#include "llvm/CodeGen/MachineMemOperand.h"
#include "llvm/CodeGen/MachineRegisterInfo.h"
#include "llvm/CodeGen/PseudoSourceValue.h"
+#include "llvm/MC/MCInstrItineraries.h"
#include "llvm/Target/TargetMachine.h"
#include "llvm/Target/TargetInstrInfo.h"
#include "llvm/Target/TargetRegisterInfo.h"
-#include "llvm/Target/TargetSubtarget.h"
+#include "llvm/Target/TargetSubtargetInfo.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/raw_ostream.h"
#include "llvm/ADT/SmallSet.h"
ScheduleDAGInstrs::ScheduleDAGInstrs(MachineFunction &mf,
const MachineLoopInfo &mli,
- const MachineDominatorTree &mdt)
+ const MachineDominatorTree &mdt,
+ bool IsPostRAFlag)
: ScheduleDAG(mf), MLI(mli), MDT(mdt), MFI(mf.getFrameInfo()),
- InstrItins(mf.getTarget().getInstrItineraryData()),
- Defs(TRI->getNumRegs()), Uses(TRI->getNumRegs()), LoopRegs(MLI, MDT) {
- DbgValueVec.clear();
+ InstrItins(mf.getTarget().getInstrItineraryData()), IsPostRA(IsPostRAFlag),
+ UnitLatencies(false), Defs(TRI->getNumRegs()), Uses(TRI->getNumRegs()),
+ LoopRegs(MLI, MDT), FirstDbgValue(0) {
+ DbgValues.clear();
}
/// Run - perform scheduling.
Begin = begin;
InsertPosIndex = endcount;
+ // Check to see if the scheduler cares about latencies.
+ UnitLatencies = ForceUnitLatencies();
+
ScheduleDAG::Run(bb, end);
}
}
void ScheduleDAGInstrs::StartBlock(MachineBasicBlock *BB) {
+ LoopRegs.Deps.clear();
if (MachineLoop *ML = MLI.getLoopFor(BB))
- if (BB == ML->getLoopLatch()) {
- MachineBasicBlock *Header = ML->getHeader();
- for (MachineBasicBlock::livein_iterator I = Header->livein_begin(),
- E = Header->livein_end(); I != E; ++I)
- LoopLiveInRegs.insert(*I);
+ if (BB == ML->getLoopLatch())
LoopRegs.VisitLoop(ML);
- }
}
/// AddSchedBarrierDeps - Add dependencies from instructions in the current
MachineInstr *ExitMI = InsertPos != BB->end() ? &*InsertPos : 0;
ExitSU.setInstr(ExitMI);
bool AllDepKnown = ExitMI &&
- (ExitMI->getDesc().isCall() || ExitMI->getDesc().isBarrier());
+ (ExitMI->isCall() || ExitMI->isBarrier());
if (ExitMI && AllDepKnown) {
// If it's a call or a barrier, add dependencies on the defs and uses of
// instruction.
unsigned Reg = MO.getReg();
if (Reg == 0) continue;
- assert(TRI->isPhysicalRegister(Reg) && "Virtual register encountered!");
- Uses[Reg].push_back(&ExitSU);
+ if (TRI->isPhysicalRegister(Reg))
+ Uses[Reg].push_back(&ExitSU);
+ else
+ assert(!IsPostRA && "Virtual register encountered after regalloc.");
}
} else {
// For others, e.g. fallthrough, conditional branch, assume the exit
}
}
+/// addPhysRegDeps - Add register dependencies (data, anti, and output) from
+/// this SUnit to following instructions in the same scheduling region that
+/// depend the physical register referenced at OperIdx.
+void ScheduleDAGInstrs::addPhysRegDeps(SUnit *SU, unsigned OperIdx) {
+ const MachineInstr *MI = SU->getInstr();
+ const MachineOperand &MO = MI->getOperand(OperIdx);
+ unsigned Reg = MO.getReg();
+
+ // Ask the target if address-backscheduling is desirable, and if so how much.
+ const TargetSubtargetInfo &ST = TM.getSubtarget<TargetSubtargetInfo>();
+ unsigned SpecialAddressLatency = ST.getSpecialAddressLatency();
+
+ // Optionally add output and anti dependencies. For anti
+ // dependencies we use a latency of 0 because for a multi-issue
+ // target we want to allow the defining instruction to issue
+ // in the same cycle as the using instruction.
+ // TODO: Using a latency of 1 here for output dependencies assumes
+ // there's no cost for reusing registers.
+ SDep::Kind Kind = MO.isUse() ? SDep::Anti : SDep::Output;
+ for (const unsigned *Alias = TRI->getOverlaps(Reg); *Alias; ++Alias) {
+ 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))) {
+ if (Kind == SDep::Anti)
+ DefSU->addPred(SDep(SU, Kind, 0, /*Reg=*/*Alias));
+ else {
+ unsigned AOLat = TII->getOutputLatency(InstrItins, MI, OperIdx,
+ DefSU->getInstr());
+ DefSU->addPred(SDep(SU, Kind, AOLat, /*Reg=*/*Alias));
+ }
+ }
+ }
+ }
+
+ // Retrieve the UseList to add data dependencies and update uses.
+ std::vector<SUnit *> &UseList = Uses[Reg];
+ if (MO.isDef()) {
+ // Update DefList. Defs are pushed in the order they are visited and
+ // never reordered.
+ std::vector<SUnit *> &DefList = Defs[Reg];
+
+ // Add any data dependencies.
+ unsigned DataLatency = SU->Latency;
+ for (unsigned i = 0, e = UseList.size(); i != e; ++i) {
+ SUnit *UseSU = UseList[i];
+ 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 MCInstrDesc &UseMCID = UseMI->getDesc();
+ int RegUseIndex = UseMI->findRegisterUseOperandIdx(Reg);
+ assert(RegUseIndex >= 0 && "UseMI doesn's 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.
+ 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)
+ 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);
+ }
+ }
+
+ // 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(Reg);
+ 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();
+ // 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);
+ }
+ }
+
+ UseList.clear();
+ if (!MO.isDead())
+ DefList.clear();
+
+ // Calls will not be reordered because of chain dependencies (see
+ // below). Since call operands are dead, calls may continue to be added
+ // to the DefList making dependence checking quadratic in the size of
+ // the block. Instead, we leave only one call at the back of the
+ // DefList.
+ if (SU->isCall) {
+ while (!DefList.empty() && DefList.back()->isCall)
+ DefList.pop_back();
+ }
+ DefList.push_back(SU);
+ } else {
+ UseList.push_back(SU);
+ }
+}
+
+/// addVRegDefDeps - Add register output and data dependencies from this SUnit
+/// to instructions that occur later in the same scheduling region if they read
+/// from or write to the virtual register defined at OperIdx.
+///
+/// TODO: Hoist loop induction variable increments. This has to be
+/// reevaluated. Generally, IV scheduling should be done before coalescing.
+void ScheduleDAGInstrs::addVRegDefDeps(SUnit *SU, unsigned OperIdx) {
+ const MachineInstr *MI = SU->getInstr();
+ unsigned Reg = MI->getOperand(OperIdx).getReg();
+
+ const TargetSubtargetInfo &ST = TM.getSubtarget<TargetSubtargetInfo>();
+
+ // Add output dependence to the next nearest def of this vreg.
+ //
+ // Unless this definition is dead, the output dependence should be
+ // transitively redundant with antidependencies from this definition's
+ // uses. We're conservative for now until we have a way to guarantee the uses
+ // are not eliminated sometime during scheduling. The output dependence edge
+ // is also useful if output latency exceeds def-use latency.
+ SUnit *DefSU = VRegDefs[Reg];
+ if (DefSU && DefSU != SU && DefSU != &ExitSU) {
+ unsigned OutLatency = TII->getOutputLatency(InstrItins, MI, OperIdx,
+ DefSU->getInstr());
+ DefSU->addPred(SDep(SU, SDep::Output, OutLatency, Reg));
+ }
+ VRegDefs[Reg] = SU;
+
+ // Add data dependence to any uses of this vreg before the next nearest def.
+ //
+ // TODO: Handle ExitSU properly.
+ //
+ // TODO: Data dependence could be handled more efficiently at the use-side.
+ std::vector<SUnit*> &UseList = VRegUses[Reg];
+ for (std::vector<SUnit*>::const_iterator UI = UseList.begin(),
+ UE = UseList.end(); UI != UE; ++UI) {
+ SUnit *UseSU = *UI;
+ if (UseSU == SU) continue;
+
+ // TODO: Handle "special" address latencies cleanly.
+ const SDep& dep = SDep(SU, SDep::Data, SU->Latency, Reg);
+ if (!UnitLatencies) {
+ // Adjust the dependence latency using operand def/use information, then
+ // allow the target to perform its own adjustments.
+ ComputeOperandLatency(SU, UseSU, const_cast<SDep &>(dep));
+ ST.adjustSchedDependency(SU, UseSU, const_cast<SDep &>(dep));
+ }
+ UseSU->addPred(dep);
+ }
+ UseList.clear();
+}
+
+/// addVRegUseDeps - Add register antidependencies from this SUnit to
+/// instructions that occur later in the same scheduling region if they
+/// write the virtual register referenced at OperIdx.
+void ScheduleDAGInstrs::addVRegUseDeps(SUnit *SU, unsigned OperIdx) {
+ unsigned Reg = SU->getInstr()->getOperand(OperIdx).getReg();
+
+ // Add antidependence to the following def of the vreg it uses.
+ SUnit *DefSU = VRegDefs[Reg];
+ if (DefSU && DefSU != SU)
+ DefSU->addPred(SDep(SU, SDep::Anti, 0, Reg));
+
+ // Add this SUnit to the use list of the vreg it uses.
+ //
+ // TODO: pinch the DAG before we see too many uses to avoid quadratic
+ // behavior. Limiting the scheduling window can accomplish the same thing.
+ VRegUses[Reg].push_back(SU);
+}
+
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();
-
- // Ask the target if address-backscheduling is desirable, and if so how much.
- 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();
+ DbgValues.clear();
+ FirstDbgValue = NULL;
// Model data dependencies between instructions being scheduled and the
// ExitSU.
assert(Defs[i].empty() && "Only BuildGraph should push/pop Defs");
}
+ // Reinitialize the large VReg vectors, while reusing the memory.
+ //
+ // Note: this can be an expensive part of DAG building. We may want to be more
+ // clever. Reevaluate after VRegUses goes away.
+ assert(VRegDefs.size() == 0 && VRegUses.size() == 0 &&
+ "Only BuildSchedGraph should access VRegDefs/Uses");
+ VRegDefs.resize(MF.getRegInfo().getNumVirtRegs());
+ VRegUses.resize(MF.getRegInfo().getNumVirtRegs());
+
// Walk the list of instructions, from bottom moving up.
+ MachineInstr *PrevMI = NULL;
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 && PrevMI) {
+ DbgValues.push_back(std::make_pair(PrevMI, MI));
+ PrevMI = NULL;
+ }
+
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);
+ PrevMI = MI;
continue;
}
- const TargetInstrDesc &TID = MI->getDesc();
- assert(!TID.isTerminator() && !MI->isLabel() &&
+
+ assert(!MI->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();
+ SU->isCall = MI->isCall();
+ SU->isCommutable = MI->isCommutable();
// Assign the Latency field of SU using target-provided information.
if (UnitLatencies)
unsigned Reg = MO.getReg();
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];
- // Defs are push in the order they are visited and never reordered.
- std::vector<SUnit *> &DefList = Defs[Reg];
- // Optionally add output and anti dependencies. For anti
- // dependencies we use a latency of 0 because for a multi-issue
- // target we want to allow the defining instruction to issue
- // in the same cycle as the using instruction.
- // TODO: Using a latency of 1 here for output dependencies assumes
- // there's no cost for reusing registers.
- SDep::Kind Kind = MO.isUse() ? SDep::Anti : SDep::Output;
- 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)))
- DefSU->addPred(SDep(SU, Kind, AOLatency, /*Reg=*/Reg));
- }
- for (const unsigned *Alias = TRI->getAliasSet(Reg); *Alias; ++Alias) {
- std::vector<SUnit *> &MemDefList = Defs[*Alias];
- for (unsigned i = 0, e = MemDefList.size(); i != e; ++i) {
- SUnit *DefSU = MemDefList[i];
- if (DefSU == &ExitSU)
- continue;
- if (DefSU != SU &&
- (Kind != SDep::Output || !MO.isDead() ||
- !DefSU->getInstr()->registerDefIsDead(*Alias)))
- DefSU->addPred(SDep(SU, Kind, AOLatency, /*Reg=*/ *Alias));
- }
- }
-
- if (MO.isDef()) {
- // Add any data dependencies.
- unsigned DataLatency = SU->Latency;
- for (unsigned i = 0, e = UseList.size(); i != e; ++i) {
- SUnit *UseSU = UseList[i];
- 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)
- 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);
- }
- }
-
- // 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(Reg);
- 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 TargetInstrDesc &UseTID = UseMI->getDesc();
- // 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 < UseTID.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 (UseTID.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 &&
- UseTID.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);
- }
- }
-
- UseList.clear();
- if (!MO.isDead())
- DefList.clear();
-
- // Calls will not be reordered because of chain dependencies (see
- // below). Since call operands are dead, calls may continue to be added
- // to the DefList making dependence checking quadratic in the size of
- // the block. Instead, we leave only one call at the back of the
- // DefList.
- if (SU->isCall) {
- while (!DefList.empty() && DefList.back()->isCall)
- DefList.pop_back();
- }
- DefList.push_back(SU);
- } else {
- UseList.push_back(SU);
+ if (TRI->isPhysicalRegister(Reg))
+ addPhysRegDeps(SU, j);
+ else {
+ assert(!IsPostRA && "Virtual register encountered!");
+ if (MO.isDef())
+ addVRegDefDeps(SU, j);
+ else
+ addVRegUseDeps(SU, j);
}
}
// produce more precise dependence information.
#define STORE_LOAD_LATENCY 1
unsigned TrueMemOrderLatency = 0;
- if (TID.isCall() || MI->hasUnmodeledSideEffects() ||
+ if (MI->isCall() || MI->hasUnmodeledSideEffects() ||
(MI->hasVolatileMemoryRef() &&
- (!TID.mayLoad() || !MI->isInvariantLoad(AA)))) {
+ (!MI->mayLoad() || !MI->isInvariantLoad(AA)))) {
// Be conservative with these and add dependencies on all memory
// references, even those that are known to not alias.
for (std::map<const Value *, SUnit *>::iterator I =
PendingLoads.clear();
AliasMemDefs.clear();
AliasMemUses.clear();
- } else if (TID.mayStore()) {
+ } else if (MI->mayStore()) {
bool MayAlias = true;
TrueMemOrderLatency = STORE_LOAD_LATENCY;
if (const Value *V = getUnderlyingObjectForInstr(MI, MFI, MayAlias)) {
/*Reg=*/0, /*isNormalMemory=*/false,
/*isMustAlias=*/false,
/*isArtificial=*/true));
- } else if (TID.mayLoad()) {
+ } else if (MI->mayLoad()) {
bool MayAlias = true;
TrueMemOrderLatency = 0;
if (MI->isInvariantLoad(AA)) {
}
}
}
+ if (PrevMI)
+ FirstDbgValue = PrevMI;
for (int i = 0, e = TRI->getNumRegs(); i != e; ++i) {
Defs[i].clear();
Uses[i].clear();
}
+ VRegDefs.clear();
+ VRegUses.clear();
PendingLoads.clear();
}
// Simplistic target-independent heuristic: assume that loads take
// extra time.
- if (SU->getInstr()->getDesc().mayLoad())
+ if (SU->getInstr()->mayLoad())
SU->Latency += 2;
} else {
SU->Latency = TII->getInstrLatency(InstrItins, SU->getInstr());
// EmitSchedule - Emit the machine code in scheduled order.
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) {
- MachineBasicBlock::iterator I = Begin;
- ++Begin;
- BB->remove(I);
- }
+ Begin = InsertPos;
- // 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]);
+ // If first instruction was a DBG_VALUE then put it back.
+ if (FirstDbgValue)
+ BB->splice(InsertPos, BB, FirstDbgValue);
// Then re-insert them according to the given schedule.
for (unsigned i = 0, e = Sequence.size(); i != e; i++) {
- SUnit *SU = Sequence[i];
- if (!SU) {
+ if (SUnit *SU = Sequence[i])
+ BB->splice(InsertPos, BB, SU->getInstr());
+ else
// Null SUnit* is a noop.
EmitNoop();
- continue;
- }
- 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 (i == 0)
+ Begin = prior(InsertPos);
}
- // Update the Begin iterator, as the first instruction in the block
- // may have been scheduled later.
- 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();
+ // Reinsert any remaining debug_values.
+ for (std::vector<std::pair<MachineInstr *, MachineInstr *> >::iterator
+ DI = DbgValues.end(), DE = DbgValues.begin(); DI != DE; --DI) {
+ std::pair<MachineInstr *, MachineInstr *> P = *prior(DI);
+ MachineInstr *DbgValue = P.first;
+ MachineBasicBlock::iterator OrigPrivMI = P.second;
+ BB->splice(++OrigPrivMI, BB, DbgValue);
+ }
+ DbgValues.clear();
+ FirstDbgValue = NULL;
return BB;
}