-//===----- ScheduleDAGList.cpp - Reg pressure reduction list scheduler ----===//
+//===----- ScheduleDAGRRList.cpp - Reg pressure reduction list scheduler --===//
//
// The LLVM Compiler Infrastructure
//
//===----------------------------------------------------------------------===//
#define DEBUG_TYPE "pre-RA-sched"
-#include "llvm/CodeGen/ScheduleDAG.h"
+#include "ScheduleDAGSDNodes.h"
+#include "llvm/InlineAsm.h"
#include "llvm/CodeGen/SchedulerRegistry.h"
-#include "llvm/Target/MRegisterInfo.h"
+#include "llvm/CodeGen/SelectionDAGISel.h"
+#include "llvm/CodeGen/ScheduleHazardRecognizer.h"
+#include "llvm/Target/TargetRegisterInfo.h"
#include "llvm/Target/TargetData.h"
#include "llvm/Target/TargetMachine.h"
#include "llvm/Target/TargetInstrInfo.h"
-#include "llvm/Support/Debug.h"
-#include "llvm/Support/Compiler.h"
-#include "llvm/ADT/SmallPtrSet.h"
+#include "llvm/Target/TargetLowering.h"
#include "llvm/ADT/SmallSet.h"
#include "llvm/ADT/Statistic.h"
+#include "llvm/ADT/STLExtras.h"
+#include "llvm/Support/Debug.h"
+#include "llvm/Support/ErrorHandling.h"
+#include "llvm/Support/raw_ostream.h"
#include <climits>
-#include <queue>
-#include "llvm/Support/CommandLine.h"
using namespace llvm;
-STATISTIC(NumBacktracks, "Number of times scheduler backtraced");
+STATISTIC(NumBacktracks, "Number of times scheduler backtracked");
STATISTIC(NumUnfolds, "Number of nodes unfolded");
STATISTIC(NumDups, "Number of duplicated nodes");
-STATISTIC(NumCCCopies, "Number of cross class copies");
+STATISTIC(NumPRCopies, "Number of physical register copies");
static RegisterScheduler
burrListDAGScheduler("list-burr",
- " Bottom-up register reduction list scheduling",
+ "Bottom-up register reduction list scheduling",
createBURRListDAGScheduler);
static RegisterScheduler
tdrListrDAGScheduler("list-tdrr",
- " Top-down register reduction list scheduling",
+ "Top-down register reduction list scheduling",
createTDRRListDAGScheduler);
+static RegisterScheduler
+ sourceListDAGScheduler("source",
+ "Similar to list-burr but schedules in source "
+ "order when possible",
+ createSourceListDAGScheduler);
+
+static RegisterScheduler
+ hybridListDAGScheduler("list-hybrid",
+ "Bottom-up register pressure aware list scheduling "
+ "which tries to balance latency and register pressure",
+ createHybridListDAGScheduler);
+
+static RegisterScheduler
+ ILPListDAGScheduler("list-ilp",
+ "Bottom-up register pressure aware list scheduling "
+ "which tries to balance ILP and register pressure",
+ createILPListDAGScheduler);
+
+static cl::opt<bool> DisableSchedCycles(
+ "disable-sched-cycles", cl::Hidden, cl::init(false),
+ cl::desc("Disable cycle-level precision during preRA scheduling"));
+
+// Temporary sched=list-ilp flags until the heuristics are robust.
+static cl::opt<bool> DisableSchedRegPressure(
+ "disable-sched-reg-pressure", cl::Hidden, cl::init(false),
+ cl::desc("Disable regpressure priority in sched=list-ilp"));
+static cl::opt<bool> DisableSchedLiveUses(
+ "disable-sched-live-uses", cl::Hidden, cl::init(false),
+ cl::desc("Disable live use priority in sched=list-ilp"));
+static cl::opt<bool> DisableSchedStalls(
+ "disable-sched-stalls", cl::Hidden, cl::init(false),
+ cl::desc("Disable no-stall priority in sched=list-ilp"));
+static cl::opt<bool> DisableSchedCriticalPath(
+ "disable-sched-critical-path", cl::Hidden, cl::init(false),
+ cl::desc("Disable critical path priority in sched=list-ilp"));
+static cl::opt<bool> DisableSchedHeight(
+ "disable-sched-height", cl::Hidden, cl::init(false),
+ cl::desc("Disable scheduled-height priority in sched=list-ilp"));
+
+static cl::opt<int> MaxReorderWindow(
+ "max-sched-reorder", cl::Hidden, cl::init(6),
+ cl::desc("Number of instructions to allow ahead of the critical path "
+ "in sched=list-ilp"));
+
+static cl::opt<unsigned> AvgIPC(
+ "sched-avg-ipc", cl::Hidden, cl::init(1),
+ cl::desc("Average inst/cycle whan no target itinerary exists."));
+
+#ifndef NDEBUG
+namespace {
+ // For sched=list-ilp, Count the number of times each factor comes into play.
+ enum { FactPressureDiff, FactRegUses, FactHeight, FactDepth, FactUllman,
+ NumFactors };
+}
+static const char *FactorName[NumFactors] =
+{"PressureDiff", "RegUses", "Height", "Depth","Ullman"};
+static int FactorCount[NumFactors];
+#endif //!NDEBUG
namespace {
//===----------------------------------------------------------------------===//
/// ScheduleDAGRRList - The actual register reduction list scheduler
/// implementation. This supports both top-down and bottom-up scheduling.
///
-class VISIBILITY_HIDDEN ScheduleDAGRRList : public ScheduleDAG {
+class ScheduleDAGRRList : public ScheduleDAGSDNodes {
private:
/// isBottomUp - This is true if the scheduling problem is bottom-up, false if
/// it is top-down.
bool isBottomUp;
-
+
+ /// NeedLatency - True if the scheduler will make use of latency information.
+ ///
+ bool NeedLatency;
+
/// AvailableQueue - The priority queue to use for the available SUnits.
- ///a
SchedulingPriorityQueue *AvailableQueue;
- /// LiveRegs / LiveRegDefs - A set of physical registers and their definition
+ /// PendingQueue - This contains all of the instructions whose operands have
+ /// been issued, but their results are not ready yet (due to the latency of
+ /// the operation). Once the operands becomes available, the instruction is
+ /// added to the AvailableQueue.
+ std::vector<SUnit*> PendingQueue;
+
+ /// HazardRec - The hazard recognizer to use.
+ ScheduleHazardRecognizer *HazardRec;
+
+ /// CurCycle - The current scheduler state corresponds to this cycle.
+ unsigned CurCycle;
+
+ /// MinAvailableCycle - Cycle of the soonest available instruction.
+ unsigned MinAvailableCycle;
+
+ /// IssueCount - Count instructions issued in this cycle
+ /// Currently valid only for bottom-up scheduling.
+ unsigned IssueCount;
+
+ /// LiveRegDefs - A set of physical registers and their definition
/// that are "live". These nodes must be scheduled before any other nodes that
/// modifies the registers can be scheduled.
- SmallSet<unsigned, 4> LiveRegs;
+ unsigned NumLiveRegs;
std::vector<SUnit*> LiveRegDefs;
- std::vector<unsigned> LiveRegCycles;
+ std::vector<SUnit*> LiveRegGens;
+
+ /// Topo - A topological ordering for SUnits which permits fast IsReachable
+ /// and similar queries.
+ ScheduleDAGTopologicalSort Topo;
public:
- ScheduleDAGRRList(SelectionDAG &dag, MachineBasicBlock *bb,
- const TargetMachine &tm, bool isbottomup,
- SchedulingPriorityQueue *availqueue)
- : ScheduleDAG(dag, bb, tm), isBottomUp(isbottomup),
- AvailableQueue(availqueue) {
- }
+ ScheduleDAGRRList(MachineFunction &mf, bool needlatency,
+ SchedulingPriorityQueue *availqueue,
+ CodeGenOpt::Level OptLevel)
+ : ScheduleDAGSDNodes(mf), isBottomUp(availqueue->isBottomUp()),
+ NeedLatency(needlatency), AvailableQueue(availqueue), CurCycle(0),
+ Topo(SUnits) {
+
+ const TargetMachine &tm = mf.getTarget();
+ if (DisableSchedCycles || !NeedLatency)
+ HazardRec = new ScheduleHazardRecognizer();
+ else
+ HazardRec = tm.getInstrInfo()->CreateTargetHazardRecognizer(&tm, this);
+ }
~ScheduleDAGRRList() {
+ delete HazardRec;
delete AvailableQueue;
}
void Schedule();
+ ScheduleHazardRecognizer *getHazardRec() { return HazardRec; }
+
+ /// IsReachable - Checks if SU is reachable from TargetSU.
+ bool IsReachable(const SUnit *SU, const SUnit *TargetSU) {
+ return Topo.IsReachable(SU, TargetSU);
+ }
+
+ /// WillCreateCycle - Returns true if adding an edge from SU to TargetSU will
+ /// create a cycle.
+ bool WillCreateCycle(SUnit *SU, SUnit *TargetSU) {
+ return Topo.WillCreateCycle(SU, TargetSU);
+ }
+
+ /// AddPred - adds a predecessor edge to SUnit SU.
+ /// This returns true if this is a new predecessor.
+ /// Updates the topological ordering if required.
+ void AddPred(SUnit *SU, const SDep &D) {
+ Topo.AddPred(SU, D.getSUnit());
+ SU->addPred(D);
+ }
+
+ /// RemovePred - removes a predecessor edge from SUnit SU.
+ /// This returns true if an edge was removed.
+ /// Updates the topological ordering if required.
+ void RemovePred(SUnit *SU, const SDep &D) {
+ Topo.RemovePred(SU, D.getSUnit());
+ SU->removePred(D);
+ }
+
private:
- void ReleasePred(SUnit*, bool, unsigned);
- void ReleaseSucc(SUnit*, bool isChain, unsigned);
- void CapturePred(SUnit*, SUnit*, bool);
- void ScheduleNodeBottomUp(SUnit*, unsigned);
- void ScheduleNodeTopDown(SUnit*, unsigned);
+ bool isReady(SUnit *SU) {
+ return DisableSchedCycles || !AvailableQueue->hasReadyFilter() ||
+ AvailableQueue->isReady(SU);
+ }
+
+ void ReleasePred(SUnit *SU, const SDep *PredEdge);
+ void ReleasePredecessors(SUnit *SU);
+ void ReleaseSucc(SUnit *SU, const SDep *SuccEdge);
+ void ReleaseSuccessors(SUnit *SU);
+ void ReleasePending();
+ void AdvanceToCycle(unsigned NextCycle);
+ void AdvancePastStalls(SUnit *SU);
+ void EmitNode(SUnit *SU);
+ void ScheduleNodeBottomUp(SUnit*);
+ void CapturePred(SDep *PredEdge);
void UnscheduleNodeBottomUp(SUnit*);
- void BacktrackBottomUp(SUnit*, unsigned, unsigned&);
+ void RestoreHazardCheckerBottomUp();
+ void BacktrackBottomUp(SUnit*, SUnit*);
SUnit *CopyAndMoveSuccessors(SUnit*);
- void InsertCCCopiesAndMoveSuccs(SUnit*, unsigned,
- const TargetRegisterClass*,
- const TargetRegisterClass*,
- SmallVector<SUnit*, 2>&);
+ void InsertCopiesAndMoveSuccs(SUnit*, unsigned,
+ const TargetRegisterClass*,
+ const TargetRegisterClass*,
+ SmallVector<SUnit*, 2>&);
bool DelayForLiveRegsBottomUp(SUnit*, SmallVector<unsigned, 4>&);
- void ListScheduleTopDown();
+
+ SUnit *PickNodeToScheduleBottomUp();
void ListScheduleBottomUp();
- void CommuteNodesToReducePressure();
+
+ void ScheduleNodeTopDown(SUnit*);
+ void ListScheduleTopDown();
+
+
+ /// CreateNewSUnit - Creates a new SUnit and returns a pointer to it.
+ /// Updates the topological ordering if required.
+ SUnit *CreateNewSUnit(SDNode *N) {
+ unsigned NumSUnits = SUnits.size();
+ SUnit *NewNode = NewSUnit(N);
+ // Update the topological ordering.
+ if (NewNode->NodeNum >= NumSUnits)
+ Topo.InitDAGTopologicalSorting();
+ return NewNode;
+ }
+
+ /// CreateClone - Creates a new SUnit from an existing one.
+ /// Updates the topological ordering if required.
+ SUnit *CreateClone(SUnit *N) {
+ unsigned NumSUnits = SUnits.size();
+ SUnit *NewNode = Clone(N);
+ // Update the topological ordering.
+ if (NewNode->NodeNum >= NumSUnits)
+ Topo.InitDAGTopologicalSorting();
+ return NewNode;
+ }
+
+ /// ForceUnitLatencies - Register-pressure-reducing scheduling doesn't
+ /// need actual latency information but the hybrid scheduler does.
+ bool ForceUnitLatencies() const {
+ return !NeedLatency;
+ }
};
} // end anonymous namespace
/// Schedule - Schedule the DAG using list scheduling.
void ScheduleDAGRRList::Schedule() {
- DOUT << "********** List Scheduling **********\n";
+ DEBUG(dbgs()
+ << "********** List Scheduling BB#" << BB->getNumber()
+ << " '" << BB->getName() << "' **********\n");
+#ifndef NDEBUG
+ for (int i = 0; i < NumFactors; ++i) {
+ FactorCount[i] = 0;
+ }
+#endif //!NDEBUG
- LiveRegDefs.resize(MRI->getNumRegs(), NULL);
- LiveRegCycles.resize(MRI->getNumRegs(), 0);
+ CurCycle = 0;
+ IssueCount = 0;
+ MinAvailableCycle = DisableSchedCycles ? 0 : UINT_MAX;
+ NumLiveRegs = 0;
+ LiveRegDefs.resize(TRI->getNumRegs(), NULL);
+ LiveRegGens.resize(TRI->getNumRegs(), NULL);
- // Build scheduling units.
- BuildSchedUnits();
+ // Build the scheduling graph.
+ BuildSchedGraph(NULL);
DEBUG(for (unsigned su = 0, e = SUnits.size(); su != e; ++su)
- SUnits[su].dumpAll(&DAG));
- CalculateDepths();
- CalculateHeights();
+ SUnits[su].dumpAll(this));
+ Topo.InitDAGTopologicalSorting();
+
+ AvailableQueue->initNodes(SUnits);
+
+ HazardRec->Reset();
- AvailableQueue->initNodes(SUnitMap, SUnits);
-
// Execute the actual scheduling loop Top-Down or Bottom-Up as appropriate.
if (isBottomUp)
ListScheduleBottomUp();
else
ListScheduleTopDown();
-
- AvailableQueue->releaseState();
-
- CommuteNodesToReducePressure();
-
- DOUT << "*** Final schedule ***\n";
- DEBUG(dumpSchedule());
- DOUT << "\n";
-
- // Emit in scheduled order
- EmitSchedule();
-}
-
-/// CommuteNodesToReducePressure - If a node is two-address and commutable, and
-/// it is not the last use of its first operand, add it to the CommuteSet if
-/// possible. It will be commuted when it is translated to a MI.
-void ScheduleDAGRRList::CommuteNodesToReducePressure() {
- SmallPtrSet<SUnit*, 4> OperandSeen;
- for (unsigned i = Sequence.size()-1; i != 0; --i) { // Ignore first node.
- SUnit *SU = Sequence[i];
- if (!SU || !SU->Node) continue;
- if (SU->isCommutable) {
- unsigned Opc = SU->Node->getTargetOpcode();
- unsigned NumRes = TII->getNumDefs(Opc);
- unsigned NumOps = CountOperands(SU->Node);
- for (unsigned j = 0; j != NumOps; ++j) {
- if (TII->getOperandConstraint(Opc, j+NumRes, TOI::TIED_TO) == -1)
- continue;
-
- SDNode *OpN = SU->Node->getOperand(j).Val;
- SUnit *OpSU = isPassiveNode(OpN) ? NULL : SUnitMap[OpN][SU->InstanceNo];
- if (OpSU && OperandSeen.count(OpSU) == 1) {
- // Ok, so SU is not the last use of OpSU, but SU is two-address so
- // it will clobber OpSU. Try to commute SU if no other source operands
- // are live below.
- bool DoCommute = true;
- for (unsigned k = 0; k < NumOps; ++k) {
- if (k != j) {
- OpN = SU->Node->getOperand(k).Val;
- OpSU = isPassiveNode(OpN) ? NULL : SUnitMap[OpN][SU->InstanceNo];
- if (OpSU && OperandSeen.count(OpSU) == 1) {
- DoCommute = false;
- break;
- }
- }
- }
- if (DoCommute)
- CommuteSet.insert(SU->Node);
- }
-
- // Only look at the first use&def node for now.
- break;
- }
- }
- for (SUnit::pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
- I != E; ++I) {
- if (!I->isCtrl)
- OperandSeen.insert(I->Dep);
- }
+#ifndef NDEBUG
+ for (int i = 0; i < NumFactors; ++i) {
+ DEBUG(dbgs() << FactorName[i] << "\t" << FactorCount[i] << "\n");
}
+#endif // !NDEBUG
+ AvailableQueue->releaseState();
}
//===----------------------------------------------------------------------===//
/// ReleasePred - Decrement the NumSuccsLeft count of a predecessor. Add it to
/// the AvailableQueue if the count reaches zero. Also update its cycle bound.
-void ScheduleDAGRRList::ReleasePred(SUnit *PredSU, bool isChain,
- unsigned CurCycle) {
- // FIXME: the distance between two nodes is not always == the predecessor's
- // latency. For example, the reader can very well read the register written
- // by the predecessor later than the issue cycle. It also depends on the
- // interrupt model (drain vs. freeze).
- PredSU->CycleBound = std::max(PredSU->CycleBound, CurCycle + PredSU->Latency);
+void ScheduleDAGRRList::ReleasePred(SUnit *SU, const SDep *PredEdge) {
+ SUnit *PredSU = PredEdge->getSUnit();
- --PredSU->NumSuccsLeft;
-
#ifndef NDEBUG
- if (PredSU->NumSuccsLeft < 0) {
- cerr << "*** List scheduling failed! ***\n";
- PredSU->dump(&DAG);
- cerr << " has been released too many times!\n";
- assert(0);
+ if (PredSU->NumSuccsLeft == 0) {
+ dbgs() << "*** Scheduling failed! ***\n";
+ PredSU->dump(this);
+ dbgs() << " has been released too many times!\n";
+ llvm_unreachable(0);
}
#endif
-
- if (PredSU->NumSuccsLeft == 0) {
- // EntryToken has to go last! Special case it here.
- if (!PredSU->Node || PredSU->Node->getOpcode() != ISD::EntryToken) {
- PredSU->isAvailable = true;
+ --PredSU->NumSuccsLeft;
+
+ if (!ForceUnitLatencies()) {
+ // Updating predecessor's height. This is now the cycle when the
+ // predecessor can be scheduled without causing a pipeline stall.
+ PredSU->setHeightToAtLeast(SU->getHeight() + PredEdge->getLatency());
+ }
+
+ // If all the node's successors are scheduled, this node is ready
+ // to be scheduled. Ignore the special EntrySU node.
+ if (PredSU->NumSuccsLeft == 0 && PredSU != &EntrySU) {
+ PredSU->isAvailable = true;
+
+ unsigned Height = PredSU->getHeight();
+ if (Height < MinAvailableCycle)
+ MinAvailableCycle = Height;
+
+ if (isReady(PredSU)) {
AvailableQueue->push(PredSU);
}
+ // CapturePred and others may have left the node in the pending queue, avoid
+ // adding it twice.
+ else if (!PredSU->isPending) {
+ PredSU->isPending = true;
+ PendingQueue.push_back(PredSU);
+ }
}
}
-/// ScheduleNodeBottomUp - Add the node to the schedule. Decrement the pending
-/// count of its predecessors. If a predecessor pending count is zero, add it to
-/// the Available queue.
-void ScheduleDAGRRList::ScheduleNodeBottomUp(SUnit *SU, unsigned CurCycle) {
- DOUT << "*** Scheduling [" << CurCycle << "]: ";
- DEBUG(SU->dump(&DAG));
- SU->Cycle = CurCycle;
-
- AvailableQueue->ScheduledNode(SU);
-
+/// Call ReleasePred for each predecessor, then update register live def/gen.
+/// Always update LiveRegDefs for a register dependence even if the current SU
+/// also defines the register. This effectively create one large live range
+/// across a sequence of two-address node. This is important because the
+/// entire chain must be scheduled together. Example:
+///
+/// flags = (3) add
+/// flags = (2) addc flags
+/// flags = (1) addc flags
+///
+/// results in
+///
+/// LiveRegDefs[flags] = 3
+/// LiveRegGens[flags] = 1
+///
+/// If (2) addc is unscheduled, then (1) addc must also be unscheduled to avoid
+/// interference on flags.
+void ScheduleDAGRRList::ReleasePredecessors(SUnit *SU) {
// Bottom up: release predecessors
for (SUnit::pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
I != E; ++I) {
- ReleasePred(I->Dep, I->isCtrl, CurCycle);
- if (I->Cost < 0) {
+ ReleasePred(SU, &*I);
+ if (I->isAssignedRegDep()) {
// This is a physical register dependency and it's impossible or
- // expensive to copy the register. Make sure nothing that can
+ // expensive to copy the register. Make sure nothing that can
// clobber the register is scheduled between the predecessor and
// this node.
- if (LiveRegs.insert(I->Reg)) {
- LiveRegDefs[I->Reg] = I->Dep;
- LiveRegCycles[I->Reg] = CurCycle;
+ SUnit *RegDef = LiveRegDefs[I->getReg()]; (void)RegDef;
+ assert((!RegDef || RegDef == SU || RegDef == I->getSUnit()) &&
+ "interference on register dependence");
+ LiveRegDefs[I->getReg()] = I->getSUnit();
+ if (!LiveRegGens[I->getReg()]) {
+ ++NumLiveRegs;
+ LiveRegGens[I->getReg()] = SU;
}
}
}
+}
+
+/// Check to see if any of the pending instructions are ready to issue. If
+/// so, add them to the available queue.
+void ScheduleDAGRRList::ReleasePending() {
+ if (DisableSchedCycles) {
+ assert(PendingQueue.empty() && "pending instrs not allowed in this mode");
+ return;
+ }
+
+ // If the available queue is empty, it is safe to reset MinAvailableCycle.
+ if (AvailableQueue->empty())
+ MinAvailableCycle = UINT_MAX;
+
+ // Check to see if any of the pending instructions are ready to issue. If
+ // so, add them to the available queue.
+ for (unsigned i = 0, e = PendingQueue.size(); i != e; ++i) {
+ unsigned ReadyCycle =
+ isBottomUp ? PendingQueue[i]->getHeight() : PendingQueue[i]->getDepth();
+ if (ReadyCycle < MinAvailableCycle)
+ MinAvailableCycle = ReadyCycle;
+
+ if (PendingQueue[i]->isAvailable) {
+ if (!isReady(PendingQueue[i]))
+ continue;
+ AvailableQueue->push(PendingQueue[i]);
+ }
+ PendingQueue[i]->isPending = false;
+ PendingQueue[i] = PendingQueue.back();
+ PendingQueue.pop_back();
+ --i; --e;
+ }
+}
+
+/// Move the scheduler state forward by the specified number of Cycles.
+void ScheduleDAGRRList::AdvanceToCycle(unsigned NextCycle) {
+ if (NextCycle <= CurCycle)
+ return;
+
+ IssueCount = 0;
+ AvailableQueue->setCurCycle(NextCycle);
+ if (!HazardRec->isEnabled()) {
+ // Bypass lots of virtual calls in case of long latency.
+ CurCycle = NextCycle;
+ }
+ else {
+ for (; CurCycle != NextCycle; ++CurCycle) {
+ if (isBottomUp)
+ HazardRec->RecedeCycle();
+ else
+ HazardRec->AdvanceCycle();
+ }
+ }
+ // FIXME: Instead of visiting the pending Q each time, set a dirty flag on the
+ // available Q to release pending nodes at least once before popping.
+ ReleasePending();
+}
+
+/// Move the scheduler state forward until the specified node's dependents are
+/// ready and can be scheduled with no resource conflicts.
+void ScheduleDAGRRList::AdvancePastStalls(SUnit *SU) {
+ if (DisableSchedCycles)
+ return;
+
+ unsigned ReadyCycle = isBottomUp ? SU->getHeight() : SU->getDepth();
+
+ // Bump CurCycle to account for latency. We assume the latency of other
+ // available instructions may be hidden by the stall (not a full pipe stall).
+ // This updates the hazard recognizer's cycle before reserving resources for
+ // this instruction.
+ AdvanceToCycle(ReadyCycle);
+
+ // Calls are scheduled in their preceding cycle, so don't conflict with
+ // hazards from instructions after the call. EmitNode will reset the
+ // scoreboard state before emitting the call.
+ if (isBottomUp && SU->isCall)
+ return;
+
+ // FIXME: For resource conflicts in very long non-pipelined stages, we
+ // should probably skip ahead here to avoid useless scoreboard checks.
+ int Stalls = 0;
+ while (true) {
+ ScheduleHazardRecognizer::HazardType HT =
+ HazardRec->getHazardType(SU, isBottomUp ? -Stalls : Stalls);
+
+ if (HT == ScheduleHazardRecognizer::NoHazard)
+ break;
+
+ ++Stalls;
+ }
+ AdvanceToCycle(CurCycle + Stalls);
+}
+
+/// Record this SUnit in the HazardRecognizer.
+/// Does not update CurCycle.
+void ScheduleDAGRRList::EmitNode(SUnit *SU) {
+ if (!HazardRec->isEnabled())
+ return;
+
+ // Check for phys reg copy.
+ if (!SU->getNode())
+ return;
+
+ switch (SU->getNode()->getOpcode()) {
+ default:
+ assert(SU->getNode()->isMachineOpcode() &&
+ "This target-independent node should not be scheduled.");
+ break;
+ case ISD::MERGE_VALUES:
+ case ISD::TokenFactor:
+ case ISD::CopyToReg:
+ case ISD::CopyFromReg:
+ case ISD::EH_LABEL:
+ // Noops don't affect the scoreboard state. Copies are likely to be
+ // removed.
+ return;
+ case ISD::INLINEASM:
+ // For inline asm, clear the pipeline state.
+ HazardRec->Reset();
+ return;
+ }
+ if (isBottomUp && SU->isCall) {
+ // Calls are scheduled with their preceding instructions. For bottom-up
+ // scheduling, clear the pipeline state before emitting.
+ HazardRec->Reset();
+ }
+
+ HazardRec->EmitInstruction(SU);
+
+ if (!isBottomUp && SU->isCall) {
+ HazardRec->Reset();
+ }
+}
+
+/// ScheduleNodeBottomUp - Add the node to the schedule. Decrement the pending
+/// count of its predecessors. If a predecessor pending count is zero, add it to
+/// the Available queue.
+void ScheduleDAGRRList::ScheduleNodeBottomUp(SUnit *SU) {
+ DEBUG(dbgs() << "\n*** Scheduling [" << CurCycle << "]: ");
+ DEBUG(SU->dump(this));
+
+#ifndef NDEBUG
+ if (CurCycle < SU->getHeight())
+ DEBUG(dbgs() << " Height [" << SU->getHeight() << "] pipeline stall!\n");
+#endif
+
+ // FIXME: Do not modify node height. It may interfere with
+ // backtracking. Instead add a "ready cycle" to SUnit. Before scheduling the
+ // node it's ready cycle can aid heuristics, and after scheduling it can
+ // indicate the scheduled cycle.
+ SU->setHeightToAtLeast(CurCycle);
+
+ // Reserve resources for the scheduled intruction.
+ EmitNode(SU);
+
+ Sequence.push_back(SU);
+
+ AvailableQueue->ScheduledNode(SU);
+
+ // If HazardRec is disabled, and each inst counts as one cycle, then
+ // advance CurCycle before ReleasePredecessors to avoid useles pushed to
+ // PendingQueue for schedulers that implement HasReadyFilter.
+ if (!HazardRec->isEnabled() && AvgIPC < 2)
+ AdvanceToCycle(CurCycle + 1);
+
+ // Update liveness of predecessors before successors to avoid treating a
+ // two-address node as a live range def.
+ ReleasePredecessors(SU);
// Release all the implicit physical register defs that are live.
for (SUnit::succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
I != E; ++I) {
- if (I->Cost < 0) {
- if (LiveRegCycles[I->Reg] == I->Dep->Cycle) {
- LiveRegs.erase(I->Reg);
- assert(LiveRegDefs[I->Reg] == SU &&
- "Physical register dependency violated?");
- LiveRegDefs[I->Reg] = NULL;
- LiveRegCycles[I->Reg] = 0;
- }
+ // LiveRegDegs[I->getReg()] != SU when SU is a two-address node.
+ if (I->isAssignedRegDep() && LiveRegDefs[I->getReg()] == SU) {
+ assert(NumLiveRegs > 0 && "NumLiveRegs is already zero!");
+ --NumLiveRegs;
+ LiveRegDefs[I->getReg()] = NULL;
+ LiveRegGens[I->getReg()] = NULL;
}
}
SU->isScheduled = true;
+
+ // Conditions under which the scheduler should eagerly advance the cycle:
+ // (1) No available instructions
+ // (2) All pipelines full, so available instructions must have hazards.
+ //
+ // If HazardRec is disabled, the cycle was advanced earlier.
+ //
+ // Check AvailableQueue after ReleasePredecessors in case of zero latency.
+ ++IssueCount;
+ if ((HazardRec->isEnabled() && HazardRec->atIssueLimit())
+ || (!HazardRec->isEnabled() && AvgIPC > 1 && IssueCount == AvgIPC)
+ || AvailableQueue->empty())
+ AdvanceToCycle(CurCycle + 1);
}
/// CapturePred - This does the opposite of ReleasePred. Since SU is being
/// unscheduled, incrcease the succ left count of its predecessors. Remove
/// them from AvailableQueue if necessary.
-void ScheduleDAGRRList::CapturePred(SUnit *PredSU, SUnit *SU, bool isChain) {
- PredSU->CycleBound = 0;
- for (SUnit::succ_iterator I = PredSU->Succs.begin(), E = PredSU->Succs.end();
- I != E; ++I) {
- if (I->Dep == SU)
- continue;
- PredSU->CycleBound = std::max(PredSU->CycleBound,
- I->Dep->Cycle + PredSU->Latency);
- }
-
+void ScheduleDAGRRList::CapturePred(SDep *PredEdge) {
+ SUnit *PredSU = PredEdge->getSUnit();
if (PredSU->isAvailable) {
PredSU->isAvailable = false;
if (!PredSU->isPending)
AvailableQueue->remove(PredSU);
}
+ assert(PredSU->NumSuccsLeft < UINT_MAX && "NumSuccsLeft will overflow!");
++PredSU->NumSuccsLeft;
}
/// UnscheduleNodeBottomUp - Remove the node from the schedule, update its and
/// its predecessor states to reflect the change.
void ScheduleDAGRRList::UnscheduleNodeBottomUp(SUnit *SU) {
- DOUT << "*** Unscheduling [" << SU->Cycle << "]: ";
- DEBUG(SU->dump(&DAG));
-
- AvailableQueue->UnscheduledNode(SU);
+ DEBUG(dbgs() << "*** Unscheduling [" << SU->getHeight() << "]: ");
+ DEBUG(SU->dump(this));
for (SUnit::pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
I != E; ++I) {
- CapturePred(I->Dep, SU, I->isCtrl);
- if (I->Cost < 0 && SU->Cycle == LiveRegCycles[I->Reg]) {
- LiveRegs.erase(I->Reg);
- assert(LiveRegDefs[I->Reg] == I->Dep &&
+ CapturePred(&*I);
+ if (I->isAssignedRegDep() && SU == LiveRegGens[I->getReg()]){
+ assert(NumLiveRegs > 0 && "NumLiveRegs is already zero!");
+ assert(LiveRegDefs[I->getReg()] == I->getSUnit() &&
"Physical register dependency violated?");
- LiveRegDefs[I->Reg] = NULL;
- LiveRegCycles[I->Reg] = 0;
+ --NumLiveRegs;
+ LiveRegDefs[I->getReg()] = NULL;
+ LiveRegGens[I->getReg()] = NULL;
}
}
for (SUnit::succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
I != E; ++I) {
- if (I->Cost < 0) {
- if (LiveRegs.insert(I->Reg)) {
- assert(!LiveRegDefs[I->Reg] &&
- "Physical register dependency violated?");
- LiveRegDefs[I->Reg] = SU;
+ if (I->isAssignedRegDep()) {
+ // This becomes the nearest def. Note that an earlier def may still be
+ // pending if this is a two-address node.
+ LiveRegDefs[I->getReg()] = SU;
+ if (!LiveRegDefs[I->getReg()]) {
+ ++NumLiveRegs;
}
- if (I->Dep->Cycle < LiveRegCycles[I->Reg])
- LiveRegCycles[I->Reg] = I->Dep->Cycle;
+ if (LiveRegGens[I->getReg()] == NULL ||
+ I->getSUnit()->getHeight() < LiveRegGens[I->getReg()]->getHeight())
+ LiveRegGens[I->getReg()] = I->getSUnit();
}
}
+ if (SU->getHeight() < MinAvailableCycle)
+ MinAvailableCycle = SU->getHeight();
- SU->Cycle = 0;
+ SU->setHeightDirty();
SU->isScheduled = false;
SU->isAvailable = true;
- AvailableQueue->push(SU);
-}
-
-// FIXME: This is probably too slow!
-static void isReachable(SUnit *SU, SUnit *TargetSU,
- SmallPtrSet<SUnit*, 32> &Visited, bool &Reached) {
- if (Reached) return;
- if (SU == TargetSU) {
- Reached = true;
- return;
+ if (!DisableSchedCycles && AvailableQueue->hasReadyFilter()) {
+ // Don't make available until backtracking is complete.
+ SU->isPending = true;
+ PendingQueue.push_back(SU);
}
- if (!Visited.insert(SU)) return;
-
- for (SUnit::pred_iterator I = SU->Preds.begin(), E = SU->Preds.end(); I != E;
- ++I)
- isReachable(I->Dep, TargetSU, Visited, Reached);
+ else {
+ AvailableQueue->push(SU);
+ }
+ AvailableQueue->UnscheduledNode(SU);
}
-static bool isReachable(SUnit *SU, SUnit *TargetSU) {
- SmallPtrSet<SUnit*, 32> Visited;
- bool Reached = false;
- isReachable(SU, TargetSU, Visited, Reached);
- return Reached;
-}
+/// After backtracking, the hazard checker needs to be restored to a state
+/// corresponding the the current cycle.
+void ScheduleDAGRRList::RestoreHazardCheckerBottomUp() {
+ HazardRec->Reset();
-/// willCreateCycle - Returns true if adding an edge from SU to TargetSU will
-/// create a cycle.
-static bool WillCreateCycle(SUnit *SU, SUnit *TargetSU) {
- if (isReachable(TargetSU, SU))
- return true;
- for (SUnit::pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
- I != E; ++I)
- if (I->Cost < 0 && isReachable(TargetSU, I->Dep))
- return true;
- return false;
+ unsigned LookAhead = std::min((unsigned)Sequence.size(),
+ HazardRec->getMaxLookAhead());
+ if (LookAhead == 0)
+ return;
+
+ std::vector<SUnit*>::const_iterator I = (Sequence.end() - LookAhead);
+ unsigned HazardCycle = (*I)->getHeight();
+ for (std::vector<SUnit*>::const_iterator E = Sequence.end(); I != E; ++I) {
+ SUnit *SU = *I;
+ for (; SU->getHeight() > HazardCycle; ++HazardCycle) {
+ HazardRec->RecedeCycle();
+ }
+ EmitNode(SU);
+ }
}
/// BacktrackBottomUp - Backtrack scheduling to a previous cycle specified in
-/// BTCycle in order to schedule a specific node. Returns the last unscheduled
-/// SUnit. Also returns if a successor is unscheduled in the process.
-void ScheduleDAGRRList::BacktrackBottomUp(SUnit *SU, unsigned BtCycle,
- unsigned &CurCycle) {
- SUnit *OldSU = NULL;
- while (CurCycle > BtCycle) {
- OldSU = Sequence.back();
+/// BTCycle in order to schedule a specific node.
+void ScheduleDAGRRList::BacktrackBottomUp(SUnit *SU, SUnit *BtSU) {
+ SUnit *OldSU = Sequence.back();
+ while (true) {
Sequence.pop_back();
if (SU->isSucc(OldSU))
// Don't try to remove SU from AvailableQueue.
SU->isAvailable = false;
+ // FIXME: use ready cycle instead of height
+ CurCycle = OldSU->getHeight();
UnscheduleNodeBottomUp(OldSU);
- --CurCycle;
+ AvailableQueue->setCurCycle(CurCycle);
+ if (OldSU == BtSU)
+ break;
+ OldSU = Sequence.back();
}
-
- if (SU->isSucc(OldSU)) {
- assert(false && "Something is wrong!");
- abort();
- }
+ assert(!SU->isSucc(OldSU) && "Something is wrong!");
+
+ RestoreHazardCheckerBottomUp();
+
+ ReleasePending();
++NumBacktracks;
}
+static bool isOperandOf(const SUnit *SU, SDNode *N) {
+ for (const SDNode *SUNode = SU->getNode(); SUNode;
+ SUNode = SUNode->getGluedNode()) {
+ if (SUNode->isOperandOf(N))
+ return true;
+ }
+ return false;
+}
+
/// CopyAndMoveSuccessors - Clone the specified node and move its scheduled
/// successors to the newly created node.
SUnit *ScheduleDAGRRList::CopyAndMoveSuccessors(SUnit *SU) {
- if (SU->FlaggedNodes.size())
+ SDNode *N = SU->getNode();
+ if (!N)
return NULL;
- SDNode *N = SU->Node;
- if (!N)
+ if (SU->getNode()->getGluedNode())
return NULL;
SUnit *NewSU;
bool TryUnfold = false;
for (unsigned i = 0, e = N->getNumValues(); i != e; ++i) {
- MVT::ValueType VT = N->getValueType(i);
- if (VT == MVT::Flag)
+ EVT VT = N->getValueType(i);
+ if (VT == MVT::Glue)
return NULL;
else if (VT == MVT::Other)
TryUnfold = true;
}
for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
- const SDOperand &Op = N->getOperand(i);
- MVT::ValueType VT = Op.Val->getValueType(Op.ResNo);
- if (VT == MVT::Flag)
+ const SDValue &Op = N->getOperand(i);
+ EVT VT = Op.getNode()->getValueType(Op.getResNo());
+ if (VT == MVT::Glue)
return NULL;
}
if (TryUnfold) {
- SmallVector<SDNode*, 4> NewNodes;
- if (!TII->unfoldMemoryOperand(DAG, N, NewNodes))
+ SmallVector<SDNode*, 2> NewNodes;
+ if (!TII->unfoldMemoryOperand(*DAG, N, NewNodes))
return NULL;
- DOUT << "Unfolding SU # " << SU->NodeNum << "\n";
+ DEBUG(dbgs() << "Unfolding SU #" << SU->NodeNum << "\n");
assert(NewNodes.size() == 2 && "Expected a load folding node!");
N = NewNodes[1];
SDNode *LoadNode = NewNodes[0];
unsigned NumVals = N->getNumValues();
- unsigned OldNumVals = SU->Node->getNumValues();
+ unsigned OldNumVals = SU->getNode()->getNumValues();
for (unsigned i = 0; i != NumVals; ++i)
- DAG.ReplaceAllUsesOfValueWith(SDOperand(SU->Node, i), SDOperand(N, i));
- DAG.ReplaceAllUsesOfValueWith(SDOperand(SU->Node, OldNumVals-1),
- SDOperand(LoadNode, 1));
-
- SUnit *NewSU = NewSUnit(N);
- SUnitMap[N].push_back(NewSU);
- const TargetInstrDescriptor *TID = &TII->get(N->getTargetOpcode());
- for (unsigned i = 0; i != TID->numOperands; ++i) {
- if (TID->getOperandConstraint(i, TOI::TIED_TO) != -1) {
- NewSU->isTwoAddress = true;
- break;
- }
- }
- if (TID->Flags & M_COMMUTABLE)
- NewSU->isCommutable = true;
- // FIXME: Calculate height / depth and propagate the changes?
- NewSU->Depth = SU->Depth;
- NewSU->Height = SU->Height;
- ComputeLatency(NewSU);
+ DAG->ReplaceAllUsesOfValueWith(SDValue(SU->getNode(), i), SDValue(N, i));
+ DAG->ReplaceAllUsesOfValueWith(SDValue(SU->getNode(), OldNumVals-1),
+ SDValue(LoadNode, 1));
// LoadNode may already exist. This can happen when there is another
// load from the same location and producing the same type of value
// but it has different alignment or volatileness.
bool isNewLoad = true;
SUnit *LoadSU;
- DenseMap<SDNode*, std::vector<SUnit*> >::iterator SMI =
- SUnitMap.find(LoadNode);
- if (SMI != SUnitMap.end()) {
- LoadSU = SMI->second.front();
+ if (LoadNode->getNodeId() != -1) {
+ LoadSU = &SUnits[LoadNode->getNodeId()];
isNewLoad = false;
} else {
- LoadSU = NewSUnit(LoadNode);
- SUnitMap[LoadNode].push_back(LoadSU);
+ LoadSU = CreateNewSUnit(LoadNode);
+ LoadNode->setNodeId(LoadSU->NodeNum);
- LoadSU->Depth = SU->Depth;
- LoadSU->Height = SU->Height;
+ InitNumRegDefsLeft(LoadSU);
ComputeLatency(LoadSU);
}
- SUnit *ChainPred = NULL;
+ SUnit *NewSU = CreateNewSUnit(N);
+ assert(N->getNodeId() == -1 && "Node already inserted!");
+ N->setNodeId(NewSU->NodeNum);
+
+ const TargetInstrDesc &TID = TII->get(N->getMachineOpcode());
+ for (unsigned i = 0; i != TID.getNumOperands(); ++i) {
+ if (TID.getOperandConstraint(i, TOI::TIED_TO) != -1) {
+ NewSU->isTwoAddress = true;
+ break;
+ }
+ }
+ if (TID.isCommutable())
+ NewSU->isCommutable = true;
+
+ InitNumRegDefsLeft(NewSU);
+ ComputeLatency(NewSU);
+
+ // Record all the edges to and from the old SU, by category.
+ SmallVector<SDep, 4> ChainPreds;
SmallVector<SDep, 4> ChainSuccs;
SmallVector<SDep, 4> LoadPreds;
SmallVector<SDep, 4> NodePreds;
SmallVector<SDep, 4> NodeSuccs;
for (SUnit::pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
I != E; ++I) {
- if (I->isCtrl)
- ChainPred = I->Dep;
- else if (I->Dep->Node && I->Dep->Node->isOperand(LoadNode))
- LoadPreds.push_back(SDep(I->Dep, I->Reg, I->Cost, false, false));
+ if (I->isCtrl())
+ ChainPreds.push_back(*I);
+ else if (isOperandOf(I->getSUnit(), LoadNode))
+ LoadPreds.push_back(*I);
else
- NodePreds.push_back(SDep(I->Dep, I->Reg, I->Cost, false, false));
+ NodePreds.push_back(*I);
}
for (SUnit::succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
I != E; ++I) {
- if (I->isCtrl)
- ChainSuccs.push_back(SDep(I->Dep, I->Reg, I->Cost,
- I->isCtrl, I->isSpecial));
+ if (I->isCtrl())
+ ChainSuccs.push_back(*I);
else
- NodeSuccs.push_back(SDep(I->Dep, I->Reg, I->Cost,
- I->isCtrl, I->isSpecial));
+ NodeSuccs.push_back(*I);
}
- SU->removePred(ChainPred, true, false);
- if (isNewLoad)
- LoadSU->addPred(ChainPred, true, false);
+ // Now assign edges to the newly-created nodes.
+ for (unsigned i = 0, e = ChainPreds.size(); i != e; ++i) {
+ const SDep &Pred = ChainPreds[i];
+ RemovePred(SU, Pred);
+ if (isNewLoad)
+ AddPred(LoadSU, Pred);
+ }
for (unsigned i = 0, e = LoadPreds.size(); i != e; ++i) {
- SDep *Pred = &LoadPreds[i];
- SU->removePred(Pred->Dep, Pred->isCtrl, Pred->isSpecial);
+ const SDep &Pred = LoadPreds[i];
+ RemovePred(SU, Pred);
if (isNewLoad)
- LoadSU->addPred(Pred->Dep, Pred->isCtrl, Pred->isSpecial,
- Pred->Reg, Pred->Cost);
+ AddPred(LoadSU, Pred);
}
for (unsigned i = 0, e = NodePreds.size(); i != e; ++i) {
- SDep *Pred = &NodePreds[i];
- SU->removePred(Pred->Dep, Pred->isCtrl, Pred->isSpecial);
- NewSU->addPred(Pred->Dep, Pred->isCtrl, Pred->isSpecial,
- Pred->Reg, Pred->Cost);
+ const SDep &Pred = NodePreds[i];
+ RemovePred(SU, Pred);
+ AddPred(NewSU, Pred);
}
for (unsigned i = 0, e = NodeSuccs.size(); i != e; ++i) {
- SDep *Succ = &NodeSuccs[i];
- Succ->Dep->removePred(SU, Succ->isCtrl, Succ->isSpecial);
- Succ->Dep->addPred(NewSU, Succ->isCtrl, Succ->isSpecial,
- Succ->Reg, Succ->Cost);
+ SDep D = NodeSuccs[i];
+ SUnit *SuccDep = D.getSUnit();
+ D.setSUnit(SU);
+ RemovePred(SuccDep, D);
+ D.setSUnit(NewSU);
+ AddPred(SuccDep, D);
+ // Balance register pressure.
+ if (AvailableQueue->tracksRegPressure() && SuccDep->isScheduled
+ && !D.isCtrl() && NewSU->NumRegDefsLeft > 0)
+ --NewSU->NumRegDefsLeft;
}
for (unsigned i = 0, e = ChainSuccs.size(); i != e; ++i) {
- SDep *Succ = &ChainSuccs[i];
- Succ->Dep->removePred(SU, Succ->isCtrl, Succ->isSpecial);
- if (isNewLoad)
- Succ->Dep->addPred(LoadSU, Succ->isCtrl, Succ->isSpecial,
- Succ->Reg, Succ->Cost);
- }
- if (isNewLoad)
- NewSU->addPred(LoadSU, false, false);
+ SDep D = ChainSuccs[i];
+ SUnit *SuccDep = D.getSUnit();
+ D.setSUnit(SU);
+ RemovePred(SuccDep, D);
+ if (isNewLoad) {
+ D.setSUnit(LoadSU);
+ AddPred(SuccDep, D);
+ }
+ }
+
+ // Add a data dependency to reflect that NewSU reads the value defined
+ // by LoadSU.
+ AddPred(NewSU, SDep(LoadSU, SDep::Data, LoadSU->Latency));
if (isNewLoad)
AvailableQueue->addNode(LoadSU);
SU = NewSU;
}
- DOUT << "Duplicating SU # " << SU->NodeNum << "\n";
- NewSU = Clone(SU);
+ DEBUG(dbgs() << " Duplicating SU #" << SU->NodeNum << "\n");
+ NewSU = CreateClone(SU);
// New SUnit has the exact same predecessors.
for (SUnit::pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
I != E; ++I)
- if (!I->isSpecial) {
- NewSU->addPred(I->Dep, I->isCtrl, false, I->Reg, I->Cost);
- NewSU->Depth = std::max(NewSU->Depth, I->Dep->Depth+1);
- }
+ if (!I->isArtificial())
+ AddPred(NewSU, *I);
// Only copy scheduled successors. Cut them from old node's successor
// list and move them over.
- SmallVector<std::pair<SUnit*, bool>, 4> DelDeps;
+ SmallVector<std::pair<SUnit *, SDep>, 4> DelDeps;
for (SUnit::succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
I != E; ++I) {
- if (I->isSpecial)
+ if (I->isArtificial())
continue;
- if (I->Dep->isScheduled) {
- NewSU->Height = std::max(NewSU->Height, I->Dep->Height+1);
- I->Dep->addPred(NewSU, I->isCtrl, false, I->Reg, I->Cost);
- DelDeps.push_back(std::make_pair(I->Dep, I->isCtrl));
+ SUnit *SuccSU = I->getSUnit();
+ if (SuccSU->isScheduled) {
+ SDep D = *I;
+ D.setSUnit(NewSU);
+ AddPred(SuccSU, D);
+ D.setSUnit(SU);
+ DelDeps.push_back(std::make_pair(SuccSU, D));
}
}
- for (unsigned i = 0, e = DelDeps.size(); i != e; ++i) {
- SUnit *Succ = DelDeps[i].first;
- bool isCtrl = DelDeps[i].second;
- Succ->removePred(SU, isCtrl, false);
- }
+ for (unsigned i = 0, e = DelDeps.size(); i != e; ++i)
+ RemovePred(DelDeps[i].first, DelDeps[i].second);
AvailableQueue->updateNode(SU);
AvailableQueue->addNode(NewSU);
return NewSU;
}
-/// InsertCCCopiesAndMoveSuccs - Insert expensive cross register class copies
-/// and move all scheduled successors of the given SUnit to the last copy.
-void ScheduleDAGRRList::InsertCCCopiesAndMoveSuccs(SUnit *SU, unsigned Reg,
- const TargetRegisterClass *DestRC,
- const TargetRegisterClass *SrcRC,
+/// InsertCopiesAndMoveSuccs - Insert register copies and move all
+/// scheduled successors of the given SUnit to the last copy.
+void ScheduleDAGRRList::InsertCopiesAndMoveSuccs(SUnit *SU, unsigned Reg,
+ const TargetRegisterClass *DestRC,
+ const TargetRegisterClass *SrcRC,
SmallVector<SUnit*, 2> &Copies) {
- SUnit *CopyFromSU = NewSUnit(NULL);
+ SUnit *CopyFromSU = CreateNewSUnit(NULL);
CopyFromSU->CopySrcRC = SrcRC;
CopyFromSU->CopyDstRC = DestRC;
- CopyFromSU->Depth = SU->Depth;
- CopyFromSU->Height = SU->Height;
- SUnit *CopyToSU = NewSUnit(NULL);
+ SUnit *CopyToSU = CreateNewSUnit(NULL);
CopyToSU->CopySrcRC = DestRC;
CopyToSU->CopyDstRC = SrcRC;
// Only copy scheduled successors. Cut them from old node's successor
// list and move them over.
- SmallVector<std::pair<SUnit*, bool>, 4> DelDeps;
+ SmallVector<std::pair<SUnit *, SDep>, 4> DelDeps;
for (SUnit::succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
I != E; ++I) {
- if (I->isSpecial)
+ if (I->isArtificial())
continue;
- if (I->Dep->isScheduled) {
- CopyToSU->Height = std::max(CopyToSU->Height, I->Dep->Height+1);
- I->Dep->addPred(CopyToSU, I->isCtrl, false, I->Reg, I->Cost);
- DelDeps.push_back(std::make_pair(I->Dep, I->isCtrl));
+ SUnit *SuccSU = I->getSUnit();
+ if (SuccSU->isScheduled) {
+ SDep D = *I;
+ D.setSUnit(CopyToSU);
+ AddPred(SuccSU, D);
+ DelDeps.push_back(std::make_pair(SuccSU, *I));
}
}
- for (unsigned i = 0, e = DelDeps.size(); i != e; ++i) {
- SUnit *Succ = DelDeps[i].first;
- bool isCtrl = DelDeps[i].second;
- Succ->removePred(SU, isCtrl, false);
- }
+ for (unsigned i = 0, e = DelDeps.size(); i != e; ++i)
+ RemovePred(DelDeps[i].first, DelDeps[i].second);
- CopyFromSU->addPred(SU, false, false, Reg, -1);
- CopyToSU->addPred(CopyFromSU, false, false, Reg, 1);
+ AddPred(CopyFromSU, SDep(SU, SDep::Data, SU->Latency, Reg));
+ AddPred(CopyToSU, SDep(CopyFromSU, SDep::Data, CopyFromSU->Latency, 0));
AvailableQueue->updateNode(SU);
AvailableQueue->addNode(CopyFromSU);
Copies.push_back(CopyFromSU);
Copies.push_back(CopyToSU);
- ++NumCCCopies;
+ ++NumPRCopies;
}
/// getPhysicalRegisterVT - Returns the ValueType of the physical register
/// definition of the specified node.
/// FIXME: Move to SelectionDAG?
-static MVT::ValueType getPhysicalRegisterVT(SDNode *N, unsigned Reg,
- const TargetInstrInfo *TII) {
- const TargetInstrDescriptor &TID = TII->get(N->getTargetOpcode());
+static EVT getPhysicalRegisterVT(SDNode *N, unsigned Reg,
+ const TargetInstrInfo *TII) {
+ const TargetInstrDesc &TID = TII->get(N->getMachineOpcode());
assert(TID.ImplicitDefs && "Physical reg def must be in implicit def list!");
- unsigned NumRes = TID.numDefs;
- for (const unsigned *ImpDef = TID.ImplicitDefs; *ImpDef; ++ImpDef) {
+ unsigned NumRes = TID.getNumDefs();
+ for (const unsigned *ImpDef = TID.getImplicitDefs(); *ImpDef; ++ImpDef) {
if (Reg == *ImpDef)
break;
++NumRes;
return N->getValueType(NumRes);
}
+/// CheckForLiveRegDef - Return true and update live register vector if the
+/// specified register def of the specified SUnit clobbers any "live" registers.
+static void CheckForLiveRegDef(SUnit *SU, unsigned Reg,
+ std::vector<SUnit*> &LiveRegDefs,
+ SmallSet<unsigned, 4> &RegAdded,
+ SmallVector<unsigned, 4> &LRegs,
+ const TargetRegisterInfo *TRI) {
+ for (const unsigned *AliasI = TRI->getOverlaps(Reg); *AliasI; ++AliasI) {
+
+ // Check if Ref is live.
+ if (!LiveRegDefs[Reg]) continue;
+
+ // Allow multiple uses of the same def.
+ if (LiveRegDefs[Reg] == SU) continue;
+
+ // Add Reg to the set of interfering live regs.
+ if (RegAdded.insert(Reg))
+ LRegs.push_back(Reg);
+ }
+}
+
/// DelayForLiveRegsBottomUp - Returns true if it is necessary to delay
/// scheduling of the given node to satisfy live physical register dependencies.
/// If the specific node is the last one that's available to schedule, do
/// whatever is necessary (i.e. backtracking or cloning) to make it possible.
-bool ScheduleDAGRRList::DelayForLiveRegsBottomUp(SUnit *SU,
- SmallVector<unsigned, 4> &LRegs){
- if (LiveRegs.empty())
+bool ScheduleDAGRRList::
+DelayForLiveRegsBottomUp(SUnit *SU, SmallVector<unsigned, 4> &LRegs) {
+ if (NumLiveRegs == 0)
return false;
SmallSet<unsigned, 4> RegAdded;
// If this node would clobber any "live" register, then it's not ready.
+ //
+ // If SU is the currently live definition of the same register that it uses,
+ // then we are free to schedule it.
for (SUnit::pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
I != E; ++I) {
- if (I->Cost < 0) {
- unsigned Reg = I->Reg;
- if (LiveRegs.count(Reg) && LiveRegDefs[Reg] != I->Dep) {
- if (RegAdded.insert(Reg))
- LRegs.push_back(Reg);
+ if (I->isAssignedRegDep() && LiveRegDefs[I->getReg()] != SU)
+ CheckForLiveRegDef(I->getSUnit(), I->getReg(), LiveRegDefs,
+ RegAdded, LRegs, TRI);
+ }
+
+ for (SDNode *Node = SU->getNode(); Node; Node = Node->getGluedNode()) {
+ if (Node->getOpcode() == ISD::INLINEASM) {
+ // Inline asm can clobber physical defs.
+ unsigned NumOps = Node->getNumOperands();
+ if (Node->getOperand(NumOps-1).getValueType() == MVT::Glue)
+ --NumOps; // Ignore the glue operand.
+
+ for (unsigned i = InlineAsm::Op_FirstOperand; i != NumOps;) {
+ unsigned Flags =
+ cast<ConstantSDNode>(Node->getOperand(i))->getZExtValue();
+ unsigned NumVals = InlineAsm::getNumOperandRegisters(Flags);
+
+ ++i; // Skip the ID value.
+ if (InlineAsm::isRegDefKind(Flags) ||
+ InlineAsm::isRegDefEarlyClobberKind(Flags)) {
+ // Check for def of register or earlyclobber register.
+ for (; NumVals; --NumVals, ++i) {
+ unsigned Reg = cast<RegisterSDNode>(Node->getOperand(i))->getReg();
+ if (TargetRegisterInfo::isPhysicalRegister(Reg))
+ CheckForLiveRegDef(SU, Reg, LiveRegDefs, RegAdded, LRegs, TRI);
+ }
+ } else
+ i += NumVals;
}
- for (const unsigned *Alias = MRI->getAliasSet(Reg);
- *Alias; ++Alias)
- if (LiveRegs.count(*Alias) && LiveRegDefs[*Alias] != I->Dep) {
- if (RegAdded.insert(*Alias))
- LRegs.push_back(*Alias);
- }
+ continue;
}
- }
- for (unsigned i = 0, e = SU->FlaggedNodes.size()+1; i != e; ++i) {
- SDNode *Node = (i == 0) ? SU->Node : SU->FlaggedNodes[i-1];
- if (!Node || !Node->isTargetOpcode())
+ if (!Node->isMachineOpcode())
continue;
- const TargetInstrDescriptor &TID = TII->get(Node->getTargetOpcode());
+ const TargetInstrDesc &TID = TII->get(Node->getMachineOpcode());
if (!TID.ImplicitDefs)
continue;
- for (const unsigned *Reg = TID.ImplicitDefs; *Reg; ++Reg) {
- if (LiveRegs.count(*Reg) && LiveRegDefs[*Reg] != SU) {
- if (RegAdded.insert(*Reg))
- LRegs.push_back(*Reg);
- }
- for (const unsigned *Alias = MRI->getAliasSet(*Reg);
- *Alias; ++Alias)
- if (LiveRegs.count(*Alias) && LiveRegDefs[*Alias] != SU) {
- if (RegAdded.insert(*Alias))
- LRegs.push_back(*Alias);
- }
- }
+ for (const unsigned *Reg = TID.ImplicitDefs; *Reg; ++Reg)
+ CheckForLiveRegDef(SU, *Reg, LiveRegDefs, RegAdded, LRegs, TRI);
}
+
return !LRegs.empty();
}
+/// Return a node that can be scheduled in this cycle. Requirements:
+/// (1) Ready: latency has been satisfied
+/// (2) No Hazards: resources are available
+/// (3) No Interferences: may unschedule to break register interferences.
+SUnit *ScheduleDAGRRList::PickNodeToScheduleBottomUp() {
+ SmallVector<SUnit*, 4> Interferences;
+ DenseMap<SUnit*, SmallVector<unsigned, 4> > LRegsMap;
+
+ SUnit *CurSU = AvailableQueue->pop();
+ while (CurSU) {
+ SmallVector<unsigned, 4> LRegs;
+ if (!DelayForLiveRegsBottomUp(CurSU, LRegs))
+ break;
+ LRegsMap.insert(std::make_pair(CurSU, LRegs));
-/// ListScheduleBottomUp - The main loop of list scheduling for bottom-up
-/// schedulers.
-void ScheduleDAGRRList::ListScheduleBottomUp() {
- unsigned CurCycle = 0;
- // Add root to Available queue.
- SUnit *RootSU = SUnitMap[DAG.getRoot().Val].front();
- RootSU->isAvailable = true;
- AvailableQueue->push(RootSU);
+ CurSU->isPending = true; // This SU is not in AvailableQueue right now.
+ Interferences.push_back(CurSU);
+ CurSU = AvailableQueue->pop();
+ }
+ if (CurSU) {
+ // Add the nodes that aren't ready back onto the available list.
+ for (unsigned i = 0, e = Interferences.size(); i != e; ++i) {
+ Interferences[i]->isPending = false;
+ assert(Interferences[i]->isAvailable && "must still be available");
+ AvailableQueue->push(Interferences[i]);
+ }
+ return CurSU;
+ }
- // While Available queue is not empty, grab the node with the highest
- // priority. If it is not ready put it back. Schedule the node.
- SmallVector<SUnit*, 4> NotReady;
- while (!AvailableQueue->empty()) {
- bool Delayed = false;
- DenseMap<SUnit*, SmallVector<unsigned, 4> > LRegsMap;
- SUnit *CurSU = AvailableQueue->pop();
- while (CurSU) {
- if (CurSU->CycleBound <= CurCycle) {
- SmallVector<unsigned, 4> LRegs;
- if (!DelayForLiveRegsBottomUp(CurSU, LRegs))
- break;
- Delayed = true;
- LRegsMap.insert(std::make_pair(CurSU, LRegs));
+ // All candidates are delayed due to live physical reg dependencies.
+ // Try backtracking, code duplication, or inserting cross class copies
+ // to resolve it.
+ for (unsigned i = 0, e = Interferences.size(); i != e; ++i) {
+ SUnit *TrySU = Interferences[i];
+ SmallVector<unsigned, 4> &LRegs = LRegsMap[TrySU];
+
+ // Try unscheduling up to the point where it's safe to schedule
+ // this node.
+ SUnit *BtSU = NULL;
+ unsigned LiveCycle = UINT_MAX;
+ for (unsigned j = 0, ee = LRegs.size(); j != ee; ++j) {
+ unsigned Reg = LRegs[j];
+ if (LiveRegGens[Reg]->getHeight() < LiveCycle) {
+ BtSU = LiveRegGens[Reg];
+ LiveCycle = BtSU->getHeight();
}
-
- CurSU->isPending = true; // This SU is not in AvailableQueue right now.
- NotReady.push_back(CurSU);
- CurSU = AvailableQueue->pop();
}
-
- // All candidates are delayed due to live physical reg dependencies.
- // Try backtracking, code duplication, or inserting cross class copies
- // to resolve it.
- if (Delayed && !CurSU) {
- for (unsigned i = 0, e = NotReady.size(); i != e; ++i) {
- SUnit *TrySU = NotReady[i];
- SmallVector<unsigned, 4> &LRegs = LRegsMap[TrySU];
-
- // Try unscheduling up to the point where it's safe to schedule
- // this node.
- unsigned LiveCycle = CurCycle;
- for (unsigned j = 0, ee = LRegs.size(); j != ee; ++j) {
- unsigned Reg = LRegs[j];
- unsigned LCycle = LiveRegCycles[Reg];
- LiveCycle = std::min(LiveCycle, LCycle);
- }
- SUnit *OldSU = Sequence[LiveCycle];
- if (!WillCreateCycle(TrySU, OldSU)) {
- BacktrackBottomUp(TrySU, LiveCycle, CurCycle);
- // Force the current node to be scheduled before the node that
- // requires the physical reg dep.
- if (OldSU->isAvailable) {
- OldSU->isAvailable = false;
- AvailableQueue->remove(OldSU);
- }
- TrySU->addPred(OldSU, true, true);
- // If one or more successors has been unscheduled, then the current
- // node is no longer avaialable. Schedule a successor that's now
- // available instead.
- if (!TrySU->isAvailable)
- CurSU = AvailableQueue->pop();
- else {
- CurSU = TrySU;
- TrySU->isPending = false;
- NotReady.erase(NotReady.begin()+i);
- }
- break;
- }
+ if (!WillCreateCycle(TrySU, BtSU)) {
+ BacktrackBottomUp(TrySU, BtSU);
+
+ // Force the current node to be scheduled before the node that
+ // requires the physical reg dep.
+ if (BtSU->isAvailable) {
+ BtSU->isAvailable = false;
+ if (!BtSU->isPending)
+ AvailableQueue->remove(BtSU);
}
-
- if (!CurSU) {
- // Can't backtrace. Try duplicating the nodes that produces these
- // "expensive to copy" values to break the dependency. In case even
- // that doesn't work, insert cross class copies.
- SUnit *TrySU = NotReady[0];
- SmallVector<unsigned, 4> &LRegs = LRegsMap[TrySU];
- assert(LRegs.size() == 1 && "Can't handle this yet!");
- unsigned Reg = LRegs[0];
- SUnit *LRDef = LiveRegDefs[Reg];
- SUnit *NewDef = CopyAndMoveSuccessors(LRDef);
- if (!NewDef) {
- // Issue expensive cross register class copies.
- MVT::ValueType VT = getPhysicalRegisterVT(LRDef->Node, Reg, TII);
- const TargetRegisterClass *RC =
- MRI->getPhysicalRegisterRegClass(VT, Reg);
- const TargetRegisterClass *DestRC = MRI->getCrossCopyRegClass(RC);
- if (!DestRC) {
- assert(false && "Don't know how to copy this physical register!");
- abort();
- }
- SmallVector<SUnit*, 2> Copies;
- InsertCCCopiesAndMoveSuccs(LRDef, Reg, DestRC, RC, Copies);
- DOUT << "Adding an edge from SU # " << TrySU->NodeNum
- << " to SU #" << Copies.front()->NodeNum << "\n";
- TrySU->addPred(Copies.front(), true, true);
- NewDef = Copies.back();
- }
-
- DOUT << "Adding an edge from SU # " << NewDef->NodeNum
- << " to SU #" << TrySU->NodeNum << "\n";
- LiveRegDefs[Reg] = NewDef;
- NewDef->addPred(TrySU, true, true);
- TrySU->isAvailable = false;
- CurSU = NewDef;
+ AddPred(TrySU, SDep(BtSU, SDep::Order, /*Latency=*/1,
+ /*Reg=*/0, /*isNormalMemory=*/false,
+ /*isMustAlias=*/false, /*isArtificial=*/true));
+
+ // If one or more successors has been unscheduled, then the current
+ // node is no longer avaialable. Schedule a successor that's now
+ // available instead.
+ if (!TrySU->isAvailable) {
+ CurSU = AvailableQueue->pop();
}
-
- if (!CurSU) {
- assert(false && "Unable to resolve live physical register dependencies!");
- abort();
+ else {
+ CurSU = TrySU;
+ TrySU->isPending = false;
+ Interferences.erase(Interferences.begin()+i);
}
+ break;
}
+ }
- // Add the nodes that aren't ready back onto the available list.
- for (unsigned i = 0, e = NotReady.size(); i != e; ++i) {
- NotReady[i]->isPending = false;
- // May no longer be available due to backtracking.
- if (NotReady[i]->isAvailable)
- AvailableQueue->push(NotReady[i]);
+ if (!CurSU) {
+ // Can't backtrack. If it's too expensive to copy the value, then try
+ // duplicate the nodes that produces these "too expensive to copy"
+ // values to break the dependency. In case even that doesn't work,
+ // insert cross class copies.
+ // If it's not too expensive, i.e. cost != -1, issue copies.
+ SUnit *TrySU = Interferences[0];
+ SmallVector<unsigned, 4> &LRegs = LRegsMap[TrySU];
+ assert(LRegs.size() == 1 && "Can't handle this yet!");
+ unsigned Reg = LRegs[0];
+ SUnit *LRDef = LiveRegDefs[Reg];
+ EVT VT = getPhysicalRegisterVT(LRDef->getNode(), Reg, TII);
+ const TargetRegisterClass *RC =
+ TRI->getMinimalPhysRegClass(Reg, VT);
+ const TargetRegisterClass *DestRC = TRI->getCrossCopyRegClass(RC);
+
+ // If cross copy register class is null, then it must be possible copy
+ // the value directly. Do not try duplicate the def.
+ SUnit *NewDef = 0;
+ if (DestRC)
+ NewDef = CopyAndMoveSuccessors(LRDef);
+ else
+ DestRC = RC;
+ if (!NewDef) {
+ // Issue copies, these can be expensive cross register class copies.
+ SmallVector<SUnit*, 2> Copies;
+ InsertCopiesAndMoveSuccs(LRDef, Reg, DestRC, RC, Copies);
+ DEBUG(dbgs() << " Adding an edge from SU #" << TrySU->NodeNum
+ << " to SU #" << Copies.front()->NodeNum << "\n");
+ AddPred(TrySU, SDep(Copies.front(), SDep::Order, /*Latency=*/1,
+ /*Reg=*/0, /*isNormalMemory=*/false,
+ /*isMustAlias=*/false,
+ /*isArtificial=*/true));
+ NewDef = Copies.back();
}
- NotReady.clear();
- if (!CurSU)
- Sequence.push_back(0);
- else {
- ScheduleNodeBottomUp(CurSU, CurCycle);
- Sequence.push_back(CurSU);
- }
- ++CurCycle;
+ DEBUG(dbgs() << " Adding an edge from SU #" << NewDef->NodeNum
+ << " to SU #" << TrySU->NodeNum << "\n");
+ LiveRegDefs[Reg] = NewDef;
+ AddPred(NewDef, SDep(TrySU, SDep::Order, /*Latency=*/1,
+ /*Reg=*/0, /*isNormalMemory=*/false,
+ /*isMustAlias=*/false,
+ /*isArtificial=*/true));
+ TrySU->isAvailable = false;
+ CurSU = NewDef;
}
- // Add entry node last
- if (DAG.getEntryNode().Val != DAG.getRoot().Val) {
- SUnit *Entry = SUnitMap[DAG.getEntryNode().Val].front();
- Sequence.push_back(Entry);
+ assert(CurSU && "Unable to resolve live physical register dependencies!");
+
+ // Add the nodes that aren't ready back onto the available list.
+ for (unsigned i = 0, e = Interferences.size(); i != e; ++i) {
+ Interferences[i]->isPending = false;
+ // May no longer be available due to backtracking.
+ if (Interferences[i]->isAvailable) {
+ AvailableQueue->push(Interferences[i]);
+ }
}
+ return CurSU;
+}
- // Reverse the order if it is bottom up.
- std::reverse(Sequence.begin(), Sequence.end());
-
-
-#ifndef NDEBUG
- // Verify that all SUnits were scheduled.
- bool AnyNotSched = false;
- for (unsigned i = 0, e = SUnits.size(); i != e; ++i) {
- if (SUnits[i].NumSuccsLeft != 0) {
- if (!AnyNotSched)
- cerr << "*** List scheduling failed! ***\n";
- SUnits[i].dump(&DAG);
- cerr << "has not been scheduled!\n";
- AnyNotSched = true;
+/// ListScheduleBottomUp - The main loop of list scheduling for bottom-up
+/// schedulers.
+void ScheduleDAGRRList::ListScheduleBottomUp() {
+ // Release any predecessors of the special Exit node.
+ ReleasePredecessors(&ExitSU);
+
+ // Add root to Available queue.
+ if (!SUnits.empty()) {
+ SUnit *RootSU = &SUnits[DAG->getRoot().getNode()->getNodeId()];
+ assert(RootSU->Succs.empty() && "Graph root shouldn't have successors!");
+ RootSU->isAvailable = true;
+ AvailableQueue->push(RootSU);
+ }
+
+ // While Available queue is not empty, grab the node with the highest
+ // priority. If it is not ready put it back. Schedule the node.
+ Sequence.reserve(SUnits.size());
+ while (!AvailableQueue->empty()) {
+ DEBUG(dbgs() << "\n*** Examining Available\n";
+ AvailableQueue->dump(this));
+
+ // Pick the best node to schedule taking all constraints into
+ // consideration.
+ SUnit *SU = PickNodeToScheduleBottomUp();
+
+ AdvancePastStalls(SU);
+
+ ScheduleNodeBottomUp(SU);
+
+ while (AvailableQueue->empty() && !PendingQueue.empty()) {
+ // Advance the cycle to free resources. Skip ahead to the next ready SU.
+ assert(MinAvailableCycle < UINT_MAX && "MinAvailableCycle uninitialized");
+ AdvanceToCycle(std::max(CurCycle + 1, MinAvailableCycle));
}
}
- assert(!AnyNotSched);
+
+ // Reverse the order if it is bottom up.
+ std::reverse(Sequence.begin(), Sequence.end());
+
+#ifndef NDEBUG
+ VerifySchedule(isBottomUp);
#endif
}
/// ReleaseSucc - Decrement the NumPredsLeft count of a successor. Add it to
/// the AvailableQueue if the count reaches zero. Also update its cycle bound.
-void ScheduleDAGRRList::ReleaseSucc(SUnit *SuccSU, bool isChain,
- unsigned CurCycle) {
- // FIXME: the distance between two nodes is not always == the predecessor's
- // latency. For example, the reader can very well read the register written
- // by the predecessor later than the issue cycle. It also depends on the
- // interrupt model (drain vs. freeze).
- SuccSU->CycleBound = std::max(SuccSU->CycleBound, CurCycle + SuccSU->Latency);
+void ScheduleDAGRRList::ReleaseSucc(SUnit *SU, const SDep *SuccEdge) {
+ SUnit *SuccSU = SuccEdge->getSUnit();
- --SuccSU->NumPredsLeft;
-
#ifndef NDEBUG
- if (SuccSU->NumPredsLeft < 0) {
- cerr << "*** List scheduling failed! ***\n";
- SuccSU->dump(&DAG);
- cerr << " has been released too many times!\n";
- assert(0);
+ if (SuccSU->NumPredsLeft == 0) {
+ dbgs() << "*** Scheduling failed! ***\n";
+ SuccSU->dump(this);
+ dbgs() << " has been released too many times!\n";
+ llvm_unreachable(0);
}
#endif
-
- if (SuccSU->NumPredsLeft == 0) {
+ --SuccSU->NumPredsLeft;
+
+ // If all the node's predecessors are scheduled, this node is ready
+ // to be scheduled. Ignore the special ExitSU node.
+ if (SuccSU->NumPredsLeft == 0 && SuccSU != &ExitSU) {
SuccSU->isAvailable = true;
AvailableQueue->push(SuccSU);
}
}
+void ScheduleDAGRRList::ReleaseSuccessors(SUnit *SU) {
+ // Top down: release successors
+ for (SUnit::succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
+ I != E; ++I) {
+ assert(!I->isAssignedRegDep() &&
+ "The list-tdrr scheduler doesn't yet support physreg dependencies!");
+
+ ReleaseSucc(SU, &*I);
+ }
+}
/// ScheduleNodeTopDown - Add the node to the schedule. Decrement the pending
/// count of its successors. If a successor pending count is zero, add it to
/// the Available queue.
-void ScheduleDAGRRList::ScheduleNodeTopDown(SUnit *SU, unsigned CurCycle) {
- DOUT << "*** Scheduling [" << CurCycle << "]: ";
- DEBUG(SU->dump(&DAG));
- SU->Cycle = CurCycle;
+void ScheduleDAGRRList::ScheduleNodeTopDown(SUnit *SU) {
+ DEBUG(dbgs() << "*** Scheduling [" << CurCycle << "]: ");
+ DEBUG(SU->dump(this));
- AvailableQueue->ScheduledNode(SU);
+ assert(CurCycle >= SU->getDepth() && "Node scheduled above its depth!");
+ SU->setDepthToAtLeast(CurCycle);
+ Sequence.push_back(SU);
- // Top down: release successors
- for (SUnit::succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
- I != E; ++I)
- ReleaseSucc(I->Dep, I->isCtrl, CurCycle);
+ ReleaseSuccessors(SU);
SU->isScheduled = true;
+ AvailableQueue->ScheduledNode(SU);
}
/// ListScheduleTopDown - The main loop of list scheduling for top-down
/// schedulers.
void ScheduleDAGRRList::ListScheduleTopDown() {
- unsigned CurCycle = 0;
- SUnit *Entry = SUnitMap[DAG.getEntryNode().Val].front();
+ AvailableQueue->setCurCycle(CurCycle);
+
+ // Release any successors of the special Entry node.
+ ReleaseSuccessors(&EntrySU);
// All leaves to Available queue.
for (unsigned i = 0, e = SUnits.size(); i != e; ++i) {
// It is available if it has no predecessors.
- if (SUnits[i].Preds.size() == 0 && &SUnits[i] != Entry) {
+ if (SUnits[i].Preds.empty()) {
AvailableQueue->push(&SUnits[i]);
SUnits[i].isAvailable = true;
}
}
-
- // Emit the entry node first.
- ScheduleNodeTopDown(Entry, CurCycle);
- Sequence.push_back(Entry);
- ++CurCycle;
// While Available queue is not empty, grab the node with the highest
// priority. If it is not ready put it back. Schedule the node.
- std::vector<SUnit*> NotReady;
+ Sequence.reserve(SUnits.size());
while (!AvailableQueue->empty()) {
SUnit *CurSU = AvailableQueue->pop();
- while (CurSU && CurSU->CycleBound > CurCycle) {
- NotReady.push_back(CurSU);
- CurSU = AvailableQueue->pop();
- }
-
- // Add the nodes that aren't ready back onto the available list.
- AvailableQueue->push_all(NotReady);
- NotReady.clear();
- if (!CurSU)
- Sequence.push_back(0);
- else {
- ScheduleNodeTopDown(CurSU, CurCycle);
- Sequence.push_back(CurSU);
- }
- CurCycle++;
+ if (CurSU)
+ ScheduleNodeTopDown(CurSU);
+ ++CurCycle;
+ AvailableQueue->setCurCycle(CurCycle);
}
-
-
+
#ifndef NDEBUG
- // Verify that all SUnits were scheduled.
- bool AnyNotSched = false;
- for (unsigned i = 0, e = SUnits.size(); i != e; ++i) {
- if (!SUnits[i].isScheduled) {
- if (!AnyNotSched)
- cerr << "*** List scheduling failed! ***\n";
- SUnits[i].dump(&DAG);
- cerr << "has not been scheduled!\n";
- AnyNotSched = true;
- }
- }
- assert(!AnyNotSched);
+ VerifySchedule(isBottomUp);
#endif
}
-
//===----------------------------------------------------------------------===//
-// RegReductionPriorityQueue Implementation
+// RegReductionPriorityQueue Definition
//===----------------------------------------------------------------------===//
//
// This is a SchedulingPriorityQueue that schedules using Sethi Ullman numbers
// to reduce register pressure.
-//
+//
namespace {
- template<class SF>
- class RegReductionPriorityQueue;
-
- /// Sorting functions for the Available queue.
- struct bu_ls_rr_sort : public std::binary_function<SUnit*, SUnit*, bool> {
- RegReductionPriorityQueue<bu_ls_rr_sort> *SPQ;
- bu_ls_rr_sort(RegReductionPriorityQueue<bu_ls_rr_sort> *spq) : SPQ(spq) {}
- bu_ls_rr_sort(const bu_ls_rr_sort &RHS) : SPQ(RHS.SPQ) {}
-
- bool operator()(const SUnit* left, const SUnit* right) const;
+class RegReductionPQBase;
+
+struct queue_sort : public std::binary_function<SUnit*, SUnit*, bool> {
+ bool isReady(SUnit* SU, unsigned CurCycle) const { return true; }
+};
+
+/// bu_ls_rr_sort - Priority function for bottom up register pressure
+// reduction scheduler.
+struct bu_ls_rr_sort : public queue_sort {
+ enum {
+ IsBottomUp = true,
+ HasReadyFilter = false
};
- struct td_ls_rr_sort : public std::binary_function<SUnit*, SUnit*, bool> {
- RegReductionPriorityQueue<td_ls_rr_sort> *SPQ;
- td_ls_rr_sort(RegReductionPriorityQueue<td_ls_rr_sort> *spq) : SPQ(spq) {}
- td_ls_rr_sort(const td_ls_rr_sort &RHS) : SPQ(RHS.SPQ) {}
-
- bool operator()(const SUnit* left, const SUnit* right) const;
+ RegReductionPQBase *SPQ;
+ bu_ls_rr_sort(RegReductionPQBase *spq) : SPQ(spq) {}
+ bu_ls_rr_sort(const bu_ls_rr_sort &RHS) : SPQ(RHS.SPQ) {}
+
+ bool operator()(SUnit* left, SUnit* right) const;
+};
+
+// td_ls_rr_sort - Priority function for top down register pressure reduction
+// scheduler.
+struct td_ls_rr_sort : public queue_sort {
+ enum {
+ IsBottomUp = false,
+ HasReadyFilter = false
};
-} // end anonymous namespace
-static inline bool isCopyFromLiveIn(const SUnit *SU) {
- SDNode *N = SU->Node;
- return N && N->getOpcode() == ISD::CopyFromReg &&
- N->getOperand(N->getNumOperands()-1).getValueType() != MVT::Flag;
-}
+ RegReductionPQBase *SPQ;
+ td_ls_rr_sort(RegReductionPQBase *spq) : SPQ(spq) {}
+ td_ls_rr_sort(const td_ls_rr_sort &RHS) : SPQ(RHS.SPQ) {}
-namespace {
- template<class SF>
- class VISIBILITY_HIDDEN RegReductionPriorityQueue
- : public SchedulingPriorityQueue {
- std::priority_queue<SUnit*, std::vector<SUnit*>, SF> Queue;
-
- public:
- RegReductionPriorityQueue() :
- Queue(SF(this)) {}
-
- virtual void initNodes(DenseMap<SDNode*, std::vector<SUnit*> > &sumap,
- std::vector<SUnit> &sunits) {}
-
- virtual void addNode(const SUnit *SU) {}
-
- virtual void updateNode(const SUnit *SU) {}
-
- virtual void releaseState() {}
-
- virtual unsigned getNodePriority(const SUnit *SU) const {
- return 0;
- }
-
- unsigned size() const { return Queue.size(); }
+ bool operator()(const SUnit* left, const SUnit* right) const;
+};
- bool empty() const { return Queue.empty(); }
-
- void push(SUnit *U) {
- Queue.push(U);
- }
- void push_all(const std::vector<SUnit *> &Nodes) {
- for (unsigned i = 0, e = Nodes.size(); i != e; ++i)
- Queue.push(Nodes[i]);
- }
-
- SUnit *pop() {
- if (empty()) return NULL;
- SUnit *V = Queue.top();
- Queue.pop();
- return V;
- }
+// src_ls_rr_sort - Priority function for source order scheduler.
+struct src_ls_rr_sort : public queue_sort {
+ enum {
+ IsBottomUp = true,
+ HasReadyFilter = false
+ };
- /// remove - This is a really inefficient way to remove a node from a
- /// priority queue. We should roll our own heap to make this better or
- /// something.
- void remove(SUnit *SU) {
- std::vector<SUnit*> Temp;
-
- assert(!Queue.empty() && "Not in queue!");
- while (Queue.top() != SU) {
- Temp.push_back(Queue.top());
- Queue.pop();
- assert(!Queue.empty() && "Not in queue!");
- }
+ RegReductionPQBase *SPQ;
+ src_ls_rr_sort(RegReductionPQBase *spq)
+ : SPQ(spq) {}
+ src_ls_rr_sort(const src_ls_rr_sort &RHS)
+ : SPQ(RHS.SPQ) {}
- // Remove the node from the PQ.
- Queue.pop();
-
- // Add all the other nodes back.
- for (unsigned i = 0, e = Temp.size(); i != e; ++i)
- Queue.push(Temp[i]);
- }
+ bool operator()(SUnit* left, SUnit* right) const;
+};
+
+// hybrid_ls_rr_sort - Priority function for hybrid scheduler.
+struct hybrid_ls_rr_sort : public queue_sort {
+ enum {
+ IsBottomUp = true,
+ HasReadyFilter = false
};
- template<class SF>
- class VISIBILITY_HIDDEN BURegReductionPriorityQueue
- : public RegReductionPriorityQueue<SF> {
- // SUnitMap SDNode to SUnit mapping (n -> n).
- DenseMap<SDNode*, std::vector<SUnit*> > *SUnitMap;
-
- // SUnits - The SUnits for the current graph.
- const std::vector<SUnit> *SUnits;
-
- // SethiUllmanNumbers - The SethiUllman number for each node.
- std::vector<unsigned> SethiUllmanNumbers;
-
- const TargetInstrInfo *TII;
- const MRegisterInfo *MRI;
- public:
- explicit BURegReductionPriorityQueue(const TargetInstrInfo *tii,
- const MRegisterInfo *mri)
- : TII(tii), MRI(mri) {}
-
- void initNodes(DenseMap<SDNode*, std::vector<SUnit*> > &sumap,
- std::vector<SUnit> &sunits) {
- SUnitMap = &sumap;
- SUnits = &sunits;
- // Add pseudo dependency edges for two-address nodes.
- AddPseudoTwoAddrDeps();
- // Calculate node priorities.
- CalculateSethiUllmanNumbers();
- }
+ RegReductionPQBase *SPQ;
+ hybrid_ls_rr_sort(RegReductionPQBase *spq)
+ : SPQ(spq) {}
+ hybrid_ls_rr_sort(const hybrid_ls_rr_sort &RHS)
+ : SPQ(RHS.SPQ) {}
- void addNode(const SUnit *SU) {
- SethiUllmanNumbers.resize(SUnits->size(), 0);
- CalcNodeSethiUllmanNumber(SU);
- }
+ bool isReady(SUnit *SU, unsigned CurCycle) const;
- void updateNode(const SUnit *SU) {
- SethiUllmanNumbers[SU->NodeNum] = 0;
- CalcNodeSethiUllmanNumber(SU);
- }
+ bool operator()(SUnit* left, SUnit* right) const;
+};
+
+// ilp_ls_rr_sort - Priority function for ILP (instruction level parallelism)
+// scheduler.
+struct ilp_ls_rr_sort : public queue_sort {
+ enum {
+ IsBottomUp = true,
+ HasReadyFilter = false
+ };
+
+ RegReductionPQBase *SPQ;
+ ilp_ls_rr_sort(RegReductionPQBase *spq)
+ : SPQ(spq) {}
+ ilp_ls_rr_sort(const ilp_ls_rr_sort &RHS)
+ : SPQ(RHS.SPQ) {}
+
+ bool isReady(SUnit *SU, unsigned CurCycle) const;
+
+ bool operator()(SUnit* left, SUnit* right) const;
+};
+
+class RegReductionPQBase : public SchedulingPriorityQueue {
+protected:
+ std::vector<SUnit*> Queue;
+ unsigned CurQueueId;
+ bool TracksRegPressure;
+
+ // SUnits - The SUnits for the current graph.
+ std::vector<SUnit> *SUnits;
- void releaseState() {
- SUnits = 0;
- SethiUllmanNumbers.clear();
+ MachineFunction &MF;
+ const TargetInstrInfo *TII;
+ const TargetRegisterInfo *TRI;
+ const TargetLowering *TLI;
+ ScheduleDAGRRList *scheduleDAG;
+
+ // SethiUllmanNumbers - The SethiUllman number for each node.
+ std::vector<unsigned> SethiUllmanNumbers;
+
+ /// RegPressure - Tracking current reg pressure per register class.
+ ///
+ std::vector<unsigned> RegPressure;
+
+ /// RegLimit - Tracking the number of allocatable registers per register
+ /// class.
+ std::vector<unsigned> RegLimit;
+
+public:
+ RegReductionPQBase(MachineFunction &mf,
+ bool hasReadyFilter,
+ bool tracksrp,
+ const TargetInstrInfo *tii,
+ const TargetRegisterInfo *tri,
+ const TargetLowering *tli)
+ : SchedulingPriorityQueue(hasReadyFilter),
+ CurQueueId(0), TracksRegPressure(tracksrp),
+ MF(mf), TII(tii), TRI(tri), TLI(tli), scheduleDAG(NULL) {
+ if (TracksRegPressure) {
+ unsigned NumRC = TRI->getNumRegClasses();
+ RegLimit.resize(NumRC);
+ RegPressure.resize(NumRC);
+ std::fill(RegLimit.begin(), RegLimit.end(), 0);
+ std::fill(RegPressure.begin(), RegPressure.end(), 0);
+ for (TargetRegisterInfo::regclass_iterator I = TRI->regclass_begin(),
+ E = TRI->regclass_end(); I != E; ++I)
+ RegLimit[(*I)->getID()] = tli->getRegPressureLimit(*I, MF);
}
+ }
+
+ void setScheduleDAG(ScheduleDAGRRList *scheduleDag) {
+ scheduleDAG = scheduleDag;
+ }
+
+ ScheduleHazardRecognizer* getHazardRec() {
+ return scheduleDAG->getHazardRec();
+ }
+
+ void initNodes(std::vector<SUnit> &sunits);
+
+ void addNode(const SUnit *SU);
+
+ void updateNode(const SUnit *SU);
+
+ void releaseState() {
+ SUnits = 0;
+ SethiUllmanNumbers.clear();
+ std::fill(RegPressure.begin(), RegPressure.end(), 0);
+ }
+
+ unsigned getNodePriority(const SUnit *SU) const;
+
+ unsigned getNodeOrdering(const SUnit *SU) const {
+ return scheduleDAG->DAG->GetOrdering(SU->getNode());
+ }
+
+ bool empty() const { return Queue.empty(); }
+
+ void push(SUnit *U) {
+ assert(!U->NodeQueueId && "Node in the queue already");
+ U->NodeQueueId = ++CurQueueId;
+ Queue.push_back(U);
+ }
+
+ void remove(SUnit *SU) {
+ assert(!Queue.empty() && "Queue is empty!");
+ assert(SU->NodeQueueId != 0 && "Not in queue!");
+ std::vector<SUnit *>::iterator I = std::find(Queue.begin(), Queue.end(),
+ SU);
+ if (I != prior(Queue.end()))
+ std::swap(*I, Queue.back());
+ Queue.pop_back();
+ SU->NodeQueueId = 0;
+ }
+
+ bool tracksRegPressure() const { return TracksRegPressure; }
+
+ void dumpRegPressure() const;
+
+ bool HighRegPressure(const SUnit *SU) const;
- unsigned getNodePriority(const SUnit *SU) const {
- assert(SU->NodeNum < SethiUllmanNumbers.size());
- unsigned Opc = SU->Node ? SU->Node->getOpcode() : 0;
- if (Opc == ISD::CopyFromReg && !isCopyFromLiveIn(SU))
- // CopyFromReg should be close to its def because it restricts
- // allocation choices. But if it is a livein then perhaps we want it
- // closer to its uses so it can be coalesced.
- return 0xffff;
- else if (Opc == ISD::TokenFactor || Opc == ISD::CopyToReg)
- // CopyToReg should be close to its uses to facilitate coalescing and
- // avoid spilling.
- return 0;
- else if (Opc == TargetInstrInfo::EXTRACT_SUBREG ||
- Opc == TargetInstrInfo::INSERT_SUBREG)
- // EXTRACT_SUBREG / INSERT_SUBREG should be close to its use to
- // facilitate coalescing.
- return 0;
- else if (SU->NumSuccs == 0)
- // If SU does not have a use, i.e. it doesn't produce a value that would
- // be consumed (e.g. store), then it terminates a chain of computation.
- // Give it a large SethiUllman number so it will be scheduled right
- // before its predecessors that it doesn't lengthen their live ranges.
- return 0xffff;
- else if (SU->NumPreds == 0)
- // If SU does not have a def, schedule it close to its uses because it
- // does not lengthen any live ranges.
- return 0;
+ bool MayReduceRegPressure(SUnit *SU) const;
+
+ int RegPressureDiff(SUnit *SU, unsigned &LiveUses) const;
+
+ void ScheduledNode(SUnit *SU);
+
+ void UnscheduledNode(SUnit *SU);
+
+protected:
+ bool canClobber(const SUnit *SU, const SUnit *Op);
+ void AddPseudoTwoAddrDeps();
+ void PrescheduleNodesWithMultipleUses();
+ void CalculateSethiUllmanNumbers();
+};
+
+template<class SF>
+class RegReductionPriorityQueue : public RegReductionPQBase {
+ static SUnit *popFromQueue(std::vector<SUnit*> &Q, SF &Picker) {
+ std::vector<SUnit *>::iterator Best = Q.begin();
+ for (std::vector<SUnit *>::iterator I = llvm::next(Q.begin()),
+ E = Q.end(); I != E; ++I)
+ if (Picker(*Best, *I))
+ Best = I;
+ SUnit *V = *Best;
+ if (Best != prior(Q.end()))
+ std::swap(*Best, Q.back());
+ Q.pop_back();
+ return V;
+ }
+
+ SF Picker;
+
+public:
+ RegReductionPriorityQueue(MachineFunction &mf,
+ bool tracksrp,
+ const TargetInstrInfo *tii,
+ const TargetRegisterInfo *tri,
+ const TargetLowering *tli)
+ : RegReductionPQBase(mf, SF::HasReadyFilter, tracksrp, tii, tri, tli),
+ Picker(this) {}
+
+ bool isBottomUp() const { return SF::IsBottomUp; }
+
+ bool isReady(SUnit *U) const {
+ return Picker.HasReadyFilter && Picker.isReady(U, getCurCycle());
+ }
+
+ SUnit *pop() {
+ if (Queue.empty()) return NULL;
+
+ SUnit *V = popFromQueue(Queue, Picker);
+ V->NodeQueueId = 0;
+ return V;
+ }
+
+ void dump(ScheduleDAG *DAG) const {
+ // Emulate pop() without clobbering NodeQueueIds.
+ std::vector<SUnit*> DumpQueue = Queue;
+ SF DumpPicker = Picker;
+ while (!DumpQueue.empty()) {
+ SUnit *SU = popFromQueue(DumpQueue, DumpPicker);
+ if (isBottomUp())
+ dbgs() << "Height " << SU->getHeight() << ": ";
else
- return SethiUllmanNumbers[SU->NodeNum];
+ dbgs() << "Depth " << SU->getDepth() << ": ";
+ SU->dump(DAG);
}
+ }
+};
- private:
- bool canClobber(SUnit *SU, SUnit *Op);
- void AddPseudoTwoAddrDeps();
- void CalculateSethiUllmanNumbers();
- unsigned CalcNodeSethiUllmanNumber(const SUnit *SU);
- };
+typedef RegReductionPriorityQueue<bu_ls_rr_sort>
+BURegReductionPriorityQueue;
+
+typedef RegReductionPriorityQueue<td_ls_rr_sort>
+TDRegReductionPriorityQueue;
+
+typedef RegReductionPriorityQueue<src_ls_rr_sort>
+SrcRegReductionPriorityQueue;
+
+typedef RegReductionPriorityQueue<hybrid_ls_rr_sort>
+HybridBURRPriorityQueue;
+
+typedef RegReductionPriorityQueue<ilp_ls_rr_sort>
+ILPBURRPriorityQueue;
+} // end anonymous namespace
+
+//===----------------------------------------------------------------------===//
+// Static Node Priority for Register Pressure Reduction
+//===----------------------------------------------------------------------===//
+
+/// CalcNodeSethiUllmanNumber - Compute Sethi Ullman number.
+/// Smaller number is the higher priority.
+static unsigned
+CalcNodeSethiUllmanNumber(const SUnit *SU, std::vector<unsigned> &SUNumbers) {
+ unsigned &SethiUllmanNumber = SUNumbers[SU->NodeNum];
+ if (SethiUllmanNumber != 0)
+ return SethiUllmanNumber;
+
+ unsigned Extra = 0;
+ for (SUnit::const_pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
+ I != E; ++I) {
+ if (I->isCtrl()) continue; // ignore chain preds
+ SUnit *PredSU = I->getSUnit();
+ unsigned PredSethiUllman = CalcNodeSethiUllmanNumber(PredSU, SUNumbers);
+ if (PredSethiUllman > SethiUllmanNumber) {
+ SethiUllmanNumber = PredSethiUllman;
+ Extra = 0;
+ } else if (PredSethiUllman == SethiUllmanNumber)
+ ++Extra;
+ }
+
+ SethiUllmanNumber += Extra;
+
+ if (SethiUllmanNumber == 0)
+ SethiUllmanNumber = 1;
+
+ return SethiUllmanNumber;
+}
+
+/// CalculateSethiUllmanNumbers - Calculate Sethi-Ullman numbers of all
+/// scheduling units.
+void RegReductionPQBase::CalculateSethiUllmanNumbers() {
+ SethiUllmanNumbers.assign(SUnits->size(), 0);
+
+ for (unsigned i = 0, e = SUnits->size(); i != e; ++i)
+ CalcNodeSethiUllmanNumber(&(*SUnits)[i], SethiUllmanNumbers);
+}
+
+void RegReductionPQBase::initNodes(std::vector<SUnit> &sunits) {
+ SUnits = &sunits;
+ // Add pseudo dependency edges for two-address nodes.
+ AddPseudoTwoAddrDeps();
+ // Reroute edges to nodes with multiple uses.
+ if (!TracksRegPressure)
+ PrescheduleNodesWithMultipleUses();
+ // Calculate node priorities.
+ CalculateSethiUllmanNumbers();
+}
+
+void RegReductionPQBase::addNode(const SUnit *SU) {
+ unsigned SUSize = SethiUllmanNumbers.size();
+ if (SUnits->size() > SUSize)
+ SethiUllmanNumbers.resize(SUSize*2, 0);
+ CalcNodeSethiUllmanNumber(SU, SethiUllmanNumbers);
+}
+
+void RegReductionPQBase::updateNode(const SUnit *SU) {
+ SethiUllmanNumbers[SU->NodeNum] = 0;
+ CalcNodeSethiUllmanNumber(SU, SethiUllmanNumbers);
+}
+// Lower priority means schedule further down. For bottom-up scheduling, lower
+// priority SUs are scheduled before higher priority SUs.
+unsigned RegReductionPQBase::getNodePriority(const SUnit *SU) const {
+ assert(SU->NodeNum < SethiUllmanNumbers.size());
+ unsigned Opc = SU->getNode() ? SU->getNode()->getOpcode() : 0;
+ if (Opc == ISD::TokenFactor || Opc == ISD::CopyToReg)
+ // CopyToReg should be close to its uses to facilitate coalescing and
+ // avoid spilling.
+ return 0;
+ if (Opc == TargetOpcode::EXTRACT_SUBREG ||
+ Opc == TargetOpcode::SUBREG_TO_REG ||
+ Opc == TargetOpcode::INSERT_SUBREG)
+ // EXTRACT_SUBREG, INSERT_SUBREG, and SUBREG_TO_REG nodes should be
+ // close to their uses to facilitate coalescing.
+ return 0;
+ if (SU->NumSuccs == 0 && SU->NumPreds != 0)
+ // If SU does not have a register use, i.e. it doesn't produce a value
+ // that would be consumed (e.g. store), then it terminates a chain of
+ // computation. Give it a large SethiUllman number so it will be
+ // scheduled right before its predecessors that it doesn't lengthen
+ // their live ranges.
+ return 0xffff;
+ if (SU->NumPreds == 0 && SU->NumSuccs != 0)
+ // If SU does not have a register def, schedule it close to its uses
+ // because it does not lengthen any live ranges.
+ return 0;
+ return SethiUllmanNumbers[SU->NodeNum];
+}
- template<class SF>
- class VISIBILITY_HIDDEN TDRegReductionPriorityQueue
- : public RegReductionPriorityQueue<SF> {
- // SUnitMap SDNode to SUnit mapping (n -> n).
- DenseMap<SDNode*, std::vector<SUnit*> > *SUnitMap;
+//===----------------------------------------------------------------------===//
+// Register Pressure Tracking
+//===----------------------------------------------------------------------===//
- // SUnits - The SUnits for the current graph.
- const std::vector<SUnit> *SUnits;
-
- // SethiUllmanNumbers - The SethiUllman number for each node.
- std::vector<unsigned> SethiUllmanNumbers;
+void RegReductionPQBase::dumpRegPressure() const {
+ for (TargetRegisterInfo::regclass_iterator I = TRI->regclass_begin(),
+ E = TRI->regclass_end(); I != E; ++I) {
+ const TargetRegisterClass *RC = *I;
+ unsigned Id = RC->getID();
+ unsigned RP = RegPressure[Id];
+ if (!RP) continue;
+ DEBUG(dbgs() << RC->getName() << ": " << RP << " / " << RegLimit[Id]
+ << '\n');
+ }
+}
- public:
- TDRegReductionPriorityQueue() {}
+bool RegReductionPQBase::HighRegPressure(const SUnit *SU) const {
+ if (!TLI)
+ return false;
- void initNodes(DenseMap<SDNode*, std::vector<SUnit*> > &sumap,
- std::vector<SUnit> &sunits) {
- SUnitMap = &sumap;
- SUnits = &sunits;
- // Calculate node priorities.
- CalculateSethiUllmanNumbers();
+ for (SUnit::const_pred_iterator I = SU->Preds.begin(),E = SU->Preds.end();
+ I != E; ++I) {
+ if (I->isCtrl())
+ continue;
+ SUnit *PredSU = I->getSUnit();
+ // NumRegDefsLeft is zero when enough uses of this node have been scheduled
+ // to cover the number of registers defined (they are all live).
+ if (PredSU->NumRegDefsLeft == 0) {
+ continue;
}
+ for (ScheduleDAGSDNodes::RegDefIter RegDefPos(PredSU, scheduleDAG);
+ RegDefPos.IsValid(); RegDefPos.Advance()) {
+ EVT VT = RegDefPos.GetValue();
+ unsigned RCId = TLI->getRepRegClassFor(VT)->getID();
+ unsigned Cost = TLI->getRepRegClassCostFor(VT);
+ if ((RegPressure[RCId] + Cost) >= RegLimit[RCId])
+ return true;
+ }
+ }
+ return false;
+}
+
+bool RegReductionPQBase::MayReduceRegPressure(SUnit *SU) const {
+ const SDNode *N = SU->getNode();
+
+ if (!N->isMachineOpcode() || !SU->NumSuccs)
+ return false;
- void addNode(const SUnit *SU) {
- SethiUllmanNumbers.resize(SUnits->size(), 0);
- CalcNodeSethiUllmanNumber(SU);
+ unsigned NumDefs = TII->get(N->getMachineOpcode()).getNumDefs();
+ for (unsigned i = 0; i != NumDefs; ++i) {
+ EVT VT = N->getValueType(i);
+ if (!N->hasAnyUseOfValue(i))
+ continue;
+ unsigned RCId = TLI->getRepRegClassFor(VT)->getID();
+ if (RegPressure[RCId] >= RegLimit[RCId])
+ return true;
+ }
+ return false;
+}
+
+// Compute the register pressure contribution by this instruction by count up
+// for uses that are not live and down for defs. Only count register classes
+// that are already under high pressure. As a side effect, compute the number of
+// uses of registers that are already live.
+//
+// FIXME: This encompasses the logic in HighRegPressure and MayReduceRegPressure
+// so could probably be factored.
+int RegReductionPQBase::RegPressureDiff(SUnit *SU, unsigned &LiveUses) const {
+ LiveUses = 0;
+ int PDiff = 0;
+ for (SUnit::const_pred_iterator I = SU->Preds.begin(),E = SU->Preds.end();
+ I != E; ++I) {
+ if (I->isCtrl())
+ continue;
+ SUnit *PredSU = I->getSUnit();
+ // NumRegDefsLeft is zero when enough uses of this node have been scheduled
+ // to cover the number of registers defined (they are all live).
+ if (PredSU->NumRegDefsLeft == 0) {
+ if (PredSU->getNode()->isMachineOpcode())
+ ++LiveUses;
+ continue;
+ }
+ for (ScheduleDAGSDNodes::RegDefIter RegDefPos(PredSU, scheduleDAG);
+ RegDefPos.IsValid(); RegDefPos.Advance()) {
+ EVT VT = RegDefPos.GetValue();
+ unsigned RCId = TLI->getRepRegClassFor(VT)->getID();
+ if (RegPressure[RCId] >= RegLimit[RCId])
+ ++PDiff;
}
+ }
+ const SDNode *N = SU->getNode();
+
+ if (!N->isMachineOpcode() || !SU->NumSuccs)
+ return PDiff;
+
+ unsigned NumDefs = TII->get(N->getMachineOpcode()).getNumDefs();
+ for (unsigned i = 0; i != NumDefs; ++i) {
+ EVT VT = N->getValueType(i);
+ if (!N->hasAnyUseOfValue(i))
+ continue;
+ unsigned RCId = TLI->getRepRegClassFor(VT)->getID();
+ if (RegPressure[RCId] >= RegLimit[RCId])
+ --PDiff;
+ }
+ return PDiff;
+}
+
+void RegReductionPQBase::ScheduledNode(SUnit *SU) {
+ if (!TracksRegPressure)
+ return;
- void updateNode(const SUnit *SU) {
- SethiUllmanNumbers[SU->NodeNum] = 0;
- CalcNodeSethiUllmanNumber(SU);
+ for (SUnit::pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
+ I != E; ++I) {
+ if (I->isCtrl())
+ continue;
+ SUnit *PredSU = I->getSUnit();
+ // NumRegDefsLeft is zero when enough uses of this node have been scheduled
+ // to cover the number of registers defined (they are all live).
+ if (PredSU->NumRegDefsLeft == 0) {
+ continue;
+ }
+ // FIXME: The ScheduleDAG currently loses information about which of a
+ // node's values is consumed by each dependence. Consequently, if the node
+ // defines multiple register classes, we don't know which to pressurize
+ // here. Instead the following loop consumes the register defs in an
+ // arbitrary order. At least it handles the common case of clustered loads
+ // to the same class. For precise liveness, each SDep needs to indicate the
+ // result number. But that tightly couples the ScheduleDAG with the
+ // SelectionDAG making updates tricky. A simpler hack would be to attach a
+ // value type or register class to SDep.
+ //
+ // The most important aspect of register tracking is balancing the increase
+ // here with the reduction further below. Note that this SU may use multiple
+ // defs in PredSU. The can't be determined here, but we've already
+ // compensated by reducing NumRegDefsLeft in PredSU during
+ // ScheduleDAGSDNodes::AddSchedEdges.
+ --PredSU->NumRegDefsLeft;
+ unsigned SkipRegDefs = PredSU->NumRegDefsLeft;
+ for (ScheduleDAGSDNodes::RegDefIter RegDefPos(PredSU, scheduleDAG);
+ RegDefPos.IsValid(); RegDefPos.Advance(), --SkipRegDefs) {
+ if (SkipRegDefs)
+ continue;
+ EVT VT = RegDefPos.GetValue();
+ unsigned RCId = TLI->getRepRegClassFor(VT)->getID();
+ RegPressure[RCId] += TLI->getRepRegClassCostFor(VT);
+ break;
}
+ }
- void releaseState() {
- SUnits = 0;
- SethiUllmanNumbers.clear();
+ // We should have this assert, but there may be dead SDNodes that never
+ // materialize as SUnits, so they don't appear to generate liveness.
+ //assert(SU->NumRegDefsLeft == 0 && "not all regdefs have scheduled uses");
+ int SkipRegDefs = (int)SU->NumRegDefsLeft;
+ for (ScheduleDAGSDNodes::RegDefIter RegDefPos(SU, scheduleDAG);
+ RegDefPos.IsValid(); RegDefPos.Advance(), --SkipRegDefs) {
+ if (SkipRegDefs > 0)
+ continue;
+ EVT VT = RegDefPos.GetValue();
+ unsigned RCId = TLI->getRepRegClassFor(VT)->getID();
+ if (RegPressure[RCId] < TLI->getRepRegClassCostFor(VT)) {
+ // Register pressure tracking is imprecise. This can happen. But we try
+ // hard not to let it happen because it likely results in poor scheduling.
+ DEBUG(dbgs() << " SU(" << SU->NodeNum << ") has too many regdefs\n");
+ RegPressure[RCId] = 0;
}
+ else {
+ RegPressure[RCId] -= TLI->getRepRegClassCostFor(VT);
+ }
+ }
+ dumpRegPressure();
+}
+
+void RegReductionPQBase::UnscheduledNode(SUnit *SU) {
+ if (!TracksRegPressure)
+ return;
+
+ const SDNode *N = SU->getNode();
+ if (!N->isMachineOpcode()) {
+ if (N->getOpcode() != ISD::CopyToReg)
+ return;
+ } else {
+ unsigned Opc = N->getMachineOpcode();
+ if (Opc == TargetOpcode::EXTRACT_SUBREG ||
+ Opc == TargetOpcode::INSERT_SUBREG ||
+ Opc == TargetOpcode::SUBREG_TO_REG ||
+ Opc == TargetOpcode::REG_SEQUENCE ||
+ Opc == TargetOpcode::IMPLICIT_DEF)
+ return;
+ }
- unsigned getNodePriority(const SUnit *SU) const {
- assert(SU->NodeNum < SethiUllmanNumbers.size());
- return SethiUllmanNumbers[SU->NodeNum];
+ for (SUnit::pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
+ I != E; ++I) {
+ if (I->isCtrl())
+ continue;
+ SUnit *PredSU = I->getSUnit();
+ // NumSuccsLeft counts all deps. Don't compare it with NumSuccs which only
+ // counts data deps.
+ if (PredSU->NumSuccsLeft != PredSU->Succs.size())
+ continue;
+ const SDNode *PN = PredSU->getNode();
+ if (!PN->isMachineOpcode()) {
+ if (PN->getOpcode() == ISD::CopyFromReg) {
+ EVT VT = PN->getValueType(0);
+ unsigned RCId = TLI->getRepRegClassFor(VT)->getID();
+ RegPressure[RCId] += TLI->getRepRegClassCostFor(VT);
+ }
+ continue;
}
+ unsigned POpc = PN->getMachineOpcode();
+ if (POpc == TargetOpcode::IMPLICIT_DEF)
+ continue;
+ if (POpc == TargetOpcode::EXTRACT_SUBREG) {
+ EVT VT = PN->getOperand(0).getValueType();
+ unsigned RCId = TLI->getRepRegClassFor(VT)->getID();
+ RegPressure[RCId] += TLI->getRepRegClassCostFor(VT);
+ continue;
+ } else if (POpc == TargetOpcode::INSERT_SUBREG ||
+ POpc == TargetOpcode::SUBREG_TO_REG) {
+ EVT VT = PN->getValueType(0);
+ unsigned RCId = TLI->getRepRegClassFor(VT)->getID();
+ RegPressure[RCId] += TLI->getRepRegClassCostFor(VT);
+ continue;
+ }
+ unsigned NumDefs = TII->get(PN->getMachineOpcode()).getNumDefs();
+ for (unsigned i = 0; i != NumDefs; ++i) {
+ EVT VT = PN->getValueType(i);
+ if (!PN->hasAnyUseOfValue(i))
+ continue;
+ unsigned RCId = TLI->getRepRegClassFor(VT)->getID();
+ if (RegPressure[RCId] < TLI->getRepRegClassCostFor(VT))
+ // Register pressure tracking is imprecise. This can happen.
+ RegPressure[RCId] = 0;
+ else
+ RegPressure[RCId] -= TLI->getRepRegClassCostFor(VT);
+ }
+ }
- private:
- void CalculateSethiUllmanNumbers();
- unsigned CalcNodeSethiUllmanNumber(const SUnit *SU);
- };
+ // Check for isMachineOpcode() as PrescheduleNodesWithMultipleUses()
+ // may transfer data dependencies to CopyToReg.
+ if (SU->NumSuccs && N->isMachineOpcode()) {
+ unsigned NumDefs = TII->get(N->getMachineOpcode()).getNumDefs();
+ for (unsigned i = NumDefs, e = N->getNumValues(); i != e; ++i) {
+ EVT VT = N->getValueType(i);
+ if (VT == MVT::Glue || VT == MVT::Other)
+ continue;
+ if (!N->hasAnyUseOfValue(i))
+ continue;
+ unsigned RCId = TLI->getRepRegClassFor(VT)->getID();
+ RegPressure[RCId] += TLI->getRepRegClassCostFor(VT);
+ }
+ }
+
+ dumpRegPressure();
}
+//===----------------------------------------------------------------------===//
+// Dynamic Node Priority for Register Pressure Reduction
+//===----------------------------------------------------------------------===//
+
/// closestSucc - Returns the scheduled cycle of the successor which is
-/// closet to the current cycle.
+/// closest to the current cycle.
static unsigned closestSucc(const SUnit *SU) {
- unsigned MaxCycle = 0;
+ unsigned MaxHeight = 0;
for (SUnit::const_succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
I != E; ++I) {
- unsigned Cycle = I->Dep->Cycle;
+ if (I->isCtrl()) continue; // ignore chain succs
+ unsigned Height = I->getSUnit()->getHeight();
// If there are bunch of CopyToRegs stacked up, they should be considered
// to be at the same position.
- if (I->Dep->Node && I->Dep->Node->getOpcode() == ISD::CopyToReg)
- Cycle = closestSucc(I->Dep)+1;
- if (Cycle > MaxCycle)
- MaxCycle = Cycle;
+ if (I->getSUnit()->getNode() &&
+ I->getSUnit()->getNode()->getOpcode() == ISD::CopyToReg)
+ Height = closestSucc(I->getSUnit())+1;
+ if (Height > MaxHeight)
+ MaxHeight = Height;
}
- return MaxCycle;
+ return MaxHeight;
}
/// calcMaxScratches - Returns an cost estimate of the worse case requirement
-/// for scratch registers. Live-in operands and live-out results don't count
-/// since they are "fixed".
+/// for scratch registers, i.e. number of data dependencies.
static unsigned calcMaxScratches(const SUnit *SU) {
unsigned Scratches = 0;
for (SUnit::const_pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
I != E; ++I) {
- if (I->isCtrl) continue; // ignore chain preds
- if (I->Dep->Node->getOpcode() != ISD::CopyFromReg)
- Scratches++;
+ if (I->isCtrl()) continue; // ignore chain preds
+ Scratches++;
}
+ return Scratches;
+}
+
+/// hasOnlyLiveOutUse - Return true if SU has a single value successor that is a
+/// CopyToReg to a virtual register. This SU def is probably a liveout and
+/// it has no other use. It should be scheduled closer to the terminator.
+static bool hasOnlyLiveOutUses(const SUnit *SU) {
+ bool RetVal = false;
for (SUnit::const_succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
I != E; ++I) {
- if (I->isCtrl) continue; // ignore chain succs
- if (I->Dep->Node->getOpcode() != ISD::CopyToReg)
- Scratches += 10;
+ if (I->isCtrl()) continue;
+ const SUnit *SuccSU = I->getSUnit();
+ if (SuccSU->getNode() && SuccSU->getNode()->getOpcode() == ISD::CopyToReg) {
+ unsigned Reg =
+ cast<RegisterSDNode>(SuccSU->getNode()->getOperand(1))->getReg();
+ if (TargetRegisterInfo::isVirtualRegister(Reg)) {
+ RetVal = true;
+ continue;
+ }
+ }
+ return false;
}
- return Scratches;
+ return RetVal;
}
-// Bottom up
-bool bu_ls_rr_sort::operator()(const SUnit *left, const SUnit *right) const {
- // There used to be a special tie breaker here that looked for
- // two-address instructions and preferred the instruction with a
- // def&use operand. The special case triggered diagnostics when
- // _GLIBCXX_DEBUG was enabled because it broke the strict weak
- // ordering that priority_queue requires. It didn't help much anyway
- // because AddPseudoTwoAddrDeps already covers many of the cases
- // where it would have applied. In addition, it's counter-intuitive
- // that a tie breaker would be the first thing attempted. There's a
- // "real" tie breaker below that is the operation of last resort.
- // The fact that the "special tie breaker" would trigger when there
- // wasn't otherwise a tie is what broke the strict weak ordering
- // constraint.
+/// UnitsSharePred - Return true if the two scheduling units share a common
+/// data predecessor.
+static bool UnitsSharePred(const SUnit *left, const SUnit *right) {
+ SmallSet<const SUnit*, 4> Preds;
+ for (SUnit::const_pred_iterator I = left->Preds.begin(),E = left->Preds.end();
+ I != E; ++I) {
+ if (I->isCtrl()) continue; // ignore chain preds
+ Preds.insert(I->getSUnit());
+ }
+ for (SUnit::const_pred_iterator I = right->Preds.begin(),E = right->Preds.end();
+ I != E; ++I) {
+ if (I->isCtrl()) continue; // ignore chain preds
+ if (Preds.count(I->getSUnit()))
+ return true;
+ }
+ return false;
+}
+
+// Check for either a dependence (latency) or resource (hazard) stall.
+//
+// Note: The ScheduleHazardRecognizer interface requires a non-const SU.
+static bool BUHasStall(SUnit *SU, int Height, RegReductionPQBase *SPQ) {
+ if ((int)SPQ->getCurCycle() < Height) return true;
+ if (SPQ->getHazardRec()->getHazardType(SU, 0)
+ != ScheduleHazardRecognizer::NoHazard)
+ return true;
+ return false;
+}
+
+// Return -1 if left has higher priority, 1 if right has higher priority.
+// Return 0 if latency-based priority is equivalent.
+static int BUCompareLatency(SUnit *left, SUnit *right, bool checkPref,
+ RegReductionPQBase *SPQ) {
+ // If the two nodes share an operand and one of them has a single
+ // use that is a live out copy, favor the one that is live out. Otherwise
+ // it will be difficult to eliminate the copy if the instruction is a
+ // loop induction variable update. e.g.
+ // BB:
+ // sub r1, r3, #1
+ // str r0, [r2, r3]
+ // mov r3, r1
+ // cmp
+ // bne BB
+ bool SharePred = UnitsSharePred(left, right);
+ // FIXME: Only adjust if BB is a loop back edge.
+ // FIXME: What's the cost of a copy?
+ int LBonus = (SharePred && hasOnlyLiveOutUses(left)) ? 1 : 0;
+ int RBonus = (SharePred && hasOnlyLiveOutUses(right)) ? 1 : 0;
+ int LHeight = (int)left->getHeight() - LBonus;
+ int RHeight = (int)right->getHeight() - RBonus;
+
+ bool LStall = (!checkPref || left->SchedulingPref == Sched::Latency) &&
+ BUHasStall(left, LHeight, SPQ);
+ bool RStall = (!checkPref || right->SchedulingPref == Sched::Latency) &&
+ BUHasStall(right, RHeight, SPQ);
+
+ // If scheduling one of the node will cause a pipeline stall, delay it.
+ // If scheduling either one of the node will cause a pipeline stall, sort
+ // them according to their height.
+ if (LStall) {
+ if (!RStall)
+ return 1;
+ if (LHeight != RHeight)
+ return LHeight > RHeight ? 1 : -1;
+ } else if (RStall)
+ return -1;
+
+ // If either node is scheduling for latency, sort them by height/depth
+ // and latency.
+ if (!checkPref || (left->SchedulingPref == Sched::Latency ||
+ right->SchedulingPref == Sched::Latency)) {
+ if (DisableSchedCycles) {
+ if (LHeight != RHeight)
+ return LHeight > RHeight ? 1 : -1;
+ }
+ else {
+ // If neither instruction stalls (!LStall && !RStall) then
+ // it's height is already covered so only its depth matters. We also reach
+ // this if both stall but have the same height.
+ unsigned LDepth = left->getDepth();
+ unsigned RDepth = right->getDepth();
+ if (LDepth != RDepth) {
+ DEBUG(dbgs() << " Comparing latency of SU (" << left->NodeNum
+ << ") depth " << LDepth << " vs SU (" << right->NodeNum
+ << ") depth " << RDepth << "\n");
+ return LDepth < RDepth ? 1 : -1;
+ }
+ }
+ if (left->Latency != right->Latency)
+ return left->Latency > right->Latency ? 1 : -1;
+ }
+ return 0;
+}
+static bool BURRSort(SUnit *left, SUnit *right, RegReductionPQBase *SPQ) {
unsigned LPriority = SPQ->getNodePriority(left);
unsigned RPriority = SPQ->getNodePriority(right);
- if (LPriority > RPriority)
+ if (LPriority != RPriority) {
+ DEBUG(++FactorCount[FactUllman]);
+ return LPriority > RPriority;
+ }
+ // Try schedule def + use closer when Sethi-Ullman numbers are the same.
+ // e.g.
+ // t1 = op t2, c1
+ // t3 = op t4, c2
+ //
+ // and the following instructions are both ready.
+ // t2 = op c3
+ // t4 = op c4
+ //
+ // Then schedule t2 = op first.
+ // i.e.
+ // t4 = op c4
+ // t2 = op c3
+ // t1 = op t2, c1
+ // t3 = op t4, c2
+ //
+ // This creates more short live intervals.
+ unsigned LDist = closestSucc(left);
+ unsigned RDist = closestSucc(right);
+ if (LDist != RDist)
+ return LDist < RDist;
+
+ // How many registers becomes live when the node is scheduled.
+ unsigned LScratch = calcMaxScratches(left);
+ unsigned RScratch = calcMaxScratches(right);
+ if (LScratch != RScratch)
+ return LScratch > RScratch;
+
+ if (!DisableSchedCycles) {
+ int result = BUCompareLatency(left, right, false /*checkPref*/, SPQ);
+ if (result != 0)
+ return result > 0;
+ }
+ else {
+ if (left->getHeight() != right->getHeight())
+ return left->getHeight() > right->getHeight();
+
+ if (left->getDepth() != right->getDepth())
+ return left->getDepth() < right->getDepth();
+ }
+
+ assert(left->NodeQueueId && right->NodeQueueId &&
+ "NodeQueueId cannot be zero");
+ return (left->NodeQueueId > right->NodeQueueId);
+}
+
+// Bottom up
+bool bu_ls_rr_sort::operator()(SUnit *left, SUnit *right) const {
+ return BURRSort(left, right, SPQ);
+}
+
+// Source order, otherwise bottom up.
+bool src_ls_rr_sort::operator()(SUnit *left, SUnit *right) const {
+ unsigned LOrder = SPQ->getNodeOrdering(left);
+ unsigned ROrder = SPQ->getNodeOrdering(right);
+
+ // Prefer an ordering where the lower the non-zero order number, the higher
+ // the preference.
+ if ((LOrder || ROrder) && LOrder != ROrder)
+ return LOrder != 0 && (LOrder < ROrder || ROrder == 0);
+
+ return BURRSort(left, right, SPQ);
+}
+
+// If the time between now and when the instruction will be ready can cover
+// the spill code, then avoid adding it to the ready queue. This gives long
+// stalls highest priority and allows hoisting across calls. It should also
+// speed up processing the available queue.
+bool hybrid_ls_rr_sort::isReady(SUnit *SU, unsigned CurCycle) const {
+ static const unsigned ReadyDelay = 3;
+
+ if (SPQ->MayReduceRegPressure(SU)) return true;
+
+ if (SU->getHeight() > (CurCycle + ReadyDelay)) return false;
+
+ if (SPQ->getHazardRec()->getHazardType(SU, -ReadyDelay)
+ != ScheduleHazardRecognizer::NoHazard)
+ return false;
+
+ return true;
+}
+
+// Return true if right should be scheduled with higher priority than left.
+bool hybrid_ls_rr_sort::operator()(SUnit *left, SUnit *right) const {
+ if (left->isCall || right->isCall)
+ // No way to compute latency of calls.
+ return BURRSort(left, right, SPQ);
+
+ bool LHigh = SPQ->HighRegPressure(left);
+ bool RHigh = SPQ->HighRegPressure(right);
+ // Avoid causing spills. If register pressure is high, schedule for
+ // register pressure reduction.
+ if (LHigh && !RHigh) {
+ DEBUG(dbgs() << " pressure SU(" << left->NodeNum << ") > SU("
+ << right->NodeNum << ")\n");
return true;
- else if (LPriority == RPriority) {
- // Try schedule def + use closer when Sethi-Ullman numbers are the same.
- // e.g.
- // t1 = op t2, c1
- // t3 = op t4, c2
- //
- // and the following instructions are both ready.
- // t2 = op c3
- // t4 = op c4
- //
- // Then schedule t2 = op first.
- // i.e.
- // t4 = op c4
- // t2 = op c3
- // t1 = op t2, c1
- // t3 = op t4, c2
- //
- // This creates more short live intervals.
- unsigned LDist = closestSucc(left);
- unsigned RDist = closestSucc(right);
- if (LDist < RDist)
- return true;
- else if (LDist == RDist) {
- // Intuitively, it's good to push down instructions whose results are
- // liveout so their long live ranges won't conflict with other values
- // which are needed inside the BB. Further prioritize liveout instructions
- // by the number of operands which are calculated within the BB.
- unsigned LScratch = calcMaxScratches(left);
- unsigned RScratch = calcMaxScratches(right);
- if (LScratch > RScratch)
- return true;
- else if (LScratch == RScratch)
- if (left->Height > right->Height)
- return true;
- else if (left->Height == right->Height)
- if (left->Depth < right->Depth)
- return true;
- else if (left->Depth == right->Depth)
- if (left->CycleBound > right->CycleBound)
- return true;
+ }
+ else if (!LHigh && RHigh) {
+ DEBUG(dbgs() << " pressure SU(" << right->NodeNum << ") > SU("
+ << left->NodeNum << ")\n");
+ return false;
+ }
+ else if (!LHigh && !RHigh) {
+ int result = BUCompareLatency(left, right, true /*checkPref*/, SPQ);
+ if (result != 0)
+ return result > 0;
+ }
+ return BURRSort(left, right, SPQ);
+}
+
+// Schedule as many instructions in each cycle as possible. So don't make an
+// instruction available unless it is ready in the current cycle.
+bool ilp_ls_rr_sort::isReady(SUnit *SU, unsigned CurCycle) const {
+ if (SU->getHeight() > CurCycle) return false;
+
+ if (SPQ->getHazardRec()->getHazardType(SU, 0)
+ != ScheduleHazardRecognizer::NoHazard)
+ return false;
+
+ return true;
+}
+
+// list-ilp is currently an experimental scheduler that allows various
+// heuristics to be enabled prior to the normal register reduction logic.
+bool ilp_ls_rr_sort::operator()(SUnit *left, SUnit *right) const {
+ if (left->isCall || right->isCall)
+ // No way to compute latency of calls.
+ return BURRSort(left, right, SPQ);
+
+ unsigned LLiveUses, RLiveUses;
+ int LPDiff = SPQ->RegPressureDiff(left, LLiveUses);
+ int RPDiff = SPQ->RegPressureDiff(right, RLiveUses);
+ if (!DisableSchedRegPressure && LPDiff != RPDiff) {
+ DEBUG(++FactorCount[FactPressureDiff]);
+ return LPDiff > RPDiff;
+ }
+
+ if (!DisableSchedLiveUses && LLiveUses != RLiveUses) {
+ DEBUG(dbgs() << "Live uses " << left->NodeNum << " = " << LLiveUses
+ << " != " << right->NodeNum << " = " << RLiveUses << "\n");
+ DEBUG(++FactorCount[FactRegUses]);
+ return LLiveUses < RLiveUses;
+ }
+
+ bool LStall = BUHasStall(left, left->getHeight(), SPQ);
+ bool RStall = BUHasStall(right, right->getHeight(), SPQ);
+ if (!DisableSchedStalls && LStall != RStall) {
+ DEBUG(++FactorCount[FactHeight]);
+ return left->getHeight() > right->getHeight();
+ }
+
+ if (!DisableSchedCriticalPath) {
+ int spread = (int)left->getDepth() - (int)right->getDepth();
+ if (std::abs(spread) > MaxReorderWindow) {
+ DEBUG(++FactorCount[FactDepth]);
+ return left->getDepth() < right->getDepth();
}
}
- return false;
+
+ if (!DisableSchedHeight && left->getHeight() != right->getHeight()) {
+ DEBUG(++FactorCount[FactHeight]);
+ return left->getHeight() > right->getHeight();
+ }
+
+ return BURRSort(left, right, SPQ);
}
-template<class SF>
-bool BURegReductionPriorityQueue<SF>::canClobber(SUnit *SU, SUnit *Op) {
+//===----------------------------------------------------------------------===//
+// Preschedule for Register Pressure
+//===----------------------------------------------------------------------===//
+
+bool RegReductionPQBase::canClobber(const SUnit *SU, const SUnit *Op) {
if (SU->isTwoAddress) {
- unsigned Opc = SU->Node->getTargetOpcode();
- unsigned NumRes = TII->getNumDefs(Opc);
- unsigned NumOps = ScheduleDAG::CountOperands(SU->Node);
+ unsigned Opc = SU->getNode()->getMachineOpcode();
+ const TargetInstrDesc &TID = TII->get(Opc);
+ unsigned NumRes = TID.getNumDefs();
+ unsigned NumOps = TID.getNumOperands() - NumRes;
for (unsigned i = 0; i != NumOps; ++i) {
- if (TII->getOperandConstraint(Opc, i+NumRes, TOI::TIED_TO) != -1) {
- SDNode *DU = SU->Node->getOperand(i).Val;
- if ((*SUnitMap).find(DU) != (*SUnitMap).end() &&
- Op == (*SUnitMap)[DU][SU->InstanceNo])
+ if (TID.getOperandConstraint(i+NumRes, TOI::TIED_TO) != -1) {
+ SDNode *DU = SU->getNode()->getOperand(i).getNode();
+ if (DU->getNodeId() != -1 &&
+ Op->OrigNode == &(*SUnits)[DU->getNodeId()])
return true;
}
}
return false;
}
-
-/// hasCopyToRegUse - Return true if SU has a value successor that is a
-/// CopyToReg node.
-static bool hasCopyToRegUse(SUnit *SU) {
- for (SUnit::const_succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
- I != E; ++I) {
- if (I->isCtrl) continue;
- SUnit *SuccSU = I->Dep;
- if (SuccSU->Node && SuccSU->Node->getOpcode() == ISD::CopyToReg)
- return true;
+/// canClobberPhysRegDefs - True if SU would clobber one of SuccSU's
+/// physical register defs.
+static bool canClobberPhysRegDefs(const SUnit *SuccSU, const SUnit *SU,
+ const TargetInstrInfo *TII,
+ const TargetRegisterInfo *TRI) {
+ SDNode *N = SuccSU->getNode();
+ unsigned NumDefs = TII->get(N->getMachineOpcode()).getNumDefs();
+ const unsigned *ImpDefs = TII->get(N->getMachineOpcode()).getImplicitDefs();
+ assert(ImpDefs && "Caller should check hasPhysRegDefs");
+ for (const SDNode *SUNode = SU->getNode(); SUNode;
+ SUNode = SUNode->getGluedNode()) {
+ if (!SUNode->isMachineOpcode())
+ continue;
+ const unsigned *SUImpDefs =
+ TII->get(SUNode->getMachineOpcode()).getImplicitDefs();
+ if (!SUImpDefs)
+ return false;
+ for (unsigned i = NumDefs, e = N->getNumValues(); i != e; ++i) {
+ EVT VT = N->getValueType(i);
+ if (VT == MVT::Glue || VT == MVT::Other)
+ continue;
+ if (!N->hasAnyUseOfValue(i))
+ continue;
+ unsigned Reg = ImpDefs[i - NumDefs];
+ for (;*SUImpDefs; ++SUImpDefs) {
+ unsigned SUReg = *SUImpDefs;
+ if (TRI->regsOverlap(Reg, SUReg))
+ return true;
+ }
+ }
}
return false;
}
-/// canClobberPhysRegDefs - True if SU would clobber one of SuccSU's
-/// physical register def.
-static bool canClobberPhysRegDefs(SUnit *SuccSU, SUnit *SU,
- const TargetInstrInfo *TII,
- const MRegisterInfo *MRI) {
- SDNode *N = SuccSU->Node;
- unsigned NumDefs = TII->getNumDefs(N->getTargetOpcode());
- const unsigned *ImpDefs = TII->getImplicitDefs(N->getTargetOpcode());
- if (!ImpDefs)
- return false;
- const unsigned *SUImpDefs = TII->getImplicitDefs(SU->Node->getTargetOpcode());
- if (!SUImpDefs)
- return false;
- for (unsigned i = NumDefs, e = N->getNumValues(); i != e; ++i) {
- MVT::ValueType VT = N->getValueType(i);
- if (VT == MVT::Flag || VT == MVT::Other)
+/// PrescheduleNodesWithMultipleUses - Nodes with multiple uses
+/// are not handled well by the general register pressure reduction
+/// heuristics. When presented with code like this:
+///
+/// N
+/// / |
+/// / |
+/// U store
+/// |
+/// ...
+///
+/// the heuristics tend to push the store up, but since the
+/// operand of the store has another use (U), this would increase
+/// the length of that other use (the U->N edge).
+///
+/// This function transforms code like the above to route U's
+/// dependence through the store when possible, like this:
+///
+/// N
+/// ||
+/// ||
+/// store
+/// |
+/// U
+/// |
+/// ...
+///
+/// This results in the store being scheduled immediately
+/// after N, which shortens the U->N live range, reducing
+/// register pressure.
+///
+void RegReductionPQBase::PrescheduleNodesWithMultipleUses() {
+ // Visit all the nodes in topological order, working top-down.
+ for (unsigned i = 0, e = SUnits->size(); i != e; ++i) {
+ SUnit *SU = &(*SUnits)[i];
+ // For now, only look at nodes with no data successors, such as stores.
+ // These are especially important, due to the heuristics in
+ // getNodePriority for nodes with no data successors.
+ if (SU->NumSuccs != 0)
continue;
- unsigned Reg = ImpDefs[i - NumDefs];
- for (;*SUImpDefs; ++SUImpDefs) {
- unsigned SUReg = *SUImpDefs;
- if (MRI->regsOverlap(Reg, SUReg))
- return true;
+ // For now, only look at nodes with exactly one data predecessor.
+ if (SU->NumPreds != 1)
+ continue;
+ // Avoid prescheduling copies to virtual registers, which don't behave
+ // like other nodes from the perspective of scheduling heuristics.
+ if (SDNode *N = SU->getNode())
+ if (N->getOpcode() == ISD::CopyToReg &&
+ TargetRegisterInfo::isVirtualRegister
+ (cast<RegisterSDNode>(N->getOperand(1))->getReg()))
+ continue;
+
+ // Locate the single data predecessor.
+ SUnit *PredSU = 0;
+ for (SUnit::const_pred_iterator II = SU->Preds.begin(),
+ EE = SU->Preds.end(); II != EE; ++II)
+ if (!II->isCtrl()) {
+ PredSU = II->getSUnit();
+ break;
+ }
+ assert(PredSU);
+
+ // Don't rewrite edges that carry physregs, because that requires additional
+ // support infrastructure.
+ if (PredSU->hasPhysRegDefs)
+ continue;
+ // Short-circuit the case where SU is PredSU's only data successor.
+ if (PredSU->NumSuccs == 1)
+ continue;
+ // Avoid prescheduling to copies from virtual registers, which don't behave
+ // like other nodes from the perspective of scheduling heuristics.
+ if (SDNode *N = SU->getNode())
+ if (N->getOpcode() == ISD::CopyFromReg &&
+ TargetRegisterInfo::isVirtualRegister
+ (cast<RegisterSDNode>(N->getOperand(1))->getReg()))
+ continue;
+
+ // Perform checks on the successors of PredSU.
+ for (SUnit::const_succ_iterator II = PredSU->Succs.begin(),
+ EE = PredSU->Succs.end(); II != EE; ++II) {
+ SUnit *PredSuccSU = II->getSUnit();
+ if (PredSuccSU == SU) continue;
+ // If PredSU has another successor with no data successors, for
+ // now don't attempt to choose either over the other.
+ if (PredSuccSU->NumSuccs == 0)
+ goto outer_loop_continue;
+ // Don't break physical register dependencies.
+ if (SU->hasPhysRegClobbers && PredSuccSU->hasPhysRegDefs)
+ if (canClobberPhysRegDefs(PredSuccSU, SU, TII, TRI))
+ goto outer_loop_continue;
+ // Don't introduce graph cycles.
+ if (scheduleDAG->IsReachable(SU, PredSuccSU))
+ goto outer_loop_continue;
}
+
+ // Ok, the transformation is safe and the heuristics suggest it is
+ // profitable. Update the graph.
+ DEBUG(dbgs() << " Prescheduling SU #" << SU->NodeNum
+ << " next to PredSU #" << PredSU->NodeNum
+ << " to guide scheduling in the presence of multiple uses\n");
+ for (unsigned i = 0; i != PredSU->Succs.size(); ++i) {
+ SDep Edge = PredSU->Succs[i];
+ assert(!Edge.isAssignedRegDep());
+ SUnit *SuccSU = Edge.getSUnit();
+ if (SuccSU != SU) {
+ Edge.setSUnit(PredSU);
+ scheduleDAG->RemovePred(SuccSU, Edge);
+ scheduleDAG->AddPred(SU, Edge);
+ Edge.setSUnit(SU);
+ scheduleDAG->AddPred(SuccSU, Edge);
+ --i;
+ }
+ }
+ outer_loop_continue:;
}
- return false;
}
/// AddPseudoTwoAddrDeps - If two nodes share an operand and one of them uses
/// one that has a CopyToReg use (more likely to be a loop induction update).
/// If both are two-address, but one is commutable while the other is not
/// commutable, favor the one that's not commutable.
-template<class SF>
-void BURegReductionPriorityQueue<SF>::AddPseudoTwoAddrDeps() {
+void RegReductionPQBase::AddPseudoTwoAddrDeps() {
for (unsigned i = 0, e = SUnits->size(); i != e; ++i) {
- SUnit *SU = (SUnit *)&((*SUnits)[i]);
+ SUnit *SU = &(*SUnits)[i];
if (!SU->isTwoAddress)
continue;
- SDNode *Node = SU->Node;
- if (!Node || !Node->isTargetOpcode() || SU->FlaggedNodes.size() > 0)
+ SDNode *Node = SU->getNode();
+ if (!Node || !Node->isMachineOpcode() || SU->getNode()->getGluedNode())
continue;
- unsigned Opc = Node->getTargetOpcode();
- unsigned NumRes = TII->getNumDefs(Opc);
- unsigned NumOps = ScheduleDAG::CountOperands(Node);
+ bool isLiveOut = hasOnlyLiveOutUses(SU);
+ unsigned Opc = Node->getMachineOpcode();
+ const TargetInstrDesc &TID = TII->get(Opc);
+ unsigned NumRes = TID.getNumDefs();
+ unsigned NumOps = TID.getNumOperands() - NumRes;
for (unsigned j = 0; j != NumOps; ++j) {
- if (TII->getOperandConstraint(Opc, j+NumRes, TOI::TIED_TO) != -1) {
- SDNode *DU = SU->Node->getOperand(j).Val;
- if ((*SUnitMap).find(DU) == (*SUnitMap).end())
+ if (TID.getOperandConstraint(j+NumRes, TOI::TIED_TO) == -1)
+ continue;
+ SDNode *DU = SU->getNode()->getOperand(j).getNode();
+ if (DU->getNodeId() == -1)
+ continue;
+ const SUnit *DUSU = &(*SUnits)[DU->getNodeId()];
+ if (!DUSU) continue;
+ for (SUnit::const_succ_iterator I = DUSU->Succs.begin(),
+ E = DUSU->Succs.end(); I != E; ++I) {
+ if (I->isCtrl()) continue;
+ SUnit *SuccSU = I->getSUnit();
+ if (SuccSU == SU)
continue;
- SUnit *DUSU = (*SUnitMap)[DU][SU->InstanceNo];
- if (!DUSU) continue;
- for (SUnit::succ_iterator I = DUSU->Succs.begin(),E = DUSU->Succs.end();
- I != E; ++I) {
- if (I->isCtrl) continue;
- SUnit *SuccSU = I->Dep;
- if (SuccSU == SU)
- continue;
- // Be conservative. Ignore if nodes aren't at roughly the same
- // depth and height.
- if (SuccSU->Height < SU->Height && (SU->Height - SuccSU->Height) > 1)
- continue;
- if (!SuccSU->Node || !SuccSU->Node->isTargetOpcode())
- continue;
- // Don't constrain nodes with physical register defs if the
- // predecessor can cloober them.
- if (SuccSU->hasPhysRegDefs) {
- if (canClobberPhysRegDefs(SuccSU, SU, TII, MRI))
- continue;
- }
- // Don't constraint extract_subreg / insert_subreg these may be
- // coalesced away. We don't them close to their uses.
- unsigned SuccOpc = SuccSU->Node->getTargetOpcode();
- if (SuccOpc == TargetInstrInfo::EXTRACT_SUBREG ||
- SuccOpc == TargetInstrInfo::INSERT_SUBREG)
+ // Be conservative. Ignore if nodes aren't at roughly the same
+ // depth and height.
+ if (SuccSU->getHeight() < SU->getHeight() &&
+ (SU->getHeight() - SuccSU->getHeight()) > 1)
+ continue;
+ // Skip past COPY_TO_REGCLASS nodes, so that the pseudo edge
+ // constrains whatever is using the copy, instead of the copy
+ // itself. In the case that the copy is coalesced, this
+ // preserves the intent of the pseudo two-address heurietics.
+ while (SuccSU->Succs.size() == 1 &&
+ SuccSU->getNode()->isMachineOpcode() &&
+ SuccSU->getNode()->getMachineOpcode() ==
+ TargetOpcode::COPY_TO_REGCLASS)
+ SuccSU = SuccSU->Succs.front().getSUnit();
+ // Don't constrain non-instruction nodes.
+ if (!SuccSU->getNode() || !SuccSU->getNode()->isMachineOpcode())
+ continue;
+ // Don't constrain nodes with physical register defs if the
+ // predecessor can clobber them.
+ if (SuccSU->hasPhysRegDefs && SU->hasPhysRegClobbers) {
+ if (canClobberPhysRegDefs(SuccSU, SU, TII, TRI))
continue;
- if ((!canClobber(SuccSU, DUSU) ||
- (hasCopyToRegUse(SU) && !hasCopyToRegUse(SuccSU)) ||
- (!SU->isCommutable && SuccSU->isCommutable)) &&
- !isReachable(SuccSU, SU)) {
- DOUT << "Adding an edge from SU # " << SU->NodeNum
- << " to SU #" << SuccSU->NodeNum << "\n";
- SU->addPred(SuccSU, true, true);
- }
+ }
+ // Don't constrain EXTRACT_SUBREG, INSERT_SUBREG, and SUBREG_TO_REG;
+ // these may be coalesced away. We want them close to their uses.
+ unsigned SuccOpc = SuccSU->getNode()->getMachineOpcode();
+ if (SuccOpc == TargetOpcode::EXTRACT_SUBREG ||
+ SuccOpc == TargetOpcode::INSERT_SUBREG ||
+ SuccOpc == TargetOpcode::SUBREG_TO_REG)
+ continue;
+ if ((!canClobber(SuccSU, DUSU) ||
+ (isLiveOut && !hasOnlyLiveOutUses(SuccSU)) ||
+ (!SU->isCommutable && SuccSU->isCommutable)) &&
+ !scheduleDAG->IsReachable(SuccSU, SU)) {
+ DEBUG(dbgs() << " Adding a pseudo-two-addr edge from SU #"
+ << SU->NodeNum << " to SU #" << SuccSU->NodeNum << "\n");
+ scheduleDAG->AddPred(SU, SDep(SuccSU, SDep::Order, /*Latency=*/0,
+ /*Reg=*/0, /*isNormalMemory=*/false,
+ /*isMustAlias=*/false,
+ /*isArtificial=*/true));
}
}
}
}
}
-/// CalcNodeSethiUllmanNumber - Priority is the Sethi Ullman number.
-/// Smaller number is the higher priority.
-template<class SF>
-unsigned BURegReductionPriorityQueue<SF>::
-CalcNodeSethiUllmanNumber(const SUnit *SU) {
- unsigned &SethiUllmanNumber = SethiUllmanNumbers[SU->NodeNum];
- if (SethiUllmanNumber != 0)
- return SethiUllmanNumber;
-
- unsigned Extra = 0;
- for (SUnit::const_pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
- I != E; ++I) {
- if (I->isCtrl) continue; // ignore chain preds
- SUnit *PredSU = I->Dep;
- unsigned PredSethiUllman = CalcNodeSethiUllmanNumber(PredSU);
- if (PredSethiUllman > SethiUllmanNumber) {
- SethiUllmanNumber = PredSethiUllman;
- Extra = 0;
- } else if (PredSethiUllman == SethiUllmanNumber && !I->isCtrl)
- ++Extra;
- }
-
- SethiUllmanNumber += Extra;
-
- if (SethiUllmanNumber == 0)
- SethiUllmanNumber = 1;
-
- return SethiUllmanNumber;
-}
-
-/// CalculateSethiUllmanNumbers - Calculate Sethi-Ullman numbers of all
-/// scheduling units.
-template<class SF>
-void BURegReductionPriorityQueue<SF>::CalculateSethiUllmanNumbers() {
- SethiUllmanNumbers.assign(SUnits->size(), 0);
-
- for (unsigned i = 0, e = SUnits->size(); i != e; ++i)
- CalcNodeSethiUllmanNumber(&(*SUnits)[i]);
-}
-
-static unsigned SumOfUnscheduledPredsOfSuccs(const SUnit *SU) {
+/// LimitedSumOfUnscheduledPredsOfSuccs - Compute the sum of the unscheduled
+/// predecessors of the successors of the SUnit SU. Stop when the provided
+/// limit is exceeded.
+static unsigned LimitedSumOfUnscheduledPredsOfSuccs(const SUnit *SU,
+ unsigned Limit) {
unsigned Sum = 0;
for (SUnit::const_succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
I != E; ++I) {
- SUnit *SuccSU = I->Dep;
+ const SUnit *SuccSU = I->getSUnit();
for (SUnit::const_pred_iterator II = SuccSU->Preds.begin(),
EE = SuccSU->Preds.end(); II != EE; ++II) {
- SUnit *PredSU = II->Dep;
+ SUnit *PredSU = II->getSUnit();
if (!PredSU->isScheduled)
- ++Sum;
+ if (++Sum > Limit)
+ return Sum;
}
}
-
return Sum;
}
bool td_ls_rr_sort::operator()(const SUnit *left, const SUnit *right) const {
unsigned LPriority = SPQ->getNodePriority(left);
unsigned RPriority = SPQ->getNodePriority(right);
- bool LIsTarget = left->Node && left->Node->isTargetOpcode();
- bool RIsTarget = right->Node && right->Node->isTargetOpcode();
+ bool LIsTarget = left->getNode() && left->getNode()->isMachineOpcode();
+ bool RIsTarget = right->getNode() && right->getNode()->isMachineOpcode();
bool LIsFloater = LIsTarget && left->NumPreds == 0;
bool RIsFloater = RIsTarget && right->NumPreds == 0;
- unsigned LBonus = (SumOfUnscheduledPredsOfSuccs(left) == 1) ? 2 : 0;
- unsigned RBonus = (SumOfUnscheduledPredsOfSuccs(right) == 1) ? 2 : 0;
+ unsigned LBonus = (LimitedSumOfUnscheduledPredsOfSuccs(left,1) == 1) ? 2 : 0;
+ unsigned RBonus = (LimitedSumOfUnscheduledPredsOfSuccs(right,1) == 1) ? 2 : 0;
if (left->NumSuccs == 0 && right->NumSuccs != 0)
return false;
else if (left->NumSuccs != 0 && right->NumSuccs == 0)
return true;
- // Special tie breaker: if two nodes share a operand, the one that use it
- // as a def&use operand is preferred.
- if (LIsTarget && RIsTarget) {
- if (left->isTwoAddress && !right->isTwoAddress) {
- SDNode *DUNode = left->Node->getOperand(0).Val;
- if (DUNode->isOperand(right->Node))
- RBonus += 2;
- }
- if (!left->isTwoAddress && right->isTwoAddress) {
- SDNode *DUNode = right->Node->getOperand(0).Val;
- if (DUNode->isOperand(left->Node))
- LBonus += 2;
- }
- }
if (LIsFloater)
LBonus -= 2;
if (RIsFloater)
if (right->NumSuccs == 1)
RBonus += 2;
- if (LPriority+LBonus < RPriority+RBonus)
- return true;
- else if (LPriority == RPriority)
- if (left->Depth < right->Depth)
- return true;
- else if (left->Depth == right->Depth)
- if (left->NumSuccsLeft > right->NumSuccsLeft)
- return true;
- else if (left->NumSuccsLeft == right->NumSuccsLeft)
- if (left->CycleBound > right->CycleBound)
- return true;
- return false;
-}
+ if (LPriority+LBonus != RPriority+RBonus)
+ return LPriority+LBonus < RPriority+RBonus;
-/// CalcNodeSethiUllmanNumber - Priority is the Sethi Ullman number.
-/// Smaller number is the higher priority.
-template<class SF>
-unsigned TDRegReductionPriorityQueue<SF>::
-CalcNodeSethiUllmanNumber(const SUnit *SU) {
- unsigned &SethiUllmanNumber = SethiUllmanNumbers[SU->NodeNum];
- if (SethiUllmanNumber != 0)
- return SethiUllmanNumber;
+ if (left->getDepth() != right->getDepth())
+ return left->getDepth() < right->getDepth();
- unsigned Opc = SU->Node ? SU->Node->getOpcode() : 0;
- if (Opc == ISD::TokenFactor || Opc == ISD::CopyToReg)
- SethiUllmanNumber = 0xffff;
- else if (SU->NumSuccsLeft == 0)
- // If SU does not have a use, i.e. it doesn't produce a value that would
- // be consumed (e.g. store), then it terminates a chain of computation.
- // Give it a small SethiUllman number so it will be scheduled right before
- // its predecessors that it doesn't lengthen their live ranges.
- SethiUllmanNumber = 0;
- else if (SU->NumPredsLeft == 0 &&
- (Opc != ISD::CopyFromReg || isCopyFromLiveIn(SU)))
- SethiUllmanNumber = 0xffff;
- else {
- int Extra = 0;
- for (SUnit::const_pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
- I != E; ++I) {
- if (I->isCtrl) continue; // ignore chain preds
- SUnit *PredSU = I->Dep;
- unsigned PredSethiUllman = CalcNodeSethiUllmanNumber(PredSU);
- if (PredSethiUllman > SethiUllmanNumber) {
- SethiUllmanNumber = PredSethiUllman;
- Extra = 0;
- } else if (PredSethiUllman == SethiUllmanNumber && !I->isCtrl)
- ++Extra;
- }
-
- SethiUllmanNumber += Extra;
- }
-
- return SethiUllmanNumber;
-}
+ if (left->NumSuccsLeft != right->NumSuccsLeft)
+ return left->NumSuccsLeft > right->NumSuccsLeft;
-/// CalculateSethiUllmanNumbers - Calculate Sethi-Ullman numbers of all
-/// scheduling units.
-template<class SF>
-void TDRegReductionPriorityQueue<SF>::CalculateSethiUllmanNumbers() {
- SethiUllmanNumbers.assign(SUnits->size(), 0);
-
- for (unsigned i = 0, e = SUnits->size(); i != e; ++i)
- CalcNodeSethiUllmanNumber(&(*SUnits)[i]);
+ assert(left->NodeQueueId && right->NodeQueueId &&
+ "NodeQueueId cannot be zero");
+ return (left->NodeQueueId > right->NodeQueueId);
}
//===----------------------------------------------------------------------===//
// Public Constructor Functions
//===----------------------------------------------------------------------===//
-llvm::ScheduleDAG* llvm::createBURRListDAGScheduler(SelectionDAGISel *IS,
- SelectionDAG *DAG,
- MachineBasicBlock *BB) {
- const TargetInstrInfo *TII = DAG->getTarget().getInstrInfo();
- const MRegisterInfo *MRI = DAG->getTarget().getRegisterInfo();
- return new ScheduleDAGRRList(*DAG, BB, DAG->getTarget(), true,
- new BURegReductionPriorityQueue<bu_ls_rr_sort>(TII, MRI));
+llvm::ScheduleDAGSDNodes *
+llvm::createBURRListDAGScheduler(SelectionDAGISel *IS,
+ CodeGenOpt::Level OptLevel) {
+ const TargetMachine &TM = IS->TM;
+ const TargetInstrInfo *TII = TM.getInstrInfo();
+ const TargetRegisterInfo *TRI = TM.getRegisterInfo();
+
+ BURegReductionPriorityQueue *PQ =
+ new BURegReductionPriorityQueue(*IS->MF, false, TII, TRI, 0);
+ ScheduleDAGRRList *SD = new ScheduleDAGRRList(*IS->MF, false, PQ, OptLevel);
+ PQ->setScheduleDAG(SD);
+ return SD;
}
-llvm::ScheduleDAG* llvm::createTDRRListDAGScheduler(SelectionDAGISel *IS,
- SelectionDAG *DAG,
- MachineBasicBlock *BB) {
- return new ScheduleDAGRRList(*DAG, BB, DAG->getTarget(), false,
- new TDRegReductionPriorityQueue<td_ls_rr_sort>());
+llvm::ScheduleDAGSDNodes *
+llvm::createTDRRListDAGScheduler(SelectionDAGISel *IS,
+ CodeGenOpt::Level OptLevel) {
+ const TargetMachine &TM = IS->TM;
+ const TargetInstrInfo *TII = TM.getInstrInfo();
+ const TargetRegisterInfo *TRI = TM.getRegisterInfo();
+
+ TDRegReductionPriorityQueue *PQ =
+ new TDRegReductionPriorityQueue(*IS->MF, false, TII, TRI, 0);
+ ScheduleDAGRRList *SD = new ScheduleDAGRRList(*IS->MF, false, PQ, OptLevel);
+ PQ->setScheduleDAG(SD);
+ return SD;
}
+llvm::ScheduleDAGSDNodes *
+llvm::createSourceListDAGScheduler(SelectionDAGISel *IS,
+ CodeGenOpt::Level OptLevel) {
+ const TargetMachine &TM = IS->TM;
+ const TargetInstrInfo *TII = TM.getInstrInfo();
+ const TargetRegisterInfo *TRI = TM.getRegisterInfo();
+
+ SrcRegReductionPriorityQueue *PQ =
+ new SrcRegReductionPriorityQueue(*IS->MF, false, TII, TRI, 0);
+ ScheduleDAGRRList *SD = new ScheduleDAGRRList(*IS->MF, false, PQ, OptLevel);
+ PQ->setScheduleDAG(SD);
+ return SD;
+}
+
+llvm::ScheduleDAGSDNodes *
+llvm::createHybridListDAGScheduler(SelectionDAGISel *IS,
+ CodeGenOpt::Level OptLevel) {
+ const TargetMachine &TM = IS->TM;
+ const TargetInstrInfo *TII = TM.getInstrInfo();
+ const TargetRegisterInfo *TRI = TM.getRegisterInfo();
+ const TargetLowering *TLI = &IS->getTargetLowering();
+
+ HybridBURRPriorityQueue *PQ =
+ new HybridBURRPriorityQueue(*IS->MF, true, TII, TRI, TLI);
+
+ ScheduleDAGRRList *SD = new ScheduleDAGRRList(*IS->MF, true, PQ, OptLevel);
+ PQ->setScheduleDAG(SD);
+ return SD;
+}
+
+llvm::ScheduleDAGSDNodes *
+llvm::createILPListDAGScheduler(SelectionDAGISel *IS,
+ CodeGenOpt::Level OptLevel) {
+ const TargetMachine &TM = IS->TM;
+ const TargetInstrInfo *TII = TM.getInstrInfo();
+ const TargetRegisterInfo *TRI = TM.getRegisterInfo();
+ const TargetLowering *TLI = &IS->getTargetLowering();
+
+ ILPBURRPriorityQueue *PQ =
+ new ILPBURRPriorityQueue(*IS->MF, true, TII, TRI, TLI);
+ ScheduleDAGRRList *SD = new ScheduleDAGRRList(*IS->MF, true, PQ, OptLevel);
+ PQ->setScheduleDAG(SD);
+ return SD;
+}
projects / oota-llvm.git / blobdiff