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