#include "llvm/ADT/PriorityQueue.h"
#include "llvm/Analysis/AliasAnalysis.h"
#include "llvm/CodeGen/LiveIntervalAnalysis.h"
+#include "llvm/CodeGen/MachineDominators.h"
+#include "llvm/CodeGen/MachineLoopInfo.h"
+#include "llvm/CodeGen/MachineRegisterInfo.h"
#include "llvm/CodeGen/Passes.h"
#include "llvm/CodeGen/RegisterClassInfo.h"
#include "llvm/CodeGen/ScheduleDFS.h"
#include "llvm/Support/ErrorHandling.h"
#include "llvm/Support/GraphWriter.h"
#include "llvm/Support/raw_ostream.h"
+#include "llvm/Target/TargetInstrInfo.h"
#include <queue>
using namespace llvm;
static bool ViewMISchedDAGs = false;
#endif // NDEBUG
-// Experimental heuristics
static cl::opt<bool> EnableLoadCluster("misched-cluster", cl::Hidden,
cl::desc("Enable load clustering."), cl::init(true));
static cl::opt<bool> EnableMacroFusion("misched-fusion", cl::Hidden,
cl::desc("Enable scheduling for macro fusion."), cl::init(true));
+static cl::opt<bool> VerifyScheduling("verify-misched", cl::Hidden,
+ cl::desc("Verify machine instrs before and after machine scheduling"));
+
// DAG subtrees must have at least this many nodes.
static const unsigned MinSubtreeSize = 8;
LIS = &getAnalysis<LiveIntervals>();
const TargetInstrInfo *TII = MF->getTarget().getInstrInfo();
+ if (VerifyScheduling) {
+ DEBUG(LIS->print(dbgs()));
+ MF->verify(this, "Before machine scheduling.");
+ }
RegClassInfo->runOnMachineFunction(*MF);
// Select the scheduler, or set the default.
}
Scheduler->finalizeSchedule();
DEBUG(LIS->print(dbgs()));
+ if (VerifyScheduling)
+ MF->verify(this, "After machine scheduling.");
return true;
}
delete SchedImpl;
}
+bool ScheduleDAGMI::canAddEdge(SUnit *SuccSU, SUnit *PredSU) {
+ return SuccSU == &ExitSU || !Topo.IsReachable(PredSU, SuccSU);
+}
+
bool ScheduleDAGMI::addEdge(SUnit *SuccSU, const SDep &PredDep) {
if (SuccSU != &ExitSU) {
// Do not use WillCreateCycle, it assumes SD scheduling.
}
}
+/// This is normally called from the main scheduler loop but may also be invoked
+/// by the scheduling strategy to perform additional code motion.
void ScheduleDAGMI::moveInstruction(MachineInstr *MI,
MachineBasicBlock::iterator InsertPos) {
// Advance RegionBegin if the first instruction moves down.
// 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;
+ const 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)
+ unsigned Limit = RegClassInfo->getRegPressureSetLimit(i);
+ if (RegionPressure[i] > Limit) {
+ DEBUG(dbgs() << TRI->getRegPressureSetName(i)
+ << " Limit " << Limit
+ << " Actual " << RegionPressure[i] << "\n");
RegionCriticalPSets.push_back(PressureElement(i, 0));
+ }
}
DEBUG(dbgs() << "Excess PSets: ";
for (unsigned i = 0, e = RegionCriticalPSets.size(); i != e; ++i)
// 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) {
+updateScheduledPressure(const 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];
}
+ DEBUG(
+ for (unsigned i = 0, e = NewMaxPressure.size(); i < e; ++i) {
+ unsigned Limit = RegClassInfo->getRegPressureSetLimit(i);
+ if (NewMaxPressure[i] > Limit ) {
+ dbgs() << " " << TRI->getRegPressureSetName(i) << ": "
+ << NewMaxPressure[i] << " > " << Limit << "\n";
+ }
+ });
}
/// schedule - Called back from MachineScheduler::runOnMachineFunction
for (std::vector<SUnit>::iterator
I = SUnits.begin(), E = SUnits.end(); I != E; ++I) {
SUnit *SU = &(*I);
+ assert(!SU->isBoundaryNode() && "Boundary node should not be in SUnits");
// Order predecessors so DFSResult follows the critical path.
SU->biasCriticalPath();
// A SUnit is ready to top schedule if it has no predecessors.
- if (!I->NumPredsLeft && SU != &EntrySU)
+ if (!I->NumPredsLeft)
TopRoots.push_back(SU);
// A SUnit is ready to bottom schedule if it has no successors.
- if (!I->NumSuccsLeft && SU != &ExitSU)
+ if (!I->NumSuccsLeft)
BotRoots.push_back(SU);
}
+ ExitSU.biasCriticalPath();
}
/// Identify DAG roots and setup scheduler queues.
}
//===----------------------------------------------------------------------===//
-// ConvergingScheduler - Implementation of the standard MachineSchedStrategy.
+// CopyConstrain - DAG post-processing to encourage copy elimination.
+//===----------------------------------------------------------------------===//
+
+namespace {
+/// \brief Post-process the DAG to create weak edges from all uses of a copy to
+/// the one use that defines the copy's source vreg, most likely an induction
+/// variable increment.
+class CopyConstrain : public ScheduleDAGMutation {
+ // Transient state.
+ SlotIndex RegionBeginIdx;
+ // RegionEndIdx is the slot index of the last non-debug instruction in the
+ // scheduling region. So we may have RegionBeginIdx == RegionEndIdx.
+ SlotIndex RegionEndIdx;
+public:
+ CopyConstrain(const TargetInstrInfo *, const TargetRegisterInfo *) {}
+
+ virtual void apply(ScheduleDAGMI *DAG);
+
+protected:
+ void constrainLocalCopy(SUnit *CopySU, ScheduleDAGMI *DAG);
+};
+} // anonymous
+
+/// constrainLocalCopy handles two possibilities:
+/// 1) Local src:
+/// I0: = dst
+/// I1: src = ...
+/// I2: = dst
+/// I3: dst = src (copy)
+/// (create pred->succ edges I0->I1, I2->I1)
+///
+/// 2) Local copy:
+/// I0: dst = src (copy)
+/// I1: = dst
+/// I2: src = ...
+/// I3: = dst
+/// (create pred->succ edges I1->I2, I3->I2)
+///
+/// Although the MachineScheduler is currently constrained to single blocks,
+/// this algorithm should handle extended blocks. An EBB is a set of
+/// contiguously numbered blocks such that the previous block in the EBB is
+/// always the single predecessor.
+void CopyConstrain::constrainLocalCopy(SUnit *CopySU, ScheduleDAGMI *DAG) {
+ LiveIntervals *LIS = DAG->getLIS();
+ MachineInstr *Copy = CopySU->getInstr();
+
+ // Check for pure vreg copies.
+ unsigned SrcReg = Copy->getOperand(1).getReg();
+ if (!TargetRegisterInfo::isVirtualRegister(SrcReg))
+ return;
+
+ unsigned DstReg = Copy->getOperand(0).getReg();
+ if (!TargetRegisterInfo::isVirtualRegister(DstReg))
+ return;
+
+ // Check if either the dest or source is local. If it's live across a back
+ // edge, it's not local. Note that if both vregs are live across the back
+ // edge, we cannot successfully contrain the copy without cyclic scheduling.
+ unsigned LocalReg = DstReg;
+ unsigned GlobalReg = SrcReg;
+ LiveInterval *LocalLI = &LIS->getInterval(LocalReg);
+ if (!LocalLI->isLocal(RegionBeginIdx, RegionEndIdx)) {
+ LocalReg = SrcReg;
+ GlobalReg = DstReg;
+ LocalLI = &LIS->getInterval(LocalReg);
+ if (!LocalLI->isLocal(RegionBeginIdx, RegionEndIdx))
+ return;
+ }
+ LiveInterval *GlobalLI = &LIS->getInterval(GlobalReg);
+
+ // Find the global segment after the start of the local LI.
+ LiveInterval::iterator GlobalSegment = GlobalLI->find(LocalLI->beginIndex());
+ // If GlobalLI does not overlap LocalLI->start, then a copy directly feeds a
+ // local live range. We could create edges from other global uses to the local
+ // start, but the coalescer should have already eliminated these cases, so
+ // don't bother dealing with it.
+ if (GlobalSegment == GlobalLI->end())
+ return;
+
+ // If GlobalSegment is killed at the LocalLI->start, the call to find()
+ // returned the next global segment. But if GlobalSegment overlaps with
+ // LocalLI->start, then advance to the next segement. If a hole in GlobalLI
+ // exists in LocalLI's vicinity, GlobalSegment will be the end of the hole.
+ if (GlobalSegment->contains(LocalLI->beginIndex()))
+ ++GlobalSegment;
+
+ if (GlobalSegment == GlobalLI->end())
+ return;
+
+ // Check if GlobalLI contains a hole in the vicinity of LocalLI.
+ if (GlobalSegment != GlobalLI->begin()) {
+ // Two address defs have no hole.
+ if (SlotIndex::isSameInstr(llvm::prior(GlobalSegment)->end,
+ GlobalSegment->start)) {
+ return;
+ }
+ // If GlobalLI has a prior segment, it must be live into the EBB. Otherwise
+ // it would be a disconnected component in the live range.
+ assert(llvm::prior(GlobalSegment)->start < LocalLI->beginIndex() &&
+ "Disconnected LRG within the scheduling region.");
+ }
+ MachineInstr *GlobalDef = LIS->getInstructionFromIndex(GlobalSegment->start);
+ if (!GlobalDef)
+ return;
+
+ SUnit *GlobalSU = DAG->getSUnit(GlobalDef);
+ if (!GlobalSU)
+ return;
+
+ // GlobalDef is the bottom of the GlobalLI hole. Open the hole by
+ // constraining the uses of the last local def to precede GlobalDef.
+ SmallVector<SUnit*,8> LocalUses;
+ const VNInfo *LastLocalVN = LocalLI->getVNInfoBefore(LocalLI->endIndex());
+ MachineInstr *LastLocalDef = LIS->getInstructionFromIndex(LastLocalVN->def);
+ SUnit *LastLocalSU = DAG->getSUnit(LastLocalDef);
+ for (SUnit::const_succ_iterator
+ I = LastLocalSU->Succs.begin(), E = LastLocalSU->Succs.end();
+ I != E; ++I) {
+ if (I->getKind() != SDep::Data || I->getReg() != LocalReg)
+ continue;
+ if (I->getSUnit() == GlobalSU)
+ continue;
+ if (!DAG->canAddEdge(GlobalSU, I->getSUnit()))
+ return;
+ LocalUses.push_back(I->getSUnit());
+ }
+ // Open the top of the GlobalLI hole by constraining any earlier global uses
+ // to precede the start of LocalLI.
+ SmallVector<SUnit*,8> GlobalUses;
+ MachineInstr *FirstLocalDef =
+ LIS->getInstructionFromIndex(LocalLI->beginIndex());
+ SUnit *FirstLocalSU = DAG->getSUnit(FirstLocalDef);
+ for (SUnit::const_pred_iterator
+ I = GlobalSU->Preds.begin(), E = GlobalSU->Preds.end(); I != E; ++I) {
+ if (I->getKind() != SDep::Anti || I->getReg() != GlobalReg)
+ continue;
+ if (I->getSUnit() == FirstLocalSU)
+ continue;
+ if (!DAG->canAddEdge(FirstLocalSU, I->getSUnit()))
+ return;
+ GlobalUses.push_back(I->getSUnit());
+ }
+ DEBUG(dbgs() << "Constraining copy SU(" << CopySU->NodeNum << ")\n");
+ // Add the weak edges.
+ for (SmallVectorImpl<SUnit*>::const_iterator
+ I = LocalUses.begin(), E = LocalUses.end(); I != E; ++I) {
+ DEBUG(dbgs() << " Local use SU(" << (*I)->NodeNum << ") -> SU("
+ << GlobalSU->NodeNum << ")\n");
+ DAG->addEdge(GlobalSU, SDep(*I, SDep::Weak));
+ }
+ for (SmallVectorImpl<SUnit*>::const_iterator
+ I = GlobalUses.begin(), E = GlobalUses.end(); I != E; ++I) {
+ DEBUG(dbgs() << " Global use SU(" << (*I)->NodeNum << ") -> SU("
+ << FirstLocalSU->NodeNum << ")\n");
+ DAG->addEdge(FirstLocalSU, SDep(*I, SDep::Weak));
+ }
+}
+
+/// \brief Callback from DAG postProcessing to create weak edges to encourage
+/// copy elimination.
+void CopyConstrain::apply(ScheduleDAGMI *DAG) {
+ MachineBasicBlock::iterator FirstPos = nextIfDebug(DAG->begin(), DAG->end());
+ if (FirstPos == DAG->end())
+ return;
+ RegionBeginIdx = DAG->getLIS()->getInstructionIndex(&*FirstPos);
+ RegionEndIdx = DAG->getLIS()->getInstructionIndex(
+ &*priorNonDebug(DAG->end(), DAG->begin()));
+
+ for (unsigned Idx = 0, End = DAG->SUnits.size(); Idx != End; ++Idx) {
+ SUnit *SU = &DAG->SUnits[Idx];
+ if (!SU->getInstr()->isCopy())
+ continue;
+
+ constrainLocalCopy(SU, DAG);
+ }
+}
+
+//===----------------------------------------------------------------------===//
+// ConvergingScheduler - Implementation of the generic MachineSchedStrategy.
//===----------------------------------------------------------------------===//
namespace {
/// Represent the type of SchedCandidate found within a single queue.
/// pickNodeBidirectional depends on these listed by decreasing priority.
enum CandReason {
- NoCand, SingleExcess, SingleCritical, Cluster,
+ NoCand, PhysRegCopy, RegExcess, RegCritical, Cluster, Weak, RegMax,
ResourceReduce, ResourceDemand, BotHeightReduce, BotPathReduce,
- TopDepthReduce, TopPathReduce, SingleMax, MultiPressure, NextDefUse,
- NodeOrder};
+ TopDepthReduce, TopPathReduce, NextDefUse, NodeOrder};
#ifndef NDEBUG
static const char *getReasonStr(ConvergingScheduler::CandReason Reason);
// The reason for this candidate.
CandReason Reason;
+ // Set of reasons that apply to multiple candidates.
+ uint32_t RepeatReasonSet;
+
// Register pressure values for the best candidate.
RegPressureDelta RPDelta;
SchedResourceDelta ResDelta;
SchedCandidate(const CandPolicy &policy)
- : Policy(policy), SU(NULL), Reason(NoCand) {}
+ : Policy(policy), SU(NULL), Reason(NoCand), RepeatReasonSet(0) {}
bool isValid() const { return SU; }
ResDelta = Best.ResDelta;
}
+ bool isRepeat(CandReason R) { return RepeatReasonSet & (1 << R); }
+ void setRepeat(CandReason R) { RepeatReasonSet |= (1 << R); }
+
void initResourceDelta(const ScheduleDAGMI *DAG,
const TargetSchedModel *SchedModel);
};
// Critical path through the DAG in expected latency.
unsigned CriticalPath;
+ // Scaled count of micro-ops left to schedule.
+ unsigned RemIssueCount;
+
// Unscheduled resources
SmallVector<unsigned, 16> RemainingCounts;
- // Critical resource for the unscheduled zone.
- unsigned CritResIdx;
- // Number of micro-ops left to schedule.
- unsigned RemainingMicroOps;
void reset() {
CriticalPath = 0;
+ RemIssueCount = 0;
RemainingCounts.clear();
- CritResIdx = 0;
- RemainingMicroOps = 0;
}
SchedRemainder() { reset(); }
void init(ScheduleDAGMI *DAG, const TargetSchedModel *SchedModel);
-
- unsigned getMaxRemainingCount(const TargetSchedModel *SchedModel) const {
- if (!SchedModel->hasInstrSchedModel())
- return 0;
-
- return std::max(
- RemainingMicroOps * SchedModel->getMicroOpFactor(),
- RemainingCounts[CritResIdx]);
- }
};
/// Each Scheduling boundary is associated with ready queues. It tracks the
ScheduleHazardRecognizer *HazardRec;
+ /// Number of cycles it takes to issue the instructions scheduled in this
+ /// zone. It is defined as: scheduled-micro-ops / issue-width + stalls.
+ /// See getStalls().
unsigned CurrCycle;
- unsigned IssueCount;
+
+ /// Micro-ops issued in the current cycle
+ unsigned CurrMOps;
/// MinReadyCycle - Cycle of the soonest available instruction.
unsigned MinReadyCycle;
// The expected latency of the critical path in this scheduled zone.
unsigned ExpectedLatency;
- // Resources used in the scheduled zone beyond this boundary.
- SmallVector<unsigned, 16> ResourceCounts;
+ // The latency of dependence chains leading into this zone.
+ // For each node scheduled top-down: DLat = max DLat, N.Depth.
+ // For each cycle scheduled: DLat -= 1.
+ unsigned DependentLatency;
+
+ /// Count the scheduled (issued) micro-ops that can be retired by
+ /// time=CurrCycle assuming the first scheduled instr is retired at time=0.
+ unsigned RetiredMOps;
+
+ // Count scheduled resources that have been executed. Resources are
+ // considered executed if they become ready in the time that it takes to
+ // saturate any resource including the one in question. Counts are scaled
+ // for direct comparison with other resources. Counts ca be compared with
+ // MOps * getMicroOpFactor and Latency * getLatencyFactor.
+ SmallVector<unsigned, 16> ExecutedResCounts;
+
+ /// Cache the max count for a single resource.
+ unsigned MaxExecutedResCount;
// Cache the critical resources ID in this scheduled zone.
- unsigned CritResIdx;
+ unsigned ZoneCritResIdx;
// Is the scheduled region resource limited vs. latency limited.
bool IsResourceLimited;
- unsigned ExpectedCount;
-
#ifndef NDEBUG
- // Remember the greatest min operand latency.
- unsigned MaxMinLatency;
+ // Remember the greatest operand latency as an upper bound on the number of
+ // times we should retry the pending queue because of a hazard.
+ unsigned MaxObservedLatency;
#endif
void reset() {
+ // A new HazardRec is created for each DAG and owned by SchedBoundary.
+ delete HazardRec;
+
Available.clear();
Pending.clear();
CheckPending = false;
NextSUs.clear();
HazardRec = 0;
CurrCycle = 0;
- IssueCount = 0;
+ CurrMOps = 0;
MinReadyCycle = UINT_MAX;
ExpectedLatency = 0;
- ResourceCounts.resize(1);
- assert(!ResourceCounts[0] && "nonzero count for bad resource");
- CritResIdx = 0;
+ DependentLatency = 0;
+ RetiredMOps = 0;
+ MaxExecutedResCount = 0;
+ ZoneCritResIdx = 0;
IsResourceLimited = false;
- ExpectedCount = 0;
#ifndef NDEBUG
- MaxMinLatency = 0;
+ MaxObservedLatency = 0;
#endif
// Reserve a zero-count for invalid CritResIdx.
- ResourceCounts.resize(1);
+ ExecutedResCounts.resize(1);
+ assert(!ExecutedResCounts[0] && "nonzero count for bad resource");
}
/// Pending queues extend the ready queues with the same ID and the
/// PendingFlag set.
SchedBoundary(unsigned ID, const Twine &Name):
DAG(0), SchedModel(0), Rem(0), Available(ID, Name+".A"),
- Pending(ID << ConvergingScheduler::LogMaxQID, Name+".P") {
+ Pending(ID << ConvergingScheduler::LogMaxQID, Name+".P"),
+ HazardRec(0) {
reset();
}
return Available.getID() == ConvergingScheduler::TopQID;
}
+#ifndef NDEBUG
+ const char *getResourceName(unsigned PIdx) {
+ if (!PIdx)
+ return "MOps";
+ return SchedModel->getProcResource(PIdx)->Name;
+ }
+#endif
+
+ /// Get the number of latency cycles "covered" by the scheduled
+ /// instructions. This is the larger of the critical path within the zone
+ /// and the number of cycles required to issue the instructions.
+ unsigned getScheduledLatency() const {
+ return std::max(ExpectedLatency, CurrCycle);
+ }
+
unsigned getUnscheduledLatency(SUnit *SU) const {
- if (isTop())
- return SU->getHeight();
- return SU->getDepth() + SU->Latency;
+ return isTop() ? SU->getHeight() : SU->getDepth();
+ }
+
+ unsigned getResourceCount(unsigned ResIdx) const {
+ return ExecutedResCounts[ResIdx];
}
+ /// Get the scaled count of scheduled micro-ops and resources, including
+ /// executed resources.
unsigned getCriticalCount() const {
- return ResourceCounts[CritResIdx];
+ if (!ZoneCritResIdx)
+ return RetiredMOps * SchedModel->getMicroOpFactor();
+ return getResourceCount(ZoneCritResIdx);
+ }
+
+ /// Get a scaled count for the minimum execution time of the scheduled
+ /// micro-ops that are ready to execute by getExecutedCount. Notice the
+ /// feedback loop.
+ unsigned getExecutedCount() const {
+ return std::max(CurrCycle * SchedModel->getLatencyFactor(),
+ MaxExecutedResCount);
}
bool checkHazard(SUnit *SU);
- void setLatencyPolicy(CandPolicy &Policy);
+ unsigned findMaxLatency(ArrayRef<SUnit*> ReadySUs);
+
+ unsigned getOtherResourceCount(unsigned &OtherCritIdx);
+
+ void setPolicy(CandPolicy &Policy, SchedBoundary &OtherZone);
void releaseNode(SUnit *SU, unsigned ReadyCycle);
- void bumpCycle();
+ void bumpCycle(unsigned NextCycle);
+
+ void incExecutedResources(unsigned PIdx, unsigned Count);
- void countResource(unsigned PIdx, unsigned Cycles);
+ unsigned countResource(unsigned PIdx, unsigned Cycles, unsigned ReadyCycle);
void bumpNode(SUnit *SU);
void removeReady(SUnit *SU);
SUnit *pickOnlyChoice();
+
+#ifndef NDEBUG
+ void dumpScheduledState();
+#endif
};
private:
virtual void registerRoots();
protected:
- void balanceZones(
- ConvergingScheduler::SchedBoundary &CriticalZone,
- ConvergingScheduler::SchedCandidate &CriticalCand,
- ConvergingScheduler::SchedBoundary &OppositeZone,
- ConvergingScheduler::SchedCandidate &OppositeCand);
-
- void checkResourceLimits(ConvergingScheduler::SchedCandidate &TopCand,
- ConvergingScheduler::SchedCandidate &BotCand);
-
void tryCandidate(SchedCandidate &Cand,
SchedCandidate &TryCand,
SchedBoundary &Zone,
const RegPressureTracker &RPTracker,
SchedCandidate &Candidate);
+ void reschedulePhysRegCopies(SUnit *SU, bool isTop);
+
#ifndef NDEBUG
- void traceCandidate(const SchedCandidate &Cand, const SchedBoundary &Zone);
+ void traceCandidate(const SchedCandidate &Cand);
#endif
};
} // namespace
for (std::vector<SUnit>::iterator
I = DAG->SUnits.begin(), E = DAG->SUnits.end(); I != E; ++I) {
const MCSchedClassDesc *SC = DAG->getSchedClass(&*I);
- RemainingMicroOps += SchedModel->getNumMicroOps(I->getInstr(), SC);
+ RemIssueCount += SchedModel->getNumMicroOps(I->getInstr(), SC)
+ * SchedModel->getMicroOpFactor();
for (TargetSchedModel::ProcResIter
PI = SchedModel->getWriteProcResBegin(SC),
PE = SchedModel->getWriteProcResEnd(SC); PI != PE; ++PI) {
RemainingCounts[PIdx] += (Factor * PI->Cycles);
}
}
- for (unsigned PIdx = 0, PEnd = SchedModel->getNumProcResourceKinds();
- PIdx != PEnd; ++PIdx) {
- if ((int)(RemainingCounts[PIdx] - RemainingCounts[CritResIdx])
- >= (int)SchedModel->getLatencyFactor()) {
- CritResIdx = PIdx;
- }
- }
}
void ConvergingScheduler::SchedBoundary::
SchedModel = smodel;
Rem = rem;
if (SchedModel->hasInstrSchedModel())
- ResourceCounts.resize(SchedModel->getNumProcResourceKinds());
+ ExecutedResCounts.resize(SchedModel->getNumProcResourceKinds());
}
void ConvergingScheduler::initialize(ScheduleDAGMI *dag) {
DAG = dag;
SchedModel = DAG->getSchedModel();
TRI = DAG->TRI;
+
Rem.init(DAG, SchedModel);
Top.init(DAG, SchedModel, &Rem);
Bot.init(DAG, SchedModel, &Rem);
- DAG->computeDFSResult();
-
// Initialize resource counts.
// Initialize the HazardRecognizers. If itineraries don't exist, are empty, or
for (SUnit::pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
I != E; ++I) {
+ if (I->isWeak())
+ continue;
unsigned PredReadyCycle = I->getSUnit()->TopReadyCycle;
- unsigned MinLatency = I->getMinLatency();
+ unsigned Latency = I->getLatency();
#ifndef NDEBUG
- Top.MaxMinLatency = std::max(MinLatency, Top.MaxMinLatency);
+ Top.MaxObservedLatency = std::max(Latency, Top.MaxObservedLatency);
#endif
- if (SU->TopReadyCycle < PredReadyCycle + MinLatency)
- SU->TopReadyCycle = PredReadyCycle + MinLatency;
+ if (SU->TopReadyCycle < PredReadyCycle + Latency)
+ SU->TopReadyCycle = PredReadyCycle + Latency;
}
Top.releaseNode(SU, SU->TopReadyCycle);
}
if (I->isWeak())
continue;
unsigned SuccReadyCycle = I->getSUnit()->BotReadyCycle;
- unsigned MinLatency = I->getMinLatency();
+ unsigned Latency = I->getLatency();
#ifndef NDEBUG
- Bot.MaxMinLatency = std::max(MinLatency, Bot.MaxMinLatency);
+ Bot.MaxObservedLatency = std::max(Latency, Bot.MaxObservedLatency);
#endif
- if (SU->BotReadyCycle < SuccReadyCycle + MinLatency)
- SU->BotReadyCycle = SuccReadyCycle + MinLatency;
+ if (SU->BotReadyCycle < SuccReadyCycle + Latency)
+ SU->BotReadyCycle = SuccReadyCycle + Latency;
}
Bot.releaseNode(SU, SU->BotReadyCycle);
}
return HazardRec->getHazardType(SU) != ScheduleHazardRecognizer::NoHazard;
unsigned uops = SchedModel->getNumMicroOps(SU->getInstr());
- if ((IssueCount > 0) && (IssueCount + uops > SchedModel->getIssueWidth())) {
+ if ((CurrMOps > 0) && (CurrMOps + uops > SchedModel->getIssueWidth())) {
DEBUG(dbgs() << " SU(" << SU->NodeNum << ") uops="
<< SchedModel->getNumMicroOps(SU->getInstr()) << '\n');
return true;
return false;
}
-/// Compute the remaining latency to determine whether ILP should be increased.
-void ConvergingScheduler::SchedBoundary::setLatencyPolicy(CandPolicy &Policy) {
- // FIXME: compile time. In all, we visit four queues here one we should only
- // need to visit the one that was last popped if we cache the result.
+// Find the unscheduled node in ReadySUs with the highest latency.
+unsigned ConvergingScheduler::SchedBoundary::
+findMaxLatency(ArrayRef<SUnit*> ReadySUs) {
+ SUnit *LateSU = 0;
unsigned RemLatency = 0;
- for (ReadyQueue::iterator I = Available.begin(), E = Available.end();
+ for (ArrayRef<SUnit*>::iterator I = ReadySUs.begin(), E = ReadySUs.end();
I != E; ++I) {
unsigned L = getUnscheduledLatency(*I);
- if (L > RemLatency)
+ if (L > RemLatency) {
RemLatency = L;
+ LateSU = *I;
+ }
}
- for (ReadyQueue::iterator I = Pending.begin(), E = Pending.end();
- I != E; ++I) {
- unsigned L = getUnscheduledLatency(*I);
- if (L > RemLatency)
- RemLatency = L;
+ if (LateSU) {
+ DEBUG(dbgs() << Available.getName() << " RemLatency SU("
+ << LateSU->NodeNum << ") " << RemLatency << "c\n");
+ }
+ return RemLatency;
+}
+
+// Count resources in this zone and the remaining unscheduled
+// instruction. Return the max count, scaled. Set OtherCritIdx to the critical
+// resource index, or zero if the zone is issue limited.
+unsigned ConvergingScheduler::SchedBoundary::
+getOtherResourceCount(unsigned &OtherCritIdx) {
+ if (!SchedModel->hasInstrSchedModel())
+ return 0;
+
+ unsigned OtherCritCount = Rem->RemIssueCount
+ + (RetiredMOps * SchedModel->getMicroOpFactor());
+ DEBUG(dbgs() << " " << Available.getName() << " + Remain MOps: "
+ << OtherCritCount / SchedModel->getMicroOpFactor() << '\n');
+ OtherCritIdx = 0;
+ for (unsigned PIdx = 1, PEnd = SchedModel->getNumProcResourceKinds();
+ PIdx != PEnd; ++PIdx) {
+ unsigned OtherCount = getResourceCount(PIdx) + Rem->RemainingCounts[PIdx];
+ if (OtherCount > OtherCritCount) {
+ OtherCritCount = OtherCount;
+ OtherCritIdx = PIdx;
+ }
}
- unsigned CriticalPathLimit = Rem->CriticalPath + SchedModel->getILPWindow();
- if (RemLatency + ExpectedLatency >= CriticalPathLimit
- && RemLatency > Rem->getMaxRemainingCount(SchedModel)) {
- Policy.ReduceLatency = true;
- DEBUG(dbgs() << "Increase ILP: " << Available.getName() << '\n');
+ if (OtherCritIdx) {
+ DEBUG(dbgs() << " " << Available.getName() << " + Remain CritRes: "
+ << OtherCritCount / SchedModel->getResourceFactor(OtherCritIdx)
+ << " " << getResourceName(OtherCritIdx) << "\n");
}
+ return OtherCritCount;
+}
+
+/// Set the CandPolicy for this zone given the current resources and latencies
+/// inside and outside the zone.
+void ConvergingScheduler::SchedBoundary::setPolicy(CandPolicy &Policy,
+ SchedBoundary &OtherZone) {
+ // Now that potential stalls have been considered, apply preemptive heuristics
+ // based on the the total latency and resources inside and outside this
+ // zone.
+
+ // Compute remaining latency. We need this both to determine whether the
+ // overall schedule has become latency-limited and whether the instructions
+ // outside this zone are resource or latency limited.
+ //
+ // The "dependent" latency is updated incrementally during scheduling as the
+ // max height/depth of scheduled nodes minus the cycles since it was
+ // scheduled:
+ // DLat = max (N.depth - (CurrCycle - N.ReadyCycle) for N in Zone
+ //
+ // The "independent" latency is the max ready queue depth:
+ // ILat = max N.depth for N in Available|Pending
+ //
+ // RemainingLatency is the greater of independent and dependent latency.
+ unsigned RemLatency = DependentLatency;
+ RemLatency = std::max(RemLatency, findMaxLatency(Available.elements()));
+ RemLatency = std::max(RemLatency, findMaxLatency(Pending.elements()));
+
+ // Compute the critical resource outside the zone.
+ unsigned OtherCritIdx;
+ unsigned OtherCount = OtherZone.getOtherResourceCount(OtherCritIdx);
+
+ bool OtherResLimited = false;
+ if (SchedModel->hasInstrSchedModel()) {
+ unsigned LFactor = SchedModel->getLatencyFactor();
+ OtherResLimited = (int)(OtherCount - (RemLatency * LFactor)) > (int)LFactor;
+ }
+ if (!OtherResLimited && (RemLatency + CurrCycle > Rem->CriticalPath)) {
+ Policy.ReduceLatency |= true;
+ DEBUG(dbgs() << " " << Available.getName() << " RemainingLatency "
+ << RemLatency << " + " << CurrCycle << "c > CritPath "
+ << Rem->CriticalPath << "\n");
+ }
+ // If the same resource is limiting inside and outside the zone, do nothing.
+ if (IsResourceLimited && OtherResLimited && (ZoneCritResIdx == OtherCritIdx))
+ return;
+
+ DEBUG(
+ if (IsResourceLimited) {
+ dbgs() << " " << Available.getName() << " ResourceLimited: "
+ << getResourceName(ZoneCritResIdx) << "\n";
+ }
+ if (OtherResLimited)
+ dbgs() << " RemainingLimit: " << getResourceName(OtherCritIdx) << "\n";
+ if (!IsResourceLimited && !OtherResLimited)
+ dbgs() << " Latency limited both directions.\n");
+
+ if (IsResourceLimited && !Policy.ReduceResIdx)
+ Policy.ReduceResIdx = ZoneCritResIdx;
+
+ if (OtherResLimited)
+ Policy.DemandResIdx = OtherCritIdx;
}
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))
+ bool IsBuffered = SchedModel->getMicroOpBufferSize() != 0;
+ if ((!IsBuffered && 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;
-
- unsigned NextCycle = CurrCycle + 1;
- assert(MinReadyCycle < UINT_MAX && "MinReadyCycle uninitialized");
- if (MinReadyCycle > NextCycle) {
- IssueCount = 0;
- NextCycle = MinReadyCycle;
- }
+void ConvergingScheduler::SchedBoundary::bumpCycle(unsigned NextCycle) {
+ if (SchedModel->getMicroOpBufferSize() == 0) {
+ assert(MinReadyCycle < UINT_MAX && "MinReadyCycle uninitialized");
+ if (MinReadyCycle > NextCycle)
+ NextCycle = MinReadyCycle;
+ }
+ // Update the current micro-ops, which will issue in the next cycle.
+ unsigned DecMOps = SchedModel->getIssueWidth() * (NextCycle - CurrCycle);
+ CurrMOps = (CurrMOps <= DecMOps) ? 0 : CurrMOps - DecMOps;
+
+ // Decrement DependentLatency based on the next cycle.
+ if ((NextCycle - CurrCycle) > DependentLatency)
+ DependentLatency = 0;
+ else
+ DependentLatency -= (NextCycle - CurrCycle);
if (!HazardRec->isEnabled()) {
// Bypass HazardRec virtual calls.
}
}
CheckPending = true;
- IsResourceLimited = getCriticalCount() > std::max(ExpectedLatency, CurrCycle);
+ unsigned LFactor = SchedModel->getLatencyFactor();
+ IsResourceLimited =
+ (int)(getCriticalCount() - (getScheduledLatency() * LFactor))
+ > (int)LFactor;
- DEBUG(dbgs() << " *** " << Available.getName() << " cycle "
- << CurrCycle << '\n');
+ DEBUG(dbgs() << "Cycle: " << CurrCycle << ' ' << Available.getName() << '\n');
+}
+
+void ConvergingScheduler::SchedBoundary::incExecutedResources(unsigned PIdx,
+ unsigned Count) {
+ ExecutedResCounts[PIdx] += Count;
+ if (ExecutedResCounts[PIdx] > MaxExecutedResCount)
+ MaxExecutedResCount = ExecutedResCounts[PIdx];
}
/// Add the given processor resource to this scheduled zone.
-void ConvergingScheduler::SchedBoundary::countResource(unsigned PIdx,
- unsigned Cycles) {
+///
+/// \param Cycles indicates the number of consecutive (non-pipelined) cycles
+/// during which this resource is consumed.
+///
+/// \return the next cycle at which the instruction may execute without
+/// oversubscribing resources.
+unsigned ConvergingScheduler::SchedBoundary::
+countResource(unsigned PIdx, unsigned Cycles, unsigned ReadyCycle) {
unsigned Factor = SchedModel->getResourceFactor(PIdx);
- DEBUG(dbgs() << " " << SchedModel->getProcResource(PIdx)->Name
- << " +(" << Cycles << "x" << Factor
- << ") / " << SchedModel->getLatencyFactor() << '\n');
-
unsigned Count = Factor * Cycles;
- ResourceCounts[PIdx] += Count;
+ DEBUG(dbgs() << " " << getResourceName(PIdx)
+ << " +" << Cycles << "x" << Factor << "u\n");
+
+ // Update Executed resources counts.
+ incExecutedResources(PIdx, Count);
assert(Rem->RemainingCounts[PIdx] >= Count && "resource double counted");
Rem->RemainingCounts[PIdx] -= Count;
// Check if this resource exceeds the current critical resource by a full
// cycle. If so, it becomes the critical resource.
- if ((int)(ResourceCounts[PIdx] - ResourceCounts[CritResIdx])
- >= (int)SchedModel->getLatencyFactor()) {
- CritResIdx = PIdx;
+ if (ZoneCritResIdx != PIdx
+ && ((int)(getResourceCount(PIdx) - getCriticalCount())
+ >= (int)SchedModel->getLatencyFactor())) {
+ ZoneCritResIdx = PIdx;
DEBUG(dbgs() << " *** Critical resource "
- << SchedModel->getProcResource(PIdx)->Name << " x"
- << ResourceCounts[PIdx] << '\n');
+ << getResourceName(PIdx) << ": "
+ << getResourceCount(PIdx) / SchedModel->getLatencyFactor() << "c\n");
}
+ // TODO: We don't yet model reserved resources. It's not hard though.
+ return CurrCycle;
}
/// Move the boundary of scheduled code by one SUnit.
}
HazardRec->EmitInstruction(SU);
}
+ const MCSchedClassDesc *SC = DAG->getSchedClass(SU);
+ unsigned IncMOps = SchedModel->getNumMicroOps(SU->getInstr());
+ CurrMOps += IncMOps;
+ // checkHazard prevents scheduling multiple instructions per cycle that exceed
+ // issue width. However, we commonly reach the maximum. In this case
+ // opportunistically bump the cycle to avoid uselessly checking everything in
+ // the readyQ. Furthermore, a single instruction may produce more than one
+ // cycle's worth of micro-ops.
+ //
+ // TODO: Also check if this SU must end a dispatch group.
+ unsigned NextCycle = CurrCycle;
+ if (CurrMOps >= SchedModel->getIssueWidth()) {
+ ++NextCycle;
+ DEBUG(dbgs() << " *** Max MOps " << CurrMOps
+ << " at cycle " << CurrCycle << '\n');
+ }
+ unsigned ReadyCycle = (isTop() ? SU->TopReadyCycle : SU->BotReadyCycle);
+ DEBUG(dbgs() << " Ready @" << ReadyCycle << "c\n");
+
+ switch (SchedModel->getMicroOpBufferSize()) {
+ case 0:
+ assert(ReadyCycle <= CurrCycle && "Broken PendingQueue");
+ break;
+ case 1:
+ if (ReadyCycle > NextCycle) {
+ NextCycle = ReadyCycle;
+ DEBUG(dbgs() << " *** Stall until: " << ReadyCycle << "\n");
+ }
+ break;
+ default:
+ // We don't currently model the OOO reorder buffer, so consider all
+ // scheduled MOps to be "retired".
+ break;
+ }
+ RetiredMOps += IncMOps;
+
// Update resource counts and critical resource.
if (SchedModel->hasInstrSchedModel()) {
- const MCSchedClassDesc *SC = DAG->getSchedClass(SU);
- Rem->RemainingMicroOps -= SchedModel->getNumMicroOps(SU->getInstr(), SC);
+ unsigned DecRemIssue = IncMOps * SchedModel->getMicroOpFactor();
+ assert(Rem->RemIssueCount >= DecRemIssue && "MOps double counted");
+ Rem->RemIssueCount -= DecRemIssue;
+ if (ZoneCritResIdx) {
+ // Scale scheduled micro-ops for comparing with the critical resource.
+ unsigned ScaledMOps =
+ RetiredMOps * SchedModel->getMicroOpFactor();
+
+ // If scaled micro-ops are now more than the previous critical resource by
+ // a full cycle, then micro-ops issue becomes critical.
+ if ((int)(ScaledMOps - getResourceCount(ZoneCritResIdx))
+ >= (int)SchedModel->getLatencyFactor()) {
+ ZoneCritResIdx = 0;
+ DEBUG(dbgs() << " *** Critical resource NumMicroOps: "
+ << ScaledMOps / SchedModel->getLatencyFactor() << "c\n");
+ }
+ }
for (TargetSchedModel::ProcResIter
PI = SchedModel->getWriteProcResBegin(SC),
PE = SchedModel->getWriteProcResEnd(SC); PI != PE; ++PI) {
- countResource(PI->ProcResourceIdx, PI->Cycles);
+ unsigned RCycle =
+ countResource(PI->ProcResourceIdx, PI->Cycles, ReadyCycle);
+ if (RCycle > NextCycle)
+ NextCycle = RCycle;
}
}
- if (isTop()) {
- if (SU->getDepth() > ExpectedLatency)
- ExpectedLatency = SU->getDepth();
+ // Update ExpectedLatency and DependentLatency.
+ unsigned &TopLatency = isTop() ? ExpectedLatency : DependentLatency;
+ unsigned &BotLatency = isTop() ? DependentLatency : ExpectedLatency;
+ if (SU->getDepth() > TopLatency) {
+ TopLatency = SU->getDepth();
+ DEBUG(dbgs() << " " << Available.getName()
+ << " TopLatency SU(" << SU->NodeNum << ") " << TopLatency << "c\n");
}
- else {
- if (SU->getHeight() > ExpectedLatency)
- ExpectedLatency = SU->getHeight();
+ if (SU->getHeight() > BotLatency) {
+ BotLatency = SU->getHeight();
+ DEBUG(dbgs() << " " << Available.getName()
+ << " BotLatency SU(" << SU->NodeNum << ") " << BotLatency << "c\n");
}
-
- IsResourceLimited = getCriticalCount() > std::max(ExpectedLatency, CurrCycle);
-
- // Check the instruction group dispatch limit.
- // TODO: Check if this SU must end a dispatch group.
- IssueCount += SchedModel->getNumMicroOps(SU->getInstr());
-
- // checkHazard prevents scheduling multiple instructions per cycle that exceed
- // issue width. However, we commonly reach the maximum. In this case
- // opportunistically bump the cycle to avoid uselessly checking everything in
- // the readyQ. Furthermore, a single instruction may produce more than one
- // cycle's worth of micro-ops.
- if (IssueCount >= SchedModel->getIssueWidth()) {
- DEBUG(dbgs() << " *** Max instrs at cycle " << CurrCycle << '\n');
- bumpCycle();
+ // If we stall for any reason, bump the cycle.
+ if (NextCycle > CurrCycle) {
+ bumpCycle(NextCycle);
}
+ else {
+ // After updating ZoneCritResIdx and ExpectedLatency, check if we're
+ // resource limited. If a stall occured, bumpCycle does this.
+ unsigned LFactor = SchedModel->getLatencyFactor();
+ IsResourceLimited =
+ (int)(getCriticalCount() - (getScheduledLatency() * LFactor))
+ > (int)LFactor;
+ }
+ DEBUG(dumpScheduledState());
}
/// Release pending ready nodes in to the available queue. This makes them
// Check to see if any of the pending instructions are ready to issue. If
// so, add them to the available queue.
+ bool IsBuffered = SchedModel->getMicroOpBufferSize() != 0;
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)
+ if (!IsBuffered && ReadyCycle > CurrCycle)
continue;
if (checkHazard(SU))
if (CheckPending)
releasePending();
- if (IssueCount > 0) {
+ if (CurrMOps > 0) {
// Defer any ready instrs that now have a hazard.
for (ReadyQueue::iterator I = Available.begin(); I != Available.end();) {
if (checkHazard(*I)) {
}
}
for (unsigned i = 0; Available.empty(); ++i) {
- assert(i <= (HazardRec->getMaxLookAhead() + MaxMinLatency) &&
+ assert(i <= (HazardRec->getMaxLookAhead() + MaxObservedLatency) &&
"permanent hazard"); (void)i;
- bumpCycle();
+ bumpCycle(CurrCycle + 1);
releasePending();
}
if (Available.size() == 1)
return NULL;
}
-/// Record the candidate policy for opposite zones with different critical
-/// resources.
-///
-/// If the CriticalZone is latency limited, don't force a policy for the
-/// candidates here. Instead, setLatencyPolicy sets ReduceLatency if needed.
-void ConvergingScheduler::balanceZones(
- ConvergingScheduler::SchedBoundary &CriticalZone,
- ConvergingScheduler::SchedCandidate &CriticalCand,
- ConvergingScheduler::SchedBoundary &OppositeZone,
- ConvergingScheduler::SchedCandidate &OppositeCand) {
-
- if (!CriticalZone.IsResourceLimited)
- return;
- assert(SchedModel->hasInstrSchedModel() && "required schedmodel");
-
- SchedRemainder *Rem = CriticalZone.Rem;
-
- // If the critical zone is overconsuming a resource relative to the
- // remainder, try to reduce it.
- unsigned RemainingCritCount =
- Rem->RemainingCounts[CriticalZone.CritResIdx];
- if ((int)(Rem->getMaxRemainingCount(SchedModel) - RemainingCritCount)
- > (int)SchedModel->getLatencyFactor()) {
- CriticalCand.Policy.ReduceResIdx = CriticalZone.CritResIdx;
- DEBUG(dbgs() << "Balance " << CriticalZone.Available.getName() << " reduce "
- << SchedModel->getProcResource(CriticalZone.CritResIdx)->Name
- << '\n');
- }
- // If the other zone is underconsuming a resource relative to the full zone,
- // try to increase it.
- unsigned OppositeCount =
- OppositeZone.ResourceCounts[CriticalZone.CritResIdx];
- if ((int)(OppositeZone.ExpectedCount - OppositeCount)
- > (int)SchedModel->getLatencyFactor()) {
- OppositeCand.Policy.DemandResIdx = CriticalZone.CritResIdx;
- DEBUG(dbgs() << "Balance " << OppositeZone.Available.getName() << " demand "
- << SchedModel->getProcResource(OppositeZone.CritResIdx)->Name
- << '\n');
- }
-}
-
-/// Determine if the scheduled zones exceed resource limits or critical path and
-/// set each candidate's ReduceHeight policy accordingly.
-void ConvergingScheduler::checkResourceLimits(
- ConvergingScheduler::SchedCandidate &TopCand,
- ConvergingScheduler::SchedCandidate &BotCand) {
-
- // Set ReduceLatency to true if needed.
- Bot.setLatencyPolicy(BotCand.Policy);
- Top.setLatencyPolicy(TopCand.Policy);
-
- // Handle resource-limited regions.
- if (Top.IsResourceLimited && Bot.IsResourceLimited
- && Top.CritResIdx == Bot.CritResIdx) {
- // If the scheduled critical resource in both zones is no longer the
- // critical remaining resource, attempt to reduce resource height both ways.
- if (Top.CritResIdx != Rem.CritResIdx) {
- TopCand.Policy.ReduceResIdx = Top.CritResIdx;
- BotCand.Policy.ReduceResIdx = Bot.CritResIdx;
- DEBUG(dbgs() << "Reduce scheduled "
- << SchedModel->getProcResource(Top.CritResIdx)->Name << '\n');
- }
- return;
- }
- // Handle latency-limited regions.
- if (!Top.IsResourceLimited && !Bot.IsResourceLimited) {
- // If the total scheduled expected latency exceeds the region's critical
- // path then reduce latency both ways.
- //
- // Just because a zone is not resource limited does not mean it is latency
- // limited. Unbuffered resource, such as max micro-ops may cause CurrCycle
- // to exceed expected latency.
- if ((Top.ExpectedLatency + Bot.ExpectedLatency >= Rem.CriticalPath)
- && (Rem.CriticalPath > Top.CurrCycle + Bot.CurrCycle)) {
- TopCand.Policy.ReduceLatency = true;
- BotCand.Policy.ReduceLatency = true;
- DEBUG(dbgs() << "Reduce scheduled latency " << Top.ExpectedLatency
- << " + " << Bot.ExpectedLatency << '\n');
- }
- return;
+#ifndef NDEBUG
+// This is useful information to dump after bumpNode.
+// Note that the Queue contents are more useful before pickNodeFromQueue.
+void ConvergingScheduler::SchedBoundary::dumpScheduledState() {
+ unsigned ResFactor;
+ unsigned ResCount;
+ if (ZoneCritResIdx) {
+ ResFactor = SchedModel->getResourceFactor(ZoneCritResIdx);
+ ResCount = getResourceCount(ZoneCritResIdx);
}
- // The critical resource is different in each zone, so request balancing.
-
- // Compute the cost of each zone.
- Top.ExpectedCount = std::max(Top.ExpectedLatency, Top.CurrCycle);
- Top.ExpectedCount = std::max(
- Top.getCriticalCount(),
- Top.ExpectedCount * SchedModel->getLatencyFactor());
- Bot.ExpectedCount = std::max(Bot.ExpectedLatency, Bot.CurrCycle);
- Bot.ExpectedCount = std::max(
- Bot.getCriticalCount(),
- Bot.ExpectedCount * SchedModel->getLatencyFactor());
-
- balanceZones(Top, TopCand, Bot, BotCand);
- balanceZones(Bot, BotCand, Top, TopCand);
+ else {
+ ResFactor = SchedModel->getMicroOpFactor();
+ ResCount = RetiredMOps * SchedModel->getMicroOpFactor();
+ }
+ unsigned LFactor = SchedModel->getLatencyFactor();
+ dbgs() << Available.getName() << " @" << CurrCycle << "c\n"
+ << " Retired: " << RetiredMOps;
+ dbgs() << "\n Executed: " << getExecutedCount() / LFactor << "c";
+ dbgs() << "\n Critical: " << ResCount / LFactor << "c, "
+ << ResCount / ResFactor << " " << getResourceName(ZoneCritResIdx)
+ << "\n ExpectedLatency: " << ExpectedLatency << "c\n"
+ << (IsResourceLimited ? " - Resource" : " - Latency")
+ << " limited.\n";
}
+#endif
void ConvergingScheduler::SchedCandidate::
initResourceDelta(const ScheduleDAGMI *DAG,
}
}
+
/// Return true if this heuristic determines order.
-static bool tryLess(unsigned TryVal, unsigned CandVal,
+static bool tryLess(int TryVal, int CandVal,
ConvergingScheduler::SchedCandidate &TryCand,
ConvergingScheduler::SchedCandidate &Cand,
ConvergingScheduler::CandReason Reason) {
Cand.Reason = Reason;
return true;
}
+ Cand.setRepeat(Reason);
return false;
}
-static bool tryGreater(unsigned TryVal, unsigned CandVal,
+static bool tryGreater(int TryVal, int CandVal,
ConvergingScheduler::SchedCandidate &TryCand,
ConvergingScheduler::SchedCandidate &Cand,
ConvergingScheduler::CandReason Reason) {
Cand.Reason = Reason;
return true;
}
+ Cand.setRepeat(Reason);
return false;
}
return (isTop) ? SU->WeakPredsLeft : SU->WeakSuccsLeft;
}
+/// Minimize physical register live ranges. Regalloc wants them adjacent to
+/// their physreg def/use.
+///
+/// FIXME: This is an unnecessary check on the critical path. Most are root/leaf
+/// copies which can be prescheduled. The rest (e.g. x86 MUL) could be bundled
+/// with the operation that produces or consumes the physreg. We'll do this when
+/// regalloc has support for parallel copies.
+static int biasPhysRegCopy(const SUnit *SU, bool isTop) {
+ const MachineInstr *MI = SU->getInstr();
+ if (!MI->isCopy())
+ return 0;
+
+ unsigned ScheduledOper = isTop ? 1 : 0;
+ unsigned UnscheduledOper = isTop ? 0 : 1;
+ // If we have already scheduled the physreg produce/consumer, immediately
+ // schedule the copy.
+ if (TargetRegisterInfo::isPhysicalRegister(
+ MI->getOperand(ScheduledOper).getReg()))
+ return 1;
+ // If the physreg is at the boundary, defer it. Otherwise schedule it
+ // immediately to free the dependent. We can hoist the copy later.
+ bool AtBoundary = isTop ? !SU->NumSuccsLeft : !SU->NumPredsLeft;
+ if (TargetRegisterInfo::isPhysicalRegister(
+ MI->getOperand(UnscheduledOper).getReg()))
+ return AtBoundary ? -1 : 1;
+ return 0;
+}
+
/// Apply a set of heursitics to a new candidate. Heuristics are currently
/// hierarchical. This may be more efficient than a graduated cost model because
/// we don't need to evaluate all aspects of the model for each node in the
TryCand.Reason = NodeOrder;
return;
}
+
+ if (tryGreater(biasPhysRegCopy(TryCand.SU, Zone.isTop()),
+ biasPhysRegCopy(Cand.SU, Zone.isTop()),
+ TryCand, Cand, PhysRegCopy))
+ return;
+
// Avoid exceeding the target's limit.
if (tryLess(TryCand.RPDelta.Excess.UnitIncrease,
- Cand.RPDelta.Excess.UnitIncrease, TryCand, Cand, SingleExcess))
+ Cand.RPDelta.Excess.UnitIncrease, TryCand, Cand, RegExcess))
return;
- if (Cand.Reason == SingleExcess)
- Cand.Reason = MultiPressure;
// Avoid increasing the max critical pressure in the scheduled region.
if (tryLess(TryCand.RPDelta.CriticalMax.UnitIncrease,
Cand.RPDelta.CriticalMax.UnitIncrease,
- TryCand, Cand, SingleCritical))
+ TryCand, Cand, RegCritical))
return;
- if (Cand.Reason == SingleCritical)
- Cand.Reason = MultiPressure;
// Keep clustered nodes together to encourage downstream peephole
// optimizations which may reduce resource requirements.
if (tryGreater(TryCand.SU == NextClusterSU, Cand.SU == NextClusterSU,
TryCand, Cand, Cluster))
return;
- // Currently, weak edges are for clustering, so we hard-code that reason.
- // However, deferring the current TryCand will not change Cand's reason.
- CandReason OrigReason = Cand.Reason;
+
+ // Weak edges are for clustering and other constraints.
if (tryLess(getWeakLeft(TryCand.SU, Zone.isTop()),
getWeakLeft(Cand.SU, Zone.isTop()),
- TryCand, Cand, Cluster)) {
- Cand.Reason = OrigReason;
+ TryCand, Cand, Weak)) {
return;
}
+ // Avoid increasing the max pressure of the entire region.
+ if (tryLess(TryCand.RPDelta.CurrentMax.UnitIncrease,
+ Cand.RPDelta.CurrentMax.UnitIncrease, TryCand, Cand, RegMax))
+ return;
+
// Avoid critical resource consumption and balance the schedule.
TryCand.initResourceDelta(DAG, SchedModel);
if (tryLess(TryCand.ResDelta.CritResources, Cand.ResDelta.CritResources,
// Avoid serializing long latency dependence chains.
if (Cand.Policy.ReduceLatency) {
if (Zone.isTop()) {
- if (Cand.SU->getDepth() * SchedModel->getLatencyFactor()
- > Zone.ExpectedCount) {
+ if (Cand.SU->getDepth() > Zone.getScheduledLatency()) {
if (tryLess(TryCand.SU->getDepth(), Cand.SU->getDepth(),
TryCand, Cand, TopDepthReduce))
return;
return;
}
else {
- if (Cand.SU->getHeight() * SchedModel->getLatencyFactor()
- > Zone.ExpectedCount) {
+ if (Cand.SU->getHeight() > Zone.getScheduledLatency()) {
if (tryLess(TryCand.SU->getHeight(), Cand.SU->getHeight(),
TryCand, Cand, BotHeightReduce))
return;
}
}
- // Avoid increasing the max pressure of the entire region.
- if (tryLess(TryCand.RPDelta.CurrentMax.UnitIncrease,
- Cand.RPDelta.CurrentMax.UnitIncrease, TryCand, Cand, SingleMax))
- return;
- if (Cand.Reason == SingleMax)
- Cand.Reason = MultiPressure;
-
// Prefer immediate defs/users of the last scheduled instruction. This is a
- // nice pressure avoidance strategy that also conserves the processor's
- // register renaming resources and keeps the machine code readable.
+ // local pressure avoidance strategy that also makes the machine code
+ // readable.
if (tryGreater(Zone.NextSUs.count(TryCand.SU), Zone.NextSUs.count(Cand.SU),
TryCand, Cand, NextDefUse))
return;
}
}
-/// 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) {
- DEBUG(dbgs() << "RP excess top - bot: "
- << (LHS.Excess.UnitIncrease - RHS.Excess.UnitIncrease) << '\n');
- return LHS.Excess.UnitIncrease < RHS.Excess.UnitIncrease;
- }
- // Avoid increasing the max critical pressure in the scheduled region.
- if (LHS.CriticalMax.UnitIncrease != RHS.CriticalMax.UnitIncrease) {
- DEBUG(dbgs() << "RP critical top - bot: "
- << (LHS.CriticalMax.UnitIncrease - RHS.CriticalMax.UnitIncrease)
- << '\n');
- return LHS.CriticalMax.UnitIncrease < RHS.CriticalMax.UnitIncrease;
- }
- // Avoid increasing the max pressure of the entire region.
- if (LHS.CurrentMax.UnitIncrease != RHS.CurrentMax.UnitIncrease) {
- DEBUG(dbgs() << "RP current top - bot: "
- << (LHS.CurrentMax.UnitIncrease - RHS.CurrentMax.UnitIncrease)
- << '\n');
- return LHS.CurrentMax.UnitIncrease < RHS.CurrentMax.UnitIncrease;
- }
- return false;
-}
-
#ifndef NDEBUG
const char *ConvergingScheduler::getReasonStr(
ConvergingScheduler::CandReason Reason) {
switch (Reason) {
case NoCand: return "NOCAND ";
- case SingleExcess: return "REG-EXCESS";
- case SingleCritical: return "REG-CRIT ";
+ case PhysRegCopy: return "PREG-COPY";
+ case RegExcess: return "REG-EXCESS";
+ case RegCritical: return "REG-CRIT ";
case Cluster: return "CLUSTER ";
- case SingleMax: return "REG-MAX ";
- case MultiPressure: return "REG-MULTI ";
+ case Weak: return "WEAK ";
+ case RegMax: return "REG-MAX ";
case ResourceReduce: return "RES-REDUCE";
case ResourceDemand: return "RES-DEMAND";
case TopDepthReduce: return "TOP-DEPTH ";
llvm_unreachable("Unknown reason!");
}
-void ConvergingScheduler::traceCandidate(const SchedCandidate &Cand,
- const SchedBoundary &Zone) {
- const char *Label = getReasonStr(Cand.Reason);
+void ConvergingScheduler::traceCandidate(const SchedCandidate &Cand) {
PressureElement P;
unsigned ResIdx = 0;
unsigned Latency = 0;
switch (Cand.Reason) {
default:
break;
- case SingleExcess:
+ case RegExcess:
P = Cand.RPDelta.Excess;
break;
- case SingleCritical:
+ case RegCritical:
P = Cand.RPDelta.CriticalMax;
break;
- case SingleMax:
+ case RegMax:
P = Cand.RPDelta.CurrentMax;
break;
case ResourceReduce:
Latency = Cand.SU->getDepth();
break;
}
- dbgs() << Label << " " << Zone.Available.getName() << " ";
+ dbgs() << " SU(" << Cand.SU->NodeNum << ") " << getReasonStr(Cand.Reason);
if (P.isValid())
- dbgs() << TRI->getRegPressureSetName(P.PSetID) << ":" << P.UnitIncrease
- << " ";
+ dbgs() << " " << TRI->getRegPressureSetName(P.PSetID)
+ << ":" << P.UnitIncrease << " ";
else
- dbgs() << " ";
+ dbgs() << " ";
if (ResIdx)
- dbgs() << SchedModel->getProcResource(ResIdx)->Name << " ";
+ dbgs() << " " << SchedModel->getProcResource(ResIdx)->Name << " ";
else
- dbgs() << " ";
+ dbgs() << " ";
if (Latency)
- dbgs() << Latency << " cycles ";
+ dbgs() << " " << Latency << " cycles ";
else
- dbgs() << " ";
- Cand.SU->dump(DAG);
+ dbgs() << " ";
+ dbgs() << '\n';
}
#endif
if (TryCand.ResDelta == SchedResourceDelta())
TryCand.initResourceDelta(DAG, SchedModel);
Cand.setBest(TryCand);
- DEBUG(traceCandidate(Cand, Zone));
+ DEBUG(traceCandidate(Cand));
}
}
}
static void tracePick(const ConvergingScheduler::SchedCandidate &Cand,
bool IsTop) {
- DEBUG(dbgs() << "Pick " << (IsTop ? "top" : "bot")
- << " SU(" << Cand.SU->NodeNum << ") "
+ DEBUG(dbgs() << "Pick " << (IsTop ? "Top " : "Bot ")
<< ConvergingScheduler::getReasonStr(Cand.Reason) << '\n');
}
// efficient, but also provides the best heuristics for CriticalPSets.
if (SUnit *SU = Bot.pickOnlyChoice()) {
IsTopNode = false;
+ DEBUG(dbgs() << "Pick Bot NOCAND\n");
return SU;
}
if (SUnit *SU = Top.pickOnlyChoice()) {
IsTopNode = true;
+ DEBUG(dbgs() << "Pick Top NOCAND\n");
return SU;
}
CandPolicy NoPolicy;
SchedCandidate BotCand(NoPolicy);
SchedCandidate TopCand(NoPolicy);
- checkResourceLimits(TopCand, BotCand);
+ Bot.setPolicy(BotCand.Policy, Top);
+ Top.setPolicy(TopCand.Policy, Bot);
// Prefer bottom scheduling when heuristics are silent.
pickNodeFromQueue(Bot, DAG->getBotRPTracker(), BotCand);
// 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 (BotCand.Reason == SingleExcess || BotCand.Reason == SingleCritical) {
+ if ((BotCand.Reason == RegExcess && !BotCand.isRepeat(RegExcess))
+ || (BotCand.Reason == RegCritical
+ && !BotCand.isRepeat(RegCritical)))
+ {
IsTopNode = false;
tracePick(BotCand, IsTopNode);
return BotCand.SU;
pickNodeFromQueue(Top, DAG->getTopRPTracker(), TopCand);
assert(TopCand.Reason != NoCand && "failed to find the first candidate");
- // If either Q has a single candidate that minimizes pressure above the
- // original region's pressure pick it.
- if (TopCand.Reason <= SingleMax || BotCand.Reason <= SingleMax) {
- if (TopCand.Reason < BotCand.Reason) {
- IsTopNode = true;
- tracePick(TopCand, IsTopNode);
- return TopCand.SU;
- }
- IsTopNode = false;
- tracePick(BotCand, IsTopNode);
- return BotCand.SU;
- }
- // Check for a salient pressure difference and pick the best from either side.
- if (compareRPDelta(TopCand.RPDelta, BotCand.RPDelta)) {
- IsTopNode = true;
- tracePick(TopCand, IsTopNode);
- return TopCand.SU;
- }
- // Otherwise prefer the bottom candidate, in node order if all else failed.
+ // Choose the queue with the most important (lowest enum) reason.
if (TopCand.Reason < BotCand.Reason) {
IsTopNode = true;
tracePick(TopCand, IsTopNode);
return TopCand.SU;
}
+ // Otherwise prefer the bottom candidate, in node order if all else failed.
IsTopNode = false;
tracePick(BotCand, IsTopNode);
return BotCand.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));
+ DEBUG(dbgs() << "Scheduling SU(" << SU->NodeNum << ") " << *SU->getInstr());
return SU;
}
+void ConvergingScheduler::reschedulePhysRegCopies(SUnit *SU, bool isTop) {
+
+ MachineBasicBlock::iterator InsertPos = SU->getInstr();
+ if (!isTop)
+ ++InsertPos;
+ SmallVectorImpl<SDep> &Deps = isTop ? SU->Preds : SU->Succs;
+
+ // Find already scheduled copies with a single physreg dependence and move
+ // them just above the scheduled instruction.
+ for (SmallVectorImpl<SDep>::iterator I = Deps.begin(), E = Deps.end();
+ I != E; ++I) {
+ if (I->getKind() != SDep::Data || !TRI->isPhysicalRegister(I->getReg()))
+ continue;
+ SUnit *DepSU = I->getSUnit();
+ if (isTop ? DepSU->Succs.size() > 1 : DepSU->Preds.size() > 1)
+ continue;
+ MachineInstr *Copy = DepSU->getInstr();
+ if (!Copy->isCopy())
+ continue;
+ DEBUG(dbgs() << " Rescheduling physreg copy ";
+ I->getSUnit()->dump(DAG));
+ DAG->moveInstruction(Copy, InsertPos);
+ }
+}
+
/// 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.
+///
+/// FIXME: Eventually, we may bundle physreg copies rather than rescheduling
+/// them here. See comments in biasPhysRegCopy.
void ConvergingScheduler::schedNode(SUnit *SU, bool IsTopNode) {
if (IsTopNode) {
- SU->TopReadyCycle = Top.CurrCycle;
+ SU->TopReadyCycle = std::max(SU->TopReadyCycle, Top.CurrCycle);
Top.bumpNode(SU);
+ if (SU->hasPhysRegUses)
+ reschedulePhysRegCopies(SU, true);
}
else {
- SU->BotReadyCycle = Bot.CurrCycle;
+ SU->BotReadyCycle = std::max(SU->BotReadyCycle, Bot.CurrCycle);
Bot.bumpNode(SU);
+ if (SU->hasPhysRegDefs)
+ reschedulePhysRegCopies(SU, false);
}
}
"-misched-topdown incompatible with -misched-bottomup");
ScheduleDAGMI *DAG = new ScheduleDAGMI(C, new ConvergingScheduler());
// Register DAG post-processors.
+ //
+ // FIXME: extend the mutation API to allow earlier mutations to instantiate
+ // data and pass it to later mutations. Have a single mutation that gathers
+ // the interesting nodes in one pass.
+ DAG->addMutation(new CopyConstrain(DAG->TII, DAG->TRI));
if (EnableLoadCluster)
DAG->addMutation(new LoadClusterMutation(DAG->TII, DAG->TRI));
if (EnableMacroFusion)
/// Callback to select the highest priority node from the ready Q.
virtual SUnit *pickNode(bool &IsTopNode) {
if (ReadyQ.empty()) return NULL;
- pop_heap(ReadyQ.begin(), ReadyQ.end(), Cmp);
+ std::pop_heap(ReadyQ.begin(), ReadyQ.end(), Cmp);
SUnit *SU = ReadyQ.back();
ReadyQ.pop_back();
IsTopNode = false;
- DEBUG(dbgs() << "*** Scheduling " << "SU(" << SU->NodeNum << "): "
- << *SU->getInstr()
+ DEBUG(dbgs() << "Pick node " << "SU(" << SU->NodeNum << ") "
<< " ILP: " << DAG->getDFSResult()->getILP(SU)
<< " Tree: " << DAG->getDFSResult()->getSubtreeID(SU) << " @"
<< DAG->getDFSResult()->getSubtreeLevel(
- DAG->getDFSResult()->getSubtreeID(SU)) << '\n');
+ DAG->getDFSResult()->getSubtreeID(SU)) << '\n'
+ << "Scheduling " << *SU->getInstr());
return SU;
}