#define DEBUG_TYPE "pre-RA-sched"
#include "ScheduleDAGSDNodes.h"
+#include "llvm/InlineAsm.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/Target/TargetInstrInfo.h"
-#include "llvm/Support/Debug.h"
-#include "llvm/Support/Compiler.h"
-#include "llvm/ADT/PriorityQueue.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>
using namespace llvm;
tdrListrDAGScheduler("list-tdrr",
"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);
namespace {
//===----------------------------------------------------------------------===//
/// ScheduleDAGRRList - The actual register reduction list scheduler
/// implementation. This supports both top-down and bottom-up scheduling.
///
-class VISIBILITY_HIDDEN ScheduleDAGRRList : public ScheduleDAGSDNodes {
+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.
SchedulingPriorityQueue *AvailableQueue;
public:
ScheduleDAGRRList(MachineFunction &mf,
- bool isbottomup,
+ bool isbottomup, bool needlatency,
SchedulingPriorityQueue *availqueue)
- : ScheduleDAGSDNodes(mf), isBottomUp(isbottomup),
+ : ScheduleDAGSDNodes(mf), isBottomUp(isbottomup), NeedLatency(needlatency),
AvailableQueue(availqueue), Topo(SUnits) {
}
return NewNode;
}
- /// ForceUnitLatencies - Return true, since register-pressure-reducing
- /// scheduling doesn't need actual latency information.
- bool ForceUnitLatencies() const { return true; }
+ /// 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");
NumLiveRegs = 0;
- LiveRegDefs.resize(TRI->getNumRegs(), NULL);
+ LiveRegDefs.resize(TRI->getNumRegs(), NULL);
LiveRegCycles.resize(TRI->getNumRegs(), 0);
// Build the scheduling graph.
- BuildSchedGraph();
+ BuildSchedGraph(NULL);
DEBUG(for (unsigned su = 0, e = SUnits.size(); su != e; ++su)
SUnits[su].dumpAll(this));
Topo.InitDAGTopologicalSorting();
AvailableQueue->initNodes(SUnits);
-
+
// Execute the actual scheduling loop Top-Down or Bottom-Up as appropriate.
if (isBottomUp)
ListScheduleBottomUp();
else
ListScheduleTopDown();
-
+
AvailableQueue->releaseState();
}
/// the AvailableQueue if the count reaches zero. Also update its cycle bound.
void ScheduleDAGRRList::ReleasePred(SUnit *SU, const SDep *PredEdge) {
SUnit *PredSU = PredEdge->getSUnit();
- --PredSU->NumSuccsLeft;
-
+
#ifndef NDEBUG
- if (PredSU->NumSuccsLeft < 0) {
- cerr << "*** Scheduling failed! ***\n";
+ if (PredSU->NumSuccsLeft == 0) {
+ dbgs() << "*** Scheduling failed! ***\n";
PredSU->dump(this);
- cerr << " has been released too many times!\n";
- assert(0);
+ dbgs() << " has been released too many times!\n";
+ llvm_unreachable(0);
}
#endif
-
+ --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) {
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 (!LiveRegDefs[I->getReg()]) {
/// 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(dbgs() << "\n*** Scheduling [" << CurCycle << "]: ");
DEBUG(SU->dump(this));
- assert(CurCycle >= SU->getHeight() && "Node scheduled below its height!");
+#ifndef NDEBUG
+ if (CurCycle < SU->getHeight())
+ DEBUG(dbgs() << " Height [" << SU->getHeight() << "] pipeline stall!\n");
+#endif
+
+ // FIXME: Handle noop hazard.
SU->setHeightToAtLeast(CurCycle);
Sequence.push_back(SU);
+ AvailableQueue->ScheduledNode(SU);
+
ReleasePredecessors(SU, CurCycle);
// Release all the implicit physical register defs that are live.
}
SU->isScheduled = true;
- AvailableQueue->ScheduledNode(SU);
}
/// 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(SDep *PredEdge) {
+void ScheduleDAGRRList::CapturePred(SDep *PredEdge) {
SUnit *PredSU = PredEdge->getSUnit();
if (PredSU->isAvailable) {
PredSU->isAvailable = false;
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->getHeight() << "]: ";
+ DEBUG(dbgs() << "*** Unscheduling [" << SU->getHeight() << "]: ");
DEBUG(SU->dump(this));
- AvailableQueue->UnscheduledNode(SU);
-
for (SUnit::pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
I != E; ++I) {
CapturePred(&*I);
- if (I->isAssignedRegDep() && SU->getHeight() == LiveRegCycles[I->getReg()]) {
+ if (I->isAssignedRegDep() && SU->getHeight() == LiveRegCycles[I->getReg()]){
assert(NumLiveRegs > 0 && "NumLiveRegs is already zero!");
assert(LiveRegDefs[I->getReg()] == I->getSUnit() &&
"Physical register dependency violated?");
SU->isScheduled = false;
SU->isAvailable = true;
AvailableQueue->push(SU);
+ AvailableQueue->UnscheduledNode(SU);
}
/// BacktrackBottomUp - Backtrack scheduling to a previous cycle specified in
SU->isAvailable = false;
UnscheduleNodeBottomUp(OldSU);
--CurCycle;
+ AvailableQueue->setCurCycle(CurCycle);
}
assert(!SU->isSucc(OldSU) && "Something is wrong!");
++NumBacktracks;
}
+static bool isOperandOf(const SUnit *SU, SDNode *N) {
+ for (const SDNode *SUNode = SU->getNode(); SUNode;
+ SUNode = SUNode->getFlaggedNode()) {
+ 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) {
SUnit *NewSU;
bool TryUnfold = false;
for (unsigned i = 0, e = N->getNumValues(); i != e; ++i) {
- MVT 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 SDValue &Op = N->getOperand(i);
- MVT VT = Op.getNode()->getValueType(Op.getResNo());
- if (VT == MVT::Flag)
+ EVT VT = Op.getNode()->getValueType(Op.getResNo());
+ if (VT == MVT::Glue)
return NULL;
}
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];
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->isCommutable = true;
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;
I != E; ++I) {
if (I->isCtrl())
ChainPreds.push_back(*I);
- else if (I->getSUnit()->getNode() &&
- I->getSUnit()->getNode()->isOperandOf(LoadNode))
+ else if (isOperandOf(I->getSUnit(), LoadNode))
LoadPreds.push_back(*I);
else
NodePreds.push_back(*I);
NodeSuccs.push_back(*I);
}
+ // 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);
D.setSUnit(LoadSU);
AddPred(SuccDep, D);
}
- }
- if (isNewLoad) {
- AddPred(NewSU, SDep(LoadSU, SDep::Order, LoadSU->Latency));
}
+ // 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);
AvailableQueue->addNode(NewSU);
SU = NewSU;
}
- DOUT << "Duplicating SU # " << SU->NodeNum << "\n";
+ DEBUG(dbgs() << " Duplicating SU #" << SU->NodeNum << "\n");
NewSU = CreateClone(SU);
// New SUnit has the exact same predecessors.
/// getPhysicalRegisterVT - Returns the ValueType of the physical register
/// definition of the specified node.
/// FIXME: Move to SelectionDAG?
-static MVT getPhysicalRegisterVT(SDNode *N, unsigned Reg,
+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!");
/// CheckForLiveRegDef - Return true and update live register vector if the
/// specified register def of the specified SUnit clobbers any "live" registers.
-static bool CheckForLiveRegDef(SUnit *SU, unsigned Reg,
+static void CheckForLiveRegDef(SUnit *SU, unsigned Reg,
std::vector<SUnit*> &LiveRegDefs,
SmallSet<unsigned, 4> &RegAdded,
SmallVector<unsigned, 4> &LRegs,
const TargetRegisterInfo *TRI) {
- bool Added = false;
if (LiveRegDefs[Reg] && LiveRegDefs[Reg] != SU) {
- if (RegAdded.insert(Reg)) {
+ if (RegAdded.insert(Reg))
LRegs.push_back(Reg);
- Added = true;
- }
}
for (const unsigned *Alias = TRI->getAliasSet(Reg); *Alias; ++Alias)
if (LiveRegDefs[*Alias] && LiveRegDefs[*Alias] != SU) {
- if (RegAdded.insert(*Alias)) {
+ if (RegAdded.insert(*Alias))
LRegs.push_back(*Alias);
- Added = true;
- }
}
- return Added;
}
/// 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){
+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->isAssignedRegDep())
+ if (I->isAssignedRegDep() && LiveRegDefs[I->getReg()] != SU)
CheckForLiveRegDef(I->getSUnit(), I->getReg(), LiveRegDefs,
RegAdded, LRegs, TRI);
}
if (Node->getOpcode() == ISD::INLINEASM) {
// Inline asm can clobber physical defs.
unsigned NumOps = Node->getNumOperands();
- if (Node->getOperand(NumOps-1).getValueType() == MVT::Flag)
+ if (Node->getOperand(NumOps-1).getValueType() == MVT::Glue)
--NumOps; // Ignore the flag operand.
- for (unsigned i = 2; i != NumOps;) {
+ for (unsigned i = InlineAsm::Op_FirstOperand; i != NumOps;) {
unsigned Flags =
cast<ConstantSDNode>(Node->getOperand(i))->getZExtValue();
- unsigned NumVals = Flags >> 3;
+ unsigned NumVals = InlineAsm::getNumOperandRegisters(Flags);
++i; // Skip the ID value.
- if ((Flags & 7) == 2 || (Flags & 7) == 6) {
+ 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();
for (const unsigned *Reg = TID.ImplicitDefs; *Reg; ++Reg)
CheckForLiveRegDef(SU, *Reg, LiveRegDefs, RegAdded, LRegs, TRI);
}
+
return !LRegs.empty();
}
assert(LRegs.size() == 1 && "Can't handle this yet!");
unsigned Reg = LRegs[0];
SUnit *LRDef = LiveRegDefs[Reg];
- MVT VT = getPhysicalRegisterVT(LRDef->getNode(), Reg, TII);
+ EVT VT = getPhysicalRegisterVT(LRDef->getNode(), Reg, TII);
const TargetRegisterClass *RC =
- TRI->getPhysicalRegisterRegClass(Reg, VT);
+ TRI->getMinimalPhysRegClass(Reg, VT);
const TargetRegisterClass *DestRC = TRI->getCrossCopyRegClass(RC);
// If cross copy register class is null, then it must be possible copy
// Issue copies, these can be expensive cross register class copies.
SmallVector<SUnit*, 2> Copies;
InsertCopiesAndMoveSuccs(LRDef, Reg, DestRC, RC, Copies);
- DOUT << "Adding an edge from SU #" << TrySU->NodeNum
- << " to SU #" << Copies.front()->NodeNum << "\n";
+ 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,
NewDef = Copies.back();
}
- DOUT << "Adding an edge from SU #" << NewDef->NodeNum
- << " to SU #" << TrySU->NodeNum << "\n";
+ 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,
if (CurSU)
ScheduleNodeBottomUp(CurSU, CurCycle);
++CurCycle;
+ AvailableQueue->setCurCycle(CurCycle);
}
// Reverse the order if it is bottom up.
std::reverse(Sequence.begin(), Sequence.end());
-
+
#ifndef NDEBUG
VerifySchedule(isBottomUp);
#endif
/// the AvailableQueue if the count reaches zero. Also update its cycle bound.
void ScheduleDAGRRList::ReleaseSucc(SUnit *SU, const SDep *SuccEdge) {
SUnit *SuccSU = SuccEdge->getSUnit();
- --SuccSU->NumPredsLeft;
-
+
#ifndef NDEBUG
- if (SuccSU->NumPredsLeft < 0) {
- cerr << "*** Scheduling failed! ***\n";
+ if (SuccSU->NumPredsLeft == 0) {
+ dbgs() << "*** Scheduling failed! ***\n";
SuccSU->dump(this);
- cerr << " has been released too many times!\n";
- assert(0);
+ dbgs() << " has been released too many times!\n";
+ llvm_unreachable(0);
}
#endif
-
+ --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) {
/// 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(dbgs() << "*** Scheduling [" << CurCycle << "]: ");
DEBUG(SU->dump(this));
assert(CurCycle >= SU->getDepth() && "Node scheduled above its depth!");
/// schedulers.
void ScheduleDAGRRList::ListScheduleTopDown() {
unsigned CurCycle = 0;
+ AvailableQueue->setCurCycle(CurCycle);
// Release any successors of the special Entry node.
ReleaseSuccessors(&EntrySU);
SUnits[i].isAvailable = true;
}
}
-
+
// 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()) {
SUnit *CurSU = AvailableQueue->pop();
-
+
if (CurSU)
ScheduleNodeTopDown(CurSU, CurCycle);
++CurCycle;
+ AvailableQueue->setCurCycle(CurCycle);
}
-
+
#ifndef NDEBUG
VerifySchedule(isBottomUp);
#endif
//
// 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.
+
+ /// bu_ls_rr_sort - Priority function for bottom up register pressure
+ // reduction scheduler.
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;
};
+ // td_ls_rr_sort - Priority function for top down register pressure reduction
+ // scheduler.
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;
+ };
+
+ // src_ls_rr_sort - Priority function for source order scheduler.
+ struct src_ls_rr_sort : public std::binary_function<SUnit*, SUnit*, bool> {
+ RegReductionPriorityQueue<src_ls_rr_sort> *SPQ;
+ src_ls_rr_sort(RegReductionPriorityQueue<src_ls_rr_sort> *spq)
+ : SPQ(spq) {}
+ src_ls_rr_sort(const src_ls_rr_sort &RHS)
+ : SPQ(RHS.SPQ) {}
+
+ bool operator()(const SUnit* left, const SUnit* right) const;
+ };
+
+ // hybrid_ls_rr_sort - Priority function for hybrid scheduler.
+ struct hybrid_ls_rr_sort : public std::binary_function<SUnit*, SUnit*, bool> {
+ RegReductionPriorityQueue<hybrid_ls_rr_sort> *SPQ;
+ hybrid_ls_rr_sort(RegReductionPriorityQueue<hybrid_ls_rr_sort> *spq)
+ : SPQ(spq) {}
+ hybrid_ls_rr_sort(const hybrid_ls_rr_sort &RHS)
+ : SPQ(RHS.SPQ) {}
+
+ bool operator()(const SUnit* left, const SUnit* right) const;
+ };
+
+ // ilp_ls_rr_sort - Priority function for ILP (instruction level parallelism)
+ // scheduler.
+ struct ilp_ls_rr_sort : public std::binary_function<SUnit*, SUnit*, bool> {
+ RegReductionPriorityQueue<ilp_ls_rr_sort> *SPQ;
+ ilp_ls_rr_sort(RegReductionPriorityQueue<ilp_ls_rr_sort> *spq)
+ : SPQ(spq) {}
+ ilp_ls_rr_sort(const ilp_ls_rr_sort &RHS)
+ : SPQ(RHS.SPQ) {}
+
bool operator()(const SUnit* left, const SUnit* right) const;
};
} // end anonymous namespace
if (SethiUllmanNumber == 0)
SethiUllmanNumber = 1;
-
+
return SethiUllmanNumber;
}
namespace {
template<class SF>
- class VISIBILITY_HIDDEN RegReductionPriorityQueue
- : public SchedulingPriorityQueue {
- PriorityQueue<SUnit*, std::vector<SUnit*>, SF> Queue;
- unsigned currentQueueId;
+ class RegReductionPriorityQueue : public SchedulingPriorityQueue {
+ std::vector<SUnit*> Queue;
+ SF Picker;
+ unsigned CurQueueId;
+ bool TracksRegPressure;
protected:
// SUnits - The SUnits for the current graph.
std::vector<SUnit> *SUnits;
-
+
+ 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:
- RegReductionPriorityQueue(const TargetInstrInfo *tii,
- const TargetRegisterInfo *tri) :
- Queue(SF(this)), currentQueueId(0),
- TII(tii), TRI(tri), scheduleDAG(NULL) {}
-
+ RegReductionPriorityQueue(MachineFunction &mf,
+ bool tracksrp,
+ const TargetInstrInfo *tii,
+ const TargetRegisterInfo *tri,
+ const TargetLowering *tli)
+ : Picker(this), 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 initNodes(std::vector<SUnit> &sunits) {
SUnits = &sunits;
// Add pseudo dependency edges for two-address nodes.
AddPseudoTwoAddrDeps();
+ // Reroute edges to nodes with multiple uses.
+ PrescheduleNodesWithMultipleUses();
// Calculate node priorities.
CalculateSethiUllmanNumbers();
}
void releaseState() {
SUnits = 0;
SethiUllmanNumbers.clear();
+ std::fill(RegPressure.begin(), RegPressure.end(), 0);
}
unsigned getNodePriority(const SUnit *SU) const {
// CopyToReg should be close to its uses to facilitate coalescing and
// avoid spilling.
return 0;
- if (Opc == TargetInstrInfo::EXTRACT_SUBREG ||
- Opc == TargetInstrInfo::INSERT_SUBREG)
- // EXTRACT_SUBREG / INSERT_SUBREG should be close to its use to
- // facilitate coalescing.
+ 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
return 0;
return SethiUllmanNumbers[SU->NodeNum];
}
-
- unsigned size() const { return Queue.size(); }
+
+ 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 = ++currentQueueId;
- Queue.push(U);
+ U->NodeQueueId = ++CurQueueId;
+ Queue.push_back(U);
}
- void push_all(const std::vector<SUnit *> &Nodes) {
- for (unsigned i = 0, e = Nodes.size(); i != e; ++i)
- push(Nodes[i]);
- }
-
SUnit *pop() {
if (empty()) return NULL;
- SUnit *V = Queue.top();
- Queue.pop();
+ std::vector<SUnit *>::iterator Best = Queue.begin();
+ for (std::vector<SUnit *>::iterator I = llvm::next(Queue.begin()),
+ E = Queue.end(); I != E; ++I)
+ if (Picker(*Best, *I))
+ Best = I;
+ SUnit *V = *Best;
+ if (Best != prior(Queue.end()))
+ std::swap(*Best, Queue.back());
+ Queue.pop_back();
V->NodeQueueId = 0;
return V;
}
void remove(SUnit *SU) {
assert(!Queue.empty() && "Queue is empty!");
assert(SU->NodeQueueId != 0 && "Not in queue!");
- Queue.erase_one(SU);
+ 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;
}
- void setScheduleDAG(ScheduleDAGRRList *scheduleDag) {
- scheduleDAG = scheduleDag;
+ bool HighRegPressure(const SUnit *SU) const {
+ if (!TLI)
+ return false;
+
+ for (SUnit::const_pred_iterator I = SU->Preds.begin(),E = SU->Preds.end();
+ I != E; ++I) {
+ if (I->isCtrl())
+ continue;
+ SUnit *PredSU = I->getSUnit();
+ const SDNode *PN = PredSU->getNode();
+ if (!PN->isMachineOpcode()) {
+ if (PN->getOpcode() == ISD::CopyFromReg) {
+ EVT VT = PN->getValueType(0);
+ unsigned RCId = TLI->getRepRegClassFor(VT)->getID();
+ unsigned Cost = TLI->getRepRegClassCostFor(VT);
+ if ((RegPressure[RCId] + Cost) >= RegLimit[RCId])
+ return true;
+ }
+ 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();
+ unsigned Cost = TLI->getRepRegClassCostFor(VT);
+ // Check if this increases register pressure of the specific register
+ // class to the point where it would cause spills.
+ if ((RegPressure[RCId] + Cost) >= RegLimit[RCId])
+ return true;
+ continue;
+ } else if (POpc == TargetOpcode::INSERT_SUBREG ||
+ POpc == TargetOpcode::SUBREG_TO_REG) {
+ EVT VT = PN->getValueType(0);
+ unsigned RCId = TLI->getRepRegClassFor(VT)->getID();
+ unsigned Cost = TLI->getRepRegClassCostFor(VT);
+ // Check if this increases register pressure of the specific register
+ // class to the point where it would cause spills.
+ if ((RegPressure[RCId] + Cost) >= RegLimit[RCId])
+ return true;
+ continue;
+ }
+ unsigned NumDefs = TII->get(PN->getMachineOpcode()).getNumDefs();
+ for (unsigned i = 0; i != NumDefs; ++i) {
+ EVT VT = PN->getValueType(i);
+ unsigned RCId = TLI->getRepRegClassFor(VT)->getID();
+ if (RegPressure[RCId] >= RegLimit[RCId])
+ return true; // Reg pressure already high.
+ unsigned Cost = TLI->getRepRegClassCostFor(VT);
+ if (!PN->hasAnyUseOfValue(i))
+ continue;
+ // Check if this increases register pressure of the specific register
+ // class to the point where it would cause spills.
+ if ((RegPressure[RCId] + Cost) >= RegLimit[RCId])
+ return true;
+ }
+ }
+
+ return false;
+ }
+
+ void ScheduledNode(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;
+ }
+
+ for (SUnit::pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
+ I != E; ++I) {
+ if (I->isCtrl())
+ continue;
+ SUnit *PredSU = I->getSUnit();
+ if (PredSU->NumSuccsLeft != PredSU->NumSuccs)
+ 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();
+ RegPressure[RCId] += TLI->getRepRegClassCostFor(VT);
+ }
+ }
+
+ // 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 = 0; i != NumDefs; ++i) {
+ EVT VT = N->getValueType(i);
+ if (!N->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);
+ }
+ }
+
+ dumpRegPressure();
+ }
+
+ void 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;
+ }
+
+ for (SUnit::pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
+ I != E; ++I) {
+ if (I->isCtrl())
+ continue;
+ SUnit *PredSU = I->getSUnit();
+ if (PredSU->NumSuccsLeft != PredSU->NumSuccs)
+ 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);
+ }
+ }
+
+ // 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();
+ }
+
+ void setScheduleDAG(ScheduleDAGRRList *scheduleDag) {
+ scheduleDAG = scheduleDag;
+ }
+
+ void 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');
+ }
}
protected:
bool canClobber(const SUnit *SU, const SUnit *Op);
void AddPseudoTwoAddrDeps();
+ void PrescheduleNodesWithMultipleUses();
void CalculateSethiUllmanNumbers();
};
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;
}
/// closestSucc - Returns the scheduled cycle of the successor which is
return Scratches;
}
-// Bottom up
-bool bu_ls_rr_sort::operator()(const SUnit *left, const SUnit *right) const {
+/// 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;
+ 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 RetVal;
+}
+
+/// 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;
+}
+
+template <typename RRSort>
+static bool BURRSort(const SUnit *left, const SUnit *right,
+ const RegReductionPriorityQueue<RRSort> *SPQ) {
unsigned LPriority = SPQ->getNodePriority(left);
unsigned RPriority = SPQ->getNodePriority(right);
if (LPriority != RPriority)
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 &&
+ assert(left->NodeQueueId && right->NodeQueueId &&
"NodeQueueId cannot be zero");
return (left->NodeQueueId > right->NodeQueueId);
}
+// Bottom up
+bool bu_ls_rr_sort::operator()(const SUnit *left, const SUnit *right) const {
+ return BURRSort(left, right, SPQ);
+}
+
+// Source order, otherwise bottom up.
+bool src_ls_rr_sort::operator()(const SUnit *left, const 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);
+}
+
+bool hybrid_ls_rr_sort::operator()(const SUnit *left, const 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)
+ return true;
+ else if (!LHigh && RHigh)
+ return false;
+ else if (!LHigh && !RHigh) {
+ // 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;
+
+ // Low register pressure situation, schedule for latency if possible.
+ bool LStall = left->SchedulingPref == Sched::Latency &&
+ (int)SPQ->getCurCycle() < LHeight;
+ bool RStall = right->SchedulingPref == Sched::Latency &&
+ (int)SPQ->getCurCycle() < RHeight;
+ // 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 true;
+ if (LHeight != RHeight)
+ return LHeight > RHeight;
+ } else if (RStall)
+ return false;
+
+ // If either node is scheduling for latency, sort them by height
+ // and latency.
+ if (left->SchedulingPref == Sched::Latency ||
+ right->SchedulingPref == Sched::Latency) {
+ if (LHeight != RHeight)
+ return LHeight > RHeight;
+ if (left->Latency != right->Latency)
+ return left->Latency > right->Latency;
+ }
+ }
+
+ return BURRSort(left, right, SPQ);
+}
+
+bool ilp_ls_rr_sort::operator()(const SUnit *left,
+ const 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)
+ return true;
+ else if (!LHigh && RHigh)
+ return false;
+ else if (!LHigh && !RHigh) {
+ // Low register pressure situation, schedule to maximize instruction level
+ // parallelism.
+ if (left->NumPreds > right->NumPreds)
+ return false;
+ else if (left->NumPreds < right->NumPreds)
+ return false;
+ }
+
+ return BURRSort(left, right, SPQ);
+}
+
template<class SF>
bool
RegReductionPriorityQueue<SF>::canClobber(const SUnit *SU, const SUnit *Op) {
return false;
}
-
-/// hasCopyToRegUse - Return true if SU has a value successor that is a
-/// CopyToReg node.
-static bool hasCopyToRegUse(const SUnit *SU) {
- for (SUnit::const_succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
- I != E; ++I) {
- if (I->isCtrl()) continue;
- const SUnit *SuccSU = I->getSUnit();
- if (SuccSU->getNode() && SuccSU->getNode()->getOpcode() == ISD::CopyToReg)
- return true;
- }
- return false;
-}
-
/// canClobberPhysRegDefs - True if SU would clobber one of SuccSU's
/// physical register defs.
static bool canClobberPhysRegDefs(const SUnit *SuccSU, const SUnit *SU,
unsigned NumDefs = TII->get(N->getMachineOpcode()).getNumDefs();
const unsigned *ImpDefs = TII->get(N->getMachineOpcode()).getImplicitDefs();
assert(ImpDefs && "Caller should check hasPhysRegDefs");
- const unsigned *SUImpDefs =
- TII->get(SU->getNode()->getMachineOpcode()).getImplicitDefs();
- if (!SUImpDefs)
- return false;
- for (unsigned i = NumDefs, e = N->getNumValues(); i != e; ++i) {
- MVT VT = N->getValueType(i);
- if (VT == MVT::Flag || VT == MVT::Other)
+ for (const SDNode *SUNode = SU->getNode(); SUNode;
+ SUNode = SUNode->getFlaggedNode()) {
+ if (!SUNode->isMachineOpcode())
continue;
- if (!N->hasAnyUseOfValue(i))
- continue;
- unsigned Reg = ImpDefs[i - NumDefs];
- for (;*SUImpDefs; ++SUImpDefs) {
- unsigned SUReg = *SUImpDefs;
- if (TRI->regsOverlap(Reg, SUReg))
- return true;
+ 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;
}
+/// 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.
+///
+template<class SF>
+void RegReductionPriorityQueue<SF>::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;
+ // 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:;
+ }
+}
+
/// AddPseudoTwoAddrDeps - If two nodes share an operand and one of them uses
/// it as a def&use operand. Add a pseudo control edge from it to the other
/// node (if it won't create a cycle) so the two-address one will be scheduled
if (!Node || !Node->isMachineOpcode() || SU->getNode()->getFlaggedNode())
continue;
+ bool isLiveOut = hasOnlyLiveOutUses(SU);
unsigned Opc = Node->getMachineOpcode();
const TargetInstrDesc &TID = TII->get(Opc);
unsigned NumRes = TID.getNumDefs();
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) {
+ if (SuccSU->hasPhysRegDefs && SU->hasPhysRegClobbers) {
if (canClobberPhysRegDefs(SuccSU, SU, TII, TRI))
continue;
}
- // Don't constrain extract_subreg / insert_subreg; these may be
- // coalesced away. We want them close to their uses.
+ // 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 == TargetInstrInfo::EXTRACT_SUBREG ||
- SuccOpc == TargetInstrInfo::INSERT_SUBREG)
+ if (SuccOpc == TargetOpcode::EXTRACT_SUBREG ||
+ SuccOpc == TargetOpcode::INSERT_SUBREG ||
+ SuccOpc == TargetOpcode::SUBREG_TO_REG)
continue;
if ((!canClobber(SuccSU, DUSU) ||
- (hasCopyToRegUse(SU) && !hasCopyToRegUse(SuccSU)) ||
+ (isLiveOut && !hasOnlyLiveOutUses(SuccSU)) ||
(!SU->isCommutable && SuccSU->isCommutable)) &&
!scheduleDAG->IsReachable(SuccSU, SU)) {
- DOUT << "Adding a pseudo-two-addr edge from SU # " << SU->NodeNum
- << " to SU #" << SuccSU->NodeNum << "\n";
+ 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,
template<class SF>
void RegReductionPriorityQueue<SF>::CalculateSethiUllmanNumbers() {
SethiUllmanNumbers.assign(SUnits->size(), 0);
-
+
for (unsigned i = 0, e = SUnits->size(); i != e; ++i)
CalcNodeSethiUllmanNumber(&(*SUnits)[i], SethiUllmanNumbers);
}
/// 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,
+static unsigned LimitedSumOfUnscheduledPredsOfSuccs(const SUnit *SU,
unsigned Limit) {
unsigned Sum = 0;
for (SUnit::const_succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
if (left->NumSuccsLeft != right->NumSuccsLeft)
return left->NumSuccsLeft > right->NumSuccsLeft;
- assert(left->NodeQueueId && right->NodeQueueId &&
+ assert(left->NodeQueueId && right->NodeQueueId &&
"NodeQueueId cannot be zero");
return (left->NodeQueueId > right->NodeQueueId);
}
//===----------------------------------------------------------------------===//
llvm::ScheduleDAGSDNodes *
-llvm::createBURRListDAGScheduler(SelectionDAGISel *IS, bool) {
+llvm::createBURRListDAGScheduler(SelectionDAGISel *IS, CodeGenOpt::Level) {
+ 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, true, false, PQ);
+ PQ->setScheduleDAG(SD);
+ return SD;
+}
+
+llvm::ScheduleDAGSDNodes *
+llvm::createTDRRListDAGScheduler(SelectionDAGISel *IS, CodeGenOpt::Level) {
+ 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, false, PQ);
+ PQ->setScheduleDAG(SD);
+ return SD;
+}
+
+llvm::ScheduleDAGSDNodes *
+llvm::createSourceListDAGScheduler(SelectionDAGISel *IS, CodeGenOpt::Level) {
+ 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, true, false, PQ);
+ PQ->setScheduleDAG(SD);
+ return SD;
+}
+
+llvm::ScheduleDAGSDNodes *
+llvm::createHybridListDAGScheduler(SelectionDAGISel *IS, CodeGenOpt::Level) {
const TargetMachine &TM = IS->TM;
const TargetInstrInfo *TII = TM.getInstrInfo();
const TargetRegisterInfo *TRI = TM.getRegisterInfo();
-
- BURegReductionPriorityQueue *PQ = new BURegReductionPriorityQueue(TII, TRI);
+ const TargetLowering *TLI = &IS->getTargetLowering();
- ScheduleDAGRRList *SD =
- new ScheduleDAGRRList(*IS->MF, true, PQ);
+ HybridBURRPriorityQueue *PQ =
+ new HybridBURRPriorityQueue(*IS->MF, true, TII, TRI, TLI);
+ ScheduleDAGRRList *SD = new ScheduleDAGRRList(*IS->MF, true, true, PQ);
PQ->setScheduleDAG(SD);
- return SD;
+ return SD;
}
llvm::ScheduleDAGSDNodes *
-llvm::createTDRRListDAGScheduler(SelectionDAGISel *IS, bool) {
+llvm::createILPListDAGScheduler(SelectionDAGISel *IS, CodeGenOpt::Level) {
const TargetMachine &TM = IS->TM;
const TargetInstrInfo *TII = TM.getInstrInfo();
const TargetRegisterInfo *TRI = TM.getRegisterInfo();
-
- TDRegReductionPriorityQueue *PQ = new TDRegReductionPriorityQueue(TII, TRI);
+ const TargetLowering *TLI = &IS->getTargetLowering();
- ScheduleDAGRRList *SD =
- new ScheduleDAGRRList(*IS->MF, false, PQ);
+ ILPBURRPriorityQueue *PQ =
+ new ILPBURRPriorityQueue(*IS->MF, true, TII, TRI, TLI);
+ ScheduleDAGRRList *SD = new ScheduleDAGRRList(*IS->MF, true, true, PQ);
PQ->setScheduleDAG(SD);
return SD;
}
projects / oota-llvm.git / blobdiff