GVN.cpp

      Val = Builder.CreateOr(OneElt, ShVal);
      ++NumBytesSet;
    }
    
    return CoerceAvailableValueToLoadType(Val, LoadTy, InsertPt, TD);
  }
 
  // Otherwise, this is a memcpy/memmove from a constant global.
  MemTransferInst *MTI = cast<MemTransferInst>(SrcInst);
  Constant *Src = cast<Constant>(MTI->getSource());

  // Otherwise, see if we can constant fold a load from the constant with the
  // offset applied as appropriate.
  Src = ConstantExpr::getBitCast(Src,
                                 llvm::Type::getInt8PtrTy(Src->getContext()));
  Constant *OffsetCst = 
  ConstantInt::get(Type::getInt64Ty(Src->getContext()), (unsigned)Offset);
  Src = ConstantExpr::getGetElementPtr(Src, &OffsetCst, 1);
  Src = ConstantExpr::getBitCast(Src, PointerType::getUnqual(LoadTy));
  return ConstantFoldLoadFromConstPtr(Src, &TD);
}

namespace {

struct AvailableValueInBlock {
  /// BB - The basic block in question.
  BasicBlock *BB;
  enum ValType {
    SimpleVal,  // A simple offsetted value that is accessed.
    LoadVal,    // A value produced by a load.
    MemIntrin   // A memory intrinsic which is loaded from.
  };
  
  /// V - The value that is live out of the block.
  PointerIntPair<Value *, 2, ValType> Val;
  
  /// Offset - The byte offset in Val that is interesting for the load query.
  unsigned Offset;
  
  static AvailableValueInBlock get(BasicBlock *BB, Value *V,
                                   unsigned Offset = 0) {
    AvailableValueInBlock Res;
    Res.BB = BB;
    Res.Val.setPointer(V);
    Res.Val.setInt(SimpleVal);
    Res.Offset = Offset;
    return Res;
  }

  static AvailableValueInBlock getMI(BasicBlock *BB, MemIntrinsic *MI,
                                     unsigned Offset = 0) {
    AvailableValueInBlock Res;
    Res.BB = BB;
    Res.Val.setPointer(MI);
    Res.Val.setInt(MemIntrin);
    Res.Offset = Offset;
    return Res;
  }
  
  static AvailableValueInBlock getLoad(BasicBlock *BB, LoadInst *LI,
                                       unsigned Offset = 0) {
    AvailableValueInBlock Res;
    Res.BB = BB;
    Res.Val.setPointer(LI);
    Res.Val.setInt(LoadVal);
    Res.Offset = Offset;
    return Res;
  }

  bool isSimpleValue() const { return Val.getInt() == SimpleVal; }
  bool isCoercedLoadValue() const { return Val.getInt() == LoadVal; }
  bool isMemIntrinValue() const { return Val.getInt() == MemIntrin; }

  Value *getSimpleValue() const {
    assert(isSimpleValue() && "Wrong accessor");
    return Val.getPointer();
  }
  
  LoadInst *getCoercedLoadValue() const {
    assert(isCoercedLoadValue() && "Wrong accessor");
    return cast<LoadInst>(Val.getPointer());
  }
  
  MemIntrinsic *getMemIntrinValue() const {
    assert(isMemIntrinValue() && "Wrong accessor");
    return cast<MemIntrinsic>(Val.getPointer());
  }
  
  /// MaterializeAdjustedValue - Emit code into this block to adjust the value
  /// defined here to the specified type.  This handles various coercion cases.
  Value *MaterializeAdjustedValue(const Type *LoadTy,
                                  const TargetData *TD,
                                  MemoryDependenceAnalysis &MD) const {
    Value *Res;
    if (isSimpleValue()) {
      Res = getSimpleValue();
      if (Res->getType() != LoadTy) {
        assert(TD && "Need target data to handle type mismatch case");
        Res = GetStoreValueForLoad(Res, Offset, LoadTy, BB->getTerminator(),
                                   *TD);
        
        DEBUG(dbgs() << "GVN COERCED NONLOCAL VAL:\nOffset: " << Offset << "  "
                     << *getSimpleValue() << '\n'
                     << *Res << '\n' << "\n\n\n");
      }
    } else if (isCoercedLoadValue()) {
      LoadInst *Load = getCoercedLoadValue();
      if (Load->getType() == LoadTy && Offset == 0) {
        Res = Load;
      } else {
        assert(TD && "Need target data to handle type mismatch case");
        Res = GetLoadValueForLoad(Load, Offset, LoadTy, BB->getTerminator(),
                                  *TD, MD);
        
        DEBUG(dbgs() << "GVN COERCED NONLOCAL LOAD:\nOffset: " << Offset << "  "
                     << *getCoercedLoadValue() << '\n'
                     << *Res << '\n' << "\n\n\n");
      }
    } else {
      Res = GetMemInstValueForLoad(getMemIntrinValue(), Offset,
                                   LoadTy, BB->getTerminator(), *TD);
      DEBUG(dbgs() << "GVN COERCED NONLOCAL MEM INTRIN:\nOffset: " << Offset
                   << "  " << *getMemIntrinValue() << '\n'
                   << *Res << '\n' << "\n\n\n");
    }
    return Res;
  }
};

} // end anonymous namespace

/// ConstructSSAForLoadSet - Given a set of loads specified by ValuesPerBlock,
/// construct SSA form, allowing us to eliminate LI.  This returns the value
/// that should be used at LI's definition site.
static Value *ConstructSSAForLoadSet(LoadInst *LI, 
                         SmallVectorImpl<AvailableValueInBlock> &ValuesPerBlock,
                                     const TargetData *TD,
                                     const DominatorTree &DT,
                                     AliasAnalysis *AA,
                                     MemoryDependenceAnalysis &MD) {
  // Check for the fully redundant, dominating load case.  In this case, we can
  // just use the dominating value directly.
  if (ValuesPerBlock.size() == 1 && 
      DT.properlyDominates(ValuesPerBlock[0].BB, LI->getParent()))
    return ValuesPerBlock[0].MaterializeAdjustedValue(LI->getType(), TD, MD);

  // Otherwise, we have to construct SSA form.
  SmallVector<PHINode*, 8> NewPHIs;
  SSAUpdater SSAUpdate(&NewPHIs);
  SSAUpdate.Initialize(LI->getType(), LI->getName());
  
  const Type *LoadTy = LI->getType();
  
  for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) {
    const AvailableValueInBlock &AV = ValuesPerBlock[i];
    BasicBlock *BB = AV.BB;
    
    if (SSAUpdate.HasValueForBlock(BB))
      continue;

    SSAUpdate.AddAvailableValue(BB, AV.MaterializeAdjustedValue(LoadTy, TD,MD));
  }
  
  // Perform PHI construction.
  Value *V = SSAUpdate.GetValueInMiddleOfBlock(LI->getParent());
  
  // If new PHI nodes were created, notify alias analysis.
  if (V->getType()->isPointerTy())
    for (unsigned i = 0, e = NewPHIs.size(); i != e; ++i)
      AA->copyValue(LI, NewPHIs[i]);
    
    // Now that we've copied information to the new PHIs, scan through
    // them again and inform alias analysis that we've added potentially
    // escaping uses to any values that are operands to these PHIs.
    for (unsigned i = 0, e = NewPHIs.size(); i != e; ++i) {
      PHINode *P = NewPHIs[i];
      for (unsigned ii = 0, ee = P->getNumIncomingValues(); ii != ee; ++ii)
        AA->addEscapingUse(P->getOperandUse(2*ii));
    }

  return V;
}

static bool isLifetimeStart(const Instruction *Inst) {
  if (const IntrinsicInst* II = dyn_cast<IntrinsicInst>(Inst))
    return II->getIntrinsicID() == Intrinsic::lifetime_start;
  return false;
}

/// 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<NonLocalDepResult, 64> Deps;
  AliasAnalysis::Location Loc = VN.getAliasAnalysis()->getLocation(LI);
  MD->getNonLocalPointerDependency(Loc, true, LI->getParent(), Deps);
  //DEBUG(dbgs() << "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].getResult().isClobber() &&
      Deps[0].getResult().getInst()->getParent() == LI->getParent()) {
    DEBUG(
      dbgs() << "GVN: non-local load ";
      WriteAsOperand(dbgs(), LI);
      dbgs() << " is clobbered by " << *Deps[0].getResult().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<AvailableValueInBlock, 16> ValuesPerBlock;
  SmallVector<BasicBlock*, 16> UnavailableBlocks;

  for (unsigned i = 0, e = Deps.size(); i != e; ++i) {
    BasicBlock *DepBB = Deps[i].getBB();
    MemDepResult DepInfo = Deps[i].getResult();

    if (DepInfo.isClobber()) {
      // The address being loaded in this non-local block may not be the same as
      // the pointer operand of the load if PHI translation occurs.  Make sure
      // to consider the right address.
      Value *Address = Deps[i].getAddress();
      
      // If the dependence is to a store that writes to a superset of the bits
      // read by the load, we can extract the bits we need for the load from the
      // stored value.
      if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInfo.getInst())) {
        if (TD && Address) {
          int Offset = AnalyzeLoadFromClobberingStore(LI->getType(), Address,
                                                      DepSI, *TD);
          if (Offset != -1) {
            ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB,
                                                       DepSI->getValueOperand(),
                                                                Offset));
            continue;
          }
        }
      }
      
      // Check to see if we have something like this:
      //    load i32* P
      //    load i8* (P+1)
      // if we have this, replace the later with an extraction from the former.
      if (LoadInst *DepLI = dyn_cast<LoadInst>(DepInfo.getInst())) {
        // If this is a clobber and L is the first instruction in its block, then
        // we have the first instruction in the entry block.
        if (DepLI != LI && Address && TD) {
          int Offset = AnalyzeLoadFromClobberingLoad(LI->getType(),
                                                     LI->getPointerOperand(),
                                                     DepLI, *TD);
          
          if (Offset != -1) {
            ValuesPerBlock.push_back(AvailableValueInBlock::getLoad(DepBB,DepLI,
                                                                    Offset));
            continue;
          }
        }
      }

      // If the clobbering value is a memset/memcpy/memmove, see if we can
      // forward a value on from it.
      if (MemIntrinsic *DepMI = dyn_cast<MemIntrinsic>(DepInfo.getInst())) {
        if (TD && Address) {
          int Offset = AnalyzeLoadFromClobberingMemInst(LI->getType(), Address,
                                                        DepMI, *TD);
          if (Offset != -1) {
            ValuesPerBlock.push_back(AvailableValueInBlock::getMI(DepBB, DepMI,
                                                                  Offset));
            continue;
          }            
        }
      }
      
      UnavailableBlocks.push_back(DepBB);
      continue;
    }

    Instruction *DepInst = DepInfo.getInst();

    // Loading the allocation -> undef.
    if (isa<AllocaInst>(DepInst) || isMalloc(DepInst) ||
        // Loading immediately after lifetime begin -> undef.
        isLifetimeStart(DepInst)) {
      ValuesPerBlock.push_back(AvailableValueInBlock::get(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->getValueOperand()->getType() != LI->getType()) {
        // If the stored value is larger or equal to the loaded value, we can
        // reuse it.
        if (TD == 0 || !CanCoerceMustAliasedValueToLoad(S->getValueOperand(),
                                                        LI->getType(), *TD)) {
          UnavailableBlocks.push_back(DepBB);
          continue;
        }
      }

      ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB,
                                                         S->getValueOperand()));
      continue;
    }
    
    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 the stored value is larger or equal to the loaded value, we can
        // reuse it.
        if (TD == 0 || !CanCoerceMustAliasedValueToLoad(LD, LI->getType(),*TD)){
          UnavailableBlocks.push_back(DepBB);
          continue;
        }          
      }
      ValuesPerBlock.push_back(AvailableValueInBlock::getLoad(DepBB, LD));
      continue;
    }
    
    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()) {
    DEBUG(dbgs() << "GVN REMOVING NONLOCAL LOAD: " << *LI << '\n');
    
    // Perform PHI construction.
    Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, TD, *DT,
                                      VN.getAliasAnalysis(), *MD);
    LI->replaceAllUsesWith(V);

    if (isa<PHINode>(V))
      V->takeName(LI);
    if (V->getType()->isPointerTy())
      MD->invalidateCachedPointerInfo(V);
    VN.erase(LI);
    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 == LoadBB) // Infinite (unreachable) loop.
      return false;
    if (Blockers.count(TmpBB))
      return false;
    
    // If any of these blocks has more than one successor (i.e. if the edge we
    // just traversed was critical), then there are other paths through this 
    // block along which the load may not be anticipated.  Hoisting the load 
    // above this block would be adding the load to execution paths along
    // which it was not previously executed.
    if (TmpBB->getTerminator()->getNumSuccessors() != 1)
      return false;
  }

  assert(TmpBB);
  LoadBB = TmpBB;

  // FIXME: It is extremely unclear what this loop is doing, other than
  // artificially restricting loadpre.
  if (isSinglePred) {
    bool isHot = false;
    for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) {
      const AvailableValueInBlock &AV = ValuesPerBlock[i];
      if (AV.isSimpleValue())
        // "Hot" Instruction is in some loop (because it dominates its dep.
        // instruction).
        if (Instruction *I = dyn_cast<Instruction>(AV.getSimpleValue()))
          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;
  }

  // Check to see how many predecessors have the loaded value fully
  // available.
  DenseMap<BasicBlock*, Value*> PredLoads;
  DenseMap<BasicBlock*, char> FullyAvailableBlocks;
  for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i)
    FullyAvailableBlocks[ValuesPerBlock[i].BB] = true;
  for (unsigned i = 0, e = UnavailableBlocks.size(); i != e; ++i)
    FullyAvailableBlocks[UnavailableBlocks[i]] = false;

  SmallVector<std::pair<TerminatorInst*, unsigned>, 4> NeedToSplit;
  for (pred_iterator PI = pred_begin(LoadBB), E = pred_end(LoadBB);
       PI != E; ++PI) {
    BasicBlock *Pred = *PI;
    if (IsValueFullyAvailableInBlock(Pred, FullyAvailableBlocks)) {
      continue;
    }
    PredLoads[Pred] = 0;

    if (Pred->getTerminator()->getNumSuccessors() != 1) {
      if (isa<IndirectBrInst>(Pred->getTerminator())) {
        DEBUG(dbgs() << "COULD NOT PRE LOAD BECAUSE OF INDBR CRITICAL EDGE '"
              << Pred->getName() << "': " << *LI << '\n');
        return false;
      }
      unsigned SuccNum = GetSuccessorNumber(Pred, LoadBB);
      NeedToSplit.push_back(std::make_pair(Pred->getTerminator(), SuccNum));
    }
  }
  if (!NeedToSplit.empty()) {
    toSplit.append(NeedToSplit.begin(), NeedToSplit.end());
    return false;
  }

  // Decide whether PRE is profitable for this load.
  unsigned NumUnavailablePreds = PredLoads.size();
  assert(NumUnavailablePreds != 0 &&
         "Fully available value should be eliminated above!");
  
  // If this load is unavailable in multiple predecessors, reject it.
  // 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.
  if (NumUnavailablePreds != 1)
      return false;

  // Check if the load can safely be moved to all the unavailable predecessors.
  bool CanDoPRE = true;
  SmallVector<Instruction*, 8> NewInsts;
  for (DenseMap<BasicBlock*, Value*>::iterator I = PredLoads.begin(),
         E = PredLoads.end(); I != E; ++I) {
    BasicBlock *UnavailablePred = I->first;

    // Do PHI translation to get its value in the predecessor if necessary.  The
    // returned pointer (if non-null) is guaranteed to dominate UnavailablePred.

    // If all preds have a single successor, then we know it is safe to insert
    // the load on the pred (?!?), so we can insert code to materialize the
    // pointer if it is not available.
    PHITransAddr Address(LI->getPointerOperand(), TD);
    Value *LoadPtr = 0;
    if (allSingleSucc) {
      LoadPtr = Address.PHITranslateWithInsertion(LoadBB, UnavailablePred,
                                                  *DT, NewInsts);
    } else {
      Address.PHITranslateValue(LoadBB, UnavailablePred, DT);
      LoadPtr = Address.getAddr();
    }

    // If we couldn't find or insert a computation of this phi translated value,
    // we fail PRE.
    if (LoadPtr == 0) {
      DEBUG(dbgs() << "COULDN'T INSERT PHI TRANSLATED VALUE OF: "
            << *LI->getPointerOperand() << "\n");
      CanDoPRE = false;
      break;
    }

    // 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 &&
        // FIXME: REEVALUTE THIS.
        !isSafeToLoadUnconditionally(LoadPtr,
                                     UnavailablePred->getTerminator(),
                                     LI->getAlignment(), TD)) {
      CanDoPRE = false;
      break;
    }

    I->second = LoadPtr;
  }

  if (!CanDoPRE) {
    while (!NewInsts.empty()) {
      Instruction *I = NewInsts.pop_back_val();
      if (MD) MD->removeInstruction(I);
      I->eraseFromParent();
    }
    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(dbgs() << "GVN REMOVING PRE LOAD: " << *LI << '\n');
  DEBUG(if (!NewInsts.empty())
          dbgs() << "INSERTED " << NewInsts.size() << " INSTS: "
                 << *NewInsts.back() << '\n');
  
  // Assign value numbers to the new instructions.
  for (unsigned i = 0, e = NewInsts.size(); i != e; ++i) {
    // FIXME: We really _ought_ to insert these value numbers into their 
    // parent's availability map.  However, in doing so, we risk getting into
    // ordering issues.  If a block hasn't been processed yet, we would be
    // marking a value as AVAIL-IN, which isn't what we intend.
    VN.lookup_or_add(NewInsts[i]);
  }

  for (DenseMap<BasicBlock*, Value*>::iterator I = PredLoads.begin(),
         E = PredLoads.end(); I != E; ++I) {
    BasicBlock *UnavailablePred = I->first;
    Value *LoadPtr = I->second;

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

    // Transfer the old load's TBAA tag to the new load.
    if (MDNode *Tag = LI->getMetadata(LLVMContext::MD_tbaa))
      NewLoad->setMetadata(LLVMContext::MD_tbaa, Tag);

    // Add the newly created load.
    ValuesPerBlock.push_back(AvailableValueInBlock::get(UnavailablePred,
                                                        NewLoad));
    MD->invalidateCachedPointerInfo(LoadPtr);
    DEBUG(dbgs() << "GVN INSERTED " << *NewLoad << '\n');
  }

  // Perform PHI construction.
  Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, TD, *DT,
                                    VN.getAliasAnalysis(), *MD);
  LI->replaceAllUsesWith(V);
  if (isa<PHINode>(V))
    V->takeName(LI);
  if (V->getType()->isPointerTy())
    MD->invalidateCachedPointerInfo(V);
  VN.erase(LI);
  toErase.push_back(LI);
  ++NumPRELoad;
  return true;
}

/// 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 (!MD)
    return false;

  if (L->isVolatile())
    return false;

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

  // If we have a clobber and target data is around, see if this is a clobber
  // that we can fix up through code synthesis.
  if (Dep.isClobber() && TD) {
    // 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.
    Value *AvailVal = 0;
    if (StoreInst *DepSI = dyn_cast<StoreInst>(Dep.getInst())) {
      int Offset = AnalyzeLoadFromClobberingStore(L->getType(),
                                                  L->getPointerOperand(),
                                                  DepSI, *TD);
      if (Offset != -1)
        AvailVal = GetStoreValueForLoad(DepSI->getValueOperand(), Offset,
                                        L->getType(), L, *TD);
    }
    
    // Check to see if we have something like this:
    //    load i32* P
    //    load i8* (P+1)
    // if we have this, replace the later with an extraction from the former.
    if (LoadInst *DepLI = dyn_cast<LoadInst>(Dep.getInst())) {
      // If this is a clobber and L is the first instruction in its block, then
      // we have the first instruction in the entry block.
      if (DepLI == L)
        return false;
      
      int Offset = AnalyzeLoadFromClobberingLoad(L->getType(),
                                                 L->getPointerOperand(),
                                                 DepLI, *TD);
      if (Offset != -1)
        AvailVal = GetLoadValueForLoad(DepLI, Offset, L->getType(), L, *TD, *MD);
    }
    
    // If the clobbering value is a memset/memcpy/memmove, see if we can forward
    // a value on from it.
    if (MemIntrinsic *DepMI = dyn_cast<MemIntrinsic>(Dep.getInst())) {
      int Offset = AnalyzeLoadFromClobberingMemInst(L->getType(),
                                                    L->getPointerOperand(),
                                                    DepMI, *TD);
      if (Offset != -1)
        AvailVal = GetMemInstValueForLoad(DepMI, Offset, L->getType(), L, *TD);
    }
        
    if (AvailVal) {
      DEBUG(dbgs() << "GVN COERCED INST:\n" << *Dep.getInst() << '\n'
            << *AvailVal << '\n' << *L << "\n\n\n");
      
      // Replace the load!
      L->replaceAllUsesWith(AvailVal);
      if (AvailVal->getType()->isPointerTy())
        MD->invalidateCachedPointerInfo(AvailVal);
      VN.erase(L);
      toErase.push_back(L);
      ++NumGVNLoad;
      return true;
    }
  }
  
  // If the value isn't available, don't do anything!
  if (Dep.isClobber()) {
    DEBUG(
      // fast print dep, using operator<< on instruction is too slow.
      dbgs() << "GVN: load ";
      WriteAsOperand(dbgs(), L);
      Instruction *I = Dep.getInst();
      dbgs() << " 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->getValueOperand();
    
    // 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).
    if (StoredVal->getType() != L->getType()) {
      if (TD) {
        StoredVal = CoerceAvailableValueToLoadType(StoredVal, L->getType(),
                                                   L, *TD);
        if (StoredVal == 0)
          return false;
        
        DEBUG(dbgs() << "GVN COERCED STORE:\n" << *DepSI << '\n' << *StoredVal
                     << '\n' << *L << "\n\n\n");
      }
      else 
        return false;
    }

    // Remove it!
    L->replaceAllUsesWith(StoredVal);
    if (StoredVal->getType()->isPointerTy())
      MD->invalidateCachedPointerInfo(StoredVal);
    VN.erase(L);
    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).
    if (DepLI->getType() != L->getType()) {
      if (TD) {
        AvailableVal = CoerceAvailableValueToLoadType(DepLI, L->getType(),
                                                      L, *TD);
        if (AvailableVal == 0)
          return false;
      
        DEBUG(dbgs() << "GVN COERCED LOAD:\n" << *DepLI << "\n" << *AvailableVal
                     << "\n" << *L << "\n\n\n");
      }
      else 
        return false;
    }
    
    // Remove it!
    L->replaceAllUsesWith(AvailableVal);
    if (DepLI->getType()->isPointerTy())
      MD->invalidateCachedPointerInfo(DepLI);
    VN.erase(L);
    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<AllocaInst>(DepInst) || isMalloc(DepInst)) {
    L->replaceAllUsesWith(UndefValue::get(L->getType()));
    VN.erase(L);
    toErase.push_back(L);
    ++NumGVNLoad;
    return true;
  }
  
  // If this load occurs either right after a lifetime begin,
  // then the loaded value is undefined.
  if (IntrinsicInst* II = dyn_cast<IntrinsicInst>(DepInst)) {
    if (II->getIntrinsicID() == Intrinsic::lifetime_start) {
      L->replaceAllUsesWith(UndefValue::get(L->getType()));
      VN.erase(L);
      toErase.push_back(L);
      ++NumGVNLoad;
      return true;
    }
  }

  return false;
}

// findLeader - In order to find a leader for a given value number at a 
// specific basic block, we first obtain the list of all Values for that number,
// and then scan the list to find one whose block dominates the block in 
// question.  This is fast because dominator tree queries consist of only
// a few comparisons of DFS numbers.
Value *GVN::findLeader(BasicBlock *BB, uint32_t num) {
  LeaderTableEntry Vals = LeaderTable[num];
  if (!Vals.Val) return 0;
  
  Value *Val = 0;
  if (DT->dominates(Vals.BB, BB)) {
    Val = Vals.Val;
    if (isa<Constant>(Val)) return Val;
  }
  
  LeaderTableEntry* Next = Vals.Next;
  while (Next) {
    if (DT->dominates(Next->BB, BB)) {
      if (isa<Constant>(Next->Val)) return Next->Val;
      if (!Val) Val = Next->Val;
    }
    
    Next = Next->Next;
  }

  return Val;
}


/// processInstruction - When calculating availability, handle an instruction
/// by inserting it into the appropriate sets
bool GVN::processInstruction(Instruction *I,
                             SmallVectorImpl<Instruction*> &toErase) {
  // Ignore dbg info intrinsics.
  if (isa<DbgInfoIntrinsic>(I))
    return false;

  // If the instruction can be easily simplified then do so now in preference
  // to value numbering it.  Value numbering often exposes redundancies, for
  // example if it determines that %y is equal to %x then the instruction
  // "%z = and i32 %x, %y" becomes "%z = and i32 %x, %x" which we now simplify.
  if (Value *V = SimplifyInstruction(I, TD, DT)) {
    I->replaceAllUsesWith(V);
    if (MD && V->getType()->isPointerTy())
      MD->invalidateCachedPointerInfo(V);
    VN.erase(I);
    toErase.push_back(I);
    return true;
  }

  if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
    bool Changed = processLoad(LI, toErase);

    if (!Changed) {
      unsigned Num = VN.lookup_or_add(LI);
      addToLeaderTable(Num, LI, LI->getParent());
    }

    return Changed;
  }

  // For conditions branches, we can perform simple conditional propagation on
  // the condition value itself.
  if (BranchInst *BI = dyn_cast<BranchInst>(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())
      addToLeaderTable(CondVN,
                   ConstantInt::getTrue(TrueSucc->getContext()),
                   TrueSucc);
    if (FalseSucc->getSinglePredecessor())
      addToLeaderTable(CondVN,
                   ConstantInt::getFalse(TrueSucc->getContext()),
                   FalseSucc);
    
    return false;
  }
  
  // Instructions with void type don't return a value, so there's
  // no point in trying to find redudancies in them.
  if (I->getType()->isVoidTy()) return false;
  
  uint32_t NextNum = VN.getNextUnusedValueNumber();
  unsigned Num = VN.lookup_or_add(I);

  // Allocations are always uniquely numbered, so we can save time and memory
  // by fast failing them.
  if (isa<AllocaInst>(I) || isa<TerminatorInst>(I) || isa<PHINode>(I)) {
    addToLeaderTable(Num, I, I->getParent());
    return false;
  }

  // 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!
  if (Num == NextNum) {
    addToLeaderTable(Num, I, I->getParent());
    return false;
  }
  
  // Perform fast-path value-number based elimination of values inherited from
  // dominators.
  Value *repl = findLeader(I->getParent(), Num);
  if (repl == 0) {
    // Failure, just remember this instance for future use.
    addToLeaderTable(Num, I, I->getParent());
    return false;
  }
  
  // Remove it!
  VN.erase(I);
  I->replaceAllUsesWith(repl);
  if (MD && repl->getType()->isPointerTy())
    MD->invalidateCachedPointerInfo(repl);
  toErase.push_back(I);
  return true;
}

/// runOnFunction - This is the main transformation entry point for a function.
bool GVN::runOnFunction(Function& F) {
  if (!NoLoads)
    MD = &getAnalysis<MemoryDependenceAnalysis>();
  DT = &getAnalysis<DominatorTree>();
  TD = getAnalysisIfAvailable<TargetData>();
  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++;
    
    bool removedBlock = MergeBlockIntoPredecessor(BB, this);
    if (removedBlock) ++NumGVNBlocks;

    Changed |= removedBlock;
  }

  unsigned Iteration = 0;
  while (ShouldContinue) {
    DEBUG(dbgs() << "GVN iteration: " << Iteration << "\n");
    ShouldContinue = iterateOnFunction(F);
    if (splitCriticalEdges())
      ShouldContinue = true;
    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(dbgs() << "GVN removed: " << **I << '\n');
      if (MD) 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;
  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<AllocaInst>(CurInst) ||
          isa<TerminatorInst>(CurInst) || isa<PHINode>(CurInst) ||
          CurInst->getType()->isVoidTy() ||
          CurInst->mayReadFromMemory() || CurInst->mayHaveSideEffects() ||