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 static cl::opt<bool> EnableGVNHoist(
138 "enable-gvn-hoist", cl::init(false), cl::Hidden,
139 cl::desc("Enable the experimental GVN Hoisting pass"));
141 PassManagerBuilder::PassManagerBuilder() {
144 LibraryInfo = nullptr;
146 ModuleSummary = nullptr;
147 DisableUnitAtATime = false;
148 DisableUnrollLoops = false;
149 BBVectorize = RunBBVectorization;
150 SLPVectorize = RunSLPVectorization;
151 LoopVectorize = RunLoopVectorization;
152 RerollLoops = RunLoopRerolling;
153 LoadCombine = RunLoadCombine;
154 DisableGVNLoadPRE = false;
156 VerifyOutput = false;
157 MergeFunctions = false;
158 PrepareForLTO = false;
159 PGOInstrGen = RunPGOInstrGen;
160 PGOInstrUse = RunPGOInstrUse;
161 PrepareForThinLTO = false;
162 PerformThinLTO = false;
165 PassManagerBuilder::~PassManagerBuilder() {
170 /// Set of global extensions, automatically added as part of the standard set.
171 static ManagedStatic<SmallVector<std::pair<PassManagerBuilder::ExtensionPointTy,
172 PassManagerBuilder::ExtensionFn>, 8> > GlobalExtensions;
174 void PassManagerBuilder::addGlobalExtension(
175 PassManagerBuilder::ExtensionPointTy Ty,
176 PassManagerBuilder::ExtensionFn Fn) {
177 GlobalExtensions->push_back(std::make_pair(Ty, std::move(Fn)));
180 void PassManagerBuilder::addExtension(ExtensionPointTy Ty, ExtensionFn Fn) {
181 Extensions.push_back(std::make_pair(Ty, std::move(Fn)));
184 void PassManagerBuilder::addExtensionsToPM(ExtensionPointTy ETy,
185 legacy::PassManagerBase &PM) const {
186 for (unsigned i = 0, e = GlobalExtensions->size(); i != e; ++i)
187 if ((*GlobalExtensions)[i].first == ETy)
188 (*GlobalExtensions)[i].second(*this, PM);
189 for (unsigned i = 0, e = Extensions.size(); i != e; ++i)
190 if (Extensions[i].first == ETy)
191 Extensions[i].second(*this, PM);
194 void PassManagerBuilder::addInitialAliasAnalysisPasses(
195 legacy::PassManagerBase &PM) const {
197 case CFLAAType::Steensgaard:
198 PM.add(createCFLSteensAAWrapperPass());
200 case CFLAAType::Andersen:
201 PM.add(createCFLAndersAAWrapperPass());
203 case CFLAAType::Both:
204 PM.add(createCFLSteensAAWrapperPass());
205 PM.add(createCFLAndersAAWrapperPass());
211 // Add TypeBasedAliasAnalysis before BasicAliasAnalysis so that
212 // BasicAliasAnalysis wins if they disagree. This is intended to help
213 // support "obvious" type-punning idioms.
214 PM.add(createTypeBasedAAWrapperPass());
215 PM.add(createScopedNoAliasAAWrapperPass());
218 void PassManagerBuilder::addInstructionCombiningPass(
219 legacy::PassManagerBase &PM) const {
220 bool ExpensiveCombines = OptLevel > 2;
221 PM.add(createInstructionCombiningPass(ExpensiveCombines));
224 void PassManagerBuilder::populateFunctionPassManager(
225 legacy::FunctionPassManager &FPM) {
226 addExtensionsToPM(EP_EarlyAsPossible, FPM);
228 // Add LibraryInfo if we have some.
230 FPM.add(new TargetLibraryInfoWrapperPass(*LibraryInfo));
232 if (OptLevel == 0) return;
234 addInitialAliasAnalysisPasses(FPM);
236 FPM.add(createCFGSimplificationPass());
237 FPM.add(createSROAPass());
238 FPM.add(createEarlyCSEPass());
240 FPM.add(createGVNHoistPass());
241 FPM.add(createLowerExpectIntrinsicPass());
244 // Do PGO instrumentation generation or use pass as the option specified.
245 void PassManagerBuilder::addPGOInstrPasses(legacy::PassManagerBase &MPM) {
246 if (PGOInstrGen.empty() && PGOInstrUse.empty())
248 // Perform the preinline and cleanup passes for O1 and above.
249 // And avoid doing them if optimizing for size.
250 if (OptLevel > 0 && SizeLevel == 0 && !DisablePreInliner) {
251 // Create preinline pass.
252 MPM.add(createFunctionInliningPass(PreInlineThreshold));
253 MPM.add(createSROAPass());
254 MPM.add(createEarlyCSEPass()); // Catch trivial redundancies
255 MPM.add(createCFGSimplificationPass()); // Merge & remove BBs
256 MPM.add(createInstructionCombiningPass()); // Combine silly seq's
257 addExtensionsToPM(EP_Peephole, MPM);
259 if (!PGOInstrGen.empty()) {
260 MPM.add(createPGOInstrumentationGenLegacyPass());
261 // Add the profile lowering pass.
262 InstrProfOptions Options;
263 Options.InstrProfileOutput = PGOInstrGen;
264 MPM.add(createInstrProfilingLegacyPass(Options));
266 if (!PGOInstrUse.empty())
267 MPM.add(createPGOInstrumentationUseLegacyPass(PGOInstrUse));
269 void PassManagerBuilder::addFunctionSimplificationPasses(
270 legacy::PassManagerBase &MPM) {
271 // Start of function pass.
272 // Break up aggregate allocas, using SSAUpdater.
273 MPM.add(createSROAPass());
274 MPM.add(createEarlyCSEPass()); // Catch trivial redundancies
275 // Speculative execution if the target has divergent branches; otherwise nop.
276 MPM.add(createSpeculativeExecutionIfHasBranchDivergencePass());
277 MPM.add(createJumpThreadingPass()); // Thread jumps.
278 MPM.add(createCorrelatedValuePropagationPass()); // Propagate conditionals
279 MPM.add(createCFGSimplificationPass()); // Merge & remove BBs
280 // Combine silly seq's
281 addInstructionCombiningPass(MPM);
282 addExtensionsToPM(EP_Peephole, MPM);
284 MPM.add(createTailCallEliminationPass()); // Eliminate tail calls
285 MPM.add(createCFGSimplificationPass()); // Merge & remove BBs
286 MPM.add(createReassociatePass()); // Reassociate expressions
287 // Rotate Loop - disable header duplication at -Oz
288 MPM.add(createLoopRotatePass(SizeLevel == 2 ? 0 : -1));
289 MPM.add(createLICMPass()); // Hoist loop invariants
290 MPM.add(createLoopUnswitchPass(SizeLevel || OptLevel < 3));
291 MPM.add(createCFGSimplificationPass());
292 addInstructionCombiningPass(MPM);
293 MPM.add(createIndVarSimplifyPass()); // Canonicalize indvars
294 MPM.add(createLoopIdiomPass()); // Recognize idioms like memset.
295 MPM.add(createLoopDeletionPass()); // Delete dead loops
296 if (EnableLoopInterchange) {
297 MPM.add(createLoopInterchangePass()); // Interchange loops
298 MPM.add(createCFGSimplificationPass());
300 if (!DisableUnrollLoops)
301 MPM.add(createSimpleLoopUnrollPass()); // Unroll small loops
302 addExtensionsToPM(EP_LoopOptimizerEnd, MPM);
306 MPM.add(createMergedLoadStoreMotionPass()); // Merge ld/st in diamonds
307 MPM.add(createGVNPass(DisableGVNLoadPRE)); // Remove redundancies
309 MPM.add(createMemCpyOptPass()); // Remove memcpy / form memset
310 MPM.add(createSCCPPass()); // Constant prop with SCCP
312 // Delete dead bit computations (instcombine runs after to fold away the dead
313 // computations, and then ADCE will run later to exploit any new DCE
314 // opportunities that creates).
315 MPM.add(createBitTrackingDCEPass()); // Delete dead bit computations
317 // Run instcombine after redundancy elimination to exploit opportunities
318 // opened up by them.
319 addInstructionCombiningPass(MPM);
320 addExtensionsToPM(EP_Peephole, MPM);
321 MPM.add(createJumpThreadingPass()); // Thread jumps
322 MPM.add(createCorrelatedValuePropagationPass());
323 MPM.add(createDeadStoreEliminationPass()); // Delete dead stores
324 MPM.add(createLICMPass());
326 addExtensionsToPM(EP_ScalarOptimizerLate, MPM);
329 MPM.add(createLoopRerollPass());
330 if (!RunSLPAfterLoopVectorization) {
332 MPM.add(createSLPVectorizerPass()); // Vectorize parallel scalar chains.
335 MPM.add(createBBVectorizePass());
336 addInstructionCombiningPass(MPM);
337 addExtensionsToPM(EP_Peephole, MPM);
338 if (OptLevel > 1 && UseGVNAfterVectorization)
339 MPM.add(createGVNPass(DisableGVNLoadPRE)); // Remove redundancies
341 MPM.add(createEarlyCSEPass()); // Catch trivial redundancies
343 // BBVectorize may have significantly shortened a loop body; unroll again.
344 if (!DisableUnrollLoops)
345 MPM.add(createLoopUnrollPass());
350 MPM.add(createLoadCombinePass());
352 MPM.add(createAggressiveDCEPass()); // Delete dead instructions
353 MPM.add(createCFGSimplificationPass()); // Merge & remove BBs
354 // Clean up after everything.
355 addInstructionCombiningPass(MPM);
356 addExtensionsToPM(EP_Peephole, MPM);
359 void PassManagerBuilder::populateModulePassManager(
360 legacy::PassManagerBase &MPM) {
361 // Allow forcing function attributes as a debugging and tuning aid.
362 MPM.add(createForceFunctionAttrsLegacyPass());
364 // If all optimizations are disabled, just run the always-inline pass and,
365 // if enabled, the function merging pass.
367 addPGOInstrPasses(MPM);
373 // FIXME: The BarrierNoopPass is a HACK! The inliner pass above implicitly
374 // creates a CGSCC pass manager, but we don't want to add extensions into
375 // that pass manager. To prevent this we insert a no-op module pass to reset
376 // the pass manager to get the same behavior as EP_OptimizerLast in non-O0
377 // builds. The function merging pass is
379 MPM.add(createMergeFunctionsPass());
380 else if (!GlobalExtensions->empty() || !Extensions.empty())
381 MPM.add(createBarrierNoopPass());
383 addExtensionsToPM(EP_EnabledOnOptLevel0, MPM);
387 // Add LibraryInfo if we have some.
389 MPM.add(new TargetLibraryInfoWrapperPass(*LibraryInfo));
391 addInitialAliasAnalysisPasses(MPM);
393 if (!DisableUnitAtATime) {
394 // Infer attributes about declarations if possible.
395 MPM.add(createInferFunctionAttrsLegacyPass());
397 addExtensionsToPM(EP_ModuleOptimizerEarly, MPM);
399 MPM.add(createIPSCCPPass()); // IP SCCP
400 MPM.add(createGlobalOptimizerPass()); // Optimize out global vars
401 // Promote any localized global vars.
402 MPM.add(createPromoteMemoryToRegisterPass());
404 MPM.add(createDeadArgEliminationPass()); // Dead argument elimination
406 addInstructionCombiningPass(MPM); // Clean up after IPCP & DAE
407 addExtensionsToPM(EP_Peephole, MPM);
408 MPM.add(createCFGSimplificationPass()); // Clean up after IPCP & DAE
411 if (!PerformThinLTO) {
412 /// PGO instrumentation is added during the compile phase for ThinLTO, do
413 /// not run it a second time
414 addPGOInstrPasses(MPM);
417 // Indirect call promotion that promotes intra-module targets only.
418 MPM.add(createPGOIndirectCallPromotionLegacyPass());
420 if (EnableNonLTOGlobalsModRef)
421 // We add a module alias analysis pass here. In part due to bugs in the
422 // analysis infrastructure this "works" in that the analysis stays alive
423 // for the entire SCC pass run below.
424 MPM.add(createGlobalsAAWrapperPass());
426 // Start of CallGraph SCC passes.
427 if (!DisableUnitAtATime)
428 MPM.add(createPruneEHPass()); // Remove dead EH info
433 if (!DisableUnitAtATime)
434 MPM.add(createPostOrderFunctionAttrsLegacyPass());
436 MPM.add(createArgumentPromotionPass()); // Scalarize uninlined fn args
438 addFunctionSimplificationPasses(MPM);
440 // FIXME: This is a HACK! The inliner pass above implicitly creates a CGSCC
441 // pass manager that we are specifically trying to avoid. To prevent this
442 // we must insert a no-op module pass to reset the pass manager.
443 MPM.add(createBarrierNoopPass());
445 if (!DisableUnitAtATime && OptLevel > 1 && !PrepareForLTO &&
447 // Remove avail extern fns and globals definitions if we aren't
448 // compiling an object file for later LTO. For LTO we want to preserve
449 // these so they are eligible for inlining at link-time. Note if they
450 // are unreferenced they will be removed by GlobalDCE later, so
451 // this only impacts referenced available externally globals.
452 // Eventually they will be suppressed during codegen, but eliminating
453 // here enables more opportunity for GlobalDCE as it may make
454 // globals referenced by available external functions dead
455 // and saves running remaining passes on the eliminated functions.
456 MPM.add(createEliminateAvailableExternallyPass());
458 if (!DisableUnitAtATime)
459 MPM.add(createReversePostOrderFunctionAttrsPass());
461 // If we are planning to perform ThinLTO later, let's not bloat the code with
462 // unrolling/vectorization/... now. We'll first run the inliner + CGSCC passes
463 // during ThinLTO and perform the rest of the optimizations afterward.
464 if (PrepareForThinLTO) {
465 // Reduce the size of the IR as much as possible.
466 MPM.add(createGlobalOptimizerPass());
467 // Rename anon function to be able to export them in the summary.
468 MPM.add(createNameAnonFunctionPass());
473 // Optimize globals now when performing ThinLTO, this enables more
474 // optimizations later.
475 MPM.add(createGlobalOptimizerPass());
477 // Scheduling LoopVersioningLICM when inlining is over, because after that
478 // we may see more accurate aliasing. Reason to run this late is that too
479 // early versioning may prevent further inlining due to increase of code
480 // size. By placing it just after inlining other optimizations which runs
481 // later might get benefit of no-alias assumption in clone loop.
482 if (UseLoopVersioningLICM) {
483 MPM.add(createLoopVersioningLICMPass()); // Do LoopVersioningLICM
484 MPM.add(createLICMPass()); // Hoist loop invariants
487 if (EnableNonLTOGlobalsModRef)
488 // We add a fresh GlobalsModRef run at this point. This is particularly
489 // useful as the above will have inlined, DCE'ed, and function-attr
490 // propagated everything. We should at this point have a reasonably minimal
491 // and richly annotated call graph. By computing aliasing and mod/ref
492 // information for all local globals here, the late loop passes and notably
493 // the vectorizer will be able to use them to help recognize vectorizable
494 // memory operations.
496 // Note that this relies on a bug in the pass manager which preserves
497 // a module analysis into a function pass pipeline (and throughout it) so
498 // long as the first function pass doesn't invalidate the module analysis.
499 // Thus both Float2Int and LoopRotate have to preserve AliasAnalysis for
500 // this to work. Fortunately, it is trivial to preserve AliasAnalysis
501 // (doing nothing preserves it as it is required to be conservatively
502 // correct in the face of IR changes).
503 MPM.add(createGlobalsAAWrapperPass());
506 MPM.add(createFloat2IntPass());
508 addExtensionsToPM(EP_VectorizerStart, MPM);
510 // Re-rotate loops in all our loop nests. These may have fallout out of
511 // rotated form due to GVN or other transformations, and the vectorizer relies
512 // on the rotated form. Disable header duplication at -Oz.
513 MPM.add(createLoopRotatePass(SizeLevel == 2 ? 0 : -1));
515 // Distribute loops to allow partial vectorization. I.e. isolate dependences
516 // into separate loop that would otherwise inhibit vectorization. This is
517 // currently only performed for loops marked with the metadata
518 // llvm.loop.distribute=true or when -enable-loop-distribute is specified.
519 MPM.add(createLoopDistributePass(/*ProcessAllLoopsByDefault=*/false));
521 MPM.add(createLoopVectorizePass(DisableUnrollLoops, LoopVectorize));
523 // Eliminate loads by forwarding stores from the previous iteration to loads
524 // of the current iteration.
525 if (EnableLoopLoadElim)
526 MPM.add(createLoopLoadEliminationPass());
528 // FIXME: Because of #pragma vectorize enable, the passes below are always
529 // inserted in the pipeline, even when the vectorizer doesn't run (ex. when
530 // on -O1 and no #pragma is found). Would be good to have these two passes
531 // as function calls, so that we can only pass them when the vectorizer
533 addInstructionCombiningPass(MPM);
534 if (OptLevel > 1 && ExtraVectorizerPasses) {
535 // At higher optimization levels, try to clean up any runtime overlap and
536 // alignment checks inserted by the vectorizer. We want to track correllated
537 // runtime checks for two inner loops in the same outer loop, fold any
538 // common computations, hoist loop-invariant aspects out of any outer loop,
539 // and unswitch the runtime checks if possible. Once hoisted, we may have
540 // dead (or speculatable) control flows or more combining opportunities.
541 MPM.add(createEarlyCSEPass());
542 MPM.add(createCorrelatedValuePropagationPass());
543 addInstructionCombiningPass(MPM);
544 MPM.add(createLICMPass());
545 MPM.add(createLoopUnswitchPass(SizeLevel || OptLevel < 3));
546 MPM.add(createCFGSimplificationPass());
547 addInstructionCombiningPass(MPM);
550 if (RunSLPAfterLoopVectorization) {
552 MPM.add(createSLPVectorizerPass()); // Vectorize parallel scalar chains.
553 if (OptLevel > 1 && ExtraVectorizerPasses) {
554 MPM.add(createEarlyCSEPass());
559 MPM.add(createBBVectorizePass());
560 addInstructionCombiningPass(MPM);
561 addExtensionsToPM(EP_Peephole, MPM);
562 if (OptLevel > 1 && UseGVNAfterVectorization)
563 MPM.add(createGVNPass(DisableGVNLoadPRE)); // Remove redundancies
565 MPM.add(createEarlyCSEPass()); // Catch trivial redundancies
567 // BBVectorize may have significantly shortened a loop body; unroll again.
568 if (!DisableUnrollLoops)
569 MPM.add(createLoopUnrollPass());
573 addExtensionsToPM(EP_Peephole, MPM);
574 MPM.add(createCFGSimplificationPass());
575 addInstructionCombiningPass(MPM);
577 if (!DisableUnrollLoops) {
578 MPM.add(createLoopUnrollPass()); // Unroll small loops
580 // LoopUnroll may generate some redundency to cleanup.
581 addInstructionCombiningPass(MPM);
583 // Runtime unrolling will introduce runtime check in loop prologue. If the
584 // unrolled loop is a inner loop, then the prologue will be inside the
585 // outer loop. LICM pass can help to promote the runtime check out if the
586 // checked value is loop invariant.
587 MPM.add(createLICMPass());
589 // Get rid of LCSSA nodes.
590 MPM.add(createInstructionSimplifierPass());
593 // After vectorization and unrolling, assume intrinsics may tell us more
594 // about pointer alignments.
595 MPM.add(createAlignmentFromAssumptionsPass());
597 if (!DisableUnitAtATime) {
598 // FIXME: We shouldn't bother with this anymore.
599 MPM.add(createStripDeadPrototypesPass()); // Get rid of dead prototypes
601 // GlobalOpt already deletes dead functions and globals, at -O2 try a
602 // late pass of GlobalDCE. It is capable of deleting dead cycles.
604 MPM.add(createGlobalDCEPass()); // Remove dead fns and globals.
605 MPM.add(createConstantMergePass()); // Merge dup global constants
610 MPM.add(createMergeFunctionsPass());
612 addExtensionsToPM(EP_OptimizerLast, MPM);
615 void PassManagerBuilder::addLTOOptimizationPasses(legacy::PassManagerBase &PM) {
616 // Remove unused virtual tables to improve the quality of code generated by
617 // whole-program devirtualization and bitset lowering.
618 PM.add(createGlobalDCEPass());
620 // Provide AliasAnalysis services for optimizations.
621 addInitialAliasAnalysisPasses(PM);
624 PM.add(createFunctionImportPass(ModuleSummary));
626 // Allow forcing function attributes as a debugging and tuning aid.
627 PM.add(createForceFunctionAttrsLegacyPass());
629 // Infer attributes about declarations if possible.
630 PM.add(createInferFunctionAttrsLegacyPass());
633 // Indirect call promotion. This should promote all the targets that are
634 // left by the earlier promotion pass that promotes intra-module targets.
635 // This two-step promotion is to save the compile time. For LTO, it should
636 // produce the same result as if we only do promotion here.
637 PM.add(createPGOIndirectCallPromotionLegacyPass(true));
639 // Propagate constants at call sites into the functions they call. This
640 // opens opportunities for globalopt (and inlining) by substituting function
641 // pointers passed as arguments to direct uses of functions.
642 PM.add(createIPSCCPPass());
645 // Infer attributes about definitions. The readnone attribute in particular is
646 // required for virtual constant propagation.
647 PM.add(createPostOrderFunctionAttrsLegacyPass());
648 PM.add(createReversePostOrderFunctionAttrsPass());
650 // Apply whole-program devirtualization and virtual constant propagation.
651 PM.add(createWholeProgramDevirtPass());
653 // That's all we need at opt level 1.
657 // Now that we internalized some globals, see if we can hack on them!
658 PM.add(createGlobalOptimizerPass());
659 // Promote any localized global vars.
660 PM.add(createPromoteMemoryToRegisterPass());
662 // Linking modules together can lead to duplicated global constants, only
663 // keep one copy of each constant.
664 PM.add(createConstantMergePass());
666 // Remove unused arguments from functions.
667 PM.add(createDeadArgEliminationPass());
669 // Reduce the code after globalopt and ipsccp. Both can open up significant
670 // simplification opportunities, and both can propagate functions through
671 // function pointers. When this happens, we often have to resolve varargs
672 // calls, etc, so let instcombine do this.
673 addInstructionCombiningPass(PM);
674 addExtensionsToPM(EP_Peephole, PM);
676 // Inline small functions
677 bool RunInliner = Inliner;
683 PM.add(createPruneEHPass()); // Remove dead EH info.
685 // Optimize globals again if we ran the inliner.
687 PM.add(createGlobalOptimizerPass());
688 PM.add(createGlobalDCEPass()); // Remove dead functions.
690 // If we didn't decide to inline a function, check to see if we can
691 // transform it to pass arguments by value instead of by reference.
692 PM.add(createArgumentPromotionPass());
694 // The IPO passes may leave cruft around. Clean up after them.
695 addInstructionCombiningPass(PM);
696 addExtensionsToPM(EP_Peephole, PM);
697 PM.add(createJumpThreadingPass());
700 PM.add(createSROAPass());
702 // Run a few AA driven optimizations here and now, to cleanup the code.
703 PM.add(createPostOrderFunctionAttrsLegacyPass()); // Add nocapture.
704 PM.add(createGlobalsAAWrapperPass()); // IP alias analysis.
706 PM.add(createLICMPass()); // Hoist loop invariants.
708 PM.add(createMergedLoadStoreMotionPass()); // Merge ld/st in diamonds.
709 PM.add(createGVNPass(DisableGVNLoadPRE)); // Remove redundancies.
710 PM.add(createMemCpyOptPass()); // Remove dead memcpys.
713 PM.add(createDeadStoreEliminationPass());
715 // More loops are countable; try to optimize them.
716 PM.add(createIndVarSimplifyPass());
717 PM.add(createLoopDeletionPass());
718 if (EnableLoopInterchange)
719 PM.add(createLoopInterchangePass());
721 if (!DisableUnrollLoops)
722 PM.add(createSimpleLoopUnrollPass()); // Unroll small loops
723 PM.add(createLoopVectorizePass(true, LoopVectorize));
724 // The vectorizer may have significantly shortened a loop body; unroll again.
725 if (!DisableUnrollLoops)
726 PM.add(createLoopUnrollPass());
728 // Now that we've optimized loops (in particular loop induction variables),
729 // we may have exposed more scalar opportunities. Run parts of the scalar
730 // optimizer again at this point.
731 addInstructionCombiningPass(PM); // Initial cleanup
732 PM.add(createCFGSimplificationPass()); // if-convert
733 PM.add(createSCCPPass()); // Propagate exposed constants
734 addInstructionCombiningPass(PM); // Clean up again
735 PM.add(createBitTrackingDCEPass());
737 // More scalar chains could be vectorized due to more alias information
738 if (RunSLPAfterLoopVectorization)
740 PM.add(createSLPVectorizerPass()); // Vectorize parallel scalar chains.
742 // After vectorization, assume intrinsics may tell us more about pointer
744 PM.add(createAlignmentFromAssumptionsPass());
747 PM.add(createLoadCombinePass());
749 // Cleanup and simplify the code after the scalar optimizations.
750 addInstructionCombiningPass(PM);
751 addExtensionsToPM(EP_Peephole, PM);
753 PM.add(createJumpThreadingPass());
756 void PassManagerBuilder::addLateLTOOptimizationPasses(
757 legacy::PassManagerBase &PM) {
758 // Delete basic blocks, which optimization passes may have killed.
759 PM.add(createCFGSimplificationPass());
761 // Drop bodies of available externally objects to improve GlobalDCE.
762 PM.add(createEliminateAvailableExternallyPass());
764 // Now that we have optimized the program, discard unreachable functions.
765 PM.add(createGlobalDCEPass());
767 // FIXME: this is profitable (for compiler time) to do at -O0 too, but
768 // currently it damages debug info.
770 PM.add(createMergeFunctionsPass());
773 void PassManagerBuilder::populateThinLTOPassManager(
774 legacy::PassManagerBase &PM) {
775 PerformThinLTO = true;
778 PM.add(createVerifierPass());
781 PM.add(createFunctionImportPass(ModuleSummary));
783 populateModulePassManager(PM);
786 PM.add(createVerifierPass());
787 PerformThinLTO = false;
790 void PassManagerBuilder::populateLTOPassManager(legacy::PassManagerBase &PM) {
792 PM.add(new TargetLibraryInfoWrapperPass(*LibraryInfo));
795 PM.add(createVerifierPass());
798 addLTOOptimizationPasses(PM);
800 // Create a function that performs CFI checks for cross-DSO calls with targets
801 // in the current module.
802 PM.add(createCrossDSOCFIPass());
804 // Lower type metadata and the type.test intrinsic. This pass supports Clang's
805 // control flow integrity mechanisms (-fsanitize=cfi*) and needs to run at
806 // link time if CFI is enabled. The pass does nothing if CFI is disabled.
807 PM.add(createLowerTypeTestsPass());
810 addLateLTOOptimizationPasses(PM);
813 PM.add(createVerifierPass());
816 inline PassManagerBuilder *unwrap(LLVMPassManagerBuilderRef P) {
817 return reinterpret_cast<PassManagerBuilder*>(P);
820 inline LLVMPassManagerBuilderRef wrap(PassManagerBuilder *P) {
821 return reinterpret_cast<LLVMPassManagerBuilderRef>(P);
824 LLVMPassManagerBuilderRef LLVMPassManagerBuilderCreate() {
825 PassManagerBuilder *PMB = new PassManagerBuilder();
829 void LLVMPassManagerBuilderDispose(LLVMPassManagerBuilderRef PMB) {
830 PassManagerBuilder *Builder = unwrap(PMB);
835 LLVMPassManagerBuilderSetOptLevel(LLVMPassManagerBuilderRef PMB,
837 PassManagerBuilder *Builder = unwrap(PMB);
838 Builder->OptLevel = OptLevel;
842 LLVMPassManagerBuilderSetSizeLevel(LLVMPassManagerBuilderRef PMB,
843 unsigned SizeLevel) {
844 PassManagerBuilder *Builder = unwrap(PMB);
845 Builder->SizeLevel = SizeLevel;
849 LLVMPassManagerBuilderSetDisableUnitAtATime(LLVMPassManagerBuilderRef PMB,
851 PassManagerBuilder *Builder = unwrap(PMB);
852 Builder->DisableUnitAtATime = Value;
856 LLVMPassManagerBuilderSetDisableUnrollLoops(LLVMPassManagerBuilderRef PMB,
858 PassManagerBuilder *Builder = unwrap(PMB);
859 Builder->DisableUnrollLoops = Value;
863 LLVMPassManagerBuilderSetDisableSimplifyLibCalls(LLVMPassManagerBuilderRef PMB,
865 // NOTE: The simplify-libcalls pass has been removed.
869 LLVMPassManagerBuilderUseInlinerWithThreshold(LLVMPassManagerBuilderRef PMB,
870 unsigned Threshold) {
871 PassManagerBuilder *Builder = unwrap(PMB);
872 Builder->Inliner = createFunctionInliningPass(Threshold);
876 LLVMPassManagerBuilderPopulateFunctionPassManager(LLVMPassManagerBuilderRef PMB,
877 LLVMPassManagerRef PM) {
878 PassManagerBuilder *Builder = unwrap(PMB);
879 legacy::FunctionPassManager *FPM = unwrap<legacy::FunctionPassManager>(PM);
880 Builder->populateFunctionPassManager(*FPM);
884 LLVMPassManagerBuilderPopulateModulePassManager(LLVMPassManagerBuilderRef PMB,
885 LLVMPassManagerRef PM) {
886 PassManagerBuilder *Builder = unwrap(PMB);
887 legacy::PassManagerBase *MPM = unwrap(PM);
888 Builder->populateModulePassManager(*MPM);
891 void LLVMPassManagerBuilderPopulateLTOPassManager(LLVMPassManagerBuilderRef PMB,
892 LLVMPassManagerRef PM,
893 LLVMBool Internalize,
894 LLVMBool RunInliner) {
895 PassManagerBuilder *Builder = unwrap(PMB);
896 legacy::PassManagerBase *LPM = unwrap(PM);
898 // A small backwards compatibility hack. populateLTOPassManager used to take
899 // an RunInliner option.
900 if (RunInliner && !Builder->Inliner)
901 Builder->Inliner = createFunctionInliningPass();
903 Builder->populateLTOPassManager(*LPM);