1 //===- PassManagerBuilder.cpp - Build Standard Pass -----------------------===// 2 // 3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4 // See https://llvm.org/LICENSE.txt for license information. 5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6 // 7 //===----------------------------------------------------------------------===// 8 // 9 // This file defines the PassManagerBuilder class, which is used to set up a 10 // "standard" optimization sequence suitable for languages like C and C++. 11 // 12 //===----------------------------------------------------------------------===// 13 14 #include "llvm/Transforms/IPO/PassManagerBuilder.h" 15 #include "llvm-c/Transforms/PassManagerBuilder.h" 16 #include "llvm/ADT/STLExtras.h" 17 #include "llvm/ADT/SmallVector.h" 18 #include "llvm/Analysis/BasicAliasAnalysis.h" 19 #include "llvm/Analysis/CFLAndersAliasAnalysis.h" 20 #include "llvm/Analysis/CFLSteensAliasAnalysis.h" 21 #include "llvm/Analysis/GlobalsModRef.h" 22 #include "llvm/Analysis/InlineCost.h" 23 #include "llvm/Analysis/Passes.h" 24 #include "llvm/Analysis/ScopedNoAliasAA.h" 25 #include "llvm/Analysis/TargetLibraryInfo.h" 26 #include "llvm/Analysis/TypeBasedAliasAnalysis.h" 27 #include "llvm/IR/DataLayout.h" 28 #include "llvm/IR/LegacyPassManager.h" 29 #include "llvm/IR/Verifier.h" 30 #include "llvm/Support/CommandLine.h" 31 #include "llvm/Support/ManagedStatic.h" 32 #include "llvm/Transforms/AggressiveInstCombine/AggressiveInstCombine.h" 33 #include "llvm/Transforms/IPO.h" 34 #include "llvm/Transforms/IPO/Attributor.h" 35 #include "llvm/Transforms/IPO/ForceFunctionAttrs.h" 36 #include "llvm/Transforms/IPO/FunctionAttrs.h" 37 #include "llvm/Transforms/IPO/InferFunctionAttrs.h" 38 #include "llvm/Transforms/InstCombine/InstCombine.h" 39 #include "llvm/Transforms/Instrumentation.h" 40 #include "llvm/Transforms/Scalar.h" 41 #include "llvm/Transforms/Scalar/GVN.h" 42 #include "llvm/Transforms/Scalar/InstSimplifyPass.h" 43 #include "llvm/Transforms/Scalar/LICM.h" 44 #include "llvm/Transforms/Scalar/LoopUnrollPass.h" 45 #include "llvm/Transforms/Scalar/SimpleLoopUnswitch.h" 46 #include "llvm/Transforms/Utils.h" 47 #include "llvm/Transforms/Vectorize.h" 48 #include "llvm/Transforms/Vectorize/LoopVectorize.h" 49 #include "llvm/Transforms/Vectorize/SLPVectorizer.h" 50 #include "llvm/Transforms/Vectorize/VectorCombine.h" 51 52 using namespace llvm; 53 54 static cl::opt<bool> 55 RunPartialInlining("enable-partial-inlining", cl::init(false), cl::Hidden, 56 cl::ZeroOrMore, cl::desc("Run Partial inlinining pass")); 57 58 static cl::opt<bool> 59 UseGVNAfterVectorization("use-gvn-after-vectorization", 60 cl::init(false), cl::Hidden, 61 cl::desc("Run GVN instead of Early CSE after vectorization passes")); 62 63 static cl::opt<bool> ExtraVectorizerPasses( 64 "extra-vectorizer-passes", cl::init(false), cl::Hidden, 65 cl::desc("Run cleanup optimization passes after vectorization.")); 66 67 static cl::opt<bool> 68 RunLoopRerolling("reroll-loops", cl::Hidden, 69 cl::desc("Run the loop rerolling pass")); 70 71 static cl::opt<bool> RunNewGVN("enable-newgvn", cl::init(false), cl::Hidden, 72 cl::desc("Run the NewGVN pass")); 73 74 // Experimental option to use CFL-AA 75 enum class CFLAAType { None, Steensgaard, Andersen, Both }; 76 static cl::opt<CFLAAType> 77 UseCFLAA("use-cfl-aa", cl::init(CFLAAType::None), cl::Hidden, 78 cl::desc("Enable the new, experimental CFL alias analysis"), 79 cl::values(clEnumValN(CFLAAType::None, "none", "Disable CFL-AA"), 80 clEnumValN(CFLAAType::Steensgaard, "steens", 81 "Enable unification-based CFL-AA"), 82 clEnumValN(CFLAAType::Andersen, "anders", 83 "Enable inclusion-based CFL-AA"), 84 clEnumValN(CFLAAType::Both, "both", 85 "Enable both variants of CFL-AA"))); 86 87 static cl::opt<bool> EnableLoopInterchange( 88 "enable-loopinterchange", cl::init(false), cl::Hidden, 89 cl::desc("Enable the new, experimental LoopInterchange Pass")); 90 91 static cl::opt<bool> EnableUnrollAndJam("enable-unroll-and-jam", 92 cl::init(false), cl::Hidden, 93 cl::desc("Enable Unroll And Jam Pass")); 94 95 static cl::opt<bool> 96 EnablePrepareForThinLTO("prepare-for-thinlto", cl::init(false), cl::Hidden, 97 cl::desc("Enable preparation for ThinLTO.")); 98 99 static cl::opt<bool> 100 EnablePerformThinLTO("perform-thinlto", cl::init(false), cl::Hidden, 101 cl::desc("Enable performing ThinLTO.")); 102 103 cl::opt<bool> EnableHotColdSplit("hot-cold-split", cl::init(false), 104 cl::ZeroOrMore, cl::desc("Enable hot-cold splitting pass")); 105 106 static cl::opt<bool> UseLoopVersioningLICM( 107 "enable-loop-versioning-licm", cl::init(false), cl::Hidden, 108 cl::desc("Enable the experimental Loop Versioning LICM pass")); 109 110 static cl::opt<bool> 111 DisablePreInliner("disable-preinline", cl::init(false), cl::Hidden, 112 cl::desc("Disable pre-instrumentation inliner")); 113 114 static cl::opt<int> PreInlineThreshold( 115 "preinline-threshold", cl::Hidden, cl::init(75), cl::ZeroOrMore, 116 cl::desc("Control the amount of inlining in pre-instrumentation inliner " 117 "(default = 75)")); 118 119 static cl::opt<bool> EnableGVNHoist( 120 "enable-gvn-hoist", cl::init(false), cl::ZeroOrMore, 121 cl::desc("Enable the GVN hoisting pass (default = off)")); 122 123 static cl::opt<bool> 124 DisableLibCallsShrinkWrap("disable-libcalls-shrinkwrap", cl::init(false), 125 cl::Hidden, 126 cl::desc("Disable shrink-wrap library calls")); 127 128 static cl::opt<bool> EnableSimpleLoopUnswitch( 129 "enable-simple-loop-unswitch", cl::init(false), cl::Hidden, 130 cl::desc("Enable the simple loop unswitch pass. Also enables independent " 131 "cleanup passes integrated into the loop pass manager pipeline.")); 132 133 static cl::opt<bool> EnableGVNSink( 134 "enable-gvn-sink", cl::init(false), cl::ZeroOrMore, 135 cl::desc("Enable the GVN sinking pass (default = off)")); 136 137 // This option is used in simplifying testing SampleFDO optimizations for 138 // profile loading. 139 static cl::opt<bool> 140 EnableCHR("enable-chr", cl::init(true), cl::Hidden, 141 cl::desc("Enable control height reduction optimization (CHR)")); 142 143 cl::opt<bool> FlattenedProfileUsed( 144 "flattened-profile-used", cl::init(false), cl::Hidden, 145 cl::desc("Indicate the sample profile being used is flattened, i.e., " 146 "no inline hierachy exists in the profile. ")); 147 148 cl::opt<bool> EnableOrderFileInstrumentation( 149 "enable-order-file-instrumentation", cl::init(false), cl::Hidden, 150 cl::desc("Enable order file instrumentation (default = off)")); 151 152 cl::opt<bool> EnableMatrix( 153 "enable-matrix", cl::init(false), cl::Hidden, 154 cl::desc("Enable lowering of the matrix intrinsics")); 155 156 cl::opt<bool> EnableConstraintElimination( 157 "enable-constraint-elimination", cl::init(false), cl::Hidden, 158 cl::desc( 159 "Enable pass to eliminate conditions based on linear constraints.")); 160 161 cl::opt<AttributorRunOption> AttributorRun( 162 "attributor-enable", cl::Hidden, cl::init(AttributorRunOption::NONE), 163 cl::desc("Enable the attributor inter-procedural deduction pass."), 164 cl::values(clEnumValN(AttributorRunOption::ALL, "all", 165 "enable all attributor runs"), 166 clEnumValN(AttributorRunOption::MODULE, "module", 167 "enable module-wide attributor runs"), 168 clEnumValN(AttributorRunOption::CGSCC, "cgscc", 169 "enable call graph SCC attributor runs"), 170 clEnumValN(AttributorRunOption::NONE, "none", 171 "disable attributor runs"))); 172 173 extern cl::opt<bool> EnableKnowledgeRetention; 174 175 PassManagerBuilder::PassManagerBuilder() { 176 OptLevel = 2; 177 SizeLevel = 0; 178 LibraryInfo = nullptr; 179 Inliner = nullptr; 180 DisableUnrollLoops = false; 181 SLPVectorize = false; 182 LoopVectorize = true; 183 LoopsInterleaved = true; 184 RerollLoops = RunLoopRerolling; 185 NewGVN = RunNewGVN; 186 LicmMssaOptCap = SetLicmMssaOptCap; 187 LicmMssaNoAccForPromotionCap = SetLicmMssaNoAccForPromotionCap; 188 DisableGVNLoadPRE = false; 189 ForgetAllSCEVInLoopUnroll = ForgetSCEVInLoopUnroll; 190 VerifyInput = false; 191 VerifyOutput = false; 192 MergeFunctions = false; 193 PrepareForLTO = false; 194 EnablePGOInstrGen = false; 195 EnablePGOCSInstrGen = false; 196 EnablePGOCSInstrUse = false; 197 PGOInstrGen = ""; 198 PGOInstrUse = ""; 199 PGOSampleUse = ""; 200 PrepareForThinLTO = EnablePrepareForThinLTO; 201 PerformThinLTO = EnablePerformThinLTO; 202 DivergentTarget = false; 203 CallGraphProfile = true; 204 } 205 206 PassManagerBuilder::~PassManagerBuilder() { 207 delete LibraryInfo; 208 delete Inliner; 209 } 210 211 /// Set of global extensions, automatically added as part of the standard set. 212 static ManagedStatic< 213 SmallVector<std::tuple<PassManagerBuilder::ExtensionPointTy, 214 PassManagerBuilder::ExtensionFn, 215 PassManagerBuilder::GlobalExtensionID>, 216 8>> 217 GlobalExtensions; 218 static PassManagerBuilder::GlobalExtensionID GlobalExtensionsCounter; 219 220 /// Check if GlobalExtensions is constructed and not empty. 221 /// Since GlobalExtensions is a managed static, calling 'empty()' will trigger 222 /// the construction of the object. 223 static bool GlobalExtensionsNotEmpty() { 224 return GlobalExtensions.isConstructed() && !GlobalExtensions->empty(); 225 } 226 227 PassManagerBuilder::GlobalExtensionID 228 PassManagerBuilder::addGlobalExtension(PassManagerBuilder::ExtensionPointTy Ty, 229 PassManagerBuilder::ExtensionFn Fn) { 230 auto ExtensionID = GlobalExtensionsCounter++; 231 GlobalExtensions->push_back(std::make_tuple(Ty, std::move(Fn), ExtensionID)); 232 return ExtensionID; 233 } 234 235 void PassManagerBuilder::removeGlobalExtension( 236 PassManagerBuilder::GlobalExtensionID ExtensionID) { 237 // RegisterStandardPasses may try to call this function after GlobalExtensions 238 // has already been destroyed; doing so should not generate an error. 239 if (!GlobalExtensions.isConstructed()) 240 return; 241 242 auto GlobalExtension = 243 llvm::find_if(*GlobalExtensions, [ExtensionID](const auto &elem) { 244 return std::get<2>(elem) == ExtensionID; 245 }); 246 assert(GlobalExtension != GlobalExtensions->end() && 247 "The extension ID to be removed should always be valid."); 248 249 GlobalExtensions->erase(GlobalExtension); 250 } 251 252 void PassManagerBuilder::addExtension(ExtensionPointTy Ty, ExtensionFn Fn) { 253 Extensions.push_back(std::make_pair(Ty, std::move(Fn))); 254 } 255 256 void PassManagerBuilder::addExtensionsToPM(ExtensionPointTy ETy, 257 legacy::PassManagerBase &PM) const { 258 if (GlobalExtensionsNotEmpty()) { 259 for (auto &Ext : *GlobalExtensions) { 260 if (std::get<0>(Ext) == ETy) 261 std::get<1>(Ext)(*this, PM); 262 } 263 } 264 for (unsigned i = 0, e = Extensions.size(); i != e; ++i) 265 if (Extensions[i].first == ETy) 266 Extensions[i].second(*this, PM); 267 } 268 269 void PassManagerBuilder::addInitialAliasAnalysisPasses( 270 legacy::PassManagerBase &PM) const { 271 switch (UseCFLAA) { 272 case CFLAAType::Steensgaard: 273 PM.add(createCFLSteensAAWrapperPass()); 274 break; 275 case CFLAAType::Andersen: 276 PM.add(createCFLAndersAAWrapperPass()); 277 break; 278 case CFLAAType::Both: 279 PM.add(createCFLSteensAAWrapperPass()); 280 PM.add(createCFLAndersAAWrapperPass()); 281 break; 282 default: 283 break; 284 } 285 286 // Add TypeBasedAliasAnalysis before BasicAliasAnalysis so that 287 // BasicAliasAnalysis wins if they disagree. This is intended to help 288 // support "obvious" type-punning idioms. 289 PM.add(createTypeBasedAAWrapperPass()); 290 PM.add(createScopedNoAliasAAWrapperPass()); 291 } 292 293 void PassManagerBuilder::populateFunctionPassManager( 294 legacy::FunctionPassManager &FPM) { 295 addExtensionsToPM(EP_EarlyAsPossible, FPM); 296 FPM.add(createEntryExitInstrumenterPass()); 297 298 // Add LibraryInfo if we have some. 299 if (LibraryInfo) 300 FPM.add(new TargetLibraryInfoWrapperPass(*LibraryInfo)); 301 302 // The backends do not handle matrix intrinsics currently. 303 // Make sure they are also lowered in O0. 304 // FIXME: A lightweight version of the pass should run in the backend 305 // pipeline on demand. 306 if (EnableMatrix && OptLevel == 0) 307 FPM.add(createLowerMatrixIntrinsicsMinimalPass()); 308 309 if (OptLevel == 0) return; 310 311 addInitialAliasAnalysisPasses(FPM); 312 313 FPM.add(createCFGSimplificationPass()); 314 FPM.add(createSROAPass()); 315 FPM.add(createEarlyCSEPass()); 316 FPM.add(createLowerExpectIntrinsicPass()); 317 } 318 319 // Do PGO instrumentation generation or use pass as the option specified. 320 void PassManagerBuilder::addPGOInstrPasses(legacy::PassManagerBase &MPM, 321 bool IsCS = false) { 322 if (IsCS) { 323 if (!EnablePGOCSInstrGen && !EnablePGOCSInstrUse) 324 return; 325 } else if (!EnablePGOInstrGen && PGOInstrUse.empty() && PGOSampleUse.empty()) 326 return; 327 328 // Perform the preinline and cleanup passes for O1 and above. 329 // And avoid doing them if optimizing for size. 330 // We will not do this inline for context sensitive PGO (when IsCS is true). 331 if (OptLevel > 0 && SizeLevel == 0 && !DisablePreInliner && 332 PGOSampleUse.empty() && !IsCS) { 333 // Create preinline pass. We construct an InlineParams object and specify 334 // the threshold here to avoid the command line options of the regular 335 // inliner to influence pre-inlining. The only fields of InlineParams we 336 // care about are DefaultThreshold and HintThreshold. 337 InlineParams IP; 338 IP.DefaultThreshold = PreInlineThreshold; 339 // FIXME: The hint threshold has the same value used by the regular inliner. 340 // This should probably be lowered after performance testing. 341 IP.HintThreshold = 325; 342 343 MPM.add(createFunctionInliningPass(IP)); 344 MPM.add(createSROAPass()); 345 MPM.add(createEarlyCSEPass()); // Catch trivial redundancies 346 MPM.add(createCFGSimplificationPass()); // Merge & remove BBs 347 MPM.add(createInstructionCombiningPass()); // Combine silly seq's 348 addExtensionsToPM(EP_Peephole, MPM); 349 } 350 if ((EnablePGOInstrGen && !IsCS) || (EnablePGOCSInstrGen && IsCS)) { 351 MPM.add(createPGOInstrumentationGenLegacyPass(IsCS)); 352 // Add the profile lowering pass. 353 InstrProfOptions Options; 354 if (!PGOInstrGen.empty()) 355 Options.InstrProfileOutput = PGOInstrGen; 356 Options.DoCounterPromotion = true; 357 Options.UseBFIInPromotion = IsCS; 358 MPM.add(createLoopRotatePass()); 359 MPM.add(createInstrProfilingLegacyPass(Options, IsCS)); 360 } 361 if (!PGOInstrUse.empty()) 362 MPM.add(createPGOInstrumentationUseLegacyPass(PGOInstrUse, IsCS)); 363 // Indirect call promotion that promotes intra-module targets only. 364 // For ThinLTO this is done earlier due to interactions with globalopt 365 // for imported functions. We don't run this at -O0. 366 if (OptLevel > 0 && !IsCS) 367 MPM.add( 368 createPGOIndirectCallPromotionLegacyPass(false, !PGOSampleUse.empty())); 369 } 370 void PassManagerBuilder::addFunctionSimplificationPasses( 371 legacy::PassManagerBase &MPM) { 372 // Start of function pass. 373 // Break up aggregate allocas, using SSAUpdater. 374 assert(OptLevel >= 1 && "Calling function optimizer with no optimization level!"); 375 MPM.add(createSROAPass()); 376 MPM.add(createEarlyCSEPass(true /* Enable mem-ssa. */)); // Catch trivial redundancies 377 if (EnableKnowledgeRetention) 378 MPM.add(createAssumeSimplifyPass()); 379 380 if (OptLevel > 1) { 381 if (EnableGVNHoist) 382 MPM.add(createGVNHoistPass()); 383 if (EnableGVNSink) { 384 MPM.add(createGVNSinkPass()); 385 MPM.add(createCFGSimplificationPass()); 386 } 387 } 388 389 if (EnableConstraintElimination) 390 MPM.add(createConstraintEliminationPass()); 391 392 if (OptLevel > 1) { 393 // Speculative execution if the target has divergent branches; otherwise nop. 394 MPM.add(createSpeculativeExecutionIfHasBranchDivergencePass()); 395 396 MPM.add(createJumpThreadingPass()); // Thread jumps. 397 MPM.add(createCorrelatedValuePropagationPass()); // Propagate conditionals 398 } 399 MPM.add(createCFGSimplificationPass()); // Merge & remove BBs 400 // Combine silly seq's 401 if (OptLevel > 2) 402 MPM.add(createAggressiveInstCombinerPass()); 403 MPM.add(createInstructionCombiningPass()); 404 if (SizeLevel == 0 && !DisableLibCallsShrinkWrap) 405 MPM.add(createLibCallsShrinkWrapPass()); 406 addExtensionsToPM(EP_Peephole, MPM); 407 408 // Optimize memory intrinsic calls based on the profiled size information. 409 if (SizeLevel == 0) 410 MPM.add(createPGOMemOPSizeOptLegacyPass()); 411 412 // TODO: Investigate the cost/benefit of tail call elimination on debugging. 413 if (OptLevel > 1) 414 MPM.add(createTailCallEliminationPass()); // Eliminate tail calls 415 MPM.add(createCFGSimplificationPass()); // Merge & remove BBs 416 MPM.add(createReassociatePass()); // Reassociate expressions 417 418 // Begin the loop pass pipeline. 419 if (EnableSimpleLoopUnswitch) { 420 // The simple loop unswitch pass relies on separate cleanup passes. Schedule 421 // them first so when we re-process a loop they run before other loop 422 // passes. 423 MPM.add(createLoopInstSimplifyPass()); 424 MPM.add(createLoopSimplifyCFGPass()); 425 } 426 // Rotate Loop - disable header duplication at -Oz 427 MPM.add(createLoopRotatePass(SizeLevel == 2 ? 0 : -1)); 428 // TODO: Investigate promotion cap for O1. 429 MPM.add(createLICMPass(LicmMssaOptCap, LicmMssaNoAccForPromotionCap)); 430 if (EnableSimpleLoopUnswitch) 431 MPM.add(createSimpleLoopUnswitchLegacyPass()); 432 else 433 MPM.add(createLoopUnswitchPass(SizeLevel || OptLevel < 3, DivergentTarget)); 434 // FIXME: We break the loop pass pipeline here in order to do full 435 // simplify-cfg. Eventually loop-simplifycfg should be enhanced to replace the 436 // need for this. 437 MPM.add(createCFGSimplificationPass()); 438 MPM.add(createInstructionCombiningPass()); 439 // We resume loop passes creating a second loop pipeline here. 440 MPM.add(createIndVarSimplifyPass()); // Canonicalize indvars 441 MPM.add(createLoopIdiomPass()); // Recognize idioms like memset. 442 addExtensionsToPM(EP_LateLoopOptimizations, MPM); 443 MPM.add(createLoopDeletionPass()); // Delete dead loops 444 445 if (EnableLoopInterchange) 446 MPM.add(createLoopInterchangePass()); // Interchange loops 447 448 // Unroll small loops 449 MPM.add(createSimpleLoopUnrollPass(OptLevel, DisableUnrollLoops, 450 ForgetAllSCEVInLoopUnroll)); 451 addExtensionsToPM(EP_LoopOptimizerEnd, MPM); 452 // This ends the loop pass pipelines. 453 454 if (OptLevel > 1) { 455 MPM.add(createMergedLoadStoreMotionPass()); // Merge ld/st in diamonds 456 MPM.add(NewGVN ? createNewGVNPass() 457 : createGVNPass(DisableGVNLoadPRE)); // Remove redundancies 458 } 459 MPM.add(createMemCpyOptPass()); // Remove memcpy / form memset 460 MPM.add(createSCCPPass()); // Constant prop with SCCP 461 462 // Delete dead bit computations (instcombine runs after to fold away the dead 463 // computations, and then ADCE will run later to exploit any new DCE 464 // opportunities that creates). 465 MPM.add(createBitTrackingDCEPass()); // Delete dead bit computations 466 467 // Run instcombine after redundancy elimination to exploit opportunities 468 // opened up by them. 469 MPM.add(createInstructionCombiningPass()); 470 addExtensionsToPM(EP_Peephole, MPM); 471 if (OptLevel > 1) { 472 MPM.add(createJumpThreadingPass()); // Thread jumps 473 MPM.add(createCorrelatedValuePropagationPass()); 474 MPM.add(createDeadStoreEliminationPass()); // Delete dead stores 475 MPM.add(createLICMPass(LicmMssaOptCap, LicmMssaNoAccForPromotionCap)); 476 } 477 478 addExtensionsToPM(EP_ScalarOptimizerLate, MPM); 479 480 if (RerollLoops) 481 MPM.add(createLoopRerollPass()); 482 483 // TODO: Investigate if this is too expensive at O1. 484 MPM.add(createAggressiveDCEPass()); // Delete dead instructions 485 MPM.add(createCFGSimplificationPass()); // Merge & remove BBs 486 // Clean up after everything. 487 MPM.add(createInstructionCombiningPass()); 488 addExtensionsToPM(EP_Peephole, MPM); 489 490 if (EnableCHR && OptLevel >= 3 && 491 (!PGOInstrUse.empty() || !PGOSampleUse.empty() || EnablePGOCSInstrGen)) 492 MPM.add(createControlHeightReductionLegacyPass()); 493 } 494 495 void PassManagerBuilder::populateModulePassManager( 496 legacy::PassManagerBase &MPM) { 497 // Whether this is a default or *LTO pre-link pipeline. The FullLTO post-link 498 // is handled separately, so just check this is not the ThinLTO post-link. 499 bool DefaultOrPreLinkPipeline = !PerformThinLTO; 500 501 if (!PGOSampleUse.empty()) { 502 MPM.add(createPruneEHPass()); 503 // In ThinLTO mode, when flattened profile is used, all the available 504 // profile information will be annotated in PreLink phase so there is 505 // no need to load the profile again in PostLink. 506 if (!(FlattenedProfileUsed && PerformThinLTO)) 507 MPM.add(createSampleProfileLoaderPass(PGOSampleUse)); 508 } 509 510 // Allow forcing function attributes as a debugging and tuning aid. 511 MPM.add(createForceFunctionAttrsLegacyPass()); 512 513 // If all optimizations are disabled, just run the always-inline pass and, 514 // if enabled, the function merging pass. 515 if (OptLevel == 0) { 516 addPGOInstrPasses(MPM); 517 if (Inliner) { 518 MPM.add(Inliner); 519 Inliner = nullptr; 520 } 521 522 // FIXME: The BarrierNoopPass is a HACK! The inliner pass above implicitly 523 // creates a CGSCC pass manager, but we don't want to add extensions into 524 // that pass manager. To prevent this we insert a no-op module pass to reset 525 // the pass manager to get the same behavior as EP_OptimizerLast in non-O0 526 // builds. The function merging pass is 527 if (MergeFunctions) 528 MPM.add(createMergeFunctionsPass()); 529 else if (GlobalExtensionsNotEmpty() || !Extensions.empty()) 530 MPM.add(createBarrierNoopPass()); 531 532 if (PerformThinLTO) { 533 MPM.add(createLowerTypeTestsPass(nullptr, nullptr, true)); 534 // Drop available_externally and unreferenced globals. This is necessary 535 // with ThinLTO in order to avoid leaving undefined references to dead 536 // globals in the object file. 537 MPM.add(createEliminateAvailableExternallyPass()); 538 MPM.add(createGlobalDCEPass()); 539 } 540 541 addExtensionsToPM(EP_EnabledOnOptLevel0, MPM); 542 543 if (PrepareForLTO || PrepareForThinLTO) { 544 MPM.add(createCanonicalizeAliasesPass()); 545 // Rename anon globals to be able to export them in the summary. 546 // This has to be done after we add the extensions to the pass manager 547 // as there could be passes (e.g. Adddress sanitizer) which introduce 548 // new unnamed globals. 549 MPM.add(createNameAnonGlobalPass()); 550 } 551 return; 552 } 553 554 // Add LibraryInfo if we have some. 555 if (LibraryInfo) 556 MPM.add(new TargetLibraryInfoWrapperPass(*LibraryInfo)); 557 558 addInitialAliasAnalysisPasses(MPM); 559 560 // For ThinLTO there are two passes of indirect call promotion. The 561 // first is during the compile phase when PerformThinLTO=false and 562 // intra-module indirect call targets are promoted. The second is during 563 // the ThinLTO backend when PerformThinLTO=true, when we promote imported 564 // inter-module indirect calls. For that we perform indirect call promotion 565 // earlier in the pass pipeline, here before globalopt. Otherwise imported 566 // available_externally functions look unreferenced and are removed. 567 if (PerformThinLTO) { 568 MPM.add(createPGOIndirectCallPromotionLegacyPass(/*InLTO = */ true, 569 !PGOSampleUse.empty())); 570 MPM.add(createLowerTypeTestsPass(nullptr, nullptr, true)); 571 } 572 573 // For SamplePGO in ThinLTO compile phase, we do not want to unroll loops 574 // as it will change the CFG too much to make the 2nd profile annotation 575 // in backend more difficult. 576 bool PrepareForThinLTOUsingPGOSampleProfile = 577 PrepareForThinLTO && !PGOSampleUse.empty(); 578 if (PrepareForThinLTOUsingPGOSampleProfile) 579 DisableUnrollLoops = true; 580 581 // Infer attributes about declarations if possible. 582 MPM.add(createInferFunctionAttrsLegacyPass()); 583 584 // Infer attributes on declarations, call sites, arguments, etc. 585 if (AttributorRun & AttributorRunOption::MODULE) 586 MPM.add(createAttributorLegacyPass()); 587 588 addExtensionsToPM(EP_ModuleOptimizerEarly, MPM); 589 590 if (OptLevel > 2) 591 MPM.add(createCallSiteSplittingPass()); 592 593 MPM.add(createIPSCCPPass()); // IP SCCP 594 MPM.add(createCalledValuePropagationPass()); 595 596 MPM.add(createGlobalOptimizerPass()); // Optimize out global vars 597 // Promote any localized global vars. 598 MPM.add(createPromoteMemoryToRegisterPass()); 599 600 MPM.add(createDeadArgEliminationPass()); // Dead argument elimination 601 602 MPM.add(createInstructionCombiningPass()); // Clean up after IPCP & DAE 603 addExtensionsToPM(EP_Peephole, MPM); 604 MPM.add(createCFGSimplificationPass()); // Clean up after IPCP & DAE 605 606 // For SamplePGO in ThinLTO compile phase, we do not want to do indirect 607 // call promotion as it will change the CFG too much to make the 2nd 608 // profile annotation in backend more difficult. 609 // PGO instrumentation is added during the compile phase for ThinLTO, do 610 // not run it a second time 611 if (DefaultOrPreLinkPipeline && !PrepareForThinLTOUsingPGOSampleProfile) 612 addPGOInstrPasses(MPM); 613 614 // Create profile COMDAT variables. Lld linker wants to see all variables 615 // before the LTO/ThinLTO link since it needs to resolve symbols/comdats. 616 if (!PerformThinLTO && EnablePGOCSInstrGen) 617 MPM.add(createPGOInstrumentationGenCreateVarLegacyPass(PGOInstrGen)); 618 619 // We add a module alias analysis pass here. In part due to bugs in the 620 // analysis infrastructure this "works" in that the analysis stays alive 621 // for the entire SCC pass run below. 622 MPM.add(createGlobalsAAWrapperPass()); 623 624 // Start of CallGraph SCC passes. 625 MPM.add(createPruneEHPass()); // Remove dead EH info 626 bool RunInliner = false; 627 if (Inliner) { 628 MPM.add(Inliner); 629 Inliner = nullptr; 630 RunInliner = true; 631 } 632 633 // Infer attributes on declarations, call sites, arguments, etc. for an SCC. 634 if (AttributorRun & AttributorRunOption::CGSCC) 635 MPM.add(createAttributorCGSCCLegacyPass()); 636 637 // Try to perform OpenMP specific optimizations. This is a (quick!) no-op if 638 // there are no OpenMP runtime calls present in the module. 639 if (OptLevel > 1) 640 MPM.add(createOpenMPOptLegacyPass()); 641 642 MPM.add(createPostOrderFunctionAttrsLegacyPass()); 643 if (OptLevel > 2) 644 MPM.add(createArgumentPromotionPass()); // Scalarize uninlined fn args 645 646 addExtensionsToPM(EP_CGSCCOptimizerLate, MPM); 647 addFunctionSimplificationPasses(MPM); 648 649 // FIXME: This is a HACK! The inliner pass above implicitly creates a CGSCC 650 // pass manager that we are specifically trying to avoid. To prevent this 651 // we must insert a no-op module pass to reset the pass manager. 652 MPM.add(createBarrierNoopPass()); 653 654 if (RunPartialInlining) 655 MPM.add(createPartialInliningPass()); 656 657 if (OptLevel > 1 && !PrepareForLTO && !PrepareForThinLTO) 658 // Remove avail extern fns and globals definitions if we aren't 659 // compiling an object file for later LTO. For LTO we want to preserve 660 // these so they are eligible for inlining at link-time. Note if they 661 // are unreferenced they will be removed by GlobalDCE later, so 662 // this only impacts referenced available externally globals. 663 // Eventually they will be suppressed during codegen, but eliminating 664 // here enables more opportunity for GlobalDCE as it may make 665 // globals referenced by available external functions dead 666 // and saves running remaining passes on the eliminated functions. 667 MPM.add(createEliminateAvailableExternallyPass()); 668 669 // CSFDO instrumentation and use pass. Don't invoke this for Prepare pass 670 // for LTO and ThinLTO -- The actual pass will be called after all inlines 671 // are performed. 672 // Need to do this after COMDAT variables have been eliminated, 673 // (i.e. after EliminateAvailableExternallyPass). 674 if (!(PrepareForLTO || PrepareForThinLTO)) 675 addPGOInstrPasses(MPM, /* IsCS */ true); 676 677 if (EnableOrderFileInstrumentation) 678 MPM.add(createInstrOrderFilePass()); 679 680 MPM.add(createReversePostOrderFunctionAttrsPass()); 681 682 // The inliner performs some kind of dead code elimination as it goes, 683 // but there are cases that are not really caught by it. We might 684 // at some point consider teaching the inliner about them, but it 685 // is OK for now to run GlobalOpt + GlobalDCE in tandem as their 686 // benefits generally outweight the cost, making the whole pipeline 687 // faster. 688 if (RunInliner) { 689 MPM.add(createGlobalOptimizerPass()); 690 MPM.add(createGlobalDCEPass()); 691 } 692 693 // If we are planning to perform ThinLTO later, let's not bloat the code with 694 // unrolling/vectorization/... now. We'll first run the inliner + CGSCC passes 695 // during ThinLTO and perform the rest of the optimizations afterward. 696 if (PrepareForThinLTO) { 697 // Ensure we perform any last passes, but do so before renaming anonymous 698 // globals in case the passes add any. 699 addExtensionsToPM(EP_OptimizerLast, MPM); 700 MPM.add(createCanonicalizeAliasesPass()); 701 // Rename anon globals to be able to export them in the summary. 702 MPM.add(createNameAnonGlobalPass()); 703 return; 704 } 705 706 if (PerformThinLTO) 707 // Optimize globals now when performing ThinLTO, this enables more 708 // optimizations later. 709 MPM.add(createGlobalOptimizerPass()); 710 711 // Scheduling LoopVersioningLICM when inlining is over, because after that 712 // we may see more accurate aliasing. Reason to run this late is that too 713 // early versioning may prevent further inlining due to increase of code 714 // size. By placing it just after inlining other optimizations which runs 715 // later might get benefit of no-alias assumption in clone loop. 716 if (UseLoopVersioningLICM) { 717 MPM.add(createLoopVersioningLICMPass()); // Do LoopVersioningLICM 718 MPM.add(createLICMPass(LicmMssaOptCap, LicmMssaNoAccForPromotionCap)); 719 } 720 721 // We add a fresh GlobalsModRef run at this point. This is particularly 722 // useful as the above will have inlined, DCE'ed, and function-attr 723 // propagated everything. We should at this point have a reasonably minimal 724 // and richly annotated call graph. By computing aliasing and mod/ref 725 // information for all local globals here, the late loop passes and notably 726 // the vectorizer will be able to use them to help recognize vectorizable 727 // memory operations. 728 // 729 // Note that this relies on a bug in the pass manager which preserves 730 // a module analysis into a function pass pipeline (and throughout it) so 731 // long as the first function pass doesn't invalidate the module analysis. 732 // Thus both Float2Int and LoopRotate have to preserve AliasAnalysis for 733 // this to work. Fortunately, it is trivial to preserve AliasAnalysis 734 // (doing nothing preserves it as it is required to be conservatively 735 // correct in the face of IR changes). 736 MPM.add(createGlobalsAAWrapperPass()); 737 738 MPM.add(createFloat2IntPass()); 739 MPM.add(createLowerConstantIntrinsicsPass()); 740 741 if (EnableMatrix) { 742 MPM.add(createLowerMatrixIntrinsicsPass()); 743 // CSE the pointer arithmetic of the column vectors. This allows alias 744 // analysis to establish no-aliasing between loads and stores of different 745 // columns of the same matrix. 746 MPM.add(createEarlyCSEPass(false)); 747 } 748 749 addExtensionsToPM(EP_VectorizerStart, MPM); 750 751 // Re-rotate loops in all our loop nests. These may have fallout out of 752 // rotated form due to GVN or other transformations, and the vectorizer relies 753 // on the rotated form. Disable header duplication at -Oz. 754 MPM.add(createLoopRotatePass(SizeLevel == 2 ? 0 : -1)); 755 756 // Distribute loops to allow partial vectorization. I.e. isolate dependences 757 // into separate loop that would otherwise inhibit vectorization. This is 758 // currently only performed for loops marked with the metadata 759 // llvm.loop.distribute=true or when -enable-loop-distribute is specified. 760 MPM.add(createLoopDistributePass()); 761 762 MPM.add(createLoopVectorizePass(!LoopsInterleaved, !LoopVectorize)); 763 764 // Eliminate loads by forwarding stores from the previous iteration to loads 765 // of the current iteration. 766 MPM.add(createLoopLoadEliminationPass()); 767 768 // FIXME: Because of #pragma vectorize enable, the passes below are always 769 // inserted in the pipeline, even when the vectorizer doesn't run (ex. when 770 // on -O1 and no #pragma is found). Would be good to have these two passes 771 // as function calls, so that we can only pass them when the vectorizer 772 // changed the code. 773 MPM.add(createInstructionCombiningPass()); 774 if (OptLevel > 1 && ExtraVectorizerPasses) { 775 // At higher optimization levels, try to clean up any runtime overlap and 776 // alignment checks inserted by the vectorizer. We want to track correllated 777 // runtime checks for two inner loops in the same outer loop, fold any 778 // common computations, hoist loop-invariant aspects out of any outer loop, 779 // and unswitch the runtime checks if possible. Once hoisted, we may have 780 // dead (or speculatable) control flows or more combining opportunities. 781 MPM.add(createEarlyCSEPass()); 782 MPM.add(createCorrelatedValuePropagationPass()); 783 MPM.add(createInstructionCombiningPass()); 784 MPM.add(createLICMPass(LicmMssaOptCap, LicmMssaNoAccForPromotionCap)); 785 MPM.add(createLoopUnswitchPass(SizeLevel || OptLevel < 3, DivergentTarget)); 786 MPM.add(createCFGSimplificationPass()); 787 MPM.add(createInstructionCombiningPass()); 788 } 789 790 // Cleanup after loop vectorization, etc. Simplification passes like CVP and 791 // GVN, loop transforms, and others have already run, so it's now better to 792 // convert to more optimized IR using more aggressive simplify CFG options. 793 // The extra sinking transform can create larger basic blocks, so do this 794 // before SLP vectorization. 795 // FIXME: study whether hoisting and/or sinking of common instructions should 796 // be delayed until after SLP vectorizer. 797 MPM.add(createCFGSimplificationPass(SimplifyCFGOptions() 798 .forwardSwitchCondToPhi(true) 799 .convertSwitchToLookupTable(true) 800 .needCanonicalLoops(false) 801 .hoistCommonInsts(true) 802 .sinkCommonInsts(true))); 803 804 if (SLPVectorize) { 805 MPM.add(createSLPVectorizerPass()); // Vectorize parallel scalar chains. 806 if (OptLevel > 1 && ExtraVectorizerPasses) { 807 MPM.add(createEarlyCSEPass()); 808 } 809 } 810 811 // Enhance/cleanup vector code. 812 MPM.add(createVectorCombinePass()); 813 814 addExtensionsToPM(EP_Peephole, MPM); 815 MPM.add(createInstructionCombiningPass()); 816 817 if (EnableUnrollAndJam && !DisableUnrollLoops) { 818 // Unroll and Jam. We do this before unroll but need to be in a separate 819 // loop pass manager in order for the outer loop to be processed by 820 // unroll and jam before the inner loop is unrolled. 821 MPM.add(createLoopUnrollAndJamPass(OptLevel)); 822 } 823 824 // Unroll small loops 825 MPM.add(createLoopUnrollPass(OptLevel, DisableUnrollLoops, 826 ForgetAllSCEVInLoopUnroll)); 827 828 if (!DisableUnrollLoops) { 829 // LoopUnroll may generate some redundency to cleanup. 830 MPM.add(createInstructionCombiningPass()); 831 832 // Runtime unrolling will introduce runtime check in loop prologue. If the 833 // unrolled loop is a inner loop, then the prologue will be inside the 834 // outer loop. LICM pass can help to promote the runtime check out if the 835 // checked value is loop invariant. 836 MPM.add(createLICMPass(LicmMssaOptCap, LicmMssaNoAccForPromotionCap)); 837 } 838 839 MPM.add(createWarnMissedTransformationsPass()); 840 841 // After vectorization and unrolling, assume intrinsics may tell us more 842 // about pointer alignments. 843 MPM.add(createAlignmentFromAssumptionsPass()); 844 845 // FIXME: We shouldn't bother with this anymore. 846 MPM.add(createStripDeadPrototypesPass()); // Get rid of dead prototypes 847 848 // GlobalOpt already deletes dead functions and globals, at -O2 try a 849 // late pass of GlobalDCE. It is capable of deleting dead cycles. 850 if (OptLevel > 1) { 851 MPM.add(createGlobalDCEPass()); // Remove dead fns and globals. 852 MPM.add(createConstantMergePass()); // Merge dup global constants 853 } 854 855 // See comment in the new PM for justification of scheduling splitting at 856 // this stage (\ref buildModuleSimplificationPipeline). 857 if (EnableHotColdSplit && !(PrepareForLTO || PrepareForThinLTO)) 858 MPM.add(createHotColdSplittingPass()); 859 860 if (MergeFunctions) 861 MPM.add(createMergeFunctionsPass()); 862 863 // Add Module flag "CG Profile" based on Branch Frequency Information. 864 if (CallGraphProfile) 865 MPM.add(createCGProfileLegacyPass()); 866 867 // LoopSink pass sinks instructions hoisted by LICM, which serves as a 868 // canonicalization pass that enables other optimizations. As a result, 869 // LoopSink pass needs to be a very late IR pass to avoid undoing LICM 870 // result too early. 871 MPM.add(createLoopSinkPass()); 872 // Get rid of LCSSA nodes. 873 MPM.add(createInstSimplifyLegacyPass()); 874 875 // This hoists/decomposes div/rem ops. It should run after other sink/hoist 876 // passes to avoid re-sinking, but before SimplifyCFG because it can allow 877 // flattening of blocks. 878 MPM.add(createDivRemPairsPass()); 879 880 // LoopSink (and other loop passes since the last simplifyCFG) might have 881 // resulted in single-entry-single-exit or empty blocks. Clean up the CFG. 882 MPM.add(createCFGSimplificationPass()); 883 884 addExtensionsToPM(EP_OptimizerLast, MPM); 885 886 if (PrepareForLTO) { 887 MPM.add(createCanonicalizeAliasesPass()); 888 // Rename anon globals to be able to handle them in the summary 889 MPM.add(createNameAnonGlobalPass()); 890 } 891 } 892 893 void PassManagerBuilder::addLTOOptimizationPasses(legacy::PassManagerBase &PM) { 894 // Load sample profile before running the LTO optimization pipeline. 895 if (!PGOSampleUse.empty()) { 896 PM.add(createPruneEHPass()); 897 PM.add(createSampleProfileLoaderPass(PGOSampleUse)); 898 } 899 900 // Remove unused virtual tables to improve the quality of code generated by 901 // whole-program devirtualization and bitset lowering. 902 PM.add(createGlobalDCEPass()); 903 904 // Provide AliasAnalysis services for optimizations. 905 addInitialAliasAnalysisPasses(PM); 906 907 // Allow forcing function attributes as a debugging and tuning aid. 908 PM.add(createForceFunctionAttrsLegacyPass()); 909 910 // Infer attributes about declarations if possible. 911 PM.add(createInferFunctionAttrsLegacyPass()); 912 913 if (OptLevel > 1) { 914 // Split call-site with more constrained arguments. 915 PM.add(createCallSiteSplittingPass()); 916 917 // Indirect call promotion. This should promote all the targets that are 918 // left by the earlier promotion pass that promotes intra-module targets. 919 // This two-step promotion is to save the compile time. For LTO, it should 920 // produce the same result as if we only do promotion here. 921 PM.add( 922 createPGOIndirectCallPromotionLegacyPass(true, !PGOSampleUse.empty())); 923 924 // Propagate constants at call sites into the functions they call. This 925 // opens opportunities for globalopt (and inlining) by substituting function 926 // pointers passed as arguments to direct uses of functions. 927 PM.add(createIPSCCPPass()); 928 929 // Attach metadata to indirect call sites indicating the set of functions 930 // they may target at run-time. This should follow IPSCCP. 931 PM.add(createCalledValuePropagationPass()); 932 933 // Infer attributes on declarations, call sites, arguments, etc. 934 if (AttributorRun & AttributorRunOption::MODULE) 935 PM.add(createAttributorLegacyPass()); 936 } 937 938 // Infer attributes about definitions. The readnone attribute in particular is 939 // required for virtual constant propagation. 940 PM.add(createPostOrderFunctionAttrsLegacyPass()); 941 PM.add(createReversePostOrderFunctionAttrsPass()); 942 943 // Split globals using inrange annotations on GEP indices. This can help 944 // improve the quality of generated code when virtual constant propagation or 945 // control flow integrity are enabled. 946 PM.add(createGlobalSplitPass()); 947 948 // Apply whole-program devirtualization and virtual constant propagation. 949 PM.add(createWholeProgramDevirtPass(ExportSummary, nullptr)); 950 951 // That's all we need at opt level 1. 952 if (OptLevel == 1) 953 return; 954 955 // Now that we internalized some globals, see if we can hack on them! 956 PM.add(createGlobalOptimizerPass()); 957 // Promote any localized global vars. 958 PM.add(createPromoteMemoryToRegisterPass()); 959 960 // Linking modules together can lead to duplicated global constants, only 961 // keep one copy of each constant. 962 PM.add(createConstantMergePass()); 963 964 // Remove unused arguments from functions. 965 PM.add(createDeadArgEliminationPass()); 966 967 // Reduce the code after globalopt and ipsccp. Both can open up significant 968 // simplification opportunities, and both can propagate functions through 969 // function pointers. When this happens, we often have to resolve varargs 970 // calls, etc, so let instcombine do this. 971 if (OptLevel > 2) 972 PM.add(createAggressiveInstCombinerPass()); 973 PM.add(createInstructionCombiningPass()); 974 addExtensionsToPM(EP_Peephole, PM); 975 976 // Inline small functions 977 bool RunInliner = Inliner; 978 if (RunInliner) { 979 PM.add(Inliner); 980 Inliner = nullptr; 981 } 982 983 PM.add(createPruneEHPass()); // Remove dead EH info. 984 985 // CSFDO instrumentation and use pass. 986 addPGOInstrPasses(PM, /* IsCS */ true); 987 988 // Infer attributes on declarations, call sites, arguments, etc. for an SCC. 989 if (AttributorRun & AttributorRunOption::CGSCC) 990 PM.add(createAttributorCGSCCLegacyPass()); 991 992 // Try to perform OpenMP specific optimizations. This is a (quick!) no-op if 993 // there are no OpenMP runtime calls present in the module. 994 if (OptLevel > 1) 995 PM.add(createOpenMPOptLegacyPass()); 996 997 // Optimize globals again if we ran the inliner. 998 if (RunInliner) 999 PM.add(createGlobalOptimizerPass()); 1000 PM.add(createGlobalDCEPass()); // Remove dead functions. 1001 1002 // If we didn't decide to inline a function, check to see if we can 1003 // transform it to pass arguments by value instead of by reference. 1004 PM.add(createArgumentPromotionPass()); 1005 1006 // The IPO passes may leave cruft around. Clean up after them. 1007 PM.add(createInstructionCombiningPass()); 1008 addExtensionsToPM(EP_Peephole, PM); 1009 PM.add(createJumpThreadingPass(/*FreezeSelectCond*/ true)); 1010 1011 // Break up allocas 1012 PM.add(createSROAPass()); 1013 1014 // LTO provides additional opportunities for tailcall elimination due to 1015 // link-time inlining, and visibility of nocapture attribute. 1016 if (OptLevel > 1) 1017 PM.add(createTailCallEliminationPass()); 1018 1019 // Infer attributes on declarations, call sites, arguments, etc. 1020 PM.add(createPostOrderFunctionAttrsLegacyPass()); // Add nocapture. 1021 // Run a few AA driven optimizations here and now, to cleanup the code. 1022 PM.add(createGlobalsAAWrapperPass()); // IP alias analysis. 1023 1024 PM.add(createLICMPass(LicmMssaOptCap, LicmMssaNoAccForPromotionCap)); 1025 PM.add(NewGVN ? createNewGVNPass() 1026 : createGVNPass(DisableGVNLoadPRE)); // Remove redundancies. 1027 PM.add(createMemCpyOptPass()); // Remove dead memcpys. 1028 1029 // Nuke dead stores. 1030 PM.add(createDeadStoreEliminationPass()); 1031 PM.add(createMergedLoadStoreMotionPass()); // Merge ld/st in diamonds. 1032 1033 // More loops are countable; try to optimize them. 1034 PM.add(createIndVarSimplifyPass()); 1035 PM.add(createLoopDeletionPass()); 1036 if (EnableLoopInterchange) 1037 PM.add(createLoopInterchangePass()); 1038 1039 // Unroll small loops 1040 PM.add(createSimpleLoopUnrollPass(OptLevel, DisableUnrollLoops, 1041 ForgetAllSCEVInLoopUnroll)); 1042 PM.add(createLoopVectorizePass(true, !LoopVectorize)); 1043 // The vectorizer may have significantly shortened a loop body; unroll again. 1044 PM.add(createLoopUnrollPass(OptLevel, DisableUnrollLoops, 1045 ForgetAllSCEVInLoopUnroll)); 1046 1047 PM.add(createWarnMissedTransformationsPass()); 1048 1049 // Now that we've optimized loops (in particular loop induction variables), 1050 // we may have exposed more scalar opportunities. Run parts of the scalar 1051 // optimizer again at this point. 1052 PM.add(createInstructionCombiningPass()); // Initial cleanup 1053 PM.add(createCFGSimplificationPass()); // if-convert 1054 PM.add(createSCCPPass()); // Propagate exposed constants 1055 PM.add(createInstructionCombiningPass()); // Clean up again 1056 PM.add(createBitTrackingDCEPass()); 1057 1058 // More scalar chains could be vectorized due to more alias information 1059 if (SLPVectorize) 1060 PM.add(createSLPVectorizerPass()); // Vectorize parallel scalar chains. 1061 1062 PM.add(createVectorCombinePass()); // Clean up partial vectorization. 1063 1064 // After vectorization, assume intrinsics may tell us more about pointer 1065 // alignments. 1066 PM.add(createAlignmentFromAssumptionsPass()); 1067 1068 // Cleanup and simplify the code after the scalar optimizations. 1069 PM.add(createInstructionCombiningPass()); 1070 addExtensionsToPM(EP_Peephole, PM); 1071 1072 PM.add(createJumpThreadingPass(/*FreezeSelectCond*/ true)); 1073 } 1074 1075 void PassManagerBuilder::addLateLTOOptimizationPasses( 1076 legacy::PassManagerBase &PM) { 1077 // See comment in the new PM for justification of scheduling splitting at 1078 // this stage (\ref buildLTODefaultPipeline). 1079 if (EnableHotColdSplit) 1080 PM.add(createHotColdSplittingPass()); 1081 1082 // Delete basic blocks, which optimization passes may have killed. 1083 PM.add(createCFGSimplificationPass()); 1084 1085 // Drop bodies of available externally objects to improve GlobalDCE. 1086 PM.add(createEliminateAvailableExternallyPass()); 1087 1088 // Now that we have optimized the program, discard unreachable functions. 1089 PM.add(createGlobalDCEPass()); 1090 1091 // FIXME: this is profitable (for compiler time) to do at -O0 too, but 1092 // currently it damages debug info. 1093 if (MergeFunctions) 1094 PM.add(createMergeFunctionsPass()); 1095 } 1096 1097 void PassManagerBuilder::populateThinLTOPassManager( 1098 legacy::PassManagerBase &PM) { 1099 PerformThinLTO = true; 1100 if (LibraryInfo) 1101 PM.add(new TargetLibraryInfoWrapperPass(*LibraryInfo)); 1102 1103 if (VerifyInput) 1104 PM.add(createVerifierPass()); 1105 1106 if (ImportSummary) { 1107 // This pass imports type identifier resolutions for whole-program 1108 // devirtualization and CFI. It must run early because other passes may 1109 // disturb the specific instruction patterns that these passes look for, 1110 // creating dependencies on resolutions that may not appear in the summary. 1111 // 1112 // For example, GVN may transform the pattern assume(type.test) appearing in 1113 // two basic blocks into assume(phi(type.test, type.test)), which would 1114 // transform a dependency on a WPD resolution into a dependency on a type 1115 // identifier resolution for CFI. 1116 // 1117 // Also, WPD has access to more precise information than ICP and can 1118 // devirtualize more effectively, so it should operate on the IR first. 1119 PM.add(createWholeProgramDevirtPass(nullptr, ImportSummary)); 1120 PM.add(createLowerTypeTestsPass(nullptr, ImportSummary)); 1121 } 1122 1123 populateModulePassManager(PM); 1124 1125 if (VerifyOutput) 1126 PM.add(createVerifierPass()); 1127 PerformThinLTO = false; 1128 } 1129 1130 void PassManagerBuilder::populateLTOPassManager(legacy::PassManagerBase &PM) { 1131 if (LibraryInfo) 1132 PM.add(new TargetLibraryInfoWrapperPass(*LibraryInfo)); 1133 1134 if (VerifyInput) 1135 PM.add(createVerifierPass()); 1136 1137 addExtensionsToPM(EP_FullLinkTimeOptimizationEarly, PM); 1138 1139 if (OptLevel != 0) 1140 addLTOOptimizationPasses(PM); 1141 else { 1142 // The whole-program-devirt pass needs to run at -O0 because only it knows 1143 // about the llvm.type.checked.load intrinsic: it needs to both lower the 1144 // intrinsic itself and handle it in the summary. 1145 PM.add(createWholeProgramDevirtPass(ExportSummary, nullptr)); 1146 } 1147 1148 // Create a function that performs CFI checks for cross-DSO calls with targets 1149 // in the current module. 1150 PM.add(createCrossDSOCFIPass()); 1151 1152 // Lower type metadata and the type.test intrinsic. This pass supports Clang's 1153 // control flow integrity mechanisms (-fsanitize=cfi*) and needs to run at 1154 // link time if CFI is enabled. The pass does nothing if CFI is disabled. 1155 PM.add(createLowerTypeTestsPass(ExportSummary, nullptr)); 1156 // Run a second time to clean up any type tests left behind by WPD for use 1157 // in ICP (which is performed earlier than this in the regular LTO pipeline). 1158 PM.add(createLowerTypeTestsPass(nullptr, nullptr, true)); 1159 1160 if (OptLevel != 0) 1161 addLateLTOOptimizationPasses(PM); 1162 1163 addExtensionsToPM(EP_FullLinkTimeOptimizationLast, PM); 1164 1165 if (VerifyOutput) 1166 PM.add(createVerifierPass()); 1167 } 1168 1169 LLVMPassManagerBuilderRef LLVMPassManagerBuilderCreate() { 1170 PassManagerBuilder *PMB = new PassManagerBuilder(); 1171 return wrap(PMB); 1172 } 1173 1174 void LLVMPassManagerBuilderDispose(LLVMPassManagerBuilderRef PMB) { 1175 PassManagerBuilder *Builder = unwrap(PMB); 1176 delete Builder; 1177 } 1178 1179 void 1180 LLVMPassManagerBuilderSetOptLevel(LLVMPassManagerBuilderRef PMB, 1181 unsigned OptLevel) { 1182 PassManagerBuilder *Builder = unwrap(PMB); 1183 Builder->OptLevel = OptLevel; 1184 } 1185 1186 void 1187 LLVMPassManagerBuilderSetSizeLevel(LLVMPassManagerBuilderRef PMB, 1188 unsigned SizeLevel) { 1189 PassManagerBuilder *Builder = unwrap(PMB); 1190 Builder->SizeLevel = SizeLevel; 1191 } 1192 1193 void 1194 LLVMPassManagerBuilderSetDisableUnitAtATime(LLVMPassManagerBuilderRef PMB, 1195 LLVMBool Value) { 1196 // NOTE: The DisableUnitAtATime switch has been removed. 1197 } 1198 1199 void 1200 LLVMPassManagerBuilderSetDisableUnrollLoops(LLVMPassManagerBuilderRef PMB, 1201 LLVMBool Value) { 1202 PassManagerBuilder *Builder = unwrap(PMB); 1203 Builder->DisableUnrollLoops = Value; 1204 } 1205 1206 void 1207 LLVMPassManagerBuilderSetDisableSimplifyLibCalls(LLVMPassManagerBuilderRef PMB, 1208 LLVMBool Value) { 1209 // NOTE: The simplify-libcalls pass has been removed. 1210 } 1211 1212 void 1213 LLVMPassManagerBuilderUseInlinerWithThreshold(LLVMPassManagerBuilderRef PMB, 1214 unsigned Threshold) { 1215 PassManagerBuilder *Builder = unwrap(PMB); 1216 Builder->Inliner = createFunctionInliningPass(Threshold); 1217 } 1218 1219 void 1220 LLVMPassManagerBuilderPopulateFunctionPassManager(LLVMPassManagerBuilderRef PMB, 1221 LLVMPassManagerRef PM) { 1222 PassManagerBuilder *Builder = unwrap(PMB); 1223 legacy::FunctionPassManager *FPM = unwrap<legacy::FunctionPassManager>(PM); 1224 Builder->populateFunctionPassManager(*FPM); 1225 } 1226 1227 void 1228 LLVMPassManagerBuilderPopulateModulePassManager(LLVMPassManagerBuilderRef PMB, 1229 LLVMPassManagerRef PM) { 1230 PassManagerBuilder *Builder = unwrap(PMB); 1231 legacy::PassManagerBase *MPM = unwrap(PM); 1232 Builder->populateModulePassManager(*MPM); 1233 } 1234 1235 void LLVMPassManagerBuilderPopulateLTOPassManager(LLVMPassManagerBuilderRef PMB, 1236 LLVMPassManagerRef PM, 1237 LLVMBool Internalize, 1238 LLVMBool RunInliner) { 1239 PassManagerBuilder *Builder = unwrap(PMB); 1240 legacy::PassManagerBase *LPM = unwrap(PM); 1241 1242 // A small backwards compatibility hack. populateLTOPassManager used to take 1243 // an RunInliner option. 1244 if (RunInliner && !Builder->Inliner) 1245 Builder->Inliner = createFunctionInliningPass(); 1246 1247 Builder->populateLTOPassManager(*LPM); 1248 } 1249