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 PassManagerBuilder::PassManagerBuilder() {
83 LibraryInfo = nullptr;
85 DisableTailCalls = false;
86 DisableUnitAtATime = false;
87 DisableUnrollLoops = false;
88 BBVectorize = RunBBVectorization;
89 SLPVectorize = RunSLPVectorization;
90 LoopVectorize = RunLoopVectorization;
91 RerollLoops = RunLoopRerolling;
92 LoadCombine = RunLoadCombine;
93 DisableGVNLoadPRE = false;
97 MergeFunctions = false;
100 PassManagerBuilder::~PassManagerBuilder() {
105 /// Set of global extensions, automatically added as part of the standard set.
106 static ManagedStatic<SmallVector<std::pair<PassManagerBuilder::ExtensionPointTy,
107 PassManagerBuilder::ExtensionFn>, 8> > GlobalExtensions;
109 void PassManagerBuilder::addGlobalExtension(
110 PassManagerBuilder::ExtensionPointTy Ty,
111 PassManagerBuilder::ExtensionFn Fn) {
112 GlobalExtensions->push_back(std::make_pair(Ty, Fn));
115 void PassManagerBuilder::addExtension(ExtensionPointTy Ty, ExtensionFn Fn) {
116 Extensions.push_back(std::make_pair(Ty, Fn));
119 void PassManagerBuilder::addExtensionsToPM(ExtensionPointTy ETy,
120 legacy::PassManagerBase &PM) const {
121 for (unsigned i = 0, e = GlobalExtensions->size(); i != e; ++i)
122 if ((*GlobalExtensions)[i].first == ETy)
123 (*GlobalExtensions)[i].second(*this, PM);
124 for (unsigned i = 0, e = Extensions.size(); i != e; ++i)
125 if (Extensions[i].first == ETy)
126 Extensions[i].second(*this, PM);
129 void PassManagerBuilder::addInitialAliasAnalysisPasses(
130 legacy::PassManagerBase &PM) const {
131 // Add TypeBasedAliasAnalysis before BasicAliasAnalysis so that
132 // BasicAliasAnalysis wins if they disagree. This is intended to help
133 // support "obvious" type-punning idioms.
135 PM.add(createCFLAliasAnalysisPass());
136 PM.add(createTypeBasedAliasAnalysisPass());
137 PM.add(createScopedNoAliasAAPass());
138 PM.add(createBasicAliasAnalysisPass());
141 void PassManagerBuilder::populateFunctionPassManager(
142 legacy::FunctionPassManager &FPM) {
143 addExtensionsToPM(EP_EarlyAsPossible, FPM);
145 // Add LibraryInfo if we have some.
147 FPM.add(new TargetLibraryInfoWrapperPass(*LibraryInfo));
149 if (OptLevel == 0) return;
151 addInitialAliasAnalysisPasses(FPM);
153 FPM.add(createCFGSimplificationPass());
155 FPM.add(createSROAPass());
157 FPM.add(createScalarReplAggregatesPass());
158 FPM.add(createEarlyCSEPass());
159 FPM.add(createLowerExpectIntrinsicPass());
162 void PassManagerBuilder::populateModulePassManager(
163 legacy::PassManagerBase &MPM) {
164 // If all optimizations are disabled, just run the always-inline pass and,
165 // if enabled, the function merging pass.
172 // FIXME: The BarrierNoopPass is a HACK! The inliner pass above implicitly
173 // creates a CGSCC pass manager, but we don't want to add extensions into
174 // that pass manager. To prevent this we insert a no-op module pass to reset
175 // the pass manager to get the same behavior as EP_OptimizerLast in non-O0
176 // builds. The function merging pass is
178 MPM.add(createMergeFunctionsPass());
179 else if (!GlobalExtensions->empty() || !Extensions.empty())
180 MPM.add(createBarrierNoopPass());
182 addExtensionsToPM(EP_EnabledOnOptLevel0, MPM);
186 // Add LibraryInfo if we have some.
188 MPM.add(new TargetLibraryInfoWrapperPass(*LibraryInfo));
190 addInitialAliasAnalysisPasses(MPM);
192 if (!DisableUnitAtATime) {
193 addExtensionsToPM(EP_ModuleOptimizerEarly, MPM);
195 MPM.add(createIPSCCPPass()); // IP SCCP
196 MPM.add(createGlobalOptimizerPass()); // Optimize out global vars
198 MPM.add(createDeadArgEliminationPass()); // Dead argument elimination
200 MPM.add(createInstructionCombiningPass());// Clean up after IPCP & DAE
201 addExtensionsToPM(EP_Peephole, MPM);
202 MPM.add(createCFGSimplificationPass()); // Clean up after IPCP & DAE
205 // Start of CallGraph SCC passes.
206 if (!DisableUnitAtATime)
207 MPM.add(createPruneEHPass()); // Remove dead EH info
212 if (!DisableUnitAtATime)
213 MPM.add(createFunctionAttrsPass()); // Set readonly/readnone attrs
215 MPM.add(createArgumentPromotionPass()); // Scalarize uninlined fn args
217 // Start of function pass.
218 // Break up aggregate allocas, using SSAUpdater.
220 MPM.add(createSROAPass(/*RequiresDomTree*/ false));
222 MPM.add(createScalarReplAggregatesPass(-1, false));
223 MPM.add(createEarlyCSEPass()); // Catch trivial redundancies
224 MPM.add(createJumpThreadingPass()); // Thread jumps.
225 MPM.add(createCorrelatedValuePropagationPass()); // Propagate conditionals
226 MPM.add(createCFGSimplificationPass()); // Merge & remove BBs
227 MPM.add(createInstructionCombiningPass()); // Combine silly seq's
228 addExtensionsToPM(EP_Peephole, MPM);
230 if (!DisableTailCalls)
231 MPM.add(createTailCallEliminationPass()); // Eliminate tail calls
232 MPM.add(createCFGSimplificationPass()); // Merge & remove BBs
233 MPM.add(createReassociatePass()); // Reassociate expressions
234 // Rotate Loop - disable header duplication at -Oz
235 MPM.add(createLoopRotatePass(SizeLevel == 2 ? 0 : -1));
236 MPM.add(createLICMPass()); // Hoist loop invariants
237 MPM.add(createLoopUnswitchPass(SizeLevel || OptLevel < 3));
238 MPM.add(createInstructionCombiningPass());
239 MPM.add(createIndVarSimplifyPass()); // Canonicalize indvars
240 MPM.add(createLoopIdiomPass()); // Recognize idioms like memset.
241 MPM.add(createLoopDeletionPass()); // Delete dead loops
243 if (!DisableUnrollLoops)
244 MPM.add(createSimpleLoopUnrollPass()); // Unroll small loops
245 addExtensionsToPM(EP_LoopOptimizerEnd, MPM);
249 MPM.add(createMergedLoadStoreMotionPass()); // Merge ld/st in diamonds
250 MPM.add(createGVNPass(DisableGVNLoadPRE)); // Remove redundancies
252 MPM.add(createMemCpyOptPass()); // Remove memcpy / form memset
253 MPM.add(createSCCPPass()); // Constant prop with SCCP
255 // Run instcombine after redundancy elimination to exploit opportunities
256 // opened up by them.
257 MPM.add(createInstructionCombiningPass());
258 addExtensionsToPM(EP_Peephole, MPM);
259 MPM.add(createJumpThreadingPass()); // Thread jumps
260 MPM.add(createCorrelatedValuePropagationPass());
261 MPM.add(createDeadStoreEliminationPass()); // Delete dead stores
263 addExtensionsToPM(EP_ScalarOptimizerLate, MPM);
266 MPM.add(createLoopRerollPass());
267 if (!RunSLPAfterLoopVectorization) {
269 MPM.add(createSLPVectorizerPass()); // Vectorize parallel scalar chains.
272 MPM.add(createBBVectorizePass());
273 MPM.add(createInstructionCombiningPass());
274 addExtensionsToPM(EP_Peephole, MPM);
275 if (OptLevel > 1 && UseGVNAfterVectorization)
276 MPM.add(createGVNPass(DisableGVNLoadPRE)); // Remove redundancies
278 MPM.add(createEarlyCSEPass()); // Catch trivial redundancies
280 // BBVectorize may have significantly shortened a loop body; unroll again.
281 if (!DisableUnrollLoops)
282 MPM.add(createLoopUnrollPass());
287 MPM.add(createLoadCombinePass());
289 MPM.add(createAggressiveDCEPass()); // Delete dead instructions
290 MPM.add(createCFGSimplificationPass()); // Merge & remove BBs
291 MPM.add(createInstructionCombiningPass()); // Clean up after everything.
292 addExtensionsToPM(EP_Peephole, MPM);
294 // FIXME: This is a HACK! The inliner pass above implicitly creates a CGSCC
295 // pass manager that we are specifically trying to avoid. To prevent this
296 // we must insert a no-op module pass to reset the pass manager.
297 MPM.add(createBarrierNoopPass());
299 // Re-rotate loops in all our loop nests. These may have fallout out of
300 // rotated form due to GVN or other transformations, and the vectorizer relies
301 // on the rotated form.
302 if (ExtraVectorizerPasses)
303 MPM.add(createLoopRotatePass());
305 MPM.add(createLoopVectorizePass(DisableUnrollLoops, LoopVectorize));
306 // FIXME: Because of #pragma vectorize enable, the passes below are always
307 // inserted in the pipeline, even when the vectorizer doesn't run (ex. when
308 // on -O1 and no #pragma is found). Would be good to have these two passes
309 // as function calls, so that we can only pass them when the vectorizer
311 MPM.add(createInstructionCombiningPass());
312 if (OptLevel > 1 && ExtraVectorizerPasses) {
313 // At higher optimization levels, try to clean up any runtime overlap and
314 // alignment checks inserted by the vectorizer. We want to track correllated
315 // runtime checks for two inner loops in the same outer loop, fold any
316 // common computations, hoist loop-invariant aspects out of any outer loop,
317 // and unswitch the runtime checks if possible. Once hoisted, we may have
318 // dead (or speculatable) control flows or more combining opportunities.
319 MPM.add(createEarlyCSEPass());
320 MPM.add(createCorrelatedValuePropagationPass());
321 MPM.add(createInstructionCombiningPass());
322 MPM.add(createLICMPass());
323 MPM.add(createLoopUnswitchPass(SizeLevel || OptLevel < 3));
324 MPM.add(createCFGSimplificationPass());
325 MPM.add(createInstructionCombiningPass());
328 if (RunSLPAfterLoopVectorization) {
330 MPM.add(createSLPVectorizerPass()); // Vectorize parallel scalar chains.
331 if (OptLevel > 1 && ExtraVectorizerPasses) {
332 MPM.add(createEarlyCSEPass());
337 MPM.add(createBBVectorizePass());
338 MPM.add(createInstructionCombiningPass());
339 addExtensionsToPM(EP_Peephole, MPM);
340 if (OptLevel > 1 && UseGVNAfterVectorization)
341 MPM.add(createGVNPass(DisableGVNLoadPRE)); // Remove redundancies
343 MPM.add(createEarlyCSEPass()); // Catch trivial redundancies
345 // BBVectorize may have significantly shortened a loop body; unroll again.
346 if (!DisableUnrollLoops)
347 MPM.add(createLoopUnrollPass());
351 addExtensionsToPM(EP_Peephole, MPM);
352 MPM.add(createCFGSimplificationPass());
353 MPM.add(createInstructionCombiningPass());
355 if (!DisableUnrollLoops)
356 MPM.add(createLoopUnrollPass()); // Unroll small loops
358 // After vectorization and unrolling, assume intrinsics may tell us more
359 // about pointer alignments.
360 MPM.add(createAlignmentFromAssumptionsPass());
362 if (!DisableUnitAtATime) {
363 // FIXME: We shouldn't bother with this anymore.
364 MPM.add(createStripDeadPrototypesPass()); // Get rid of dead prototypes
366 // GlobalOpt already deletes dead functions and globals, at -O2 try a
367 // late pass of GlobalDCE. It is capable of deleting dead cycles.
369 MPM.add(createGlobalDCEPass()); // Remove dead fns and globals.
370 MPM.add(createConstantMergePass()); // Merge dup global constants
375 MPM.add(createMergeFunctionsPass());
377 addExtensionsToPM(EP_OptimizerLast, MPM);
380 void PassManagerBuilder::addLTOOptimizationPasses(legacy::PassManagerBase &PM) {
381 // Provide AliasAnalysis services for optimizations.
382 addInitialAliasAnalysisPasses(PM);
384 // Propagate constants at call sites into the functions they call. This
385 // opens opportunities for globalopt (and inlining) by substituting function
386 // pointers passed as arguments to direct uses of functions.
387 PM.add(createIPSCCPPass());
389 // Now that we internalized some globals, see if we can hack on them!
390 PM.add(createGlobalOptimizerPass());
392 // Linking modules together can lead to duplicated global constants, only
393 // keep one copy of each constant.
394 PM.add(createConstantMergePass());
396 // Remove unused arguments from functions.
397 PM.add(createDeadArgEliminationPass());
399 // Reduce the code after globalopt and ipsccp. Both can open up significant
400 // simplification opportunities, and both can propagate functions through
401 // function pointers. When this happens, we often have to resolve varargs
402 // calls, etc, so let instcombine do this.
403 PM.add(createInstructionCombiningPass());
404 addExtensionsToPM(EP_Peephole, PM);
406 // Inline small functions
407 bool RunInliner = Inliner;
413 PM.add(createPruneEHPass()); // Remove dead EH info.
415 // Optimize globals again if we ran the inliner.
417 PM.add(createGlobalOptimizerPass());
418 PM.add(createGlobalDCEPass()); // Remove dead functions.
420 // If we didn't decide to inline a function, check to see if we can
421 // transform it to pass arguments by value instead of by reference.
422 PM.add(createArgumentPromotionPass());
424 // The IPO passes may leave cruft around. Clean up after them.
425 PM.add(createInstructionCombiningPass());
426 addExtensionsToPM(EP_Peephole, PM);
427 PM.add(createJumpThreadingPass());
431 PM.add(createSROAPass());
433 PM.add(createScalarReplAggregatesPass());
435 // Run a few AA driven optimizations here and now, to cleanup the code.
436 PM.add(createFunctionAttrsPass()); // Add nocapture.
437 PM.add(createGlobalsModRefPass()); // IP alias analysis.
439 PM.add(createLICMPass()); // Hoist loop invariants.
441 PM.add(createMergedLoadStoreMotionPass()); // Merge ld/st in diamonds.
442 PM.add(createGVNPass(DisableGVNLoadPRE)); // Remove redundancies.
443 PM.add(createMemCpyOptPass()); // Remove dead memcpys.
446 PM.add(createDeadStoreEliminationPass());
448 // More loops are countable; try to optimize them.
449 PM.add(createIndVarSimplifyPass());
450 PM.add(createLoopDeletionPass());
451 PM.add(createLoopVectorizePass(true, LoopVectorize));
453 // More scalar chains could be vectorized due to more alias information
454 if (RunSLPAfterLoopVectorization)
456 PM.add(createSLPVectorizerPass()); // Vectorize parallel scalar chains.
458 // After vectorization, assume intrinsics may tell us more about pointer
460 PM.add(createAlignmentFromAssumptionsPass());
463 PM.add(createLoadCombinePass());
465 // Cleanup and simplify the code after the scalar optimizations.
466 PM.add(createInstructionCombiningPass());
467 addExtensionsToPM(EP_Peephole, PM);
469 PM.add(createJumpThreadingPass());
471 // Delete basic blocks, which optimization passes may have killed.
472 PM.add(createCFGSimplificationPass());
474 // Now that we have optimized the program, discard unreachable functions.
475 PM.add(createGlobalDCEPass());
477 // FIXME: this is profitable (for compiler time) to do at -O0 too, but
478 // currently it damages debug info.
480 PM.add(createMergeFunctionsPass());
483 void PassManagerBuilder::populateLTOPassManager(legacy::PassManagerBase &PM) {
485 PM.add(new TargetLibraryInfoWrapperPass(*LibraryInfo));
488 PM.add(createVerifierPass());
491 PM.add(createStripSymbolsPass(true));
494 PM.add(createDebugInfoVerifierPass());
497 addLTOOptimizationPasses(PM);
500 PM.add(createVerifierPass());
501 PM.add(createDebugInfoVerifierPass());
505 inline PassManagerBuilder *unwrap(LLVMPassManagerBuilderRef P) {
506 return reinterpret_cast<PassManagerBuilder*>(P);
509 inline LLVMPassManagerBuilderRef wrap(PassManagerBuilder *P) {
510 return reinterpret_cast<LLVMPassManagerBuilderRef>(P);
513 LLVMPassManagerBuilderRef LLVMPassManagerBuilderCreate() {
514 PassManagerBuilder *PMB = new PassManagerBuilder();
518 void LLVMPassManagerBuilderDispose(LLVMPassManagerBuilderRef PMB) {
519 PassManagerBuilder *Builder = unwrap(PMB);
524 LLVMPassManagerBuilderSetOptLevel(LLVMPassManagerBuilderRef PMB,
526 PassManagerBuilder *Builder = unwrap(PMB);
527 Builder->OptLevel = OptLevel;
531 LLVMPassManagerBuilderSetSizeLevel(LLVMPassManagerBuilderRef PMB,
532 unsigned SizeLevel) {
533 PassManagerBuilder *Builder = unwrap(PMB);
534 Builder->SizeLevel = SizeLevel;
538 LLVMPassManagerBuilderSetDisableUnitAtATime(LLVMPassManagerBuilderRef PMB,
540 PassManagerBuilder *Builder = unwrap(PMB);
541 Builder->DisableUnitAtATime = Value;
545 LLVMPassManagerBuilderSetDisableUnrollLoops(LLVMPassManagerBuilderRef PMB,
547 PassManagerBuilder *Builder = unwrap(PMB);
548 Builder->DisableUnrollLoops = Value;
552 LLVMPassManagerBuilderSetDisableSimplifyLibCalls(LLVMPassManagerBuilderRef PMB,
554 // NOTE: The simplify-libcalls pass has been removed.
558 LLVMPassManagerBuilderUseInlinerWithThreshold(LLVMPassManagerBuilderRef PMB,
559 unsigned Threshold) {
560 PassManagerBuilder *Builder = unwrap(PMB);
561 Builder->Inliner = createFunctionInliningPass(Threshold);
565 LLVMPassManagerBuilderPopulateFunctionPassManager(LLVMPassManagerBuilderRef PMB,
566 LLVMPassManagerRef PM) {
567 PassManagerBuilder *Builder = unwrap(PMB);
568 legacy::FunctionPassManager *FPM = unwrap<legacy::FunctionPassManager>(PM);
569 Builder->populateFunctionPassManager(*FPM);
573 LLVMPassManagerBuilderPopulateModulePassManager(LLVMPassManagerBuilderRef PMB,
574 LLVMPassManagerRef PM) {
575 PassManagerBuilder *Builder = unwrap(PMB);
576 legacy::PassManagerBase *MPM = unwrap(PM);
577 Builder->populateModulePassManager(*MPM);
580 void LLVMPassManagerBuilderPopulateLTOPassManager(LLVMPassManagerBuilderRef PMB,
581 LLVMPassManagerRef PM,
582 LLVMBool Internalize,
583 LLVMBool RunInliner) {
584 PassManagerBuilder *Builder = unwrap(PMB);
585 legacy::PassManagerBase *LPM = unwrap(PM);
587 // A small backwards compatibility hack. populateLTOPassManager used to take
588 // an RunInliner option.
589 if (RunInliner && !Builder->Inliner)
590 Builder->Inliner = createFunctionInliningPass();
592 Builder->populateLTOPassManager(*LPM);