1 //===- Construction of pass pipelines -------------------------------------===//
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
7 //===----------------------------------------------------------------------===//
10 /// This file provides the implementation of the PassBuilder based on our
11 /// static pass registry as well as related functionality. It also provides
12 /// helpers to aid in analyzing, debugging, and testing passes and pass
15 //===----------------------------------------------------------------------===//
17 #include "llvm/Analysis/AliasAnalysis.h"
18 #include "llvm/Analysis/BasicAliasAnalysis.h"
19 #include "llvm/Analysis/CGSCCPassManager.h"
20 #include "llvm/Analysis/GlobalsModRef.h"
21 #include "llvm/Analysis/InlineAdvisor.h"
22 #include "llvm/Analysis/OptimizationRemarkEmitter.h"
23 #include "llvm/Analysis/ProfileSummaryInfo.h"
24 #include "llvm/Analysis/ScopedNoAliasAA.h"
25 #include "llvm/Analysis/TypeBasedAliasAnalysis.h"
26 #include "llvm/IR/PassManager.h"
27 #include "llvm/Passes/OptimizationLevel.h"
28 #include "llvm/Passes/PassBuilder.h"
29 #include "llvm/Support/CommandLine.h"
30 #include "llvm/Support/ErrorHandling.h"
31 #include "llvm/Support/PGOOptions.h"
32 #include "llvm/Target/TargetMachine.h"
33 #include "llvm/Transforms/AggressiveInstCombine/AggressiveInstCombine.h"
34 #include "llvm/Transforms/Coroutines/CoroCleanup.h"
35 #include "llvm/Transforms/Coroutines/CoroEarly.h"
36 #include "llvm/Transforms/Coroutines/CoroElide.h"
37 #include "llvm/Transforms/Coroutines/CoroSplit.h"
38 #include "llvm/Transforms/IPO/AlwaysInliner.h"
39 #include "llvm/Transforms/IPO/Annotation2Metadata.h"
40 #include "llvm/Transforms/IPO/ArgumentPromotion.h"
41 #include "llvm/Transforms/IPO/Attributor.h"
42 #include "llvm/Transforms/IPO/CalledValuePropagation.h"
43 #include "llvm/Transforms/IPO/ConstantMerge.h"
44 #include "llvm/Transforms/IPO/CrossDSOCFI.h"
45 #include "llvm/Transforms/IPO/DeadArgumentElimination.h"
46 #include "llvm/Transforms/IPO/ElimAvailExtern.h"
47 #include "llvm/Transforms/IPO/ForceFunctionAttrs.h"
48 #include "llvm/Transforms/IPO/FunctionAttrs.h"
49 #include "llvm/Transforms/IPO/GlobalDCE.h"
50 #include "llvm/Transforms/IPO/GlobalOpt.h"
51 #include "llvm/Transforms/IPO/GlobalSplit.h"
52 #include "llvm/Transforms/IPO/HotColdSplitting.h"
53 #include "llvm/Transforms/IPO/IROutliner.h"
54 #include "llvm/Transforms/IPO/InferFunctionAttrs.h"
55 #include "llvm/Transforms/IPO/Inliner.h"
56 #include "llvm/Transforms/IPO/LowerTypeTests.h"
57 #include "llvm/Transforms/IPO/MergeFunctions.h"
58 #include "llvm/Transforms/IPO/ModuleInliner.h"
59 #include "llvm/Transforms/IPO/OpenMPOpt.h"
60 #include "llvm/Transforms/IPO/PartialInlining.h"
61 #include "llvm/Transforms/IPO/SCCP.h"
62 #include "llvm/Transforms/IPO/SampleProfile.h"
63 #include "llvm/Transforms/IPO/SampleProfileProbe.h"
64 #include "llvm/Transforms/IPO/SyntheticCountsPropagation.h"
65 #include "llvm/Transforms/IPO/WholeProgramDevirt.h"
66 #include "llvm/Transforms/InstCombine/InstCombine.h"
67 #include "llvm/Transforms/Instrumentation/CGProfile.h"
68 #include "llvm/Transforms/Instrumentation/ControlHeightReduction.h"
69 #include "llvm/Transforms/Instrumentation/InstrOrderFile.h"
70 #include "llvm/Transforms/Instrumentation/InstrProfiling.h"
71 #include "llvm/Transforms/Instrumentation/MemProfiler.h"
72 #include "llvm/Transforms/Instrumentation/PGOInstrumentation.h"
73 #include "llvm/Transforms/Scalar/ADCE.h"
74 #include "llvm/Transforms/Scalar/AlignmentFromAssumptions.h"
75 #include "llvm/Transforms/Scalar/AnnotationRemarks.h"
76 #include "llvm/Transforms/Scalar/BDCE.h"
77 #include "llvm/Transforms/Scalar/CallSiteSplitting.h"
78 #include "llvm/Transforms/Scalar/ConstraintElimination.h"
79 #include "llvm/Transforms/Scalar/CorrelatedValuePropagation.h"
80 #include "llvm/Transforms/Scalar/DFAJumpThreading.h"
81 #include "llvm/Transforms/Scalar/DeadStoreElimination.h"
82 #include "llvm/Transforms/Scalar/DivRemPairs.h"
83 #include "llvm/Transforms/Scalar/EarlyCSE.h"
84 #include "llvm/Transforms/Scalar/Float2Int.h"
85 #include "llvm/Transforms/Scalar/GVN.h"
86 #include "llvm/Transforms/Scalar/IndVarSimplify.h"
87 #include "llvm/Transforms/Scalar/InstSimplifyPass.h"
88 #include "llvm/Transforms/Scalar/JumpThreading.h"
89 #include "llvm/Transforms/Scalar/LICM.h"
90 #include "llvm/Transforms/Scalar/LoopDeletion.h"
91 #include "llvm/Transforms/Scalar/LoopDistribute.h"
92 #include "llvm/Transforms/Scalar/LoopFlatten.h"
93 #include "llvm/Transforms/Scalar/LoopIdiomRecognize.h"
94 #include "llvm/Transforms/Scalar/LoopInstSimplify.h"
95 #include "llvm/Transforms/Scalar/LoopInterchange.h"
96 #include "llvm/Transforms/Scalar/LoopLoadElimination.h"
97 #include "llvm/Transforms/Scalar/LoopPassManager.h"
98 #include "llvm/Transforms/Scalar/LoopRotation.h"
99 #include "llvm/Transforms/Scalar/LoopSimplifyCFG.h"
100 #include "llvm/Transforms/Scalar/LoopSink.h"
101 #include "llvm/Transforms/Scalar/LoopUnrollAndJamPass.h"
102 #include "llvm/Transforms/Scalar/LoopUnrollPass.h"
103 #include "llvm/Transforms/Scalar/LowerConstantIntrinsics.h"
104 #include "llvm/Transforms/Scalar/LowerExpectIntrinsic.h"
105 #include "llvm/Transforms/Scalar/LowerMatrixIntrinsics.h"
106 #include "llvm/Transforms/Scalar/MemCpyOptimizer.h"
107 #include "llvm/Transforms/Scalar/MergedLoadStoreMotion.h"
108 #include "llvm/Transforms/Scalar/NewGVN.h"
109 #include "llvm/Transforms/Scalar/Reassociate.h"
110 #include "llvm/Transforms/Scalar/SCCP.h"
111 #include "llvm/Transforms/Scalar/SROA.h"
112 #include "llvm/Transforms/Scalar/SimpleLoopUnswitch.h"
113 #include "llvm/Transforms/Scalar/SimplifyCFG.h"
114 #include "llvm/Transforms/Scalar/SpeculativeExecution.h"
115 #include "llvm/Transforms/Scalar/TailRecursionElimination.h"
116 #include "llvm/Transforms/Scalar/WarnMissedTransforms.h"
117 #include "llvm/Transforms/Utils/AddDiscriminators.h"
118 #include "llvm/Transforms/Utils/AssumeBundleBuilder.h"
119 #include "llvm/Transforms/Utils/CanonicalizeAliases.h"
120 #include "llvm/Transforms/Utils/InjectTLIMappings.h"
121 #include "llvm/Transforms/Utils/LibCallsShrinkWrap.h"
122 #include "llvm/Transforms/Utils/Mem2Reg.h"
123 #include "llvm/Transforms/Utils/NameAnonGlobals.h"
124 #include "llvm/Transforms/Utils/RelLookupTableConverter.h"
125 #include "llvm/Transforms/Utils/SimplifyCFGOptions.h"
126 #include "llvm/Transforms/Vectorize/LoopVectorize.h"
127 #include "llvm/Transforms/Vectorize/SLPVectorizer.h"
128 #include "llvm/Transforms/Vectorize/VectorCombine.h"
130 using namespace llvm;
132 static cl::opt<InliningAdvisorMode> UseInlineAdvisor(
133 "enable-ml-inliner", cl::init(InliningAdvisorMode::Default), cl::Hidden,
134 cl::desc("Enable ML policy for inliner. Currently trained for -Oz only"),
135 cl::values(clEnumValN(InliningAdvisorMode::Default, "default",
136 "Heuristics-based inliner version."),
137 clEnumValN(InliningAdvisorMode::Development, "development",
138 "Use development mode (runtime-loadable model)."),
139 clEnumValN(InliningAdvisorMode::Release, "release",
140 "Use release mode (AOT-compiled model).")));
142 static cl::opt<bool> EnableSyntheticCounts(
143 "enable-npm-synthetic-counts", cl::init(false), cl::Hidden, cl::ZeroOrMore,
144 cl::desc("Run synthetic function entry count generation "
147 /// Flag to enable inline deferral during PGO.
149 EnablePGOInlineDeferral("enable-npm-pgo-inline-deferral", cl::init(true),
151 cl::desc("Enable inline deferral during PGO"));
153 static cl::opt<bool> EnableMemProfiler("enable-mem-prof", cl::init(false),
154 cl::Hidden, cl::ZeroOrMore,
155 cl::desc("Enable memory profiler"));
157 static cl::opt<bool> EnableModuleInliner("enable-module-inliner",
158 cl::init(false), cl::Hidden,
159 cl::desc("Enable module inliner"));
161 static cl::opt<bool> PerformMandatoryInliningsFirst(
162 "mandatory-inlining-first", cl::init(true), cl::Hidden, cl::ZeroOrMore,
163 cl::desc("Perform mandatory inlinings module-wide, before performing "
166 static cl::opt<bool> EnableO3NonTrivialUnswitching(
167 "enable-npm-O3-nontrivial-unswitch", cl::init(true), cl::Hidden,
168 cl::ZeroOrMore, cl::desc("Enable non-trivial loop unswitching for -O3"));
170 static cl::opt<bool> EnableEagerlyInvalidateAnalyses(
171 "eagerly-invalidate-analyses", cl::init(true), cl::Hidden,
172 cl::desc("Eagerly invalidate more analyses in default pipelines"));
174 static cl::opt<bool> EnableNoRerunSimplificationPipeline(
175 "enable-no-rerun-simplification-pipeline", cl::init(false), cl::Hidden,
177 "Prevent running the simplification pipeline on a function more "
178 "than once in the case that SCC mutations cause a function to be "
179 "visited multiple times as long as the function has not been changed"));
181 static cl::opt<bool> EnableMergeFunctions(
182 "enable-merge-functions", cl::init(false), cl::Hidden,
183 cl::desc("Enable function merging as part of the optimization pipeline"));
185 PipelineTuningOptions::PipelineTuningOptions() {
186 LoopInterleaving = true;
187 LoopVectorization = true;
188 SLPVectorization = false;
189 LoopUnrolling = true;
190 ForgetAllSCEVInLoopUnroll = ForgetSCEVInLoopUnroll;
191 LicmMssaOptCap = SetLicmMssaOptCap;
192 LicmMssaNoAccForPromotionCap = SetLicmMssaNoAccForPromotionCap;
193 CallGraphProfile = true;
194 MergeFunctions = EnableMergeFunctions;
195 EagerlyInvalidateAnalyses = EnableEagerlyInvalidateAnalyses;
200 extern cl::opt<unsigned> MaxDevirtIterations;
201 extern cl::opt<bool> EnableConstraintElimination;
202 extern cl::opt<bool> EnableFunctionSpecialization;
203 extern cl::opt<bool> EnableGVNHoist;
204 extern cl::opt<bool> EnableGVNSink;
205 extern cl::opt<bool> EnableHotColdSplit;
206 extern cl::opt<bool> EnableIROutliner;
207 extern cl::opt<bool> EnableOrderFileInstrumentation;
208 extern cl::opt<bool> EnableCHR;
209 extern cl::opt<bool> EnableLoopInterchange;
210 extern cl::opt<bool> EnableUnrollAndJam;
211 extern cl::opt<bool> EnableLoopFlatten;
212 extern cl::opt<bool> EnableDFAJumpThreading;
213 extern cl::opt<bool> RunNewGVN;
214 extern cl::opt<bool> RunPartialInlining;
215 extern cl::opt<bool> ExtraVectorizerPasses;
217 extern cl::opt<bool> FlattenedProfileUsed;
219 extern cl::opt<AttributorRunOption> AttributorRun;
220 extern cl::opt<bool> EnableKnowledgeRetention;
222 extern cl::opt<bool> EnableMatrix;
224 extern cl::opt<bool> DisablePreInliner;
225 extern cl::opt<int> PreInlineThreshold;
228 void PassBuilder::invokePeepholeEPCallbacks(FunctionPassManager &FPM,
229 OptimizationLevel Level) {
230 for (auto &C : PeepholeEPCallbacks)
234 // Helper to add AnnotationRemarksPass.
235 static void addAnnotationRemarksPass(ModulePassManager &MPM) {
236 FunctionPassManager FPM;
237 FPM.addPass(AnnotationRemarksPass());
238 MPM.addPass(createModuleToFunctionPassAdaptor(std::move(FPM)));
241 // Helper to check if the current compilation phase is preparing for LTO
242 static bool isLTOPreLink(ThinOrFullLTOPhase Phase) {
243 return Phase == ThinOrFullLTOPhase::ThinLTOPreLink ||
244 Phase == ThinOrFullLTOPhase::FullLTOPreLink;
247 // TODO: Investigate the cost/benefit of tail call elimination on debugging.
249 PassBuilder::buildO1FunctionSimplificationPipeline(OptimizationLevel Level,
250 ThinOrFullLTOPhase Phase) {
252 FunctionPassManager FPM;
254 // Form SSA out of local memory accesses after breaking apart aggregates into
256 FPM.addPass(SROAPass());
258 // Catch trivial redundancies
259 FPM.addPass(EarlyCSEPass(true /* Enable mem-ssa. */));
261 // Hoisting of scalars and load expressions.
263 SimplifyCFGPass(SimplifyCFGOptions().convertSwitchRangeToICmp(true)));
264 FPM.addPass(InstCombinePass());
266 FPM.addPass(LibCallsShrinkWrapPass());
268 invokePeepholeEPCallbacks(FPM, Level);
271 SimplifyCFGPass(SimplifyCFGOptions().convertSwitchRangeToICmp(true)));
273 // Form canonically associated expression trees, and simplify the trees using
274 // basic mathematical properties. For example, this will form (nearly)
275 // minimal multiplication trees.
276 FPM.addPass(ReassociatePass());
278 // Add the primary loop simplification pipeline.
279 // FIXME: Currently this is split into two loop pass pipelines because we run
280 // some function passes in between them. These can and should be removed
281 // and/or replaced by scheduling the loop pass equivalents in the correct
282 // positions. But those equivalent passes aren't powerful enough yet.
283 // Specifically, `SimplifyCFGPass` and `InstCombinePass` are currently still
284 // used. We have `LoopSimplifyCFGPass` which isn't yet powerful enough yet to
285 // fully replace `SimplifyCFGPass`, and the closest to the other we have is
286 // `LoopInstSimplify`.
287 LoopPassManager LPM1, LPM2;
289 // Simplify the loop body. We do this initially to clean up after other loop
290 // passes run, either when iterating on a loop or on inner loops with
291 // implications on the outer loop.
292 LPM1.addPass(LoopInstSimplifyPass());
293 LPM1.addPass(LoopSimplifyCFGPass());
295 // Try to remove as much code from the loop header as possible,
296 // to reduce amount of IR that will have to be duplicated. However,
297 // do not perform speculative hoisting the first time as LICM
298 // will destroy metadata that may not need to be destroyed if run
299 // after loop rotation.
300 // TODO: Investigate promotion cap for O1.
301 LPM1.addPass(LICMPass(PTO.LicmMssaOptCap, PTO.LicmMssaNoAccForPromotionCap,
302 /*AllowSpeculation=*/false));
304 LPM1.addPass(LoopRotatePass(/* Disable header duplication */ true,
305 isLTOPreLink(Phase)));
306 // TODO: Investigate promotion cap for O1.
307 LPM1.addPass(LICMPass(PTO.LicmMssaOptCap, PTO.LicmMssaNoAccForPromotionCap,
308 /*AllowSpeculation=*/true));
309 LPM1.addPass(SimpleLoopUnswitchPass());
310 if (EnableLoopFlatten)
311 LPM1.addPass(LoopFlattenPass());
313 LPM2.addPass(LoopIdiomRecognizePass());
314 LPM2.addPass(IndVarSimplifyPass());
316 for (auto &C : LateLoopOptimizationsEPCallbacks)
319 LPM2.addPass(LoopDeletionPass());
321 if (EnableLoopInterchange)
322 LPM2.addPass(LoopInterchangePass());
324 // Do not enable unrolling in PreLinkThinLTO phase during sample PGO
325 // because it changes IR to makes profile annotation in back compile
326 // inaccurate. The normal unroller doesn't pay attention to forced full unroll
327 // attributes so we need to make sure and allow the full unroll pass to pay
329 if (Phase != ThinOrFullLTOPhase::ThinLTOPreLink || !PGOOpt ||
330 PGOOpt->Action != PGOOptions::SampleUse)
331 LPM2.addPass(LoopFullUnrollPass(Level.getSpeedupLevel(),
332 /* OnlyWhenForced= */ !PTO.LoopUnrolling,
333 PTO.ForgetAllSCEVInLoopUnroll));
335 for (auto &C : LoopOptimizerEndEPCallbacks)
338 // We provide the opt remark emitter pass for LICM to use. We only need to do
339 // this once as it is immutable.
341 RequireAnalysisPass<OptimizationRemarkEmitterAnalysis, Function>());
342 FPM.addPass(createFunctionToLoopPassAdaptor(std::move(LPM1),
343 /*UseMemorySSA=*/true,
344 /*UseBlockFrequencyInfo=*/true));
346 SimplifyCFGPass(SimplifyCFGOptions().convertSwitchRangeToICmp(true)));
347 FPM.addPass(InstCombinePass());
348 // The loop passes in LPM2 (LoopFullUnrollPass) do not preserve MemorySSA.
349 // *All* loop passes must preserve it, in order to be able to use it.
350 FPM.addPass(createFunctionToLoopPassAdaptor(std::move(LPM2),
351 /*UseMemorySSA=*/false,
352 /*UseBlockFrequencyInfo=*/false));
354 // Delete small array after loop unroll.
355 FPM.addPass(SROAPass());
357 // Specially optimize memory movement as it doesn't look like dataflow in SSA.
358 FPM.addPass(MemCpyOptPass());
360 // Sparse conditional constant propagation.
361 // FIXME: It isn't clear why we do this *after* loop passes rather than
363 FPM.addPass(SCCPPass());
365 // Delete dead bit computations (instcombine runs after to fold away the dead
366 // computations, and then ADCE will run later to exploit any new DCE
367 // opportunities that creates).
368 FPM.addPass(BDCEPass());
370 // Run instcombine after redundancy and dead bit elimination to exploit
371 // opportunities opened up by them.
372 FPM.addPass(InstCombinePass());
373 invokePeepholeEPCallbacks(FPM, Level);
375 FPM.addPass(CoroElidePass());
377 for (auto &C : ScalarOptimizerLateEPCallbacks)
380 // Finally, do an expensive DCE pass to catch all the dead code exposed by
381 // the simplifications and basic cleanup after all the simplifications.
382 // TODO: Investigate if this is too expensive.
383 FPM.addPass(ADCEPass());
385 SimplifyCFGPass(SimplifyCFGOptions().convertSwitchRangeToICmp(true)));
386 FPM.addPass(InstCombinePass());
387 invokePeepholeEPCallbacks(FPM, Level);
393 PassBuilder::buildFunctionSimplificationPipeline(OptimizationLevel Level,
394 ThinOrFullLTOPhase Phase) {
395 assert(Level != OptimizationLevel::O0 && "Must request optimizations!");
397 // The O1 pipeline has a separate pipeline creation function to simplify
398 // construction readability.
399 if (Level.getSpeedupLevel() == 1)
400 return buildO1FunctionSimplificationPipeline(Level, Phase);
402 FunctionPassManager FPM;
404 // Form SSA out of local memory accesses after breaking apart aggregates into
406 FPM.addPass(SROAPass());
408 // Catch trivial redundancies
409 FPM.addPass(EarlyCSEPass(true /* Enable mem-ssa. */));
410 if (EnableKnowledgeRetention)
411 FPM.addPass(AssumeSimplifyPass());
413 // Hoisting of scalars and load expressions.
415 FPM.addPass(GVNHoistPass());
417 // Global value numbering based sinking.
419 FPM.addPass(GVNSinkPass());
421 SimplifyCFGPass(SimplifyCFGOptions().convertSwitchRangeToICmp(true)));
424 if (EnableConstraintElimination)
425 FPM.addPass(ConstraintEliminationPass());
427 // Speculative execution if the target has divergent branches; otherwise nop.
428 FPM.addPass(SpeculativeExecutionPass(/* OnlyIfDivergentTarget =*/true));
430 // Optimize based on known information about branches, and cleanup afterward.
431 FPM.addPass(JumpThreadingPass());
432 FPM.addPass(CorrelatedValuePropagationPass());
435 SimplifyCFGPass(SimplifyCFGOptions().convertSwitchRangeToICmp(true)));
436 FPM.addPass(InstCombinePass());
437 if (Level == OptimizationLevel::O3)
438 FPM.addPass(AggressiveInstCombinePass());
440 if (!Level.isOptimizingForSize())
441 FPM.addPass(LibCallsShrinkWrapPass());
443 invokePeepholeEPCallbacks(FPM, Level);
445 // For PGO use pipeline, try to optimize memory intrinsics such as memcpy
446 // using the size value profile. Don't perform this when optimizing for size.
447 if (PGOOpt && PGOOpt->Action == PGOOptions::IRUse &&
448 !Level.isOptimizingForSize())
449 FPM.addPass(PGOMemOPSizeOpt());
451 FPM.addPass(TailCallElimPass());
453 SimplifyCFGPass(SimplifyCFGOptions().convertSwitchRangeToICmp(true)));
455 // Form canonically associated expression trees, and simplify the trees using
456 // basic mathematical properties. For example, this will form (nearly)
457 // minimal multiplication trees.
458 FPM.addPass(ReassociatePass());
460 // Add the primary loop simplification pipeline.
461 // FIXME: Currently this is split into two loop pass pipelines because we run
462 // some function passes in between them. These can and should be removed
463 // and/or replaced by scheduling the loop pass equivalents in the correct
464 // positions. But those equivalent passes aren't powerful enough yet.
465 // Specifically, `SimplifyCFGPass` and `InstCombinePass` are currently still
466 // used. We have `LoopSimplifyCFGPass` which isn't yet powerful enough yet to
467 // fully replace `SimplifyCFGPass`, and the closest to the other we have is
468 // `LoopInstSimplify`.
469 LoopPassManager LPM1, LPM2;
471 // Simplify the loop body. We do this initially to clean up after other loop
472 // passes run, either when iterating on a loop or on inner loops with
473 // implications on the outer loop.
474 LPM1.addPass(LoopInstSimplifyPass());
475 LPM1.addPass(LoopSimplifyCFGPass());
477 // Try to remove as much code from the loop header as possible,
478 // to reduce amount of IR that will have to be duplicated. However,
479 // do not perform speculative hoisting the first time as LICM
480 // will destroy metadata that may not need to be destroyed if run
481 // after loop rotation.
482 // TODO: Investigate promotion cap for O1.
483 LPM1.addPass(LICMPass(PTO.LicmMssaOptCap, PTO.LicmMssaNoAccForPromotionCap,
484 /*AllowSpeculation=*/false));
486 // Disable header duplication in loop rotation at -Oz.
488 LoopRotatePass(Level != OptimizationLevel::Oz, isLTOPreLink(Phase)));
489 // TODO: Investigate promotion cap for O1.
490 LPM1.addPass(LICMPass(PTO.LicmMssaOptCap, PTO.LicmMssaNoAccForPromotionCap,
491 /*AllowSpeculation=*/true));
493 SimpleLoopUnswitchPass(/* NonTrivial */ Level == OptimizationLevel::O3 &&
494 EnableO3NonTrivialUnswitching));
495 if (EnableLoopFlatten)
496 LPM1.addPass(LoopFlattenPass());
498 LPM2.addPass(LoopIdiomRecognizePass());
499 LPM2.addPass(IndVarSimplifyPass());
501 for (auto &C : LateLoopOptimizationsEPCallbacks)
504 LPM2.addPass(LoopDeletionPass());
506 if (EnableLoopInterchange)
507 LPM2.addPass(LoopInterchangePass());
509 // Do not enable unrolling in PreLinkThinLTO phase during sample PGO
510 // because it changes IR to makes profile annotation in back compile
511 // inaccurate. The normal unroller doesn't pay attention to forced full unroll
512 // attributes so we need to make sure and allow the full unroll pass to pay
514 if (Phase != ThinOrFullLTOPhase::ThinLTOPreLink || !PGOOpt ||
515 PGOOpt->Action != PGOOptions::SampleUse)
516 LPM2.addPass(LoopFullUnrollPass(Level.getSpeedupLevel(),
517 /* OnlyWhenForced= */ !PTO.LoopUnrolling,
518 PTO.ForgetAllSCEVInLoopUnroll));
520 for (auto &C : LoopOptimizerEndEPCallbacks)
523 // We provide the opt remark emitter pass for LICM to use. We only need to do
524 // this once as it is immutable.
526 RequireAnalysisPass<OptimizationRemarkEmitterAnalysis, Function>());
527 FPM.addPass(createFunctionToLoopPassAdaptor(std::move(LPM1),
528 /*UseMemorySSA=*/true,
529 /*UseBlockFrequencyInfo=*/true));
531 SimplifyCFGPass(SimplifyCFGOptions().convertSwitchRangeToICmp(true)));
532 FPM.addPass(InstCombinePass());
533 // The loop passes in LPM2 (LoopIdiomRecognizePass, IndVarSimplifyPass,
534 // LoopDeletionPass and LoopFullUnrollPass) do not preserve MemorySSA.
535 // *All* loop passes must preserve it, in order to be able to use it.
536 FPM.addPass(createFunctionToLoopPassAdaptor(std::move(LPM2),
537 /*UseMemorySSA=*/false,
538 /*UseBlockFrequencyInfo=*/false));
540 // Delete small array after loop unroll.
541 FPM.addPass(SROAPass());
543 // The matrix extension can introduce large vector operations early, which can
544 // benefit from running vector-combine early on.
546 FPM.addPass(VectorCombinePass(/*ScalarizationOnly=*/true));
548 // Eliminate redundancies.
549 FPM.addPass(MergedLoadStoreMotionPass());
551 FPM.addPass(NewGVNPass());
553 FPM.addPass(GVNPass());
555 // Sparse conditional constant propagation.
556 // FIXME: It isn't clear why we do this *after* loop passes rather than
558 FPM.addPass(SCCPPass());
560 // Delete dead bit computations (instcombine runs after to fold away the dead
561 // computations, and then ADCE will run later to exploit any new DCE
562 // opportunities that creates).
563 FPM.addPass(BDCEPass());
565 // Run instcombine after redundancy and dead bit elimination to exploit
566 // opportunities opened up by them.
567 FPM.addPass(InstCombinePass());
568 invokePeepholeEPCallbacks(FPM, Level);
570 // Re-consider control flow based optimizations after redundancy elimination,
572 if (EnableDFAJumpThreading && Level.getSizeLevel() == 0)
573 FPM.addPass(DFAJumpThreadingPass());
575 FPM.addPass(JumpThreadingPass());
576 FPM.addPass(CorrelatedValuePropagationPass());
578 // Finally, do an expensive DCE pass to catch all the dead code exposed by
579 // the simplifications and basic cleanup after all the simplifications.
580 // TODO: Investigate if this is too expensive.
581 FPM.addPass(ADCEPass());
583 // Specially optimize memory movement as it doesn't look like dataflow in SSA.
584 FPM.addPass(MemCpyOptPass());
586 FPM.addPass(DSEPass());
587 FPM.addPass(createFunctionToLoopPassAdaptor(
588 LICMPass(PTO.LicmMssaOptCap, PTO.LicmMssaNoAccForPromotionCap,
589 /*AllowSpeculation=*/true),
590 /*UseMemorySSA=*/true, /*UseBlockFrequencyInfo=*/true));
592 FPM.addPass(CoroElidePass());
594 for (auto &C : ScalarOptimizerLateEPCallbacks)
597 FPM.addPass(SimplifyCFGPass(SimplifyCFGOptions()
598 .convertSwitchRangeToICmp(true)
599 .hoistCommonInsts(true)
600 .sinkCommonInsts(true)));
601 FPM.addPass(InstCombinePass());
602 invokePeepholeEPCallbacks(FPM, Level);
604 if (EnableCHR && Level == OptimizationLevel::O3 && PGOOpt &&
605 (PGOOpt->Action == PGOOptions::IRUse ||
606 PGOOpt->Action == PGOOptions::SampleUse))
607 FPM.addPass(ControlHeightReductionPass());
612 void PassBuilder::addRequiredLTOPreLinkPasses(ModulePassManager &MPM) {
613 MPM.addPass(CanonicalizeAliasesPass());
614 MPM.addPass(NameAnonGlobalPass());
617 void PassBuilder::addPGOInstrPasses(ModulePassManager &MPM,
618 OptimizationLevel Level, bool RunProfileGen,
619 bool IsCS, std::string ProfileFile,
620 std::string ProfileRemappingFile) {
621 assert(Level != OptimizationLevel::O0 && "Not expecting O0 here!");
622 if (!IsCS && !DisablePreInliner) {
625 IP.DefaultThreshold = PreInlineThreshold;
627 // FIXME: The hint threshold has the same value used by the regular inliner
628 // when not optimzing for size. This should probably be lowered after
629 // performance testing.
630 // FIXME: this comment is cargo culted from the old pass manager, revisit).
631 IP.HintThreshold = Level.isOptimizingForSize() ? PreInlineThreshold : 325;
632 ModuleInlinerWrapperPass MIWP(IP);
633 CGSCCPassManager &CGPipeline = MIWP.getPM();
635 FunctionPassManager FPM;
636 FPM.addPass(SROAPass());
637 FPM.addPass(EarlyCSEPass()); // Catch trivial redundancies.
638 FPM.addPass(SimplifyCFGPass(SimplifyCFGOptions().convertSwitchRangeToICmp(
639 true))); // Merge & remove basic blocks.
640 FPM.addPass(InstCombinePass()); // Combine silly sequences.
641 invokePeepholeEPCallbacks(FPM, Level);
643 CGPipeline.addPass(createCGSCCToFunctionPassAdaptor(
644 std::move(FPM), PTO.EagerlyInvalidateAnalyses));
646 MPM.addPass(std::move(MIWP));
648 // Delete anything that is now dead to make sure that we don't instrument
649 // dead code. Instrumentation can end up keeping dead code around and
650 // dramatically increase code size.
651 MPM.addPass(GlobalDCEPass());
654 if (!RunProfileGen) {
655 assert(!ProfileFile.empty() && "Profile use expecting a profile file!");
656 MPM.addPass(PGOInstrumentationUse(ProfileFile, ProfileRemappingFile, IsCS));
657 // Cache ProfileSummaryAnalysis once to avoid the potential need to insert
658 // RequireAnalysisPass for PSI before subsequent non-module passes.
659 MPM.addPass(RequireAnalysisPass<ProfileSummaryAnalysis, Module>());
663 // Perform PGO instrumentation.
664 MPM.addPass(PGOInstrumentationGen(IsCS));
666 FunctionPassManager FPM;
667 // Disable header duplication in loop rotation at -Oz.
668 FPM.addPass(createFunctionToLoopPassAdaptor(
669 LoopRotatePass(Level != OptimizationLevel::Oz), /*UseMemorySSA=*/false,
670 /*UseBlockFrequencyInfo=*/false));
671 MPM.addPass(createModuleToFunctionPassAdaptor(std::move(FPM),
672 PTO.EagerlyInvalidateAnalyses));
674 // Add the profile lowering pass.
675 InstrProfOptions Options;
676 if (!ProfileFile.empty())
677 Options.InstrProfileOutput = ProfileFile;
678 // Do counter promotion at Level greater than O0.
679 Options.DoCounterPromotion = true;
680 Options.UseBFIInPromotion = IsCS;
681 MPM.addPass(InstrProfiling(Options, IsCS));
684 void PassBuilder::addPGOInstrPassesForO0(ModulePassManager &MPM,
685 bool RunProfileGen, bool IsCS,
686 std::string ProfileFile,
687 std::string ProfileRemappingFile) {
688 if (!RunProfileGen) {
689 assert(!ProfileFile.empty() && "Profile use expecting a profile file!");
690 MPM.addPass(PGOInstrumentationUse(ProfileFile, ProfileRemappingFile, IsCS));
691 // Cache ProfileSummaryAnalysis once to avoid the potential need to insert
692 // RequireAnalysisPass for PSI before subsequent non-module passes.
693 MPM.addPass(RequireAnalysisPass<ProfileSummaryAnalysis, Module>());
697 // Perform PGO instrumentation.
698 MPM.addPass(PGOInstrumentationGen(IsCS));
699 // Add the profile lowering pass.
700 InstrProfOptions Options;
701 if (!ProfileFile.empty())
702 Options.InstrProfileOutput = ProfileFile;
703 // Do not do counter promotion at O0.
704 Options.DoCounterPromotion = false;
705 Options.UseBFIInPromotion = IsCS;
706 MPM.addPass(InstrProfiling(Options, IsCS));
709 static InlineParams getInlineParamsFromOptLevel(OptimizationLevel Level) {
710 return getInlineParams(Level.getSpeedupLevel(), Level.getSizeLevel());
713 ModuleInlinerWrapperPass
714 PassBuilder::buildInlinerPipeline(OptimizationLevel Level,
715 ThinOrFullLTOPhase Phase) {
716 InlineParams IP = getInlineParamsFromOptLevel(Level);
717 if (Phase == ThinOrFullLTOPhase::ThinLTOPreLink && PGOOpt &&
718 PGOOpt->Action == PGOOptions::SampleUse)
719 IP.HotCallSiteThreshold = 0;
722 IP.EnableDeferral = EnablePGOInlineDeferral;
724 ModuleInlinerWrapperPass MIWP(IP, PerformMandatoryInliningsFirst,
725 UseInlineAdvisor, MaxDevirtIterations);
727 // Require the GlobalsAA analysis for the module so we can query it within
728 // the CGSCC pipeline.
729 MIWP.addModulePass(RequireAnalysisPass<GlobalsAA, Module>());
730 // Invalidate AAManager so it can be recreated and pick up the newly available
733 createModuleToFunctionPassAdaptor(InvalidateAnalysisPass<AAManager>()));
735 // Require the ProfileSummaryAnalysis for the module so we can query it within
737 MIWP.addModulePass(RequireAnalysisPass<ProfileSummaryAnalysis, Module>());
739 // Now begin the main postorder CGSCC pipeline.
740 // FIXME: The current CGSCC pipeline has its origins in the legacy pass
741 // manager and trying to emulate its precise behavior. Much of this doesn't
742 // make a lot of sense and we should revisit the core CGSCC structure.
743 CGSCCPassManager &MainCGPipeline = MIWP.getPM();
745 // Note: historically, the PruneEH pass was run first to deduce nounwind and
746 // generally clean up exception handling overhead. It isn't clear this is
747 // valuable as the inliner doesn't currently care whether it is inlining an
750 if (AttributorRun & AttributorRunOption::CGSCC)
751 MainCGPipeline.addPass(AttributorCGSCCPass());
753 // Now deduce any function attributes based in the current code.
754 MainCGPipeline.addPass(PostOrderFunctionAttrsPass());
756 // When at O3 add argument promotion to the pass pipeline.
757 // FIXME: It isn't at all clear why this should be limited to O3.
758 if (Level == OptimizationLevel::O3)
759 MainCGPipeline.addPass(ArgumentPromotionPass());
761 // Try to perform OpenMP specific optimizations. This is a (quick!) no-op if
762 // there are no OpenMP runtime calls present in the module.
763 if (Level == OptimizationLevel::O2 || Level == OptimizationLevel::O3)
764 MainCGPipeline.addPass(OpenMPOptCGSCCPass());
766 for (auto &C : CGSCCOptimizerLateEPCallbacks)
767 C(MainCGPipeline, Level);
769 // Lastly, add the core function simplification pipeline nested inside the
771 MainCGPipeline.addPass(createCGSCCToFunctionPassAdaptor(
772 buildFunctionSimplificationPipeline(Level, Phase),
773 PTO.EagerlyInvalidateAnalyses, EnableNoRerunSimplificationPipeline));
775 MainCGPipeline.addPass(CoroSplitPass(Level != OptimizationLevel::O0));
777 if (EnableNoRerunSimplificationPipeline)
778 MIWP.addLateModulePass(createModuleToFunctionPassAdaptor(
779 InvalidateAnalysisPass<ShouldNotRunFunctionPassesAnalysis>()));
785 PassBuilder::buildModuleInlinerPipeline(OptimizationLevel Level,
786 ThinOrFullLTOPhase Phase) {
787 ModulePassManager MPM;
789 InlineParams IP = getInlineParamsFromOptLevel(Level);
790 if (Phase == ThinOrFullLTOPhase::ThinLTOPreLink && PGOOpt &&
791 PGOOpt->Action == PGOOptions::SampleUse)
792 IP.HotCallSiteThreshold = 0;
795 IP.EnableDeferral = EnablePGOInlineDeferral;
797 // The inline deferral logic is used to avoid losing some
798 // inlining chance in future. It is helpful in SCC inliner, in which
799 // inlining is processed in bottom-up order.
800 // While in module inliner, the inlining order is a priority-based order
801 // by default. The inline deferral is unnecessary there. So we disable the
802 // inline deferral logic in module inliner.
803 IP.EnableDeferral = false;
805 MPM.addPass(ModuleInlinerPass(IP, UseInlineAdvisor));
807 MPM.addPass(createModuleToFunctionPassAdaptor(
808 buildFunctionSimplificationPipeline(Level, Phase),
809 PTO.EagerlyInvalidateAnalyses));
811 MPM.addPass(createModuleToPostOrderCGSCCPassAdaptor(
812 CoroSplitPass(Level != OptimizationLevel::O0)));
818 PassBuilder::buildModuleSimplificationPipeline(OptimizationLevel Level,
819 ThinOrFullLTOPhase Phase) {
820 ModulePassManager MPM;
822 // Place pseudo probe instrumentation as the first pass of the pipeline to
823 // minimize the impact of optimization changes.
824 if (PGOOpt && PGOOpt->PseudoProbeForProfiling &&
825 Phase != ThinOrFullLTOPhase::ThinLTOPostLink)
826 MPM.addPass(SampleProfileProbePass(TM));
828 bool HasSampleProfile = PGOOpt && (PGOOpt->Action == PGOOptions::SampleUse);
830 // In ThinLTO mode, when flattened profile is used, all the available
831 // profile information will be annotated in PreLink phase so there is
832 // no need to load the profile again in PostLink.
833 bool LoadSampleProfile =
835 !(FlattenedProfileUsed && Phase == ThinOrFullLTOPhase::ThinLTOPostLink);
837 // During the ThinLTO backend phase we perform early indirect call promotion
838 // here, before globalopt. Otherwise imported available_externally functions
839 // look unreferenced and are removed. If we are going to load the sample
840 // profile then defer until later.
841 // TODO: See if we can move later and consolidate with the location where
842 // we perform ICP when we are loading a sample profile.
843 // TODO: We pass HasSampleProfile (whether there was a sample profile file
844 // passed to the compile) to the SamplePGO flag of ICP. This is used to
845 // determine whether the new direct calls are annotated with prof metadata.
846 // Ideally this should be determined from whether the IR is annotated with
847 // sample profile, and not whether the a sample profile was provided on the
848 // command line. E.g. for flattened profiles where we will not be reloading
849 // the sample profile in the ThinLTO backend, we ideally shouldn't have to
850 // provide the sample profile file.
851 if (Phase == ThinOrFullLTOPhase::ThinLTOPostLink && !LoadSampleProfile)
852 MPM.addPass(PGOIndirectCallPromotion(true /* InLTO */, HasSampleProfile));
854 // Do basic inference of function attributes from known properties of system
855 // libraries and other oracles.
856 MPM.addPass(InferFunctionAttrsPass());
858 // Create an early function pass manager to cleanup the output of the
860 FunctionPassManager EarlyFPM;
861 // Lower llvm.expect to metadata before attempting transforms.
862 // Compare/branch metadata may alter the behavior of passes like SimplifyCFG.
863 EarlyFPM.addPass(LowerExpectIntrinsicPass());
864 EarlyFPM.addPass(SimplifyCFGPass());
865 EarlyFPM.addPass(SROAPass());
866 EarlyFPM.addPass(EarlyCSEPass());
867 EarlyFPM.addPass(CoroEarlyPass());
868 if (Level == OptimizationLevel::O3)
869 EarlyFPM.addPass(CallSiteSplittingPass());
871 // In SamplePGO ThinLTO backend, we need instcombine before profile annotation
872 // to convert bitcast to direct calls so that they can be inlined during the
873 // profile annotation prepration step.
874 // More details about SamplePGO design can be found in:
875 // https://research.google.com/pubs/pub45290.html
876 // FIXME: revisit how SampleProfileLoad/Inliner/ICP is structured.
877 if (LoadSampleProfile)
878 EarlyFPM.addPass(InstCombinePass());
879 MPM.addPass(createModuleToFunctionPassAdaptor(std::move(EarlyFPM),
880 PTO.EagerlyInvalidateAnalyses));
882 if (LoadSampleProfile) {
883 // Annotate sample profile right after early FPM to ensure freshness of
885 MPM.addPass(SampleProfileLoaderPass(PGOOpt->ProfileFile,
886 PGOOpt->ProfileRemappingFile, Phase));
887 // Cache ProfileSummaryAnalysis once to avoid the potential need to insert
888 // RequireAnalysisPass for PSI before subsequent non-module passes.
889 MPM.addPass(RequireAnalysisPass<ProfileSummaryAnalysis, Module>());
890 // Do not invoke ICP in the LTOPrelink phase as it makes it hard
891 // for the profile annotation to be accurate in the LTO backend.
892 if (Phase != ThinOrFullLTOPhase::ThinLTOPreLink &&
893 Phase != ThinOrFullLTOPhase::FullLTOPreLink)
894 // We perform early indirect call promotion here, before globalopt.
895 // This is important for the ThinLTO backend phase because otherwise
896 // imported available_externally functions look unreferenced and are
899 PGOIndirectCallPromotion(true /* IsInLTO */, true /* SamplePGO */));
902 // Try to perform OpenMP specific optimizations on the module. This is a
903 // (quick!) no-op if there are no OpenMP runtime calls present in the module.
904 if (Level != OptimizationLevel::O0)
905 MPM.addPass(OpenMPOptPass());
907 if (AttributorRun & AttributorRunOption::MODULE)
908 MPM.addPass(AttributorPass());
910 // Lower type metadata and the type.test intrinsic in the ThinLTO
911 // post link pipeline after ICP. This is to enable usage of the type
912 // tests in ICP sequences.
913 if (Phase == ThinOrFullLTOPhase::ThinLTOPostLink)
914 MPM.addPass(LowerTypeTestsPass(nullptr, nullptr, true));
916 for (auto &C : PipelineEarlySimplificationEPCallbacks)
919 // Specialize functions with IPSCCP.
920 if (EnableFunctionSpecialization && Level == OptimizationLevel::O3)
921 MPM.addPass(FunctionSpecializationPass());
923 // Interprocedural constant propagation now that basic cleanup has occurred
924 // and prior to optimizing globals.
925 // FIXME: This position in the pipeline hasn't been carefully considered in
926 // years, it should be re-analyzed.
927 MPM.addPass(IPSCCPPass());
929 // Attach metadata to indirect call sites indicating the set of functions
930 // they may target at run-time. This should follow IPSCCP.
931 MPM.addPass(CalledValuePropagationPass());
933 // Optimize globals to try and fold them into constants.
934 MPM.addPass(GlobalOptPass());
936 // Promote any localized globals to SSA registers.
937 // FIXME: Should this instead by a run of SROA?
938 // FIXME: We should probably run instcombine and simplifycfg afterward to
939 // delete control flows that are dead once globals have been folded to
941 MPM.addPass(createModuleToFunctionPassAdaptor(PromotePass()));
943 // Remove any dead arguments exposed by cleanups and constant folding
945 MPM.addPass(DeadArgumentEliminationPass());
947 // Create a small function pass pipeline to cleanup after all the global
949 FunctionPassManager GlobalCleanupPM;
950 GlobalCleanupPM.addPass(InstCombinePass());
951 invokePeepholeEPCallbacks(GlobalCleanupPM, Level);
953 GlobalCleanupPM.addPass(
954 SimplifyCFGPass(SimplifyCFGOptions().convertSwitchRangeToICmp(true)));
955 MPM.addPass(createModuleToFunctionPassAdaptor(std::move(GlobalCleanupPM),
956 PTO.EagerlyInvalidateAnalyses));
958 // Add all the requested passes for instrumentation PGO, if requested.
959 if (PGOOpt && Phase != ThinOrFullLTOPhase::ThinLTOPostLink &&
960 (PGOOpt->Action == PGOOptions::IRInstr ||
961 PGOOpt->Action == PGOOptions::IRUse)) {
962 addPGOInstrPasses(MPM, Level,
963 /* RunProfileGen */ PGOOpt->Action == PGOOptions::IRInstr,
964 /* IsCS */ false, PGOOpt->ProfileFile,
965 PGOOpt->ProfileRemappingFile);
966 MPM.addPass(PGOIndirectCallPromotion(false, false));
968 if (PGOOpt && Phase != ThinOrFullLTOPhase::ThinLTOPostLink &&
969 PGOOpt->CSAction == PGOOptions::CSIRInstr)
970 MPM.addPass(PGOInstrumentationGenCreateVar(PGOOpt->CSProfileGenFile));
972 // Synthesize function entry counts for non-PGO compilation.
973 if (EnableSyntheticCounts && !PGOOpt)
974 MPM.addPass(SyntheticCountsPropagation());
976 if (EnableModuleInliner)
977 MPM.addPass(buildModuleInlinerPipeline(Level, Phase));
979 MPM.addPass(buildInlinerPipeline(Level, Phase));
981 if (EnableMemProfiler && Phase != ThinOrFullLTOPhase::ThinLTOPreLink) {
982 MPM.addPass(createModuleToFunctionPassAdaptor(MemProfilerPass()));
983 MPM.addPass(ModuleMemProfilerPass());
989 /// TODO: Should LTO cause any differences to this set of passes?
990 void PassBuilder::addVectorPasses(OptimizationLevel Level,
991 FunctionPassManager &FPM, bool IsFullLTO) {
992 FPM.addPass(LoopVectorizePass(
993 LoopVectorizeOptions(!PTO.LoopInterleaving, !PTO.LoopVectorization)));
996 // The vectorizer may have significantly shortened a loop body; unroll
997 // again. Unroll small loops to hide loop backedge latency and saturate any
998 // parallel execution resources of an out-of-order processor. We also then
999 // need to clean up redundancies and loop invariant code.
1000 // FIXME: It would be really good to use a loop-integrated instruction
1001 // combiner for cleanup here so that the unrolling and LICM can be pipelined
1002 // across the loop nests.
1003 // We do UnrollAndJam in a separate LPM to ensure it happens before unroll
1004 if (EnableUnrollAndJam && PTO.LoopUnrolling)
1005 FPM.addPass(createFunctionToLoopPassAdaptor(
1006 LoopUnrollAndJamPass(Level.getSpeedupLevel())));
1007 FPM.addPass(LoopUnrollPass(LoopUnrollOptions(
1008 Level.getSpeedupLevel(), /*OnlyWhenForced=*/!PTO.LoopUnrolling,
1009 PTO.ForgetAllSCEVInLoopUnroll)));
1010 FPM.addPass(WarnMissedTransformationsPass());
1014 // Eliminate loads by forwarding stores from the previous iteration to loads
1015 // of the current iteration.
1016 FPM.addPass(LoopLoadEliminationPass());
1018 // Cleanup after the loop optimization passes.
1019 FPM.addPass(InstCombinePass());
1021 if (Level.getSpeedupLevel() > 1 && ExtraVectorizerPasses) {
1022 ExtraVectorPassManager ExtraPasses;
1023 // At higher optimization levels, try to clean up any runtime overlap and
1024 // alignment checks inserted by the vectorizer. We want to track correlated
1025 // runtime checks for two inner loops in the same outer loop, fold any
1026 // common computations, hoist loop-invariant aspects out of any outer loop,
1027 // and unswitch the runtime checks if possible. Once hoisted, we may have
1028 // dead (or speculatable) control flows or more combining opportunities.
1029 ExtraPasses.addPass(EarlyCSEPass());
1030 ExtraPasses.addPass(CorrelatedValuePropagationPass());
1031 ExtraPasses.addPass(InstCombinePass());
1032 LoopPassManager LPM;
1033 LPM.addPass(LICMPass(PTO.LicmMssaOptCap, PTO.LicmMssaNoAccForPromotionCap,
1034 /*AllowSpeculation=*/true));
1035 LPM.addPass(SimpleLoopUnswitchPass(/* NonTrivial */ Level ==
1036 OptimizationLevel::O3));
1037 ExtraPasses.addPass(
1038 RequireAnalysisPass<OptimizationRemarkEmitterAnalysis, Function>());
1039 ExtraPasses.addPass(
1040 createFunctionToLoopPassAdaptor(std::move(LPM), /*UseMemorySSA=*/true,
1041 /*UseBlockFrequencyInfo=*/true));
1042 ExtraPasses.addPass(
1043 SimplifyCFGPass(SimplifyCFGOptions().convertSwitchRangeToICmp(true)));
1044 ExtraPasses.addPass(InstCombinePass());
1045 FPM.addPass(std::move(ExtraPasses));
1048 // Now that we've formed fast to execute loop structures, we do further
1049 // optimizations. These are run afterward as they might block doing complex
1050 // analyses and transforms such as what are needed for loop vectorization.
1052 // Cleanup after loop vectorization, etc. Simplification passes like CVP and
1053 // GVN, loop transforms, and others have already run, so it's now better to
1054 // convert to more optimized IR using more aggressive simplify CFG options.
1055 // The extra sinking transform can create larger basic blocks, so do this
1056 // before SLP vectorization.
1057 FPM.addPass(SimplifyCFGPass(SimplifyCFGOptions()
1058 .forwardSwitchCondToPhi(true)
1059 .convertSwitchRangeToICmp(true)
1060 .convertSwitchToLookupTable(true)
1061 .needCanonicalLoops(false)
1062 .hoistCommonInsts(true)
1063 .sinkCommonInsts(true)));
1066 FPM.addPass(SCCPPass());
1067 FPM.addPass(InstCombinePass());
1068 FPM.addPass(BDCEPass());
1071 // Optimize parallel scalar instruction chains into SIMD instructions.
1072 if (PTO.SLPVectorization) {
1073 FPM.addPass(SLPVectorizerPass());
1074 if (Level.getSpeedupLevel() > 1 && ExtraVectorizerPasses) {
1075 FPM.addPass(EarlyCSEPass());
1078 // Enhance/cleanup vector code.
1079 FPM.addPass(VectorCombinePass());
1082 FPM.addPass(InstCombinePass());
1083 // Unroll small loops to hide loop backedge latency and saturate any
1084 // parallel execution resources of an out-of-order processor. We also then
1085 // need to clean up redundancies and loop invariant code.
1086 // FIXME: It would be really good to use a loop-integrated instruction
1087 // combiner for cleanup here so that the unrolling and LICM can be pipelined
1088 // across the loop nests.
1089 // We do UnrollAndJam in a separate LPM to ensure it happens before unroll
1090 if (EnableUnrollAndJam && PTO.LoopUnrolling) {
1091 FPM.addPass(createFunctionToLoopPassAdaptor(
1092 LoopUnrollAndJamPass(Level.getSpeedupLevel())));
1094 FPM.addPass(LoopUnrollPass(LoopUnrollOptions(
1095 Level.getSpeedupLevel(), /*OnlyWhenForced=*/!PTO.LoopUnrolling,
1096 PTO.ForgetAllSCEVInLoopUnroll)));
1097 FPM.addPass(WarnMissedTransformationsPass());
1098 FPM.addPass(InstCombinePass());
1100 RequireAnalysisPass<OptimizationRemarkEmitterAnalysis, Function>());
1101 FPM.addPass(createFunctionToLoopPassAdaptor(
1102 LICMPass(PTO.LicmMssaOptCap, PTO.LicmMssaNoAccForPromotionCap,
1103 /*AllowSpeculation=*/true),
1104 /*UseMemorySSA=*/true, /*UseBlockFrequencyInfo=*/true));
1107 // Now that we've vectorized and unrolled loops, we may have more refined
1108 // alignment information, try to re-derive it here.
1109 FPM.addPass(AlignmentFromAssumptionsPass());
1112 FPM.addPass(InstCombinePass());
1116 PassBuilder::buildModuleOptimizationPipeline(OptimizationLevel Level,
1118 ModulePassManager MPM;
1120 // Optimize globals now that the module is fully simplified.
1121 MPM.addPass(GlobalOptPass());
1122 MPM.addPass(GlobalDCEPass());
1124 // Run partial inlining pass to partially inline functions that have
1126 if (RunPartialInlining)
1127 MPM.addPass(PartialInlinerPass());
1129 // Remove avail extern fns and globals definitions since we aren't compiling
1130 // an object file for later LTO. For LTO we want to preserve these so they
1131 // are eligible for inlining at link-time. Note if they are unreferenced they
1132 // will be removed by GlobalDCE later, so this only impacts referenced
1133 // available externally globals. Eventually they will be suppressed during
1134 // codegen, but eliminating here enables more opportunity for GlobalDCE as it
1135 // may make globals referenced by available external functions dead and saves
1136 // running remaining passes on the eliminated functions. These should be
1137 // preserved during prelinking for link-time inlining decisions.
1139 MPM.addPass(EliminateAvailableExternallyPass());
1141 if (EnableOrderFileInstrumentation)
1142 MPM.addPass(InstrOrderFilePass());
1144 // Do RPO function attribute inference across the module to forward-propagate
1145 // attributes where applicable.
1146 // FIXME: Is this really an optimization rather than a canonicalization?
1147 MPM.addPass(ReversePostOrderFunctionAttrsPass());
1149 // Do a post inline PGO instrumentation and use pass. This is a context
1150 // sensitive PGO pass. We don't want to do this in LTOPreLink phrase as
1151 // cross-module inline has not been done yet. The context sensitive
1152 // instrumentation is after all the inlines are done.
1153 if (!LTOPreLink && PGOOpt) {
1154 if (PGOOpt->CSAction == PGOOptions::CSIRInstr)
1155 addPGOInstrPasses(MPM, Level, /* RunProfileGen */ true,
1156 /* IsCS */ true, PGOOpt->CSProfileGenFile,
1157 PGOOpt->ProfileRemappingFile);
1158 else if (PGOOpt->CSAction == PGOOptions::CSIRUse)
1159 addPGOInstrPasses(MPM, Level, /* RunProfileGen */ false,
1160 /* IsCS */ true, PGOOpt->ProfileFile,
1161 PGOOpt->ProfileRemappingFile);
1164 // Re-require GloblasAA here prior to function passes. This is particularly
1165 // useful as the above will have inlined, DCE'ed, and function-attr
1166 // propagated everything. We should at this point have a reasonably minimal
1167 // and richly annotated call graph. By computing aliasing and mod/ref
1168 // information for all local globals here, the late loop passes and notably
1169 // the vectorizer will be able to use them to help recognize vectorizable
1170 // memory operations.
1171 MPM.addPass(RequireAnalysisPass<GlobalsAA, Module>());
1173 FunctionPassManager OptimizePM;
1174 OptimizePM.addPass(Float2IntPass());
1175 OptimizePM.addPass(LowerConstantIntrinsicsPass());
1178 OptimizePM.addPass(LowerMatrixIntrinsicsPass());
1179 OptimizePM.addPass(EarlyCSEPass());
1182 // FIXME: We need to run some loop optimizations to re-rotate loops after
1183 // simplifycfg and others undo their rotation.
1185 // Optimize the loop execution. These passes operate on entire loop nests
1186 // rather than on each loop in an inside-out manner, and so they are actually
1189 for (auto &C : VectorizerStartEPCallbacks)
1190 C(OptimizePM, Level);
1192 LoopPassManager LPM;
1193 // First rotate loops that may have been un-rotated by prior passes.
1194 // Disable header duplication at -Oz.
1195 LPM.addPass(LoopRotatePass(Level != OptimizationLevel::Oz, LTOPreLink));
1196 // Some loops may have become dead by now. Try to delete them.
1197 // FIXME: see discussion in https://reviews.llvm.org/D112851,
1198 // this may need to be revisited once we run GVN before loop deletion
1199 // in the simplification pipeline.
1200 LPM.addPass(LoopDeletionPass());
1201 OptimizePM.addPass(createFunctionToLoopPassAdaptor(
1202 std::move(LPM), /*UseMemorySSA=*/false, /*UseBlockFrequencyInfo=*/false));
1204 // Distribute loops to allow partial vectorization. I.e. isolate dependences
1205 // into separate loop that would otherwise inhibit vectorization. This is
1206 // currently only performed for loops marked with the metadata
1207 // llvm.loop.distribute=true or when -enable-loop-distribute is specified.
1208 OptimizePM.addPass(LoopDistributePass());
1210 // Populates the VFABI attribute with the scalar-to-vector mappings
1211 // from the TargetLibraryInfo.
1212 OptimizePM.addPass(InjectTLIMappings());
1214 addVectorPasses(Level, OptimizePM, /* IsFullLTO */ false);
1216 // LoopSink pass sinks instructions hoisted by LICM, which serves as a
1217 // canonicalization pass that enables other optimizations. As a result,
1218 // LoopSink pass needs to be a very late IR pass to avoid undoing LICM
1219 // result too early.
1220 OptimizePM.addPass(LoopSinkPass());
1222 // And finally clean up LCSSA form before generating code.
1223 OptimizePM.addPass(InstSimplifyPass());
1225 // This hoists/decomposes div/rem ops. It should run after other sink/hoist
1226 // passes to avoid re-sinking, but before SimplifyCFG because it can allow
1227 // flattening of blocks.
1228 OptimizePM.addPass(DivRemPairsPass());
1230 // LoopSink (and other loop passes since the last simplifyCFG) might have
1231 // resulted in single-entry-single-exit or empty blocks. Clean up the CFG.
1233 SimplifyCFGPass(SimplifyCFGOptions().convertSwitchRangeToICmp(true)));
1235 OptimizePM.addPass(CoroCleanupPass());
1237 // Add the core optimizing pipeline.
1238 MPM.addPass(createModuleToFunctionPassAdaptor(std::move(OptimizePM),
1239 PTO.EagerlyInvalidateAnalyses));
1241 for (auto &C : OptimizerLastEPCallbacks)
1244 // Split out cold code. Splitting is done late to avoid hiding context from
1245 // other optimizations and inadvertently regressing performance. The tradeoff
1246 // is that this has a higher code size cost than splitting early.
1247 if (EnableHotColdSplit && !LTOPreLink)
1248 MPM.addPass(HotColdSplittingPass());
1250 // Search the code for similar regions of code. If enough similar regions can
1251 // be found where extracting the regions into their own function will decrease
1252 // the size of the program, we extract the regions, a deduplicate the
1253 // structurally similar regions.
1254 if (EnableIROutliner)
1255 MPM.addPass(IROutlinerPass());
1257 // Merge functions if requested.
1258 if (PTO.MergeFunctions)
1259 MPM.addPass(MergeFunctionsPass());
1261 if (PTO.CallGraphProfile)
1262 MPM.addPass(CGProfilePass());
1264 // Now we need to do some global optimization transforms.
1265 // FIXME: It would seem like these should come first in the optimization
1266 // pipeline and maybe be the bottom of the canonicalization pipeline? Weird
1268 MPM.addPass(GlobalDCEPass());
1269 MPM.addPass(ConstantMergePass());
1271 // TODO: Relative look table converter pass caused an issue when full lto is
1272 // enabled. See https://reviews.llvm.org/D94355 for more details.
1273 // Until the issue fixed, disable this pass during pre-linking phase.
1275 MPM.addPass(RelLookupTableConverterPass());
1281 PassBuilder::buildPerModuleDefaultPipeline(OptimizationLevel Level,
1283 assert(Level != OptimizationLevel::O0 &&
1284 "Must request optimizations for the default pipeline!");
1286 ModulePassManager MPM;
1288 // Convert @llvm.global.annotations to !annotation metadata.
1289 MPM.addPass(Annotation2MetadataPass());
1291 // Force any function attributes we want the rest of the pipeline to observe.
1292 MPM.addPass(ForceFunctionAttrsPass());
1294 // Apply module pipeline start EP callback.
1295 for (auto &C : PipelineStartEPCallbacks)
1298 if (PGOOpt && PGOOpt->DebugInfoForProfiling)
1299 MPM.addPass(createModuleToFunctionPassAdaptor(AddDiscriminatorsPass()));
1301 // Add the core simplification pipeline.
1302 MPM.addPass(buildModuleSimplificationPipeline(
1303 Level, LTOPreLink ? ThinOrFullLTOPhase::FullLTOPreLink
1304 : ThinOrFullLTOPhase::None));
1306 // Now add the optimization pipeline.
1307 MPM.addPass(buildModuleOptimizationPipeline(Level, LTOPreLink));
1309 if (PGOOpt && PGOOpt->PseudoProbeForProfiling &&
1310 PGOOpt->Action == PGOOptions::SampleUse)
1311 MPM.addPass(PseudoProbeUpdatePass());
1313 // Emit annotation remarks.
1314 addAnnotationRemarksPass(MPM);
1317 addRequiredLTOPreLinkPasses(MPM);
1323 PassBuilder::buildThinLTOPreLinkDefaultPipeline(OptimizationLevel Level) {
1324 assert(Level != OptimizationLevel::O0 &&
1325 "Must request optimizations for the default pipeline!");
1327 ModulePassManager MPM;
1329 // Convert @llvm.global.annotations to !annotation metadata.
1330 MPM.addPass(Annotation2MetadataPass());
1332 // Force any function attributes we want the rest of the pipeline to observe.
1333 MPM.addPass(ForceFunctionAttrsPass());
1335 if (PGOOpt && PGOOpt->DebugInfoForProfiling)
1336 MPM.addPass(createModuleToFunctionPassAdaptor(AddDiscriminatorsPass()));
1338 // Apply module pipeline start EP callback.
1339 for (auto &C : PipelineStartEPCallbacks)
1342 // If we are planning to perform ThinLTO later, we don't bloat the code with
1343 // unrolling/vectorization/... now. Just simplify the module as much as we
1345 MPM.addPass(buildModuleSimplificationPipeline(
1346 Level, ThinOrFullLTOPhase::ThinLTOPreLink));
1348 // Run partial inlining pass to partially inline functions that have
1350 // FIXME: It isn't clear whether this is really the right place to run this
1351 // in ThinLTO. Because there is another canonicalization and simplification
1352 // phase that will run after the thin link, running this here ends up with
1353 // less information than will be available later and it may grow functions in
1354 // ways that aren't beneficial.
1355 if (RunPartialInlining)
1356 MPM.addPass(PartialInlinerPass());
1358 // Reduce the size of the IR as much as possible.
1359 MPM.addPass(GlobalOptPass());
1361 // Module simplification splits coroutines, but does not fully clean up
1362 // coroutine intrinsics. To ensure ThinLTO optimization passes don't trip up
1363 // on these, we schedule the cleanup here.
1364 MPM.addPass(createModuleToFunctionPassAdaptor(CoroCleanupPass()));
1366 if (PGOOpt && PGOOpt->PseudoProbeForProfiling &&
1367 PGOOpt->Action == PGOOptions::SampleUse)
1368 MPM.addPass(PseudoProbeUpdatePass());
1370 // Handle OptimizerLastEPCallbacks added by clang on PreLink. Actual
1371 // optimization is going to be done in PostLink stage, but clang can't
1372 // add callbacks there in case of in-process ThinLTO called by linker.
1373 for (auto &C : OptimizerLastEPCallbacks)
1376 // Emit annotation remarks.
1377 addAnnotationRemarksPass(MPM);
1379 addRequiredLTOPreLinkPasses(MPM);
1384 ModulePassManager PassBuilder::buildThinLTODefaultPipeline(
1385 OptimizationLevel Level, const ModuleSummaryIndex *ImportSummary) {
1386 ModulePassManager MPM;
1388 // Convert @llvm.global.annotations to !annotation metadata.
1389 MPM.addPass(Annotation2MetadataPass());
1391 if (ImportSummary) {
1392 // These passes import type identifier resolutions for whole-program
1393 // devirtualization and CFI. They must run early because other passes may
1394 // disturb the specific instruction patterns that these passes look for,
1395 // creating dependencies on resolutions that may not appear in the summary.
1397 // For example, GVN may transform the pattern assume(type.test) appearing in
1398 // two basic blocks into assume(phi(type.test, type.test)), which would
1399 // transform a dependency on a WPD resolution into a dependency on a type
1400 // identifier resolution for CFI.
1402 // Also, WPD has access to more precise information than ICP and can
1403 // devirtualize more effectively, so it should operate on the IR first.
1405 // The WPD and LowerTypeTest passes need to run at -O0 to lower type
1406 // metadata and intrinsics.
1407 MPM.addPass(WholeProgramDevirtPass(nullptr, ImportSummary));
1408 MPM.addPass(LowerTypeTestsPass(nullptr, ImportSummary));
1411 if (Level == OptimizationLevel::O0) {
1412 // Run a second time to clean up any type tests left behind by WPD for use
1414 MPM.addPass(LowerTypeTestsPass(nullptr, nullptr, true));
1415 // Drop available_externally and unreferenced globals. This is necessary
1416 // with ThinLTO in order to avoid leaving undefined references to dead
1417 // globals in the object file.
1418 MPM.addPass(EliminateAvailableExternallyPass());
1419 MPM.addPass(GlobalDCEPass());
1423 // Force any function attributes we want the rest of the pipeline to observe.
1424 MPM.addPass(ForceFunctionAttrsPass());
1426 // Add the core simplification pipeline.
1427 MPM.addPass(buildModuleSimplificationPipeline(
1428 Level, ThinOrFullLTOPhase::ThinLTOPostLink));
1430 // Now add the optimization pipeline.
1431 MPM.addPass(buildModuleOptimizationPipeline(Level));
1433 // Emit annotation remarks.
1434 addAnnotationRemarksPass(MPM);
1440 PassBuilder::buildLTOPreLinkDefaultPipeline(OptimizationLevel Level) {
1441 assert(Level != OptimizationLevel::O0 &&
1442 "Must request optimizations for the default pipeline!");
1443 // FIXME: We should use a customized pre-link pipeline!
1444 return buildPerModuleDefaultPipeline(Level,
1445 /* LTOPreLink */ true);
1449 PassBuilder::buildLTODefaultPipeline(OptimizationLevel Level,
1450 ModuleSummaryIndex *ExportSummary) {
1451 ModulePassManager MPM;
1453 // Convert @llvm.global.annotations to !annotation metadata.
1454 MPM.addPass(Annotation2MetadataPass());
1456 // Create a function that performs CFI checks for cross-DSO calls with targets
1457 // in the current module.
1458 MPM.addPass(CrossDSOCFIPass());
1460 if (Level == OptimizationLevel::O0) {
1461 // The WPD and LowerTypeTest passes need to run at -O0 to lower type
1462 // metadata and intrinsics.
1463 MPM.addPass(WholeProgramDevirtPass(ExportSummary, nullptr));
1464 MPM.addPass(LowerTypeTestsPass(ExportSummary, nullptr));
1465 // Run a second time to clean up any type tests left behind by WPD for use
1467 MPM.addPass(LowerTypeTestsPass(nullptr, nullptr, true));
1469 // Emit annotation remarks.
1470 addAnnotationRemarksPass(MPM);
1475 if (PGOOpt && PGOOpt->Action == PGOOptions::SampleUse) {
1476 // Load sample profile before running the LTO optimization pipeline.
1477 MPM.addPass(SampleProfileLoaderPass(PGOOpt->ProfileFile,
1478 PGOOpt->ProfileRemappingFile,
1479 ThinOrFullLTOPhase::FullLTOPostLink));
1480 // Cache ProfileSummaryAnalysis once to avoid the potential need to insert
1481 // RequireAnalysisPass for PSI before subsequent non-module passes.
1482 MPM.addPass(RequireAnalysisPass<ProfileSummaryAnalysis, Module>());
1485 // Try to run OpenMP optimizations, quick no-op if no OpenMP metadata present.
1486 MPM.addPass(OpenMPOptPass());
1488 // Remove unused virtual tables to improve the quality of code generated by
1489 // whole-program devirtualization and bitset lowering.
1490 MPM.addPass(GlobalDCEPass());
1492 // Force any function attributes we want the rest of the pipeline to observe.
1493 MPM.addPass(ForceFunctionAttrsPass());
1495 // Do basic inference of function attributes from known properties of system
1496 // libraries and other oracles.
1497 MPM.addPass(InferFunctionAttrsPass());
1499 if (Level.getSpeedupLevel() > 1) {
1500 FunctionPassManager EarlyFPM;
1501 EarlyFPM.addPass(CallSiteSplittingPass());
1502 MPM.addPass(createModuleToFunctionPassAdaptor(
1503 std::move(EarlyFPM), PTO.EagerlyInvalidateAnalyses));
1505 // Indirect call promotion. This should promote all the targets that are
1506 // left by the earlier promotion pass that promotes intra-module targets.
1507 // This two-step promotion is to save the compile time. For LTO, it should
1508 // produce the same result as if we only do promotion here.
1509 MPM.addPass(PGOIndirectCallPromotion(
1510 true /* InLTO */, PGOOpt && PGOOpt->Action == PGOOptions::SampleUse));
1512 if (EnableFunctionSpecialization && Level == OptimizationLevel::O3)
1513 MPM.addPass(FunctionSpecializationPass());
1514 // Propagate constants at call sites into the functions they call. This
1515 // opens opportunities for globalopt (and inlining) by substituting function
1516 // pointers passed as arguments to direct uses of functions.
1517 MPM.addPass(IPSCCPPass());
1519 // Attach metadata to indirect call sites indicating the set of functions
1520 // they may target at run-time. This should follow IPSCCP.
1521 MPM.addPass(CalledValuePropagationPass());
1524 // Now deduce any function attributes based in the current code.
1526 createModuleToPostOrderCGSCCPassAdaptor(PostOrderFunctionAttrsPass()));
1528 // Do RPO function attribute inference across the module to forward-propagate
1529 // attributes where applicable.
1530 // FIXME: Is this really an optimization rather than a canonicalization?
1531 MPM.addPass(ReversePostOrderFunctionAttrsPass());
1533 // Use in-range annotations on GEP indices to split globals where beneficial.
1534 MPM.addPass(GlobalSplitPass());
1536 // Run whole program optimization of virtual call when the list of callees
1538 MPM.addPass(WholeProgramDevirtPass(ExportSummary, nullptr));
1540 // Stop here at -O1.
1541 if (Level == OptimizationLevel::O1) {
1542 // The LowerTypeTestsPass needs to run to lower type metadata and the
1543 // type.test intrinsics. The pass does nothing if CFI is disabled.
1544 MPM.addPass(LowerTypeTestsPass(ExportSummary, nullptr));
1545 // Run a second time to clean up any type tests left behind by WPD for use
1546 // in ICP (which is performed earlier than this in the regular LTO
1548 MPM.addPass(LowerTypeTestsPass(nullptr, nullptr, true));
1550 // Emit annotation remarks.
1551 addAnnotationRemarksPass(MPM);
1556 // Optimize globals to try and fold them into constants.
1557 MPM.addPass(GlobalOptPass());
1559 // Promote any localized globals to SSA registers.
1560 MPM.addPass(createModuleToFunctionPassAdaptor(PromotePass()));
1562 // Linking modules together can lead to duplicate global constant, only
1563 // keep one copy of each constant.
1564 MPM.addPass(ConstantMergePass());
1566 // Remove unused arguments from functions.
1567 MPM.addPass(DeadArgumentEliminationPass());
1569 // Reduce the code after globalopt and ipsccp. Both can open up significant
1570 // simplification opportunities, and both can propagate functions through
1571 // function pointers. When this happens, we often have to resolve varargs
1572 // calls, etc, so let instcombine do this.
1573 FunctionPassManager PeepholeFPM;
1574 PeepholeFPM.addPass(InstCombinePass());
1575 if (Level == OptimizationLevel::O3)
1576 PeepholeFPM.addPass(AggressiveInstCombinePass());
1577 invokePeepholeEPCallbacks(PeepholeFPM, Level);
1579 MPM.addPass(createModuleToFunctionPassAdaptor(std::move(PeepholeFPM),
1580 PTO.EagerlyInvalidateAnalyses));
1582 // Note: historically, the PruneEH pass was run first to deduce nounwind and
1583 // generally clean up exception handling overhead. It isn't clear this is
1584 // valuable as the inliner doesn't currently care whether it is inlining an
1585 // invoke or a call.
1586 // Run the inliner now.
1587 MPM.addPass(ModuleInlinerWrapperPass(getInlineParamsFromOptLevel(Level)));
1589 // Optimize globals again after we ran the inliner.
1590 MPM.addPass(GlobalOptPass());
1592 // Garbage collect dead functions.
1593 MPM.addPass(GlobalDCEPass());
1595 // If we didn't decide to inline a function, check to see if we can
1596 // transform it to pass arguments by value instead of by reference.
1597 MPM.addPass(createModuleToPostOrderCGSCCPassAdaptor(ArgumentPromotionPass()));
1599 FunctionPassManager FPM;
1600 // The IPO Passes may leave cruft around. Clean up after them.
1601 FPM.addPass(InstCombinePass());
1602 invokePeepholeEPCallbacks(FPM, Level);
1604 FPM.addPass(JumpThreadingPass(/*InsertFreezeWhenUnfoldingSelect*/ true));
1606 // Do a post inline PGO instrumentation and use pass. This is a context
1607 // sensitive PGO pass.
1609 if (PGOOpt->CSAction == PGOOptions::CSIRInstr)
1610 addPGOInstrPasses(MPM, Level, /* RunProfileGen */ true,
1611 /* IsCS */ true, PGOOpt->CSProfileGenFile,
1612 PGOOpt->ProfileRemappingFile);
1613 else if (PGOOpt->CSAction == PGOOptions::CSIRUse)
1614 addPGOInstrPasses(MPM, Level, /* RunProfileGen */ false,
1615 /* IsCS */ true, PGOOpt->ProfileFile,
1616 PGOOpt->ProfileRemappingFile);
1620 FPM.addPass(SROAPass());
1622 // LTO provides additional opportunities for tailcall elimination due to
1623 // link-time inlining, and visibility of nocapture attribute.
1624 FPM.addPass(TailCallElimPass());
1626 // Run a few AA driver optimizations here and now to cleanup the code.
1627 MPM.addPass(createModuleToFunctionPassAdaptor(std::move(FPM),
1628 PTO.EagerlyInvalidateAnalyses));
1631 createModuleToPostOrderCGSCCPassAdaptor(PostOrderFunctionAttrsPass()));
1633 // Require the GlobalsAA analysis for the module so we can query it within
1635 MPM.addPass(RequireAnalysisPass<GlobalsAA, Module>());
1636 // Invalidate AAManager so it can be recreated and pick up the newly available
1639 createModuleToFunctionPassAdaptor(InvalidateAnalysisPass<AAManager>()));
1641 FunctionPassManager MainFPM;
1642 MainFPM.addPass(createFunctionToLoopPassAdaptor(
1643 LICMPass(PTO.LicmMssaOptCap, PTO.LicmMssaNoAccForPromotionCap,
1644 /*AllowSpeculation=*/true),
1645 /*USeMemorySSA=*/true, /*UseBlockFrequencyInfo=*/true));
1648 MainFPM.addPass(NewGVNPass());
1650 MainFPM.addPass(GVNPass());
1652 // Remove dead memcpy()'s.
1653 MainFPM.addPass(MemCpyOptPass());
1655 // Nuke dead stores.
1656 MainFPM.addPass(DSEPass());
1657 MainFPM.addPass(MergedLoadStoreMotionPass());
1660 if (EnableConstraintElimination)
1661 MainFPM.addPass(ConstraintEliminationPass());
1663 LoopPassManager LPM;
1664 if (EnableLoopFlatten && Level.getSpeedupLevel() > 1)
1665 LPM.addPass(LoopFlattenPass());
1666 LPM.addPass(IndVarSimplifyPass());
1667 LPM.addPass(LoopDeletionPass());
1668 // FIXME: Add loop interchange.
1670 // Unroll small loops and perform peeling.
1671 LPM.addPass(LoopFullUnrollPass(Level.getSpeedupLevel(),
1672 /* OnlyWhenForced= */ !PTO.LoopUnrolling,
1673 PTO.ForgetAllSCEVInLoopUnroll));
1674 // The loop passes in LPM (LoopFullUnrollPass) do not preserve MemorySSA.
1675 // *All* loop passes must preserve it, in order to be able to use it.
1676 MainFPM.addPass(createFunctionToLoopPassAdaptor(
1677 std::move(LPM), /*UseMemorySSA=*/false, /*UseBlockFrequencyInfo=*/true));
1679 MainFPM.addPass(LoopDistributePass());
1681 addVectorPasses(Level, MainFPM, /* IsFullLTO */ true);
1683 // Run the OpenMPOpt CGSCC pass again late.
1685 createModuleToPostOrderCGSCCPassAdaptor(OpenMPOptCGSCCPass()));
1687 invokePeepholeEPCallbacks(MainFPM, Level);
1688 MainFPM.addPass(JumpThreadingPass(/*InsertFreezeWhenUnfoldingSelect*/ true));
1689 MPM.addPass(createModuleToFunctionPassAdaptor(std::move(MainFPM),
1690 PTO.EagerlyInvalidateAnalyses));
1692 // Lower type metadata and the type.test intrinsic. This pass supports
1693 // clang's control flow integrity mechanisms (-fsanitize=cfi*) and needs
1694 // to be run at link time if CFI is enabled. This pass does nothing if
1696 MPM.addPass(LowerTypeTestsPass(ExportSummary, nullptr));
1697 // Run a second time to clean up any type tests left behind by WPD for use
1698 // in ICP (which is performed earlier than this in the regular LTO pipeline).
1699 MPM.addPass(LowerTypeTestsPass(nullptr, nullptr, true));
1701 // Enable splitting late in the FullLTO post-link pipeline. This is done in
1702 // the same stage in the old pass manager (\ref addLateLTOOptimizationPasses).
1703 if (EnableHotColdSplit)
1704 MPM.addPass(HotColdSplittingPass());
1706 // Add late LTO optimization passes.
1707 // Delete basic blocks, which optimization passes may have killed.
1708 MPM.addPass(createModuleToFunctionPassAdaptor(SimplifyCFGPass(
1709 SimplifyCFGOptions().convertSwitchRangeToICmp(true).hoistCommonInsts(
1712 // Drop bodies of available eternally objects to improve GlobalDCE.
1713 MPM.addPass(EliminateAvailableExternallyPass());
1715 // Now that we have optimized the program, discard unreachable functions.
1716 MPM.addPass(GlobalDCEPass());
1718 if (PTO.MergeFunctions)
1719 MPM.addPass(MergeFunctionsPass());
1721 // Emit annotation remarks.
1722 addAnnotationRemarksPass(MPM);
1727 ModulePassManager PassBuilder::buildO0DefaultPipeline(OptimizationLevel Level,
1729 assert(Level == OptimizationLevel::O0 &&
1730 "buildO0DefaultPipeline should only be used with O0");
1732 ModulePassManager MPM;
1734 // Perform pseudo probe instrumentation in O0 mode. This is for the
1735 // consistency between different build modes. For example, a LTO build can be
1736 // mixed with an O0 prelink and an O2 postlink. Loading a sample profile in
1737 // the postlink will require pseudo probe instrumentation in the prelink.
1738 if (PGOOpt && PGOOpt->PseudoProbeForProfiling)
1739 MPM.addPass(SampleProfileProbePass(TM));
1741 if (PGOOpt && (PGOOpt->Action == PGOOptions::IRInstr ||
1742 PGOOpt->Action == PGOOptions::IRUse))
1743 addPGOInstrPassesForO0(
1745 /* RunProfileGen */ (PGOOpt->Action == PGOOptions::IRInstr),
1746 /* IsCS */ false, PGOOpt->ProfileFile, PGOOpt->ProfileRemappingFile);
1748 for (auto &C : PipelineStartEPCallbacks)
1751 if (PGOOpt && PGOOpt->DebugInfoForProfiling)
1752 MPM.addPass(createModuleToFunctionPassAdaptor(AddDiscriminatorsPass()));
1754 for (auto &C : PipelineEarlySimplificationEPCallbacks)
1757 // Build a minimal pipeline based on the semantics required by LLVM,
1758 // which is just that always inlining occurs. Further, disable generating
1759 // lifetime intrinsics to avoid enabling further optimizations during
1761 MPM.addPass(AlwaysInlinerPass(
1762 /*InsertLifetimeIntrinsics=*/false));
1764 if (PTO.MergeFunctions)
1765 MPM.addPass(MergeFunctionsPass());
1769 createModuleToFunctionPassAdaptor(LowerMatrixIntrinsicsPass(true)));
1771 if (!CGSCCOptimizerLateEPCallbacks.empty()) {
1772 CGSCCPassManager CGPM;
1773 for (auto &C : CGSCCOptimizerLateEPCallbacks)
1775 if (!CGPM.isEmpty())
1776 MPM.addPass(createModuleToPostOrderCGSCCPassAdaptor(std::move(CGPM)));
1778 if (!LateLoopOptimizationsEPCallbacks.empty()) {
1779 LoopPassManager LPM;
1780 for (auto &C : LateLoopOptimizationsEPCallbacks)
1782 if (!LPM.isEmpty()) {
1783 MPM.addPass(createModuleToFunctionPassAdaptor(
1784 createFunctionToLoopPassAdaptor(std::move(LPM))));
1787 if (!LoopOptimizerEndEPCallbacks.empty()) {
1788 LoopPassManager LPM;
1789 for (auto &C : LoopOptimizerEndEPCallbacks)
1791 if (!LPM.isEmpty()) {
1792 MPM.addPass(createModuleToFunctionPassAdaptor(
1793 createFunctionToLoopPassAdaptor(std::move(LPM))));
1796 if (!ScalarOptimizerLateEPCallbacks.empty()) {
1797 FunctionPassManager FPM;
1798 for (auto &C : ScalarOptimizerLateEPCallbacks)
1801 MPM.addPass(createModuleToFunctionPassAdaptor(std::move(FPM)));
1803 if (!VectorizerStartEPCallbacks.empty()) {
1804 FunctionPassManager FPM;
1805 for (auto &C : VectorizerStartEPCallbacks)
1808 MPM.addPass(createModuleToFunctionPassAdaptor(std::move(FPM)));
1811 MPM.addPass(createModuleToFunctionPassAdaptor(CoroEarlyPass()));
1812 CGSCCPassManager CGPM;
1813 CGPM.addPass(CoroSplitPass());
1814 MPM.addPass(createModuleToPostOrderCGSCCPassAdaptor(std::move(CGPM)));
1815 MPM.addPass(createModuleToFunctionPassAdaptor(CoroCleanupPass()));
1817 for (auto &C : OptimizerLastEPCallbacks)
1821 addRequiredLTOPreLinkPasses(MPM);
1823 MPM.addPass(createModuleToFunctionPassAdaptor(AnnotationRemarksPass()));
1828 AAManager PassBuilder::buildDefaultAAPipeline() {
1831 // The order in which these are registered determines their priority when
1834 // First we register the basic alias analysis that provides the majority of
1835 // per-function local AA logic. This is a stateless, on-demand local set of
1837 AA.registerFunctionAnalysis<BasicAA>();
1839 // Next we query fast, specialized alias analyses that wrap IR-embedded
1840 // information about aliasing.
1841 AA.registerFunctionAnalysis<ScopedNoAliasAA>();
1842 AA.registerFunctionAnalysis<TypeBasedAA>();
1844 // Add support for querying global aliasing information when available.
1845 // Because the `AAManager` is a function analysis and `GlobalsAA` is a module
1846 // analysis, all that the `AAManager` can do is query for any *cached*
1847 // results from `GlobalsAA` through a readonly proxy.
1848 AA.registerModuleAnalysis<GlobalsAA>();
1850 // Add target-specific alias analyses.
1852 TM->registerDefaultAliasAnalyses(AA);