const MachineDominatorTree &mdt,
bool IsPostRAFlag,
LiveIntervals *lis)
- : ScheduleDAG(mf), MLI(mli), MDT(mdt), MFI(mf.getFrameInfo()),
- InstrItins(mf.getTarget().getInstrItineraryData()), LIS(lis),
- IsPostRA(IsPostRAFlag), UnitLatencies(false), CanHandleTerminators(false),
- LoopRegs(MDT), FirstDbgValue(0) {
+ : ScheduleDAG(mf), MLI(mli), MDT(mdt), MFI(mf.getFrameInfo()), LIS(lis),
+ IsPostRA(IsPostRAFlag), CanHandleTerminators(false), FirstDbgValue(0) {
assert((IsPostRA || LIS) && "PreRA scheduling requires LiveIntervals");
DbgValues.clear();
assert(!(IsPostRA && MRI.getNumVirtRegs()) &&
"Virtual registers must be removed prior to PostRA scheduling");
+
+ const TargetSubtargetInfo &ST = TM.getSubtarget<TargetSubtargetInfo>();
+ SchedModel.init(*ST.getSchedModel(), &ST, TII);
}
/// getUnderlyingObjectFromInt - This is the function that does the work of
void ScheduleDAGInstrs::startBlock(MachineBasicBlock *bb) {
BB = bb;
- LoopRegs.Deps.clear();
- if (MachineLoop *ML = MLI.getLoopFor(BB))
- if (BB == ML->getLoopLatch())
- LoopRegs.VisitLoop(ML);
}
void ScheduleDAGInstrs::finishBlock() {
EndIndex = endcount;
MISUnitMap.clear();
- // Check to see if the scheduler cares about latencies.
- UnitLatencies = forceUnitLatencies();
-
ScheduleDAG::clearDAG();
}
if (Reg == 0) continue;
if (TRI->isPhysicalRegister(Reg))
- Uses[Reg].push_back(&ExitSU);
+ Uses[Reg].push_back(PhysRegSUOper(&ExitSU, -1));
else {
assert(!IsPostRA && "Virtual register encountered after regalloc.");
addVRegUseDeps(&ExitSU, i);
E = (*SI)->livein_end(); I != E; ++I) {
unsigned Reg = *I;
if (!Uses.contains(Reg))
- Uses[Reg].push_back(&ExitSU);
+ Uses[Reg].push_back(PhysRegSUOper(&ExitSU, -1));
}
}
}
/// MO is an operand of SU's instruction that defines a physical register. Add
/// data dependencies from SU to any uses of the physical register.
-void ScheduleDAGInstrs::addPhysRegDataDeps(SUnit *SU,
- const MachineOperand &MO) {
+void ScheduleDAGInstrs::addPhysRegDataDeps(SUnit *SU, unsigned OperIdx) {
+ const MachineOperand &MO = SU->getInstr()->getOperand(OperIdx);
assert(MO.isDef() && "expect physreg def");
// Ask the target if address-backscheduling is desirable, and if so how much.
const TargetSubtargetInfo &ST = TM.getSubtarget<TargetSubtargetInfo>();
- unsigned SpecialAddressLatency = ST.getSpecialAddressLatency();
- unsigned DataLatency = SU->Latency;
for (MCRegAliasIterator Alias(MO.getReg(), TRI, true);
Alias.isValid(); ++Alias) {
if (!Uses.contains(*Alias))
continue;
- std::vector<SUnit*> &UseList = Uses[*Alias];
+ std::vector<PhysRegSUOper> &UseList = Uses[*Alias];
for (unsigned i = 0, e = UseList.size(); i != e; ++i) {
- SUnit *UseSU = UseList[i];
+ SUnit *UseSU = UseList[i].SU;
if (UseSU == SU)
continue;
- unsigned LDataLatency = DataLatency;
- // Optionally add in a special extra latency for nodes that
- // feed addresses.
- // 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 MCInstrDesc &UseMCID = UseMI->getDesc();
- int RegUseIndex = UseMI->findRegisterUseOperandIdx(*Alias);
- assert(RegUseIndex >= 0 && "UseMI doesn't use register!");
- if (RegUseIndex >= 0 &&
- (UseMI->mayLoad() || UseMI->mayStore()) &&
- (unsigned)RegUseIndex < UseMCID.getNumOperands() &&
- UseMCID.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.
- SDep dep(SU, SDep::Data, LDataLatency, *Alias);
- if (!UnitLatencies) {
- unsigned Latency = computeOperandLatency(SU, UseSU, dep);
- dep.setLatency(Latency);
-
- ST.adjustSchedDependency(SU, UseSU, dep);
- }
+
+ SDep dep(SU, SDep::Data, 1, *Alias);
+
+ // Adjust the dependence latency using operand def/use information,
+ // then allow the target to perform its own adjustments.
+ int UseOp = UseList[i].OpIdx;
+ MachineInstr *RegUse = UseOp < 0 ? 0 : UseSU->getInstr();
+ dep.setLatency(
+ SchedModel.computeOperandLatency(SU->getInstr(), OperIdx,
+ RegUse, UseOp, /*FindMin=*/false));
+ dep.setMinLatency(
+ SchedModel.computeOperandLatency(SU->getInstr(), OperIdx,
+ RegUse, UseOp, /*FindMin=*/true));
+
+ ST.adjustSchedDependency(SU, UseSU, dep);
UseSU->addPred(dep);
}
}
Alias.isValid(); ++Alias) {
if (!Defs.contains(*Alias))
continue;
- std::vector<SUnit *> &DefList = Defs[*Alias];
+ std::vector<PhysRegSUOper> &DefList = Defs[*Alias];
for (unsigned i = 0, e = DefList.size(); i != e; ++i) {
- SUnit *DefSU = DefList[i];
+ SUnit *DefSU = DefList[i].SU;
if (DefSU == &ExitSU)
continue;
if (DefSU != SU &&
if (Kind == SDep::Anti)
DefSU->addPred(SDep(SU, Kind, 0, /*Reg=*/*Alias));
else {
- unsigned AOLat = TII->getOutputLatency(InstrItins, MI, OperIdx,
- DefSU->getInstr());
+ unsigned AOLat =
+ SchedModel.computeOutputLatency(MI, OperIdx, DefSU->getInstr());
DefSU->addPred(SDep(SU, Kind, AOLat, /*Reg=*/*Alias));
}
}
// Either insert a new Reg2SUnits entry with an empty SUnits list, or
// retrieve the existing SUnits list for this register's uses.
// Push this SUnit on the use list.
- Uses[MO.getReg()].push_back(SU);
+ Uses[MO.getReg()].push_back(PhysRegSUOper(SU, OperIdx));
}
else {
- addPhysRegDataDeps(SU, MO);
+ addPhysRegDataDeps(SU, OperIdx);
// Either insert a new Reg2SUnits entry with an empty SUnits list, or
// retrieve the existing SUnits list for this register's defs.
- std::vector<SUnit *> &DefList = Defs[MO.getReg()];
-
- // If a def is going to wrap back around to the top of the loop,
- // backschedule it.
- if (!UnitLatencies && DefList.empty()) {
- LoopDependencies::LoopDeps::iterator I = LoopRegs.Deps.find(MO.getReg());
- if (I != LoopRegs.Deps.end()) {
- const MachineOperand *UseMO = I->second.first;
- unsigned Count = I->second.second;
- const MachineInstr *UseMI = UseMO->getParent();
- unsigned UseMOIdx = UseMO - &UseMI->getOperand(0);
- const MCInstrDesc &UseMCID = UseMI->getDesc();
- const TargetSubtargetInfo &ST =
- TM.getSubtarget<TargetSubtargetInfo>();
- unsigned SpecialAddressLatency = ST.getSpecialAddressLatency();
- // TODO: If we knew the total depth of the region here, we could
- // handle the case where the whole loop is inside the region but
- // is large enough that the isScheduleHigh trick isn't needed.
- if (UseMOIdx < UseMCID.getNumOperands()) {
- // Currently, we only support scheduling regions consisting of
- // single basic blocks. Check to see if the instruction is in
- // the same region by checking to see if it has the same parent.
- if (UseMI->getParent() != MI->getParent()) {
- unsigned Latency = SU->Latency;
- if (UseMCID.OpInfo[UseMOIdx].isLookupPtrRegClass())
- Latency += SpecialAddressLatency;
- // This is a wild guess as to the portion of the latency which
- // will be overlapped by work done outside the current
- // scheduling region.
- Latency -= std::min(Latency, Count);
- // Add the artificial edge.
- ExitSU.addPred(SDep(SU, SDep::Order, Latency,
- /*Reg=*/0, /*isNormalMemory=*/false,
- /*isMustAlias=*/false,
- /*isArtificial=*/true));
- } else if (SpecialAddressLatency > 0 &&
- UseMCID.OpInfo[UseMOIdx].isLookupPtrRegClass()) {
- // The entire loop body is within the current scheduling region
- // and the latency of this operation is assumed to be greater
- // than the latency of the loop.
- // TODO: Recursively mark data-edge predecessors as
- // isScheduleHigh too.
- SU->isScheduleHigh = true;
- }
- }
- LoopRegs.Deps.erase(I);
- }
- }
+ std::vector<PhysRegSUOper> &DefList = Defs[MO.getReg()];
// clear this register's use list
if (Uses.contains(MO.getReg()))
// the block. Instead, we leave only one call at the back of the
// DefList.
if (SU->isCall) {
- while (!DefList.empty() && DefList.back()->isCall)
+ while (!DefList.empty() && DefList.back().SU->isCall)
DefList.pop_back();
}
// Defs are pushed in the order they are visited and never reordered.
- DefList.push_back(SU);
+ DefList.push_back(PhysRegSUOper(SU, OperIdx));
}
}
const MachineInstr *MI = SU->getInstr();
unsigned Reg = MI->getOperand(OperIdx).getReg();
- // SSA defs do not have output/anti dependencies.
+ // Singly defined vregs do not have output/anti dependencies.
// The current operand is a def, so we have at least one.
- if (llvm::next(MRI.def_begin(Reg)) == MRI.def_end())
+ // Check here if there are any others...
+ if (MRI.hasOneDef(Reg))
return;
// Add output dependence to the next nearest def of this vreg.
else {
SUnit *DefSU = DefI->SU;
if (DefSU != SU && DefSU != &ExitSU) {
- unsigned OutLatency = TII->getOutputLatency(InstrItins, MI, OperIdx,
- DefSU->getInstr());
+ unsigned OutLatency =
+ SchedModel.computeOutputLatency(MI, OperIdx, DefSU->getInstr());
DefSU->addPred(SDep(SU, SDep::Output, OutLatency, Reg));
}
DefI->SU = SU;
if (DefSU) {
// The reaching Def lives within this scheduling region.
// Create a data dependence.
- //
- // TODO: Handle "special" address latencies cleanly.
- SDep dep(DefSU, SDep::Data, DefSU->Latency, Reg);
- if (!UnitLatencies) {
- // Adjust the dependence latency using operand def/use information, then
- // allow the target to perform its own adjustments.
- unsigned Latency = computeOperandLatency(DefSU, SU, const_cast<SDep &>(dep));
- dep.setLatency(Latency);
-
- const TargetSubtargetInfo &ST = TM.getSubtarget<TargetSubtargetInfo>();
- ST.adjustSchedDependency(DefSU, SU, const_cast<SDep &>(dep));
- }
+ SDep dep(DefSU, SDep::Data, 1, Reg);
+ // Adjust the dependence latency using operand def/use information, then
+ // allow the target to perform its own adjustments.
+ int DefOp = Def->findRegisterDefOperandIdx(Reg);
+ dep.setLatency(
+ SchedModel.computeOperandLatency(Def, DefOp, MI, OperIdx, false));
+ dep.setMinLatency(
+ SchedModel.computeOperandLatency(Def, DefOp, MI, OperIdx, true));
+
+ const TargetSubtargetInfo &ST = TM.getSubtarget<TargetSubtargetInfo>();
+ ST.adjustSchedDependency(DefSU, SU, const_cast<SDep &>(dep));
SU->addPred(dep);
}
}
/// (like a call or something with unmodeled side effects).
static inline bool isGlobalMemoryObject(AliasAnalysis *AA, MachineInstr *MI) {
if (MI->isCall() || MI->hasUnmodeledSideEffects() ||
- (MI->hasVolatileMemoryRef() &&
+ (MI->hasOrderedMemoryRef() &&
(!MI->mayLoad() || !MI->isInvariantLoad(AA))))
return true;
return false;
/// checks whether SU can be aliasing any node dominated
/// by it.
static void adjustChainDeps(AliasAnalysis *AA, const MachineFrameInfo *MFI,
- SUnit *SU, SUnit *ExitSU, std::set<SUnit *> &CheckList) {
+ SUnit *SU, SUnit *ExitSU, std::set<SUnit *> &CheckList,
+ unsigned LatencyToLoad) {
if (!SU)
return;
I != IE; ++I) {
if (SU == *I)
continue;
- if (MIsNeedChainEdge(AA, MFI, SU->getInstr(), (*I)->getInstr()))
- (*I)->addPred(SDep(SU, SDep::Order, /*Latency=*/0, /*Reg=*/0,
+ if (MIsNeedChainEdge(AA, MFI, SU->getInstr(), (*I)->getInstr())) {
+ unsigned Latency = ((*I)->getInstr()->mayLoad()) ? LatencyToLoad : 0;
+ (*I)->addPred(SDep(SU, SDep::Order, Latency, /*Reg=*/0,
/*isNormalMemory=*/true));
+ }
// Now go through all the chain successors and iterate from them.
// Keep track of visited nodes.
for (SUnit::const_succ_iterator J = (*I)->Succs.begin(),
SU->isCommutable = MI->isCommutable();
// Assign the Latency field of SU using target-provided information.
- if (UnitLatencies)
- SU->Latency = 1;
- else
- computeLatency(SU);
+ SU->Latency = SchedModel.computeInstrLatency(SU->getInstr());
}
}
// after stack slots are lowered to actual addresses.
// TODO: Use an AliasAnalysis and do real alias-analysis queries, and
// produce more precise dependence information.
-#define STORE_LOAD_LATENCY 1
- unsigned TrueMemOrderLatency = 0;
+ unsigned TrueMemOrderLatency = MI->mayStore() ? 1 : 0;
if (isGlobalMemoryObject(AA, MI)) {
// Be conservative with these and add dependencies on all memory
// references, even those that are known to not alias.
BarrierChain = SU;
// This is a barrier event that acts as a pivotal node in the DAG,
// so it is safe to clear list of exposed nodes.
- adjustChainDeps(AA, MFI, SU, &ExitSU, RejectMemNodes);
+ adjustChainDeps(AA, MFI, SU, &ExitSU, RejectMemNodes,
+ TrueMemOrderLatency);
RejectMemNodes.clear();
NonAliasMemDefs.clear();
NonAliasMemUses.clear();
// fall-through
new_alias_chain:
// Chain all possibly aliasing memory references though SU.
- if (AliasChain)
- addChainDependency(AA, MFI, SU, AliasChain, RejectMemNodes);
+ if (AliasChain) {
+ unsigned ChainLatency = 0;
+ if (AliasChain->getInstr()->mayLoad())
+ ChainLatency = TrueMemOrderLatency;
+ addChainDependency(AA, MFI, SU, AliasChain, RejectMemNodes,
+ ChainLatency);
+ }
AliasChain = SU;
for (unsigned k = 0, m = PendingLoads.size(); k != m; ++k)
addChainDependency(AA, MFI, SU, PendingLoads[k], RejectMemNodes,
addChainDependency(AA, MFI, SU, I->second[i], RejectMemNodes,
TrueMemOrderLatency);
}
- adjustChainDeps(AA, MFI, SU, &ExitSU, RejectMemNodes);
+ adjustChainDeps(AA, MFI, SU, &ExitSU, RejectMemNodes,
+ TrueMemOrderLatency);
PendingLoads.clear();
AliasMemDefs.clear();
AliasMemUses.clear();
} else if (MI->mayStore()) {
bool MayAlias = true;
- TrueMemOrderLatency = STORE_LOAD_LATENCY;
if (const Value *V = getUnderlyingObjectForInstr(MI, MFI, MayAlias)) {
// A store to a specific PseudoSourceValue. Add precise dependencies.
// Record the def in MemDefs, first adding a dep if there is
addChainDependency(AA, MFI, SU, AliasChain, RejectMemNodes);
// But we also should check dependent instructions for the
// SU in question.
- adjustChainDeps(AA, MFI, SU, &ExitSU, RejectMemNodes);
+ adjustChainDeps(AA, MFI, SU, &ExitSU, RejectMemNodes,
+ TrueMemOrderLatency);
}
// Add dependence on barrier chain, if needed.
// There is no point to check aliasing on barrier event. Even if
/*isArtificial=*/true));
} else if (MI->mayLoad()) {
bool MayAlias = true;
- TrueMemOrderLatency = 0;
if (MI->isInvariantLoad(AA)) {
// Invariant load, no chain dependencies needed!
} else {
MayAlias = true;
}
if (MayAlias)
- adjustChainDeps(AA, MFI, SU, &ExitSU, RejectMemNodes);
+ adjustChainDeps(AA, MFI, SU, &ExitSU, RejectMemNodes, /*Latency=*/0);
// Add dependencies on alias and barrier chains, if needed.
if (MayAlias && AliasChain)
addChainDependency(AA, MFI, SU, AliasChain, RejectMemNodes);
PendingLoads.clear();
}
-void ScheduleDAGInstrs::computeLatency(SUnit *SU) {
- // Compute the latency for the node. We only provide a default for missing
- // itineraries. Empty itineraries still have latency properties.
- if (!InstrItins) {
- SU->Latency = 1;
-
- // Simplistic target-independent heuristic: assume that loads take
- // extra time.
- if (SU->getInstr()->mayLoad())
- SU->Latency += 2;
- } else {
- SU->Latency = TII->getInstrLatency(InstrItins, SU->getInstr());
- }
-}
-
-unsigned ScheduleDAGInstrs::computeOperandLatency(SUnit *Def, SUnit *Use,
- const SDep& dep,
- bool FindMin) const {
- // For a data dependency with a known register...
- if ((dep.getKind() != SDep::Data) || (dep.getReg() == 0))
- return 1;
-
- return TII->computeOperandLatency(InstrItins, TRI, Def->getInstr(),
- Use->getInstr(), dep.getReg(), FindMin);
-}
-
void ScheduleDAGInstrs::dumpNode(const SUnit *SU) const {
+#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
SU->getInstr()->dump();
+#endif
}
std::string ScheduleDAGInstrs::getGraphNodeLabel(const SUnit *SU) const {
projects / oota-llvm.git / blobdiff