RegAllocGreedy.cpp

    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");

  SmallVector<LiveInterval*, 4> SpillRegs;
  LiveRangeEdit LREdit(VirtReg, NewVRegs, SpillRegs);
  SplitEditor SE(*SA, *LIS, *VRM, *DomTree, LREdit);

  SE.openIntv();
  SlotIndex SegStart = SE.enterIntvBefore(Uses[BestBefore]);
  SlotIndex SegStop  = SE.leaveIntvAfter(Uses[BestAfter]);
  SE.useIntv(SegStart, SegStop);
  SE.closeIntv();
  SE.finish();
  ++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) {
  SA->analyze(&VirtReg);

  // Local intervals are handled separately.
  if (LIS->intervalIsInOneMBB(VirtReg)) {
    NamedRegionTimer T("Local Splitting", TimerGroupName, TimePassesIsEnabled);
    return tryLocalSplit(VirtReg, Order, NewVRegs);
  }

  NamedRegionTimer T("Global Splitting", TimerGroupName, TimePassesIsEnabled);

  // 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 (VerifyEnabled)
      MF->verify(this, "After splitting live range around basic blocks");
  }

  // 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;
}

/// 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;

  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;
  }

  // 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];
      unassign(*VReg, *AI);
    }
  }

  // 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;
}


//===----------------------------------------------------------------------===//
//                            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;
  }

  // Try to reassign interferences.
  if (unsigned PhysReg = tryReassignOrEvict(VirtReg, Order, NewVRegs))
    return PhysReg;

  assert(NewVRegs.empty() && "Cannot append to existing NewVRegs");

  // Try splitting VirtReg or interferences.
  unsigned PhysReg = trySplit(VirtReg, Order, NewVRegs);
  if (PhysReg || !NewVRegs.empty())
    return PhysReg;

  // Try to spill another interfering reg with less spill weight.
  PhysReg = trySpillInterferences(VirtReg, Order, NewVRegs);
  if (PhysReg)
    return PhysReg;

  // Finally spill VirtReg itself.
  NamedRegionTimer T("Spiller", TimerGroupName, TimePassesIsEnabled);
  SmallVector<LiveInterval*, 1> 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.
  return 0;
}

bool RAGreedy::runOnMachineFunction(MachineFunction &mf) {
  DEBUG(dbgs() << "********** GREEDY REGISTER ALLOCATION **********\n"
               << "********** Function: "
               << ((Value*)mf.getFunction())->getName() << '\n');

  MF = &mf;
  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(createInlineSpiller(*this, *MF, *VRM));
  Loops = &getAnalysis<MachineLoopInfo>();
  LoopRanges = &getAnalysis<MachineLoopRanges>();
  Bundles = &getAnalysis<EdgeBundles>();
  SpillPlacer = &getAnalysis<SpillPlacement>();

  SA.reset(new SplitAnalysis(*VRM, *LIS, *Loops));

  allocatePhysRegs();
  addMBBLiveIns(MF);
  LIS->addKillFlags();

  // Run rewriter
  {
    NamedRegionTimer T("Rewriter", TimerGroupName, TimePassesIsEnabled);
    VRM->rewrite(Indexes);
  }

  // The pass output is in VirtRegMap. Release all the transient data.
  releaseMemory();

  return true;
}