#include "ScheduleDAGInstrs.h"
#include "llvm/Operator.h"
#include "llvm/Analysis/AliasAnalysis.h"
+#include "llvm/Analysis/ValueTracking.h"
#include "llvm/CodeGen/MachineFunctionPass.h"
#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)
- : ScheduleDAG(mf), MLI(mli), MDT(mdt), LoopRegs(MLI, MDT) {}
+ const MachineDominatorTree &mdt,
+ bool IsPostRAFlag)
+ : ScheduleDAG(mf), MLI(mli), MDT(mdt), MFI(mf.getFrameInfo()),
+ 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);
}
} else {
return V;
}
- assert(isa<IntegerType>(V->getType()) && "Unexpected operand type!");
+ assert(V->getType()->isIntegerTy() && "Unexpected operand type!");
} while (1);
}
-/// getUnderlyingObject - This is a wrapper around Value::getUnderlyingObject
+/// getUnderlyingObject - This is a wrapper around GetUnderlyingObject
/// and adds support for basic ptrtoint+arithmetic+inttoptr sequences.
static const Value *getUnderlyingObject(const Value *V) {
// First just call Value::getUnderlyingObject to let it do what it does.
do {
- V = V->getUnderlyingObject();
+ V = GetUnderlyingObject(V);
// If it found an inttoptr, use special code to continue climing.
if (Operator::getOpcode(V) != Instruction::IntToPtr)
break;
const Value *O = getUnderlyingObjectFromInt(cast<User>(V)->getOperand(0));
// If that succeeded in finding a pointer, continue the search.
- if (!isa<PointerType>(O->getType()))
+ if (!O->getType()->isPointerTy())
break;
V = O;
} while (1);
/// getUnderlyingObjectForInstr - If this machine instr has memory reference
/// information and it can be tracked to a normal reference to a known
/// object, return the Value for that object. Otherwise return null.
-static const Value *getUnderlyingObjectForInstr(const MachineInstr *MI) {
+static const Value *getUnderlyingObjectForInstr(const MachineInstr *MI,
+ const MachineFrameInfo *MFI,
+ bool &MayAlias) {
+ MayAlias = true;
if (!MI->hasOneMemOperand() ||
!(*MI->memoperands_begin())->getValue() ||
(*MI->memoperands_begin())->isVolatile())
return 0;
V = getUnderlyingObject(V);
- if (!isa<PseudoSourceValue>(V) && !isIdentifiedObject(V))
- return 0;
+ if (const PseudoSourceValue *PSV = dyn_cast<PseudoSourceValue>(V)) {
+ // For now, ignore PseudoSourceValues which may alias LLVM IR values
+ // because the code that uses this function has no way to cope with
+ // such aliases.
+ if (PSV->isAliased(MFI))
+ return 0;
+
+ MayAlias = PSV->mayAlias(MFI);
+ return V;
+ }
- return V;
+ if (isIdentifiedObject(V))
+ return V;
+
+ return 0;
}
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
+/// 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->isCall() || ExitMI->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;
+
+ 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
+ // 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);
+ }
+ }
+}
+
+/// 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);
+ }
}
-void ScheduleDAGInstrs::BuildSchedGraph() {
+/// 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.
SUnits.reserve(BB->size());
// We build scheduling units by walking a block's instruction list from bottom
// to top.
- // Remember where a generic side-effecting instruction is as we procede. If
- // ChainMMO is null, this is assumed to have arbitrary side-effects. If
- // ChainMMO is non-null, then Chain makes only a single memory reference.
- SUnit *Chain = 0;
- MachineMemOperand *ChainMMO = 0;
+ // Remember where a generic side-effecting instruction is as we procede.
+ SUnit *BarrierChain = 0, *AliasChain = 0;
- // Memory references to specific known memory locations are tracked so that
- // they can be given more precise dependencies.
- std::map<const Value *, SUnit *> MemDefs;
- std::map<const Value *, std::vector<SUnit *> > MemUses;
+ // Memory references to specific known memory locations are tracked
+ // so that they can be given more precise dependencies. We track
+ // separately the known memory locations that may alias and those
+ // that are known not to alias
+ std::map<const Value *, SUnit *> AliasMemDefs, NonAliasMemDefs;
+ std::map<const Value *, std::vector<SUnit *> > AliasMemUses, NonAliasMemUses;
- // Check to see if the scheduler cares about latencies.
- bool UnitLatencies = ForceUnitLatencies();
+ // Remove any stale debug info; sometimes BuildSchedGraph is called again
+ // without emitting the info from the previous call.
+ DbgValues.clear();
+ FirstDbgValue = NULL;
- // Ask the target if address-backscheduling is desirable, and if so how much.
- const TargetSubtarget &ST = TM.getSubtarget<TargetSubtarget>();
- unsigned SpecialAddressLatency = ST.getSpecialAddressLatency();
+ // Model data dependencies between instructions being scheduled and the
+ // ExitSU.
+ AddSchedBarrierDeps();
+
+ for (int i = 0, e = TRI->getNumRegs(); i != e; ++i) {
+ 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);
- const TargetInstrDesc &TID = MI->getDesc();
- assert(!TID.isTerminator() && !MI->isLabel() &&
+ if (MI && PrevMI) {
+ DbgValues.push_back(std::make_pair(PrevMI, MI));
+ PrevMI = NULL;
+ }
+
+ if (MI->isDebugValue()) {
+ PrevMI = MI;
+ continue;
+ }
+
+ assert(!MI->isTerminator() && !MI->isLabel() &&
"Cannot schedule terminators or labels!");
// Create the SUnit for this MI.
SUnit *SU = NewSUnit(MI);
+ 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!");
- std::vector<SUnit *> &UseList = Uses[Reg];
- 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 != 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 *> &DefList = Defs[*Alias];
- for (unsigned i = 0, e = DefList.size(); i != e; ++i) {
- SUnit *DefSU = DefList[i];
- if (DefSU != SU &&
- (Kind != SDep::Output || !MO.isDead() ||
- !DefSU->getInstr()->registerDefIsDead(Reg)))
- 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) {
- 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);
- }
- }
- 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 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 artifical 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();
- 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);
}
}
// Add chain dependencies.
+ // Chain dependencies used to enforce memory order should have
+ // latency of 0 (except for true dependency of Store followed by
+ // aliased Load... we estimate that with a single cycle of latency
+ // assuming the hardware will bypass)
// Note that isStoreToStackSlot and isLoadFromStackSLot are not usable
// after stack slots are lowered to actual addresses.
// TODO: Use an AliasAnalysis and do real alias-analysis queries, and
// produce more precise dependence information.
- if (TID.isCall() || TID.hasUnmodeledSideEffects()) {
- new_chain:
- // This is the conservative case. Add dependencies on all memory
- // references.
- if (Chain)
- Chain->addPred(SDep(SU, SDep::Order, SU->Latency));
- Chain = SU;
+#define STORE_LOAD_LATENCY 1
+ unsigned TrueMemOrderLatency = 0;
+ if (MI->isCall() || MI->hasUnmodeledSideEffects() ||
+ (MI->hasVolatileMemoryRef() &&
+ (!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 =
+ NonAliasMemDefs.begin(), E = NonAliasMemDefs.end(); I != E; ++I) {
+ I->second->addPred(SDep(SU, SDep::Order, /*Latency=*/0));
+ }
+ for (std::map<const Value *, std::vector<SUnit *> >::iterator I =
+ NonAliasMemUses.begin(), E = NonAliasMemUses.end(); I != E; ++I) {
+ for (unsigned i = 0, e = I->second.size(); i != e; ++i)
+ I->second[i]->addPred(SDep(SU, SDep::Order, TrueMemOrderLatency));
+ }
+ NonAliasMemDefs.clear();
+ NonAliasMemUses.clear();
+ // Add SU to the barrier chain.
+ if (BarrierChain)
+ BarrierChain->addPred(SDep(SU, SDep::Order, /*Latency=*/0));
+ BarrierChain = SU;
+
+ // fall-through
+ new_alias_chain:
+ // Chain all possibly aliasing memory references though SU.
+ if (AliasChain)
+ AliasChain->addPred(SDep(SU, SDep::Order, /*Latency=*/0));
+ AliasChain = SU;
for (unsigned k = 0, m = PendingLoads.size(); k != m; ++k)
- PendingLoads[k]->addPred(SDep(SU, SDep::Order, SU->Latency));
- PendingLoads.clear();
- for (std::map<const Value *, SUnit *>::iterator I = MemDefs.begin(),
- E = MemDefs.end(); I != E; ++I) {
- I->second->addPred(SDep(SU, SDep::Order, SU->Latency));
- I->second = SU;
+ PendingLoads[k]->addPred(SDep(SU, SDep::Order, TrueMemOrderLatency));
+ for (std::map<const Value *, SUnit *>::iterator I = AliasMemDefs.begin(),
+ E = AliasMemDefs.end(); I != E; ++I) {
+ I->second->addPred(SDep(SU, SDep::Order, /*Latency=*/0));
}
for (std::map<const Value *, std::vector<SUnit *> >::iterator I =
- MemUses.begin(), E = MemUses.end(); I != E; ++I) {
+ AliasMemUses.begin(), E = AliasMemUses.end(); I != E; ++I) {
for (unsigned i = 0, e = I->second.size(); i != e; ++i)
- I->second[i]->addPred(SDep(SU, SDep::Order, SU->Latency));
- I->second.clear();
+ I->second[i]->addPred(SDep(SU, SDep::Order, TrueMemOrderLatency));
}
- // See if it is known to just have a single memory reference.
- MachineInstr *ChainMI = Chain->getInstr();
- const TargetInstrDesc &ChainTID = ChainMI->getDesc();
- if (!ChainTID.isCall() &&
- !ChainTID.hasUnmodeledSideEffects() &&
- ChainMI->hasOneMemOperand() &&
- !(*ChainMI->memoperands_begin())->isVolatile() &&
- (*ChainMI->memoperands_begin())->getValue())
- // We know that the Chain accesses one specific memory location.
- ChainMMO = *ChainMI->memoperands_begin();
- else
- // Unknown memory accesses. Assume the worst.
- ChainMMO = 0;
- } else if (TID.mayStore()) {
- if (const Value *V = getUnderlyingObjectForInstr(MI)) {
+ 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.
- // Handle the def in MemDefs, if there is one.
- std::map<const Value *, SUnit *>::iterator I = MemDefs.find(V);
- if (I != MemDefs.end()) {
- I->second->addPred(SDep(SU, SDep::Order, SU->Latency, /*Reg=*/0,
+ // Record the def in MemDefs, first adding a dep if there is
+ // an existing def.
+ std::map<const Value *, SUnit *>::iterator I =
+ ((MayAlias) ? AliasMemDefs.find(V) : NonAliasMemDefs.find(V));
+ std::map<const Value *, SUnit *>::iterator IE =
+ ((MayAlias) ? AliasMemDefs.end() : NonAliasMemDefs.end());
+ if (I != IE) {
+ I->second->addPred(SDep(SU, SDep::Order, /*Latency=*/0, /*Reg=*/0,
/*isNormalMemory=*/true));
I->second = SU;
} else {
- MemDefs[V] = SU;
+ if (MayAlias)
+ AliasMemDefs[V] = SU;
+ else
+ NonAliasMemDefs[V] = SU;
}
// Handle the uses in MemUses, if there are any.
std::map<const Value *, std::vector<SUnit *> >::iterator J =
- MemUses.find(V);
- if (J != MemUses.end()) {
+ ((MayAlias) ? AliasMemUses.find(V) : NonAliasMemUses.find(V));
+ std::map<const Value *, std::vector<SUnit *> >::iterator JE =
+ ((MayAlias) ? AliasMemUses.end() : NonAliasMemUses.end());
+ if (J != JE) {
for (unsigned i = 0, e = J->second.size(); i != e; ++i)
- J->second[i]->addPred(SDep(SU, SDep::Order, SU->Latency, /*Reg=*/0,
- /*isNormalMemory=*/true));
+ J->second[i]->addPred(SDep(SU, SDep::Order, TrueMemOrderLatency,
+ /*Reg=*/0, /*isNormalMemory=*/true));
J->second.clear();
}
- // Add dependencies from all the PendingLoads, since without
- // memoperands we must assume they alias anything.
- for (unsigned k = 0, m = PendingLoads.size(); k != m; ++k)
- PendingLoads[k]->addPred(SDep(SU, SDep::Order, SU->Latency));
- // Add a general dependence too, if needed.
- if (Chain)
- Chain->addPred(SDep(SU, SDep::Order, SU->Latency));
- } else
+ if (MayAlias) {
+ // Add dependencies from all the PendingLoads, i.e. loads
+ // with no underlying object.
+ for (unsigned k = 0, m = PendingLoads.size(); k != m; ++k)
+ PendingLoads[k]->addPred(SDep(SU, SDep::Order, TrueMemOrderLatency));
+ // Add dependence on alias chain, if needed.
+ if (AliasChain)
+ AliasChain->addPred(SDep(SU, SDep::Order, /*Latency=*/0));
+ }
+ // Add dependence on barrier chain, if needed.
+ if (BarrierChain)
+ BarrierChain->addPred(SDep(SU, SDep::Order, /*Latency=*/0));
+ } else {
// Treat all other stores conservatively.
- goto new_chain;
- } else if (TID.mayLoad()) {
- if (MI->isInvariantLoad()) {
+ 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 (MI->mayLoad()) {
+ bool MayAlias = true;
+ TrueMemOrderLatency = 0;
+ if (MI->isInvariantLoad(AA)) {
// Invariant load, no chain dependencies needed!
- } else if (const Value *V = getUnderlyingObjectForInstr(MI)) {
- // A load from a specific PseudoSourceValue. Add precise dependencies.
- std::map<const Value *, SUnit *>::iterator I = MemDefs.find(V);
- if (I != MemDefs.end())
- I->second->addPred(SDep(SU, SDep::Order, SU->Latency, /*Reg=*/0,
- /*isNormalMemory=*/true));
- MemUses[V].push_back(SU);
-
- // Add a general dependence too, if needed.
- if (Chain && (!ChainMMO ||
- (ChainMMO->isStore() || ChainMMO->isVolatile())))
- Chain->addPred(SDep(SU, SDep::Order, SU->Latency));
- } else if (MI->hasVolatileMemoryRef()) {
- // Treat volatile loads conservatively. Note that this includes
- // cases where memoperand information is unavailable.
- goto new_chain;
} else {
- // A normal load. Depend on the general chain, as well as on
- // all stores. In the absense of MachineMemOperand information,
- // we can't even assume that the load doesn't alias well-behaved
- // memory locations.
- if (Chain)
- Chain->addPred(SDep(SU, SDep::Order, SU->Latency));
- for (std::map<const Value *, SUnit *>::iterator I = MemDefs.begin(),
- E = MemDefs.end(); I != E; ++I)
- I->second->addPred(SDep(SU, SDep::Order, SU->Latency));
- PendingLoads.push_back(SU);
+ if (const Value *V =
+ getUnderlyingObjectForInstr(MI, MFI, MayAlias)) {
+ // A load from a specific PseudoSourceValue. Add precise dependencies.
+ std::map<const Value *, SUnit *>::iterator I =
+ ((MayAlias) ? AliasMemDefs.find(V) : NonAliasMemDefs.find(V));
+ std::map<const Value *, SUnit *>::iterator IE =
+ ((MayAlias) ? AliasMemDefs.end() : NonAliasMemDefs.end());
+ if (I != IE)
+ I->second->addPred(SDep(SU, SDep::Order, /*Latency=*/0, /*Reg=*/0,
+ /*isNormalMemory=*/true));
+ if (MayAlias)
+ AliasMemUses[V].push_back(SU);
+ else
+ NonAliasMemUses[V].push_back(SU);
+ } else {
+ // A load with no underlying object. Depend on all
+ // potentially aliasing stores.
+ for (std::map<const Value *, SUnit *>::iterator I =
+ AliasMemDefs.begin(), E = AliasMemDefs.end(); I != E; ++I)
+ I->second->addPred(SDep(SU, SDep::Order, /*Latency=*/0));
+
+ PendingLoads.push_back(SU);
+ MayAlias = true;
+ }
+
+ // Add dependencies on alias and barrier chains, if needed.
+ if (MayAlias && AliasChain)
+ AliasChain->addPred(SDep(SU, SDep::Order, /*Latency=*/0));
+ if (BarrierChain)
+ BarrierChain->addPred(SDep(SU, SDep::Order, /*Latency=*/0));
}
}
}
+ 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();
}
}
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())
- if (SU->getInstr()->getDesc().mayLoad())
+ // 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());
+ }
}
-void ScheduleDAGInstrs::ComputeOperandLatency(SUnit *Def, SUnit *Use,
+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...
if ((dep.getKind() != SDep::Data) || (dep.getReg() == 0))
return;
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) {
- // 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);
- }
+MachineBasicBlock *ScheduleDAGInstrs::EmitSchedule() {
+ Begin = InsertPos;
+
+ // 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());
+ // 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 (!Sequence.empty())
- Begin = Sequence[0]->getInstr();
-
+ // 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;
}