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. We can only fold two
21 // functions when we know that the definition of one of them is not
24 //===----------------------------------------------------------------------===//
28 // * fold vector<T*>::push_back and vector<S*>::push_back.
30 // These two functions have different types, but in a way that doesn't matter
31 // to us. As long as we never see an S or T itself, using S* and S** is the
32 // same as using a T* and T**.
34 // * virtual functions.
36 // Many functions have their address taken by the virtual function table for
37 // the object they belong to. However, as long as it's only used for a lookup
38 // and call, this is irrelevant, and we'd like to fold such implementations.
40 //===----------------------------------------------------------------------===//
42 #define DEBUG_TYPE "mergefunc"
43 #include "llvm/Transforms/IPO.h"
44 #include "llvm/ADT/DenseMap.h"
45 #include "llvm/ADT/FoldingSet.h"
46 #include "llvm/ADT/Statistic.h"
47 #include "llvm/Constants.h"
48 #include "llvm/InlineAsm.h"
49 #include "llvm/Instructions.h"
50 #include "llvm/Module.h"
51 #include "llvm/Pass.h"
52 #include "llvm/Support/CallSite.h"
53 #include "llvm/Support/Compiler.h"
54 #include "llvm/Support/Debug.h"
59 STATISTIC(NumFunctionsMerged, "Number of functions merged");
62 struct VISIBILITY_HIDDEN MergeFunctions : public ModulePass {
63 static char ID; // Pass identification, replacement for typeid
64 MergeFunctions() : ModulePass((intptr_t)&ID) {}
66 bool runOnModule(Module &M);
70 char MergeFunctions::ID = 0;
71 static RegisterPass<MergeFunctions>
72 X("mergefunc", "Merge Functions");
74 ModulePass *llvm::createMergeFunctionsPass() {
75 return new MergeFunctions();
78 // ===----------------------------------------------------------------------===
79 // Comparison of functions
80 // ===----------------------------------------------------------------------===
82 static unsigned long hash(const Function *F) {
83 const FunctionType *FTy = F->getFunctionType();
86 ID.AddInteger(F->size());
87 ID.AddInteger(F->getCallingConv());
88 ID.AddBoolean(F->hasGC());
89 ID.AddBoolean(FTy->isVarArg());
90 ID.AddInteger(FTy->getReturnType()->getTypeID());
91 for (unsigned i = 0, e = FTy->getNumParams(); i != e; ++i)
92 ID.AddInteger(FTy->getParamType(i)->getTypeID());
93 return ID.ComputeHash();
96 /// IgnoreBitcasts - given a bitcast, returns the first non-bitcast found by
97 /// walking the chain of cast operands. Otherwise, returns the argument.
98 static Value* IgnoreBitcasts(Value *V) {
99 while (BitCastInst *BC = dyn_cast<BitCastInst>(V))
100 V = BC->getOperand(0);
105 /// isEquivalentType - any two pointers are equivalent. Otherwise, standard
106 /// type equivalence rules apply.
107 static bool isEquivalentType(const Type *Ty1, const Type *Ty2) {
110 if (Ty1->getTypeID() != Ty2->getTypeID())
113 switch(Ty1->getTypeID()) {
115 case Type::FloatTyID:
116 case Type::DoubleTyID:
117 case Type::X86_FP80TyID:
118 case Type::FP128TyID:
119 case Type::PPC_FP128TyID:
120 case Type::LabelTyID:
121 case Type::MetadataTyID:
124 case Type::IntegerTyID:
125 case Type::OpaqueTyID:
126 // Ty1 == Ty2 would have returned true earlier.
130 assert(0 && "Unknown type!");
133 case Type::PointerTyID: {
134 const PointerType *PTy1 = cast<PointerType>(Ty1);
135 const PointerType *PTy2 = cast<PointerType>(Ty2);
136 return PTy1->getAddressSpace() == PTy2->getAddressSpace();
139 case Type::StructTyID: {
140 const StructType *STy1 = cast<StructType>(Ty1);
141 const StructType *STy2 = cast<StructType>(Ty2);
142 if (STy1->getNumElements() != STy2->getNumElements())
145 if (STy1->isPacked() != STy2->isPacked())
148 for (unsigned i = 0, e = STy1->getNumElements(); i != e; ++i) {
149 if (!isEquivalentType(STy1->getElementType(i), STy2->getElementType(i)))
155 case Type::FunctionTyID: {
156 const FunctionType *FTy1 = cast<FunctionType>(Ty1);
157 const FunctionType *FTy2 = cast<FunctionType>(Ty2);
158 if (FTy1->getNumParams() != FTy2->getNumParams() ||
159 FTy1->isVarArg() != FTy2->isVarArg())
162 if (!isEquivalentType(FTy1->getReturnType(), FTy2->getReturnType()))
165 for (unsigned i = 0, e = FTy1->getNumParams(); i != e; ++i) {
166 if (!isEquivalentType(FTy1->getParamType(i), FTy2->getParamType(i)))
172 case Type::ArrayTyID:
173 case Type::VectorTyID: {
174 const SequentialType *STy1 = cast<SequentialType>(Ty1);
175 const SequentialType *STy2 = cast<SequentialType>(Ty2);
176 return isEquivalentType(STy1->getElementType(), STy2->getElementType());
181 /// isEquivalentOperation - determine whether the two operations are the same
182 /// except that pointer-to-A and pointer-to-B are equivalent. This should be
183 /// kept in sync with Instruction::isSameOperandAs.
184 static bool isEquivalentOperation(const Instruction *I1, const Instruction *I2) {
185 if (I1->getOpcode() != I2->getOpcode() ||
186 I1->getNumOperands() != I2->getNumOperands() ||
187 !isEquivalentType(I1->getType(), I2->getType()))
190 // We have two instructions of identical opcode and #operands. Check to see
191 // if all operands are the same type
192 for (unsigned i = 0, e = I1->getNumOperands(); i != e; ++i)
193 if (!isEquivalentType(I1->getOperand(i)->getType(),
194 I2->getOperand(i)->getType()))
197 // Check special state that is a part of some instructions.
198 if (const LoadInst *LI = dyn_cast<LoadInst>(I1))
199 return LI->isVolatile() == cast<LoadInst>(I2)->isVolatile() &&
200 LI->getAlignment() == cast<LoadInst>(I2)->getAlignment();
201 if (const StoreInst *SI = dyn_cast<StoreInst>(I1))
202 return SI->isVolatile() == cast<StoreInst>(I2)->isVolatile() &&
203 SI->getAlignment() == cast<StoreInst>(I2)->getAlignment();
204 if (const CmpInst *CI = dyn_cast<CmpInst>(I1))
205 return CI->getPredicate() == cast<CmpInst>(I2)->getPredicate();
206 if (const CallInst *CI = dyn_cast<CallInst>(I1))
207 return CI->isTailCall() == cast<CallInst>(I2)->isTailCall() &&
208 CI->getCallingConv() == cast<CallInst>(I2)->getCallingConv() &&
209 CI->getAttributes().getRawPointer() ==
210 cast<CallInst>(I2)->getAttributes().getRawPointer();
211 if (const InvokeInst *CI = dyn_cast<InvokeInst>(I1))
212 return CI->getCallingConv() == cast<InvokeInst>(I2)->getCallingConv() &&
213 CI->getAttributes().getRawPointer() ==
214 cast<InvokeInst>(I2)->getAttributes().getRawPointer();
215 if (const InsertValueInst *IVI = dyn_cast<InsertValueInst>(I1)) {
216 if (IVI->getNumIndices() != cast<InsertValueInst>(I2)->getNumIndices())
218 for (unsigned i = 0, e = IVI->getNumIndices(); i != e; ++i)
219 if (IVI->idx_begin()[i] != cast<InsertValueInst>(I2)->idx_begin()[i])
223 if (const ExtractValueInst *EVI = dyn_cast<ExtractValueInst>(I1)) {
224 if (EVI->getNumIndices() != cast<ExtractValueInst>(I2)->getNumIndices())
226 for (unsigned i = 0, e = EVI->getNumIndices(); i != e; ++i)
227 if (EVI->idx_begin()[i] != cast<ExtractValueInst>(I2)->idx_begin()[i])
235 static bool compare(const Value *V, const Value *U) {
236 assert(!isa<BasicBlock>(V) && !isa<BasicBlock>(U) &&
237 "Must not compare basic blocks.");
239 assert(isEquivalentType(V->getType(), U->getType()) &&
240 "Two of the same operation have operands of different type.");
242 // TODO: If the constant is an expression of F, we should accept that it's
243 // equal to the same expression in terms of G.
244 if (isa<Constant>(V))
247 // The caller has ensured that ValueMap[V] != U. Since Arguments are
248 // pre-loaded into the ValueMap, and Instructions are added as we go, we know
249 // that this can only be a mis-match.
250 if (isa<Instruction>(V) || isa<Argument>(V))
253 if (isa<InlineAsm>(V) && isa<InlineAsm>(U)) {
254 const InlineAsm *IAF = cast<InlineAsm>(V);
255 const InlineAsm *IAG = cast<InlineAsm>(U);
256 return IAF->getAsmString() == IAG->getAsmString() &&
257 IAF->getConstraintString() == IAG->getConstraintString();
263 static bool equals(const BasicBlock *BB1, const BasicBlock *BB2,
264 DenseMap<const Value *, const Value *> &ValueMap,
265 DenseMap<const Value *, const Value *> &SpeculationMap) {
266 // Speculatively add it anyways. If it's false, we'll notice a difference
267 // later, and this won't matter.
270 BasicBlock::const_iterator FI = BB1->begin(), FE = BB1->end();
271 BasicBlock::const_iterator GI = BB2->begin(), GE = BB2->end();
274 if (isa<BitCastInst>(FI)) {
278 if (isa<BitCastInst>(GI)) {
283 if (!isEquivalentOperation(FI, GI))
286 if (ValueMap[FI] == GI) {
291 if (ValueMap[FI] != NULL)
294 for (unsigned i = 0, e = FI->getNumOperands(); i != e; ++i) {
295 Value *OpF = IgnoreBitcasts(FI->getOperand(i));
296 Value *OpG = IgnoreBitcasts(GI->getOperand(i));
298 if (ValueMap[OpF] == OpG)
301 if (ValueMap[OpF] != NULL)
304 if (OpF->getValueID() != OpG->getValueID() ||
305 !isEquivalentType(OpF->getType(), OpG->getType()))
308 if (isa<PHINode>(FI)) {
309 if (SpeculationMap[OpF] == NULL)
310 SpeculationMap[OpF] = OpG;
311 else if (SpeculationMap[OpF] != OpG)
314 } else if (isa<BasicBlock>(OpF)) {
315 assert(isa<TerminatorInst>(FI) &&
316 "BasicBlock referenced by non-Terminator non-PHI");
317 // This call changes the ValueMap, hence we can't use
318 // Value *& = ValueMap[...]
319 if (!equals(cast<BasicBlock>(OpF), cast<BasicBlock>(OpG), ValueMap,
323 if (!compare(OpF, OpG))
332 } while (FI != FE && GI != GE);
334 return FI == FE && GI == GE;
337 static bool equals(const Function *F, const Function *G) {
338 // We need to recheck everything, but check the things that weren't included
339 // in the hash first.
341 if (F->getAttributes() != G->getAttributes())
344 if (F->hasGC() != G->hasGC())
347 if (F->hasGC() && F->getGC() != G->getGC())
350 if (F->hasSection() != G->hasSection())
353 if (F->hasSection() && F->getSection() != G->getSection())
356 if (F->isVarArg() != G->isVarArg())
359 // TODO: if it's internal and only used in direct calls, we could handle this
361 if (F->getCallingConv() != G->getCallingConv())
364 if (!isEquivalentType(F->getFunctionType(), G->getFunctionType()))
367 DenseMap<const Value *, const Value *> ValueMap;
368 DenseMap<const Value *, const Value *> SpeculationMap;
371 assert(F->arg_size() == G->arg_size() &&
372 "Identical functions have a different number of args.");
374 for (Function::const_arg_iterator fi = F->arg_begin(), gi = G->arg_begin(),
375 fe = F->arg_end(); fi != fe; ++fi, ++gi)
378 if (!equals(&F->getEntryBlock(), &G->getEntryBlock(), ValueMap,
382 for (DenseMap<const Value *, const Value *>::iterator
383 I = SpeculationMap.begin(), E = SpeculationMap.end(); I != E; ++I) {
384 if (ValueMap[I->first] != I->second)
391 // ===----------------------------------------------------------------------===
392 // Folding of functions
393 // ===----------------------------------------------------------------------===
396 // * F is external strong, G is external strong:
397 // turn G into a thunk to F (1)
398 // * F is external strong, G is external weak:
399 // turn G into a thunk to F (1)
400 // * F is external weak, G is external weak:
402 // * F is external strong, G is internal:
403 // address of G taken:
404 // turn G into a thunk to F (1)
405 // address of G not taken:
406 // make G an alias to F (2)
407 // * F is internal, G is external weak
408 // address of F is taken:
409 // turn G into a thunk to F (1)
410 // address of F is not taken:
411 // make G an alias of F (2)
412 // * F is internal, G is internal:
413 // address of F and G are taken:
414 // turn G into a thunk to F (1)
415 // address of G is not taken:
416 // make G an alias to F (2)
418 // alias requires linkage == (external,local,weak) fallback to creating a thunk
419 // external means 'externally visible' linkage != (internal,private)
420 // internal means linkage == (internal,private)
421 // weak means linkage mayBeOverridable
422 // being external implies that the address is taken
424 // 1. turn G into a thunk to F
425 // 2. make G an alias to F
427 enum LinkageCategory {
433 static LinkageCategory categorize(const Function *F) {
434 switch (F->getLinkage()) {
435 case GlobalValue::InternalLinkage:
436 case GlobalValue::PrivateLinkage:
439 case GlobalValue::WeakAnyLinkage:
440 case GlobalValue::WeakODRLinkage:
441 case GlobalValue::ExternalWeakLinkage:
444 case GlobalValue::ExternalLinkage:
445 case GlobalValue::AvailableExternallyLinkage:
446 case GlobalValue::LinkOnceAnyLinkage:
447 case GlobalValue::LinkOnceODRLinkage:
448 case GlobalValue::AppendingLinkage:
449 case GlobalValue::DLLImportLinkage:
450 case GlobalValue::DLLExportLinkage:
451 case GlobalValue::GhostLinkage:
452 case GlobalValue::CommonLinkage:
453 return ExternalStrong;
456 assert(0 && "Unknown LinkageType.");
460 static void ThunkGToF(Function *F, Function *G) {
461 Function *NewG = Function::Create(G->getFunctionType(), G->getLinkage(), "",
463 BasicBlock *BB = BasicBlock::Create("", NewG);
465 std::vector<Value *> Args;
467 const FunctionType *FFTy = F->getFunctionType();
468 for (Function::arg_iterator AI = NewG->arg_begin(), AE = NewG->arg_end();
470 if (FFTy->getParamType(i) == AI->getType())
473 Value *BCI = new BitCastInst(AI, FFTy->getParamType(i), "", BB);
479 CallInst *CI = CallInst::Create(F, Args.begin(), Args.end(), "", BB);
481 CI->setCallingConv(F->getCallingConv());
482 if (NewG->getReturnType() == Type::VoidTy) {
483 ReturnInst::Create(BB);
484 } else if (CI->getType() != NewG->getReturnType()) {
485 Value *BCI = new BitCastInst(CI, NewG->getReturnType(), "", BB);
486 ReturnInst::Create(BCI, BB);
488 ReturnInst::Create(CI, BB);
491 NewG->copyAttributesFrom(G);
493 G->replaceAllUsesWith(NewG);
494 G->eraseFromParent();
496 // TODO: look at direct callers to G and make them all direct callers to F.
499 static void AliasGToF(Function *F, Function *G) {
500 if (!G->hasExternalLinkage() && !G->hasLocalLinkage() && !G->hasWeakLinkage())
501 return ThunkGToF(F, G);
503 GlobalAlias *GA = new GlobalAlias(
504 G->getType(), G->getLinkage(), "",
505 ConstantExpr::getBitCast(F, G->getType()), G->getParent());
506 F->setAlignment(std::max(F->getAlignment(), G->getAlignment()));
508 GA->setVisibility(G->getVisibility());
509 G->replaceAllUsesWith(GA);
510 G->eraseFromParent();
513 static bool fold(std::vector<Function *> &FnVec, unsigned i, unsigned j) {
514 Function *F = FnVec[i];
515 Function *G = FnVec[j];
517 LinkageCategory catF = categorize(F);
518 LinkageCategory catG = categorize(G);
520 if (catF == ExternalWeak || (catF == Internal && catG == ExternalStrong)) {
521 std::swap(FnVec[i], FnVec[j]);
523 std::swap(catF, catG);
534 if (G->hasAddressTaken())
543 assert(catG == ExternalWeak);
545 // Make them both thunks to the same internal function.
546 F->setAlignment(std::max(F->getAlignment(), G->getAlignment()));
547 Function *H = Function::Create(F->getFunctionType(), F->getLinkage(), "",
549 H->copyAttributesFrom(F);
551 F->replaceAllUsesWith(H);
556 F->setLinkage(GlobalValue::InternalLinkage);
565 if (F->hasAddressTaken())
571 bool addrTakenF = F->hasAddressTaken();
572 bool addrTakenG = G->hasAddressTaken();
573 if (!addrTakenF && addrTakenG) {
574 std::swap(FnVec[i], FnVec[j]);
576 std::swap(addrTakenF, addrTakenG);
579 if (addrTakenF && addrTakenG) {
590 ++NumFunctionsMerged;
594 // ===----------------------------------------------------------------------===
596 // ===----------------------------------------------------------------------===
598 bool MergeFunctions::runOnModule(Module &M) {
599 bool Changed = false;
601 std::map<unsigned long, std::vector<Function *> > FnMap;
603 for (Module::iterator F = M.begin(), E = M.end(); F != E; ++F) {
604 if (F->isDeclaration() || F->isIntrinsic())
607 FnMap[hash(F)].push_back(F);
610 // TODO: instead of running in a loop, we could also fold functions in
611 // callgraph order. Constructing the CFG probably isn't cheaper than just
612 // running in a loop, unless it happened to already be available.
616 LocalChanged = false;
617 DOUT << "size: " << FnMap.size() << "\n";
618 for (std::map<unsigned long, std::vector<Function *> >::iterator
619 I = FnMap.begin(), E = FnMap.end(); I != E; ++I) {
620 std::vector<Function *> &FnVec = I->second;
621 DOUT << "hash (" << I->first << "): " << FnVec.size() << "\n";
623 for (int i = 0, e = FnVec.size(); i != e; ++i) {
624 for (int j = i + 1; j != e; ++j) {
625 bool isEqual = equals(FnVec[i], FnVec[j]);
627 DOUT << " " << FnVec[i]->getName()
628 << (isEqual ? " == " : " != ")
629 << FnVec[j]->getName() << "\n";
632 if (fold(FnVec, i, j)) {
634 FnVec.erase(FnVec.begin() + j);
642 Changed |= LocalChanged;
643 } while (LocalChanged);