#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/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;
/// 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(); }
-
+
void push(SUnit *U) {
Queue.push_back(U);
}
//===----------------------------------------------------------------------===//
/// 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;
bool DelayForLiveRegsBottomUp(SUnit*, SmallVector<unsigned, 4>&);
void ListScheduleBottomUp();
- /// ForceUnitLatencies - The fast scheduler doesn't care about real latencies.
- bool ForceUnitLatencies() const { return true; }
+ /// 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() {
- DEBUG(errs() << "********** List Scheduling **********\n");
+ DEBUG(dbgs() << "********** List Scheduling **********\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));
/// 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) {
- errs() << "*** Scheduling failed! ***\n";
+ if (PredSU->NumSuccsLeft == 0) {
+ dbgs() << "*** Scheduling failed! ***\n";
PredSU->dump(this);
- errs() << " has been released too many times!\n";
+ dbgs() << " has been released too many times!\n";
llvm_unreachable(0);
}
#endif
-
+ --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) {
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 ScheduleDAGFast::ScheduleNodeBottomUp(SUnit *SU, unsigned CurCycle) {
- DEBUG(errs() << "*** Scheduling [" << CurCycle << "]: ");
+ DEBUG(dbgs() << "*** Scheduling [" << CurCycle << "]: ");
DEBUG(SU->dump(this));
assert(CurCycle >= SU->getHeight() && "Node scheduled below its height!");
/// 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();
bool TryUnfold = false;
for (unsigned i = 0, e = N->getNumValues(); i != e; ++i) {
EVT VT = N->getValueType(i);
- if (VT == MVT::Flag)
+ 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);
EVT VT = Op.getNode()->getValueType(Op.getResNo());
- if (VT == MVT::Flag)
+ if (VT == MVT::Glue)
return NULL;
}
if (!TII->unfoldMemoryOperand(*DAG, N, NewNodes))
return NULL;
- DEBUG(errs() << "Unfolding SU # " << SU->NodeNum << "\n");
+ DEBUG(dbgs() << "Unfolding SU # " << SU->NodeNum << "\n");
assert(NewNodes.size() == 2 && "Expected a load folding node!");
N = NewNodes[1];
DAG->ReplaceAllUsesOfValueWith(SDValue(SU->getNode(), OldNumVals-1),
SDValue(LoadNode, 1));
- SUnit *NewSU = NewSUnit(N);
+ SUnit *NewSU = newSUnit(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) {
+
+ const MCInstrDesc &MCID = TII->get(N->getMachineOpcode());
+ for (unsigned i = 0; i != MCID.getNumOperands(); ++i) {
+ if (MCID.getOperandConstraint(i, MCOI::TIED_TO) != -1) {
NewSU->isTwoAddress = true;
break;
}
}
- if (TID.isCommutable())
+ if (MCID.isCommutable())
NewSU->isCommutable = true;
// LoadNode may already exist. This can happen when there is another
LoadSU = &SUnits[LoadNode->getNodeId()];
isNewLoad = false;
} else {
- LoadSU = NewSUnit(LoadNode);
+ LoadSU = newSUnit(LoadNode);
LoadNode->setNodeId(LoadSU->NodeNum);
}
D.setSUnit(LoadSU);
AddPred(SuccDep, D);
}
- }
+ }
if (isNewLoad) {
AddPred(NewSU, SDep(LoadSU, SDep::Order, LoadSU->Latency));
}
SU = NewSU;
}
- DEBUG(errs() << "Duplicating SU # " << SU->NodeNum << "\n");
+ DEBUG(dbgs() << "Duplicating SU # " << SU->NodeNum << "\n");
NewSU = Clone(SU);
// New SUnit has the exact same predecessors.
const TargetRegisterClass *DestRC,
const TargetRegisterClass *SrcRC,
SmallVector<SUnit*, 2> &Copies) {
- SUnit *CopyFromSU = NewSUnit(static_cast<SDNode *>(NULL));
+ SUnit *CopyFromSU = newSUnit(static_cast<SDNode *>(NULL));
CopyFromSU->CopySrcRC = SrcRC;
CopyFromSU->CopyDstRC = DestRC;
- SUnit *CopyToSU = NewSUnit(static_cast<SDNode *>(NULL));
+ SUnit *CopyToSU = newSUnit(static_cast<SDNode *>(NULL));
CopyToSU->CopySrcRC = DestRC;
CopyToSU->CopyDstRC = SrcRC;
/// FIXME: Move to SelectionDAG?
static EVT getPhysicalRegisterVT(SDNode *N, unsigned Reg,
const TargetInstrInfo *TII) {
- const TargetInstrDesc &TID = TII->get(N->getMachineOpcode());
- assert(TID.ImplicitDefs && "Physical reg def must be in implicit def list!");
- unsigned NumRes = TID.getNumDefs();
- for (const unsigned *ImpDef = TID.getImplicitDefs(); *ImpDef; ++ImpDef) {
+ 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 uint16_t *ImpDef = MCID.getImplicitDefs(); *ImpDef; ++ImpDef) {
if (Reg == *ImpDef)
break;
++NumRes;
return N->getValueType(NumRes);
}
+/// CheckForLiveRegDef - Return true and update live register vector if the
+/// specified register def of the specified SUnit clobbers any "live" registers.
+static bool CheckForLiveRegDef(SUnit *SU, unsigned Reg,
+ std::vector<SUnit*> &LiveRegDefs,
+ SmallSet<unsigned, 4> &RegAdded,
+ SmallVector<unsigned, 4> &LRegs,
+ const TargetRegisterInfo *TRI) {
+ bool Added = false;
+ for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI) {
+ if (LiveRegDefs[*AI] && LiveRegDefs[*AI] != SU) {
+ if (RegAdded.insert(*AI)) {
+ LRegs.push_back(*AI);
+ 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) ||
+ InlineAsm::isClobberKind(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)
+ const MCInstrDesc &MCID = TII->get(Node->getMachineOpcode());
+ if (!MCID.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);
- }
+ for (const uint16_t *Reg = MCID.getImplicitDefs(); *Reg; ++Reg) {
+ CheckForLiveRegDef(SU, *Reg, LiveRegDefs, RegAdded, LRegs, TRI);
}
}
return !LRegs.empty();
SUnit *LRDef = LiveRegDefs[Reg];
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
- // the value directly. Do not try duplicate the def.
+ // 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)
+ if (DestRC != RC) {
NewDef = CopyAndMoveSuccessors(LRDef);
- else
- DestRC = RC;
+ if (!DestRC && !NewDef)
+ report_fatal_error("Can't handle live physical "
+ "register dependency!");
+ }
if (!NewDef) {
// Issue copies, these can be expensive cross register class copies.
SmallVector<SUnit*, 2> Copies;
InsertCopiesAndMoveSuccs(LRDef, Reg, DestRC, RC, Copies);
- DEBUG(errs() << "Adding an edge from SU # " << TrySU->NodeNum
+ 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,
NewDef = Copies.back();
}
- DEBUG(errs() << "Adding an edge from SU # " << NewDef->NodeNum
+ 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,
std::reverse(Sequence.begin(), Sequence.end());
#ifndef NDEBUG
- VerifySchedule(/*isBottomUp=*/true);
+ VerifyScheduledSequence(/*isBottomUp=*/true);
#endif
}