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/TargetRegisterInfo.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/PriorityQueue.h"
32 #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(SUnits);
99 ListScheduleTopDown();
101 AvailableQueue->releaseState();
104 //===----------------------------------------------------------------------===//
105 // Top-Down Scheduling
106 //===----------------------------------------------------------------------===//
108 /// ReleaseSucc - Decrement the NumPredsLeft count of a successor. Add it to
109 /// the PendingQueue if the count reaches zero.
110 void ScheduleDAGList::ReleaseSucc(SUnit *SuccSU, bool isChain) {
111 SuccSU->NumPredsLeft--;
113 assert(SuccSU->NumPredsLeft >= 0 &&
114 "List scheduling internal error");
116 if (SuccSU->NumPredsLeft == 0) {
117 // Compute how many cycles it will be before this actually becomes
118 // available. This is the max of the start time of all predecessors plus
120 unsigned AvailableCycle = 0;
121 for (SUnit::pred_iterator I = SuccSU->Preds.begin(),
122 E = SuccSU->Preds.end(); I != E; ++I) {
123 // If this is a token edge, we don't need to wait for the latency of the
124 // preceeding instruction (e.g. a long-latency load) unless there is also
125 // some other data dependence.
126 SUnit &Pred = *I->Dep;
127 unsigned PredDoneCycle = Pred.Cycle;
129 PredDoneCycle += Pred.Latency;
130 else if (Pred.Latency)
133 AvailableCycle = std::max(AvailableCycle, PredDoneCycle);
136 PendingQueue.push_back(std::make_pair(AvailableCycle, SuccSU));
140 /// ScheduleNodeTopDown - Add the node to the schedule. Decrement the pending
141 /// count of its successors. If a successor pending count is zero, add it to
142 /// the Available queue.
143 void ScheduleDAGList::ScheduleNodeTopDown(SUnit *SU, unsigned CurCycle) {
144 DOUT << "*** Scheduling [" << CurCycle << "]: ";
145 DEBUG(SU->dump(&DAG));
147 Sequence.push_back(SU);
148 SU->Cycle = CurCycle;
150 // Bottom up: release successors.
151 for (SUnit::succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
153 ReleaseSucc(I->Dep, I->isCtrl);
156 /// ListScheduleTopDown - The main loop of list scheduling for top-down
158 void ScheduleDAGList::ListScheduleTopDown() {
159 unsigned CurCycle = 0;
161 // All leaves to Available queue.
162 for (unsigned i = 0, e = SUnits.size(); i != e; ++i) {
163 // It is available if it has no predecessors.
164 if (SUnits[i].Preds.empty()) {
165 AvailableQueue->push(&SUnits[i]);
166 SUnits[i].isAvailable = SUnits[i].isPending = true;
170 // While Available queue is not empty, grab the node with the highest
171 // priority. If it is not ready put it back. Schedule the node.
172 std::vector<SUnit*> NotReady;
173 Sequence.reserve(SUnits.size());
174 while (!AvailableQueue->empty() || !PendingQueue.empty()) {
175 // Check to see if any of the pending instructions are ready to issue. If
176 // so, add them to the available queue.
177 for (unsigned i = 0, e = PendingQueue.size(); i != e; ++i) {
178 if (PendingQueue[i].first == CurCycle) {
179 AvailableQueue->push(PendingQueue[i].second);
180 PendingQueue[i].second->isAvailable = true;
181 PendingQueue[i] = PendingQueue.back();
182 PendingQueue.pop_back();
185 assert(PendingQueue[i].first > CurCycle && "Negative latency?");
189 // If there are no instructions available, don't try to issue anything, and
190 // don't advance the hazard recognizer.
191 if (AvailableQueue->empty()) {
196 SUnit *FoundSUnit = 0;
197 SDNode *FoundNode = 0;
199 bool HasNoopHazards = false;
200 while (!AvailableQueue->empty()) {
201 SUnit *CurSUnit = AvailableQueue->pop();
203 // Get the node represented by this SUnit.
204 FoundNode = CurSUnit->Node;
206 // If this is a pseudo op, like copyfromreg, look to see if there is a
207 // real target node flagged to it. If so, use the target node.
208 for (unsigned i = 0, e = CurSUnit->FlaggedNodes.size();
209 !FoundNode->isMachineOpcode() && i != e; ++i)
210 FoundNode = CurSUnit->FlaggedNodes[i];
212 HazardRecognizer::HazardType HT = HazardRec->getHazardType(FoundNode);
213 if (HT == HazardRecognizer::NoHazard) {
214 FoundSUnit = CurSUnit;
218 // Remember if this is a noop hazard.
219 HasNoopHazards |= HT == HazardRecognizer::NoopHazard;
221 NotReady.push_back(CurSUnit);
224 // Add the nodes that aren't ready back onto the available list.
225 if (!NotReady.empty()) {
226 AvailableQueue->push_all(NotReady);
230 // If we found a node to schedule, do it now.
232 ScheduleNodeTopDown(FoundSUnit, CurCycle);
233 HazardRec->EmitInstruction(FoundNode);
234 FoundSUnit->isScheduled = true;
235 AvailableQueue->ScheduledNode(FoundSUnit);
237 // If this is a pseudo-op node, we don't want to increment the current
239 if (FoundSUnit->Latency) // Don't increment CurCycle for pseudo-ops!
241 } else if (!HasNoopHazards) {
242 // Otherwise, we have a pipeline stall, but no other problem, just advance
243 // the current cycle and try again.
244 DOUT << "*** Advancing cycle, no work to do\n";
245 HazardRec->AdvanceCycle();
249 // Otherwise, we have no instructions to issue and we have instructions
250 // that will fault if we don't do this right. This is the case for
251 // processors without pipeline interlocks and other cases.
252 DOUT << "*** Emitting noop\n";
253 HazardRec->EmitNoop();
254 Sequence.push_back(0); // NULL SUnit* -> noop
261 // Verify that all SUnits were scheduled.
262 bool AnyNotSched = false;
263 for (unsigned i = 0, e = SUnits.size(); i != e; ++i) {
264 if (SUnits[i].NumPredsLeft != 0) {
266 cerr << "*** List scheduling failed! ***\n";
267 SUnits[i].dump(&DAG);
268 cerr << "has not been scheduled!\n";
272 assert(!AnyNotSched);
276 //===----------------------------------------------------------------------===//
277 // LatencyPriorityQueue Implementation
278 //===----------------------------------------------------------------------===//
280 // This is a SchedulingPriorityQueue that schedules using latency information to
281 // reduce the length of the critical path through the basic block.
284 class LatencyPriorityQueue;
286 /// Sorting functions for the Available queue.
287 struct latency_sort : public std::binary_function<SUnit*, SUnit*, bool> {
288 LatencyPriorityQueue *PQ;
289 latency_sort(LatencyPriorityQueue *pq) : PQ(pq) {}
290 latency_sort(const latency_sort &RHS) : PQ(RHS.PQ) {}
292 bool operator()(const SUnit* left, const SUnit* right) const;
294 } // end anonymous namespace
297 class LatencyPriorityQueue : public SchedulingPriorityQueue {
298 // SUnits - The SUnits for the current graph.
299 std::vector<SUnit> *SUnits;
301 // Latencies - The latency (max of latency from this node to the bb exit)
303 std::vector<int> Latencies;
305 /// NumNodesSolelyBlocking - This vector contains, for every node in the
306 /// Queue, the number of nodes that the node is the sole unscheduled
307 /// predecessor for. This is used as a tie-breaker heuristic for better
309 std::vector<unsigned> NumNodesSolelyBlocking;
311 PriorityQueue<SUnit*, std::vector<SUnit*>, latency_sort> Queue;
313 LatencyPriorityQueue() : Queue(latency_sort(this)) {
316 void initNodes(std::vector<SUnit> &sunits) {
318 // Calculate node priorities.
319 CalculatePriorities();
322 void addNode(const SUnit *SU) {
323 Latencies.resize(SUnits->size(), -1);
324 NumNodesSolelyBlocking.resize(SUnits->size(), 0);
328 void updateNode(const SUnit *SU) {
329 Latencies[SU->NodeNum] = -1;
333 void releaseState() {
338 unsigned getLatency(unsigned NodeNum) const {
339 assert(NodeNum < Latencies.size());
340 return Latencies[NodeNum];
343 unsigned getNumSolelyBlockNodes(unsigned NodeNum) const {
344 assert(NodeNum < NumNodesSolelyBlocking.size());
345 return NumNodesSolelyBlocking[NodeNum];
348 unsigned size() const { return Queue.size(); }
350 bool empty() const { return Queue.empty(); }
352 virtual void push(SUnit *U) {
355 void push_impl(SUnit *U);
357 void push_all(const std::vector<SUnit *> &Nodes) {
358 for (unsigned i = 0, e = Nodes.size(); i != e; ++i)
363 if (empty()) return NULL;
364 SUnit *V = Queue.top();
369 void remove(SUnit *SU) {
370 assert(!Queue.empty() && "Not in queue!");
374 // ScheduledNode - As nodes are scheduled, we look to see if there are any
375 // successor nodes that have a single unscheduled predecessor. If so, that
376 // single predecessor has a higher priority, since scheduling it will make
377 // the node available.
378 void ScheduledNode(SUnit *Node);
381 void CalculatePriorities();
382 int CalcLatency(const SUnit &SU);
383 void AdjustPriorityOfUnscheduledPreds(SUnit *SU);
384 SUnit *getSingleUnscheduledPred(SUnit *SU);
388 bool latency_sort::operator()(const SUnit *LHS, const SUnit *RHS) const {
389 unsigned LHSNum = LHS->NodeNum;
390 unsigned RHSNum = RHS->NodeNum;
392 // The most important heuristic is scheduling the critical path.
393 unsigned LHSLatency = PQ->getLatency(LHSNum);
394 unsigned RHSLatency = PQ->getLatency(RHSNum);
395 if (LHSLatency < RHSLatency) return true;
396 if (LHSLatency > RHSLatency) return false;
398 // After that, if two nodes have identical latencies, look to see if one will
399 // unblock more other nodes than the other.
400 unsigned LHSBlocked = PQ->getNumSolelyBlockNodes(LHSNum);
401 unsigned RHSBlocked = PQ->getNumSolelyBlockNodes(RHSNum);
402 if (LHSBlocked < RHSBlocked) return true;
403 if (LHSBlocked > RHSBlocked) return false;
405 // Finally, just to provide a stable ordering, use the node number as a
407 return LHSNum < RHSNum;
411 /// CalcNodePriority - Calculate the maximal path from the node to the exit.
413 int LatencyPriorityQueue::CalcLatency(const SUnit &SU) {
414 int &Latency = Latencies[SU.NodeNum];
418 std::vector<const SUnit*> WorkList;
419 WorkList.push_back(&SU);
420 while (!WorkList.empty()) {
421 const SUnit *Cur = WorkList.back();
423 int MaxSuccLatency = 0;
424 for (SUnit::const_succ_iterator I = Cur->Succs.begin(),E = Cur->Succs.end();
426 int SuccLatency = Latencies[I->Dep->NodeNum];
427 if (SuccLatency == -1) {
429 WorkList.push_back(I->Dep);
431 MaxSuccLatency = std::max(MaxSuccLatency, SuccLatency);
435 Latencies[Cur->NodeNum] = MaxSuccLatency + Cur->Latency;
443 /// CalculatePriorities - Calculate priorities of all scheduling units.
444 void LatencyPriorityQueue::CalculatePriorities() {
445 Latencies.assign(SUnits->size(), -1);
446 NumNodesSolelyBlocking.assign(SUnits->size(), 0);
448 // For each node, calculate the maximal path from the node to the exit.
449 std::vector<std::pair<const SUnit*, unsigned> > WorkList;
450 for (unsigned i = 0, e = SUnits->size(); i != e; ++i) {
451 const SUnit *SU = &(*SUnits)[i];
452 if (SU->Succs.empty())
453 WorkList.push_back(std::make_pair(SU, 0U));
456 while (!WorkList.empty()) {
457 const SUnit *SU = WorkList.back().first;
458 unsigned SuccLat = WorkList.back().second;
460 int &Latency = Latencies[SU->NodeNum];
461 if (Latency == -1 || (SU->Latency + SuccLat) > (unsigned)Latency) {
462 Latency = SU->Latency + SuccLat;
463 for (SUnit::const_pred_iterator I = SU->Preds.begin(),E = SU->Preds.end();
465 WorkList.push_back(std::make_pair(I->Dep, Latency));
470 /// getSingleUnscheduledPred - If there is exactly one unscheduled predecessor
471 /// of SU, return it, otherwise return null.
472 SUnit *LatencyPriorityQueue::getSingleUnscheduledPred(SUnit *SU) {
473 SUnit *OnlyAvailablePred = 0;
474 for (SUnit::const_pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
476 SUnit &Pred = *I->Dep;
477 if (!Pred.isScheduled) {
478 // We found an available, but not scheduled, predecessor. If it's the
479 // only one we have found, keep track of it... otherwise give up.
480 if (OnlyAvailablePred && OnlyAvailablePred != &Pred)
482 OnlyAvailablePred = &Pred;
486 return OnlyAvailablePred;
489 void LatencyPriorityQueue::push_impl(SUnit *SU) {
490 // Look at all of the successors of this node. Count the number of nodes that
491 // this node is the sole unscheduled node for.
492 unsigned NumNodesBlocking = 0;
493 for (SUnit::const_succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
495 if (getSingleUnscheduledPred(I->Dep) == SU)
497 NumNodesSolelyBlocking[SU->NodeNum] = NumNodesBlocking;
503 // ScheduledNode - As nodes are scheduled, we look to see if there are any
504 // successor nodes that have a single unscheduled predecessor. If so, that
505 // single predecessor has a higher priority, since scheduling it will make
506 // the node available.
507 void LatencyPriorityQueue::ScheduledNode(SUnit *SU) {
508 for (SUnit::const_succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
510 AdjustPriorityOfUnscheduledPreds(I->Dep);
513 /// AdjustPriorityOfUnscheduledPreds - One of the predecessors of SU was just
514 /// scheduled. If SU is not itself available, then there is at least one
515 /// predecessor node that has not been scheduled yet. If SU has exactly ONE
516 /// unscheduled predecessor, we want to increase its priority: it getting
517 /// scheduled will make this node available, so it is better than some other
518 /// node of the same priority that will not make a node available.
519 void LatencyPriorityQueue::AdjustPriorityOfUnscheduledPreds(SUnit *SU) {
520 if (SU->isPending) return; // All preds scheduled.
522 SUnit *OnlyAvailablePred = getSingleUnscheduledPred(SU);
523 if (OnlyAvailablePred == 0 || !OnlyAvailablePred->isAvailable) return;
525 // Okay, we found a single predecessor that is available, but not scheduled.
526 // Since it is available, it must be in the priority queue. First remove it.
527 remove(OnlyAvailablePred);
529 // Reinsert the node into the priority queue, which recomputes its
530 // NumNodesSolelyBlocking value.
531 push(OnlyAvailablePred);
535 //===----------------------------------------------------------------------===//
536 // Public Constructor Functions
537 //===----------------------------------------------------------------------===//
539 /// createTDListDAGScheduler - This creates a top-down list scheduler with a
540 /// new hazard recognizer. This scheduler takes ownership of the hazard
541 /// recognizer and deletes it when done.
542 ScheduleDAG* llvm::createTDListDAGScheduler(SelectionDAGISel *IS,
544 MachineBasicBlock *BB, bool Fast) {
545 return new ScheduleDAGList(*DAG, BB, DAG->getTarget(),
546 new LatencyPriorityQueue(),
547 IS->CreateTargetHazardRecognizer());