#define DEBUG_TYPE "pre-RA-sched"
#include "llvm/CodeGen/ScheduleDAG.h"
+#include "llvm/CodeGen/ScheduleHazardRecognizer.h"
#include "llvm/Target/TargetMachine.h"
#include "llvm/Target/TargetInstrInfo.h"
#include "llvm/Target/TargetRegisterInfo.h"
#include "llvm/Support/Debug.h"
+#include "llvm/Support/raw_ostream.h"
#include <climits>
using namespace llvm;
-ScheduleDAG::ScheduleDAG(SelectionDAG *dag, MachineBasicBlock *bb,
- const TargetMachine &tm)
- : DAG(dag), BB(bb), TM(tm), MRI(BB->getParent()->getRegInfo()) {
- TII = TM.getInstrInfo();
- MF = BB->getParent();
- TRI = TM.getRegisterInfo();
- TLI = TM.getTargetLowering();
- ConstPool = MF->getConstantPool();
+ScheduleDAG::ScheduleDAG(MachineFunction &mf)
+ : TM(mf.getTarget()),
+ TII(TM.getInstrInfo()),
+ TRI(TM.getRegisterInfo()),
+ TLI(TM.getTargetLowering()),
+ MF(mf), MRI(mf.getRegInfo()),
+ ConstPool(MF.getConstantPool()),
+ EntrySU(), ExitSU() {
}
ScheduleDAG::~ScheduleDAG() {}
-/// CalculateDepths - compute depths using algorithms for the longest
-/// paths in the DAG
-void ScheduleDAG::CalculateDepths() {
- unsigned DAGSize = SUnits.size();
- std::vector<SUnit*> WorkList;
- WorkList.reserve(DAGSize);
+/// dump - dump the schedule.
+void ScheduleDAG::dumpSchedule() const {
+ for (unsigned i = 0, e = Sequence.size(); i != e; i++) {
+ if (SUnit *SU = Sequence[i])
+ SU->dump(this);
+ else
+ errs() << "**** NOOP ****\n";
+ }
+}
- // Initialize the data structures
- for (unsigned i = 0, e = DAGSize; i != e; ++i) {
- SUnit *SU = &SUnits[i];
- unsigned Degree = SU->Preds.size();
- // Temporarily use the Depth field as scratch space for the degree count.
- SU->Depth = Degree;
- // Is it a node without dependencies?
- if (Degree == 0) {
- assert(SU->Preds.empty() && "SUnit should have no predecessors");
- // Collect leaf nodes
- WorkList.push_back(SU);
- }
- }
+/// Run - perform scheduling.
+///
+void ScheduleDAG::Run(MachineBasicBlock *bb,
+ MachineBasicBlock::iterator insertPos) {
+ BB = bb;
+ InsertPos = insertPos;
- // Process nodes in the topological order
- while (!WorkList.empty()) {
- SUnit *SU = WorkList.back();
- WorkList.pop_back();
- unsigned SUDepth = 0;
+ SUnits.clear();
+ Sequence.clear();
+ EntrySU = SUnit();
+ ExitSU = SUnit();
- // Use dynamic programming:
- // When current node is being processed, all of its dependencies
- // are already processed.
- // So, just iterate over all predecessors and take the longest path
- for (SUnit::const_pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
- I != E; ++I) {
- unsigned PredDepth = I->Dep->Depth;
- if (PredDepth+1 > SUDepth) {
- SUDepth = PredDepth + 1;
- }
- }
+ Schedule();
- SU->Depth = SUDepth;
+ DEBUG({
+ errs() << "*** Final schedule ***\n";
+ dumpSchedule();
+ errs() << '\n';
+ });
+}
- // Update degrees of all nodes depending on current SUnit
- for (SUnit::const_succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
- I != E; ++I) {
- SUnit *SU = I->Dep;
- if (!--SU->Depth)
- // If all dependencies of the node are processed already,
- // then the longest path for the node can be computed now
- WorkList.push_back(SU);
- }
+/// addPred - This adds the specified edge as a pred of the current node if
+/// not already. It also adds the current node as a successor of the
+/// specified node.
+void SUnit::addPred(const SDep &D) {
+ // If this node already has this depenence, don't add a redundant one.
+ for (SmallVector<SDep, 4>::const_iterator I = Preds.begin(), E = Preds.end();
+ I != E; ++I)
+ if (*I == D)
+ return;
+ // Now add a corresponding succ to N.
+ SDep P = D;
+ P.setSUnit(this);
+ SUnit *N = D.getSUnit();
+ // Update the bookkeeping.
+ if (D.getKind() == SDep::Data) {
+ assert(NumPreds < UINT_MAX && "NumPreds will overflow!");
+ assert(N->NumSuccs < UINT_MAX && "NumSuccs will overflow!");
+ ++NumPreds;
+ ++N->NumSuccs;
+ }
+ if (!N->isScheduled) {
+ assert(NumPredsLeft < UINT_MAX && "NumPredsLeft will overflow!");
+ ++NumPredsLeft;
+ }
+ if (!isScheduled) {
+ assert(N->NumSuccsLeft < UINT_MAX && "NumSuccsLeft will overflow!");
+ ++N->NumSuccsLeft;
+ }
+ Preds.push_back(D);
+ N->Succs.push_back(P);
+ if (P.getLatency() != 0) {
+ this->setDepthDirty();
+ N->setHeightDirty();
}
}
-/// CalculateHeights - compute heights using algorithms for the longest
-/// paths in the DAG
-void ScheduleDAG::CalculateHeights() {
- unsigned DAGSize = SUnits.size();
- std::vector<SUnit*> WorkList;
- WorkList.reserve(DAGSize);
+/// removePred - This removes the specified edge as a pred of the current
+/// node if it exists. It also removes the current node as a successor of
+/// the specified node.
+void SUnit::removePred(const SDep &D) {
+ // Find the matching predecessor.
+ for (SmallVector<SDep, 4>::iterator I = Preds.begin(), E = Preds.end();
+ I != E; ++I)
+ if (*I == D) {
+ bool FoundSucc = false;
+ // Find the corresponding successor in N.
+ SDep P = D;
+ P.setSUnit(this);
+ SUnit *N = D.getSUnit();
+ for (SmallVector<SDep, 4>::iterator II = N->Succs.begin(),
+ EE = N->Succs.end(); II != EE; ++II)
+ if (*II == P) {
+ FoundSucc = true;
+ N->Succs.erase(II);
+ break;
+ }
+ assert(FoundSucc && "Mismatching preds / succs lists!");
+ Preds.erase(I);
+ // Update the bookkeeping.
+ if (P.getKind() == SDep::Data) {
+ assert(NumPreds > 0 && "NumPreds will underflow!");
+ assert(N->NumSuccs > 0 && "NumSuccs will underflow!");
+ --NumPreds;
+ --N->NumSuccs;
+ }
+ if (!N->isScheduled) {
+ assert(NumPredsLeft > 0 && "NumPredsLeft will underflow!");
+ --NumPredsLeft;
+ }
+ if (!isScheduled) {
+ assert(N->NumSuccsLeft > 0 && "NumSuccsLeft will underflow!");
+ --N->NumSuccsLeft;
+ }
+ if (P.getLatency() != 0) {
+ this->setDepthDirty();
+ N->setHeightDirty();
+ }
+ return;
+ }
+}
- // Initialize the data structures
- for (unsigned i = 0, e = DAGSize; i != e; ++i) {
- SUnit *SU = &SUnits[i];
- unsigned Degree = SU->Succs.size();
- // Temporarily use the Height field as scratch space for the degree count.
- SU->Height = Degree;
+void SUnit::setDepthDirty() {
+ if (!isDepthCurrent) return;
+ SmallVector<SUnit*, 8> WorkList;
+ WorkList.push_back(this);
+ do {
+ SUnit *SU = WorkList.pop_back_val();
+ SU->isDepthCurrent = false;
+ for (SUnit::const_succ_iterator I = SU->Succs.begin(),
+ E = SU->Succs.end(); I != E; ++I) {
+ SUnit *SuccSU = I->getSUnit();
+ if (SuccSU->isDepthCurrent)
+ WorkList.push_back(SuccSU);
+ }
+ } while (!WorkList.empty());
+}
- // Is it a node without dependencies?
- if (Degree == 0) {
- assert(SU->Succs.empty() && "Something wrong");
- assert(WorkList.empty() && "Should be empty");
- // Collect leaf nodes
- WorkList.push_back(SU);
+void SUnit::setHeightDirty() {
+ if (!isHeightCurrent) return;
+ SmallVector<SUnit*, 8> WorkList;
+ WorkList.push_back(this);
+ do {
+ SUnit *SU = WorkList.pop_back_val();
+ SU->isHeightCurrent = false;
+ for (SUnit::const_pred_iterator I = SU->Preds.begin(),
+ E = SU->Preds.end(); I != E; ++I) {
+ SUnit *PredSU = I->getSUnit();
+ if (PredSU->isHeightCurrent)
+ WorkList.push_back(PredSU);
}
- }
+ } while (!WorkList.empty());
+}
- // Process nodes in the topological order
- while (!WorkList.empty()) {
- SUnit *SU = WorkList.back();
- WorkList.pop_back();
- unsigned SUHeight = 0;
+/// setDepthToAtLeast - Update this node's successors to reflect the
+/// fact that this node's depth just increased.
+///
+void SUnit::setDepthToAtLeast(unsigned NewDepth) {
+ if (NewDepth <= getDepth())
+ return;
+ setDepthDirty();
+ Depth = NewDepth;
+ isDepthCurrent = true;
+}
- // Use dynamic programming:
- // When current node is being processed, all of its dependencies
- // are already processed.
- // So, just iterate over all successors and take the longest path
- for (SUnit::const_succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
- I != E; ++I) {
- unsigned SuccHeight = I->Dep->Height;
- if (SuccHeight+1 > SUHeight) {
- SUHeight = SuccHeight + 1;
+/// setHeightToAtLeast - Update this node's predecessors to reflect the
+/// fact that this node's height just increased.
+///
+void SUnit::setHeightToAtLeast(unsigned NewHeight) {
+ if (NewHeight <= getHeight())
+ return;
+ setHeightDirty();
+ Height = NewHeight;
+ isHeightCurrent = true;
+}
+
+/// ComputeDepth - Calculate the maximal path from the node to the exit.
+///
+void SUnit::ComputeDepth() {
+ SmallVector<SUnit*, 8> WorkList;
+ WorkList.push_back(this);
+ do {
+ SUnit *Cur = WorkList.back();
+
+ bool Done = true;
+ unsigned MaxPredDepth = 0;
+ for (SUnit::const_pred_iterator I = Cur->Preds.begin(),
+ E = Cur->Preds.end(); I != E; ++I) {
+ SUnit *PredSU = I->getSUnit();
+ if (PredSU->isDepthCurrent)
+ MaxPredDepth = std::max(MaxPredDepth,
+ PredSU->Depth + I->getLatency());
+ else {
+ Done = false;
+ WorkList.push_back(PredSU);
}
}
- SU->Height = SUHeight;
-
- // Update degrees of all nodes depending on current SUnit
- for (SUnit::const_pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
- I != E; ++I) {
- SUnit *SU = I->Dep;
- if (!--SU->Height)
- // If all dependencies of the node are processed already,
- // then the longest path for the node can be computed now
- WorkList.push_back(SU);
+ if (Done) {
+ WorkList.pop_back();
+ if (MaxPredDepth != Cur->Depth) {
+ Cur->setDepthDirty();
+ Cur->Depth = MaxPredDepth;
+ }
+ Cur->isDepthCurrent = true;
}
- }
-}
-
-/// dump - dump the schedule.
-void ScheduleDAG::dumpSchedule() const {
- for (unsigned i = 0, e = Sequence.size(); i != e; i++) {
- if (SUnit *SU = Sequence[i])
- SU->dump(this);
- else
- cerr << "**** NOOP ****\n";
- }
+ } while (!WorkList.empty());
}
-
-/// Run - perform scheduling.
+/// ComputeHeight - Calculate the maximal path from the node to the entry.
///
-void ScheduleDAG::Run() {
- Schedule();
-
- DOUT << "*** Final schedule ***\n";
- DEBUG(dumpSchedule());
- DOUT << "\n";
+void SUnit::ComputeHeight() {
+ SmallVector<SUnit*, 8> WorkList;
+ WorkList.push_back(this);
+ do {
+ SUnit *Cur = WorkList.back();
+
+ bool Done = true;
+ unsigned MaxSuccHeight = 0;
+ for (SUnit::const_succ_iterator I = Cur->Succs.begin(),
+ E = Cur->Succs.end(); I != E; ++I) {
+ SUnit *SuccSU = I->getSUnit();
+ if (SuccSU->isHeightCurrent)
+ MaxSuccHeight = std::max(MaxSuccHeight,
+ SuccSU->Height + I->getLatency());
+ else {
+ Done = false;
+ WorkList.push_back(SuccSU);
+ }
+ }
+
+ if (Done) {
+ WorkList.pop_back();
+ if (MaxSuccHeight != Cur->Height) {
+ Cur->setHeightDirty();
+ Cur->Height = MaxSuccHeight;
+ }
+ Cur->isHeightCurrent = true;
+ }
+ } while (!WorkList.empty());
}
/// SUnit - Scheduling unit. It's an wrapper around either a single SDNode or
/// a group of nodes flagged together.
void SUnit::dump(const ScheduleDAG *G) const {
- cerr << "SU(" << NodeNum << "): ";
+ errs() << "SU(" << NodeNum << "): ";
G->dumpNode(this);
}
void SUnit::dumpAll(const ScheduleDAG *G) const {
dump(G);
- cerr << " # preds left : " << NumPredsLeft << "\n";
- cerr << " # succs left : " << NumSuccsLeft << "\n";
- cerr << " Latency : " << Latency << "\n";
- cerr << " Depth : " << Depth << "\n";
- cerr << " Height : " << Height << "\n";
+ errs() << " # preds left : " << NumPredsLeft << "\n";
+ errs() << " # succs left : " << NumSuccsLeft << "\n";
+ errs() << " Latency : " << Latency << "\n";
+ errs() << " Depth : " << Depth << "\n";
+ errs() << " Height : " << Height << "\n";
if (Preds.size() != 0) {
- cerr << " Predecessors:\n";
+ errs() << " Predecessors:\n";
for (SUnit::const_succ_iterator I = Preds.begin(), E = Preds.end();
I != E; ++I) {
- if (I->isCtrl)
- cerr << " ch #";
- else
- cerr << " val #";
- cerr << I->Dep << " - SU(" << I->Dep->NodeNum << ")";
- if (I->isArtificial)
- cerr << " *";
- cerr << "\n";
+ errs() << " ";
+ switch (I->getKind()) {
+ case SDep::Data: errs() << "val "; break;
+ case SDep::Anti: errs() << "anti"; break;
+ case SDep::Output: errs() << "out "; break;
+ case SDep::Order: errs() << "ch "; break;
+ }
+ errs() << "#";
+ errs() << I->getSUnit() << " - SU(" << I->getSUnit()->NodeNum << ")";
+ if (I->isArtificial())
+ errs() << " *";
+ errs() << ": Latency=" << I->getLatency();
+ errs() << "\n";
}
}
if (Succs.size() != 0) {
- cerr << " Successors:\n";
+ errs() << " Successors:\n";
for (SUnit::const_succ_iterator I = Succs.begin(), E = Succs.end();
I != E; ++I) {
- if (I->isCtrl)
- cerr << " ch #";
- else
- cerr << " val #";
- cerr << I->Dep << " - SU(" << I->Dep->NodeNum << ")";
- if (I->isArtificial)
- cerr << " *";
- cerr << "\n";
+ errs() << " ";
+ switch (I->getKind()) {
+ case SDep::Data: errs() << "val "; break;
+ case SDep::Anti: errs() << "anti"; break;
+ case SDep::Output: errs() << "out "; break;
+ case SDep::Order: errs() << "ch "; break;
+ }
+ errs() << "#";
+ errs() << I->getSUnit() << " - SU(" << I->getSUnit()->NodeNum << ")";
+ if (I->isArtificial())
+ errs() << " *";
+ errs() << ": Latency=" << I->getLatency();
+ errs() << "\n";
}
}
- cerr << "\n";
+ errs() << "\n";
}
#ifndef NDEBUG
continue;
}
if (!AnyNotSched)
- cerr << "*** Scheduling failed! ***\n";
+ errs() << "*** Scheduling failed! ***\n";
SUnits[i].dump(this);
- cerr << "has not been scheduled!\n";
+ errs() << "has not been scheduled!\n";
AnyNotSched = true;
}
- if (SUnits[i].isScheduled && SUnits[i].Cycle > (unsigned)INT_MAX) {
+ if (SUnits[i].isScheduled &&
+ (isBottomUp ? SUnits[i].getHeight() : SUnits[i].getDepth()) >
+ unsigned(INT_MAX)) {
if (!AnyNotSched)
- cerr << "*** Scheduling failed! ***\n";
+ errs() << "*** Scheduling failed! ***\n";
SUnits[i].dump(this);
- cerr << "has an unexpected Cycle value!\n";
+ errs() << "has an unexpected "
+ << (isBottomUp ? "Height" : "Depth") << " value!\n";
AnyNotSched = true;
}
if (isBottomUp) {
if (SUnits[i].NumSuccsLeft != 0) {
if (!AnyNotSched)
- cerr << "*** Scheduling failed! ***\n";
+ errs() << "*** Scheduling failed! ***\n";
SUnits[i].dump(this);
- cerr << "has successors left!\n";
+ errs() << "has successors left!\n";
AnyNotSched = true;
}
} else {
if (SUnits[i].NumPredsLeft != 0) {
if (!AnyNotSched)
- cerr << "*** Scheduling failed! ***\n";
+ errs() << "*** Scheduling failed! ***\n";
SUnits[i].dump(this);
- cerr << "has predecessors left!\n";
+ errs() << "has predecessors left!\n";
AnyNotSched = true;
}
}
Allocate(SU->NodeNum, --Id);
for (SUnit::const_pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
I != E; ++I) {
- SUnit *SU = I->Dep;
+ SUnit *SU = I->getSUnit();
if (!--Node2Index[SU->NodeNum])
// If all dependencies of the node are processed already,
// then the node can be computed now.
SUnit *SU = &SUnits[i];
for (SUnit::const_pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
I != E; ++I) {
- assert(Node2Index[SU->NodeNum] > Node2Index[I->Dep->NodeNum] &&
+ assert(Node2Index[SU->NodeNum] > Node2Index[I->getSUnit()->NodeNum] &&
"Wrong topological sorting");
}
}
WorkList.reserve(SUnits.size());
WorkList.push_back(SU);
- while (!WorkList.empty()) {
+ do {
SU = WorkList.back();
WorkList.pop_back();
Visited.set(SU->NodeNum);
for (int I = SU->Succs.size()-1; I >= 0; --I) {
- int s = SU->Succs[I].Dep->NodeNum;
+ int s = SU->Succs[I].getSUnit()->NodeNum;
if (Node2Index[s] == UpperBound) {
HasLoop = true;
return;
}
// Visit successors if not already and in affected region.
if (!Visited.test(s) && Node2Index[s] < UpperBound) {
- WorkList.push_back(SU->Succs[I].Dep);
+ WorkList.push_back(SU->Succs[I].getSUnit());
}
}
- }
+ } while (!WorkList.empty());
}
/// Shift - Renumber the nodes so that the topological ordering is
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))
+ if (I->isAssignedRegDep() &&
+ IsReachable(TargetSU, I->getSUnit()))
return true;
return false;
}
ScheduleDAGTopologicalSort::ScheduleDAGTopologicalSort(
std::vector<SUnit> &sunits)
: SUnits(sunits) {}
+
+ScheduleHazardRecognizer::~ScheduleHazardRecognizer() {}