Fix a bug that David Greene found in the DAGCombiner's logic
[oota-llvm.git] / lib / CodeGen / PreAllocSplitting.cpp
index f4540889f6aa04758a84676911940d8b85449164..9e955bd69690d7b169bc56dda7af64b26d833977 100644 (file)
@@ -50,6 +50,7 @@ namespace {
     MachineFunction       *CurrMF;
     const TargetMachine   *TM;
     const TargetInstrInfo *TII;
+    const TargetRegisterInfo* TRI;
     MachineFrameInfo      *MFI;
     MachineRegisterInfo   *MRI;
     LiveIntervals         *LIs;
@@ -163,16 +164,24 @@ namespace {
     bool removeDeadSpills(SmallPtrSet<LiveInterval*, 8>& split);
     unsigned getNumberOfNonSpills(SmallPtrSet<MachineInstr*, 4>& MIs,
                                unsigned Reg, int FrameIndex, bool& TwoAddr);
-    VNInfo* PerformPHIConstruction(MachineBasicBlock::iterator use,
-                                   MachineBasicBlock* MBB,
-                                   LiveInterval* LI,
+    VNInfo* PerformPHIConstruction(MachineBasicBlock::iterator Use,
+                                   MachineBasicBlock* MBB, LiveInterval* LI,
                                    SmallPtrSet<MachineInstr*, 4>& Visited,
             DenseMap<MachineBasicBlock*, SmallPtrSet<MachineInstr*, 2> >& Defs,
             DenseMap<MachineBasicBlock*, SmallPtrSet<MachineInstr*, 2> >& Uses,
                                       DenseMap<MachineInstr*, VNInfo*>& NewVNs,
                                 DenseMap<MachineBasicBlock*, VNInfo*>& LiveOut,
                                 DenseMap<MachineBasicBlock*, VNInfo*>& Phis,
-                                        bool toplevel, bool intrablock);
+                                        bool IsTopLevel, bool IsIntraBlock);
+    VNInfo* PerformPHIConstructionFallBack(MachineBasicBlock::iterator Use,
+                                   MachineBasicBlock* MBB, LiveInterval* LI,
+                                   SmallPtrSet<MachineInstr*, 4>& Visited,
+            DenseMap<MachineBasicBlock*, SmallPtrSet<MachineInstr*, 2> >& Defs,
+            DenseMap<MachineBasicBlock*, SmallPtrSet<MachineInstr*, 2> >& Uses,
+                                      DenseMap<MachineInstr*, VNInfo*>& NewVNs,
+                                DenseMap<MachineBasicBlock*, VNInfo*>& LiveOut,
+                                DenseMap<MachineBasicBlock*, VNInfo*>& Phis,
+                                        bool IsTopLevel, bool IsIntraBlock);
 };
 } // end anonymous namespace
 
@@ -215,11 +224,28 @@ PreAllocSplitting::findSpillPoint(MachineBasicBlock *MBB, MachineInstr *MI,
   if (RefsInMBB.empty() && !DefMI) {
     MachineBasicBlock::iterator MII = MBB->begin();
     MachineBasicBlock::iterator EndPt = MI;
+    
+    if (MII == EndPt) return Pt;
+    
     do {
       ++MII;
       unsigned Index = LIs->getInstructionIndex(MII);
       unsigned Gap = LIs->findGapBeforeInstr(Index);
-      if (Gap) {
+      
+      // We can't insert the spill between the barrier (a call), and its
+      // corresponding call frame setup/teardown.
+      if (prior(MII)->getOpcode() == TRI->getCallFrameSetupOpcode()) {
+        bool reachedBarrier = false;
+        do {
+          if (MII == EndPt) {
+            reachedBarrier = true;
+            break;
+          }
+          ++MII;
+        } while (MII->getOpcode() != TRI->getCallFrameDestroyOpcode());
+        
+        if (reachedBarrier) break;
+      } else if (Gap) {
         Pt = MII;
         SpillIndex = Gap;
         break;
@@ -229,9 +255,28 @@ PreAllocSplitting::findSpillPoint(MachineBasicBlock *MBB, MachineInstr *MI,
     MachineBasicBlock::iterator MII = MI;
     MachineBasicBlock::iterator EndPt = DefMI
       ? MachineBasicBlock::iterator(DefMI) : MBB->begin();
+    
     while (MII != EndPt && !RefsInMBB.count(MII)) {
       unsigned Index = LIs->getInstructionIndex(MII);
-      if (LIs->hasGapBeforeInstr(Index)) {
+      
+      // We can't insert the spill between the barrier (a call), and its
+      // corresponding call frame setup.
+      if (prior(MII)->getOpcode() == TRI->getCallFrameSetupOpcode()) {
+        --MII;
+        continue;
+      } if (MII->getOpcode() == TRI->getCallFrameDestroyOpcode()) {
+        bool reachedBarrier = false;
+        while (MII->getOpcode() != TRI->getCallFrameSetupOpcode()) {
+          --MII;
+          if (MII == EndPt) {
+            reachedBarrier = true;
+            break;
+          }
+        }
+        
+        if (reachedBarrier) break;
+        else continue;
+      } else if (LIs->hasGapBeforeInstr(Index)) {
         Pt = MII;
         SpillIndex = LIs->findGapBeforeInstr(Index, true);
       }
@@ -261,31 +306,70 @@ PreAllocSplitting::findRestorePoint(MachineBasicBlock *MBB, MachineInstr *MI,
   if (RefsInMBB.empty() && LastIdx >= EndIdx) {
     MachineBasicBlock::iterator MII = MBB->getFirstTerminator();
     MachineBasicBlock::iterator EndPt = MI;
+    
+    if (MII == EndPt) return Pt;
+    
     --MII;
     do {
       unsigned Index = LIs->getInstructionIndex(MII);
       unsigned Gap = LIs->findGapBeforeInstr(Index);
-      if (Gap) {
+      
+      // We can't insert a restore between the barrier (a call) and its 
+      // corresponding call frame teardown.
+      if (MII->getOpcode() == TRI->getCallFrameDestroyOpcode()) {
+        bool reachedBarrier = false;
+        while (MII->getOpcode() != TRI->getCallFrameSetupOpcode()) {
+          --MII;
+          if (MII == EndPt) {
+            reachedBarrier = true;
+            break;
+          }
+        }
+        
+        if (reachedBarrier) break;
+        else continue;
+      } else if (Gap) {
         Pt = MII;
         RestoreIndex = Gap;
         break;
       }
+      
       --MII;
     } while (MII != EndPt);
   } else {
     MachineBasicBlock::iterator MII = MI;
     MII = ++MII;
+    
     // FIXME: Limit the number of instructions to examine to reduce
     // compile time?
-    while (MII != MBB->end()) {
+    while (MII != MBB->getFirstTerminator()) {
       unsigned Index = LIs->getInstructionIndex(MII);
       if (Index > LastIdx)
         break;
       unsigned Gap = LIs->findGapBeforeInstr(Index);
-      if (Gap) {
+      
+      // We can't insert a restore between the barrier (a call) and its 
+      // corresponding call frame teardown.
+      if (MII->getOpcode() == TRI->getCallFrameDestroyOpcode()) {
+        ++MII;
+        continue;
+      } else if (prior(MII)->getOpcode() == TRI->getCallFrameSetupOpcode()) {
+        bool reachedBarrier = false;
+        do {
+          if (MII == MBB->getFirstTerminator() || RefsInMBB.count(MII)) {
+            reachedBarrier = true;
+            break;
+          }
+          
+          ++MII;
+        } while (MII->getOpcode() != TRI->getCallFrameDestroyOpcode());
+        
+        if (reachedBarrier) break;
+      } else if (Gap) {
         Pt = MII;
         RestoreIndex = Gap;
       }
+      
       if (RefsInMBB.count(MII))
         break;
       ++MII;
@@ -411,94 +495,35 @@ PreAllocSplitting::UpdateSpillSlotInterval(VNInfo *ValNo, unsigned SpillIndex,
 
 /// PerformPHIConstruction - From properly set up use and def lists, use a PHI
 /// construction algorithm to compute the ranges and valnos for an interval.
-VNInfo* PreAllocSplitting::PerformPHIConstruction(
-                                                MachineBasicBlock::iterator use,
-                                                         MachineBasicBlock* MBB,
-                                                               LiveInterval* LI,
+VNInfo*
+PreAllocSplitting::PerformPHIConstruction(MachineBasicBlock::iterator UseI,
+                                       MachineBasicBlock* MBB, LiveInterval* LI,
                                        SmallPtrSet<MachineInstr*, 4>& Visited,
              DenseMap<MachineBasicBlock*, SmallPtrSet<MachineInstr*, 2> >& Defs,
              DenseMap<MachineBasicBlock*, SmallPtrSet<MachineInstr*, 2> >& Uses,
                                        DenseMap<MachineInstr*, VNInfo*>& NewVNs,
                                  DenseMap<MachineBasicBlock*, VNInfo*>& LiveOut,
                                  DenseMap<MachineBasicBlock*, VNInfo*>& Phis,
-                                              bool toplevel, bool intrablock) {
+                                           bool IsTopLevel, bool IsIntraBlock) {
   // Return memoized result if it's available.
-  if (toplevel && Visited.count(use) && NewVNs.count(use))
-    return NewVNs[use];
-  else if (!toplevel && intrablock && NewVNs.count(use))
-    return NewVNs[use];
-  else if (!intrablock && LiveOut.count(MBB))
+  if (IsTopLevel && Visited.count(UseI) && NewVNs.count(UseI))
+    return NewVNs[UseI];
+  else if (!IsTopLevel && IsIntraBlock && NewVNs.count(UseI))
+    return NewVNs[UseI];
+  else if (!IsIntraBlock && LiveOut.count(MBB))
     return LiveOut[MBB];
   
-  typedef DenseMap<MachineBasicBlock*, SmallPtrSet<MachineInstr*, 2> > RegMap;
-  
   // Check if our block contains any uses or defs.
   bool ContainsDefs = Defs.count(MBB);
   bool ContainsUses = Uses.count(MBB);
   
-  VNInfo* ret = 0;
+  VNInfo* RetVNI = 0;
   
   // Enumerate the cases of use/def contaning blocks.
   if (!ContainsDefs && !ContainsUses) {
-  Fallback:
-    // NOTE: Because this is the fallback case from other cases, we do NOT
-    // assume that we are not intrablock here.
-    if (Phis.count(MBB)) return Phis[MBB];
-    
-    unsigned StartIndex = LIs->getMBBStartIdx(MBB);
-    
-    if (MBB->pred_size() == 1) {
-      Phis[MBB] = ret = PerformPHIConstruction((*MBB->pred_begin())->end(),
-                                          *(MBB->pred_begin()), LI, Visited,
-                                          Defs, Uses, NewVNs, LiveOut, Phis,
-                                          false, false);
-      unsigned EndIndex = 0;
-      if (intrablock) {
-        EndIndex = LIs->getInstructionIndex(use);
-        EndIndex = LiveIntervals::getUseIndex(EndIndex);
-      } else
-        EndIndex = LIs->getMBBEndIdx(MBB);
-      
-      LI->addRange(LiveRange(StartIndex, EndIndex+1, ret));
-      if (intrablock)
-        LI->addKill(ret, EndIndex);
-    } else {
-      Phis[MBB] = ret = LI->getNextValue(~0U, /*FIXME*/ 0,
-                                          LIs->getVNInfoAllocator());
-      if (!intrablock) LiveOut[MBB] = ret;
-    
-      // If there are no uses or defs between our starting point and the
-      // beginning of the block, then recursive perform phi construction
-      // on our predecessors.
-      DenseMap<MachineBasicBlock*, VNInfo*> IncomingVNs;
-      for (MachineBasicBlock::pred_iterator PI = MBB->pred_begin(),
-           PE = MBB->pred_end(); PI != PE; ++PI) {
-        VNInfo* Incoming = PerformPHIConstruction((*PI)->end(), *PI, LI, 
-                                            Visited, Defs, Uses, NewVNs,
-                                            LiveOut, Phis, false, false);
-        if (Incoming != 0)
-          IncomingVNs[*PI] = Incoming;
-      }
-    
-      // Otherwise, merge the incoming VNInfos with a phi join.  Create a new
-      // VNInfo to represent the joined value.
-      for (DenseMap<MachineBasicBlock*, VNInfo*>::iterator I =
-           IncomingVNs.begin(), E = IncomingVNs.end(); I != E; ++I) {
-        I->second->hasPHIKill = true;
-        unsigned KillIndex = LIs->getMBBEndIdx(I->first);
-        LI->addKill(I->second, KillIndex);
-      }
-      
-      unsigned EndIndex = 0;
-      if (intrablock) {
-        EndIndex = LIs->getInstructionIndex(use);
-        EndIndex = LiveIntervals::getUseIndex(EndIndex);
-      } else
-        EndIndex = LIs->getMBBEndIdx(MBB);
-      LI->addRange(LiveRange(StartIndex, EndIndex+1, ret));
-      if (intrablock)
-        LI->addKill(ret, EndIndex);
-    }
+    return PerformPHIConstructionFallBack(UseI, MBB, LI, Visited, Defs, Uses,
+                                          NewVNs, LiveOut, Phis,
+                                          IsTopLevel, IsIntraBlock);
   } else if (ContainsDefs && !ContainsUses) {
     SmallPtrSet<MachineInstr*, 2>& BlockDefs = Defs[MBB];
 
@@ -506,71 +531,75 @@ VNInfo* PreAllocSplitting::PerformPHIConstruction(
     // instruction we care about, go to the fallback case.  Note that that
     // should never happen: this cannot be intrablock, so use should
     // always be an end() iterator.
-    assert(use == MBB->end() && "No use marked in intrablock");
+    assert(UseI == MBB->end() && "No use marked in intrablock");
     
-    MachineBasicBlock::iterator walker = use;
-    --walker;
-    while (walker != MBB->begin())
-      if (BlockDefs.count(walker)) {
+    MachineBasicBlock::iterator Walker = UseI;
+    --Walker;
+    while (Walker != MBB->begin()) {
+      if (BlockDefs.count(Walker))
         break;
-      } else
-        --walker;
+      --Walker;
+    }
     
     // Once we've found it, extend its VNInfo to our instruction.
-    unsigned DefIndex = LIs->getInstructionIndex(walker);
+    unsigned DefIndex = LIs->getInstructionIndex(Walker);
     DefIndex = LiveIntervals::getDefIndex(DefIndex);
     unsigned EndIndex = LIs->getMBBEndIdx(MBB);
     
-    ret = NewVNs[walker];
-    LI->addRange(LiveRange(DefIndex, EndIndex+1, ret));
+    RetVNI = NewVNs[Walker];
+    LI->addRange(LiveRange(DefIndex, EndIndex+1, RetVNI));
   } else if (!ContainsDefs && ContainsUses) {
     SmallPtrSet<MachineInstr*, 2>& BlockUses = Uses[MBB];
     
     // Search for the use in this block that precedes the instruction we care 
-    // about, going to the fallback case if we don't find it.
+    // about, going to the fallback case if we don't find it.    
+    if (UseI == MBB->begin())
+      return PerformPHIConstructionFallBack(UseI, MBB, LI, Visited, Defs,
+                                            Uses, NewVNs, LiveOut, Phis,
+                                            IsTopLevel, IsIntraBlock);
     
-    if (use == MBB->begin())
-      goto Fallback;
-    
-    MachineBasicBlock::iterator walker = use;
-    --walker;
+    MachineBasicBlock::iterator Walker = UseI;
+    --Walker;
     bool found = false;
-    while (walker != MBB->begin())
-      if (BlockUses.count(walker)) {
+    while (Walker != MBB->begin()) {
+      if (BlockUses.count(Walker)) {
         found = true;
         break;
-      } else
-        --walker;
+      }
+      --Walker;
+    }
         
     // Must check begin() too.
     if (!found) {
-      if (BlockUses.count(walker))
+      if (BlockUses.count(Walker))
         found = true;
       else
-        goto Fallback;
+        return PerformPHIConstructionFallBack(UseI, MBB, LI, Visited, Defs,
+                                              Uses, NewVNs, LiveOut, Phis,
+                                              IsTopLevel, IsIntraBlock);
     }
 
-    unsigned UseIndex = LIs->getInstructionIndex(walker);
+    unsigned UseIndex = LIs->getInstructionIndex(Walker);
     UseIndex = LiveIntervals::getUseIndex(UseIndex);
     unsigned EndIndex = 0;
-    if (intrablock) {
-      EndIndex = LIs->getInstructionIndex(use);
+    if (IsIntraBlock) {
+      EndIndex = LIs->getInstructionIndex(UseI);
       EndIndex = LiveIntervals::getUseIndex(EndIndex);
     } else
       EndIndex = LIs->getMBBEndIdx(MBB);
 
     // Now, recursively phi construct the VNInfo for the use we found,
     // and then extend it to include the instruction we care about
-    ret = PerformPHIConstruction(walker, MBB, LI, Visited, Defs, Uses,
-                                 NewVNs, LiveOut, Phis, false, true);
+    RetVNI = PerformPHIConstruction(Walker, MBB, LI, Visited, Defs, Uses,
+                                    NewVNs, LiveOut, Phis, false, true);
     
-    LI->addRange(LiveRange(UseIndex, EndIndex+1, ret));
+    LI->addRange(LiveRange(UseIndex, EndIndex+1, RetVNI));
     
     // FIXME: Need to set kills properly for inter-block stuff.
-    if (LI->isKill(ret, UseIndex)) LI->removeKill(ret, UseIndex);
-    if (intrablock)
-      LI->addKill(ret, EndIndex);
-  } else if (ContainsDefs && ContainsUses){
+    if (LI->isKill(RetVNI, UseIndex)) LI->removeKill(RetVNI, UseIndex);
+    if (IsIntraBlock)
+      LI->addKill(RetVNI, EndIndex);
+  } else if (ContainsDefs && ContainsUses) {
     SmallPtrSet<MachineInstr*, 2>& BlockDefs = Defs[MBB];
     SmallPtrSet<MachineInstr*, 2>& BlockUses = Uses[MBB];
     
@@ -578,67 +607,159 @@ VNInfo* PreAllocSplitting::PerformPHIConstruction(
     // special note that checking for defs must take precedence over checking
     // for uses, because of two-address instructions.
     
-    if (use == MBB->begin())
-      goto Fallback;
+    if (UseI == MBB->begin())
+      return PerformPHIConstructionFallBack(UseI, MBB, LI, Visited, Defs, Uses,
+                                            NewVNs, LiveOut, Phis,
+                                            IsTopLevel, IsIntraBlock);
     
-    MachineBasicBlock::iterator walker = use;
-    --walker;
+    MachineBasicBlock::iterator Walker = UseI;
+    --Walker;
     bool foundDef = false;
     bool foundUse = false;
-    while (walker != MBB->begin())
-      if (BlockDefs.count(walker)) {
+    while (Walker != MBB->begin()) {
+      if (BlockDefs.count(Walker)) {
         foundDef = true;
         break;
-      } else if (BlockUses.count(walker)) {
+      } else if (BlockUses.count(Walker)) {
         foundUse = true;
         break;
-      } else
-        --walker;
+      }
+      --Walker;
+    }
         
     // Must check begin() too.
     if (!foundDef && !foundUse) {
-      if (BlockDefs.count(walker))
+      if (BlockDefs.count(Walker))
         foundDef = true;
-      else if (BlockUses.count(walker))
+      else if (BlockUses.count(Walker))
         foundUse = true;
       else
-        goto Fallback;
+        return PerformPHIConstructionFallBack(UseI, MBB, LI, Visited, Defs,
+                                              Uses, NewVNs, LiveOut, Phis,
+                                              IsTopLevel, IsIntraBlock);
     }
 
-    unsigned StartIndex = LIs->getInstructionIndex(walker);
+    unsigned StartIndex = LIs->getInstructionIndex(Walker);
     StartIndex = foundDef ? LiveIntervals::getDefIndex(StartIndex) :
                             LiveIntervals::getUseIndex(StartIndex);
     unsigned EndIndex = 0;
-    if (intrablock) {
-      EndIndex = LIs->getInstructionIndex(use);
+    if (IsIntraBlock) {
+      EndIndex = LIs->getInstructionIndex(UseI);
       EndIndex = LiveIntervals::getUseIndex(EndIndex);
     } else
       EndIndex = LIs->getMBBEndIdx(MBB);
 
     if (foundDef)
-      ret = NewVNs[walker];
+      RetVNI = NewVNs[Walker];
     else
-      ret = PerformPHIConstruction(walker, MBB, LI, Visited, Defs, Uses,
-                                   NewVNs, LiveOut, Phis, false, true);
+      RetVNI = PerformPHIConstruction(Walker, MBB, LI, Visited, Defs, Uses,
+                                      NewVNs, LiveOut, Phis, false, true);
 
-    LI->addRange(LiveRange(StartIndex, EndIndex+1, ret));
+    LI->addRange(LiveRange(StartIndex, EndIndex+1, RetVNI));
     
-    if (foundUse && LI->isKill(ret, StartIndex))
-      LI->removeKill(ret, StartIndex);
-    if (intrablock) {
-      LI->addKill(ret, EndIndex);
+    if (foundUse && LI->isKill(RetVNI, StartIndex))
+      LI->removeKill(RetVNI, StartIndex);
+    if (IsIntraBlock) {
+      LI->addKill(RetVNI, EndIndex);
     }
   }
   
   // Memoize results so we don't have to recompute them.
-  if (!intrablock) LiveOut[MBB] = ret;
+  if (!IsIntraBlock) LiveOut[MBB] = RetVNI;
   else {
-    if (!NewVNs.count(use))
-      NewVNs[use] = ret;
-    Visited.insert(use);
+    if (!NewVNs.count(UseI))
+      NewVNs[UseI] = RetVNI;
+    Visited.insert(UseI);
   }
 
-  return ret;
+  return RetVNI;
+}
+
+/// PerformPHIConstructionFallBack - PerformPHIConstruction fall back path.
+///
+VNInfo*
+PreAllocSplitting::PerformPHIConstructionFallBack(MachineBasicBlock::iterator UseI,
+                                       MachineBasicBlock* MBB, LiveInterval* LI,
+                                       SmallPtrSet<MachineInstr*, 4>& Visited,
+             DenseMap<MachineBasicBlock*, SmallPtrSet<MachineInstr*, 2> >& Defs,
+             DenseMap<MachineBasicBlock*, SmallPtrSet<MachineInstr*, 2> >& Uses,
+                                       DenseMap<MachineInstr*, VNInfo*>& NewVNs,
+                                 DenseMap<MachineBasicBlock*, VNInfo*>& LiveOut,
+                                 DenseMap<MachineBasicBlock*, VNInfo*>& Phis,
+                                           bool IsTopLevel, bool IsIntraBlock) {
+  // NOTE: Because this is the fallback case from other cases, we do NOT
+  // assume that we are not intrablock here.
+  if (Phis.count(MBB)) return Phis[MBB]; 
+
+  unsigned StartIndex = LIs->getMBBStartIdx(MBB);
+  VNInfo *RetVNI = Phis[MBB] = LI->getNextValue(~0U, /*FIXME*/ 0,
+                                                LIs->getVNInfoAllocator());
+  if (!IsIntraBlock) LiveOut[MBB] = RetVNI;
+    
+  // If there are no uses or defs between our starting point and the
+  // beginning of the block, then recursive perform phi construction
+  // on our predecessors.
+  DenseMap<MachineBasicBlock*, VNInfo*> IncomingVNs;
+  for (MachineBasicBlock::pred_iterator PI = MBB->pred_begin(),
+         PE = MBB->pred_end(); PI != PE; ++PI) {
+    VNInfo* Incoming = PerformPHIConstruction((*PI)->end(), *PI, LI, 
+                                              Visited, Defs, Uses, NewVNs,
+                                              LiveOut, Phis, false, false);
+    if (Incoming != 0)
+      IncomingVNs[*PI] = Incoming;
+  }
+    
+  if (MBB->pred_size() == 1 && !RetVNI->hasPHIKill) {
+    VNInfo* OldVN = RetVNI;
+    VNInfo* NewVN = IncomingVNs.begin()->second;
+    VNInfo* MergedVN = LI->MergeValueNumberInto(OldVN, NewVN);
+    if (MergedVN == OldVN) std::swap(OldVN, NewVN);
+    
+    for (DenseMap<MachineBasicBlock*, VNInfo*>::iterator LOI = LiveOut.begin(),
+         LOE = LiveOut.end(); LOI != LOE; ++LOI)
+      if (LOI->second == OldVN)
+        LOI->second = MergedVN;
+    for (DenseMap<MachineInstr*, VNInfo*>::iterator NVI = NewVNs.begin(),
+         NVE = NewVNs.end(); NVI != NVE; ++NVI)
+      if (NVI->second == OldVN)
+        NVI->second = MergedVN;
+    for (DenseMap<MachineBasicBlock*, VNInfo*>::iterator PI = Phis.begin(),
+         PE = Phis.end(); PI != PE; ++PI)
+      if (PI->second == OldVN)
+        PI->second = MergedVN;
+    RetVNI = MergedVN;
+  } else {
+    // Otherwise, merge the incoming VNInfos with a phi join.  Create a new
+    // VNInfo to represent the joined value.
+    for (DenseMap<MachineBasicBlock*, VNInfo*>::iterator I =
+           IncomingVNs.begin(), E = IncomingVNs.end(); I != E; ++I) {
+      I->second->hasPHIKill = true;
+      unsigned KillIndex = LIs->getMBBEndIdx(I->first);
+      if (!LiveInterval::isKill(I->second, KillIndex))
+        LI->addKill(I->second, KillIndex);
+    }
+  }
+      
+  unsigned EndIndex = 0;
+  if (IsIntraBlock) {
+    EndIndex = LIs->getInstructionIndex(UseI);
+    EndIndex = LiveIntervals::getUseIndex(EndIndex);
+  } else
+    EndIndex = LIs->getMBBEndIdx(MBB);
+  LI->addRange(LiveRange(StartIndex, EndIndex+1, RetVNI));
+  if (IsIntraBlock)
+    LI->addKill(RetVNI, EndIndex);
+
+  // Memoize results so we don't have to recompute them.
+  if (!IsIntraBlock)
+    LiveOut[MBB] = RetVNI;
+  else {
+    if (!NewVNs.count(UseI))
+      NewVNs[UseI] = RetVNI;
+    Visited.insert(UseI);
+  }
+
+  return RetVNI;
 }
 
 /// ReconstructLiveInterval - Recompute a live interval from scratch.
@@ -1013,7 +1134,10 @@ PreAllocSplitting::SplitRegLiveIntervals(const TargetRegisterClass **RCs,
   // by the current barrier.
   SmallVector<LiveInterval*, 8> Intervals;
   for (const TargetRegisterClass **RC = RCs; *RC; ++RC) {
-    if (TII->IgnoreRegisterClassBarriers(*RC))
+    // FIXME: If it's not safe to move any instruction that defines the barrier
+    // register class, then it means there are some special dependencies which
+    // codegen is not modelling. Ignore these barriers for now.
+    if (!TII->isSafeToMoveRegClassDefs(*RC))
       continue;
     std::vector<unsigned> &VRs = MRI->getRegClassVirtRegs(*RC);
     for (unsigned i = 0, e = VRs.size(); i != e; ++i) {
@@ -1070,10 +1194,14 @@ unsigned PreAllocSplitting::getNumberOfNonSpills(
 bool PreAllocSplitting::removeDeadSpills(SmallPtrSet<LiveInterval*, 8>& split) {
   bool changed = false;
   
+  // Walk over all of the live intervals that were touched by the splitter,
+  // and see if we can do any DCE and/or folding.
   for (SmallPtrSet<LiveInterval*, 8>::iterator LI = split.begin(),
        LE = split.end(); LI != LE; ++LI) {
     DenseMap<VNInfo*, SmallPtrSet<MachineInstr*, 4> > VNUseCount;
     
+    // First, collect all the uses of the vreg, and sort them by their
+    // reaching definition (VNInfo).
     for (MachineRegisterInfo::use_iterator UI = MRI->use_begin((*LI)->reg),
          UE = MRI->use_end(); UI != UE; ++UI) {
       unsigned index = LIs->getInstructionIndex(&*UI);
@@ -1083,6 +1211,8 @@ bool PreAllocSplitting::removeDeadSpills(SmallPtrSet<LiveInterval*, 8>& split) {
       VNUseCount[LR->valno].insert(&*UI);
     }
     
+    // Now, take the definitions (VNInfo's) one at a time and try to DCE 
+    // and/or fold them away.
     for (LiveInterval::vni_iterator VI = (*LI)->vni_begin(),
          VE = (*LI)->vni_end(); VI != VE; ++VI) {
       
@@ -1090,30 +1220,45 @@ bool PreAllocSplitting::removeDeadSpills(SmallPtrSet<LiveInterval*, 8>& split) {
         return changed;
       
       VNInfo* CurrVN = *VI;
+      
+      // We don't currently try to handle definitions with PHI kills, because
+      // it would involve processing more than one VNInfo at once.
       if (CurrVN->hasPHIKill) continue;
       
+      // We also don't try to handle the results of PHI joins, since there's
+      // no defining instruction to analyze.
       unsigned DefIdx = CurrVN->def;
       if (DefIdx == ~0U || DefIdx == ~1U) continue;
     
+      // We're only interested in eliminating cruft introduced by the splitter,
+      // is of the form load-use or load-use-store.  First, check that the
+      // definition is a load, and remember what stack slot we loaded it from.
       MachineInstr* DefMI = LIs->getInstructionFromIndex(DefIdx);
       int FrameIndex;
       if (!TII->isLoadFromStackSlot(DefMI, FrameIndex)) continue;
       
+      // If the definition has no uses at all, just DCE it.
       if (VNUseCount[CurrVN].size() == 0) {
         LIs->RemoveMachineInstrFromMaps(DefMI);
         (*LI)->removeValNo(CurrVN);
         DefMI->eraseFromParent();
+        VNUseCount.erase(CurrVN);
         NumDeadSpills++;
         changed = true;
         continue;
       }
       
+      // Second, get the number of non-store uses of the definition, as well as
+      // a flag indicating whether it feeds into a later two-address definition.
       bool FeedsTwoAddr = false;
       unsigned NonSpillCount = getNumberOfNonSpills(VNUseCount[CurrVN],
                                                     (*LI)->reg, FrameIndex,
                                                     FeedsTwoAddr);
       
+      // If there's one non-store use and it doesn't feed a two-addr, then
+      // this is a load-use-store case that we can try to fold.
       if (NonSpillCount == 1 && !FeedsTwoAddr) {
+        // Start by finding the non-store use MachineInstr.
         SmallPtrSet<MachineInstr*, 4>::iterator UI = VNUseCount[CurrVN].begin();
         int StoreFrameIndex;
         unsigned StoreVReg = TII->isStoreToStackSlot(*UI, StoreFrameIndex);
@@ -1123,17 +1268,15 @@ bool PreAllocSplitting::removeDeadSpills(SmallPtrSet<LiveInterval*, 8>& split) {
           if (UI != VNUseCount[CurrVN].end())
             StoreVReg = TII->isStoreToStackSlot(*UI, StoreFrameIndex);
         }
-        
         if (UI == VNUseCount[CurrVN].end()) continue;
         
         MachineInstr* use = *UI;
         
+        // Attempt to fold it away!
         int OpIdx = use->findRegisterUseOperandIdx((*LI)->reg, false);
         if (OpIdx == -1) continue;
-
         SmallVector<unsigned, 1> Ops;
         Ops.push_back(OpIdx);
-
         if (!TII->canFoldMemoryOperand(use, Ops)) continue;
 
         MachineInstr* NewMI =
@@ -1142,6 +1285,7 @@ bool PreAllocSplitting::removeDeadSpills(SmallPtrSet<LiveInterval*, 8>& split) {
 
         if (!NewMI) continue;
 
+        // Update relevant analyses.
         LIs->RemoveMachineInstrFromMaps(DefMI);
         LIs->ReplaceMachineInstrInMaps(use, NewMI);
         (*LI)->removeValNo(CurrVN);
@@ -1151,12 +1295,21 @@ bool PreAllocSplitting::removeDeadSpills(SmallPtrSet<LiveInterval*, 8>& split) {
         NewMI = MBB->insert(MBB->erase(use), NewMI);
         VNUseCount[CurrVN].erase(use);
         
+        // Remove deleted instructions.  Note that we need to remove them from 
+        // the VNInfo->use map as well, just to be safe.
         for (SmallPtrSet<MachineInstr*, 4>::iterator II = 
              VNUseCount[CurrVN].begin(), IE = VNUseCount[CurrVN].end();
              II != IE; ++II) {
+          for (DenseMap<VNInfo*, SmallPtrSet<MachineInstr*, 4> >::iterator
+               VNI = VNUseCount.begin(), VNE = VNUseCount.end(); VNI != VNE; 
+               ++VNI)
+            if (VNI->first != CurrVN)
+              VNI->second.erase(*II);
           LIs->RemoveMachineInstrFromMaps(*II);
           (*II)->eraseFromParent();
         }
+        
+        VNUseCount.erase(CurrVN);
 
         for (DenseMap<VNInfo*, SmallPtrSet<MachineInstr*, 4> >::iterator
              VI = VNUseCount.begin(), VE = VNUseCount.end(); VI != VE; ++VI)
@@ -1168,8 +1321,11 @@ bool PreAllocSplitting::removeDeadSpills(SmallPtrSet<LiveInterval*, 8>& split) {
         continue;
       }
       
+      // If there's more than one non-store instruction, we can't profitably
+      // fold it, so bail.
       if (NonSpillCount) continue;
         
+      // Otherwise, this is a load-store case, so DCE them.
       for (SmallPtrSet<MachineInstr*, 4>::iterator UI = 
            VNUseCount[CurrVN].begin(), UE = VNUseCount[CurrVN].end();
            UI != UI; ++UI) {
@@ -1177,6 +1333,8 @@ bool PreAllocSplitting::removeDeadSpills(SmallPtrSet<LiveInterval*, 8>& split) {
         (*UI)->eraseFromParent();
       }
         
+      VNUseCount.erase(CurrVN);
+        
       LIs->RemoveMachineInstrFromMaps(DefMI);
       (*LI)->removeValNo(CurrVN);
       DefMI->eraseFromParent();
@@ -1256,6 +1414,7 @@ bool PreAllocSplitting::createsNewJoin(LiveRange* LR,
 bool PreAllocSplitting::runOnMachineFunction(MachineFunction &MF) {
   CurrMF = &MF;
   TM     = &MF.getTarget();
+  TRI    = TM->getRegisterInfo();
   TII    = TM->getInstrInfo();
   MFI    = MF.getFrameInfo();
   MRI    = &MF.getRegInfo();