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 #include "llvm/CodeGen/MachineScheduler.h"
16 #include "llvm/ADT/PriorityQueue.h"
17 #include "llvm/Analysis/AliasAnalysis.h"
18 #include "llvm/CodeGen/LiveIntervalAnalysis.h"
19 #include "llvm/CodeGen/MachineDominators.h"
20 #include "llvm/CodeGen/MachineLoopInfo.h"
21 #include "llvm/CodeGen/MachineRegisterInfo.h"
22 #include "llvm/CodeGen/Passes.h"
23 #include "llvm/CodeGen/RegisterClassInfo.h"
24 #include "llvm/CodeGen/ScheduleDFS.h"
25 #include "llvm/CodeGen/ScheduleHazardRecognizer.h"
26 #include "llvm/Support/CommandLine.h"
27 #include "llvm/Support/Debug.h"
28 #include "llvm/Support/ErrorHandling.h"
29 #include "llvm/Support/GraphWriter.h"
30 #include "llvm/Support/raw_ostream.h"
31 #include "llvm/Target/TargetInstrInfo.h"
36 #define DEBUG_TYPE "misched"
39 cl::opt<bool> ForceTopDown("misched-topdown", cl::Hidden,
40 cl::desc("Force top-down list scheduling"));
41 cl::opt<bool> ForceBottomUp("misched-bottomup", cl::Hidden,
42 cl::desc("Force bottom-up list scheduling"));
46 static cl::opt<bool> ViewMISchedDAGs("view-misched-dags", cl::Hidden,
47 cl::desc("Pop up a window to show MISched dags after they are processed"));
49 static cl::opt<unsigned> MISchedCutoff("misched-cutoff", cl::Hidden,
50 cl::desc("Stop scheduling after N instructions"), cl::init(~0U));
52 static cl::opt<std::string> SchedOnlyFunc("misched-only-func", cl::Hidden,
53 cl::desc("Only schedule this function"));
54 static cl::opt<unsigned> SchedOnlyBlock("misched-only-block", cl::Hidden,
55 cl::desc("Only schedule this MBB#"));
57 static bool ViewMISchedDAGs = false;
60 static cl::opt<bool> EnableRegPressure("misched-regpressure", cl::Hidden,
61 cl::desc("Enable register pressure scheduling."), cl::init(true));
63 static cl::opt<bool> EnableCyclicPath("misched-cyclicpath", cl::Hidden,
64 cl::desc("Enable cyclic critical path analysis."), cl::init(true));
66 static cl::opt<bool> EnableLoadCluster("misched-cluster", cl::Hidden,
67 cl::desc("Enable load clustering."), cl::init(true));
69 // Experimental heuristics
70 static cl::opt<bool> EnableMacroFusion("misched-fusion", cl::Hidden,
71 cl::desc("Enable scheduling for macro fusion."), cl::init(true));
73 static cl::opt<bool> VerifyScheduling("verify-misched", cl::Hidden,
74 cl::desc("Verify machine instrs before and after machine scheduling"));
76 // DAG subtrees must have at least this many nodes.
77 static const unsigned MinSubtreeSize = 8;
79 // Pin the vtables to this file.
80 void MachineSchedStrategy::anchor() {}
81 void ScheduleDAGMutation::anchor() {}
83 //===----------------------------------------------------------------------===//
84 // Machine Instruction Scheduling Pass and Registry
85 //===----------------------------------------------------------------------===//
87 MachineSchedContext::MachineSchedContext():
88 MF(nullptr), MLI(nullptr), MDT(nullptr), PassConfig(nullptr), AA(nullptr), LIS(nullptr) {
89 RegClassInfo = new RegisterClassInfo();
92 MachineSchedContext::~MachineSchedContext() {
97 /// Base class for a machine scheduler class that can run at any point.
98 class MachineSchedulerBase : public MachineSchedContext,
99 public MachineFunctionPass {
101 MachineSchedulerBase(char &ID): MachineFunctionPass(ID) {}
103 void print(raw_ostream &O, const Module* = nullptr) const override;
106 void scheduleRegions(ScheduleDAGInstrs &Scheduler);
109 /// MachineScheduler runs after coalescing and before register allocation.
110 class MachineScheduler : public MachineSchedulerBase {
114 void getAnalysisUsage(AnalysisUsage &AU) const override;
116 bool runOnMachineFunction(MachineFunction&) override;
118 static char ID; // Class identification, replacement for typeinfo
121 ScheduleDAGInstrs *createMachineScheduler();
124 /// PostMachineScheduler runs after shortly before code emission.
125 class PostMachineScheduler : public MachineSchedulerBase {
127 PostMachineScheduler();
129 void getAnalysisUsage(AnalysisUsage &AU) const override;
131 bool runOnMachineFunction(MachineFunction&) override;
133 static char ID; // Class identification, replacement for typeinfo
136 ScheduleDAGInstrs *createPostMachineScheduler();
140 char MachineScheduler::ID = 0;
142 char &llvm::MachineSchedulerID = MachineScheduler::ID;
144 INITIALIZE_PASS_BEGIN(MachineScheduler, "misched",
145 "Machine Instruction Scheduler", false, false)
146 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
147 INITIALIZE_PASS_DEPENDENCY(SlotIndexes)
148 INITIALIZE_PASS_DEPENDENCY(LiveIntervals)
149 INITIALIZE_PASS_END(MachineScheduler, "misched",
150 "Machine Instruction Scheduler", false, false)
152 MachineScheduler::MachineScheduler()
153 : MachineSchedulerBase(ID) {
154 initializeMachineSchedulerPass(*PassRegistry::getPassRegistry());
157 void MachineScheduler::getAnalysisUsage(AnalysisUsage &AU) const {
158 AU.setPreservesCFG();
159 AU.addRequiredID(MachineDominatorsID);
160 AU.addRequired<MachineLoopInfo>();
161 AU.addRequired<AliasAnalysis>();
162 AU.addRequired<TargetPassConfig>();
163 AU.addRequired<SlotIndexes>();
164 AU.addPreserved<SlotIndexes>();
165 AU.addRequired<LiveIntervals>();
166 AU.addPreserved<LiveIntervals>();
167 MachineFunctionPass::getAnalysisUsage(AU);
170 char PostMachineScheduler::ID = 0;
172 char &llvm::PostMachineSchedulerID = PostMachineScheduler::ID;
174 INITIALIZE_PASS(PostMachineScheduler, "postmisched",
175 "PostRA Machine Instruction Scheduler", false, false)
177 PostMachineScheduler::PostMachineScheduler()
178 : MachineSchedulerBase(ID) {
179 initializePostMachineSchedulerPass(*PassRegistry::getPassRegistry());
182 void PostMachineScheduler::getAnalysisUsage(AnalysisUsage &AU) const {
183 AU.setPreservesCFG();
184 AU.addRequiredID(MachineDominatorsID);
185 AU.addRequired<MachineLoopInfo>();
186 AU.addRequired<TargetPassConfig>();
187 MachineFunctionPass::getAnalysisUsage(AU);
190 MachinePassRegistry MachineSchedRegistry::Registry;
192 /// A dummy default scheduler factory indicates whether the scheduler
193 /// is overridden on the command line.
194 static ScheduleDAGInstrs *useDefaultMachineSched(MachineSchedContext *C) {
198 /// MachineSchedOpt allows command line selection of the scheduler.
199 static cl::opt<MachineSchedRegistry::ScheduleDAGCtor, false,
200 RegisterPassParser<MachineSchedRegistry> >
201 MachineSchedOpt("misched",
202 cl::init(&useDefaultMachineSched), cl::Hidden,
203 cl::desc("Machine instruction scheduler to use"));
205 static MachineSchedRegistry
206 DefaultSchedRegistry("default", "Use the target's default scheduler choice.",
207 useDefaultMachineSched);
209 /// Forward declare the standard machine scheduler. This will be used as the
210 /// default scheduler if the target does not set a default.
211 static ScheduleDAGInstrs *createGenericSchedLive(MachineSchedContext *C);
212 static ScheduleDAGInstrs *createGenericSchedPostRA(MachineSchedContext *C);
214 /// Decrement this iterator until reaching the top or a non-debug instr.
215 static MachineBasicBlock::const_iterator
216 priorNonDebug(MachineBasicBlock::const_iterator I,
217 MachineBasicBlock::const_iterator Beg) {
218 assert(I != Beg && "reached the top of the region, cannot decrement");
220 if (!I->isDebugValue())
226 /// Non-const version.
227 static MachineBasicBlock::iterator
228 priorNonDebug(MachineBasicBlock::iterator I,
229 MachineBasicBlock::const_iterator Beg) {
230 return const_cast<MachineInstr*>(
231 &*priorNonDebug(MachineBasicBlock::const_iterator(I), Beg));
234 /// If this iterator is a debug value, increment until reaching the End or a
235 /// non-debug instruction.
236 static MachineBasicBlock::const_iterator
237 nextIfDebug(MachineBasicBlock::const_iterator I,
238 MachineBasicBlock::const_iterator End) {
239 for(; I != End; ++I) {
240 if (!I->isDebugValue())
246 /// Non-const version.
247 static MachineBasicBlock::iterator
248 nextIfDebug(MachineBasicBlock::iterator I,
249 MachineBasicBlock::const_iterator End) {
250 // Cast the return value to nonconst MachineInstr, then cast to an
251 // instr_iterator, which does not check for null, finally return a
253 return MachineBasicBlock::instr_iterator(
254 const_cast<MachineInstr*>(
255 &*nextIfDebug(MachineBasicBlock::const_iterator(I), End)));
258 /// Instantiate a ScheduleDAGInstrs that will be owned by the caller.
259 ScheduleDAGInstrs *MachineScheduler::createMachineScheduler() {
260 // Select the scheduler, or set the default.
261 MachineSchedRegistry::ScheduleDAGCtor Ctor = MachineSchedOpt;
262 if (Ctor != useDefaultMachineSched)
265 // Get the default scheduler set by the target for this function.
266 ScheduleDAGInstrs *Scheduler = PassConfig->createMachineScheduler(this);
270 // Default to GenericScheduler.
271 return createGenericSchedLive(this);
274 /// Instantiate a ScheduleDAGInstrs for PostRA scheduling that will be owned by
275 /// the caller. We don't have a command line option to override the postRA
276 /// scheduler. The Target must configure it.
277 ScheduleDAGInstrs *PostMachineScheduler::createPostMachineScheduler() {
278 // Get the postRA scheduler set by the target for this function.
279 ScheduleDAGInstrs *Scheduler = PassConfig->createPostMachineScheduler(this);
283 // Default to GenericScheduler.
284 return createGenericSchedPostRA(this);
287 /// Top-level MachineScheduler pass driver.
289 /// Visit blocks in function order. Divide each block into scheduling regions
290 /// and visit them bottom-up. Visiting regions bottom-up is not required, but is
291 /// consistent with the DAG builder, which traverses the interior of the
292 /// scheduling regions bottom-up.
294 /// This design avoids exposing scheduling boundaries to the DAG builder,
295 /// simplifying the DAG builder's support for "special" target instructions.
296 /// At the same time the design allows target schedulers to operate across
297 /// scheduling boundaries, for example to bundle the boudary instructions
298 /// without reordering them. This creates complexity, because the target
299 /// scheduler must update the RegionBegin and RegionEnd positions cached by
300 /// ScheduleDAGInstrs whenever adding or removing instructions. A much simpler
301 /// design would be to split blocks at scheduling boundaries, but LLVM has a
302 /// general bias against block splitting purely for implementation simplicity.
303 bool MachineScheduler::runOnMachineFunction(MachineFunction &mf) {
304 DEBUG(dbgs() << "Before MISsched:\n"; mf.print(dbgs()));
306 // Initialize the context of the pass.
308 MLI = &getAnalysis<MachineLoopInfo>();
309 MDT = &getAnalysis<MachineDominatorTree>();
310 PassConfig = &getAnalysis<TargetPassConfig>();
311 AA = &getAnalysis<AliasAnalysis>();
313 LIS = &getAnalysis<LiveIntervals>();
315 if (VerifyScheduling) {
317 MF->verify(this, "Before machine scheduling.");
319 RegClassInfo->runOnMachineFunction(*MF);
321 // Instantiate the selected scheduler for this target, function, and
322 // optimization level.
323 std::unique_ptr<ScheduleDAGInstrs> Scheduler(createMachineScheduler());
324 scheduleRegions(*Scheduler);
327 if (VerifyScheduling)
328 MF->verify(this, "After machine scheduling.");
332 bool PostMachineScheduler::runOnMachineFunction(MachineFunction &mf) {
333 if (skipOptnoneFunction(*mf.getFunction()))
336 DEBUG(dbgs() << "Before post-MI-sched:\n"; mf.print(dbgs()));
338 // Initialize the context of the pass.
340 PassConfig = &getAnalysis<TargetPassConfig>();
342 if (VerifyScheduling)
343 MF->verify(this, "Before post machine scheduling.");
345 // Instantiate the selected scheduler for this target, function, and
346 // optimization level.
347 std::unique_ptr<ScheduleDAGInstrs> Scheduler(createPostMachineScheduler());
348 scheduleRegions(*Scheduler);
350 if (VerifyScheduling)
351 MF->verify(this, "After post machine scheduling.");
355 /// Return true of the given instruction should not be included in a scheduling
358 /// MachineScheduler does not currently support scheduling across calls. To
359 /// handle calls, the DAG builder needs to be modified to create register
360 /// anti/output dependencies on the registers clobbered by the call's regmask
361 /// operand. In PreRA scheduling, the stack pointer adjustment already prevents
362 /// scheduling across calls. In PostRA scheduling, we need the isCall to enforce
363 /// the boundary, but there would be no benefit to postRA scheduling across
364 /// calls this late anyway.
365 static bool isSchedBoundary(MachineBasicBlock::iterator MI,
366 MachineBasicBlock *MBB,
368 const TargetInstrInfo *TII,
370 return MI->isCall() || TII->isSchedulingBoundary(MI, MBB, *MF);
373 /// Main driver for both MachineScheduler and PostMachineScheduler.
374 void MachineSchedulerBase::scheduleRegions(ScheduleDAGInstrs &Scheduler) {
375 const TargetInstrInfo *TII = MF->getTarget().getInstrInfo();
376 bool IsPostRA = Scheduler.isPostRA();
378 // Visit all machine basic blocks.
380 // TODO: Visit blocks in global postorder or postorder within the bottom-up
381 // loop tree. Then we can optionally compute global RegPressure.
382 for (MachineFunction::iterator MBB = MF->begin(), MBBEnd = MF->end();
383 MBB != MBBEnd; ++MBB) {
385 Scheduler.startBlock(MBB);
388 if (SchedOnlyFunc.getNumOccurrences() && SchedOnlyFunc != MF->getName())
390 if (SchedOnlyBlock.getNumOccurrences()
391 && (int)SchedOnlyBlock != MBB->getNumber())
395 // Break the block into scheduling regions [I, RegionEnd), and schedule each
396 // region as soon as it is discovered. RegionEnd points the scheduling
397 // boundary at the bottom of the region. The DAG does not include RegionEnd,
398 // but the region does (i.e. the next RegionEnd is above the previous
399 // RegionBegin). If the current block has no terminator then RegionEnd ==
400 // MBB->end() for the bottom region.
402 // The Scheduler may insert instructions during either schedule() or
403 // exitRegion(), even for empty regions. So the local iterators 'I' and
404 // 'RegionEnd' are invalid across these calls.
406 // MBB::size() uses instr_iterator to count. Here we need a bundle to count
407 // as a single instruction.
408 unsigned RemainingInstrs = std::distance(MBB->begin(), MBB->end());
409 for(MachineBasicBlock::iterator RegionEnd = MBB->end();
410 RegionEnd != MBB->begin(); RegionEnd = Scheduler.begin()) {
412 // Avoid decrementing RegionEnd for blocks with no terminator.
413 if (RegionEnd != MBB->end() ||
414 isSchedBoundary(std::prev(RegionEnd), MBB, MF, TII, IsPostRA)) {
416 // Count the boundary instruction.
420 // The next region starts above the previous region. Look backward in the
421 // instruction stream until we find the nearest boundary.
422 unsigned NumRegionInstrs = 0;
423 MachineBasicBlock::iterator I = RegionEnd;
424 for(;I != MBB->begin(); --I, --RemainingInstrs, ++NumRegionInstrs) {
425 if (isSchedBoundary(std::prev(I), MBB, MF, TII, IsPostRA))
428 // Notify the scheduler of the region, even if we may skip scheduling
429 // it. Perhaps it still needs to be bundled.
430 Scheduler.enterRegion(MBB, I, RegionEnd, NumRegionInstrs);
432 // Skip empty scheduling regions (0 or 1 schedulable instructions).
433 if (I == RegionEnd || I == std::prev(RegionEnd)) {
434 // Close the current region. Bundle the terminator if needed.
435 // This invalidates 'RegionEnd' and 'I'.
436 Scheduler.exitRegion();
439 DEBUG(dbgs() << "********** " << ((Scheduler.isPostRA()) ? "PostRA " : "")
440 << "MI Scheduling **********\n");
441 DEBUG(dbgs() << MF->getName()
442 << ":BB#" << MBB->getNumber() << " " << MBB->getName()
443 << "\n From: " << *I << " To: ";
444 if (RegionEnd != MBB->end()) dbgs() << *RegionEnd;
445 else dbgs() << "End";
446 dbgs() << " RegionInstrs: " << NumRegionInstrs
447 << " Remaining: " << RemainingInstrs << "\n");
449 // Schedule a region: possibly reorder instructions.
450 // This invalidates 'RegionEnd' and 'I'.
451 Scheduler.schedule();
453 // Close the current region.
454 Scheduler.exitRegion();
456 // Scheduling has invalidated the current iterator 'I'. Ask the
457 // scheduler for the top of it's scheduled region.
458 RegionEnd = Scheduler.begin();
460 assert(RemainingInstrs == 0 && "Instruction count mismatch!");
461 Scheduler.finishBlock();
462 if (Scheduler.isPostRA()) {
463 // FIXME: Ideally, no further passes should rely on kill flags. However,
464 // thumb2 size reduction is currently an exception.
465 Scheduler.fixupKills(MBB);
468 Scheduler.finalizeSchedule();
471 void MachineSchedulerBase::print(raw_ostream &O, const Module* m) const {
475 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
476 void ReadyQueue::dump() {
477 dbgs() << Name << ": ";
478 for (unsigned i = 0, e = Queue.size(); i < e; ++i)
479 dbgs() << Queue[i]->NodeNum << " ";
484 //===----------------------------------------------------------------------===//
485 // ScheduleDAGMI - Basic machine instruction scheduling. This is
486 // independent of PreRA/PostRA scheduling and involves no extra book-keeping for
487 // virtual registers.
488 // ===----------------------------------------------------------------------===/
490 // Provide a vtable anchor.
491 ScheduleDAGMI::~ScheduleDAGMI() {
494 bool ScheduleDAGMI::canAddEdge(SUnit *SuccSU, SUnit *PredSU) {
495 return SuccSU == &ExitSU || !Topo.IsReachable(PredSU, SuccSU);
498 bool ScheduleDAGMI::addEdge(SUnit *SuccSU, const SDep &PredDep) {
499 if (SuccSU != &ExitSU) {
500 // Do not use WillCreateCycle, it assumes SD scheduling.
501 // If Pred is reachable from Succ, then the edge creates a cycle.
502 if (Topo.IsReachable(PredDep.getSUnit(), SuccSU))
504 Topo.AddPred(SuccSU, PredDep.getSUnit());
506 SuccSU->addPred(PredDep, /*Required=*/!PredDep.isArtificial());
507 // Return true regardless of whether a new edge needed to be inserted.
511 /// ReleaseSucc - Decrement the NumPredsLeft count of a successor. When
512 /// NumPredsLeft reaches zero, release the successor node.
514 /// FIXME: Adjust SuccSU height based on MinLatency.
515 void ScheduleDAGMI::releaseSucc(SUnit *SU, SDep *SuccEdge) {
516 SUnit *SuccSU = SuccEdge->getSUnit();
518 if (SuccEdge->isWeak()) {
519 --SuccSU->WeakPredsLeft;
520 if (SuccEdge->isCluster())
521 NextClusterSucc = SuccSU;
525 if (SuccSU->NumPredsLeft == 0) {
526 dbgs() << "*** Scheduling failed! ***\n";
528 dbgs() << " has been released too many times!\n";
529 llvm_unreachable(nullptr);
532 --SuccSU->NumPredsLeft;
533 if (SuccSU->NumPredsLeft == 0 && SuccSU != &ExitSU)
534 SchedImpl->releaseTopNode(SuccSU);
537 /// releaseSuccessors - Call releaseSucc on each of SU's successors.
538 void ScheduleDAGMI::releaseSuccessors(SUnit *SU) {
539 for (SUnit::succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
541 releaseSucc(SU, &*I);
545 /// ReleasePred - Decrement the NumSuccsLeft count of a predecessor. When
546 /// NumSuccsLeft reaches zero, release the predecessor node.
548 /// FIXME: Adjust PredSU height based on MinLatency.
549 void ScheduleDAGMI::releasePred(SUnit *SU, SDep *PredEdge) {
550 SUnit *PredSU = PredEdge->getSUnit();
552 if (PredEdge->isWeak()) {
553 --PredSU->WeakSuccsLeft;
554 if (PredEdge->isCluster())
555 NextClusterPred = PredSU;
559 if (PredSU->NumSuccsLeft == 0) {
560 dbgs() << "*** Scheduling failed! ***\n";
562 dbgs() << " has been released too many times!\n";
563 llvm_unreachable(nullptr);
566 --PredSU->NumSuccsLeft;
567 if (PredSU->NumSuccsLeft == 0 && PredSU != &EntrySU)
568 SchedImpl->releaseBottomNode(PredSU);
571 /// releasePredecessors - Call releasePred on each of SU's predecessors.
572 void ScheduleDAGMI::releasePredecessors(SUnit *SU) {
573 for (SUnit::pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
575 releasePred(SU, &*I);
579 /// enterRegion - Called back from MachineScheduler::runOnMachineFunction after
580 /// crossing a scheduling boundary. [begin, end) includes all instructions in
581 /// the region, including the boundary itself and single-instruction regions
582 /// that don't get scheduled.
583 void ScheduleDAGMI::enterRegion(MachineBasicBlock *bb,
584 MachineBasicBlock::iterator begin,
585 MachineBasicBlock::iterator end,
586 unsigned regioninstrs)
588 ScheduleDAGInstrs::enterRegion(bb, begin, end, regioninstrs);
590 SchedImpl->initPolicy(begin, end, regioninstrs);
593 /// This is normally called from the main scheduler loop but may also be invoked
594 /// by the scheduling strategy to perform additional code motion.
595 void ScheduleDAGMI::moveInstruction(
596 MachineInstr *MI, MachineBasicBlock::iterator InsertPos) {
597 // Advance RegionBegin if the first instruction moves down.
598 if (&*RegionBegin == MI)
601 // Update the instruction stream.
602 BB->splice(InsertPos, BB, MI);
604 // Update LiveIntervals
606 LIS->handleMove(MI, /*UpdateFlags=*/true);
608 // Recede RegionBegin if an instruction moves above the first.
609 if (RegionBegin == InsertPos)
613 bool ScheduleDAGMI::checkSchedLimit() {
615 if (NumInstrsScheduled == MISchedCutoff && MISchedCutoff != ~0U) {
616 CurrentTop = CurrentBottom;
619 ++NumInstrsScheduled;
624 /// Per-region scheduling driver, called back from
625 /// MachineScheduler::runOnMachineFunction. This is a simplified driver that
626 /// does not consider liveness or register pressure. It is useful for PostRA
627 /// scheduling and potentially other custom schedulers.
628 void ScheduleDAGMI::schedule() {
632 Topo.InitDAGTopologicalSorting();
636 SmallVector<SUnit*, 8> TopRoots, BotRoots;
637 findRootsAndBiasEdges(TopRoots, BotRoots);
639 // Initialize the strategy before modifying the DAG.
640 // This may initialize a DFSResult to be used for queue priority.
641 SchedImpl->initialize(this);
643 DEBUG(for (unsigned su = 0, e = SUnits.size(); su != e; ++su)
644 SUnits[su].dumpAll(this));
645 if (ViewMISchedDAGs) viewGraph();
647 // Initialize ready queues now that the DAG and priority data are finalized.
648 initQueues(TopRoots, BotRoots);
650 bool IsTopNode = false;
651 while (SUnit *SU = SchedImpl->pickNode(IsTopNode)) {
652 assert(!SU->isScheduled && "Node already scheduled");
653 if (!checkSchedLimit())
656 MachineInstr *MI = SU->getInstr();
658 assert(SU->isTopReady() && "node still has unscheduled dependencies");
659 if (&*CurrentTop == MI)
660 CurrentTop = nextIfDebug(++CurrentTop, CurrentBottom);
662 moveInstruction(MI, CurrentTop);
665 assert(SU->isBottomReady() && "node still has unscheduled dependencies");
666 MachineBasicBlock::iterator priorII =
667 priorNonDebug(CurrentBottom, CurrentTop);
669 CurrentBottom = priorII;
671 if (&*CurrentTop == MI)
672 CurrentTop = nextIfDebug(++CurrentTop, priorII);
673 moveInstruction(MI, CurrentBottom);
677 updateQueues(SU, IsTopNode);
679 // Notify the scheduling strategy after updating the DAG.
680 SchedImpl->schedNode(SU, IsTopNode);
682 assert(CurrentTop == CurrentBottom && "Nonempty unscheduled zone.");
687 unsigned BBNum = begin()->getParent()->getNumber();
688 dbgs() << "*** Final schedule for BB#" << BBNum << " ***\n";
694 /// Apply each ScheduleDAGMutation step in order.
695 void ScheduleDAGMI::postprocessDAG() {
696 for (unsigned i = 0, e = Mutations.size(); i < e; ++i) {
697 Mutations[i]->apply(this);
702 findRootsAndBiasEdges(SmallVectorImpl<SUnit*> &TopRoots,
703 SmallVectorImpl<SUnit*> &BotRoots) {
704 for (std::vector<SUnit>::iterator
705 I = SUnits.begin(), E = SUnits.end(); I != E; ++I) {
707 assert(!SU->isBoundaryNode() && "Boundary node should not be in SUnits");
709 // Order predecessors so DFSResult follows the critical path.
710 SU->biasCriticalPath();
712 // A SUnit is ready to top schedule if it has no predecessors.
713 if (!I->NumPredsLeft)
714 TopRoots.push_back(SU);
715 // A SUnit is ready to bottom schedule if it has no successors.
716 if (!I->NumSuccsLeft)
717 BotRoots.push_back(SU);
719 ExitSU.biasCriticalPath();
722 /// Identify DAG roots and setup scheduler queues.
723 void ScheduleDAGMI::initQueues(ArrayRef<SUnit*> TopRoots,
724 ArrayRef<SUnit*> BotRoots) {
725 NextClusterSucc = nullptr;
726 NextClusterPred = nullptr;
728 // Release all DAG roots for scheduling, not including EntrySU/ExitSU.
730 // Nodes with unreleased weak edges can still be roots.
731 // Release top roots in forward order.
732 for (SmallVectorImpl<SUnit*>::const_iterator
733 I = TopRoots.begin(), E = TopRoots.end(); I != E; ++I) {
734 SchedImpl->releaseTopNode(*I);
736 // Release bottom roots in reverse order so the higher priority nodes appear
737 // first. This is more natural and slightly more efficient.
738 for (SmallVectorImpl<SUnit*>::const_reverse_iterator
739 I = BotRoots.rbegin(), E = BotRoots.rend(); I != E; ++I) {
740 SchedImpl->releaseBottomNode(*I);
743 releaseSuccessors(&EntrySU);
744 releasePredecessors(&ExitSU);
746 SchedImpl->registerRoots();
748 // Advance past initial DebugValues.
749 CurrentTop = nextIfDebug(RegionBegin, RegionEnd);
750 CurrentBottom = RegionEnd;
753 /// Update scheduler queues after scheduling an instruction.
754 void ScheduleDAGMI::updateQueues(SUnit *SU, bool IsTopNode) {
755 // Release dependent instructions for scheduling.
757 releaseSuccessors(SU);
759 releasePredecessors(SU);
761 SU->isScheduled = true;
764 /// Reinsert any remaining debug_values, just like the PostRA scheduler.
765 void ScheduleDAGMI::placeDebugValues() {
766 // If first instruction was a DBG_VALUE then put it back.
768 BB->splice(RegionBegin, BB, FirstDbgValue);
769 RegionBegin = FirstDbgValue;
772 for (std::vector<std::pair<MachineInstr *, MachineInstr *> >::iterator
773 DI = DbgValues.end(), DE = DbgValues.begin(); DI != DE; --DI) {
774 std::pair<MachineInstr *, MachineInstr *> P = *std::prev(DI);
775 MachineInstr *DbgValue = P.first;
776 MachineBasicBlock::iterator OrigPrevMI = P.second;
777 if (&*RegionBegin == DbgValue)
779 BB->splice(++OrigPrevMI, BB, DbgValue);
780 if (OrigPrevMI == std::prev(RegionEnd))
781 RegionEnd = DbgValue;
784 FirstDbgValue = nullptr;
787 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
788 void ScheduleDAGMI::dumpSchedule() const {
789 for (MachineBasicBlock::iterator MI = begin(), ME = end(); MI != ME; ++MI) {
790 if (SUnit *SU = getSUnit(&(*MI)))
793 dbgs() << "Missing SUnit\n";
798 //===----------------------------------------------------------------------===//
799 // ScheduleDAGMILive - Base class for MachineInstr scheduling with LiveIntervals
801 //===----------------------------------------------------------------------===//
803 ScheduleDAGMILive::~ScheduleDAGMILive() {
807 /// enterRegion - Called back from MachineScheduler::runOnMachineFunction after
808 /// crossing a scheduling boundary. [begin, end) includes all instructions in
809 /// the region, including the boundary itself and single-instruction regions
810 /// that don't get scheduled.
811 void ScheduleDAGMILive::enterRegion(MachineBasicBlock *bb,
812 MachineBasicBlock::iterator begin,
813 MachineBasicBlock::iterator end,
814 unsigned regioninstrs)
816 // ScheduleDAGMI initializes SchedImpl's per-region policy.
817 ScheduleDAGMI::enterRegion(bb, begin, end, regioninstrs);
819 // For convenience remember the end of the liveness region.
820 LiveRegionEnd = (RegionEnd == bb->end()) ? RegionEnd : std::next(RegionEnd);
822 SUPressureDiffs.clear();
824 ShouldTrackPressure = SchedImpl->shouldTrackPressure();
827 // Setup the register pressure trackers for the top scheduled top and bottom
828 // scheduled regions.
829 void ScheduleDAGMILive::initRegPressure() {
830 TopRPTracker.init(&MF, RegClassInfo, LIS, BB, RegionBegin);
831 BotRPTracker.init(&MF, RegClassInfo, LIS, BB, LiveRegionEnd);
833 // Close the RPTracker to finalize live ins.
834 RPTracker.closeRegion();
836 DEBUG(RPTracker.dump());
838 // Initialize the live ins and live outs.
839 TopRPTracker.addLiveRegs(RPTracker.getPressure().LiveInRegs);
840 BotRPTracker.addLiveRegs(RPTracker.getPressure().LiveOutRegs);
842 // Close one end of the tracker so we can call
843 // getMaxUpward/DownwardPressureDelta before advancing across any
844 // instructions. This converts currently live regs into live ins/outs.
845 TopRPTracker.closeTop();
846 BotRPTracker.closeBottom();
848 BotRPTracker.initLiveThru(RPTracker);
849 if (!BotRPTracker.getLiveThru().empty()) {
850 TopRPTracker.initLiveThru(BotRPTracker.getLiveThru());
851 DEBUG(dbgs() << "Live Thru: ";
852 dumpRegSetPressure(BotRPTracker.getLiveThru(), TRI));
855 // For each live out vreg reduce the pressure change associated with other
856 // uses of the same vreg below the live-out reaching def.
857 updatePressureDiffs(RPTracker.getPressure().LiveOutRegs);
859 // Account for liveness generated by the region boundary.
860 if (LiveRegionEnd != RegionEnd) {
861 SmallVector<unsigned, 8> LiveUses;
862 BotRPTracker.recede(&LiveUses);
863 updatePressureDiffs(LiveUses);
866 assert(BotRPTracker.getPos() == RegionEnd && "Can't find the region bottom");
868 // Cache the list of excess pressure sets in this region. This will also track
869 // the max pressure in the scheduled code for these sets.
870 RegionCriticalPSets.clear();
871 const std::vector<unsigned> &RegionPressure =
872 RPTracker.getPressure().MaxSetPressure;
873 for (unsigned i = 0, e = RegionPressure.size(); i < e; ++i) {
874 unsigned Limit = RegClassInfo->getRegPressureSetLimit(i);
875 if (RegionPressure[i] > Limit) {
876 DEBUG(dbgs() << TRI->getRegPressureSetName(i)
877 << " Limit " << Limit
878 << " Actual " << RegionPressure[i] << "\n");
879 RegionCriticalPSets.push_back(PressureChange(i));
882 DEBUG(dbgs() << "Excess PSets: ";
883 for (unsigned i = 0, e = RegionCriticalPSets.size(); i != e; ++i)
884 dbgs() << TRI->getRegPressureSetName(
885 RegionCriticalPSets[i].getPSet()) << " ";
889 void ScheduleDAGMILive::
890 updateScheduledPressure(const SUnit *SU,
891 const std::vector<unsigned> &NewMaxPressure) {
892 const PressureDiff &PDiff = getPressureDiff(SU);
893 unsigned CritIdx = 0, CritEnd = RegionCriticalPSets.size();
894 for (PressureDiff::const_iterator I = PDiff.begin(), E = PDiff.end();
898 unsigned ID = I->getPSet();
899 while (CritIdx != CritEnd && RegionCriticalPSets[CritIdx].getPSet() < ID)
901 if (CritIdx != CritEnd && RegionCriticalPSets[CritIdx].getPSet() == ID) {
902 if ((int)NewMaxPressure[ID] > RegionCriticalPSets[CritIdx].getUnitInc()
903 && NewMaxPressure[ID] <= INT16_MAX)
904 RegionCriticalPSets[CritIdx].setUnitInc(NewMaxPressure[ID]);
906 unsigned Limit = RegClassInfo->getRegPressureSetLimit(ID);
907 if (NewMaxPressure[ID] >= Limit - 2) {
908 DEBUG(dbgs() << " " << TRI->getRegPressureSetName(ID) << ": "
909 << NewMaxPressure[ID] << " > " << Limit << "(+ "
910 << BotRPTracker.getLiveThru()[ID] << " livethru)\n");
915 /// Update the PressureDiff array for liveness after scheduling this
917 void ScheduleDAGMILive::updatePressureDiffs(ArrayRef<unsigned> LiveUses) {
918 for (unsigned LUIdx = 0, LUEnd = LiveUses.size(); LUIdx != LUEnd; ++LUIdx) {
919 /// FIXME: Currently assuming single-use physregs.
920 unsigned Reg = LiveUses[LUIdx];
921 DEBUG(dbgs() << " LiveReg: " << PrintVRegOrUnit(Reg, TRI) << "\n");
922 if (!TRI->isVirtualRegister(Reg))
925 // This may be called before CurrentBottom has been initialized. However,
926 // BotRPTracker must have a valid position. We want the value live into the
927 // instruction or live out of the block, so ask for the previous
928 // instruction's live-out.
929 const LiveInterval &LI = LIS->getInterval(Reg);
931 MachineBasicBlock::const_iterator I =
932 nextIfDebug(BotRPTracker.getPos(), BB->end());
934 VNI = LI.getVNInfoBefore(LIS->getMBBEndIdx(BB));
936 LiveQueryResult LRQ = LI.Query(LIS->getInstructionIndex(I));
939 // RegisterPressureTracker guarantees that readsReg is true for LiveUses.
940 assert(VNI && "No live value at use.");
941 for (VReg2UseMap::iterator
942 UI = VRegUses.find(Reg); UI != VRegUses.end(); ++UI) {
944 DEBUG(dbgs() << " UpdateRegP: SU(" << SU->NodeNum << ") "
946 // If this use comes before the reaching def, it cannot be a last use, so
947 // descrease its pressure change.
948 if (!SU->isScheduled && SU != &ExitSU) {
950 = LI.Query(LIS->getInstructionIndex(SU->getInstr()));
951 if (LRQ.valueIn() == VNI)
952 getPressureDiff(SU).addPressureChange(Reg, true, &MRI);
958 /// schedule - Called back from MachineScheduler::runOnMachineFunction
959 /// after setting up the current scheduling region. [RegionBegin, RegionEnd)
960 /// only includes instructions that have DAG nodes, not scheduling boundaries.
962 /// This is a skeletal driver, with all the functionality pushed into helpers,
963 /// so that it can be easilly extended by experimental schedulers. Generally,
964 /// implementing MachineSchedStrategy should be sufficient to implement a new
965 /// scheduling algorithm. However, if a scheduler further subclasses
966 /// ScheduleDAGMILive then it will want to override this virtual method in order
967 /// to update any specialized state.
968 void ScheduleDAGMILive::schedule() {
969 buildDAGWithRegPressure();
971 Topo.InitDAGTopologicalSorting();
975 SmallVector<SUnit*, 8> TopRoots, BotRoots;
976 findRootsAndBiasEdges(TopRoots, BotRoots);
978 // Initialize the strategy before modifying the DAG.
979 // This may initialize a DFSResult to be used for queue priority.
980 SchedImpl->initialize(this);
982 DEBUG(for (unsigned su = 0, e = SUnits.size(); su != e; ++su)
983 SUnits[su].dumpAll(this));
984 if (ViewMISchedDAGs) viewGraph();
986 // Initialize ready queues now that the DAG and priority data are finalized.
987 initQueues(TopRoots, BotRoots);
989 if (ShouldTrackPressure) {
990 assert(TopRPTracker.getPos() == RegionBegin && "bad initial Top tracker");
991 TopRPTracker.setPos(CurrentTop);
994 bool IsTopNode = false;
995 while (SUnit *SU = SchedImpl->pickNode(IsTopNode)) {
996 assert(!SU->isScheduled && "Node already scheduled");
997 if (!checkSchedLimit())
1000 scheduleMI(SU, IsTopNode);
1002 updateQueues(SU, IsTopNode);
1005 unsigned SubtreeID = DFSResult->getSubtreeID(SU);
1006 if (!ScheduledTrees.test(SubtreeID)) {
1007 ScheduledTrees.set(SubtreeID);
1008 DFSResult->scheduleTree(SubtreeID);
1009 SchedImpl->scheduleTree(SubtreeID);
1013 // Notify the scheduling strategy after updating the DAG.
1014 SchedImpl->schedNode(SU, IsTopNode);
1016 assert(CurrentTop == CurrentBottom && "Nonempty unscheduled zone.");
1021 unsigned BBNum = begin()->getParent()->getNumber();
1022 dbgs() << "*** Final schedule for BB#" << BBNum << " ***\n";
1028 /// Build the DAG and setup three register pressure trackers.
1029 void ScheduleDAGMILive::buildDAGWithRegPressure() {
1030 if (!ShouldTrackPressure) {
1032 RegionCriticalPSets.clear();
1033 buildSchedGraph(AA);
1037 // Initialize the register pressure tracker used by buildSchedGraph.
1038 RPTracker.init(&MF, RegClassInfo, LIS, BB, LiveRegionEnd,
1039 /*TrackUntiedDefs=*/true);
1041 // Account for liveness generate by the region boundary.
1042 if (LiveRegionEnd != RegionEnd)
1045 // Build the DAG, and compute current register pressure.
1046 buildSchedGraph(AA, &RPTracker, &SUPressureDiffs);
1048 // Initialize top/bottom trackers after computing region pressure.
1052 void ScheduleDAGMILive::computeDFSResult() {
1054 DFSResult = new SchedDFSResult(/*BottomU*/true, MinSubtreeSize);
1056 ScheduledTrees.clear();
1057 DFSResult->resize(SUnits.size());
1058 DFSResult->compute(SUnits);
1059 ScheduledTrees.resize(DFSResult->getNumSubtrees());
1062 /// Compute the max cyclic critical path through the DAG. The scheduling DAG
1063 /// only provides the critical path for single block loops. To handle loops that
1064 /// span blocks, we could use the vreg path latencies provided by
1065 /// MachineTraceMetrics instead. However, MachineTraceMetrics is not currently
1066 /// available for use in the scheduler.
1068 /// The cyclic path estimation identifies a def-use pair that crosses the back
1069 /// edge and considers the depth and height of the nodes. For example, consider
1070 /// the following instruction sequence where each instruction has unit latency
1071 /// and defines an epomymous virtual register:
1073 /// a->b(a,c)->c(b)->d(c)->exit
1075 /// The cyclic critical path is a two cycles: b->c->b
1076 /// The acyclic critical path is four cycles: a->b->c->d->exit
1077 /// LiveOutHeight = height(c) = len(c->d->exit) = 2
1078 /// LiveOutDepth = depth(c) + 1 = len(a->b->c) + 1 = 3
1079 /// LiveInHeight = height(b) + 1 = len(b->c->d->exit) + 1 = 4
1080 /// LiveInDepth = depth(b) = len(a->b) = 1
1082 /// LiveOutDepth - LiveInDepth = 3 - 1 = 2
1083 /// LiveInHeight - LiveOutHeight = 4 - 2 = 2
1084 /// CyclicCriticalPath = min(2, 2) = 2
1086 /// This could be relevant to PostRA scheduling, but is currently implemented
1087 /// assuming LiveIntervals.
1088 unsigned ScheduleDAGMILive::computeCyclicCriticalPath() {
1089 // This only applies to single block loop.
1090 if (!BB->isSuccessor(BB))
1093 unsigned MaxCyclicLatency = 0;
1094 // Visit each live out vreg def to find def/use pairs that cross iterations.
1095 ArrayRef<unsigned> LiveOuts = RPTracker.getPressure().LiveOutRegs;
1096 for (ArrayRef<unsigned>::iterator RI = LiveOuts.begin(), RE = LiveOuts.end();
1099 if (!TRI->isVirtualRegister(Reg))
1101 const LiveInterval &LI = LIS->getInterval(Reg);
1102 const VNInfo *DefVNI = LI.getVNInfoBefore(LIS->getMBBEndIdx(BB));
1106 MachineInstr *DefMI = LIS->getInstructionFromIndex(DefVNI->def);
1107 const SUnit *DefSU = getSUnit(DefMI);
1111 unsigned LiveOutHeight = DefSU->getHeight();
1112 unsigned LiveOutDepth = DefSU->getDepth() + DefSU->Latency;
1113 // Visit all local users of the vreg def.
1114 for (VReg2UseMap::iterator
1115 UI = VRegUses.find(Reg); UI != VRegUses.end(); ++UI) {
1116 if (UI->SU == &ExitSU)
1119 // Only consider uses of the phi.
1120 LiveQueryResult LRQ =
1121 LI.Query(LIS->getInstructionIndex(UI->SU->getInstr()));
1122 if (!LRQ.valueIn()->isPHIDef())
1125 // Assume that a path spanning two iterations is a cycle, which could
1126 // overestimate in strange cases. This allows cyclic latency to be
1127 // estimated as the minimum slack of the vreg's depth or height.
1128 unsigned CyclicLatency = 0;
1129 if (LiveOutDepth > UI->SU->getDepth())
1130 CyclicLatency = LiveOutDepth - UI->SU->getDepth();
1132 unsigned LiveInHeight = UI->SU->getHeight() + DefSU->Latency;
1133 if (LiveInHeight > LiveOutHeight) {
1134 if (LiveInHeight - LiveOutHeight < CyclicLatency)
1135 CyclicLatency = LiveInHeight - LiveOutHeight;
1140 DEBUG(dbgs() << "Cyclic Path: SU(" << DefSU->NodeNum << ") -> SU("
1141 << UI->SU->NodeNum << ") = " << CyclicLatency << "c\n");
1142 if (CyclicLatency > MaxCyclicLatency)
1143 MaxCyclicLatency = CyclicLatency;
1146 DEBUG(dbgs() << "Cyclic Critical Path: " << MaxCyclicLatency << "c\n");
1147 return MaxCyclicLatency;
1150 /// Move an instruction and update register pressure.
1151 void ScheduleDAGMILive::scheduleMI(SUnit *SU, bool IsTopNode) {
1152 // Move the instruction to its new location in the instruction stream.
1153 MachineInstr *MI = SU->getInstr();
1156 assert(SU->isTopReady() && "node still has unscheduled dependencies");
1157 if (&*CurrentTop == MI)
1158 CurrentTop = nextIfDebug(++CurrentTop, CurrentBottom);
1160 moveInstruction(MI, CurrentTop);
1161 TopRPTracker.setPos(MI);
1164 if (ShouldTrackPressure) {
1165 // Update top scheduled pressure.
1166 TopRPTracker.advance();
1167 assert(TopRPTracker.getPos() == CurrentTop && "out of sync");
1168 updateScheduledPressure(SU, TopRPTracker.getPressure().MaxSetPressure);
1172 assert(SU->isBottomReady() && "node still has unscheduled dependencies");
1173 MachineBasicBlock::iterator priorII =
1174 priorNonDebug(CurrentBottom, CurrentTop);
1175 if (&*priorII == MI)
1176 CurrentBottom = priorII;
1178 if (&*CurrentTop == MI) {
1179 CurrentTop = nextIfDebug(++CurrentTop, priorII);
1180 TopRPTracker.setPos(CurrentTop);
1182 moveInstruction(MI, CurrentBottom);
1185 if (ShouldTrackPressure) {
1186 // Update bottom scheduled pressure.
1187 SmallVector<unsigned, 8> LiveUses;
1188 BotRPTracker.recede(&LiveUses);
1189 assert(BotRPTracker.getPos() == CurrentBottom && "out of sync");
1190 updateScheduledPressure(SU, BotRPTracker.getPressure().MaxSetPressure);
1191 updatePressureDiffs(LiveUses);
1196 //===----------------------------------------------------------------------===//
1197 // LoadClusterMutation - DAG post-processing to cluster loads.
1198 //===----------------------------------------------------------------------===//
1201 /// \brief Post-process the DAG to create cluster edges between neighboring
1203 class LoadClusterMutation : public ScheduleDAGMutation {
1208 LoadInfo(SUnit *su, unsigned reg, unsigned ofs)
1209 : SU(su), BaseReg(reg), Offset(ofs) {}
1211 bool operator<(const LoadInfo &RHS) const {
1212 return std::tie(BaseReg, Offset) < std::tie(RHS.BaseReg, RHS.Offset);
1216 const TargetInstrInfo *TII;
1217 const TargetRegisterInfo *TRI;
1219 LoadClusterMutation(const TargetInstrInfo *tii,
1220 const TargetRegisterInfo *tri)
1221 : TII(tii), TRI(tri) {}
1223 void apply(ScheduleDAGMI *DAG) override;
1225 void clusterNeighboringLoads(ArrayRef<SUnit*> Loads, ScheduleDAGMI *DAG);
1229 void LoadClusterMutation::clusterNeighboringLoads(ArrayRef<SUnit*> Loads,
1230 ScheduleDAGMI *DAG) {
1231 SmallVector<LoadClusterMutation::LoadInfo,32> LoadRecords;
1232 for (unsigned Idx = 0, End = Loads.size(); Idx != End; ++Idx) {
1233 SUnit *SU = Loads[Idx];
1236 if (TII->getLdStBaseRegImmOfs(SU->getInstr(), BaseReg, Offset, TRI))
1237 LoadRecords.push_back(LoadInfo(SU, BaseReg, Offset));
1239 if (LoadRecords.size() < 2)
1241 std::sort(LoadRecords.begin(), LoadRecords.end());
1242 unsigned ClusterLength = 1;
1243 for (unsigned Idx = 0, End = LoadRecords.size(); Idx < (End - 1); ++Idx) {
1244 if (LoadRecords[Idx].BaseReg != LoadRecords[Idx+1].BaseReg) {
1249 SUnit *SUa = LoadRecords[Idx].SU;
1250 SUnit *SUb = LoadRecords[Idx+1].SU;
1251 if (TII->shouldClusterLoads(SUa->getInstr(), SUb->getInstr(), ClusterLength)
1252 && DAG->addEdge(SUb, SDep(SUa, SDep::Cluster))) {
1254 DEBUG(dbgs() << "Cluster loads SU(" << SUa->NodeNum << ") - SU("
1255 << SUb->NodeNum << ")\n");
1256 // Copy successor edges from SUa to SUb. Interleaving computation
1257 // dependent on SUa can prevent load combining due to register reuse.
1258 // Predecessor edges do not need to be copied from SUb to SUa since nearby
1259 // loads should have effectively the same inputs.
1260 for (SUnit::const_succ_iterator
1261 SI = SUa->Succs.begin(), SE = SUa->Succs.end(); SI != SE; ++SI) {
1262 if (SI->getSUnit() == SUb)
1264 DEBUG(dbgs() << " Copy Succ SU(" << SI->getSUnit()->NodeNum << ")\n");
1265 DAG->addEdge(SI->getSUnit(), SDep(SUb, SDep::Artificial));
1274 /// \brief Callback from DAG postProcessing to create cluster edges for loads.
1275 void LoadClusterMutation::apply(ScheduleDAGMI *DAG) {
1276 // Map DAG NodeNum to store chain ID.
1277 DenseMap<unsigned, unsigned> StoreChainIDs;
1278 // Map each store chain to a set of dependent loads.
1279 SmallVector<SmallVector<SUnit*,4>, 32> StoreChainDependents;
1280 for (unsigned Idx = 0, End = DAG->SUnits.size(); Idx != End; ++Idx) {
1281 SUnit *SU = &DAG->SUnits[Idx];
1282 if (!SU->getInstr()->mayLoad())
1284 unsigned ChainPredID = DAG->SUnits.size();
1285 for (SUnit::const_pred_iterator
1286 PI = SU->Preds.begin(), PE = SU->Preds.end(); PI != PE; ++PI) {
1288 ChainPredID = PI->getSUnit()->NodeNum;
1292 // Check if this chain-like pred has been seen
1293 // before. ChainPredID==MaxNodeID for loads at the top of the schedule.
1294 unsigned NumChains = StoreChainDependents.size();
1295 std::pair<DenseMap<unsigned, unsigned>::iterator, bool> Result =
1296 StoreChainIDs.insert(std::make_pair(ChainPredID, NumChains));
1298 StoreChainDependents.resize(NumChains + 1);
1299 StoreChainDependents[Result.first->second].push_back(SU);
1301 // Iterate over the store chains.
1302 for (unsigned Idx = 0, End = StoreChainDependents.size(); Idx != End; ++Idx)
1303 clusterNeighboringLoads(StoreChainDependents[Idx], DAG);
1306 //===----------------------------------------------------------------------===//
1307 // MacroFusion - DAG post-processing to encourage fusion of macro ops.
1308 //===----------------------------------------------------------------------===//
1311 /// \brief Post-process the DAG to create cluster edges between instructions
1312 /// that may be fused by the processor into a single operation.
1313 class MacroFusion : public ScheduleDAGMutation {
1314 const TargetInstrInfo *TII;
1316 MacroFusion(const TargetInstrInfo *tii): TII(tii) {}
1318 void apply(ScheduleDAGMI *DAG) override;
1322 /// \brief Callback from DAG postProcessing to create cluster edges to encourage
1323 /// fused operations.
1324 void MacroFusion::apply(ScheduleDAGMI *DAG) {
1325 // For now, assume targets can only fuse with the branch.
1326 MachineInstr *Branch = DAG->ExitSU.getInstr();
1330 for (unsigned Idx = DAG->SUnits.size(); Idx > 0;) {
1331 SUnit *SU = &DAG->SUnits[--Idx];
1332 if (!TII->shouldScheduleAdjacent(SU->getInstr(), Branch))
1335 // Create a single weak edge from SU to ExitSU. The only effect is to cause
1336 // bottom-up scheduling to heavily prioritize the clustered SU. There is no
1337 // need to copy predecessor edges from ExitSU to SU, since top-down
1338 // scheduling cannot prioritize ExitSU anyway. To defer top-down scheduling
1339 // of SU, we could create an artificial edge from the deepest root, but it
1340 // hasn't been needed yet.
1341 bool Success = DAG->addEdge(&DAG->ExitSU, SDep(SU, SDep::Cluster));
1343 assert(Success && "No DAG nodes should be reachable from ExitSU");
1345 DEBUG(dbgs() << "Macro Fuse SU(" << SU->NodeNum << ")\n");
1350 //===----------------------------------------------------------------------===//
1351 // CopyConstrain - DAG post-processing to encourage copy elimination.
1352 //===----------------------------------------------------------------------===//
1355 /// \brief Post-process the DAG to create weak edges from all uses of a copy to
1356 /// the one use that defines the copy's source vreg, most likely an induction
1357 /// variable increment.
1358 class CopyConstrain : public ScheduleDAGMutation {
1360 SlotIndex RegionBeginIdx;
1361 // RegionEndIdx is the slot index of the last non-debug instruction in the
1362 // scheduling region. So we may have RegionBeginIdx == RegionEndIdx.
1363 SlotIndex RegionEndIdx;
1365 CopyConstrain(const TargetInstrInfo *, const TargetRegisterInfo *) {}
1367 void apply(ScheduleDAGMI *DAG) override;
1370 void constrainLocalCopy(SUnit *CopySU, ScheduleDAGMILive *DAG);
1374 /// constrainLocalCopy handles two possibilities:
1379 /// I3: dst = src (copy)
1380 /// (create pred->succ edges I0->I1, I2->I1)
1383 /// I0: dst = src (copy)
1387 /// (create pred->succ edges I1->I2, I3->I2)
1389 /// Although the MachineScheduler is currently constrained to single blocks,
1390 /// this algorithm should handle extended blocks. An EBB is a set of
1391 /// contiguously numbered blocks such that the previous block in the EBB is
1392 /// always the single predecessor.
1393 void CopyConstrain::constrainLocalCopy(SUnit *CopySU, ScheduleDAGMILive *DAG) {
1394 LiveIntervals *LIS = DAG->getLIS();
1395 MachineInstr *Copy = CopySU->getInstr();
1397 // Check for pure vreg copies.
1398 unsigned SrcReg = Copy->getOperand(1).getReg();
1399 if (!TargetRegisterInfo::isVirtualRegister(SrcReg))
1402 unsigned DstReg = Copy->getOperand(0).getReg();
1403 if (!TargetRegisterInfo::isVirtualRegister(DstReg))
1406 // Check if either the dest or source is local. If it's live across a back
1407 // edge, it's not local. Note that if both vregs are live across the back
1408 // edge, we cannot successfully contrain the copy without cyclic scheduling.
1409 unsigned LocalReg = DstReg;
1410 unsigned GlobalReg = SrcReg;
1411 LiveInterval *LocalLI = &LIS->getInterval(LocalReg);
1412 if (!LocalLI->isLocal(RegionBeginIdx, RegionEndIdx)) {
1415 LocalLI = &LIS->getInterval(LocalReg);
1416 if (!LocalLI->isLocal(RegionBeginIdx, RegionEndIdx))
1419 LiveInterval *GlobalLI = &LIS->getInterval(GlobalReg);
1421 // Find the global segment after the start of the local LI.
1422 LiveInterval::iterator GlobalSegment = GlobalLI->find(LocalLI->beginIndex());
1423 // If GlobalLI does not overlap LocalLI->start, then a copy directly feeds a
1424 // local live range. We could create edges from other global uses to the local
1425 // start, but the coalescer should have already eliminated these cases, so
1426 // don't bother dealing with it.
1427 if (GlobalSegment == GlobalLI->end())
1430 // If GlobalSegment is killed at the LocalLI->start, the call to find()
1431 // returned the next global segment. But if GlobalSegment overlaps with
1432 // LocalLI->start, then advance to the next segement. If a hole in GlobalLI
1433 // exists in LocalLI's vicinity, GlobalSegment will be the end of the hole.
1434 if (GlobalSegment->contains(LocalLI->beginIndex()))
1437 if (GlobalSegment == GlobalLI->end())
1440 // Check if GlobalLI contains a hole in the vicinity of LocalLI.
1441 if (GlobalSegment != GlobalLI->begin()) {
1442 // Two address defs have no hole.
1443 if (SlotIndex::isSameInstr(std::prev(GlobalSegment)->end,
1444 GlobalSegment->start)) {
1447 // If the prior global segment may be defined by the same two-address
1448 // instruction that also defines LocalLI, then can't make a hole here.
1449 if (SlotIndex::isSameInstr(std::prev(GlobalSegment)->start,
1450 LocalLI->beginIndex())) {
1453 // If GlobalLI has a prior segment, it must be live into the EBB. Otherwise
1454 // it would be a disconnected component in the live range.
1455 assert(std::prev(GlobalSegment)->start < LocalLI->beginIndex() &&
1456 "Disconnected LRG within the scheduling region.");
1458 MachineInstr *GlobalDef = LIS->getInstructionFromIndex(GlobalSegment->start);
1462 SUnit *GlobalSU = DAG->getSUnit(GlobalDef);
1466 // GlobalDef is the bottom of the GlobalLI hole. Open the hole by
1467 // constraining the uses of the last local def to precede GlobalDef.
1468 SmallVector<SUnit*,8> LocalUses;
1469 const VNInfo *LastLocalVN = LocalLI->getVNInfoBefore(LocalLI->endIndex());
1470 MachineInstr *LastLocalDef = LIS->getInstructionFromIndex(LastLocalVN->def);
1471 SUnit *LastLocalSU = DAG->getSUnit(LastLocalDef);
1472 for (SUnit::const_succ_iterator
1473 I = LastLocalSU->Succs.begin(), E = LastLocalSU->Succs.end();
1475 if (I->getKind() != SDep::Data || I->getReg() != LocalReg)
1477 if (I->getSUnit() == GlobalSU)
1479 if (!DAG->canAddEdge(GlobalSU, I->getSUnit()))
1481 LocalUses.push_back(I->getSUnit());
1483 // Open the top of the GlobalLI hole by constraining any earlier global uses
1484 // to precede the start of LocalLI.
1485 SmallVector<SUnit*,8> GlobalUses;
1486 MachineInstr *FirstLocalDef =
1487 LIS->getInstructionFromIndex(LocalLI->beginIndex());
1488 SUnit *FirstLocalSU = DAG->getSUnit(FirstLocalDef);
1489 for (SUnit::const_pred_iterator
1490 I = GlobalSU->Preds.begin(), E = GlobalSU->Preds.end(); I != E; ++I) {
1491 if (I->getKind() != SDep::Anti || I->getReg() != GlobalReg)
1493 if (I->getSUnit() == FirstLocalSU)
1495 if (!DAG->canAddEdge(FirstLocalSU, I->getSUnit()))
1497 GlobalUses.push_back(I->getSUnit());
1499 DEBUG(dbgs() << "Constraining copy SU(" << CopySU->NodeNum << ")\n");
1500 // Add the weak edges.
1501 for (SmallVectorImpl<SUnit*>::const_iterator
1502 I = LocalUses.begin(), E = LocalUses.end(); I != E; ++I) {
1503 DEBUG(dbgs() << " Local use SU(" << (*I)->NodeNum << ") -> SU("
1504 << GlobalSU->NodeNum << ")\n");
1505 DAG->addEdge(GlobalSU, SDep(*I, SDep::Weak));
1507 for (SmallVectorImpl<SUnit*>::const_iterator
1508 I = GlobalUses.begin(), E = GlobalUses.end(); I != E; ++I) {
1509 DEBUG(dbgs() << " Global use SU(" << (*I)->NodeNum << ") -> SU("
1510 << FirstLocalSU->NodeNum << ")\n");
1511 DAG->addEdge(FirstLocalSU, SDep(*I, SDep::Weak));
1515 /// \brief Callback from DAG postProcessing to create weak edges to encourage
1516 /// copy elimination.
1517 void CopyConstrain::apply(ScheduleDAGMI *DAG) {
1518 assert(DAG->hasVRegLiveness() && "Expect VRegs with LiveIntervals");
1520 MachineBasicBlock::iterator FirstPos = nextIfDebug(DAG->begin(), DAG->end());
1521 if (FirstPos == DAG->end())
1523 RegionBeginIdx = DAG->getLIS()->getInstructionIndex(&*FirstPos);
1524 RegionEndIdx = DAG->getLIS()->getInstructionIndex(
1525 &*priorNonDebug(DAG->end(), DAG->begin()));
1527 for (unsigned Idx = 0, End = DAG->SUnits.size(); Idx != End; ++Idx) {
1528 SUnit *SU = &DAG->SUnits[Idx];
1529 if (!SU->getInstr()->isCopy())
1532 constrainLocalCopy(SU, static_cast<ScheduleDAGMILive*>(DAG));
1536 //===----------------------------------------------------------------------===//
1537 // MachineSchedStrategy helpers used by GenericScheduler, GenericPostScheduler
1538 // and possibly other custom schedulers.
1539 //===----------------------------------------------------------------------===//
1541 static const unsigned InvalidCycle = ~0U;
1543 SchedBoundary::~SchedBoundary() { delete HazardRec; }
1545 void SchedBoundary::reset() {
1546 // A new HazardRec is created for each DAG and owned by SchedBoundary.
1547 // Destroying and reconstructing it is very expensive though. So keep
1548 // invalid, placeholder HazardRecs.
1549 if (HazardRec && HazardRec->isEnabled()) {
1551 HazardRec = nullptr;
1555 CheckPending = false;
1559 MinReadyCycle = UINT_MAX;
1560 ExpectedLatency = 0;
1561 DependentLatency = 0;
1563 MaxExecutedResCount = 0;
1565 IsResourceLimited = false;
1566 ReservedCycles.clear();
1568 // Track the maximum number of stall cycles that could arise either from the
1569 // latency of a DAG edge or the number of cycles that a processor resource is
1570 // reserved (SchedBoundary::ReservedCycles).
1571 MaxObservedLatency = 0;
1573 // Reserve a zero-count for invalid CritResIdx.
1574 ExecutedResCounts.resize(1);
1575 assert(!ExecutedResCounts[0] && "nonzero count for bad resource");
1578 void SchedRemainder::
1579 init(ScheduleDAGMI *DAG, const TargetSchedModel *SchedModel) {
1581 if (!SchedModel->hasInstrSchedModel())
1583 RemainingCounts.resize(SchedModel->getNumProcResourceKinds());
1584 for (std::vector<SUnit>::iterator
1585 I = DAG->SUnits.begin(), E = DAG->SUnits.end(); I != E; ++I) {
1586 const MCSchedClassDesc *SC = DAG->getSchedClass(&*I);
1587 RemIssueCount += SchedModel->getNumMicroOps(I->getInstr(), SC)
1588 * SchedModel->getMicroOpFactor();
1589 for (TargetSchedModel::ProcResIter
1590 PI = SchedModel->getWriteProcResBegin(SC),
1591 PE = SchedModel->getWriteProcResEnd(SC); PI != PE; ++PI) {
1592 unsigned PIdx = PI->ProcResourceIdx;
1593 unsigned Factor = SchedModel->getResourceFactor(PIdx);
1594 RemainingCounts[PIdx] += (Factor * PI->Cycles);
1599 void SchedBoundary::
1600 init(ScheduleDAGMI *dag, const TargetSchedModel *smodel, SchedRemainder *rem) {
1603 SchedModel = smodel;
1605 if (SchedModel->hasInstrSchedModel()) {
1606 ExecutedResCounts.resize(SchedModel->getNumProcResourceKinds());
1607 ReservedCycles.resize(SchedModel->getNumProcResourceKinds(), InvalidCycle);
1611 /// Compute the stall cycles based on this SUnit's ready time. Heuristics treat
1612 /// these "soft stalls" differently than the hard stall cycles based on CPU
1613 /// resources and computed by checkHazard(). A fully in-order model
1614 /// (MicroOpBufferSize==0) will not make use of this since instructions are not
1615 /// available for scheduling until they are ready. However, a weaker in-order
1616 /// model may use this for heuristics. For example, if a processor has in-order
1617 /// behavior when reading certain resources, this may come into play.
1618 unsigned SchedBoundary::getLatencyStallCycles(SUnit *SU) {
1619 if (!SU->isUnbuffered)
1622 unsigned ReadyCycle = (isTop() ? SU->TopReadyCycle : SU->BotReadyCycle);
1623 if (ReadyCycle > CurrCycle)
1624 return ReadyCycle - CurrCycle;
1628 /// Compute the next cycle at which the given processor resource can be
1630 unsigned SchedBoundary::
1631 getNextResourceCycle(unsigned PIdx, unsigned Cycles) {
1632 unsigned NextUnreserved = ReservedCycles[PIdx];
1633 // If this resource has never been used, always return cycle zero.
1634 if (NextUnreserved == InvalidCycle)
1636 // For bottom-up scheduling add the cycles needed for the current operation.
1638 NextUnreserved += Cycles;
1639 return NextUnreserved;
1642 /// Does this SU have a hazard within the current instruction group.
1644 /// The scheduler supports two modes of hazard recognition. The first is the
1645 /// ScheduleHazardRecognizer API. It is a fully general hazard recognizer that
1646 /// supports highly complicated in-order reservation tables
1647 /// (ScoreboardHazardRecognizer) and arbitraty target-specific logic.
1649 /// The second is a streamlined mechanism that checks for hazards based on
1650 /// simple counters that the scheduler itself maintains. It explicitly checks
1651 /// for instruction dispatch limitations, including the number of micro-ops that
1652 /// can dispatch per cycle.
1654 /// TODO: Also check whether the SU must start a new group.
1655 bool SchedBoundary::checkHazard(SUnit *SU) {
1656 if (HazardRec->isEnabled()
1657 && HazardRec->getHazardType(SU) != ScheduleHazardRecognizer::NoHazard) {
1660 unsigned uops = SchedModel->getNumMicroOps(SU->getInstr());
1661 if ((CurrMOps > 0) && (CurrMOps + uops > SchedModel->getIssueWidth())) {
1662 DEBUG(dbgs() << " SU(" << SU->NodeNum << ") uops="
1663 << SchedModel->getNumMicroOps(SU->getInstr()) << '\n');
1666 if (SchedModel->hasInstrSchedModel() && SU->hasReservedResource) {
1667 const MCSchedClassDesc *SC = DAG->getSchedClass(SU);
1668 for (TargetSchedModel::ProcResIter
1669 PI = SchedModel->getWriteProcResBegin(SC),
1670 PE = SchedModel->getWriteProcResEnd(SC); PI != PE; ++PI) {
1671 if (getNextResourceCycle(PI->ProcResourceIdx, PI->Cycles) > CurrCycle)
1678 // Find the unscheduled node in ReadySUs with the highest latency.
1679 unsigned SchedBoundary::
1680 findMaxLatency(ArrayRef<SUnit*> ReadySUs) {
1681 SUnit *LateSU = nullptr;
1682 unsigned RemLatency = 0;
1683 for (ArrayRef<SUnit*>::iterator I = ReadySUs.begin(), E = ReadySUs.end();
1685 unsigned L = getUnscheduledLatency(*I);
1686 if (L > RemLatency) {
1692 DEBUG(dbgs() << Available.getName() << " RemLatency SU("
1693 << LateSU->NodeNum << ") " << RemLatency << "c\n");
1698 // Count resources in this zone and the remaining unscheduled
1699 // instruction. Return the max count, scaled. Set OtherCritIdx to the critical
1700 // resource index, or zero if the zone is issue limited.
1701 unsigned SchedBoundary::
1702 getOtherResourceCount(unsigned &OtherCritIdx) {
1704 if (!SchedModel->hasInstrSchedModel())
1707 unsigned OtherCritCount = Rem->RemIssueCount
1708 + (RetiredMOps * SchedModel->getMicroOpFactor());
1709 DEBUG(dbgs() << " " << Available.getName() << " + Remain MOps: "
1710 << OtherCritCount / SchedModel->getMicroOpFactor() << '\n');
1711 for (unsigned PIdx = 1, PEnd = SchedModel->getNumProcResourceKinds();
1712 PIdx != PEnd; ++PIdx) {
1713 unsigned OtherCount = getResourceCount(PIdx) + Rem->RemainingCounts[PIdx];
1714 if (OtherCount > OtherCritCount) {
1715 OtherCritCount = OtherCount;
1716 OtherCritIdx = PIdx;
1720 DEBUG(dbgs() << " " << Available.getName() << " + Remain CritRes: "
1721 << OtherCritCount / SchedModel->getResourceFactor(OtherCritIdx)
1722 << " " << SchedModel->getResourceName(OtherCritIdx) << "\n");
1724 return OtherCritCount;
1727 void SchedBoundary::releaseNode(SUnit *SU, unsigned ReadyCycle) {
1728 if (ReadyCycle < MinReadyCycle)
1729 MinReadyCycle = ReadyCycle;
1731 // Check for interlocks first. For the purpose of other heuristics, an
1732 // instruction that cannot issue appears as if it's not in the ReadyQueue.
1733 bool IsBuffered = SchedModel->getMicroOpBufferSize() != 0;
1734 if ((!IsBuffered && ReadyCycle > CurrCycle) || checkHazard(SU))
1739 // Record this node as an immediate dependent of the scheduled node.
1743 void SchedBoundary::releaseTopNode(SUnit *SU) {
1744 if (SU->isScheduled)
1747 for (SUnit::pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
1751 unsigned PredReadyCycle = I->getSUnit()->TopReadyCycle;
1752 unsigned Latency = I->getLatency();
1754 MaxObservedLatency = std::max(Latency, MaxObservedLatency);
1756 if (SU->TopReadyCycle < PredReadyCycle + Latency)
1757 SU->TopReadyCycle = PredReadyCycle + Latency;
1759 releaseNode(SU, SU->TopReadyCycle);
1762 void SchedBoundary::releaseBottomNode(SUnit *SU) {
1763 if (SU->isScheduled)
1766 assert(SU->getInstr() && "Scheduled SUnit must have instr");
1768 for (SUnit::succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
1772 unsigned SuccReadyCycle = I->getSUnit()->BotReadyCycle;
1773 unsigned Latency = I->getLatency();
1775 MaxObservedLatency = std::max(Latency, MaxObservedLatency);
1777 if (SU->BotReadyCycle < SuccReadyCycle + Latency)
1778 SU->BotReadyCycle = SuccReadyCycle + Latency;
1780 releaseNode(SU, SU->BotReadyCycle);
1783 /// Move the boundary of scheduled code by one cycle.
1784 void SchedBoundary::bumpCycle(unsigned NextCycle) {
1785 if (SchedModel->getMicroOpBufferSize() == 0) {
1786 assert(MinReadyCycle < UINT_MAX && "MinReadyCycle uninitialized");
1787 if (MinReadyCycle > NextCycle)
1788 NextCycle = MinReadyCycle;
1790 // Update the current micro-ops, which will issue in the next cycle.
1791 unsigned DecMOps = SchedModel->getIssueWidth() * (NextCycle - CurrCycle);
1792 CurrMOps = (CurrMOps <= DecMOps) ? 0 : CurrMOps - DecMOps;
1794 // Decrement DependentLatency based on the next cycle.
1795 if ((NextCycle - CurrCycle) > DependentLatency)
1796 DependentLatency = 0;
1798 DependentLatency -= (NextCycle - CurrCycle);
1800 if (!HazardRec->isEnabled()) {
1801 // Bypass HazardRec virtual calls.
1802 CurrCycle = NextCycle;
1805 // Bypass getHazardType calls in case of long latency.
1806 for (; CurrCycle != NextCycle; ++CurrCycle) {
1808 HazardRec->AdvanceCycle();
1810 HazardRec->RecedeCycle();
1813 CheckPending = true;
1814 unsigned LFactor = SchedModel->getLatencyFactor();
1816 (int)(getCriticalCount() - (getScheduledLatency() * LFactor))
1819 DEBUG(dbgs() << "Cycle: " << CurrCycle << ' ' << Available.getName() << '\n');
1822 void SchedBoundary::incExecutedResources(unsigned PIdx, unsigned Count) {
1823 ExecutedResCounts[PIdx] += Count;
1824 if (ExecutedResCounts[PIdx] > MaxExecutedResCount)
1825 MaxExecutedResCount = ExecutedResCounts[PIdx];
1828 /// Add the given processor resource to this scheduled zone.
1830 /// \param Cycles indicates the number of consecutive (non-pipelined) cycles
1831 /// during which this resource is consumed.
1833 /// \return the next cycle at which the instruction may execute without
1834 /// oversubscribing resources.
1835 unsigned SchedBoundary::
1836 countResource(unsigned PIdx, unsigned Cycles, unsigned NextCycle) {
1837 unsigned Factor = SchedModel->getResourceFactor(PIdx);
1838 unsigned Count = Factor * Cycles;
1839 DEBUG(dbgs() << " " << SchedModel->getResourceName(PIdx)
1840 << " +" << Cycles << "x" << Factor << "u\n");
1842 // Update Executed resources counts.
1843 incExecutedResources(PIdx, Count);
1844 assert(Rem->RemainingCounts[PIdx] >= Count && "resource double counted");
1845 Rem->RemainingCounts[PIdx] -= Count;
1847 // Check if this resource exceeds the current critical resource. If so, it
1848 // becomes the critical resource.
1849 if (ZoneCritResIdx != PIdx && (getResourceCount(PIdx) > getCriticalCount())) {
1850 ZoneCritResIdx = PIdx;
1851 DEBUG(dbgs() << " *** Critical resource "
1852 << SchedModel->getResourceName(PIdx) << ": "
1853 << getResourceCount(PIdx) / SchedModel->getLatencyFactor() << "c\n");
1855 // For reserved resources, record the highest cycle using the resource.
1856 unsigned NextAvailable = getNextResourceCycle(PIdx, Cycles);
1857 if (NextAvailable > CurrCycle) {
1858 DEBUG(dbgs() << " Resource conflict: "
1859 << SchedModel->getProcResource(PIdx)->Name << " reserved until @"
1860 << NextAvailable << "\n");
1862 return NextAvailable;
1865 /// Move the boundary of scheduled code by one SUnit.
1866 void SchedBoundary::bumpNode(SUnit *SU) {
1867 // Update the reservation table.
1868 if (HazardRec->isEnabled()) {
1869 if (!isTop() && SU->isCall) {
1870 // Calls are scheduled with their preceding instructions. For bottom-up
1871 // scheduling, clear the pipeline state before emitting.
1874 HazardRec->EmitInstruction(SU);
1876 // checkHazard should prevent scheduling multiple instructions per cycle that
1877 // exceed the issue width.
1878 const MCSchedClassDesc *SC = DAG->getSchedClass(SU);
1879 unsigned IncMOps = SchedModel->getNumMicroOps(SU->getInstr());
1881 (CurrMOps == 0 || (CurrMOps + IncMOps) <= SchedModel->getIssueWidth()) &&
1882 "Cannot schedule this instruction's MicroOps in the current cycle.");
1884 unsigned ReadyCycle = (isTop() ? SU->TopReadyCycle : SU->BotReadyCycle);
1885 DEBUG(dbgs() << " Ready @" << ReadyCycle << "c\n");
1887 unsigned NextCycle = CurrCycle;
1888 switch (SchedModel->getMicroOpBufferSize()) {
1890 assert(ReadyCycle <= CurrCycle && "Broken PendingQueue");
1893 if (ReadyCycle > NextCycle) {
1894 NextCycle = ReadyCycle;
1895 DEBUG(dbgs() << " *** Stall until: " << ReadyCycle << "\n");
1899 // We don't currently model the OOO reorder buffer, so consider all
1900 // scheduled MOps to be "retired". We do loosely model in-order resource
1901 // latency. If this instruction uses an in-order resource, account for any
1902 // likely stall cycles.
1903 if (SU->isUnbuffered && ReadyCycle > NextCycle)
1904 NextCycle = ReadyCycle;
1907 RetiredMOps += IncMOps;
1909 // Update resource counts and critical resource.
1910 if (SchedModel->hasInstrSchedModel()) {
1911 unsigned DecRemIssue = IncMOps * SchedModel->getMicroOpFactor();
1912 assert(Rem->RemIssueCount >= DecRemIssue && "MOps double counted");
1913 Rem->RemIssueCount -= DecRemIssue;
1914 if (ZoneCritResIdx) {
1915 // Scale scheduled micro-ops for comparing with the critical resource.
1916 unsigned ScaledMOps =
1917 RetiredMOps * SchedModel->getMicroOpFactor();
1919 // If scaled micro-ops are now more than the previous critical resource by
1920 // a full cycle, then micro-ops issue becomes critical.
1921 if ((int)(ScaledMOps - getResourceCount(ZoneCritResIdx))
1922 >= (int)SchedModel->getLatencyFactor()) {
1924 DEBUG(dbgs() << " *** Critical resource NumMicroOps: "
1925 << ScaledMOps / SchedModel->getLatencyFactor() << "c\n");
1928 for (TargetSchedModel::ProcResIter
1929 PI = SchedModel->getWriteProcResBegin(SC),
1930 PE = SchedModel->getWriteProcResEnd(SC); PI != PE; ++PI) {
1932 countResource(PI->ProcResourceIdx, PI->Cycles, NextCycle);
1933 if (RCycle > NextCycle)
1936 if (SU->hasReservedResource) {
1937 // For reserved resources, record the highest cycle using the resource.
1938 // For top-down scheduling, this is the cycle in which we schedule this
1939 // instruction plus the number of cycles the operations reserves the
1940 // resource. For bottom-up is it simply the instruction's cycle.
1941 for (TargetSchedModel::ProcResIter
1942 PI = SchedModel->getWriteProcResBegin(SC),
1943 PE = SchedModel->getWriteProcResEnd(SC); PI != PE; ++PI) {
1944 unsigned PIdx = PI->ProcResourceIdx;
1945 if (SchedModel->getProcResource(PIdx)->BufferSize == 0) {
1946 ReservedCycles[PIdx] = isTop() ? NextCycle + PI->Cycles : NextCycle;
1948 MaxObservedLatency = std::max(PI->Cycles, MaxObservedLatency);
1954 // Update ExpectedLatency and DependentLatency.
1955 unsigned &TopLatency = isTop() ? ExpectedLatency : DependentLatency;
1956 unsigned &BotLatency = isTop() ? DependentLatency : ExpectedLatency;
1957 if (SU->getDepth() > TopLatency) {
1958 TopLatency = SU->getDepth();
1959 DEBUG(dbgs() << " " << Available.getName()
1960 << " TopLatency SU(" << SU->NodeNum << ") " << TopLatency << "c\n");
1962 if (SU->getHeight() > BotLatency) {
1963 BotLatency = SU->getHeight();
1964 DEBUG(dbgs() << " " << Available.getName()
1965 << " BotLatency SU(" << SU->NodeNum << ") " << BotLatency << "c\n");
1967 // If we stall for any reason, bump the cycle.
1968 if (NextCycle > CurrCycle) {
1969 bumpCycle(NextCycle);
1972 // After updating ZoneCritResIdx and ExpectedLatency, check if we're
1973 // resource limited. If a stall occurred, bumpCycle does this.
1974 unsigned LFactor = SchedModel->getLatencyFactor();
1976 (int)(getCriticalCount() - (getScheduledLatency() * LFactor))
1979 // Update CurrMOps after calling bumpCycle to handle stalls, since bumpCycle
1980 // resets CurrMOps. Loop to handle instructions with more MOps than issue in
1981 // one cycle. Since we commonly reach the max MOps here, opportunistically
1982 // bump the cycle to avoid uselessly checking everything in the readyQ.
1983 CurrMOps += IncMOps;
1984 while (CurrMOps >= SchedModel->getIssueWidth()) {
1985 DEBUG(dbgs() << " *** Max MOps " << CurrMOps
1986 << " at cycle " << CurrCycle << '\n');
1987 bumpCycle(++NextCycle);
1989 DEBUG(dumpScheduledState());
1992 /// Release pending ready nodes in to the available queue. This makes them
1993 /// visible to heuristics.
1994 void SchedBoundary::releasePending() {
1995 // If the available queue is empty, it is safe to reset MinReadyCycle.
1996 if (Available.empty())
1997 MinReadyCycle = UINT_MAX;
1999 // Check to see if any of the pending instructions are ready to issue. If
2000 // so, add them to the available queue.
2001 bool IsBuffered = SchedModel->getMicroOpBufferSize() != 0;
2002 for (unsigned i = 0, e = Pending.size(); i != e; ++i) {
2003 SUnit *SU = *(Pending.begin()+i);
2004 unsigned ReadyCycle = isTop() ? SU->TopReadyCycle : SU->BotReadyCycle;
2006 if (ReadyCycle < MinReadyCycle)
2007 MinReadyCycle = ReadyCycle;
2009 if (!IsBuffered && ReadyCycle > CurrCycle)
2012 if (checkHazard(SU))
2016 Pending.remove(Pending.begin()+i);
2019 DEBUG(if (!Pending.empty()) Pending.dump());
2020 CheckPending = false;
2023 /// Remove SU from the ready set for this boundary.
2024 void SchedBoundary::removeReady(SUnit *SU) {
2025 if (Available.isInQueue(SU))
2026 Available.remove(Available.find(SU));
2028 assert(Pending.isInQueue(SU) && "bad ready count");
2029 Pending.remove(Pending.find(SU));
2033 /// If this queue only has one ready candidate, return it. As a side effect,
2034 /// defer any nodes that now hit a hazard, and advance the cycle until at least
2035 /// one node is ready. If multiple instructions are ready, return NULL.
2036 SUnit *SchedBoundary::pickOnlyChoice() {
2041 // Defer any ready instrs that now have a hazard.
2042 for (ReadyQueue::iterator I = Available.begin(); I != Available.end();) {
2043 if (checkHazard(*I)) {
2045 I = Available.remove(I);
2051 for (unsigned i = 0; Available.empty(); ++i) {
2052 assert(i <= (HazardRec->getMaxLookAhead() + MaxObservedLatency) &&
2053 "permanent hazard"); (void)i;
2054 bumpCycle(CurrCycle + 1);
2057 if (Available.size() == 1)
2058 return *Available.begin();
2063 // This is useful information to dump after bumpNode.
2064 // Note that the Queue contents are more useful before pickNodeFromQueue.
2065 void SchedBoundary::dumpScheduledState() {
2068 if (ZoneCritResIdx) {
2069 ResFactor = SchedModel->getResourceFactor(ZoneCritResIdx);
2070 ResCount = getResourceCount(ZoneCritResIdx);
2073 ResFactor = SchedModel->getMicroOpFactor();
2074 ResCount = RetiredMOps * SchedModel->getMicroOpFactor();
2076 unsigned LFactor = SchedModel->getLatencyFactor();
2077 dbgs() << Available.getName() << " @" << CurrCycle << "c\n"
2078 << " Retired: " << RetiredMOps;
2079 dbgs() << "\n Executed: " << getExecutedCount() / LFactor << "c";
2080 dbgs() << "\n Critical: " << ResCount / LFactor << "c, "
2081 << ResCount / ResFactor << " "
2082 << SchedModel->getResourceName(ZoneCritResIdx)
2083 << "\n ExpectedLatency: " << ExpectedLatency << "c\n"
2084 << (IsResourceLimited ? " - Resource" : " - Latency")
2089 //===----------------------------------------------------------------------===//
2090 // GenericScheduler - Generic implementation of MachineSchedStrategy.
2091 //===----------------------------------------------------------------------===//
2093 void GenericSchedulerBase::SchedCandidate::
2094 initResourceDelta(const ScheduleDAGMI *DAG,
2095 const TargetSchedModel *SchedModel) {
2096 if (!Policy.ReduceResIdx && !Policy.DemandResIdx)
2099 const MCSchedClassDesc *SC = DAG->getSchedClass(SU);
2100 for (TargetSchedModel::ProcResIter
2101 PI = SchedModel->getWriteProcResBegin(SC),
2102 PE = SchedModel->getWriteProcResEnd(SC); PI != PE; ++PI) {
2103 if (PI->ProcResourceIdx == Policy.ReduceResIdx)
2104 ResDelta.CritResources += PI->Cycles;
2105 if (PI->ProcResourceIdx == Policy.DemandResIdx)
2106 ResDelta.DemandedResources += PI->Cycles;
2110 /// Set the CandPolicy given a scheduling zone given the current resources and
2111 /// latencies inside and outside the zone.
2112 void GenericSchedulerBase::setPolicy(CandPolicy &Policy,
2114 SchedBoundary &CurrZone,
2115 SchedBoundary *OtherZone) {
2116 // Apply preemptive heuristics based on the the total latency and resources
2117 // inside and outside this zone. Potential stalls should be considered before
2118 // following this policy.
2120 // Compute remaining latency. We need this both to determine whether the
2121 // overall schedule has become latency-limited and whether the instructions
2122 // outside this zone are resource or latency limited.
2124 // The "dependent" latency is updated incrementally during scheduling as the
2125 // max height/depth of scheduled nodes minus the cycles since it was
2127 // DLat = max (N.depth - (CurrCycle - N.ReadyCycle) for N in Zone
2129 // The "independent" latency is the max ready queue depth:
2130 // ILat = max N.depth for N in Available|Pending
2132 // RemainingLatency is the greater of independent and dependent latency.
2133 unsigned RemLatency = CurrZone.getDependentLatency();
2134 RemLatency = std::max(RemLatency,
2135 CurrZone.findMaxLatency(CurrZone.Available.elements()));
2136 RemLatency = std::max(RemLatency,
2137 CurrZone.findMaxLatency(CurrZone.Pending.elements()));
2139 // Compute the critical resource outside the zone.
2140 unsigned OtherCritIdx = 0;
2141 unsigned OtherCount =
2142 OtherZone ? OtherZone->getOtherResourceCount(OtherCritIdx) : 0;
2144 bool OtherResLimited = false;
2145 if (SchedModel->hasInstrSchedModel()) {
2146 unsigned LFactor = SchedModel->getLatencyFactor();
2147 OtherResLimited = (int)(OtherCount - (RemLatency * LFactor)) > (int)LFactor;
2149 // Schedule aggressively for latency in PostRA mode. We don't check for
2150 // acyclic latency during PostRA, and highly out-of-order processors will
2151 // skip PostRA scheduling.
2152 if (!OtherResLimited) {
2153 if (IsPostRA || (RemLatency + CurrZone.getCurrCycle() > Rem.CriticalPath)) {
2154 Policy.ReduceLatency |= true;
2155 DEBUG(dbgs() << " " << CurrZone.Available.getName()
2156 << " RemainingLatency " << RemLatency << " + "
2157 << CurrZone.getCurrCycle() << "c > CritPath "
2158 << Rem.CriticalPath << "\n");
2161 // If the same resource is limiting inside and outside the zone, do nothing.
2162 if (CurrZone.getZoneCritResIdx() == OtherCritIdx)
2166 if (CurrZone.isResourceLimited()) {
2167 dbgs() << " " << CurrZone.Available.getName() << " ResourceLimited: "
2168 << SchedModel->getResourceName(CurrZone.getZoneCritResIdx())
2171 if (OtherResLimited)
2172 dbgs() << " RemainingLimit: "
2173 << SchedModel->getResourceName(OtherCritIdx) << "\n";
2174 if (!CurrZone.isResourceLimited() && !OtherResLimited)
2175 dbgs() << " Latency limited both directions.\n");
2177 if (CurrZone.isResourceLimited() && !Policy.ReduceResIdx)
2178 Policy.ReduceResIdx = CurrZone.getZoneCritResIdx();
2180 if (OtherResLimited)
2181 Policy.DemandResIdx = OtherCritIdx;
2185 const char *GenericSchedulerBase::getReasonStr(
2186 GenericSchedulerBase::CandReason Reason) {
2188 case NoCand: return "NOCAND ";
2189 case PhysRegCopy: return "PREG-COPY";
2190 case RegExcess: return "REG-EXCESS";
2191 case RegCritical: return "REG-CRIT ";
2192 case Stall: return "STALL ";
2193 case Cluster: return "CLUSTER ";
2194 case Weak: return "WEAK ";
2195 case RegMax: return "REG-MAX ";
2196 case ResourceReduce: return "RES-REDUCE";
2197 case ResourceDemand: return "RES-DEMAND";
2198 case TopDepthReduce: return "TOP-DEPTH ";
2199 case TopPathReduce: return "TOP-PATH ";
2200 case BotHeightReduce:return "BOT-HEIGHT";
2201 case BotPathReduce: return "BOT-PATH ";
2202 case NextDefUse: return "DEF-USE ";
2203 case NodeOrder: return "ORDER ";
2205 llvm_unreachable("Unknown reason!");
2208 void GenericSchedulerBase::traceCandidate(const SchedCandidate &Cand) {
2210 unsigned ResIdx = 0;
2211 unsigned Latency = 0;
2212 switch (Cand.Reason) {
2216 P = Cand.RPDelta.Excess;
2219 P = Cand.RPDelta.CriticalMax;
2222 P = Cand.RPDelta.CurrentMax;
2224 case ResourceReduce:
2225 ResIdx = Cand.Policy.ReduceResIdx;
2227 case ResourceDemand:
2228 ResIdx = Cand.Policy.DemandResIdx;
2230 case TopDepthReduce:
2231 Latency = Cand.SU->getDepth();
2234 Latency = Cand.SU->getHeight();
2236 case BotHeightReduce:
2237 Latency = Cand.SU->getHeight();
2240 Latency = Cand.SU->getDepth();
2243 dbgs() << " SU(" << Cand.SU->NodeNum << ") " << getReasonStr(Cand.Reason);
2245 dbgs() << " " << TRI->getRegPressureSetName(P.getPSet())
2246 << ":" << P.getUnitInc() << " ";
2250 dbgs() << " " << SchedModel->getProcResource(ResIdx)->Name << " ";
2254 dbgs() << " " << Latency << " cycles ";
2261 /// Return true if this heuristic determines order.
2262 static bool tryLess(int TryVal, int CandVal,
2263 GenericSchedulerBase::SchedCandidate &TryCand,
2264 GenericSchedulerBase::SchedCandidate &Cand,
2265 GenericSchedulerBase::CandReason Reason) {
2266 if (TryVal < CandVal) {
2267 TryCand.Reason = Reason;
2270 if (TryVal > CandVal) {
2271 if (Cand.Reason > Reason)
2272 Cand.Reason = Reason;
2275 Cand.setRepeat(Reason);
2279 static bool tryGreater(int TryVal, int CandVal,
2280 GenericSchedulerBase::SchedCandidate &TryCand,
2281 GenericSchedulerBase::SchedCandidate &Cand,
2282 GenericSchedulerBase::CandReason Reason) {
2283 if (TryVal > CandVal) {
2284 TryCand.Reason = Reason;
2287 if (TryVal < CandVal) {
2288 if (Cand.Reason > Reason)
2289 Cand.Reason = Reason;
2292 Cand.setRepeat(Reason);
2296 static bool tryLatency(GenericSchedulerBase::SchedCandidate &TryCand,
2297 GenericSchedulerBase::SchedCandidate &Cand,
2298 SchedBoundary &Zone) {
2300 if (Cand.SU->getDepth() > Zone.getScheduledLatency()) {
2301 if (tryLess(TryCand.SU->getDepth(), Cand.SU->getDepth(),
2302 TryCand, Cand, GenericSchedulerBase::TopDepthReduce))
2305 if (tryGreater(TryCand.SU->getHeight(), Cand.SU->getHeight(),
2306 TryCand, Cand, GenericSchedulerBase::TopPathReduce))
2310 if (Cand.SU->getHeight() > Zone.getScheduledLatency()) {
2311 if (tryLess(TryCand.SU->getHeight(), Cand.SU->getHeight(),
2312 TryCand, Cand, GenericSchedulerBase::BotHeightReduce))
2315 if (tryGreater(TryCand.SU->getDepth(), Cand.SU->getDepth(),
2316 TryCand, Cand, GenericSchedulerBase::BotPathReduce))
2322 static void tracePick(const GenericSchedulerBase::SchedCandidate &Cand,
2324 DEBUG(dbgs() << "Pick " << (IsTop ? "Top " : "Bot ")
2325 << GenericSchedulerBase::getReasonStr(Cand.Reason) << '\n');
2328 void GenericScheduler::initialize(ScheduleDAGMI *dag) {
2329 assert(dag->hasVRegLiveness() &&
2330 "(PreRA)GenericScheduler needs vreg liveness");
2331 DAG = static_cast<ScheduleDAGMILive*>(dag);
2332 SchedModel = DAG->getSchedModel();
2335 Rem.init(DAG, SchedModel);
2336 Top.init(DAG, SchedModel, &Rem);
2337 Bot.init(DAG, SchedModel, &Rem);
2339 // Initialize resource counts.
2341 // Initialize the HazardRecognizers. If itineraries don't exist, are empty, or
2342 // are disabled, then these HazardRecs will be disabled.
2343 const InstrItineraryData *Itin = SchedModel->getInstrItineraries();
2344 const TargetMachine &TM = DAG->MF.getTarget();
2345 if (!Top.HazardRec) {
2347 TM.getInstrInfo()->CreateTargetMIHazardRecognizer(Itin, DAG);
2349 if (!Bot.HazardRec) {
2351 TM.getInstrInfo()->CreateTargetMIHazardRecognizer(Itin, DAG);
2355 /// Initialize the per-region scheduling policy.
2356 void GenericScheduler::initPolicy(MachineBasicBlock::iterator Begin,
2357 MachineBasicBlock::iterator End,
2358 unsigned NumRegionInstrs) {
2359 const TargetMachine &TM = Context->MF->getTarget();
2360 const TargetLowering *TLI = TM.getTargetLowering();
2362 // Avoid setting up the register pressure tracker for small regions to save
2363 // compile time. As a rough heuristic, only track pressure when the number of
2364 // schedulable instructions exceeds half the integer register file.
2365 RegionPolicy.ShouldTrackPressure = true;
2366 for (unsigned VT = MVT::i32; VT > (unsigned)MVT::i1; --VT) {
2367 MVT::SimpleValueType LegalIntVT = (MVT::SimpleValueType)VT;
2368 if (TLI->isTypeLegal(LegalIntVT)) {
2369 unsigned NIntRegs = Context->RegClassInfo->getNumAllocatableRegs(
2370 TLI->getRegClassFor(LegalIntVT));
2371 RegionPolicy.ShouldTrackPressure = NumRegionInstrs > (NIntRegs / 2);
2375 // For generic targets, we default to bottom-up, because it's simpler and more
2376 // compile-time optimizations have been implemented in that direction.
2377 RegionPolicy.OnlyBottomUp = true;
2379 // Allow the subtarget to override default policy.
2380 const TargetSubtargetInfo &ST = TM.getSubtarget<TargetSubtargetInfo>();
2381 ST.overrideSchedPolicy(RegionPolicy, Begin, End, NumRegionInstrs);
2383 // After subtarget overrides, apply command line options.
2384 if (!EnableRegPressure)
2385 RegionPolicy.ShouldTrackPressure = false;
2387 // Check -misched-topdown/bottomup can force or unforce scheduling direction.
2388 // e.g. -misched-bottomup=false allows scheduling in both directions.
2389 assert((!ForceTopDown || !ForceBottomUp) &&
2390 "-misched-topdown incompatible with -misched-bottomup");
2391 if (ForceBottomUp.getNumOccurrences() > 0) {
2392 RegionPolicy.OnlyBottomUp = ForceBottomUp;
2393 if (RegionPolicy.OnlyBottomUp)
2394 RegionPolicy.OnlyTopDown = false;
2396 if (ForceTopDown.getNumOccurrences() > 0) {
2397 RegionPolicy.OnlyTopDown = ForceTopDown;
2398 if (RegionPolicy.OnlyTopDown)
2399 RegionPolicy.OnlyBottomUp = false;
2403 /// Set IsAcyclicLatencyLimited if the acyclic path is longer than the cyclic
2404 /// critical path by more cycles than it takes to drain the instruction buffer.
2405 /// We estimate an upper bounds on in-flight instructions as:
2407 /// CyclesPerIteration = max( CyclicPath, Loop-Resource-Height )
2408 /// InFlightIterations = AcyclicPath / CyclesPerIteration
2409 /// InFlightResources = InFlightIterations * LoopResources
2411 /// TODO: Check execution resources in addition to IssueCount.
2412 void GenericScheduler::checkAcyclicLatency() {
2413 if (Rem.CyclicCritPath == 0 || Rem.CyclicCritPath >= Rem.CriticalPath)
2416 // Scaled number of cycles per loop iteration.
2417 unsigned IterCount =
2418 std::max(Rem.CyclicCritPath * SchedModel->getLatencyFactor(),
2420 // Scaled acyclic critical path.
2421 unsigned AcyclicCount = Rem.CriticalPath * SchedModel->getLatencyFactor();
2422 // InFlightCount = (AcyclicPath / IterCycles) * InstrPerLoop
2423 unsigned InFlightCount =
2424 (AcyclicCount * Rem.RemIssueCount + IterCount-1) / IterCount;
2425 unsigned BufferLimit =
2426 SchedModel->getMicroOpBufferSize() * SchedModel->getMicroOpFactor();
2428 Rem.IsAcyclicLatencyLimited = InFlightCount > BufferLimit;
2430 DEBUG(dbgs() << "IssueCycles="
2431 << Rem.RemIssueCount / SchedModel->getLatencyFactor() << "c "
2432 << "IterCycles=" << IterCount / SchedModel->getLatencyFactor()
2433 << "c NumIters=" << (AcyclicCount + IterCount-1) / IterCount
2434 << " InFlight=" << InFlightCount / SchedModel->getMicroOpFactor()
2435 << "m BufferLim=" << SchedModel->getMicroOpBufferSize() << "m\n";
2436 if (Rem.IsAcyclicLatencyLimited)
2437 dbgs() << " ACYCLIC LATENCY LIMIT\n");
2440 void GenericScheduler::registerRoots() {
2441 Rem.CriticalPath = DAG->ExitSU.getDepth();
2443 // Some roots may not feed into ExitSU. Check all of them in case.
2444 for (std::vector<SUnit*>::const_iterator
2445 I = Bot.Available.begin(), E = Bot.Available.end(); I != E; ++I) {
2446 if ((*I)->getDepth() > Rem.CriticalPath)
2447 Rem.CriticalPath = (*I)->getDepth();
2449 DEBUG(dbgs() << "Critical Path: " << Rem.CriticalPath << '\n');
2451 if (EnableCyclicPath) {
2452 Rem.CyclicCritPath = DAG->computeCyclicCriticalPath();
2453 checkAcyclicLatency();
2457 static bool tryPressure(const PressureChange &TryP,
2458 const PressureChange &CandP,
2459 GenericSchedulerBase::SchedCandidate &TryCand,
2460 GenericSchedulerBase::SchedCandidate &Cand,
2461 GenericSchedulerBase::CandReason Reason) {
2462 int TryRank = TryP.getPSetOrMax();
2463 int CandRank = CandP.getPSetOrMax();
2464 // If both candidates affect the same set, go with the smallest increase.
2465 if (TryRank == CandRank) {
2466 return tryLess(TryP.getUnitInc(), CandP.getUnitInc(), TryCand, Cand,
2469 // If one candidate decreases and the other increases, go with it.
2470 // Invalid candidates have UnitInc==0.
2471 if (tryLess(TryP.getUnitInc() < 0, CandP.getUnitInc() < 0, TryCand, Cand,
2475 // If the candidates are decreasing pressure, reverse priority.
2476 if (TryP.getUnitInc() < 0)
2477 std::swap(TryRank, CandRank);
2478 return tryGreater(TryRank, CandRank, TryCand, Cand, Reason);
2481 static unsigned getWeakLeft(const SUnit *SU, bool isTop) {
2482 return (isTop) ? SU->WeakPredsLeft : SU->WeakSuccsLeft;
2485 /// Minimize physical register live ranges. Regalloc wants them adjacent to
2486 /// their physreg def/use.
2488 /// FIXME: This is an unnecessary check on the critical path. Most are root/leaf
2489 /// copies which can be prescheduled. The rest (e.g. x86 MUL) could be bundled
2490 /// with the operation that produces or consumes the physreg. We'll do this when
2491 /// regalloc has support for parallel copies.
2492 static int biasPhysRegCopy(const SUnit *SU, bool isTop) {
2493 const MachineInstr *MI = SU->getInstr();
2497 unsigned ScheduledOper = isTop ? 1 : 0;
2498 unsigned UnscheduledOper = isTop ? 0 : 1;
2499 // If we have already scheduled the physreg produce/consumer, immediately
2500 // schedule the copy.
2501 if (TargetRegisterInfo::isPhysicalRegister(
2502 MI->getOperand(ScheduledOper).getReg()))
2504 // If the physreg is at the boundary, defer it. Otherwise schedule it
2505 // immediately to free the dependent. We can hoist the copy later.
2506 bool AtBoundary = isTop ? !SU->NumSuccsLeft : !SU->NumPredsLeft;
2507 if (TargetRegisterInfo::isPhysicalRegister(
2508 MI->getOperand(UnscheduledOper).getReg()))
2509 return AtBoundary ? -1 : 1;
2513 /// Apply a set of heursitics to a new candidate. Heuristics are currently
2514 /// hierarchical. This may be more efficient than a graduated cost model because
2515 /// we don't need to evaluate all aspects of the model for each node in the
2516 /// queue. But it's really done to make the heuristics easier to debug and
2517 /// statistically analyze.
2519 /// \param Cand provides the policy and current best candidate.
2520 /// \param TryCand refers to the next SUnit candidate, otherwise uninitialized.
2521 /// \param Zone describes the scheduled zone that we are extending.
2522 /// \param RPTracker describes reg pressure within the scheduled zone.
2523 /// \param TempTracker is a scratch pressure tracker to reuse in queries.
2524 void GenericScheduler::tryCandidate(SchedCandidate &Cand,
2525 SchedCandidate &TryCand,
2526 SchedBoundary &Zone,
2527 const RegPressureTracker &RPTracker,
2528 RegPressureTracker &TempTracker) {
2530 if (DAG->isTrackingPressure()) {
2531 // Always initialize TryCand's RPDelta.
2533 TempTracker.getMaxDownwardPressureDelta(
2534 TryCand.SU->getInstr(),
2536 DAG->getRegionCriticalPSets(),
2537 DAG->getRegPressure().MaxSetPressure);
2540 if (VerifyScheduling) {
2541 TempTracker.getMaxUpwardPressureDelta(
2542 TryCand.SU->getInstr(),
2543 &DAG->getPressureDiff(TryCand.SU),
2545 DAG->getRegionCriticalPSets(),
2546 DAG->getRegPressure().MaxSetPressure);
2549 RPTracker.getUpwardPressureDelta(
2550 TryCand.SU->getInstr(),
2551 DAG->getPressureDiff(TryCand.SU),
2553 DAG->getRegionCriticalPSets(),
2554 DAG->getRegPressure().MaxSetPressure);
2558 DEBUG(if (TryCand.RPDelta.Excess.isValid())
2559 dbgs() << " SU(" << TryCand.SU->NodeNum << ") "
2560 << TRI->getRegPressureSetName(TryCand.RPDelta.Excess.getPSet())
2561 << ":" << TryCand.RPDelta.Excess.getUnitInc() << "\n");
2563 // Initialize the candidate if needed.
2564 if (!Cand.isValid()) {
2565 TryCand.Reason = NodeOrder;
2569 if (tryGreater(biasPhysRegCopy(TryCand.SU, Zone.isTop()),
2570 biasPhysRegCopy(Cand.SU, Zone.isTop()),
2571 TryCand, Cand, PhysRegCopy))
2574 // Avoid exceeding the target's limit. If signed PSetID is negative, it is
2575 // invalid; convert it to INT_MAX to give it lowest priority.
2576 if (DAG->isTrackingPressure() && tryPressure(TryCand.RPDelta.Excess,
2577 Cand.RPDelta.Excess,
2578 TryCand, Cand, RegExcess))
2581 // Avoid increasing the max critical pressure in the scheduled region.
2582 if (DAG->isTrackingPressure() && tryPressure(TryCand.RPDelta.CriticalMax,
2583 Cand.RPDelta.CriticalMax,
2584 TryCand, Cand, RegCritical))
2587 // For loops that are acyclic path limited, aggressively schedule for latency.
2588 // This can result in very long dependence chains scheduled in sequence, so
2589 // once every cycle (when CurrMOps == 0), switch to normal heuristics.
2590 if (Rem.IsAcyclicLatencyLimited && !Zone.getCurrMOps()
2591 && tryLatency(TryCand, Cand, Zone))
2594 // Prioritize instructions that read unbuffered resources by stall cycles.
2595 if (tryLess(Zone.getLatencyStallCycles(TryCand.SU),
2596 Zone.getLatencyStallCycles(Cand.SU), TryCand, Cand, Stall))
2599 // Keep clustered nodes together to encourage downstream peephole
2600 // optimizations which may reduce resource requirements.
2602 // This is a best effort to set things up for a post-RA pass. Optimizations
2603 // like generating loads of multiple registers should ideally be done within
2604 // the scheduler pass by combining the loads during DAG postprocessing.
2605 const SUnit *NextClusterSU =
2606 Zone.isTop() ? DAG->getNextClusterSucc() : DAG->getNextClusterPred();
2607 if (tryGreater(TryCand.SU == NextClusterSU, Cand.SU == NextClusterSU,
2608 TryCand, Cand, Cluster))
2611 // Weak edges are for clustering and other constraints.
2612 if (tryLess(getWeakLeft(TryCand.SU, Zone.isTop()),
2613 getWeakLeft(Cand.SU, Zone.isTop()),
2614 TryCand, Cand, Weak)) {
2617 // Avoid increasing the max pressure of the entire region.
2618 if (DAG->isTrackingPressure() && tryPressure(TryCand.RPDelta.CurrentMax,
2619 Cand.RPDelta.CurrentMax,
2620 TryCand, Cand, RegMax))
2623 // Avoid critical resource consumption and balance the schedule.
2624 TryCand.initResourceDelta(DAG, SchedModel);
2625 if (tryLess(TryCand.ResDelta.CritResources, Cand.ResDelta.CritResources,
2626 TryCand, Cand, ResourceReduce))
2628 if (tryGreater(TryCand.ResDelta.DemandedResources,
2629 Cand.ResDelta.DemandedResources,
2630 TryCand, Cand, ResourceDemand))
2633 // Avoid serializing long latency dependence chains.
2634 // For acyclic path limited loops, latency was already checked above.
2635 if (Cand.Policy.ReduceLatency && !Rem.IsAcyclicLatencyLimited
2636 && tryLatency(TryCand, Cand, Zone)) {
2640 // Prefer immediate defs/users of the last scheduled instruction. This is a
2641 // local pressure avoidance strategy that also makes the machine code
2643 if (tryGreater(Zone.isNextSU(TryCand.SU), Zone.isNextSU(Cand.SU),
2644 TryCand, Cand, NextDefUse))
2647 // Fall through to original instruction order.
2648 if ((Zone.isTop() && TryCand.SU->NodeNum < Cand.SU->NodeNum)
2649 || (!Zone.isTop() && TryCand.SU->NodeNum > Cand.SU->NodeNum)) {
2650 TryCand.Reason = NodeOrder;
2654 /// Pick the best candidate from the queue.
2656 /// TODO: getMaxPressureDelta results can be mostly cached for each SUnit during
2657 /// DAG building. To adjust for the current scheduling location we need to
2658 /// maintain the number of vreg uses remaining to be top-scheduled.
2659 void GenericScheduler::pickNodeFromQueue(SchedBoundary &Zone,
2660 const RegPressureTracker &RPTracker,
2661 SchedCandidate &Cand) {
2662 ReadyQueue &Q = Zone.Available;
2666 // getMaxPressureDelta temporarily modifies the tracker.
2667 RegPressureTracker &TempTracker = const_cast<RegPressureTracker&>(RPTracker);
2669 for (ReadyQueue::iterator I = Q.begin(), E = Q.end(); I != E; ++I) {
2671 SchedCandidate TryCand(Cand.Policy);
2673 tryCandidate(Cand, TryCand, Zone, RPTracker, TempTracker);
2674 if (TryCand.Reason != NoCand) {
2675 // Initialize resource delta if needed in case future heuristics query it.
2676 if (TryCand.ResDelta == SchedResourceDelta())
2677 TryCand.initResourceDelta(DAG, SchedModel);
2678 Cand.setBest(TryCand);
2679 DEBUG(traceCandidate(Cand));
2684 /// Pick the best candidate node from either the top or bottom queue.
2685 SUnit *GenericScheduler::pickNodeBidirectional(bool &IsTopNode) {
2686 // Schedule as far as possible in the direction of no choice. This is most
2687 // efficient, but also provides the best heuristics for CriticalPSets.
2688 if (SUnit *SU = Bot.pickOnlyChoice()) {
2690 DEBUG(dbgs() << "Pick Bot NOCAND\n");
2693 if (SUnit *SU = Top.pickOnlyChoice()) {
2695 DEBUG(dbgs() << "Pick Top NOCAND\n");
2698 CandPolicy NoPolicy;
2699 SchedCandidate BotCand(NoPolicy);
2700 SchedCandidate TopCand(NoPolicy);
2701 // Set the bottom-up policy based on the state of the current bottom zone and
2702 // the instructions outside the zone, including the top zone.
2703 setPolicy(BotCand.Policy, /*IsPostRA=*/false, Bot, &Top);
2704 // Set the top-down policy based on the state of the current top zone and
2705 // the instructions outside the zone, including the bottom zone.
2706 setPolicy(TopCand.Policy, /*IsPostRA=*/false, Top, &Bot);
2708 // Prefer bottom scheduling when heuristics are silent.
2709 pickNodeFromQueue(Bot, DAG->getBotRPTracker(), BotCand);
2710 assert(BotCand.Reason != NoCand && "failed to find the first candidate");
2712 // If either Q has a single candidate that provides the least increase in
2713 // Excess pressure, we can immediately schedule from that Q.
2715 // RegionCriticalPSets summarizes the pressure within the scheduled region and
2716 // affects picking from either Q. If scheduling in one direction must
2717 // increase pressure for one of the excess PSets, then schedule in that
2718 // direction first to provide more freedom in the other direction.
2719 if ((BotCand.Reason == RegExcess && !BotCand.isRepeat(RegExcess))
2720 || (BotCand.Reason == RegCritical
2721 && !BotCand.isRepeat(RegCritical)))
2724 tracePick(BotCand, IsTopNode);
2727 // Check if the top Q has a better candidate.
2728 pickNodeFromQueue(Top, DAG->getTopRPTracker(), TopCand);
2729 assert(TopCand.Reason != NoCand && "failed to find the first candidate");
2731 // Choose the queue with the most important (lowest enum) reason.
2732 if (TopCand.Reason < BotCand.Reason) {
2734 tracePick(TopCand, IsTopNode);
2737 // Otherwise prefer the bottom candidate, in node order if all else failed.
2739 tracePick(BotCand, IsTopNode);
2743 /// Pick the best node to balance the schedule. Implements MachineSchedStrategy.
2744 SUnit *GenericScheduler::pickNode(bool &IsTopNode) {
2745 if (DAG->top() == DAG->bottom()) {
2746 assert(Top.Available.empty() && Top.Pending.empty() &&
2747 Bot.Available.empty() && Bot.Pending.empty() && "ReadyQ garbage");
2752 if (RegionPolicy.OnlyTopDown) {
2753 SU = Top.pickOnlyChoice();
2755 CandPolicy NoPolicy;
2756 SchedCandidate TopCand(NoPolicy);
2757 pickNodeFromQueue(Top, DAG->getTopRPTracker(), TopCand);
2758 assert(TopCand.Reason != NoCand && "failed to find a candidate");
2759 tracePick(TopCand, true);
2764 else if (RegionPolicy.OnlyBottomUp) {
2765 SU = Bot.pickOnlyChoice();
2767 CandPolicy NoPolicy;
2768 SchedCandidate BotCand(NoPolicy);
2769 pickNodeFromQueue(Bot, DAG->getBotRPTracker(), BotCand);
2770 assert(BotCand.Reason != NoCand && "failed to find a candidate");
2771 tracePick(BotCand, false);
2777 SU = pickNodeBidirectional(IsTopNode);
2779 } while (SU->isScheduled);
2781 if (SU->isTopReady())
2782 Top.removeReady(SU);
2783 if (SU->isBottomReady())
2784 Bot.removeReady(SU);
2786 DEBUG(dbgs() << "Scheduling SU(" << SU->NodeNum << ") " << *SU->getInstr());
2790 void GenericScheduler::reschedulePhysRegCopies(SUnit *SU, bool isTop) {
2792 MachineBasicBlock::iterator InsertPos = SU->getInstr();
2795 SmallVectorImpl<SDep> &Deps = isTop ? SU->Preds : SU->Succs;
2797 // Find already scheduled copies with a single physreg dependence and move
2798 // them just above the scheduled instruction.
2799 for (SmallVectorImpl<SDep>::iterator I = Deps.begin(), E = Deps.end();
2801 if (I->getKind() != SDep::Data || !TRI->isPhysicalRegister(I->getReg()))
2803 SUnit *DepSU = I->getSUnit();
2804 if (isTop ? DepSU->Succs.size() > 1 : DepSU->Preds.size() > 1)
2806 MachineInstr *Copy = DepSU->getInstr();
2807 if (!Copy->isCopy())
2809 DEBUG(dbgs() << " Rescheduling physreg copy ";
2810 I->getSUnit()->dump(DAG));
2811 DAG->moveInstruction(Copy, InsertPos);
2815 /// Update the scheduler's state after scheduling a node. This is the same node
2816 /// that was just returned by pickNode(). However, ScheduleDAGMILive needs to
2817 /// update it's state based on the current cycle before MachineSchedStrategy
2820 /// FIXME: Eventually, we may bundle physreg copies rather than rescheduling
2821 /// them here. See comments in biasPhysRegCopy.
2822 void GenericScheduler::schedNode(SUnit *SU, bool IsTopNode) {
2824 SU->TopReadyCycle = std::max(SU->TopReadyCycle, Top.getCurrCycle());
2826 if (SU->hasPhysRegUses)
2827 reschedulePhysRegCopies(SU, true);
2830 SU->BotReadyCycle = std::max(SU->BotReadyCycle, Bot.getCurrCycle());
2832 if (SU->hasPhysRegDefs)
2833 reschedulePhysRegCopies(SU, false);
2837 /// Create the standard converging machine scheduler. This will be used as the
2838 /// default scheduler if the target does not set a default.
2839 static ScheduleDAGInstrs *createGenericSchedLive(MachineSchedContext *C) {
2840 ScheduleDAGMILive *DAG = new ScheduleDAGMILive(C, make_unique<GenericScheduler>(C));
2841 // Register DAG post-processors.
2843 // FIXME: extend the mutation API to allow earlier mutations to instantiate
2844 // data and pass it to later mutations. Have a single mutation that gathers
2845 // the interesting nodes in one pass.
2846 DAG->addMutation(make_unique<CopyConstrain>(DAG->TII, DAG->TRI));
2847 if (EnableLoadCluster && DAG->TII->enableClusterLoads())
2848 DAG->addMutation(make_unique<LoadClusterMutation>(DAG->TII, DAG->TRI));
2849 if (EnableMacroFusion)
2850 DAG->addMutation(make_unique<MacroFusion>(DAG->TII));
2854 static MachineSchedRegistry
2855 GenericSchedRegistry("converge", "Standard converging scheduler.",
2856 createGenericSchedLive);
2858 //===----------------------------------------------------------------------===//
2859 // PostGenericScheduler - Generic PostRA implementation of MachineSchedStrategy.
2860 //===----------------------------------------------------------------------===//
2862 void PostGenericScheduler::initialize(ScheduleDAGMI *Dag) {
2864 SchedModel = DAG->getSchedModel();
2867 Rem.init(DAG, SchedModel);
2868 Top.init(DAG, SchedModel, &Rem);
2871 // Initialize the HazardRecognizers. If itineraries don't exist, are empty,
2872 // or are disabled, then these HazardRecs will be disabled.
2873 const InstrItineraryData *Itin = SchedModel->getInstrItineraries();
2874 const TargetMachine &TM = DAG->MF.getTarget();
2875 if (!Top.HazardRec) {
2877 TM.getInstrInfo()->CreateTargetMIHazardRecognizer(Itin, DAG);
2882 void PostGenericScheduler::registerRoots() {
2883 Rem.CriticalPath = DAG->ExitSU.getDepth();
2885 // Some roots may not feed into ExitSU. Check all of them in case.
2886 for (SmallVectorImpl<SUnit*>::const_iterator
2887 I = BotRoots.begin(), E = BotRoots.end(); I != E; ++I) {
2888 if ((*I)->getDepth() > Rem.CriticalPath)
2889 Rem.CriticalPath = (*I)->getDepth();
2891 DEBUG(dbgs() << "Critical Path: " << Rem.CriticalPath << '\n');
2894 /// Apply a set of heursitics to a new candidate for PostRA scheduling.
2896 /// \param Cand provides the policy and current best candidate.
2897 /// \param TryCand refers to the next SUnit candidate, otherwise uninitialized.
2898 void PostGenericScheduler::tryCandidate(SchedCandidate &Cand,
2899 SchedCandidate &TryCand) {
2901 // Initialize the candidate if needed.
2902 if (!Cand.isValid()) {
2903 TryCand.Reason = NodeOrder;
2907 // Prioritize instructions that read unbuffered resources by stall cycles.
2908 if (tryLess(Top.getLatencyStallCycles(TryCand.SU),
2909 Top.getLatencyStallCycles(Cand.SU), TryCand, Cand, Stall))
2912 // Avoid critical resource consumption and balance the schedule.
2913 if (tryLess(TryCand.ResDelta.CritResources, Cand.ResDelta.CritResources,
2914 TryCand, Cand, ResourceReduce))
2916 if (tryGreater(TryCand.ResDelta.DemandedResources,
2917 Cand.ResDelta.DemandedResources,
2918 TryCand, Cand, ResourceDemand))
2921 // Avoid serializing long latency dependence chains.
2922 if (Cand.Policy.ReduceLatency && tryLatency(TryCand, Cand, Top)) {
2926 // Fall through to original instruction order.
2927 if (TryCand.SU->NodeNum < Cand.SU->NodeNum)
2928 TryCand.Reason = NodeOrder;
2931 void PostGenericScheduler::pickNodeFromQueue(SchedCandidate &Cand) {
2932 ReadyQueue &Q = Top.Available;
2936 for (ReadyQueue::iterator I = Q.begin(), E = Q.end(); I != E; ++I) {
2937 SchedCandidate TryCand(Cand.Policy);
2939 TryCand.initResourceDelta(DAG, SchedModel);
2940 tryCandidate(Cand, TryCand);
2941 if (TryCand.Reason != NoCand) {
2942 Cand.setBest(TryCand);
2943 DEBUG(traceCandidate(Cand));
2948 /// Pick the next node to schedule.
2949 SUnit *PostGenericScheduler::pickNode(bool &IsTopNode) {
2950 if (DAG->top() == DAG->bottom()) {
2951 assert(Top.Available.empty() && Top.Pending.empty() && "ReadyQ garbage");
2956 SU = Top.pickOnlyChoice();
2958 CandPolicy NoPolicy;
2959 SchedCandidate TopCand(NoPolicy);
2960 // Set the top-down policy based on the state of the current top zone and
2961 // the instructions outside the zone, including the bottom zone.
2962 setPolicy(TopCand.Policy, /*IsPostRA=*/true, Top, nullptr);
2963 pickNodeFromQueue(TopCand);
2964 assert(TopCand.Reason != NoCand && "failed to find a candidate");
2965 tracePick(TopCand, true);
2968 } while (SU->isScheduled);
2971 Top.removeReady(SU);
2973 DEBUG(dbgs() << "Scheduling SU(" << SU->NodeNum << ") " << *SU->getInstr());
2977 /// Called after ScheduleDAGMI has scheduled an instruction and updated
2978 /// scheduled/remaining flags in the DAG nodes.
2979 void PostGenericScheduler::schedNode(SUnit *SU, bool IsTopNode) {
2980 SU->TopReadyCycle = std::max(SU->TopReadyCycle, Top.getCurrCycle());
2984 /// Create a generic scheduler with no vreg liveness or DAG mutation passes.
2985 static ScheduleDAGInstrs *createGenericSchedPostRA(MachineSchedContext *C) {
2986 return new ScheduleDAGMI(C, make_unique<PostGenericScheduler>(C), /*IsPostRA=*/true);
2989 //===----------------------------------------------------------------------===//
2990 // ILP Scheduler. Currently for experimental analysis of heuristics.
2991 //===----------------------------------------------------------------------===//
2994 /// \brief Order nodes by the ILP metric.
2996 const SchedDFSResult *DFSResult;
2997 const BitVector *ScheduledTrees;
3000 ILPOrder(bool MaxILP)
3001 : DFSResult(nullptr), ScheduledTrees(nullptr), MaximizeILP(MaxILP) {}
3003 /// \brief Apply a less-than relation on node priority.
3005 /// (Return true if A comes after B in the Q.)
3006 bool operator()(const SUnit *A, const SUnit *B) const {
3007 unsigned SchedTreeA = DFSResult->getSubtreeID(A);
3008 unsigned SchedTreeB = DFSResult->getSubtreeID(B);
3009 if (SchedTreeA != SchedTreeB) {
3010 // Unscheduled trees have lower priority.
3011 if (ScheduledTrees->test(SchedTreeA) != ScheduledTrees->test(SchedTreeB))
3012 return ScheduledTrees->test(SchedTreeB);
3014 // Trees with shallower connections have have lower priority.
3015 if (DFSResult->getSubtreeLevel(SchedTreeA)
3016 != DFSResult->getSubtreeLevel(SchedTreeB)) {
3017 return DFSResult->getSubtreeLevel(SchedTreeA)
3018 < DFSResult->getSubtreeLevel(SchedTreeB);
3022 return DFSResult->getILP(A) < DFSResult->getILP(B);
3024 return DFSResult->getILP(A) > DFSResult->getILP(B);
3028 /// \brief Schedule based on the ILP metric.
3029 class ILPScheduler : public MachineSchedStrategy {
3030 ScheduleDAGMILive *DAG;
3033 std::vector<SUnit*> ReadyQ;
3035 ILPScheduler(bool MaximizeILP): DAG(nullptr), Cmp(MaximizeILP) {}
3037 void initialize(ScheduleDAGMI *dag) override {
3038 assert(dag->hasVRegLiveness() && "ILPScheduler needs vreg liveness");
3039 DAG = static_cast<ScheduleDAGMILive*>(dag);
3040 DAG->computeDFSResult();
3041 Cmp.DFSResult = DAG->getDFSResult();
3042 Cmp.ScheduledTrees = &DAG->getScheduledTrees();
3046 void registerRoots() override {
3047 // Restore the heap in ReadyQ with the updated DFS results.
3048 std::make_heap(ReadyQ.begin(), ReadyQ.end(), Cmp);
3051 /// Implement MachineSchedStrategy interface.
3052 /// -----------------------------------------
3054 /// Callback to select the highest priority node from the ready Q.
3055 SUnit *pickNode(bool &IsTopNode) override {
3056 if (ReadyQ.empty()) return nullptr;
3057 std::pop_heap(ReadyQ.begin(), ReadyQ.end(), Cmp);
3058 SUnit *SU = ReadyQ.back();
3061 DEBUG(dbgs() << "Pick node " << "SU(" << SU->NodeNum << ") "
3062 << " ILP: " << DAG->getDFSResult()->getILP(SU)
3063 << " Tree: " << DAG->getDFSResult()->getSubtreeID(SU) << " @"
3064 << DAG->getDFSResult()->getSubtreeLevel(
3065 DAG->getDFSResult()->getSubtreeID(SU)) << '\n'
3066 << "Scheduling " << *SU->getInstr());
3070 /// \brief Scheduler callback to notify that a new subtree is scheduled.
3071 void scheduleTree(unsigned SubtreeID) override {
3072 std::make_heap(ReadyQ.begin(), ReadyQ.end(), Cmp);
3075 /// Callback after a node is scheduled. Mark a newly scheduled tree, notify
3076 /// DFSResults, and resort the priority Q.
3077 void schedNode(SUnit *SU, bool IsTopNode) override {
3078 assert(!IsTopNode && "SchedDFSResult needs bottom-up");
3081 void releaseTopNode(SUnit *) override { /*only called for top roots*/ }
3083 void releaseBottomNode(SUnit *SU) override {
3084 ReadyQ.push_back(SU);
3085 std::push_heap(ReadyQ.begin(), ReadyQ.end(), Cmp);
3090 static ScheduleDAGInstrs *createILPMaxScheduler(MachineSchedContext *C) {
3091 return new ScheduleDAGMILive(C, make_unique<ILPScheduler>(true));
3093 static ScheduleDAGInstrs *createILPMinScheduler(MachineSchedContext *C) {
3094 return new ScheduleDAGMILive(C, make_unique<ILPScheduler>(false));
3096 static MachineSchedRegistry ILPMaxRegistry(
3097 "ilpmax", "Schedule bottom-up for max ILP", createILPMaxScheduler);
3098 static MachineSchedRegistry ILPMinRegistry(
3099 "ilpmin", "Schedule bottom-up for min ILP", createILPMinScheduler);
3101 //===----------------------------------------------------------------------===//
3102 // Machine Instruction Shuffler for Correctness Testing
3103 //===----------------------------------------------------------------------===//
3107 /// Apply a less-than relation on the node order, which corresponds to the
3108 /// instruction order prior to scheduling. IsReverse implements greater-than.
3109 template<bool IsReverse>
3111 bool operator()(SUnit *A, SUnit *B) const {
3113 return A->NodeNum > B->NodeNum;
3115 return A->NodeNum < B->NodeNum;
3119 /// Reorder instructions as much as possible.
3120 class InstructionShuffler : public MachineSchedStrategy {
3124 // Using a less-than relation (SUnitOrder<false>) for the TopQ priority
3125 // gives nodes with a higher number higher priority causing the latest
3126 // instructions to be scheduled first.
3127 PriorityQueue<SUnit*, std::vector<SUnit*>, SUnitOrder<false> >
3129 // When scheduling bottom-up, use greater-than as the queue priority.
3130 PriorityQueue<SUnit*, std::vector<SUnit*>, SUnitOrder<true> >
3133 InstructionShuffler(bool alternate, bool topdown)
3134 : IsAlternating(alternate), IsTopDown(topdown) {}
3136 void initialize(ScheduleDAGMI*) override {
3141 /// Implement MachineSchedStrategy interface.
3142 /// -----------------------------------------
3144 SUnit *pickNode(bool &IsTopNode) override {
3148 if (TopQ.empty()) return nullptr;
3151 } while (SU->isScheduled);
3156 if (BottomQ.empty()) return nullptr;
3159 } while (SU->isScheduled);
3163 IsTopDown = !IsTopDown;
3167 void schedNode(SUnit *SU, bool IsTopNode) override {}
3169 void releaseTopNode(SUnit *SU) override {
3172 void releaseBottomNode(SUnit *SU) override {
3178 static ScheduleDAGInstrs *createInstructionShuffler(MachineSchedContext *C) {
3179 bool Alternate = !ForceTopDown && !ForceBottomUp;
3180 bool TopDown = !ForceBottomUp;
3181 assert((TopDown || !ForceTopDown) &&
3182 "-misched-topdown incompatible with -misched-bottomup");
3183 return new ScheduleDAGMILive(C, make_unique<InstructionShuffler>(Alternate, TopDown));
3185 static MachineSchedRegistry ShufflerRegistry(
3186 "shuffle", "Shuffle machine instructions alternating directions",
3187 createInstructionShuffler);
3190 //===----------------------------------------------------------------------===//
3191 // GraphWriter support for ScheduleDAGMILive.
3192 //===----------------------------------------------------------------------===//
3197 template<> struct GraphTraits<
3198 ScheduleDAGMI*> : public GraphTraits<ScheduleDAG*> {};
3201 struct DOTGraphTraits<ScheduleDAGMI*> : public DefaultDOTGraphTraits {
3203 DOTGraphTraits (bool isSimple=false) : DefaultDOTGraphTraits(isSimple) {}
3205 static std::string getGraphName(const ScheduleDAG *G) {
3206 return G->MF.getName();
3209 static bool renderGraphFromBottomUp() {
3213 static bool isNodeHidden(const SUnit *Node) {
3214 return (Node->Preds.size() > 10 || Node->Succs.size() > 10);
3217 static bool hasNodeAddressLabel(const SUnit *Node,
3218 const ScheduleDAG *Graph) {
3222 /// If you want to override the dot attributes printed for a particular
3223 /// edge, override this method.
3224 static std::string getEdgeAttributes(const SUnit *Node,
3226 const ScheduleDAG *Graph) {
3227 if (EI.isArtificialDep())
3228 return "color=cyan,style=dashed";
3230 return "color=blue,style=dashed";
3234 static std::string getNodeLabel(const SUnit *SU, const ScheduleDAG *G) {
3236 raw_string_ostream SS(Str);
3237 const ScheduleDAGMI *DAG = static_cast<const ScheduleDAGMI*>(G);
3238 const SchedDFSResult *DFS = DAG->hasVRegLiveness() ?
3239 static_cast<const ScheduleDAGMILive*>(G)->getDFSResult() : nullptr;
3240 SS << "SU:" << SU->NodeNum;
3242 SS << " I:" << DFS->getNumInstrs(SU);
3245 static std::string getNodeDescription(const SUnit *SU, const ScheduleDAG *G) {
3246 return G->getGraphNodeLabel(SU);
3249 static std::string getNodeAttributes(const SUnit *N, const ScheduleDAG *G) {
3250 std::string Str("shape=Mrecord");
3251 const ScheduleDAGMI *DAG = static_cast<const ScheduleDAGMI*>(G);
3252 const SchedDFSResult *DFS = DAG->hasVRegLiveness() ?
3253 static_cast<const ScheduleDAGMILive*>(G)->getDFSResult() : nullptr;
3255 Str += ",style=filled,fillcolor=\"#";
3256 Str += DOT::getColorString(DFS->getSubtreeID(N));
3265 /// viewGraph - Pop up a ghostview window with the reachable parts of the DAG
3266 /// rendered using 'dot'.
3268 void ScheduleDAGMI::viewGraph(const Twine &Name, const Twine &Title) {
3270 ViewGraph(this, Name, false, Title);
3272 errs() << "ScheduleDAGMI::viewGraph is only available in debug builds on "
3273 << "systems with Graphviz or gv!\n";
3277 /// Out-of-line implementation with no arguments is handy for gdb.
3278 void ScheduleDAGMI::viewGraph() {
3279 viewGraph(getDAGName(), "Scheduling-Units Graph for " + getDAGName());