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/DenseMap.h"
49 #include "llvm/ADT/FoldingSet.h"
50 #include "llvm/ADT/SmallSet.h"
51 #include "llvm/ADT/Statistic.h"
52 #include "llvm/Constants.h"
53 #include "llvm/InlineAsm.h"
54 #include "llvm/Instructions.h"
55 #include "llvm/LLVMContext.h"
56 #include "llvm/Module.h"
57 #include "llvm/Pass.h"
58 #include "llvm/Support/CallSite.h"
59 #include "llvm/Support/Debug.h"
60 #include "llvm/Support/ErrorHandling.h"
61 #include "llvm/Support/IRBuilder.h"
62 #include "llvm/Support/raw_ostream.h"
63 #include "llvm/Target/TargetData.h"
68 STATISTIC(NumFunctionsMerged, "Number of functions merged");
71 /// MergeFunctions finds functions which will generate identical machine code,
72 /// by considering all pointer types to be equivalent. Once identified,
73 /// MergeFunctions will fold them by replacing a call to one to a call to a
74 /// bitcast of the other.
76 class MergeFunctions : public ModulePass {
79 MergeFunctions() : ModulePass(ID) {}
81 bool runOnModule(Module &M);
84 /// PairwiseCompareAndMerge - Given a list of functions, compare each pair
85 /// and merge the pairs of equivalent functions.
86 bool PairwiseCompareAndMerge(std::vector<Function *> &FnVec);
88 /// MergeTwoFunctions - Merge two equivalent functions. Upon completion,
89 /// FnVec[j] should never be visited again.
90 void MergeTwoFunctions(std::vector<Function *> &FnVec,
91 unsigned i, unsigned j) const;
93 /// WriteThunk - Replace G with a simple tail call to bitcast(F). Also
94 /// replace direct uses of G with bitcast(F).
95 void WriteThunk(Function *F, Function *G) const;
101 char MergeFunctions::ID = 0;
102 INITIALIZE_PASS(MergeFunctions, "mergefunc", "Merge Functions", false, false);
104 ModulePass *llvm::createMergeFunctionsPass() {
105 return new MergeFunctions();
109 /// FunctionComparator - Compares two functions to determine whether or not
110 /// they will generate machine code with the same behaviour. TargetData is
111 /// used if available. The comparator always fails conservatively (erring on the
112 /// side of claiming that two functions are different).
113 class FunctionComparator {
115 FunctionComparator(TargetData *TD, Function *F1, Function *F2)
116 : F1(F1), F2(F2), TD(TD), IDMap1Count(0), IDMap2Count(0) {}
118 /// Compare - test whether the two functions have equivalent behaviour.
122 /// Compare - test whether two basic blocks have equivalent behaviour.
123 bool Compare(const BasicBlock *BB1, const BasicBlock *BB2);
125 /// Enumerate - Assign or look up previously assigned numbers for the two
126 /// values, and return whether the numbers are equal. Numbers are assigned in
127 /// the order visited.
128 bool Enumerate(const Value *V1, const Value *V2);
130 /// isEquivalentOperation - Compare two Instructions for equivalence, similar
131 /// to Instruction::isSameOperationAs but with modifications to the type
133 bool isEquivalentOperation(const Instruction *I1,
134 const Instruction *I2) const;
136 /// isEquivalentGEP - Compare two GEPs for equivalent pointer arithmetic.
137 bool isEquivalentGEP(const GEPOperator *GEP1, const GEPOperator *GEP2);
138 bool isEquivalentGEP(const GetElementPtrInst *GEP1,
139 const GetElementPtrInst *GEP2) {
140 return isEquivalentGEP(cast<GEPOperator>(GEP1), cast<GEPOperator>(GEP2));
143 /// isEquivalentType - Compare two Types, treating all pointer types as equal.
144 bool isEquivalentType(const Type *Ty1, const Type *Ty2) const;
146 // The two functions undergoing comparison.
151 typedef DenseMap<const Value *, unsigned long> IDMap;
153 unsigned long IDMap1Count, IDMap2Count;
157 /// Compute a hash guaranteed to be equal for two equivalent functions, but
158 /// very likely to be different for different functions.
159 static unsigned long ProfileFunction(const Function *F) {
160 const FunctionType *FTy = F->getFunctionType();
163 ID.AddInteger(F->size());
164 ID.AddInteger(F->getCallingConv());
165 ID.AddBoolean(F->hasGC());
166 ID.AddBoolean(FTy->isVarArg());
167 ID.AddInteger(FTy->getReturnType()->getTypeID());
168 for (unsigned i = 0, e = FTy->getNumParams(); i != e; ++i)
169 ID.AddInteger(FTy->getParamType(i)->getTypeID());
170 return ID.ComputeHash();
173 /// isEquivalentType - any two pointers in the same address space are
174 /// equivalent. Otherwise, standard type equivalence rules apply.
175 bool FunctionComparator::isEquivalentType(const Type *Ty1,
176 const Type *Ty2) const {
179 if (Ty1->getTypeID() != Ty2->getTypeID())
182 switch(Ty1->getTypeID()) {
184 llvm_unreachable("Unknown type!");
185 // Fall through in Release mode.
186 case Type::IntegerTyID:
187 case Type::OpaqueTyID:
188 // Ty1 == Ty2 would have returned true earlier.
192 case Type::FloatTyID:
193 case Type::DoubleTyID:
194 case Type::X86_FP80TyID:
195 case Type::FP128TyID:
196 case Type::PPC_FP128TyID:
197 case Type::LabelTyID:
198 case Type::MetadataTyID:
201 case Type::PointerTyID: {
202 const PointerType *PTy1 = cast<PointerType>(Ty1);
203 const PointerType *PTy2 = cast<PointerType>(Ty2);
204 return PTy1->getAddressSpace() == PTy2->getAddressSpace();
207 case Type::StructTyID: {
208 const StructType *STy1 = cast<StructType>(Ty1);
209 const StructType *STy2 = cast<StructType>(Ty2);
210 if (STy1->getNumElements() != STy2->getNumElements())
213 if (STy1->isPacked() != STy2->isPacked())
216 for (unsigned i = 0, e = STy1->getNumElements(); i != e; ++i) {
217 if (!isEquivalentType(STy1->getElementType(i), STy2->getElementType(i)))
223 case Type::FunctionTyID: {
224 const FunctionType *FTy1 = cast<FunctionType>(Ty1);
225 const FunctionType *FTy2 = cast<FunctionType>(Ty2);
226 if (FTy1->getNumParams() != FTy2->getNumParams() ||
227 FTy1->isVarArg() != FTy2->isVarArg())
230 if (!isEquivalentType(FTy1->getReturnType(), FTy2->getReturnType()))
233 for (unsigned i = 0, e = FTy1->getNumParams(); i != e; ++i) {
234 if (!isEquivalentType(FTy1->getParamType(i), FTy2->getParamType(i)))
240 case Type::ArrayTyID: {
241 const ArrayType *ATy1 = cast<ArrayType>(Ty1);
242 const ArrayType *ATy2 = cast<ArrayType>(Ty2);
243 return ATy1->getNumElements() == ATy2->getNumElements() &&
244 isEquivalentType(ATy1->getElementType(), ATy2->getElementType());
247 case Type::VectorTyID: {
248 const VectorType *VTy1 = cast<VectorType>(Ty1);
249 const VectorType *VTy2 = cast<VectorType>(Ty2);
250 return VTy1->getNumElements() == VTy2->getNumElements() &&
251 isEquivalentType(VTy1->getElementType(), VTy2->getElementType());
256 /// isEquivalentOperation - determine whether the two operations are the same
257 /// except that pointer-to-A and pointer-to-B are equivalent. This should be
258 /// kept in sync with Instruction::isSameOperationAs.
259 bool FunctionComparator::isEquivalentOperation(const Instruction *I1,
260 const Instruction *I2) const {
261 if (I1->getOpcode() != I2->getOpcode() ||
262 I1->getNumOperands() != I2->getNumOperands() ||
263 !isEquivalentType(I1->getType(), I2->getType()) ||
264 !I1->hasSameSubclassOptionalData(I2))
267 // We have two instructions of identical opcode and #operands. Check to see
268 // if all operands are the same type
269 for (unsigned i = 0, e = I1->getNumOperands(); i != e; ++i)
270 if (!isEquivalentType(I1->getOperand(i)->getType(),
271 I2->getOperand(i)->getType()))
274 // Check special state that is a part of some instructions.
275 if (const LoadInst *LI = dyn_cast<LoadInst>(I1))
276 return LI->isVolatile() == cast<LoadInst>(I2)->isVolatile() &&
277 LI->getAlignment() == cast<LoadInst>(I2)->getAlignment();
278 if (const StoreInst *SI = dyn_cast<StoreInst>(I1))
279 return SI->isVolatile() == cast<StoreInst>(I2)->isVolatile() &&
280 SI->getAlignment() == cast<StoreInst>(I2)->getAlignment();
281 if (const CmpInst *CI = dyn_cast<CmpInst>(I1))
282 return CI->getPredicate() == cast<CmpInst>(I2)->getPredicate();
283 if (const CallInst *CI = dyn_cast<CallInst>(I1))
284 return CI->isTailCall() == cast<CallInst>(I2)->isTailCall() &&
285 CI->getCallingConv() == cast<CallInst>(I2)->getCallingConv() &&
286 CI->getAttributes().getRawPointer() ==
287 cast<CallInst>(I2)->getAttributes().getRawPointer();
288 if (const InvokeInst *CI = dyn_cast<InvokeInst>(I1))
289 return CI->getCallingConv() == cast<InvokeInst>(I2)->getCallingConv() &&
290 CI->getAttributes().getRawPointer() ==
291 cast<InvokeInst>(I2)->getAttributes().getRawPointer();
292 if (const InsertValueInst *IVI = dyn_cast<InsertValueInst>(I1)) {
293 if (IVI->getNumIndices() != cast<InsertValueInst>(I2)->getNumIndices())
295 for (unsigned i = 0, e = IVI->getNumIndices(); i != e; ++i)
296 if (IVI->idx_begin()[i] != cast<InsertValueInst>(I2)->idx_begin()[i])
300 if (const ExtractValueInst *EVI = dyn_cast<ExtractValueInst>(I1)) {
301 if (EVI->getNumIndices() != cast<ExtractValueInst>(I2)->getNumIndices())
303 for (unsigned i = 0, e = EVI->getNumIndices(); i != e; ++i)
304 if (EVI->idx_begin()[i] != cast<ExtractValueInst>(I2)->idx_begin()[i])
312 /// isEquivalentGEP - determine whether two GEP operations perform the same
313 /// underlying arithmetic.
314 bool FunctionComparator::isEquivalentGEP(const GEPOperator *GEP1,
315 const GEPOperator *GEP2) {
316 // When we have target data, we can reduce the GEP down to the value in bytes
317 // added to the address.
318 if (TD && GEP1->hasAllConstantIndices() && GEP2->hasAllConstantIndices()) {
319 SmallVector<Value *, 8> Indices1(GEP1->idx_begin(), GEP1->idx_end());
320 SmallVector<Value *, 8> Indices2(GEP2->idx_begin(), GEP2->idx_end());
321 uint64_t Offset1 = TD->getIndexedOffset(GEP1->getPointerOperandType(),
322 Indices1.data(), Indices1.size());
323 uint64_t Offset2 = TD->getIndexedOffset(GEP2->getPointerOperandType(),
324 Indices2.data(), Indices2.size());
325 return Offset1 == Offset2;
328 if (GEP1->getPointerOperand()->getType() !=
329 GEP2->getPointerOperand()->getType())
332 if (GEP1->getNumOperands() != GEP2->getNumOperands())
335 for (unsigned i = 0, e = GEP1->getNumOperands(); i != e; ++i) {
336 if (!Enumerate(GEP1->getOperand(i), GEP2->getOperand(i)))
343 /// Enumerate - Compare two values used by the two functions under pair-wise
344 /// comparison. If this is the first time the values are seen, they're added to
345 /// the mapping so that we will detect mismatches on next use.
346 bool FunctionComparator::Enumerate(const Value *V1, const Value *V2) {
347 // Check for function @f1 referring to itself and function @f2 referring to
348 // itself, or referring to each other, or both referring to either of them.
349 // They're all equivalent if the two functions are otherwise equivalent.
350 if (V1 == F1 && V2 == F2)
352 if (V1 == F2 && V2 == F1)
355 // TODO: constant expressions with GEP or references to F1 or F2.
356 if (isa<Constant>(V1))
359 if (isa<InlineAsm>(V1) && isa<InlineAsm>(V2)) {
360 const InlineAsm *IA1 = cast<InlineAsm>(V1);
361 const InlineAsm *IA2 = cast<InlineAsm>(V2);
362 return IA1->getAsmString() == IA2->getAsmString() &&
363 IA1->getConstraintString() == IA2->getConstraintString();
366 unsigned long &ID1 = Map1[V1];
370 unsigned long &ID2 = Map2[V2];
377 /// Compare - test whether two basic blocks have equivalent behaviour.
378 bool FunctionComparator::Compare(const BasicBlock *BB1, const BasicBlock *BB2) {
379 BasicBlock::const_iterator F1I = BB1->begin(), F1E = BB1->end();
380 BasicBlock::const_iterator F2I = BB2->begin(), F2E = BB2->end();
383 if (!Enumerate(F1I, F2I))
386 if (const GetElementPtrInst *GEP1 = dyn_cast<GetElementPtrInst>(F1I)) {
387 const GetElementPtrInst *GEP2 = dyn_cast<GetElementPtrInst>(F2I);
391 if (!Enumerate(GEP1->getPointerOperand(), GEP2->getPointerOperand()))
394 if (!isEquivalentGEP(GEP1, GEP2))
397 if (!isEquivalentOperation(F1I, F2I))
400 assert(F1I->getNumOperands() == F2I->getNumOperands());
401 for (unsigned i = 0, e = F1I->getNumOperands(); i != e; ++i) {
402 Value *OpF1 = F1I->getOperand(i);
403 Value *OpF2 = F2I->getOperand(i);
405 if (!Enumerate(OpF1, OpF2))
408 if (OpF1->getValueID() != OpF2->getValueID() ||
409 !isEquivalentType(OpF1->getType(), OpF2->getType()))
415 } while (F1I != F1E && F2I != F2E);
417 return F1I == F1E && F2I == F2E;
420 /// Compare - test whether the two functions have equivalent behaviour.
421 bool FunctionComparator::Compare() {
422 // We need to recheck everything, but check the things that weren't included
423 // in the hash first.
425 if (F1->getAttributes() != F2->getAttributes())
428 if (F1->hasGC() != F2->hasGC())
431 if (F1->hasGC() && F1->getGC() != F2->getGC())
434 if (F1->hasSection() != F2->hasSection())
437 if (F1->hasSection() && F1->getSection() != F2->getSection())
440 if (F1->isVarArg() != F2->isVarArg())
443 // TODO: if it's internal and only used in direct calls, we could handle this
445 if (F1->getCallingConv() != F2->getCallingConv())
448 if (!isEquivalentType(F1->getFunctionType(), F2->getFunctionType()))
451 assert(F1->arg_size() == F2->arg_size() &&
452 "Identical functions have a different number of args.");
454 // Visit the arguments so that they get enumerated in the order they're
456 for (Function::const_arg_iterator f1i = F1->arg_begin(),
457 f2i = F2->arg_begin(), f1e = F1->arg_end(); f1i != f1e; ++f1i, ++f2i) {
458 if (!Enumerate(f1i, f2i))
459 llvm_unreachable("Arguments repeat");
462 // We do a CFG-ordered walk since the actual ordering of the blocks in the
463 // linked list is immaterial. Our walk starts at the entry block for both
464 // functions, then takes each block from each terminator in order. As an
465 // artifact, this also means that unreachable blocks are ignored.
466 SmallVector<const BasicBlock *, 8> F1BBs, F2BBs;
467 SmallSet<const BasicBlock *, 128> VisitedBBs; // in terms of F1.
469 F1BBs.push_back(&F1->getEntryBlock());
470 F2BBs.push_back(&F2->getEntryBlock());
472 VisitedBBs.insert(F1BBs[0]);
473 while (!F1BBs.empty()) {
474 const BasicBlock *F1BB = F1BBs.pop_back_val();
475 const BasicBlock *F2BB = F2BBs.pop_back_val();
477 if (!Enumerate(F1BB, F2BB) || !Compare(F1BB, F2BB))
480 const TerminatorInst *F1TI = F1BB->getTerminator();
481 const TerminatorInst *F2TI = F2BB->getTerminator();
483 assert(F1TI->getNumSuccessors() == F2TI->getNumSuccessors());
484 for (unsigned i = 0, e = F1TI->getNumSuccessors(); i != e; ++i) {
485 if (!VisitedBBs.insert(F1TI->getSuccessor(i)))
488 F1BBs.push_back(F1TI->getSuccessor(i));
489 F2BBs.push_back(F2TI->getSuccessor(i));
495 /// WriteThunk - Replace G with a simple tail call to bitcast(F). Also replace
496 /// direct uses of G with bitcast(F).
497 void MergeFunctions::WriteThunk(Function *F, Function *G) const {
498 if (!G->mayBeOverridden()) {
499 // Redirect direct callers of G to F.
500 Constant *BitcastF = ConstantExpr::getBitCast(F, G->getType());
501 for (Value::use_iterator UI = G->use_begin(), UE = G->use_end();
503 Value::use_iterator TheIter = UI;
505 CallSite CS(*TheIter);
506 if (CS && CS.isCallee(TheIter))
507 TheIter.getUse().set(BitcastF);
511 // If G was internal then we may have replaced all uses if G with F. If so,
512 // stop here and delete G. There's no need for a thunk.
513 if (G->hasLocalLinkage() && G->use_empty()) {
514 G->eraseFromParent();
518 Function *NewG = Function::Create(G->getFunctionType(), G->getLinkage(), "",
520 BasicBlock *BB = BasicBlock::Create(F->getContext(), "", NewG);
521 IRBuilder<false> Builder(BB);
523 SmallVector<Value *, 16> Args;
525 const FunctionType *FFTy = F->getFunctionType();
526 for (Function::arg_iterator AI = NewG->arg_begin(), AE = NewG->arg_end();
528 Args.push_back(Builder.CreateBitCast(AI, FFTy->getParamType(i)));
532 CallInst *CI = Builder.CreateCall(F, Args.begin(), Args.end());
534 CI->setCallingConv(F->getCallingConv());
535 if (NewG->getReturnType()->isVoidTy()) {
536 Builder.CreateRetVoid();
538 Builder.CreateRet(Builder.CreateBitCast(CI, NewG->getReturnType()));
541 NewG->copyAttributesFrom(G);
543 G->replaceAllUsesWith(NewG);
544 G->eraseFromParent();
547 /// MergeTwoFunctions - Merge two equivalent functions. Upon completion,
548 /// FnVec[j] is deleted but not removed from the vector.
549 void MergeFunctions::MergeTwoFunctions(std::vector<Function *> &FnVec,
550 unsigned i, unsigned j) const {
551 Function *F = FnVec[i];
552 Function *G = FnVec[j];
554 if (F->isWeakForLinker() && !G->isWeakForLinker()) {
555 std::swap(FnVec[i], FnVec[j]);
559 if (F->isWeakForLinker()) {
560 assert(G->isWeakForLinker());
562 // Make them both thunks to the same internal function.
563 Function *H = Function::Create(F->getFunctionType(), F->getLinkage(), "",
565 H->copyAttributesFrom(F);
567 F->replaceAllUsesWith(H);
569 unsigned MaxAlignment = std::max(G->getAlignment(), H->getAlignment());
574 F->setAlignment(MaxAlignment);
575 F->setLinkage(GlobalValue::InternalLinkage);
580 ++NumFunctionsMerged;
583 /// PairwiseCompareAndMerge - Given a list of functions, compare each pair and
584 /// merge the pairs of equivalent functions.
585 bool MergeFunctions::PairwiseCompareAndMerge(std::vector<Function *> &FnVec) {
586 bool Changed = false;
587 for (int i = 0, e = FnVec.size(); i != e; ++i) {
588 for (int j = i + 1; j != e; ++j) {
589 bool isEqual = FunctionComparator(TD, FnVec[i], FnVec[j]).Compare();
591 DEBUG(dbgs() << " " << FnVec[i]->getName()
592 << (isEqual ? " == " : " != ") << FnVec[j]->getName() << "\n");
595 MergeTwoFunctions(FnVec, i, j);
597 FnVec.erase(FnVec.begin() + j);
605 bool MergeFunctions::runOnModule(Module &M) {
606 bool Changed = false;
608 std::map<unsigned long, std::vector<Function *> > FnMap;
610 for (Module::iterator F = M.begin(), E = M.end(); F != E; ++F) {
611 if (F->isDeclaration() || F->hasAvailableExternallyLinkage())
614 FnMap[ProfileFunction(F)].push_back(F);
617 TD = getAnalysisIfAvailable<TargetData>();
621 LocalChanged = false;
622 DEBUG(dbgs() << "size: " << FnMap.size() << "\n");
623 for (std::map<unsigned long, std::vector<Function *> >::iterator
624 I = FnMap.begin(), E = FnMap.end(); I != E; ++I) {
625 std::vector<Function *> &FnVec = I->second;
626 DEBUG(dbgs() << "hash (" << I->first << "): " << FnVec.size() << "\n");
627 LocalChanged |= PairwiseCompareAndMerge(FnVec);
629 Changed |= LocalChanged;
630 } while (LocalChanged);