Inline check that's used only once.
[oota-llvm.git] / lib / CodeGen / RegAllocGreedy.cpp
index 5f2be811aa6554207ed089ba2b8775000744df41..77a172870a3d12dc120b17ca7f213f93eb2eba00 100644 (file)
 //===----------------------------------------------------------------------===//
 
 #define DEBUG_TYPE "regalloc"
+#include "AllocationOrder.h"
 #include "LiveIntervalUnion.h"
+#include "LiveRangeEdit.h"
 #include "RegAllocBase.h"
 #include "Spiller.h"
+#include "SpillPlacement.h"
+#include "SplitKit.h"
 #include "VirtRegMap.h"
-#include "VirtRegRewriter.h"
+#include "llvm/ADT/Statistic.h"
 #include "llvm/Analysis/AliasAnalysis.h"
 #include "llvm/Function.h"
 #include "llvm/PassAnalysisSupport.h"
 #include "llvm/CodeGen/CalcSpillWeights.h"
+#include "llvm/CodeGen/EdgeBundles.h"
 #include "llvm/CodeGen/LiveIntervalAnalysis.h"
 #include "llvm/CodeGen/LiveStackAnalysis.h"
+#include "llvm/CodeGen/MachineDominators.h"
 #include "llvm/CodeGen/MachineFunctionPass.h"
 #include "llvm/CodeGen/MachineLoopInfo.h"
+#include "llvm/CodeGen/MachineLoopRanges.h"
 #include "llvm/CodeGen/MachineRegisterInfo.h"
 #include "llvm/CodeGen/Passes.h"
 #include "llvm/CodeGen/RegAllocRegistry.h"
 #include "llvm/Support/Debug.h"
 #include "llvm/Support/ErrorHandling.h"
 #include "llvm/Support/raw_ostream.h"
+#include "llvm/Support/Timer.h"
+
+#include <queue>
 
 using namespace llvm;
 
+STATISTIC(NumGlobalSplits, "Number of split global live ranges");
+STATISTIC(NumLocalSplits,  "Number of split local live ranges");
+STATISTIC(NumEvicted,      "Number of interferences evicted");
+
 static RegisterRegAlloc greedyRegAlloc("greedy", "greedy register allocator",
                                        createGreedyRegisterAllocator);
 
 namespace {
-class RAGreedy : public MachineFunctionPass, public RegAllocBase {
+class RAGreedy : public MachineFunctionPass,
+                 public RegAllocBase,
+                 private LiveRangeEdit::Delegate {
+
   // context
   MachineFunction *MF;
-  const TargetMachine *TM;
-  MachineRegisterInfo *MRI;
-
   BitVector ReservedRegs;
 
   // analyses
+  SlotIndexes *Indexes;
   LiveStacks *LS;
+  MachineDominatorTree *DomTree;
+  MachineLoopInfo *Loops;
+  MachineLoopRanges *LoopRanges;
+  EdgeBundles *Bundles;
+  SpillPlacement *SpillPlacer;
 
   // state
   std::auto_ptr<Spiller> SpillerInstance;
+  std::priority_queue<std::pair<unsigned, unsigned> > Queue;
+
+  // Live ranges pass through a number of stages as we try to allocate them.
+  // Some of the stages may also create new live ranges:
+  //
+  // - Region splitting.
+  // - Per-block splitting.
+  // - Local splitting.
+  // - Spilling.
+  //
+  // Ranges produced by one of the stages skip the previous stages when they are
+  // dequeued. This improves performance because we can skip interference checks
+  // that are unlikely to give any results. It also guarantees that the live
+  // range splitting algorithm terminates, something that is otherwise hard to
+  // ensure.
+  enum LiveRangeStage {
+    RS_Original, ///< Never seen before, never split.
+    RS_Second,   ///< Second time in the queue.
+    RS_Region,   ///< Produced by region splitting.
+    RS_Block,    ///< Produced by per-block splitting.
+    RS_Local,    ///< Produced by local splitting.
+    RS_Spill     ///< Produced by spilling.
+  };
+
+  IndexedMap<unsigned char, VirtReg2IndexFunctor> LRStage;
+
+  LiveRangeStage getStage(const LiveInterval &VirtReg) const {
+    return LiveRangeStage(LRStage[VirtReg.reg]);
+  }
+
+  template<typename Iterator>
+  void setStage(Iterator Begin, Iterator End, LiveRangeStage NewStage) {
+    LRStage.resize(MRI->getNumVirtRegs());
+    for (;Begin != End; ++Begin)
+      LRStage[(*Begin)->reg] = NewStage;
+  }
+
+  // splitting state.
+  std::auto_ptr<SplitAnalysis> SA;
+  std::auto_ptr<SplitEditor> SE;
+
+  /// All basic blocks where the current register is live.
+  SmallVector<SpillPlacement::BlockConstraint, 8> SplitConstraints;
+
+  typedef std::pair<SlotIndex, SlotIndex> IndexPair;
+
+  /// Global live range splitting candidate info.
+  struct GlobalSplitCandidate {
+    unsigned PhysReg;
+    SmallVector<IndexPair, 8> Interference;
+    BitVector LiveBundles;
+  };
+
+  /// Candidate info for for each PhysReg in AllocationOrder.
+  /// This vector never shrinks, but grows to the size of the largest register
+  /// class.
+  SmallVector<GlobalSplitCandidate, 32> GlobalCand;
+
+  /// For every instruction in SA->UseSlots, store the previous non-copy
+  /// instruction.
+  SmallVector<SlotIndex, 8> PrevSlot;
 
 public:
   RAGreedy();
 
   /// Return the pass name.
   virtual const char* getPassName() const {
-    return "Basic Register Allocator";
+    return "Greedy Register Allocator";
   }
 
   /// RAGreedy analysis usage.
   virtual void getAnalysisUsage(AnalysisUsage &AU) const;
-
   virtual void releaseMemory();
-
   virtual Spiller &spiller() { return *SpillerInstance; }
-
-  virtual float getPriority(LiveInterval *LI) { return LI->weight; }
-
-  virtual unsigned selectOrSplit(LiveInterval &VirtReg,
-                                 SmallVectorImpl<LiveInterval*> &SplitVRegs);
+  virtual void enqueue(LiveInterval *LI);
+  virtual LiveInterval *dequeue();
+  virtual unsigned selectOrSplit(LiveInterval&,
+                                 SmallVectorImpl<LiveInterval*>&);
 
   /// Perform register allocation.
   virtual bool runOnMachineFunction(MachineFunction &mf);
 
   static char ID;
+
+private:
+  void LRE_WillEraseInstruction(MachineInstr*);
+  bool LRE_CanEraseVirtReg(unsigned);
+  void LRE_WillShrinkVirtReg(unsigned);
+
+  void mapGlobalInterference(unsigned, SmallVectorImpl<IndexPair>&);
+  float calcSplitConstraints(const SmallVectorImpl<IndexPair>&);
+
+  float calcGlobalSplitCost(const BitVector&);
+  void splitAroundRegion(LiveInterval&, unsigned, const BitVector&,
+                         SmallVectorImpl<LiveInterval*>&);
+  void calcGapWeights(unsigned, SmallVectorImpl<float>&);
+  SlotIndex getPrevMappedIndex(const MachineInstr*);
+  void calcPrevSlots();
+  unsigned nextSplitPoint(unsigned);
+  bool canEvictInterference(LiveInterval&, unsigned, float&);
+
+  unsigned tryEvict(LiveInterval&, AllocationOrder&,
+                    SmallVectorImpl<LiveInterval*>&);
+  unsigned tryRegionSplit(LiveInterval&, AllocationOrder&,
+                          SmallVectorImpl<LiveInterval*>&);
+  unsigned tryLocalSplit(LiveInterval&, AllocationOrder&,
+    SmallVectorImpl<LiveInterval*>&);
+  unsigned trySplit(LiveInterval&, AllocationOrder&,
+                    SmallVectorImpl<LiveInterval*>&);
 };
 } // end anonymous namespace
 
@@ -88,7 +192,8 @@ FunctionPass* llvm::createGreedyRegisterAllocator() {
   return new RAGreedy();
 }
 
-RAGreedy::RAGreedy(): MachineFunctionPass(ID) {
+RAGreedy::RAGreedy(): MachineFunctionPass(ID), LRStage(RS_Original) {
+  initializeSlotIndexesPass(*PassRegistry::getPassRegistry());
   initializeLiveIntervalsPass(*PassRegistry::getPassRegistry());
   initializeSlotIndexesPass(*PassRegistry::getPassRegistry());
   initializeStrongPHIEliminationPass(*PassRegistry::getPassRegistry());
@@ -97,7 +202,10 @@ RAGreedy::RAGreedy(): MachineFunctionPass(ID) {
   initializeLiveStacksPass(*PassRegistry::getPassRegistry());
   initializeMachineDominatorTreePass(*PassRegistry::getPassRegistry());
   initializeMachineLoopInfoPass(*PassRegistry::getPassRegistry());
+  initializeMachineLoopRangesPass(*PassRegistry::getPassRegistry());
   initializeVirtRegMapPass(*PassRegistry::getPassRegistry());
+  initializeEdgeBundlesPass(*PassRegistry::getPassRegistry());
+  initializeSpillPlacementPass(*PassRegistry::getPassRegistry());
 }
 
 void RAGreedy::getAnalysisUsage(AnalysisUsage &AU) const {
@@ -105,6 +213,7 @@ void RAGreedy::getAnalysisUsage(AnalysisUsage &AU) const {
   AU.addRequired<AliasAnalysis>();
   AU.addPreserved<AliasAnalysis>();
   AU.addRequired<LiveIntervals>();
+  AU.addRequired<SlotIndexes>();
   AU.addPreserved<SlotIndexes>();
   if (StrongPHIElim)
     AU.addRequiredID(StrongPHIEliminationID);
@@ -112,70 +221,952 @@ void RAGreedy::getAnalysisUsage(AnalysisUsage &AU) const {
   AU.addRequired<CalculateSpillWeights>();
   AU.addRequired<LiveStacks>();
   AU.addPreserved<LiveStacks>();
-  AU.addRequiredID(MachineDominatorsID);
-  AU.addPreservedID(MachineDominatorsID);
+  AU.addRequired<MachineDominatorTree>();
+  AU.addPreserved<MachineDominatorTree>();
   AU.addRequired<MachineLoopInfo>();
   AU.addPreserved<MachineLoopInfo>();
+  AU.addRequired<MachineLoopRanges>();
+  AU.addPreserved<MachineLoopRanges>();
   AU.addRequired<VirtRegMap>();
   AU.addPreserved<VirtRegMap>();
+  AU.addRequired<EdgeBundles>();
+  AU.addRequired<SpillPlacement>();
   MachineFunctionPass::getAnalysisUsage(AU);
 }
 
+
+//===----------------------------------------------------------------------===//
+//                     LiveRangeEdit delegate methods
+//===----------------------------------------------------------------------===//
+
+void RAGreedy::LRE_WillEraseInstruction(MachineInstr *MI) {
+  // LRE itself will remove from SlotIndexes and parent basic block.
+  VRM->RemoveMachineInstrFromMaps(MI);
+}
+
+bool RAGreedy::LRE_CanEraseVirtReg(unsigned VirtReg) {
+  if (unsigned PhysReg = VRM->getPhys(VirtReg)) {
+    unassign(LIS->getInterval(VirtReg), PhysReg);
+    return true;
+  }
+  // Unassigned virtreg is probably in the priority queue.
+  // RegAllocBase will erase it after dequeueing.
+  return false;
+}
+
+void RAGreedy::LRE_WillShrinkVirtReg(unsigned VirtReg) {
+  unsigned PhysReg = VRM->getPhys(VirtReg);
+  if (!PhysReg)
+    return;
+
+  // Register is assigned, put it back on the queue for reassignment.
+  LiveInterval &LI = LIS->getInterval(VirtReg);
+  unassign(LI, PhysReg);
+  enqueue(&LI);
+}
+
 void RAGreedy::releaseMemory() {
   SpillerInstance.reset(0);
+  LRStage.clear();
   RegAllocBase::releaseMemory();
 }
 
-unsigned RAGreedy::selectOrSplit(LiveInterval &VirtReg,
-                                SmallVectorImpl<LiveInterval*> &SplitVRegs) {
-  // Populate a list of physical register spill candidates.
-  SmallVector<unsigned, 8> PhysRegSpillCands;
-
-  // Check for an available register in this class.
-  const TargetRegisterClass *TRC = MRI->getRegClass(VirtReg.reg);
-  DEBUG(dbgs() << "RegClass: " << TRC->getName() << ' ');
-
-  for (TargetRegisterClass::iterator I = TRC->allocation_order_begin(*MF),
-         E = TRC->allocation_order_end(*MF);
-       I != E; ++I) {
-
-    unsigned PhysReg = *I;
-    if (ReservedRegs.test(PhysReg)) continue;
-
-    // Check interference and as a side effect, intialize queries for this
-    // VirtReg and its aliases.
-    unsigned interfReg = checkPhysRegInterference(VirtReg, PhysReg);
-    if (interfReg == 0) {
-      // Found an available register.
-      return PhysReg;
+void RAGreedy::enqueue(LiveInterval *LI) {
+  // Prioritize live ranges by size, assigning larger ranges first.
+  // The queue holds (size, reg) pairs.
+  const unsigned Size = LI->getSize();
+  const unsigned Reg = LI->reg;
+  assert(TargetRegisterInfo::isVirtualRegister(Reg) &&
+         "Can only enqueue virtual registers");
+  unsigned Prio;
+
+  LRStage.grow(Reg);
+  if (LRStage[Reg] == RS_Second)
+    // Unsplit ranges that couldn't be allocated immediately are deferred until
+    // everything else has been allocated. Long ranges are allocated last so
+    // they are split against realistic interference.
+    Prio = (1u << 31) - Size;
+  else {
+    // Everything else is allocated in long->short order. Long ranges that don't
+    // fit should be spilled ASAP so they don't create interference.
+    Prio = (1u << 31) + Size;
+
+    // Boost ranges that have a physical register hint.
+    if (TargetRegisterInfo::isPhysicalRegister(VRM->getRegAllocPref(Reg)))
+      Prio |= (1u << 30);
+  }
+
+  Queue.push(std::make_pair(Prio, Reg));
+}
+
+LiveInterval *RAGreedy::dequeue() {
+  if (Queue.empty())
+    return 0;
+  LiveInterval *LI = &LIS->getInterval(Queue.top().second);
+  Queue.pop();
+  return LI;
+}
+
+//===----------------------------------------------------------------------===//
+//                         Interference eviction
+//===----------------------------------------------------------------------===//
+
+/// canEvict - Return true if all interferences between VirtReg and PhysReg can
+/// be evicted. Set maxWeight to the maximal spill weight of an interference.
+bool RAGreedy::canEvictInterference(LiveInterval &VirtReg, unsigned PhysReg,
+                                    float &MaxWeight) {
+  float Weight = 0;
+  for (const unsigned *AliasI = TRI->getOverlaps(PhysReg); *AliasI; ++AliasI) {
+    LiveIntervalUnion::Query &Q = query(VirtReg, *AliasI);
+    // If there is 10 or more interferences, chances are one is smaller.
+    if (Q.collectInterferingVRegs(10) >= 10)
+      return false;
+
+    // Check if any interfering live range is heavier than VirtReg.
+    for (unsigned i = 0, e = Q.interferingVRegs().size(); i != e; ++i) {
+      LiveInterval *Intf = Q.interferingVRegs()[i];
+      if (TargetRegisterInfo::isPhysicalRegister(Intf->reg))
+        return false;
+      if (Intf->weight >= VirtReg.weight)
+        return false;
+      Weight = std::max(Weight, Intf->weight);
     }
-    LiveInterval *interferingVirtReg =
-      Queries[interfReg].firstInterference().liveUnionPos().value();
+  }
+  MaxWeight = Weight;
+  return true;
+}
+
+/// tryEvict - Try to evict all interferences for a physreg.
+/// @param  VirtReg Currently unassigned virtual register.
+/// @param  Order   Physregs to try.
+/// @return         Physreg to assign VirtReg, or 0.
+unsigned RAGreedy::tryEvict(LiveInterval &VirtReg,
+                            AllocationOrder &Order,
+                            SmallVectorImpl<LiveInterval*> &NewVRegs){
+  NamedRegionTimer T("Evict", TimerGroupName, TimePassesIsEnabled);
 
-    // The current VirtReg must either spillable, or one of its interferences
-    // must have less spill weight.
-    if (interferingVirtReg->weight < VirtReg.weight ) {
-      PhysRegSpillCands.push_back(PhysReg);
+  // Keep track of the lightest single interference seen so far.
+  float BestWeight = 0;
+  unsigned BestPhys = 0;
+
+  Order.rewind();
+  while (unsigned PhysReg = Order.next()) {
+    float Weight = 0;
+    if (!canEvictInterference(VirtReg, PhysReg, Weight))
+      continue;
+
+    // This is an eviction candidate.
+    DEBUG(dbgs() << "max " << PrintReg(PhysReg, TRI) << " interference = "
+                 << Weight << '\n');
+    if (BestPhys && Weight >= BestWeight)
+      continue;
+
+    // Best so far.
+    BestPhys = PhysReg;
+    BestWeight = Weight;
+    // Stop if the hint can be used.
+    if (Order.isHint(PhysReg))
+      break;
+  }
+
+  if (!BestPhys)
+    return 0;
+
+  DEBUG(dbgs() << "evicting " << PrintReg(BestPhys, TRI) << " interference\n");
+  for (const unsigned *AliasI = TRI->getOverlaps(BestPhys); *AliasI; ++AliasI) {
+    LiveIntervalUnion::Query &Q = query(VirtReg, *AliasI);
+    assert(Q.seenAllInterferences() && "Didn't check all interfererences.");
+    for (unsigned i = 0, e = Q.interferingVRegs().size(); i != e; ++i) {
+      LiveInterval *Intf = Q.interferingVRegs()[i];
+      unassign(*Intf, VRM->getPhys(Intf->reg));
+      ++NumEvicted;
+      NewVRegs.push_back(Intf);
     }
   }
-  // Try to spill another interfering reg with less spill weight.
-  //
-  // FIXME: RAGreedy will sort this list by spill weight.
-  for (SmallVectorImpl<unsigned>::iterator PhysRegI = PhysRegSpillCands.begin(),
-         PhysRegE = PhysRegSpillCands.end(); PhysRegI != PhysRegE; ++PhysRegI) {
+  return BestPhys;
+}
+
+
+//===----------------------------------------------------------------------===//
+//                              Region Splitting
+//===----------------------------------------------------------------------===//
+
+/// mapGlobalInterference - Compute a map of the interference from PhysReg and
+/// its aliases in each block in SA->LiveBlocks.
+/// If LiveBlocks[i] is live-in, Ranges[i].first is the first interference.
+/// If LiveBlocks[i] is live-out, Ranges[i].second is the last interference.
+void RAGreedy::mapGlobalInterference(unsigned PhysReg,
+                                     SmallVectorImpl<IndexPair> &Ranges) {
+  Ranges.assign(SA->LiveBlocks.size(), IndexPair());
+  LiveInterval &VirtReg = const_cast<LiveInterval&>(SA->getParent());
+  for (const unsigned *AI = TRI->getOverlaps(PhysReg); *AI; ++AI) {
+    if (!query(VirtReg, *AI).checkInterference())
+      continue;
+    LiveIntervalUnion::SegmentIter IntI =
+      PhysReg2LiveUnion[*AI].find(VirtReg.beginIndex());
+    if (!IntI.valid())
+      continue;
+    for (unsigned i = 0, e = SA->LiveBlocks.size(); i != e; ++i) {
+      const SplitAnalysis::BlockInfo &BI = SA->LiveBlocks[i];
+      IndexPair &IP = Ranges[i];
+
+      // Skip interference-free blocks.
+      if (IntI.start() >= BI.Stop)
+        continue;
+
+      // First interference in block.
+      if (BI.LiveIn) {
+        IntI.advanceTo(BI.Start);
+        if (!IntI.valid())
+          break;
+        if (IntI.start() >= BI.Stop)
+          continue;
+        if (!IP.first.isValid() || IntI.start() < IP.first)
+          IP.first = IntI.start();
+      }
+
+      // Last interference in block.
+      if (BI.LiveOut) {
+        IntI.advanceTo(BI.Stop);
+        if (!IntI.valid() || IntI.start() >= BI.Stop)
+          --IntI;
+        if (IntI.stop() <= BI.Start)
+          continue;
+        if (!IP.second.isValid() || IntI.stop() > IP.second)
+          IP.second = IntI.stop();
+      }
+    }
+  }
+}
+
+/// calcSplitConstraints - Fill out the SplitConstraints vector based on the
+/// interference pattern in Intf. Return the static cost of this split,
+/// assuming that all preferences in SplitConstraints are met.
+float RAGreedy::calcSplitConstraints(const SmallVectorImpl<IndexPair> &Intf) {
+  // Reset interference dependent info.
+  SplitConstraints.resize(SA->LiveBlocks.size());
+  float StaticCost = 0;
+  for (unsigned i = 0, e = SA->LiveBlocks.size(); i != e; ++i) {
+    SplitAnalysis::BlockInfo &BI = SA->LiveBlocks[i];
+    SpillPlacement::BlockConstraint &BC = SplitConstraints[i];
+    IndexPair IP = Intf[i];
+
+    BC.Number = BI.MBB->getNumber();
+    BC.Entry = (BI.Uses && BI.LiveIn) ?
+      SpillPlacement::PrefReg : SpillPlacement::DontCare;
+    BC.Exit = (BI.Uses && BI.LiveOut) ?
+      SpillPlacement::PrefReg : SpillPlacement::DontCare;
+
+    // Number of spill code instructions to insert.
+    unsigned Ins = 0;
+
+    // Interference for the live-in value.
+    if (IP.first.isValid()) {
+      if (IP.first <= BI.Start)
+        BC.Entry = SpillPlacement::MustSpill, Ins += BI.Uses;
+      else if (!BI.Uses)
+        BC.Entry = SpillPlacement::PrefSpill;
+      else if (IP.first < BI.FirstUse)
+        BC.Entry = SpillPlacement::PrefSpill, ++Ins;
+      else if (IP.first < (BI.LiveThrough ? BI.LastUse : BI.Kill))
+        ++Ins;
+    }
+
+    // Interference for the live-out value.
+    if (IP.second.isValid()) {
+      if (IP.second >= BI.LastSplitPoint)
+        BC.Exit = SpillPlacement::MustSpill, Ins += BI.Uses;
+      else if (!BI.Uses)
+        BC.Exit = SpillPlacement::PrefSpill;
+      else if (IP.second > BI.LastUse)
+        BC.Exit = SpillPlacement::PrefSpill, ++Ins;
+      else if (IP.second > (BI.LiveThrough ? BI.FirstUse : BI.Def))
+        ++Ins;
+    }
+
+    // Accumulate the total frequency of inserted spill code.
+    if (Ins)
+      StaticCost += Ins * SpillPlacer->getBlockFrequency(BC.Number);
+  }
+  return StaticCost;
+}
+
+
+/// calcGlobalSplitCost - Return the global split cost of following the split
+/// pattern in LiveBundles. This cost should be added to the local cost of the
+/// interference pattern in SplitConstraints.
+///
+float RAGreedy::calcGlobalSplitCost(const BitVector &LiveBundles) {
+  float GlobalCost = 0;
+  for (unsigned i = 0, e = SA->LiveBlocks.size(); i != e; ++i) {
+    SplitAnalysis::BlockInfo &BI = SA->LiveBlocks[i];
+    SpillPlacement::BlockConstraint &BC = SplitConstraints[i];
+    bool RegIn  = LiveBundles[Bundles->getBundle(BC.Number, 0)];
+    bool RegOut = LiveBundles[Bundles->getBundle(BC.Number, 1)];
+    unsigned Ins = 0;
+
+    if (!BI.Uses)
+      Ins += RegIn != RegOut;
+    else {
+      if (BI.LiveIn)
+        Ins += RegIn != (BC.Entry == SpillPlacement::PrefReg);
+      if (BI.LiveOut)
+        Ins += RegOut != (BC.Exit == SpillPlacement::PrefReg);
+    }
+    if (Ins)
+      GlobalCost += Ins * SpillPlacer->getBlockFrequency(BC.Number);
+  }
+  return GlobalCost;
+}
+
+/// splitAroundRegion - Split VirtReg around the region determined by
+/// LiveBundles. Make an effort to avoid interference from PhysReg.
+///
+/// The 'register' interval is going to contain as many uses as possible while
+/// avoiding interference. The 'stack' interval is the complement constructed by
+/// SplitEditor. It will contain the rest.
+///
+void RAGreedy::splitAroundRegion(LiveInterval &VirtReg, unsigned PhysReg,
+                                 const BitVector &LiveBundles,
+                                 SmallVectorImpl<LiveInterval*> &NewVRegs) {
+  DEBUG({
+    dbgs() << "Splitting around region for " << PrintReg(PhysReg, TRI)
+           << " with bundles";
+    for (int i = LiveBundles.find_first(); i>=0; i = LiveBundles.find_next(i))
+      dbgs() << " EB#" << i;
+    dbgs() << ".\n";
+  });
+
+  // First compute interference ranges in the live blocks.
+  SmallVector<IndexPair, 8> InterferenceRanges;
+  mapGlobalInterference(PhysReg, InterferenceRanges);
+
+  LiveRangeEdit LREdit(VirtReg, NewVRegs, this);
+  SE->reset(LREdit);
+
+  // Create the main cross-block interval.
+  SE->openIntv();
+
+  // First add all defs that are live out of a block.
+  for (unsigned i = 0, e = SA->LiveBlocks.size(); i != e; ++i) {
+    SplitAnalysis::BlockInfo &BI = SA->LiveBlocks[i];
+    bool RegIn  = LiveBundles[Bundles->getBundle(BI.MBB->getNumber(), 0)];
+    bool RegOut = LiveBundles[Bundles->getBundle(BI.MBB->getNumber(), 1)];
+
+    // Should the register be live out?
+    if (!BI.LiveOut || !RegOut)
+      continue;
+
+    IndexPair &IP = InterferenceRanges[i];
+    DEBUG(dbgs() << "BB#" << BI.MBB->getNumber() << " -> EB#"
+                 << Bundles->getBundle(BI.MBB->getNumber(), 1)
+                 << " [" << BI.Start << ';' << BI.LastSplitPoint << '-'
+                 << BI.Stop << ") intf [" << IP.first << ';' << IP.second
+                 << ')');
 
-    if (!spillInterferences(VirtReg, *PhysRegI, SplitVRegs)) continue;
+    // The interference interval should either be invalid or overlap MBB.
+    assert((!IP.first.isValid() || IP.first < BI.Stop) && "Bad interference");
+    assert((!IP.second.isValid() || IP.second > BI.Start)
+           && "Bad interference");
 
-    assert(checkPhysRegInterference(VirtReg, *PhysRegI) == 0 &&
-           "Interference after spill.");
-    // Tell the caller to allocate to this newly freed physical register.
-    return *PhysRegI;
+    // Check interference leaving the block.
+    if (!IP.second.isValid()) {
+      // Block is interference-free.
+      DEBUG(dbgs() << ", no interference");
+      if (!BI.Uses) {
+        assert(BI.LiveThrough && "No uses, but not live through block?");
+        // Block is live-through without interference.
+        DEBUG(dbgs() << ", no uses"
+                     << (RegIn ? ", live-through.\n" : ", stack in.\n"));
+        if (!RegIn)
+          SE->enterIntvAtEnd(*BI.MBB);
+        continue;
+      }
+      if (!BI.LiveThrough) {
+        DEBUG(dbgs() << ", not live-through.\n");
+        SE->useIntv(SE->enterIntvBefore(BI.Def), BI.Stop);
+        continue;
+      }
+      if (!RegIn) {
+        // Block is live-through, but entry bundle is on the stack.
+        // Reload just before the first use.
+        DEBUG(dbgs() << ", not live-in, enter before first use.\n");
+        SE->useIntv(SE->enterIntvBefore(BI.FirstUse), BI.Stop);
+        continue;
+      }
+      DEBUG(dbgs() << ", live-through.\n");
+      continue;
+    }
+
+    // Block has interference.
+    DEBUG(dbgs() << ", interference to " << IP.second);
+
+    if (!BI.LiveThrough && IP.second <= BI.Def) {
+      // The interference doesn't reach the outgoing segment.
+      DEBUG(dbgs() << " doesn't affect def from " << BI.Def << '\n');
+      SE->useIntv(BI.Def, BI.Stop);
+      continue;
+    }
+
+
+    if (!BI.Uses) {
+      // No uses in block, avoid interference by reloading as late as possible.
+      DEBUG(dbgs() << ", no uses.\n");
+      SlotIndex SegStart = SE->enterIntvAtEnd(*BI.MBB);
+      assert(SegStart >= IP.second && "Couldn't avoid interference");
+      continue;
+    }
+
+    if (IP.second.getBoundaryIndex() < BI.LastUse) {
+      // There are interference-free uses at the end of the block.
+      // Find the first use that can get the live-out register.
+      SmallVectorImpl<SlotIndex>::const_iterator UI =
+        std::lower_bound(SA->UseSlots.begin(), SA->UseSlots.end(),
+                         IP.second.getBoundaryIndex());
+      assert(UI != SA->UseSlots.end() && "Couldn't find last use");
+      SlotIndex Use = *UI;
+      assert(Use <= BI.LastUse && "Couldn't find last use");
+      // Only attempt a split befroe the last split point.
+      if (Use.getBaseIndex() <= BI.LastSplitPoint) {
+        DEBUG(dbgs() << ", free use at " << Use << ".\n");
+        SlotIndex SegStart = SE->enterIntvBefore(Use);
+        assert(SegStart >= IP.second && "Couldn't avoid interference");
+        assert(SegStart < BI.LastSplitPoint && "Impossible split point");
+        SE->useIntv(SegStart, BI.Stop);
+        continue;
+      }
+    }
+
+    // Interference is after the last use.
+    DEBUG(dbgs() << " after last use.\n");
+    SlotIndex SegStart = SE->enterIntvAtEnd(*BI.MBB);
+    assert(SegStart >= IP.second && "Couldn't avoid interference");
   }
-  // No other spill candidates were found, so spill the current VirtReg.
-  DEBUG(dbgs() << "spilling: " << VirtReg << '\n');
-  SmallVector<LiveInterval*, 1> pendingSpills;
 
-  spiller().spill(&VirtReg, SplitVRegs, pendingSpills);
+  // Now all defs leading to live bundles are handled, do everything else.
+  for (unsigned i = 0, e = SA->LiveBlocks.size(); i != e; ++i) {
+    SplitAnalysis::BlockInfo &BI = SA->LiveBlocks[i];
+    bool RegIn  = LiveBundles[Bundles->getBundle(BI.MBB->getNumber(), 0)];
+    bool RegOut = LiveBundles[Bundles->getBundle(BI.MBB->getNumber(), 1)];
+
+    // Is the register live-in?
+    if (!BI.LiveIn || !RegIn)
+      continue;
+
+    // We have an incoming register. Check for interference.
+    IndexPair &IP = InterferenceRanges[i];
+
+    DEBUG(dbgs() << "EB#" << Bundles->getBundle(BI.MBB->getNumber(), 0)
+                 << " -> BB#" << BI.MBB->getNumber() << " [" << BI.Start << ';'
+                 << BI.LastSplitPoint << '-' << BI.Stop << ')');
+
+    // Check interference entering the block.
+    if (!IP.first.isValid()) {
+      // Block is interference-free.
+      DEBUG(dbgs() << ", no interference");
+      if (!BI.Uses) {
+        assert(BI.LiveThrough && "No uses, but not live through block?");
+        // Block is live-through without interference.
+        if (RegOut) {
+          DEBUG(dbgs() << ", no uses, live-through.\n");
+          SE->useIntv(BI.Start, BI.Stop);
+        } else {
+          DEBUG(dbgs() << ", no uses, stack-out.\n");
+          SE->leaveIntvAtTop(*BI.MBB);
+        }
+        continue;
+      }
+      if (!BI.LiveThrough) {
+        DEBUG(dbgs() << ", killed in block.\n");
+        SE->useIntv(BI.Start, SE->leaveIntvAfter(BI.Kill));
+        continue;
+      }
+      if (!RegOut) {
+        // Block is live-through, but exit bundle is on the stack.
+        // Spill immediately after the last use.
+        if (BI.LastUse < BI.LastSplitPoint) {
+          DEBUG(dbgs() << ", uses, stack-out.\n");
+          SE->useIntv(BI.Start, SE->leaveIntvAfter(BI.LastUse));
+          continue;
+        }
+        // The last use is after the last split point, it is probably an
+        // indirect jump.
+        DEBUG(dbgs() << ", uses at " << BI.LastUse << " after split point "
+                     << BI.LastSplitPoint << ", stack-out.\n");
+        SlotIndex SegEnd = SE->leaveIntvBefore(BI.LastSplitPoint);
+        SE->useIntv(BI.Start, SegEnd);
+        // Run a double interval from the split to the last use.
+        // This makes it possible to spill the complement without affecting the
+        // indirect branch.
+        SE->overlapIntv(SegEnd, BI.LastUse);
+        continue;
+      }
+      // Register is live-through.
+      DEBUG(dbgs() << ", uses, live-through.\n");
+      SE->useIntv(BI.Start, BI.Stop);
+      continue;
+    }
+
+    // Block has interference.
+    DEBUG(dbgs() << ", interference from " << IP.first);
+
+    if (!BI.LiveThrough && IP.first >= BI.Kill) {
+      // The interference doesn't reach the outgoing segment.
+      DEBUG(dbgs() << " doesn't affect kill at " << BI.Kill << '\n');
+      SE->useIntv(BI.Start, BI.Kill);
+      continue;
+    }
+
+    if (!BI.Uses) {
+      // No uses in block, avoid interference by spilling as soon as possible.
+      DEBUG(dbgs() << ", no uses.\n");
+      SlotIndex SegEnd = SE->leaveIntvAtTop(*BI.MBB);
+      assert(SegEnd <= IP.first && "Couldn't avoid interference");
+      continue;
+    }
+    if (IP.first.getBaseIndex() > BI.FirstUse) {
+      // There are interference-free uses at the beginning of the block.
+      // Find the last use that can get the register.
+      SmallVectorImpl<SlotIndex>::const_iterator UI =
+        std::lower_bound(SA->UseSlots.begin(), SA->UseSlots.end(),
+                         IP.first.getBaseIndex());
+      assert(UI != SA->UseSlots.begin() && "Couldn't find first use");
+      SlotIndex Use = (--UI)->getBoundaryIndex();
+      DEBUG(dbgs() << ", free use at " << *UI << ".\n");
+      SlotIndex SegEnd = SE->leaveIntvAfter(Use);
+      assert(SegEnd <= IP.first && "Couldn't avoid interference");
+      SE->useIntv(BI.Start, SegEnd);
+      continue;
+    }
+
+    // Interference is before the first use.
+    DEBUG(dbgs() << " before first use.\n");
+    SlotIndex SegEnd = SE->leaveIntvAtTop(*BI.MBB);
+    assert(SegEnd <= IP.first && "Couldn't avoid interference");
+  }
+
+  SE->closeIntv();
+
+  // FIXME: Should we be more aggressive about splitting the stack region into
+  // per-block segments? The current approach allows the stack region to
+  // separate into connected components. Some components may be allocatable.
+  SE->finish();
+  ++NumGlobalSplits;
+
+  if (VerifyEnabled)
+    MF->verify(this, "After splitting live range around region");
+}
+
+unsigned RAGreedy::tryRegionSplit(LiveInterval &VirtReg, AllocationOrder &Order,
+                                  SmallVectorImpl<LiveInterval*> &NewVRegs) {
+  BitVector LiveBundles, BestBundles;
+  float BestCost = 0;
+  unsigned BestReg = 0;
+
+  Order.rewind();
+  for (unsigned Cand = 0; unsigned PhysReg = Order.next(); ++Cand) {
+    if (GlobalCand.size() <= Cand)
+      GlobalCand.resize(Cand+1);
+    GlobalCand[Cand].PhysReg = PhysReg;
+
+    mapGlobalInterference(PhysReg, GlobalCand[Cand].Interference);
+    float Cost = calcSplitConstraints(GlobalCand[Cand].Interference);
+    DEBUG(dbgs() << PrintReg(PhysReg, TRI) << "\tstatic = " << Cost);
+    if (BestReg && Cost >= BestCost) {
+      DEBUG(dbgs() << " higher.\n");
+      continue;
+    }
+
+    SpillPlacer->placeSpills(SplitConstraints, LiveBundles);
+    // No live bundles, defer to splitSingleBlocks().
+    if (!LiveBundles.any()) {
+      DEBUG(dbgs() << " no bundles.\n");
+      continue;
+    }
+
+    Cost += calcGlobalSplitCost(LiveBundles);
+    DEBUG({
+      dbgs() << ", total = " << Cost << " with bundles";
+      for (int i = LiveBundles.find_first(); i>=0; i = LiveBundles.find_next(i))
+        dbgs() << " EB#" << i;
+      dbgs() << ".\n";
+    });
+    if (!BestReg || Cost < BestCost) {
+      BestReg = PhysReg;
+      BestCost = 0.98f * Cost; // Prevent rounding effects.
+      BestBundles.swap(LiveBundles);
+    }
+  }
+
+  if (!BestReg)
+    return 0;
+
+  splitAroundRegion(VirtReg, BestReg, BestBundles, NewVRegs);
+  setStage(NewVRegs.begin(), NewVRegs.end(), RS_Region);
+  return 0;
+}
+
+
+//===----------------------------------------------------------------------===//
+//                             Local Splitting
+//===----------------------------------------------------------------------===//
+
+
+/// calcGapWeights - Compute the maximum spill weight that needs to be evicted
+/// in order to use PhysReg between two entries in SA->UseSlots.
+///
+/// GapWeight[i] represents the gap between UseSlots[i] and UseSlots[i+1].
+///
+void RAGreedy::calcGapWeights(unsigned PhysReg,
+                              SmallVectorImpl<float> &GapWeight) {
+  assert(SA->LiveBlocks.size() == 1 && "Not a local interval");
+  const SplitAnalysis::BlockInfo &BI = SA->LiveBlocks.front();
+  const SmallVectorImpl<SlotIndex> &Uses = SA->UseSlots;
+  const unsigned NumGaps = Uses.size()-1;
+
+  // Start and end points for the interference check.
+  SlotIndex StartIdx = BI.LiveIn ? BI.FirstUse.getBaseIndex() : BI.FirstUse;
+  SlotIndex StopIdx = BI.LiveOut ? BI.LastUse.getBoundaryIndex() : BI.LastUse;
+
+  GapWeight.assign(NumGaps, 0.0f);
+
+  // Add interference from each overlapping register.
+  for (const unsigned *AI = TRI->getOverlaps(PhysReg); *AI; ++AI) {
+    if (!query(const_cast<LiveInterval&>(SA->getParent()), *AI)
+           .checkInterference())
+      continue;
+
+    // We know that VirtReg is a continuous interval from FirstUse to LastUse,
+    // so we don't need InterferenceQuery.
+    //
+    // Interference that overlaps an instruction is counted in both gaps
+    // surrounding the instruction. The exception is interference before
+    // StartIdx and after StopIdx.
+    //
+    LiveIntervalUnion::SegmentIter IntI = PhysReg2LiveUnion[*AI].find(StartIdx);
+    for (unsigned Gap = 0; IntI.valid() && IntI.start() < StopIdx; ++IntI) {
+      // Skip the gaps before IntI.
+      while (Uses[Gap+1].getBoundaryIndex() < IntI.start())
+        if (++Gap == NumGaps)
+          break;
+      if (Gap == NumGaps)
+        break;
+
+      // Update the gaps covered by IntI.
+      const float weight = IntI.value()->weight;
+      for (; Gap != NumGaps; ++Gap) {
+        GapWeight[Gap] = std::max(GapWeight[Gap], weight);
+        if (Uses[Gap+1].getBaseIndex() >= IntI.stop())
+          break;
+      }
+      if (Gap == NumGaps)
+        break;
+    }
+  }
+}
+
+/// getPrevMappedIndex - Return the slot index of the last non-copy instruction
+/// before MI that has a slot index. If MI is the first mapped instruction in
+/// its block, return the block start index instead.
+///
+SlotIndex RAGreedy::getPrevMappedIndex(const MachineInstr *MI) {
+  assert(MI && "Missing MachineInstr");
+  const MachineBasicBlock *MBB = MI->getParent();
+  MachineBasicBlock::const_iterator B = MBB->begin(), I = MI;
+  while (I != B)
+    if (!(--I)->isDebugValue() && !I->isCopy())
+      return Indexes->getInstructionIndex(I);
+  return Indexes->getMBBStartIdx(MBB);
+}
+
+/// calcPrevSlots - Fill in the PrevSlot array with the index of the previous
+/// real non-copy instruction for each instruction in SA->UseSlots.
+///
+void RAGreedy::calcPrevSlots() {
+  const SmallVectorImpl<SlotIndex> &Uses = SA->UseSlots;
+  PrevSlot.clear();
+  PrevSlot.reserve(Uses.size());
+  for (unsigned i = 0, e = Uses.size(); i != e; ++i) {
+    const MachineInstr *MI = Indexes->getInstructionFromIndex(Uses[i]);
+    PrevSlot.push_back(getPrevMappedIndex(MI).getDefIndex());
+  }
+}
+
+/// nextSplitPoint - Find the next index into SA->UseSlots > i such that it may
+/// be beneficial to split before UseSlots[i].
+///
+/// 0 is always a valid split point
+unsigned RAGreedy::nextSplitPoint(unsigned i) {
+  const SmallVectorImpl<SlotIndex> &Uses = SA->UseSlots;
+  const unsigned Size = Uses.size();
+  assert(i != Size && "No split points after the end");
+  // Allow split before i when Uses[i] is not adjacent to the previous use.
+  while (++i != Size && PrevSlot[i].getBaseIndex() <= Uses[i-1].getBaseIndex())
+    ;
+  return i;
+}
+
+/// tryLocalSplit - Try to split VirtReg into smaller intervals inside its only
+/// basic block.
+///
+unsigned RAGreedy::tryLocalSplit(LiveInterval &VirtReg, AllocationOrder &Order,
+                                 SmallVectorImpl<LiveInterval*> &NewVRegs) {
+  assert(SA->LiveBlocks.size() == 1 && "Not a local interval");
+  const SplitAnalysis::BlockInfo &BI = SA->LiveBlocks.front();
+
+  // Note that it is possible to have an interval that is live-in or live-out
+  // while only covering a single block - A phi-def can use undef values from
+  // predecessors, and the block could be a single-block loop.
+  // We don't bother doing anything clever about such a case, we simply assume
+  // that the interval is continuous from FirstUse to LastUse. We should make
+  // sure that we don't do anything illegal to such an interval, though.
+
+  const SmallVectorImpl<SlotIndex> &Uses = SA->UseSlots;
+  if (Uses.size() <= 2)
+    return 0;
+  const unsigned NumGaps = Uses.size()-1;
+
+  DEBUG({
+    dbgs() << "tryLocalSplit: ";
+    for (unsigned i = 0, e = Uses.size(); i != e; ++i)
+      dbgs() << ' ' << SA->UseSlots[i];
+    dbgs() << '\n';
+  });
+
+  // For every use, find the previous mapped non-copy instruction.
+  // We use this to detect valid split points, and to estimate new interval
+  // sizes.
+  calcPrevSlots();
+
+  unsigned BestBefore = NumGaps;
+  unsigned BestAfter = 0;
+  float BestDiff = 0;
+
+  const float blockFreq = SpillPlacer->getBlockFrequency(BI.MBB->getNumber());
+  SmallVector<float, 8> GapWeight;
+
+  Order.rewind();
+  while (unsigned PhysReg = Order.next()) {
+    // Keep track of the largest spill weight that would need to be evicted in
+    // order to make use of PhysReg between UseSlots[i] and UseSlots[i+1].
+    calcGapWeights(PhysReg, GapWeight);
+
+    // Try to find the best sequence of gaps to close.
+    // The new spill weight must be larger than any gap interference.
+
+    // We will split before Uses[SplitBefore] and after Uses[SplitAfter].
+    unsigned SplitBefore = 0, SplitAfter = nextSplitPoint(1) - 1;
+
+    // MaxGap should always be max(GapWeight[SplitBefore..SplitAfter-1]).
+    // It is the spill weight that needs to be evicted.
+    float MaxGap = GapWeight[0];
+    for (unsigned i = 1; i != SplitAfter; ++i)
+      MaxGap = std::max(MaxGap, GapWeight[i]);
+
+    for (;;) {
+      // Live before/after split?
+      const bool LiveBefore = SplitBefore != 0 || BI.LiveIn;
+      const bool LiveAfter = SplitAfter != NumGaps || BI.LiveOut;
+
+      DEBUG(dbgs() << PrintReg(PhysReg, TRI) << ' '
+                   << Uses[SplitBefore] << '-' << Uses[SplitAfter]
+                   << " i=" << MaxGap);
+
+      // Stop before the interval gets so big we wouldn't be making progress.
+      if (!LiveBefore && !LiveAfter) {
+        DEBUG(dbgs() << " all\n");
+        break;
+      }
+      // Should the interval be extended or shrunk?
+      bool Shrink = true;
+      if (MaxGap < HUGE_VALF) {
+        // Estimate the new spill weight.
+        //
+        // Each instruction reads and writes the register, except the first
+        // instr doesn't read when !FirstLive, and the last instr doesn't write
+        // when !LastLive.
+        //
+        // We will be inserting copies before and after, so the total number of
+        // reads and writes is 2 * EstUses.
+        //
+        const unsigned EstUses = 2*(SplitAfter - SplitBefore) +
+                                 2*(LiveBefore + LiveAfter);
+
+        // Try to guess the size of the new interval. This should be trivial,
+        // but the slot index of an inserted copy can be a lot smaller than the
+        // instruction it is inserted before if there are many dead indexes
+        // between them.
+        //
+        // We measure the distance from the instruction before SplitBefore to
+        // get a conservative estimate.
+        //
+        // The final distance can still be different if inserting copies
+        // triggers a slot index renumbering.
+        //
+        const float EstWeight = normalizeSpillWeight(blockFreq * EstUses,
+                              PrevSlot[SplitBefore].distance(Uses[SplitAfter]));
+        // Would this split be possible to allocate?
+        // Never allocate all gaps, we wouldn't be making progress.
+        float Diff = EstWeight - MaxGap;
+        DEBUG(dbgs() << " w=" << EstWeight << " d=" << Diff);
+        if (Diff > 0) {
+          Shrink = false;
+          if (Diff > BestDiff) {
+            DEBUG(dbgs() << " (best)");
+            BestDiff = Diff;
+            BestBefore = SplitBefore;
+            BestAfter = SplitAfter;
+          }
+        }
+      }
+
+      // Try to shrink.
+      if (Shrink) {
+        SplitBefore = nextSplitPoint(SplitBefore);
+        if (SplitBefore < SplitAfter) {
+          DEBUG(dbgs() << " shrink\n");
+          // Recompute the max when necessary.
+          if (GapWeight[SplitBefore - 1] >= MaxGap) {
+            MaxGap = GapWeight[SplitBefore];
+            for (unsigned i = SplitBefore + 1; i != SplitAfter; ++i)
+              MaxGap = std::max(MaxGap, GapWeight[i]);
+          }
+          continue;
+        }
+        MaxGap = 0;
+      }
+
+      // Try to extend the interval.
+      if (SplitAfter >= NumGaps) {
+        DEBUG(dbgs() << " end\n");
+        break;
+      }
+
+      DEBUG(dbgs() << " extend\n");
+      for (unsigned e = nextSplitPoint(SplitAfter + 1) - 1;
+           SplitAfter != e; ++SplitAfter)
+        MaxGap = std::max(MaxGap, GapWeight[SplitAfter]);
+          continue;
+    }
+  }
+
+  // Didn't find any candidates?
+  if (BestBefore == NumGaps)
+    return 0;
+
+  DEBUG(dbgs() << "Best local split range: " << Uses[BestBefore]
+               << '-' << Uses[BestAfter] << ", " << BestDiff
+               << ", " << (BestAfter - BestBefore + 1) << " instrs\n");
+
+  LiveRangeEdit LREdit(VirtReg, NewVRegs, this);
+  SE->reset(LREdit);
+
+  SE->openIntv();
+  SlotIndex SegStart = SE->enterIntvBefore(Uses[BestBefore]);
+  SlotIndex SegStop  = SE->leaveIntvAfter(Uses[BestAfter]);
+  SE->useIntv(SegStart, SegStop);
+  SE->closeIntv();
+  SE->finish();
+  setStage(NewVRegs.begin(), NewVRegs.end(), RS_Local);
+  ++NumLocalSplits;
+
+  return 0;
+}
+
+//===----------------------------------------------------------------------===//
+//                          Live Range Splitting
+//===----------------------------------------------------------------------===//
+
+/// trySplit - Try to split VirtReg or one of its interferences, making it
+/// assignable.
+/// @return Physreg when VirtReg may be assigned and/or new NewVRegs.
+unsigned RAGreedy::trySplit(LiveInterval &VirtReg, AllocationOrder &Order,
+                            SmallVectorImpl<LiveInterval*>&NewVRegs) {
+  // Local intervals are handled separately.
+  if (LIS->intervalIsInOneMBB(VirtReg)) {
+    NamedRegionTimer T("Local Splitting", TimerGroupName, TimePassesIsEnabled);
+    SA->analyze(&VirtReg);
+    return tryLocalSplit(VirtReg, Order, NewVRegs);
+  }
+
+  NamedRegionTimer T("Global Splitting", TimerGroupName, TimePassesIsEnabled);
+
+  // Don't iterate global splitting.
+  // Move straight to spilling if this range was produced by a global split.
+  LiveRangeStage Stage = getStage(VirtReg);
+  if (Stage >= RS_Block)
+    return 0;
+
+  SA->analyze(&VirtReg);
+
+  // First try to split around a region spanning multiple blocks.
+  if (Stage < RS_Region) {
+    unsigned PhysReg = tryRegionSplit(VirtReg, Order, NewVRegs);
+    if (PhysReg || !NewVRegs.empty())
+      return PhysReg;
+  }
+
+  // Then isolate blocks with multiple uses.
+  if (Stage < RS_Block) {
+    SplitAnalysis::BlockPtrSet Blocks;
+    if (SA->getMultiUseBlocks(Blocks)) {
+      LiveRangeEdit LREdit(VirtReg, NewVRegs, this);
+      SE->reset(LREdit);
+      SE->splitSingleBlocks(Blocks);
+      setStage(NewVRegs.begin(), NewVRegs.end(), RS_Block);
+      if (VerifyEnabled)
+        MF->verify(this, "After splitting live range around basic blocks");
+    }
+  }
+
+  // Don't assign any physregs.
+  return 0;
+}
+
+
+//===----------------------------------------------------------------------===//
+//                            Main Entry Point
+//===----------------------------------------------------------------------===//
+
+unsigned RAGreedy::selectOrSplit(LiveInterval &VirtReg,
+                                 SmallVectorImpl<LiveInterval*> &NewVRegs) {
+  // First try assigning a free register.
+  AllocationOrder Order(VirtReg.reg, *VRM, ReservedRegs);
+  while (unsigned PhysReg = Order.next()) {
+    if (!checkPhysRegInterference(VirtReg, PhysReg))
+      return PhysReg;
+  }
+
+  if (unsigned PhysReg = tryEvict(VirtReg, Order, NewVRegs))
+    return PhysReg;
+
+  assert(NewVRegs.empty() && "Cannot append to existing NewVRegs");
+
+  // The first time we see a live range, don't try to split or spill.
+  // Wait until the second time, when all smaller ranges have been allocated.
+  // This gives a better picture of the interference to split around.
+  LiveRangeStage Stage = getStage(VirtReg);
+  if (Stage == RS_Original) {
+    LRStage[VirtReg.reg] = RS_Second;
+    DEBUG(dbgs() << "wait for second round\n");
+    NewVRegs.push_back(&VirtReg);
+    return 0;
+  }
+
+  assert(Stage < RS_Spill && "Cannot allocate after spilling");
+
+  // Try splitting VirtReg or interferences.
+  unsigned PhysReg = trySplit(VirtReg, Order, NewVRegs);
+  if (PhysReg || !NewVRegs.empty())
+    return PhysReg;
+
+  // Finally spill VirtReg itself.
+  NamedRegionTimer T("Spiller", TimerGroupName, TimePassesIsEnabled);
+  LiveRangeEdit LRE(VirtReg, NewVRegs, this);
+  spiller().spill(LRE);
+
+  if (VerifyEnabled)
+    MF->verify(this, "After spilling");
 
   // The live virtual register requesting allocation was spilled, so tell
   // the caller not to allocate anything during this round.
@@ -188,25 +1179,36 @@ bool RAGreedy::runOnMachineFunction(MachineFunction &mf) {
                << ((Value*)mf.getFunction())->getName() << '\n');
 
   MF = &mf;
-  TM = &mf.getTarget();
-  MRI = &mf.getRegInfo();
-
-  const TargetRegisterInfo *TRI = TM->getRegisterInfo();
-  RegAllocBase::init(*TRI, getAnalysis<VirtRegMap>(),
-                     getAnalysis<LiveIntervals>());
+  if (VerifyEnabled)
+    MF->verify(this, "Before greedy register allocator");
 
+  RegAllocBase::init(getAnalysis<VirtRegMap>(), getAnalysis<LiveIntervals>());
+  Indexes = &getAnalysis<SlotIndexes>();
+  DomTree = &getAnalysis<MachineDominatorTree>();
   ReservedRegs = TRI->getReservedRegs(*MF);
-  SpillerInstance.reset(createSpiller(*this, *MF, *VRM));
+  SpillerInstance.reset(createInlineSpiller(*this, *MF, *VRM));
+  Loops = &getAnalysis<MachineLoopInfo>();
+  LoopRanges = &getAnalysis<MachineLoopRanges>();
+  Bundles = &getAnalysis<EdgeBundles>();
+  SpillPlacer = &getAnalysis<SpillPlacement>();
+
+  SA.reset(new SplitAnalysis(*VRM, *LIS, *Loops));
+  SE.reset(new SplitEditor(*SA, *LIS, *VRM, *DomTree));
+  LRStage.clear();
+  LRStage.resize(MRI->getNumVirtRegs());
+
   allocatePhysRegs();
   addMBBLiveIns(MF);
+  LIS->addKillFlags();
 
   // Run rewriter
-  std::auto_ptr<VirtRegRewriter> rewriter(createVirtRegRewriter());
-  rewriter->runOnMachineFunction(*MF, *VRM, LIS);
+  {
+    NamedRegionTimer T("Rewriter", TimerGroupName, TimePassesIsEnabled);
+    VRM->rewrite(Indexes);
+  }
 
   // The pass output is in VirtRegMap. Release all the transient data.
   releaseMemory();
 
   return true;
 }
-