/// EmitSchedule - Emit the machine code in scheduled order.
MachineBasicBlock *ScheduleDAGSDNodes::EmitSchedule() {
DenseMap<SDValue, unsigned> VRBaseMap;
DenseMap<SUnit*, unsigned> CopyVRBaseMap;
for (unsigned i = 0, e = Sequence.size(); i != e; i++) {
SUnit *SU = Sequence[i];
if (!SU) {
// Null SUnit* is a noop.
EmitNoop();
continue;
}
// For pre-regalloc scheduling, create instructions corresponding to the
// SDNode and any flagged SDNodes and append them to the block.
if (!SU->getNode()) {
// Emit a copy.
EmitPhysRegCopy(SU, CopyVRBaseMap);
continue;
}
SmallVector<SDNode *, 4> FlaggedNodes;
for (SDNode *N = SU->getNode()->getFlaggedNode(); N;
N = N->getFlaggedNode())
FlaggedNodes.push_back(N);
while (!FlaggedNodes.empty()) {
EmitNode(FlaggedNodes.back(), SU->OrigNode != SU, SU->isCloned,VRBaseMap);
FlaggedNodes.pop_back();
}
EmitNode(SU->getNode(), SU->OrigNode != SU, SU->isCloned, VRBaseMap);
}
return BB;
}
void ScheduleDAGSDNodes::dumpNode(const SUnit &SU) const {
#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
dumpNodeName(SU);
dbgs() << ": ";
if (!SU.getNode()) {
dbgs() << "PHYS REG COPY\n";
return;
}
SU.getNode()->dump(DAG);
dbgs() << "\n";
SmallVector<SDNode *, 4> GluedNodes;
for (SDNode *N = SU.getNode()->getGluedNode(); N; N = N->getGluedNode())
GluedNodes.push_back(N);
while (!GluedNodes.empty()) {
dbgs() << " ";
GluedNodes.back()->dump(DAG);
dbgs() << "\n";
GluedNodes.pop_back();
}
#endif
}
unsigned
ResourcePriorityQueue::numberRCValPredInSU(SUnit *SU, unsigned RCId) {
unsigned NumberDeps = 0;
for (SUnit::pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
I != E; ++I) {
if (I->isCtrl())
continue;
SUnit *PredSU = I->getSUnit();
const SDNode *ScegN = PredSU->getNode();
if (!ScegN)
continue;
// If value is passed to CopyToReg, it is probably
// live outside BB.
switch (ScegN->getOpcode()) {
default:
break;
case ISD::TokenFactor:
break;
case ISD::CopyFromReg:
NumberDeps++;
break;
case ISD::CopyToReg:
break;
case ISD::INLINEASM:
break;
}
if (!ScegN->isMachineOpcode())
continue;
for (unsigned i = 0, e = ScegN->getNumValues(); i != e; ++i) {
MVT VT = ScegN->getSimpleValueType(i);
if (TLI->isTypeLegal(VT)
&& (TLI->getRegClassFor(VT)->getID() == RCId)) {
NumberDeps++;
break;
}
}
}
return NumberDeps;
}
unsigned ResourcePriorityQueue::numberRCValSuccInSU(SUnit *SU,
unsigned RCId) {
unsigned NumberDeps = 0;
for (const SDep &Succ : SU->Succs) {
if (Succ.isCtrl())
continue;
SUnit *SuccSU = Succ.getSUnit();
const SDNode *ScegN = SuccSU->getNode();
if (!ScegN)
continue;
// If value is passed to CopyToReg, it is probably
// live outside BB.
switch (ScegN->getOpcode()) {
default: break;
case ISD::TokenFactor: break;
case ISD::CopyFromReg: break;
case ISD::CopyToReg: NumberDeps++; break;
case ISD::INLINEASM: break;
}
if (!ScegN->isMachineOpcode())
continue;
for (unsigned i = 0, e = ScegN->getNumOperands(); i != e; ++i) {
const SDValue &Op = ScegN->getOperand(i);
MVT VT = Op.getNode()->getSimpleValueType(Op.getResNo());
if (TLI->isTypeLegal(VT)
&& (TLI->getRegClassFor(VT)->getID() == RCId)) {
NumberDeps++;
break;
}
}
}
return NumberDeps;
}
/// ListScheduleBottomUp - The main loop of list scheduling for bottom-up
/// schedulers.
void ScheduleDAGFast::ListScheduleBottomUp() {
unsigned CurCycle = 0;
// Release any predecessors of the special Exit node.
ReleasePredecessors(&ExitSU, CurCycle);
// Add root to Available queue.
if (!SUnits.empty()) {
SUnit *RootSU = &SUnits[DAG->getRoot().getNode()->getNodeId()];
assert(RootSU->Succs.empty() && "Graph root shouldn't have successors!");
RootSU->isAvailable = true;
AvailableQueue.push(RootSU);
}
// While Available queue is not empty, grab the node with the highest
// priority. If it is not ready put it back. Schedule the node.
SmallVector<SUnit*, 4> NotReady;
DenseMap<SUnit*, SmallVector<unsigned, 4> > LRegsMap;
Sequence.reserve(SUnits.size());
while (!AvailableQueue.empty()) {
bool Delayed = false;
LRegsMap.clear();
SUnit *CurSU = AvailableQueue.pop();
while (CurSU) {
SmallVector<unsigned, 4> LRegs;
if (!DelayForLiveRegsBottomUp(CurSU, LRegs))
break;
Delayed = true;
LRegsMap.insert(std::make_pair(CurSU, LRegs));
CurSU->isPending = true; // This SU is not in AvailableQueue right now.
NotReady.push_back(CurSU);
CurSU = AvailableQueue.pop();
}
// All candidates are delayed due to live physical reg dependencies.
// Try code duplication or inserting cross class copies
// to resolve it.
if (Delayed && !CurSU) {
if (!CurSU) {
// Try duplicating the nodes that produces these
// "expensive to copy" values to break the dependency. In case even
// that doesn't work, insert cross class copies.
SUnit *TrySU = NotReady[0];
SmallVectorImpl<unsigned> &LRegs = LRegsMap[TrySU];
assert(LRegs.size() == 1 && "Can't handle this yet!");
unsigned Reg = LRegs[0];
SUnit *LRDef = LiveRegDefs[Reg];
MVT VT = getPhysicalRegisterVT(LRDef->getNode(), Reg, TII);
const TargetRegisterClass *RC =
TRI->getMinimalPhysRegClass(Reg, VT);
const TargetRegisterClass *DestRC = TRI->getCrossCopyRegClass(RC);
// If cross copy register class is the same as RC, then it must be
// possible copy the value directly. Do not try duplicate the def.
// If cross copy register class is not the same as RC, then it's
// possible to copy the value but it require cross register class copies
// and it is expensive.
// If cross copy register class is null, then it's not possible to copy
// the value at all.
SUnit *NewDef = nullptr;
if (DestRC != RC) {
NewDef = CopyAndMoveSuccessors(LRDef);
if (!DestRC && !NewDef)
report_fatal_error("Can't handle live physical "
"register dependency!");
}
if (!NewDef) {
// Issue copies, these can be expensive cross register class copies.
SmallVector<SUnit*, 2> Copies;
InsertCopiesAndMoveSuccs(LRDef, Reg, DestRC, RC, Copies);
LLVM_DEBUG(dbgs() << "Adding an edge from SU # " << TrySU->NodeNum
<< " to SU #" << Copies.front()->NodeNum << "\n");
AddPred(TrySU, SDep(Copies.front(), SDep::Artificial));
NewDef = Copies.back();
}
LLVM_DEBUG(dbgs() << "Adding an edge from SU # " << NewDef->NodeNum
<< " to SU #" << TrySU->NodeNum << "\n");
LiveRegDefs[Reg] = NewDef;
AddPred(NewDef, SDep(TrySU, SDep::Artificial));
TrySU->isAvailable = false;
CurSU = NewDef;
}
if (!CurSU) {
llvm_unreachable("Unable to resolve live physical register dependencies!");
}
}
// Add the nodes that aren't ready back onto the available list.
for (unsigned i = 0, e = NotReady.size(); i != e; ++i) {
NotReady[i]->isPending = false;
// May no longer be available due to backtracking.
if (NotReady[i]->isAvailable)
AvailableQueue.push(NotReady[i]);
}
NotReady.clear();
if (CurSU)
ScheduleNodeBottomUp(CurSU, CurCycle);
++CurCycle;
}
// Reverse the order since it is bottom up.
std::reverse(Sequence.begin(), Sequence.end());
#ifndef NDEBUG
VerifyScheduledSequence(/*isBottomUp=*/true);
#endif
}
/// EmitSchedule - Emit the machine code in scheduled order. Return the new
/// InsertPos and MachineBasicBlock that contains this insertion
/// point. ScheduleDAGSDNodes holds a BB pointer for convenience, but this does
/// not necessarily refer to returned BB. The emitter may split blocks.
MachineBasicBlock *ScheduleDAGSDNodes::
EmitSchedule(MachineBasicBlock::iterator &InsertPos) {
InstrEmitter Emitter(BB, InsertPos);
DenseMap<SDValue, unsigned> VRBaseMap;
DenseMap<SUnit*, unsigned> CopyVRBaseMap;
SmallVector<std::pair<unsigned, MachineInstr*>, 32> Orders;
SmallSet<unsigned, 8> Seen;
bool HasDbg = DAG->hasDebugValues();
// If this is the first BB, emit byval parameter dbg_value's.
if (HasDbg && BB->getParent()->begin() == MachineFunction::iterator(BB)) {
SDDbgInfo::DbgIterator PDI = DAG->ByvalParmDbgBegin();
SDDbgInfo::DbgIterator PDE = DAG->ByvalParmDbgEnd();
for (; PDI != PDE; ++PDI) {
MachineInstr *DbgMI= Emitter.EmitDbgValue(*PDI, VRBaseMap);
if (DbgMI)
BB->insert(InsertPos, DbgMI);
}
}
for (unsigned i = 0, e = Sequence.size(); i != e; i++) {
SUnit *SU = Sequence[i];
if (!SU) {
// Null SUnit* is a noop.
TII->insertNoop(*Emitter.getBlock(), InsertPos);
continue;
}
// For pre-regalloc scheduling, create instructions corresponding to the
// SDNode and any glued SDNodes and append them to the block.
if (!SU->getNode()) {
// Emit a copy.
EmitPhysRegCopy(SU, CopyVRBaseMap, InsertPos);
continue;
}
SmallVector<SDNode *, 4> GluedNodes;
for (SDNode *N = SU->getNode()->getGluedNode(); N;
N = N->getGluedNode())
GluedNodes.push_back(N);
while (!GluedNodes.empty()) {
SDNode *N = GluedNodes.back();
Emitter.EmitNode(GluedNodes.back(), SU->OrigNode != SU, SU->isCloned,
VRBaseMap);
// Remember the source order of the inserted instruction.
if (HasDbg)
ProcessSourceNode(N, DAG, Emitter, VRBaseMap, Orders, Seen);
GluedNodes.pop_back();
}
Emitter.EmitNode(SU->getNode(), SU->OrigNode != SU, SU->isCloned,
VRBaseMap);
// Remember the source order of the inserted instruction.
if (HasDbg)
ProcessSourceNode(SU->getNode(), DAG, Emitter, VRBaseMap, Orders,
Seen);
}
// Insert all the dbg_values which have not already been inserted in source
// order sequence.
if (HasDbg) {
MachineBasicBlock::iterator BBBegin = BB->getFirstNonPHI();
// Sort the source order instructions and use the order to insert debug
// values.
std::sort(Orders.begin(), Orders.end(), OrderSorter());
SDDbgInfo::DbgIterator DI = DAG->DbgBegin();
SDDbgInfo::DbgIterator DE = DAG->DbgEnd();
// Now emit the rest according to source order.
unsigned LastOrder = 0;
for (unsigned i = 0, e = Orders.size(); i != e && DI != DE; ++i) {
unsigned Order = Orders[i].first;
MachineInstr *MI = Orders[i].second;
// Insert all SDDbgValue's whose order(s) are before "Order".
if (!MI)
continue;
for (; DI != DE &&
(*DI)->getOrder() >= LastOrder && (*DI)->getOrder() < Order; ++DI) {
if ((*DI)->isInvalidated())
continue;
MachineInstr *DbgMI = Emitter.EmitDbgValue(*DI, VRBaseMap);
if (DbgMI) {
if (!LastOrder)
// Insert to start of the BB (after PHIs).
BB->insert(BBBegin, DbgMI);
else {
// Insert at the instruction, which may be in a different
// block, if the block was split by a custom inserter.
MachineBasicBlock::iterator Pos = MI;
MI->getParent()->insert(llvm::next(Pos), DbgMI);
}
}
}
LastOrder = Order;
}
// Add trailing DbgValue's before the terminator. FIXME: May want to add
// some of them before one or more conditional branches?
SmallVector<MachineInstr*, 8> DbgMIs;
while (DI != DE) {
if (!(*DI)->isInvalidated())
if (MachineInstr *DbgMI = Emitter.EmitDbgValue(*DI, VRBaseMap))
DbgMIs.push_back(DbgMI);
++DI;
}
MachineBasicBlock *InsertBB = Emitter.getBlock();
MachineBasicBlock::iterator Pos = InsertBB->getFirstTerminator();
InsertBB->insert(Pos, DbgMIs.begin(), DbgMIs.end());
}
InsertPos = Emitter.getInsertPos();
return Emitter.getBlock();
}
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);
}
}
}
}
/// EmitSchedule - Emit the machine code in scheduled order.
MachineBasicBlock *ScheduleDAGSDNodes::
EmitSchedule(DenseMap<MachineBasicBlock*, MachineBasicBlock*> *EM) {
InstrEmitter Emitter(BB, InsertPos);
DenseMap<SDValue, unsigned> VRBaseMap;
DenseMap<SUnit*, unsigned> CopyVRBaseMap;
SmallVector<std::pair<unsigned, MachineInstr*>, 32> Orders;
SmallSet<unsigned, 8> Seen;
bool HasDbg = DAG->hasDebugValues();
for (unsigned i = 0, e = Sequence.size(); i != e; i++) {
SUnit *SU = Sequence[i];
if (!SU) {
// Null SUnit* is a noop.
EmitNoop();
continue;
}
// For pre-regalloc scheduling, create instructions corresponding to the
// SDNode and any flagged SDNodes and append them to the block.
if (!SU->getNode()) {
// Emit a copy.
EmitPhysRegCopy(SU, CopyVRBaseMap);
continue;
}
SmallVector<SDNode *, 4> FlaggedNodes;
for (SDNode *N = SU->getNode()->getFlaggedNode(); N;
N = N->getFlaggedNode())
FlaggedNodes.push_back(N);
while (!FlaggedNodes.empty()) {
SDNode *N = FlaggedNodes.back();
Emitter.EmitNode(FlaggedNodes.back(), SU->OrigNode != SU, SU->isCloned,
VRBaseMap, EM);
// Remember the the source order of the inserted instruction.
if (HasDbg)
ProcessSourceNode(N, DAG, Emitter, EM, VRBaseMap, Orders, Seen);
FlaggedNodes.pop_back();
}
Emitter.EmitNode(SU->getNode(), SU->OrigNode != SU, SU->isCloned,
VRBaseMap, EM);
// Remember the the source order of the inserted instruction.
if (HasDbg)
ProcessSourceNode(SU->getNode(), DAG, Emitter, EM, VRBaseMap, Orders,
Seen);
}
// Insert all the dbg_value which have not already been inserted in source
// order sequence.
if (HasDbg) {
MachineBasicBlock::iterator BBBegin = BB->empty() ? BB->end() : BB->begin();
while (BBBegin != BB->end() && BBBegin->isPHI())
++BBBegin;
// Sort the source order instructions and use the order to insert debug
// values.
std::sort(Orders.begin(), Orders.end(), OrderSorter());
SDDbgInfo::DbgIterator DI = DAG->DbgBegin();
SDDbgInfo::DbgIterator DE = DAG->DbgEnd();
// Now emit the rest according to source order.
unsigned LastOrder = 0;
MachineInstr *LastMI = 0;
for (unsigned i = 0, e = Orders.size(); i != e && DI != DE; ++i) {
unsigned Order = Orders[i].first;
MachineInstr *MI = Orders[i].second;
// Insert all SDDbgValue's whose order(s) are before "Order".
if (!MI)
continue;
MachineBasicBlock *MIBB = MI->getParent();
#ifndef NDEBUG
unsigned LastDIOrder = 0;
#endif
for (; DI != DE &&
(*DI)->getOrder() >= LastOrder && (*DI)->getOrder() < Order; ++DI) {
#ifndef NDEBUG
assert((*DI)->getOrder() >= LastDIOrder &&
"SDDbgValue nodes must be in source order!");
LastDIOrder = (*DI)->getOrder();
#endif
if ((*DI)->isInvalidated())
continue;
MachineInstr *DbgMI = Emitter.EmitDbgValue(*DI, MIBB, VRBaseMap, EM);
if (!LastOrder)
// Insert to start of the BB (after PHIs).
BB->insert(BBBegin, DbgMI);
else {
MachineBasicBlock::iterator Pos = MI;
MIBB->insert(llvm::next(Pos), DbgMI);
}
}
LastOrder = Order;
LastMI = MI;
}
// Add trailing DbgValue's before the terminator. FIXME: May want to add
// some of them before one or more conditional branches?
while (DI != DE) {
MachineBasicBlock *InsertBB = Emitter.getBlock();
MachineBasicBlock::iterator Pos= Emitter.getBlock()->getFirstTerminator();
if (!(*DI)->isInvalidated()) {
MachineInstr *DbgMI= Emitter.EmitDbgValue(*DI, InsertBB, VRBaseMap, EM);
InsertBB->insert(Pos, DbgMI);
}
++DI;
}
}
BB = Emitter.getBlock();
InsertPos = Emitter.getInsertPos();
return BB;
}
/// EmitSchedule - Emit the machine code in scheduled order. Return the new
/// InsertPos and MachineBasicBlock that contains this insertion
/// point. ScheduleDAGSDNodes holds a BB pointer for convenience, but this does
/// not necessarily refer to returned BB. The emitter may split blocks.
MachineBasicBlock *ScheduleDAGSDNodes::
EmitSchedule(MachineBasicBlock::iterator &InsertPos) {
InstrEmitter Emitter(BB, InsertPos);
DenseMap<SDValue, unsigned> VRBaseMap;
DenseMap<SUnit*, unsigned> CopyVRBaseMap;
SmallVector<std::pair<unsigned, MachineInstr*>, 32> Orders;
SmallSet<unsigned, 8> Seen;
bool HasDbg = DAG->hasDebugValues();
// Emit a node, and determine where its first instruction is for debuginfo.
// Zero, one, or multiple instructions can be created when emitting a node.
auto EmitNode =
[&](SDNode *Node, bool IsClone, bool IsCloned,
DenseMap<SDValue, unsigned> &VRBaseMap) -> MachineInstr * {
// Fetch instruction prior to this, or end() if nonexistant.
auto GetPrevInsn = [&](MachineBasicBlock::iterator I) {
if (I == BB->begin())
return BB->end();
else
return std::prev(Emitter.getInsertPos());
};
MachineBasicBlock::iterator Before = GetPrevInsn(Emitter.getInsertPos());
Emitter.EmitNode(Node, IsClone, IsCloned, VRBaseMap);
MachineBasicBlock::iterator After = GetPrevInsn(Emitter.getInsertPos());
// If the iterator did not change, no instructions were inserted.
if (Before == After)
return nullptr;
if (Before == BB->end()) {
// There were no prior instructions; the new ones must start at the
// beginning of the block.
return &Emitter.getBlock()->instr_front();
} else {
// Return first instruction after the pre-existing instructions.
return &*std::next(Before);
}
};
// If this is the first BB, emit byval parameter dbg_value's.
if (HasDbg && BB->getParent()->begin() == MachineFunction::iterator(BB)) {
SDDbgInfo::DbgIterator PDI = DAG->ByvalParmDbgBegin();
SDDbgInfo::DbgIterator PDE = DAG->ByvalParmDbgEnd();
for (; PDI != PDE; ++PDI) {
MachineInstr *DbgMI= Emitter.EmitDbgValue(*PDI, VRBaseMap);
if (DbgMI) {
BB->insert(InsertPos, DbgMI);
// We re-emit the dbg_value closer to its use, too, after instructions
// are emitted to the BB.
(*PDI)->clearIsEmitted();
}
}
}
for (unsigned i = 0, e = Sequence.size(); i != e; i++) {
SUnit *SU = Sequence[i];
if (!SU) {
// Null SUnit* is a noop.
TII->insertNoop(*Emitter.getBlock(), InsertPos);
continue;
}
// For pre-regalloc scheduling, create instructions corresponding to the
// SDNode and any glued SDNodes and append them to the block.
if (!SU->getNode()) {
// Emit a copy.
EmitPhysRegCopy(SU, CopyVRBaseMap, InsertPos);
continue;
}
SmallVector<SDNode *, 4> GluedNodes;
for (SDNode *N = SU->getNode()->getGluedNode(); N; N = N->getGluedNode())
GluedNodes.push_back(N);
while (!GluedNodes.empty()) {
SDNode *N = GluedNodes.back();
auto NewInsn = EmitNode(N, SU->OrigNode != SU, SU->isCloned, VRBaseMap);
// Remember the source order of the inserted instruction.
if (HasDbg)
ProcessSourceNode(N, DAG, Emitter, VRBaseMap, Orders, Seen, NewInsn);
GluedNodes.pop_back();
}
auto NewInsn =
EmitNode(SU->getNode(), SU->OrigNode != SU, SU->isCloned, VRBaseMap);
// Remember the source order of the inserted instruction.
if (HasDbg)
ProcessSourceNode(SU->getNode(), DAG, Emitter, VRBaseMap, Orders, Seen,
NewInsn);
}
// Insert all the dbg_values which have not already been inserted in source
// order sequence.
if (HasDbg) {
MachineBasicBlock::iterator BBBegin = BB->getFirstNonPHI();
// Sort the source order instructions and use the order to insert debug
// values. Use stable_sort so that DBG_VALUEs are inserted in the same order
// regardless of the host's implementation fo std::sort.
std::stable_sort(Orders.begin(), Orders.end(), less_first());
std::stable_sort(DAG->DbgBegin(), DAG->DbgEnd(),
[](const SDDbgValue *LHS, const SDDbgValue *RHS) {
return LHS->getOrder() < RHS->getOrder();
});
SDDbgInfo::DbgIterator DI = DAG->DbgBegin();
SDDbgInfo::DbgIterator DE = DAG->DbgEnd();
// Now emit the rest according to source order.
unsigned LastOrder = 0;
for (unsigned i = 0, e = Orders.size(); i != e && DI != DE; ++i) {
unsigned Order = Orders[i].first;
MachineInstr *MI = Orders[i].second;
// Insert all SDDbgValue's whose order(s) are before "Order".
assert(MI);
for (; DI != DE; ++DI) {
if ((*DI)->getOrder() < LastOrder || (*DI)->getOrder() >= Order)
break;
if ((*DI)->isEmitted())
continue;
MachineInstr *DbgMI = Emitter.EmitDbgValue(*DI, VRBaseMap);
if (DbgMI) {
if (!LastOrder)
// Insert to start of the BB (after PHIs).
BB->insert(BBBegin, DbgMI);
else {
// Insert at the instruction, which may be in a different
// block, if the block was split by a custom inserter.
MachineBasicBlock::iterator Pos = MI;
MI->getParent()->insert(Pos, DbgMI);
}
}
}
LastOrder = Order;
}
// Add trailing DbgValue's before the terminator. FIXME: May want to add
// some of them before one or more conditional branches?
SmallVector<MachineInstr*, 8> DbgMIs;
for (; DI != DE; ++DI) {
if ((*DI)->isEmitted())
continue;
assert((*DI)->getOrder() >= LastOrder &&
"emitting DBG_VALUE out of order");
if (MachineInstr *DbgMI = Emitter.EmitDbgValue(*DI, VRBaseMap))
DbgMIs.push_back(DbgMI);
}
MachineBasicBlock *InsertBB = Emitter.getBlock();
MachineBasicBlock::iterator Pos = InsertBB->getFirstTerminator();
InsertBB->insert(Pos, DbgMIs.begin(), DbgMIs.end());
SDDbgInfo::DbgLabelIterator DLI = DAG->DbgLabelBegin();
SDDbgInfo::DbgLabelIterator DLE = DAG->DbgLabelEnd();
// Now emit the rest according to source order.
LastOrder = 0;
for (const auto &InstrOrder : Orders) {
unsigned Order = InstrOrder.first;
MachineInstr *MI = InstrOrder.second;
if (!MI)
continue;
// Insert all SDDbgLabel's whose order(s) are before "Order".
for (; DLI != DLE &&
(*DLI)->getOrder() >= LastOrder && (*DLI)->getOrder() < Order;
++DLI) {
MachineInstr *DbgMI = Emitter.EmitDbgLabel(*DLI);
if (DbgMI) {
if (!LastOrder)
// Insert to start of the BB (after PHIs).
BB->insert(BBBegin, DbgMI);
else {
// Insert at the instruction, which may be in a different
// block, if the block was split by a custom inserter.
MachineBasicBlock::iterator Pos = MI;
MI->getParent()->insert(Pos, DbgMI);
}
}
}
if (DLI == DLE)
break;
LastOrder = Order;
}
}
InsertPos = Emitter.getInsertPos();
return Emitter.getBlock();
}