STATISTIC(LoadsClustered, "Number of loads clustered together");
ScheduleDAGSDNodes::ScheduleDAGSDNodes(MachineFunction &mf)
- : ScheduleDAG(mf) {
-}
+ : ScheduleDAG(mf),
+ InstrItins(mf.getTarget().getInstrItineraryData()) {}
/// Run - perform scheduling.
///
SUnits.back().OrigNode = &SUnits.back();
SUnit *SU = &SUnits.back();
const TargetLowering &TLI = DAG->getTargetLoweringInfo();
- if (N->isMachineOpcode() &&
- N->getMachineOpcode() == TargetOpcode::IMPLICIT_DEF)
+ if (!N ||
+ (N->isMachineOpcode() &&
+ N->getMachineOpcode() == TargetOpcode::IMPLICIT_DEF))
SU->SchedulingPref = Sched::None;
else
SU->SchedulingPref = TLI.getSchedulingPreference(N);
II.ImplicitDefs[ResNo - II.getNumDefs()] == Reg) {
PhysReg = Reg;
const TargetRegisterClass *RC =
- TRI->getPhysicalRegisterRegClass(Reg, Def->getValueType(ResNo));
+ TRI->getMinimalPhysRegClass(Reg, Def->getValueType(ResNo));
Cost = RC->getCopyCost();
}
}
static void AddFlags(SDNode *N, SDValue Flag, bool AddFlag,
SelectionDAG *DAG) {
SmallVector<EVT, 4> VTs;
+ SDNode *FlagDestNode = Flag.getNode();
+
+ // Don't add a flag from a node to itself.
+ if (FlagDestNode == N) return;
+
+ // Don't add a flag to something which already has a flag.
+ if (N->getValueType(N->getNumValues() - 1) == MVT::Flag) return;
- for (unsigned i = 0, e = N->getNumValues(); i != e; ++i)
- VTs.push_back(N->getValueType(i));
+ for (unsigned I = 0, E = N->getNumValues(); I != E; ++I)
+ VTs.push_back(N->getValueType(I));
if (AddFlag)
VTs.push_back(MVT::Flag);
SmallVector<SDValue, 4> Ops;
- for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
- Ops.push_back(N->getOperand(i));
+ for (unsigned I = 0, E = N->getNumOperands(); I != E; ++I)
+ Ops.push_back(N->getOperand(I));
- if (Flag.getNode())
+ if (FlagDestNode)
Ops.push_back(Flag);
SDVTList VTList = DAG->getVTList(&VTs[0], VTs.size());
DenseMap<long long, SDNode*> O2SMap; // Map from offset to SDNode.
bool Cluster = false;
SDNode *Base = Node;
- int64_t BaseOffset;
for (SDNode::use_iterator I = Chain->use_begin(), E = Chain->use_end();
I != E; ++I) {
SDNode *User = *I;
Offsets.push_back(Offset1);
O2SMap.insert(std::make_pair(Offset2, User));
Offsets.push_back(Offset2);
- if (Offset2 < Offset1) {
+ if (Offset2 < Offset1)
Base = User;
- BaseOffset = Offset2;
- } else {
- BaseOffset = Offset1;
- }
Cluster = true;
}
// Cluster loads by adding MVT::Flag outputs and inputs. This also
// ensure they are scheduled in order of increasing addresses.
SDNode *Lead = Loads[0];
- AddFlags(Lead, SDValue(0,0), true, DAG);
+ AddFlags(Lead, SDValue(0, 0), true, DAG);
+
+ SDValue InFlag = SDValue(Lead, Lead->getNumValues() - 1);
+ for (unsigned I = 1, E = Loads.size(); I != E; ++I) {
+ bool OutFlag = I < E - 1;
+ SDNode *Load = Loads[I];
- SDValue InFlag = SDValue(Lead, Lead->getNumValues()-1);
- for (unsigned i = 1, e = Loads.size(); i != e; ++i) {
- bool OutFlag = i < e-1;
- SDNode *Load = Loads[i];
AddFlags(Load, InFlag, OutFlag, DAG);
if (OutFlag)
- InFlag = SDValue(Load, Load->getNumValues()-1);
+ InFlag = SDValue(Load, Load->getNumValues() - 1);
++LoadsClustered;
}
return;
}
- const InstrItineraryData &InstrItins = TM.getInstrItineraryData();
- if (InstrItins.isEmpty()) {
+ if (!InstrItins || InstrItins->isEmpty()) {
SU->Latency = 1;
return;
}
SU->Latency = 0;
for (SDNode *N = SU->getNode(); N; N = N->getFlaggedNode())
if (N->isMachineOpcode()) {
- SU->Latency += InstrItins.
+ SU->Latency += InstrItins->
getStageLatency(TII->get(N->getMachineOpcode()).getSchedClass());
}
}
if (ForceUnitLatencies())
return;
- const InstrItineraryData &InstrItins = TM.getInstrItineraryData();
- if (InstrItins.isEmpty())
- return;
-
if (dep.getKind() != SDep::Data)
return;
unsigned DefIdx = Use->getOperand(OpIdx).getResNo();
- if (Def->isMachineOpcode()) {
- const TargetInstrDesc &II = TII->get(Def->getMachineOpcode());
- if (DefIdx >= II.getNumDefs())
- return;
- int DefCycle = InstrItins.getOperandCycle(II.getSchedClass(), DefIdx);
- if (DefCycle < 0)
- return;
- int UseCycle = 1;
- if (Use->isMachineOpcode()) {
- const unsigned UseClass = TII->get(Use->getMachineOpcode()).getSchedClass();
- UseCycle = InstrItins.getOperandCycle(UseClass, OpIdx);
- }
- if (UseCycle >= 0) {
- int Latency = DefCycle - UseCycle + 1;
- if (Latency >= 0)
- dep.setLatency(Latency);
- }
+ if (Use->isMachineOpcode())
+ // Adjust the use operand index by num of defs.
+ OpIdx += TII->get(Use->getMachineOpcode()).getNumDefs();
+ int Latency = TII->getOperandLatency(InstrItins, Def, DefIdx, Use, OpIdx);
+ if (Latency > 1 && Use->getOpcode() == ISD::CopyToReg &&
+ !BB->succ_empty()) {
+ unsigned Reg = cast<RegisterSDNode>(Use->getOperand(1))->getReg();
+ if (TargetRegisterInfo::isVirtualRegister(Reg))
+ // This copy is a liveout value. It is likely coalesced, so reduce the
+ // latency so not to penalize the def.
+ // FIXME: need target specific adjustment here?
+ Latency = (Latency > 1) ? Latency - 1 : 1;
}
+ if (Latency >= 0)
+ dep.setLatency(Latency);
}
void ScheduleDAGSDNodes::dumpNode(const SUnit *SU) const {
}
// ProcessSourceNode - Process nodes with source order numbers. These are added
-// to a vector which EmitSchedule use to determine how to insert dbg_value
+// to a vector which EmitSchedule uses to determine how to insert dbg_value
// instructions in the right order.
static void ProcessSourceNode(SDNode *N, SelectionDAG *DAG,
InstrEmitter &Emitter,
return;
MachineBasicBlock *BB = Emitter.getBlock();
- if (BB->empty() || BB->back().isPHI()) {
+ if (Emitter.getInsertPos() == BB->begin() || BB->back().isPHI()) {
// Did not insert any instruction.
Orders.push_back(std::make_pair(Order, (MachineInstr*)0));
return;
}
- Orders.push_back(std::make_pair(Order, &BB->back()));
+ Orders.push_back(std::make_pair(Order, prior(Emitter.getInsertPos())));
if (!N->getHasDebugValue())
return;
// Opportunistically insert immediate dbg_value uses, i.e. those with source
for (; PDI != PDE; ++PDI) {
MachineInstr *DbgMI= Emitter.EmitDbgValue(*PDI, VRBaseMap);
if (DbgMI)
- BB->push_back(DbgMI);
+ BB->insert(InsertPos, DbgMI);
}
}
// Insert all the dbg_values which have not already been inserted in source
// order sequence.
if (HasDbg) {
- MachineBasicBlock::iterator BBBegin = BB->empty() ? BB->end() : BB->begin();
- while (BBBegin != BB->end() && BBBegin->isPHI())
- ++BBBegin;
+ MachineBasicBlock::iterator BBBegin = BB->getFirstNonPHI();
// Sort the source order instructions and use the order to insert debug
// values.
SDDbgInfo::DbgIterator DE = DAG->DbgEnd();
// Now emit the rest according to source order.
unsigned LastOrder = 0;
- MachineInstr *LastMI = 0;
for (unsigned i = 0, e = Orders.size(); i != e && DI != DE; ++i) {
unsigned Order = Orders[i].first;
MachineInstr *MI = Orders[i].second;
// Insert all SDDbgValue's whose order(s) are before "Order".
if (!MI)
continue;
- MachineBasicBlock *MIBB = MI->getParent();
#ifndef NDEBUG
unsigned LastDIOrder = 0;
#endif
// Insert to start of the BB (after PHIs).
BB->insert(BBBegin, DbgMI);
else {
+ // Insert at the instruction, which may be in a different
+ // block, if the block was split by a custom inserter.
MachineBasicBlock::iterator Pos = MI;
- MIBB->insert(llvm::next(Pos), DbgMI);
+ MI->getParent()->insert(llvm::next(Pos), DbgMI);
}
}
}
LastOrder = Order;
- LastMI = MI;
}
// Add trailing DbgValue's before the terminator. FIXME: May want to add
// some of them before one or more conditional branches?