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 /// Creates a hash-code for the function which is the same for any two
75 /// functions that will compare equal, without looking at the instructions
76 /// 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 /// ComparableFunction - A struct that pairs together functions with a
94 /// TargetData so that we can keep them together as elements in the DenseSet.
95 class ComparableFunction {
97 static const ComparableFunction EmptyKey;
98 static const ComparableFunction TombstoneKey;
100 ComparableFunction(Function *Func, TargetData *TD)
101 : Func(Func), Hash(profileFunction(Func)), TD(TD) {}
103 Function *getFunc() const { return Func; }
104 unsigned getHash() const { return Hash; }
105 TargetData *getTD() const { return TD; }
107 // Drops AssertingVH reference to the function. Outside of debug mode, this
111 "Attempted to release function twice, or release empty/tombstone!");
115 bool &getOrInsertCachedComparison(const ComparableFunction &Other,
116 bool &inserted) const {
117 typedef DenseMap<Function *, bool>::iterator iterator;
118 std::pair<iterator, bool> p =
119 CompareResultCache.insert(std::make_pair(Other.getFunc(), false));
121 return p.first->second;
125 explicit ComparableFunction(unsigned Hash)
126 : Func(NULL), Hash(Hash), TD(NULL) {}
128 // DenseMap::grow() triggers a recomparison of all keys in the map, which is
129 // wildly expensive. This cache tries to preserve known results.
130 mutable DenseMap<Function *, bool> CompareResultCache;
132 AssertingVH<Function> Func;
137 const ComparableFunction ComparableFunction::EmptyKey = ComparableFunction(0);
138 const ComparableFunction ComparableFunction::TombstoneKey =
139 ComparableFunction(1);
145 struct DenseMapInfo<ComparableFunction> {
146 static ComparableFunction getEmptyKey() {
147 return ComparableFunction::EmptyKey;
149 static ComparableFunction getTombstoneKey() {
150 return ComparableFunction::TombstoneKey;
152 static unsigned getHashValue(const ComparableFunction &CF) {
155 static bool isEqual(const ComparableFunction &LHS,
156 const ComparableFunction &RHS);
162 /// FunctionComparator - Compares two functions to determine whether or not
163 /// they will generate machine code with the same behaviour. TargetData is
164 /// used if available. The comparator always fails conservatively (erring on the
165 /// side of claiming that two functions are different).
166 class FunctionComparator {
168 FunctionComparator(const TargetData *TD, const Function *F1,
170 : F1(F1), F2(F2), TD(TD), IDMap1Count(0), IDMap2Count(0) {}
172 /// Test whether the two functions have equivalent behaviour.
176 /// Test whether two basic blocks have equivalent behaviour.
177 bool compare(const BasicBlock *BB1, const BasicBlock *BB2);
179 /// Assign or look up previously assigned numbers for the two values, and
180 /// return whether the numbers are equal. Numbers are assigned in the order
182 bool enumerate(const Value *V1, const Value *V2);
184 /// Compare two Instructions for equivalence, similar to
185 /// Instruction::isSameOperationAs but with modifications to the type
187 bool isEquivalentOperation(const Instruction *I1,
188 const Instruction *I2) const;
190 /// Compare two GEPs for equivalent pointer arithmetic.
191 bool isEquivalentGEP(const GEPOperator *GEP1, const GEPOperator *GEP2);
192 bool isEquivalentGEP(const GetElementPtrInst *GEP1,
193 const GetElementPtrInst *GEP2) {
194 return isEquivalentGEP(cast<GEPOperator>(GEP1), cast<GEPOperator>(GEP2));
197 /// Compare two Types, treating all pointer types as equal.
198 bool isEquivalentType(const Type *Ty1, const Type *Ty2) const;
200 // The two functions undergoing comparison.
201 const Function *F1, *F2;
203 const TargetData *TD;
205 typedef DenseMap<const Value *, unsigned long> IDMap;
207 unsigned long IDMap1Count, IDMap2Count;
212 // Any two pointers in the same address space are equivalent, intptr_t and
213 // pointers are equivalent. Otherwise, standard type equivalence rules apply.
214 bool FunctionComparator::isEquivalentType(const Type *Ty1,
215 const Type *Ty2) const {
218 if (Ty1->getTypeID() != Ty2->getTypeID()) {
220 LLVMContext &Ctx = Ty1->getContext();
221 if (isa<PointerType>(Ty1) && Ty2 == TD->getIntPtrType(Ctx)) return true;
222 if (isa<PointerType>(Ty2) && Ty1 == TD->getIntPtrType(Ctx)) return true;
227 switch(Ty1->getTypeID()) {
229 llvm_unreachable("Unknown type!");
230 // Fall through in Release mode.
231 case Type::IntegerTyID:
232 case Type::OpaqueTyID:
233 case Type::VectorTyID:
234 // Ty1 == Ty2 would have returned true earlier.
238 case Type::FloatTyID:
239 case Type::DoubleTyID:
240 case Type::X86_FP80TyID:
241 case Type::FP128TyID:
242 case Type::PPC_FP128TyID:
243 case Type::LabelTyID:
244 case Type::MetadataTyID:
247 case Type::PointerTyID: {
248 const PointerType *PTy1 = cast<PointerType>(Ty1);
249 const PointerType *PTy2 = cast<PointerType>(Ty2);
250 return PTy1->getAddressSpace() == PTy2->getAddressSpace();
253 case Type::StructTyID: {
254 const StructType *STy1 = cast<StructType>(Ty1);
255 const StructType *STy2 = cast<StructType>(Ty2);
256 if (STy1->getNumElements() != STy2->getNumElements())
259 if (STy1->isPacked() != STy2->isPacked())
262 for (unsigned i = 0, e = STy1->getNumElements(); i != e; ++i) {
263 if (!isEquivalentType(STy1->getElementType(i), STy2->getElementType(i)))
269 case Type::FunctionTyID: {
270 const FunctionType *FTy1 = cast<FunctionType>(Ty1);
271 const FunctionType *FTy2 = cast<FunctionType>(Ty2);
272 if (FTy1->getNumParams() != FTy2->getNumParams() ||
273 FTy1->isVarArg() != FTy2->isVarArg())
276 if (!isEquivalentType(FTy1->getReturnType(), FTy2->getReturnType()))
279 for (unsigned i = 0, e = FTy1->getNumParams(); i != e; ++i) {
280 if (!isEquivalentType(FTy1->getParamType(i), FTy2->getParamType(i)))
286 case Type::ArrayTyID: {
287 const ArrayType *ATy1 = cast<ArrayType>(Ty1);
288 const ArrayType *ATy2 = cast<ArrayType>(Ty2);
289 return ATy1->getNumElements() == ATy2->getNumElements() &&
290 isEquivalentType(ATy1->getElementType(), ATy2->getElementType());
295 // Determine whether the two operations are the same except that pointer-to-A
296 // and pointer-to-B are equivalent. This should be kept in sync with
297 // Instruction::isSameOperationAs.
298 bool FunctionComparator::isEquivalentOperation(const Instruction *I1,
299 const Instruction *I2) const {
300 if (I1->getOpcode() != I2->getOpcode() ||
301 I1->getNumOperands() != I2->getNumOperands() ||
302 !isEquivalentType(I1->getType(), I2->getType()) ||
303 !I1->hasSameSubclassOptionalData(I2))
306 // We have two instructions of identical opcode and #operands. Check to see
307 // if all operands are the same type
308 for (unsigned i = 0, e = I1->getNumOperands(); i != e; ++i)
309 if (!isEquivalentType(I1->getOperand(i)->getType(),
310 I2->getOperand(i)->getType()))
313 // Check special state that is a part of some instructions.
314 if (const LoadInst *LI = dyn_cast<LoadInst>(I1))
315 return LI->isVolatile() == cast<LoadInst>(I2)->isVolatile() &&
316 LI->getAlignment() == cast<LoadInst>(I2)->getAlignment();
317 if (const StoreInst *SI = dyn_cast<StoreInst>(I1))
318 return SI->isVolatile() == cast<StoreInst>(I2)->isVolatile() &&
319 SI->getAlignment() == cast<StoreInst>(I2)->getAlignment();
320 if (const CmpInst *CI = dyn_cast<CmpInst>(I1))
321 return CI->getPredicate() == cast<CmpInst>(I2)->getPredicate();
322 if (const CallInst *CI = dyn_cast<CallInst>(I1))
323 return CI->isTailCall() == cast<CallInst>(I2)->isTailCall() &&
324 CI->getCallingConv() == cast<CallInst>(I2)->getCallingConv() &&
325 CI->getAttributes() == cast<CallInst>(I2)->getAttributes();
326 if (const InvokeInst *CI = dyn_cast<InvokeInst>(I1))
327 return CI->getCallingConv() == cast<InvokeInst>(I2)->getCallingConv() &&
328 CI->getAttributes() == cast<InvokeInst>(I2)->getAttributes();
329 if (const InsertValueInst *IVI = dyn_cast<InsertValueInst>(I1)) {
330 if (IVI->getNumIndices() != cast<InsertValueInst>(I2)->getNumIndices())
332 for (unsigned i = 0, e = IVI->getNumIndices(); i != e; ++i)
333 if (IVI->idx_begin()[i] != cast<InsertValueInst>(I2)->idx_begin()[i])
337 if (const ExtractValueInst *EVI = dyn_cast<ExtractValueInst>(I1)) {
338 if (EVI->getNumIndices() != cast<ExtractValueInst>(I2)->getNumIndices())
340 for (unsigned i = 0, e = EVI->getNumIndices(); i != e; ++i)
341 if (EVI->idx_begin()[i] != cast<ExtractValueInst>(I2)->idx_begin()[i])
349 // Determine whether two GEP operations perform the same underlying arithmetic.
350 bool FunctionComparator::isEquivalentGEP(const GEPOperator *GEP1,
351 const GEPOperator *GEP2) {
352 // When we have target data, we can reduce the GEP down to the value in bytes
353 // added to the address.
354 if (TD && GEP1->hasAllConstantIndices() && GEP2->hasAllConstantIndices()) {
355 SmallVector<Value *, 8> Indices1(GEP1->idx_begin(), GEP1->idx_end());
356 SmallVector<Value *, 8> Indices2(GEP2->idx_begin(), GEP2->idx_end());
357 uint64_t Offset1 = TD->getIndexedOffset(GEP1->getPointerOperandType(),
358 Indices1.data(), Indices1.size());
359 uint64_t Offset2 = TD->getIndexedOffset(GEP2->getPointerOperandType(),
360 Indices2.data(), Indices2.size());
361 return Offset1 == Offset2;
364 if (GEP1->getPointerOperand()->getType() !=
365 GEP2->getPointerOperand()->getType())
368 if (GEP1->getNumOperands() != GEP2->getNumOperands())
371 for (unsigned i = 0, e = GEP1->getNumOperands(); i != e; ++i) {
372 if (!enumerate(GEP1->getOperand(i), GEP2->getOperand(i)))
379 // Compare two values used by the two functions under pair-wise comparison. If
380 // this is the first time the values are seen, they're added to the mapping so
381 // that we will detect mismatches on next use.
382 bool FunctionComparator::enumerate(const Value *V1, const Value *V2) {
383 // Check for function @f1 referring to itself and function @f2 referring to
384 // itself, or referring to each other, or both referring to either of them.
385 // They're all equivalent if the two functions are otherwise equivalent.
386 if (V1 == F1 && V2 == F2)
388 if (V1 == F2 && V2 == F1)
391 if (const Constant *C1 = dyn_cast<Constant>(V1)) {
392 if (V1 == V2) return true;
393 const Constant *C2 = dyn_cast<Constant>(V2);
394 if (!C2) return false;
395 // TODO: constant expressions with GEP or references to F1 or F2.
396 if (C1->isNullValue() && C2->isNullValue() &&
397 isEquivalentType(C1->getType(), C2->getType()))
399 // Try bitcasting C2 to C1's type. If the bitcast is legal and returns C1
400 // then they must have equal bit patterns.
401 return C1->getType()->canLosslesslyBitCastTo(C2->getType()) &&
402 C1 == ConstantExpr::getBitCast(const_cast<Constant*>(C2), C1->getType());
405 if (isa<InlineAsm>(V1) && isa<InlineAsm>(V2)) {
406 const InlineAsm *IA1 = cast<InlineAsm>(V1);
407 const InlineAsm *IA2 = cast<InlineAsm>(V2);
408 return IA1->getAsmString() == IA2->getAsmString() &&
409 IA1->getConstraintString() == IA2->getConstraintString();
412 unsigned long &ID1 = Map1[V1];
416 unsigned long &ID2 = Map2[V2];
423 // Test whether two basic blocks have equivalent behaviour.
424 bool FunctionComparator::compare(const BasicBlock *BB1, const BasicBlock *BB2) {
425 BasicBlock::const_iterator F1I = BB1->begin(), F1E = BB1->end();
426 BasicBlock::const_iterator F2I = BB2->begin(), F2E = BB2->end();
429 if (!enumerate(F1I, F2I))
432 if (const GetElementPtrInst *GEP1 = dyn_cast<GetElementPtrInst>(F1I)) {
433 const GetElementPtrInst *GEP2 = dyn_cast<GetElementPtrInst>(F2I);
437 if (!enumerate(GEP1->getPointerOperand(), GEP2->getPointerOperand()))
440 if (!isEquivalentGEP(GEP1, GEP2))
443 if (!isEquivalentOperation(F1I, F2I))
446 assert(F1I->getNumOperands() == F2I->getNumOperands());
447 for (unsigned i = 0, e = F1I->getNumOperands(); i != e; ++i) {
448 Value *OpF1 = F1I->getOperand(i);
449 Value *OpF2 = F2I->getOperand(i);
451 if (!enumerate(OpF1, OpF2))
454 if (OpF1->getValueID() != OpF2->getValueID() ||
455 !isEquivalentType(OpF1->getType(), OpF2->getType()))
461 } while (F1I != F1E && F2I != F2E);
463 return F1I == F1E && F2I == F2E;
466 // Test whether the two functions have equivalent behaviour.
467 bool FunctionComparator::compare() {
468 // We need to recheck everything, but check the things that weren't included
469 // in the hash first.
471 if (F1->getAttributes() != F2->getAttributes())
474 if (F1->hasGC() != F2->hasGC())
477 if (F1->hasGC() && F1->getGC() != F2->getGC())
480 if (F1->hasSection() != F2->hasSection())
483 if (F1->hasSection() && F1->getSection() != F2->getSection())
486 if (F1->isVarArg() != F2->isVarArg())
489 // TODO: if it's internal and only used in direct calls, we could handle this
491 if (F1->getCallingConv() != F2->getCallingConv())
494 if (!isEquivalentType(F1->getFunctionType(), F2->getFunctionType()))
497 assert(F1->arg_size() == F2->arg_size() &&
498 "Identically typed functions have different numbers of args!");
500 // Visit the arguments so that they get enumerated in the order they're
502 for (Function::const_arg_iterator f1i = F1->arg_begin(),
503 f2i = F2->arg_begin(), f1e = F1->arg_end(); f1i != f1e; ++f1i, ++f2i) {
504 if (!enumerate(f1i, f2i))
505 llvm_unreachable("Arguments repeat!");
508 // We do a CFG-ordered walk since the actual ordering of the blocks in the
509 // linked list is immaterial. Our walk starts at the entry block for both
510 // functions, then takes each block from each terminator in order. As an
511 // artifact, this also means that unreachable blocks are ignored.
512 SmallVector<const BasicBlock *, 8> F1BBs, F2BBs;
513 SmallSet<const BasicBlock *, 128> VisitedBBs; // in terms of F1.
515 F1BBs.push_back(&F1->getEntryBlock());
516 F2BBs.push_back(&F2->getEntryBlock());
518 VisitedBBs.insert(F1BBs[0]);
519 while (!F1BBs.empty()) {
520 const BasicBlock *F1BB = F1BBs.pop_back_val();
521 const BasicBlock *F2BB = F2BBs.pop_back_val();
523 if (!enumerate(F1BB, F2BB) || !compare(F1BB, F2BB))
526 const TerminatorInst *F1TI = F1BB->getTerminator();
527 const TerminatorInst *F2TI = F2BB->getTerminator();
529 assert(F1TI->getNumSuccessors() == F2TI->getNumSuccessors());
530 for (unsigned i = 0, e = F1TI->getNumSuccessors(); i != e; ++i) {
531 if (!VisitedBBs.insert(F1TI->getSuccessor(i)))
534 F1BBs.push_back(F1TI->getSuccessor(i));
535 F2BBs.push_back(F2TI->getSuccessor(i));
543 /// MergeFunctions finds functions which will generate identical machine code,
544 /// by considering all pointer types to be equivalent. Once identified,
545 /// MergeFunctions will fold them by replacing a call to one to a call to a
546 /// bitcast of the other.
548 class MergeFunctions : public ModulePass {
552 : ModulePass(ID), HasGlobalAliases(false) {
553 initializeMergeFunctionsPass(*PassRegistry::getPassRegistry());
556 bool runOnModule(Module &M);
559 typedef DenseSet<ComparableFunction> FnSetType;
561 /// A work queue of functions that may have been modified and should be
563 std::vector<WeakVH> Deferred;
565 /// Insert a ComparableFunction into the FnSet, or merge it away if it's
566 /// equal to one that's already present.
567 bool insert(ComparableFunction &NewF);
569 /// Remove a Function from the FnSet and queue it up for a second sweep of
571 void remove(Function *F);
573 /// Find the functions that use this Value and remove them from FnSet and
574 /// queue the functions.
575 void removeUsers(Value *V);
577 /// Replace all direct calls of Old with calls of New. Will bitcast New if
578 /// necessary to make types match.
579 void replaceDirectCallers(Function *Old, Function *New);
581 /// Merge two equivalent functions. Upon completion, G may be deleted, or may
582 /// be converted into a thunk. In either case, it should never be visited
584 void mergeTwoFunctions(Function *F, Function *G);
586 /// Replace G with a thunk or an alias to F. Deletes G.
587 void writeThunkOrAlias(Function *F, Function *G);
589 /// Replace G with a simple tail call to bitcast(F). Also replace direct uses
590 /// of G with bitcast(F). Deletes G.
591 void writeThunk(Function *F, Function *G);
593 /// Replace G with an alias to F. Deletes G.
594 void writeAlias(Function *F, Function *G);
596 /// The set of all distinct functions. Use the insert() and remove() methods
600 /// TargetData for more accurate GEP comparisons. May be NULL.
603 /// Whether or not the target supports global aliases.
604 bool HasGlobalAliases;
607 } // end anonymous namespace
609 char MergeFunctions::ID = 0;
610 INITIALIZE_PASS(MergeFunctions, "mergefunc", "Merge Functions", false, false)
612 ModulePass *llvm::createMergeFunctionsPass() {
613 return new MergeFunctions();
616 bool MergeFunctions::runOnModule(Module &M) {
617 bool Changed = false;
618 TD = getAnalysisIfAvailable<TargetData>();
620 for (Module::iterator I = M.begin(), E = M.end(); I != E; ++I) {
621 if (!I->isDeclaration() && !I->hasAvailableExternallyLinkage())
622 Deferred.push_back(WeakVH(I));
624 FnSet.resize(Deferred.size());
627 std::vector<WeakVH> Worklist;
628 Deferred.swap(Worklist);
630 DEBUG(dbgs() << "size of module: " << M.size() << '\n');
631 DEBUG(dbgs() << "size of worklist: " << Worklist.size() << '\n');
633 // Insert only strong functions and merge them. Strong function merging
634 // always deletes one of them.
635 for (std::vector<WeakVH>::iterator I = Worklist.begin(),
636 E = Worklist.end(); I != E; ++I) {
638 Function *F = cast<Function>(*I);
639 if (!F->isDeclaration() && !F->hasAvailableExternallyLinkage() &&
640 !F->mayBeOverridden()) {
641 ComparableFunction CF = ComparableFunction(F, TD);
642 Changed |= insert(CF);
646 // Insert only weak functions and merge them. By doing these second we
647 // create thunks to the strong function when possible. When two weak
648 // functions are identical, we create a new strong function with two weak
649 // weak thunks to it which are identical but not mergable.
650 for (std::vector<WeakVH>::iterator I = Worklist.begin(),
651 E = Worklist.end(); I != E; ++I) {
653 Function *F = cast<Function>(*I);
654 if (!F->isDeclaration() && !F->hasAvailableExternallyLinkage() &&
655 F->mayBeOverridden()) {
656 ComparableFunction CF = ComparableFunction(F, TD);
657 Changed |= insert(CF);
660 DEBUG(dbgs() << "size of FnSet: " << FnSet.size() << '\n');
661 } while (!Deferred.empty());
668 bool DenseMapInfo<ComparableFunction>::isEqual(const ComparableFunction &LHS,
669 const ComparableFunction &RHS) {
670 if (LHS.getFunc() == RHS.getFunc() &&
671 LHS.getHash() == RHS.getHash())
673 if (!LHS.getFunc() || !RHS.getFunc())
675 assert(LHS.getTD() == RHS.getTD() &&
676 "Comparing functions for different targets");
679 bool &result1 = LHS.getOrInsertCachedComparison(RHS, inserted);
682 bool &result2 = RHS.getOrInsertCachedComparison(LHS, inserted);
684 return result1 = result2;
686 return result1 = result2 = FunctionComparator(LHS.getTD(), LHS.getFunc(),
687 RHS.getFunc()).compare();
690 // Replace direct callers of Old with New.
691 void MergeFunctions::replaceDirectCallers(Function *Old, Function *New) {
692 Constant *BitcastNew = ConstantExpr::getBitCast(New, Old->getType());
693 for (Value::use_iterator UI = Old->use_begin(), UE = Old->use_end();
695 Value::use_iterator TheIter = UI;
697 CallSite CS(*TheIter);
698 if (CS && CS.isCallee(TheIter)) {
699 remove(CS.getInstruction()->getParent()->getParent());
700 TheIter.getUse().set(BitcastNew);
705 // Replace G with an alias to F if possible, or else a thunk to F. Deletes G.
706 void MergeFunctions::writeThunkOrAlias(Function *F, Function *G) {
707 if (HasGlobalAliases && G->hasUnnamedAddr()) {
708 if (G->hasExternalLinkage() || G->hasLocalLinkage() ||
709 G->hasWeakLinkage()) {
718 // Replace G with a simple tail call to bitcast(F). Also replace direct uses
719 // of G with bitcast(F). Deletes G.
720 void MergeFunctions::writeThunk(Function *F, Function *G) {
721 if (!G->mayBeOverridden()) {
722 // Redirect direct callers of G to F.
723 replaceDirectCallers(G, F);
726 // If G was internal then we may have replaced all uses of G with F. If so,
727 // stop here and delete G. There's no need for a thunk.
728 if (G->hasLocalLinkage() && G->use_empty()) {
729 G->eraseFromParent();
733 Function *NewG = Function::Create(G->getFunctionType(), G->getLinkage(), "",
735 BasicBlock *BB = BasicBlock::Create(F->getContext(), "", NewG);
736 IRBuilder<false> Builder(BB);
738 SmallVector<Value *, 16> Args;
740 const FunctionType *FFTy = F->getFunctionType();
741 for (Function::arg_iterator AI = NewG->arg_begin(), AE = NewG->arg_end();
743 Args.push_back(Builder.CreateBitCast(AI, FFTy->getParamType(i)));
747 CallInst *CI = Builder.CreateCall(F, Args.begin(), Args.end());
749 CI->setCallingConv(F->getCallingConv());
750 if (NewG->getReturnType()->isVoidTy()) {
751 Builder.CreateRetVoid();
753 Builder.CreateRet(Builder.CreateBitCast(CI, NewG->getReturnType()));
756 NewG->copyAttributesFrom(G);
759 G->replaceAllUsesWith(NewG);
760 G->eraseFromParent();
762 DEBUG(dbgs() << "writeThunk: " << NewG->getName() << '\n');
766 // Replace G with an alias to F and delete G.
767 void MergeFunctions::writeAlias(Function *F, Function *G) {
768 Constant *BitcastF = ConstantExpr::getBitCast(F, G->getType());
769 GlobalAlias *GA = new GlobalAlias(G->getType(), G->getLinkage(), "",
770 BitcastF, G->getParent());
771 F->setAlignment(std::max(F->getAlignment(), G->getAlignment()));
773 GA->setVisibility(G->getVisibility());
775 G->replaceAllUsesWith(GA);
776 G->eraseFromParent();
778 DEBUG(dbgs() << "writeAlias: " << GA->getName() << '\n');
782 // Merge two equivalent functions. Upon completion, Function G is deleted.
783 void MergeFunctions::mergeTwoFunctions(Function *F, Function *G) {
784 if (F->mayBeOverridden()) {
785 assert(G->mayBeOverridden());
787 if (HasGlobalAliases) {
788 // Make them both thunks to the same internal function.
789 Function *H = Function::Create(F->getFunctionType(), F->getLinkage(), "",
791 H->copyAttributesFrom(F);
794 F->replaceAllUsesWith(H);
796 unsigned MaxAlignment = std::max(G->getAlignment(), H->getAlignment());
801 F->setAlignment(MaxAlignment);
802 F->setLinkage(GlobalValue::PrivateLinkage);
804 // We can't merge them. Instead, pick one and update all direct callers
805 // to call it and hope that we improve the instruction cache hit rate.
806 replaceDirectCallers(G, F);
811 writeThunkOrAlias(F, G);
814 ++NumFunctionsMerged;
817 // Insert a ComparableFunction into the FnSet, or merge it away if equal to one
818 // that was already inserted.
819 bool MergeFunctions::insert(ComparableFunction &NewF) {
820 std::pair<FnSetType::iterator, bool> Result = FnSet.insert(NewF);
824 const ComparableFunction &OldF = *Result.first;
826 // Never thunk a strong function to a weak function.
827 assert(!OldF.getFunc()->mayBeOverridden() ||
828 NewF.getFunc()->mayBeOverridden());
830 DEBUG(dbgs() << " " << OldF.getFunc()->getName() << " == "
831 << NewF.getFunc()->getName() << '\n');
833 Function *DeleteF = NewF.getFunc();
835 mergeTwoFunctions(OldF.getFunc(), DeleteF);
839 // Remove a function from FnSet. If it was already in FnSet, add it to Deferred
840 // so that we'll look at it in the next round.
841 void MergeFunctions::remove(Function *F) {
842 ComparableFunction CF = ComparableFunction(F, TD);
843 if (FnSet.erase(CF)) {
844 Deferred.push_back(F);
848 // For each instruction used by the value, remove() the function that contains
849 // the instruction. This should happen right before a call to RAUW.
850 void MergeFunctions::removeUsers(Value *V) {
851 std::vector<Value *> Worklist;
852 Worklist.push_back(V);
853 while (!Worklist.empty()) {
854 Value *V = Worklist.back();
857 for (Value::use_iterator UI = V->use_begin(), UE = V->use_end();
859 Use &U = UI.getUse();
860 if (Instruction *I = dyn_cast<Instruction>(U.getUser())) {
861 remove(I->getParent()->getParent());
862 } else if (isa<GlobalValue>(U.getUser())) {
864 } else if (Constant *C = dyn_cast<Constant>(U.getUser())) {
865 for (Value::use_iterator CUI = C->use_begin(), CUE = C->use_end();
867 Worklist.push_back(*CUI);