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 // Rotate Loop - disable header duplication at -Oz
232 MPM.add(createLoopRotatePass(SizeLevel == 2 ? 0 : -1));
233 MPM.add(createLICMPass()); // Hoist loop invariants
234 MPM.add(createLoopUnswitchPass(SizeLevel || OptLevel < 3));
235 MPM.add(createInstructionCombiningPass());
236 MPM.add(createIndVarSimplifyPass()); // Canonicalize indvars
237 MPM.add(createLoopIdiomPass()); // Recognize idioms like memset.
238 MPM.add(createLoopDeletionPass()); // Delete dead loops
240 if (!DisableUnrollLoops)
241 MPM.add(createSimpleLoopUnrollPass()); // Unroll small loops
242 addExtensionsToPM(EP_LoopOptimizerEnd, MPM);
246 MPM.add(createMergedLoadStoreMotionPass()); // Merge ld/st in diamonds
247 MPM.add(createGVNPass(DisableGVNLoadPRE)); // Remove redundancies
249 MPM.add(createMemCpyOptPass()); // Remove memcpy / form memset
250 MPM.add(createSCCPPass()); // Constant prop with SCCP
252 // Run instcombine after redundancy elimination to exploit opportunities
253 // opened up by them.
254 MPM.add(createInstructionCombiningPass());
255 addExtensionsToPM(EP_Peephole, MPM);
256 MPM.add(createJumpThreadingPass()); // Thread jumps
257 MPM.add(createCorrelatedValuePropagationPass());
258 MPM.add(createDeadStoreEliminationPass()); // Delete dead stores
260 addExtensionsToPM(EP_ScalarOptimizerLate, MPM);
263 MPM.add(createLoopRerollPass());
264 if (!RunSLPAfterLoopVectorization) {
266 MPM.add(createSLPVectorizerPass()); // Vectorize parallel scalar chains.
269 MPM.add(createBBVectorizePass());
270 MPM.add(createInstructionCombiningPass());
271 addExtensionsToPM(EP_Peephole, MPM);
272 if (OptLevel > 1 && UseGVNAfterVectorization)
273 MPM.add(createGVNPass(DisableGVNLoadPRE)); // Remove redundancies
275 MPM.add(createEarlyCSEPass()); // Catch trivial redundancies
277 // BBVectorize may have significantly shortened a loop body; unroll again.
278 if (!DisableUnrollLoops)
279 MPM.add(createLoopUnrollPass());
284 MPM.add(createLoadCombinePass());
286 MPM.add(createAggressiveDCEPass()); // Delete dead instructions
287 MPM.add(createCFGSimplificationPass()); // Merge & remove BBs
288 MPM.add(createInstructionCombiningPass()); // Clean up after everything.
289 addExtensionsToPM(EP_Peephole, MPM);
291 // FIXME: This is a HACK! The inliner pass above implicitly creates a CGSCC
292 // pass manager that we are specifically trying to avoid. To prevent this
293 // we must insert a no-op module pass to reset the pass manager.
294 MPM.add(createBarrierNoopPass());
296 // Re-rotate loops in all our loop nests. These may have fallout out of
297 // rotated form due to GVN or other transformations, and the vectorizer relies
298 // on the rotated form.
299 if (ExtraVectorizerPasses)
300 MPM.add(createLoopRotatePass());
302 MPM.add(createLoopVectorizePass(DisableUnrollLoops, LoopVectorize));
303 // FIXME: Because of #pragma vectorize enable, the passes below are always
304 // inserted in the pipeline, even when the vectorizer doesn't run (ex. when
305 // on -O1 and no #pragma is found). Would be good to have these two passes
306 // as function calls, so that we can only pass them when the vectorizer
308 MPM.add(createInstructionCombiningPass());
309 if (OptLevel > 1 && ExtraVectorizerPasses) {
310 // At higher optimization levels, try to clean up any runtime overlap and
311 // alignment checks inserted by the vectorizer. We want to track correllated
312 // runtime checks for two inner loops in the same outer loop, fold any
313 // common computations, hoist loop-invariant aspects out of any outer loop,
314 // and unswitch the runtime checks if possible. Once hoisted, we may have
315 // dead (or speculatable) control flows or more combining opportunities.
316 MPM.add(createEarlyCSEPass());
317 MPM.add(createCorrelatedValuePropagationPass());
318 MPM.add(createInstructionCombiningPass());
319 MPM.add(createLICMPass());
320 MPM.add(createLoopUnswitchPass(SizeLevel || OptLevel < 3));
321 MPM.add(createCFGSimplificationPass());
322 MPM.add(createInstructionCombiningPass());
325 if (RunSLPAfterLoopVectorization) {
327 MPM.add(createSLPVectorizerPass()); // Vectorize parallel scalar chains.
328 if (OptLevel > 1 && ExtraVectorizerPasses) {
329 MPM.add(createEarlyCSEPass());
334 MPM.add(createBBVectorizePass());
335 MPM.add(createInstructionCombiningPass());
336 addExtensionsToPM(EP_Peephole, MPM);
337 if (OptLevel > 1 && UseGVNAfterVectorization)
338 MPM.add(createGVNPass(DisableGVNLoadPRE)); // Remove redundancies
340 MPM.add(createEarlyCSEPass()); // Catch trivial redundancies
342 // BBVectorize may have significantly shortened a loop body; unroll again.
343 if (!DisableUnrollLoops)
344 MPM.add(createLoopUnrollPass());
348 addExtensionsToPM(EP_Peephole, MPM);
349 MPM.add(createCFGSimplificationPass());
350 MPM.add(createInstructionCombiningPass());
352 if (!DisableUnrollLoops)
353 MPM.add(createLoopUnrollPass()); // Unroll small loops
355 // After vectorization and unrolling, assume intrinsics may tell us more
356 // about pointer alignments.
357 MPM.add(createAlignmentFromAssumptionsPass());
359 if (!DisableUnitAtATime) {
360 // FIXME: We shouldn't bother with this anymore.
361 MPM.add(createStripDeadPrototypesPass()); // Get rid of dead prototypes
363 // GlobalOpt already deletes dead functions and globals, at -O2 try a
364 // late pass of GlobalDCE. It is capable of deleting dead cycles.
366 MPM.add(createGlobalDCEPass()); // Remove dead fns and globals.
367 MPM.add(createConstantMergePass()); // Merge dup global constants
372 MPM.add(createMergeFunctionsPass());
374 addExtensionsToPM(EP_OptimizerLast, MPM);
377 void PassManagerBuilder::addLTOOptimizationPasses(PassManagerBase &PM) {
378 // Provide AliasAnalysis services for optimizations.
379 addInitialAliasAnalysisPasses(PM);
381 // Propagate constants at call sites into the functions they call. This
382 // opens opportunities for globalopt (and inlining) by substituting function
383 // pointers passed as arguments to direct uses of functions.
384 PM.add(createIPSCCPPass());
386 // Now that we internalized some globals, see if we can hack on them!
387 PM.add(createGlobalOptimizerPass());
389 // Linking modules together can lead to duplicated global constants, only
390 // keep one copy of each constant.
391 PM.add(createConstantMergePass());
393 // Remove unused arguments from functions.
394 PM.add(createDeadArgEliminationPass());
396 // Reduce the code after globalopt and ipsccp. Both can open up significant
397 // simplification opportunities, and both can propagate functions through
398 // function pointers. When this happens, we often have to resolve varargs
399 // calls, etc, so let instcombine do this.
400 PM.add(createInstructionCombiningPass());
401 addExtensionsToPM(EP_Peephole, PM);
403 // Inline small functions
404 bool RunInliner = Inliner;
410 PM.add(createPruneEHPass()); // Remove dead EH info.
412 // Optimize globals again if we ran the inliner.
414 PM.add(createGlobalOptimizerPass());
415 PM.add(createGlobalDCEPass()); // Remove dead functions.
417 // If we didn't decide to inline a function, check to see if we can
418 // transform it to pass arguments by value instead of by reference.
419 PM.add(createArgumentPromotionPass());
421 // The IPO passes may leave cruft around. Clean up after them.
422 PM.add(createInstructionCombiningPass());
423 addExtensionsToPM(EP_Peephole, PM);
424 PM.add(createJumpThreadingPass());
428 PM.add(createSROAPass());
430 PM.add(createScalarReplAggregatesPass());
432 // Run a few AA driven optimizations here and now, to cleanup the code.
433 PM.add(createFunctionAttrsPass()); // Add nocapture.
434 PM.add(createGlobalsModRefPass()); // IP alias analysis.
436 PM.add(createLICMPass()); // Hoist loop invariants.
438 PM.add(createMergedLoadStoreMotionPass()); // Merge ld/st in diamonds.
439 PM.add(createGVNPass(DisableGVNLoadPRE)); // Remove redundancies.
440 PM.add(createMemCpyOptPass()); // Remove dead memcpys.
443 PM.add(createDeadStoreEliminationPass());
445 // More loops are countable; try to optimize them.
446 PM.add(createIndVarSimplifyPass());
447 PM.add(createLoopDeletionPass());
448 PM.add(createLoopVectorizePass(true, LoopVectorize));
450 // More scalar chains could be vectorized due to more alias information
451 if (RunSLPAfterLoopVectorization)
453 PM.add(createSLPVectorizerPass()); // Vectorize parallel scalar chains.
455 // After vectorization, assume intrinsics may tell us more about pointer
457 PM.add(createAlignmentFromAssumptionsPass());
460 PM.add(createLoadCombinePass());
462 // Cleanup and simplify the code after the scalar optimizations.
463 PM.add(createInstructionCombiningPass());
464 addExtensionsToPM(EP_Peephole, PM);
466 PM.add(createJumpThreadingPass());
468 // Delete basic blocks, which optimization passes may have killed.
469 PM.add(createCFGSimplificationPass());
471 // Now that we have optimized the program, discard unreachable functions.
472 PM.add(createGlobalDCEPass());
474 // FIXME: this is profitable (for compiler time) to do at -O0 too, but
475 // currently it damages debug info.
477 PM.add(createMergeFunctionsPass());
480 void PassManagerBuilder::populateLTOPassManager(PassManagerBase &PM,
483 PM.add(new DataLayoutPass());
484 TM->addAnalysisPasses(PM);
488 PM.add(new TargetLibraryInfo(*LibraryInfo));
491 PM.add(createVerifierPass());
494 PM.add(createStripSymbolsPass(true));
497 PM.add(createDebugInfoVerifierPass());
500 addLTOOptimizationPasses(PM);
503 PM.add(createVerifierPass());
504 PM.add(createDebugInfoVerifierPass());
508 inline PassManagerBuilder *unwrap(LLVMPassManagerBuilderRef P) {
509 return reinterpret_cast<PassManagerBuilder*>(P);
512 inline LLVMPassManagerBuilderRef wrap(PassManagerBuilder *P) {
513 return reinterpret_cast<LLVMPassManagerBuilderRef>(P);
516 LLVMPassManagerBuilderRef LLVMPassManagerBuilderCreate() {
517 PassManagerBuilder *PMB = new PassManagerBuilder();
521 void LLVMPassManagerBuilderDispose(LLVMPassManagerBuilderRef PMB) {
522 PassManagerBuilder *Builder = unwrap(PMB);
527 LLVMPassManagerBuilderSetOptLevel(LLVMPassManagerBuilderRef PMB,
529 PassManagerBuilder *Builder = unwrap(PMB);
530 Builder->OptLevel = OptLevel;
534 LLVMPassManagerBuilderSetSizeLevel(LLVMPassManagerBuilderRef PMB,
535 unsigned SizeLevel) {
536 PassManagerBuilder *Builder = unwrap(PMB);
537 Builder->SizeLevel = SizeLevel;
541 LLVMPassManagerBuilderSetDisableUnitAtATime(LLVMPassManagerBuilderRef PMB,
543 PassManagerBuilder *Builder = unwrap(PMB);
544 Builder->DisableUnitAtATime = Value;
548 LLVMPassManagerBuilderSetDisableUnrollLoops(LLVMPassManagerBuilderRef PMB,
550 PassManagerBuilder *Builder = unwrap(PMB);
551 Builder->DisableUnrollLoops = Value;
555 LLVMPassManagerBuilderSetDisableSimplifyLibCalls(LLVMPassManagerBuilderRef PMB,
557 // NOTE: The simplify-libcalls pass has been removed.
561 LLVMPassManagerBuilderUseInlinerWithThreshold(LLVMPassManagerBuilderRef PMB,
562 unsigned Threshold) {
563 PassManagerBuilder *Builder = unwrap(PMB);
564 Builder->Inliner = createFunctionInliningPass(Threshold);
568 LLVMPassManagerBuilderPopulateFunctionPassManager(LLVMPassManagerBuilderRef PMB,
569 LLVMPassManagerRef PM) {
570 PassManagerBuilder *Builder = unwrap(PMB);
571 FunctionPassManager *FPM = unwrap<FunctionPassManager>(PM);
572 Builder->populateFunctionPassManager(*FPM);
576 LLVMPassManagerBuilderPopulateModulePassManager(LLVMPassManagerBuilderRef PMB,
577 LLVMPassManagerRef PM) {
578 PassManagerBuilder *Builder = unwrap(PMB);
579 PassManagerBase *MPM = unwrap(PM);
580 Builder->populateModulePassManager(*MPM);
583 void LLVMPassManagerBuilderPopulateLTOPassManager(LLVMPassManagerBuilderRef PMB,
584 LLVMPassManagerRef PM,
585 LLVMBool Internalize,
586 LLVMBool RunInliner) {
587 PassManagerBuilder *Builder = unwrap(PMB);
588 PassManagerBase *LPM = unwrap(PM);
590 // A small backwards compatibility hack. populateLTOPassManager used to take
591 // an RunInliner option.
592 if (RunInliner && !Builder->Inliner)
593 Builder->Inliner = createFunctionInliningPass();
595 Builder->populateLTOPassManager(*LPM);