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 //===----------------------------------------------------------------------===//
16 #include "llvm/Transforms/IPO/PassManagerBuilder.h"
17 #include "llvm-c/Transforms/PassManagerBuilder.h"
18 #include "llvm/ADT/SmallVector.h"
19 #include "llvm/Analysis/Passes.h"
20 #include "llvm/IR/DataLayout.h"
21 #include "llvm/IR/Verifier.h"
22 #include "llvm/IR/LegacyPassManager.h"
23 #include "llvm/Support/CommandLine.h"
24 #include "llvm/Support/ManagedStatic.h"
25 #include "llvm/Analysis/TargetLibraryInfo.h"
26 #include "llvm/Target/TargetMachine.h"
27 #include "llvm/Transforms/IPO.h"
28 #include "llvm/Transforms/Scalar.h"
29 #include "llvm/Transforms/Vectorize.h"
34 RunLoopVectorization("vectorize-loops", cl::Hidden,
35 cl::desc("Run the Loop vectorization passes"));
38 RunSLPVectorization("vectorize-slp", cl::Hidden,
39 cl::desc("Run the SLP vectorization passes"));
42 RunBBVectorization("vectorize-slp-aggressive", cl::Hidden,
43 cl::desc("Run the BB vectorization passes"));
46 UseGVNAfterVectorization("use-gvn-after-vectorization",
47 cl::init(false), cl::Hidden,
48 cl::desc("Run GVN instead of Early CSE after vectorization passes"));
50 static cl::opt<bool> ExtraVectorizerPasses(
51 "extra-vectorizer-passes", cl::init(false), cl::Hidden,
52 cl::desc("Run cleanup optimization passes after vectorization."));
54 static cl::opt<bool> UseNewSROA("use-new-sroa",
55 cl::init(true), cl::Hidden,
56 cl::desc("Enable the new, experimental SROA pass"));
59 RunLoopRerolling("reroll-loops", cl::Hidden,
60 cl::desc("Run the loop rerolling pass"));
62 static cl::opt<bool> RunLoadCombine("combine-loads", cl::init(false),
64 cl::desc("Run the load combining pass"));
67 RunSLPAfterLoopVectorization("run-slp-after-loop-vectorization",
68 cl::init(true), cl::Hidden,
69 cl::desc("Run the SLP vectorizer (and BB vectorizer) after the Loop "
70 "vectorizer instead of before"));
72 static cl::opt<bool> UseCFLAA("use-cfl-aa",
73 cl::init(false), cl::Hidden,
74 cl::desc("Enable the new, experimental CFL alias analysis"));
77 EnableMLSM("mlsm", cl::init(true), cl::Hidden,
78 cl::desc("Enable motion of merged load and store"));
80 static cl::opt<bool> EnableLoopInterchange(
81 "enable-loopinterchange", cl::init(false), cl::Hidden,
82 cl::desc("Enable the new, experimental LoopInterchange Pass"));
84 PassManagerBuilder::PassManagerBuilder() {
87 LibraryInfo = nullptr;
89 DisableTailCalls = false;
90 DisableUnitAtATime = false;
91 DisableUnrollLoops = false;
92 BBVectorize = RunBBVectorization;
93 SLPVectorize = RunSLPVectorization;
94 LoopVectorize = RunLoopVectorization;
95 RerollLoops = RunLoopRerolling;
96 LoadCombine = RunLoadCombine;
97 DisableGVNLoadPRE = false;
101 MergeFunctions = false;
104 PassManagerBuilder::~PassManagerBuilder() {
109 /// Set of global extensions, automatically added as part of the standard set.
110 static ManagedStatic<SmallVector<std::pair<PassManagerBuilder::ExtensionPointTy,
111 PassManagerBuilder::ExtensionFn>, 8> > GlobalExtensions;
113 void PassManagerBuilder::addGlobalExtension(
114 PassManagerBuilder::ExtensionPointTy Ty,
115 PassManagerBuilder::ExtensionFn Fn) {
116 GlobalExtensions->push_back(std::make_pair(Ty, Fn));
119 void PassManagerBuilder::addExtension(ExtensionPointTy Ty, ExtensionFn Fn) {
120 Extensions.push_back(std::make_pair(Ty, Fn));
123 void PassManagerBuilder::addExtensionsToPM(ExtensionPointTy ETy,
124 legacy::PassManagerBase &PM) const {
125 for (unsigned i = 0, e = GlobalExtensions->size(); i != e; ++i)
126 if ((*GlobalExtensions)[i].first == ETy)
127 (*GlobalExtensions)[i].second(*this, PM);
128 for (unsigned i = 0, e = Extensions.size(); i != e; ++i)
129 if (Extensions[i].first == ETy)
130 Extensions[i].second(*this, PM);
133 void PassManagerBuilder::addInitialAliasAnalysisPasses(
134 legacy::PassManagerBase &PM) const {
135 // Add TypeBasedAliasAnalysis before BasicAliasAnalysis so that
136 // BasicAliasAnalysis wins if they disagree. This is intended to help
137 // support "obvious" type-punning idioms.
139 PM.add(createCFLAliasAnalysisPass());
140 PM.add(createTypeBasedAliasAnalysisPass());
141 PM.add(createScopedNoAliasAAPass());
142 PM.add(createBasicAliasAnalysisPass());
145 void PassManagerBuilder::populateFunctionPassManager(
146 legacy::FunctionPassManager &FPM) {
147 addExtensionsToPM(EP_EarlyAsPossible, FPM);
149 // Add LibraryInfo if we have some.
151 FPM.add(new TargetLibraryInfoWrapperPass(*LibraryInfo));
153 if (OptLevel == 0) return;
155 addInitialAliasAnalysisPasses(FPM);
157 FPM.add(createCFGSimplificationPass());
159 FPM.add(createSROAPass());
161 FPM.add(createScalarReplAggregatesPass());
162 FPM.add(createEarlyCSEPass());
163 FPM.add(createLowerExpectIntrinsicPass());
166 void PassManagerBuilder::populateModulePassManager(
167 legacy::PassManagerBase &MPM) {
168 // If all optimizations are disabled, just run the always-inline pass and,
169 // if enabled, the function merging pass.
176 // FIXME: The BarrierNoopPass is a HACK! The inliner pass above implicitly
177 // creates a CGSCC pass manager, but we don't want to add extensions into
178 // that pass manager. To prevent this we insert a no-op module pass to reset
179 // the pass manager to get the same behavior as EP_OptimizerLast in non-O0
180 // builds. The function merging pass is
182 MPM.add(createMergeFunctionsPass());
183 else if (!GlobalExtensions->empty() || !Extensions.empty())
184 MPM.add(createBarrierNoopPass());
186 addExtensionsToPM(EP_EnabledOnOptLevel0, MPM);
190 // Add LibraryInfo if we have some.
192 MPM.add(new TargetLibraryInfoWrapperPass(*LibraryInfo));
194 addInitialAliasAnalysisPasses(MPM);
196 if (!DisableUnitAtATime) {
197 addExtensionsToPM(EP_ModuleOptimizerEarly, MPM);
199 MPM.add(createIPSCCPPass()); // IP SCCP
200 MPM.add(createGlobalOptimizerPass()); // Optimize out global vars
202 MPM.add(createDeadArgEliminationPass()); // Dead argument elimination
204 MPM.add(createInstructionCombiningPass());// Clean up after IPCP & DAE
205 addExtensionsToPM(EP_Peephole, MPM);
206 MPM.add(createCFGSimplificationPass()); // Clean up after IPCP & DAE
209 // Start of CallGraph SCC passes.
210 if (!DisableUnitAtATime)
211 MPM.add(createPruneEHPass()); // Remove dead EH info
216 if (!DisableUnitAtATime)
217 MPM.add(createFunctionAttrsPass()); // Set readonly/readnone attrs
219 MPM.add(createArgumentPromotionPass()); // Scalarize uninlined fn args
221 // Start of function pass.
222 // Break up aggregate allocas, using SSAUpdater.
224 MPM.add(createSROAPass(/*RequiresDomTree*/ false));
226 MPM.add(createScalarReplAggregatesPass(-1, false));
227 MPM.add(createEarlyCSEPass()); // Catch trivial redundancies
228 MPM.add(createJumpThreadingPass()); // Thread jumps.
229 MPM.add(createCorrelatedValuePropagationPass()); // Propagate conditionals
230 MPM.add(createCFGSimplificationPass()); // Merge & remove BBs
231 MPM.add(createInstructionCombiningPass()); // Combine silly seq's
232 addExtensionsToPM(EP_Peephole, MPM);
234 if (!DisableTailCalls)
235 MPM.add(createTailCallEliminationPass()); // Eliminate tail calls
236 MPM.add(createCFGSimplificationPass()); // Merge & remove BBs
237 MPM.add(createReassociatePass()); // Reassociate expressions
238 // Rotate Loop - disable header duplication at -Oz
239 MPM.add(createLoopRotatePass(SizeLevel == 2 ? 0 : -1));
240 MPM.add(createLICMPass()); // Hoist loop invariants
241 MPM.add(createLoopUnswitchPass(SizeLevel || OptLevel < 3));
242 MPM.add(createInstructionCombiningPass());
243 MPM.add(createIndVarSimplifyPass()); // Canonicalize indvars
244 MPM.add(createLoopIdiomPass()); // Recognize idioms like memset.
245 MPM.add(createLoopDeletionPass()); // Delete dead loops
246 if (EnableLoopInterchange)
247 MPM.add(createLoopInterchangePass()); // Interchange loops
249 if (!DisableUnrollLoops)
250 MPM.add(createSimpleLoopUnrollPass()); // Unroll small loops
251 addExtensionsToPM(EP_LoopOptimizerEnd, MPM);
255 MPM.add(createMergedLoadStoreMotionPass()); // Merge ld/st in diamonds
256 MPM.add(createGVNPass(DisableGVNLoadPRE)); // Remove redundancies
258 MPM.add(createMemCpyOptPass()); // Remove memcpy / form memset
259 MPM.add(createSCCPPass()); // Constant prop with SCCP
261 // Delete dead bit computations (instcombine runs after to fold away the dead
262 // computations, and then ADCE will run later to exploit any new DCE
263 // opportunities that creates).
264 MPM.add(createBitTrackingDCEPass()); // Delete dead bit computations
266 // Run instcombine after redundancy elimination to exploit opportunities
267 // opened up by them.
268 MPM.add(createInstructionCombiningPass());
269 addExtensionsToPM(EP_Peephole, MPM);
270 MPM.add(createJumpThreadingPass()); // Thread jumps
271 MPM.add(createCorrelatedValuePropagationPass());
272 MPM.add(createDeadStoreEliminationPass()); // Delete dead stores
273 MPM.add(createLICMPass());
275 addExtensionsToPM(EP_ScalarOptimizerLate, MPM);
278 MPM.add(createLoopRerollPass());
279 if (!RunSLPAfterLoopVectorization) {
281 MPM.add(createSLPVectorizerPass()); // Vectorize parallel scalar chains.
284 MPM.add(createBBVectorizePass());
285 MPM.add(createInstructionCombiningPass());
286 addExtensionsToPM(EP_Peephole, MPM);
287 if (OptLevel > 1 && UseGVNAfterVectorization)
288 MPM.add(createGVNPass(DisableGVNLoadPRE)); // Remove redundancies
290 MPM.add(createEarlyCSEPass()); // Catch trivial redundancies
292 // BBVectorize may have significantly shortened a loop body; unroll again.
293 if (!DisableUnrollLoops)
294 MPM.add(createLoopUnrollPass());
299 MPM.add(createLoadCombinePass());
301 MPM.add(createAggressiveDCEPass()); // Delete dead instructions
302 MPM.add(createCFGSimplificationPass()); // Merge & remove BBs
303 MPM.add(createInstructionCombiningPass()); // Clean up after everything.
304 addExtensionsToPM(EP_Peephole, MPM);
306 // FIXME: This is a HACK! The inliner pass above implicitly creates a CGSCC
307 // pass manager that we are specifically trying to avoid. To prevent this
308 // we must insert a no-op module pass to reset the pass manager.
309 MPM.add(createBarrierNoopPass());
311 // Re-rotate loops in all our loop nests. These may have fallout out of
312 // rotated form due to GVN or other transformations, and the vectorizer relies
313 // on the rotated form.
314 if (ExtraVectorizerPasses)
315 MPM.add(createLoopRotatePass());
317 MPM.add(createLoopVectorizePass(DisableUnrollLoops, LoopVectorize));
318 // FIXME: Because of #pragma vectorize enable, the passes below are always
319 // inserted in the pipeline, even when the vectorizer doesn't run (ex. when
320 // on -O1 and no #pragma is found). Would be good to have these two passes
321 // as function calls, so that we can only pass them when the vectorizer
323 MPM.add(createInstructionCombiningPass());
324 if (OptLevel > 1 && ExtraVectorizerPasses) {
325 // At higher optimization levels, try to clean up any runtime overlap and
326 // alignment checks inserted by the vectorizer. We want to track correllated
327 // runtime checks for two inner loops in the same outer loop, fold any
328 // common computations, hoist loop-invariant aspects out of any outer loop,
329 // and unswitch the runtime checks if possible. Once hoisted, we may have
330 // dead (or speculatable) control flows or more combining opportunities.
331 MPM.add(createEarlyCSEPass());
332 MPM.add(createCorrelatedValuePropagationPass());
333 MPM.add(createInstructionCombiningPass());
334 MPM.add(createLICMPass());
335 MPM.add(createLoopUnswitchPass(SizeLevel || OptLevel < 3));
336 MPM.add(createCFGSimplificationPass());
337 MPM.add(createInstructionCombiningPass());
340 if (RunSLPAfterLoopVectorization) {
342 MPM.add(createSLPVectorizerPass()); // Vectorize parallel scalar chains.
343 if (OptLevel > 1 && ExtraVectorizerPasses) {
344 MPM.add(createEarlyCSEPass());
349 MPM.add(createBBVectorizePass());
350 MPM.add(createInstructionCombiningPass());
351 addExtensionsToPM(EP_Peephole, MPM);
352 if (OptLevel > 1 && UseGVNAfterVectorization)
353 MPM.add(createGVNPass(DisableGVNLoadPRE)); // Remove redundancies
355 MPM.add(createEarlyCSEPass()); // Catch trivial redundancies
357 // BBVectorize may have significantly shortened a loop body; unroll again.
358 if (!DisableUnrollLoops)
359 MPM.add(createLoopUnrollPass());
363 addExtensionsToPM(EP_Peephole, MPM);
364 MPM.add(createCFGSimplificationPass());
365 MPM.add(createInstructionCombiningPass());
367 if (!DisableUnrollLoops)
368 MPM.add(createLoopUnrollPass()); // Unroll small loops
370 // After vectorization and unrolling, assume intrinsics may tell us more
371 // about pointer alignments.
372 MPM.add(createAlignmentFromAssumptionsPass());
374 if (!DisableUnitAtATime) {
375 // FIXME: We shouldn't bother with this anymore.
376 MPM.add(createStripDeadPrototypesPass()); // Get rid of dead prototypes
378 // GlobalOpt already deletes dead functions and globals, at -O2 try a
379 // late pass of GlobalDCE. It is capable of deleting dead cycles.
381 MPM.add(createGlobalDCEPass()); // Remove dead fns and globals.
382 MPM.add(createConstantMergePass()); // Merge dup global constants
387 MPM.add(createMergeFunctionsPass());
389 addExtensionsToPM(EP_OptimizerLast, MPM);
392 void PassManagerBuilder::addLTOOptimizationPasses(legacy::PassManagerBase &PM) {
393 // Provide AliasAnalysis services for optimizations.
394 addInitialAliasAnalysisPasses(PM);
396 // Propagate constants at call sites into the functions they call. This
397 // opens opportunities for globalopt (and inlining) by substituting function
398 // pointers passed as arguments to direct uses of functions.
399 PM.add(createIPSCCPPass());
401 // Now that we internalized some globals, see if we can hack on them!
402 PM.add(createGlobalOptimizerPass());
404 // Linking modules together can lead to duplicated global constants, only
405 // keep one copy of each constant.
406 PM.add(createConstantMergePass());
408 // Remove unused arguments from functions.
409 PM.add(createDeadArgEliminationPass());
411 // Reduce the code after globalopt and ipsccp. Both can open up significant
412 // simplification opportunities, and both can propagate functions through
413 // function pointers. When this happens, we often have to resolve varargs
414 // calls, etc, so let instcombine do this.
415 PM.add(createInstructionCombiningPass());
416 addExtensionsToPM(EP_Peephole, PM);
418 // Inline small functions
419 bool RunInliner = Inliner;
425 PM.add(createPruneEHPass()); // Remove dead EH info.
427 // Optimize globals again if we ran the inliner.
429 PM.add(createGlobalOptimizerPass());
430 PM.add(createGlobalDCEPass()); // Remove dead functions.
432 // If we didn't decide to inline a function, check to see if we can
433 // transform it to pass arguments by value instead of by reference.
434 PM.add(createArgumentPromotionPass());
436 // The IPO passes may leave cruft around. Clean up after them.
437 PM.add(createInstructionCombiningPass());
438 addExtensionsToPM(EP_Peephole, PM);
439 PM.add(createJumpThreadingPass());
443 PM.add(createSROAPass());
445 PM.add(createScalarReplAggregatesPass());
447 // Run a few AA driven optimizations here and now, to cleanup the code.
448 PM.add(createFunctionAttrsPass()); // Add nocapture.
449 PM.add(createGlobalsModRefPass()); // IP alias analysis.
451 PM.add(createLICMPass()); // Hoist loop invariants.
453 PM.add(createMergedLoadStoreMotionPass()); // Merge ld/st in diamonds.
454 PM.add(createGVNPass(DisableGVNLoadPRE)); // Remove redundancies.
455 PM.add(createMemCpyOptPass()); // Remove dead memcpys.
458 PM.add(createDeadStoreEliminationPass());
460 // More loops are countable; try to optimize them.
461 PM.add(createIndVarSimplifyPass());
462 PM.add(createLoopDeletionPass());
463 if (EnableLoopInterchange)
464 PM.add(createLoopInterchangePass());
466 PM.add(createLoopVectorizePass(true, LoopVectorize));
468 // More scalar chains could be vectorized due to more alias information
469 if (RunSLPAfterLoopVectorization)
471 PM.add(createSLPVectorizerPass()); // Vectorize parallel scalar chains.
473 // After vectorization, assume intrinsics may tell us more about pointer
475 PM.add(createAlignmentFromAssumptionsPass());
478 PM.add(createLoadCombinePass());
480 // Cleanup and simplify the code after the scalar optimizations.
481 PM.add(createInstructionCombiningPass());
482 addExtensionsToPM(EP_Peephole, PM);
484 PM.add(createJumpThreadingPass());
486 // Lower bitset metadata to bitsets.
487 PM.add(createLowerBitSetsPass());
489 // Delete basic blocks, which optimization passes may have killed.
490 PM.add(createCFGSimplificationPass());
492 // Now that we have optimized the program, discard unreachable functions.
493 PM.add(createGlobalDCEPass());
495 // FIXME: this is profitable (for compiler time) to do at -O0 too, but
496 // currently it damages debug info.
498 PM.add(createMergeFunctionsPass());
501 void PassManagerBuilder::populateLTOPassManager(legacy::PassManagerBase &PM) {
503 PM.add(new TargetLibraryInfoWrapperPass(*LibraryInfo));
506 PM.add(createVerifierPass());
509 PM.add(createStripSymbolsPass(true));
512 PM.add(createDebugInfoVerifierPass());
515 addLTOOptimizationPasses(PM);
518 PM.add(createVerifierPass());
519 PM.add(createDebugInfoVerifierPass());
523 inline PassManagerBuilder *unwrap(LLVMPassManagerBuilderRef P) {
524 return reinterpret_cast<PassManagerBuilder*>(P);
527 inline LLVMPassManagerBuilderRef wrap(PassManagerBuilder *P) {
528 return reinterpret_cast<LLVMPassManagerBuilderRef>(P);
531 LLVMPassManagerBuilderRef LLVMPassManagerBuilderCreate() {
532 PassManagerBuilder *PMB = new PassManagerBuilder();
536 void LLVMPassManagerBuilderDispose(LLVMPassManagerBuilderRef PMB) {
537 PassManagerBuilder *Builder = unwrap(PMB);
542 LLVMPassManagerBuilderSetOptLevel(LLVMPassManagerBuilderRef PMB,
544 PassManagerBuilder *Builder = unwrap(PMB);
545 Builder->OptLevel = OptLevel;
549 LLVMPassManagerBuilderSetSizeLevel(LLVMPassManagerBuilderRef PMB,
550 unsigned SizeLevel) {
551 PassManagerBuilder *Builder = unwrap(PMB);
552 Builder->SizeLevel = SizeLevel;
556 LLVMPassManagerBuilderSetDisableUnitAtATime(LLVMPassManagerBuilderRef PMB,
558 PassManagerBuilder *Builder = unwrap(PMB);
559 Builder->DisableUnitAtATime = Value;
563 LLVMPassManagerBuilderSetDisableUnrollLoops(LLVMPassManagerBuilderRef PMB,
565 PassManagerBuilder *Builder = unwrap(PMB);
566 Builder->DisableUnrollLoops = Value;
570 LLVMPassManagerBuilderSetDisableSimplifyLibCalls(LLVMPassManagerBuilderRef PMB,
572 // NOTE: The simplify-libcalls pass has been removed.
576 LLVMPassManagerBuilderUseInlinerWithThreshold(LLVMPassManagerBuilderRef PMB,
577 unsigned Threshold) {
578 PassManagerBuilder *Builder = unwrap(PMB);
579 Builder->Inliner = createFunctionInliningPass(Threshold);
583 LLVMPassManagerBuilderPopulateFunctionPassManager(LLVMPassManagerBuilderRef PMB,
584 LLVMPassManagerRef PM) {
585 PassManagerBuilder *Builder = unwrap(PMB);
586 legacy::FunctionPassManager *FPM = unwrap<legacy::FunctionPassManager>(PM);
587 Builder->populateFunctionPassManager(*FPM);
591 LLVMPassManagerBuilderPopulateModulePassManager(LLVMPassManagerBuilderRef PMB,
592 LLVMPassManagerRef PM) {
593 PassManagerBuilder *Builder = unwrap(PMB);
594 legacy::PassManagerBase *MPM = unwrap(PM);
595 Builder->populateModulePassManager(*MPM);
598 void LLVMPassManagerBuilderPopulateLTOPassManager(LLVMPassManagerBuilderRef PMB,
599 LLVMPassManagerRef PM,
600 LLVMBool Internalize,
601 LLVMBool RunInliner) {
602 PassManagerBuilder *Builder = unwrap(PMB);
603 legacy::PassManagerBase *LPM = unwrap(PM);
605 // A small backwards compatibility hack. populateLTOPassManager used to take
606 // an RunInliner option.
607 if (RunInliner && !Builder->Inliner)
608 Builder->Inliner = createFunctionInliningPass();
610 Builder->populateLTOPassManager(*LPM);