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/STLExtras.h"
51 #include "llvm/ADT/SmallSet.h"
52 #include "llvm/ADT/Statistic.h"
53 #include "llvm/IR/CallSite.h"
54 #include "llvm/IR/Constants.h"
55 #include "llvm/IR/DataLayout.h"
56 #include "llvm/IR/IRBuilder.h"
57 #include "llvm/IR/InlineAsm.h"
58 #include "llvm/IR/Instructions.h"
59 #include "llvm/IR/LLVMContext.h"
60 #include "llvm/IR/Module.h"
61 #include "llvm/IR/Operator.h"
62 #include "llvm/IR/ValueHandle.h"
63 #include "llvm/Pass.h"
64 #include "llvm/Support/Debug.h"
65 #include "llvm/Support/ErrorHandling.h"
66 #include "llvm/Support/raw_ostream.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 /// Returns the type id for a type to be hashed. We turn pointer types into
76 /// integers here because the actual compare logic below considers pointers and
77 /// integers of the same size as equal.
78 static Type::TypeID getTypeIDForHash(Type *Ty) {
79 if (Ty->isPointerTy())
80 return Type::IntegerTyID;
81 return Ty->getTypeID();
84 /// Creates a hash-code for the function which is the same for any two
85 /// functions that will compare equal, without looking at the instructions
86 /// inside the function.
87 static unsigned profileFunction(const Function *F) {
88 FunctionType *FTy = F->getFunctionType();
91 ID.AddInteger(F->size());
92 ID.AddInteger(F->getCallingConv());
93 ID.AddBoolean(F->hasGC());
94 ID.AddBoolean(FTy->isVarArg());
95 ID.AddInteger(getTypeIDForHash(FTy->getReturnType()));
96 for (unsigned i = 0, e = FTy->getNumParams(); i != e; ++i)
97 ID.AddInteger(getTypeIDForHash(FTy->getParamType(i)));
98 return ID.ComputeHash();
103 /// ComparableFunction - A struct that pairs together functions with a
104 /// DataLayout so that we can keep them together as elements in the DenseSet.
105 class ComparableFunction {
107 static const ComparableFunction EmptyKey;
108 static const ComparableFunction TombstoneKey;
109 static DataLayout * const LookupOnly;
111 ComparableFunction(Function *Func, const DataLayout *DL)
112 : Func(Func), Hash(profileFunction(Func)), DL(DL) {}
114 Function *getFunc() const { return Func; }
115 unsigned getHash() const { return Hash; }
116 const DataLayout *getDataLayout() const { return DL; }
118 // Drops AssertingVH reference to the function. Outside of debug mode, this
122 "Attempted to release function twice, or release empty/tombstone!");
127 explicit ComparableFunction(unsigned Hash)
128 : Func(NULL), Hash(Hash), DL(NULL) {}
130 AssertingVH<Function> Func;
132 const DataLayout *DL;
135 const ComparableFunction ComparableFunction::EmptyKey = ComparableFunction(0);
136 const ComparableFunction ComparableFunction::TombstoneKey =
137 ComparableFunction(1);
138 DataLayout *const ComparableFunction::LookupOnly = (DataLayout*)(-1);
144 struct DenseMapInfo<ComparableFunction> {
145 static ComparableFunction getEmptyKey() {
146 return ComparableFunction::EmptyKey;
148 static ComparableFunction getTombstoneKey() {
149 return ComparableFunction::TombstoneKey;
151 static unsigned getHashValue(const ComparableFunction &CF) {
154 static bool isEqual(const ComparableFunction &LHS,
155 const ComparableFunction &RHS);
161 /// FunctionComparator - Compares two functions to determine whether or not
162 /// they will generate machine code with the same behaviour. DataLayout is
163 /// used if available. The comparator always fails conservatively (erring on the
164 /// side of claiming that two functions are different).
165 class FunctionComparator {
167 FunctionComparator(const DataLayout *DL, const Function *F1,
169 : F1(F1), F2(F2), DL(DL) {}
171 /// Test whether the two functions have equivalent behaviour.
175 /// Test whether two basic blocks have equivalent behaviour.
176 bool compare(const BasicBlock *BB1, const BasicBlock *BB2);
178 /// Assign or look up previously assigned numbers for the two values, and
179 /// return whether the numbers are equal. Numbers are assigned in the order
181 bool enumerate(const Value *V1, const Value *V2);
183 /// Compare two Instructions for equivalence, similar to
184 /// Instruction::isSameOperationAs but with modifications to the type
186 bool isEquivalentOperation(const Instruction *I1,
187 const Instruction *I2) const;
189 /// Compare two GEPs for equivalent pointer arithmetic.
190 bool isEquivalentGEP(const GEPOperator *GEP1, const GEPOperator *GEP2);
191 bool isEquivalentGEP(const GetElementPtrInst *GEP1,
192 const GetElementPtrInst *GEP2) {
193 return isEquivalentGEP(cast<GEPOperator>(GEP1), cast<GEPOperator>(GEP2));
196 /// Compare two Types, treating all pointer types as equal.
197 bool isEquivalentType(Type *Ty1, Type *Ty2) const;
199 // The two functions undergoing comparison.
200 const Function *F1, *F2;
202 const DataLayout *DL;
204 DenseMap<const Value *, const Value *> id_map;
205 DenseSet<const Value *> seen_values;
210 // Any two pointers in the same address space are equivalent, intptr_t and
211 // pointers are equivalent. Otherwise, standard type equivalence rules apply.
212 bool FunctionComparator::isEquivalentType(Type *Ty1, Type *Ty2) const {
214 PointerType *PTy1 = dyn_cast<PointerType>(Ty1);
215 PointerType *PTy2 = dyn_cast<PointerType>(Ty2);
218 if (PTy1 && PTy1->getAddressSpace() == 0) Ty1 = DL->getIntPtrType(Ty1);
219 if (PTy2 && PTy2->getAddressSpace() == 0) Ty2 = DL->getIntPtrType(Ty2);
225 if (Ty1->getTypeID() != Ty2->getTypeID())
228 switch (Ty1->getTypeID()) {
230 llvm_unreachable("Unknown type!");
231 // Fall through in Release mode.
232 case Type::IntegerTyID:
233 case Type::VectorTyID:
234 // Ty1 == Ty2 would have returned true earlier.
238 case Type::FloatTyID:
239 case Type::DoubleTyID:
240 case Type::X86_FP80TyID:
241 case Type::FP128TyID:
242 case Type::PPC_FP128TyID:
243 case Type::LabelTyID:
244 case Type::MetadataTyID:
247 case Type::PointerTyID: {
248 assert(PTy1 && PTy2 && "Both types must be pointers here.");
249 return PTy1->getAddressSpace() == PTy2->getAddressSpace();
252 case Type::StructTyID: {
253 StructType *STy1 = cast<StructType>(Ty1);
254 StructType *STy2 = cast<StructType>(Ty2);
255 if (STy1->getNumElements() != STy2->getNumElements())
258 if (STy1->isPacked() != STy2->isPacked())
261 for (unsigned i = 0, e = STy1->getNumElements(); i != e; ++i) {
262 if (!isEquivalentType(STy1->getElementType(i), STy2->getElementType(i)))
268 case Type::FunctionTyID: {
269 FunctionType *FTy1 = cast<FunctionType>(Ty1);
270 FunctionType *FTy2 = cast<FunctionType>(Ty2);
271 if (FTy1->getNumParams() != FTy2->getNumParams() ||
272 FTy1->isVarArg() != FTy2->isVarArg())
275 if (!isEquivalentType(FTy1->getReturnType(), FTy2->getReturnType()))
278 for (unsigned i = 0, e = FTy1->getNumParams(); i != e; ++i) {
279 if (!isEquivalentType(FTy1->getParamType(i), FTy2->getParamType(i)))
285 case Type::ArrayTyID: {
286 ArrayType *ATy1 = cast<ArrayType>(Ty1);
287 ArrayType *ATy2 = cast<ArrayType>(Ty2);
288 return ATy1->getNumElements() == ATy2->getNumElements() &&
289 isEquivalentType(ATy1->getElementType(), ATy2->getElementType());
294 // Determine whether the two operations are the same except that pointer-to-A
295 // and pointer-to-B are equivalent. This should be kept in sync with
296 // Instruction::isSameOperationAs.
297 bool FunctionComparator::isEquivalentOperation(const Instruction *I1,
298 const Instruction *I2) const {
299 // Differences from Instruction::isSameOperationAs:
300 // * replace type comparison with calls to isEquivalentType.
301 // * we test for I->hasSameSubclassOptionalData (nuw/nsw/tail) at the top
302 // * because of the above, we don't test for the tail bit on calls later on
303 if (I1->getOpcode() != I2->getOpcode() ||
304 I1->getNumOperands() != I2->getNumOperands() ||
305 !isEquivalentType(I1->getType(), I2->getType()) ||
306 !I1->hasSameSubclassOptionalData(I2))
309 // We have two instructions of identical opcode and #operands. Check to see
310 // if all operands are the same type
311 for (unsigned i = 0, e = I1->getNumOperands(); i != e; ++i)
312 if (!isEquivalentType(I1->getOperand(i)->getType(),
313 I2->getOperand(i)->getType()))
316 // Check special state that is a part of some instructions.
317 if (const LoadInst *LI = dyn_cast<LoadInst>(I1))
318 return LI->isVolatile() == cast<LoadInst>(I2)->isVolatile() &&
319 LI->getAlignment() == cast<LoadInst>(I2)->getAlignment() &&
320 LI->getOrdering() == cast<LoadInst>(I2)->getOrdering() &&
321 LI->getSynchScope() == cast<LoadInst>(I2)->getSynchScope();
322 if (const StoreInst *SI = dyn_cast<StoreInst>(I1))
323 return SI->isVolatile() == cast<StoreInst>(I2)->isVolatile() &&
324 SI->getAlignment() == cast<StoreInst>(I2)->getAlignment() &&
325 SI->getOrdering() == cast<StoreInst>(I2)->getOrdering() &&
326 SI->getSynchScope() == cast<StoreInst>(I2)->getSynchScope();
327 if (const CmpInst *CI = dyn_cast<CmpInst>(I1))
328 return CI->getPredicate() == cast<CmpInst>(I2)->getPredicate();
329 if (const CallInst *CI = dyn_cast<CallInst>(I1))
330 return CI->getCallingConv() == cast<CallInst>(I2)->getCallingConv() &&
331 CI->getAttributes() == cast<CallInst>(I2)->getAttributes();
332 if (const InvokeInst *CI = dyn_cast<InvokeInst>(I1))
333 return CI->getCallingConv() == cast<InvokeInst>(I2)->getCallingConv() &&
334 CI->getAttributes() == cast<InvokeInst>(I2)->getAttributes();
335 if (const InsertValueInst *IVI = dyn_cast<InsertValueInst>(I1))
336 return IVI->getIndices() == cast<InsertValueInst>(I2)->getIndices();
337 if (const ExtractValueInst *EVI = dyn_cast<ExtractValueInst>(I1))
338 return EVI->getIndices() == cast<ExtractValueInst>(I2)->getIndices();
339 if (const FenceInst *FI = dyn_cast<FenceInst>(I1))
340 return FI->getOrdering() == cast<FenceInst>(I2)->getOrdering() &&
341 FI->getSynchScope() == cast<FenceInst>(I2)->getSynchScope();
342 if (const AtomicCmpXchgInst *CXI = dyn_cast<AtomicCmpXchgInst>(I1))
343 return CXI->isVolatile() == cast<AtomicCmpXchgInst>(I2)->isVolatile() &&
344 CXI->getSuccessOrdering() ==
345 cast<AtomicCmpXchgInst>(I2)->getSuccessOrdering() &&
346 CXI->getFailureOrdering() ==
347 cast<AtomicCmpXchgInst>(I2)->getFailureOrdering() &&
348 CXI->getSynchScope() == cast<AtomicCmpXchgInst>(I2)->getSynchScope();
349 if (const AtomicRMWInst *RMWI = dyn_cast<AtomicRMWInst>(I1))
350 return RMWI->getOperation() == cast<AtomicRMWInst>(I2)->getOperation() &&
351 RMWI->isVolatile() == cast<AtomicRMWInst>(I2)->isVolatile() &&
352 RMWI->getOrdering() == cast<AtomicRMWInst>(I2)->getOrdering() &&
353 RMWI->getSynchScope() == cast<AtomicRMWInst>(I2)->getSynchScope();
358 // Determine whether two GEP operations perform the same underlying arithmetic.
359 bool FunctionComparator::isEquivalentGEP(const GEPOperator *GEP1,
360 const GEPOperator *GEP2) {
361 unsigned AS = GEP1->getPointerAddressSpace();
362 if (AS != GEP2->getPointerAddressSpace())
366 // When we have target data, we can reduce the GEP down to the value in bytes
367 // added to the address.
368 unsigned BitWidth = DL ? DL->getPointerSizeInBits(AS) : 1;
369 APInt Offset1(BitWidth, 0), Offset2(BitWidth, 0);
370 if (GEP1->accumulateConstantOffset(*DL, Offset1) &&
371 GEP2->accumulateConstantOffset(*DL, Offset2)) {
372 return Offset1 == Offset2;
376 if (GEP1->getPointerOperand()->getType() !=
377 GEP2->getPointerOperand()->getType())
380 if (GEP1->getNumOperands() != GEP2->getNumOperands())
383 for (unsigned i = 0, e = GEP1->getNumOperands(); i != e; ++i) {
384 if (!enumerate(GEP1->getOperand(i), GEP2->getOperand(i)))
391 // Compare two values used by the two functions under pair-wise comparison. If
392 // this is the first time the values are seen, they're added to the mapping so
393 // that we will detect mismatches on next use.
394 bool FunctionComparator::enumerate(const Value *V1, const Value *V2) {
395 // Check for function @f1 referring to itself and function @f2 referring to
396 // itself, or referring to each other, or both referring to either of them.
397 // They're all equivalent if the two functions are otherwise equivalent.
398 if (V1 == F1 && V2 == F2)
400 if (V1 == F2 && V2 == F1)
403 if (const Constant *C1 = dyn_cast<Constant>(V1)) {
404 if (V1 == V2) return true;
405 const Constant *C2 = dyn_cast<Constant>(V2);
406 if (!C2) return false;
407 // TODO: constant expressions with GEP or references to F1 or F2.
408 if (C1->isNullValue() && C2->isNullValue() &&
409 isEquivalentType(C1->getType(), C2->getType()))
411 // Try bitcasting C2 to C1's type. If the bitcast is legal and returns C1
412 // then they must have equal bit patterns.
413 return C1->getType()->canLosslesslyBitCastTo(C2->getType()) &&
414 C1 == ConstantExpr::getBitCast(const_cast<Constant*>(C2), C1->getType());
417 if (isa<InlineAsm>(V1) || isa<InlineAsm>(V2))
420 // Check that V1 maps to V2. If we find a value that V1 maps to then we simply
421 // check whether it's equal to V2. When there is no mapping then we need to
422 // ensure that V2 isn't already equivalent to something else. For this
423 // purpose, we track the V2 values in a set.
425 const Value *&map_elem = id_map[V1];
427 return map_elem == V2;
428 if (!seen_values.insert(V2).second)
434 // Test whether two basic blocks have equivalent behaviour.
435 bool FunctionComparator::compare(const BasicBlock *BB1, const BasicBlock *BB2) {
436 BasicBlock::const_iterator F1I = BB1->begin(), F1E = BB1->end();
437 BasicBlock::const_iterator F2I = BB2->begin(), F2E = BB2->end();
440 if (!enumerate(F1I, F2I))
443 if (const GetElementPtrInst *GEP1 = dyn_cast<GetElementPtrInst>(F1I)) {
444 const GetElementPtrInst *GEP2 = dyn_cast<GetElementPtrInst>(F2I);
448 if (!enumerate(GEP1->getPointerOperand(), GEP2->getPointerOperand()))
451 if (!isEquivalentGEP(GEP1, GEP2))
454 if (!isEquivalentOperation(F1I, F2I))
457 assert(F1I->getNumOperands() == F2I->getNumOperands());
458 for (unsigned i = 0, e = F1I->getNumOperands(); i != e; ++i) {
459 Value *OpF1 = F1I->getOperand(i);
460 Value *OpF2 = F2I->getOperand(i);
462 if (!enumerate(OpF1, OpF2))
465 if (OpF1->getValueID() != OpF2->getValueID() ||
466 !isEquivalentType(OpF1->getType(), OpF2->getType()))
472 } while (F1I != F1E && F2I != F2E);
474 return F1I == F1E && F2I == F2E;
477 // Test whether the two functions have equivalent behaviour.
478 bool FunctionComparator::compare() {
479 // We need to recheck everything, but check the things that weren't included
480 // in the hash first.
482 if (F1->getAttributes() != F2->getAttributes())
485 if (F1->hasGC() != F2->hasGC())
488 if (F1->hasGC() && F1->getGC() != F2->getGC())
491 if (F1->hasSection() != F2->hasSection())
494 if (F1->hasSection() && F1->getSection() != F2->getSection())
497 if (F1->isVarArg() != F2->isVarArg())
500 // TODO: if it's internal and only used in direct calls, we could handle this
502 if (F1->getCallingConv() != F2->getCallingConv())
505 if (!isEquivalentType(F1->getFunctionType(), F2->getFunctionType()))
508 assert(F1->arg_size() == F2->arg_size() &&
509 "Identically typed functions have different numbers of args!");
511 // Visit the arguments so that they get enumerated in the order they're
513 for (Function::const_arg_iterator f1i = F1->arg_begin(),
514 f2i = F2->arg_begin(), f1e = F1->arg_end(); f1i != f1e; ++f1i, ++f2i) {
515 if (!enumerate(f1i, f2i))
516 llvm_unreachable("Arguments repeat!");
519 // We do a CFG-ordered walk since the actual ordering of the blocks in the
520 // linked list is immaterial. Our walk starts at the entry block for both
521 // functions, then takes each block from each terminator in order. As an
522 // artifact, this also means that unreachable blocks are ignored.
523 SmallVector<const BasicBlock *, 8> F1BBs, F2BBs;
524 SmallSet<const BasicBlock *, 128> VisitedBBs; // in terms of F1.
526 F1BBs.push_back(&F1->getEntryBlock());
527 F2BBs.push_back(&F2->getEntryBlock());
529 VisitedBBs.insert(F1BBs[0]);
530 while (!F1BBs.empty()) {
531 const BasicBlock *F1BB = F1BBs.pop_back_val();
532 const BasicBlock *F2BB = F2BBs.pop_back_val();
534 if (!enumerate(F1BB, F2BB) || !compare(F1BB, F2BB))
537 const TerminatorInst *F1TI = F1BB->getTerminator();
538 const TerminatorInst *F2TI = F2BB->getTerminator();
540 assert(F1TI->getNumSuccessors() == F2TI->getNumSuccessors());
541 for (unsigned i = 0, e = F1TI->getNumSuccessors(); i != e; ++i) {
542 if (!VisitedBBs.insert(F1TI->getSuccessor(i)))
545 F1BBs.push_back(F1TI->getSuccessor(i));
546 F2BBs.push_back(F2TI->getSuccessor(i));
554 /// MergeFunctions finds functions which will generate identical machine code,
555 /// by considering all pointer types to be equivalent. Once identified,
556 /// MergeFunctions will fold them by replacing a call to one to a call to a
557 /// bitcast of the other.
559 class MergeFunctions : public ModulePass {
563 : ModulePass(ID), HasGlobalAliases(false) {
564 initializeMergeFunctionsPass(*PassRegistry::getPassRegistry());
567 bool runOnModule(Module &M) override;
570 typedef DenseSet<ComparableFunction> FnSetType;
572 /// A work queue of functions that may have been modified and should be
574 std::vector<WeakVH> Deferred;
576 /// Insert a ComparableFunction into the FnSet, or merge it away if it's
577 /// equal to one that's already present.
578 bool insert(ComparableFunction &NewF);
580 /// Remove a Function from the FnSet and queue it up for a second sweep of
582 void remove(Function *F);
584 /// Find the functions that use this Value and remove them from FnSet and
585 /// queue the functions.
586 void removeUsers(Value *V);
588 /// Replace all direct calls of Old with calls of New. Will bitcast New if
589 /// necessary to make types match.
590 void replaceDirectCallers(Function *Old, Function *New);
592 /// Merge two equivalent functions. Upon completion, G may be deleted, or may
593 /// be converted into a thunk. In either case, it should never be visited
595 void mergeTwoFunctions(Function *F, Function *G);
597 /// Replace G with a thunk or an alias to F. Deletes G.
598 void writeThunkOrAlias(Function *F, Function *G);
600 /// Replace G with a simple tail call to bitcast(F). Also replace direct uses
601 /// of G with bitcast(F). Deletes G.
602 void writeThunk(Function *F, Function *G);
604 /// Replace G with an alias to F. Deletes G.
605 void writeAlias(Function *F, Function *G);
607 /// The set of all distinct functions. Use the insert() and remove() methods
611 /// DataLayout for more accurate GEP comparisons. May be NULL.
612 const DataLayout *DL;
614 /// Whether or not the target supports global aliases.
615 bool HasGlobalAliases;
618 } // end anonymous namespace
620 char MergeFunctions::ID = 0;
621 INITIALIZE_PASS(MergeFunctions, "mergefunc", "Merge Functions", false, false)
623 ModulePass *llvm::createMergeFunctionsPass() {
624 return new MergeFunctions();
627 bool MergeFunctions::runOnModule(Module &M) {
628 bool Changed = false;
629 DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>();
630 DL = DLP ? &DLP->getDataLayout() : 0;
632 for (Module::iterator I = M.begin(), E = M.end(); I != E; ++I) {
633 if (!I->isDeclaration() && !I->hasAvailableExternallyLinkage())
634 Deferred.push_back(WeakVH(I));
636 FnSet.resize(Deferred.size());
639 std::vector<WeakVH> Worklist;
640 Deferred.swap(Worklist);
642 DEBUG(dbgs() << "size of module: " << M.size() << '\n');
643 DEBUG(dbgs() << "size of worklist: " << Worklist.size() << '\n');
645 // Insert only strong functions and merge them. Strong function merging
646 // always deletes one of them.
647 for (std::vector<WeakVH>::iterator I = Worklist.begin(),
648 E = Worklist.end(); I != E; ++I) {
650 Function *F = cast<Function>(*I);
651 if (!F->isDeclaration() && !F->hasAvailableExternallyLinkage() &&
652 !F->mayBeOverridden()) {
653 ComparableFunction CF = ComparableFunction(F, DL);
654 Changed |= insert(CF);
658 // Insert only weak functions and merge them. By doing these second we
659 // create thunks to the strong function when possible. When two weak
660 // functions are identical, we create a new strong function with two weak
661 // weak thunks to it which are identical but not mergable.
662 for (std::vector<WeakVH>::iterator I = Worklist.begin(),
663 E = Worklist.end(); I != E; ++I) {
665 Function *F = cast<Function>(*I);
666 if (!F->isDeclaration() && !F->hasAvailableExternallyLinkage() &&
667 F->mayBeOverridden()) {
668 ComparableFunction CF = ComparableFunction(F, DL);
669 Changed |= insert(CF);
672 DEBUG(dbgs() << "size of FnSet: " << FnSet.size() << '\n');
673 } while (!Deferred.empty());
680 bool DenseMapInfo<ComparableFunction>::isEqual(const ComparableFunction &LHS,
681 const ComparableFunction &RHS) {
682 if (LHS.getFunc() == RHS.getFunc() &&
683 LHS.getHash() == RHS.getHash())
685 if (!LHS.getFunc() || !RHS.getFunc())
688 // One of these is a special "underlying pointer comparison only" object.
689 if (LHS.getDataLayout() == ComparableFunction::LookupOnly ||
690 RHS.getDataLayout() == ComparableFunction::LookupOnly)
693 assert(LHS.getDataLayout() == RHS.getDataLayout() &&
694 "Comparing functions for different targets");
696 return FunctionComparator(LHS.getDataLayout(), LHS.getFunc(),
697 RHS.getFunc()).compare();
700 // Replace direct callers of Old with New.
701 void MergeFunctions::replaceDirectCallers(Function *Old, Function *New) {
702 Constant *BitcastNew = ConstantExpr::getBitCast(New, Old->getType());
703 for (auto UI = Old->use_begin(), UE = Old->use_end(); UI != UE;) {
706 CallSite CS(U->getUser());
707 if (CS && CS.isCallee(U)) {
708 remove(CS.getInstruction()->getParent()->getParent());
714 // Replace G with an alias to F if possible, or else a thunk to F. Deletes G.
715 void MergeFunctions::writeThunkOrAlias(Function *F, Function *G) {
716 if (HasGlobalAliases && G->hasUnnamedAddr()) {
717 if (G->hasExternalLinkage() || G->hasLocalLinkage() ||
718 G->hasWeakLinkage()) {
727 // Helper for writeThunk,
728 // Selects proper bitcast operation,
729 // but a bit simpler then CastInst::getCastOpcode.
730 static Value* createCast(IRBuilder<false> &Builder, Value *V, Type *DestTy) {
731 Type *SrcTy = V->getType();
732 if (SrcTy->isIntegerTy() && DestTy->isPointerTy())
733 return Builder.CreateIntToPtr(V, DestTy);
734 else if (SrcTy->isPointerTy() && DestTy->isIntegerTy())
735 return Builder.CreatePtrToInt(V, DestTy);
737 return Builder.CreateBitCast(V, DestTy);
740 // Replace G with a simple tail call to bitcast(F). Also replace direct uses
741 // of G with bitcast(F). Deletes G.
742 void MergeFunctions::writeThunk(Function *F, Function *G) {
743 if (!G->mayBeOverridden()) {
744 // Redirect direct callers of G to F.
745 replaceDirectCallers(G, F);
748 // If G was internal then we may have replaced all uses of G with F. If so,
749 // stop here and delete G. There's no need for a thunk.
750 if (G->hasLocalLinkage() && G->use_empty()) {
751 G->eraseFromParent();
755 Function *NewG = Function::Create(G->getFunctionType(), G->getLinkage(), "",
757 BasicBlock *BB = BasicBlock::Create(F->getContext(), "", NewG);
758 IRBuilder<false> Builder(BB);
760 SmallVector<Value *, 16> Args;
762 FunctionType *FFTy = F->getFunctionType();
763 for (Function::arg_iterator AI = NewG->arg_begin(), AE = NewG->arg_end();
765 Args.push_back(createCast(Builder, (Value*)AI, FFTy->getParamType(i)));
769 CallInst *CI = Builder.CreateCall(F, Args);
771 CI->setCallingConv(F->getCallingConv());
772 if (NewG->getReturnType()->isVoidTy()) {
773 Builder.CreateRetVoid();
775 Builder.CreateRet(createCast(Builder, CI, NewG->getReturnType()));
778 NewG->copyAttributesFrom(G);
781 G->replaceAllUsesWith(NewG);
782 G->eraseFromParent();
784 DEBUG(dbgs() << "writeThunk: " << NewG->getName() << '\n');
788 // Replace G with an alias to F and delete G.
789 void MergeFunctions::writeAlias(Function *F, Function *G) {
790 Constant *BitcastF = ConstantExpr::getBitCast(F, G->getType());
791 GlobalAlias *GA = new GlobalAlias(G->getType(), G->getLinkage(), "",
792 BitcastF, G->getParent());
793 F->setAlignment(std::max(F->getAlignment(), G->getAlignment()));
795 GA->setVisibility(G->getVisibility());
797 G->replaceAllUsesWith(GA);
798 G->eraseFromParent();
800 DEBUG(dbgs() << "writeAlias: " << GA->getName() << '\n');
804 // Merge two equivalent functions. Upon completion, Function G is deleted.
805 void MergeFunctions::mergeTwoFunctions(Function *F, Function *G) {
806 if (F->mayBeOverridden()) {
807 assert(G->mayBeOverridden());
809 if (HasGlobalAliases) {
810 // Make them both thunks to the same internal function.
811 Function *H = Function::Create(F->getFunctionType(), F->getLinkage(), "",
813 H->copyAttributesFrom(F);
816 F->replaceAllUsesWith(H);
818 unsigned MaxAlignment = std::max(G->getAlignment(), H->getAlignment());
823 F->setAlignment(MaxAlignment);
824 F->setLinkage(GlobalValue::PrivateLinkage);
826 // We can't merge them. Instead, pick one and update all direct callers
827 // to call it and hope that we improve the instruction cache hit rate.
828 replaceDirectCallers(G, F);
833 writeThunkOrAlias(F, G);
836 ++NumFunctionsMerged;
839 // Insert a ComparableFunction into the FnSet, or merge it away if equal to one
840 // that was already inserted.
841 bool MergeFunctions::insert(ComparableFunction &NewF) {
842 std::pair<FnSetType::iterator, bool> Result = FnSet.insert(NewF);
844 DEBUG(dbgs() << "Inserting as unique: " << NewF.getFunc()->getName() << '\n');
848 const ComparableFunction &OldF = *Result.first;
850 // Don't merge tiny functions, since it can just end up making the function
852 // FIXME: Should still merge them if they are unnamed_addr and produce an
854 if (NewF.getFunc()->size() == 1) {
855 if (NewF.getFunc()->front().size() <= 2) {
856 DEBUG(dbgs() << NewF.getFunc()->getName()
857 << " is to small to bother merging\n");
862 // Never thunk a strong function to a weak function.
863 assert(!OldF.getFunc()->mayBeOverridden() ||
864 NewF.getFunc()->mayBeOverridden());
866 DEBUG(dbgs() << " " << OldF.getFunc()->getName() << " == "
867 << NewF.getFunc()->getName() << '\n');
869 Function *DeleteF = NewF.getFunc();
871 mergeTwoFunctions(OldF.getFunc(), DeleteF);
875 // Remove a function from FnSet. If it was already in FnSet, add it to Deferred
876 // so that we'll look at it in the next round.
877 void MergeFunctions::remove(Function *F) {
878 // We need to make sure we remove F, not a function "equal" to F per the
879 // function equality comparator.
881 // The special "lookup only" ComparableFunction bypasses the expensive
882 // function comparison in favour of a pointer comparison on the underlying
884 ComparableFunction CF = ComparableFunction(F, ComparableFunction::LookupOnly);
885 if (FnSet.erase(CF)) {
886 DEBUG(dbgs() << "Removed " << F->getName() << " from set and deferred it.\n");
887 Deferred.push_back(F);
891 // For each instruction used by the value, remove() the function that contains
892 // the instruction. This should happen right before a call to RAUW.
893 void MergeFunctions::removeUsers(Value *V) {
894 std::vector<Value *> Worklist;
895 Worklist.push_back(V);
896 while (!Worklist.empty()) {
897 Value *V = Worklist.back();
900 for (User *U : V->users()) {
901 if (Instruction *I = dyn_cast<Instruction>(U)) {
902 remove(I->getParent()->getParent());
903 } else if (isa<GlobalValue>(U)) {
905 } else if (Constant *C = dyn_cast<Constant>(U)) {
906 for (User *UU : C->users())
907 Worklist.push_back(UU);