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 cl::opt<std::string> SchedOnlyFunc("misched-only-func", cl::Hidden,
54 cl::desc("Only schedule this function"));
55 static cl::opt<unsigned> SchedOnlyBlock("misched-only-block", cl::Hidden,
56 cl::desc("Only schedule this MBB#"));
58 static bool ViewMISchedDAGs = false;
61 static cl::opt<bool> EnableRegPressure("misched-regpressure", cl::Hidden,
62 cl::desc("Enable register pressure scheduling."), cl::init(true));
64 static cl::opt<bool> EnableCyclicPath("misched-cyclicpath", cl::Hidden,
65 cl::desc("Enable cyclic critical path analysis."), cl::init(true));
67 static cl::opt<bool> EnableLoadCluster("misched-cluster", cl::Hidden,
68 cl::desc("Enable load clustering."), cl::init(true));
70 // Experimental heuristics
71 static cl::opt<bool> EnableMacroFusion("misched-fusion", cl::Hidden,
72 cl::desc("Enable scheduling for macro fusion."), cl::init(true));
74 static cl::opt<bool> VerifyScheduling("verify-misched", cl::Hidden,
75 cl::desc("Verify machine instrs before and after machine scheduling"));
77 // DAG subtrees must have at least this many nodes.
78 static const unsigned MinSubtreeSize = 8;
80 // Pin the vtables to this file.
81 void MachineSchedStrategy::anchor() {}
82 void ScheduleDAGMutation::anchor() {}
84 //===----------------------------------------------------------------------===//
85 // Machine Instruction Scheduling Pass and Registry
86 //===----------------------------------------------------------------------===//
88 MachineSchedContext::MachineSchedContext():
89 MF(0), MLI(0), MDT(0), PassConfig(0), AA(0), LIS(0) {
90 RegClassInfo = new RegisterClassInfo();
93 MachineSchedContext::~MachineSchedContext() {
98 /// Base class for a machine scheduler class that can run at any point.
99 class MachineSchedulerBase : public MachineSchedContext,
100 public MachineFunctionPass {
102 MachineSchedulerBase(char &ID): MachineFunctionPass(ID) {}
104 virtual void print(raw_ostream &O, const Module* = 0) const;
107 void scheduleRegions(ScheduleDAGInstrs &Scheduler);
110 /// MachineScheduler runs after coalescing and before register allocation.
111 class MachineScheduler : public MachineSchedulerBase {
115 virtual void getAnalysisUsage(AnalysisUsage &AU) const;
117 virtual bool runOnMachineFunction(MachineFunction&);
119 static char ID; // Class identification, replacement for typeinfo
122 ScheduleDAGInstrs *createMachineScheduler();
125 /// PostMachineScheduler runs after shortly before code emission.
126 class PostMachineScheduler : public MachineSchedulerBase {
128 PostMachineScheduler();
130 virtual void getAnalysisUsage(AnalysisUsage &AU) const;
132 virtual bool runOnMachineFunction(MachineFunction&);
134 static char ID; // Class identification, replacement for typeinfo
137 ScheduleDAGInstrs *createPostMachineScheduler();
141 char MachineScheduler::ID = 0;
143 char &llvm::MachineSchedulerID = MachineScheduler::ID;
145 INITIALIZE_PASS_BEGIN(MachineScheduler, "misched",
146 "Machine Instruction Scheduler", false, false)
147 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
148 INITIALIZE_PASS_DEPENDENCY(SlotIndexes)
149 INITIALIZE_PASS_DEPENDENCY(LiveIntervals)
150 INITIALIZE_PASS_END(MachineScheduler, "misched",
151 "Machine Instruction Scheduler", false, false)
153 MachineScheduler::MachineScheduler()
154 : MachineSchedulerBase(ID) {
155 initializeMachineSchedulerPass(*PassRegistry::getPassRegistry());
158 void MachineScheduler::getAnalysisUsage(AnalysisUsage &AU) const {
159 AU.setPreservesCFG();
160 AU.addRequiredID(MachineDominatorsID);
161 AU.addRequired<MachineLoopInfo>();
162 AU.addRequired<AliasAnalysis>();
163 AU.addRequired<TargetPassConfig>();
164 AU.addRequired<SlotIndexes>();
165 AU.addPreserved<SlotIndexes>();
166 AU.addRequired<LiveIntervals>();
167 AU.addPreserved<LiveIntervals>();
168 MachineFunctionPass::getAnalysisUsage(AU);
171 char PostMachineScheduler::ID = 0;
173 char &llvm::PostMachineSchedulerID = PostMachineScheduler::ID;
175 INITIALIZE_PASS(PostMachineScheduler, "postmisched",
176 "PostRA Machine Instruction Scheduler", false, false)
178 PostMachineScheduler::PostMachineScheduler()
179 : MachineSchedulerBase(ID) {
180 initializePostMachineSchedulerPass(*PassRegistry::getPassRegistry());
183 void PostMachineScheduler::getAnalysisUsage(AnalysisUsage &AU) const {
184 AU.setPreservesCFG();
185 AU.addRequiredID(MachineDominatorsID);
186 AU.addRequired<MachineLoopInfo>();
187 AU.addRequired<TargetPassConfig>();
188 MachineFunctionPass::getAnalysisUsage(AU);
191 MachinePassRegistry MachineSchedRegistry::Registry;
193 /// A dummy default scheduler factory indicates whether the scheduler
194 /// is overridden on the command line.
195 static ScheduleDAGInstrs *useDefaultMachineSched(MachineSchedContext *C) {
199 /// MachineSchedOpt allows command line selection of the scheduler.
200 static cl::opt<MachineSchedRegistry::ScheduleDAGCtor, false,
201 RegisterPassParser<MachineSchedRegistry> >
202 MachineSchedOpt("misched",
203 cl::init(&useDefaultMachineSched), cl::Hidden,
204 cl::desc("Machine instruction scheduler to use"));
206 static MachineSchedRegistry
207 DefaultSchedRegistry("default", "Use the target's default scheduler choice.",
208 useDefaultMachineSched);
210 /// Forward declare the standard machine scheduler. This will be used as the
211 /// default scheduler if the target does not set a default.
212 static ScheduleDAGInstrs *createGenericSchedLive(MachineSchedContext *C);
213 static ScheduleDAGInstrs *createGenericSchedPostRA(MachineSchedContext *C);
215 /// Decrement this iterator until reaching the top or a non-debug instr.
216 static MachineBasicBlock::const_iterator
217 priorNonDebug(MachineBasicBlock::const_iterator I,
218 MachineBasicBlock::const_iterator Beg) {
219 assert(I != Beg && "reached the top of the region, cannot decrement");
221 if (!I->isDebugValue())
227 /// Non-const version.
228 static MachineBasicBlock::iterator
229 priorNonDebug(MachineBasicBlock::iterator I,
230 MachineBasicBlock::const_iterator Beg) {
231 return const_cast<MachineInstr*>(
232 &*priorNonDebug(MachineBasicBlock::const_iterator(I), Beg));
235 /// If this iterator is a debug value, increment until reaching the End or a
236 /// non-debug instruction.
237 static MachineBasicBlock::const_iterator
238 nextIfDebug(MachineBasicBlock::const_iterator I,
239 MachineBasicBlock::const_iterator End) {
240 for(; I != End; ++I) {
241 if (!I->isDebugValue())
247 /// Non-const version.
248 static MachineBasicBlock::iterator
249 nextIfDebug(MachineBasicBlock::iterator I,
250 MachineBasicBlock::const_iterator End) {
251 // Cast the return value to nonconst MachineInstr, then cast to an
252 // instr_iterator, which does not check for null, finally return a
254 return MachineBasicBlock::instr_iterator(
255 const_cast<MachineInstr*>(
256 &*nextIfDebug(MachineBasicBlock::const_iterator(I), End)));
259 /// Instantiate a ScheduleDAGInstrs that will be owned by the caller.
260 ScheduleDAGInstrs *MachineScheduler::createMachineScheduler() {
261 // Select the scheduler, or set the default.
262 MachineSchedRegistry::ScheduleDAGCtor Ctor = MachineSchedOpt;
263 if (Ctor != useDefaultMachineSched)
266 // Get the default scheduler set by the target for this function.
267 ScheduleDAGInstrs *Scheduler = PassConfig->createMachineScheduler(this);
271 // Default to GenericScheduler.
272 return createGenericSchedLive(this);
275 /// Instantiate a ScheduleDAGInstrs for PostRA scheduling that will be owned by
276 /// the caller. We don't have a command line option to override the postRA
277 /// scheduler. The Target must configure it.
278 ScheduleDAGInstrs *PostMachineScheduler::createPostMachineScheduler() {
279 // Get the postRA scheduler set by the target for this function.
280 ScheduleDAGInstrs *Scheduler = PassConfig->createPostMachineScheduler(this);
284 // Default to GenericScheduler.
285 return createGenericSchedPostRA(this);
288 /// Top-level MachineScheduler pass driver.
290 /// Visit blocks in function order. Divide each block into scheduling regions
291 /// and visit them bottom-up. Visiting regions bottom-up is not required, but is
292 /// consistent with the DAG builder, which traverses the interior of the
293 /// scheduling regions bottom-up.
295 /// This design avoids exposing scheduling boundaries to the DAG builder,
296 /// simplifying the DAG builder's support for "special" target instructions.
297 /// At the same time the design allows target schedulers to operate across
298 /// scheduling boundaries, for example to bundle the boudary instructions
299 /// without reordering them. This creates complexity, because the target
300 /// scheduler must update the RegionBegin and RegionEnd positions cached by
301 /// ScheduleDAGInstrs whenever adding or removing instructions. A much simpler
302 /// design would be to split blocks at scheduling boundaries, but LLVM has a
303 /// general bias against block splitting purely for implementation simplicity.
304 bool MachineScheduler::runOnMachineFunction(MachineFunction &mf) {
305 DEBUG(dbgs() << "Before MISsched:\n"; mf.print(dbgs()));
307 // Initialize the context of the pass.
309 MLI = &getAnalysis<MachineLoopInfo>();
310 MDT = &getAnalysis<MachineDominatorTree>();
311 PassConfig = &getAnalysis<TargetPassConfig>();
312 AA = &getAnalysis<AliasAnalysis>();
314 LIS = &getAnalysis<LiveIntervals>();
316 if (VerifyScheduling) {
318 MF->verify(this, "Before machine scheduling.");
320 RegClassInfo->runOnMachineFunction(*MF);
322 // Instantiate the selected scheduler for this target, function, and
323 // optimization level.
324 OwningPtr<ScheduleDAGInstrs> Scheduler(createMachineScheduler());
325 scheduleRegions(*Scheduler);
328 if (VerifyScheduling)
329 MF->verify(this, "After machine scheduling.");
333 bool PostMachineScheduler::runOnMachineFunction(MachineFunction &mf) {
334 DEBUG(dbgs() << "Before post-MI-sched:\n"; mf.print(dbgs()));
336 // Initialize the context of the pass.
338 PassConfig = &getAnalysis<TargetPassConfig>();
340 if (VerifyScheduling)
341 MF->verify(this, "Before post machine scheduling.");
343 // Instantiate the selected scheduler for this target, function, and
344 // optimization level.
345 OwningPtr<ScheduleDAGInstrs> Scheduler(createPostMachineScheduler());
346 scheduleRegions(*Scheduler);
348 if (VerifyScheduling)
349 MF->verify(this, "After post machine scheduling.");
353 /// Return true of the given instruction should not be included in a scheduling
356 /// MachineScheduler does not currently support scheduling across calls. To
357 /// handle calls, the DAG builder needs to be modified to create register
358 /// anti/output dependencies on the registers clobbered by the call's regmask
359 /// operand. In PreRA scheduling, the stack pointer adjustment already prevents
360 /// scheduling across calls. In PostRA scheduling, we need the isCall to enforce
361 /// the boundary, but there would be no benefit to postRA scheduling across
362 /// calls this late anyway.
363 static bool isSchedBoundary(MachineBasicBlock::iterator MI,
364 MachineBasicBlock *MBB,
366 const TargetInstrInfo *TII,
368 return MI->isCall() || TII->isSchedulingBoundary(MI, MBB, *MF);
371 /// Main driver for both MachineScheduler and PostMachineScheduler.
372 void MachineSchedulerBase::scheduleRegions(ScheduleDAGInstrs &Scheduler) {
373 const TargetInstrInfo *TII = MF->getTarget().getInstrInfo();
374 bool IsPostRA = Scheduler.isPostRA();
376 // Visit all machine basic blocks.
378 // TODO: Visit blocks in global postorder or postorder within the bottom-up
379 // loop tree. Then we can optionally compute global RegPressure.
380 for (MachineFunction::iterator MBB = MF->begin(), MBBEnd = MF->end();
381 MBB != MBBEnd; ++MBB) {
383 Scheduler.startBlock(MBB);
386 if (SchedOnlyFunc.getNumOccurrences() && SchedOnlyFunc != MF->getName())
388 if (SchedOnlyBlock.getNumOccurrences()
389 && (int)SchedOnlyBlock != MBB->getNumber())
393 // Break the block into scheduling regions [I, RegionEnd), and schedule each
394 // region as soon as it is discovered. RegionEnd points the scheduling
395 // boundary at the bottom of the region. The DAG does not include RegionEnd,
396 // but the region does (i.e. the next RegionEnd is above the previous
397 // RegionBegin). If the current block has no terminator then RegionEnd ==
398 // MBB->end() for the bottom region.
400 // The Scheduler may insert instructions during either schedule() or
401 // exitRegion(), even for empty regions. So the local iterators 'I' and
402 // 'RegionEnd' are invalid across these calls.
404 // MBB::size() uses instr_iterator to count. Here we need a bundle to count
405 // as a single instruction.
406 unsigned RemainingInstrs = std::distance(MBB->begin(), MBB->end());
407 for(MachineBasicBlock::iterator RegionEnd = MBB->end();
408 RegionEnd != MBB->begin(); RegionEnd = Scheduler.begin()) {
410 // Avoid decrementing RegionEnd for blocks with no terminator.
411 if (RegionEnd != MBB->end() ||
412 isSchedBoundary(std::prev(RegionEnd), MBB, MF, TII, IsPostRA)) {
414 // Count the boundary instruction.
418 // The next region starts above the previous region. Look backward in the
419 // instruction stream until we find the nearest boundary.
420 unsigned NumRegionInstrs = 0;
421 MachineBasicBlock::iterator I = RegionEnd;
422 for(;I != MBB->begin(); --I, --RemainingInstrs, ++NumRegionInstrs) {
423 if (isSchedBoundary(std::prev(I), MBB, MF, TII, IsPostRA))
426 // Notify the scheduler of the region, even if we may skip scheduling
427 // it. Perhaps it still needs to be bundled.
428 Scheduler.enterRegion(MBB, I, RegionEnd, NumRegionInstrs);
430 // Skip empty scheduling regions (0 or 1 schedulable instructions).
431 if (I == RegionEnd || I == std::prev(RegionEnd)) {
432 // Close the current region. Bundle the terminator if needed.
433 // This invalidates 'RegionEnd' and 'I'.
434 Scheduler.exitRegion();
437 DEBUG(dbgs() << "********** " << ((Scheduler.isPostRA()) ? "PostRA " : "")
438 << "MI Scheduling **********\n");
439 DEBUG(dbgs() << MF->getName()
440 << ":BB#" << MBB->getNumber() << " " << MBB->getName()
441 << "\n From: " << *I << " To: ";
442 if (RegionEnd != MBB->end()) dbgs() << *RegionEnd;
443 else dbgs() << "End";
444 dbgs() << " RegionInstrs: " << NumRegionInstrs
445 << " Remaining: " << RemainingInstrs << "\n");
447 // Schedule a region: possibly reorder instructions.
448 // This invalidates 'RegionEnd' and 'I'.
449 Scheduler.schedule();
451 // Close the current region.
452 Scheduler.exitRegion();
454 // Scheduling has invalidated the current iterator 'I'. Ask the
455 // scheduler for the top of it's scheduled region.
456 RegionEnd = Scheduler.begin();
458 assert(RemainingInstrs == 0 && "Instruction count mismatch!");
459 Scheduler.finishBlock();
460 if (Scheduler.isPostRA()) {
461 // FIXME: Ideally, no further passes should rely on kill flags. However,
462 // thumb2 size reduction is currently an exception.
463 Scheduler.fixupKills(MBB);
466 Scheduler.finalizeSchedule();
469 void MachineSchedulerBase::print(raw_ostream &O, const Module* m) const {
473 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
474 void ReadyQueue::dump() {
475 dbgs() << Name << ": ";
476 for (unsigned i = 0, e = Queue.size(); i < e; ++i)
477 dbgs() << Queue[i]->NodeNum << " ";
482 //===----------------------------------------------------------------------===//
483 // ScheduleDAGMI - Basic machine instruction scheduling. This is
484 // independent of PreRA/PostRA scheduling and involves no extra book-keeping for
485 // virtual registers.
486 // ===----------------------------------------------------------------------===/
488 ScheduleDAGMI::~ScheduleDAGMI() {
489 DeleteContainerPointers(Mutations);
493 bool ScheduleDAGMI::canAddEdge(SUnit *SuccSU, SUnit *PredSU) {
494 return SuccSU == &ExitSU || !Topo.IsReachable(PredSU, SuccSU);
497 bool ScheduleDAGMI::addEdge(SUnit *SuccSU, const SDep &PredDep) {
498 if (SuccSU != &ExitSU) {
499 // Do not use WillCreateCycle, it assumes SD scheduling.
500 // If Pred is reachable from Succ, then the edge creates a cycle.
501 if (Topo.IsReachable(PredDep.getSUnit(), SuccSU))
503 Topo.AddPred(SuccSU, PredDep.getSUnit());
505 SuccSU->addPred(PredDep, /*Required=*/!PredDep.isArtificial());
506 // Return true regardless of whether a new edge needed to be inserted.
510 /// ReleaseSucc - Decrement the NumPredsLeft count of a successor. When
511 /// NumPredsLeft reaches zero, release the successor node.
513 /// FIXME: Adjust SuccSU height based on MinLatency.
514 void ScheduleDAGMI::releaseSucc(SUnit *SU, SDep *SuccEdge) {
515 SUnit *SuccSU = SuccEdge->getSUnit();
517 if (SuccEdge->isWeak()) {
518 --SuccSU->WeakPredsLeft;
519 if (SuccEdge->isCluster())
520 NextClusterSucc = SuccSU;
524 if (SuccSU->NumPredsLeft == 0) {
525 dbgs() << "*** Scheduling failed! ***\n";
527 dbgs() << " has been released too many times!\n";
531 --SuccSU->NumPredsLeft;
532 if (SuccSU->NumPredsLeft == 0 && SuccSU != &ExitSU)
533 SchedImpl->releaseTopNode(SuccSU);
536 /// releaseSuccessors - Call releaseSucc on each of SU's successors.
537 void ScheduleDAGMI::releaseSuccessors(SUnit *SU) {
538 for (SUnit::succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
540 releaseSucc(SU, &*I);
544 /// ReleasePred - Decrement the NumSuccsLeft count of a predecessor. When
545 /// NumSuccsLeft reaches zero, release the predecessor node.
547 /// FIXME: Adjust PredSU height based on MinLatency.
548 void ScheduleDAGMI::releasePred(SUnit *SU, SDep *PredEdge) {
549 SUnit *PredSU = PredEdge->getSUnit();
551 if (PredEdge->isWeak()) {
552 --PredSU->WeakSuccsLeft;
553 if (PredEdge->isCluster())
554 NextClusterPred = PredSU;
558 if (PredSU->NumSuccsLeft == 0) {
559 dbgs() << "*** Scheduling failed! ***\n";
561 dbgs() << " has been released too many times!\n";
565 --PredSU->NumSuccsLeft;
566 if (PredSU->NumSuccsLeft == 0 && PredSU != &EntrySU)
567 SchedImpl->releaseBottomNode(PredSU);
570 /// releasePredecessors - Call releasePred on each of SU's predecessors.
571 void ScheduleDAGMI::releasePredecessors(SUnit *SU) {
572 for (SUnit::pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
574 releasePred(SU, &*I);
578 /// enterRegion - Called back from MachineScheduler::runOnMachineFunction after
579 /// crossing a scheduling boundary. [begin, end) includes all instructions in
580 /// the region, including the boundary itself and single-instruction regions
581 /// that don't get scheduled.
582 void ScheduleDAGMI::enterRegion(MachineBasicBlock *bb,
583 MachineBasicBlock::iterator begin,
584 MachineBasicBlock::iterator end,
585 unsigned regioninstrs)
587 ScheduleDAGInstrs::enterRegion(bb, begin, end, regioninstrs);
589 SchedImpl->initPolicy(begin, end, regioninstrs);
592 /// This is normally called from the main scheduler loop but may also be invoked
593 /// by the scheduling strategy to perform additional code motion.
594 void ScheduleDAGMI::moveInstruction(
595 MachineInstr *MI, MachineBasicBlock::iterator InsertPos) {
596 // Advance RegionBegin if the first instruction moves down.
597 if (&*RegionBegin == MI)
600 // Update the instruction stream.
601 BB->splice(InsertPos, BB, MI);
603 // Update LiveIntervals
605 LIS->handleMove(MI, /*UpdateFlags=*/true);
607 // Recede RegionBegin if an instruction moves above the first.
608 if (RegionBegin == InsertPos)
612 bool ScheduleDAGMI::checkSchedLimit() {
614 if (NumInstrsScheduled == MISchedCutoff && MISchedCutoff != ~0U) {
615 CurrentTop = CurrentBottom;
618 ++NumInstrsScheduled;
623 /// Per-region scheduling driver, called back from
624 /// MachineScheduler::runOnMachineFunction. This is a simplified driver that
625 /// does not consider liveness or register pressure. It is useful for PostRA
626 /// scheduling and potentially other custom schedulers.
627 void ScheduleDAGMI::schedule() {
631 Topo.InitDAGTopologicalSorting();
635 SmallVector<SUnit*, 8> TopRoots, BotRoots;
636 findRootsAndBiasEdges(TopRoots, BotRoots);
638 // Initialize the strategy before modifying the DAG.
639 // This may initialize a DFSResult to be used for queue priority.
640 SchedImpl->initialize(this);
642 DEBUG(for (unsigned su = 0, e = SUnits.size(); su != e; ++su)
643 SUnits[su].dumpAll(this));
644 if (ViewMISchedDAGs) viewGraph();
646 // Initialize ready queues now that the DAG and priority data are finalized.
647 initQueues(TopRoots, BotRoots);
649 bool IsTopNode = false;
650 while (SUnit *SU = SchedImpl->pickNode(IsTopNode)) {
651 assert(!SU->isScheduled && "Node already scheduled");
652 if (!checkSchedLimit())
655 MachineInstr *MI = SU->getInstr();
657 assert(SU->isTopReady() && "node still has unscheduled dependencies");
658 if (&*CurrentTop == MI)
659 CurrentTop = nextIfDebug(++CurrentTop, CurrentBottom);
661 moveInstruction(MI, CurrentTop);
664 assert(SU->isBottomReady() && "node still has unscheduled dependencies");
665 MachineBasicBlock::iterator priorII =
666 priorNonDebug(CurrentBottom, CurrentTop);
668 CurrentBottom = priorII;
670 if (&*CurrentTop == MI)
671 CurrentTop = nextIfDebug(++CurrentTop, priorII);
672 moveInstruction(MI, CurrentBottom);
676 updateQueues(SU, IsTopNode);
678 // Notify the scheduling strategy after updating the DAG.
679 SchedImpl->schedNode(SU, IsTopNode);
681 assert(CurrentTop == CurrentBottom && "Nonempty unscheduled zone.");
686 unsigned BBNum = begin()->getParent()->getNumber();
687 dbgs() << "*** Final schedule for BB#" << BBNum << " ***\n";
693 /// Apply each ScheduleDAGMutation step in order.
694 void ScheduleDAGMI::postprocessDAG() {
695 for (unsigned i = 0, e = Mutations.size(); i < e; ++i) {
696 Mutations[i]->apply(this);
701 findRootsAndBiasEdges(SmallVectorImpl<SUnit*> &TopRoots,
702 SmallVectorImpl<SUnit*> &BotRoots) {
703 for (std::vector<SUnit>::iterator
704 I = SUnits.begin(), E = SUnits.end(); I != E; ++I) {
706 assert(!SU->isBoundaryNode() && "Boundary node should not be in SUnits");
708 // Order predecessors so DFSResult follows the critical path.
709 SU->biasCriticalPath();
711 // A SUnit is ready to top schedule if it has no predecessors.
712 if (!I->NumPredsLeft)
713 TopRoots.push_back(SU);
714 // A SUnit is ready to bottom schedule if it has no successors.
715 if (!I->NumSuccsLeft)
716 BotRoots.push_back(SU);
718 ExitSU.biasCriticalPath();
721 /// Identify DAG roots and setup scheduler queues.
722 void ScheduleDAGMI::initQueues(ArrayRef<SUnit*> TopRoots,
723 ArrayRef<SUnit*> BotRoots) {
724 NextClusterSucc = NULL;
725 NextClusterPred = NULL;
727 // Release all DAG roots for scheduling, not including EntrySU/ExitSU.
729 // Nodes with unreleased weak edges can still be roots.
730 // Release top roots in forward order.
731 for (SmallVectorImpl<SUnit*>::const_iterator
732 I = TopRoots.begin(), E = TopRoots.end(); I != E; ++I) {
733 SchedImpl->releaseTopNode(*I);
735 // Release bottom roots in reverse order so the higher priority nodes appear
736 // first. This is more natural and slightly more efficient.
737 for (SmallVectorImpl<SUnit*>::const_reverse_iterator
738 I = BotRoots.rbegin(), E = BotRoots.rend(); I != E; ++I) {
739 SchedImpl->releaseBottomNode(*I);
742 releaseSuccessors(&EntrySU);
743 releasePredecessors(&ExitSU);
745 SchedImpl->registerRoots();
747 // Advance past initial DebugValues.
748 CurrentTop = nextIfDebug(RegionBegin, RegionEnd);
749 CurrentBottom = RegionEnd;
752 /// Update scheduler queues after scheduling an instruction.
753 void ScheduleDAGMI::updateQueues(SUnit *SU, bool IsTopNode) {
754 // Release dependent instructions for scheduling.
756 releaseSuccessors(SU);
758 releasePredecessors(SU);
760 SU->isScheduled = true;
763 /// Reinsert any remaining debug_values, just like the PostRA scheduler.
764 void ScheduleDAGMI::placeDebugValues() {
765 // If first instruction was a DBG_VALUE then put it back.
767 BB->splice(RegionBegin, BB, FirstDbgValue);
768 RegionBegin = FirstDbgValue;
771 for (std::vector<std::pair<MachineInstr *, MachineInstr *> >::iterator
772 DI = DbgValues.end(), DE = DbgValues.begin(); DI != DE; --DI) {
773 std::pair<MachineInstr *, MachineInstr *> P = *std::prev(DI);
774 MachineInstr *DbgValue = P.first;
775 MachineBasicBlock::iterator OrigPrevMI = P.second;
776 if (&*RegionBegin == DbgValue)
778 BB->splice(++OrigPrevMI, BB, DbgValue);
779 if (OrigPrevMI == std::prev(RegionEnd))
780 RegionEnd = DbgValue;
783 FirstDbgValue = NULL;
786 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
787 void ScheduleDAGMI::dumpSchedule() const {
788 for (MachineBasicBlock::iterator MI = begin(), ME = end(); MI != ME; ++MI) {
789 if (SUnit *SU = getSUnit(&(*MI)))
792 dbgs() << "Missing SUnit\n";
797 //===----------------------------------------------------------------------===//
798 // ScheduleDAGMILive - Base class for MachineInstr scheduling with LiveIntervals
800 //===----------------------------------------------------------------------===//
802 ScheduleDAGMILive::~ScheduleDAGMILive() {
806 /// enterRegion - Called back from MachineScheduler::runOnMachineFunction after
807 /// crossing a scheduling boundary. [begin, end) includes all instructions in
808 /// the region, including the boundary itself and single-instruction regions
809 /// that don't get scheduled.
810 void ScheduleDAGMILive::enterRegion(MachineBasicBlock *bb,
811 MachineBasicBlock::iterator begin,
812 MachineBasicBlock::iterator end,
813 unsigned regioninstrs)
815 // ScheduleDAGMI initializes SchedImpl's per-region policy.
816 ScheduleDAGMI::enterRegion(bb, begin, end, regioninstrs);
818 // For convenience remember the end of the liveness region.
819 LiveRegionEnd = (RegionEnd == bb->end()) ? RegionEnd : std::next(RegionEnd);
821 SUPressureDiffs.clear();
823 ShouldTrackPressure = SchedImpl->shouldTrackPressure();
826 // Setup the register pressure trackers for the top scheduled top and bottom
827 // scheduled regions.
828 void ScheduleDAGMILive::initRegPressure() {
829 TopRPTracker.init(&MF, RegClassInfo, LIS, BB, RegionBegin);
830 BotRPTracker.init(&MF, RegClassInfo, LIS, BB, LiveRegionEnd);
832 // Close the RPTracker to finalize live ins.
833 RPTracker.closeRegion();
835 DEBUG(RPTracker.dump());
837 // Initialize the live ins and live outs.
838 TopRPTracker.addLiveRegs(RPTracker.getPressure().LiveInRegs);
839 BotRPTracker.addLiveRegs(RPTracker.getPressure().LiveOutRegs);
841 // Close one end of the tracker so we can call
842 // getMaxUpward/DownwardPressureDelta before advancing across any
843 // instructions. This converts currently live regs into live ins/outs.
844 TopRPTracker.closeTop();
845 BotRPTracker.closeBottom();
847 BotRPTracker.initLiveThru(RPTracker);
848 if (!BotRPTracker.getLiveThru().empty()) {
849 TopRPTracker.initLiveThru(BotRPTracker.getLiveThru());
850 DEBUG(dbgs() << "Live Thru: ";
851 dumpRegSetPressure(BotRPTracker.getLiveThru(), TRI));
854 // For each live out vreg reduce the pressure change associated with other
855 // uses of the same vreg below the live-out reaching def.
856 updatePressureDiffs(RPTracker.getPressure().LiveOutRegs);
858 // Account for liveness generated by the region boundary.
859 if (LiveRegionEnd != RegionEnd) {
860 SmallVector<unsigned, 8> LiveUses;
861 BotRPTracker.recede(&LiveUses);
862 updatePressureDiffs(LiveUses);
865 assert(BotRPTracker.getPos() == RegionEnd && "Can't find the region bottom");
867 // Cache the list of excess pressure sets in this region. This will also track
868 // the max pressure in the scheduled code for these sets.
869 RegionCriticalPSets.clear();
870 const std::vector<unsigned> &RegionPressure =
871 RPTracker.getPressure().MaxSetPressure;
872 for (unsigned i = 0, e = RegionPressure.size(); i < e; ++i) {
873 unsigned Limit = RegClassInfo->getRegPressureSetLimit(i);
874 if (RegionPressure[i] > Limit) {
875 DEBUG(dbgs() << TRI->getRegPressureSetName(i)
876 << " Limit " << Limit
877 << " Actual " << RegionPressure[i] << "\n");
878 RegionCriticalPSets.push_back(PressureChange(i));
881 DEBUG(dbgs() << "Excess PSets: ";
882 for (unsigned i = 0, e = RegionCriticalPSets.size(); i != e; ++i)
883 dbgs() << TRI->getRegPressureSetName(
884 RegionCriticalPSets[i].getPSet()) << " ";
888 void ScheduleDAGMILive::
889 updateScheduledPressure(const SUnit *SU,
890 const std::vector<unsigned> &NewMaxPressure) {
891 const PressureDiff &PDiff = getPressureDiff(SU);
892 unsigned CritIdx = 0, CritEnd = RegionCriticalPSets.size();
893 for (PressureDiff::const_iterator I = PDiff.begin(), E = PDiff.end();
897 unsigned ID = I->getPSet();
898 while (CritIdx != CritEnd && RegionCriticalPSets[CritIdx].getPSet() < ID)
900 if (CritIdx != CritEnd && RegionCriticalPSets[CritIdx].getPSet() == ID) {
901 if ((int)NewMaxPressure[ID] > RegionCriticalPSets[CritIdx].getUnitInc()
902 && NewMaxPressure[ID] <= INT16_MAX)
903 RegionCriticalPSets[CritIdx].setUnitInc(NewMaxPressure[ID]);
905 unsigned Limit = RegClassInfo->getRegPressureSetLimit(ID);
906 if (NewMaxPressure[ID] >= Limit - 2) {
907 DEBUG(dbgs() << " " << TRI->getRegPressureSetName(ID) << ": "
908 << NewMaxPressure[ID] << " > " << Limit << "(+ "
909 << BotRPTracker.getLiveThru()[ID] << " livethru)\n");
914 /// Update the PressureDiff array for liveness after scheduling this
916 void ScheduleDAGMILive::updatePressureDiffs(ArrayRef<unsigned> LiveUses) {
917 for (unsigned LUIdx = 0, LUEnd = LiveUses.size(); LUIdx != LUEnd; ++LUIdx) {
918 /// FIXME: Currently assuming single-use physregs.
919 unsigned Reg = LiveUses[LUIdx];
920 DEBUG(dbgs() << " LiveReg: " << PrintVRegOrUnit(Reg, TRI) << "\n");
921 if (!TRI->isVirtualRegister(Reg))
924 // This may be called before CurrentBottom has been initialized. However,
925 // BotRPTracker must have a valid position. We want the value live into the
926 // instruction or live out of the block, so ask for the previous
927 // instruction's live-out.
928 const LiveInterval &LI = LIS->getInterval(Reg);
930 MachineBasicBlock::const_iterator I =
931 nextIfDebug(BotRPTracker.getPos(), BB->end());
933 VNI = LI.getVNInfoBefore(LIS->getMBBEndIdx(BB));
935 LiveQueryResult LRQ = LI.Query(LIS->getInstructionIndex(I));
938 // RegisterPressureTracker guarantees that readsReg is true for LiveUses.
939 assert(VNI && "No live value at use.");
940 for (VReg2UseMap::iterator
941 UI = VRegUses.find(Reg); UI != VRegUses.end(); ++UI) {
943 DEBUG(dbgs() << " UpdateRegP: SU(" << SU->NodeNum << ") "
945 // If this use comes before the reaching def, it cannot be a last use, so
946 // descrease its pressure change.
947 if (!SU->isScheduled && SU != &ExitSU) {
949 = LI.Query(LIS->getInstructionIndex(SU->getInstr()));
950 if (LRQ.valueIn() == VNI)
951 getPressureDiff(SU).addPressureChange(Reg, true, &MRI);
957 /// schedule - Called back from MachineScheduler::runOnMachineFunction
958 /// after setting up the current scheduling region. [RegionBegin, RegionEnd)
959 /// only includes instructions that have DAG nodes, not scheduling boundaries.
961 /// This is a skeletal driver, with all the functionality pushed into helpers,
962 /// so that it can be easilly extended by experimental schedulers. Generally,
963 /// implementing MachineSchedStrategy should be sufficient to implement a new
964 /// scheduling algorithm. However, if a scheduler further subclasses
965 /// ScheduleDAGMILive then it will want to override this virtual method in order
966 /// to update any specialized state.
967 void ScheduleDAGMILive::schedule() {
968 buildDAGWithRegPressure();
970 Topo.InitDAGTopologicalSorting();
974 SmallVector<SUnit*, 8> TopRoots, BotRoots;
975 findRootsAndBiasEdges(TopRoots, BotRoots);
977 // Initialize the strategy before modifying the DAG.
978 // This may initialize a DFSResult to be used for queue priority.
979 SchedImpl->initialize(this);
981 DEBUG(for (unsigned su = 0, e = SUnits.size(); su != e; ++su)
982 SUnits[su].dumpAll(this));
983 if (ViewMISchedDAGs) viewGraph();
985 // Initialize ready queues now that the DAG and priority data are finalized.
986 initQueues(TopRoots, BotRoots);
988 if (ShouldTrackPressure) {
989 assert(TopRPTracker.getPos() == RegionBegin && "bad initial Top tracker");
990 TopRPTracker.setPos(CurrentTop);
993 bool IsTopNode = false;
994 while (SUnit *SU = SchedImpl->pickNode(IsTopNode)) {
995 assert(!SU->isScheduled && "Node already scheduled");
996 if (!checkSchedLimit())
999 scheduleMI(SU, IsTopNode);
1001 updateQueues(SU, IsTopNode);
1004 unsigned SubtreeID = DFSResult->getSubtreeID(SU);
1005 if (!ScheduledTrees.test(SubtreeID)) {
1006 ScheduledTrees.set(SubtreeID);
1007 DFSResult->scheduleTree(SubtreeID);
1008 SchedImpl->scheduleTree(SubtreeID);
1012 // Notify the scheduling strategy after updating the DAG.
1013 SchedImpl->schedNode(SU, IsTopNode);
1015 assert(CurrentTop == CurrentBottom && "Nonempty unscheduled zone.");
1020 unsigned BBNum = begin()->getParent()->getNumber();
1021 dbgs() << "*** Final schedule for BB#" << BBNum << " ***\n";
1027 /// Build the DAG and setup three register pressure trackers.
1028 void ScheduleDAGMILive::buildDAGWithRegPressure() {
1029 if (!ShouldTrackPressure) {
1031 RegionCriticalPSets.clear();
1032 buildSchedGraph(AA);
1036 // Initialize the register pressure tracker used by buildSchedGraph.
1037 RPTracker.init(&MF, RegClassInfo, LIS, BB, LiveRegionEnd,
1038 /*TrackUntiedDefs=*/true);
1040 // Account for liveness generate by the region boundary.
1041 if (LiveRegionEnd != RegionEnd)
1044 // Build the DAG, and compute current register pressure.
1045 buildSchedGraph(AA, &RPTracker, &SUPressureDiffs);
1047 // Initialize top/bottom trackers after computing region pressure.
1051 void ScheduleDAGMILive::computeDFSResult() {
1053 DFSResult = new SchedDFSResult(/*BottomU*/true, MinSubtreeSize);
1055 ScheduledTrees.clear();
1056 DFSResult->resize(SUnits.size());
1057 DFSResult->compute(SUnits);
1058 ScheduledTrees.resize(DFSResult->getNumSubtrees());
1061 /// Compute the max cyclic critical path through the DAG. The scheduling DAG
1062 /// only provides the critical path for single block loops. To handle loops that
1063 /// span blocks, we could use the vreg path latencies provided by
1064 /// MachineTraceMetrics instead. However, MachineTraceMetrics is not currently
1065 /// available for use in the scheduler.
1067 /// The cyclic path estimation identifies a def-use pair that crosses the back
1068 /// edge and considers the depth and height of the nodes. For example, consider
1069 /// the following instruction sequence where each instruction has unit latency
1070 /// and defines an epomymous virtual register:
1072 /// a->b(a,c)->c(b)->d(c)->exit
1074 /// The cyclic critical path is a two cycles: b->c->b
1075 /// The acyclic critical path is four cycles: a->b->c->d->exit
1076 /// LiveOutHeight = height(c) = len(c->d->exit) = 2
1077 /// LiveOutDepth = depth(c) + 1 = len(a->b->c) + 1 = 3
1078 /// LiveInHeight = height(b) + 1 = len(b->c->d->exit) + 1 = 4
1079 /// LiveInDepth = depth(b) = len(a->b) = 1
1081 /// LiveOutDepth - LiveInDepth = 3 - 1 = 2
1082 /// LiveInHeight - LiveOutHeight = 4 - 2 = 2
1083 /// CyclicCriticalPath = min(2, 2) = 2
1085 /// This could be relevant to PostRA scheduling, but is currently implemented
1086 /// assuming LiveIntervals.
1087 unsigned ScheduleDAGMILive::computeCyclicCriticalPath() {
1088 // This only applies to single block loop.
1089 if (!BB->isSuccessor(BB))
1092 unsigned MaxCyclicLatency = 0;
1093 // Visit each live out vreg def to find def/use pairs that cross iterations.
1094 ArrayRef<unsigned> LiveOuts = RPTracker.getPressure().LiveOutRegs;
1095 for (ArrayRef<unsigned>::iterator RI = LiveOuts.begin(), RE = LiveOuts.end();
1098 if (!TRI->isVirtualRegister(Reg))
1100 const LiveInterval &LI = LIS->getInterval(Reg);
1101 const VNInfo *DefVNI = LI.getVNInfoBefore(LIS->getMBBEndIdx(BB));
1105 MachineInstr *DefMI = LIS->getInstructionFromIndex(DefVNI->def);
1106 const SUnit *DefSU = getSUnit(DefMI);
1110 unsigned LiveOutHeight = DefSU->getHeight();
1111 unsigned LiveOutDepth = DefSU->getDepth() + DefSU->Latency;
1112 // Visit all local users of the vreg def.
1113 for (VReg2UseMap::iterator
1114 UI = VRegUses.find(Reg); UI != VRegUses.end(); ++UI) {
1115 if (UI->SU == &ExitSU)
1118 // Only consider uses of the phi.
1119 LiveQueryResult LRQ =
1120 LI.Query(LIS->getInstructionIndex(UI->SU->getInstr()));
1121 if (!LRQ.valueIn()->isPHIDef())
1124 // Assume that a path spanning two iterations is a cycle, which could
1125 // overestimate in strange cases. This allows cyclic latency to be
1126 // estimated as the minimum slack of the vreg's depth or height.
1127 unsigned CyclicLatency = 0;
1128 if (LiveOutDepth > UI->SU->getDepth())
1129 CyclicLatency = LiveOutDepth - UI->SU->getDepth();
1131 unsigned LiveInHeight = UI->SU->getHeight() + DefSU->Latency;
1132 if (LiveInHeight > LiveOutHeight) {
1133 if (LiveInHeight - LiveOutHeight < CyclicLatency)
1134 CyclicLatency = LiveInHeight - LiveOutHeight;
1139 DEBUG(dbgs() << "Cyclic Path: SU(" << DefSU->NodeNum << ") -> SU("
1140 << UI->SU->NodeNum << ") = " << CyclicLatency << "c\n");
1141 if (CyclicLatency > MaxCyclicLatency)
1142 MaxCyclicLatency = CyclicLatency;
1145 DEBUG(dbgs() << "Cyclic Critical Path: " << MaxCyclicLatency << "c\n");
1146 return MaxCyclicLatency;
1149 /// Move an instruction and update register pressure.
1150 void ScheduleDAGMILive::scheduleMI(SUnit *SU, bool IsTopNode) {
1151 // Move the instruction to its new location in the instruction stream.
1152 MachineInstr *MI = SU->getInstr();
1155 assert(SU->isTopReady() && "node still has unscheduled dependencies");
1156 if (&*CurrentTop == MI)
1157 CurrentTop = nextIfDebug(++CurrentTop, CurrentBottom);
1159 moveInstruction(MI, CurrentTop);
1160 TopRPTracker.setPos(MI);
1163 if (ShouldTrackPressure) {
1164 // Update top scheduled pressure.
1165 TopRPTracker.advance();
1166 assert(TopRPTracker.getPos() == CurrentTop && "out of sync");
1167 updateScheduledPressure(SU, TopRPTracker.getPressure().MaxSetPressure);
1171 assert(SU->isBottomReady() && "node still has unscheduled dependencies");
1172 MachineBasicBlock::iterator priorII =
1173 priorNonDebug(CurrentBottom, CurrentTop);
1174 if (&*priorII == MI)
1175 CurrentBottom = priorII;
1177 if (&*CurrentTop == MI) {
1178 CurrentTop = nextIfDebug(++CurrentTop, priorII);
1179 TopRPTracker.setPos(CurrentTop);
1181 moveInstruction(MI, CurrentBottom);
1184 if (ShouldTrackPressure) {
1185 // Update bottom scheduled pressure.
1186 SmallVector<unsigned, 8> LiveUses;
1187 BotRPTracker.recede(&LiveUses);
1188 assert(BotRPTracker.getPos() == CurrentBottom && "out of sync");
1189 updateScheduledPressure(SU, BotRPTracker.getPressure().MaxSetPressure);
1190 updatePressureDiffs(LiveUses);
1195 //===----------------------------------------------------------------------===//
1196 // LoadClusterMutation - DAG post-processing to cluster loads.
1197 //===----------------------------------------------------------------------===//
1200 /// \brief Post-process the DAG to create cluster edges between neighboring
1202 class LoadClusterMutation : public ScheduleDAGMutation {
1207 LoadInfo(SUnit *su, unsigned reg, unsigned ofs)
1208 : SU(su), BaseReg(reg), Offset(ofs) {}
1210 static bool LoadInfoLess(const LoadClusterMutation::LoadInfo &LHS,
1211 const LoadClusterMutation::LoadInfo &RHS);
1213 const TargetInstrInfo *TII;
1214 const TargetRegisterInfo *TRI;
1216 LoadClusterMutation(const TargetInstrInfo *tii,
1217 const TargetRegisterInfo *tri)
1218 : TII(tii), TRI(tri) {}
1220 virtual void apply(ScheduleDAGMI *DAG);
1222 void clusterNeighboringLoads(ArrayRef<SUnit*> Loads, ScheduleDAGMI *DAG);
1226 bool LoadClusterMutation::LoadInfoLess(
1227 const LoadClusterMutation::LoadInfo &LHS,
1228 const LoadClusterMutation::LoadInfo &RHS) {
1229 if (LHS.BaseReg != RHS.BaseReg)
1230 return LHS.BaseReg < RHS.BaseReg;
1231 return LHS.Offset < RHS.Offset;
1234 void LoadClusterMutation::clusterNeighboringLoads(ArrayRef<SUnit*> Loads,
1235 ScheduleDAGMI *DAG) {
1236 SmallVector<LoadClusterMutation::LoadInfo,32> LoadRecords;
1237 for (unsigned Idx = 0, End = Loads.size(); Idx != End; ++Idx) {
1238 SUnit *SU = Loads[Idx];
1241 if (TII->getLdStBaseRegImmOfs(SU->getInstr(), BaseReg, Offset, TRI))
1242 LoadRecords.push_back(LoadInfo(SU, BaseReg, Offset));
1244 if (LoadRecords.size() < 2)
1246 std::sort(LoadRecords.begin(), LoadRecords.end(), LoadInfoLess);
1247 unsigned ClusterLength = 1;
1248 for (unsigned Idx = 0, End = LoadRecords.size(); Idx < (End - 1); ++Idx) {
1249 if (LoadRecords[Idx].BaseReg != LoadRecords[Idx+1].BaseReg) {
1254 SUnit *SUa = LoadRecords[Idx].SU;
1255 SUnit *SUb = LoadRecords[Idx+1].SU;
1256 if (TII->shouldClusterLoads(SUa->getInstr(), SUb->getInstr(), ClusterLength)
1257 && DAG->addEdge(SUb, SDep(SUa, SDep::Cluster))) {
1259 DEBUG(dbgs() << "Cluster loads SU(" << SUa->NodeNum << ") - SU("
1260 << SUb->NodeNum << ")\n");
1261 // Copy successor edges from SUa to SUb. Interleaving computation
1262 // dependent on SUa can prevent load combining due to register reuse.
1263 // Predecessor edges do not need to be copied from SUb to SUa since nearby
1264 // loads should have effectively the same inputs.
1265 for (SUnit::const_succ_iterator
1266 SI = SUa->Succs.begin(), SE = SUa->Succs.end(); SI != SE; ++SI) {
1267 if (SI->getSUnit() == SUb)
1269 DEBUG(dbgs() << " Copy Succ SU(" << SI->getSUnit()->NodeNum << ")\n");
1270 DAG->addEdge(SI->getSUnit(), SDep(SUb, SDep::Artificial));
1279 /// \brief Callback from DAG postProcessing to create cluster edges for loads.
1280 void LoadClusterMutation::apply(ScheduleDAGMI *DAG) {
1281 // Map DAG NodeNum to store chain ID.
1282 DenseMap<unsigned, unsigned> StoreChainIDs;
1283 // Map each store chain to a set of dependent loads.
1284 SmallVector<SmallVector<SUnit*,4>, 32> StoreChainDependents;
1285 for (unsigned Idx = 0, End = DAG->SUnits.size(); Idx != End; ++Idx) {
1286 SUnit *SU = &DAG->SUnits[Idx];
1287 if (!SU->getInstr()->mayLoad())
1289 unsigned ChainPredID = DAG->SUnits.size();
1290 for (SUnit::const_pred_iterator
1291 PI = SU->Preds.begin(), PE = SU->Preds.end(); PI != PE; ++PI) {
1293 ChainPredID = PI->getSUnit()->NodeNum;
1297 // Check if this chain-like pred has been seen
1298 // before. ChainPredID==MaxNodeID for loads at the top of the schedule.
1299 unsigned NumChains = StoreChainDependents.size();
1300 std::pair<DenseMap<unsigned, unsigned>::iterator, bool> Result =
1301 StoreChainIDs.insert(std::make_pair(ChainPredID, NumChains));
1303 StoreChainDependents.resize(NumChains + 1);
1304 StoreChainDependents[Result.first->second].push_back(SU);
1306 // Iterate over the store chains.
1307 for (unsigned Idx = 0, End = StoreChainDependents.size(); Idx != End; ++Idx)
1308 clusterNeighboringLoads(StoreChainDependents[Idx], DAG);
1311 //===----------------------------------------------------------------------===//
1312 // MacroFusion - DAG post-processing to encourage fusion of macro ops.
1313 //===----------------------------------------------------------------------===//
1316 /// \brief Post-process the DAG to create cluster edges between instructions
1317 /// that may be fused by the processor into a single operation.
1318 class MacroFusion : public ScheduleDAGMutation {
1319 const TargetInstrInfo *TII;
1321 MacroFusion(const TargetInstrInfo *tii): TII(tii) {}
1323 virtual void apply(ScheduleDAGMI *DAG);
1327 /// \brief Callback from DAG postProcessing to create cluster edges to encourage
1328 /// fused operations.
1329 void MacroFusion::apply(ScheduleDAGMI *DAG) {
1330 // For now, assume targets can only fuse with the branch.
1331 MachineInstr *Branch = DAG->ExitSU.getInstr();
1335 for (unsigned Idx = DAG->SUnits.size(); Idx > 0;) {
1336 SUnit *SU = &DAG->SUnits[--Idx];
1337 if (!TII->shouldScheduleAdjacent(SU->getInstr(), Branch))
1340 // Create a single weak edge from SU to ExitSU. The only effect is to cause
1341 // bottom-up scheduling to heavily prioritize the clustered SU. There is no
1342 // need to copy predecessor edges from ExitSU to SU, since top-down
1343 // scheduling cannot prioritize ExitSU anyway. To defer top-down scheduling
1344 // of SU, we could create an artificial edge from the deepest root, but it
1345 // hasn't been needed yet.
1346 bool Success = DAG->addEdge(&DAG->ExitSU, SDep(SU, SDep::Cluster));
1348 assert(Success && "No DAG nodes should be reachable from ExitSU");
1350 DEBUG(dbgs() << "Macro Fuse SU(" << SU->NodeNum << ")\n");
1355 //===----------------------------------------------------------------------===//
1356 // CopyConstrain - DAG post-processing to encourage copy elimination.
1357 //===----------------------------------------------------------------------===//
1360 /// \brief Post-process the DAG to create weak edges from all uses of a copy to
1361 /// the one use that defines the copy's source vreg, most likely an induction
1362 /// variable increment.
1363 class CopyConstrain : public ScheduleDAGMutation {
1365 SlotIndex RegionBeginIdx;
1366 // RegionEndIdx is the slot index of the last non-debug instruction in the
1367 // scheduling region. So we may have RegionBeginIdx == RegionEndIdx.
1368 SlotIndex RegionEndIdx;
1370 CopyConstrain(const TargetInstrInfo *, const TargetRegisterInfo *) {}
1372 virtual void apply(ScheduleDAGMI *DAG);
1375 void constrainLocalCopy(SUnit *CopySU, ScheduleDAGMILive *DAG);
1379 /// constrainLocalCopy handles two possibilities:
1384 /// I3: dst = src (copy)
1385 /// (create pred->succ edges I0->I1, I2->I1)
1388 /// I0: dst = src (copy)
1392 /// (create pred->succ edges I1->I2, I3->I2)
1394 /// Although the MachineScheduler is currently constrained to single blocks,
1395 /// this algorithm should handle extended blocks. An EBB is a set of
1396 /// contiguously numbered blocks such that the previous block in the EBB is
1397 /// always the single predecessor.
1398 void CopyConstrain::constrainLocalCopy(SUnit *CopySU, ScheduleDAGMILive *DAG) {
1399 LiveIntervals *LIS = DAG->getLIS();
1400 MachineInstr *Copy = CopySU->getInstr();
1402 // Check for pure vreg copies.
1403 unsigned SrcReg = Copy->getOperand(1).getReg();
1404 if (!TargetRegisterInfo::isVirtualRegister(SrcReg))
1407 unsigned DstReg = Copy->getOperand(0).getReg();
1408 if (!TargetRegisterInfo::isVirtualRegister(DstReg))
1411 // Check if either the dest or source is local. If it's live across a back
1412 // edge, it's not local. Note that if both vregs are live across the back
1413 // edge, we cannot successfully contrain the copy without cyclic scheduling.
1414 unsigned LocalReg = DstReg;
1415 unsigned GlobalReg = SrcReg;
1416 LiveInterval *LocalLI = &LIS->getInterval(LocalReg);
1417 if (!LocalLI->isLocal(RegionBeginIdx, RegionEndIdx)) {
1420 LocalLI = &LIS->getInterval(LocalReg);
1421 if (!LocalLI->isLocal(RegionBeginIdx, RegionEndIdx))
1424 LiveInterval *GlobalLI = &LIS->getInterval(GlobalReg);
1426 // Find the global segment after the start of the local LI.
1427 LiveInterval::iterator GlobalSegment = GlobalLI->find(LocalLI->beginIndex());
1428 // If GlobalLI does not overlap LocalLI->start, then a copy directly feeds a
1429 // local live range. We could create edges from other global uses to the local
1430 // start, but the coalescer should have already eliminated these cases, so
1431 // don't bother dealing with it.
1432 if (GlobalSegment == GlobalLI->end())
1435 // If GlobalSegment is killed at the LocalLI->start, the call to find()
1436 // returned the next global segment. But if GlobalSegment overlaps with
1437 // LocalLI->start, then advance to the next segement. If a hole in GlobalLI
1438 // exists in LocalLI's vicinity, GlobalSegment will be the end of the hole.
1439 if (GlobalSegment->contains(LocalLI->beginIndex()))
1442 if (GlobalSegment == GlobalLI->end())
1445 // Check if GlobalLI contains a hole in the vicinity of LocalLI.
1446 if (GlobalSegment != GlobalLI->begin()) {
1447 // Two address defs have no hole.
1448 if (SlotIndex::isSameInstr(std::prev(GlobalSegment)->end,
1449 GlobalSegment->start)) {
1452 // If the prior global segment may be defined by the same two-address
1453 // instruction that also defines LocalLI, then can't make a hole here.
1454 if (SlotIndex::isSameInstr(std::prev(GlobalSegment)->start,
1455 LocalLI->beginIndex())) {
1458 // If GlobalLI has a prior segment, it must be live into the EBB. Otherwise
1459 // it would be a disconnected component in the live range.
1460 assert(std::prev(GlobalSegment)->start < LocalLI->beginIndex() &&
1461 "Disconnected LRG within the scheduling region.");
1463 MachineInstr *GlobalDef = LIS->getInstructionFromIndex(GlobalSegment->start);
1467 SUnit *GlobalSU = DAG->getSUnit(GlobalDef);
1471 // GlobalDef is the bottom of the GlobalLI hole. Open the hole by
1472 // constraining the uses of the last local def to precede GlobalDef.
1473 SmallVector<SUnit*,8> LocalUses;
1474 const VNInfo *LastLocalVN = LocalLI->getVNInfoBefore(LocalLI->endIndex());
1475 MachineInstr *LastLocalDef = LIS->getInstructionFromIndex(LastLocalVN->def);
1476 SUnit *LastLocalSU = DAG->getSUnit(LastLocalDef);
1477 for (SUnit::const_succ_iterator
1478 I = LastLocalSU->Succs.begin(), E = LastLocalSU->Succs.end();
1480 if (I->getKind() != SDep::Data || I->getReg() != LocalReg)
1482 if (I->getSUnit() == GlobalSU)
1484 if (!DAG->canAddEdge(GlobalSU, I->getSUnit()))
1486 LocalUses.push_back(I->getSUnit());
1488 // Open the top of the GlobalLI hole by constraining any earlier global uses
1489 // to precede the start of LocalLI.
1490 SmallVector<SUnit*,8> GlobalUses;
1491 MachineInstr *FirstLocalDef =
1492 LIS->getInstructionFromIndex(LocalLI->beginIndex());
1493 SUnit *FirstLocalSU = DAG->getSUnit(FirstLocalDef);
1494 for (SUnit::const_pred_iterator
1495 I = GlobalSU->Preds.begin(), E = GlobalSU->Preds.end(); I != E; ++I) {
1496 if (I->getKind() != SDep::Anti || I->getReg() != GlobalReg)
1498 if (I->getSUnit() == FirstLocalSU)
1500 if (!DAG->canAddEdge(FirstLocalSU, I->getSUnit()))
1502 GlobalUses.push_back(I->getSUnit());
1504 DEBUG(dbgs() << "Constraining copy SU(" << CopySU->NodeNum << ")\n");
1505 // Add the weak edges.
1506 for (SmallVectorImpl<SUnit*>::const_iterator
1507 I = LocalUses.begin(), E = LocalUses.end(); I != E; ++I) {
1508 DEBUG(dbgs() << " Local use SU(" << (*I)->NodeNum << ") -> SU("
1509 << GlobalSU->NodeNum << ")\n");
1510 DAG->addEdge(GlobalSU, SDep(*I, SDep::Weak));
1512 for (SmallVectorImpl<SUnit*>::const_iterator
1513 I = GlobalUses.begin(), E = GlobalUses.end(); I != E; ++I) {
1514 DEBUG(dbgs() << " Global use SU(" << (*I)->NodeNum << ") -> SU("
1515 << FirstLocalSU->NodeNum << ")\n");
1516 DAG->addEdge(FirstLocalSU, SDep(*I, SDep::Weak));
1520 /// \brief Callback from DAG postProcessing to create weak edges to encourage
1521 /// copy elimination.
1522 void CopyConstrain::apply(ScheduleDAGMI *DAG) {
1523 assert(DAG->hasVRegLiveness() && "Expect VRegs with LiveIntervals");
1525 MachineBasicBlock::iterator FirstPos = nextIfDebug(DAG->begin(), DAG->end());
1526 if (FirstPos == DAG->end())
1528 RegionBeginIdx = DAG->getLIS()->getInstructionIndex(&*FirstPos);
1529 RegionEndIdx = DAG->getLIS()->getInstructionIndex(
1530 &*priorNonDebug(DAG->end(), DAG->begin()));
1532 for (unsigned Idx = 0, End = DAG->SUnits.size(); Idx != End; ++Idx) {
1533 SUnit *SU = &DAG->SUnits[Idx];
1534 if (!SU->getInstr()->isCopy())
1537 constrainLocalCopy(SU, static_cast<ScheduleDAGMILive*>(DAG));
1541 //===----------------------------------------------------------------------===//
1542 // MachineSchedStrategy helpers used by GenericScheduler, GenericPostScheduler
1543 // and possibly other custom schedulers.
1544 //===----------------------------------------------------------------------===//
1546 static const unsigned InvalidCycle = ~0U;
1548 SchedBoundary::~SchedBoundary() { delete HazardRec; }
1550 void SchedBoundary::reset() {
1551 // A new HazardRec is created for each DAG and owned by SchedBoundary.
1552 // Destroying and reconstructing it is very expensive though. So keep
1553 // invalid, placeholder HazardRecs.
1554 if (HazardRec && HazardRec->isEnabled()) {
1560 CheckPending = false;
1564 MinReadyCycle = UINT_MAX;
1565 ExpectedLatency = 0;
1566 DependentLatency = 0;
1568 MaxExecutedResCount = 0;
1570 IsResourceLimited = false;
1571 ReservedCycles.clear();
1573 // Track the maximum number of stall cycles that could arise either from the
1574 // latency of a DAG edge or the number of cycles that a processor resource is
1575 // reserved (SchedBoundary::ReservedCycles).
1576 MaxObservedLatency = 0;
1578 // Reserve a zero-count for invalid CritResIdx.
1579 ExecutedResCounts.resize(1);
1580 assert(!ExecutedResCounts[0] && "nonzero count for bad resource");
1583 void SchedRemainder::
1584 init(ScheduleDAGMI *DAG, const TargetSchedModel *SchedModel) {
1586 if (!SchedModel->hasInstrSchedModel())
1588 RemainingCounts.resize(SchedModel->getNumProcResourceKinds());
1589 for (std::vector<SUnit>::iterator
1590 I = DAG->SUnits.begin(), E = DAG->SUnits.end(); I != E; ++I) {
1591 const MCSchedClassDesc *SC = DAG->getSchedClass(&*I);
1592 RemIssueCount += SchedModel->getNumMicroOps(I->getInstr(), SC)
1593 * SchedModel->getMicroOpFactor();
1594 for (TargetSchedModel::ProcResIter
1595 PI = SchedModel->getWriteProcResBegin(SC),
1596 PE = SchedModel->getWriteProcResEnd(SC); PI != PE; ++PI) {
1597 unsigned PIdx = PI->ProcResourceIdx;
1598 unsigned Factor = SchedModel->getResourceFactor(PIdx);
1599 RemainingCounts[PIdx] += (Factor * PI->Cycles);
1604 void SchedBoundary::
1605 init(ScheduleDAGMI *dag, const TargetSchedModel *smodel, SchedRemainder *rem) {
1608 SchedModel = smodel;
1610 if (SchedModel->hasInstrSchedModel()) {
1611 ExecutedResCounts.resize(SchedModel->getNumProcResourceKinds());
1612 ReservedCycles.resize(SchedModel->getNumProcResourceKinds(), InvalidCycle);
1616 /// Compute the stall cycles based on this SUnit's ready time. Heuristics treat
1617 /// these "soft stalls" differently than the hard stall cycles based on CPU
1618 /// resources and computed by checkHazard(). A fully in-order model
1619 /// (MicroOpBufferSize==0) will not make use of this since instructions are not
1620 /// available for scheduling until they are ready. However, a weaker in-order
1621 /// model may use this for heuristics. For example, if a processor has in-order
1622 /// behavior when reading certain resources, this may come into play.
1623 unsigned SchedBoundary::getLatencyStallCycles(SUnit *SU) {
1624 if (!SU->isUnbuffered)
1627 unsigned ReadyCycle = (isTop() ? SU->TopReadyCycle : SU->BotReadyCycle);
1628 if (ReadyCycle > CurrCycle)
1629 return ReadyCycle - CurrCycle;
1633 /// Compute the next cycle at which the given processor resource can be
1635 unsigned SchedBoundary::
1636 getNextResourceCycle(unsigned PIdx, unsigned Cycles) {
1637 unsigned NextUnreserved = ReservedCycles[PIdx];
1638 // If this resource has never been used, always return cycle zero.
1639 if (NextUnreserved == InvalidCycle)
1641 // For bottom-up scheduling add the cycles needed for the current operation.
1643 NextUnreserved += Cycles;
1644 return NextUnreserved;
1647 /// Does this SU have a hazard within the current instruction group.
1649 /// The scheduler supports two modes of hazard recognition. The first is the
1650 /// ScheduleHazardRecognizer API. It is a fully general hazard recognizer that
1651 /// supports highly complicated in-order reservation tables
1652 /// (ScoreboardHazardRecognizer) and arbitraty target-specific logic.
1654 /// The second is a streamlined mechanism that checks for hazards based on
1655 /// simple counters that the scheduler itself maintains. It explicitly checks
1656 /// for instruction dispatch limitations, including the number of micro-ops that
1657 /// can dispatch per cycle.
1659 /// TODO: Also check whether the SU must start a new group.
1660 bool SchedBoundary::checkHazard(SUnit *SU) {
1661 if (HazardRec->isEnabled()
1662 && HazardRec->getHazardType(SU) != ScheduleHazardRecognizer::NoHazard) {
1665 unsigned uops = SchedModel->getNumMicroOps(SU->getInstr());
1666 if ((CurrMOps > 0) && (CurrMOps + uops > SchedModel->getIssueWidth())) {
1667 DEBUG(dbgs() << " SU(" << SU->NodeNum << ") uops="
1668 << SchedModel->getNumMicroOps(SU->getInstr()) << '\n');
1671 if (SchedModel->hasInstrSchedModel() && SU->hasReservedResource) {
1672 const MCSchedClassDesc *SC = DAG->getSchedClass(SU);
1673 for (TargetSchedModel::ProcResIter
1674 PI = SchedModel->getWriteProcResBegin(SC),
1675 PE = SchedModel->getWriteProcResEnd(SC); PI != PE; ++PI) {
1676 if (getNextResourceCycle(PI->ProcResourceIdx, PI->Cycles) > CurrCycle)
1683 // Find the unscheduled node in ReadySUs with the highest latency.
1684 unsigned SchedBoundary::
1685 findMaxLatency(ArrayRef<SUnit*> ReadySUs) {
1687 unsigned RemLatency = 0;
1688 for (ArrayRef<SUnit*>::iterator I = ReadySUs.begin(), E = ReadySUs.end();
1690 unsigned L = getUnscheduledLatency(*I);
1691 if (L > RemLatency) {
1697 DEBUG(dbgs() << Available.getName() << " RemLatency SU("
1698 << LateSU->NodeNum << ") " << RemLatency << "c\n");
1703 // Count resources in this zone and the remaining unscheduled
1704 // instruction. Return the max count, scaled. Set OtherCritIdx to the critical
1705 // resource index, or zero if the zone is issue limited.
1706 unsigned SchedBoundary::
1707 getOtherResourceCount(unsigned &OtherCritIdx) {
1709 if (!SchedModel->hasInstrSchedModel())
1712 unsigned OtherCritCount = Rem->RemIssueCount
1713 + (RetiredMOps * SchedModel->getMicroOpFactor());
1714 DEBUG(dbgs() << " " << Available.getName() << " + Remain MOps: "
1715 << OtherCritCount / SchedModel->getMicroOpFactor() << '\n');
1716 for (unsigned PIdx = 1, PEnd = SchedModel->getNumProcResourceKinds();
1717 PIdx != PEnd; ++PIdx) {
1718 unsigned OtherCount = getResourceCount(PIdx) + Rem->RemainingCounts[PIdx];
1719 if (OtherCount > OtherCritCount) {
1720 OtherCritCount = OtherCount;
1721 OtherCritIdx = PIdx;
1725 DEBUG(dbgs() << " " << Available.getName() << " + Remain CritRes: "
1726 << OtherCritCount / SchedModel->getResourceFactor(OtherCritIdx)
1727 << " " << SchedModel->getResourceName(OtherCritIdx) << "\n");
1729 return OtherCritCount;
1732 void SchedBoundary::releaseNode(SUnit *SU, unsigned ReadyCycle) {
1733 if (ReadyCycle < MinReadyCycle)
1734 MinReadyCycle = ReadyCycle;
1736 // Check for interlocks first. For the purpose of other heuristics, an
1737 // instruction that cannot issue appears as if it's not in the ReadyQueue.
1738 bool IsBuffered = SchedModel->getMicroOpBufferSize() != 0;
1739 if ((!IsBuffered && ReadyCycle > CurrCycle) || checkHazard(SU))
1744 // Record this node as an immediate dependent of the scheduled node.
1748 void SchedBoundary::releaseTopNode(SUnit *SU) {
1749 if (SU->isScheduled)
1752 for (SUnit::pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
1756 unsigned PredReadyCycle = I->getSUnit()->TopReadyCycle;
1757 unsigned Latency = I->getLatency();
1759 MaxObservedLatency = std::max(Latency, MaxObservedLatency);
1761 if (SU->TopReadyCycle < PredReadyCycle + Latency)
1762 SU->TopReadyCycle = PredReadyCycle + Latency;
1764 releaseNode(SU, SU->TopReadyCycle);
1767 void SchedBoundary::releaseBottomNode(SUnit *SU) {
1768 if (SU->isScheduled)
1771 assert(SU->getInstr() && "Scheduled SUnit must have instr");
1773 for (SUnit::succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
1777 unsigned SuccReadyCycle = I->getSUnit()->BotReadyCycle;
1778 unsigned Latency = I->getLatency();
1780 MaxObservedLatency = std::max(Latency, MaxObservedLatency);
1782 if (SU->BotReadyCycle < SuccReadyCycle + Latency)
1783 SU->BotReadyCycle = SuccReadyCycle + Latency;
1785 releaseNode(SU, SU->BotReadyCycle);
1788 /// Move the boundary of scheduled code by one cycle.
1789 void SchedBoundary::bumpCycle(unsigned NextCycle) {
1790 if (SchedModel->getMicroOpBufferSize() == 0) {
1791 assert(MinReadyCycle < UINT_MAX && "MinReadyCycle uninitialized");
1792 if (MinReadyCycle > NextCycle)
1793 NextCycle = MinReadyCycle;
1795 // Update the current micro-ops, which will issue in the next cycle.
1796 unsigned DecMOps = SchedModel->getIssueWidth() * (NextCycle - CurrCycle);
1797 CurrMOps = (CurrMOps <= DecMOps) ? 0 : CurrMOps - DecMOps;
1799 // Decrement DependentLatency based on the next cycle.
1800 if ((NextCycle - CurrCycle) > DependentLatency)
1801 DependentLatency = 0;
1803 DependentLatency -= (NextCycle - CurrCycle);
1805 if (!HazardRec->isEnabled()) {
1806 // Bypass HazardRec virtual calls.
1807 CurrCycle = NextCycle;
1810 // Bypass getHazardType calls in case of long latency.
1811 for (; CurrCycle != NextCycle; ++CurrCycle) {
1813 HazardRec->AdvanceCycle();
1815 HazardRec->RecedeCycle();
1818 CheckPending = true;
1819 unsigned LFactor = SchedModel->getLatencyFactor();
1821 (int)(getCriticalCount() - (getScheduledLatency() * LFactor))
1824 DEBUG(dbgs() << "Cycle: " << CurrCycle << ' ' << Available.getName() << '\n');
1827 void SchedBoundary::incExecutedResources(unsigned PIdx, unsigned Count) {
1828 ExecutedResCounts[PIdx] += Count;
1829 if (ExecutedResCounts[PIdx] > MaxExecutedResCount)
1830 MaxExecutedResCount = ExecutedResCounts[PIdx];
1833 /// Add the given processor resource to this scheduled zone.
1835 /// \param Cycles indicates the number of consecutive (non-pipelined) cycles
1836 /// during which this resource is consumed.
1838 /// \return the next cycle at which the instruction may execute without
1839 /// oversubscribing resources.
1840 unsigned SchedBoundary::
1841 countResource(unsigned PIdx, unsigned Cycles, unsigned NextCycle) {
1842 unsigned Factor = SchedModel->getResourceFactor(PIdx);
1843 unsigned Count = Factor * Cycles;
1844 DEBUG(dbgs() << " " << SchedModel->getResourceName(PIdx)
1845 << " +" << Cycles << "x" << Factor << "u\n");
1847 // Update Executed resources counts.
1848 incExecutedResources(PIdx, Count);
1849 assert(Rem->RemainingCounts[PIdx] >= Count && "resource double counted");
1850 Rem->RemainingCounts[PIdx] -= Count;
1852 // Check if this resource exceeds the current critical resource. If so, it
1853 // becomes the critical resource.
1854 if (ZoneCritResIdx != PIdx && (getResourceCount(PIdx) > getCriticalCount())) {
1855 ZoneCritResIdx = PIdx;
1856 DEBUG(dbgs() << " *** Critical resource "
1857 << SchedModel->getResourceName(PIdx) << ": "
1858 << getResourceCount(PIdx) / SchedModel->getLatencyFactor() << "c\n");
1860 // For reserved resources, record the highest cycle using the resource.
1861 unsigned NextAvailable = getNextResourceCycle(PIdx, Cycles);
1862 if (NextAvailable > CurrCycle) {
1863 DEBUG(dbgs() << " Resource conflict: "
1864 << SchedModel->getProcResource(PIdx)->Name << " reserved until @"
1865 << NextAvailable << "\n");
1867 return NextAvailable;
1870 /// Move the boundary of scheduled code by one SUnit.
1871 void SchedBoundary::bumpNode(SUnit *SU) {
1872 // Update the reservation table.
1873 if (HazardRec->isEnabled()) {
1874 if (!isTop() && SU->isCall) {
1875 // Calls are scheduled with their preceding instructions. For bottom-up
1876 // scheduling, clear the pipeline state before emitting.
1879 HazardRec->EmitInstruction(SU);
1881 // checkHazard should prevent scheduling multiple instructions per cycle that
1882 // exceed the issue width.
1883 const MCSchedClassDesc *SC = DAG->getSchedClass(SU);
1884 unsigned IncMOps = SchedModel->getNumMicroOps(SU->getInstr());
1886 (CurrMOps == 0 || (CurrMOps + IncMOps) <= SchedModel->getIssueWidth()) &&
1887 "Cannot schedule this instruction's MicroOps in the current cycle.");
1889 unsigned ReadyCycle = (isTop() ? SU->TopReadyCycle : SU->BotReadyCycle);
1890 DEBUG(dbgs() << " Ready @" << ReadyCycle << "c\n");
1892 unsigned NextCycle = CurrCycle;
1893 switch (SchedModel->getMicroOpBufferSize()) {
1895 assert(ReadyCycle <= CurrCycle && "Broken PendingQueue");
1898 if (ReadyCycle > NextCycle) {
1899 NextCycle = ReadyCycle;
1900 DEBUG(dbgs() << " *** Stall until: " << ReadyCycle << "\n");
1904 // We don't currently model the OOO reorder buffer, so consider all
1905 // scheduled MOps to be "retired". We do loosely model in-order resource
1906 // latency. If this instruction uses an in-order resource, account for any
1907 // likely stall cycles.
1908 if (SU->isUnbuffered && ReadyCycle > NextCycle)
1909 NextCycle = ReadyCycle;
1912 RetiredMOps += IncMOps;
1914 // Update resource counts and critical resource.
1915 if (SchedModel->hasInstrSchedModel()) {
1916 unsigned DecRemIssue = IncMOps * SchedModel->getMicroOpFactor();
1917 assert(Rem->RemIssueCount >= DecRemIssue && "MOps double counted");
1918 Rem->RemIssueCount -= DecRemIssue;
1919 if (ZoneCritResIdx) {
1920 // Scale scheduled micro-ops for comparing with the critical resource.
1921 unsigned ScaledMOps =
1922 RetiredMOps * SchedModel->getMicroOpFactor();
1924 // If scaled micro-ops are now more than the previous critical resource by
1925 // a full cycle, then micro-ops issue becomes critical.
1926 if ((int)(ScaledMOps - getResourceCount(ZoneCritResIdx))
1927 >= (int)SchedModel->getLatencyFactor()) {
1929 DEBUG(dbgs() << " *** Critical resource NumMicroOps: "
1930 << ScaledMOps / SchedModel->getLatencyFactor() << "c\n");
1933 for (TargetSchedModel::ProcResIter
1934 PI = SchedModel->getWriteProcResBegin(SC),
1935 PE = SchedModel->getWriteProcResEnd(SC); PI != PE; ++PI) {
1937 countResource(PI->ProcResourceIdx, PI->Cycles, NextCycle);
1938 if (RCycle > NextCycle)
1941 if (SU->hasReservedResource) {
1942 // For reserved resources, record the highest cycle using the resource.
1943 // For top-down scheduling, this is the cycle in which we schedule this
1944 // instruction plus the number of cycles the operations reserves the
1945 // resource. For bottom-up is it simply the instruction's cycle.
1946 for (TargetSchedModel::ProcResIter
1947 PI = SchedModel->getWriteProcResBegin(SC),
1948 PE = SchedModel->getWriteProcResEnd(SC); PI != PE; ++PI) {
1949 unsigned PIdx = PI->ProcResourceIdx;
1950 if (SchedModel->getProcResource(PIdx)->BufferSize == 0) {
1951 ReservedCycles[PIdx] = isTop() ? NextCycle + PI->Cycles : NextCycle;
1953 MaxObservedLatency = std::max(PI->Cycles, MaxObservedLatency);
1959 // Update ExpectedLatency and DependentLatency.
1960 unsigned &TopLatency = isTop() ? ExpectedLatency : DependentLatency;
1961 unsigned &BotLatency = isTop() ? DependentLatency : ExpectedLatency;
1962 if (SU->getDepth() > TopLatency) {
1963 TopLatency = SU->getDepth();
1964 DEBUG(dbgs() << " " << Available.getName()
1965 << " TopLatency SU(" << SU->NodeNum << ") " << TopLatency << "c\n");
1967 if (SU->getHeight() > BotLatency) {
1968 BotLatency = SU->getHeight();
1969 DEBUG(dbgs() << " " << Available.getName()
1970 << " BotLatency SU(" << SU->NodeNum << ") " << BotLatency << "c\n");
1972 // If we stall for any reason, bump the cycle.
1973 if (NextCycle > CurrCycle) {
1974 bumpCycle(NextCycle);
1977 // After updating ZoneCritResIdx and ExpectedLatency, check if we're
1978 // resource limited. If a stall occurred, bumpCycle does this.
1979 unsigned LFactor = SchedModel->getLatencyFactor();
1981 (int)(getCriticalCount() - (getScheduledLatency() * LFactor))
1984 // Update CurrMOps after calling bumpCycle to handle stalls, since bumpCycle
1985 // resets CurrMOps. Loop to handle instructions with more MOps than issue in
1986 // one cycle. Since we commonly reach the max MOps here, opportunistically
1987 // bump the cycle to avoid uselessly checking everything in the readyQ.
1988 CurrMOps += IncMOps;
1989 while (CurrMOps >= SchedModel->getIssueWidth()) {
1990 DEBUG(dbgs() << " *** Max MOps " << CurrMOps
1991 << " at cycle " << CurrCycle << '\n');
1992 bumpCycle(++NextCycle);
1994 DEBUG(dumpScheduledState());
1997 /// Release pending ready nodes in to the available queue. This makes them
1998 /// visible to heuristics.
1999 void SchedBoundary::releasePending() {
2000 // If the available queue is empty, it is safe to reset MinReadyCycle.
2001 if (Available.empty())
2002 MinReadyCycle = UINT_MAX;
2004 // Check to see if any of the pending instructions are ready to issue. If
2005 // so, add them to the available queue.
2006 bool IsBuffered = SchedModel->getMicroOpBufferSize() != 0;
2007 for (unsigned i = 0, e = Pending.size(); i != e; ++i) {
2008 SUnit *SU = *(Pending.begin()+i);
2009 unsigned ReadyCycle = isTop() ? SU->TopReadyCycle : SU->BotReadyCycle;
2011 if (ReadyCycle < MinReadyCycle)
2012 MinReadyCycle = ReadyCycle;
2014 if (!IsBuffered && ReadyCycle > CurrCycle)
2017 if (checkHazard(SU))
2021 Pending.remove(Pending.begin()+i);
2024 DEBUG(if (!Pending.empty()) Pending.dump());
2025 CheckPending = false;
2028 /// Remove SU from the ready set for this boundary.
2029 void SchedBoundary::removeReady(SUnit *SU) {
2030 if (Available.isInQueue(SU))
2031 Available.remove(Available.find(SU));
2033 assert(Pending.isInQueue(SU) && "bad ready count");
2034 Pending.remove(Pending.find(SU));
2038 /// If this queue only has one ready candidate, return it. As a side effect,
2039 /// defer any nodes that now hit a hazard, and advance the cycle until at least
2040 /// one node is ready. If multiple instructions are ready, return NULL.
2041 SUnit *SchedBoundary::pickOnlyChoice() {
2046 // Defer any ready instrs that now have a hazard.
2047 for (ReadyQueue::iterator I = Available.begin(); I != Available.end();) {
2048 if (checkHazard(*I)) {
2050 I = Available.remove(I);
2056 for (unsigned i = 0; Available.empty(); ++i) {
2057 assert(i <= (HazardRec->getMaxLookAhead() + MaxObservedLatency) &&
2058 "permanent hazard"); (void)i;
2059 bumpCycle(CurrCycle + 1);
2062 if (Available.size() == 1)
2063 return *Available.begin();
2068 // This is useful information to dump after bumpNode.
2069 // Note that the Queue contents are more useful before pickNodeFromQueue.
2070 void SchedBoundary::dumpScheduledState() {
2073 if (ZoneCritResIdx) {
2074 ResFactor = SchedModel->getResourceFactor(ZoneCritResIdx);
2075 ResCount = getResourceCount(ZoneCritResIdx);
2078 ResFactor = SchedModel->getMicroOpFactor();
2079 ResCount = RetiredMOps * SchedModel->getMicroOpFactor();
2081 unsigned LFactor = SchedModel->getLatencyFactor();
2082 dbgs() << Available.getName() << " @" << CurrCycle << "c\n"
2083 << " Retired: " << RetiredMOps;
2084 dbgs() << "\n Executed: " << getExecutedCount() / LFactor << "c";
2085 dbgs() << "\n Critical: " << ResCount / LFactor << "c, "
2086 << ResCount / ResFactor << " "
2087 << SchedModel->getResourceName(ZoneCritResIdx)
2088 << "\n ExpectedLatency: " << ExpectedLatency << "c\n"
2089 << (IsResourceLimited ? " - Resource" : " - Latency")
2094 //===----------------------------------------------------------------------===//
2095 // GenericScheduler - Generic implementation of MachineSchedStrategy.
2096 //===----------------------------------------------------------------------===//
2099 /// Base class for GenericScheduler. This class maintains information about
2100 /// scheduling candidates based on TargetSchedModel making it easy to implement
2101 /// heuristics for either preRA or postRA scheduling.
2102 class GenericSchedulerBase : public MachineSchedStrategy {
2104 /// Represent the type of SchedCandidate found within a single queue.
2105 /// pickNodeBidirectional depends on these listed by decreasing priority.
2107 NoCand, PhysRegCopy, RegExcess, RegCritical, Stall, Cluster, Weak, RegMax,
2108 ResourceReduce, ResourceDemand, BotHeightReduce, BotPathReduce,
2109 TopDepthReduce, TopPathReduce, NextDefUse, NodeOrder};
2112 static const char *getReasonStr(GenericSchedulerBase::CandReason Reason);
2115 /// Policy for scheduling the next instruction in the candidate's zone.
2118 unsigned ReduceResIdx;
2119 unsigned DemandResIdx;
2121 CandPolicy(): ReduceLatency(false), ReduceResIdx(0), DemandResIdx(0) {}
2124 /// Status of an instruction's critical resource consumption.
2125 struct SchedResourceDelta {
2126 // Count critical resources in the scheduled region required by SU.
2127 unsigned CritResources;
2129 // Count critical resources from another region consumed by SU.
2130 unsigned DemandedResources;
2132 SchedResourceDelta(): CritResources(0), DemandedResources(0) {}
2134 bool operator==(const SchedResourceDelta &RHS) const {
2135 return CritResources == RHS.CritResources
2136 && DemandedResources == RHS.DemandedResources;
2138 bool operator!=(const SchedResourceDelta &RHS) const {
2139 return !operator==(RHS);
2143 /// Store the state used by GenericScheduler heuristics, required for the
2144 /// lifetime of one invocation of pickNode().
2145 struct SchedCandidate {
2148 // The best SUnit candidate.
2151 // The reason for this candidate.
2154 // Set of reasons that apply to multiple candidates.
2155 uint32_t RepeatReasonSet;
2157 // Register pressure values for the best candidate.
2158 RegPressureDelta RPDelta;
2160 // Critical resource consumption of the best candidate.
2161 SchedResourceDelta ResDelta;
2163 SchedCandidate(const CandPolicy &policy)
2164 : Policy(policy), SU(NULL), Reason(NoCand), RepeatReasonSet(0) {}
2166 bool isValid() const { return SU; }
2168 // Copy the status of another candidate without changing policy.
2169 void setBest(SchedCandidate &Best) {
2170 assert(Best.Reason != NoCand && "uninitialized Sched candidate");
2172 Reason = Best.Reason;
2173 RPDelta = Best.RPDelta;
2174 ResDelta = Best.ResDelta;
2177 bool isRepeat(CandReason R) { return RepeatReasonSet & (1 << R); }
2178 void setRepeat(CandReason R) { RepeatReasonSet |= (1 << R); }
2180 void initResourceDelta(const ScheduleDAGMI *DAG,
2181 const TargetSchedModel *SchedModel);
2185 const MachineSchedContext *Context;
2186 const TargetSchedModel *SchedModel;
2187 const TargetRegisterInfo *TRI;
2191 GenericSchedulerBase(const MachineSchedContext *C):
2192 Context(C), SchedModel(0), TRI(0) {}
2194 void setPolicy(CandPolicy &Policy, bool IsPostRA, SchedBoundary &CurrZone,
2195 SchedBoundary *OtherZone);
2198 void traceCandidate(const SchedCandidate &Cand);
2203 void GenericSchedulerBase::SchedCandidate::
2204 initResourceDelta(const ScheduleDAGMI *DAG,
2205 const TargetSchedModel *SchedModel) {
2206 if (!Policy.ReduceResIdx && !Policy.DemandResIdx)
2209 const MCSchedClassDesc *SC = DAG->getSchedClass(SU);
2210 for (TargetSchedModel::ProcResIter
2211 PI = SchedModel->getWriteProcResBegin(SC),
2212 PE = SchedModel->getWriteProcResEnd(SC); PI != PE; ++PI) {
2213 if (PI->ProcResourceIdx == Policy.ReduceResIdx)
2214 ResDelta.CritResources += PI->Cycles;
2215 if (PI->ProcResourceIdx == Policy.DemandResIdx)
2216 ResDelta.DemandedResources += PI->Cycles;
2220 /// Set the CandPolicy given a scheduling zone given the current resources and
2221 /// latencies inside and outside the zone.
2222 void GenericSchedulerBase::setPolicy(CandPolicy &Policy,
2224 SchedBoundary &CurrZone,
2225 SchedBoundary *OtherZone) {
2226 // Apply preemptive heuristics based on the the total latency and resources
2227 // inside and outside this zone. Potential stalls should be considered before
2228 // following this policy.
2230 // Compute remaining latency. We need this both to determine whether the
2231 // overall schedule has become latency-limited and whether the instructions
2232 // outside this zone are resource or latency limited.
2234 // The "dependent" latency is updated incrementally during scheduling as the
2235 // max height/depth of scheduled nodes minus the cycles since it was
2237 // DLat = max (N.depth - (CurrCycle - N.ReadyCycle) for N in Zone
2239 // The "independent" latency is the max ready queue depth:
2240 // ILat = max N.depth for N in Available|Pending
2242 // RemainingLatency is the greater of independent and dependent latency.
2243 unsigned RemLatency = CurrZone.getDependentLatency();
2244 RemLatency = std::max(RemLatency,
2245 CurrZone.findMaxLatency(CurrZone.Available.elements()));
2246 RemLatency = std::max(RemLatency,
2247 CurrZone.findMaxLatency(CurrZone.Pending.elements()));
2249 // Compute the critical resource outside the zone.
2250 unsigned OtherCritIdx = 0;
2251 unsigned OtherCount =
2252 OtherZone ? OtherZone->getOtherResourceCount(OtherCritIdx) : 0;
2254 bool OtherResLimited = false;
2255 if (SchedModel->hasInstrSchedModel()) {
2256 unsigned LFactor = SchedModel->getLatencyFactor();
2257 OtherResLimited = (int)(OtherCount - (RemLatency * LFactor)) > (int)LFactor;
2259 // Schedule aggressively for latency in PostRA mode. We don't check for
2260 // acyclic latency during PostRA, and highly out-of-order processors will
2261 // skip PostRA scheduling.
2262 if (!OtherResLimited) {
2263 if (IsPostRA || (RemLatency + CurrZone.getCurrCycle() > Rem.CriticalPath)) {
2264 Policy.ReduceLatency |= true;
2265 DEBUG(dbgs() << " " << CurrZone.Available.getName()
2266 << " RemainingLatency " << RemLatency << " + "
2267 << CurrZone.getCurrCycle() << "c > CritPath "
2268 << Rem.CriticalPath << "\n");
2271 // If the same resource is limiting inside and outside the zone, do nothing.
2272 if (CurrZone.getZoneCritResIdx() == OtherCritIdx)
2276 if (CurrZone.isResourceLimited()) {
2277 dbgs() << " " << CurrZone.Available.getName() << " ResourceLimited: "
2278 << SchedModel->getResourceName(CurrZone.getZoneCritResIdx())
2281 if (OtherResLimited)
2282 dbgs() << " RemainingLimit: "
2283 << SchedModel->getResourceName(OtherCritIdx) << "\n";
2284 if (!CurrZone.isResourceLimited() && !OtherResLimited)
2285 dbgs() << " Latency limited both directions.\n");
2287 if (CurrZone.isResourceLimited() && !Policy.ReduceResIdx)
2288 Policy.ReduceResIdx = CurrZone.getZoneCritResIdx();
2290 if (OtherResLimited)
2291 Policy.DemandResIdx = OtherCritIdx;
2295 const char *GenericSchedulerBase::getReasonStr(
2296 GenericSchedulerBase::CandReason Reason) {
2298 case NoCand: return "NOCAND ";
2299 case PhysRegCopy: return "PREG-COPY";
2300 case RegExcess: return "REG-EXCESS";
2301 case RegCritical: return "REG-CRIT ";
2302 case Stall: return "STALL ";
2303 case Cluster: return "CLUSTER ";
2304 case Weak: return "WEAK ";
2305 case RegMax: return "REG-MAX ";
2306 case ResourceReduce: return "RES-REDUCE";
2307 case ResourceDemand: return "RES-DEMAND";
2308 case TopDepthReduce: return "TOP-DEPTH ";
2309 case TopPathReduce: return "TOP-PATH ";
2310 case BotHeightReduce:return "BOT-HEIGHT";
2311 case BotPathReduce: return "BOT-PATH ";
2312 case NextDefUse: return "DEF-USE ";
2313 case NodeOrder: return "ORDER ";
2315 llvm_unreachable("Unknown reason!");
2318 void GenericSchedulerBase::traceCandidate(const SchedCandidate &Cand) {
2320 unsigned ResIdx = 0;
2321 unsigned Latency = 0;
2322 switch (Cand.Reason) {
2326 P = Cand.RPDelta.Excess;
2329 P = Cand.RPDelta.CriticalMax;
2332 P = Cand.RPDelta.CurrentMax;
2334 case ResourceReduce:
2335 ResIdx = Cand.Policy.ReduceResIdx;
2337 case ResourceDemand:
2338 ResIdx = Cand.Policy.DemandResIdx;
2340 case TopDepthReduce:
2341 Latency = Cand.SU->getDepth();
2344 Latency = Cand.SU->getHeight();
2346 case BotHeightReduce:
2347 Latency = Cand.SU->getHeight();
2350 Latency = Cand.SU->getDepth();
2353 dbgs() << " SU(" << Cand.SU->NodeNum << ") " << getReasonStr(Cand.Reason);
2355 dbgs() << " " << TRI->getRegPressureSetName(P.getPSet())
2356 << ":" << P.getUnitInc() << " ";
2360 dbgs() << " " << SchedModel->getProcResource(ResIdx)->Name << " ";
2364 dbgs() << " " << Latency << " cycles ";
2371 /// Return true if this heuristic determines order.
2372 static bool tryLess(int TryVal, int CandVal,
2373 GenericSchedulerBase::SchedCandidate &TryCand,
2374 GenericSchedulerBase::SchedCandidate &Cand,
2375 GenericSchedulerBase::CandReason Reason) {
2376 if (TryVal < CandVal) {
2377 TryCand.Reason = Reason;
2380 if (TryVal > CandVal) {
2381 if (Cand.Reason > Reason)
2382 Cand.Reason = Reason;
2385 Cand.setRepeat(Reason);
2389 static bool tryGreater(int TryVal, int CandVal,
2390 GenericSchedulerBase::SchedCandidate &TryCand,
2391 GenericSchedulerBase::SchedCandidate &Cand,
2392 GenericSchedulerBase::CandReason Reason) {
2393 if (TryVal > CandVal) {
2394 TryCand.Reason = Reason;
2397 if (TryVal < CandVal) {
2398 if (Cand.Reason > Reason)
2399 Cand.Reason = Reason;
2402 Cand.setRepeat(Reason);
2406 static bool tryLatency(GenericSchedulerBase::SchedCandidate &TryCand,
2407 GenericSchedulerBase::SchedCandidate &Cand,
2408 SchedBoundary &Zone) {
2410 if (Cand.SU->getDepth() > Zone.getScheduledLatency()) {
2411 if (tryLess(TryCand.SU->getDepth(), Cand.SU->getDepth(),
2412 TryCand, Cand, GenericSchedulerBase::TopDepthReduce))
2415 if (tryGreater(TryCand.SU->getHeight(), Cand.SU->getHeight(),
2416 TryCand, Cand, GenericSchedulerBase::TopPathReduce))
2420 if (Cand.SU->getHeight() > Zone.getScheduledLatency()) {
2421 if (tryLess(TryCand.SU->getHeight(), Cand.SU->getHeight(),
2422 TryCand, Cand, GenericSchedulerBase::BotHeightReduce))
2425 if (tryGreater(TryCand.SU->getDepth(), Cand.SU->getDepth(),
2426 TryCand, Cand, GenericSchedulerBase::BotPathReduce))
2432 static void tracePick(const GenericSchedulerBase::SchedCandidate &Cand,
2434 DEBUG(dbgs() << "Pick " << (IsTop ? "Top " : "Bot ")
2435 << GenericSchedulerBase::getReasonStr(Cand.Reason) << '\n');
2439 /// GenericScheduler shrinks the unscheduled zone using heuristics to balance
2441 class GenericScheduler : public GenericSchedulerBase {
2442 ScheduleDAGMILive *DAG;
2444 // State of the top and bottom scheduled instruction boundaries.
2448 MachineSchedPolicy RegionPolicy;
2450 GenericScheduler(const MachineSchedContext *C):
2451 GenericSchedulerBase(C), DAG(0), Top(SchedBoundary::TopQID, "TopQ"),
2452 Bot(SchedBoundary::BotQID, "BotQ") {}
2454 virtual void initPolicy(MachineBasicBlock::iterator Begin,
2455 MachineBasicBlock::iterator End,
2456 unsigned NumRegionInstrs) override;
2458 virtual bool shouldTrackPressure() const override {
2459 return RegionPolicy.ShouldTrackPressure;
2462 virtual void initialize(ScheduleDAGMI *dag) override;
2464 virtual SUnit *pickNode(bool &IsTopNode) override;
2466 virtual void schedNode(SUnit *SU, bool IsTopNode) override;
2468 virtual void releaseTopNode(SUnit *SU) override {
2469 Top.releaseTopNode(SU);
2472 virtual void releaseBottomNode(SUnit *SU) override {
2473 Bot.releaseBottomNode(SU);
2476 virtual void registerRoots() override;
2479 void checkAcyclicLatency();
2481 void tryCandidate(SchedCandidate &Cand,
2482 SchedCandidate &TryCand,
2483 SchedBoundary &Zone,
2484 const RegPressureTracker &RPTracker,
2485 RegPressureTracker &TempTracker);
2487 SUnit *pickNodeBidirectional(bool &IsTopNode);
2489 void pickNodeFromQueue(SchedBoundary &Zone,
2490 const RegPressureTracker &RPTracker,
2491 SchedCandidate &Candidate);
2493 void reschedulePhysRegCopies(SUnit *SU, bool isTop);
2497 void GenericScheduler::initialize(ScheduleDAGMI *dag) {
2498 assert(dag->hasVRegLiveness() &&
2499 "(PreRA)GenericScheduler needs vreg liveness");
2500 DAG = static_cast<ScheduleDAGMILive*>(dag);
2501 SchedModel = DAG->getSchedModel();
2504 Rem.init(DAG, SchedModel);
2505 Top.init(DAG, SchedModel, &Rem);
2506 Bot.init(DAG, SchedModel, &Rem);
2508 // Initialize resource counts.
2510 // Initialize the HazardRecognizers. If itineraries don't exist, are empty, or
2511 // are disabled, then these HazardRecs will be disabled.
2512 const InstrItineraryData *Itin = SchedModel->getInstrItineraries();
2513 const TargetMachine &TM = DAG->MF.getTarget();
2514 if (!Top.HazardRec) {
2516 TM.getInstrInfo()->CreateTargetMIHazardRecognizer(Itin, DAG);
2518 if (!Bot.HazardRec) {
2520 TM.getInstrInfo()->CreateTargetMIHazardRecognizer(Itin, DAG);
2524 /// Initialize the per-region scheduling policy.
2525 void GenericScheduler::initPolicy(MachineBasicBlock::iterator Begin,
2526 MachineBasicBlock::iterator End,
2527 unsigned NumRegionInstrs) {
2528 const TargetMachine &TM = Context->MF->getTarget();
2529 const TargetLowering *TLI = TM.getTargetLowering();
2531 // Avoid setting up the register pressure tracker for small regions to save
2532 // compile time. As a rough heuristic, only track pressure when the number of
2533 // schedulable instructions exceeds half the integer register file.
2534 RegionPolicy.ShouldTrackPressure = true;
2535 for (unsigned VT = MVT::i32; VT > (unsigned)MVT::i1; --VT) {
2536 MVT::SimpleValueType LegalIntVT = (MVT::SimpleValueType)VT;
2537 if (TLI->isTypeLegal(LegalIntVT)) {
2538 unsigned NIntRegs = Context->RegClassInfo->getNumAllocatableRegs(
2539 TLI->getRegClassFor(LegalIntVT));
2540 RegionPolicy.ShouldTrackPressure = NumRegionInstrs > (NIntRegs / 2);
2544 // For generic targets, we default to bottom-up, because it's simpler and more
2545 // compile-time optimizations have been implemented in that direction.
2546 RegionPolicy.OnlyBottomUp = true;
2548 // Allow the subtarget to override default policy.
2549 const TargetSubtargetInfo &ST = TM.getSubtarget<TargetSubtargetInfo>();
2550 ST.overrideSchedPolicy(RegionPolicy, Begin, End, NumRegionInstrs);
2552 // After subtarget overrides, apply command line options.
2553 if (!EnableRegPressure)
2554 RegionPolicy.ShouldTrackPressure = false;
2556 // Check -misched-topdown/bottomup can force or unforce scheduling direction.
2557 // e.g. -misched-bottomup=false allows scheduling in both directions.
2558 assert((!ForceTopDown || !ForceBottomUp) &&
2559 "-misched-topdown incompatible with -misched-bottomup");
2560 if (ForceBottomUp.getNumOccurrences() > 0) {
2561 RegionPolicy.OnlyBottomUp = ForceBottomUp;
2562 if (RegionPolicy.OnlyBottomUp)
2563 RegionPolicy.OnlyTopDown = false;
2565 if (ForceTopDown.getNumOccurrences() > 0) {
2566 RegionPolicy.OnlyTopDown = ForceTopDown;
2567 if (RegionPolicy.OnlyTopDown)
2568 RegionPolicy.OnlyBottomUp = false;
2572 /// Set IsAcyclicLatencyLimited if the acyclic path is longer than the cyclic
2573 /// critical path by more cycles than it takes to drain the instruction buffer.
2574 /// We estimate an upper bounds on in-flight instructions as:
2576 /// CyclesPerIteration = max( CyclicPath, Loop-Resource-Height )
2577 /// InFlightIterations = AcyclicPath / CyclesPerIteration
2578 /// InFlightResources = InFlightIterations * LoopResources
2580 /// TODO: Check execution resources in addition to IssueCount.
2581 void GenericScheduler::checkAcyclicLatency() {
2582 if (Rem.CyclicCritPath == 0 || Rem.CyclicCritPath >= Rem.CriticalPath)
2585 // Scaled number of cycles per loop iteration.
2586 unsigned IterCount =
2587 std::max(Rem.CyclicCritPath * SchedModel->getLatencyFactor(),
2589 // Scaled acyclic critical path.
2590 unsigned AcyclicCount = Rem.CriticalPath * SchedModel->getLatencyFactor();
2591 // InFlightCount = (AcyclicPath / IterCycles) * InstrPerLoop
2592 unsigned InFlightCount =
2593 (AcyclicCount * Rem.RemIssueCount + IterCount-1) / IterCount;
2594 unsigned BufferLimit =
2595 SchedModel->getMicroOpBufferSize() * SchedModel->getMicroOpFactor();
2597 Rem.IsAcyclicLatencyLimited = InFlightCount > BufferLimit;
2599 DEBUG(dbgs() << "IssueCycles="
2600 << Rem.RemIssueCount / SchedModel->getLatencyFactor() << "c "
2601 << "IterCycles=" << IterCount / SchedModel->getLatencyFactor()
2602 << "c NumIters=" << (AcyclicCount + IterCount-1) / IterCount
2603 << " InFlight=" << InFlightCount / SchedModel->getMicroOpFactor()
2604 << "m BufferLim=" << SchedModel->getMicroOpBufferSize() << "m\n";
2605 if (Rem.IsAcyclicLatencyLimited)
2606 dbgs() << " ACYCLIC LATENCY LIMIT\n");
2609 void GenericScheduler::registerRoots() {
2610 Rem.CriticalPath = DAG->ExitSU.getDepth();
2612 // Some roots may not feed into ExitSU. Check all of them in case.
2613 for (std::vector<SUnit*>::const_iterator
2614 I = Bot.Available.begin(), E = Bot.Available.end(); I != E; ++I) {
2615 if ((*I)->getDepth() > Rem.CriticalPath)
2616 Rem.CriticalPath = (*I)->getDepth();
2618 DEBUG(dbgs() << "Critical Path: " << Rem.CriticalPath << '\n');
2620 if (EnableCyclicPath) {
2621 Rem.CyclicCritPath = DAG->computeCyclicCriticalPath();
2622 checkAcyclicLatency();
2626 static bool tryPressure(const PressureChange &TryP,
2627 const PressureChange &CandP,
2628 GenericSchedulerBase::SchedCandidate &TryCand,
2629 GenericSchedulerBase::SchedCandidate &Cand,
2630 GenericSchedulerBase::CandReason Reason) {
2631 int TryRank = TryP.getPSetOrMax();
2632 int CandRank = CandP.getPSetOrMax();
2633 // If both candidates affect the same set, go with the smallest increase.
2634 if (TryRank == CandRank) {
2635 return tryLess(TryP.getUnitInc(), CandP.getUnitInc(), TryCand, Cand,
2638 // If one candidate decreases and the other increases, go with it.
2639 // Invalid candidates have UnitInc==0.
2640 if (tryLess(TryP.getUnitInc() < 0, CandP.getUnitInc() < 0, TryCand, Cand,
2644 // If the candidates are decreasing pressure, reverse priority.
2645 if (TryP.getUnitInc() < 0)
2646 std::swap(TryRank, CandRank);
2647 return tryGreater(TryRank, CandRank, TryCand, Cand, Reason);
2650 static unsigned getWeakLeft(const SUnit *SU, bool isTop) {
2651 return (isTop) ? SU->WeakPredsLeft : SU->WeakSuccsLeft;
2654 /// Minimize physical register live ranges. Regalloc wants them adjacent to
2655 /// their physreg def/use.
2657 /// FIXME: This is an unnecessary check on the critical path. Most are root/leaf
2658 /// copies which can be prescheduled. The rest (e.g. x86 MUL) could be bundled
2659 /// with the operation that produces or consumes the physreg. We'll do this when
2660 /// regalloc has support for parallel copies.
2661 static int biasPhysRegCopy(const SUnit *SU, bool isTop) {
2662 const MachineInstr *MI = SU->getInstr();
2666 unsigned ScheduledOper = isTop ? 1 : 0;
2667 unsigned UnscheduledOper = isTop ? 0 : 1;
2668 // If we have already scheduled the physreg produce/consumer, immediately
2669 // schedule the copy.
2670 if (TargetRegisterInfo::isPhysicalRegister(
2671 MI->getOperand(ScheduledOper).getReg()))
2673 // If the physreg is at the boundary, defer it. Otherwise schedule it
2674 // immediately to free the dependent. We can hoist the copy later.
2675 bool AtBoundary = isTop ? !SU->NumSuccsLeft : !SU->NumPredsLeft;
2676 if (TargetRegisterInfo::isPhysicalRegister(
2677 MI->getOperand(UnscheduledOper).getReg()))
2678 return AtBoundary ? -1 : 1;
2682 /// Apply a set of heursitics to a new candidate. Heuristics are currently
2683 /// hierarchical. This may be more efficient than a graduated cost model because
2684 /// we don't need to evaluate all aspects of the model for each node in the
2685 /// queue. But it's really done to make the heuristics easier to debug and
2686 /// statistically analyze.
2688 /// \param Cand provides the policy and current best candidate.
2689 /// \param TryCand refers to the next SUnit candidate, otherwise uninitialized.
2690 /// \param Zone describes the scheduled zone that we are extending.
2691 /// \param RPTracker describes reg pressure within the scheduled zone.
2692 /// \param TempTracker is a scratch pressure tracker to reuse in queries.
2693 void GenericScheduler::tryCandidate(SchedCandidate &Cand,
2694 SchedCandidate &TryCand,
2695 SchedBoundary &Zone,
2696 const RegPressureTracker &RPTracker,
2697 RegPressureTracker &TempTracker) {
2699 if (DAG->isTrackingPressure()) {
2700 // Always initialize TryCand's RPDelta.
2702 TempTracker.getMaxDownwardPressureDelta(
2703 TryCand.SU->getInstr(),
2705 DAG->getRegionCriticalPSets(),
2706 DAG->getRegPressure().MaxSetPressure);
2709 if (VerifyScheduling) {
2710 TempTracker.getMaxUpwardPressureDelta(
2711 TryCand.SU->getInstr(),
2712 &DAG->getPressureDiff(TryCand.SU),
2714 DAG->getRegionCriticalPSets(),
2715 DAG->getRegPressure().MaxSetPressure);
2718 RPTracker.getUpwardPressureDelta(
2719 TryCand.SU->getInstr(),
2720 DAG->getPressureDiff(TryCand.SU),
2722 DAG->getRegionCriticalPSets(),
2723 DAG->getRegPressure().MaxSetPressure);
2727 DEBUG(if (TryCand.RPDelta.Excess.isValid())
2728 dbgs() << " SU(" << TryCand.SU->NodeNum << ") "
2729 << TRI->getRegPressureSetName(TryCand.RPDelta.Excess.getPSet())
2730 << ":" << TryCand.RPDelta.Excess.getUnitInc() << "\n");
2732 // Initialize the candidate if needed.
2733 if (!Cand.isValid()) {
2734 TryCand.Reason = NodeOrder;
2738 if (tryGreater(biasPhysRegCopy(TryCand.SU, Zone.isTop()),
2739 biasPhysRegCopy(Cand.SU, Zone.isTop()),
2740 TryCand, Cand, PhysRegCopy))
2743 // Avoid exceeding the target's limit. If signed PSetID is negative, it is
2744 // invalid; convert it to INT_MAX to give it lowest priority.
2745 if (DAG->isTrackingPressure() && tryPressure(TryCand.RPDelta.Excess,
2746 Cand.RPDelta.Excess,
2747 TryCand, Cand, RegExcess))
2750 // Avoid increasing the max critical pressure in the scheduled region.
2751 if (DAG->isTrackingPressure() && tryPressure(TryCand.RPDelta.CriticalMax,
2752 Cand.RPDelta.CriticalMax,
2753 TryCand, Cand, RegCritical))
2756 // For loops that are acyclic path limited, aggressively schedule for latency.
2757 // This can result in very long dependence chains scheduled in sequence, so
2758 // once every cycle (when CurrMOps == 0), switch to normal heuristics.
2759 if (Rem.IsAcyclicLatencyLimited && !Zone.getCurrMOps()
2760 && tryLatency(TryCand, Cand, Zone))
2763 // Prioritize instructions that read unbuffered resources by stall cycles.
2764 if (tryLess(Zone.getLatencyStallCycles(TryCand.SU),
2765 Zone.getLatencyStallCycles(Cand.SU), TryCand, Cand, Stall))
2768 // Keep clustered nodes together to encourage downstream peephole
2769 // optimizations which may reduce resource requirements.
2771 // This is a best effort to set things up for a post-RA pass. Optimizations
2772 // like generating loads of multiple registers should ideally be done within
2773 // the scheduler pass by combining the loads during DAG postprocessing.
2774 const SUnit *NextClusterSU =
2775 Zone.isTop() ? DAG->getNextClusterSucc() : DAG->getNextClusterPred();
2776 if (tryGreater(TryCand.SU == NextClusterSU, Cand.SU == NextClusterSU,
2777 TryCand, Cand, Cluster))
2780 // Weak edges are for clustering and other constraints.
2781 if (tryLess(getWeakLeft(TryCand.SU, Zone.isTop()),
2782 getWeakLeft(Cand.SU, Zone.isTop()),
2783 TryCand, Cand, Weak)) {
2786 // Avoid increasing the max pressure of the entire region.
2787 if (DAG->isTrackingPressure() && tryPressure(TryCand.RPDelta.CurrentMax,
2788 Cand.RPDelta.CurrentMax,
2789 TryCand, Cand, RegMax))
2792 // Avoid critical resource consumption and balance the schedule.
2793 TryCand.initResourceDelta(DAG, SchedModel);
2794 if (tryLess(TryCand.ResDelta.CritResources, Cand.ResDelta.CritResources,
2795 TryCand, Cand, ResourceReduce))
2797 if (tryGreater(TryCand.ResDelta.DemandedResources,
2798 Cand.ResDelta.DemandedResources,
2799 TryCand, Cand, ResourceDemand))
2802 // Avoid serializing long latency dependence chains.
2803 // For acyclic path limited loops, latency was already checked above.
2804 if (Cand.Policy.ReduceLatency && !Rem.IsAcyclicLatencyLimited
2805 && tryLatency(TryCand, Cand, Zone)) {
2809 // Prefer immediate defs/users of the last scheduled instruction. This is a
2810 // local pressure avoidance strategy that also makes the machine code
2812 if (tryGreater(Zone.isNextSU(TryCand.SU), Zone.isNextSU(Cand.SU),
2813 TryCand, Cand, NextDefUse))
2816 // Fall through to original instruction order.
2817 if ((Zone.isTop() && TryCand.SU->NodeNum < Cand.SU->NodeNum)
2818 || (!Zone.isTop() && TryCand.SU->NodeNum > Cand.SU->NodeNum)) {
2819 TryCand.Reason = NodeOrder;
2823 /// Pick the best candidate from the queue.
2825 /// TODO: getMaxPressureDelta results can be mostly cached for each SUnit during
2826 /// DAG building. To adjust for the current scheduling location we need to
2827 /// maintain the number of vreg uses remaining to be top-scheduled.
2828 void GenericScheduler::pickNodeFromQueue(SchedBoundary &Zone,
2829 const RegPressureTracker &RPTracker,
2830 SchedCandidate &Cand) {
2831 ReadyQueue &Q = Zone.Available;
2835 // getMaxPressureDelta temporarily modifies the tracker.
2836 RegPressureTracker &TempTracker = const_cast<RegPressureTracker&>(RPTracker);
2838 for (ReadyQueue::iterator I = Q.begin(), E = Q.end(); I != E; ++I) {
2840 SchedCandidate TryCand(Cand.Policy);
2842 tryCandidate(Cand, TryCand, Zone, RPTracker, TempTracker);
2843 if (TryCand.Reason != NoCand) {
2844 // Initialize resource delta if needed in case future heuristics query it.
2845 if (TryCand.ResDelta == SchedResourceDelta())
2846 TryCand.initResourceDelta(DAG, SchedModel);
2847 Cand.setBest(TryCand);
2848 DEBUG(traceCandidate(Cand));
2853 /// Pick the best candidate node from either the top or bottom queue.
2854 SUnit *GenericScheduler::pickNodeBidirectional(bool &IsTopNode) {
2855 // Schedule as far as possible in the direction of no choice. This is most
2856 // efficient, but also provides the best heuristics for CriticalPSets.
2857 if (SUnit *SU = Bot.pickOnlyChoice()) {
2859 DEBUG(dbgs() << "Pick Bot NOCAND\n");
2862 if (SUnit *SU = Top.pickOnlyChoice()) {
2864 DEBUG(dbgs() << "Pick Top NOCAND\n");
2867 CandPolicy NoPolicy;
2868 SchedCandidate BotCand(NoPolicy);
2869 SchedCandidate TopCand(NoPolicy);
2870 // Set the bottom-up policy based on the state of the current bottom zone and
2871 // the instructions outside the zone, including the top zone.
2872 setPolicy(BotCand.Policy, /*IsPostRA=*/false, Bot, &Top);
2873 // Set the top-down policy based on the state of the current top zone and
2874 // the instructions outside the zone, including the bottom zone.
2875 setPolicy(TopCand.Policy, /*IsPostRA=*/false, Top, &Bot);
2877 // Prefer bottom scheduling when heuristics are silent.
2878 pickNodeFromQueue(Bot, DAG->getBotRPTracker(), BotCand);
2879 assert(BotCand.Reason != NoCand && "failed to find the first candidate");
2881 // If either Q has a single candidate that provides the least increase in
2882 // Excess pressure, we can immediately schedule from that Q.
2884 // RegionCriticalPSets summarizes the pressure within the scheduled region and
2885 // affects picking from either Q. If scheduling in one direction must
2886 // increase pressure for one of the excess PSets, then schedule in that
2887 // direction first to provide more freedom in the other direction.
2888 if ((BotCand.Reason == RegExcess && !BotCand.isRepeat(RegExcess))
2889 || (BotCand.Reason == RegCritical
2890 && !BotCand.isRepeat(RegCritical)))
2893 tracePick(BotCand, IsTopNode);
2896 // Check if the top Q has a better candidate.
2897 pickNodeFromQueue(Top, DAG->getTopRPTracker(), TopCand);
2898 assert(TopCand.Reason != NoCand && "failed to find the first candidate");
2900 // Choose the queue with the most important (lowest enum) reason.
2901 if (TopCand.Reason < BotCand.Reason) {
2903 tracePick(TopCand, IsTopNode);
2906 // Otherwise prefer the bottom candidate, in node order if all else failed.
2908 tracePick(BotCand, IsTopNode);
2912 /// Pick the best node to balance the schedule. Implements MachineSchedStrategy.
2913 SUnit *GenericScheduler::pickNode(bool &IsTopNode) {
2914 if (DAG->top() == DAG->bottom()) {
2915 assert(Top.Available.empty() && Top.Pending.empty() &&
2916 Bot.Available.empty() && Bot.Pending.empty() && "ReadyQ garbage");
2921 if (RegionPolicy.OnlyTopDown) {
2922 SU = Top.pickOnlyChoice();
2924 CandPolicy NoPolicy;
2925 SchedCandidate TopCand(NoPolicy);
2926 pickNodeFromQueue(Top, DAG->getTopRPTracker(), TopCand);
2927 assert(TopCand.Reason != NoCand && "failed to find a candidate");
2928 tracePick(TopCand, true);
2933 else if (RegionPolicy.OnlyBottomUp) {
2934 SU = Bot.pickOnlyChoice();
2936 CandPolicy NoPolicy;
2937 SchedCandidate BotCand(NoPolicy);
2938 pickNodeFromQueue(Bot, DAG->getBotRPTracker(), BotCand);
2939 assert(BotCand.Reason != NoCand && "failed to find a candidate");
2940 tracePick(BotCand, false);
2946 SU = pickNodeBidirectional(IsTopNode);
2948 } while (SU->isScheduled);
2950 if (SU->isTopReady())
2951 Top.removeReady(SU);
2952 if (SU->isBottomReady())
2953 Bot.removeReady(SU);
2955 DEBUG(dbgs() << "Scheduling SU(" << SU->NodeNum << ") " << *SU->getInstr());
2959 void GenericScheduler::reschedulePhysRegCopies(SUnit *SU, bool isTop) {
2961 MachineBasicBlock::iterator InsertPos = SU->getInstr();
2964 SmallVectorImpl<SDep> &Deps = isTop ? SU->Preds : SU->Succs;
2966 // Find already scheduled copies with a single physreg dependence and move
2967 // them just above the scheduled instruction.
2968 for (SmallVectorImpl<SDep>::iterator I = Deps.begin(), E = Deps.end();
2970 if (I->getKind() != SDep::Data || !TRI->isPhysicalRegister(I->getReg()))
2972 SUnit *DepSU = I->getSUnit();
2973 if (isTop ? DepSU->Succs.size() > 1 : DepSU->Preds.size() > 1)
2975 MachineInstr *Copy = DepSU->getInstr();
2976 if (!Copy->isCopy())
2978 DEBUG(dbgs() << " Rescheduling physreg copy ";
2979 I->getSUnit()->dump(DAG));
2980 DAG->moveInstruction(Copy, InsertPos);
2984 /// Update the scheduler's state after scheduling a node. This is the same node
2985 /// that was just returned by pickNode(). However, ScheduleDAGMILive needs to
2986 /// update it's state based on the current cycle before MachineSchedStrategy
2989 /// FIXME: Eventually, we may bundle physreg copies rather than rescheduling
2990 /// them here. See comments in biasPhysRegCopy.
2991 void GenericScheduler::schedNode(SUnit *SU, bool IsTopNode) {
2993 SU->TopReadyCycle = std::max(SU->TopReadyCycle, Top.getCurrCycle());
2995 if (SU->hasPhysRegUses)
2996 reschedulePhysRegCopies(SU, true);
2999 SU->BotReadyCycle = std::max(SU->BotReadyCycle, Bot.getCurrCycle());
3001 if (SU->hasPhysRegDefs)
3002 reschedulePhysRegCopies(SU, false);
3006 /// Create the standard converging machine scheduler. This will be used as the
3007 /// default scheduler if the target does not set a default.
3008 static ScheduleDAGInstrs *createGenericSchedLive(MachineSchedContext *C) {
3009 ScheduleDAGMILive *DAG = new ScheduleDAGMILive(C, new GenericScheduler(C));
3010 // Register DAG post-processors.
3012 // FIXME: extend the mutation API to allow earlier mutations to instantiate
3013 // data and pass it to later mutations. Have a single mutation that gathers
3014 // the interesting nodes in one pass.
3015 DAG->addMutation(new CopyConstrain(DAG->TII, DAG->TRI));
3016 if (EnableLoadCluster && DAG->TII->enableClusterLoads())
3017 DAG->addMutation(new LoadClusterMutation(DAG->TII, DAG->TRI));
3018 if (EnableMacroFusion)
3019 DAG->addMutation(new MacroFusion(DAG->TII));
3023 static MachineSchedRegistry
3024 GenericSchedRegistry("converge", "Standard converging scheduler.",
3025 createGenericSchedLive);
3027 //===----------------------------------------------------------------------===//
3028 // PostGenericScheduler - Generic PostRA implementation of MachineSchedStrategy.
3029 //===----------------------------------------------------------------------===//
3032 /// PostGenericScheduler - Interface to the scheduling algorithm used by
3035 /// Callbacks from ScheduleDAGMI:
3036 /// initPolicy -> initialize(DAG) -> registerRoots -> pickNode ...
3037 class PostGenericScheduler : public GenericSchedulerBase {
3040 SmallVector<SUnit*, 8> BotRoots;
3042 PostGenericScheduler(const MachineSchedContext *C):
3043 GenericSchedulerBase(C), Top(SchedBoundary::TopQID, "TopQ") {}
3045 virtual ~PostGenericScheduler() {}
3047 virtual void initPolicy(MachineBasicBlock::iterator Begin,
3048 MachineBasicBlock::iterator End,
3049 unsigned NumRegionInstrs) override {
3050 /* no configurable policy */
3053 /// PostRA scheduling does not track pressure.
3054 virtual bool shouldTrackPressure() const override { return false; }
3056 virtual void initialize(ScheduleDAGMI *Dag) override {
3058 SchedModel = DAG->getSchedModel();
3061 Rem.init(DAG, SchedModel);
3062 Top.init(DAG, SchedModel, &Rem);
3065 // Initialize the HazardRecognizers. If itineraries don't exist, are empty,
3066 // or are disabled, then these HazardRecs will be disabled.
3067 const InstrItineraryData *Itin = SchedModel->getInstrItineraries();
3068 const TargetMachine &TM = DAG->MF.getTarget();
3069 if (!Top.HazardRec) {
3071 TM.getInstrInfo()->CreateTargetMIHazardRecognizer(Itin, DAG);
3075 virtual void registerRoots() override;
3077 virtual SUnit *pickNode(bool &IsTopNode) override;
3079 virtual void scheduleTree(unsigned SubtreeID) override {
3080 llvm_unreachable("PostRA scheduler does not support subtree analysis.");
3083 virtual void schedNode(SUnit *SU, bool IsTopNode) override;
3085 virtual void releaseTopNode(SUnit *SU) override {
3086 Top.releaseTopNode(SU);
3089 // Only called for roots.
3090 virtual void releaseBottomNode(SUnit *SU) override {
3091 BotRoots.push_back(SU);
3095 void tryCandidate(SchedCandidate &Cand, SchedCandidate &TryCand);
3097 void pickNodeFromQueue(SchedCandidate &Cand);
3101 void PostGenericScheduler::registerRoots() {
3102 Rem.CriticalPath = DAG->ExitSU.getDepth();
3104 // Some roots may not feed into ExitSU. Check all of them in case.
3105 for (SmallVectorImpl<SUnit*>::const_iterator
3106 I = BotRoots.begin(), E = BotRoots.end(); I != E; ++I) {
3107 if ((*I)->getDepth() > Rem.CriticalPath)
3108 Rem.CriticalPath = (*I)->getDepth();
3110 DEBUG(dbgs() << "Critical Path: " << Rem.CriticalPath << '\n');
3113 /// Apply a set of heursitics to a new candidate for PostRA scheduling.
3115 /// \param Cand provides the policy and current best candidate.
3116 /// \param TryCand refers to the next SUnit candidate, otherwise uninitialized.
3117 void PostGenericScheduler::tryCandidate(SchedCandidate &Cand,
3118 SchedCandidate &TryCand) {
3120 // Initialize the candidate if needed.
3121 if (!Cand.isValid()) {
3122 TryCand.Reason = NodeOrder;
3126 // Prioritize instructions that read unbuffered resources by stall cycles.
3127 if (tryLess(Top.getLatencyStallCycles(TryCand.SU),
3128 Top.getLatencyStallCycles(Cand.SU), TryCand, Cand, Stall))
3131 // Avoid critical resource consumption and balance the schedule.
3132 if (tryLess(TryCand.ResDelta.CritResources, Cand.ResDelta.CritResources,
3133 TryCand, Cand, ResourceReduce))
3135 if (tryGreater(TryCand.ResDelta.DemandedResources,
3136 Cand.ResDelta.DemandedResources,
3137 TryCand, Cand, ResourceDemand))
3140 // Avoid serializing long latency dependence chains.
3141 if (Cand.Policy.ReduceLatency && tryLatency(TryCand, Cand, Top)) {
3145 // Fall through to original instruction order.
3146 if (TryCand.SU->NodeNum < Cand.SU->NodeNum)
3147 TryCand.Reason = NodeOrder;
3150 void PostGenericScheduler::pickNodeFromQueue(SchedCandidate &Cand) {
3151 ReadyQueue &Q = Top.Available;
3155 for (ReadyQueue::iterator I = Q.begin(), E = Q.end(); I != E; ++I) {
3156 SchedCandidate TryCand(Cand.Policy);
3158 TryCand.initResourceDelta(DAG, SchedModel);
3159 tryCandidate(Cand, TryCand);
3160 if (TryCand.Reason != NoCand) {
3161 Cand.setBest(TryCand);
3162 DEBUG(traceCandidate(Cand));
3167 /// Pick the next node to schedule.
3168 SUnit *PostGenericScheduler::pickNode(bool &IsTopNode) {
3169 if (DAG->top() == DAG->bottom()) {
3170 assert(Top.Available.empty() && Top.Pending.empty() && "ReadyQ garbage");
3175 SU = Top.pickOnlyChoice();
3177 CandPolicy NoPolicy;
3178 SchedCandidate TopCand(NoPolicy);
3179 // Set the top-down policy based on the state of the current top zone and
3180 // the instructions outside the zone, including the bottom zone.
3181 setPolicy(TopCand.Policy, /*IsPostRA=*/true, Top, NULL);
3182 pickNodeFromQueue(TopCand);
3183 assert(TopCand.Reason != NoCand && "failed to find a candidate");
3184 tracePick(TopCand, true);
3187 } while (SU->isScheduled);
3190 Top.removeReady(SU);
3192 DEBUG(dbgs() << "Scheduling SU(" << SU->NodeNum << ") " << *SU->getInstr());
3196 /// Called after ScheduleDAGMI has scheduled an instruction and updated
3197 /// scheduled/remaining flags in the DAG nodes.
3198 void PostGenericScheduler::schedNode(SUnit *SU, bool IsTopNode) {
3199 SU->TopReadyCycle = std::max(SU->TopReadyCycle, Top.getCurrCycle());
3203 /// Create a generic scheduler with no vreg liveness or DAG mutation passes.
3204 static ScheduleDAGInstrs *createGenericSchedPostRA(MachineSchedContext *C) {
3205 return new ScheduleDAGMI(C, new PostGenericScheduler(C), /*IsPostRA=*/true);
3208 //===----------------------------------------------------------------------===//
3209 // ILP Scheduler. Currently for experimental analysis of heuristics.
3210 //===----------------------------------------------------------------------===//
3213 /// \brief Order nodes by the ILP metric.
3215 const SchedDFSResult *DFSResult;
3216 const BitVector *ScheduledTrees;
3219 ILPOrder(bool MaxILP): DFSResult(0), ScheduledTrees(0), MaximizeILP(MaxILP) {}
3221 /// \brief Apply a less-than relation on node priority.
3223 /// (Return true if A comes after B in the Q.)
3224 bool operator()(const SUnit *A, const SUnit *B) const {
3225 unsigned SchedTreeA = DFSResult->getSubtreeID(A);
3226 unsigned SchedTreeB = DFSResult->getSubtreeID(B);
3227 if (SchedTreeA != SchedTreeB) {
3228 // Unscheduled trees have lower priority.
3229 if (ScheduledTrees->test(SchedTreeA) != ScheduledTrees->test(SchedTreeB))
3230 return ScheduledTrees->test(SchedTreeB);
3232 // Trees with shallower connections have have lower priority.
3233 if (DFSResult->getSubtreeLevel(SchedTreeA)
3234 != DFSResult->getSubtreeLevel(SchedTreeB)) {
3235 return DFSResult->getSubtreeLevel(SchedTreeA)
3236 < DFSResult->getSubtreeLevel(SchedTreeB);
3240 return DFSResult->getILP(A) < DFSResult->getILP(B);
3242 return DFSResult->getILP(A) > DFSResult->getILP(B);
3246 /// \brief Schedule based on the ILP metric.
3247 class ILPScheduler : public MachineSchedStrategy {
3248 ScheduleDAGMILive *DAG;
3251 std::vector<SUnit*> ReadyQ;
3253 ILPScheduler(bool MaximizeILP): DAG(0), Cmp(MaximizeILP) {}
3255 virtual void initialize(ScheduleDAGMI *dag) {
3256 assert(dag->hasVRegLiveness() && "ILPScheduler needs vreg liveness");
3257 DAG = static_cast<ScheduleDAGMILive*>(dag);
3258 DAG->computeDFSResult();
3259 Cmp.DFSResult = DAG->getDFSResult();
3260 Cmp.ScheduledTrees = &DAG->getScheduledTrees();
3264 virtual void registerRoots() {
3265 // Restore the heap in ReadyQ with the updated DFS results.
3266 std::make_heap(ReadyQ.begin(), ReadyQ.end(), Cmp);
3269 /// Implement MachineSchedStrategy interface.
3270 /// -----------------------------------------
3272 /// Callback to select the highest priority node from the ready Q.
3273 virtual SUnit *pickNode(bool &IsTopNode) {
3274 if (ReadyQ.empty()) return NULL;
3275 std::pop_heap(ReadyQ.begin(), ReadyQ.end(), Cmp);
3276 SUnit *SU = ReadyQ.back();
3279 DEBUG(dbgs() << "Pick node " << "SU(" << SU->NodeNum << ") "
3280 << " ILP: " << DAG->getDFSResult()->getILP(SU)
3281 << " Tree: " << DAG->getDFSResult()->getSubtreeID(SU) << " @"
3282 << DAG->getDFSResult()->getSubtreeLevel(
3283 DAG->getDFSResult()->getSubtreeID(SU)) << '\n'
3284 << "Scheduling " << *SU->getInstr());
3288 /// \brief Scheduler callback to notify that a new subtree is scheduled.
3289 virtual void scheduleTree(unsigned SubtreeID) {
3290 std::make_heap(ReadyQ.begin(), ReadyQ.end(), Cmp);
3293 /// Callback after a node is scheduled. Mark a newly scheduled tree, notify
3294 /// DFSResults, and resort the priority Q.
3295 virtual void schedNode(SUnit *SU, bool IsTopNode) {
3296 assert(!IsTopNode && "SchedDFSResult needs bottom-up");
3299 virtual void releaseTopNode(SUnit *) { /*only called for top roots*/ }
3301 virtual void releaseBottomNode(SUnit *SU) {
3302 ReadyQ.push_back(SU);
3303 std::push_heap(ReadyQ.begin(), ReadyQ.end(), Cmp);
3308 static ScheduleDAGInstrs *createILPMaxScheduler(MachineSchedContext *C) {
3309 return new ScheduleDAGMILive(C, new ILPScheduler(true));
3311 static ScheduleDAGInstrs *createILPMinScheduler(MachineSchedContext *C) {
3312 return new ScheduleDAGMILive(C, new ILPScheduler(false));
3314 static MachineSchedRegistry ILPMaxRegistry(
3315 "ilpmax", "Schedule bottom-up for max ILP", createILPMaxScheduler);
3316 static MachineSchedRegistry ILPMinRegistry(
3317 "ilpmin", "Schedule bottom-up for min ILP", createILPMinScheduler);
3319 //===----------------------------------------------------------------------===//
3320 // Machine Instruction Shuffler for Correctness Testing
3321 //===----------------------------------------------------------------------===//
3325 /// Apply a less-than relation on the node order, which corresponds to the
3326 /// instruction order prior to scheduling. IsReverse implements greater-than.
3327 template<bool IsReverse>
3329 bool operator()(SUnit *A, SUnit *B) const {
3331 return A->NodeNum > B->NodeNum;
3333 return A->NodeNum < B->NodeNum;
3337 /// Reorder instructions as much as possible.
3338 class InstructionShuffler : public MachineSchedStrategy {
3342 // Using a less-than relation (SUnitOrder<false>) for the TopQ priority
3343 // gives nodes with a higher number higher priority causing the latest
3344 // instructions to be scheduled first.
3345 PriorityQueue<SUnit*, std::vector<SUnit*>, SUnitOrder<false> >
3347 // When scheduling bottom-up, use greater-than as the queue priority.
3348 PriorityQueue<SUnit*, std::vector<SUnit*>, SUnitOrder<true> >
3351 InstructionShuffler(bool alternate, bool topdown)
3352 : IsAlternating(alternate), IsTopDown(topdown) {}
3354 virtual void initialize(ScheduleDAGMI*) {
3359 /// Implement MachineSchedStrategy interface.
3360 /// -----------------------------------------
3362 virtual SUnit *pickNode(bool &IsTopNode) {
3366 if (TopQ.empty()) return NULL;
3369 } while (SU->isScheduled);
3374 if (BottomQ.empty()) return NULL;
3377 } while (SU->isScheduled);
3381 IsTopDown = !IsTopDown;
3385 virtual void schedNode(SUnit *SU, bool IsTopNode) {}
3387 virtual void releaseTopNode(SUnit *SU) {
3390 virtual void releaseBottomNode(SUnit *SU) {
3396 static ScheduleDAGInstrs *createInstructionShuffler(MachineSchedContext *C) {
3397 bool Alternate = !ForceTopDown && !ForceBottomUp;
3398 bool TopDown = !ForceBottomUp;
3399 assert((TopDown || !ForceTopDown) &&
3400 "-misched-topdown incompatible with -misched-bottomup");
3401 return new ScheduleDAGMILive(C, new InstructionShuffler(Alternate, TopDown));
3403 static MachineSchedRegistry ShufflerRegistry(
3404 "shuffle", "Shuffle machine instructions alternating directions",
3405 createInstructionShuffler);
3408 //===----------------------------------------------------------------------===//
3409 // GraphWriter support for ScheduleDAGMILive.
3410 //===----------------------------------------------------------------------===//
3415 template<> struct GraphTraits<
3416 ScheduleDAGMI*> : public GraphTraits<ScheduleDAG*> {};
3419 struct DOTGraphTraits<ScheduleDAGMI*> : public DefaultDOTGraphTraits {
3421 DOTGraphTraits (bool isSimple=false) : DefaultDOTGraphTraits(isSimple) {}
3423 static std::string getGraphName(const ScheduleDAG *G) {
3424 return G->MF.getName();
3427 static bool renderGraphFromBottomUp() {
3431 static bool isNodeHidden(const SUnit *Node) {
3432 return (Node->Preds.size() > 10 || Node->Succs.size() > 10);
3435 static bool hasNodeAddressLabel(const SUnit *Node,
3436 const ScheduleDAG *Graph) {
3440 /// If you want to override the dot attributes printed for a particular
3441 /// edge, override this method.
3442 static std::string getEdgeAttributes(const SUnit *Node,
3444 const ScheduleDAG *Graph) {
3445 if (EI.isArtificialDep())
3446 return "color=cyan,style=dashed";
3448 return "color=blue,style=dashed";
3452 static std::string getNodeLabel(const SUnit *SU, const ScheduleDAG *G) {
3454 raw_string_ostream SS(Str);
3455 const ScheduleDAGMI *DAG = static_cast<const ScheduleDAGMI*>(G);
3456 const SchedDFSResult *DFS = DAG->hasVRegLiveness() ?
3457 static_cast<const ScheduleDAGMILive*>(G)->getDFSResult() : 0;
3458 SS << "SU:" << SU->NodeNum;
3460 SS << " I:" << DFS->getNumInstrs(SU);
3463 static std::string getNodeDescription(const SUnit *SU, const ScheduleDAG *G) {
3464 return G->getGraphNodeLabel(SU);
3467 static std::string getNodeAttributes(const SUnit *N, const ScheduleDAG *G) {
3468 std::string Str("shape=Mrecord");
3469 const ScheduleDAGMI *DAG = static_cast<const ScheduleDAGMI*>(G);
3470 const SchedDFSResult *DFS = DAG->hasVRegLiveness() ?
3471 static_cast<const ScheduleDAGMILive*>(G)->getDFSResult() : 0;
3473 Str += ",style=filled,fillcolor=\"#";
3474 Str += DOT::getColorString(DFS->getSubtreeID(N));
3483 /// viewGraph - Pop up a ghostview window with the reachable parts of the DAG
3484 /// rendered using 'dot'.
3486 void ScheduleDAGMI::viewGraph(const Twine &Name, const Twine &Title) {
3488 ViewGraph(this, Name, false, Title);
3490 errs() << "ScheduleDAGMI::viewGraph is only available in debug builds on "
3491 << "systems with Graphviz or gv!\n";
3495 /// Out-of-line implementation with no arguments is handy for gdb.
3496 void ScheduleDAGMI::viewGraph() {
3497 viewGraph(getDAGName(), "Scheduling-Units Graph for " + getDAGName());