GVN.cpp

  }
  
  // If this is a big-endian system, we need to shift the value down to the low
  // bits so that a truncate will work.
  if (TD.isBigEndian()) {
    Constant *Val = ConstantInt::get(StoredVal->getType(), StoreSize-LoadSize);
    StoredVal = BinaryOperator::CreateLShr(StoredVal, Val, "tmp", InsertPt);
  }
  
  // Truncate the integer to the right size now.
  const Type *NewIntTy = IntegerType::get(StoredValTy->getContext(), LoadSize);
  StoredVal = new TruncInst(StoredVal, NewIntTy, "trunc", InsertPt);
  
  if (LoadedTy == NewIntTy)
    return StoredVal;
  
  // If the result is a pointer, inttoptr.
  if (isa<PointerType>(LoadedTy))
    return new IntToPtrInst(StoredVal, LoadedTy, "inttoptr", InsertPt);
  
  // Otherwise, bitcast.
  return new BitCastInst(StoredVal, LoadedTy, "bitcast", InsertPt);
}

static void 
GetAvailableBlockValues(DenseMap<BasicBlock*, Value*> &BlockReplValues,
                        SmallVector<std::pair<BasicBlock*,
                                    Value*>, 16> &ValuesPerBlock,
                        const Type *LoadTy,
                        const TargetData *TD) {

  for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) {
    BasicBlock *BB = ValuesPerBlock[i].first;
    Value *AvailableVal = ValuesPerBlock[i].second;
    
    Value *&BlockEntry = BlockReplValues[BB];
    if (BlockEntry) continue;
    
    if (AvailableVal->getType() != LoadTy) {
      assert(TD && "Need target data to handle type mismatch case");
      AvailableVal = CoerceAvailableValueToLoadType(AvailableVal, LoadTy,
                                                    BB->getTerminator(), *TD);
      DEBUG(errs() << "GVN COERCED NONLOCAL VAL:\n"
                   << *ValuesPerBlock[i].second << '\n'
                   << *AvailableVal << '\n' << "\n\n\n");
    }
    BlockEntry = AvailableVal;
  }
}

/// processNonLocalLoad - Attempt to eliminate a load whose dependencies are
/// non-local by performing PHI construction.
bool GVN::processNonLocalLoad(LoadInst *LI,
                              SmallVectorImpl<Instruction*> &toErase) {
  // Find the non-local dependencies of the load.
  SmallVector<MemoryDependenceAnalysis::NonLocalDepEntry, 64> Deps;
  MD->getNonLocalPointerDependency(LI->getOperand(0), true, LI->getParent(),
                                   Deps);
  //DEBUG(errs() << "INVESTIGATING NONLOCAL LOAD: "
  //             << Deps.size() << *LI << '\n');

  // If we had to process more than one hundred blocks to find the
  // dependencies, this load isn't worth worrying about.  Optimizing
  // it will be too expensive.
  if (Deps.size() > 100)
    return false;

  // If we had a phi translation failure, we'll have a single entry which is a
  // clobber in the current block.  Reject this early.
  if (Deps.size() == 1 && Deps[0].second.isClobber()) {
    DEBUG(
      errs() << "GVN: non-local load ";
      WriteAsOperand(errs(), LI);
      errs() << " is clobbered by " << *Deps[0].second.getInst() << '\n';
    );
    return false;
  }

  // Filter out useless results (non-locals, etc).  Keep track of the blocks
  // where we have a value available in repl, also keep track of whether we see
  // dependencies that produce an unknown value for the load (such as a call
  // that could potentially clobber the load).
  SmallVector<std::pair<BasicBlock*, Value*>, 16> ValuesPerBlock;
  SmallVector<BasicBlock*, 16> UnavailableBlocks;

  const TargetData *TD = 0;
  
  for (unsigned i = 0, e = Deps.size(); i != e; ++i) {
    BasicBlock *DepBB = Deps[i].first;
    MemDepResult DepInfo = Deps[i].second;

    if (DepInfo.isClobber()) {
      UnavailableBlocks.push_back(DepBB);
      continue;
    }

    Instruction *DepInst = DepInfo.getInst();

    // Loading the allocation -> undef.
    if (isa<AllocationInst>(DepInst) || isMalloc(DepInst)) {
      ValuesPerBlock.push_back(std::make_pair(DepBB,
                               UndefValue::get(LI->getType())));
      continue;
    }

    if (StoreInst* S = dyn_cast<StoreInst>(DepInst)) {
      // Reject loads and stores that are to the same address but are of
      // different types if we have to.
      if (S->getOperand(0)->getType() != LI->getType()) {
        if (TD == 0)
          TD = getAnalysisIfAvailable<TargetData>();
        
        // If the stored value is larger or equal to the loaded value, we can
        // reuse it.
        if (TD == 0 || 
            TD->getTypeSizeInBits(S->getOperand(0)->getType()) <
              TD->getTypeSizeInBits(LI->getType())) {
          UnavailableBlocks.push_back(DepBB);
          continue;
        }
      }

      ValuesPerBlock.push_back(std::make_pair(DepBB, S->getOperand(0)));

    } else if (LoadInst* LD = dyn_cast<LoadInst>(DepInst)) {
      // If the types mismatch and we can't handle it, reject reuse of the load.
      if (LD->getType() != LI->getType()) {
        if (TD == 0)
          TD = getAnalysisIfAvailable<TargetData>();
        
        // If the stored value is larger or equal to the loaded value, we can
        // reuse it.
        if (TD == 0 || 
            TD->getTypeSizeInBits(LD->getType()) <
               TD->getTypeSizeInBits(LI->getType())) {
          UnavailableBlocks.push_back(DepBB);
          continue;
        }          
      }
      ValuesPerBlock.push_back(std::make_pair(DepBB, LD));
    } else {
      // FIXME: Handle memset/memcpy.
      UnavailableBlocks.push_back(DepBB);
      continue;
    }
  }

  // If we have no predecessors that produce a known value for this load, exit
  // early.
  if (ValuesPerBlock.empty()) return false;

  // If all of the instructions we depend on produce a known value for this
  // load, then it is fully redundant and we can use PHI insertion to compute
  // its value.  Insert PHIs and remove the fully redundant value now.
  if (UnavailableBlocks.empty()) {
    // Use cached PHI construction information from previous runs
    SmallPtrSet<Instruction*, 4> &p = phiMap[LI->getPointerOperand()];
    // FIXME: What does phiMap do? Are we positive it isn't getting invalidated?
    for (SmallPtrSet<Instruction*, 4>::iterator I = p.begin(), E = p.end();
         I != E; ++I) {
      if ((*I)->getParent() == LI->getParent()) {
        DEBUG(errs() << "GVN REMOVING NONLOCAL LOAD #1: " << *LI << '\n');
        LI->replaceAllUsesWith(*I);
        if (isa<PointerType>((*I)->getType()))
          MD->invalidateCachedPointerInfo(*I);
        toErase.push_back(LI);
        NumGVNLoad++;
        return true;
      }

      ValuesPerBlock.push_back(std::make_pair((*I)->getParent(), *I));
    }

    DEBUG(errs() << "GVN REMOVING NONLOCAL LOAD: " << *LI << '\n');

    // Convert the block information to a map, and insert coersions as needed.
    DenseMap<BasicBlock*, Value*> BlockReplValues;
    GetAvailableBlockValues(BlockReplValues, ValuesPerBlock, LI->getType(), TD);
    
    // Perform PHI construction.
    Value *V = GetValueForBlock(LI->getParent(), LI, BlockReplValues, true);
    LI->replaceAllUsesWith(V);

    if (isa<PHINode>(V))
      V->takeName(LI);
    if (isa<PointerType>(V->getType()))
      MD->invalidateCachedPointerInfo(V);
    toErase.push_back(LI);
    NumGVNLoad++;
    return true;
  }

  if (!EnablePRE || !EnableLoadPRE)
    return false;

  // Okay, we have *some* definitions of the value.  This means that the value
  // is available in some of our (transitive) predecessors.  Lets think about
  // doing PRE of this load.  This will involve inserting a new load into the
  // predecessor when it's not available.  We could do this in general, but
  // prefer to not increase code size.  As such, we only do this when we know
  // that we only have to insert *one* load (which means we're basically moving
  // the load, not inserting a new one).

  SmallPtrSet<BasicBlock *, 4> Blockers;
  for (unsigned i = 0, e = UnavailableBlocks.size(); i != e; ++i)
    Blockers.insert(UnavailableBlocks[i]);

  // Lets find first basic block with more than one predecessor.  Walk backwards
  // through predecessors if needed.
  BasicBlock *LoadBB = LI->getParent();
  BasicBlock *TmpBB = LoadBB;

  bool isSinglePred = false;
  bool allSingleSucc = true;
  while (TmpBB->getSinglePredecessor()) {
    isSinglePred = true;
    TmpBB = TmpBB->getSinglePredecessor();
    if (!TmpBB) // If haven't found any, bail now.
      return false;
    if (TmpBB == LoadBB) // Infinite (unreachable) loop.
      return false;
    if (Blockers.count(TmpBB))
      return false;
    if (TmpBB->getTerminator()->getNumSuccessors() != 1)
      allSingleSucc = false;
  }

  assert(TmpBB);
  LoadBB = TmpBB;

  // If we have a repl set with LI itself in it, this means we have a loop where
  // at least one of the values is LI.  Since this means that we won't be able
  // to eliminate LI even if we insert uses in the other predecessors, we will
  // end up increasing code size.  Reject this by scanning for LI.
  for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i)
    if (ValuesPerBlock[i].second == LI)
      return false;

  if (isSinglePred) {
    bool isHot = false;
    for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i)
      if (Instruction *I = dyn_cast<Instruction>(ValuesPerBlock[i].second))
        // "Hot" Instruction is in some loop (because it dominates its dep.
        // instruction).
        if (DT->dominates(LI, I)) {
          isHot = true;
          break;
        }

    // We are interested only in "hot" instructions. We don't want to do any
    // mis-optimizations here.
    if (!isHot)
      return false;
  }

  // Okay, we have some hope :).  Check to see if the loaded value is fully
  // available in all but one predecessor.
  // FIXME: If we could restructure the CFG, we could make a common pred with
  // all the preds that don't have an available LI and insert a new load into
  // that one block.
  BasicBlock *UnavailablePred = 0;

  DenseMap<BasicBlock*, char> FullyAvailableBlocks;
  for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i)
    FullyAvailableBlocks[ValuesPerBlock[i].first] = true;
  for (unsigned i = 0, e = UnavailableBlocks.size(); i != e; ++i)
    FullyAvailableBlocks[UnavailableBlocks[i]] = false;

  for (pred_iterator PI = pred_begin(LoadBB), E = pred_end(LoadBB);
       PI != E; ++PI) {
    if (IsValueFullyAvailableInBlock(*PI, FullyAvailableBlocks))
      continue;

    // If this load is not available in multiple predecessors, reject it.
    if (UnavailablePred && UnavailablePred != *PI)
      return false;
    UnavailablePred = *PI;
  }

  assert(UnavailablePred != 0 &&
         "Fully available value should be eliminated above!");

  // If the loaded pointer is PHI node defined in this block, do PHI translation
  // to get its value in the predecessor.
  Value *LoadPtr = LI->getOperand(0)->DoPHITranslation(LoadBB, UnavailablePred);

  // Make sure the value is live in the predecessor.  If it was defined by a
  // non-PHI instruction in this block, we don't know how to recompute it above.
  if (Instruction *LPInst = dyn_cast<Instruction>(LoadPtr))
    if (!DT->dominates(LPInst->getParent(), UnavailablePred)) {
      DEBUG(errs() << "COULDN'T PRE LOAD BECAUSE PTR IS UNAVAILABLE IN PRED: "
                   << *LPInst << '\n' << *LI << "\n");
      return false;
    }

  // We don't currently handle critical edges :(
  if (UnavailablePred->getTerminator()->getNumSuccessors() != 1) {
    DEBUG(errs() << "COULD NOT PRE LOAD BECAUSE OF CRITICAL EDGE '"
                 << UnavailablePred->getName() << "': " << *LI << '\n');
    return false;
  }

  // Make sure it is valid to move this load here.  We have to watch out for:
  //  @1 = getelementptr (i8* p, ...
  //  test p and branch if == 0
  //  load @1
  // It is valid to have the getelementptr before the test, even if p can be 0,
  // as getelementptr only does address arithmetic.
  // If we are not pushing the value through any multiple-successor blocks
  // we do not have this case.  Otherwise, check that the load is safe to
  // put anywhere; this can be improved, but should be conservatively safe.
  if (!allSingleSucc &&
      !isSafeToLoadUnconditionally(LoadPtr, UnavailablePred->getTerminator()))
    return false;

  // Okay, we can eliminate this load by inserting a reload in the predecessor
  // and using PHI construction to get the value in the other predecessors, do
  // it.
  DEBUG(errs() << "GVN REMOVING PRE LOAD: " << *LI << '\n');

  Value *NewLoad = new LoadInst(LoadPtr, LI->getName()+".pre", false,
                                LI->getAlignment(),
                                UnavailablePred->getTerminator());

  SmallPtrSet<Instruction*, 4> &p = phiMap[LI->getPointerOperand()];
  for (SmallPtrSet<Instruction*, 4>::iterator I = p.begin(), E = p.end();
       I != E; ++I)
    ValuesPerBlock.push_back(std::make_pair((*I)->getParent(), *I));

  DenseMap<BasicBlock*, Value*> BlockReplValues;
  GetAvailableBlockValues(BlockReplValues, ValuesPerBlock, LI->getType(), TD);
  BlockReplValues[UnavailablePred] = NewLoad;

  // Perform PHI construction.
  Value *V = GetValueForBlock(LI->getParent(), LI, BlockReplValues, true);
  LI->replaceAllUsesWith(V);
  if (isa<PHINode>(V))
    V->takeName(LI);
  if (isa<PointerType>(V->getType()))
    MD->invalidateCachedPointerInfo(V);
  toErase.push_back(LI);
  NumPRELoad++;
  return true;
}

/// GetBaseWithConstantOffset - Analyze the specified pointer to see if it can
/// be expressed as a base pointer plus a constant offset.  Return the base and
/// offset to the caller.
static Value *GetBaseWithConstantOffset(Value *Ptr, int64_t &Offset,
                                        const TargetData *TD) {
  Operator *PtrOp = dyn_cast<Operator>(Ptr);
  if (PtrOp == 0) return Ptr;
  
  // Just look through bitcasts.
  if (PtrOp->getOpcode() == Instruction::BitCast)
    return GetBaseWithConstantOffset(PtrOp->getOperand(0), Offset, TD);
  
  // If this is a GEP with constant indices, we can look through it.
  GEPOperator *GEP = dyn_cast<GEPOperator>(PtrOp);
  if (GEP == 0 || !GEP->hasAllConstantIndices()) return Ptr;
  
  gep_type_iterator GTI = gep_type_begin(GEP);
  for (User::op_iterator I = GEP->idx_begin(), E = GEP->idx_end(); I != E;
       ++I, ++GTI) {
    ConstantInt *OpC = cast<ConstantInt>(*I);
    if (OpC->isZero()) continue;
    
    // Handle a struct and array indices which add their offset to the pointer.
    if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
      Offset += TD->getStructLayout(STy)->getElementOffset(OpC->getZExtValue());
    } else {
      uint64_t Size = TD->getTypeAllocSize(GTI.getIndexedType());
      Offset += OpC->getSExtValue()*Size;
    }
  }
  
  // Re-sign extend from the pointer size if needed to get overflow edge cases
  // right.
  unsigned PtrSize = TD->getPointerSizeInBits();
  if (PtrSize < 64)
    Offset = (Offset << (64-PtrSize)) >> (64-PtrSize);
  
  return GetBaseWithConstantOffset(GEP->getPointerOperand(), Offset, TD);
}


/// AnalyzeLoadFromClobberingStore - This function is called when we have a
/// memdep query of a load that ends up being a clobbering store.  This means
/// that the store *may* provide bits used by the load but we can't be sure
/// because the pointers don't mustalias.  Check this case to see if there is
/// anything more we can do before we give up.  This returns -1 if we have to
/// give up, or a byte number in the stored value of the piece that feeds the
/// load.
static int AnalyzeLoadFromClobberingStore(LoadInst *L, StoreInst *DepSI,
                                          const TargetData *TD) {
  int64_t StoreOffset = 0, LoadOffset = 0;
  Value *StoreBase = 
    GetBaseWithConstantOffset(DepSI->getPointerOperand(), StoreOffset, TD);
  Value *LoadBase = 
    GetBaseWithConstantOffset(L->getPointerOperand(), LoadOffset, TD);
  if (StoreBase != LoadBase)
    return -1;

  // If the load and store are to the exact same address, they should have been
  // a must alias.  AA must have gotten confused.
  // FIXME: Study to see if/when this happens.
  if (LoadOffset == StoreOffset) {
#if 0
    errs() << "STORE/LOAD DEP WITH COMMON POINTER MISSED:\n"
    << "Base       = " << *StoreBase << "\n"
    << "Store Ptr  = " << *DepSI->getPointerOperand() << "\n"
    << "Store Offs = " << StoreOffset << " - " << *DepSI << "\n"
    << "Load Ptr   = " << *L->getPointerOperand() << "\n"
    << "Load Offs  = " << LoadOffset << " - " << *L << "\n\n";
    errs() << "'" << L->getParent()->getParent()->getName() << "'"
    << *L->getParent();
#endif
    return -1;
  }
  
  // If the load and store don't overlap at all, the store doesn't provide
  // anything to the load.  In this case, they really don't alias at all, AA
  // must have gotten confused.
  // FIXME: Investigate cases where this bails out, e.g. rdar://7238614. Then
  // remove this check, as it is duplicated with what we have below.
  uint64_t StoreSize = TD->getTypeSizeInBits(DepSI->getOperand(0)->getType());
  uint64_t LoadSize = TD->getTypeSizeInBits(L->getType());
  
  if ((StoreSize & 7) | (LoadSize & 7))
    return -1;
  StoreSize >>= 3;  // Convert to bytes.
  LoadSize >>= 3;
  
  
  bool isAAFailure = false;
  if (StoreOffset < LoadOffset) {
    isAAFailure = StoreOffset+int64_t(StoreSize) <= LoadOffset;
  } else {
    isAAFailure = LoadOffset+int64_t(LoadSize) <= StoreOffset;
  }
  if (isAAFailure) {
#if 0
    errs() << "STORE LOAD DEP WITH COMMON BASE:\n"
           << "Base       = " << *StoreBase << "\n"
           << "Store Ptr  = " << *DepSI->getPointerOperand() << "\n"
           << "Store Offs = " << StoreOffset << " - " << *DepSI << "\n"
           << "Load Ptr   = " << *L->getPointerOperand() << "\n"
           << "Load Offs  = " << LoadOffset << " - " << *L << "\n\n";
    errs() << "'" << L->getParent()->getParent()->getName() << "'"
           << *L->getParent();
#endif
    return -1;
  }
  
  // If the Load isn't completely contained within the stored bits, we don't
  // have all the bits to feed it.  We could do something crazy in the future
  // (issue a smaller load then merge the bits in) but this seems unlikely to be
  // valuable.
  if (StoreOffset > LoadOffset ||
      StoreOffset+StoreSize < LoadOffset+LoadSize)
    return -1;
  
  // Okay, we can do this transformation.  Return the number of bytes into the
  // store that the load is.
  return LoadOffset-StoreOffset;
}  
  
  
/// GetStoreValueForLoad - This function is called when we have a
/// memdep query of a load that ends up being a clobbering store.  This means
/// that the store *may* provide bits used by the load but we can't be sure
/// because the pointers don't mustalias.  Check this case to see if there is
/// anything more we can do before we give up.
static Value *GetStoreValueForLoad(Value *SrcVal, int Offset,const Type *LoadTy,
                                   Instruction *InsertPt, const TargetData *TD){
  LLVMContext &Ctx = SrcVal->getType()->getContext();

  uint64_t StoreSize = TD->getTypeSizeInBits(SrcVal->getType())/8;
  uint64_t LoadSize = TD->getTypeSizeInBits(LoadTy)/8;

  
  // Compute which bits of the stored value are being used by the load.  Convert
  // to an integer type to start with.
  if (isa<PointerType>(SrcVal->getType()))
    SrcVal = new PtrToIntInst(SrcVal, TD->getIntPtrType(Ctx), "tmp", InsertPt);
  if (!isa<IntegerType>(SrcVal->getType()))
    SrcVal = new BitCastInst(SrcVal, IntegerType::get(Ctx, StoreSize*8),
                             "tmp", InsertPt);
  
  // Shift the bits to the least significant depending on endianness.
  unsigned ShiftAmt;
  if (TD->isLittleEndian()) {
    ShiftAmt = Offset*8;
  } else {
    ShiftAmt = StoreSize-LoadSize-Offset;
  }

  SrcVal = BinaryOperator::CreateLShr(SrcVal,
                ConstantInt::get(SrcVal->getType(), ShiftAmt), "tmp", InsertPt);

  SrcVal = new TruncInst(SrcVal, IntegerType::get(Ctx, LoadSize*8),
                         "tmp", InsertPt);
  
  return CoerceAvailableValueToLoadType(SrcVal, LoadTy, InsertPt, *TD);
}



/// processLoad - Attempt to eliminate a load, first by eliminating it
/// locally, and then attempting non-local elimination if that fails.
bool GVN::processLoad(LoadInst *L, SmallVectorImpl<Instruction*> &toErase) {
  if (L->isVolatile())
    return false;

  // ... to a pointer that has been loaded from before...
  MemDepResult Dep = MD->getDependency(L);

  // If the value isn't available, don't do anything!
  if (Dep.isClobber()) {
    // FIXME: We should handle memset/memcpy/memmove as dependent instructions
    // to forward the value if available.
    //if (isa<MemIntrinsic>(Dep.getInst()))
    //errs() << "LOAD DEPENDS ON MEM: " << *L << "\n" << *Dep.getInst()<<"\n\n";
    
    // Check to see if we have something like this:
    //   store i32 123, i32* %P
    //   %A = bitcast i32* %P to i8*
    //   %B = gep i8* %A, i32 1
    //   %C = load i8* %B
    //
    // We could do that by recognizing if the clobber instructions are obviously
    // a common base + constant offset, and if the previous store (or memset)
    // completely covers this load.  This sort of thing can happen in bitfield
    // access code.
    if (StoreInst *DepSI = dyn_cast<StoreInst>(Dep.getInst()))
      if (const TargetData *TD = getAnalysisIfAvailable<TargetData>()) {
        int Offset = AnalyzeLoadFromClobberingStore(L, DepSI, TD);
        if (Offset != -1) {
          Value *AvailVal = GetStoreValueForLoad(DepSI->getOperand(0), Offset,
                                                 L->getType(), L, TD);
          DEBUG(errs() << "GVN COERCED STORE BITS:\n" << *DepSI << '\n'
                       << *AvailVal << '\n' << *L << "\n\n\n");
    
          // Replace the load!
          L->replaceAllUsesWith(AvailVal);
          if (isa<PointerType>(AvailVal->getType()))
            MD->invalidateCachedPointerInfo(AvailVal);
          toErase.push_back(L);
          NumGVNLoad++;
          return true;
        }
      }
    
    DEBUG(
      // fast print dep, using operator<< on instruction would be too slow
      errs() << "GVN: load ";
      WriteAsOperand(errs(), L);
      Instruction *I = Dep.getInst();
      errs() << " is clobbered by " << *I << '\n';
    );
    return false;
  }

  // If it is defined in another block, try harder.
  if (Dep.isNonLocal())
    return processNonLocalLoad(L, toErase);

  Instruction *DepInst = Dep.getInst();
  if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInst)) {
    Value *StoredVal = DepSI->getOperand(0);
    
    // The store and load are to a must-aliased pointer, but they may not
    // actually have the same type.  See if we know how to reuse the stored
    // value (depending on its type).
    const TargetData *TD = 0;
    if (StoredVal->getType() != L->getType() &&
        (TD = getAnalysisIfAvailable<TargetData>())) {
      StoredVal = CoerceAvailableValueToLoadType(StoredVal, L->getType(), L, *TD);
      if (StoredVal == 0)
        return false;
      
      DEBUG(errs() << "GVN COERCED STORE:\n" << *DepSI << '\n' << *StoredVal
                   << '\n' << *L << "\n\n\n");
    }

    // Remove it!
    L->replaceAllUsesWith(StoredVal);
    if (isa<PointerType>(StoredVal->getType()))
      MD->invalidateCachedPointerInfo(StoredVal);
    toErase.push_back(L);
    NumGVNLoad++;
    return true;
  }

  if (LoadInst *DepLI = dyn_cast<LoadInst>(DepInst)) {
    Value *AvailableVal = DepLI;
    
    // The loads are of a must-aliased pointer, but they may not actually have
    // the same type.  See if we know how to reuse the previously loaded value
    // (depending on its type).
    const TargetData *TD = 0;
    if (DepLI->getType() != L->getType() &&
        (TD = getAnalysisIfAvailable<TargetData>())) {
      AvailableVal = CoerceAvailableValueToLoadType(DepLI, L->getType(), L, *TD);
      if (AvailableVal == 0)
        return false;
      
      DEBUG(errs() << "GVN COERCED LOAD:\n" << *DepLI << "\n" << *AvailableVal
                   << "\n" << *L << "\n\n\n");
    }
    
    // Remove it!
    L->replaceAllUsesWith(AvailableVal);
    if (isa<PointerType>(DepLI->getType()))
      MD->invalidateCachedPointerInfo(DepLI);
    toErase.push_back(L);
    NumGVNLoad++;
    return true;
  }

  // If this load really doesn't depend on anything, then we must be loading an
  // undef value.  This can happen when loading for a fresh allocation with no
  // intervening stores, for example.
  if (isa<AllocationInst>(DepInst) || isMalloc(DepInst)) {
    L->replaceAllUsesWith(UndefValue::get(L->getType()));
    toErase.push_back(L);
    NumGVNLoad++;
    return true;
  }

  return false;
}

Value *GVN::lookupNumber(BasicBlock *BB, uint32_t num) {
  DenseMap<BasicBlock*, ValueNumberScope*>::iterator I = localAvail.find(BB);
  if (I == localAvail.end())
    return 0;

  ValueNumberScope *Locals = I->second;
  while (Locals) {
    DenseMap<uint32_t, Value*>::iterator I = Locals->table.find(num);
    if (I != Locals->table.end())
      return I->second;
    Locals = Locals->parent;
  }

  return 0;
}

/// AttemptRedundancyElimination - If the "fast path" of redundancy elimination
/// by inheritance from the dominator fails, see if we can perform phi
/// construction to eliminate the redundancy.
Value *GVN::AttemptRedundancyElimination(Instruction *orig, unsigned valno) {
  BasicBlock *BaseBlock = orig->getParent();

  SmallPtrSet<BasicBlock*, 4> Visited;
  SmallVector<BasicBlock*, 8> Stack;
  Stack.push_back(BaseBlock);

  DenseMap<BasicBlock*, Value*> Results;

  // Walk backwards through our predecessors, looking for instances of the
  // value number we're looking for.  Instances are recorded in the Results
  // map, which is then used to perform phi construction.
  while (!Stack.empty()) {
    BasicBlock *Current = Stack.back();
    Stack.pop_back();

    // If we've walked all the way to a proper dominator, then give up. Cases
    // where the instance is in the dominator will have been caught by the fast
    // path, and any cases that require phi construction further than this are
    // probably not worth it anyways.  Note that this is a SIGNIFICANT compile
    // time improvement.
    if (DT->properlyDominates(Current, orig->getParent())) return 0;

    DenseMap<BasicBlock*, ValueNumberScope*>::iterator LA =
                                                       localAvail.find(Current);
    if (LA == localAvail.end()) return 0;
    DenseMap<uint32_t, Value*>::iterator V = LA->second->table.find(valno);

    if (V != LA->second->table.end()) {
      // Found an instance, record it.
      Results.insert(std::make_pair(Current, V->second));
      continue;
    }

    // If we reach the beginning of the function, then give up.
    if (pred_begin(Current) == pred_end(Current))
      return 0;

    for (pred_iterator PI = pred_begin(Current), PE = pred_end(Current);
         PI != PE; ++PI)
      if (Visited.insert(*PI))
        Stack.push_back(*PI);
  }

  // If we didn't find instances, give up.  Otherwise, perform phi construction.
  if (Results.size() == 0)
    return 0;
  else
    return GetValueForBlock(BaseBlock, orig, Results, true);
}

/// processInstruction - When calculating availability, handle an instruction
/// by inserting it into the appropriate sets
bool GVN::processInstruction(Instruction *I,
                             SmallVectorImpl<Instruction*> &toErase) {
  if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
    bool Changed = processLoad(LI, toErase);

    if (!Changed) {
      unsigned Num = VN.lookup_or_add(LI);
      localAvail[I->getParent()]->table.insert(std::make_pair(Num, LI));
    }

    return Changed;
  }

  uint32_t NextNum = VN.getNextUnusedValueNumber();
  unsigned Num = VN.lookup_or_add(I);

  if (BranchInst *BI = dyn_cast<BranchInst>(I)) {
    localAvail[I->getParent()]->table.insert(std::make_pair(Num, I));

    if (!BI->isConditional() || isa<Constant>(BI->getCondition()))
      return false;

    Value *BranchCond = BI->getCondition();
    uint32_t CondVN = VN.lookup_or_add(BranchCond);

    BasicBlock *TrueSucc = BI->getSuccessor(0);
    BasicBlock *FalseSucc = BI->getSuccessor(1);

    if (TrueSucc->getSinglePredecessor())
      localAvail[TrueSucc]->table[CondVN] =
        ConstantInt::getTrue(TrueSucc->getContext());
    if (FalseSucc->getSinglePredecessor())
      localAvail[FalseSucc]->table[CondVN] =
        ConstantInt::getFalse(TrueSucc->getContext());

    return false;

  // Allocations are always uniquely numbered, so we can save time and memory
  // by fast failing them.
  } else if (isa<AllocationInst>(I) || isMalloc(I) || isa<TerminatorInst>(I)) {
    localAvail[I->getParent()]->table.insert(std::make_pair(Num, I));
    return false;
  }

  // Collapse PHI nodes
  if (PHINode* p = dyn_cast<PHINode>(I)) {
    Value *constVal = CollapsePhi(p);

    if (constVal) {
      for (PhiMapType::iterator PI = phiMap.begin(), PE = phiMap.end();
           PI != PE; ++PI)
        PI->second.erase(p);

      p->replaceAllUsesWith(constVal);
      if (isa<PointerType>(constVal->getType()))
        MD->invalidateCachedPointerInfo(constVal);
      VN.erase(p);

      toErase.push_back(p);
    } else {
      localAvail[I->getParent()]->table.insert(std::make_pair(Num, I));
    }

  // If the number we were assigned was a brand new VN, then we don't
  // need to do a lookup to see if the number already exists
  // somewhere in the domtree: it can't!
  } else if (Num == NextNum) {
    localAvail[I->getParent()]->table.insert(std::make_pair(Num, I));

  // Perform fast-path value-number based elimination of values inherited from
  // dominators.
  } else if (Value *repl = lookupNumber(I->getParent(), Num)) {
    // Remove it!
    VN.erase(I);
    I->replaceAllUsesWith(repl);
    if (isa<PointerType>(repl->getType()))
      MD->invalidateCachedPointerInfo(repl);
    toErase.push_back(I);
    return true;

#if 0
  // Perform slow-pathvalue-number based elimination with phi construction.
  } else if (Value *repl = AttemptRedundancyElimination(I, Num)) {
    // Remove it!
    VN.erase(I);
    I->replaceAllUsesWith(repl);
    if (isa<PointerType>(repl->getType()))
      MD->invalidateCachedPointerInfo(repl);
    toErase.push_back(I);
    return true;
#endif
  } else {
    localAvail[I->getParent()]->table.insert(std::make_pair(Num, I));
  }

  return false;
}

/// runOnFunction - This is the main transformation entry point for a function.
bool GVN::runOnFunction(Function& F) {
  MD = &getAnalysis<MemoryDependenceAnalysis>();
  DT = &getAnalysis<DominatorTree>();
  VN.setAliasAnalysis(&getAnalysis<AliasAnalysis>());
  VN.setMemDep(MD);
  VN.setDomTree(DT);

  bool Changed = false;
  bool ShouldContinue = true;

  // Merge unconditional branches, allowing PRE to catch more
  // optimization opportunities.
  for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; ) {
    BasicBlock *BB = FI;
    ++FI;
    bool removedBlock = MergeBlockIntoPredecessor(BB, this);
    if (removedBlock) NumGVNBlocks++;

    Changed |= removedBlock;
  }

  unsigned Iteration = 0;

  while (ShouldContinue) {
    DEBUG(errs() << "GVN iteration: " << Iteration << "\n");
    ShouldContinue = iterateOnFunction(F);
    Changed |= ShouldContinue;
    ++Iteration;
  }

  if (EnablePRE) {
    bool PREChanged = true;
    while (PREChanged) {
      PREChanged = performPRE(F);
      Changed |= PREChanged;
    }
  }
  // FIXME: Should perform GVN again after PRE does something.  PRE can move
  // computations into blocks where they become fully redundant.  Note that
  // we can't do this until PRE's critical edge splitting updates memdep.
  // Actually, when this happens, we should just fully integrate PRE into GVN.

  cleanupGlobalSets();

  return Changed;
}


bool GVN::processBlock(BasicBlock *BB) {
  // FIXME: Kill off toErase by doing erasing eagerly in a helper function (and
  // incrementing BI before processing an instruction).
  SmallVector<Instruction*, 8> toErase;
  bool ChangedFunction = false;

  for (BasicBlock::iterator BI = BB->begin(), BE = BB->end();
       BI != BE;) {
    ChangedFunction |= processInstruction(BI, toErase);
    if (toErase.empty()) {
      ++BI;
      continue;
    }

    // If we need some instructions deleted, do it now.
    NumGVNInstr += toErase.size();

    // Avoid iterator invalidation.
    bool AtStart = BI == BB->begin();
    if (!AtStart)
      --BI;

    for (SmallVector<Instruction*, 4>::iterator I = toErase.begin(),
         E = toErase.end(); I != E; ++I) {
      DEBUG(errs() << "GVN removed: " << **I << '\n');
      MD->removeInstruction(*I);
      (*I)->eraseFromParent();
      DEBUG(verifyRemoved(*I));
    }
    toErase.clear();

    if (AtStart)
      BI = BB->begin();
    else
      ++BI;
  }

  return ChangedFunction;
}

/// performPRE - Perform a purely local form of PRE that looks for diamond
/// control flow patterns and attempts to perform simple PRE at the join point.
bool GVN::performPRE(Function& F) {
  bool Changed = false;
  SmallVector<std::pair<TerminatorInst*, unsigned>, 4> toSplit;
  DenseMap<BasicBlock*, Value*> predMap;
  for (df_iterator<BasicBlock*> DI = df_begin(&F.getEntryBlock()),
       DE = df_end(&F.getEntryBlock()); DI != DE; ++DI) {
    BasicBlock *CurrentBlock = *DI;

    // Nothing to PRE in the entry block.
    if (CurrentBlock == &F.getEntryBlock()) continue;

    for (BasicBlock::iterator BI = CurrentBlock->begin(),
         BE = CurrentBlock->end(); BI != BE; ) {
      Instruction *CurInst = BI++;

      if (isa<AllocationInst>(CurInst) || isMalloc(CurInst) ||
          isa<TerminatorInst>(CurInst) || isa<PHINode>(CurInst) ||
          (CurInst->getType() == Type::getVoidTy(F.getContext())) ||
          CurInst->mayReadFromMemory() || CurInst->mayHaveSideEffects() ||
          isa<DbgInfoIntrinsic>(CurInst))
        continue;

      uint32_t ValNo = VN.lookup(CurInst);

      // Look for the predecessors for PRE opportunities.  We're
      // only trying to solve the basic diamond case, where
      // a value is computed in the successor and one predecessor,
      // but not the other.  We also explicitly disallow cases
      // where the successor is its own predecessor, because they're
      // more complicated to get right.
      unsigned NumWith = 0;
      unsigned NumWithout = 0;
      BasicBlock *PREPred = 0;
      predMap.clear();

      for (pred_iterator PI = pred_begin(CurrentBlock),
           PE = pred_end(CurrentBlock); PI != PE; ++PI) {
        // We're not interested in PRE where the block is its
        // own predecessor, on in blocks with predecessors
        // that are not reachable.
        if (*PI == CurrentBlock) {
          NumWithout = 2;
          break;
        } else if (!localAvail.count(*PI))  {
          NumWithout = 2;
          break;
        }

        DenseMap<uint32_t, Value*>::iterator predV =
                                            localAvail[*PI]->table.find(ValNo);
        if (predV == localAvail[*PI]->table.end()) {
          PREPred = *PI;
          NumWithout++;
        } else if (predV->second == CurInst) {
          NumWithout = 2;
        } else {
          predMap[*PI] = predV->second;
          NumWith++;
        }
      }

      // Don't do PRE when it might increase code size, i.e. when
      // we would need to insert instructions in more than one pred.
      if (NumWithout != 1 || NumWith == 0)
        continue;

      // We can't do PRE safely on a critical edge, so instead we schedule
      // the edge to be split and perform the PRE the next time we iterate
      // on the function.
      unsigned SuccNum = 0;
      for (unsigned i = 0, e = PREPred->getTerminator()->getNumSuccessors();
           i != e; ++i)
        if (PREPred->getTerminator()->getSuccessor(i) == CurrentBlock) {
          SuccNum = i;
          break;
        }

      if (isCriticalEdge(PREPred->getTerminator(), SuccNum)) {
        toSplit.push_back(std::make_pair(PREPred->getTerminator(), SuccNum));
        continue;
      }

      // Instantiate the expression the in predecessor that lacked it.
      // Because we are going top-down through the block, all value numbers
      // will be available in the predecessor by the time we need them.  Any
      // that weren't original present will have been instantiated earlier
      // in this loop.
      Instruction *PREInstr = CurInst->clone(CurInst->getContext());
      bool success = true;
      for (unsigned i = 0, e = CurInst->getNumOperands(); i != e; ++i) {
        Value *Op = PREInstr->getOperand(i);
        if (isa<Argument>(Op) || isa<Constant>(Op) || isa<GlobalValue>(Op))
          continue;

        if (Value *V = lookupNumber(PREPred, VN.lookup(Op))) {
          PREInstr->setOperand(i, V);
        } else {
          success = false;
          break;
        }
      }

      // Fail out if we encounter an operand that is not available in
      // the PRE predecessor.  This is typically because of loads which
      // are not value numbered precisely.
      if (!success) {