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"
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 /// MergeTwoFunctions - Merge two equivalent functions. Upon completion, G
86 void MergeTwoFunctions(Function *F, Function *G) const;
88 /// WriteThunk - Replace G with a simple tail call to bitcast(F). Also
89 /// replace direct uses of G with bitcast(F).
90 void WriteThunk(Function *F, Function *G) const;
96 char MergeFunctions::ID = 0;
97 INITIALIZE_PASS(MergeFunctions, "mergefunc", "Merge Functions", false, false);
99 ModulePass *llvm::createMergeFunctionsPass() {
100 return new MergeFunctions();
104 /// FunctionComparator - Compares two functions to determine whether or not
105 /// they will generate machine code with the same behaviour. TargetData is
106 /// used if available. The comparator always fails conservatively (erring on the
107 /// side of claiming that two functions are different).
108 class FunctionComparator {
110 FunctionComparator(const TargetData *TD, const Function *F1,
112 : F1(F1), F2(F2), TD(TD), IDMap1Count(0), IDMap2Count(0) {}
114 /// Compare - test whether the two functions have equivalent behaviour.
118 /// Compare - test whether two basic blocks have equivalent behaviour.
119 bool Compare(const BasicBlock *BB1, const BasicBlock *BB2);
121 /// Enumerate - Assign or look up previously assigned numbers for the two
122 /// values, and return whether the numbers are equal. Numbers are assigned in
123 /// the order visited.
124 bool Enumerate(const Value *V1, const Value *V2);
126 /// isEquivalentOperation - Compare two Instructions for equivalence, similar
127 /// to Instruction::isSameOperationAs but with modifications to the type
129 bool isEquivalentOperation(const Instruction *I1,
130 const Instruction *I2) const;
132 /// isEquivalentGEP - Compare two GEPs for equivalent pointer arithmetic.
133 bool isEquivalentGEP(const GEPOperator *GEP1, const GEPOperator *GEP2);
134 bool isEquivalentGEP(const GetElementPtrInst *GEP1,
135 const GetElementPtrInst *GEP2) {
136 return isEquivalentGEP(cast<GEPOperator>(GEP1), cast<GEPOperator>(GEP2));
139 /// isEquivalentType - Compare two Types, treating all pointer types as equal.
140 bool isEquivalentType(const Type *Ty1, const Type *Ty2) const;
142 // The two functions undergoing comparison.
143 const Function *F1, *F2;
145 const TargetData *TD;
147 typedef DenseMap<const Value *, unsigned long> IDMap;
149 unsigned long IDMap1Count, IDMap2Count;
153 /// isEquivalentType - any two pointers in the same address space are
154 /// equivalent. Otherwise, standard type equivalence rules apply.
155 bool FunctionComparator::isEquivalentType(const Type *Ty1,
156 const Type *Ty2) const {
159 if (Ty1->getTypeID() != Ty2->getTypeID())
162 switch(Ty1->getTypeID()) {
164 llvm_unreachable("Unknown type!");
165 // Fall through in Release mode.
166 case Type::IntegerTyID:
167 case Type::OpaqueTyID:
168 // Ty1 == Ty2 would have returned true earlier.
172 case Type::FloatTyID:
173 case Type::DoubleTyID:
174 case Type::X86_FP80TyID:
175 case Type::FP128TyID:
176 case Type::PPC_FP128TyID:
177 case Type::LabelTyID:
178 case Type::MetadataTyID:
181 case Type::PointerTyID: {
182 const PointerType *PTy1 = cast<PointerType>(Ty1);
183 const PointerType *PTy2 = cast<PointerType>(Ty2);
184 return PTy1->getAddressSpace() == PTy2->getAddressSpace();
187 case Type::StructTyID: {
188 const StructType *STy1 = cast<StructType>(Ty1);
189 const StructType *STy2 = cast<StructType>(Ty2);
190 if (STy1->getNumElements() != STy2->getNumElements())
193 if (STy1->isPacked() != STy2->isPacked())
196 for (unsigned i = 0, e = STy1->getNumElements(); i != e; ++i) {
197 if (!isEquivalentType(STy1->getElementType(i), STy2->getElementType(i)))
203 case Type::FunctionTyID: {
204 const FunctionType *FTy1 = cast<FunctionType>(Ty1);
205 const FunctionType *FTy2 = cast<FunctionType>(Ty2);
206 if (FTy1->getNumParams() != FTy2->getNumParams() ||
207 FTy1->isVarArg() != FTy2->isVarArg())
210 if (!isEquivalentType(FTy1->getReturnType(), FTy2->getReturnType()))
213 for (unsigned i = 0, e = FTy1->getNumParams(); i != e; ++i) {
214 if (!isEquivalentType(FTy1->getParamType(i), FTy2->getParamType(i)))
220 case Type::ArrayTyID: {
221 const ArrayType *ATy1 = cast<ArrayType>(Ty1);
222 const ArrayType *ATy2 = cast<ArrayType>(Ty2);
223 return ATy1->getNumElements() == ATy2->getNumElements() &&
224 isEquivalentType(ATy1->getElementType(), ATy2->getElementType());
227 case Type::VectorTyID: {
228 const VectorType *VTy1 = cast<VectorType>(Ty1);
229 const VectorType *VTy2 = cast<VectorType>(Ty2);
230 return VTy1->getNumElements() == VTy2->getNumElements() &&
231 isEquivalentType(VTy1->getElementType(), VTy2->getElementType());
236 /// isEquivalentOperation - determine whether the two operations are the same
237 /// except that pointer-to-A and pointer-to-B are equivalent. This should be
238 /// kept in sync with Instruction::isSameOperationAs.
239 bool FunctionComparator::isEquivalentOperation(const Instruction *I1,
240 const Instruction *I2) const {
241 if (I1->getOpcode() != I2->getOpcode() ||
242 I1->getNumOperands() != I2->getNumOperands() ||
243 !isEquivalentType(I1->getType(), I2->getType()) ||
244 !I1->hasSameSubclassOptionalData(I2))
247 // We have two instructions of identical opcode and #operands. Check to see
248 // if all operands are the same type
249 for (unsigned i = 0, e = I1->getNumOperands(); i != e; ++i)
250 if (!isEquivalentType(I1->getOperand(i)->getType(),
251 I2->getOperand(i)->getType()))
254 // Check special state that is a part of some instructions.
255 if (const LoadInst *LI = dyn_cast<LoadInst>(I1))
256 return LI->isVolatile() == cast<LoadInst>(I2)->isVolatile() &&
257 LI->getAlignment() == cast<LoadInst>(I2)->getAlignment();
258 if (const StoreInst *SI = dyn_cast<StoreInst>(I1))
259 return SI->isVolatile() == cast<StoreInst>(I2)->isVolatile() &&
260 SI->getAlignment() == cast<StoreInst>(I2)->getAlignment();
261 if (const CmpInst *CI = dyn_cast<CmpInst>(I1))
262 return CI->getPredicate() == cast<CmpInst>(I2)->getPredicate();
263 if (const CallInst *CI = dyn_cast<CallInst>(I1))
264 return CI->isTailCall() == cast<CallInst>(I2)->isTailCall() &&
265 CI->getCallingConv() == cast<CallInst>(I2)->getCallingConv() &&
266 CI->getAttributes().getRawPointer() ==
267 cast<CallInst>(I2)->getAttributes().getRawPointer();
268 if (const InvokeInst *CI = dyn_cast<InvokeInst>(I1))
269 return CI->getCallingConv() == cast<InvokeInst>(I2)->getCallingConv() &&
270 CI->getAttributes().getRawPointer() ==
271 cast<InvokeInst>(I2)->getAttributes().getRawPointer();
272 if (const InsertValueInst *IVI = dyn_cast<InsertValueInst>(I1)) {
273 if (IVI->getNumIndices() != cast<InsertValueInst>(I2)->getNumIndices())
275 for (unsigned i = 0, e = IVI->getNumIndices(); i != e; ++i)
276 if (IVI->idx_begin()[i] != cast<InsertValueInst>(I2)->idx_begin()[i])
280 if (const ExtractValueInst *EVI = dyn_cast<ExtractValueInst>(I1)) {
281 if (EVI->getNumIndices() != cast<ExtractValueInst>(I2)->getNumIndices())
283 for (unsigned i = 0, e = EVI->getNumIndices(); i != e; ++i)
284 if (EVI->idx_begin()[i] != cast<ExtractValueInst>(I2)->idx_begin()[i])
292 /// isEquivalentGEP - determine whether two GEP operations perform the same
293 /// underlying arithmetic.
294 bool FunctionComparator::isEquivalentGEP(const GEPOperator *GEP1,
295 const GEPOperator *GEP2) {
296 // When we have target data, we can reduce the GEP down to the value in bytes
297 // added to the address.
298 if (TD && GEP1->hasAllConstantIndices() && GEP2->hasAllConstantIndices()) {
299 SmallVector<Value *, 8> Indices1(GEP1->idx_begin(), GEP1->idx_end());
300 SmallVector<Value *, 8> Indices2(GEP2->idx_begin(), GEP2->idx_end());
301 uint64_t Offset1 = TD->getIndexedOffset(GEP1->getPointerOperandType(),
302 Indices1.data(), Indices1.size());
303 uint64_t Offset2 = TD->getIndexedOffset(GEP2->getPointerOperandType(),
304 Indices2.data(), Indices2.size());
305 return Offset1 == Offset2;
308 if (GEP1->getPointerOperand()->getType() !=
309 GEP2->getPointerOperand()->getType())
312 if (GEP1->getNumOperands() != GEP2->getNumOperands())
315 for (unsigned i = 0, e = GEP1->getNumOperands(); i != e; ++i) {
316 if (!Enumerate(GEP1->getOperand(i), GEP2->getOperand(i)))
323 /// Enumerate - Compare two values used by the two functions under pair-wise
324 /// comparison. If this is the first time the values are seen, they're added to
325 /// the mapping so that we will detect mismatches on next use.
326 bool FunctionComparator::Enumerate(const Value *V1, const Value *V2) {
327 // Check for function @f1 referring to itself and function @f2 referring to
328 // itself, or referring to each other, or both referring to either of them.
329 // They're all equivalent if the two functions are otherwise equivalent.
330 if (V1 == F1 && V2 == F2)
332 if (V1 == F2 && V2 == F1)
335 // TODO: constant expressions with GEP or references to F1 or F2.
336 if (isa<Constant>(V1))
339 if (isa<InlineAsm>(V1) && isa<InlineAsm>(V2)) {
340 const InlineAsm *IA1 = cast<InlineAsm>(V1);
341 const InlineAsm *IA2 = cast<InlineAsm>(V2);
342 return IA1->getAsmString() == IA2->getAsmString() &&
343 IA1->getConstraintString() == IA2->getConstraintString();
346 unsigned long &ID1 = Map1[V1];
350 unsigned long &ID2 = Map2[V2];
357 /// Compare - test whether two basic blocks have equivalent behaviour.
358 bool FunctionComparator::Compare(const BasicBlock *BB1, const BasicBlock *BB2) {
359 BasicBlock::const_iterator F1I = BB1->begin(), F1E = BB1->end();
360 BasicBlock::const_iterator F2I = BB2->begin(), F2E = BB2->end();
363 if (!Enumerate(F1I, F2I))
366 if (const GetElementPtrInst *GEP1 = dyn_cast<GetElementPtrInst>(F1I)) {
367 const GetElementPtrInst *GEP2 = dyn_cast<GetElementPtrInst>(F2I);
371 if (!Enumerate(GEP1->getPointerOperand(), GEP2->getPointerOperand()))
374 if (!isEquivalentGEP(GEP1, GEP2))
377 if (!isEquivalentOperation(F1I, F2I))
380 assert(F1I->getNumOperands() == F2I->getNumOperands());
381 for (unsigned i = 0, e = F1I->getNumOperands(); i != e; ++i) {
382 Value *OpF1 = F1I->getOperand(i);
383 Value *OpF2 = F2I->getOperand(i);
385 if (!Enumerate(OpF1, OpF2))
388 if (OpF1->getValueID() != OpF2->getValueID() ||
389 !isEquivalentType(OpF1->getType(), OpF2->getType()))
395 } while (F1I != F1E && F2I != F2E);
397 return F1I == F1E && F2I == F2E;
400 /// Compare - test whether the two functions have equivalent behaviour.
401 bool FunctionComparator::Compare() {
402 // We need to recheck everything, but check the things that weren't included
403 // in the hash first.
405 if (F1->getAttributes() != F2->getAttributes())
408 if (F1->hasGC() != F2->hasGC())
411 if (F1->hasGC() && F1->getGC() != F2->getGC())
414 if (F1->hasSection() != F2->hasSection())
417 if (F1->hasSection() && F1->getSection() != F2->getSection())
420 if (F1->isVarArg() != F2->isVarArg())
423 // TODO: if it's internal and only used in direct calls, we could handle this
425 if (F1->getCallingConv() != F2->getCallingConv())
428 if (!isEquivalentType(F1->getFunctionType(), F2->getFunctionType()))
431 assert(F1->arg_size() == F2->arg_size() &&
432 "Identical functions have a different number of args.");
434 // Visit the arguments so that they get enumerated in the order they're
436 for (Function::const_arg_iterator f1i = F1->arg_begin(),
437 f2i = F2->arg_begin(), f1e = F1->arg_end(); f1i != f1e; ++f1i, ++f2i) {
438 if (!Enumerate(f1i, f2i))
439 llvm_unreachable("Arguments repeat");
442 // We do a CFG-ordered walk since the actual ordering of the blocks in the
443 // linked list is immaterial. Our walk starts at the entry block for both
444 // functions, then takes each block from each terminator in order. As an
445 // artifact, this also means that unreachable blocks are ignored.
446 SmallVector<const BasicBlock *, 8> F1BBs, F2BBs;
447 SmallSet<const BasicBlock *, 128> VisitedBBs; // in terms of F1.
449 F1BBs.push_back(&F1->getEntryBlock());
450 F2BBs.push_back(&F2->getEntryBlock());
452 VisitedBBs.insert(F1BBs[0]);
453 while (!F1BBs.empty()) {
454 const BasicBlock *F1BB = F1BBs.pop_back_val();
455 const BasicBlock *F2BB = F2BBs.pop_back_val();
457 if (!Enumerate(F1BB, F2BB) || !Compare(F1BB, F2BB))
460 const TerminatorInst *F1TI = F1BB->getTerminator();
461 const TerminatorInst *F2TI = F2BB->getTerminator();
463 assert(F1TI->getNumSuccessors() == F2TI->getNumSuccessors());
464 for (unsigned i = 0, e = F1TI->getNumSuccessors(); i != e; ++i) {
465 if (!VisitedBBs.insert(F1TI->getSuccessor(i)))
468 F1BBs.push_back(F1TI->getSuccessor(i));
469 F2BBs.push_back(F2TI->getSuccessor(i));
475 /// WriteThunk - Replace G with a simple tail call to bitcast(F). Also replace
476 /// direct uses of G with bitcast(F).
477 void MergeFunctions::WriteThunk(Function *F, Function *G) const {
478 if (!G->mayBeOverridden()) {
479 // Redirect direct callers of G to F.
480 Constant *BitcastF = ConstantExpr::getBitCast(F, G->getType());
481 for (Value::use_iterator UI = G->use_begin(), UE = G->use_end();
483 Value::use_iterator TheIter = UI;
485 CallSite CS(*TheIter);
486 if (CS && CS.isCallee(TheIter))
487 TheIter.getUse().set(BitcastF);
491 // If G was internal then we may have replaced all uses if G with F. If so,
492 // stop here and delete G. There's no need for a thunk.
493 if (G->hasLocalLinkage() && G->use_empty()) {
494 G->eraseFromParent();
498 Function *NewG = Function::Create(G->getFunctionType(), G->getLinkage(), "",
500 BasicBlock *BB = BasicBlock::Create(F->getContext(), "", NewG);
501 IRBuilder<false> Builder(BB);
503 SmallVector<Value *, 16> Args;
505 const FunctionType *FFTy = F->getFunctionType();
506 for (Function::arg_iterator AI = NewG->arg_begin(), AE = NewG->arg_end();
508 Args.push_back(Builder.CreateBitCast(AI, FFTy->getParamType(i)));
512 CallInst *CI = Builder.CreateCall(F, Args.begin(), Args.end());
514 CI->setCallingConv(F->getCallingConv());
515 if (NewG->getReturnType()->isVoidTy()) {
516 Builder.CreateRetVoid();
518 Builder.CreateRet(Builder.CreateBitCast(CI, NewG->getReturnType()));
521 NewG->copyAttributesFrom(G);
523 G->replaceAllUsesWith(NewG);
524 G->eraseFromParent();
527 /// MergeTwoFunctions - Merge two equivalent functions. Upon completion,
528 /// Function G is deleted.
529 void MergeFunctions::MergeTwoFunctions(Function *F, Function *G) const {
530 if (F->isWeakForLinker()) {
531 assert(G->isWeakForLinker());
533 // Make them both thunks to the same internal function.
534 Function *H = Function::Create(F->getFunctionType(), F->getLinkage(), "",
536 H->copyAttributesFrom(F);
538 F->replaceAllUsesWith(H);
540 unsigned MaxAlignment = std::max(G->getAlignment(), H->getAlignment());
545 F->setAlignment(MaxAlignment);
546 F->setLinkage(GlobalValue::InternalLinkage);
551 ++NumFunctionsMerged;
554 static unsigned ProfileFunction(const Function *F) {
555 const FunctionType *FTy = F->getFunctionType();
558 ID.AddInteger(F->size());
559 ID.AddInteger(F->getCallingConv());
560 ID.AddBoolean(F->hasGC());
561 ID.AddBoolean(FTy->isVarArg());
562 ID.AddInteger(FTy->getReturnType()->getTypeID());
563 for (unsigned i = 0, e = FTy->getNumParams(); i != e; ++i)
564 ID.AddInteger(FTy->getParamType(i)->getTypeID());
565 return ID.ComputeHash();
568 class ComparableFunction {
570 ComparableFunction(Function *Func, TargetData *TD)
571 : Func(Func), Hash(ProfileFunction(Func)), TD(TD) {}
573 AssertingVH<Function> const Func;
575 TargetData * const TD;
578 struct MergeFunctionsEqualityInfo {
579 static ComparableFunction *getEmptyKey() {
580 return reinterpret_cast<ComparableFunction*>(0);
582 static ComparableFunction *getTombstoneKey() {
583 return reinterpret_cast<ComparableFunction*>(-1);
585 static unsigned getHashValue(const ComparableFunction *CF) {
588 static bool isEqual(const ComparableFunction *LHS,
589 const ComparableFunction *RHS) {
592 if (LHS == getEmptyKey() || LHS == getTombstoneKey() ||
593 RHS == getEmptyKey() || RHS == getTombstoneKey())
595 assert(LHS->TD == RHS->TD && "Comparing functions for different targets");
596 return FunctionComparator(LHS->TD, LHS->Func, RHS->Func).Compare();
600 bool MergeFunctions::runOnModule(Module &M) {
601 bool Changed = false;
603 TD = getAnalysisIfAvailable<TargetData>();
605 typedef DenseSet<ComparableFunction *, MergeFunctionsEqualityInfo> FnSetType;
607 for (Module::iterator F = M.begin(), E = M.end(); F != E;) {
608 if (F->isDeclaration() || F->hasAvailableExternallyLinkage()) {
613 ComparableFunction *NewF = new ComparableFunction(F, TD);
615 std::pair<FnSetType::iterator, bool> Result = FnSet.insert(NewF);
616 if (!Result.second) {
617 ComparableFunction *&OldF = *Result.first;
618 assert(OldF && "Expected a hash collision");
620 // NewF will be deleted in favour of OldF unless NewF is strong and OldF
621 // is weak in which case swap them to keep the strong definition.
623 if (OldF->Func->isWeakForLinker() && !NewF->Func->isWeakForLinker())
624 std::swap(OldF, NewF);
626 DEBUG(dbgs() << " " << OldF->Func->getName() << " == "
627 << NewF->Func->getName() << '\n');
630 Function *DeleteF = NewF->Func;
632 MergeTwoFunctions(OldF->Func, DeleteF);
636 DeleteContainerPointers(FnSet);