Fix batch of converting RegisterPass<> to INTIALIZE_PASS().
[oota-llvm.git] / lib / Transforms / Scalar / DeadStoreElimination.cpp
index bcf674ed37ff0ee02794c4c7b4decc9dd0385c8d..dddefeda083a58740f7ae3d27646537136343d7f 100644 (file)
@@ -2,8 +2,8 @@
 //
 //                     The LLVM Compiler Infrastructure
 //
-// This file was developed by Owen Anderson and is distributed under
-// the University of Illinois Open Source License. See LICENSE.TXT for details.
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
 //
 //===----------------------------------------------------------------------===//
 //
 #include "llvm/Constants.h"
 #include "llvm/Function.h"
 #include "llvm/Instructions.h"
+#include "llvm/IntrinsicInst.h"
 #include "llvm/Pass.h"
-#include "llvm/ADT/SetVector.h"
 #include "llvm/ADT/SmallPtrSet.h"
 #include "llvm/ADT/Statistic.h"
 #include "llvm/Analysis/AliasAnalysis.h"
+#include "llvm/Analysis/Dominators.h"
+#include "llvm/Analysis/MemoryBuiltins.h"
 #include "llvm/Analysis/MemoryDependenceAnalysis.h"
 #include "llvm/Target/TargetData.h"
 #include "llvm/Transforms/Utils/Local.h"
-#include "llvm/Support/Compiler.h"
 using namespace llvm;
 
 STATISTIC(NumFastStores, "Number of stores deleted");
 STATISTIC(NumFastOther , "Number of other instrs removed");
 
 namespace {
-  struct VISIBILITY_HIDDEN DSE : public FunctionPass {
+  struct DSE : public FunctionPass {
+    TargetData *TD;
+
     static char ID; // Pass identification, replacement for typeid
-    DSE() : FunctionPass((intptr_t)&ID) {}
+    DSE() : FunctionPass(&ID) {}
 
     virtual bool runOnFunction(Function &F) {
       bool Changed = false;
+      
+      DominatorTree &DT = getAnalysis<DominatorTree>();
+      
       for (Function::iterator I = F.begin(), E = F.end(); I != E; ++I)
-        Changed |= runOnBasicBlock(*I);
+        // Only check non-dead blocks.  Dead blocks may have strange pointer
+        // cycles that will confuse alias analysis.
+        if (DT.isReachableFromEntry(I))
+          Changed |= runOnBasicBlock(*I);
       return Changed;
     }
-
+    
     bool runOnBasicBlock(BasicBlock &BB);
-    bool handleFreeWithNonTrivialDependency(FreeInst* F,
-                                            Instruction* dependency,
-                                        SetVector<Instruction*>& possiblyDead);
-    bool handleEndBlock(BasicBlock& BB, SetVector<Instruction*>& possiblyDead);
-    bool RemoveUndeadPointers(Value* pointer,
-                              BasicBlock::iterator& BBI,
-                              SmallPtrSet<AllocaInst*, 64>& deadPointers, 
-                              SetVector<Instruction*>& possiblyDead);
-    void DeleteDeadInstructionChains(Instruction *I,
-                                     SetVector<Instruction*> &DeadInsts);
+    bool handleFreeWithNonTrivialDependency(const CallInst *F,
+                                            MemDepResult Dep);
+    bool handleEndBlock(BasicBlock &BB);
+    bool RemoveUndeadPointers(Value *Ptr, uint64_t killPointerSize,
+                              BasicBlock::iterator &BBI,
+                              SmallPtrSet<Value*, 64> &deadPointers);
+    void DeleteDeadInstruction(Instruction *I,
+                               SmallPtrSet<Value*, 64> *deadPointers = 0);
     
-    /// Find the base pointer that a pointer came from
-    /// Because this is used to find pointers that originate
-    /// from allocas, it is safe to ignore GEP indices, since
-    /// either the store will be in the alloca, and thus dead,
-    /// or beyond the end of the alloca, and thus undefined.
-    void TranslatePointerBitCasts(Value*& v) {
-      assert(isa<PointerType>(v->getType()) &&
-             "Translating a non-pointer type?");
-      while (true) {
-        if (BitCastInst* C = dyn_cast<BitCastInst>(v))
-          v = C->getOperand(0);
-        else if (GetElementPtrInst* G = dyn_cast<GetElementPtrInst>(v))
-          v = G->getOperand(0);
-        else
-          break;
-      }
-    }
 
     // getAnalysisUsage - We require post dominance frontiers (aka Control
     // Dependence Graph)
     virtual void getAnalysisUsage(AnalysisUsage &AU) const {
       AU.setPreservesCFG();
-      AU.addRequired<TargetData>();
+      AU.addRequired<DominatorTree>();
       AU.addRequired<AliasAnalysis>();
       AU.addRequired<MemoryDependenceAnalysis>();
-      AU.addPreserved<AliasAnalysis>();
+      AU.addPreserved<DominatorTree>();
       AU.addPreserved<MemoryDependenceAnalysis>();
     }
+
+    unsigned getPointerSize(Value *V) const;
   };
-  char DSE::ID = 0;
-  RegisterPass<DSE> X("dse", "Dead Store Elimination");
 }
 
+char DSE::ID = 0;
+INITIALIZE_PASS(DSE, "dse", "Dead Store Elimination", false, false);
+
 FunctionPass *llvm::createDeadStoreEliminationPass() { return new DSE(); }
 
-bool DSE::runOnBasicBlock(BasicBlock &BB) {
-  MemoryDependenceAnalysis& MD = getAnalysis<MemoryDependenceAnalysis>();
+/// doesClobberMemory - Does this instruction clobber (write without reading)
+/// some memory?
+static bool doesClobberMemory(Instruction *I) {
+  if (isa<StoreInst>(I))
+    return true;
+  if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
+    switch (II->getIntrinsicID()) {
+    default:
+      return false;
+    case Intrinsic::memset:
+    case Intrinsic::memmove:
+    case Intrinsic::memcpy:
+    case Intrinsic::init_trampoline:
+    case Intrinsic::lifetime_end:
+      return true;
+    }
+  }
+  return false;
+}
+
+/// isElidable - If the value of this instruction and the memory it writes to is
+/// unused, may we delete this instrtction?
+static bool isElidable(Instruction *I) {
+  assert(doesClobberMemory(I));
+  if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I))
+    return II->getIntrinsicID() != Intrinsic::lifetime_end;
+  if (StoreInst *SI = dyn_cast<StoreInst>(I))
+    return !SI->isVolatile();
+  return true;
+}
+
+/// getPointerOperand - Return the pointer that is being clobbered.
+static Value *getPointerOperand(Instruction *I) {
+  assert(doesClobberMemory(I));
+  if (StoreInst *SI = dyn_cast<StoreInst>(I))
+    return SI->getPointerOperand();
+  if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(I))
+    return MI->getArgOperand(0);
+
+  IntrinsicInst *II = cast<IntrinsicInst>(I);
+  switch (II->getIntrinsicID()) {
+  default: assert(false && "Unexpected intrinsic!");
+  case Intrinsic::init_trampoline:
+    return II->getArgOperand(0);
+  case Intrinsic::lifetime_end:
+    return II->getArgOperand(1);
+  }
+}
+
+/// getStoreSize - Return the length in bytes of the write by the clobbering
+/// instruction. If variable or unknown, returns -1.
+static unsigned getStoreSize(Instruction *I, const TargetData *TD) {
+  assert(doesClobberMemory(I));
+  if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
+    if (!TD) return -1u;
+    return TD->getTypeStoreSize(SI->getOperand(0)->getType());
+  }
+
+  Value *Len;
+  if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(I)) {
+    Len = MI->getLength();
+  } else {
+    IntrinsicInst *II = cast<IntrinsicInst>(I);
+    switch (II->getIntrinsicID()) {
+    default: assert(false && "Unexpected intrinsic!");
+    case Intrinsic::init_trampoline:
+      return -1u;
+    case Intrinsic::lifetime_end:
+      Len = II->getArgOperand(0);
+      break;
+    }
+  }
+  if (ConstantInt *LenCI = dyn_cast<ConstantInt>(Len))
+    if (!LenCI->isAllOnesValue())
+      return LenCI->getZExtValue();
+  return -1u;
+}
+
+/// isStoreAtLeastAsWideAs - Return true if the size of the store in I1 is
+/// greater than or equal to the store in I2.  This returns false if we don't
+/// know.
+///
+static bool isStoreAtLeastAsWideAs(Instruction *I1, Instruction *I2,
+                                   const TargetData *TD) {
+  const Type *I1Ty = getPointerOperand(I1)->getType();
+  const Type *I2Ty = getPointerOperand(I2)->getType();
   
-  // Record the last-seen store to this pointer
-  DenseMap<Value*, StoreInst*> lastStore;
-  // Record instructions possibly made dead by deleting a store
-  SetVector<Instruction*> possiblyDead;
+  // Exactly the same type, must have exactly the same size.
+  if (I1Ty == I2Ty) return true;
   
+  int I1Size = getStoreSize(I1, TD);
+  int I2Size = getStoreSize(I2, TD);
+  
+  return I1Size != -1 && I2Size != -1 && I1Size >= I2Size;
+}
+
+bool DSE::runOnBasicBlock(BasicBlock &BB) {
+  MemoryDependenceAnalysis &MD = getAnalysis<MemoryDependenceAnalysis>();
+  TD = getAnalysisIfAvailable<TargetData>();
+
   bool MadeChange = false;
   
-  // Do a top-down walk on the BB
-  for (BasicBlock::iterator BBI = BB.begin(), BBE = BB.end();
-       BBI != BBE; ++BBI) {
-    // If we find a store or a free...
-    if (!isa<StoreInst>(BBI) && !isa<FreeInst>(BBI))
+  // Do a top-down walk on the BB.
+  for (BasicBlock::iterator BBI = BB.begin(), BBE = BB.end(); BBI != BBE; ) {
+    Instruction *Inst = BBI++;
+    
+    // If we find a store or a free, get its memory dependence.
+    if (!doesClobberMemory(Inst) && !isFreeCall(Inst))
       continue;
-      
-    Value* pointer = 0;
-    if (StoreInst* S = dyn_cast<StoreInst>(BBI))
-      pointer = S->getPointerOperand();
-    else
-      pointer = cast<FreeInst>(BBI)->getPointerOperand();
-      
-    StoreInst*& last = lastStore[pointer];
-    bool deletedStore = false;
-      
-    // ... to a pointer that has been stored to before...
-    if (last) {
-      Instruction* dep = MD.getDependency(BBI);
-        
-      // ... and no other memory dependencies are between them....
-      while (dep != MemoryDependenceAnalysis::None &&
-             dep != MemoryDependenceAnalysis::NonLocal &&
-             isa<StoreInst>(dep)) {
-        if (dep != last) {
-          dep = MD.getDependency(BBI, dep);
-          continue;
-        }
-        
-        // Remove it!
-        MD.removeInstruction(last);
-          
-        // DCE instructions only used to calculate that store
-        if (Instruction* D = dyn_cast<Instruction>(last->getOperand(0)))
-          possiblyDead.insert(D);
-        if (Instruction* D = dyn_cast<Instruction>(last->getOperand(1)))
-          possiblyDead.insert(D);
-          
-        last->eraseFromParent();
-        NumFastStores++;
-        deletedStore = true;
+    
+    MemDepResult InstDep = MD.getDependency(Inst);
+    
+    // Ignore non-local stores.
+    // FIXME: cross-block DSE would be fun. :)
+    if (InstDep.isNonLocal()) continue;
+  
+    // Handle frees whose dependencies are non-trivial.
+    if (const CallInst *F = isFreeCall(Inst)) {
+      MadeChange |= handleFreeWithNonTrivialDependency(F, InstDep);
+      continue;
+    }
+    
+    // If not a definite must-alias dependency, ignore it.
+    if (!InstDep.isDef())
+      continue;
+    
+    // If this is a store-store dependence, then the previous store is dead so
+    // long as this store is at least as big as it.
+    if (doesClobberMemory(InstDep.getInst())) {
+      Instruction *DepStore = InstDep.getInst();
+      if (isStoreAtLeastAsWideAs(Inst, DepStore, TD) &&
+          isElidable(DepStore)) {
+        // Delete the store and now-dead instructions that feed it.
+        DeleteDeadInstruction(DepStore);
+        ++NumFastStores;
         MadeChange = true;
+
+        // DeleteDeadInstruction can delete the current instruction in loop
+        // cases, reset BBI.
+        BBI = Inst;
+        if (BBI != BB.begin())
+          --BBI;
+        continue;
+      }
+    }
+    
+    if (!isElidable(Inst))
+      continue;
+    
+    // If we're storing the same value back to a pointer that we just
+    // loaded from, then the store can be removed.
+    if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
+      if (LoadInst *DepLoad = dyn_cast<LoadInst>(InstDep.getInst())) {
+        if (SI->getPointerOperand() == DepLoad->getPointerOperand() &&
+            SI->getOperand(0) == DepLoad) {
+          // DeleteDeadInstruction can delete the current instruction.  Save BBI
+          // in case we need it.
+          WeakVH NextInst(BBI);
+          
+          DeleteDeadInstruction(SI);
           
-        break;
+          if (NextInst == 0)  // Next instruction deleted.
+            BBI = BB.begin();
+          else if (BBI != BB.begin())  // Revisit this instruction if possible.
+            --BBI;
+          ++NumFastStores;
+          MadeChange = true;
+          continue;
+        }
       }
     }
     
-    // Handle frees whose dependencies are non-trivial.
-    if (FreeInst* F = dyn_cast<FreeInst>(BBI)) {
-      if (!deletedStore)
-        MadeChange |= handleFreeWithNonTrivialDependency(F,
-                                                         MD.getDependency(F),
-                                                         possiblyDead);
-      // No known stores after the free
-      last = 0;
-    } else {
-      // Update our most-recent-store map.
-      last = cast<StoreInst>(BBI);
+    // If this is a lifetime end marker, we can throw away the store.
+    if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(InstDep.getInst())) {
+      if (II->getIntrinsicID() == Intrinsic::lifetime_end) {
+        // Delete the store and now-dead instructions that feed it.
+        // DeleteDeadInstruction can delete the current instruction.  Save BBI
+        // in case we need it.
+        WeakVH NextInst(BBI);
+        
+        DeleteDeadInstruction(Inst);
+        
+        if (NextInst == 0)  // Next instruction deleted.
+          BBI = BB.begin();
+        else if (BBI != BB.begin())  // Revisit this instruction if possible.
+          --BBI;
+        ++NumFastStores;
+        MadeChange = true;
+        continue;
+      }
     }
   }
   
-  // If this block ends in a return, unwind, unreachable, and eventually
-  // tailcall, then all allocas are dead at its end.
+  // If this block ends in a return, unwind, or unreachable, all allocas are
+  // dead at its end, which means stores to them are also dead.
   if (BB.getTerminator()->getNumSuccessors() == 0)
-    MadeChange |= handleEndBlock(BB, possiblyDead);
-  
-  // Do a trivial DCE
-  while (!possiblyDead.empty()) {
-    Instruction *I = possiblyDead.back();
-    possiblyDead.pop_back();
-    DeleteDeadInstructionChains(I, possiblyDead);
-  }
+    MadeChange |= handleEndBlock(BB);
   
   return MadeChange;
 }
 
 /// handleFreeWithNonTrivialDependency - Handle frees of entire structures whose
-/// dependency is a store to a field of that structure
-bool DSE::handleFreeWithNonTrivialDependency(FreeInst* F, Instruction* dep,
-                                       SetVector<Instruction*>& possiblyDead) {
-  TargetData &TD = getAnalysis<TargetData>();
+/// dependency is a store to a field of that structure.
+bool DSE::handleFreeWithNonTrivialDependency(const CallInst *F,
+                                             MemDepResult Dep) {
   AliasAnalysis &AA = getAnalysis<AliasAnalysis>();
-  MemoryDependenceAnalysis& MD = getAnalysis<MemoryDependenceAnalysis>();
   
-  if (dep == MemoryDependenceAnalysis::None ||
-      dep == MemoryDependenceAnalysis::NonLocal)
+  Instruction *Dependency = Dep.getInst();
+  if (!Dependency || !doesClobberMemory(Dependency) || !isElidable(Dependency))
     return false;
   
-  StoreInst* dependency = dyn_cast<StoreInst>(dep);
-  if (!dependency)
+  Value *DepPointer = getPointerOperand(Dependency)->getUnderlyingObject();
+
+  // Check for aliasing.
+  if (AA.alias(F->getArgOperand(0), 1, DepPointer, 1) !=
+         AliasAnalysis::MustAlias)
     return false;
   
-  Value* depPointer = dependency->getPointerOperand();
-  const Type* depType = dependency->getOperand(0)->getType();
-  unsigned depPointerSize = TD.getTypeSize(depType);
-  
-  // Check for aliasing
-  AliasAnalysis::AliasResult A = AA.alias(F->getPointerOperand(), ~0UL,
-                                          depPointer, depPointerSize);
-    
-  if (A == AliasAnalysis::MustAlias) {
-    // Remove it!
-    MD.removeInstruction(dependency);
-
-    // DCE instructions only used to calculate that store
-    if (Instruction* D = dyn_cast<Instruction>(dependency->getOperand(0)))
-      possiblyDead.insert(D);
-    if (Instruction* D = dyn_cast<Instruction>(dependency->getOperand(1)))
-      possiblyDead.insert(D);
-
-    dependency->eraseFromParent();
-    NumFastStores++;
-    return true;
-  }
-  
-  return false;
+  // DCE instructions only used to calculate that store
+  DeleteDeadInstruction(Dependency);
+  ++NumFastStores;
+  return true;
 }
 
 /// handleEndBlock - Remove dead stores to stack-allocated locations in the
@@ -227,190 +316,261 @@ bool DSE::handleFreeWithNonTrivialDependency(FreeInst* F, Instruction* dep,
 /// ...
 /// store i32 1, i32* %A
 /// ret void
-bool DSE::handleEndBlock(BasicBlock& BB,
-                         SetVector<Instruction*>& possiblyDead) {
-  TargetData &TD = getAnalysis<TargetData>();
+bool DSE::handleEndBlock(BasicBlock &BB) {
   AliasAnalysis &AA = getAnalysis<AliasAnalysis>();
-  MemoryDependenceAnalysis& MD = getAnalysis<MemoryDependenceAnalysis>();
   
   bool MadeChange = false;
   
   // Pointers alloca'd in this function are dead in the end block
-  SmallPtrSet<AllocaInst*, 64> deadPointers;
+  SmallPtrSet<Value*, 64> deadPointers;
   
-  // Find all of the alloca'd pointers in the entry block
+  // Find all of the alloca'd pointers in the entry block.
   BasicBlock *Entry = BB.getParent()->begin();
   for (BasicBlock::iterator I = Entry->begin(), E = Entry->end(); I != E; ++I)
     if (AllocaInst *AI = dyn_cast<AllocaInst>(I))
       deadPointers.insert(AI);
   
+  // Treat byval arguments the same, stores to them are dead at the end of the
+  // function.
+  for (Function::arg_iterator AI = BB.getParent()->arg_begin(),
+       AE = BB.getParent()->arg_end(); AI != AE; ++AI)
+    if (AI->hasByValAttr())
+      deadPointers.insert(AI);
+  
   // Scan the basic block backwards
   for (BasicBlock::iterator BBI = BB.end(); BBI != BB.begin(); ){
     --BBI;
     
-    if (deadPointers.empty())
-      break;
-    
-    // If we find a store whose pointer is dead...
-    if (StoreInst* S = dyn_cast<StoreInst>(BBI)) {
-      Value* pointerOperand = S->getPointerOperand();
-      // See through pointer-to-pointer bitcasts
-      TranslatePointerBitCasts(pointerOperand);
-      
-      if (deadPointers.count(pointerOperand)){
-        // Remove it!
-        MD.removeInstruction(S);
-        
-        // DCE instructions only used to calculate that store
-        if (Instruction* D = dyn_cast<Instruction>(S->getOperand(0)))
-          possiblyDead.insert(D);
-        if (Instruction* D = dyn_cast<Instruction>(S->getOperand(1)))
-          possiblyDead.insert(D);
-        
-        BBI++;
-        S->eraseFromParent();
-        NumFastStores++;
-        MadeChange = true;
+    // If we find a store whose pointer is dead.
+    if (doesClobberMemory(BBI)) {
+      if (isElidable(BBI)) {
+        // See through pointer-to-pointer bitcasts
+        Value *pointerOperand = getPointerOperand(BBI)->getUnderlyingObject();
+
+        // Alloca'd pointers or byval arguments (which are functionally like
+        // alloca's) are valid candidates for removal.
+        if (deadPointers.count(pointerOperand)) {
+          // DCE instructions only used to calculate that store.
+          Instruction *Dead = BBI;
+          ++BBI;
+          DeleteDeadInstruction(Dead, &deadPointers);
+          ++NumFastStores;
+          MadeChange = true;
+          continue;
+        }
       }
       
-      continue;
+      // Because a memcpy or memmove is also a load, we can't skip it if we
+      // didn't remove it.
+      if (!isa<MemTransferInst>(BBI))
+        continue;
     }
     
-    Value* killPointer = 0;
+    Value *killPointer = 0;
+    uint64_t killPointerSize = ~0UL;
     
     // If we encounter a use of the pointer, it is no longer considered dead
-    if (LoadInst* L = dyn_cast<LoadInst>(BBI)) {
+    if (LoadInst *L = dyn_cast<LoadInst>(BBI)) {
+      // However, if this load is unused and not volatile, we can go ahead and
+      // remove it, and not have to worry about it making our pointer undead!
+      if (L->use_empty() && !L->isVolatile()) {
+        ++BBI;
+        DeleteDeadInstruction(L, &deadPointers);
+        ++NumFastOther;
+        MadeChange = true;
+        continue;
+      }
+      
       killPointer = L->getPointerOperand();
-    } else if (VAArgInstV = dyn_cast<VAArgInst>(BBI)) {
+    } else if (VAArgInst *V = dyn_cast<VAArgInst>(BBI)) {
       killPointer = V->getOperand(0);
-    } else if (AllocaInst* A = dyn_cast<AllocaInst>(BBI)) {
+    } else if (isa<MemTransferInst>(BBI) &&
+               isa<ConstantInt>(cast<MemTransferInst>(BBI)->getLength())) {
+      killPointer = cast<MemTransferInst>(BBI)->getSource();
+      killPointerSize = cast<ConstantInt>(
+                       cast<MemTransferInst>(BBI)->getLength())->getZExtValue();
+    } else if (AllocaInst *A = dyn_cast<AllocaInst>(BBI)) {
       deadPointers.erase(A);
+      
+      // Dead alloca's can be DCE'd when we reach them
+      if (A->use_empty()) {
+        ++BBI;
+        DeleteDeadInstruction(A, &deadPointers);
+        ++NumFastOther;
+        MadeChange = true;
+      }
+      
       continue;
     } else if (CallSite::get(BBI).getInstruction() != 0) {
       // If this call does not access memory, it can't
       // be undeadifying any of our pointers.
       CallSite CS = CallSite::get(BBI);
-      if (CS.getCalledFunction() &&
-          AA.doesNotAccessMemory(CS.getCalledFunction()))
+      if (AA.doesNotAccessMemory(CS))
         continue;
       
+      unsigned modRef = 0;
+      unsigned other = 0;
+      
       // Remove any pointers made undead by the call from the dead set
-      std::vector<Instruction*> dead;
-      for (SmallPtrSet<AllocaInst*, 64>::iterator I = deadPointers.begin(),
+      std::vector<Value*> dead;
+      for (SmallPtrSet<Value*, 64>::iterator I = deadPointers.begin(),
            E = deadPointers.end(); I != E; ++I) {
-        // Get size information for the alloca
-        unsigned pointerSize = ~0UL;
-        if (ConstantInt* C = dyn_cast<ConstantInt>((*I)->getArraySize()))
-          pointerSize = C->getZExtValue() * \
-                        TD.getTypeSize((*I)->getAllocatedType());     
+        // HACK: if we detect that our AA is imprecise, it's not
+        // worth it to scan the rest of the deadPointers set.  Just
+        // assume that the AA will return ModRef for everything, and
+        // go ahead and bail.
+        if (modRef >= 16 && other == 0) {
+          deadPointers.clear();
+          return MadeChange;
+        }
         
         // See if the call site touches it
-        AliasAnalysis::ModRefResult A = AA.getModRefInfo(CS, *I, pointerSize);
+        AliasAnalysis::ModRefResult A = AA.getModRefInfo(CS, *I,
+                                                         getPointerSize(*I));
+        
+        if (A == AliasAnalysis::ModRef)
+          ++modRef;
+        else
+          ++other;
+        
         if (A == AliasAnalysis::ModRef || A == AliasAnalysis::Ref)
           dead.push_back(*I);
       }
 
-      for (std::vector<Instruction*>::iterator I = dead.begin(), E = dead.end();
+      for (std::vector<Value*>::iterator I = dead.begin(), E = dead.end();
            I != E; ++I)
         deadPointers.erase(*I);
       
       continue;
+    } else if (isInstructionTriviallyDead(BBI)) {
+      // For any non-memory-affecting non-terminators, DCE them as we reach them
+      Instruction *Inst = BBI;
+      ++BBI;
+      DeleteDeadInstruction(Inst, &deadPointers);
+      ++NumFastOther;
+      MadeChange = true;
+      continue;
     }
     
     if (!killPointer)
       continue;
-    
+
+    killPointer = killPointer->getUnderlyingObject();
+
     // Deal with undead pointers
-    MadeChange |= RemoveUndeadPointers(killPointer, BBI,
-                                       deadPointers, possiblyDead);
+    MadeChange |= RemoveUndeadPointers(killPointer, killPointerSize, BBI,
+                                       deadPointers);
   }
   
   return MadeChange;
 }
 
-/// RemoveUndeadPointers - takes an instruction and a setvector of
-/// dead instructions.  If I is dead, it is erased, and its operands are
-/// checked for deadness.  If they are dead, they are added to the dead
-/// setvector.
-bool DSE::RemoveUndeadPointers(Value* killPointer,
-                                BasicBlock::iterator& BBI,
-                                SmallPtrSet<AllocaInst*, 64>& deadPointers, 
-                                SetVector<Instruction*>& possiblyDead) {
-  TargetData &TD = getAnalysis<TargetData>();
+/// RemoveUndeadPointers - check for uses of a pointer that make it
+/// undead when scanning for dead stores to alloca's.
+bool DSE::RemoveUndeadPointers(Value *killPointer, uint64_t killPointerSize,
+                               BasicBlock::iterator &BBI,
+                               SmallPtrSet<Value*, 64> &deadPointers) {
   AliasAnalysis &AA = getAnalysis<AliasAnalysis>();
-  MemoryDependenceAnalysis& MD = getAnalysis<MemoryDependenceAnalysis>();
-                                  
+
+  // If the kill pointer can be easily reduced to an alloca,
+  // don't bother doing extraneous AA queries.
+  if (deadPointers.count(killPointer)) {
+    deadPointers.erase(killPointer);
+    return false;
+  }
+  
+  // A global can't be in the dead pointer set.
+  if (isa<GlobalValue>(killPointer))
+    return false;
+  
   bool MadeChange = false;
   
-  std::vector<Instruction*> undead;
-    
-  for (SmallPtrSet<AllocaInst*, 64>::iterator I = deadPointers.begin(),
-      E = deadPointers.end(); I != E; ++I) {
-    // Get size information for the alloca
-    unsigned pointerSize = ~0UL;
-    if (ConstantInt* C = dyn_cast<ConstantInt>((*I)->getArraySize()))
-      pointerSize = C->getZExtValue() * \
-                    TD.getTypeSize((*I)->getAllocatedType());     
-      
+  SmallVector<Value*, 16> undead;
+  
+  for (SmallPtrSet<Value*, 64>::iterator I = deadPointers.begin(),
+       E = deadPointers.end(); I != E; ++I) {
     // See if this pointer could alias it
-    AliasAnalysis::AliasResult A = AA.alias(*I, pointerSize,
-                                            killPointer, ~0UL);
+    AliasAnalysis::AliasResult A = AA.alias(*I, getPointerSize(*I),
+                                            killPointer, killPointerSize);
 
     // If it must-alias and a store, we can delete it
     if (isa<StoreInst>(BBI) && A == AliasAnalysis::MustAlias) {
-      StoreInstS = cast<StoreInst>(BBI);
+      StoreInst *S = cast<StoreInst>(BBI);
 
       // Remove it!
-      MD.removeInstruction(S);
-
-      // DCE instructions only used to calculate that store
-      if (Instruction* D = dyn_cast<Instruction>(S->getOperand(0)))
-        possiblyDead.insert(D);
-      if (Instruction* D = dyn_cast<Instruction>(S->getOperand(1)))
-        possiblyDead.insert(D);
-
-      BBI++;
-      S->eraseFromParent();
-      NumFastStores++;
+      ++BBI;
+      DeleteDeadInstruction(S, &deadPointers);
+      ++NumFastStores;
       MadeChange = true;
 
       continue;
 
       // Otherwise, it is undead
-      } else if (A != AliasAnalysis::NoAlias)
-        undead.push_back(*I);
+    } else if (A != AliasAnalysis::NoAlias)
+      undead.push_back(*I);
   }
 
-  for (std::vector<Instruction*>::iterator I = undead.begin(), E = undead.end();
+  for (SmallVector<Value*, 16>::iterator I = undead.begin(), E = undead.end();
        I != E; ++I)
-    deadPointers.erase(*I);
+      deadPointers.erase(*I);
   
   return MadeChange;
 }
 
-/// DeleteDeadInstructionChains - takes an instruction and a setvector of
-/// dead instructions.  If I is dead, it is erased, and its operands are
-/// checked for deadness.  If they are dead, they are added to the dead
-/// setvector.
-void DSE::DeleteDeadInstructionChains(Instruction *I,
-                                      SetVector<Instruction*> &DeadInsts) {
-  // Instruction must be dead.
-  if (!I->use_empty() || !isInstructionTriviallyDead(I)) return;
-
-  // Let the memory dependence know
-  getAnalysis<MemoryDependenceAnalysis>().removeInstruction(I);
-
-  // See if this made any operands dead.  We do it this way in case the
-  // instruction uses the same operand twice.  We don't want to delete a
-  // value then reference it.
-  for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
-    if (I->getOperand(i)->hasOneUse())
-      if (Instruction* Op = dyn_cast<Instruction>(I->getOperand(i)))
-        DeadInsts.insert(Op);      // Attempt to nuke it later.
+/// DeleteDeadInstruction - Delete this instruction.  Before we do, go through
+/// and zero out all the operands of this instruction.  If any of them become
+/// dead, delete them and the computation tree that feeds them.
+///
+/// If ValueSet is non-null, remove any deleted instructions from it as well.
+///
+void DSE::DeleteDeadInstruction(Instruction *I,
+                                SmallPtrSet<Value*, 64> *ValueSet) {
+  SmallVector<Instruction*, 32> NowDeadInsts;
+  
+  NowDeadInsts.push_back(I);
+  --NumFastOther;
+
+  // Before we touch this instruction, remove it from memdep!
+  MemoryDependenceAnalysis &MDA = getAnalysis<MemoryDependenceAnalysis>();
+  do {
+    Instruction *DeadInst = NowDeadInsts.pop_back_val();
     
-    I->setOperand(i, 0);         // Drop from the operand list.
-  }
+    ++NumFastOther;
+    
+    // This instruction is dead, zap it, in stages.  Start by removing it from
+    // MemDep, which needs to know the operands and needs it to be in the
+    // function.
+    MDA.removeInstruction(DeadInst);
+    
+    for (unsigned op = 0, e = DeadInst->getNumOperands(); op != e; ++op) {
+      Value *Op = DeadInst->getOperand(op);
+      DeadInst->setOperand(op, 0);
+      
+      // If this operand just became dead, add it to the NowDeadInsts list.
+      if (!Op->use_empty()) continue;
+      
+      if (Instruction *OpI = dyn_cast<Instruction>(Op))
+        if (isInstructionTriviallyDead(OpI))
+          NowDeadInsts.push_back(OpI);
+    }
+    
+    DeadInst->eraseFromParent();
+    
+    if (ValueSet) ValueSet->erase(DeadInst);
+  } while (!NowDeadInsts.empty());
+}
 
-  I->eraseFromParent();
-  ++NumFastOther;
+unsigned DSE::getPointerSize(Value *V) const {
+  if (TD) {
+    if (AllocaInst *A = dyn_cast<AllocaInst>(V)) {
+      // Get size information for the alloca
+      if (ConstantInt *C = dyn_cast<ConstantInt>(A->getArraySize()))
+        return C->getZExtValue() * TD->getTypeAllocSize(A->getAllocatedType());
+    } else {
+      assert(isa<Argument>(V) && "Expected AllocaInst or Argument!");
+      const PointerType *PT = cast<PointerType>(V->getType());
+      return TD->getTypeAllocSize(PT->getElementType());
+    }
+  }
+  return ~0U;
 }