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"));
63 RunFloat2Int("float-to-int", cl::Hidden, cl::init(true),
64 cl::desc("Run the float2int (float demotion) pass"));
66 static cl::opt<bool> RunLoadCombine("combine-loads", cl::init(false),
68 cl::desc("Run the load combining pass"));
71 RunSLPAfterLoopVectorization("run-slp-after-loop-vectorization",
72 cl::init(true), cl::Hidden,
73 cl::desc("Run the SLP vectorizer (and BB vectorizer) after the Loop "
74 "vectorizer instead of before"));
76 static cl::opt<bool> UseCFLAA("use-cfl-aa",
77 cl::init(false), cl::Hidden,
78 cl::desc("Enable the new, experimental CFL alias analysis"));
81 EnableMLSM("mlsm", cl::init(true), cl::Hidden,
82 cl::desc("Enable motion of merged load and store"));
84 static cl::opt<bool> EnableLoopInterchange(
85 "enable-loopinterchange", cl::init(false), cl::Hidden,
86 cl::desc("Enable the new, experimental LoopInterchange Pass"));
88 static cl::opt<bool> EnableLoopDistribute(
89 "enable-loop-distribute", cl::init(false), cl::Hidden,
90 cl::desc("Enable the new, experimental LoopDistribution Pass"));
92 PassManagerBuilder::PassManagerBuilder() {
95 LibraryInfo = nullptr;
97 DisableTailCalls = false;
98 DisableUnitAtATime = false;
99 DisableUnrollLoops = false;
100 BBVectorize = RunBBVectorization;
101 SLPVectorize = RunSLPVectorization;
102 LoopVectorize = RunLoopVectorization;
103 RerollLoops = RunLoopRerolling;
104 LoadCombine = RunLoadCombine;
105 DisableGVNLoadPRE = false;
107 VerifyOutput = false;
108 MergeFunctions = false;
111 PassManagerBuilder::~PassManagerBuilder() {
116 /// Set of global extensions, automatically added as part of the standard set.
117 static ManagedStatic<SmallVector<std::pair<PassManagerBuilder::ExtensionPointTy,
118 PassManagerBuilder::ExtensionFn>, 8> > GlobalExtensions;
120 void PassManagerBuilder::addGlobalExtension(
121 PassManagerBuilder::ExtensionPointTy Ty,
122 PassManagerBuilder::ExtensionFn Fn) {
123 GlobalExtensions->push_back(std::make_pair(Ty, Fn));
126 void PassManagerBuilder::addExtension(ExtensionPointTy Ty, ExtensionFn Fn) {
127 Extensions.push_back(std::make_pair(Ty, Fn));
130 void PassManagerBuilder::addExtensionsToPM(ExtensionPointTy ETy,
131 legacy::PassManagerBase &PM) const {
132 for (unsigned i = 0, e = GlobalExtensions->size(); i != e; ++i)
133 if ((*GlobalExtensions)[i].first == ETy)
134 (*GlobalExtensions)[i].second(*this, PM);
135 for (unsigned i = 0, e = Extensions.size(); i != e; ++i)
136 if (Extensions[i].first == ETy)
137 Extensions[i].second(*this, PM);
140 void PassManagerBuilder::addInitialAliasAnalysisPasses(
141 legacy::PassManagerBase &PM) const {
142 // Add TypeBasedAliasAnalysis before BasicAliasAnalysis so that
143 // BasicAliasAnalysis wins if they disagree. This is intended to help
144 // support "obvious" type-punning idioms.
146 PM.add(createCFLAliasAnalysisPass());
147 PM.add(createTypeBasedAliasAnalysisPass());
148 PM.add(createScopedNoAliasAAPass());
149 PM.add(createBasicAliasAnalysisPass());
152 void PassManagerBuilder::populateFunctionPassManager(
153 legacy::FunctionPassManager &FPM) {
154 addExtensionsToPM(EP_EarlyAsPossible, FPM);
156 // Add LibraryInfo if we have some.
158 FPM.add(new TargetLibraryInfoWrapperPass(*LibraryInfo));
160 if (OptLevel == 0) return;
162 addInitialAliasAnalysisPasses(FPM);
164 FPM.add(createCFGSimplificationPass());
166 FPM.add(createSROAPass());
168 FPM.add(createScalarReplAggregatesPass());
169 FPM.add(createEarlyCSEPass());
170 FPM.add(createLowerExpectIntrinsicPass());
173 void PassManagerBuilder::populateModulePassManager(
174 legacy::PassManagerBase &MPM) {
175 // If all optimizations are disabled, just run the always-inline pass and,
176 // if enabled, the function merging pass.
183 // FIXME: The BarrierNoopPass is a HACK! The inliner pass above implicitly
184 // creates a CGSCC pass manager, but we don't want to add extensions into
185 // that pass manager. To prevent this we insert a no-op module pass to reset
186 // the pass manager to get the same behavior as EP_OptimizerLast in non-O0
187 // builds. The function merging pass is
189 MPM.add(createMergeFunctionsPass());
190 else if (!GlobalExtensions->empty() || !Extensions.empty())
191 MPM.add(createBarrierNoopPass());
193 addExtensionsToPM(EP_EnabledOnOptLevel0, MPM);
197 // Add LibraryInfo if we have some.
199 MPM.add(new TargetLibraryInfoWrapperPass(*LibraryInfo));
201 addInitialAliasAnalysisPasses(MPM);
203 if (!DisableUnitAtATime) {
204 addExtensionsToPM(EP_ModuleOptimizerEarly, MPM);
206 MPM.add(createIPSCCPPass()); // IP SCCP
207 MPM.add(createGlobalOptimizerPass()); // Optimize out global vars
209 MPM.add(createDeadArgEliminationPass()); // Dead argument elimination
211 MPM.add(createInstructionCombiningPass());// Clean up after IPCP & DAE
212 addExtensionsToPM(EP_Peephole, MPM);
213 MPM.add(createCFGSimplificationPass()); // Clean up after IPCP & DAE
216 // Start of CallGraph SCC passes.
217 if (!DisableUnitAtATime)
218 MPM.add(createPruneEHPass()); // Remove dead EH info
223 if (!DisableUnitAtATime)
224 MPM.add(createFunctionAttrsPass()); // Set readonly/readnone attrs
226 MPM.add(createArgumentPromotionPass()); // Scalarize uninlined fn args
228 // Start of function pass.
229 // Break up aggregate allocas, using SSAUpdater.
231 MPM.add(createSROAPass(/*RequiresDomTree*/ false));
233 MPM.add(createScalarReplAggregatesPass(-1, false));
234 MPM.add(createEarlyCSEPass()); // Catch trivial redundancies
235 MPM.add(createJumpThreadingPass()); // Thread jumps.
236 MPM.add(createCorrelatedValuePropagationPass()); // Propagate conditionals
237 MPM.add(createCFGSimplificationPass()); // Merge & remove BBs
238 MPM.add(createInstructionCombiningPass()); // Combine silly seq's
239 addExtensionsToPM(EP_Peephole, MPM);
241 if (!DisableTailCalls)
242 MPM.add(createTailCallEliminationPass()); // Eliminate tail calls
243 MPM.add(createCFGSimplificationPass()); // Merge & remove BBs
244 MPM.add(createReassociatePass()); // Reassociate expressions
245 // Rotate Loop - disable header duplication at -Oz
246 MPM.add(createLoopRotatePass(SizeLevel == 2 ? 0 : -1));
247 MPM.add(createLICMPass()); // Hoist loop invariants
248 MPM.add(createLoopUnswitchPass(SizeLevel || OptLevel < 3));
249 MPM.add(createInstructionCombiningPass());
250 MPM.add(createIndVarSimplifyPass()); // Canonicalize indvars
251 MPM.add(createLoopIdiomPass()); // Recognize idioms like memset.
252 MPM.add(createLoopDeletionPass()); // Delete dead loops
253 if (EnableLoopInterchange) {
254 MPM.add(createLoopInterchangePass()); // Interchange loops
255 MPM.add(createCFGSimplificationPass());
257 if (!DisableUnrollLoops)
258 MPM.add(createSimpleLoopUnrollPass()); // Unroll small loops
259 addExtensionsToPM(EP_LoopOptimizerEnd, MPM);
263 MPM.add(createMergedLoadStoreMotionPass()); // Merge ld/st in diamonds
264 MPM.add(createGVNPass(DisableGVNLoadPRE)); // Remove redundancies
266 MPM.add(createMemCpyOptPass()); // Remove memcpy / form memset
267 MPM.add(createSCCPPass()); // Constant prop with SCCP
269 // Delete dead bit computations (instcombine runs after to fold away the dead
270 // computations, and then ADCE will run later to exploit any new DCE
271 // opportunities that creates).
272 MPM.add(createBitTrackingDCEPass()); // Delete dead bit computations
274 // Run instcombine after redundancy elimination to exploit opportunities
275 // opened up by them.
276 MPM.add(createInstructionCombiningPass());
277 addExtensionsToPM(EP_Peephole, MPM);
278 MPM.add(createJumpThreadingPass()); // Thread jumps
279 MPM.add(createCorrelatedValuePropagationPass());
280 MPM.add(createDeadStoreEliminationPass()); // Delete dead stores
281 MPM.add(createLICMPass());
283 addExtensionsToPM(EP_ScalarOptimizerLate, MPM);
286 MPM.add(createLoopRerollPass());
287 if (!RunSLPAfterLoopVectorization) {
289 MPM.add(createSLPVectorizerPass()); // Vectorize parallel scalar chains.
292 MPM.add(createBBVectorizePass());
293 MPM.add(createInstructionCombiningPass());
294 addExtensionsToPM(EP_Peephole, MPM);
295 if (OptLevel > 1 && UseGVNAfterVectorization)
296 MPM.add(createGVNPass(DisableGVNLoadPRE)); // Remove redundancies
298 MPM.add(createEarlyCSEPass()); // Catch trivial redundancies
300 // BBVectorize may have significantly shortened a loop body; unroll again.
301 if (!DisableUnrollLoops)
302 MPM.add(createLoopUnrollPass());
307 MPM.add(createLoadCombinePass());
309 MPM.add(createAggressiveDCEPass()); // Delete dead instructions
310 MPM.add(createCFGSimplificationPass()); // Merge & remove BBs
311 MPM.add(createInstructionCombiningPass()); // Clean up after everything.
312 addExtensionsToPM(EP_Peephole, MPM);
314 // FIXME: This is a HACK! The inliner pass above implicitly creates a CGSCC
315 // pass manager that we are specifically trying to avoid. To prevent this
316 // we must insert a no-op module pass to reset the pass manager.
317 MPM.add(createBarrierNoopPass());
320 MPM.add(createFloat2IntPass());
322 // Re-rotate loops in all our loop nests. These may have fallout out of
323 // rotated form due to GVN or other transformations, and the vectorizer relies
324 // on the rotated form.
325 MPM.add(createLoopRotatePass());
327 // Distribute loops to allow partial vectorization. I.e. isolate dependences
328 // into separate loop that would otherwise inhibit vectorization.
329 if (EnableLoopDistribute)
330 MPM.add(createLoopDistributePass());
332 MPM.add(createLoopVectorizePass(DisableUnrollLoops, LoopVectorize));
333 // FIXME: Because of #pragma vectorize enable, the passes below are always
334 // inserted in the pipeline, even when the vectorizer doesn't run (ex. when
335 // on -O1 and no #pragma is found). Would be good to have these two passes
336 // as function calls, so that we can only pass them when the vectorizer
338 MPM.add(createInstructionCombiningPass());
339 if (OptLevel > 1 && ExtraVectorizerPasses) {
340 // At higher optimization levels, try to clean up any runtime overlap and
341 // alignment checks inserted by the vectorizer. We want to track correllated
342 // runtime checks for two inner loops in the same outer loop, fold any
343 // common computations, hoist loop-invariant aspects out of any outer loop,
344 // and unswitch the runtime checks if possible. Once hoisted, we may have
345 // dead (or speculatable) control flows or more combining opportunities.
346 MPM.add(createEarlyCSEPass());
347 MPM.add(createCorrelatedValuePropagationPass());
348 MPM.add(createInstructionCombiningPass());
349 MPM.add(createLICMPass());
350 MPM.add(createLoopUnswitchPass(SizeLevel || OptLevel < 3));
351 MPM.add(createCFGSimplificationPass());
352 MPM.add(createInstructionCombiningPass());
355 if (RunSLPAfterLoopVectorization) {
357 MPM.add(createSLPVectorizerPass()); // Vectorize parallel scalar chains.
358 if (OptLevel > 1 && ExtraVectorizerPasses) {
359 MPM.add(createEarlyCSEPass());
364 MPM.add(createBBVectorizePass());
365 MPM.add(createInstructionCombiningPass());
366 addExtensionsToPM(EP_Peephole, MPM);
367 if (OptLevel > 1 && UseGVNAfterVectorization)
368 MPM.add(createGVNPass(DisableGVNLoadPRE)); // Remove redundancies
370 MPM.add(createEarlyCSEPass()); // Catch trivial redundancies
372 // BBVectorize may have significantly shortened a loop body; unroll again.
373 if (!DisableUnrollLoops)
374 MPM.add(createLoopUnrollPass());
378 addExtensionsToPM(EP_Peephole, MPM);
379 MPM.add(createCFGSimplificationPass());
380 MPM.add(createInstructionCombiningPass());
382 if (!DisableUnrollLoops) {
383 MPM.add(createLoopUnrollPass()); // Unroll small loops
385 // This is a barrier pass to avoid combine LICM pass and loop unroll pass
386 // within same loop pass manager.
387 MPM.add(createInstructionSimplifierPass());
389 // Runtime unrolling will introduce runtime check in loop prologue. If the
390 // unrolled loop is a inner loop, then the prologue will be inside the
391 // outer loop. LICM pass can help to promote the runtime check out if the
392 // checked value is loop invariant.
393 MPM.add(createLICMPass());
396 // After vectorization and unrolling, assume intrinsics may tell us more
397 // about pointer alignments.
398 MPM.add(createAlignmentFromAssumptionsPass());
400 if (!DisableUnitAtATime) {
401 // FIXME: We shouldn't bother with this anymore.
402 MPM.add(createStripDeadPrototypesPass()); // Get rid of dead prototypes
404 // GlobalOpt already deletes dead functions and globals, at -O2 try a
405 // late pass of GlobalDCE. It is capable of deleting dead cycles.
407 MPM.add(createGlobalDCEPass()); // Remove dead fns and globals.
408 MPM.add(createConstantMergePass()); // Merge dup global constants
413 MPM.add(createMergeFunctionsPass());
415 addExtensionsToPM(EP_OptimizerLast, MPM);
418 void PassManagerBuilder::addLTOOptimizationPasses(legacy::PassManagerBase &PM) {
419 // Provide AliasAnalysis services for optimizations.
420 addInitialAliasAnalysisPasses(PM);
422 // Propagate constants at call sites into the functions they call. This
423 // opens opportunities for globalopt (and inlining) by substituting function
424 // pointers passed as arguments to direct uses of functions.
425 PM.add(createIPSCCPPass());
427 // Now that we internalized some globals, see if we can hack on them!
428 PM.add(createGlobalOptimizerPass());
430 // Linking modules together can lead to duplicated global constants, only
431 // keep one copy of each constant.
432 PM.add(createConstantMergePass());
434 // Remove unused arguments from functions.
435 PM.add(createDeadArgEliminationPass());
437 // Reduce the code after globalopt and ipsccp. Both can open up significant
438 // simplification opportunities, and both can propagate functions through
439 // function pointers. When this happens, we often have to resolve varargs
440 // calls, etc, so let instcombine do this.
441 PM.add(createInstructionCombiningPass());
442 addExtensionsToPM(EP_Peephole, PM);
444 // Inline small functions
445 bool RunInliner = Inliner;
451 PM.add(createPruneEHPass()); // Remove dead EH info.
453 // Optimize globals again if we ran the inliner.
455 PM.add(createGlobalOptimizerPass());
456 PM.add(createGlobalDCEPass()); // Remove dead functions.
458 // If we didn't decide to inline a function, check to see if we can
459 // transform it to pass arguments by value instead of by reference.
460 PM.add(createArgumentPromotionPass());
462 // The IPO passes may leave cruft around. Clean up after them.
463 PM.add(createInstructionCombiningPass());
464 addExtensionsToPM(EP_Peephole, PM);
465 PM.add(createJumpThreadingPass());
469 PM.add(createSROAPass());
471 PM.add(createScalarReplAggregatesPass());
473 // Run a few AA driven optimizations here and now, to cleanup the code.
474 PM.add(createFunctionAttrsPass()); // Add nocapture.
475 PM.add(createGlobalsModRefPass()); // IP alias analysis.
477 PM.add(createLICMPass()); // Hoist loop invariants.
479 PM.add(createMergedLoadStoreMotionPass()); // Merge ld/st in diamonds.
480 PM.add(createGVNPass(DisableGVNLoadPRE)); // Remove redundancies.
481 PM.add(createMemCpyOptPass()); // Remove dead memcpys.
484 PM.add(createDeadStoreEliminationPass());
486 // More loops are countable; try to optimize them.
487 PM.add(createIndVarSimplifyPass());
488 PM.add(createLoopDeletionPass());
489 if (EnableLoopInterchange)
490 PM.add(createLoopInterchangePass());
492 PM.add(createLoopVectorizePass(true, LoopVectorize));
494 // More scalar chains could be vectorized due to more alias information
495 if (RunSLPAfterLoopVectorization)
497 PM.add(createSLPVectorizerPass()); // Vectorize parallel scalar chains.
499 // After vectorization, assume intrinsics may tell us more about pointer
501 PM.add(createAlignmentFromAssumptionsPass());
504 PM.add(createLoadCombinePass());
506 // Cleanup and simplify the code after the scalar optimizations.
507 PM.add(createInstructionCombiningPass());
508 addExtensionsToPM(EP_Peephole, PM);
510 PM.add(createJumpThreadingPass());
513 void PassManagerBuilder::addLateLTOOptimizationPasses(
514 legacy::PassManagerBase &PM) {
515 // Delete basic blocks, which optimization passes may have killed.
516 PM.add(createCFGSimplificationPass());
518 // Now that we have optimized the program, discard unreachable functions.
519 PM.add(createGlobalDCEPass());
521 // FIXME: this is profitable (for compiler time) to do at -O0 too, but
522 // currently it damages debug info.
524 PM.add(createMergeFunctionsPass());
527 void PassManagerBuilder::populateLTOPassManager(legacy::PassManagerBase &PM) {
529 PM.add(new TargetLibraryInfoWrapperPass(*LibraryInfo));
532 PM.add(createVerifierPass());
535 addLTOOptimizationPasses(PM);
537 // Lower bit sets to globals. This pass supports Clang's control flow
538 // integrity mechanisms (-fsanitize=cfi*) and needs to run at link time if CFI
539 // is enabled. The pass does nothing if CFI is disabled.
540 PM.add(createLowerBitSetsPass());
543 addLateLTOOptimizationPasses(PM);
546 PM.add(createVerifierPass());
549 inline PassManagerBuilder *unwrap(LLVMPassManagerBuilderRef P) {
550 return reinterpret_cast<PassManagerBuilder*>(P);
553 inline LLVMPassManagerBuilderRef wrap(PassManagerBuilder *P) {
554 return reinterpret_cast<LLVMPassManagerBuilderRef>(P);
557 LLVMPassManagerBuilderRef LLVMPassManagerBuilderCreate() {
558 PassManagerBuilder *PMB = new PassManagerBuilder();
562 void LLVMPassManagerBuilderDispose(LLVMPassManagerBuilderRef PMB) {
563 PassManagerBuilder *Builder = unwrap(PMB);
568 LLVMPassManagerBuilderSetOptLevel(LLVMPassManagerBuilderRef PMB,
570 PassManagerBuilder *Builder = unwrap(PMB);
571 Builder->OptLevel = OptLevel;
575 LLVMPassManagerBuilderSetSizeLevel(LLVMPassManagerBuilderRef PMB,
576 unsigned SizeLevel) {
577 PassManagerBuilder *Builder = unwrap(PMB);
578 Builder->SizeLevel = SizeLevel;
582 LLVMPassManagerBuilderSetDisableUnitAtATime(LLVMPassManagerBuilderRef PMB,
584 PassManagerBuilder *Builder = unwrap(PMB);
585 Builder->DisableUnitAtATime = Value;
589 LLVMPassManagerBuilderSetDisableUnrollLoops(LLVMPassManagerBuilderRef PMB,
591 PassManagerBuilder *Builder = unwrap(PMB);
592 Builder->DisableUnrollLoops = Value;
596 LLVMPassManagerBuilderSetDisableSimplifyLibCalls(LLVMPassManagerBuilderRef PMB,
598 // NOTE: The simplify-libcalls pass has been removed.
602 LLVMPassManagerBuilderUseInlinerWithThreshold(LLVMPassManagerBuilderRef PMB,
603 unsigned Threshold) {
604 PassManagerBuilder *Builder = unwrap(PMB);
605 Builder->Inliner = createFunctionInliningPass(Threshold);
609 LLVMPassManagerBuilderPopulateFunctionPassManager(LLVMPassManagerBuilderRef PMB,
610 LLVMPassManagerRef PM) {
611 PassManagerBuilder *Builder = unwrap(PMB);
612 legacy::FunctionPassManager *FPM = unwrap<legacy::FunctionPassManager>(PM);
613 Builder->populateFunctionPassManager(*FPM);
617 LLVMPassManagerBuilderPopulateModulePassManager(LLVMPassManagerBuilderRef PMB,
618 LLVMPassManagerRef PM) {
619 PassManagerBuilder *Builder = unwrap(PMB);
620 legacy::PassManagerBase *MPM = unwrap(PM);
621 Builder->populateModulePassManager(*MPM);
624 void LLVMPassManagerBuilderPopulateLTOPassManager(LLVMPassManagerBuilderRef PMB,
625 LLVMPassManagerRef PM,
626 LLVMBool Internalize,
627 LLVMBool RunInliner) {
628 PassManagerBuilder *Builder = unwrap(PMB);
629 legacy::PassManagerBase *LPM = unwrap(PM);
631 // A small backwards compatibility hack. populateLTOPassManager used to take
632 // an RunInliner option.
633 if (RunInliner && !Builder->Inliner)
634 Builder->Inliner = createFunctionInliningPass();
636 Builder->populateLTOPassManager(*LPM);