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 PipelineTuningOptions::PipelineTuningOptions() {
182 LoopInterleaving = true;
183 LoopVectorization = true;
184 SLPVectorization = false;
185 LoopUnrolling = true;
186 ForgetAllSCEVInLoopUnroll = ForgetSCEVInLoopUnroll;
187 LicmMssaOptCap = SetLicmMssaOptCap;
188 LicmMssaNoAccForPromotionCap = SetLicmMssaNoAccForPromotionCap;
189 CallGraphProfile = true;
190 MergeFunctions = false;
191 EagerlyInvalidateAnalyses = EnableEagerlyInvalidateAnalyses;
196 extern cl::opt<unsigned> MaxDevirtIterations;
197 extern cl::opt<bool> EnableConstraintElimination;
198 extern cl::opt<bool> EnableFunctionSpecialization;
199 extern cl::opt<bool> EnableGVNHoist;
200 extern cl::opt<bool> EnableGVNSink;
201 extern cl::opt<bool> EnableHotColdSplit;
202 extern cl::opt<bool> EnableIROutliner;
203 extern cl::opt<bool> EnableOrderFileInstrumentation;
204 extern cl::opt<bool> EnableCHR;
205 extern cl::opt<bool> EnableLoopInterchange;
206 extern cl::opt<bool> EnableUnrollAndJam;
207 extern cl::opt<bool> EnableLoopFlatten;
208 extern cl::opt<bool> EnableDFAJumpThreading;
209 extern cl::opt<bool> RunNewGVN;
210 extern cl::opt<bool> RunPartialInlining;
211 extern cl::opt<bool> ExtraVectorizerPasses;
213 extern cl::opt<bool> FlattenedProfileUsed;
215 extern cl::opt<AttributorRunOption> AttributorRun;
216 extern cl::opt<bool> EnableKnowledgeRetention;
218 extern cl::opt<bool> EnableMatrix;
220 extern cl::opt<bool> DisablePreInliner;
221 extern cl::opt<int> PreInlineThreshold;
224 void PassBuilder::invokePeepholeEPCallbacks(FunctionPassManager &FPM,
225 OptimizationLevel Level) {
226 for (auto &C : PeepholeEPCallbacks)
230 // Helper to add AnnotationRemarksPass.
231 static void addAnnotationRemarksPass(ModulePassManager &MPM) {
232 FunctionPassManager FPM;
233 FPM.addPass(AnnotationRemarksPass());
234 MPM.addPass(createModuleToFunctionPassAdaptor(std::move(FPM)));
237 // Helper to check if the current compilation phase is preparing for LTO
238 static bool isLTOPreLink(ThinOrFullLTOPhase Phase) {
239 return Phase == ThinOrFullLTOPhase::ThinLTOPreLink ||
240 Phase == ThinOrFullLTOPhase::FullLTOPreLink;
243 // TODO: Investigate the cost/benefit of tail call elimination on debugging.
245 PassBuilder::buildO1FunctionSimplificationPipeline(OptimizationLevel Level,
246 ThinOrFullLTOPhase Phase) {
248 FunctionPassManager FPM;
250 // Form SSA out of local memory accesses after breaking apart aggregates into
252 FPM.addPass(SROAPass());
254 // Catch trivial redundancies
255 FPM.addPass(EarlyCSEPass(true /* Enable mem-ssa. */));
257 // Hoisting of scalars and load expressions.
258 FPM.addPass(SimplifyCFGPass());
259 FPM.addPass(InstCombinePass());
261 FPM.addPass(LibCallsShrinkWrapPass());
263 invokePeepholeEPCallbacks(FPM, Level);
265 FPM.addPass(SimplifyCFGPass());
267 // Form canonically associated expression trees, and simplify the trees using
268 // basic mathematical properties. For example, this will form (nearly)
269 // minimal multiplication trees.
270 FPM.addPass(ReassociatePass());
272 // Add the primary loop simplification pipeline.
273 // FIXME: Currently this is split into two loop pass pipelines because we run
274 // some function passes in between them. These can and should be removed
275 // and/or replaced by scheduling the loop pass equivalents in the correct
276 // positions. But those equivalent passes aren't powerful enough yet.
277 // Specifically, `SimplifyCFGPass` and `InstCombinePass` are currently still
278 // used. We have `LoopSimplifyCFGPass` which isn't yet powerful enough yet to
279 // fully replace `SimplifyCFGPass`, and the closest to the other we have is
280 // `LoopInstSimplify`.
281 LoopPassManager LPM1, LPM2;
283 // Simplify the loop body. We do this initially to clean up after other loop
284 // passes run, either when iterating on a loop or on inner loops with
285 // implications on the outer loop.
286 LPM1.addPass(LoopInstSimplifyPass());
287 LPM1.addPass(LoopSimplifyCFGPass());
289 // Try to remove as much code from the loop header as possible,
290 // to reduce amount of IR that will have to be duplicated.
291 // TODO: Investigate promotion cap for O1.
292 LPM1.addPass(LICMPass(PTO.LicmMssaOptCap, PTO.LicmMssaNoAccForPromotionCap));
294 LPM1.addPass(LoopRotatePass(/* Disable header duplication */ true,
295 isLTOPreLink(Phase)));
296 // TODO: Investigate promotion cap for O1.
297 LPM1.addPass(LICMPass(PTO.LicmMssaOptCap, PTO.LicmMssaNoAccForPromotionCap));
298 LPM1.addPass(SimpleLoopUnswitchPass());
300 LPM2.addPass(LoopIdiomRecognizePass());
301 LPM2.addPass(IndVarSimplifyPass());
303 for (auto &C : LateLoopOptimizationsEPCallbacks)
306 LPM2.addPass(LoopDeletionPass());
308 if (EnableLoopInterchange)
309 LPM2.addPass(LoopInterchangePass());
311 // Do not enable unrolling in PreLinkThinLTO phase during sample PGO
312 // because it changes IR to makes profile annotation in back compile
313 // inaccurate. The normal unroller doesn't pay attention to forced full unroll
314 // attributes so we need to make sure and allow the full unroll pass to pay
316 if (Phase != ThinOrFullLTOPhase::ThinLTOPreLink || !PGOOpt ||
317 PGOOpt->Action != PGOOptions::SampleUse)
318 LPM2.addPass(LoopFullUnrollPass(Level.getSpeedupLevel(),
319 /* OnlyWhenForced= */ !PTO.LoopUnrolling,
320 PTO.ForgetAllSCEVInLoopUnroll));
322 for (auto &C : LoopOptimizerEndEPCallbacks)
325 // We provide the opt remark emitter pass for LICM to use. We only need to do
326 // this once as it is immutable.
328 RequireAnalysisPass<OptimizationRemarkEmitterAnalysis, Function>());
329 FPM.addPass(createFunctionToLoopPassAdaptor(std::move(LPM1),
330 /*UseMemorySSA=*/true,
331 /*UseBlockFrequencyInfo=*/true));
332 FPM.addPass(SimplifyCFGPass());
333 FPM.addPass(InstCombinePass());
334 if (EnableLoopFlatten)
335 FPM.addPass(createFunctionToLoopPassAdaptor(LoopFlattenPass()));
336 // The loop passes in LPM2 (LoopFullUnrollPass) do not preserve MemorySSA.
337 // *All* loop passes must preserve it, in order to be able to use it.
338 FPM.addPass(createFunctionToLoopPassAdaptor(std::move(LPM2),
339 /*UseMemorySSA=*/false,
340 /*UseBlockFrequencyInfo=*/false));
342 // Delete small array after loop unroll.
343 FPM.addPass(SROAPass());
345 // Specially optimize memory movement as it doesn't look like dataflow in SSA.
346 FPM.addPass(MemCpyOptPass());
348 // Sparse conditional constant propagation.
349 // FIXME: It isn't clear why we do this *after* loop passes rather than
351 FPM.addPass(SCCPPass());
353 // Delete dead bit computations (instcombine runs after to fold away the dead
354 // computations, and then ADCE will run later to exploit any new DCE
355 // opportunities that creates).
356 FPM.addPass(BDCEPass());
358 // Run instcombine after redundancy and dead bit elimination to exploit
359 // opportunities opened up by them.
360 FPM.addPass(InstCombinePass());
361 invokePeepholeEPCallbacks(FPM, Level);
363 FPM.addPass(CoroElidePass());
365 for (auto &C : ScalarOptimizerLateEPCallbacks)
368 // Finally, do an expensive DCE pass to catch all the dead code exposed by
369 // the simplifications and basic cleanup after all the simplifications.
370 // TODO: Investigate if this is too expensive.
371 FPM.addPass(ADCEPass());
372 FPM.addPass(SimplifyCFGPass());
373 FPM.addPass(InstCombinePass());
374 invokePeepholeEPCallbacks(FPM, Level);
380 PassBuilder::buildFunctionSimplificationPipeline(OptimizationLevel Level,
381 ThinOrFullLTOPhase Phase) {
382 assert(Level != OptimizationLevel::O0 && "Must request optimizations!");
384 // The O1 pipeline has a separate pipeline creation function to simplify
385 // construction readability.
386 if (Level.getSpeedupLevel() == 1)
387 return buildO1FunctionSimplificationPipeline(Level, Phase);
389 FunctionPassManager FPM;
391 // Form SSA out of local memory accesses after breaking apart aggregates into
393 FPM.addPass(SROAPass());
395 // Catch trivial redundancies
396 FPM.addPass(EarlyCSEPass(true /* Enable mem-ssa. */));
397 if (EnableKnowledgeRetention)
398 FPM.addPass(AssumeSimplifyPass());
400 // Hoisting of scalars and load expressions.
402 FPM.addPass(GVNHoistPass());
404 // Global value numbering based sinking.
406 FPM.addPass(GVNSinkPass());
407 FPM.addPass(SimplifyCFGPass());
410 if (EnableConstraintElimination)
411 FPM.addPass(ConstraintEliminationPass());
413 // Speculative execution if the target has divergent branches; otherwise nop.
414 FPM.addPass(SpeculativeExecutionPass(/* OnlyIfDivergentTarget =*/true));
416 // Optimize based on known information about branches, and cleanup afterward.
417 FPM.addPass(JumpThreadingPass());
418 FPM.addPass(CorrelatedValuePropagationPass());
420 FPM.addPass(SimplifyCFGPass());
421 if (Level == OptimizationLevel::O3)
422 FPM.addPass(AggressiveInstCombinePass());
423 FPM.addPass(InstCombinePass());
425 if (!Level.isOptimizingForSize())
426 FPM.addPass(LibCallsShrinkWrapPass());
428 invokePeepholeEPCallbacks(FPM, Level);
430 // For PGO use pipeline, try to optimize memory intrinsics such as memcpy
431 // using the size value profile. Don't perform this when optimizing for size.
432 if (PGOOpt && PGOOpt->Action == PGOOptions::IRUse &&
433 !Level.isOptimizingForSize())
434 FPM.addPass(PGOMemOPSizeOpt());
436 FPM.addPass(TailCallElimPass());
437 FPM.addPass(SimplifyCFGPass());
439 // Form canonically associated expression trees, and simplify the trees using
440 // basic mathematical properties. For example, this will form (nearly)
441 // minimal multiplication trees.
442 FPM.addPass(ReassociatePass());
444 // Add the primary loop simplification pipeline.
445 // FIXME: Currently this is split into two loop pass pipelines because we run
446 // some function passes in between them. These can and should be removed
447 // and/or replaced by scheduling the loop pass equivalents in the correct
448 // positions. But those equivalent passes aren't powerful enough yet.
449 // Specifically, `SimplifyCFGPass` and `InstCombinePass` are currently still
450 // used. We have `LoopSimplifyCFGPass` which isn't yet powerful enough yet to
451 // fully replace `SimplifyCFGPass`, and the closest to the other we have is
452 // `LoopInstSimplify`.
453 LoopPassManager LPM1, LPM2;
455 // Simplify the loop body. We do this initially to clean up after other loop
456 // passes run, either when iterating on a loop or on inner loops with
457 // implications on the outer loop.
458 LPM1.addPass(LoopInstSimplifyPass());
459 LPM1.addPass(LoopSimplifyCFGPass());
461 // Try to remove as much code from the loop header as possible,
462 // to reduce amount of IR that will have to be duplicated.
463 // TODO: Investigate promotion cap for O1.
464 LPM1.addPass(LICMPass(PTO.LicmMssaOptCap, PTO.LicmMssaNoAccForPromotionCap));
466 // Disable header duplication in loop rotation at -Oz.
468 LoopRotatePass(Level != OptimizationLevel::Oz, isLTOPreLink(Phase)));
469 // TODO: Investigate promotion cap for O1.
470 LPM1.addPass(LICMPass(PTO.LicmMssaOptCap, PTO.LicmMssaNoAccForPromotionCap));
472 SimpleLoopUnswitchPass(/* NonTrivial */ Level == OptimizationLevel::O3 &&
473 EnableO3NonTrivialUnswitching));
474 LPM2.addPass(LoopIdiomRecognizePass());
475 LPM2.addPass(IndVarSimplifyPass());
477 for (auto &C : LateLoopOptimizationsEPCallbacks)
480 LPM2.addPass(LoopDeletionPass());
482 if (EnableLoopInterchange)
483 LPM2.addPass(LoopInterchangePass());
485 // Do not enable unrolling in PreLinkThinLTO phase during sample PGO
486 // because it changes IR to makes profile annotation in back compile
487 // inaccurate. The normal unroller doesn't pay attention to forced full unroll
488 // attributes so we need to make sure and allow the full unroll pass to pay
490 if (Phase != ThinOrFullLTOPhase::ThinLTOPreLink || !PGOOpt ||
491 PGOOpt->Action != PGOOptions::SampleUse)
492 LPM2.addPass(LoopFullUnrollPass(Level.getSpeedupLevel(),
493 /* OnlyWhenForced= */ !PTO.LoopUnrolling,
494 PTO.ForgetAllSCEVInLoopUnroll));
496 for (auto &C : LoopOptimizerEndEPCallbacks)
499 // We provide the opt remark emitter pass for LICM to use. We only need to do
500 // this once as it is immutable.
502 RequireAnalysisPass<OptimizationRemarkEmitterAnalysis, Function>());
503 FPM.addPass(createFunctionToLoopPassAdaptor(std::move(LPM1),
504 /*UseMemorySSA=*/true,
505 /*UseBlockFrequencyInfo=*/true));
506 FPM.addPass(SimplifyCFGPass());
507 FPM.addPass(InstCombinePass());
508 if (EnableLoopFlatten)
509 FPM.addPass(createFunctionToLoopPassAdaptor(LoopFlattenPass()));
510 // The loop passes in LPM2 (LoopIdiomRecognizePass, IndVarSimplifyPass,
511 // LoopDeletionPass and LoopFullUnrollPass) do not preserve MemorySSA.
512 // *All* loop passes must preserve it, in order to be able to use it.
513 FPM.addPass(createFunctionToLoopPassAdaptor(std::move(LPM2),
514 /*UseMemorySSA=*/false,
515 /*UseBlockFrequencyInfo=*/false));
517 // Delete small array after loop unroll.
518 FPM.addPass(SROAPass());
520 // The matrix extension can introduce large vector operations early, which can
521 // benefit from running vector-combine early on.
523 FPM.addPass(VectorCombinePass(/*ScalarizationOnly=*/true));
525 // Eliminate redundancies.
526 FPM.addPass(MergedLoadStoreMotionPass());
528 FPM.addPass(NewGVNPass());
530 FPM.addPass(GVNPass());
532 // Sparse conditional constant propagation.
533 // FIXME: It isn't clear why we do this *after* loop passes rather than
535 FPM.addPass(SCCPPass());
537 // Delete dead bit computations (instcombine runs after to fold away the dead
538 // computations, and then ADCE will run later to exploit any new DCE
539 // opportunities that creates).
540 FPM.addPass(BDCEPass());
542 // Run instcombine after redundancy and dead bit elimination to exploit
543 // opportunities opened up by them.
544 FPM.addPass(InstCombinePass());
545 invokePeepholeEPCallbacks(FPM, Level);
547 // Re-consider control flow based optimizations after redundancy elimination,
549 if (EnableDFAJumpThreading && Level.getSizeLevel() == 0)
550 FPM.addPass(DFAJumpThreadingPass());
552 FPM.addPass(JumpThreadingPass());
553 FPM.addPass(CorrelatedValuePropagationPass());
555 // Finally, do an expensive DCE pass to catch all the dead code exposed by
556 // the simplifications and basic cleanup after all the simplifications.
557 // TODO: Investigate if this is too expensive.
558 FPM.addPass(ADCEPass());
560 // Specially optimize memory movement as it doesn't look like dataflow in SSA.
561 FPM.addPass(MemCpyOptPass());
563 FPM.addPass(DSEPass());
564 FPM.addPass(createFunctionToLoopPassAdaptor(
565 LICMPass(PTO.LicmMssaOptCap, PTO.LicmMssaNoAccForPromotionCap),
566 /*UseMemorySSA=*/true, /*UseBlockFrequencyInfo=*/true));
568 FPM.addPass(CoroElidePass());
570 for (auto &C : ScalarOptimizerLateEPCallbacks)
573 FPM.addPass(SimplifyCFGPass(
574 SimplifyCFGOptions().hoistCommonInsts(true).sinkCommonInsts(true)));
575 FPM.addPass(InstCombinePass());
576 invokePeepholeEPCallbacks(FPM, Level);
578 if (EnableCHR && Level == OptimizationLevel::O3 && PGOOpt &&
579 (PGOOpt->Action == PGOOptions::IRUse ||
580 PGOOpt->Action == PGOOptions::SampleUse))
581 FPM.addPass(ControlHeightReductionPass());
586 void PassBuilder::addRequiredLTOPreLinkPasses(ModulePassManager &MPM) {
587 MPM.addPass(CanonicalizeAliasesPass());
588 MPM.addPass(NameAnonGlobalPass());
591 void PassBuilder::addPGOInstrPasses(ModulePassManager &MPM,
592 OptimizationLevel Level, bool RunProfileGen,
593 bool IsCS, std::string ProfileFile,
594 std::string ProfileRemappingFile) {
595 assert(Level != OptimizationLevel::O0 && "Not expecting O0 here!");
596 if (!IsCS && !DisablePreInliner) {
599 IP.DefaultThreshold = PreInlineThreshold;
601 // FIXME: The hint threshold has the same value used by the regular inliner
602 // when not optimzing for size. This should probably be lowered after
603 // performance testing.
604 // FIXME: this comment is cargo culted from the old pass manager, revisit).
605 IP.HintThreshold = Level.isOptimizingForSize() ? PreInlineThreshold : 325;
606 ModuleInlinerWrapperPass MIWP(IP);
607 CGSCCPassManager &CGPipeline = MIWP.getPM();
609 FunctionPassManager FPM;
610 FPM.addPass(SROAPass());
611 FPM.addPass(EarlyCSEPass()); // Catch trivial redundancies.
612 FPM.addPass(SimplifyCFGPass()); // Merge & remove basic blocks.
613 FPM.addPass(InstCombinePass()); // Combine silly sequences.
614 invokePeepholeEPCallbacks(FPM, Level);
616 CGPipeline.addPass(createCGSCCToFunctionPassAdaptor(
617 std::move(FPM), PTO.EagerlyInvalidateAnalyses));
619 MPM.addPass(std::move(MIWP));
621 // Delete anything that is now dead to make sure that we don't instrument
622 // dead code. Instrumentation can end up keeping dead code around and
623 // dramatically increase code size.
624 MPM.addPass(GlobalDCEPass());
627 if (!RunProfileGen) {
628 assert(!ProfileFile.empty() && "Profile use expecting a profile file!");
629 MPM.addPass(PGOInstrumentationUse(ProfileFile, ProfileRemappingFile, IsCS));
630 // Cache ProfileSummaryAnalysis once to avoid the potential need to insert
631 // RequireAnalysisPass for PSI before subsequent non-module passes.
632 MPM.addPass(RequireAnalysisPass<ProfileSummaryAnalysis, Module>());
636 // Perform PGO instrumentation.
637 MPM.addPass(PGOInstrumentationGen(IsCS));
639 FunctionPassManager FPM;
640 // Disable header duplication in loop rotation at -Oz.
641 FPM.addPass(createFunctionToLoopPassAdaptor(
642 LoopRotatePass(Level != OptimizationLevel::Oz), /*UseMemorySSA=*/false,
643 /*UseBlockFrequencyInfo=*/false));
644 MPM.addPass(createModuleToFunctionPassAdaptor(std::move(FPM),
645 PTO.EagerlyInvalidateAnalyses));
647 // Add the profile lowering pass.
648 InstrProfOptions Options;
649 if (!ProfileFile.empty())
650 Options.InstrProfileOutput = ProfileFile;
651 // Do counter promotion at Level greater than O0.
652 Options.DoCounterPromotion = true;
653 Options.UseBFIInPromotion = IsCS;
654 MPM.addPass(InstrProfiling(Options, IsCS));
657 void PassBuilder::addPGOInstrPassesForO0(ModulePassManager &MPM,
658 bool RunProfileGen, bool IsCS,
659 std::string ProfileFile,
660 std::string ProfileRemappingFile) {
661 if (!RunProfileGen) {
662 assert(!ProfileFile.empty() && "Profile use expecting a profile file!");
663 MPM.addPass(PGOInstrumentationUse(ProfileFile, ProfileRemappingFile, IsCS));
664 // Cache ProfileSummaryAnalysis once to avoid the potential need to insert
665 // RequireAnalysisPass for PSI before subsequent non-module passes.
666 MPM.addPass(RequireAnalysisPass<ProfileSummaryAnalysis, Module>());
670 // Perform PGO instrumentation.
671 MPM.addPass(PGOInstrumentationGen(IsCS));
672 // Add the profile lowering pass.
673 InstrProfOptions Options;
674 if (!ProfileFile.empty())
675 Options.InstrProfileOutput = ProfileFile;
676 // Do not do counter promotion at O0.
677 Options.DoCounterPromotion = false;
678 Options.UseBFIInPromotion = IsCS;
679 MPM.addPass(InstrProfiling(Options, IsCS));
682 static InlineParams getInlineParamsFromOptLevel(OptimizationLevel Level) {
683 return getInlineParams(Level.getSpeedupLevel(), Level.getSizeLevel());
686 ModuleInlinerWrapperPass
687 PassBuilder::buildInlinerPipeline(OptimizationLevel Level,
688 ThinOrFullLTOPhase Phase) {
689 InlineParams IP = getInlineParamsFromOptLevel(Level);
690 if (Phase == ThinOrFullLTOPhase::ThinLTOPreLink && PGOOpt &&
691 PGOOpt->Action == PGOOptions::SampleUse)
692 IP.HotCallSiteThreshold = 0;
695 IP.EnableDeferral = EnablePGOInlineDeferral;
697 ModuleInlinerWrapperPass MIWP(IP, PerformMandatoryInliningsFirst,
698 UseInlineAdvisor, MaxDevirtIterations);
700 // Require the GlobalsAA analysis for the module so we can query it within
701 // the CGSCC pipeline.
702 MIWP.addModulePass(RequireAnalysisPass<GlobalsAA, Module>());
703 // Invalidate AAManager so it can be recreated and pick up the newly available
706 createModuleToFunctionPassAdaptor(InvalidateAnalysisPass<AAManager>()));
708 // Require the ProfileSummaryAnalysis for the module so we can query it within
710 MIWP.addModulePass(RequireAnalysisPass<ProfileSummaryAnalysis, Module>());
712 // Now begin the main postorder CGSCC pipeline.
713 // FIXME: The current CGSCC pipeline has its origins in the legacy pass
714 // manager and trying to emulate its precise behavior. Much of this doesn't
715 // make a lot of sense and we should revisit the core CGSCC structure.
716 CGSCCPassManager &MainCGPipeline = MIWP.getPM();
718 // Note: historically, the PruneEH pass was run first to deduce nounwind and
719 // generally clean up exception handling overhead. It isn't clear this is
720 // valuable as the inliner doesn't currently care whether it is inlining an
723 if (AttributorRun & AttributorRunOption::CGSCC)
724 MainCGPipeline.addPass(AttributorCGSCCPass());
726 // Now deduce any function attributes based in the current code.
727 MainCGPipeline.addPass(PostOrderFunctionAttrsPass());
729 // When at O3 add argument promotion to the pass pipeline.
730 // FIXME: It isn't at all clear why this should be limited to O3.
731 if (Level == OptimizationLevel::O3)
732 MainCGPipeline.addPass(ArgumentPromotionPass());
734 // Try to perform OpenMP specific optimizations. This is a (quick!) no-op if
735 // there are no OpenMP runtime calls present in the module.
736 if (Level == OptimizationLevel::O2 || Level == OptimizationLevel::O3)
737 MainCGPipeline.addPass(OpenMPOptCGSCCPass());
739 for (auto &C : CGSCCOptimizerLateEPCallbacks)
740 C(MainCGPipeline, Level);
742 // Lastly, add the core function simplification pipeline nested inside the
744 MainCGPipeline.addPass(createCGSCCToFunctionPassAdaptor(
745 buildFunctionSimplificationPipeline(Level, Phase),
746 PTO.EagerlyInvalidateAnalyses, EnableNoRerunSimplificationPipeline));
748 MainCGPipeline.addPass(CoroSplitPass(Level != OptimizationLevel::O0));
750 if (EnableNoRerunSimplificationPipeline)
751 MIWP.addLateModulePass(createModuleToFunctionPassAdaptor(
752 InvalidateAnalysisPass<ShouldNotRunFunctionPassesAnalysis>()));
758 PassBuilder::buildModuleInlinerPipeline(OptimizationLevel Level,
759 ThinOrFullLTOPhase Phase) {
760 InlineParams IP = getInlineParamsFromOptLevel(Level);
761 if (Phase == ThinOrFullLTOPhase::ThinLTOPreLink && PGOOpt &&
762 PGOOpt->Action == PGOOptions::SampleUse)
763 IP.HotCallSiteThreshold = 0;
766 IP.EnableDeferral = EnablePGOInlineDeferral;
768 // The inline deferral logic is used to avoid losing some
769 // inlining chance in future. It is helpful in SCC inliner, in which
770 // inlining is processed in bottom-up order.
771 // While in module inliner, the inlining order is a priority-based order
772 // by default. The inline deferral is unnecessary there. So we disable the
773 // inline deferral logic in module inliner.
774 IP.EnableDeferral = false;
776 return ModuleInlinerPass(IP, UseInlineAdvisor);
780 PassBuilder::buildModuleSimplificationPipeline(OptimizationLevel Level,
781 ThinOrFullLTOPhase Phase) {
782 ModulePassManager MPM;
784 // Place pseudo probe instrumentation as the first pass of the pipeline to
785 // minimize the impact of optimization changes.
786 if (PGOOpt && PGOOpt->PseudoProbeForProfiling &&
787 Phase != ThinOrFullLTOPhase::ThinLTOPostLink)
788 MPM.addPass(SampleProfileProbePass(TM));
790 bool HasSampleProfile = PGOOpt && (PGOOpt->Action == PGOOptions::SampleUse);
792 // In ThinLTO mode, when flattened profile is used, all the available
793 // profile information will be annotated in PreLink phase so there is
794 // no need to load the profile again in PostLink.
795 bool LoadSampleProfile =
797 !(FlattenedProfileUsed && Phase == ThinOrFullLTOPhase::ThinLTOPostLink);
799 // During the ThinLTO backend phase we perform early indirect call promotion
800 // here, before globalopt. Otherwise imported available_externally functions
801 // look unreferenced and are removed. If we are going to load the sample
802 // profile then defer until later.
803 // TODO: See if we can move later and consolidate with the location where
804 // we perform ICP when we are loading a sample profile.
805 // TODO: We pass HasSampleProfile (whether there was a sample profile file
806 // passed to the compile) to the SamplePGO flag of ICP. This is used to
807 // determine whether the new direct calls are annotated with prof metadata.
808 // Ideally this should be determined from whether the IR is annotated with
809 // sample profile, and not whether the a sample profile was provided on the
810 // command line. E.g. for flattened profiles where we will not be reloading
811 // the sample profile in the ThinLTO backend, we ideally shouldn't have to
812 // provide the sample profile file.
813 if (Phase == ThinOrFullLTOPhase::ThinLTOPostLink && !LoadSampleProfile)
814 MPM.addPass(PGOIndirectCallPromotion(true /* InLTO */, HasSampleProfile));
816 // Do basic inference of function attributes from known properties of system
817 // libraries and other oracles.
818 MPM.addPass(InferFunctionAttrsPass());
820 // Create an early function pass manager to cleanup the output of the
822 FunctionPassManager EarlyFPM;
823 // Lower llvm.expect to metadata before attempting transforms.
824 // Compare/branch metadata may alter the behavior of passes like SimplifyCFG.
825 EarlyFPM.addPass(LowerExpectIntrinsicPass());
826 EarlyFPM.addPass(SimplifyCFGPass());
827 EarlyFPM.addPass(SROAPass());
828 EarlyFPM.addPass(EarlyCSEPass());
829 EarlyFPM.addPass(CoroEarlyPass());
830 if (Level == OptimizationLevel::O3)
831 EarlyFPM.addPass(CallSiteSplittingPass());
833 // In SamplePGO ThinLTO backend, we need instcombine before profile annotation
834 // to convert bitcast to direct calls so that they can be inlined during the
835 // profile annotation prepration step.
836 // More details about SamplePGO design can be found in:
837 // https://research.google.com/pubs/pub45290.html
838 // FIXME: revisit how SampleProfileLoad/Inliner/ICP is structured.
839 if (LoadSampleProfile)
840 EarlyFPM.addPass(InstCombinePass());
841 MPM.addPass(createModuleToFunctionPassAdaptor(std::move(EarlyFPM),
842 PTO.EagerlyInvalidateAnalyses));
844 if (LoadSampleProfile) {
845 // Annotate sample profile right after early FPM to ensure freshness of
847 MPM.addPass(SampleProfileLoaderPass(PGOOpt->ProfileFile,
848 PGOOpt->ProfileRemappingFile, Phase));
849 // Cache ProfileSummaryAnalysis once to avoid the potential need to insert
850 // RequireAnalysisPass for PSI before subsequent non-module passes.
851 MPM.addPass(RequireAnalysisPass<ProfileSummaryAnalysis, Module>());
852 // Do not invoke ICP in the LTOPrelink phase as it makes it hard
853 // for the profile annotation to be accurate in the LTO backend.
854 if (Phase != ThinOrFullLTOPhase::ThinLTOPreLink &&
855 Phase != ThinOrFullLTOPhase::FullLTOPreLink)
856 // We perform early indirect call promotion here, before globalopt.
857 // This is important for the ThinLTO backend phase because otherwise
858 // imported available_externally functions look unreferenced and are
861 PGOIndirectCallPromotion(true /* IsInLTO */, true /* SamplePGO */));
864 // Try to perform OpenMP specific optimizations on the module. This is a
865 // (quick!) no-op if there are no OpenMP runtime calls present in the module.
866 if (Level != OptimizationLevel::O0)
867 MPM.addPass(OpenMPOptPass());
869 if (AttributorRun & AttributorRunOption::MODULE)
870 MPM.addPass(AttributorPass());
872 // Lower type metadata and the type.test intrinsic in the ThinLTO
873 // post link pipeline after ICP. This is to enable usage of the type
874 // tests in ICP sequences.
875 if (Phase == ThinOrFullLTOPhase::ThinLTOPostLink)
876 MPM.addPass(LowerTypeTestsPass(nullptr, nullptr, true));
878 for (auto &C : PipelineEarlySimplificationEPCallbacks)
881 // Specialize functions with IPSCCP.
882 if (EnableFunctionSpecialization && Level == OptimizationLevel::O3)
883 MPM.addPass(FunctionSpecializationPass());
885 // Interprocedural constant propagation now that basic cleanup has occurred
886 // and prior to optimizing globals.
887 // FIXME: This position in the pipeline hasn't been carefully considered in
888 // years, it should be re-analyzed.
889 MPM.addPass(IPSCCPPass());
891 // Attach metadata to indirect call sites indicating the set of functions
892 // they may target at run-time. This should follow IPSCCP.
893 MPM.addPass(CalledValuePropagationPass());
895 // Optimize globals to try and fold them into constants.
896 MPM.addPass(GlobalOptPass());
898 // Promote any localized globals to SSA registers.
899 // FIXME: Should this instead by a run of SROA?
900 // FIXME: We should probably run instcombine and simplifycfg afterward to
901 // delete control flows that are dead once globals have been folded to
903 MPM.addPass(createModuleToFunctionPassAdaptor(PromotePass()));
905 // Remove any dead arguments exposed by cleanups and constant folding
907 MPM.addPass(DeadArgumentEliminationPass());
909 // Create a small function pass pipeline to cleanup after all the global
911 FunctionPassManager GlobalCleanupPM;
912 GlobalCleanupPM.addPass(InstCombinePass());
913 invokePeepholeEPCallbacks(GlobalCleanupPM, Level);
915 GlobalCleanupPM.addPass(SimplifyCFGPass());
916 MPM.addPass(createModuleToFunctionPassAdaptor(std::move(GlobalCleanupPM),
917 PTO.EagerlyInvalidateAnalyses));
919 // Add all the requested passes for instrumentation PGO, if requested.
920 if (PGOOpt && Phase != ThinOrFullLTOPhase::ThinLTOPostLink &&
921 (PGOOpt->Action == PGOOptions::IRInstr ||
922 PGOOpt->Action == PGOOptions::IRUse)) {
923 addPGOInstrPasses(MPM, Level,
924 /* RunProfileGen */ PGOOpt->Action == PGOOptions::IRInstr,
925 /* IsCS */ false, PGOOpt->ProfileFile,
926 PGOOpt->ProfileRemappingFile);
927 MPM.addPass(PGOIndirectCallPromotion(false, false));
929 if (PGOOpt && Phase != ThinOrFullLTOPhase::ThinLTOPostLink &&
930 PGOOpt->CSAction == PGOOptions::CSIRInstr)
931 MPM.addPass(PGOInstrumentationGenCreateVar(PGOOpt->CSProfileGenFile));
933 // Synthesize function entry counts for non-PGO compilation.
934 if (EnableSyntheticCounts && !PGOOpt)
935 MPM.addPass(SyntheticCountsPropagation());
937 if (EnableModuleInliner)
938 MPM.addPass(buildModuleInlinerPipeline(Level, Phase));
940 MPM.addPass(buildInlinerPipeline(Level, Phase));
942 if (EnableMemProfiler && Phase != ThinOrFullLTOPhase::ThinLTOPreLink) {
943 MPM.addPass(createModuleToFunctionPassAdaptor(MemProfilerPass()));
944 MPM.addPass(ModuleMemProfilerPass());
950 /// TODO: Should LTO cause any differences to this set of passes?
951 void PassBuilder::addVectorPasses(OptimizationLevel Level,
952 FunctionPassManager &FPM, bool IsFullLTO) {
953 FPM.addPass(LoopVectorizePass(
954 LoopVectorizeOptions(!PTO.LoopInterleaving, !PTO.LoopVectorization)));
957 // The vectorizer may have significantly shortened a loop body; unroll
958 // again. Unroll small loops to hide loop backedge latency and saturate any
959 // parallel execution resources of an out-of-order processor. We also then
960 // need to clean up redundancies and loop invariant code.
961 // FIXME: It would be really good to use a loop-integrated instruction
962 // combiner for cleanup here so that the unrolling and LICM can be pipelined
963 // across the loop nests.
964 // We do UnrollAndJam in a separate LPM to ensure it happens before unroll
965 if (EnableUnrollAndJam && PTO.LoopUnrolling)
966 FPM.addPass(createFunctionToLoopPassAdaptor(
967 LoopUnrollAndJamPass(Level.getSpeedupLevel())));
968 FPM.addPass(LoopUnrollPass(LoopUnrollOptions(
969 Level.getSpeedupLevel(), /*OnlyWhenForced=*/!PTO.LoopUnrolling,
970 PTO.ForgetAllSCEVInLoopUnroll)));
971 FPM.addPass(WarnMissedTransformationsPass());
975 // Eliminate loads by forwarding stores from the previous iteration to loads
976 // of the current iteration.
977 FPM.addPass(LoopLoadEliminationPass());
979 // Cleanup after the loop optimization passes.
980 FPM.addPass(InstCombinePass());
982 if (Level.getSpeedupLevel() > 1 && ExtraVectorizerPasses) {
983 // At higher optimization levels, try to clean up any runtime overlap and
984 // alignment checks inserted by the vectorizer. We want to track correlated
985 // runtime checks for two inner loops in the same outer loop, fold any
986 // common computations, hoist loop-invariant aspects out of any outer loop,
987 // and unswitch the runtime checks if possible. Once hoisted, we may have
988 // dead (or speculatable) control flows or more combining opportunities.
989 FPM.addPass(EarlyCSEPass());
990 FPM.addPass(CorrelatedValuePropagationPass());
991 FPM.addPass(InstCombinePass());
993 LPM.addPass(LICMPass(PTO.LicmMssaOptCap, PTO.LicmMssaNoAccForPromotionCap));
994 LPM.addPass(SimpleLoopUnswitchPass(/* NonTrivial */ Level ==
995 OptimizationLevel::O3));
997 RequireAnalysisPass<OptimizationRemarkEmitterAnalysis, Function>());
999 createFunctionToLoopPassAdaptor(std::move(LPM), /*UseMemorySSA=*/true,
1000 /*UseBlockFrequencyInfo=*/true));
1001 FPM.addPass(SimplifyCFGPass());
1002 FPM.addPass(InstCombinePass());
1005 // Now that we've formed fast to execute loop structures, we do further
1006 // optimizations. These are run afterward as they might block doing complex
1007 // analyses and transforms such as what are needed for loop vectorization.
1009 // Cleanup after loop vectorization, etc. Simplification passes like CVP and
1010 // GVN, loop transforms, and others have already run, so it's now better to
1011 // convert to more optimized IR using more aggressive simplify CFG options.
1012 // The extra sinking transform can create larger basic blocks, so do this
1013 // before SLP vectorization.
1014 FPM.addPass(SimplifyCFGPass(SimplifyCFGOptions()
1015 .forwardSwitchCondToPhi(true)
1016 .convertSwitchToLookupTable(true)
1017 .needCanonicalLoops(false)
1018 .hoistCommonInsts(true)
1019 .sinkCommonInsts(true)));
1022 FPM.addPass(SCCPPass());
1023 FPM.addPass(InstCombinePass());
1024 FPM.addPass(BDCEPass());
1027 // Optimize parallel scalar instruction chains into SIMD instructions.
1028 if (PTO.SLPVectorization) {
1029 FPM.addPass(SLPVectorizerPass());
1030 if (Level.getSpeedupLevel() > 1 && ExtraVectorizerPasses) {
1031 FPM.addPass(EarlyCSEPass());
1034 // Enhance/cleanup vector code.
1035 FPM.addPass(VectorCombinePass());
1038 FPM.addPass(InstCombinePass());
1039 // Unroll small loops to hide loop backedge latency and saturate any
1040 // parallel execution resources of an out-of-order processor. We also then
1041 // need to clean up redundancies and loop invariant code.
1042 // FIXME: It would be really good to use a loop-integrated instruction
1043 // combiner for cleanup here so that the unrolling and LICM can be pipelined
1044 // across the loop nests.
1045 // We do UnrollAndJam in a separate LPM to ensure it happens before unroll
1046 if (EnableUnrollAndJam && PTO.LoopUnrolling) {
1047 FPM.addPass(createFunctionToLoopPassAdaptor(
1048 LoopUnrollAndJamPass(Level.getSpeedupLevel())));
1050 FPM.addPass(LoopUnrollPass(LoopUnrollOptions(
1051 Level.getSpeedupLevel(), /*OnlyWhenForced=*/!PTO.LoopUnrolling,
1052 PTO.ForgetAllSCEVInLoopUnroll)));
1053 FPM.addPass(WarnMissedTransformationsPass());
1054 FPM.addPass(InstCombinePass());
1056 RequireAnalysisPass<OptimizationRemarkEmitterAnalysis, Function>());
1057 FPM.addPass(createFunctionToLoopPassAdaptor(
1058 LICMPass(PTO.LicmMssaOptCap, PTO.LicmMssaNoAccForPromotionCap),
1059 /*UseMemorySSA=*/true, /*UseBlockFrequencyInfo=*/true));
1062 // Now that we've vectorized and unrolled loops, we may have more refined
1063 // alignment information, try to re-derive it here.
1064 FPM.addPass(AlignmentFromAssumptionsPass());
1067 FPM.addPass(InstCombinePass());
1071 PassBuilder::buildModuleOptimizationPipeline(OptimizationLevel Level,
1073 ModulePassManager MPM;
1075 // Optimize globals now that the module is fully simplified.
1076 MPM.addPass(GlobalOptPass());
1077 MPM.addPass(GlobalDCEPass());
1079 // Run partial inlining pass to partially inline functions that have
1081 if (RunPartialInlining)
1082 MPM.addPass(PartialInlinerPass());
1084 // Remove avail extern fns and globals definitions since we aren't compiling
1085 // an object file for later LTO. For LTO we want to preserve these so they
1086 // are eligible for inlining at link-time. Note if they are unreferenced they
1087 // will be removed by GlobalDCE later, so this only impacts referenced
1088 // available externally globals. Eventually they will be suppressed during
1089 // codegen, but eliminating here enables more opportunity for GlobalDCE as it
1090 // may make globals referenced by available external functions dead and saves
1091 // running remaining passes on the eliminated functions. These should be
1092 // preserved during prelinking for link-time inlining decisions.
1094 MPM.addPass(EliminateAvailableExternallyPass());
1096 if (EnableOrderFileInstrumentation)
1097 MPM.addPass(InstrOrderFilePass());
1099 // Do RPO function attribute inference across the module to forward-propagate
1100 // attributes where applicable.
1101 // FIXME: Is this really an optimization rather than a canonicalization?
1102 MPM.addPass(ReversePostOrderFunctionAttrsPass());
1104 // Do a post inline PGO instrumentation and use pass. This is a context
1105 // sensitive PGO pass. We don't want to do this in LTOPreLink phrase as
1106 // cross-module inline has not been done yet. The context sensitive
1107 // instrumentation is after all the inlines are done.
1108 if (!LTOPreLink && PGOOpt) {
1109 if (PGOOpt->CSAction == PGOOptions::CSIRInstr)
1110 addPGOInstrPasses(MPM, Level, /* RunProfileGen */ true,
1111 /* IsCS */ true, PGOOpt->CSProfileGenFile,
1112 PGOOpt->ProfileRemappingFile);
1113 else if (PGOOpt->CSAction == PGOOptions::CSIRUse)
1114 addPGOInstrPasses(MPM, Level, /* RunProfileGen */ false,
1115 /* IsCS */ true, PGOOpt->ProfileFile,
1116 PGOOpt->ProfileRemappingFile);
1119 // Re-require GloblasAA here prior to function passes. This is particularly
1120 // useful as the above will have inlined, DCE'ed, and function-attr
1121 // propagated everything. We should at this point have a reasonably minimal
1122 // and richly annotated call graph. By computing aliasing and mod/ref
1123 // information for all local globals here, the late loop passes and notably
1124 // the vectorizer will be able to use them to help recognize vectorizable
1125 // memory operations.
1126 MPM.addPass(RequireAnalysisPass<GlobalsAA, Module>());
1128 FunctionPassManager OptimizePM;
1129 OptimizePM.addPass(Float2IntPass());
1130 OptimizePM.addPass(LowerConstantIntrinsicsPass());
1133 OptimizePM.addPass(LowerMatrixIntrinsicsPass());
1134 OptimizePM.addPass(EarlyCSEPass());
1137 // FIXME: We need to run some loop optimizations to re-rotate loops after
1138 // simplifycfg and others undo their rotation.
1140 // Optimize the loop execution. These passes operate on entire loop nests
1141 // rather than on each loop in an inside-out manner, and so they are actually
1144 for (auto &C : VectorizerStartEPCallbacks)
1145 C(OptimizePM, Level);
1147 LoopPassManager LPM;
1148 // First rotate loops that may have been un-rotated by prior passes.
1149 // Disable header duplication at -Oz.
1150 LPM.addPass(LoopRotatePass(Level != OptimizationLevel::Oz, LTOPreLink));
1151 // Some loops may have become dead by now. Try to delete them.
1152 // FIXME: see disscussion in https://reviews.llvm.org/D112851
1153 // this may need to be revisited once GVN is more powerful.
1154 LPM.addPass(LoopDeletionPass());
1155 OptimizePM.addPass(createFunctionToLoopPassAdaptor(
1156 std::move(LPM), /*UseMemorySSA=*/false, /*UseBlockFrequencyInfo=*/false));
1158 // Distribute loops to allow partial vectorization. I.e. isolate dependences
1159 // into separate loop that would otherwise inhibit vectorization. This is
1160 // currently only performed for loops marked with the metadata
1161 // llvm.loop.distribute=true or when -enable-loop-distribute is specified.
1162 OptimizePM.addPass(LoopDistributePass());
1164 // Populates the VFABI attribute with the scalar-to-vector mappings
1165 // from the TargetLibraryInfo.
1166 OptimizePM.addPass(InjectTLIMappings());
1168 addVectorPasses(Level, OptimizePM, /* IsFullLTO */ false);
1170 // Split out cold code. Splitting is done late to avoid hiding context from
1171 // other optimizations and inadvertently regressing performance. The tradeoff
1172 // is that this has a higher code size cost than splitting early.
1173 if (EnableHotColdSplit && !LTOPreLink)
1174 MPM.addPass(HotColdSplittingPass());
1176 // Search the code for similar regions of code. If enough similar regions can
1177 // be found where extracting the regions into their own function will decrease
1178 // the size of the program, we extract the regions, a deduplicate the
1179 // structurally similar regions.
1180 if (EnableIROutliner)
1181 MPM.addPass(IROutlinerPass());
1183 // Merge functions if requested.
1184 if (PTO.MergeFunctions)
1185 MPM.addPass(MergeFunctionsPass());
1187 // LoopSink pass sinks instructions hoisted by LICM, which serves as a
1188 // canonicalization pass that enables other optimizations. As a result,
1189 // LoopSink pass needs to be a very late IR pass to avoid undoing LICM
1190 // result too early.
1191 OptimizePM.addPass(LoopSinkPass());
1193 // And finally clean up LCSSA form before generating code.
1194 OptimizePM.addPass(InstSimplifyPass());
1196 // This hoists/decomposes div/rem ops. It should run after other sink/hoist
1197 // passes to avoid re-sinking, but before SimplifyCFG because it can allow
1198 // flattening of blocks.
1199 OptimizePM.addPass(DivRemPairsPass());
1201 // LoopSink (and other loop passes since the last simplifyCFG) might have
1202 // resulted in single-entry-single-exit or empty blocks. Clean up the CFG.
1203 OptimizePM.addPass(SimplifyCFGPass());
1205 OptimizePM.addPass(CoroCleanupPass());
1207 // Add the core optimizing pipeline.
1208 MPM.addPass(createModuleToFunctionPassAdaptor(std::move(OptimizePM),
1209 PTO.EagerlyInvalidateAnalyses));
1211 for (auto &C : OptimizerLastEPCallbacks)
1214 if (PTO.CallGraphProfile)
1215 MPM.addPass(CGProfilePass());
1217 // Now we need to do some global optimization transforms.
1218 // FIXME: It would seem like these should come first in the optimization
1219 // pipeline and maybe be the bottom of the canonicalization pipeline? Weird
1221 MPM.addPass(GlobalDCEPass());
1222 MPM.addPass(ConstantMergePass());
1224 // TODO: Relative look table converter pass caused an issue when full lto is
1225 // enabled. See https://reviews.llvm.org/D94355 for more details.
1226 // Until the issue fixed, disable this pass during pre-linking phase.
1228 MPM.addPass(RelLookupTableConverterPass());
1234 PassBuilder::buildPerModuleDefaultPipeline(OptimizationLevel Level,
1236 assert(Level != OptimizationLevel::O0 &&
1237 "Must request optimizations for the default pipeline!");
1239 ModulePassManager MPM;
1241 // Convert @llvm.global.annotations to !annotation metadata.
1242 MPM.addPass(Annotation2MetadataPass());
1244 // Force any function attributes we want the rest of the pipeline to observe.
1245 MPM.addPass(ForceFunctionAttrsPass());
1247 // Apply module pipeline start EP callback.
1248 for (auto &C : PipelineStartEPCallbacks)
1251 if (PGOOpt && PGOOpt->DebugInfoForProfiling)
1252 MPM.addPass(createModuleToFunctionPassAdaptor(AddDiscriminatorsPass()));
1254 // Add the core simplification pipeline.
1255 MPM.addPass(buildModuleSimplificationPipeline(
1256 Level, LTOPreLink ? ThinOrFullLTOPhase::FullLTOPreLink
1257 : ThinOrFullLTOPhase::None));
1259 // Now add the optimization pipeline.
1260 MPM.addPass(buildModuleOptimizationPipeline(Level, LTOPreLink));
1262 if (PGOOpt && PGOOpt->PseudoProbeForProfiling &&
1263 PGOOpt->Action == PGOOptions::SampleUse)
1264 MPM.addPass(PseudoProbeUpdatePass());
1266 // Emit annotation remarks.
1267 addAnnotationRemarksPass(MPM);
1270 addRequiredLTOPreLinkPasses(MPM);
1276 PassBuilder::buildThinLTOPreLinkDefaultPipeline(OptimizationLevel Level) {
1277 assert(Level != OptimizationLevel::O0 &&
1278 "Must request optimizations for the default pipeline!");
1280 ModulePassManager MPM;
1282 // Convert @llvm.global.annotations to !annotation metadata.
1283 MPM.addPass(Annotation2MetadataPass());
1285 // Force any function attributes we want the rest of the pipeline to observe.
1286 MPM.addPass(ForceFunctionAttrsPass());
1288 if (PGOOpt && PGOOpt->DebugInfoForProfiling)
1289 MPM.addPass(createModuleToFunctionPassAdaptor(AddDiscriminatorsPass()));
1291 // Apply module pipeline start EP callback.
1292 for (auto &C : PipelineStartEPCallbacks)
1295 // If we are planning to perform ThinLTO later, we don't bloat the code with
1296 // unrolling/vectorization/... now. Just simplify the module as much as we
1298 MPM.addPass(buildModuleSimplificationPipeline(
1299 Level, ThinOrFullLTOPhase::ThinLTOPreLink));
1301 // Run partial inlining pass to partially inline functions that have
1303 // FIXME: It isn't clear whether this is really the right place to run this
1304 // in ThinLTO. Because there is another canonicalization and simplification
1305 // phase that will run after the thin link, running this here ends up with
1306 // less information than will be available later and it may grow functions in
1307 // ways that aren't beneficial.
1308 if (RunPartialInlining)
1309 MPM.addPass(PartialInlinerPass());
1311 // Reduce the size of the IR as much as possible.
1312 MPM.addPass(GlobalOptPass());
1314 // Module simplification splits coroutines, but does not fully clean up
1315 // coroutine intrinsics. To ensure ThinLTO optimization passes don't trip up
1316 // on these, we schedule the cleanup here.
1317 MPM.addPass(createModuleToFunctionPassAdaptor(CoroCleanupPass()));
1319 if (PGOOpt && PGOOpt->PseudoProbeForProfiling &&
1320 PGOOpt->Action == PGOOptions::SampleUse)
1321 MPM.addPass(PseudoProbeUpdatePass());
1323 // Handle OptimizerLastEPCallbacks added by clang on PreLink. Actual
1324 // optimization is going to be done in PostLink stage, but clang can't
1325 // add callbacks there in case of in-process ThinLTO called by linker.
1326 for (auto &C : OptimizerLastEPCallbacks)
1329 // Emit annotation remarks.
1330 addAnnotationRemarksPass(MPM);
1332 addRequiredLTOPreLinkPasses(MPM);
1337 ModulePassManager PassBuilder::buildThinLTODefaultPipeline(
1338 OptimizationLevel Level, const ModuleSummaryIndex *ImportSummary) {
1339 ModulePassManager MPM;
1341 // Convert @llvm.global.annotations to !annotation metadata.
1342 MPM.addPass(Annotation2MetadataPass());
1344 if (ImportSummary) {
1345 // These passes import type identifier resolutions for whole-program
1346 // devirtualization and CFI. They must run early because other passes may
1347 // disturb the specific instruction patterns that these passes look for,
1348 // creating dependencies on resolutions that may not appear in the summary.
1350 // For example, GVN may transform the pattern assume(type.test) appearing in
1351 // two basic blocks into assume(phi(type.test, type.test)), which would
1352 // transform a dependency on a WPD resolution into a dependency on a type
1353 // identifier resolution for CFI.
1355 // Also, WPD has access to more precise information than ICP and can
1356 // devirtualize more effectively, so it should operate on the IR first.
1358 // The WPD and LowerTypeTest passes need to run at -O0 to lower type
1359 // metadata and intrinsics.
1360 MPM.addPass(WholeProgramDevirtPass(nullptr, ImportSummary));
1361 MPM.addPass(LowerTypeTestsPass(nullptr, ImportSummary));
1364 if (Level == OptimizationLevel::O0) {
1365 // Run a second time to clean up any type tests left behind by WPD for use
1367 MPM.addPass(LowerTypeTestsPass(nullptr, nullptr, true));
1368 // Drop available_externally and unreferenced globals. This is necessary
1369 // with ThinLTO in order to avoid leaving undefined references to dead
1370 // globals in the object file.
1371 MPM.addPass(EliminateAvailableExternallyPass());
1372 MPM.addPass(GlobalDCEPass());
1376 // Force any function attributes we want the rest of the pipeline to observe.
1377 MPM.addPass(ForceFunctionAttrsPass());
1379 // Add the core simplification pipeline.
1380 MPM.addPass(buildModuleSimplificationPipeline(
1381 Level, ThinOrFullLTOPhase::ThinLTOPostLink));
1383 // Now add the optimization pipeline.
1384 MPM.addPass(buildModuleOptimizationPipeline(Level));
1386 // Emit annotation remarks.
1387 addAnnotationRemarksPass(MPM);
1393 PassBuilder::buildLTOPreLinkDefaultPipeline(OptimizationLevel Level) {
1394 assert(Level != OptimizationLevel::O0 &&
1395 "Must request optimizations for the default pipeline!");
1396 // FIXME: We should use a customized pre-link pipeline!
1397 return buildPerModuleDefaultPipeline(Level,
1398 /* LTOPreLink */ true);
1402 PassBuilder::buildLTODefaultPipeline(OptimizationLevel Level,
1403 ModuleSummaryIndex *ExportSummary) {
1404 ModulePassManager MPM;
1406 // Convert @llvm.global.annotations to !annotation metadata.
1407 MPM.addPass(Annotation2MetadataPass());
1409 // Create a function that performs CFI checks for cross-DSO calls with targets
1410 // in the current module.
1411 MPM.addPass(CrossDSOCFIPass());
1413 if (Level == OptimizationLevel::O0) {
1414 // The WPD and LowerTypeTest passes need to run at -O0 to lower type
1415 // metadata and intrinsics.
1416 MPM.addPass(WholeProgramDevirtPass(ExportSummary, nullptr));
1417 MPM.addPass(LowerTypeTestsPass(ExportSummary, nullptr));
1418 // Run a second time to clean up any type tests left behind by WPD for use
1420 MPM.addPass(LowerTypeTestsPass(nullptr, nullptr, true));
1422 // Emit annotation remarks.
1423 addAnnotationRemarksPass(MPM);
1428 if (PGOOpt && PGOOpt->Action == PGOOptions::SampleUse) {
1429 // Load sample profile before running the LTO optimization pipeline.
1430 MPM.addPass(SampleProfileLoaderPass(PGOOpt->ProfileFile,
1431 PGOOpt->ProfileRemappingFile,
1432 ThinOrFullLTOPhase::FullLTOPostLink));
1433 // Cache ProfileSummaryAnalysis once to avoid the potential need to insert
1434 // RequireAnalysisPass for PSI before subsequent non-module passes.
1435 MPM.addPass(RequireAnalysisPass<ProfileSummaryAnalysis, Module>());
1438 // Remove unused virtual tables to improve the quality of code generated by
1439 // whole-program devirtualization and bitset lowering.
1440 MPM.addPass(GlobalDCEPass());
1442 // Force any function attributes we want the rest of the pipeline to observe.
1443 MPM.addPass(ForceFunctionAttrsPass());
1445 // Do basic inference of function attributes from known properties of system
1446 // libraries and other oracles.
1447 MPM.addPass(InferFunctionAttrsPass());
1449 if (Level.getSpeedupLevel() > 1) {
1450 FunctionPassManager EarlyFPM;
1451 EarlyFPM.addPass(CallSiteSplittingPass());
1452 MPM.addPass(createModuleToFunctionPassAdaptor(
1453 std::move(EarlyFPM), PTO.EagerlyInvalidateAnalyses));
1455 // Indirect call promotion. This should promote all the targets that are
1456 // left by the earlier promotion pass that promotes intra-module targets.
1457 // This two-step promotion is to save the compile time. For LTO, it should
1458 // produce the same result as if we only do promotion here.
1459 MPM.addPass(PGOIndirectCallPromotion(
1460 true /* InLTO */, PGOOpt && PGOOpt->Action == PGOOptions::SampleUse));
1462 if (EnableFunctionSpecialization && Level == OptimizationLevel::O3)
1463 MPM.addPass(FunctionSpecializationPass());
1464 // Propagate constants at call sites into the functions they call. This
1465 // opens opportunities for globalopt (and inlining) by substituting function
1466 // pointers passed as arguments to direct uses of functions.
1467 MPM.addPass(IPSCCPPass());
1469 // Attach metadata to indirect call sites indicating the set of functions
1470 // they may target at run-time. This should follow IPSCCP.
1471 MPM.addPass(CalledValuePropagationPass());
1474 // Now deduce any function attributes based in the current code.
1476 createModuleToPostOrderCGSCCPassAdaptor(PostOrderFunctionAttrsPass()));
1478 // Do RPO function attribute inference across the module to forward-propagate
1479 // attributes where applicable.
1480 // FIXME: Is this really an optimization rather than a canonicalization?
1481 MPM.addPass(ReversePostOrderFunctionAttrsPass());
1483 // Use in-range annotations on GEP indices to split globals where beneficial.
1484 MPM.addPass(GlobalSplitPass());
1486 // Run whole program optimization of virtual call when the list of callees
1488 MPM.addPass(WholeProgramDevirtPass(ExportSummary, nullptr));
1490 // Stop here at -O1.
1491 if (Level == OptimizationLevel::O1) {
1492 // The LowerTypeTestsPass needs to run to lower type metadata and the
1493 // type.test intrinsics. The pass does nothing if CFI is disabled.
1494 MPM.addPass(LowerTypeTestsPass(ExportSummary, nullptr));
1495 // Run a second time to clean up any type tests left behind by WPD for use
1496 // in ICP (which is performed earlier than this in the regular LTO
1498 MPM.addPass(LowerTypeTestsPass(nullptr, nullptr, true));
1500 // Emit annotation remarks.
1501 addAnnotationRemarksPass(MPM);
1506 // Optimize globals to try and fold them into constants.
1507 MPM.addPass(GlobalOptPass());
1509 // Promote any localized globals to SSA registers.
1510 MPM.addPass(createModuleToFunctionPassAdaptor(PromotePass()));
1512 // Linking modules together can lead to duplicate global constant, only
1513 // keep one copy of each constant.
1514 MPM.addPass(ConstantMergePass());
1516 // Remove unused arguments from functions.
1517 MPM.addPass(DeadArgumentEliminationPass());
1519 // Reduce the code after globalopt and ipsccp. Both can open up significant
1520 // simplification opportunities, and both can propagate functions through
1521 // function pointers. When this happens, we often have to resolve varargs
1522 // calls, etc, so let instcombine do this.
1523 FunctionPassManager PeepholeFPM;
1524 if (Level == OptimizationLevel::O3)
1525 PeepholeFPM.addPass(AggressiveInstCombinePass());
1526 PeepholeFPM.addPass(InstCombinePass());
1527 invokePeepholeEPCallbacks(PeepholeFPM, Level);
1529 MPM.addPass(createModuleToFunctionPassAdaptor(std::move(PeepholeFPM),
1530 PTO.EagerlyInvalidateAnalyses));
1532 // Note: historically, the PruneEH pass was run first to deduce nounwind and
1533 // generally clean up exception handling overhead. It isn't clear this is
1534 // valuable as the inliner doesn't currently care whether it is inlining an
1535 // invoke or a call.
1536 // Run the inliner now.
1537 MPM.addPass(ModuleInlinerWrapperPass(getInlineParamsFromOptLevel(Level)));
1539 // Optimize globals again after we ran the inliner.
1540 MPM.addPass(GlobalOptPass());
1542 // Garbage collect dead functions.
1543 MPM.addPass(GlobalDCEPass());
1545 // If we didn't decide to inline a function, check to see if we can
1546 // transform it to pass arguments by value instead of by reference.
1547 MPM.addPass(createModuleToPostOrderCGSCCPassAdaptor(ArgumentPromotionPass()));
1549 FunctionPassManager FPM;
1550 // The IPO Passes may leave cruft around. Clean up after them.
1551 FPM.addPass(InstCombinePass());
1552 invokePeepholeEPCallbacks(FPM, Level);
1554 FPM.addPass(JumpThreadingPass(/*InsertFreezeWhenUnfoldingSelect*/ true));
1556 // Do a post inline PGO instrumentation and use pass. This is a context
1557 // sensitive PGO pass.
1559 if (PGOOpt->CSAction == PGOOptions::CSIRInstr)
1560 addPGOInstrPasses(MPM, Level, /* RunProfileGen */ true,
1561 /* IsCS */ true, PGOOpt->CSProfileGenFile,
1562 PGOOpt->ProfileRemappingFile);
1563 else if (PGOOpt->CSAction == PGOOptions::CSIRUse)
1564 addPGOInstrPasses(MPM, Level, /* RunProfileGen */ false,
1565 /* IsCS */ true, PGOOpt->ProfileFile,
1566 PGOOpt->ProfileRemappingFile);
1570 FPM.addPass(SROAPass());
1572 // LTO provides additional opportunities for tailcall elimination due to
1573 // link-time inlining, and visibility of nocapture attribute.
1574 FPM.addPass(TailCallElimPass());
1576 // Run a few AA driver optimizations here and now to cleanup the code.
1577 MPM.addPass(createModuleToFunctionPassAdaptor(std::move(FPM),
1578 PTO.EagerlyInvalidateAnalyses));
1581 createModuleToPostOrderCGSCCPassAdaptor(PostOrderFunctionAttrsPass()));
1583 // Require the GlobalsAA analysis for the module so we can query it within
1585 MPM.addPass(RequireAnalysisPass<GlobalsAA, Module>());
1586 // Invalidate AAManager so it can be recreated and pick up the newly available
1589 createModuleToFunctionPassAdaptor(InvalidateAnalysisPass<AAManager>()));
1591 FunctionPassManager MainFPM;
1592 MainFPM.addPass(createFunctionToLoopPassAdaptor(
1593 LICMPass(PTO.LicmMssaOptCap, PTO.LicmMssaNoAccForPromotionCap),
1594 /*USeMemorySSA=*/true, /*UseBlockFrequencyInfo=*/true));
1597 MainFPM.addPass(NewGVNPass());
1599 MainFPM.addPass(GVNPass());
1601 // Remove dead memcpy()'s.
1602 MainFPM.addPass(MemCpyOptPass());
1604 // Nuke dead stores.
1605 MainFPM.addPass(DSEPass());
1606 MainFPM.addPass(MergedLoadStoreMotionPass());
1608 // More loops are countable; try to optimize them.
1609 if (EnableLoopFlatten && Level.getSpeedupLevel() > 1)
1610 MainFPM.addPass(createFunctionToLoopPassAdaptor(LoopFlattenPass()));
1612 if (EnableConstraintElimination)
1613 MainFPM.addPass(ConstraintEliminationPass());
1615 LoopPassManager LPM;
1616 LPM.addPass(IndVarSimplifyPass());
1617 LPM.addPass(LoopDeletionPass());
1618 // FIXME: Add loop interchange.
1620 // Unroll small loops and perform peeling.
1621 LPM.addPass(LoopFullUnrollPass(Level.getSpeedupLevel(),
1622 /* OnlyWhenForced= */ !PTO.LoopUnrolling,
1623 PTO.ForgetAllSCEVInLoopUnroll));
1624 // The loop passes in LPM (LoopFullUnrollPass) do not preserve MemorySSA.
1625 // *All* loop passes must preserve it, in order to be able to use it.
1626 MainFPM.addPass(createFunctionToLoopPassAdaptor(
1627 std::move(LPM), /*UseMemorySSA=*/false, /*UseBlockFrequencyInfo=*/true));
1629 MainFPM.addPass(LoopDistributePass());
1631 addVectorPasses(Level, MainFPM, /* IsFullLTO */ true);
1633 invokePeepholeEPCallbacks(MainFPM, Level);
1634 MainFPM.addPass(JumpThreadingPass(/*InsertFreezeWhenUnfoldingSelect*/ true));
1635 MPM.addPass(createModuleToFunctionPassAdaptor(std::move(MainFPM),
1636 PTO.EagerlyInvalidateAnalyses));
1638 // Lower type metadata and the type.test intrinsic. This pass supports
1639 // clang's control flow integrity mechanisms (-fsanitize=cfi*) and needs
1640 // to be run at link time if CFI is enabled. This pass does nothing if
1642 MPM.addPass(LowerTypeTestsPass(ExportSummary, nullptr));
1643 // Run a second time to clean up any type tests left behind by WPD for use
1644 // in ICP (which is performed earlier than this in the regular LTO pipeline).
1645 MPM.addPass(LowerTypeTestsPass(nullptr, nullptr, true));
1647 // Enable splitting late in the FullLTO post-link pipeline. This is done in
1648 // the same stage in the old pass manager (\ref addLateLTOOptimizationPasses).
1649 if (EnableHotColdSplit)
1650 MPM.addPass(HotColdSplittingPass());
1652 // Add late LTO optimization passes.
1653 // Delete basic blocks, which optimization passes may have killed.
1654 MPM.addPass(createModuleToFunctionPassAdaptor(
1655 SimplifyCFGPass(SimplifyCFGOptions().hoistCommonInsts(true))));
1657 // Drop bodies of available eternally objects to improve GlobalDCE.
1658 MPM.addPass(EliminateAvailableExternallyPass());
1660 // Now that we have optimized the program, discard unreachable functions.
1661 MPM.addPass(GlobalDCEPass());
1663 if (PTO.MergeFunctions)
1664 MPM.addPass(MergeFunctionsPass());
1666 // Emit annotation remarks.
1667 addAnnotationRemarksPass(MPM);
1672 ModulePassManager PassBuilder::buildO0DefaultPipeline(OptimizationLevel Level,
1674 assert(Level == OptimizationLevel::O0 &&
1675 "buildO0DefaultPipeline should only be used with O0");
1677 ModulePassManager MPM;
1679 // Perform pseudo probe instrumentation in O0 mode. This is for the
1680 // consistency between different build modes. For example, a LTO build can be
1681 // mixed with an O0 prelink and an O2 postlink. Loading a sample profile in
1682 // the postlink will require pseudo probe instrumentation in the prelink.
1683 if (PGOOpt && PGOOpt->PseudoProbeForProfiling)
1684 MPM.addPass(SampleProfileProbePass(TM));
1686 if (PGOOpt && (PGOOpt->Action == PGOOptions::IRInstr ||
1687 PGOOpt->Action == PGOOptions::IRUse))
1688 addPGOInstrPassesForO0(
1690 /* RunProfileGen */ (PGOOpt->Action == PGOOptions::IRInstr),
1691 /* IsCS */ false, PGOOpt->ProfileFile, PGOOpt->ProfileRemappingFile);
1693 for (auto &C : PipelineStartEPCallbacks)
1696 if (PGOOpt && PGOOpt->DebugInfoForProfiling)
1697 MPM.addPass(createModuleToFunctionPassAdaptor(AddDiscriminatorsPass()));
1699 for (auto &C : PipelineEarlySimplificationEPCallbacks)
1702 // Build a minimal pipeline based on the semantics required by LLVM,
1703 // which is just that always inlining occurs. Further, disable generating
1704 // lifetime intrinsics to avoid enabling further optimizations during
1706 MPM.addPass(AlwaysInlinerPass(
1707 /*InsertLifetimeIntrinsics=*/false));
1709 if (PTO.MergeFunctions)
1710 MPM.addPass(MergeFunctionsPass());
1714 createModuleToFunctionPassAdaptor(LowerMatrixIntrinsicsPass(true)));
1716 if (!CGSCCOptimizerLateEPCallbacks.empty()) {
1717 CGSCCPassManager CGPM;
1718 for (auto &C : CGSCCOptimizerLateEPCallbacks)
1720 if (!CGPM.isEmpty())
1721 MPM.addPass(createModuleToPostOrderCGSCCPassAdaptor(std::move(CGPM)));
1723 if (!LateLoopOptimizationsEPCallbacks.empty()) {
1724 LoopPassManager LPM;
1725 for (auto &C : LateLoopOptimizationsEPCallbacks)
1727 if (!LPM.isEmpty()) {
1728 MPM.addPass(createModuleToFunctionPassAdaptor(
1729 createFunctionToLoopPassAdaptor(std::move(LPM))));
1732 if (!LoopOptimizerEndEPCallbacks.empty()) {
1733 LoopPassManager LPM;
1734 for (auto &C : LoopOptimizerEndEPCallbacks)
1736 if (!LPM.isEmpty()) {
1737 MPM.addPass(createModuleToFunctionPassAdaptor(
1738 createFunctionToLoopPassAdaptor(std::move(LPM))));
1741 if (!ScalarOptimizerLateEPCallbacks.empty()) {
1742 FunctionPassManager FPM;
1743 for (auto &C : ScalarOptimizerLateEPCallbacks)
1746 MPM.addPass(createModuleToFunctionPassAdaptor(std::move(FPM)));
1748 if (!VectorizerStartEPCallbacks.empty()) {
1749 FunctionPassManager FPM;
1750 for (auto &C : VectorizerStartEPCallbacks)
1753 MPM.addPass(createModuleToFunctionPassAdaptor(std::move(FPM)));
1756 MPM.addPass(createModuleToFunctionPassAdaptor(CoroEarlyPass()));
1757 CGSCCPassManager CGPM;
1758 CGPM.addPass(CoroSplitPass());
1759 MPM.addPass(createModuleToPostOrderCGSCCPassAdaptor(std::move(CGPM)));
1760 MPM.addPass(createModuleToFunctionPassAdaptor(CoroCleanupPass()));
1762 for (auto &C : OptimizerLastEPCallbacks)
1766 addRequiredLTOPreLinkPasses(MPM);
1768 MPM.addPass(createModuleToFunctionPassAdaptor(AnnotationRemarksPass()));
1773 AAManager PassBuilder::buildDefaultAAPipeline() {
1776 // The order in which these are registered determines their priority when
1779 // First we register the basic alias analysis that provides the majority of
1780 // per-function local AA logic. This is a stateless, on-demand local set of
1782 AA.registerFunctionAnalysis<BasicAA>();
1784 // Next we query fast, specialized alias analyses that wrap IR-embedded
1785 // information about aliasing.
1786 AA.registerFunctionAnalysis<ScopedNoAliasAA>();
1787 AA.registerFunctionAnalysis<TypeBasedAA>();
1789 // Add support for querying global aliasing information when available.
1790 // Because the `AAManager` is a function analysis and `GlobalsAA` is a module
1791 // analysis, all that the `AAManager` can do is query for any *cached*
1792 // results from `GlobalsAA` through a readonly proxy.
1793 AA.registerModuleAnalysis<GlobalsAA>();
1795 // Add target-specific alias analyses.
1797 TM->registerDefaultAliasAnalyses(AA);