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/LiveIntervalAnalysis.h"
18 #include "llvm/CodeGen/MachineScheduler.h"
19 #include "llvm/CodeGen/Passes.h"
20 #include "llvm/CodeGen/RegisterClassInfo.h"
21 #include "llvm/CodeGen/ScheduleHazardRecognizer.h"
22 #include "llvm/Analysis/AliasAnalysis.h"
23 #include "llvm/Support/CommandLine.h"
24 #include "llvm/Support/Debug.h"
25 #include "llvm/Support/ErrorHandling.h"
26 #include "llvm/Support/raw_ostream.h"
27 #include "llvm/ADT/OwningPtr.h"
28 #include "llvm/ADT/PriorityQueue.h"
35 cl::opt<bool> ForceTopDown("misched-topdown", cl::Hidden,
36 cl::desc("Force top-down list scheduling"));
37 cl::opt<bool> ForceBottomUp("misched-bottomup", cl::Hidden,
38 cl::desc("Force bottom-up list scheduling"));
42 static cl::opt<bool> ViewMISchedDAGs("view-misched-dags", cl::Hidden,
43 cl::desc("Pop up a window to show MISched dags after they are processed"));
45 static cl::opt<unsigned> MISchedCutoff("misched-cutoff", cl::Hidden,
46 cl::desc("Stop scheduling after N instructions"), cl::init(~0U));
48 static bool ViewMISchedDAGs = false;
51 //===----------------------------------------------------------------------===//
52 // Machine Instruction Scheduling Pass and Registry
53 //===----------------------------------------------------------------------===//
55 MachineSchedContext::MachineSchedContext():
56 MF(0), MLI(0), MDT(0), PassConfig(0), AA(0), LIS(0) {
57 RegClassInfo = new RegisterClassInfo();
60 MachineSchedContext::~MachineSchedContext() {
65 /// MachineScheduler runs after coalescing and before register allocation.
66 class MachineScheduler : public MachineSchedContext,
67 public MachineFunctionPass {
71 virtual void getAnalysisUsage(AnalysisUsage &AU) const;
73 virtual void releaseMemory() {}
75 virtual bool runOnMachineFunction(MachineFunction&);
77 virtual void print(raw_ostream &O, const Module* = 0) const;
79 static char ID; // Class identification, replacement for typeinfo
83 char MachineScheduler::ID = 0;
85 char &llvm::MachineSchedulerID = MachineScheduler::ID;
87 INITIALIZE_PASS_BEGIN(MachineScheduler, "misched",
88 "Machine Instruction Scheduler", false, false)
89 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
90 INITIALIZE_PASS_DEPENDENCY(SlotIndexes)
91 INITIALIZE_PASS_DEPENDENCY(LiveIntervals)
92 INITIALIZE_PASS_END(MachineScheduler, "misched",
93 "Machine Instruction Scheduler", false, false)
95 MachineScheduler::MachineScheduler()
96 : MachineFunctionPass(ID) {
97 initializeMachineSchedulerPass(*PassRegistry::getPassRegistry());
100 void MachineScheduler::getAnalysisUsage(AnalysisUsage &AU) const {
101 AU.setPreservesCFG();
102 AU.addRequiredID(MachineDominatorsID);
103 AU.addRequired<MachineLoopInfo>();
104 AU.addRequired<AliasAnalysis>();
105 AU.addRequired<TargetPassConfig>();
106 AU.addRequired<SlotIndexes>();
107 AU.addPreserved<SlotIndexes>();
108 AU.addRequired<LiveIntervals>();
109 AU.addPreserved<LiveIntervals>();
110 MachineFunctionPass::getAnalysisUsage(AU);
113 MachinePassRegistry MachineSchedRegistry::Registry;
115 /// A dummy default scheduler factory indicates whether the scheduler
116 /// is overridden on the command line.
117 static ScheduleDAGInstrs *useDefaultMachineSched(MachineSchedContext *C) {
121 /// MachineSchedOpt allows command line selection of the scheduler.
122 static cl::opt<MachineSchedRegistry::ScheduleDAGCtor, false,
123 RegisterPassParser<MachineSchedRegistry> >
124 MachineSchedOpt("misched",
125 cl::init(&useDefaultMachineSched), cl::Hidden,
126 cl::desc("Machine instruction scheduler to use"));
128 static MachineSchedRegistry
129 DefaultSchedRegistry("default", "Use the target's default scheduler choice.",
130 useDefaultMachineSched);
132 /// Forward declare the standard machine scheduler. This will be used as the
133 /// default scheduler if the target does not set a default.
134 static ScheduleDAGInstrs *createConvergingSched(MachineSchedContext *C);
137 /// Decrement this iterator until reaching the top or a non-debug instr.
138 static MachineBasicBlock::iterator
139 priorNonDebug(MachineBasicBlock::iterator I, MachineBasicBlock::iterator Beg) {
140 assert(I != Beg && "reached the top of the region, cannot decrement");
142 if (!I->isDebugValue())
148 /// If this iterator is a debug value, increment until reaching the End or a
149 /// non-debug instruction.
150 static MachineBasicBlock::iterator
151 nextIfDebug(MachineBasicBlock::iterator I, MachineBasicBlock::iterator End) {
152 for(; I != End; ++I) {
153 if (!I->isDebugValue())
159 /// Top-level MachineScheduler pass driver.
161 /// Visit blocks in function order. Divide each block into scheduling regions
162 /// and visit them bottom-up. Visiting regions bottom-up is not required, but is
163 /// consistent with the DAG builder, which traverses the interior of the
164 /// scheduling regions bottom-up.
166 /// This design avoids exposing scheduling boundaries to the DAG builder,
167 /// simplifying the DAG builder's support for "special" target instructions.
168 /// At the same time the design allows target schedulers to operate across
169 /// scheduling boundaries, for example to bundle the boudary instructions
170 /// without reordering them. This creates complexity, because the target
171 /// scheduler must update the RegionBegin and RegionEnd positions cached by
172 /// ScheduleDAGInstrs whenever adding or removing instructions. A much simpler
173 /// design would be to split blocks at scheduling boundaries, but LLVM has a
174 /// general bias against block splitting purely for implementation simplicity.
175 bool MachineScheduler::runOnMachineFunction(MachineFunction &mf) {
176 DEBUG(dbgs() << "Before MISsched:\n"; mf.print(dbgs()));
178 // Initialize the context of the pass.
180 MLI = &getAnalysis<MachineLoopInfo>();
181 MDT = &getAnalysis<MachineDominatorTree>();
182 PassConfig = &getAnalysis<TargetPassConfig>();
183 AA = &getAnalysis<AliasAnalysis>();
185 LIS = &getAnalysis<LiveIntervals>();
186 const TargetInstrInfo *TII = MF->getTarget().getInstrInfo();
188 RegClassInfo->runOnMachineFunction(*MF);
190 // Select the scheduler, or set the default.
191 MachineSchedRegistry::ScheduleDAGCtor Ctor = MachineSchedOpt;
192 if (Ctor == useDefaultMachineSched) {
193 // Get the default scheduler set by the target.
194 Ctor = MachineSchedRegistry::getDefault();
196 Ctor = createConvergingSched;
197 MachineSchedRegistry::setDefault(Ctor);
200 // Instantiate the selected scheduler.
201 OwningPtr<ScheduleDAGInstrs> Scheduler(Ctor(this));
203 // Visit all machine basic blocks.
205 // TODO: Visit blocks in global postorder or postorder within the bottom-up
206 // loop tree. Then we can optionally compute global RegPressure.
207 for (MachineFunction::iterator MBB = MF->begin(), MBBEnd = MF->end();
208 MBB != MBBEnd; ++MBB) {
210 Scheduler->startBlock(MBB);
212 // Break the block into scheduling regions [I, RegionEnd), and schedule each
213 // region as soon as it is discovered. RegionEnd points the scheduling
214 // boundary at the bottom of the region. The DAG does not include RegionEnd,
215 // but the region does (i.e. the next RegionEnd is above the previous
216 // RegionBegin). If the current block has no terminator then RegionEnd ==
217 // MBB->end() for the bottom region.
219 // The Scheduler may insert instructions during either schedule() or
220 // exitRegion(), even for empty regions. So the local iterators 'I' and
221 // 'RegionEnd' are invalid across these calls.
222 unsigned RemainingCount = MBB->size();
223 for(MachineBasicBlock::iterator RegionEnd = MBB->end();
224 RegionEnd != MBB->begin(); RegionEnd = Scheduler->begin()) {
226 // Avoid decrementing RegionEnd for blocks with no terminator.
227 if (RegionEnd != MBB->end()
228 || TII->isSchedulingBoundary(llvm::prior(RegionEnd), MBB, *MF)) {
230 // Count the boundary instruction.
234 // The next region starts above the previous region. Look backward in the
235 // instruction stream until we find the nearest boundary.
236 MachineBasicBlock::iterator I = RegionEnd;
237 for(;I != MBB->begin(); --I, --RemainingCount) {
238 if (TII->isSchedulingBoundary(llvm::prior(I), MBB, *MF))
241 // Notify the scheduler of the region, even if we may skip scheduling
242 // it. Perhaps it still needs to be bundled.
243 Scheduler->enterRegion(MBB, I, RegionEnd, RemainingCount);
245 // Skip empty scheduling regions (0 or 1 schedulable instructions).
246 if (I == RegionEnd || I == llvm::prior(RegionEnd)) {
247 // Close the current region. Bundle the terminator if needed.
248 // This invalidates 'RegionEnd' and 'I'.
249 Scheduler->exitRegion();
252 DEBUG(dbgs() << "********** MI Scheduling **********\n");
253 DEBUG(dbgs() << MF->getName()
254 << ":BB#" << MBB->getNumber() << "\n From: " << *I << " To: ";
255 if (RegionEnd != MBB->end()) dbgs() << *RegionEnd;
256 else dbgs() << "End";
257 dbgs() << " Remaining: " << RemainingCount << "\n");
259 // Schedule a region: possibly reorder instructions.
260 // This invalidates 'RegionEnd' and 'I'.
261 Scheduler->schedule();
263 // Close the current region.
264 Scheduler->exitRegion();
266 // Scheduling has invalidated the current iterator 'I'. Ask the
267 // scheduler for the top of it's scheduled region.
268 RegionEnd = Scheduler->begin();
270 assert(RemainingCount == 0 && "Instruction count mismatch!");
271 Scheduler->finishBlock();
273 Scheduler->finalizeSchedule();
274 DEBUG(LIS->print(dbgs()));
278 void MachineScheduler::print(raw_ostream &O, const Module* m) const {
283 void ReadyQueue::dump() {
284 dbgs() << Name << ": ";
285 for (unsigned i = 0, e = Queue.size(); i < e; ++i)
286 dbgs() << Queue[i]->NodeNum << " ";
291 //===----------------------------------------------------------------------===//
292 // ScheduleDAGMI - Base class for MachineInstr scheduling with LiveIntervals
294 //===----------------------------------------------------------------------===//
296 /// ReleaseSucc - Decrement the NumPredsLeft count of a successor. When
297 /// NumPredsLeft reaches zero, release the successor node.
299 /// FIXME: Adjust SuccSU height based on MinLatency.
300 void ScheduleDAGMI::releaseSucc(SUnit *SU, SDep *SuccEdge) {
301 SUnit *SuccSU = SuccEdge->getSUnit();
304 if (SuccSU->NumPredsLeft == 0) {
305 dbgs() << "*** Scheduling failed! ***\n";
307 dbgs() << " has been released too many times!\n";
311 --SuccSU->NumPredsLeft;
312 if (SuccSU->NumPredsLeft == 0 && SuccSU != &ExitSU)
313 SchedImpl->releaseTopNode(SuccSU);
316 /// releaseSuccessors - Call releaseSucc on each of SU's successors.
317 void ScheduleDAGMI::releaseSuccessors(SUnit *SU) {
318 for (SUnit::succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
320 releaseSucc(SU, &*I);
324 /// ReleasePred - Decrement the NumSuccsLeft count of a predecessor. When
325 /// NumSuccsLeft reaches zero, release the predecessor node.
327 /// FIXME: Adjust PredSU height based on MinLatency.
328 void ScheduleDAGMI::releasePred(SUnit *SU, SDep *PredEdge) {
329 SUnit *PredSU = PredEdge->getSUnit();
332 if (PredSU->NumSuccsLeft == 0) {
333 dbgs() << "*** Scheduling failed! ***\n";
335 dbgs() << " has been released too many times!\n";
339 --PredSU->NumSuccsLeft;
340 if (PredSU->NumSuccsLeft == 0 && PredSU != &EntrySU)
341 SchedImpl->releaseBottomNode(PredSU);
344 /// releasePredecessors - Call releasePred on each of SU's predecessors.
345 void ScheduleDAGMI::releasePredecessors(SUnit *SU) {
346 for (SUnit::pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
348 releasePred(SU, &*I);
352 void ScheduleDAGMI::moveInstruction(MachineInstr *MI,
353 MachineBasicBlock::iterator InsertPos) {
354 // Advance RegionBegin if the first instruction moves down.
355 if (&*RegionBegin == MI)
358 // Update the instruction stream.
359 BB->splice(InsertPos, BB, MI);
361 // Update LiveIntervals
364 // Recede RegionBegin if an instruction moves above the first.
365 if (RegionBegin == InsertPos)
369 bool ScheduleDAGMI::checkSchedLimit() {
371 if (NumInstrsScheduled == MISchedCutoff && MISchedCutoff != ~0U) {
372 CurrentTop = CurrentBottom;
375 ++NumInstrsScheduled;
380 /// enterRegion - Called back from MachineScheduler::runOnMachineFunction after
381 /// crossing a scheduling boundary. [begin, end) includes all instructions in
382 /// the region, including the boundary itself and single-instruction regions
383 /// that don't get scheduled.
384 void ScheduleDAGMI::enterRegion(MachineBasicBlock *bb,
385 MachineBasicBlock::iterator begin,
386 MachineBasicBlock::iterator end,
389 ScheduleDAGInstrs::enterRegion(bb, begin, end, endcount);
391 // For convenience remember the end of the liveness region.
393 (RegionEnd == bb->end()) ? RegionEnd : llvm::next(RegionEnd);
396 // Setup the register pressure trackers for the top scheduled top and bottom
397 // scheduled regions.
398 void ScheduleDAGMI::initRegPressure() {
399 TopRPTracker.init(&MF, RegClassInfo, LIS, BB, RegionBegin);
400 BotRPTracker.init(&MF, RegClassInfo, LIS, BB, LiveRegionEnd);
402 // Close the RPTracker to finalize live ins.
403 RPTracker.closeRegion();
405 DEBUG(RPTracker.getPressure().dump(TRI));
407 // Initialize the live ins and live outs.
408 TopRPTracker.addLiveRegs(RPTracker.getPressure().LiveInRegs);
409 BotRPTracker.addLiveRegs(RPTracker.getPressure().LiveOutRegs);
411 // Close one end of the tracker so we can call
412 // getMaxUpward/DownwardPressureDelta before advancing across any
413 // instructions. This converts currently live regs into live ins/outs.
414 TopRPTracker.closeTop();
415 BotRPTracker.closeBottom();
417 // Account for liveness generated by the region boundary.
418 if (LiveRegionEnd != RegionEnd)
419 BotRPTracker.recede();
421 assert(BotRPTracker.getPos() == RegionEnd && "Can't find the region bottom");
423 // Cache the list of excess pressure sets in this region. This will also track
424 // the max pressure in the scheduled code for these sets.
425 RegionCriticalPSets.clear();
426 std::vector<unsigned> RegionPressure = RPTracker.getPressure().MaxSetPressure;
427 for (unsigned i = 0, e = RegionPressure.size(); i < e; ++i) {
428 unsigned Limit = TRI->getRegPressureSetLimit(i);
429 DEBUG(dbgs() << TRI->getRegPressureSetName(i)
431 << " Actual " << RegionPressure[i] << "\n");
432 if (RegionPressure[i] > Limit)
433 RegionCriticalPSets.push_back(PressureElement(i, 0));
435 DEBUG(dbgs() << "Excess PSets: ";
436 for (unsigned i = 0, e = RegionCriticalPSets.size(); i != e; ++i)
437 dbgs() << TRI->getRegPressureSetName(
438 RegionCriticalPSets[i].PSetID) << " ";
442 // FIXME: When the pressure tracker deals in pressure differences then we won't
443 // iterate over all RegionCriticalPSets[i].
445 updateScheduledPressure(std::vector<unsigned> NewMaxPressure) {
446 for (unsigned i = 0, e = RegionCriticalPSets.size(); i < e; ++i) {
447 unsigned ID = RegionCriticalPSets[i].PSetID;
448 int &MaxUnits = RegionCriticalPSets[i].UnitIncrease;
449 if ((int)NewMaxPressure[ID] > MaxUnits)
450 MaxUnits = NewMaxPressure[ID];
454 // Release all DAG roots for scheduling.
455 void ScheduleDAGMI::releaseRoots() {
456 SmallVector<SUnit*, 16> BotRoots;
458 for (std::vector<SUnit>::iterator
459 I = SUnits.begin(), E = SUnits.end(); I != E; ++I) {
460 // A SUnit is ready to top schedule if it has no predecessors.
461 if (I->Preds.empty())
462 SchedImpl->releaseTopNode(&(*I));
463 // A SUnit is ready to bottom schedule if it has no successors.
464 if (I->Succs.empty())
465 BotRoots.push_back(&(*I));
467 // Release bottom roots in reverse order so the higher priority nodes appear
468 // first. This is more natural and slightly more efficient.
469 for (SmallVectorImpl<SUnit*>::const_reverse_iterator
470 I = BotRoots.rbegin(), E = BotRoots.rend(); I != E; ++I)
471 SchedImpl->releaseBottomNode(*I);
474 /// schedule - Called back from MachineScheduler::runOnMachineFunction
475 /// after setting up the current scheduling region. [RegionBegin, RegionEnd)
476 /// only includes instructions that have DAG nodes, not scheduling boundaries.
478 /// This is a skeletal driver, with all the functionality pushed into helpers,
479 /// so that it can be easilly extended by experimental schedulers. Generally,
480 /// implementing MachineSchedStrategy should be sufficient to implement a new
481 /// scheduling algorithm. However, if a scheduler further subclasses
482 /// ScheduleDAGMI then it will want to override this virtual method in order to
483 /// update any specialized state.
484 void ScheduleDAGMI::schedule() {
485 buildDAGWithRegPressure();
487 DEBUG(for (unsigned su = 0, e = SUnits.size(); su != e; ++su)
488 SUnits[su].dumpAll(this));
490 if (ViewMISchedDAGs) viewGraph();
494 bool IsTopNode = false;
495 while (SUnit *SU = SchedImpl->pickNode(IsTopNode)) {
496 if (!checkSchedLimit())
499 scheduleMI(SU, IsTopNode);
501 updateQueues(SU, IsTopNode);
503 assert(CurrentTop == CurrentBottom && "Nonempty unscheduled zone.");
508 /// Build the DAG and setup three register pressure trackers.
509 void ScheduleDAGMI::buildDAGWithRegPressure() {
510 // Initialize the register pressure tracker used by buildSchedGraph.
511 RPTracker.init(&MF, RegClassInfo, LIS, BB, LiveRegionEnd);
513 // Account for liveness generate by the region boundary.
514 if (LiveRegionEnd != RegionEnd)
517 // Build the DAG, and compute current register pressure.
518 buildSchedGraph(AA, &RPTracker);
519 if (ViewMISchedDAGs) viewGraph();
521 // Initialize top/bottom trackers after computing region pressure.
525 /// Identify DAG roots and setup scheduler queues.
526 void ScheduleDAGMI::initQueues() {
527 // Initialize the strategy before modifying the DAG.
528 SchedImpl->initialize(this);
530 // Release edges from the special Entry node or to the special Exit node.
531 releaseSuccessors(&EntrySU);
532 releasePredecessors(&ExitSU);
534 // Release all DAG roots for scheduling.
537 CurrentTop = nextIfDebug(RegionBegin, RegionEnd);
538 CurrentBottom = RegionEnd;
541 /// Move an instruction and update register pressure.
542 void ScheduleDAGMI::scheduleMI(SUnit *SU, bool IsTopNode) {
543 // Move the instruction to its new location in the instruction stream.
544 MachineInstr *MI = SU->getInstr();
547 assert(SU->isTopReady() && "node still has unscheduled dependencies");
548 if (&*CurrentTop == MI)
549 CurrentTop = nextIfDebug(++CurrentTop, CurrentBottom);
551 moveInstruction(MI, CurrentTop);
552 TopRPTracker.setPos(MI);
555 // Update top scheduled pressure.
556 TopRPTracker.advance();
557 assert(TopRPTracker.getPos() == CurrentTop && "out of sync");
558 updateScheduledPressure(TopRPTracker.getPressure().MaxSetPressure);
561 assert(SU->isBottomReady() && "node still has unscheduled dependencies");
562 MachineBasicBlock::iterator priorII =
563 priorNonDebug(CurrentBottom, CurrentTop);
565 CurrentBottom = priorII;
567 if (&*CurrentTop == MI) {
568 CurrentTop = nextIfDebug(++CurrentTop, priorII);
569 TopRPTracker.setPos(CurrentTop);
571 moveInstruction(MI, CurrentBottom);
574 // Update bottom scheduled pressure.
575 BotRPTracker.recede();
576 assert(BotRPTracker.getPos() == CurrentBottom && "out of sync");
577 updateScheduledPressure(BotRPTracker.getPressure().MaxSetPressure);
581 /// Update scheduler queues after scheduling an instruction.
582 void ScheduleDAGMI::updateQueues(SUnit *SU, bool IsTopNode) {
583 // Release dependent instructions for scheduling.
585 releaseSuccessors(SU);
587 releasePredecessors(SU);
589 SU->isScheduled = true;
591 // Notify the scheduling strategy after updating the DAG.
592 SchedImpl->schedNode(SU, IsTopNode);
595 /// Reinsert any remaining debug_values, just like the PostRA scheduler.
596 void ScheduleDAGMI::placeDebugValues() {
597 // If first instruction was a DBG_VALUE then put it back.
599 BB->splice(RegionBegin, BB, FirstDbgValue);
600 RegionBegin = FirstDbgValue;
603 for (std::vector<std::pair<MachineInstr *, MachineInstr *> >::iterator
604 DI = DbgValues.end(), DE = DbgValues.begin(); DI != DE; --DI) {
605 std::pair<MachineInstr *, MachineInstr *> P = *prior(DI);
606 MachineInstr *DbgValue = P.first;
607 MachineBasicBlock::iterator OrigPrevMI = P.second;
608 BB->splice(++OrigPrevMI, BB, DbgValue);
609 if (OrigPrevMI == llvm::prior(RegionEnd))
610 RegionEnd = DbgValue;
613 FirstDbgValue = NULL;
616 //===----------------------------------------------------------------------===//
617 // ConvergingScheduler - Implementation of the standard MachineSchedStrategy.
618 //===----------------------------------------------------------------------===//
621 /// ConvergingScheduler shrinks the unscheduled zone using heuristics to balance
623 class ConvergingScheduler : public MachineSchedStrategy {
625 /// Store the state used by ConvergingScheduler heuristics, required for the
626 /// lifetime of one invocation of pickNode().
627 struct SchedCandidate {
628 // The best SUnit candidate.
631 // Register pressure values for the best candidate.
632 RegPressureDelta RPDelta;
634 SchedCandidate(): SU(NULL) {}
636 /// Represent the type of SchedCandidate found within a single queue.
638 NoCand, NodeOrder, SingleExcess, SingleCritical, SingleMax, MultiPressure };
640 /// Each Scheduling boundary is associated with ready queues. It tracks the
641 /// current cycle in whichever direction at has moved, and maintains the state
642 /// of "hazards" and other interlocks at the current cycle.
643 struct SchedBoundary {
646 ReadyQueue Available;
650 ScheduleHazardRecognizer *HazardRec;
655 /// MinReadyCycle - Cycle of the soonest available instruction.
656 unsigned MinReadyCycle;
658 // Remember the greatest min operand latency.
659 unsigned MaxMinLatency;
661 /// Pending queues extend the ready queues with the same ID and the
663 SchedBoundary(unsigned ID, const Twine &Name):
664 DAG(0), Available(ID, Name+".A"),
665 Pending(ID << ConvergingScheduler::LogMaxQID, Name+".P"),
666 CheckPending(false), HazardRec(0), CurrCycle(0), IssueCount(0),
667 MinReadyCycle(UINT_MAX), MaxMinLatency(0) {}
669 ~SchedBoundary() { delete HazardRec; }
672 return Available.getID() == ConvergingScheduler::TopQID;
675 bool checkHazard(SUnit *SU);
677 void releaseNode(SUnit *SU, unsigned ReadyCycle);
681 void bumpNode(SUnit *SU);
683 void releasePending();
685 void removeReady(SUnit *SU);
687 SUnit *pickOnlyChoice();
691 const TargetRegisterInfo *TRI;
693 // State of the top and bottom scheduled instruction boundaries.
698 /// SUnit::NodeQueueId: 0 (none), 1 (top), 2 (bot), 3 (both)
705 ConvergingScheduler():
706 DAG(0), TRI(0), Top(TopQID, "TopQ"), Bot(BotQID, "BotQ") {}
708 virtual void initialize(ScheduleDAGMI *dag);
710 virtual SUnit *pickNode(bool &IsTopNode);
712 virtual void schedNode(SUnit *SU, bool IsTopNode);
714 virtual void releaseTopNode(SUnit *SU);
716 virtual void releaseBottomNode(SUnit *SU);
719 SUnit *pickNodeBidrectional(bool &IsTopNode);
721 CandResult pickNodeFromQueue(ReadyQueue &Q,
722 const RegPressureTracker &RPTracker,
723 SchedCandidate &Candidate);
725 void traceCandidate(const char *Label, const ReadyQueue &Q, SUnit *SU,
726 PressureElement P = PressureElement());
731 void ConvergingScheduler::initialize(ScheduleDAGMI *dag) {
737 // Initialize the HazardRecognizers.
738 const TargetMachine &TM = DAG->MF.getTarget();
739 const InstrItineraryData *Itin = TM.getInstrItineraryData();
740 Top.HazardRec = TM.getInstrInfo()->CreateTargetMIHazardRecognizer(Itin, DAG);
741 Bot.HazardRec = TM.getInstrInfo()->CreateTargetMIHazardRecognizer(Itin, DAG);
743 assert((!ForceTopDown || !ForceBottomUp) &&
744 "-misched-topdown incompatible with -misched-bottomup");
747 void ConvergingScheduler::releaseTopNode(SUnit *SU) {
751 for (SUnit::succ_iterator I = SU->Preds.begin(), E = SU->Preds.end();
753 unsigned PredReadyCycle = I->getSUnit()->TopReadyCycle;
754 unsigned MinLatency = I->getMinLatency();
756 Top.MaxMinLatency = std::max(MinLatency, Top.MaxMinLatency);
758 if (SU->TopReadyCycle < PredReadyCycle + MinLatency)
759 SU->TopReadyCycle = PredReadyCycle + MinLatency;
761 Top.releaseNode(SU, SU->TopReadyCycle);
764 void ConvergingScheduler::releaseBottomNode(SUnit *SU) {
768 assert(SU->getInstr() && "Scheduled SUnit must have instr");
770 for (SUnit::succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
772 unsigned SuccReadyCycle = I->getSUnit()->BotReadyCycle;
773 unsigned MinLatency = I->getMinLatency();
775 Bot.MaxMinLatency = std::max(MinLatency, Bot.MaxMinLatency);
777 if (SU->BotReadyCycle < SuccReadyCycle + MinLatency)
778 SU->BotReadyCycle = SuccReadyCycle + MinLatency;
780 Bot.releaseNode(SU, SU->BotReadyCycle);
783 /// Does this SU have a hazard within the current instruction group.
785 /// The scheduler supports two modes of hazard recognition. The first is the
786 /// ScheduleHazardRecognizer API. It is a fully general hazard recognizer that
787 /// supports highly complicated in-order reservation tables
788 /// (ScoreboardHazardRecognizer) and arbitraty target-specific logic.
790 /// The second is a streamlined mechanism that checks for hazards based on
791 /// simple counters that the scheduler itself maintains. It explicitly checks
792 /// for instruction dispatch limitations, including the number of micro-ops that
793 /// can dispatch per cycle.
795 /// TODO: Also check whether the SU must start a new group.
796 bool ConvergingScheduler::SchedBoundary::checkHazard(SUnit *SU) {
797 if (HazardRec->isEnabled())
798 return HazardRec->getHazardType(SU) != ScheduleHazardRecognizer::NoHazard;
800 if (IssueCount + DAG->getNumMicroOps(SU->getInstr()) > DAG->getIssueWidth())
806 void ConvergingScheduler::SchedBoundary::releaseNode(SUnit *SU,
807 unsigned ReadyCycle) {
808 if (ReadyCycle < MinReadyCycle)
809 MinReadyCycle = ReadyCycle;
811 // Check for interlocks first. For the purpose of other heuristics, an
812 // instruction that cannot issue appears as if it's not in the ReadyQueue.
813 if (ReadyCycle > CurrCycle || checkHazard(SU))
819 /// Move the boundary of scheduled code by one cycle.
820 void ConvergingScheduler::SchedBoundary::bumpCycle() {
821 unsigned Width = DAG->getIssueWidth();
822 IssueCount = (IssueCount <= Width) ? 0 : IssueCount - Width;
824 assert(MinReadyCycle < UINT_MAX && "MinReadyCycle uninitialized");
825 unsigned NextCycle = std::max(CurrCycle + 1, MinReadyCycle);
827 if (!HazardRec->isEnabled()) {
828 // Bypass HazardRec virtual calls.
829 CurrCycle = NextCycle;
832 // Bypass getHazardType calls in case of long latency.
833 for (; CurrCycle != NextCycle; ++CurrCycle) {
835 HazardRec->AdvanceCycle();
837 HazardRec->RecedeCycle();
842 DEBUG(dbgs() << "*** " << Available.getName() << " cycle "
843 << CurrCycle << '\n');
846 /// Move the boundary of scheduled code by one SUnit.
847 void ConvergingScheduler::SchedBoundary::bumpNode(SUnit *SU) {
848 // Update the reservation table.
849 if (HazardRec->isEnabled()) {
850 if (!isTop() && SU->isCall) {
851 // Calls are scheduled with their preceding instructions. For bottom-up
852 // scheduling, clear the pipeline state before emitting.
855 HazardRec->EmitInstruction(SU);
857 // Check the instruction group dispatch limit.
858 // TODO: Check if this SU must end a dispatch group.
859 IssueCount += DAG->getNumMicroOps(SU->getInstr());
860 if (IssueCount >= DAG->getIssueWidth()) {
861 DEBUG(dbgs() << "*** Max instrs at cycle " << CurrCycle << '\n');
866 /// Release pending ready nodes in to the available queue. This makes them
867 /// visible to heuristics.
868 void ConvergingScheduler::SchedBoundary::releasePending() {
869 // If the available queue is empty, it is safe to reset MinReadyCycle.
870 if (Available.empty())
871 MinReadyCycle = UINT_MAX;
873 // Check to see if any of the pending instructions are ready to issue. If
874 // so, add them to the available queue.
875 for (unsigned i = 0, e = Pending.size(); i != e; ++i) {
876 SUnit *SU = *(Pending.begin()+i);
877 unsigned ReadyCycle = isTop() ? SU->TopReadyCycle : SU->BotReadyCycle;
879 if (ReadyCycle < MinReadyCycle)
880 MinReadyCycle = ReadyCycle;
882 if (ReadyCycle > CurrCycle)
889 Pending.remove(Pending.begin()+i);
892 CheckPending = false;
895 /// Remove SU from the ready set for this boundary.
896 void ConvergingScheduler::SchedBoundary::removeReady(SUnit *SU) {
897 if (Available.isInQueue(SU))
898 Available.remove(Available.find(SU));
900 assert(Pending.isInQueue(SU) && "bad ready count");
901 Pending.remove(Pending.find(SU));
905 /// If this queue only has one ready candidate, return it. As a side effect,
906 /// advance the cycle until at least one node is ready. If multiple instructions
907 /// are ready, return NULL.
908 SUnit *ConvergingScheduler::SchedBoundary::pickOnlyChoice() {
912 for (unsigned i = 0; Available.empty(); ++i) {
913 assert(i <= (HazardRec->getMaxLookAhead() + MaxMinLatency) &&
914 "permanent hazard"); (void)i;
918 if (Available.size() == 1)
919 return *Available.begin();
924 void ConvergingScheduler::traceCandidate(const char *Label, const ReadyQueue &Q,
925 SUnit *SU, PressureElement P) {
926 dbgs() << Label << " " << Q.getName() << " ";
928 dbgs() << TRI->getRegPressureSetName(P.PSetID) << ":" << P.UnitIncrease
936 /// pickNodeFromQueue helper that returns true if the LHS reg pressure effect is
937 /// more desirable than RHS from scheduling standpoint.
938 static bool compareRPDelta(const RegPressureDelta &LHS,
939 const RegPressureDelta &RHS) {
940 // Compare each component of pressure in decreasing order of importance
941 // without checking if any are valid. Invalid PressureElements are assumed to
942 // have UnitIncrease==0, so are neutral.
944 // Avoid increasing the max critical pressure in the scheduled region.
945 if (LHS.Excess.UnitIncrease != RHS.Excess.UnitIncrease)
946 return LHS.Excess.UnitIncrease < RHS.Excess.UnitIncrease;
948 // Avoid increasing the max critical pressure in the scheduled region.
949 if (LHS.CriticalMax.UnitIncrease != RHS.CriticalMax.UnitIncrease)
950 return LHS.CriticalMax.UnitIncrease < RHS.CriticalMax.UnitIncrease;
952 // Avoid increasing the max pressure of the entire region.
953 if (LHS.CurrentMax.UnitIncrease != RHS.CurrentMax.UnitIncrease)
954 return LHS.CurrentMax.UnitIncrease < RHS.CurrentMax.UnitIncrease;
959 /// Pick the best candidate from the top queue.
961 /// TODO: getMaxPressureDelta results can be mostly cached for each SUnit during
962 /// DAG building. To adjust for the current scheduling location we need to
963 /// maintain the number of vreg uses remaining to be top-scheduled.
964 ConvergingScheduler::CandResult ConvergingScheduler::
965 pickNodeFromQueue(ReadyQueue &Q, const RegPressureTracker &RPTracker,
966 SchedCandidate &Candidate) {
969 // getMaxPressureDelta temporarily modifies the tracker.
970 RegPressureTracker &TempTracker = const_cast<RegPressureTracker&>(RPTracker);
972 // BestSU remains NULL if no top candidates beat the best existing candidate.
973 CandResult FoundCandidate = NoCand;
974 for (ReadyQueue::iterator I = Q.begin(), E = Q.end(); I != E; ++I) {
975 RegPressureDelta RPDelta;
976 TempTracker.getMaxPressureDelta((*I)->getInstr(), RPDelta,
977 DAG->getRegionCriticalPSets(),
978 DAG->getRegPressure().MaxSetPressure);
980 // Initialize the candidate if needed.
983 Candidate.RPDelta = RPDelta;
984 FoundCandidate = NodeOrder;
987 // Avoid exceeding the target's limit.
988 if (RPDelta.Excess.UnitIncrease < Candidate.RPDelta.Excess.UnitIncrease) {
989 DEBUG(traceCandidate("ECAND", Q, *I, RPDelta.Excess));
991 Candidate.RPDelta = RPDelta;
992 FoundCandidate = SingleExcess;
995 if (RPDelta.Excess.UnitIncrease > Candidate.RPDelta.Excess.UnitIncrease)
997 if (FoundCandidate == SingleExcess)
998 FoundCandidate = MultiPressure;
1000 // Avoid increasing the max critical pressure in the scheduled region.
1001 if (RPDelta.CriticalMax.UnitIncrease
1002 < Candidate.RPDelta.CriticalMax.UnitIncrease) {
1003 DEBUG(traceCandidate("PCAND", Q, *I, RPDelta.CriticalMax));
1005 Candidate.RPDelta = RPDelta;
1006 FoundCandidate = SingleCritical;
1009 if (RPDelta.CriticalMax.UnitIncrease
1010 > Candidate.RPDelta.CriticalMax.UnitIncrease)
1012 if (FoundCandidate == SingleCritical)
1013 FoundCandidate = MultiPressure;
1015 // Avoid increasing the max pressure of the entire region.
1016 if (RPDelta.CurrentMax.UnitIncrease
1017 < Candidate.RPDelta.CurrentMax.UnitIncrease) {
1018 DEBUG(traceCandidate("MCAND", Q, *I, RPDelta.CurrentMax));
1020 Candidate.RPDelta = RPDelta;
1021 FoundCandidate = SingleMax;
1024 if (RPDelta.CurrentMax.UnitIncrease
1025 > Candidate.RPDelta.CurrentMax.UnitIncrease)
1027 if (FoundCandidate == SingleMax)
1028 FoundCandidate = MultiPressure;
1030 // Fall through to original instruction order.
1031 // Only consider node order if Candidate was chosen from this Q.
1032 if (FoundCandidate == NoCand)
1035 if ((Q.getID() == TopQID && (*I)->NodeNum < Candidate.SU->NodeNum)
1036 || (Q.getID() == BotQID && (*I)->NodeNum > Candidate.SU->NodeNum)) {
1037 DEBUG(traceCandidate("NCAND", Q, *I));
1039 Candidate.RPDelta = RPDelta;
1040 FoundCandidate = NodeOrder;
1043 return FoundCandidate;
1046 /// Pick the best candidate node from either the top or bottom queue.
1047 SUnit *ConvergingScheduler::pickNodeBidrectional(bool &IsTopNode) {
1048 // Schedule as far as possible in the direction of no choice. This is most
1049 // efficient, but also provides the best heuristics for CriticalPSets.
1050 if (SUnit *SU = Bot.pickOnlyChoice()) {
1054 if (SUnit *SU = Top.pickOnlyChoice()) {
1058 SchedCandidate BotCand;
1059 // Prefer bottom scheduling when heuristics are silent.
1060 CandResult BotResult = pickNodeFromQueue(Bot.Available,
1061 DAG->getBotRPTracker(), BotCand);
1062 assert(BotResult != NoCand && "failed to find the first candidate");
1064 // If either Q has a single candidate that provides the least increase in
1065 // Excess pressure, we can immediately schedule from that Q.
1067 // RegionCriticalPSets summarizes the pressure within the scheduled region and
1068 // affects picking from either Q. If scheduling in one direction must
1069 // increase pressure for one of the excess PSets, then schedule in that
1070 // direction first to provide more freedom in the other direction.
1071 if (BotResult == SingleExcess || BotResult == SingleCritical) {
1075 // Check if the top Q has a better candidate.
1076 SchedCandidate TopCand;
1077 CandResult TopResult = pickNodeFromQueue(Top.Available,
1078 DAG->getTopRPTracker(), TopCand);
1079 assert(TopResult != NoCand && "failed to find the first candidate");
1081 if (TopResult == SingleExcess || TopResult == SingleCritical) {
1085 // If either Q has a single candidate that minimizes pressure above the
1086 // original region's pressure pick it.
1087 if (BotResult == SingleMax) {
1091 if (TopResult == SingleMax) {
1095 // Check for a salient pressure difference and pick the best from either side.
1096 if (compareRPDelta(TopCand.RPDelta, BotCand.RPDelta)) {
1100 // Otherwise prefer the bottom candidate in node order.
1105 /// Pick the best node to balance the schedule. Implements MachineSchedStrategy.
1106 SUnit *ConvergingScheduler::pickNode(bool &IsTopNode) {
1107 if (DAG->top() == DAG->bottom()) {
1108 assert(Top.Available.empty() && Top.Pending.empty() &&
1109 Bot.Available.empty() && Bot.Pending.empty() && "ReadyQ garbage");
1114 SU = Top.pickOnlyChoice();
1116 SchedCandidate TopCand;
1117 CandResult TopResult =
1118 pickNodeFromQueue(Top.Available, DAG->getTopRPTracker(), TopCand);
1119 assert(TopResult != NoCand && "failed to find the first candidate");
1125 else if (ForceBottomUp) {
1126 SU = Bot.pickOnlyChoice();
1128 SchedCandidate BotCand;
1129 CandResult BotResult =
1130 pickNodeFromQueue(Bot.Available, DAG->getBotRPTracker(), BotCand);
1131 assert(BotResult != NoCand && "failed to find the first candidate");
1138 SU = pickNodeBidrectional(IsTopNode);
1140 if (SU->isTopReady())
1141 Top.removeReady(SU);
1142 if (SU->isBottomReady())
1143 Bot.removeReady(SU);
1145 DEBUG(dbgs() << "*** " << (IsTopNode ? "Top" : "Bottom")
1146 << " Scheduling Instruction in cycle "
1147 << (IsTopNode ? Top.CurrCycle : Bot.CurrCycle) << '\n';
1152 /// Update the scheduler's state after scheduling a node. This is the same node
1153 /// that was just returned by pickNode(). However, ScheduleDAGMI needs to update
1154 /// it's state based on the current cycle before MachineSchedStrategy does.
1155 void ConvergingScheduler::schedNode(SUnit *SU, bool IsTopNode) {
1157 SU->TopReadyCycle = Top.CurrCycle;
1161 SU->BotReadyCycle = Bot.CurrCycle;
1166 /// Create the standard converging machine scheduler. This will be used as the
1167 /// default scheduler if the target does not set a default.
1168 static ScheduleDAGInstrs *createConvergingSched(MachineSchedContext *C) {
1169 assert((!ForceTopDown || !ForceBottomUp) &&
1170 "-misched-topdown incompatible with -misched-bottomup");
1171 return new ScheduleDAGMI(C, new ConvergingScheduler());
1173 static MachineSchedRegistry
1174 ConvergingSchedRegistry("converge", "Standard converging scheduler.",
1175 createConvergingSched);
1177 //===----------------------------------------------------------------------===//
1178 // Machine Instruction Shuffler for Correctness Testing
1179 //===----------------------------------------------------------------------===//
1183 /// Apply a less-than relation on the node order, which corresponds to the
1184 /// instruction order prior to scheduling. IsReverse implements greater-than.
1185 template<bool IsReverse>
1187 bool operator()(SUnit *A, SUnit *B) const {
1189 return A->NodeNum > B->NodeNum;
1191 return A->NodeNum < B->NodeNum;
1195 /// Reorder instructions as much as possible.
1196 class InstructionShuffler : public MachineSchedStrategy {
1200 // Using a less-than relation (SUnitOrder<false>) for the TopQ priority
1201 // gives nodes with a higher number higher priority causing the latest
1202 // instructions to be scheduled first.
1203 PriorityQueue<SUnit*, std::vector<SUnit*>, SUnitOrder<false> >
1205 // When scheduling bottom-up, use greater-than as the queue priority.
1206 PriorityQueue<SUnit*, std::vector<SUnit*>, SUnitOrder<true> >
1209 InstructionShuffler(bool alternate, bool topdown)
1210 : IsAlternating(alternate), IsTopDown(topdown) {}
1212 virtual void initialize(ScheduleDAGMI *) {
1217 /// Implement MachineSchedStrategy interface.
1218 /// -----------------------------------------
1220 virtual SUnit *pickNode(bool &IsTopNode) {
1224 if (TopQ.empty()) return NULL;
1227 } while (SU->isScheduled);
1232 if (BottomQ.empty()) return NULL;
1235 } while (SU->isScheduled);
1239 IsTopDown = !IsTopDown;
1243 virtual void schedNode(SUnit *SU, bool IsTopNode) {}
1245 virtual void releaseTopNode(SUnit *SU) {
1248 virtual void releaseBottomNode(SUnit *SU) {
1254 static ScheduleDAGInstrs *createInstructionShuffler(MachineSchedContext *C) {
1255 bool Alternate = !ForceTopDown && !ForceBottomUp;
1256 bool TopDown = !ForceBottomUp;
1257 assert((TopDown || !ForceTopDown) &&
1258 "-misched-topdown incompatible with -misched-bottomup");
1259 return new ScheduleDAGMI(C, new InstructionShuffler(Alternate, TopDown));
1261 static MachineSchedRegistry ShufflerRegistry(
1262 "shuffle", "Shuffle machine instructions alternating directions",
1263 createInstructionShuffler);