1 //===- PassManagerBuilder.cpp - Build Standard Pass -----------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file defines the PassManagerBuilder class, which is used to set up a
11 // "standard" optimization sequence suitable for languages like C and C++.
13 //===----------------------------------------------------------------------===//
15 #include "llvm/Transforms/IPO/PassManagerBuilder.h"
16 #include "llvm-c/Transforms/PassManagerBuilder.h"
17 #include "llvm/ADT/SmallVector.h"
18 #include "llvm/Analysis/BasicAliasAnalysis.h"
19 #include "llvm/Analysis/CFLAndersAliasAnalysis.h"
20 #include "llvm/Analysis/CFLSteensAliasAnalysis.h"
21 #include "llvm/Analysis/GlobalsModRef.h"
22 #include "llvm/Analysis/Passes.h"
23 #include "llvm/Analysis/ScopedNoAliasAA.h"
24 #include "llvm/Analysis/TargetLibraryInfo.h"
25 #include "llvm/Analysis/TypeBasedAliasAnalysis.h"
26 #include "llvm/IR/DataLayout.h"
27 #include "llvm/IR/LegacyPassManager.h"
28 #include "llvm/IR/ModuleSummaryIndex.h"
29 #include "llvm/IR/Verifier.h"
30 #include "llvm/Support/CommandLine.h"
31 #include "llvm/Support/ManagedStatic.h"
32 #include "llvm/Target/TargetMachine.h"
33 #include "llvm/Transforms/IPO.h"
34 #include "llvm/Transforms/IPO/ForceFunctionAttrs.h"
35 #include "llvm/Transforms/IPO/FunctionAttrs.h"
36 #include "llvm/Transforms/IPO/InferFunctionAttrs.h"
37 #include "llvm/Transforms/Instrumentation.h"
38 #include "llvm/Transforms/Scalar.h"
39 #include "llvm/Transforms/Scalar/GVN.h"
40 #include "llvm/Transforms/Vectorize.h"
45 RunLoopVectorization("vectorize-loops", cl::Hidden,
46 cl::desc("Run the Loop vectorization passes"));
49 RunSLPVectorization("vectorize-slp", cl::Hidden,
50 cl::desc("Run the SLP vectorization passes"));
53 RunBBVectorization("vectorize-slp-aggressive", cl::Hidden,
54 cl::desc("Run the BB vectorization passes"));
57 UseGVNAfterVectorization("use-gvn-after-vectorization",
58 cl::init(false), cl::Hidden,
59 cl::desc("Run GVN instead of Early CSE after vectorization passes"));
61 static cl::opt<bool> ExtraVectorizerPasses(
62 "extra-vectorizer-passes", cl::init(false), cl::Hidden,
63 cl::desc("Run cleanup optimization passes after vectorization."));
66 RunLoopRerolling("reroll-loops", cl::Hidden,
67 cl::desc("Run the loop rerolling pass"));
70 RunFloat2Int("float-to-int", cl::Hidden, cl::init(true),
71 cl::desc("Run the float2int (float demotion) pass"));
73 static cl::opt<bool> RunLoadCombine("combine-loads", cl::init(false),
75 cl::desc("Run the load combining pass"));
78 RunSLPAfterLoopVectorization("run-slp-after-loop-vectorization",
79 cl::init(true), cl::Hidden,
80 cl::desc("Run the SLP vectorizer (and BB vectorizer) after the Loop "
81 "vectorizer instead of before"));
83 // Experimental option to use CFL-AA
84 enum class CFLAAType { None, Steensgaard, Andersen, Both };
85 static cl::opt<CFLAAType>
86 UseCFLAA("use-cfl-aa", cl::init(CFLAAType::None), cl::Hidden,
87 cl::desc("Enable the new, experimental CFL alias analysis"),
88 cl::values(clEnumValN(CFLAAType::None, "none", "Disable CFL-AA"),
89 clEnumValN(CFLAAType::Steensgaard, "steens",
90 "Enable unification-based CFL-AA"),
91 clEnumValN(CFLAAType::Andersen, "anders",
92 "Enable inclusion-based CFL-AA"),
93 clEnumValN(CFLAAType::Both, "both",
94 "Enable both variants of CFL-aa"),
98 EnableMLSM("mlsm", cl::init(true), cl::Hidden,
99 cl::desc("Enable motion of merged load and store"));
101 static cl::opt<bool> EnableLoopInterchange(
102 "enable-loopinterchange", cl::init(false), cl::Hidden,
103 cl::desc("Enable the new, experimental LoopInterchange Pass"));
105 static cl::opt<bool> EnableNonLTOGlobalsModRef(
106 "enable-non-lto-gmr", cl::init(true), cl::Hidden,
108 "Enable the GlobalsModRef AliasAnalysis outside of the LTO pipeline."));
110 static cl::opt<bool> EnableLoopLoadElim(
111 "enable-loop-load-elim", cl::init(true), cl::Hidden,
112 cl::desc("Enable the LoopLoadElimination Pass"));
114 static cl::opt<std::string> RunPGOInstrGen(
115 "profile-generate", cl::init(""), cl::Hidden,
116 cl::desc("Enable generation phase of PGO instrumentation and specify the "
117 "path of profile data file"));
119 static cl::opt<std::string> RunPGOInstrUse(
120 "profile-use", cl::init(""), cl::Hidden, cl::value_desc("filename"),
121 cl::desc("Enable use phase of PGO instrumentation and specify the path "
122 "of profile data file"));
124 static cl::opt<bool> UseLoopVersioningLICM(
125 "enable-loop-versioning-licm", cl::init(false), cl::Hidden,
126 cl::desc("Enable the experimental Loop Versioning LICM pass"));
129 DisablePreInliner("disable-preinline", cl::init(false), cl::Hidden,
130 cl::desc("Disable pre-instrumentation inliner"));
132 static cl::opt<int> PreInlineThreshold(
133 "preinline-threshold", cl::Hidden, cl::init(75), cl::ZeroOrMore,
134 cl::desc("Control the amount of inlining in pre-instrumentation inliner "
137 PassManagerBuilder::PassManagerBuilder() {
140 LibraryInfo = nullptr;
142 ModuleSummary = nullptr;
143 DisableUnitAtATime = false;
144 DisableUnrollLoops = false;
145 BBVectorize = RunBBVectorization;
146 SLPVectorize = RunSLPVectorization;
147 LoopVectorize = RunLoopVectorization;
148 RerollLoops = RunLoopRerolling;
149 LoadCombine = RunLoadCombine;
150 DisableGVNLoadPRE = false;
152 VerifyOutput = false;
153 MergeFunctions = false;
154 PrepareForLTO = false;
155 PGOInstrGen = RunPGOInstrGen;
156 PGOInstrUse = RunPGOInstrUse;
157 PrepareForThinLTO = false;
158 PerformThinLTO = false;
161 PassManagerBuilder::~PassManagerBuilder() {
166 /// Set of global extensions, automatically added as part of the standard set.
167 static ManagedStatic<SmallVector<std::pair<PassManagerBuilder::ExtensionPointTy,
168 PassManagerBuilder::ExtensionFn>, 8> > GlobalExtensions;
170 void PassManagerBuilder::addGlobalExtension(
171 PassManagerBuilder::ExtensionPointTy Ty,
172 PassManagerBuilder::ExtensionFn Fn) {
173 GlobalExtensions->push_back(std::make_pair(Ty, std::move(Fn)));
176 void PassManagerBuilder::addExtension(ExtensionPointTy Ty, ExtensionFn Fn) {
177 Extensions.push_back(std::make_pair(Ty, std::move(Fn)));
180 void PassManagerBuilder::addExtensionsToPM(ExtensionPointTy ETy,
181 legacy::PassManagerBase &PM) const {
182 for (unsigned i = 0, e = GlobalExtensions->size(); i != e; ++i)
183 if ((*GlobalExtensions)[i].first == ETy)
184 (*GlobalExtensions)[i].second(*this, PM);
185 for (unsigned i = 0, e = Extensions.size(); i != e; ++i)
186 if (Extensions[i].first == ETy)
187 Extensions[i].second(*this, PM);
190 void PassManagerBuilder::addInitialAliasAnalysisPasses(
191 legacy::PassManagerBase &PM) const {
193 case CFLAAType::Steensgaard:
194 PM.add(createCFLSteensAAWrapperPass());
196 case CFLAAType::Andersen:
197 PM.add(createCFLAndersAAWrapperPass());
199 case CFLAAType::Both:
200 PM.add(createCFLSteensAAWrapperPass());
201 PM.add(createCFLAndersAAWrapperPass());
207 // Add TypeBasedAliasAnalysis before BasicAliasAnalysis so that
208 // BasicAliasAnalysis wins if they disagree. This is intended to help
209 // support "obvious" type-punning idioms.
210 PM.add(createTypeBasedAAWrapperPass());
211 PM.add(createScopedNoAliasAAWrapperPass());
214 void PassManagerBuilder::addInstructionCombiningPass(
215 legacy::PassManagerBase &PM) const {
216 bool ExpensiveCombines = OptLevel > 2;
217 PM.add(createInstructionCombiningPass(ExpensiveCombines));
220 void PassManagerBuilder::populateFunctionPassManager(
221 legacy::FunctionPassManager &FPM) {
222 addExtensionsToPM(EP_EarlyAsPossible, FPM);
224 // Add LibraryInfo if we have some.
226 FPM.add(new TargetLibraryInfoWrapperPass(*LibraryInfo));
228 if (OptLevel == 0) return;
230 addInitialAliasAnalysisPasses(FPM);
232 FPM.add(createCFGSimplificationPass());
233 FPM.add(createSROAPass());
234 FPM.add(createEarlyCSEPass());
235 FPM.add(createGVNHoistPass());
236 FPM.add(createLowerExpectIntrinsicPass());
239 // Do PGO instrumentation generation or use pass as the option specified.
240 void PassManagerBuilder::addPGOInstrPasses(legacy::PassManagerBase &MPM) {
241 if (PGOInstrGen.empty() && PGOInstrUse.empty())
243 // Perform the preinline and cleanup passes for O1 and above.
244 // And avoid doing them if optimizing for size.
245 if (OptLevel > 0 && SizeLevel == 0 && !DisablePreInliner) {
246 // Create preinline pass.
247 MPM.add(createFunctionInliningPass(PreInlineThreshold));
248 MPM.add(createSROAPass());
249 MPM.add(createEarlyCSEPass()); // Catch trivial redundancies
250 MPM.add(createCFGSimplificationPass()); // Merge & remove BBs
251 MPM.add(createInstructionCombiningPass()); // Combine silly seq's
252 addExtensionsToPM(EP_Peephole, MPM);
254 if (!PGOInstrGen.empty()) {
255 MPM.add(createPGOInstrumentationGenLegacyPass());
256 // Add the profile lowering pass.
257 InstrProfOptions Options;
258 Options.InstrProfileOutput = PGOInstrGen;
259 MPM.add(createInstrProfilingLegacyPass(Options));
261 if (!PGOInstrUse.empty())
262 MPM.add(createPGOInstrumentationUseLegacyPass(PGOInstrUse));
264 void PassManagerBuilder::addFunctionSimplificationPasses(
265 legacy::PassManagerBase &MPM) {
266 // Start of function pass.
267 // Break up aggregate allocas, using SSAUpdater.
268 MPM.add(createSROAPass());
269 MPM.add(createEarlyCSEPass()); // Catch trivial redundancies
270 // Speculative execution if the target has divergent branches; otherwise nop.
271 MPM.add(createSpeculativeExecutionIfHasBranchDivergencePass());
272 MPM.add(createJumpThreadingPass()); // Thread jumps.
273 MPM.add(createCorrelatedValuePropagationPass()); // Propagate conditionals
274 MPM.add(createCFGSimplificationPass()); // Merge & remove BBs
275 // Combine silly seq's
276 addInstructionCombiningPass(MPM);
277 addExtensionsToPM(EP_Peephole, MPM);
279 MPM.add(createTailCallEliminationPass()); // Eliminate tail calls
280 MPM.add(createCFGSimplificationPass()); // Merge & remove BBs
281 MPM.add(createReassociatePass()); // Reassociate expressions
282 // Rotate Loop - disable header duplication at -Oz
283 MPM.add(createLoopRotatePass(SizeLevel == 2 ? 0 : -1));
284 MPM.add(createLICMPass()); // Hoist loop invariants
285 MPM.add(createLoopUnswitchPass(SizeLevel || OptLevel < 3));
286 MPM.add(createCFGSimplificationPass());
287 addInstructionCombiningPass(MPM);
288 MPM.add(createIndVarSimplifyPass()); // Canonicalize indvars
289 MPM.add(createLoopIdiomPass()); // Recognize idioms like memset.
290 MPM.add(createLoopDeletionPass()); // Delete dead loops
291 if (EnableLoopInterchange) {
292 MPM.add(createLoopInterchangePass()); // Interchange loops
293 MPM.add(createCFGSimplificationPass());
295 if (!DisableUnrollLoops)
296 MPM.add(createSimpleLoopUnrollPass()); // Unroll small loops
297 addExtensionsToPM(EP_LoopOptimizerEnd, MPM);
301 MPM.add(createMergedLoadStoreMotionPass()); // Merge ld/st in diamonds
302 MPM.add(createGVNPass(DisableGVNLoadPRE)); // Remove redundancies
304 MPM.add(createMemCpyOptPass()); // Remove memcpy / form memset
305 MPM.add(createSCCPPass()); // Constant prop with SCCP
307 // Delete dead bit computations (instcombine runs after to fold away the dead
308 // computations, and then ADCE will run later to exploit any new DCE
309 // opportunities that creates).
310 MPM.add(createBitTrackingDCEPass()); // Delete dead bit computations
312 // Run instcombine after redundancy elimination to exploit opportunities
313 // opened up by them.
314 addInstructionCombiningPass(MPM);
315 addExtensionsToPM(EP_Peephole, MPM);
316 MPM.add(createJumpThreadingPass()); // Thread jumps
317 MPM.add(createCorrelatedValuePropagationPass());
318 MPM.add(createDeadStoreEliminationPass()); // Delete dead stores
319 MPM.add(createLICMPass());
321 addExtensionsToPM(EP_ScalarOptimizerLate, MPM);
324 MPM.add(createLoopRerollPass());
325 if (!RunSLPAfterLoopVectorization) {
327 MPM.add(createSLPVectorizerPass()); // Vectorize parallel scalar chains.
330 MPM.add(createBBVectorizePass());
331 addInstructionCombiningPass(MPM);
332 addExtensionsToPM(EP_Peephole, MPM);
333 if (OptLevel > 1 && UseGVNAfterVectorization)
334 MPM.add(createGVNPass(DisableGVNLoadPRE)); // Remove redundancies
336 MPM.add(createEarlyCSEPass()); // Catch trivial redundancies
338 // BBVectorize may have significantly shortened a loop body; unroll again.
339 if (!DisableUnrollLoops)
340 MPM.add(createLoopUnrollPass());
345 MPM.add(createLoadCombinePass());
347 MPM.add(createAggressiveDCEPass()); // Delete dead instructions
348 MPM.add(createCFGSimplificationPass()); // Merge & remove BBs
349 // Clean up after everything.
350 addInstructionCombiningPass(MPM);
351 addExtensionsToPM(EP_Peephole, MPM);
354 void PassManagerBuilder::populateModulePassManager(
355 legacy::PassManagerBase &MPM) {
356 // Allow forcing function attributes as a debugging and tuning aid.
357 MPM.add(createForceFunctionAttrsLegacyPass());
359 // If all optimizations are disabled, just run the always-inline pass and,
360 // if enabled, the function merging pass.
362 addPGOInstrPasses(MPM);
368 // FIXME: The BarrierNoopPass is a HACK! The inliner pass above implicitly
369 // creates a CGSCC pass manager, but we don't want to add extensions into
370 // that pass manager. To prevent this we insert a no-op module pass to reset
371 // the pass manager to get the same behavior as EP_OptimizerLast in non-O0
372 // builds. The function merging pass is
374 MPM.add(createMergeFunctionsPass());
375 else if (!GlobalExtensions->empty() || !Extensions.empty())
376 MPM.add(createBarrierNoopPass());
378 addExtensionsToPM(EP_EnabledOnOptLevel0, MPM);
382 // Add LibraryInfo if we have some.
384 MPM.add(new TargetLibraryInfoWrapperPass(*LibraryInfo));
386 addInitialAliasAnalysisPasses(MPM);
388 if (!DisableUnitAtATime) {
389 // Infer attributes about declarations if possible.
390 MPM.add(createInferFunctionAttrsLegacyPass());
392 addExtensionsToPM(EP_ModuleOptimizerEarly, MPM);
394 MPM.add(createIPSCCPPass()); // IP SCCP
395 MPM.add(createGlobalOptimizerPass()); // Optimize out global vars
396 // Promote any localized global vars.
397 MPM.add(createPromoteMemoryToRegisterPass());
399 MPM.add(createDeadArgEliminationPass()); // Dead argument elimination
401 addInstructionCombiningPass(MPM); // Clean up after IPCP & DAE
402 addExtensionsToPM(EP_Peephole, MPM);
403 MPM.add(createCFGSimplificationPass()); // Clean up after IPCP & DAE
406 if (!PerformThinLTO) {
407 /// PGO instrumentation is added during the compile phase for ThinLTO, do
408 /// not run it a second time
409 addPGOInstrPasses(MPM);
412 // Indirect call promotion that promotes intra-module targets only.
413 MPM.add(createPGOIndirectCallPromotionLegacyPass());
415 if (EnableNonLTOGlobalsModRef)
416 // We add a module alias analysis pass here. In part due to bugs in the
417 // analysis infrastructure this "works" in that the analysis stays alive
418 // for the entire SCC pass run below.
419 MPM.add(createGlobalsAAWrapperPass());
421 // Start of CallGraph SCC passes.
422 if (!DisableUnitAtATime)
423 MPM.add(createPruneEHPass()); // Remove dead EH info
428 if (!DisableUnitAtATime)
429 MPM.add(createPostOrderFunctionAttrsLegacyPass());
431 MPM.add(createArgumentPromotionPass()); // Scalarize uninlined fn args
433 addFunctionSimplificationPasses(MPM);
435 // FIXME: This is a HACK! The inliner pass above implicitly creates a CGSCC
436 // pass manager that we are specifically trying to avoid. To prevent this
437 // we must insert a no-op module pass to reset the pass manager.
438 MPM.add(createBarrierNoopPass());
440 if (!DisableUnitAtATime && OptLevel > 1 && !PrepareForLTO &&
442 // Remove avail extern fns and globals definitions if we aren't
443 // compiling an object file for later LTO. For LTO we want to preserve
444 // these so they are eligible for inlining at link-time. Note if they
445 // are unreferenced they will be removed by GlobalDCE later, so
446 // this only impacts referenced available externally globals.
447 // Eventually they will be suppressed during codegen, but eliminating
448 // here enables more opportunity for GlobalDCE as it may make
449 // globals referenced by available external functions dead
450 // and saves running remaining passes on the eliminated functions.
451 MPM.add(createEliminateAvailableExternallyPass());
453 if (!DisableUnitAtATime)
454 MPM.add(createReversePostOrderFunctionAttrsPass());
456 // If we are planning to perform ThinLTO later, let's not bloat the code with
457 // unrolling/vectorization/... now. We'll first run the inliner + CGSCC passes
458 // during ThinLTO and perform the rest of the optimizations afterward.
459 if (PrepareForThinLTO) {
460 // Reduce the size of the IR as much as possible.
461 MPM.add(createGlobalOptimizerPass());
462 // Rename anon function to be able to export them in the summary.
463 MPM.add(createNameAnonFunctionPass());
468 // Optimize globals now when performing ThinLTO, this enables more
469 // optimizations later.
470 MPM.add(createGlobalOptimizerPass());
472 // Scheduling LoopVersioningLICM when inlining is over, because after that
473 // we may see more accurate aliasing. Reason to run this late is that too
474 // early versioning may prevent further inlining due to increase of code
475 // size. By placing it just after inlining other optimizations which runs
476 // later might get benefit of no-alias assumption in clone loop.
477 if (UseLoopVersioningLICM) {
478 MPM.add(createLoopVersioningLICMPass()); // Do LoopVersioningLICM
479 MPM.add(createLICMPass()); // Hoist loop invariants
482 if (EnableNonLTOGlobalsModRef)
483 // We add a fresh GlobalsModRef run at this point. This is particularly
484 // useful as the above will have inlined, DCE'ed, and function-attr
485 // propagated everything. We should at this point have a reasonably minimal
486 // and richly annotated call graph. By computing aliasing and mod/ref
487 // information for all local globals here, the late loop passes and notably
488 // the vectorizer will be able to use them to help recognize vectorizable
489 // memory operations.
491 // Note that this relies on a bug in the pass manager which preserves
492 // a module analysis into a function pass pipeline (and throughout it) so
493 // long as the first function pass doesn't invalidate the module analysis.
494 // Thus both Float2Int and LoopRotate have to preserve AliasAnalysis for
495 // this to work. Fortunately, it is trivial to preserve AliasAnalysis
496 // (doing nothing preserves it as it is required to be conservatively
497 // correct in the face of IR changes).
498 MPM.add(createGlobalsAAWrapperPass());
501 MPM.add(createFloat2IntPass());
503 addExtensionsToPM(EP_VectorizerStart, MPM);
505 // Re-rotate loops in all our loop nests. These may have fallout out of
506 // rotated form due to GVN or other transformations, and the vectorizer relies
507 // on the rotated form. Disable header duplication at -Oz.
508 MPM.add(createLoopRotatePass(SizeLevel == 2 ? 0 : -1));
510 // Distribute loops to allow partial vectorization. I.e. isolate dependences
511 // into separate loop that would otherwise inhibit vectorization. This is
512 // currently only performed for loops marked with the metadata
513 // llvm.loop.distribute=true or when -enable-loop-distribute is specified.
514 MPM.add(createLoopDistributePass(/*ProcessAllLoopsByDefault=*/false));
516 MPM.add(createLoopVectorizePass(DisableUnrollLoops, LoopVectorize));
518 // Eliminate loads by forwarding stores from the previous iteration to loads
519 // of the current iteration.
520 if (EnableLoopLoadElim)
521 MPM.add(createLoopLoadEliminationPass());
523 // FIXME: Because of #pragma vectorize enable, the passes below are always
524 // inserted in the pipeline, even when the vectorizer doesn't run (ex. when
525 // on -O1 and no #pragma is found). Would be good to have these two passes
526 // as function calls, so that we can only pass them when the vectorizer
528 addInstructionCombiningPass(MPM);
529 if (OptLevel > 1 && ExtraVectorizerPasses) {
530 // At higher optimization levels, try to clean up any runtime overlap and
531 // alignment checks inserted by the vectorizer. We want to track correllated
532 // runtime checks for two inner loops in the same outer loop, fold any
533 // common computations, hoist loop-invariant aspects out of any outer loop,
534 // and unswitch the runtime checks if possible. Once hoisted, we may have
535 // dead (or speculatable) control flows or more combining opportunities.
536 MPM.add(createEarlyCSEPass());
537 MPM.add(createCorrelatedValuePropagationPass());
538 addInstructionCombiningPass(MPM);
539 MPM.add(createLICMPass());
540 MPM.add(createLoopUnswitchPass(SizeLevel || OptLevel < 3));
541 MPM.add(createCFGSimplificationPass());
542 addInstructionCombiningPass(MPM);
545 if (RunSLPAfterLoopVectorization) {
547 MPM.add(createSLPVectorizerPass()); // Vectorize parallel scalar chains.
548 if (OptLevel > 1 && ExtraVectorizerPasses) {
549 MPM.add(createEarlyCSEPass());
554 MPM.add(createBBVectorizePass());
555 addInstructionCombiningPass(MPM);
556 addExtensionsToPM(EP_Peephole, MPM);
557 if (OptLevel > 1 && UseGVNAfterVectorization)
558 MPM.add(createGVNPass(DisableGVNLoadPRE)); // Remove redundancies
560 MPM.add(createEarlyCSEPass()); // Catch trivial redundancies
562 // BBVectorize may have significantly shortened a loop body; unroll again.
563 if (!DisableUnrollLoops)
564 MPM.add(createLoopUnrollPass());
568 addExtensionsToPM(EP_Peephole, MPM);
569 MPM.add(createCFGSimplificationPass());
570 addInstructionCombiningPass(MPM);
572 if (!DisableUnrollLoops) {
573 MPM.add(createLoopUnrollPass()); // Unroll small loops
575 // LoopUnroll may generate some redundency to cleanup.
576 addInstructionCombiningPass(MPM);
578 // Runtime unrolling will introduce runtime check in loop prologue. If the
579 // unrolled loop is a inner loop, then the prologue will be inside the
580 // outer loop. LICM pass can help to promote the runtime check out if the
581 // checked value is loop invariant.
582 MPM.add(createLICMPass());
584 // Get rid of LCSSA nodes.
585 MPM.add(createInstructionSimplifierPass());
588 // After vectorization and unrolling, assume intrinsics may tell us more
589 // about pointer alignments.
590 MPM.add(createAlignmentFromAssumptionsPass());
592 if (!DisableUnitAtATime) {
593 // FIXME: We shouldn't bother with this anymore.
594 MPM.add(createStripDeadPrototypesPass()); // Get rid of dead prototypes
596 // GlobalOpt already deletes dead functions and globals, at -O2 try a
597 // late pass of GlobalDCE. It is capable of deleting dead cycles.
599 MPM.add(createGlobalDCEPass()); // Remove dead fns and globals.
600 MPM.add(createConstantMergePass()); // Merge dup global constants
605 MPM.add(createMergeFunctionsPass());
607 addExtensionsToPM(EP_OptimizerLast, MPM);
610 void PassManagerBuilder::addLTOOptimizationPasses(legacy::PassManagerBase &PM) {
611 // Remove unused virtual tables to improve the quality of code generated by
612 // whole-program devirtualization and bitset lowering.
613 PM.add(createGlobalDCEPass());
615 // Provide AliasAnalysis services for optimizations.
616 addInitialAliasAnalysisPasses(PM);
619 PM.add(createFunctionImportPass(ModuleSummary));
621 // Allow forcing function attributes as a debugging and tuning aid.
622 PM.add(createForceFunctionAttrsLegacyPass());
624 // Infer attributes about declarations if possible.
625 PM.add(createInferFunctionAttrsLegacyPass());
628 // Indirect call promotion. This should promote all the targets that are
629 // left by the earlier promotion pass that promotes intra-module targets.
630 // This two-step promotion is to save the compile time. For LTO, it should
631 // produce the same result as if we only do promotion here.
632 PM.add(createPGOIndirectCallPromotionLegacyPass(true));
634 // Propagate constants at call sites into the functions they call. This
635 // opens opportunities for globalopt (and inlining) by substituting function
636 // pointers passed as arguments to direct uses of functions.
637 PM.add(createIPSCCPPass());
640 // Infer attributes about definitions. The readnone attribute in particular is
641 // required for virtual constant propagation.
642 PM.add(createPostOrderFunctionAttrsLegacyPass());
643 PM.add(createReversePostOrderFunctionAttrsPass());
645 // Apply whole-program devirtualization and virtual constant propagation.
646 PM.add(createWholeProgramDevirtPass());
648 // That's all we need at opt level 1.
652 // Now that we internalized some globals, see if we can hack on them!
653 PM.add(createGlobalOptimizerPass());
654 // Promote any localized global vars.
655 PM.add(createPromoteMemoryToRegisterPass());
657 // Linking modules together can lead to duplicated global constants, only
658 // keep one copy of each constant.
659 PM.add(createConstantMergePass());
661 // Remove unused arguments from functions.
662 PM.add(createDeadArgEliminationPass());
664 // Reduce the code after globalopt and ipsccp. Both can open up significant
665 // simplification opportunities, and both can propagate functions through
666 // function pointers. When this happens, we often have to resolve varargs
667 // calls, etc, so let instcombine do this.
668 addInstructionCombiningPass(PM);
669 addExtensionsToPM(EP_Peephole, PM);
671 // Inline small functions
672 bool RunInliner = Inliner;
678 PM.add(createPruneEHPass()); // Remove dead EH info.
680 // Optimize globals again if we ran the inliner.
682 PM.add(createGlobalOptimizerPass());
683 PM.add(createGlobalDCEPass()); // Remove dead functions.
685 // If we didn't decide to inline a function, check to see if we can
686 // transform it to pass arguments by value instead of by reference.
687 PM.add(createArgumentPromotionPass());
689 // The IPO passes may leave cruft around. Clean up after them.
690 addInstructionCombiningPass(PM);
691 addExtensionsToPM(EP_Peephole, PM);
692 PM.add(createJumpThreadingPass());
695 PM.add(createSROAPass());
697 // Run a few AA driven optimizations here and now, to cleanup the code.
698 PM.add(createPostOrderFunctionAttrsLegacyPass()); // Add nocapture.
699 PM.add(createGlobalsAAWrapperPass()); // IP alias analysis.
701 PM.add(createLICMPass()); // Hoist loop invariants.
703 PM.add(createMergedLoadStoreMotionPass()); // Merge ld/st in diamonds.
704 PM.add(createGVNPass(DisableGVNLoadPRE)); // Remove redundancies.
705 PM.add(createMemCpyOptPass()); // Remove dead memcpys.
708 PM.add(createDeadStoreEliminationPass());
710 // More loops are countable; try to optimize them.
711 PM.add(createIndVarSimplifyPass());
712 PM.add(createLoopDeletionPass());
713 if (EnableLoopInterchange)
714 PM.add(createLoopInterchangePass());
716 if (!DisableUnrollLoops)
717 PM.add(createSimpleLoopUnrollPass()); // Unroll small loops
718 PM.add(createLoopVectorizePass(true, LoopVectorize));
719 // The vectorizer may have significantly shortened a loop body; unroll again.
720 if (!DisableUnrollLoops)
721 PM.add(createLoopUnrollPass());
723 // Now that we've optimized loops (in particular loop induction variables),
724 // we may have exposed more scalar opportunities. Run parts of the scalar
725 // optimizer again at this point.
726 addInstructionCombiningPass(PM); // Initial cleanup
727 PM.add(createCFGSimplificationPass()); // if-convert
728 PM.add(createSCCPPass()); // Propagate exposed constants
729 addInstructionCombiningPass(PM); // Clean up again
730 PM.add(createBitTrackingDCEPass());
732 // More scalar chains could be vectorized due to more alias information
733 if (RunSLPAfterLoopVectorization)
735 PM.add(createSLPVectorizerPass()); // Vectorize parallel scalar chains.
737 // After vectorization, assume intrinsics may tell us more about pointer
739 PM.add(createAlignmentFromAssumptionsPass());
742 PM.add(createLoadCombinePass());
744 // Cleanup and simplify the code after the scalar optimizations.
745 addInstructionCombiningPass(PM);
746 addExtensionsToPM(EP_Peephole, PM);
748 PM.add(createJumpThreadingPass());
751 void PassManagerBuilder::addLateLTOOptimizationPasses(
752 legacy::PassManagerBase &PM) {
753 // Delete basic blocks, which optimization passes may have killed.
754 PM.add(createCFGSimplificationPass());
756 // Drop bodies of available externally objects to improve GlobalDCE.
757 PM.add(createEliminateAvailableExternallyPass());
759 // Now that we have optimized the program, discard unreachable functions.
760 PM.add(createGlobalDCEPass());
762 // FIXME: this is profitable (for compiler time) to do at -O0 too, but
763 // currently it damages debug info.
765 PM.add(createMergeFunctionsPass());
768 void PassManagerBuilder::populateThinLTOPassManager(
769 legacy::PassManagerBase &PM) {
770 PerformThinLTO = true;
773 PM.add(createVerifierPass());
776 PM.add(createFunctionImportPass(ModuleSummary));
778 populateModulePassManager(PM);
781 PM.add(createVerifierPass());
782 PerformThinLTO = false;
785 void PassManagerBuilder::populateLTOPassManager(legacy::PassManagerBase &PM) {
787 PM.add(new TargetLibraryInfoWrapperPass(*LibraryInfo));
790 PM.add(createVerifierPass());
793 addLTOOptimizationPasses(PM);
795 // Create a function that performs CFI checks for cross-DSO calls with targets
796 // in the current module.
797 PM.add(createCrossDSOCFIPass());
799 // Lower type metadata and the type.test intrinsic. This pass supports Clang's
800 // control flow integrity mechanisms (-fsanitize=cfi*) and needs to run at
801 // link time if CFI is enabled. The pass does nothing if CFI is disabled.
802 PM.add(createLowerTypeTestsPass());
805 addLateLTOOptimizationPasses(PM);
808 PM.add(createVerifierPass());
811 inline PassManagerBuilder *unwrap(LLVMPassManagerBuilderRef P) {
812 return reinterpret_cast<PassManagerBuilder*>(P);
815 inline LLVMPassManagerBuilderRef wrap(PassManagerBuilder *P) {
816 return reinterpret_cast<LLVMPassManagerBuilderRef>(P);
819 LLVMPassManagerBuilderRef LLVMPassManagerBuilderCreate() {
820 PassManagerBuilder *PMB = new PassManagerBuilder();
824 void LLVMPassManagerBuilderDispose(LLVMPassManagerBuilderRef PMB) {
825 PassManagerBuilder *Builder = unwrap(PMB);
830 LLVMPassManagerBuilderSetOptLevel(LLVMPassManagerBuilderRef PMB,
832 PassManagerBuilder *Builder = unwrap(PMB);
833 Builder->OptLevel = OptLevel;
837 LLVMPassManagerBuilderSetSizeLevel(LLVMPassManagerBuilderRef PMB,
838 unsigned SizeLevel) {
839 PassManagerBuilder *Builder = unwrap(PMB);
840 Builder->SizeLevel = SizeLevel;
844 LLVMPassManagerBuilderSetDisableUnitAtATime(LLVMPassManagerBuilderRef PMB,
846 PassManagerBuilder *Builder = unwrap(PMB);
847 Builder->DisableUnitAtATime = Value;
851 LLVMPassManagerBuilderSetDisableUnrollLoops(LLVMPassManagerBuilderRef PMB,
853 PassManagerBuilder *Builder = unwrap(PMB);
854 Builder->DisableUnrollLoops = Value;
858 LLVMPassManagerBuilderSetDisableSimplifyLibCalls(LLVMPassManagerBuilderRef PMB,
860 // NOTE: The simplify-libcalls pass has been removed.
864 LLVMPassManagerBuilderUseInlinerWithThreshold(LLVMPassManagerBuilderRef PMB,
865 unsigned Threshold) {
866 PassManagerBuilder *Builder = unwrap(PMB);
867 Builder->Inliner = createFunctionInliningPass(Threshold);
871 LLVMPassManagerBuilderPopulateFunctionPassManager(LLVMPassManagerBuilderRef PMB,
872 LLVMPassManagerRef PM) {
873 PassManagerBuilder *Builder = unwrap(PMB);
874 legacy::FunctionPassManager *FPM = unwrap<legacy::FunctionPassManager>(PM);
875 Builder->populateFunctionPassManager(*FPM);
879 LLVMPassManagerBuilderPopulateModulePassManager(LLVMPassManagerBuilderRef PMB,
880 LLVMPassManagerRef PM) {
881 PassManagerBuilder *Builder = unwrap(PMB);
882 legacy::PassManagerBase *MPM = unwrap(PM);
883 Builder->populateModulePassManager(*MPM);
886 void LLVMPassManagerBuilderPopulateLTOPassManager(LLVMPassManagerBuilderRef PMB,
887 LLVMPassManagerRef PM,
888 LLVMBool Internalize,
889 LLVMBool RunInliner) {
890 PassManagerBuilder *Builder = unwrap(PMB);
891 legacy::PassManagerBase *LPM = unwrap(PM);
893 // A small backwards compatibility hack. populateLTOPassManager used to take
894 // an RunInliner option.
895 if (RunInliner && !Builder->Inliner)
896 Builder->Inliner = createFunctionInliningPass();
898 Builder->populateLTOPassManager(*LPM);