PreAllocSplitting.cpp

        UMII2 = UseMIs.find(UseMBB);
      UMII->second.push_back(&UseMO);
      UMII2->second.insert(UseMI);
    } else {
      SmallVector<MachineOperand*, 4> Ops;
      Ops.push_back(&UseMO);
      Uses.insert(std::make_pair(UseMBB, Ops));
      SmallPtrSet<MachineInstr*, 4> MIs;
      MIs.insert(UseMI);
      UseMIs.insert(std::make_pair(UseMBB, MIs));
    }
  }

  // Walk up the predecessor chains.
  SmallPtrSet<MachineBasicBlock*, 8> Visited;
  ShrinkWrapLiveInterval(ValNo, BarrierMBB, NULL, DefMI->getParent(), Visited,
                         Uses, UseMIs, UseMBBs);

  // Remove live range from barrier to the restore. FIXME: Find a better
  // point to re-start the live interval.
  VNInfo* AfterValNo = UpdateRegisterInterval(ValNo,
                                              LIs->getUseIndex(BarrierIdx)+1,
                                              LIs->getDefIndex(RestoreIdx));
  
  // Attempt to renumber the new valno into a new vreg.
  RenumberValno(AfterValNo);
}

/// RenumberValno - Split the given valno out into a new vreg, allowing it to
/// be allocated to a different register.  This function creates a new vreg,
/// copies the valno and its live ranges over to the new vreg's interval,
/// removes them from the old interval, and rewrites all uses and defs of
/// the original reg to the new vreg within those ranges.
void PreAllocSplitting::RenumberValno(VNInfo* VN) {
  SmallVector<VNInfo*, 4> Stack;
  SmallVector<VNInfo*, 4> VNsToCopy;
  Stack.push_back(VN);

  // Walk through and copy the valno we care about, and any other valnos
  // that are two-address redefinitions of the one we care about.  These
  // will need to be rewritten as well.  We also check for safety of the 
  // renumbering here, by making sure that none of the valno involved has
  // phi kills.
  while (!Stack.empty()) {
    VNInfo* OldVN = Stack.back();
    Stack.pop_back();
    
    // Bail out if we ever encounter a valno that has a PHI kill.  We can't
    // renumber these.
    if (OldVN->hasPHIKill) return;
    
    VNsToCopy.push_back(OldVN);
    
    // Locate two-address redefinitions
    for (SmallVector<unsigned, 4>::iterator KI = OldVN->kills.begin(),
         KE = OldVN->kills.end(); KI != KE; ++KI) {
      MachineInstr* MI = LIs->getInstructionFromIndex(*KI);
      //if (!MI) continue;
      unsigned DefIdx = MI->findRegisterDefOperandIdx(CurrLI->reg);
      if (DefIdx == ~0U) continue;
      if (MI->isRegReDefinedByTwoAddr(DefIdx)) {
        VNInfo* NextVN =
                     CurrLI->findDefinedVNInfo(LiveIntervals::getDefIndex(*KI));
        Stack.push_back(NextVN);
      }
    }
  }
  
  // Create the new vreg
  unsigned NewVReg = MRI->createVirtualRegister(MRI->getRegClass(CurrLI->reg));
  
  // Create the new live interval
  LiveInterval& NewLI = LIs->getOrCreateInterval(NewVReg);
  
  for (SmallVector<VNInfo*, 4>::iterator OI = VNsToCopy.begin(), OE = 
       VNsToCopy.end(); OI != OE; ++OI) {
    VNInfo* OldVN = *OI;
    
    // Copy the valno over
    VNInfo* NewVN = NewLI.getNextValue(OldVN->def, OldVN->copy, 
                                       LIs->getVNInfoAllocator());
    NewLI.copyValNumInfo(NewVN, OldVN);
    NewLI.MergeValueInAsValue(*CurrLI, OldVN, NewVN);

    // Remove the valno from the old interval
    CurrLI->removeValNo(OldVN);
  }
  
  // Rewrite defs and uses.  This is done in two stages to avoid invalidating
  // the reg_iterator.
  SmallVector<std::pair<MachineInstr*, unsigned>, 8> OpsToChange;
  
  for (MachineRegisterInfo::reg_iterator I = MRI->reg_begin(CurrLI->reg),
         E = MRI->reg_end(); I != E; ++I) {
    MachineOperand& MO = I.getOperand();
    unsigned InstrIdx = LIs->getInstructionIndex(&*I);
    
    if ((MO.isUse() && NewLI.liveAt(LiveIntervals::getUseIndex(InstrIdx))) ||
        (MO.isDef() && NewLI.liveAt(LiveIntervals::getDefIndex(InstrIdx))))
      OpsToChange.push_back(std::make_pair(&*I, I.getOperandNo()));
  }
  
  for (SmallVector<std::pair<MachineInstr*, unsigned>, 8>::iterator I =
       OpsToChange.begin(), E = OpsToChange.end(); I != E; ++I) {
    MachineInstr* Inst = I->first;
    unsigned OpIdx = I->second;
    MachineOperand& MO = Inst->getOperand(OpIdx);
    MO.setReg(NewVReg);
  }
  
  NumRenumbers++;
}

bool PreAllocSplitting::Rematerialize(unsigned vreg, VNInfo* ValNo,
                                      MachineInstr* DefMI,
                                      MachineBasicBlock::iterator RestorePt,
                                      unsigned RestoreIdx,
                                    SmallPtrSet<MachineInstr*, 4>& RefsInMBB) {
  MachineBasicBlock& MBB = *RestorePt->getParent();
  
  MachineBasicBlock::iterator KillPt = BarrierMBB->end();
  unsigned KillIdx = 0;
  if (ValNo->def == ~0U || DefMI->getParent() == BarrierMBB)
    KillPt = findSpillPoint(BarrierMBB, Barrier, NULL, RefsInMBB, KillIdx);
  else
    KillPt = findNextEmptySlot(DefMI->getParent(), DefMI, KillIdx);
  
  if (KillPt == DefMI->getParent()->end())
    return false;
  
  TII->reMaterialize(MBB, RestorePt, vreg, DefMI);
  LIs->InsertMachineInstrInMaps(prior(RestorePt), RestoreIdx);
  
  ReconstructLiveInterval(CurrLI);
  unsigned RematIdx = LIs->getInstructionIndex(prior(RestorePt));
  RematIdx = LiveIntervals::getDefIndex(RematIdx);
  RenumberValno(CurrLI->findDefinedVNInfo(RematIdx));
  
  ++NumSplits;
  ++NumRemats;
  return true;  
}

MachineInstr* PreAllocSplitting::FoldSpill(unsigned vreg, 
                                           const TargetRegisterClass* RC,
                                           MachineInstr* DefMI,
                                           MachineInstr* Barrier,
                                           MachineBasicBlock* MBB,
                                           int& SS,
                                    SmallPtrSet<MachineInstr*, 4>& RefsInMBB) {
  MachineBasicBlock::iterator Pt = MBB->begin();

  // Go top down if RefsInMBB is empty.
  if (RefsInMBB.empty())
    return 0;
  
  MachineBasicBlock::iterator FoldPt = Barrier;
  while (&*FoldPt != DefMI && FoldPt != MBB->begin() &&
         !RefsInMBB.count(FoldPt))
    --FoldPt;
  
  int OpIdx = FoldPt->findRegisterDefOperandIdx(vreg, false);
  if (OpIdx == -1)
    return 0;
  
  SmallVector<unsigned, 1> Ops;
  Ops.push_back(OpIdx);
  
  if (!TII->canFoldMemoryOperand(FoldPt, Ops))
    return 0;
  
  DenseMap<unsigned, int>::iterator I = IntervalSSMap.find(vreg);
  if (I != IntervalSSMap.end()) {
    SS = I->second;
  } else {
    SS = MFI->CreateStackObject(RC->getSize(), RC->getAlignment());
    
  }
  
  MachineInstr* FMI = TII->foldMemoryOperand(*MBB->getParent(),
                                             FoldPt, Ops, SS);
  
  if (FMI) {
    LIs->ReplaceMachineInstrInMaps(FoldPt, FMI);
    FMI = MBB->insert(MBB->erase(FoldPt), FMI);
    ++NumFolds;
    
    IntervalSSMap[vreg] = SS;
    CurrSLI = &LSs->getOrCreateInterval(SS);
    if (CurrSLI->hasAtLeastOneValue())
      CurrSValNo = CurrSLI->getValNumInfo(0);
    else
      CurrSValNo = CurrSLI->getNextValue(~0U, 0, LSs->getVNInfoAllocator());
  }
  
  return FMI;
}

/// SplitRegLiveInterval - Split (spill and restore) the given live interval
/// so it would not cross the barrier that's being processed. Shrink wrap
/// (minimize) the live interval to the last uses.
bool PreAllocSplitting::SplitRegLiveInterval(LiveInterval *LI) {
  CurrLI = LI;

  // Find live range where current interval cross the barrier.
  LiveInterval::iterator LR =
    CurrLI->FindLiveRangeContaining(LIs->getUseIndex(BarrierIdx));
  VNInfo *ValNo = LR->valno;

  if (ValNo->def == ~1U) {
    // Defined by a dead def? How can this be?
    assert(0 && "Val# is defined by a dead def?");
    abort();
  }

  MachineInstr *DefMI = (ValNo->def != ~0U)
    ? LIs->getInstructionFromIndex(ValNo->def) : NULL;

  // If this would create a new join point, do not split.
  if (DefMI && createsNewJoin(LR, DefMI->getParent(), Barrier->getParent()))
    return false;

  // Find all references in the barrier mbb.
  SmallPtrSet<MachineInstr*, 4> RefsInMBB;
  for (MachineRegisterInfo::reg_iterator I = MRI->reg_begin(CurrLI->reg),
         E = MRI->reg_end(); I != E; ++I) {
    MachineInstr *RefMI = &*I;
    if (RefMI->getParent() == BarrierMBB)
      RefsInMBB.insert(RefMI);
  }

  // Find a point to restore the value after the barrier.
  unsigned RestoreIndex = 0;
  MachineBasicBlock::iterator RestorePt =
    findRestorePoint(BarrierMBB, Barrier, LR->end, RefsInMBB, RestoreIndex);
  if (RestorePt == BarrierMBB->end())
    return false;

  if (DefMI && LIs->isReMaterializable(*LI, ValNo, DefMI))
    if (Rematerialize(LI->reg, ValNo, DefMI, RestorePt,
                      RestoreIndex, RefsInMBB))
    return true;

  // Add a spill either before the barrier or after the definition.
  MachineBasicBlock *DefMBB = DefMI ? DefMI->getParent() : NULL;
  const TargetRegisterClass *RC = MRI->getRegClass(CurrLI->reg);
  unsigned SpillIndex = 0;
  MachineInstr *SpillMI = NULL;
  int SS = -1;
  if (ValNo->def == ~0U) {
    // If it's defined by a phi, we must split just before the barrier.
    if ((SpillMI = FoldSpill(LI->reg, RC, 0, Barrier,
                            BarrierMBB, SS, RefsInMBB))) {
      SpillIndex = LIs->getInstructionIndex(SpillMI);
    } else {
      MachineBasicBlock::iterator SpillPt = 
        findSpillPoint(BarrierMBB, Barrier, NULL, RefsInMBB, SpillIndex);
      if (SpillPt == BarrierMBB->begin())
        return false; // No gap to insert spill.
      // Add spill.
    
      SS = CreateSpillStackSlot(CurrLI->reg, RC);
      TII->storeRegToStackSlot(*BarrierMBB, SpillPt, CurrLI->reg, true, SS, RC);
      SpillMI = prior(SpillPt);
      LIs->InsertMachineInstrInMaps(SpillMI, SpillIndex);
    }
  } else if (!IsAvailableInStack(DefMBB, CurrLI->reg, ValNo->def,
                                 RestoreIndex, SpillIndex, SS)) {
    // If it's already split, just restore the value. There is no need to spill
    // the def again.
    if (!DefMI)
      return false; // Def is dead. Do nothing.
    
    if ((SpillMI = FoldSpill(LI->reg, RC, DefMI, Barrier,
                            BarrierMBB, SS, RefsInMBB))) {
      SpillIndex = LIs->getInstructionIndex(SpillMI);
    } else {
      // Check if it's possible to insert a spill after the def MI.
      MachineBasicBlock::iterator SpillPt;
      if (DefMBB == BarrierMBB) {
        // Add spill after the def and the last use before the barrier.
        SpillPt = findSpillPoint(BarrierMBB, Barrier, DefMI,
                                 RefsInMBB, SpillIndex);
        if (SpillPt == DefMBB->begin())
          return false; // No gap to insert spill.
      } else {
        SpillPt = findNextEmptySlot(DefMBB, DefMI, SpillIndex);
        if (SpillPt == DefMBB->end())
          return false; // No gap to insert spill.
      }
      // Add spill. The store instruction kills the register if def is before
      // the barrier in the barrier block.
      SS = CreateSpillStackSlot(CurrLI->reg, RC);
      TII->storeRegToStackSlot(*DefMBB, SpillPt, CurrLI->reg,
                               DefMBB == BarrierMBB, SS, RC);
      SpillMI = prior(SpillPt);
      LIs->InsertMachineInstrInMaps(SpillMI, SpillIndex);
    }
  }

  // Remember def instruction index to spill index mapping.
  if (DefMI && SpillMI)
    Def2SpillMap[ValNo->def] = SpillIndex;

  // Add restore.
  TII->loadRegFromStackSlot(*BarrierMBB, RestorePt, CurrLI->reg, SS, RC);
  MachineInstr *LoadMI = prior(RestorePt);
  LIs->InsertMachineInstrInMaps(LoadMI, RestoreIndex);

  // Update spill stack slot live interval.
  UpdateSpillSlotInterval(ValNo, LIs->getUseIndex(SpillIndex)+1,
                          LIs->getDefIndex(RestoreIndex));

  ReconstructLiveInterval(CurrLI);
  unsigned RestoreIdx = LIs->getInstructionIndex(prior(RestorePt));
  RestoreIdx = LiveIntervals::getDefIndex(RestoreIdx);
  RenumberValno(CurrLI->findDefinedVNInfo(RestoreIdx));
  
  ++NumSplits;
  return true;
}

/// SplitRegLiveIntervals - Split all register live intervals that cross the
/// barrier that's being processed.
bool
PreAllocSplitting::SplitRegLiveIntervals(const TargetRegisterClass **RCs,
                                         SmallPtrSet<LiveInterval*, 8>& Split) {
  // First find all the virtual registers whose live intervals are intercepted
  // by the current barrier.
  SmallVector<LiveInterval*, 8> Intervals;
  for (const TargetRegisterClass **RC = RCs; *RC; ++RC) {
    if (TII->IgnoreRegisterClassBarriers(*RC))
      continue;
    std::vector<unsigned> &VRs = MRI->getRegClassVirtRegs(*RC);
    for (unsigned i = 0, e = VRs.size(); i != e; ++i) {
      unsigned Reg = VRs[i];
      if (!LIs->hasInterval(Reg))
        continue;
      LiveInterval *LI = &LIs->getInterval(Reg);
      if (LI->liveAt(BarrierIdx) && !Barrier->readsRegister(Reg))
        // Virtual register live interval is intercepted by the barrier. We
        // should split and shrink wrap its interval if possible.
        Intervals.push_back(LI);
    }
  }

  // Process the affected live intervals.
  bool Change = false;
  while (!Intervals.empty()) {
    if (PreSplitLimit != -1 && (int)NumSplits == PreSplitLimit)
      break;
    else if (NumSplits == 4)
      Change |= Change;
    LiveInterval *LI = Intervals.back();
    Intervals.pop_back();
    bool result = SplitRegLiveInterval(LI);
    if (result) Split.insert(LI);
    Change |= result;
  }

  return Change;
}

unsigned PreAllocSplitting::getNumberOfSpills(
                                  SmallPtrSet<MachineInstr*, 4>& MIs,
                                  unsigned Reg, int FrameIndex) {
  unsigned Spills = 0;
  for (SmallPtrSet<MachineInstr*, 4>::iterator UI = MIs.begin(), UE = MIs.end();
       UI != UI; ++UI) {
    int StoreFrameIndex;
    unsigned StoreVReg = TII->isStoreToStackSlot(*UI, StoreFrameIndex);
    if (StoreVReg == Reg && StoreFrameIndex == FrameIndex)
      Spills++;
  }
  
  return Spills;
}

/// removeDeadSpills - After doing splitting, filter through all intervals we've
/// split, and see if any of the spills are unnecessary.  If so, remove them.
bool PreAllocSplitting::removeDeadSpills(SmallPtrSet<LiveInterval*, 8>& split) {
  bool changed = false;
  
  for (SmallPtrSet<LiveInterval*, 8>::iterator LI = split.begin(),
       LE = split.end(); LI != LE; ++LI) {
    DenseMap<VNInfo*, SmallPtrSet<MachineInstr*, 4> > VNUseCount;
    
    for (MachineRegisterInfo::use_iterator UI = MRI->use_begin((*LI)->reg),
         UE = MRI->use_end(); UI != UE; ++UI) {
      unsigned index = LIs->getInstructionIndex(&*UI);
      index = LiveIntervals::getUseIndex(index);
      
      const LiveRange* LR = (*LI)->getLiveRangeContaining(index);
      VNUseCount[LR->valno].insert(&*UI);
    }
    
    for (LiveInterval::vni_iterator VI = (*LI)->vni_begin(),
         VE = (*LI)->vni_end(); VI != VE; ++VI) {
      VNInfo* CurrVN = *VI;
      if (CurrVN->hasPHIKill) continue;
      
      unsigned DefIdx = CurrVN->def;
      if (DefIdx == ~0U || DefIdx == ~1U) continue;
    
      MachineInstr* DefMI = LIs->getInstructionFromIndex(DefIdx);
      int FrameIndex;
      if (!TII->isLoadFromStackSlot(DefMI, FrameIndex)) continue;
      
      if (VNUseCount[CurrVN].size() == 0) {
        LIs->RemoveMachineInstrFromMaps(DefMI);
        (*LI)->removeValNo(CurrVN);
        DefMI->eraseFromParent();
        NumDeadSpills++;
        changed = true;
        continue;
      }
      
      unsigned SpillCount = getNumberOfSpills(VNUseCount[CurrVN],
                                              (*LI)->reg, FrameIndex);
      if (SpillCount != VNUseCount[CurrVN].size()) continue;
        
      for (SmallPtrSet<MachineInstr*, 4>::iterator UI = 
           VNUseCount[CurrVN].begin(), UE = VNUseCount[CurrVN].end();
           UI != UI; ++UI) {
        LIs->RemoveMachineInstrFromMaps(*UI);
        (*UI)->eraseFromParent();
      }
        
      LIs->RemoveMachineInstrFromMaps(DefMI);
      (*LI)->removeValNo(CurrVN);
      DefMI->eraseFromParent();
      NumDeadSpills++;
      changed = true;
    }
  }
  
  return changed;
}

bool PreAllocSplitting::createsNewJoin(LiveRange* LR,
                                       MachineBasicBlock* DefMBB,
                                       MachineBasicBlock* BarrierMBB) {
  if (DefMBB == BarrierMBB)
    return false;
  
  if (LR->valno->hasPHIKill)
    return false;
  
  unsigned MBBEnd = LIs->getMBBEndIdx(BarrierMBB);
  if (LR->end < MBBEnd)
    return false;
  
  MachineLoopInfo& MLI = getAnalysis<MachineLoopInfo>();
  if (MLI.getLoopFor(DefMBB) != MLI.getLoopFor(BarrierMBB))
    return true;
  
  MachineDominatorTree& MDT = getAnalysis<MachineDominatorTree>();
  SmallPtrSet<MachineBasicBlock*, 4> Visited;
  typedef std::pair<MachineBasicBlock*,
                    MachineBasicBlock::succ_iterator> ItPair;
  SmallVector<ItPair, 4> Stack;
  Stack.push_back(std::make_pair(BarrierMBB, BarrierMBB->succ_begin()));
  
  while (!Stack.empty()) {
    ItPair P = Stack.back();
    Stack.pop_back();
    
    MachineBasicBlock* PredMBB = P.first;
    MachineBasicBlock::succ_iterator S = P.second;
    
    if (S == PredMBB->succ_end())
      continue;
    else if (Visited.count(*S)) {
      Stack.push_back(std::make_pair(PredMBB, ++S));
      continue;
    } else
      Stack.push_back(std::make_pair(PredMBB, S+1));
    
    MachineBasicBlock* MBB = *S;
    Visited.insert(MBB);
    
    if (MBB == BarrierMBB)
      return true;
    
    MachineDomTreeNode* DefMDTN = MDT.getNode(DefMBB);
    MachineDomTreeNode* BarrierMDTN = MDT.getNode(BarrierMBB);
    MachineDomTreeNode* MDTN = MDT.getNode(MBB)->getIDom();
    while (MDTN) {
      if (MDTN == DefMDTN)
        return true;
      else if (MDTN == BarrierMDTN)
        break;
      MDTN = MDTN->getIDom();
    }
    
    MBBEnd = LIs->getMBBEndIdx(MBB);
    if (LR->end > MBBEnd)
      Stack.push_back(std::make_pair(MBB, MBB->succ_begin()));
  }
  
  return false;
} 
  

bool PreAllocSplitting::runOnMachineFunction(MachineFunction &MF) {
  CurrMF = &MF;
  TM     = &MF.getTarget();
  TII    = TM->getInstrInfo();
  MFI    = MF.getFrameInfo();
  MRI    = &MF.getRegInfo();
  LIs    = &getAnalysis<LiveIntervals>();
  LSs    = &getAnalysis<LiveStacks>();

  bool MadeChange = false;

  // Make sure blocks are numbered in order.
  MF.RenumberBlocks();

  MachineBasicBlock *Entry = MF.begin();
  SmallPtrSet<MachineBasicBlock*,16> Visited;

  SmallPtrSet<LiveInterval*, 8> Split;

  for (df_ext_iterator<MachineBasicBlock*, SmallPtrSet<MachineBasicBlock*,16> >
         DFI = df_ext_begin(Entry, Visited), E = df_ext_end(Entry, Visited);
       DFI != E; ++DFI) {
    BarrierMBB = *DFI;
    for (MachineBasicBlock::iterator I = BarrierMBB->begin(),
           E = BarrierMBB->end(); I != E; ++I) {
      Barrier = &*I;
      const TargetRegisterClass **BarrierRCs =
        Barrier->getDesc().getRegClassBarriers();
      if (!BarrierRCs)
        continue;
      BarrierIdx = LIs->getInstructionIndex(Barrier);
      MadeChange |= SplitRegLiveIntervals(BarrierRCs, Split);
    }
  }

  MadeChange |= removeDeadSpills(Split);

  return MadeChange;
}