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/InlineCost.h"
23 #include "llvm/Analysis/Passes.h"
24 #include "llvm/Analysis/ScopedNoAliasAA.h"
25 #include "llvm/Analysis/TargetLibraryInfo.h"
26 #include "llvm/Analysis/TypeBasedAliasAnalysis.h"
27 #include "llvm/IR/DataLayout.h"
28 #include "llvm/IR/LegacyPassManager.h"
29 #include "llvm/IR/ModuleSummaryIndex.h"
30 #include "llvm/IR/Verifier.h"
31 #include "llvm/Support/CommandLine.h"
32 #include "llvm/Support/ManagedStatic.h"
33 #include "llvm/Target/TargetMachine.h"
34 #include "llvm/Transforms/IPO.h"
35 #include "llvm/Transforms/IPO/ForceFunctionAttrs.h"
36 #include "llvm/Transforms/IPO/FunctionAttrs.h"
37 #include "llvm/Transforms/IPO/InferFunctionAttrs.h"
38 #include "llvm/Transforms/Instrumentation.h"
39 #include "llvm/Transforms/Scalar.h"
40 #include "llvm/Transforms/Scalar/GVN.h"
41 #include "llvm/Transforms/Vectorize.h"
46 RunLoopVectorization("vectorize-loops", cl::Hidden,
47 cl::desc("Run the Loop vectorization passes"));
50 RunSLPVectorization("vectorize-slp", cl::Hidden,
51 cl::desc("Run the SLP vectorization passes"));
54 RunBBVectorization("vectorize-slp-aggressive", cl::Hidden,
55 cl::desc("Run the BB vectorization passes"));
58 UseGVNAfterVectorization("use-gvn-after-vectorization",
59 cl::init(false), cl::Hidden,
60 cl::desc("Run GVN instead of Early CSE after vectorization passes"));
62 static cl::opt<bool> ExtraVectorizerPasses(
63 "extra-vectorizer-passes", cl::init(false), cl::Hidden,
64 cl::desc("Run cleanup optimization passes after vectorization."));
67 RunLoopRerolling("reroll-loops", cl::Hidden,
68 cl::desc("Run the loop rerolling pass"));
70 static cl::opt<bool> RunLoadCombine("combine-loads", cl::init(false),
72 cl::desc("Run the load combining pass"));
74 static cl::opt<bool> RunNewGVN("enable-newgvn", cl::init(false), cl::Hidden,
75 cl::desc("Run the NewGVN 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")));
97 EnableMLSM("mlsm", cl::init(true), cl::Hidden,
98 cl::desc("Enable motion of merged load and store"));
100 static cl::opt<bool> EnableLoopInterchange(
101 "enable-loopinterchange", cl::init(false), cl::Hidden,
102 cl::desc("Enable the new, experimental LoopInterchange Pass"));
104 static cl::opt<bool> EnableNonLTOGlobalsModRef(
105 "enable-non-lto-gmr", cl::init(true), cl::Hidden,
107 "Enable the GlobalsModRef AliasAnalysis outside of the LTO pipeline."));
109 static cl::opt<bool> EnableLoopLoadElim(
110 "enable-loop-load-elim", cl::init(true), cl::Hidden,
111 cl::desc("Enable the LoopLoadElimination Pass"));
114 EnablePrepareForThinLTO("prepare-for-thinlto", cl::init(false), cl::Hidden,
115 cl::desc("Enable preparation for ThinLTO."));
117 static cl::opt<bool> RunPGOInstrGen(
118 "profile-generate", cl::init(false), cl::Hidden,
119 cl::desc("Enable PGO instrumentation."));
121 static cl::opt<std::string>
122 PGOOutputFile("profile-generate-file", cl::init(""), cl::Hidden,
123 cl::desc("Specify the path of profile data file."));
125 static cl::opt<std::string> RunPGOInstrUse(
126 "profile-use", cl::init(""), cl::Hidden, cl::value_desc("filename"),
127 cl::desc("Enable use phase of PGO instrumentation and specify the path "
128 "of profile data file"));
130 static cl::opt<bool> UseLoopVersioningLICM(
131 "enable-loop-versioning-licm", cl::init(false), cl::Hidden,
132 cl::desc("Enable the experimental Loop Versioning LICM pass"));
135 DisablePreInliner("disable-preinline", cl::init(false), cl::Hidden,
136 cl::desc("Disable pre-instrumentation inliner"));
138 static cl::opt<int> PreInlineThreshold(
139 "preinline-threshold", cl::Hidden, cl::init(75), cl::ZeroOrMore,
140 cl::desc("Control the amount of inlining in pre-instrumentation inliner "
143 static cl::opt<bool> EnableGVNHoist(
144 "enable-gvn-hoist", cl::init(true), cl::Hidden,
145 cl::desc("Enable the GVN hoisting pass (default = on)"));
148 DisableLibCallsShrinkWrap("disable-libcalls-shrinkwrap", cl::init(false),
150 cl::desc("Disable shrink-wrap library calls"));
152 PassManagerBuilder::PassManagerBuilder() {
155 LibraryInfo = nullptr;
157 DisableUnitAtATime = false;
158 DisableUnrollLoops = false;
159 BBVectorize = RunBBVectorization;
160 SLPVectorize = RunSLPVectorization;
161 LoopVectorize = RunLoopVectorization;
162 RerollLoops = RunLoopRerolling;
163 LoadCombine = RunLoadCombine;
165 DisableGVNLoadPRE = false;
167 VerifyOutput = false;
168 MergeFunctions = false;
169 PrepareForLTO = false;
170 EnablePGOInstrGen = RunPGOInstrGen;
171 PGOInstrGen = PGOOutputFile;
172 PGOInstrUse = RunPGOInstrUse;
173 PrepareForThinLTO = EnablePrepareForThinLTO;
174 PerformThinLTO = false;
177 PassManagerBuilder::~PassManagerBuilder() {
182 /// Set of global extensions, automatically added as part of the standard set.
183 static ManagedStatic<SmallVector<std::pair<PassManagerBuilder::ExtensionPointTy,
184 PassManagerBuilder::ExtensionFn>, 8> > GlobalExtensions;
186 void PassManagerBuilder::addGlobalExtension(
187 PassManagerBuilder::ExtensionPointTy Ty,
188 PassManagerBuilder::ExtensionFn Fn) {
189 GlobalExtensions->push_back(std::make_pair(Ty, std::move(Fn)));
192 void PassManagerBuilder::addExtension(ExtensionPointTy Ty, ExtensionFn Fn) {
193 Extensions.push_back(std::make_pair(Ty, std::move(Fn)));
196 void PassManagerBuilder::addExtensionsToPM(ExtensionPointTy ETy,
197 legacy::PassManagerBase &PM) const {
198 for (unsigned i = 0, e = GlobalExtensions->size(); i != e; ++i)
199 if ((*GlobalExtensions)[i].first == ETy)
200 (*GlobalExtensions)[i].second(*this, PM);
201 for (unsigned i = 0, e = Extensions.size(); i != e; ++i)
202 if (Extensions[i].first == ETy)
203 Extensions[i].second(*this, PM);
206 void PassManagerBuilder::addInitialAliasAnalysisPasses(
207 legacy::PassManagerBase &PM) const {
209 case CFLAAType::Steensgaard:
210 PM.add(createCFLSteensAAWrapperPass());
212 case CFLAAType::Andersen:
213 PM.add(createCFLAndersAAWrapperPass());
215 case CFLAAType::Both:
216 PM.add(createCFLSteensAAWrapperPass());
217 PM.add(createCFLAndersAAWrapperPass());
223 // Add TypeBasedAliasAnalysis before BasicAliasAnalysis so that
224 // BasicAliasAnalysis wins if they disagree. This is intended to help
225 // support "obvious" type-punning idioms.
226 PM.add(createTypeBasedAAWrapperPass());
227 PM.add(createScopedNoAliasAAWrapperPass());
230 void PassManagerBuilder::addInstructionCombiningPass(
231 legacy::PassManagerBase &PM) const {
232 bool ExpensiveCombines = OptLevel > 2;
233 PM.add(createInstructionCombiningPass(ExpensiveCombines));
236 void PassManagerBuilder::populateFunctionPassManager(
237 legacy::FunctionPassManager &FPM) {
238 addExtensionsToPM(EP_EarlyAsPossible, FPM);
240 // Add LibraryInfo if we have some.
242 FPM.add(new TargetLibraryInfoWrapperPass(*LibraryInfo));
244 if (OptLevel == 0) return;
246 addInitialAliasAnalysisPasses(FPM);
248 FPM.add(createCFGSimplificationPass());
249 FPM.add(createSROAPass());
250 FPM.add(createEarlyCSEPass());
252 FPM.add(createGVNHoistPass());
253 FPM.add(createLowerExpectIntrinsicPass());
256 // Do PGO instrumentation generation or use pass as the option specified.
257 void PassManagerBuilder::addPGOInstrPasses(legacy::PassManagerBase &MPM) {
258 if (!EnablePGOInstrGen && PGOInstrUse.empty())
260 // Perform the preinline and cleanup passes for O1 and above.
261 // And avoid doing them if optimizing for size.
262 if (OptLevel > 0 && SizeLevel == 0 && !DisablePreInliner) {
263 // Create preinline pass. We construct an InlineParams object and specify
264 // the threshold here to avoid the command line options of the regular
265 // inliner to influence pre-inlining. The only fields of InlineParams we
266 // care about are DefaultThreshold and HintThreshold.
268 IP.DefaultThreshold = PreInlineThreshold;
269 // FIXME: The hint threshold has the same value used by the regular inliner.
270 // This should probably be lowered after performance testing.
271 IP.HintThreshold = 325;
273 MPM.add(createFunctionInliningPass(IP));
274 MPM.add(createSROAPass());
275 MPM.add(createEarlyCSEPass()); // Catch trivial redundancies
276 MPM.add(createCFGSimplificationPass()); // Merge & remove BBs
277 MPM.add(createInstructionCombiningPass()); // Combine silly seq's
278 addExtensionsToPM(EP_Peephole, MPM);
280 if (EnablePGOInstrGen) {
281 MPM.add(createPGOInstrumentationGenLegacyPass());
282 // Add the profile lowering pass.
283 InstrProfOptions Options;
284 if (!PGOInstrGen.empty())
285 Options.InstrProfileOutput = PGOInstrGen;
286 MPM.add(createInstrProfilingLegacyPass(Options));
288 if (!PGOInstrUse.empty())
289 MPM.add(createPGOInstrumentationUseLegacyPass(PGOInstrUse));
291 void PassManagerBuilder::addFunctionSimplificationPasses(
292 legacy::PassManagerBase &MPM) {
293 // Start of function pass.
294 // Break up aggregate allocas, using SSAUpdater.
295 MPM.add(createSROAPass());
296 MPM.add(createEarlyCSEPass()); // Catch trivial redundancies
297 // Speculative execution if the target has divergent branches; otherwise nop.
298 MPM.add(createSpeculativeExecutionIfHasBranchDivergencePass());
299 MPM.add(createJumpThreadingPass()); // Thread jumps.
300 MPM.add(createCorrelatedValuePropagationPass()); // Propagate conditionals
301 MPM.add(createCFGSimplificationPass()); // Merge & remove BBs
302 // Combine silly seq's
303 addInstructionCombiningPass(MPM);
304 if (SizeLevel == 0 && !DisableLibCallsShrinkWrap)
305 MPM.add(createLibCallsShrinkWrapPass());
306 addExtensionsToPM(EP_Peephole, MPM);
308 MPM.add(createTailCallEliminationPass()); // Eliminate tail calls
309 MPM.add(createCFGSimplificationPass()); // Merge & remove BBs
310 MPM.add(createReassociatePass()); // Reassociate expressions
311 // Rotate Loop - disable header duplication at -Oz
312 MPM.add(createLoopRotatePass(SizeLevel == 2 ? 0 : -1));
313 MPM.add(createLICMPass()); // Hoist loop invariants
314 MPM.add(createLoopUnswitchPass(SizeLevel || OptLevel < 3));
315 MPM.add(createCFGSimplificationPass());
316 addInstructionCombiningPass(MPM);
317 MPM.add(createIndVarSimplifyPass()); // Canonicalize indvars
318 MPM.add(createLoopIdiomPass()); // Recognize idioms like memset.
319 MPM.add(createLoopDeletionPass()); // Delete dead loops
320 if (EnableLoopInterchange) {
321 MPM.add(createLoopInterchangePass()); // Interchange loops
322 MPM.add(createCFGSimplificationPass());
324 if (!DisableUnrollLoops)
325 MPM.add(createSimpleLoopUnrollPass()); // Unroll small loops
326 addExtensionsToPM(EP_LoopOptimizerEnd, MPM);
330 MPM.add(createMergedLoadStoreMotionPass()); // Merge ld/st in diamonds
331 MPM.add(NewGVN ? createNewGVNPass()
332 : createGVNPass(DisableGVNLoadPRE)); // Remove redundancies
334 MPM.add(createMemCpyOptPass()); // Remove memcpy / form memset
335 MPM.add(createSCCPPass()); // Constant prop with SCCP
337 // Delete dead bit computations (instcombine runs after to fold away the dead
338 // computations, and then ADCE will run later to exploit any new DCE
339 // opportunities that creates).
340 MPM.add(createBitTrackingDCEPass()); // Delete dead bit computations
342 // Run instcombine after redundancy elimination to exploit opportunities
343 // opened up by them.
344 addInstructionCombiningPass(MPM);
345 addExtensionsToPM(EP_Peephole, MPM);
346 MPM.add(createJumpThreadingPass()); // Thread jumps
347 MPM.add(createCorrelatedValuePropagationPass());
348 MPM.add(createDeadStoreEliminationPass()); // Delete dead stores
349 MPM.add(createLICMPass());
351 addExtensionsToPM(EP_ScalarOptimizerLate, MPM);
354 MPM.add(createLoopRerollPass());
355 if (!RunSLPAfterLoopVectorization) {
357 MPM.add(createSLPVectorizerPass()); // Vectorize parallel scalar chains.
360 MPM.add(createBBVectorizePass());
361 addInstructionCombiningPass(MPM);
362 addExtensionsToPM(EP_Peephole, MPM);
363 if (OptLevel > 1 && UseGVNAfterVectorization)
366 : createGVNPass(DisableGVNLoadPRE)); // Remove redundancies
368 MPM.add(createEarlyCSEPass()); // Catch trivial redundancies
370 // BBVectorize may have significantly shortened a loop body; unroll again.
371 if (!DisableUnrollLoops)
372 MPM.add(createLoopUnrollPass());
377 MPM.add(createLoadCombinePass());
379 MPM.add(createAggressiveDCEPass()); // Delete dead instructions
380 MPM.add(createCFGSimplificationPass()); // Merge & remove BBs
381 // Clean up after everything.
382 addInstructionCombiningPass(MPM);
383 addExtensionsToPM(EP_Peephole, MPM);
386 void PassManagerBuilder::populateModulePassManager(
387 legacy::PassManagerBase &MPM) {
388 if (!PGOSampleUse.empty()) {
389 MPM.add(createPruneEHPass());
390 MPM.add(createSampleProfileLoaderPass(PGOSampleUse));
393 // Allow forcing function attributes as a debugging and tuning aid.
394 MPM.add(createForceFunctionAttrsLegacyPass());
396 // If all optimizations are disabled, just run the always-inline pass and,
397 // if enabled, the function merging pass.
399 addPGOInstrPasses(MPM);
405 // FIXME: The BarrierNoopPass is a HACK! The inliner pass above implicitly
406 // creates a CGSCC pass manager, but we don't want to add extensions into
407 // that pass manager. To prevent this we insert a no-op module pass to reset
408 // the pass manager to get the same behavior as EP_OptimizerLast in non-O0
409 // builds. The function merging pass is
411 MPM.add(createMergeFunctionsPass());
412 else if (!GlobalExtensions->empty() || !Extensions.empty())
413 MPM.add(createBarrierNoopPass());
415 if (PrepareForThinLTO)
416 // Rename anon globals to be able to export them in the summary.
417 MPM.add(createNameAnonGlobalPass());
419 addExtensionsToPM(EP_EnabledOnOptLevel0, MPM);
423 // Add LibraryInfo if we have some.
425 MPM.add(new TargetLibraryInfoWrapperPass(*LibraryInfo));
427 addInitialAliasAnalysisPasses(MPM);
429 // For ThinLTO there are two passes of indirect call promotion. The
430 // first is during the compile phase when PerformThinLTO=false and
431 // intra-module indirect call targets are promoted. The second is during
432 // the ThinLTO backend when PerformThinLTO=true, when we promote imported
433 // inter-module indirect calls. For that we perform indirect call promotion
434 // earlier in the pass pipeline, here before globalopt. Otherwise imported
435 // available_externally functions look unreferenced and are removed.
437 MPM.add(createPGOIndirectCallPromotionLegacyPass(/*InLTO = */ true));
439 if (!DisableUnitAtATime) {
440 // Infer attributes about declarations if possible.
441 MPM.add(createInferFunctionAttrsLegacyPass());
443 addExtensionsToPM(EP_ModuleOptimizerEarly, MPM);
445 MPM.add(createIPSCCPPass()); // IP SCCP
446 MPM.add(createGlobalOptimizerPass()); // Optimize out global vars
447 // Promote any localized global vars.
448 MPM.add(createPromoteMemoryToRegisterPass());
450 MPM.add(createDeadArgEliminationPass()); // Dead argument elimination
452 addInstructionCombiningPass(MPM); // Clean up after IPCP & DAE
453 addExtensionsToPM(EP_Peephole, MPM);
454 MPM.add(createCFGSimplificationPass()); // Clean up after IPCP & DAE
457 if (!PerformThinLTO) {
458 /// PGO instrumentation is added during the compile phase for ThinLTO, do
459 /// not run it a second time
460 addPGOInstrPasses(MPM);
461 // Indirect call promotion that promotes intra-module targets only.
462 // For ThinLTO this is done earlier due to interactions with globalopt
463 // for imported functions.
464 MPM.add(createPGOIndirectCallPromotionLegacyPass());
467 if (EnableNonLTOGlobalsModRef)
468 // We add a module alias analysis pass here. In part due to bugs in the
469 // analysis infrastructure this "works" in that the analysis stays alive
470 // for the entire SCC pass run below.
471 MPM.add(createGlobalsAAWrapperPass());
473 // Start of CallGraph SCC passes.
474 if (!DisableUnitAtATime)
475 MPM.add(createPruneEHPass()); // Remove dead EH info
480 if (!DisableUnitAtATime)
481 MPM.add(createPostOrderFunctionAttrsLegacyPass());
483 MPM.add(createArgumentPromotionPass()); // Scalarize uninlined fn args
485 addExtensionsToPM(EP_CGSCCOptimizerLate, MPM);
486 addFunctionSimplificationPasses(MPM);
488 // FIXME: This is a HACK! The inliner pass above implicitly creates a CGSCC
489 // pass manager that we are specifically trying to avoid. To prevent this
490 // we must insert a no-op module pass to reset the pass manager.
491 MPM.add(createBarrierNoopPass());
493 if (!DisableUnitAtATime && OptLevel > 1 && !PrepareForLTO &&
495 // Remove avail extern fns and globals definitions if we aren't
496 // compiling an object file for later LTO. For LTO we want to preserve
497 // these so they are eligible for inlining at link-time. Note if they
498 // are unreferenced they will be removed by GlobalDCE later, so
499 // this only impacts referenced available externally globals.
500 // Eventually they will be suppressed during codegen, but eliminating
501 // here enables more opportunity for GlobalDCE as it may make
502 // globals referenced by available external functions dead
503 // and saves running remaining passes on the eliminated functions.
504 MPM.add(createEliminateAvailableExternallyPass());
506 if (!DisableUnitAtATime)
507 MPM.add(createReversePostOrderFunctionAttrsPass());
509 // If we are planning to perform ThinLTO later, let's not bloat the code with
510 // unrolling/vectorization/... now. We'll first run the inliner + CGSCC passes
511 // during ThinLTO and perform the rest of the optimizations afterward.
512 if (PrepareForThinLTO) {
513 // Reduce the size of the IR as much as possible.
514 MPM.add(createGlobalOptimizerPass());
515 // Rename anon globals to be able to export them in the summary.
516 MPM.add(createNameAnonGlobalPass());
521 // Optimize globals now when performing ThinLTO, this enables more
522 // optimizations later.
523 MPM.add(createGlobalOptimizerPass());
525 // Scheduling LoopVersioningLICM when inlining is over, because after that
526 // we may see more accurate aliasing. Reason to run this late is that too
527 // early versioning may prevent further inlining due to increase of code
528 // size. By placing it just after inlining other optimizations which runs
529 // later might get benefit of no-alias assumption in clone loop.
530 if (UseLoopVersioningLICM) {
531 MPM.add(createLoopVersioningLICMPass()); // Do LoopVersioningLICM
532 MPM.add(createLICMPass()); // Hoist loop invariants
535 if (EnableNonLTOGlobalsModRef)
536 // We add a fresh GlobalsModRef run at this point. This is particularly
537 // useful as the above will have inlined, DCE'ed, and function-attr
538 // propagated everything. We should at this point have a reasonably minimal
539 // and richly annotated call graph. By computing aliasing and mod/ref
540 // information for all local globals here, the late loop passes and notably
541 // the vectorizer will be able to use them to help recognize vectorizable
542 // memory operations.
544 // Note that this relies on a bug in the pass manager which preserves
545 // a module analysis into a function pass pipeline (and throughout it) so
546 // long as the first function pass doesn't invalidate the module analysis.
547 // Thus both Float2Int and LoopRotate have to preserve AliasAnalysis for
548 // this to work. Fortunately, it is trivial to preserve AliasAnalysis
549 // (doing nothing preserves it as it is required to be conservatively
550 // correct in the face of IR changes).
551 MPM.add(createGlobalsAAWrapperPass());
553 MPM.add(createFloat2IntPass());
555 addExtensionsToPM(EP_VectorizerStart, MPM);
557 // Re-rotate loops in all our loop nests. These may have fallout out of
558 // rotated form due to GVN or other transformations, and the vectorizer relies
559 // on the rotated form. Disable header duplication at -Oz.
560 MPM.add(createLoopRotatePass(SizeLevel == 2 ? 0 : -1));
562 // Distribute loops to allow partial vectorization. I.e. isolate dependences
563 // into separate loop that would otherwise inhibit vectorization. This is
564 // currently only performed for loops marked with the metadata
565 // llvm.loop.distribute=true or when -enable-loop-distribute is specified.
566 MPM.add(createLoopDistributePass());
568 MPM.add(createLoopVectorizePass(DisableUnrollLoops, LoopVectorize));
570 // Eliminate loads by forwarding stores from the previous iteration to loads
571 // of the current iteration.
572 if (EnableLoopLoadElim)
573 MPM.add(createLoopLoadEliminationPass());
575 // FIXME: Because of #pragma vectorize enable, the passes below are always
576 // inserted in the pipeline, even when the vectorizer doesn't run (ex. when
577 // on -O1 and no #pragma is found). Would be good to have these two passes
578 // as function calls, so that we can only pass them when the vectorizer
580 addInstructionCombiningPass(MPM);
581 if (OptLevel > 1 && ExtraVectorizerPasses) {
582 // At higher optimization levels, try to clean up any runtime overlap and
583 // alignment checks inserted by the vectorizer. We want to track correllated
584 // runtime checks for two inner loops in the same outer loop, fold any
585 // common computations, hoist loop-invariant aspects out of any outer loop,
586 // and unswitch the runtime checks if possible. Once hoisted, we may have
587 // dead (or speculatable) control flows or more combining opportunities.
588 MPM.add(createEarlyCSEPass());
589 MPM.add(createCorrelatedValuePropagationPass());
590 addInstructionCombiningPass(MPM);
591 MPM.add(createLICMPass());
592 MPM.add(createLoopUnswitchPass(SizeLevel || OptLevel < 3));
593 MPM.add(createCFGSimplificationPass());
594 addInstructionCombiningPass(MPM);
597 if (RunSLPAfterLoopVectorization) {
599 MPM.add(createSLPVectorizerPass()); // Vectorize parallel scalar chains.
600 if (OptLevel > 1 && ExtraVectorizerPasses) {
601 MPM.add(createEarlyCSEPass());
606 MPM.add(createBBVectorizePass());
607 addInstructionCombiningPass(MPM);
608 addExtensionsToPM(EP_Peephole, MPM);
609 if (OptLevel > 1 && UseGVNAfterVectorization)
612 : createGVNPass(DisableGVNLoadPRE)); // Remove redundancies
614 MPM.add(createEarlyCSEPass()); // Catch trivial redundancies
616 // BBVectorize may have significantly shortened a loop body; unroll again.
617 if (!DisableUnrollLoops)
618 MPM.add(createLoopUnrollPass());
622 addExtensionsToPM(EP_Peephole, MPM);
623 MPM.add(createCFGSimplificationPass());
624 addInstructionCombiningPass(MPM);
626 if (!DisableUnrollLoops) {
627 MPM.add(createLoopUnrollPass()); // Unroll small loops
629 // LoopUnroll may generate some redundency to cleanup.
630 addInstructionCombiningPass(MPM);
632 // Runtime unrolling will introduce runtime check in loop prologue. If the
633 // unrolled loop is a inner loop, then the prologue will be inside the
634 // outer loop. LICM pass can help to promote the runtime check out if the
635 // checked value is loop invariant.
636 MPM.add(createLICMPass());
639 // After vectorization and unrolling, assume intrinsics may tell us more
640 // about pointer alignments.
641 MPM.add(createAlignmentFromAssumptionsPass());
643 if (!DisableUnitAtATime) {
644 // FIXME: We shouldn't bother with this anymore.
645 MPM.add(createStripDeadPrototypesPass()); // Get rid of dead prototypes
647 // GlobalOpt already deletes dead functions and globals, at -O2 try a
648 // late pass of GlobalDCE. It is capable of deleting dead cycles.
650 MPM.add(createGlobalDCEPass()); // Remove dead fns and globals.
651 MPM.add(createConstantMergePass()); // Merge dup global constants
656 MPM.add(createMergeFunctionsPass());
658 // LoopSink pass sinks instructions hoisted by LICM, which serves as a
659 // canonicalization pass that enables other optimizations. As a result,
660 // LoopSink pass needs to be a very late IR pass to avoid undoing LICM
662 MPM.add(createLoopSinkPass());
663 // Get rid of LCSSA nodes.
664 MPM.add(createInstructionSimplifierPass());
665 addExtensionsToPM(EP_OptimizerLast, MPM);
668 void PassManagerBuilder::addLTOOptimizationPasses(legacy::PassManagerBase &PM) {
669 // Remove unused virtual tables to improve the quality of code generated by
670 // whole-program devirtualization and bitset lowering.
671 PM.add(createGlobalDCEPass());
673 // Provide AliasAnalysis services for optimizations.
674 addInitialAliasAnalysisPasses(PM);
676 // Allow forcing function attributes as a debugging and tuning aid.
677 PM.add(createForceFunctionAttrsLegacyPass());
679 // Infer attributes about declarations if possible.
680 PM.add(createInferFunctionAttrsLegacyPass());
683 // Indirect call promotion. This should promote all the targets that are
684 // left by the earlier promotion pass that promotes intra-module targets.
685 // This two-step promotion is to save the compile time. For LTO, it should
686 // produce the same result as if we only do promotion here.
687 PM.add(createPGOIndirectCallPromotionLegacyPass(true));
689 // Propagate constants at call sites into the functions they call. This
690 // opens opportunities for globalopt (and inlining) by substituting function
691 // pointers passed as arguments to direct uses of functions.
692 PM.add(createIPSCCPPass());
695 // Infer attributes about definitions. The readnone attribute in particular is
696 // required for virtual constant propagation.
697 PM.add(createPostOrderFunctionAttrsLegacyPass());
698 PM.add(createReversePostOrderFunctionAttrsPass());
700 // Split globals using inrange annotations on GEP indices. This can help
701 // improve the quality of generated code when virtual constant propagation or
702 // control flow integrity are enabled.
703 PM.add(createGlobalSplitPass());
705 // Apply whole-program devirtualization and virtual constant propagation.
706 PM.add(createWholeProgramDevirtPass());
708 // That's all we need at opt level 1.
712 // Now that we internalized some globals, see if we can hack on them!
713 PM.add(createGlobalOptimizerPass());
714 // Promote any localized global vars.
715 PM.add(createPromoteMemoryToRegisterPass());
717 // Linking modules together can lead to duplicated global constants, only
718 // keep one copy of each constant.
719 PM.add(createConstantMergePass());
721 // Remove unused arguments from functions.
722 PM.add(createDeadArgEliminationPass());
724 // Reduce the code after globalopt and ipsccp. Both can open up significant
725 // simplification opportunities, and both can propagate functions through
726 // function pointers. When this happens, we often have to resolve varargs
727 // calls, etc, so let instcombine do this.
728 addInstructionCombiningPass(PM);
729 addExtensionsToPM(EP_Peephole, PM);
731 // Inline small functions
732 bool RunInliner = Inliner;
738 PM.add(createPruneEHPass()); // Remove dead EH info.
740 // Optimize globals again if we ran the inliner.
742 PM.add(createGlobalOptimizerPass());
743 PM.add(createGlobalDCEPass()); // Remove dead functions.
745 // If we didn't decide to inline a function, check to see if we can
746 // transform it to pass arguments by value instead of by reference.
747 PM.add(createArgumentPromotionPass());
749 // The IPO passes may leave cruft around. Clean up after them.
750 addInstructionCombiningPass(PM);
751 addExtensionsToPM(EP_Peephole, PM);
752 PM.add(createJumpThreadingPass());
755 PM.add(createSROAPass());
757 // Run a few AA driven optimizations here and now, to cleanup the code.
758 PM.add(createPostOrderFunctionAttrsLegacyPass()); // Add nocapture.
759 PM.add(createGlobalsAAWrapperPass()); // IP alias analysis.
761 PM.add(createLICMPass()); // Hoist loop invariants.
763 PM.add(createMergedLoadStoreMotionPass()); // Merge ld/st in diamonds.
764 PM.add(NewGVN ? createNewGVNPass()
765 : createGVNPass(DisableGVNLoadPRE)); // Remove redundancies.
766 PM.add(createMemCpyOptPass()); // Remove dead memcpys.
769 PM.add(createDeadStoreEliminationPass());
771 // More loops are countable; try to optimize them.
772 PM.add(createIndVarSimplifyPass());
773 PM.add(createLoopDeletionPass());
774 if (EnableLoopInterchange)
775 PM.add(createLoopInterchangePass());
777 if (!DisableUnrollLoops)
778 PM.add(createSimpleLoopUnrollPass()); // Unroll small loops
779 PM.add(createLoopVectorizePass(true, LoopVectorize));
780 // The vectorizer may have significantly shortened a loop body; unroll again.
781 if (!DisableUnrollLoops)
782 PM.add(createLoopUnrollPass());
784 // Now that we've optimized loops (in particular loop induction variables),
785 // we may have exposed more scalar opportunities. Run parts of the scalar
786 // optimizer again at this point.
787 addInstructionCombiningPass(PM); // Initial cleanup
788 PM.add(createCFGSimplificationPass()); // if-convert
789 PM.add(createSCCPPass()); // Propagate exposed constants
790 addInstructionCombiningPass(PM); // Clean up again
791 PM.add(createBitTrackingDCEPass());
793 // More scalar chains could be vectorized due to more alias information
794 if (RunSLPAfterLoopVectorization)
796 PM.add(createSLPVectorizerPass()); // Vectorize parallel scalar chains.
798 // After vectorization, assume intrinsics may tell us more about pointer
800 PM.add(createAlignmentFromAssumptionsPass());
803 PM.add(createLoadCombinePass());
805 // Cleanup and simplify the code after the scalar optimizations.
806 addInstructionCombiningPass(PM);
807 addExtensionsToPM(EP_Peephole, PM);
809 PM.add(createJumpThreadingPass());
812 void PassManagerBuilder::addLateLTOOptimizationPasses(
813 legacy::PassManagerBase &PM) {
814 // Delete basic blocks, which optimization passes may have killed.
815 PM.add(createCFGSimplificationPass());
817 // Drop bodies of available externally objects to improve GlobalDCE.
818 PM.add(createEliminateAvailableExternallyPass());
820 // Now that we have optimized the program, discard unreachable functions.
821 PM.add(createGlobalDCEPass());
823 // FIXME: this is profitable (for compiler time) to do at -O0 too, but
824 // currently it damages debug info.
826 PM.add(createMergeFunctionsPass());
829 void PassManagerBuilder::populateThinLTOPassManager(
830 legacy::PassManagerBase &PM) {
831 PerformThinLTO = true;
834 PM.add(createVerifierPass());
836 populateModulePassManager(PM);
839 PM.add(createVerifierPass());
840 PerformThinLTO = false;
843 void PassManagerBuilder::populateLTOPassManager(legacy::PassManagerBase &PM) {
845 PM.add(new TargetLibraryInfoWrapperPass(*LibraryInfo));
848 PM.add(createVerifierPass());
851 addLTOOptimizationPasses(PM);
853 // Create a function that performs CFI checks for cross-DSO calls with targets
854 // in the current module.
855 PM.add(createCrossDSOCFIPass());
857 // Lower type metadata and the type.test intrinsic. This pass supports Clang's
858 // control flow integrity mechanisms (-fsanitize=cfi*) and needs to run at
859 // link time if CFI is enabled. The pass does nothing if CFI is disabled.
860 PM.add(createLowerTypeTestsPass());
863 addLateLTOOptimizationPasses(PM);
866 PM.add(createVerifierPass());
869 inline PassManagerBuilder *unwrap(LLVMPassManagerBuilderRef P) {
870 return reinterpret_cast<PassManagerBuilder*>(P);
873 inline LLVMPassManagerBuilderRef wrap(PassManagerBuilder *P) {
874 return reinterpret_cast<LLVMPassManagerBuilderRef>(P);
877 LLVMPassManagerBuilderRef LLVMPassManagerBuilderCreate() {
878 PassManagerBuilder *PMB = new PassManagerBuilder();
882 void LLVMPassManagerBuilderDispose(LLVMPassManagerBuilderRef PMB) {
883 PassManagerBuilder *Builder = unwrap(PMB);
888 LLVMPassManagerBuilderSetOptLevel(LLVMPassManagerBuilderRef PMB,
890 PassManagerBuilder *Builder = unwrap(PMB);
891 Builder->OptLevel = OptLevel;
895 LLVMPassManagerBuilderSetSizeLevel(LLVMPassManagerBuilderRef PMB,
896 unsigned SizeLevel) {
897 PassManagerBuilder *Builder = unwrap(PMB);
898 Builder->SizeLevel = SizeLevel;
902 LLVMPassManagerBuilderSetDisableUnitAtATime(LLVMPassManagerBuilderRef PMB,
904 PassManagerBuilder *Builder = unwrap(PMB);
905 Builder->DisableUnitAtATime = Value;
909 LLVMPassManagerBuilderSetDisableUnrollLoops(LLVMPassManagerBuilderRef PMB,
911 PassManagerBuilder *Builder = unwrap(PMB);
912 Builder->DisableUnrollLoops = Value;
916 LLVMPassManagerBuilderSetDisableSimplifyLibCalls(LLVMPassManagerBuilderRef PMB,
918 // NOTE: The simplify-libcalls pass has been removed.
922 LLVMPassManagerBuilderUseInlinerWithThreshold(LLVMPassManagerBuilderRef PMB,
923 unsigned Threshold) {
924 PassManagerBuilder *Builder = unwrap(PMB);
925 Builder->Inliner = createFunctionInliningPass(Threshold);
929 LLVMPassManagerBuilderPopulateFunctionPassManager(LLVMPassManagerBuilderRef PMB,
930 LLVMPassManagerRef PM) {
931 PassManagerBuilder *Builder = unwrap(PMB);
932 legacy::FunctionPassManager *FPM = unwrap<legacy::FunctionPassManager>(PM);
933 Builder->populateFunctionPassManager(*FPM);
937 LLVMPassManagerBuilderPopulateModulePassManager(LLVMPassManagerBuilderRef PMB,
938 LLVMPassManagerRef PM) {
939 PassManagerBuilder *Builder = unwrap(PMB);
940 legacy::PassManagerBase *MPM = unwrap(PM);
941 Builder->populateModulePassManager(*MPM);
944 void LLVMPassManagerBuilderPopulateLTOPassManager(LLVMPassManagerBuilderRef PMB,
945 LLVMPassManagerRef PM,
946 LLVMBool Internalize,
947 LLVMBool RunInliner) {
948 PassManagerBuilder *Builder = unwrap(PMB);
949 legacy::PassManagerBase *LPM = unwrap(PM);
951 // A small backwards compatibility hack. populateLTOPassManager used to take
952 // an RunInliner option.
953 if (RunInliner && !Builder->Inliner)
954 Builder->Inliner = createFunctionInliningPass();
956 Builder->populateLTOPassManager(*LPM);