1 //===---- ScheduleDAGList.cpp - Implement a list scheduler for isel DAG ---===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This implements a top-down list scheduler, using standard algorithms.
11 // The basic approach uses a priority queue of available nodes to schedule.
12 // One at a time, nodes are taken from the priority queue (thus in priority
13 // order), checked for legality to schedule, and emitted if legal.
15 // Nodes may not be legal to schedule either due to structural hazards (e.g.
16 // pipeline or resource constraints) or because an input to the instruction has
17 // not completed execution.
19 //===----------------------------------------------------------------------===//
21 #define DEBUG_TYPE "pre-RA-sched"
22 #include "llvm/CodeGen/ScheduleDAG.h"
23 #include "llvm/CodeGen/SchedulerRegistry.h"
24 #include "llvm/CodeGen/SelectionDAGISel.h"
25 #include "llvm/Target/MRegisterInfo.h"
26 #include "llvm/Target/TargetData.h"
27 #include "llvm/Target/TargetMachine.h"
28 #include "llvm/Target/TargetInstrInfo.h"
29 #include "llvm/Support/Debug.h"
30 #include "llvm/Support/Compiler.h"
31 #include "llvm/ADT/Statistic.h"
36 STATISTIC(NumNoops , "Number of noops inserted");
37 STATISTIC(NumStalls, "Number of pipeline stalls");
39 static RegisterScheduler
40 tdListDAGScheduler("list-td", " Top-down list scheduler",
41 createTDListDAGScheduler);
44 //===----------------------------------------------------------------------===//
45 /// ScheduleDAGList - The actual list scheduler implementation. This supports
46 /// top-down scheduling.
48 class VISIBILITY_HIDDEN ScheduleDAGList : public ScheduleDAG {
50 /// AvailableQueue - The priority queue to use for the available SUnits.
52 SchedulingPriorityQueue *AvailableQueue;
54 /// PendingQueue - This contains all of the instructions whose operands have
55 /// been issued, but their results are not ready yet (due to the latency of
56 /// the operation). Once the operands becomes available, the instruction is
57 /// added to the AvailableQueue. This keeps track of each SUnit and the
58 /// number of cycles left to execute before the operation is available.
59 std::vector<std::pair<unsigned, SUnit*> > PendingQueue;
61 /// HazardRec - The hazard recognizer to use.
62 HazardRecognizer *HazardRec;
65 ScheduleDAGList(SelectionDAG &dag, MachineBasicBlock *bb,
66 const TargetMachine &tm,
67 SchedulingPriorityQueue *availqueue,
69 : ScheduleDAG(dag, bb, tm),
70 AvailableQueue(availqueue), HazardRec(HR) {
75 delete AvailableQueue;
81 void ReleaseSucc(SUnit *SuccSU, bool isChain);
82 void ScheduleNodeTopDown(SUnit *SU, unsigned CurCycle);
83 void ListScheduleTopDown();
85 } // end anonymous namespace
87 HazardRecognizer::~HazardRecognizer() {}
90 /// Schedule - Schedule the DAG using list scheduling.
91 void ScheduleDAGList::Schedule() {
92 DOUT << "********** List Scheduling **********\n";
94 // Build scheduling units.
97 AvailableQueue->initNodes(SUnitMap, SUnits);
99 ListScheduleTopDown();
101 AvailableQueue->releaseState();
103 DOUT << "*** Final schedule ***\n";
104 DEBUG(dumpSchedule());
107 // Emit in scheduled order
111 //===----------------------------------------------------------------------===//
112 // Top-Down Scheduling
113 //===----------------------------------------------------------------------===//
115 /// ReleaseSucc - Decrement the NumPredsLeft count of a successor. Add it to
116 /// the PendingQueue if the count reaches zero.
117 void ScheduleDAGList::ReleaseSucc(SUnit *SuccSU, bool isChain) {
118 SuccSU->NumPredsLeft--;
120 assert(SuccSU->NumPredsLeft >= 0 &&
121 "List scheduling internal error");
123 if (SuccSU->NumPredsLeft == 0) {
124 // Compute how many cycles it will be before this actually becomes
125 // available. This is the max of the start time of all predecessors plus
127 unsigned AvailableCycle = 0;
128 for (SUnit::pred_iterator I = SuccSU->Preds.begin(),
129 E = SuccSU->Preds.end(); I != E; ++I) {
130 // If this is a token edge, we don't need to wait for the latency of the
131 // preceeding instruction (e.g. a long-latency load) unless there is also
132 // some other data dependence.
133 SUnit &Pred = *I->Dep;
134 unsigned PredDoneCycle = Pred.Cycle;
136 PredDoneCycle += Pred.Latency;
137 else if (Pred.Latency)
140 AvailableCycle = std::max(AvailableCycle, PredDoneCycle);
143 PendingQueue.push_back(std::make_pair(AvailableCycle, SuccSU));
147 /// ScheduleNodeTopDown - Add the node to the schedule. Decrement the pending
148 /// count of its successors. If a successor pending count is zero, add it to
149 /// the Available queue.
150 void ScheduleDAGList::ScheduleNodeTopDown(SUnit *SU, unsigned CurCycle) {
151 DOUT << "*** Scheduling [" << CurCycle << "]: ";
152 DEBUG(SU->dump(&DAG));
154 Sequence.push_back(SU);
155 SU->Cycle = CurCycle;
157 // Bottom up: release successors.
158 for (SUnit::succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
160 ReleaseSucc(I->Dep, I->isCtrl);
163 /// ListScheduleTopDown - The main loop of list scheduling for top-down
165 void ScheduleDAGList::ListScheduleTopDown() {
166 unsigned CurCycle = 0;
167 SUnit *Entry = SUnitMap[DAG.getEntryNode().Val].front();
169 // All leaves to Available queue.
170 for (unsigned i = 0, e = SUnits.size(); i != e; ++i) {
171 // It is available if it has no predecessors.
172 if (SUnits[i].Preds.empty() && &SUnits[i] != Entry) {
173 AvailableQueue->push(&SUnits[i]);
174 SUnits[i].isAvailable = SUnits[i].isPending = true;
178 // Emit the entry node first.
179 ScheduleNodeTopDown(Entry, CurCycle);
180 HazardRec->EmitInstruction(Entry->Node);
182 // While Available queue is not empty, grab the node with the highest
183 // priority. If it is not ready put it back. Schedule the node.
184 std::vector<SUnit*> NotReady;
185 while (!AvailableQueue->empty() || !PendingQueue.empty()) {
186 // Check to see if any of the pending instructions are ready to issue. If
187 // so, add them to the available queue.
188 for (unsigned i = 0, e = PendingQueue.size(); i != e; ++i) {
189 if (PendingQueue[i].first == CurCycle) {
190 AvailableQueue->push(PendingQueue[i].second);
191 PendingQueue[i].second->isAvailable = true;
192 PendingQueue[i] = PendingQueue.back();
193 PendingQueue.pop_back();
196 assert(PendingQueue[i].first > CurCycle && "Negative latency?");
200 // If there are no instructions available, don't try to issue anything, and
201 // don't advance the hazard recognizer.
202 if (AvailableQueue->empty()) {
207 SUnit *FoundSUnit = 0;
208 SDNode *FoundNode = 0;
210 bool HasNoopHazards = false;
211 while (!AvailableQueue->empty()) {
212 SUnit *CurSUnit = AvailableQueue->pop();
214 // Get the node represented by this SUnit.
215 FoundNode = CurSUnit->Node;
217 // If this is a pseudo op, like copyfromreg, look to see if there is a
218 // real target node flagged to it. If so, use the target node.
219 for (unsigned i = 0, e = CurSUnit->FlaggedNodes.size();
220 FoundNode->getOpcode() < ISD::BUILTIN_OP_END && i != e; ++i)
221 FoundNode = CurSUnit->FlaggedNodes[i];
223 HazardRecognizer::HazardType HT = HazardRec->getHazardType(FoundNode);
224 if (HT == HazardRecognizer::NoHazard) {
225 FoundSUnit = CurSUnit;
229 // Remember if this is a noop hazard.
230 HasNoopHazards |= HT == HazardRecognizer::NoopHazard;
232 NotReady.push_back(CurSUnit);
235 // Add the nodes that aren't ready back onto the available list.
236 if (!NotReady.empty()) {
237 AvailableQueue->push_all(NotReady);
241 // If we found a node to schedule, do it now.
243 ScheduleNodeTopDown(FoundSUnit, CurCycle);
244 HazardRec->EmitInstruction(FoundNode);
245 FoundSUnit->isScheduled = true;
246 AvailableQueue->ScheduledNode(FoundSUnit);
248 // If this is a pseudo-op node, we don't want to increment the current
250 if (FoundSUnit->Latency) // Don't increment CurCycle for pseudo-ops!
252 } else if (!HasNoopHazards) {
253 // Otherwise, we have a pipeline stall, but no other problem, just advance
254 // the current cycle and try again.
255 DOUT << "*** Advancing cycle, no work to do\n";
256 HazardRec->AdvanceCycle();
260 // Otherwise, we have no instructions to issue and we have instructions
261 // that will fault if we don't do this right. This is the case for
262 // processors without pipeline interlocks and other cases.
263 DOUT << "*** Emitting noop\n";
264 HazardRec->EmitNoop();
265 Sequence.push_back(0); // NULL SUnit* -> noop
272 // Verify that all SUnits were scheduled.
273 bool AnyNotSched = false;
274 for (unsigned i = 0, e = SUnits.size(); i != e; ++i) {
275 if (SUnits[i].NumPredsLeft != 0) {
277 cerr << "*** List scheduling failed! ***\n";
278 SUnits[i].dump(&DAG);
279 cerr << "has not been scheduled!\n";
283 assert(!AnyNotSched);
287 //===----------------------------------------------------------------------===//
288 // LatencyPriorityQueue Implementation
289 //===----------------------------------------------------------------------===//
291 // This is a SchedulingPriorityQueue that schedules using latency information to
292 // reduce the length of the critical path through the basic block.
295 class LatencyPriorityQueue;
297 /// Sorting functions for the Available queue.
298 struct latency_sort : public std::binary_function<SUnit*, SUnit*, bool> {
299 LatencyPriorityQueue *PQ;
300 latency_sort(LatencyPriorityQueue *pq) : PQ(pq) {}
301 latency_sort(const latency_sort &RHS) : PQ(RHS.PQ) {}
303 bool operator()(const SUnit* left, const SUnit* right) const;
305 } // end anonymous namespace
308 class LatencyPriorityQueue : public SchedulingPriorityQueue {
309 // SUnits - The SUnits for the current graph.
310 std::vector<SUnit> *SUnits;
312 // Latencies - The latency (max of latency from this node to the bb exit)
314 std::vector<int> Latencies;
316 /// NumNodesSolelyBlocking - This vector contains, for every node in the
317 /// Queue, the number of nodes that the node is the sole unscheduled
318 /// predecessor for. This is used as a tie-breaker heuristic for better
320 std::vector<unsigned> NumNodesSolelyBlocking;
322 std::priority_queue<SUnit*, std::vector<SUnit*>, latency_sort> Queue;
324 LatencyPriorityQueue() : Queue(latency_sort(this)) {
327 void initNodes(DenseMap<SDNode*, std::vector<SUnit*> > &sumap,
328 std::vector<SUnit> &sunits) {
330 // Calculate node priorities.
331 CalculatePriorities();
334 void addNode(const SUnit *SU) {
335 Latencies.resize(SUnits->size(), -1);
336 NumNodesSolelyBlocking.resize(SUnits->size(), 0);
340 void updateNode(const SUnit *SU) {
341 Latencies[SU->NodeNum] = -1;
345 void releaseState() {
350 unsigned getLatency(unsigned NodeNum) const {
351 assert(NodeNum < Latencies.size());
352 return Latencies[NodeNum];
355 unsigned getNumSolelyBlockNodes(unsigned NodeNum) const {
356 assert(NodeNum < NumNodesSolelyBlocking.size());
357 return NumNodesSolelyBlocking[NodeNum];
360 unsigned size() const { return Queue.size(); }
362 bool empty() const { return Queue.empty(); }
364 virtual void push(SUnit *U) {
367 void push_impl(SUnit *U);
369 void push_all(const std::vector<SUnit *> &Nodes) {
370 for (unsigned i = 0, e = Nodes.size(); i != e; ++i)
375 if (empty()) return NULL;
376 SUnit *V = Queue.top();
381 /// remove - This is a really inefficient way to remove a node from a
382 /// priority queue. We should roll our own heap to make this better or
384 void remove(SUnit *SU) {
385 std::vector<SUnit*> Temp;
387 assert(!Queue.empty() && "Not in queue!");
388 while (Queue.top() != SU) {
389 Temp.push_back(Queue.top());
391 assert(!Queue.empty() && "Not in queue!");
394 // Remove the node from the PQ.
397 // Add all the other nodes back.
398 for (unsigned i = 0, e = Temp.size(); i != e; ++i)
402 // ScheduledNode - As nodes are scheduled, we look to see if there are any
403 // successor nodes that have a single unscheduled predecessor. If so, that
404 // single predecessor has a higher priority, since scheduling it will make
405 // the node available.
406 void ScheduledNode(SUnit *Node);
409 void CalculatePriorities();
410 int CalcLatency(const SUnit &SU);
411 void AdjustPriorityOfUnscheduledPreds(SUnit *SU);
412 SUnit *getSingleUnscheduledPred(SUnit *SU);
416 bool latency_sort::operator()(const SUnit *LHS, const SUnit *RHS) const {
417 unsigned LHSNum = LHS->NodeNum;
418 unsigned RHSNum = RHS->NodeNum;
420 // The most important heuristic is scheduling the critical path.
421 unsigned LHSLatency = PQ->getLatency(LHSNum);
422 unsigned RHSLatency = PQ->getLatency(RHSNum);
423 if (LHSLatency < RHSLatency) return true;
424 if (LHSLatency > RHSLatency) return false;
426 // After that, if two nodes have identical latencies, look to see if one will
427 // unblock more other nodes than the other.
428 unsigned LHSBlocked = PQ->getNumSolelyBlockNodes(LHSNum);
429 unsigned RHSBlocked = PQ->getNumSolelyBlockNodes(RHSNum);
430 if (LHSBlocked < RHSBlocked) return true;
431 if (LHSBlocked > RHSBlocked) return false;
433 // Finally, just to provide a stable ordering, use the node number as a
435 return LHSNum < RHSNum;
439 /// CalcNodePriority - Calculate the maximal path from the node to the exit.
441 int LatencyPriorityQueue::CalcLatency(const SUnit &SU) {
442 int &Latency = Latencies[SU.NodeNum];
446 std::vector<const SUnit*> WorkList;
447 WorkList.push_back(&SU);
448 while (!WorkList.empty()) {
449 const SUnit *Cur = WorkList.back();
451 int MaxSuccLatency = 0;
452 for (SUnit::const_succ_iterator I = Cur->Succs.begin(),E = Cur->Succs.end();
454 int SuccLatency = Latencies[I->Dep->NodeNum];
455 if (SuccLatency == -1) {
457 WorkList.push_back(I->Dep);
459 MaxSuccLatency = std::max(MaxSuccLatency, SuccLatency);
463 Latencies[Cur->NodeNum] = MaxSuccLatency + Cur->Latency;
471 /// CalculatePriorities - Calculate priorities of all scheduling units.
472 void LatencyPriorityQueue::CalculatePriorities() {
473 Latencies.assign(SUnits->size(), -1);
474 NumNodesSolelyBlocking.assign(SUnits->size(), 0);
476 // For each node, calculate the maximal path from the node to the exit.
477 std::vector<std::pair<const SUnit*, unsigned> > WorkList;
478 for (unsigned i = 0, e = SUnits->size(); i != e; ++i) {
479 const SUnit *SU = &(*SUnits)[i];
480 if (SU->Succs.empty())
481 WorkList.push_back(std::make_pair(SU, 0U));
484 while (!WorkList.empty()) {
485 const SUnit *SU = WorkList.back().first;
486 unsigned SuccLat = WorkList.back().second;
488 int &Latency = Latencies[SU->NodeNum];
489 if (Latency == -1 || (SU->Latency + SuccLat) > (unsigned)Latency) {
490 Latency = SU->Latency + SuccLat;
491 for (SUnit::const_pred_iterator I = SU->Preds.begin(),E = SU->Preds.end();
493 WorkList.push_back(std::make_pair(I->Dep, Latency));
498 /// getSingleUnscheduledPred - If there is exactly one unscheduled predecessor
499 /// of SU, return it, otherwise return null.
500 SUnit *LatencyPriorityQueue::getSingleUnscheduledPred(SUnit *SU) {
501 SUnit *OnlyAvailablePred = 0;
502 for (SUnit::const_pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
504 SUnit &Pred = *I->Dep;
505 if (!Pred.isScheduled) {
506 // We found an available, but not scheduled, predecessor. If it's the
507 // only one we have found, keep track of it... otherwise give up.
508 if (OnlyAvailablePred && OnlyAvailablePred != &Pred)
510 OnlyAvailablePred = &Pred;
514 return OnlyAvailablePred;
517 void LatencyPriorityQueue::push_impl(SUnit *SU) {
518 // Look at all of the successors of this node. Count the number of nodes that
519 // this node is the sole unscheduled node for.
520 unsigned NumNodesBlocking = 0;
521 for (SUnit::const_succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
523 if (getSingleUnscheduledPred(I->Dep) == SU)
525 NumNodesSolelyBlocking[SU->NodeNum] = NumNodesBlocking;
531 // ScheduledNode - As nodes are scheduled, we look to see if there are any
532 // successor nodes that have a single unscheduled predecessor. If so, that
533 // single predecessor has a higher priority, since scheduling it will make
534 // the node available.
535 void LatencyPriorityQueue::ScheduledNode(SUnit *SU) {
536 for (SUnit::const_succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
538 AdjustPriorityOfUnscheduledPreds(I->Dep);
541 /// AdjustPriorityOfUnscheduledPreds - One of the predecessors of SU was just
542 /// scheduled. If SU is not itself available, then there is at least one
543 /// predecessor node that has not been scheduled yet. If SU has exactly ONE
544 /// unscheduled predecessor, we want to increase its priority: it getting
545 /// scheduled will make this node available, so it is better than some other
546 /// node of the same priority that will not make a node available.
547 void LatencyPriorityQueue::AdjustPriorityOfUnscheduledPreds(SUnit *SU) {
548 if (SU->isPending) return; // All preds scheduled.
550 SUnit *OnlyAvailablePred = getSingleUnscheduledPred(SU);
551 if (OnlyAvailablePred == 0 || !OnlyAvailablePred->isAvailable) return;
553 // Okay, we found a single predecessor that is available, but not scheduled.
554 // Since it is available, it must be in the priority queue. First remove it.
555 remove(OnlyAvailablePred);
557 // Reinsert the node into the priority queue, which recomputes its
558 // NumNodesSolelyBlocking value.
559 push(OnlyAvailablePred);
563 //===----------------------------------------------------------------------===//
564 // Public Constructor Functions
565 //===----------------------------------------------------------------------===//
567 /// createTDListDAGScheduler - This creates a top-down list scheduler with a
568 /// new hazard recognizer. This scheduler takes ownership of the hazard
569 /// recognizer and deletes it when done.
570 ScheduleDAG* llvm::createTDListDAGScheduler(SelectionDAGISel *IS,
572 MachineBasicBlock *BB) {
573 return new ScheduleDAGList(*DAG, BB, DAG->getTarget(),
574 new LatencyPriorityQueue(),
575 IS->CreateTargetHazardRecognizer());