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/MachineRegisterInfo.h"
25 #include "llvm/CodeGen/Passes.h"
26 #include "llvm/CodeGen/RegisterClassInfo.h"
27 #include "llvm/CodeGen/ScheduleDFS.h"
28 #include "llvm/CodeGen/ScheduleHazardRecognizer.h"
29 #include "llvm/Support/CommandLine.h"
30 #include "llvm/Support/Debug.h"
31 #include "llvm/Support/ErrorHandling.h"
32 #include "llvm/Support/GraphWriter.h"
33 #include "llvm/Support/raw_ostream.h"
34 #include "llvm/Target/TargetInstrInfo.h"
40 cl::opt<bool> ForceTopDown("misched-topdown", cl::Hidden,
41 cl::desc("Force top-down list scheduling"));
42 cl::opt<bool> ForceBottomUp("misched-bottomup", cl::Hidden,
43 cl::desc("Force bottom-up list scheduling"));
47 static cl::opt<bool> ViewMISchedDAGs("view-misched-dags", cl::Hidden,
48 cl::desc("Pop up a window to show MISched dags after they are processed"));
50 static cl::opt<unsigned> MISchedCutoff("misched-cutoff", cl::Hidden,
51 cl::desc("Stop scheduling after N instructions"), cl::init(~0U));
53 static bool ViewMISchedDAGs = false;
56 static cl::opt<bool> EnableRegPressure("misched-regpressure", cl::Hidden,
57 cl::desc("Enable register pressure scheduling."), cl::init(true));
59 static cl::opt<bool> EnableCyclicPath("misched-cyclicpath", cl::Hidden,
60 cl::desc("Enable cyclic critical path analysis."), cl::init(false));
62 static cl::opt<bool> EnableLoadCluster("misched-cluster", cl::Hidden,
63 cl::desc("Enable load clustering."), cl::init(true));
65 // Experimental heuristics
66 static cl::opt<bool> EnableMacroFusion("misched-fusion", cl::Hidden,
67 cl::desc("Enable scheduling for macro fusion."), cl::init(true));
69 static cl::opt<bool> VerifyScheduling("verify-misched", cl::Hidden,
70 cl::desc("Verify machine instrs before and after machine scheduling"));
72 // DAG subtrees must have at least this many nodes.
73 static const unsigned MinSubtreeSize = 8;
75 //===----------------------------------------------------------------------===//
76 // Machine Instruction Scheduling Pass and Registry
77 //===----------------------------------------------------------------------===//
79 MachineSchedContext::MachineSchedContext():
80 MF(0), MLI(0), MDT(0), PassConfig(0), AA(0), LIS(0) {
81 RegClassInfo = new RegisterClassInfo();
84 MachineSchedContext::~MachineSchedContext() {
89 /// MachineScheduler runs after coalescing and before register allocation.
90 class MachineScheduler : public MachineSchedContext,
91 public MachineFunctionPass {
95 virtual void getAnalysisUsage(AnalysisUsage &AU) const;
97 virtual void releaseMemory() {}
99 virtual bool runOnMachineFunction(MachineFunction&);
101 virtual void print(raw_ostream &O, const Module* = 0) const;
103 static char ID; // Class identification, replacement for typeinfo
107 char MachineScheduler::ID = 0;
109 char &llvm::MachineSchedulerID = MachineScheduler::ID;
111 INITIALIZE_PASS_BEGIN(MachineScheduler, "misched",
112 "Machine Instruction Scheduler", false, false)
113 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
114 INITIALIZE_PASS_DEPENDENCY(SlotIndexes)
115 INITIALIZE_PASS_DEPENDENCY(LiveIntervals)
116 INITIALIZE_PASS_END(MachineScheduler, "misched",
117 "Machine Instruction Scheduler", false, false)
119 MachineScheduler::MachineScheduler()
120 : MachineFunctionPass(ID) {
121 initializeMachineSchedulerPass(*PassRegistry::getPassRegistry());
124 void MachineScheduler::getAnalysisUsage(AnalysisUsage &AU) const {
125 AU.setPreservesCFG();
126 AU.addRequiredID(MachineDominatorsID);
127 AU.addRequired<MachineLoopInfo>();
128 AU.addRequired<AliasAnalysis>();
129 AU.addRequired<TargetPassConfig>();
130 AU.addRequired<SlotIndexes>();
131 AU.addPreserved<SlotIndexes>();
132 AU.addRequired<LiveIntervals>();
133 AU.addPreserved<LiveIntervals>();
134 MachineFunctionPass::getAnalysisUsage(AU);
137 MachinePassRegistry MachineSchedRegistry::Registry;
139 /// A dummy default scheduler factory indicates whether the scheduler
140 /// is overridden on the command line.
141 static ScheduleDAGInstrs *useDefaultMachineSched(MachineSchedContext *C) {
145 /// MachineSchedOpt allows command line selection of the scheduler.
146 static cl::opt<MachineSchedRegistry::ScheduleDAGCtor, false,
147 RegisterPassParser<MachineSchedRegistry> >
148 MachineSchedOpt("misched",
149 cl::init(&useDefaultMachineSched), cl::Hidden,
150 cl::desc("Machine instruction scheduler to use"));
152 static MachineSchedRegistry
153 DefaultSchedRegistry("default", "Use the target's default scheduler choice.",
154 useDefaultMachineSched);
156 /// Forward declare the standard machine scheduler. This will be used as the
157 /// default scheduler if the target does not set a default.
158 static ScheduleDAGInstrs *createConvergingSched(MachineSchedContext *C);
161 /// Decrement this iterator until reaching the top or a non-debug instr.
162 static MachineBasicBlock::const_iterator
163 priorNonDebug(MachineBasicBlock::const_iterator I,
164 MachineBasicBlock::const_iterator Beg) {
165 assert(I != Beg && "reached the top of the region, cannot decrement");
167 if (!I->isDebugValue())
173 /// Non-const version.
174 static MachineBasicBlock::iterator
175 priorNonDebug(MachineBasicBlock::iterator I,
176 MachineBasicBlock::const_iterator Beg) {
177 return const_cast<MachineInstr*>(
178 &*priorNonDebug(MachineBasicBlock::const_iterator(I), Beg));
181 /// If this iterator is a debug value, increment until reaching the End or a
182 /// non-debug instruction.
183 static MachineBasicBlock::const_iterator
184 nextIfDebug(MachineBasicBlock::const_iterator I,
185 MachineBasicBlock::const_iterator End) {
186 for(; I != End; ++I) {
187 if (!I->isDebugValue())
193 /// Non-const version.
194 static MachineBasicBlock::iterator
195 nextIfDebug(MachineBasicBlock::iterator I,
196 MachineBasicBlock::const_iterator End) {
197 // Cast the return value to nonconst MachineInstr, then cast to an
198 // instr_iterator, which does not check for null, finally return a
200 return MachineBasicBlock::instr_iterator(
201 const_cast<MachineInstr*>(
202 &*nextIfDebug(MachineBasicBlock::const_iterator(I), End)));
205 /// Top-level MachineScheduler pass driver.
207 /// Visit blocks in function order. Divide each block into scheduling regions
208 /// and visit them bottom-up. Visiting regions bottom-up is not required, but is
209 /// consistent with the DAG builder, which traverses the interior of the
210 /// scheduling regions bottom-up.
212 /// This design avoids exposing scheduling boundaries to the DAG builder,
213 /// simplifying the DAG builder's support for "special" target instructions.
214 /// At the same time the design allows target schedulers to operate across
215 /// scheduling boundaries, for example to bundle the boudary instructions
216 /// without reordering them. This creates complexity, because the target
217 /// scheduler must update the RegionBegin and RegionEnd positions cached by
218 /// ScheduleDAGInstrs whenever adding or removing instructions. A much simpler
219 /// design would be to split blocks at scheduling boundaries, but LLVM has a
220 /// general bias against block splitting purely for implementation simplicity.
221 bool MachineScheduler::runOnMachineFunction(MachineFunction &mf) {
222 DEBUG(dbgs() << "Before MISsched:\n"; mf.print(dbgs()));
224 // Initialize the context of the pass.
226 MLI = &getAnalysis<MachineLoopInfo>();
227 MDT = &getAnalysis<MachineDominatorTree>();
228 PassConfig = &getAnalysis<TargetPassConfig>();
229 AA = &getAnalysis<AliasAnalysis>();
231 LIS = &getAnalysis<LiveIntervals>();
232 const TargetInstrInfo *TII = MF->getTarget().getInstrInfo();
234 if (VerifyScheduling) {
236 MF->verify(this, "Before machine scheduling.");
238 RegClassInfo->runOnMachineFunction(*MF);
240 // Select the scheduler, or set the default.
241 MachineSchedRegistry::ScheduleDAGCtor Ctor = MachineSchedOpt;
242 if (Ctor == useDefaultMachineSched) {
243 // Get the default scheduler set by the target.
244 Ctor = MachineSchedRegistry::getDefault();
246 Ctor = createConvergingSched;
247 MachineSchedRegistry::setDefault(Ctor);
250 // Instantiate the selected scheduler.
251 OwningPtr<ScheduleDAGInstrs> Scheduler(Ctor(this));
253 // Visit all machine basic blocks.
255 // TODO: Visit blocks in global postorder or postorder within the bottom-up
256 // loop tree. Then we can optionally compute global RegPressure.
257 for (MachineFunction::iterator MBB = MF->begin(), MBBEnd = MF->end();
258 MBB != MBBEnd; ++MBB) {
260 Scheduler->startBlock(MBB);
262 // Break the block into scheduling regions [I, RegionEnd), and schedule each
263 // region as soon as it is discovered. RegionEnd points the scheduling
264 // boundary at the bottom of the region. The DAG does not include RegionEnd,
265 // but the region does (i.e. the next RegionEnd is above the previous
266 // RegionBegin). If the current block has no terminator then RegionEnd ==
267 // MBB->end() for the bottom region.
269 // The Scheduler may insert instructions during either schedule() or
270 // exitRegion(), even for empty regions. So the local iterators 'I' and
271 // 'RegionEnd' are invalid across these calls.
272 unsigned RemainingInstrs = MBB->size();
273 for(MachineBasicBlock::iterator RegionEnd = MBB->end();
274 RegionEnd != MBB->begin(); RegionEnd = Scheduler->begin()) {
276 // Avoid decrementing RegionEnd for blocks with no terminator.
277 if (RegionEnd != MBB->end()
278 || TII->isSchedulingBoundary(llvm::prior(RegionEnd), MBB, *MF)) {
280 // Count the boundary instruction.
284 // The next region starts above the previous region. Look backward in the
285 // instruction stream until we find the nearest boundary.
286 unsigned NumRegionInstrs = 0;
287 MachineBasicBlock::iterator I = RegionEnd;
288 for(;I != MBB->begin(); --I, --RemainingInstrs, ++NumRegionInstrs) {
289 if (TII->isSchedulingBoundary(llvm::prior(I), MBB, *MF))
292 // Notify the scheduler of the region, even if we may skip scheduling
293 // it. Perhaps it still needs to be bundled.
294 Scheduler->enterRegion(MBB, I, RegionEnd, NumRegionInstrs);
296 // Skip empty scheduling regions (0 or 1 schedulable instructions).
297 if (I == RegionEnd || I == llvm::prior(RegionEnd)) {
298 // Close the current region. Bundle the terminator if needed.
299 // This invalidates 'RegionEnd' and 'I'.
300 Scheduler->exitRegion();
303 DEBUG(dbgs() << "********** MI Scheduling **********\n");
304 DEBUG(dbgs() << MF->getName()
305 << ":BB#" << MBB->getNumber() << " " << MBB->getName()
306 << "\n From: " << *I << " To: ";
307 if (RegionEnd != MBB->end()) dbgs() << *RegionEnd;
308 else dbgs() << "End";
309 dbgs() << " RegionInstrs: " << NumRegionInstrs
310 << " Remaining: " << RemainingInstrs << "\n");
312 // Schedule a region: possibly reorder instructions.
313 // This invalidates 'RegionEnd' and 'I'.
314 Scheduler->schedule();
316 // Close the current region.
317 Scheduler->exitRegion();
319 // Scheduling has invalidated the current iterator 'I'. Ask the
320 // scheduler for the top of it's scheduled region.
321 RegionEnd = Scheduler->begin();
323 assert(RemainingInstrs == 0 && "Instruction count mismatch!");
324 Scheduler->finishBlock();
326 Scheduler->finalizeSchedule();
328 if (VerifyScheduling)
329 MF->verify(this, "After machine scheduling.");
333 void MachineScheduler::print(raw_ostream &O, const Module* m) const {
337 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
338 void ReadyQueue::dump() {
339 dbgs() << Name << ": ";
340 for (unsigned i = 0, e = Queue.size(); i < e; ++i)
341 dbgs() << Queue[i]->NodeNum << " ";
346 //===----------------------------------------------------------------------===//
347 // ScheduleDAGMI - Base class for MachineInstr scheduling with LiveIntervals
349 //===----------------------------------------------------------------------===//
351 ScheduleDAGMI::~ScheduleDAGMI() {
353 DeleteContainerPointers(Mutations);
357 bool ScheduleDAGMI::canAddEdge(SUnit *SuccSU, SUnit *PredSU) {
358 return SuccSU == &ExitSU || !Topo.IsReachable(PredSU, SuccSU);
361 bool ScheduleDAGMI::addEdge(SUnit *SuccSU, const SDep &PredDep) {
362 if (SuccSU != &ExitSU) {
363 // Do not use WillCreateCycle, it assumes SD scheduling.
364 // If Pred is reachable from Succ, then the edge creates a cycle.
365 if (Topo.IsReachable(PredDep.getSUnit(), SuccSU))
367 Topo.AddPred(SuccSU, PredDep.getSUnit());
369 SuccSU->addPred(PredDep, /*Required=*/!PredDep.isArtificial());
370 // Return true regardless of whether a new edge needed to be inserted.
374 /// ReleaseSucc - Decrement the NumPredsLeft count of a successor. When
375 /// NumPredsLeft reaches zero, release the successor node.
377 /// FIXME: Adjust SuccSU height based on MinLatency.
378 void ScheduleDAGMI::releaseSucc(SUnit *SU, SDep *SuccEdge) {
379 SUnit *SuccSU = SuccEdge->getSUnit();
381 if (SuccEdge->isWeak()) {
382 --SuccSU->WeakPredsLeft;
383 if (SuccEdge->isCluster())
384 NextClusterSucc = SuccSU;
388 if (SuccSU->NumPredsLeft == 0) {
389 dbgs() << "*** Scheduling failed! ***\n";
391 dbgs() << " has been released too many times!\n";
395 --SuccSU->NumPredsLeft;
396 if (SuccSU->NumPredsLeft == 0 && SuccSU != &ExitSU)
397 SchedImpl->releaseTopNode(SuccSU);
400 /// releaseSuccessors - Call releaseSucc on each of SU's successors.
401 void ScheduleDAGMI::releaseSuccessors(SUnit *SU) {
402 for (SUnit::succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
404 releaseSucc(SU, &*I);
408 /// ReleasePred - Decrement the NumSuccsLeft count of a predecessor. When
409 /// NumSuccsLeft reaches zero, release the predecessor node.
411 /// FIXME: Adjust PredSU height based on MinLatency.
412 void ScheduleDAGMI::releasePred(SUnit *SU, SDep *PredEdge) {
413 SUnit *PredSU = PredEdge->getSUnit();
415 if (PredEdge->isWeak()) {
416 --PredSU->WeakSuccsLeft;
417 if (PredEdge->isCluster())
418 NextClusterPred = PredSU;
422 if (PredSU->NumSuccsLeft == 0) {
423 dbgs() << "*** Scheduling failed! ***\n";
425 dbgs() << " has been released too many times!\n";
429 --PredSU->NumSuccsLeft;
430 if (PredSU->NumSuccsLeft == 0 && PredSU != &EntrySU)
431 SchedImpl->releaseBottomNode(PredSU);
434 /// releasePredecessors - Call releasePred on each of SU's predecessors.
435 void ScheduleDAGMI::releasePredecessors(SUnit *SU) {
436 for (SUnit::pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
438 releasePred(SU, &*I);
442 /// This is normally called from the main scheduler loop but may also be invoked
443 /// by the scheduling strategy to perform additional code motion.
444 void ScheduleDAGMI::moveInstruction(MachineInstr *MI,
445 MachineBasicBlock::iterator InsertPos) {
446 // Advance RegionBegin if the first instruction moves down.
447 if (&*RegionBegin == MI)
450 // Update the instruction stream.
451 BB->splice(InsertPos, BB, MI);
453 // Update LiveIntervals
454 LIS->handleMove(MI, /*UpdateFlags=*/true);
456 // Recede RegionBegin if an instruction moves above the first.
457 if (RegionBegin == InsertPos)
461 bool ScheduleDAGMI::checkSchedLimit() {
463 if (NumInstrsScheduled == MISchedCutoff && MISchedCutoff != ~0U) {
464 CurrentTop = CurrentBottom;
467 ++NumInstrsScheduled;
472 /// enterRegion - Called back from MachineScheduler::runOnMachineFunction after
473 /// crossing a scheduling boundary. [begin, end) includes all instructions in
474 /// the region, including the boundary itself and single-instruction regions
475 /// that don't get scheduled.
476 void ScheduleDAGMI::enterRegion(MachineBasicBlock *bb,
477 MachineBasicBlock::iterator begin,
478 MachineBasicBlock::iterator end,
479 unsigned regioninstrs)
481 ScheduleDAGInstrs::enterRegion(bb, begin, end, regioninstrs);
483 ShouldTrackPressure =
484 EnableRegPressure && SchedImpl->shouldTrackPressure(regioninstrs);
486 // For convenience remember the end of the liveness region.
488 (RegionEnd == bb->end()) ? RegionEnd : llvm::next(RegionEnd);
491 // Setup the register pressure trackers for the top scheduled top and bottom
492 // scheduled regions.
493 void ScheduleDAGMI::initRegPressure() {
494 TopRPTracker.init(&MF, RegClassInfo, LIS, BB, RegionBegin);
495 BotRPTracker.init(&MF, RegClassInfo, LIS, BB, LiveRegionEnd);
497 // Close the RPTracker to finalize live ins.
498 RPTracker.closeRegion();
500 DEBUG(RPTracker.dump());
502 // Initialize the live ins and live outs.
503 TopRPTracker.addLiveRegs(RPTracker.getPressure().LiveInRegs);
504 BotRPTracker.addLiveRegs(RPTracker.getPressure().LiveOutRegs);
506 // Close one end of the tracker so we can call
507 // getMaxUpward/DownwardPressureDelta before advancing across any
508 // instructions. This converts currently live regs into live ins/outs.
509 TopRPTracker.closeTop();
510 BotRPTracker.closeBottom();
512 BotRPTracker.initLiveThru(RPTracker);
513 if (!BotRPTracker.getLiveThru().empty()) {
514 TopRPTracker.initLiveThru(BotRPTracker.getLiveThru());
515 DEBUG(dbgs() << "Live Thru: ";
516 dumpRegSetPressure(BotRPTracker.getLiveThru(), TRI));
519 // For each live out vreg reduce the pressure change associated with other
520 // uses of the same vreg below the live-out reaching def.
521 updatePressureDiffs(RPTracker.getPressure().LiveOutRegs);
523 // Account for liveness generated by the region boundary.
524 if (LiveRegionEnd != RegionEnd) {
525 SmallVector<unsigned, 8> LiveUses;
526 BotRPTracker.recede(&LiveUses);
527 updatePressureDiffs(LiveUses);
530 assert(BotRPTracker.getPos() == RegionEnd && "Can't find the region bottom");
532 // Cache the list of excess pressure sets in this region. This will also track
533 // the max pressure in the scheduled code for these sets.
534 RegionCriticalPSets.clear();
535 const std::vector<unsigned> &RegionPressure =
536 RPTracker.getPressure().MaxSetPressure;
537 for (unsigned i = 0, e = RegionPressure.size(); i < e; ++i) {
538 unsigned Limit = RegClassInfo->getRegPressureSetLimit(i);
539 if (RegionPressure[i] > Limit) {
540 DEBUG(dbgs() << TRI->getRegPressureSetName(i)
541 << " Limit " << Limit
542 << " Actual " << RegionPressure[i] << "\n");
543 RegionCriticalPSets.push_back(PressureChange(i));
546 DEBUG(dbgs() << "Excess PSets: ";
547 for (unsigned i = 0, e = RegionCriticalPSets.size(); i != e; ++i)
548 dbgs() << TRI->getRegPressureSetName(
549 RegionCriticalPSets[i].getPSet()) << " ";
553 // FIXME: When the pressure tracker deals in pressure differences then we won't
554 // iterate over all RegionCriticalPSets[i].
556 updateScheduledPressure(const std::vector<unsigned> &NewMaxPressure) {
557 for (unsigned i = 0, e = RegionCriticalPSets.size(); i < e; ++i) {
558 unsigned ID = RegionCriticalPSets[i].getPSet();
559 if ((int)NewMaxPressure[ID] > RegionCriticalPSets[i].getUnitInc()
560 && NewMaxPressure[ID] <= INT16_MAX)
561 RegionCriticalPSets[i].setUnitInc(NewMaxPressure[ID]);
564 for (unsigned i = 0, e = NewMaxPressure.size(); i < e; ++i) {
565 unsigned Limit = RegClassInfo->getRegPressureSetLimit(i);
566 if (NewMaxPressure[i] > Limit ) {
567 dbgs() << " " << TRI->getRegPressureSetName(i) << ": "
568 << NewMaxPressure[i] << " > " << Limit << "\n";
573 /// Update the PressureDiff array for liveness after scheduling this
575 void ScheduleDAGMI::updatePressureDiffs(ArrayRef<unsigned> LiveUses) {
576 for (unsigned LUIdx = 0, LUEnd = LiveUses.size(); LUIdx != LUEnd; ++LUIdx) {
577 /// FIXME: Currently assuming single-use physregs.
578 unsigned Reg = LiveUses[LUIdx];
579 if (!TRI->isVirtualRegister(Reg))
581 // This may be called before CurrentBottom has been initialized. However,
582 // BotRPTracker must have a valid position. We want the value live into the
583 // instruction or live out of the block, so ask for the previous
584 // instruction's live-out.
585 const LiveInterval &LI = LIS->getInterval(Reg);
587 MachineBasicBlock::const_iterator I =
588 nextIfDebug(BotRPTracker.getPos(), BB->end());
590 VNI = LI.getVNInfoBefore(LIS->getMBBEndIdx(BB));
592 LiveRangeQuery LRQ(LI, LIS->getInstructionIndex(I));
595 // RegisterPressureTracker guarantees that readsReg is true for LiveUses.
596 assert(VNI && "No live value at use.");
597 for (VReg2UseMap::iterator
598 UI = VRegUses.find(Reg); UI != VRegUses.end(); ++UI) {
600 // If this use comes before the reaching def, it cannot be a last use, so
601 // descrease its pressure change.
602 if (!SU->isScheduled && SU != &ExitSU) {
603 LiveRangeQuery LRQ(LI, LIS->getInstructionIndex(SU->getInstr()));
604 if (LRQ.valueIn() == VNI)
605 getPressureDiff(SU).addPressureChange(Reg, true, &MRI);
611 /// schedule - Called back from MachineScheduler::runOnMachineFunction
612 /// after setting up the current scheduling region. [RegionBegin, RegionEnd)
613 /// only includes instructions that have DAG nodes, not scheduling boundaries.
615 /// This is a skeletal driver, with all the functionality pushed into helpers,
616 /// so that it can be easilly extended by experimental schedulers. Generally,
617 /// implementing MachineSchedStrategy should be sufficient to implement a new
618 /// scheduling algorithm. However, if a scheduler further subclasses
619 /// ScheduleDAGMI then it will want to override this virtual method in order to
620 /// update any specialized state.
621 void ScheduleDAGMI::schedule() {
622 buildDAGWithRegPressure();
624 Topo.InitDAGTopologicalSorting();
628 SmallVector<SUnit*, 8> TopRoots, BotRoots;
629 findRootsAndBiasEdges(TopRoots, BotRoots);
631 // Initialize the strategy before modifying the DAG.
632 // This may initialize a DFSResult to be used for queue priority.
633 SchedImpl->initialize(this);
635 DEBUG(for (unsigned su = 0, e = SUnits.size(); su != e; ++su)
636 SUnits[su].dumpAll(this));
637 if (ViewMISchedDAGs) viewGraph();
639 // Initialize ready queues now that the DAG and priority data are finalized.
640 initQueues(TopRoots, BotRoots);
642 bool IsTopNode = false;
643 while (SUnit *SU = SchedImpl->pickNode(IsTopNode)) {
644 assert(!SU->isScheduled && "Node already scheduled");
645 if (!checkSchedLimit())
648 scheduleMI(SU, IsTopNode);
650 updateQueues(SU, IsTopNode);
652 assert(CurrentTop == CurrentBottom && "Nonempty unscheduled zone.");
657 unsigned BBNum = begin()->getParent()->getNumber();
658 dbgs() << "*** Final schedule for BB#" << BBNum << " ***\n";
664 /// Build the DAG and setup three register pressure trackers.
665 void ScheduleDAGMI::buildDAGWithRegPressure() {
666 if (!ShouldTrackPressure) {
668 RegionCriticalPSets.clear();
673 // Initialize the register pressure tracker used by buildSchedGraph.
674 RPTracker.init(&MF, RegClassInfo, LIS, BB, LiveRegionEnd,
675 /*TrackUntiedDefs=*/true);
677 // Account for liveness generate by the region boundary.
678 if (LiveRegionEnd != RegionEnd)
681 // Build the DAG, and compute current register pressure.
682 buildSchedGraph(AA, &RPTracker, &SUPressureDiffs);
684 // Initialize top/bottom trackers after computing region pressure.
688 /// Apply each ScheduleDAGMutation step in order.
689 void ScheduleDAGMI::postprocessDAG() {
690 for (unsigned i = 0, e = Mutations.size(); i < e; ++i) {
691 Mutations[i]->apply(this);
695 void ScheduleDAGMI::computeDFSResult() {
697 DFSResult = new SchedDFSResult(/*BottomU*/true, MinSubtreeSize);
699 ScheduledTrees.clear();
700 DFSResult->resize(SUnits.size());
701 DFSResult->compute(SUnits);
702 ScheduledTrees.resize(DFSResult->getNumSubtrees());
705 void ScheduleDAGMI::findRootsAndBiasEdges(SmallVectorImpl<SUnit*> &TopRoots,
706 SmallVectorImpl<SUnit*> &BotRoots) {
707 for (std::vector<SUnit>::iterator
708 I = SUnits.begin(), E = SUnits.end(); I != E; ++I) {
710 assert(!SU->isBoundaryNode() && "Boundary node should not be in SUnits");
712 // Order predecessors so DFSResult follows the critical path.
713 SU->biasCriticalPath();
715 // A SUnit is ready to top schedule if it has no predecessors.
716 if (!I->NumPredsLeft)
717 TopRoots.push_back(SU);
718 // A SUnit is ready to bottom schedule if it has no successors.
719 if (!I->NumSuccsLeft)
720 BotRoots.push_back(SU);
722 ExitSU.biasCriticalPath();
725 /// Compute the max cyclic critical path through the DAG. The scheduling DAG
726 /// only provides the critical path for single block loops. To handle loops that
727 /// span blocks, we could use the vreg path latencies provided by
728 /// MachineTraceMetrics instead. However, MachineTraceMetrics is not currently
729 /// available for use in the scheduler.
731 /// The cyclic path estimation identifies a def-use pair that crosses the back
732 /// edge and considers the depth and height of the nodes. For example, consider
733 /// the following instruction sequence where each instruction has unit latency
734 /// and defines an epomymous virtual register:
736 /// a->b(a,c)->c(b)->d(c)->exit
738 /// The cyclic critical path is a two cycles: b->c->b
739 /// The acyclic critical path is four cycles: a->b->c->d->exit
740 /// LiveOutHeight = height(c) = len(c->d->exit) = 2
741 /// LiveOutDepth = depth(c) + 1 = len(a->b->c) + 1 = 3
742 /// LiveInHeight = height(b) + 1 = len(b->c->d->exit) + 1 = 4
743 /// LiveInDepth = depth(b) = len(a->b) = 1
745 /// LiveOutDepth - LiveInDepth = 3 - 1 = 2
746 /// LiveInHeight - LiveOutHeight = 4 - 2 = 2
747 /// CyclicCriticalPath = min(2, 2) = 2
748 unsigned ScheduleDAGMI::computeCyclicCriticalPath() {
749 // This only applies to single block loop.
750 if (!BB->isSuccessor(BB))
753 unsigned MaxCyclicLatency = 0;
754 // Visit each live out vreg def to find def/use pairs that cross iterations.
755 ArrayRef<unsigned> LiveOuts = RPTracker.getPressure().LiveOutRegs;
756 for (ArrayRef<unsigned>::iterator RI = LiveOuts.begin(), RE = LiveOuts.end();
759 if (!TRI->isVirtualRegister(Reg))
761 const LiveInterval &LI = LIS->getInterval(Reg);
762 const VNInfo *DefVNI = LI.getVNInfoBefore(LIS->getMBBEndIdx(BB));
766 MachineInstr *DefMI = LIS->getInstructionFromIndex(DefVNI->def);
767 const SUnit *DefSU = getSUnit(DefMI);
771 unsigned LiveOutHeight = DefSU->getHeight();
772 unsigned LiveOutDepth = DefSU->getDepth() + DefSU->Latency;
773 // Visit all local users of the vreg def.
774 for (VReg2UseMap::iterator
775 UI = VRegUses.find(Reg); UI != VRegUses.end(); ++UI) {
776 if (UI->SU == &ExitSU)
779 // Only consider uses of the phi.
780 LiveRangeQuery LRQ(LI, LIS->getInstructionIndex(UI->SU->getInstr()));
781 if (!LRQ.valueIn()->isPHIDef())
784 // Assume that a path spanning two iterations is a cycle, which could
785 // overestimate in strange cases. This allows cyclic latency to be
786 // estimated as the minimum slack of the vreg's depth or height.
787 unsigned CyclicLatency = 0;
788 if (LiveOutDepth > UI->SU->getDepth())
789 CyclicLatency = LiveOutDepth - UI->SU->getDepth();
791 unsigned LiveInHeight = UI->SU->getHeight() + DefSU->Latency;
792 if (LiveInHeight > LiveOutHeight) {
793 if (LiveInHeight - LiveOutHeight < CyclicLatency)
794 CyclicLatency = LiveInHeight - LiveOutHeight;
799 DEBUG(dbgs() << "Cyclic Path: SU(" << DefSU->NodeNum << ") -> SU("
800 << UI->SU->NodeNum << ") = " << CyclicLatency << "c\n");
801 if (CyclicLatency > MaxCyclicLatency)
802 MaxCyclicLatency = CyclicLatency;
805 DEBUG(dbgs() << "Cyclic Critical Path: " << MaxCyclicLatency << "c\n");
806 return MaxCyclicLatency;
809 /// Identify DAG roots and setup scheduler queues.
810 void ScheduleDAGMI::initQueues(ArrayRef<SUnit*> TopRoots,
811 ArrayRef<SUnit*> BotRoots) {
812 NextClusterSucc = NULL;
813 NextClusterPred = NULL;
815 // Release all DAG roots for scheduling, not including EntrySU/ExitSU.
817 // Nodes with unreleased weak edges can still be roots.
818 // Release top roots in forward order.
819 for (SmallVectorImpl<SUnit*>::const_iterator
820 I = TopRoots.begin(), E = TopRoots.end(); I != E; ++I) {
821 SchedImpl->releaseTopNode(*I);
823 // Release bottom roots in reverse order so the higher priority nodes appear
824 // first. This is more natural and slightly more efficient.
825 for (SmallVectorImpl<SUnit*>::const_reverse_iterator
826 I = BotRoots.rbegin(), E = BotRoots.rend(); I != E; ++I) {
827 SchedImpl->releaseBottomNode(*I);
830 releaseSuccessors(&EntrySU);
831 releasePredecessors(&ExitSU);
833 SchedImpl->registerRoots();
835 // Advance past initial DebugValues.
836 CurrentTop = nextIfDebug(RegionBegin, RegionEnd);
837 CurrentBottom = RegionEnd;
839 if (ShouldTrackPressure) {
840 assert(TopRPTracker.getPos() == RegionBegin && "bad initial Top tracker");
841 TopRPTracker.setPos(CurrentTop);
845 /// Move an instruction and update register pressure.
846 void ScheduleDAGMI::scheduleMI(SUnit *SU, bool IsTopNode) {
847 // Move the instruction to its new location in the instruction stream.
848 MachineInstr *MI = SU->getInstr();
851 assert(SU->isTopReady() && "node still has unscheduled dependencies");
852 if (&*CurrentTop == MI)
853 CurrentTop = nextIfDebug(++CurrentTop, CurrentBottom);
855 moveInstruction(MI, CurrentTop);
856 TopRPTracker.setPos(MI);
859 if (ShouldTrackPressure) {
860 // Update top scheduled pressure.
861 TopRPTracker.advance();
862 assert(TopRPTracker.getPos() == CurrentTop && "out of sync");
863 updateScheduledPressure(TopRPTracker.getPressure().MaxSetPressure);
867 assert(SU->isBottomReady() && "node still has unscheduled dependencies");
868 MachineBasicBlock::iterator priorII =
869 priorNonDebug(CurrentBottom, CurrentTop);
871 CurrentBottom = priorII;
873 if (&*CurrentTop == MI) {
874 CurrentTop = nextIfDebug(++CurrentTop, priorII);
875 TopRPTracker.setPos(CurrentTop);
877 moveInstruction(MI, CurrentBottom);
880 if (ShouldTrackPressure) {
881 // Update bottom scheduled pressure.
882 SmallVector<unsigned, 8> LiveUses;
883 BotRPTracker.recede(&LiveUses);
884 assert(BotRPTracker.getPos() == CurrentBottom && "out of sync");
885 updatePressureDiffs(LiveUses);
886 updateScheduledPressure(BotRPTracker.getPressure().MaxSetPressure);
891 /// Update scheduler queues after scheduling an instruction.
892 void ScheduleDAGMI::updateQueues(SUnit *SU, bool IsTopNode) {
893 // Release dependent instructions for scheduling.
895 releaseSuccessors(SU);
897 releasePredecessors(SU);
899 SU->isScheduled = true;
902 unsigned SubtreeID = DFSResult->getSubtreeID(SU);
903 if (!ScheduledTrees.test(SubtreeID)) {
904 ScheduledTrees.set(SubtreeID);
905 DFSResult->scheduleTree(SubtreeID);
906 SchedImpl->scheduleTree(SubtreeID);
910 // Notify the scheduling strategy after updating the DAG.
911 SchedImpl->schedNode(SU, IsTopNode);
914 /// Reinsert any remaining debug_values, just like the PostRA scheduler.
915 void ScheduleDAGMI::placeDebugValues() {
916 // If first instruction was a DBG_VALUE then put it back.
918 BB->splice(RegionBegin, BB, FirstDbgValue);
919 RegionBegin = FirstDbgValue;
922 for (std::vector<std::pair<MachineInstr *, MachineInstr *> >::iterator
923 DI = DbgValues.end(), DE = DbgValues.begin(); DI != DE; --DI) {
924 std::pair<MachineInstr *, MachineInstr *> P = *prior(DI);
925 MachineInstr *DbgValue = P.first;
926 MachineBasicBlock::iterator OrigPrevMI = P.second;
927 if (&*RegionBegin == DbgValue)
929 BB->splice(++OrigPrevMI, BB, DbgValue);
930 if (OrigPrevMI == llvm::prior(RegionEnd))
931 RegionEnd = DbgValue;
934 FirstDbgValue = NULL;
937 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
938 void ScheduleDAGMI::dumpSchedule() const {
939 for (MachineBasicBlock::iterator MI = begin(), ME = end(); MI != ME; ++MI) {
940 if (SUnit *SU = getSUnit(&(*MI)))
943 dbgs() << "Missing SUnit\n";
948 //===----------------------------------------------------------------------===//
949 // LoadClusterMutation - DAG post-processing to cluster loads.
950 //===----------------------------------------------------------------------===//
953 /// \brief Post-process the DAG to create cluster edges between neighboring
955 class LoadClusterMutation : public ScheduleDAGMutation {
960 LoadInfo(SUnit *su, unsigned reg, unsigned ofs)
961 : SU(su), BaseReg(reg), Offset(ofs) {}
963 static bool LoadInfoLess(const LoadClusterMutation::LoadInfo &LHS,
964 const LoadClusterMutation::LoadInfo &RHS);
966 const TargetInstrInfo *TII;
967 const TargetRegisterInfo *TRI;
969 LoadClusterMutation(const TargetInstrInfo *tii,
970 const TargetRegisterInfo *tri)
971 : TII(tii), TRI(tri) {}
973 virtual void apply(ScheduleDAGMI *DAG);
975 void clusterNeighboringLoads(ArrayRef<SUnit*> Loads, ScheduleDAGMI *DAG);
979 bool LoadClusterMutation::LoadInfoLess(
980 const LoadClusterMutation::LoadInfo &LHS,
981 const LoadClusterMutation::LoadInfo &RHS) {
982 if (LHS.BaseReg != RHS.BaseReg)
983 return LHS.BaseReg < RHS.BaseReg;
984 return LHS.Offset < RHS.Offset;
987 void LoadClusterMutation::clusterNeighboringLoads(ArrayRef<SUnit*> Loads,
988 ScheduleDAGMI *DAG) {
989 SmallVector<LoadClusterMutation::LoadInfo,32> LoadRecords;
990 for (unsigned Idx = 0, End = Loads.size(); Idx != End; ++Idx) {
991 SUnit *SU = Loads[Idx];
994 if (TII->getLdStBaseRegImmOfs(SU->getInstr(), BaseReg, Offset, TRI))
995 LoadRecords.push_back(LoadInfo(SU, BaseReg, Offset));
997 if (LoadRecords.size() < 2)
999 std::sort(LoadRecords.begin(), LoadRecords.end(), LoadInfoLess);
1000 unsigned ClusterLength = 1;
1001 for (unsigned Idx = 0, End = LoadRecords.size(); Idx < (End - 1); ++Idx) {
1002 if (LoadRecords[Idx].BaseReg != LoadRecords[Idx+1].BaseReg) {
1007 SUnit *SUa = LoadRecords[Idx].SU;
1008 SUnit *SUb = LoadRecords[Idx+1].SU;
1009 if (TII->shouldClusterLoads(SUa->getInstr(), SUb->getInstr(), ClusterLength)
1010 && DAG->addEdge(SUb, SDep(SUa, SDep::Cluster))) {
1012 DEBUG(dbgs() << "Cluster loads SU(" << SUa->NodeNum << ") - SU("
1013 << SUb->NodeNum << ")\n");
1014 // Copy successor edges from SUa to SUb. Interleaving computation
1015 // dependent on SUa can prevent load combining due to register reuse.
1016 // Predecessor edges do not need to be copied from SUb to SUa since nearby
1017 // loads should have effectively the same inputs.
1018 for (SUnit::const_succ_iterator
1019 SI = SUa->Succs.begin(), SE = SUa->Succs.end(); SI != SE; ++SI) {
1020 if (SI->getSUnit() == SUb)
1022 DEBUG(dbgs() << " Copy Succ SU(" << SI->getSUnit()->NodeNum << ")\n");
1023 DAG->addEdge(SI->getSUnit(), SDep(SUb, SDep::Artificial));
1032 /// \brief Callback from DAG postProcessing to create cluster edges for loads.
1033 void LoadClusterMutation::apply(ScheduleDAGMI *DAG) {
1034 // Map DAG NodeNum to store chain ID.
1035 DenseMap<unsigned, unsigned> StoreChainIDs;
1036 // Map each store chain to a set of dependent loads.
1037 SmallVector<SmallVector<SUnit*,4>, 32> StoreChainDependents;
1038 for (unsigned Idx = 0, End = DAG->SUnits.size(); Idx != End; ++Idx) {
1039 SUnit *SU = &DAG->SUnits[Idx];
1040 if (!SU->getInstr()->mayLoad())
1042 unsigned ChainPredID = DAG->SUnits.size();
1043 for (SUnit::const_pred_iterator
1044 PI = SU->Preds.begin(), PE = SU->Preds.end(); PI != PE; ++PI) {
1046 ChainPredID = PI->getSUnit()->NodeNum;
1050 // Check if this chain-like pred has been seen
1051 // before. ChainPredID==MaxNodeID for loads at the top of the schedule.
1052 unsigned NumChains = StoreChainDependents.size();
1053 std::pair<DenseMap<unsigned, unsigned>::iterator, bool> Result =
1054 StoreChainIDs.insert(std::make_pair(ChainPredID, NumChains));
1056 StoreChainDependents.resize(NumChains + 1);
1057 StoreChainDependents[Result.first->second].push_back(SU);
1059 // Iterate over the store chains.
1060 for (unsigned Idx = 0, End = StoreChainDependents.size(); Idx != End; ++Idx)
1061 clusterNeighboringLoads(StoreChainDependents[Idx], DAG);
1064 //===----------------------------------------------------------------------===//
1065 // MacroFusion - DAG post-processing to encourage fusion of macro ops.
1066 //===----------------------------------------------------------------------===//
1069 /// \brief Post-process the DAG to create cluster edges between instructions
1070 /// that may be fused by the processor into a single operation.
1071 class MacroFusion : public ScheduleDAGMutation {
1072 const TargetInstrInfo *TII;
1074 MacroFusion(const TargetInstrInfo *tii): TII(tii) {}
1076 virtual void apply(ScheduleDAGMI *DAG);
1080 /// \brief Callback from DAG postProcessing to create cluster edges to encourage
1081 /// fused operations.
1082 void MacroFusion::apply(ScheduleDAGMI *DAG) {
1083 // For now, assume targets can only fuse with the branch.
1084 MachineInstr *Branch = DAG->ExitSU.getInstr();
1088 for (unsigned Idx = DAG->SUnits.size(); Idx > 0;) {
1089 SUnit *SU = &DAG->SUnits[--Idx];
1090 if (!TII->shouldScheduleAdjacent(SU->getInstr(), Branch))
1093 // Create a single weak edge from SU to ExitSU. The only effect is to cause
1094 // bottom-up scheduling to heavily prioritize the clustered SU. There is no
1095 // need to copy predecessor edges from ExitSU to SU, since top-down
1096 // scheduling cannot prioritize ExitSU anyway. To defer top-down scheduling
1097 // of SU, we could create an artificial edge from the deepest root, but it
1098 // hasn't been needed yet.
1099 bool Success = DAG->addEdge(&DAG->ExitSU, SDep(SU, SDep::Cluster));
1101 assert(Success && "No DAG nodes should be reachable from ExitSU");
1103 DEBUG(dbgs() << "Macro Fuse SU(" << SU->NodeNum << ")\n");
1108 //===----------------------------------------------------------------------===//
1109 // CopyConstrain - DAG post-processing to encourage copy elimination.
1110 //===----------------------------------------------------------------------===//
1113 /// \brief Post-process the DAG to create weak edges from all uses of a copy to
1114 /// the one use that defines the copy's source vreg, most likely an induction
1115 /// variable increment.
1116 class CopyConstrain : public ScheduleDAGMutation {
1118 SlotIndex RegionBeginIdx;
1119 // RegionEndIdx is the slot index of the last non-debug instruction in the
1120 // scheduling region. So we may have RegionBeginIdx == RegionEndIdx.
1121 SlotIndex RegionEndIdx;
1123 CopyConstrain(const TargetInstrInfo *, const TargetRegisterInfo *) {}
1125 virtual void apply(ScheduleDAGMI *DAG);
1128 void constrainLocalCopy(SUnit *CopySU, ScheduleDAGMI *DAG);
1132 /// constrainLocalCopy handles two possibilities:
1137 /// I3: dst = src (copy)
1138 /// (create pred->succ edges I0->I1, I2->I1)
1141 /// I0: dst = src (copy)
1145 /// (create pred->succ edges I1->I2, I3->I2)
1147 /// Although the MachineScheduler is currently constrained to single blocks,
1148 /// this algorithm should handle extended blocks. An EBB is a set of
1149 /// contiguously numbered blocks such that the previous block in the EBB is
1150 /// always the single predecessor.
1151 void CopyConstrain::constrainLocalCopy(SUnit *CopySU, ScheduleDAGMI *DAG) {
1152 LiveIntervals *LIS = DAG->getLIS();
1153 MachineInstr *Copy = CopySU->getInstr();
1155 // Check for pure vreg copies.
1156 unsigned SrcReg = Copy->getOperand(1).getReg();
1157 if (!TargetRegisterInfo::isVirtualRegister(SrcReg))
1160 unsigned DstReg = Copy->getOperand(0).getReg();
1161 if (!TargetRegisterInfo::isVirtualRegister(DstReg))
1164 // Check if either the dest or source is local. If it's live across a back
1165 // edge, it's not local. Note that if both vregs are live across the back
1166 // edge, we cannot successfully contrain the copy without cyclic scheduling.
1167 unsigned LocalReg = DstReg;
1168 unsigned GlobalReg = SrcReg;
1169 LiveInterval *LocalLI = &LIS->getInterval(LocalReg);
1170 if (!LocalLI->isLocal(RegionBeginIdx, RegionEndIdx)) {
1173 LocalLI = &LIS->getInterval(LocalReg);
1174 if (!LocalLI->isLocal(RegionBeginIdx, RegionEndIdx))
1177 LiveInterval *GlobalLI = &LIS->getInterval(GlobalReg);
1179 // Find the global segment after the start of the local LI.
1180 LiveInterval::iterator GlobalSegment = GlobalLI->find(LocalLI->beginIndex());
1181 // If GlobalLI does not overlap LocalLI->start, then a copy directly feeds a
1182 // local live range. We could create edges from other global uses to the local
1183 // start, but the coalescer should have already eliminated these cases, so
1184 // don't bother dealing with it.
1185 if (GlobalSegment == GlobalLI->end())
1188 // If GlobalSegment is killed at the LocalLI->start, the call to find()
1189 // returned the next global segment. But if GlobalSegment overlaps with
1190 // LocalLI->start, then advance to the next segement. If a hole in GlobalLI
1191 // exists in LocalLI's vicinity, GlobalSegment will be the end of the hole.
1192 if (GlobalSegment->contains(LocalLI->beginIndex()))
1195 if (GlobalSegment == GlobalLI->end())
1198 // Check if GlobalLI contains a hole in the vicinity of LocalLI.
1199 if (GlobalSegment != GlobalLI->begin()) {
1200 // Two address defs have no hole.
1201 if (SlotIndex::isSameInstr(llvm::prior(GlobalSegment)->end,
1202 GlobalSegment->start)) {
1205 // If the prior global segment may be defined by the same two-address
1206 // instruction that also defines LocalLI, then can't make a hole here.
1207 if (SlotIndex::isSameInstr(llvm::prior(GlobalSegment)->start,
1208 LocalLI->beginIndex())) {
1211 // If GlobalLI has a prior segment, it must be live into the EBB. Otherwise
1212 // it would be a disconnected component in the live range.
1213 assert(llvm::prior(GlobalSegment)->start < LocalLI->beginIndex() &&
1214 "Disconnected LRG within the scheduling region.");
1216 MachineInstr *GlobalDef = LIS->getInstructionFromIndex(GlobalSegment->start);
1220 SUnit *GlobalSU = DAG->getSUnit(GlobalDef);
1224 // GlobalDef is the bottom of the GlobalLI hole. Open the hole by
1225 // constraining the uses of the last local def to precede GlobalDef.
1226 SmallVector<SUnit*,8> LocalUses;
1227 const VNInfo *LastLocalVN = LocalLI->getVNInfoBefore(LocalLI->endIndex());
1228 MachineInstr *LastLocalDef = LIS->getInstructionFromIndex(LastLocalVN->def);
1229 SUnit *LastLocalSU = DAG->getSUnit(LastLocalDef);
1230 for (SUnit::const_succ_iterator
1231 I = LastLocalSU->Succs.begin(), E = LastLocalSU->Succs.end();
1233 if (I->getKind() != SDep::Data || I->getReg() != LocalReg)
1235 if (I->getSUnit() == GlobalSU)
1237 if (!DAG->canAddEdge(GlobalSU, I->getSUnit()))
1239 LocalUses.push_back(I->getSUnit());
1241 // Open the top of the GlobalLI hole by constraining any earlier global uses
1242 // to precede the start of LocalLI.
1243 SmallVector<SUnit*,8> GlobalUses;
1244 MachineInstr *FirstLocalDef =
1245 LIS->getInstructionFromIndex(LocalLI->beginIndex());
1246 SUnit *FirstLocalSU = DAG->getSUnit(FirstLocalDef);
1247 for (SUnit::const_pred_iterator
1248 I = GlobalSU->Preds.begin(), E = GlobalSU->Preds.end(); I != E; ++I) {
1249 if (I->getKind() != SDep::Anti || I->getReg() != GlobalReg)
1251 if (I->getSUnit() == FirstLocalSU)
1253 if (!DAG->canAddEdge(FirstLocalSU, I->getSUnit()))
1255 GlobalUses.push_back(I->getSUnit());
1257 DEBUG(dbgs() << "Constraining copy SU(" << CopySU->NodeNum << ")\n");
1258 // Add the weak edges.
1259 for (SmallVectorImpl<SUnit*>::const_iterator
1260 I = LocalUses.begin(), E = LocalUses.end(); I != E; ++I) {
1261 DEBUG(dbgs() << " Local use SU(" << (*I)->NodeNum << ") -> SU("
1262 << GlobalSU->NodeNum << ")\n");
1263 DAG->addEdge(GlobalSU, SDep(*I, SDep::Weak));
1265 for (SmallVectorImpl<SUnit*>::const_iterator
1266 I = GlobalUses.begin(), E = GlobalUses.end(); I != E; ++I) {
1267 DEBUG(dbgs() << " Global use SU(" << (*I)->NodeNum << ") -> SU("
1268 << FirstLocalSU->NodeNum << ")\n");
1269 DAG->addEdge(FirstLocalSU, SDep(*I, SDep::Weak));
1273 /// \brief Callback from DAG postProcessing to create weak edges to encourage
1274 /// copy elimination.
1275 void CopyConstrain::apply(ScheduleDAGMI *DAG) {
1276 MachineBasicBlock::iterator FirstPos = nextIfDebug(DAG->begin(), DAG->end());
1277 if (FirstPos == DAG->end())
1279 RegionBeginIdx = DAG->getLIS()->getInstructionIndex(&*FirstPos);
1280 RegionEndIdx = DAG->getLIS()->getInstructionIndex(
1281 &*priorNonDebug(DAG->end(), DAG->begin()));
1283 for (unsigned Idx = 0, End = DAG->SUnits.size(); Idx != End; ++Idx) {
1284 SUnit *SU = &DAG->SUnits[Idx];
1285 if (!SU->getInstr()->isCopy())
1288 constrainLocalCopy(SU, DAG);
1292 //===----------------------------------------------------------------------===//
1293 // ConvergingScheduler - Implementation of the generic MachineSchedStrategy.
1294 //===----------------------------------------------------------------------===//
1297 /// ConvergingScheduler shrinks the unscheduled zone using heuristics to balance
1299 class ConvergingScheduler : public MachineSchedStrategy {
1301 /// Represent the type of SchedCandidate found within a single queue.
1302 /// pickNodeBidirectional depends on these listed by decreasing priority.
1304 NoCand, PhysRegCopy, RegExcess, RegCritical, Cluster, Weak, RegMax,
1305 ResourceReduce, ResourceDemand, BotHeightReduce, BotPathReduce,
1306 TopDepthReduce, TopPathReduce, NextDefUse, NodeOrder};
1309 static const char *getReasonStr(ConvergingScheduler::CandReason Reason);
1312 /// Policy for scheduling the next instruction in the candidate's zone.
1315 unsigned ReduceResIdx;
1316 unsigned DemandResIdx;
1318 CandPolicy(): ReduceLatency(false), ReduceResIdx(0), DemandResIdx(0) {}
1321 /// Status of an instruction's critical resource consumption.
1322 struct SchedResourceDelta {
1323 // Count critical resources in the scheduled region required by SU.
1324 unsigned CritResources;
1326 // Count critical resources from another region consumed by SU.
1327 unsigned DemandedResources;
1329 SchedResourceDelta(): CritResources(0), DemandedResources(0) {}
1331 bool operator==(const SchedResourceDelta &RHS) const {
1332 return CritResources == RHS.CritResources
1333 && DemandedResources == RHS.DemandedResources;
1335 bool operator!=(const SchedResourceDelta &RHS) const {
1336 return !operator==(RHS);
1340 /// Store the state used by ConvergingScheduler heuristics, required for the
1341 /// lifetime of one invocation of pickNode().
1342 struct SchedCandidate {
1345 // The best SUnit candidate.
1348 // The reason for this candidate.
1351 // Set of reasons that apply to multiple candidates.
1352 uint32_t RepeatReasonSet;
1354 // Register pressure values for the best candidate.
1355 RegPressureDelta RPDelta;
1357 // Critical resource consumption of the best candidate.
1358 SchedResourceDelta ResDelta;
1360 SchedCandidate(const CandPolicy &policy)
1361 : Policy(policy), SU(NULL), Reason(NoCand), RepeatReasonSet(0) {}
1363 bool isValid() const { return SU; }
1365 // Copy the status of another candidate without changing policy.
1366 void setBest(SchedCandidate &Best) {
1367 assert(Best.Reason != NoCand && "uninitialized Sched candidate");
1369 Reason = Best.Reason;
1370 RPDelta = Best.RPDelta;
1371 ResDelta = Best.ResDelta;
1374 bool isRepeat(CandReason R) { return RepeatReasonSet & (1 << R); }
1375 void setRepeat(CandReason R) { RepeatReasonSet |= (1 << R); }
1377 void initResourceDelta(const ScheduleDAGMI *DAG,
1378 const TargetSchedModel *SchedModel);
1381 /// Summarize the unscheduled region.
1382 struct SchedRemainder {
1383 // Critical path through the DAG in expected latency.
1384 unsigned CriticalPath;
1385 unsigned CyclicCritPath;
1387 // Scaled count of micro-ops left to schedule.
1388 unsigned RemIssueCount;
1390 bool IsAcyclicLatencyLimited;
1392 // Unscheduled resources
1393 SmallVector<unsigned, 16> RemainingCounts;
1399 IsAcyclicLatencyLimited = false;
1400 RemainingCounts.clear();
1403 SchedRemainder() { reset(); }
1405 void init(ScheduleDAGMI *DAG, const TargetSchedModel *SchedModel);
1408 /// Each Scheduling boundary is associated with ready queues. It tracks the
1409 /// current cycle in the direction of movement, and maintains the state
1410 /// of "hazards" and other interlocks at the current cycle.
1411 struct SchedBoundary {
1413 const TargetSchedModel *SchedModel;
1414 SchedRemainder *Rem;
1416 ReadyQueue Available;
1420 // For heuristics, keep a list of the nodes that immediately depend on the
1421 // most recently scheduled node.
1422 SmallPtrSet<const SUnit*, 8> NextSUs;
1424 ScheduleHazardRecognizer *HazardRec;
1426 /// Number of cycles it takes to issue the instructions scheduled in this
1427 /// zone. It is defined as: scheduled-micro-ops / issue-width + stalls.
1428 /// See getStalls().
1431 /// Micro-ops issued in the current cycle
1434 /// MinReadyCycle - Cycle of the soonest available instruction.
1435 unsigned MinReadyCycle;
1437 // The expected latency of the critical path in this scheduled zone.
1438 unsigned ExpectedLatency;
1440 // The latency of dependence chains leading into this zone.
1441 // For each node scheduled bottom-up: DLat = max DLat, N.Depth.
1442 // For each cycle scheduled: DLat -= 1.
1443 unsigned DependentLatency;
1445 /// Count the scheduled (issued) micro-ops that can be retired by
1446 /// time=CurrCycle assuming the first scheduled instr is retired at time=0.
1447 unsigned RetiredMOps;
1449 // Count scheduled resources that have been executed. Resources are
1450 // considered executed if they become ready in the time that it takes to
1451 // saturate any resource including the one in question. Counts are scaled
1452 // for direct comparison with other resources. Counts can be compared with
1453 // MOps * getMicroOpFactor and Latency * getLatencyFactor.
1454 SmallVector<unsigned, 16> ExecutedResCounts;
1456 /// Cache the max count for a single resource.
1457 unsigned MaxExecutedResCount;
1459 // Cache the critical resources ID in this scheduled zone.
1460 unsigned ZoneCritResIdx;
1462 // Is the scheduled region resource limited vs. latency limited.
1463 bool IsResourceLimited;
1466 // Remember the greatest operand latency as an upper bound on the number of
1467 // times we should retry the pending queue because of a hazard.
1468 unsigned MaxObservedLatency;
1472 // A new HazardRec is created for each DAG and owned by SchedBoundary.
1473 // Destroying and reconstructing it is very expensive though. So keep
1474 // invalid, placeholder HazardRecs.
1475 if (HazardRec && HazardRec->isEnabled()) {
1481 CheckPending = false;
1485 MinReadyCycle = UINT_MAX;
1486 ExpectedLatency = 0;
1487 DependentLatency = 0;
1489 MaxExecutedResCount = 0;
1491 IsResourceLimited = false;
1493 MaxObservedLatency = 0;
1495 // Reserve a zero-count for invalid CritResIdx.
1496 ExecutedResCounts.resize(1);
1497 assert(!ExecutedResCounts[0] && "nonzero count for bad resource");
1500 /// Pending queues extend the ready queues with the same ID and the
1501 /// PendingFlag set.
1502 SchedBoundary(unsigned ID, const Twine &Name):
1503 DAG(0), SchedModel(0), Rem(0), Available(ID, Name+".A"),
1504 Pending(ID << ConvergingScheduler::LogMaxQID, Name+".P"),
1509 ~SchedBoundary() { delete HazardRec; }
1511 void init(ScheduleDAGMI *dag, const TargetSchedModel *smodel,
1512 SchedRemainder *rem);
1514 bool isTop() const {
1515 return Available.getID() == ConvergingScheduler::TopQID;
1519 const char *getResourceName(unsigned PIdx) {
1522 return SchedModel->getProcResource(PIdx)->Name;
1526 /// Get the number of latency cycles "covered" by the scheduled
1527 /// instructions. This is the larger of the critical path within the zone
1528 /// and the number of cycles required to issue the instructions.
1529 unsigned getScheduledLatency() const {
1530 return std::max(ExpectedLatency, CurrCycle);
1533 unsigned getUnscheduledLatency(SUnit *SU) const {
1534 return isTop() ? SU->getHeight() : SU->getDepth();
1537 unsigned getResourceCount(unsigned ResIdx) const {
1538 return ExecutedResCounts[ResIdx];
1541 /// Get the scaled count of scheduled micro-ops and resources, including
1542 /// executed resources.
1543 unsigned getCriticalCount() const {
1544 if (!ZoneCritResIdx)
1545 return RetiredMOps * SchedModel->getMicroOpFactor();
1546 return getResourceCount(ZoneCritResIdx);
1549 /// Get a scaled count for the minimum execution time of the scheduled
1550 /// micro-ops that are ready to execute by getExecutedCount. Notice the
1552 unsigned getExecutedCount() const {
1553 return std::max(CurrCycle * SchedModel->getLatencyFactor(),
1554 MaxExecutedResCount);
1557 bool checkHazard(SUnit *SU);
1559 unsigned findMaxLatency(ArrayRef<SUnit*> ReadySUs);
1561 unsigned getOtherResourceCount(unsigned &OtherCritIdx);
1563 void setPolicy(CandPolicy &Policy, SchedBoundary &OtherZone);
1565 void releaseNode(SUnit *SU, unsigned ReadyCycle);
1567 void bumpCycle(unsigned NextCycle);
1569 void incExecutedResources(unsigned PIdx, unsigned Count);
1571 unsigned countResource(unsigned PIdx, unsigned Cycles, unsigned ReadyCycle);
1573 void bumpNode(SUnit *SU);
1575 void releasePending();
1577 void removeReady(SUnit *SU);
1579 SUnit *pickOnlyChoice();
1582 void dumpScheduledState();
1587 const MachineSchedContext *Context;
1589 const TargetSchedModel *SchedModel;
1590 const TargetRegisterInfo *TRI;
1592 // State of the top and bottom scheduled instruction boundaries.
1597 // Allow the driver to force top-down or bottom-up scheduling. If neither is
1598 // true, the scheduler runs in both directions and converges. For generic
1599 // targets, we default to bottom-up, because it's simpler and more
1600 // compile-time optimizations have been implemented in that direction.
1604 /// SUnit::NodeQueueId: 0 (none), 1 (top), 2 (bot), 3 (both)
1611 ConvergingScheduler(const MachineSchedContext *C):
1612 Context(C), DAG(0), SchedModel(0), TRI(0),
1613 Top(TopQID, "TopQ"), Bot(BotQID, "BotQ"),
1614 OnlyBottomUp(true), OnlyTopDown(false) {}
1616 virtual bool shouldTrackPressure(unsigned NumRegionInstrs);
1618 virtual void initialize(ScheduleDAGMI *dag);
1620 virtual SUnit *pickNode(bool &IsTopNode);
1622 virtual void schedNode(SUnit *SU, bool IsTopNode);
1624 virtual void releaseTopNode(SUnit *SU);
1626 virtual void releaseBottomNode(SUnit *SU);
1628 virtual void registerRoots();
1631 void checkAcyclicLatency();
1633 void tryCandidate(SchedCandidate &Cand,
1634 SchedCandidate &TryCand,
1635 SchedBoundary &Zone,
1636 const RegPressureTracker &RPTracker,
1637 RegPressureTracker &TempTracker);
1639 SUnit *pickNodeBidirectional(bool &IsTopNode);
1641 void pickNodeFromQueue(SchedBoundary &Zone,
1642 const RegPressureTracker &RPTracker,
1643 SchedCandidate &Candidate);
1645 void reschedulePhysRegCopies(SUnit *SU, bool isTop);
1648 void traceCandidate(const SchedCandidate &Cand);
1653 void ConvergingScheduler::SchedRemainder::
1654 init(ScheduleDAGMI *DAG, const TargetSchedModel *SchedModel) {
1656 if (!SchedModel->hasInstrSchedModel())
1658 RemainingCounts.resize(SchedModel->getNumProcResourceKinds());
1659 for (std::vector<SUnit>::iterator
1660 I = DAG->SUnits.begin(), E = DAG->SUnits.end(); I != E; ++I) {
1661 const MCSchedClassDesc *SC = DAG->getSchedClass(&*I);
1662 RemIssueCount += SchedModel->getNumMicroOps(I->getInstr(), SC)
1663 * SchedModel->getMicroOpFactor();
1664 for (TargetSchedModel::ProcResIter
1665 PI = SchedModel->getWriteProcResBegin(SC),
1666 PE = SchedModel->getWriteProcResEnd(SC); PI != PE; ++PI) {
1667 unsigned PIdx = PI->ProcResourceIdx;
1668 unsigned Factor = SchedModel->getResourceFactor(PIdx);
1669 RemainingCounts[PIdx] += (Factor * PI->Cycles);
1674 void ConvergingScheduler::SchedBoundary::
1675 init(ScheduleDAGMI *dag, const TargetSchedModel *smodel, SchedRemainder *rem) {
1678 SchedModel = smodel;
1680 if (SchedModel->hasInstrSchedModel())
1681 ExecutedResCounts.resize(SchedModel->getNumProcResourceKinds());
1684 /// Avoid setting up the register pressure tracker for small regions to save
1685 /// compile time. As a rough heuristic, only track pressure when the number
1686 /// of schedulable instructions exceeds half the integer register file.
1687 bool ConvergingScheduler::shouldTrackPressure(unsigned NumRegionInstrs) {
1688 unsigned NIntRegs = Context->RegClassInfo->getNumAllocatableRegs(
1689 Context->MF->getTarget().getTargetLowering()->getRegClassFor(MVT::i32));
1691 return NumRegionInstrs > (NIntRegs / 2);
1694 void ConvergingScheduler::initialize(ScheduleDAGMI *dag) {
1696 SchedModel = DAG->getSchedModel();
1699 Rem.init(DAG, SchedModel);
1700 Top.init(DAG, SchedModel, &Rem);
1701 Bot.init(DAG, SchedModel, &Rem);
1703 // Initialize resource counts.
1705 // Initialize the HazardRecognizers. If itineraries don't exist, are empty, or
1706 // are disabled, then these HazardRecs will be disabled.
1707 const InstrItineraryData *Itin = SchedModel->getInstrItineraries();
1708 const TargetMachine &TM = DAG->MF.getTarget();
1709 if (!Top.HazardRec) {
1711 TM.getInstrInfo()->CreateTargetMIHazardRecognizer(Itin, DAG);
1713 if (!Bot.HazardRec) {
1715 TM.getInstrInfo()->CreateTargetMIHazardRecognizer(Itin, DAG);
1717 assert((!ForceTopDown || !ForceBottomUp) &&
1718 "-misched-topdown incompatible with -misched-bottomup");
1720 // Check -misched-topdown/bottomup can force or unforce scheduling direction.
1721 // e.g. -misched-bottomup=false allows scheduling in both directions.
1722 if (ForceBottomUp.getNumOccurrences() > 0) {
1723 OnlyBottomUp = ForceBottomUp;
1725 OnlyTopDown = false;
1727 if (ForceTopDown.getNumOccurrences() > 0) {
1728 OnlyTopDown = ForceTopDown;
1730 OnlyBottomUp = false;
1734 void ConvergingScheduler::releaseTopNode(SUnit *SU) {
1735 if (SU->isScheduled)
1738 for (SUnit::pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
1742 unsigned PredReadyCycle = I->getSUnit()->TopReadyCycle;
1743 unsigned Latency = I->getLatency();
1745 Top.MaxObservedLatency = std::max(Latency, Top.MaxObservedLatency);
1747 if (SU->TopReadyCycle < PredReadyCycle + Latency)
1748 SU->TopReadyCycle = PredReadyCycle + Latency;
1750 Top.releaseNode(SU, SU->TopReadyCycle);
1753 void ConvergingScheduler::releaseBottomNode(SUnit *SU) {
1754 if (SU->isScheduled)
1757 assert(SU->getInstr() && "Scheduled SUnit must have instr");
1759 for (SUnit::succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
1763 unsigned SuccReadyCycle = I->getSUnit()->BotReadyCycle;
1764 unsigned Latency = I->getLatency();
1766 Bot.MaxObservedLatency = std::max(Latency, Bot.MaxObservedLatency);
1768 if (SU->BotReadyCycle < SuccReadyCycle + Latency)
1769 SU->BotReadyCycle = SuccReadyCycle + Latency;
1771 Bot.releaseNode(SU, SU->BotReadyCycle);
1774 /// Set IsAcyclicLatencyLimited if the acyclic path is longer than the cyclic
1775 /// critical path by more cycles than it takes to drain the instruction buffer.
1776 /// We estimate an upper bounds on in-flight instructions as:
1778 /// CyclesPerIteration = max( CyclicPath, Loop-Resource-Height )
1779 /// InFlightIterations = AcyclicPath / CyclesPerIteration
1780 /// InFlightResources = InFlightIterations * LoopResources
1782 /// TODO: Check execution resources in addition to IssueCount.
1783 void ConvergingScheduler::checkAcyclicLatency() {
1784 if (Rem.CyclicCritPath == 0 || Rem.CyclicCritPath >= Rem.CriticalPath)
1787 // Scaled number of cycles per loop iteration.
1788 unsigned IterCount =
1789 std::max(Rem.CyclicCritPath * SchedModel->getLatencyFactor(),
1791 // Scaled acyclic critical path.
1792 unsigned AcyclicCount = Rem.CriticalPath * SchedModel->getLatencyFactor();
1793 // InFlightCount = (AcyclicPath / IterCycles) * InstrPerLoop
1794 unsigned InFlightCount =
1795 (AcyclicCount * Rem.RemIssueCount + IterCount-1) / IterCount;
1796 unsigned BufferLimit =
1797 SchedModel->getMicroOpBufferSize() * SchedModel->getMicroOpFactor();
1799 Rem.IsAcyclicLatencyLimited = InFlightCount > BufferLimit;
1801 DEBUG(dbgs() << "IssueCycles="
1802 << Rem.RemIssueCount / SchedModel->getLatencyFactor() << "c "
1803 << "IterCycles=" << IterCount / SchedModel->getLatencyFactor()
1804 << "c NumIters=" << (AcyclicCount + IterCount-1) / IterCount
1805 << " InFlight=" << InFlightCount / SchedModel->getMicroOpFactor()
1806 << "m BufferLim=" << SchedModel->getMicroOpBufferSize() << "m\n";
1807 if (Rem.IsAcyclicLatencyLimited)
1808 dbgs() << " ACYCLIC LATENCY LIMIT\n");
1811 void ConvergingScheduler::registerRoots() {
1812 Rem.CriticalPath = DAG->ExitSU.getDepth();
1814 // Some roots may not feed into ExitSU. Check all of them in case.
1815 for (std::vector<SUnit*>::const_iterator
1816 I = Bot.Available.begin(), E = Bot.Available.end(); I != E; ++I) {
1817 if ((*I)->getDepth() > Rem.CriticalPath)
1818 Rem.CriticalPath = (*I)->getDepth();
1820 DEBUG(dbgs() << "Critical Path: " << Rem.CriticalPath << '\n');
1822 if (EnableCyclicPath) {
1823 Rem.CyclicCritPath = DAG->computeCyclicCriticalPath();
1824 checkAcyclicLatency();
1828 /// Does this SU have a hazard within the current instruction group.
1830 /// The scheduler supports two modes of hazard recognition. The first is the
1831 /// ScheduleHazardRecognizer API. It is a fully general hazard recognizer that
1832 /// supports highly complicated in-order reservation tables
1833 /// (ScoreboardHazardRecognizer) and arbitraty target-specific logic.
1835 /// The second is a streamlined mechanism that checks for hazards based on
1836 /// simple counters that the scheduler itself maintains. It explicitly checks
1837 /// for instruction dispatch limitations, including the number of micro-ops that
1838 /// can dispatch per cycle.
1840 /// TODO: Also check whether the SU must start a new group.
1841 bool ConvergingScheduler::SchedBoundary::checkHazard(SUnit *SU) {
1842 if (HazardRec->isEnabled())
1843 return HazardRec->getHazardType(SU) != ScheduleHazardRecognizer::NoHazard;
1845 unsigned uops = SchedModel->getNumMicroOps(SU->getInstr());
1846 if ((CurrMOps > 0) && (CurrMOps + uops > SchedModel->getIssueWidth())) {
1847 DEBUG(dbgs() << " SU(" << SU->NodeNum << ") uops="
1848 << SchedModel->getNumMicroOps(SU->getInstr()) << '\n');
1854 // Find the unscheduled node in ReadySUs with the highest latency.
1855 unsigned ConvergingScheduler::SchedBoundary::
1856 findMaxLatency(ArrayRef<SUnit*> ReadySUs) {
1858 unsigned RemLatency = 0;
1859 for (ArrayRef<SUnit*>::iterator I = ReadySUs.begin(), E = ReadySUs.end();
1861 unsigned L = getUnscheduledLatency(*I);
1862 if (L > RemLatency) {
1868 DEBUG(dbgs() << Available.getName() << " RemLatency SU("
1869 << LateSU->NodeNum << ") " << RemLatency << "c\n");
1874 // Count resources in this zone and the remaining unscheduled
1875 // instruction. Return the max count, scaled. Set OtherCritIdx to the critical
1876 // resource index, or zero if the zone is issue limited.
1877 unsigned ConvergingScheduler::SchedBoundary::
1878 getOtherResourceCount(unsigned &OtherCritIdx) {
1880 if (!SchedModel->hasInstrSchedModel())
1883 unsigned OtherCritCount = Rem->RemIssueCount
1884 + (RetiredMOps * SchedModel->getMicroOpFactor());
1885 DEBUG(dbgs() << " " << Available.getName() << " + Remain MOps: "
1886 << OtherCritCount / SchedModel->getMicroOpFactor() << '\n');
1887 for (unsigned PIdx = 1, PEnd = SchedModel->getNumProcResourceKinds();
1888 PIdx != PEnd; ++PIdx) {
1889 unsigned OtherCount = getResourceCount(PIdx) + Rem->RemainingCounts[PIdx];
1890 if (OtherCount > OtherCritCount) {
1891 OtherCritCount = OtherCount;
1892 OtherCritIdx = PIdx;
1896 DEBUG(dbgs() << " " << Available.getName() << " + Remain CritRes: "
1897 << OtherCritCount / SchedModel->getResourceFactor(OtherCritIdx)
1898 << " " << getResourceName(OtherCritIdx) << "\n");
1900 return OtherCritCount;
1903 /// Set the CandPolicy for this zone given the current resources and latencies
1904 /// inside and outside the zone.
1905 void ConvergingScheduler::SchedBoundary::setPolicy(CandPolicy &Policy,
1906 SchedBoundary &OtherZone) {
1907 // Now that potential stalls have been considered, apply preemptive heuristics
1908 // based on the the total latency and resources inside and outside this
1911 // Compute remaining latency. We need this both to determine whether the
1912 // overall schedule has become latency-limited and whether the instructions
1913 // outside this zone are resource or latency limited.
1915 // The "dependent" latency is updated incrementally during scheduling as the
1916 // max height/depth of scheduled nodes minus the cycles since it was
1918 // DLat = max (N.depth - (CurrCycle - N.ReadyCycle) for N in Zone
1920 // The "independent" latency is the max ready queue depth:
1921 // ILat = max N.depth for N in Available|Pending
1923 // RemainingLatency is the greater of independent and dependent latency.
1924 unsigned RemLatency = DependentLatency;
1925 RemLatency = std::max(RemLatency, findMaxLatency(Available.elements()));
1926 RemLatency = std::max(RemLatency, findMaxLatency(Pending.elements()));
1928 // Compute the critical resource outside the zone.
1929 unsigned OtherCritIdx;
1930 unsigned OtherCount = OtherZone.getOtherResourceCount(OtherCritIdx);
1932 bool OtherResLimited = false;
1933 if (SchedModel->hasInstrSchedModel()) {
1934 unsigned LFactor = SchedModel->getLatencyFactor();
1935 OtherResLimited = (int)(OtherCount - (RemLatency * LFactor)) > (int)LFactor;
1937 if (!OtherResLimited && (RemLatency + CurrCycle > Rem->CriticalPath)) {
1938 Policy.ReduceLatency |= true;
1939 DEBUG(dbgs() << " " << Available.getName() << " RemainingLatency "
1940 << RemLatency << " + " << CurrCycle << "c > CritPath "
1941 << Rem->CriticalPath << "\n");
1943 // If the same resource is limiting inside and outside the zone, do nothing.
1944 if (ZoneCritResIdx == OtherCritIdx)
1948 if (IsResourceLimited) {
1949 dbgs() << " " << Available.getName() << " ResourceLimited: "
1950 << getResourceName(ZoneCritResIdx) << "\n";
1952 if (OtherResLimited)
1953 dbgs() << " RemainingLimit: " << getResourceName(OtherCritIdx) << "\n";
1954 if (!IsResourceLimited && !OtherResLimited)
1955 dbgs() << " Latency limited both directions.\n");
1957 if (IsResourceLimited && !Policy.ReduceResIdx)
1958 Policy.ReduceResIdx = ZoneCritResIdx;
1960 if (OtherResLimited)
1961 Policy.DemandResIdx = OtherCritIdx;
1964 void ConvergingScheduler::SchedBoundary::releaseNode(SUnit *SU,
1965 unsigned ReadyCycle) {
1966 if (ReadyCycle < MinReadyCycle)
1967 MinReadyCycle = ReadyCycle;
1969 // Check for interlocks first. For the purpose of other heuristics, an
1970 // instruction that cannot issue appears as if it's not in the ReadyQueue.
1971 bool IsBuffered = SchedModel->getMicroOpBufferSize() != 0;
1972 if ((!IsBuffered && ReadyCycle > CurrCycle) || checkHazard(SU))
1977 // Record this node as an immediate dependent of the scheduled node.
1981 /// Move the boundary of scheduled code by one cycle.
1982 void ConvergingScheduler::SchedBoundary::bumpCycle(unsigned NextCycle) {
1983 if (SchedModel->getMicroOpBufferSize() == 0) {
1984 assert(MinReadyCycle < UINT_MAX && "MinReadyCycle uninitialized");
1985 if (MinReadyCycle > NextCycle)
1986 NextCycle = MinReadyCycle;
1988 // Update the current micro-ops, which will issue in the next cycle.
1989 unsigned DecMOps = SchedModel->getIssueWidth() * (NextCycle - CurrCycle);
1990 CurrMOps = (CurrMOps <= DecMOps) ? 0 : CurrMOps - DecMOps;
1992 // Decrement DependentLatency based on the next cycle.
1993 if ((NextCycle - CurrCycle) > DependentLatency)
1994 DependentLatency = 0;
1996 DependentLatency -= (NextCycle - CurrCycle);
1998 if (!HazardRec->isEnabled()) {
1999 // Bypass HazardRec virtual calls.
2000 CurrCycle = NextCycle;
2003 // Bypass getHazardType calls in case of long latency.
2004 for (; CurrCycle != NextCycle; ++CurrCycle) {
2006 HazardRec->AdvanceCycle();
2008 HazardRec->RecedeCycle();
2011 CheckPending = true;
2012 unsigned LFactor = SchedModel->getLatencyFactor();
2014 (int)(getCriticalCount() - (getScheduledLatency() * LFactor))
2017 DEBUG(dbgs() << "Cycle: " << CurrCycle << ' ' << Available.getName() << '\n');
2020 void ConvergingScheduler::SchedBoundary::incExecutedResources(unsigned PIdx,
2022 ExecutedResCounts[PIdx] += Count;
2023 if (ExecutedResCounts[PIdx] > MaxExecutedResCount)
2024 MaxExecutedResCount = ExecutedResCounts[PIdx];
2027 /// Add the given processor resource to this scheduled zone.
2029 /// \param Cycles indicates the number of consecutive (non-pipelined) cycles
2030 /// during which this resource is consumed.
2032 /// \return the next cycle at which the instruction may execute without
2033 /// oversubscribing resources.
2034 unsigned ConvergingScheduler::SchedBoundary::
2035 countResource(unsigned PIdx, unsigned Cycles, unsigned ReadyCycle) {
2036 unsigned Factor = SchedModel->getResourceFactor(PIdx);
2037 unsigned Count = Factor * Cycles;
2038 DEBUG(dbgs() << " " << getResourceName(PIdx)
2039 << " +" << Cycles << "x" << Factor << "u\n");
2041 // Update Executed resources counts.
2042 incExecutedResources(PIdx, Count);
2043 assert(Rem->RemainingCounts[PIdx] >= Count && "resource double counted");
2044 Rem->RemainingCounts[PIdx] -= Count;
2046 // Check if this resource exceeds the current critical resource. If so, it
2047 // becomes the critical resource.
2048 if (ZoneCritResIdx != PIdx && (getResourceCount(PIdx) > getCriticalCount())) {
2049 ZoneCritResIdx = PIdx;
2050 DEBUG(dbgs() << " *** Critical resource "
2051 << getResourceName(PIdx) << ": "
2052 << getResourceCount(PIdx) / SchedModel->getLatencyFactor() << "c\n");
2054 // TODO: We don't yet model reserved resources. It's not hard though.
2058 /// Move the boundary of scheduled code by one SUnit.
2059 void ConvergingScheduler::SchedBoundary::bumpNode(SUnit *SU) {
2060 // Update the reservation table.
2061 if (HazardRec->isEnabled()) {
2062 if (!isTop() && SU->isCall) {
2063 // Calls are scheduled with their preceding instructions. For bottom-up
2064 // scheduling, clear the pipeline state before emitting.
2067 HazardRec->EmitInstruction(SU);
2069 const MCSchedClassDesc *SC = DAG->getSchedClass(SU);
2070 unsigned IncMOps = SchedModel->getNumMicroOps(SU->getInstr());
2071 CurrMOps += IncMOps;
2072 // checkHazard prevents scheduling multiple instructions per cycle that exceed
2073 // issue width. However, we commonly reach the maximum. In this case
2074 // opportunistically bump the cycle to avoid uselessly checking everything in
2075 // the readyQ. Furthermore, a single instruction may produce more than one
2076 // cycle's worth of micro-ops.
2078 // TODO: Also check if this SU must end a dispatch group.
2079 unsigned NextCycle = CurrCycle;
2080 if (CurrMOps >= SchedModel->getIssueWidth()) {
2082 DEBUG(dbgs() << " *** Max MOps " << CurrMOps
2083 << " at cycle " << CurrCycle << '\n');
2085 unsigned ReadyCycle = (isTop() ? SU->TopReadyCycle : SU->BotReadyCycle);
2086 DEBUG(dbgs() << " Ready @" << ReadyCycle << "c\n");
2088 switch (SchedModel->getMicroOpBufferSize()) {
2090 assert(ReadyCycle <= CurrCycle && "Broken PendingQueue");
2093 if (ReadyCycle > NextCycle) {
2094 NextCycle = ReadyCycle;
2095 DEBUG(dbgs() << " *** Stall until: " << ReadyCycle << "\n");
2099 // We don't currently model the OOO reorder buffer, so consider all
2100 // scheduled MOps to be "retired".
2103 RetiredMOps += IncMOps;
2105 // Update resource counts and critical resource.
2106 if (SchedModel->hasInstrSchedModel()) {
2107 unsigned DecRemIssue = IncMOps * SchedModel->getMicroOpFactor();
2108 assert(Rem->RemIssueCount >= DecRemIssue && "MOps double counted");
2109 Rem->RemIssueCount -= DecRemIssue;
2110 if (ZoneCritResIdx) {
2111 // Scale scheduled micro-ops for comparing with the critical resource.
2112 unsigned ScaledMOps =
2113 RetiredMOps * SchedModel->getMicroOpFactor();
2115 // If scaled micro-ops are now more than the previous critical resource by
2116 // a full cycle, then micro-ops issue becomes critical.
2117 if ((int)(ScaledMOps - getResourceCount(ZoneCritResIdx))
2118 >= (int)SchedModel->getLatencyFactor()) {
2120 DEBUG(dbgs() << " *** Critical resource NumMicroOps: "
2121 << ScaledMOps / SchedModel->getLatencyFactor() << "c\n");
2124 for (TargetSchedModel::ProcResIter
2125 PI = SchedModel->getWriteProcResBegin(SC),
2126 PE = SchedModel->getWriteProcResEnd(SC); PI != PE; ++PI) {
2128 countResource(PI->ProcResourceIdx, PI->Cycles, ReadyCycle);
2129 if (RCycle > NextCycle)
2133 // Update ExpectedLatency and DependentLatency.
2134 unsigned &TopLatency = isTop() ? ExpectedLatency : DependentLatency;
2135 unsigned &BotLatency = isTop() ? DependentLatency : ExpectedLatency;
2136 if (SU->getDepth() > TopLatency) {
2137 TopLatency = SU->getDepth();
2138 DEBUG(dbgs() << " " << Available.getName()
2139 << " TopLatency SU(" << SU->NodeNum << ") " << TopLatency << "c\n");
2141 if (SU->getHeight() > BotLatency) {
2142 BotLatency = SU->getHeight();
2143 DEBUG(dbgs() << " " << Available.getName()
2144 << " BotLatency SU(" << SU->NodeNum << ") " << BotLatency << "c\n");
2146 // If we stall for any reason, bump the cycle.
2147 if (NextCycle > CurrCycle) {
2148 bumpCycle(NextCycle);
2151 // After updating ZoneCritResIdx and ExpectedLatency, check if we're
2152 // resource limited. If a stall occured, bumpCycle does this.
2153 unsigned LFactor = SchedModel->getLatencyFactor();
2155 (int)(getCriticalCount() - (getScheduledLatency() * LFactor))
2158 DEBUG(dumpScheduledState());
2161 /// Release pending ready nodes in to the available queue. This makes them
2162 /// visible to heuristics.
2163 void ConvergingScheduler::SchedBoundary::releasePending() {
2164 // If the available queue is empty, it is safe to reset MinReadyCycle.
2165 if (Available.empty())
2166 MinReadyCycle = UINT_MAX;
2168 // Check to see if any of the pending instructions are ready to issue. If
2169 // so, add them to the available queue.
2170 bool IsBuffered = SchedModel->getMicroOpBufferSize() != 0;
2171 for (unsigned i = 0, e = Pending.size(); i != e; ++i) {
2172 SUnit *SU = *(Pending.begin()+i);
2173 unsigned ReadyCycle = isTop() ? SU->TopReadyCycle : SU->BotReadyCycle;
2175 if (ReadyCycle < MinReadyCycle)
2176 MinReadyCycle = ReadyCycle;
2178 if (!IsBuffered && ReadyCycle > CurrCycle)
2181 if (checkHazard(SU))
2185 Pending.remove(Pending.begin()+i);
2188 DEBUG(if (!Pending.empty()) Pending.dump());
2189 CheckPending = false;
2192 /// Remove SU from the ready set for this boundary.
2193 void ConvergingScheduler::SchedBoundary::removeReady(SUnit *SU) {
2194 if (Available.isInQueue(SU))
2195 Available.remove(Available.find(SU));
2197 assert(Pending.isInQueue(SU) && "bad ready count");
2198 Pending.remove(Pending.find(SU));
2202 /// If this queue only has one ready candidate, return it. As a side effect,
2203 /// defer any nodes that now hit a hazard, and advance the cycle until at least
2204 /// one node is ready. If multiple instructions are ready, return NULL.
2205 SUnit *ConvergingScheduler::SchedBoundary::pickOnlyChoice() {
2210 // Defer any ready instrs that now have a hazard.
2211 for (ReadyQueue::iterator I = Available.begin(); I != Available.end();) {
2212 if (checkHazard(*I)) {
2214 I = Available.remove(I);
2220 for (unsigned i = 0; Available.empty(); ++i) {
2221 assert(i <= (HazardRec->getMaxLookAhead() + MaxObservedLatency) &&
2222 "permanent hazard"); (void)i;
2223 bumpCycle(CurrCycle + 1);
2226 if (Available.size() == 1)
2227 return *Available.begin();
2232 // This is useful information to dump after bumpNode.
2233 // Note that the Queue contents are more useful before pickNodeFromQueue.
2234 void ConvergingScheduler::SchedBoundary::dumpScheduledState() {
2237 if (ZoneCritResIdx) {
2238 ResFactor = SchedModel->getResourceFactor(ZoneCritResIdx);
2239 ResCount = getResourceCount(ZoneCritResIdx);
2242 ResFactor = SchedModel->getMicroOpFactor();
2243 ResCount = RetiredMOps * SchedModel->getMicroOpFactor();
2245 unsigned LFactor = SchedModel->getLatencyFactor();
2246 dbgs() << Available.getName() << " @" << CurrCycle << "c\n"
2247 << " Retired: " << RetiredMOps;
2248 dbgs() << "\n Executed: " << getExecutedCount() / LFactor << "c";
2249 dbgs() << "\n Critical: " << ResCount / LFactor << "c, "
2250 << ResCount / ResFactor << " " << getResourceName(ZoneCritResIdx)
2251 << "\n ExpectedLatency: " << ExpectedLatency << "c\n"
2252 << (IsResourceLimited ? " - Resource" : " - Latency")
2257 void ConvergingScheduler::SchedCandidate::
2258 initResourceDelta(const ScheduleDAGMI *DAG,
2259 const TargetSchedModel *SchedModel) {
2260 if (!Policy.ReduceResIdx && !Policy.DemandResIdx)
2263 const MCSchedClassDesc *SC = DAG->getSchedClass(SU);
2264 for (TargetSchedModel::ProcResIter
2265 PI = SchedModel->getWriteProcResBegin(SC),
2266 PE = SchedModel->getWriteProcResEnd(SC); PI != PE; ++PI) {
2267 if (PI->ProcResourceIdx == Policy.ReduceResIdx)
2268 ResDelta.CritResources += PI->Cycles;
2269 if (PI->ProcResourceIdx == Policy.DemandResIdx)
2270 ResDelta.DemandedResources += PI->Cycles;
2275 /// Return true if this heuristic determines order.
2276 static bool tryLess(int TryVal, int CandVal,
2277 ConvergingScheduler::SchedCandidate &TryCand,
2278 ConvergingScheduler::SchedCandidate &Cand,
2279 ConvergingScheduler::CandReason Reason) {
2280 if (TryVal < CandVal) {
2281 TryCand.Reason = Reason;
2284 if (TryVal > CandVal) {
2285 if (Cand.Reason > Reason)
2286 Cand.Reason = Reason;
2289 Cand.setRepeat(Reason);
2293 static bool tryGreater(int TryVal, int CandVal,
2294 ConvergingScheduler::SchedCandidate &TryCand,
2295 ConvergingScheduler::SchedCandidate &Cand,
2296 ConvergingScheduler::CandReason Reason) {
2297 if (TryVal > CandVal) {
2298 TryCand.Reason = Reason;
2301 if (TryVal < CandVal) {
2302 if (Cand.Reason > Reason)
2303 Cand.Reason = Reason;
2306 Cand.setRepeat(Reason);
2310 static bool tryPressure(const PressureChange &TryP,
2311 const PressureChange &CandP,
2312 ConvergingScheduler::SchedCandidate &TryCand,
2313 ConvergingScheduler::SchedCandidate &Cand,
2314 ConvergingScheduler::CandReason Reason) {
2315 int TryRank = TryP.getPSetOrMax();
2316 int CandRank = CandP.getPSetOrMax();
2317 // If both candidates affect the same set, go with the smallest increase.
2318 if (TryRank == CandRank) {
2319 return tryLess(TryP.getUnitInc(), CandP.getUnitInc(), TryCand, Cand,
2322 // If one candidate decreases and the other increases, go with it.
2323 // Invalid candidates have UnitInc==0.
2324 if (tryLess(TryP.getUnitInc() < 0, CandP.getUnitInc() < 0, TryCand, Cand,
2328 // If the candidates are decreasing pressure, reverse priority.
2329 if (TryP.getUnitInc() < 0)
2330 std::swap(TryRank, CandRank);
2331 return tryGreater(TryRank, CandRank, TryCand, Cand, Reason);
2334 static unsigned getWeakLeft(const SUnit *SU, bool isTop) {
2335 return (isTop) ? SU->WeakPredsLeft : SU->WeakSuccsLeft;
2338 /// Minimize physical register live ranges. Regalloc wants them adjacent to
2339 /// their physreg def/use.
2341 /// FIXME: This is an unnecessary check on the critical path. Most are root/leaf
2342 /// copies which can be prescheduled. The rest (e.g. x86 MUL) could be bundled
2343 /// with the operation that produces or consumes the physreg. We'll do this when
2344 /// regalloc has support for parallel copies.
2345 static int biasPhysRegCopy(const SUnit *SU, bool isTop) {
2346 const MachineInstr *MI = SU->getInstr();
2350 unsigned ScheduledOper = isTop ? 1 : 0;
2351 unsigned UnscheduledOper = isTop ? 0 : 1;
2352 // If we have already scheduled the physreg produce/consumer, immediately
2353 // schedule the copy.
2354 if (TargetRegisterInfo::isPhysicalRegister(
2355 MI->getOperand(ScheduledOper).getReg()))
2357 // If the physreg is at the boundary, defer it. Otherwise schedule it
2358 // immediately to free the dependent. We can hoist the copy later.
2359 bool AtBoundary = isTop ? !SU->NumSuccsLeft : !SU->NumPredsLeft;
2360 if (TargetRegisterInfo::isPhysicalRegister(
2361 MI->getOperand(UnscheduledOper).getReg()))
2362 return AtBoundary ? -1 : 1;
2366 static bool tryLatency(ConvergingScheduler::SchedCandidate &TryCand,
2367 ConvergingScheduler::SchedCandidate &Cand,
2368 ConvergingScheduler::SchedBoundary &Zone) {
2370 if (Cand.SU->getDepth() > Zone.getScheduledLatency()) {
2371 if (tryLess(TryCand.SU->getDepth(), Cand.SU->getDepth(),
2372 TryCand, Cand, ConvergingScheduler::TopDepthReduce))
2375 if (tryGreater(TryCand.SU->getHeight(), Cand.SU->getHeight(),
2376 TryCand, Cand, ConvergingScheduler::TopPathReduce))
2380 if (Cand.SU->getHeight() > Zone.getScheduledLatency()) {
2381 if (tryLess(TryCand.SU->getHeight(), Cand.SU->getHeight(),
2382 TryCand, Cand, ConvergingScheduler::BotHeightReduce))
2385 if (tryGreater(TryCand.SU->getDepth(), Cand.SU->getDepth(),
2386 TryCand, Cand, ConvergingScheduler::BotPathReduce))
2392 /// Apply a set of heursitics to a new candidate. Heuristics are currently
2393 /// hierarchical. This may be more efficient than a graduated cost model because
2394 /// we don't need to evaluate all aspects of the model for each node in the
2395 /// queue. But it's really done to make the heuristics easier to debug and
2396 /// statistically analyze.
2398 /// \param Cand provides the policy and current best candidate.
2399 /// \param TryCand refers to the next SUnit candidate, otherwise uninitialized.
2400 /// \param Zone describes the scheduled zone that we are extending.
2401 /// \param RPTracker describes reg pressure within the scheduled zone.
2402 /// \param TempTracker is a scratch pressure tracker to reuse in queries.
2403 void ConvergingScheduler::tryCandidate(SchedCandidate &Cand,
2404 SchedCandidate &TryCand,
2405 SchedBoundary &Zone,
2406 const RegPressureTracker &RPTracker,
2407 RegPressureTracker &TempTracker) {
2409 if (DAG->isTrackingPressure()) {
2410 // Always initialize TryCand's RPDelta.
2412 TempTracker.getMaxDownwardPressureDelta(
2413 TryCand.SU->getInstr(),
2415 DAG->getRegionCriticalPSets(),
2416 DAG->getRegPressure().MaxSetPressure);
2419 if (VerifyScheduling) {
2420 TempTracker.getMaxUpwardPressureDelta(
2421 TryCand.SU->getInstr(),
2422 &DAG->getPressureDiff(TryCand.SU),
2424 DAG->getRegionCriticalPSets(),
2425 DAG->getRegPressure().MaxSetPressure);
2428 RPTracker.getUpwardPressureDelta(
2429 TryCand.SU->getInstr(),
2430 DAG->getPressureDiff(TryCand.SU),
2432 DAG->getRegionCriticalPSets(),
2433 DAG->getRegPressure().MaxSetPressure);
2438 // Initialize the candidate if needed.
2439 if (!Cand.isValid()) {
2440 TryCand.Reason = NodeOrder;
2444 if (tryGreater(biasPhysRegCopy(TryCand.SU, Zone.isTop()),
2445 biasPhysRegCopy(Cand.SU, Zone.isTop()),
2446 TryCand, Cand, PhysRegCopy))
2449 // Avoid exceeding the target's limit. If signed PSetID is negative, it is
2450 // invalid; convert it to INT_MAX to give it lowest priority.
2451 if (DAG->isTrackingPressure() && tryPressure(TryCand.RPDelta.Excess,
2452 Cand.RPDelta.Excess,
2453 TryCand, Cand, RegExcess))
2456 // For loops that are acyclic path limited, aggressively schedule for latency.
2457 if (Rem.IsAcyclicLatencyLimited && tryLatency(TryCand, Cand, Zone))
2460 // Avoid increasing the max critical pressure in the scheduled region.
2461 if (DAG->isTrackingPressure() && tryPressure(TryCand.RPDelta.CriticalMax,
2462 Cand.RPDelta.CriticalMax,
2463 TryCand, Cand, RegCritical))
2466 // Keep clustered nodes together to encourage downstream peephole
2467 // optimizations which may reduce resource requirements.
2469 // This is a best effort to set things up for a post-RA pass. Optimizations
2470 // like generating loads of multiple registers should ideally be done within
2471 // the scheduler pass by combining the loads during DAG postprocessing.
2472 const SUnit *NextClusterSU =
2473 Zone.isTop() ? DAG->getNextClusterSucc() : DAG->getNextClusterPred();
2474 if (tryGreater(TryCand.SU == NextClusterSU, Cand.SU == NextClusterSU,
2475 TryCand, Cand, Cluster))
2478 // Weak edges are for clustering and other constraints.
2479 if (tryLess(getWeakLeft(TryCand.SU, Zone.isTop()),
2480 getWeakLeft(Cand.SU, Zone.isTop()),
2481 TryCand, Cand, Weak)) {
2484 // Avoid increasing the max pressure of the entire region.
2485 if (DAG->isTrackingPressure() && tryPressure(TryCand.RPDelta.CurrentMax,
2486 Cand.RPDelta.CurrentMax,
2487 TryCand, Cand, RegMax))
2490 // Avoid critical resource consumption and balance the schedule.
2491 TryCand.initResourceDelta(DAG, SchedModel);
2492 if (tryLess(TryCand.ResDelta.CritResources, Cand.ResDelta.CritResources,
2493 TryCand, Cand, ResourceReduce))
2495 if (tryGreater(TryCand.ResDelta.DemandedResources,
2496 Cand.ResDelta.DemandedResources,
2497 TryCand, Cand, ResourceDemand))
2500 // Avoid serializing long latency dependence chains.
2501 // For acyclic path limited loops, latency was already checked above.
2502 if (Cand.Policy.ReduceLatency && !Rem.IsAcyclicLatencyLimited
2503 && tryLatency(TryCand, Cand, Zone)) {
2507 // Prefer immediate defs/users of the last scheduled instruction. This is a
2508 // local pressure avoidance strategy that also makes the machine code
2510 if (tryGreater(Zone.NextSUs.count(TryCand.SU), Zone.NextSUs.count(Cand.SU),
2511 TryCand, Cand, NextDefUse))
2514 // Fall through to original instruction order.
2515 if ((Zone.isTop() && TryCand.SU->NodeNum < Cand.SU->NodeNum)
2516 || (!Zone.isTop() && TryCand.SU->NodeNum > Cand.SU->NodeNum)) {
2517 TryCand.Reason = NodeOrder;
2522 const char *ConvergingScheduler::getReasonStr(
2523 ConvergingScheduler::CandReason Reason) {
2525 case NoCand: return "NOCAND ";
2526 case PhysRegCopy: return "PREG-COPY";
2527 case RegExcess: return "REG-EXCESS";
2528 case RegCritical: return "REG-CRIT ";
2529 case Cluster: return "CLUSTER ";
2530 case Weak: return "WEAK ";
2531 case RegMax: return "REG-MAX ";
2532 case ResourceReduce: return "RES-REDUCE";
2533 case ResourceDemand: return "RES-DEMAND";
2534 case TopDepthReduce: return "TOP-DEPTH ";
2535 case TopPathReduce: return "TOP-PATH ";
2536 case BotHeightReduce:return "BOT-HEIGHT";
2537 case BotPathReduce: return "BOT-PATH ";
2538 case NextDefUse: return "DEF-USE ";
2539 case NodeOrder: return "ORDER ";
2541 llvm_unreachable("Unknown reason!");
2544 void ConvergingScheduler::traceCandidate(const SchedCandidate &Cand) {
2546 unsigned ResIdx = 0;
2547 unsigned Latency = 0;
2548 switch (Cand.Reason) {
2552 P = Cand.RPDelta.Excess;
2555 P = Cand.RPDelta.CriticalMax;
2558 P = Cand.RPDelta.CurrentMax;
2560 case ResourceReduce:
2561 ResIdx = Cand.Policy.ReduceResIdx;
2563 case ResourceDemand:
2564 ResIdx = Cand.Policy.DemandResIdx;
2566 case TopDepthReduce:
2567 Latency = Cand.SU->getDepth();
2570 Latency = Cand.SU->getHeight();
2572 case BotHeightReduce:
2573 Latency = Cand.SU->getHeight();
2576 Latency = Cand.SU->getDepth();
2579 dbgs() << " SU(" << Cand.SU->NodeNum << ") " << getReasonStr(Cand.Reason);
2581 dbgs() << " " << TRI->getRegPressureSetName(P.getPSet())
2582 << ":" << P.getUnitInc() << " ";
2586 dbgs() << " " << SchedModel->getProcResource(ResIdx)->Name << " ";
2590 dbgs() << " " << Latency << " cycles ";
2597 /// Pick the best candidate from the top queue.
2599 /// TODO: getMaxPressureDelta results can be mostly cached for each SUnit during
2600 /// DAG building. To adjust for the current scheduling location we need to
2601 /// maintain the number of vreg uses remaining to be top-scheduled.
2602 void ConvergingScheduler::pickNodeFromQueue(SchedBoundary &Zone,
2603 const RegPressureTracker &RPTracker,
2604 SchedCandidate &Cand) {
2605 ReadyQueue &Q = Zone.Available;
2609 // getMaxPressureDelta temporarily modifies the tracker.
2610 RegPressureTracker &TempTracker = const_cast<RegPressureTracker&>(RPTracker);
2612 for (ReadyQueue::iterator I = Q.begin(), E = Q.end(); I != E; ++I) {
2614 SchedCandidate TryCand(Cand.Policy);
2616 tryCandidate(Cand, TryCand, Zone, RPTracker, TempTracker);
2617 if (TryCand.Reason != NoCand) {
2618 // Initialize resource delta if needed in case future heuristics query it.
2619 if (TryCand.ResDelta == SchedResourceDelta())
2620 TryCand.initResourceDelta(DAG, SchedModel);
2621 Cand.setBest(TryCand);
2622 DEBUG(traceCandidate(Cand));
2627 static void tracePick(const ConvergingScheduler::SchedCandidate &Cand,
2629 DEBUG(dbgs() << "Pick " << (IsTop ? "Top " : "Bot ")
2630 << ConvergingScheduler::getReasonStr(Cand.Reason) << '\n');
2633 /// Pick the best candidate node from either the top or bottom queue.
2634 SUnit *ConvergingScheduler::pickNodeBidirectional(bool &IsTopNode) {
2635 // Schedule as far as possible in the direction of no choice. This is most
2636 // efficient, but also provides the best heuristics for CriticalPSets.
2637 if (SUnit *SU = Bot.pickOnlyChoice()) {
2639 DEBUG(dbgs() << "Pick Bot NOCAND\n");
2642 if (SUnit *SU = Top.pickOnlyChoice()) {
2644 DEBUG(dbgs() << "Pick Top NOCAND\n");
2647 CandPolicy NoPolicy;
2648 SchedCandidate BotCand(NoPolicy);
2649 SchedCandidate TopCand(NoPolicy);
2650 Bot.setPolicy(BotCand.Policy, Top);
2651 Top.setPolicy(TopCand.Policy, Bot);
2653 // Prefer bottom scheduling when heuristics are silent.
2654 pickNodeFromQueue(Bot, DAG->getBotRPTracker(), BotCand);
2655 assert(BotCand.Reason != NoCand && "failed to find the first candidate");
2657 // If either Q has a single candidate that provides the least increase in
2658 // Excess pressure, we can immediately schedule from that Q.
2660 // RegionCriticalPSets summarizes the pressure within the scheduled region and
2661 // affects picking from either Q. If scheduling in one direction must
2662 // increase pressure for one of the excess PSets, then schedule in that
2663 // direction first to provide more freedom in the other direction.
2664 if ((BotCand.Reason == RegExcess && !BotCand.isRepeat(RegExcess))
2665 || (BotCand.Reason == RegCritical
2666 && !BotCand.isRepeat(RegCritical)))
2669 tracePick(BotCand, IsTopNode);
2672 // Check if the top Q has a better candidate.
2673 pickNodeFromQueue(Top, DAG->getTopRPTracker(), TopCand);
2674 assert(TopCand.Reason != NoCand && "failed to find the first candidate");
2676 // Choose the queue with the most important (lowest enum) reason.
2677 if (TopCand.Reason < BotCand.Reason) {
2679 tracePick(TopCand, IsTopNode);
2682 // Otherwise prefer the bottom candidate, in node order if all else failed.
2684 tracePick(BotCand, IsTopNode);
2688 /// Pick the best node to balance the schedule. Implements MachineSchedStrategy.
2689 SUnit *ConvergingScheduler::pickNode(bool &IsTopNode) {
2690 if (DAG->top() == DAG->bottom()) {
2691 assert(Top.Available.empty() && Top.Pending.empty() &&
2692 Bot.Available.empty() && Bot.Pending.empty() && "ReadyQ garbage");
2698 SU = Top.pickOnlyChoice();
2700 CandPolicy NoPolicy;
2701 SchedCandidate TopCand(NoPolicy);
2702 pickNodeFromQueue(Top, DAG->getTopRPTracker(), TopCand);
2703 assert(TopCand.Reason != NoCand && "failed to find a candidate");
2704 tracePick(TopCand, true);
2709 else if (OnlyBottomUp) {
2710 SU = Bot.pickOnlyChoice();
2712 CandPolicy NoPolicy;
2713 SchedCandidate BotCand(NoPolicy);
2714 pickNodeFromQueue(Bot, DAG->getBotRPTracker(), BotCand);
2715 assert(BotCand.Reason != NoCand && "failed to find a candidate");
2716 tracePick(BotCand, false);
2722 SU = pickNodeBidirectional(IsTopNode);
2724 } while (SU->isScheduled);
2726 if (SU->isTopReady())
2727 Top.removeReady(SU);
2728 if (SU->isBottomReady())
2729 Bot.removeReady(SU);
2731 DEBUG(dbgs() << "Scheduling SU(" << SU->NodeNum << ") " << *SU->getInstr());
2735 void ConvergingScheduler::reschedulePhysRegCopies(SUnit *SU, bool isTop) {
2737 MachineBasicBlock::iterator InsertPos = SU->getInstr();
2740 SmallVectorImpl<SDep> &Deps = isTop ? SU->Preds : SU->Succs;
2742 // Find already scheduled copies with a single physreg dependence and move
2743 // them just above the scheduled instruction.
2744 for (SmallVectorImpl<SDep>::iterator I = Deps.begin(), E = Deps.end();
2746 if (I->getKind() != SDep::Data || !TRI->isPhysicalRegister(I->getReg()))
2748 SUnit *DepSU = I->getSUnit();
2749 if (isTop ? DepSU->Succs.size() > 1 : DepSU->Preds.size() > 1)
2751 MachineInstr *Copy = DepSU->getInstr();
2752 if (!Copy->isCopy())
2754 DEBUG(dbgs() << " Rescheduling physreg copy ";
2755 I->getSUnit()->dump(DAG));
2756 DAG->moveInstruction(Copy, InsertPos);
2760 /// Update the scheduler's state after scheduling a node. This is the same node
2761 /// that was just returned by pickNode(). However, ScheduleDAGMI needs to update
2762 /// it's state based on the current cycle before MachineSchedStrategy does.
2764 /// FIXME: Eventually, we may bundle physreg copies rather than rescheduling
2765 /// them here. See comments in biasPhysRegCopy.
2766 void ConvergingScheduler::schedNode(SUnit *SU, bool IsTopNode) {
2768 SU->TopReadyCycle = std::max(SU->TopReadyCycle, Top.CurrCycle);
2770 if (SU->hasPhysRegUses)
2771 reschedulePhysRegCopies(SU, true);
2774 SU->BotReadyCycle = std::max(SU->BotReadyCycle, Bot.CurrCycle);
2776 if (SU->hasPhysRegDefs)
2777 reschedulePhysRegCopies(SU, false);
2781 /// Create the standard converging machine scheduler. This will be used as the
2782 /// default scheduler if the target does not set a default.
2783 static ScheduleDAGInstrs *createConvergingSched(MachineSchedContext *C) {
2784 ScheduleDAGMI *DAG = new ScheduleDAGMI(C, new ConvergingScheduler(C));
2785 // Register DAG post-processors.
2787 // FIXME: extend the mutation API to allow earlier mutations to instantiate
2788 // data and pass it to later mutations. Have a single mutation that gathers
2789 // the interesting nodes in one pass.
2790 DAG->addMutation(new CopyConstrain(DAG->TII, DAG->TRI));
2791 if (EnableLoadCluster && DAG->TII->enableClusterLoads())
2792 DAG->addMutation(new LoadClusterMutation(DAG->TII, DAG->TRI));
2793 if (EnableMacroFusion)
2794 DAG->addMutation(new MacroFusion(DAG->TII));
2797 static MachineSchedRegistry
2798 ConvergingSchedRegistry("converge", "Standard converging scheduler.",
2799 createConvergingSched);
2801 //===----------------------------------------------------------------------===//
2802 // ILP Scheduler. Currently for experimental analysis of heuristics.
2803 //===----------------------------------------------------------------------===//
2806 /// \brief Order nodes by the ILP metric.
2808 const SchedDFSResult *DFSResult;
2809 const BitVector *ScheduledTrees;
2812 ILPOrder(bool MaxILP): DFSResult(0), ScheduledTrees(0), MaximizeILP(MaxILP) {}
2814 /// \brief Apply a less-than relation on node priority.
2816 /// (Return true if A comes after B in the Q.)
2817 bool operator()(const SUnit *A, const SUnit *B) const {
2818 unsigned SchedTreeA = DFSResult->getSubtreeID(A);
2819 unsigned SchedTreeB = DFSResult->getSubtreeID(B);
2820 if (SchedTreeA != SchedTreeB) {
2821 // Unscheduled trees have lower priority.
2822 if (ScheduledTrees->test(SchedTreeA) != ScheduledTrees->test(SchedTreeB))
2823 return ScheduledTrees->test(SchedTreeB);
2825 // Trees with shallower connections have have lower priority.
2826 if (DFSResult->getSubtreeLevel(SchedTreeA)
2827 != DFSResult->getSubtreeLevel(SchedTreeB)) {
2828 return DFSResult->getSubtreeLevel(SchedTreeA)
2829 < DFSResult->getSubtreeLevel(SchedTreeB);
2833 return DFSResult->getILP(A) < DFSResult->getILP(B);
2835 return DFSResult->getILP(A) > DFSResult->getILP(B);
2839 /// \brief Schedule based on the ILP metric.
2840 class ILPScheduler : public MachineSchedStrategy {
2844 std::vector<SUnit*> ReadyQ;
2846 ILPScheduler(bool MaximizeILP): DAG(0), Cmp(MaximizeILP) {}
2848 virtual void initialize(ScheduleDAGMI *dag) {
2850 DAG->computeDFSResult();
2851 Cmp.DFSResult = DAG->getDFSResult();
2852 Cmp.ScheduledTrees = &DAG->getScheduledTrees();
2856 virtual void registerRoots() {
2857 // Restore the heap in ReadyQ with the updated DFS results.
2858 std::make_heap(ReadyQ.begin(), ReadyQ.end(), Cmp);
2861 /// Implement MachineSchedStrategy interface.
2862 /// -----------------------------------------
2864 /// Callback to select the highest priority node from the ready Q.
2865 virtual SUnit *pickNode(bool &IsTopNode) {
2866 if (ReadyQ.empty()) return NULL;
2867 std::pop_heap(ReadyQ.begin(), ReadyQ.end(), Cmp);
2868 SUnit *SU = ReadyQ.back();
2871 DEBUG(dbgs() << "Pick node " << "SU(" << SU->NodeNum << ") "
2872 << " ILP: " << DAG->getDFSResult()->getILP(SU)
2873 << " Tree: " << DAG->getDFSResult()->getSubtreeID(SU) << " @"
2874 << DAG->getDFSResult()->getSubtreeLevel(
2875 DAG->getDFSResult()->getSubtreeID(SU)) << '\n'
2876 << "Scheduling " << *SU->getInstr());
2880 /// \brief Scheduler callback to notify that a new subtree is scheduled.
2881 virtual void scheduleTree(unsigned SubtreeID) {
2882 std::make_heap(ReadyQ.begin(), ReadyQ.end(), Cmp);
2885 /// Callback after a node is scheduled. Mark a newly scheduled tree, notify
2886 /// DFSResults, and resort the priority Q.
2887 virtual void schedNode(SUnit *SU, bool IsTopNode) {
2888 assert(!IsTopNode && "SchedDFSResult needs bottom-up");
2891 virtual void releaseTopNode(SUnit *) { /*only called for top roots*/ }
2893 virtual void releaseBottomNode(SUnit *SU) {
2894 ReadyQ.push_back(SU);
2895 std::push_heap(ReadyQ.begin(), ReadyQ.end(), Cmp);
2900 static ScheduleDAGInstrs *createILPMaxScheduler(MachineSchedContext *C) {
2901 return new ScheduleDAGMI(C, new ILPScheduler(true));
2903 static ScheduleDAGInstrs *createILPMinScheduler(MachineSchedContext *C) {
2904 return new ScheduleDAGMI(C, new ILPScheduler(false));
2906 static MachineSchedRegistry ILPMaxRegistry(
2907 "ilpmax", "Schedule bottom-up for max ILP", createILPMaxScheduler);
2908 static MachineSchedRegistry ILPMinRegistry(
2909 "ilpmin", "Schedule bottom-up for min ILP", createILPMinScheduler);
2911 //===----------------------------------------------------------------------===//
2912 // Machine Instruction Shuffler for Correctness Testing
2913 //===----------------------------------------------------------------------===//
2917 /// Apply a less-than relation on the node order, which corresponds to the
2918 /// instruction order prior to scheduling. IsReverse implements greater-than.
2919 template<bool IsReverse>
2921 bool operator()(SUnit *A, SUnit *B) const {
2923 return A->NodeNum > B->NodeNum;
2925 return A->NodeNum < B->NodeNum;
2929 /// Reorder instructions as much as possible.
2930 class InstructionShuffler : public MachineSchedStrategy {
2934 // Using a less-than relation (SUnitOrder<false>) for the TopQ priority
2935 // gives nodes with a higher number higher priority causing the latest
2936 // instructions to be scheduled first.
2937 PriorityQueue<SUnit*, std::vector<SUnit*>, SUnitOrder<false> >
2939 // When scheduling bottom-up, use greater-than as the queue priority.
2940 PriorityQueue<SUnit*, std::vector<SUnit*>, SUnitOrder<true> >
2943 InstructionShuffler(bool alternate, bool topdown)
2944 : IsAlternating(alternate), IsTopDown(topdown) {}
2946 virtual void initialize(ScheduleDAGMI *) {
2951 /// Implement MachineSchedStrategy interface.
2952 /// -----------------------------------------
2954 virtual SUnit *pickNode(bool &IsTopNode) {
2958 if (TopQ.empty()) return NULL;
2961 } while (SU->isScheduled);
2966 if (BottomQ.empty()) return NULL;
2969 } while (SU->isScheduled);
2973 IsTopDown = !IsTopDown;
2977 virtual void schedNode(SUnit *SU, bool IsTopNode) {}
2979 virtual void releaseTopNode(SUnit *SU) {
2982 virtual void releaseBottomNode(SUnit *SU) {
2988 static ScheduleDAGInstrs *createInstructionShuffler(MachineSchedContext *C) {
2989 bool Alternate = !ForceTopDown && !ForceBottomUp;
2990 bool TopDown = !ForceBottomUp;
2991 assert((TopDown || !ForceTopDown) &&
2992 "-misched-topdown incompatible with -misched-bottomup");
2993 return new ScheduleDAGMI(C, new InstructionShuffler(Alternate, TopDown));
2995 static MachineSchedRegistry ShufflerRegistry(
2996 "shuffle", "Shuffle machine instructions alternating directions",
2997 createInstructionShuffler);
3000 //===----------------------------------------------------------------------===//
3001 // GraphWriter support for ScheduleDAGMI.
3002 //===----------------------------------------------------------------------===//
3007 template<> struct GraphTraits<
3008 ScheduleDAGMI*> : public GraphTraits<ScheduleDAG*> {};
3011 struct DOTGraphTraits<ScheduleDAGMI*> : public DefaultDOTGraphTraits {
3013 DOTGraphTraits (bool isSimple=false) : DefaultDOTGraphTraits(isSimple) {}
3015 static std::string getGraphName(const ScheduleDAG *G) {
3016 return G->MF.getName();
3019 static bool renderGraphFromBottomUp() {
3023 static bool isNodeHidden(const SUnit *Node) {
3024 return (Node->Preds.size() > 10 || Node->Succs.size() > 10);
3027 static bool hasNodeAddressLabel(const SUnit *Node,
3028 const ScheduleDAG *Graph) {
3032 /// If you want to override the dot attributes printed for a particular
3033 /// edge, override this method.
3034 static std::string getEdgeAttributes(const SUnit *Node,
3036 const ScheduleDAG *Graph) {
3037 if (EI.isArtificialDep())
3038 return "color=cyan,style=dashed";
3040 return "color=blue,style=dashed";
3044 static std::string getNodeLabel(const SUnit *SU, const ScheduleDAG *G) {
3046 raw_string_ostream SS(Str);
3047 SS << "SU(" << SU->NodeNum << ')';
3050 static std::string getNodeDescription(const SUnit *SU, const ScheduleDAG *G) {
3051 return G->getGraphNodeLabel(SU);
3054 static std::string getNodeAttributes(const SUnit *N,
3055 const ScheduleDAG *Graph) {
3056 std::string Str("shape=Mrecord");
3057 const SchedDFSResult *DFS =
3058 static_cast<const ScheduleDAGMI*>(Graph)->getDFSResult();
3060 Str += ",style=filled,fillcolor=\"#";
3061 Str += DOT::getColorString(DFS->getSubtreeID(N));
3070 /// viewGraph - Pop up a ghostview window with the reachable parts of the DAG
3071 /// rendered using 'dot'.
3073 void ScheduleDAGMI::viewGraph(const Twine &Name, const Twine &Title) {
3075 ViewGraph(this, Name, false, Title);
3077 errs() << "ScheduleDAGMI::viewGraph is only available in debug builds on "
3078 << "systems with Graphviz or gv!\n";
3082 /// Out-of-line implementation with no arguments is handy for gdb.
3083 void ScheduleDAGMI::viewGraph() {
3084 viewGraph(getDAGName(), "Scheduling-Units Graph for " + getDAGName());