+#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
+void ReadyQueue::dump() {
+ dbgs() << Name << ": ";
+ for (unsigned i = 0, e = Queue.size(); i < e; ++i)
+ dbgs() << Queue[i]->NodeNum << " ";
+ dbgs() << "\n";
+}
+#endif
+
+//===----------------------------------------------------------------------===//
+// ScheduleDAGMI - Base class for MachineInstr scheduling with LiveIntervals
+// preservation.
+//===----------------------------------------------------------------------===//
+
+/// ReleaseSucc - Decrement the NumPredsLeft count of a successor. When
+/// NumPredsLeft reaches zero, release the successor node.
+///
+/// FIXME: Adjust SuccSU height based on MinLatency.
+void ScheduleDAGMI::releaseSucc(SUnit *SU, SDep *SuccEdge) {
+ SUnit *SuccSU = SuccEdge->getSUnit();
+
+#ifndef NDEBUG
+ if (SuccSU->NumPredsLeft == 0) {
+ dbgs() << "*** Scheduling failed! ***\n";
+ SuccSU->dump(this);
+ dbgs() << " has been released too many times!\n";
+ llvm_unreachable(0);
+ }
+#endif
+ --SuccSU->NumPredsLeft;
+ if (SuccSU->NumPredsLeft == 0 && SuccSU != &ExitSU)
+ SchedImpl->releaseTopNode(SuccSU);
+}
+
+/// releaseSuccessors - Call releaseSucc on each of SU's successors.
+void ScheduleDAGMI::releaseSuccessors(SUnit *SU) {
+ for (SUnit::succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
+ I != E; ++I) {
+ releaseSucc(SU, &*I);
+ }
+}
+
+/// ReleasePred - Decrement the NumSuccsLeft count of a predecessor. When
+/// NumSuccsLeft reaches zero, release the predecessor node.
+///
+/// FIXME: Adjust PredSU height based on MinLatency.
+void ScheduleDAGMI::releasePred(SUnit *SU, SDep *PredEdge) {
+ SUnit *PredSU = PredEdge->getSUnit();
+
+#ifndef NDEBUG
+ if (PredSU->NumSuccsLeft == 0) {
+ dbgs() << "*** Scheduling failed! ***\n";
+ PredSU->dump(this);
+ dbgs() << " has been released too many times!\n";
+ llvm_unreachable(0);
+ }
+#endif
+ --PredSU->NumSuccsLeft;
+ if (PredSU->NumSuccsLeft == 0 && PredSU != &EntrySU)
+ SchedImpl->releaseBottomNode(PredSU);
+}
+
+/// releasePredecessors - Call releasePred on each of SU's predecessors.
+void ScheduleDAGMI::releasePredecessors(SUnit *SU) {
+ for (SUnit::pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
+ I != E; ++I) {
+ releasePred(SU, &*I);
+ }
+}
+
+void ScheduleDAGMI::moveInstruction(MachineInstr *MI,
+ MachineBasicBlock::iterator InsertPos) {
+ // Advance RegionBegin if the first instruction moves down.
+ if (&*RegionBegin == MI)
+ ++RegionBegin;
+
+ // Update the instruction stream.
+ BB->splice(InsertPos, BB, MI);
+
+ // Update LiveIntervals
+ LIS->handleMove(MI);
+
+ // Recede RegionBegin if an instruction moves above the first.
+ if (RegionBegin == InsertPos)
+ RegionBegin = MI;
+}
+
+bool ScheduleDAGMI::checkSchedLimit() {
+#ifndef NDEBUG
+ if (NumInstrsScheduled == MISchedCutoff && MISchedCutoff != ~0U) {
+ CurrentTop = CurrentBottom;
+ return false;
+ }
+ ++NumInstrsScheduled;
+#endif
+ return true;
+}
+
+/// enterRegion - Called back from MachineScheduler::runOnMachineFunction after
+/// crossing a scheduling boundary. [begin, end) includes all instructions in
+/// the region, including the boundary itself and single-instruction regions
+/// that don't get scheduled.
+void ScheduleDAGMI::enterRegion(MachineBasicBlock *bb,
+ MachineBasicBlock::iterator begin,
+ MachineBasicBlock::iterator end,
+ unsigned endcount)
+{
+ ScheduleDAGInstrs::enterRegion(bb, begin, end, endcount);
+
+ // For convenience remember the end of the liveness region.
+ LiveRegionEnd =
+ (RegionEnd == bb->end()) ? RegionEnd : llvm::next(RegionEnd);
+}
+
+// Setup the register pressure trackers for the top scheduled top and bottom
+// scheduled regions.
+void ScheduleDAGMI::initRegPressure() {
+ TopRPTracker.init(&MF, RegClassInfo, LIS, BB, RegionBegin);
+ BotRPTracker.init(&MF, RegClassInfo, LIS, BB, LiveRegionEnd);
+
+ // Close the RPTracker to finalize live ins.
+ RPTracker.closeRegion();
+
+ DEBUG(RPTracker.getPressure().dump(TRI));
+
+ // Initialize the live ins and live outs.
+ TopRPTracker.addLiveRegs(RPTracker.getPressure().LiveInRegs);
+ BotRPTracker.addLiveRegs(RPTracker.getPressure().LiveOutRegs);
+
+ // Close one end of the tracker so we can call
+ // getMaxUpward/DownwardPressureDelta before advancing across any
+ // instructions. This converts currently live regs into live ins/outs.
+ TopRPTracker.closeTop();
+ BotRPTracker.closeBottom();
+
+ // Account for liveness generated by the region boundary.
+ if (LiveRegionEnd != RegionEnd)
+ BotRPTracker.recede();
+
+ assert(BotRPTracker.getPos() == RegionEnd && "Can't find the region bottom");
+
+ // Cache the list of excess pressure sets in this region. This will also track
+ // the max pressure in the scheduled code for these sets.
+ RegionCriticalPSets.clear();
+ std::vector<unsigned> RegionPressure = RPTracker.getPressure().MaxSetPressure;
+ for (unsigned i = 0, e = RegionPressure.size(); i < e; ++i) {
+ unsigned Limit = TRI->getRegPressureSetLimit(i);
+ DEBUG(dbgs() << TRI->getRegPressureSetName(i)
+ << "Limit " << Limit
+ << " Actual " << RegionPressure[i] << "\n");
+ if (RegionPressure[i] > Limit)
+ RegionCriticalPSets.push_back(PressureElement(i, 0));
+ }
+ DEBUG(dbgs() << "Excess PSets: ";
+ for (unsigned i = 0, e = RegionCriticalPSets.size(); i != e; ++i)
+ dbgs() << TRI->getRegPressureSetName(
+ RegionCriticalPSets[i].PSetID) << " ";
+ dbgs() << "\n");
+}
+
+// FIXME: When the pressure tracker deals in pressure differences then we won't
+// iterate over all RegionCriticalPSets[i].
+void ScheduleDAGMI::
+updateScheduledPressure(std::vector<unsigned> NewMaxPressure) {
+ for (unsigned i = 0, e = RegionCriticalPSets.size(); i < e; ++i) {
+ unsigned ID = RegionCriticalPSets[i].PSetID;
+ int &MaxUnits = RegionCriticalPSets[i].UnitIncrease;
+ if ((int)NewMaxPressure[ID] > MaxUnits)
+ MaxUnits = NewMaxPressure[ID];
+ }
+}
+
+// Release all DAG roots for scheduling.
+void ScheduleDAGMI::releaseRoots() {
+ SmallVector<SUnit*, 16> BotRoots;
+
+ for (std::vector<SUnit>::iterator
+ I = SUnits.begin(), E = SUnits.end(); I != E; ++I) {
+ // A SUnit is ready to top schedule if it has no predecessors.
+ if (I->Preds.empty())
+ SchedImpl->releaseTopNode(&(*I));
+ // A SUnit is ready to bottom schedule if it has no successors.
+ if (I->Succs.empty())
+ BotRoots.push_back(&(*I));
+ }
+ // Release bottom roots in reverse order so the higher priority nodes appear
+ // first. This is more natural and slightly more efficient.
+ for (SmallVectorImpl<SUnit*>::const_reverse_iterator
+ I = BotRoots.rbegin(), E = BotRoots.rend(); I != E; ++I)
+ SchedImpl->releaseBottomNode(*I);
+}
+
+/// schedule - Called back from MachineScheduler::runOnMachineFunction
+/// after setting up the current scheduling region. [RegionBegin, RegionEnd)
+/// only includes instructions that have DAG nodes, not scheduling boundaries.
+///
+/// This is a skeletal driver, with all the functionality pushed into helpers,
+/// so that it can be easilly extended by experimental schedulers. Generally,
+/// implementing MachineSchedStrategy should be sufficient to implement a new
+/// scheduling algorithm. However, if a scheduler further subclasses
+/// ScheduleDAGMI then it will want to override this virtual method in order to
+/// update any specialized state.
+void ScheduleDAGMI::schedule() {
+ buildDAGWithRegPressure();
+
+ postprocessDAG();
+
+ DEBUG(for (unsigned su = 0, e = SUnits.size(); su != e; ++su)
+ SUnits[su].dumpAll(this));
+
+ if (ViewMISchedDAGs) viewGraph();
+
+ initQueues();
+
+ bool IsTopNode = false;
+ while (SUnit *SU = SchedImpl->pickNode(IsTopNode)) {
+ assert(!SU->isScheduled && "Node already scheduled");
+ if (!checkSchedLimit())
+ break;
+
+ scheduleMI(SU, IsTopNode);
+
+ updateQueues(SU, IsTopNode);
+ }
+ assert(CurrentTop == CurrentBottom && "Nonempty unscheduled zone.");
+
+ placeDebugValues();
+}
+
+/// Build the DAG and setup three register pressure trackers.
+void ScheduleDAGMI::buildDAGWithRegPressure() {
+ // Initialize the register pressure tracker used by buildSchedGraph.
+ RPTracker.init(&MF, RegClassInfo, LIS, BB, LiveRegionEnd);
+
+ // Account for liveness generate by the region boundary.
+ if (LiveRegionEnd != RegionEnd)
+ RPTracker.recede();
+
+ // Build the DAG, and compute current register pressure.
+ buildSchedGraph(AA, &RPTracker);
+ if (ViewMISchedDAGs) viewGraph();
+
+ // Initialize top/bottom trackers after computing region pressure.
+ initRegPressure();
+}
+
+/// Apply each ScheduleDAGMutation step in order.
+void ScheduleDAGMI::postprocessDAG() {
+ for (unsigned i = 0, e = Mutations.size(); i < e; ++i) {
+ Mutations[i]->apply(this);
+ }
+}
+
+/// Identify DAG roots and setup scheduler queues.
+void ScheduleDAGMI::initQueues() {
+ // Initialize the strategy before modifying the DAG.
+ SchedImpl->initialize(this);
+
+ // Release edges from the special Entry node or to the special Exit node.
+ releaseSuccessors(&EntrySU);
+ releasePredecessors(&ExitSU);
+
+ // Release all DAG roots for scheduling.
+ releaseRoots();
+
+ CurrentTop = nextIfDebug(RegionBegin, RegionEnd);
+ CurrentBottom = RegionEnd;
+}
+
+/// Move an instruction and update register pressure.
+void ScheduleDAGMI::scheduleMI(SUnit *SU, bool IsTopNode) {
+ // Move the instruction to its new location in the instruction stream.
+ MachineInstr *MI = SU->getInstr();
+
+ if (IsTopNode) {
+ assert(SU->isTopReady() && "node still has unscheduled dependencies");
+ if (&*CurrentTop == MI)
+ CurrentTop = nextIfDebug(++CurrentTop, CurrentBottom);
+ else {
+ moveInstruction(MI, CurrentTop);
+ TopRPTracker.setPos(MI);
+ }
+
+ // Update top scheduled pressure.
+ TopRPTracker.advance();
+ assert(TopRPTracker.getPos() == CurrentTop && "out of sync");
+ updateScheduledPressure(TopRPTracker.getPressure().MaxSetPressure);
+ }
+ else {
+ assert(SU->isBottomReady() && "node still has unscheduled dependencies");
+ MachineBasicBlock::iterator priorII =
+ priorNonDebug(CurrentBottom, CurrentTop);
+ if (&*priorII == MI)
+ CurrentBottom = priorII;
+ else {
+ if (&*CurrentTop == MI) {
+ CurrentTop = nextIfDebug(++CurrentTop, priorII);
+ TopRPTracker.setPos(CurrentTop);
+ }
+ moveInstruction(MI, CurrentBottom);
+ CurrentBottom = MI;
+ }
+ // Update bottom scheduled pressure.
+ BotRPTracker.recede();
+ assert(BotRPTracker.getPos() == CurrentBottom && "out of sync");
+ updateScheduledPressure(BotRPTracker.getPressure().MaxSetPressure);
+ }
+}
+
+/// Update scheduler queues after scheduling an instruction.
+void ScheduleDAGMI::updateQueues(SUnit *SU, bool IsTopNode) {
+ // Release dependent instructions for scheduling.
+ if (IsTopNode)
+ releaseSuccessors(SU);
+ else
+ releasePredecessors(SU);
+
+ SU->isScheduled = true;
+
+ // Notify the scheduling strategy after updating the DAG.
+ SchedImpl->schedNode(SU, IsTopNode);
+}
+
+/// Reinsert any remaining debug_values, just like the PostRA scheduler.
+void ScheduleDAGMI::placeDebugValues() {
+ // If first instruction was a DBG_VALUE then put it back.
+ if (FirstDbgValue) {
+ BB->splice(RegionBegin, BB, FirstDbgValue);
+ RegionBegin = FirstDbgValue;
+ }
+
+ for (std::vector<std::pair<MachineInstr *, MachineInstr *> >::iterator
+ DI = DbgValues.end(), DE = DbgValues.begin(); DI != DE; --DI) {
+ std::pair<MachineInstr *, MachineInstr *> P = *prior(DI);
+ MachineInstr *DbgValue = P.first;
+ MachineBasicBlock::iterator OrigPrevMI = P.second;
+ BB->splice(++OrigPrevMI, BB, DbgValue);
+ if (OrigPrevMI == llvm::prior(RegionEnd))
+ RegionEnd = DbgValue;
+ }
+ DbgValues.clear();
+ FirstDbgValue = NULL;
+}
+
+//===----------------------------------------------------------------------===//
+// ConvergingScheduler - Implementation of the standard MachineSchedStrategy.
+//===----------------------------------------------------------------------===//
+
+namespace {
+/// ConvergingScheduler shrinks the unscheduled zone using heuristics to balance
+/// the schedule.
+class ConvergingScheduler : public MachineSchedStrategy {
+
+ /// Store the state used by ConvergingScheduler heuristics, required for the
+ /// lifetime of one invocation of pickNode().
+ struct SchedCandidate {
+ // The best SUnit candidate.
+ SUnit *SU;
+
+ // Register pressure values for the best candidate.
+ RegPressureDelta RPDelta;
+
+ SchedCandidate(): SU(NULL) {}
+ };
+ /// Represent the type of SchedCandidate found within a single queue.
+ enum CandResult {
+ NoCand, NodeOrder, SingleExcess, SingleCritical, SingleMax, MultiPressure };
+
+ /// Each Scheduling boundary is associated with ready queues. It tracks the
+ /// current cycle in whichever direction at has moved, and maintains the state
+ /// of "hazards" and other interlocks at the current cycle.
+ struct SchedBoundary {
+ ScheduleDAGMI *DAG;
+ const TargetSchedModel *SchedModel;
+
+ ReadyQueue Available;
+ ReadyQueue Pending;
+ bool CheckPending;
+
+ ScheduleHazardRecognizer *HazardRec;
+
+ unsigned CurrCycle;
+ unsigned IssueCount;
+
+ /// MinReadyCycle - Cycle of the soonest available instruction.
+ unsigned MinReadyCycle;
+
+ // Remember the greatest min operand latency.
+ unsigned MaxMinLatency;
+
+ /// Pending queues extend the ready queues with the same ID and the
+ /// PendingFlag set.
+ SchedBoundary(unsigned ID, const Twine &Name):
+ DAG(0), SchedModel(0), Available(ID, Name+".A"),
+ Pending(ID << ConvergingScheduler::LogMaxQID, Name+".P"),
+ CheckPending(false), HazardRec(0), CurrCycle(0), IssueCount(0),
+ MinReadyCycle(UINT_MAX), MaxMinLatency(0) {}
+
+ ~SchedBoundary() { delete HazardRec; }
+
+ void init(ScheduleDAGMI *dag, const TargetSchedModel *smodel) {
+ DAG = dag;
+ SchedModel = smodel;
+ }
+
+ bool isTop() const {
+ return Available.getID() == ConvergingScheduler::TopQID;
+ }
+
+ bool checkHazard(SUnit *SU);
+
+ void releaseNode(SUnit *SU, unsigned ReadyCycle);
+
+ void bumpCycle();
+
+ void bumpNode(SUnit *SU);
+
+ void releasePending();
+
+ void removeReady(SUnit *SU);
+
+ SUnit *pickOnlyChoice();
+ };
+
+ ScheduleDAGMI *DAG;
+ const TargetSchedModel *SchedModel;
+ const TargetRegisterInfo *TRI;
+
+ // State of the top and bottom scheduled instruction boundaries.
+ SchedBoundary Top;
+ SchedBoundary Bot;
+
+public:
+ /// SUnit::NodeQueueId: 0 (none), 1 (top), 2 (bot), 3 (both)
+ enum {
+ TopQID = 1,
+ BotQID = 2,
+ LogMaxQID = 2
+ };
+
+ ConvergingScheduler():
+ DAG(0), SchedModel(0), TRI(0), Top(TopQID, "TopQ"), Bot(BotQID, "BotQ") {}
+
+ virtual void initialize(ScheduleDAGMI *dag);
+
+ virtual SUnit *pickNode(bool &IsTopNode);
+
+ virtual void schedNode(SUnit *SU, bool IsTopNode);
+
+ virtual void releaseTopNode(SUnit *SU);
+
+ virtual void releaseBottomNode(SUnit *SU);
+
+protected:
+ SUnit *pickNodeBidrectional(bool &IsTopNode);
+
+ CandResult pickNodeFromQueue(ReadyQueue &Q,
+ const RegPressureTracker &RPTracker,
+ SchedCandidate &Candidate);
+#ifndef NDEBUG
+ void traceCandidate(const char *Label, const ReadyQueue &Q, SUnit *SU,
+ PressureElement P = PressureElement());
+#endif
+};
+} // namespace
+
+void ConvergingScheduler::initialize(ScheduleDAGMI *dag) {
+ DAG = dag;
+ SchedModel = DAG->getSchedModel();
+ TRI = DAG->TRI;
+ Top.init(DAG, SchedModel);
+ Bot.init(DAG, SchedModel);
+
+ // Initialize the HazardRecognizers. If itineraries don't exist, are empty, or
+ // are disabled, then these HazardRecs will be disabled.
+ const InstrItineraryData *Itin = SchedModel->getInstrItineraries();
+ const TargetMachine &TM = DAG->MF.getTarget();
+ Top.HazardRec = TM.getInstrInfo()->CreateTargetMIHazardRecognizer(Itin, DAG);
+ Bot.HazardRec = TM.getInstrInfo()->CreateTargetMIHazardRecognizer(Itin, DAG);
+
+ assert((!ForceTopDown || !ForceBottomUp) &&
+ "-misched-topdown incompatible with -misched-bottomup");
+}
+
+void ConvergingScheduler::releaseTopNode(SUnit *SU) {
+ if (SU->isScheduled)
+ return;
+
+ for (SUnit::succ_iterator I = SU->Preds.begin(), E = SU->Preds.end();
+ I != E; ++I) {
+ unsigned PredReadyCycle = I->getSUnit()->TopReadyCycle;
+ unsigned MinLatency = I->getMinLatency();
+#ifndef NDEBUG
+ Top.MaxMinLatency = std::max(MinLatency, Top.MaxMinLatency);
+#endif
+ if (SU->TopReadyCycle < PredReadyCycle + MinLatency)
+ SU->TopReadyCycle = PredReadyCycle + MinLatency;
+ }
+ Top.releaseNode(SU, SU->TopReadyCycle);
+}
+
+void ConvergingScheduler::releaseBottomNode(SUnit *SU) {
+ if (SU->isScheduled)
+ return;
+
+ assert(SU->getInstr() && "Scheduled SUnit must have instr");
+
+ for (SUnit::succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
+ I != E; ++I) {
+ unsigned SuccReadyCycle = I->getSUnit()->BotReadyCycle;
+ unsigned MinLatency = I->getMinLatency();
+#ifndef NDEBUG
+ Bot.MaxMinLatency = std::max(MinLatency, Bot.MaxMinLatency);
+#endif
+ if (SU->BotReadyCycle < SuccReadyCycle + MinLatency)
+ SU->BotReadyCycle = SuccReadyCycle + MinLatency;
+ }
+ Bot.releaseNode(SU, SU->BotReadyCycle);
+}
+
+/// Does this SU have a hazard within the current instruction group.
+///
+/// The scheduler supports two modes of hazard recognition. The first is the
+/// ScheduleHazardRecognizer API. It is a fully general hazard recognizer that
+/// supports highly complicated in-order reservation tables
+/// (ScoreboardHazardRecognizer) and arbitraty target-specific logic.
+///
+/// The second is a streamlined mechanism that checks for hazards based on
+/// simple counters that the scheduler itself maintains. It explicitly checks
+/// for instruction dispatch limitations, including the number of micro-ops that
+/// can dispatch per cycle.
+///
+/// TODO: Also check whether the SU must start a new group.
+bool ConvergingScheduler::SchedBoundary::checkHazard(SUnit *SU) {
+ if (HazardRec->isEnabled())
+ return HazardRec->getHazardType(SU) != ScheduleHazardRecognizer::NoHazard;
+
+ unsigned uops = SchedModel->getNumMicroOps(SU->getInstr());
+ if (IssueCount + uops > SchedModel->getIssueWidth())
+ return true;
+
+ return false;
+}
+
+void ConvergingScheduler::SchedBoundary::releaseNode(SUnit *SU,
+ unsigned ReadyCycle) {
+ if (ReadyCycle < MinReadyCycle)
+ MinReadyCycle = ReadyCycle;
+
+ // Check for interlocks first. For the purpose of other heuristics, an
+ // instruction that cannot issue appears as if it's not in the ReadyQueue.
+ if (ReadyCycle > CurrCycle || checkHazard(SU))
+ Pending.push(SU);
+ else
+ Available.push(SU);
+}
+
+/// Move the boundary of scheduled code by one cycle.
+void ConvergingScheduler::SchedBoundary::bumpCycle() {
+ unsigned Width = SchedModel->getIssueWidth();
+ IssueCount = (IssueCount <= Width) ? 0 : IssueCount - Width;
+
+ assert(MinReadyCycle < UINT_MAX && "MinReadyCycle uninitialized");
+ unsigned NextCycle = std::max(CurrCycle + 1, MinReadyCycle);
+
+ if (!HazardRec->isEnabled()) {
+ // Bypass HazardRec virtual calls.
+ CurrCycle = NextCycle;
+ }
+ else {
+ // Bypass getHazardType calls in case of long latency.
+ for (; CurrCycle != NextCycle; ++CurrCycle) {
+ if (isTop())
+ HazardRec->AdvanceCycle();
+ else
+ HazardRec->RecedeCycle();
+ }
+ }
+ CheckPending = true;
+
+ DEBUG(dbgs() << "*** " << Available.getName() << " cycle "
+ << CurrCycle << '\n');
+}
+
+/// Move the boundary of scheduled code by one SUnit.
+void ConvergingScheduler::SchedBoundary::bumpNode(SUnit *SU) {
+ // Update the reservation table.
+ if (HazardRec->isEnabled()) {
+ if (!isTop() && SU->isCall) {
+ // Calls are scheduled with their preceding instructions. For bottom-up
+ // scheduling, clear the pipeline state before emitting.
+ HazardRec->Reset();
+ }
+ HazardRec->EmitInstruction(SU);
+ }
+ // Check the instruction group dispatch limit.
+ // TODO: Check if this SU must end a dispatch group.
+ IssueCount += SchedModel->getNumMicroOps(SU->getInstr());
+ if (IssueCount >= SchedModel->getIssueWidth()) {
+ DEBUG(dbgs() << "*** Max instrs at cycle " << CurrCycle << '\n');
+ bumpCycle();
+ }
+}
+
+/// Release pending ready nodes in to the available queue. This makes them
+/// visible to heuristics.
+void ConvergingScheduler::SchedBoundary::releasePending() {
+ // If the available queue is empty, it is safe to reset MinReadyCycle.
+ if (Available.empty())
+ MinReadyCycle = UINT_MAX;
+
+ // Check to see if any of the pending instructions are ready to issue. If
+ // so, add them to the available queue.
+ for (unsigned i = 0, e = Pending.size(); i != e; ++i) {
+ SUnit *SU = *(Pending.begin()+i);
+ unsigned ReadyCycle = isTop() ? SU->TopReadyCycle : SU->BotReadyCycle;
+
+ if (ReadyCycle < MinReadyCycle)
+ MinReadyCycle = ReadyCycle;
+
+ if (ReadyCycle > CurrCycle)
+ continue;
+
+ if (checkHazard(SU))
+ continue;
+
+ Available.push(SU);
+ Pending.remove(Pending.begin()+i);
+ --i; --e;
+ }
+ CheckPending = false;
+}
+
+/// Remove SU from the ready set for this boundary.
+void ConvergingScheduler::SchedBoundary::removeReady(SUnit *SU) {
+ if (Available.isInQueue(SU))
+ Available.remove(Available.find(SU));
+ else {
+ assert(Pending.isInQueue(SU) && "bad ready count");
+ Pending.remove(Pending.find(SU));
+ }
+}
+
+/// If this queue only has one ready candidate, return it. As a side effect,
+/// advance the cycle until at least one node is ready. If multiple instructions
+/// are ready, return NULL.
+SUnit *ConvergingScheduler::SchedBoundary::pickOnlyChoice() {
+ if (CheckPending)
+ releasePending();
+
+ for (unsigned i = 0; Available.empty(); ++i) {
+ assert(i <= (HazardRec->getMaxLookAhead() + MaxMinLatency) &&
+ "permanent hazard"); (void)i;
+ bumpCycle();
+ releasePending();
+ }
+ if (Available.size() == 1)
+ return *Available.begin();
+ return NULL;
+}
+
+#ifndef NDEBUG
+void ConvergingScheduler::traceCandidate(const char *Label, const ReadyQueue &Q,
+ SUnit *SU, PressureElement P) {
+ dbgs() << Label << " " << Q.getName() << " ";
+ if (P.isValid())
+ dbgs() << TRI->getRegPressureSetName(P.PSetID) << ":" << P.UnitIncrease
+ << " ";
+ else
+ dbgs() << " ";
+ SU->dump(DAG);
+}
+#endif
+
+/// pickNodeFromQueue helper that returns true if the LHS reg pressure effect is
+/// more desirable than RHS from scheduling standpoint.
+static bool compareRPDelta(const RegPressureDelta &LHS,
+ const RegPressureDelta &RHS) {
+ // Compare each component of pressure in decreasing order of importance
+ // without checking if any are valid. Invalid PressureElements are assumed to
+ // have UnitIncrease==0, so are neutral.
+
+ // Avoid increasing the max critical pressure in the scheduled region.
+ if (LHS.Excess.UnitIncrease != RHS.Excess.UnitIncrease)
+ return LHS.Excess.UnitIncrease < RHS.Excess.UnitIncrease;
+
+ // Avoid increasing the max critical pressure in the scheduled region.
+ if (LHS.CriticalMax.UnitIncrease != RHS.CriticalMax.UnitIncrease)
+ return LHS.CriticalMax.UnitIncrease < RHS.CriticalMax.UnitIncrease;
+
+ // Avoid increasing the max pressure of the entire region.
+ if (LHS.CurrentMax.UnitIncrease != RHS.CurrentMax.UnitIncrease)
+ return LHS.CurrentMax.UnitIncrease < RHS.CurrentMax.UnitIncrease;
+
+ return false;
+}
+
+/// Pick the best candidate from the top queue.
+///
+/// TODO: getMaxPressureDelta results can be mostly cached for each SUnit during
+/// DAG building. To adjust for the current scheduling location we need to
+/// maintain the number of vreg uses remaining to be top-scheduled.
+ConvergingScheduler::CandResult ConvergingScheduler::
+pickNodeFromQueue(ReadyQueue &Q, const RegPressureTracker &RPTracker,
+ SchedCandidate &Candidate) {
+ DEBUG(Q.dump());
+
+ // getMaxPressureDelta temporarily modifies the tracker.
+ RegPressureTracker &TempTracker = const_cast<RegPressureTracker&>(RPTracker);
+
+ // BestSU remains NULL if no top candidates beat the best existing candidate.
+ CandResult FoundCandidate = NoCand;
+ for (ReadyQueue::iterator I = Q.begin(), E = Q.end(); I != E; ++I) {
+ RegPressureDelta RPDelta;
+ TempTracker.getMaxPressureDelta((*I)->getInstr(), RPDelta,
+ DAG->getRegionCriticalPSets(),
+ DAG->getRegPressure().MaxSetPressure);
+
+ // Initialize the candidate if needed.
+ if (!Candidate.SU) {
+ Candidate.SU = *I;
+ Candidate.RPDelta = RPDelta;
+ FoundCandidate = NodeOrder;
+ continue;
+ }
+ // Avoid exceeding the target's limit.
+ if (RPDelta.Excess.UnitIncrease < Candidate.RPDelta.Excess.UnitIncrease) {
+ DEBUG(traceCandidate("ECAND", Q, *I, RPDelta.Excess));
+ Candidate.SU = *I;
+ Candidate.RPDelta = RPDelta;
+ FoundCandidate = SingleExcess;
+ continue;
+ }
+ if (RPDelta.Excess.UnitIncrease > Candidate.RPDelta.Excess.UnitIncrease)
+ continue;
+ if (FoundCandidate == SingleExcess)
+ FoundCandidate = MultiPressure;
+
+ // Avoid increasing the max critical pressure in the scheduled region.
+ if (RPDelta.CriticalMax.UnitIncrease
+ < Candidate.RPDelta.CriticalMax.UnitIncrease) {
+ DEBUG(traceCandidate("PCAND", Q, *I, RPDelta.CriticalMax));
+ Candidate.SU = *I;
+ Candidate.RPDelta = RPDelta;
+ FoundCandidate = SingleCritical;
+ continue;
+ }
+ if (RPDelta.CriticalMax.UnitIncrease
+ > Candidate.RPDelta.CriticalMax.UnitIncrease)
+ continue;
+ if (FoundCandidate == SingleCritical)
+ FoundCandidate = MultiPressure;
+
+ // Avoid increasing the max pressure of the entire region.
+ if (RPDelta.CurrentMax.UnitIncrease
+ < Candidate.RPDelta.CurrentMax.UnitIncrease) {
+ DEBUG(traceCandidate("MCAND", Q, *I, RPDelta.CurrentMax));
+ Candidate.SU = *I;
+ Candidate.RPDelta = RPDelta;
+ FoundCandidate = SingleMax;
+ continue;
+ }
+ if (RPDelta.CurrentMax.UnitIncrease
+ > Candidate.RPDelta.CurrentMax.UnitIncrease)
+ continue;
+ if (FoundCandidate == SingleMax)
+ FoundCandidate = MultiPressure;
+
+ // Fall through to original instruction order.
+ // Only consider node order if Candidate was chosen from this Q.
+ if (FoundCandidate == NoCand)
+ continue;
+
+ if ((Q.getID() == TopQID && (*I)->NodeNum < Candidate.SU->NodeNum)
+ || (Q.getID() == BotQID && (*I)->NodeNum > Candidate.SU->NodeNum)) {
+ DEBUG(traceCandidate("NCAND", Q, *I));
+ Candidate.SU = *I;
+ Candidate.RPDelta = RPDelta;
+ FoundCandidate = NodeOrder;
+ }
+ }
+ return FoundCandidate;
+}
+
+/// Pick the best candidate node from either the top or bottom queue.
+SUnit *ConvergingScheduler::pickNodeBidrectional(bool &IsTopNode) {
+ // Schedule as far as possible in the direction of no choice. This is most
+ // efficient, but also provides the best heuristics for CriticalPSets.
+ if (SUnit *SU = Bot.pickOnlyChoice()) {
+ IsTopNode = false;
+ return SU;
+ }
+ if (SUnit *SU = Top.pickOnlyChoice()) {
+ IsTopNode = true;
+ return SU;
+ }
+ SchedCandidate BotCand;
+ // Prefer bottom scheduling when heuristics are silent.
+ CandResult BotResult = pickNodeFromQueue(Bot.Available,
+ DAG->getBotRPTracker(), BotCand);
+ assert(BotResult != NoCand && "failed to find the first candidate");
+
+ // If either Q has a single candidate that provides the least increase in
+ // Excess pressure, we can immediately schedule from that Q.
+ //
+ // RegionCriticalPSets summarizes the pressure within the scheduled region and
+ // affects picking from either Q. If scheduling in one direction must
+ // increase pressure for one of the excess PSets, then schedule in that
+ // direction first to provide more freedom in the other direction.
+ if (BotResult == SingleExcess || BotResult == SingleCritical) {
+ IsTopNode = false;
+ return BotCand.SU;
+ }
+ // Check if the top Q has a better candidate.
+ SchedCandidate TopCand;
+ CandResult TopResult = pickNodeFromQueue(Top.Available,
+ DAG->getTopRPTracker(), TopCand);
+ assert(TopResult != NoCand && "failed to find the first candidate");
+
+ if (TopResult == SingleExcess || TopResult == SingleCritical) {
+ IsTopNode = true;
+ return TopCand.SU;
+ }
+ // If either Q has a single candidate that minimizes pressure above the
+ // original region's pressure pick it.
+ if (BotResult == SingleMax) {
+ IsTopNode = false;
+ return BotCand.SU;
+ }
+ if (TopResult == SingleMax) {
+ IsTopNode = true;
+ return TopCand.SU;
+ }
+ // Check for a salient pressure difference and pick the best from either side.
+ if (compareRPDelta(TopCand.RPDelta, BotCand.RPDelta)) {
+ IsTopNode = true;
+ return TopCand.SU;
+ }
+ // Otherwise prefer the bottom candidate in node order.
+ IsTopNode = false;
+ return BotCand.SU;
+}
+
+/// Pick the best node to balance the schedule. Implements MachineSchedStrategy.
+SUnit *ConvergingScheduler::pickNode(bool &IsTopNode) {
+ if (DAG->top() == DAG->bottom()) {
+ assert(Top.Available.empty() && Top.Pending.empty() &&
+ Bot.Available.empty() && Bot.Pending.empty() && "ReadyQ garbage");
+ return NULL;
+ }
+ SUnit *SU;
+ do {
+ if (ForceTopDown) {
+ SU = Top.pickOnlyChoice();
+ if (!SU) {
+ SchedCandidate TopCand;
+ CandResult TopResult =
+ pickNodeFromQueue(Top.Available, DAG->getTopRPTracker(), TopCand);
+ assert(TopResult != NoCand && "failed to find the first candidate");
+ (void)TopResult;
+ SU = TopCand.SU;
+ }
+ IsTopNode = true;
+ }
+ else if (ForceBottomUp) {
+ SU = Bot.pickOnlyChoice();
+ if (!SU) {
+ SchedCandidate BotCand;
+ CandResult BotResult =
+ pickNodeFromQueue(Bot.Available, DAG->getBotRPTracker(), BotCand);
+ assert(BotResult != NoCand && "failed to find the first candidate");
+ (void)BotResult;
+ SU = BotCand.SU;
+ }
+ IsTopNode = false;
+ }
+ else {
+ SU = pickNodeBidrectional(IsTopNode);
+ }
+ } while (SU->isScheduled);
+
+ if (SU->isTopReady())
+ Top.removeReady(SU);
+ if (SU->isBottomReady())
+ Bot.removeReady(SU);
+
+ DEBUG(dbgs() << "*** " << (IsTopNode ? "Top" : "Bottom")
+ << " Scheduling Instruction in cycle "
+ << (IsTopNode ? Top.CurrCycle : Bot.CurrCycle) << '\n';
+ SU->dump(DAG));
+ return SU;
+}
+
+/// Update the scheduler's state after scheduling a node. This is the same node
+/// that was just returned by pickNode(). However, ScheduleDAGMI needs to update
+/// it's state based on the current cycle before MachineSchedStrategy does.
+void ConvergingScheduler::schedNode(SUnit *SU, bool IsTopNode) {
+ if (IsTopNode) {
+ SU->TopReadyCycle = Top.CurrCycle;
+ Top.bumpNode(SU);
+ }
+ else {
+ SU->BotReadyCycle = Bot.CurrCycle;
+ Bot.bumpNode(SU);
+ }
+}
+
+/// Create the standard converging machine scheduler. This will be used as the
+/// default scheduler if the target does not set a default.
+static ScheduleDAGInstrs *createConvergingSched(MachineSchedContext *C) {
+ assert((!ForceTopDown || !ForceBottomUp) &&
+ "-misched-topdown incompatible with -misched-bottomup");
+ return new ScheduleDAGMI(C, new ConvergingScheduler());
+}
+static MachineSchedRegistry
+ConvergingSchedRegistry("converge", "Standard converging scheduler.",
+ createConvergingSched);
+