//
//===----------------------------------------------------------------------===//
//
-// This implements a simple two pass scheduler. The first pass attempts to push
-// backward any lengthy instructions and critical paths. The second pass packs
-// instructions into semi-optimal time slots.
+// This implements a top-down list scheduler, using standard algorithms.
+// The basic approach uses a priority queue of available nodes to schedule.
+// One at a time, nodes are taken from the priority queue (thus in priority
+// order), checked for legality to schedule, and emitted if legal.
+//
+// Nodes may not be legal to schedule either due to structural hazards (e.g.
+// pipeline or resource constraints) or because an input to the instruction has
+// not completed execution.
//
//===----------------------------------------------------------------------===//
#define DEBUG_TYPE "sched"
#include "llvm/CodeGen/ScheduleDAG.h"
-#include "llvm/CodeGen/SelectionDAG.h"
+#include "llvm/CodeGen/SchedulerRegistry.h"
+#include "llvm/CodeGen/SelectionDAGISel.h"
+#include "llvm/CodeGen/SSARegMap.h"
+#include "llvm/Target/MRegisterInfo.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/Statistic.h"
#include <climits>
-#include <iostream>
#include <queue>
-#include <set>
-#include <vector>
using namespace llvm;
-namespace {
- Statistic<> NumNoops ("scheduler", "Number of noops inserted");
- Statistic<> NumStalls("scheduler", "Number of pipeline stalls");
-
-/// SUnit - Scheduling unit. It's an wrapper around either a single SDNode or a
-/// group of nodes flagged together.
-struct SUnit {
- SDNode *Node; // Representative node.
- std::vector<SDNode*> FlaggedNodes; // All nodes flagged to Node.
- std::set<SUnit*> Preds; // All real predecessors.
- std::set<SUnit*> ChainPreds; // All chain predecessors.
- std::set<SUnit*> Succs; // All real successors.
- std::set<SUnit*> ChainSuccs; // All chain successors.
- int NumPredsLeft; // # of preds not scheduled.
- int NumSuccsLeft; // # of succs not scheduled.
- int NumChainPredsLeft; // # of chain preds not scheduled.
- int NumChainSuccsLeft; // # of chain succs not scheduled.
- int Priority1; // Scheduling priority 1.
- int Priority2; // Scheduling priority 2.
- bool isTwoAddress; // Is a two-address instruction.
- bool isDefNUseOperand; // Is a def&use operand.
- unsigned Latency; // Node latency.
- unsigned CycleBound; // Upper/lower cycle to be scheduled at.
- unsigned Slot; // Cycle node is scheduled at.
- SUnit *Next;
-
- SUnit(SDNode *node)
- : Node(node), NumPredsLeft(0), NumSuccsLeft(0),
- NumChainPredsLeft(0), NumChainSuccsLeft(0),
- Priority1(INT_MIN), Priority2(INT_MIN),
- isTwoAddress(false), isDefNUseOperand(false),
- Latency(0), CycleBound(0), Slot(0), Next(NULL) {}
-
- void dump(const SelectionDAG *G, bool All=true) const;
-};
-
-void SUnit::dump(const SelectionDAG *G, bool All) const {
- std::cerr << "SU: ";
- Node->dump(G);
- std::cerr << "\n";
- if (FlaggedNodes.size() != 0) {
- for (unsigned i = 0, e = FlaggedNodes.size(); i != e; i++) {
- std::cerr << " ";
- FlaggedNodes[i]->dump(G);
- std::cerr << "\n";
- }
- }
-
- if (All) {
- std::cerr << " # preds left : " << NumPredsLeft << "\n";
- std::cerr << " # succs left : " << NumSuccsLeft << "\n";
- std::cerr << " # chain preds left : " << NumChainPredsLeft << "\n";
- std::cerr << " # chain succs left : " << NumChainSuccsLeft << "\n";
- std::cerr << " Latency : " << Latency << "\n";
- std::cerr << " Priority : " << Priority1 << " , "
- << Priority2 << "\n";
-
- if (Preds.size() != 0) {
- std::cerr << " Predecessors:\n";
- for (std::set<SUnit*>::const_iterator I = Preds.begin(),
- E = Preds.end(); I != E; ++I) {
- std::cerr << " ";
- (*I)->dump(G, false);
- }
- }
- if (ChainPreds.size() != 0) {
- std::cerr << " Chained Preds:\n";
- for (std::set<SUnit*>::const_iterator I = ChainPreds.begin(),
- E = ChainPreds.end(); I != E; ++I) {
- std::cerr << " ";
- (*I)->dump(G, false);
- }
- }
- if (Succs.size() != 0) {
- std::cerr << " Successors:\n";
- for (std::set<SUnit*>::const_iterator I = Succs.begin(),
- E = Succs.end(); I != E; ++I) {
- std::cerr << " ";
- (*I)->dump(G, false);
- }
- }
- if (ChainSuccs.size() != 0) {
- std::cerr << " Chained succs:\n";
- for (std::set<SUnit*>::const_iterator I = ChainSuccs.begin(),
- E = ChainSuccs.end(); I != E; ++I) {
- std::cerr << " ";
- (*I)->dump(G, false);
- }
- }
- }
-}
-
-/// Sorting functions for the Available queue.
-struct ls_rr_sort : public std::binary_function<SUnit*, SUnit*, bool> {
- bool operator()(const SUnit* left, const SUnit* right) const {
- bool LFloater = (left ->Preds.size() == 0);
- bool RFloater = (right->Preds.size() == 0);
- int LBonus = (int)left ->isDefNUseOperand;
- int RBonus = (int)right->isDefNUseOperand;
-
- // Special tie breaker: if two nodes share a operand, the one that
- // use it as a def&use operand is preferred.
- if (left->isTwoAddress && !right->isTwoAddress) {
- SDNode *DUNode = left->Node->getOperand(0).Val;
- if (DUNode->isOperand(right->Node))
- LBonus++;
- }
- if (!left->isTwoAddress && right->isTwoAddress) {
- SDNode *DUNode = right->Node->getOperand(0).Val;
- if (DUNode->isOperand(left->Node))
- RBonus++;
- }
+STATISTIC(NumNoops , "Number of noops inserted");
+STATISTIC(NumStalls, "Number of pipeline stalls");
- int LPriority1 = left ->Priority1 - LBonus;
- int RPriority1 = right->Priority1 - RBonus;
- int LPriority2 = left ->Priority2 + LBonus;
- int RPriority2 = right->Priority2 + RBonus;
-
- // Favor floaters (i.e. node with no non-passive predecessors):
- // e.g. MOV32ri.
- if (!LFloater && RFloater)
- return true;
- else if (LFloater == RFloater)
- if (LPriority1 > RPriority1)
- return true;
- else if (LPriority1 == RPriority1)
- if (LPriority2 < RPriority2)
- return true;
- else if (LPriority1 == RPriority1)
- if (left->CycleBound > right->CycleBound)
- return true;
-
- return false;
- }
-};
-
-
-/// HazardRecognizer - This determines whether or not an instruction can be
-/// issued this cycle, and whether or not a noop needs to be inserted to handle
-/// the hazard.
+static RegisterScheduler
+ tdListDAGScheduler("list-td", " Top-down list scheduler",
+ createTDListDAGScheduler);
+
namespace {
- class HazardRecognizer {
- public:
- virtual ~HazardRecognizer() {}
-
- enum HazardType {
- NoHazard, // This instruction can be emitted at this cycle.
- Hazard, // This instruction can't be emitted at this cycle.
- NoopHazard, // This instruction can't be emitted, and needs noops.
- };
-
- /// getHazardType - Return the hazard type of emitting this node. There are
- /// three possible results. Either:
- /// * NoHazard: it is legal to issue this instruction on this cycle.
- /// * Hazard: issuing this instruction would stall the machine. If some
- /// other instruction is available, issue it first.
- /// * NoopHazard: issuing this instruction would break the program. If
- /// some other instruction can be issued, do so, otherwise issue a noop.
- virtual HazardType getHazardType(SDNode *Node) {
- return NoHazard;
- }
-
- /// EmitInstruction - This callback is invoked when an instruction is
- /// emitted, to advance the hazard state.
- virtual void EmitInstruction(SDNode *Node) {
- }
-
- /// AdvanceCycle - This callback is invoked when no instructions can be
- /// issued on this cycle without a hazard. This should increment the
- /// internal state of the hazard recognizer so that previously "Hazard"
- /// instructions will now not be hazards.
- virtual void AdvanceCycle() {
- }
-
- /// EmitNoop - This callback is invoked when a noop was added to the
- /// instruction stream.
- virtual void EmitNoop() {
- }
- };
-}
-
-
-/// ScheduleDAGList - List scheduler.
-class ScheduleDAGList : public ScheduleDAG {
+//===----------------------------------------------------------------------===//
+/// ScheduleDAGList - The actual list scheduler implementation. This supports
+/// top-down scheduling.
+///
+class VISIBILITY_HIDDEN ScheduleDAGList : public ScheduleDAG {
private:
- // SDNode to SUnit mapping (many to one).
- std::map<SDNode*, SUnit*> SUnitMap;
- // The schedule.
- std::vector<SUnit*> Sequence;
- // Current scheduling cycle.
- unsigned CurrCycle;
- // First and last SUnit created.
- SUnit *HeadSUnit, *TailSUnit;
-
- /// isBottomUp - This is true if the scheduling problem is bottom-up, false if
- /// it is top-down.
- bool isBottomUp;
+ /// AvailableQueue - The priority queue to use for the available SUnits.
+ ///
+ SchedulingPriorityQueue *AvailableQueue;
+ /// PendingQueue - This contains all of the instructions whose operands have
+ /// been issued, but their results are not ready yet (due to the latency of
+ /// the operation). Once the operands becomes available, the instruction is
+ /// added to the AvailableQueue. This keeps track of each SUnit and the
+ /// number of cycles left to execute before the operation is available.
+ std::vector<std::pair<unsigned, SUnit*> > PendingQueue;
+
/// HazardRec - The hazard recognizer to use.
HazardRecognizer *HazardRec;
-
- typedef std::priority_queue<SUnit*, std::vector<SUnit*>, ls_rr_sort>
- AvailableQueueTy;
public:
ScheduleDAGList(SelectionDAG &dag, MachineBasicBlock *bb,
- const TargetMachine &tm, bool isbottomup,
- HazardRecognizer *HR = 0)
- : ScheduleDAG(listSchedulingBURR, dag, bb, tm),
- CurrCycle(0), HeadSUnit(NULL), TailSUnit(NULL), isBottomUp(isbottomup) {
- if (HR == 0) HR = new HazardRecognizer();
- HazardRec = HR;
+ const TargetMachine &tm,
+ SchedulingPriorityQueue *availqueue,
+ HazardRecognizer *HR)
+ : ScheduleDAG(dag, bb, tm),
+ AvailableQueue(availqueue), HazardRec(HR) {
}
~ScheduleDAGList() {
- SUnit *SU = HeadSUnit;
- while (SU) {
- SUnit *NextSU = SU->Next;
- delete SU;
- SU = NextSU;
- }
-
delete HazardRec;
+ delete AvailableQueue;
}
void Schedule();
- void dump() const;
-
private:
- SUnit *NewSUnit(SDNode *N);
- void ReleasePred(AvailableQueueTy &Avail,SUnit *PredSU, bool isChain = false);
- void ReleaseSucc(AvailableQueueTy &Avail,SUnit *SuccSU, bool isChain = false);
- void ScheduleNodeBottomUp(AvailableQueueTy &Avail, SUnit *SU);
- void ScheduleNodeTopDown(AvailableQueueTy &Avail, SUnit *SU);
- int CalcNodePriority(SUnit *SU);
- void CalculatePriorities();
+ void ReleaseSucc(SUnit *SuccSU, bool isChain);
+ void ScheduleNodeTopDown(SUnit *SU, unsigned CurCycle);
void ListScheduleTopDown();
- void ListScheduleBottomUp();
- void BuildSchedUnits();
- void EmitSchedule();
};
-} // end namespace
+} // end anonymous namespace
+HazardRecognizer::~HazardRecognizer() {}
-/// NewSUnit - Creates a new SUnit and return a ptr to it.
-SUnit *ScheduleDAGList::NewSUnit(SDNode *N) {
- SUnit *CurrSUnit = new SUnit(N);
- if (HeadSUnit == NULL)
- HeadSUnit = CurrSUnit;
- if (TailSUnit != NULL)
- TailSUnit->Next = CurrSUnit;
- TailSUnit = CurrSUnit;
-
- return CurrSUnit;
-}
+/// Schedule - Schedule the DAG using list scheduling.
+void ScheduleDAGList::Schedule() {
+ DOUT << "********** List Scheduling **********\n";
+
+ // Build scheduling units.
+ BuildSchedUnits();
-/// ReleasePred - Decrement the NumSuccsLeft count of a predecessor. Add it to
-/// the Available queue is the count reaches zero. Also update its cycle bound.
-void ScheduleDAGList::ReleasePred(AvailableQueueTy &Available,
- SUnit *PredSU, bool isChain) {
- // FIXME: the distance between two nodes is not always == the predecessor's
- // latency. For example, the reader can very well read the register written
- // by the predecessor later than the issue cycle. It also depends on the
- // interrupt model (drain vs. freeze).
- PredSU->CycleBound = std::max(PredSU->CycleBound, CurrCycle + PredSU->Latency);
-
- if (!isChain) {
- PredSU->NumSuccsLeft--;
- PredSU->Priority1++;
- } else
- PredSU->NumChainSuccsLeft--;
+ AvailableQueue->initNodes(SUnitMap, SUnits);
-#ifndef NDEBUG
- if (PredSU->NumSuccsLeft < 0 || PredSU->NumChainSuccsLeft < 0) {
- std::cerr << "*** List scheduling failed! ***\n";
- PredSU->dump(&DAG);
- std::cerr << " has been released too many times!\n";
- assert(0);
- }
-#endif
+ ListScheduleTopDown();
- if ((PredSU->NumSuccsLeft + PredSU->NumChainSuccsLeft) == 0) {
- // EntryToken has to go last! Special case it here.
- if (PredSU->Node->getOpcode() != ISD::EntryToken)
- Available.push(PredSU);
- }
+ AvailableQueue->releaseState();
+
+ DOUT << "*** Final schedule ***\n";
+ DEBUG(dumpSchedule());
+ DOUT << "\n";
+
+ // Emit in scheduled order
+ EmitSchedule();
}
+//===----------------------------------------------------------------------===//
+// Top-Down Scheduling
+//===----------------------------------------------------------------------===//
+
/// ReleaseSucc - Decrement the NumPredsLeft count of a successor. Add it to
-/// the Available queue is the count reaches zero. Also update its cycle bound.
-void ScheduleDAGList::ReleaseSucc(AvailableQueueTy &Available,
- SUnit *SuccSU, bool isChain) {
- // FIXME: the distance between two nodes is not always == the predecessor's
- // latency. For example, the reader can very well read the register written
- // by the predecessor later than the issue cycle. It also depends on the
- // interrupt model (drain vs. freeze).
- SuccSU->CycleBound = std::max(SuccSU->CycleBound, CurrCycle + SuccSU->Latency);
-
- if (!isChain) {
+/// the PendingQueue if the count reaches zero.
+void ScheduleDAGList::ReleaseSucc(SUnit *SuccSU, bool isChain) {
+ if (!isChain)
SuccSU->NumPredsLeft--;
- SuccSU->Priority1++; // FIXME: ??
- } else
+ else
SuccSU->NumChainPredsLeft--;
-#ifndef NDEBUG
- if (SuccSU->NumPredsLeft < 0 || SuccSU->NumChainPredsLeft < 0) {
- std::cerr << "*** List scheduling failed! ***\n";
- SuccSU->dump(&DAG);
- std::cerr << " has been released too many times!\n";
- abort();
- }
-#endif
+ assert(SuccSU->NumPredsLeft >= 0 && SuccSU->NumChainPredsLeft >= 0 &&
+ "List scheduling internal error");
- if ((SuccSU->NumPredsLeft + SuccSU->NumChainPredsLeft) == 0)
- Available.push(SuccSU);
-}
-
-/// ScheduleNodeBottomUp - Add the node to the schedule. Decrement the pending
-/// count of its predecessors. If a predecessor pending count is zero, add it to
-/// the Available queue.
-void ScheduleDAGList::ScheduleNodeBottomUp(AvailableQueueTy &Available,
- SUnit *SU) {
- DEBUG(std::cerr << "*** Scheduling: ");
- DEBUG(SU->dump(&DAG, false));
-
- Sequence.push_back(SU);
- SU->Slot = CurrCycle;
-
- // Bottom up: release predecessors
- for (std::set<SUnit*>::iterator I1 = SU->Preds.begin(),
- E1 = SU->Preds.end(); I1 != E1; ++I1) {
- ReleasePred(Available, *I1);
- SU->NumPredsLeft--;
- SU->Priority1--;
+ if ((SuccSU->NumPredsLeft + SuccSU->NumChainPredsLeft) == 0) {
+ // Compute how many cycles it will be before this actually becomes
+ // available. This is the max of the start time of all predecessors plus
+ // their latencies.
+ unsigned AvailableCycle = 0;
+ for (SUnit::pred_iterator I = SuccSU->Preds.begin(),
+ E = SuccSU->Preds.end(); I != E; ++I) {
+ // If this is a token edge, we don't need to wait for the latency of the
+ // preceeding instruction (e.g. a long-latency load) unless there is also
+ // some other data dependence.
+ SUnit &Pred = *I->first;
+ unsigned PredDoneCycle = Pred.Cycle;
+ if (!I->second)
+ PredDoneCycle += Pred.Latency;
+ else if (Pred.Latency)
+ PredDoneCycle += 1;
+
+ AvailableCycle = std::max(AvailableCycle, PredDoneCycle);
+ }
+
+ PendingQueue.push_back(std::make_pair(AvailableCycle, SuccSU));
}
- for (std::set<SUnit*>::iterator I2 = SU->ChainPreds.begin(),
- E2 = SU->ChainPreds.end(); I2 != E2; ++I2)
- ReleasePred(Available, *I2, true);
-
- CurrCycle++;
}
/// ScheduleNodeTopDown - Add the node to the schedule. Decrement the pending
/// count of its successors. If a successor pending count is zero, add it to
/// the Available queue.
-void ScheduleDAGList::ScheduleNodeTopDown(AvailableQueueTy &Available,
- SUnit *SU) {
- DEBUG(std::cerr << "*** Scheduling: ");
- DEBUG(SU->dump(&DAG, false));
+void ScheduleDAGList::ScheduleNodeTopDown(SUnit *SU, unsigned CurCycle) {
+ DOUT << "*** Scheduling [" << CurCycle << "]: ";
+ DEBUG(SU->dump(&DAG));
Sequence.push_back(SU);
- SU->Slot = CurrCycle;
+ SU->Cycle = CurCycle;
// Bottom up: release successors.
- for (std::set<SUnit*>::iterator I1 = SU->Succs.begin(),
- E1 = SU->Succs.end(); I1 != E1; ++I1) {
- ReleaseSucc(Available, *I1);
- SU->NumSuccsLeft--;
- SU->Priority1--; // FIXME: what is this??
- }
- for (std::set<SUnit*>::iterator I2 = SU->ChainSuccs.begin(),
- E2 = SU->ChainSuccs.end(); I2 != E2; ++I2)
- ReleaseSucc(Available, *I2, true);
-
- CurrCycle++;
-}
-
-/// isReady - True if node's lower cycle bound is less or equal to the current
-/// scheduling cycle. Always true if all nodes have uniform latency 1.
-static inline bool isReady(SUnit *SU, unsigned CurrCycle) {
- return SU->CycleBound <= CurrCycle;
-}
-
-/// ListScheduleBottomUp - The main loop of list scheduling for bottom-up
-/// schedulers.
-void ScheduleDAGList::ListScheduleBottomUp() {
- // Available queue.
- AvailableQueueTy Available;
-
- // Add root to Available queue.
- Available.push(SUnitMap[DAG.getRoot().Val]);
-
- // While Available queue is not empty, grab the node with the highest
- // priority. If it is not ready put it back. Schedule the node.
- std::vector<SUnit*> NotReady;
- while (!Available.empty()) {
- SUnit *CurrNode = Available.top();
- Available.pop();
-
- while (!isReady(CurrNode, CurrCycle)) {
- NotReady.push_back(CurrNode);
- CurrNode = Available.top();
- Available.pop();
- }
-
- // Add the nodes that aren't ready back onto the available list.
- while (!NotReady.empty()) {
- Available.push(NotReady.back());
- NotReady.pop_back();
- }
-
- ScheduleNodeBottomUp(Available, CurrNode);
- }
-
- // Add entry node last
- if (DAG.getEntryNode().Val != DAG.getRoot().Val) {
- SUnit *Entry = SUnitMap[DAG.getEntryNode().Val];
- Entry->Slot = CurrCycle;
- Sequence.push_back(Entry);
- }
-
- // Reverse the order if it is bottom up.
- std::reverse(Sequence.begin(), Sequence.end());
-
-
-#ifndef NDEBUG
- // Verify that all SUnits were scheduled.
- bool AnyNotSched = false;
- for (SUnit *SU = HeadSUnit; SU != NULL; SU = SU->Next) {
- if (SU->NumSuccsLeft != 0 || SU->NumChainSuccsLeft != 0) {
- if (!AnyNotSched)
- std::cerr << "*** List scheduling failed! ***\n";
- SU->dump(&DAG);
- std::cerr << "has not been scheduled!\n";
- AnyNotSched = true;
- }
- }
- assert(!AnyNotSched);
-#endif
+ for (SUnit::succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
+ I != E; ++I)
+ ReleaseSucc(I->first, I->second);
}
/// ListScheduleTopDown - The main loop of list scheduling for top-down
/// schedulers.
void ScheduleDAGList::ListScheduleTopDown() {
- // Available queue.
- AvailableQueueTy Available;
-
- // Emit the entry node first.
+ unsigned CurCycle = 0;
SUnit *Entry = SUnitMap[DAG.getEntryNode().Val];
- ScheduleNodeTopDown(Available, Entry);
- HazardRec->EmitInstruction(Entry->Node);
-
+
// All leaves to Available queue.
- for (SUnit *SU = HeadSUnit; SU != NULL; SU = SU->Next) {
+ for (unsigned i = 0, e = SUnits.size(); i != e; ++i) {
// It is available if it has no predecessors.
- if ((SU->Preds.size() + SU->ChainPreds.size()) == 0 && SU != Entry)
- Available.push(SU);
+ if (SUnits[i].Preds.size() == 0 && &SUnits[i] != Entry) {
+ AvailableQueue->push(&SUnits[i]);
+ SUnits[i].isAvailable = SUnits[i].isPending = true;
+ }
}
+ // Emit the entry node first.
+ ScheduleNodeTopDown(Entry, CurCycle);
+ HazardRec->EmitInstruction(Entry->Node);
+
// While Available queue is not empty, grab the node with the highest
// priority. If it is not ready put it back. Schedule the node.
std::vector<SUnit*> NotReady;
- while (!Available.empty()) {
- SUnit *FoundNode = 0;
+ while (!AvailableQueue->empty() || !PendingQueue.empty()) {
+ // Check to see if any of the pending instructions are ready to issue. If
+ // so, add them to the available queue.
+ for (unsigned i = 0, e = PendingQueue.size(); i != e; ++i) {
+ if (PendingQueue[i].first == CurCycle) {
+ AvailableQueue->push(PendingQueue[i].second);
+ PendingQueue[i].second->isAvailable = true;
+ PendingQueue[i] = PendingQueue.back();
+ PendingQueue.pop_back();
+ --i; --e;
+ } else {
+ assert(PendingQueue[i].first > CurCycle && "Negative latency?");
+ }
+ }
+
+ // If there are no instructions available, don't try to issue anything, and
+ // don't advance the hazard recognizer.
+ if (AvailableQueue->empty()) {
+ ++CurCycle;
+ continue;
+ }
+ SUnit *FoundSUnit = 0;
+ SDNode *FoundNode = 0;
+
bool HasNoopHazards = false;
- do {
- SUnit *CurrNode = Available.top();
- Available.pop();
- HazardRecognizer::HazardType HT =
- HazardRec->getHazardType(CurrNode->Node);
+ while (!AvailableQueue->empty()) {
+ SUnit *CurSUnit = AvailableQueue->pop();
+
+ // Get the node represented by this SUnit.
+ FoundNode = CurSUnit->Node;
+
+ // If this is a pseudo op, like copyfromreg, look to see if there is a
+ // real target node flagged to it. If so, use the target node.
+ for (unsigned i = 0, e = CurSUnit->FlaggedNodes.size();
+ FoundNode->getOpcode() < ISD::BUILTIN_OP_END && i != e; ++i)
+ FoundNode = CurSUnit->FlaggedNodes[i];
+
+ HazardRecognizer::HazardType HT = HazardRec->getHazardType(FoundNode);
if (HT == HazardRecognizer::NoHazard) {
- FoundNode = CurrNode;
+ FoundSUnit = CurSUnit;
break;
}
// Remember if this is a noop hazard.
HasNoopHazards |= HT == HazardRecognizer::NoopHazard;
- NotReady.push_back(CurrNode);
- } while (!Available.empty());
+ NotReady.push_back(CurSUnit);
+ }
// Add the nodes that aren't ready back onto the available list.
- while (!NotReady.empty()) {
- Available.push(NotReady.back());
- NotReady.pop_back();
+ if (!NotReady.empty()) {
+ AvailableQueue->push_all(NotReady);
+ NotReady.clear();
}
// If we found a node to schedule, do it now.
- if (FoundNode) {
- ScheduleNodeTopDown(Available, FoundNode);
- HazardRec->EmitInstruction(FoundNode->Node);
+ if (FoundSUnit) {
+ ScheduleNodeTopDown(FoundSUnit, CurCycle);
+ HazardRec->EmitInstruction(FoundNode);
+ FoundSUnit->isScheduled = true;
+ AvailableQueue->ScheduledNode(FoundSUnit);
+
+ // If this is a pseudo-op node, we don't want to increment the current
+ // cycle.
+ if (FoundSUnit->Latency) // Don't increment CurCycle for pseudo-ops!
+ ++CurCycle;
} else if (!HasNoopHazards) {
// Otherwise, we have a pipeline stall, but no other problem, just advance
// the current cycle and try again.
- DEBUG(std::cerr << "*** Advancing cycle, no work to do");
+ DOUT << "*** Advancing cycle, no work to do\n";
HazardRec->AdvanceCycle();
++NumStalls;
+ ++CurCycle;
} else {
// Otherwise, we have no instructions to issue and we have instructions
// that will fault if we don't do this right. This is the case for
// processors without pipeline interlocks and other cases.
- DEBUG(std::cerr << "*** Emitting noop");
+ DOUT << "*** Emitting noop\n";
HazardRec->EmitNoop();
- // FIXME: Add a noop to the schedule!!
+ Sequence.push_back(0); // NULL SUnit* -> noop
++NumNoops;
+ ++CurCycle;
}
}
#ifndef NDEBUG
// Verify that all SUnits were scheduled.
bool AnyNotSched = false;
- for (SUnit *SU = HeadSUnit; SU != NULL; SU = SU->Next) {
- if (SU->NumPredsLeft != 0 || SU->NumChainPredsLeft != 0) {
+ for (unsigned i = 0, e = SUnits.size(); i != e; ++i) {
+ if (SUnits[i].NumPredsLeft != 0 || SUnits[i].NumChainPredsLeft != 0) {
if (!AnyNotSched)
- std::cerr << "*** List scheduling failed! ***\n";
- SU->dump(&DAG);
- std::cerr << "has not been scheduled!\n";
+ cerr << "*** List scheduling failed! ***\n";
+ SUnits[i].dump(&DAG);
+ cerr << "has not been scheduled!\n";
AnyNotSched = true;
}
}
#endif
}
+//===----------------------------------------------------------------------===//
+// LatencyPriorityQueue Implementation
+//===----------------------------------------------------------------------===//
+//
+// This is a SchedulingPriorityQueue that schedules using latency information to
+// reduce the length of the critical path through the basic block.
+//
+namespace {
+ class LatencyPriorityQueue;
+
+ /// Sorting functions for the Available queue.
+ struct latency_sort : public std::binary_function<SUnit*, SUnit*, bool> {
+ LatencyPriorityQueue *PQ;
+ latency_sort(LatencyPriorityQueue *pq) : PQ(pq) {}
+ latency_sort(const latency_sort &RHS) : PQ(RHS.PQ) {}
+
+ bool operator()(const SUnit* left, const SUnit* right) const;
+ };
+} // end anonymous namespace
+
+namespace {
+ class LatencyPriorityQueue : public SchedulingPriorityQueue {
+ // SUnits - The SUnits for the current graph.
+ std::vector<SUnit> *SUnits;
+
+ // Latencies - The latency (max of latency from this node to the bb exit)
+ // for each node.
+ std::vector<int> Latencies;
-/// CalcNodePriority - Priority1 is just the number of live range genned -
-/// number of live range killed. Priority2 is the Sethi Ullman number. It
-/// returns Priority2 since it is calculated recursively.
-/// Smaller number is the higher priority for Priority2. Reverse is true for
-/// Priority1.
-int ScheduleDAGList::CalcNodePriority(SUnit *SU) {
- if (SU->Priority2 != INT_MIN)
- return SU->Priority2;
-
- SU->Priority1 = SU->NumPredsLeft - SU->NumSuccsLeft;
-
- if (SU->Preds.size() == 0) {
- SU->Priority2 = 1;
- } else {
- int Extra = 0;
- for (std::set<SUnit*>::iterator I = SU->Preds.begin(),
- E = SU->Preds.end(); I != E; ++I) {
- SUnit *PredSU = *I;
- int PredPriority = CalcNodePriority(PredSU);
- if (PredPriority > SU->Priority2) {
- SU->Priority2 = PredPriority;
- Extra = 0;
- } else if (PredPriority == SU->Priority2)
- Extra++;
+ /// NumNodesSolelyBlocking - This vector contains, for every node in the
+ /// Queue, the number of nodes that the node is the sole unscheduled
+ /// predecessor for. This is used as a tie-breaker heuristic for better
+ /// mobility.
+ std::vector<unsigned> NumNodesSolelyBlocking;
+
+ std::priority_queue<SUnit*, std::vector<SUnit*>, latency_sort> Queue;
+public:
+ LatencyPriorityQueue() : Queue(latency_sort(this)) {
+ }
+
+ void initNodes(DenseMap<SDNode*, SUnit*> &sumap,
+ std::vector<SUnit> &sunits) {
+ SUnits = &sunits;
+ // Calculate node priorities.
+ CalculatePriorities();
+ }
+ void releaseState() {
+ SUnits = 0;
+ Latencies.clear();
+ }
+
+ unsigned getLatency(unsigned NodeNum) const {
+ assert(NodeNum < Latencies.size());
+ return Latencies[NodeNum];
+ }
+
+ unsigned getNumSolelyBlockNodes(unsigned NodeNum) const {
+ assert(NodeNum < NumNodesSolelyBlocking.size());
+ return NumNodesSolelyBlocking[NodeNum];
+ }
+
+ bool empty() const { return Queue.empty(); }
+
+ virtual void push(SUnit *U) {
+ push_impl(U);
+ }
+ void push_impl(SUnit *U);
+
+ void push_all(const std::vector<SUnit *> &Nodes) {
+ for (unsigned i = 0, e = Nodes.size(); i != e; ++i)
+ push_impl(Nodes[i]);
+ }
+
+ SUnit *pop() {
+ if (empty()) return NULL;
+ SUnit *V = Queue.top();
+ Queue.pop();
+ return V;
}
- if (SU->Node->getOpcode() != ISD::TokenFactor)
- SU->Priority2 += Extra;
- else
- SU->Priority2 = (Extra == 1) ? 0 : Extra-1;
- }
+ // ScheduledNode - As nodes are scheduled, we look to see if there are any
+ // successor nodes that have a single unscheduled predecessor. If so, that
+ // single predecessor has a higher priority, since scheduling it will make
+ // the node available.
+ void ScheduledNode(SUnit *Node);
+
+private:
+ void CalculatePriorities();
+ int CalcLatency(const SUnit &SU);
+ void AdjustPriorityOfUnscheduledPreds(SUnit *SU);
+ SUnit *getSingleUnscheduledPred(SUnit *SU);
+
+ /// RemoveFromPriorityQueue - This is a really inefficient way to remove a
+ /// node from a priority queue. We should roll our own heap to make this
+ /// better or something.
+ void RemoveFromPriorityQueue(SUnit *SU) {
+ std::vector<SUnit*> Temp;
+
+ assert(!Queue.empty() && "Not in queue!");
+ while (Queue.top() != SU) {
+ Temp.push_back(Queue.top());
+ Queue.pop();
+ assert(!Queue.empty() && "Not in queue!");
+ }
- return SU->Priority2;
+ // Remove the node from the PQ.
+ Queue.pop();
+
+ // Add all the other nodes back.
+ for (unsigned i = 0, e = Temp.size(); i != e; ++i)
+ Queue.push(Temp[i]);
+ }
+ };
}
-/// CalculatePriorities - Calculate priorities of all scheduling units.
-void ScheduleDAGList::CalculatePriorities() {
- for (SUnit *SU = HeadSUnit; SU != NULL; SU = SU->Next) {
- // FIXME: assumes uniform latency for now.
- SU->Latency = 1;
- (void)CalcNodePriority(SU);
- DEBUG(SU->dump(&DAG));
- DEBUG(std::cerr << "\n");
- }
+bool latency_sort::operator()(const SUnit *LHS, const SUnit *RHS) const {
+ unsigned LHSNum = LHS->NodeNum;
+ unsigned RHSNum = RHS->NodeNum;
+
+ // The most important heuristic is scheduling the critical path.
+ unsigned LHSLatency = PQ->getLatency(LHSNum);
+ unsigned RHSLatency = PQ->getLatency(RHSNum);
+ if (LHSLatency < RHSLatency) return true;
+ if (LHSLatency > RHSLatency) return false;
+
+ // After that, if two nodes have identical latencies, look to see if one will
+ // unblock more other nodes than the other.
+ unsigned LHSBlocked = PQ->getNumSolelyBlockNodes(LHSNum);
+ unsigned RHSBlocked = PQ->getNumSolelyBlockNodes(RHSNum);
+ if (LHSBlocked < RHSBlocked) return true;
+ if (LHSBlocked > RHSBlocked) return false;
+
+ // Finally, just to provide a stable ordering, use the node number as a
+ // deciding factor.
+ return LHSNum < RHSNum;
}
-void ScheduleDAGList::BuildSchedUnits() {
- // Pass 1: create the SUnit's.
- for (unsigned i = 0, NC = NodeCount; i < NC; i++) {
- NodeInfo *NI = &Info[i];
- SDNode *N = NI->Node;
- if (isPassiveNode(N))
- continue;
- SUnit *SU;
- if (NI->isInGroup()) {
- if (NI != NI->Group->getBottom()) // Bottom up, so only look at bottom
- continue; // node of the NodeGroup
-
- SU = NewSUnit(N);
- // Find the flagged nodes.
- SDOperand FlagOp = N->getOperand(N->getNumOperands() - 1);
- SDNode *Flag = FlagOp.Val;
- unsigned ResNo = FlagOp.ResNo;
- while (Flag->getValueType(ResNo) == MVT::Flag) {
- NodeInfo *FNI = getNI(Flag);
- assert(FNI->Group == NI->Group);
- SU->FlaggedNodes.insert(SU->FlaggedNodes.begin(), Flag);
- SUnitMap[Flag] = SU;
-
- FlagOp = Flag->getOperand(Flag->getNumOperands() - 1);
- Flag = FlagOp.Val;
- ResNo = FlagOp.ResNo;
- }
- } else {
- SU = NewSUnit(N);
- }
- SUnitMap[N] = SU;
- }
+/// CalcNodePriority - Calculate the maximal path from the node to the exit.
+///
+int LatencyPriorityQueue::CalcLatency(const SUnit &SU) {
+ int &Latency = Latencies[SU.NodeNum];
+ if (Latency != -1)
+ return Latency;
+
+ int MaxSuccLatency = 0;
+ for (SUnit::const_succ_iterator I = SU.Succs.begin(), E = SU.Succs.end();
+ I != E; ++I)
+ MaxSuccLatency = std::max(MaxSuccLatency, CalcLatency(*I->first));
- // Pass 2: add the preds, succs, etc.
- for (SUnit *SU = HeadSUnit; SU != NULL; SU = SU->Next) {
- SDNode *N = SU->Node;
- NodeInfo *NI = getNI(N);
-
- if (N->isTargetOpcode() && TII->isTwoAddrInstr(N->getTargetOpcode()))
- SU->isTwoAddress = true;
-
- if (NI->isInGroup()) {
- // Find all predecessors (of the group).
- NodeGroupOpIterator NGOI(NI);
- while (!NGOI.isEnd()) {
- SDOperand Op = NGOI.next();
- SDNode *OpN = Op.Val;
- MVT::ValueType VT = OpN->getValueType(Op.ResNo);
- NodeInfo *OpNI = getNI(OpN);
- if (OpNI->Group != NI->Group && !isPassiveNode(OpN)) {
- assert(VT != MVT::Flag);
- SUnit *OpSU = SUnitMap[OpN];
- if (VT == MVT::Other) {
- if (SU->ChainPreds.insert(OpSU).second)
- SU->NumChainPredsLeft++;
- if (OpSU->ChainSuccs.insert(SU).second)
- OpSU->NumChainSuccsLeft++;
- } else {
- if (SU->Preds.insert(OpSU).second)
- SU->NumPredsLeft++;
- if (OpSU->Succs.insert(SU).second)
- OpSU->NumSuccsLeft++;
- }
- }
- }
- } else {
- // Find node predecessors.
- for (unsigned j = 0, e = N->getNumOperands(); j != e; j++) {
- SDOperand Op = N->getOperand(j);
- SDNode *OpN = Op.Val;
- MVT::ValueType VT = OpN->getValueType(Op.ResNo);
- if (!isPassiveNode(OpN)) {
- assert(VT != MVT::Flag);
- SUnit *OpSU = SUnitMap[OpN];
- if (VT == MVT::Other) {
- if (SU->ChainPreds.insert(OpSU).second)
- SU->NumChainPredsLeft++;
- if (OpSU->ChainSuccs.insert(SU).second)
- OpSU->NumChainSuccsLeft++;
- } else {
- if (SU->Preds.insert(OpSU).second)
- SU->NumPredsLeft++;
- if (OpSU->Succs.insert(SU).second)
- OpSU->NumSuccsLeft++;
- if (j == 0 && SU->isTwoAddress)
- OpSU->isDefNUseOperand = true;
- }
- }
- }
- }
- }
+ return Latency = MaxSuccLatency + SU.Latency;
}
-/// EmitSchedule - Emit the machine code in scheduled order.
-void ScheduleDAGList::EmitSchedule() {
- for (unsigned i = 0, e = Sequence.size(); i != e; i++) {
- SDNode *N;
- SUnit *SU = Sequence[i];
- for (unsigned j = 0, ee = SU->FlaggedNodes.size(); j != ee; j++) {
- N = SU->FlaggedNodes[j];
- EmitNode(getNI(N));
- }
- EmitNode(getNI(SU->Node));
- }
+/// CalculatePriorities - Calculate priorities of all scheduling units.
+void LatencyPriorityQueue::CalculatePriorities() {
+ Latencies.assign(SUnits->size(), -1);
+ NumNodesSolelyBlocking.assign(SUnits->size(), 0);
+
+ for (unsigned i = 0, e = SUnits->size(); i != e; ++i)
+ CalcLatency((*SUnits)[i]);
}
-/// dump - dump the schedule.
-void ScheduleDAGList::dump() const {
- for (unsigned i = 0, e = Sequence.size(); i != e; i++) {
- SUnit *SU = Sequence[i];
- SU->dump(&DAG, false);
+/// getSingleUnscheduledPred - If there is exactly one unscheduled predecessor
+/// of SU, return it, otherwise return null.
+SUnit *LatencyPriorityQueue::getSingleUnscheduledPred(SUnit *SU) {
+ SUnit *OnlyAvailablePred = 0;
+ for (SUnit::const_pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
+ I != E; ++I) {
+ SUnit &Pred = *I->first;
+ if (!Pred.isScheduled) {
+ // We found an available, but not scheduled, predecessor. If it's the
+ // only one we have found, keep track of it... otherwise give up.
+ if (OnlyAvailablePred && OnlyAvailablePred != &Pred)
+ return 0;
+ OnlyAvailablePred = &Pred;
+ }
}
+
+ return OnlyAvailablePred;
}
-/// Schedule - Schedule the DAG using list scheduling.
-/// FIXME: Right now it only supports the burr (bottom up register reducing)
-/// heuristic.
-void ScheduleDAGList::Schedule() {
- DEBUG(std::cerr << "********** List Scheduling **********\n");
+void LatencyPriorityQueue::push_impl(SUnit *SU) {
+ // Look at all of the successors of this node. Count the number of nodes that
+ // this node is the sole unscheduled node for.
+ unsigned NumNodesBlocking = 0;
+ for (SUnit::const_succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
+ I != E; ++I)
+ if (getSingleUnscheduledPred(I->first) == SU)
+ ++NumNodesBlocking;
+ NumNodesSolelyBlocking[SU->NodeNum] = NumNodesBlocking;
+
+ Queue.push(SU);
+}
- // Build scheduling units.
- BuildSchedUnits();
- // Calculate node prirorities.
- CalculatePriorities();
+// ScheduledNode - As nodes are scheduled, we look to see if there are any
+// successor nodes that have a single unscheduled predecessor. If so, that
+// single predecessor has a higher priority, since scheduling it will make
+// the node available.
+void LatencyPriorityQueue::ScheduledNode(SUnit *SU) {
+ for (SUnit::const_succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
+ I != E; ++I)
+ AdjustPriorityOfUnscheduledPreds(I->first);
+}
- // Execute the actual scheduling loop Top-Down or Bottom-Up as appropriate.
- if (isBottomUp)
- ListScheduleBottomUp();
- else
- ListScheduleTopDown();
+/// AdjustPriorityOfUnscheduledPreds - One of the predecessors of SU was just
+/// scheduled. If SU is not itself available, then there is at least one
+/// predecessor node that has not been scheduled yet. If SU has exactly ONE
+/// unscheduled predecessor, we want to increase its priority: it getting
+/// scheduled will make this node available, so it is better than some other
+/// node of the same priority that will not make a node available.
+void LatencyPriorityQueue::AdjustPriorityOfUnscheduledPreds(SUnit *SU) {
+ if (SU->isPending) return; // All preds scheduled.
- DEBUG(std::cerr << "*** Final schedule ***\n");
- DEBUG(dump());
- DEBUG(std::cerr << "\n");
+ SUnit *OnlyAvailablePred = getSingleUnscheduledPred(SU);
+ if (OnlyAvailablePred == 0 || !OnlyAvailablePred->isAvailable) return;
- // Emit in scheduled order
- EmitSchedule();
-}
+ // Okay, we found a single predecessor that is available, but not scheduled.
+ // Since it is available, it must be in the priority queue. First remove it.
+ RemoveFromPriorityQueue(OnlyAvailablePred);
-llvm::ScheduleDAG* llvm::createBURRListDAGScheduler(SelectionDAG &DAG,
- MachineBasicBlock *BB) {
- return new ScheduleDAGList(DAG, BB, DAG.getTarget(), true);
+ // Reinsert the node into the priority queue, which recomputes its
+ // NumNodesSolelyBlocking value.
+ push(OnlyAvailablePred);
}
-/// G5HazardRecognizer - A hazard recognizer for the PowerPC G5 processor.
-/// FIXME: Move to the PowerPC backend.
-class G5HazardRecognizer : public HazardRecognizer {
- // Totally bogus hazard recognizer, used to test noop insertion. This requires
- // a noop between copyfromreg's.
- unsigned EmittedCopyFromReg;
-public:
- G5HazardRecognizer() {
- EmittedCopyFromReg = 0;
- }
-
- virtual HazardType getHazardType(SDNode *Node) {
- if (Node->getOpcode() == ISD::CopyFromReg && EmittedCopyFromReg)
- return NoopHazard;
- return NoHazard;
- }
-
- /// EmitInstruction - This callback is invoked when an instruction is
- /// emitted, to advance the hazard state.
- virtual void EmitInstruction(SDNode *Node) {
- if (Node->getOpcode() == ISD::CopyFromReg) {
- EmittedCopyFromReg = 5;
- } else if (EmittedCopyFromReg) {
- --EmittedCopyFromReg;
- }
- }
-
- /// AdvanceCycle - This callback is invoked when no instructions can be
- /// issued on this cycle without a hazard. This should increment the
- /// internal state of the hazard recognizer so that previously "Hazard"
- /// instructions will now not be hazards.
- virtual void AdvanceCycle() {
- }
-
- /// EmitNoop - This callback is invoked when a noop was added to the
- /// instruction stream.
- virtual void EmitNoop() {
- --EmittedCopyFromReg;
- }
-};
+//===----------------------------------------------------------------------===//
+// Public Constructor Functions
+//===----------------------------------------------------------------------===//
-/// createTDG5ListDAGScheduler - This creates a top-down list scheduler for
-/// the PowerPC G5. FIXME: pull the priority function out into the PPC
-/// backend!
-ScheduleDAG* llvm::createTDG5ListDAGScheduler(SelectionDAG &DAG,
- MachineBasicBlock *BB) {
- return new ScheduleDAGList(DAG, BB, DAG.getTarget(), false,
- new G5HazardRecognizer());
+/// createTDListDAGScheduler - This creates a top-down list scheduler with a
+/// new hazard recognizer. This scheduler takes ownership of the hazard
+/// recognizer and deletes it when done.
+ScheduleDAG* llvm::createTDListDAGScheduler(SelectionDAGISel *IS,
+ SelectionDAG *DAG,
+ MachineBasicBlock *BB) {
+ return new ScheduleDAGList(*DAG, BB, DAG->getTarget(),
+ new LatencyPriorityQueue(),
+ IS->CreateTargetHazardRecognizer());
}