1 //===- MachineScheduler.cpp - Machine Instruction Scheduler ---------------===//
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 // MachineScheduler schedules machine instructions after phi elimination. It
11 // preserves LiveIntervals so it can be invoked before register allocation.
13 //===----------------------------------------------------------------------===//
15 #define DEBUG_TYPE "misched"
17 #include "llvm/CodeGen/MachineScheduler.h"
18 #include "llvm/ADT/OwningPtr.h"
19 #include "llvm/ADT/PriorityQueue.h"
20 #include "llvm/Analysis/AliasAnalysis.h"
21 #include "llvm/CodeGen/LiveIntervalAnalysis.h"
22 #include "llvm/CodeGen/Passes.h"
23 #include "llvm/CodeGen/RegisterClassInfo.h"
24 #include "llvm/CodeGen/ScheduleDFS.h"
25 #include "llvm/CodeGen/ScheduleHazardRecognizer.h"
26 #include "llvm/Support/CommandLine.h"
27 #include "llvm/Support/Debug.h"
28 #include "llvm/Support/ErrorHandling.h"
29 #include "llvm/Support/GraphWriter.h"
30 #include "llvm/Support/raw_ostream.h"
36 cl::opt<bool> ForceTopDown("misched-topdown", cl::Hidden,
37 cl::desc("Force top-down list scheduling"));
38 cl::opt<bool> ForceBottomUp("misched-bottomup", cl::Hidden,
39 cl::desc("Force bottom-up list scheduling"));
43 static cl::opt<bool> ViewMISchedDAGs("view-misched-dags", cl::Hidden,
44 cl::desc("Pop up a window to show MISched dags after they are processed"));
46 static cl::opt<unsigned> MISchedCutoff("misched-cutoff", cl::Hidden,
47 cl::desc("Stop scheduling after N instructions"), cl::init(~0U));
49 static bool ViewMISchedDAGs = false;
52 // Experimental heuristics
53 static cl::opt<bool> EnableLoadCluster("misched-cluster", cl::Hidden,
54 cl::desc("Enable load clustering."), cl::init(true));
56 // Experimental heuristics
57 static cl::opt<bool> EnableMacroFusion("misched-fusion", cl::Hidden,
58 cl::desc("Enable scheduling for macro fusion."), cl::init(true));
60 // DAG subtrees must have at least this many nodes.
61 static const unsigned MinSubtreeSize = 8;
63 //===----------------------------------------------------------------------===//
64 // Machine Instruction Scheduling Pass and Registry
65 //===----------------------------------------------------------------------===//
67 MachineSchedContext::MachineSchedContext():
68 MF(0), MLI(0), MDT(0), PassConfig(0), AA(0), LIS(0) {
69 RegClassInfo = new RegisterClassInfo();
72 MachineSchedContext::~MachineSchedContext() {
77 /// MachineScheduler runs after coalescing and before register allocation.
78 class MachineScheduler : public MachineSchedContext,
79 public MachineFunctionPass {
83 virtual void getAnalysisUsage(AnalysisUsage &AU) const;
85 virtual void releaseMemory() {}
87 virtual bool runOnMachineFunction(MachineFunction&);
89 virtual void print(raw_ostream &O, const Module* = 0) const;
91 static char ID; // Class identification, replacement for typeinfo
95 char MachineScheduler::ID = 0;
97 char &llvm::MachineSchedulerID = MachineScheduler::ID;
99 INITIALIZE_PASS_BEGIN(MachineScheduler, "misched",
100 "Machine Instruction Scheduler", false, false)
101 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
102 INITIALIZE_PASS_DEPENDENCY(SlotIndexes)
103 INITIALIZE_PASS_DEPENDENCY(LiveIntervals)
104 INITIALIZE_PASS_END(MachineScheduler, "misched",
105 "Machine Instruction Scheduler", false, false)
107 MachineScheduler::MachineScheduler()
108 : MachineFunctionPass(ID) {
109 initializeMachineSchedulerPass(*PassRegistry::getPassRegistry());
112 void MachineScheduler::getAnalysisUsage(AnalysisUsage &AU) const {
113 AU.setPreservesCFG();
114 AU.addRequiredID(MachineDominatorsID);
115 AU.addRequired<MachineLoopInfo>();
116 AU.addRequired<AliasAnalysis>();
117 AU.addRequired<TargetPassConfig>();
118 AU.addRequired<SlotIndexes>();
119 AU.addPreserved<SlotIndexes>();
120 AU.addRequired<LiveIntervals>();
121 AU.addPreserved<LiveIntervals>();
122 MachineFunctionPass::getAnalysisUsage(AU);
125 MachinePassRegistry MachineSchedRegistry::Registry;
127 /// A dummy default scheduler factory indicates whether the scheduler
128 /// is overridden on the command line.
129 static ScheduleDAGInstrs *useDefaultMachineSched(MachineSchedContext *C) {
133 /// MachineSchedOpt allows command line selection of the scheduler.
134 static cl::opt<MachineSchedRegistry::ScheduleDAGCtor, false,
135 RegisterPassParser<MachineSchedRegistry> >
136 MachineSchedOpt("misched",
137 cl::init(&useDefaultMachineSched), cl::Hidden,
138 cl::desc("Machine instruction scheduler to use"));
140 static MachineSchedRegistry
141 DefaultSchedRegistry("default", "Use the target's default scheduler choice.",
142 useDefaultMachineSched);
144 /// Forward declare the standard machine scheduler. This will be used as the
145 /// default scheduler if the target does not set a default.
146 static ScheduleDAGInstrs *createConvergingSched(MachineSchedContext *C);
149 /// Decrement this iterator until reaching the top or a non-debug instr.
150 static MachineBasicBlock::iterator
151 priorNonDebug(MachineBasicBlock::iterator I, MachineBasicBlock::iterator Beg) {
152 assert(I != Beg && "reached the top of the region, cannot decrement");
154 if (!I->isDebugValue())
160 /// If this iterator is a debug value, increment until reaching the End or a
161 /// non-debug instruction.
162 static MachineBasicBlock::iterator
163 nextIfDebug(MachineBasicBlock::iterator I, MachineBasicBlock::iterator End) {
164 for(; I != End; ++I) {
165 if (!I->isDebugValue())
171 /// Top-level MachineScheduler pass driver.
173 /// Visit blocks in function order. Divide each block into scheduling regions
174 /// and visit them bottom-up. Visiting regions bottom-up is not required, but is
175 /// consistent with the DAG builder, which traverses the interior of the
176 /// scheduling regions bottom-up.
178 /// This design avoids exposing scheduling boundaries to the DAG builder,
179 /// simplifying the DAG builder's support for "special" target instructions.
180 /// At the same time the design allows target schedulers to operate across
181 /// scheduling boundaries, for example to bundle the boudary instructions
182 /// without reordering them. This creates complexity, because the target
183 /// scheduler must update the RegionBegin and RegionEnd positions cached by
184 /// ScheduleDAGInstrs whenever adding or removing instructions. A much simpler
185 /// design would be to split blocks at scheduling boundaries, but LLVM has a
186 /// general bias against block splitting purely for implementation simplicity.
187 bool MachineScheduler::runOnMachineFunction(MachineFunction &mf) {
188 DEBUG(dbgs() << "Before MISsched:\n"; mf.print(dbgs()));
190 // Initialize the context of the pass.
192 MLI = &getAnalysis<MachineLoopInfo>();
193 MDT = &getAnalysis<MachineDominatorTree>();
194 PassConfig = &getAnalysis<TargetPassConfig>();
195 AA = &getAnalysis<AliasAnalysis>();
197 LIS = &getAnalysis<LiveIntervals>();
198 const TargetInstrInfo *TII = MF->getTarget().getInstrInfo();
200 RegClassInfo->runOnMachineFunction(*MF);
202 // Select the scheduler, or set the default.
203 MachineSchedRegistry::ScheduleDAGCtor Ctor = MachineSchedOpt;
204 if (Ctor == useDefaultMachineSched) {
205 // Get the default scheduler set by the target.
206 Ctor = MachineSchedRegistry::getDefault();
208 Ctor = createConvergingSched;
209 MachineSchedRegistry::setDefault(Ctor);
212 // Instantiate the selected scheduler.
213 OwningPtr<ScheduleDAGInstrs> Scheduler(Ctor(this));
215 // Visit all machine basic blocks.
217 // TODO: Visit blocks in global postorder or postorder within the bottom-up
218 // loop tree. Then we can optionally compute global RegPressure.
219 for (MachineFunction::iterator MBB = MF->begin(), MBBEnd = MF->end();
220 MBB != MBBEnd; ++MBB) {
222 Scheduler->startBlock(MBB);
224 // Break the block into scheduling regions [I, RegionEnd), and schedule each
225 // region as soon as it is discovered. RegionEnd points the scheduling
226 // boundary at the bottom of the region. The DAG does not include RegionEnd,
227 // but the region does (i.e. the next RegionEnd is above the previous
228 // RegionBegin). If the current block has no terminator then RegionEnd ==
229 // MBB->end() for the bottom region.
231 // The Scheduler may insert instructions during either schedule() or
232 // exitRegion(), even for empty regions. So the local iterators 'I' and
233 // 'RegionEnd' are invalid across these calls.
234 unsigned RemainingInstrs = MBB->size();
235 for(MachineBasicBlock::iterator RegionEnd = MBB->end();
236 RegionEnd != MBB->begin(); RegionEnd = Scheduler->begin()) {
238 // Avoid decrementing RegionEnd for blocks with no terminator.
239 if (RegionEnd != MBB->end()
240 || TII->isSchedulingBoundary(llvm::prior(RegionEnd), MBB, *MF)) {
242 // Count the boundary instruction.
246 // The next region starts above the previous region. Look backward in the
247 // instruction stream until we find the nearest boundary.
248 MachineBasicBlock::iterator I = RegionEnd;
249 for(;I != MBB->begin(); --I, --RemainingInstrs) {
250 if (TII->isSchedulingBoundary(llvm::prior(I), MBB, *MF))
253 // Notify the scheduler of the region, even if we may skip scheduling
254 // it. Perhaps it still needs to be bundled.
255 Scheduler->enterRegion(MBB, I, RegionEnd, RemainingInstrs);
257 // Skip empty scheduling regions (0 or 1 schedulable instructions).
258 if (I == RegionEnd || I == llvm::prior(RegionEnd)) {
259 // Close the current region. Bundle the terminator if needed.
260 // This invalidates 'RegionEnd' and 'I'.
261 Scheduler->exitRegion();
264 DEBUG(dbgs() << "********** MI Scheduling **********\n");
265 DEBUG(dbgs() << MF->getName()
266 << ":BB#" << MBB->getNumber() << " " << MBB->getName()
267 << "\n From: " << *I << " To: ";
268 if (RegionEnd != MBB->end()) dbgs() << *RegionEnd;
269 else dbgs() << "End";
270 dbgs() << " Remaining: " << RemainingInstrs << "\n");
272 // Schedule a region: possibly reorder instructions.
273 // This invalidates 'RegionEnd' and 'I'.
274 Scheduler->schedule();
276 // Close the current region.
277 Scheduler->exitRegion();
279 // Scheduling has invalidated the current iterator 'I'. Ask the
280 // scheduler for the top of it's scheduled region.
281 RegionEnd = Scheduler->begin();
283 assert(RemainingInstrs == 0 && "Instruction count mismatch!");
284 Scheduler->finishBlock();
286 Scheduler->finalizeSchedule();
287 DEBUG(LIS->print(dbgs()));
291 void MachineScheduler::print(raw_ostream &O, const Module* m) const {
295 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
296 void ReadyQueue::dump() {
297 dbgs() << Name << ": ";
298 for (unsigned i = 0, e = Queue.size(); i < e; ++i)
299 dbgs() << Queue[i]->NodeNum << " ";
304 //===----------------------------------------------------------------------===//
305 // ScheduleDAGMI - Base class for MachineInstr scheduling with LiveIntervals
307 //===----------------------------------------------------------------------===//
309 ScheduleDAGMI::~ScheduleDAGMI() {
311 DeleteContainerPointers(Mutations);
315 bool ScheduleDAGMI::addEdge(SUnit *SuccSU, const SDep &PredDep) {
316 if (SuccSU != &ExitSU) {
317 // Do not use WillCreateCycle, it assumes SD scheduling.
318 // If Pred is reachable from Succ, then the edge creates a cycle.
319 if (Topo.IsReachable(PredDep.getSUnit(), SuccSU))
321 Topo.AddPred(SuccSU, PredDep.getSUnit());
323 SuccSU->addPred(PredDep, /*Required=*/!PredDep.isArtificial());
324 // Return true regardless of whether a new edge needed to be inserted.
328 /// ReleaseSucc - Decrement the NumPredsLeft count of a successor. When
329 /// NumPredsLeft reaches zero, release the successor node.
331 /// FIXME: Adjust SuccSU height based on MinLatency.
332 void ScheduleDAGMI::releaseSucc(SUnit *SU, SDep *SuccEdge) {
333 SUnit *SuccSU = SuccEdge->getSUnit();
335 if (SuccEdge->isWeak()) {
336 --SuccSU->WeakPredsLeft;
337 if (SuccEdge->isCluster())
338 NextClusterSucc = SuccSU;
342 if (SuccSU->NumPredsLeft == 0) {
343 dbgs() << "*** Scheduling failed! ***\n";
345 dbgs() << " has been released too many times!\n";
349 --SuccSU->NumPredsLeft;
350 if (SuccSU->NumPredsLeft == 0 && SuccSU != &ExitSU)
351 SchedImpl->releaseTopNode(SuccSU);
354 /// releaseSuccessors - Call releaseSucc on each of SU's successors.
355 void ScheduleDAGMI::releaseSuccessors(SUnit *SU) {
356 for (SUnit::succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
358 releaseSucc(SU, &*I);
362 /// ReleasePred - Decrement the NumSuccsLeft count of a predecessor. When
363 /// NumSuccsLeft reaches zero, release the predecessor node.
365 /// FIXME: Adjust PredSU height based on MinLatency.
366 void ScheduleDAGMI::releasePred(SUnit *SU, SDep *PredEdge) {
367 SUnit *PredSU = PredEdge->getSUnit();
369 if (PredEdge->isWeak()) {
370 --PredSU->WeakSuccsLeft;
371 if (PredEdge->isCluster())
372 NextClusterPred = PredSU;
376 if (PredSU->NumSuccsLeft == 0) {
377 dbgs() << "*** Scheduling failed! ***\n";
379 dbgs() << " has been released too many times!\n";
383 --PredSU->NumSuccsLeft;
384 if (PredSU->NumSuccsLeft == 0 && PredSU != &EntrySU)
385 SchedImpl->releaseBottomNode(PredSU);
388 /// releasePredecessors - Call releasePred on each of SU's predecessors.
389 void ScheduleDAGMI::releasePredecessors(SUnit *SU) {
390 for (SUnit::pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
392 releasePred(SU, &*I);
396 void ScheduleDAGMI::moveInstruction(MachineInstr *MI,
397 MachineBasicBlock::iterator InsertPos) {
398 // Advance RegionBegin if the first instruction moves down.
399 if (&*RegionBegin == MI)
402 // Update the instruction stream.
403 BB->splice(InsertPos, BB, MI);
405 // Update LiveIntervals
406 LIS->handleMove(MI, /*UpdateFlags=*/true);
408 // Recede RegionBegin if an instruction moves above the first.
409 if (RegionBegin == InsertPos)
413 bool ScheduleDAGMI::checkSchedLimit() {
415 if (NumInstrsScheduled == MISchedCutoff && MISchedCutoff != ~0U) {
416 CurrentTop = CurrentBottom;
419 ++NumInstrsScheduled;
424 /// enterRegion - Called back from MachineScheduler::runOnMachineFunction after
425 /// crossing a scheduling boundary. [begin, end) includes all instructions in
426 /// the region, including the boundary itself and single-instruction regions
427 /// that don't get scheduled.
428 void ScheduleDAGMI::enterRegion(MachineBasicBlock *bb,
429 MachineBasicBlock::iterator begin,
430 MachineBasicBlock::iterator end,
433 ScheduleDAGInstrs::enterRegion(bb, begin, end, endcount);
435 // For convenience remember the end of the liveness region.
437 (RegionEnd == bb->end()) ? RegionEnd : llvm::next(RegionEnd);
440 // Setup the register pressure trackers for the top scheduled top and bottom
441 // scheduled regions.
442 void ScheduleDAGMI::initRegPressure() {
443 TopRPTracker.init(&MF, RegClassInfo, LIS, BB, RegionBegin);
444 BotRPTracker.init(&MF, RegClassInfo, LIS, BB, LiveRegionEnd);
446 // Close the RPTracker to finalize live ins.
447 RPTracker.closeRegion();
449 DEBUG(RPTracker.getPressure().dump(TRI));
451 // Initialize the live ins and live outs.
452 TopRPTracker.addLiveRegs(RPTracker.getPressure().LiveInRegs);
453 BotRPTracker.addLiveRegs(RPTracker.getPressure().LiveOutRegs);
455 // Close one end of the tracker so we can call
456 // getMaxUpward/DownwardPressureDelta before advancing across any
457 // instructions. This converts currently live regs into live ins/outs.
458 TopRPTracker.closeTop();
459 BotRPTracker.closeBottom();
461 // Account for liveness generated by the region boundary.
462 if (LiveRegionEnd != RegionEnd)
463 BotRPTracker.recede();
465 assert(BotRPTracker.getPos() == RegionEnd && "Can't find the region bottom");
467 // Cache the list of excess pressure sets in this region. This will also track
468 // the max pressure in the scheduled code for these sets.
469 RegionCriticalPSets.clear();
470 const std::vector<unsigned> &RegionPressure =
471 RPTracker.getPressure().MaxSetPressure;
472 for (unsigned i = 0, e = RegionPressure.size(); i < e; ++i) {
473 unsigned Limit = TRI->getRegPressureSetLimit(i);
474 DEBUG(dbgs() << TRI->getRegPressureSetName(i)
476 << " Actual " << RegionPressure[i] << "\n");
477 if (RegionPressure[i] > Limit)
478 RegionCriticalPSets.push_back(PressureElement(i, 0));
480 DEBUG(dbgs() << "Excess PSets: ";
481 for (unsigned i = 0, e = RegionCriticalPSets.size(); i != e; ++i)
482 dbgs() << TRI->getRegPressureSetName(
483 RegionCriticalPSets[i].PSetID) << " ";
487 // FIXME: When the pressure tracker deals in pressure differences then we won't
488 // iterate over all RegionCriticalPSets[i].
490 updateScheduledPressure(std::vector<unsigned> NewMaxPressure) {
491 for (unsigned i = 0, e = RegionCriticalPSets.size(); i < e; ++i) {
492 unsigned ID = RegionCriticalPSets[i].PSetID;
493 int &MaxUnits = RegionCriticalPSets[i].UnitIncrease;
494 if ((int)NewMaxPressure[ID] > MaxUnits)
495 MaxUnits = NewMaxPressure[ID];
499 /// schedule - Called back from MachineScheduler::runOnMachineFunction
500 /// after setting up the current scheduling region. [RegionBegin, RegionEnd)
501 /// only includes instructions that have DAG nodes, not scheduling boundaries.
503 /// This is a skeletal driver, with all the functionality pushed into helpers,
504 /// so that it can be easilly extended by experimental schedulers. Generally,
505 /// implementing MachineSchedStrategy should be sufficient to implement a new
506 /// scheduling algorithm. However, if a scheduler further subclasses
507 /// ScheduleDAGMI then it will want to override this virtual method in order to
508 /// update any specialized state.
509 void ScheduleDAGMI::schedule() {
510 buildDAGWithRegPressure();
512 Topo.InitDAGTopologicalSorting();
516 SmallVector<SUnit*, 8> TopRoots, BotRoots;
517 findRootsAndBiasEdges(TopRoots, BotRoots);
519 // Initialize the strategy before modifying the DAG.
520 // This may initialize a DFSResult to be used for queue priority.
521 SchedImpl->initialize(this);
523 DEBUG(for (unsigned su = 0, e = SUnits.size(); su != e; ++su)
524 SUnits[su].dumpAll(this));
525 if (ViewMISchedDAGs) viewGraph();
527 // Initialize ready queues now that the DAG and priority data are finalized.
528 initQueues(TopRoots, BotRoots);
530 bool IsTopNode = false;
531 while (SUnit *SU = SchedImpl->pickNode(IsTopNode)) {
532 assert(!SU->isScheduled && "Node already scheduled");
533 if (!checkSchedLimit())
536 scheduleMI(SU, IsTopNode);
538 updateQueues(SU, IsTopNode);
540 assert(CurrentTop == CurrentBottom && "Nonempty unscheduled zone.");
545 unsigned BBNum = begin()->getParent()->getNumber();
546 dbgs() << "*** Final schedule for BB#" << BBNum << " ***\n";
552 /// Build the DAG and setup three register pressure trackers.
553 void ScheduleDAGMI::buildDAGWithRegPressure() {
554 // Initialize the register pressure tracker used by buildSchedGraph.
555 RPTracker.init(&MF, RegClassInfo, LIS, BB, LiveRegionEnd);
557 // Account for liveness generate by the region boundary.
558 if (LiveRegionEnd != RegionEnd)
561 // Build the DAG, and compute current register pressure.
562 buildSchedGraph(AA, &RPTracker);
564 // Initialize top/bottom trackers after computing region pressure.
568 /// Apply each ScheduleDAGMutation step in order.
569 void ScheduleDAGMI::postprocessDAG() {
570 for (unsigned i = 0, e = Mutations.size(); i < e; ++i) {
571 Mutations[i]->apply(this);
575 void ScheduleDAGMI::computeDFSResult() {
577 DFSResult = new SchedDFSResult(/*BottomU*/true, MinSubtreeSize);
579 ScheduledTrees.clear();
580 DFSResult->resize(SUnits.size());
581 DFSResult->compute(SUnits);
582 ScheduledTrees.resize(DFSResult->getNumSubtrees());
585 void ScheduleDAGMI::findRootsAndBiasEdges(SmallVectorImpl<SUnit*> &TopRoots,
586 SmallVectorImpl<SUnit*> &BotRoots) {
587 for (std::vector<SUnit>::iterator
588 I = SUnits.begin(), E = SUnits.end(); I != E; ++I) {
591 // Order predecessors so DFSResult follows the critical path.
592 SU->biasCriticalPath();
594 // A SUnit is ready to top schedule if it has no predecessors.
595 if (!I->NumPredsLeft && SU != &EntrySU)
596 TopRoots.push_back(SU);
597 // A SUnit is ready to bottom schedule if it has no successors.
598 if (!I->NumSuccsLeft && SU != &ExitSU)
599 BotRoots.push_back(SU);
603 /// Identify DAG roots and setup scheduler queues.
604 void ScheduleDAGMI::initQueues(ArrayRef<SUnit*> TopRoots,
605 ArrayRef<SUnit*> BotRoots) {
606 NextClusterSucc = NULL;
607 NextClusterPred = NULL;
609 // Release all DAG roots for scheduling, not including EntrySU/ExitSU.
611 // Nodes with unreleased weak edges can still be roots.
612 // Release top roots in forward order.
613 for (SmallVectorImpl<SUnit*>::const_iterator
614 I = TopRoots.begin(), E = TopRoots.end(); I != E; ++I) {
615 SchedImpl->releaseTopNode(*I);
617 // Release bottom roots in reverse order so the higher priority nodes appear
618 // first. This is more natural and slightly more efficient.
619 for (SmallVectorImpl<SUnit*>::const_reverse_iterator
620 I = BotRoots.rbegin(), E = BotRoots.rend(); I != E; ++I) {
621 SchedImpl->releaseBottomNode(*I);
624 releaseSuccessors(&EntrySU);
625 releasePredecessors(&ExitSU);
627 SchedImpl->registerRoots();
629 // Advance past initial DebugValues.
630 assert(TopRPTracker.getPos() == RegionBegin && "bad initial Top tracker");
631 CurrentTop = nextIfDebug(RegionBegin, RegionEnd);
632 TopRPTracker.setPos(CurrentTop);
634 CurrentBottom = RegionEnd;
637 /// Move an instruction and update register pressure.
638 void ScheduleDAGMI::scheduleMI(SUnit *SU, bool IsTopNode) {
639 // Move the instruction to its new location in the instruction stream.
640 MachineInstr *MI = SU->getInstr();
643 assert(SU->isTopReady() && "node still has unscheduled dependencies");
644 if (&*CurrentTop == MI)
645 CurrentTop = nextIfDebug(++CurrentTop, CurrentBottom);
647 moveInstruction(MI, CurrentTop);
648 TopRPTracker.setPos(MI);
651 // Update top scheduled pressure.
652 TopRPTracker.advance();
653 assert(TopRPTracker.getPos() == CurrentTop && "out of sync");
654 updateScheduledPressure(TopRPTracker.getPressure().MaxSetPressure);
657 assert(SU->isBottomReady() && "node still has unscheduled dependencies");
658 MachineBasicBlock::iterator priorII =
659 priorNonDebug(CurrentBottom, CurrentTop);
661 CurrentBottom = priorII;
663 if (&*CurrentTop == MI) {
664 CurrentTop = nextIfDebug(++CurrentTop, priorII);
665 TopRPTracker.setPos(CurrentTop);
667 moveInstruction(MI, CurrentBottom);
670 // Update bottom scheduled pressure.
671 BotRPTracker.recede();
672 assert(BotRPTracker.getPos() == CurrentBottom && "out of sync");
673 updateScheduledPressure(BotRPTracker.getPressure().MaxSetPressure);
677 /// Update scheduler queues after scheduling an instruction.
678 void ScheduleDAGMI::updateQueues(SUnit *SU, bool IsTopNode) {
679 // Release dependent instructions for scheduling.
681 releaseSuccessors(SU);
683 releasePredecessors(SU);
685 SU->isScheduled = true;
688 unsigned SubtreeID = DFSResult->getSubtreeID(SU);
689 if (!ScheduledTrees.test(SubtreeID)) {
690 ScheduledTrees.set(SubtreeID);
691 DFSResult->scheduleTree(SubtreeID);
692 SchedImpl->scheduleTree(SubtreeID);
696 // Notify the scheduling strategy after updating the DAG.
697 SchedImpl->schedNode(SU, IsTopNode);
700 /// Reinsert any remaining debug_values, just like the PostRA scheduler.
701 void ScheduleDAGMI::placeDebugValues() {
702 // If first instruction was a DBG_VALUE then put it back.
704 BB->splice(RegionBegin, BB, FirstDbgValue);
705 RegionBegin = FirstDbgValue;
708 for (std::vector<std::pair<MachineInstr *, MachineInstr *> >::iterator
709 DI = DbgValues.end(), DE = DbgValues.begin(); DI != DE; --DI) {
710 std::pair<MachineInstr *, MachineInstr *> P = *prior(DI);
711 MachineInstr *DbgValue = P.first;
712 MachineBasicBlock::iterator OrigPrevMI = P.second;
713 if (&*RegionBegin == DbgValue)
715 BB->splice(++OrigPrevMI, BB, DbgValue);
716 if (OrigPrevMI == llvm::prior(RegionEnd))
717 RegionEnd = DbgValue;
720 FirstDbgValue = NULL;
723 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
724 void ScheduleDAGMI::dumpSchedule() const {
725 for (MachineBasicBlock::iterator MI = begin(), ME = end(); MI != ME; ++MI) {
726 if (SUnit *SU = getSUnit(&(*MI)))
729 dbgs() << "Missing SUnit\n";
734 //===----------------------------------------------------------------------===//
735 // LoadClusterMutation - DAG post-processing to cluster loads.
736 //===----------------------------------------------------------------------===//
739 /// \brief Post-process the DAG to create cluster edges between neighboring
741 class LoadClusterMutation : public ScheduleDAGMutation {
746 LoadInfo(SUnit *su, unsigned reg, unsigned ofs)
747 : SU(su), BaseReg(reg), Offset(ofs) {}
749 static bool LoadInfoLess(const LoadClusterMutation::LoadInfo &LHS,
750 const LoadClusterMutation::LoadInfo &RHS);
752 const TargetInstrInfo *TII;
753 const TargetRegisterInfo *TRI;
755 LoadClusterMutation(const TargetInstrInfo *tii,
756 const TargetRegisterInfo *tri)
757 : TII(tii), TRI(tri) {}
759 virtual void apply(ScheduleDAGMI *DAG);
761 void clusterNeighboringLoads(ArrayRef<SUnit*> Loads, ScheduleDAGMI *DAG);
765 bool LoadClusterMutation::LoadInfoLess(
766 const LoadClusterMutation::LoadInfo &LHS,
767 const LoadClusterMutation::LoadInfo &RHS) {
768 if (LHS.BaseReg != RHS.BaseReg)
769 return LHS.BaseReg < RHS.BaseReg;
770 return LHS.Offset < RHS.Offset;
773 void LoadClusterMutation::clusterNeighboringLoads(ArrayRef<SUnit*> Loads,
774 ScheduleDAGMI *DAG) {
775 SmallVector<LoadClusterMutation::LoadInfo,32> LoadRecords;
776 for (unsigned Idx = 0, End = Loads.size(); Idx != End; ++Idx) {
777 SUnit *SU = Loads[Idx];
780 if (TII->getLdStBaseRegImmOfs(SU->getInstr(), BaseReg, Offset, TRI))
781 LoadRecords.push_back(LoadInfo(SU, BaseReg, Offset));
783 if (LoadRecords.size() < 2)
785 std::sort(LoadRecords.begin(), LoadRecords.end(), LoadInfoLess);
786 unsigned ClusterLength = 1;
787 for (unsigned Idx = 0, End = LoadRecords.size(); Idx < (End - 1); ++Idx) {
788 if (LoadRecords[Idx].BaseReg != LoadRecords[Idx+1].BaseReg) {
793 SUnit *SUa = LoadRecords[Idx].SU;
794 SUnit *SUb = LoadRecords[Idx+1].SU;
795 if (TII->shouldClusterLoads(SUa->getInstr(), SUb->getInstr(), ClusterLength)
796 && DAG->addEdge(SUb, SDep(SUa, SDep::Cluster))) {
798 DEBUG(dbgs() << "Cluster loads SU(" << SUa->NodeNum << ") - SU("
799 << SUb->NodeNum << ")\n");
800 // Copy successor edges from SUa to SUb. Interleaving computation
801 // dependent on SUa can prevent load combining due to register reuse.
802 // Predecessor edges do not need to be copied from SUb to SUa since nearby
803 // loads should have effectively the same inputs.
804 for (SUnit::const_succ_iterator
805 SI = SUa->Succs.begin(), SE = SUa->Succs.end(); SI != SE; ++SI) {
806 if (SI->getSUnit() == SUb)
808 DEBUG(dbgs() << " Copy Succ SU(" << SI->getSUnit()->NodeNum << ")\n");
809 DAG->addEdge(SI->getSUnit(), SDep(SUb, SDep::Artificial));
818 /// \brief Callback from DAG postProcessing to create cluster edges for loads.
819 void LoadClusterMutation::apply(ScheduleDAGMI *DAG) {
820 // Map DAG NodeNum to store chain ID.
821 DenseMap<unsigned, unsigned> StoreChainIDs;
822 // Map each store chain to a set of dependent loads.
823 SmallVector<SmallVector<SUnit*,4>, 32> StoreChainDependents;
824 for (unsigned Idx = 0, End = DAG->SUnits.size(); Idx != End; ++Idx) {
825 SUnit *SU = &DAG->SUnits[Idx];
826 if (!SU->getInstr()->mayLoad())
828 unsigned ChainPredID = DAG->SUnits.size();
829 for (SUnit::const_pred_iterator
830 PI = SU->Preds.begin(), PE = SU->Preds.end(); PI != PE; ++PI) {
832 ChainPredID = PI->getSUnit()->NodeNum;
836 // Check if this chain-like pred has been seen
837 // before. ChainPredID==MaxNodeID for loads at the top of the schedule.
838 unsigned NumChains = StoreChainDependents.size();
839 std::pair<DenseMap<unsigned, unsigned>::iterator, bool> Result =
840 StoreChainIDs.insert(std::make_pair(ChainPredID, NumChains));
842 StoreChainDependents.resize(NumChains + 1);
843 StoreChainDependents[Result.first->second].push_back(SU);
845 // Iterate over the store chains.
846 for (unsigned Idx = 0, End = StoreChainDependents.size(); Idx != End; ++Idx)
847 clusterNeighboringLoads(StoreChainDependents[Idx], DAG);
850 //===----------------------------------------------------------------------===//
851 // MacroFusion - DAG post-processing to encourage fusion of macro ops.
852 //===----------------------------------------------------------------------===//
855 /// \brief Post-process the DAG to create cluster edges between instructions
856 /// that may be fused by the processor into a single operation.
857 class MacroFusion : public ScheduleDAGMutation {
858 const TargetInstrInfo *TII;
860 MacroFusion(const TargetInstrInfo *tii): TII(tii) {}
862 virtual void apply(ScheduleDAGMI *DAG);
866 /// \brief Callback from DAG postProcessing to create cluster edges to encourage
867 /// fused operations.
868 void MacroFusion::apply(ScheduleDAGMI *DAG) {
869 // For now, assume targets can only fuse with the branch.
870 MachineInstr *Branch = DAG->ExitSU.getInstr();
874 for (unsigned Idx = DAG->SUnits.size(); Idx > 0;) {
875 SUnit *SU = &DAG->SUnits[--Idx];
876 if (!TII->shouldScheduleAdjacent(SU->getInstr(), Branch))
879 // Create a single weak edge from SU to ExitSU. The only effect is to cause
880 // bottom-up scheduling to heavily prioritize the clustered SU. There is no
881 // need to copy predecessor edges from ExitSU to SU, since top-down
882 // scheduling cannot prioritize ExitSU anyway. To defer top-down scheduling
883 // of SU, we could create an artificial edge from the deepest root, but it
884 // hasn't been needed yet.
885 bool Success = DAG->addEdge(&DAG->ExitSU, SDep(SU, SDep::Cluster));
887 assert(Success && "No DAG nodes should be reachable from ExitSU");
889 DEBUG(dbgs() << "Macro Fuse SU(" << SU->NodeNum << ")\n");
894 //===----------------------------------------------------------------------===//
895 // ConvergingScheduler - Implementation of the standard MachineSchedStrategy.
896 //===----------------------------------------------------------------------===//
899 /// ConvergingScheduler shrinks the unscheduled zone using heuristics to balance
901 class ConvergingScheduler : public MachineSchedStrategy {
903 /// Represent the type of SchedCandidate found within a single queue.
904 /// pickNodeBidirectional depends on these listed by decreasing priority.
906 NoCand, SingleExcess, SingleCritical, Cluster,
907 ResourceReduce, ResourceDemand, BotHeightReduce, BotPathReduce,
908 TopDepthReduce, TopPathReduce, SingleMax, MultiPressure, NextDefUse,
912 static const char *getReasonStr(ConvergingScheduler::CandReason Reason);
915 /// Policy for scheduling the next instruction in the candidate's zone.
918 unsigned ReduceResIdx;
919 unsigned DemandResIdx;
921 CandPolicy(): ReduceLatency(false), ReduceResIdx(0), DemandResIdx(0) {}
924 /// Status of an instruction's critical resource consumption.
925 struct SchedResourceDelta {
926 // Count critical resources in the scheduled region required by SU.
927 unsigned CritResources;
929 // Count critical resources from another region consumed by SU.
930 unsigned DemandedResources;
932 SchedResourceDelta(): CritResources(0), DemandedResources(0) {}
934 bool operator==(const SchedResourceDelta &RHS) const {
935 return CritResources == RHS.CritResources
936 && DemandedResources == RHS.DemandedResources;
938 bool operator!=(const SchedResourceDelta &RHS) const {
939 return !operator==(RHS);
943 /// Store the state used by ConvergingScheduler heuristics, required for the
944 /// lifetime of one invocation of pickNode().
945 struct SchedCandidate {
948 // The best SUnit candidate.
951 // The reason for this candidate.
954 // Register pressure values for the best candidate.
955 RegPressureDelta RPDelta;
957 // Critical resource consumption of the best candidate.
958 SchedResourceDelta ResDelta;
960 SchedCandidate(const CandPolicy &policy)
961 : Policy(policy), SU(NULL), Reason(NoCand) {}
963 bool isValid() const { return SU; }
965 // Copy the status of another candidate without changing policy.
966 void setBest(SchedCandidate &Best) {
967 assert(Best.Reason != NoCand && "uninitialized Sched candidate");
969 Reason = Best.Reason;
970 RPDelta = Best.RPDelta;
971 ResDelta = Best.ResDelta;
974 void initResourceDelta(const ScheduleDAGMI *DAG,
975 const TargetSchedModel *SchedModel);
978 /// Summarize the unscheduled region.
979 struct SchedRemainder {
980 // Critical path through the DAG in expected latency.
981 unsigned CriticalPath;
983 // Unscheduled resources
984 SmallVector<unsigned, 16> RemainingCounts;
985 // Critical resource for the unscheduled zone.
987 // Number of micro-ops left to schedule.
988 unsigned RemainingMicroOps;
992 RemainingCounts.clear();
994 RemainingMicroOps = 0;
997 SchedRemainder() { reset(); }
999 void init(ScheduleDAGMI *DAG, const TargetSchedModel *SchedModel);
1001 unsigned getMaxRemainingCount(const TargetSchedModel *SchedModel) const {
1002 if (!SchedModel->hasInstrSchedModel())
1006 RemainingMicroOps * SchedModel->getMicroOpFactor(),
1007 RemainingCounts[CritResIdx]);
1011 /// Each Scheduling boundary is associated with ready queues. It tracks the
1012 /// current cycle in the direction of movement, and maintains the state
1013 /// of "hazards" and other interlocks at the current cycle.
1014 struct SchedBoundary {
1016 const TargetSchedModel *SchedModel;
1017 SchedRemainder *Rem;
1019 ReadyQueue Available;
1023 // For heuristics, keep a list of the nodes that immediately depend on the
1024 // most recently scheduled node.
1025 SmallPtrSet<const SUnit*, 8> NextSUs;
1027 ScheduleHazardRecognizer *HazardRec;
1030 unsigned IssueCount;
1032 /// MinReadyCycle - Cycle of the soonest available instruction.
1033 unsigned MinReadyCycle;
1035 // The expected latency of the critical path in this scheduled zone.
1036 unsigned ExpectedLatency;
1038 // Resources used in the scheduled zone beyond this boundary.
1039 SmallVector<unsigned, 16> ResourceCounts;
1041 // Cache the critical resources ID in this scheduled zone.
1042 unsigned CritResIdx;
1044 // Is the scheduled region resource limited vs. latency limited.
1045 bool IsResourceLimited;
1047 unsigned ExpectedCount;
1050 // Remember the greatest min operand latency.
1051 unsigned MaxMinLatency;
1057 CheckPending = false;
1062 MinReadyCycle = UINT_MAX;
1063 ExpectedLatency = 0;
1064 ResourceCounts.resize(1);
1065 assert(!ResourceCounts[0] && "nonzero count for bad resource");
1067 IsResourceLimited = false;
1072 // Reserve a zero-count for invalid CritResIdx.
1073 ResourceCounts.resize(1);
1076 /// Pending queues extend the ready queues with the same ID and the
1077 /// PendingFlag set.
1078 SchedBoundary(unsigned ID, const Twine &Name):
1079 DAG(0), SchedModel(0), Rem(0), Available(ID, Name+".A"),
1080 Pending(ID << ConvergingScheduler::LogMaxQID, Name+".P") {
1084 ~SchedBoundary() { delete HazardRec; }
1086 void init(ScheduleDAGMI *dag, const TargetSchedModel *smodel,
1087 SchedRemainder *rem);
1089 bool isTop() const {
1090 return Available.getID() == ConvergingScheduler::TopQID;
1093 unsigned getUnscheduledLatency(SUnit *SU) const {
1095 return SU->getHeight();
1096 return SU->getDepth() + SU->Latency;
1099 unsigned getCriticalCount() const {
1100 return ResourceCounts[CritResIdx];
1103 bool checkHazard(SUnit *SU);
1105 void setLatencyPolicy(CandPolicy &Policy);
1107 void releaseNode(SUnit *SU, unsigned ReadyCycle);
1111 void countResource(unsigned PIdx, unsigned Cycles);
1113 void bumpNode(SUnit *SU);
1115 void releasePending();
1117 void removeReady(SUnit *SU);
1119 SUnit *pickOnlyChoice();
1124 const TargetSchedModel *SchedModel;
1125 const TargetRegisterInfo *TRI;
1127 // State of the top and bottom scheduled instruction boundaries.
1133 /// SUnit::NodeQueueId: 0 (none), 1 (top), 2 (bot), 3 (both)
1140 ConvergingScheduler():
1141 DAG(0), SchedModel(0), TRI(0), Top(TopQID, "TopQ"), Bot(BotQID, "BotQ") {}
1143 virtual void initialize(ScheduleDAGMI *dag);
1145 virtual SUnit *pickNode(bool &IsTopNode);
1147 virtual void schedNode(SUnit *SU, bool IsTopNode);
1149 virtual void releaseTopNode(SUnit *SU);
1151 virtual void releaseBottomNode(SUnit *SU);
1153 virtual void registerRoots();
1157 ConvergingScheduler::SchedBoundary &CriticalZone,
1158 ConvergingScheduler::SchedCandidate &CriticalCand,
1159 ConvergingScheduler::SchedBoundary &OppositeZone,
1160 ConvergingScheduler::SchedCandidate &OppositeCand);
1162 void checkResourceLimits(ConvergingScheduler::SchedCandidate &TopCand,
1163 ConvergingScheduler::SchedCandidate &BotCand);
1165 void tryCandidate(SchedCandidate &Cand,
1166 SchedCandidate &TryCand,
1167 SchedBoundary &Zone,
1168 const RegPressureTracker &RPTracker,
1169 RegPressureTracker &TempTracker);
1171 SUnit *pickNodeBidirectional(bool &IsTopNode);
1173 void pickNodeFromQueue(SchedBoundary &Zone,
1174 const RegPressureTracker &RPTracker,
1175 SchedCandidate &Candidate);
1178 void traceCandidate(const SchedCandidate &Cand, const SchedBoundary &Zone);
1183 void ConvergingScheduler::SchedRemainder::
1184 init(ScheduleDAGMI *DAG, const TargetSchedModel *SchedModel) {
1186 if (!SchedModel->hasInstrSchedModel())
1188 RemainingCounts.resize(SchedModel->getNumProcResourceKinds());
1189 for (std::vector<SUnit>::iterator
1190 I = DAG->SUnits.begin(), E = DAG->SUnits.end(); I != E; ++I) {
1191 const MCSchedClassDesc *SC = DAG->getSchedClass(&*I);
1192 RemainingMicroOps += SchedModel->getNumMicroOps(I->getInstr(), SC);
1193 for (TargetSchedModel::ProcResIter
1194 PI = SchedModel->getWriteProcResBegin(SC),
1195 PE = SchedModel->getWriteProcResEnd(SC); PI != PE; ++PI) {
1196 unsigned PIdx = PI->ProcResourceIdx;
1197 unsigned Factor = SchedModel->getResourceFactor(PIdx);
1198 RemainingCounts[PIdx] += (Factor * PI->Cycles);
1201 for (unsigned PIdx = 0, PEnd = SchedModel->getNumProcResourceKinds();
1202 PIdx != PEnd; ++PIdx) {
1203 if ((int)(RemainingCounts[PIdx] - RemainingCounts[CritResIdx])
1204 >= (int)SchedModel->getLatencyFactor()) {
1210 void ConvergingScheduler::SchedBoundary::
1211 init(ScheduleDAGMI *dag, const TargetSchedModel *smodel, SchedRemainder *rem) {
1214 SchedModel = smodel;
1216 if (SchedModel->hasInstrSchedModel())
1217 ResourceCounts.resize(SchedModel->getNumProcResourceKinds());
1220 void ConvergingScheduler::initialize(ScheduleDAGMI *dag) {
1222 SchedModel = DAG->getSchedModel();
1224 Rem.init(DAG, SchedModel);
1225 Top.init(DAG, SchedModel, &Rem);
1226 Bot.init(DAG, SchedModel, &Rem);
1228 DAG->computeDFSResult();
1230 // Initialize resource counts.
1232 // Initialize the HazardRecognizers. If itineraries don't exist, are empty, or
1233 // are disabled, then these HazardRecs will be disabled.
1234 const InstrItineraryData *Itin = SchedModel->getInstrItineraries();
1235 const TargetMachine &TM = DAG->MF.getTarget();
1236 Top.HazardRec = TM.getInstrInfo()->CreateTargetMIHazardRecognizer(Itin, DAG);
1237 Bot.HazardRec = TM.getInstrInfo()->CreateTargetMIHazardRecognizer(Itin, DAG);
1239 assert((!ForceTopDown || !ForceBottomUp) &&
1240 "-misched-topdown incompatible with -misched-bottomup");
1243 void ConvergingScheduler::releaseTopNode(SUnit *SU) {
1244 if (SU->isScheduled)
1247 for (SUnit::pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
1249 unsigned PredReadyCycle = I->getSUnit()->TopReadyCycle;
1250 unsigned MinLatency = I->getMinLatency();
1252 Top.MaxMinLatency = std::max(MinLatency, Top.MaxMinLatency);
1254 if (SU->TopReadyCycle < PredReadyCycle + MinLatency)
1255 SU->TopReadyCycle = PredReadyCycle + MinLatency;
1257 Top.releaseNode(SU, SU->TopReadyCycle);
1260 void ConvergingScheduler::releaseBottomNode(SUnit *SU) {
1261 if (SU->isScheduled)
1264 assert(SU->getInstr() && "Scheduled SUnit must have instr");
1266 for (SUnit::succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
1270 unsigned SuccReadyCycle = I->getSUnit()->BotReadyCycle;
1271 unsigned MinLatency = I->getMinLatency();
1273 Bot.MaxMinLatency = std::max(MinLatency, Bot.MaxMinLatency);
1275 if (SU->BotReadyCycle < SuccReadyCycle + MinLatency)
1276 SU->BotReadyCycle = SuccReadyCycle + MinLatency;
1278 Bot.releaseNode(SU, SU->BotReadyCycle);
1281 void ConvergingScheduler::registerRoots() {
1282 Rem.CriticalPath = DAG->ExitSU.getDepth();
1283 // Some roots may not feed into ExitSU. Check all of them in case.
1284 for (std::vector<SUnit*>::const_iterator
1285 I = Bot.Available.begin(), E = Bot.Available.end(); I != E; ++I) {
1286 if ((*I)->getDepth() > Rem.CriticalPath)
1287 Rem.CriticalPath = (*I)->getDepth();
1289 DEBUG(dbgs() << "Critical Path: " << Rem.CriticalPath << '\n');
1292 /// Does this SU have a hazard within the current instruction group.
1294 /// The scheduler supports two modes of hazard recognition. The first is the
1295 /// ScheduleHazardRecognizer API. It is a fully general hazard recognizer that
1296 /// supports highly complicated in-order reservation tables
1297 /// (ScoreboardHazardRecognizer) and arbitraty target-specific logic.
1299 /// The second is a streamlined mechanism that checks for hazards based on
1300 /// simple counters that the scheduler itself maintains. It explicitly checks
1301 /// for instruction dispatch limitations, including the number of micro-ops that
1302 /// can dispatch per cycle.
1304 /// TODO: Also check whether the SU must start a new group.
1305 bool ConvergingScheduler::SchedBoundary::checkHazard(SUnit *SU) {
1306 if (HazardRec->isEnabled())
1307 return HazardRec->getHazardType(SU) != ScheduleHazardRecognizer::NoHazard;
1309 unsigned uops = SchedModel->getNumMicroOps(SU->getInstr());
1310 if ((IssueCount > 0) && (IssueCount + uops > SchedModel->getIssueWidth())) {
1311 DEBUG(dbgs() << " SU(" << SU->NodeNum << ") uops="
1312 << SchedModel->getNumMicroOps(SU->getInstr()) << '\n');
1318 /// Compute the remaining latency to determine whether ILP should be increased.
1319 void ConvergingScheduler::SchedBoundary::setLatencyPolicy(CandPolicy &Policy) {
1320 // FIXME: compile time. In all, we visit four queues here one we should only
1321 // need to visit the one that was last popped if we cache the result.
1322 unsigned RemLatency = 0;
1323 for (ReadyQueue::iterator I = Available.begin(), E = Available.end();
1325 unsigned L = getUnscheduledLatency(*I);
1329 for (ReadyQueue::iterator I = Pending.begin(), E = Pending.end();
1331 unsigned L = getUnscheduledLatency(*I);
1335 unsigned CriticalPathLimit = Rem->CriticalPath + SchedModel->getILPWindow();
1336 if (RemLatency + ExpectedLatency >= CriticalPathLimit
1337 && RemLatency > Rem->getMaxRemainingCount(SchedModel)) {
1338 Policy.ReduceLatency = true;
1339 DEBUG(dbgs() << "Increase ILP: " << Available.getName() << '\n');
1343 void ConvergingScheduler::SchedBoundary::releaseNode(SUnit *SU,
1344 unsigned ReadyCycle) {
1346 if (ReadyCycle < MinReadyCycle)
1347 MinReadyCycle = ReadyCycle;
1349 // Check for interlocks first. For the purpose of other heuristics, an
1350 // instruction that cannot issue appears as if it's not in the ReadyQueue.
1351 if (ReadyCycle > CurrCycle || checkHazard(SU))
1356 // Record this node as an immediate dependent of the scheduled node.
1360 /// Move the boundary of scheduled code by one cycle.
1361 void ConvergingScheduler::SchedBoundary::bumpCycle() {
1362 unsigned Width = SchedModel->getIssueWidth();
1363 IssueCount = (IssueCount <= Width) ? 0 : IssueCount - Width;
1365 unsigned NextCycle = CurrCycle + 1;
1366 assert(MinReadyCycle < UINT_MAX && "MinReadyCycle uninitialized");
1367 if (MinReadyCycle > NextCycle) {
1369 NextCycle = MinReadyCycle;
1372 if (!HazardRec->isEnabled()) {
1373 // Bypass HazardRec virtual calls.
1374 CurrCycle = NextCycle;
1377 // Bypass getHazardType calls in case of long latency.
1378 for (; CurrCycle != NextCycle; ++CurrCycle) {
1380 HazardRec->AdvanceCycle();
1382 HazardRec->RecedeCycle();
1385 CheckPending = true;
1386 IsResourceLimited = getCriticalCount() > std::max(ExpectedLatency, CurrCycle);
1388 DEBUG(dbgs() << " *** " << Available.getName() << " cycle "
1389 << CurrCycle << '\n');
1392 /// Add the given processor resource to this scheduled zone.
1393 void ConvergingScheduler::SchedBoundary::countResource(unsigned PIdx,
1395 unsigned Factor = SchedModel->getResourceFactor(PIdx);
1396 DEBUG(dbgs() << " " << SchedModel->getProcResource(PIdx)->Name
1397 << " +(" << Cycles << "x" << Factor
1398 << ") / " << SchedModel->getLatencyFactor() << '\n');
1400 unsigned Count = Factor * Cycles;
1401 ResourceCounts[PIdx] += Count;
1402 assert(Rem->RemainingCounts[PIdx] >= Count && "resource double counted");
1403 Rem->RemainingCounts[PIdx] -= Count;
1405 // Check if this resource exceeds the current critical resource by a full
1406 // cycle. If so, it becomes the critical resource.
1407 if ((int)(ResourceCounts[PIdx] - ResourceCounts[CritResIdx])
1408 >= (int)SchedModel->getLatencyFactor()) {
1410 DEBUG(dbgs() << " *** Critical resource "
1411 << SchedModel->getProcResource(PIdx)->Name << " x"
1412 << ResourceCounts[PIdx] << '\n');
1416 /// Move the boundary of scheduled code by one SUnit.
1417 void ConvergingScheduler::SchedBoundary::bumpNode(SUnit *SU) {
1418 // Update the reservation table.
1419 if (HazardRec->isEnabled()) {
1420 if (!isTop() && SU->isCall) {
1421 // Calls are scheduled with their preceding instructions. For bottom-up
1422 // scheduling, clear the pipeline state before emitting.
1425 HazardRec->EmitInstruction(SU);
1427 // Update resource counts and critical resource.
1428 if (SchedModel->hasInstrSchedModel()) {
1429 const MCSchedClassDesc *SC = DAG->getSchedClass(SU);
1430 Rem->RemainingMicroOps -= SchedModel->getNumMicroOps(SU->getInstr(), SC);
1431 for (TargetSchedModel::ProcResIter
1432 PI = SchedModel->getWriteProcResBegin(SC),
1433 PE = SchedModel->getWriteProcResEnd(SC); PI != PE; ++PI) {
1434 countResource(PI->ProcResourceIdx, PI->Cycles);
1438 if (SU->getDepth() > ExpectedLatency)
1439 ExpectedLatency = SU->getDepth();
1442 if (SU->getHeight() > ExpectedLatency)
1443 ExpectedLatency = SU->getHeight();
1446 IsResourceLimited = getCriticalCount() > std::max(ExpectedLatency, CurrCycle);
1448 // Check the instruction group dispatch limit.
1449 // TODO: Check if this SU must end a dispatch group.
1450 IssueCount += SchedModel->getNumMicroOps(SU->getInstr());
1452 // checkHazard prevents scheduling multiple instructions per cycle that exceed
1453 // issue width. However, we commonly reach the maximum. In this case
1454 // opportunistically bump the cycle to avoid uselessly checking everything in
1455 // the readyQ. Furthermore, a single instruction may produce more than one
1456 // cycle's worth of micro-ops.
1457 if (IssueCount >= SchedModel->getIssueWidth()) {
1458 DEBUG(dbgs() << " *** Max instrs at cycle " << CurrCycle << '\n');
1463 /// Release pending ready nodes in to the available queue. This makes them
1464 /// visible to heuristics.
1465 void ConvergingScheduler::SchedBoundary::releasePending() {
1466 // If the available queue is empty, it is safe to reset MinReadyCycle.
1467 if (Available.empty())
1468 MinReadyCycle = UINT_MAX;
1470 // Check to see if any of the pending instructions are ready to issue. If
1471 // so, add them to the available queue.
1472 for (unsigned i = 0, e = Pending.size(); i != e; ++i) {
1473 SUnit *SU = *(Pending.begin()+i);
1474 unsigned ReadyCycle = isTop() ? SU->TopReadyCycle : SU->BotReadyCycle;
1476 if (ReadyCycle < MinReadyCycle)
1477 MinReadyCycle = ReadyCycle;
1479 if (ReadyCycle > CurrCycle)
1482 if (checkHazard(SU))
1486 Pending.remove(Pending.begin()+i);
1489 DEBUG(if (!Pending.empty()) Pending.dump());
1490 CheckPending = false;
1493 /// Remove SU from the ready set for this boundary.
1494 void ConvergingScheduler::SchedBoundary::removeReady(SUnit *SU) {
1495 if (Available.isInQueue(SU))
1496 Available.remove(Available.find(SU));
1498 assert(Pending.isInQueue(SU) && "bad ready count");
1499 Pending.remove(Pending.find(SU));
1503 /// If this queue only has one ready candidate, return it. As a side effect,
1504 /// defer any nodes that now hit a hazard, and advance the cycle until at least
1505 /// one node is ready. If multiple instructions are ready, return NULL.
1506 SUnit *ConvergingScheduler::SchedBoundary::pickOnlyChoice() {
1510 if (IssueCount > 0) {
1511 // Defer any ready instrs that now have a hazard.
1512 for (ReadyQueue::iterator I = Available.begin(); I != Available.end();) {
1513 if (checkHazard(*I)) {
1515 I = Available.remove(I);
1521 for (unsigned i = 0; Available.empty(); ++i) {
1522 assert(i <= (HazardRec->getMaxLookAhead() + MaxMinLatency) &&
1523 "permanent hazard"); (void)i;
1527 if (Available.size() == 1)
1528 return *Available.begin();
1532 /// Record the candidate policy for opposite zones with different critical
1535 /// If the CriticalZone is latency limited, don't force a policy for the
1536 /// candidates here. Instead, setLatencyPolicy sets ReduceLatency if needed.
1537 void ConvergingScheduler::balanceZones(
1538 ConvergingScheduler::SchedBoundary &CriticalZone,
1539 ConvergingScheduler::SchedCandidate &CriticalCand,
1540 ConvergingScheduler::SchedBoundary &OppositeZone,
1541 ConvergingScheduler::SchedCandidate &OppositeCand) {
1543 if (!CriticalZone.IsResourceLimited)
1545 assert(SchedModel->hasInstrSchedModel() && "required schedmodel");
1547 SchedRemainder *Rem = CriticalZone.Rem;
1549 // If the critical zone is overconsuming a resource relative to the
1550 // remainder, try to reduce it.
1551 unsigned RemainingCritCount =
1552 Rem->RemainingCounts[CriticalZone.CritResIdx];
1553 if ((int)(Rem->getMaxRemainingCount(SchedModel) - RemainingCritCount)
1554 > (int)SchedModel->getLatencyFactor()) {
1555 CriticalCand.Policy.ReduceResIdx = CriticalZone.CritResIdx;
1556 DEBUG(dbgs() << "Balance " << CriticalZone.Available.getName() << " reduce "
1557 << SchedModel->getProcResource(CriticalZone.CritResIdx)->Name
1560 // If the other zone is underconsuming a resource relative to the full zone,
1561 // try to increase it.
1562 unsigned OppositeCount =
1563 OppositeZone.ResourceCounts[CriticalZone.CritResIdx];
1564 if ((int)(OppositeZone.ExpectedCount - OppositeCount)
1565 > (int)SchedModel->getLatencyFactor()) {
1566 OppositeCand.Policy.DemandResIdx = CriticalZone.CritResIdx;
1567 DEBUG(dbgs() << "Balance " << OppositeZone.Available.getName() << " demand "
1568 << SchedModel->getProcResource(OppositeZone.CritResIdx)->Name
1573 /// Determine if the scheduled zones exceed resource limits or critical path and
1574 /// set each candidate's ReduceHeight policy accordingly.
1575 void ConvergingScheduler::checkResourceLimits(
1576 ConvergingScheduler::SchedCandidate &TopCand,
1577 ConvergingScheduler::SchedCandidate &BotCand) {
1579 // Set ReduceLatency to true if needed.
1580 Bot.setLatencyPolicy(BotCand.Policy);
1581 Top.setLatencyPolicy(TopCand.Policy);
1583 // Handle resource-limited regions.
1584 if (Top.IsResourceLimited && Bot.IsResourceLimited
1585 && Top.CritResIdx == Bot.CritResIdx) {
1586 // If the scheduled critical resource in both zones is no longer the
1587 // critical remaining resource, attempt to reduce resource height both ways.
1588 if (Top.CritResIdx != Rem.CritResIdx) {
1589 TopCand.Policy.ReduceResIdx = Top.CritResIdx;
1590 BotCand.Policy.ReduceResIdx = Bot.CritResIdx;
1591 DEBUG(dbgs() << "Reduce scheduled "
1592 << SchedModel->getProcResource(Top.CritResIdx)->Name << '\n');
1596 // Handle latency-limited regions.
1597 if (!Top.IsResourceLimited && !Bot.IsResourceLimited) {
1598 // If the total scheduled expected latency exceeds the region's critical
1599 // path then reduce latency both ways.
1601 // Just because a zone is not resource limited does not mean it is latency
1602 // limited. Unbuffered resource, such as max micro-ops may cause CurrCycle
1603 // to exceed expected latency.
1604 if ((Top.ExpectedLatency + Bot.ExpectedLatency >= Rem.CriticalPath)
1605 && (Rem.CriticalPath > Top.CurrCycle + Bot.CurrCycle)) {
1606 TopCand.Policy.ReduceLatency = true;
1607 BotCand.Policy.ReduceLatency = true;
1608 DEBUG(dbgs() << "Reduce scheduled latency " << Top.ExpectedLatency
1609 << " + " << Bot.ExpectedLatency << '\n');
1613 // The critical resource is different in each zone, so request balancing.
1615 // Compute the cost of each zone.
1616 Top.ExpectedCount = std::max(Top.ExpectedLatency, Top.CurrCycle);
1617 Top.ExpectedCount = std::max(
1618 Top.getCriticalCount(),
1619 Top.ExpectedCount * SchedModel->getLatencyFactor());
1620 Bot.ExpectedCount = std::max(Bot.ExpectedLatency, Bot.CurrCycle);
1621 Bot.ExpectedCount = std::max(
1622 Bot.getCriticalCount(),
1623 Bot.ExpectedCount * SchedModel->getLatencyFactor());
1625 balanceZones(Top, TopCand, Bot, BotCand);
1626 balanceZones(Bot, BotCand, Top, TopCand);
1629 void ConvergingScheduler::SchedCandidate::
1630 initResourceDelta(const ScheduleDAGMI *DAG,
1631 const TargetSchedModel *SchedModel) {
1632 if (!Policy.ReduceResIdx && !Policy.DemandResIdx)
1635 const MCSchedClassDesc *SC = DAG->getSchedClass(SU);
1636 for (TargetSchedModel::ProcResIter
1637 PI = SchedModel->getWriteProcResBegin(SC),
1638 PE = SchedModel->getWriteProcResEnd(SC); PI != PE; ++PI) {
1639 if (PI->ProcResourceIdx == Policy.ReduceResIdx)
1640 ResDelta.CritResources += PI->Cycles;
1641 if (PI->ProcResourceIdx == Policy.DemandResIdx)
1642 ResDelta.DemandedResources += PI->Cycles;
1646 /// Return true if this heuristic determines order.
1647 static bool tryLess(unsigned TryVal, unsigned CandVal,
1648 ConvergingScheduler::SchedCandidate &TryCand,
1649 ConvergingScheduler::SchedCandidate &Cand,
1650 ConvergingScheduler::CandReason Reason) {
1651 if (TryVal < CandVal) {
1652 TryCand.Reason = Reason;
1655 if (TryVal > CandVal) {
1656 if (Cand.Reason > Reason)
1657 Cand.Reason = Reason;
1663 static bool tryGreater(unsigned TryVal, unsigned CandVal,
1664 ConvergingScheduler::SchedCandidate &TryCand,
1665 ConvergingScheduler::SchedCandidate &Cand,
1666 ConvergingScheduler::CandReason Reason) {
1667 if (TryVal > CandVal) {
1668 TryCand.Reason = Reason;
1671 if (TryVal < CandVal) {
1672 if (Cand.Reason > Reason)
1673 Cand.Reason = Reason;
1679 static unsigned getWeakLeft(const SUnit *SU, bool isTop) {
1680 return (isTop) ? SU->WeakPredsLeft : SU->WeakSuccsLeft;
1683 /// Apply a set of heursitics to a new candidate. Heuristics are currently
1684 /// hierarchical. This may be more efficient than a graduated cost model because
1685 /// we don't need to evaluate all aspects of the model for each node in the
1686 /// queue. But it's really done to make the heuristics easier to debug and
1687 /// statistically analyze.
1689 /// \param Cand provides the policy and current best candidate.
1690 /// \param TryCand refers to the next SUnit candidate, otherwise uninitialized.
1691 /// \param Zone describes the scheduled zone that we are extending.
1692 /// \param RPTracker describes reg pressure within the scheduled zone.
1693 /// \param TempTracker is a scratch pressure tracker to reuse in queries.
1694 void ConvergingScheduler::tryCandidate(SchedCandidate &Cand,
1695 SchedCandidate &TryCand,
1696 SchedBoundary &Zone,
1697 const RegPressureTracker &RPTracker,
1698 RegPressureTracker &TempTracker) {
1700 // Always initialize TryCand's RPDelta.
1701 TempTracker.getMaxPressureDelta(TryCand.SU->getInstr(), TryCand.RPDelta,
1702 DAG->getRegionCriticalPSets(),
1703 DAG->getRegPressure().MaxSetPressure);
1705 // Initialize the candidate if needed.
1706 if (!Cand.isValid()) {
1707 TryCand.Reason = NodeOrder;
1710 // Avoid exceeding the target's limit.
1711 if (tryLess(TryCand.RPDelta.Excess.UnitIncrease,
1712 Cand.RPDelta.Excess.UnitIncrease, TryCand, Cand, SingleExcess))
1714 if (Cand.Reason == SingleExcess)
1715 Cand.Reason = MultiPressure;
1717 // Avoid increasing the max critical pressure in the scheduled region.
1718 if (tryLess(TryCand.RPDelta.CriticalMax.UnitIncrease,
1719 Cand.RPDelta.CriticalMax.UnitIncrease,
1720 TryCand, Cand, SingleCritical))
1722 if (Cand.Reason == SingleCritical)
1723 Cand.Reason = MultiPressure;
1725 // Keep clustered nodes together to encourage downstream peephole
1726 // optimizations which may reduce resource requirements.
1728 // This is a best effort to set things up for a post-RA pass. Optimizations
1729 // like generating loads of multiple registers should ideally be done within
1730 // the scheduler pass by combining the loads during DAG postprocessing.
1731 const SUnit *NextClusterSU =
1732 Zone.isTop() ? DAG->getNextClusterSucc() : DAG->getNextClusterPred();
1733 if (tryGreater(TryCand.SU == NextClusterSU, Cand.SU == NextClusterSU,
1734 TryCand, Cand, Cluster))
1736 // Currently, weak edges are for clustering, so we hard-code that reason.
1737 // However, deferring the current TryCand will not change Cand's reason.
1738 CandReason OrigReason = Cand.Reason;
1739 if (tryLess(getWeakLeft(TryCand.SU, Zone.isTop()),
1740 getWeakLeft(Cand.SU, Zone.isTop()),
1741 TryCand, Cand, Cluster)) {
1742 Cand.Reason = OrigReason;
1745 // Avoid critical resource consumption and balance the schedule.
1746 TryCand.initResourceDelta(DAG, SchedModel);
1747 if (tryLess(TryCand.ResDelta.CritResources, Cand.ResDelta.CritResources,
1748 TryCand, Cand, ResourceReduce))
1750 if (tryGreater(TryCand.ResDelta.DemandedResources,
1751 Cand.ResDelta.DemandedResources,
1752 TryCand, Cand, ResourceDemand))
1755 // Avoid serializing long latency dependence chains.
1756 if (Cand.Policy.ReduceLatency) {
1758 if (Cand.SU->getDepth() * SchedModel->getLatencyFactor()
1759 > Zone.ExpectedCount) {
1760 if (tryLess(TryCand.SU->getDepth(), Cand.SU->getDepth(),
1761 TryCand, Cand, TopDepthReduce))
1764 if (tryGreater(TryCand.SU->getHeight(), Cand.SU->getHeight(),
1765 TryCand, Cand, TopPathReduce))
1769 if (Cand.SU->getHeight() * SchedModel->getLatencyFactor()
1770 > Zone.ExpectedCount) {
1771 if (tryLess(TryCand.SU->getHeight(), Cand.SU->getHeight(),
1772 TryCand, Cand, BotHeightReduce))
1775 if (tryGreater(TryCand.SU->getDepth(), Cand.SU->getDepth(),
1776 TryCand, Cand, BotPathReduce))
1781 // Avoid increasing the max pressure of the entire region.
1782 if (tryLess(TryCand.RPDelta.CurrentMax.UnitIncrease,
1783 Cand.RPDelta.CurrentMax.UnitIncrease, TryCand, Cand, SingleMax))
1785 if (Cand.Reason == SingleMax)
1786 Cand.Reason = MultiPressure;
1788 // Prefer immediate defs/users of the last scheduled instruction. This is a
1789 // nice pressure avoidance strategy that also conserves the processor's
1790 // register renaming resources and keeps the machine code readable.
1791 if (tryGreater(Zone.NextSUs.count(TryCand.SU), Zone.NextSUs.count(Cand.SU),
1792 TryCand, Cand, NextDefUse))
1795 // Fall through to original instruction order.
1796 if ((Zone.isTop() && TryCand.SU->NodeNum < Cand.SU->NodeNum)
1797 || (!Zone.isTop() && TryCand.SU->NodeNum > Cand.SU->NodeNum)) {
1798 TryCand.Reason = NodeOrder;
1802 /// pickNodeFromQueue helper that returns true if the LHS reg pressure effect is
1803 /// more desirable than RHS from scheduling standpoint.
1804 static bool compareRPDelta(const RegPressureDelta &LHS,
1805 const RegPressureDelta &RHS) {
1806 // Compare each component of pressure in decreasing order of importance
1807 // without checking if any are valid. Invalid PressureElements are assumed to
1808 // have UnitIncrease==0, so are neutral.
1810 // Avoid increasing the max critical pressure in the scheduled region.
1811 if (LHS.Excess.UnitIncrease != RHS.Excess.UnitIncrease) {
1812 DEBUG(dbgs() << "RP excess top - bot: "
1813 << (LHS.Excess.UnitIncrease - RHS.Excess.UnitIncrease) << '\n');
1814 return LHS.Excess.UnitIncrease < RHS.Excess.UnitIncrease;
1816 // Avoid increasing the max critical pressure in the scheduled region.
1817 if (LHS.CriticalMax.UnitIncrease != RHS.CriticalMax.UnitIncrease) {
1818 DEBUG(dbgs() << "RP critical top - bot: "
1819 << (LHS.CriticalMax.UnitIncrease - RHS.CriticalMax.UnitIncrease)
1821 return LHS.CriticalMax.UnitIncrease < RHS.CriticalMax.UnitIncrease;
1823 // Avoid increasing the max pressure of the entire region.
1824 if (LHS.CurrentMax.UnitIncrease != RHS.CurrentMax.UnitIncrease) {
1825 DEBUG(dbgs() << "RP current top - bot: "
1826 << (LHS.CurrentMax.UnitIncrease - RHS.CurrentMax.UnitIncrease)
1828 return LHS.CurrentMax.UnitIncrease < RHS.CurrentMax.UnitIncrease;
1834 const char *ConvergingScheduler::getReasonStr(
1835 ConvergingScheduler::CandReason Reason) {
1837 case NoCand: return "NOCAND ";
1838 case SingleExcess: return "REG-EXCESS";
1839 case SingleCritical: return "REG-CRIT ";
1840 case Cluster: return "CLUSTER ";
1841 case SingleMax: return "REG-MAX ";
1842 case MultiPressure: return "REG-MULTI ";
1843 case ResourceReduce: return "RES-REDUCE";
1844 case ResourceDemand: return "RES-DEMAND";
1845 case TopDepthReduce: return "TOP-DEPTH ";
1846 case TopPathReduce: return "TOP-PATH ";
1847 case BotHeightReduce:return "BOT-HEIGHT";
1848 case BotPathReduce: return "BOT-PATH ";
1849 case NextDefUse: return "DEF-USE ";
1850 case NodeOrder: return "ORDER ";
1852 llvm_unreachable("Unknown reason!");
1855 void ConvergingScheduler::traceCandidate(const SchedCandidate &Cand,
1856 const SchedBoundary &Zone) {
1857 const char *Label = getReasonStr(Cand.Reason);
1859 unsigned ResIdx = 0;
1860 unsigned Latency = 0;
1861 switch (Cand.Reason) {
1865 P = Cand.RPDelta.Excess;
1867 case SingleCritical:
1868 P = Cand.RPDelta.CriticalMax;
1871 P = Cand.RPDelta.CurrentMax;
1873 case ResourceReduce:
1874 ResIdx = Cand.Policy.ReduceResIdx;
1876 case ResourceDemand:
1877 ResIdx = Cand.Policy.DemandResIdx;
1879 case TopDepthReduce:
1880 Latency = Cand.SU->getDepth();
1883 Latency = Cand.SU->getHeight();
1885 case BotHeightReduce:
1886 Latency = Cand.SU->getHeight();
1889 Latency = Cand.SU->getDepth();
1892 dbgs() << Label << " " << Zone.Available.getName() << " ";
1894 dbgs() << TRI->getRegPressureSetName(P.PSetID) << ":" << P.UnitIncrease
1899 dbgs() << SchedModel->getProcResource(ResIdx)->Name << " ";
1903 dbgs() << Latency << " cycles ";
1910 /// Pick the best candidate from the top queue.
1912 /// TODO: getMaxPressureDelta results can be mostly cached for each SUnit during
1913 /// DAG building. To adjust for the current scheduling location we need to
1914 /// maintain the number of vreg uses remaining to be top-scheduled.
1915 void ConvergingScheduler::pickNodeFromQueue(SchedBoundary &Zone,
1916 const RegPressureTracker &RPTracker,
1917 SchedCandidate &Cand) {
1918 ReadyQueue &Q = Zone.Available;
1922 // getMaxPressureDelta temporarily modifies the tracker.
1923 RegPressureTracker &TempTracker = const_cast<RegPressureTracker&>(RPTracker);
1925 for (ReadyQueue::iterator I = Q.begin(), E = Q.end(); I != E; ++I) {
1927 SchedCandidate TryCand(Cand.Policy);
1929 tryCandidate(Cand, TryCand, Zone, RPTracker, TempTracker);
1930 if (TryCand.Reason != NoCand) {
1931 // Initialize resource delta if needed in case future heuristics query it.
1932 if (TryCand.ResDelta == SchedResourceDelta())
1933 TryCand.initResourceDelta(DAG, SchedModel);
1934 Cand.setBest(TryCand);
1935 DEBUG(traceCandidate(Cand, Zone));
1940 static void tracePick(const ConvergingScheduler::SchedCandidate &Cand,
1942 DEBUG(dbgs() << "Pick " << (IsTop ? "top" : "bot")
1943 << " SU(" << Cand.SU->NodeNum << ") "
1944 << ConvergingScheduler::getReasonStr(Cand.Reason) << '\n');
1947 /// Pick the best candidate node from either the top or bottom queue.
1948 SUnit *ConvergingScheduler::pickNodeBidirectional(bool &IsTopNode) {
1949 // Schedule as far as possible in the direction of no choice. This is most
1950 // efficient, but also provides the best heuristics for CriticalPSets.
1951 if (SUnit *SU = Bot.pickOnlyChoice()) {
1955 if (SUnit *SU = Top.pickOnlyChoice()) {
1959 CandPolicy NoPolicy;
1960 SchedCandidate BotCand(NoPolicy);
1961 SchedCandidate TopCand(NoPolicy);
1962 checkResourceLimits(TopCand, BotCand);
1964 // Prefer bottom scheduling when heuristics are silent.
1965 pickNodeFromQueue(Bot, DAG->getBotRPTracker(), BotCand);
1966 assert(BotCand.Reason != NoCand && "failed to find the first candidate");
1968 // If either Q has a single candidate that provides the least increase in
1969 // Excess pressure, we can immediately schedule from that Q.
1971 // RegionCriticalPSets summarizes the pressure within the scheduled region and
1972 // affects picking from either Q. If scheduling in one direction must
1973 // increase pressure for one of the excess PSets, then schedule in that
1974 // direction first to provide more freedom in the other direction.
1975 if (BotCand.Reason == SingleExcess || BotCand.Reason == SingleCritical) {
1977 tracePick(BotCand, IsTopNode);
1980 // Check if the top Q has a better candidate.
1981 pickNodeFromQueue(Top, DAG->getTopRPTracker(), TopCand);
1982 assert(TopCand.Reason != NoCand && "failed to find the first candidate");
1984 // If either Q has a single candidate that minimizes pressure above the
1985 // original region's pressure pick it.
1986 if (TopCand.Reason <= SingleMax || BotCand.Reason <= SingleMax) {
1987 if (TopCand.Reason < BotCand.Reason) {
1989 tracePick(TopCand, IsTopNode);
1993 tracePick(BotCand, IsTopNode);
1996 // Check for a salient pressure difference and pick the best from either side.
1997 if (compareRPDelta(TopCand.RPDelta, BotCand.RPDelta)) {
1999 tracePick(TopCand, IsTopNode);
2002 // Otherwise prefer the bottom candidate, in node order if all else failed.
2003 if (TopCand.Reason < BotCand.Reason) {
2005 tracePick(TopCand, IsTopNode);
2009 tracePick(BotCand, IsTopNode);
2013 /// Pick the best node to balance the schedule. Implements MachineSchedStrategy.
2014 SUnit *ConvergingScheduler::pickNode(bool &IsTopNode) {
2015 if (DAG->top() == DAG->bottom()) {
2016 assert(Top.Available.empty() && Top.Pending.empty() &&
2017 Bot.Available.empty() && Bot.Pending.empty() && "ReadyQ garbage");
2023 SU = Top.pickOnlyChoice();
2025 CandPolicy NoPolicy;
2026 SchedCandidate TopCand(NoPolicy);
2027 pickNodeFromQueue(Top, DAG->getTopRPTracker(), TopCand);
2028 assert(TopCand.Reason != NoCand && "failed to find the first candidate");
2033 else if (ForceBottomUp) {
2034 SU = Bot.pickOnlyChoice();
2036 CandPolicy NoPolicy;
2037 SchedCandidate BotCand(NoPolicy);
2038 pickNodeFromQueue(Bot, DAG->getBotRPTracker(), BotCand);
2039 assert(BotCand.Reason != NoCand && "failed to find the first candidate");
2045 SU = pickNodeBidirectional(IsTopNode);
2047 } while (SU->isScheduled);
2049 if (SU->isTopReady())
2050 Top.removeReady(SU);
2051 if (SU->isBottomReady())
2052 Bot.removeReady(SU);
2054 DEBUG(dbgs() << "*** " << (IsTopNode ? "Top" : "Bottom")
2055 << " Scheduling Instruction in cycle "
2056 << (IsTopNode ? Top.CurrCycle : Bot.CurrCycle) << '\n';
2061 /// Update the scheduler's state after scheduling a node. This is the same node
2062 /// that was just returned by pickNode(). However, ScheduleDAGMI needs to update
2063 /// it's state based on the current cycle before MachineSchedStrategy does.
2064 void ConvergingScheduler::schedNode(SUnit *SU, bool IsTopNode) {
2066 SU->TopReadyCycle = Top.CurrCycle;
2070 SU->BotReadyCycle = Bot.CurrCycle;
2075 /// Create the standard converging machine scheduler. This will be used as the
2076 /// default scheduler if the target does not set a default.
2077 static ScheduleDAGInstrs *createConvergingSched(MachineSchedContext *C) {
2078 assert((!ForceTopDown || !ForceBottomUp) &&
2079 "-misched-topdown incompatible with -misched-bottomup");
2080 ScheduleDAGMI *DAG = new ScheduleDAGMI(C, new ConvergingScheduler());
2081 // Register DAG post-processors.
2082 if (EnableLoadCluster)
2083 DAG->addMutation(new LoadClusterMutation(DAG->TII, DAG->TRI));
2084 if (EnableMacroFusion)
2085 DAG->addMutation(new MacroFusion(DAG->TII));
2088 static MachineSchedRegistry
2089 ConvergingSchedRegistry("converge", "Standard converging scheduler.",
2090 createConvergingSched);
2092 //===----------------------------------------------------------------------===//
2093 // ILP Scheduler. Currently for experimental analysis of heuristics.
2094 //===----------------------------------------------------------------------===//
2097 /// \brief Order nodes by the ILP metric.
2099 const SchedDFSResult *DFSResult;
2100 const BitVector *ScheduledTrees;
2103 ILPOrder(bool MaxILP): DFSResult(0), ScheduledTrees(0), MaximizeILP(MaxILP) {}
2105 /// \brief Apply a less-than relation on node priority.
2107 /// (Return true if A comes after B in the Q.)
2108 bool operator()(const SUnit *A, const SUnit *B) const {
2109 unsigned SchedTreeA = DFSResult->getSubtreeID(A);
2110 unsigned SchedTreeB = DFSResult->getSubtreeID(B);
2111 if (SchedTreeA != SchedTreeB) {
2112 // Unscheduled trees have lower priority.
2113 if (ScheduledTrees->test(SchedTreeA) != ScheduledTrees->test(SchedTreeB))
2114 return ScheduledTrees->test(SchedTreeB);
2116 // Trees with shallower connections have have lower priority.
2117 if (DFSResult->getSubtreeLevel(SchedTreeA)
2118 != DFSResult->getSubtreeLevel(SchedTreeB)) {
2119 return DFSResult->getSubtreeLevel(SchedTreeA)
2120 < DFSResult->getSubtreeLevel(SchedTreeB);
2124 return DFSResult->getILP(A) < DFSResult->getILP(B);
2126 return DFSResult->getILP(A) > DFSResult->getILP(B);
2130 /// \brief Schedule based on the ILP metric.
2131 class ILPScheduler : public MachineSchedStrategy {
2132 /// In case all subtrees are eventually connected to a common root through
2133 /// data dependence (e.g. reduction), place an upper limit on their size.
2135 /// FIXME: A subtree limit is generally good, but in the situation commented
2136 /// above, where multiple similar subtrees feed a common root, we should
2137 /// only split at a point where the resulting subtrees will be balanced.
2138 /// (a motivating test case must be found).
2139 static const unsigned SubtreeLimit = 16;
2144 std::vector<SUnit*> ReadyQ;
2146 ILPScheduler(bool MaximizeILP): DAG(0), Cmp(MaximizeILP) {}
2148 virtual void initialize(ScheduleDAGMI *dag) {
2150 DAG->computeDFSResult();
2151 Cmp.DFSResult = DAG->getDFSResult();
2152 Cmp.ScheduledTrees = &DAG->getScheduledTrees();
2156 virtual void registerRoots() {
2157 // Restore the heap in ReadyQ with the updated DFS results.
2158 std::make_heap(ReadyQ.begin(), ReadyQ.end(), Cmp);
2161 /// Implement MachineSchedStrategy interface.
2162 /// -----------------------------------------
2164 /// Callback to select the highest priority node from the ready Q.
2165 virtual SUnit *pickNode(bool &IsTopNode) {
2166 if (ReadyQ.empty()) return NULL;
2167 pop_heap(ReadyQ.begin(), ReadyQ.end(), Cmp);
2168 SUnit *SU = ReadyQ.back();
2171 DEBUG(dbgs() << "*** Scheduling " << "SU(" << SU->NodeNum << "): "
2173 << " ILP: " << DAG->getDFSResult()->getILP(SU)
2174 << " Tree: " << DAG->getDFSResult()->getSubtreeID(SU) << " @"
2175 << DAG->getDFSResult()->getSubtreeLevel(
2176 DAG->getDFSResult()->getSubtreeID(SU)) << '\n');
2180 /// \brief Scheduler callback to notify that a new subtree is scheduled.
2181 virtual void scheduleTree(unsigned SubtreeID) {
2182 std::make_heap(ReadyQ.begin(), ReadyQ.end(), Cmp);
2185 /// Callback after a node is scheduled. Mark a newly scheduled tree, notify
2186 /// DFSResults, and resort the priority Q.
2187 virtual void schedNode(SUnit *SU, bool IsTopNode) {
2188 assert(!IsTopNode && "SchedDFSResult needs bottom-up");
2191 virtual void releaseTopNode(SUnit *) { /*only called for top roots*/ }
2193 virtual void releaseBottomNode(SUnit *SU) {
2194 ReadyQ.push_back(SU);
2195 std::push_heap(ReadyQ.begin(), ReadyQ.end(), Cmp);
2200 static ScheduleDAGInstrs *createILPMaxScheduler(MachineSchedContext *C) {
2201 return new ScheduleDAGMI(C, new ILPScheduler(true));
2203 static ScheduleDAGInstrs *createILPMinScheduler(MachineSchedContext *C) {
2204 return new ScheduleDAGMI(C, new ILPScheduler(false));
2206 static MachineSchedRegistry ILPMaxRegistry(
2207 "ilpmax", "Schedule bottom-up for max ILP", createILPMaxScheduler);
2208 static MachineSchedRegistry ILPMinRegistry(
2209 "ilpmin", "Schedule bottom-up for min ILP", createILPMinScheduler);
2211 //===----------------------------------------------------------------------===//
2212 // Machine Instruction Shuffler for Correctness Testing
2213 //===----------------------------------------------------------------------===//
2217 /// Apply a less-than relation on the node order, which corresponds to the
2218 /// instruction order prior to scheduling. IsReverse implements greater-than.
2219 template<bool IsReverse>
2221 bool operator()(SUnit *A, SUnit *B) const {
2223 return A->NodeNum > B->NodeNum;
2225 return A->NodeNum < B->NodeNum;
2229 /// Reorder instructions as much as possible.
2230 class InstructionShuffler : public MachineSchedStrategy {
2234 // Using a less-than relation (SUnitOrder<false>) for the TopQ priority
2235 // gives nodes with a higher number higher priority causing the latest
2236 // instructions to be scheduled first.
2237 PriorityQueue<SUnit*, std::vector<SUnit*>, SUnitOrder<false> >
2239 // When scheduling bottom-up, use greater-than as the queue priority.
2240 PriorityQueue<SUnit*, std::vector<SUnit*>, SUnitOrder<true> >
2243 InstructionShuffler(bool alternate, bool topdown)
2244 : IsAlternating(alternate), IsTopDown(topdown) {}
2246 virtual void initialize(ScheduleDAGMI *) {
2251 /// Implement MachineSchedStrategy interface.
2252 /// -----------------------------------------
2254 virtual SUnit *pickNode(bool &IsTopNode) {
2258 if (TopQ.empty()) return NULL;
2261 } while (SU->isScheduled);
2266 if (BottomQ.empty()) return NULL;
2269 } while (SU->isScheduled);
2273 IsTopDown = !IsTopDown;
2277 virtual void schedNode(SUnit *SU, bool IsTopNode) {}
2279 virtual void releaseTopNode(SUnit *SU) {
2282 virtual void releaseBottomNode(SUnit *SU) {
2288 static ScheduleDAGInstrs *createInstructionShuffler(MachineSchedContext *C) {
2289 bool Alternate = !ForceTopDown && !ForceBottomUp;
2290 bool TopDown = !ForceBottomUp;
2291 assert((TopDown || !ForceTopDown) &&
2292 "-misched-topdown incompatible with -misched-bottomup");
2293 return new ScheduleDAGMI(C, new InstructionShuffler(Alternate, TopDown));
2295 static MachineSchedRegistry ShufflerRegistry(
2296 "shuffle", "Shuffle machine instructions alternating directions",
2297 createInstructionShuffler);
2300 //===----------------------------------------------------------------------===//
2301 // GraphWriter support for ScheduleDAGMI.
2302 //===----------------------------------------------------------------------===//
2307 template<> struct GraphTraits<
2308 ScheduleDAGMI*> : public GraphTraits<ScheduleDAG*> {};
2311 struct DOTGraphTraits<ScheduleDAGMI*> : public DefaultDOTGraphTraits {
2313 DOTGraphTraits (bool isSimple=false) : DefaultDOTGraphTraits(isSimple) {}
2315 static std::string getGraphName(const ScheduleDAG *G) {
2316 return G->MF.getName();
2319 static bool renderGraphFromBottomUp() {
2323 static bool isNodeHidden(const SUnit *Node) {
2324 return (Node->NumPreds > 10 || Node->NumSuccs > 10);
2327 static bool hasNodeAddressLabel(const SUnit *Node,
2328 const ScheduleDAG *Graph) {
2332 /// If you want to override the dot attributes printed for a particular
2333 /// edge, override this method.
2334 static std::string getEdgeAttributes(const SUnit *Node,
2336 const ScheduleDAG *Graph) {
2337 if (EI.isArtificialDep())
2338 return "color=cyan,style=dashed";
2340 return "color=blue,style=dashed";
2344 static std::string getNodeLabel(const SUnit *SU, const ScheduleDAG *G) {
2346 raw_string_ostream SS(Str);
2347 SS << "SU(" << SU->NodeNum << ')';
2350 static std::string getNodeDescription(const SUnit *SU, const ScheduleDAG *G) {
2351 return G->getGraphNodeLabel(SU);
2354 static std::string getNodeAttributes(const SUnit *N,
2355 const ScheduleDAG *Graph) {
2356 std::string Str("shape=Mrecord");
2357 const SchedDFSResult *DFS =
2358 static_cast<const ScheduleDAGMI*>(Graph)->getDFSResult();
2360 Str += ",style=filled,fillcolor=\"#";
2361 Str += DOT::getColorString(DFS->getSubtreeID(N));
2370 /// viewGraph - Pop up a ghostview window with the reachable parts of the DAG
2371 /// rendered using 'dot'.
2373 void ScheduleDAGMI::viewGraph(const Twine &Name, const Twine &Title) {
2375 ViewGraph(this, Name, false, Title);
2377 errs() << "ScheduleDAGMI::viewGraph is only available in debug builds on "
2378 << "systems with Graphviz or gv!\n";
2382 /// Out-of-line implementation with no arguments is handy for gdb.
2383 void ScheduleDAGMI::viewGraph() {
2384 viewGraph(getDAGName(), "Scheduling-Units Graph for " + getDAGName());