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 #include "llvm/Transforms/IPO.h"
47 #include "llvm/ADT/DenseSet.h"
48 #include "llvm/ADT/FoldingSet.h"
49 #include "llvm/ADT/STLExtras.h"
50 #include "llvm/ADT/SmallSet.h"
51 #include "llvm/ADT/Statistic.h"
52 #include "llvm/IR/CallSite.h"
53 #include "llvm/IR/Constants.h"
54 #include "llvm/IR/DataLayout.h"
55 #include "llvm/IR/IRBuilder.h"
56 #include "llvm/IR/InlineAsm.h"
57 #include "llvm/IR/Instructions.h"
58 #include "llvm/IR/LLVMContext.h"
59 #include "llvm/IR/Module.h"
60 #include "llvm/IR/Operator.h"
61 #include "llvm/IR/ValueHandle.h"
62 #include "llvm/Pass.h"
63 #include "llvm/Support/Debug.h"
64 #include "llvm/Support/ErrorHandling.h"
65 #include "llvm/Support/raw_ostream.h"
69 #define DEBUG_TYPE "mergefunc"
71 STATISTIC(NumFunctionsMerged, "Number of functions merged");
72 STATISTIC(NumThunksWritten, "Number of thunks generated");
73 STATISTIC(NumAliasesWritten, "Number of aliases generated");
74 STATISTIC(NumDoubleWeak, "Number of new functions created");
76 /// Returns the type id for a type to be hashed. We turn pointer types into
77 /// integers here because the actual compare logic below considers pointers and
78 /// integers of the same size as equal.
79 static Type::TypeID getTypeIDForHash(Type *Ty) {
80 if (Ty->isPointerTy())
81 return Type::IntegerTyID;
82 return Ty->getTypeID();
85 /// Creates a hash-code for the function which is the same for any two
86 /// functions that will compare equal, without looking at the instructions
87 /// inside the function.
88 static unsigned profileFunction(const Function *F) {
89 FunctionType *FTy = F->getFunctionType();
92 ID.AddInteger(F->size());
93 ID.AddInteger(F->getCallingConv());
94 ID.AddBoolean(F->hasGC());
95 ID.AddBoolean(FTy->isVarArg());
96 ID.AddInteger(getTypeIDForHash(FTy->getReturnType()));
97 for (unsigned i = 0, e = FTy->getNumParams(); i != e; ++i)
98 ID.AddInteger(getTypeIDForHash(FTy->getParamType(i)));
99 return ID.ComputeHash();
104 /// ComparableFunction - A struct that pairs together functions with a
105 /// DataLayout so that we can keep them together as elements in the DenseSet.
106 class ComparableFunction {
108 static const ComparableFunction EmptyKey;
109 static const ComparableFunction TombstoneKey;
110 static DataLayout * const LookupOnly;
112 ComparableFunction(Function *Func, const DataLayout *DL)
113 : Func(Func), Hash(profileFunction(Func)), DL(DL) {}
115 Function *getFunc() const { return Func; }
116 unsigned getHash() const { return Hash; }
117 const DataLayout *getDataLayout() const { return DL; }
119 // Drops AssertingVH reference to the function. Outside of debug mode, this
123 "Attempted to release function twice, or release empty/tombstone!");
128 explicit ComparableFunction(unsigned Hash)
129 : Func(nullptr), Hash(Hash), DL(nullptr) {}
131 AssertingVH<Function> Func;
133 const DataLayout *DL;
136 const ComparableFunction ComparableFunction::EmptyKey = ComparableFunction(0);
137 const ComparableFunction ComparableFunction::TombstoneKey =
138 ComparableFunction(1);
139 DataLayout *const ComparableFunction::LookupOnly = (DataLayout*)(-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. DataLayout 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 DataLayout *DL, const Function *F1,
170 : F1(F1), F2(F2), DL(DL) {}
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 /// cmpType - compares two types,
198 /// defines total ordering among the types set.
201 /// 0 if types are equal,
202 /// -1 if Left is less than Right,
203 /// +1 if Left is greater than Right.
206 /// Comparison is broken onto stages. Like in lexicographical comparison
207 /// stage coming first has higher priority.
208 /// On each explanation stage keep in mind total ordering properties.
210 /// 0. Before comparison we coerce pointer types of 0 address space to
212 /// We also don't bother with same type at left and right, so
213 /// just return 0 in this case.
215 /// 1. If types are of different kind (different type IDs).
216 /// Return result of type IDs comparison, treating them as numbers.
217 /// 2. If types are vectors or integers, compare Type* values as numbers.
218 /// 3. Types has same ID, so check whether they belongs to the next group:
227 /// If so - return 0, yes - we can treat these types as equal only because
228 /// their IDs are same.
229 /// 4. If Left and Right are pointers, return result of address space
230 /// comparison (numbers comparison). We can treat pointer types of same
231 /// address space as equal.
232 /// 5. If types are complex.
233 /// Then both Left and Right are to be expanded and their element types will
234 /// be checked with the same way. If we get Res != 0 on some stage, return it.
235 /// Otherwise return 0.
236 /// 6. For all other cases put llvm_unreachable.
237 int cmpType(Type *TyL, Type *TyR) const;
239 bool isEquivalentType(Type *Ty1, Type *Ty2) const {
240 return cmpType(Ty1, Ty2) == 0;
243 int cmpNumbers(uint64_t L, uint64_t R) const;
245 // The two functions undergoing comparison.
246 const Function *F1, *F2;
248 const DataLayout *DL;
250 DenseMap<const Value *, const Value *> id_map;
251 DenseSet<const Value *> seen_values;
256 int FunctionComparator::cmpNumbers(uint64_t L, uint64_t R) const {
257 if (L < R) return -1;
262 /// cmpType - compares two types,
263 /// defines total ordering among the types set.
264 /// See method declaration comments for more details.
265 int FunctionComparator::cmpType(Type *TyL, Type *TyR) const {
267 PointerType *PTyL = dyn_cast<PointerType>(TyL);
268 PointerType *PTyR = dyn_cast<PointerType>(TyR);
271 if (PTyL && PTyL->getAddressSpace() == 0) TyL = DL->getIntPtrType(TyL);
272 if (PTyR && PTyR->getAddressSpace() == 0) TyR = DL->getIntPtrType(TyR);
278 if (int Res = cmpNumbers(TyL->getTypeID(), TyR->getTypeID()))
281 switch (TyL->getTypeID()) {
283 llvm_unreachable("Unknown type!");
284 // Fall through in Release mode.
285 case Type::IntegerTyID:
286 case Type::VectorTyID:
287 // TyL == TyR would have returned true earlier.
288 return cmpNumbers((uint64_t)TyL, (uint64_t)TyR);
291 case Type::FloatTyID:
292 case Type::DoubleTyID:
293 case Type::X86_FP80TyID:
294 case Type::FP128TyID:
295 case Type::PPC_FP128TyID:
296 case Type::LabelTyID:
297 case Type::MetadataTyID:
300 case Type::PointerTyID: {
301 assert(PTyL && PTyR && "Both types must be pointers here.");
302 return cmpNumbers(PTyL->getAddressSpace(), PTyR->getAddressSpace());
305 case Type::StructTyID: {
306 StructType *STyL = cast<StructType>(TyL);
307 StructType *STyR = cast<StructType>(TyR);
308 if (STyL->getNumElements() != STyR->getNumElements())
309 return cmpNumbers(STyL->getNumElements(), STyR->getNumElements());
311 if (STyL->isPacked() != STyR->isPacked())
312 return cmpNumbers(STyL->isPacked(), STyR->isPacked());
314 for (unsigned i = 0, e = STyL->getNumElements(); i != e; ++i) {
315 if (int Res = cmpType(STyL->getElementType(i),
316 STyR->getElementType(i)))
322 case Type::FunctionTyID: {
323 FunctionType *FTyL = cast<FunctionType>(TyL);
324 FunctionType *FTyR = cast<FunctionType>(TyR);
325 if (FTyL->getNumParams() != FTyR->getNumParams())
326 return cmpNumbers(FTyL->getNumParams(), FTyR->getNumParams());
328 if (FTyL->isVarArg() != FTyR->isVarArg())
329 return cmpNumbers(FTyL->isVarArg(), FTyR->isVarArg());
331 if (int Res = cmpType(FTyL->getReturnType(), FTyR->getReturnType()))
334 for (unsigned i = 0, e = FTyL->getNumParams(); i != e; ++i) {
335 if (int Res = cmpType(FTyL->getParamType(i), FTyR->getParamType(i)))
341 case Type::ArrayTyID: {
342 ArrayType *ATyL = cast<ArrayType>(TyL);
343 ArrayType *ATyR = cast<ArrayType>(TyR);
344 if (ATyL->getNumElements() != ATyR->getNumElements())
345 return cmpNumbers(ATyL->getNumElements(), ATyR->getNumElements());
346 return cmpType(ATyL->getElementType(), ATyR->getElementType());
351 // Determine whether the two operations are the same except that pointer-to-A
352 // and pointer-to-B are equivalent. This should be kept in sync with
353 // Instruction::isSameOperationAs.
354 bool FunctionComparator::isEquivalentOperation(const Instruction *I1,
355 const Instruction *I2) const {
356 // Differences from Instruction::isSameOperationAs:
357 // * replace type comparison with calls to isEquivalentType.
358 // * we test for I->hasSameSubclassOptionalData (nuw/nsw/tail) at the top
359 // * because of the above, we don't test for the tail bit on calls later on
360 if (I1->getOpcode() != I2->getOpcode() ||
361 I1->getNumOperands() != I2->getNumOperands() ||
362 !isEquivalentType(I1->getType(), I2->getType()) ||
363 !I1->hasSameSubclassOptionalData(I2))
366 // We have two instructions of identical opcode and #operands. Check to see
367 // if all operands are the same type
368 for (unsigned i = 0, e = I1->getNumOperands(); i != e; ++i)
369 if (!isEquivalentType(I1->getOperand(i)->getType(),
370 I2->getOperand(i)->getType()))
373 // Check special state that is a part of some instructions.
374 if (const LoadInst *LI = dyn_cast<LoadInst>(I1))
375 return LI->isVolatile() == cast<LoadInst>(I2)->isVolatile() &&
376 LI->getAlignment() == cast<LoadInst>(I2)->getAlignment() &&
377 LI->getOrdering() == cast<LoadInst>(I2)->getOrdering() &&
378 LI->getSynchScope() == cast<LoadInst>(I2)->getSynchScope();
379 if (const StoreInst *SI = dyn_cast<StoreInst>(I1))
380 return SI->isVolatile() == cast<StoreInst>(I2)->isVolatile() &&
381 SI->getAlignment() == cast<StoreInst>(I2)->getAlignment() &&
382 SI->getOrdering() == cast<StoreInst>(I2)->getOrdering() &&
383 SI->getSynchScope() == cast<StoreInst>(I2)->getSynchScope();
384 if (const CmpInst *CI = dyn_cast<CmpInst>(I1))
385 return CI->getPredicate() == cast<CmpInst>(I2)->getPredicate();
386 if (const CallInst *CI = dyn_cast<CallInst>(I1))
387 return CI->getCallingConv() == cast<CallInst>(I2)->getCallingConv() &&
388 CI->getAttributes() == cast<CallInst>(I2)->getAttributes();
389 if (const InvokeInst *CI = dyn_cast<InvokeInst>(I1))
390 return CI->getCallingConv() == cast<InvokeInst>(I2)->getCallingConv() &&
391 CI->getAttributes() == cast<InvokeInst>(I2)->getAttributes();
392 if (const InsertValueInst *IVI = dyn_cast<InsertValueInst>(I1))
393 return IVI->getIndices() == cast<InsertValueInst>(I2)->getIndices();
394 if (const ExtractValueInst *EVI = dyn_cast<ExtractValueInst>(I1))
395 return EVI->getIndices() == cast<ExtractValueInst>(I2)->getIndices();
396 if (const FenceInst *FI = dyn_cast<FenceInst>(I1))
397 return FI->getOrdering() == cast<FenceInst>(I2)->getOrdering() &&
398 FI->getSynchScope() == cast<FenceInst>(I2)->getSynchScope();
399 if (const AtomicCmpXchgInst *CXI = dyn_cast<AtomicCmpXchgInst>(I1))
400 return CXI->isVolatile() == cast<AtomicCmpXchgInst>(I2)->isVolatile() &&
401 CXI->getSuccessOrdering() ==
402 cast<AtomicCmpXchgInst>(I2)->getSuccessOrdering() &&
403 CXI->getFailureOrdering() ==
404 cast<AtomicCmpXchgInst>(I2)->getFailureOrdering() &&
405 CXI->getSynchScope() == cast<AtomicCmpXchgInst>(I2)->getSynchScope();
406 if (const AtomicRMWInst *RMWI = dyn_cast<AtomicRMWInst>(I1))
407 return RMWI->getOperation() == cast<AtomicRMWInst>(I2)->getOperation() &&
408 RMWI->isVolatile() == cast<AtomicRMWInst>(I2)->isVolatile() &&
409 RMWI->getOrdering() == cast<AtomicRMWInst>(I2)->getOrdering() &&
410 RMWI->getSynchScope() == cast<AtomicRMWInst>(I2)->getSynchScope();
415 // Determine whether two GEP operations perform the same underlying arithmetic.
416 bool FunctionComparator::isEquivalentGEP(const GEPOperator *GEP1,
417 const GEPOperator *GEP2) {
418 unsigned AS = GEP1->getPointerAddressSpace();
419 if (AS != GEP2->getPointerAddressSpace())
423 // When we have target data, we can reduce the GEP down to the value in bytes
424 // added to the address.
425 unsigned BitWidth = DL ? DL->getPointerSizeInBits(AS) : 1;
426 APInt Offset1(BitWidth, 0), Offset2(BitWidth, 0);
427 if (GEP1->accumulateConstantOffset(*DL, Offset1) &&
428 GEP2->accumulateConstantOffset(*DL, Offset2)) {
429 return Offset1 == Offset2;
433 if (GEP1->getPointerOperand()->getType() !=
434 GEP2->getPointerOperand()->getType())
437 if (GEP1->getNumOperands() != GEP2->getNumOperands())
440 for (unsigned i = 0, e = GEP1->getNumOperands(); i != e; ++i) {
441 if (!enumerate(GEP1->getOperand(i), GEP2->getOperand(i)))
448 // Compare two values used by the two functions under pair-wise comparison. If
449 // this is the first time the values are seen, they're added to the mapping so
450 // that we will detect mismatches on next use.
451 bool FunctionComparator::enumerate(const Value *V1, const Value *V2) {
452 // Check for function @f1 referring to itself and function @f2 referring to
453 // itself, or referring to each other, or both referring to either of them.
454 // They're all equivalent if the two functions are otherwise equivalent.
455 if (V1 == F1 && V2 == F2)
457 if (V1 == F2 && V2 == F1)
460 if (const Constant *C1 = dyn_cast<Constant>(V1)) {
461 if (V1 == V2) return true;
462 const Constant *C2 = dyn_cast<Constant>(V2);
463 if (!C2) return false;
464 // TODO: constant expressions with GEP or references to F1 or F2.
465 if (C1->isNullValue() && C2->isNullValue() &&
466 isEquivalentType(C1->getType(), C2->getType()))
468 // Try bitcasting C2 to C1's type. If the bitcast is legal and returns C1
469 // then they must have equal bit patterns.
470 return C1->getType()->canLosslesslyBitCastTo(C2->getType()) &&
471 C1 == ConstantExpr::getBitCast(const_cast<Constant*>(C2), C1->getType());
474 if (isa<InlineAsm>(V1) || isa<InlineAsm>(V2))
477 // Check that V1 maps to V2. If we find a value that V1 maps to then we simply
478 // check whether it's equal to V2. When there is no mapping then we need to
479 // ensure that V2 isn't already equivalent to something else. For this
480 // purpose, we track the V2 values in a set.
482 const Value *&map_elem = id_map[V1];
484 return map_elem == V2;
485 if (!seen_values.insert(V2).second)
491 // Test whether two basic blocks have equivalent behaviour.
492 bool FunctionComparator::compare(const BasicBlock *BB1, const BasicBlock *BB2) {
493 BasicBlock::const_iterator F1I = BB1->begin(), F1E = BB1->end();
494 BasicBlock::const_iterator F2I = BB2->begin(), F2E = BB2->end();
497 if (!enumerate(F1I, F2I))
500 if (const GetElementPtrInst *GEP1 = dyn_cast<GetElementPtrInst>(F1I)) {
501 const GetElementPtrInst *GEP2 = dyn_cast<GetElementPtrInst>(F2I);
505 if (!enumerate(GEP1->getPointerOperand(), GEP2->getPointerOperand()))
508 if (!isEquivalentGEP(GEP1, GEP2))
511 if (!isEquivalentOperation(F1I, F2I))
514 assert(F1I->getNumOperands() == F2I->getNumOperands());
515 for (unsigned i = 0, e = F1I->getNumOperands(); i != e; ++i) {
516 Value *OpF1 = F1I->getOperand(i);
517 Value *OpF2 = F2I->getOperand(i);
519 if (!enumerate(OpF1, OpF2))
522 if (OpF1->getValueID() != OpF2->getValueID() ||
523 !isEquivalentType(OpF1->getType(), OpF2->getType()))
529 } while (F1I != F1E && F2I != F2E);
531 return F1I == F1E && F2I == F2E;
534 // Test whether the two functions have equivalent behaviour.
535 bool FunctionComparator::compare() {
536 // We need to recheck everything, but check the things that weren't included
537 // in the hash first.
539 if (F1->getAttributes() != F2->getAttributes())
542 if (F1->hasGC() != F2->hasGC())
545 if (F1->hasGC() && F1->getGC() != F2->getGC())
548 if (F1->hasSection() != F2->hasSection())
551 if (F1->hasSection() && F1->getSection() != F2->getSection())
554 if (F1->isVarArg() != F2->isVarArg())
557 // TODO: if it's internal and only used in direct calls, we could handle this
559 if (F1->getCallingConv() != F2->getCallingConv())
562 if (!isEquivalentType(F1->getFunctionType(), F2->getFunctionType()))
565 assert(F1->arg_size() == F2->arg_size() &&
566 "Identically typed functions have different numbers of args!");
568 // Visit the arguments so that they get enumerated in the order they're
570 for (Function::const_arg_iterator f1i = F1->arg_begin(),
571 f2i = F2->arg_begin(), f1e = F1->arg_end(); f1i != f1e; ++f1i, ++f2i) {
572 if (!enumerate(f1i, f2i))
573 llvm_unreachable("Arguments repeat!");
576 // We do a CFG-ordered walk since the actual ordering of the blocks in the
577 // linked list is immaterial. Our walk starts at the entry block for both
578 // functions, then takes each block from each terminator in order. As an
579 // artifact, this also means that unreachable blocks are ignored.
580 SmallVector<const BasicBlock *, 8> F1BBs, F2BBs;
581 SmallSet<const BasicBlock *, 128> VisitedBBs; // in terms of F1.
583 F1BBs.push_back(&F1->getEntryBlock());
584 F2BBs.push_back(&F2->getEntryBlock());
586 VisitedBBs.insert(F1BBs[0]);
587 while (!F1BBs.empty()) {
588 const BasicBlock *F1BB = F1BBs.pop_back_val();
589 const BasicBlock *F2BB = F2BBs.pop_back_val();
591 if (!enumerate(F1BB, F2BB) || !compare(F1BB, F2BB))
594 const TerminatorInst *F1TI = F1BB->getTerminator();
595 const TerminatorInst *F2TI = F2BB->getTerminator();
597 assert(F1TI->getNumSuccessors() == F2TI->getNumSuccessors());
598 for (unsigned i = 0, e = F1TI->getNumSuccessors(); i != e; ++i) {
599 if (!VisitedBBs.insert(F1TI->getSuccessor(i)))
602 F1BBs.push_back(F1TI->getSuccessor(i));
603 F2BBs.push_back(F2TI->getSuccessor(i));
611 /// MergeFunctions finds functions which will generate identical machine code,
612 /// by considering all pointer types to be equivalent. Once identified,
613 /// MergeFunctions will fold them by replacing a call to one to a call to a
614 /// bitcast of the other.
616 class MergeFunctions : public ModulePass {
620 : ModulePass(ID), HasGlobalAliases(false) {
621 initializeMergeFunctionsPass(*PassRegistry::getPassRegistry());
624 bool runOnModule(Module &M) override;
627 typedef DenseSet<ComparableFunction> FnSetType;
629 /// A work queue of functions that may have been modified and should be
631 std::vector<WeakVH> Deferred;
633 /// Insert a ComparableFunction into the FnSet, or merge it away if it's
634 /// equal to one that's already present.
635 bool insert(ComparableFunction &NewF);
637 /// Remove a Function from the FnSet and queue it up for a second sweep of
639 void remove(Function *F);
641 /// Find the functions that use this Value and remove them from FnSet and
642 /// queue the functions.
643 void removeUsers(Value *V);
645 /// Replace all direct calls of Old with calls of New. Will bitcast New if
646 /// necessary to make types match.
647 void replaceDirectCallers(Function *Old, Function *New);
649 /// Merge two equivalent functions. Upon completion, G may be deleted, or may
650 /// be converted into a thunk. In either case, it should never be visited
652 void mergeTwoFunctions(Function *F, Function *G);
654 /// Replace G with a thunk or an alias to F. Deletes G.
655 void writeThunkOrAlias(Function *F, Function *G);
657 /// Replace G with a simple tail call to bitcast(F). Also replace direct uses
658 /// of G with bitcast(F). Deletes G.
659 void writeThunk(Function *F, Function *G);
661 /// Replace G with an alias to F. Deletes G.
662 void writeAlias(Function *F, Function *G);
664 /// The set of all distinct functions. Use the insert() and remove() methods
668 /// DataLayout for more accurate GEP comparisons. May be NULL.
669 const DataLayout *DL;
671 /// Whether or not the target supports global aliases.
672 bool HasGlobalAliases;
675 } // end anonymous namespace
677 char MergeFunctions::ID = 0;
678 INITIALIZE_PASS(MergeFunctions, "mergefunc", "Merge Functions", false, false)
680 ModulePass *llvm::createMergeFunctionsPass() {
681 return new MergeFunctions();
684 bool MergeFunctions::runOnModule(Module &M) {
685 bool Changed = false;
686 DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>();
687 DL = DLP ? &DLP->getDataLayout() : nullptr;
689 for (Module::iterator I = M.begin(), E = M.end(); I != E; ++I) {
690 if (!I->isDeclaration() && !I->hasAvailableExternallyLinkage())
691 Deferred.push_back(WeakVH(I));
693 FnSet.resize(Deferred.size());
696 std::vector<WeakVH> Worklist;
697 Deferred.swap(Worklist);
699 DEBUG(dbgs() << "size of module: " << M.size() << '\n');
700 DEBUG(dbgs() << "size of worklist: " << Worklist.size() << '\n');
702 // Insert only strong functions and merge them. Strong function merging
703 // always deletes one of them.
704 for (std::vector<WeakVH>::iterator I = Worklist.begin(),
705 E = Worklist.end(); I != E; ++I) {
707 Function *F = cast<Function>(*I);
708 if (!F->isDeclaration() && !F->hasAvailableExternallyLinkage() &&
709 !F->mayBeOverridden()) {
710 ComparableFunction CF = ComparableFunction(F, DL);
711 Changed |= insert(CF);
715 // Insert only weak functions and merge them. By doing these second we
716 // create thunks to the strong function when possible. When two weak
717 // functions are identical, we create a new strong function with two weak
718 // weak thunks to it which are identical but not mergable.
719 for (std::vector<WeakVH>::iterator I = Worklist.begin(),
720 E = Worklist.end(); I != E; ++I) {
722 Function *F = cast<Function>(*I);
723 if (!F->isDeclaration() && !F->hasAvailableExternallyLinkage() &&
724 F->mayBeOverridden()) {
725 ComparableFunction CF = ComparableFunction(F, DL);
726 Changed |= insert(CF);
729 DEBUG(dbgs() << "size of FnSet: " << FnSet.size() << '\n');
730 } while (!Deferred.empty());
737 bool DenseMapInfo<ComparableFunction>::isEqual(const ComparableFunction &LHS,
738 const ComparableFunction &RHS) {
739 if (LHS.getFunc() == RHS.getFunc() &&
740 LHS.getHash() == RHS.getHash())
742 if (!LHS.getFunc() || !RHS.getFunc())
745 // One of these is a special "underlying pointer comparison only" object.
746 if (LHS.getDataLayout() == ComparableFunction::LookupOnly ||
747 RHS.getDataLayout() == ComparableFunction::LookupOnly)
750 assert(LHS.getDataLayout() == RHS.getDataLayout() &&
751 "Comparing functions for different targets");
753 return FunctionComparator(LHS.getDataLayout(), LHS.getFunc(),
754 RHS.getFunc()).compare();
757 // Replace direct callers of Old with New.
758 void MergeFunctions::replaceDirectCallers(Function *Old, Function *New) {
759 Constant *BitcastNew = ConstantExpr::getBitCast(New, Old->getType());
760 for (auto UI = Old->use_begin(), UE = Old->use_end(); UI != UE;) {
763 CallSite CS(U->getUser());
764 if (CS && CS.isCallee(U)) {
765 remove(CS.getInstruction()->getParent()->getParent());
771 // Replace G with an alias to F if possible, or else a thunk to F. Deletes G.
772 void MergeFunctions::writeThunkOrAlias(Function *F, Function *G) {
773 if (HasGlobalAliases && G->hasUnnamedAddr()) {
774 if (G->hasExternalLinkage() || G->hasLocalLinkage() ||
775 G->hasWeakLinkage()) {
784 // Helper for writeThunk,
785 // Selects proper bitcast operation,
786 // but a bit simpler then CastInst::getCastOpcode.
787 static Value* createCast(IRBuilder<false> &Builder, Value *V, Type *DestTy) {
788 Type *SrcTy = V->getType();
789 if (SrcTy->isIntegerTy() && DestTy->isPointerTy())
790 return Builder.CreateIntToPtr(V, DestTy);
791 else if (SrcTy->isPointerTy() && DestTy->isIntegerTy())
792 return Builder.CreatePtrToInt(V, DestTy);
794 return Builder.CreateBitCast(V, DestTy);
797 // Replace G with a simple tail call to bitcast(F). Also replace direct uses
798 // of G with bitcast(F). Deletes G.
799 void MergeFunctions::writeThunk(Function *F, Function *G) {
800 if (!G->mayBeOverridden()) {
801 // Redirect direct callers of G to F.
802 replaceDirectCallers(G, F);
805 // If G was internal then we may have replaced all uses of G with F. If so,
806 // stop here and delete G. There's no need for a thunk.
807 if (G->hasLocalLinkage() && G->use_empty()) {
808 G->eraseFromParent();
812 Function *NewG = Function::Create(G->getFunctionType(), G->getLinkage(), "",
814 BasicBlock *BB = BasicBlock::Create(F->getContext(), "", NewG);
815 IRBuilder<false> Builder(BB);
817 SmallVector<Value *, 16> Args;
819 FunctionType *FFTy = F->getFunctionType();
820 for (Function::arg_iterator AI = NewG->arg_begin(), AE = NewG->arg_end();
822 Args.push_back(createCast(Builder, (Value*)AI, FFTy->getParamType(i)));
826 CallInst *CI = Builder.CreateCall(F, Args);
828 CI->setCallingConv(F->getCallingConv());
829 if (NewG->getReturnType()->isVoidTy()) {
830 Builder.CreateRetVoid();
832 Builder.CreateRet(createCast(Builder, CI, NewG->getReturnType()));
835 NewG->copyAttributesFrom(G);
838 G->replaceAllUsesWith(NewG);
839 G->eraseFromParent();
841 DEBUG(dbgs() << "writeThunk: " << NewG->getName() << '\n');
845 // Replace G with an alias to F and delete G.
846 void MergeFunctions::writeAlias(Function *F, Function *G) {
847 Constant *BitcastF = ConstantExpr::getBitCast(F, G->getType());
848 GlobalAlias *GA = new GlobalAlias(G->getType(), G->getLinkage(), "",
849 BitcastF, G->getParent());
850 F->setAlignment(std::max(F->getAlignment(), G->getAlignment()));
852 GA->setVisibility(G->getVisibility());
854 G->replaceAllUsesWith(GA);
855 G->eraseFromParent();
857 DEBUG(dbgs() << "writeAlias: " << GA->getName() << '\n');
861 // Merge two equivalent functions. Upon completion, Function G is deleted.
862 void MergeFunctions::mergeTwoFunctions(Function *F, Function *G) {
863 if (F->mayBeOverridden()) {
864 assert(G->mayBeOverridden());
866 if (HasGlobalAliases) {
867 // Make them both thunks to the same internal function.
868 Function *H = Function::Create(F->getFunctionType(), F->getLinkage(), "",
870 H->copyAttributesFrom(F);
873 F->replaceAllUsesWith(H);
875 unsigned MaxAlignment = std::max(G->getAlignment(), H->getAlignment());
880 F->setAlignment(MaxAlignment);
881 F->setLinkage(GlobalValue::PrivateLinkage);
883 // We can't merge them. Instead, pick one and update all direct callers
884 // to call it and hope that we improve the instruction cache hit rate.
885 replaceDirectCallers(G, F);
890 writeThunkOrAlias(F, G);
893 ++NumFunctionsMerged;
896 // Insert a ComparableFunction into the FnSet, or merge it away if equal to one
897 // that was already inserted.
898 bool MergeFunctions::insert(ComparableFunction &NewF) {
899 std::pair<FnSetType::iterator, bool> Result = FnSet.insert(NewF);
901 DEBUG(dbgs() << "Inserting as unique: " << NewF.getFunc()->getName() << '\n');
905 const ComparableFunction &OldF = *Result.first;
907 // Don't merge tiny functions, since it can just end up making the function
909 // FIXME: Should still merge them if they are unnamed_addr and produce an
911 if (NewF.getFunc()->size() == 1) {
912 if (NewF.getFunc()->front().size() <= 2) {
913 DEBUG(dbgs() << NewF.getFunc()->getName()
914 << " is to small to bother merging\n");
919 // Never thunk a strong function to a weak function.
920 assert(!OldF.getFunc()->mayBeOverridden() ||
921 NewF.getFunc()->mayBeOverridden());
923 DEBUG(dbgs() << " " << OldF.getFunc()->getName() << " == "
924 << NewF.getFunc()->getName() << '\n');
926 Function *DeleteF = NewF.getFunc();
928 mergeTwoFunctions(OldF.getFunc(), DeleteF);
932 // Remove a function from FnSet. If it was already in FnSet, add it to Deferred
933 // so that we'll look at it in the next round.
934 void MergeFunctions::remove(Function *F) {
935 // We need to make sure we remove F, not a function "equal" to F per the
936 // function equality comparator.
938 // The special "lookup only" ComparableFunction bypasses the expensive
939 // function comparison in favour of a pointer comparison on the underlying
941 ComparableFunction CF = ComparableFunction(F, ComparableFunction::LookupOnly);
942 if (FnSet.erase(CF)) {
943 DEBUG(dbgs() << "Removed " << F->getName() << " from set and deferred it.\n");
944 Deferred.push_back(F);
948 // For each instruction used by the value, remove() the function that contains
949 // the instruction. This should happen right before a call to RAUW.
950 void MergeFunctions::removeUsers(Value *V) {
951 std::vector<Value *> Worklist;
952 Worklist.push_back(V);
953 while (!Worklist.empty()) {
954 Value *V = Worklist.back();
957 for (User *U : V->users()) {
958 if (Instruction *I = dyn_cast<Instruction>(U)) {
959 remove(I->getParent()->getParent());
960 } else if (isa<GlobalValue>(U)) {
962 } else if (Constant *C = dyn_cast<Constant>(U)) {
963 for (User *UU : C->users())
964 Worklist.push_back(UU);