RegAllocLinearScan.cpp

          if (SrcSubReg)
            Reg = tri_->getSubReg(Reg, SrcSubReg);
          if (DstSubReg)
            Reg = tri_->getMatchingSuperReg(Reg, DstSubReg, RC);
          if (Reg && allocatableRegs_[Reg] && RC->contains(Reg))
            cur->preference = Reg;
        }
      }
    }
  }

  // For every interval in inactive we overlap with, mark the
  // register as not free and update spill weights.
  for (IntervalPtrs::const_iterator i = inactive_.begin(),
         e = inactive_.end(); i != e; ++i) {
    unsigned Reg = i->first->reg;
    assert(TargetRegisterInfo::isVirtualRegister(Reg) &&
           "Can only allocate virtual registers!");
    const TargetRegisterClass *RegRC = mri_->getRegClass(Reg);
    // If this is not in a related reg class to the register we're allocating, 
    // don't check it.
    if (RelatedRegClasses.getLeaderValue(RegRC) == RCLeader &&
        cur->overlapsFrom(*i->first, i->second-1)) {
      Reg = vrm_->getPhys(Reg);
      addRegUse(Reg);
      SpillWeightsToAdd.push_back(std::make_pair(Reg, i->first->weight));
    }
  }
  
  // Speculatively check to see if we can get a register right now.  If not,
  // we know we won't be able to by adding more constraints.  If so, we can
  // check to see if it is valid.  Doing an exhaustive search of the fixed_ list
  // is very bad (it contains all callee clobbered registers for any functions
  // with a call), so we want to avoid doing that if possible.
  unsigned physReg = getFreePhysReg(cur);
  unsigned BestPhysReg = physReg;
  if (physReg) {
    // We got a register.  However, if it's in the fixed_ list, we might
    // conflict with it.  Check to see if we conflict with it or any of its
    // aliases.
    SmallSet<unsigned, 8> RegAliases;
    for (const unsigned *AS = tri_->getAliasSet(physReg); *AS; ++AS)
      RegAliases.insert(*AS);
    
    bool ConflictsWithFixed = false;
    for (unsigned i = 0, e = fixed_.size(); i != e; ++i) {
      IntervalPtr &IP = fixed_[i];
      if (physReg == IP.first->reg || RegAliases.count(IP.first->reg)) {
        // Okay, this reg is on the fixed list.  Check to see if we actually
        // conflict.
        LiveInterval *I = IP.first;
        if (I->endNumber() > StartPosition) {
          LiveInterval::iterator II = I->advanceTo(IP.second, StartPosition);
          IP.second = II;
          if (II != I->begin() && II->start > StartPosition)
            --II;
          if (cur->overlapsFrom(*I, II)) {
            ConflictsWithFixed = true;
            break;
          }
        }
      }
    }
    
    // Okay, the register picked by our speculative getFreePhysReg call turned
    // out to be in use.  Actually add all of the conflicting fixed registers to
    // regUse_ so we can do an accurate query.
    if (ConflictsWithFixed) {
      // For every interval in fixed we overlap with, mark the register as not
      // free and update spill weights.
      for (unsigned i = 0, e = fixed_.size(); i != e; ++i) {
        IntervalPtr &IP = fixed_[i];
        LiveInterval *I = IP.first;

        const TargetRegisterClass *RegRC = OneClassForEachPhysReg[I->reg];
        if (RelatedRegClasses.getLeaderValue(RegRC) == RCLeader &&       
            I->endNumber() > StartPosition) {
          LiveInterval::iterator II = I->advanceTo(IP.second, StartPosition);
          IP.second = II;
          if (II != I->begin() && II->start > StartPosition)
            --II;
          if (cur->overlapsFrom(*I, II)) {
            unsigned reg = I->reg;
            addRegUse(reg);
            SpillWeightsToAdd.push_back(std::make_pair(reg, I->weight));
          }
        }
      }

      // Using the newly updated regUse_ object, which includes conflicts in the
      // future, see if there are any registers available.
      physReg = getFreePhysReg(cur);
    }
  }
    
  // Restore the physical register tracker, removing information about the
  // future.
  restoreRegUses();
  
  // If we find a free register, we are done: assign this virtual to
  // the free physical register and add this interval to the active
  // list.
  if (physReg) {
    DOUT <<  tri_->getName(physReg) << '\n';
    vrm_->assignVirt2Phys(cur->reg, physReg);
    addRegUse(physReg);
    active_.push_back(std::make_pair(cur, cur->begin()));
    handled_.push_back(cur);

    // "Upgrade" the physical register since it has been allocated.
    UpgradeRegister(physReg);
    if (LiveInterval *NextReloadLI = hasNextReloadInterval(cur)) {
      // "Downgrade" physReg to try to keep physReg from being allocated until
      // the next reload from the same SS is allocated. 
      NextReloadLI->preference = physReg;
      DowngradeRegister(cur, physReg);
    }
    return;
  }
  DOUT << "no free registers\n";

  // Compile the spill weights into an array that is better for scanning.
  std::vector<float> SpillWeights(tri_->getNumRegs(), 0.0f);
  for (std::vector<std::pair<unsigned, float> >::iterator
       I = SpillWeightsToAdd.begin(), E = SpillWeightsToAdd.end(); I != E; ++I)
    updateSpillWeights(SpillWeights, I->first, I->second, RC);
  
  // for each interval in active, update spill weights.
  for (IntervalPtrs::const_iterator i = active_.begin(), e = active_.end();
       i != e; ++i) {
    unsigned reg = i->first->reg;
    assert(TargetRegisterInfo::isVirtualRegister(reg) &&
           "Can only allocate virtual registers!");
    reg = vrm_->getPhys(reg);
    updateSpillWeights(SpillWeights, reg, i->first->weight, RC);
  }
 
  DOUT << "\tassigning stack slot at interval "<< *cur << ":\n";

  // Find a register to spill.
  float minWeight = HUGE_VALF;
  unsigned minReg = 0; /*cur->preference*/;  // Try the pref register first.

  bool Found = false;
  std::vector<std::pair<unsigned,float> > RegsWeights;
  if (!minReg || SpillWeights[minReg] == HUGE_VALF)
    for (TargetRegisterClass::iterator i = RC->allocation_order_begin(*mf_),
           e = RC->allocation_order_end(*mf_); i != e; ++i) {
      unsigned reg = *i;
      float regWeight = SpillWeights[reg];
      if (minWeight > regWeight)
        Found = true;
      RegsWeights.push_back(std::make_pair(reg, regWeight));
    }
  
  // If we didn't find a register that is spillable, try aliases?
  if (!Found) {
    for (TargetRegisterClass::iterator i = RC->allocation_order_begin(*mf_),
           e = RC->allocation_order_end(*mf_); i != e; ++i) {
      unsigned reg = *i;
      // No need to worry about if the alias register size < regsize of RC.
      // We are going to spill all registers that alias it anyway.
      for (const unsigned* as = tri_->getAliasSet(reg); *as; ++as)
        RegsWeights.push_back(std::make_pair(*as, SpillWeights[*as]));
    }
  }

  // Sort all potential spill candidates by weight.
  std::sort(RegsWeights.begin(), RegsWeights.end(), WeightCompare());
  minReg = RegsWeights[0].first;
  minWeight = RegsWeights[0].second;
  if (minWeight == HUGE_VALF) {
    // All registers must have inf weight. Just grab one!
    minReg = BestPhysReg ? BestPhysReg : *RC->allocation_order_begin(*mf_);
    if (cur->weight == HUGE_VALF ||
        li_->getApproximateInstructionCount(*cur) == 0) {
      // Spill a physical register around defs and uses.
      if (li_->spillPhysRegAroundRegDefsUses(*cur, minReg, *vrm_)) {
        // spillPhysRegAroundRegDefsUses may have invalidated iterator stored
        // in fixed_. Reset them.
        for (unsigned i = 0, e = fixed_.size(); i != e; ++i) {
          IntervalPtr &IP = fixed_[i];
          LiveInterval *I = IP.first;
          if (I->reg == minReg || tri_->isSubRegister(minReg, I->reg))
            IP.second = I->advanceTo(I->begin(), StartPosition);
        }

        DowngradedRegs.clear();
        assignRegOrStackSlotAtInterval(cur);
      } else {
        cerr << "Ran out of registers during register allocation!\n";
        exit(1);
      }
      return;
    }
  }

  // Find up to 3 registers to consider as spill candidates.
  unsigned LastCandidate = RegsWeights.size() >= 3 ? 3 : 1;
  while (LastCandidate > 1) {
    if (weightsAreClose(RegsWeights[LastCandidate-1].second, minWeight))
      break;
    --LastCandidate;
  }

  DOUT << "\t\tregister(s) with min weight(s): ";
  DEBUG(for (unsigned i = 0; i != LastCandidate; ++i)
          DOUT << tri_->getName(RegsWeights[i].first)
               << " (" << RegsWeights[i].second << ")\n");

  // If the current has the minimum weight, we need to spill it and
  // add any added intervals back to unhandled, and restart
  // linearscan.
  if (cur->weight != HUGE_VALF && cur->weight <= minWeight) {
    DOUT << "\t\t\tspilling(c): " << *cur << '\n';
    SmallVector<LiveInterval*, 8> spillIs;
    std::vector<LiveInterval*> added;
    
    if (!NewSpillFramework) {
      added = li_->addIntervalsForSpills(*cur, spillIs, loopInfo, *vrm_);
    } else {
      added = spiller_->spill(cur); 
    }

    std::sort(added.begin(), added.end(), LISorter());
    addStackInterval(cur, ls_, li_, mri_, *vrm_);
    if (added.empty())
      return;  // Early exit if all spills were folded.

    // Merge added with unhandled.  Note that we have already sorted
    // intervals returned by addIntervalsForSpills by their starting
    // point.
    // This also update the NextReloadMap. That is, it adds mapping from a
    // register defined by a reload from SS to the next reload from SS in the
    // same basic block.
    MachineBasicBlock *LastReloadMBB = 0;
    LiveInterval *LastReload = 0;
    int LastReloadSS = VirtRegMap::NO_STACK_SLOT;
    for (unsigned i = 0, e = added.size(); i != e; ++i) {
      LiveInterval *ReloadLi = added[i];
      if (ReloadLi->weight == HUGE_VALF &&
          li_->getApproximateInstructionCount(*ReloadLi) == 0) {
        unsigned ReloadIdx = ReloadLi->beginNumber();
        MachineBasicBlock *ReloadMBB = li_->getMBBFromIndex(ReloadIdx);
        int ReloadSS = vrm_->getStackSlot(ReloadLi->reg);
        if (LastReloadMBB == ReloadMBB && LastReloadSS == ReloadSS) {
          // Last reload of same SS is in the same MBB. We want to try to
          // allocate both reloads the same register and make sure the reg
          // isn't clobbered in between if at all possible.
          assert(LastReload->beginNumber() < ReloadIdx);
          NextReloadMap.insert(std::make_pair(LastReload->reg, ReloadLi->reg));
        }
        LastReloadMBB = ReloadMBB;
        LastReload = ReloadLi;
        LastReloadSS = ReloadSS;
      }
      unhandled_.push(ReloadLi);
    }
    return;
  }

  ++NumBacktracks;

  // Push the current interval back to unhandled since we are going
  // to re-run at least this iteration. Since we didn't modify it it
  // should go back right in the front of the list
  unhandled_.push(cur);

  assert(TargetRegisterInfo::isPhysicalRegister(minReg) &&
         "did not choose a register to spill?");

  // We spill all intervals aliasing the register with
  // minimum weight, rollback to the interval with the earliest
  // start point and let the linear scan algorithm run again
  SmallVector<LiveInterval*, 8> spillIs;

  // Determine which intervals have to be spilled.
  findIntervalsToSpill(cur, RegsWeights, LastCandidate, spillIs);

  // Set of spilled vregs (used later to rollback properly)
  SmallSet<unsigned, 8> spilled;

  // The earliest start of a Spilled interval indicates up to where
  // in handled we need to roll back
  
  unsigned earliestStart = cur->beginNumber();
  LiveInterval *earliestStartInterval = cur;

  // Spill live intervals of virtual regs mapped to the physical register we
  // want to clear (and its aliases).  We only spill those that overlap with the
  // current interval as the rest do not affect its allocation. we also keep
  // track of the earliest start of all spilled live intervals since this will
  // mark our rollback point.
  std::vector<LiveInterval*> added;
  while (!spillIs.empty()) {
    bool epicFail = false;
    LiveInterval *sli = spillIs.back();
    spillIs.pop_back();
    DOUT << "\t\t\tspilling(a): " << *sli << '\n';
    earliestStart = std::min(earliestStart, sli->beginNumber());
    earliestStartInterval =
      (earliestStartInterval->beginNumber() < sli->beginNumber()) ?
         earliestStartInterval : sli;
    
    if (earliestStartInterval->beginNumber()!=earliestStart) {
      epicFail |= true;
      std::cerr << "What the 1 - "
      		<< "earliestStart = " << earliestStart
      		<< "earliestStartInterval = " << earliestStartInterval->beginNumber()
      		<< "\n";
    }
   
    std::vector<LiveInterval*> newIs;
    if (!NewSpillFramework) {
      newIs = li_->addIntervalsForSpills(*sli, spillIs, loopInfo, *vrm_);
    } else {
      newIs = spiller_->spill(sli);
    }
    addStackInterval(sli, ls_, li_, mri_, *vrm_);
    std::copy(newIs.begin(), newIs.end(), std::back_inserter(added));
    spilled.insert(sli->reg);

    if (earliestStartInterval->beginNumber()!=earliestStart) {
      epicFail |= true;
      std::cerr << "What the 2 - "
      		<< "earliestStart = " << earliestStart
      		<< "earliestStartInterval = " << earliestStartInterval->beginNumber()
      		<< "\n";
    }

    if (epicFail) {
      //abort();
    }
  }

  earliestStart = earliestStartInterval->beginNumber();

  DOUT << "\t\trolling back to: " << earliestStart << '\n';

  // Scan handled in reverse order up to the earliest start of a
  // spilled live interval and undo each one, restoring the state of
  // unhandled.
  while (!handled_.empty()) {
    LiveInterval* i = handled_.back();
    // If this interval starts before t we are done.
    if (i->beginNumber() < earliestStart)
      break;
    DOUT << "\t\t\tundo changes for: " << *i << '\n';
    handled_.pop_back();

    // When undoing a live interval allocation we must know if it is active or
    // inactive to properly update regUse_ and the VirtRegMap.
    IntervalPtrs::iterator it;
    if ((it = FindIntervalInVector(active_, i)) != active_.end()) {
      active_.erase(it);
      assert(!TargetRegisterInfo::isPhysicalRegister(i->reg));
      if (!spilled.count(i->reg))
        unhandled_.push(i);
      delRegUse(vrm_->getPhys(i->reg));
      vrm_->clearVirt(i->reg);
    } else if ((it = FindIntervalInVector(inactive_, i)) != inactive_.end()) {
      inactive_.erase(it);
      assert(!TargetRegisterInfo::isPhysicalRegister(i->reg));
      if (!spilled.count(i->reg))
        unhandled_.push(i);
      vrm_->clearVirt(i->reg);
    } else {
      assert(TargetRegisterInfo::isVirtualRegister(i->reg) &&
             "Can only allocate virtual registers!");
      vrm_->clearVirt(i->reg);
      unhandled_.push(i);
    }

    DenseMap<unsigned, unsigned>::iterator ii = DowngradeMap.find(i->reg);
    if (ii == DowngradeMap.end())
      // It interval has a preference, it must be defined by a copy. Clear the
      // preference now since the source interval allocation may have been
      // undone as well.
      i->preference = 0;
    else {
      UpgradeRegister(ii->second);
    }
  }

  // Rewind the iterators in the active, inactive, and fixed lists back to the
  // point we reverted to.
  RevertVectorIteratorsTo(active_, earliestStart);
  RevertVectorIteratorsTo(inactive_, earliestStart);
  RevertVectorIteratorsTo(fixed_, earliestStart);

  // Scan the rest and undo each interval that expired after t and
  // insert it in active (the next iteration of the algorithm will
  // put it in inactive if required)
  for (unsigned i = 0, e = handled_.size(); i != e; ++i) {
    LiveInterval *HI = handled_[i];
    if (!HI->expiredAt(earliestStart) &&
        HI->expiredAt(cur->beginNumber())) {
      DOUT << "\t\t\tundo changes for: " << *HI << '\n';
      active_.push_back(std::make_pair(HI, HI->begin()));
      assert(!TargetRegisterInfo::isPhysicalRegister(HI->reg));
      addRegUse(vrm_->getPhys(HI->reg));
    }
  }

  // Merge added with unhandled.
  // This also update the NextReloadMap. That is, it adds mapping from a
  // register defined by a reload from SS to the next reload from SS in the
  // same basic block.
  MachineBasicBlock *LastReloadMBB = 0;
  LiveInterval *LastReload = 0;
  int LastReloadSS = VirtRegMap::NO_STACK_SLOT;
  std::sort(added.begin(), added.end(), LISorter());
  for (unsigned i = 0, e = added.size(); i != e; ++i) {
    LiveInterval *ReloadLi = added[i];
    if (ReloadLi->weight == HUGE_VALF &&
        li_->getApproximateInstructionCount(*ReloadLi) == 0) {
      unsigned ReloadIdx = ReloadLi->beginNumber();
      MachineBasicBlock *ReloadMBB = li_->getMBBFromIndex(ReloadIdx);
      int ReloadSS = vrm_->getStackSlot(ReloadLi->reg);
      if (LastReloadMBB == ReloadMBB && LastReloadSS == ReloadSS) {
        // Last reload of same SS is in the same MBB. We want to try to
        // allocate both reloads the same register and make sure the reg
        // isn't clobbered in between if at all possible.
        assert(LastReload->beginNumber() < ReloadIdx);
        NextReloadMap.insert(std::make_pair(LastReload->reg, ReloadLi->reg));
      }
      LastReloadMBB = ReloadMBB;
      LastReload = ReloadLi;
      LastReloadSS = ReloadSS;
    }
    unhandled_.push(ReloadLi);
  }
}

unsigned RALinScan::getFreePhysReg(const TargetRegisterClass *RC,
                                   unsigned MaxInactiveCount,
                                   SmallVector<unsigned, 256> &inactiveCounts,
                                   bool SkipDGRegs) {
  unsigned FreeReg = 0;
  unsigned FreeRegInactiveCount = 0;

  TargetRegisterClass::iterator I = RC->allocation_order_begin(*mf_);
  TargetRegisterClass::iterator E = RC->allocation_order_end(*mf_);
  assert(I != E && "No allocatable register in this register class!");

  // Scan for the first available register.
  for (; I != E; ++I) {
    unsigned Reg = *I;
    // Ignore "downgraded" registers.
    if (SkipDGRegs && DowngradedRegs.count(Reg))
      continue;
    if (isRegAvail(Reg)) {
      FreeReg = Reg;
      if (FreeReg < inactiveCounts.size())
        FreeRegInactiveCount = inactiveCounts[FreeReg];
      else
        FreeRegInactiveCount = 0;
      break;
    }
  }

  // If there are no free regs, or if this reg has the max inactive count,
  // return this register.
  if (FreeReg == 0 || FreeRegInactiveCount == MaxInactiveCount)
    return FreeReg;
  
  // Continue scanning the registers, looking for the one with the highest
  // inactive count.  Alkis found that this reduced register pressure very
  // slightly on X86 (in rev 1.94 of this file), though this should probably be
  // reevaluated now.
  for (; I != E; ++I) {
    unsigned Reg = *I;
    // Ignore "downgraded" registers.
    if (SkipDGRegs && DowngradedRegs.count(Reg))
      continue;
    if (isRegAvail(Reg) && Reg < inactiveCounts.size() &&
        FreeRegInactiveCount < inactiveCounts[Reg]) {
      FreeReg = Reg;
      FreeRegInactiveCount = inactiveCounts[Reg];
      if (FreeRegInactiveCount == MaxInactiveCount)
        break;    // We found the one with the max inactive count.
    }
  }

  return FreeReg;
}

/// getFreePhysReg - return a free physical register for this virtual register
/// interval if we have one, otherwise return 0.
unsigned RALinScan::getFreePhysReg(LiveInterval *cur) {
  SmallVector<unsigned, 256> inactiveCounts;
  unsigned MaxInactiveCount = 0;
  
  const TargetRegisterClass *RC = mri_->getRegClass(cur->reg);
  const TargetRegisterClass *RCLeader = RelatedRegClasses.getLeaderValue(RC);
 
  for (IntervalPtrs::iterator i = inactive_.begin(), e = inactive_.end();
       i != e; ++i) {
    unsigned reg = i->first->reg;
    assert(TargetRegisterInfo::isVirtualRegister(reg) &&
           "Can only allocate virtual registers!");

    // If this is not in a related reg class to the register we're allocating, 
    // don't check it.
    const TargetRegisterClass *RegRC = mri_->getRegClass(reg);
    if (RelatedRegClasses.getLeaderValue(RegRC) == RCLeader) {
      reg = vrm_->getPhys(reg);
      if (inactiveCounts.size() <= reg)
        inactiveCounts.resize(reg+1);
      ++inactiveCounts[reg];
      MaxInactiveCount = std::max(MaxInactiveCount, inactiveCounts[reg]);
    }
  }

  // If copy coalescer has assigned a "preferred" register, check if it's
  // available first.
  if (cur->preference) {
    DOUT << "(preferred: " << tri_->getName(cur->preference) << ") ";
    if (isRegAvail(cur->preference) && 
        RC->contains(cur->preference))
      return cur->preference;
  }

  if (!DowngradedRegs.empty()) {
    unsigned FreeReg = getFreePhysReg(RC, MaxInactiveCount, inactiveCounts,
                                      true);
    if (FreeReg)
      return FreeReg;
  }
  return getFreePhysReg(RC, MaxInactiveCount, inactiveCounts, false);
}

FunctionPass* llvm::createLinearScanRegisterAllocator() {
  return new RALinScan();
}