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 /// FunctionComparator - Compares two functions to determine whether or not
161 /// they will generate machine code with the same behaviour. TargetData is
162 /// used if available. The comparator always fails conservatively (erring on the
163 /// side of claiming that two functions are different).
164 class FunctionComparator {
166 FunctionComparator(const TargetData *TD, const Function *F1,
168 : F1(F1), F2(F2), TD(TD), IDMap1Count(0), IDMap2Count(0) {}
170 /// Compare - test whether the two functions have equivalent behaviour.
174 /// Compare - test whether two basic blocks have equivalent behaviour.
175 bool Compare(const BasicBlock *BB1, const BasicBlock *BB2);
177 /// Enumerate - Assign or look up previously assigned numbers for the two
178 /// values, and return whether the numbers are equal. Numbers are assigned in
179 /// the order visited.
180 bool Enumerate(const Value *V1, const Value *V2);
182 /// isEquivalentOperation - Compare two Instructions for equivalence, similar
183 /// to Instruction::isSameOperationAs but with modifications to the type
185 bool isEquivalentOperation(const Instruction *I1,
186 const Instruction *I2) const;
188 /// isEquivalentGEP - Compare two GEPs for equivalent pointer arithmetic.
189 bool isEquivalentGEP(const GEPOperator *GEP1, const GEPOperator *GEP2);
190 bool isEquivalentGEP(const GetElementPtrInst *GEP1,
191 const GetElementPtrInst *GEP2) {
192 return isEquivalentGEP(cast<GEPOperator>(GEP1), cast<GEPOperator>(GEP2));
195 /// isEquivalentType - Compare two Types, treating all pointer types as equal.
196 bool isEquivalentType(const Type *Ty1, const Type *Ty2) const;
198 // The two functions undergoing comparison.
199 const Function *F1, *F2;
201 const TargetData *TD;
203 typedef DenseMap<const Value *, unsigned long> IDMap;
205 unsigned long IDMap1Count, IDMap2Count;
210 /// isEquivalentType - any two pointers in the same address space are
211 /// equivalent. Otherwise, standard type equivalence rules apply.
212 bool FunctionComparator::isEquivalentType(const Type *Ty1,
213 const Type *Ty2) const {
216 if (Ty1->getTypeID() != Ty2->getTypeID()) {
218 LLVMContext &Ctx = Ty1->getContext();
219 if (isa<PointerType>(Ty1) && Ty2 == TD->getIntPtrType(Ctx)) return true;
220 if (isa<PointerType>(Ty2) && Ty1 == TD->getIntPtrType(Ctx)) return true;
225 switch(Ty1->getTypeID()) {
227 llvm_unreachable("Unknown type!");
228 // Fall through in Release mode.
229 case Type::IntegerTyID:
230 case Type::OpaqueTyID:
231 case Type::VectorTyID:
232 // Ty1 == Ty2 would have returned true earlier.
236 case Type::FloatTyID:
237 case Type::DoubleTyID:
238 case Type::X86_FP80TyID:
239 case Type::FP128TyID:
240 case Type::PPC_FP128TyID:
241 case Type::LabelTyID:
242 case Type::MetadataTyID:
245 case Type::PointerTyID: {
246 const PointerType *PTy1 = cast<PointerType>(Ty1);
247 const PointerType *PTy2 = cast<PointerType>(Ty2);
248 return PTy1->getAddressSpace() == PTy2->getAddressSpace();
251 case Type::StructTyID: {
252 const StructType *STy1 = cast<StructType>(Ty1);
253 const StructType *STy2 = cast<StructType>(Ty2);
254 if (STy1->getNumElements() != STy2->getNumElements())
257 if (STy1->isPacked() != STy2->isPacked())
260 for (unsigned i = 0, e = STy1->getNumElements(); i != e; ++i) {
261 if (!isEquivalentType(STy1->getElementType(i), STy2->getElementType(i)))
267 case Type::FunctionTyID: {
268 const FunctionType *FTy1 = cast<FunctionType>(Ty1);
269 const FunctionType *FTy2 = cast<FunctionType>(Ty2);
270 if (FTy1->getNumParams() != FTy2->getNumParams() ||
271 FTy1->isVarArg() != FTy2->isVarArg())
274 if (!isEquivalentType(FTy1->getReturnType(), FTy2->getReturnType()))
277 for (unsigned i = 0, e = FTy1->getNumParams(); i != e; ++i) {
278 if (!isEquivalentType(FTy1->getParamType(i), FTy2->getParamType(i)))
284 case Type::ArrayTyID: {
285 const ArrayType *ATy1 = cast<ArrayType>(Ty1);
286 const ArrayType *ATy2 = cast<ArrayType>(Ty2);
287 return ATy1->getNumElements() == ATy2->getNumElements() &&
288 isEquivalentType(ATy1->getElementType(), ATy2->getElementType());
293 /// isEquivalentOperation - determine whether the two operations are the same
294 /// except that pointer-to-A and pointer-to-B are equivalent. This should be
295 /// kept in sync with Instruction::isSameOperationAs.
296 bool FunctionComparator::isEquivalentOperation(const Instruction *I1,
297 const Instruction *I2) const {
298 if (I1->getOpcode() != I2->getOpcode() ||
299 I1->getNumOperands() != I2->getNumOperands() ||
300 !isEquivalentType(I1->getType(), I2->getType()) ||
301 !I1->hasSameSubclassOptionalData(I2))
304 // We have two instructions of identical opcode and #operands. Check to see
305 // if all operands are the same type
306 for (unsigned i = 0, e = I1->getNumOperands(); i != e; ++i)
307 if (!isEquivalentType(I1->getOperand(i)->getType(),
308 I2->getOperand(i)->getType()))
311 // Check special state that is a part of some instructions.
312 if (const LoadInst *LI = dyn_cast<LoadInst>(I1))
313 return LI->isVolatile() == cast<LoadInst>(I2)->isVolatile() &&
314 LI->getAlignment() == cast<LoadInst>(I2)->getAlignment();
315 if (const StoreInst *SI = dyn_cast<StoreInst>(I1))
316 return SI->isVolatile() == cast<StoreInst>(I2)->isVolatile() &&
317 SI->getAlignment() == cast<StoreInst>(I2)->getAlignment();
318 if (const CmpInst *CI = dyn_cast<CmpInst>(I1))
319 return CI->getPredicate() == cast<CmpInst>(I2)->getPredicate();
320 if (const CallInst *CI = dyn_cast<CallInst>(I1))
321 return CI->isTailCall() == cast<CallInst>(I2)->isTailCall() &&
322 CI->getCallingConv() == cast<CallInst>(I2)->getCallingConv() &&
323 CI->getAttributes() == cast<CallInst>(I2)->getAttributes();
324 if (const InvokeInst *CI = dyn_cast<InvokeInst>(I1))
325 return CI->getCallingConv() == cast<InvokeInst>(I2)->getCallingConv() &&
326 CI->getAttributes() == cast<InvokeInst>(I2)->getAttributes();
327 if (const InsertValueInst *IVI = dyn_cast<InsertValueInst>(I1)) {
328 if (IVI->getNumIndices() != cast<InsertValueInst>(I2)->getNumIndices())
330 for (unsigned i = 0, e = IVI->getNumIndices(); i != e; ++i)
331 if (IVI->idx_begin()[i] != cast<InsertValueInst>(I2)->idx_begin()[i])
335 if (const ExtractValueInst *EVI = dyn_cast<ExtractValueInst>(I1)) {
336 if (EVI->getNumIndices() != cast<ExtractValueInst>(I2)->getNumIndices())
338 for (unsigned i = 0, e = EVI->getNumIndices(); i != e; ++i)
339 if (EVI->idx_begin()[i] != cast<ExtractValueInst>(I2)->idx_begin()[i])
347 /// isEquivalentGEP - determine whether two GEP operations perform the same
348 /// underlying arithmetic.
349 bool FunctionComparator::isEquivalentGEP(const GEPOperator *GEP1,
350 const GEPOperator *GEP2) {
351 // When we have target data, we can reduce the GEP down to the value in bytes
352 // added to the address.
353 if (TD && GEP1->hasAllConstantIndices() && GEP2->hasAllConstantIndices()) {
354 SmallVector<Value *, 8> Indices1(GEP1->idx_begin(), GEP1->idx_end());
355 SmallVector<Value *, 8> Indices2(GEP2->idx_begin(), GEP2->idx_end());
356 uint64_t Offset1 = TD->getIndexedOffset(GEP1->getPointerOperandType(),
357 Indices1.data(), Indices1.size());
358 uint64_t Offset2 = TD->getIndexedOffset(GEP2->getPointerOperandType(),
359 Indices2.data(), Indices2.size());
360 return Offset1 == Offset2;
363 if (GEP1->getPointerOperand()->getType() !=
364 GEP2->getPointerOperand()->getType())
367 if (GEP1->getNumOperands() != GEP2->getNumOperands())
370 for (unsigned i = 0, e = GEP1->getNumOperands(); i != e; ++i) {
371 if (!Enumerate(GEP1->getOperand(i), GEP2->getOperand(i)))
378 /// Enumerate - Compare two values used by the two functions under pair-wise
379 /// comparison. If this is the first time the values are seen, they're added to
380 /// the mapping so that we will detect mismatches on next use.
381 bool FunctionComparator::Enumerate(const Value *V1, const Value *V2) {
382 // Check for function @f1 referring to itself and function @f2 referring to
383 // itself, or referring to each other, or both referring to either of them.
384 // They're all equivalent if the two functions are otherwise equivalent.
385 if (V1 == F1 && V2 == F2)
387 if (V1 == F2 && V2 == F1)
390 if (const Constant *C1 = dyn_cast<Constant>(V1)) {
391 if (V1 == V2) return true;
392 const Constant *C2 = dyn_cast<Constant>(V2);
393 if (!C2) return false;
394 // TODO: constant expressions with GEP or references to F1 or F2.
395 if (C1->isNullValue() && C2->isNullValue() &&
396 isEquivalentType(C1->getType(), C2->getType()))
398 // Try bitcasting C2 to C1's type. If the bitcast is legal and returns C1
399 // then they must have equal bit patterns.
400 return C1->getType()->canLosslesslyBitCastTo(C2->getType()) &&
401 C1 == ConstantExpr::getBitCast(const_cast<Constant*>(C2), C1->getType());
404 if (isa<InlineAsm>(V1) && isa<InlineAsm>(V2)) {
405 const InlineAsm *IA1 = cast<InlineAsm>(V1);
406 const InlineAsm *IA2 = cast<InlineAsm>(V2);
407 return IA1->getAsmString() == IA2->getAsmString() &&
408 IA1->getConstraintString() == IA2->getConstraintString();
411 unsigned long &ID1 = Map1[V1];
415 unsigned long &ID2 = Map2[V2];
422 /// Compare - test whether two basic blocks have equivalent behaviour.
423 bool FunctionComparator::Compare(const BasicBlock *BB1, const BasicBlock *BB2) {
424 BasicBlock::const_iterator F1I = BB1->begin(), F1E = BB1->end();
425 BasicBlock::const_iterator F2I = BB2->begin(), F2E = BB2->end();
428 if (!Enumerate(F1I, F2I))
431 if (const GetElementPtrInst *GEP1 = dyn_cast<GetElementPtrInst>(F1I)) {
432 const GetElementPtrInst *GEP2 = dyn_cast<GetElementPtrInst>(F2I);
436 if (!Enumerate(GEP1->getPointerOperand(), GEP2->getPointerOperand()))
439 if (!isEquivalentGEP(GEP1, GEP2))
442 if (!isEquivalentOperation(F1I, F2I))
445 assert(F1I->getNumOperands() == F2I->getNumOperands());
446 for (unsigned i = 0, e = F1I->getNumOperands(); i != e; ++i) {
447 Value *OpF1 = F1I->getOperand(i);
448 Value *OpF2 = F2I->getOperand(i);
450 if (!Enumerate(OpF1, OpF2))
453 if (OpF1->getValueID() != OpF2->getValueID() ||
454 !isEquivalentType(OpF1->getType(), OpF2->getType()))
460 } while (F1I != F1E && F2I != F2E);
462 return F1I == F1E && F2I == F2E;
465 /// Compare - test whether the two functions have equivalent behaviour.
466 bool FunctionComparator::Compare() {
467 // We need to recheck everything, but check the things that weren't included
468 // in the hash first.
470 if (F1->getAttributes() != F2->getAttributes())
473 if (F1->hasGC() != F2->hasGC())
476 if (F1->hasGC() && F1->getGC() != F2->getGC())
479 if (F1->hasSection() != F2->hasSection())
482 if (F1->hasSection() && F1->getSection() != F2->getSection())
485 if (F1->isVarArg() != F2->isVarArg())
488 // TODO: if it's internal and only used in direct calls, we could handle this
490 if (F1->getCallingConv() != F2->getCallingConv())
493 if (!isEquivalentType(F1->getFunctionType(), F2->getFunctionType()))
496 assert(F1->arg_size() == F2->arg_size() &&
497 "Identically typed functions have different numbers of args!");
499 // Visit the arguments so that they get enumerated in the order they're
501 for (Function::const_arg_iterator f1i = F1->arg_begin(),
502 f2i = F2->arg_begin(), f1e = F1->arg_end(); f1i != f1e; ++f1i, ++f2i) {
503 if (!Enumerate(f1i, f2i))
504 llvm_unreachable("Arguments repeat!");
507 // We do a CFG-ordered walk since the actual ordering of the blocks in the
508 // linked list is immaterial. Our walk starts at the entry block for both
509 // functions, then takes each block from each terminator in order. As an
510 // artifact, this also means that unreachable blocks are ignored.
511 SmallVector<const BasicBlock *, 8> F1BBs, F2BBs;
512 SmallSet<const BasicBlock *, 128> VisitedBBs; // in terms of F1.
514 F1BBs.push_back(&F1->getEntryBlock());
515 F2BBs.push_back(&F2->getEntryBlock());
517 VisitedBBs.insert(F1BBs[0]);
518 while (!F1BBs.empty()) {
519 const BasicBlock *F1BB = F1BBs.pop_back_val();
520 const BasicBlock *F2BB = F2BBs.pop_back_val();
522 if (!Enumerate(F1BB, F2BB) || !Compare(F1BB, F2BB))
525 const TerminatorInst *F1TI = F1BB->getTerminator();
526 const TerminatorInst *F2TI = F2BB->getTerminator();
528 assert(F1TI->getNumSuccessors() == F2TI->getNumSuccessors());
529 for (unsigned i = 0, e = F1TI->getNumSuccessors(); i != e; ++i) {
530 if (!VisitedBBs.insert(F1TI->getSuccessor(i)))
533 F1BBs.push_back(F1TI->getSuccessor(i));
534 F2BBs.push_back(F2TI->getSuccessor(i));
542 /// MergeFunctions finds functions which will generate identical machine code,
543 /// by considering all pointer types to be equivalent. Once identified,
544 /// MergeFunctions will fold them by replacing a call to one to a call to a
545 /// bitcast of the other.
547 class MergeFunctions : public ModulePass {
551 : ModulePass(ID), HasGlobalAliases(false) {
552 initializeMergeFunctionsPass(*PassRegistry::getPassRegistry());
555 bool runOnModule(Module &M);
558 typedef DenseSet<ComparableFunction> FnSetType;
560 /// A work queue of functions that may have been modified and should be
562 std::vector<WeakVH> Deferred;
564 /// Insert a ComparableFunction into the FnSet, or merge it away if it's
565 /// equal to one that's already present.
566 bool Insert(ComparableFunction &NewF);
568 /// Remove a Function from the FnSet and queue it up for a second sweep of
570 void Remove(Function *F);
572 /// Find the functions that use this Value and remove them from FnSet and
573 /// queue the functions.
574 void RemoveUsers(Value *V);
576 /// Replace all direct calls of Old with calls of New. Will bitcast New if
577 /// necessary to make types match.
578 void replaceDirectCallers(Function *Old, Function *New);
580 /// MergeTwoFunctions - Merge two equivalent functions. Upon completion, G
581 /// may be deleted, or may be converted into a thunk. In either case, it
582 /// should never be visited again.
583 void MergeTwoFunctions(Function *F, Function *G);
585 /// WriteThunkOrAlias - Replace G with a thunk or an alias to F. Deletes G.
586 void WriteThunkOrAlias(Function *F, Function *G);
588 /// WriteThunk - Replace G with a simple tail call to bitcast(F). Also
589 /// replace direct uses of G with bitcast(F). Deletes G.
590 void WriteThunk(Function *F, Function *G);
592 /// WriteAlias - Replace G with an alias to F. Deletes G.
593 void WriteAlias(Function *F, Function *G);
595 /// The set of all distinct functions. Use the Insert and Remove methods to
599 /// TargetData for more accurate GEP comparisons. May be NULL.
602 /// Whether or not the target supports global aliases.
603 bool HasGlobalAliases;
606 } // end anonymous namespace
608 char MergeFunctions::ID = 0;
609 INITIALIZE_PASS(MergeFunctions, "mergefunc", "Merge Functions", false, false)
611 ModulePass *llvm::createMergeFunctionsPass() {
612 return new MergeFunctions();
615 bool MergeFunctions::runOnModule(Module &M) {
616 bool Changed = false;
617 TD = getAnalysisIfAvailable<TargetData>();
619 for (Module::iterator I = M.begin(), E = M.end(); I != E; ++I) {
620 if (!I->isDeclaration() && !I->hasAvailableExternallyLinkage())
621 Deferred.push_back(WeakVH(I));
623 FnSet.resize(Deferred.size());
626 std::vector<WeakVH> Worklist;
627 Deferred.swap(Worklist);
629 DEBUG(dbgs() << "size of module: " << M.size() << '\n');
630 DEBUG(dbgs() << "size of worklist: " << Worklist.size() << '\n');
632 // Insert only strong functions and merge them. Strong function merging
633 // always deletes one of them.
634 for (std::vector<WeakVH>::iterator I = Worklist.begin(),
635 E = Worklist.end(); I != E; ++I) {
637 Function *F = cast<Function>(*I);
638 if (!F->isDeclaration() && !F->hasAvailableExternallyLinkage() &&
639 !F->mayBeOverridden()) {
640 ComparableFunction CF = ComparableFunction(F, TD);
641 Changed |= Insert(CF);
645 // Insert only weak functions and merge them. By doing these second we
646 // create thunks to the strong function when possible. When two weak
647 // functions are identical, we create a new strong function with two weak
648 // weak thunks to it which are identical but not mergable.
649 for (std::vector<WeakVH>::iterator I = Worklist.begin(),
650 E = Worklist.end(); I != E; ++I) {
652 Function *F = cast<Function>(*I);
653 if (!F->isDeclaration() && !F->hasAvailableExternallyLinkage() &&
654 F->mayBeOverridden()) {
655 ComparableFunction CF = ComparableFunction(F, TD);
656 Changed |= Insert(CF);
659 DEBUG(dbgs() << "size of FnSet: " << FnSet.size() << '\n');
660 } while (!Deferred.empty());
667 bool DenseMapInfo<ComparableFunction>::isEqual(const ComparableFunction &LHS,
668 const ComparableFunction &RHS) {
669 if (LHS.getFunc() == RHS.getFunc() &&
670 LHS.getHash() == RHS.getHash())
672 if (!LHS.getFunc() || !RHS.getFunc())
674 assert(LHS.getTD() == RHS.getTD() &&
675 "Comparing functions for different targets");
678 bool &result1 = LHS.getOrInsertCachedComparison(RHS, inserted);
681 bool &result2 = RHS.getOrInsertCachedComparison(LHS, inserted);
683 return result1 = result2;
685 return result1 = result2 = FunctionComparator(LHS.getTD(), LHS.getFunc(),
686 RHS.getFunc()).Compare();
689 /// Replace direct callers of Old with New.
690 void MergeFunctions::replaceDirectCallers(Function *Old, Function *New) {
691 Constant *BitcastNew = ConstantExpr::getBitCast(New, Old->getType());
692 for (Value::use_iterator UI = Old->use_begin(), UE = Old->use_end();
694 Value::use_iterator TheIter = UI;
696 CallSite CS(*TheIter);
697 if (CS && CS.isCallee(TheIter)) {
698 Remove(CS.getInstruction()->getParent()->getParent());
699 TheIter.getUse().set(BitcastNew);
704 void MergeFunctions::WriteThunkOrAlias(Function *F, Function *G) {
705 if (HasGlobalAliases && G->hasUnnamedAddr()) {
706 if (G->hasExternalLinkage() || G->hasLocalLinkage() ||
707 G->hasWeakLinkage()) {
716 /// WriteThunk - Replace G with a simple tail call to bitcast(F). Also replace
717 /// direct uses of G with bitcast(F). Deletes G.
718 void MergeFunctions::WriteThunk(Function *F, Function *G) {
719 if (!G->mayBeOverridden()) {
720 // Redirect direct callers of G to F.
721 replaceDirectCallers(G, F);
724 // If G was internal then we may have replaced all uses of G with F. If so,
725 // stop here and delete G. There's no need for a thunk.
726 if (G->hasLocalLinkage() && G->use_empty()) {
727 G->eraseFromParent();
731 Function *NewG = Function::Create(G->getFunctionType(), G->getLinkage(), "",
733 BasicBlock *BB = BasicBlock::Create(F->getContext(), "", NewG);
734 IRBuilder<false> Builder(BB);
736 SmallVector<Value *, 16> Args;
738 const FunctionType *FFTy = F->getFunctionType();
739 for (Function::arg_iterator AI = NewG->arg_begin(), AE = NewG->arg_end();
741 Args.push_back(Builder.CreateBitCast(AI, FFTy->getParamType(i)));
745 CallInst *CI = Builder.CreateCall(F, Args.begin(), Args.end());
747 CI->setCallingConv(F->getCallingConv());
748 if (NewG->getReturnType()->isVoidTy()) {
749 Builder.CreateRetVoid();
751 Builder.CreateRet(Builder.CreateBitCast(CI, NewG->getReturnType()));
754 NewG->copyAttributesFrom(G);
757 G->replaceAllUsesWith(NewG);
758 G->eraseFromParent();
760 DEBUG(dbgs() << "WriteThunk: " << NewG->getName() << '\n');
764 /// WriteAlias - Replace G with an alias to F and delete G.
765 void MergeFunctions::WriteAlias(Function *F, Function *G) {
766 Constant *BitcastF = ConstantExpr::getBitCast(F, G->getType());
767 GlobalAlias *GA = new GlobalAlias(G->getType(), G->getLinkage(), "",
768 BitcastF, G->getParent());
769 F->setAlignment(std::max(F->getAlignment(), G->getAlignment()));
771 GA->setVisibility(G->getVisibility());
773 G->replaceAllUsesWith(GA);
774 G->eraseFromParent();
776 DEBUG(dbgs() << "WriteAlias: " << GA->getName() << '\n');
780 /// MergeTwoFunctions - Merge two equivalent functions. Upon completion,
781 /// Function G is deleted.
782 void MergeFunctions::MergeTwoFunctions(Function *F, Function *G) {
783 if (F->mayBeOverridden()) {
784 assert(G->mayBeOverridden());
786 if (HasGlobalAliases) {
787 // Make them both thunks to the same internal function.
788 Function *H = Function::Create(F->getFunctionType(), F->getLinkage(), "",
790 H->copyAttributesFrom(F);
793 F->replaceAllUsesWith(H);
795 unsigned MaxAlignment = std::max(G->getAlignment(), H->getAlignment());
800 F->setAlignment(MaxAlignment);
801 F->setLinkage(GlobalValue::PrivateLinkage);
803 // We can't merge them. Instead, pick one and update all direct callers
804 // to call it and hope that we improve the instruction cache hit rate.
805 replaceDirectCallers(G, F);
810 WriteThunkOrAlias(F, G);
813 ++NumFunctionsMerged;
816 // Insert - Insert a ComparableFunction into the FnSet, or merge it away if
817 // equal to one that's already inserted.
818 bool MergeFunctions::Insert(ComparableFunction &NewF) {
819 std::pair<FnSetType::iterator, bool> Result = FnSet.insert(NewF);
823 const ComparableFunction &OldF = *Result.first;
825 // Never thunk a strong function to a weak function.
826 assert(!OldF.getFunc()->mayBeOverridden() ||
827 NewF.getFunc()->mayBeOverridden());
829 DEBUG(dbgs() << " " << OldF.getFunc()->getName() << " == "
830 << NewF.getFunc()->getName() << '\n');
832 Function *DeleteF = NewF.getFunc();
834 MergeTwoFunctions(OldF.getFunc(), DeleteF);
838 // Remove - Remove a function from FnSet. If it was already in FnSet, add it to
839 // Deferred so that we'll look at it in the next round.
840 void MergeFunctions::Remove(Function *F) {
841 ComparableFunction CF = ComparableFunction(F, TD);
842 if (FnSet.erase(CF)) {
843 Deferred.push_back(F);
847 // RemoveUsers - For each instruction used by the value, Remove() the function
848 // that contains the instruction. This should happen right before a call to RAUW.
849 void MergeFunctions::RemoveUsers(Value *V) {
850 std::vector<Value *> Worklist;
851 Worklist.push_back(V);
852 while (!Worklist.empty()) {
853 Value *V = Worklist.back();
856 for (Value::use_iterator UI = V->use_begin(), UE = V->use_end();
858 Use &U = UI.getUse();
859 if (Instruction *I = dyn_cast<Instruction>(U.getUser())) {
860 Remove(I->getParent()->getParent());
861 } else if (isa<GlobalValue>(U.getUser())) {
863 } else if (Constant *C = dyn_cast<Constant>(U.getUser())) {
864 for (Value::use_iterator CUI = C->use_begin(), CUE = C->use_end();
866 Worklist.push_back(*CUI);