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