void ScheduleDAGLinearize::ScheduleNode(SDNode *N) {
if (N->getNodeId() != 0)
llvm_unreachable(nullptr);
if (!N->isMachineOpcode() &&
(N->getOpcode() == ISD::EntryToken || isPassiveNode(N)))
// These nodes do not need to be translated into MIs.
return;
LLVM_DEBUG(dbgs() << "\n*** Scheduling: ");
LLVM_DEBUG(N->dump(DAG));
Sequence.push_back(N);
unsigned NumOps = N->getNumOperands();
if (unsigned NumLeft = NumOps) {
SDNode *GluedOpN = nullptr;
do {
const SDValue &Op = N->getOperand(NumLeft-1);
SDNode *OpN = Op.getNode();
if (NumLeft == NumOps && Op.getValueType() == MVT::Glue) {
// Schedule glue operand right above N.
GluedOpN = OpN;
assert(OpN->getNodeId() != 0 && "Glue operand not ready?");
OpN->setNodeId(0);
ScheduleNode(OpN);
continue;
}
if (OpN == GluedOpN)
// Glue operand is already scheduled.
continue;
DenseMap<SDNode*, SDNode*>::iterator DI = GluedMap.find(OpN);
if (DI != GluedMap.end() && DI->second != N)
// Users of glues are counted against the glued users.
OpN = DI->second;
unsigned Degree = OpN->getNodeId();
assert(Degree > 0 && "Predecessor over-released!");
OpN->setNodeId(--Degree);
if (Degree == 0)
ScheduleNode(OpN);
} while (--NumLeft);
}
}
void ScheduleDAGLinearize::Schedule() {
LLVM_DEBUG(dbgs() << "********** DAG Linearization **********\n");
SmallVector<SDNode*, 8> Glues;
unsigned DAGSize = 0;
for (SDNode &Node : DAG->allnodes()) {
SDNode *N = &Node;
// Use node id to record degree.
unsigned Degree = N->use_size();
N->setNodeId(Degree);
unsigned NumVals = N->getNumValues();
if (NumVals && N->getValueType(NumVals-1) == MVT::Glue &&
N->hasAnyUseOfValue(NumVals-1)) {
SDNode *User = findGluedUser(N);
if (User) {
Glues.push_back(N);
GluedMap.insert(std::make_pair(N, User));
}
}
if (N->isMachineOpcode() ||
(N->getOpcode() != ISD::EntryToken && !isPassiveNode(N)))
++DAGSize;
}
for (unsigned i = 0, e = Glues.size(); i != e; ++i) {
SDNode *Glue = Glues[i];
SDNode *GUser = GluedMap[Glue];
unsigned Degree = Glue->getNodeId();
unsigned UDegree = GUser->getNodeId();
// Glue user must be scheduled together with the glue operand. So other
// users of the glue operand must be treated as its users.
SDNode *ImmGUser = Glue->getGluedUser();
for (const SDNode *U : Glue->uses())
if (U == ImmGUser)
--Degree;
GUser->setNodeId(UDegree + Degree);
Glue->setNodeId(1);
}
Sequence.reserve(DAGSize);
ScheduleNode(DAG->getRoot().getNode());
}
/// This is the main entry point for the type legalizer. This does a top-down
/// traversal of the dag, legalizing types as it goes. Returns "true" if it made
/// any changes.
bool DAGTypeLegalizer::run() {
bool Changed = false;
// Create a dummy node (which is not added to allnodes), that adds a reference
// to the root node, preventing it from being deleted, and tracking any
// changes of the root.
HandleSDNode Dummy(DAG.getRoot());
Dummy.setNodeId(Unanalyzed);
// The root of the dag may dangle to deleted nodes until the type legalizer is
// done. Set it to null to avoid confusion.
DAG.setRoot(SDValue());
// Walk all nodes in the graph, assigning them a NodeId of 'ReadyToProcess'
// (and remembering them) if they are leaves and assigning 'Unanalyzed' if
// non-leaves.
for (SDNode &Node : DAG.allnodes()) {
if (Node.getNumOperands() == 0) {
AddToWorklist(&Node);
} else {
Node.setNodeId(Unanalyzed);
}
}
// Now that we have a set of nodes to process, handle them all.
while (!Worklist.empty()) {
#ifndef EXPENSIVE_CHECKS
if (EnableExpensiveChecks)
#endif
PerformExpensiveChecks();
SDNode *N = Worklist.back();
Worklist.pop_back();
assert(N->getNodeId() == ReadyToProcess &&
"Node should be ready if on worklist!");
if (IgnoreNodeResults(N))
goto ScanOperands;
// Scan the values produced by the node, checking to see if any result
// types are illegal.
for (unsigned i = 0, NumResults = N->getNumValues(); i < NumResults; ++i) {
EVT ResultVT = N->getValueType(i);
switch (getTypeAction(ResultVT)) {
case TargetLowering::TypeLegal:
break;
// The following calls must take care of *all* of the node's results,
// not just the illegal result they were passed (this includes results
// with a legal type). Results can be remapped using ReplaceValueWith,
// or their promoted/expanded/etc values registered in PromotedIntegers,
// ExpandedIntegers etc.
case TargetLowering::TypePromoteInteger:
PromoteIntegerResult(N, i);
Changed = true;
goto NodeDone;
case TargetLowering::TypeExpandInteger:
ExpandIntegerResult(N, i);
Changed = true;
goto NodeDone;
case TargetLowering::TypeSoftenFloat:
Changed = SoftenFloatResult(N, i);
if (Changed)
goto NodeDone;
// If not changed, the result type should be legally in register.
assert(isLegalInHWReg(ResultVT) &&
"Unchanged SoftenFloatResult should be legal in register!");
goto ScanOperands;
case TargetLowering::TypeExpandFloat:
ExpandFloatResult(N, i);
Changed = true;
goto NodeDone;
case TargetLowering::TypeScalarizeVector:
ScalarizeVectorResult(N, i);
Changed = true;
goto NodeDone;
case TargetLowering::TypeSplitVector:
SplitVectorResult(N, i);
Changed = true;
goto NodeDone;
case TargetLowering::TypeWidenVector:
WidenVectorResult(N, i);
Changed = true;
goto NodeDone;
case TargetLowering::TypePromoteFloat:
PromoteFloatResult(N, i);
Changed = true;
goto NodeDone;
}
}
ScanOperands:
// Scan the operand list for the node, handling any nodes with operands that
// are illegal.
{
unsigned NumOperands = N->getNumOperands();
bool NeedsReanalyzing = false;
unsigned i;
for (i = 0; i != NumOperands; ++i) {
if (IgnoreNodeResults(N->getOperand(i).getNode()))
continue;
EVT OpVT = N->getOperand(i).getValueType();
switch (getTypeAction(OpVT)) {
case TargetLowering::TypeLegal:
continue;
// The following calls must either replace all of the node's results
// using ReplaceValueWith, and return "false"; or update the node's
// operands in place, and return "true".
case TargetLowering::TypePromoteInteger:
NeedsReanalyzing = PromoteIntegerOperand(N, i);
Changed = true;
break;
case TargetLowering::TypeExpandInteger:
NeedsReanalyzing = ExpandIntegerOperand(N, i);
Changed = true;
break;
case TargetLowering::TypeSoftenFloat:
NeedsReanalyzing = SoftenFloatOperand(N, i);
Changed = true;
break;
case TargetLowering::TypeExpandFloat:
NeedsReanalyzing = ExpandFloatOperand(N, i);
Changed = true;
break;
case TargetLowering::TypeScalarizeVector:
NeedsReanalyzing = ScalarizeVectorOperand(N, i);
Changed = true;
break;
case TargetLowering::TypeSplitVector:
NeedsReanalyzing = SplitVectorOperand(N, i);
Changed = true;
break;
case TargetLowering::TypeWidenVector:
NeedsReanalyzing = WidenVectorOperand(N, i);
Changed = true;
break;
case TargetLowering::TypePromoteFloat:
NeedsReanalyzing = PromoteFloatOperand(N, i);
Changed = true;
break;
}
break;
}
// The sub-method updated N in place. Check to see if any operands are new,
// and if so, mark them. If the node needs revisiting, don't add all users
// to the worklist etc.
if (NeedsReanalyzing) {
assert(N->getNodeId() == ReadyToProcess && "Node ID recalculated?");
N->setNodeId(NewNode);
// Recompute the NodeId and correct processed operands, adding the node to
// the worklist if ready.
SDNode *M = AnalyzeNewNode(N);
if (M == N)
// The node didn't morph - nothing special to do, it will be revisited.
continue;
// The node morphed - this is equivalent to legalizing by replacing every
// value of N with the corresponding value of M. So do that now.
assert(N->getNumValues() == M->getNumValues() &&
"Node morphing changed the number of results!");
for (unsigned i = 0, e = N->getNumValues(); i != e; ++i)
// Replacing the value takes care of remapping the new value.
ReplaceValueWith(SDValue(N, i), SDValue(M, i));
assert(N->getNodeId() == NewNode && "Unexpected node state!");
// The node continues to live on as part of the NewNode fungus that
// grows on top of the useful nodes. Nothing more needs to be done
// with it - move on to the next node.
continue;
}
if (i == NumOperands) {
DEBUG(dbgs() << "Legally typed node: "; N->dump(&DAG); dbgs() << "\n");
}
}
NodeDone:
// If we reach here, the node was processed, potentially creating new nodes.
// Mark it as processed and add its users to the worklist as appropriate.
assert(N->getNodeId() == ReadyToProcess && "Node ID recalculated?");
N->setNodeId(Processed);
for (SDNode::use_iterator UI = N->use_begin(), E = N->use_end();
UI != E; ++UI) {
SDNode *User = *UI;
int NodeId = User->getNodeId();
// This node has two options: it can either be a new node or its Node ID
// may be a count of the number of operands it has that are not ready.
if (NodeId > 0) {
User->setNodeId(NodeId-1);
// If this was the last use it was waiting on, add it to the ready list.
if (NodeId-1 == ReadyToProcess)
Worklist.push_back(User);
continue;
}
// If this is an unreachable new node, then ignore it. If it ever becomes
// reachable by being used by a newly created node then it will be handled
// by AnalyzeNewNode.
if (NodeId == NewNode)
continue;
// Otherwise, this node is new: this is the first operand of it that
// became ready. Its new NodeId is the number of operands it has minus 1
// (as this node is now processed).
assert(NodeId == Unanalyzed && "Unknown node ID!");
User->setNodeId(User->getNumOperands() - 1);
// If the node only has a single operand, it is now ready.
if (User->getNumOperands() == 1)
Worklist.push_back(User);
}
}
/// CopyAndMoveSuccessors - Clone the specified node and move its scheduled
/// successors to the newly created node.
SUnit *ScheduleDAGFast::CopyAndMoveSuccessors(SUnit *SU) {
if (SU->getNode()->getGluedNode())
return nullptr;
SDNode *N = SU->getNode();
if (!N)
return nullptr;
SUnit *NewSU;
bool TryUnfold = false;
for (unsigned i = 0, e = N->getNumValues(); i != e; ++i) {
MVT VT = N->getSimpleValueType(i);
if (VT == MVT::Glue)
return nullptr;
else if (VT == MVT::Other)
TryUnfold = true;
}
for (const SDValue &Op : N->op_values()) {
MVT VT = Op.getNode()->getSimpleValueType(Op.getResNo());
if (VT == MVT::Glue)
return nullptr;
}
if (TryUnfold) {
SmallVector<SDNode*, 2> NewNodes;
if (!TII->unfoldMemoryOperand(*DAG, N, NewNodes))
return nullptr;
LLVM_DEBUG(dbgs() << "Unfolding SU # " << SU->NodeNum << "\n");
assert(NewNodes.size() == 2 && "Expected a load folding node!");
N = NewNodes[1];
SDNode *LoadNode = NewNodes[0];
unsigned NumVals = N->getNumValues();
unsigned OldNumVals = SU->getNode()->getNumValues();
for (unsigned i = 0; i != NumVals; ++i)
DAG->ReplaceAllUsesOfValueWith(SDValue(SU->getNode(), i), SDValue(N, i));
DAG->ReplaceAllUsesOfValueWith(SDValue(SU->getNode(), OldNumVals-1),
SDValue(LoadNode, 1));
SUnit *NewSU = newSUnit(N);
assert(N->getNodeId() == -1 && "Node already inserted!");
N->setNodeId(NewSU->NodeNum);
const MCInstrDesc &MCID = TII->get(N->getMachineOpcode());
for (unsigned i = 0; i != MCID.getNumOperands(); ++i) {
if (MCID.getOperandConstraint(i, MCOI::TIED_TO) != -1) {
NewSU->isTwoAddress = true;
break;
}
}
if (MCID.isCommutable())
NewSU->isCommutable = true;
// LoadNode may already exist. This can happen when there is another
// load from the same location and producing the same type of value
// but it has different alignment or volatileness.
bool isNewLoad = true;
SUnit *LoadSU;
if (LoadNode->getNodeId() != -1) {
LoadSU = &SUnits[LoadNode->getNodeId()];
isNewLoad = false;
} else {
LoadSU = newSUnit(LoadNode);
LoadNode->setNodeId(LoadSU->NodeNum);
}
SDep ChainPred;
SmallVector<SDep, 4> ChainSuccs;
SmallVector<SDep, 4> LoadPreds;
SmallVector<SDep, 4> NodePreds;
SmallVector<SDep, 4> NodeSuccs;
for (SDep &Pred : SU->Preds) {
if (Pred.isCtrl())
ChainPred = Pred;
else if (Pred.getSUnit()->getNode() &&
Pred.getSUnit()->getNode()->isOperandOf(LoadNode))
LoadPreds.push_back(Pred);
else
NodePreds.push_back(Pred);
}
for (SDep &Succ : SU->Succs) {
if (Succ.isCtrl())
ChainSuccs.push_back(Succ);
else
NodeSuccs.push_back(Succ);
}
if (ChainPred.getSUnit()) {
RemovePred(SU, ChainPred);
if (isNewLoad)
AddPred(LoadSU, ChainPred);
}
for (unsigned i = 0, e = LoadPreds.size(); i != e; ++i) {
const SDep &Pred = LoadPreds[i];
RemovePred(SU, Pred);
if (isNewLoad) {
AddPred(LoadSU, Pred);
}
}
for (unsigned i = 0, e = NodePreds.size(); i != e; ++i) {
const SDep &Pred = NodePreds[i];
RemovePred(SU, Pred);
AddPred(NewSU, Pred);
}
for (unsigned i = 0, e = NodeSuccs.size(); i != e; ++i) {
SDep D = NodeSuccs[i];
SUnit *SuccDep = D.getSUnit();
D.setSUnit(SU);
RemovePred(SuccDep, D);
D.setSUnit(NewSU);
AddPred(SuccDep, D);
}
for (unsigned i = 0, e = ChainSuccs.size(); i != e; ++i) {
SDep D = ChainSuccs[i];
SUnit *SuccDep = D.getSUnit();
D.setSUnit(SU);
RemovePred(SuccDep, D);
if (isNewLoad) {
D.setSUnit(LoadSU);
AddPred(SuccDep, D);
}
}
if (isNewLoad) {
SDep D(LoadSU, SDep::Barrier);
D.setLatency(LoadSU->Latency);
AddPred(NewSU, D);
}
++NumUnfolds;
if (NewSU->NumSuccsLeft == 0) {
NewSU->isAvailable = true;
return NewSU;
}
SU = NewSU;
}
LLVM_DEBUG(dbgs() << "Duplicating SU # " << SU->NodeNum << "\n");
NewSU = Clone(SU);
// New SUnit has the exact same predecessors.
for (SDep &Pred : SU->Preds)
if (!Pred.isArtificial())
AddPred(NewSU, Pred);
// Only copy scheduled successors. Cut them from old node's successor
// list and move them over.
SmallVector<std::pair<SUnit *, SDep>, 4> DelDeps;
for (SDep &Succ : SU->Succs) {
if (Succ.isArtificial())
continue;
SUnit *SuccSU = Succ.getSUnit();
if (SuccSU->isScheduled) {
SDep D = Succ;
D.setSUnit(NewSU);
AddPred(SuccSU, D);
D.setSUnit(SU);
DelDeps.push_back(std::make_pair(SuccSU, D));
}
}
for (unsigned i = 0, e = DelDeps.size(); i != e; ++i)
RemovePred(DelDeps[i].first, DelDeps[i].second);
++NumDups;
return NewSU;
}
void ScheduleDAGSDNodes::AddSchedEdges() {
const TargetSubtargetInfo &ST = TM.getSubtarget<TargetSubtargetInfo>();
// Check to see if the scheduler cares about latencies.
bool UnitLatencies = forceUnitLatencies();
// Pass 2: add the preds, succs, etc.
for (unsigned su = 0, e = SUnits.size(); su != e; ++su) {
SUnit *SU = &SUnits[su];
SDNode *MainNode = SU->getNode();
if (MainNode->isMachineOpcode()) {
unsigned Opc = MainNode->getMachineOpcode();
const MCInstrDesc &MCID = TII->get(Opc);
for (unsigned i = 0; i != MCID.getNumOperands(); ++i) {
if (MCID.getOperandConstraint(i, MCOI::TIED_TO) != -1) {
SU->isTwoAddress = true;
break;
}
}
if (MCID.isCommutable())
SU->isCommutable = true;
}
// Find all predecessors and successors of the group.
for (SDNode *N = SU->getNode(); N; N = N->getGluedNode()) {
if (N->isMachineOpcode() &&
TII->get(N->getMachineOpcode()).getImplicitDefs()) {
SU->hasPhysRegClobbers = true;
unsigned NumUsed = InstrEmitter::CountResults(N);
while (NumUsed != 0 && !N->hasAnyUseOfValue(NumUsed - 1))
--NumUsed; // Skip over unused values at the end.
if (NumUsed > TII->get(N->getMachineOpcode()).getNumDefs())
SU->hasPhysRegDefs = true;
}
for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
SDNode *OpN = N->getOperand(i).getNode();
if (isPassiveNode(OpN)) continue; // Not scheduled.
SUnit *OpSU = &SUnits[OpN->getNodeId()];
assert(OpSU && "Node has no SUnit!");
if (OpSU == SU) continue; // In the same group.
EVT OpVT = N->getOperand(i).getValueType();
assert(OpVT != MVT::Glue && "Glued nodes should be in same sunit!");
bool isChain = OpVT == MVT::Other;
unsigned PhysReg = 0;
int Cost = 1;
// Determine if this is a physical register dependency.
CheckForPhysRegDependency(OpN, N, i, TRI, TII, PhysReg, Cost);
assert((PhysReg == 0 || !isChain) &&
"Chain dependence via physreg data?");
// FIXME: See ScheduleDAGSDNodes::EmitCopyFromReg. For now, scheduler
// emits a copy from the physical register to a virtual register unless
// it requires a cross class copy (cost < 0). That means we are only
// treating "expensive to copy" register dependency as physical register
// dependency. This may change in the future though.
if (Cost >= 0 && !StressSched)
PhysReg = 0;
// If this is a ctrl dep, latency is 1.
unsigned OpLatency = isChain ? 1 : OpSU->Latency;
// Special-case TokenFactor chains as zero-latency.
if(isChain && OpN->getOpcode() == ISD::TokenFactor)
OpLatency = 0;
const SDep &dep = SDep(OpSU, isChain ? SDep::Order : SDep::Data,
OpLatency, PhysReg);
if (!isChain && !UnitLatencies) {
computeOperandLatency(OpN, N, i, const_cast<SDep &>(dep));
ST.adjustSchedDependency(OpSU, SU, const_cast<SDep &>(dep));
}
if (!SU->addPred(dep) && !dep.isCtrl() && OpSU->NumRegDefsLeft > 1) {
// Multiple register uses are combined in the same SUnit. For example,
// we could have a set of glued nodes with all their defs consumed by
// another set of glued nodes. Register pressure tracking sees this as
// a single use, so to keep pressure balanced we reduce the defs.
//
// We can't tell (without more book-keeping) if this results from
// glued nodes or duplicate operands. As long as we don't reduce
// NumRegDefsLeft to zero, we handle the common cases well.
--OpSU->NumRegDefsLeft;
}
}
}
}
}
void ScheduleDAGSDNodes::BuildSchedUnits() {
// During scheduling, the NodeId field of SDNode is used to map SDNodes
// to their associated SUnits by holding SUnits table indices. A value
// of -1 means the SDNode does not yet have an associated SUnit.
unsigned NumNodes = 0;
for (SelectionDAG::allnodes_iterator NI = DAG->allnodes_begin(),
E = DAG->allnodes_end(); NI != E; ++NI) {
NI->setNodeId(-1);
++NumNodes;
}
// Reserve entries in the vector for each of the SUnits we are creating. This
// ensure that reallocation of the vector won't happen, so SUnit*'s won't get
// invalidated.
// FIXME: Multiply by 2 because we may clone nodes during scheduling.
// This is a temporary workaround.
SUnits.reserve(NumNodes * 2);
// Add all nodes in depth first order.
SmallVector<SDNode*, 64> Worklist;
SmallPtrSet<SDNode*, 64> Visited;
Worklist.push_back(DAG->getRoot().getNode());
Visited.insert(DAG->getRoot().getNode());
SmallVector<SUnit*, 8> CallSUnits;
while (!Worklist.empty()) {
SDNode *NI = Worklist.pop_back_val();
// Add all operands to the worklist unless they've already been added.
for (unsigned i = 0, e = NI->getNumOperands(); i != e; ++i)
if (Visited.insert(NI->getOperand(i).getNode()))
Worklist.push_back(NI->getOperand(i).getNode());
if (isPassiveNode(NI)) // Leaf node, e.g. a TargetImmediate.
continue;
// If this node has already been processed, stop now.
if (NI->getNodeId() != -1) continue;
SUnit *NodeSUnit = newSUnit(NI);
// See if anything is glued to this node, if so, add them to glued
// nodes. Nodes can have at most one glue input and one glue output. Glue
// is required to be the last operand and result of a node.
// Scan up to find glued preds.
SDNode *N = NI;
while (N->getNumOperands() &&
N->getOperand(N->getNumOperands()-1).getValueType() == MVT::Glue) {
N = N->getOperand(N->getNumOperands()-1).getNode();
assert(N->getNodeId() == -1 && "Node already inserted!");
N->setNodeId(NodeSUnit->NodeNum);
if (N->isMachineOpcode() && TII->get(N->getMachineOpcode()).isCall())
NodeSUnit->isCall = true;
}
// Scan down to find any glued succs.
N = NI;
while (N->getValueType(N->getNumValues()-1) == MVT::Glue) {
SDValue GlueVal(N, N->getNumValues()-1);
// There are either zero or one users of the Glue result.
bool HasGlueUse = false;
for (SDNode::use_iterator UI = N->use_begin(), E = N->use_end();
UI != E; ++UI)
if (GlueVal.isOperandOf(*UI)) {
HasGlueUse = true;
assert(N->getNodeId() == -1 && "Node already inserted!");
N->setNodeId(NodeSUnit->NodeNum);
N = *UI;
if (N->isMachineOpcode() && TII->get(N->getMachineOpcode()).isCall())
NodeSUnit->isCall = true;
break;
}
if (!HasGlueUse) break;
}
if (NodeSUnit->isCall)
CallSUnits.push_back(NodeSUnit);
// Schedule zero-latency TokenFactor below any nodes that may increase the
// schedule height. Otherwise, ancestors of the TokenFactor may appear to
// have false stalls.
if (NI->getOpcode() == ISD::TokenFactor)
NodeSUnit->isScheduleLow = true;
// If there are glue operands involved, N is now the bottom-most node
// of the sequence of nodes that are glued together.
// Update the SUnit.
NodeSUnit->setNode(N);
assert(N->getNodeId() == -1 && "Node already inserted!");
N->setNodeId(NodeSUnit->NodeNum);
// Compute NumRegDefsLeft. This must be done before AddSchedEdges.
InitNumRegDefsLeft(NodeSUnit);
// Assign the Latency field of NodeSUnit using target-provided information.
computeLatency(NodeSUnit);
}
// Find all call operands.
while (!CallSUnits.empty()) {
SUnit *SU = CallSUnits.pop_back_val();
for (const SDNode *SUNode = SU->getNode(); SUNode;
SUNode = SUNode->getGluedNode()) {
if (SUNode->getOpcode() != ISD::CopyToReg)
continue;
SDNode *SrcN = SUNode->getOperand(2).getNode();
if (isPassiveNode(SrcN)) continue; // Not scheduled.
SUnit *SrcSU = &SUnits[SrcN->getNodeId()];
SrcSU->isCallOp = true;
}
}
}
void ScheduleDAGSDNodes::AddSchedEdges() {
const TargetSubtarget &ST = TM.getSubtarget<TargetSubtarget>();
// Check to see if the scheduler cares about latencies.
bool UnitLatencies = ForceUnitLatencies();
// Pass 2: add the preds, succs, etc.
for (unsigned su = 0, e = SUnits.size(); su != e; ++su) {
SUnit *SU = &SUnits[su];
SDNode *MainNode = SU->getNode();
if (MainNode->isMachineOpcode()) {
unsigned Opc = MainNode->getMachineOpcode();
const TargetInstrDesc &TID = TII->get(Opc);
for (unsigned i = 0; i != TID.getNumOperands(); ++i) {
if (TID.getOperandConstraint(i, TOI::TIED_TO) != -1) {
SU->isTwoAddress = true;
break;
}
}
if (TID.isCommutable())
SU->isCommutable = true;
}
// Find all predecessors and successors of the group.
for (SDNode *N = SU->getNode(); N; N = N->getGluedNode()) {
if (N->isMachineOpcode() &&
TII->get(N->getMachineOpcode()).getImplicitDefs()) {
SU->hasPhysRegClobbers = true;
unsigned NumUsed = InstrEmitter::CountResults(N);
while (NumUsed != 0 && !N->hasAnyUseOfValue(NumUsed - 1))
--NumUsed; // Skip over unused values at the end.
if (NumUsed > TII->get(N->getMachineOpcode()).getNumDefs())
SU->hasPhysRegDefs = true;
}
for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
SDNode *OpN = N->getOperand(i).getNode();
if (isPassiveNode(OpN)) continue; // Not scheduled.
SUnit *OpSU = &SUnits[OpN->getNodeId()];
assert(OpSU && "Node has no SUnit!");
if (OpSU == SU) continue; // In the same group.
EVT OpVT = N->getOperand(i).getValueType();
assert(OpVT != MVT::Glue && "Glued nodes should be in same sunit!");
bool isChain = OpVT == MVT::Other;
unsigned PhysReg = 0;
int Cost = 1;
// Determine if this is a physical register dependency.
CheckForPhysRegDependency(OpN, N, i, TRI, TII, PhysReg, Cost);
assert((PhysReg == 0 || !isChain) &&
"Chain dependence via physreg data?");
// FIXME: See ScheduleDAGSDNodes::EmitCopyFromReg. For now, scheduler
// emits a copy from the physical register to a virtual register unless
// it requires a cross class copy (cost < 0). That means we are only
// treating "expensive to copy" register dependency as physical register
// dependency. This may change in the future though.
if (Cost >= 0)
PhysReg = 0;
// If this is a ctrl dep, latency is 1.
unsigned OpLatency = isChain ? 1 : OpSU->Latency;
const SDep &dep = SDep(OpSU, isChain ? SDep::Order : SDep::Data,
OpLatency, PhysReg);
if (!isChain && !UnitLatencies) {
ComputeOperandLatency(OpN, N, i, const_cast<SDep &>(dep));
ST.adjustSchedDependency(OpSU, SU, const_cast<SDep &>(dep));
}
SU->addPred(dep);
}
}
}
}
void ScheduleDAGSDNodes::BuildSchedUnits() {
// During scheduling, the NodeId field of SDNode is used to map SDNodes
// to their associated SUnits by holding SUnits table indices. A value
// of -1 means the SDNode does not yet have an associated SUnit.
unsigned NumNodes = 0;
for (SelectionDAG::allnodes_iterator NI = DAG->allnodes_begin(),
E = DAG->allnodes_end(); NI != E; ++NI) {
NI->setNodeId(-1);
++NumNodes;
}
// Reserve entries in the vector for each of the SUnits we are creating. This
// ensure that reallocation of the vector won't happen, so SUnit*'s won't get
// invalidated.
// FIXME: Multiply by 2 because we may clone nodes during scheduling.
// This is a temporary workaround.
SUnits.reserve(NumNodes * 2);
// Check to see if the scheduler cares about latencies.
bool UnitLatencies = ForceUnitLatencies();
// Add all nodes in depth first order.
SmallVector<SDNode*, 64> Worklist;
SmallPtrSet<SDNode*, 64> Visited;
Worklist.push_back(DAG->getRoot().getNode());
Visited.insert(DAG->getRoot().getNode());
while (!Worklist.empty()) {
SDNode *NI = Worklist.pop_back_val();
// Add all operands to the worklist unless they've already been added.
for (unsigned i = 0, e = NI->getNumOperands(); i != e; ++i)
if (Visited.insert(NI->getOperand(i).getNode()))
Worklist.push_back(NI->getOperand(i).getNode());
if (isPassiveNode(NI)) // Leaf node, e.g. a TargetImmediate.
continue;
// If this node has already been processed, stop now.
if (NI->getNodeId() != -1) continue;
SUnit *NodeSUnit = NewSUnit(NI);
// See if anything is flagged to this node, if so, add them to flagged
// nodes. Nodes can have at most one flag input and one flag output. Flags
// are required to be the last operand and result of a node.
// Scan up to find flagged preds.
SDNode *N = NI;
while (N->getNumOperands() &&
N->getOperand(N->getNumOperands()-1).getValueType() == MVT::Flag) {
N = N->getOperand(N->getNumOperands()-1).getNode();
assert(N->getNodeId() == -1 && "Node already inserted!");
N->setNodeId(NodeSUnit->NodeNum);
}
// Scan down to find any flagged succs.
N = NI;
while (N->getValueType(N->getNumValues()-1) == MVT::Flag) {
SDValue FlagVal(N, N->getNumValues()-1);
// There are either zero or one users of the Flag result.
bool HasFlagUse = false;
for (SDNode::use_iterator UI = N->use_begin(), E = N->use_end();
UI != E; ++UI)
if (FlagVal.isOperandOf(*UI)) {
HasFlagUse = true;
assert(N->getNodeId() == -1 && "Node already inserted!");
N->setNodeId(NodeSUnit->NodeNum);
N = *UI;
break;
}
if (!HasFlagUse) break;
}
// If there are flag operands involved, N is now the bottom-most node
// of the sequence of nodes that are flagged together.
// Update the SUnit.
NodeSUnit->setNode(N);
assert(N->getNodeId() == -1 && "Node already inserted!");
N->setNodeId(NodeSUnit->NodeNum);
// Assign the Latency field of NodeSUnit using target-provided information.
if (UnitLatencies)
NodeSUnit->Latency = 1;
else
ComputeLatency(NodeSUnit);
}
}
