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/Constants.h"
49 #include "llvm/IRBuilder.h"
50 #include "llvm/InlineAsm.h"
51 #include "llvm/Instructions.h"
52 #include "llvm/LLVMContext.h"
53 #include "llvm/Module.h"
54 #include "llvm/Operator.h"
55 #include "llvm/Pass.h"
56 #include "llvm/ADT/DenseSet.h"
57 #include "llvm/ADT/FoldingSet.h"
58 #include "llvm/ADT/STLExtras.h"
59 #include "llvm/ADT/SmallSet.h"
60 #include "llvm/ADT/Statistic.h"
61 #include "llvm/Support/CallSite.h"
62 #include "llvm/Support/Debug.h"
63 #include "llvm/Support/ErrorHandling.h"
64 #include "llvm/Support/ValueHandle.h"
65 #include "llvm/Support/raw_ostream.h"
66 #include "llvm/DataLayout.h"
70 STATISTIC(NumFunctionsMerged, "Number of functions merged");
71 STATISTIC(NumThunksWritten, "Number of thunks generated");
72 STATISTIC(NumAliasesWritten, "Number of aliases generated");
73 STATISTIC(NumDoubleWeak, "Number of new functions created");
75 /// Creates a hash-code for the function which is the same for any two
76 /// functions that will compare equal, without looking at the instructions
77 /// inside the function.
78 static unsigned profileFunction(const Function *F) {
79 FunctionType *FTy = F->getFunctionType();
82 ID.AddInteger(F->size());
83 ID.AddInteger(F->getCallingConv());
84 ID.AddBoolean(F->hasGC());
85 ID.AddBoolean(FTy->isVarArg());
86 ID.AddInteger(FTy->getReturnType()->getTypeID());
87 for (unsigned i = 0, e = FTy->getNumParams(); i != e; ++i)
88 ID.AddInteger(FTy->getParamType(i)->getTypeID());
89 return ID.ComputeHash();
94 /// ComparableFunction - A struct that pairs together functions with a
95 /// DataLayout so that we can keep them together as elements in the DenseSet.
96 class ComparableFunction {
98 static const ComparableFunction EmptyKey;
99 static const ComparableFunction TombstoneKey;
100 static DataLayout * const LookupOnly;
102 ComparableFunction(Function *Func, DataLayout *TD)
103 : Func(Func), Hash(profileFunction(Func)), TD(TD) {}
105 Function *getFunc() const { return Func; }
106 unsigned getHash() const { return Hash; }
107 DataLayout *getTD() const { return TD; }
109 // Drops AssertingVH reference to the function. Outside of debug mode, this
113 "Attempted to release function twice, or release empty/tombstone!");
118 explicit ComparableFunction(unsigned Hash)
119 : Func(NULL), Hash(Hash), TD(NULL) {}
121 AssertingVH<Function> Func;
126 const ComparableFunction ComparableFunction::EmptyKey = ComparableFunction(0);
127 const ComparableFunction ComparableFunction::TombstoneKey =
128 ComparableFunction(1);
129 DataLayout *const ComparableFunction::LookupOnly = (DataLayout*)(-1);
135 struct DenseMapInfo<ComparableFunction> {
136 static ComparableFunction getEmptyKey() {
137 return ComparableFunction::EmptyKey;
139 static ComparableFunction getTombstoneKey() {
140 return ComparableFunction::TombstoneKey;
142 static unsigned getHashValue(const ComparableFunction &CF) {
145 static bool isEqual(const ComparableFunction &LHS,
146 const ComparableFunction &RHS);
152 /// FunctionComparator - Compares two functions to determine whether or not
153 /// they will generate machine code with the same behaviour. DataLayout is
154 /// used if available. The comparator always fails conservatively (erring on the
155 /// side of claiming that two functions are different).
156 class FunctionComparator {
158 FunctionComparator(const DataLayout *TD, const Function *F1,
160 : F1(F1), F2(F2), TD(TD) {}
162 /// Test whether the two functions have equivalent behaviour.
166 /// Test whether two basic blocks have equivalent behaviour.
167 bool compare(const BasicBlock *BB1, const BasicBlock *BB2);
169 /// Assign or look up previously assigned numbers for the two values, and
170 /// return whether the numbers are equal. Numbers are assigned in the order
172 bool enumerate(const Value *V1, const Value *V2);
174 /// Compare two Instructions for equivalence, similar to
175 /// Instruction::isSameOperationAs but with modifications to the type
177 bool isEquivalentOperation(const Instruction *I1,
178 const Instruction *I2) const;
180 /// Compare two GEPs for equivalent pointer arithmetic.
181 bool isEquivalentGEP(const GEPOperator *GEP1, const GEPOperator *GEP2);
182 bool isEquivalentGEP(const GetElementPtrInst *GEP1,
183 const GetElementPtrInst *GEP2) {
184 return isEquivalentGEP(cast<GEPOperator>(GEP1), cast<GEPOperator>(GEP2));
187 /// Compare two Types, treating all pointer types as equal.
188 bool isEquivalentType(Type *Ty1, Type *Ty2) const;
190 // The two functions undergoing comparison.
191 const Function *F1, *F2;
193 const DataLayout *TD;
195 DenseMap<const Value *, const Value *> id_map;
196 DenseSet<const Value *> seen_values;
201 // Any two pointers in the same address space are equivalent, intptr_t and
202 // pointers are equivalent. Otherwise, standard type equivalence rules apply.
203 bool FunctionComparator::isEquivalentType(Type *Ty1,
207 if (Ty1->getTypeID() != Ty2->getTypeID()) {
209 if (isa<PointerType>(Ty1) && Ty2 == TD->getIntPtrType(Ty1)) return true;
210 if (isa<PointerType>(Ty2) && Ty1 == TD->getIntPtrType(Ty2)) return true;
215 switch (Ty1->getTypeID()) {
217 llvm_unreachable("Unknown type!");
218 // Fall through in Release mode.
219 case Type::IntegerTyID:
220 case Type::VectorTyID:
221 // Ty1 == Ty2 would have returned true earlier.
225 case Type::FloatTyID:
226 case Type::DoubleTyID:
227 case Type::X86_FP80TyID:
228 case Type::FP128TyID:
229 case Type::PPC_FP128TyID:
230 case Type::LabelTyID:
231 case Type::MetadataTyID:
234 case Type::PointerTyID: {
235 PointerType *PTy1 = cast<PointerType>(Ty1);
236 PointerType *PTy2 = cast<PointerType>(Ty2);
237 return PTy1->getAddressSpace() == PTy2->getAddressSpace();
240 case Type::StructTyID: {
241 StructType *STy1 = cast<StructType>(Ty1);
242 StructType *STy2 = cast<StructType>(Ty2);
243 if (STy1->getNumElements() != STy2->getNumElements())
246 if (STy1->isPacked() != STy2->isPacked())
249 for (unsigned i = 0, e = STy1->getNumElements(); i != e; ++i) {
250 if (!isEquivalentType(STy1->getElementType(i), STy2->getElementType(i)))
256 case Type::FunctionTyID: {
257 FunctionType *FTy1 = cast<FunctionType>(Ty1);
258 FunctionType *FTy2 = cast<FunctionType>(Ty2);
259 if (FTy1->getNumParams() != FTy2->getNumParams() ||
260 FTy1->isVarArg() != FTy2->isVarArg())
263 if (!isEquivalentType(FTy1->getReturnType(), FTy2->getReturnType()))
266 for (unsigned i = 0, e = FTy1->getNumParams(); i != e; ++i) {
267 if (!isEquivalentType(FTy1->getParamType(i), FTy2->getParamType(i)))
273 case Type::ArrayTyID: {
274 ArrayType *ATy1 = cast<ArrayType>(Ty1);
275 ArrayType *ATy2 = cast<ArrayType>(Ty2);
276 return ATy1->getNumElements() == ATy2->getNumElements() &&
277 isEquivalentType(ATy1->getElementType(), ATy2->getElementType());
282 // Determine whether the two operations are the same except that pointer-to-A
283 // and pointer-to-B are equivalent. This should be kept in sync with
284 // Instruction::isSameOperationAs.
285 bool FunctionComparator::isEquivalentOperation(const Instruction *I1,
286 const Instruction *I2) const {
287 // Differences from Instruction::isSameOperationAs:
288 // * replace type comparison with calls to isEquivalentType.
289 // * we test for I->hasSameSubclassOptionalData (nuw/nsw/tail) at the top
290 // * because of the above, we don't test for the tail bit on calls later on
291 if (I1->getOpcode() != I2->getOpcode() ||
292 I1->getNumOperands() != I2->getNumOperands() ||
293 !isEquivalentType(I1->getType(), I2->getType()) ||
294 !I1->hasSameSubclassOptionalData(I2))
297 // We have two instructions of identical opcode and #operands. Check to see
298 // if all operands are the same type
299 for (unsigned i = 0, e = I1->getNumOperands(); i != e; ++i)
300 if (!isEquivalentType(I1->getOperand(i)->getType(),
301 I2->getOperand(i)->getType()))
304 // Check special state that is a part of some instructions.
305 if (const LoadInst *LI = dyn_cast<LoadInst>(I1))
306 return LI->isVolatile() == cast<LoadInst>(I2)->isVolatile() &&
307 LI->getAlignment() == cast<LoadInst>(I2)->getAlignment() &&
308 LI->getOrdering() == cast<LoadInst>(I2)->getOrdering() &&
309 LI->getSynchScope() == cast<LoadInst>(I2)->getSynchScope();
310 if (const StoreInst *SI = dyn_cast<StoreInst>(I1))
311 return SI->isVolatile() == cast<StoreInst>(I2)->isVolatile() &&
312 SI->getAlignment() == cast<StoreInst>(I2)->getAlignment() &&
313 SI->getOrdering() == cast<StoreInst>(I2)->getOrdering() &&
314 SI->getSynchScope() == cast<StoreInst>(I2)->getSynchScope();
315 if (const CmpInst *CI = dyn_cast<CmpInst>(I1))
316 return CI->getPredicate() == cast<CmpInst>(I2)->getPredicate();
317 if (const CallInst *CI = dyn_cast<CallInst>(I1))
318 return CI->getCallingConv() == cast<CallInst>(I2)->getCallingConv() &&
319 CI->getAttributes() == cast<CallInst>(I2)->getAttributes();
320 if (const InvokeInst *CI = dyn_cast<InvokeInst>(I1))
321 return CI->getCallingConv() == cast<InvokeInst>(I2)->getCallingConv() &&
322 CI->getAttributes() == cast<InvokeInst>(I2)->getAttributes();
323 if (const InsertValueInst *IVI = dyn_cast<InsertValueInst>(I1))
324 return IVI->getIndices() == cast<InsertValueInst>(I2)->getIndices();
325 if (const ExtractValueInst *EVI = dyn_cast<ExtractValueInst>(I1))
326 return EVI->getIndices() == cast<ExtractValueInst>(I2)->getIndices();
327 if (const FenceInst *FI = dyn_cast<FenceInst>(I1))
328 return FI->getOrdering() == cast<FenceInst>(I2)->getOrdering() &&
329 FI->getSynchScope() == cast<FenceInst>(I2)->getSynchScope();
330 if (const AtomicCmpXchgInst *CXI = dyn_cast<AtomicCmpXchgInst>(I1))
331 return CXI->isVolatile() == cast<AtomicCmpXchgInst>(I2)->isVolatile() &&
332 CXI->getOrdering() == cast<AtomicCmpXchgInst>(I2)->getOrdering() &&
333 CXI->getSynchScope() == cast<AtomicCmpXchgInst>(I2)->getSynchScope();
334 if (const AtomicRMWInst *RMWI = dyn_cast<AtomicRMWInst>(I1))
335 return RMWI->getOperation() == cast<AtomicRMWInst>(I2)->getOperation() &&
336 RMWI->isVolatile() == cast<AtomicRMWInst>(I2)->isVolatile() &&
337 RMWI->getOrdering() == cast<AtomicRMWInst>(I2)->getOrdering() &&
338 RMWI->getSynchScope() == cast<AtomicRMWInst>(I2)->getSynchScope();
343 // Determine whether two GEP operations perform the same underlying arithmetic.
344 bool FunctionComparator::isEquivalentGEP(const GEPOperator *GEP1,
345 const GEPOperator *GEP2) {
346 // When we have target data, we can reduce the GEP down to the value in bytes
347 // added to the address.
348 if (TD && GEP1->hasAllConstantIndices() && GEP2->hasAllConstantIndices()) {
349 SmallVector<Value *, 8> Indices1(GEP1->idx_begin(), GEP1->idx_end());
350 SmallVector<Value *, 8> Indices2(GEP2->idx_begin(), GEP2->idx_end());
351 uint64_t Offset1 = TD->getIndexedOffset(GEP1->getPointerOperandType(),
353 uint64_t Offset2 = TD->getIndexedOffset(GEP2->getPointerOperandType(),
355 return Offset1 == Offset2;
358 if (GEP1->getPointerOperand()->getType() !=
359 GEP2->getPointerOperand()->getType())
362 if (GEP1->getNumOperands() != GEP2->getNumOperands())
365 for (unsigned i = 0, e = GEP1->getNumOperands(); i != e; ++i) {
366 if (!enumerate(GEP1->getOperand(i), GEP2->getOperand(i)))
373 // Compare two values used by the two functions under pair-wise comparison. If
374 // this is the first time the values are seen, they're added to the mapping so
375 // that we will detect mismatches on next use.
376 bool FunctionComparator::enumerate(const Value *V1, const Value *V2) {
377 // Check for function @f1 referring to itself and function @f2 referring to
378 // itself, or referring to each other, or both referring to either of them.
379 // They're all equivalent if the two functions are otherwise equivalent.
380 if (V1 == F1 && V2 == F2)
382 if (V1 == F2 && V2 == F1)
385 if (const Constant *C1 = dyn_cast<Constant>(V1)) {
386 if (V1 == V2) return true;
387 const Constant *C2 = dyn_cast<Constant>(V2);
388 if (!C2) return false;
389 // TODO: constant expressions with GEP or references to F1 or F2.
390 if (C1->isNullValue() && C2->isNullValue() &&
391 isEquivalentType(C1->getType(), C2->getType()))
393 // Try bitcasting C2 to C1's type. If the bitcast is legal and returns C1
394 // then they must have equal bit patterns.
395 return C1->getType()->canLosslesslyBitCastTo(C2->getType()) &&
396 C1 == ConstantExpr::getBitCast(const_cast<Constant*>(C2), C1->getType());
399 if (isa<InlineAsm>(V1) || isa<InlineAsm>(V2))
402 // Check that V1 maps to V2. If we find a value that V1 maps to then we simply
403 // check whether it's equal to V2. When there is no mapping then we need to
404 // ensure that V2 isn't already equivalent to something else. For this
405 // purpose, we track the V2 values in a set.
407 const Value *&map_elem = id_map[V1];
409 return map_elem == V2;
410 if (!seen_values.insert(V2).second)
416 // Test whether two basic blocks have equivalent behaviour.
417 bool FunctionComparator::compare(const BasicBlock *BB1, const BasicBlock *BB2) {
418 BasicBlock::const_iterator F1I = BB1->begin(), F1E = BB1->end();
419 BasicBlock::const_iterator F2I = BB2->begin(), F2E = BB2->end();
422 if (!enumerate(F1I, F2I))
425 if (const GetElementPtrInst *GEP1 = dyn_cast<GetElementPtrInst>(F1I)) {
426 const GetElementPtrInst *GEP2 = dyn_cast<GetElementPtrInst>(F2I);
430 if (!enumerate(GEP1->getPointerOperand(), GEP2->getPointerOperand()))
433 if (!isEquivalentGEP(GEP1, GEP2))
436 if (!isEquivalentOperation(F1I, F2I))
439 assert(F1I->getNumOperands() == F2I->getNumOperands());
440 for (unsigned i = 0, e = F1I->getNumOperands(); i != e; ++i) {
441 Value *OpF1 = F1I->getOperand(i);
442 Value *OpF2 = F2I->getOperand(i);
444 if (!enumerate(OpF1, OpF2))
447 if (OpF1->getValueID() != OpF2->getValueID() ||
448 !isEquivalentType(OpF1->getType(), OpF2->getType()))
454 } while (F1I != F1E && F2I != F2E);
456 return F1I == F1E && F2I == F2E;
459 // Test whether the two functions have equivalent behaviour.
460 bool FunctionComparator::compare() {
461 // We need to recheck everything, but check the things that weren't included
462 // in the hash first.
464 if (F1->getAttributes() != F2->getAttributes())
467 if (F1->hasGC() != F2->hasGC())
470 if (F1->hasGC() && F1->getGC() != F2->getGC())
473 if (F1->hasSection() != F2->hasSection())
476 if (F1->hasSection() && F1->getSection() != F2->getSection())
479 if (F1->isVarArg() != F2->isVarArg())
482 // TODO: if it's internal and only used in direct calls, we could handle this
484 if (F1->getCallingConv() != F2->getCallingConv())
487 if (!isEquivalentType(F1->getFunctionType(), F2->getFunctionType()))
490 assert(F1->arg_size() == F2->arg_size() &&
491 "Identically typed functions have different numbers of args!");
493 // Visit the arguments so that they get enumerated in the order they're
495 for (Function::const_arg_iterator f1i = F1->arg_begin(),
496 f2i = F2->arg_begin(), f1e = F1->arg_end(); f1i != f1e; ++f1i, ++f2i) {
497 if (!enumerate(f1i, f2i))
498 llvm_unreachable("Arguments repeat!");
501 // We do a CFG-ordered walk since the actual ordering of the blocks in the
502 // linked list is immaterial. Our walk starts at the entry block for both
503 // functions, then takes each block from each terminator in order. As an
504 // artifact, this also means that unreachable blocks are ignored.
505 SmallVector<const BasicBlock *, 8> F1BBs, F2BBs;
506 SmallSet<const BasicBlock *, 128> VisitedBBs; // in terms of F1.
508 F1BBs.push_back(&F1->getEntryBlock());
509 F2BBs.push_back(&F2->getEntryBlock());
511 VisitedBBs.insert(F1BBs[0]);
512 while (!F1BBs.empty()) {
513 const BasicBlock *F1BB = F1BBs.pop_back_val();
514 const BasicBlock *F2BB = F2BBs.pop_back_val();
516 if (!enumerate(F1BB, F2BB) || !compare(F1BB, F2BB))
519 const TerminatorInst *F1TI = F1BB->getTerminator();
520 const TerminatorInst *F2TI = F2BB->getTerminator();
522 assert(F1TI->getNumSuccessors() == F2TI->getNumSuccessors());
523 for (unsigned i = 0, e = F1TI->getNumSuccessors(); i != e; ++i) {
524 if (!VisitedBBs.insert(F1TI->getSuccessor(i)))
527 F1BBs.push_back(F1TI->getSuccessor(i));
528 F2BBs.push_back(F2TI->getSuccessor(i));
536 /// MergeFunctions finds functions which will generate identical machine code,
537 /// by considering all pointer types to be equivalent. Once identified,
538 /// MergeFunctions will fold them by replacing a call to one to a call to a
539 /// bitcast of the other.
541 class MergeFunctions : public ModulePass {
545 : ModulePass(ID), HasGlobalAliases(false) {
546 initializeMergeFunctionsPass(*PassRegistry::getPassRegistry());
549 bool runOnModule(Module &M);
552 typedef DenseSet<ComparableFunction> FnSetType;
554 /// A work queue of functions that may have been modified and should be
556 std::vector<WeakVH> Deferred;
558 /// Insert a ComparableFunction into the FnSet, or merge it away if it's
559 /// equal to one that's already present.
560 bool insert(ComparableFunction &NewF);
562 /// Remove a Function from the FnSet and queue it up for a second sweep of
564 void remove(Function *F);
566 /// Find the functions that use this Value and remove them from FnSet and
567 /// queue the functions.
568 void removeUsers(Value *V);
570 /// Replace all direct calls of Old with calls of New. Will bitcast New if
571 /// necessary to make types match.
572 void replaceDirectCallers(Function *Old, Function *New);
574 /// Merge two equivalent functions. Upon completion, G may be deleted, or may
575 /// be converted into a thunk. In either case, it should never be visited
577 void mergeTwoFunctions(Function *F, Function *G);
579 /// Replace G with a thunk or an alias to F. Deletes G.
580 void writeThunkOrAlias(Function *F, Function *G);
582 /// Replace G with a simple tail call to bitcast(F). Also replace direct uses
583 /// of G with bitcast(F). Deletes G.
584 void writeThunk(Function *F, Function *G);
586 /// Replace G with an alias to F. Deletes G.
587 void writeAlias(Function *F, Function *G);
589 /// The set of all distinct functions. Use the insert() and remove() methods
593 /// DataLayout for more accurate GEP comparisons. May be NULL.
596 /// Whether or not the target supports global aliases.
597 bool HasGlobalAliases;
600 } // end anonymous namespace
602 char MergeFunctions::ID = 0;
603 INITIALIZE_PASS(MergeFunctions, "mergefunc", "Merge Functions", false, false)
605 ModulePass *llvm::createMergeFunctionsPass() {
606 return new MergeFunctions();
609 bool MergeFunctions::runOnModule(Module &M) {
610 bool Changed = false;
611 TD = getAnalysisIfAvailable<DataLayout>();
613 for (Module::iterator I = M.begin(), E = M.end(); I != E; ++I) {
614 if (!I->isDeclaration() && !I->hasAvailableExternallyLinkage())
615 Deferred.push_back(WeakVH(I));
617 FnSet.resize(Deferred.size());
620 std::vector<WeakVH> Worklist;
621 Deferred.swap(Worklist);
623 DEBUG(dbgs() << "size of module: " << M.size() << '\n');
624 DEBUG(dbgs() << "size of worklist: " << Worklist.size() << '\n');
626 // Insert only strong functions and merge them. Strong function merging
627 // always deletes one of them.
628 for (std::vector<WeakVH>::iterator I = Worklist.begin(),
629 E = Worklist.end(); I != E; ++I) {
631 Function *F = cast<Function>(*I);
632 if (!F->isDeclaration() && !F->hasAvailableExternallyLinkage() &&
633 !F->mayBeOverridden()) {
634 ComparableFunction CF = ComparableFunction(F, TD);
635 Changed |= insert(CF);
639 // Insert only weak functions and merge them. By doing these second we
640 // create thunks to the strong function when possible. When two weak
641 // functions are identical, we create a new strong function with two weak
642 // weak thunks to it which are identical but not mergable.
643 for (std::vector<WeakVH>::iterator I = Worklist.begin(),
644 E = Worklist.end(); I != E; ++I) {
646 Function *F = cast<Function>(*I);
647 if (!F->isDeclaration() && !F->hasAvailableExternallyLinkage() &&
648 F->mayBeOverridden()) {
649 ComparableFunction CF = ComparableFunction(F, TD);
650 Changed |= insert(CF);
653 DEBUG(dbgs() << "size of FnSet: " << FnSet.size() << '\n');
654 } while (!Deferred.empty());
661 bool DenseMapInfo<ComparableFunction>::isEqual(const ComparableFunction &LHS,
662 const ComparableFunction &RHS) {
663 if (LHS.getFunc() == RHS.getFunc() &&
664 LHS.getHash() == RHS.getHash())
666 if (!LHS.getFunc() || !RHS.getFunc())
669 // One of these is a special "underlying pointer comparison only" object.
670 if (LHS.getTD() == ComparableFunction::LookupOnly ||
671 RHS.getTD() == ComparableFunction::LookupOnly)
674 assert(LHS.getTD() == RHS.getTD() &&
675 "Comparing functions for different targets");
677 return FunctionComparator(LHS.getTD(), LHS.getFunc(),
678 RHS.getFunc()).compare();
681 // Replace direct callers of Old with New.
682 void MergeFunctions::replaceDirectCallers(Function *Old, Function *New) {
683 Constant *BitcastNew = ConstantExpr::getBitCast(New, Old->getType());
684 for (Value::use_iterator UI = Old->use_begin(), UE = Old->use_end();
686 Value::use_iterator TheIter = UI;
688 CallSite CS(*TheIter);
689 if (CS && CS.isCallee(TheIter)) {
690 remove(CS.getInstruction()->getParent()->getParent());
691 TheIter.getUse().set(BitcastNew);
696 // Replace G with an alias to F if possible, or else a thunk to F. Deletes G.
697 void MergeFunctions::writeThunkOrAlias(Function *F, Function *G) {
698 if (HasGlobalAliases && G->hasUnnamedAddr()) {
699 if (G->hasExternalLinkage() || G->hasLocalLinkage() ||
700 G->hasWeakLinkage()) {
709 // Replace G with a simple tail call to bitcast(F). Also replace direct uses
710 // of G with bitcast(F). Deletes G.
711 void MergeFunctions::writeThunk(Function *F, Function *G) {
712 if (!G->mayBeOverridden()) {
713 // Redirect direct callers of G to F.
714 replaceDirectCallers(G, F);
717 // If G was internal then we may have replaced all uses of G with F. If so,
718 // stop here and delete G. There's no need for a thunk.
719 if (G->hasLocalLinkage() && G->use_empty()) {
720 G->eraseFromParent();
724 Function *NewG = Function::Create(G->getFunctionType(), G->getLinkage(), "",
726 BasicBlock *BB = BasicBlock::Create(F->getContext(), "", NewG);
727 IRBuilder<false> Builder(BB);
729 SmallVector<Value *, 16> Args;
731 FunctionType *FFTy = F->getFunctionType();
732 for (Function::arg_iterator AI = NewG->arg_begin(), AE = NewG->arg_end();
734 Args.push_back(Builder.CreateBitCast(AI, FFTy->getParamType(i)));
738 CallInst *CI = Builder.CreateCall(F, Args);
740 CI->setCallingConv(F->getCallingConv());
741 if (NewG->getReturnType()->isVoidTy()) {
742 Builder.CreateRetVoid();
744 Builder.CreateRet(Builder.CreateBitCast(CI, NewG->getReturnType()));
747 NewG->copyAttributesFrom(G);
750 G->replaceAllUsesWith(NewG);
751 G->eraseFromParent();
753 DEBUG(dbgs() << "writeThunk: " << NewG->getName() << '\n');
757 // Replace G with an alias to F and delete G.
758 void MergeFunctions::writeAlias(Function *F, Function *G) {
759 Constant *BitcastF = ConstantExpr::getBitCast(F, G->getType());
760 GlobalAlias *GA = new GlobalAlias(G->getType(), G->getLinkage(), "",
761 BitcastF, G->getParent());
762 F->setAlignment(std::max(F->getAlignment(), G->getAlignment()));
764 GA->setVisibility(G->getVisibility());
766 G->replaceAllUsesWith(GA);
767 G->eraseFromParent();
769 DEBUG(dbgs() << "writeAlias: " << GA->getName() << '\n');
773 // Merge two equivalent functions. Upon completion, Function G is deleted.
774 void MergeFunctions::mergeTwoFunctions(Function *F, Function *G) {
775 if (F->mayBeOverridden()) {
776 assert(G->mayBeOverridden());
778 if (HasGlobalAliases) {
779 // Make them both thunks to the same internal function.
780 Function *H = Function::Create(F->getFunctionType(), F->getLinkage(), "",
782 H->copyAttributesFrom(F);
785 F->replaceAllUsesWith(H);
787 unsigned MaxAlignment = std::max(G->getAlignment(), H->getAlignment());
792 F->setAlignment(MaxAlignment);
793 F->setLinkage(GlobalValue::PrivateLinkage);
795 // We can't merge them. Instead, pick one and update all direct callers
796 // to call it and hope that we improve the instruction cache hit rate.
797 replaceDirectCallers(G, F);
802 writeThunkOrAlias(F, G);
805 ++NumFunctionsMerged;
808 // Insert a ComparableFunction into the FnSet, or merge it away if equal to one
809 // that was already inserted.
810 bool MergeFunctions::insert(ComparableFunction &NewF) {
811 std::pair<FnSetType::iterator, bool> Result = FnSet.insert(NewF);
813 DEBUG(dbgs() << "Inserting as unique: " << NewF.getFunc()->getName() << '\n');
817 const ComparableFunction &OldF = *Result.first;
819 // Never thunk a strong function to a weak function.
820 assert(!OldF.getFunc()->mayBeOverridden() ||
821 NewF.getFunc()->mayBeOverridden());
823 DEBUG(dbgs() << " " << OldF.getFunc()->getName() << " == "
824 << NewF.getFunc()->getName() << '\n');
826 Function *DeleteF = NewF.getFunc();
828 mergeTwoFunctions(OldF.getFunc(), DeleteF);
832 // Remove a function from FnSet. If it was already in FnSet, add it to Deferred
833 // so that we'll look at it in the next round.
834 void MergeFunctions::remove(Function *F) {
835 // We need to make sure we remove F, not a function "equal" to F per the
836 // function equality comparator.
838 // The special "lookup only" ComparableFunction bypasses the expensive
839 // function comparison in favour of a pointer comparison on the underlying
841 ComparableFunction CF = ComparableFunction(F, ComparableFunction::LookupOnly);
842 if (FnSet.erase(CF)) {
843 DEBUG(dbgs() << "Removed " << F->getName() << " from set and deferred it.\n");
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);