X-Git-Url: http://plrg.eecs.uci.edu/git/?a=blobdiff_plain;f=lib%2FCodeGen%2FScheduleDAG.cpp;h=9a65071001709970097d3f04f8077bb38f59f3dd;hb=c2945e44d5d485ce080a09900fbd5245e01165fc;hp=8630cfee7d7c5a89a532d893bc140b9c456df0d1;hpb=8044e9b3af6e62afe89cd60bf76de22ae7138306;p=oota-llvm.git diff --git a/lib/CodeGen/ScheduleDAG.cpp b/lib/CodeGen/ScheduleDAG.cpp index 8630cfee7d7..9a650710017 100644 --- a/lib/CodeGen/ScheduleDAG.cpp +++ b/lib/CodeGen/ScheduleDAG.cpp @@ -14,71 +14,103 @@ #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 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(); +#ifndef NDEBUG +static cl::opt StressSchedOpt( + "stress-sched", cl::Hidden, cl::init(false), + cl::desc("Stress test instruction scheduling")); +#endif + +void SchedulingPriorityQueue::anchor() { } + +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 } 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; + for (SmallVector::iterator I = Preds.begin(), E = Preds.end(); + I != E; ++I) { + 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 (SmallVector::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); 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) + 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; - N->Succs.push_back(P); + } Preds.push_back(D); - this->setDepthDirty(); - N->setHeightDirty(); + 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 @@ -102,18 +134,27 @@ void SUnit::removePred(const SDep &D) { 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; - this->setDepthDirty(); - N->setHeightDirty(); + } + if (P.getLatency() != 0) { + this->setDepthDirty(); + N->setHeightDirty(); + } return; } } @@ -177,7 +218,7 @@ void SUnit::setHeightToAtLeast(unsigned NewHeight) { void SUnit::ComputeDepth() { SmallVector WorkList; WorkList.push_back(this); - while (!WorkList.empty()) { + do { SUnit *Cur = WorkList.back(); bool Done = true; @@ -202,7 +243,7 @@ void SUnit::ComputeDepth() { } Cur->isDepthCurrent = true; } - } + } while (!WorkList.empty()); } /// ComputeHeight - Calculate the maximal path from the node to the entry. @@ -210,7 +251,7 @@ void SUnit::ComputeDepth() { void SUnit::ComputeHeight() { SmallVector WorkList; WorkList.push_back(this); - while (!WorkList.empty()) { + do { SUnit *Cur = WorkList.back(); bool Done = true; @@ -235,72 +276,76 @@ void SUnit::ComputeHeight() { } Cur->isHeightCurrent = true; } - } + } while (!WorkList.empty()); } +#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 { - 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"; } +#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) { @@ -308,68 +353,64 @@ void ScheduleDAG::VerifySchedule(bool isBottomUp) { 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 && - (isBottomUp ? SUnits[i].getHeight() : SUnits[i].getHeight()) > + (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 " + 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 @@ -401,7 +442,7 @@ void ScheduleDAGTopologicalSort::InitDAGTopologicalSorting() { // Collect leaf nodes. WorkList.push_back(SU); } - } + } int Id = DAGSize; while (!WorkList.empty()) { @@ -426,14 +467,14 @@ void ScheduleDAGTopologicalSort::InitDAGTopologicalSorting() { 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; @@ -451,7 +492,7 @@ void ScheduleDAGTopologicalSort::AddPred(SUnit *Y, SUnit *X) { } } -/// 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) { @@ -462,32 +503,32 @@ 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 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 L; int shift = 0; @@ -538,7 +579,7 @@ bool ScheduleDAGTopologicalSort::IsReachable(const SUnit *SU, // 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; @@ -550,6 +591,7 @@ void ScheduleDAGTopologicalSort::Allocate(int n, int index) { Index2Node[index] = n; } -ScheduleDAGTopologicalSort::ScheduleDAGTopologicalSort( - std::vector &sunits) - : SUnits(sunits) {} +ScheduleDAGTopologicalSort:: +ScheduleDAGTopologicalSort(std::vector &sunits) : SUnits(sunits) {} + +ScheduleHazardRecognizer::~ScheduleHazardRecognizer() {}