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 /// cmpType - compares two types,
197 /// defines total ordering among the types set.
200 /// 0 if types are equal,
201 /// -1 if Left is less than Right,
202 /// +1 if Left is greater than Right.
205 /// Comparison is broken onto stages. Like in lexicographical comparison
206 /// stage coming first has higher priority.
207 /// On each explanation stage keep in mind total ordering properties.
209 /// 0. Before comparison we coerce pointer types of 0 address space to integer.
210 /// We also don't bother with same type at left and right, so
211 /// just return 0 in this case.
213 /// 1. If types are of different kind (different type IDs).
214 /// Return result of type IDs comparison, treating them as numbers.
215 /// 2. If types are vectors or integers, compare Type* values as numbers.
216 /// 3. Types has same ID, so check whether they belongs to the next group:
225 /// If so - return 0, yes - we can treat these types as equal only because
226 /// their IDs are same.
227 /// 4. If Left and Right are pointers, return result of address space
228 /// comparison (numbers comparison). We can treat pointer types of same
229 /// address space as equal.
230 /// 5. If types are complex.
231 /// Then both Left and Right are to be expanded and their element types will
232 /// be checked with the same way. If we get Res != 0 on some stage, return it.
233 /// Otherwise return 0.
234 /// 6. For all other cases put llvm_unreachable.
235 int cmpType(Type *TyL, Type *TyR) const;
237 bool isEquivalentType(Type *Ty1, Type *Ty2) const {
238 return cmpType(Ty1, Ty2) == 0;
241 int cmpNumbers(uint64_t L, uint64_t R) const;
243 // The two functions undergoing comparison.
244 const Function *F1, *F2;
246 const DataLayout *DL;
248 DenseMap<const Value *, const Value *> id_map;
249 DenseSet<const Value *> seen_values;
254 int FunctionComparator::cmpNumbers(uint64_t L, uint64_t R) const {
255 if (L < R) return -1;
260 /// cmpType - compares two types,
261 /// defines total ordering among the types set.
262 /// See method declaration comments for more details.
263 int FunctionComparator::cmpType(Type *TyL, Type *TyR) const {
265 PointerType *PTy1 = dyn_cast<PointerType>(TyL);
266 PointerType *PTy2 = dyn_cast<PointerType>(TyR);
269 if (PTy1 && PTy1->getAddressSpace() == 0) TyL = DL->getIntPtrType(TyL);
270 if (PTy2 && PTy2->getAddressSpace() == 0) TyR = DL->getIntPtrType(TyR);
276 if (int Res = cmpNumbers(TyL->getTypeID(), TyR->getTypeID()))
279 switch (TyL->getTypeID()) {
281 llvm_unreachable("Unknown type!");
282 // Fall through in Release mode.
283 case Type::IntegerTyID:
284 case Type::VectorTyID:
285 // TyL == TyR would have returned true earlier.
286 return cmpNumbers((uint64_t)TyL, (uint64_t)TyR);
289 case Type::FloatTyID:
290 case Type::DoubleTyID:
291 case Type::X86_FP80TyID:
292 case Type::FP128TyID:
293 case Type::PPC_FP128TyID:
294 case Type::LabelTyID:
295 case Type::MetadataTyID:
298 case Type::PointerTyID: {
299 assert(PTy1 && PTy2 && "Both types must be pointers here.");
300 return cmpNumbers(PTy1->getAddressSpace(), PTy2->getAddressSpace());
303 case Type::StructTyID: {
304 StructType *STy1 = cast<StructType>(TyL);
305 StructType *STy2 = cast<StructType>(TyR);
306 if (STy1->getNumElements() != STy2->getNumElements())
307 return cmpNumbers(STy1->getNumElements(), STy2->getNumElements());
309 if (STy1->isPacked() != STy2->isPacked())
310 return cmpNumbers(STy1->isPacked(), STy2->isPacked());
312 for (unsigned i = 0, e = STy1->getNumElements(); i != e; ++i) {
313 if (int Res = cmpType(STy1->getElementType(i),
314 STy2->getElementType(i)))
320 case Type::FunctionTyID: {
321 FunctionType *FTy1 = cast<FunctionType>(TyL);
322 FunctionType *FTy2 = cast<FunctionType>(TyR);
323 if (FTy1->getNumParams() != FTy2->getNumParams())
324 return cmpNumbers(FTy1->getNumParams(), FTy2->getNumParams());
326 if (FTy1->isVarArg() != FTy2->isVarArg())
327 return cmpNumbers(FTy1->isVarArg(), FTy2->isVarArg());
329 if (int Res = cmpType(FTy1->getReturnType(), FTy2->getReturnType()))
332 for (unsigned i = 0, e = FTy1->getNumParams(); i != e; ++i) {
333 if (int Res = cmpType(FTy1->getParamType(i), FTy2->getParamType(i)))
339 case Type::ArrayTyID: {
340 ArrayType *ATy1 = cast<ArrayType>(TyL);
341 ArrayType *ATy2 = cast<ArrayType>(TyR);
342 if (ATy1->getNumElements() != ATy2->getNumElements())
343 return cmpNumbers(ATy1->getNumElements(), ATy2->getNumElements());
344 return cmpType(ATy1->getElementType(), ATy2->getElementType());
349 // Determine whether the two operations are the same except that pointer-to-A
350 // and pointer-to-B are equivalent. This should be kept in sync with
351 // Instruction::isSameOperationAs.
352 bool FunctionComparator::isEquivalentOperation(const Instruction *I1,
353 const Instruction *I2) const {
354 // Differences from Instruction::isSameOperationAs:
355 // * replace type comparison with calls to isEquivalentType.
356 // * we test for I->hasSameSubclassOptionalData (nuw/nsw/tail) at the top
357 // * because of the above, we don't test for the tail bit on calls later on
358 if (I1->getOpcode() != I2->getOpcode() ||
359 I1->getNumOperands() != I2->getNumOperands() ||
360 !isEquivalentType(I1->getType(), I2->getType()) ||
361 !I1->hasSameSubclassOptionalData(I2))
364 // We have two instructions of identical opcode and #operands. Check to see
365 // if all operands are the same type
366 for (unsigned i = 0, e = I1->getNumOperands(); i != e; ++i)
367 if (!isEquivalentType(I1->getOperand(i)->getType(),
368 I2->getOperand(i)->getType()))
371 // Check special state that is a part of some instructions.
372 if (const LoadInst *LI = dyn_cast<LoadInst>(I1))
373 return LI->isVolatile() == cast<LoadInst>(I2)->isVolatile() &&
374 LI->getAlignment() == cast<LoadInst>(I2)->getAlignment() &&
375 LI->getOrdering() == cast<LoadInst>(I2)->getOrdering() &&
376 LI->getSynchScope() == cast<LoadInst>(I2)->getSynchScope();
377 if (const StoreInst *SI = dyn_cast<StoreInst>(I1))
378 return SI->isVolatile() == cast<StoreInst>(I2)->isVolatile() &&
379 SI->getAlignment() == cast<StoreInst>(I2)->getAlignment() &&
380 SI->getOrdering() == cast<StoreInst>(I2)->getOrdering() &&
381 SI->getSynchScope() == cast<StoreInst>(I2)->getSynchScope();
382 if (const CmpInst *CI = dyn_cast<CmpInst>(I1))
383 return CI->getPredicate() == cast<CmpInst>(I2)->getPredicate();
384 if (const CallInst *CI = dyn_cast<CallInst>(I1))
385 return CI->getCallingConv() == cast<CallInst>(I2)->getCallingConv() &&
386 CI->getAttributes() == cast<CallInst>(I2)->getAttributes();
387 if (const InvokeInst *CI = dyn_cast<InvokeInst>(I1))
388 return CI->getCallingConv() == cast<InvokeInst>(I2)->getCallingConv() &&
389 CI->getAttributes() == cast<InvokeInst>(I2)->getAttributes();
390 if (const InsertValueInst *IVI = dyn_cast<InsertValueInst>(I1))
391 return IVI->getIndices() == cast<InsertValueInst>(I2)->getIndices();
392 if (const ExtractValueInst *EVI = dyn_cast<ExtractValueInst>(I1))
393 return EVI->getIndices() == cast<ExtractValueInst>(I2)->getIndices();
394 if (const FenceInst *FI = dyn_cast<FenceInst>(I1))
395 return FI->getOrdering() == cast<FenceInst>(I2)->getOrdering() &&
396 FI->getSynchScope() == cast<FenceInst>(I2)->getSynchScope();
397 if (const AtomicCmpXchgInst *CXI = dyn_cast<AtomicCmpXchgInst>(I1))
398 return CXI->isVolatile() == cast<AtomicCmpXchgInst>(I2)->isVolatile() &&
399 CXI->getSuccessOrdering() ==
400 cast<AtomicCmpXchgInst>(I2)->getSuccessOrdering() &&
401 CXI->getFailureOrdering() ==
402 cast<AtomicCmpXchgInst>(I2)->getFailureOrdering() &&
403 CXI->getSynchScope() == cast<AtomicCmpXchgInst>(I2)->getSynchScope();
404 if (const AtomicRMWInst *RMWI = dyn_cast<AtomicRMWInst>(I1))
405 return RMWI->getOperation() == cast<AtomicRMWInst>(I2)->getOperation() &&
406 RMWI->isVolatile() == cast<AtomicRMWInst>(I2)->isVolatile() &&
407 RMWI->getOrdering() == cast<AtomicRMWInst>(I2)->getOrdering() &&
408 RMWI->getSynchScope() == cast<AtomicRMWInst>(I2)->getSynchScope();
413 // Determine whether two GEP operations perform the same underlying arithmetic.
414 bool FunctionComparator::isEquivalentGEP(const GEPOperator *GEP1,
415 const GEPOperator *GEP2) {
416 unsigned AS = GEP1->getPointerAddressSpace();
417 if (AS != GEP2->getPointerAddressSpace())
421 // When we have target data, we can reduce the GEP down to the value in bytes
422 // added to the address.
423 unsigned BitWidth = DL ? DL->getPointerSizeInBits(AS) : 1;
424 APInt Offset1(BitWidth, 0), Offset2(BitWidth, 0);
425 if (GEP1->accumulateConstantOffset(*DL, Offset1) &&
426 GEP2->accumulateConstantOffset(*DL, Offset2)) {
427 return Offset1 == Offset2;
431 if (GEP1->getPointerOperand()->getType() !=
432 GEP2->getPointerOperand()->getType())
435 if (GEP1->getNumOperands() != GEP2->getNumOperands())
438 for (unsigned i = 0, e = GEP1->getNumOperands(); i != e; ++i) {
439 if (!enumerate(GEP1->getOperand(i), GEP2->getOperand(i)))
446 // Compare two values used by the two functions under pair-wise comparison. If
447 // this is the first time the values are seen, they're added to the mapping so
448 // that we will detect mismatches on next use.
449 bool FunctionComparator::enumerate(const Value *V1, const Value *V2) {
450 // Check for function @f1 referring to itself and function @f2 referring to
451 // itself, or referring to each other, or both referring to either of them.
452 // They're all equivalent if the two functions are otherwise equivalent.
453 if (V1 == F1 && V2 == F2)
455 if (V1 == F2 && V2 == F1)
458 if (const Constant *C1 = dyn_cast<Constant>(V1)) {
459 if (V1 == V2) return true;
460 const Constant *C2 = dyn_cast<Constant>(V2);
461 if (!C2) return false;
462 // TODO: constant expressions with GEP or references to F1 or F2.
463 if (C1->isNullValue() && C2->isNullValue() &&
464 isEquivalentType(C1->getType(), C2->getType()))
466 // Try bitcasting C2 to C1's type. If the bitcast is legal and returns C1
467 // then they must have equal bit patterns.
468 return C1->getType()->canLosslesslyBitCastTo(C2->getType()) &&
469 C1 == ConstantExpr::getBitCast(const_cast<Constant*>(C2), C1->getType());
472 if (isa<InlineAsm>(V1) || isa<InlineAsm>(V2))
475 // Check that V1 maps to V2. If we find a value that V1 maps to then we simply
476 // check whether it's equal to V2. When there is no mapping then we need to
477 // ensure that V2 isn't already equivalent to something else. For this
478 // purpose, we track the V2 values in a set.
480 const Value *&map_elem = id_map[V1];
482 return map_elem == V2;
483 if (!seen_values.insert(V2).second)
489 // Test whether two basic blocks have equivalent behaviour.
490 bool FunctionComparator::compare(const BasicBlock *BB1, const BasicBlock *BB2) {
491 BasicBlock::const_iterator F1I = BB1->begin(), F1E = BB1->end();
492 BasicBlock::const_iterator F2I = BB2->begin(), F2E = BB2->end();
495 if (!enumerate(F1I, F2I))
498 if (const GetElementPtrInst *GEP1 = dyn_cast<GetElementPtrInst>(F1I)) {
499 const GetElementPtrInst *GEP2 = dyn_cast<GetElementPtrInst>(F2I);
503 if (!enumerate(GEP1->getPointerOperand(), GEP2->getPointerOperand()))
506 if (!isEquivalentGEP(GEP1, GEP2))
509 if (!isEquivalentOperation(F1I, F2I))
512 assert(F1I->getNumOperands() == F2I->getNumOperands());
513 for (unsigned i = 0, e = F1I->getNumOperands(); i != e; ++i) {
514 Value *OpF1 = F1I->getOperand(i);
515 Value *OpF2 = F2I->getOperand(i);
517 if (!enumerate(OpF1, OpF2))
520 if (OpF1->getValueID() != OpF2->getValueID() ||
521 !isEquivalentType(OpF1->getType(), OpF2->getType()))
527 } while (F1I != F1E && F2I != F2E);
529 return F1I == F1E && F2I == F2E;
532 // Test whether the two functions have equivalent behaviour.
533 bool FunctionComparator::compare() {
534 // We need to recheck everything, but check the things that weren't included
535 // in the hash first.
537 if (F1->getAttributes() != F2->getAttributes())
540 if (F1->hasGC() != F2->hasGC())
543 if (F1->hasGC() && F1->getGC() != F2->getGC())
546 if (F1->hasSection() != F2->hasSection())
549 if (F1->hasSection() && F1->getSection() != F2->getSection())
552 if (F1->isVarArg() != F2->isVarArg())
555 // TODO: if it's internal and only used in direct calls, we could handle this
557 if (F1->getCallingConv() != F2->getCallingConv())
560 if (!isEquivalentType(F1->getFunctionType(), F2->getFunctionType()))
563 assert(F1->arg_size() == F2->arg_size() &&
564 "Identically typed functions have different numbers of args!");
566 // Visit the arguments so that they get enumerated in the order they're
568 for (Function::const_arg_iterator f1i = F1->arg_begin(),
569 f2i = F2->arg_begin(), f1e = F1->arg_end(); f1i != f1e; ++f1i, ++f2i) {
570 if (!enumerate(f1i, f2i))
571 llvm_unreachable("Arguments repeat!");
574 // We do a CFG-ordered walk since the actual ordering of the blocks in the
575 // linked list is immaterial. Our walk starts at the entry block for both
576 // functions, then takes each block from each terminator in order. As an
577 // artifact, this also means that unreachable blocks are ignored.
578 SmallVector<const BasicBlock *, 8> F1BBs, F2BBs;
579 SmallSet<const BasicBlock *, 128> VisitedBBs; // in terms of F1.
581 F1BBs.push_back(&F1->getEntryBlock());
582 F2BBs.push_back(&F2->getEntryBlock());
584 VisitedBBs.insert(F1BBs[0]);
585 while (!F1BBs.empty()) {
586 const BasicBlock *F1BB = F1BBs.pop_back_val();
587 const BasicBlock *F2BB = F2BBs.pop_back_val();
589 if (!enumerate(F1BB, F2BB) || !compare(F1BB, F2BB))
592 const TerminatorInst *F1TI = F1BB->getTerminator();
593 const TerminatorInst *F2TI = F2BB->getTerminator();
595 assert(F1TI->getNumSuccessors() == F2TI->getNumSuccessors());
596 for (unsigned i = 0, e = F1TI->getNumSuccessors(); i != e; ++i) {
597 if (!VisitedBBs.insert(F1TI->getSuccessor(i)))
600 F1BBs.push_back(F1TI->getSuccessor(i));
601 F2BBs.push_back(F2TI->getSuccessor(i));
609 /// MergeFunctions finds functions which will generate identical machine code,
610 /// by considering all pointer types to be equivalent. Once identified,
611 /// MergeFunctions will fold them by replacing a call to one to a call to a
612 /// bitcast of the other.
614 class MergeFunctions : public ModulePass {
618 : ModulePass(ID), HasGlobalAliases(false) {
619 initializeMergeFunctionsPass(*PassRegistry::getPassRegistry());
622 bool runOnModule(Module &M) override;
625 typedef DenseSet<ComparableFunction> FnSetType;
627 /// A work queue of functions that may have been modified and should be
629 std::vector<WeakVH> Deferred;
631 /// Insert a ComparableFunction into the FnSet, or merge it away if it's
632 /// equal to one that's already present.
633 bool insert(ComparableFunction &NewF);
635 /// Remove a Function from the FnSet and queue it up for a second sweep of
637 void remove(Function *F);
639 /// Find the functions that use this Value and remove them from FnSet and
640 /// queue the functions.
641 void removeUsers(Value *V);
643 /// Replace all direct calls of Old with calls of New. Will bitcast New if
644 /// necessary to make types match.
645 void replaceDirectCallers(Function *Old, Function *New);
647 /// Merge two equivalent functions. Upon completion, G may be deleted, or may
648 /// be converted into a thunk. In either case, it should never be visited
650 void mergeTwoFunctions(Function *F, Function *G);
652 /// Replace G with a thunk or an alias to F. Deletes G.
653 void writeThunkOrAlias(Function *F, Function *G);
655 /// Replace G with a simple tail call to bitcast(F). Also replace direct uses
656 /// of G with bitcast(F). Deletes G.
657 void writeThunk(Function *F, Function *G);
659 /// Replace G with an alias to F. Deletes G.
660 void writeAlias(Function *F, Function *G);
662 /// The set of all distinct functions. Use the insert() and remove() methods
666 /// DataLayout for more accurate GEP comparisons. May be NULL.
667 const DataLayout *DL;
669 /// Whether or not the target supports global aliases.
670 bool HasGlobalAliases;
673 } // end anonymous namespace
675 char MergeFunctions::ID = 0;
676 INITIALIZE_PASS(MergeFunctions, "mergefunc", "Merge Functions", false, false)
678 ModulePass *llvm::createMergeFunctionsPass() {
679 return new MergeFunctions();
682 bool MergeFunctions::runOnModule(Module &M) {
683 bool Changed = false;
684 DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>();
685 DL = DLP ? &DLP->getDataLayout() : 0;
687 for (Module::iterator I = M.begin(), E = M.end(); I != E; ++I) {
688 if (!I->isDeclaration() && !I->hasAvailableExternallyLinkage())
689 Deferred.push_back(WeakVH(I));
691 FnSet.resize(Deferred.size());
694 std::vector<WeakVH> Worklist;
695 Deferred.swap(Worklist);
697 DEBUG(dbgs() << "size of module: " << M.size() << '\n');
698 DEBUG(dbgs() << "size of worklist: " << Worklist.size() << '\n');
700 // Insert only strong functions and merge them. Strong function merging
701 // always deletes one of them.
702 for (std::vector<WeakVH>::iterator I = Worklist.begin(),
703 E = Worklist.end(); I != E; ++I) {
705 Function *F = cast<Function>(*I);
706 if (!F->isDeclaration() && !F->hasAvailableExternallyLinkage() &&
707 !F->mayBeOverridden()) {
708 ComparableFunction CF = ComparableFunction(F, DL);
709 Changed |= insert(CF);
713 // Insert only weak functions and merge them. By doing these second we
714 // create thunks to the strong function when possible. When two weak
715 // functions are identical, we create a new strong function with two weak
716 // weak thunks to it which are identical but not mergable.
717 for (std::vector<WeakVH>::iterator I = Worklist.begin(),
718 E = Worklist.end(); I != E; ++I) {
720 Function *F = cast<Function>(*I);
721 if (!F->isDeclaration() && !F->hasAvailableExternallyLinkage() &&
722 F->mayBeOverridden()) {
723 ComparableFunction CF = ComparableFunction(F, DL);
724 Changed |= insert(CF);
727 DEBUG(dbgs() << "size of FnSet: " << FnSet.size() << '\n');
728 } while (!Deferred.empty());
735 bool DenseMapInfo<ComparableFunction>::isEqual(const ComparableFunction &LHS,
736 const ComparableFunction &RHS) {
737 if (LHS.getFunc() == RHS.getFunc() &&
738 LHS.getHash() == RHS.getHash())
740 if (!LHS.getFunc() || !RHS.getFunc())
743 // One of these is a special "underlying pointer comparison only" object.
744 if (LHS.getDataLayout() == ComparableFunction::LookupOnly ||
745 RHS.getDataLayout() == ComparableFunction::LookupOnly)
748 assert(LHS.getDataLayout() == RHS.getDataLayout() &&
749 "Comparing functions for different targets");
751 return FunctionComparator(LHS.getDataLayout(), LHS.getFunc(),
752 RHS.getFunc()).compare();
755 // Replace direct callers of Old with New.
756 void MergeFunctions::replaceDirectCallers(Function *Old, Function *New) {
757 Constant *BitcastNew = ConstantExpr::getBitCast(New, Old->getType());
758 for (auto UI = Old->use_begin(), UE = Old->use_end(); UI != UE;) {
761 CallSite CS(U->getUser());
762 if (CS && CS.isCallee(U)) {
763 remove(CS.getInstruction()->getParent()->getParent());
769 // Replace G with an alias to F if possible, or else a thunk to F. Deletes G.
770 void MergeFunctions::writeThunkOrAlias(Function *F, Function *G) {
771 if (HasGlobalAliases && G->hasUnnamedAddr()) {
772 if (G->hasExternalLinkage() || G->hasLocalLinkage() ||
773 G->hasWeakLinkage()) {
782 // Helper for writeThunk,
783 // Selects proper bitcast operation,
784 // but a bit simpler then CastInst::getCastOpcode.
785 static Value* createCast(IRBuilder<false> &Builder, Value *V, Type *DestTy) {
786 Type *SrcTy = V->getType();
787 if (SrcTy->isIntegerTy() && DestTy->isPointerTy())
788 return Builder.CreateIntToPtr(V, DestTy);
789 else if (SrcTy->isPointerTy() && DestTy->isIntegerTy())
790 return Builder.CreatePtrToInt(V, DestTy);
792 return Builder.CreateBitCast(V, DestTy);
795 // Replace G with a simple tail call to bitcast(F). Also replace direct uses
796 // of G with bitcast(F). Deletes G.
797 void MergeFunctions::writeThunk(Function *F, Function *G) {
798 if (!G->mayBeOverridden()) {
799 // Redirect direct callers of G to F.
800 replaceDirectCallers(G, F);
803 // If G was internal then we may have replaced all uses of G with F. If so,
804 // stop here and delete G. There's no need for a thunk.
805 if (G->hasLocalLinkage() && G->use_empty()) {
806 G->eraseFromParent();
810 Function *NewG = Function::Create(G->getFunctionType(), G->getLinkage(), "",
812 BasicBlock *BB = BasicBlock::Create(F->getContext(), "", NewG);
813 IRBuilder<false> Builder(BB);
815 SmallVector<Value *, 16> Args;
817 FunctionType *FFTy = F->getFunctionType();
818 for (Function::arg_iterator AI = NewG->arg_begin(), AE = NewG->arg_end();
820 Args.push_back(createCast(Builder, (Value*)AI, FFTy->getParamType(i)));
824 CallInst *CI = Builder.CreateCall(F, Args);
826 CI->setCallingConv(F->getCallingConv());
827 if (NewG->getReturnType()->isVoidTy()) {
828 Builder.CreateRetVoid();
830 Builder.CreateRet(createCast(Builder, CI, NewG->getReturnType()));
833 NewG->copyAttributesFrom(G);
836 G->replaceAllUsesWith(NewG);
837 G->eraseFromParent();
839 DEBUG(dbgs() << "writeThunk: " << NewG->getName() << '\n');
843 // Replace G with an alias to F and delete G.
844 void MergeFunctions::writeAlias(Function *F, Function *G) {
845 Constant *BitcastF = ConstantExpr::getBitCast(F, G->getType());
846 GlobalAlias *GA = new GlobalAlias(G->getType(), G->getLinkage(), "",
847 BitcastF, G->getParent());
848 F->setAlignment(std::max(F->getAlignment(), G->getAlignment()));
850 GA->setVisibility(G->getVisibility());
852 G->replaceAllUsesWith(GA);
853 G->eraseFromParent();
855 DEBUG(dbgs() << "writeAlias: " << GA->getName() << '\n');
859 // Merge two equivalent functions. Upon completion, Function G is deleted.
860 void MergeFunctions::mergeTwoFunctions(Function *F, Function *G) {
861 if (F->mayBeOverridden()) {
862 assert(G->mayBeOverridden());
864 if (HasGlobalAliases) {
865 // Make them both thunks to the same internal function.
866 Function *H = Function::Create(F->getFunctionType(), F->getLinkage(), "",
868 H->copyAttributesFrom(F);
871 F->replaceAllUsesWith(H);
873 unsigned MaxAlignment = std::max(G->getAlignment(), H->getAlignment());
878 F->setAlignment(MaxAlignment);
879 F->setLinkage(GlobalValue::PrivateLinkage);
881 // We can't merge them. Instead, pick one and update all direct callers
882 // to call it and hope that we improve the instruction cache hit rate.
883 replaceDirectCallers(G, F);
888 writeThunkOrAlias(F, G);
891 ++NumFunctionsMerged;
894 // Insert a ComparableFunction into the FnSet, or merge it away if equal to one
895 // that was already inserted.
896 bool MergeFunctions::insert(ComparableFunction &NewF) {
897 std::pair<FnSetType::iterator, bool> Result = FnSet.insert(NewF);
899 DEBUG(dbgs() << "Inserting as unique: " << NewF.getFunc()->getName() << '\n');
903 const ComparableFunction &OldF = *Result.first;
905 // Don't merge tiny functions, since it can just end up making the function
907 // FIXME: Should still merge them if they are unnamed_addr and produce an
909 if (NewF.getFunc()->size() == 1) {
910 if (NewF.getFunc()->front().size() <= 2) {
911 DEBUG(dbgs() << NewF.getFunc()->getName()
912 << " is to small to bother merging\n");
917 // Never thunk a strong function to a weak function.
918 assert(!OldF.getFunc()->mayBeOverridden() ||
919 NewF.getFunc()->mayBeOverridden());
921 DEBUG(dbgs() << " " << OldF.getFunc()->getName() << " == "
922 << NewF.getFunc()->getName() << '\n');
924 Function *DeleteF = NewF.getFunc();
926 mergeTwoFunctions(OldF.getFunc(), DeleteF);
930 // Remove a function from FnSet. If it was already in FnSet, add it to Deferred
931 // so that we'll look at it in the next round.
932 void MergeFunctions::remove(Function *F) {
933 // We need to make sure we remove F, not a function "equal" to F per the
934 // function equality comparator.
936 // The special "lookup only" ComparableFunction bypasses the expensive
937 // function comparison in favour of a pointer comparison on the underlying
939 ComparableFunction CF = ComparableFunction(F, ComparableFunction::LookupOnly);
940 if (FnSet.erase(CF)) {
941 DEBUG(dbgs() << "Removed " << F->getName() << " from set and deferred it.\n");
942 Deferred.push_back(F);
946 // For each instruction used by the value, remove() the function that contains
947 // the instruction. This should happen right before a call to RAUW.
948 void MergeFunctions::removeUsers(Value *V) {
949 std::vector<Value *> Worklist;
950 Worklist.push_back(V);
951 while (!Worklist.empty()) {
952 Value *V = Worklist.back();
955 for (User *U : V->users()) {
956 if (Instruction *I = dyn_cast<Instruction>(U)) {
957 remove(I->getParent()->getParent());
958 } else if (isa<GlobalValue>(U)) {
960 } else if (Constant *C = dyn_cast<Constant>(U)) {
961 for (User *UU : C->users())
962 Worklist.push_back(UU);