//===----------------------------------------------------------------------===//
#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/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/Target/TargetLowering.h"
+#include "ScheduleDAGSDNodes.h"
+#include "llvm/ADT/STLExtras.h"
#include "llvm/ADT/SmallSet.h"
#include "llvm/ADT/Statistic.h"
-#include "llvm/ADT/STLExtras.h"
+#include "llvm/CodeGen/ScheduleHazardRecognizer.h"
+#include "llvm/CodeGen/SelectionDAGISel.h"
+#include "llvm/IR/DataLayout.h"
+#include "llvm/IR/InlineAsm.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/ErrorHandling.h"
#include "llvm/Support/raw_ostream.h"
+#include "llvm/Target/TargetInstrInfo.h"
+#include "llvm/Target/TargetLowering.h"
+#include "llvm/Target/TargetMachine.h"
+#include "llvm/Target/TargetRegisterInfo.h"
#include <climits>
using namespace llvm;
"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, FactStall, FactHeight, FactDepth,
- FactStatic, FactOther, NumFactors };
-}
-static const char *FactorName[NumFactors] =
-{"PressureDiff", "RegUses", "Stall", "Height", "Depth","Static", "Other"};
-static int FactorCount[NumFactors];
-#endif //!NDEBUG
-
namespace {
//===----------------------------------------------------------------------===//
/// ScheduleDAGRRList - The actual register reduction list scheduler
/// and similar queries.
ScheduleDAGTopologicalSort Topo;
+ // Hack to keep track of the inverse of FindCallSeqStart without more crazy
+ // DAG crawling.
+ DenseMap<SUnit*, SUnit*> CallSeqEndForStart;
+
public:
ScheduleDAGRRList(MachineFunction &mf, bool needlatency,
SchedulingPriorityQueue *availqueue,
CodeGenOpt::Level OptLevel)
: ScheduleDAGSDNodes(mf),
NeedLatency(needlatency), AvailableQueue(availqueue), CurCycle(0),
- Topo(SUnits) {
+ Topo(SUnits, NULL) {
const TargetMachine &tm = mf.getTarget();
if (DisableSchedCycles || !NeedLatency)
/// Updates the topological ordering if required.
SUnit *CreateNewSUnit(SDNode *N) {
unsigned NumSUnits = SUnits.size();
- SUnit *NewNode = NewSUnit(N);
+ SUnit *NewNode = newSUnit(N);
// Update the topological ordering.
if (NewNode->NodeNum >= NumSUnits)
Topo.InitDAGTopologicalSorting();
return NewNode;
}
- /// ForceUnitLatencies - Register-pressure-reducing scheduling doesn't
+ /// forceUnitLatencies - Register-pressure-reducing scheduling doesn't
/// need actual latency information but the hybrid scheduler does.
- bool ForceUnitLatencies() const {
+ bool forceUnitLatencies() const {
return !NeedLatency;
}
};
/// GetCostForDef - Looks up the register class and cost for a given definition.
/// Typically this just means looking up the representative register class,
-/// but for untyped values (MVT::untyped) it means inspecting the node's
+/// but for untyped values (MVT::Untyped) it means inspecting the node's
/// opcode to determine what register class is being generated.
static void GetCostForDef(const ScheduleDAGSDNodes::RegDefIter &RegDefPos,
const TargetLowering *TLI,
const TargetInstrInfo *TII,
const TargetRegisterInfo *TRI,
- unsigned &RegClass, unsigned &Cost) {
- EVT VT = RegDefPos.GetValue();
+ unsigned &RegClass, unsigned &Cost,
+ const MachineFunction &MF) {
+ MVT VT = RegDefPos.GetValue();
// Special handling for untyped values. These values can only come from
// the expansion of custom DAG-to-DAG patterns.
- if (VT == MVT::untyped) {
+ if (VT == MVT::Untyped) {
const SDNode *Node = RegDefPos.GetNode();
unsigned Opcode = Node->getMachineOpcode();
unsigned Idx = RegDefPos.GetIdx();
const MCInstrDesc Desc = TII->get(Opcode);
- const TargetRegisterClass *RC = TII->getRegClass(Desc, Idx, TRI);
+ const TargetRegisterClass *RC = TII->getRegClass(Desc, Idx, TRI, MF);
RegClass = RC->getID();
// FIXME: Cost arbitrarily set to 1 because there doesn't seem to be a
// better way to determine it.
DEBUG(dbgs()
<< "********** List Scheduling BB#" << BB->getNumber()
<< " '" << BB->getName() << "' **********\n");
-#ifndef NDEBUG
- for (int i = 0; i < NumFactors; ++i) {
- FactorCount[i] = 0;
- }
-#endif //!NDEBUG
CurCycle = 0;
IssueCount = 0;
// to track the virtual resource of a calling sequence.
LiveRegDefs.resize(TRI->getNumRegs() + 1, NULL);
LiveRegGens.resize(TRI->getNumRegs() + 1, NULL);
+ CallSeqEndForStart.clear();
// Build the scheduling graph.
BuildSchedGraph(NULL);
// Execute the actual scheduling loop.
ListScheduleBottomUp();
-#ifndef NDEBUG
- for (int i = 0; i < NumFactors; ++i) {
- DEBUG(dbgs() << FactorName[i] << "\t" << FactorCount[i] << "\n");
- }
-#endif // !NDEBUG
AvailableQueue->releaseState();
+
+ DEBUG({
+ dbgs() << "*** Final schedule ***\n";
+ dumpSchedule();
+ dbgs() << '\n';
+ });
}
//===----------------------------------------------------------------------===//
#endif
--PredSU->NumSuccsLeft;
- if (!ForceUnitLatencies()) {
+ 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());
SDNode *N = FindCallSeqStart(Node, NestLevel, MaxNest, TII);
SUnit *Def = &SUnits[N->getNodeId()];
+ CallSeqEndForStart[Def] = SU;
+
++NumLiveRegs;
LiveRegDefs[CallResource] = Def;
LiveRegGens[CallResource] = SU;
break;
case ISD::MERGE_VALUES:
case ISD::TokenFactor:
+ case ISD::LIFETIME_START:
+ case ISD::LIFETIME_END:
case ISD::CopyToReg:
case ISD::CopyFromReg:
case ISD::EH_LABEL:
Sequence.push_back(SU);
- AvailableQueue->ScheduledNode(SU);
+ AvailableQueue->scheduledNode(SU);
// If HazardRec is disabled, and each inst counts as one cycle, then
// advance CurCycle before ReleasePredecessors to avoid useless pushes to
SUNode->getMachineOpcode() == (unsigned)TII->getCallFrameSetupOpcode()) {
++NumLiveRegs;
LiveRegDefs[CallResource] = SU;
- LiveRegGens[CallResource] = NULL;
+ LiveRegGens[CallResource] = CallSeqEndForStart[SU];
}
}
for (SUnit::succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
I != E; ++I) {
if (I->isAssignedRegDep()) {
+ if (!LiveRegDefs[I->getReg()])
+ ++NumLiveRegs;
// 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 (LiveRegGens[I->getReg()] == NULL ||
I->getSUnit()->getHeight() < LiveRegGens[I->getReg()]->getHeight())
LiveRegGens[I->getReg()] = I->getSUnit();
else {
AvailableQueue->push(SU);
}
- AvailableQueue->UnscheduledNode(SU);
+ AvailableQueue->unscheduledNode(SU);
}
/// After backtracking, the hazard checker needs to be restored to a state
-/// corresponding the the current cycle.
+/// corresponding the current cycle.
void ScheduleDAGRRList::RestoreHazardCheckerBottomUp() {
HazardRec->Reset();
LoadNode->setNodeId(LoadSU->NodeNum);
InitNumRegDefsLeft(LoadSU);
- ComputeLatency(LoadSU);
+ computeLatency(LoadSU);
}
SUnit *NewSU = CreateNewSUnit(N);
NewSU->isCommutable = true;
InitNumRegDefsLeft(NewSU);
- ComputeLatency(NewSU);
+ computeLatency(NewSU);
// Record all the edges to and from the old SU, by category.
SmallVector<SDep, 4> ChainPreds;
// Add a data dependency to reflect that NewSU reads the value defined
// by LoadSU.
- AddPred(NewSU, SDep(LoadSU, SDep::Data, LoadSU->Latency));
+ SDep D(LoadSU, SDep::Data, 0);
+ D.setLatency(LoadSU->Latency);
+ AddPred(NewSU, D);
if (isNewLoad)
AvailableQueue->addNode(LoadSU);
// Avoid scheduling the def-side copy before other successors. Otherwise
// we could introduce another physreg interference on the copy and
// continue inserting copies indefinitely.
- SDep D(CopyFromSU, SDep::Order, /*Latency=*/0,
- /*Reg=*/0, /*isNormalMemory=*/false,
- /*isMustAlias=*/false, /*isArtificial=*/true);
- AddPred(SuccSU, D);
+ AddPred(SuccSU, SDep(CopyFromSU, SDep::Artificial));
}
}
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));
+ SDep FromDep(SU, SDep::Data, Reg);
+ FromDep.setLatency(SU->Latency);
+ AddPred(CopyFromSU, FromDep);
+ SDep ToDep(CopyFromSU, SDep::Data, 0);
+ ToDep.setLatency(CopyFromSU->Latency);
+ AddPred(CopyToSU, ToDep);
AvailableQueue->updateNode(SU);
AvailableQueue->addNode(CopyFromSU);
const MCInstrDesc &MCID = TII->get(N->getMachineOpcode());
assert(MCID.ImplicitDefs && "Physical reg def must be in implicit def list!");
unsigned NumRes = MCID.getNumDefs();
- for (const unsigned *ImpDef = MCID.getImplicitDefs(); *ImpDef; ++ImpDef) {
+ for (const uint16_t *ImpDef = MCID.getImplicitDefs(); *ImpDef; ++ImpDef) {
if (Reg == *ImpDef)
break;
++NumRes;
SmallSet<unsigned, 4> &RegAdded,
SmallVector<unsigned, 4> &LRegs,
const TargetRegisterInfo *TRI) {
- for (const unsigned *AliasI = TRI->getOverlaps(Reg); *AliasI; ++AliasI) {
+ for (MCRegAliasIterator AliasI(Reg, TRI, true); AliasI.isValid(); ++AliasI) {
// Check if Ref is live.
if (!LiveRegDefs[*AliasI]) continue;
}
}
+/// CheckForLiveRegDefMasked - Check for any live physregs that are clobbered
+/// by RegMask, and add them to LRegs.
+static void CheckForLiveRegDefMasked(SUnit *SU, const uint32_t *RegMask,
+ std::vector<SUnit*> &LiveRegDefs,
+ SmallSet<unsigned, 4> &RegAdded,
+ SmallVector<unsigned, 4> &LRegs) {
+ // Look at all live registers. Skip Reg0 and the special CallResource.
+ for (unsigned i = 1, e = LiveRegDefs.size()-1; i != e; ++i) {
+ if (!LiveRegDefs[i]) continue;
+ if (LiveRegDefs[i] == SU) continue;
+ if (!MachineOperand::clobbersPhysReg(RegMask, i)) continue;
+ if (RegAdded.insert(i))
+ LRegs.push_back(i);
+ }
+}
+
+/// getNodeRegMask - Returns the register mask attached to an SDNode, if any.
+static const uint32_t *getNodeRegMask(const SDNode *N) {
+ for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
+ if (const RegisterMaskSDNode *Op =
+ dyn_cast<RegisterMaskSDNode>(N->getOperand(i).getNode()))
+ return Op->getRegMask();
+ return NULL;
+}
+
/// 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
LRegs.push_back(CallResource);
}
}
+ if (const uint32_t *RegMask = getNodeRegMask(Node))
+ CheckForLiveRegDefMasked(SU, RegMask, LiveRegDefs, RegAdded, LRegs);
+
const MCInstrDesc &MCID = TII->get(Node->getMachineOpcode());
if (!MCID.ImplicitDefs)
continue;
- for (const unsigned *Reg = MCID.ImplicitDefs; *Reg; ++Reg)
+ for (const uint16_t *Reg = MCID.getImplicitDefs(); *Reg; ++Reg)
CheckForLiveRegDef(SU, *Reg, LiveRegDefs, RegAdded, LRegs, TRI);
}
if (!BtSU->isPending)
AvailableQueue->remove(BtSU);
}
- AddPred(TrySU, SDep(BtSU, SDep::Order, /*Latency=*/1,
- /*Reg=*/0, /*isNormalMemory=*/false,
- /*isMustAlias=*/false, /*isArtificial=*/true));
+ AddPred(TrySU, SDep(BtSU, SDep::Artificial));
// If one or more successors has been unscheduled, then the current
// node is no longer avaialable. Schedule a successor that's now
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));
+ AddPred(TrySU, SDep(Copies.front(), SDep::Artificial));
NewDef = Copies.back();
}
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));
+ AddPred(NewDef, SDep(TrySU, SDep::Artificial));
TrySU->isAvailable = false;
CurSU = NewDef;
}
std::reverse(Sequence.begin(), Sequence.end());
#ifndef NDEBUG
- VerifySchedule(/*isBottomUp=*/true);
+ VerifyScheduledSequence(/*isBottomUp=*/true);
#endif
}
std::vector<SUnit*> Queue;
unsigned CurQueueId;
bool TracksRegPressure;
+ bool SrcOrder;
// SUnits - The SUnits for the current graph.
std::vector<SUnit> *SUnits;
RegReductionPQBase(MachineFunction &mf,
bool hasReadyFilter,
bool tracksrp,
+ bool srcorder,
const TargetInstrInfo *tii,
const TargetRegisterInfo *tri,
const TargetLowering *tli)
: SchedulingPriorityQueue(hasReadyFilter),
- CurQueueId(0), TracksRegPressure(tracksrp),
+ CurQueueId(0), TracksRegPressure(tracksrp), SrcOrder(srcorder),
MF(mf), TII(tii), TRI(tri), TLI(tli), scheduleDAG(NULL) {
if (TracksRegPressure) {
unsigned NumRC = TRI->getNumRegClasses();
int RegPressureDiff(SUnit *SU, unsigned &LiveUses) const;
- void ScheduledNode(SUnit *SU);
+ void scheduledNode(SUnit *SU);
- void UnscheduledNode(SUnit *SU);
+ void unscheduledNode(SUnit *SU);
protected:
bool canClobber(const SUnit *SU, const SUnit *Op);
public:
RegReductionPriorityQueue(MachineFunction &mf,
bool tracksrp,
+ bool srcorder,
const TargetInstrInfo *tii,
const TargetRegisterInfo *tri,
const TargetLowering *tli)
- : RegReductionPQBase(mf, SF::HasReadyFilter, tracksrp, tii, tri, tli),
+ : RegReductionPQBase(mf, SF::HasReadyFilter, tracksrp, srcorder,
+ tii, tri, tli),
Picker(this) {}
bool isBottomUp() const { return SF::IsBottomUp; }
return V;
}
+#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
void dump(ScheduleDAG *DAG) const {
// Emulate pop() without clobbering NodeQueueIds.
std::vector<SUnit*> DumpQueue = Queue;
SU->dump(DAG);
}
}
+#endif
};
typedef RegReductionPriorityQueue<bu_ls_rr_sort>
//===----------------------------------------------------------------------===//
void RegReductionPQBase::dumpRegPressure() const {
+#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
for (TargetRegisterInfo::regclass_iterator I = TRI->regclass_begin(),
E = TRI->regclass_end(); I != E; ++I) {
const TargetRegisterClass *RC = *I;
DEBUG(dbgs() << RC->getName() << ": " << RP << " / " << RegLimit[Id]
<< '\n');
}
+#endif
}
bool RegReductionPQBase::HighRegPressure(const SUnit *SU) const {
for (ScheduleDAGSDNodes::RegDefIter RegDefPos(PredSU, scheduleDAG);
RegDefPos.IsValid(); RegDefPos.Advance()) {
unsigned RCId, Cost;
- GetCostForDef(RegDefPos, TLI, TII, TRI, RCId, Cost);
+ GetCostForDef(RegDefPos, TLI, TII, TRI, RCId, Cost, MF);
if ((RegPressure[RCId] + Cost) >= RegLimit[RCId])
return true;
unsigned NumDefs = TII->get(N->getMachineOpcode()).getNumDefs();
for (unsigned i = 0; i != NumDefs; ++i) {
- EVT VT = N->getValueType(i);
+ MVT VT = N->getSimpleValueType(i);
if (!N->hasAnyUseOfValue(i))
continue;
unsigned RCId = TLI->getRepRegClassFor(VT)->getID();
}
for (ScheduleDAGSDNodes::RegDefIter RegDefPos(PredSU, scheduleDAG);
RegDefPos.IsValid(); RegDefPos.Advance()) {
- EVT VT = RegDefPos.GetValue();
+ MVT VT = RegDefPos.GetValue();
unsigned RCId = TLI->getRepRegClassFor(VT)->getID();
if (RegPressure[RCId] >= RegLimit[RCId])
++PDiff;
unsigned NumDefs = TII->get(N->getMachineOpcode()).getNumDefs();
for (unsigned i = 0; i != NumDefs; ++i) {
- EVT VT = N->getValueType(i);
+ MVT VT = N->getSimpleValueType(i);
if (!N->hasAnyUseOfValue(i))
continue;
unsigned RCId = TLI->getRepRegClassFor(VT)->getID();
return PDiff;
}
-void RegReductionPQBase::ScheduledNode(SUnit *SU) {
+void RegReductionPQBase::scheduledNode(SUnit *SU) {
if (!TracksRegPressure)
return;
continue;
unsigned RCId, Cost;
- GetCostForDef(RegDefPos, TLI, TII, TRI, RCId, Cost);
+ GetCostForDef(RegDefPos, TLI, TII, TRI, RCId, Cost, MF);
RegPressure[RCId] += Cost;
break;
}
if (SkipRegDefs > 0)
continue;
unsigned RCId, Cost;
- GetCostForDef(RegDefPos, TLI, TII, TRI, RCId, Cost);
+ GetCostForDef(RegDefPos, TLI, TII, TRI, RCId, Cost, MF);
if (RegPressure[RCId] < Cost) {
// Register pressure tracking is imprecise. This can happen. But we try
// hard not to let it happen because it likely results in poor scheduling.
dumpRegPressure();
}
-void RegReductionPQBase::UnscheduledNode(SUnit *SU) {
+void RegReductionPQBase::unscheduledNode(SUnit *SU) {
if (!TracksRegPressure)
return;
const SDNode *PN = PredSU->getNode();
if (!PN->isMachineOpcode()) {
if (PN->getOpcode() == ISD::CopyFromReg) {
- EVT VT = PN->getValueType(0);
+ MVT VT = PN->getSimpleValueType(0);
unsigned RCId = TLI->getRepRegClassFor(VT)->getID();
RegPressure[RCId] += TLI->getRepRegClassCostFor(VT);
}
if (POpc == TargetOpcode::EXTRACT_SUBREG ||
POpc == TargetOpcode::INSERT_SUBREG ||
POpc == TargetOpcode::SUBREG_TO_REG) {
- EVT VT = PN->getValueType(0);
+ MVT VT = PN->getSimpleValueType(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);
+ MVT VT = PN->getSimpleValueType(i);
if (!PN->hasAnyUseOfValue(i))
continue;
unsigned RCId = TLI->getRepRegClassFor(VT)->getID();
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);
+ MVT VT = N->getSimpleValueType(i);
if (VT == MVT::Glue || VT == MVT::Other)
continue;
if (!N->hasAnyUseOfValue(i))
// If scheduling either one of the node will cause a pipeline stall, sort
// them according to their height.
if (LStall) {
- if (!RStall) {
- DEBUG(++FactorCount[FactStall]);
+ if (!RStall)
return 1;
- }
- if (LHeight != RHeight) {
- DEBUG(++FactorCount[FactStall]);
+ if (LHeight != RHeight)
return LHeight > RHeight ? 1 : -1;
- }
- } else if (RStall) {
- DEBUG(++FactorCount[FactStall]);
+ } else if (RStall)
return -1;
- }
// If either node is scheduling for latency, sort them by height/depth
// and latency.
if (!checkPref || (left->SchedulingPref == Sched::ILP ||
right->SchedulingPref == Sched::ILP)) {
- if (DisableSchedCycles) {
- if (LHeight != RHeight) {
- DEBUG(++FactorCount[FactHeight]);
+ // If neither instruction stalls (!LStall && !RStall) and HazardRecognizer
+ // is enabled, grouping instructions by cycle, then its height is already
+ // covered so only its depth matters. We also reach this point if both stall
+ // but have the same height.
+ if (!SPQ->getHazardRec()->isEnabled()) {
+ if (LHeight != RHeight)
return LHeight > RHeight ? 1 : -1;
- }
}
- else {
- // If neither instruction stalls (!LStall && !RStall) then
- // its height is already covered so only its depth matters. We also reach
- // this if both stall but have the same height.
- int LDepth = left->getDepth() - LPenalty;
- int RDepth = right->getDepth() - RPenalty;
- if (LDepth != RDepth) {
- DEBUG(++FactorCount[FactDepth]);
- DEBUG(dbgs() << " Comparing latency of SU (" << left->NodeNum
- << ") depth " << LDepth << " vs SU (" << right->NodeNum
- << ") depth " << RDepth << "\n");
- return LDepth < RDepth ? 1 : -1;
- }
+ int LDepth = left->getDepth() - LPenalty;
+ int RDepth = right->getDepth() - RPenalty;
+ 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) {
- DEBUG(++FactorCount[FactOther]);
+ if (left->Latency != right->Latency)
return left->Latency > right->Latency ? 1 : -1;
- }
}
return 0;
}
bool LHasPhysReg = left->hasPhysRegDefs;
bool RHasPhysReg = right->hasPhysRegDefs;
if (LHasPhysReg != RHasPhysReg) {
- DEBUG(++FactorCount[FactRegUses]);
#ifndef NDEBUG
- const char *PhysRegMsg[] = {" has no physreg", " defines a physreg"};
+ const char *const PhysRegMsg[] = {" has no physreg"," defines a physreg"};
#endif
DEBUG(dbgs() << " SU (" << left->NodeNum << ") "
<< PhysRegMsg[LHasPhysReg] << " SU(" << right->NodeNum << ") "
LPriority = (LPriority > LNumVals) ? (LPriority - LNumVals) : 0;
}
- if (LPriority != RPriority) {
- DEBUG(++FactorCount[FactStatic]);
+ if (LPriority != RPriority)
return LPriority > RPriority;
- }
// One or both of the nodes are calls and their sethi-ullman numbers are the
// same, then keep source order.
// This creates more short live intervals.
unsigned LDist = closestSucc(left);
unsigned RDist = closestSucc(right);
- if (LDist != RDist) {
- DEBUG(++FactorCount[FactOther]);
+ 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) {
- DEBUG(++FactorCount[FactOther]);
+ if (LScratch != RScratch)
return LScratch > RScratch;
- }
// Comparing latency against a call makes little sense unless the node
// is register pressure-neutral.
return result > 0;
}
else {
- if (left->getHeight() != right->getHeight()) {
- DEBUG(++FactorCount[FactHeight]);
+ if (left->getHeight() != right->getHeight())
return left->getHeight() > right->getHeight();
- }
- if (left->getDepth() != right->getDepth()) {
- DEBUG(++FactorCount[FactDepth]);
+ if (left->getDepth() != right->getDepth())
return left->getDepth() < right->getDepth();
- }
}
assert(left->NodeQueueId && right->NodeQueueId &&
"NodeQueueId cannot be zero");
- DEBUG(++FactorCount[FactOther]);
return (left->NodeQueueId > right->NodeQueueId);
}
// Avoid causing spills. If register pressure is high, schedule for
// register pressure reduction.
if (LHigh && !RHigh) {
- DEBUG(++FactorCount[FactPressureDiff]);
DEBUG(dbgs() << " pressure SU(" << left->NodeNum << ") > SU("
<< right->NodeNum << ")\n");
return true;
}
else if (!LHigh && RHigh) {
- DEBUG(++FactorCount[FactPressureDiff]);
DEBUG(dbgs() << " pressure SU(" << right->NodeNum << ") > SU("
<< left->NodeNum << ")\n");
return false;
RPDiff = SPQ->RegPressureDiff(right, RLiveUses);
}
if (!DisableSchedRegPressure && LPDiff != RPDiff) {
- DEBUG(++FactorCount[FactPressureDiff]);
DEBUG(dbgs() << "RegPressureDiff SU(" << left->NodeNum << "): " << LPDiff
<< " != SU(" << right->NodeNum << "): " << RPDiff << "\n");
return LPDiff > RPDiff;
if (!DisableSchedRegPressure && (LPDiff > 0 || RPDiff > 0)) {
bool LReduce = canEnableCoalescing(left);
bool RReduce = canEnableCoalescing(right);
- DEBUG(if (LReduce != RReduce) ++FactorCount[FactPressureDiff]);
if (LReduce && !RReduce) return false;
if (RReduce && !LReduce) return true;
}
if (!DisableSchedLiveUses && (LLiveUses != RLiveUses)) {
DEBUG(dbgs() << "Live uses SU(" << left->NodeNum << "): " << LLiveUses
<< " != SU(" << right->NodeNum << "): " << RLiveUses << "\n");
- DEBUG(++FactorCount[FactRegUses]);
return LLiveUses < RLiveUses;
}
if (!DisableSchedStalls) {
bool LStall = BUHasStall(left, left->getHeight(), SPQ);
bool RStall = BUHasStall(right, right->getHeight(), SPQ);
- if (LStall != RStall) {
- DEBUG(++FactorCount[FactHeight]);
+ if (LStall != RStall)
return left->getHeight() > right->getHeight();
- }
}
if (!DisableSchedCriticalPath) {
DEBUG(dbgs() << "Depth of SU(" << left->NodeNum << "): "
<< left->getDepth() << " != SU(" << right->NodeNum << "): "
<< right->getDepth() << "\n");
- DEBUG(++FactorCount[FactDepth]);
return left->getDepth() < right->getDepth();
}
}
if (!DisableSchedHeight && left->getHeight() != right->getHeight()) {
int spread = (int)left->getHeight() - (int)right->getHeight();
- if (std::abs(spread) > MaxReorderWindow) {
- DEBUG(++FactorCount[FactHeight]);
+ if (std::abs(spread) > MaxReorderWindow)
return left->getHeight() > right->getHeight();
- }
}
return BURRSort(left, right, SPQ);
if (!Disable2AddrHack)
AddPseudoTwoAddrDeps();
// Reroute edges to nodes with multiple uses.
- if (!TracksRegPressure)
+ if (!TracksRegPressure && !SrcOrder)
PrescheduleNodesWithMultipleUses();
// Calculate node priorities.
CalculateSethiUllmanNumbers();
ScheduleDAGRRList *scheduleDAG,
const TargetInstrInfo *TII,
const TargetRegisterInfo *TRI) {
- const unsigned *ImpDefs
+ const uint16_t *ImpDefs
= TII->get(SU->getNode()->getMachineOpcode()).getImplicitDefs();
- if(!ImpDefs)
+ const uint32_t *RegMask = getNodeRegMask(SU->getNode());
+ if(!ImpDefs && !RegMask)
return false;
for (SUnit::const_succ_iterator SI = SU->Succs.begin(), SE = SU->Succs.end();
if (!PI->isAssignedRegDep())
continue;
- for (const unsigned *ImpDef = ImpDefs; *ImpDef; ++ImpDef) {
- // Return true if SU clobbers this physical register use and the
- // definition of the register reaches from DepSU. IsReachable queries a
- // topological forward sort of the DAG (following the successors).
- if (TRI->regsOverlap(*ImpDef, PI->getReg()) &&
- scheduleDAG->IsReachable(DepSU, PI->getSUnit()))
- return true;
- }
+ if (RegMask && MachineOperand::clobbersPhysReg(RegMask, PI->getReg()) &&
+ scheduleDAG->IsReachable(DepSU, PI->getSUnit()))
+ return true;
+
+ if (ImpDefs)
+ for (const uint16_t *ImpDef = ImpDefs; *ImpDef; ++ImpDef)
+ // Return true if SU clobbers this physical register use and the
+ // definition of the register reaches from DepSU. IsReachable queries
+ // a topological forward sort of the DAG (following the successors).
+ if (TRI->regsOverlap(*ImpDef, PI->getReg()) &&
+ scheduleDAG->IsReachable(DepSU, PI->getSUnit()))
+ return true;
}
}
return false;
const TargetRegisterInfo *TRI) {
SDNode *N = SuccSU->getNode();
unsigned NumDefs = TII->get(N->getMachineOpcode()).getNumDefs();
- const unsigned *ImpDefs = TII->get(N->getMachineOpcode()).getImplicitDefs();
+ const uint16_t *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 =
+ const uint16_t *SUImpDefs =
TII->get(SUNode->getMachineOpcode()).getImplicitDefs();
- if (!SUImpDefs)
- return false;
+ const uint32_t *SURegMask = getNodeRegMask(SUNode);
+ if (!SUImpDefs && !SURegMask)
+ continue;
for (unsigned i = NumDefs, e = N->getNumValues(); i != e; ++i) {
EVT VT = N->getValueType(i);
if (VT == MVT::Glue || VT == MVT::Other)
if (!N->hasAnyUseOfValue(i))
continue;
unsigned Reg = ImpDefs[i - NumDefs];
+ if (SURegMask && MachineOperand::clobbersPhysReg(SURegMask, Reg))
+ return true;
+ if (!SUImpDefs)
+ continue;
for (;*SUImpDefs; ++SUImpDefs) {
unsigned SUReg = *SUImpDefs;
if (TRI->regsOverlap(Reg, SUReg))
!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));
+ scheduleDAG->AddPred(SU, SDep(SuccSU, SDep::Artificial));
}
}
}
const TargetRegisterInfo *TRI = TM.getRegisterInfo();
BURegReductionPriorityQueue *PQ =
- new BURegReductionPriorityQueue(*IS->MF, false, TII, TRI, 0);
+ new BURegReductionPriorityQueue(*IS->MF, false, false, TII, TRI, 0);
ScheduleDAGRRList *SD = new ScheduleDAGRRList(*IS->MF, false, PQ, OptLevel);
PQ->setScheduleDAG(SD);
return SD;
const TargetRegisterInfo *TRI = TM.getRegisterInfo();
SrcRegReductionPriorityQueue *PQ =
- new SrcRegReductionPriorityQueue(*IS->MF, false, TII, TRI, 0);
+ new SrcRegReductionPriorityQueue(*IS->MF, false, true, TII, TRI, 0);
ScheduleDAGRRList *SD = new ScheduleDAGRRList(*IS->MF, false, PQ, OptLevel);
PQ->setScheduleDAG(SD);
return SD;
const TargetLowering *TLI = &IS->getTargetLowering();
HybridBURRPriorityQueue *PQ =
- new HybridBURRPriorityQueue(*IS->MF, true, TII, TRI, TLI);
+ new HybridBURRPriorityQueue(*IS->MF, true, false, TII, TRI, TLI);
ScheduleDAGRRList *SD = new ScheduleDAGRRList(*IS->MF, true, PQ, OptLevel);
PQ->setScheduleDAG(SD);
const TargetLowering *TLI = &IS->getTargetLowering();
ILPBURRPriorityQueue *PQ =
- new ILPBURRPriorityQueue(*IS->MF, true, TII, TRI, TLI);
+ new ILPBURRPriorityQueue(*IS->MF, true, false, TII, TRI, TLI);
ScheduleDAGRRList *SD = new ScheduleDAGRRList(*IS->MF, true, PQ, OptLevel);
PQ->setScheduleDAG(SD);
return SD;
projects / oota-llvm.git / blobdiff