#define DEBUG_TYPE "misched"
-#include "llvm/CodeGen/LiveIntervalAnalysis.h"
#include "llvm/CodeGen/MachineScheduler.h"
+#include "llvm/ADT/OwningPtr.h"
+#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/Passes.h"
#include "llvm/CodeGen/RegisterClassInfo.h"
-#include "llvm/CodeGen/ScheduleDAGILP.h"
+#include "llvm/CodeGen/ScheduleDFS.h"
#include "llvm/CodeGen/ScheduleHazardRecognizer.h"
-#include "llvm/Analysis/AliasAnalysis.h"
#include "llvm/Support/CommandLine.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/ErrorHandling.h"
+#include "llvm/Support/GraphWriter.h"
#include "llvm/Support/raw_ostream.h"
-#include "llvm/ADT/OwningPtr.h"
-#include "llvm/ADT/PriorityQueue.h"
-
#include <queue>
using namespace llvm;
static bool ViewMISchedDAGs = false;
#endif // NDEBUG
-// Threshold to very roughly model an out-of-order processor's instruction
-// buffers. If the actual value of this threshold matters much in practice, then
-// it can be specified by the machine model. For now, it's an experimental
-// tuning knob to determine when and if it matters.
-static cl::opt<unsigned> ILPWindow("ilp-window", cl::Hidden,
- cl::desc("Allow expected latency to exceed the critical path by N cycles "
- "before attempting to balance ILP"),
- cl::init(10U));
-
// Experimental heuristics
static cl::opt<bool> EnableLoadCluster("misched-cluster", cl::Hidden,
- cl::desc("Enable load clustering."));
+ cl::desc("Enable load clustering."), cl::init(true));
// Experimental heuristics
static cl::opt<bool> EnableMacroFusion("misched-fusion", cl::Hidden,
- cl::desc("Enable scheduling for macro fusion."));
+ 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;
//===----------------------------------------------------------------------===//
// Machine Instruction Scheduling Pass and Registry
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.
}
DEBUG(dbgs() << "********** MI Scheduling **********\n");
DEBUG(dbgs() << MF->getName()
- << ":BB#" << MBB->getNumber() << "\n From: " << *I << " To: ";
+ << ":BB#" << MBB->getNumber() << " " << MBB->getName()
+ << "\n From: " << *I << " To: ";
if (RegionEnd != MBB->end()) dbgs() << *RegionEnd;
else dbgs() << "End";
dbgs() << " Remaining: " << RemainingInstrs << "\n");
}
Scheduler->finalizeSchedule();
DEBUG(LIS->print(dbgs()));
+ if (VerifyScheduling)
+ MF->verify(this, "After machine scheduling.");
return true;
}
#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
void ReadyQueue::dump() {
- dbgs() << Name << ": ";
+ dbgs() << " " << Name << ": ";
for (unsigned i = 0, e = Queue.size(); i < e; ++i)
dbgs() << Queue[i]->NodeNum << " ";
dbgs() << "\n";
// preservation.
//===----------------------------------------------------------------------===//
+ScheduleDAGMI::~ScheduleDAGMI() {
+ delete DFSResult;
+ DeleteContainerPointers(Mutations);
+ delete SchedImpl;
+}
+
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)
// 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;
postprocessDAG();
+ SmallVector<SUnit*, 8> TopRoots, BotRoots;
+ findRootsAndBiasEdges(TopRoots, BotRoots);
+
+ // Initialize the strategy before modifying the DAG.
+ // This may initialize a DFSResult to be used for queue priority.
+ SchedImpl->initialize(this);
+
DEBUG(for (unsigned su = 0, e = SUnits.size(); su != e; ++su)
SUnits[su].dumpAll(this));
-
if (ViewMISchedDAGs) viewGraph();
- initQueues();
+ // Initialize ready queues now that the DAG and priority data are finalized.
+ initQueues(TopRoots, BotRoots);
bool IsTopNode = false;
while (SUnit *SU = SchedImpl->pickNode(IsTopNode)) {
placeDebugValues();
DEBUG({
- unsigned BBNum = top()->getParent()->getNumber();
+ unsigned BBNum = begin()->getParent()->getNumber();
dbgs() << "*** Final schedule for BB#" << BBNum << " ***\n";
dumpSchedule();
dbgs() << '\n';
// Build the DAG, and compute current register pressure.
buildSchedGraph(AA, &RPTracker);
- if (ViewMISchedDAGs) viewGraph();
// Initialize top/bottom trackers after computing region pressure.
initRegPressure();
}
}
-// Release all DAG roots for scheduling.
-//
-// Nodes with unreleased weak edges can still be roots.
-void ScheduleDAGMI::releaseRoots() {
- SmallVector<SUnit*, 16> BotRoots;
+void ScheduleDAGMI::computeDFSResult() {
+ if (!DFSResult)
+ DFSResult = new SchedDFSResult(/*BottomU*/true, MinSubtreeSize);
+ DFSResult->clear();
+ ScheduledTrees.clear();
+ DFSResult->resize(SUnits.size());
+ DFSResult->compute(SUnits);
+ ScheduledTrees.resize(DFSResult->getNumSubtrees());
+}
+void ScheduleDAGMI::findRootsAndBiasEdges(SmallVectorImpl<SUnit*> &TopRoots,
+ SmallVectorImpl<SUnit*> &BotRoots) {
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)
- SchedImpl->releaseTopNode(SU);
+ 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);
}
- // 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);
+ ExitSU.biasCriticalPath();
}
/// Identify DAG roots and setup scheduler queues.
-void ScheduleDAGMI::initQueues() {
+void ScheduleDAGMI::initQueues(ArrayRef<SUnit*> TopRoots,
+ ArrayRef<SUnit*> BotRoots) {
NextClusterSucc = NULL;
NextClusterPred = NULL;
- // Initialize the strategy before modifying the DAG.
- SchedImpl->initialize(this);
-
// Release all DAG roots for scheduling, not including EntrySU/ExitSU.
- releaseRoots();
+ //
+ // Nodes with unreleased weak edges can still be roots.
+ // Release top roots in forward order.
+ for (SmallVectorImpl<SUnit*>::const_iterator
+ I = TopRoots.begin(), E = TopRoots.end(); I != E; ++I) {
+ SchedImpl->releaseTopNode(*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);
+ }
releaseSuccessors(&EntrySU);
releasePredecessors(&ExitSU);
SchedImpl->registerRoots();
+ // Advance past initial DebugValues.
+ assert(TopRPTracker.getPos() == RegionBegin && "bad initial Top tracker");
CurrentTop = nextIfDebug(RegionBegin, RegionEnd);
+ TopRPTracker.setPos(CurrentTop);
+
CurrentBottom = RegionEnd;
}
SU->isScheduled = true;
+ if (DFSResult) {
+ unsigned SubtreeID = DFSResult->getSubtreeID(SU);
+ if (!ScheduledTrees.test(SubtreeID)) {
+ ScheduledTrees.set(SubtreeID);
+ DFSResult->scheduleTree(SubtreeID);
+ SchedImpl->scheduleTree(SubtreeID);
+ }
+ }
+
// Notify the scheduling strategy after updating the DAG.
SchedImpl->schedNode(SU, IsTopNode);
}
std::pair<MachineInstr *, MachineInstr *> P = *prior(DI);
MachineInstr *DbgValue = P.first;
MachineBasicBlock::iterator OrigPrevMI = P.second;
+ if (&*RegionBegin == DbgValue)
+ ++RegionBegin;
BB->splice(++OrigPrevMI, BB, DbgValue);
if (OrigPrevMI == llvm::prior(RegionEnd))
RegionEnd = DbgValue;
SUnit *SUa = LoadRecords[Idx].SU;
SUnit *SUb = LoadRecords[Idx+1].SU;
- if (TII->shouldScheduleLoadsNear(SUa->getInstr(), SUb->getInstr(),
- ClusterLength)
+ if (TII->shouldClusterLoads(SUa->getInstr(), SUb->getInstr(), ClusterLength)
&& DAG->addEdge(SUb, SDep(SUa, SDep::Cluster))) {
DEBUG(dbgs() << "Cluster loads SU(" << SUa->NodeNum << ") - SU("
/// 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, SingleExcess, SingleCritical, Cluster,
ResourceReduce, ResourceDemand, BotHeightReduce, BotPathReduce,
TopDepthReduce, TopPathReduce, SingleMax, MultiPressure, NextDefUse,
NodeOrder};
unsigned CritResIdx;
// Number of micro-ops left to schedule.
unsigned RemainingMicroOps;
- // Is the unscheduled zone resource limited.
- bool IsResourceLimited;
-
- unsigned MaxRemainingCount;
void reset() {
CriticalPath = 0;
RemainingCounts.clear();
CritResIdx = 0;
RemainingMicroOps = 0;
- IsResourceLimited = false;
- MaxRemainingCount = 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
unsigned ExpectedCount;
- // Policy flag: attempt to find ILP until expected latency is covered.
- bool ShouldIncreaseILP;
-
#ifndef NDEBUG
// Remember the greatest min operand latency.
unsigned MaxMinLatency;
#endif
void reset() {
+ // A new HazardRec is created for each DAG and owned by SchedBoundary.
+ delete HazardRec;
+
Available.clear();
Pending.clear();
CheckPending = false;
CritResIdx = 0;
IsResourceLimited = false;
ExpectedCount = 0;
- ShouldIncreaseILP = false;
#ifndef NDEBUG
MaxMinLatency = 0;
#endif
/// 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();
}
unsigned getUnscheduledLatency(SUnit *SU) const {
if (isTop())
return SU->getHeight();
- return SU->getDepth();
+ return SU->getDepth() + SU->Latency;
}
unsigned getCriticalCount() const {
bool checkHazard(SUnit *SU);
- void checkILPPolicy();
+ void setLatencyPolicy(CandPolicy &Policy);
void releaseNode(SUnit *SU, unsigned ReadyCycle);
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
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::
DAG = dag;
SchedModel = DAG->getSchedModel();
TRI = DAG->TRI;
+
Rem.init(DAG, SchedModel);
Top.init(DAG, SchedModel, &Rem);
Bot.init(DAG, SchedModel, &Rem);
if (SU->isScheduled)
return;
- for (SUnit::succ_iterator I = SU->Preds.begin(), E = SU->Preds.end();
+ for (SUnit::pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
I != E; ++I) {
unsigned PredReadyCycle = I->getSUnit()->TopReadyCycle;
unsigned MinLatency = I->getMinLatency();
return false;
}
-/// If expected latency is covered, disable ILP policy.
-void ConvergingScheduler::SchedBoundary::checkILPPolicy() {
- if (ShouldIncreaseILP
- && (IsResourceLimited || ExpectedLatency <= CurrCycle)) {
- ShouldIncreaseILP = false;
- DEBUG(dbgs() << "Disable ILP: " << Available.getName() << '\n');
+/// 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.
+ unsigned RemLatency = 0;
+ for (ReadyQueue::iterator I = Available.begin(), E = Available.end();
+ I != E; ++I) {
+ unsigned L = getUnscheduledLatency(*I);
+ DEBUG(dbgs() << " " << Available.getName()
+ << " RemLatency SU(" << (*I)->NodeNum << ") " << L << '\n');
+ if (L > RemLatency)
+ RemLatency = L;
+ }
+ for (ReadyQueue::iterator I = Pending.begin(), E = Pending.end();
+ I != E; ++I) {
+ unsigned L = getUnscheduledLatency(*I);
+ if (L > RemLatency)
+ RemLatency = L;
+ }
+ unsigned CriticalPathLimit = Rem->CriticalPath + SchedModel->getILPWindow();
+ DEBUG(dbgs() << " " << Available.getName()
+ << " ExpectedLatency " << ExpectedLatency
+ << " CP Limit " << CriticalPathLimit << '\n');
+ if (RemLatency + ExpectedLatency >= CriticalPathLimit
+ && RemLatency > Rem->getMaxRemainingCount(SchedModel)) {
+ Policy.ReduceLatency = true;
+ DEBUG(dbgs() << " Increase ILP: " << Available.getName() << '\n');
}
}
// Record this node as an immediate dependent of the scheduled node.
NextSUs.insert(SU);
-
- // If CriticalPath has been computed, then check if the unscheduled nodes
- // exceed the ILP window. Before registerRoots, CriticalPath==0.
- if (Rem->CriticalPath && (ExpectedLatency + getUnscheduledLatency(SU)
- > Rem->CriticalPath + ILPWindow)) {
- ShouldIncreaseILP = true;
- DEBUG(dbgs() << "Increase ILP: " << Available.getName() << " "
- << ExpectedLatency << " + " << getUnscheduledLatency(SU) << '\n');
- }
}
/// Move the boundary of scheduled code by one cycle.
CheckPending = true;
IsResourceLimited = getCriticalCount() > std::max(ExpectedLatency, CurrCycle);
- DEBUG(dbgs() << " *** " << Available.getName() << " cycle "
- << CurrCycle << '\n');
+ DEBUG(dbgs() << " " << Available.getName()
+ << " Cycle: " << CurrCycle << '\n');
}
/// Add the given processor resource to this scheduled zone.
assert(Rem->RemainingCounts[PIdx] >= Count && "resource double counted");
Rem->RemainingCounts[PIdx] -= Count;
- // Reset MaxRemainingCount for sanity.
- Rem->MaxRemainingCount = 0;
-
// 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])
/// resources.
///
/// If the CriticalZone is latency limited, don't force a policy for the
-/// candidates here. Instead, When releasing each candidate, releaseNode
-/// compares the region's critical path to the candidate's height or depth and
-/// the scheduled zone's expected latency then sets ShouldIncreaseILP.
+/// candidates here. Instead, setLatencyPolicy sets ReduceLatency if needed.
void ConvergingScheduler::balanceZones(
ConvergingScheduler::SchedBoundary &CriticalZone,
ConvergingScheduler::SchedCandidate &CriticalCand,
if (!CriticalZone.IsResourceLimited)
return;
+ assert(SchedModel->hasInstrSchedModel() && "required schedmodel");
SchedRemainder *Rem = CriticalZone.Rem;
// remainder, try to reduce it.
unsigned RemainingCritCount =
Rem->RemainingCounts[CriticalZone.CritResIdx];
- if ((int)(Rem->MaxRemainingCount - RemainingCritCount)
+ if ((int)(Rem->getMaxRemainingCount(SchedModel) - RemainingCritCount)
> (int)SchedModel->getLatencyFactor()) {
CriticalCand.Policy.ReduceResIdx = CriticalZone.CritResIdx;
- DEBUG(dbgs() << "Balance " << CriticalZone.Available.getName() << " reduce "
+ DEBUG(dbgs() << " Balance " << CriticalZone.Available.getName()
+ << " reduce "
<< SchedModel->getProcResource(CriticalZone.CritResIdx)->Name
<< '\n');
}
if ((int)(OppositeZone.ExpectedCount - OppositeCount)
> (int)SchedModel->getLatencyFactor()) {
OppositeCand.Policy.DemandResIdx = CriticalZone.CritResIdx;
- DEBUG(dbgs() << "Balance " << OppositeZone.Available.getName() << " demand "
+ DEBUG(dbgs() << " Balance " << OppositeZone.Available.getName()
+ << " demand "
<< SchedModel->getProcResource(OppositeZone.CritResIdx)->Name
<< '\n');
}
ConvergingScheduler::SchedCandidate &TopCand,
ConvergingScheduler::SchedCandidate &BotCand) {
- Bot.checkILPPolicy();
- Top.checkILPPolicy();
- if (Bot.ShouldIncreaseILP)
- BotCand.Policy.ReduceLatency = true;
- if (Top.ShouldIncreaseILP)
- TopCand.Policy.ReduceLatency = true;
+ // Set ReduceLatency to true if needed.
+ Bot.setLatencyPolicy(BotCand.Policy);
+ Top.setLatencyPolicy(TopCand.Policy);
// Handle resource-limited regions.
if (Top.IsResourceLimited && Bot.IsResourceLimited
if (Top.CritResIdx != Rem.CritResIdx) {
TopCand.Policy.ReduceResIdx = Top.CritResIdx;
BotCand.Policy.ReduceResIdx = Bot.CritResIdx;
- DEBUG(dbgs() << "Reduce scheduled "
+ DEBUG(dbgs() << " Reduce scheduled "
<< SchedModel->getProcResource(Top.CritResIdx)->Name << '\n');
}
return;
&& (Rem.CriticalPath > Top.CurrCycle + Bot.CurrCycle)) {
TopCand.Policy.ReduceLatency = true;
BotCand.Policy.ReduceLatency = true;
- DEBUG(dbgs() << "Reduce scheduled latency " << Top.ExpectedLatency
+ DEBUG(dbgs() << " Reduce scheduled latency " << Top.ExpectedLatency
<< " + " << Bot.ExpectedLatency << '\n');
}
return;
// The critical resource is different in each zone, so request balancing.
// Compute the cost of each zone.
- Rem.MaxRemainingCount = std::max(
- Rem.RemainingMicroOps * SchedModel->getMicroOpFactor(),
- Rem.RemainingCounts[Rem.CritResIdx]);
Top.ExpectedCount = std::max(Top.ExpectedLatency, Top.CurrCycle);
Top.ExpectedCount = std::max(
Top.getCriticalCount(),
}
/// 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) {
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) {
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))
// Avoid increasing the max critical pressure in the scheduled region.
if (LHS.Excess.UnitIncrease != RHS.Excess.UnitIncrease) {
- DEBUG(dbgs() << "RP excess top - bot: "
+ 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: "
+ 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: "
+ DEBUG(dbgs() << " RP current top - bot: "
<< (LHS.CurrentMax.UnitIncrease - RHS.CurrentMax.UnitIncrease)
<< '\n');
return LHS.CurrentMax.UnitIncrease < RHS.CurrentMax.UnitIncrease;
ConvergingScheduler::CandReason Reason) {
switch (Reason) {
case NoCand: return "NOCAND ";
+ case PhysRegCopy: return "PREG-COPY";
case SingleExcess: return "REG-EXCESS";
case SingleCritical: return "REG-CRIT ";
case Cluster: return "CLUSTER ";
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;
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));
}
- TryCand.SU = *I;
}
}
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 Top NOCAND\n");
return SU;
}
if (SUnit *SU = Top.pickOnlyChoice()) {
IsTopNode = true;
+ DEBUG(dbgs() << "Pick Bot NOCAND\n");
return SU;
}
CandPolicy NoPolicy;
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;
Top.bumpNode(SU);
+ if (SU->hasPhysRegUses)
+ reschedulePhysRegCopies(SU, true);
}
else {
SU->BotReadyCycle = Bot.CurrCycle;
Bot.bumpNode(SU);
+ if (SU->hasPhysRegDefs)
+ reschedulePhysRegCopies(SU, false);
}
}
namespace {
/// \brief Order nodes by the ILP metric.
struct ILPOrder {
- ScheduleDAGILP *ILP;
+ const SchedDFSResult *DFSResult;
+ const BitVector *ScheduledTrees;
bool MaximizeILP;
- ILPOrder(ScheduleDAGILP *ilp, bool MaxILP): ILP(ilp), MaximizeILP(MaxILP) {}
+ ILPOrder(bool MaxILP): DFSResult(0), ScheduledTrees(0), MaximizeILP(MaxILP) {}
/// \brief Apply a less-than relation on node priority.
+ ///
+ /// (Return true if A comes after B in the Q.)
bool operator()(const SUnit *A, const SUnit *B) const {
- // Return true if A comes after B in the Q.
+ unsigned SchedTreeA = DFSResult->getSubtreeID(A);
+ unsigned SchedTreeB = DFSResult->getSubtreeID(B);
+ if (SchedTreeA != SchedTreeB) {
+ // Unscheduled trees have lower priority.
+ if (ScheduledTrees->test(SchedTreeA) != ScheduledTrees->test(SchedTreeB))
+ return ScheduledTrees->test(SchedTreeB);
+
+ // Trees with shallower connections have have lower priority.
+ if (DFSResult->getSubtreeLevel(SchedTreeA)
+ != DFSResult->getSubtreeLevel(SchedTreeB)) {
+ return DFSResult->getSubtreeLevel(SchedTreeA)
+ < DFSResult->getSubtreeLevel(SchedTreeB);
+ }
+ }
if (MaximizeILP)
- return ILP->getILP(A) < ILP->getILP(B);
+ return DFSResult->getILP(A) < DFSResult->getILP(B);
else
- return ILP->getILP(A) > ILP->getILP(B);
+ return DFSResult->getILP(A) > DFSResult->getILP(B);
}
};
/// \brief Schedule based on the ILP metric.
class ILPScheduler : public MachineSchedStrategy {
- ScheduleDAGILP ILP;
+ /// 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;
std::vector<SUnit*> ReadyQ;
public:
- ILPScheduler(bool MaximizeILP)
- : ILP(/*BottomUp=*/true), Cmp(&ILP, MaximizeILP) {}
+ ILPScheduler(bool MaximizeILP): DAG(0), Cmp(MaximizeILP) {}
- virtual void initialize(ScheduleDAGMI *DAG) {
+ virtual void initialize(ScheduleDAGMI *dag) {
+ DAG = dag;
+ DAG->computeDFSResult();
+ Cmp.DFSResult = DAG->getDFSResult();
+ Cmp.ScheduledTrees = &DAG->getScheduledTrees();
ReadyQ.clear();
- ILP.resize(DAG->SUnits.size());
}
virtual void registerRoots() {
- for (std::vector<SUnit*>::const_iterator
- I = ReadyQ.begin(), E = ReadyQ.end(); I != E; ++I) {
- ILP.computeILP(*I);
- }
+ // Restore the heap in ReadyQ with the updated DFS results.
+ std::make_heap(ReadyQ.begin(), ReadyQ.end(), Cmp);
}
/// Implement MachineSchedStrategy interface.
/// -----------------------------------------
+ /// 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->getInstr()
- << " ILP: " << ILP.getILP(SU) << '\n');
+ 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'
+ << "Scheduling " << *SU->getInstr());
return SU;
}
- virtual void schedNode(SUnit *, bool) {}
+ /// \brief Scheduler callback to notify that a new subtree is scheduled.
+ virtual void scheduleTree(unsigned SubtreeID) {
+ std::make_heap(ReadyQ.begin(), ReadyQ.end(), Cmp);
+ }
+
+ /// Callback after a node is scheduled. Mark a newly scheduled tree, notify
+ /// DFSResults, and resort the priority Q.
+ virtual void schedNode(SUnit *SU, bool IsTopNode) {
+ assert(!IsTopNode && "SchedDFSResult needs bottom-up");
+ }
virtual void releaseTopNode(SUnit *) { /*only called for top roots*/ }
"shuffle", "Shuffle machine instructions alternating directions",
createInstructionShuffler);
#endif // !NDEBUG
+
+//===----------------------------------------------------------------------===//
+// GraphWriter support for ScheduleDAGMI.
+//===----------------------------------------------------------------------===//
+
+#ifndef NDEBUG
+namespace llvm {
+
+template<> struct GraphTraits<
+ ScheduleDAGMI*> : public GraphTraits<ScheduleDAG*> {};
+
+template<>
+struct DOTGraphTraits<ScheduleDAGMI*> : public DefaultDOTGraphTraits {
+
+ DOTGraphTraits (bool isSimple=false) : DefaultDOTGraphTraits(isSimple) {}
+
+ static std::string getGraphName(const ScheduleDAG *G) {
+ return G->MF.getName();
+ }
+
+ static bool renderGraphFromBottomUp() {
+ return true;
+ }
+
+ static bool isNodeHidden(const SUnit *Node) {
+ return (Node->NumPreds > 10 || Node->NumSuccs > 10);
+ }
+
+ static bool hasNodeAddressLabel(const SUnit *Node,
+ const ScheduleDAG *Graph) {
+ return false;
+ }
+
+ /// If you want to override the dot attributes printed for a particular
+ /// edge, override this method.
+ static std::string getEdgeAttributes(const SUnit *Node,
+ SUnitIterator EI,
+ const ScheduleDAG *Graph) {
+ if (EI.isArtificialDep())
+ return "color=cyan,style=dashed";
+ if (EI.isCtrlDep())
+ return "color=blue,style=dashed";
+ return "";
+ }
+
+ static std::string getNodeLabel(const SUnit *SU, const ScheduleDAG *G) {
+ std::string Str;
+ raw_string_ostream SS(Str);
+ SS << "SU(" << SU->NodeNum << ')';
+ return SS.str();
+ }
+ static std::string getNodeDescription(const SUnit *SU, const ScheduleDAG *G) {
+ return G->getGraphNodeLabel(SU);
+ }
+
+ static std::string getNodeAttributes(const SUnit *N,
+ const ScheduleDAG *Graph) {
+ std::string Str("shape=Mrecord");
+ const SchedDFSResult *DFS =
+ static_cast<const ScheduleDAGMI*>(Graph)->getDFSResult();
+ if (DFS) {
+ Str += ",style=filled,fillcolor=\"#";
+ Str += DOT::getColorString(DFS->getSubtreeID(N));
+ Str += '"';
+ }
+ return Str;
+ }
+};
+} // namespace llvm
+#endif // NDEBUG
+
+/// viewGraph - Pop up a ghostview window with the reachable parts of the DAG
+/// rendered using 'dot'.
+///
+void ScheduleDAGMI::viewGraph(const Twine &Name, const Twine &Title) {
+#ifndef NDEBUG
+ ViewGraph(this, Name, false, Title);
+#else
+ errs() << "ScheduleDAGMI::viewGraph is only available in debug builds on "
+ << "systems with Graphviz or gv!\n";
+#endif // NDEBUG
+}
+
+/// Out-of-line implementation with no arguments is handy for gdb.
+void ScheduleDAGMI::viewGraph() {
+ viewGraph(getDAGName(), "Scheduling-Units Graph for " + getDAGName());
+}
projects / oota-llvm.git / blobdiff