//
//===----------------------------------------------------------------------===//
-#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/CodeGen/SelectionDAGNodes.h"
+#include "llvm/Support/CommandLine.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/raw_ostream.h"
+#include "llvm/Target/TargetInstrInfo.h"
+#include "llvm/Target/TargetMachine.h"
+#include "llvm/Target/TargetRegisterInfo.h"
+#include "llvm/Target/TargetSubtargetInfo.h"
#include <climits>
using namespace llvm;
-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() {
-}
+#define DEBUG_TYPE "pre-RA-sched"
-ScheduleDAG::~ScheduleDAG() {}
+#ifndef NDEBUG
+static cl::opt<bool> StressSchedOpt(
+ "stress-sched", cl::Hidden, cl::init(false),
+ cl::desc("Stress test instruction scheduling"));
+#endif
-/// 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";
- }
-}
+void SchedulingPriorityQueue::anchor() { }
+ScheduleDAG::ScheduleDAG(MachineFunction &mf)
+ : TM(mf.getTarget()), TII(TM.getSubtargetImpl()->getInstrInfo()),
+ TRI(TM.getSubtargetImpl()->getRegisterInfo()), MF(mf),
+ MRI(mf.getRegInfo()), EntrySU(), ExitSU() {
+#ifndef NDEBUG
+ StressSched = StressSchedOpt;
+#endif
+}
-/// Run - perform scheduling.
-///
-void ScheduleDAG::Run(MachineBasicBlock *bb,
- MachineBasicBlock::iterator insertPos) {
- BB = bb;
- InsertPos = insertPos;
+ScheduleDAG::~ScheduleDAG() {}
+/// Clear the DAG state (e.g. between scheduling regions).
+void ScheduleDAG::clearDAG() {
SUnits.clear();
- Sequence.clear();
EntrySU = SUnit();
ExitSU = SUnit();
+}
- Schedule();
-
- DEBUG({
- errs() << "*** Final schedule ***\n";
- dumpSchedule();
- errs() << '\n';
- });
+/// getInstrDesc helper to handle SDNodes.
+const MCInstrDesc *ScheduleDAG::getNodeDesc(const SDNode *Node) const {
+ if (!Node || !Node->isMachineOpcode()) return nullptr;
+ 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) {
- // 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;
+bool SUnit::addPred(const SDep &D, bool Required) {
+ // If this node already has this dependence, don't add a redundant one.
+ for (SmallVectorImpl<SDep>::iterator I = Preds.begin(), E = Preds.end();
+ I != E; ++I) {
+ // Zero-latency weak edges may be added purely for heuristic ordering. Don't
+ // add them if another kind of edge already exists.
+ if (!Required && I->getSUnit() == D.getSUnit())
+ return false;
+ if (I->overlaps(D)) {
+ // Extend the latency if needed. Equivalent to removePred(I) + addPred(D).
+ if (I->getLatency() < D.getLatency()) {
+ SUnit *PredSU = I->getSUnit();
+ // Find the corresponding successor in N.
+ SDep ForwardD = *I;
+ ForwardD.setSUnit(this);
+ for (SmallVectorImpl<SDep>::iterator II = PredSU->Succs.begin(),
+ EE = PredSU->Succs.end(); II != EE; ++II) {
+ if (*II == ForwardD) {
+ II->setLatency(D.getLatency());
+ break;
+ }
+ }
+ I->setLatency(D.getLatency());
+ }
+ return false;
+ }
+ }
// Now add a corresponding succ to N.
SDep P = D;
P.setSUnit(this);
++N->NumSuccs;
}
if (!N->isScheduled) {
- assert(NumPredsLeft < UINT_MAX && "NumPredsLeft will overflow!");
- ++NumPredsLeft;
+ if (D.isWeak()) {
+ ++WeakPredsLeft;
+ }
+ else {
+ assert(NumPredsLeft < UINT_MAX && "NumPredsLeft will overflow!");
+ ++NumPredsLeft;
+ }
}
if (!isScheduled) {
- assert(N->NumSuccsLeft < UINT_MAX && "NumSuccsLeft will overflow!");
- ++N->NumSuccsLeft;
+ if (D.isWeak()) {
+ ++N->WeakSuccsLeft;
+ }
+ else {
+ assert(N->NumSuccsLeft < UINT_MAX && "NumSuccsLeft will overflow!");
+ ++N->NumSuccsLeft;
+ }
}
Preds.push_back(D);
N->Succs.push_back(P);
this->setDepthDirty();
N->setHeightDirty();
}
+ return true;
}
/// removePred - This removes the specified edge as a pred of the current
/// 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)
+ for (SmallVectorImpl<SDep>::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!");
+ SmallVectorImpl<SDep>::iterator Succ = std::find(N->Succs.begin(),
+ N->Succs.end(), P);
+ assert(Succ != N->Succs.end() && "Mismatching preds / succs lists!");
+ N->Succs.erase(Succ);
Preds.erase(I);
// Update the bookkeeping.
if (P.getKind() == SDep::Data) {
--N->NumSuccs;
}
if (!N->isScheduled) {
- assert(NumPredsLeft > 0 && "NumPredsLeft will underflow!");
- --NumPredsLeft;
+ if (D.isWeak())
+ --WeakPredsLeft;
+ else {
+ assert(NumPredsLeft > 0 && "NumPredsLeft will underflow!");
+ --NumPredsLeft;
+ }
}
if (!isScheduled) {
- assert(N->NumSuccsLeft > 0 && "NumSuccsLeft will underflow!");
- --N->NumSuccsLeft;
+ if (D.isWeak())
+ --N->WeakSuccsLeft;
+ else {
+ assert(N->NumSuccsLeft > 0 && "NumSuccsLeft will underflow!");
+ --N->NumSuccsLeft;
+ }
}
if (P.getLatency() != 0) {
this->setDepthDirty();
/// setDepthToAtLeast - Update this node's successors to reflect the
/// fact that this node's depth just increased.
///
-void SUnit::setDepthToAtLeast(unsigned NewDepth, bool IgnoreAntiDep) {
- if (NewDepth <= getDepth(IgnoreAntiDep))
+void SUnit::setDepthToAtLeast(unsigned NewDepth) {
+ if (NewDepth <= getDepth())
return;
setDepthDirty();
Depth = NewDepth;
/// setHeightToAtLeast - Update this node's predecessors to reflect the
/// fact that this node's height just increased.
///
-void SUnit::setHeightToAtLeast(unsigned NewHeight, bool IgnoreAntiDep) {
- if (NewHeight <= getHeight(IgnoreAntiDep))
+void SUnit::setHeightToAtLeast(unsigned NewHeight) {
+ if (NewHeight <= getHeight())
return;
setHeightDirty();
Height = NewHeight;
/// ComputeDepth - Calculate the maximal path from the node to the exit.
///
-void SUnit::ComputeDepth(bool IgnoreAntiDep) {
+void SUnit::ComputeDepth() {
SmallVector<SUnit*, 8> WorkList;
WorkList.push_back(this);
do {
unsigned MaxPredDepth = 0;
for (SUnit::const_pred_iterator I = Cur->Preds.begin(),
E = Cur->Preds.end(); I != E; ++I) {
- if (IgnoreAntiDep && (I->getKind() == SDep::Anti)) continue;
SUnit *PredSU = I->getSUnit();
if (PredSU->isDepthCurrent)
MaxPredDepth = std::max(MaxPredDepth,
/// ComputeHeight - Calculate the maximal path from the node to the entry.
///
-void SUnit::ComputeHeight(bool IgnoreAntiDep) {
+void SUnit::ComputeHeight() {
SmallVector<SUnit*, 8> WorkList;
WorkList.push_back(this);
do {
unsigned MaxSuccHeight = 0;
for (SUnit::const_succ_iterator I = Cur->Succs.begin(),
E = Cur->Succs.end(); I != E; ++I) {
- if (IgnoreAntiDep && (I->getKind() == SDep::Anti)) continue;
SUnit *SuccSU = I->getSUnit();
if (SuccSU->isHeightCurrent)
MaxSuccHeight = std::max(MaxSuccHeight,
} while (!WorkList.empty());
}
+void SUnit::biasCriticalPath() {
+ if (NumPreds < 2)
+ return;
+
+ SUnit::pred_iterator BestI = Preds.begin();
+ unsigned MaxDepth = BestI->getSUnit()->getDepth();
+ for (SUnit::pred_iterator I = std::next(BestI), E = Preds.end(); I != E;
+ ++I) {
+ if (I->getKind() == SDep::Data && I->getSUnit()->getDepth() > MaxDepth)
+ BestI = I;
+ }
+ if (BestI != Preds.begin())
+ std::swap(*Preds.begin(), *BestI);
+}
+
+#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
/// 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 {
- errs() << "SU(" << NodeNum << "): ";
+ dbgs() << "SU(" << NodeNum << "): ";
G->dumpNode(this);
}
void SUnit::dumpAll(const ScheduleDAG *G) const {
dump(G);
- errs() << " # preds left : " << NumPredsLeft << "\n";
- errs() << " # succs left : " << NumSuccsLeft << "\n";
- errs() << " Latency : " << Latency << "\n";
- errs() << " Depth : " << Depth << "\n";
- errs() << " Height : " << Height << "\n";
+ dbgs() << " # preds left : " << NumPredsLeft << "\n";
+ dbgs() << " # succs left : " << NumSuccsLeft << "\n";
+ if (WeakPredsLeft)
+ dbgs() << " # weak preds left : " << WeakPredsLeft << "\n";
+ if (WeakSuccsLeft)
+ dbgs() << " # weak succs left : " << WeakSuccsLeft << "\n";
+ dbgs() << " # rdefs left : " << NumRegDefsLeft << "\n";
+ dbgs() << " Latency : " << Latency << "\n";
+ dbgs() << " Depth : " << getDepth() << "\n";
+ dbgs() << " Height : " << getHeight() << "\n";
if (Preds.size() != 0) {
- errs() << " Predecessors:\n";
+ dbgs() << " Predecessors:\n";
for (SUnit::const_succ_iterator I = Preds.begin(), E = Preds.end();
I != E; ++I) {
- errs() << " ";
+ dbgs() << " ";
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;
+ case SDep::Data: dbgs() << "val "; break;
+ case SDep::Anti: dbgs() << "anti"; break;
+ case SDep::Output: dbgs() << "out "; break;
+ case SDep::Order: dbgs() << "ch "; break;
}
- errs() << "#";
- errs() << I->getSUnit() << " - SU(" << I->getSUnit()->NodeNum << ")";
+ dbgs() << "SU(" << I->getSUnit()->NodeNum << ")";
if (I->isArtificial())
- errs() << " *";
- errs() << ": Latency=" << I->getLatency();
- errs() << "\n";
+ dbgs() << " *";
+ dbgs() << ": Latency=" << I->getLatency();
+ if (I->isAssignedRegDep())
+ dbgs() << " Reg=" << PrintReg(I->getReg(), G->TRI);
+ dbgs() << "\n";
}
}
if (Succs.size() != 0) {
- errs() << " Successors:\n";
+ dbgs() << " Successors:\n";
for (SUnit::const_succ_iterator I = Succs.begin(), E = Succs.end();
I != E; ++I) {
- errs() << " ";
+ dbgs() << " ";
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;
+ case SDep::Data: dbgs() << "val "; break;
+ case SDep::Anti: dbgs() << "anti"; break;
+ case SDep::Output: dbgs() << "out "; break;
+ case SDep::Order: dbgs() << "ch "; break;
}
- errs() << "#";
- errs() << I->getSUnit() << " - SU(" << I->getSUnit()->NodeNum << ")";
+ dbgs() << "SU(" << I->getSUnit()->NodeNum << ")";
if (I->isArtificial())
- errs() << " *";
- errs() << ": Latency=" << I->getLatency();
- errs() << "\n";
+ dbgs() << " *";
+ dbgs() << ": Latency=" << I->getLatency();
+ if (I->isAssignedRegDep())
+ dbgs() << " Reg=" << PrintReg(I->getReg(), G->TRI);
+ dbgs() << "\n";
}
}
- errs() << "\n";
+ dbgs() << "\n";
}
+#endif
#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)
- errs() << "*** Scheduling failed! ***\n";
+ dbgs() << "*** Scheduling failed! ***\n";
SUnits[i].dump(this);
- errs() << "has not been scheduled!\n";
+ dbgs() << "has not been scheduled!\n";
AnyNotSched = true;
}
if (SUnits[i].isScheduled &&
(isBottomUp ? SUnits[i].getHeight() : SUnits[i].getDepth()) >
unsigned(INT_MAX)) {
if (!AnyNotSched)
- errs() << "*** Scheduling failed! ***\n";
+ dbgs() << "*** Scheduling failed! ***\n";
SUnits[i].dump(this);
- errs() << "has an unexpected "
+ dbgs() << "has an unexpected "
<< (isBottomUp ? "Height" : "Depth") << " value!\n";
AnyNotSched = true;
}
if (isBottomUp) {
if (SUnits[i].NumSuccsLeft != 0) {
if (!AnyNotSched)
- errs() << "*** Scheduling failed! ***\n";
+ dbgs() << "*** Scheduling failed! ***\n";
SUnits[i].dump(this);
- errs() << "has successors left!\n";
+ dbgs() << "has successors left!\n";
AnyNotSched = true;
}
} else {
if (SUnits[i].NumPredsLeft != 0) {
if (!AnyNotSched)
- errs() << "*** Scheduling failed! ***\n";
+ dbgs() << "*** Scheduling failed! ***\n";
SUnits[i].dump(this);
- errs() << "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
Node2Index.resize(DAGSize);
// Initialize the data structures.
+ if (ExitSU)
+ WorkList.push_back(ExitSU);
for (unsigned i = 0, e = DAGSize; i != e; ++i) {
SUnit *SU = &SUnits[i];
int NodeNum = SU->NodeNum;
// Collect leaf nodes.
WorkList.push_back(SU);
}
- }
+ }
int Id = DAGSize;
while (!WorkList.empty()) {
SUnit *SU = WorkList.back();
WorkList.pop_back();
- Allocate(SU->NodeNum, --Id);
+ if (SU->NodeNum < DAGSize)
+ Allocate(SU->NodeNum, --Id);
for (SUnit::const_pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
I != E; ++I) {
SUnit *SU = I->getSUnit();
- if (!--Node2Index[SU->NodeNum])
+ if (SU->NodeNum < DAGSize && !--Node2Index[SU->NodeNum])
// If all dependencies of the node are processed already,
// then the node can be computed now.
WorkList.push_back(SU);
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);
do {
WorkList.pop_back();
Visited.set(SU->NodeNum);
for (int I = SU->Succs.size()-1; I >= 0; --I) {
- int s = SU->Succs[I].getSUnit()->NodeNum;
+ unsigned s = SU->Succs[I].getSUnit()->NodeNum;
+ // Edges to non-SUnits are allowed but ignored (e.g. ExitSU).
+ if (s >= Node2Index.size())
+ continue;
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;
}
-/// WillCreateCycle - Returns true if adding an edge from SU to TargetSU will
-/// create a cycle.
-bool ScheduleDAGTopologicalSort::WillCreateCycle(SUnit *SU, SUnit *TargetSU) {
- if (IsReachable(TargetSU, SU))
+/// WillCreateCycle - Returns true if adding an edge to TargetSU from SU will
+/// create a cycle. If so, it is not safe to call AddPred(TargetSU, SU).
+bool ScheduleDAGTopologicalSort::WillCreateCycle(SUnit *TargetSU, SUnit *SU) {
+ // Is SU reachable from TargetSU via successor edges?
+ if (IsReachable(SU, TargetSU))
return true;
- for (SUnit::pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
- I != E; ++I)
+ for (SUnit::pred_iterator
+ I = TargetSU->Preds.begin(), E = TargetSU->Preds.end(); I != E; ++I)
if (I->isAssignedRegDep() &&
- IsReachable(TargetSU, I->getSUnit()))
+ IsReachable(SU, I->getSUnit()))
return true;
return false;
}
// 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, SUnit *exitsu)
+ : SUnits(sunits), ExitSU(exitsu) {}
ScheduleHazardRecognizer::~ScheduleHazardRecognizer() {}
projects / oota-llvm.git / blobdiff