static bool ViewMISchedDAGs = false;
#endif // NDEBUG
+static cl::opt<bool> EnableRegPressure("misched-regpressure", cl::Hidden,
+ cl::desc("Enable register pressure scheduling."), cl::init(true));
+
static cl::opt<bool> EnableCyclicPath("misched-cyclicpath", cl::Hidden,
cl::desc("Enable cyclic critical path analysis."), cl::init(false));
/// Decrement this iterator until reaching the top or a non-debug instr.
-static MachineBasicBlock::iterator
-priorNonDebug(MachineBasicBlock::iterator I, MachineBasicBlock::iterator Beg) {
+static MachineBasicBlock::const_iterator
+priorNonDebug(MachineBasicBlock::const_iterator I,
+ MachineBasicBlock::const_iterator Beg) {
assert(I != Beg && "reached the top of the region, cannot decrement");
while (--I != Beg) {
if (!I->isDebugValue())
return I;
}
+/// Non-const version.
+static MachineBasicBlock::iterator
+priorNonDebug(MachineBasicBlock::iterator I,
+ MachineBasicBlock::const_iterator Beg) {
+ return const_cast<MachineInstr*>(
+ &*priorNonDebug(MachineBasicBlock::const_iterator(I), Beg));
+}
+
/// If this iterator is a debug value, increment until reaching the End or a
/// non-debug instruction.
-static MachineBasicBlock::iterator
-nextIfDebug(MachineBasicBlock::iterator I, MachineBasicBlock::iterator End) {
+static MachineBasicBlock::const_iterator
+nextIfDebug(MachineBasicBlock::const_iterator I,
+ MachineBasicBlock::const_iterator End) {
for(; I != End; ++I) {
if (!I->isDebugValue())
break;
return I;
}
+/// Non-const version.
+static MachineBasicBlock::iterator
+nextIfDebug(MachineBasicBlock::iterator I,
+ MachineBasicBlock::const_iterator End) {
+ // Cast the return value to nonconst MachineInstr, then cast to an
+ // instr_iterator, which does not check for null, finally return a
+ // bundle_iterator.
+ return MachineBasicBlock::instr_iterator(
+ const_cast<MachineInstr*>(
+ &*nextIfDebug(MachineBasicBlock::const_iterator(I), End)));
+}
+
/// Top-level MachineScheduler pass driver.
///
/// Visit blocks in function order. Divide each block into scheduling regions
// For convenience remember the end of the liveness region.
LiveRegionEnd =
(RegionEnd == bb->end()) ? RegionEnd : llvm::next(RegionEnd);
+
+ SUPressureDiffs.clear();
+
+ SchedImpl->initPolicy(begin, end, regioninstrs);
+
+ ShouldTrackPressure = SchedImpl->shouldTrackPressure();
}
// Setup the register pressure trackers for the top scheduled top and bottom
dumpRegSetPressure(BotRPTracker.getLiveThru(), TRI));
};
+ // For each live out vreg reduce the pressure change associated with other
+ // uses of the same vreg below the live-out reaching def.
+ updatePressureDiffs(RPTracker.getPressure().LiveOutRegs);
+
// Account for liveness generated by the region boundary.
- if (LiveRegionEnd != RegionEnd)
- BotRPTracker.recede();
+ if (LiveRegionEnd != RegionEnd) {
+ SmallVector<unsigned, 8> LiveUses;
+ BotRPTracker.recede(&LiveUses);
+ updatePressureDiffs(LiveUses);
+ }
assert(BotRPTracker.getPos() == RegionEnd && "Can't find the region bottom");
DEBUG(dbgs() << TRI->getRegPressureSetName(i)
<< " Limit " << Limit
<< " Actual " << RegionPressure[i] << "\n");
- RegionCriticalPSets.push_back(PressureElement(i, 0));
+ RegionCriticalPSets.push_back(PressureChange(i));
}
}
DEBUG(dbgs() << "Excess PSets: ";
for (unsigned i = 0, e = RegionCriticalPSets.size(); i != e; ++i)
dbgs() << TRI->getRegPressureSetName(
- RegionCriticalPSets[i].PSetID) << " ";
+ RegionCriticalPSets[i].getPSet()) << " ";
dbgs() << "\n");
}
-// FIXME: When the pressure tracker deals in pressure differences then we won't
-// iterate over all RegionCriticalPSets[i].
void ScheduleDAGMI::
-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];
+updateScheduledPressure(const SUnit *SU,
+ const std::vector<unsigned> &NewMaxPressure) {
+ const PressureDiff &PDiff = getPressureDiff(SU);
+ unsigned CritIdx = 0, CritEnd = RegionCriticalPSets.size();
+ for (PressureDiff::const_iterator I = PDiff.begin(), E = PDiff.end();
+ I != E; ++I) {
+ if (!I->isValid())
+ break;
+ unsigned ID = I->getPSet();
+ while (CritIdx != CritEnd && RegionCriticalPSets[CritIdx].getPSet() < ID)
+ ++CritIdx;
+ if (CritIdx != CritEnd && RegionCriticalPSets[CritIdx].getPSet() == ID) {
+ if ((int)NewMaxPressure[ID] > RegionCriticalPSets[CritIdx].getUnitInc()
+ && NewMaxPressure[ID] <= INT16_MAX)
+ RegionCriticalPSets[CritIdx].setUnitInc(NewMaxPressure[ID]);
+ }
+ unsigned Limit = RegClassInfo->getRegPressureSetLimit(ID);
+ if (NewMaxPressure[ID] >= Limit - 2) {
+ DEBUG(dbgs() << " " << TRI->getRegPressureSetName(ID) << ": "
+ << NewMaxPressure[ID] << " > " << Limit << "(+ "
+ << BotRPTracker.getLiveThru()[ID] << " livethru)\n");
+ }
}
- 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";
+}
+
+/// Update the PressureDiff array for liveness after scheduling this
+/// instruction.
+void ScheduleDAGMI::updatePressureDiffs(ArrayRef<unsigned> LiveUses) {
+ for (unsigned LUIdx = 0, LUEnd = LiveUses.size(); LUIdx != LUEnd; ++LUIdx) {
+ /// FIXME: Currently assuming single-use physregs.
+ unsigned Reg = LiveUses[LUIdx];
+ DEBUG(dbgs() << " LiveReg: " << PrintVRegOrUnit(Reg, TRI) << "\n");
+ if (!TRI->isVirtualRegister(Reg))
+ continue;
+
+ // This may be called before CurrentBottom has been initialized. However,
+ // BotRPTracker must have a valid position. We want the value live into the
+ // instruction or live out of the block, so ask for the previous
+ // instruction's live-out.
+ const LiveInterval &LI = LIS->getInterval(Reg);
+ VNInfo *VNI;
+ MachineBasicBlock::const_iterator I =
+ nextIfDebug(BotRPTracker.getPos(), BB->end());
+ if (I == BB->end())
+ VNI = LI.getVNInfoBefore(LIS->getMBBEndIdx(BB));
+ else {
+ LiveRangeQuery LRQ(LI, LIS->getInstructionIndex(I));
+ VNI = LRQ.valueIn();
+ }
+ // RegisterPressureTracker guarantees that readsReg is true for LiveUses.
+ assert(VNI && "No live value at use.");
+ for (VReg2UseMap::iterator
+ UI = VRegUses.find(Reg); UI != VRegUses.end(); ++UI) {
+ SUnit *SU = UI->SU;
+ DEBUG(dbgs() << " UpdateRegP: SU(" << SU->NodeNum << ") "
+ << *SU->getInstr());
+ // If this use comes before the reaching def, it cannot be a last use, so
+ // descrease its pressure change.
+ if (!SU->isScheduled && SU != &ExitSU) {
+ LiveRangeQuery LRQ(LI, LIS->getInstructionIndex(SU->getInstr()));
+ if (LRQ.valueIn() == VNI)
+ getPressureDiff(SU).addPressureChange(Reg, true, &MRI);
}
- });
+ }
+ }
}
/// schedule - Called back from MachineScheduler::runOnMachineFunction
/// Build the DAG and setup three register pressure trackers.
void ScheduleDAGMI::buildDAGWithRegPressure() {
+ if (!ShouldTrackPressure) {
+ RPTracker.reset();
+ RegionCriticalPSets.clear();
+ buildSchedGraph(AA);
+ return;
+ }
+
// Initialize the register pressure tracker used by buildSchedGraph.
RPTracker.init(&MF, RegClassInfo, LIS, BB, LiveRegionEnd,
/*TrackUntiedDefs=*/true);
RPTracker.recede();
// Build the DAG, and compute current register pressure.
- buildSchedGraph(AA, &RPTracker);
+ buildSchedGraph(AA, &RPTracker, &SUPressureDiffs);
// Initialize top/bottom trackers after computing region pressure.
initRegPressure();
/// available for use in the scheduler.
///
/// The cyclic path estimation identifies a def-use pair that crosses the back
-/// end and considers the depth and height of the nodes. For example, consider
+/// edge and considers the depth and height of the nodes. For example, consider
/// the following instruction sequence where each instruction has unit latency
/// and defines an epomymous virtual register:
///
SchedImpl->registerRoots();
// Advance past initial DebugValues.
- assert(TopRPTracker.getPos() == RegionBegin && "bad initial Top tracker");
CurrentTop = nextIfDebug(RegionBegin, RegionEnd);
- TopRPTracker.setPos(CurrentTop);
-
CurrentBottom = RegionEnd;
+
+ if (ShouldTrackPressure) {
+ assert(TopRPTracker.getPos() == RegionBegin && "bad initial Top tracker");
+ TopRPTracker.setPos(CurrentTop);
+ }
}
/// Move an instruction and update register pressure.
TopRPTracker.setPos(MI);
}
- // Update top scheduled pressure.
- TopRPTracker.advance();
- assert(TopRPTracker.getPos() == CurrentTop && "out of sync");
- updateScheduledPressure(TopRPTracker.getPressure().MaxSetPressure);
+ if (ShouldTrackPressure) {
+ // Update top scheduled pressure.
+ TopRPTracker.advance();
+ assert(TopRPTracker.getPos() == CurrentTop && "out of sync");
+ updateScheduledPressure(SU, TopRPTracker.getPressure().MaxSetPressure);
+ }
}
else {
assert(SU->isBottomReady() && "node still has unscheduled dependencies");
moveInstruction(MI, CurrentBottom);
CurrentBottom = MI;
}
- // Update bottom scheduled pressure.
- BotRPTracker.recede();
- assert(BotRPTracker.getPos() == CurrentBottom && "out of sync");
- updateScheduledPressure(BotRPTracker.getPressure().MaxSetPressure);
+ if (ShouldTrackPressure) {
+ // Update bottom scheduled pressure.
+ SmallVector<unsigned, 8> LiveUses;
+ BotRPTracker.recede(&LiveUses);
+ assert(BotRPTracker.getPos() == CurrentBottom && "out of sync");
+ updateScheduledPressure(SU, BotRPTracker.getPressure().MaxSetPressure);
+ updatePressureDiffs(LiveUses);
+ }
}
}
void reset() {
// A new HazardRec is created for each DAG and owned by SchedBoundary.
- delete HazardRec;
-
+ // Destroying and reconstructing it is very expensive though. So keep
+ // invalid, placeholder HazardRecs.
+ if (HazardRec && HazardRec->isEnabled()) {
+ delete HazardRec;
+ HazardRec = 0;
+ }
Available.clear();
Pending.clear();
CheckPending = false;
NextSUs.clear();
- HazardRec = 0;
CurrCycle = 0;
CurrMOps = 0;
MinReadyCycle = UINT_MAX;
};
private:
+ const MachineSchedContext *Context;
ScheduleDAGMI *DAG;
const TargetSchedModel *SchedModel;
const TargetRegisterInfo *TRI;
SchedBoundary Top;
SchedBoundary Bot;
+ MachineSchedPolicy RegionPolicy;
public:
/// SUnit::NodeQueueId: 0 (none), 1 (top), 2 (bot), 3 (both)
enum {
LogMaxQID = 2
};
- ConvergingScheduler():
- DAG(0), SchedModel(0), TRI(0), Top(TopQID, "TopQ"), Bot(BotQID, "BotQ") {}
+ ConvergingScheduler(const MachineSchedContext *C):
+ Context(C), DAG(0), SchedModel(0), TRI(0),
+ Top(TopQID, "TopQ"), Bot(BotQID, "BotQ") {}
+
+ virtual void initPolicy(MachineBasicBlock::iterator Begin,
+ MachineBasicBlock::iterator End,
+ unsigned NumRegionInstrs);
+
+ bool shouldTrackPressure() const { return RegionPolicy.ShouldTrackPressure; }
virtual void initialize(ScheduleDAGMI *dag);
ExecutedResCounts.resize(SchedModel->getNumProcResourceKinds());
}
+/// Initialize the per-region scheduling policy.
+void ConvergingScheduler::initPolicy(MachineBasicBlock::iterator Begin,
+ MachineBasicBlock::iterator End,
+ unsigned NumRegionInstrs) {
+ const TargetMachine &TM = Context->MF->getTarget();
+
+ // Avoid setting up the register pressure tracker for small regions to save
+ // compile time. As a rough heuristic, only track pressure when the number of
+ // schedulable instructions exceeds half the integer register file.
+ unsigned NIntRegs = Context->RegClassInfo->getNumAllocatableRegs(
+ TM.getTargetLowering()->getRegClassFor(MVT::i32));
+
+ RegionPolicy.ShouldTrackPressure = NumRegionInstrs > (NIntRegs / 2);
+
+ // For generic targets, we default to bottom-up, because it's simpler and more
+ // compile-time optimizations have been implemented in that direction.
+ RegionPolicy.OnlyBottomUp = true;
+
+ // Allow the subtarget to override default policy.
+ const TargetSubtargetInfo &ST = TM.getSubtarget<TargetSubtargetInfo>();
+ ST.overrideSchedPolicy(RegionPolicy, Begin, End, NumRegionInstrs);
+
+ // After subtarget overrides, apply command line options.
+ if (!EnableRegPressure)
+ RegionPolicy.ShouldTrackPressure = false;
+
+ // Check -misched-topdown/bottomup can force or unforce scheduling direction.
+ // e.g. -misched-bottomup=false allows scheduling in both directions.
+ assert((!ForceTopDown || !ForceBottomUp) &&
+ "-misched-topdown incompatible with -misched-bottomup");
+ if (ForceBottomUp.getNumOccurrences() > 0) {
+ RegionPolicy.OnlyBottomUp = ForceBottomUp;
+ if (RegionPolicy.OnlyBottomUp)
+ RegionPolicy.OnlyTopDown = false;
+ }
+ if (ForceTopDown.getNumOccurrences() > 0) {
+ RegionPolicy.OnlyTopDown = ForceTopDown;
+ if (RegionPolicy.OnlyTopDown)
+ RegionPolicy.OnlyBottomUp = false;
+ }
+}
+
void ConvergingScheduler::initialize(ScheduleDAGMI *dag) {
DAG = dag;
SchedModel = DAG->getSchedModel();
// 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");
+ if (!Top.HazardRec) {
+ Top.HazardRec =
+ TM.getInstrInfo()->CreateTargetMIHazardRecognizer(Itin, DAG);
+ }
+ if (!Bot.HazardRec) {
+ Bot.HazardRec =
+ TM.getInstrInfo()->CreateTargetMIHazardRecognizer(Itin, DAG);
+ }
}
void ConvergingScheduler::releaseTopNode(SUnit *SU) {
return false;
}
-static bool tryPressure(const PressureElement &TryP,
- const PressureElement &CandP,
+static bool tryPressure(const PressureChange &TryP,
+ const PressureChange &CandP,
ConvergingScheduler::SchedCandidate &TryCand,
ConvergingScheduler::SchedCandidate &Cand,
ConvergingScheduler::CandReason Reason) {
+ int TryRank = TryP.getPSetOrMax();
+ int CandRank = CandP.getPSetOrMax();
// If both candidates affect the same set, go with the smallest increase.
- if (TryP.PSetID == CandP.PSetID) {
- return tryLess(TryP.UnitIncrease, CandP.UnitIncrease, TryCand, Cand,
+ if (TryRank == CandRank) {
+ return tryLess(TryP.getUnitInc(), CandP.getUnitInc(), TryCand, Cand,
Reason);
}
// If one candidate decreases and the other increases, go with it.
- if (tryLess(TryP.UnitIncrease < 0, CandP.UnitIncrease < 0, TryCand, Cand,
+ // Invalid candidates have UnitInc==0.
+ if (tryLess(TryP.getUnitInc() < 0, CandP.getUnitInc() < 0, TryCand, Cand,
Reason)) {
return true;
}
- // If TryP has lower Rank, it has a higher priority.
- int TryRank = TryP.PSetRank();
- int CandRank = CandP.PSetRank();
// If the candidates are decreasing pressure, reverse priority.
- if (TryP.UnitIncrease < 0)
+ if (TryP.getUnitInc() < 0)
std::swap(TryRank, CandRank);
return tryGreater(TryRank, CandRank, TryCand, Cand, Reason);
}
const RegPressureTracker &RPTracker,
RegPressureTracker &TempTracker) {
- // Always initialize TryCand's RPDelta.
- TempTracker.getMaxPressureDelta(TryCand.SU->getInstr(), TryCand.RPDelta,
- DAG->getRegionCriticalPSets(),
- DAG->getRegPressure().MaxSetPressure);
+ if (DAG->isTrackingPressure()) {
+ // Always initialize TryCand's RPDelta.
+ if (Zone.isTop()) {
+ TempTracker.getMaxDownwardPressureDelta(
+ TryCand.SU->getInstr(),
+ TryCand.RPDelta,
+ DAG->getRegionCriticalPSets(),
+ DAG->getRegPressure().MaxSetPressure);
+ }
+ else {
+ if (VerifyScheduling) {
+ TempTracker.getMaxUpwardPressureDelta(
+ TryCand.SU->getInstr(),
+ &DAG->getPressureDiff(TryCand.SU),
+ TryCand.RPDelta,
+ DAG->getRegionCriticalPSets(),
+ DAG->getRegPressure().MaxSetPressure);
+ }
+ else {
+ RPTracker.getUpwardPressureDelta(
+ TryCand.SU->getInstr(),
+ DAG->getPressureDiff(TryCand.SU),
+ TryCand.RPDelta,
+ DAG->getRegionCriticalPSets(),
+ DAG->getRegPressure().MaxSetPressure);
+ }
+ }
+ }
+ DEBUG(if (TryCand.RPDelta.Excess.isValid())
+ dbgs() << " SU(" << TryCand.SU->NodeNum << ") "
+ << TRI->getRegPressureSetName(TryCand.RPDelta.Excess.getPSet())
+ << ":" << TryCand.RPDelta.Excess.getUnitInc() << "\n");
// Initialize the candidate if needed.
if (!Cand.isValid()) {
// Avoid exceeding the target's limit. If signed PSetID is negative, it is
// invalid; convert it to INT_MAX to give it lowest priority.
- if (tryPressure(TryCand.RPDelta.Excess, Cand.RPDelta.Excess, TryCand, Cand,
- RegExcess))
+ if (DAG->isTrackingPressure() && tryPressure(TryCand.RPDelta.Excess,
+ Cand.RPDelta.Excess,
+ TryCand, Cand, RegExcess))
return;
- // For loops that are acyclic path limited, aggressively schedule for latency.
- if (Rem.IsAcyclicLatencyLimited && tryLatency(TryCand, Cand, Zone))
+ // Avoid increasing the max critical pressure in the scheduled region.
+ if (DAG->isTrackingPressure() && tryPressure(TryCand.RPDelta.CriticalMax,
+ Cand.RPDelta.CriticalMax,
+ TryCand, Cand, RegCritical))
return;
- // Avoid increasing the max critical pressure in the scheduled region.
- if (tryPressure(TryCand.RPDelta.CriticalMax, Cand.RPDelta.CriticalMax,
- TryCand, Cand, RegCritical))
+ // For loops that are acyclic path limited, aggressively schedule for latency.
+ // This can result in very long dependence chains scheduled in sequence, so
+ // once every cycle (when CurrMOps == 0), switch to normal heuristics.
+ if (Rem.IsAcyclicLatencyLimited && !Zone.CurrMOps
+ && tryLatency(TryCand, Cand, Zone))
return;
// Keep clustered nodes together to encourage downstream peephole
return;
}
// Avoid increasing the max pressure of the entire region.
- if (tryPressure(TryCand.RPDelta.CurrentMax, Cand.RPDelta.CurrentMax,
- TryCand, Cand, RegMax))
+ if (DAG->isTrackingPressure() && tryPressure(TryCand.RPDelta.CurrentMax,
+ Cand.RPDelta.CurrentMax,
+ TryCand, Cand, RegMax))
return;
// Avoid critical resource consumption and balance the schedule.
}
void ConvergingScheduler::traceCandidate(const SchedCandidate &Cand) {
- PressureElement P;
+ PressureChange P;
unsigned ResIdx = 0;
unsigned Latency = 0;
switch (Cand.Reason) {
}
dbgs() << " SU(" << Cand.SU->NodeNum << ") " << getReasonStr(Cand.Reason);
if (P.isValid())
- dbgs() << " " << TRI->getRegPressureSetName(P.PSetID)
- << ":" << P.UnitIncrease << " ";
+ dbgs() << " " << TRI->getRegPressureSetName(P.getPSet())
+ << ":" << P.getUnitInc() << " ";
else
dbgs() << " ";
if (ResIdx)
}
#endif
-/// Pick the best candidate from the top queue.
+/// Pick the best candidate from the queue.
///
/// TODO: getMaxPressureDelta results can be mostly cached for each SUnit during
/// DAG building. To adjust for the current scheduling location we need to
}
SUnit *SU;
do {
- if (ForceTopDown) {
+ if (RegionPolicy.OnlyTopDown) {
SU = Top.pickOnlyChoice();
if (!SU) {
CandPolicy NoPolicy;
SchedCandidate TopCand(NoPolicy);
pickNodeFromQueue(Top, DAG->getTopRPTracker(), TopCand);
- assert(TopCand.Reason != NoCand && "failed to find the first candidate");
+ assert(TopCand.Reason != NoCand && "failed to find a candidate");
+ tracePick(TopCand, true);
SU = TopCand.SU;
}
IsTopNode = true;
}
- else if (ForceBottomUp) {
+ else if (RegionPolicy.OnlyBottomUp) {
SU = Bot.pickOnlyChoice();
if (!SU) {
CandPolicy NoPolicy;
SchedCandidate BotCand(NoPolicy);
pickNodeFromQueue(Bot, DAG->getBotRPTracker(), BotCand);
- assert(BotCand.Reason != NoCand && "failed to find the first candidate");
+ assert(BotCand.Reason != NoCand && "failed to find a candidate");
+ tracePick(BotCand, false);
SU = BotCand.SU;
}
IsTopNode = false;
/// 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");
- ScheduleDAGMI *DAG = new ScheduleDAGMI(C, new ConvergingScheduler());
+ ScheduleDAGMI *DAG = new ScheduleDAGMI(C, new ConvergingScheduler(C));
// 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)
+ if (EnableLoadCluster && DAG->TII->enableClusterLoads())
DAG->addMutation(new LoadClusterMutation(DAG->TII, DAG->TRI));
if (EnableMacroFusion)
DAG->addMutation(new MacroFusion(DAG->TII));
/// \brief Schedule based on the ILP metric.
class ILPScheduler : public MachineSchedStrategy {
- /// In case all subtrees are eventually connected to a common root through
- /// data dependence (e.g. reduction), place an upper limit on their size.
- ///
- /// FIXME: A subtree limit is generally good, but in the situation commented
- /// above, where multiple similar subtrees feed a common root, we should
- /// only split at a point where the resulting subtrees will be balanced.
- /// (a motivating test case must be found).
- static const unsigned SubtreeLimit = 16;
-
ScheduleDAGMI *DAG;
ILPOrder Cmp;
}
static bool isNodeHidden(const SUnit *Node) {
- return (Node->NumPreds > 10 || Node->NumSuccs > 10);
+ return (Node->Preds.size() > 10 || Node->Succs.size() > 10);
}
static bool hasNodeAddressLabel(const SUnit *Node,
static std::string getNodeLabel(const SUnit *SU, const ScheduleDAG *G) {
std::string Str;
raw_string_ostream SS(Str);
- SS << "SU(" << SU->NodeNum << ')';
+ const SchedDFSResult *DFS =
+ static_cast<const ScheduleDAGMI*>(G)->getDFSResult();
+ SS << "SU:" << SU->NodeNum;
+ if (DFS)
+ SS << " I:" << DFS->getNumInstrs(SU);
return SS.str();
}
static std::string getNodeDescription(const SUnit *SU, const ScheduleDAG *G) {