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"
33 #include "llvm/Target/TargetInstrInfo.h"
39 cl::opt<bool> ForceTopDown("misched-topdown", cl::Hidden,
40 cl::desc("Force top-down list scheduling"));
41 cl::opt<bool> ForceBottomUp("misched-bottomup", cl::Hidden,
42 cl::desc("Force bottom-up list scheduling"));
46 static cl::opt<bool> ViewMISchedDAGs("view-misched-dags", cl::Hidden,
47 cl::desc("Pop up a window to show MISched dags after they are processed"));
49 static cl::opt<unsigned> MISchedCutoff("misched-cutoff", cl::Hidden,
50 cl::desc("Stop scheduling after N instructions"), cl::init(~0U));
52 static bool ViewMISchedDAGs = false;
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::canAddEdge(SUnit *SuccSU, SUnit *PredSU) {
327 return SuccSU == &ExitSU || !Topo.IsReachable(PredSU, SuccSU);
330 bool ScheduleDAGMI::addEdge(SUnit *SuccSU, const SDep &PredDep) {
331 if (SuccSU != &ExitSU) {
332 // Do not use WillCreateCycle, it assumes SD scheduling.
333 // If Pred is reachable from Succ, then the edge creates a cycle.
334 if (Topo.IsReachable(PredDep.getSUnit(), SuccSU))
336 Topo.AddPred(SuccSU, PredDep.getSUnit());
338 SuccSU->addPred(PredDep, /*Required=*/!PredDep.isArtificial());
339 // Return true regardless of whether a new edge needed to be inserted.
343 /// ReleaseSucc - Decrement the NumPredsLeft count of a successor. When
344 /// NumPredsLeft reaches zero, release the successor node.
346 /// FIXME: Adjust SuccSU height based on MinLatency.
347 void ScheduleDAGMI::releaseSucc(SUnit *SU, SDep *SuccEdge) {
348 SUnit *SuccSU = SuccEdge->getSUnit();
350 if (SuccEdge->isWeak()) {
351 --SuccSU->WeakPredsLeft;
352 if (SuccEdge->isCluster())
353 NextClusterSucc = SuccSU;
357 if (SuccSU->NumPredsLeft == 0) {
358 dbgs() << "*** Scheduling failed! ***\n";
360 dbgs() << " has been released too many times!\n";
364 --SuccSU->NumPredsLeft;
365 if (SuccSU->NumPredsLeft == 0 && SuccSU != &ExitSU)
366 SchedImpl->releaseTopNode(SuccSU);
369 /// releaseSuccessors - Call releaseSucc on each of SU's successors.
370 void ScheduleDAGMI::releaseSuccessors(SUnit *SU) {
371 for (SUnit::succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
373 releaseSucc(SU, &*I);
377 /// ReleasePred - Decrement the NumSuccsLeft count of a predecessor. When
378 /// NumSuccsLeft reaches zero, release the predecessor node.
380 /// FIXME: Adjust PredSU height based on MinLatency.
381 void ScheduleDAGMI::releasePred(SUnit *SU, SDep *PredEdge) {
382 SUnit *PredSU = PredEdge->getSUnit();
384 if (PredEdge->isWeak()) {
385 --PredSU->WeakSuccsLeft;
386 if (PredEdge->isCluster())
387 NextClusterPred = PredSU;
391 if (PredSU->NumSuccsLeft == 0) {
392 dbgs() << "*** Scheduling failed! ***\n";
394 dbgs() << " has been released too many times!\n";
398 --PredSU->NumSuccsLeft;
399 if (PredSU->NumSuccsLeft == 0 && PredSU != &EntrySU)
400 SchedImpl->releaseBottomNode(PredSU);
403 /// releasePredecessors - Call releasePred on each of SU's predecessors.
404 void ScheduleDAGMI::releasePredecessors(SUnit *SU) {
405 for (SUnit::pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
407 releasePred(SU, &*I);
411 /// This is normally called from the main scheduler loop but may also be invoked
412 /// by the scheduling strategy to perform additional code motion.
413 void ScheduleDAGMI::moveInstruction(MachineInstr *MI,
414 MachineBasicBlock::iterator InsertPos) {
415 // Advance RegionBegin if the first instruction moves down.
416 if (&*RegionBegin == MI)
419 // Update the instruction stream.
420 BB->splice(InsertPos, BB, MI);
422 // Update LiveIntervals
423 LIS->handleMove(MI, /*UpdateFlags=*/true);
425 // Recede RegionBegin if an instruction moves above the first.
426 if (RegionBegin == InsertPos)
430 bool ScheduleDAGMI::checkSchedLimit() {
432 if (NumInstrsScheduled == MISchedCutoff && MISchedCutoff != ~0U) {
433 CurrentTop = CurrentBottom;
436 ++NumInstrsScheduled;
441 /// enterRegion - Called back from MachineScheduler::runOnMachineFunction after
442 /// crossing a scheduling boundary. [begin, end) includes all instructions in
443 /// the region, including the boundary itself and single-instruction regions
444 /// that don't get scheduled.
445 void ScheduleDAGMI::enterRegion(MachineBasicBlock *bb,
446 MachineBasicBlock::iterator begin,
447 MachineBasicBlock::iterator end,
450 ScheduleDAGInstrs::enterRegion(bb, begin, end, endcount);
452 // For convenience remember the end of the liveness region.
454 (RegionEnd == bb->end()) ? RegionEnd : llvm::next(RegionEnd);
457 // Setup the register pressure trackers for the top scheduled top and bottom
458 // scheduled regions.
459 void ScheduleDAGMI::initRegPressure() {
460 TopRPTracker.init(&MF, RegClassInfo, LIS, BB, RegionBegin);
461 BotRPTracker.init(&MF, RegClassInfo, LIS, BB, LiveRegionEnd);
463 // Close the RPTracker to finalize live ins.
464 RPTracker.closeRegion();
466 DEBUG(RPTracker.getPressure().dump(TRI));
468 // Initialize the live ins and live outs.
469 TopRPTracker.addLiveRegs(RPTracker.getPressure().LiveInRegs);
470 BotRPTracker.addLiveRegs(RPTracker.getPressure().LiveOutRegs);
472 // Close one end of the tracker so we can call
473 // getMaxUpward/DownwardPressureDelta before advancing across any
474 // instructions. This converts currently live regs into live ins/outs.
475 TopRPTracker.closeTop();
476 BotRPTracker.closeBottom();
478 // Account for liveness generated by the region boundary.
479 if (LiveRegionEnd != RegionEnd)
480 BotRPTracker.recede();
482 assert(BotRPTracker.getPos() == RegionEnd && "Can't find the region bottom");
484 // Cache the list of excess pressure sets in this region. This will also track
485 // the max pressure in the scheduled code for these sets.
486 RegionCriticalPSets.clear();
487 const std::vector<unsigned> &RegionPressure =
488 RPTracker.getPressure().MaxSetPressure;
489 for (unsigned i = 0, e = RegionPressure.size(); i < e; ++i) {
490 unsigned Limit = TRI->getRegPressureSetLimit(i);
491 DEBUG(dbgs() << TRI->getRegPressureSetName(i)
493 << " Actual " << RegionPressure[i] << "\n");
494 if (RegionPressure[i] > Limit)
495 RegionCriticalPSets.push_back(PressureElement(i, 0));
497 DEBUG(dbgs() << "Excess PSets: ";
498 for (unsigned i = 0, e = RegionCriticalPSets.size(); i != e; ++i)
499 dbgs() << TRI->getRegPressureSetName(
500 RegionCriticalPSets[i].PSetID) << " ";
504 // FIXME: When the pressure tracker deals in pressure differences then we won't
505 // iterate over all RegionCriticalPSets[i].
507 updateScheduledPressure(const std::vector<unsigned> &NewMaxPressure) {
508 for (unsigned i = 0, e = RegionCriticalPSets.size(); i < e; ++i) {
509 unsigned ID = RegionCriticalPSets[i].PSetID;
510 int &MaxUnits = RegionCriticalPSets[i].UnitIncrease;
511 if ((int)NewMaxPressure[ID] > MaxUnits)
512 MaxUnits = NewMaxPressure[ID];
515 for (unsigned i = 0, e = NewMaxPressure.size(); i < e; ++i) {
516 unsigned Limit = TRI->getRegPressureSetLimit(i);
517 if (NewMaxPressure[i] > Limit ) {
518 dbgs() << " " << TRI->getRegPressureSetName(i) << ": "
519 << NewMaxPressure[i] << " > " << Limit << "\n";
524 /// schedule - Called back from MachineScheduler::runOnMachineFunction
525 /// after setting up the current scheduling region. [RegionBegin, RegionEnd)
526 /// only includes instructions that have DAG nodes, not scheduling boundaries.
528 /// This is a skeletal driver, with all the functionality pushed into helpers,
529 /// so that it can be easilly extended by experimental schedulers. Generally,
530 /// implementing MachineSchedStrategy should be sufficient to implement a new
531 /// scheduling algorithm. However, if a scheduler further subclasses
532 /// ScheduleDAGMI then it will want to override this virtual method in order to
533 /// update any specialized state.
534 void ScheduleDAGMI::schedule() {
535 buildDAGWithRegPressure();
537 Topo.InitDAGTopologicalSorting();
541 SmallVector<SUnit*, 8> TopRoots, BotRoots;
542 findRootsAndBiasEdges(TopRoots, BotRoots);
544 // Initialize the strategy before modifying the DAG.
545 // This may initialize a DFSResult to be used for queue priority.
546 SchedImpl->initialize(this);
548 DEBUG(for (unsigned su = 0, e = SUnits.size(); su != e; ++su)
549 SUnits[su].dumpAll(this));
550 if (ViewMISchedDAGs) viewGraph();
552 // Initialize ready queues now that the DAG and priority data are finalized.
553 initQueues(TopRoots, BotRoots);
555 bool IsTopNode = false;
556 while (SUnit *SU = SchedImpl->pickNode(IsTopNode)) {
557 assert(!SU->isScheduled && "Node already scheduled");
558 if (!checkSchedLimit())
561 scheduleMI(SU, IsTopNode);
563 updateQueues(SU, IsTopNode);
565 assert(CurrentTop == CurrentBottom && "Nonempty unscheduled zone.");
570 unsigned BBNum = begin()->getParent()->getNumber();
571 dbgs() << "*** Final schedule for BB#" << BBNum << " ***\n";
577 /// Build the DAG and setup three register pressure trackers.
578 void ScheduleDAGMI::buildDAGWithRegPressure() {
579 // Initialize the register pressure tracker used by buildSchedGraph.
580 RPTracker.init(&MF, RegClassInfo, LIS, BB, LiveRegionEnd);
582 // Account for liveness generate by the region boundary.
583 if (LiveRegionEnd != RegionEnd)
586 // Build the DAG, and compute current register pressure.
587 buildSchedGraph(AA, &RPTracker);
589 // Initialize top/bottom trackers after computing region pressure.
593 /// Apply each ScheduleDAGMutation step in order.
594 void ScheduleDAGMI::postprocessDAG() {
595 for (unsigned i = 0, e = Mutations.size(); i < e; ++i) {
596 Mutations[i]->apply(this);
600 void ScheduleDAGMI::computeDFSResult() {
602 DFSResult = new SchedDFSResult(/*BottomU*/true, MinSubtreeSize);
604 ScheduledTrees.clear();
605 DFSResult->resize(SUnits.size());
606 DFSResult->compute(SUnits);
607 ScheduledTrees.resize(DFSResult->getNumSubtrees());
610 void ScheduleDAGMI::findRootsAndBiasEdges(SmallVectorImpl<SUnit*> &TopRoots,
611 SmallVectorImpl<SUnit*> &BotRoots) {
612 for (std::vector<SUnit>::iterator
613 I = SUnits.begin(), E = SUnits.end(); I != E; ++I) {
615 assert(!SU->isBoundaryNode() && "Boundary node should not be in SUnits");
617 // Order predecessors so DFSResult follows the critical path.
618 SU->biasCriticalPath();
620 // A SUnit is ready to top schedule if it has no predecessors.
621 if (!I->NumPredsLeft)
622 TopRoots.push_back(SU);
623 // A SUnit is ready to bottom schedule if it has no successors.
624 if (!I->NumSuccsLeft)
625 BotRoots.push_back(SU);
627 ExitSU.biasCriticalPath();
630 /// Identify DAG roots and setup scheduler queues.
631 void ScheduleDAGMI::initQueues(ArrayRef<SUnit*> TopRoots,
632 ArrayRef<SUnit*> BotRoots) {
633 NextClusterSucc = NULL;
634 NextClusterPred = NULL;
636 // Release all DAG roots for scheduling, not including EntrySU/ExitSU.
638 // Nodes with unreleased weak edges can still be roots.
639 // Release top roots in forward order.
640 for (SmallVectorImpl<SUnit*>::const_iterator
641 I = TopRoots.begin(), E = TopRoots.end(); I != E; ++I) {
642 SchedImpl->releaseTopNode(*I);
644 // Release bottom roots in reverse order so the higher priority nodes appear
645 // first. This is more natural and slightly more efficient.
646 for (SmallVectorImpl<SUnit*>::const_reverse_iterator
647 I = BotRoots.rbegin(), E = BotRoots.rend(); I != E; ++I) {
648 SchedImpl->releaseBottomNode(*I);
651 releaseSuccessors(&EntrySU);
652 releasePredecessors(&ExitSU);
654 SchedImpl->registerRoots();
656 // Advance past initial DebugValues.
657 assert(TopRPTracker.getPos() == RegionBegin && "bad initial Top tracker");
658 CurrentTop = nextIfDebug(RegionBegin, RegionEnd);
659 TopRPTracker.setPos(CurrentTop);
661 CurrentBottom = RegionEnd;
664 /// Move an instruction and update register pressure.
665 void ScheduleDAGMI::scheduleMI(SUnit *SU, bool IsTopNode) {
666 // Move the instruction to its new location in the instruction stream.
667 MachineInstr *MI = SU->getInstr();
670 assert(SU->isTopReady() && "node still has unscheduled dependencies");
671 if (&*CurrentTop == MI)
672 CurrentTop = nextIfDebug(++CurrentTop, CurrentBottom);
674 moveInstruction(MI, CurrentTop);
675 TopRPTracker.setPos(MI);
678 // Update top scheduled pressure.
679 TopRPTracker.advance();
680 assert(TopRPTracker.getPos() == CurrentTop && "out of sync");
681 updateScheduledPressure(TopRPTracker.getPressure().MaxSetPressure);
684 assert(SU->isBottomReady() && "node still has unscheduled dependencies");
685 MachineBasicBlock::iterator priorII =
686 priorNonDebug(CurrentBottom, CurrentTop);
688 CurrentBottom = priorII;
690 if (&*CurrentTop == MI) {
691 CurrentTop = nextIfDebug(++CurrentTop, priorII);
692 TopRPTracker.setPos(CurrentTop);
694 moveInstruction(MI, CurrentBottom);
697 // Update bottom scheduled pressure.
698 BotRPTracker.recede();
699 assert(BotRPTracker.getPos() == CurrentBottom && "out of sync");
700 updateScheduledPressure(BotRPTracker.getPressure().MaxSetPressure);
704 /// Update scheduler queues after scheduling an instruction.
705 void ScheduleDAGMI::updateQueues(SUnit *SU, bool IsTopNode) {
706 // Release dependent instructions for scheduling.
708 releaseSuccessors(SU);
710 releasePredecessors(SU);
712 SU->isScheduled = true;
715 unsigned SubtreeID = DFSResult->getSubtreeID(SU);
716 if (!ScheduledTrees.test(SubtreeID)) {
717 ScheduledTrees.set(SubtreeID);
718 DFSResult->scheduleTree(SubtreeID);
719 SchedImpl->scheduleTree(SubtreeID);
723 // Notify the scheduling strategy after updating the DAG.
724 SchedImpl->schedNode(SU, IsTopNode);
727 /// Reinsert any remaining debug_values, just like the PostRA scheduler.
728 void ScheduleDAGMI::placeDebugValues() {
729 // If first instruction was a DBG_VALUE then put it back.
731 BB->splice(RegionBegin, BB, FirstDbgValue);
732 RegionBegin = FirstDbgValue;
735 for (std::vector<std::pair<MachineInstr *, MachineInstr *> >::iterator
736 DI = DbgValues.end(), DE = DbgValues.begin(); DI != DE; --DI) {
737 std::pair<MachineInstr *, MachineInstr *> P = *prior(DI);
738 MachineInstr *DbgValue = P.first;
739 MachineBasicBlock::iterator OrigPrevMI = P.second;
740 if (&*RegionBegin == DbgValue)
742 BB->splice(++OrigPrevMI, BB, DbgValue);
743 if (OrigPrevMI == llvm::prior(RegionEnd))
744 RegionEnd = DbgValue;
747 FirstDbgValue = NULL;
750 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
751 void ScheduleDAGMI::dumpSchedule() const {
752 for (MachineBasicBlock::iterator MI = begin(), ME = end(); MI != ME; ++MI) {
753 if (SUnit *SU = getSUnit(&(*MI)))
756 dbgs() << "Missing SUnit\n";
761 //===----------------------------------------------------------------------===//
762 // LoadClusterMutation - DAG post-processing to cluster loads.
763 //===----------------------------------------------------------------------===//
766 /// \brief Post-process the DAG to create cluster edges between neighboring
768 class LoadClusterMutation : public ScheduleDAGMutation {
773 LoadInfo(SUnit *su, unsigned reg, unsigned ofs)
774 : SU(su), BaseReg(reg), Offset(ofs) {}
776 static bool LoadInfoLess(const LoadClusterMutation::LoadInfo &LHS,
777 const LoadClusterMutation::LoadInfo &RHS);
779 const TargetInstrInfo *TII;
780 const TargetRegisterInfo *TRI;
782 LoadClusterMutation(const TargetInstrInfo *tii,
783 const TargetRegisterInfo *tri)
784 : TII(tii), TRI(tri) {}
786 virtual void apply(ScheduleDAGMI *DAG);
788 void clusterNeighboringLoads(ArrayRef<SUnit*> Loads, ScheduleDAGMI *DAG);
792 bool LoadClusterMutation::LoadInfoLess(
793 const LoadClusterMutation::LoadInfo &LHS,
794 const LoadClusterMutation::LoadInfo &RHS) {
795 if (LHS.BaseReg != RHS.BaseReg)
796 return LHS.BaseReg < RHS.BaseReg;
797 return LHS.Offset < RHS.Offset;
800 void LoadClusterMutation::clusterNeighboringLoads(ArrayRef<SUnit*> Loads,
801 ScheduleDAGMI *DAG) {
802 SmallVector<LoadClusterMutation::LoadInfo,32> LoadRecords;
803 for (unsigned Idx = 0, End = Loads.size(); Idx != End; ++Idx) {
804 SUnit *SU = Loads[Idx];
807 if (TII->getLdStBaseRegImmOfs(SU->getInstr(), BaseReg, Offset, TRI))
808 LoadRecords.push_back(LoadInfo(SU, BaseReg, Offset));
810 if (LoadRecords.size() < 2)
812 std::sort(LoadRecords.begin(), LoadRecords.end(), LoadInfoLess);
813 unsigned ClusterLength = 1;
814 for (unsigned Idx = 0, End = LoadRecords.size(); Idx < (End - 1); ++Idx) {
815 if (LoadRecords[Idx].BaseReg != LoadRecords[Idx+1].BaseReg) {
820 SUnit *SUa = LoadRecords[Idx].SU;
821 SUnit *SUb = LoadRecords[Idx+1].SU;
822 if (TII->shouldClusterLoads(SUa->getInstr(), SUb->getInstr(), ClusterLength)
823 && DAG->addEdge(SUb, SDep(SUa, SDep::Cluster))) {
825 DEBUG(dbgs() << "Cluster loads SU(" << SUa->NodeNum << ") - SU("
826 << SUb->NodeNum << ")\n");
827 // Copy successor edges from SUa to SUb. Interleaving computation
828 // dependent on SUa can prevent load combining due to register reuse.
829 // Predecessor edges do not need to be copied from SUb to SUa since nearby
830 // loads should have effectively the same inputs.
831 for (SUnit::const_succ_iterator
832 SI = SUa->Succs.begin(), SE = SUa->Succs.end(); SI != SE; ++SI) {
833 if (SI->getSUnit() == SUb)
835 DEBUG(dbgs() << " Copy Succ SU(" << SI->getSUnit()->NodeNum << ")\n");
836 DAG->addEdge(SI->getSUnit(), SDep(SUb, SDep::Artificial));
845 /// \brief Callback from DAG postProcessing to create cluster edges for loads.
846 void LoadClusterMutation::apply(ScheduleDAGMI *DAG) {
847 // Map DAG NodeNum to store chain ID.
848 DenseMap<unsigned, unsigned> StoreChainIDs;
849 // Map each store chain to a set of dependent loads.
850 SmallVector<SmallVector<SUnit*,4>, 32> StoreChainDependents;
851 for (unsigned Idx = 0, End = DAG->SUnits.size(); Idx != End; ++Idx) {
852 SUnit *SU = &DAG->SUnits[Idx];
853 if (!SU->getInstr()->mayLoad())
855 unsigned ChainPredID = DAG->SUnits.size();
856 for (SUnit::const_pred_iterator
857 PI = SU->Preds.begin(), PE = SU->Preds.end(); PI != PE; ++PI) {
859 ChainPredID = PI->getSUnit()->NodeNum;
863 // Check if this chain-like pred has been seen
864 // before. ChainPredID==MaxNodeID for loads at the top of the schedule.
865 unsigned NumChains = StoreChainDependents.size();
866 std::pair<DenseMap<unsigned, unsigned>::iterator, bool> Result =
867 StoreChainIDs.insert(std::make_pair(ChainPredID, NumChains));
869 StoreChainDependents.resize(NumChains + 1);
870 StoreChainDependents[Result.first->second].push_back(SU);
872 // Iterate over the store chains.
873 for (unsigned Idx = 0, End = StoreChainDependents.size(); Idx != End; ++Idx)
874 clusterNeighboringLoads(StoreChainDependents[Idx], DAG);
877 //===----------------------------------------------------------------------===//
878 // MacroFusion - DAG post-processing to encourage fusion of macro ops.
879 //===----------------------------------------------------------------------===//
882 /// \brief Post-process the DAG to create cluster edges between instructions
883 /// that may be fused by the processor into a single operation.
884 class MacroFusion : public ScheduleDAGMutation {
885 const TargetInstrInfo *TII;
887 MacroFusion(const TargetInstrInfo *tii): TII(tii) {}
889 virtual void apply(ScheduleDAGMI *DAG);
893 /// \brief Callback from DAG postProcessing to create cluster edges to encourage
894 /// fused operations.
895 void MacroFusion::apply(ScheduleDAGMI *DAG) {
896 // For now, assume targets can only fuse with the branch.
897 MachineInstr *Branch = DAG->ExitSU.getInstr();
901 for (unsigned Idx = DAG->SUnits.size(); Idx > 0;) {
902 SUnit *SU = &DAG->SUnits[--Idx];
903 if (!TII->shouldScheduleAdjacent(SU->getInstr(), Branch))
906 // Create a single weak edge from SU to ExitSU. The only effect is to cause
907 // bottom-up scheduling to heavily prioritize the clustered SU. There is no
908 // need to copy predecessor edges from ExitSU to SU, since top-down
909 // scheduling cannot prioritize ExitSU anyway. To defer top-down scheduling
910 // of SU, we could create an artificial edge from the deepest root, but it
911 // hasn't been needed yet.
912 bool Success = DAG->addEdge(&DAG->ExitSU, SDep(SU, SDep::Cluster));
914 assert(Success && "No DAG nodes should be reachable from ExitSU");
916 DEBUG(dbgs() << "Macro Fuse SU(" << SU->NodeNum << ")\n");
921 //===----------------------------------------------------------------------===//
922 // CopyConstrain - DAG post-processing to encourage copy elimination.
923 //===----------------------------------------------------------------------===//
926 /// \brief Post-process the DAG to create weak edges from all uses of a copy to
927 /// the one use that defines the copy's source vreg, most likely an induction
928 /// variable increment.
929 class CopyConstrain : public ScheduleDAGMutation {
931 SlotIndex RegionBeginIdx;
932 // RegionEndIdx is the slot index of the last non-debug instruction in the
933 // scheduling region. So we may have RegionBeginIdx == RegionEndIdx.
934 SlotIndex RegionEndIdx;
936 CopyConstrain(const TargetInstrInfo *, const TargetRegisterInfo *) {}
938 virtual void apply(ScheduleDAGMI *DAG);
941 void constrainLocalCopy(SUnit *CopySU, ScheduleDAGMI *DAG);
945 /// constrainLocalCopy handles two possibilities:
950 /// I3: dst = src (copy)
951 /// (create pred->succ edges I0->I1, I2->I1)
954 /// I0: dst = src (copy)
958 /// (create pred->succ edges I1->I2, I3->I2)
960 /// Although the MachineScheduler is currently constrained to single blocks,
961 /// this algorithm should handle extended blocks. An EBB is a set of
962 /// contiguously numbered blocks such that the previous block in the EBB is
963 /// always the single predecessor.
964 void CopyConstrain::constrainLocalCopy(SUnit *CopySU, ScheduleDAGMI *DAG) {
965 LiveIntervals *LIS = DAG->getLIS();
966 MachineInstr *Copy = CopySU->getInstr();
968 // Check for pure vreg copies.
969 unsigned SrcReg = Copy->getOperand(1).getReg();
970 if (!TargetRegisterInfo::isVirtualRegister(SrcReg))
973 unsigned DstReg = Copy->getOperand(0).getReg();
974 if (!TargetRegisterInfo::isVirtualRegister(DstReg))
977 // Check if either the dest or source is local. If it's live across a back
978 // edge, it's not local. Note that if both vregs are live across the back
979 // edge, we cannot successfully contrain the copy without cyclic scheduling.
980 unsigned LocalReg = DstReg;
981 unsigned GlobalReg = SrcReg;
982 LiveInterval *LocalLI = &LIS->getInterval(LocalReg);
983 if (!LocalLI->isLocal(RegionBeginIdx, RegionEndIdx)) {
986 LocalLI = &LIS->getInterval(LocalReg);
987 if (!LocalLI->isLocal(RegionBeginIdx, RegionEndIdx))
990 LiveInterval *GlobalLI = &LIS->getInterval(GlobalReg);
992 // Find the global segment after the start of the local LI.
993 LiveInterval::iterator GlobalSegment = GlobalLI->find(LocalLI->beginIndex());
994 // If GlobalLI does not overlap LocalLI->start, then a copy directly feeds a
995 // local live range. We could create edges from other global uses to the local
996 // start, but the coalescer should have already eliminated these cases, so
997 // don't bother dealing with it.
998 if (GlobalSegment == GlobalLI->end())
1001 // If GlobalSegment is killed at the LocalLI->start, the call to find()
1002 // returned the next global segment. But if GlobalSegment overlaps with
1003 // LocalLI->start, then advance to the next segement. If a hole in GlobalLI
1004 // exists in LocalLI's vicinity, GlobalSegment will be the end of the hole.
1005 if (GlobalSegment->contains(LocalLI->beginIndex()))
1008 if (GlobalSegment == GlobalLI->end())
1011 // Check if GlobalLI contains a hole in the vicinity of LocalLI.
1012 if (GlobalSegment != GlobalLI->begin()) {
1013 // Two address defs have no hole.
1014 if (SlotIndex::isSameInstr(llvm::prior(GlobalSegment)->end,
1015 GlobalSegment->start)) {
1018 // If GlobalLI has a prior segment, it must be live into the EBB. Otherwise
1019 // it would be a disconnected component in the live range.
1020 assert(llvm::prior(GlobalSegment)->start < LocalLI->beginIndex() &&
1021 "Disconnected LRG within the scheduling region.");
1023 MachineInstr *GlobalDef = LIS->getInstructionFromIndex(GlobalSegment->start);
1027 SUnit *GlobalSU = DAG->getSUnit(GlobalDef);
1031 // GlobalDef is the bottom of the GlobalLI hole. Open the hole by
1032 // constraining the uses of the last local def to precede GlobalDef.
1033 SmallVector<SUnit*,8> LocalUses;
1034 const VNInfo *LastLocalVN = LocalLI->getVNInfoBefore(LocalLI->endIndex());
1035 MachineInstr *LastLocalDef = LIS->getInstructionFromIndex(LastLocalVN->def);
1036 SUnit *LastLocalSU = DAG->getSUnit(LastLocalDef);
1037 for (SUnit::const_succ_iterator
1038 I = LastLocalSU->Succs.begin(), E = LastLocalSU->Succs.end();
1040 if (I->getKind() != SDep::Data || I->getReg() != LocalReg)
1042 if (I->getSUnit() == GlobalSU)
1044 if (!DAG->canAddEdge(GlobalSU, I->getSUnit()))
1046 LocalUses.push_back(I->getSUnit());
1048 // Open the top of the GlobalLI hole by constraining any earlier global uses
1049 // to precede the start of LocalLI.
1050 SmallVector<SUnit*,8> GlobalUses;
1051 MachineInstr *FirstLocalDef =
1052 LIS->getInstructionFromIndex(LocalLI->beginIndex());
1053 SUnit *FirstLocalSU = DAG->getSUnit(FirstLocalDef);
1054 for (SUnit::const_pred_iterator
1055 I = GlobalSU->Preds.begin(), E = GlobalSU->Preds.end(); I != E; ++I) {
1056 if (I->getKind() != SDep::Anti || I->getReg() != GlobalReg)
1058 if (I->getSUnit() == FirstLocalSU)
1060 if (!DAG->canAddEdge(FirstLocalSU, I->getSUnit()))
1062 GlobalUses.push_back(I->getSUnit());
1064 DEBUG(dbgs() << "Constraining copy SU(" << CopySU->NodeNum << ")\n");
1065 // Add the weak edges.
1066 for (SmallVectorImpl<SUnit*>::const_iterator
1067 I = LocalUses.begin(), E = LocalUses.end(); I != E; ++I) {
1068 DEBUG(dbgs() << " Local use SU(" << (*I)->NodeNum << ") -> SU("
1069 << GlobalSU->NodeNum << ")\n");
1070 DAG->addEdge(GlobalSU, SDep(*I, SDep::Weak));
1072 for (SmallVectorImpl<SUnit*>::const_iterator
1073 I = GlobalUses.begin(), E = GlobalUses.end(); I != E; ++I) {
1074 DEBUG(dbgs() << " Global use SU(" << (*I)->NodeNum << ") -> SU("
1075 << FirstLocalSU->NodeNum << ")\n");
1076 DAG->addEdge(FirstLocalSU, SDep(*I, SDep::Weak));
1080 /// \brief Callback from DAG postProcessing to create weak edges to encourage
1081 /// copy elimination.
1082 void CopyConstrain::apply(ScheduleDAGMI *DAG) {
1083 MachineBasicBlock::iterator FirstPos = nextIfDebug(DAG->begin(), DAG->end());
1084 if (FirstPos == DAG->end())
1086 RegionBeginIdx = DAG->getLIS()->getInstructionIndex(&*FirstPos);
1087 RegionEndIdx = DAG->getLIS()->getInstructionIndex(
1088 &*priorNonDebug(DAG->end(), DAG->begin()));
1090 for (unsigned Idx = 0, End = DAG->SUnits.size(); Idx != End; ++Idx) {
1091 SUnit *SU = &DAG->SUnits[Idx];
1092 if (!SU->getInstr()->isCopy())
1095 constrainLocalCopy(SU, DAG);
1099 //===----------------------------------------------------------------------===//
1100 // ConvergingScheduler - Implementation of the standard MachineSchedStrategy.
1101 //===----------------------------------------------------------------------===//
1104 /// ConvergingScheduler shrinks the unscheduled zone using heuristics to balance
1106 class ConvergingScheduler : public MachineSchedStrategy {
1108 /// Represent the type of SchedCandidate found within a single queue.
1109 /// pickNodeBidirectional depends on these listed by decreasing priority.
1111 NoCand, PhysRegCopy, SingleExcess, SingleCritical, Cluster, Weak,
1112 ResourceReduce, ResourceDemand, BotHeightReduce, BotPathReduce,
1113 TopDepthReduce, TopPathReduce, SingleMax, MultiPressure, NextDefUse,
1117 static const char *getReasonStr(ConvergingScheduler::CandReason Reason);
1120 /// Policy for scheduling the next instruction in the candidate's zone.
1123 unsigned ReduceResIdx;
1124 unsigned DemandResIdx;
1126 CandPolicy(): ReduceLatency(false), ReduceResIdx(0), DemandResIdx(0) {}
1129 /// Status of an instruction's critical resource consumption.
1130 struct SchedResourceDelta {
1131 // Count critical resources in the scheduled region required by SU.
1132 unsigned CritResources;
1134 // Count critical resources from another region consumed by SU.
1135 unsigned DemandedResources;
1137 SchedResourceDelta(): CritResources(0), DemandedResources(0) {}
1139 bool operator==(const SchedResourceDelta &RHS) const {
1140 return CritResources == RHS.CritResources
1141 && DemandedResources == RHS.DemandedResources;
1143 bool operator!=(const SchedResourceDelta &RHS) const {
1144 return !operator==(RHS);
1148 /// Store the state used by ConvergingScheduler heuristics, required for the
1149 /// lifetime of one invocation of pickNode().
1150 struct SchedCandidate {
1153 // The best SUnit candidate.
1156 // The reason for this candidate.
1159 // Register pressure values for the best candidate.
1160 RegPressureDelta RPDelta;
1162 // Critical resource consumption of the best candidate.
1163 SchedResourceDelta ResDelta;
1165 SchedCandidate(const CandPolicy &policy)
1166 : Policy(policy), SU(NULL), Reason(NoCand) {}
1168 bool isValid() const { return SU; }
1170 // Copy the status of another candidate without changing policy.
1171 void setBest(SchedCandidate &Best) {
1172 assert(Best.Reason != NoCand && "uninitialized Sched candidate");
1174 Reason = Best.Reason;
1175 RPDelta = Best.RPDelta;
1176 ResDelta = Best.ResDelta;
1179 void initResourceDelta(const ScheduleDAGMI *DAG,
1180 const TargetSchedModel *SchedModel);
1183 /// Summarize the unscheduled region.
1184 struct SchedRemainder {
1185 // Critical path through the DAG in expected latency.
1186 unsigned CriticalPath;
1188 // Unscheduled resources
1189 SmallVector<unsigned, 16> RemainingCounts;
1190 // Critical resource for the unscheduled zone.
1191 unsigned CritResIdx;
1192 // Number of micro-ops left to schedule.
1193 unsigned RemainingMicroOps;
1197 RemainingCounts.clear();
1199 RemainingMicroOps = 0;
1202 SchedRemainder() { reset(); }
1204 void init(ScheduleDAGMI *DAG, const TargetSchedModel *SchedModel);
1206 unsigned getMaxRemainingCount(const TargetSchedModel *SchedModel) const {
1207 if (!SchedModel->hasInstrSchedModel())
1211 RemainingMicroOps * SchedModel->getMicroOpFactor(),
1212 RemainingCounts[CritResIdx]);
1216 /// Each Scheduling boundary is associated with ready queues. It tracks the
1217 /// current cycle in the direction of movement, and maintains the state
1218 /// of "hazards" and other interlocks at the current cycle.
1219 struct SchedBoundary {
1221 const TargetSchedModel *SchedModel;
1222 SchedRemainder *Rem;
1224 ReadyQueue Available;
1228 // For heuristics, keep a list of the nodes that immediately depend on the
1229 // most recently scheduled node.
1230 SmallPtrSet<const SUnit*, 8> NextSUs;
1232 ScheduleHazardRecognizer *HazardRec;
1235 unsigned IssueCount;
1237 /// MinReadyCycle - Cycle of the soonest available instruction.
1238 unsigned MinReadyCycle;
1240 // The expected latency of the critical path in this scheduled zone.
1241 unsigned ExpectedLatency;
1243 // The latency of dependence chains leading into this zone.
1244 // For each node scheduled: DLat = max DLat, N.Depth.
1245 // For each cycle scheduled: DLat -= 1.
1246 unsigned DependentLatency;
1248 // Resources used in the scheduled zone beyond this boundary.
1249 SmallVector<unsigned, 16> ResourceCounts;
1251 // Cache the critical resources ID in this scheduled zone.
1252 unsigned CritResIdx;
1254 // Is the scheduled region resource limited vs. latency limited.
1255 bool IsResourceLimited;
1257 unsigned ExpectedCount;
1260 // Remember the greatest min operand latency.
1261 unsigned MaxMinLatency;
1265 // A new HazardRec is created for each DAG and owned by SchedBoundary.
1270 CheckPending = false;
1275 MinReadyCycle = UINT_MAX;
1276 ExpectedLatency = 0;
1277 DependentLatency = 0;
1278 ResourceCounts.resize(1);
1279 assert(!ResourceCounts[0] && "nonzero count for bad resource");
1281 IsResourceLimited = false;
1286 // Reserve a zero-count for invalid CritResIdx.
1287 ResourceCounts.resize(1);
1290 /// Pending queues extend the ready queues with the same ID and the
1291 /// PendingFlag set.
1292 SchedBoundary(unsigned ID, const Twine &Name):
1293 DAG(0), SchedModel(0), Rem(0), Available(ID, Name+".A"),
1294 Pending(ID << ConvergingScheduler::LogMaxQID, Name+".P"),
1299 ~SchedBoundary() { delete HazardRec; }
1301 void init(ScheduleDAGMI *dag, const TargetSchedModel *smodel,
1302 SchedRemainder *rem);
1304 bool isTop() const {
1305 return Available.getID() == ConvergingScheduler::TopQID;
1308 unsigned getUnscheduledLatency(SUnit *SU) const {
1310 return SU->getHeight();
1311 return SU->getDepth() + SU->Latency;
1314 unsigned getCriticalCount() const {
1315 return ResourceCounts[CritResIdx];
1318 bool checkHazard(SUnit *SU);
1320 void setLatencyPolicy(CandPolicy &Policy);
1322 void releaseNode(SUnit *SU, unsigned ReadyCycle);
1326 void countResource(unsigned PIdx, unsigned Cycles);
1328 void bumpNode(SUnit *SU);
1330 void releasePending();
1332 void removeReady(SUnit *SU);
1334 SUnit *pickOnlyChoice();
1339 const TargetSchedModel *SchedModel;
1340 const TargetRegisterInfo *TRI;
1342 // State of the top and bottom scheduled instruction boundaries.
1348 /// SUnit::NodeQueueId: 0 (none), 1 (top), 2 (bot), 3 (both)
1355 ConvergingScheduler():
1356 DAG(0), SchedModel(0), TRI(0), Top(TopQID, "TopQ"), Bot(BotQID, "BotQ") {}
1358 virtual void initialize(ScheduleDAGMI *dag);
1360 virtual SUnit *pickNode(bool &IsTopNode);
1362 virtual void schedNode(SUnit *SU, bool IsTopNode);
1364 virtual void releaseTopNode(SUnit *SU);
1366 virtual void releaseBottomNode(SUnit *SU);
1368 virtual void registerRoots();
1372 ConvergingScheduler::SchedBoundary &CriticalZone,
1373 ConvergingScheduler::SchedCandidate &CriticalCand,
1374 ConvergingScheduler::SchedBoundary &OppositeZone,
1375 ConvergingScheduler::SchedCandidate &OppositeCand);
1377 void checkResourceLimits(ConvergingScheduler::SchedCandidate &TopCand,
1378 ConvergingScheduler::SchedCandidate &BotCand);
1380 void tryCandidate(SchedCandidate &Cand,
1381 SchedCandidate &TryCand,
1382 SchedBoundary &Zone,
1383 const RegPressureTracker &RPTracker,
1384 RegPressureTracker &TempTracker);
1386 SUnit *pickNodeBidirectional(bool &IsTopNode);
1388 void pickNodeFromQueue(SchedBoundary &Zone,
1389 const RegPressureTracker &RPTracker,
1390 SchedCandidate &Candidate);
1392 void reschedulePhysRegCopies(SUnit *SU, bool isTop);
1395 void traceCandidate(const SchedCandidate &Cand);
1400 void ConvergingScheduler::SchedRemainder::
1401 init(ScheduleDAGMI *DAG, const TargetSchedModel *SchedModel) {
1403 if (!SchedModel->hasInstrSchedModel())
1405 RemainingCounts.resize(SchedModel->getNumProcResourceKinds());
1406 for (std::vector<SUnit>::iterator
1407 I = DAG->SUnits.begin(), E = DAG->SUnits.end(); I != E; ++I) {
1408 const MCSchedClassDesc *SC = DAG->getSchedClass(&*I);
1409 RemainingMicroOps += SchedModel->getNumMicroOps(I->getInstr(), SC);
1410 for (TargetSchedModel::ProcResIter
1411 PI = SchedModel->getWriteProcResBegin(SC),
1412 PE = SchedModel->getWriteProcResEnd(SC); PI != PE; ++PI) {
1413 unsigned PIdx = PI->ProcResourceIdx;
1414 unsigned Factor = SchedModel->getResourceFactor(PIdx);
1415 RemainingCounts[PIdx] += (Factor * PI->Cycles);
1418 for (unsigned PIdx = 0, PEnd = SchedModel->getNumProcResourceKinds();
1419 PIdx != PEnd; ++PIdx) {
1420 if ((int)(RemainingCounts[PIdx] - RemainingCounts[CritResIdx])
1421 >= (int)SchedModel->getLatencyFactor()) {
1425 DEBUG(dbgs() << "Critical Resource: "
1426 << SchedModel->getProcResource(CritResIdx)->Name
1427 << ": " << RemainingCounts[CritResIdx]
1428 << " / " << SchedModel->getLatencyFactor() << '\n');
1431 void ConvergingScheduler::SchedBoundary::
1432 init(ScheduleDAGMI *dag, const TargetSchedModel *smodel, SchedRemainder *rem) {
1435 SchedModel = smodel;
1437 if (SchedModel->hasInstrSchedModel())
1438 ResourceCounts.resize(SchedModel->getNumProcResourceKinds());
1441 void ConvergingScheduler::initialize(ScheduleDAGMI *dag) {
1443 SchedModel = DAG->getSchedModel();
1446 Rem.init(DAG, SchedModel);
1447 Top.init(DAG, SchedModel, &Rem);
1448 Bot.init(DAG, SchedModel, &Rem);
1450 // Initialize resource counts.
1452 // Initialize the HazardRecognizers. If itineraries don't exist, are empty, or
1453 // are disabled, then these HazardRecs will be disabled.
1454 const InstrItineraryData *Itin = SchedModel->getInstrItineraries();
1455 const TargetMachine &TM = DAG->MF.getTarget();
1456 Top.HazardRec = TM.getInstrInfo()->CreateTargetMIHazardRecognizer(Itin, DAG);
1457 Bot.HazardRec = TM.getInstrInfo()->CreateTargetMIHazardRecognizer(Itin, DAG);
1459 assert((!ForceTopDown || !ForceBottomUp) &&
1460 "-misched-topdown incompatible with -misched-bottomup");
1463 void ConvergingScheduler::releaseTopNode(SUnit *SU) {
1464 if (SU->isScheduled)
1467 for (SUnit::pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
1469 unsigned PredReadyCycle = I->getSUnit()->TopReadyCycle;
1470 unsigned MinLatency = I->getMinLatency();
1472 Top.MaxMinLatency = std::max(MinLatency, Top.MaxMinLatency);
1474 if (SU->TopReadyCycle < PredReadyCycle + MinLatency)
1475 SU->TopReadyCycle = PredReadyCycle + MinLatency;
1477 Top.releaseNode(SU, SU->TopReadyCycle);
1480 void ConvergingScheduler::releaseBottomNode(SUnit *SU) {
1481 if (SU->isScheduled)
1484 assert(SU->getInstr() && "Scheduled SUnit must have instr");
1486 for (SUnit::succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
1490 unsigned SuccReadyCycle = I->getSUnit()->BotReadyCycle;
1491 unsigned MinLatency = I->getMinLatency();
1493 Bot.MaxMinLatency = std::max(MinLatency, Bot.MaxMinLatency);
1495 if (SU->BotReadyCycle < SuccReadyCycle + MinLatency)
1496 SU->BotReadyCycle = SuccReadyCycle + MinLatency;
1498 Bot.releaseNode(SU, SU->BotReadyCycle);
1501 void ConvergingScheduler::registerRoots() {
1502 Rem.CriticalPath = DAG->ExitSU.getDepth();
1503 // Some roots may not feed into ExitSU. Check all of them in case.
1504 for (std::vector<SUnit*>::const_iterator
1505 I = Bot.Available.begin(), E = Bot.Available.end(); I != E; ++I) {
1506 if ((*I)->getDepth() > Rem.CriticalPath)
1507 Rem.CriticalPath = (*I)->getDepth();
1509 DEBUG(dbgs() << "Critical Path: " << Rem.CriticalPath << '\n');
1512 /// Does this SU have a hazard within the current instruction group.
1514 /// The scheduler supports two modes of hazard recognition. The first is the
1515 /// ScheduleHazardRecognizer API. It is a fully general hazard recognizer that
1516 /// supports highly complicated in-order reservation tables
1517 /// (ScoreboardHazardRecognizer) and arbitraty target-specific logic.
1519 /// The second is a streamlined mechanism that checks for hazards based on
1520 /// simple counters that the scheduler itself maintains. It explicitly checks
1521 /// for instruction dispatch limitations, including the number of micro-ops that
1522 /// can dispatch per cycle.
1524 /// TODO: Also check whether the SU must start a new group.
1525 bool ConvergingScheduler::SchedBoundary::checkHazard(SUnit *SU) {
1526 if (HazardRec->isEnabled())
1527 return HazardRec->getHazardType(SU) != ScheduleHazardRecognizer::NoHazard;
1529 unsigned uops = SchedModel->getNumMicroOps(SU->getInstr());
1530 if ((IssueCount > 0) && (IssueCount + uops > SchedModel->getIssueWidth())) {
1531 DEBUG(dbgs() << " SU(" << SU->NodeNum << ") uops="
1532 << SchedModel->getNumMicroOps(SU->getInstr()) << '\n');
1538 /// Compute the remaining latency to determine whether ILP should be increased.
1539 void ConvergingScheduler::SchedBoundary::setLatencyPolicy(CandPolicy &Policy) {
1540 DEBUG(dbgs() << " " << Available.getName()
1541 << " DependentLatency " << DependentLatency << '\n');
1543 // FIXME: compile time. In all, we visit four queues here one we should only
1544 // need to visit the one that was last popped if we cache the result.
1545 unsigned RemLatency = DependentLatency;
1546 for (ReadyQueue::iterator I = Available.begin(), E = Available.end();
1548 unsigned L = getUnscheduledLatency(*I);
1549 if (L > RemLatency) {
1550 DEBUG(dbgs() << " " << Available.getName()
1551 << " RemLatency SU(" << (*I)->NodeNum << ") " << L << '\n');
1555 for (ReadyQueue::iterator I = Pending.begin(), E = Pending.end();
1557 unsigned L = getUnscheduledLatency(*I);
1561 unsigned CriticalPathLimit = Rem->CriticalPath + SchedModel->getILPWindow();
1562 DEBUG(dbgs() << " " << Available.getName()
1563 << " ExpectedLatency " << ExpectedLatency
1564 << " CP Limit " << CriticalPathLimit << '\n');
1566 if (RemLatency + std::max(ExpectedLatency, CurrCycle) >= CriticalPathLimit
1567 && RemLatency > Rem->getMaxRemainingCount(SchedModel)) {
1568 Policy.ReduceLatency = true;
1569 DEBUG(dbgs() << " Increase ILP: " << Available.getName() << '\n');
1573 void ConvergingScheduler::SchedBoundary::releaseNode(SUnit *SU,
1574 unsigned ReadyCycle) {
1576 if (ReadyCycle < MinReadyCycle)
1577 MinReadyCycle = ReadyCycle;
1579 // Check for interlocks first. For the purpose of other heuristics, an
1580 // instruction that cannot issue appears as if it's not in the ReadyQueue.
1581 if (ReadyCycle > CurrCycle || checkHazard(SU))
1586 // Record this node as an immediate dependent of the scheduled node.
1590 /// Move the boundary of scheduled code by one cycle.
1591 void ConvergingScheduler::SchedBoundary::bumpCycle() {
1592 unsigned Width = SchedModel->getIssueWidth();
1593 IssueCount = (IssueCount <= Width) ? 0 : IssueCount - Width;
1595 unsigned NextCycle = CurrCycle + 1;
1596 assert(MinReadyCycle < UINT_MAX && "MinReadyCycle uninitialized");
1597 if (MinReadyCycle > NextCycle) {
1599 NextCycle = MinReadyCycle;
1601 if ((NextCycle - CurrCycle) > DependentLatency)
1602 DependentLatency = 0;
1604 DependentLatency -= (NextCycle - CurrCycle);
1606 if (!HazardRec->isEnabled()) {
1607 // Bypass HazardRec virtual calls.
1608 CurrCycle = NextCycle;
1611 // Bypass getHazardType calls in case of long latency.
1612 for (; CurrCycle != NextCycle; ++CurrCycle) {
1614 HazardRec->AdvanceCycle();
1616 HazardRec->RecedeCycle();
1619 CheckPending = true;
1620 IsResourceLimited = getCriticalCount() > std::max(ExpectedLatency, CurrCycle);
1622 DEBUG(dbgs() << " " << Available.getName()
1623 << " Cycle: " << CurrCycle << '\n');
1626 /// Add the given processor resource to this scheduled zone.
1627 void ConvergingScheduler::SchedBoundary::countResource(unsigned PIdx,
1629 unsigned Factor = SchedModel->getResourceFactor(PIdx);
1630 DEBUG(dbgs() << " " << SchedModel->getProcResource(PIdx)->Name
1631 << " +(" << Cycles << "x" << Factor
1632 << ") / " << SchedModel->getLatencyFactor() << '\n');
1634 unsigned Count = Factor * Cycles;
1635 ResourceCounts[PIdx] += Count;
1636 assert(Rem->RemainingCounts[PIdx] >= Count && "resource double counted");
1637 Rem->RemainingCounts[PIdx] -= Count;
1639 // Check if this resource exceeds the current critical resource by a full
1640 // cycle. If so, it becomes the critical resource.
1641 if ((int)(ResourceCounts[PIdx] - ResourceCounts[CritResIdx])
1642 >= (int)SchedModel->getLatencyFactor()) {
1644 DEBUG(dbgs() << " *** Critical resource "
1645 << SchedModel->getProcResource(PIdx)->Name << " x"
1646 << ResourceCounts[PIdx] << '\n');
1650 /// Move the boundary of scheduled code by one SUnit.
1651 void ConvergingScheduler::SchedBoundary::bumpNode(SUnit *SU) {
1652 // Update the reservation table.
1653 if (HazardRec->isEnabled()) {
1654 if (!isTop() && SU->isCall) {
1655 // Calls are scheduled with their preceding instructions. For bottom-up
1656 // scheduling, clear the pipeline state before emitting.
1659 HazardRec->EmitInstruction(SU);
1661 // Update resource counts and critical resource.
1662 if (SchedModel->hasInstrSchedModel()) {
1663 const MCSchedClassDesc *SC = DAG->getSchedClass(SU);
1664 Rem->RemainingMicroOps -= SchedModel->getNumMicroOps(SU->getInstr(), SC);
1665 for (TargetSchedModel::ProcResIter
1666 PI = SchedModel->getWriteProcResBegin(SC),
1667 PE = SchedModel->getWriteProcResEnd(SC); PI != PE; ++PI) {
1668 countResource(PI->ProcResourceIdx, PI->Cycles);
1671 unsigned &TopLatency = isTop() ? ExpectedLatency : DependentLatency;
1672 unsigned &BotLatency = isTop() ? DependentLatency : ExpectedLatency;
1673 if (SU->getDepth() > TopLatency)
1674 TopLatency = SU->getDepth();
1675 if (SU->getHeight() > BotLatency)
1676 BotLatency = SU->getHeight();
1678 IsResourceLimited = getCriticalCount() > std::max(ExpectedLatency, CurrCycle);
1680 // Check the instruction group dispatch limit.
1681 // TODO: Check if this SU must end a dispatch group.
1682 IssueCount += SchedModel->getNumMicroOps(SU->getInstr());
1684 // checkHazard prevents scheduling multiple instructions per cycle that exceed
1685 // issue width. However, we commonly reach the maximum. In this case
1686 // opportunistically bump the cycle to avoid uselessly checking everything in
1687 // the readyQ. Furthermore, a single instruction may produce more than one
1688 // cycle's worth of micro-ops.
1689 if (IssueCount >= SchedModel->getIssueWidth()) {
1690 DEBUG(dbgs() << " *** Max instrs at cycle " << CurrCycle << '\n');
1695 /// Release pending ready nodes in to the available queue. This makes them
1696 /// visible to heuristics.
1697 void ConvergingScheduler::SchedBoundary::releasePending() {
1698 // If the available queue is empty, it is safe to reset MinReadyCycle.
1699 if (Available.empty())
1700 MinReadyCycle = UINT_MAX;
1702 // Check to see if any of the pending instructions are ready to issue. If
1703 // so, add them to the available queue.
1704 for (unsigned i = 0, e = Pending.size(); i != e; ++i) {
1705 SUnit *SU = *(Pending.begin()+i);
1706 unsigned ReadyCycle = isTop() ? SU->TopReadyCycle : SU->BotReadyCycle;
1708 if (ReadyCycle < MinReadyCycle)
1709 MinReadyCycle = ReadyCycle;
1711 if (ReadyCycle > CurrCycle)
1714 if (checkHazard(SU))
1718 Pending.remove(Pending.begin()+i);
1721 DEBUG(if (!Pending.empty()) Pending.dump());
1722 CheckPending = false;
1725 /// Remove SU from the ready set for this boundary.
1726 void ConvergingScheduler::SchedBoundary::removeReady(SUnit *SU) {
1727 if (Available.isInQueue(SU))
1728 Available.remove(Available.find(SU));
1730 assert(Pending.isInQueue(SU) && "bad ready count");
1731 Pending.remove(Pending.find(SU));
1735 /// If this queue only has one ready candidate, return it. As a side effect,
1736 /// defer any nodes that now hit a hazard, and advance the cycle until at least
1737 /// one node is ready. If multiple instructions are ready, return NULL.
1738 SUnit *ConvergingScheduler::SchedBoundary::pickOnlyChoice() {
1742 if (IssueCount > 0) {
1743 // Defer any ready instrs that now have a hazard.
1744 for (ReadyQueue::iterator I = Available.begin(); I != Available.end();) {
1745 if (checkHazard(*I)) {
1747 I = Available.remove(I);
1753 for (unsigned i = 0; Available.empty(); ++i) {
1754 assert(i <= (HazardRec->getMaxLookAhead() + MaxMinLatency) &&
1755 "permanent hazard"); (void)i;
1759 if (Available.size() == 1)
1760 return *Available.begin();
1764 /// Record the candidate policy for opposite zones with different critical
1767 /// If the CriticalZone is latency limited, don't force a policy for the
1768 /// candidates here. Instead, setLatencyPolicy sets ReduceLatency if needed.
1769 void ConvergingScheduler::balanceZones(
1770 ConvergingScheduler::SchedBoundary &CriticalZone,
1771 ConvergingScheduler::SchedCandidate &CriticalCand,
1772 ConvergingScheduler::SchedBoundary &OppositeZone,
1773 ConvergingScheduler::SchedCandidate &OppositeCand) {
1775 if (!CriticalZone.IsResourceLimited)
1777 assert(SchedModel->hasInstrSchedModel() && "required schedmodel");
1779 SchedRemainder *Rem = CriticalZone.Rem;
1781 // If the critical zone is overconsuming a resource relative to the
1782 // remainder, try to reduce it.
1783 unsigned RemainingCritCount =
1784 Rem->RemainingCounts[CriticalZone.CritResIdx];
1785 if ((int)(Rem->getMaxRemainingCount(SchedModel) - RemainingCritCount)
1786 > (int)SchedModel->getLatencyFactor()) {
1787 CriticalCand.Policy.ReduceResIdx = CriticalZone.CritResIdx;
1788 DEBUG(dbgs() << " Balance " << CriticalZone.Available.getName()
1790 << SchedModel->getProcResource(CriticalZone.CritResIdx)->Name
1793 // If the other zone is underconsuming a resource relative to the full zone,
1794 // try to increase it.
1795 unsigned OppositeCount =
1796 OppositeZone.ResourceCounts[CriticalZone.CritResIdx];
1797 if ((int)(OppositeZone.ExpectedCount - OppositeCount)
1798 > (int)SchedModel->getLatencyFactor()) {
1799 OppositeCand.Policy.DemandResIdx = CriticalZone.CritResIdx;
1800 DEBUG(dbgs() << " Balance " << OppositeZone.Available.getName()
1802 << SchedModel->getProcResource(OppositeZone.CritResIdx)->Name
1807 /// Determine if the scheduled zones exceed resource limits or critical path and
1808 /// set each candidate's ReduceHeight policy accordingly.
1809 void ConvergingScheduler::checkResourceLimits(
1810 ConvergingScheduler::SchedCandidate &TopCand,
1811 ConvergingScheduler::SchedCandidate &BotCand) {
1813 // Set ReduceLatency to true if needed.
1814 Bot.setLatencyPolicy(BotCand.Policy);
1815 Top.setLatencyPolicy(TopCand.Policy);
1817 // Handle resource-limited regions.
1818 if (Top.IsResourceLimited && Bot.IsResourceLimited
1819 && Top.CritResIdx == Bot.CritResIdx) {
1820 // If the scheduled critical resource in both zones is no longer the
1821 // critical remaining resource, attempt to reduce resource height both ways.
1822 if (Top.CritResIdx != Rem.CritResIdx) {
1823 TopCand.Policy.ReduceResIdx = Top.CritResIdx;
1824 BotCand.Policy.ReduceResIdx = Bot.CritResIdx;
1825 DEBUG(dbgs() << " Reduce scheduled "
1826 << SchedModel->getProcResource(Top.CritResIdx)->Name << '\n');
1830 // Handle latency-limited regions.
1831 if (!Top.IsResourceLimited && !Bot.IsResourceLimited) {
1832 // If the total scheduled expected latency exceeds the region's critical
1833 // path then reduce latency both ways.
1835 // Just because a zone is not resource limited does not mean it is latency
1836 // limited. Unbuffered resource, such as max micro-ops may cause CurrCycle
1837 // to exceed expected latency.
1838 if ((Top.ExpectedLatency + Bot.ExpectedLatency >= Rem.CriticalPath)
1839 && (Rem.CriticalPath > Top.CurrCycle + Bot.CurrCycle)) {
1840 TopCand.Policy.ReduceLatency = true;
1841 BotCand.Policy.ReduceLatency = true;
1842 DEBUG(dbgs() << " Reduce scheduled latency " << Top.ExpectedLatency
1843 << " + " << Bot.ExpectedLatency << '\n');
1847 // The critical resource is different in each zone, so request balancing.
1849 // Compute the cost of each zone.
1850 Top.ExpectedCount = std::max(Top.ExpectedLatency, Top.CurrCycle);
1851 Top.ExpectedCount = std::max(
1852 Top.getCriticalCount(),
1853 Top.ExpectedCount * SchedModel->getLatencyFactor());
1854 Bot.ExpectedCount = std::max(Bot.ExpectedLatency, Bot.CurrCycle);
1855 Bot.ExpectedCount = std::max(
1856 Bot.getCriticalCount(),
1857 Bot.ExpectedCount * SchedModel->getLatencyFactor());
1859 balanceZones(Top, TopCand, Bot, BotCand);
1860 balanceZones(Bot, BotCand, Top, TopCand);
1863 void ConvergingScheduler::SchedCandidate::
1864 initResourceDelta(const ScheduleDAGMI *DAG,
1865 const TargetSchedModel *SchedModel) {
1866 if (!Policy.ReduceResIdx && !Policy.DemandResIdx)
1869 const MCSchedClassDesc *SC = DAG->getSchedClass(SU);
1870 for (TargetSchedModel::ProcResIter
1871 PI = SchedModel->getWriteProcResBegin(SC),
1872 PE = SchedModel->getWriteProcResEnd(SC); PI != PE; ++PI) {
1873 if (PI->ProcResourceIdx == Policy.ReduceResIdx)
1874 ResDelta.CritResources += PI->Cycles;
1875 if (PI->ProcResourceIdx == Policy.DemandResIdx)
1876 ResDelta.DemandedResources += PI->Cycles;
1880 /// Return true if this heuristic determines order.
1881 static bool tryLess(int TryVal, int CandVal,
1882 ConvergingScheduler::SchedCandidate &TryCand,
1883 ConvergingScheduler::SchedCandidate &Cand,
1884 ConvergingScheduler::CandReason Reason) {
1885 if (TryVal < CandVal) {
1886 TryCand.Reason = Reason;
1889 if (TryVal > CandVal) {
1890 if (Cand.Reason > Reason)
1891 Cand.Reason = Reason;
1897 static bool tryGreater(int TryVal, int CandVal,
1898 ConvergingScheduler::SchedCandidate &TryCand,
1899 ConvergingScheduler::SchedCandidate &Cand,
1900 ConvergingScheduler::CandReason Reason) {
1901 if (TryVal > CandVal) {
1902 TryCand.Reason = Reason;
1905 if (TryVal < CandVal) {
1906 if (Cand.Reason > Reason)
1907 Cand.Reason = Reason;
1913 static unsigned getWeakLeft(const SUnit *SU, bool isTop) {
1914 return (isTop) ? SU->WeakPredsLeft : SU->WeakSuccsLeft;
1917 /// Minimize physical register live ranges. Regalloc wants them adjacent to
1918 /// their physreg def/use.
1920 /// FIXME: This is an unnecessary check on the critical path. Most are root/leaf
1921 /// copies which can be prescheduled. The rest (e.g. x86 MUL) could be bundled
1922 /// with the operation that produces or consumes the physreg. We'll do this when
1923 /// regalloc has support for parallel copies.
1924 static int biasPhysRegCopy(const SUnit *SU, bool isTop) {
1925 const MachineInstr *MI = SU->getInstr();
1929 unsigned ScheduledOper = isTop ? 1 : 0;
1930 unsigned UnscheduledOper = isTop ? 0 : 1;
1931 // If we have already scheduled the physreg produce/consumer, immediately
1932 // schedule the copy.
1933 if (TargetRegisterInfo::isPhysicalRegister(
1934 MI->getOperand(ScheduledOper).getReg()))
1936 // If the physreg is at the boundary, defer it. Otherwise schedule it
1937 // immediately to free the dependent. We can hoist the copy later.
1938 bool AtBoundary = isTop ? !SU->NumSuccsLeft : !SU->NumPredsLeft;
1939 if (TargetRegisterInfo::isPhysicalRegister(
1940 MI->getOperand(UnscheduledOper).getReg()))
1941 return AtBoundary ? -1 : 1;
1945 /// Apply a set of heursitics to a new candidate. Heuristics are currently
1946 /// hierarchical. This may be more efficient than a graduated cost model because
1947 /// we don't need to evaluate all aspects of the model for each node in the
1948 /// queue. But it's really done to make the heuristics easier to debug and
1949 /// statistically analyze.
1951 /// \param Cand provides the policy and current best candidate.
1952 /// \param TryCand refers to the next SUnit candidate, otherwise uninitialized.
1953 /// \param Zone describes the scheduled zone that we are extending.
1954 /// \param RPTracker describes reg pressure within the scheduled zone.
1955 /// \param TempTracker is a scratch pressure tracker to reuse in queries.
1956 void ConvergingScheduler::tryCandidate(SchedCandidate &Cand,
1957 SchedCandidate &TryCand,
1958 SchedBoundary &Zone,
1959 const RegPressureTracker &RPTracker,
1960 RegPressureTracker &TempTracker) {
1962 // Always initialize TryCand's RPDelta.
1963 TempTracker.getMaxPressureDelta(TryCand.SU->getInstr(), TryCand.RPDelta,
1964 DAG->getRegionCriticalPSets(),
1965 DAG->getRegPressure().MaxSetPressure);
1967 // Initialize the candidate if needed.
1968 if (!Cand.isValid()) {
1969 TryCand.Reason = NodeOrder;
1973 if (tryGreater(biasPhysRegCopy(TryCand.SU, Zone.isTop()),
1974 biasPhysRegCopy(Cand.SU, Zone.isTop()),
1975 TryCand, Cand, PhysRegCopy))
1978 // Avoid exceeding the target's limit.
1979 if (tryLess(TryCand.RPDelta.Excess.UnitIncrease,
1980 Cand.RPDelta.Excess.UnitIncrease, TryCand, Cand, SingleExcess))
1982 if (Cand.Reason == SingleExcess)
1983 Cand.Reason = MultiPressure;
1985 // Avoid increasing the max critical pressure in the scheduled region.
1986 if (tryLess(TryCand.RPDelta.CriticalMax.UnitIncrease,
1987 Cand.RPDelta.CriticalMax.UnitIncrease,
1988 TryCand, Cand, SingleCritical))
1990 if (Cand.Reason == SingleCritical)
1991 Cand.Reason = MultiPressure;
1993 // Keep clustered nodes together to encourage downstream peephole
1994 // optimizations which may reduce resource requirements.
1996 // This is a best effort to set things up for a post-RA pass. Optimizations
1997 // like generating loads of multiple registers should ideally be done within
1998 // the scheduler pass by combining the loads during DAG postprocessing.
1999 const SUnit *NextClusterSU =
2000 Zone.isTop() ? DAG->getNextClusterSucc() : DAG->getNextClusterPred();
2001 if (tryGreater(TryCand.SU == NextClusterSU, Cand.SU == NextClusterSU,
2002 TryCand, Cand, Cluster))
2005 // Weak edges are for clustering and other constraints.
2007 // Deferring TryCand here does not change Cand's reason. This is good in the
2008 // sense that a bad candidate shouldn't affect a previous candidate's
2009 // goodness, but bad in that it is assymetric and depends on queue order.
2010 CandReason OrigReason = Cand.Reason;
2011 if (tryLess(getWeakLeft(TryCand.SU, Zone.isTop()),
2012 getWeakLeft(Cand.SU, Zone.isTop()),
2013 TryCand, Cand, Weak)) {
2014 Cand.Reason = OrigReason;
2017 // Avoid critical resource consumption and balance the schedule.
2018 TryCand.initResourceDelta(DAG, SchedModel);
2019 if (tryLess(TryCand.ResDelta.CritResources, Cand.ResDelta.CritResources,
2020 TryCand, Cand, ResourceReduce))
2022 if (tryGreater(TryCand.ResDelta.DemandedResources,
2023 Cand.ResDelta.DemandedResources,
2024 TryCand, Cand, ResourceDemand))
2027 // Avoid serializing long latency dependence chains.
2028 if (Cand.Policy.ReduceLatency) {
2030 if (Cand.SU->getDepth() * SchedModel->getLatencyFactor()
2031 > Zone.ExpectedCount) {
2032 if (tryLess(TryCand.SU->getDepth(), Cand.SU->getDepth(),
2033 TryCand, Cand, TopDepthReduce))
2036 if (tryGreater(TryCand.SU->getHeight(), Cand.SU->getHeight(),
2037 TryCand, Cand, TopPathReduce))
2041 if (Cand.SU->getHeight() * SchedModel->getLatencyFactor()
2042 > Zone.ExpectedCount) {
2043 if (tryLess(TryCand.SU->getHeight(), Cand.SU->getHeight(),
2044 TryCand, Cand, BotHeightReduce))
2047 if (tryGreater(TryCand.SU->getDepth(), Cand.SU->getDepth(),
2048 TryCand, Cand, BotPathReduce))
2053 // Avoid increasing the max pressure of the entire region.
2054 if (tryLess(TryCand.RPDelta.CurrentMax.UnitIncrease,
2055 Cand.RPDelta.CurrentMax.UnitIncrease, TryCand, Cand, SingleMax))
2057 if (Cand.Reason == SingleMax)
2058 Cand.Reason = MultiPressure;
2060 // Prefer immediate defs/users of the last scheduled instruction. This is a
2061 // nice pressure avoidance strategy that also conserves the processor's
2062 // register renaming resources and keeps the machine code readable.
2063 if (tryGreater(Zone.NextSUs.count(TryCand.SU), Zone.NextSUs.count(Cand.SU),
2064 TryCand, Cand, NextDefUse))
2067 // Fall through to original instruction order.
2068 if ((Zone.isTop() && TryCand.SU->NodeNum < Cand.SU->NodeNum)
2069 || (!Zone.isTop() && TryCand.SU->NodeNum > Cand.SU->NodeNum)) {
2070 TryCand.Reason = NodeOrder;
2074 /// pickNodeFromQueue helper that returns true if the LHS reg pressure effect is
2075 /// more desirable than RHS from scheduling standpoint.
2076 static bool compareRPDelta(const RegPressureDelta &LHS,
2077 const RegPressureDelta &RHS) {
2078 // Compare each component of pressure in decreasing order of importance
2079 // without checking if any are valid. Invalid PressureElements are assumed to
2080 // have UnitIncrease==0, so are neutral.
2082 // Avoid increasing the max critical pressure in the scheduled region.
2083 if (LHS.Excess.UnitIncrease != RHS.Excess.UnitIncrease) {
2084 DEBUG(dbgs() << " RP excess top - bot: "
2085 << (LHS.Excess.UnitIncrease - RHS.Excess.UnitIncrease) << '\n');
2086 return LHS.Excess.UnitIncrease < RHS.Excess.UnitIncrease;
2088 // Avoid increasing the max critical pressure in the scheduled region.
2089 if (LHS.CriticalMax.UnitIncrease != RHS.CriticalMax.UnitIncrease) {
2090 DEBUG(dbgs() << " RP critical top - bot: "
2091 << (LHS.CriticalMax.UnitIncrease - RHS.CriticalMax.UnitIncrease)
2093 return LHS.CriticalMax.UnitIncrease < RHS.CriticalMax.UnitIncrease;
2095 // Avoid increasing the max pressure of the entire region.
2096 if (LHS.CurrentMax.UnitIncrease != RHS.CurrentMax.UnitIncrease) {
2097 DEBUG(dbgs() << " RP current top - bot: "
2098 << (LHS.CurrentMax.UnitIncrease - RHS.CurrentMax.UnitIncrease)
2100 return LHS.CurrentMax.UnitIncrease < RHS.CurrentMax.UnitIncrease;
2106 const char *ConvergingScheduler::getReasonStr(
2107 ConvergingScheduler::CandReason Reason) {
2109 case NoCand: return "NOCAND ";
2110 case PhysRegCopy: return "PREG-COPY";
2111 case SingleExcess: return "REG-EXCESS";
2112 case SingleCritical: return "REG-CRIT ";
2113 case Cluster: return "CLUSTER ";
2114 case Weak: return "WEAK ";
2115 case SingleMax: return "REG-MAX ";
2116 case MultiPressure: return "REG-MULTI ";
2117 case ResourceReduce: return "RES-REDUCE";
2118 case ResourceDemand: return "RES-DEMAND";
2119 case TopDepthReduce: return "TOP-DEPTH ";
2120 case TopPathReduce: return "TOP-PATH ";
2121 case BotHeightReduce:return "BOT-HEIGHT";
2122 case BotPathReduce: return "BOT-PATH ";
2123 case NextDefUse: return "DEF-USE ";
2124 case NodeOrder: return "ORDER ";
2126 llvm_unreachable("Unknown reason!");
2129 void ConvergingScheduler::traceCandidate(const SchedCandidate &Cand) {
2131 unsigned ResIdx = 0;
2132 unsigned Latency = 0;
2133 switch (Cand.Reason) {
2137 P = Cand.RPDelta.Excess;
2139 case SingleCritical:
2140 P = Cand.RPDelta.CriticalMax;
2143 P = Cand.RPDelta.CurrentMax;
2145 case ResourceReduce:
2146 ResIdx = Cand.Policy.ReduceResIdx;
2148 case ResourceDemand:
2149 ResIdx = Cand.Policy.DemandResIdx;
2151 case TopDepthReduce:
2152 Latency = Cand.SU->getDepth();
2155 Latency = Cand.SU->getHeight();
2157 case BotHeightReduce:
2158 Latency = Cand.SU->getHeight();
2161 Latency = Cand.SU->getDepth();
2164 dbgs() << " SU(" << Cand.SU->NodeNum << ") " << getReasonStr(Cand.Reason);
2166 dbgs() << " " << TRI->getRegPressureSetName(P.PSetID)
2167 << ":" << P.UnitIncrease << " ";
2171 dbgs() << " " << SchedModel->getProcResource(ResIdx)->Name << " ";
2175 dbgs() << " " << Latency << " cycles ";
2182 /// Pick the best candidate from the top queue.
2184 /// TODO: getMaxPressureDelta results can be mostly cached for each SUnit during
2185 /// DAG building. To adjust for the current scheduling location we need to
2186 /// maintain the number of vreg uses remaining to be top-scheduled.
2187 void ConvergingScheduler::pickNodeFromQueue(SchedBoundary &Zone,
2188 const RegPressureTracker &RPTracker,
2189 SchedCandidate &Cand) {
2190 ReadyQueue &Q = Zone.Available;
2194 // getMaxPressureDelta temporarily modifies the tracker.
2195 RegPressureTracker &TempTracker = const_cast<RegPressureTracker&>(RPTracker);
2197 for (ReadyQueue::iterator I = Q.begin(), E = Q.end(); I != E; ++I) {
2199 SchedCandidate TryCand(Cand.Policy);
2201 tryCandidate(Cand, TryCand, Zone, RPTracker, TempTracker);
2202 if (TryCand.Reason != NoCand) {
2203 // Initialize resource delta if needed in case future heuristics query it.
2204 if (TryCand.ResDelta == SchedResourceDelta())
2205 TryCand.initResourceDelta(DAG, SchedModel);
2206 Cand.setBest(TryCand);
2207 DEBUG(traceCandidate(Cand));
2212 static void tracePick(const ConvergingScheduler::SchedCandidate &Cand,
2214 DEBUG(dbgs() << "Pick " << (IsTop ? "Top " : "Bot ")
2215 << ConvergingScheduler::getReasonStr(Cand.Reason) << '\n');
2218 /// Pick the best candidate node from either the top or bottom queue.
2219 SUnit *ConvergingScheduler::pickNodeBidirectional(bool &IsTopNode) {
2220 // Schedule as far as possible in the direction of no choice. This is most
2221 // efficient, but also provides the best heuristics for CriticalPSets.
2222 if (SUnit *SU = Bot.pickOnlyChoice()) {
2224 DEBUG(dbgs() << "Pick Top NOCAND\n");
2227 if (SUnit *SU = Top.pickOnlyChoice()) {
2229 DEBUG(dbgs() << "Pick Bot NOCAND\n");
2232 CandPolicy NoPolicy;
2233 SchedCandidate BotCand(NoPolicy);
2234 SchedCandidate TopCand(NoPolicy);
2235 checkResourceLimits(TopCand, BotCand);
2237 // Prefer bottom scheduling when heuristics are silent.
2238 pickNodeFromQueue(Bot, DAG->getBotRPTracker(), BotCand);
2239 assert(BotCand.Reason != NoCand && "failed to find the first candidate");
2241 // If either Q has a single candidate that provides the least increase in
2242 // Excess pressure, we can immediately schedule from that Q.
2244 // RegionCriticalPSets summarizes the pressure within the scheduled region and
2245 // affects picking from either Q. If scheduling in one direction must
2246 // increase pressure for one of the excess PSets, then schedule in that
2247 // direction first to provide more freedom in the other direction.
2248 if (BotCand.Reason == SingleExcess || BotCand.Reason == SingleCritical) {
2250 tracePick(BotCand, IsTopNode);
2253 // Check if the top Q has a better candidate.
2254 pickNodeFromQueue(Top, DAG->getTopRPTracker(), TopCand);
2255 assert(TopCand.Reason != NoCand && "failed to find the first candidate");
2257 // If either Q has a single candidate that minimizes pressure above the
2258 // original region's pressure pick it.
2259 if (TopCand.Reason <= SingleMax || BotCand.Reason <= SingleMax) {
2260 if (TopCand.Reason < BotCand.Reason) {
2262 tracePick(TopCand, IsTopNode);
2266 tracePick(BotCand, IsTopNode);
2269 // Check for a salient pressure difference and pick the best from either side.
2270 if (compareRPDelta(TopCand.RPDelta, BotCand.RPDelta)) {
2272 tracePick(TopCand, IsTopNode);
2275 // Otherwise prefer the bottom candidate, in node order if all else failed.
2276 if (TopCand.Reason < BotCand.Reason) {
2278 tracePick(TopCand, IsTopNode);
2282 tracePick(BotCand, IsTopNode);
2286 /// Pick the best node to balance the schedule. Implements MachineSchedStrategy.
2287 SUnit *ConvergingScheduler::pickNode(bool &IsTopNode) {
2288 if (DAG->top() == DAG->bottom()) {
2289 assert(Top.Available.empty() && Top.Pending.empty() &&
2290 Bot.Available.empty() && Bot.Pending.empty() && "ReadyQ garbage");
2296 SU = Top.pickOnlyChoice();
2298 CandPolicy NoPolicy;
2299 SchedCandidate TopCand(NoPolicy);
2300 pickNodeFromQueue(Top, DAG->getTopRPTracker(), TopCand);
2301 assert(TopCand.Reason != NoCand && "failed to find the first candidate");
2306 else if (ForceBottomUp) {
2307 SU = Bot.pickOnlyChoice();
2309 CandPolicy NoPolicy;
2310 SchedCandidate BotCand(NoPolicy);
2311 pickNodeFromQueue(Bot, DAG->getBotRPTracker(), BotCand);
2312 assert(BotCand.Reason != NoCand && "failed to find the first candidate");
2318 SU = pickNodeBidirectional(IsTopNode);
2320 } while (SU->isScheduled);
2322 if (SU->isTopReady())
2323 Top.removeReady(SU);
2324 if (SU->isBottomReady())
2325 Bot.removeReady(SU);
2327 DEBUG(dbgs() << "Scheduling SU(" << SU->NodeNum << ") " << *SU->getInstr());
2331 void ConvergingScheduler::reschedulePhysRegCopies(SUnit *SU, bool isTop) {
2333 MachineBasicBlock::iterator InsertPos = SU->getInstr();
2336 SmallVectorImpl<SDep> &Deps = isTop ? SU->Preds : SU->Succs;
2338 // Find already scheduled copies with a single physreg dependence and move
2339 // them just above the scheduled instruction.
2340 for (SmallVectorImpl<SDep>::iterator I = Deps.begin(), E = Deps.end();
2342 if (I->getKind() != SDep::Data || !TRI->isPhysicalRegister(I->getReg()))
2344 SUnit *DepSU = I->getSUnit();
2345 if (isTop ? DepSU->Succs.size() > 1 : DepSU->Preds.size() > 1)
2347 MachineInstr *Copy = DepSU->getInstr();
2348 if (!Copy->isCopy())
2350 DEBUG(dbgs() << " Rescheduling physreg copy ";
2351 I->getSUnit()->dump(DAG));
2352 DAG->moveInstruction(Copy, InsertPos);
2356 /// Update the scheduler's state after scheduling a node. This is the same node
2357 /// that was just returned by pickNode(). However, ScheduleDAGMI needs to update
2358 /// it's state based on the current cycle before MachineSchedStrategy does.
2360 /// FIXME: Eventually, we may bundle physreg copies rather than rescheduling
2361 /// them here. See comments in biasPhysRegCopy.
2362 void ConvergingScheduler::schedNode(SUnit *SU, bool IsTopNode) {
2364 SU->TopReadyCycle = Top.CurrCycle;
2366 if (SU->hasPhysRegUses)
2367 reschedulePhysRegCopies(SU, true);
2370 SU->BotReadyCycle = Bot.CurrCycle;
2372 if (SU->hasPhysRegDefs)
2373 reschedulePhysRegCopies(SU, false);
2377 /// Create the standard converging machine scheduler. This will be used as the
2378 /// default scheduler if the target does not set a default.
2379 static ScheduleDAGInstrs *createConvergingSched(MachineSchedContext *C) {
2380 assert((!ForceTopDown || !ForceBottomUp) &&
2381 "-misched-topdown incompatible with -misched-bottomup");
2382 ScheduleDAGMI *DAG = new ScheduleDAGMI(C, new ConvergingScheduler());
2383 // Register DAG post-processors.
2385 // FIXME: extend the mutation API to allow earlier mutations to instantiate
2386 // data and pass it to later mutations. Have a single mutation that gathers
2387 // the interesting nodes in one pass.
2388 DAG->addMutation(new CopyConstrain(DAG->TII, DAG->TRI));
2389 if (EnableLoadCluster)
2390 DAG->addMutation(new LoadClusterMutation(DAG->TII, DAG->TRI));
2391 if (EnableMacroFusion)
2392 DAG->addMutation(new MacroFusion(DAG->TII));
2395 static MachineSchedRegistry
2396 ConvergingSchedRegistry("converge", "Standard converging scheduler.",
2397 createConvergingSched);
2399 //===----------------------------------------------------------------------===//
2400 // ILP Scheduler. Currently for experimental analysis of heuristics.
2401 //===----------------------------------------------------------------------===//
2404 /// \brief Order nodes by the ILP metric.
2406 const SchedDFSResult *DFSResult;
2407 const BitVector *ScheduledTrees;
2410 ILPOrder(bool MaxILP): DFSResult(0), ScheduledTrees(0), MaximizeILP(MaxILP) {}
2412 /// \brief Apply a less-than relation on node priority.
2414 /// (Return true if A comes after B in the Q.)
2415 bool operator()(const SUnit *A, const SUnit *B) const {
2416 unsigned SchedTreeA = DFSResult->getSubtreeID(A);
2417 unsigned SchedTreeB = DFSResult->getSubtreeID(B);
2418 if (SchedTreeA != SchedTreeB) {
2419 // Unscheduled trees have lower priority.
2420 if (ScheduledTrees->test(SchedTreeA) != ScheduledTrees->test(SchedTreeB))
2421 return ScheduledTrees->test(SchedTreeB);
2423 // Trees with shallower connections have have lower priority.
2424 if (DFSResult->getSubtreeLevel(SchedTreeA)
2425 != DFSResult->getSubtreeLevel(SchedTreeB)) {
2426 return DFSResult->getSubtreeLevel(SchedTreeA)
2427 < DFSResult->getSubtreeLevel(SchedTreeB);
2431 return DFSResult->getILP(A) < DFSResult->getILP(B);
2433 return DFSResult->getILP(A) > DFSResult->getILP(B);
2437 /// \brief Schedule based on the ILP metric.
2438 class ILPScheduler : public MachineSchedStrategy {
2439 /// In case all subtrees are eventually connected to a common root through
2440 /// data dependence (e.g. reduction), place an upper limit on their size.
2442 /// FIXME: A subtree limit is generally good, but in the situation commented
2443 /// above, where multiple similar subtrees feed a common root, we should
2444 /// only split at a point where the resulting subtrees will be balanced.
2445 /// (a motivating test case must be found).
2446 static const unsigned SubtreeLimit = 16;
2451 std::vector<SUnit*> ReadyQ;
2453 ILPScheduler(bool MaximizeILP): DAG(0), Cmp(MaximizeILP) {}
2455 virtual void initialize(ScheduleDAGMI *dag) {
2457 DAG->computeDFSResult();
2458 Cmp.DFSResult = DAG->getDFSResult();
2459 Cmp.ScheduledTrees = &DAG->getScheduledTrees();
2463 virtual void registerRoots() {
2464 // Restore the heap in ReadyQ with the updated DFS results.
2465 std::make_heap(ReadyQ.begin(), ReadyQ.end(), Cmp);
2468 /// Implement MachineSchedStrategy interface.
2469 /// -----------------------------------------
2471 /// Callback to select the highest priority node from the ready Q.
2472 virtual SUnit *pickNode(bool &IsTopNode) {
2473 if (ReadyQ.empty()) return NULL;
2474 std::pop_heap(ReadyQ.begin(), ReadyQ.end(), Cmp);
2475 SUnit *SU = ReadyQ.back();
2478 DEBUG(dbgs() << "Pick node " << "SU(" << SU->NodeNum << ") "
2479 << " ILP: " << DAG->getDFSResult()->getILP(SU)
2480 << " Tree: " << DAG->getDFSResult()->getSubtreeID(SU) << " @"
2481 << DAG->getDFSResult()->getSubtreeLevel(
2482 DAG->getDFSResult()->getSubtreeID(SU)) << '\n'
2483 << "Scheduling " << *SU->getInstr());
2487 /// \brief Scheduler callback to notify that a new subtree is scheduled.
2488 virtual void scheduleTree(unsigned SubtreeID) {
2489 std::make_heap(ReadyQ.begin(), ReadyQ.end(), Cmp);
2492 /// Callback after a node is scheduled. Mark a newly scheduled tree, notify
2493 /// DFSResults, and resort the priority Q.
2494 virtual void schedNode(SUnit *SU, bool IsTopNode) {
2495 assert(!IsTopNode && "SchedDFSResult needs bottom-up");
2498 virtual void releaseTopNode(SUnit *) { /*only called for top roots*/ }
2500 virtual void releaseBottomNode(SUnit *SU) {
2501 ReadyQ.push_back(SU);
2502 std::push_heap(ReadyQ.begin(), ReadyQ.end(), Cmp);
2507 static ScheduleDAGInstrs *createILPMaxScheduler(MachineSchedContext *C) {
2508 return new ScheduleDAGMI(C, new ILPScheduler(true));
2510 static ScheduleDAGInstrs *createILPMinScheduler(MachineSchedContext *C) {
2511 return new ScheduleDAGMI(C, new ILPScheduler(false));
2513 static MachineSchedRegistry ILPMaxRegistry(
2514 "ilpmax", "Schedule bottom-up for max ILP", createILPMaxScheduler);
2515 static MachineSchedRegistry ILPMinRegistry(
2516 "ilpmin", "Schedule bottom-up for min ILP", createILPMinScheduler);
2518 //===----------------------------------------------------------------------===//
2519 // Machine Instruction Shuffler for Correctness Testing
2520 //===----------------------------------------------------------------------===//
2524 /// Apply a less-than relation on the node order, which corresponds to the
2525 /// instruction order prior to scheduling. IsReverse implements greater-than.
2526 template<bool IsReverse>
2528 bool operator()(SUnit *A, SUnit *B) const {
2530 return A->NodeNum > B->NodeNum;
2532 return A->NodeNum < B->NodeNum;
2536 /// Reorder instructions as much as possible.
2537 class InstructionShuffler : public MachineSchedStrategy {
2541 // Using a less-than relation (SUnitOrder<false>) for the TopQ priority
2542 // gives nodes with a higher number higher priority causing the latest
2543 // instructions to be scheduled first.
2544 PriorityQueue<SUnit*, std::vector<SUnit*>, SUnitOrder<false> >
2546 // When scheduling bottom-up, use greater-than as the queue priority.
2547 PriorityQueue<SUnit*, std::vector<SUnit*>, SUnitOrder<true> >
2550 InstructionShuffler(bool alternate, bool topdown)
2551 : IsAlternating(alternate), IsTopDown(topdown) {}
2553 virtual void initialize(ScheduleDAGMI *) {
2558 /// Implement MachineSchedStrategy interface.
2559 /// -----------------------------------------
2561 virtual SUnit *pickNode(bool &IsTopNode) {
2565 if (TopQ.empty()) return NULL;
2568 } while (SU->isScheduled);
2573 if (BottomQ.empty()) return NULL;
2576 } while (SU->isScheduled);
2580 IsTopDown = !IsTopDown;
2584 virtual void schedNode(SUnit *SU, bool IsTopNode) {}
2586 virtual void releaseTopNode(SUnit *SU) {
2589 virtual void releaseBottomNode(SUnit *SU) {
2595 static ScheduleDAGInstrs *createInstructionShuffler(MachineSchedContext *C) {
2596 bool Alternate = !ForceTopDown && !ForceBottomUp;
2597 bool TopDown = !ForceBottomUp;
2598 assert((TopDown || !ForceTopDown) &&
2599 "-misched-topdown incompatible with -misched-bottomup");
2600 return new ScheduleDAGMI(C, new InstructionShuffler(Alternate, TopDown));
2602 static MachineSchedRegistry ShufflerRegistry(
2603 "shuffle", "Shuffle machine instructions alternating directions",
2604 createInstructionShuffler);
2607 //===----------------------------------------------------------------------===//
2608 // GraphWriter support for ScheduleDAGMI.
2609 //===----------------------------------------------------------------------===//
2614 template<> struct GraphTraits<
2615 ScheduleDAGMI*> : public GraphTraits<ScheduleDAG*> {};
2618 struct DOTGraphTraits<ScheduleDAGMI*> : public DefaultDOTGraphTraits {
2620 DOTGraphTraits (bool isSimple=false) : DefaultDOTGraphTraits(isSimple) {}
2622 static std::string getGraphName(const ScheduleDAG *G) {
2623 return G->MF.getName();
2626 static bool renderGraphFromBottomUp() {
2630 static bool isNodeHidden(const SUnit *Node) {
2631 return (Node->NumPreds > 10 || Node->NumSuccs > 10);
2634 static bool hasNodeAddressLabel(const SUnit *Node,
2635 const ScheduleDAG *Graph) {
2639 /// If you want to override the dot attributes printed for a particular
2640 /// edge, override this method.
2641 static std::string getEdgeAttributes(const SUnit *Node,
2643 const ScheduleDAG *Graph) {
2644 if (EI.isArtificialDep())
2645 return "color=cyan,style=dashed";
2647 return "color=blue,style=dashed";
2651 static std::string getNodeLabel(const SUnit *SU, const ScheduleDAG *G) {
2653 raw_string_ostream SS(Str);
2654 SS << "SU(" << SU->NodeNum << ')';
2657 static std::string getNodeDescription(const SUnit *SU, const ScheduleDAG *G) {
2658 return G->getGraphNodeLabel(SU);
2661 static std::string getNodeAttributes(const SUnit *N,
2662 const ScheduleDAG *Graph) {
2663 std::string Str("shape=Mrecord");
2664 const SchedDFSResult *DFS =
2665 static_cast<const ScheduleDAGMI*>(Graph)->getDFSResult();
2667 Str += ",style=filled,fillcolor=\"#";
2668 Str += DOT::getColorString(DFS->getSubtreeID(N));
2677 /// viewGraph - Pop up a ghostview window with the reachable parts of the DAG
2678 /// rendered using 'dot'.
2680 void ScheduleDAGMI::viewGraph(const Twine &Name, const Twine &Title) {
2682 ViewGraph(this, Name, false, Title);
2684 errs() << "ScheduleDAGMI::viewGraph is only available in debug builds on "
2685 << "systems with Graphviz or gv!\n";
2689 /// Out-of-line implementation with no arguments is handy for gdb.
2690 void ScheduleDAGMI::viewGraph() {
2691 viewGraph(getDAGName(), "Scheduling-Units Graph for " + getDAGName());