1 //===---- ScheduleDAGList.cpp - Implement a list scheduler for isel DAG ---===//
3 // The LLVM Compiler Infrastructure
5 // This file was developed by Evan Cheng and is distributed under the
6 // University of Illinois Open Source 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 "sched"
22 #include "llvm/CodeGen/ScheduleDAG.h"
23 #include "llvm/CodeGen/SSARegMap.h"
24 #include "llvm/Target/MRegisterInfo.h"
25 #include "llvm/Target/TargetData.h"
26 #include "llvm/Target/TargetMachine.h"
27 #include "llvm/Target/TargetInstrInfo.h"
28 #include "llvm/Support/Debug.h"
29 #include "llvm/ADT/Statistic.h"
36 Statistic<> NumNoops ("scheduler", "Number of noops inserted");
37 Statistic<> NumStalls("scheduler", "Number of pipeline stalls");
41 //===----------------------------------------------------------------------===//
42 /// ScheduleDAGList - The actual list scheduler implementation. This supports
43 /// top-down scheduling.
45 class ScheduleDAGList : public ScheduleDAG {
47 /// AvailableQueue - The priority queue to use for the available SUnits.
49 SchedulingPriorityQueue *AvailableQueue;
51 /// PendingQueue - This contains all of the instructions whose operands have
52 /// been issued, but their results are not ready yet (due to the latency of
53 /// the operation). Once the operands becomes available, the instruction is
54 /// added to the AvailableQueue. This keeps track of each SUnit and the
55 /// number of cycles left to execute before the operation is available.
56 std::vector<std::pair<unsigned, SUnit*> > PendingQueue;
58 /// HazardRec - The hazard recognizer to use.
59 HazardRecognizer *HazardRec;
62 ScheduleDAGList(SelectionDAG &dag, MachineBasicBlock *bb,
63 const TargetMachine &tm,
64 SchedulingPriorityQueue *availqueue,
66 : ScheduleDAG(dag, bb, tm),
67 AvailableQueue(availqueue), HazardRec(HR) {
72 delete AvailableQueue;
78 void ReleaseSucc(SUnit *SuccSU, bool isChain);
79 void ScheduleNodeTopDown(SUnit *SU, unsigned CurCycle);
80 void ListScheduleTopDown();
82 } // end anonymous namespace
84 HazardRecognizer::~HazardRecognizer() {}
87 /// Schedule - Schedule the DAG using list scheduling.
88 void ScheduleDAGList::Schedule() {
89 DEBUG(std::cerr << "********** List Scheduling **********\n");
91 // Build scheduling units.
94 AvailableQueue->initNodes(SUnits);
96 ListScheduleTopDown();
98 AvailableQueue->releaseState();
100 DEBUG(std::cerr << "*** Final schedule ***\n");
101 DEBUG(dumpSchedule());
102 DEBUG(std::cerr << "\n");
104 // Emit in scheduled order
108 //===----------------------------------------------------------------------===//
109 // Top-Down Scheduling
110 //===----------------------------------------------------------------------===//
112 /// ReleaseSucc - Decrement the NumPredsLeft count of a successor. Add it to
113 /// the PendingQueue if the count reaches zero.
114 void ScheduleDAGList::ReleaseSucc(SUnit *SuccSU, bool isChain) {
116 SuccSU->NumPredsLeft--;
118 SuccSU->NumChainPredsLeft--;
120 assert(SuccSU->NumPredsLeft >= 0 && SuccSU->NumChainPredsLeft >= 0 &&
121 "List scheduling internal error");
123 if ((SuccSU->NumPredsLeft + SuccSU->NumChainPredsLeft) == 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 (std::set<std::pair<SUnit*, bool> >::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 unsigned PredDoneCycle = I->first->Cycle;
135 PredDoneCycle += I->first->Latency;
136 else if (I->first->Latency)
139 AvailableCycle = std::max(AvailableCycle, PredDoneCycle);
142 PendingQueue.push_back(std::make_pair(AvailableCycle, SuccSU));
146 /// ScheduleNodeTopDown - Add the node to the schedule. Decrement the pending
147 /// count of its successors. If a successor pending count is zero, add it to
148 /// the Available queue.
149 void ScheduleDAGList::ScheduleNodeTopDown(SUnit *SU, unsigned CurCycle) {
150 DEBUG(std::cerr << "*** Scheduling [" << CurCycle << "]: ");
151 DEBUG(SU->dump(&DAG));
153 Sequence.push_back(SU);
154 SU->Cycle = CurCycle;
156 // Bottom up: release successors.
157 for (std::set<std::pair<SUnit*, bool> >::iterator I = SU->Succs.begin(),
158 E = SU->Succs.end(); I != E; ++I)
159 ReleaseSucc(I->first, I->second);
162 /// ListScheduleTopDown - The main loop of list scheduling for top-down
164 void ScheduleDAGList::ListScheduleTopDown() {
165 unsigned CurCycle = 0;
166 SUnit *Entry = SUnitMap[DAG.getEntryNode().Val];
168 // All leaves to Available queue.
169 for (unsigned i = 0, e = SUnits.size(); i != e; ++i) {
170 // It is available if it has no predecessors.
171 if (SUnits[i].Preds.size() == 0 && &SUnits[i] != Entry) {
172 AvailableQueue->push(&SUnits[i]);
173 SUnits[i].isAvailable = SUnits[i].isPending = true;
177 // Emit the entry node first.
178 ScheduleNodeTopDown(Entry, CurCycle);
179 HazardRec->EmitInstruction(Entry->Node);
181 // While Available queue is not empty, grab the node with the highest
182 // priority. If it is not ready put it back. Schedule the node.
183 std::vector<SUnit*> NotReady;
184 while (!AvailableQueue->empty() || !PendingQueue.empty()) {
185 // Check to see if any of the pending instructions are ready to issue. If
186 // so, add them to the available queue.
187 for (unsigned i = 0, e = PendingQueue.size(); i != e; ++i) {
188 if (PendingQueue[i].first == CurCycle) {
189 AvailableQueue->push(PendingQueue[i].second);
190 PendingQueue[i].second->isAvailable = true;
191 PendingQueue[i] = PendingQueue.back();
192 PendingQueue.pop_back();
195 assert(PendingQueue[i].first > CurCycle && "Negative latency?");
199 // If there are no instructions available, don't try to issue anything, and
200 // don't advance the hazard recognizer.
201 if (AvailableQueue->empty()) {
206 SUnit *FoundSUnit = 0;
207 SDNode *FoundNode = 0;
209 bool HasNoopHazards = false;
210 while (!AvailableQueue->empty()) {
211 SUnit *CurSUnit = AvailableQueue->pop();
213 // Get the node represented by this SUnit.
214 FoundNode = CurSUnit->Node;
216 // If this is a pseudo op, like copyfromreg, look to see if there is a
217 // real target node flagged to it. If so, use the target node.
218 for (unsigned i = 0, e = CurSUnit->FlaggedNodes.size();
219 FoundNode->getOpcode() < ISD::BUILTIN_OP_END && i != e; ++i)
220 FoundNode = CurSUnit->FlaggedNodes[i];
222 HazardRecognizer::HazardType HT = HazardRec->getHazardType(FoundNode);
223 if (HT == HazardRecognizer::NoHazard) {
224 FoundSUnit = CurSUnit;
228 // Remember if this is a noop hazard.
229 HasNoopHazards |= HT == HazardRecognizer::NoopHazard;
231 NotReady.push_back(CurSUnit);
234 // Add the nodes that aren't ready back onto the available list.
235 if (!NotReady.empty()) {
236 AvailableQueue->push_all(NotReady);
240 // If we found a node to schedule, do it now.
242 ScheduleNodeTopDown(FoundSUnit, CurCycle);
243 HazardRec->EmitInstruction(FoundNode);
244 FoundSUnit->isScheduled = true;
245 AvailableQueue->ScheduledNode(FoundSUnit);
247 // If this is a pseudo-op node, we don't want to increment the current
249 if (FoundSUnit->Latency) // Don't increment CurCycle for pseudo-ops!
251 } else if (!HasNoopHazards) {
252 // Otherwise, we have a pipeline stall, but no other problem, just advance
253 // the current cycle and try again.
254 DEBUG(std::cerr << "*** Advancing cycle, no work to do\n");
255 HazardRec->AdvanceCycle();
259 // Otherwise, we have no instructions to issue and we have instructions
260 // that will fault if we don't do this right. This is the case for
261 // processors without pipeline interlocks and other cases.
262 DEBUG(std::cerr << "*** Emitting noop\n");
263 HazardRec->EmitNoop();
264 Sequence.push_back(0); // NULL SUnit* -> noop
271 // Verify that all SUnits were scheduled.
272 bool AnyNotSched = false;
273 for (unsigned i = 0, e = SUnits.size(); i != e; ++i) {
274 if (SUnits[i].NumPredsLeft != 0 || SUnits[i].NumChainPredsLeft != 0) {
276 std::cerr << "*** List scheduling failed! ***\n";
277 SUnits[i].dump(&DAG);
278 std::cerr << "has not been scheduled!\n";
282 assert(!AnyNotSched);
286 //===----------------------------------------------------------------------===//
287 // LatencyPriorityQueue Implementation
288 //===----------------------------------------------------------------------===//
290 // This is a SchedulingPriorityQueue that schedules using latency information to
291 // reduce the length of the critical path through the basic block.
294 class LatencyPriorityQueue;
296 /// Sorting functions for the Available queue.
297 struct latency_sort : public std::binary_function<SUnit*, SUnit*, bool> {
298 LatencyPriorityQueue *PQ;
299 latency_sort(LatencyPriorityQueue *pq) : PQ(pq) {}
300 latency_sort(const latency_sort &RHS) : PQ(RHS.PQ) {}
302 bool operator()(const SUnit* left, const SUnit* right) const;
304 } // end anonymous namespace
307 class LatencyPriorityQueue : public SchedulingPriorityQueue {
308 // SUnits - The SUnits for the current graph.
309 const std::vector<SUnit> *SUnits;
311 // Latencies - The latency (max of latency from this node to the bb exit)
313 std::vector<int> Latencies;
315 /// NumNodesSolelyBlocking - This vector contains, for every node in the
316 /// Queue, the number of nodes that the node is the sole unscheduled
317 /// predecessor for. This is used as a tie-breaker heuristic for better
319 std::vector<unsigned> NumNodesSolelyBlocking;
321 std::priority_queue<SUnit*, std::vector<SUnit*>, latency_sort> Queue;
323 LatencyPriorityQueue() : Queue(latency_sort(this)) {
326 void initNodes(const std::vector<SUnit> &sunits) {
328 // Calculate node priorities.
329 CalculatePriorities();
331 void releaseState() {
336 unsigned getLatency(unsigned NodeNum) const {
337 assert(NodeNum < Latencies.size());
338 return Latencies[NodeNum];
341 unsigned getNumSolelyBlockNodes(unsigned NodeNum) const {
342 assert(NodeNum < NumNodesSolelyBlocking.size());
343 return NumNodesSolelyBlocking[NodeNum];
346 bool empty() const { return Queue.empty(); }
348 virtual void push(SUnit *U) {
351 void push_impl(SUnit *U);
353 void push_all(const std::vector<SUnit *> &Nodes) {
354 for (unsigned i = 0, e = Nodes.size(); i != e; ++i)
359 SUnit *V = Queue.top();
364 // ScheduledNode - As nodes are scheduled, we look to see if there are any
365 // successor nodes that have a single unscheduled predecessor. If so, that
366 // single predecessor has a higher priority, since scheduling it will make
367 // the node available.
368 void ScheduledNode(SUnit *Node);
371 void CalculatePriorities();
372 int CalcLatency(const SUnit &SU);
373 void AdjustPriorityOfUnscheduledPreds(SUnit *SU);
375 /// RemoveFromPriorityQueue - This is a really inefficient way to remove a
376 /// node from a priority queue. We should roll our own heap to make this
377 /// better or something.
378 void RemoveFromPriorityQueue(SUnit *SU) {
379 std::vector<SUnit*> Temp;
381 assert(!Queue.empty() && "Not in queue!");
382 while (Queue.top() != SU) {
383 Temp.push_back(Queue.top());
385 assert(!Queue.empty() && "Not in queue!");
388 // Remove the node from the PQ.
391 // Add all the other nodes back.
392 for (unsigned i = 0, e = Temp.size(); i != e; ++i)
398 bool latency_sort::operator()(const SUnit *LHS, const SUnit *RHS) const {
399 unsigned LHSNum = LHS->NodeNum;
400 unsigned RHSNum = RHS->NodeNum;
402 // The most important heuristic is scheduling the critical path.
403 unsigned LHSLatency = PQ->getLatency(LHSNum);
404 unsigned RHSLatency = PQ->getLatency(RHSNum);
405 if (LHSLatency < RHSLatency) return true;
406 if (LHSLatency > RHSLatency) return false;
408 // After that, if two nodes have identical latencies, look to see if one will
409 // unblock more other nodes than the other.
410 unsigned LHSBlocked = PQ->getNumSolelyBlockNodes(LHSNum);
411 unsigned RHSBlocked = PQ->getNumSolelyBlockNodes(RHSNum);
412 if (LHSBlocked < RHSBlocked) return true;
413 if (LHSBlocked > RHSBlocked) return false;
415 // Finally, just to provide a stable ordering, use the node number as a
417 return LHSNum < RHSNum;
421 /// CalcNodePriority - Calculate the maximal path from the node to the exit.
423 int LatencyPriorityQueue::CalcLatency(const SUnit &SU) {
424 int &Latency = Latencies[SU.NodeNum];
428 int MaxSuccLatency = 0;
429 for (std::set<std::pair<SUnit*, bool> >::const_iterator I = SU.Succs.begin(),
430 E = SU.Succs.end(); I != E; ++I)
431 MaxSuccLatency = std::max(MaxSuccLatency, CalcLatency(*I->first));
433 return Latency = MaxSuccLatency + SU.Latency;
436 /// CalculatePriorities - Calculate priorities of all scheduling units.
437 void LatencyPriorityQueue::CalculatePriorities() {
438 Latencies.assign(SUnits->size(), -1);
439 NumNodesSolelyBlocking.assign(SUnits->size(), 0);
441 for (unsigned i = 0, e = SUnits->size(); i != e; ++i)
442 CalcLatency((*SUnits)[i]);
445 /// getSingleUnscheduledPred - If there is exactly one unscheduled predecessor
446 /// of SU, return it, otherwise return null.
447 static SUnit *getSingleUnscheduledPred(SUnit *SU) {
448 SUnit *OnlyAvailablePred = 0;
449 for (std::set<std::pair<SUnit*, bool> >::const_iterator I = SU->Preds.begin(),
450 E = SU->Preds.end(); I != E; ++I)
451 if (!I->first->isScheduled) {
452 // We found an available, but not scheduled, predecessor. If it's the
453 // only one we have found, keep track of it... otherwise give up.
454 if (OnlyAvailablePred && OnlyAvailablePred != I->first)
456 OnlyAvailablePred = I->first;
459 return OnlyAvailablePred;
462 void LatencyPriorityQueue::push_impl(SUnit *SU) {
463 // Look at all of the successors of this node. Count the number of nodes that
464 // this node is the sole unscheduled node for.
465 unsigned NumNodesBlocking = 0;
466 for (std::set<std::pair<SUnit*, bool> >::const_iterator I = SU->Succs.begin(),
467 E = SU->Succs.end(); I != E; ++I)
468 if (getSingleUnscheduledPred(I->first) == SU)
470 NumNodesSolelyBlocking[SU->NodeNum] = NumNodesBlocking;
476 // ScheduledNode - As nodes are scheduled, we look to see if there are any
477 // successor nodes that have a single unscheduled predecessor. If so, that
478 // single predecessor has a higher priority, since scheduling it will make
479 // the node available.
480 void LatencyPriorityQueue::ScheduledNode(SUnit *SU) {
481 for (std::set<std::pair<SUnit*, bool> >::const_iterator I = SU->Succs.begin(),
482 E = SU->Succs.end(); I != E; ++I)
483 AdjustPriorityOfUnscheduledPreds(I->first);
486 /// AdjustPriorityOfUnscheduledPreds - One of the predecessors of SU was just
487 /// scheduled. If SU is not itself available, then there is at least one
488 /// predecessor node that has not been scheduled yet. If SU has exactly ONE
489 /// unscheduled predecessor, we want to increase its priority: it getting
490 /// scheduled will make this node available, so it is better than some other
491 /// node of the same priority that will not make a node available.
492 void LatencyPriorityQueue::AdjustPriorityOfUnscheduledPreds(SUnit *SU) {
493 if (SU->isPending) return; // All preds scheduled.
495 SUnit *OnlyAvailablePred = getSingleUnscheduledPred(SU);
496 if (OnlyAvailablePred == 0 || !OnlyAvailablePred->isAvailable) return;
498 // Okay, we found a single predecessor that is available, but not scheduled.
499 // Since it is available, it must be in the priority queue. First remove it.
500 RemoveFromPriorityQueue(OnlyAvailablePred);
502 // Reinsert the node into the priority queue, which recomputes its
503 // NumNodesSolelyBlocking value.
504 push(OnlyAvailablePred);
508 //===----------------------------------------------------------------------===//
509 // Public Constructor Functions
510 //===----------------------------------------------------------------------===//
512 /// createTDListDAGScheduler - This creates a top-down list scheduler with the
513 /// specified hazard recognizer.
514 ScheduleDAG* llvm::createTDListDAGScheduler(SelectionDAG &DAG,
515 MachineBasicBlock *BB,
516 HazardRecognizer *HR) {
517 return new ScheduleDAGList(DAG, BB, DAG.getTarget(),
518 new LatencyPriorityQueue(),