//===----------------------------------------------------------------------===//
#define DEBUG_TYPE "pre-RA-sched"
-#include "llvm/CodeGen/ScheduleDAGSDNodes.h"
+#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/SmallSet.h"
#include "llvm/ADT/Statistic.h"
#include "llvm/ADT/STLExtras.h"
-#include "llvm/Support/CommandLine.h"
+#include "llvm/Support/ErrorHandling.h"
+#include "llvm/Support/raw_ostream.h"
using namespace llvm;
STATISTIC(NumUnfolds, "Number of nodes unfolded");
STATISTIC(NumDups, "Number of duplicated nodes");
-STATISTIC(NumCCCopies, "Number of cross class copies");
+STATISTIC(NumPRCopies, "Number of physical copies");
static RegisterScheduler
fastDAGScheduler("fast", "Fast suboptimal list scheduling",
/// FastPriorityQueue - A degenerate priority queue that considers
/// all nodes to have the same priority.
///
- struct VISIBILITY_HIDDEN FastPriorityQueue {
+ struct FastPriorityQueue {
SmallVector<SUnit *, 16> Queue;
bool empty() const { return Queue.empty(); }
//===----------------------------------------------------------------------===//
/// ScheduleDAGFast - The actual "fast" list scheduler implementation.
///
-class VISIBILITY_HIDDEN ScheduleDAGFast : public ScheduleDAGSDNodes {
+class ScheduleDAGFast : public ScheduleDAGSDNodes {
private:
/// AvailableQueue - The priority queue to use for the available SUnits.
FastPriorityQueue AvailableQueue;
std::vector<unsigned> LiveRegCycles;
public:
- ScheduleDAGFast(SelectionDAG *dag, MachineBasicBlock *bb,
- const TargetMachine &tm)
- : ScheduleDAGSDNodes(dag, bb, tm) {}
+ ScheduleDAGFast(MachineFunction &mf)
+ : ScheduleDAGSDNodes(mf) {}
void Schedule();
/// AddPred - adds a predecessor edge to SUnit SU.
/// This returns true if this is a new predecessor.
- bool AddPred(SUnit *SU, const SDep &D) {
- return SU->addPred(D);
+ void AddPred(SUnit *SU, const SDep &D) {
+ SU->addPred(D);
}
/// RemovePred - removes a predecessor edge from SUnit SU.
/// This returns true if an edge was removed.
- bool RemovePred(SUnit *SU, const SDep &D) {
- return SU->removePred(D);
+ void RemovePred(SUnit *SU, const SDep &D) {
+ SU->removePred(D);
}
private:
void ReleasePred(SUnit *SU, SDep *PredEdge);
+ void ReleasePredecessors(SUnit *SU, unsigned CurCycle);
void ScheduleNodeBottomUp(SUnit*, 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 ListScheduleBottomUp();
+
+ /// ForceUnitLatencies - The fast scheduler doesn't care about real latencies.
+ bool ForceUnitLatencies() const { return true; }
};
} // end anonymous namespace
/// Schedule - Schedule the DAG using list scheduling.
void ScheduleDAGFast::Schedule() {
- DOUT << "********** List Scheduling **********\n";
+ DEBUG(dbgs() << "********** List Scheduling **********\n");
NumLiveRegs = 0;
LiveRegDefs.resize(TRI->getNumRegs(), NULL);
LiveRegCycles.resize(TRI->getNumRegs(), 0);
- // Build scheduling units.
- BuildSchedUnits();
+ // Build the scheduling graph.
+ BuildSchedGraph(NULL);
DEBUG(for (unsigned su = 0, e = SUnits.size(); su != e; ++su)
SUnits[su].dumpAll(this));
/// the AvailableQueue if the count reaches zero. Also update its cycle bound.
void ScheduleDAGFast::ReleasePred(SUnit *SU, 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
-
- if (PredSU->NumSuccsLeft == 0) {
+ --PredSU->NumSuccsLeft;
+
+ // 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 ScheduleDAGFast::ScheduleNodeBottomUp(SUnit *SU, unsigned CurCycle) {
- DOUT << "*** Scheduling [" << CurCycle << "]: ";
- DEBUG(SU->dump(this));
-
- SU->Cycle = CurCycle;
- Sequence.push_back(SU);
-
+void ScheduleDAGFast::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 ScheduleDAGFast::ScheduleNodeBottomUp(SUnit *SU, unsigned CurCycle) {
+ DEBUG(dbgs() << "*** 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();
I != E; ++I) {
if (I->isAssignedRegDep()) {
- if (LiveRegCycles[I->getReg()] == I->getSUnit()->Cycle) {
+ if (LiveRegCycles[I->getReg()] == I->getSUnit()->getHeight()) {
assert(NumLiveRegs > 0 && "NumLiveRegs is already zero!");
assert(LiveRegDefs[I->getReg()] == SU &&
"Physical register dependency violated?");
/// CopyAndMoveSuccessors - Clone the specified node and move its scheduled
/// successors to the newly created node.
SUnit *ScheduleDAGFast::CopyAndMoveSuccessors(SUnit *SU) {
- if (SU->getNode()->getFlaggedNode())
+ if (SU->getNode()->getGluedNode())
return NULL;
SDNode *N = SU->getNode();
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];
}
if (TID.isCommutable())
NewSU->isCommutable = true;
- // FIXME: Calculate height / depth and propagate the changes?
- NewSU->Depth = SU->Depth;
- NewSU->Height = SU->Height;
// LoadNode may already exist. This can happen when there is another
// load from the same location and producing the same type of value
} else {
LoadSU = NewSUnit(LoadNode);
LoadNode->setNodeId(LoadSU->NodeNum);
-
- LoadSU->Depth = SU->Depth;
- LoadSU->Height = SU->Height;
}
SDep ChainPred;
SU = NewSU;
}
- DOUT << "Duplicating SU # " << SU->NodeNum << "\n";
+ DEBUG(dbgs() << "Duplicating SU # " << SU->NodeNum << "\n");
NewSU = Clone(SU);
// New SUnit has the exact same predecessors.
for (SUnit::pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
I != E; ++I)
- if (!I->isArtificial()) {
+ if (!I->isArtificial())
AddPred(NewSU, *I);
- NewSU->Depth = std::max(NewSU->Depth, I->getSUnit()->Depth+1);
- }
// Only copy scheduled successors. Cut them from old node's successor
// list and move them over.
continue;
SUnit *SuccSU = I->getSUnit();
if (SuccSU->isScheduled) {
- NewSU->Height = std::max(NewSU->Height, SuccSU->Height+1);
SDep D = *I;
D.setSUnit(NewSU);
AddPred(SuccSU, D);
DelDeps.push_back(std::make_pair(SuccSU, D));
}
}
- 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);
- }
++NumDups;
return NewSU;
}
-/// InsertCCCopiesAndMoveSuccs - Insert expensive cross register class copies
-/// and move all scheduled successors of the given SUnit to the last copy.
-void ScheduleDAGFast::InsertCCCopiesAndMoveSuccs(SUnit *SU, unsigned Reg,
+/// InsertCopiesAndMoveSuccs - Insert register copies and move all
+/// scheduled successors of the given SUnit to the last copy.
+void ScheduleDAGFast::InsertCopiesAndMoveSuccs(SUnit *SU, unsigned Reg,
const TargetRegisterClass *DestRC,
const TargetRegisterClass *SrcRC,
SmallVector<SUnit*, 2> &Copies) {
Copies.push_back(CopyFromSU);
Copies.push_back(CopyToSU);
- ++NumCCCopies;
+ ++NumPRCopies;
}
/// getPhysicalRegisterVT - Returns the ValueType of the physical register
/// definition of the specified node.
/// FIXME: Move to SelectionDAG?
-static MVT 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!");
return N->getValueType(NumRes);
}
+/// CheckForLiveRegDef - Return true and update live register vector if the
+/// specified register def of the specified SUnit clobbers any "live" registers.
+static bool 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)) {
+ 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)) {
+ 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
for (SUnit::pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
I != E; ++I) {
if (I->isAssignedRegDep()) {
- unsigned Reg = I->getReg();
- if (LiveRegDefs[Reg] && LiveRegDefs[Reg] != I->getSUnit()) {
- if (RegAdded.insert(Reg))
- LRegs.push_back(Reg);
- }
- for (const unsigned *Alias = TRI->getAliasSet(Reg);
- *Alias; ++Alias)
- if (LiveRegDefs[*Alias] && LiveRegDefs[*Alias] != I->getSUnit()) {
- if (RegAdded.insert(*Alias))
- LRegs.push_back(*Alias);
- }
+ CheckForLiveRegDef(I->getSUnit(), I->getReg(), LiveRegDefs,
+ RegAdded, LRegs, TRI);
}
}
- for (SDNode *Node = SU->getNode(); Node; Node = Node->getFlaggedNode()) {
+ for (SDNode *Node = SU->getNode(); Node; Node = Node->getGluedNode()) {
+ if (Node->getOpcode() == ISD::INLINEASM) {
+ // Inline asm can clobber physical defs.
+ unsigned NumOps = Node->getNumOperands();
+ if (Node->getOperand(NumOps-1).getValueType() == MVT::Glue)
+ --NumOps; // Ignore the glue operand.
+
+ for (unsigned i = InlineAsm::Op_FirstOperand; i != NumOps;) {
+ unsigned Flags =
+ cast<ConstantSDNode>(Node->getOperand(i))->getZExtValue();
+ unsigned NumVals = InlineAsm::getNumOperandRegisters(Flags);
+
+ ++i; // Skip the ID value.
+ if (InlineAsm::isRegDefKind(Flags) ||
+ InlineAsm::isRegDefEarlyClobberKind(Flags)) {
+ // Check for def of register or earlyclobber register.
+ for (; NumVals; --NumVals, ++i) {
+ unsigned Reg = cast<RegisterSDNode>(Node->getOperand(i))->getReg();
+ if (TargetRegisterInfo::isPhysicalRegister(Reg))
+ CheckForLiveRegDef(SU, Reg, LiveRegDefs, RegAdded, LRegs, TRI);
+ }
+ } else
+ i += NumVals;
+ }
+ continue;
+ }
if (!Node->isMachineOpcode())
continue;
const TargetInstrDesc &TID = TII->get(Node->getMachineOpcode());
if (!TID.ImplicitDefs)
continue;
for (const unsigned *Reg = TID.ImplicitDefs; *Reg; ++Reg) {
- if (LiveRegDefs[*Reg] && LiveRegDefs[*Reg] != SU) {
- if (RegAdded.insert(*Reg))
- LRegs.push_back(*Reg);
- }
- for (const unsigned *Alias = TRI->getAliasSet(*Reg);
- *Alias; ++Alias)
- if (LiveRegDefs[*Alias] && LiveRegDefs[*Alias] != SU) {
- if (RegAdded.insert(*Alias))
- LRegs.push_back(*Alias);
- }
+ CheckForLiveRegDef(SU, *Reg, LiveRegDefs, RegAdded, LRegs, TRI);
}
}
return !LRegs.empty();
/// schedulers.
void ScheduleDAGFast::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()];
assert(LRegs.size() == 1 && "Can't handle this yet!");
unsigned Reg = LRegs[0];
SUnit *LRDef = LiveRegDefs[Reg];
- SUnit *NewDef = CopyAndMoveSuccessors(LRDef);
+ EVT VT = getPhysicalRegisterVT(LRDef->getNode(), Reg, TII);
+ const TargetRegisterClass *RC =
+ TRI->getMinimalPhysRegClass(Reg, VT);
+ const TargetRegisterClass *DestRC = TRI->getCrossCopyRegClass(RC);
+
+ // If cross copy register class is the same as RC, then it must be
+ // possible copy the value directly. Do not try duplicate the def.
+ // If cross copy register class is not the same as RC, then it's
+ // possible to copy the value but it require cross register class copies
+ // and it is expensive.
+ // If cross copy register class is null, then it's not possible to copy
+ // the value at all.
+ SUnit *NewDef = 0;
+ if (DestRC != RC) {
+ NewDef = CopyAndMoveSuccessors(LRDef);
+ if (!DestRC && !NewDef)
+ report_fatal_error("Can't handle live physical "
+ "register dependency!");
+ }
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
- << " to SU #" << Copies.front()->NodeNum << "\n";
+ InsertCopiesAndMoveSuccs(LRDef, Reg, DestRC, RC, Copies);
+ DEBUG(dbgs() << "Adding an edge from SU # " << TrySU->NodeNum
+ << " to SU #" << Copies.front()->NodeNum << "\n");
AddPred(TrySU, SDep(Copies.front(), SDep::Order, /*Latency=*/1,
/*Reg=*/0, /*isNormalMemory=*/false,
/*isMustAlias=*/false, /*isArtificial=*/true));
NewDef = Copies.back();
}
- 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) {
- assert(false && "Unable to resolve live physical register dependencies!");
- abort();
+ llvm_unreachable("Unable to resolve live physical register dependencies!");
}
}
++CurCycle;
}
- // Reverse the order if it is bottom up.
+ // Reverse the order since it is bottom up.
std::reverse(Sequence.begin(), Sequence.end());
-
-
+
#ifndef NDEBUG
- // Verify that all SUnits were scheduled.
- bool AnyNotSched = false;
- unsigned DeadNodes = 0;
- unsigned Noops = 0;
- for (unsigned i = 0, e = SUnits.size(); i != e; ++i) {
- if (!SUnits[i].isScheduled) {
- if (SUnits[i].NumPreds == 0 && SUnits[i].NumSuccs == 0) {
- ++DeadNodes;
- continue;
- }
- if (!AnyNotSched)
- cerr << "*** List scheduling failed! ***\n";
- SUnits[i].dump(this);
- cerr << "has not been scheduled!\n";
- AnyNotSched = true;
- }
- if (SUnits[i].NumSuccsLeft != 0) {
- if (!AnyNotSched)
- cerr << "*** List scheduling failed! ***\n";
- SUnits[i].dump(this);
- cerr << "has successors left!\n";
- AnyNotSched = true;
- }
- }
- for (unsigned i = 0, e = Sequence.size(); i != e; ++i)
- if (!Sequence[i])
- ++Noops;
- assert(!AnyNotSched);
- assert(Sequence.size() + DeadNodes - Noops == SUnits.size() &&
- "The number of nodes scheduled doesn't match the expected number!");
+ VerifySchedule(/*isBottomUp=*/true);
#endif
}
// Public Constructor Functions
//===----------------------------------------------------------------------===//
-llvm::ScheduleDAG* llvm::createFastDAGScheduler(SelectionDAGISel *IS,
- SelectionDAG *DAG,
- const TargetMachine *TM,
- MachineBasicBlock *BB, bool) {
- return new ScheduleDAGFast(DAG, BB, *TM);
+llvm::ScheduleDAGSDNodes *
+llvm::createFastDAGScheduler(SelectionDAGISel *IS, CodeGenOpt::Level) {
+ return new ScheduleDAGFast(*IS->MF);
}