loop nests. 976 // We do UnrollAndJam in a separate LPM to ensure it happens before unroll 977 if (EnableUnrollAndJam && PTO.LoopUnrolling) { 978 OptimizePM.addPass(LoopUnrollAndJamPass(Level)); 979 } 980 OptimizePM.addPass(LoopUnrollPass( 981 LoopUnrollOptions(Level, /*OnlyWhenForced=*/!PTO.LoopUnrolling, 982 PTO.ForgetAllSCEVInLoopUnroll))); 983 OptimizePM.addPass(WarnMissedTransformationsPass()); 984 OptimizePM.addPass(InstCombinePass()); 985 OptimizePM.addPass(RequireAnalysisPass<OptimizationRemarkEmitterAnalysis, Function>()); 986 OptimizePM.addPass(createFunctionToLoopPassAdaptor( 987 LICMPass(PTO.LicmMssaOptCap, PTO.LicmMssaNoAccForPromotionCap), 988 EnableMSSALoopDependency, DebugLogging)); 989 990 // Now that we've vectorized and unrolled loops, we may have more refined 991 // alignment information, try to re-derive it here. 992 OptimizePM.addPass(AlignmentFromAssumptionsPass()); 993 994 // Split out cold code. Splitting is done late to avoid hiding context from 995 // other optimizations and inadvertently regressing performance. The tradeoff 996 // is that this has a higher code size cost than splitting early. 997 if (EnableHotColdSplit && !LTOPreLink) 998 MPM.addPass(HotColdSplittingPass()); 999 1000 // LoopSink pass sinks instructions hoisted by LICM, which serves as a 1001 // canonicalization pass that enables other optimizations. As a result, 1002 // LoopSink pass needs to be a very late IR pass to avoid undoing LICM 1003 // result too early. 1004 OptimizePM.addPass(LoopSinkPass()); 1005 1006 // And finally clean up LCSSA form before generating code. 1007 OptimizePM.addPass(InstSimplifyPass()); 1008 1009 // This hoists/decomposes div/rem ops. It should run after other sink/hoist 1010 // passes to avoid re-sinking, but before SimplifyCFG because it can allow 1011 // flattening of blocks. 1012 OptimizePM.addPass(DivRemPairsPass()); 1013 1014 // LoopSink (and other loop passes since the last simplifyCFG) might have 1015 // resulted in single-entry-single-exit or empty blocks. Clean up the CFG. 1016 OptimizePM.addPass(SimplifyCFGPass()); 1017 1018 // Optimize PHIs by speculating around them when profitable. Note that this 1019 // pass needs to be run after any PRE or similar pass as it is essentially 1020 // inserting redundancies into the program. This even includes SimplifyCFG. 1021 OptimizePM.addPass(SpeculateAroundPHIsPass()); 1022 1023 for (auto &C : OptimizerLastEPCallbacks) 1024 C(OptimizePM, Level); 1025 1026 // Add the core optimizing pipeline. 1027 MPM.addPass(createModuleToFunctionPassAdaptor(std::move(OptimizePM))); 1028 1029 MPM.addPass(CGProfilePass()); 1030 1031 // Now we need to do some global optimization transforms. 1032 // FIXME: It would seem like these should come first in the optimization 1033 // pipeline and maybe be the bottom of the canonicalization pipeline? Weird 1034 // ordering here. 1035 MPM.addPass(GlobalDCEPass()); 1036 MPM.addPass(ConstantMergePass()); 1037 1038 return MPM; 1039 } 1040 1041 ModulePassManager 1042 PassBuilder::buildPerModuleDefaultPipeline(OptimizationLevel Level, 1043 bool DebugLogging, bool LTOPreLink) { 1044 assert(Level != O0 && "Must request optimizations for the default pipeline!"); 1045 1046 ModulePassManager MPM(DebugLogging); 1047 1048 // Force any function attributes we want the rest of the pipeline to observe. 1049 MPM.addPass(ForceFunctionAttrsPass()); 1050 1051 // Apply module pipeline start EP callback. 1052 for (auto &C : PipelineStartEPCallbacks) 1053 C(MPM); 1054 1055 if (PGOOpt && PGOOpt->SamplePGOSupport) 1056 MPM.addPass(createModuleToFunctionPassAdaptor(AddDiscriminatorsPass())); 1057 1058 // Add the core simplification pipeline. 1059 MPM.addPass(buildModuleSimplificationPipeline(Level, ThinLTOPhase::None, 1060 DebugLogging)); 1061 1062 // Now add the optimization pipeline. 1063 MPM.addPass(buildModuleOptimizationPipeline(Level, DebugLogging, LTOPreLink)); 1064 1065 return MPM; 1066 } 1067 1068 ModulePassManager 1069 PassBuilder::buildThinLTOPreLinkDefaultPipeline(OptimizationLevel Level, 1070 bool DebugLogging) { 1071 assert(Level != O0 && "Must request optimizations for the default pipeline!"); 1072 1073 ModulePassManager MPM(DebugLogging); 1074 1075 // Force any function attributes we want the rest of the pipeline to observe. 1076 MPM.addPass(ForceFunctionAttrsPass()); 1077 1078 if (PGOOpt && PGOOpt->SamplePGOSupport) 1079 MPM.addPass(createModuleToFunctionPassAdaptor(AddDiscriminatorsPass())); 1080 1081 // Apply module pipeline start EP callback. 1082 for (auto &C : PipelineStartEPCallbacks) 1083 C(MPM); 1084 1085 // If we are planning to perform ThinLTO later, we don't bloat the code with 1086 // unrolling/vectorization/... now. Just simplify the module as much as we 1087 // can. 1088 MPM.addPass(buildModuleSimplificationPipeline(Level, ThinLTOPhase::PreLink, 1089 DebugLogging)); 1090 1091 // Run partial inlining pass to partially inline functions that have 1092 // large bodies. 1093 // FIXME: It isn't clear whether this is really the right place to run this 1094 // in ThinLTO. Because there is another canonicalization and simplification 1095 // phase that will run after the thin link, running this here ends up with 1096 // less information than will be available later and it may grow functions in 1097 // ways that aren't beneficial. 1098 if (RunPartialInlining) 1099 MPM.addPass(PartialInlinerPass()); 1100 1101 // Reduce the size of the IR as much as possible. 1102 MPM.addPass(GlobalOptPass()); 1103 1104 return MPM; 1105 } 1106 1107 ModulePassManager PassBuilder::buildThinLTODefaultPipeline( 1108 OptimizationLevel Level, bool DebugLogging, 1109 const ModuleSummaryIndex *ImportSummary) { 1110 ModulePassManager MPM(DebugLogging); 1111 1112 if (ImportSummary) { 1113 // These passes import type identifier resolutions for whole-program 1114 // devirtualization and CFI. They must run early because other passes may 1115 // disturb the specific instruction patterns that these passes look for, 1116 // creating dependencies on resolutions that may not appear in the summary. 1117 // 1118 // For example, GVN may transform the pattern assume(type.test) appearing in 1119 // two basic blocks into assume(phi(type.test, type.test)), which would 1120 // transform a dependency on a WPD resolution into a dependency on a type 1121 // identifier resolution for CFI. 1122 // 1123 // Also, WPD has access to more precise information than ICP and can 1124 // devirtualize more effectively, so it should operate on the IR first. 1125 // 1126 // The WPD and LowerTypeTest passes need to run at -O0 to lower type 1127 // metadata and intrinsics. 1128 MPM.addPass(WholeProgramDevirtPass(nullptr, ImportSummary)); 1129 MPM.addPass(LowerTypeTestsPass(nullptr, ImportSummary)); 1130 } 1131 1132 if (Level == O0) 1133 return MPM; 1134 1135 // Force any function attributes we want the rest of the pipeline to observe. 1136 MPM.addPass(ForceFunctionAttrsPass()); 1137 1138 // Add the core simplification pipeline. 1139 MPM.addPass(buildModuleSimplificationPipeline(Level, ThinLTOPhase::PostLink, 1140 DebugLogging)); 1141 1142 // Now add the optimization pipeline. 1143 MPM.addPass(buildModuleOptimizationPipeline(Level, DebugLogging)); 1144 1145 return MPM; 1146 } 1147 1148 ModulePassManager 1149 PassBuilder::buildLTOPreLinkDefaultPipeline(OptimizationLevel Level, 1150 bool DebugLogging) { 1151 assert(Level != O0 && "Must request optimizations for the default pipeline!"); 1152 // FIXME: We should use a customized pre-link pipeline! 1153 return buildPerModuleDefaultPipeline(Level, DebugLogging, 1154 /* LTOPreLink */true); 1155 } 1156 1157 ModulePassManager 1158 PassBuilder::buildLTODefaultPipeline(OptimizationLevel Level, bool DebugLogging, 1159 ModuleSummaryIndex *ExportSummary) { 1160 ModulePassManager MPM(DebugLogging); 1161 1162 if (Level == O0) { 1163 // The WPD and LowerTypeTest passes need to run at -O0 to lower type 1164 // metadata and intrinsics. 1165 MPM.addPass(WholeProgramDevirtPass(ExportSummary, nullptr)); 1166 MPM.addPass(LowerTypeTestsPass(ExportSummary, nullptr)); 1167 return MPM; 1168 } 1169 1170 if (PGOOpt && PGOOpt->Action == PGOOptions::SampleUse) { 1171 // Load sample profile before running the LTO optimization pipeline. 1172 MPM.addPass(SampleProfileLoaderPass(PGOOpt->ProfileFile, 1173 PGOOpt->ProfileRemappingFile, 1174 false /* ThinLTOPhase::PreLink */)); 1175 // Cache ProfileSummaryAnalysis once to avoid the potential need to insert 1176 // RequireAnalysisPass for PSI before subsequent non-module passes. 1177 MPM.addPass(RequireAnalysisPass<ProfileSummaryAnalysis, Module>()); 1178 } 1179 1180 // Remove unused virtual tables to improve the quality of code generated by 1181 // whole-program devirtualization and bitset lowering. 1182 MPM.addPass(GlobalDCEPass()); 1183 1184 // Force any function attributes we want the rest of the pipeline to observe. 1185 MPM.addPass(ForceFunctionAttrsPass()); 1186 1187 // Do basic inference of function attributes from known properties of system 1188 // libraries and other oracles. 1189 MPM.addPass(InferFunctionAttrsPass()); 1190 1191 if (Level > 1) { 1192 FunctionPassManager EarlyFPM(DebugLogging); 1193 EarlyFPM.addPass(CallSiteSplittingPass()); 1194 MPM.addPass(createModuleToFunctionPassAdaptor(std::move(EarlyFPM))); 1195 1196 // Indirect call promotion. This should promote all the targets that are 1197 // left by the earlier promotion pass that promotes intra-module targets. 1198 // This two-step promotion is to save the compile time. For LTO, it should 1199 // produce the same result as if we only do promotion here. 1200 MPM.addPass(PGOIndirectCallPromotion( 1201 true /* InLTO */, PGOOpt && PGOOpt->Action == PGOOptions::SampleUse)); 1202 // Propagate constants at call sites into the functions they call. This 1203 // opens opportunities for globalopt (and inlining) by substituting function 1204 // pointers passed as arguments to direct uses of functions. 1205 MPM.addPass(IPSCCPPass()); 1206 1207 // Attach metadata to indirect call sites indicating the set of functions 1208 // they may target at run-time. This should follow IPSCCP. 1209 MPM.addPass(CalledValuePropagationPass()); 1210 } 1211 1212 // Now deduce any function attributes based in the current code. 1213 MPM.addPass(createModuleToPostOrderCGSCCPassAdaptor( 1214 PostOrderFunctionAttrsPass())); 1215 1216 // Do RPO function attribute inference across the module to forward-propagate 1217 // attributes where applicable. 1218 // FIXME: Is this really an optimization rather than a canonicalization? 1219 MPM.addPass(ReversePostOrderFunctionAttrsPass()); 1220 1221 // Use in-range annotations on GEP indices to split globals where beneficial. 1222 MPM.addPass(GlobalSplitPass()); 1223 1224 // Run whole program optimization of virtual call when the list of callees 1225 // is fixed. 1226 MPM.addPass(WholeProgramDevirtPass(ExportSummary, nullptr)); 1227 1228 // Stop here at -O1. 1229 if (Level == 1) { 1230 // The LowerTypeTestsPass needs to run to lower type metadata and the 1231 // type.test intrinsics. The pass does nothing if CFI is disabled. 1232 MPM.addPass(LowerTypeTestsPass(ExportSummary, nullptr)); 1233 return MPM; 1234 } 1235 1236 // Optimize globals to try and fold them into constants. 1237 MPM.addPass(GlobalOptPass()); 1238 1239 // Promote any localized globals to SSA registers. 1240 MPM.addPass(createModuleToFunctionPassAdaptor(PromotePass())); 1241 1242 // Linking modules together can lead to duplicate global constant, only 1243 // keep one copy of each constant. 1244 MPM.addPass(ConstantMergePass()); 1245 1246 // Remove unused arguments from functions. 1247 MPM.addPass(DeadArgumentEliminationPass()); 1248 1249 // Reduce the code after globalopt and ipsccp. Both can open up significant 1250 // simplification opportunities, and both can propagate functions through 1251 // function pointers. When this happens, we often have to resolve varargs 1252 // calls, etc, so let instcombine do this. 1253 FunctionPassManager PeepholeFPM(DebugLogging); 1254 if (Level == O3) 1255 PeepholeFPM.addPass(AggressiveInstCombinePass()); 1256 PeepholeFPM.addPass(InstCombinePass()); 1257 invokePeepholeEPCallbacks(PeepholeFPM, Level); 1258 1259 MPM.addPass(createModuleToFunctionPassAdaptor(std::move(PeepholeFPM))); 1260 1261 // Note: historically, the PruneEH pass was run first to deduce nounwind and 1262 // generally clean up exception handling overhead. It isn't clear this is 1263 // valuable as the inliner doesn't currently care whether it is inlining an 1264 // invoke or a call. 1265 // Run the inliner now. 1266 MPM.addPass(createModuleToPostOrderCGSCCPassAdaptor( 1267 InlinerPass(getInlineParamsFromOptLevel(Level)))); 1268 1269 // Optimize globals again after we ran the inliner. 1270 MPM.addPass(GlobalOptPass()); 1271 1272 // Garbage collect dead functions. 1273 // FIXME: Add ArgumentPromotion pass after once it's ported. 1274 MPM.addPass(GlobalDCEPass()); 1275 1276 FunctionPassManager FPM(DebugLogging); 1277 // The IPO Passes may leave cruft around. Clean up after them. 1278 FPM.addPass(InstCombinePass()); 1279 invokePeepholeEPCallbacks(FPM, Level); 1280 1281 FPM.addPass(JumpThreadingPass()); 1282 1283 // Do a post inline PGO instrumentation and use pass. This is a context 1284 // sensitive PGO pass. 1285 if (PGOOpt) { 1286 if (PGOOpt->CSAction == PGOOptions::CSIRInstr) 1287 addPGOInstrPasses(MPM, DebugLogging, Level, /* RunProfileGen */ true, 1288 /* IsCS */ true, PGOOpt->CSProfileGenFile, 1289 PGOOpt->ProfileRemappingFile); 1290 else if (PGOOpt->CSAction == PGOOptions::CSIRUse) 1291 addPGOInstrPasses(MPM, DebugLogging, Level, /* RunProfileGen */ false, 1292 /* IsCS */ true, PGOOpt->ProfileFile, 1293 PGOOpt->ProfileRemappingFile); 1294 } 1295 1296 // Break up allocas 1297 FPM.addPass(SROA()); 1298 1299 // LTO provides additional opportunities for tailcall elimination due to 1300 // link-time inlining, and visibility of nocapture attribute. 1301 FPM.addPass(TailCallElimPass()); 1302 1303 // Run a few AA driver optimizations here and now to cleanup the code. 1304 MPM.addPass(createModuleToFunctionPassAdaptor(std::move(FPM))); 1305 1306 MPM.addPass(createModuleToPostOrderCGSCCPassAdaptor( 1307 PostOrderFunctionAttrsPass())); 1308 // FIXME: here we run IP alias analysis in the legacy PM. 1309 1310 FunctionPassManager MainFPM; 1311 1312 // FIXME: once we fix LoopPass Manager, add LICM here. 1313 // FIXME: once we provide support for enabling MLSM, add it here. 1314 if (RunNewGVN) 1315 MainFPM.addPass(NewGVNPass()); 1316 else 1317 MainFPM.addPass(GVN()); 1318 1319 // Remove dead memcpy()'s. 1320 MainFPM.addPass(MemCpyOptPass()); 1321 1322 // Nuke dead stores. 1323 MainFPM.addPass(DSEPass()); 1324 1325 // FIXME: at this point, we run a bunch of loop passes: 1326 // indVarSimplify, loopDeletion, loopInterchange, loopUnroll, 1327 // loopVectorize. Enable them once the remaining issue with LPM 1328 // are sorted out. 1329 1330 MainFPM.addPass(InstCombinePass()); 1331 MainFPM.addPass(SimplifyCFGPass()); 1332 MainFPM.addPass(SCCPPass()); 1333 MainFPM.addPass(InstCombinePass()); 1334 MainFPM.addPass(BDCEPass()); 1335 1336 // FIXME: We may want to run SLPVectorizer here. 1337 // After vectorization, assume intrinsics may tell us more 1338 // about pointer alignments. 1339 #if 0 1340 MainFPM.add(AlignmentFromAssumptionsPass()); 1341 #endif 1342 1343 // FIXME: Conditionally run LoadCombine here, after it's ported 1344 // (in case we still have this pass, given its questionable usefulness). 1345 1346 MainFPM.addPass(InstCombinePass()); 1347 invokePeepholeEPCallbacks(MainFPM, Level); 1348 MainFPM.addPass(JumpThreadingPass()); 1349 MPM.addPass(createModuleToFunctionPassAdaptor(std::move(MainFPM))); 1350 1351 // Create a function that performs CFI checks for cross-DSO calls with 1352 // targets in the current module. 1353 MPM.addPass(CrossDSOCFIPass()); 1354 1355 // Lower type metadata and the type.test intrinsic. This pass supports 1356 // clang's control flow integrity mechanisms (-fsanitize=cfi*) and needs 1357 // to be run at link time if CFI is enabled. This pass does nothing if 1358 // CFI is disabled. 1359 MPM.addPass(LowerTypeTestsPass(ExportSummary, nullptr)); 1360 1361 // Enable splitting late in the FullLTO post-link pipeline. This is done in 1362 // the same stage in the old pass manager (\ref addLateLTOOptimizationPasses). 1363 if (EnableHotColdSplit) 1364 MPM.addPass(HotColdSplittingPass()); 1365 1366 // Add late LTO optimization passes. 1367 // Delete basic blocks, which optimization passes may have killed. 1368 MPM.addPass(createModuleToFunctionPassAdaptor(SimplifyCFGPass())); 1369 1370 // Drop bodies of available eternally objects to improve GlobalDCE. 1371 MPM.addPass(EliminateAvailableExternallyPass()); 1372 1373 // Now that we have optimized the program, discard unreachable functions. 1374 MPM.addPass(GlobalDCEPass()); 1375 1376 // FIXME: Maybe enable MergeFuncs conditionally after it's ported. 1377 return MPM; 1378 } 1379 1380 AAManager PassBuilder::buildDefaultAAPipeline() { 1381 AAManager AA; 1382 1383 // The order in which these are registered determines their priority when 1384 // being queried. 1385 1386 // First we register the basic alias analysis that provides the majority of 1387 // per-function local AA logic. This is a stateless, on-demand local set of 1388 // AA techniques. 1389 AA.registerFunctionAnalysis<BasicAA>(); 1390 1391 // Next we query fast, specialized alias analyses that wrap IR-embedded 1392 // information about aliasing. 1393 AA.registerFunctionAnalysis<ScopedNoAliasAA>(); 1394 AA.registerFunctionAnalysis<TypeBasedAA>(); 1395 1396 // Add support for querying global aliasing information when available. 1397 // Because the `AAManager` is a function analysis and `GlobalsAA` is a module 1398 // analysis, all that the `AAManager` can do is query for any *cached* 1399 // results from `GlobalsAA` through a readonly proxy. 1400 AA.registerModuleAnalysis<GlobalsAA>(); 1401 1402 return AA; 1403 } 1404 1405 static Optional<int> parseRepeatPassName(StringRef Name) { 1406 if (!Name.consume_front("repeat<") || !Name.consume_back(">")) 1407 return None; 1408 int Count; 1409 if (Name.getAsInteger(0, Count) || Count <= 0) 1410 return None; 1411 return Count; 1412 } 1413 1414 static Optional<int> parseDevirtPassName(StringRef Name) { 1415 if (!Name.consume_front("devirt<") || !Name.consume_back(">")) 1416 return None; 1417 int Count; 1418 if (Name.getAsInteger(0, Count) || Count <= 0) 1419 return None; 1420 return Count; 1421 } 1422 1423 static bool checkParametrizedPassName(StringRef Name, StringRef PassName) { 1424 if (!Name.consume_front(PassName)) 1425 return false; 1426 // normal pass name w/o parameters == default parameters 1427 if (Name.empty()) 1428 return true; 1429 return Name.startswith("<") && Name.endswith(">"); 1430 } 1431 1432 namespace { 1433 1434 /// This performs customized parsing of pass name with parameters. 1435 /// 1436 /// We do not need parametrization of passes in textual pipeline very often, 1437 /// yet on a rare occasion ability to specify parameters right there can be 1438 /// useful. 1439 /// 1440 /// \p Name - parameterized specification of a pass from a textual pipeline 1441 /// is a string in a form of : 1442 /// PassName '<' parameter-list '>' 1443 /// 1444 /// Parameter list is being parsed by the parser callable argument, \p Parser, 1445 /// It takes a string-ref of parameters and returns either StringError or a 1446 /// parameter list in a form of a custom parameters type, all wrapped into 1447 /// Expected<> template class. 1448 /// 1449 template <typename ParametersParseCallableT> 1450 auto parsePassParameters(ParametersParseCallableT &&Parser, StringRef Name, 1451 StringRef PassName) -> decltype(Parser(StringRef{})) { 1452 using ParametersT = typename decltype(Parser(StringRef{}))::value_type; 1453 1454 StringRef Params = Name; 1455 if (!Params.consume_front(PassName)) { 1456 assert(false && 1457 "unable to strip pass name from parametrized pass specification"); 1458 } 1459 if (Params.empty()) 1460 return ParametersT{}; 1461 if (!Params.consume_front("<") || !Params.consume_back(">")) { 1462 assert(false && "invalid format for parametrized pass name"); 1463 } 1464 1465 Expected<ParametersT> Result = Parser(Params); 1466 assert((Result || Result.template errorIsA<StringError>()) && 1467 "Pass parameter parser can only return StringErrors."); 1468 return Result; 1469 } 1470 1471 /// Parser of parameters for LoopUnroll pass. 1472 Expected<LoopUnrollOptions> parseLoopUnrollOptions(StringRef Params) { 1473 LoopUnrollOptions UnrollOpts; 1474 while (!Params.empty()) { 1475 StringRef ParamName; 1476 std::tie(ParamName, Params) = Params.split(';'); 1477 int OptLevel = StringSwitch<int>(ParamName) 1478 .Case("O0", 0) 1479 .Case("O1", 1) 1480 .Case("O2", 2) 1481 .Case("O3", 3) 1482 .Default(-1); 1483 if (OptLevel >= 0) { 1484 UnrollOpts.setOptLevel(OptLevel); 1485 continue; 1486 } 1487 if (ParamName.consume_front("full-unroll-max=")) { 1488 int Count; 1489 if (ParamName.getAsInteger(0, Count)) 1490 return make_error<StringError>( 1491 formatv("invalid LoopUnrollPass parameter '{0}' ", ParamName).str(), 1492 inconvertibleErrorCode()); 1493 UnrollOpts.setFullUnrollMaxCount(Count); 1494 continue; 1495 } 1496 1497 bool Enable = !ParamName.consume_front("no-"); 1498 if (ParamName == "partial") { 1499 UnrollOpts.setPartial(Enable); 1500 } else if (ParamName == "peeling") { 1501 UnrollOpts.setPeeling(Enable); 1502 } else if (ParamName == "profile-peeling") { 1503 UnrollOpts.setProfileBasedPeeling(Enable); 1504 } else if (ParamName == "runtime") { 1505 UnrollOpts.setRuntime(Enable); 1506 } else if (ParamName == "upperbound") { 1507 UnrollOpts.setUpperBound(Enable); 1508 } else { 1509 return make_error<StringError>( 1510 formatv("invalid LoopUnrollPass parameter '{0}' ", ParamName).str(), 1511 inconvertibleErrorCode()); 1512 } 1513 } 1514 return UnrollOpts; 1515 } 1516 1517 Expected<MemorySanitizerOptions> parseMSanPassOptions(StringRef Params) { 1518 MemorySanitizerOptions Result; 1519 while (!Params.empty()) { 1520 StringRef ParamName; 1521 std::tie(ParamName, Params) = Params.split(';'); 1522 1523 if (ParamName == "recover") { 1524 Result.Recover = true; 1525 } else if (ParamName == "kernel") { 1526 Result.Kernel = true; 1527 } else if (ParamName.consume_front("track-origins=")) { 1528 if (ParamName.getAsInteger(0, Result.TrackOrigins)) 1529 return make_error<StringError>( 1530 formatv("invalid argument to MemorySanitizer pass track-origins " 1531 "parameter: '{0}' ", 1532 ParamName) 1533 .str(), 1534 inconvertibleErrorCode()); 1535 } else { 1536 return make_error<StringError>( 1537 formatv("invalid MemorySanitizer pass parameter '{0}' ", ParamName) 1538 .str(), 1539 inconvertibleErrorCode()); 1540 } 1541 } 1542 return Result; 1543 } 1544 1545 /// Parser of parameters for SimplifyCFG pass. 1546 Expected<SimplifyCFGOptions> parseSimplifyCFGOptions(StringRef Params) { 1547 SimplifyCFGOptions Result; 1548 while (!Params.empty()) { 1549 StringRef ParamName; 1550 std::tie(ParamName, Params) = Params.split(';'); 1551 1552 bool Enable = !ParamName.consume_front("no-"); 1553 if (ParamName == "forward-switch-cond") { 1554 Result.forwardSwitchCondToPhi(Enable); 1555 } else if (ParamName == "switch-to-lookup") { 1556 Result.convertSwitchToLookupTable(Enable); 1557 } else if (ParamName == "keep-loops") { 1558 Result.needCanonicalLoops(Enable); 1559 } else if (ParamName == "sink-common-insts") { 1560 Result.sinkCommonInsts(Enable); 1561 } else if (Enable && ParamName.consume_front("bonus-inst-threshold=")) { 1562 APInt BonusInstThreshold; 1563 if (ParamName.getAsInteger(0, BonusInstThreshold)) 1564 return make_error<StringError>( 1565 formatv("invalid argument to SimplifyCFG pass bonus-threshold " 1566 "parameter: '{0}' ", 1567 ParamName).str(), 1568 inconvertibleErrorCode()); 1569 Result.bonusInstThreshold(BonusInstThreshold.getSExtValue()); 1570 } else { 1571 return make_error<StringError>( 1572 formatv("invalid SimplifyCFG pass parameter '{0}' ", ParamName).str(), 1573 inconvertibleErrorCode()); 1574 } 1575 } 1576 return Result; 1577 } 1578 1579 /// Parser of parameters for LoopVectorize pass. 1580 Expected<LoopVectorizeOptions> parseLoopVectorizeOptions(StringRef Params) { 1581 LoopVectorizeOptions Opts; 1582 while (!Params.empty()) { 1583 StringRef ParamName; 1584 std::tie(ParamName, Params) = Params.split(';'); 1585 1586 bool Enable = !ParamName.consume_front("no-"); 1587 if (ParamName == "interleave-forced-only") { 1588 Opts.setInterleaveOnlyWhenForced(Enable); 1589 } else if (ParamName == "vectorize-forced-only") { 1590 Opts.setVectorizeOnlyWhenForced(Enable); 1591 } else { 1592 return make_error<StringError>( 1593 formatv("invalid LoopVectorize parameter '{0}' ", ParamName).str(), 1594 inconvertibleErrorCode()); 1595 } 1596 } 1597 return Opts; 1598 } 1599 1600 Expected<bool> parseLoopUnswitchOptions(StringRef Params) { 1601 bool Result = false; 1602 while (!Params.empty()) { 1603 StringRef ParamName; 1604 std::tie(ParamName, Params) = Params.split(';'); 1605 1606 bool Enable = !ParamName.consume_front("no-"); 1607 if (ParamName == "nontrivial") { 1608 Result = Enable; 1609 } else { 1610 return make_error<StringError>( 1611 formatv("invalid LoopUnswitch pass parameter '{0}' ", ParamName) 1612 .str(), 1613 inconvertibleErrorCode()); 1614 } 1615 } 1616 return Result; 1617 } 1618 1619 Expected<bool> parseMergedLoadStoreMotionOptions(StringRef Params) { 1620 bool Result = false; 1621 while (!Params.empty()) { 1622 StringRef ParamName; 1623 std::tie(ParamName, Params) = Params.split(';'); 1624 1625 bool Enable = !ParamName.consume_front("no-"); 1626 if (ParamName == "split-footer-bb") { 1627 Result = Enable; 1628 } else { 1629 return make_error<StringError>( 1630 formatv("invalid MergedLoadStoreMotion pass parameter '{0}' ", 1631 ParamName) 1632 .str(), 1633 inconvertibleErrorCode()); 1634 } 1635 } 1636 return Result; 1637 } 1638 } // namespace 1639 1640 /// Tests whether a pass name starts with a valid prefix for a default pipeline 1641 /// alias. 1642 static bool startsWithDefaultPipelineAliasPrefix(StringRef Name) { 1643 return Name.startswith("default") || Name.startswith("thinlto") || 1644 Name.startswith("lto"); 1645 } 1646 1647 /// Tests whether registered callbacks will accept a given pass name. 1648 /// 1649 /// When parsing a pipeline text, the type of the outermost pipeline may be 1650 /// omitted, in which case the type is automatically determined from the first 1651 /// pass name in the text. This may be a name that is handled through one of the 1652 /// callbacks. We check this through the oridinary parsing callbacks by setting 1653 /// up a dummy PassManager in order to not force the client to also handle this 1654 /// type of query. 1655 template <typename PassManagerT, typename CallbacksT> 1656 static bool callbacksAcceptPassName(StringRef Name, CallbacksT &Callbacks) { 1657 if (!Callbacks.empty()) { 1658 PassManagerT DummyPM; 1659 for (auto &CB : Callbacks) 1660 if (CB(Name, DummyPM, {})) 1661 return true; 1662 } 1663 return false; 1664 } 1665 1666 template <typename CallbacksT> 1667 static bool isModulePassName(StringRef Name, CallbacksT &Callbacks) { 1668 // Manually handle aliases for pre-configured pipeline fragments. 1669 if (startsWithDefaultPipelineAliasPrefix(Name)) 1670 return DefaultAliasRegex.match(Name); 1671 1672 // Explicitly handle pass manager names. 1673 if (Name == "module") 1674 return true; 1675 if (Name == "cgscc") 1676 return true; 1677 if (Name == "function") 1678 return true; 1679 1680 // Explicitly handle custom-parsed pass names. 1681 if (parseRepeatPassName(Name)) 1682 return true; 1683 1684 #define MODULE_PASS(NAME, CREATE_PASS) \ 1685 if (Name == NAME) \ 1686 return true; 1687 #define MODULE_ANALYSIS(NAME, CREATE_PASS) \ 1688 if (Name == "require<" NAME ">" || Name == "invalidate<" NAME ">") \ 1689 return true; 1690 #include "PassRegistry.def" 1691 1692 return callbacksAcceptPassName<ModulePassManager>(Name, Callbacks); 1693 } 1694 1695 template <typename CallbacksT> 1696 static bool isCGSCCPassName(StringRef Name, CallbacksT &Callbacks) { 1697 // Explicitly handle pass manager names. 1698 if (Name == "cgscc") 1699 return true; 1700 if (Name == "function") 1701 return true; 1702 1703 // Explicitly handle custom-parsed pass names. 1704 if (parseRepeatPassName(Name)) 1705 return true; 1706 if (parseDevirtPassName(Name)) 1707 return true; 1708 1709 #define CGSCC_PASS(NAME, CREATE_PASS) \ 1710 if (Name == NAME) \ 1711 return true; 1712 #define CGSCC_ANALYSIS(NAME, CREATE_PASS) \ 1713 if (Name == "require<" NAME ">" || Name == "invalidate<" NAME ">") \ 1714 return true; 1715 #include "PassRegistry.def" 1716 1717 return callbacksAcceptPassName<CGSCCPassManager>(Name, Callbacks); 1718 } 1719 1720 template <typename CallbacksT> 1721 static bool isFunctionPassName(StringRef Name, CallbacksT &Callbacks) { 1722 // Explicitly handle pass manager names. 1723 if (Name == "function") 1724 return true; 1725 if (Name == "loop" || Name == "loop-mssa") 1726 return true; 1727 1728 // Explicitly handle custom-parsed pass names. 1729 if (parseRepeatPassName(Name)) 1730 return true; 1731 1732 #define FUNCTION_PASS(NAME, CREATE_PASS) \ 1733 if (Name == NAME) \ 1734 return true; 1735 #define FUNCTION_PASS_WITH_PARAMS(NAME, CREATE_PASS, PARSER) \ 1736 if (checkParametrizedPassName(Name, NAME)) \ 1737 return true; 1738 #define FUNCTION_ANALYSIS(NAME, CREATE_PASS) \ 1739 if (Name == "require<" NAME ">" || Name == "invalidate<" NAME ">") \ 1740 return true; 1741 #include "PassRegistry.def" 1742 1743 return callbacksAcceptPassName<FunctionPassManager>(Name, Callbacks); 1744 } 1745 1746 template <typename CallbacksT> 1747 static bool isLoopPassName(StringRef Name, CallbacksT &Callbacks) { 1748 // Explicitly handle pass manager names. 1749 if (Name == "loop" || Name == "loop-mssa") 1750 return true; 1751 1752 // Explicitly handle custom-parsed pass names. 1753 if (parseRepeatPassName(Name)) 1754 return true; 1755 1756 #define LOOP_PASS(NAME, CREATE_PASS) \ 1757 if (Name == NAME) \ 1758 return true; 1759 #define LOOP_PASS_WITH_PARAMS(NAME, CREATE_PASS, PARSER) \ 1760 if (checkParametrizedPassName(Name, NAME)) \ 1761 return true; 1762 #define LOOP_ANALYSIS(NAME, CREATE_PASS) \ 1763 if (Name == "require<" NAME ">" || Name == "invalidate<" NAME ">") \ 1764 return true; 1765 #include "PassRegistry.def" 1766 1767 return callbacksAcceptPassName<LoopPassManager>(Name, Callbacks); 1768 } 1769 1770 Optional<std::vector<PassBuilder::PipelineElement>> 1771 PassBuilder::parsePipelineText(StringRef Text) { 1772 std::vector<PipelineElement> ResultPipeline; 1773 1774 SmallVector<std::vector<PipelineElement> *, 4> PipelineStack = { 1775 &ResultPipeline}; 1776 for (;;) { 1777 std::vector<PipelineElement> &Pipeline = *PipelineStack.back(); 1778 size_t Pos = Text.find_first_of(",()"); 1779 Pipeline.push_back({Text.substr(0, Pos), {}}); 1780 1781 // If we have a single terminating name, we're done. 1782 if (Pos == Text.npos) 1783 break; 1784 1785 char Sep = Text[Pos]; 1786 Text = Text.substr(Pos + 1); 1787 if (Sep == ',') 1788 // Just a name ending in a comma, continue. 1789 continue; 1790 1791 if (Sep == '(') { 1792 // Push the inner pipeline onto the stack to continue processing. 1793 PipelineStack.push_back(&Pipeline.back().InnerPipeline); 1794 continue; 1795 } 1796 1797 assert(Sep == ')' && "Bogus separator!"); 1798 // When handling the close parenthesis, we greedily consume them to avoid 1799 // empty strings in the pipeline. 1800 do { 1801 // If we try to pop the outer pipeline we have unbalanced parentheses. 1802 if (PipelineStack.size() == 1) 1803 return None; 1804 1805 PipelineStack.pop_back(); 1806 } while (Text.consume_front(")")); 1807 1808 // Check if we've finished parsing. 1809 if (Text.empty()) 1810 break; 1811 1812 // Otherwise, the end of an inner pipeline always has to be followed by 1813 // a comma, and then we can continue. 1814 if (!Text.consume_front(",")) 1815 return None; 1816 } 1817 1818 if (PipelineStack.size() > 1) 1819 // Unbalanced paretheses. 1820 return None; 1821 1822 assert(PipelineStack.back() == &ResultPipeline && 1823 "Wrong pipeline at the bottom of the stack!"); 1824 return {std::move(ResultPipeline)}; 1825 } 1826 1827 Error PassBuilder::parseModulePass(ModulePassManager &MPM, 1828 const PipelineElement &E, 1829 bool VerifyEachPass, bool DebugLogging) { 1830 auto &Name = E.Name; 1831 auto &InnerPipeline = E.InnerPipeline; 1832 1833 // First handle complex passes like the pass managers which carry pipelines. 1834 if (!InnerPipeline.empty()) { 1835 if (Name == "module") { 1836 ModulePassManager NestedMPM(DebugLogging); 1837 if (auto Err = parseModulePassPipeline(NestedMPM, InnerPipeline, 1838 VerifyEachPass, DebugLogging)) 1839 return Err; 1840 MPM.addPass(std::move(NestedMPM)); 1841 return Error::success(); 1842 } 1843 if (Name == "cgscc") { 1844 CGSCCPassManager CGPM(DebugLogging); 1845 if (auto Err = parseCGSCCPassPipeline(CGPM, InnerPipeline, VerifyEachPass, 1846 DebugLogging)) 1847 return Err; 1848 MPM.addPass(createModuleToPostOrderCGSCCPassAdaptor(std::move(CGPM))); 1849 return Error::success(); 1850 } 1851 if (Name == "function") { 1852 FunctionPassManager FPM(DebugLogging); 1853 if (auto Err = parseFunctionPassPipeline(FPM, InnerPipeline, 1854 VerifyEachPass, DebugLogging)) 1855 return Err; 1856 MPM.addPass(createModuleToFunctionPassAdaptor(std::move(FPM))); 1857 return Error::success(); 1858 } 1859 if (auto Count = parseRepeatPassName(Name)) { 1860 ModulePassManager NestedMPM(DebugLogging); 1861 if (auto Err = parseModulePassPipeline(NestedMPM, InnerPipeline, 1862 VerifyEachPass, DebugLogging)) 1863 return Err; 1864 MPM.addPass(createRepeatedPass(*Count, std::move(NestedMPM))); 1865 return Error::success(); 1866 } 1867 1868 for (auto &C : ModulePipelineParsingCallbacks) 1869 if (C(Name, MPM, InnerPipeline)) 1870 return Error::success(); 1871 1872 // Normal passes can't have pipelines. 1873 return make_error<StringError>( 1874 formatv("invalid use of '{0}' pass as module pipeline", Name).str(), 1875 inconvertibleErrorCode()); 1876 ; 1877 } 1878 1879 // Manually handle aliases for pre-configured pipeline fragments. 1880 if (startsWithDefaultPipelineAliasPrefix(Name)) { 1881 SmallVector<StringRef, 3> Matches; 1882 if (!DefaultAliasRegex.match(Name, &Matches)) 1883 return make_error<StringError>( 1884 formatv("unknown default pipeline alias '{0}'", Name).str(), 1885 inconvertibleErrorCode()); 1886 1887 assert(Matches.size() == 3 && "Must capture two matched strings!"); 1888 1889 OptimizationLevel L = StringSwitch<OptimizationLevel>(Matches[2]) 1890 .Case("O0", O0) 1891 .Case("O1", O1) 1892 .Case("O2", O2) 1893 .Case("O3", O3) 1894 .Case("Os", Os) 1895 .Case("Oz", Oz); 1896 if (L == O0) { 1897 // Add instrumentation PGO passes -- at O0 we can still do PGO. 1898 if (PGOOpt && Matches[1] != "thinlto" && 1899 (PGOOpt->Action == PGOOptions::IRInstr || 1900 PGOOpt->Action == PGOOptions::IRUse)) 1901 addPGOInstrPassesForO0( 1902 MPM, DebugLogging, 1903 /* RunProfileGen */ (PGOOpt->Action == PGOOptions::IRInstr), 1904 /* IsCS */ false, PGOOpt->ProfileFile, 1905 PGOOpt->ProfileRemappingFile); 1906 // Do nothing else at all! 1907 return Error::success(); 1908 } 1909 1910 // This is consistent with old pass manager invoked via opt, but 1911 // inconsistent with clang. Clang doesn't enable loop vectorization 1912 // but does enable slp vectorization at Oz. 1913 PTO.LoopVectorization = L > O1 && L < Oz; 1914 PTO.SLPVectorization = L > O1 && L < Oz; 1915 1916 if (Matches[1] == "default") { 1917 MPM.addPass(buildPerModuleDefaultPipeline(L, DebugLogging)); 1918 } else if (Matches[1] == "thinlto-pre-link") { 1919 MPM.addPass(buildThinLTOPreLinkDefaultPipeline(L, DebugLogging)); 1920 } else if (Matches[1] == "thinlto") { 1921 MPM.addPass(buildThinLTODefaultPipeline(L, DebugLogging, nullptr)); 1922 } else if (Matches[1] == "lto-pre-link") { 1923 MPM.addPass(buildLTOPreLinkDefaultPipeline(L, DebugLogging)); 1924 } else { 1925 assert(Matches[1] == "lto" && "Not one of the matched options!"); 1926 MPM.addPass(buildLTODefaultPipeline(L, DebugLogging, nullptr)); 1927 } 1928 return Error::success(); 1929 } 1930 1931 // Finally expand the basic registered passes from the .inc file. 1932 #define MODULE_PASS(NAME, CREATE_PASS) \ 1933 if (Name == NAME) { \ 1934 MPM.addPass(CREATE_PASS); \ 1935 return Error::success(); \ 1936 } 1937 #define MODULE_ANALYSIS(NAME, CREATE_PASS) \ 1938 if (Name == "require<" NAME ">") { \ 1939 MPM.addPass( \ 1940 RequireAnalysisPass< \ 1941 std::remove_reference<decltype(CREATE_PASS)>::type, Module>()); \ 1942 return Error::success(); \ 1943 } \ 1944 if (Name == "invalidate<" NAME ">") { \ 1945 MPM.addPass(InvalidateAnalysisPass< \ 1946 std::remove_reference<decltype(CREATE_PASS)>::type>()); \ 1947 return Error::success(); \ 1948 } 1949 #include "PassRegistry.def" 1950 1951 for (auto &C : ModulePipelineParsingCallbacks) 1952 if (C(Name, MPM, InnerPipeline)) 1953 return Error::success(); 1954 return make_error<StringError>( 1955 formatv("unknown module pass '{0}'", Name).str(), 1956 inconvertibleErrorCode()); 1957 } 1958 1959 Error PassBuilder::parseCGSCCPass(CGSCCPassManager &CGPM, 1960 const PipelineElement &E, bool VerifyEachPass, 1961 bool DebugLogging) { 1962 auto &Name = E.Name; 1963 auto &InnerPipeline = E.InnerPipeline; 1964 1965 // First handle complex passes like the pass managers which carry pipelines. 1966 if (!InnerPipeline.empty()) { 1967 if (Name == "cgscc") { 1968 CGSCCPassManager NestedCGPM(DebugLogging); 1969 if (auto Err = parseCGSCCPassPipeline(NestedCGPM, InnerPipeline, 1970 VerifyEachPass, DebugLogging)) 1971 return Err; 1972 // Add the nested pass manager with the appropriate adaptor. 1973 CGPM.addPass(std::move(NestedCGPM)); 1974 return Error::success(); 1975 } 1976 if (Name == "function") { 1977 FunctionPassManager FPM(DebugLogging); 1978 if (auto Err = parseFunctionPassPipeline(FPM, InnerPipeline, 1979 VerifyEachPass, DebugLogging)) 1980 return Err; 1981 // Add the nested pass manager with the appropriate adaptor. 1982 CGPM.addPass(createCGSCCToFunctionPassAdaptor(std::move(FPM))); 1983 return Error::success(); 1984 } 1985 if (auto Count = parseRepeatPassName(Name)) { 1986 CGSCCPassManager NestedCGPM(DebugLogging); 1987 if (auto Err = parseCGSCCPassPipeline(NestedCGPM, InnerPipeline, 1988 VerifyEachPass, DebugLogging)) 1989 return Err; 1990 CGPM.addPass(createRepeatedPass(*Count, std::move(NestedCGPM))); 1991 return Error::success(); 1992 } 1993 if (auto MaxRepetitions = parseDevirtPassName(Name)) { 1994 CGSCCPassManager NestedCGPM(DebugLogging); 1995 if (auto Err = parseCGSCCPassPipeline(NestedCGPM, InnerPipeline, 1996 VerifyEachPass, DebugLogging)) 1997 return Err; 1998 CGPM.addPass( 1999 createDevirtSCCRepeatedPass(std::move(NestedCGPM), *MaxRepetitions)); 2000 return Error::success(); 2001 } 2002 2003 for (auto &C : CGSCCPipelineParsingCallbacks) 2004 if (C(Name, CGPM, InnerPipeline)) 2005 return Error::success(); 2006 2007 // Normal passes can't have pipelines. 2008 return make_error<StringError>( 2009 formatv("invalid use of '{0}' pass as cgscc pipeline", Name).str(), 2010 inconvertibleErrorCode()); 2011 } 2012 2013 // Now expand the basic registered passes from the .inc file. 2014 #define CGSCC_PASS(NAME, CREATE_PASS) \ 2015 if (Name == NAME) { \ 2016 CGPM.addPass(CREATE_PASS); \ 2017 return Error::success(); \ 2018 } 2019 #define CGSCC_ANALYSIS(NAME, CREATE_PASS) \ 2020 if (Name == "require<" NAME ">") { \ 2021 CGPM.addPass(RequireAnalysisPass< \ 2022 std::remove_reference<decltype(CREATE_PASS)>::type, \ 2023 LazyCallGraph::SCC, CGSCCAnalysisManager, LazyCallGraph &, \ 2024 CGSCCUpdateResult &>()); \ 2025 return Error::success(); \ 2026 } \ 2027 if (Name == "invalidate<" NAME ">") { \ 2028 CGPM.addPass(InvalidateAnalysisPass< \ 2029 std::remove_reference<decltype(CREATE_PASS)>::type>()); \ 2030 return Error::success(); \ 2031 } 2032 #include "PassRegistry.def" 2033 2034 for (auto &C : CGSCCPipelineParsingCallbacks) 2035 if (C(Name, CGPM, InnerPipeline)) 2036 return Error::success(); 2037 return make_error<StringError>( 2038 formatv("unknown cgscc pass '{0}'", Name).str(), 2039 inconvertibleErrorCode()); 2040 } 2041 2042 Error PassBuilder::parseFunctionPass(FunctionPassManager &FPM, 2043 const PipelineElement &E, 2044 bool VerifyEachPass, bool DebugLogging) { 2045 auto &Name = E.Name; 2046 auto &InnerPipeline = E.InnerPipeline; 2047 2048 // First handle complex passes like the pass managers which carry pipelines. 2049 if (!InnerPipeline.empty()) { 2050 if (Name == "function") { 2051 FunctionPassManager NestedFPM(DebugLogging); 2052 if (auto Err = parseFunctionPassPipeline(NestedFPM, InnerPipeline, 2053 VerifyEachPass, DebugLogging)) 2054 return Err; 2055 // Add the nested pass manager with the appropriate adaptor. 2056 FPM.addPass(std::move(NestedFPM)); 2057 return Error::success(); 2058 } 2059 if (Name == "loop" || Name == "loop-mssa") { 2060 LoopPassManager LPM(DebugLogging); 2061 if (auto Err = parseLoopPassPipeline(LPM, InnerPipeline, VerifyEachPass, 2062 DebugLogging)) 2063 return Err; 2064 // Add the nested pass manager with the appropriate adaptor. 2065 bool UseMemorySSA = (Name == "loop-mssa"); 2066 FPM.addPass(createFunctionToLoopPassAdaptor(std::move(LPM), UseMemorySSA, 2067 DebugLogging)); 2068 return Error::success(); 2069 } 2070 if (auto Count = parseRepeatPassName(Name)) { 2071 FunctionPassManager NestedFPM(DebugLogging); 2072 if (auto Err = parseFunctionPassPipeline(NestedFPM, InnerPipeline, 2073 VerifyEachPass, DebugLogging)) 2074 return Err; 2075 FPM.addPass(createRepeatedPass(*Count, std::move(NestedFPM))); 2076 return Error::success(); 2077 } 2078 2079 for (auto &C : FunctionPipelineParsingCallbacks) 2080 if (C(Name, FPM, InnerPipeline)) 2081 return Error::success(); 2082 2083 // Normal passes can't have pipelines. 2084 return make_error<StringError>( 2085 formatv("invalid use of '{0}' pass as function pipeline", Name).str(), 2086 inconvertibleErrorCode()); 2087 } 2088 2089 // Now expand the basic registered passes from the .inc file. 2090 #define FUNCTION_PASS(NAME, CREATE_PASS) \ 2091 if (Name == NAME) { \ 2092 FPM.addPass(CREATE_PASS); \ 2093 return Error::success(); \ 2094 } 2095 #define FUNCTION_PASS_WITH_PARAMS(NAME, CREATE_PASS, PARSER) \ 2096 if (checkParametrizedPassName(Name, NAME)) { \ 2097 auto Params = parsePassParameters(PARSER, Name, NAME); \ 2098 if (!Params) \ 2099 return Params.takeError(); \ 2100 FPM.addPass(CREATE_PASS(Params.get())); \ 2101 return Error::success(); \ 2102 } 2103 #define FUNCTION_ANALYSIS(NAME, CREATE_PASS) \ 2104 if (Name == "require<" NAME ">") { \ 2105 FPM.addPass( \ 2106 RequireAnalysisPass< \ 2107 std::remove_reference<decltype(CREATE_PASS)>::type, Function>()); \ 2108 return Error::success(); \ 2109 } \ 2110 if (Name == "invalidate<" NAME ">") { \ 2111 FPM.addPass(InvalidateAnalysisPass< \ 2112 std::remove_reference<decltype(CREATE_PASS)>::type>()); \ 2113 return Error::success(); \ 2114 } 2115 #include "PassRegistry.def" 2116 2117 for (auto &C : FunctionPipelineParsingCallbacks) 2118 if (C(Name, FPM, InnerPipeline)) 2119 return Error::success(); 2120 return make_error<StringError>( 2121 formatv("unknown function pass '{0}'", Name).str(), 2122 inconvertibleErrorCode()); 2123 } 2124 2125 Error PassBuilder::parseLoopPass(LoopPassManager &LPM, const PipelineElement &E, 2126 bool VerifyEachPass, bool DebugLogging) { 2127 StringRef Name = E.Name; 2128 auto &InnerPipeline = E.InnerPipeline; 2129 2130 // First handle complex passes like the pass managers which carry pipelines. 2131 if (!InnerPipeline.empty()) { 2132 if (Name == "loop") { 2133 LoopPassManager NestedLPM(DebugLogging); 2134 if (auto Err = parseLoopPassPipeline(NestedLPM, InnerPipeline, 2135 VerifyEachPass, DebugLogging)) 2136 return Err; 2137 // Add the nested pass manager with the appropriate adaptor. 2138 LPM.addPass(std::move(NestedLPM)); 2139 return Error::success(); 2140 } 2141 if (auto Count = parseRepeatPassName(Name)) { 2142 LoopPassManager NestedLPM(DebugLogging); 2143 if (auto Err = parseLoopPassPipeline(NestedLPM, InnerPipeline, 2144 VerifyEachPass, DebugLogging)) 2145 return Err; 2146 LPM.addPass(createRepeatedPass(*Count, std::move(NestedLPM))); 2147 return Error::success(); 2148 } 2149 2150 for (auto &C : LoopPipelineParsingCallbacks) 2151 if (C(Name, LPM, InnerPipeline)) 2152 return Error::success(); 2153 2154 // Normal passes can't have pipelines. 2155 return make_error<StringError>( 2156 formatv("invalid use of '{0}' pass as loop pipeline", Name).str(), 2157 inconvertibleErrorCode()); 2158 } 2159 2160 // Now expand the basic registered passes from the .inc file. 2161 #define LOOP_PASS(NAME, CREATE_PASS) \ 2162 if (Name == NAME) { \ 2163 LPM.addPass(CREATE_PASS); \ 2164 return Error::success(); \ 2165 } 2166 #define LOOP_PASS_WITH_PARAMS(NAME, CREATE_PASS, PARSER) \ 2167 if (checkParametrizedPassName(Name, NAME)) { \ 2168 auto Params = parsePassParameters(PARSER, Name, NAME); \ 2169 if (!Params) \ 2170 return Params.takeError(); \ 2171 LPM.addPass(CREATE_PASS(Params.get())); \ 2172 return Error::success(); \ 2173 } 2174 #define LOOP_ANALYSIS(NAME, CREATE_PASS) \ 2175 if (Name == "require<" NAME ">") { \ 2176 LPM.addPass(RequireAnalysisPass< \ 2177 std::remove_reference<decltype(CREATE_PASS)>::type, Loop, \ 2178 LoopAnalysisManager, LoopStandardAnalysisResults &, \ 2179 LPMUpdater &>()); \ 2180 return Error::success(); \ 2181 } \ 2182 if (Name == "invalidate<" NAME ">") { \ 2183 LPM.addPass(InvalidateAnalysisPass< \ 2184 std::remove_reference<decltype(CREATE_PASS)>::type>()); \ 2185 return Error::success(); \ 2186 } 2187 #include "PassRegistry.def" 2188 2189 for (auto &C : LoopPipelineParsingCallbacks) 2190 if (C(Name, LPM, InnerPipeline)) 2191 return Error::success(); 2192 return make_error<StringError>(formatv("unknown loop pass '{0}'", Name).str(), 2193 inconvertibleErrorCode()); 2194 } 2195 2196 bool PassBuilder::parseAAPassName(AAManager &AA, StringRef Name) { 2197 #define MODULE_ALIAS_ANALYSIS(NAME, CREATE_PASS) \ 2198 if (Name == NAME) { \ 2199 AA.registerModuleAnalysis< \ 2200 std::remove_reference<decltype(CREATE_PASS)>::type>(); \ 2201 return true; \ 2202 } 2203 #define FUNCTION_ALIAS_ANALYSIS(NAME, CREATE_PASS) \ 2204 if (Name == NAME) { \ 2205 AA.registerFunctionAnalysis< \ 2206 std::remove_reference<decltype(CREATE_PASS)>::type>(); \ 2207 return true; \ 2208 } 2209 #include "PassRegistry.def" 2210 2211 for (auto &C : AAParsingCallbacks) 2212 if (C(Name, AA)) 2213 return true; 2214 return false; 2215 } 2216 2217 Error PassBuilder::parseLoopPassPipeline(LoopPassManager &LPM, 2218 ArrayRef<PipelineElement> Pipeline, 2219 bool VerifyEachPass, 2220 bool DebugLogging) { 2221 for (const auto &Element : Pipeline) { 2222 if (auto Err = parseLoopPass(LPM, Element, VerifyEachPass, DebugLogging)) 2223 return Err; 2224 // FIXME: No verifier support for Loop passes! 2225 } 2226 return Error::success(); 2227 } 2228 2229 Error PassBuilder::parseFunctionPassPipeline(FunctionPassManager &FPM, 2230 ArrayRef<PipelineElement> Pipeline, 2231 bool VerifyEachPass, 2232 bool DebugLogging) { 2233 for (const auto &Element : Pipeline) { 2234 if (auto Err = 2235 parseFunctionPass(FPM, Element, VerifyEachPass, DebugLogging)) 2236 return Err; 2237 if (VerifyEachPass) 2238 FPM.addPass(VerifierPass()); 2239 } 2240 return Error::success(); 2241 } 2242 2243 Error PassBuilder::parseCGSCCPassPipeline(CGSCCPassManager &CGPM, 2244 ArrayRef<PipelineElement> Pipeline, 2245 bool VerifyEachPass, 2246 bool DebugLogging) { 2247 for (const auto &Element : Pipeline) { 2248 if (auto Err = parseCGSCCPass(CGPM, Element, VerifyEachPass, DebugLogging)) 2249 return Err; 2250 // FIXME: No verifier support for CGSCC passes! 2251 } 2252 return Error::success(); 2253 } 2254 2255 void PassBuilder::crossRegisterProxies(LoopAnalysisManager &LAM, 2256 FunctionAnalysisManager &FAM, 2257 CGSCCAnalysisManager &CGAM, 2258 ModuleAnalysisManager &MAM) { 2259 MAM.registerPass([&] { return FunctionAnalysisManagerModuleProxy(FAM); }); 2260 MAM.registerPass([&] { return CGSCCAnalysisManagerModuleProxy(CGAM); }); 2261 CGAM.registerPass([&] { return ModuleAnalysisManagerCGSCCProxy(MAM); }); 2262 FAM.registerPass([&] { return CGSCCAnalysisManagerFunctionProxy(CGAM); }); 2263 FAM.registerPass([&] { return ModuleAnalysisManagerFunctionProxy(MAM); }); 2264 FAM.registerPass([&] { return LoopAnalysisManagerFunctionProxy(LAM); }); 2265 LAM.registerPass([&] { return FunctionAnalysisManagerLoopProxy(FAM); }); 2266 } 2267 2268 Error PassBuilder::parseModulePassPipeline(ModulePassManager &MPM, 2269 ArrayRef<PipelineElement> Pipeline, 2270 bool VerifyEachPass, 2271 bool DebugLogging) { 2272 for (const auto &Element : Pipeline) { 2273 if (auto Err = parseModulePass(MPM, Element, VerifyEachPass, DebugLogging)) 2274 return Err; 2275 if (VerifyEachPass) 2276 MPM.addPass(VerifierPass()); 2277 } 2278 return Error::success(); 2279 } 2280 2281 // Primary pass pipeline description parsing routine for a \c ModulePassManager 2282 // FIXME: Should this routine accept a TargetMachine or require the caller to 2283 // pre-populate the analysis managers with target-specific stuff? 2284 Error PassBuilder::parsePassPipeline(ModulePassManager &MPM, 2285 StringRef PipelineText, 2286 bool VerifyEachPass, bool DebugLogging) { 2287 auto Pipeline = parsePipelineText(PipelineText); 2288 if (!Pipeline || Pipeline->empty()) 2289 return make_error<StringError>( 2290 formatv("invalid pipeline '{0}'", PipelineText).str(), 2291 inconvertibleErrorCode()); 2292 2293 // If the first name isn't at the module layer, wrap the pipeline up 2294 // automatically. 2295 StringRef FirstName = Pipeline->front().Name; 2296 2297 if (!isModulePassName(FirstName, ModulePipelineParsingCallbacks)) { 2298 if (isCGSCCPassName(FirstName, CGSCCPipelineParsingCallbacks)) { 2299 Pipeline = {{"cgscc", std::move(*Pipeline)}}; 2300 } else if (isFunctionPassName(FirstName, 2301 FunctionPipelineParsingCallbacks)) { 2302 Pipeline = {{"function", std::move(*Pipeline)}}; 2303 } else if (isLoopPassName(FirstName, LoopPipelineParsingCallbacks)) { 2304 Pipeline = {{"function", {{"loop", std::move(*Pipeline)}}}}; 2305 } else { 2306 for (auto &C : TopLevelPipelineParsingCallbacks) 2307 if (C(MPM, *Pipeline, VerifyEachPass, DebugLogging)) 2308 return Error::success(); 2309 2310 // Unknown pass or pipeline name! 2311 auto &InnerPipeline = Pipeline->front().InnerPipeline; 2312 return make_error<StringError>( 2313 formatv("unknown {0} name '{1}'", 2314 (InnerPipeline.empty() ? "pass" : "pipeline"), FirstName) 2315 .str(), 2316 inconvertibleErrorCode()); 2317 } 2318 } 2319 2320 if (auto Err = 2321 parseModulePassPipeline(MPM, *Pipeline, VerifyEachPass, DebugLogging)) 2322 return Err; 2323 return Error::success(); 2324 } 2325 2326 // Primary pass pipeline description parsing routine for a \c CGSCCPassManager 2327 Error PassBuilder::parsePassPipeline(CGSCCPassManager &CGPM, 2328 StringRef PipelineText, 2329 bool VerifyEachPass, bool DebugLogging) { 2330 auto Pipeline = parsePipelineText(PipelineText); 2331 if (!Pipeline || Pipeline->empty()) 2332 return make_error<StringError>( 2333 formatv("invalid pipeline '{0}'", PipelineText).str(), 2334 inconvertibleErrorCode()); 2335 2336 StringRef FirstName = Pipeline->front().Name; 2337 if (!isCGSCCPassName(FirstName, CGSCCPipelineParsingCallbacks)) 2338 return make_error<StringError>( 2339 formatv("unknown cgscc pass '{0}' in pipeline '{1}'", FirstName, 2340 PipelineText) 2341 .str(), 2342 inconvertibleErrorCode()); 2343 2344 if (auto Err = 2345 parseCGSCCPassPipeline(CGPM, *Pipeline, VerifyEachPass, DebugLogging)) 2346 return Err; 2347 return Error::success(); 2348 } 2349 2350 // Primary pass pipeline description parsing routine for a \c 2351 // FunctionPassManager 2352 Error PassBuilder::parsePassPipeline(FunctionPassManager &FPM, 2353 StringRef PipelineText, 2354 bool VerifyEachPass, bool DebugLogging) { 2355 auto Pipeline = parsePipelineText(PipelineText); 2356 if (!Pipeline || Pipeline->empty()) 2357 return make_error<StringError>( 2358 formatv("invalid pipeline '{0}'", PipelineText).str(), 2359 inconvertibleErrorCode()); 2360 2361 StringRef FirstName = Pipeline->front().Name; 2362 if (!isFunctionPassName(FirstName, FunctionPipelineParsingCallbacks)) 2363 return make_error<StringError>( 2364 formatv("unknown function pass '{0}' in pipeline '{1}'", FirstName, 2365 PipelineText) 2366 .str(), 2367 inconvertibleErrorCode()); 2368 2369 if (auto Err = parseFunctionPassPipeline(FPM, *Pipeline, VerifyEachPass, 2370 DebugLogging)) 2371 return Err; 2372 return Error::success(); 2373 } 2374 2375 // Primary pass pipeline description parsing routine for a \c LoopPassManager 2376 Error PassBuilder::parsePassPipeline(LoopPassManager &CGPM, 2377 StringRef PipelineText, 2378 bool VerifyEachPass, bool DebugLogging) { 2379 auto Pipeline = parsePipelineText(PipelineText); 2380 if (!Pipeline || Pipeline->empty()) 2381 return make_error<StringError>( 2382 formatv("invalid pipeline '{0}'", PipelineText).str(), 2383 inconvertibleErrorCode()); 2384 2385 if (auto Err = 2386 parseLoopPassPipeline(CGPM, *Pipeline, VerifyEachPass, DebugLogging)) 2387 return Err; 2388 2389 return Error::success(); 2390 } 2391 2392 Error PassBuilder::parseAAPipeline(AAManager &AA, StringRef PipelineText) { 2393 // If the pipeline just consists of the word 'default' just replace the AA 2394 // manager with our default one. 2395 if (PipelineText == "default") { 2396 AA = buildDefaultAAPipeline(); 2397 return Error::success(); 2398 } 2399 2400 while (!PipelineText.empty()) { 2401 StringRef Name; 2402 std::tie(Name, PipelineText) = PipelineText.split(','); 2403 if (!parseAAPassName(AA, Name)) 2404 return make_error<StringError>( 2405 formatv("unknown alias analysis name '{0}'", Name).str(), 2406 inconvertibleErrorCode()); 2407 } 2408 2409 return Error::success(); 2410 } 2411