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.
262 FPM.addPass(SimplifyCFGPass());
263 FPM.addPass(InstCombinePass());
265 FPM.addPass(LibCallsShrinkWrapPass());
267 invokePeepholeEPCallbacks(FPM, Level);
269 FPM.addPass(SimplifyCFGPass());
271 // Form canonically associated expression trees, and simplify the trees using
272 // basic mathematical properties. For example, this will form (nearly)
273 // minimal multiplication trees.
274 FPM.addPass(ReassociatePass());
276 // Add the primary loop simplification pipeline.
277 // FIXME: Currently this is split into two loop pass pipelines because we run
278 // some function passes in between them. These can and should be removed
279 // and/or replaced by scheduling the loop pass equivalents in the correct
280 // positions. But those equivalent passes aren't powerful enough yet.
281 // Specifically, `SimplifyCFGPass` and `InstCombinePass` are currently still
282 // used. We have `LoopSimplifyCFGPass` which isn't yet powerful enough yet to
283 // fully replace `SimplifyCFGPass`, and the closest to the other we have is
284 // `LoopInstSimplify`.
285 LoopPassManager LPM1, LPM2;
287 // Simplify the loop body. We do this initially to clean up after other loop
288 // passes run, either when iterating on a loop or on inner loops with
289 // implications on the outer loop.
290 LPM1.addPass(LoopInstSimplifyPass());
291 LPM1.addPass(LoopSimplifyCFGPass());
293 // Try to remove as much code from the loop header as possible,
294 // to reduce amount of IR that will have to be duplicated.
295 // TODO: Investigate promotion cap for O1.
296 LPM1.addPass(LICMPass(PTO.LicmMssaOptCap, PTO.LicmMssaNoAccForPromotionCap));
298 LPM1.addPass(LoopRotatePass(/* Disable header duplication */ true,
299 isLTOPreLink(Phase)));
300 // TODO: Investigate promotion cap for O1.
301 LPM1.addPass(LICMPass(PTO.LicmMssaOptCap, PTO.LicmMssaNoAccForPromotionCap));
302 LPM1.addPass(SimpleLoopUnswitchPass());
304 LPM2.addPass(LoopIdiomRecognizePass());
305 LPM2.addPass(IndVarSimplifyPass());
307 for (auto &C : LateLoopOptimizationsEPCallbacks)
310 LPM2.addPass(LoopDeletionPass());
312 if (EnableLoopInterchange)
313 LPM2.addPass(LoopInterchangePass());
315 // Do not enable unrolling in PreLinkThinLTO phase during sample PGO
316 // because it changes IR to makes profile annotation in back compile
317 // inaccurate. The normal unroller doesn't pay attention to forced full unroll
318 // attributes so we need to make sure and allow the full unroll pass to pay
320 if (Phase != ThinOrFullLTOPhase::ThinLTOPreLink || !PGOOpt ||
321 PGOOpt->Action != PGOOptions::SampleUse)
322 LPM2.addPass(LoopFullUnrollPass(Level.getSpeedupLevel(),
323 /* OnlyWhenForced= */ !PTO.LoopUnrolling,
324 PTO.ForgetAllSCEVInLoopUnroll));
326 for (auto &C : LoopOptimizerEndEPCallbacks)
329 // We provide the opt remark emitter pass for LICM to use. We only need to do
330 // this once as it is immutable.
332 RequireAnalysisPass<OptimizationRemarkEmitterAnalysis, Function>());
333 FPM.addPass(createFunctionToLoopPassAdaptor(std::move(LPM1),
334 /*UseMemorySSA=*/true,
335 /*UseBlockFrequencyInfo=*/true));
336 FPM.addPass(SimplifyCFGPass());
337 FPM.addPass(InstCombinePass());
338 if (EnableLoopFlatten)
339 FPM.addPass(createFunctionToLoopPassAdaptor(LoopFlattenPass()));
340 // The loop passes in LPM2 (LoopFullUnrollPass) do not preserve MemorySSA.
341 // *All* loop passes must preserve it, in order to be able to use it.
342 FPM.addPass(createFunctionToLoopPassAdaptor(std::move(LPM2),
343 /*UseMemorySSA=*/false,
344 /*UseBlockFrequencyInfo=*/false));
346 // Delete small array after loop unroll.
347 FPM.addPass(SROAPass());
349 // Specially optimize memory movement as it doesn't look like dataflow in SSA.
350 FPM.addPass(MemCpyOptPass());
352 // Sparse conditional constant propagation.
353 // FIXME: It isn't clear why we do this *after* loop passes rather than
355 FPM.addPass(SCCPPass());
357 // Delete dead bit computations (instcombine runs after to fold away the dead
358 // computations, and then ADCE will run later to exploit any new DCE
359 // opportunities that creates).
360 FPM.addPass(BDCEPass());
362 // Run instcombine after redundancy and dead bit elimination to exploit
363 // opportunities opened up by them.
364 FPM.addPass(InstCombinePass());
365 invokePeepholeEPCallbacks(FPM, Level);
367 FPM.addPass(CoroElidePass());
369 for (auto &C : ScalarOptimizerLateEPCallbacks)
372 // Finally, do an expensive DCE pass to catch all the dead code exposed by
373 // the simplifications and basic cleanup after all the simplifications.
374 // TODO: Investigate if this is too expensive.
375 FPM.addPass(ADCEPass());
376 FPM.addPass(SimplifyCFGPass());
377 FPM.addPass(InstCombinePass());
378 invokePeepholeEPCallbacks(FPM, Level);
384 PassBuilder::buildFunctionSimplificationPipeline(OptimizationLevel Level,
385 ThinOrFullLTOPhase Phase) {
386 assert(Level != OptimizationLevel::O0 && "Must request optimizations!");
388 // The O1 pipeline has a separate pipeline creation function to simplify
389 // construction readability.
390 if (Level.getSpeedupLevel() == 1)
391 return buildO1FunctionSimplificationPipeline(Level, Phase);
393 FunctionPassManager FPM;
395 // Form SSA out of local memory accesses after breaking apart aggregates into
397 FPM.addPass(SROAPass());
399 // Catch trivial redundancies
400 FPM.addPass(EarlyCSEPass(true /* Enable mem-ssa. */));
401 if (EnableKnowledgeRetention)
402 FPM.addPass(AssumeSimplifyPass());
404 // Hoisting of scalars and load expressions.
406 FPM.addPass(GVNHoistPass());
408 // Global value numbering based sinking.
410 FPM.addPass(GVNSinkPass());
411 FPM.addPass(SimplifyCFGPass());
414 if (EnableConstraintElimination)
415 FPM.addPass(ConstraintEliminationPass());
417 // Speculative execution if the target has divergent branches; otherwise nop.
418 FPM.addPass(SpeculativeExecutionPass(/* OnlyIfDivergentTarget =*/true));
420 // Optimize based on known information about branches, and cleanup afterward.
421 FPM.addPass(JumpThreadingPass());
422 FPM.addPass(CorrelatedValuePropagationPass());
424 FPM.addPass(SimplifyCFGPass());
425 FPM.addPass(InstCombinePass());
426 if (Level == OptimizationLevel::O3)
427 FPM.addPass(AggressiveInstCombinePass());
429 if (!Level.isOptimizingForSize())
430 FPM.addPass(LibCallsShrinkWrapPass());
432 invokePeepholeEPCallbacks(FPM, Level);
434 // For PGO use pipeline, try to optimize memory intrinsics such as memcpy
435 // using the size value profile. Don't perform this when optimizing for size.
436 if (PGOOpt && PGOOpt->Action == PGOOptions::IRUse &&
437 !Level.isOptimizingForSize())
438 FPM.addPass(PGOMemOPSizeOpt());
440 FPM.addPass(TailCallElimPass());
441 FPM.addPass(SimplifyCFGPass());
443 // Form canonically associated expression trees, and simplify the trees using
444 // basic mathematical properties. For example, this will form (nearly)
445 // minimal multiplication trees.
446 FPM.addPass(ReassociatePass());
448 // Add the primary loop simplification pipeline.
449 // FIXME: Currently this is split into two loop pass pipelines because we run
450 // some function passes in between them. These can and should be removed
451 // and/or replaced by scheduling the loop pass equivalents in the correct
452 // positions. But those equivalent passes aren't powerful enough yet.
453 // Specifically, `SimplifyCFGPass` and `InstCombinePass` are currently still
454 // used. We have `LoopSimplifyCFGPass` which isn't yet powerful enough yet to
455 // fully replace `SimplifyCFGPass`, and the closest to the other we have is
456 // `LoopInstSimplify`.
457 LoopPassManager LPM1, LPM2;
459 // Simplify the loop body. We do this initially to clean up after other loop
460 // passes run, either when iterating on a loop or on inner loops with
461 // implications on the outer loop.
462 LPM1.addPass(LoopInstSimplifyPass());
463 LPM1.addPass(LoopSimplifyCFGPass());
465 // Try to remove as much code from the loop header as possible,
466 // to reduce amount of IR that will have to be duplicated.
467 // TODO: Investigate promotion cap for O1.
468 LPM1.addPass(LICMPass(PTO.LicmMssaOptCap, PTO.LicmMssaNoAccForPromotionCap));
470 // Disable header duplication in loop rotation at -Oz.
472 LoopRotatePass(Level != OptimizationLevel::Oz, isLTOPreLink(Phase)));
473 // TODO: Investigate promotion cap for O1.
474 LPM1.addPass(LICMPass(PTO.LicmMssaOptCap, PTO.LicmMssaNoAccForPromotionCap));
476 SimpleLoopUnswitchPass(/* NonTrivial */ Level == OptimizationLevel::O3 &&
477 EnableO3NonTrivialUnswitching));
478 LPM2.addPass(LoopIdiomRecognizePass());
479 LPM2.addPass(IndVarSimplifyPass());
481 for (auto &C : LateLoopOptimizationsEPCallbacks)
484 LPM2.addPass(LoopDeletionPass());
486 if (EnableLoopInterchange)
487 LPM2.addPass(LoopInterchangePass());
489 // Do not enable unrolling in PreLinkThinLTO phase during sample PGO
490 // because it changes IR to makes profile annotation in back compile
491 // inaccurate. The normal unroller doesn't pay attention to forced full unroll
492 // attributes so we need to make sure and allow the full unroll pass to pay
494 if (Phase != ThinOrFullLTOPhase::ThinLTOPreLink || !PGOOpt ||
495 PGOOpt->Action != PGOOptions::SampleUse)
496 LPM2.addPass(LoopFullUnrollPass(Level.getSpeedupLevel(),
497 /* OnlyWhenForced= */ !PTO.LoopUnrolling,
498 PTO.ForgetAllSCEVInLoopUnroll));
500 for (auto &C : LoopOptimizerEndEPCallbacks)
503 // We provide the opt remark emitter pass for LICM to use. We only need to do
504 // this once as it is immutable.
506 RequireAnalysisPass<OptimizationRemarkEmitterAnalysis, Function>());
507 FPM.addPass(createFunctionToLoopPassAdaptor(std::move(LPM1),
508 /*UseMemorySSA=*/true,
509 /*UseBlockFrequencyInfo=*/true));
510 FPM.addPass(SimplifyCFGPass());
511 FPM.addPass(InstCombinePass());
512 if (EnableLoopFlatten)
513 FPM.addPass(createFunctionToLoopPassAdaptor(LoopFlattenPass()));
514 // The loop passes in LPM2 (LoopIdiomRecognizePass, IndVarSimplifyPass,
515 // LoopDeletionPass and LoopFullUnrollPass) do not preserve MemorySSA.
516 // *All* loop passes must preserve it, in order to be able to use it.
517 FPM.addPass(createFunctionToLoopPassAdaptor(std::move(LPM2),
518 /*UseMemorySSA=*/false,
519 /*UseBlockFrequencyInfo=*/false));
521 // Delete small array after loop unroll.
522 FPM.addPass(SROAPass());
524 // The matrix extension can introduce large vector operations early, which can
525 // benefit from running vector-combine early on.
527 FPM.addPass(VectorCombinePass(/*ScalarizationOnly=*/true));
529 // Eliminate redundancies.
530 FPM.addPass(MergedLoadStoreMotionPass());
532 FPM.addPass(NewGVNPass());
534 FPM.addPass(GVNPass());
536 // Sparse conditional constant propagation.
537 // FIXME: It isn't clear why we do this *after* loop passes rather than
539 FPM.addPass(SCCPPass());
541 // Delete dead bit computations (instcombine runs after to fold away the dead
542 // computations, and then ADCE will run later to exploit any new DCE
543 // opportunities that creates).
544 FPM.addPass(BDCEPass());
546 // Run instcombine after redundancy and dead bit elimination to exploit
547 // opportunities opened up by them.
548 FPM.addPass(InstCombinePass());
549 invokePeepholeEPCallbacks(FPM, Level);
551 // Re-consider control flow based optimizations after redundancy elimination,
553 if (EnableDFAJumpThreading && Level.getSizeLevel() == 0)
554 FPM.addPass(DFAJumpThreadingPass());
556 FPM.addPass(JumpThreadingPass());
557 FPM.addPass(CorrelatedValuePropagationPass());
559 // Finally, do an expensive DCE pass to catch all the dead code exposed by
560 // the simplifications and basic cleanup after all the simplifications.
561 // TODO: Investigate if this is too expensive.
562 FPM.addPass(ADCEPass());
564 // Specially optimize memory movement as it doesn't look like dataflow in SSA.
565 FPM.addPass(MemCpyOptPass());
567 FPM.addPass(DSEPass());
568 FPM.addPass(createFunctionToLoopPassAdaptor(
569 LICMPass(PTO.LicmMssaOptCap, PTO.LicmMssaNoAccForPromotionCap),
570 /*UseMemorySSA=*/true, /*UseBlockFrequencyInfo=*/true));
572 FPM.addPass(CoroElidePass());
574 for (auto &C : ScalarOptimizerLateEPCallbacks)
577 FPM.addPass(SimplifyCFGPass(
578 SimplifyCFGOptions().hoistCommonInsts(true).sinkCommonInsts(true)));
579 FPM.addPass(InstCombinePass());
580 invokePeepholeEPCallbacks(FPM, Level);
582 if (EnableCHR && Level == OptimizationLevel::O3 && PGOOpt &&
583 (PGOOpt->Action == PGOOptions::IRUse ||
584 PGOOpt->Action == PGOOptions::SampleUse))
585 FPM.addPass(ControlHeightReductionPass());
590 void PassBuilder::addRequiredLTOPreLinkPasses(ModulePassManager &MPM) {
591 MPM.addPass(CanonicalizeAliasesPass());
592 MPM.addPass(NameAnonGlobalPass());
595 void PassBuilder::addPGOInstrPasses(ModulePassManager &MPM,
596 OptimizationLevel Level, bool RunProfileGen,
597 bool IsCS, std::string ProfileFile,
598 std::string ProfileRemappingFile) {
599 assert(Level != OptimizationLevel::O0 && "Not expecting O0 here!");
600 if (!IsCS && !DisablePreInliner) {
603 IP.DefaultThreshold = PreInlineThreshold;
605 // FIXME: The hint threshold has the same value used by the regular inliner
606 // when not optimzing for size. This should probably be lowered after
607 // performance testing.
608 // FIXME: this comment is cargo culted from the old pass manager, revisit).
609 IP.HintThreshold = Level.isOptimizingForSize() ? PreInlineThreshold : 325;
610 ModuleInlinerWrapperPass MIWP(IP);
611 CGSCCPassManager &CGPipeline = MIWP.getPM();
613 FunctionPassManager FPM;
614 FPM.addPass(SROAPass());
615 FPM.addPass(EarlyCSEPass()); // Catch trivial redundancies.
616 FPM.addPass(SimplifyCFGPass()); // Merge & remove basic blocks.
617 FPM.addPass(InstCombinePass()); // Combine silly sequences.
618 invokePeepholeEPCallbacks(FPM, Level);
620 CGPipeline.addPass(createCGSCCToFunctionPassAdaptor(
621 std::move(FPM), PTO.EagerlyInvalidateAnalyses));
623 MPM.addPass(std::move(MIWP));
625 // Delete anything that is now dead to make sure that we don't instrument
626 // dead code. Instrumentation can end up keeping dead code around and
627 // dramatically increase code size.
628 MPM.addPass(GlobalDCEPass());
631 if (!RunProfileGen) {
632 assert(!ProfileFile.empty() && "Profile use expecting a profile file!");
633 MPM.addPass(PGOInstrumentationUse(ProfileFile, ProfileRemappingFile, IsCS));
634 // Cache ProfileSummaryAnalysis once to avoid the potential need to insert
635 // RequireAnalysisPass for PSI before subsequent non-module passes.
636 MPM.addPass(RequireAnalysisPass<ProfileSummaryAnalysis, Module>());
640 // Perform PGO instrumentation.
641 MPM.addPass(PGOInstrumentationGen(IsCS));
643 FunctionPassManager FPM;
644 // Disable header duplication in loop rotation at -Oz.
645 FPM.addPass(createFunctionToLoopPassAdaptor(
646 LoopRotatePass(Level != OptimizationLevel::Oz), /*UseMemorySSA=*/false,
647 /*UseBlockFrequencyInfo=*/false));
648 MPM.addPass(createModuleToFunctionPassAdaptor(std::move(FPM),
649 PTO.EagerlyInvalidateAnalyses));
651 // Add the profile lowering pass.
652 InstrProfOptions Options;
653 if (!ProfileFile.empty())
654 Options.InstrProfileOutput = ProfileFile;
655 // Do counter promotion at Level greater than O0.
656 Options.DoCounterPromotion = true;
657 Options.UseBFIInPromotion = IsCS;
658 MPM.addPass(InstrProfiling(Options, IsCS));
661 void PassBuilder::addPGOInstrPassesForO0(ModulePassManager &MPM,
662 bool RunProfileGen, bool IsCS,
663 std::string ProfileFile,
664 std::string ProfileRemappingFile) {
665 if (!RunProfileGen) {
666 assert(!ProfileFile.empty() && "Profile use expecting a profile file!");
667 MPM.addPass(PGOInstrumentationUse(ProfileFile, ProfileRemappingFile, IsCS));
668 // Cache ProfileSummaryAnalysis once to avoid the potential need to insert
669 // RequireAnalysisPass for PSI before subsequent non-module passes.
670 MPM.addPass(RequireAnalysisPass<ProfileSummaryAnalysis, Module>());
674 // Perform PGO instrumentation.
675 MPM.addPass(PGOInstrumentationGen(IsCS));
676 // Add the profile lowering pass.
677 InstrProfOptions Options;
678 if (!ProfileFile.empty())
679 Options.InstrProfileOutput = ProfileFile;
680 // Do not do counter promotion at O0.
681 Options.DoCounterPromotion = false;
682 Options.UseBFIInPromotion = IsCS;
683 MPM.addPass(InstrProfiling(Options, IsCS));
686 static InlineParams getInlineParamsFromOptLevel(OptimizationLevel Level) {
687 return getInlineParams(Level.getSpeedupLevel(), Level.getSizeLevel());
690 ModuleInlinerWrapperPass
691 PassBuilder::buildInlinerPipeline(OptimizationLevel Level,
692 ThinOrFullLTOPhase Phase) {
693 InlineParams IP = getInlineParamsFromOptLevel(Level);
694 if (Phase == ThinOrFullLTOPhase::ThinLTOPreLink && PGOOpt &&
695 PGOOpt->Action == PGOOptions::SampleUse)
696 IP.HotCallSiteThreshold = 0;
699 IP.EnableDeferral = EnablePGOInlineDeferral;
701 ModuleInlinerWrapperPass MIWP(IP, PerformMandatoryInliningsFirst,
702 UseInlineAdvisor, MaxDevirtIterations);
704 // Require the GlobalsAA analysis for the module so we can query it within
705 // the CGSCC pipeline.
706 MIWP.addModulePass(RequireAnalysisPass<GlobalsAA, Module>());
707 // Invalidate AAManager so it can be recreated and pick up the newly available
710 createModuleToFunctionPassAdaptor(InvalidateAnalysisPass<AAManager>()));
712 // Require the ProfileSummaryAnalysis for the module so we can query it within
714 MIWP.addModulePass(RequireAnalysisPass<ProfileSummaryAnalysis, Module>());
716 // Now begin the main postorder CGSCC pipeline.
717 // FIXME: The current CGSCC pipeline has its origins in the legacy pass
718 // manager and trying to emulate its precise behavior. Much of this doesn't
719 // make a lot of sense and we should revisit the core CGSCC structure.
720 CGSCCPassManager &MainCGPipeline = MIWP.getPM();
722 // Note: historically, the PruneEH pass was run first to deduce nounwind and
723 // generally clean up exception handling overhead. It isn't clear this is
724 // valuable as the inliner doesn't currently care whether it is inlining an
727 if (AttributorRun & AttributorRunOption::CGSCC)
728 MainCGPipeline.addPass(AttributorCGSCCPass());
730 // Now deduce any function attributes based in the current code.
731 MainCGPipeline.addPass(PostOrderFunctionAttrsPass());
733 // When at O3 add argument promotion to the pass pipeline.
734 // FIXME: It isn't at all clear why this should be limited to O3.
735 if (Level == OptimizationLevel::O3)
736 MainCGPipeline.addPass(ArgumentPromotionPass());
738 // Try to perform OpenMP specific optimizations. This is a (quick!) no-op if
739 // there are no OpenMP runtime calls present in the module.
740 if (Level == OptimizationLevel::O2 || Level == OptimizationLevel::O3)
741 MainCGPipeline.addPass(OpenMPOptCGSCCPass());
743 for (auto &C : CGSCCOptimizerLateEPCallbacks)
744 C(MainCGPipeline, Level);
746 // Lastly, add the core function simplification pipeline nested inside the
748 MainCGPipeline.addPass(createCGSCCToFunctionPassAdaptor(
749 buildFunctionSimplificationPipeline(Level, Phase),
750 PTO.EagerlyInvalidateAnalyses, EnableNoRerunSimplificationPipeline));
752 MainCGPipeline.addPass(CoroSplitPass(Level != OptimizationLevel::O0));
754 if (EnableNoRerunSimplificationPipeline)
755 MIWP.addLateModulePass(createModuleToFunctionPassAdaptor(
756 InvalidateAnalysisPass<ShouldNotRunFunctionPassesAnalysis>()));
762 PassBuilder::buildModuleInlinerPipeline(OptimizationLevel Level,
763 ThinOrFullLTOPhase Phase) {
764 ModulePassManager MPM;
766 InlineParams IP = getInlineParamsFromOptLevel(Level);
767 if (Phase == ThinOrFullLTOPhase::ThinLTOPreLink && PGOOpt &&
768 PGOOpt->Action == PGOOptions::SampleUse)
769 IP.HotCallSiteThreshold = 0;
772 IP.EnableDeferral = EnablePGOInlineDeferral;
774 // The inline deferral logic is used to avoid losing some
775 // inlining chance in future. It is helpful in SCC inliner, in which
776 // inlining is processed in bottom-up order.
777 // While in module inliner, the inlining order is a priority-based order
778 // by default. The inline deferral is unnecessary there. So we disable the
779 // inline deferral logic in module inliner.
780 IP.EnableDeferral = false;
782 MPM.addPass(ModuleInlinerPass(IP, UseInlineAdvisor));
784 MPM.addPass(createModuleToFunctionPassAdaptor(
785 buildFunctionSimplificationPipeline(Level, Phase),
786 PTO.EagerlyInvalidateAnalyses));
788 MPM.addPass(createModuleToPostOrderCGSCCPassAdaptor(
789 CoroSplitPass(Level != OptimizationLevel::O0)));
795 PassBuilder::buildModuleSimplificationPipeline(OptimizationLevel Level,
796 ThinOrFullLTOPhase Phase) {
797 ModulePassManager MPM;
799 // Place pseudo probe instrumentation as the first pass of the pipeline to
800 // minimize the impact of optimization changes.
801 if (PGOOpt && PGOOpt->PseudoProbeForProfiling &&
802 Phase != ThinOrFullLTOPhase::ThinLTOPostLink)
803 MPM.addPass(SampleProfileProbePass(TM));
805 bool HasSampleProfile = PGOOpt && (PGOOpt->Action == PGOOptions::SampleUse);
807 // In ThinLTO mode, when flattened profile is used, all the available
808 // profile information will be annotated in PreLink phase so there is
809 // no need to load the profile again in PostLink.
810 bool LoadSampleProfile =
812 !(FlattenedProfileUsed && Phase == ThinOrFullLTOPhase::ThinLTOPostLink);
814 // During the ThinLTO backend phase we perform early indirect call promotion
815 // here, before globalopt. Otherwise imported available_externally functions
816 // look unreferenced and are removed. If we are going to load the sample
817 // profile then defer until later.
818 // TODO: See if we can move later and consolidate with the location where
819 // we perform ICP when we are loading a sample profile.
820 // TODO: We pass HasSampleProfile (whether there was a sample profile file
821 // passed to the compile) to the SamplePGO flag of ICP. This is used to
822 // determine whether the new direct calls are annotated with prof metadata.
823 // Ideally this should be determined from whether the IR is annotated with
824 // sample profile, and not whether the a sample profile was provided on the
825 // command line. E.g. for flattened profiles where we will not be reloading
826 // the sample profile in the ThinLTO backend, we ideally shouldn't have to
827 // provide the sample profile file.
828 if (Phase == ThinOrFullLTOPhase::ThinLTOPostLink && !LoadSampleProfile)
829 MPM.addPass(PGOIndirectCallPromotion(true /* InLTO */, HasSampleProfile));
831 // Do basic inference of function attributes from known properties of system
832 // libraries and other oracles.
833 MPM.addPass(InferFunctionAttrsPass());
835 // Create an early function pass manager to cleanup the output of the
837 FunctionPassManager EarlyFPM;
838 // Lower llvm.expect to metadata before attempting transforms.
839 // Compare/branch metadata may alter the behavior of passes like SimplifyCFG.
840 EarlyFPM.addPass(LowerExpectIntrinsicPass());
841 EarlyFPM.addPass(SimplifyCFGPass());
842 EarlyFPM.addPass(SROAPass());
843 EarlyFPM.addPass(EarlyCSEPass());
844 EarlyFPM.addPass(CoroEarlyPass());
845 if (Level == OptimizationLevel::O3)
846 EarlyFPM.addPass(CallSiteSplittingPass());
848 // In SamplePGO ThinLTO backend, we need instcombine before profile annotation
849 // to convert bitcast to direct calls so that they can be inlined during the
850 // profile annotation prepration step.
851 // More details about SamplePGO design can be found in:
852 // https://research.google.com/pubs/pub45290.html
853 // FIXME: revisit how SampleProfileLoad/Inliner/ICP is structured.
854 if (LoadSampleProfile)
855 EarlyFPM.addPass(InstCombinePass());
856 MPM.addPass(createModuleToFunctionPassAdaptor(std::move(EarlyFPM),
857 PTO.EagerlyInvalidateAnalyses));
859 if (LoadSampleProfile) {
860 // Annotate sample profile right after early FPM to ensure freshness of
862 MPM.addPass(SampleProfileLoaderPass(PGOOpt->ProfileFile,
863 PGOOpt->ProfileRemappingFile, Phase));
864 // Cache ProfileSummaryAnalysis once to avoid the potential need to insert
865 // RequireAnalysisPass for PSI before subsequent non-module passes.
866 MPM.addPass(RequireAnalysisPass<ProfileSummaryAnalysis, Module>());
867 // Do not invoke ICP in the LTOPrelink phase as it makes it hard
868 // for the profile annotation to be accurate in the LTO backend.
869 if (Phase != ThinOrFullLTOPhase::ThinLTOPreLink &&
870 Phase != ThinOrFullLTOPhase::FullLTOPreLink)
871 // We perform early indirect call promotion here, before globalopt.
872 // This is important for the ThinLTO backend phase because otherwise
873 // imported available_externally functions look unreferenced and are
876 PGOIndirectCallPromotion(true /* IsInLTO */, true /* SamplePGO */));
879 // Try to perform OpenMP specific optimizations on the module. This is a
880 // (quick!) no-op if there are no OpenMP runtime calls present in the module.
881 if (Level != OptimizationLevel::O0)
882 MPM.addPass(OpenMPOptPass());
884 if (AttributorRun & AttributorRunOption::MODULE)
885 MPM.addPass(AttributorPass());
887 // Lower type metadata and the type.test intrinsic in the ThinLTO
888 // post link pipeline after ICP. This is to enable usage of the type
889 // tests in ICP sequences.
890 if (Phase == ThinOrFullLTOPhase::ThinLTOPostLink)
891 MPM.addPass(LowerTypeTestsPass(nullptr, nullptr, true));
893 for (auto &C : PipelineEarlySimplificationEPCallbacks)
896 // Specialize functions with IPSCCP.
897 if (EnableFunctionSpecialization && Level == OptimizationLevel::O3)
898 MPM.addPass(FunctionSpecializationPass());
900 // Interprocedural constant propagation now that basic cleanup has occurred
901 // and prior to optimizing globals.
902 // FIXME: This position in the pipeline hasn't been carefully considered in
903 // years, it should be re-analyzed.
904 MPM.addPass(IPSCCPPass());
906 // Attach metadata to indirect call sites indicating the set of functions
907 // they may target at run-time. This should follow IPSCCP.
908 MPM.addPass(CalledValuePropagationPass());
910 // Optimize globals to try and fold them into constants.
911 MPM.addPass(GlobalOptPass());
913 // Promote any localized globals to SSA registers.
914 // FIXME: Should this instead by a run of SROA?
915 // FIXME: We should probably run instcombine and simplifycfg afterward to
916 // delete control flows that are dead once globals have been folded to
918 MPM.addPass(createModuleToFunctionPassAdaptor(PromotePass()));
920 // Remove any dead arguments exposed by cleanups and constant folding
922 MPM.addPass(DeadArgumentEliminationPass());
924 // Create a small function pass pipeline to cleanup after all the global
926 FunctionPassManager GlobalCleanupPM;
927 GlobalCleanupPM.addPass(InstCombinePass());
928 invokePeepholeEPCallbacks(GlobalCleanupPM, Level);
930 GlobalCleanupPM.addPass(SimplifyCFGPass());
931 MPM.addPass(createModuleToFunctionPassAdaptor(std::move(GlobalCleanupPM),
932 PTO.EagerlyInvalidateAnalyses));
934 // Add all the requested passes for instrumentation PGO, if requested.
935 if (PGOOpt && Phase != ThinOrFullLTOPhase::ThinLTOPostLink &&
936 (PGOOpt->Action == PGOOptions::IRInstr ||
937 PGOOpt->Action == PGOOptions::IRUse)) {
938 addPGOInstrPasses(MPM, Level,
939 /* RunProfileGen */ PGOOpt->Action == PGOOptions::IRInstr,
940 /* IsCS */ false, PGOOpt->ProfileFile,
941 PGOOpt->ProfileRemappingFile);
942 MPM.addPass(PGOIndirectCallPromotion(false, false));
944 if (PGOOpt && Phase != ThinOrFullLTOPhase::ThinLTOPostLink &&
945 PGOOpt->CSAction == PGOOptions::CSIRInstr)
946 MPM.addPass(PGOInstrumentationGenCreateVar(PGOOpt->CSProfileGenFile));
948 // Synthesize function entry counts for non-PGO compilation.
949 if (EnableSyntheticCounts && !PGOOpt)
950 MPM.addPass(SyntheticCountsPropagation());
952 if (EnableModuleInliner)
953 MPM.addPass(buildModuleInlinerPipeline(Level, Phase));
955 MPM.addPass(buildInlinerPipeline(Level, Phase));
957 if (EnableMemProfiler && Phase != ThinOrFullLTOPhase::ThinLTOPreLink) {
958 MPM.addPass(createModuleToFunctionPassAdaptor(MemProfilerPass()));
959 MPM.addPass(ModuleMemProfilerPass());
965 /// TODO: Should LTO cause any differences to this set of passes?
966 void PassBuilder::addVectorPasses(OptimizationLevel Level,
967 FunctionPassManager &FPM, bool IsFullLTO) {
968 FPM.addPass(LoopVectorizePass(
969 LoopVectorizeOptions(!PTO.LoopInterleaving, !PTO.LoopVectorization)));
972 // The vectorizer may have significantly shortened a loop body; unroll
973 // again. Unroll small loops to hide loop backedge latency and saturate any
974 // parallel execution resources of an out-of-order processor. We also then
975 // need to clean up redundancies and loop invariant code.
976 // FIXME: It would be really good to use a loop-integrated instruction
977 // combiner for cleanup here so that the unrolling and LICM can be pipelined
978 // across the loop nests.
979 // We do UnrollAndJam in a separate LPM to ensure it happens before unroll
980 if (EnableUnrollAndJam && PTO.LoopUnrolling)
981 FPM.addPass(createFunctionToLoopPassAdaptor(
982 LoopUnrollAndJamPass(Level.getSpeedupLevel())));
983 FPM.addPass(LoopUnrollPass(LoopUnrollOptions(
984 Level.getSpeedupLevel(), /*OnlyWhenForced=*/!PTO.LoopUnrolling,
985 PTO.ForgetAllSCEVInLoopUnroll)));
986 FPM.addPass(WarnMissedTransformationsPass());
990 // Eliminate loads by forwarding stores from the previous iteration to loads
991 // of the current iteration.
992 FPM.addPass(LoopLoadEliminationPass());
994 // Cleanup after the loop optimization passes.
995 FPM.addPass(InstCombinePass());
997 if (Level.getSpeedupLevel() > 1 && ExtraVectorizerPasses) {
998 ExtraVectorPassManager ExtraPasses;
999 // At higher optimization levels, try to clean up any runtime overlap and
1000 // alignment checks inserted by the vectorizer. We want to track correlated
1001 // runtime checks for two inner loops in the same outer loop, fold any
1002 // common computations, hoist loop-invariant aspects out of any outer loop,
1003 // and unswitch the runtime checks if possible. Once hoisted, we may have
1004 // dead (or speculatable) control flows or more combining opportunities.
1005 ExtraPasses.addPass(EarlyCSEPass());
1006 ExtraPasses.addPass(CorrelatedValuePropagationPass());
1007 ExtraPasses.addPass(InstCombinePass());
1008 LoopPassManager LPM;
1009 LPM.addPass(LICMPass(PTO.LicmMssaOptCap, PTO.LicmMssaNoAccForPromotionCap));
1010 LPM.addPass(SimpleLoopUnswitchPass(/* NonTrivial */ Level ==
1011 OptimizationLevel::O3));
1012 ExtraPasses.addPass(
1013 RequireAnalysisPass<OptimizationRemarkEmitterAnalysis, Function>());
1014 ExtraPasses.addPass(
1015 createFunctionToLoopPassAdaptor(std::move(LPM), /*UseMemorySSA=*/true,
1016 /*UseBlockFrequencyInfo=*/true));
1017 ExtraPasses.addPass(SimplifyCFGPass());
1018 ExtraPasses.addPass(InstCombinePass());
1019 FPM.addPass(std::move(ExtraPasses));
1022 // Now that we've formed fast to execute loop structures, we do further
1023 // optimizations. These are run afterward as they might block doing complex
1024 // analyses and transforms such as what are needed for loop vectorization.
1026 // Cleanup after loop vectorization, etc. Simplification passes like CVP and
1027 // GVN, loop transforms, and others have already run, so it's now better to
1028 // convert to more optimized IR using more aggressive simplify CFG options.
1029 // The extra sinking transform can create larger basic blocks, so do this
1030 // before SLP vectorization.
1031 FPM.addPass(SimplifyCFGPass(SimplifyCFGOptions()
1032 .forwardSwitchCondToPhi(true)
1033 .convertSwitchToLookupTable(true)
1034 .needCanonicalLoops(false)
1035 .hoistCommonInsts(true)
1036 .sinkCommonInsts(true)));
1039 FPM.addPass(SCCPPass());
1040 FPM.addPass(InstCombinePass());
1041 FPM.addPass(BDCEPass());
1044 // Optimize parallel scalar instruction chains into SIMD instructions.
1045 if (PTO.SLPVectorization) {
1046 FPM.addPass(SLPVectorizerPass());
1047 if (Level.getSpeedupLevel() > 1 && ExtraVectorizerPasses) {
1048 FPM.addPass(EarlyCSEPass());
1051 // Enhance/cleanup vector code.
1052 FPM.addPass(VectorCombinePass());
1055 FPM.addPass(InstCombinePass());
1056 // Unroll small loops to hide loop backedge latency and saturate any
1057 // parallel execution resources of an out-of-order processor. We also then
1058 // need to clean up redundancies and loop invariant code.
1059 // FIXME: It would be really good to use a loop-integrated instruction
1060 // combiner for cleanup here so that the unrolling and LICM can be pipelined
1061 // across the loop nests.
1062 // We do UnrollAndJam in a separate LPM to ensure it happens before unroll
1063 if (EnableUnrollAndJam && PTO.LoopUnrolling) {
1064 FPM.addPass(createFunctionToLoopPassAdaptor(
1065 LoopUnrollAndJamPass(Level.getSpeedupLevel())));
1067 FPM.addPass(LoopUnrollPass(LoopUnrollOptions(
1068 Level.getSpeedupLevel(), /*OnlyWhenForced=*/!PTO.LoopUnrolling,
1069 PTO.ForgetAllSCEVInLoopUnroll)));
1070 FPM.addPass(WarnMissedTransformationsPass());
1071 FPM.addPass(InstCombinePass());
1073 RequireAnalysisPass<OptimizationRemarkEmitterAnalysis, Function>());
1074 FPM.addPass(createFunctionToLoopPassAdaptor(
1075 LICMPass(PTO.LicmMssaOptCap, PTO.LicmMssaNoAccForPromotionCap),
1076 /*UseMemorySSA=*/true, /*UseBlockFrequencyInfo=*/true));
1079 // Now that we've vectorized and unrolled loops, we may have more refined
1080 // alignment information, try to re-derive it here.
1081 FPM.addPass(AlignmentFromAssumptionsPass());
1084 FPM.addPass(InstCombinePass());
1088 PassBuilder::buildModuleOptimizationPipeline(OptimizationLevel Level,
1090 ModulePassManager MPM;
1092 // Optimize globals now that the module is fully simplified.
1093 MPM.addPass(GlobalOptPass());
1094 MPM.addPass(GlobalDCEPass());
1096 // Run partial inlining pass to partially inline functions that have
1098 if (RunPartialInlining)
1099 MPM.addPass(PartialInlinerPass());
1101 // Remove avail extern fns and globals definitions since we aren't compiling
1102 // an object file for later LTO. For LTO we want to preserve these so they
1103 // are eligible for inlining at link-time. Note if they are unreferenced they
1104 // will be removed by GlobalDCE later, so this only impacts referenced
1105 // available externally globals. Eventually they will be suppressed during
1106 // codegen, but eliminating here enables more opportunity for GlobalDCE as it
1107 // may make globals referenced by available external functions dead and saves
1108 // running remaining passes on the eliminated functions. These should be
1109 // preserved during prelinking for link-time inlining decisions.
1111 MPM.addPass(EliminateAvailableExternallyPass());
1113 if (EnableOrderFileInstrumentation)
1114 MPM.addPass(InstrOrderFilePass());
1116 // Do RPO function attribute inference across the module to forward-propagate
1117 // attributes where applicable.
1118 // FIXME: Is this really an optimization rather than a canonicalization?
1119 MPM.addPass(ReversePostOrderFunctionAttrsPass());
1121 // Do a post inline PGO instrumentation and use pass. This is a context
1122 // sensitive PGO pass. We don't want to do this in LTOPreLink phrase as
1123 // cross-module inline has not been done yet. The context sensitive
1124 // instrumentation is after all the inlines are done.
1125 if (!LTOPreLink && PGOOpt) {
1126 if (PGOOpt->CSAction == PGOOptions::CSIRInstr)
1127 addPGOInstrPasses(MPM, Level, /* RunProfileGen */ true,
1128 /* IsCS */ true, PGOOpt->CSProfileGenFile,
1129 PGOOpt->ProfileRemappingFile);
1130 else if (PGOOpt->CSAction == PGOOptions::CSIRUse)
1131 addPGOInstrPasses(MPM, Level, /* RunProfileGen */ false,
1132 /* IsCS */ true, PGOOpt->ProfileFile,
1133 PGOOpt->ProfileRemappingFile);
1136 // Re-require GloblasAA here prior to function passes. This is particularly
1137 // useful as the above will have inlined, DCE'ed, and function-attr
1138 // propagated everything. We should at this point have a reasonably minimal
1139 // and richly annotated call graph. By computing aliasing and mod/ref
1140 // information for all local globals here, the late loop passes and notably
1141 // the vectorizer will be able to use them to help recognize vectorizable
1142 // memory operations.
1143 MPM.addPass(RequireAnalysisPass<GlobalsAA, Module>());
1145 FunctionPassManager OptimizePM;
1146 OptimizePM.addPass(Float2IntPass());
1147 OptimizePM.addPass(LowerConstantIntrinsicsPass());
1150 OptimizePM.addPass(LowerMatrixIntrinsicsPass());
1151 OptimizePM.addPass(EarlyCSEPass());
1154 // FIXME: We need to run some loop optimizations to re-rotate loops after
1155 // simplifycfg and others undo their rotation.
1157 // Optimize the loop execution. These passes operate on entire loop nests
1158 // rather than on each loop in an inside-out manner, and so they are actually
1161 for (auto &C : VectorizerStartEPCallbacks)
1162 C(OptimizePM, Level);
1164 LoopPassManager LPM;
1165 // First rotate loops that may have been un-rotated by prior passes.
1166 // Disable header duplication at -Oz.
1167 LPM.addPass(LoopRotatePass(Level != OptimizationLevel::Oz, LTOPreLink));
1168 // Some loops may have become dead by now. Try to delete them.
1169 // FIXME: see discussion in https://reviews.llvm.org/D112851,
1170 // this may need to be revisited once we run GVN before loop deletion
1171 // in the simplification pipeline.
1172 LPM.addPass(LoopDeletionPass());
1173 OptimizePM.addPass(createFunctionToLoopPassAdaptor(
1174 std::move(LPM), /*UseMemorySSA=*/false, /*UseBlockFrequencyInfo=*/false));
1176 // Distribute loops to allow partial vectorization. I.e. isolate dependences
1177 // into separate loop that would otherwise inhibit vectorization. This is
1178 // currently only performed for loops marked with the metadata
1179 // llvm.loop.distribute=true or when -enable-loop-distribute is specified.
1180 OptimizePM.addPass(LoopDistributePass());
1182 // Populates the VFABI attribute with the scalar-to-vector mappings
1183 // from the TargetLibraryInfo.
1184 OptimizePM.addPass(InjectTLIMappings());
1186 addVectorPasses(Level, OptimizePM, /* IsFullLTO */ false);
1188 // LoopSink pass sinks instructions hoisted by LICM, which serves as a
1189 // canonicalization pass that enables other optimizations. As a result,
1190 // LoopSink pass needs to be a very late IR pass to avoid undoing LICM
1191 // result too early.
1192 OptimizePM.addPass(LoopSinkPass());
1194 // And finally clean up LCSSA form before generating code.
1195 OptimizePM.addPass(InstSimplifyPass());
1197 // This hoists/decomposes div/rem ops. It should run after other sink/hoist
1198 // passes to avoid re-sinking, but before SimplifyCFG because it can allow
1199 // flattening of blocks.
1200 OptimizePM.addPass(DivRemPairsPass());
1202 // LoopSink (and other loop passes since the last simplifyCFG) might have
1203 // resulted in single-entry-single-exit or empty blocks. Clean up the CFG.
1204 OptimizePM.addPass(SimplifyCFGPass());
1206 OptimizePM.addPass(CoroCleanupPass());
1208 // Add the core optimizing pipeline.
1209 MPM.addPass(createModuleToFunctionPassAdaptor(std::move(OptimizePM),
1210 PTO.EagerlyInvalidateAnalyses));
1212 for (auto &C : OptimizerLastEPCallbacks)
1215 // Split out cold code. Splitting is done late to avoid hiding context from
1216 // other optimizations and inadvertently regressing performance. The tradeoff
1217 // is that this has a higher code size cost than splitting early.
1218 if (EnableHotColdSplit && !LTOPreLink)
1219 MPM.addPass(HotColdSplittingPass());
1221 // Search the code for similar regions of code. If enough similar regions can
1222 // be found where extracting the regions into their own function will decrease
1223 // the size of the program, we extract the regions, a deduplicate the
1224 // structurally similar regions.
1225 if (EnableIROutliner)
1226 MPM.addPass(IROutlinerPass());
1228 // Merge functions if requested.
1229 if (PTO.MergeFunctions)
1230 MPM.addPass(MergeFunctionsPass());
1232 if (PTO.CallGraphProfile)
1233 MPM.addPass(CGProfilePass());
1235 // Now we need to do some global optimization transforms.
1236 // FIXME: It would seem like these should come first in the optimization
1237 // pipeline and maybe be the bottom of the canonicalization pipeline? Weird
1239 MPM.addPass(GlobalDCEPass());
1240 MPM.addPass(ConstantMergePass());
1242 // TODO: Relative look table converter pass caused an issue when full lto is
1243 // enabled. See https://reviews.llvm.org/D94355 for more details.
1244 // Until the issue fixed, disable this pass during pre-linking phase.
1246 MPM.addPass(RelLookupTableConverterPass());
1252 PassBuilder::buildPerModuleDefaultPipeline(OptimizationLevel Level,
1254 assert(Level != OptimizationLevel::O0 &&
1255 "Must request optimizations for the default pipeline!");
1257 ModulePassManager MPM;
1259 // Convert @llvm.global.annotations to !annotation metadata.
1260 MPM.addPass(Annotation2MetadataPass());
1262 // Force any function attributes we want the rest of the pipeline to observe.
1263 MPM.addPass(ForceFunctionAttrsPass());
1265 // Apply module pipeline start EP callback.
1266 for (auto &C : PipelineStartEPCallbacks)
1269 if (PGOOpt && PGOOpt->DebugInfoForProfiling)
1270 MPM.addPass(createModuleToFunctionPassAdaptor(AddDiscriminatorsPass()));
1272 // Add the core simplification pipeline.
1273 MPM.addPass(buildModuleSimplificationPipeline(
1274 Level, LTOPreLink ? ThinOrFullLTOPhase::FullLTOPreLink
1275 : ThinOrFullLTOPhase::None));
1277 // Now add the optimization pipeline.
1278 MPM.addPass(buildModuleOptimizationPipeline(Level, LTOPreLink));
1280 if (PGOOpt && PGOOpt->PseudoProbeForProfiling &&
1281 PGOOpt->Action == PGOOptions::SampleUse)
1282 MPM.addPass(PseudoProbeUpdatePass());
1284 // Emit annotation remarks.
1285 addAnnotationRemarksPass(MPM);
1288 addRequiredLTOPreLinkPasses(MPM);
1294 PassBuilder::buildThinLTOPreLinkDefaultPipeline(OptimizationLevel Level) {
1295 assert(Level != OptimizationLevel::O0 &&
1296 "Must request optimizations for the default pipeline!");
1298 ModulePassManager MPM;
1300 // Convert @llvm.global.annotations to !annotation metadata.
1301 MPM.addPass(Annotation2MetadataPass());
1303 // Force any function attributes we want the rest of the pipeline to observe.
1304 MPM.addPass(ForceFunctionAttrsPass());
1306 if (PGOOpt && PGOOpt->DebugInfoForProfiling)
1307 MPM.addPass(createModuleToFunctionPassAdaptor(AddDiscriminatorsPass()));
1309 // Apply module pipeline start EP callback.
1310 for (auto &C : PipelineStartEPCallbacks)
1313 // If we are planning to perform ThinLTO later, we don't bloat the code with
1314 // unrolling/vectorization/... now. Just simplify the module as much as we
1316 MPM.addPass(buildModuleSimplificationPipeline(
1317 Level, ThinOrFullLTOPhase::ThinLTOPreLink));
1319 // Run partial inlining pass to partially inline functions that have
1321 // FIXME: It isn't clear whether this is really the right place to run this
1322 // in ThinLTO. Because there is another canonicalization and simplification
1323 // phase that will run after the thin link, running this here ends up with
1324 // less information than will be available later and it may grow functions in
1325 // ways that aren't beneficial.
1326 if (RunPartialInlining)
1327 MPM.addPass(PartialInlinerPass());
1329 // Reduce the size of the IR as much as possible.
1330 MPM.addPass(GlobalOptPass());
1332 // Module simplification splits coroutines, but does not fully clean up
1333 // coroutine intrinsics. To ensure ThinLTO optimization passes don't trip up
1334 // on these, we schedule the cleanup here.
1335 MPM.addPass(createModuleToFunctionPassAdaptor(CoroCleanupPass()));
1337 if (PGOOpt && PGOOpt->PseudoProbeForProfiling &&
1338 PGOOpt->Action == PGOOptions::SampleUse)
1339 MPM.addPass(PseudoProbeUpdatePass());
1341 // Handle OptimizerLastEPCallbacks added by clang on PreLink. Actual
1342 // optimization is going to be done in PostLink stage, but clang can't
1343 // add callbacks there in case of in-process ThinLTO called by linker.
1344 for (auto &C : OptimizerLastEPCallbacks)
1347 // Emit annotation remarks.
1348 addAnnotationRemarksPass(MPM);
1350 addRequiredLTOPreLinkPasses(MPM);
1355 ModulePassManager PassBuilder::buildThinLTODefaultPipeline(
1356 OptimizationLevel Level, const ModuleSummaryIndex *ImportSummary) {
1357 ModulePassManager MPM;
1359 // Convert @llvm.global.annotations to !annotation metadata.
1360 MPM.addPass(Annotation2MetadataPass());
1362 if (ImportSummary) {
1363 // These passes import type identifier resolutions for whole-program
1364 // devirtualization and CFI. They must run early because other passes may
1365 // disturb the specific instruction patterns that these passes look for,
1366 // creating dependencies on resolutions that may not appear in the summary.
1368 // For example, GVN may transform the pattern assume(type.test) appearing in
1369 // two basic blocks into assume(phi(type.test, type.test)), which would
1370 // transform a dependency on a WPD resolution into a dependency on a type
1371 // identifier resolution for CFI.
1373 // Also, WPD has access to more precise information than ICP and can
1374 // devirtualize more effectively, so it should operate on the IR first.
1376 // The WPD and LowerTypeTest passes need to run at -O0 to lower type
1377 // metadata and intrinsics.
1378 MPM.addPass(WholeProgramDevirtPass(nullptr, ImportSummary));
1379 MPM.addPass(LowerTypeTestsPass(nullptr, ImportSummary));
1382 if (Level == OptimizationLevel::O0) {
1383 // Run a second time to clean up any type tests left behind by WPD for use
1385 MPM.addPass(LowerTypeTestsPass(nullptr, nullptr, true));
1386 // Drop available_externally and unreferenced globals. This is necessary
1387 // with ThinLTO in order to avoid leaving undefined references to dead
1388 // globals in the object file.
1389 MPM.addPass(EliminateAvailableExternallyPass());
1390 MPM.addPass(GlobalDCEPass());
1394 // Force any function attributes we want the rest of the pipeline to observe.
1395 MPM.addPass(ForceFunctionAttrsPass());
1397 // Add the core simplification pipeline.
1398 MPM.addPass(buildModuleSimplificationPipeline(
1399 Level, ThinOrFullLTOPhase::ThinLTOPostLink));
1401 // Now add the optimization pipeline.
1402 MPM.addPass(buildModuleOptimizationPipeline(Level));
1404 // Emit annotation remarks.
1405 addAnnotationRemarksPass(MPM);
1411 PassBuilder::buildLTOPreLinkDefaultPipeline(OptimizationLevel Level) {
1412 assert(Level != OptimizationLevel::O0 &&
1413 "Must request optimizations for the default pipeline!");
1414 // FIXME: We should use a customized pre-link pipeline!
1415 return buildPerModuleDefaultPipeline(Level,
1416 /* LTOPreLink */ true);
1420 PassBuilder::buildLTODefaultPipeline(OptimizationLevel Level,
1421 ModuleSummaryIndex *ExportSummary) {
1422 ModulePassManager MPM;
1424 // Convert @llvm.global.annotations to !annotation metadata.
1425 MPM.addPass(Annotation2MetadataPass());
1427 // Create a function that performs CFI checks for cross-DSO calls with targets
1428 // in the current module.
1429 MPM.addPass(CrossDSOCFIPass());
1431 if (Level == OptimizationLevel::O0) {
1432 // The WPD and LowerTypeTest passes need to run at -O0 to lower type
1433 // metadata and intrinsics.
1434 MPM.addPass(WholeProgramDevirtPass(ExportSummary, nullptr));
1435 MPM.addPass(LowerTypeTestsPass(ExportSummary, nullptr));
1436 // Run a second time to clean up any type tests left behind by WPD for use
1438 MPM.addPass(LowerTypeTestsPass(nullptr, nullptr, true));
1440 // Emit annotation remarks.
1441 addAnnotationRemarksPass(MPM);
1446 if (PGOOpt && PGOOpt->Action == PGOOptions::SampleUse) {
1447 // Load sample profile before running the LTO optimization pipeline.
1448 MPM.addPass(SampleProfileLoaderPass(PGOOpt->ProfileFile,
1449 PGOOpt->ProfileRemappingFile,
1450 ThinOrFullLTOPhase::FullLTOPostLink));
1451 // Cache ProfileSummaryAnalysis once to avoid the potential need to insert
1452 // RequireAnalysisPass for PSI before subsequent non-module passes.
1453 MPM.addPass(RequireAnalysisPass<ProfileSummaryAnalysis, Module>());
1456 // Remove unused virtual tables to improve the quality of code generated by
1457 // whole-program devirtualization and bitset lowering.
1458 MPM.addPass(GlobalDCEPass());
1460 // Force any function attributes we want the rest of the pipeline to observe.
1461 MPM.addPass(ForceFunctionAttrsPass());
1463 // Do basic inference of function attributes from known properties of system
1464 // libraries and other oracles.
1465 MPM.addPass(InferFunctionAttrsPass());
1467 if (Level.getSpeedupLevel() > 1) {
1468 FunctionPassManager EarlyFPM;
1469 EarlyFPM.addPass(CallSiteSplittingPass());
1470 MPM.addPass(createModuleToFunctionPassAdaptor(
1471 std::move(EarlyFPM), PTO.EagerlyInvalidateAnalyses));
1473 // Indirect call promotion. This should promote all the targets that are
1474 // left by the earlier promotion pass that promotes intra-module targets.
1475 // This two-step promotion is to save the compile time. For LTO, it should
1476 // produce the same result as if we only do promotion here.
1477 MPM.addPass(PGOIndirectCallPromotion(
1478 true /* InLTO */, PGOOpt && PGOOpt->Action == PGOOptions::SampleUse));
1480 if (EnableFunctionSpecialization && Level == OptimizationLevel::O3)
1481 MPM.addPass(FunctionSpecializationPass());
1482 // Propagate constants at call sites into the functions they call. This
1483 // opens opportunities for globalopt (and inlining) by substituting function
1484 // pointers passed as arguments to direct uses of functions.
1485 MPM.addPass(IPSCCPPass());
1487 // Attach metadata to indirect call sites indicating the set of functions
1488 // they may target at run-time. This should follow IPSCCP.
1489 MPM.addPass(CalledValuePropagationPass());
1492 // Now deduce any function attributes based in the current code.
1494 createModuleToPostOrderCGSCCPassAdaptor(PostOrderFunctionAttrsPass()));
1496 // Do RPO function attribute inference across the module to forward-propagate
1497 // attributes where applicable.
1498 // FIXME: Is this really an optimization rather than a canonicalization?
1499 MPM.addPass(ReversePostOrderFunctionAttrsPass());
1501 // Use in-range annotations on GEP indices to split globals where beneficial.
1502 MPM.addPass(GlobalSplitPass());
1504 // Run whole program optimization of virtual call when the list of callees
1506 MPM.addPass(WholeProgramDevirtPass(ExportSummary, nullptr));
1508 // Stop here at -O1.
1509 if (Level == OptimizationLevel::O1) {
1510 // The LowerTypeTestsPass needs to run to lower type metadata and the
1511 // type.test intrinsics. The pass does nothing if CFI is disabled.
1512 MPM.addPass(LowerTypeTestsPass(ExportSummary, nullptr));
1513 // Run a second time to clean up any type tests left behind by WPD for use
1514 // in ICP (which is performed earlier than this in the regular LTO
1516 MPM.addPass(LowerTypeTestsPass(nullptr, nullptr, true));
1518 // Emit annotation remarks.
1519 addAnnotationRemarksPass(MPM);
1524 // Optimize globals to try and fold them into constants.
1525 MPM.addPass(GlobalOptPass());
1527 // Promote any localized globals to SSA registers.
1528 MPM.addPass(createModuleToFunctionPassAdaptor(PromotePass()));
1530 // Linking modules together can lead to duplicate global constant, only
1531 // keep one copy of each constant.
1532 MPM.addPass(ConstantMergePass());
1534 // Remove unused arguments from functions.
1535 MPM.addPass(DeadArgumentEliminationPass());
1537 // Reduce the code after globalopt and ipsccp. Both can open up significant
1538 // simplification opportunities, and both can propagate functions through
1539 // function pointers. When this happens, we often have to resolve varargs
1540 // calls, etc, so let instcombine do this.
1541 FunctionPassManager PeepholeFPM;
1542 PeepholeFPM.addPass(InstCombinePass());
1543 if (Level == OptimizationLevel::O3)
1544 PeepholeFPM.addPass(AggressiveInstCombinePass());
1545 invokePeepholeEPCallbacks(PeepholeFPM, Level);
1547 MPM.addPass(createModuleToFunctionPassAdaptor(std::move(PeepholeFPM),
1548 PTO.EagerlyInvalidateAnalyses));
1550 // Note: historically, the PruneEH pass was run first to deduce nounwind and
1551 // generally clean up exception handling overhead. It isn't clear this is
1552 // valuable as the inliner doesn't currently care whether it is inlining an
1553 // invoke or a call.
1554 // Run the inliner now.
1555 MPM.addPass(ModuleInlinerWrapperPass(getInlineParamsFromOptLevel(Level)));
1557 // Optimize globals again after we ran the inliner.
1558 MPM.addPass(GlobalOptPass());
1560 // Garbage collect dead functions.
1561 MPM.addPass(GlobalDCEPass());
1563 // If we didn't decide to inline a function, check to see if we can
1564 // transform it to pass arguments by value instead of by reference.
1565 MPM.addPass(createModuleToPostOrderCGSCCPassAdaptor(ArgumentPromotionPass()));
1567 FunctionPassManager FPM;
1568 // The IPO Passes may leave cruft around. Clean up after them.
1569 FPM.addPass(InstCombinePass());
1570 invokePeepholeEPCallbacks(FPM, Level);
1572 FPM.addPass(JumpThreadingPass(/*InsertFreezeWhenUnfoldingSelect*/ true));
1574 // Do a post inline PGO instrumentation and use pass. This is a context
1575 // sensitive PGO pass.
1577 if (PGOOpt->CSAction == PGOOptions::CSIRInstr)
1578 addPGOInstrPasses(MPM, Level, /* RunProfileGen */ true,
1579 /* IsCS */ true, PGOOpt->CSProfileGenFile,
1580 PGOOpt->ProfileRemappingFile);
1581 else if (PGOOpt->CSAction == PGOOptions::CSIRUse)
1582 addPGOInstrPasses(MPM, Level, /* RunProfileGen */ false,
1583 /* IsCS */ true, PGOOpt->ProfileFile,
1584 PGOOpt->ProfileRemappingFile);
1588 FPM.addPass(SROAPass());
1590 // LTO provides additional opportunities for tailcall elimination due to
1591 // link-time inlining, and visibility of nocapture attribute.
1592 FPM.addPass(TailCallElimPass());
1594 // Run a few AA driver optimizations here and now to cleanup the code.
1595 MPM.addPass(createModuleToFunctionPassAdaptor(std::move(FPM),
1596 PTO.EagerlyInvalidateAnalyses));
1599 createModuleToPostOrderCGSCCPassAdaptor(PostOrderFunctionAttrsPass()));
1601 // Require the GlobalsAA analysis for the module so we can query it within
1603 MPM.addPass(RequireAnalysisPass<GlobalsAA, Module>());
1604 // Invalidate AAManager so it can be recreated and pick up the newly available
1607 createModuleToFunctionPassAdaptor(InvalidateAnalysisPass<AAManager>()));
1609 FunctionPassManager MainFPM;
1610 MainFPM.addPass(createFunctionToLoopPassAdaptor(
1611 LICMPass(PTO.LicmMssaOptCap, PTO.LicmMssaNoAccForPromotionCap),
1612 /*USeMemorySSA=*/true, /*UseBlockFrequencyInfo=*/true));
1615 MainFPM.addPass(NewGVNPass());
1617 MainFPM.addPass(GVNPass());
1619 // Remove dead memcpy()'s.
1620 MainFPM.addPass(MemCpyOptPass());
1622 // Nuke dead stores.
1623 MainFPM.addPass(DSEPass());
1624 MainFPM.addPass(MergedLoadStoreMotionPass());
1626 // More loops are countable; try to optimize them.
1627 if (EnableLoopFlatten && Level.getSpeedupLevel() > 1)
1628 MainFPM.addPass(createFunctionToLoopPassAdaptor(LoopFlattenPass()));
1630 if (EnableConstraintElimination)
1631 MainFPM.addPass(ConstraintEliminationPass());
1633 LoopPassManager LPM;
1634 LPM.addPass(IndVarSimplifyPass());
1635 LPM.addPass(LoopDeletionPass());
1636 // FIXME: Add loop interchange.
1638 // Unroll small loops and perform peeling.
1639 LPM.addPass(LoopFullUnrollPass(Level.getSpeedupLevel(),
1640 /* OnlyWhenForced= */ !PTO.LoopUnrolling,
1641 PTO.ForgetAllSCEVInLoopUnroll));
1642 // The loop passes in LPM (LoopFullUnrollPass) do not preserve MemorySSA.
1643 // *All* loop passes must preserve it, in order to be able to use it.
1644 MainFPM.addPass(createFunctionToLoopPassAdaptor(
1645 std::move(LPM), /*UseMemorySSA=*/false, /*UseBlockFrequencyInfo=*/true));
1647 MainFPM.addPass(LoopDistributePass());
1649 addVectorPasses(Level, MainFPM, /* IsFullLTO */ true);
1651 invokePeepholeEPCallbacks(MainFPM, Level);
1652 MainFPM.addPass(JumpThreadingPass(/*InsertFreezeWhenUnfoldingSelect*/ true));
1653 MPM.addPass(createModuleToFunctionPassAdaptor(std::move(MainFPM),
1654 PTO.EagerlyInvalidateAnalyses));
1656 // Lower type metadata and the type.test intrinsic. This pass supports
1657 // clang's control flow integrity mechanisms (-fsanitize=cfi*) and needs
1658 // to be run at link time if CFI is enabled. This pass does nothing if
1660 MPM.addPass(LowerTypeTestsPass(ExportSummary, nullptr));
1661 // Run a second time to clean up any type tests left behind by WPD for use
1662 // in ICP (which is performed earlier than this in the regular LTO pipeline).
1663 MPM.addPass(LowerTypeTestsPass(nullptr, nullptr, true));
1665 // Enable splitting late in the FullLTO post-link pipeline. This is done in
1666 // the same stage in the old pass manager (\ref addLateLTOOptimizationPasses).
1667 if (EnableHotColdSplit)
1668 MPM.addPass(HotColdSplittingPass());
1670 // Add late LTO optimization passes.
1671 // Delete basic blocks, which optimization passes may have killed.
1672 MPM.addPass(createModuleToFunctionPassAdaptor(
1673 SimplifyCFGPass(SimplifyCFGOptions().hoistCommonInsts(true))));
1675 // Drop bodies of available eternally objects to improve GlobalDCE.
1676 MPM.addPass(EliminateAvailableExternallyPass());
1678 // Now that we have optimized the program, discard unreachable functions.
1679 MPM.addPass(GlobalDCEPass());
1681 if (PTO.MergeFunctions)
1682 MPM.addPass(MergeFunctionsPass());
1684 // Emit annotation remarks.
1685 addAnnotationRemarksPass(MPM);
1690 ModulePassManager PassBuilder::buildO0DefaultPipeline(OptimizationLevel Level,
1692 assert(Level == OptimizationLevel::O0 &&
1693 "buildO0DefaultPipeline should only be used with O0");
1695 ModulePassManager MPM;
1697 // Perform pseudo probe instrumentation in O0 mode. This is for the
1698 // consistency between different build modes. For example, a LTO build can be
1699 // mixed with an O0 prelink and an O2 postlink. Loading a sample profile in
1700 // the postlink will require pseudo probe instrumentation in the prelink.
1701 if (PGOOpt && PGOOpt->PseudoProbeForProfiling)
1702 MPM.addPass(SampleProfileProbePass(TM));
1704 if (PGOOpt && (PGOOpt->Action == PGOOptions::IRInstr ||
1705 PGOOpt->Action == PGOOptions::IRUse))
1706 addPGOInstrPassesForO0(
1708 /* RunProfileGen */ (PGOOpt->Action == PGOOptions::IRInstr),
1709 /* IsCS */ false, PGOOpt->ProfileFile, PGOOpt->ProfileRemappingFile);
1711 for (auto &C : PipelineStartEPCallbacks)
1714 if (PGOOpt && PGOOpt->DebugInfoForProfiling)
1715 MPM.addPass(createModuleToFunctionPassAdaptor(AddDiscriminatorsPass()));
1717 for (auto &C : PipelineEarlySimplificationEPCallbacks)
1720 // Build a minimal pipeline based on the semantics required by LLVM,
1721 // which is just that always inlining occurs. Further, disable generating
1722 // lifetime intrinsics to avoid enabling further optimizations during
1724 MPM.addPass(AlwaysInlinerPass(
1725 /*InsertLifetimeIntrinsics=*/false));
1727 if (PTO.MergeFunctions)
1728 MPM.addPass(MergeFunctionsPass());
1732 createModuleToFunctionPassAdaptor(LowerMatrixIntrinsicsPass(true)));
1734 if (!CGSCCOptimizerLateEPCallbacks.empty()) {
1735 CGSCCPassManager CGPM;
1736 for (auto &C : CGSCCOptimizerLateEPCallbacks)
1738 if (!CGPM.isEmpty())
1739 MPM.addPass(createModuleToPostOrderCGSCCPassAdaptor(std::move(CGPM)));
1741 if (!LateLoopOptimizationsEPCallbacks.empty()) {
1742 LoopPassManager LPM;
1743 for (auto &C : LateLoopOptimizationsEPCallbacks)
1745 if (!LPM.isEmpty()) {
1746 MPM.addPass(createModuleToFunctionPassAdaptor(
1747 createFunctionToLoopPassAdaptor(std::move(LPM))));
1750 if (!LoopOptimizerEndEPCallbacks.empty()) {
1751 LoopPassManager LPM;
1752 for (auto &C : LoopOptimizerEndEPCallbacks)
1754 if (!LPM.isEmpty()) {
1755 MPM.addPass(createModuleToFunctionPassAdaptor(
1756 createFunctionToLoopPassAdaptor(std::move(LPM))));
1759 if (!ScalarOptimizerLateEPCallbacks.empty()) {
1760 FunctionPassManager FPM;
1761 for (auto &C : ScalarOptimizerLateEPCallbacks)
1764 MPM.addPass(createModuleToFunctionPassAdaptor(std::move(FPM)));
1766 if (!VectorizerStartEPCallbacks.empty()) {
1767 FunctionPassManager FPM;
1768 for (auto &C : VectorizerStartEPCallbacks)
1771 MPM.addPass(createModuleToFunctionPassAdaptor(std::move(FPM)));
1774 MPM.addPass(createModuleToFunctionPassAdaptor(CoroEarlyPass()));
1775 CGSCCPassManager CGPM;
1776 CGPM.addPass(CoroSplitPass());
1777 MPM.addPass(createModuleToPostOrderCGSCCPassAdaptor(std::move(CGPM)));
1778 MPM.addPass(createModuleToFunctionPassAdaptor(CoroCleanupPass()));
1780 for (auto &C : OptimizerLastEPCallbacks)
1784 addRequiredLTOPreLinkPasses(MPM);
1786 MPM.addPass(createModuleToFunctionPassAdaptor(AnnotationRemarksPass()));
1791 AAManager PassBuilder::buildDefaultAAPipeline() {
1794 // The order in which these are registered determines their priority when
1797 // First we register the basic alias analysis that provides the majority of
1798 // per-function local AA logic. This is a stateless, on-demand local set of
1800 AA.registerFunctionAnalysis<BasicAA>();
1802 // Next we query fast, specialized alias analyses that wrap IR-embedded
1803 // information about aliasing.
1804 AA.registerFunctionAnalysis<ScopedNoAliasAA>();
1805 AA.registerFunctionAnalysis<TypeBasedAA>();
1807 // Add support for querying global aliasing information when available.
1808 // Because the `AAManager` is a function analysis and `GlobalsAA` is a module
1809 // analysis, all that the `AAManager` can do is query for any *cached*
1810 // results from `GlobalsAA` through a readonly proxy.
1811 AA.registerModuleAnalysis<GlobalsAA>();
1813 // Add target-specific alias analyses.
1815 TM->registerDefaultAliasAnalyses(AA);