1 //===- MergeFunctions.cpp - Merge identical functions ---------------------===//
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 pass looks for equivalent functions that are mergable and folds them.
12 // A hash is computed from the function, based on its type and number of
15 // Once all hashes are computed, we perform an expensive equality comparison
16 // on each function pair. This takes n^2/2 comparisons per bucket, so it's
17 // important that the hash function be high quality. The equality comparison
18 // iterates through each instruction in each basic block.
20 // When a match is found the functions are folded. If both functions are
21 // overridable, we move the functionality into a new internal function and
22 // leave two overridable thunks to it.
24 //===----------------------------------------------------------------------===//
28 // * virtual functions.
30 // Many functions have their address taken by the virtual function table for
31 // the object they belong to. However, as long as it's only used for a lookup
32 // and call, this is irrelevant, and we'd like to fold such functions.
34 // * switch from n^2 pair-wise comparisons to an n-way comparison for each
37 // * be smarter about bitcasts.
39 // In order to fold functions, we will sometimes add either bitcast instructions
40 // or bitcast constant expressions. Unfortunately, this can confound further
41 // analysis since the two functions differ where one has a bitcast and the
42 // other doesn't. We should learn to look through bitcasts.
44 //===----------------------------------------------------------------------===//
46 #define DEBUG_TYPE "mergefunc"
47 #include "llvm/Transforms/IPO.h"
48 #include "llvm/ADT/DenseSet.h"
49 #include "llvm/ADT/FoldingSet.h"
50 #include "llvm/ADT/SmallSet.h"
51 #include "llvm/ADT/Statistic.h"
52 #include "llvm/ADT/STLExtras.h"
53 #include "llvm/Constants.h"
54 #include "llvm/InlineAsm.h"
55 #include "llvm/Instructions.h"
56 #include "llvm/LLVMContext.h"
57 #include "llvm/Module.h"
58 #include "llvm/Pass.h"
59 #include "llvm/Support/CallSite.h"
60 #include "llvm/Support/Debug.h"
61 #include "llvm/Support/ErrorHandling.h"
62 #include "llvm/Support/IRBuilder.h"
63 #include "llvm/Support/ValueHandle.h"
64 #include "llvm/Support/raw_ostream.h"
65 #include "llvm/Target/TargetData.h"
69 STATISTIC(NumFunctionsMerged, "Number of functions merged");
70 STATISTIC(NumThunksWritten, "Number of thunks generated");
71 STATISTIC(NumAliasesWritten, "Number of aliases generated");
72 STATISTIC(NumDoubleWeak, "Number of new functions created");
74 /// ProfileFunction - Creates a hash-code for the function which is the same
75 /// for any two functions that will compare equal, without looking at the
76 /// instructions inside the function.
77 static unsigned ProfileFunction(const Function *F) {
78 const FunctionType *FTy = F->getFunctionType();
81 ID.AddInteger(F->size());
82 ID.AddInteger(F->getCallingConv());
83 ID.AddBoolean(F->hasGC());
84 ID.AddBoolean(FTy->isVarArg());
85 ID.AddInteger(FTy->getReturnType()->getTypeID());
86 for (unsigned i = 0, e = FTy->getNumParams(); i != e; ++i)
87 ID.AddInteger(FTy->getParamType(i)->getTypeID());
88 return ID.ComputeHash();
93 class ComparableFunction {
95 static const ComparableFunction EmptyKey;
96 static const ComparableFunction TombstoneKey;
98 ComparableFunction(Function *Func, TargetData *TD)
99 : Func(Func), Hash(ProfileFunction(Func)), TD(TD) {}
101 Function *getFunc() const { return Func; }
102 unsigned getHash() const { return Hash; }
103 TargetData *getTD() const { return TD; }
105 // Drops AssertingVH reference to the function. Outside of debug mode, this
109 "Attempted to release function twice, or release empty/tombstone!");
113 bool &getOrInsertCachedComparison(const ComparableFunction &Other,
114 bool &inserted) const {
115 typedef DenseMap<Function *, bool>::iterator iterator;
116 std::pair<iterator, bool> p =
117 CompareResultCache.insert(std::make_pair(Other.getFunc(), false));
119 return p.first->second;
123 explicit ComparableFunction(unsigned Hash)
124 : Func(NULL), Hash(Hash), TD(NULL) {}
126 // DenseMap::grow() triggers a recomparison of all keys in the map, which is
127 // wildly expensive. This cache tries to preserve known results.
128 mutable DenseMap<Function *, bool> CompareResultCache;
130 AssertingVH<Function> Func;
135 const ComparableFunction ComparableFunction::EmptyKey = ComparableFunction(0);
136 const ComparableFunction ComparableFunction::TombstoneKey =
137 ComparableFunction(1);
143 struct DenseMapInfo<ComparableFunction> {
144 static ComparableFunction getEmptyKey() {
145 return ComparableFunction::EmptyKey;
147 static ComparableFunction getTombstoneKey() {
148 return ComparableFunction::TombstoneKey;
150 static unsigned getHashValue(const ComparableFunction &CF) {
153 static bool isEqual(const ComparableFunction &LHS,
154 const ComparableFunction &RHS);
160 /// MergeFunctions finds functions which will generate identical machine code,
161 /// by considering all pointer types to be equivalent. Once identified,
162 /// MergeFunctions will fold them by replacing a call to one to a call to a
163 /// bitcast of the other.
165 class MergeFunctions : public ModulePass {
169 : ModulePass(ID), HasGlobalAliases(false) {
170 initializeMergeFunctionsPass(*PassRegistry::getPassRegistry());
173 bool runOnModule(Module &M);
176 typedef DenseSet<ComparableFunction> FnSetType;
178 /// A work queue of functions that may have been modified and should be
180 std::vector<WeakVH> Deferred;
182 /// Insert a ComparableFunction into the FnSet, or merge it away if it's
183 /// equal to one that's already present.
184 bool Insert(ComparableFunction &NewF);
186 /// Remove a Function from the FnSet and queue it up for a second sweep of
188 void Remove(Function *F);
190 /// Find the functions that use this Value and remove them from FnSet and
191 /// queue the functions.
192 void RemoveUsers(Value *V);
194 /// Replace all direct calls of Old with calls of New. Will bitcast New if
195 /// necessary to make types match.
196 void replaceDirectCallers(Function *Old, Function *New);
198 /// MergeTwoFunctions - Merge two equivalent functions. Upon completion, G
199 /// may be deleted, or may be converted into a thunk. In either case, it
200 /// should never be visited again.
201 void MergeTwoFunctions(Function *F, Function *G);
203 /// WriteThunkOrAlias - Replace G with a thunk or an alias to F. Deletes G.
204 void WriteThunkOrAlias(Function *F, Function *G);
206 /// WriteThunk - Replace G with a simple tail call to bitcast(F). Also
207 /// replace direct uses of G with bitcast(F). Deletes G.
208 void WriteThunk(Function *F, Function *G);
210 /// WriteAlias - Replace G with an alias to F. Deletes G.
211 void WriteAlias(Function *F, Function *G);
213 /// The set of all distinct functions. Use the Insert and Remove methods to
217 /// TargetData for more accurate GEP comparisons. May be NULL.
220 /// Whether or not the target supports global aliases.
221 bool HasGlobalAliases;
224 } // end anonymous namespace
226 char MergeFunctions::ID = 0;
227 INITIALIZE_PASS(MergeFunctions, "mergefunc", "Merge Functions", false, false)
229 ModulePass *llvm::createMergeFunctionsPass() {
230 return new MergeFunctions();
234 /// FunctionComparator - Compares two functions to determine whether or not
235 /// they will generate machine code with the same behaviour. TargetData is
236 /// used if available. The comparator always fails conservatively (erring on the
237 /// side of claiming that two functions are different).
238 class FunctionComparator {
240 FunctionComparator(const TargetData *TD, const Function *F1,
242 : F1(F1), F2(F2), TD(TD), IDMap1Count(0), IDMap2Count(0) {}
244 /// Compare - test whether the two functions have equivalent behaviour.
248 /// Compare - test whether two basic blocks have equivalent behaviour.
249 bool Compare(const BasicBlock *BB1, const BasicBlock *BB2);
251 /// Enumerate - Assign or look up previously assigned numbers for the two
252 /// values, and return whether the numbers are equal. Numbers are assigned in
253 /// the order visited.
254 bool Enumerate(const Value *V1, const Value *V2);
256 /// isEquivalentOperation - Compare two Instructions for equivalence, similar
257 /// to Instruction::isSameOperationAs but with modifications to the type
259 bool isEquivalentOperation(const Instruction *I1,
260 const Instruction *I2) const;
262 /// isEquivalentGEP - Compare two GEPs for equivalent pointer arithmetic.
263 bool isEquivalentGEP(const GEPOperator *GEP1, const GEPOperator *GEP2);
264 bool isEquivalentGEP(const GetElementPtrInst *GEP1,
265 const GetElementPtrInst *GEP2) {
266 return isEquivalentGEP(cast<GEPOperator>(GEP1), cast<GEPOperator>(GEP2));
269 /// isEquivalentType - Compare two Types, treating all pointer types as equal.
270 bool isEquivalentType(const Type *Ty1, const Type *Ty2) const;
272 // The two functions undergoing comparison.
273 const Function *F1, *F2;
275 const TargetData *TD;
277 typedef DenseMap<const Value *, unsigned long> IDMap;
279 unsigned long IDMap1Count, IDMap2Count;
283 /// isEquivalentType - any two pointers in the same address space are
284 /// equivalent. Otherwise, standard type equivalence rules apply.
285 bool FunctionComparator::isEquivalentType(const Type *Ty1,
286 const Type *Ty2) const {
289 if (Ty1->getTypeID() != Ty2->getTypeID())
292 switch(Ty1->getTypeID()) {
294 llvm_unreachable("Unknown type!");
295 // Fall through in Release mode.
296 case Type::IntegerTyID:
297 case Type::OpaqueTyID:
298 case Type::VectorTyID:
299 // Ty1 == Ty2 would have returned true earlier.
303 case Type::FloatTyID:
304 case Type::DoubleTyID:
305 case Type::X86_FP80TyID:
306 case Type::FP128TyID:
307 case Type::PPC_FP128TyID:
308 case Type::LabelTyID:
309 case Type::MetadataTyID:
312 case Type::PointerTyID: {
313 const PointerType *PTy1 = cast<PointerType>(Ty1);
314 const PointerType *PTy2 = cast<PointerType>(Ty2);
315 return PTy1->getAddressSpace() == PTy2->getAddressSpace();
318 case Type::StructTyID: {
319 const StructType *STy1 = cast<StructType>(Ty1);
320 const StructType *STy2 = cast<StructType>(Ty2);
321 if (STy1->getNumElements() != STy2->getNumElements())
324 if (STy1->isPacked() != STy2->isPacked())
327 for (unsigned i = 0, e = STy1->getNumElements(); i != e; ++i) {
328 if (!isEquivalentType(STy1->getElementType(i), STy2->getElementType(i)))
334 case Type::FunctionTyID: {
335 const FunctionType *FTy1 = cast<FunctionType>(Ty1);
336 const FunctionType *FTy2 = cast<FunctionType>(Ty2);
337 if (FTy1->getNumParams() != FTy2->getNumParams() ||
338 FTy1->isVarArg() != FTy2->isVarArg())
341 if (!isEquivalentType(FTy1->getReturnType(), FTy2->getReturnType()))
344 for (unsigned i = 0, e = FTy1->getNumParams(); i != e; ++i) {
345 if (!isEquivalentType(FTy1->getParamType(i), FTy2->getParamType(i)))
351 case Type::ArrayTyID: {
352 const ArrayType *ATy1 = cast<ArrayType>(Ty1);
353 const ArrayType *ATy2 = cast<ArrayType>(Ty2);
354 return ATy1->getNumElements() == ATy2->getNumElements() &&
355 isEquivalentType(ATy1->getElementType(), ATy2->getElementType());
360 /// isEquivalentOperation - determine whether the two operations are the same
361 /// except that pointer-to-A and pointer-to-B are equivalent. This should be
362 /// kept in sync with Instruction::isSameOperationAs.
363 bool FunctionComparator::isEquivalentOperation(const Instruction *I1,
364 const Instruction *I2) const {
365 if (I1->getOpcode() != I2->getOpcode() ||
366 I1->getNumOperands() != I2->getNumOperands() ||
367 !isEquivalentType(I1->getType(), I2->getType()) ||
368 !I1->hasSameSubclassOptionalData(I2))
371 // We have two instructions of identical opcode and #operands. Check to see
372 // if all operands are the same type
373 for (unsigned i = 0, e = I1->getNumOperands(); i != e; ++i)
374 if (!isEquivalentType(I1->getOperand(i)->getType(),
375 I2->getOperand(i)->getType()))
378 // Check special state that is a part of some instructions.
379 if (const LoadInst *LI = dyn_cast<LoadInst>(I1))
380 return LI->isVolatile() == cast<LoadInst>(I2)->isVolatile() &&
381 LI->getAlignment() == cast<LoadInst>(I2)->getAlignment();
382 if (const StoreInst *SI = dyn_cast<StoreInst>(I1))
383 return SI->isVolatile() == cast<StoreInst>(I2)->isVolatile() &&
384 SI->getAlignment() == cast<StoreInst>(I2)->getAlignment();
385 if (const CmpInst *CI = dyn_cast<CmpInst>(I1))
386 return CI->getPredicate() == cast<CmpInst>(I2)->getPredicate();
387 if (const CallInst *CI = dyn_cast<CallInst>(I1))
388 return CI->isTailCall() == cast<CallInst>(I2)->isTailCall() &&
389 CI->getCallingConv() == cast<CallInst>(I2)->getCallingConv() &&
390 CI->getAttributes().getRawPointer() ==
391 cast<CallInst>(I2)->getAttributes().getRawPointer();
392 if (const InvokeInst *CI = dyn_cast<InvokeInst>(I1))
393 return CI->getCallingConv() == cast<InvokeInst>(I2)->getCallingConv() &&
394 CI->getAttributes().getRawPointer() ==
395 cast<InvokeInst>(I2)->getAttributes().getRawPointer();
396 if (const InsertValueInst *IVI = dyn_cast<InsertValueInst>(I1)) {
397 if (IVI->getNumIndices() != cast<InsertValueInst>(I2)->getNumIndices())
399 for (unsigned i = 0, e = IVI->getNumIndices(); i != e; ++i)
400 if (IVI->idx_begin()[i] != cast<InsertValueInst>(I2)->idx_begin()[i])
404 if (const ExtractValueInst *EVI = dyn_cast<ExtractValueInst>(I1)) {
405 if (EVI->getNumIndices() != cast<ExtractValueInst>(I2)->getNumIndices())
407 for (unsigned i = 0, e = EVI->getNumIndices(); i != e; ++i)
408 if (EVI->idx_begin()[i] != cast<ExtractValueInst>(I2)->idx_begin()[i])
416 /// isEquivalentGEP - determine whether two GEP operations perform the same
417 /// underlying arithmetic.
418 bool FunctionComparator::isEquivalentGEP(const GEPOperator *GEP1,
419 const GEPOperator *GEP2) {
420 // When we have target data, we can reduce the GEP down to the value in bytes
421 // added to the address.
422 if (TD && GEP1->hasAllConstantIndices() && GEP2->hasAllConstantIndices()) {
423 SmallVector<Value *, 8> Indices1(GEP1->idx_begin(), GEP1->idx_end());
424 SmallVector<Value *, 8> Indices2(GEP2->idx_begin(), GEP2->idx_end());
425 uint64_t Offset1 = TD->getIndexedOffset(GEP1->getPointerOperandType(),
426 Indices1.data(), Indices1.size());
427 uint64_t Offset2 = TD->getIndexedOffset(GEP2->getPointerOperandType(),
428 Indices2.data(), Indices2.size());
429 return Offset1 == Offset2;
432 if (GEP1->getPointerOperand()->getType() !=
433 GEP2->getPointerOperand()->getType())
436 if (GEP1->getNumOperands() != GEP2->getNumOperands())
439 for (unsigned i = 0, e = GEP1->getNumOperands(); i != e; ++i) {
440 if (!Enumerate(GEP1->getOperand(i), GEP2->getOperand(i)))
447 /// Enumerate - Compare two values used by the two functions under pair-wise
448 /// comparison. If this is the first time the values are seen, they're added to
449 /// the mapping so that we will detect mismatches on next use.
450 bool FunctionComparator::Enumerate(const Value *V1, const Value *V2) {
451 // Check for function @f1 referring to itself and function @f2 referring to
452 // itself, or referring to each other, or both referring to either of them.
453 // They're all equivalent if the two functions are otherwise equivalent.
454 if (V1 == F1 && V2 == F2)
456 if (V1 == F2 && V2 == F1)
459 // TODO: constant expressions with GEP or references to F1 or F2.
460 if (isa<Constant>(V1))
463 if (isa<InlineAsm>(V1) && isa<InlineAsm>(V2)) {
464 const InlineAsm *IA1 = cast<InlineAsm>(V1);
465 const InlineAsm *IA2 = cast<InlineAsm>(V2);
466 return IA1->getAsmString() == IA2->getAsmString() &&
467 IA1->getConstraintString() == IA2->getConstraintString();
470 unsigned long &ID1 = Map1[V1];
474 unsigned long &ID2 = Map2[V2];
481 /// Compare - test whether two basic blocks have equivalent behaviour.
482 bool FunctionComparator::Compare(const BasicBlock *BB1, const BasicBlock *BB2) {
483 BasicBlock::const_iterator F1I = BB1->begin(), F1E = BB1->end();
484 BasicBlock::const_iterator F2I = BB2->begin(), F2E = BB2->end();
487 if (!Enumerate(F1I, F2I))
490 if (const GetElementPtrInst *GEP1 = dyn_cast<GetElementPtrInst>(F1I)) {
491 const GetElementPtrInst *GEP2 = dyn_cast<GetElementPtrInst>(F2I);
495 if (!Enumerate(GEP1->getPointerOperand(), GEP2->getPointerOperand()))
498 if (!isEquivalentGEP(GEP1, GEP2))
501 if (!isEquivalentOperation(F1I, F2I))
504 assert(F1I->getNumOperands() == F2I->getNumOperands());
505 for (unsigned i = 0, e = F1I->getNumOperands(); i != e; ++i) {
506 Value *OpF1 = F1I->getOperand(i);
507 Value *OpF2 = F2I->getOperand(i);
509 if (!Enumerate(OpF1, OpF2))
512 if (OpF1->getValueID() != OpF2->getValueID() ||
513 !isEquivalentType(OpF1->getType(), OpF2->getType()))
519 } while (F1I != F1E && F2I != F2E);
521 return F1I == F1E && F2I == F2E;
524 /// Compare - test whether the two functions have equivalent behaviour.
525 bool FunctionComparator::Compare() {
526 // We need to recheck everything, but check the things that weren't included
527 // in the hash first.
529 if (F1->getAttributes() != F2->getAttributes())
532 if (F1->hasGC() != F2->hasGC())
535 if (F1->hasGC() && F1->getGC() != F2->getGC())
538 if (F1->hasSection() != F2->hasSection())
541 if (F1->hasSection() && F1->getSection() != F2->getSection())
544 if (F1->isVarArg() != F2->isVarArg())
547 // TODO: if it's internal and only used in direct calls, we could handle this
549 if (F1->getCallingConv() != F2->getCallingConv())
552 if (!isEquivalentType(F1->getFunctionType(), F2->getFunctionType()))
555 assert(F1->arg_size() == F2->arg_size() &&
556 "Identically typed functions have different numbers of args!");
558 // Visit the arguments so that they get enumerated in the order they're
560 for (Function::const_arg_iterator f1i = F1->arg_begin(),
561 f2i = F2->arg_begin(), f1e = F1->arg_end(); f1i != f1e; ++f1i, ++f2i) {
562 if (!Enumerate(f1i, f2i))
563 llvm_unreachable("Arguments repeat!");
566 // We do a CFG-ordered walk since the actual ordering of the blocks in the
567 // linked list is immaterial. Our walk starts at the entry block for both
568 // functions, then takes each block from each terminator in order. As an
569 // artifact, this also means that unreachable blocks are ignored.
570 SmallVector<const BasicBlock *, 8> F1BBs, F2BBs;
571 SmallSet<const BasicBlock *, 128> VisitedBBs; // in terms of F1.
573 F1BBs.push_back(&F1->getEntryBlock());
574 F2BBs.push_back(&F2->getEntryBlock());
576 VisitedBBs.insert(F1BBs[0]);
577 while (!F1BBs.empty()) {
578 const BasicBlock *F1BB = F1BBs.pop_back_val();
579 const BasicBlock *F2BB = F2BBs.pop_back_val();
581 if (!Enumerate(F1BB, F2BB) || !Compare(F1BB, F2BB))
584 const TerminatorInst *F1TI = F1BB->getTerminator();
585 const TerminatorInst *F2TI = F2BB->getTerminator();
587 assert(F1TI->getNumSuccessors() == F2TI->getNumSuccessors());
588 for (unsigned i = 0, e = F1TI->getNumSuccessors(); i != e; ++i) {
589 if (!VisitedBBs.insert(F1TI->getSuccessor(i)))
592 F1BBs.push_back(F1TI->getSuccessor(i));
593 F2BBs.push_back(F2TI->getSuccessor(i));
599 /// Replace direct callers of Old with New.
600 void MergeFunctions::replaceDirectCallers(Function *Old, Function *New) {
601 Constant *BitcastNew = ConstantExpr::getBitCast(New, Old->getType());
602 for (Value::use_iterator UI = Old->use_begin(), UE = Old->use_end();
604 Value::use_iterator TheIter = UI;
606 CallSite CS(*TheIter);
607 if (CS && CS.isCallee(TheIter)) {
608 Remove(CS.getInstruction()->getParent()->getParent());
609 TheIter.getUse().set(BitcastNew);
614 void MergeFunctions::WriteThunkOrAlias(Function *F, Function *G) {
615 if (HasGlobalAliases && G->hasUnnamedAddr()) {
616 if (G->hasExternalLinkage() || G->hasLocalLinkage() ||
617 G->hasWeakLinkage()) {
626 /// WriteThunk - Replace G with a simple tail call to bitcast(F). Also replace
627 /// direct uses of G with bitcast(F). Deletes G.
628 void MergeFunctions::WriteThunk(Function *F, Function *G) {
629 if (!G->mayBeOverridden()) {
630 // Redirect direct callers of G to F.
631 replaceDirectCallers(G, F);
634 // If G was internal then we may have replaced all uses of G with F. If so,
635 // stop here and delete G. There's no need for a thunk.
636 if (G->hasLocalLinkage() && G->use_empty()) {
637 G->eraseFromParent();
641 Function *NewG = Function::Create(G->getFunctionType(), G->getLinkage(), "",
643 BasicBlock *BB = BasicBlock::Create(F->getContext(), "", NewG);
644 IRBuilder<false> Builder(BB);
646 SmallVector<Value *, 16> Args;
648 const FunctionType *FFTy = F->getFunctionType();
649 for (Function::arg_iterator AI = NewG->arg_begin(), AE = NewG->arg_end();
651 Args.push_back(Builder.CreateBitCast(AI, FFTy->getParamType(i)));
655 CallInst *CI = Builder.CreateCall(F, Args.begin(), Args.end());
657 CI->setCallingConv(F->getCallingConv());
658 if (NewG->getReturnType()->isVoidTy()) {
659 Builder.CreateRetVoid();
661 Builder.CreateRet(Builder.CreateBitCast(CI, NewG->getReturnType()));
664 NewG->copyAttributesFrom(G);
667 G->replaceAllUsesWith(NewG);
668 G->eraseFromParent();
670 DEBUG(dbgs() << "WriteThunk: " << NewG->getName() << '\n');
674 /// WriteAlias - Replace G with an alias to F and delete G.
675 void MergeFunctions::WriteAlias(Function *F, Function *G) {
676 Constant *BitcastF = ConstantExpr::getBitCast(F, G->getType());
677 GlobalAlias *GA = new GlobalAlias(G->getType(), G->getLinkage(), "",
678 BitcastF, G->getParent());
679 F->setAlignment(std::max(F->getAlignment(), G->getAlignment()));
681 GA->setVisibility(G->getVisibility());
683 G->replaceAllUsesWith(GA);
684 G->eraseFromParent();
686 DEBUG(dbgs() << "WriteAlias: " << GA->getName() << '\n');
690 /// MergeTwoFunctions - Merge two equivalent functions. Upon completion,
691 /// Function G is deleted.
692 void MergeFunctions::MergeTwoFunctions(Function *F, Function *G) {
693 if (F->mayBeOverridden()) {
694 assert(G->mayBeOverridden());
696 if (HasGlobalAliases) {
697 // Make them both thunks to the same internal function.
698 Function *H = Function::Create(F->getFunctionType(), F->getLinkage(), "",
700 H->copyAttributesFrom(F);
703 F->replaceAllUsesWith(H);
705 unsigned MaxAlignment = std::max(G->getAlignment(), H->getAlignment());
710 F->setAlignment(MaxAlignment);
711 F->setLinkage(GlobalValue::PrivateLinkage);
713 // We can't merge them. Instead, pick one and update all direct callers
714 // to call it and hope that we improve the instruction cache hit rate.
715 replaceDirectCallers(G, F);
720 WriteThunkOrAlias(F, G);
723 ++NumFunctionsMerged;
726 // Insert - Insert a ComparableFunction into the FnSet, or merge it away if
727 // equal to one that's already inserted.
728 bool MergeFunctions::Insert(ComparableFunction &NewF) {
729 std::pair<FnSetType::iterator, bool> Result = FnSet.insert(NewF);
733 const ComparableFunction &OldF = *Result.first;
735 // Never thunk a strong function to a weak function.
736 assert(!OldF.getFunc()->mayBeOverridden() ||
737 NewF.getFunc()->mayBeOverridden());
739 DEBUG(dbgs() << " " << OldF.getFunc()->getName() << " == "
740 << NewF.getFunc()->getName() << '\n');
742 Function *DeleteF = NewF.getFunc();
744 MergeTwoFunctions(OldF.getFunc(), DeleteF);
748 // Remove - Remove a function from FnSet. If it was already in FnSet, add it to
749 // Deferred so that we'll look at it in the next round.
750 void MergeFunctions::Remove(Function *F) {
751 ComparableFunction CF = ComparableFunction(F, TD);
752 if (FnSet.erase(CF)) {
753 Deferred.push_back(F);
757 // RemoveUsers - For each instruction used by the value, Remove() the function
758 // that contains the instruction. This should happen right before a call to RAUW.
759 void MergeFunctions::RemoveUsers(Value *V) {
760 std::vector<Value *> Worklist;
761 Worklist.push_back(V);
762 while (!Worklist.empty()) {
763 Value *V = Worklist.back();
766 for (Value::use_iterator UI = V->use_begin(), UE = V->use_end();
768 Use &U = UI.getUse();
769 if (Instruction *I = dyn_cast<Instruction>(U.getUser())) {
770 Remove(I->getParent()->getParent());
771 } else if (isa<GlobalValue>(U.getUser())) {
773 } else if (Constant *C = dyn_cast<Constant>(U.getUser())) {
774 for (Value::use_iterator CUI = C->use_begin(), CUE = C->use_end();
776 Worklist.push_back(*CUI);
782 bool MergeFunctions::runOnModule(Module &M) {
783 bool Changed = false;
784 TD = getAnalysisIfAvailable<TargetData>();
786 for (Module::iterator I = M.begin(), E = M.end(); I != E; ++I) {
787 Deferred.push_back(WeakVH(I));
791 std::vector<WeakVH> Worklist;
792 Deferred.swap(Worklist);
794 DEBUG(dbgs() << "size of module: " << M.size() << '\n');
795 DEBUG(dbgs() << "size of worklist: " << Worklist.size() << '\n');
797 // Insert only strong functions and merge them. Strong function merging
798 // always deletes one of them.
799 for (std::vector<WeakVH>::iterator I = Worklist.begin(),
800 E = Worklist.end(); I != E; ++I) {
802 Function *F = cast<Function>(*I);
803 if (!F->isDeclaration() && !F->hasAvailableExternallyLinkage() &&
804 !F->mayBeOverridden()) {
805 ComparableFunction CF = ComparableFunction(F, TD);
806 Changed |= Insert(CF);
810 // Insert only weak functions and merge them. By doing these second we
811 // create thunks to the strong function when possible. When two weak
812 // functions are identical, we create a new strong function with two weak
813 // weak thunks to it which are identical but not mergable.
814 for (std::vector<WeakVH>::iterator I = Worklist.begin(),
815 E = Worklist.end(); I != E; ++I) {
817 Function *F = cast<Function>(*I);
818 if (!F->isDeclaration() && !F->hasAvailableExternallyLinkage() &&
819 F->mayBeOverridden()) {
820 ComparableFunction CF = ComparableFunction(F, TD);
821 Changed |= Insert(CF);
824 DEBUG(dbgs() << "size of FnSet: " << FnSet.size() << '\n');
825 } while (!Deferred.empty());
832 bool DenseMapInfo<ComparableFunction>::isEqual(const ComparableFunction &LHS,
833 const ComparableFunction &RHS) {
834 if (LHS.getFunc() == RHS.getFunc() &&
835 LHS.getHash() == RHS.getHash())
837 if (!LHS.getFunc() || !RHS.getFunc())
839 assert(LHS.getTD() == RHS.getTD() &&
840 "Comparing functions for different targets");
843 bool &result1 = LHS.getOrInsertCachedComparison(RHS, inserted);
846 bool &result2 = RHS.getOrInsertCachedComparison(LHS, inserted);
848 return result1 = result2;
850 return result1 = result2 = FunctionComparator(LHS.getTD(), LHS.getFunc(),
851 RHS.getFunc()).Compare();