//
// The LLVM Compiler Infrastructure
//
-// This file was developed by Evan Cheng and is distributed under the
-// University of Illinois Open Source License. See LICENSE.TXT for details.
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------------===//
//
#define DEBUG_TYPE "pre-RA-sched"
#include "llvm/CodeGen/ScheduleDAG.h"
#include "llvm/CodeGen/SchedulerRegistry.h"
-#include "llvm/CodeGen/SSARegMap.h"
-#include "llvm/Target/MRegisterInfo.h"
+#include "llvm/Target/TargetRegisterInfo.h"
#include "llvm/Target/TargetData.h"
#include "llvm/Target/TargetMachine.h"
#include "llvm/Target/TargetInstrInfo.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/Compiler.h"
+#include "llvm/ADT/SmallPtrSet.h"
#include "llvm/ADT/SmallSet.h"
#include "llvm/ADT/Statistic.h"
#include <climits>
#include "llvm/Support/CommandLine.h"
using namespace llvm;
+STATISTIC(NumBacktracks, "Number of times scheduler backtraced");
+STATISTIC(NumUnfolds, "Number of nodes unfolded");
+STATISTIC(NumDups, "Number of duplicated nodes");
+STATISTIC(NumCCCopies, "Number of cross class copies");
+
static RegisterScheduler
burrListDAGScheduler("list-burr",
" Bottom-up register reduction list scheduling",
bool isBottomUp;
/// AvailableQueue - The priority queue to use for the available SUnits.
- ///a
SchedulingPriorityQueue *AvailableQueue;
/// LiveRegs / LiveRegDefs - A set of physical registers and their definition
void Schedule();
private:
- void ReleasePred(SUnit *PredSU, bool isChain, unsigned CurCycle);
- void ReleaseSucc(SUnit *SuccSU, bool isChain, unsigned CurCycle);
- void CapturePred(SUnit *PredSU, SUnit *SU, bool isChain);
- void ScheduleNodeBottomUp(SUnit *SU, unsigned CurCycle);
- void ScheduleNodeTopDown(SUnit *SU, unsigned CurCycle);
- void UnscheduleNodeBottomUp(SUnit *SU);
- SUnit *BackTrackBottomUp(SUnit*, unsigned, unsigned&, bool&);
- SUnit *CopyAndMoveSuccessors(SUnit *SU);
- bool DelayForLiveRegsBottomUp(SUnit *SU, unsigned &CurCycle);
+ void ReleasePred(SUnit*, bool, unsigned);
+ void ReleaseSucc(SUnit*, bool isChain, unsigned);
+ void CapturePred(SUnit*, SUnit*, bool);
+ void ScheduleNodeBottomUp(SUnit*, unsigned);
+ void ScheduleNodeTopDown(SUnit*, unsigned);
+ void UnscheduleNodeBottomUp(SUnit*);
+ void BacktrackBottomUp(SUnit*, unsigned, unsigned&);
+ SUnit *CopyAndMoveSuccessors(SUnit*);
+ void InsertCCCopiesAndMoveSuccs(SUnit*, unsigned,
+ const TargetRegisterClass*,
+ const TargetRegisterClass*,
+ SmallVector<SUnit*, 2>&);
+ bool DelayForLiveRegsBottomUp(SUnit*, SmallVector<unsigned, 4>&);
void ListScheduleTopDown();
void ListScheduleBottomUp();
void CommuteNodesToReducePressure();
void ScheduleDAGRRList::Schedule() {
DOUT << "********** List Scheduling **********\n";
- LiveRegDefs.resize(MRI->getNumRegs(), NULL);
- LiveRegCycles.resize(MRI->getNumRegs(), 0);
+ LiveRegDefs.resize(TRI->getNumRegs(), NULL);
+ LiveRegCycles.resize(TRI->getNumRegs(), 0);
// Build scheduling units.
BuildSchedUnits();
SmallPtrSet<SUnit*, 4> OperandSeen;
for (unsigned i = Sequence.size()-1; i != 0; --i) { // Ignore first node.
SUnit *SU = Sequence[i];
- if (!SU) continue;
+ if (!SU || !SU->Node) continue;
if (SU->isCommutable) {
unsigned Opc = SU->Node->getTargetOpcode();
- unsigned NumRes = TII->getNumDefs(Opc);
- unsigned NumOps = CountOperands(SU->Node);
+ const TargetInstrDesc &TID = TII->get(Opc);
+ unsigned NumRes = TID.getNumDefs();
+ unsigned NumOps = TID.getNumOperands() - NumRes;
for (unsigned j = 0; j != NumOps; ++j) {
- if (TII->getOperandConstraint(Opc, j+NumRes, TOI::TIED_TO) == -1)
+ if (TID.getOperandConstraint(j+NumRes, TOI::TIED_TO) == -1)
continue;
SDNode *OpN = SU->Node->getOperand(j).Val;
- SUnit *OpSU = SUnitMap[OpN][SU->InstanceNo];
+ SUnit *OpSU = isPassiveNode(OpN) ? NULL : SUnitMap[OpN][SU->InstanceNo];
if (OpSU && OperandSeen.count(OpSU) == 1) {
// Ok, so SU is not the last use of OpSU, but SU is two-address so
// it will clobber OpSU. Try to commute SU if no other source operands
for (unsigned k = 0; k < NumOps; ++k) {
if (k != j) {
OpN = SU->Node->getOperand(k).Val;
- OpSU = SUnitMap[OpN][SU->InstanceNo];
+ OpSU = isPassiveNode(OpN) ? NULL : SUnitMap[OpN][SU->InstanceNo];
if (OpSU && OperandSeen.count(OpSU) == 1) {
DoCommute = false;
break;
// interrupt model (drain vs. freeze).
PredSU->CycleBound = std::max(PredSU->CycleBound, CurCycle + PredSU->Latency);
- if (!isChain)
- --PredSU->NumSuccsLeft;
- else
- --PredSU->NumChainSuccsLeft;
+ --PredSU->NumSuccsLeft;
#ifndef NDEBUG
- if (PredSU->NumSuccsLeft < 0 || PredSU->NumChainSuccsLeft < 0) {
+ if (PredSU->NumSuccsLeft < 0) {
cerr << "*** List scheduling failed! ***\n";
PredSU->dump(&DAG);
cerr << " has been released too many times!\n";
}
#endif
- if ((PredSU->NumSuccsLeft + PredSU->NumChainSuccsLeft) == 0) {
+ if (PredSU->NumSuccsLeft == 0) {
// EntryToken has to go last! Special case it here.
- if (PredSU->Node->getOpcode() != ISD::EntryToken) {
+ if (!PredSU->Node || PredSU->Node->getOpcode() != ISD::EntryToken) {
PredSU->isAvailable = true;
AvailableQueue->push(PredSU);
}
AvailableQueue->remove(PredSU);
}
- if (!isChain)
- ++PredSU->NumSuccsLeft;
- else
- ++PredSU->NumChainSuccsLeft;
+ ++PredSU->NumSuccsLeft;
}
/// UnscheduleNodeBottomUp - Remove the node from the schedule, update its and
AvailableQueue->push(SU);
}
-/// BackTrackBottomUp - Back track scheduling to a previous cycle specified in
+// FIXME: This is probably too slow!
+static void isReachable(SUnit *SU, SUnit *TargetSU,
+ SmallPtrSet<SUnit*, 32> &Visited, bool &Reached) {
+ if (Reached) return;
+ if (SU == TargetSU) {
+ Reached = true;
+ return;
+ }
+ if (!Visited.insert(SU)) return;
+
+ for (SUnit::pred_iterator I = SU->Preds.begin(), E = SU->Preds.end(); I != E;
+ ++I)
+ isReachable(I->Dep, TargetSU, Visited, Reached);
+}
+
+static bool isReachable(SUnit *SU, SUnit *TargetSU) {
+ SmallPtrSet<SUnit*, 32> Visited;
+ bool Reached = false;
+ isReachable(SU, TargetSU, Visited, Reached);
+ return Reached;
+}
+
+/// willCreateCycle - Returns true if adding an edge from SU to TargetSU will
+/// create a cycle.
+static bool WillCreateCycle(SUnit *SU, SUnit *TargetSU) {
+ if (isReachable(TargetSU, SU))
+ return true;
+ for (SUnit::pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
+ I != E; ++I)
+ if (I->Cost < 0 && isReachable(TargetSU, I->Dep))
+ return true;
+ return false;
+}
+
+/// BacktrackBottomUp - Backtrack scheduling to a previous cycle specified in
/// BTCycle in order to schedule a specific node. Returns the last unscheduled
/// SUnit. Also returns if a successor is unscheduled in the process.
-SUnit *ScheduleDAGRRList::BackTrackBottomUp(SUnit *SU, unsigned BTCycle,
- unsigned &CurCycle, bool &SuccUnsched) {
- SuccUnsched = false;
+void ScheduleDAGRRList::BacktrackBottomUp(SUnit *SU, unsigned BtCycle,
+ unsigned &CurCycle) {
SUnit *OldSU = NULL;
- while (CurCycle > BTCycle) {
+ while (CurCycle > BtCycle) {
OldSU = Sequence.back();
Sequence.pop_back();
if (SU->isSucc(OldSU))
- SuccUnsched = true;
+ // Don't try to remove SU from AvailableQueue.
+ SU->isAvailable = false;
UnscheduleNodeBottomUp(OldSU);
--CurCycle;
}
abort();
}
- return OldSU;
+ ++NumBacktracks;
}
-/// isSafeToCopy - True if the SUnit for the given SDNode can safely cloned,
-/// i.e. the node does not produce a flag, it does not read a flag and it does
-/// not have an incoming chain.
-static bool isSafeToCopy(SDNode *N) {
- for (unsigned i = 0, e = N->getNumValues(); i != e; ++i)
- if (N->getValueType(i) == MVT::Flag)
- return false;
+/// CopyAndMoveSuccessors - Clone the specified node and move its scheduled
+/// successors to the newly created node.
+SUnit *ScheduleDAGRRList::CopyAndMoveSuccessors(SUnit *SU) {
+ if (SU->FlaggedNodes.size())
+ return NULL;
+
+ SDNode *N = SU->Node;
+ if (!N)
+ return NULL;
+
+ SUnit *NewSU;
+ bool TryUnfold = false;
+ for (unsigned i = 0, e = N->getNumValues(); i != e; ++i) {
+ MVT::ValueType VT = N->getValueType(i);
+ if (VT == MVT::Flag)
+ return NULL;
+ else if (VT == MVT::Other)
+ TryUnfold = true;
+ }
for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
const SDOperand &Op = N->getOperand(i);
MVT::ValueType VT = Op.Val->getValueType(Op.ResNo);
- if (VT == MVT::Other || VT == MVT::Flag)
- return false;
+ if (VT == MVT::Flag)
+ return NULL;
}
- return true;
-}
+ if (TryUnfold) {
+ SmallVector<SDNode*, 4> NewNodes;
+ if (!TII->unfoldMemoryOperand(DAG, N, NewNodes))
+ return NULL;
+
+ DOUT << "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->Node->getNumValues();
+ for (unsigned i = 0; i != NumVals; ++i)
+ DAG.ReplaceAllUsesOfValueWith(SDOperand(SU->Node, i), SDOperand(N, i));
+ DAG.ReplaceAllUsesOfValueWith(SDOperand(SU->Node, OldNumVals-1),
+ SDOperand(LoadNode, 1));
+
+ SUnit *NewSU = NewSUnit(N);
+ SUnitMap[N].push_back(NewSU);
+ const TargetInstrDesc &TID = TII->get(N->getTargetOpcode());
+ for (unsigned i = 0; i != TID.getNumOperands(); ++i) {
+ if (TID.getOperandConstraint(i, TOI::TIED_TO) != -1) {
+ NewSU->isTwoAddress = true;
+ break;
+ }
+ }
+ if (TID.isCommutable())
+ NewSU->isCommutable = true;
+ // FIXME: Calculate height / depth and propagate the changes?
+ NewSU->Depth = SU->Depth;
+ NewSU->Height = SU->Height;
+ ComputeLatency(NewSU);
+
+ // 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;
+ DenseMap<SDNode*, std::vector<SUnit*> >::iterator SMI =
+ SUnitMap.find(LoadNode);
+ if (SMI != SUnitMap.end()) {
+ LoadSU = SMI->second.front();
+ isNewLoad = false;
+ } else {
+ LoadSU = NewSUnit(LoadNode);
+ SUnitMap[LoadNode].push_back(LoadSU);
-/// CopyAndMoveSuccessors - Clone the specified node and move its scheduled
-/// successors to the newly created node.
-SUnit *ScheduleDAGRRList::CopyAndMoveSuccessors(SUnit *SU) {
- SUnit *NewSU = Clone(SU);
+ LoadSU->Depth = SU->Depth;
+ LoadSU->Height = SU->Height;
+ ComputeLatency(LoadSU);
+ }
+
+ SUnit *ChainPred = NULL;
+ SmallVector<SDep, 4> ChainSuccs;
+ SmallVector<SDep, 4> LoadPreds;
+ SmallVector<SDep, 4> NodePreds;
+ SmallVector<SDep, 4> NodeSuccs;
+ for (SUnit::pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
+ I != E; ++I) {
+ if (I->isCtrl)
+ ChainPred = I->Dep;
+ else if (I->Dep->Node && I->Dep->Node->isOperand(LoadNode))
+ LoadPreds.push_back(SDep(I->Dep, I->Reg, I->Cost, false, false));
+ else
+ NodePreds.push_back(SDep(I->Dep, I->Reg, I->Cost, false, false));
+ }
+ for (SUnit::succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
+ I != E; ++I) {
+ if (I->isCtrl)
+ ChainSuccs.push_back(SDep(I->Dep, I->Reg, I->Cost,
+ I->isCtrl, I->isSpecial));
+ else
+ NodeSuccs.push_back(SDep(I->Dep, I->Reg, I->Cost,
+ I->isCtrl, I->isSpecial));
+ }
+
+ SU->removePred(ChainPred, true, false);
+ if (isNewLoad)
+ LoadSU->addPred(ChainPred, true, false);
+ for (unsigned i = 0, e = LoadPreds.size(); i != e; ++i) {
+ SDep *Pred = &LoadPreds[i];
+ SU->removePred(Pred->Dep, Pred->isCtrl, Pred->isSpecial);
+ if (isNewLoad)
+ LoadSU->addPred(Pred->Dep, Pred->isCtrl, Pred->isSpecial,
+ Pred->Reg, Pred->Cost);
+ }
+ for (unsigned i = 0, e = NodePreds.size(); i != e; ++i) {
+ SDep *Pred = &NodePreds[i];
+ SU->removePred(Pred->Dep, Pred->isCtrl, Pred->isSpecial);
+ NewSU->addPred(Pred->Dep, Pred->isCtrl, Pred->isSpecial,
+ Pred->Reg, Pred->Cost);
+ }
+ for (unsigned i = 0, e = NodeSuccs.size(); i != e; ++i) {
+ SDep *Succ = &NodeSuccs[i];
+ Succ->Dep->removePred(SU, Succ->isCtrl, Succ->isSpecial);
+ Succ->Dep->addPred(NewSU, Succ->isCtrl, Succ->isSpecial,
+ Succ->Reg, Succ->Cost);
+ }
+ for (unsigned i = 0, e = ChainSuccs.size(); i != e; ++i) {
+ SDep *Succ = &ChainSuccs[i];
+ Succ->Dep->removePred(SU, Succ->isCtrl, Succ->isSpecial);
+ if (isNewLoad)
+ Succ->Dep->addPred(LoadSU, Succ->isCtrl, Succ->isSpecial,
+ Succ->Reg, Succ->Cost);
+ }
+ if (isNewLoad)
+ NewSU->addPred(LoadSU, false, false);
+
+ if (isNewLoad)
+ AvailableQueue->addNode(LoadSU);
+ AvailableQueue->addNode(NewSU);
+
+ ++NumUnfolds;
+
+ if (NewSU->NumSuccsLeft == 0) {
+ NewSU->isAvailable = true;
+ return NewSU;
+ }
+ SU = NewSU;
+ }
+
+ DOUT << "Duplicating SU # " << SU->NodeNum << "\n";
+ NewSU = Clone(SU);
// New SUnit has the exact same predecessors.
for (SUnit::pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
// Only copy scheduled successors. Cut them from old node's successor
// list and move them over.
- SmallVector<SDep*, 2> DelDeps;
+ SmallVector<std::pair<SUnit*, bool>, 4> DelDeps;
for (SUnit::succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
I != E; ++I) {
if (I->isSpecial)
continue;
- NewSU->Height = std::max(NewSU->Height, I->Dep->Height+1);
if (I->Dep->isScheduled) {
+ NewSU->Height = std::max(NewSU->Height, I->Dep->Height+1);
I->Dep->addPred(NewSU, I->isCtrl, false, I->Reg, I->Cost);
- DelDeps.push_back(I);
+ DelDeps.push_back(std::make_pair(I->Dep, I->isCtrl));
}
}
for (unsigned i = 0, e = DelDeps.size(); i != e; ++i) {
- SUnit *Succ = DelDeps[i]->Dep;
- bool isCtrl = DelDeps[i]->isCtrl;
+ SUnit *Succ = DelDeps[i].first;
+ bool isCtrl = DelDeps[i].second;
Succ->removePred(SU, isCtrl, false);
}
AvailableQueue->updateNode(SU);
AvailableQueue->addNode(NewSU);
+ ++NumDups;
return NewSU;
}
+/// InsertCCCopiesAndMoveSuccs - Insert expensive cross register class copies
+/// and move all scheduled successors of the given SUnit to the last copy.
+void ScheduleDAGRRList::InsertCCCopiesAndMoveSuccs(SUnit *SU, unsigned Reg,
+ const TargetRegisterClass *DestRC,
+ const TargetRegisterClass *SrcRC,
+ SmallVector<SUnit*, 2> &Copies) {
+ SUnit *CopyFromSU = NewSUnit(NULL);
+ CopyFromSU->CopySrcRC = SrcRC;
+ CopyFromSU->CopyDstRC = DestRC;
+ CopyFromSU->Depth = SU->Depth;
+ CopyFromSU->Height = SU->Height;
+
+ SUnit *CopyToSU = NewSUnit(NULL);
+ CopyToSU->CopySrcRC = DestRC;
+ CopyToSU->CopyDstRC = SrcRC;
+
+ // Only copy scheduled successors. Cut them from old node's successor
+ // list and move them over.
+ SmallVector<std::pair<SUnit*, bool>, 4> DelDeps;
+ for (SUnit::succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
+ I != E; ++I) {
+ if (I->isSpecial)
+ continue;
+ if (I->Dep->isScheduled) {
+ CopyToSU->Height = std::max(CopyToSU->Height, I->Dep->Height+1);
+ I->Dep->addPred(CopyToSU, I->isCtrl, false, I->Reg, I->Cost);
+ DelDeps.push_back(std::make_pair(I->Dep, I->isCtrl));
+ }
+ }
+ for (unsigned i = 0, e = DelDeps.size(); i != e; ++i) {
+ SUnit *Succ = DelDeps[i].first;
+ bool isCtrl = DelDeps[i].second;
+ Succ->removePred(SU, isCtrl, false);
+ }
+
+ CopyFromSU->addPred(SU, false, false, Reg, -1);
+ CopyToSU->addPred(CopyFromSU, false, false, Reg, 1);
+
+ AvailableQueue->updateNode(SU);
+ AvailableQueue->addNode(CopyFromSU);
+ AvailableQueue->addNode(CopyToSU);
+ Copies.push_back(CopyFromSU);
+ Copies.push_back(CopyToSU);
+
+ ++NumCCCopies;
+}
+
+/// getPhysicalRegisterVT - Returns the ValueType of the physical register
+/// definition of the specified node.
+/// FIXME: Move to SelectionDAG?
+static MVT::ValueType getPhysicalRegisterVT(SDNode *N, unsigned Reg,
+ const TargetInstrInfo *TII) {
+ const TargetInstrDesc &TID = TII->get(N->getTargetOpcode());
+ assert(TID.ImplicitDefs && "Physical reg def must be in implicit def list!");
+ unsigned NumRes = TID.getNumDefs();
+ for (const unsigned *ImpDef = TID.getImplicitDefs(); *ImpDef; ++ImpDef) {
+ if (Reg == *ImpDef)
+ break;
+ ++NumRes;
+ }
+ return N->getValueType(NumRes);
+}
+
/// DelayForLiveRegsBottomUp - Returns true if it is necessary to delay
/// scheduling of the given node to satisfy live physical register dependencies.
/// If the specific node is the last one that's available to schedule, do
/// whatever is necessary (i.e. backtracking or cloning) to make it possible.
-bool ScheduleDAGRRList::DelayForLiveRegsBottomUp(SUnit *SU, unsigned &CurCycle){
+bool ScheduleDAGRRList::DelayForLiveRegsBottomUp(SUnit *SU,
+ SmallVector<unsigned, 4> &LRegs){
if (LiveRegs.empty())
return false;
+ SmallSet<unsigned, 4> RegAdded;
// If this node would clobber any "live" register, then it's not ready.
- // However, if this is the last "available" node, then we may have to
- // backtrack.
- bool MustSched = AvailableQueue->empty();
- SmallVector<unsigned, 4> LRegs;
for (SUnit::pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
I != E; ++I) {
if (I->Cost < 0) {
unsigned Reg = I->Reg;
- if (LiveRegs.count(Reg) && LiveRegDefs[Reg] != I->Dep)
- LRegs.push_back(Reg);
- for (const unsigned *Alias = MRI->getAliasSet(Reg);
+ if (LiveRegs.count(Reg) && LiveRegDefs[Reg] != I->Dep) {
+ if (RegAdded.insert(Reg))
+ LRegs.push_back(Reg);
+ }
+ for (const unsigned *Alias = TRI->getAliasSet(Reg);
*Alias; ++Alias)
- if (LiveRegs.count(*Alias) && LiveRegDefs[*Alias] != I->Dep)
- LRegs.push_back(*Alias);
+ if (LiveRegs.count(*Alias) && LiveRegDefs[*Alias] != I->Dep) {
+ if (RegAdded.insert(*Alias))
+ LRegs.push_back(*Alias);
+ }
}
}
for (unsigned i = 0, e = SU->FlaggedNodes.size()+1; i != e; ++i) {
SDNode *Node = (i == 0) ? SU->Node : SU->FlaggedNodes[i-1];
- if (!Node->isTargetOpcode())
+ if (!Node || !Node->isTargetOpcode())
continue;
- const TargetInstrDescriptor &TID = TII->get(Node->getTargetOpcode());
+ const TargetInstrDesc &TID = TII->get(Node->getTargetOpcode());
if (!TID.ImplicitDefs)
continue;
for (const unsigned *Reg = TID.ImplicitDefs; *Reg; ++Reg) {
- if (LiveRegs.count(*Reg) && LiveRegDefs[*Reg] != SU)
- LRegs.push_back(*Reg);
- for (const unsigned *Alias = MRI->getAliasSet(*Reg);
- *Alias; ++Alias)
- if (LiveRegs.count(*Alias) && LiveRegDefs[*Alias] != SU)
- LRegs.push_back(*Alias);
- }
- }
-
- if (MustSched && !LRegs.empty()) {
- // We have made a mistake by scheduling some nodes too early. Now we must
- // schedule the current node which will end up clobbering some live
- // registers that are expensive / impossible to copy. Try unscheduling
- // up to the point where it's safe to schedule the current node.
- unsigned LiveCycle = CurCycle;
- for (unsigned i = 0, e = LRegs.size(); i != e; ++i) {
- unsigned Reg = LRegs[i];
- unsigned LCycle = LiveRegCycles[Reg];
- LiveCycle = std::min(LiveCycle, LCycle);
- }
-
- if (SU->CycleBound < LiveCycle) {
- bool SuccUnsched = false;
- SUnit *OldSU = BackTrackBottomUp(SU, LiveCycle, CurCycle, SuccUnsched);
- // Force the current node to be scheduled before the node that
- // requires the physical reg dep.
- if (OldSU->isAvailable) {
- OldSU->isAvailable = false;
- AvailableQueue->remove(OldSU);
+ if (LiveRegs.count(*Reg) && LiveRegDefs[*Reg] != SU) {
+ if (RegAdded.insert(*Reg))
+ LRegs.push_back(*Reg);
}
- SU->addPred(OldSU, true, true);
- // If a successor has been unscheduled, then it's not possible to
- // schedule the current node.
- return SuccUnsched;
- } else {
- // Try duplicating the nodes that produces these "expensive to copy"
- // values to break the dependency.
- for (unsigned i = 0, e = LRegs.size(); i != e; ++i) {
- unsigned Reg = LRegs[i];
- SUnit *LRDef = LiveRegDefs[Reg];
- if (isSafeToCopy(LRDef->Node)) {
- SUnit *NewDef = CopyAndMoveSuccessors(LRDef);
- LiveRegDefs[Reg] = NewDef;
- NewDef->addPred(SU, true, true);
- SU->isAvailable = false;
- AvailableQueue->push(NewDef);
- } else {
- assert(false && "Expensive copying is required?");
- abort();
+ for (const unsigned *Alias = TRI->getAliasSet(*Reg);
+ *Alias; ++Alias)
+ if (LiveRegs.count(*Alias) && LiveRegDefs[*Alias] != SU) {
+ if (RegAdded.insert(*Alias))
+ LRegs.push_back(*Alias);
}
- }
- return true;
}
}
return !LRegs.empty();
}
+
/// ListScheduleBottomUp - The main loop of list scheduling for bottom-up
/// schedulers.
void ScheduleDAGRRList::ListScheduleBottomUp() {
// priority. If it is not ready put it back. Schedule the node.
SmallVector<SUnit*, 4> NotReady;
while (!AvailableQueue->empty()) {
+ bool Delayed = false;
+ DenseMap<SUnit*, SmallVector<unsigned, 4> > LRegsMap;
SUnit *CurSU = AvailableQueue->pop();
while (CurSU) {
- if (CurSU->CycleBound <= CurCycle)
- if (!DelayForLiveRegsBottomUp(CurSU, CurCycle))
+ if (CurSU->CycleBound <= CurCycle) {
+ SmallVector<unsigned, 4> LRegs;
+ if (!DelayForLiveRegsBottomUp(CurSU, LRegs))
break;
-
- // Verify node is still ready. It may not be in case the
- // scheduler has backtracked.
- if (CurSU->isAvailable) {
- CurSU->isPending = true;
- NotReady.push_back(CurSU);
+ 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 backtracking, code duplication, or inserting cross class copies
+ // to resolve it.
+ if (Delayed && !CurSU) {
+ for (unsigned i = 0, e = NotReady.size(); i != e; ++i) {
+ SUnit *TrySU = NotReady[i];
+ SmallVector<unsigned, 4> &LRegs = LRegsMap[TrySU];
+
+ // Try unscheduling up to the point where it's safe to schedule
+ // this node.
+ unsigned LiveCycle = CurCycle;
+ for (unsigned j = 0, ee = LRegs.size(); j != ee; ++j) {
+ unsigned Reg = LRegs[j];
+ unsigned LCycle = LiveRegCycles[Reg];
+ LiveCycle = std::min(LiveCycle, LCycle);
+ }
+ SUnit *OldSU = Sequence[LiveCycle];
+ if (!WillCreateCycle(TrySU, OldSU)) {
+ BacktrackBottomUp(TrySU, LiveCycle, CurCycle);
+ // Force the current node to be scheduled before the node that
+ // requires the physical reg dep.
+ if (OldSU->isAvailable) {
+ OldSU->isAvailable = false;
+ AvailableQueue->remove(OldSU);
+ }
+ TrySU->addPred(OldSU, true, true);
+ // If one or more successors has been unscheduled, then the current
+ // node is no longer avaialable. Schedule a successor that's now
+ // available instead.
+ if (!TrySU->isAvailable)
+ CurSU = AvailableQueue->pop();
+ else {
+ CurSU = TrySU;
+ TrySU->isPending = false;
+ NotReady.erase(NotReady.begin()+i);
+ }
+ break;
+ }
+ }
+
+ if (!CurSU) {
+ // Can't backtrace. 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];
+ SmallVector<unsigned, 4> &LRegs = LRegsMap[TrySU];
+ assert(LRegs.size() == 1 && "Can't handle this yet!");
+ unsigned Reg = LRegs[0];
+ SUnit *LRDef = LiveRegDefs[Reg];
+ SUnit *NewDef = CopyAndMoveSuccessors(LRDef);
+ if (!NewDef) {
+ // Issue expensive cross register class copies.
+ MVT::ValueType VT = getPhysicalRegisterVT(LRDef->Node, Reg, TII);
+ const TargetRegisterClass *RC =
+ TRI->getPhysicalRegisterRegClass(VT, Reg);
+ const TargetRegisterClass *DestRC = TRI->getCrossCopyRegClass(RC);
+ if (!DestRC) {
+ assert(false && "Don't know how to copy this physical register!");
+ abort();
+ }
+ SmallVector<SUnit*, 2> Copies;
+ InsertCCCopiesAndMoveSuccs(LRDef, Reg, DestRC, RC, Copies);
+ DOUT << "Adding an edge from SU # " << TrySU->NodeNum
+ << " to SU #" << Copies.front()->NodeNum << "\n";
+ TrySU->addPred(Copies.front(), true, true);
+ NewDef = Copies.back();
+ }
+
+ DOUT << "Adding an edge from SU # " << NewDef->NodeNum
+ << " to SU #" << TrySU->NodeNum << "\n";
+ LiveRegDefs[Reg] = NewDef;
+ NewDef->addPred(TrySU, true, true);
+ TrySU->isAvailable = false;
+ CurSU = NewDef;
+ }
+
+ if (!CurSU) {
+ assert(false && "Unable to resolve live physical register dependencies!");
+ abort();
+ }
+ }
+
// 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]);
}
// Verify that all SUnits were scheduled.
bool AnyNotSched = false;
for (unsigned i = 0, e = SUnits.size(); i != e; ++i) {
- if (SUnits[i].NumSuccsLeft != 0 || SUnits[i].NumChainSuccsLeft != 0) {
+ if (SUnits[i].NumSuccsLeft != 0) {
if (!AnyNotSched)
cerr << "*** List scheduling failed! ***\n";
SUnits[i].dump(&DAG);
// interrupt model (drain vs. freeze).
SuccSU->CycleBound = std::max(SuccSU->CycleBound, CurCycle + SuccSU->Latency);
- if (!isChain)
- --SuccSU->NumPredsLeft;
- else
- --SuccSU->NumChainPredsLeft;
+ --SuccSU->NumPredsLeft;
#ifndef NDEBUG
- if (SuccSU->NumPredsLeft < 0 || SuccSU->NumChainPredsLeft < 0) {
+ if (SuccSU->NumPredsLeft < 0) {
cerr << "*** List scheduling failed! ***\n";
SuccSU->dump(&DAG);
cerr << " has been released too many times!\n";
}
#endif
- if ((SuccSU->NumPredsLeft + SuccSU->NumChainPredsLeft) == 0) {
+ if (SuccSU->NumPredsLeft == 0) {
SuccSU->isAvailable = true;
AvailableQueue->push(SuccSU);
}
// All leaves to Available queue.
for (unsigned i = 0, e = SUnits.size(); i != e; ++i) {
// It is available if it has no predecessors.
- if (SUnits[i].Preds.size() == 0 && &SUnits[i] != Entry) {
+ if (SUnits[i].Preds.empty() && &SUnits[i] != Entry) {
AvailableQueue->push(&SUnits[i]);
SUnits[i].isAvailable = true;
}
static inline bool isCopyFromLiveIn(const SUnit *SU) {
SDNode *N = SU->Node;
- return N->getOpcode() == ISD::CopyFromReg &&
+ return N && N->getOpcode() == ISD::CopyFromReg &&
N->getOperand(N->getNumOperands()-1).getValueType() != MVT::Flag;
}
std::vector<unsigned> SethiUllmanNumbers;
const TargetInstrInfo *TII;
+ const TargetRegisterInfo *TRI;
public:
- explicit BURegReductionPriorityQueue(const TargetInstrInfo *tii)
- : TII(tii) {}
+ explicit BURegReductionPriorityQueue(const TargetInstrInfo *tii,
+ const TargetRegisterInfo *tri)
+ : TII(tii), TRI(tri) {}
void initNodes(DenseMap<SDNode*, std::vector<SUnit*> > &sumap,
std::vector<SUnit> &sunits) {
unsigned getNodePriority(const SUnit *SU) const {
assert(SU->NodeNum < SethiUllmanNumbers.size());
- unsigned Opc = SU->Node->getOpcode();
+ unsigned Opc = SU->Node ? SU->Node->getOpcode() : 0;
if (Opc == ISD::CopyFromReg && !isCopyFromLiveIn(SU))
// CopyFromReg should be close to its def because it restricts
// allocation choices. But if it is a livein then perhaps we want it
// CopyToReg should be close to its uses to facilitate coalescing and
// avoid spilling.
return 0;
+ else if (Opc == TargetInstrInfo::EXTRACT_SUBREG ||
+ Opc == TargetInstrInfo::INSERT_SUBREG)
+ // EXTRACT_SUBREG / INSERT_SUBREG should be close to its use to
+ // facilitate coalescing.
+ return 0;
else if (SU->NumSuccs == 0)
// If SU does not have a use, i.e. it doesn't produce a value that would
// be consumed (e.g. store), then it terminates a chain of computation.
unsigned Cycle = I->Dep->Cycle;
// If there are bunch of CopyToRegs stacked up, they should be considered
// to be at the same position.
- if (I->Dep->Node->getOpcode() == ISD::CopyToReg)
+ if (I->Dep->Node && I->Dep->Node->getOpcode() == ISD::CopyToReg)
Cycle = closestSucc(I->Dep)+1;
if (Cycle > MaxCycle)
MaxCycle = Cycle;
for (SUnit::const_pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
I != E; ++I) {
if (I->isCtrl) continue; // ignore chain preds
- if (I->Dep->Node->getOpcode() != ISD::CopyFromReg)
+ if (!I->Dep->Node || I->Dep->Node->getOpcode() != ISD::CopyFromReg)
Scratches++;
}
for (SUnit::const_succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
I != E; ++I) {
if (I->isCtrl) continue; // ignore chain succs
- if (I->Dep->Node->getOpcode() != ISD::CopyToReg)
+ if (!I->Dep->Node || I->Dep->Node->getOpcode() != ISD::CopyToReg)
Scratches += 10;
}
return Scratches;
return false;
}
-// FIXME: This is probably too slow!
-static void isReachable(SUnit *SU, SUnit *TargetSU,
- SmallPtrSet<SUnit*, 32> &Visited, bool &Reached) {
- if (Reached) return;
- if (SU == TargetSU) {
- Reached = true;
- return;
- }
- if (!Visited.insert(SU)) return;
-
- for (SUnit::pred_iterator I = SU->Preds.begin(), E = SU->Preds.end(); I != E;
- ++I)
- isReachable(I->Dep, TargetSU, Visited, Reached);
-}
-
-static bool isReachable(SUnit *SU, SUnit *TargetSU) {
- SmallPtrSet<SUnit*, 32> Visited;
- bool Reached = false;
- isReachable(SU, TargetSU, Visited, Reached);
- return Reached;
-}
-
template<class SF>
bool BURegReductionPriorityQueue<SF>::canClobber(SUnit *SU, SUnit *Op) {
if (SU->isTwoAddress) {
unsigned Opc = SU->Node->getTargetOpcode();
- unsigned NumRes = TII->getNumDefs(Opc);
- unsigned NumOps = ScheduleDAG::CountOperands(SU->Node);
+ const TargetInstrDesc &TID = TII->get(Opc);
+ unsigned NumRes = TID.getNumDefs();
+ unsigned NumOps = TID.getNumOperands() - NumRes;
for (unsigned i = 0; i != NumOps; ++i) {
- if (TII->getOperandConstraint(Opc, i+NumRes, TOI::TIED_TO) != -1) {
+ if (TID.getOperandConstraint(i+NumRes, TOI::TIED_TO) != -1) {
SDNode *DU = SU->Node->getOperand(i).Val;
- if (Op == (*SUnitMap)[DU][SU->InstanceNo])
+ if ((*SUnitMap).find(DU) != (*SUnitMap).end() &&
+ Op == (*SUnitMap)[DU][SU->InstanceNo])
return true;
}
}
}
+/// hasCopyToRegUse - Return true if SU has a value successor that is a
+/// CopyToReg node.
+static bool hasCopyToRegUse(SUnit *SU) {
+ for (SUnit::const_succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
+ I != E; ++I) {
+ if (I->isCtrl) continue;
+ SUnit *SuccSU = I->Dep;
+ if (SuccSU->Node && SuccSU->Node->getOpcode() == ISD::CopyToReg)
+ return true;
+ }
+ return false;
+}
+
+/// canClobberPhysRegDefs - True if SU would clobber one of SuccSU's
+/// physical register def.
+static bool canClobberPhysRegDefs(SUnit *SuccSU, SUnit *SU,
+ const TargetInstrInfo *TII,
+ const TargetRegisterInfo *TRI) {
+ SDNode *N = SuccSU->Node;
+ unsigned NumDefs = TII->get(N->getTargetOpcode()).getNumDefs();
+ const unsigned *ImpDefs = TII->get(N->getTargetOpcode()).getImplicitDefs();
+ if (!ImpDefs)
+ return false;
+ const unsigned *SUImpDefs =
+ TII->get(SU->Node->getTargetOpcode()).getImplicitDefs();
+ if (!SUImpDefs)
+ return false;
+ for (unsigned i = NumDefs, e = N->getNumValues(); i != e; ++i) {
+ MVT::ValueType VT = N->getValueType(i);
+ if (VT == MVT::Flag || VT == MVT::Other)
+ continue;
+ unsigned Reg = ImpDefs[i - NumDefs];
+ for (;*SUImpDefs; ++SUImpDefs) {
+ unsigned SUReg = *SUImpDefs;
+ if (TRI->regsOverlap(Reg, SUReg))
+ return true;
+ }
+ }
+ return false;
+}
+
/// AddPseudoTwoAddrDeps - If two nodes share an operand and one of them uses
/// it as a def&use operand. Add a pseudo control edge from it to the other
/// node (if it won't create a cycle) so the two-address one will be scheduled
-/// first (lower in the schedule).
+/// first (lower in the schedule). If both nodes are two-address, favor the
+/// one that has a CopyToReg use (more likely to be a loop induction update).
+/// If both are two-address, but one is commutable while the other is not
+/// commutable, favor the one that's not commutable.
template<class SF>
void BURegReductionPriorityQueue<SF>::AddPseudoTwoAddrDeps() {
for (unsigned i = 0, e = SUnits->size(); i != e; ++i) {
continue;
SDNode *Node = SU->Node;
- if (!Node->isTargetOpcode())
+ if (!Node || !Node->isTargetOpcode() || SU->FlaggedNodes.size() > 0)
continue;
unsigned Opc = Node->getTargetOpcode();
- unsigned NumRes = TII->getNumDefs(Opc);
- unsigned NumOps = ScheduleDAG::CountOperands(Node);
+ const TargetInstrDesc &TID = TII->get(Opc);
+ unsigned NumRes = TID.getNumDefs();
+ unsigned NumOps = TID.getNumOperands() - NumRes;
for (unsigned j = 0; j != NumOps; ++j) {
- if (TII->getOperandConstraint(Opc, j+NumRes, TOI::TIED_TO) != -1) {
+ if (TID.getOperandConstraint(j+NumRes, TOI::TIED_TO) != -1) {
SDNode *DU = SU->Node->getOperand(j).Val;
+ if ((*SUnitMap).find(DU) == (*SUnitMap).end())
+ continue;
SUnit *DUSU = (*SUnitMap)[DU][SU->InstanceNo];
if (!DUSU) continue;
for (SUnit::succ_iterator I = DUSU->Succs.begin(),E = DUSU->Succs.end();
I != E; ++I) {
if (I->isCtrl) continue;
SUnit *SuccSU = I->Dep;
- // Don't constraint nodes with implicit defs. It can create cycles
- // plus it may increase register pressures.
- if (SuccSU == SU || SuccSU->hasImplicitDefs)
+ if (SuccSU == SU)
+ continue;
+ // Be conservative. Ignore if nodes aren't at roughly the same
+ // depth and height.
+ if (SuccSU->Height < SU->Height && (SU->Height - SuccSU->Height) > 1)
continue;
- // Be conservative. Ignore if nodes aren't at the same depth.
- if (SuccSU->Depth != SU->Depth)
+ if (!SuccSU->Node || !SuccSU->Node->isTargetOpcode())
+ continue;
+ // Don't constrain nodes with physical register defs if the
+ // predecessor can clobber them.
+ if (SuccSU->hasPhysRegDefs) {
+ if (canClobberPhysRegDefs(SuccSU, SU, TII, TRI))
+ continue;
+ }
+ // Don't constraint extract_subreg / insert_subreg these may be
+ // coalesced away. We don't them close to their uses.
+ unsigned SuccOpc = SuccSU->Node->getTargetOpcode();
+ if (SuccOpc == TargetInstrInfo::EXTRACT_SUBREG ||
+ SuccOpc == TargetInstrInfo::INSERT_SUBREG)
continue;
if ((!canClobber(SuccSU, DUSU) ||
+ (hasCopyToRegUse(SU) && !hasCopyToRegUse(SuccSU)) ||
(!SU->isCommutable && SuccSU->isCommutable)) &&
!isReachable(SuccSU, SU)) {
DOUT << "Adding an edge from SU # " << SU->NodeNum
bool td_ls_rr_sort::operator()(const SUnit *left, const SUnit *right) const {
unsigned LPriority = SPQ->getNodePriority(left);
unsigned RPriority = SPQ->getNodePriority(right);
- bool LIsTarget = left->Node->isTargetOpcode();
- bool RIsTarget = right->Node->isTargetOpcode();
+ bool LIsTarget = left->Node && left->Node->isTargetOpcode();
+ bool RIsTarget = right->Node && right->Node->isTargetOpcode();
bool LIsFloater = LIsTarget && left->NumPreds == 0;
bool RIsFloater = RIsTarget && right->NumPreds == 0;
unsigned LBonus = (SumOfUnscheduledPredsOfSuccs(left) == 1) ? 2 : 0;
if (SethiUllmanNumber != 0)
return SethiUllmanNumber;
- unsigned Opc = SU->Node->getOpcode();
+ unsigned Opc = SU->Node ? SU->Node->getOpcode() : 0;
if (Opc == ISD::TokenFactor || Opc == ISD::CopyToReg)
SethiUllmanNumber = 0xffff;
else if (SU->NumSuccsLeft == 0)
SelectionDAG *DAG,
MachineBasicBlock *BB) {
const TargetInstrInfo *TII = DAG->getTarget().getInstrInfo();
+ const TargetRegisterInfo *TRI = DAG->getTarget().getRegisterInfo();
return new ScheduleDAGRRList(*DAG, BB, DAG->getTarget(), true,
- new BURegReductionPriorityQueue<bu_ls_rr_sort>(TII));
+ new BURegReductionPriorityQueue<bu_ls_rr_sort>(TII, TRI));
}
llvm::ScheduleDAG* llvm::createTDRRListDAGScheduler(SelectionDAGISel *IS,