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/ScheduleDFS.h"
22 #include "llvm/CodeGen/ScheduleHazardRecognizer.h"
23 #include "llvm/Analysis/AliasAnalysis.h"
24 #include "llvm/Support/CommandLine.h"
25 #include "llvm/Support/Debug.h"
26 #include "llvm/Support/ErrorHandling.h"
27 #include "llvm/Support/raw_ostream.h"
28 #include "llvm/ADT/OwningPtr.h"
29 #include "llvm/ADT/PriorityQueue.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 // Threshold to very roughly model an out-of-order processor's instruction
53 // buffers. If the actual value of this threshold matters much in practice, then
54 // it can be specified by the machine model. For now, it's an experimental
55 // tuning knob to determine when and if it matters.
56 static cl::opt<unsigned> ILPWindow("ilp-window", cl::Hidden,
57 cl::desc("Allow expected latency to exceed the critical path by N cycles "
58 "before attempting to balance ILP"),
61 // Experimental heuristics
62 static cl::opt<bool> EnableLoadCluster("misched-cluster", cl::Hidden,
63 cl::desc("Enable load clustering."), cl::init(true));
65 // Experimental heuristics
66 static cl::opt<bool> EnableMacroFusion("misched-fusion", cl::Hidden,
67 cl::desc("Enable scheduling for macro fusion."), cl::init(true));
69 //===----------------------------------------------------------------------===//
70 // Machine Instruction Scheduling Pass and Registry
71 //===----------------------------------------------------------------------===//
73 MachineSchedContext::MachineSchedContext():
74 MF(0), MLI(0), MDT(0), PassConfig(0), AA(0), LIS(0) {
75 RegClassInfo = new RegisterClassInfo();
78 MachineSchedContext::~MachineSchedContext() {
83 /// MachineScheduler runs after coalescing and before register allocation.
84 class MachineScheduler : public MachineSchedContext,
85 public MachineFunctionPass {
89 virtual void getAnalysisUsage(AnalysisUsage &AU) const;
91 virtual void releaseMemory() {}
93 virtual bool runOnMachineFunction(MachineFunction&);
95 virtual void print(raw_ostream &O, const Module* = 0) const;
97 static char ID; // Class identification, replacement for typeinfo
101 char MachineScheduler::ID = 0;
103 char &llvm::MachineSchedulerID = MachineScheduler::ID;
105 INITIALIZE_PASS_BEGIN(MachineScheduler, "misched",
106 "Machine Instruction Scheduler", false, false)
107 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
108 INITIALIZE_PASS_DEPENDENCY(SlotIndexes)
109 INITIALIZE_PASS_DEPENDENCY(LiveIntervals)
110 INITIALIZE_PASS_END(MachineScheduler, "misched",
111 "Machine Instruction Scheduler", false, false)
113 MachineScheduler::MachineScheduler()
114 : MachineFunctionPass(ID) {
115 initializeMachineSchedulerPass(*PassRegistry::getPassRegistry());
118 void MachineScheduler::getAnalysisUsage(AnalysisUsage &AU) const {
119 AU.setPreservesCFG();
120 AU.addRequiredID(MachineDominatorsID);
121 AU.addRequired<MachineLoopInfo>();
122 AU.addRequired<AliasAnalysis>();
123 AU.addRequired<TargetPassConfig>();
124 AU.addRequired<SlotIndexes>();
125 AU.addPreserved<SlotIndexes>();
126 AU.addRequired<LiveIntervals>();
127 AU.addPreserved<LiveIntervals>();
128 MachineFunctionPass::getAnalysisUsage(AU);
131 MachinePassRegistry MachineSchedRegistry::Registry;
133 /// A dummy default scheduler factory indicates whether the scheduler
134 /// is overridden on the command line.
135 static ScheduleDAGInstrs *useDefaultMachineSched(MachineSchedContext *C) {
139 /// MachineSchedOpt allows command line selection of the scheduler.
140 static cl::opt<MachineSchedRegistry::ScheduleDAGCtor, false,
141 RegisterPassParser<MachineSchedRegistry> >
142 MachineSchedOpt("misched",
143 cl::init(&useDefaultMachineSched), cl::Hidden,
144 cl::desc("Machine instruction scheduler to use"));
146 static MachineSchedRegistry
147 DefaultSchedRegistry("default", "Use the target's default scheduler choice.",
148 useDefaultMachineSched);
150 /// Forward declare the standard machine scheduler. This will be used as the
151 /// default scheduler if the target does not set a default.
152 static ScheduleDAGInstrs *createConvergingSched(MachineSchedContext *C);
155 /// Decrement this iterator until reaching the top or a non-debug instr.
156 static MachineBasicBlock::iterator
157 priorNonDebug(MachineBasicBlock::iterator I, MachineBasicBlock::iterator Beg) {
158 assert(I != Beg && "reached the top of the region, cannot decrement");
160 if (!I->isDebugValue())
166 /// If this iterator is a debug value, increment until reaching the End or a
167 /// non-debug instruction.
168 static MachineBasicBlock::iterator
169 nextIfDebug(MachineBasicBlock::iterator I, MachineBasicBlock::iterator End) {
170 for(; I != End; ++I) {
171 if (!I->isDebugValue())
177 /// Top-level MachineScheduler pass driver.
179 /// Visit blocks in function order. Divide each block into scheduling regions
180 /// and visit them bottom-up. Visiting regions bottom-up is not required, but is
181 /// consistent with the DAG builder, which traverses the interior of the
182 /// scheduling regions bottom-up.
184 /// This design avoids exposing scheduling boundaries to the DAG builder,
185 /// simplifying the DAG builder's support for "special" target instructions.
186 /// At the same time the design allows target schedulers to operate across
187 /// scheduling boundaries, for example to bundle the boudary instructions
188 /// without reordering them. This creates complexity, because the target
189 /// scheduler must update the RegionBegin and RegionEnd positions cached by
190 /// ScheduleDAGInstrs whenever adding or removing instructions. A much simpler
191 /// design would be to split blocks at scheduling boundaries, but LLVM has a
192 /// general bias against block splitting purely for implementation simplicity.
193 bool MachineScheduler::runOnMachineFunction(MachineFunction &mf) {
194 DEBUG(dbgs() << "Before MISsched:\n"; mf.print(dbgs()));
196 // Initialize the context of the pass.
198 MLI = &getAnalysis<MachineLoopInfo>();
199 MDT = &getAnalysis<MachineDominatorTree>();
200 PassConfig = &getAnalysis<TargetPassConfig>();
201 AA = &getAnalysis<AliasAnalysis>();
203 LIS = &getAnalysis<LiveIntervals>();
204 const TargetInstrInfo *TII = MF->getTarget().getInstrInfo();
206 RegClassInfo->runOnMachineFunction(*MF);
208 // Select the scheduler, or set the default.
209 MachineSchedRegistry::ScheduleDAGCtor Ctor = MachineSchedOpt;
210 if (Ctor == useDefaultMachineSched) {
211 // Get the default scheduler set by the target.
212 Ctor = MachineSchedRegistry::getDefault();
214 Ctor = createConvergingSched;
215 MachineSchedRegistry::setDefault(Ctor);
218 // Instantiate the selected scheduler.
219 OwningPtr<ScheduleDAGInstrs> Scheduler(Ctor(this));
221 // Visit all machine basic blocks.
223 // TODO: Visit blocks in global postorder or postorder within the bottom-up
224 // loop tree. Then we can optionally compute global RegPressure.
225 for (MachineFunction::iterator MBB = MF->begin(), MBBEnd = MF->end();
226 MBB != MBBEnd; ++MBB) {
228 Scheduler->startBlock(MBB);
230 // Break the block into scheduling regions [I, RegionEnd), and schedule each
231 // region as soon as it is discovered. RegionEnd points the scheduling
232 // boundary at the bottom of the region. The DAG does not include RegionEnd,
233 // but the region does (i.e. the next RegionEnd is above the previous
234 // RegionBegin). If the current block has no terminator then RegionEnd ==
235 // MBB->end() for the bottom region.
237 // The Scheduler may insert instructions during either schedule() or
238 // exitRegion(), even for empty regions. So the local iterators 'I' and
239 // 'RegionEnd' are invalid across these calls.
240 unsigned RemainingInstrs = MBB->size();
241 for(MachineBasicBlock::iterator RegionEnd = MBB->end();
242 RegionEnd != MBB->begin(); RegionEnd = Scheduler->begin()) {
244 // Avoid decrementing RegionEnd for blocks with no terminator.
245 if (RegionEnd != MBB->end()
246 || TII->isSchedulingBoundary(llvm::prior(RegionEnd), MBB, *MF)) {
248 // Count the boundary instruction.
252 // The next region starts above the previous region. Look backward in the
253 // instruction stream until we find the nearest boundary.
254 MachineBasicBlock::iterator I = RegionEnd;
255 for(;I != MBB->begin(); --I, --RemainingInstrs) {
256 if (TII->isSchedulingBoundary(llvm::prior(I), MBB, *MF))
259 // Notify the scheduler of the region, even if we may skip scheduling
260 // it. Perhaps it still needs to be bundled.
261 Scheduler->enterRegion(MBB, I, RegionEnd, RemainingInstrs);
263 // Skip empty scheduling regions (0 or 1 schedulable instructions).
264 if (I == RegionEnd || I == llvm::prior(RegionEnd)) {
265 // Close the current region. Bundle the terminator if needed.
266 // This invalidates 'RegionEnd' and 'I'.
267 Scheduler->exitRegion();
270 DEBUG(dbgs() << "********** MI Scheduling **********\n");
271 DEBUG(dbgs() << MF->getName()
272 << ":BB#" << MBB->getNumber() << "\n From: " << *I << " To: ";
273 if (RegionEnd != MBB->end()) dbgs() << *RegionEnd;
274 else dbgs() << "End";
275 dbgs() << " Remaining: " << RemainingInstrs << "\n");
277 // Schedule a region: possibly reorder instructions.
278 // This invalidates 'RegionEnd' and 'I'.
279 Scheduler->schedule();
281 // Close the current region.
282 Scheduler->exitRegion();
284 // Scheduling has invalidated the current iterator 'I'. Ask the
285 // scheduler for the top of it's scheduled region.
286 RegionEnd = Scheduler->begin();
288 assert(RemainingInstrs == 0 && "Instruction count mismatch!");
289 Scheduler->finishBlock();
291 Scheduler->finalizeSchedule();
292 DEBUG(LIS->print(dbgs()));
296 void MachineScheduler::print(raw_ostream &O, const Module* m) const {
300 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
301 void ReadyQueue::dump() {
302 dbgs() << Name << ": ";
303 for (unsigned i = 0, e = Queue.size(); i < e; ++i)
304 dbgs() << Queue[i]->NodeNum << " ";
309 //===----------------------------------------------------------------------===//
310 // ScheduleDAGMI - Base class for MachineInstr scheduling with LiveIntervals
312 //===----------------------------------------------------------------------===//
314 bool ScheduleDAGMI::addEdge(SUnit *SuccSU, const SDep &PredDep) {
315 if (SuccSU != &ExitSU) {
316 // Do not use WillCreateCycle, it assumes SD scheduling.
317 // If Pred is reachable from Succ, then the edge creates a cycle.
318 if (Topo.IsReachable(PredDep.getSUnit(), SuccSU))
320 Topo.AddPred(SuccSU, PredDep.getSUnit());
322 SuccSU->addPred(PredDep, /*Required=*/!PredDep.isArtificial());
323 // Return true regardless of whether a new edge needed to be inserted.
327 /// ReleaseSucc - Decrement the NumPredsLeft count of a successor. When
328 /// NumPredsLeft reaches zero, release the successor node.
330 /// FIXME: Adjust SuccSU height based on MinLatency.
331 void ScheduleDAGMI::releaseSucc(SUnit *SU, SDep *SuccEdge) {
332 SUnit *SuccSU = SuccEdge->getSUnit();
334 if (SuccEdge->isWeak()) {
335 --SuccSU->WeakPredsLeft;
336 if (SuccEdge->isCluster())
337 NextClusterSucc = SuccSU;
341 if (SuccSU->NumPredsLeft == 0) {
342 dbgs() << "*** Scheduling failed! ***\n";
344 dbgs() << " has been released too many times!\n";
348 --SuccSU->NumPredsLeft;
349 if (SuccSU->NumPredsLeft == 0 && SuccSU != &ExitSU)
350 SchedImpl->releaseTopNode(SuccSU);
353 /// releaseSuccessors - Call releaseSucc on each of SU's successors.
354 void ScheduleDAGMI::releaseSuccessors(SUnit *SU) {
355 for (SUnit::succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
357 releaseSucc(SU, &*I);
361 /// ReleasePred - Decrement the NumSuccsLeft count of a predecessor. When
362 /// NumSuccsLeft reaches zero, release the predecessor node.
364 /// FIXME: Adjust PredSU height based on MinLatency.
365 void ScheduleDAGMI::releasePred(SUnit *SU, SDep *PredEdge) {
366 SUnit *PredSU = PredEdge->getSUnit();
368 if (PredEdge->isWeak()) {
369 --PredSU->WeakSuccsLeft;
370 if (PredEdge->isCluster())
371 NextClusterPred = PredSU;
375 if (PredSU->NumSuccsLeft == 0) {
376 dbgs() << "*** Scheduling failed! ***\n";
378 dbgs() << " has been released too many times!\n";
382 --PredSU->NumSuccsLeft;
383 if (PredSU->NumSuccsLeft == 0 && PredSU != &EntrySU)
384 SchedImpl->releaseBottomNode(PredSU);
387 /// releasePredecessors - Call releasePred on each of SU's predecessors.
388 void ScheduleDAGMI::releasePredecessors(SUnit *SU) {
389 for (SUnit::pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
391 releasePred(SU, &*I);
395 void ScheduleDAGMI::moveInstruction(MachineInstr *MI,
396 MachineBasicBlock::iterator InsertPos) {
397 // Advance RegionBegin if the first instruction moves down.
398 if (&*RegionBegin == MI)
401 // Update the instruction stream.
402 BB->splice(InsertPos, BB, MI);
404 // Update LiveIntervals
405 LIS->handleMove(MI, /*UpdateFlags=*/true);
407 // Recede RegionBegin if an instruction moves above the first.
408 if (RegionBegin == InsertPos)
412 bool ScheduleDAGMI::checkSchedLimit() {
414 if (NumInstrsScheduled == MISchedCutoff && MISchedCutoff != ~0U) {
415 CurrentTop = CurrentBottom;
418 ++NumInstrsScheduled;
423 /// enterRegion - Called back from MachineScheduler::runOnMachineFunction after
424 /// crossing a scheduling boundary. [begin, end) includes all instructions in
425 /// the region, including the boundary itself and single-instruction regions
426 /// that don't get scheduled.
427 void ScheduleDAGMI::enterRegion(MachineBasicBlock *bb,
428 MachineBasicBlock::iterator begin,
429 MachineBasicBlock::iterator end,
432 ScheduleDAGInstrs::enterRegion(bb, begin, end, endcount);
434 // For convenience remember the end of the liveness region.
436 (RegionEnd == bb->end()) ? RegionEnd : llvm::next(RegionEnd);
439 // Setup the register pressure trackers for the top scheduled top and bottom
440 // scheduled regions.
441 void ScheduleDAGMI::initRegPressure() {
442 TopRPTracker.init(&MF, RegClassInfo, LIS, BB, RegionBegin);
443 BotRPTracker.init(&MF, RegClassInfo, LIS, BB, LiveRegionEnd);
445 // Close the RPTracker to finalize live ins.
446 RPTracker.closeRegion();
448 DEBUG(RPTracker.getPressure().dump(TRI));
450 // Initialize the live ins and live outs.
451 TopRPTracker.addLiveRegs(RPTracker.getPressure().LiveInRegs);
452 BotRPTracker.addLiveRegs(RPTracker.getPressure().LiveOutRegs);
454 // Close one end of the tracker so we can call
455 // getMaxUpward/DownwardPressureDelta before advancing across any
456 // instructions. This converts currently live regs into live ins/outs.
457 TopRPTracker.closeTop();
458 BotRPTracker.closeBottom();
460 // Account for liveness generated by the region boundary.
461 if (LiveRegionEnd != RegionEnd)
462 BotRPTracker.recede();
464 assert(BotRPTracker.getPos() == RegionEnd && "Can't find the region bottom");
466 // Cache the list of excess pressure sets in this region. This will also track
467 // the max pressure in the scheduled code for these sets.
468 RegionCriticalPSets.clear();
469 std::vector<unsigned> RegionPressure = RPTracker.getPressure().MaxSetPressure;
470 for (unsigned i = 0, e = RegionPressure.size(); i < e; ++i) {
471 unsigned Limit = TRI->getRegPressureSetLimit(i);
472 DEBUG(dbgs() << TRI->getRegPressureSetName(i)
474 << " Actual " << RegionPressure[i] << "\n");
475 if (RegionPressure[i] > Limit)
476 RegionCriticalPSets.push_back(PressureElement(i, 0));
478 DEBUG(dbgs() << "Excess PSets: ";
479 for (unsigned i = 0, e = RegionCriticalPSets.size(); i != e; ++i)
480 dbgs() << TRI->getRegPressureSetName(
481 RegionCriticalPSets[i].PSetID) << " ";
485 // FIXME: When the pressure tracker deals in pressure differences then we won't
486 // iterate over all RegionCriticalPSets[i].
488 updateScheduledPressure(std::vector<unsigned> NewMaxPressure) {
489 for (unsigned i = 0, e = RegionCriticalPSets.size(); i < e; ++i) {
490 unsigned ID = RegionCriticalPSets[i].PSetID;
491 int &MaxUnits = RegionCriticalPSets[i].UnitIncrease;
492 if ((int)NewMaxPressure[ID] > MaxUnits)
493 MaxUnits = NewMaxPressure[ID];
497 /// schedule - Called back from MachineScheduler::runOnMachineFunction
498 /// after setting up the current scheduling region. [RegionBegin, RegionEnd)
499 /// only includes instructions that have DAG nodes, not scheduling boundaries.
501 /// This is a skeletal driver, with all the functionality pushed into helpers,
502 /// so that it can be easilly extended by experimental schedulers. Generally,
503 /// implementing MachineSchedStrategy should be sufficient to implement a new
504 /// scheduling algorithm. However, if a scheduler further subclasses
505 /// ScheduleDAGMI then it will want to override this virtual method in order to
506 /// update any specialized state.
507 void ScheduleDAGMI::schedule() {
508 buildDAGWithRegPressure();
510 Topo.InitDAGTopologicalSorting();
514 DEBUG(for (unsigned su = 0, e = SUnits.size(); su != e; ++su)
515 SUnits[su].dumpAll(this));
517 if (ViewMISchedDAGs) viewGraph();
521 bool IsTopNode = false;
522 while (SUnit *SU = SchedImpl->pickNode(IsTopNode)) {
523 assert(!SU->isScheduled && "Node already scheduled");
524 if (!checkSchedLimit())
527 scheduleMI(SU, IsTopNode);
529 updateQueues(SU, IsTopNode);
531 assert(CurrentTop == CurrentBottom && "Nonempty unscheduled zone.");
536 unsigned BBNum = begin()->getParent()->getNumber();
537 dbgs() << "*** Final schedule for BB#" << BBNum << " ***\n";
543 /// Build the DAG and setup three register pressure trackers.
544 void ScheduleDAGMI::buildDAGWithRegPressure() {
545 // Initialize the register pressure tracker used by buildSchedGraph.
546 RPTracker.init(&MF, RegClassInfo, LIS, BB, LiveRegionEnd);
548 // Account for liveness generate by the region boundary.
549 if (LiveRegionEnd != RegionEnd)
552 // Build the DAG, and compute current register pressure.
553 buildSchedGraph(AA, &RPTracker);
554 if (ViewMISchedDAGs) viewGraph();
556 // Initialize top/bottom trackers after computing region pressure.
560 /// Apply each ScheduleDAGMutation step in order.
561 void ScheduleDAGMI::postprocessDAG() {
562 for (unsigned i = 0, e = Mutations.size(); i < e; ++i) {
563 Mutations[i]->apply(this);
567 // Release all DAG roots for scheduling.
569 // Nodes with unreleased weak edges can still be roots.
570 void ScheduleDAGMI::releaseRoots() {
571 SmallVector<SUnit*, 16> BotRoots;
573 for (std::vector<SUnit>::iterator
574 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->NumPredsLeft && SU != &EntrySU)
578 SchedImpl->releaseTopNode(SU);
579 // A SUnit is ready to bottom schedule if it has no successors.
580 if (!I->NumSuccsLeft && SU != &ExitSU)
581 BotRoots.push_back(SU);
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 /// Identify DAG roots and setup scheduler queues.
591 void ScheduleDAGMI::initQueues() {
592 NextClusterSucc = NULL;
593 NextClusterPred = NULL;
595 // Initialize the strategy before modifying the DAG.
596 SchedImpl->initialize(this);
598 // Release all DAG roots for scheduling, not including EntrySU/ExitSU.
601 releaseSuccessors(&EntrySU);
602 releasePredecessors(&ExitSU);
604 SchedImpl->registerRoots();
606 CurrentTop = nextIfDebug(RegionBegin, RegionEnd);
607 CurrentBottom = RegionEnd;
610 /// Move an instruction and update register pressure.
611 void ScheduleDAGMI::scheduleMI(SUnit *SU, bool IsTopNode) {
612 // Move the instruction to its new location in the instruction stream.
613 MachineInstr *MI = SU->getInstr();
616 assert(SU->isTopReady() && "node still has unscheduled dependencies");
617 if (&*CurrentTop == MI)
618 CurrentTop = nextIfDebug(++CurrentTop, CurrentBottom);
620 moveInstruction(MI, CurrentTop);
621 TopRPTracker.setPos(MI);
624 // Update top scheduled pressure.
625 TopRPTracker.advance();
626 assert(TopRPTracker.getPos() == CurrentTop && "out of sync");
627 updateScheduledPressure(TopRPTracker.getPressure().MaxSetPressure);
630 assert(SU->isBottomReady() && "node still has unscheduled dependencies");
631 MachineBasicBlock::iterator priorII =
632 priorNonDebug(CurrentBottom, CurrentTop);
634 CurrentBottom = priorII;
636 if (&*CurrentTop == MI) {
637 CurrentTop = nextIfDebug(++CurrentTop, priorII);
638 TopRPTracker.setPos(CurrentTop);
640 moveInstruction(MI, CurrentBottom);
643 // Update bottom scheduled pressure.
644 BotRPTracker.recede();
645 assert(BotRPTracker.getPos() == CurrentBottom && "out of sync");
646 updateScheduledPressure(BotRPTracker.getPressure().MaxSetPressure);
650 /// Update scheduler queues after scheduling an instruction.
651 void ScheduleDAGMI::updateQueues(SUnit *SU, bool IsTopNode) {
652 // Release dependent instructions for scheduling.
654 releaseSuccessors(SU);
656 releasePredecessors(SU);
658 SU->isScheduled = true;
660 // Notify the scheduling strategy after updating the DAG.
661 SchedImpl->schedNode(SU, IsTopNode);
664 /// Reinsert any remaining debug_values, just like the PostRA scheduler.
665 void ScheduleDAGMI::placeDebugValues() {
666 // If first instruction was a DBG_VALUE then put it back.
668 BB->splice(RegionBegin, BB, FirstDbgValue);
669 RegionBegin = FirstDbgValue;
672 for (std::vector<std::pair<MachineInstr *, MachineInstr *> >::iterator
673 DI = DbgValues.end(), DE = DbgValues.begin(); DI != DE; --DI) {
674 std::pair<MachineInstr *, MachineInstr *> P = *prior(DI);
675 MachineInstr *DbgValue = P.first;
676 MachineBasicBlock::iterator OrigPrevMI = P.second;
677 if (&*RegionBegin == DbgValue)
679 BB->splice(++OrigPrevMI, BB, DbgValue);
680 if (OrigPrevMI == llvm::prior(RegionEnd))
681 RegionEnd = DbgValue;
684 FirstDbgValue = NULL;
687 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
688 void ScheduleDAGMI::dumpSchedule() const {
689 for (MachineBasicBlock::iterator MI = begin(), ME = end(); MI != ME; ++MI) {
690 if (SUnit *SU = getSUnit(&(*MI)))
693 dbgs() << "Missing SUnit\n";
698 //===----------------------------------------------------------------------===//
699 // LoadClusterMutation - DAG post-processing to cluster loads.
700 //===----------------------------------------------------------------------===//
703 /// \brief Post-process the DAG to create cluster edges between neighboring
705 class LoadClusterMutation : public ScheduleDAGMutation {
710 LoadInfo(SUnit *su, unsigned reg, unsigned ofs)
711 : SU(su), BaseReg(reg), Offset(ofs) {}
713 static bool LoadInfoLess(const LoadClusterMutation::LoadInfo &LHS,
714 const LoadClusterMutation::LoadInfo &RHS);
716 const TargetInstrInfo *TII;
717 const TargetRegisterInfo *TRI;
719 LoadClusterMutation(const TargetInstrInfo *tii,
720 const TargetRegisterInfo *tri)
721 : TII(tii), TRI(tri) {}
723 virtual void apply(ScheduleDAGMI *DAG);
725 void clusterNeighboringLoads(ArrayRef<SUnit*> Loads, ScheduleDAGMI *DAG);
729 bool LoadClusterMutation::LoadInfoLess(
730 const LoadClusterMutation::LoadInfo &LHS,
731 const LoadClusterMutation::LoadInfo &RHS) {
732 if (LHS.BaseReg != RHS.BaseReg)
733 return LHS.BaseReg < RHS.BaseReg;
734 return LHS.Offset < RHS.Offset;
737 void LoadClusterMutation::clusterNeighboringLoads(ArrayRef<SUnit*> Loads,
738 ScheduleDAGMI *DAG) {
739 SmallVector<LoadClusterMutation::LoadInfo,32> LoadRecords;
740 for (unsigned Idx = 0, End = Loads.size(); Idx != End; ++Idx) {
741 SUnit *SU = Loads[Idx];
744 if (TII->getLdStBaseRegImmOfs(SU->getInstr(), BaseReg, Offset, TRI))
745 LoadRecords.push_back(LoadInfo(SU, BaseReg, Offset));
747 if (LoadRecords.size() < 2)
749 std::sort(LoadRecords.begin(), LoadRecords.end(), LoadInfoLess);
750 unsigned ClusterLength = 1;
751 for (unsigned Idx = 0, End = LoadRecords.size(); Idx < (End - 1); ++Idx) {
752 if (LoadRecords[Idx].BaseReg != LoadRecords[Idx+1].BaseReg) {
757 SUnit *SUa = LoadRecords[Idx].SU;
758 SUnit *SUb = LoadRecords[Idx+1].SU;
759 if (TII->shouldClusterLoads(SUa->getInstr(), SUb->getInstr(), ClusterLength)
760 && DAG->addEdge(SUb, SDep(SUa, SDep::Cluster))) {
762 DEBUG(dbgs() << "Cluster loads SU(" << SUa->NodeNum << ") - SU("
763 << SUb->NodeNum << ")\n");
764 // Copy successor edges from SUa to SUb. Interleaving computation
765 // dependent on SUa can prevent load combining due to register reuse.
766 // Predecessor edges do not need to be copied from SUb to SUa since nearby
767 // loads should have effectively the same inputs.
768 for (SUnit::const_succ_iterator
769 SI = SUa->Succs.begin(), SE = SUa->Succs.end(); SI != SE; ++SI) {
770 if (SI->getSUnit() == SUb)
772 DEBUG(dbgs() << " Copy Succ SU(" << SI->getSUnit()->NodeNum << ")\n");
773 DAG->addEdge(SI->getSUnit(), SDep(SUb, SDep::Artificial));
782 /// \brief Callback from DAG postProcessing to create cluster edges for loads.
783 void LoadClusterMutation::apply(ScheduleDAGMI *DAG) {
784 // Map DAG NodeNum to store chain ID.
785 DenseMap<unsigned, unsigned> StoreChainIDs;
786 // Map each store chain to a set of dependent loads.
787 SmallVector<SmallVector<SUnit*,4>, 32> StoreChainDependents;
788 for (unsigned Idx = 0, End = DAG->SUnits.size(); Idx != End; ++Idx) {
789 SUnit *SU = &DAG->SUnits[Idx];
790 if (!SU->getInstr()->mayLoad())
792 unsigned ChainPredID = DAG->SUnits.size();
793 for (SUnit::const_pred_iterator
794 PI = SU->Preds.begin(), PE = SU->Preds.end(); PI != PE; ++PI) {
796 ChainPredID = PI->getSUnit()->NodeNum;
800 // Check if this chain-like pred has been seen
801 // before. ChainPredID==MaxNodeID for loads at the top of the schedule.
802 unsigned NumChains = StoreChainDependents.size();
803 std::pair<DenseMap<unsigned, unsigned>::iterator, bool> Result =
804 StoreChainIDs.insert(std::make_pair(ChainPredID, NumChains));
806 StoreChainDependents.resize(NumChains + 1);
807 StoreChainDependents[Result.first->second].push_back(SU);
809 // Iterate over the store chains.
810 for (unsigned Idx = 0, End = StoreChainDependents.size(); Idx != End; ++Idx)
811 clusterNeighboringLoads(StoreChainDependents[Idx], DAG);
814 //===----------------------------------------------------------------------===//
815 // MacroFusion - DAG post-processing to encourage fusion of macro ops.
816 //===----------------------------------------------------------------------===//
819 /// \brief Post-process the DAG to create cluster edges between instructions
820 /// that may be fused by the processor into a single operation.
821 class MacroFusion : public ScheduleDAGMutation {
822 const TargetInstrInfo *TII;
824 MacroFusion(const TargetInstrInfo *tii): TII(tii) {}
826 virtual void apply(ScheduleDAGMI *DAG);
830 /// \brief Callback from DAG postProcessing to create cluster edges to encourage
831 /// fused operations.
832 void MacroFusion::apply(ScheduleDAGMI *DAG) {
833 // For now, assume targets can only fuse with the branch.
834 MachineInstr *Branch = DAG->ExitSU.getInstr();
838 for (unsigned Idx = DAG->SUnits.size(); Idx > 0;) {
839 SUnit *SU = &DAG->SUnits[--Idx];
840 if (!TII->shouldScheduleAdjacent(SU->getInstr(), Branch))
843 // Create a single weak edge from SU to ExitSU. The only effect is to cause
844 // bottom-up scheduling to heavily prioritize the clustered SU. There is no
845 // need to copy predecessor edges from ExitSU to SU, since top-down
846 // scheduling cannot prioritize ExitSU anyway. To defer top-down scheduling
847 // of SU, we could create an artificial edge from the deepest root, but it
848 // hasn't been needed yet.
849 bool Success = DAG->addEdge(&DAG->ExitSU, SDep(SU, SDep::Cluster));
851 assert(Success && "No DAG nodes should be reachable from ExitSU");
853 DEBUG(dbgs() << "Macro Fuse SU(" << SU->NodeNum << ")\n");
858 //===----------------------------------------------------------------------===//
859 // ConvergingScheduler - Implementation of the standard MachineSchedStrategy.
860 //===----------------------------------------------------------------------===//
863 /// ConvergingScheduler shrinks the unscheduled zone using heuristics to balance
865 class ConvergingScheduler : public MachineSchedStrategy {
867 /// Represent the type of SchedCandidate found within a single queue.
868 /// pickNodeBidirectional depends on these listed by decreasing priority.
870 NoCand, SingleExcess, SingleCritical, Cluster,
871 ResourceReduce, ResourceDemand, BotHeightReduce, BotPathReduce,
872 TopDepthReduce, TopPathReduce, SingleMax, MultiPressure, NextDefUse,
876 static const char *getReasonStr(ConvergingScheduler::CandReason Reason);
879 /// Policy for scheduling the next instruction in the candidate's zone.
882 unsigned ReduceResIdx;
883 unsigned DemandResIdx;
885 CandPolicy(): ReduceLatency(false), ReduceResIdx(0), DemandResIdx(0) {}
888 /// Status of an instruction's critical resource consumption.
889 struct SchedResourceDelta {
890 // Count critical resources in the scheduled region required by SU.
891 unsigned CritResources;
893 // Count critical resources from another region consumed by SU.
894 unsigned DemandedResources;
896 SchedResourceDelta(): CritResources(0), DemandedResources(0) {}
898 bool operator==(const SchedResourceDelta &RHS) const {
899 return CritResources == RHS.CritResources
900 && DemandedResources == RHS.DemandedResources;
902 bool operator!=(const SchedResourceDelta &RHS) const {
903 return !operator==(RHS);
907 /// Store the state used by ConvergingScheduler heuristics, required for the
908 /// lifetime of one invocation of pickNode().
909 struct SchedCandidate {
912 // The best SUnit candidate.
915 // The reason for this candidate.
918 // Register pressure values for the best candidate.
919 RegPressureDelta RPDelta;
921 // Critical resource consumption of the best candidate.
922 SchedResourceDelta ResDelta;
924 SchedCandidate(const CandPolicy &policy)
925 : Policy(policy), SU(NULL), Reason(NoCand) {}
927 bool isValid() const { return SU; }
929 // Copy the status of another candidate without changing policy.
930 void setBest(SchedCandidate &Best) {
931 assert(Best.Reason != NoCand && "uninitialized Sched candidate");
933 Reason = Best.Reason;
934 RPDelta = Best.RPDelta;
935 ResDelta = Best.ResDelta;
938 void initResourceDelta(const ScheduleDAGMI *DAG,
939 const TargetSchedModel *SchedModel);
942 /// Summarize the unscheduled region.
943 struct SchedRemainder {
944 // Critical path through the DAG in expected latency.
945 unsigned CriticalPath;
947 // Unscheduled resources
948 SmallVector<unsigned, 16> RemainingCounts;
949 // Critical resource for the unscheduled zone.
951 // Number of micro-ops left to schedule.
952 unsigned RemainingMicroOps;
953 // Is the unscheduled zone resource limited.
954 bool IsResourceLimited;
956 unsigned MaxRemainingCount;
960 RemainingCounts.clear();
962 RemainingMicroOps = 0;
963 IsResourceLimited = false;
964 MaxRemainingCount = 0;
967 SchedRemainder() { reset(); }
969 void init(ScheduleDAGMI *DAG, const TargetSchedModel *SchedModel);
972 /// Each Scheduling boundary is associated with ready queues. It tracks the
973 /// current cycle in the direction of movement, and maintains the state
974 /// of "hazards" and other interlocks at the current cycle.
975 struct SchedBoundary {
977 const TargetSchedModel *SchedModel;
980 ReadyQueue Available;
984 // For heuristics, keep a list of the nodes that immediately depend on the
985 // most recently scheduled node.
986 SmallPtrSet<const SUnit*, 8> NextSUs;
988 ScheduleHazardRecognizer *HazardRec;
993 /// MinReadyCycle - Cycle of the soonest available instruction.
994 unsigned MinReadyCycle;
996 // The expected latency of the critical path in this scheduled zone.
997 unsigned ExpectedLatency;
999 // Resources used in the scheduled zone beyond this boundary.
1000 SmallVector<unsigned, 16> ResourceCounts;
1002 // Cache the critical resources ID in this scheduled zone.
1003 unsigned CritResIdx;
1005 // Is the scheduled region resource limited vs. latency limited.
1006 bool IsResourceLimited;
1008 unsigned ExpectedCount;
1010 // Policy flag: attempt to find ILP until expected latency is covered.
1011 bool ShouldIncreaseILP;
1014 // Remember the greatest min operand latency.
1015 unsigned MaxMinLatency;
1021 CheckPending = false;
1026 MinReadyCycle = UINT_MAX;
1027 ExpectedLatency = 0;
1028 ResourceCounts.resize(1);
1029 assert(!ResourceCounts[0] && "nonzero count for bad resource");
1031 IsResourceLimited = false;
1033 ShouldIncreaseILP = false;
1037 // Reserve a zero-count for invalid CritResIdx.
1038 ResourceCounts.resize(1);
1041 /// Pending queues extend the ready queues with the same ID and the
1042 /// PendingFlag set.
1043 SchedBoundary(unsigned ID, const Twine &Name):
1044 DAG(0), SchedModel(0), Rem(0), Available(ID, Name+".A"),
1045 Pending(ID << ConvergingScheduler::LogMaxQID, Name+".P") {
1049 ~SchedBoundary() { delete HazardRec; }
1051 void init(ScheduleDAGMI *dag, const TargetSchedModel *smodel,
1052 SchedRemainder *rem);
1054 bool isTop() const {
1055 return Available.getID() == ConvergingScheduler::TopQID;
1058 unsigned getUnscheduledLatency(SUnit *SU) const {
1060 return SU->getHeight();
1061 return SU->getDepth();
1064 unsigned getCriticalCount() const {
1065 return ResourceCounts[CritResIdx];
1068 bool checkHazard(SUnit *SU);
1070 void checkILPPolicy();
1072 void releaseNode(SUnit *SU, unsigned ReadyCycle);
1076 void countResource(unsigned PIdx, unsigned Cycles);
1078 void bumpNode(SUnit *SU);
1080 void releasePending();
1082 void removeReady(SUnit *SU);
1084 SUnit *pickOnlyChoice();
1089 const TargetSchedModel *SchedModel;
1090 const TargetRegisterInfo *TRI;
1092 // State of the top and bottom scheduled instruction boundaries.
1098 /// SUnit::NodeQueueId: 0 (none), 1 (top), 2 (bot), 3 (both)
1105 ConvergingScheduler():
1106 DAG(0), SchedModel(0), TRI(0), Top(TopQID, "TopQ"), Bot(BotQID, "BotQ") {}
1108 virtual void initialize(ScheduleDAGMI *dag);
1110 virtual SUnit *pickNode(bool &IsTopNode);
1112 virtual void schedNode(SUnit *SU, bool IsTopNode);
1114 virtual void releaseTopNode(SUnit *SU);
1116 virtual void releaseBottomNode(SUnit *SU);
1118 virtual void registerRoots();
1122 ConvergingScheduler::SchedBoundary &CriticalZone,
1123 ConvergingScheduler::SchedCandidate &CriticalCand,
1124 ConvergingScheduler::SchedBoundary &OppositeZone,
1125 ConvergingScheduler::SchedCandidate &OppositeCand);
1127 void checkResourceLimits(ConvergingScheduler::SchedCandidate &TopCand,
1128 ConvergingScheduler::SchedCandidate &BotCand);
1130 void tryCandidate(SchedCandidate &Cand,
1131 SchedCandidate &TryCand,
1132 SchedBoundary &Zone,
1133 const RegPressureTracker &RPTracker,
1134 RegPressureTracker &TempTracker);
1136 SUnit *pickNodeBidirectional(bool &IsTopNode);
1138 void pickNodeFromQueue(SchedBoundary &Zone,
1139 const RegPressureTracker &RPTracker,
1140 SchedCandidate &Candidate);
1143 void traceCandidate(const SchedCandidate &Cand, const SchedBoundary &Zone);
1148 void ConvergingScheduler::SchedRemainder::
1149 init(ScheduleDAGMI *DAG, const TargetSchedModel *SchedModel) {
1151 if (!SchedModel->hasInstrSchedModel())
1153 RemainingCounts.resize(SchedModel->getNumProcResourceKinds());
1154 for (std::vector<SUnit>::iterator
1155 I = DAG->SUnits.begin(), E = DAG->SUnits.end(); I != E; ++I) {
1156 const MCSchedClassDesc *SC = DAG->getSchedClass(&*I);
1157 RemainingMicroOps += SchedModel->getNumMicroOps(I->getInstr(), SC);
1158 for (TargetSchedModel::ProcResIter
1159 PI = SchedModel->getWriteProcResBegin(SC),
1160 PE = SchedModel->getWriteProcResEnd(SC); PI != PE; ++PI) {
1161 unsigned PIdx = PI->ProcResourceIdx;
1162 unsigned Factor = SchedModel->getResourceFactor(PIdx);
1163 RemainingCounts[PIdx] += (Factor * PI->Cycles);
1168 void ConvergingScheduler::SchedBoundary::
1169 init(ScheduleDAGMI *dag, const TargetSchedModel *smodel, SchedRemainder *rem) {
1172 SchedModel = smodel;
1174 if (SchedModel->hasInstrSchedModel())
1175 ResourceCounts.resize(SchedModel->getNumProcResourceKinds());
1178 void ConvergingScheduler::initialize(ScheduleDAGMI *dag) {
1180 SchedModel = DAG->getSchedModel();
1182 Rem.init(DAG, SchedModel);
1183 Top.init(DAG, SchedModel, &Rem);
1184 Bot.init(DAG, SchedModel, &Rem);
1186 // Initialize resource counts.
1188 // Initialize the HazardRecognizers. If itineraries don't exist, are empty, or
1189 // are disabled, then these HazardRecs will be disabled.
1190 const InstrItineraryData *Itin = SchedModel->getInstrItineraries();
1191 const TargetMachine &TM = DAG->MF.getTarget();
1192 Top.HazardRec = TM.getInstrInfo()->CreateTargetMIHazardRecognizer(Itin, DAG);
1193 Bot.HazardRec = TM.getInstrInfo()->CreateTargetMIHazardRecognizer(Itin, DAG);
1195 assert((!ForceTopDown || !ForceBottomUp) &&
1196 "-misched-topdown incompatible with -misched-bottomup");
1199 void ConvergingScheduler::releaseTopNode(SUnit *SU) {
1200 if (SU->isScheduled)
1203 for (SUnit::succ_iterator I = SU->Preds.begin(), E = SU->Preds.end();
1205 unsigned PredReadyCycle = I->getSUnit()->TopReadyCycle;
1206 unsigned MinLatency = I->getMinLatency();
1208 Top.MaxMinLatency = std::max(MinLatency, Top.MaxMinLatency);
1210 if (SU->TopReadyCycle < PredReadyCycle + MinLatency)
1211 SU->TopReadyCycle = PredReadyCycle + MinLatency;
1213 Top.releaseNode(SU, SU->TopReadyCycle);
1216 void ConvergingScheduler::releaseBottomNode(SUnit *SU) {
1217 if (SU->isScheduled)
1220 assert(SU->getInstr() && "Scheduled SUnit must have instr");
1222 for (SUnit::succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
1226 unsigned SuccReadyCycle = I->getSUnit()->BotReadyCycle;
1227 unsigned MinLatency = I->getMinLatency();
1229 Bot.MaxMinLatency = std::max(MinLatency, Bot.MaxMinLatency);
1231 if (SU->BotReadyCycle < SuccReadyCycle + MinLatency)
1232 SU->BotReadyCycle = SuccReadyCycle + MinLatency;
1234 Bot.releaseNode(SU, SU->BotReadyCycle);
1237 void ConvergingScheduler::registerRoots() {
1238 Rem.CriticalPath = DAG->ExitSU.getDepth();
1239 // Some roots may not feed into ExitSU. Check all of them in case.
1240 for (std::vector<SUnit*>::const_iterator
1241 I = Bot.Available.begin(), E = Bot.Available.end(); I != E; ++I) {
1242 if ((*I)->getDepth() > Rem.CriticalPath)
1243 Rem.CriticalPath = (*I)->getDepth();
1245 DEBUG(dbgs() << "Critical Path: " << Rem.CriticalPath << '\n');
1248 /// Does this SU have a hazard within the current instruction group.
1250 /// The scheduler supports two modes of hazard recognition. The first is the
1251 /// ScheduleHazardRecognizer API. It is a fully general hazard recognizer that
1252 /// supports highly complicated in-order reservation tables
1253 /// (ScoreboardHazardRecognizer) and arbitraty target-specific logic.
1255 /// The second is a streamlined mechanism that checks for hazards based on
1256 /// simple counters that the scheduler itself maintains. It explicitly checks
1257 /// for instruction dispatch limitations, including the number of micro-ops that
1258 /// can dispatch per cycle.
1260 /// TODO: Also check whether the SU must start a new group.
1261 bool ConvergingScheduler::SchedBoundary::checkHazard(SUnit *SU) {
1262 if (HazardRec->isEnabled())
1263 return HazardRec->getHazardType(SU) != ScheduleHazardRecognizer::NoHazard;
1265 unsigned uops = SchedModel->getNumMicroOps(SU->getInstr());
1266 if ((IssueCount > 0) && (IssueCount + uops > SchedModel->getIssueWidth())) {
1267 DEBUG(dbgs() << " SU(" << SU->NodeNum << ") uops="
1268 << SchedModel->getNumMicroOps(SU->getInstr()) << '\n');
1274 /// If expected latency is covered, disable ILP policy.
1275 void ConvergingScheduler::SchedBoundary::checkILPPolicy() {
1276 if (ShouldIncreaseILP
1277 && (IsResourceLimited || ExpectedLatency <= CurrCycle)) {
1278 ShouldIncreaseILP = false;
1279 DEBUG(dbgs() << "Disable ILP: " << Available.getName() << '\n');
1283 void ConvergingScheduler::SchedBoundary::releaseNode(SUnit *SU,
1284 unsigned ReadyCycle) {
1286 if (ReadyCycle < MinReadyCycle)
1287 MinReadyCycle = ReadyCycle;
1289 // Check for interlocks first. For the purpose of other heuristics, an
1290 // instruction that cannot issue appears as if it's not in the ReadyQueue.
1291 if (ReadyCycle > CurrCycle || checkHazard(SU))
1296 // Record this node as an immediate dependent of the scheduled node.
1299 // If CriticalPath has been computed, then check if the unscheduled nodes
1300 // exceed the ILP window. Before registerRoots, CriticalPath==0.
1301 if (Rem->CriticalPath && (ExpectedLatency + getUnscheduledLatency(SU)
1302 > Rem->CriticalPath + ILPWindow)) {
1303 ShouldIncreaseILP = true;
1304 DEBUG(dbgs() << "Increase ILP: " << Available.getName() << " "
1305 << ExpectedLatency << " + " << getUnscheduledLatency(SU) << '\n');
1309 /// Move the boundary of scheduled code by one cycle.
1310 void ConvergingScheduler::SchedBoundary::bumpCycle() {
1311 unsigned Width = SchedModel->getIssueWidth();
1312 IssueCount = (IssueCount <= Width) ? 0 : IssueCount - Width;
1314 unsigned NextCycle = CurrCycle + 1;
1315 assert(MinReadyCycle < UINT_MAX && "MinReadyCycle uninitialized");
1316 if (MinReadyCycle > NextCycle) {
1318 NextCycle = MinReadyCycle;
1321 if (!HazardRec->isEnabled()) {
1322 // Bypass HazardRec virtual calls.
1323 CurrCycle = NextCycle;
1326 // Bypass getHazardType calls in case of long latency.
1327 for (; CurrCycle != NextCycle; ++CurrCycle) {
1329 HazardRec->AdvanceCycle();
1331 HazardRec->RecedeCycle();
1334 CheckPending = true;
1335 IsResourceLimited = getCriticalCount() > std::max(ExpectedLatency, CurrCycle);
1337 DEBUG(dbgs() << " *** " << Available.getName() << " cycle "
1338 << CurrCycle << '\n');
1341 /// Add the given processor resource to this scheduled zone.
1342 void ConvergingScheduler::SchedBoundary::countResource(unsigned PIdx,
1344 unsigned Factor = SchedModel->getResourceFactor(PIdx);
1345 DEBUG(dbgs() << " " << SchedModel->getProcResource(PIdx)->Name
1346 << " +(" << Cycles << "x" << Factor
1347 << ") / " << SchedModel->getLatencyFactor() << '\n');
1349 unsigned Count = Factor * Cycles;
1350 ResourceCounts[PIdx] += Count;
1351 assert(Rem->RemainingCounts[PIdx] >= Count && "resource double counted");
1352 Rem->RemainingCounts[PIdx] -= Count;
1354 // Reset MaxRemainingCount for sanity.
1355 Rem->MaxRemainingCount = 0;
1357 // Check if this resource exceeds the current critical resource by a full
1358 // cycle. If so, it becomes the critical resource.
1359 if ((int)(ResourceCounts[PIdx] - ResourceCounts[CritResIdx])
1360 >= (int)SchedModel->getLatencyFactor()) {
1362 DEBUG(dbgs() << " *** Critical resource "
1363 << SchedModel->getProcResource(PIdx)->Name << " x"
1364 << ResourceCounts[PIdx] << '\n');
1368 /// Move the boundary of scheduled code by one SUnit.
1369 void ConvergingScheduler::SchedBoundary::bumpNode(SUnit *SU) {
1370 // Update the reservation table.
1371 if (HazardRec->isEnabled()) {
1372 if (!isTop() && SU->isCall) {
1373 // Calls are scheduled with their preceding instructions. For bottom-up
1374 // scheduling, clear the pipeline state before emitting.
1377 HazardRec->EmitInstruction(SU);
1379 // Update resource counts and critical resource.
1380 if (SchedModel->hasInstrSchedModel()) {
1381 const MCSchedClassDesc *SC = DAG->getSchedClass(SU);
1382 Rem->RemainingMicroOps -= SchedModel->getNumMicroOps(SU->getInstr(), SC);
1383 for (TargetSchedModel::ProcResIter
1384 PI = SchedModel->getWriteProcResBegin(SC),
1385 PE = SchedModel->getWriteProcResEnd(SC); PI != PE; ++PI) {
1386 countResource(PI->ProcResourceIdx, PI->Cycles);
1390 if (SU->getDepth() > ExpectedLatency)
1391 ExpectedLatency = SU->getDepth();
1394 if (SU->getHeight() > ExpectedLatency)
1395 ExpectedLatency = SU->getHeight();
1398 IsResourceLimited = getCriticalCount() > std::max(ExpectedLatency, CurrCycle);
1400 // Check the instruction group dispatch limit.
1401 // TODO: Check if this SU must end a dispatch group.
1402 IssueCount += SchedModel->getNumMicroOps(SU->getInstr());
1404 // checkHazard prevents scheduling multiple instructions per cycle that exceed
1405 // issue width. However, we commonly reach the maximum. In this case
1406 // opportunistically bump the cycle to avoid uselessly checking everything in
1407 // the readyQ. Furthermore, a single instruction may produce more than one
1408 // cycle's worth of micro-ops.
1409 if (IssueCount >= SchedModel->getIssueWidth()) {
1410 DEBUG(dbgs() << " *** Max instrs at cycle " << CurrCycle << '\n');
1415 /// Release pending ready nodes in to the available queue. This makes them
1416 /// visible to heuristics.
1417 void ConvergingScheduler::SchedBoundary::releasePending() {
1418 // If the available queue is empty, it is safe to reset MinReadyCycle.
1419 if (Available.empty())
1420 MinReadyCycle = UINT_MAX;
1422 // Check to see if any of the pending instructions are ready to issue. If
1423 // so, add them to the available queue.
1424 for (unsigned i = 0, e = Pending.size(); i != e; ++i) {
1425 SUnit *SU = *(Pending.begin()+i);
1426 unsigned ReadyCycle = isTop() ? SU->TopReadyCycle : SU->BotReadyCycle;
1428 if (ReadyCycle < MinReadyCycle)
1429 MinReadyCycle = ReadyCycle;
1431 if (ReadyCycle > CurrCycle)
1434 if (checkHazard(SU))
1438 Pending.remove(Pending.begin()+i);
1441 DEBUG(if (!Pending.empty()) Pending.dump());
1442 CheckPending = false;
1445 /// Remove SU from the ready set for this boundary.
1446 void ConvergingScheduler::SchedBoundary::removeReady(SUnit *SU) {
1447 if (Available.isInQueue(SU))
1448 Available.remove(Available.find(SU));
1450 assert(Pending.isInQueue(SU) && "bad ready count");
1451 Pending.remove(Pending.find(SU));
1455 /// If this queue only has one ready candidate, return it. As a side effect,
1456 /// defer any nodes that now hit a hazard, and advance the cycle until at least
1457 /// one node is ready. If multiple instructions are ready, return NULL.
1458 SUnit *ConvergingScheduler::SchedBoundary::pickOnlyChoice() {
1462 if (IssueCount > 0) {
1463 // Defer any ready instrs that now have a hazard.
1464 for (ReadyQueue::iterator I = Available.begin(); I != Available.end();) {
1465 if (checkHazard(*I)) {
1467 I = Available.remove(I);
1473 for (unsigned i = 0; Available.empty(); ++i) {
1474 assert(i <= (HazardRec->getMaxLookAhead() + MaxMinLatency) &&
1475 "permanent hazard"); (void)i;
1479 if (Available.size() == 1)
1480 return *Available.begin();
1484 /// Record the candidate policy for opposite zones with different critical
1487 /// If the CriticalZone is latency limited, don't force a policy for the
1488 /// candidates here. Instead, When releasing each candidate, releaseNode
1489 /// compares the region's critical path to the candidate's height or depth and
1490 /// the scheduled zone's expected latency then sets ShouldIncreaseILP.
1491 void ConvergingScheduler::balanceZones(
1492 ConvergingScheduler::SchedBoundary &CriticalZone,
1493 ConvergingScheduler::SchedCandidate &CriticalCand,
1494 ConvergingScheduler::SchedBoundary &OppositeZone,
1495 ConvergingScheduler::SchedCandidate &OppositeCand) {
1497 if (!CriticalZone.IsResourceLimited)
1500 SchedRemainder *Rem = CriticalZone.Rem;
1502 // If the critical zone is overconsuming a resource relative to the
1503 // remainder, try to reduce it.
1504 unsigned RemainingCritCount =
1505 Rem->RemainingCounts[CriticalZone.CritResIdx];
1506 if ((int)(Rem->MaxRemainingCount - RemainingCritCount)
1507 > (int)SchedModel->getLatencyFactor()) {
1508 CriticalCand.Policy.ReduceResIdx = CriticalZone.CritResIdx;
1509 DEBUG(dbgs() << "Balance " << CriticalZone.Available.getName() << " reduce "
1510 << SchedModel->getProcResource(CriticalZone.CritResIdx)->Name
1513 // If the other zone is underconsuming a resource relative to the full zone,
1514 // try to increase it.
1515 unsigned OppositeCount =
1516 OppositeZone.ResourceCounts[CriticalZone.CritResIdx];
1517 if ((int)(OppositeZone.ExpectedCount - OppositeCount)
1518 > (int)SchedModel->getLatencyFactor()) {
1519 OppositeCand.Policy.DemandResIdx = CriticalZone.CritResIdx;
1520 DEBUG(dbgs() << "Balance " << OppositeZone.Available.getName() << " demand "
1521 << SchedModel->getProcResource(OppositeZone.CritResIdx)->Name
1526 /// Determine if the scheduled zones exceed resource limits or critical path and
1527 /// set each candidate's ReduceHeight policy accordingly.
1528 void ConvergingScheduler::checkResourceLimits(
1529 ConvergingScheduler::SchedCandidate &TopCand,
1530 ConvergingScheduler::SchedCandidate &BotCand) {
1532 Bot.checkILPPolicy();
1533 Top.checkILPPolicy();
1534 if (Bot.ShouldIncreaseILP)
1535 BotCand.Policy.ReduceLatency = true;
1536 if (Top.ShouldIncreaseILP)
1537 TopCand.Policy.ReduceLatency = true;
1539 // Handle resource-limited regions.
1540 if (Top.IsResourceLimited && Bot.IsResourceLimited
1541 && Top.CritResIdx == Bot.CritResIdx) {
1542 // If the scheduled critical resource in both zones is no longer the
1543 // critical remaining resource, attempt to reduce resource height both ways.
1544 if (Top.CritResIdx != Rem.CritResIdx) {
1545 TopCand.Policy.ReduceResIdx = Top.CritResIdx;
1546 BotCand.Policy.ReduceResIdx = Bot.CritResIdx;
1547 DEBUG(dbgs() << "Reduce scheduled "
1548 << SchedModel->getProcResource(Top.CritResIdx)->Name << '\n');
1552 // Handle latency-limited regions.
1553 if (!Top.IsResourceLimited && !Bot.IsResourceLimited) {
1554 // If the total scheduled expected latency exceeds the region's critical
1555 // path then reduce latency both ways.
1557 // Just because a zone is not resource limited does not mean it is latency
1558 // limited. Unbuffered resource, such as max micro-ops may cause CurrCycle
1559 // to exceed expected latency.
1560 if ((Top.ExpectedLatency + Bot.ExpectedLatency >= Rem.CriticalPath)
1561 && (Rem.CriticalPath > Top.CurrCycle + Bot.CurrCycle)) {
1562 TopCand.Policy.ReduceLatency = true;
1563 BotCand.Policy.ReduceLatency = true;
1564 DEBUG(dbgs() << "Reduce scheduled latency " << Top.ExpectedLatency
1565 << " + " << Bot.ExpectedLatency << '\n');
1569 // The critical resource is different in each zone, so request balancing.
1571 // Compute the cost of each zone.
1572 Rem.MaxRemainingCount = std::max(
1573 Rem.RemainingMicroOps * SchedModel->getMicroOpFactor(),
1574 Rem.RemainingCounts[Rem.CritResIdx]);
1575 Top.ExpectedCount = std::max(Top.ExpectedLatency, Top.CurrCycle);
1576 Top.ExpectedCount = std::max(
1577 Top.getCriticalCount(),
1578 Top.ExpectedCount * SchedModel->getLatencyFactor());
1579 Bot.ExpectedCount = std::max(Bot.ExpectedLatency, Bot.CurrCycle);
1580 Bot.ExpectedCount = std::max(
1581 Bot.getCriticalCount(),
1582 Bot.ExpectedCount * SchedModel->getLatencyFactor());
1584 balanceZones(Top, TopCand, Bot, BotCand);
1585 balanceZones(Bot, BotCand, Top, TopCand);
1588 void ConvergingScheduler::SchedCandidate::
1589 initResourceDelta(const ScheduleDAGMI *DAG,
1590 const TargetSchedModel *SchedModel) {
1591 if (!Policy.ReduceResIdx && !Policy.DemandResIdx)
1594 const MCSchedClassDesc *SC = DAG->getSchedClass(SU);
1595 for (TargetSchedModel::ProcResIter
1596 PI = SchedModel->getWriteProcResBegin(SC),
1597 PE = SchedModel->getWriteProcResEnd(SC); PI != PE; ++PI) {
1598 if (PI->ProcResourceIdx == Policy.ReduceResIdx)
1599 ResDelta.CritResources += PI->Cycles;
1600 if (PI->ProcResourceIdx == Policy.DemandResIdx)
1601 ResDelta.DemandedResources += PI->Cycles;
1605 /// Return true if this heuristic determines order.
1606 static bool tryLess(unsigned TryVal, unsigned CandVal,
1607 ConvergingScheduler::SchedCandidate &TryCand,
1608 ConvergingScheduler::SchedCandidate &Cand,
1609 ConvergingScheduler::CandReason Reason) {
1610 if (TryVal < CandVal) {
1611 TryCand.Reason = Reason;
1614 if (TryVal > CandVal) {
1615 if (Cand.Reason > Reason)
1616 Cand.Reason = Reason;
1622 static bool tryGreater(unsigned TryVal, unsigned CandVal,
1623 ConvergingScheduler::SchedCandidate &TryCand,
1624 ConvergingScheduler::SchedCandidate &Cand,
1625 ConvergingScheduler::CandReason Reason) {
1626 if (TryVal > CandVal) {
1627 TryCand.Reason = Reason;
1630 if (TryVal < CandVal) {
1631 if (Cand.Reason > Reason)
1632 Cand.Reason = Reason;
1638 static unsigned getWeakLeft(const SUnit *SU, bool isTop) {
1639 return (isTop) ? SU->WeakPredsLeft : SU->WeakSuccsLeft;
1642 /// Apply a set of heursitics to a new candidate. Heuristics are currently
1643 /// hierarchical. This may be more efficient than a graduated cost model because
1644 /// we don't need to evaluate all aspects of the model for each node in the
1645 /// queue. But it's really done to make the heuristics easier to debug and
1646 /// statistically analyze.
1648 /// \param Cand provides the policy and current best candidate.
1649 /// \param TryCand refers to the next SUnit candidate, otherwise uninitialized.
1650 /// \param Zone describes the scheduled zone that we are extending.
1651 /// \param RPTracker describes reg pressure within the scheduled zone.
1652 /// \param TempTracker is a scratch pressure tracker to reuse in queries.
1653 void ConvergingScheduler::tryCandidate(SchedCandidate &Cand,
1654 SchedCandidate &TryCand,
1655 SchedBoundary &Zone,
1656 const RegPressureTracker &RPTracker,
1657 RegPressureTracker &TempTracker) {
1659 // Always initialize TryCand's RPDelta.
1660 TempTracker.getMaxPressureDelta(TryCand.SU->getInstr(), TryCand.RPDelta,
1661 DAG->getRegionCriticalPSets(),
1662 DAG->getRegPressure().MaxSetPressure);
1664 // Initialize the candidate if needed.
1665 if (!Cand.isValid()) {
1666 TryCand.Reason = NodeOrder;
1669 // Avoid exceeding the target's limit.
1670 if (tryLess(TryCand.RPDelta.Excess.UnitIncrease,
1671 Cand.RPDelta.Excess.UnitIncrease, TryCand, Cand, SingleExcess))
1673 if (Cand.Reason == SingleExcess)
1674 Cand.Reason = MultiPressure;
1676 // Avoid increasing the max critical pressure in the scheduled region.
1677 if (tryLess(TryCand.RPDelta.CriticalMax.UnitIncrease,
1678 Cand.RPDelta.CriticalMax.UnitIncrease,
1679 TryCand, Cand, SingleCritical))
1681 if (Cand.Reason == SingleCritical)
1682 Cand.Reason = MultiPressure;
1684 // Keep clustered nodes together to encourage downstream peephole
1685 // optimizations which may reduce resource requirements.
1687 // This is a best effort to set things up for a post-RA pass. Optimizations
1688 // like generating loads of multiple registers should ideally be done within
1689 // the scheduler pass by combining the loads during DAG postprocessing.
1690 const SUnit *NextClusterSU =
1691 Zone.isTop() ? DAG->getNextClusterSucc() : DAG->getNextClusterPred();
1692 if (tryGreater(TryCand.SU == NextClusterSU, Cand.SU == NextClusterSU,
1693 TryCand, Cand, Cluster))
1695 // Currently, weak edges are for clustering, so we hard-code that reason.
1696 // However, deferring the current TryCand will not change Cand's reason.
1697 CandReason OrigReason = Cand.Reason;
1698 if (tryLess(getWeakLeft(TryCand.SU, Zone.isTop()),
1699 getWeakLeft(Cand.SU, Zone.isTop()),
1700 TryCand, Cand, Cluster)) {
1701 Cand.Reason = OrigReason;
1704 // Avoid critical resource consumption and balance the schedule.
1705 TryCand.initResourceDelta(DAG, SchedModel);
1706 if (tryLess(TryCand.ResDelta.CritResources, Cand.ResDelta.CritResources,
1707 TryCand, Cand, ResourceReduce))
1709 if (tryGreater(TryCand.ResDelta.DemandedResources,
1710 Cand.ResDelta.DemandedResources,
1711 TryCand, Cand, ResourceDemand))
1714 // Avoid serializing long latency dependence chains.
1715 if (Cand.Policy.ReduceLatency) {
1717 if (Cand.SU->getDepth() * SchedModel->getLatencyFactor()
1718 > Zone.ExpectedCount) {
1719 if (tryLess(TryCand.SU->getDepth(), Cand.SU->getDepth(),
1720 TryCand, Cand, TopDepthReduce))
1723 if (tryGreater(TryCand.SU->getHeight(), Cand.SU->getHeight(),
1724 TryCand, Cand, TopPathReduce))
1728 if (Cand.SU->getHeight() * SchedModel->getLatencyFactor()
1729 > Zone.ExpectedCount) {
1730 if (tryLess(TryCand.SU->getHeight(), Cand.SU->getHeight(),
1731 TryCand, Cand, BotHeightReduce))
1734 if (tryGreater(TryCand.SU->getDepth(), Cand.SU->getDepth(),
1735 TryCand, Cand, BotPathReduce))
1740 // Avoid increasing the max pressure of the entire region.
1741 if (tryLess(TryCand.RPDelta.CurrentMax.UnitIncrease,
1742 Cand.RPDelta.CurrentMax.UnitIncrease, TryCand, Cand, SingleMax))
1744 if (Cand.Reason == SingleMax)
1745 Cand.Reason = MultiPressure;
1747 // Prefer immediate defs/users of the last scheduled instruction. This is a
1748 // nice pressure avoidance strategy that also conserves the processor's
1749 // register renaming resources and keeps the machine code readable.
1750 if (tryGreater(Zone.NextSUs.count(TryCand.SU), Zone.NextSUs.count(Cand.SU),
1751 TryCand, Cand, NextDefUse))
1754 // Fall through to original instruction order.
1755 if ((Zone.isTop() && TryCand.SU->NodeNum < Cand.SU->NodeNum)
1756 || (!Zone.isTop() && TryCand.SU->NodeNum > Cand.SU->NodeNum)) {
1757 TryCand.Reason = NodeOrder;
1761 /// pickNodeFromQueue helper that returns true if the LHS reg pressure effect is
1762 /// more desirable than RHS from scheduling standpoint.
1763 static bool compareRPDelta(const RegPressureDelta &LHS,
1764 const RegPressureDelta &RHS) {
1765 // Compare each component of pressure in decreasing order of importance
1766 // without checking if any are valid. Invalid PressureElements are assumed to
1767 // have UnitIncrease==0, so are neutral.
1769 // Avoid increasing the max critical pressure in the scheduled region.
1770 if (LHS.Excess.UnitIncrease != RHS.Excess.UnitIncrease) {
1771 DEBUG(dbgs() << "RP excess top - bot: "
1772 << (LHS.Excess.UnitIncrease - RHS.Excess.UnitIncrease) << '\n');
1773 return LHS.Excess.UnitIncrease < RHS.Excess.UnitIncrease;
1775 // Avoid increasing the max critical pressure in the scheduled region.
1776 if (LHS.CriticalMax.UnitIncrease != RHS.CriticalMax.UnitIncrease) {
1777 DEBUG(dbgs() << "RP critical top - bot: "
1778 << (LHS.CriticalMax.UnitIncrease - RHS.CriticalMax.UnitIncrease)
1780 return LHS.CriticalMax.UnitIncrease < RHS.CriticalMax.UnitIncrease;
1782 // Avoid increasing the max pressure of the entire region.
1783 if (LHS.CurrentMax.UnitIncrease != RHS.CurrentMax.UnitIncrease) {
1784 DEBUG(dbgs() << "RP current top - bot: "
1785 << (LHS.CurrentMax.UnitIncrease - RHS.CurrentMax.UnitIncrease)
1787 return LHS.CurrentMax.UnitIncrease < RHS.CurrentMax.UnitIncrease;
1793 const char *ConvergingScheduler::getReasonStr(
1794 ConvergingScheduler::CandReason Reason) {
1796 case NoCand: return "NOCAND ";
1797 case SingleExcess: return "REG-EXCESS";
1798 case SingleCritical: return "REG-CRIT ";
1799 case Cluster: return "CLUSTER ";
1800 case SingleMax: return "REG-MAX ";
1801 case MultiPressure: return "REG-MULTI ";
1802 case ResourceReduce: return "RES-REDUCE";
1803 case ResourceDemand: return "RES-DEMAND";
1804 case TopDepthReduce: return "TOP-DEPTH ";
1805 case TopPathReduce: return "TOP-PATH ";
1806 case BotHeightReduce:return "BOT-HEIGHT";
1807 case BotPathReduce: return "BOT-PATH ";
1808 case NextDefUse: return "DEF-USE ";
1809 case NodeOrder: return "ORDER ";
1811 llvm_unreachable("Unknown reason!");
1814 void ConvergingScheduler::traceCandidate(const SchedCandidate &Cand,
1815 const SchedBoundary &Zone) {
1816 const char *Label = getReasonStr(Cand.Reason);
1818 unsigned ResIdx = 0;
1819 unsigned Latency = 0;
1820 switch (Cand.Reason) {
1824 P = Cand.RPDelta.Excess;
1826 case SingleCritical:
1827 P = Cand.RPDelta.CriticalMax;
1830 P = Cand.RPDelta.CurrentMax;
1832 case ResourceReduce:
1833 ResIdx = Cand.Policy.ReduceResIdx;
1835 case ResourceDemand:
1836 ResIdx = Cand.Policy.DemandResIdx;
1838 case TopDepthReduce:
1839 Latency = Cand.SU->getDepth();
1842 Latency = Cand.SU->getHeight();
1844 case BotHeightReduce:
1845 Latency = Cand.SU->getHeight();
1848 Latency = Cand.SU->getDepth();
1851 dbgs() << Label << " " << Zone.Available.getName() << " ";
1853 dbgs() << TRI->getRegPressureSetName(P.PSetID) << ":" << P.UnitIncrease
1858 dbgs() << SchedModel->getProcResource(ResIdx)->Name << " ";
1862 dbgs() << Latency << " cycles ";
1869 /// Pick the best candidate from the top queue.
1871 /// TODO: getMaxPressureDelta results can be mostly cached for each SUnit during
1872 /// DAG building. To adjust for the current scheduling location we need to
1873 /// maintain the number of vreg uses remaining to be top-scheduled.
1874 void ConvergingScheduler::pickNodeFromQueue(SchedBoundary &Zone,
1875 const RegPressureTracker &RPTracker,
1876 SchedCandidate &Cand) {
1877 ReadyQueue &Q = Zone.Available;
1881 // getMaxPressureDelta temporarily modifies the tracker.
1882 RegPressureTracker &TempTracker = const_cast<RegPressureTracker&>(RPTracker);
1884 for (ReadyQueue::iterator I = Q.begin(), E = Q.end(); I != E; ++I) {
1886 SchedCandidate TryCand(Cand.Policy);
1888 tryCandidate(Cand, TryCand, Zone, RPTracker, TempTracker);
1889 if (TryCand.Reason != NoCand) {
1890 // Initialize resource delta if needed in case future heuristics query it.
1891 if (TryCand.ResDelta == SchedResourceDelta())
1892 TryCand.initResourceDelta(DAG, SchedModel);
1893 Cand.setBest(TryCand);
1894 DEBUG(traceCandidate(Cand, Zone));
1900 static void tracePick(const ConvergingScheduler::SchedCandidate &Cand,
1902 DEBUG(dbgs() << "Pick " << (IsTop ? "top" : "bot")
1903 << " SU(" << Cand.SU->NodeNum << ") "
1904 << ConvergingScheduler::getReasonStr(Cand.Reason) << '\n');
1907 /// Pick the best candidate node from either the top or bottom queue.
1908 SUnit *ConvergingScheduler::pickNodeBidirectional(bool &IsTopNode) {
1909 // Schedule as far as possible in the direction of no choice. This is most
1910 // efficient, but also provides the best heuristics for CriticalPSets.
1911 if (SUnit *SU = Bot.pickOnlyChoice()) {
1915 if (SUnit *SU = Top.pickOnlyChoice()) {
1919 CandPolicy NoPolicy;
1920 SchedCandidate BotCand(NoPolicy);
1921 SchedCandidate TopCand(NoPolicy);
1922 checkResourceLimits(TopCand, BotCand);
1924 // Prefer bottom scheduling when heuristics are silent.
1925 pickNodeFromQueue(Bot, DAG->getBotRPTracker(), BotCand);
1926 assert(BotCand.Reason != NoCand && "failed to find the first candidate");
1928 // If either Q has a single candidate that provides the least increase in
1929 // Excess pressure, we can immediately schedule from that Q.
1931 // RegionCriticalPSets summarizes the pressure within the scheduled region and
1932 // affects picking from either Q. If scheduling in one direction must
1933 // increase pressure for one of the excess PSets, then schedule in that
1934 // direction first to provide more freedom in the other direction.
1935 if (BotCand.Reason == SingleExcess || BotCand.Reason == SingleCritical) {
1937 tracePick(BotCand, IsTopNode);
1940 // Check if the top Q has a better candidate.
1941 pickNodeFromQueue(Top, DAG->getTopRPTracker(), TopCand);
1942 assert(TopCand.Reason != NoCand && "failed to find the first candidate");
1944 // If either Q has a single candidate that minimizes pressure above the
1945 // original region's pressure pick it.
1946 if (TopCand.Reason <= SingleMax || BotCand.Reason <= SingleMax) {
1947 if (TopCand.Reason < BotCand.Reason) {
1949 tracePick(TopCand, IsTopNode);
1953 tracePick(BotCand, IsTopNode);
1956 // Check for a salient pressure difference and pick the best from either side.
1957 if (compareRPDelta(TopCand.RPDelta, BotCand.RPDelta)) {
1959 tracePick(TopCand, IsTopNode);
1962 // Otherwise prefer the bottom candidate, in node order if all else failed.
1963 if (TopCand.Reason < BotCand.Reason) {
1965 tracePick(TopCand, IsTopNode);
1969 tracePick(BotCand, IsTopNode);
1973 /// Pick the best node to balance the schedule. Implements MachineSchedStrategy.
1974 SUnit *ConvergingScheduler::pickNode(bool &IsTopNode) {
1975 if (DAG->top() == DAG->bottom()) {
1976 assert(Top.Available.empty() && Top.Pending.empty() &&
1977 Bot.Available.empty() && Bot.Pending.empty() && "ReadyQ garbage");
1983 SU = Top.pickOnlyChoice();
1985 CandPolicy NoPolicy;
1986 SchedCandidate TopCand(NoPolicy);
1987 pickNodeFromQueue(Top, DAG->getTopRPTracker(), TopCand);
1988 assert(TopCand.Reason != NoCand && "failed to find the first candidate");
1993 else if (ForceBottomUp) {
1994 SU = Bot.pickOnlyChoice();
1996 CandPolicy NoPolicy;
1997 SchedCandidate BotCand(NoPolicy);
1998 pickNodeFromQueue(Bot, DAG->getBotRPTracker(), BotCand);
1999 assert(BotCand.Reason != NoCand && "failed to find the first candidate");
2005 SU = pickNodeBidirectional(IsTopNode);
2007 } while (SU->isScheduled);
2009 if (SU->isTopReady())
2010 Top.removeReady(SU);
2011 if (SU->isBottomReady())
2012 Bot.removeReady(SU);
2014 DEBUG(dbgs() << "*** " << (IsTopNode ? "Top" : "Bottom")
2015 << " Scheduling Instruction in cycle "
2016 << (IsTopNode ? Top.CurrCycle : Bot.CurrCycle) << '\n';
2021 /// Update the scheduler's state after scheduling a node. This is the same node
2022 /// that was just returned by pickNode(). However, ScheduleDAGMI needs to update
2023 /// it's state based on the current cycle before MachineSchedStrategy does.
2024 void ConvergingScheduler::schedNode(SUnit *SU, bool IsTopNode) {
2026 SU->TopReadyCycle = Top.CurrCycle;
2030 SU->BotReadyCycle = Bot.CurrCycle;
2035 /// Create the standard converging machine scheduler. This will be used as the
2036 /// default scheduler if the target does not set a default.
2037 static ScheduleDAGInstrs *createConvergingSched(MachineSchedContext *C) {
2038 assert((!ForceTopDown || !ForceBottomUp) &&
2039 "-misched-topdown incompatible with -misched-bottomup");
2040 ScheduleDAGMI *DAG = new ScheduleDAGMI(C, new ConvergingScheduler());
2041 // Register DAG post-processors.
2042 if (EnableLoadCluster)
2043 DAG->addMutation(new LoadClusterMutation(DAG->TII, DAG->TRI));
2044 if (EnableMacroFusion)
2045 DAG->addMutation(new MacroFusion(DAG->TII));
2048 static MachineSchedRegistry
2049 ConvergingSchedRegistry("converge", "Standard converging scheduler.",
2050 createConvergingSched);
2052 //===----------------------------------------------------------------------===//
2053 // ILP Scheduler. Currently for experimental analysis of heuristics.
2054 //===----------------------------------------------------------------------===//
2057 /// \brief Order nodes by the ILP metric.
2059 SchedDFSResult *DFSResult;
2060 BitVector *ScheduledTrees;
2063 ILPOrder(SchedDFSResult *dfs, BitVector *schedtrees, bool MaxILP)
2064 : DFSResult(dfs), ScheduledTrees(schedtrees), MaximizeILP(MaxILP) {}
2066 /// \brief Apply a less-than relation on node priority.
2068 /// (Return true if A comes after B in the Q.)
2069 bool operator()(const SUnit *A, const SUnit *B) const {
2070 unsigned SchedTreeA = DFSResult->getSubtreeID(A);
2071 unsigned SchedTreeB = DFSResult->getSubtreeID(B);
2072 if (SchedTreeA != SchedTreeB) {
2073 // Unscheduled trees have lower priority.
2074 if (ScheduledTrees->test(SchedTreeA) != ScheduledTrees->test(SchedTreeB))
2075 return ScheduledTrees->test(SchedTreeB);
2077 // Trees with shallower connections have have lower priority.
2078 if (DFSResult->getSubtreeLevel(SchedTreeA)
2079 != DFSResult->getSubtreeLevel(SchedTreeB)) {
2080 return DFSResult->getSubtreeLevel(SchedTreeA)
2081 < DFSResult->getSubtreeLevel(SchedTreeB);
2085 return DFSResult->getILP(A) < DFSResult->getILP(B);
2087 return DFSResult->getILP(A) > DFSResult->getILP(B);
2091 /// \brief Schedule based on the ILP metric.
2092 class ILPScheduler : public MachineSchedStrategy {
2093 /// In case all subtrees are eventually connected to a common root through
2094 /// data dependence (e.g. reduction), place an upper limit on their size.
2096 /// FIXME: A subtree limit is generally good, but in the situation commented
2097 /// above, where multiple similar subtrees feed a common root, we should
2098 /// only split at a point where the resulting subtrees will be balanced.
2099 /// (a motivating test case must be found).
2100 static const unsigned SubtreeLimit = 16;
2102 SchedDFSResult DFSResult;
2103 BitVector ScheduledTrees;
2106 std::vector<SUnit*> ReadyQ;
2108 ILPScheduler(bool MaximizeILP)
2109 : DFSResult(/*BottomUp=*/true, SubtreeLimit),
2110 Cmp(&DFSResult, &ScheduledTrees, MaximizeILP) {}
2112 virtual void initialize(ScheduleDAGMI *DAG) {
2115 DFSResult.resize(DAG->SUnits.size());
2116 ScheduledTrees.clear();
2119 virtual void registerRoots() {
2120 DFSResult.compute(ReadyQ);
2121 ScheduledTrees.resize(DFSResult.getNumSubtrees());
2122 // Restore the heap in ReadyQ with the updated DFS results.
2123 std::make_heap(ReadyQ.begin(), ReadyQ.end(), Cmp);
2126 /// Implement MachineSchedStrategy interface.
2127 /// -----------------------------------------
2129 /// Callback to select the highest priority node from the ready Q.
2130 virtual SUnit *pickNode(bool &IsTopNode) {
2131 if (ReadyQ.empty()) return NULL;
2132 pop_heap(ReadyQ.begin(), ReadyQ.end(), Cmp);
2133 SUnit *SU = ReadyQ.back();
2136 DEBUG(dbgs() << "*** Scheduling " << "SU(" << SU->NodeNum << "): "
2138 << " ILP: " << DFSResult.getILP(SU)
2139 << " Tree: " << DFSResult.getSubtreeID(SU) << " @"
2140 << DFSResult.getSubtreeLevel(DFSResult.getSubtreeID(SU))<< '\n');
2144 /// Callback after a node is scheduled. Mark a newly scheduled tree, notify
2145 /// DFSResults, and resort the priority Q.
2146 virtual void schedNode(SUnit *SU, bool IsTopNode) {
2147 assert(!IsTopNode && "SchedDFSResult needs bottom-up");
2148 if (!ScheduledTrees.test(DFSResult.getSubtreeID(SU))) {
2149 ScheduledTrees.set(DFSResult.getSubtreeID(SU));
2150 DFSResult.scheduleTree(DFSResult.getSubtreeID(SU));
2151 std::make_heap(ReadyQ.begin(), ReadyQ.end(), Cmp);
2155 virtual void releaseTopNode(SUnit *) { /*only called for top roots*/ }
2157 virtual void releaseBottomNode(SUnit *SU) {
2158 ReadyQ.push_back(SU);
2159 std::push_heap(ReadyQ.begin(), ReadyQ.end(), Cmp);
2164 static ScheduleDAGInstrs *createILPMaxScheduler(MachineSchedContext *C) {
2165 return new ScheduleDAGMI(C, new ILPScheduler(true));
2167 static ScheduleDAGInstrs *createILPMinScheduler(MachineSchedContext *C) {
2168 return new ScheduleDAGMI(C, new ILPScheduler(false));
2170 static MachineSchedRegistry ILPMaxRegistry(
2171 "ilpmax", "Schedule bottom-up for max ILP", createILPMaxScheduler);
2172 static MachineSchedRegistry ILPMinRegistry(
2173 "ilpmin", "Schedule bottom-up for min ILP", createILPMinScheduler);
2175 //===----------------------------------------------------------------------===//
2176 // Machine Instruction Shuffler for Correctness Testing
2177 //===----------------------------------------------------------------------===//
2181 /// Apply a less-than relation on the node order, which corresponds to the
2182 /// instruction order prior to scheduling. IsReverse implements greater-than.
2183 template<bool IsReverse>
2185 bool operator()(SUnit *A, SUnit *B) const {
2187 return A->NodeNum > B->NodeNum;
2189 return A->NodeNum < B->NodeNum;
2193 /// Reorder instructions as much as possible.
2194 class InstructionShuffler : public MachineSchedStrategy {
2198 // Using a less-than relation (SUnitOrder<false>) for the TopQ priority
2199 // gives nodes with a higher number higher priority causing the latest
2200 // instructions to be scheduled first.
2201 PriorityQueue<SUnit*, std::vector<SUnit*>, SUnitOrder<false> >
2203 // When scheduling bottom-up, use greater-than as the queue priority.
2204 PriorityQueue<SUnit*, std::vector<SUnit*>, SUnitOrder<true> >
2207 InstructionShuffler(bool alternate, bool topdown)
2208 : IsAlternating(alternate), IsTopDown(topdown) {}
2210 virtual void initialize(ScheduleDAGMI *) {
2215 /// Implement MachineSchedStrategy interface.
2216 /// -----------------------------------------
2218 virtual SUnit *pickNode(bool &IsTopNode) {
2222 if (TopQ.empty()) return NULL;
2225 } while (SU->isScheduled);
2230 if (BottomQ.empty()) return NULL;
2233 } while (SU->isScheduled);
2237 IsTopDown = !IsTopDown;
2241 virtual void schedNode(SUnit *SU, bool IsTopNode) {}
2243 virtual void releaseTopNode(SUnit *SU) {
2246 virtual void releaseBottomNode(SUnit *SU) {
2252 static ScheduleDAGInstrs *createInstructionShuffler(MachineSchedContext *C) {
2253 bool Alternate = !ForceTopDown && !ForceBottomUp;
2254 bool TopDown = !ForceBottomUp;
2255 assert((TopDown || !ForceTopDown) &&
2256 "-misched-topdown incompatible with -misched-bottomup");
2257 return new ScheduleDAGMI(C, new InstructionShuffler(Alternate, TopDown));
2259 static MachineSchedRegistry ShufflerRegistry(
2260 "shuffle", "Shuffle machine instructions alternating directions",
2261 createInstructionShuffler);