Implement RAGreedy::splitAroundRegion and remove loop splitting.
[oota-llvm.git] / lib / CodeGen / RegAllocGreedy.cpp
index 9ba67d9c642f233df2d53babaf8f5701b6601e24..fa3b5871fcefe4a87fc291868a21b1f7da40581a 100644 (file)
@@ -18,6 +18,7 @@
 #include "LiveRangeEdit.h"
 #include "RegAllocBase.h"
 #include "Spiller.h"
+#include "SpillPlacement.h"
 #include "SplitKit.h"
 #include "VirtRegMap.h"
 #include "VirtRegRewriter.h"
@@ -25,6 +26,7 @@
 #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"
@@ -53,15 +55,53 @@ class RAGreedy : public MachineFunctionPass, public RegAllocBase {
   BitVector ReservedRegs;
 
   // analyses
+  SlotIndexes *Indexes;
   LiveStacks *LS;
   MachineDominatorTree *DomTree;
   MachineLoopInfo *Loops;
   MachineLoopRanges *LoopRanges;
+  EdgeBundles *Bundles;
+  SpillPlacement *SpillPlacer;
 
   // state
   std::auto_ptr<Spiller> SpillerInstance;
   std::auto_ptr<SplitAnalysis> SA;
 
+  // splitting state.
+
+  /// All basic blocks where the current register is live.
+  SmallVector<SpillPlacement::BlockConstraint, 8> SpillConstraints;
+
+  /// Additional information about basic blocks where the current variable is
+  /// live. Such a block will look like one of these templates:
+  ///
+  ///  1. |   o---x   | Internal to block. Variable is only live in this block.
+  ///  2. |---x       | Live-in, kill.
+  ///  3. |       o---| Def, live-out.
+  ///  4. |---x   o---| Live-in, kill, def, live-out.
+  ///  5. |---o---o---| Live-through with uses or defs.
+  ///  6. |-----------| Live-through without uses. Transparent.
+  ///
+  struct BlockInfo {
+    MachineBasicBlock *MBB;
+    SlotIndex FirstUse;   ///< First instr using current reg.
+    SlotIndex LastUse;    ///< Last instr using current reg.
+    SlotIndex Kill;       ///< Interval end point inside block.
+    SlotIndex Def;        ///< Interval start point inside block.
+    bool Uses;            ///< Current reg has uses or defs in block.
+    bool LiveThrough;     ///< Live in whole block (Templ 5. or 6. above).
+    bool LiveIn;          ///< Current reg is live in.
+    bool LiveOut;         ///< Current reg is live out.
+
+    // Per-interference pattern scratch data.
+    bool OverlapEntry;    ///< Interference overlaps entering interval.
+    bool OverlapExit;     ///< Interference overlaps exiting interval.
+  };
+
+  /// Basic blocks where var is live. This array is parallel to
+  /// SpillConstraints.
+  SmallVector<BlockInfo, 8> LiveBlocks;
+
 public:
   RAGreedy();
 
@@ -79,8 +119,8 @@ public:
 
   virtual float getPriority(LiveInterval *LI);
 
-  virtual unsigned selectOrSplit(LiveInterval &VirtReg,
-                                 SmallVectorImpl<LiveInterval*> &SplitVRegs);
+  virtual unsigned selectOrSplit(LiveInterval&,
+                                 SmallVectorImpl<LiveInterval*>&);
 
   /// Perform register allocation.
   virtual bool runOnMachineFunction(MachineFunction &mf);
@@ -92,12 +132,20 @@ private:
   LiveInterval *getSingleInterference(LiveInterval&, unsigned);
   bool reassignVReg(LiveInterval &InterferingVReg, unsigned OldPhysReg);
   bool reassignInterferences(LiveInterval &VirtReg, unsigned PhysReg);
-  unsigned findInterferenceFreeReg(MachineLoopRange*,
-                                   LiveInterval&, AllocationOrder&);
+  float calcInterferenceWeight(LiveInterval&, unsigned);
+  void calcLiveBlockInfo(LiveInterval&);
+  float calcInterferenceInfo(LiveInterval&, unsigned);
+  float calcGlobalSplitCost(const BitVector&);
+  void splitAroundRegion(LiveInterval&, unsigned, const BitVector&,
+                         SmallVectorImpl<LiveInterval*>&);
 
   unsigned tryReassign(LiveInterval&, AllocationOrder&);
+  unsigned tryRegionSplit(LiveInterval&, AllocationOrder&,
+                          SmallVectorImpl<LiveInterval*>&);
   unsigned trySplit(LiveInterval&, AllocationOrder&,
                     SmallVectorImpl<LiveInterval*>&);
+  unsigned trySpillInterferences(LiveInterval&, AllocationOrder&,
+                                 SmallVectorImpl<LiveInterval*>&);
 };
 } // end anonymous namespace
 
@@ -108,6 +156,7 @@ FunctionPass* llvm::createGreedyRegisterAllocator() {
 }
 
 RAGreedy::RAGreedy(): MachineFunctionPass(ID) {
+  initializeSlotIndexesPass(*PassRegistry::getPassRegistry());
   initializeLiveIntervalsPass(*PassRegistry::getPassRegistry());
   initializeSlotIndexesPass(*PassRegistry::getPassRegistry());
   initializeStrongPHIEliminationPass(*PassRegistry::getPassRegistry());
@@ -118,6 +167,8 @@ RAGreedy::RAGreedy(): MachineFunctionPass(ID) {
   initializeMachineLoopInfoPass(*PassRegistry::getPassRegistry());
   initializeMachineLoopRangesPass(*PassRegistry::getPassRegistry());
   initializeVirtRegMapPass(*PassRegistry::getPassRegistry());
+  initializeEdgeBundlesPass(*PassRegistry::getPassRegistry());
+  initializeSpillPlacementPass(*PassRegistry::getPassRegistry());
 }
 
 void RAGreedy::getAnalysisUsage(AnalysisUsage &AU) const {
@@ -125,6 +176,7 @@ 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);
@@ -140,6 +192,8 @@ void RAGreedy::getAnalysisUsage(AnalysisUsage &AU) const {
   AU.addPreserved<MachineLoopRanges>();
   AU.addRequired<VirtRegMap>();
   AU.addPreserved<VirtRegMap>();
+  AU.addRequired<EdgeBundles>();
+  AU.addRequired<SpillPlacement>();
   MachineFunctionPass::getAnalysisUsage(AU);
 }
 
@@ -164,6 +218,11 @@ float RAGreedy::getPriority(LiveInterval *LI) {
   return Priority;
 }
 
+
+//===----------------------------------------------------------------------===//
+//                         Register Reassignment
+//===----------------------------------------------------------------------===//
+
 // Check interference without using the cache.
 bool RAGreedy::checkUncachedInterference(LiveInterval &VirtReg,
                                          unsigned PhysReg) {
@@ -263,137 +322,658 @@ unsigned RAGreedy::tryReassign(LiveInterval &VirtReg, AllocationOrder &Order) {
   return 0;
 }
 
-/// findInterferenceFreeReg - Find a physical register in Order where Loop has
-/// no interferences with VirtReg.
-unsigned RAGreedy::findInterferenceFreeReg(MachineLoopRange *Loop,
-                                           LiveInterval &VirtReg,
-                                           AllocationOrder &Order) {
-  Order.rewind();
-  while (unsigned PhysReg = Order.next()) {
-    bool interference = false;
-    for (const unsigned *AI = TRI->getOverlaps(PhysReg); *AI; ++AI) {
-      if (query(VirtReg, *AI).checkLoopInterference(Loop)) {
-        interference = true;
+
+//===----------------------------------------------------------------------===//
+//                              Region Splitting
+//===----------------------------------------------------------------------===//
+
+/// calcLiveBlockInfo - Fill the LiveBlocks array with information about blocks
+/// where VirtReg is live.
+/// The SpillConstraints array is minimally initialized with MBB->getNumber().
+void RAGreedy::calcLiveBlockInfo(LiveInterval &VirtReg) {
+  LiveBlocks.clear();
+  SpillConstraints.clear();
+
+  assert(!VirtReg.empty() && "Cannot allocate an empty interval");
+  LiveInterval::const_iterator LVI = VirtReg.begin();
+  LiveInterval::const_iterator LVE = VirtReg.end();
+
+  SmallVectorImpl<SlotIndex>::const_iterator UseI, UseE;
+  UseI = SA->UseSlots.begin();
+  UseE = SA->UseSlots.end();
+
+  // Loop over basic blocks where VirtReg is live.
+  MachineFunction::iterator MFI = Indexes->getMBBFromIndex(LVI->start);
+  for (;;) {
+    // Block constraints depend on the interference pattern.
+    // Just allocate them here, don't compute anything.
+    SpillPlacement::BlockConstraint BC;
+    BC.Number = MFI->getNumber();
+    SpillConstraints.push_back(BC);
+
+    BlockInfo BI;
+    BI.MBB = MFI;
+    SlotIndex Start, Stop;
+    tie(Start, Stop) = Indexes->getMBBRange(BI.MBB);
+
+    // LVI is the first live segment overlapping MBB.
+    BI.LiveIn = LVI->start <= Start;
+    if (!BI.LiveIn)
+      BI.Def = LVI->start;
+
+    // Find the first and last uses in the block.
+    BI.Uses = SA->hasUses(MFI);
+    if (BI.Uses && UseI != UseE) {
+      BI.FirstUse = *UseI;
+      assert(BI.FirstUse >= Start);
+      do ++UseI;
+      while (UseI != UseE && *UseI < Stop);
+      BI.LastUse = UseI[-1];
+      assert(BI.LastUse < Stop);
+    }
+
+    // Look for gaps in the live range.
+    bool hasGap = false;
+    BI.LiveOut = true;
+    while (LVI->end < Stop) {
+      SlotIndex LastStop = LVI->end;
+      if (++LVI == LVE || LVI->start >= Stop) {
+        BI.Kill = LastStop;
+        BI.LiveOut = false;
         break;
       }
+      if (LastStop < LVI->start) {
+        hasGap = true;
+        BI.Kill = LastStop;
+        BI.Def = LVI->start;
+      }
     }
-    if (!interference)
-      return PhysReg;
+
+    // Don't set LiveThrough when the block has a gap.
+    BI.LiveThrough = !hasGap && BI.LiveIn && BI.LiveOut;
+    LiveBlocks.push_back(BI);
+
+    // LVI is now at LVE or LVI->end >= Stop.
+    if (LVI == LVE)
+      break;
+
+    // Live segment ends exactly at Stop. Move to the next segment.
+    if (LVI->end == Stop && ++LVI == LVE)
+      break;
+
+    // Pick the next basic block.
+    if (LVI->start < Stop)
+      ++MFI;
+    else
+      MFI = Indexes->getMBBFromIndex(LVI->start);
+  }
+}
+
+/// calcInterferenceInfo - Compute per-block outgoing and ingoing constraints
+/// when considering interference from PhysReg. Also compute an optimistic local
+/// cost of this interference pattern.
+///
+/// The final cost of a split is the local cost + global cost of preferences
+/// broken by SpillPlacement.
+///
+float RAGreedy::calcInterferenceInfo(LiveInterval &VirtReg, unsigned PhysReg) {
+  // Reset interference dependent info.
+  for (unsigned i = 0, e = LiveBlocks.size(); i != e; ++i) {
+    BlockInfo &BI = LiveBlocks[i];
+    SpillPlacement::BlockConstraint &BC = SpillConstraints[i];
+    BC.Entry = (BI.Uses && BI.LiveIn) ?
+      SpillPlacement::PrefReg : SpillPlacement::DontCare;
+    BC.Exit = (BI.Uses && BI.LiveOut) ?
+      SpillPlacement::PrefReg : SpillPlacement::DontCare;
+    BI.OverlapEntry = BI.OverlapExit = false;
+  }
+
+  // Add interference info from each PhysReg alias.
+  for (const unsigned *AI = TRI->getOverlaps(PhysReg); *AI; ++AI) {
+    if (!query(VirtReg, *AI).checkInterference())
+      continue;
+    DEBUG(PhysReg2LiveUnion[*AI].print(dbgs(), TRI));
+    LiveIntervalUnion::SegmentIter IntI =
+      PhysReg2LiveUnion[*AI].find(VirtReg.beginIndex());
+    if (!IntI.valid())
+      continue;
+
+    for (unsigned i = 0, e = LiveBlocks.size(); i != e; ++i) {
+      BlockInfo &BI = LiveBlocks[i];
+      SpillPlacement::BlockConstraint &BC = SpillConstraints[i];
+      SlotIndex Start, Stop;
+      tie(Start, Stop) = Indexes->getMBBRange(BI.MBB);
+
+      // Skip interference-free blocks.
+      if (IntI.start() >= Stop)
+        continue;
+
+      // Handle transparent blocks with interference separately.
+      // Transparent blocks never incur any fixed cost.
+      if (BI.LiveThrough && !BI.Uses) {
+        // Check if interference is live-in - force spill.
+        if (BC.Entry != SpillPlacement::MustSpill) {
+          BC.Entry = SpillPlacement::PrefSpill;
+          IntI.advanceTo(Start);
+          if (IntI.valid() && IntI.start() <= Start)
+            BC.Entry = SpillPlacement::MustSpill;
+        }
+
+        // Check if interference is live-out - force spill.
+        if (BC.Exit != SpillPlacement::MustSpill) {
+          BC.Exit = SpillPlacement::PrefSpill;
+          IntI.advanceTo(Stop);
+          if (IntI.valid() && IntI.start() < Stop)
+            BC.Exit = SpillPlacement::MustSpill;
+        }
+
+        // Nothing more to do for this transparent block.
+        if (!IntI.valid())
+          break;
+        continue;
+      }
+
+      // Now we only have blocks with uses left.
+      // Check if the interference overlaps the uses.
+      assert(BI.Uses && "Non-transparent block without any uses");
+
+      // Check interference on entry.
+      if (BI.LiveIn && BC.Entry != SpillPlacement::MustSpill) {
+        IntI.advanceTo(Start);
+        if (!IntI.valid())
+          break;
+
+        // Interference is live-in - force spill.
+        if (IntI.start() <= Start)
+          BC.Entry = SpillPlacement::MustSpill;
+        // Not live in, but before the first use.
+        else if (IntI.start() < BI.FirstUse)
+          BC.Entry = SpillPlacement::PrefSpill;
+      }
+
+      // Does interference overlap the uses in the entry segment
+      // [FirstUse;Kill)?
+      if (BI.LiveIn && !BI.OverlapEntry) {
+        IntI.advanceTo(BI.FirstUse);
+        if (!IntI.valid())
+          break;
+        // A live-through interval has no kill.
+        // Check [FirstUse;LastUse) instead.
+        if (IntI.start() < (BI.LiveThrough ? BI.LastUse : BI.Kill))
+          BI.OverlapEntry = true;
+      }
+
+      // Does interference overlap the uses in the exit segment [Def;LastUse)?
+      if (BI.LiveOut && !BI.LiveThrough && !BI.OverlapExit) {
+        IntI.advanceTo(BI.Def);
+        if (!IntI.valid())
+          break;
+        if (IntI.start() < BI.LastUse)
+          BI.OverlapExit = true;
+      }
+
+      // Check interference on exit.
+      if (BI.LiveOut && BC.Exit != SpillPlacement::MustSpill) {
+        // Check interference between LastUse and Stop.
+        if (BC.Exit != SpillPlacement::PrefSpill) {
+          IntI.advanceTo(BI.LastUse);
+          if (!IntI.valid())
+            break;
+          if (IntI.start() < Stop)
+            BC.Exit = SpillPlacement::PrefSpill;
+        }
+        // Is the interference live-out?
+        IntI.advanceTo(Stop);
+        if (!IntI.valid())
+          break;
+        if (IntI.start() < Stop)
+          BC.Exit = SpillPlacement::MustSpill;
+      }
+    }
+  }
+
+  // Accumulate a local cost of this interference pattern.
+  float LocalCost = 0;
+  for (unsigned i = 0, e = LiveBlocks.size(); i != e; ++i) {
+    BlockInfo &BI = LiveBlocks[i];
+    if (!BI.Uses)
+      continue;
+    SpillPlacement::BlockConstraint &BC = SpillConstraints[i];
+    unsigned Inserts = 0;
+
+    // Do we need spill code for the entry segment?
+    if (BI.LiveIn)
+      Inserts += BI.OverlapEntry || BC.Entry != SpillPlacement::PrefReg;
+
+    // For the exit segment?
+    if (BI.LiveOut)
+      Inserts += BI.OverlapExit || BC.Exit != SpillPlacement::PrefReg;
+
+    // The local cost of spill code in this block is the block frequency times
+    // the number of spill instructions inserted.
+    if (Inserts)
+      LocalCost += Inserts * SpillPlacer->getBlockFrequency(BI.MBB);
+  }
+  DEBUG(dbgs() << "Local cost of " << PrintReg(PhysReg, TRI) << " = "
+               << LocalCost << '\n');
+  return LocalCost;
+}
+
+/// 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 SpillConstraints.
+///
+float RAGreedy::calcGlobalSplitCost(const BitVector &LiveBundles) {
+  float GlobalCost = 0;
+  for (unsigned i = 0, e = LiveBlocks.size(); i != e; ++i) {
+    SpillPlacement::BlockConstraint &BC = SpillConstraints[i];
+    unsigned Inserts = 0;
+    // Broken entry preference?
+    Inserts += LiveBundles[Bundles->getBundle(BC.Number, 0)] !=
+                 (BC.Entry == SpillPlacement::PrefReg);
+    // Broken exit preference?
+    Inserts += LiveBundles[Bundles->getBundle(BC.Number, 1)] !=
+                 (BC.Exit == SpillPlacement::PrefReg);
+    if (Inserts)
+      GlobalCost += Inserts * SpillPlacer->getBlockFrequency(LiveBlocks[i].MBB);
+  }
+  DEBUG(dbgs() << "Global cost = " << GlobalCost << '\n');
+  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.
+  typedef std::pair<SlotIndex, SlotIndex> IndexPair;
+  SmallVector<IndexPair, 8> InterferenceRanges;
+  InterferenceRanges.resize(LiveBlocks.size());
+  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 = LiveBlocks.size(); i != e; ++i) {
+      BlockInfo &BI = LiveBlocks[i];
+      if (!BI.Uses)
+        continue;
+      IndexPair &IP = InterferenceRanges[i];
+      SlotIndex Start, Stop;
+      tie(Start, Stop) = Indexes->getMBBRange(BI.MBB);
+      // Skip interference-free blocks.
+      if (IntI.start() >= Stop)
+        continue;
+
+      // First interference in block.
+      if (BI.LiveIn) {
+        IntI.advanceTo(Start);
+        if (!IntI.valid())
+          break;
+        if (!IP.first.isValid() || IntI.start() < IP.first)
+          IP.first = IntI.start();
+      }
+
+      // Last interference in block.
+      if (BI.LiveOut) {
+        IntI.advanceTo(Stop);
+        if (!IntI.valid() || IntI.start() >= Stop)
+          --IntI;
+        if (!IP.second.isValid() || IntI.stop() > IP.second)
+          IP.second = IntI.stop();
+      }
+    }
+  }
+
+  SmallVector<LiveInterval*, 4> SpillRegs;
+  LiveRangeEdit LREdit(VirtReg, NewVRegs, SpillRegs);
+  SplitEditor SE(*SA, *LIS, *VRM, *DomTree, 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 = LiveBlocks.size(); i != e; ++i) {
+    BlockInfo &BI = 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];
+    SlotIndex Start, Stop;
+    tie(Start, Stop) = Indexes->getMBBRange(BI.MBB);
+
+    DEBUG(dbgs() << "BB#" << BI.MBB->getNumber() << " -> EB#"
+                 << Bundles->getBundle(BI.MBB->getNumber(), 1));
+
+    // Check interference leaving the block.
+    if (!IP.second.isValid() || IP.second < Start) {
+      // 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.enterIntvBefore(BI.Def);
+        SE.useIntv(BI.Def, 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.enterIntvBefore(BI.FirstUse);
+        SE.useIntv(BI.FirstUse, Stop);
+        continue;
+      }
+      DEBUG(dbgs() << ", live-through.\n");
+      continue;
+    }
+
+    // Block has interference.
+    DEBUG(dbgs() << ", interference to " << IP.second);
+    if (!BI.Uses) {
+      // No uses in block, avoid interference by reloading as late as possible.
+      DEBUG(dbgs() << ", no uses.\n");
+      SE.enterIntvAtEnd(*BI.MBB);
+      continue;
+    }
+    if (IP.second < BI.LastUse) {
+      // There are interference-free uses at the end of the block.
+      // Find the first use that can get the live-out register.
+      SlotIndex Use = *std::lower_bound(SA->UseSlots.begin(), SA->UseSlots.end(),
+                                        IP.second);
+      DEBUG(dbgs() << ", free use at " << Use << ".\n");
+      assert(Use > IP.second && Use <= BI.LastUse);
+      SE.enterIntvBefore(Use);
+      SE.useIntv(Use, Stop);
+      continue;
+    }
+
+    // Interference is after the last use.
+    DEBUG(dbgs() << " after last use.\n");
+    SE.enterIntvAtEnd(*BI.MBB);
   }
-  // No physreg found.
+
+  // Now all defs leading to live bundles are handled, do everything else.
+  for (unsigned i = 0, e = LiveBlocks.size(); i != e; ++i) {
+    BlockInfo &BI = 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];
+    SlotIndex Start, Stop;
+    tie(Start, Stop) = Indexes->getMBBRange(BI.MBB);
+
+    DEBUG(dbgs() << "EB#" << Bundles->getBundle(BI.MBB->getNumber(), 0)
+                 << " -> BB#" << BI.MBB->getNumber());
+
+    // Check interference entering the block.
+    if (!IP.first.isValid() || IP.first > Stop) {
+      // 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(Start, 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(Start, BI.Kill);
+        SE.leaveIntvAfter(BI.Kill);
+        continue;
+      }
+      if (!RegOut) {
+        // Block is live-through, but exit bundle is on the stack.
+        // Spill immediately after the last use.
+        DEBUG(dbgs() << ", uses, stack-out.\n");
+        SE.useIntv(Start, BI.LastUse);
+        SE.leaveIntvAfter(BI.LastUse);
+        continue;
+      }
+      // Register is live-through.
+      DEBUG(dbgs() << ", uses, live-through.\n");
+      SE.useIntv(Start, Stop);
+      continue;
+    }
+
+    // Block has interference.
+    DEBUG(dbgs() << ", interference from " << IP.first);
+    if (!BI.Uses) {
+      // No uses in block, avoid interference by spilling as soon as possible.
+      DEBUG(dbgs() << ", no uses.\n");
+      SE.leaveIntvAtTop(*BI.MBB);
+      continue;
+    }
+    if (IP.first > BI.FirstUse) {
+      // There are interference-free uses at the beginning of the block.
+      // Find the last use that can get the register.
+      SlotIndex Use = std::lower_bound(SA->UseSlots.begin(), SA->UseSlots.end(),
+                                       IP.second)[-1];
+      DEBUG(dbgs() << ", free use at " << Use << ".\n");
+      assert(Use >= BI.FirstUse && Use < IP.first);
+      SE.useIntv(Start, Use);
+      SE.leaveIntvAfter(Use);
+      continue;
+    }
+
+    // Interference is before the first use.
+    DEBUG(dbgs() << " before first use.\n");
+    SE.leaveIntvAtTop(*BI.MBB);
+  }
+
+  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();
+
+  if (VerifyEnabled)
+    MF->verify(this, "After splitting live range around region");
+}
+
+unsigned RAGreedy::tryRegionSplit(LiveInterval &VirtReg, AllocationOrder &Order,
+                                  SmallVectorImpl<LiveInterval*> &NewVRegs) {
+  calcLiveBlockInfo(VirtReg);
+  BitVector LiveBundles, BestBundles;
+  float BestCost = 0;
+  unsigned BestReg = 0;
+  Order.rewind();
+  while (unsigned PhysReg = Order.next()) {
+    float Cost = calcInterferenceInfo(VirtReg, PhysReg);
+    if (BestReg && Cost >= BestCost)
+      continue;
+
+    SpillPlacer->placeSpills(SpillConstraints, LiveBundles);
+    // No live bundles, defer to splitSingleBlocks().
+    if (!LiveBundles.any())
+      continue;
+
+    Cost += calcGlobalSplitCost(LiveBundles);
+    if (!BestReg || Cost < BestCost) {
+      BestReg = PhysReg;
+      BestCost = Cost;
+      BestBundles.swap(LiveBundles);
+    }
+  }
+
+  if (!BestReg)
+    return 0;
+
+  splitAroundRegion(VirtReg, BestReg, BestBundles, NewVRegs);
   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 SplitVRegs.
+/// @return Physreg when VirtReg may be assigned and/or new NewVRegs.
 unsigned RAGreedy::trySplit(LiveInterval &VirtReg, AllocationOrder &Order,
-                            SmallVectorImpl<LiveInterval*>&SplitVRegs) {
+                            SmallVectorImpl<LiveInterval*>&NewVRegs) {
   NamedRegionTimer T("Splitter", TimerGroupName, TimePassesIsEnabled);
   SA->analyze(&VirtReg);
 
-  // Get the set of loops that have VirtReg uses and are splittable.
-  SplitAnalysis::LoopPtrSet SplitLoopSet;
-  SA->getSplitLoops(SplitLoopSet);
-
-  // Order loops by descending area.
-  SmallVector<MachineLoopRange*, 8> SplitLoops;
-  for (SplitAnalysis::LoopPtrSet::const_iterator I = SplitLoopSet.begin(),
-         E = SplitLoopSet.end(); I != E; ++I)
-    SplitLoops.push_back(LoopRanges->getLoopRange(*I));
-  array_pod_sort(SplitLoops.begin(), SplitLoops.end(),
-                 MachineLoopRange::byAreaDesc);
-
-  // Find the first loop that is interference-free for some register in the
-  // allocation order.
-  MachineLoopRange *Loop = 0;
-  for (unsigned i = 0, e = SplitLoops.size(); i != e; ++i) {
-    DEBUG(dbgs() << "  Checking " << *SplitLoops[i]);
-    if (unsigned PhysReg = findInterferenceFreeReg(SplitLoops[i],
-                                                   VirtReg, Order)) {
-      (void)PhysReg;
-      Loop = SplitLoops[i];
-      DEBUG(dbgs() << ": Use %" << TRI->getName(PhysReg) << '\n');
-      break;
-    } else {
-      DEBUG(dbgs() << ": Interference.\n");
-    }
+  // Don't attempt splitting on local intervals for now. TBD.
+  if (LIS->intervalIsInOneMBB(VirtReg))
+    return 0;
+
+  // First try to split around a region spanning multiple blocks.
+  unsigned PhysReg = tryRegionSplit(VirtReg, Order, NewVRegs);
+  if (PhysReg || !NewVRegs.empty())
+    return PhysReg;
+
+  // Then isolate blocks with multiple uses.
+  SplitAnalysis::BlockPtrSet Blocks;
+  if (SA->getMultiUseBlocks(Blocks)) {
+    SmallVector<LiveInterval*, 4> SpillRegs;
+    LiveRangeEdit LREdit(VirtReg, NewVRegs, SpillRegs);
+    SplitEditor(*SA, *LIS, *VRM, *DomTree, LREdit).splitSingleBlocks(Blocks);
   }
 
-  if (!Loop) {
-    DEBUG(dbgs() << "  All candidate loops have interference.\n");
-    return 0;
+  // Don't assign any physregs.
+  return 0;
+}
+
+
+//===----------------------------------------------------------------------===//
+//                                Spilling
+//===----------------------------------------------------------------------===//
+
+/// calcInterferenceWeight - Calculate the combined spill weight of
+/// interferences when assigning VirtReg to PhysReg.
+float RAGreedy::calcInterferenceWeight(LiveInterval &VirtReg, unsigned PhysReg){
+  float Sum = 0;
+  for (const unsigned *AI = TRI->getOverlaps(PhysReg); *AI; ++AI) {
+    LiveIntervalUnion::Query &Q = query(VirtReg, *AI);
+    Q.collectInterferingVRegs();
+    if (Q.seenUnspillableVReg())
+      return HUGE_VALF;
+    for (unsigned i = 0, e = Q.interferingVRegs().size(); i != e; ++i)
+      Sum += Q.interferingVRegs()[i]->weight;
   }
+  return Sum;
+}
 
-  // Execute the split around Loop.
-  SmallVector<LiveInterval*, 4> SpillRegs;
-  LiveRangeEdit LREdit(VirtReg, SplitVRegs, SpillRegs);
-  SplitEditor(*SA, *LIS, *VRM, *DomTree, LREdit)
-    .splitAroundLoop(Loop->getLoop());
+/// trySpillInterferences - Try to spill interfering registers instead of the
+/// current one. Only do it if the accumulated spill weight is smaller than the
+/// current spill weight.
+unsigned RAGreedy::trySpillInterferences(LiveInterval &VirtReg,
+                                         AllocationOrder &Order,
+                                     SmallVectorImpl<LiveInterval*> &NewVRegs) {
+  NamedRegionTimer T("Spill Interference", TimerGroupName, TimePassesIsEnabled);
+  unsigned BestPhys = 0;
+  float BestWeight = 0;
 
-  if (VerifyEnabled)
-    MF->verify(this, "After splitting live range around loop");
+  Order.rewind();
+  while (unsigned PhysReg = Order.next()) {
+    float Weight = calcInterferenceWeight(VirtReg, PhysReg);
+    if (Weight == HUGE_VALF || Weight >= VirtReg.weight)
+      continue;
+    if (!BestPhys || Weight < BestWeight)
+      BestPhys = PhysReg, BestWeight = Weight;
+  }
 
-  // We have new split regs, don't assign anything.
-  return 0;
+  // No candidates found.
+  if (!BestPhys)
+    return 0;
+
+  // Collect all interfering registers.
+  SmallVector<LiveInterval*, 8> Spills;
+  for (const unsigned *AI = TRI->getOverlaps(BestPhys); *AI; ++AI) {
+    LiveIntervalUnion::Query &Q = query(VirtReg, *AI);
+    Spills.append(Q.interferingVRegs().begin(), Q.interferingVRegs().end());
+    for (unsigned i = 0, e = Q.interferingVRegs().size(); i != e; ++i) {
+      LiveInterval *VReg = Q.interferingVRegs()[i];
+      PhysReg2LiveUnion[*AI].extract(*VReg);
+      VRM->clearVirt(VReg->reg);
+    }
+  }
+
+  // Spill them all.
+  DEBUG(dbgs() << "spilling " << Spills.size() << " interferences with weight "
+               << BestWeight << '\n');
+  for (unsigned i = 0, e = Spills.size(); i != e; ++i)
+    spiller().spill(Spills[i], NewVRegs, Spills);
+  return BestPhys;
 }
 
-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.
+//===----------------------------------------------------------------------===//
+//                            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()) {
-    // 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.
+    if (!checkPhysRegInterference(VirtReg, PhysReg))
       return PhysReg;
-    }
-    assert(!VirtReg.empty() && "Empty VirtReg has interference");
-    LiveInterval *InterferingVirtReg =
-      Queries[InterfReg].firstInterference().liveUnionPos().value();
-
-    // The current VirtReg must either be spillable, or one of its interferences
-    // must have less spill weight.
-    if (InterferingVirtReg->weight < VirtReg.weight )
-      PhysRegSpillCands.push_back(PhysReg);
   }
 
   // Try to reassign interferences.
   if (unsigned PhysReg = tryReassign(VirtReg, Order))
     return PhysReg;
 
+  assert(NewVRegs.empty() && "Cannot append to existing NewVRegs");
+
   // Try splitting VirtReg or interferences.
-  unsigned PhysReg = trySplit(VirtReg, Order, SplitVRegs);
-  if (PhysReg || !SplitVRegs.empty())
+  unsigned PhysReg = trySplit(VirtReg, Order, NewVRegs);
+  if (PhysReg || !NewVRegs.empty())
     return PhysReg;
 
   // Try to spill another interfering reg with less spill weight.
-  NamedRegionTimer T("Spiller", TimerGroupName, TimePassesIsEnabled);
-  //
-  // FIXME: do this in two steps: (1) check for unspillable interferences while
-  // accumulating spill weight; (2) spill the interferences with lowest
-  // aggregate spill weight.
-  for (SmallVectorImpl<unsigned>::iterator PhysRegI = PhysRegSpillCands.begin(),
-         PhysRegE = PhysRegSpillCands.end(); PhysRegI != PhysRegE; ++PhysRegI) {
-
-    if (!spillInterferences(VirtReg, *PhysRegI, SplitVRegs)) continue;
-
-    assert(checkPhysRegInterference(VirtReg, *PhysRegI) == 0 &&
-           "Interference after spill.");
-    // Tell the caller to allocate to this newly freed physical register.
-    return *PhysRegI;
-  }
+  PhysReg = trySpillInterferences(VirtReg, Order, NewVRegs);
+  if (PhysReg)
+    return PhysReg;
 
-  // No other spill candidates were found, so spill the current VirtReg.
-  DEBUG(dbgs() << "spilling: " << VirtReg << '\n');
+  // Finally spill VirtReg itself.
+  NamedRegionTimer T("Spiller", TimerGroupName, TimePassesIsEnabled);
   SmallVector<LiveInterval*, 1> pendingSpills;
-
-  spiller().spill(&VirtReg, SplitVRegs, pendingSpills);
+  spiller().spill(&VirtReg, NewVRegs, pendingSpills);
 
   // The live virtual register requesting allocation was spilled, so tell
   // the caller not to allocate anything during this round.
@@ -410,11 +990,15 @@ bool RAGreedy::runOnMachineFunction(MachineFunction &mf) {
     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(createInlineSpiller(*this, *MF, *VRM));
   Loops = &getAnalysis<MachineLoopInfo>();
   LoopRanges = &getAnalysis<MachineLoopRanges>();
+  Bundles = &getAnalysis<EdgeBundles>();
+  SpillPlacer = &getAnalysis<SpillPlacement>();
+
   SA.reset(new SplitAnalysis(*MF, *LIS, *Loops));
 
   allocatePhysRegs();