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/PassManager.h"
23 #include "llvm/Support/CommandLine.h"
24 #include "llvm/Support/ManagedStatic.h"
25 #include "llvm/Target/TargetLibraryInfo.h"
26 #include "llvm/Target/TargetMachine.h"
27 #include "llvm/Target/TargetSubtargetInfo.h"
28 #include "llvm/Transforms/IPO.h"
29 #include "llvm/Transforms/Scalar.h"
30 #include "llvm/Transforms/Vectorize.h"
35 RunLoopVectorization("vectorize-loops", cl::Hidden,
36 cl::desc("Run the Loop vectorization passes"));
39 RunSLPVectorization("vectorize-slp", cl::Hidden,
40 cl::desc("Run the SLP vectorization passes"));
43 RunBBVectorization("vectorize-slp-aggressive", cl::Hidden,
44 cl::desc("Run the BB vectorization passes"));
47 UseGVNAfterVectorization("use-gvn-after-vectorization",
48 cl::init(false), cl::Hidden,
49 cl::desc("Run GVN instead of Early CSE after vectorization passes"));
51 static cl::opt<bool> ExtraVectorizerPasses(
52 "extra-vectorizer-passes", cl::init(false), cl::Hidden,
53 cl::desc("Run cleanup optimization passes after vectorization."));
55 static cl::opt<bool> UseNewSROA("use-new-sroa",
56 cl::init(true), cl::Hidden,
57 cl::desc("Enable the new, experimental SROA pass"));
60 RunLoopRerolling("reroll-loops", cl::Hidden,
61 cl::desc("Run the loop rerolling pass"));
63 static cl::opt<bool> RunLoadCombine("combine-loads", cl::init(false),
65 cl::desc("Run the load combining pass"));
68 RunSLPAfterLoopVectorization("run-slp-after-loop-vectorization",
69 cl::init(true), cl::Hidden,
70 cl::desc("Run the SLP vectorizer (and BB vectorizer) after the Loop "
71 "vectorizer instead of before"));
73 static cl::opt<bool> UseCFLAA("use-cfl-aa",
74 cl::init(false), cl::Hidden,
75 cl::desc("Enable the new, experimental CFL alias analysis"));
78 EnableMLSM("mlsm", cl::init(true), cl::Hidden,
79 cl::desc("Enable motion of merged load and store"));
81 PassManagerBuilder::PassManagerBuilder() {
84 LibraryInfo = nullptr;
86 DisableTailCalls = false;
87 DisableUnitAtATime = false;
88 DisableUnrollLoops = false;
89 BBVectorize = RunBBVectorization;
90 SLPVectorize = RunSLPVectorization;
91 LoopVectorize = RunLoopVectorization;
92 RerollLoops = RunLoopRerolling;
93 LoadCombine = RunLoadCombine;
94 DisableGVNLoadPRE = false;
98 MergeFunctions = false;
101 PassManagerBuilder::~PassManagerBuilder() {
106 /// Set of global extensions, automatically added as part of the standard set.
107 static ManagedStatic<SmallVector<std::pair<PassManagerBuilder::ExtensionPointTy,
108 PassManagerBuilder::ExtensionFn>, 8> > GlobalExtensions;
110 void PassManagerBuilder::addGlobalExtension(
111 PassManagerBuilder::ExtensionPointTy Ty,
112 PassManagerBuilder::ExtensionFn Fn) {
113 GlobalExtensions->push_back(std::make_pair(Ty, Fn));
116 void PassManagerBuilder::addExtension(ExtensionPointTy Ty, ExtensionFn Fn) {
117 Extensions.push_back(std::make_pair(Ty, Fn));
120 void PassManagerBuilder::addExtensionsToPM(ExtensionPointTy ETy,
121 PassManagerBase &PM) const {
122 for (unsigned i = 0, e = GlobalExtensions->size(); i != e; ++i)
123 if ((*GlobalExtensions)[i].first == ETy)
124 (*GlobalExtensions)[i].second(*this, PM);
125 for (unsigned i = 0, e = Extensions.size(); i != e; ++i)
126 if (Extensions[i].first == ETy)
127 Extensions[i].second(*this, PM);
131 PassManagerBuilder::addInitialAliasAnalysisPasses(PassManagerBase &PM) const {
132 // Add TypeBasedAliasAnalysis before BasicAliasAnalysis so that
133 // BasicAliasAnalysis wins if they disagree. This is intended to help
134 // support "obvious" type-punning idioms.
136 PM.add(createCFLAliasAnalysisPass());
137 PM.add(createTypeBasedAliasAnalysisPass());
138 PM.add(createScopedNoAliasAAPass());
139 PM.add(createBasicAliasAnalysisPass());
142 void PassManagerBuilder::populateFunctionPassManager(FunctionPassManager &FPM) {
143 addExtensionsToPM(EP_EarlyAsPossible, FPM);
145 // Add LibraryInfo if we have some.
146 if (LibraryInfo) FPM.add(new TargetLibraryInfo(*LibraryInfo));
148 if (OptLevel == 0) return;
150 addInitialAliasAnalysisPasses(FPM);
152 FPM.add(createCFGSimplificationPass());
154 FPM.add(createSROAPass());
156 FPM.add(createScalarReplAggregatesPass());
157 FPM.add(createEarlyCSEPass());
158 FPM.add(createLowerExpectIntrinsicPass());
161 void PassManagerBuilder::populateModulePassManager(PassManagerBase &MPM) {
162 // If all optimizations are disabled, just run the always-inline pass and,
163 // if enabled, the function merging pass.
170 // FIXME: The BarrierNoopPass is a HACK! The inliner pass above implicitly
171 // creates a CGSCC pass manager, but we don't want to add extensions into
172 // that pass manager. To prevent this we insert a no-op module pass to reset
173 // the pass manager to get the same behavior as EP_OptimizerLast in non-O0
174 // builds. The function merging pass is
176 MPM.add(createMergeFunctionsPass());
177 else if (!GlobalExtensions->empty() || !Extensions.empty())
178 MPM.add(createBarrierNoopPass());
180 addExtensionsToPM(EP_EnabledOnOptLevel0, MPM);
184 // Add LibraryInfo if we have some.
185 if (LibraryInfo) MPM.add(new TargetLibraryInfo(*LibraryInfo));
187 addInitialAliasAnalysisPasses(MPM);
189 if (!DisableUnitAtATime) {
190 addExtensionsToPM(EP_ModuleOptimizerEarly, MPM);
192 MPM.add(createIPSCCPPass()); // IP SCCP
193 MPM.add(createGlobalOptimizerPass()); // Optimize out global vars
195 MPM.add(createDeadArgEliminationPass()); // Dead argument elimination
197 MPM.add(createInstructionCombiningPass());// Clean up after IPCP & DAE
198 addExtensionsToPM(EP_Peephole, MPM);
199 MPM.add(createCFGSimplificationPass()); // Clean up after IPCP & DAE
202 // Start of CallGraph SCC passes.
203 if (!DisableUnitAtATime)
204 MPM.add(createPruneEHPass()); // Remove dead EH info
209 if (!DisableUnitAtATime)
210 MPM.add(createFunctionAttrsPass()); // Set readonly/readnone attrs
212 MPM.add(createArgumentPromotionPass()); // Scalarize uninlined fn args
214 // Start of function pass.
215 // Break up aggregate allocas, using SSAUpdater.
217 MPM.add(createSROAPass(/*RequiresDomTree*/ false));
219 MPM.add(createScalarReplAggregatesPass(-1, false));
220 MPM.add(createEarlyCSEPass()); // Catch trivial redundancies
221 MPM.add(createJumpThreadingPass()); // Thread jumps.
222 MPM.add(createCorrelatedValuePropagationPass()); // Propagate conditionals
223 MPM.add(createCFGSimplificationPass()); // Merge & remove BBs
224 MPM.add(createInstructionCombiningPass()); // Combine silly seq's
225 addExtensionsToPM(EP_Peephole, MPM);
227 if (!DisableTailCalls)
228 MPM.add(createTailCallEliminationPass()); // Eliminate tail calls
229 MPM.add(createCFGSimplificationPass()); // Merge & remove BBs
230 MPM.add(createReassociatePass()); // Reassociate expressions
231 MPM.add(createLoopRotatePass()); // Rotate Loop
232 MPM.add(createLICMPass()); // Hoist loop invariants
233 MPM.add(createLoopUnswitchPass(SizeLevel || OptLevel < 3));
234 MPM.add(createInstructionCombiningPass());
235 MPM.add(createIndVarSimplifyPass()); // Canonicalize indvars
236 MPM.add(createLoopIdiomPass()); // Recognize idioms like memset.
237 MPM.add(createLoopDeletionPass()); // Delete dead loops
239 if (!DisableUnrollLoops)
240 MPM.add(createSimpleLoopUnrollPass()); // Unroll small loops
241 addExtensionsToPM(EP_LoopOptimizerEnd, MPM);
245 MPM.add(createMergedLoadStoreMotionPass()); // Merge ld/st in diamonds
246 MPM.add(createGVNPass(DisableGVNLoadPRE)); // Remove redundancies
248 MPM.add(createMemCpyOptPass()); // Remove memcpy / form memset
249 MPM.add(createSCCPPass()); // Constant prop with SCCP
251 // Run instcombine after redundancy elimination to exploit opportunities
252 // opened up by them.
253 MPM.add(createInstructionCombiningPass());
254 addExtensionsToPM(EP_Peephole, MPM);
255 MPM.add(createJumpThreadingPass()); // Thread jumps
256 MPM.add(createCorrelatedValuePropagationPass());
257 MPM.add(createDeadStoreEliminationPass()); // Delete dead stores
259 addExtensionsToPM(EP_ScalarOptimizerLate, MPM);
262 MPM.add(createLoopRerollPass());
263 if (!RunSLPAfterLoopVectorization) {
265 MPM.add(createSLPVectorizerPass()); // Vectorize parallel scalar chains.
268 MPM.add(createBBVectorizePass());
269 MPM.add(createInstructionCombiningPass());
270 addExtensionsToPM(EP_Peephole, MPM);
271 if (OptLevel > 1 && UseGVNAfterVectorization)
272 MPM.add(createGVNPass(DisableGVNLoadPRE)); // Remove redundancies
274 MPM.add(createEarlyCSEPass()); // Catch trivial redundancies
276 // BBVectorize may have significantly shortened a loop body; unroll again.
277 if (!DisableUnrollLoops)
278 MPM.add(createLoopUnrollPass());
283 MPM.add(createLoadCombinePass());
285 MPM.add(createAggressiveDCEPass()); // Delete dead instructions
286 MPM.add(createCFGSimplificationPass()); // Merge & remove BBs
287 MPM.add(createInstructionCombiningPass()); // Clean up after everything.
288 addExtensionsToPM(EP_Peephole, MPM);
290 // FIXME: This is a HACK! The inliner pass above implicitly creates a CGSCC
291 // pass manager that we are specifically trying to avoid. To prevent this
292 // we must insert a no-op module pass to reset the pass manager.
293 MPM.add(createBarrierNoopPass());
295 // Re-rotate loops in all our loop nests. These may have fallout out of
296 // rotated form due to GVN or other transformations, and the vectorizer relies
297 // on the rotated form.
298 if (ExtraVectorizerPasses)
299 MPM.add(createLoopRotatePass());
301 MPM.add(createLoopVectorizePass(DisableUnrollLoops, LoopVectorize));
302 // FIXME: Because of #pragma vectorize enable, the passes below are always
303 // inserted in the pipeline, even when the vectorizer doesn't run (ex. when
304 // on -O1 and no #pragma is found). Would be good to have these two passes
305 // as function calls, so that we can only pass them when the vectorizer
307 MPM.add(createInstructionCombiningPass());
308 if (OptLevel > 1 && ExtraVectorizerPasses) {
309 // At higher optimization levels, try to clean up any runtime overlap and
310 // alignment checks inserted by the vectorizer. We want to track correllated
311 // runtime checks for two inner loops in the same outer loop, fold any
312 // common computations, hoist loop-invariant aspects out of any outer loop,
313 // and unswitch the runtime checks if possible. Once hoisted, we may have
314 // dead (or speculatable) control flows or more combining opportunities.
315 MPM.add(createEarlyCSEPass());
316 MPM.add(createCorrelatedValuePropagationPass());
317 MPM.add(createInstructionCombiningPass());
318 MPM.add(createLICMPass());
319 MPM.add(createLoopUnswitchPass(SizeLevel || OptLevel < 3));
320 MPM.add(createCFGSimplificationPass());
321 MPM.add(createInstructionCombiningPass());
324 if (RunSLPAfterLoopVectorization) {
326 MPM.add(createSLPVectorizerPass()); // Vectorize parallel scalar chains.
327 if (OptLevel > 1 && ExtraVectorizerPasses) {
328 MPM.add(createEarlyCSEPass());
333 MPM.add(createBBVectorizePass());
334 MPM.add(createInstructionCombiningPass());
335 addExtensionsToPM(EP_Peephole, MPM);
336 if (OptLevel > 1 && UseGVNAfterVectorization)
337 MPM.add(createGVNPass(DisableGVNLoadPRE)); // Remove redundancies
339 MPM.add(createEarlyCSEPass()); // Catch trivial redundancies
341 // BBVectorize may have significantly shortened a loop body; unroll again.
342 if (!DisableUnrollLoops)
343 MPM.add(createLoopUnrollPass());
347 addExtensionsToPM(EP_Peephole, MPM);
348 MPM.add(createCFGSimplificationPass());
349 MPM.add(createInstructionCombiningPass());
351 if (!DisableUnrollLoops)
352 MPM.add(createLoopUnrollPass()); // Unroll small loops
354 // After vectorization and unrolling, assume intrinsics may tell us more
355 // about pointer alignments.
356 MPM.add(createAlignmentFromAssumptionsPass());
358 if (!DisableUnitAtATime) {
359 // FIXME: We shouldn't bother with this anymore.
360 MPM.add(createStripDeadPrototypesPass()); // Get rid of dead prototypes
362 // GlobalOpt already deletes dead functions and globals, at -O2 try a
363 // late pass of GlobalDCE. It is capable of deleting dead cycles.
365 MPM.add(createGlobalDCEPass()); // Remove dead fns and globals.
366 MPM.add(createConstantMergePass()); // Merge dup global constants
371 MPM.add(createMergeFunctionsPass());
373 addExtensionsToPM(EP_OptimizerLast, MPM);
376 void PassManagerBuilder::addLTOOptimizationPasses(PassManagerBase &PM) {
377 // Provide AliasAnalysis services for optimizations.
378 addInitialAliasAnalysisPasses(PM);
380 // Propagate constants at call sites into the functions they call. This
381 // opens opportunities for globalopt (and inlining) by substituting function
382 // pointers passed as arguments to direct uses of functions.
383 PM.add(createIPSCCPPass());
385 // Now that we internalized some globals, see if we can hack on them!
386 PM.add(createGlobalOptimizerPass());
388 // Linking modules together can lead to duplicated global constants, only
389 // keep one copy of each constant.
390 PM.add(createConstantMergePass());
392 // Remove unused arguments from functions.
393 PM.add(createDeadArgEliminationPass());
395 // Reduce the code after globalopt and ipsccp. Both can open up significant
396 // simplification opportunities, and both can propagate functions through
397 // function pointers. When this happens, we often have to resolve varargs
398 // calls, etc, so let instcombine do this.
399 PM.add(createInstructionCombiningPass());
400 addExtensionsToPM(EP_Peephole, PM);
402 // Inline small functions
403 bool RunInliner = Inliner;
409 PM.add(createPruneEHPass()); // Remove dead EH info.
411 // Optimize globals again if we ran the inliner.
413 PM.add(createGlobalOptimizerPass());
414 PM.add(createGlobalDCEPass()); // Remove dead functions.
416 // If we didn't decide to inline a function, check to see if we can
417 // transform it to pass arguments by value instead of by reference.
418 PM.add(createArgumentPromotionPass());
420 // The IPO passes may leave cruft around. Clean up after them.
421 PM.add(createInstructionCombiningPass());
422 addExtensionsToPM(EP_Peephole, PM);
423 PM.add(createJumpThreadingPass());
427 PM.add(createSROAPass());
429 PM.add(createScalarReplAggregatesPass());
431 // Run a few AA driven optimizations here and now, to cleanup the code.
432 PM.add(createFunctionAttrsPass()); // Add nocapture.
433 PM.add(createGlobalsModRefPass()); // IP alias analysis.
435 PM.add(createLICMPass()); // Hoist loop invariants.
437 PM.add(createMergedLoadStoreMotionPass()); // Merge ld/st in diamonds.
438 PM.add(createGVNPass(DisableGVNLoadPRE)); // Remove redundancies.
439 PM.add(createMemCpyOptPass()); // Remove dead memcpys.
442 PM.add(createDeadStoreEliminationPass());
444 // More loops are countable; try to optimize them.
445 PM.add(createIndVarSimplifyPass());
446 PM.add(createLoopDeletionPass());
447 PM.add(createLoopVectorizePass(DisableUnrollLoops, LoopVectorize));
449 // More scalar chains could be vectorized due to more alias information
450 if (RunSLPAfterLoopVectorization)
452 PM.add(createSLPVectorizerPass()); // Vectorize parallel scalar chains.
454 // After vectorization, assume intrinsics may tell us more about pointer
456 PM.add(createAlignmentFromAssumptionsPass());
459 PM.add(createLoadCombinePass());
461 // Cleanup and simplify the code after the scalar optimizations.
462 PM.add(createInstructionCombiningPass());
463 addExtensionsToPM(EP_Peephole, PM);
465 PM.add(createJumpThreadingPass());
467 // Delete basic blocks, which optimization passes may have killed.
468 PM.add(createCFGSimplificationPass());
470 // Now that we have optimized the program, discard unreachable functions.
471 PM.add(createGlobalDCEPass());
473 // FIXME: this is profitable (for compiler time) to do at -O0 too, but
474 // currently it damages debug info.
476 PM.add(createMergeFunctionsPass());
479 void PassManagerBuilder::populateLTOPassManager(PassManagerBase &PM,
482 PM.add(new DataLayoutPass());
483 TM->addAnalysisPasses(PM);
487 PM.add(new TargetLibraryInfo(*LibraryInfo));
490 PM.add(createVerifierPass());
493 PM.add(createStripSymbolsPass(true));
496 PM.add(createDebugInfoVerifierPass());
499 addLTOOptimizationPasses(PM);
502 PM.add(createVerifierPass());
503 PM.add(createDebugInfoVerifierPass());
507 inline PassManagerBuilder *unwrap(LLVMPassManagerBuilderRef P) {
508 return reinterpret_cast<PassManagerBuilder*>(P);
511 inline LLVMPassManagerBuilderRef wrap(PassManagerBuilder *P) {
512 return reinterpret_cast<LLVMPassManagerBuilderRef>(P);
515 LLVMPassManagerBuilderRef LLVMPassManagerBuilderCreate() {
516 PassManagerBuilder *PMB = new PassManagerBuilder();
520 void LLVMPassManagerBuilderDispose(LLVMPassManagerBuilderRef PMB) {
521 PassManagerBuilder *Builder = unwrap(PMB);
526 LLVMPassManagerBuilderSetOptLevel(LLVMPassManagerBuilderRef PMB,
528 PassManagerBuilder *Builder = unwrap(PMB);
529 Builder->OptLevel = OptLevel;
533 LLVMPassManagerBuilderSetSizeLevel(LLVMPassManagerBuilderRef PMB,
534 unsigned SizeLevel) {
535 PassManagerBuilder *Builder = unwrap(PMB);
536 Builder->SizeLevel = SizeLevel;
540 LLVMPassManagerBuilderSetDisableUnitAtATime(LLVMPassManagerBuilderRef PMB,
542 PassManagerBuilder *Builder = unwrap(PMB);
543 Builder->DisableUnitAtATime = Value;
547 LLVMPassManagerBuilderSetDisableUnrollLoops(LLVMPassManagerBuilderRef PMB,
549 PassManagerBuilder *Builder = unwrap(PMB);
550 Builder->DisableUnrollLoops = Value;
554 LLVMPassManagerBuilderSetDisableSimplifyLibCalls(LLVMPassManagerBuilderRef PMB,
556 // NOTE: The simplify-libcalls pass has been removed.
560 LLVMPassManagerBuilderUseInlinerWithThreshold(LLVMPassManagerBuilderRef PMB,
561 unsigned Threshold) {
562 PassManagerBuilder *Builder = unwrap(PMB);
563 Builder->Inliner = createFunctionInliningPass(Threshold);
567 LLVMPassManagerBuilderPopulateFunctionPassManager(LLVMPassManagerBuilderRef PMB,
568 LLVMPassManagerRef PM) {
569 PassManagerBuilder *Builder = unwrap(PMB);
570 FunctionPassManager *FPM = unwrap<FunctionPassManager>(PM);
571 Builder->populateFunctionPassManager(*FPM);
575 LLVMPassManagerBuilderPopulateModulePassManager(LLVMPassManagerBuilderRef PMB,
576 LLVMPassManagerRef PM) {
577 PassManagerBuilder *Builder = unwrap(PMB);
578 PassManagerBase *MPM = unwrap(PM);
579 Builder->populateModulePassManager(*MPM);
582 void LLVMPassManagerBuilderPopulateLTOPassManager(LLVMPassManagerBuilderRef PMB,
583 LLVMPassManagerRef PM,
584 LLVMBool Internalize,
585 LLVMBool RunInliner) {
586 PassManagerBuilder *Builder = unwrap(PMB);
587 PassManagerBase *LPM = unwrap(PM);
589 // A small backwards compatibility hack. populateLTOPassManager used to take
590 // an RunInliner option.
591 if (RunInliner && !Builder->Inliner)
592 Builder->Inliner = createFunctionInliningPass();
594 Builder->populateLTOPassManager(*LPM);