1 //===- MachineScheduler.cpp - Machine Instruction Scheduler ---------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // MachineScheduler schedules machine instructions after phi elimination. It
11 // preserves LiveIntervals so it can be invoked before register allocation.
13 //===----------------------------------------------------------------------===//
15 #define DEBUG_TYPE "misched"
17 #include "llvm/CodeGen/MachineScheduler.h"
18 #include "llvm/ADT/OwningPtr.h"
19 #include "llvm/ADT/PriorityQueue.h"
20 #include "llvm/Analysis/AliasAnalysis.h"
21 #include "llvm/CodeGen/LiveIntervalAnalysis.h"
22 #include "llvm/CodeGen/MachineDominators.h"
23 #include "llvm/CodeGen/MachineLoopInfo.h"
24 #include "llvm/CodeGen/Passes.h"
25 #include "llvm/CodeGen/RegisterClassInfo.h"
26 #include "llvm/CodeGen/ScheduleDFS.h"
27 #include "llvm/CodeGen/ScheduleHazardRecognizer.h"
28 #include "llvm/Support/CommandLine.h"
29 #include "llvm/Support/Debug.h"
30 #include "llvm/Support/ErrorHandling.h"
31 #include "llvm/Support/GraphWriter.h"
32 #include "llvm/Support/raw_ostream.h"
38 cl::opt<bool> ForceTopDown("misched-topdown", cl::Hidden,
39 cl::desc("Force top-down list scheduling"));
40 cl::opt<bool> ForceBottomUp("misched-bottomup", cl::Hidden,
41 cl::desc("Force bottom-up list scheduling"));
45 static cl::opt<bool> ViewMISchedDAGs("view-misched-dags", cl::Hidden,
46 cl::desc("Pop up a window to show MISched dags after they are processed"));
48 static cl::opt<unsigned> MISchedCutoff("misched-cutoff", cl::Hidden,
49 cl::desc("Stop scheduling after N instructions"), cl::init(~0U));
51 static bool ViewMISchedDAGs = false;
54 // Experimental heuristics
55 static cl::opt<bool> EnableLoadCluster("misched-cluster", cl::Hidden,
56 cl::desc("Enable load clustering."), cl::init(true));
58 // Experimental heuristics
59 static cl::opt<bool> EnableMacroFusion("misched-fusion", cl::Hidden,
60 cl::desc("Enable scheduling for macro fusion."), cl::init(true));
62 static cl::opt<bool> VerifyScheduling("verify-misched", cl::Hidden,
63 cl::desc("Verify machine instrs before and after machine scheduling"));
65 // DAG subtrees must have at least this many nodes.
66 static const unsigned MinSubtreeSize = 8;
68 //===----------------------------------------------------------------------===//
69 // Machine Instruction Scheduling Pass and Registry
70 //===----------------------------------------------------------------------===//
72 MachineSchedContext::MachineSchedContext():
73 MF(0), MLI(0), MDT(0), PassConfig(0), AA(0), LIS(0) {
74 RegClassInfo = new RegisterClassInfo();
77 MachineSchedContext::~MachineSchedContext() {
82 /// MachineScheduler runs after coalescing and before register allocation.
83 class MachineScheduler : public MachineSchedContext,
84 public MachineFunctionPass {
88 virtual void getAnalysisUsage(AnalysisUsage &AU) const;
90 virtual void releaseMemory() {}
92 virtual bool runOnMachineFunction(MachineFunction&);
94 virtual void print(raw_ostream &O, const Module* = 0) const;
96 static char ID; // Class identification, replacement for typeinfo
100 char MachineScheduler::ID = 0;
102 char &llvm::MachineSchedulerID = MachineScheduler::ID;
104 INITIALIZE_PASS_BEGIN(MachineScheduler, "misched",
105 "Machine Instruction Scheduler", false, false)
106 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
107 INITIALIZE_PASS_DEPENDENCY(SlotIndexes)
108 INITIALIZE_PASS_DEPENDENCY(LiveIntervals)
109 INITIALIZE_PASS_END(MachineScheduler, "misched",
110 "Machine Instruction Scheduler", false, false)
112 MachineScheduler::MachineScheduler()
113 : MachineFunctionPass(ID) {
114 initializeMachineSchedulerPass(*PassRegistry::getPassRegistry());
117 void MachineScheduler::getAnalysisUsage(AnalysisUsage &AU) const {
118 AU.setPreservesCFG();
119 AU.addRequiredID(MachineDominatorsID);
120 AU.addRequired<MachineLoopInfo>();
121 AU.addRequired<AliasAnalysis>();
122 AU.addRequired<TargetPassConfig>();
123 AU.addRequired<SlotIndexes>();
124 AU.addPreserved<SlotIndexes>();
125 AU.addRequired<LiveIntervals>();
126 AU.addPreserved<LiveIntervals>();
127 MachineFunctionPass::getAnalysisUsage(AU);
130 MachinePassRegistry MachineSchedRegistry::Registry;
132 /// A dummy default scheduler factory indicates whether the scheduler
133 /// is overridden on the command line.
134 static ScheduleDAGInstrs *useDefaultMachineSched(MachineSchedContext *C) {
138 /// MachineSchedOpt allows command line selection of the scheduler.
139 static cl::opt<MachineSchedRegistry::ScheduleDAGCtor, false,
140 RegisterPassParser<MachineSchedRegistry> >
141 MachineSchedOpt("misched",
142 cl::init(&useDefaultMachineSched), cl::Hidden,
143 cl::desc("Machine instruction scheduler to use"));
145 static MachineSchedRegistry
146 DefaultSchedRegistry("default", "Use the target's default scheduler choice.",
147 useDefaultMachineSched);
149 /// Forward declare the standard machine scheduler. This will be used as the
150 /// default scheduler if the target does not set a default.
151 static ScheduleDAGInstrs *createConvergingSched(MachineSchedContext *C);
154 /// Decrement this iterator until reaching the top or a non-debug instr.
155 static MachineBasicBlock::iterator
156 priorNonDebug(MachineBasicBlock::iterator I, MachineBasicBlock::iterator Beg) {
157 assert(I != Beg && "reached the top of the region, cannot decrement");
159 if (!I->isDebugValue())
165 /// If this iterator is a debug value, increment until reaching the End or a
166 /// non-debug instruction.
167 static MachineBasicBlock::iterator
168 nextIfDebug(MachineBasicBlock::iterator I, MachineBasicBlock::iterator End) {
169 for(; I != End; ++I) {
170 if (!I->isDebugValue())
176 /// Top-level MachineScheduler pass driver.
178 /// Visit blocks in function order. Divide each block into scheduling regions
179 /// and visit them bottom-up. Visiting regions bottom-up is not required, but is
180 /// consistent with the DAG builder, which traverses the interior of the
181 /// scheduling regions bottom-up.
183 /// This design avoids exposing scheduling boundaries to the DAG builder,
184 /// simplifying the DAG builder's support for "special" target instructions.
185 /// At the same time the design allows target schedulers to operate across
186 /// scheduling boundaries, for example to bundle the boudary instructions
187 /// without reordering them. This creates complexity, because the target
188 /// scheduler must update the RegionBegin and RegionEnd positions cached by
189 /// ScheduleDAGInstrs whenever adding or removing instructions. A much simpler
190 /// design would be to split blocks at scheduling boundaries, but LLVM has a
191 /// general bias against block splitting purely for implementation simplicity.
192 bool MachineScheduler::runOnMachineFunction(MachineFunction &mf) {
193 DEBUG(dbgs() << "Before MISsched:\n"; mf.print(dbgs()));
195 // Initialize the context of the pass.
197 MLI = &getAnalysis<MachineLoopInfo>();
198 MDT = &getAnalysis<MachineDominatorTree>();
199 PassConfig = &getAnalysis<TargetPassConfig>();
200 AA = &getAnalysis<AliasAnalysis>();
202 LIS = &getAnalysis<LiveIntervals>();
203 const TargetInstrInfo *TII = MF->getTarget().getInstrInfo();
205 if (VerifyScheduling) {
206 DEBUG(LIS->print(dbgs()));
207 MF->verify(this, "Before machine scheduling.");
209 RegClassInfo->runOnMachineFunction(*MF);
211 // Select the scheduler, or set the default.
212 MachineSchedRegistry::ScheduleDAGCtor Ctor = MachineSchedOpt;
213 if (Ctor == useDefaultMachineSched) {
214 // Get the default scheduler set by the target.
215 Ctor = MachineSchedRegistry::getDefault();
217 Ctor = createConvergingSched;
218 MachineSchedRegistry::setDefault(Ctor);
221 // Instantiate the selected scheduler.
222 OwningPtr<ScheduleDAGInstrs> Scheduler(Ctor(this));
224 // Visit all machine basic blocks.
226 // TODO: Visit blocks in global postorder or postorder within the bottom-up
227 // loop tree. Then we can optionally compute global RegPressure.
228 for (MachineFunction::iterator MBB = MF->begin(), MBBEnd = MF->end();
229 MBB != MBBEnd; ++MBB) {
231 Scheduler->startBlock(MBB);
233 // Break the block into scheduling regions [I, RegionEnd), and schedule each
234 // region as soon as it is discovered. RegionEnd points the scheduling
235 // boundary at the bottom of the region. The DAG does not include RegionEnd,
236 // but the region does (i.e. the next RegionEnd is above the previous
237 // RegionBegin). If the current block has no terminator then RegionEnd ==
238 // MBB->end() for the bottom region.
240 // The Scheduler may insert instructions during either schedule() or
241 // exitRegion(), even for empty regions. So the local iterators 'I' and
242 // 'RegionEnd' are invalid across these calls.
243 unsigned RemainingInstrs = MBB->size();
244 for(MachineBasicBlock::iterator RegionEnd = MBB->end();
245 RegionEnd != MBB->begin(); RegionEnd = Scheduler->begin()) {
247 // Avoid decrementing RegionEnd for blocks with no terminator.
248 if (RegionEnd != MBB->end()
249 || TII->isSchedulingBoundary(llvm::prior(RegionEnd), MBB, *MF)) {
251 // Count the boundary instruction.
255 // The next region starts above the previous region. Look backward in the
256 // instruction stream until we find the nearest boundary.
257 MachineBasicBlock::iterator I = RegionEnd;
258 for(;I != MBB->begin(); --I, --RemainingInstrs) {
259 if (TII->isSchedulingBoundary(llvm::prior(I), MBB, *MF))
262 // Notify the scheduler of the region, even if we may skip scheduling
263 // it. Perhaps it still needs to be bundled.
264 Scheduler->enterRegion(MBB, I, RegionEnd, RemainingInstrs);
266 // Skip empty scheduling regions (0 or 1 schedulable instructions).
267 if (I == RegionEnd || I == llvm::prior(RegionEnd)) {
268 // Close the current region. Bundle the terminator if needed.
269 // This invalidates 'RegionEnd' and 'I'.
270 Scheduler->exitRegion();
273 DEBUG(dbgs() << "********** MI Scheduling **********\n");
274 DEBUG(dbgs() << MF->getName()
275 << ":BB#" << MBB->getNumber() << " " << MBB->getName()
276 << "\n From: " << *I << " To: ";
277 if (RegionEnd != MBB->end()) dbgs() << *RegionEnd;
278 else dbgs() << "End";
279 dbgs() << " Remaining: " << RemainingInstrs << "\n");
281 // Schedule a region: possibly reorder instructions.
282 // This invalidates 'RegionEnd' and 'I'.
283 Scheduler->schedule();
285 // Close the current region.
286 Scheduler->exitRegion();
288 // Scheduling has invalidated the current iterator 'I'. Ask the
289 // scheduler for the top of it's scheduled region.
290 RegionEnd = Scheduler->begin();
292 assert(RemainingInstrs == 0 && "Instruction count mismatch!");
293 Scheduler->finishBlock();
295 Scheduler->finalizeSchedule();
296 DEBUG(LIS->print(dbgs()));
297 if (VerifyScheduling)
298 MF->verify(this, "After machine scheduling.");
302 void MachineScheduler::print(raw_ostream &O, const Module* m) const {
306 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
307 void ReadyQueue::dump() {
308 dbgs() << Name << ": ";
309 for (unsigned i = 0, e = Queue.size(); i < e; ++i)
310 dbgs() << Queue[i]->NodeNum << " ";
315 //===----------------------------------------------------------------------===//
316 // ScheduleDAGMI - Base class for MachineInstr scheduling with LiveIntervals
318 //===----------------------------------------------------------------------===//
320 ScheduleDAGMI::~ScheduleDAGMI() {
322 DeleteContainerPointers(Mutations);
326 bool ScheduleDAGMI::addEdge(SUnit *SuccSU, const SDep &PredDep) {
327 if (SuccSU != &ExitSU) {
328 // Do not use WillCreateCycle, it assumes SD scheduling.
329 // If Pred is reachable from Succ, then the edge creates a cycle.
330 if (Topo.IsReachable(PredDep.getSUnit(), SuccSU))
332 Topo.AddPred(SuccSU, PredDep.getSUnit());
334 SuccSU->addPred(PredDep, /*Required=*/!PredDep.isArtificial());
335 // Return true regardless of whether a new edge needed to be inserted.
339 /// ReleaseSucc - Decrement the NumPredsLeft count of a successor. When
340 /// NumPredsLeft reaches zero, release the successor node.
342 /// FIXME: Adjust SuccSU height based on MinLatency.
343 void ScheduleDAGMI::releaseSucc(SUnit *SU, SDep *SuccEdge) {
344 SUnit *SuccSU = SuccEdge->getSUnit();
346 if (SuccEdge->isWeak()) {
347 --SuccSU->WeakPredsLeft;
348 if (SuccEdge->isCluster())
349 NextClusterSucc = SuccSU;
353 if (SuccSU->NumPredsLeft == 0) {
354 dbgs() << "*** Scheduling failed! ***\n";
356 dbgs() << " has been released too many times!\n";
360 --SuccSU->NumPredsLeft;
361 if (SuccSU->NumPredsLeft == 0 && SuccSU != &ExitSU)
362 SchedImpl->releaseTopNode(SuccSU);
365 /// releaseSuccessors - Call releaseSucc on each of SU's successors.
366 void ScheduleDAGMI::releaseSuccessors(SUnit *SU) {
367 for (SUnit::succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
369 releaseSucc(SU, &*I);
373 /// ReleasePred - Decrement the NumSuccsLeft count of a predecessor. When
374 /// NumSuccsLeft reaches zero, release the predecessor node.
376 /// FIXME: Adjust PredSU height based on MinLatency.
377 void ScheduleDAGMI::releasePred(SUnit *SU, SDep *PredEdge) {
378 SUnit *PredSU = PredEdge->getSUnit();
380 if (PredEdge->isWeak()) {
381 --PredSU->WeakSuccsLeft;
382 if (PredEdge->isCluster())
383 NextClusterPred = PredSU;
387 if (PredSU->NumSuccsLeft == 0) {
388 dbgs() << "*** Scheduling failed! ***\n";
390 dbgs() << " has been released too many times!\n";
394 --PredSU->NumSuccsLeft;
395 if (PredSU->NumSuccsLeft == 0 && PredSU != &EntrySU)
396 SchedImpl->releaseBottomNode(PredSU);
399 /// releasePredecessors - Call releasePred on each of SU's predecessors.
400 void ScheduleDAGMI::releasePredecessors(SUnit *SU) {
401 for (SUnit::pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
403 releasePred(SU, &*I);
407 void ScheduleDAGMI::moveInstruction(MachineInstr *MI,
408 MachineBasicBlock::iterator InsertPos) {
409 // Advance RegionBegin if the first instruction moves down.
410 if (&*RegionBegin == MI)
413 // Update the instruction stream.
414 BB->splice(InsertPos, BB, MI);
416 // Update LiveIntervals
417 LIS->handleMove(MI, /*UpdateFlags=*/true);
419 // Recede RegionBegin if an instruction moves above the first.
420 if (RegionBegin == InsertPos)
424 bool ScheduleDAGMI::checkSchedLimit() {
426 if (NumInstrsScheduled == MISchedCutoff && MISchedCutoff != ~0U) {
427 CurrentTop = CurrentBottom;
430 ++NumInstrsScheduled;
435 /// enterRegion - Called back from MachineScheduler::runOnMachineFunction after
436 /// crossing a scheduling boundary. [begin, end) includes all instructions in
437 /// the region, including the boundary itself and single-instruction regions
438 /// that don't get scheduled.
439 void ScheduleDAGMI::enterRegion(MachineBasicBlock *bb,
440 MachineBasicBlock::iterator begin,
441 MachineBasicBlock::iterator end,
444 ScheduleDAGInstrs::enterRegion(bb, begin, end, endcount);
446 // For convenience remember the end of the liveness region.
448 (RegionEnd == bb->end()) ? RegionEnd : llvm::next(RegionEnd);
451 // Setup the register pressure trackers for the top scheduled top and bottom
452 // scheduled regions.
453 void ScheduleDAGMI::initRegPressure() {
454 TopRPTracker.init(&MF, RegClassInfo, LIS, BB, RegionBegin);
455 BotRPTracker.init(&MF, RegClassInfo, LIS, BB, LiveRegionEnd);
457 // Close the RPTracker to finalize live ins.
458 RPTracker.closeRegion();
460 DEBUG(RPTracker.getPressure().dump(TRI));
462 // Initialize the live ins and live outs.
463 TopRPTracker.addLiveRegs(RPTracker.getPressure().LiveInRegs);
464 BotRPTracker.addLiveRegs(RPTracker.getPressure().LiveOutRegs);
466 // Close one end of the tracker so we can call
467 // getMaxUpward/DownwardPressureDelta before advancing across any
468 // instructions. This converts currently live regs into live ins/outs.
469 TopRPTracker.closeTop();
470 BotRPTracker.closeBottom();
472 // Account for liveness generated by the region boundary.
473 if (LiveRegionEnd != RegionEnd)
474 BotRPTracker.recede();
476 assert(BotRPTracker.getPos() == RegionEnd && "Can't find the region bottom");
478 // Cache the list of excess pressure sets in this region. This will also track
479 // the max pressure in the scheduled code for these sets.
480 RegionCriticalPSets.clear();
481 const std::vector<unsigned> &RegionPressure =
482 RPTracker.getPressure().MaxSetPressure;
483 for (unsigned i = 0, e = RegionPressure.size(); i < e; ++i) {
484 unsigned Limit = TRI->getRegPressureSetLimit(i);
485 DEBUG(dbgs() << TRI->getRegPressureSetName(i)
487 << " Actual " << RegionPressure[i] << "\n");
488 if (RegionPressure[i] > Limit)
489 RegionCriticalPSets.push_back(PressureElement(i, 0));
491 DEBUG(dbgs() << "Excess PSets: ";
492 for (unsigned i = 0, e = RegionCriticalPSets.size(); i != e; ++i)
493 dbgs() << TRI->getRegPressureSetName(
494 RegionCriticalPSets[i].PSetID) << " ";
498 // FIXME: When the pressure tracker deals in pressure differences then we won't
499 // iterate over all RegionCriticalPSets[i].
501 updateScheduledPressure(const std::vector<unsigned> &NewMaxPressure) {
502 for (unsigned i = 0, e = RegionCriticalPSets.size(); i < e; ++i) {
503 unsigned ID = RegionCriticalPSets[i].PSetID;
504 int &MaxUnits = RegionCriticalPSets[i].UnitIncrease;
505 if ((int)NewMaxPressure[ID] > MaxUnits)
506 MaxUnits = NewMaxPressure[ID];
510 /// schedule - Called back from MachineScheduler::runOnMachineFunction
511 /// after setting up the current scheduling region. [RegionBegin, RegionEnd)
512 /// only includes instructions that have DAG nodes, not scheduling boundaries.
514 /// This is a skeletal driver, with all the functionality pushed into helpers,
515 /// so that it can be easilly extended by experimental schedulers. Generally,
516 /// implementing MachineSchedStrategy should be sufficient to implement a new
517 /// scheduling algorithm. However, if a scheduler further subclasses
518 /// ScheduleDAGMI then it will want to override this virtual method in order to
519 /// update any specialized state.
520 void ScheduleDAGMI::schedule() {
521 buildDAGWithRegPressure();
523 Topo.InitDAGTopologicalSorting();
527 SmallVector<SUnit*, 8> TopRoots, BotRoots;
528 findRootsAndBiasEdges(TopRoots, BotRoots);
530 // Initialize the strategy before modifying the DAG.
531 // This may initialize a DFSResult to be used for queue priority.
532 SchedImpl->initialize(this);
534 DEBUG(for (unsigned su = 0, e = SUnits.size(); su != e; ++su)
535 SUnits[su].dumpAll(this));
536 if (ViewMISchedDAGs) viewGraph();
538 // Initialize ready queues now that the DAG and priority data are finalized.
539 initQueues(TopRoots, BotRoots);
541 bool IsTopNode = false;
542 while (SUnit *SU = SchedImpl->pickNode(IsTopNode)) {
543 assert(!SU->isScheduled && "Node already scheduled");
544 if (!checkSchedLimit())
547 scheduleMI(SU, IsTopNode);
549 updateQueues(SU, IsTopNode);
551 assert(CurrentTop == CurrentBottom && "Nonempty unscheduled zone.");
556 unsigned BBNum = begin()->getParent()->getNumber();
557 dbgs() << "*** Final schedule for BB#" << BBNum << " ***\n";
563 /// Build the DAG and setup three register pressure trackers.
564 void ScheduleDAGMI::buildDAGWithRegPressure() {
565 // Initialize the register pressure tracker used by buildSchedGraph.
566 RPTracker.init(&MF, RegClassInfo, LIS, BB, LiveRegionEnd);
568 // Account for liveness generate by the region boundary.
569 if (LiveRegionEnd != RegionEnd)
572 // Build the DAG, and compute current register pressure.
573 buildSchedGraph(AA, &RPTracker);
575 // Initialize top/bottom trackers after computing region pressure.
579 /// Apply each ScheduleDAGMutation step in order.
580 void ScheduleDAGMI::postprocessDAG() {
581 for (unsigned i = 0, e = Mutations.size(); i < e; ++i) {
582 Mutations[i]->apply(this);
586 void ScheduleDAGMI::computeDFSResult() {
588 DFSResult = new SchedDFSResult(/*BottomU*/true, MinSubtreeSize);
590 ScheduledTrees.clear();
591 DFSResult->resize(SUnits.size());
592 DFSResult->compute(SUnits);
593 ScheduledTrees.resize(DFSResult->getNumSubtrees());
596 void ScheduleDAGMI::findRootsAndBiasEdges(SmallVectorImpl<SUnit*> &TopRoots,
597 SmallVectorImpl<SUnit*> &BotRoots) {
598 for (std::vector<SUnit>::iterator
599 I = SUnits.begin(), E = SUnits.end(); I != E; ++I) {
601 assert(!SU->isBoundaryNode() && "Boundary node should not be in SUnits");
603 // Order predecessors so DFSResult follows the critical path.
604 SU->biasCriticalPath();
606 // A SUnit is ready to top schedule if it has no predecessors.
607 if (!I->NumPredsLeft)
608 TopRoots.push_back(SU);
609 // A SUnit is ready to bottom schedule if it has no successors.
610 if (!I->NumSuccsLeft)
611 BotRoots.push_back(SU);
613 ExitSU.biasCriticalPath();
616 /// Identify DAG roots and setup scheduler queues.
617 void ScheduleDAGMI::initQueues(ArrayRef<SUnit*> TopRoots,
618 ArrayRef<SUnit*> BotRoots) {
619 NextClusterSucc = NULL;
620 NextClusterPred = NULL;
622 // Release all DAG roots for scheduling, not including EntrySU/ExitSU.
624 // Nodes with unreleased weak edges can still be roots.
625 // Release top roots in forward order.
626 for (SmallVectorImpl<SUnit*>::const_iterator
627 I = TopRoots.begin(), E = TopRoots.end(); I != E; ++I) {
628 SchedImpl->releaseTopNode(*I);
630 // Release bottom roots in reverse order so the higher priority nodes appear
631 // first. This is more natural and slightly more efficient.
632 for (SmallVectorImpl<SUnit*>::const_reverse_iterator
633 I = BotRoots.rbegin(), E = BotRoots.rend(); I != E; ++I) {
634 SchedImpl->releaseBottomNode(*I);
637 releaseSuccessors(&EntrySU);
638 releasePredecessors(&ExitSU);
640 SchedImpl->registerRoots();
642 // Advance past initial DebugValues.
643 assert(TopRPTracker.getPos() == RegionBegin && "bad initial Top tracker");
644 CurrentTop = nextIfDebug(RegionBegin, RegionEnd);
645 TopRPTracker.setPos(CurrentTop);
647 CurrentBottom = RegionEnd;
650 /// Move an instruction and update register pressure.
651 void ScheduleDAGMI::scheduleMI(SUnit *SU, bool IsTopNode) {
652 // Move the instruction to its new location in the instruction stream.
653 MachineInstr *MI = SU->getInstr();
656 assert(SU->isTopReady() && "node still has unscheduled dependencies");
657 if (&*CurrentTop == MI)
658 CurrentTop = nextIfDebug(++CurrentTop, CurrentBottom);
660 moveInstruction(MI, CurrentTop);
661 TopRPTracker.setPos(MI);
664 // Update top scheduled pressure.
665 TopRPTracker.advance();
666 assert(TopRPTracker.getPos() == CurrentTop && "out of sync");
667 updateScheduledPressure(TopRPTracker.getPressure().MaxSetPressure);
670 assert(SU->isBottomReady() && "node still has unscheduled dependencies");
671 MachineBasicBlock::iterator priorII =
672 priorNonDebug(CurrentBottom, CurrentTop);
674 CurrentBottom = priorII;
676 if (&*CurrentTop == MI) {
677 CurrentTop = nextIfDebug(++CurrentTop, priorII);
678 TopRPTracker.setPos(CurrentTop);
680 moveInstruction(MI, CurrentBottom);
683 // Update bottom scheduled pressure.
684 BotRPTracker.recede();
685 assert(BotRPTracker.getPos() == CurrentBottom && "out of sync");
686 updateScheduledPressure(BotRPTracker.getPressure().MaxSetPressure);
690 /// Update scheduler queues after scheduling an instruction.
691 void ScheduleDAGMI::updateQueues(SUnit *SU, bool IsTopNode) {
692 // Release dependent instructions for scheduling.
694 releaseSuccessors(SU);
696 releasePredecessors(SU);
698 SU->isScheduled = true;
701 unsigned SubtreeID = DFSResult->getSubtreeID(SU);
702 if (!ScheduledTrees.test(SubtreeID)) {
703 ScheduledTrees.set(SubtreeID);
704 DFSResult->scheduleTree(SubtreeID);
705 SchedImpl->scheduleTree(SubtreeID);
709 // Notify the scheduling strategy after updating the DAG.
710 SchedImpl->schedNode(SU, IsTopNode);
713 /// Reinsert any remaining debug_values, just like the PostRA scheduler.
714 void ScheduleDAGMI::placeDebugValues() {
715 // If first instruction was a DBG_VALUE then put it back.
717 BB->splice(RegionBegin, BB, FirstDbgValue);
718 RegionBegin = FirstDbgValue;
721 for (std::vector<std::pair<MachineInstr *, MachineInstr *> >::iterator
722 DI = DbgValues.end(), DE = DbgValues.begin(); DI != DE; --DI) {
723 std::pair<MachineInstr *, MachineInstr *> P = *prior(DI);
724 MachineInstr *DbgValue = P.first;
725 MachineBasicBlock::iterator OrigPrevMI = P.second;
726 if (&*RegionBegin == DbgValue)
728 BB->splice(++OrigPrevMI, BB, DbgValue);
729 if (OrigPrevMI == llvm::prior(RegionEnd))
730 RegionEnd = DbgValue;
733 FirstDbgValue = NULL;
736 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
737 void ScheduleDAGMI::dumpSchedule() const {
738 for (MachineBasicBlock::iterator MI = begin(), ME = end(); MI != ME; ++MI) {
739 if (SUnit *SU = getSUnit(&(*MI)))
742 dbgs() << "Missing SUnit\n";
747 //===----------------------------------------------------------------------===//
748 // LoadClusterMutation - DAG post-processing to cluster loads.
749 //===----------------------------------------------------------------------===//
752 /// \brief Post-process the DAG to create cluster edges between neighboring
754 class LoadClusterMutation : public ScheduleDAGMutation {
759 LoadInfo(SUnit *su, unsigned reg, unsigned ofs)
760 : SU(su), BaseReg(reg), Offset(ofs) {}
762 static bool LoadInfoLess(const LoadClusterMutation::LoadInfo &LHS,
763 const LoadClusterMutation::LoadInfo &RHS);
765 const TargetInstrInfo *TII;
766 const TargetRegisterInfo *TRI;
768 LoadClusterMutation(const TargetInstrInfo *tii,
769 const TargetRegisterInfo *tri)
770 : TII(tii), TRI(tri) {}
772 virtual void apply(ScheduleDAGMI *DAG);
774 void clusterNeighboringLoads(ArrayRef<SUnit*> Loads, ScheduleDAGMI *DAG);
778 bool LoadClusterMutation::LoadInfoLess(
779 const LoadClusterMutation::LoadInfo &LHS,
780 const LoadClusterMutation::LoadInfo &RHS) {
781 if (LHS.BaseReg != RHS.BaseReg)
782 return LHS.BaseReg < RHS.BaseReg;
783 return LHS.Offset < RHS.Offset;
786 void LoadClusterMutation::clusterNeighboringLoads(ArrayRef<SUnit*> Loads,
787 ScheduleDAGMI *DAG) {
788 SmallVector<LoadClusterMutation::LoadInfo,32> LoadRecords;
789 for (unsigned Idx = 0, End = Loads.size(); Idx != End; ++Idx) {
790 SUnit *SU = Loads[Idx];
793 if (TII->getLdStBaseRegImmOfs(SU->getInstr(), BaseReg, Offset, TRI))
794 LoadRecords.push_back(LoadInfo(SU, BaseReg, Offset));
796 if (LoadRecords.size() < 2)
798 std::sort(LoadRecords.begin(), LoadRecords.end(), LoadInfoLess);
799 unsigned ClusterLength = 1;
800 for (unsigned Idx = 0, End = LoadRecords.size(); Idx < (End - 1); ++Idx) {
801 if (LoadRecords[Idx].BaseReg != LoadRecords[Idx+1].BaseReg) {
806 SUnit *SUa = LoadRecords[Idx].SU;
807 SUnit *SUb = LoadRecords[Idx+1].SU;
808 if (TII->shouldClusterLoads(SUa->getInstr(), SUb->getInstr(), ClusterLength)
809 && DAG->addEdge(SUb, SDep(SUa, SDep::Cluster))) {
811 DEBUG(dbgs() << "Cluster loads SU(" << SUa->NodeNum << ") - SU("
812 << SUb->NodeNum << ")\n");
813 // Copy successor edges from SUa to SUb. Interleaving computation
814 // dependent on SUa can prevent load combining due to register reuse.
815 // Predecessor edges do not need to be copied from SUb to SUa since nearby
816 // loads should have effectively the same inputs.
817 for (SUnit::const_succ_iterator
818 SI = SUa->Succs.begin(), SE = SUa->Succs.end(); SI != SE; ++SI) {
819 if (SI->getSUnit() == SUb)
821 DEBUG(dbgs() << " Copy Succ SU(" << SI->getSUnit()->NodeNum << ")\n");
822 DAG->addEdge(SI->getSUnit(), SDep(SUb, SDep::Artificial));
831 /// \brief Callback from DAG postProcessing to create cluster edges for loads.
832 void LoadClusterMutation::apply(ScheduleDAGMI *DAG) {
833 // Map DAG NodeNum to store chain ID.
834 DenseMap<unsigned, unsigned> StoreChainIDs;
835 // Map each store chain to a set of dependent loads.
836 SmallVector<SmallVector<SUnit*,4>, 32> StoreChainDependents;
837 for (unsigned Idx = 0, End = DAG->SUnits.size(); Idx != End; ++Idx) {
838 SUnit *SU = &DAG->SUnits[Idx];
839 if (!SU->getInstr()->mayLoad())
841 unsigned ChainPredID = DAG->SUnits.size();
842 for (SUnit::const_pred_iterator
843 PI = SU->Preds.begin(), PE = SU->Preds.end(); PI != PE; ++PI) {
845 ChainPredID = PI->getSUnit()->NodeNum;
849 // Check if this chain-like pred has been seen
850 // before. ChainPredID==MaxNodeID for loads at the top of the schedule.
851 unsigned NumChains = StoreChainDependents.size();
852 std::pair<DenseMap<unsigned, unsigned>::iterator, bool> Result =
853 StoreChainIDs.insert(std::make_pair(ChainPredID, NumChains));
855 StoreChainDependents.resize(NumChains + 1);
856 StoreChainDependents[Result.first->second].push_back(SU);
858 // Iterate over the store chains.
859 for (unsigned Idx = 0, End = StoreChainDependents.size(); Idx != End; ++Idx)
860 clusterNeighboringLoads(StoreChainDependents[Idx], DAG);
863 //===----------------------------------------------------------------------===//
864 // MacroFusion - DAG post-processing to encourage fusion of macro ops.
865 //===----------------------------------------------------------------------===//
868 /// \brief Post-process the DAG to create cluster edges between instructions
869 /// that may be fused by the processor into a single operation.
870 class MacroFusion : public ScheduleDAGMutation {
871 const TargetInstrInfo *TII;
873 MacroFusion(const TargetInstrInfo *tii): TII(tii) {}
875 virtual void apply(ScheduleDAGMI *DAG);
879 /// \brief Callback from DAG postProcessing to create cluster edges to encourage
880 /// fused operations.
881 void MacroFusion::apply(ScheduleDAGMI *DAG) {
882 // For now, assume targets can only fuse with the branch.
883 MachineInstr *Branch = DAG->ExitSU.getInstr();
887 for (unsigned Idx = DAG->SUnits.size(); Idx > 0;) {
888 SUnit *SU = &DAG->SUnits[--Idx];
889 if (!TII->shouldScheduleAdjacent(SU->getInstr(), Branch))
892 // Create a single weak edge from SU to ExitSU. The only effect is to cause
893 // bottom-up scheduling to heavily prioritize the clustered SU. There is no
894 // need to copy predecessor edges from ExitSU to SU, since top-down
895 // scheduling cannot prioritize ExitSU anyway. To defer top-down scheduling
896 // of SU, we could create an artificial edge from the deepest root, but it
897 // hasn't been needed yet.
898 bool Success = DAG->addEdge(&DAG->ExitSU, SDep(SU, SDep::Cluster));
900 assert(Success && "No DAG nodes should be reachable from ExitSU");
902 DEBUG(dbgs() << "Macro Fuse SU(" << SU->NodeNum << ")\n");
907 //===----------------------------------------------------------------------===//
908 // ConvergingScheduler - Implementation of the standard MachineSchedStrategy.
909 //===----------------------------------------------------------------------===//
912 /// ConvergingScheduler shrinks the unscheduled zone using heuristics to balance
914 class ConvergingScheduler : public MachineSchedStrategy {
916 /// Represent the type of SchedCandidate found within a single queue.
917 /// pickNodeBidirectional depends on these listed by decreasing priority.
919 NoCand, SingleExcess, SingleCritical, Cluster,
920 ResourceReduce, ResourceDemand, BotHeightReduce, BotPathReduce,
921 TopDepthReduce, TopPathReduce, SingleMax, MultiPressure, NextDefUse,
925 static const char *getReasonStr(ConvergingScheduler::CandReason Reason);
928 /// Policy for scheduling the next instruction in the candidate's zone.
931 unsigned ReduceResIdx;
932 unsigned DemandResIdx;
934 CandPolicy(): ReduceLatency(false), ReduceResIdx(0), DemandResIdx(0) {}
937 /// Status of an instruction's critical resource consumption.
938 struct SchedResourceDelta {
939 // Count critical resources in the scheduled region required by SU.
940 unsigned CritResources;
942 // Count critical resources from another region consumed by SU.
943 unsigned DemandedResources;
945 SchedResourceDelta(): CritResources(0), DemandedResources(0) {}
947 bool operator==(const SchedResourceDelta &RHS) const {
948 return CritResources == RHS.CritResources
949 && DemandedResources == RHS.DemandedResources;
951 bool operator!=(const SchedResourceDelta &RHS) const {
952 return !operator==(RHS);
956 /// Store the state used by ConvergingScheduler heuristics, required for the
957 /// lifetime of one invocation of pickNode().
958 struct SchedCandidate {
961 // The best SUnit candidate.
964 // The reason for this candidate.
967 // Register pressure values for the best candidate.
968 RegPressureDelta RPDelta;
970 // Critical resource consumption of the best candidate.
971 SchedResourceDelta ResDelta;
973 SchedCandidate(const CandPolicy &policy)
974 : Policy(policy), SU(NULL), Reason(NoCand) {}
976 bool isValid() const { return SU; }
978 // Copy the status of another candidate without changing policy.
979 void setBest(SchedCandidate &Best) {
980 assert(Best.Reason != NoCand && "uninitialized Sched candidate");
982 Reason = Best.Reason;
983 RPDelta = Best.RPDelta;
984 ResDelta = Best.ResDelta;
987 void initResourceDelta(const ScheduleDAGMI *DAG,
988 const TargetSchedModel *SchedModel);
991 /// Summarize the unscheduled region.
992 struct SchedRemainder {
993 // Critical path through the DAG in expected latency.
994 unsigned CriticalPath;
996 // Unscheduled resources
997 SmallVector<unsigned, 16> RemainingCounts;
998 // Critical resource for the unscheduled zone.
1000 // Number of micro-ops left to schedule.
1001 unsigned RemainingMicroOps;
1005 RemainingCounts.clear();
1007 RemainingMicroOps = 0;
1010 SchedRemainder() { reset(); }
1012 void init(ScheduleDAGMI *DAG, const TargetSchedModel *SchedModel);
1014 unsigned getMaxRemainingCount(const TargetSchedModel *SchedModel) const {
1015 if (!SchedModel->hasInstrSchedModel())
1019 RemainingMicroOps * SchedModel->getMicroOpFactor(),
1020 RemainingCounts[CritResIdx]);
1024 /// Each Scheduling boundary is associated with ready queues. It tracks the
1025 /// current cycle in the direction of movement, and maintains the state
1026 /// of "hazards" and other interlocks at the current cycle.
1027 struct SchedBoundary {
1029 const TargetSchedModel *SchedModel;
1030 SchedRemainder *Rem;
1032 ReadyQueue Available;
1036 // For heuristics, keep a list of the nodes that immediately depend on the
1037 // most recently scheduled node.
1038 SmallPtrSet<const SUnit*, 8> NextSUs;
1040 ScheduleHazardRecognizer *HazardRec;
1043 unsigned IssueCount;
1045 /// MinReadyCycle - Cycle of the soonest available instruction.
1046 unsigned MinReadyCycle;
1048 // The expected latency of the critical path in this scheduled zone.
1049 unsigned ExpectedLatency;
1051 // Resources used in the scheduled zone beyond this boundary.
1052 SmallVector<unsigned, 16> ResourceCounts;
1054 // Cache the critical resources ID in this scheduled zone.
1055 unsigned CritResIdx;
1057 // Is the scheduled region resource limited vs. latency limited.
1058 bool IsResourceLimited;
1060 unsigned ExpectedCount;
1063 // Remember the greatest min operand latency.
1064 unsigned MaxMinLatency;
1068 // A new HazardRec is created for each DAG and owned by SchedBoundary.
1073 CheckPending = false;
1078 MinReadyCycle = UINT_MAX;
1079 ExpectedLatency = 0;
1080 ResourceCounts.resize(1);
1081 assert(!ResourceCounts[0] && "nonzero count for bad resource");
1083 IsResourceLimited = false;
1088 // Reserve a zero-count for invalid CritResIdx.
1089 ResourceCounts.resize(1);
1092 /// Pending queues extend the ready queues with the same ID and the
1093 /// PendingFlag set.
1094 SchedBoundary(unsigned ID, const Twine &Name):
1095 DAG(0), SchedModel(0), Rem(0), Available(ID, Name+".A"),
1096 Pending(ID << ConvergingScheduler::LogMaxQID, Name+".P"),
1101 ~SchedBoundary() { delete HazardRec; }
1103 void init(ScheduleDAGMI *dag, const TargetSchedModel *smodel,
1104 SchedRemainder *rem);
1106 bool isTop() const {
1107 return Available.getID() == ConvergingScheduler::TopQID;
1110 unsigned getUnscheduledLatency(SUnit *SU) const {
1112 return SU->getHeight();
1113 return SU->getDepth() + SU->Latency;
1116 unsigned getCriticalCount() const {
1117 return ResourceCounts[CritResIdx];
1120 bool checkHazard(SUnit *SU);
1122 void setLatencyPolicy(CandPolicy &Policy);
1124 void releaseNode(SUnit *SU, unsigned ReadyCycle);
1128 void countResource(unsigned PIdx, unsigned Cycles);
1130 void bumpNode(SUnit *SU);
1132 void releasePending();
1134 void removeReady(SUnit *SU);
1136 SUnit *pickOnlyChoice();
1141 const TargetSchedModel *SchedModel;
1142 const TargetRegisterInfo *TRI;
1144 // State of the top and bottom scheduled instruction boundaries.
1150 /// SUnit::NodeQueueId: 0 (none), 1 (top), 2 (bot), 3 (both)
1157 ConvergingScheduler():
1158 DAG(0), SchedModel(0), TRI(0), Top(TopQID, "TopQ"), Bot(BotQID, "BotQ") {}
1160 virtual void initialize(ScheduleDAGMI *dag);
1162 virtual SUnit *pickNode(bool &IsTopNode);
1164 virtual void schedNode(SUnit *SU, bool IsTopNode);
1166 virtual void releaseTopNode(SUnit *SU);
1168 virtual void releaseBottomNode(SUnit *SU);
1170 virtual void registerRoots();
1174 ConvergingScheduler::SchedBoundary &CriticalZone,
1175 ConvergingScheduler::SchedCandidate &CriticalCand,
1176 ConvergingScheduler::SchedBoundary &OppositeZone,
1177 ConvergingScheduler::SchedCandidate &OppositeCand);
1179 void checkResourceLimits(ConvergingScheduler::SchedCandidate &TopCand,
1180 ConvergingScheduler::SchedCandidate &BotCand);
1182 void tryCandidate(SchedCandidate &Cand,
1183 SchedCandidate &TryCand,
1184 SchedBoundary &Zone,
1185 const RegPressureTracker &RPTracker,
1186 RegPressureTracker &TempTracker);
1188 SUnit *pickNodeBidirectional(bool &IsTopNode);
1190 void pickNodeFromQueue(SchedBoundary &Zone,
1191 const RegPressureTracker &RPTracker,
1192 SchedCandidate &Candidate);
1195 void traceCandidate(const SchedCandidate &Cand, const SchedBoundary &Zone);
1200 void ConvergingScheduler::SchedRemainder::
1201 init(ScheduleDAGMI *DAG, const TargetSchedModel *SchedModel) {
1203 if (!SchedModel->hasInstrSchedModel())
1205 RemainingCounts.resize(SchedModel->getNumProcResourceKinds());
1206 for (std::vector<SUnit>::iterator
1207 I = DAG->SUnits.begin(), E = DAG->SUnits.end(); I != E; ++I) {
1208 const MCSchedClassDesc *SC = DAG->getSchedClass(&*I);
1209 RemainingMicroOps += SchedModel->getNumMicroOps(I->getInstr(), SC);
1210 for (TargetSchedModel::ProcResIter
1211 PI = SchedModel->getWriteProcResBegin(SC),
1212 PE = SchedModel->getWriteProcResEnd(SC); PI != PE; ++PI) {
1213 unsigned PIdx = PI->ProcResourceIdx;
1214 unsigned Factor = SchedModel->getResourceFactor(PIdx);
1215 RemainingCounts[PIdx] += (Factor * PI->Cycles);
1218 for (unsigned PIdx = 0, PEnd = SchedModel->getNumProcResourceKinds();
1219 PIdx != PEnd; ++PIdx) {
1220 if ((int)(RemainingCounts[PIdx] - RemainingCounts[CritResIdx])
1221 >= (int)SchedModel->getLatencyFactor()) {
1227 void ConvergingScheduler::SchedBoundary::
1228 init(ScheduleDAGMI *dag, const TargetSchedModel *smodel, SchedRemainder *rem) {
1231 SchedModel = smodel;
1233 if (SchedModel->hasInstrSchedModel())
1234 ResourceCounts.resize(SchedModel->getNumProcResourceKinds());
1237 void ConvergingScheduler::initialize(ScheduleDAGMI *dag) {
1239 SchedModel = DAG->getSchedModel();
1242 Rem.init(DAG, SchedModel);
1243 Top.init(DAG, SchedModel, &Rem);
1244 Bot.init(DAG, SchedModel, &Rem);
1246 DAG->computeDFSResult();
1248 // Initialize resource counts.
1250 // Initialize the HazardRecognizers. If itineraries don't exist, are empty, or
1251 // are disabled, then these HazardRecs will be disabled.
1252 const InstrItineraryData *Itin = SchedModel->getInstrItineraries();
1253 const TargetMachine &TM = DAG->MF.getTarget();
1254 Top.HazardRec = TM.getInstrInfo()->CreateTargetMIHazardRecognizer(Itin, DAG);
1255 Bot.HazardRec = TM.getInstrInfo()->CreateTargetMIHazardRecognizer(Itin, DAG);
1257 assert((!ForceTopDown || !ForceBottomUp) &&
1258 "-misched-topdown incompatible with -misched-bottomup");
1261 void ConvergingScheduler::releaseTopNode(SUnit *SU) {
1262 if (SU->isScheduled)
1265 for (SUnit::pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
1267 unsigned PredReadyCycle = I->getSUnit()->TopReadyCycle;
1268 unsigned MinLatency = I->getMinLatency();
1270 Top.MaxMinLatency = std::max(MinLatency, Top.MaxMinLatency);
1272 if (SU->TopReadyCycle < PredReadyCycle + MinLatency)
1273 SU->TopReadyCycle = PredReadyCycle + MinLatency;
1275 Top.releaseNode(SU, SU->TopReadyCycle);
1278 void ConvergingScheduler::releaseBottomNode(SUnit *SU) {
1279 if (SU->isScheduled)
1282 assert(SU->getInstr() && "Scheduled SUnit must have instr");
1284 for (SUnit::succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
1288 unsigned SuccReadyCycle = I->getSUnit()->BotReadyCycle;
1289 unsigned MinLatency = I->getMinLatency();
1291 Bot.MaxMinLatency = std::max(MinLatency, Bot.MaxMinLatency);
1293 if (SU->BotReadyCycle < SuccReadyCycle + MinLatency)
1294 SU->BotReadyCycle = SuccReadyCycle + MinLatency;
1296 Bot.releaseNode(SU, SU->BotReadyCycle);
1299 void ConvergingScheduler::registerRoots() {
1300 Rem.CriticalPath = DAG->ExitSU.getDepth();
1301 // Some roots may not feed into ExitSU. Check all of them in case.
1302 for (std::vector<SUnit*>::const_iterator
1303 I = Bot.Available.begin(), E = Bot.Available.end(); I != E; ++I) {
1304 if ((*I)->getDepth() > Rem.CriticalPath)
1305 Rem.CriticalPath = (*I)->getDepth();
1307 DEBUG(dbgs() << "Critical Path: " << Rem.CriticalPath << '\n');
1310 /// Does this SU have a hazard within the current instruction group.
1312 /// The scheduler supports two modes of hazard recognition. The first is the
1313 /// ScheduleHazardRecognizer API. It is a fully general hazard recognizer that
1314 /// supports highly complicated in-order reservation tables
1315 /// (ScoreboardHazardRecognizer) and arbitraty target-specific logic.
1317 /// The second is a streamlined mechanism that checks for hazards based on
1318 /// simple counters that the scheduler itself maintains. It explicitly checks
1319 /// for instruction dispatch limitations, including the number of micro-ops that
1320 /// can dispatch per cycle.
1322 /// TODO: Also check whether the SU must start a new group.
1323 bool ConvergingScheduler::SchedBoundary::checkHazard(SUnit *SU) {
1324 if (HazardRec->isEnabled())
1325 return HazardRec->getHazardType(SU) != ScheduleHazardRecognizer::NoHazard;
1327 unsigned uops = SchedModel->getNumMicroOps(SU->getInstr());
1328 if ((IssueCount > 0) && (IssueCount + uops > SchedModel->getIssueWidth())) {
1329 DEBUG(dbgs() << " SU(" << SU->NodeNum << ") uops="
1330 << SchedModel->getNumMicroOps(SU->getInstr()) << '\n');
1336 /// Compute the remaining latency to determine whether ILP should be increased.
1337 void ConvergingScheduler::SchedBoundary::setLatencyPolicy(CandPolicy &Policy) {
1338 // FIXME: compile time. In all, we visit four queues here one we should only
1339 // need to visit the one that was last popped if we cache the result.
1340 unsigned RemLatency = 0;
1341 for (ReadyQueue::iterator I = Available.begin(), E = Available.end();
1343 unsigned L = getUnscheduledLatency(*I);
1347 for (ReadyQueue::iterator I = Pending.begin(), E = Pending.end();
1349 unsigned L = getUnscheduledLatency(*I);
1353 unsigned CriticalPathLimit = Rem->CriticalPath + SchedModel->getILPWindow();
1354 if (RemLatency + ExpectedLatency >= CriticalPathLimit
1355 && RemLatency > Rem->getMaxRemainingCount(SchedModel)) {
1356 Policy.ReduceLatency = true;
1357 DEBUG(dbgs() << "Increase ILP: " << Available.getName() << '\n');
1361 void ConvergingScheduler::SchedBoundary::releaseNode(SUnit *SU,
1362 unsigned ReadyCycle) {
1364 if (ReadyCycle < MinReadyCycle)
1365 MinReadyCycle = ReadyCycle;
1367 // Check for interlocks first. For the purpose of other heuristics, an
1368 // instruction that cannot issue appears as if it's not in the ReadyQueue.
1369 if (ReadyCycle > CurrCycle || checkHazard(SU))
1374 // Record this node as an immediate dependent of the scheduled node.
1378 /// Move the boundary of scheduled code by one cycle.
1379 void ConvergingScheduler::SchedBoundary::bumpCycle() {
1380 unsigned Width = SchedModel->getIssueWidth();
1381 IssueCount = (IssueCount <= Width) ? 0 : IssueCount - Width;
1383 unsigned NextCycle = CurrCycle + 1;
1384 assert(MinReadyCycle < UINT_MAX && "MinReadyCycle uninitialized");
1385 if (MinReadyCycle > NextCycle) {
1387 NextCycle = MinReadyCycle;
1390 if (!HazardRec->isEnabled()) {
1391 // Bypass HazardRec virtual calls.
1392 CurrCycle = NextCycle;
1395 // Bypass getHazardType calls in case of long latency.
1396 for (; CurrCycle != NextCycle; ++CurrCycle) {
1398 HazardRec->AdvanceCycle();
1400 HazardRec->RecedeCycle();
1403 CheckPending = true;
1404 IsResourceLimited = getCriticalCount() > std::max(ExpectedLatency, CurrCycle);
1406 DEBUG(dbgs() << " *** " << Available.getName() << " cycle "
1407 << CurrCycle << '\n');
1410 /// Add the given processor resource to this scheduled zone.
1411 void ConvergingScheduler::SchedBoundary::countResource(unsigned PIdx,
1413 unsigned Factor = SchedModel->getResourceFactor(PIdx);
1414 DEBUG(dbgs() << " " << SchedModel->getProcResource(PIdx)->Name
1415 << " +(" << Cycles << "x" << Factor
1416 << ") / " << SchedModel->getLatencyFactor() << '\n');
1418 unsigned Count = Factor * Cycles;
1419 ResourceCounts[PIdx] += Count;
1420 assert(Rem->RemainingCounts[PIdx] >= Count && "resource double counted");
1421 Rem->RemainingCounts[PIdx] -= Count;
1423 // Check if this resource exceeds the current critical resource by a full
1424 // cycle. If so, it becomes the critical resource.
1425 if ((int)(ResourceCounts[PIdx] - ResourceCounts[CritResIdx])
1426 >= (int)SchedModel->getLatencyFactor()) {
1428 DEBUG(dbgs() << " *** Critical resource "
1429 << SchedModel->getProcResource(PIdx)->Name << " x"
1430 << ResourceCounts[PIdx] << '\n');
1434 /// Move the boundary of scheduled code by one SUnit.
1435 void ConvergingScheduler::SchedBoundary::bumpNode(SUnit *SU) {
1436 // Update the reservation table.
1437 if (HazardRec->isEnabled()) {
1438 if (!isTop() && SU->isCall) {
1439 // Calls are scheduled with their preceding instructions. For bottom-up
1440 // scheduling, clear the pipeline state before emitting.
1443 HazardRec->EmitInstruction(SU);
1445 // Update resource counts and critical resource.
1446 if (SchedModel->hasInstrSchedModel()) {
1447 const MCSchedClassDesc *SC = DAG->getSchedClass(SU);
1448 Rem->RemainingMicroOps -= SchedModel->getNumMicroOps(SU->getInstr(), SC);
1449 for (TargetSchedModel::ProcResIter
1450 PI = SchedModel->getWriteProcResBegin(SC),
1451 PE = SchedModel->getWriteProcResEnd(SC); PI != PE; ++PI) {
1452 countResource(PI->ProcResourceIdx, PI->Cycles);
1456 if (SU->getDepth() > ExpectedLatency)
1457 ExpectedLatency = SU->getDepth();
1460 if (SU->getHeight() > ExpectedLatency)
1461 ExpectedLatency = SU->getHeight();
1464 IsResourceLimited = getCriticalCount() > std::max(ExpectedLatency, CurrCycle);
1466 // Check the instruction group dispatch limit.
1467 // TODO: Check if this SU must end a dispatch group.
1468 IssueCount += SchedModel->getNumMicroOps(SU->getInstr());
1470 // checkHazard prevents scheduling multiple instructions per cycle that exceed
1471 // issue width. However, we commonly reach the maximum. In this case
1472 // opportunistically bump the cycle to avoid uselessly checking everything in
1473 // the readyQ. Furthermore, a single instruction may produce more than one
1474 // cycle's worth of micro-ops.
1475 if (IssueCount >= SchedModel->getIssueWidth()) {
1476 DEBUG(dbgs() << " *** Max instrs at cycle " << CurrCycle << '\n');
1481 /// Release pending ready nodes in to the available queue. This makes them
1482 /// visible to heuristics.
1483 void ConvergingScheduler::SchedBoundary::releasePending() {
1484 // If the available queue is empty, it is safe to reset MinReadyCycle.
1485 if (Available.empty())
1486 MinReadyCycle = UINT_MAX;
1488 // Check to see if any of the pending instructions are ready to issue. If
1489 // so, add them to the available queue.
1490 for (unsigned i = 0, e = Pending.size(); i != e; ++i) {
1491 SUnit *SU = *(Pending.begin()+i);
1492 unsigned ReadyCycle = isTop() ? SU->TopReadyCycle : SU->BotReadyCycle;
1494 if (ReadyCycle < MinReadyCycle)
1495 MinReadyCycle = ReadyCycle;
1497 if (ReadyCycle > CurrCycle)
1500 if (checkHazard(SU))
1504 Pending.remove(Pending.begin()+i);
1507 DEBUG(if (!Pending.empty()) Pending.dump());
1508 CheckPending = false;
1511 /// Remove SU from the ready set for this boundary.
1512 void ConvergingScheduler::SchedBoundary::removeReady(SUnit *SU) {
1513 if (Available.isInQueue(SU))
1514 Available.remove(Available.find(SU));
1516 assert(Pending.isInQueue(SU) && "bad ready count");
1517 Pending.remove(Pending.find(SU));
1521 /// If this queue only has one ready candidate, return it. As a side effect,
1522 /// defer any nodes that now hit a hazard, and advance the cycle until at least
1523 /// one node is ready. If multiple instructions are ready, return NULL.
1524 SUnit *ConvergingScheduler::SchedBoundary::pickOnlyChoice() {
1528 if (IssueCount > 0) {
1529 // Defer any ready instrs that now have a hazard.
1530 for (ReadyQueue::iterator I = Available.begin(); I != Available.end();) {
1531 if (checkHazard(*I)) {
1533 I = Available.remove(I);
1539 for (unsigned i = 0; Available.empty(); ++i) {
1540 assert(i <= (HazardRec->getMaxLookAhead() + MaxMinLatency) &&
1541 "permanent hazard"); (void)i;
1545 if (Available.size() == 1)
1546 return *Available.begin();
1550 /// Record the candidate policy for opposite zones with different critical
1553 /// If the CriticalZone is latency limited, don't force a policy for the
1554 /// candidates here. Instead, setLatencyPolicy sets ReduceLatency if needed.
1555 void ConvergingScheduler::balanceZones(
1556 ConvergingScheduler::SchedBoundary &CriticalZone,
1557 ConvergingScheduler::SchedCandidate &CriticalCand,
1558 ConvergingScheduler::SchedBoundary &OppositeZone,
1559 ConvergingScheduler::SchedCandidate &OppositeCand) {
1561 if (!CriticalZone.IsResourceLimited)
1563 assert(SchedModel->hasInstrSchedModel() && "required schedmodel");
1565 SchedRemainder *Rem = CriticalZone.Rem;
1567 // If the critical zone is overconsuming a resource relative to the
1568 // remainder, try to reduce it.
1569 unsigned RemainingCritCount =
1570 Rem->RemainingCounts[CriticalZone.CritResIdx];
1571 if ((int)(Rem->getMaxRemainingCount(SchedModel) - RemainingCritCount)
1572 > (int)SchedModel->getLatencyFactor()) {
1573 CriticalCand.Policy.ReduceResIdx = CriticalZone.CritResIdx;
1574 DEBUG(dbgs() << "Balance " << CriticalZone.Available.getName() << " reduce "
1575 << SchedModel->getProcResource(CriticalZone.CritResIdx)->Name
1578 // If the other zone is underconsuming a resource relative to the full zone,
1579 // try to increase it.
1580 unsigned OppositeCount =
1581 OppositeZone.ResourceCounts[CriticalZone.CritResIdx];
1582 if ((int)(OppositeZone.ExpectedCount - OppositeCount)
1583 > (int)SchedModel->getLatencyFactor()) {
1584 OppositeCand.Policy.DemandResIdx = CriticalZone.CritResIdx;
1585 DEBUG(dbgs() << "Balance " << OppositeZone.Available.getName() << " demand "
1586 << SchedModel->getProcResource(OppositeZone.CritResIdx)->Name
1591 /// Determine if the scheduled zones exceed resource limits or critical path and
1592 /// set each candidate's ReduceHeight policy accordingly.
1593 void ConvergingScheduler::checkResourceLimits(
1594 ConvergingScheduler::SchedCandidate &TopCand,
1595 ConvergingScheduler::SchedCandidate &BotCand) {
1597 // Set ReduceLatency to true if needed.
1598 Bot.setLatencyPolicy(BotCand.Policy);
1599 Top.setLatencyPolicy(TopCand.Policy);
1601 // Handle resource-limited regions.
1602 if (Top.IsResourceLimited && Bot.IsResourceLimited
1603 && Top.CritResIdx == Bot.CritResIdx) {
1604 // If the scheduled critical resource in both zones is no longer the
1605 // critical remaining resource, attempt to reduce resource height both ways.
1606 if (Top.CritResIdx != Rem.CritResIdx) {
1607 TopCand.Policy.ReduceResIdx = Top.CritResIdx;
1608 BotCand.Policy.ReduceResIdx = Bot.CritResIdx;
1609 DEBUG(dbgs() << "Reduce scheduled "
1610 << SchedModel->getProcResource(Top.CritResIdx)->Name << '\n');
1614 // Handle latency-limited regions.
1615 if (!Top.IsResourceLimited && !Bot.IsResourceLimited) {
1616 // If the total scheduled expected latency exceeds the region's critical
1617 // path then reduce latency both ways.
1619 // Just because a zone is not resource limited does not mean it is latency
1620 // limited. Unbuffered resource, such as max micro-ops may cause CurrCycle
1621 // to exceed expected latency.
1622 if ((Top.ExpectedLatency + Bot.ExpectedLatency >= Rem.CriticalPath)
1623 && (Rem.CriticalPath > Top.CurrCycle + Bot.CurrCycle)) {
1624 TopCand.Policy.ReduceLatency = true;
1625 BotCand.Policy.ReduceLatency = true;
1626 DEBUG(dbgs() << "Reduce scheduled latency " << Top.ExpectedLatency
1627 << " + " << Bot.ExpectedLatency << '\n');
1631 // The critical resource is different in each zone, so request balancing.
1633 // Compute the cost of each zone.
1634 Top.ExpectedCount = std::max(Top.ExpectedLatency, Top.CurrCycle);
1635 Top.ExpectedCount = std::max(
1636 Top.getCriticalCount(),
1637 Top.ExpectedCount * SchedModel->getLatencyFactor());
1638 Bot.ExpectedCount = std::max(Bot.ExpectedLatency, Bot.CurrCycle);
1639 Bot.ExpectedCount = std::max(
1640 Bot.getCriticalCount(),
1641 Bot.ExpectedCount * SchedModel->getLatencyFactor());
1643 balanceZones(Top, TopCand, Bot, BotCand);
1644 balanceZones(Bot, BotCand, Top, TopCand);
1647 void ConvergingScheduler::SchedCandidate::
1648 initResourceDelta(const ScheduleDAGMI *DAG,
1649 const TargetSchedModel *SchedModel) {
1650 if (!Policy.ReduceResIdx && !Policy.DemandResIdx)
1653 const MCSchedClassDesc *SC = DAG->getSchedClass(SU);
1654 for (TargetSchedModel::ProcResIter
1655 PI = SchedModel->getWriteProcResBegin(SC),
1656 PE = SchedModel->getWriteProcResEnd(SC); PI != PE; ++PI) {
1657 if (PI->ProcResourceIdx == Policy.ReduceResIdx)
1658 ResDelta.CritResources += PI->Cycles;
1659 if (PI->ProcResourceIdx == Policy.DemandResIdx)
1660 ResDelta.DemandedResources += PI->Cycles;
1664 /// Return true if this heuristic determines order.
1665 static bool tryLess(unsigned TryVal, unsigned CandVal,
1666 ConvergingScheduler::SchedCandidate &TryCand,
1667 ConvergingScheduler::SchedCandidate &Cand,
1668 ConvergingScheduler::CandReason Reason) {
1669 if (TryVal < CandVal) {
1670 TryCand.Reason = Reason;
1673 if (TryVal > CandVal) {
1674 if (Cand.Reason > Reason)
1675 Cand.Reason = Reason;
1681 static bool tryGreater(unsigned TryVal, unsigned CandVal,
1682 ConvergingScheduler::SchedCandidate &TryCand,
1683 ConvergingScheduler::SchedCandidate &Cand,
1684 ConvergingScheduler::CandReason Reason) {
1685 if (TryVal > CandVal) {
1686 TryCand.Reason = Reason;
1689 if (TryVal < CandVal) {
1690 if (Cand.Reason > Reason)
1691 Cand.Reason = Reason;
1697 static unsigned getWeakLeft(const SUnit *SU, bool isTop) {
1698 return (isTop) ? SU->WeakPredsLeft : SU->WeakSuccsLeft;
1701 /// Apply a set of heursitics to a new candidate. Heuristics are currently
1702 /// hierarchical. This may be more efficient than a graduated cost model because
1703 /// we don't need to evaluate all aspects of the model for each node in the
1704 /// queue. But it's really done to make the heuristics easier to debug and
1705 /// statistically analyze.
1707 /// \param Cand provides the policy and current best candidate.
1708 /// \param TryCand refers to the next SUnit candidate, otherwise uninitialized.
1709 /// \param Zone describes the scheduled zone that we are extending.
1710 /// \param RPTracker describes reg pressure within the scheduled zone.
1711 /// \param TempTracker is a scratch pressure tracker to reuse in queries.
1712 void ConvergingScheduler::tryCandidate(SchedCandidate &Cand,
1713 SchedCandidate &TryCand,
1714 SchedBoundary &Zone,
1715 const RegPressureTracker &RPTracker,
1716 RegPressureTracker &TempTracker) {
1718 // Always initialize TryCand's RPDelta.
1719 TempTracker.getMaxPressureDelta(TryCand.SU->getInstr(), TryCand.RPDelta,
1720 DAG->getRegionCriticalPSets(),
1721 DAG->getRegPressure().MaxSetPressure);
1723 // Initialize the candidate if needed.
1724 if (!Cand.isValid()) {
1725 TryCand.Reason = NodeOrder;
1728 // Avoid exceeding the target's limit.
1729 if (tryLess(TryCand.RPDelta.Excess.UnitIncrease,
1730 Cand.RPDelta.Excess.UnitIncrease, TryCand, Cand, SingleExcess))
1732 if (Cand.Reason == SingleExcess)
1733 Cand.Reason = MultiPressure;
1735 // Avoid increasing the max critical pressure in the scheduled region.
1736 if (tryLess(TryCand.RPDelta.CriticalMax.UnitIncrease,
1737 Cand.RPDelta.CriticalMax.UnitIncrease,
1738 TryCand, Cand, SingleCritical))
1740 if (Cand.Reason == SingleCritical)
1741 Cand.Reason = MultiPressure;
1743 // Keep clustered nodes together to encourage downstream peephole
1744 // optimizations which may reduce resource requirements.
1746 // This is a best effort to set things up for a post-RA pass. Optimizations
1747 // like generating loads of multiple registers should ideally be done within
1748 // the scheduler pass by combining the loads during DAG postprocessing.
1749 const SUnit *NextClusterSU =
1750 Zone.isTop() ? DAG->getNextClusterSucc() : DAG->getNextClusterPred();
1751 if (tryGreater(TryCand.SU == NextClusterSU, Cand.SU == NextClusterSU,
1752 TryCand, Cand, Cluster))
1754 // Currently, weak edges are for clustering, so we hard-code that reason.
1755 // However, deferring the current TryCand will not change Cand's reason.
1756 CandReason OrigReason = Cand.Reason;
1757 if (tryLess(getWeakLeft(TryCand.SU, Zone.isTop()),
1758 getWeakLeft(Cand.SU, Zone.isTop()),
1759 TryCand, Cand, Cluster)) {
1760 Cand.Reason = OrigReason;
1763 // Avoid critical resource consumption and balance the schedule.
1764 TryCand.initResourceDelta(DAG, SchedModel);
1765 if (tryLess(TryCand.ResDelta.CritResources, Cand.ResDelta.CritResources,
1766 TryCand, Cand, ResourceReduce))
1768 if (tryGreater(TryCand.ResDelta.DemandedResources,
1769 Cand.ResDelta.DemandedResources,
1770 TryCand, Cand, ResourceDemand))
1773 // Avoid serializing long latency dependence chains.
1774 if (Cand.Policy.ReduceLatency) {
1776 if (Cand.SU->getDepth() * SchedModel->getLatencyFactor()
1777 > Zone.ExpectedCount) {
1778 if (tryLess(TryCand.SU->getDepth(), Cand.SU->getDepth(),
1779 TryCand, Cand, TopDepthReduce))
1782 if (tryGreater(TryCand.SU->getHeight(), Cand.SU->getHeight(),
1783 TryCand, Cand, TopPathReduce))
1787 if (Cand.SU->getHeight() * SchedModel->getLatencyFactor()
1788 > Zone.ExpectedCount) {
1789 if (tryLess(TryCand.SU->getHeight(), Cand.SU->getHeight(),
1790 TryCand, Cand, BotHeightReduce))
1793 if (tryGreater(TryCand.SU->getDepth(), Cand.SU->getDepth(),
1794 TryCand, Cand, BotPathReduce))
1799 // Avoid increasing the max pressure of the entire region.
1800 if (tryLess(TryCand.RPDelta.CurrentMax.UnitIncrease,
1801 Cand.RPDelta.CurrentMax.UnitIncrease, TryCand, Cand, SingleMax))
1803 if (Cand.Reason == SingleMax)
1804 Cand.Reason = MultiPressure;
1806 // Prefer immediate defs/users of the last scheduled instruction. This is a
1807 // nice pressure avoidance strategy that also conserves the processor's
1808 // register renaming resources and keeps the machine code readable.
1809 if (tryGreater(Zone.NextSUs.count(TryCand.SU), Zone.NextSUs.count(Cand.SU),
1810 TryCand, Cand, NextDefUse))
1813 // Fall through to original instruction order.
1814 if ((Zone.isTop() && TryCand.SU->NodeNum < Cand.SU->NodeNum)
1815 || (!Zone.isTop() && TryCand.SU->NodeNum > Cand.SU->NodeNum)) {
1816 TryCand.Reason = NodeOrder;
1820 /// pickNodeFromQueue helper that returns true if the LHS reg pressure effect is
1821 /// more desirable than RHS from scheduling standpoint.
1822 static bool compareRPDelta(const RegPressureDelta &LHS,
1823 const RegPressureDelta &RHS) {
1824 // Compare each component of pressure in decreasing order of importance
1825 // without checking if any are valid. Invalid PressureElements are assumed to
1826 // have UnitIncrease==0, so are neutral.
1828 // Avoid increasing the max critical pressure in the scheduled region.
1829 if (LHS.Excess.UnitIncrease != RHS.Excess.UnitIncrease) {
1830 DEBUG(dbgs() << "RP excess top - bot: "
1831 << (LHS.Excess.UnitIncrease - RHS.Excess.UnitIncrease) << '\n');
1832 return LHS.Excess.UnitIncrease < RHS.Excess.UnitIncrease;
1834 // Avoid increasing the max critical pressure in the scheduled region.
1835 if (LHS.CriticalMax.UnitIncrease != RHS.CriticalMax.UnitIncrease) {
1836 DEBUG(dbgs() << "RP critical top - bot: "
1837 << (LHS.CriticalMax.UnitIncrease - RHS.CriticalMax.UnitIncrease)
1839 return LHS.CriticalMax.UnitIncrease < RHS.CriticalMax.UnitIncrease;
1841 // Avoid increasing the max pressure of the entire region.
1842 if (LHS.CurrentMax.UnitIncrease != RHS.CurrentMax.UnitIncrease) {
1843 DEBUG(dbgs() << "RP current top - bot: "
1844 << (LHS.CurrentMax.UnitIncrease - RHS.CurrentMax.UnitIncrease)
1846 return LHS.CurrentMax.UnitIncrease < RHS.CurrentMax.UnitIncrease;
1852 const char *ConvergingScheduler::getReasonStr(
1853 ConvergingScheduler::CandReason Reason) {
1855 case NoCand: return "NOCAND ";
1856 case SingleExcess: return "REG-EXCESS";
1857 case SingleCritical: return "REG-CRIT ";
1858 case Cluster: return "CLUSTER ";
1859 case SingleMax: return "REG-MAX ";
1860 case MultiPressure: return "REG-MULTI ";
1861 case ResourceReduce: return "RES-REDUCE";
1862 case ResourceDemand: return "RES-DEMAND";
1863 case TopDepthReduce: return "TOP-DEPTH ";
1864 case TopPathReduce: return "TOP-PATH ";
1865 case BotHeightReduce:return "BOT-HEIGHT";
1866 case BotPathReduce: return "BOT-PATH ";
1867 case NextDefUse: return "DEF-USE ";
1868 case NodeOrder: return "ORDER ";
1870 llvm_unreachable("Unknown reason!");
1873 void ConvergingScheduler::traceCandidate(const SchedCandidate &Cand,
1874 const SchedBoundary &Zone) {
1875 const char *Label = getReasonStr(Cand.Reason);
1877 unsigned ResIdx = 0;
1878 unsigned Latency = 0;
1879 switch (Cand.Reason) {
1883 P = Cand.RPDelta.Excess;
1885 case SingleCritical:
1886 P = Cand.RPDelta.CriticalMax;
1889 P = Cand.RPDelta.CurrentMax;
1891 case ResourceReduce:
1892 ResIdx = Cand.Policy.ReduceResIdx;
1894 case ResourceDemand:
1895 ResIdx = Cand.Policy.DemandResIdx;
1897 case TopDepthReduce:
1898 Latency = Cand.SU->getDepth();
1901 Latency = Cand.SU->getHeight();
1903 case BotHeightReduce:
1904 Latency = Cand.SU->getHeight();
1907 Latency = Cand.SU->getDepth();
1910 dbgs() << Label << " " << Zone.Available.getName() << " ";
1912 dbgs() << TRI->getRegPressureSetName(P.PSetID) << ":" << P.UnitIncrease
1917 dbgs() << SchedModel->getProcResource(ResIdx)->Name << " ";
1921 dbgs() << Latency << " cycles ";
1928 /// Pick the best candidate from the top queue.
1930 /// TODO: getMaxPressureDelta results can be mostly cached for each SUnit during
1931 /// DAG building. To adjust for the current scheduling location we need to
1932 /// maintain the number of vreg uses remaining to be top-scheduled.
1933 void ConvergingScheduler::pickNodeFromQueue(SchedBoundary &Zone,
1934 const RegPressureTracker &RPTracker,
1935 SchedCandidate &Cand) {
1936 ReadyQueue &Q = Zone.Available;
1940 // getMaxPressureDelta temporarily modifies the tracker.
1941 RegPressureTracker &TempTracker = const_cast<RegPressureTracker&>(RPTracker);
1943 for (ReadyQueue::iterator I = Q.begin(), E = Q.end(); I != E; ++I) {
1945 SchedCandidate TryCand(Cand.Policy);
1947 tryCandidate(Cand, TryCand, Zone, RPTracker, TempTracker);
1948 if (TryCand.Reason != NoCand) {
1949 // Initialize resource delta if needed in case future heuristics query it.
1950 if (TryCand.ResDelta == SchedResourceDelta())
1951 TryCand.initResourceDelta(DAG, SchedModel);
1952 Cand.setBest(TryCand);
1953 DEBUG(traceCandidate(Cand, Zone));
1958 static void tracePick(const ConvergingScheduler::SchedCandidate &Cand,
1960 DEBUG(dbgs() << "Pick " << (IsTop ? "top" : "bot")
1961 << " SU(" << Cand.SU->NodeNum << ") "
1962 << ConvergingScheduler::getReasonStr(Cand.Reason) << '\n');
1965 /// Pick the best candidate node from either the top or bottom queue.
1966 SUnit *ConvergingScheduler::pickNodeBidirectional(bool &IsTopNode) {
1967 // Schedule as far as possible in the direction of no choice. This is most
1968 // efficient, but also provides the best heuristics for CriticalPSets.
1969 if (SUnit *SU = Bot.pickOnlyChoice()) {
1973 if (SUnit *SU = Top.pickOnlyChoice()) {
1977 CandPolicy NoPolicy;
1978 SchedCandidate BotCand(NoPolicy);
1979 SchedCandidate TopCand(NoPolicy);
1980 checkResourceLimits(TopCand, BotCand);
1982 // Prefer bottom scheduling when heuristics are silent.
1983 pickNodeFromQueue(Bot, DAG->getBotRPTracker(), BotCand);
1984 assert(BotCand.Reason != NoCand && "failed to find the first candidate");
1986 // If either Q has a single candidate that provides the least increase in
1987 // Excess pressure, we can immediately schedule from that Q.
1989 // RegionCriticalPSets summarizes the pressure within the scheduled region and
1990 // affects picking from either Q. If scheduling in one direction must
1991 // increase pressure for one of the excess PSets, then schedule in that
1992 // direction first to provide more freedom in the other direction.
1993 if (BotCand.Reason == SingleExcess || BotCand.Reason == SingleCritical) {
1995 tracePick(BotCand, IsTopNode);
1998 // Check if the top Q has a better candidate.
1999 pickNodeFromQueue(Top, DAG->getTopRPTracker(), TopCand);
2000 assert(TopCand.Reason != NoCand && "failed to find the first candidate");
2002 // If either Q has a single candidate that minimizes pressure above the
2003 // original region's pressure pick it.
2004 if (TopCand.Reason <= SingleMax || BotCand.Reason <= SingleMax) {
2005 if (TopCand.Reason < BotCand.Reason) {
2007 tracePick(TopCand, IsTopNode);
2011 tracePick(BotCand, IsTopNode);
2014 // Check for a salient pressure difference and pick the best from either side.
2015 if (compareRPDelta(TopCand.RPDelta, BotCand.RPDelta)) {
2017 tracePick(TopCand, IsTopNode);
2020 // Otherwise prefer the bottom candidate, in node order if all else failed.
2021 if (TopCand.Reason < BotCand.Reason) {
2023 tracePick(TopCand, IsTopNode);
2027 tracePick(BotCand, IsTopNode);
2031 /// Pick the best node to balance the schedule. Implements MachineSchedStrategy.
2032 SUnit *ConvergingScheduler::pickNode(bool &IsTopNode) {
2033 if (DAG->top() == DAG->bottom()) {
2034 assert(Top.Available.empty() && Top.Pending.empty() &&
2035 Bot.Available.empty() && Bot.Pending.empty() && "ReadyQ garbage");
2041 SU = Top.pickOnlyChoice();
2043 CandPolicy NoPolicy;
2044 SchedCandidate TopCand(NoPolicy);
2045 pickNodeFromQueue(Top, DAG->getTopRPTracker(), TopCand);
2046 assert(TopCand.Reason != NoCand && "failed to find the first candidate");
2051 else if (ForceBottomUp) {
2052 SU = Bot.pickOnlyChoice();
2054 CandPolicy NoPolicy;
2055 SchedCandidate BotCand(NoPolicy);
2056 pickNodeFromQueue(Bot, DAG->getBotRPTracker(), BotCand);
2057 assert(BotCand.Reason != NoCand && "failed to find the first candidate");
2063 SU = pickNodeBidirectional(IsTopNode);
2065 } while (SU->isScheduled);
2067 if (SU->isTopReady())
2068 Top.removeReady(SU);
2069 if (SU->isBottomReady())
2070 Bot.removeReady(SU);
2072 DEBUG(dbgs() << "*** " << (IsTopNode ? "Top" : "Bottom")
2073 << " Scheduling Instruction in cycle "
2074 << (IsTopNode ? Top.CurrCycle : Bot.CurrCycle) << '\n';
2079 /// Update the scheduler's state after scheduling a node. This is the same node
2080 /// that was just returned by pickNode(). However, ScheduleDAGMI needs to update
2081 /// it's state based on the current cycle before MachineSchedStrategy does.
2082 void ConvergingScheduler::schedNode(SUnit *SU, bool IsTopNode) {
2084 SU->TopReadyCycle = Top.CurrCycle;
2088 SU->BotReadyCycle = Bot.CurrCycle;
2093 /// Create the standard converging machine scheduler. This will be used as the
2094 /// default scheduler if the target does not set a default.
2095 static ScheduleDAGInstrs *createConvergingSched(MachineSchedContext *C) {
2096 assert((!ForceTopDown || !ForceBottomUp) &&
2097 "-misched-topdown incompatible with -misched-bottomup");
2098 ScheduleDAGMI *DAG = new ScheduleDAGMI(C, new ConvergingScheduler());
2099 // Register DAG post-processors.
2100 if (EnableLoadCluster)
2101 DAG->addMutation(new LoadClusterMutation(DAG->TII, DAG->TRI));
2102 if (EnableMacroFusion)
2103 DAG->addMutation(new MacroFusion(DAG->TII));
2106 static MachineSchedRegistry
2107 ConvergingSchedRegistry("converge", "Standard converging scheduler.",
2108 createConvergingSched);
2110 //===----------------------------------------------------------------------===//
2111 // ILP Scheduler. Currently for experimental analysis of heuristics.
2112 //===----------------------------------------------------------------------===//
2115 /// \brief Order nodes by the ILP metric.
2117 const SchedDFSResult *DFSResult;
2118 const BitVector *ScheduledTrees;
2121 ILPOrder(bool MaxILP): DFSResult(0), ScheduledTrees(0), MaximizeILP(MaxILP) {}
2123 /// \brief Apply a less-than relation on node priority.
2125 /// (Return true if A comes after B in the Q.)
2126 bool operator()(const SUnit *A, const SUnit *B) const {
2127 unsigned SchedTreeA = DFSResult->getSubtreeID(A);
2128 unsigned SchedTreeB = DFSResult->getSubtreeID(B);
2129 if (SchedTreeA != SchedTreeB) {
2130 // Unscheduled trees have lower priority.
2131 if (ScheduledTrees->test(SchedTreeA) != ScheduledTrees->test(SchedTreeB))
2132 return ScheduledTrees->test(SchedTreeB);
2134 // Trees with shallower connections have have lower priority.
2135 if (DFSResult->getSubtreeLevel(SchedTreeA)
2136 != DFSResult->getSubtreeLevel(SchedTreeB)) {
2137 return DFSResult->getSubtreeLevel(SchedTreeA)
2138 < DFSResult->getSubtreeLevel(SchedTreeB);
2142 return DFSResult->getILP(A) < DFSResult->getILP(B);
2144 return DFSResult->getILP(A) > DFSResult->getILP(B);
2148 /// \brief Schedule based on the ILP metric.
2149 class ILPScheduler : public MachineSchedStrategy {
2150 /// In case all subtrees are eventually connected to a common root through
2151 /// data dependence (e.g. reduction), place an upper limit on their size.
2153 /// FIXME: A subtree limit is generally good, but in the situation commented
2154 /// above, where multiple similar subtrees feed a common root, we should
2155 /// only split at a point where the resulting subtrees will be balanced.
2156 /// (a motivating test case must be found).
2157 static const unsigned SubtreeLimit = 16;
2162 std::vector<SUnit*> ReadyQ;
2164 ILPScheduler(bool MaximizeILP): DAG(0), Cmp(MaximizeILP) {}
2166 virtual void initialize(ScheduleDAGMI *dag) {
2168 DAG->computeDFSResult();
2169 Cmp.DFSResult = DAG->getDFSResult();
2170 Cmp.ScheduledTrees = &DAG->getScheduledTrees();
2174 virtual void registerRoots() {
2175 // Restore the heap in ReadyQ with the updated DFS results.
2176 std::make_heap(ReadyQ.begin(), ReadyQ.end(), Cmp);
2179 /// Implement MachineSchedStrategy interface.
2180 /// -----------------------------------------
2182 /// Callback to select the highest priority node from the ready Q.
2183 virtual SUnit *pickNode(bool &IsTopNode) {
2184 if (ReadyQ.empty()) return NULL;
2185 std::pop_heap(ReadyQ.begin(), ReadyQ.end(), Cmp);
2186 SUnit *SU = ReadyQ.back();
2189 DEBUG(dbgs() << "*** Scheduling " << "SU(" << SU->NodeNum << "): "
2191 << " ILP: " << DAG->getDFSResult()->getILP(SU)
2192 << " Tree: " << DAG->getDFSResult()->getSubtreeID(SU) << " @"
2193 << DAG->getDFSResult()->getSubtreeLevel(
2194 DAG->getDFSResult()->getSubtreeID(SU)) << '\n');
2198 /// \brief Scheduler callback to notify that a new subtree is scheduled.
2199 virtual void scheduleTree(unsigned SubtreeID) {
2200 std::make_heap(ReadyQ.begin(), ReadyQ.end(), Cmp);
2203 /// Callback after a node is scheduled. Mark a newly scheduled tree, notify
2204 /// DFSResults, and resort the priority Q.
2205 virtual void schedNode(SUnit *SU, bool IsTopNode) {
2206 assert(!IsTopNode && "SchedDFSResult needs bottom-up");
2209 virtual void releaseTopNode(SUnit *) { /*only called for top roots*/ }
2211 virtual void releaseBottomNode(SUnit *SU) {
2212 ReadyQ.push_back(SU);
2213 std::push_heap(ReadyQ.begin(), ReadyQ.end(), Cmp);
2218 static ScheduleDAGInstrs *createILPMaxScheduler(MachineSchedContext *C) {
2219 return new ScheduleDAGMI(C, new ILPScheduler(true));
2221 static ScheduleDAGInstrs *createILPMinScheduler(MachineSchedContext *C) {
2222 return new ScheduleDAGMI(C, new ILPScheduler(false));
2224 static MachineSchedRegistry ILPMaxRegistry(
2225 "ilpmax", "Schedule bottom-up for max ILP", createILPMaxScheduler);
2226 static MachineSchedRegistry ILPMinRegistry(
2227 "ilpmin", "Schedule bottom-up for min ILP", createILPMinScheduler);
2229 //===----------------------------------------------------------------------===//
2230 // Machine Instruction Shuffler for Correctness Testing
2231 //===----------------------------------------------------------------------===//
2235 /// Apply a less-than relation on the node order, which corresponds to the
2236 /// instruction order prior to scheduling. IsReverse implements greater-than.
2237 template<bool IsReverse>
2239 bool operator()(SUnit *A, SUnit *B) const {
2241 return A->NodeNum > B->NodeNum;
2243 return A->NodeNum < B->NodeNum;
2247 /// Reorder instructions as much as possible.
2248 class InstructionShuffler : public MachineSchedStrategy {
2252 // Using a less-than relation (SUnitOrder<false>) for the TopQ priority
2253 // gives nodes with a higher number higher priority causing the latest
2254 // instructions to be scheduled first.
2255 PriorityQueue<SUnit*, std::vector<SUnit*>, SUnitOrder<false> >
2257 // When scheduling bottom-up, use greater-than as the queue priority.
2258 PriorityQueue<SUnit*, std::vector<SUnit*>, SUnitOrder<true> >
2261 InstructionShuffler(bool alternate, bool topdown)
2262 : IsAlternating(alternate), IsTopDown(topdown) {}
2264 virtual void initialize(ScheduleDAGMI *) {
2269 /// Implement MachineSchedStrategy interface.
2270 /// -----------------------------------------
2272 virtual SUnit *pickNode(bool &IsTopNode) {
2276 if (TopQ.empty()) return NULL;
2279 } while (SU->isScheduled);
2284 if (BottomQ.empty()) return NULL;
2287 } while (SU->isScheduled);
2291 IsTopDown = !IsTopDown;
2295 virtual void schedNode(SUnit *SU, bool IsTopNode) {}
2297 virtual void releaseTopNode(SUnit *SU) {
2300 virtual void releaseBottomNode(SUnit *SU) {
2306 static ScheduleDAGInstrs *createInstructionShuffler(MachineSchedContext *C) {
2307 bool Alternate = !ForceTopDown && !ForceBottomUp;
2308 bool TopDown = !ForceBottomUp;
2309 assert((TopDown || !ForceTopDown) &&
2310 "-misched-topdown incompatible with -misched-bottomup");
2311 return new ScheduleDAGMI(C, new InstructionShuffler(Alternate, TopDown));
2313 static MachineSchedRegistry ShufflerRegistry(
2314 "shuffle", "Shuffle machine instructions alternating directions",
2315 createInstructionShuffler);
2318 //===----------------------------------------------------------------------===//
2319 // GraphWriter support for ScheduleDAGMI.
2320 //===----------------------------------------------------------------------===//
2325 template<> struct GraphTraits<
2326 ScheduleDAGMI*> : public GraphTraits<ScheduleDAG*> {};
2329 struct DOTGraphTraits<ScheduleDAGMI*> : public DefaultDOTGraphTraits {
2331 DOTGraphTraits (bool isSimple=false) : DefaultDOTGraphTraits(isSimple) {}
2333 static std::string getGraphName(const ScheduleDAG *G) {
2334 return G->MF.getName();
2337 static bool renderGraphFromBottomUp() {
2341 static bool isNodeHidden(const SUnit *Node) {
2342 return (Node->NumPreds > 10 || Node->NumSuccs > 10);
2345 static bool hasNodeAddressLabel(const SUnit *Node,
2346 const ScheduleDAG *Graph) {
2350 /// If you want to override the dot attributes printed for a particular
2351 /// edge, override this method.
2352 static std::string getEdgeAttributes(const SUnit *Node,
2354 const ScheduleDAG *Graph) {
2355 if (EI.isArtificialDep())
2356 return "color=cyan,style=dashed";
2358 return "color=blue,style=dashed";
2362 static std::string getNodeLabel(const SUnit *SU, const ScheduleDAG *G) {
2364 raw_string_ostream SS(Str);
2365 SS << "SU(" << SU->NodeNum << ')';
2368 static std::string getNodeDescription(const SUnit *SU, const ScheduleDAG *G) {
2369 return G->getGraphNodeLabel(SU);
2372 static std::string getNodeAttributes(const SUnit *N,
2373 const ScheduleDAG *Graph) {
2374 std::string Str("shape=Mrecord");
2375 const SchedDFSResult *DFS =
2376 static_cast<const ScheduleDAGMI*>(Graph)->getDFSResult();
2378 Str += ",style=filled,fillcolor=\"#";
2379 Str += DOT::getColorString(DFS->getSubtreeID(N));
2388 /// viewGraph - Pop up a ghostview window with the reachable parts of the DAG
2389 /// rendered using 'dot'.
2391 void ScheduleDAGMI::viewGraph(const Twine &Name, const Twine &Title) {
2393 ViewGraph(this, Name, false, Title);
2395 errs() << "ScheduleDAGMI::viewGraph is only available in debug builds on "
2396 << "systems with Graphviz or gv!\n";
2400 /// Out-of-line implementation with no arguments is handy for gdb.
2401 void ScheduleDAGMI::viewGraph() {
2402 viewGraph(getDAGName(), "Scheduling-Units Graph for " + getDAGName());