#define DEBUG_TYPE "pre-RA-sched"
#include "llvm/CodeGen/ScheduleDAG.h"
+#include "llvm/CodeGen/ScheduleHazardRecognizer.h"
+#include "llvm/CodeGen/SelectionDAGNodes.h"
#include "llvm/Target/TargetMachine.h"
#include "llvm/Target/TargetInstrInfo.h"
#include "llvm/Target/TargetRegisterInfo.h"
+#include "llvm/Support/CommandLine.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() {}
-
-/// 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);
-
- // 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);
- }
- }
-
- // Process nodes in the topological order
- while (!WorkList.empty()) {
- SUnit *SU = WorkList.back();
- WorkList.pop_back();
- unsigned SUDepth = 0;
-
- // 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->getSUnit()->Depth;
- if (PredDepth+1 > SUDepth) {
- SUDepth = PredDepth + 1;
- }
- }
+#ifndef NDEBUG
+static cl::opt<bool> StressSchedOpt(
+ "stress-sched", cl::Hidden, cl::init(false),
+ cl::desc("Stress test instruction scheduling"));
+#endif
- SU->Depth = SUDepth;
+void SchedulingPriorityQueue::anchor() { }
- // 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->getSUnit();
- 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);
- }
- }
+ScheduleDAG::ScheduleDAG(MachineFunction &mf)
+ : TM(mf.getTarget()),
+ TII(TM.getInstrInfo()),
+ TRI(TM.getRegisterInfo()),
+ MF(mf), MRI(mf.getRegInfo()),
+ EntrySU(), ExitSU() {
+#ifndef NDEBUG
+ StressSched = StressSchedOpt;
+#endif
}
-/// 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);
-
- // 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;
-
- // 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);
- }
- }
-
- // Process nodes in the topological order
- while (!WorkList.empty()) {
- SUnit *SU = WorkList.back();
- WorkList.pop_back();
- unsigned SUHeight = 0;
-
- // 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->getSUnit()->Height;
- if (SuccHeight+1 > SUHeight) {
- SUHeight = SuccHeight + 1;
- }
- }
-
- 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->getSUnit();
- 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);
- }
- }
-}
+ScheduleDAG::~ScheduleDAG() {}
-/// 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";
- }
+/// Clear the DAG state (e.g. between scheduling regions).
+void ScheduleDAG::clearDAG() {
+ SUnits.clear();
+ EntrySU = SUnit();
+ ExitSU = SUnit();
}
-
-/// Run - perform scheduling.
-///
-void ScheduleDAG::Run() {
- Schedule();
-
- DOUT << "*** Final schedule ***\n";
- DEBUG(dumpSchedule());
- DOUT << "\n";
+/// getInstrDesc helper to handle SDNodes.
+const MCInstrDesc *ScheduleDAG::getNodeDesc(const SDNode *Node) const {
+ if (!Node || !Node->isMachineOpcode()) return NULL;
+ return &TII->get(Node->getMachineOpcode());
}
/// 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) {
+bool SUnit::addPred(const SDep &D) {
// If this node already has this depenence, don't add a redundant one.
- for (unsigned i = 0, e = (unsigned)Preds.size(); i != e; ++i)
- if (Preds[i] == D)
- return;
- // Add a pred to this SUnit.
- Preds.push_back(D);
+ for (SmallVector<SDep, 4>::const_iterator I = Preds.begin(), E = Preds.end();
+ I != E; ++I)
+ if (*I == D)
+ return false;
// Now add a corresponding succ to N.
SDep P = D;
P.setSUnit(this);
SUnit *N = D.getSUnit();
- N->Succs.push_back(P);
// 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)
+ if (!N->isScheduled) {
+ assert(NumPredsLeft < UINT_MAX && "NumPredsLeft will overflow!");
++NumPredsLeft;
- if (!isScheduled)
+ }
+ 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();
+ }
+ return true;
}
/// removePred - This removes the specified edge as a pred of the current
break;
}
assert(FoundSucc && "Mismatching preds / succs lists!");
+ (void)FoundSucc;
Preds.erase(I);
- // Update the bookkeeping;
- if (D.getKind() == SDep::Data) {
+ // 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)
+ if (!N->isScheduled) {
+ assert(NumPredsLeft > 0 && "NumPredsLeft will underflow!");
--NumPredsLeft;
- if (!isScheduled)
+ }
+ if (!isScheduled) {
+ assert(N->NumSuccsLeft > 0 && "NumSuccsLeft will underflow!");
--N->NumSuccsLeft;
+ }
+ if (P.getLatency() != 0) {
+ this->setDepthDirty();
+ N->setHeightDirty();
+ }
return;
}
}
+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());
+}
+
+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());
+}
+
+/// 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;
+}
+
+/// 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);
+ }
+ }
+
+ if (Done) {
+ WorkList.pop_back();
+ if (MaxPredDepth != Cur->Depth) {
+ Cur->setDepthDirty();
+ Cur->Depth = MaxPredDepth;
+ }
+ Cur->isDepthCurrent = true;
+ }
+ } while (!WorkList.empty());
+}
+
+/// ComputeHeight - Calculate the maximal path from the node to the entry.
+///
+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 << "): ";
+ dbgs() << "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";
+ dbgs() << " # preds left : " << NumPredsLeft << "\n";
+ dbgs() << " # succs left : " << NumSuccsLeft << "\n";
+ dbgs() << " # rdefs left : " << NumRegDefsLeft << "\n";
+ dbgs() << " Latency : " << Latency << "\n";
+ dbgs() << " Depth : " << Depth << "\n";
+ dbgs() << " Height : " << Height << "\n";
if (Preds.size() != 0) {
- cerr << " Predecessors:\n";
+ dbgs() << " Predecessors:\n";
for (SUnit::const_succ_iterator I = Preds.begin(), E = Preds.end();
I != E; ++I) {
- cerr << " ";
+ dbgs() << " ";
switch (I->getKind()) {
- case SDep::Data: cerr << "val "; break;
- case SDep::Anti: cerr << "anti"; break;
- case SDep::Output: cerr << "out "; break;
- case SDep::Order: cerr << "ch "; break;
+ case SDep::Data: dbgs() << "val "; break;
+ case SDep::Anti: dbgs() << "anti"; break;
+ case SDep::Output: dbgs() << "out "; break;
+ case SDep::Order: dbgs() << "ch "; break;
}
- cerr << "#";
- cerr << I->getSUnit() << " - SU(" << I->getSUnit()->NodeNum << ")";
+ dbgs() << "SU(" << I->getSUnit()->NodeNum << ")";
if (I->isArtificial())
- cerr << " *";
- cerr << "\n";
+ dbgs() << " *";
+ dbgs() << ": Latency=" << I->getLatency();
+ if (I->isAssignedRegDep())
+ dbgs() << " Reg=" << PrintReg(I->getReg(), G->TRI);
+ dbgs() << "\n";
}
}
if (Succs.size() != 0) {
- cerr << " Successors:\n";
+ dbgs() << " Successors:\n";
for (SUnit::const_succ_iterator I = Succs.begin(), E = Succs.end();
I != E; ++I) {
- cerr << " ";
+ dbgs() << " ";
switch (I->getKind()) {
- case SDep::Data: cerr << "val "; break;
- case SDep::Anti: cerr << "anti"; break;
- case SDep::Output: cerr << "out "; break;
- case SDep::Order: cerr << "ch "; break;
+ case SDep::Data: dbgs() << "val "; break;
+ case SDep::Anti: dbgs() << "anti"; break;
+ case SDep::Output: dbgs() << "out "; break;
+ case SDep::Order: dbgs() << "ch "; break;
}
- cerr << "#";
- cerr << I->getSUnit() << " - SU(" << I->getSUnit()->NodeNum << ")";
+ dbgs() << "SU(" << I->getSUnit()->NodeNum << ")";
if (I->isArtificial())
- cerr << " *";
- cerr << "\n";
+ dbgs() << " *";
+ dbgs() << ": Latency=" << I->getLatency();
+ dbgs() << "\n";
}
}
- cerr << "\n";
+ dbgs() << "\n";
}
#ifndef NDEBUG
-/// VerifySchedule - Verify that all SUnits were scheduled and that
-/// their state is consistent.
+/// VerifyScheduledDAG - Verify that all SUnits were scheduled and that
+/// their state is consistent. Return the number of scheduled nodes.
///
-void ScheduleDAG::VerifySchedule(bool isBottomUp) {
+unsigned ScheduleDAG::VerifyScheduledDAG(bool isBottomUp) {
bool AnyNotSched = false;
unsigned DeadNodes = 0;
- unsigned Noops = 0;
for (unsigned i = 0, e = SUnits.size(); i != e; ++i) {
if (!SUnits[i].isScheduled) {
if (SUnits[i].NumPreds == 0 && SUnits[i].NumSuccs == 0) {
continue;
}
if (!AnyNotSched)
- cerr << "*** Scheduling failed! ***\n";
+ dbgs() << "*** Scheduling failed! ***\n";
SUnits[i].dump(this);
- cerr << "has not been scheduled!\n";
+ dbgs() << "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";
+ dbgs() << "*** Scheduling failed! ***\n";
SUnits[i].dump(this);
- cerr << "has an unexpected Cycle value!\n";
+ dbgs() << "has an unexpected "
+ << (isBottomUp ? "Height" : "Depth") << " value!\n";
AnyNotSched = true;
}
if (isBottomUp) {
if (SUnits[i].NumSuccsLeft != 0) {
if (!AnyNotSched)
- cerr << "*** Scheduling failed! ***\n";
+ dbgs() << "*** Scheduling failed! ***\n";
SUnits[i].dump(this);
- cerr << "has successors left!\n";
+ dbgs() << "has successors left!\n";
AnyNotSched = true;
}
} else {
if (SUnits[i].NumPredsLeft != 0) {
if (!AnyNotSched)
- cerr << "*** Scheduling failed! ***\n";
+ dbgs() << "*** Scheduling failed! ***\n";
SUnits[i].dump(this);
- cerr << "has predecessors left!\n";
+ dbgs() << "has predecessors left!\n";
AnyNotSched = true;
}
}
}
- for (unsigned i = 0, e = Sequence.size(); i != e; ++i)
- if (!Sequence[i])
- ++Noops;
assert(!AnyNotSched);
- assert(Sequence.size() + DeadNodes - Noops == SUnits.size() &&
- "The number of nodes scheduled doesn't match the expected number!");
+ return SUnits.size() - DeadNodes;
}
#endif
-/// InitDAGTopologicalSorting - create the initial topological
+/// InitDAGTopologicalSorting - create the initial topological
/// ordering from the DAG to be scheduled.
///
-/// The idea of the algorithm is taken from
+/// The idea of the algorithm is taken from
/// "Online algorithms for managing the topological order of
/// a directed acyclic graph" by David J. Pearce and Paul H.J. Kelly
-/// This is the MNR algorithm, which was first introduced by
-/// A. Marchetti-Spaccamela, U. Nanni and H. Rohnert in
+/// This is the MNR algorithm, which was first introduced by
+/// A. Marchetti-Spaccamela, U. Nanni and H. Rohnert in
/// "Maintaining a topological order under edge insertions".
///
-/// Short description of the algorithm:
+/// Short description of the algorithm:
///
/// Topological ordering, ord, of a DAG maps each node to a topological
/// index so that for all edges X->Y it is the case that ord(X) < ord(Y).
///
-/// This means that if there is a path from the node X to the node Z,
+/// This means that if there is a path from the node X to the node Z,
/// then ord(X) < ord(Z).
///
/// This property can be used to check for reachability of nodes:
-/// if Z is reachable from X, then an insertion of the edge Z->X would
+/// if Z is reachable from X, then an insertion of the edge Z->X would
/// create a cycle.
///
/// The algorithm first computes a topological ordering for the DAG by
// Collect leaf nodes.
WorkList.push_back(SU);
}
- }
+ }
int Id = DAGSize;
while (!WorkList.empty()) {
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->getSUnit()->NodeNum] &&
+ assert(Node2Index[SU->NodeNum] > Node2Index[I->getSUnit()->NodeNum] &&
"Wrong topological sorting");
}
}
#endif
}
-/// AddPred - Updates the topological ordering to accomodate an edge
+/// AddPred - Updates the topological ordering to accommodate an edge
/// to be added from SUnit X to SUnit Y.
void ScheduleDAGTopologicalSort::AddPred(SUnit *Y, SUnit *X) {
int UpperBound, LowerBound;
}
}
-/// RemovePred - Updates the topological ordering to accomodate an
+/// RemovePred - Updates the topological ordering to accommodate an
/// an edge to be removed from the specified node N from the predecessors
/// of the current node M.
void ScheduleDAGTopologicalSort::RemovePred(SUnit *M, SUnit *N) {
/// all nodes affected by the edge insertion. These nodes will later get new
/// topological indexes by means of the Shift method.
void ScheduleDAGTopologicalSort::DFS(const SUnit *SU, int UpperBound,
- bool& HasLoop) {
+ bool &HasLoop) {
std::vector<const SUnit*> WorkList;
- WorkList.reserve(SUnits.size());
+ 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].getSUnit()->NodeNum;
if (Node2Index[s] == UpperBound) {
- HasLoop = true;
+ 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].getSUnit());
- }
- }
- }
+ }
+ }
+ } while (!WorkList.empty());
}
-/// Shift - Renumber the nodes so that the topological ordering is
+/// Shift - Renumber the nodes so that the topological ordering is
/// preserved.
-void ScheduleDAGTopologicalSort::Shift(BitVector& Visited, int LowerBound,
+void ScheduleDAGTopologicalSort::Shift(BitVector& Visited, int LowerBound,
int UpperBound) {
std::vector<int> L;
int shift = 0;
// Is Ord(TargetSU) < Ord(SU) ?
if (LowerBound < UpperBound) {
Visited.reset();
- // There may be a path from TargetSU to SU. Check for it.
+ // There may be a path from TargetSU to SU. Check for it.
DFS(TargetSU, UpperBound, HasLoop);
}
return HasLoop;
Index2Node[index] = n;
}
-ScheduleDAGTopologicalSort::ScheduleDAGTopologicalSort(
- std::vector<SUnit> &sunits)
- : SUnits(sunits) {}
+ScheduleDAGTopologicalSort::
+ScheduleDAGTopologicalSort(std::vector<SUnit> &sunits) : SUnits(sunits) {}
+
+ScheduleHazardRecognizer::~ScheduleHazardRecognizer() {}