//===----------------------------------------------------------------------===//
#define DEBUG_TYPE "pre-RA-sched"
-#include "llvm/CodeGen/ScheduleDAGSDNodes.h"
+#include "ScheduleDAGSDNodes.h"
#include "llvm/CodeGen/SchedulerRegistry.h"
+#include "llvm/CodeGen/SelectionDAGISel.h"
#include "llvm/Target/TargetRegisterInfo.h"
#include "llvm/Target/TargetData.h"
#include "llvm/Target/TargetMachine.h"
#include "llvm/ADT/Statistic.h"
#include "llvm/ADT/STLExtras.h"
#include <climits>
-#include "llvm/Support/CommandLine.h"
using namespace llvm;
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",
ScheduleDAGTopologicalSort Topo;
public:
- ScheduleDAGRRList(SelectionDAG *dag, MachineBasicBlock *bb,
- const TargetMachine &tm, bool isbottomup,
+ ScheduleDAGRRList(MachineFunction &mf,
+ bool isbottomup,
SchedulingPriorityQueue *availqueue)
- : ScheduleDAGSDNodes(dag, bb, tm), isBottomUp(isbottomup),
+ : ScheduleDAGSDNodes(mf), isBottomUp(isbottomup),
AvailableQueue(availqueue), Topo(SUnits) {
}
return Topo.IsReachable(SU, TargetSU);
}
- /// willCreateCycle - Returns true if adding an edge from SU to TargetSU will
+ /// 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);
}
private:
- void ReleasePred(SUnit *SU, SDep *PredEdge);
- void ReleaseSucc(SUnit *SU, SDep *SuccEdge);
+ void ReleasePred(SUnit *SU, const SDep *PredEdge);
+ void ReleasePredecessors(SUnit *SU, unsigned CurCycle);
+ void ReleaseSucc(SUnit *SU, const SDep *SuccEdge);
+ void ReleaseSuccessors(SUnit *SU);
void CapturePred(SDep *PredEdge);
void ScheduleNodeBottomUp(SUnit*, unsigned);
void ScheduleNodeTopDown(SUnit*, unsigned);
void UnscheduleNodeBottomUp(SUnit*);
void BacktrackBottomUp(SUnit*, unsigned, unsigned&);
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();
void ListScheduleBottomUp();
/// 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 *SU, SDep *PredEdge) {
+void ScheduleDAGRRList::ReleasePred(SUnit *SU, const SDep *PredEdge) {
SUnit *PredSU = PredEdge->getSUnit();
--PredSU->NumSuccsLeft;
}
#endif
- if (PredSU->NumSuccsLeft == 0) {
+ // 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;
AvailableQueue->push(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(this));
-
- assert(CurCycle >= SU->getHeight() && "Node scheduled below its height!");
- SU->setHeightToAtLeast(CurCycle);
- Sequence.push_back(SU);
-
+void ScheduleDAGRRList::ReleasePredecessors(SUnit *SU, unsigned CurCycle) {
// Bottom up: release predecessors
for (SUnit::pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
I != E; ++I) {
}
}
}
+}
+
+/// 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(this));
+
+ assert(CurCycle >= SU->getHeight() && "Node scheduled below its height!");
+ SU->setHeightToAtLeast(CurCycle);
+ Sequence.push_back(SU);
+
+ ReleasePredecessors(SU, CurCycle);
// Release all the implicit physical register defs that are live.
for (SUnit::succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
}
/// 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.
+/// BTCycle in order to schedule a specific node.
void ScheduleDAGRRList::BacktrackBottomUp(SUnit *SU, unsigned BtCycle,
unsigned &CurCycle) {
SUnit *OldSU = NULL;
--CurCycle;
}
-
- if (SU->isSucc(OldSU)) {
- assert(false && "Something is wrong!");
- abort();
- }
+ assert(!SU->isSucc(OldSU) && "Something is wrong!");
++NumBacktracks;
}
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 = CreateNewSUnit(NULL);
CopyFromSU->CopySrcRC = SrcRC;
DelDeps.push_back(std::make_pair(SuccSU, *I));
}
}
- for (unsigned i = 0, e = DelDeps.size(); i != e; ++i) {
+ for (unsigned i = 0, e = DelDeps.size(); i != e; ++i)
RemovePred(DelDeps[i].first, DelDeps[i].second);
- }
AddPred(CopyFromSU, SDep(SU, SDep::Data, SU->Latency, Reg));
AddPred(CopyToSU, SDep(CopyFromSU, SDep::Data, CopyFromSU->Latency, 0));
Copies.push_back(CopyFromSU);
Copies.push_back(CopyToSU);
- ++NumCCCopies;
+ ++NumPRCopies;
}
/// getPhysicalRegisterVT - Returns the ValueType of the physical register
/// schedulers.
void ScheduleDAGRRList::ListScheduleBottomUp() {
unsigned CurCycle = 0;
+
+ // Release any predecessors of the special Exit node.
+ ReleasePredecessors(&ExitSU, CurCycle);
+
// Add root to Available queue.
if (!SUnits.empty()) {
SUnit *RootSU = &SUnits[DAG->getRoot().getNode()->getNodeId()];
}
if (!CurSU) {
- // Can't backtrack. 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.
+ // 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 = 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);
+ MVT VT = getPhysicalRegisterVT(LRDef->getNode(), Reg, TII);
+ const TargetRegisterClass *RC =
+ TRI->getPhysicalRegisterRegClass(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 expensive cross register class copies.
- MVT VT = getPhysicalRegisterVT(LRDef->getNode(), Reg, TII);
- const TargetRegisterClass *RC =
- TRI->getPhysicalRegisterRegClass(Reg, VT);
- const TargetRegisterClass *DestRC = TRI->getCrossCopyRegClass(RC);
- if (!DestRC) {
- assert(false && "Don't know how to copy this physical register!");
- abort();
- }
+ // Issue copies, these can be expensive cross register class copies.
SmallVector<SUnit*, 2> Copies;
- InsertCCCopiesAndMoveSuccs(LRDef, Reg, DestRC, RC, Copies);
- DOUT << "Adding an edge from SU # " << TrySU->NodeNum
+ InsertCopiesAndMoveSuccs(LRDef, Reg, DestRC, RC, Copies);
+ DOUT << "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,
NewDef = Copies.back();
}
- DOUT << "Adding an edge from SU # " << NewDef->NodeNum
+ DOUT << "Adding an edge from SU #" << NewDef->NodeNum
<< " to SU #" << TrySU->NodeNum << "\n";
LiveRegDefs[Reg] = NewDef;
AddPred(NewDef, SDep(TrySU, SDep::Order, /*Latency=*/1,
CurSU = NewDef;
}
- if (!CurSU) {
- assert(false && "Unable to resolve live physical register dependencies!");
- abort();
- }
+ assert(CurSU && "Unable to resolve live physical register dependencies!");
}
// Add the nodes that aren't ready back onto the available list.
/// 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 *SU, SDep *SuccEdge) {
+void ScheduleDAGRRList::ReleaseSucc(SUnit *SU, const SDep *SuccEdge) {
SUnit *SuccSU = SuccEdge->getSUnit();
--SuccSU->NumPredsLeft;
}
#endif
- if (SuccSU->NumPredsLeft == 0) {
+ // 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.
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(SU, &*I);
-
+ ReleaseSuccessors(SU);
SU->isScheduled = true;
AvailableQueue->ScheduledNode(SU);
}
void ScheduleDAGRRList::ListScheduleTopDown() {
unsigned CurCycle = 0;
+ // 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.
};
} // end anonymous namespace
-static inline bool isCopyFromLiveIn(const SUnit *SU) {
- SDNode *N = SU->getNode();
- return N && N->getOpcode() == ISD::CopyFromReg &&
- N->getOperand(N->getNumOperands()-1).getValueType() != MVT::Flag;
-}
-
/// CalcNodeSethiUllmanNumber - Compute Sethi Ullman number.
/// Smaller number is the higher priority.
static unsigned
if (PredSethiUllman > SethiUllmanNumber) {
SethiUllmanNumber = PredSethiUllman;
Extra = 0;
- } else if (PredSethiUllman == SethiUllmanNumber && !I->isCtrl())
+ } else if (PredSethiUllman == SethiUllmanNumber)
++Extra;
}
unsigned getNodePriority(const SUnit *SU) const {
assert(SU->NodeNum < SethiUllmanNumbers.size());
unsigned Opc = SU->getNode() ? SU->getNode()->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;
if (Opc == ISD::TokenFactor || Opc == ISD::CopyToReg)
// CopyToReg should be close to its uses to facilitate coalescing and
// avoid spilling.
// EXTRACT_SUBREG / INSERT_SUBREG should be close to its use to
// facilitate coalescing.
return 0;
- 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.
+ 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)
- // If SU does not have a def, schedule it close to its uses because it
- // does not lengthen any live ranges.
+ 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];
}
unsigned MaxHeight = 0;
for (SUnit::const_succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
I != E; ++I) {
+ 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.
}
/// 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->getSUnit()->getNode() ||
- I->getSUnit()->getNode()->getOpcode() != ISD::CopyFromReg)
- Scratches++;
- }
- 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->getSUnit()->getNode() ||
- I->getSUnit()->getNode()->getOpcode() != ISD::CopyToReg)
- Scratches += 10;
+ Scratches++;
}
return Scratches;
}
if (LDist != RDist)
return 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.
+ // How many registers becomes live when the node is scheduled.
unsigned LScratch = calcMaxScratches(left);
unsigned RScratch = calcMaxScratches(right);
if (LScratch != RScratch)
if (canClobberPhysRegDefs(SuccSU, SU, TII, TRI))
continue;
}
- // Don't constraint extract_subreg / insert_subreg these may be
- // coalesced away. We don't them close to their uses.
+ // Don't constrain extract_subreg / insert_subreg; these may be
+ // coalesced away. We want them close to their uses.
unsigned SuccOpc = SuccSU->getNode()->getMachineOpcode();
if (SuccOpc == TargetInstrInfo::EXTRACT_SUBREG ||
SuccOpc == TargetInstrInfo::INSERT_SUBREG)
// Public Constructor Functions
//===----------------------------------------------------------------------===//
-llvm::ScheduleDAG* llvm::createBURRListDAGScheduler(SelectionDAGISel *IS,
- SelectionDAG *DAG,
- const TargetMachine *TM,
- MachineBasicBlock *BB,
- bool) {
- const TargetInstrInfo *TII = TM->getInstrInfo();
- const TargetRegisterInfo *TRI = TM->getRegisterInfo();
+llvm::ScheduleDAGSDNodes *
+llvm::createBURRListDAGScheduler(SelectionDAGISel *IS, bool) {
+ const TargetMachine &TM = IS->TM;
+ const TargetInstrInfo *TII = TM.getInstrInfo();
+ const TargetRegisterInfo *TRI = TM.getRegisterInfo();
BURegReductionPriorityQueue *PQ = new BURegReductionPriorityQueue(TII, TRI);
ScheduleDAGRRList *SD =
- new ScheduleDAGRRList(DAG, BB, *TM, true, PQ);
+ new ScheduleDAGRRList(*IS->MF, true, PQ);
PQ->setScheduleDAG(SD);
return SD;
}
-llvm::ScheduleDAG* llvm::createTDRRListDAGScheduler(SelectionDAGISel *IS,
- SelectionDAG *DAG,
- const TargetMachine *TM,
- MachineBasicBlock *BB,
- bool) {
- const TargetInstrInfo *TII = TM->getInstrInfo();
- const TargetRegisterInfo *TRI = TM->getRegisterInfo();
+llvm::ScheduleDAGSDNodes *
+llvm::createTDRRListDAGScheduler(SelectionDAGISel *IS, bool) {
+ const TargetMachine &TM = IS->TM;
+ const TargetInstrInfo *TII = TM.getInstrInfo();
+ const TargetRegisterInfo *TRI = TM.getRegisterInfo();
TDRegReductionPriorityQueue *PQ = new TDRegReductionPriorityQueue(TII, TRI);
- ScheduleDAGRRList *SD = new ScheduleDAGRRList(DAG, BB, *TM, false, PQ);
+ ScheduleDAGRRList *SD =
+ new ScheduleDAGRRList(*IS->MF, false, PQ);
PQ->setScheduleDAG(SD);
return SD;
}
projects / oota-llvm.git / blobdiff