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 "RegisterClassInfo.h"
18 #include "RegisterPressure.h"
19 #include "llvm/CodeGen/LiveIntervalAnalysis.h"
20 #include "llvm/CodeGen/MachineScheduler.h"
21 #include "llvm/CodeGen/Passes.h"
22 #include "llvm/CodeGen/ScheduleDAGInstrs.h"
23 #include "llvm/CodeGen/ScheduleHazardRecognizer.h"
24 #include "llvm/Analysis/AliasAnalysis.h"
25 #include "llvm/Target/TargetInstrInfo.h"
26 #include "llvm/Support/CommandLine.h"
27 #include "llvm/Support/Debug.h"
28 #include "llvm/Support/ErrorHandling.h"
29 #include "llvm/Support/raw_ostream.h"
30 #include "llvm/ADT/OwningPtr.h"
31 #include "llvm/ADT/PriorityQueue.h"
37 static cl::opt<bool> ForceTopDown("misched-topdown", cl::Hidden,
38 cl::desc("Force top-down list scheduling"));
39 static cl::opt<bool> ForceBottomUp("misched-bottomup", cl::Hidden,
40 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 //===----------------------------------------------------------------------===//
53 // Machine Instruction Scheduling Pass and Registry
54 //===----------------------------------------------------------------------===//
56 MachineSchedContext::MachineSchedContext():
57 MF(0), MLI(0), MDT(0), PassConfig(0), AA(0), LIS(0) {
58 RegClassInfo = new RegisterClassInfo();
61 MachineSchedContext::~MachineSchedContext() {
66 /// MachineScheduler runs after coalescing and before register allocation.
67 class MachineScheduler : public MachineSchedContext,
68 public MachineFunctionPass {
72 virtual void getAnalysisUsage(AnalysisUsage &AU) const;
74 virtual void releaseMemory() {}
76 virtual bool runOnMachineFunction(MachineFunction&);
78 virtual void print(raw_ostream &O, const Module* = 0) const;
80 static char ID; // Class identification, replacement for typeinfo
84 char MachineScheduler::ID = 0;
86 char &llvm::MachineSchedulerID = MachineScheduler::ID;
88 INITIALIZE_PASS_BEGIN(MachineScheduler, "misched",
89 "Machine Instruction Scheduler", false, false)
90 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
91 INITIALIZE_PASS_DEPENDENCY(SlotIndexes)
92 INITIALIZE_PASS_DEPENDENCY(LiveIntervals)
93 INITIALIZE_PASS_END(MachineScheduler, "misched",
94 "Machine Instruction Scheduler", false, false)
96 MachineScheduler::MachineScheduler()
97 : MachineFunctionPass(ID) {
98 initializeMachineSchedulerPass(*PassRegistry::getPassRegistry());
101 void MachineScheduler::getAnalysisUsage(AnalysisUsage &AU) const {
102 AU.setPreservesCFG();
103 AU.addRequiredID(MachineDominatorsID);
104 AU.addRequired<MachineLoopInfo>();
105 AU.addRequired<AliasAnalysis>();
106 AU.addRequired<TargetPassConfig>();
107 AU.addRequired<SlotIndexes>();
108 AU.addPreserved<SlotIndexes>();
109 AU.addRequired<LiveIntervals>();
110 AU.addPreserved<LiveIntervals>();
111 MachineFunctionPass::getAnalysisUsage(AU);
114 MachinePassRegistry MachineSchedRegistry::Registry;
116 /// A dummy default scheduler factory indicates whether the scheduler
117 /// is overridden on the command line.
118 static ScheduleDAGInstrs *useDefaultMachineSched(MachineSchedContext *C) {
122 /// MachineSchedOpt allows command line selection of the scheduler.
123 static cl::opt<MachineSchedRegistry::ScheduleDAGCtor, false,
124 RegisterPassParser<MachineSchedRegistry> >
125 MachineSchedOpt("misched",
126 cl::init(&useDefaultMachineSched), cl::Hidden,
127 cl::desc("Machine instruction scheduler to use"));
129 static MachineSchedRegistry
130 DefaultSchedRegistry("default", "Use the target's default scheduler choice.",
131 useDefaultMachineSched);
133 /// Forward declare the standard machine scheduler. This will be used as the
134 /// default scheduler if the target does not set a default.
135 static ScheduleDAGInstrs *createConvergingSched(MachineSchedContext *C);
138 /// Decrement this iterator until reaching the top or a non-debug instr.
139 static MachineBasicBlock::iterator
140 priorNonDebug(MachineBasicBlock::iterator I, MachineBasicBlock::iterator Beg) {
141 assert(I != Beg && "reached the top of the region, cannot decrement");
143 if (!I->isDebugValue())
149 /// If this iterator is a debug value, increment until reaching the End or a
150 /// non-debug instruction.
151 static MachineBasicBlock::iterator
152 nextIfDebug(MachineBasicBlock::iterator I, MachineBasicBlock::iterator End) {
153 for(; I != End; ++I) {
154 if (!I->isDebugValue())
160 /// Top-level MachineScheduler pass driver.
162 /// Visit blocks in function order. Divide each block into scheduling regions
163 /// and visit them bottom-up. Visiting regions bottom-up is not required, but is
164 /// consistent with the DAG builder, which traverses the interior of the
165 /// scheduling regions bottom-up.
167 /// This design avoids exposing scheduling boundaries to the DAG builder,
168 /// simplifying the DAG builder's support for "special" target instructions.
169 /// At the same time the design allows target schedulers to operate across
170 /// scheduling boundaries, for example to bundle the boudary instructions
171 /// without reordering them. This creates complexity, because the target
172 /// scheduler must update the RegionBegin and RegionEnd positions cached by
173 /// ScheduleDAGInstrs whenever adding or removing instructions. A much simpler
174 /// design would be to split blocks at scheduling boundaries, but LLVM has a
175 /// general bias against block splitting purely for implementation simplicity.
176 bool MachineScheduler::runOnMachineFunction(MachineFunction &mf) {
177 DEBUG(dbgs() << "Before MISsched:\n"; mf.print(dbgs()));
179 // Initialize the context of the pass.
181 MLI = &getAnalysis<MachineLoopInfo>();
182 MDT = &getAnalysis<MachineDominatorTree>();
183 PassConfig = &getAnalysis<TargetPassConfig>();
184 AA = &getAnalysis<AliasAnalysis>();
186 LIS = &getAnalysis<LiveIntervals>();
187 const TargetInstrInfo *TII = MF->getTarget().getInstrInfo();
189 RegClassInfo->runOnMachineFunction(*MF);
191 // Select the scheduler, or set the default.
192 MachineSchedRegistry::ScheduleDAGCtor Ctor = MachineSchedOpt;
193 if (Ctor == useDefaultMachineSched) {
194 // Get the default scheduler set by the target.
195 Ctor = MachineSchedRegistry::getDefault();
197 Ctor = createConvergingSched;
198 MachineSchedRegistry::setDefault(Ctor);
201 // Instantiate the selected scheduler.
202 OwningPtr<ScheduleDAGInstrs> Scheduler(Ctor(this));
204 // Visit all machine basic blocks.
206 // TODO: Visit blocks in global postorder or postorder within the bottom-up
207 // loop tree. Then we can optionally compute global RegPressure.
208 for (MachineFunction::iterator MBB = MF->begin(), MBBEnd = MF->end();
209 MBB != MBBEnd; ++MBB) {
211 Scheduler->startBlock(MBB);
213 // Break the block into scheduling regions [I, RegionEnd), and schedule each
214 // region as soon as it is discovered. RegionEnd points the the scheduling
215 // boundary at the bottom of the region. The DAG does not include RegionEnd,
216 // but the region does (i.e. the next RegionEnd is above the previous
217 // RegionBegin). If the current block has no terminator then RegionEnd ==
218 // MBB->end() for the bottom region.
220 // The Scheduler may insert instructions during either schedule() or
221 // exitRegion(), even for empty regions. So the local iterators 'I' and
222 // 'RegionEnd' are invalid across these calls.
223 unsigned RemainingCount = MBB->size();
224 for(MachineBasicBlock::iterator RegionEnd = MBB->end();
225 RegionEnd != MBB->begin(); RegionEnd = Scheduler->begin()) {
227 // Avoid decrementing RegionEnd for blocks with no terminator.
228 if (RegionEnd != MBB->end()
229 || TII->isSchedulingBoundary(llvm::prior(RegionEnd), MBB, *MF)) {
231 // Count the boundary instruction.
235 // The next region starts above the previous region. Look backward in the
236 // instruction stream until we find the nearest boundary.
237 MachineBasicBlock::iterator I = RegionEnd;
238 for(;I != MBB->begin(); --I, --RemainingCount) {
239 if (TII->isSchedulingBoundary(llvm::prior(I), MBB, *MF))
242 // Notify the scheduler of the region, even if we may skip scheduling
243 // it. Perhaps it still needs to be bundled.
244 Scheduler->enterRegion(MBB, I, RegionEnd, RemainingCount);
246 // Skip empty scheduling regions (0 or 1 schedulable instructions).
247 if (I == RegionEnd || I == llvm::prior(RegionEnd)) {
248 // Close the current region. Bundle the terminator if needed.
249 // This invalidates 'RegionEnd' and 'I'.
250 Scheduler->exitRegion();
253 DEBUG(dbgs() << "********** MI Scheduling **********\n");
254 DEBUG(dbgs() << MF->getFunction()->getName()
255 << ":BB#" << MBB->getNumber() << "\n From: " << *I << " To: ";
256 if (RegionEnd != MBB->end()) dbgs() << *RegionEnd;
257 else dbgs() << "End";
258 dbgs() << " Remaining: " << RemainingCount << "\n");
260 // Schedule a region: possibly reorder instructions.
261 // This invalidates 'RegionEnd' and 'I'.
262 Scheduler->schedule();
264 // Close the current region.
265 Scheduler->exitRegion();
267 // Scheduling has invalidated the current iterator 'I'. Ask the
268 // scheduler for the top of it's scheduled region.
269 RegionEnd = Scheduler->begin();
271 assert(RemainingCount == 0 && "Instruction count mismatch!");
272 Scheduler->finishBlock();
274 Scheduler->finalizeSchedule();
275 DEBUG(LIS->print(dbgs()));
279 void MachineScheduler::print(raw_ostream &O, const Module* m) const {
283 //===----------------------------------------------------------------------===//
284 // MachineSchedStrategy - Interface to a machine scheduling algorithm.
285 //===----------------------------------------------------------------------===//
290 /// MachineSchedStrategy - Interface used by ScheduleDAGMI to drive the selected
291 /// scheduling algorithm.
293 /// If this works well and targets wish to reuse ScheduleDAGMI, we may expose it
294 /// in ScheduleDAGInstrs.h
295 class MachineSchedStrategy {
297 virtual ~MachineSchedStrategy() {}
299 /// Initialize the strategy after building the DAG for a new region.
300 virtual void initialize(ScheduleDAGMI *DAG) = 0;
302 /// Pick the next node to schedule, or return NULL. Set IsTopNode to true to
303 /// schedule the node at the top of the unscheduled region. Otherwise it will
304 /// be scheduled at the bottom.
305 virtual SUnit *pickNode(bool &IsTopNode) = 0;
307 /// Notify MachineSchedStrategy that ScheduleDAGMI has scheduled a node.
308 virtual void schedNode(SUnit *SU, bool IsTopNode) = 0;
310 /// When all predecessor dependencies have been resolved, free this node for
311 /// top-down scheduling.
312 virtual void releaseTopNode(SUnit *SU) = 0;
313 /// When all successor dependencies have been resolved, free this node for
314 /// bottom-up scheduling.
315 virtual void releaseBottomNode(SUnit *SU) = 0;
319 //===----------------------------------------------------------------------===//
320 // ScheduleDAGMI - Base class for MachineInstr scheduling with LiveIntervals
322 //===----------------------------------------------------------------------===//
325 /// ScheduleDAGMI is an implementation of ScheduleDAGInstrs that schedules
326 /// machine instructions while updating LiveIntervals.
327 class ScheduleDAGMI : public ScheduleDAGInstrs {
329 RegisterClassInfo *RegClassInfo;
330 MachineSchedStrategy *SchedImpl;
332 MachineBasicBlock::iterator LiveRegionEnd;
334 /// Register pressure in this region computed by buildSchedGraph.
335 IntervalPressure RegPressure;
336 RegPressureTracker RPTracker;
338 /// List of pressure sets that exceed the target's pressure limit before
339 /// scheduling, listed in increasing set ID order. Each pressure set is paired
340 /// with its max pressure in the currently scheduled regions.
341 std::vector<PressureElement> RegionCriticalPSets;
343 /// The top of the unscheduled zone.
344 MachineBasicBlock::iterator CurrentTop;
345 IntervalPressure TopPressure;
346 RegPressureTracker TopRPTracker;
348 /// The bottom of the unscheduled zone.
349 MachineBasicBlock::iterator CurrentBottom;
350 IntervalPressure BotPressure;
351 RegPressureTracker BotRPTracker;
353 /// The number of instructions scheduled so far. Used to cut off the
354 /// scheduler at the point determined by misched-cutoff.
355 unsigned NumInstrsScheduled;
357 ScheduleDAGMI(MachineSchedContext *C, MachineSchedStrategy *S):
358 ScheduleDAGInstrs(*C->MF, *C->MLI, *C->MDT, /*IsPostRA=*/false, C->LIS),
359 AA(C->AA), RegClassInfo(C->RegClassInfo), SchedImpl(S),
360 RPTracker(RegPressure), CurrentTop(), TopRPTracker(TopPressure),
361 CurrentBottom(), BotRPTracker(BotPressure), NumInstrsScheduled(0) {}
367 MachineBasicBlock::iterator top() const { return CurrentTop; }
368 MachineBasicBlock::iterator bottom() const { return CurrentBottom; }
370 /// Implement the ScheduleDAGInstrs interface for handling the next scheduling
371 /// region. This covers all instructions in a block, while schedule() may only
373 void enterRegion(MachineBasicBlock *bb,
374 MachineBasicBlock::iterator begin,
375 MachineBasicBlock::iterator end,
378 /// Implement ScheduleDAGInstrs interface for scheduling a sequence of
379 /// reorderable instructions.
382 /// Get current register pressure for the top scheduled instructions.
383 const IntervalPressure &getTopPressure() const { return TopPressure; }
384 const RegPressureTracker &getTopRPTracker() const { return TopRPTracker; }
386 /// Get current register pressure for the bottom scheduled instructions.
387 const IntervalPressure &getBotPressure() const { return BotPressure; }
388 const RegPressureTracker &getBotRPTracker() const { return BotRPTracker; }
390 /// Get register pressure for the entire scheduling region before scheduling.
391 const IntervalPressure &getRegPressure() const { return RegPressure; }
393 const std::vector<PressureElement> &getRegionCriticalPSets() const {
394 return RegionCriticalPSets;
398 void initRegPressure();
399 void updateScheduledPressure(std::vector<unsigned> NewMaxPressure);
401 void moveInstruction(MachineInstr *MI, MachineBasicBlock::iterator InsertPos);
402 bool checkSchedLimit();
406 void releaseSucc(SUnit *SU, SDep *SuccEdge);
407 void releaseSuccessors(SUnit *SU);
408 void releasePred(SUnit *SU, SDep *PredEdge);
409 void releasePredecessors(SUnit *SU);
411 void placeDebugValues();
415 /// ReleaseSucc - Decrement the NumPredsLeft count of a successor. When
416 /// NumPredsLeft reaches zero, release the successor node.
418 /// FIXME: Adjust SuccSU height based on MinLatency.
419 void ScheduleDAGMI::releaseSucc(SUnit *SU, SDep *SuccEdge) {
420 SUnit *SuccSU = SuccEdge->getSUnit();
423 if (SuccSU->NumPredsLeft == 0) {
424 dbgs() << "*** Scheduling failed! ***\n";
426 dbgs() << " has been released too many times!\n";
430 --SuccSU->NumPredsLeft;
431 if (SuccSU->NumPredsLeft == 0 && SuccSU != &ExitSU)
432 SchedImpl->releaseTopNode(SuccSU);
435 /// releaseSuccessors - Call releaseSucc on each of SU's successors.
436 void ScheduleDAGMI::releaseSuccessors(SUnit *SU) {
437 for (SUnit::succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
439 releaseSucc(SU, &*I);
443 /// ReleasePred - Decrement the NumSuccsLeft count of a predecessor. When
444 /// NumSuccsLeft reaches zero, release the predecessor node.
446 /// FIXME: Adjust PredSU height based on MinLatency.
447 void ScheduleDAGMI::releasePred(SUnit *SU, SDep *PredEdge) {
448 SUnit *PredSU = PredEdge->getSUnit();
451 if (PredSU->NumSuccsLeft == 0) {
452 dbgs() << "*** Scheduling failed! ***\n";
454 dbgs() << " has been released too many times!\n";
458 --PredSU->NumSuccsLeft;
459 if (PredSU->NumSuccsLeft == 0 && PredSU != &EntrySU)
460 SchedImpl->releaseBottomNode(PredSU);
463 /// releasePredecessors - Call releasePred on each of SU's predecessors.
464 void ScheduleDAGMI::releasePredecessors(SUnit *SU) {
465 for (SUnit::pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
467 releasePred(SU, &*I);
471 void ScheduleDAGMI::moveInstruction(MachineInstr *MI,
472 MachineBasicBlock::iterator InsertPos) {
473 // Advance RegionBegin if the first instruction moves down.
474 if (&*RegionBegin == MI)
477 // Update the instruction stream.
478 BB->splice(InsertPos, BB, MI);
480 // Update LiveIntervals
483 // Recede RegionBegin if an instruction moves above the first.
484 if (RegionBegin == InsertPos)
488 bool ScheduleDAGMI::checkSchedLimit() {
490 if (NumInstrsScheduled == MISchedCutoff && MISchedCutoff != ~0U) {
491 CurrentTop = CurrentBottom;
494 ++NumInstrsScheduled;
499 /// enterRegion - Called back from MachineScheduler::runOnMachineFunction after
500 /// crossing a scheduling boundary. [begin, end) includes all instructions in
501 /// the region, including the boundary itself and single-instruction regions
502 /// that don't get scheduled.
503 void ScheduleDAGMI::enterRegion(MachineBasicBlock *bb,
504 MachineBasicBlock::iterator begin,
505 MachineBasicBlock::iterator end,
508 ScheduleDAGInstrs::enterRegion(bb, begin, end, endcount);
510 // For convenience remember the end of the liveness region.
512 (RegionEnd == bb->end()) ? RegionEnd : llvm::next(RegionEnd);
515 // Setup the register pressure trackers for the top scheduled top and bottom
516 // scheduled regions.
517 void ScheduleDAGMI::initRegPressure() {
518 TopRPTracker.init(&MF, RegClassInfo, LIS, BB, RegionBegin);
519 BotRPTracker.init(&MF, RegClassInfo, LIS, BB, LiveRegionEnd);
521 // Close the RPTracker to finalize live ins.
522 RPTracker.closeRegion();
524 DEBUG(RPTracker.getPressure().dump(TRI));
526 // Initialize the live ins and live outs.
527 TopRPTracker.addLiveRegs(RPTracker.getPressure().LiveInRegs);
528 BotRPTracker.addLiveRegs(RPTracker.getPressure().LiveOutRegs);
530 // Close one end of the tracker so we can call
531 // getMaxUpward/DownwardPressureDelta before advancing across any
532 // instructions. This converts currently live regs into live ins/outs.
533 TopRPTracker.closeTop();
534 BotRPTracker.closeBottom();
536 // Account for liveness generated by the region boundary.
537 if (LiveRegionEnd != RegionEnd)
538 BotRPTracker.recede();
540 assert(BotRPTracker.getPos() == RegionEnd && "Can't find the region bottom");
542 // Cache the list of excess pressure sets in this region. This will also track
543 // the max pressure in the scheduled code for these sets.
544 RegionCriticalPSets.clear();
545 std::vector<unsigned> RegionPressure = RPTracker.getPressure().MaxSetPressure;
546 for (unsigned i = 0, e = RegionPressure.size(); i < e; ++i) {
547 unsigned Limit = TRI->getRegPressureSetLimit(i);
548 if (RegionPressure[i] > Limit)
549 RegionCriticalPSets.push_back(PressureElement(i, 0));
551 DEBUG(dbgs() << "Excess PSets: ";
552 for (unsigned i = 0, e = RegionCriticalPSets.size(); i != e; ++i)
553 dbgs() << TRI->getRegPressureSetName(
554 RegionCriticalPSets[i].PSetID) << " ";
558 // FIXME: When the pressure tracker deals in pressure differences then we won't
559 // iterate over all RegionCriticalPSets[i].
561 updateScheduledPressure(std::vector<unsigned> NewMaxPressure) {
562 for (unsigned i = 0, e = RegionCriticalPSets.size(); i < e; ++i) {
563 unsigned ID = RegionCriticalPSets[i].PSetID;
564 int &MaxUnits = RegionCriticalPSets[i].UnitIncrease;
565 if ((int)NewMaxPressure[ID] > MaxUnits)
566 MaxUnits = NewMaxPressure[ID];
570 // Release all DAG roots for scheduling.
571 void ScheduleDAGMI::releaseRoots() {
572 SmallVector<SUnit*, 16> BotRoots;
574 for (std::vector<SUnit>::iterator
575 I = SUnits.begin(), E = SUnits.end(); I != E; ++I) {
576 // A SUnit is ready to top schedule if it has no predecessors.
577 if (I->Preds.empty())
578 SchedImpl->releaseTopNode(&(*I));
579 // A SUnit is ready to bottom schedule if it has no successors.
580 if (I->Succs.empty())
581 BotRoots.push_back(&(*I));
583 // Release bottom roots in reverse order so the higher priority nodes appear
584 // first. This is more natural and slightly more efficient.
585 for (SmallVectorImpl<SUnit*>::const_reverse_iterator
586 I = BotRoots.rbegin(), E = BotRoots.rend(); I != E; ++I)
587 SchedImpl->releaseBottomNode(*I);
590 /// schedule - Called back from MachineScheduler::runOnMachineFunction
591 /// after setting up the current scheduling region. [RegionBegin, RegionEnd)
592 /// only includes instructions that have DAG nodes, not scheduling boundaries.
593 void ScheduleDAGMI::schedule() {
594 // Initialize the register pressure tracker used by buildSchedGraph.
595 RPTracker.init(&MF, RegClassInfo, LIS, BB, LiveRegionEnd);
597 // Account for liveness generate by the region boundary.
598 if (LiveRegionEnd != RegionEnd)
601 // Build the DAG, and compute current register pressure.
602 buildSchedGraph(AA, &RPTracker);
604 // Initialize top/bottom trackers after computing region pressure.
607 DEBUG(for (unsigned su = 0, e = SUnits.size(); su != e; ++su)
608 SUnits[su].dumpAll(this));
610 if (ViewMISchedDAGs) viewGraph();
612 SchedImpl->initialize(this);
614 // Release edges from the special Entry node or to the special Exit node.
615 releaseSuccessors(&EntrySU);
616 releasePredecessors(&ExitSU);
618 // Release all DAG roots for scheduling.
621 CurrentTop = nextIfDebug(RegionBegin, RegionEnd);
622 CurrentBottom = RegionEnd;
623 bool IsTopNode = false;
624 while (SUnit *SU = SchedImpl->pickNode(IsTopNode)) {
625 DEBUG(dbgs() << "*** " << (IsTopNode ? "Top" : "Bottom")
626 << " Scheduling Instruction");
627 if (!checkSchedLimit())
630 // Move the instruction to its new location in the instruction stream.
631 MachineInstr *MI = SU->getInstr();
634 assert(SU->isTopReady() && "node still has unscheduled dependencies");
635 if (&*CurrentTop == MI)
636 CurrentTop = nextIfDebug(++CurrentTop, CurrentBottom);
638 moveInstruction(MI, CurrentTop);
639 TopRPTracker.setPos(MI);
642 // Update top scheduled pressure.
643 TopRPTracker.advance();
644 assert(TopRPTracker.getPos() == CurrentTop && "out of sync");
645 updateScheduledPressure(TopRPTracker.getPressure().MaxSetPressure);
647 // Release dependent instructions for scheduling.
648 releaseSuccessors(SU);
651 assert(SU->isBottomReady() && "node still has unscheduled dependencies");
652 MachineBasicBlock::iterator priorII =
653 priorNonDebug(CurrentBottom, CurrentTop);
655 CurrentBottom = priorII;
657 if (&*CurrentTop == MI) {
658 CurrentTop = nextIfDebug(++CurrentTop, priorII);
659 TopRPTracker.setPos(CurrentTop);
661 moveInstruction(MI, CurrentBottom);
664 // Update bottom scheduled pressure.
665 BotRPTracker.recede();
666 assert(BotRPTracker.getPos() == CurrentBottom && "out of sync");
667 updateScheduledPressure(BotRPTracker.getPressure().MaxSetPressure);
669 // Release dependent instructions for scheduling.
670 releasePredecessors(SU);
672 SU->isScheduled = true;
673 SchedImpl->schedNode(SU, IsTopNode);
674 DEBUG(SU->dump(this));
676 assert(CurrentTop == CurrentBottom && "Nonempty unscheduled zone.");
681 /// Reinsert any remaining debug_values, just like the PostRA scheduler.
682 void ScheduleDAGMI::placeDebugValues() {
683 // If first instruction was a DBG_VALUE then put it back.
685 BB->splice(RegionBegin, BB, FirstDbgValue);
686 RegionBegin = FirstDbgValue;
689 for (std::vector<std::pair<MachineInstr *, MachineInstr *> >::iterator
690 DI = DbgValues.end(), DE = DbgValues.begin(); DI != DE; --DI) {
691 std::pair<MachineInstr *, MachineInstr *> P = *prior(DI);
692 MachineInstr *DbgValue = P.first;
693 MachineBasicBlock::iterator OrigPrevMI = P.second;
694 BB->splice(++OrigPrevMI, BB, DbgValue);
695 if (OrigPrevMI == llvm::prior(RegionEnd))
696 RegionEnd = DbgValue;
699 FirstDbgValue = NULL;
702 //===----------------------------------------------------------------------===//
703 // ConvergingScheduler - Implementation of the standard MachineSchedStrategy.
704 //===----------------------------------------------------------------------===//
707 /// ReadyQ encapsulates vector of "ready" SUnits with basic convenience methods
708 /// for pushing and removing nodes. ReadyQ's are uniquely identified by an
709 /// ID. SUnit::NodeQueueId us a mask of the ReadyQs that the SUnit is in.
713 std::vector<SUnit*> Queue;
716 ReadyQueue(unsigned id, const Twine &name): ID(id), Name(name.str()) {}
718 unsigned getID() const { return ID; }
720 StringRef getName() const { return Name; }
722 // SU is in this queue if it's NodeQueueID is a superset of this ID.
723 bool isInQueue(SUnit *SU) const { return (SU->NodeQueueId & ID); }
725 bool empty() const { return Queue.empty(); }
727 unsigned size() const { return Queue.size(); }
729 typedef std::vector<SUnit*>::iterator iterator;
731 iterator begin() { return Queue.begin(); }
733 iterator end() { return Queue.end(); }
735 iterator find(SUnit *SU) {
736 return std::find(Queue.begin(), Queue.end(), SU);
739 void push(SUnit *SU) {
741 SU->NodeQueueId |= ID;
744 void remove(iterator I) {
745 (*I)->NodeQueueId &= ~ID;
751 dbgs() << Name << ": ";
752 for (unsigned i = 0, e = Queue.size(); i < e; ++i)
753 dbgs() << Queue[i]->NodeNum << " ";
758 /// ConvergingScheduler shrinks the unscheduled zone using heuristics to balance
760 class ConvergingScheduler : public MachineSchedStrategy {
762 /// Store the state used by ConvergingScheduler heuristics, required for the
763 /// lifetime of one invocation of pickNode().
764 struct SchedCandidate {
765 // The best SUnit candidate.
768 // Register pressure values for the best candidate.
769 RegPressureDelta RPDelta;
771 SchedCandidate(): SU(NULL) {}
773 /// Represent the type of SchedCandidate found within a single queue.
775 NoCand, NodeOrder, SingleExcess, SingleCritical, SingleMax, MultiPressure };
777 /// Each Scheduling boundary is associated with ready queues. It tracks the
778 /// current cycle in whichever direction at has moved, and maintains the state
779 /// of "hazards" and other interlocks at the current cycle.
780 struct SchedBoundary {
781 ReadyQueue Available;
785 ScheduleHazardRecognizer *HazardRec;
790 /// MinReadyCycle - Cycle of the soonest available instruction.
791 unsigned MinReadyCycle;
793 /// Pending queues extend the ready queues with the same ID and the
795 SchedBoundary(unsigned ID, const Twine &Name):
796 Available(ID, Name+".A"),
797 Pending(ID << ConvergingScheduler::LogMaxQID, Name+".P"),
798 CheckPending(false), HazardRec(0), CurrCycle(0), IssueCount(0),
799 MinReadyCycle(UINT_MAX) {}
801 ~SchedBoundary() { delete HazardRec; }
804 return Available.getID() == ConvergingScheduler::TopQID;
807 void releaseNode(SUnit *SU, unsigned ReadyCycle);
811 void releasePending();
813 void removeReady(SUnit *SU);
815 SUnit *pickOnlyChoice();
819 const TargetRegisterInfo *TRI;
821 // State of the top and bottom scheduled instruction boundaries.
826 /// SUnit::NodeQueueId: 0 (none), 1 (top), 2 (bot), 3 (both)
833 ConvergingScheduler():
834 DAG(0), TRI(0), Top(TopQID, "TopQ"), Bot(BotQID, "BotQ") {}
836 virtual void initialize(ScheduleDAGMI *dag);
838 virtual SUnit *pickNode(bool &IsTopNode);
840 virtual void schedNode(SUnit *SU, bool IsTopNode);
842 virtual void releaseTopNode(SUnit *SU);
844 virtual void releaseBottomNode(SUnit *SU);
847 SUnit *pickNodeBidrectional(bool &IsTopNode);
849 CandResult pickNodeFromQueue(ReadyQueue &Q,
850 const RegPressureTracker &RPTracker,
851 SchedCandidate &Candidate);
853 void traceCandidate(const char *Label, const ReadyQueue &Q, SUnit *SU,
854 PressureElement P = PressureElement());
859 void ConvergingScheduler::initialize(ScheduleDAGMI *dag) {
863 // Initialize the HazardRecognizers.
864 const TargetMachine &TM = DAG->MF.getTarget();
865 const InstrItineraryData *Itin = TM.getInstrItineraryData();
866 Top.HazardRec = TM.getInstrInfo()->CreateTargetMIHazardRecognizer(Itin, DAG);
867 Bot.HazardRec = TM.getInstrInfo()->CreateTargetMIHazardRecognizer(Itin, DAG);
869 assert((!ForceTopDown || !ForceBottomUp) &&
870 "-misched-topdown incompatible with -misched-bottomup");
873 void ConvergingScheduler::releaseTopNode(SUnit *SU) {
874 Top.releaseNode(SU, SU->getDepth());
877 void ConvergingScheduler::releaseBottomNode(SUnit *SU) {
878 Bot.releaseNode(SU, SU->getHeight());
881 void ConvergingScheduler::SchedBoundary::releaseNode(SUnit *SU,
882 unsigned ReadyCycle) {
886 if (ReadyCycle < MinReadyCycle)
887 MinReadyCycle = ReadyCycle;
889 // Check for interlocks first. For the purpose of other heuristics, an
890 // instruction that cannot issue appears as if it's not in the ReadyQueue.
891 if (HazardRec->isEnabled()
892 && HazardRec->getHazardType(SU) != ScheduleHazardRecognizer::NoHazard)
898 /// Move the boundary of scheduled code by one cycle.
899 void ConvergingScheduler::SchedBoundary::bumpCycle() {
902 assert(MinReadyCycle < UINT_MAX && "MinReadyCycle uninitialized");
903 unsigned NextCycle = std::max(CurrCycle + 1, MinReadyCycle);
905 if (!HazardRec->isEnabled()) {
906 // Bypass lots of virtual calls in case of long latency.
907 CurrCycle = NextCycle;
910 for (; CurrCycle != NextCycle; ++CurrCycle) {
912 HazardRec->AdvanceCycle();
914 HazardRec->RecedeCycle();
919 DEBUG(dbgs() << "*** " << Available.getName() << " cycle "
920 << CurrCycle << '\n');
923 /// Release pending ready nodes in to the available queue. This makes them
924 /// visible to heuristics.
925 void ConvergingScheduler::SchedBoundary::releasePending() {
926 // If the available queue is empty, it is safe to reset MinReadyCycle.
927 if (Available.empty())
928 MinReadyCycle = UINT_MAX;
930 // Check to see if any of the pending instructions are ready to issue. If
931 // so, add them to the available queue.
932 for (unsigned i = 0, e = Pending.size(); i != e; ++i) {
933 SUnit *SU = *(Pending.begin()+i);
934 unsigned ReadyCycle = isTop() ? SU->getHeight() : SU->getDepth();
936 if (ReadyCycle < MinReadyCycle)
937 MinReadyCycle = ReadyCycle;
939 if (ReadyCycle > CurrCycle)
942 if (HazardRec->isEnabled()
943 && HazardRec->getHazardType(SU) != ScheduleHazardRecognizer::NoHazard)
947 Pending.remove(Pending.begin()+i);
950 CheckPending = false;
953 /// Remove SU from the ready set for this boundary.
954 void ConvergingScheduler::SchedBoundary::removeReady(SUnit *SU) {
955 if (Available.isInQueue(SU))
956 Available.remove(Available.find(SU));
958 assert(Pending.isInQueue(SU) && "bad ready count");
959 Pending.remove(Pending.find(SU));
963 /// If this queue only has one ready candidate, return it. As a side effect,
964 /// advance the cycle until at least one node is ready. If multiple instructions
965 /// are ready, return NULL.
966 SUnit *ConvergingScheduler::SchedBoundary::pickOnlyChoice() {
970 for (unsigned i = 0; Available.empty(); ++i) {
971 assert(i <= HazardRec->getMaxLookAhead() && "permanent hazard"); (void)i;
975 if (Available.size() == 1)
976 return *Available.begin();
981 void ConvergingScheduler::traceCandidate(const char *Label, const ReadyQueue &Q,
982 SUnit *SU, PressureElement P) {
983 dbgs() << Label << " " << Q.getName() << " ";
985 dbgs() << TRI->getRegPressureSetName(P.PSetID) << ":" << P.UnitIncrease
993 /// pickNodeFromQueue helper that returns true if the LHS reg pressure effect is
994 /// more desirable than RHS from scheduling standpoint.
995 static bool compareRPDelta(const RegPressureDelta &LHS,
996 const RegPressureDelta &RHS) {
997 // Compare each component of pressure in decreasing order of importance
998 // without checking if any are valid. Invalid PressureElements are assumed to
999 // have UnitIncrease==0, so are neutral.
1001 // Avoid increasing the max critical pressure in the scheduled region.
1002 if (LHS.Excess.UnitIncrease != RHS.Excess.UnitIncrease)
1003 return LHS.Excess.UnitIncrease < RHS.Excess.UnitIncrease;
1005 // Avoid increasing the max critical pressure in the scheduled region.
1006 if (LHS.CriticalMax.UnitIncrease != RHS.CriticalMax.UnitIncrease)
1007 return LHS.CriticalMax.UnitIncrease < RHS.CriticalMax.UnitIncrease;
1009 // Avoid increasing the max pressure of the entire region.
1010 if (LHS.CurrentMax.UnitIncrease != RHS.CurrentMax.UnitIncrease)
1011 return LHS.CurrentMax.UnitIncrease < RHS.CurrentMax.UnitIncrease;
1016 /// Pick the best candidate from the top queue.
1018 /// TODO: getMaxPressureDelta results can be mostly cached for each SUnit during
1019 /// DAG building. To adjust for the current scheduling location we need to
1020 /// maintain the number of vreg uses remaining to be top-scheduled.
1021 ConvergingScheduler::CandResult ConvergingScheduler::
1022 pickNodeFromQueue(ReadyQueue &Q, const RegPressureTracker &RPTracker,
1023 SchedCandidate &Candidate) {
1026 // getMaxPressureDelta temporarily modifies the tracker.
1027 RegPressureTracker &TempTracker = const_cast<RegPressureTracker&>(RPTracker);
1029 // BestSU remains NULL if no top candidates beat the best existing candidate.
1030 CandResult FoundCandidate = NoCand;
1031 for (ReadyQueue::iterator I = Q.begin(), E = Q.end(); I != E; ++I) {
1032 RegPressureDelta RPDelta;
1033 TempTracker.getMaxPressureDelta((*I)->getInstr(), RPDelta,
1034 DAG->getRegionCriticalPSets(),
1035 DAG->getRegPressure().MaxSetPressure);
1037 // Initialize the candidate if needed.
1038 if (!Candidate.SU) {
1040 Candidate.RPDelta = RPDelta;
1041 FoundCandidate = NodeOrder;
1044 // Avoid exceeding the target's limit.
1045 if (RPDelta.Excess.UnitIncrease < Candidate.RPDelta.Excess.UnitIncrease) {
1046 DEBUG(traceCandidate("ECAND", Q, *I, RPDelta.Excess));
1048 Candidate.RPDelta = RPDelta;
1049 FoundCandidate = SingleExcess;
1052 if (RPDelta.Excess.UnitIncrease > Candidate.RPDelta.Excess.UnitIncrease)
1054 if (FoundCandidate == SingleExcess)
1055 FoundCandidate = MultiPressure;
1057 // Avoid increasing the max critical pressure in the scheduled region.
1058 if (RPDelta.CriticalMax.UnitIncrease
1059 < Candidate.RPDelta.CriticalMax.UnitIncrease) {
1060 DEBUG(traceCandidate("PCAND", Q, *I, RPDelta.CriticalMax));
1062 Candidate.RPDelta = RPDelta;
1063 FoundCandidate = SingleCritical;
1066 if (RPDelta.CriticalMax.UnitIncrease
1067 > Candidate.RPDelta.CriticalMax.UnitIncrease)
1069 if (FoundCandidate == SingleCritical)
1070 FoundCandidate = MultiPressure;
1072 // Avoid increasing the max pressure of the entire region.
1073 if (RPDelta.CurrentMax.UnitIncrease
1074 < Candidate.RPDelta.CurrentMax.UnitIncrease) {
1075 DEBUG(traceCandidate("MCAND", Q, *I, RPDelta.CurrentMax));
1077 Candidate.RPDelta = RPDelta;
1078 FoundCandidate = SingleMax;
1081 if (RPDelta.CurrentMax.UnitIncrease
1082 > Candidate.RPDelta.CurrentMax.UnitIncrease)
1084 if (FoundCandidate == SingleMax)
1085 FoundCandidate = MultiPressure;
1087 // Fall through to original instruction order.
1088 // Only consider node order if Candidate was chosen from this Q.
1089 if (FoundCandidate == NoCand)
1092 if ((Q.getID() == TopQID && (*I)->NodeNum < Candidate.SU->NodeNum)
1093 || (Q.getID() == BotQID && (*I)->NodeNum > Candidate.SU->NodeNum)) {
1094 DEBUG(traceCandidate("NCAND", Q, *I));
1096 Candidate.RPDelta = RPDelta;
1097 FoundCandidate = NodeOrder;
1100 return FoundCandidate;
1103 /// Pick the best candidate node from either the top or bottom queue.
1104 SUnit *ConvergingScheduler::pickNodeBidrectional(bool &IsTopNode) {
1105 // Schedule as far as possible in the direction of no choice. This is most
1106 // efficient, but also provides the best heuristics for CriticalPSets.
1107 if (SUnit *SU = Bot.pickOnlyChoice()) {
1111 if (SUnit *SU = Top.pickOnlyChoice()) {
1115 SchedCandidate BotCand;
1116 // Prefer bottom scheduling when heuristics are silent.
1117 CandResult BotResult = pickNodeFromQueue(Bot.Available,
1118 DAG->getBotRPTracker(), BotCand);
1119 assert(BotResult != NoCand && "failed to find the first candidate");
1121 // If either Q has a single candidate that provides the least increase in
1122 // Excess pressure, we can immediately schedule from that Q.
1124 // RegionCriticalPSets summarizes the pressure within the scheduled region and
1125 // affects picking from either Q. If scheduling in one direction must
1126 // increase pressure for one of the excess PSets, then schedule in that
1127 // direction first to provide more freedom in the other direction.
1128 if (BotResult == SingleExcess || BotResult == SingleCritical) {
1132 // Check if the top Q has a better candidate.
1133 SchedCandidate TopCand;
1134 CandResult TopResult = pickNodeFromQueue(Top.Available,
1135 DAG->getTopRPTracker(), TopCand);
1136 assert(TopResult != NoCand && "failed to find the first candidate");
1138 if (TopResult == SingleExcess || TopResult == SingleCritical) {
1142 // If either Q has a single candidate that minimizes pressure above the
1143 // original region's pressure pick it.
1144 if (BotResult == SingleMax) {
1148 if (TopResult == SingleMax) {
1152 // Check for a salient pressure difference and pick the best from either side.
1153 if (compareRPDelta(TopCand.RPDelta, BotCand.RPDelta)) {
1157 // Otherwise prefer the bottom candidate in node order.
1162 /// Pick the best node to balance the schedule. Implements MachineSchedStrategy.
1163 SUnit *ConvergingScheduler::pickNode(bool &IsTopNode) {
1164 if (DAG->top() == DAG->bottom()) {
1165 assert(Top.Available.empty() && Top.Pending.empty() &&
1166 Bot.Available.empty() && Bot.Pending.empty() && "ReadyQ garbage");
1171 SU = Top.pickOnlyChoice();
1173 SchedCandidate TopCand;
1174 CandResult TopResult =
1175 pickNodeFromQueue(Top.Available, DAG->getTopRPTracker(), TopCand);
1176 assert(TopResult != NoCand && "failed to find the first candidate");
1182 else if (ForceBottomUp) {
1183 SU = Bot.pickOnlyChoice();
1185 SchedCandidate BotCand;
1186 CandResult BotResult =
1187 pickNodeFromQueue(Bot.Available, DAG->getBotRPTracker(), BotCand);
1188 assert(BotResult != NoCand && "failed to find the first candidate");
1195 SU = pickNodeBidrectional(IsTopNode);
1197 if (SU->isTopReady())
1198 Top.removeReady(SU);
1199 if (SU->isBottomReady())
1200 Bot.removeReady(SU);
1204 /// Update the scheduler's state after scheduling a node. This is the same node
1205 /// that was just returned by pickNode(). However, ScheduleDAGMI needs to update
1206 /// it's state based on the current cycle before MachineSchedStrategy.
1207 void ConvergingScheduler::schedNode(SUnit *SU, bool IsTopNode) {
1208 DEBUG(dbgs() << " in cycle " << (IsTopNode ? Top.CurrCycle : Bot.CurrCycle)
1211 // Update the reservation table.
1212 if (IsTopNode && Top.HazardRec->isEnabled()) {
1213 Top.HazardRec->EmitInstruction(SU);
1214 if (Top.HazardRec->atIssueLimit()) {
1215 DEBUG(dbgs() << "*** Max instrs at cycle " << Top.CurrCycle << '\n');
1219 else if (Bot.HazardRec->isEnabled()) {
1221 // Calls are scheduled with their preceding instructions. For bottom-up
1222 // scheduling, clear the pipeline state before emitting.
1223 Bot.HazardRec->Reset();
1225 Bot.HazardRec->EmitInstruction(SU);
1226 if (Bot.HazardRec->atIssueLimit()) {
1227 DEBUG(dbgs() << "*** Max instrs at cycle " << Bot.CurrCycle << '\n');
1233 /// Create the standard converging machine scheduler. This will be used as the
1234 /// default scheduler if the target does not set a default.
1235 static ScheduleDAGInstrs *createConvergingSched(MachineSchedContext *C) {
1236 assert((!ForceTopDown || !ForceBottomUp) &&
1237 "-misched-topdown incompatible with -misched-bottomup");
1238 return new ScheduleDAGMI(C, new ConvergingScheduler());
1240 static MachineSchedRegistry
1241 ConvergingSchedRegistry("converge", "Standard converging scheduler.",
1242 createConvergingSched);
1244 //===----------------------------------------------------------------------===//
1245 // Machine Instruction Shuffler for Correctness Testing
1246 //===----------------------------------------------------------------------===//
1250 /// Apply a less-than relation on the node order, which corresponds to the
1251 /// instruction order prior to scheduling. IsReverse implements greater-than.
1252 template<bool IsReverse>
1254 bool operator()(SUnit *A, SUnit *B) const {
1256 return A->NodeNum > B->NodeNum;
1258 return A->NodeNum < B->NodeNum;
1262 /// Reorder instructions as much as possible.
1263 class InstructionShuffler : public MachineSchedStrategy {
1267 // Using a less-than relation (SUnitOrder<false>) for the TopQ priority
1268 // gives nodes with a higher number higher priority causing the latest
1269 // instructions to be scheduled first.
1270 PriorityQueue<SUnit*, std::vector<SUnit*>, SUnitOrder<false> >
1272 // When scheduling bottom-up, use greater-than as the queue priority.
1273 PriorityQueue<SUnit*, std::vector<SUnit*>, SUnitOrder<true> >
1276 InstructionShuffler(bool alternate, bool topdown)
1277 : IsAlternating(alternate), IsTopDown(topdown) {}
1279 virtual void initialize(ScheduleDAGMI *) {
1284 /// Implement MachineSchedStrategy interface.
1285 /// -----------------------------------------
1287 virtual SUnit *pickNode(bool &IsTopNode) {
1291 if (TopQ.empty()) return NULL;
1294 } while (SU->isScheduled);
1299 if (BottomQ.empty()) return NULL;
1302 } while (SU->isScheduled);
1306 IsTopDown = !IsTopDown;
1310 virtual void schedNode(SUnit *SU, bool IsTopNode) {}
1312 virtual void releaseTopNode(SUnit *SU) {
1315 virtual void releaseBottomNode(SUnit *SU) {
1321 static ScheduleDAGInstrs *createInstructionShuffler(MachineSchedContext *C) {
1322 bool Alternate = !ForceTopDown && !ForceBottomUp;
1323 bool TopDown = !ForceBottomUp;
1324 assert((TopDown || !ForceTopDown) &&
1325 "-misched-topdown incompatible with -misched-bottomup");
1326 return new ScheduleDAGMI(C, new InstructionShuffler(Alternate, TopDown));
1328 static MachineSchedRegistry ShufflerRegistry(
1329 "shuffle", "Shuffle machine instructions alternating directions",
1330 createInstructionShuffler);