ScheduleDAGInstrs::ScheduleDAGInstrs(MachineFunction &mf,
const MachineLoopInfo &mli,
const MachineDominatorTree &mdt)
- : ScheduleDAG(mf), MLI(mli), MDT(mdt), LoopRegs(MLI, MDT) {
- MFI = mf.getFrameInfo();
+ : ScheduleDAG(mf), MLI(mli), MDT(mdt), MFI(mf.getFrameInfo()),
+ InstrItins(mf.getTarget().getInstrItineraryData()),
+ Defs(TRI->getNumRegs()), Uses(TRI->getNumRegs()), LoopRegs(MLI, MDT) {
+ DbgValueVec.clear();
}
/// Run - perform scheduling.
}
}
+/// AddSchedBarrierDeps - Add dependencies from instructions in the current
+/// list of instructions being scheduled to scheduling barrier by adding
+/// the exit SU to the register defs and use list. This is because we want to
+/// make sure instructions which define registers that are either used by
+/// the terminator or are live-out are properly scheduled. This is
+/// especially important when the definition latency of the return value(s)
+/// are too high to be hidden by the branch or when the liveout registers
+/// used by instructions in the fallthrough block.
+void ScheduleDAGInstrs::AddSchedBarrierDeps() {
+ MachineInstr *ExitMI = InsertPos != BB->end() ? &*InsertPos : 0;
+ ExitSU.setInstr(ExitMI);
+ bool AllDepKnown = ExitMI &&
+ (ExitMI->getDesc().isCall() || ExitMI->getDesc().isBarrier());
+ if (ExitMI && AllDepKnown) {
+ // If it's a call or a barrier, add dependencies on the defs and uses of
+ // instruction.
+ for (unsigned i = 0, e = ExitMI->getNumOperands(); i != e; ++i) {
+ const MachineOperand &MO = ExitMI->getOperand(i);
+ if (!MO.isReg() || MO.isDef()) continue;
+ unsigned Reg = MO.getReg();
+ if (Reg == 0) continue;
+
+ assert(TRI->isPhysicalRegister(Reg) && "Virtual register encountered!");
+ Uses[Reg].push_back(&ExitSU);
+ }
+ } else {
+ // For others, e.g. fallthrough, conditional branch, assume the exit
+ // uses all the registers that are livein to the successor blocks.
+ SmallSet<unsigned, 8> Seen;
+ for (MachineBasicBlock::succ_iterator SI = BB->succ_begin(),
+ SE = BB->succ_end(); SI != SE; ++SI)
+ for (MachineBasicBlock::livein_iterator I = (*SI)->livein_begin(),
+ E = (*SI)->livein_end(); I != E; ++I) {
+ unsigned Reg = *I;
+ if (Seen.insert(Reg))
+ Uses[Reg].push_back(&ExitSU);
+ }
+ }
+}
+
void ScheduleDAGInstrs::BuildSchedGraph(AliasAnalysis *AA) {
// We'll be allocating one SUnit for each instruction, plus one for
// the region exit node.
std::map<const Value *, SUnit *> AliasMemDefs, NonAliasMemDefs;
std::map<const Value *, std::vector<SUnit *> > AliasMemUses, NonAliasMemUses;
+ // Keep track of dangling debug references to registers.
+ std::vector<std::pair<MachineInstr*, unsigned> >
+ DanglingDebugValue(TRI->getNumRegs(),
+ std::make_pair(static_cast<MachineInstr*>(0), 0));
+
// Check to see if the scheduler cares about latencies.
bool UnitLatencies = ForceUnitLatencies();
const TargetSubtarget &ST = TM.getSubtarget<TargetSubtarget>();
unsigned SpecialAddressLatency = ST.getSpecialAddressLatency();
+ // Remove any stale debug info; sometimes BuildSchedGraph is called again
+ // without emitting the info from the previous call.
+ DbgValueVec.clear();
+
+ // Model data dependencies between instructions being scheduled and the
+ // ExitSU.
+ AddSchedBarrierDeps();
+
// Walk the list of instructions, from bottom moving up.
for (MachineBasicBlock::iterator MII = InsertPos, MIE = Begin;
MII != MIE; --MII) {
MachineInstr *MI = prior(MII);
+ // DBG_VALUE does not have SUnit's built, so just remember these for later
+ // reinsertion.
+ if (MI->isDebugValue()) {
+ if (MI->getNumOperands()==3 && MI->getOperand(0).isReg() &&
+ MI->getOperand(0).getReg())
+ DanglingDebugValue[MI->getOperand(0).getReg()] =
+ std::make_pair(MI, DbgValueVec.size());
+ DbgValueVec.push_back(MI);
+ continue;
+ }
const TargetInstrDesc &TID = MI->getDesc();
assert(!TID.isTerminator() && !MI->isLabel() &&
"Cannot schedule terminators or labels!");
// Create the SUnit for this MI.
SUnit *SU = NewSUnit(MI);
+ SU->isCall = TID.isCall();
+ SU->isCommutable = TID.isCommutable();
// Assign the Latency field of SU using target-provided information.
if (UnitLatencies)
if (Reg == 0) continue;
assert(TRI->isPhysicalRegister(Reg) && "Virtual register encountered!");
+
+ if (MO.isDef() && DanglingDebugValue[Reg].first!=0) {
+ SU->DbgInstrList.push_back(DanglingDebugValue[Reg].first);
+ DbgValueVec[DanglingDebugValue[Reg].second] = 0;
+ DanglingDebugValue[Reg] = std::make_pair((MachineInstr*)0, 0);
+ }
+
std::vector<SUnit *> &UseList = Uses[Reg];
std::vector<SUnit *> &DefList = Defs[Reg];
// Optionally add output and anti dependencies. For anti
unsigned AOLatency = (Kind == SDep::Anti) ? 0 : 1;
for (unsigned i = 0, e = DefList.size(); i != e; ++i) {
SUnit *DefSU = DefList[i];
+ if (DefSU == &ExitSU)
+ continue;
if (DefSU != SU &&
(Kind != SDep::Output || !MO.isDead() ||
!DefSU->getInstr()->registerDefIsDead(Reg)))
std::vector<SUnit *> &DefList = Defs[*Alias];
for (unsigned i = 0, e = DefList.size(); i != e; ++i) {
SUnit *DefSU = DefList[i];
+ if (DefSU == &ExitSU)
+ continue;
if (DefSU != SU &&
(Kind != SDep::Output || !MO.isDead() ||
!DefSU->getInstr()->registerDefIsDead(*Alias)))
unsigned DataLatency = SU->Latency;
for (unsigned i = 0, e = UseList.size(); i != e; ++i) {
SUnit *UseSU = UseList[i];
- if (UseSU != SU) {
- unsigned LDataLatency = DataLatency;
- // Optionally add in a special extra latency for nodes that
- // feed addresses.
- // TODO: Do this for register aliases too.
- // TODO: Perhaps we should get rid of
- // SpecialAddressLatency and just move this into
- // adjustSchedDependency for the targets that care about
- // it.
- if (SpecialAddressLatency != 0 && !UnitLatencies) {
- MachineInstr *UseMI = UseSU->getInstr();
- const TargetInstrDesc &UseTID = UseMI->getDesc();
- int RegUseIndex = UseMI->findRegisterUseOperandIdx(Reg);
- assert(RegUseIndex >= 0 && "UseMI doesn's use register!");
- if ((UseTID.mayLoad() || UseTID.mayStore()) &&
- (unsigned)RegUseIndex < UseTID.getNumOperands() &&
- UseTID.OpInfo[RegUseIndex].isLookupPtrRegClass())
- LDataLatency += SpecialAddressLatency;
- }
- // Adjust the dependence latency using operand def/use
- // information (if any), and then allow the target to
- // perform its own adjustments.
- const SDep& dep = SDep(SU, SDep::Data, LDataLatency, Reg);
- if (!UnitLatencies) {
- ComputeOperandLatency(SU, UseSU, (SDep &)dep);
- ST.adjustSchedDependency(SU, UseSU, (SDep &)dep);
- }
- UseSU->addPred(dep);
+ if (UseSU == SU)
+ continue;
+ unsigned LDataLatency = DataLatency;
+ // Optionally add in a special extra latency for nodes that
+ // feed addresses.
+ // TODO: Do this for register aliases too.
+ // TODO: Perhaps we should get rid of
+ // SpecialAddressLatency and just move this into
+ // adjustSchedDependency for the targets that care about it.
+ if (SpecialAddressLatency != 0 && !UnitLatencies &&
+ UseSU != &ExitSU) {
+ MachineInstr *UseMI = UseSU->getInstr();
+ const TargetInstrDesc &UseTID = UseMI->getDesc();
+ int RegUseIndex = UseMI->findRegisterUseOperandIdx(Reg);
+ assert(RegUseIndex >= 0 && "UseMI doesn's use register!");
+ if (RegUseIndex >= 0 &&
+ (UseTID.mayLoad() || UseTID.mayStore()) &&
+ (unsigned)RegUseIndex < UseTID.getNumOperands() &&
+ UseTID.OpInfo[RegUseIndex].isLookupPtrRegClass())
+ LDataLatency += SpecialAddressLatency;
+ }
+ // Adjust the dependence latency using operand def/use
+ // information (if any), and then allow the target to
+ // perform its own adjustments.
+ const SDep& dep = SDep(SU, SDep::Data, LDataLatency, Reg);
+ if (!UnitLatencies) {
+ ComputeOperandLatency(SU, UseSU, const_cast<SDep &>(dep));
+ ST.adjustSchedDependency(SU, UseSU, const_cast<SDep &>(dep));
}
+ UseSU->addPred(dep);
}
for (const unsigned *Alias = TRI->getAliasSet(Reg); *Alias; ++Alias) {
std::vector<SUnit *> &UseList = Uses[*Alias];
for (unsigned i = 0, e = UseList.size(); i != e; ++i) {
SUnit *UseSU = UseList[i];
- if (UseSU != SU) {
- const SDep& dep = SDep(SU, SDep::Data, DataLatency, *Alias);
- if (!UnitLatencies) {
- ComputeOperandLatency(SU, UseSU, (SDep &)dep);
- ST.adjustSchedDependency(SU, UseSU, (SDep &)dep);
- }
- UseSU->addPred(dep);
+ if (UseSU == SU)
+ continue;
+ const SDep& dep = SDep(SU, SDep::Data, DataLatency, *Alias);
+ if (!UnitLatencies) {
+ ComputeOperandLatency(SU, UseSU, const_cast<SDep &>(dep));
+ ST.adjustSchedDependency(SU, UseSU, const_cast<SDep &>(dep));
}
+ UseSU->addPred(dep);
}
}
// Treat all other stores conservatively.
goto new_alias_chain;
}
+
+ if (!ExitSU.isPred(SU))
+ // Push store's up a bit to avoid them getting in between cmp
+ // and branches.
+ ExitSU.addPred(SDep(SU, SDep::Order, 0,
+ /*Reg=*/0, /*isNormalMemory=*/false,
+ /*isMustAlias=*/false,
+ /*isArtificial=*/true));
} else if (TID.mayLoad()) {
bool MayAlias = true;
TrueMemOrderLatency = 0;
}
void ScheduleDAGInstrs::ComputeLatency(SUnit *SU) {
- const InstrItineraryData &InstrItins = TM.getInstrItineraryData();
-
// Compute the latency for the node.
- SU->Latency =
- InstrItins.getStageLatency(SU->getInstr()->getDesc().getSchedClass());
+ if (!InstrItins || InstrItins->isEmpty()) {
+ SU->Latency = 1;
- // Simplistic target-independent heuristic: assume that loads take
- // extra time.
- if (InstrItins.isEmpty())
+ // Simplistic target-independent heuristic: assume that loads take
+ // extra time.
if (SU->getInstr()->getDesc().mayLoad())
SU->Latency += 2;
+ } else {
+ SU->Latency = TII->getInstrLatency(InstrItins, SU->getInstr());
+ }
}
void ScheduleDAGInstrs::ComputeOperandLatency(SUnit *Def, SUnit *Use,
SDep& dep) const {
- const InstrItineraryData &InstrItins = TM.getInstrItineraryData();
- if (InstrItins.isEmpty())
+ if (!InstrItins || InstrItins->isEmpty())
return;
// For a data dependency with a known register...
MachineInstr *DefMI = Def->getInstr();
int DefIdx = DefMI->findRegisterDefOperandIdx(Reg);
if (DefIdx != -1) {
- int DefCycle = InstrItins.getOperandCycle(DefMI->getDesc().getSchedClass(), DefIdx);
- if (DefCycle >= 0) {
- MachineInstr *UseMI = Use->getInstr();
- const unsigned UseClass = UseMI->getDesc().getSchedClass();
-
- // For all uses of the register, calculate the maxmimum latency
- int Latency = -1;
+ const MachineOperand &MO = DefMI->getOperand(DefIdx);
+ if (MO.isReg() && MO.isImplicit() &&
+ DefIdx >= (int)DefMI->getDesc().getNumOperands()) {
+ // This is an implicit def, getOperandLatency() won't return the correct
+ // latency. e.g.
+ // %D6<def>, %D7<def> = VLD1q16 %R2<kill>, 0, ..., %Q3<imp-def>
+ // %Q1<def> = VMULv8i16 %Q1<kill>, %Q3<kill>, ...
+ // What we want is to compute latency between def of %D6/%D7 and use of
+ // %Q3 instead.
+ DefIdx = DefMI->findRegisterDefOperandIdx(Reg, false, true, TRI);
+ }
+ MachineInstr *UseMI = Use->getInstr();
+ // For all uses of the register, calculate the maxmimum latency
+ int Latency = -1;
+ if (UseMI) {
for (unsigned i = 0, e = UseMI->getNumOperands(); i != e; ++i) {
const MachineOperand &MO = UseMI->getOperand(i);
if (!MO.isReg() || !MO.isUse())
if (MOReg != Reg)
continue;
- int UseCycle = InstrItins.getOperandCycle(UseClass, i);
- if (UseCycle >= 0)
- Latency = std::max(Latency, DefCycle - UseCycle + 1);
+ int UseCycle = TII->getOperandLatency(InstrItins, DefMI, DefIdx,
+ UseMI, i);
+ Latency = std::max(Latency, UseCycle);
}
-
- // If we found a latency, then replace the existing dependence latency.
- if (Latency >= 0)
- dep.setLatency(Latency);
+ } else {
+ // UseMI is null, then it must be a scheduling barrier.
+ if (!InstrItins || InstrItins->isEmpty())
+ return;
+ unsigned DefClass = DefMI->getDesc().getSchedClass();
+ Latency = InstrItins->getOperandCycle(DefClass, DefIdx);
}
+
+ // If we found a latency, then replace the existing dependence latency.
+ if (Latency >= 0)
+ dep.setLatency(Latency);
}
}
}
// EmitSchedule - Emit the machine code in scheduled order.
-MachineBasicBlock *ScheduleDAGInstrs::
-EmitSchedule(DenseMap<MachineBasicBlock*, MachineBasicBlock*> *EM) {
+MachineBasicBlock *ScheduleDAGInstrs::EmitSchedule() {
// For MachineInstr-based scheduling, we're rescheduling the instructions in
// the block, so start by removing them from the block.
while (Begin != InsertPos) {
BB->remove(I);
}
+ // First reinsert any remaining debug_values; these are either constants,
+ // or refer to live-in registers. The beginning of the block is the right
+ // place for the latter. The former might reasonably be placed elsewhere
+ // using some kind of ordering algorithm, but right now it doesn't matter.
+ for (int i = DbgValueVec.size()-1; i>=0; --i)
+ if (DbgValueVec[i])
+ BB->insert(InsertPos, DbgValueVec[i]);
+
// Then re-insert them according to the given schedule.
for (unsigned i = 0, e = Sequence.size(); i != e; i++) {
SUnit *SU = Sequence[i];
}
BB->insert(InsertPos, SU->getInstr());
+ for (unsigned i = 0, e = SU->DbgInstrList.size() ; i < e ; ++i)
+ BB->insert(InsertPos, SU->DbgInstrList[i]);
}
// Update the Begin iterator, as the first instruction in the block
// may have been scheduled later.
- if (!Sequence.empty())
+ if (!DbgValueVec.empty()) {
+ for (int i = DbgValueVec.size()-1; i>=0; --i)
+ if (DbgValueVec[i]!=0) {
+ Begin = DbgValueVec[DbgValueVec.size()-1];
+ break;
+ }
+ } else if (!Sequence.empty())
Begin = Sequence[0]->getInstr();
+ DbgValueVec.clear();
return BB;
}