lib/Transforms/IPO/PassManagerBuilder.cpp

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