#include "llvm/Target/TargetMachine.h"
#include "llvm/Target/TargetRegisterInfo.h"
#include "llvm/Target/TargetSubtargetInfo.h"
+#include <queue>
+
using namespace llvm;
static cl::opt<bool> EnableAASchedMI("enable-aa-sched-mi", cl::Hidden,
SmallVector<Value *, 4> Objs;
GetUnderlyingObjects(const_cast<Value *>(V), Objs);
- for (SmallVector<Value *, 4>::iterator I = Objs.begin(), IE = Objs.end();
+ for (SmallVectorImpl<Value *>::iterator I = Objs.begin(), IE = Objs.end();
I != IE; ++I) {
V = *I;
if (!Visited.insert(V))
} while (!Working.empty());
}
+typedef SmallVector<PointerIntPair<const Value *, 1, bool>, 4>
+UnderlyingObjectsVector;
+
/// getUnderlyingObjectsForInstr - 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.
static void getUnderlyingObjectsForInstr(const MachineInstr *MI,
- const MachineFrameInfo *MFI,
- SmallVectorImpl<std::pair<const Value *, bool> > &Objects) {
+ const MachineFrameInfo *MFI,
+ UnderlyingObjectsVector &Objects) {
if (!MI->hasOneMemOperand() ||
!(*MI->memoperands_begin())->getValue() ||
(*MI->memoperands_begin())->isVolatile())
SmallVector<Value *, 4> Objs;
getUnderlyingObjects(V, Objs);
- for (SmallVector<Value *, 4>::iterator I = Objs.begin(), IE = Objs.end();
- I != IE; ++I) {
+ for (SmallVectorImpl<Value *>::iterator I = Objs.begin(), IE = Objs.end();
+ I != IE; ++I) {
bool MayAlias = true;
V = *I;
return;
}
- Objects.push_back(std::make_pair(V, MayAlias));
+ Objects.push_back(UnderlyingObjectsVector::value_type(V, MayAlias));
}
}
void ScheduleDAGInstrs::enterRegion(MachineBasicBlock *bb,
MachineBasicBlock::iterator begin,
MachineBasicBlock::iterator end,
- unsigned endcount) {
+ unsigned regioninstrs) {
assert(bb == BB && "startBlock should set BB");
RegionBegin = begin;
RegionEnd = end;
- EndIndex = endcount;
- MISUnitMap.clear();
-
- ScheduleDAG::clearDAG();
+ NumRegionInstrs = regioninstrs;
}
/// Close the current scheduling region. Don't clear any state in case the
if (UseOp < 0)
Dep = SDep(SU, SDep::Artificial);
else {
+ // Set the hasPhysRegDefs only for physreg defs that have a use within
+ // the scheduling region.
+ SU->hasPhysRegDefs = true;
Dep = SDep(SU, SDep::Data, *Alias);
RegUse = UseSU->getInstr();
- Dep.setMinLatency(
- SchedModel.computeOperandLatency(SU->getInstr(), OperIdx,
- RegUse, UseOp, /*FindMin=*/true));
}
Dep.setLatency(
- SchedModel.computeOperandLatency(SU->getInstr(), OperIdx,
- RegUse, UseOp, /*FindMin=*/false));
+ SchedModel.computeOperandLatency(SU->getInstr(), OperIdx, RegUse,
+ UseOp));
ST.adjustSchedDependency(SU, UseSU, Dep);
UseSU->addPred(Dep);
DefSU->addPred(SDep(SU, Kind, /*Reg=*/*Alias));
else {
SDep Dep(SU, Kind, /*Reg=*/*Alias);
- unsigned OutLatency =
- SchedModel.computeOutputLatency(MI, OperIdx, DefSU->getInstr());
- Dep.setMinLatency(OutLatency);
- Dep.setLatency(OutLatency);
+ Dep.setLatency(
+ SchedModel.computeOutputLatency(MI, OperIdx, DefSU->getInstr()));
DefSU->addPred(Dep);
}
}
}
if (!MO.isDef()) {
+ SU->hasPhysRegUses = true;
// 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.
SUnit *DefSU = DefI->SU;
if (DefSU != SU && DefSU != &ExitSU) {
SDep Dep(SU, SDep::Output, Reg);
- unsigned OutLatency =
- SchedModel.computeOutputLatency(MI, OperIdx, DefSU->getInstr());
- Dep.setMinLatency(OutLatency);
- Dep.setLatency(OutLatency);
+ Dep.setLatency(
+ SchedModel.computeOutputLatency(MI, OperIdx, DefSU->getInstr()));
DefSU->addPred(Dep);
}
DefI->SU = SU;
MachineInstr *MI = SU->getInstr();
unsigned Reg = MI->getOperand(OperIdx).getReg();
+ // Record this local VReg use.
+ VReg2UseMap::iterator UI = VRegUses.find(Reg);
+ for (; UI != VRegUses.end(); ++UI) {
+ if (UI->SU == SU)
+ break;
+ }
+ if (UI == VRegUses.end())
+ VRegUses.insert(VReg2SUnit(Reg, SU));
+
// Lookup this operand's reaching definition.
assert(LIS && "vreg dependencies requires LiveIntervals");
LiveRangeQuery LRQ(LIS->getInterval(Reg), LIS->getInstructionIndex(MI));
// 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));
+ dep.setLatency(SchedModel.computeOperandLatency(Def, DefOp, MI, OperIdx));
const TargetSubtargetInfo &ST = TM.getSubtarget<TargetSubtargetInfo>();
ST.adjustSchedDependency(DefSU, SU, const_cast<SDep &>(dep));
SmallVector<Value *, 4> Objs;
getUnderlyingObjects(V, Objs);
- for (SmallVector<Value *, 4>::iterator I = Objs.begin(),
- IE = Objs.end(); I != IE; ++I) {
+ for (SmallVectorImpl<Value *>::iterator I = Objs.begin(),
+ IE = Objs.end(); I != IE; ++I) {
V = *I;
if (const PseudoSourceValue *PSV = dyn_cast<PseudoSourceValue>(V)) {
bool isNormalMemory = false) {
// If this is a false dependency,
// do not add the edge, but rememeber the rejected node.
- if (!EnableAASchedMI ||
- MIsNeedChainEdge(AA, MFI, SUa->getInstr(), SUb->getInstr())) {
+ if (!AA || MIsNeedChainEdge(AA, MFI, SUa->getInstr(), SUb->getInstr())) {
SDep Dep(SUa, isNormalMemory ? SDep::MayAliasMem : SDep::Barrier);
Dep.setLatency(TrueMemOrderLatency);
SUb->addPred(Dep);
void ScheduleDAGInstrs::initSUnits() {
// We'll be allocating one SUnit for each real instruction in the region,
// which is contained within a basic block.
- SUnits.reserve(BB->size());
+ SUnits.reserve(NumRegionInstrs);
for (MachineBasicBlock::iterator I = RegionBegin; I != RegionEnd; ++I) {
MachineInstr *MI = I;
/// DAG builder is an efficient place to do it because it already visits
/// operands.
void ScheduleDAGInstrs::buildSchedGraph(AliasAnalysis *AA,
- RegPressureTracker *RPTracker) {
+ RegPressureTracker *RPTracker,
+ PressureDiffs *PDiffs) {
+ const TargetSubtargetInfo &ST = TM.getSubtarget<TargetSubtargetInfo>();
+ bool UseAA = EnableAASchedMI.getNumOccurrences() > 0 ? EnableAASchedMI
+ : ST.useAA();
+ AliasAnalysis *AAForDep = UseAA ? AA : 0;
+
+ MISUnitMap.clear();
+ ScheduleDAG::clearDAG();
+
// Create an SUnit for each real instruction.
initSUnits();
+ if (PDiffs)
+ PDiffs->init(SUnits.size());
+
// We build scheduling units by walking a block's instruction list from bottom
// to top.
Uses.setUniverse(TRI->getNumRegs());
assert(VRegDefs.empty() && "Only BuildSchedGraph may access VRegDefs");
- // FIXME: Allow SparseSet to reserve space for the creation of virtual
- // registers during scheduling. Don't artificially inflate the Universe
- // because we want to assert that vregs are not created during DAG building.
+ VRegUses.clear();
VRegDefs.setUniverse(MRI.getNumVirtRegs());
+ VRegUses.setUniverse(MRI.getNumVirtRegs());
// Model data dependencies between instructions being scheduled and the
// ExitSU.
DbgMI = MI;
continue;
}
+ SUnit *SU = MISUnitMap[MI];
+ assert(SU && "No SUnit mapped to this MI");
+
if (RPTracker) {
- RPTracker->recede();
+ PressureDiff *PDiff = PDiffs ? &(*PDiffs)[SU->NodeNum] : 0;
+ RPTracker->recede(/*LiveUses=*/0, PDiff);
assert(RPTracker->getPos() == prior(MII) && "RPTracker can't find MI");
}
- assert((!MI->isTerminator() || CanHandleTerminators) && !MI->isLabel() &&
+ assert((CanHandleTerminators || (!MI->isTerminator() && !MI->isLabel())) &&
"Cannot schedule terminators or labels!");
- SUnit *SU = MISUnitMap[MI];
- assert(SU && "No SUnit mapped to this MI");
-
// Add register-based dependencies (data, anti, and output).
bool HasVRegDef = false;
for (unsigned j = 0, n = MI->getNumOperands(); j != n; ++j) {
unsigned ChainLatency = 0;
if (AliasChain->getInstr()->mayLoad())
ChainLatency = TrueMemOrderLatency;
- addChainDependency(AA, MFI, SU, AliasChain, RejectMemNodes,
+ addChainDependency(AAForDep, 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(AAForDep, MFI, SU, PendingLoads[k], RejectMemNodes,
TrueMemOrderLatency);
for (MapVector<const Value *, SUnit *>::iterator I = AliasMemDefs.begin(),
E = AliasMemDefs.end(); I != E; ++I)
- addChainDependency(AA, MFI, SU, I->second, RejectMemNodes);
+ addChainDependency(AAForDep, MFI, SU, I->second, RejectMemNodes);
for (MapVector<const Value *, std::vector<SUnit *> >::iterator I =
AliasMemUses.begin(), E = AliasMemUses.end(); I != E; ++I) {
for (unsigned i = 0, e = I->second.size(); i != e; ++i)
- addChainDependency(AA, MFI, SU, I->second[i], RejectMemNodes,
+ addChainDependency(AAForDep, MFI, SU, I->second[i], RejectMemNodes,
TrueMemOrderLatency);
}
adjustChainDeps(AA, MFI, SU, &ExitSU, RejectMemNodes,
AliasMemDefs.clear();
AliasMemUses.clear();
} else if (MI->mayStore()) {
- SmallVector<std::pair<const Value *, bool>, 4> Objs;
+ UnderlyingObjectsVector Objs;
getUnderlyingObjectsForInstr(MI, MFI, Objs);
if (Objs.empty()) {
}
bool MayAlias = false;
- for (SmallVector<std::pair<const Value *, bool>, 4>::iterator
- K = Objs.begin(), KE = Objs.end(); K != KE; ++K) {
- const Value *V = K->first;
- bool ThisMayAlias = K->second;
+ for (UnderlyingObjectsVector::iterator K = Objs.begin(), KE = Objs.end();
+ K != KE; ++K) {
+ const Value *V = K->getPointer();
+ bool ThisMayAlias = K->getInt();
if (ThisMayAlias)
MayAlias = true;
MapVector<const Value *, SUnit *>::iterator IE =
((ThisMayAlias) ? AliasMemDefs.end() : NonAliasMemDefs.end());
if (I != IE) {
- addChainDependency(AA, MFI, SU, I->second, RejectMemNodes, 0, true);
+ addChainDependency(AAForDep, MFI, SU, I->second, RejectMemNodes,
+ 0, true);
I->second = SU;
} else {
if (ThisMayAlias)
((ThisMayAlias) ? AliasMemUses.end() : NonAliasMemUses.end());
if (J != JE) {
for (unsigned i = 0, e = J->second.size(); i != e; ++i)
- addChainDependency(AA, MFI, SU, J->second[i], RejectMemNodes,
+ addChainDependency(AAForDep, MFI, SU, J->second[i], RejectMemNodes,
TrueMemOrderLatency, true);
J->second.clear();
}
// Add dependencies from all the PendingLoads, i.e. loads
// with no underlying object.
for (unsigned k = 0, m = PendingLoads.size(); k != m; ++k)
- addChainDependency(AA, MFI, SU, PendingLoads[k], RejectMemNodes,
+ addChainDependency(AAForDep, MFI, SU, PendingLoads[k], RejectMemNodes,
TrueMemOrderLatency);
// Add dependence on alias chain, if needed.
if (AliasChain)
- addChainDependency(AA, MFI, SU, AliasChain, RejectMemNodes);
+ addChainDependency(AAForDep, MFI, SU, AliasChain, RejectMemNodes);
// But we also should check dependent instructions for the
// SU in question.
adjustChainDeps(AA, MFI, SU, &ExitSU, RejectMemNodes,
if (MI->isInvariantLoad(AA)) {
// Invariant load, no chain dependencies needed!
} else {
- SmallVector<std::pair<const Value *, bool>, 4> Objs;
+ UnderlyingObjectsVector Objs;
getUnderlyingObjectsForInstr(MI, MFI, Objs);
if (Objs.empty()) {
// potentially aliasing stores.
for (MapVector<const Value *, SUnit *>::iterator I =
AliasMemDefs.begin(), E = AliasMemDefs.end(); I != E; ++I)
- addChainDependency(AA, MFI, SU, I->second, RejectMemNodes);
+ addChainDependency(AAForDep, MFI, SU, I->second, RejectMemNodes);
PendingLoads.push_back(SU);
MayAlias = true;
MayAlias = false;
}
- for (SmallVector<std::pair<const Value *, bool>, 4>::iterator
+ for (UnderlyingObjectsVector::iterator
J = Objs.begin(), JE = Objs.end(); J != JE; ++J) {
- const Value *V = J->first;
- bool ThisMayAlias = J->second;
+ const Value *V = J->getPointer();
+ bool ThisMayAlias = J->getInt();
if (ThisMayAlias)
MayAlias = true;
MapVector<const Value *, SUnit *>::iterator IE =
((ThisMayAlias) ? AliasMemDefs.end() : NonAliasMemDefs.end());
if (I != IE)
- addChainDependency(AA, MFI, SU, I->second, RejectMemNodes, 0, true);
+ addChainDependency(AAForDep, MFI, SU, I->second, RejectMemNodes,
+ 0, true);
if (ThisMayAlias)
AliasMemUses[V].push_back(SU);
else
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);
+ addChainDependency(AAForDep, MFI, SU, AliasChain, RejectMemNodes);
if (BarrierChain)
BarrierChain->addPred(SDep(SU, SDep::Barrier));
}
else if (SU == &ExitSU)
oss << "<exit>";
else
- SU->getInstr()->print(oss);
+ SU->getInstr()->print(oss, &TM, /*SkipOpers=*/true);
return oss.str();
}
/// List PredSU, SuccSU pairs that represent data edges between subtrees.
std::vector<std::pair<const SUnit*, const SUnit*> > ConnectionPairs;
+ struct RootData {
+ unsigned NodeID;
+ unsigned ParentNodeID; // Parent node (member of the parent subtree).
+ unsigned SubInstrCount; // Instr count in this tree only, not children.
+
+ RootData(unsigned id): NodeID(id),
+ ParentNodeID(SchedDFSResult::InvalidSubtreeID),
+ SubInstrCount(0) {}
+
+ unsigned getSparseSetIndex() const { return NodeID; }
+ };
+
+ SparseSet<RootData> RootSet;
+
public:
- SchedDFSImpl(SchedDFSResult &r): R(r), SubtreeClasses(R.DFSData.size()) {}
+ SchedDFSImpl(SchedDFSResult &r): R(r), SubtreeClasses(R.DFSNodeData.size()) {
+ RootSet.setUniverse(R.DFSNodeData.size());
+ }
- /// SubtreID is initialized to zero, set to itself to flag the root of a
- /// subtree, set to the parent to indicate an interior node,
- /// then set to a representative subtree ID during finalization.
+ /// Return true if this node been visited by the DFS traversal.
+ ///
+ /// During visitPostorderNode the Node's SubtreeID is assigned to the Node
+ /// ID. Later, SubtreeID is updated but remains valid.
bool isVisited(const SUnit *SU) const {
- return R.DFSData[SU->NodeNum].SubtreeID;
+ return R.DFSNodeData[SU->NodeNum].SubtreeID
+ != SchedDFSResult::InvalidSubtreeID;
}
/// Initialize this node's instruction count. We don't need to flag the node
/// visited until visitPostorder because the DAG cannot have cycles.
void visitPreorder(const SUnit *SU) {
- R.DFSData[SU->NodeNum].InstrCount = SU->getInstr()->isTransient() ? 0 : 1;
+ R.DFSNodeData[SU->NodeNum].InstrCount =
+ SU->getInstr()->isTransient() ? 0 : 1;
}
- /// Mark this node as either the root of a subtree or an interior
- /// node. Increment the parent node's instruction count.
- void visitPostorder(const SUnit *SU, const SDep *PredDep, const SUnit *Parent) {
- R.DFSData[SU->NodeNum].SubtreeID = SU->NodeNum;
-
- if (!Parent)
- return;
- assert(PredDep && "PredDep required for non-root node");
-
- joinPredSubtree(*PredDep, Parent);
+ /// Called once for each node after all predecessors are visited. Revisit this
+ /// node's predecessors and potentially join them now that we know the ILP of
+ /// the other predecessors.
+ void visitPostorderNode(const SUnit *SU) {
+ // Mark this node as the root of a subtree. It may be joined with its
+ // successors later.
+ R.DFSNodeData[SU->NodeNum].SubtreeID = SU->NodeNum;
+ RootData RData(SU->NodeNum);
+ RData.SubInstrCount = SU->getInstr()->isTransient() ? 0 : 1;
+
+ // If any predecessors are still in their own subtree, they either cannot be
+ // joined or are large enough to remain separate. If this parent node's
+ // total instruction count is not greater than a child subtree by at least
+ // the subtree limit, then try to join it now since splitting subtrees is
+ // only useful if multiple high-pressure paths are possible.
+ unsigned InstrCount = R.DFSNodeData[SU->NodeNum].InstrCount;
+ for (SUnit::const_pred_iterator
+ PI = SU->Preds.begin(), PE = SU->Preds.end(); PI != PE; ++PI) {
+ if (PI->getKind() != SDep::Data)
+ continue;
+ unsigned PredNum = PI->getSUnit()->NodeNum;
+ if ((InstrCount - R.DFSNodeData[PredNum].InstrCount) < R.SubtreeLimit)
+ joinPredSubtree(*PI, SU, /*CheckLimit=*/false);
+
+ // Either link or merge the TreeData entry from the child to the parent.
+ if (R.DFSNodeData[PredNum].SubtreeID == PredNum) {
+ // If the predecessor's parent is invalid, this is a tree edge and the
+ // current node is the parent.
+ if (RootSet[PredNum].ParentNodeID == SchedDFSResult::InvalidSubtreeID)
+ RootSet[PredNum].ParentNodeID = SU->NodeNum;
+ }
+ else if (RootSet.count(PredNum)) {
+ // The predecessor is not a root, but is still in the root set. This
+ // must be the new parent that it was just joined to. Note that
+ // RootSet[PredNum].ParentNodeID may either be invalid or may still be
+ // set to the original parent.
+ RData.SubInstrCount += RootSet[PredNum].SubInstrCount;
+ RootSet.erase(PredNum);
+ }
+ }
+ RootSet[SU->NodeNum] = RData;
}
- /// Determine whether the DFS cross edge should be considered a subtree edge
- /// or a connection between subtrees.
- void visitCross(const SDep &PredDep, const SUnit *Succ) {
+ /// Called once for each tree edge after calling visitPostOrderNode on the
+ /// predecessor. Increment the parent node's instruction count and
+ /// preemptively join this subtree to its parent's if it is small enough.
+ void visitPostorderEdge(const SDep &PredDep, const SUnit *Succ) {
+ R.DFSNodeData[Succ->NodeNum].InstrCount
+ += R.DFSNodeData[PredDep.getSUnit()->NodeNum].InstrCount;
joinPredSubtree(PredDep, Succ);
+ }
+
+ /// Add a connection for cross edges.
+ void visitCrossEdge(const SDep &PredDep, const SUnit *Succ) {
ConnectionPairs.push_back(std::make_pair(PredDep.getSUnit(), Succ));
}
/// between trees.
void finalize() {
SubtreeClasses.compress();
+ R.DFSTreeData.resize(SubtreeClasses.getNumClasses());
+ assert(SubtreeClasses.getNumClasses() == RootSet.size()
+ && "number of roots should match trees");
+ for (SparseSet<RootData>::const_iterator
+ RI = RootSet.begin(), RE = RootSet.end(); RI != RE; ++RI) {
+ unsigned TreeID = SubtreeClasses[RI->NodeID];
+ if (RI->ParentNodeID != SchedDFSResult::InvalidSubtreeID)
+ R.DFSTreeData[TreeID].ParentTreeID = SubtreeClasses[RI->ParentNodeID];
+ R.DFSTreeData[TreeID].SubInstrCount = RI->SubInstrCount;
+ // Note that SubInstrCount may be greater than InstrCount if we joined
+ // subtrees across a cross edge. InstrCount will be attributed to the
+ // original parent, while SubInstrCount will be attributed to the joined
+ // parent.
+ }
R.SubtreeConnections.resize(SubtreeClasses.getNumClasses());
R.SubtreeConnectLevels.resize(SubtreeClasses.getNumClasses());
DEBUG(dbgs() << R.getNumSubtrees() << " subtrees:\n");
- for (unsigned Idx = 0, End = R.DFSData.size(); Idx != End; ++Idx) {
- R.DFSData[Idx].SubtreeID = SubtreeClasses[Idx];
+ for (unsigned Idx = 0, End = R.DFSNodeData.size(); Idx != End; ++Idx) {
+ R.DFSNodeData[Idx].SubtreeID = SubtreeClasses[Idx];
DEBUG(dbgs() << " SU(" << Idx << ") in tree "
- << R.DFSData[Idx].SubtreeID << '\n');
+ << R.DFSNodeData[Idx].SubtreeID << '\n');
}
for (std::vector<std::pair<const SUnit*, const SUnit*> >::const_iterator
I = ConnectionPairs.begin(), E = ConnectionPairs.end();
}
protected:
- void joinPredSubtree(const SDep &PredDep, const SUnit *Succ) {
- // Join the child to its parent if they are connected via data dependence.
- if (PredDep.getKind() != SDep::Data)
- return;
+ /// Join the predecessor subtree with the successor that is its DFS
+ /// parent. Apply some heuristics before joining.
+ bool joinPredSubtree(const SDep &PredDep, const SUnit *Succ,
+ bool CheckLimit = true) {
+ assert(PredDep.getKind() == SDep::Data && "Subtrees are for data edges");
+
+ // Check if the predecessor is already joined.
+ const SUnit *PredSU = PredDep.getSUnit();
+ unsigned PredNum = PredSU->NodeNum;
+ if (R.DFSNodeData[PredNum].SubtreeID != PredNum)
+ return false;
// Four is the magic number of successors before a node is considered a
// pinch point.
unsigned NumDataSucs = 0;
- const SUnit *PredSU = PredDep.getSUnit();
for (SUnit::const_succ_iterator SI = PredSU->Succs.begin(),
SE = PredSU->Succs.end(); SI != SE; ++SI) {
if (SI->getKind() == SDep::Data) {
if (++NumDataSucs >= 4)
- return;
+ return false;
}
}
- // If this is a cross edge to a root, join the subtrees. This happens when
- // the root was first reached by a non-data dependence.
- unsigned NodeNum = PredSU->NodeNum;
- unsigned PredCnt = R.DFSData[NodeNum].InstrCount;
- if (R.DFSData[NodeNum].SubtreeID == NodeNum && PredCnt < R.SubtreeLimit) {
- R.DFSData[NodeNum].SubtreeID = Succ->NodeNum;
- R.DFSData[Succ->NodeNum].InstrCount += PredCnt;
- SubtreeClasses.join(Succ->NodeNum, NodeNum);
- return;
- }
+ if (CheckLimit && R.DFSNodeData[PredNum].InstrCount > R.SubtreeLimit)
+ return false;
+ R.DFSNodeData[PredNum].SubtreeID = Succ->NodeNum;
+ SubtreeClasses.join(Succ->NodeNum, PredNum);
+ return true;
}
/// Called by finalize() to record a connection between trees.
if (!Depth)
return;
- SmallVectorImpl<SchedDFSResult::Connection> &Connections =
- R.SubtreeConnections[FromTree];
- for (SmallVectorImpl<SchedDFSResult::Connection>::iterator
- I = Connections.begin(), E = Connections.end(); I != E; ++I) {
- if (I->TreeID == ToTree) {
- I->Level = std::max(I->Level, Depth);
- return;
+ do {
+ SmallVectorImpl<SchedDFSResult::Connection> &Connections =
+ R.SubtreeConnections[FromTree];
+ for (SmallVectorImpl<SchedDFSResult::Connection>::iterator
+ I = Connections.begin(), E = Connections.end(); I != E; ++I) {
+ if (I->TreeID == ToTree) {
+ I->Level = std::max(I->Level, Depth);
+ return;
+ }
}
- }
- Connections.push_back(SchedDFSResult::Connection(ToTree, Depth));
+ Connections.push_back(SchedDFSResult::Connection(ToTree, Depth));
+ FromTree = R.DFSTreeData[FromTree].ParentTreeID;
+ } while (FromTree != SchedDFSResult::InvalidSubtreeID);
}
};
} // namespace llvm
};
} // anonymous
+static bool hasDataSucc(const SUnit *SU) {
+ for (SUnit::const_succ_iterator
+ SI = SU->Succs.begin(), SE = SU->Succs.end(); SI != SE; ++SI) {
+ if (SI->getKind() == SDep::Data && !SI->getSUnit()->isBoundaryNode())
+ return true;
+ }
+ return false;
+}
+
/// Compute an ILP metric for all nodes in the subDAG reachable via depth-first
/// search from this root.
-void SchedDFSResult::compute(ArrayRef<SUnit *> Roots) {
+void SchedDFSResult::compute(ArrayRef<SUnit> SUnits) {
if (!IsBottomUp)
llvm_unreachable("Top-down ILP metric is unimplemnted");
SchedDFSImpl Impl(*this);
- for (ArrayRef<const SUnit*>::const_iterator
- RootI = Roots.begin(), RootE = Roots.end(); RootI != RootE; ++RootI) {
+ for (ArrayRef<SUnit>::const_iterator
+ SI = SUnits.begin(), SE = SUnits.end(); SI != SE; ++SI) {
+ const SUnit *SU = &*SI;
+ if (Impl.isVisited(SU) || hasDataSucc(SU))
+ continue;
+
SchedDAGReverseDFS DFS;
- Impl.visitPreorder(*RootI);
- DFS.follow(*RootI);
+ Impl.visitPreorder(SU);
+ DFS.follow(SU);
for (;;) {
// Traverse the leftmost path as far as possible.
while (DFS.getPred() != DFS.getPredEnd()) {
const SDep &PredDep = *DFS.getPred();
DFS.advance();
- // If the pred is already valid, skip it. We may preorder visit a node
- // with InstrCount==0 more than once, but it won't affect heuristics
- // because we don't care about cross edges to leaf copies.
+ // Ignore non-data edges.
+ if (PredDep.getKind() != SDep::Data
+ || PredDep.getSUnit()->isBoundaryNode()) {
+ continue;
+ }
+ // An already visited edge is a cross edge, assuming an acyclic DAG.
if (Impl.isVisited(PredDep.getSUnit())) {
- Impl.visitCross(PredDep, DFS.getCurr());
+ Impl.visitCrossEdge(PredDep, DFS.getCurr());
continue;
}
Impl.visitPreorder(PredDep.getSUnit());
// Visit the top of the stack in postorder and backtrack.
const SUnit *Child = DFS.getCurr();
const SDep *PredDep = DFS.backtrack();
- Impl.visitPostorder(Child, PredDep, PredDep ? DFS.getCurr() : 0);
+ Impl.visitPostorderNode(Child);
+ if (PredDep)
+ Impl.visitPostorderEdge(*PredDep, DFS.getCurr());
if (DFS.isComplete())
break;
}