1//===- MergeFunctions.cpp - Merge identical functions ---------------------===//
2//
3//                     The LLVM Compiler Infrastructure
4//
5// This file is distributed under the University of Illinois Open Source
6// License. See LICENSE.TXT for details.
7//
8//===----------------------------------------------------------------------===//
9//
10// This pass looks for equivalent functions that are mergable and folds them.
11//
12// A hash is computed from the function, based on its type and number of
13// basic blocks.
14//
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.
19//
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.
23//
24//===----------------------------------------------------------------------===//
25//
26// Future work:
27//
28// * virtual functions.
29//
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.
33//
34// * switch from n^2 pair-wise comparisons to an n-way comparison for each
35// bucket.
36//
37// * be smarter about bitcasts.
38//
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.
43//
44//===----------------------------------------------------------------------===//
45
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/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/Pass.h"
62#include "llvm/Support/CallSite.h"
63#include "llvm/Support/Debug.h"
64#include "llvm/Support/ErrorHandling.h"
65#include "llvm/Support/ValueHandle.h"
66#include "llvm/Support/raw_ostream.h"
67#include <vector>
68using namespace llvm;
69
70STATISTIC(NumFunctionsMerged, "Number of functions merged");
71STATISTIC(NumThunksWritten, "Number of thunks generated");
72STATISTIC(NumAliasesWritten, "Number of aliases generated");
73STATISTIC(NumDoubleWeak, "Number of new functions created");
74
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.
78static Type::TypeID getTypeIDForHash(Type *Ty) {
79  if (Ty->isPointerTy())
80    return Type::IntegerTyID;
81  return Ty->getTypeID();
82}
83
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.
87static unsigned profileFunction(const Function *F) {
88  FunctionType *FTy = F->getFunctionType();
89
90  FoldingSetNodeID ID;
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();
99}
100
101namespace {
102
103/// ComparableFunction - A struct that pairs together functions with a
104/// DataLayout so that we can keep them together as elements in the DenseSet.
105class ComparableFunction {
106public:
107  static const ComparableFunction EmptyKey;
108  static const ComparableFunction TombstoneKey;
109  static DataLayout * const LookupOnly;
110
111  ComparableFunction(Function *Func, DataLayout *TD)
112    : Func(Func), Hash(profileFunction(Func)), TD(TD) {}
113
114  Function *getFunc() const { return Func; }
115  unsigned getHash() const { return Hash; }
116  DataLayout *getTD() const { return TD; }
117
118  // Drops AssertingVH reference to the function. Outside of debug mode, this
119  // does nothing.
120  void release() {
121    assert(Func &&
122           "Attempted to release function twice, or release empty/tombstone!");
123    Func = NULL;
124  }
125
126private:
127  explicit ComparableFunction(unsigned Hash)
128    : Func(NULL), Hash(Hash), TD(NULL) {}
129
130  AssertingVH<Function> Func;
131  unsigned Hash;
132  DataLayout *TD;
133};
134
135const ComparableFunction ComparableFunction::EmptyKey = ComparableFunction(0);
136const ComparableFunction ComparableFunction::TombstoneKey =
137    ComparableFunction(1);
138DataLayout *const ComparableFunction::LookupOnly = (DataLayout*)(-1);
139
140}
141
142namespace llvm {
143  template <>
144  struct DenseMapInfo<ComparableFunction> {
145    static ComparableFunction getEmptyKey() {
146      return ComparableFunction::EmptyKey;
147    }
148    static ComparableFunction getTombstoneKey() {
149      return ComparableFunction::TombstoneKey;
150    }
151    static unsigned getHashValue(const ComparableFunction &CF) {
152      return CF.getHash();
153    }
154    static bool isEqual(const ComparableFunction &LHS,
155                        const ComparableFunction &RHS);
156  };
157}
158
159namespace {
160
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).
165class FunctionComparator {
166public:
167  FunctionComparator(const DataLayout *TD, const Function *F1,
168                     const Function *F2)
169    : F1(F1), F2(F2), TD(TD) {}
170
171  /// Test whether the two functions have equivalent behaviour.
172  bool compare();
173
174private:
175  /// Test whether two basic blocks have equivalent behaviour.
176  bool compare(const BasicBlock *BB1, const BasicBlock *BB2);
177
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
180  /// visited.
181  bool enumerate(const Value *V1, const Value *V2);
182
183  /// Compare two Instructions for equivalence, similar to
184  /// Instruction::isSameOperationAs but with modifications to the type
185  /// comparison.
186  bool isEquivalentOperation(const Instruction *I1,
187                             const Instruction *I2) const;
188
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));
194  }
195
196  /// Compare two Types, treating all pointer types as equal.
197  bool isEquivalentType(Type *Ty1, Type *Ty2) const;
198
199  // The two functions undergoing comparison.
200  const Function *F1, *F2;
201
202  const DataLayout *TD;
203
204  DenseMap<const Value *, const Value *> id_map;
205  DenseSet<const Value *> seen_values;
206};
207
208}
209
210// Any two pointers in the same address space are equivalent, intptr_t and
211// pointers are equivalent. Otherwise, standard type equivalence rules apply.
212bool FunctionComparator::isEquivalentType(Type *Ty1, Type *Ty2) const {
213  if (Ty1 == Ty2)
214    return true;
215  if (Ty1->getTypeID() != Ty2->getTypeID()) {
216    if (TD) {
217      LLVMContext &Ctx = Ty1->getContext();
218      if (isa<PointerType>(Ty1) && Ty2 == TD->getIntPtrType(Ctx)) return true;
219      if (isa<PointerType>(Ty2) && Ty1 == TD->getIntPtrType(Ctx)) return true;
220    }
221    return false;
222  }
223
224  switch (Ty1->getTypeID()) {
225  default:
226    llvm_unreachable("Unknown type!");
227    // Fall through in Release mode.
228  case Type::IntegerTyID:
229  case Type::VectorTyID:
230    // Ty1 == Ty2 would have returned true earlier.
231    return false;
232
233  case Type::VoidTyID:
234  case Type::FloatTyID:
235  case Type::DoubleTyID:
236  case Type::X86_FP80TyID:
237  case Type::FP128TyID:
238  case Type::PPC_FP128TyID:
239  case Type::LabelTyID:
240  case Type::MetadataTyID:
241    return true;
242
243  case Type::PointerTyID: {
244    PointerType *PTy1 = cast<PointerType>(Ty1);
245    PointerType *PTy2 = cast<PointerType>(Ty2);
246    return PTy1->getAddressSpace() == PTy2->getAddressSpace();
247  }
248
249  case Type::StructTyID: {
250    StructType *STy1 = cast<StructType>(Ty1);
251    StructType *STy2 = cast<StructType>(Ty2);
252    if (STy1->getNumElements() != STy2->getNumElements())
253      return false;
254
255    if (STy1->isPacked() != STy2->isPacked())
256      return false;
257
258    for (unsigned i = 0, e = STy1->getNumElements(); i != e; ++i) {
259      if (!isEquivalentType(STy1->getElementType(i), STy2->getElementType(i)))
260        return false;
261    }
262    return true;
263  }
264
265  case Type::FunctionTyID: {
266    FunctionType *FTy1 = cast<FunctionType>(Ty1);
267    FunctionType *FTy2 = cast<FunctionType>(Ty2);
268    if (FTy1->getNumParams() != FTy2->getNumParams() ||
269        FTy1->isVarArg() != FTy2->isVarArg())
270      return false;
271
272    if (!isEquivalentType(FTy1->getReturnType(), FTy2->getReturnType()))
273      return false;
274
275    for (unsigned i = 0, e = FTy1->getNumParams(); i != e; ++i) {
276      if (!isEquivalentType(FTy1->getParamType(i), FTy2->getParamType(i)))
277        return false;
278    }
279    return true;
280  }
281
282  case Type::ArrayTyID: {
283    ArrayType *ATy1 = cast<ArrayType>(Ty1);
284    ArrayType *ATy2 = cast<ArrayType>(Ty2);
285    return ATy1->getNumElements() == ATy2->getNumElements() &&
286           isEquivalentType(ATy1->getElementType(), ATy2->getElementType());
287  }
288  }
289}
290
291// Determine whether the two operations are the same except that pointer-to-A
292// and pointer-to-B are equivalent. This should be kept in sync with
293// Instruction::isSameOperationAs.
294bool FunctionComparator::isEquivalentOperation(const Instruction *I1,
295                                               const Instruction *I2) const {
296  // Differences from Instruction::isSameOperationAs:
297  //  * replace type comparison with calls to isEquivalentType.
298  //  * we test for I->hasSameSubclassOptionalData (nuw/nsw/tail) at the top
299  //  * because of the above, we don't test for the tail bit on calls later on
300  if (I1->getOpcode() != I2->getOpcode() ||
301      I1->getNumOperands() != I2->getNumOperands() ||
302      !isEquivalentType(I1->getType(), I2->getType()) ||
303      !I1->hasSameSubclassOptionalData(I2))
304    return false;
305
306  // We have two instructions of identical opcode and #operands.  Check to see
307  // if all operands are the same type
308  for (unsigned i = 0, e = I1->getNumOperands(); i != e; ++i)
309    if (!isEquivalentType(I1->getOperand(i)->getType(),
310                          I2->getOperand(i)->getType()))
311      return false;
312
313  // Check special state that is a part of some instructions.
314  if (const LoadInst *LI = dyn_cast<LoadInst>(I1))
315    return LI->isVolatile() == cast<LoadInst>(I2)->isVolatile() &&
316           LI->getAlignment() == cast<LoadInst>(I2)->getAlignment() &&
317           LI->getOrdering() == cast<LoadInst>(I2)->getOrdering() &&
318           LI->getSynchScope() == cast<LoadInst>(I2)->getSynchScope();
319  if (const StoreInst *SI = dyn_cast<StoreInst>(I1))
320    return SI->isVolatile() == cast<StoreInst>(I2)->isVolatile() &&
321           SI->getAlignment() == cast<StoreInst>(I2)->getAlignment() &&
322           SI->getOrdering() == cast<StoreInst>(I2)->getOrdering() &&
323           SI->getSynchScope() == cast<StoreInst>(I2)->getSynchScope();
324  if (const CmpInst *CI = dyn_cast<CmpInst>(I1))
325    return CI->getPredicate() == cast<CmpInst>(I2)->getPredicate();
326  if (const CallInst *CI = dyn_cast<CallInst>(I1))
327    return CI->getCallingConv() == cast<CallInst>(I2)->getCallingConv() &&
328           CI->getAttributes() == cast<CallInst>(I2)->getAttributes();
329  if (const InvokeInst *CI = dyn_cast<InvokeInst>(I1))
330    return CI->getCallingConv() == cast<InvokeInst>(I2)->getCallingConv() &&
331           CI->getAttributes() == cast<InvokeInst>(I2)->getAttributes();
332  if (const InsertValueInst *IVI = dyn_cast<InsertValueInst>(I1))
333    return IVI->getIndices() == cast<InsertValueInst>(I2)->getIndices();
334  if (const ExtractValueInst *EVI = dyn_cast<ExtractValueInst>(I1))
335    return EVI->getIndices() == cast<ExtractValueInst>(I2)->getIndices();
336  if (const FenceInst *FI = dyn_cast<FenceInst>(I1))
337    return FI->getOrdering() == cast<FenceInst>(I2)->getOrdering() &&
338           FI->getSynchScope() == cast<FenceInst>(I2)->getSynchScope();
339  if (const AtomicCmpXchgInst *CXI = dyn_cast<AtomicCmpXchgInst>(I1))
340    return CXI->isVolatile() == cast<AtomicCmpXchgInst>(I2)->isVolatile() &&
341           CXI->getOrdering() == cast<AtomicCmpXchgInst>(I2)->getOrdering() &&
342           CXI->getSynchScope() == cast<AtomicCmpXchgInst>(I2)->getSynchScope();
343  if (const AtomicRMWInst *RMWI = dyn_cast<AtomicRMWInst>(I1))
344    return RMWI->getOperation() == cast<AtomicRMWInst>(I2)->getOperation() &&
345           RMWI->isVolatile() == cast<AtomicRMWInst>(I2)->isVolatile() &&
346           RMWI->getOrdering() == cast<AtomicRMWInst>(I2)->getOrdering() &&
347           RMWI->getSynchScope() == cast<AtomicRMWInst>(I2)->getSynchScope();
348
349  return true;
350}
351
352// Determine whether two GEP operations perform the same underlying arithmetic.
353bool FunctionComparator::isEquivalentGEP(const GEPOperator *GEP1,
354                                         const GEPOperator *GEP2) {
355  // When we have target data, we can reduce the GEP down to the value in bytes
356  // added to the address.
357  unsigned BitWidth = TD ? TD->getPointerSizeInBits() : 1;
358  APInt Offset1(BitWidth, 0), Offset2(BitWidth, 0);
359  if (TD &&
360      GEP1->accumulateConstantOffset(*TD, Offset1) &&
361      GEP2->accumulateConstantOffset(*TD, Offset2)) {
362    return Offset1 == Offset2;
363  }
364
365  if (GEP1->getPointerOperand()->getType() !=
366      GEP2->getPointerOperand()->getType())
367    return false;
368
369  if (GEP1->getNumOperands() != GEP2->getNumOperands())
370    return false;
371
372  for (unsigned i = 0, e = GEP1->getNumOperands(); i != e; ++i) {
373    if (!enumerate(GEP1->getOperand(i), GEP2->getOperand(i)))
374      return false;
375  }
376
377  return true;
378}
379
380// Compare two values used by the two functions under pair-wise comparison. If
381// this is the first time the values are seen, they're added to the mapping so
382// that we will detect mismatches on next use.
383bool FunctionComparator::enumerate(const Value *V1, const Value *V2) {
384  // Check for function @f1 referring to itself and function @f2 referring to
385  // itself, or referring to each other, or both referring to either of them.
386  // They're all equivalent if the two functions are otherwise equivalent.
387  if (V1 == F1 && V2 == F2)
388    return true;
389  if (V1 == F2 && V2 == F1)
390    return true;
391
392  if (const Constant *C1 = dyn_cast<Constant>(V1)) {
393    if (V1 == V2) return true;
394    const Constant *C2 = dyn_cast<Constant>(V2);
395    if (!C2) return false;
396    // TODO: constant expressions with GEP or references to F1 or F2.
397    if (C1->isNullValue() && C2->isNullValue() &&
398        isEquivalentType(C1->getType(), C2->getType()))
399      return true;
400    // Try bitcasting C2 to C1's type. If the bitcast is legal and returns C1
401    // then they must have equal bit patterns.
402    return C1->getType()->canLosslesslyBitCastTo(C2->getType()) &&
403      C1 == ConstantExpr::getBitCast(const_cast<Constant*>(C2), C1->getType());
404  }
405
406  if (isa<InlineAsm>(V1) || isa<InlineAsm>(V2))
407    return V1 == V2;
408
409  // Check that V1 maps to V2. If we find a value that V1 maps to then we simply
410  // check whether it's equal to V2. When there is no mapping then we need to
411  // ensure that V2 isn't already equivalent to something else. For this
412  // purpose, we track the V2 values in a set.
413
414  const Value *&map_elem = id_map[V1];
415  if (map_elem)
416    return map_elem == V2;
417  if (!seen_values.insert(V2).second)
418    return false;
419  map_elem = V2;
420  return true;
421}
422
423// Test whether two basic blocks have equivalent behaviour.
424bool FunctionComparator::compare(const BasicBlock *BB1, const BasicBlock *BB2) {
425  BasicBlock::const_iterator F1I = BB1->begin(), F1E = BB1->end();
426  BasicBlock::const_iterator F2I = BB2->begin(), F2E = BB2->end();
427
428  do {
429    if (!enumerate(F1I, F2I))
430      return false;
431
432    if (const GetElementPtrInst *GEP1 = dyn_cast<GetElementPtrInst>(F1I)) {
433      const GetElementPtrInst *GEP2 = dyn_cast<GetElementPtrInst>(F2I);
434      if (!GEP2)
435        return false;
436
437      if (!enumerate(GEP1->getPointerOperand(), GEP2->getPointerOperand()))
438        return false;
439
440      if (!isEquivalentGEP(GEP1, GEP2))
441        return false;
442    } else {
443      if (!isEquivalentOperation(F1I, F2I))
444        return false;
445
446      assert(F1I->getNumOperands() == F2I->getNumOperands());
447      for (unsigned i = 0, e = F1I->getNumOperands(); i != e; ++i) {
448        Value *OpF1 = F1I->getOperand(i);
449        Value *OpF2 = F2I->getOperand(i);
450
451        if (!enumerate(OpF1, OpF2))
452          return false;
453
454        if (OpF1->getValueID() != OpF2->getValueID() ||
455            !isEquivalentType(OpF1->getType(), OpF2->getType()))
456          return false;
457      }
458    }
459
460    ++F1I, ++F2I;
461  } while (F1I != F1E && F2I != F2E);
462
463  return F1I == F1E && F2I == F2E;
464}
465
466// Test whether the two functions have equivalent behaviour.
467bool FunctionComparator::compare() {
468  // We need to recheck everything, but check the things that weren't included
469  // in the hash first.
470
471  if (F1->getAttributes() != F2->getAttributes())
472    return false;
473
474  if (F1->hasGC() != F2->hasGC())
475    return false;
476
477  if (F1->hasGC() && F1->getGC() != F2->getGC())
478    return false;
479
480  if (F1->hasSection() != F2->hasSection())
481    return false;
482
483  if (F1->hasSection() && F1->getSection() != F2->getSection())
484    return false;
485
486  if (F1->isVarArg() != F2->isVarArg())
487    return false;
488
489  // TODO: if it's internal and only used in direct calls, we could handle this
490  // case too.
491  if (F1->getCallingConv() != F2->getCallingConv())
492    return false;
493
494  if (!isEquivalentType(F1->getFunctionType(), F2->getFunctionType()))
495    return false;
496
497  assert(F1->arg_size() == F2->arg_size() &&
498         "Identically typed functions have different numbers of args!");
499
500  // Visit the arguments so that they get enumerated in the order they're
501  // passed in.
502  for (Function::const_arg_iterator f1i = F1->arg_begin(),
503         f2i = F2->arg_begin(), f1e = F1->arg_end(); f1i != f1e; ++f1i, ++f2i) {
504    if (!enumerate(f1i, f2i))
505      llvm_unreachable("Arguments repeat!");
506  }
507
508  // We do a CFG-ordered walk since the actual ordering of the blocks in the
509  // linked list is immaterial. Our walk starts at the entry block for both
510  // functions, then takes each block from each terminator in order. As an
511  // artifact, this also means that unreachable blocks are ignored.
512  SmallVector<const BasicBlock *, 8> F1BBs, F2BBs;
513  SmallSet<const BasicBlock *, 128> VisitedBBs; // in terms of F1.
514
515  F1BBs.push_back(&F1->getEntryBlock());
516  F2BBs.push_back(&F2->getEntryBlock());
517
518  VisitedBBs.insert(F1BBs[0]);
519  while (!F1BBs.empty()) {
520    const BasicBlock *F1BB = F1BBs.pop_back_val();
521    const BasicBlock *F2BB = F2BBs.pop_back_val();
522
523    if (!enumerate(F1BB, F2BB) || !compare(F1BB, F2BB))
524      return false;
525
526    const TerminatorInst *F1TI = F1BB->getTerminator();
527    const TerminatorInst *F2TI = F2BB->getTerminator();
528
529    assert(F1TI->getNumSuccessors() == F2TI->getNumSuccessors());
530    for (unsigned i = 0, e = F1TI->getNumSuccessors(); i != e; ++i) {
531      if (!VisitedBBs.insert(F1TI->getSuccessor(i)))
532        continue;
533
534      F1BBs.push_back(F1TI->getSuccessor(i));
535      F2BBs.push_back(F2TI->getSuccessor(i));
536    }
537  }
538  return true;
539}
540
541namespace {
542
543/// MergeFunctions finds functions which will generate identical machine code,
544/// by considering all pointer types to be equivalent. Once identified,
545/// MergeFunctions will fold them by replacing a call to one to a call to a
546/// bitcast of the other.
547///
548class MergeFunctions : public ModulePass {
549public:
550  static char ID;
551  MergeFunctions()
552    : ModulePass(ID), HasGlobalAliases(false) {
553    initializeMergeFunctionsPass(*PassRegistry::getPassRegistry());
554  }
555
556  bool runOnModule(Module &M);
557
558private:
559  typedef DenseSet<ComparableFunction> FnSetType;
560
561  /// A work queue of functions that may have been modified and should be
562  /// analyzed again.
563  std::vector<WeakVH> Deferred;
564
565  /// Insert a ComparableFunction into the FnSet, or merge it away if it's
566  /// equal to one that's already present.
567  bool insert(ComparableFunction &NewF);
568
569  /// Remove a Function from the FnSet and queue it up for a second sweep of
570  /// analysis.
571  void remove(Function *F);
572
573  /// Find the functions that use this Value and remove them from FnSet and
574  /// queue the functions.
575  void removeUsers(Value *V);
576
577  /// Replace all direct calls of Old with calls of New. Will bitcast New if
578  /// necessary to make types match.
579  void replaceDirectCallers(Function *Old, Function *New);
580
581  /// Merge two equivalent functions. Upon completion, G may be deleted, or may
582  /// be converted into a thunk. In either case, it should never be visited
583  /// again.
584  void mergeTwoFunctions(Function *F, Function *G);
585
586  /// Replace G with a thunk or an alias to F. Deletes G.
587  void writeThunkOrAlias(Function *F, Function *G);
588
589  /// Replace G with a simple tail call to bitcast(F). Also replace direct uses
590  /// of G with bitcast(F). Deletes G.
591  void writeThunk(Function *F, Function *G);
592
593  /// Replace G with an alias to F. Deletes G.
594  void writeAlias(Function *F, Function *G);
595
596  /// The set of all distinct functions. Use the insert() and remove() methods
597  /// to modify it.
598  FnSetType FnSet;
599
600  /// DataLayout for more accurate GEP comparisons. May be NULL.
601  DataLayout *TD;
602
603  /// Whether or not the target supports global aliases.
604  bool HasGlobalAliases;
605};
606
607}  // end anonymous namespace
608
609char MergeFunctions::ID = 0;
610INITIALIZE_PASS(MergeFunctions, "mergefunc", "Merge Functions", false, false)
611
612ModulePass *llvm::createMergeFunctionsPass() {
613  return new MergeFunctions();
614}
615
616bool MergeFunctions::runOnModule(Module &M) {
617  bool Changed = false;
618  TD = getAnalysisIfAvailable<DataLayout>();
619
620  for (Module::iterator I = M.begin(), E = M.end(); I != E; ++I) {
621    if (!I->isDeclaration() && !I->hasAvailableExternallyLinkage())
622      Deferred.push_back(WeakVH(I));
623  }
624  FnSet.resize(Deferred.size());
625
626  do {
627    std::vector<WeakVH> Worklist;
628    Deferred.swap(Worklist);
629
630    DEBUG(dbgs() << "size of module: " << M.size() << '\n');
631    DEBUG(dbgs() << "size of worklist: " << Worklist.size() << '\n');
632
633    // Insert only strong functions and merge them. Strong function merging
634    // always deletes one of them.
635    for (std::vector<WeakVH>::iterator I = Worklist.begin(),
636           E = Worklist.end(); I != E; ++I) {
637      if (!*I) continue;
638      Function *F = cast<Function>(*I);
639      if (!F->isDeclaration() && !F->hasAvailableExternallyLinkage() &&
640          !F->mayBeOverridden()) {
641        ComparableFunction CF = ComparableFunction(F, TD);
642        Changed |= insert(CF);
643      }
644    }
645
646    // Insert only weak functions and merge them. By doing these second we
647    // create thunks to the strong function when possible. When two weak
648    // functions are identical, we create a new strong function with two weak
649    // weak thunks to it which are identical but not mergable.
650    for (std::vector<WeakVH>::iterator I = Worklist.begin(),
651           E = Worklist.end(); I != E; ++I) {
652      if (!*I) continue;
653      Function *F = cast<Function>(*I);
654      if (!F->isDeclaration() && !F->hasAvailableExternallyLinkage() &&
655          F->mayBeOverridden()) {
656        ComparableFunction CF = ComparableFunction(F, TD);
657        Changed |= insert(CF);
658      }
659    }
660    DEBUG(dbgs() << "size of FnSet: " << FnSet.size() << '\n');
661  } while (!Deferred.empty());
662
663  FnSet.clear();
664
665  return Changed;
666}
667
668bool DenseMapInfo<ComparableFunction>::isEqual(const ComparableFunction &LHS,
669                                               const ComparableFunction &RHS) {
670  if (LHS.getFunc() == RHS.getFunc() &&
671      LHS.getHash() == RHS.getHash())
672    return true;
673  if (!LHS.getFunc() || !RHS.getFunc())
674    return false;
675
676  // One of these is a special "underlying pointer comparison only" object.
677  if (LHS.getTD() == ComparableFunction::LookupOnly ||
678      RHS.getTD() == ComparableFunction::LookupOnly)
679    return false;
680
681  assert(LHS.getTD() == RHS.getTD() &&
682         "Comparing functions for different targets");
683
684  return FunctionComparator(LHS.getTD(), LHS.getFunc(),
685                            RHS.getFunc()).compare();
686}
687
688// Replace direct callers of Old with New.
689void MergeFunctions::replaceDirectCallers(Function *Old, Function *New) {
690  Constant *BitcastNew = ConstantExpr::getBitCast(New, Old->getType());
691  for (Value::use_iterator UI = Old->use_begin(), UE = Old->use_end();
692       UI != UE;) {
693    Value::use_iterator TheIter = UI;
694    ++UI;
695    CallSite CS(*TheIter);
696    if (CS && CS.isCallee(TheIter)) {
697      remove(CS.getInstruction()->getParent()->getParent());
698      TheIter.getUse().set(BitcastNew);
699    }
700  }
701}
702
703// Replace G with an alias to F if possible, or else a thunk to F. Deletes G.
704void MergeFunctions::writeThunkOrAlias(Function *F, Function *G) {
705  if (HasGlobalAliases && G->hasUnnamedAddr()) {
706    if (G->hasExternalLinkage() || G->hasLocalLinkage() ||
707        G->hasWeakLinkage()) {
708      writeAlias(F, G);
709      return;
710    }
711  }
712
713  writeThunk(F, G);
714}
715
716// Replace G with a simple tail call to bitcast(F). Also replace direct uses
717// of G with bitcast(F). Deletes G.
718void MergeFunctions::writeThunk(Function *F, Function *G) {
719  if (!G->mayBeOverridden()) {
720    // Redirect direct callers of G to F.
721    replaceDirectCallers(G, F);
722  }
723
724  // If G was internal then we may have replaced all uses of G with F. If so,
725  // stop here and delete G. There's no need for a thunk.
726  if (G->hasLocalLinkage() && G->use_empty()) {
727    G->eraseFromParent();
728    return;
729  }
730
731  Function *NewG = Function::Create(G->getFunctionType(), G->getLinkage(), "",
732                                    G->getParent());
733  BasicBlock *BB = BasicBlock::Create(F->getContext(), "", NewG);
734  IRBuilder<false> Builder(BB);
735
736  SmallVector<Value *, 16> Args;
737  unsigned i = 0;
738  FunctionType *FFTy = F->getFunctionType();
739  for (Function::arg_iterator AI = NewG->arg_begin(), AE = NewG->arg_end();
740       AI != AE; ++AI) {
741    Args.push_back(Builder.CreateBitCast(AI, FFTy->getParamType(i)));
742    ++i;
743  }
744
745  CallInst *CI = Builder.CreateCall(F, Args);
746  CI->setTailCall();
747  CI->setCallingConv(F->getCallingConv());
748  if (NewG->getReturnType()->isVoidTy()) {
749    Builder.CreateRetVoid();
750  } else {
751    Type *RetTy = NewG->getReturnType();
752    if (CI->getType()->isIntegerTy() && RetTy->isPointerTy())
753      Builder.CreateRet(Builder.CreateIntToPtr(CI, RetTy));
754    else if (CI->getType()->isPointerTy() && RetTy->isIntegerTy())
755      Builder.CreateRet(Builder.CreatePtrToInt(CI, RetTy));
756    else
757      Builder.CreateRet(Builder.CreateBitCast(CI, RetTy));
758  }
759
760  NewG->copyAttributesFrom(G);
761  NewG->takeName(G);
762  removeUsers(G);
763  G->replaceAllUsesWith(NewG);
764  G->eraseFromParent();
765
766  DEBUG(dbgs() << "writeThunk: " << NewG->getName() << '\n');
767  ++NumThunksWritten;
768}
769
770// Replace G with an alias to F and delete G.
771void MergeFunctions::writeAlias(Function *F, Function *G) {
772  Constant *BitcastF = ConstantExpr::getBitCast(F, G->getType());
773  GlobalAlias *GA = new GlobalAlias(G->getType(), G->getLinkage(), "",
774                                    BitcastF, G->getParent());
775  F->setAlignment(std::max(F->getAlignment(), G->getAlignment()));
776  GA->takeName(G);
777  GA->setVisibility(G->getVisibility());
778  removeUsers(G);
779  G->replaceAllUsesWith(GA);
780  G->eraseFromParent();
781
782  DEBUG(dbgs() << "writeAlias: " << GA->getName() << '\n');
783  ++NumAliasesWritten;
784}
785
786// Merge two equivalent functions. Upon completion, Function G is deleted.
787void MergeFunctions::mergeTwoFunctions(Function *F, Function *G) {
788  if (F->mayBeOverridden()) {
789    assert(G->mayBeOverridden());
790
791    if (HasGlobalAliases) {
792      // Make them both thunks to the same internal function.
793      Function *H = Function::Create(F->getFunctionType(), F->getLinkage(), "",
794                                     F->getParent());
795      H->copyAttributesFrom(F);
796      H->takeName(F);
797      removeUsers(F);
798      F->replaceAllUsesWith(H);
799
800      unsigned MaxAlignment = std::max(G->getAlignment(), H->getAlignment());
801
802      writeAlias(F, G);
803      writeAlias(F, H);
804
805      F->setAlignment(MaxAlignment);
806      F->setLinkage(GlobalValue::PrivateLinkage);
807    } else {
808      // We can't merge them. Instead, pick one and update all direct callers
809      // to call it and hope that we improve the instruction cache hit rate.
810      replaceDirectCallers(G, F);
811    }
812
813    ++NumDoubleWeak;
814  } else {
815    writeThunkOrAlias(F, G);
816  }
817
818  ++NumFunctionsMerged;
819}
820
821// Insert a ComparableFunction into the FnSet, or merge it away if equal to one
822// that was already inserted.
823bool MergeFunctions::insert(ComparableFunction &NewF) {
824  std::pair<FnSetType::iterator, bool> Result = FnSet.insert(NewF);
825  if (Result.second) {
826    DEBUG(dbgs() << "Inserting as unique: " << NewF.getFunc()->getName() << '\n');
827    return false;
828  }
829
830  const ComparableFunction &OldF = *Result.first;
831
832  // Never thunk a strong function to a weak function.
833  assert(!OldF.getFunc()->mayBeOverridden() ||
834         NewF.getFunc()->mayBeOverridden());
835
836  DEBUG(dbgs() << "  " << OldF.getFunc()->getName() << " == "
837               << NewF.getFunc()->getName() << '\n');
838
839  Function *DeleteF = NewF.getFunc();
840  NewF.release();
841  mergeTwoFunctions(OldF.getFunc(), DeleteF);
842  return true;
843}
844
845// Remove a function from FnSet. If it was already in FnSet, add it to Deferred
846// so that we'll look at it in the next round.
847void MergeFunctions::remove(Function *F) {
848  // We need to make sure we remove F, not a function "equal" to F per the
849  // function equality comparator.
850  //
851  // The special "lookup only" ComparableFunction bypasses the expensive
852  // function comparison in favour of a pointer comparison on the underlying
853  // Function*'s.
854  ComparableFunction CF = ComparableFunction(F, ComparableFunction::LookupOnly);
855  if (FnSet.erase(CF)) {
856    DEBUG(dbgs() << "Removed " << F->getName() << " from set and deferred it.\n");
857    Deferred.push_back(F);
858  }
859}
860
861// For each instruction used by the value, remove() the function that contains
862// the instruction. This should happen right before a call to RAUW.
863void MergeFunctions::removeUsers(Value *V) {
864  std::vector<Value *> Worklist;
865  Worklist.push_back(V);
866  while (!Worklist.empty()) {
867    Value *V = Worklist.back();
868    Worklist.pop_back();
869
870    for (Value::use_iterator UI = V->use_begin(), UE = V->use_end();
871         UI != UE; ++UI) {
872      Use &U = UI.getUse();
873      if (Instruction *I = dyn_cast<Instruction>(U.getUser())) {
874        remove(I->getParent()->getParent());
875      } else if (isa<GlobalValue>(U.getUser())) {
876        // do nothing
877      } else if (Constant *C = dyn_cast<Constant>(U.getUser())) {
878        for (Value::use_iterator CUI = C->use_begin(), CUE = C->use_end();
879             CUI != CUE; ++CUI)
880          Worklist.push_back(*CUI);
881      }
882    }
883  }
884}
885