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/CodeGen/TargetPassConfig.h"
27 #include "llvm/Support/CommandLine.h"
28 #include "llvm/Support/Debug.h"
29 #include "llvm/Support/ErrorHandling.h"
30 #include "llvm/Support/GraphWriter.h"
31 #include "llvm/Support/raw_ostream.h"
32 #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"));
44 DumpCriticalPathLength("misched-dcpl", cl::Hidden,
45 cl::desc("Print critical path length to stdout"));
49 static cl::opt<bool> ViewMISchedDAGs("view-misched-dags", cl::Hidden,
50 cl::desc("Pop up a window to show MISched dags after they are processed"));
52 /// In some situations a few uninteresting nodes depend on nearly all other
53 /// nodes in the graph, provide a cutoff to hide them.
54 static cl::opt<unsigned> ViewMISchedCutoff("view-misched-cutoff", cl::Hidden,
55 cl::desc("Hide nodes with more predecessor/successor than cutoff"));
57 static cl::opt<unsigned> MISchedCutoff("misched-cutoff", cl::Hidden,
58 cl::desc("Stop scheduling after N instructions"), cl::init(~0U));
60 static cl::opt<std::string> SchedOnlyFunc("misched-only-func", cl::Hidden,
61 cl::desc("Only schedule this function"));
62 static cl::opt<unsigned> SchedOnlyBlock("misched-only-block", cl::Hidden,
63 cl::desc("Only schedule this MBB#"));
65 static bool ViewMISchedDAGs = false;
68 /// Avoid quadratic complexity in unusually large basic blocks by limiting the
69 /// size of the ready lists.
70 static cl::opt<unsigned> ReadyListLimit("misched-limit", cl::Hidden,
71 cl::desc("Limit ready list to N instructions"), cl::init(256));
73 static cl::opt<bool> EnableRegPressure("misched-regpressure", cl::Hidden,
74 cl::desc("Enable register pressure scheduling."), cl::init(true));
76 static cl::opt<bool> EnableCyclicPath("misched-cyclicpath", cl::Hidden,
77 cl::desc("Enable cyclic critical path analysis."), cl::init(true));
79 static cl::opt<bool> EnableMemOpCluster("misched-cluster", cl::Hidden,
80 cl::desc("Enable memop clustering."),
83 // Experimental heuristics
84 static cl::opt<bool> EnableMacroFusion("misched-fusion", cl::Hidden,
85 cl::desc("Enable scheduling for macro fusion."), cl::init(true));
87 static cl::opt<bool> VerifyScheduling("verify-misched", cl::Hidden,
88 cl::desc("Verify machine instrs before and after machine scheduling"));
90 // DAG subtrees must have at least this many nodes.
91 static const unsigned MinSubtreeSize = 8;
93 // Pin the vtables to this file.
94 void MachineSchedStrategy::anchor() {}
95 void ScheduleDAGMutation::anchor() {}
97 //===----------------------------------------------------------------------===//
98 // Machine Instruction Scheduling Pass and Registry
99 //===----------------------------------------------------------------------===//
101 MachineSchedContext::MachineSchedContext():
102 MF(nullptr), MLI(nullptr), MDT(nullptr), PassConfig(nullptr), AA(nullptr), LIS(nullptr) {
103 RegClassInfo = new RegisterClassInfo();
106 MachineSchedContext::~MachineSchedContext() {
111 /// Base class for a machine scheduler class that can run at any point.
112 class MachineSchedulerBase : public MachineSchedContext,
113 public MachineFunctionPass {
115 MachineSchedulerBase(char &ID): MachineFunctionPass(ID) {}
117 void print(raw_ostream &O, const Module* = nullptr) const override;
120 void scheduleRegions(ScheduleDAGInstrs &Scheduler, bool FixKillFlags);
123 /// MachineScheduler runs after coalescing and before register allocation.
124 class MachineScheduler : public MachineSchedulerBase {
128 void getAnalysisUsage(AnalysisUsage &AU) const override;
130 bool runOnMachineFunction(MachineFunction&) override;
132 static char ID; // Class identification, replacement for typeinfo
135 ScheduleDAGInstrs *createMachineScheduler();
138 /// PostMachineScheduler runs after shortly before code emission.
139 class PostMachineScheduler : public MachineSchedulerBase {
141 PostMachineScheduler();
143 void getAnalysisUsage(AnalysisUsage &AU) const override;
145 bool runOnMachineFunction(MachineFunction&) override;
147 static char ID; // Class identification, replacement for typeinfo
150 ScheduleDAGInstrs *createPostMachineScheduler();
154 char MachineScheduler::ID = 0;
156 char &llvm::MachineSchedulerID = MachineScheduler::ID;
158 INITIALIZE_PASS_BEGIN(MachineScheduler, "machine-scheduler",
159 "Machine Instruction Scheduler", false, false)
160 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)
161 INITIALIZE_PASS_DEPENDENCY(SlotIndexes)
162 INITIALIZE_PASS_DEPENDENCY(LiveIntervals)
163 INITIALIZE_PASS_END(MachineScheduler, "machine-scheduler",
164 "Machine Instruction Scheduler", false, false)
166 MachineScheduler::MachineScheduler()
167 : MachineSchedulerBase(ID) {
168 initializeMachineSchedulerPass(*PassRegistry::getPassRegistry());
171 void MachineScheduler::getAnalysisUsage(AnalysisUsage &AU) const {
172 AU.setPreservesCFG();
173 AU.addRequiredID(MachineDominatorsID);
174 AU.addRequired<MachineLoopInfo>();
175 AU.addRequired<AAResultsWrapperPass>();
176 AU.addRequired<TargetPassConfig>();
177 AU.addRequired<SlotIndexes>();
178 AU.addPreserved<SlotIndexes>();
179 AU.addRequired<LiveIntervals>();
180 AU.addPreserved<LiveIntervals>();
181 MachineFunctionPass::getAnalysisUsage(AU);
184 char PostMachineScheduler::ID = 0;
186 char &llvm::PostMachineSchedulerID = PostMachineScheduler::ID;
188 INITIALIZE_PASS(PostMachineScheduler, "postmisched",
189 "PostRA Machine Instruction Scheduler", false, false)
191 PostMachineScheduler::PostMachineScheduler()
192 : MachineSchedulerBase(ID) {
193 initializePostMachineSchedulerPass(*PassRegistry::getPassRegistry());
196 void PostMachineScheduler::getAnalysisUsage(AnalysisUsage &AU) const {
197 AU.setPreservesCFG();
198 AU.addRequiredID(MachineDominatorsID);
199 AU.addRequired<MachineLoopInfo>();
200 AU.addRequired<TargetPassConfig>();
201 MachineFunctionPass::getAnalysisUsage(AU);
204 MachinePassRegistry MachineSchedRegistry::Registry;
206 /// A dummy default scheduler factory indicates whether the scheduler
207 /// is overridden on the command line.
208 static ScheduleDAGInstrs *useDefaultMachineSched(MachineSchedContext *C) {
212 /// MachineSchedOpt allows command line selection of the scheduler.
213 static cl::opt<MachineSchedRegistry::ScheduleDAGCtor, false,
214 RegisterPassParser<MachineSchedRegistry> >
215 MachineSchedOpt("misched",
216 cl::init(&useDefaultMachineSched), cl::Hidden,
217 cl::desc("Machine instruction scheduler to use"));
219 static MachineSchedRegistry
220 DefaultSchedRegistry("default", "Use the target's default scheduler choice.",
221 useDefaultMachineSched);
223 static cl::opt<bool> EnableMachineSched(
225 cl::desc("Enable the machine instruction scheduling pass."), cl::init(true),
228 static cl::opt<bool> EnablePostRAMachineSched(
229 "enable-post-misched",
230 cl::desc("Enable the post-ra machine instruction scheduling pass."),
231 cl::init(true), cl::Hidden);
233 /// Decrement this iterator until reaching the top or a non-debug instr.
234 static MachineBasicBlock::const_iterator
235 priorNonDebug(MachineBasicBlock::const_iterator I,
236 MachineBasicBlock::const_iterator Beg) {
237 assert(I != Beg && "reached the top of the region, cannot decrement");
239 if (!I->isDebugValue())
245 /// Non-const version.
246 static MachineBasicBlock::iterator
247 priorNonDebug(MachineBasicBlock::iterator I,
248 MachineBasicBlock::const_iterator Beg) {
249 return priorNonDebug(MachineBasicBlock::const_iterator(I), Beg)
250 .getNonConstIterator();
253 /// If this iterator is a debug value, increment until reaching the End or a
254 /// non-debug instruction.
255 static MachineBasicBlock::const_iterator
256 nextIfDebug(MachineBasicBlock::const_iterator I,
257 MachineBasicBlock::const_iterator End) {
258 for(; I != End; ++I) {
259 if (!I->isDebugValue())
265 /// Non-const version.
266 static MachineBasicBlock::iterator
267 nextIfDebug(MachineBasicBlock::iterator I,
268 MachineBasicBlock::const_iterator End) {
269 return nextIfDebug(MachineBasicBlock::const_iterator(I), End)
270 .getNonConstIterator();
273 /// Instantiate a ScheduleDAGInstrs that will be owned by the caller.
274 ScheduleDAGInstrs *MachineScheduler::createMachineScheduler() {
275 // Select the scheduler, or set the default.
276 MachineSchedRegistry::ScheduleDAGCtor Ctor = MachineSchedOpt;
277 if (Ctor != useDefaultMachineSched)
280 // Get the default scheduler set by the target for this function.
281 ScheduleDAGInstrs *Scheduler = PassConfig->createMachineScheduler(this);
285 // Default to GenericScheduler.
286 return createGenericSchedLive(this);
289 /// Instantiate a ScheduleDAGInstrs for PostRA scheduling that will be owned by
290 /// the caller. We don't have a command line option to override the postRA
291 /// scheduler. The Target must configure it.
292 ScheduleDAGInstrs *PostMachineScheduler::createPostMachineScheduler() {
293 // Get the postRA scheduler set by the target for this function.
294 ScheduleDAGInstrs *Scheduler = PassConfig->createPostMachineScheduler(this);
298 // Default to GenericScheduler.
299 return createGenericSchedPostRA(this);
302 /// Top-level MachineScheduler pass driver.
304 /// Visit blocks in function order. Divide each block into scheduling regions
305 /// and visit them bottom-up. Visiting regions bottom-up is not required, but is
306 /// consistent with the DAG builder, which traverses the interior of the
307 /// scheduling regions bottom-up.
309 /// This design avoids exposing scheduling boundaries to the DAG builder,
310 /// simplifying the DAG builder's support for "special" target instructions.
311 /// At the same time the design allows target schedulers to operate across
312 /// scheduling boundaries, for example to bundle the boudary instructions
313 /// without reordering them. This creates complexity, because the target
314 /// scheduler must update the RegionBegin and RegionEnd positions cached by
315 /// ScheduleDAGInstrs whenever adding or removing instructions. A much simpler
316 /// design would be to split blocks at scheduling boundaries, but LLVM has a
317 /// general bias against block splitting purely for implementation simplicity.
318 bool MachineScheduler::runOnMachineFunction(MachineFunction &mf) {
319 if (skipFunction(*mf.getFunction()))
322 if (EnableMachineSched.getNumOccurrences()) {
323 if (!EnableMachineSched)
325 } else if (!mf.getSubtarget().enableMachineScheduler())
328 DEBUG(dbgs() << "Before MISched:\n"; mf.print(dbgs()));
330 // Initialize the context of the pass.
332 MLI = &getAnalysis<MachineLoopInfo>();
333 MDT = &getAnalysis<MachineDominatorTree>();
334 PassConfig = &getAnalysis<TargetPassConfig>();
335 AA = &getAnalysis<AAResultsWrapperPass>().getAAResults();
337 LIS = &getAnalysis<LiveIntervals>();
339 if (VerifyScheduling) {
341 MF->verify(this, "Before machine scheduling.");
343 RegClassInfo->runOnMachineFunction(*MF);
345 // Instantiate the selected scheduler for this target, function, and
346 // optimization level.
347 std::unique_ptr<ScheduleDAGInstrs> Scheduler(createMachineScheduler());
348 scheduleRegions(*Scheduler, false);
351 if (VerifyScheduling)
352 MF->verify(this, "After machine scheduling.");
356 bool PostMachineScheduler::runOnMachineFunction(MachineFunction &mf) {
357 if (skipFunction(*mf.getFunction()))
360 if (EnablePostRAMachineSched.getNumOccurrences()) {
361 if (!EnablePostRAMachineSched)
363 } else if (!mf.getSubtarget().enablePostRAScheduler()) {
364 DEBUG(dbgs() << "Subtarget disables post-MI-sched.\n");
367 DEBUG(dbgs() << "Before post-MI-sched:\n"; mf.print(dbgs()));
369 // Initialize the context of the pass.
371 PassConfig = &getAnalysis<TargetPassConfig>();
373 if (VerifyScheduling)
374 MF->verify(this, "Before post machine scheduling.");
376 // Instantiate the selected scheduler for this target, function, and
377 // optimization level.
378 std::unique_ptr<ScheduleDAGInstrs> Scheduler(createPostMachineScheduler());
379 scheduleRegions(*Scheduler, true);
381 if (VerifyScheduling)
382 MF->verify(this, "After post machine scheduling.");
386 /// Return true of the given instruction should not be included in a scheduling
389 /// MachineScheduler does not currently support scheduling across calls. To
390 /// handle calls, the DAG builder needs to be modified to create register
391 /// anti/output dependencies on the registers clobbered by the call's regmask
392 /// operand. In PreRA scheduling, the stack pointer adjustment already prevents
393 /// scheduling across calls. In PostRA scheduling, we need the isCall to enforce
394 /// the boundary, but there would be no benefit to postRA scheduling across
395 /// calls this late anyway.
396 static bool isSchedBoundary(MachineBasicBlock::iterator MI,
397 MachineBasicBlock *MBB,
399 const TargetInstrInfo *TII) {
400 return MI->isCall() || TII->isSchedulingBoundary(*MI, MBB, *MF);
403 /// Main driver for both MachineScheduler and PostMachineScheduler.
404 void MachineSchedulerBase::scheduleRegions(ScheduleDAGInstrs &Scheduler,
406 const TargetInstrInfo *TII = MF->getSubtarget().getInstrInfo();
408 // Visit all machine basic blocks.
410 // TODO: Visit blocks in global postorder or postorder within the bottom-up
411 // loop tree. Then we can optionally compute global RegPressure.
412 for (MachineFunction::iterator MBB = MF->begin(), MBBEnd = MF->end();
413 MBB != MBBEnd; ++MBB) {
415 Scheduler.startBlock(&*MBB);
418 if (SchedOnlyFunc.getNumOccurrences() && SchedOnlyFunc != MF->getName())
420 if (SchedOnlyBlock.getNumOccurrences()
421 && (int)SchedOnlyBlock != MBB->getNumber())
425 // Break the block into scheduling regions [I, RegionEnd), and schedule each
426 // region as soon as it is discovered. RegionEnd points the scheduling
427 // boundary at the bottom of the region. The DAG does not include RegionEnd,
428 // but the region does (i.e. the next RegionEnd is above the previous
429 // RegionBegin). If the current block has no terminator then RegionEnd ==
430 // MBB->end() for the bottom region.
432 // The Scheduler may insert instructions during either schedule() or
433 // exitRegion(), even for empty regions. So the local iterators 'I' and
434 // 'RegionEnd' are invalid across these calls.
436 // MBB::size() uses instr_iterator to count. Here we need a bundle to count
437 // as a single instruction.
438 for(MachineBasicBlock::iterator RegionEnd = MBB->end();
439 RegionEnd != MBB->begin(); RegionEnd = Scheduler.begin()) {
441 // Avoid decrementing RegionEnd for blocks with no terminator.
442 if (RegionEnd != MBB->end() ||
443 isSchedBoundary(&*std::prev(RegionEnd), &*MBB, MF, TII)) {
447 // The next region starts above the previous region. Look backward in the
448 // instruction stream until we find the nearest boundary.
449 unsigned NumRegionInstrs = 0;
450 MachineBasicBlock::iterator I = RegionEnd;
451 for (;I != MBB->begin(); --I) {
452 MachineInstr &MI = *std::prev(I);
453 if (isSchedBoundary(&MI, &*MBB, MF, TII))
455 if (!MI.isDebugValue())
458 // Notify the scheduler of the region, even if we may skip scheduling
459 // it. Perhaps it still needs to be bundled.
460 Scheduler.enterRegion(&*MBB, I, RegionEnd, NumRegionInstrs);
462 // Skip empty scheduling regions (0 or 1 schedulable instructions).
463 if (I == RegionEnd || I == std::prev(RegionEnd)) {
464 // Close the current region. Bundle the terminator if needed.
465 // This invalidates 'RegionEnd' and 'I'.
466 Scheduler.exitRegion();
469 DEBUG(dbgs() << "********** MI Scheduling **********\n");
470 DEBUG(dbgs() << MF->getName()
471 << ":BB#" << MBB->getNumber() << " " << MBB->getName()
472 << "\n From: " << *I << " To: ";
473 if (RegionEnd != MBB->end()) dbgs() << *RegionEnd;
474 else dbgs() << "End";
475 dbgs() << " RegionInstrs: " << NumRegionInstrs << '\n');
476 if (DumpCriticalPathLength) {
477 errs() << MF->getName();
478 errs() << ":BB# " << MBB->getNumber();
479 errs() << " " << MBB->getName() << " \n";
482 // Schedule a region: possibly reorder instructions.
483 // This invalidates 'RegionEnd' and 'I'.
484 Scheduler.schedule();
486 // Close the current region.
487 Scheduler.exitRegion();
489 // Scheduling has invalidated the current iterator 'I'. Ask the
490 // scheduler for the top of it's scheduled region.
491 RegionEnd = Scheduler.begin();
493 Scheduler.finishBlock();
494 // FIXME: Ideally, no further passes should rely on kill flags. However,
495 // thumb2 size reduction is currently an exception, so the PostMIScheduler
498 Scheduler.fixupKills(&*MBB);
500 Scheduler.finalizeSchedule();
503 void MachineSchedulerBase::print(raw_ostream &O, const Module* m) const {
508 void ReadyQueue::dump() {
509 dbgs() << "Queue " << Name << ": ";
510 for (unsigned i = 0, e = Queue.size(); i < e; ++i)
511 dbgs() << Queue[i]->NodeNum << " ";
515 //===----------------------------------------------------------------------===//
516 // ScheduleDAGMI - Basic machine instruction scheduling. This is
517 // independent of PreRA/PostRA scheduling and involves no extra book-keeping for
518 // virtual registers.
519 // ===----------------------------------------------------------------------===/
521 // Provide a vtable anchor.
522 ScheduleDAGMI::~ScheduleDAGMI() {
525 bool ScheduleDAGMI::canAddEdge(SUnit *SuccSU, SUnit *PredSU) {
526 return SuccSU == &ExitSU || !Topo.IsReachable(PredSU, SuccSU);
529 bool ScheduleDAGMI::addEdge(SUnit *SuccSU, const SDep &PredDep) {
530 if (SuccSU != &ExitSU) {
531 // Do not use WillCreateCycle, it assumes SD scheduling.
532 // If Pred is reachable from Succ, then the edge creates a cycle.
533 if (Topo.IsReachable(PredDep.getSUnit(), SuccSU))
535 Topo.AddPred(SuccSU, PredDep.getSUnit());
537 SuccSU->addPred(PredDep, /*Required=*/!PredDep.isArtificial());
538 // Return true regardless of whether a new edge needed to be inserted.
542 /// ReleaseSucc - Decrement the NumPredsLeft count of a successor. When
543 /// NumPredsLeft reaches zero, release the successor node.
545 /// FIXME: Adjust SuccSU height based on MinLatency.
546 void ScheduleDAGMI::releaseSucc(SUnit *SU, SDep *SuccEdge) {
547 SUnit *SuccSU = SuccEdge->getSUnit();
549 if (SuccEdge->isWeak()) {
550 --SuccSU->WeakPredsLeft;
551 if (SuccEdge->isCluster())
552 NextClusterSucc = SuccSU;
556 if (SuccSU->NumPredsLeft == 0) {
557 dbgs() << "*** Scheduling failed! ***\n";
559 dbgs() << " has been released too many times!\n";
560 llvm_unreachable(nullptr);
563 // SU->TopReadyCycle was set to CurrCycle when it was scheduled. However,
564 // CurrCycle may have advanced since then.
565 if (SuccSU->TopReadyCycle < SU->TopReadyCycle + SuccEdge->getLatency())
566 SuccSU->TopReadyCycle = SU->TopReadyCycle + SuccEdge->getLatency();
568 --SuccSU->NumPredsLeft;
569 if (SuccSU->NumPredsLeft == 0 && SuccSU != &ExitSU)
570 SchedImpl->releaseTopNode(SuccSU);
573 /// releaseSuccessors - Call releaseSucc on each of SU's successors.
574 void ScheduleDAGMI::releaseSuccessors(SUnit *SU) {
575 for (SUnit::succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
577 releaseSucc(SU, &*I);
581 /// ReleasePred - Decrement the NumSuccsLeft count of a predecessor. When
582 /// NumSuccsLeft reaches zero, release the predecessor node.
584 /// FIXME: Adjust PredSU height based on MinLatency.
585 void ScheduleDAGMI::releasePred(SUnit *SU, SDep *PredEdge) {
586 SUnit *PredSU = PredEdge->getSUnit();
588 if (PredEdge->isWeak()) {
589 --PredSU->WeakSuccsLeft;
590 if (PredEdge->isCluster())
591 NextClusterPred = PredSU;
595 if (PredSU->NumSuccsLeft == 0) {
596 dbgs() << "*** Scheduling failed! ***\n";
598 dbgs() << " has been released too many times!\n";
599 llvm_unreachable(nullptr);
602 // SU->BotReadyCycle was set to CurrCycle when it was scheduled. However,
603 // CurrCycle may have advanced since then.
604 if (PredSU->BotReadyCycle < SU->BotReadyCycle + PredEdge->getLatency())
605 PredSU->BotReadyCycle = SU->BotReadyCycle + PredEdge->getLatency();
607 --PredSU->NumSuccsLeft;
608 if (PredSU->NumSuccsLeft == 0 && PredSU != &EntrySU)
609 SchedImpl->releaseBottomNode(PredSU);
612 /// releasePredecessors - Call releasePred on each of SU's predecessors.
613 void ScheduleDAGMI::releasePredecessors(SUnit *SU) {
614 for (SUnit::pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
616 releasePred(SU, &*I);
620 /// enterRegion - Called back from MachineScheduler::runOnMachineFunction after
621 /// crossing a scheduling boundary. [begin, end) includes all instructions in
622 /// the region, including the boundary itself and single-instruction regions
623 /// that don't get scheduled.
624 void ScheduleDAGMI::enterRegion(MachineBasicBlock *bb,
625 MachineBasicBlock::iterator begin,
626 MachineBasicBlock::iterator end,
627 unsigned regioninstrs)
629 ScheduleDAGInstrs::enterRegion(bb, begin, end, regioninstrs);
631 SchedImpl->initPolicy(begin, end, regioninstrs);
634 /// This is normally called from the main scheduler loop but may also be invoked
635 /// by the scheduling strategy to perform additional code motion.
636 void ScheduleDAGMI::moveInstruction(
637 MachineInstr *MI, MachineBasicBlock::iterator InsertPos) {
638 // Advance RegionBegin if the first instruction moves down.
639 if (&*RegionBegin == MI)
642 // Update the instruction stream.
643 BB->splice(InsertPos, BB, MI);
645 // Update LiveIntervals
647 LIS->handleMove(*MI, /*UpdateFlags=*/true);
649 // Recede RegionBegin if an instruction moves above the first.
650 if (RegionBegin == InsertPos)
654 bool ScheduleDAGMI::checkSchedLimit() {
656 if (NumInstrsScheduled == MISchedCutoff && MISchedCutoff != ~0U) {
657 CurrentTop = CurrentBottom;
660 ++NumInstrsScheduled;
665 /// Per-region scheduling driver, called back from
666 /// MachineScheduler::runOnMachineFunction. This is a simplified driver that
667 /// does not consider liveness or register pressure. It is useful for PostRA
668 /// scheduling and potentially other custom schedulers.
669 void ScheduleDAGMI::schedule() {
670 DEBUG(dbgs() << "ScheduleDAGMI::schedule starting\n");
671 DEBUG(SchedImpl->dumpPolicy());
676 Topo.InitDAGTopologicalSorting();
680 SmallVector<SUnit*, 8> TopRoots, BotRoots;
681 findRootsAndBiasEdges(TopRoots, BotRoots);
683 // Initialize the strategy before modifying the DAG.
684 // This may initialize a DFSResult to be used for queue priority.
685 SchedImpl->initialize(this);
688 if (EntrySU.getInstr() != nullptr)
689 EntrySU.dumpAll(this);
690 for (unsigned su = 0, e = SUnits.size(); su != e; ++su)
691 SUnits[su].dumpAll(this);
692 if (ExitSU.getInstr() != nullptr)
693 ExitSU.dumpAll(this);
695 if (ViewMISchedDAGs) viewGraph();
697 // Initialize ready queues now that the DAG and priority data are finalized.
698 initQueues(TopRoots, BotRoots);
700 bool IsTopNode = false;
702 DEBUG(dbgs() << "** ScheduleDAGMI::schedule picking next node\n");
703 SUnit *SU = SchedImpl->pickNode(IsTopNode);
706 assert(!SU->isScheduled && "Node already scheduled");
707 if (!checkSchedLimit())
710 MachineInstr *MI = SU->getInstr();
712 assert(SU->isTopReady() && "node still has unscheduled dependencies");
713 if (&*CurrentTop == MI)
714 CurrentTop = nextIfDebug(++CurrentTop, CurrentBottom);
716 moveInstruction(MI, CurrentTop);
718 assert(SU->isBottomReady() && "node still has unscheduled dependencies");
719 MachineBasicBlock::iterator priorII =
720 priorNonDebug(CurrentBottom, CurrentTop);
722 CurrentBottom = priorII;
724 if (&*CurrentTop == MI)
725 CurrentTop = nextIfDebug(++CurrentTop, priorII);
726 moveInstruction(MI, CurrentBottom);
730 // Notify the scheduling strategy before updating the DAG.
731 // This sets the scheduled node's ReadyCycle to CurrCycle. When updateQueues
732 // runs, it can then use the accurate ReadyCycle time to determine whether
733 // newly released nodes can move to the readyQ.
734 SchedImpl->schedNode(SU, IsTopNode);
736 updateQueues(SU, IsTopNode);
738 assert(CurrentTop == CurrentBottom && "Nonempty unscheduled zone.");
743 unsigned BBNum = begin()->getParent()->getNumber();
744 dbgs() << "*** Final schedule for BB#" << BBNum << " ***\n";
750 /// Apply each ScheduleDAGMutation step in order.
751 void ScheduleDAGMI::postprocessDAG() {
752 for (unsigned i = 0, e = Mutations.size(); i < e; ++i) {
753 Mutations[i]->apply(this);
758 findRootsAndBiasEdges(SmallVectorImpl<SUnit*> &TopRoots,
759 SmallVectorImpl<SUnit*> &BotRoots) {
760 for (std::vector<SUnit>::iterator
761 I = SUnits.begin(), E = SUnits.end(); I != E; ++I) {
763 assert(!SU->isBoundaryNode() && "Boundary node should not be in SUnits");
765 // Order predecessors so DFSResult follows the critical path.
766 SU->biasCriticalPath();
768 // A SUnit is ready to top schedule if it has no predecessors.
769 if (!I->NumPredsLeft)
770 TopRoots.push_back(SU);
771 // A SUnit is ready to bottom schedule if it has no successors.
772 if (!I->NumSuccsLeft)
773 BotRoots.push_back(SU);
775 ExitSU.biasCriticalPath();
778 /// Identify DAG roots and setup scheduler queues.
779 void ScheduleDAGMI::initQueues(ArrayRef<SUnit*> TopRoots,
780 ArrayRef<SUnit*> BotRoots) {
781 NextClusterSucc = nullptr;
782 NextClusterPred = nullptr;
784 // Release all DAG roots for scheduling, not including EntrySU/ExitSU.
786 // Nodes with unreleased weak edges can still be roots.
787 // Release top roots in forward order.
788 for (SmallVectorImpl<SUnit*>::const_iterator
789 I = TopRoots.begin(), E = TopRoots.end(); I != E; ++I) {
790 SchedImpl->releaseTopNode(*I);
792 // Release bottom roots in reverse order so the higher priority nodes appear
793 // first. This is more natural and slightly more efficient.
794 for (SmallVectorImpl<SUnit*>::const_reverse_iterator
795 I = BotRoots.rbegin(), E = BotRoots.rend(); I != E; ++I) {
796 SchedImpl->releaseBottomNode(*I);
799 releaseSuccessors(&EntrySU);
800 releasePredecessors(&ExitSU);
802 SchedImpl->registerRoots();
804 // Advance past initial DebugValues.
805 CurrentTop = nextIfDebug(RegionBegin, RegionEnd);
806 CurrentBottom = RegionEnd;
809 /// Update scheduler queues after scheduling an instruction.
810 void ScheduleDAGMI::updateQueues(SUnit *SU, bool IsTopNode) {
811 // Release dependent instructions for scheduling.
813 releaseSuccessors(SU);
815 releasePredecessors(SU);
817 SU->isScheduled = true;
820 /// Reinsert any remaining debug_values, just like the PostRA scheduler.
821 void ScheduleDAGMI::placeDebugValues() {
822 // If first instruction was a DBG_VALUE then put it back.
824 BB->splice(RegionBegin, BB, FirstDbgValue);
825 RegionBegin = FirstDbgValue;
828 for (std::vector<std::pair<MachineInstr *, MachineInstr *> >::iterator
829 DI = DbgValues.end(), DE = DbgValues.begin(); DI != DE; --DI) {
830 std::pair<MachineInstr *, MachineInstr *> P = *std::prev(DI);
831 MachineInstr *DbgValue = P.first;
832 MachineBasicBlock::iterator OrigPrevMI = P.second;
833 if (&*RegionBegin == DbgValue)
835 BB->splice(++OrigPrevMI, BB, DbgValue);
836 if (OrigPrevMI == std::prev(RegionEnd))
837 RegionEnd = DbgValue;
840 FirstDbgValue = nullptr;
843 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
844 void ScheduleDAGMI::dumpSchedule() const {
845 for (MachineBasicBlock::iterator MI = begin(), ME = end(); MI != ME; ++MI) {
846 if (SUnit *SU = getSUnit(&(*MI)))
849 dbgs() << "Missing SUnit\n";
854 //===----------------------------------------------------------------------===//
855 // ScheduleDAGMILive - Base class for MachineInstr scheduling with LiveIntervals
857 //===----------------------------------------------------------------------===//
859 ScheduleDAGMILive::~ScheduleDAGMILive() {
863 void ScheduleDAGMILive::collectVRegUses(SUnit &SU) {
864 const MachineInstr &MI = *SU.getInstr();
865 for (const MachineOperand &MO : MI.operands()) {
870 if (TrackLaneMasks && !MO.isUse())
873 unsigned Reg = MO.getReg();
874 if (!TargetRegisterInfo::isVirtualRegister(Reg))
878 if (TrackLaneMasks) {
879 bool FoundDef = false;
880 for (const MachineOperand &MO2 : MI.operands()) {
881 if (MO2.isReg() && MO2.isDef() && MO2.getReg() == Reg && !MO2.isDead()) {
890 // Record this local VReg use.
891 VReg2SUnitMultiMap::iterator UI = VRegUses.find(Reg);
892 for (; UI != VRegUses.end(); ++UI) {
896 if (UI == VRegUses.end())
897 VRegUses.insert(VReg2SUnit(Reg, LaneBitmask::getNone(), &SU));
901 /// enterRegion - Called back from MachineScheduler::runOnMachineFunction after
902 /// crossing a scheduling boundary. [begin, end) includes all instructions in
903 /// the region, including the boundary itself and single-instruction regions
904 /// that don't get scheduled.
905 void ScheduleDAGMILive::enterRegion(MachineBasicBlock *bb,
906 MachineBasicBlock::iterator begin,
907 MachineBasicBlock::iterator end,
908 unsigned regioninstrs)
910 // ScheduleDAGMI initializes SchedImpl's per-region policy.
911 ScheduleDAGMI::enterRegion(bb, begin, end, regioninstrs);
913 // For convenience remember the end of the liveness region.
914 LiveRegionEnd = (RegionEnd == bb->end()) ? RegionEnd : std::next(RegionEnd);
916 SUPressureDiffs.clear();
918 ShouldTrackPressure = SchedImpl->shouldTrackPressure();
919 ShouldTrackLaneMasks = SchedImpl->shouldTrackLaneMasks();
921 assert((!ShouldTrackLaneMasks || ShouldTrackPressure) &&
922 "ShouldTrackLaneMasks requires ShouldTrackPressure");
925 // Setup the register pressure trackers for the top scheduled top and bottom
926 // scheduled regions.
927 void ScheduleDAGMILive::initRegPressure() {
929 VRegUses.setUniverse(MRI.getNumVirtRegs());
930 for (SUnit &SU : SUnits)
933 TopRPTracker.init(&MF, RegClassInfo, LIS, BB, RegionBegin,
934 ShouldTrackLaneMasks, false);
935 BotRPTracker.init(&MF, RegClassInfo, LIS, BB, LiveRegionEnd,
936 ShouldTrackLaneMasks, false);
938 // Close the RPTracker to finalize live ins.
939 RPTracker.closeRegion();
941 DEBUG(RPTracker.dump());
943 // Initialize the live ins and live outs.
944 TopRPTracker.addLiveRegs(RPTracker.getPressure().LiveInRegs);
945 BotRPTracker.addLiveRegs(RPTracker.getPressure().LiveOutRegs);
947 // Close one end of the tracker so we can call
948 // getMaxUpward/DownwardPressureDelta before advancing across any
949 // instructions. This converts currently live regs into live ins/outs.
950 TopRPTracker.closeTop();
951 BotRPTracker.closeBottom();
953 BotRPTracker.initLiveThru(RPTracker);
954 if (!BotRPTracker.getLiveThru().empty()) {
955 TopRPTracker.initLiveThru(BotRPTracker.getLiveThru());
956 DEBUG(dbgs() << "Live Thru: ";
957 dumpRegSetPressure(BotRPTracker.getLiveThru(), TRI));
960 // For each live out vreg reduce the pressure change associated with other
961 // uses of the same vreg below the live-out reaching def.
962 updatePressureDiffs(RPTracker.getPressure().LiveOutRegs);
964 // Account for liveness generated by the region boundary.
965 if (LiveRegionEnd != RegionEnd) {
966 SmallVector<RegisterMaskPair, 8> LiveUses;
967 BotRPTracker.recede(&LiveUses);
968 updatePressureDiffs(LiveUses);
972 dbgs() << "Top Pressure:\n";
973 dumpRegSetPressure(TopRPTracker.getRegSetPressureAtPos(), TRI);
974 dbgs() << "Bottom Pressure:\n";
975 dumpRegSetPressure(BotRPTracker.getRegSetPressureAtPos(), TRI);
978 assert(BotRPTracker.getPos() == RegionEnd && "Can't find the region bottom");
980 // Cache the list of excess pressure sets in this region. This will also track
981 // the max pressure in the scheduled code for these sets.
982 RegionCriticalPSets.clear();
983 const std::vector<unsigned> &RegionPressure =
984 RPTracker.getPressure().MaxSetPressure;
985 for (unsigned i = 0, e = RegionPressure.size(); i < e; ++i) {
986 unsigned Limit = RegClassInfo->getRegPressureSetLimit(i);
987 if (RegionPressure[i] > Limit) {
988 DEBUG(dbgs() << TRI->getRegPressureSetName(i)
989 << " Limit " << Limit
990 << " Actual " << RegionPressure[i] << "\n");
991 RegionCriticalPSets.push_back(PressureChange(i));
994 DEBUG(dbgs() << "Excess PSets: ";
995 for (unsigned i = 0, e = RegionCriticalPSets.size(); i != e; ++i)
996 dbgs() << TRI->getRegPressureSetName(
997 RegionCriticalPSets[i].getPSet()) << " ";
1001 void ScheduleDAGMILive::
1002 updateScheduledPressure(const SUnit *SU,
1003 const std::vector<unsigned> &NewMaxPressure) {
1004 const PressureDiff &PDiff = getPressureDiff(SU);
1005 unsigned CritIdx = 0, CritEnd = RegionCriticalPSets.size();
1006 for (PressureDiff::const_iterator I = PDiff.begin(), E = PDiff.end();
1010 unsigned ID = I->getPSet();
1011 while (CritIdx != CritEnd && RegionCriticalPSets[CritIdx].getPSet() < ID)
1013 if (CritIdx != CritEnd && RegionCriticalPSets[CritIdx].getPSet() == ID) {
1014 if ((int)NewMaxPressure[ID] > RegionCriticalPSets[CritIdx].getUnitInc()
1015 && NewMaxPressure[ID] <= INT16_MAX)
1016 RegionCriticalPSets[CritIdx].setUnitInc(NewMaxPressure[ID]);
1018 unsigned Limit = RegClassInfo->getRegPressureSetLimit(ID);
1019 if (NewMaxPressure[ID] >= Limit - 2) {
1020 DEBUG(dbgs() << " " << TRI->getRegPressureSetName(ID) << ": "
1021 << NewMaxPressure[ID]
1022 << ((NewMaxPressure[ID] > Limit) ? " > " : " <= ") << Limit
1023 << "(+ " << BotRPTracker.getLiveThru()[ID] << " livethru)\n");
1028 /// Update the PressureDiff array for liveness after scheduling this
1030 void ScheduleDAGMILive::updatePressureDiffs(
1031 ArrayRef<RegisterMaskPair> LiveUses) {
1032 for (const RegisterMaskPair &P : LiveUses) {
1033 unsigned Reg = P.RegUnit;
1034 /// FIXME: Currently assuming single-use physregs.
1035 if (!TRI->isVirtualRegister(Reg))
1038 if (ShouldTrackLaneMasks) {
1039 // If the register has just become live then other uses won't change
1040 // this fact anymore => decrement pressure.
1041 // If the register has just become dead then other uses make it come
1042 // back to life => increment pressure.
1043 bool Decrement = P.LaneMask.any();
1045 for (const VReg2SUnit &V2SU
1046 : make_range(VRegUses.find(Reg), VRegUses.end())) {
1047 SUnit &SU = *V2SU.SU;
1048 if (SU.isScheduled || &SU == &ExitSU)
1051 PressureDiff &PDiff = getPressureDiff(&SU);
1052 PDiff.addPressureChange(Reg, Decrement, &MRI);
1054 dbgs() << " UpdateRegP: SU(" << SU.NodeNum << ") "
1055 << PrintReg(Reg, TRI) << ':' << PrintLaneMask(P.LaneMask)
1056 << ' ' << *SU.getInstr();
1062 assert(P.LaneMask.any());
1063 DEBUG(dbgs() << " LiveReg: " << PrintVRegOrUnit(Reg, TRI) << "\n");
1064 // This may be called before CurrentBottom has been initialized. However,
1065 // BotRPTracker must have a valid position. We want the value live into the
1066 // instruction or live out of the block, so ask for the previous
1067 // instruction's live-out.
1068 const LiveInterval &LI = LIS->getInterval(Reg);
1070 MachineBasicBlock::const_iterator I =
1071 nextIfDebug(BotRPTracker.getPos(), BB->end());
1073 VNI = LI.getVNInfoBefore(LIS->getMBBEndIdx(BB));
1075 LiveQueryResult LRQ = LI.Query(LIS->getInstructionIndex(*I));
1076 VNI = LRQ.valueIn();
1078 // RegisterPressureTracker guarantees that readsReg is true for LiveUses.
1079 assert(VNI && "No live value at use.");
1080 for (const VReg2SUnit &V2SU
1081 : make_range(VRegUses.find(Reg), VRegUses.end())) {
1082 SUnit *SU = V2SU.SU;
1083 // If this use comes before the reaching def, it cannot be a last use,
1084 // so decrease its pressure change.
1085 if (!SU->isScheduled && SU != &ExitSU) {
1086 LiveQueryResult LRQ =
1087 LI.Query(LIS->getInstructionIndex(*SU->getInstr()));
1088 if (LRQ.valueIn() == VNI) {
1089 PressureDiff &PDiff = getPressureDiff(SU);
1090 PDiff.addPressureChange(Reg, true, &MRI);
1092 dbgs() << " UpdateRegP: SU(" << SU->NodeNum << ") "
1104 /// schedule - Called back from MachineScheduler::runOnMachineFunction
1105 /// after setting up the current scheduling region. [RegionBegin, RegionEnd)
1106 /// only includes instructions that have DAG nodes, not scheduling boundaries.
1108 /// This is a skeletal driver, with all the functionality pushed into helpers,
1109 /// so that it can be easily extended by experimental schedulers. Generally,
1110 /// implementing MachineSchedStrategy should be sufficient to implement a new
1111 /// scheduling algorithm. However, if a scheduler further subclasses
1112 /// ScheduleDAGMILive then it will want to override this virtual method in order
1113 /// to update any specialized state.
1114 void ScheduleDAGMILive::schedule() {
1115 DEBUG(dbgs() << "ScheduleDAGMILive::schedule starting\n");
1116 DEBUG(SchedImpl->dumpPolicy());
1117 buildDAGWithRegPressure();
1119 Topo.InitDAGTopologicalSorting();
1123 SmallVector<SUnit*, 8> TopRoots, BotRoots;
1124 findRootsAndBiasEdges(TopRoots, BotRoots);
1126 // Initialize the strategy before modifying the DAG.
1127 // This may initialize a DFSResult to be used for queue priority.
1128 SchedImpl->initialize(this);
1131 if (EntrySU.getInstr() != nullptr)
1132 EntrySU.dumpAll(this);
1133 for (const SUnit &SU : SUnits) {
1135 if (ShouldTrackPressure) {
1136 dbgs() << " Pressure Diff : ";
1137 getPressureDiff(&SU).dump(*TRI);
1141 if (ExitSU.getInstr() != nullptr)
1142 ExitSU.dumpAll(this);
1144 if (ViewMISchedDAGs) viewGraph();
1146 // Initialize ready queues now that the DAG and priority data are finalized.
1147 initQueues(TopRoots, BotRoots);
1149 bool IsTopNode = false;
1151 DEBUG(dbgs() << "** ScheduleDAGMILive::schedule picking next node\n");
1152 SUnit *SU = SchedImpl->pickNode(IsTopNode);
1155 assert(!SU->isScheduled && "Node already scheduled");
1156 if (!checkSchedLimit())
1159 scheduleMI(SU, IsTopNode);
1162 unsigned SubtreeID = DFSResult->getSubtreeID(SU);
1163 if (!ScheduledTrees.test(SubtreeID)) {
1164 ScheduledTrees.set(SubtreeID);
1165 DFSResult->scheduleTree(SubtreeID);
1166 SchedImpl->scheduleTree(SubtreeID);
1170 // Notify the scheduling strategy after updating the DAG.
1171 SchedImpl->schedNode(SU, IsTopNode);
1173 updateQueues(SU, IsTopNode);
1175 assert(CurrentTop == CurrentBottom && "Nonempty unscheduled zone.");
1180 unsigned BBNum = begin()->getParent()->getNumber();
1181 dbgs() << "*** Final schedule for BB#" << BBNum << " ***\n";
1187 /// Build the DAG and setup three register pressure trackers.
1188 void ScheduleDAGMILive::buildDAGWithRegPressure() {
1189 if (!ShouldTrackPressure) {
1191 RegionCriticalPSets.clear();
1192 buildSchedGraph(AA);
1196 // Initialize the register pressure tracker used by buildSchedGraph.
1197 RPTracker.init(&MF, RegClassInfo, LIS, BB, LiveRegionEnd,
1198 ShouldTrackLaneMasks, /*TrackUntiedDefs=*/true);
1200 // Account for liveness generate by the region boundary.
1201 if (LiveRegionEnd != RegionEnd)
1204 // Build the DAG, and compute current register pressure.
1205 buildSchedGraph(AA, &RPTracker, &SUPressureDiffs, LIS, ShouldTrackLaneMasks);
1207 // Initialize top/bottom trackers after computing region pressure.
1211 void ScheduleDAGMILive::computeDFSResult() {
1213 DFSResult = new SchedDFSResult(/*BottomU*/true, MinSubtreeSize);
1215 ScheduledTrees.clear();
1216 DFSResult->resize(SUnits.size());
1217 DFSResult->compute(SUnits);
1218 ScheduledTrees.resize(DFSResult->getNumSubtrees());
1221 /// Compute the max cyclic critical path through the DAG. The scheduling DAG
1222 /// only provides the critical path for single block loops. To handle loops that
1223 /// span blocks, we could use the vreg path latencies provided by
1224 /// MachineTraceMetrics instead. However, MachineTraceMetrics is not currently
1225 /// available for use in the scheduler.
1227 /// The cyclic path estimation identifies a def-use pair that crosses the back
1228 /// edge and considers the depth and height of the nodes. For example, consider
1229 /// the following instruction sequence where each instruction has unit latency
1230 /// and defines an epomymous virtual register:
1232 /// a->b(a,c)->c(b)->d(c)->exit
1234 /// The cyclic critical path is a two cycles: b->c->b
1235 /// The acyclic critical path is four cycles: a->b->c->d->exit
1236 /// LiveOutHeight = height(c) = len(c->d->exit) = 2
1237 /// LiveOutDepth = depth(c) + 1 = len(a->b->c) + 1 = 3
1238 /// LiveInHeight = height(b) + 1 = len(b->c->d->exit) + 1 = 4
1239 /// LiveInDepth = depth(b) = len(a->b) = 1
1241 /// LiveOutDepth - LiveInDepth = 3 - 1 = 2
1242 /// LiveInHeight - LiveOutHeight = 4 - 2 = 2
1243 /// CyclicCriticalPath = min(2, 2) = 2
1245 /// This could be relevant to PostRA scheduling, but is currently implemented
1246 /// assuming LiveIntervals.
1247 unsigned ScheduleDAGMILive::computeCyclicCriticalPath() {
1248 // This only applies to single block loop.
1249 if (!BB->isSuccessor(BB))
1252 unsigned MaxCyclicLatency = 0;
1253 // Visit each live out vreg def to find def/use pairs that cross iterations.
1254 for (const RegisterMaskPair &P : RPTracker.getPressure().LiveOutRegs) {
1255 unsigned Reg = P.RegUnit;
1256 if (!TRI->isVirtualRegister(Reg))
1258 const LiveInterval &LI = LIS->getInterval(Reg);
1259 const VNInfo *DefVNI = LI.getVNInfoBefore(LIS->getMBBEndIdx(BB));
1263 MachineInstr *DefMI = LIS->getInstructionFromIndex(DefVNI->def);
1264 const SUnit *DefSU = getSUnit(DefMI);
1268 unsigned LiveOutHeight = DefSU->getHeight();
1269 unsigned LiveOutDepth = DefSU->getDepth() + DefSU->Latency;
1270 // Visit all local users of the vreg def.
1271 for (const VReg2SUnit &V2SU
1272 : make_range(VRegUses.find(Reg), VRegUses.end())) {
1273 SUnit *SU = V2SU.SU;
1277 // Only consider uses of the phi.
1278 LiveQueryResult LRQ = LI.Query(LIS->getInstructionIndex(*SU->getInstr()));
1279 if (!LRQ.valueIn()->isPHIDef())
1282 // Assume that a path spanning two iterations is a cycle, which could
1283 // overestimate in strange cases. This allows cyclic latency to be
1284 // estimated as the minimum slack of the vreg's depth or height.
1285 unsigned CyclicLatency = 0;
1286 if (LiveOutDepth > SU->getDepth())
1287 CyclicLatency = LiveOutDepth - SU->getDepth();
1289 unsigned LiveInHeight = SU->getHeight() + DefSU->Latency;
1290 if (LiveInHeight > LiveOutHeight) {
1291 if (LiveInHeight - LiveOutHeight < CyclicLatency)
1292 CyclicLatency = LiveInHeight - LiveOutHeight;
1296 DEBUG(dbgs() << "Cyclic Path: SU(" << DefSU->NodeNum << ") -> SU("
1297 << SU->NodeNum << ") = " << CyclicLatency << "c\n");
1298 if (CyclicLatency > MaxCyclicLatency)
1299 MaxCyclicLatency = CyclicLatency;
1302 DEBUG(dbgs() << "Cyclic Critical Path: " << MaxCyclicLatency << "c\n");
1303 return MaxCyclicLatency;
1306 /// Release ExitSU predecessors and setup scheduler queues. Re-position
1307 /// the Top RP tracker in case the region beginning has changed.
1308 void ScheduleDAGMILive::initQueues(ArrayRef<SUnit*> TopRoots,
1309 ArrayRef<SUnit*> BotRoots) {
1310 ScheduleDAGMI::initQueues(TopRoots, BotRoots);
1311 if (ShouldTrackPressure) {
1312 assert(TopRPTracker.getPos() == RegionBegin && "bad initial Top tracker");
1313 TopRPTracker.setPos(CurrentTop);
1317 /// Move an instruction and update register pressure.
1318 void ScheduleDAGMILive::scheduleMI(SUnit *SU, bool IsTopNode) {
1319 // Move the instruction to its new location in the instruction stream.
1320 MachineInstr *MI = SU->getInstr();
1323 assert(SU->isTopReady() && "node still has unscheduled dependencies");
1324 if (&*CurrentTop == MI)
1325 CurrentTop = nextIfDebug(++CurrentTop, CurrentBottom);
1327 moveInstruction(MI, CurrentTop);
1328 TopRPTracker.setPos(MI);
1331 if (ShouldTrackPressure) {
1332 // Update top scheduled pressure.
1333 RegisterOperands RegOpers;
1334 RegOpers.collect(*MI, *TRI, MRI, ShouldTrackLaneMasks, false);
1335 if (ShouldTrackLaneMasks) {
1336 // Adjust liveness and add missing dead+read-undef flags.
1337 SlotIndex SlotIdx = LIS->getInstructionIndex(*MI).getRegSlot();
1338 RegOpers.adjustLaneLiveness(*LIS, MRI, SlotIdx, MI);
1340 // Adjust for missing dead-def flags.
1341 RegOpers.detectDeadDefs(*MI, *LIS);
1344 TopRPTracker.advance(RegOpers);
1345 assert(TopRPTracker.getPos() == CurrentTop && "out of sync");
1347 dbgs() << "Top Pressure:\n";
1348 dumpRegSetPressure(TopRPTracker.getRegSetPressureAtPos(), TRI);
1351 updateScheduledPressure(SU, TopRPTracker.getPressure().MaxSetPressure);
1354 assert(SU->isBottomReady() && "node still has unscheduled dependencies");
1355 MachineBasicBlock::iterator priorII =
1356 priorNonDebug(CurrentBottom, CurrentTop);
1357 if (&*priorII == MI)
1358 CurrentBottom = priorII;
1360 if (&*CurrentTop == MI) {
1361 CurrentTop = nextIfDebug(++CurrentTop, priorII);
1362 TopRPTracker.setPos(CurrentTop);
1364 moveInstruction(MI, CurrentBottom);
1367 if (ShouldTrackPressure) {
1368 RegisterOperands RegOpers;
1369 RegOpers.collect(*MI, *TRI, MRI, ShouldTrackLaneMasks, false);
1370 if (ShouldTrackLaneMasks) {
1371 // Adjust liveness and add missing dead+read-undef flags.
1372 SlotIndex SlotIdx = LIS->getInstructionIndex(*MI).getRegSlot();
1373 RegOpers.adjustLaneLiveness(*LIS, MRI, SlotIdx, MI);
1375 // Adjust for missing dead-def flags.
1376 RegOpers.detectDeadDefs(*MI, *LIS);
1379 BotRPTracker.recedeSkipDebugValues();
1380 SmallVector<RegisterMaskPair, 8> LiveUses;
1381 BotRPTracker.recede(RegOpers, &LiveUses);
1382 assert(BotRPTracker.getPos() == CurrentBottom && "out of sync");
1384 dbgs() << "Bottom Pressure:\n";
1385 dumpRegSetPressure(BotRPTracker.getRegSetPressureAtPos(), TRI);
1388 updateScheduledPressure(SU, BotRPTracker.getPressure().MaxSetPressure);
1389 updatePressureDiffs(LiveUses);
1394 //===----------------------------------------------------------------------===//
1395 // BaseMemOpClusterMutation - DAG post-processing to cluster loads or stores.
1396 //===----------------------------------------------------------------------===//
1399 /// \brief Post-process the DAG to create cluster edges between neighboring
1400 /// loads or between neighboring stores.
1401 class BaseMemOpClusterMutation : public ScheduleDAGMutation {
1406 MemOpInfo(SUnit *su, unsigned reg, int64_t ofs)
1407 : SU(su), BaseReg(reg), Offset(ofs) {}
1409 bool operator<(const MemOpInfo&RHS) const {
1410 return std::tie(BaseReg, Offset, SU->NodeNum) <
1411 std::tie(RHS.BaseReg, RHS.Offset, RHS.SU->NodeNum);
1415 const TargetInstrInfo *TII;
1416 const TargetRegisterInfo *TRI;
1420 BaseMemOpClusterMutation(const TargetInstrInfo *tii,
1421 const TargetRegisterInfo *tri, bool IsLoad)
1422 : TII(tii), TRI(tri), IsLoad(IsLoad) {}
1424 void apply(ScheduleDAGInstrs *DAGInstrs) override;
1427 void clusterNeighboringMemOps(ArrayRef<SUnit *> MemOps, ScheduleDAGMI *DAG);
1430 class StoreClusterMutation : public BaseMemOpClusterMutation {
1432 StoreClusterMutation(const TargetInstrInfo *tii,
1433 const TargetRegisterInfo *tri)
1434 : BaseMemOpClusterMutation(tii, tri, false) {}
1437 class LoadClusterMutation : public BaseMemOpClusterMutation {
1439 LoadClusterMutation(const TargetInstrInfo *tii, const TargetRegisterInfo *tri)
1440 : BaseMemOpClusterMutation(tii, tri, true) {}
1446 std::unique_ptr<ScheduleDAGMutation>
1447 createLoadClusterDAGMutation(const TargetInstrInfo *TII,
1448 const TargetRegisterInfo *TRI) {
1449 return EnableMemOpCluster ? make_unique<LoadClusterMutation>(TII, TRI)
1453 std::unique_ptr<ScheduleDAGMutation>
1454 createStoreClusterDAGMutation(const TargetInstrInfo *TII,
1455 const TargetRegisterInfo *TRI) {
1456 return EnableMemOpCluster ? make_unique<StoreClusterMutation>(TII, TRI)
1462 void BaseMemOpClusterMutation::clusterNeighboringMemOps(
1463 ArrayRef<SUnit *> MemOps, ScheduleDAGMI *DAG) {
1464 SmallVector<MemOpInfo, 32> MemOpRecords;
1465 for (unsigned Idx = 0, End = MemOps.size(); Idx != End; ++Idx) {
1466 SUnit *SU = MemOps[Idx];
1469 if (TII->getMemOpBaseRegImmOfs(*SU->getInstr(), BaseReg, Offset, TRI))
1470 MemOpRecords.push_back(MemOpInfo(SU, BaseReg, Offset));
1472 if (MemOpRecords.size() < 2)
1475 std::sort(MemOpRecords.begin(), MemOpRecords.end());
1476 unsigned ClusterLength = 1;
1477 for (unsigned Idx = 0, End = MemOpRecords.size(); Idx < (End - 1); ++Idx) {
1478 if (MemOpRecords[Idx].BaseReg != MemOpRecords[Idx+1].BaseReg) {
1483 SUnit *SUa = MemOpRecords[Idx].SU;
1484 SUnit *SUb = MemOpRecords[Idx+1].SU;
1485 if (TII->shouldClusterMemOps(*SUa->getInstr(), *SUb->getInstr(),
1487 DAG->addEdge(SUb, SDep(SUa, SDep::Cluster))) {
1488 DEBUG(dbgs() << "Cluster ld/st SU(" << SUa->NodeNum << ") - SU("
1489 << SUb->NodeNum << ")\n");
1490 // Copy successor edges from SUa to SUb. Interleaving computation
1491 // dependent on SUa can prevent load combining due to register reuse.
1492 // Predecessor edges do not need to be copied from SUb to SUa since nearby
1493 // loads should have effectively the same inputs.
1494 for (SUnit::const_succ_iterator
1495 SI = SUa->Succs.begin(), SE = SUa->Succs.end(); SI != SE; ++SI) {
1496 if (SI->getSUnit() == SUb)
1498 DEBUG(dbgs() << " Copy Succ SU(" << SI->getSUnit()->NodeNum << ")\n");
1499 DAG->addEdge(SI->getSUnit(), SDep(SUb, SDep::Artificial));
1507 /// \brief Callback from DAG postProcessing to create cluster edges for loads.
1508 void BaseMemOpClusterMutation::apply(ScheduleDAGInstrs *DAGInstrs) {
1510 ScheduleDAGMI *DAG = static_cast<ScheduleDAGMI*>(DAGInstrs);
1512 // Map DAG NodeNum to store chain ID.
1513 DenseMap<unsigned, unsigned> StoreChainIDs;
1514 // Map each store chain to a set of dependent MemOps.
1515 SmallVector<SmallVector<SUnit*,4>, 32> StoreChainDependents;
1516 for (unsigned Idx = 0, End = DAG->SUnits.size(); Idx != End; ++Idx) {
1517 SUnit *SU = &DAG->SUnits[Idx];
1518 if ((IsLoad && !SU->getInstr()->mayLoad()) ||
1519 (!IsLoad && !SU->getInstr()->mayStore()))
1522 unsigned ChainPredID = DAG->SUnits.size();
1523 for (SUnit::const_pred_iterator
1524 PI = SU->Preds.begin(), PE = SU->Preds.end(); PI != PE; ++PI) {
1526 ChainPredID = PI->getSUnit()->NodeNum;
1530 // Check if this chain-like pred has been seen
1531 // before. ChainPredID==MaxNodeID at the top of the schedule.
1532 unsigned NumChains = StoreChainDependents.size();
1533 std::pair<DenseMap<unsigned, unsigned>::iterator, bool> Result =
1534 StoreChainIDs.insert(std::make_pair(ChainPredID, NumChains));
1536 StoreChainDependents.resize(NumChains + 1);
1537 StoreChainDependents[Result.first->second].push_back(SU);
1540 // Iterate over the store chains.
1541 for (unsigned Idx = 0, End = StoreChainDependents.size(); Idx != End; ++Idx)
1542 clusterNeighboringMemOps(StoreChainDependents[Idx], DAG);
1545 //===----------------------------------------------------------------------===//
1546 // MacroFusion - DAG post-processing to encourage fusion of macro ops.
1547 //===----------------------------------------------------------------------===//
1550 /// \brief Post-process the DAG to create cluster edges between instructions
1551 /// that may be fused by the processor into a single operation.
1552 class MacroFusion : public ScheduleDAGMutation {
1553 const TargetInstrInfo &TII;
1555 MacroFusion(const TargetInstrInfo &TII)
1558 void apply(ScheduleDAGInstrs *DAGInstrs) override;
1564 std::unique_ptr<ScheduleDAGMutation>
1565 createMacroFusionDAGMutation(const TargetInstrInfo *TII) {
1566 return EnableMacroFusion ? make_unique<MacroFusion>(*TII) : nullptr;
1571 /// \brief Callback from DAG postProcessing to create cluster edges to encourage
1572 /// fused operations.
1573 void MacroFusion::apply(ScheduleDAGInstrs *DAGInstrs) {
1574 ScheduleDAGMI *DAG = static_cast<ScheduleDAGMI*>(DAGInstrs);
1576 // For now, assume targets can only fuse with the branch.
1577 SUnit &ExitSU = DAG->ExitSU;
1578 MachineInstr *Branch = ExitSU.getInstr();
1582 for (SDep &PredDep : ExitSU.Preds) {
1583 if (PredDep.isWeak())
1585 SUnit &SU = *PredDep.getSUnit();
1586 MachineInstr &Pred = *SU.getInstr();
1587 if (!TII.shouldScheduleAdjacent(Pred, *Branch))
1590 // Create a single weak edge from SU to ExitSU. The only effect is to cause
1591 // bottom-up scheduling to heavily prioritize the clustered SU. There is no
1592 // need to copy predecessor edges from ExitSU to SU, since top-down
1593 // scheduling cannot prioritize ExitSU anyway. To defer top-down scheduling
1594 // of SU, we could create an artificial edge from the deepest root, but it
1595 // hasn't been needed yet.
1596 bool Success = DAG->addEdge(&ExitSU, SDep(&SU, SDep::Cluster));
1598 assert(Success && "No DAG nodes should be reachable from ExitSU");
1600 // Adjust latency of data deps between the nodes.
1601 for (SDep &PredDep : ExitSU.Preds) {
1602 if (PredDep.getSUnit() == &SU)
1603 PredDep.setLatency(0);
1605 for (SDep &SuccDep : SU.Succs) {
1606 if (SuccDep.getSUnit() == &ExitSU)
1607 SuccDep.setLatency(0);
1610 DEBUG(dbgs() << "Macro Fuse SU(" << SU.NodeNum << ")\n");
1615 //===----------------------------------------------------------------------===//
1616 // CopyConstrain - DAG post-processing to encourage copy elimination.
1617 //===----------------------------------------------------------------------===//
1620 /// \brief Post-process the DAG to create weak edges from all uses of a copy to
1621 /// the one use that defines the copy's source vreg, most likely an induction
1622 /// variable increment.
1623 class CopyConstrain : public ScheduleDAGMutation {
1625 SlotIndex RegionBeginIdx;
1626 // RegionEndIdx is the slot index of the last non-debug instruction in the
1627 // scheduling region. So we may have RegionBeginIdx == RegionEndIdx.
1628 SlotIndex RegionEndIdx;
1630 CopyConstrain(const TargetInstrInfo *, const TargetRegisterInfo *) {}
1632 void apply(ScheduleDAGInstrs *DAGInstrs) override;
1635 void constrainLocalCopy(SUnit *CopySU, ScheduleDAGMILive *DAG);
1641 std::unique_ptr<ScheduleDAGMutation>
1642 createCopyConstrainDAGMutation(const TargetInstrInfo *TII,
1643 const TargetRegisterInfo *TRI) {
1644 return make_unique<CopyConstrain>(TII, TRI);
1649 /// constrainLocalCopy handles two possibilities:
1654 /// I3: dst = src (copy)
1655 /// (create pred->succ edges I0->I1, I2->I1)
1658 /// I0: dst = src (copy)
1662 /// (create pred->succ edges I1->I2, I3->I2)
1664 /// Although the MachineScheduler is currently constrained to single blocks,
1665 /// this algorithm should handle extended blocks. An EBB is a set of
1666 /// contiguously numbered blocks such that the previous block in the EBB is
1667 /// always the single predecessor.
1668 void CopyConstrain::constrainLocalCopy(SUnit *CopySU, ScheduleDAGMILive *DAG) {
1669 LiveIntervals *LIS = DAG->getLIS();
1670 MachineInstr *Copy = CopySU->getInstr();
1672 // Check for pure vreg copies.
1673 const MachineOperand &SrcOp = Copy->getOperand(1);
1674 unsigned SrcReg = SrcOp.getReg();
1675 if (!TargetRegisterInfo::isVirtualRegister(SrcReg) || !SrcOp.readsReg())
1678 const MachineOperand &DstOp = Copy->getOperand(0);
1679 unsigned DstReg = DstOp.getReg();
1680 if (!TargetRegisterInfo::isVirtualRegister(DstReg) || DstOp.isDead())
1683 // Check if either the dest or source is local. If it's live across a back
1684 // edge, it's not local. Note that if both vregs are live across the back
1685 // edge, we cannot successfully contrain the copy without cyclic scheduling.
1686 // If both the copy's source and dest are local live intervals, then we
1687 // should treat the dest as the global for the purpose of adding
1688 // constraints. This adds edges from source's other uses to the copy.
1689 unsigned LocalReg = SrcReg;
1690 unsigned GlobalReg = DstReg;
1691 LiveInterval *LocalLI = &LIS->getInterval(LocalReg);
1692 if (!LocalLI->isLocal(RegionBeginIdx, RegionEndIdx)) {
1695 LocalLI = &LIS->getInterval(LocalReg);
1696 if (!LocalLI->isLocal(RegionBeginIdx, RegionEndIdx))
1699 LiveInterval *GlobalLI = &LIS->getInterval(GlobalReg);
1701 // Find the global segment after the start of the local LI.
1702 LiveInterval::iterator GlobalSegment = GlobalLI->find(LocalLI->beginIndex());
1703 // If GlobalLI does not overlap LocalLI->start, then a copy directly feeds a
1704 // local live range. We could create edges from other global uses to the local
1705 // start, but the coalescer should have already eliminated these cases, so
1706 // don't bother dealing with it.
1707 if (GlobalSegment == GlobalLI->end())
1710 // If GlobalSegment is killed at the LocalLI->start, the call to find()
1711 // returned the next global segment. But if GlobalSegment overlaps with
1712 // LocalLI->start, then advance to the next segement. If a hole in GlobalLI
1713 // exists in LocalLI's vicinity, GlobalSegment will be the end of the hole.
1714 if (GlobalSegment->contains(LocalLI->beginIndex()))
1717 if (GlobalSegment == GlobalLI->end())
1720 // Check if GlobalLI contains a hole in the vicinity of LocalLI.
1721 if (GlobalSegment != GlobalLI->begin()) {
1722 // Two address defs have no hole.
1723 if (SlotIndex::isSameInstr(std::prev(GlobalSegment)->end,
1724 GlobalSegment->start)) {
1727 // If the prior global segment may be defined by the same two-address
1728 // instruction that also defines LocalLI, then can't make a hole here.
1729 if (SlotIndex::isSameInstr(std::prev(GlobalSegment)->start,
1730 LocalLI->beginIndex())) {
1733 // If GlobalLI has a prior segment, it must be live into the EBB. Otherwise
1734 // it would be a disconnected component in the live range.
1735 assert(std::prev(GlobalSegment)->start < LocalLI->beginIndex() &&
1736 "Disconnected LRG within the scheduling region.");
1738 MachineInstr *GlobalDef = LIS->getInstructionFromIndex(GlobalSegment->start);
1742 SUnit *GlobalSU = DAG->getSUnit(GlobalDef);
1746 // GlobalDef is the bottom of the GlobalLI hole. Open the hole by
1747 // constraining the uses of the last local def to precede GlobalDef.
1748 SmallVector<SUnit*,8> LocalUses;
1749 const VNInfo *LastLocalVN = LocalLI->getVNInfoBefore(LocalLI->endIndex());
1750 MachineInstr *LastLocalDef = LIS->getInstructionFromIndex(LastLocalVN->def);
1751 SUnit *LastLocalSU = DAG->getSUnit(LastLocalDef);
1752 for (SUnit::const_succ_iterator
1753 I = LastLocalSU->Succs.begin(), E = LastLocalSU->Succs.end();
1755 if (I->getKind() != SDep::Data || I->getReg() != LocalReg)
1757 if (I->getSUnit() == GlobalSU)
1759 if (!DAG->canAddEdge(GlobalSU, I->getSUnit()))
1761 LocalUses.push_back(I->getSUnit());
1763 // Open the top of the GlobalLI hole by constraining any earlier global uses
1764 // to precede the start of LocalLI.
1765 SmallVector<SUnit*,8> GlobalUses;
1766 MachineInstr *FirstLocalDef =
1767 LIS->getInstructionFromIndex(LocalLI->beginIndex());
1768 SUnit *FirstLocalSU = DAG->getSUnit(FirstLocalDef);
1769 for (SUnit::const_pred_iterator
1770 I = GlobalSU->Preds.begin(), E = GlobalSU->Preds.end(); I != E; ++I) {
1771 if (I->getKind() != SDep::Anti || I->getReg() != GlobalReg)
1773 if (I->getSUnit() == FirstLocalSU)
1775 if (!DAG->canAddEdge(FirstLocalSU, I->getSUnit()))
1777 GlobalUses.push_back(I->getSUnit());
1779 DEBUG(dbgs() << "Constraining copy SU(" << CopySU->NodeNum << ")\n");
1780 // Add the weak edges.
1781 for (SmallVectorImpl<SUnit*>::const_iterator
1782 I = LocalUses.begin(), E = LocalUses.end(); I != E; ++I) {
1783 DEBUG(dbgs() << " Local use SU(" << (*I)->NodeNum << ") -> SU("
1784 << GlobalSU->NodeNum << ")\n");
1785 DAG->addEdge(GlobalSU, SDep(*I, SDep::Weak));
1787 for (SmallVectorImpl<SUnit*>::const_iterator
1788 I = GlobalUses.begin(), E = GlobalUses.end(); I != E; ++I) {
1789 DEBUG(dbgs() << " Global use SU(" << (*I)->NodeNum << ") -> SU("
1790 << FirstLocalSU->NodeNum << ")\n");
1791 DAG->addEdge(FirstLocalSU, SDep(*I, SDep::Weak));
1795 /// \brief Callback from DAG postProcessing to create weak edges to encourage
1796 /// copy elimination.
1797 void CopyConstrain::apply(ScheduleDAGInstrs *DAGInstrs) {
1798 ScheduleDAGMI *DAG = static_cast<ScheduleDAGMI*>(DAGInstrs);
1799 assert(DAG->hasVRegLiveness() && "Expect VRegs with LiveIntervals");
1801 MachineBasicBlock::iterator FirstPos = nextIfDebug(DAG->begin(), DAG->end());
1802 if (FirstPos == DAG->end())
1804 RegionBeginIdx = DAG->getLIS()->getInstructionIndex(*FirstPos);
1805 RegionEndIdx = DAG->getLIS()->getInstructionIndex(
1806 *priorNonDebug(DAG->end(), DAG->begin()));
1808 for (unsigned Idx = 0, End = DAG->SUnits.size(); Idx != End; ++Idx) {
1809 SUnit *SU = &DAG->SUnits[Idx];
1810 if (!SU->getInstr()->isCopy())
1813 constrainLocalCopy(SU, static_cast<ScheduleDAGMILive*>(DAG));
1817 //===----------------------------------------------------------------------===//
1818 // MachineSchedStrategy helpers used by GenericScheduler, GenericPostScheduler
1819 // and possibly other custom schedulers.
1820 //===----------------------------------------------------------------------===//
1822 static const unsigned InvalidCycle = ~0U;
1824 SchedBoundary::~SchedBoundary() { delete HazardRec; }
1826 void SchedBoundary::reset() {
1827 // A new HazardRec is created for each DAG and owned by SchedBoundary.
1828 // Destroying and reconstructing it is very expensive though. So keep
1829 // invalid, placeholder HazardRecs.
1830 if (HazardRec && HazardRec->isEnabled()) {
1832 HazardRec = nullptr;
1836 CheckPending = false;
1839 MinReadyCycle = UINT_MAX;
1840 ExpectedLatency = 0;
1841 DependentLatency = 0;
1843 MaxExecutedResCount = 0;
1845 IsResourceLimited = false;
1846 ReservedCycles.clear();
1848 // Track the maximum number of stall cycles that could arise either from the
1849 // latency of a DAG edge or the number of cycles that a processor resource is
1850 // reserved (SchedBoundary::ReservedCycles).
1851 MaxObservedStall = 0;
1853 // Reserve a zero-count for invalid CritResIdx.
1854 ExecutedResCounts.resize(1);
1855 assert(!ExecutedResCounts[0] && "nonzero count for bad resource");
1858 void SchedRemainder::
1859 init(ScheduleDAGMI *DAG, const TargetSchedModel *SchedModel) {
1861 if (!SchedModel->hasInstrSchedModel())
1863 RemainingCounts.resize(SchedModel->getNumProcResourceKinds());
1864 for (std::vector<SUnit>::iterator
1865 I = DAG->SUnits.begin(), E = DAG->SUnits.end(); I != E; ++I) {
1866 const MCSchedClassDesc *SC = DAG->getSchedClass(&*I);
1867 RemIssueCount += SchedModel->getNumMicroOps(I->getInstr(), SC)
1868 * SchedModel->getMicroOpFactor();
1869 for (TargetSchedModel::ProcResIter
1870 PI = SchedModel->getWriteProcResBegin(SC),
1871 PE = SchedModel->getWriteProcResEnd(SC); PI != PE; ++PI) {
1872 unsigned PIdx = PI->ProcResourceIdx;
1873 unsigned Factor = SchedModel->getResourceFactor(PIdx);
1874 RemainingCounts[PIdx] += (Factor * PI->Cycles);
1879 void SchedBoundary::
1880 init(ScheduleDAGMI *dag, const TargetSchedModel *smodel, SchedRemainder *rem) {
1883 SchedModel = smodel;
1885 if (SchedModel->hasInstrSchedModel()) {
1886 ExecutedResCounts.resize(SchedModel->getNumProcResourceKinds());
1887 ReservedCycles.resize(SchedModel->getNumProcResourceKinds(), InvalidCycle);
1891 /// Compute the stall cycles based on this SUnit's ready time. Heuristics treat
1892 /// these "soft stalls" differently than the hard stall cycles based on CPU
1893 /// resources and computed by checkHazard(). A fully in-order model
1894 /// (MicroOpBufferSize==0) will not make use of this since instructions are not
1895 /// available for scheduling until they are ready. However, a weaker in-order
1896 /// model may use this for heuristics. For example, if a processor has in-order
1897 /// behavior when reading certain resources, this may come into play.
1898 unsigned SchedBoundary::getLatencyStallCycles(SUnit *SU) {
1899 if (!SU->isUnbuffered)
1902 unsigned ReadyCycle = (isTop() ? SU->TopReadyCycle : SU->BotReadyCycle);
1903 if (ReadyCycle > CurrCycle)
1904 return ReadyCycle - CurrCycle;
1908 /// Compute the next cycle at which the given processor resource can be
1910 unsigned SchedBoundary::
1911 getNextResourceCycle(unsigned PIdx, unsigned Cycles) {
1912 unsigned NextUnreserved = ReservedCycles[PIdx];
1913 // If this resource has never been used, always return cycle zero.
1914 if (NextUnreserved == InvalidCycle)
1916 // For bottom-up scheduling add the cycles needed for the current operation.
1918 NextUnreserved += Cycles;
1919 return NextUnreserved;
1922 /// Does this SU have a hazard within the current instruction group.
1924 /// The scheduler supports two modes of hazard recognition. The first is the
1925 /// ScheduleHazardRecognizer API. It is a fully general hazard recognizer that
1926 /// supports highly complicated in-order reservation tables
1927 /// (ScoreboardHazardRecognizer) and arbitraty target-specific logic.
1929 /// The second is a streamlined mechanism that checks for hazards based on
1930 /// simple counters that the scheduler itself maintains. It explicitly checks
1931 /// for instruction dispatch limitations, including the number of micro-ops that
1932 /// can dispatch per cycle.
1934 /// TODO: Also check whether the SU must start a new group.
1935 bool SchedBoundary::checkHazard(SUnit *SU) {
1936 if (HazardRec->isEnabled()
1937 && HazardRec->getHazardType(SU) != ScheduleHazardRecognizer::NoHazard) {
1940 unsigned uops = SchedModel->getNumMicroOps(SU->getInstr());
1941 if ((CurrMOps > 0) && (CurrMOps + uops > SchedModel->getIssueWidth())) {
1942 DEBUG(dbgs() << " SU(" << SU->NodeNum << ") uops="
1943 << SchedModel->getNumMicroOps(SU->getInstr()) << '\n');
1946 if (SchedModel->hasInstrSchedModel() && SU->hasReservedResource) {
1947 const MCSchedClassDesc *SC = DAG->getSchedClass(SU);
1948 for (TargetSchedModel::ProcResIter
1949 PI = SchedModel->getWriteProcResBegin(SC),
1950 PE = SchedModel->getWriteProcResEnd(SC); PI != PE; ++PI) {
1951 unsigned NRCycle = getNextResourceCycle(PI->ProcResourceIdx, PI->Cycles);
1952 if (NRCycle > CurrCycle) {
1954 MaxObservedStall = std::max(PI->Cycles, MaxObservedStall);
1956 DEBUG(dbgs() << " SU(" << SU->NodeNum << ") "
1957 << SchedModel->getResourceName(PI->ProcResourceIdx)
1958 << "=" << NRCycle << "c\n");
1966 // Find the unscheduled node in ReadySUs with the highest latency.
1967 unsigned SchedBoundary::
1968 findMaxLatency(ArrayRef<SUnit*> ReadySUs) {
1969 SUnit *LateSU = nullptr;
1970 unsigned RemLatency = 0;
1971 for (ArrayRef<SUnit*>::iterator I = ReadySUs.begin(), E = ReadySUs.end();
1973 unsigned L = getUnscheduledLatency(*I);
1974 if (L > RemLatency) {
1980 DEBUG(dbgs() << Available.getName() << " RemLatency SU("
1981 << LateSU->NodeNum << ") " << RemLatency << "c\n");
1986 // Count resources in this zone and the remaining unscheduled
1987 // instruction. Return the max count, scaled. Set OtherCritIdx to the critical
1988 // resource index, or zero if the zone is issue limited.
1989 unsigned SchedBoundary::
1990 getOtherResourceCount(unsigned &OtherCritIdx) {
1992 if (!SchedModel->hasInstrSchedModel())
1995 unsigned OtherCritCount = Rem->RemIssueCount
1996 + (RetiredMOps * SchedModel->getMicroOpFactor());
1997 DEBUG(dbgs() << " " << Available.getName() << " + Remain MOps: "
1998 << OtherCritCount / SchedModel->getMicroOpFactor() << '\n');
1999 for (unsigned PIdx = 1, PEnd = SchedModel->getNumProcResourceKinds();
2000 PIdx != PEnd; ++PIdx) {
2001 unsigned OtherCount = getResourceCount(PIdx) + Rem->RemainingCounts[PIdx];
2002 if (OtherCount > OtherCritCount) {
2003 OtherCritCount = OtherCount;
2004 OtherCritIdx = PIdx;
2008 DEBUG(dbgs() << " " << Available.getName() << " + Remain CritRes: "
2009 << OtherCritCount / SchedModel->getResourceFactor(OtherCritIdx)
2010 << " " << SchedModel->getResourceName(OtherCritIdx) << "\n");
2012 return OtherCritCount;
2015 void SchedBoundary::releaseNode(SUnit *SU, unsigned ReadyCycle) {
2016 assert(SU->getInstr() && "Scheduled SUnit must have instr");
2019 // ReadyCycle was been bumped up to the CurrCycle when this node was
2020 // scheduled, but CurrCycle may have been eagerly advanced immediately after
2021 // scheduling, so may now be greater than ReadyCycle.
2022 if (ReadyCycle > CurrCycle)
2023 MaxObservedStall = std::max(ReadyCycle - CurrCycle, MaxObservedStall);
2026 if (ReadyCycle < MinReadyCycle)
2027 MinReadyCycle = ReadyCycle;
2029 // Check for interlocks first. For the purpose of other heuristics, an
2030 // instruction that cannot issue appears as if it's not in the ReadyQueue.
2031 bool IsBuffered = SchedModel->getMicroOpBufferSize() != 0;
2032 if ((!IsBuffered && ReadyCycle > CurrCycle) || checkHazard(SU) ||
2033 Available.size() >= ReadyListLimit)
2039 /// Move the boundary of scheduled code by one cycle.
2040 void SchedBoundary::bumpCycle(unsigned NextCycle) {
2041 if (SchedModel->getMicroOpBufferSize() == 0) {
2042 assert(MinReadyCycle < UINT_MAX && "MinReadyCycle uninitialized");
2043 if (MinReadyCycle > NextCycle)
2044 NextCycle = MinReadyCycle;
2046 // Update the current micro-ops, which will issue in the next cycle.
2047 unsigned DecMOps = SchedModel->getIssueWidth() * (NextCycle - CurrCycle);
2048 CurrMOps = (CurrMOps <= DecMOps) ? 0 : CurrMOps - DecMOps;
2050 // Decrement DependentLatency based on the next cycle.
2051 if ((NextCycle - CurrCycle) > DependentLatency)
2052 DependentLatency = 0;
2054 DependentLatency -= (NextCycle - CurrCycle);
2056 if (!HazardRec->isEnabled()) {
2057 // Bypass HazardRec virtual calls.
2058 CurrCycle = NextCycle;
2060 // Bypass getHazardType calls in case of long latency.
2061 for (; CurrCycle != NextCycle; ++CurrCycle) {
2063 HazardRec->AdvanceCycle();
2065 HazardRec->RecedeCycle();
2068 CheckPending = true;
2069 unsigned LFactor = SchedModel->getLatencyFactor();
2071 (int)(getCriticalCount() - (getScheduledLatency() * LFactor))
2074 DEBUG(dbgs() << "Cycle: " << CurrCycle << ' ' << Available.getName() << '\n');
2077 void SchedBoundary::incExecutedResources(unsigned PIdx, unsigned Count) {
2078 ExecutedResCounts[PIdx] += Count;
2079 if (ExecutedResCounts[PIdx] > MaxExecutedResCount)
2080 MaxExecutedResCount = ExecutedResCounts[PIdx];
2083 /// Add the given processor resource to this scheduled zone.
2085 /// \param Cycles indicates the number of consecutive (non-pipelined) cycles
2086 /// during which this resource is consumed.
2088 /// \return the next cycle at which the instruction may execute without
2089 /// oversubscribing resources.
2090 unsigned SchedBoundary::
2091 countResource(unsigned PIdx, unsigned Cycles, unsigned NextCycle) {
2092 unsigned Factor = SchedModel->getResourceFactor(PIdx);
2093 unsigned Count = Factor * Cycles;
2094 DEBUG(dbgs() << " " << SchedModel->getResourceName(PIdx)
2095 << " +" << Cycles << "x" << Factor << "u\n");
2097 // Update Executed resources counts.
2098 incExecutedResources(PIdx, Count);
2099 assert(Rem->RemainingCounts[PIdx] >= Count && "resource double counted");
2100 Rem->RemainingCounts[PIdx] -= Count;
2102 // Check if this resource exceeds the current critical resource. If so, it
2103 // becomes the critical resource.
2104 if (ZoneCritResIdx != PIdx && (getResourceCount(PIdx) > getCriticalCount())) {
2105 ZoneCritResIdx = PIdx;
2106 DEBUG(dbgs() << " *** Critical resource "
2107 << SchedModel->getResourceName(PIdx) << ": "
2108 << getResourceCount(PIdx) / SchedModel->getLatencyFactor() << "c\n");
2110 // For reserved resources, record the highest cycle using the resource.
2111 unsigned NextAvailable = getNextResourceCycle(PIdx, Cycles);
2112 if (NextAvailable > CurrCycle) {
2113 DEBUG(dbgs() << " Resource conflict: "
2114 << SchedModel->getProcResource(PIdx)->Name << " reserved until @"
2115 << NextAvailable << "\n");
2117 return NextAvailable;
2120 /// Move the boundary of scheduled code by one SUnit.
2121 void SchedBoundary::bumpNode(SUnit *SU) {
2122 // Update the reservation table.
2123 if (HazardRec->isEnabled()) {
2124 if (!isTop() && SU->isCall) {
2125 // Calls are scheduled with their preceding instructions. For bottom-up
2126 // scheduling, clear the pipeline state before emitting.
2129 HazardRec->EmitInstruction(SU);
2131 // checkHazard should prevent scheduling multiple instructions per cycle that
2132 // exceed the issue width.
2133 const MCSchedClassDesc *SC = DAG->getSchedClass(SU);
2134 unsigned IncMOps = SchedModel->getNumMicroOps(SU->getInstr());
2136 (CurrMOps == 0 || (CurrMOps + IncMOps) <= SchedModel->getIssueWidth()) &&
2137 "Cannot schedule this instruction's MicroOps in the current cycle.");
2139 unsigned ReadyCycle = (isTop() ? SU->TopReadyCycle : SU->BotReadyCycle);
2140 DEBUG(dbgs() << " Ready @" << ReadyCycle << "c\n");
2142 unsigned NextCycle = CurrCycle;
2143 switch (SchedModel->getMicroOpBufferSize()) {
2145 assert(ReadyCycle <= CurrCycle && "Broken PendingQueue");
2148 if (ReadyCycle > NextCycle) {
2149 NextCycle = ReadyCycle;
2150 DEBUG(dbgs() << " *** Stall until: " << ReadyCycle << "\n");
2154 // We don't currently model the OOO reorder buffer, so consider all
2155 // scheduled MOps to be "retired". We do loosely model in-order resource
2156 // latency. If this instruction uses an in-order resource, account for any
2157 // likely stall cycles.
2158 if (SU->isUnbuffered && ReadyCycle > NextCycle)
2159 NextCycle = ReadyCycle;
2162 RetiredMOps += IncMOps;
2164 // Update resource counts and critical resource.
2165 if (SchedModel->hasInstrSchedModel()) {
2166 unsigned DecRemIssue = IncMOps * SchedModel->getMicroOpFactor();
2167 assert(Rem->RemIssueCount >= DecRemIssue && "MOps double counted");
2168 Rem->RemIssueCount -= DecRemIssue;
2169 if (ZoneCritResIdx) {
2170 // Scale scheduled micro-ops for comparing with the critical resource.
2171 unsigned ScaledMOps =
2172 RetiredMOps * SchedModel->getMicroOpFactor();
2174 // If scaled micro-ops are now more than the previous critical resource by
2175 // a full cycle, then micro-ops issue becomes critical.
2176 if ((int)(ScaledMOps - getResourceCount(ZoneCritResIdx))
2177 >= (int)SchedModel->getLatencyFactor()) {
2179 DEBUG(dbgs() << " *** Critical resource NumMicroOps: "
2180 << ScaledMOps / SchedModel->getLatencyFactor() << "c\n");
2183 for (TargetSchedModel::ProcResIter
2184 PI = SchedModel->getWriteProcResBegin(SC),
2185 PE = SchedModel->getWriteProcResEnd(SC); PI != PE; ++PI) {
2187 countResource(PI->ProcResourceIdx, PI->Cycles, NextCycle);
2188 if (RCycle > NextCycle)
2191 if (SU->hasReservedResource) {
2192 // For reserved resources, record the highest cycle using the resource.
2193 // For top-down scheduling, this is the cycle in which we schedule this
2194 // instruction plus the number of cycles the operations reserves the
2195 // resource. For bottom-up is it simply the instruction's cycle.
2196 for (TargetSchedModel::ProcResIter
2197 PI = SchedModel->getWriteProcResBegin(SC),
2198 PE = SchedModel->getWriteProcResEnd(SC); PI != PE; ++PI) {
2199 unsigned PIdx = PI->ProcResourceIdx;
2200 if (SchedModel->getProcResource(PIdx)->BufferSize == 0) {
2202 ReservedCycles[PIdx] =
2203 std::max(getNextResourceCycle(PIdx, 0), NextCycle + PI->Cycles);
2206 ReservedCycles[PIdx] = NextCycle;
2211 // Update ExpectedLatency and DependentLatency.
2212 unsigned &TopLatency = isTop() ? ExpectedLatency : DependentLatency;
2213 unsigned &BotLatency = isTop() ? DependentLatency : ExpectedLatency;
2214 if (SU->getDepth() > TopLatency) {
2215 TopLatency = SU->getDepth();
2216 DEBUG(dbgs() << " " << Available.getName()
2217 << " TopLatency SU(" << SU->NodeNum << ") " << TopLatency << "c\n");
2219 if (SU->getHeight() > BotLatency) {
2220 BotLatency = SU->getHeight();
2221 DEBUG(dbgs() << " " << Available.getName()
2222 << " BotLatency SU(" << SU->NodeNum << ") " << BotLatency << "c\n");
2224 // If we stall for any reason, bump the cycle.
2225 if (NextCycle > CurrCycle) {
2226 bumpCycle(NextCycle);
2228 // After updating ZoneCritResIdx and ExpectedLatency, check if we're
2229 // resource limited. If a stall occurred, bumpCycle does this.
2230 unsigned LFactor = SchedModel->getLatencyFactor();
2232 (int)(getCriticalCount() - (getScheduledLatency() * LFactor))
2235 // Update CurrMOps after calling bumpCycle to handle stalls, since bumpCycle
2236 // resets CurrMOps. Loop to handle instructions with more MOps than issue in
2237 // one cycle. Since we commonly reach the max MOps here, opportunistically
2238 // bump the cycle to avoid uselessly checking everything in the readyQ.
2239 CurrMOps += IncMOps;
2240 while (CurrMOps >= SchedModel->getIssueWidth()) {
2241 DEBUG(dbgs() << " *** Max MOps " << CurrMOps
2242 << " at cycle " << CurrCycle << '\n');
2243 bumpCycle(++NextCycle);
2245 DEBUG(dumpScheduledState());
2248 /// Release pending ready nodes in to the available queue. This makes them
2249 /// visible to heuristics.
2250 void SchedBoundary::releasePending() {
2251 // If the available queue is empty, it is safe to reset MinReadyCycle.
2252 if (Available.empty())
2253 MinReadyCycle = UINT_MAX;
2255 // Check to see if any of the pending instructions are ready to issue. If
2256 // so, add them to the available queue.
2257 bool IsBuffered = SchedModel->getMicroOpBufferSize() != 0;
2258 for (unsigned i = 0, e = Pending.size(); i != e; ++i) {
2259 SUnit *SU = *(Pending.begin()+i);
2260 unsigned ReadyCycle = isTop() ? SU->TopReadyCycle : SU->BotReadyCycle;
2262 if (ReadyCycle < MinReadyCycle)
2263 MinReadyCycle = ReadyCycle;
2265 if (!IsBuffered && ReadyCycle > CurrCycle)
2268 if (checkHazard(SU))
2271 if (Available.size() >= ReadyListLimit)
2275 Pending.remove(Pending.begin()+i);
2278 CheckPending = false;
2281 /// Remove SU from the ready set for this boundary.
2282 void SchedBoundary::removeReady(SUnit *SU) {
2283 if (Available.isInQueue(SU))
2284 Available.remove(Available.find(SU));
2286 assert(Pending.isInQueue(SU) && "bad ready count");
2287 Pending.remove(Pending.find(SU));
2291 /// If this queue only has one ready candidate, return it. As a side effect,
2292 /// defer any nodes that now hit a hazard, and advance the cycle until at least
2293 /// one node is ready. If multiple instructions are ready, return NULL.
2294 SUnit *SchedBoundary::pickOnlyChoice() {
2299 // Defer any ready instrs that now have a hazard.
2300 for (ReadyQueue::iterator I = Available.begin(); I != Available.end();) {
2301 if (checkHazard(*I)) {
2303 I = Available.remove(I);
2309 for (unsigned i = 0; Available.empty(); ++i) {
2310 // FIXME: Re-enable assert once PR20057 is resolved.
2311 // assert(i <= (HazardRec->getMaxLookAhead() + MaxObservedStall) &&
2312 // "permanent hazard");
2314 bumpCycle(CurrCycle + 1);
2318 DEBUG(Pending.dump());
2319 DEBUG(Available.dump());
2321 if (Available.size() == 1)
2322 return *Available.begin();
2327 // This is useful information to dump after bumpNode.
2328 // Note that the Queue contents are more useful before pickNodeFromQueue.
2329 void SchedBoundary::dumpScheduledState() {
2332 if (ZoneCritResIdx) {
2333 ResFactor = SchedModel->getResourceFactor(ZoneCritResIdx);
2334 ResCount = getResourceCount(ZoneCritResIdx);
2336 ResFactor = SchedModel->getMicroOpFactor();
2337 ResCount = RetiredMOps * SchedModel->getMicroOpFactor();
2339 unsigned LFactor = SchedModel->getLatencyFactor();
2340 dbgs() << Available.getName() << " @" << CurrCycle << "c\n"
2341 << " Retired: " << RetiredMOps;
2342 dbgs() << "\n Executed: " << getExecutedCount() / LFactor << "c";
2343 dbgs() << "\n Critical: " << ResCount / LFactor << "c, "
2344 << ResCount / ResFactor << " "
2345 << SchedModel->getResourceName(ZoneCritResIdx)
2346 << "\n ExpectedLatency: " << ExpectedLatency << "c\n"
2347 << (IsResourceLimited ? " - Resource" : " - Latency")
2352 //===----------------------------------------------------------------------===//
2353 // GenericScheduler - Generic implementation of MachineSchedStrategy.
2354 //===----------------------------------------------------------------------===//
2356 void GenericSchedulerBase::SchedCandidate::
2357 initResourceDelta(const ScheduleDAGMI *DAG,
2358 const TargetSchedModel *SchedModel) {
2359 if (!Policy.ReduceResIdx && !Policy.DemandResIdx)
2362 const MCSchedClassDesc *SC = DAG->getSchedClass(SU);
2363 for (TargetSchedModel::ProcResIter
2364 PI = SchedModel->getWriteProcResBegin(SC),
2365 PE = SchedModel->getWriteProcResEnd(SC); PI != PE; ++PI) {
2366 if (PI->ProcResourceIdx == Policy.ReduceResIdx)
2367 ResDelta.CritResources += PI->Cycles;
2368 if (PI->ProcResourceIdx == Policy.DemandResIdx)
2369 ResDelta.DemandedResources += PI->Cycles;
2373 /// Set the CandPolicy given a scheduling zone given the current resources and
2374 /// latencies inside and outside the zone.
2375 void GenericSchedulerBase::setPolicy(CandPolicy &Policy, bool IsPostRA,
2376 SchedBoundary &CurrZone,
2377 SchedBoundary *OtherZone) {
2378 // Apply preemptive heuristics based on the total latency and resources
2379 // inside and outside this zone. Potential stalls should be considered before
2380 // following this policy.
2382 // Compute remaining latency. We need this both to determine whether the
2383 // overall schedule has become latency-limited and whether the instructions
2384 // outside this zone are resource or latency limited.
2386 // The "dependent" latency is updated incrementally during scheduling as the
2387 // max height/depth of scheduled nodes minus the cycles since it was
2389 // DLat = max (N.depth - (CurrCycle - N.ReadyCycle) for N in Zone
2391 // The "independent" latency is the max ready queue depth:
2392 // ILat = max N.depth for N in Available|Pending
2394 // RemainingLatency is the greater of independent and dependent latency.
2395 unsigned RemLatency = CurrZone.getDependentLatency();
2396 RemLatency = std::max(RemLatency,
2397 CurrZone.findMaxLatency(CurrZone.Available.elements()));
2398 RemLatency = std::max(RemLatency,
2399 CurrZone.findMaxLatency(CurrZone.Pending.elements()));
2401 // Compute the critical resource outside the zone.
2402 unsigned OtherCritIdx = 0;
2403 unsigned OtherCount =
2404 OtherZone ? OtherZone->getOtherResourceCount(OtherCritIdx) : 0;
2406 bool OtherResLimited = false;
2407 if (SchedModel->hasInstrSchedModel()) {
2408 unsigned LFactor = SchedModel->getLatencyFactor();
2409 OtherResLimited = (int)(OtherCount - (RemLatency * LFactor)) > (int)LFactor;
2411 // Schedule aggressively for latency in PostRA mode. We don't check for
2412 // acyclic latency during PostRA, and highly out-of-order processors will
2413 // skip PostRA scheduling.
2414 if (!OtherResLimited) {
2415 if (IsPostRA || (RemLatency + CurrZone.getCurrCycle() > Rem.CriticalPath)) {
2416 Policy.ReduceLatency |= true;
2417 DEBUG(dbgs() << " " << CurrZone.Available.getName()
2418 << " RemainingLatency " << RemLatency << " + "
2419 << CurrZone.getCurrCycle() << "c > CritPath "
2420 << Rem.CriticalPath << "\n");
2423 // If the same resource is limiting inside and outside the zone, do nothing.
2424 if (CurrZone.getZoneCritResIdx() == OtherCritIdx)
2428 if (CurrZone.isResourceLimited()) {
2429 dbgs() << " " << CurrZone.Available.getName() << " ResourceLimited: "
2430 << SchedModel->getResourceName(CurrZone.getZoneCritResIdx())
2433 if (OtherResLimited)
2434 dbgs() << " RemainingLimit: "
2435 << SchedModel->getResourceName(OtherCritIdx) << "\n";
2436 if (!CurrZone.isResourceLimited() && !OtherResLimited)
2437 dbgs() << " Latency limited both directions.\n");
2439 if (CurrZone.isResourceLimited() && !Policy.ReduceResIdx)
2440 Policy.ReduceResIdx = CurrZone.getZoneCritResIdx();
2442 if (OtherResLimited)
2443 Policy.DemandResIdx = OtherCritIdx;
2447 const char *GenericSchedulerBase::getReasonStr(
2448 GenericSchedulerBase::CandReason Reason) {
2450 case NoCand: return "NOCAND ";
2451 case Only1: return "ONLY1 ";
2452 case PhysRegCopy: return "PREG-COPY ";
2453 case RegExcess: return "REG-EXCESS";
2454 case RegCritical: return "REG-CRIT ";
2455 case Stall: return "STALL ";
2456 case Cluster: return "CLUSTER ";
2457 case Weak: return "WEAK ";
2458 case RegMax: return "REG-MAX ";
2459 case ResourceReduce: return "RES-REDUCE";
2460 case ResourceDemand: return "RES-DEMAND";
2461 case TopDepthReduce: return "TOP-DEPTH ";
2462 case TopPathReduce: return "TOP-PATH ";
2463 case BotHeightReduce:return "BOT-HEIGHT";
2464 case BotPathReduce: return "BOT-PATH ";
2465 case NextDefUse: return "DEF-USE ";
2466 case NodeOrder: return "ORDER ";
2468 llvm_unreachable("Unknown reason!");
2471 void GenericSchedulerBase::traceCandidate(const SchedCandidate &Cand) {
2473 unsigned ResIdx = 0;
2474 unsigned Latency = 0;
2475 switch (Cand.Reason) {
2479 P = Cand.RPDelta.Excess;
2482 P = Cand.RPDelta.CriticalMax;
2485 P = Cand.RPDelta.CurrentMax;
2487 case ResourceReduce:
2488 ResIdx = Cand.Policy.ReduceResIdx;
2490 case ResourceDemand:
2491 ResIdx = Cand.Policy.DemandResIdx;
2493 case TopDepthReduce:
2494 Latency = Cand.SU->getDepth();
2497 Latency = Cand.SU->getHeight();
2499 case BotHeightReduce:
2500 Latency = Cand.SU->getHeight();
2503 Latency = Cand.SU->getDepth();
2506 dbgs() << " Cand SU(" << Cand.SU->NodeNum << ") " << getReasonStr(Cand.Reason);
2508 dbgs() << " " << TRI->getRegPressureSetName(P.getPSet())
2509 << ":" << P.getUnitInc() << " ";
2513 dbgs() << " " << SchedModel->getProcResource(ResIdx)->Name << " ";
2517 dbgs() << " " << Latency << " cycles ";
2524 /// Return true if this heuristic determines order.
2525 static bool tryLess(int TryVal, int CandVal,
2526 GenericSchedulerBase::SchedCandidate &TryCand,
2527 GenericSchedulerBase::SchedCandidate &Cand,
2528 GenericSchedulerBase::CandReason Reason) {
2529 if (TryVal < CandVal) {
2530 TryCand.Reason = Reason;
2533 if (TryVal > CandVal) {
2534 if (Cand.Reason > Reason)
2535 Cand.Reason = Reason;
2541 static bool tryGreater(int TryVal, int CandVal,
2542 GenericSchedulerBase::SchedCandidate &TryCand,
2543 GenericSchedulerBase::SchedCandidate &Cand,
2544 GenericSchedulerBase::CandReason Reason) {
2545 if (TryVal > CandVal) {
2546 TryCand.Reason = Reason;
2549 if (TryVal < CandVal) {
2550 if (Cand.Reason > Reason)
2551 Cand.Reason = Reason;
2557 static bool tryLatency(GenericSchedulerBase::SchedCandidate &TryCand,
2558 GenericSchedulerBase::SchedCandidate &Cand,
2559 SchedBoundary &Zone) {
2561 if (Cand.SU->getDepth() > Zone.getScheduledLatency()) {
2562 if (tryLess(TryCand.SU->getDepth(), Cand.SU->getDepth(),
2563 TryCand, Cand, GenericSchedulerBase::TopDepthReduce))
2566 if (tryGreater(TryCand.SU->getHeight(), Cand.SU->getHeight(),
2567 TryCand, Cand, GenericSchedulerBase::TopPathReduce))
2570 if (Cand.SU->getHeight() > Zone.getScheduledLatency()) {
2571 if (tryLess(TryCand.SU->getHeight(), Cand.SU->getHeight(),
2572 TryCand, Cand, GenericSchedulerBase::BotHeightReduce))
2575 if (tryGreater(TryCand.SU->getDepth(), Cand.SU->getDepth(),
2576 TryCand, Cand, GenericSchedulerBase::BotPathReduce))
2582 static void tracePick(GenericSchedulerBase::CandReason Reason, bool IsTop) {
2583 DEBUG(dbgs() << "Pick " << (IsTop ? "Top " : "Bot ")
2584 << GenericSchedulerBase::getReasonStr(Reason) << '\n');
2587 static void tracePick(const GenericSchedulerBase::SchedCandidate &Cand) {
2588 tracePick(Cand.Reason, Cand.AtTop);
2591 void GenericScheduler::initialize(ScheduleDAGMI *dag) {
2592 assert(dag->hasVRegLiveness() &&
2593 "(PreRA)GenericScheduler needs vreg liveness");
2594 DAG = static_cast<ScheduleDAGMILive*>(dag);
2595 SchedModel = DAG->getSchedModel();
2598 Rem.init(DAG, SchedModel);
2599 Top.init(DAG, SchedModel, &Rem);
2600 Bot.init(DAG, SchedModel, &Rem);
2602 // Initialize resource counts.
2604 // Initialize the HazardRecognizers. If itineraries don't exist, are empty, or
2605 // are disabled, then these HazardRecs will be disabled.
2606 const InstrItineraryData *Itin = SchedModel->getInstrItineraries();
2607 if (!Top.HazardRec) {
2609 DAG->MF.getSubtarget().getInstrInfo()->CreateTargetMIHazardRecognizer(
2612 if (!Bot.HazardRec) {
2614 DAG->MF.getSubtarget().getInstrInfo()->CreateTargetMIHazardRecognizer(
2617 TopCand.SU = nullptr;
2618 BotCand.SU = nullptr;
2621 /// Initialize the per-region scheduling policy.
2622 void GenericScheduler::initPolicy(MachineBasicBlock::iterator Begin,
2623 MachineBasicBlock::iterator End,
2624 unsigned NumRegionInstrs) {
2625 const MachineFunction &MF = *Begin->getParent()->getParent();
2626 const TargetLowering *TLI = MF.getSubtarget().getTargetLowering();
2628 // Avoid setting up the register pressure tracker for small regions to save
2629 // compile time. As a rough heuristic, only track pressure when the number of
2630 // schedulable instructions exceeds half the integer register file.
2631 RegionPolicy.ShouldTrackPressure = true;
2632 for (unsigned VT = MVT::i32; VT > (unsigned)MVT::i1; --VT) {
2633 MVT::SimpleValueType LegalIntVT = (MVT::SimpleValueType)VT;
2634 if (TLI->isTypeLegal(LegalIntVT)) {
2635 unsigned NIntRegs = Context->RegClassInfo->getNumAllocatableRegs(
2636 TLI->getRegClassFor(LegalIntVT));
2637 RegionPolicy.ShouldTrackPressure = NumRegionInstrs > (NIntRegs / 2);
2641 // For generic targets, we default to bottom-up, because it's simpler and more
2642 // compile-time optimizations have been implemented in that direction.
2643 RegionPolicy.OnlyBottomUp = true;
2645 // Allow the subtarget to override default policy.
2646 MF.getSubtarget().overrideSchedPolicy(RegionPolicy, NumRegionInstrs);
2648 // After subtarget overrides, apply command line options.
2649 if (!EnableRegPressure)
2650 RegionPolicy.ShouldTrackPressure = false;
2652 // Check -misched-topdown/bottomup can force or unforce scheduling direction.
2653 // e.g. -misched-bottomup=false allows scheduling in both directions.
2654 assert((!ForceTopDown || !ForceBottomUp) &&
2655 "-misched-topdown incompatible with -misched-bottomup");
2656 if (ForceBottomUp.getNumOccurrences() > 0) {
2657 RegionPolicy.OnlyBottomUp = ForceBottomUp;
2658 if (RegionPolicy.OnlyBottomUp)
2659 RegionPolicy.OnlyTopDown = false;
2661 if (ForceTopDown.getNumOccurrences() > 0) {
2662 RegionPolicy.OnlyTopDown = ForceTopDown;
2663 if (RegionPolicy.OnlyTopDown)
2664 RegionPolicy.OnlyBottomUp = false;
2668 void GenericScheduler::dumpPolicy() {
2669 dbgs() << "GenericScheduler RegionPolicy: "
2670 << " ShouldTrackPressure=" << RegionPolicy.ShouldTrackPressure
2671 << " OnlyTopDown=" << RegionPolicy.OnlyTopDown
2672 << " OnlyBottomUp=" << RegionPolicy.OnlyBottomUp
2676 /// Set IsAcyclicLatencyLimited if the acyclic path is longer than the cyclic
2677 /// critical path by more cycles than it takes to drain the instruction buffer.
2678 /// We estimate an upper bounds on in-flight instructions as:
2680 /// CyclesPerIteration = max( CyclicPath, Loop-Resource-Height )
2681 /// InFlightIterations = AcyclicPath / CyclesPerIteration
2682 /// InFlightResources = InFlightIterations * LoopResources
2684 /// TODO: Check execution resources in addition to IssueCount.
2685 void GenericScheduler::checkAcyclicLatency() {
2686 if (Rem.CyclicCritPath == 0 || Rem.CyclicCritPath >= Rem.CriticalPath)
2689 // Scaled number of cycles per loop iteration.
2690 unsigned IterCount =
2691 std::max(Rem.CyclicCritPath * SchedModel->getLatencyFactor(),
2693 // Scaled acyclic critical path.
2694 unsigned AcyclicCount = Rem.CriticalPath * SchedModel->getLatencyFactor();
2695 // InFlightCount = (AcyclicPath / IterCycles) * InstrPerLoop
2696 unsigned InFlightCount =
2697 (AcyclicCount * Rem.RemIssueCount + IterCount-1) / IterCount;
2698 unsigned BufferLimit =
2699 SchedModel->getMicroOpBufferSize() * SchedModel->getMicroOpFactor();
2701 Rem.IsAcyclicLatencyLimited = InFlightCount > BufferLimit;
2703 DEBUG(dbgs() << "IssueCycles="
2704 << Rem.RemIssueCount / SchedModel->getLatencyFactor() << "c "
2705 << "IterCycles=" << IterCount / SchedModel->getLatencyFactor()
2706 << "c NumIters=" << (AcyclicCount + IterCount-1) / IterCount
2707 << " InFlight=" << InFlightCount / SchedModel->getMicroOpFactor()
2708 << "m BufferLim=" << SchedModel->getMicroOpBufferSize() << "m\n";
2709 if (Rem.IsAcyclicLatencyLimited)
2710 dbgs() << " ACYCLIC LATENCY LIMIT\n");
2713 void GenericScheduler::registerRoots() {
2714 Rem.CriticalPath = DAG->ExitSU.getDepth();
2716 // Some roots may not feed into ExitSU. Check all of them in case.
2717 for (std::vector<SUnit*>::const_iterator
2718 I = Bot.Available.begin(), E = Bot.Available.end(); I != E; ++I) {
2719 if ((*I)->getDepth() > Rem.CriticalPath)
2720 Rem.CriticalPath = (*I)->getDepth();
2722 DEBUG(dbgs() << "Critical Path(GS-RR ): " << Rem.CriticalPath << '\n');
2723 if (DumpCriticalPathLength) {
2724 errs() << "Critical Path(GS-RR ): " << Rem.CriticalPath << " \n";
2727 if (EnableCyclicPath) {
2728 Rem.CyclicCritPath = DAG->computeCyclicCriticalPath();
2729 checkAcyclicLatency();
2733 static bool tryPressure(const PressureChange &TryP,
2734 const PressureChange &CandP,
2735 GenericSchedulerBase::SchedCandidate &TryCand,
2736 GenericSchedulerBase::SchedCandidate &Cand,
2737 GenericSchedulerBase::CandReason Reason,
2738 const TargetRegisterInfo *TRI,
2739 const MachineFunction &MF) {
2740 // If one candidate decreases and the other increases, go with it.
2741 // Invalid candidates have UnitInc==0.
2742 if (tryGreater(TryP.getUnitInc() < 0, CandP.getUnitInc() < 0, TryCand, Cand,
2746 // Do not compare the magnitude of pressure changes between top and bottom
2748 if (Cand.AtTop != TryCand.AtTop)
2751 // If both candidates affect the same set in the same boundary, go with the
2752 // smallest increase.
2753 unsigned TryPSet = TryP.getPSetOrMax();
2754 unsigned CandPSet = CandP.getPSetOrMax();
2755 if (TryPSet == CandPSet) {
2756 return tryLess(TryP.getUnitInc(), CandP.getUnitInc(), TryCand, Cand,
2760 int TryRank = TryP.isValid() ? TRI->getRegPressureSetScore(MF, TryPSet) :
2761 std::numeric_limits<int>::max();
2763 int CandRank = CandP.isValid() ? TRI->getRegPressureSetScore(MF, CandPSet) :
2764 std::numeric_limits<int>::max();
2766 // If the candidates are decreasing pressure, reverse priority.
2767 if (TryP.getUnitInc() < 0)
2768 std::swap(TryRank, CandRank);
2769 return tryGreater(TryRank, CandRank, TryCand, Cand, Reason);
2772 static unsigned getWeakLeft(const SUnit *SU, bool isTop) {
2773 return (isTop) ? SU->WeakPredsLeft : SU->WeakSuccsLeft;
2776 /// Minimize physical register live ranges. Regalloc wants them adjacent to
2777 /// their physreg def/use.
2779 /// FIXME: This is an unnecessary check on the critical path. Most are root/leaf
2780 /// copies which can be prescheduled. The rest (e.g. x86 MUL) could be bundled
2781 /// with the operation that produces or consumes the physreg. We'll do this when
2782 /// regalloc has support for parallel copies.
2783 static int biasPhysRegCopy(const SUnit *SU, bool isTop) {
2784 const MachineInstr *MI = SU->getInstr();
2788 unsigned ScheduledOper = isTop ? 1 : 0;
2789 unsigned UnscheduledOper = isTop ? 0 : 1;
2790 // If we have already scheduled the physreg produce/consumer, immediately
2791 // schedule the copy.
2792 if (TargetRegisterInfo::isPhysicalRegister(
2793 MI->getOperand(ScheduledOper).getReg()))
2795 // If the physreg is at the boundary, defer it. Otherwise schedule it
2796 // immediately to free the dependent. We can hoist the copy later.
2797 bool AtBoundary = isTop ? !SU->NumSuccsLeft : !SU->NumPredsLeft;
2798 if (TargetRegisterInfo::isPhysicalRegister(
2799 MI->getOperand(UnscheduledOper).getReg()))
2800 return AtBoundary ? -1 : 1;
2804 void GenericScheduler::initCandidate(SchedCandidate &Cand, SUnit *SU,
2806 const RegPressureTracker &RPTracker,
2807 RegPressureTracker &TempTracker) {
2810 if (DAG->isTrackingPressure()) {
2812 TempTracker.getMaxDownwardPressureDelta(
2813 Cand.SU->getInstr(),
2815 DAG->getRegionCriticalPSets(),
2816 DAG->getRegPressure().MaxSetPressure);
2818 if (VerifyScheduling) {
2819 TempTracker.getMaxUpwardPressureDelta(
2820 Cand.SU->getInstr(),
2821 &DAG->getPressureDiff(Cand.SU),
2823 DAG->getRegionCriticalPSets(),
2824 DAG->getRegPressure().MaxSetPressure);
2826 RPTracker.getUpwardPressureDelta(
2827 Cand.SU->getInstr(),
2828 DAG->getPressureDiff(Cand.SU),
2830 DAG->getRegionCriticalPSets(),
2831 DAG->getRegPressure().MaxSetPressure);
2835 DEBUG(if (Cand.RPDelta.Excess.isValid())
2836 dbgs() << " Try SU(" << Cand.SU->NodeNum << ") "
2837 << TRI->getRegPressureSetName(Cand.RPDelta.Excess.getPSet())
2838 << ":" << Cand.RPDelta.Excess.getUnitInc() << "\n");
2841 /// Apply a set of heursitics to a new candidate. Heuristics are currently
2842 /// hierarchical. This may be more efficient than a graduated cost model because
2843 /// we don't need to evaluate all aspects of the model for each node in the
2844 /// queue. But it's really done to make the heuristics easier to debug and
2845 /// statistically analyze.
2847 /// \param Cand provides the policy and current best candidate.
2848 /// \param TryCand refers to the next SUnit candidate, otherwise uninitialized.
2849 /// \param Zone describes the scheduled zone that we are extending, or nullptr
2850 // if Cand is from a different zone than TryCand.
2851 void GenericScheduler::tryCandidate(SchedCandidate &Cand,
2852 SchedCandidate &TryCand,
2853 SchedBoundary *Zone) {
2854 // Initialize the candidate if needed.
2855 if (!Cand.isValid()) {
2856 TryCand.Reason = NodeOrder;
2860 if (tryGreater(biasPhysRegCopy(TryCand.SU, TryCand.AtTop),
2861 biasPhysRegCopy(Cand.SU, Cand.AtTop),
2862 TryCand, Cand, PhysRegCopy))
2865 // Avoid exceeding the target's limit.
2866 if (DAG->isTrackingPressure() && tryPressure(TryCand.RPDelta.Excess,
2867 Cand.RPDelta.Excess,
2868 TryCand, Cand, RegExcess, TRI,
2872 // Avoid increasing the max critical pressure in the scheduled region.
2873 if (DAG->isTrackingPressure() && tryPressure(TryCand.RPDelta.CriticalMax,
2874 Cand.RPDelta.CriticalMax,
2875 TryCand, Cand, RegCritical, TRI,
2879 // We only compare a subset of features when comparing nodes between
2880 // Top and Bottom boundary. Some properties are simply incomparable, in many
2881 // other instances we should only override the other boundary if something
2882 // is a clear good pick on one boundary. Skip heuristics that are more
2883 // "tie-breaking" in nature.
2884 bool SameBoundary = Zone != nullptr;
2886 // For loops that are acyclic path limited, aggressively schedule for
2887 // latency. Within an single cycle, whenever CurrMOps > 0, allow normal
2888 // heuristics to take precedence.
2889 if (Rem.IsAcyclicLatencyLimited && !Zone->getCurrMOps() &&
2890 tryLatency(TryCand, Cand, *Zone))
2893 // Prioritize instructions that read unbuffered resources by stall cycles.
2894 if (tryLess(Zone->getLatencyStallCycles(TryCand.SU),
2895 Zone->getLatencyStallCycles(Cand.SU), TryCand, Cand, Stall))
2899 // Keep clustered nodes together to encourage downstream peephole
2900 // optimizations which may reduce resource requirements.
2902 // This is a best effort to set things up for a post-RA pass. Optimizations
2903 // like generating loads of multiple registers should ideally be done within
2904 // the scheduler pass by combining the loads during DAG postprocessing.
2905 const SUnit *CandNextClusterSU =
2906 Cand.AtTop ? DAG->getNextClusterSucc() : DAG->getNextClusterPred();
2907 const SUnit *TryCandNextClusterSU =
2908 TryCand.AtTop ? DAG->getNextClusterSucc() : DAG->getNextClusterPred();
2909 if (tryGreater(TryCand.SU == TryCandNextClusterSU,
2910 Cand.SU == CandNextClusterSU,
2911 TryCand, Cand, Cluster))
2915 // Weak edges are for clustering and other constraints.
2916 if (tryLess(getWeakLeft(TryCand.SU, TryCand.AtTop),
2917 getWeakLeft(Cand.SU, Cand.AtTop),
2918 TryCand, Cand, Weak))
2922 // Avoid increasing the max pressure of the entire region.
2923 if (DAG->isTrackingPressure() && tryPressure(TryCand.RPDelta.CurrentMax,
2924 Cand.RPDelta.CurrentMax,
2925 TryCand, Cand, RegMax, TRI,
2930 // Avoid critical resource consumption and balance the schedule.
2931 TryCand.initResourceDelta(DAG, SchedModel);
2932 if (tryLess(TryCand.ResDelta.CritResources, Cand.ResDelta.CritResources,
2933 TryCand, Cand, ResourceReduce))
2935 if (tryGreater(TryCand.ResDelta.DemandedResources,
2936 Cand.ResDelta.DemandedResources,
2937 TryCand, Cand, ResourceDemand))
2940 // Avoid serializing long latency dependence chains.
2941 // For acyclic path limited loops, latency was already checked above.
2942 if (!RegionPolicy.DisableLatencyHeuristic && TryCand.Policy.ReduceLatency &&
2943 !Rem.IsAcyclicLatencyLimited && tryLatency(TryCand, Cand, *Zone))
2946 // Fall through to original instruction order.
2947 if ((Zone->isTop() && TryCand.SU->NodeNum < Cand.SU->NodeNum)
2948 || (!Zone->isTop() && TryCand.SU->NodeNum > Cand.SU->NodeNum)) {
2949 TryCand.Reason = NodeOrder;
2954 /// Pick the best candidate from the queue.
2956 /// TODO: getMaxPressureDelta results can be mostly cached for each SUnit during
2957 /// DAG building. To adjust for the current scheduling location we need to
2958 /// maintain the number of vreg uses remaining to be top-scheduled.
2959 void GenericScheduler::pickNodeFromQueue(SchedBoundary &Zone,
2960 const CandPolicy &ZonePolicy,
2961 const RegPressureTracker &RPTracker,
2962 SchedCandidate &Cand) {
2963 // getMaxPressureDelta temporarily modifies the tracker.
2964 RegPressureTracker &TempTracker = const_cast<RegPressureTracker&>(RPTracker);
2966 ReadyQueue &Q = Zone.Available;
2967 for (ReadyQueue::iterator I = Q.begin(), E = Q.end(); I != E; ++I) {
2969 SchedCandidate TryCand(ZonePolicy);
2970 initCandidate(TryCand, *I, Zone.isTop(), RPTracker, TempTracker);
2971 // Pass SchedBoundary only when comparing nodes from the same boundary.
2972 SchedBoundary *ZoneArg = Cand.AtTop == TryCand.AtTop ? &Zone : nullptr;
2973 tryCandidate(Cand, TryCand, ZoneArg);
2974 if (TryCand.Reason != NoCand) {
2975 // Initialize resource delta if needed in case future heuristics query it.
2976 if (TryCand.ResDelta == SchedResourceDelta())
2977 TryCand.initResourceDelta(DAG, SchedModel);
2978 Cand.setBest(TryCand);
2979 DEBUG(traceCandidate(Cand));
2984 /// Pick the best candidate node from either the top or bottom queue.
2985 SUnit *GenericScheduler::pickNodeBidirectional(bool &IsTopNode) {
2986 // Schedule as far as possible in the direction of no choice. This is most
2987 // efficient, but also provides the best heuristics for CriticalPSets.
2988 if (SUnit *SU = Bot.pickOnlyChoice()) {
2990 tracePick(Only1, false);
2993 if (SUnit *SU = Top.pickOnlyChoice()) {
2995 tracePick(Only1, true);
2998 // Set the bottom-up policy based on the state of the current bottom zone and
2999 // the instructions outside the zone, including the top zone.
3000 CandPolicy BotPolicy;
3001 setPolicy(BotPolicy, /*IsPostRA=*/false, Bot, &Top);
3002 // Set the top-down policy based on the state of the current top zone and
3003 // the instructions outside the zone, including the bottom zone.
3004 CandPolicy TopPolicy;
3005 setPolicy(TopPolicy, /*IsPostRA=*/false, Top, &Bot);
3007 // See if BotCand is still valid (because we previously scheduled from Top).
3008 DEBUG(dbgs() << "Picking from Bot:\n");
3009 if (!BotCand.isValid() || BotCand.SU->isScheduled ||
3010 BotCand.Policy != BotPolicy) {
3011 BotCand.reset(CandPolicy());
3012 pickNodeFromQueue(Bot, BotPolicy, DAG->getBotRPTracker(), BotCand);
3013 assert(BotCand.Reason != NoCand && "failed to find the first candidate");
3015 DEBUG(traceCandidate(BotCand));
3017 if (VerifyScheduling) {
3018 SchedCandidate TCand;
3019 TCand.reset(CandPolicy());
3020 pickNodeFromQueue(Bot, BotPolicy, DAG->getBotRPTracker(), TCand);
3021 assert(TCand.SU == BotCand.SU &&
3022 "Last pick result should correspond to re-picking right now");
3027 // Check if the top Q has a better candidate.
3028 DEBUG(dbgs() << "Picking from Top:\n");
3029 if (!TopCand.isValid() || TopCand.SU->isScheduled ||
3030 TopCand.Policy != TopPolicy) {
3031 TopCand.reset(CandPolicy());
3032 pickNodeFromQueue(Top, TopPolicy, DAG->getTopRPTracker(), TopCand);
3033 assert(TopCand.Reason != NoCand && "failed to find the first candidate");
3035 DEBUG(traceCandidate(TopCand));
3037 if (VerifyScheduling) {
3038 SchedCandidate TCand;
3039 TCand.reset(CandPolicy());
3040 pickNodeFromQueue(Top, TopPolicy, DAG->getTopRPTracker(), TCand);
3041 assert(TCand.SU == TopCand.SU &&
3042 "Last pick result should correspond to re-picking right now");
3047 // Pick best from BotCand and TopCand.
3048 assert(BotCand.isValid());
3049 assert(TopCand.isValid());
3050 SchedCandidate Cand = BotCand;
3051 TopCand.Reason = NoCand;
3052 tryCandidate(Cand, TopCand, nullptr);
3053 if (TopCand.Reason != NoCand) {
3054 Cand.setBest(TopCand);
3055 DEBUG(traceCandidate(Cand));
3058 IsTopNode = Cand.AtTop;
3063 /// Pick the best node to balance the schedule. Implements MachineSchedStrategy.
3064 SUnit *GenericScheduler::pickNode(bool &IsTopNode) {
3065 if (DAG->top() == DAG->bottom()) {
3066 assert(Top.Available.empty() && Top.Pending.empty() &&
3067 Bot.Available.empty() && Bot.Pending.empty() && "ReadyQ garbage");
3072 if (RegionPolicy.OnlyTopDown) {
3073 SU = Top.pickOnlyChoice();
3075 CandPolicy NoPolicy;
3076 TopCand.reset(NoPolicy);
3077 pickNodeFromQueue(Top, NoPolicy, DAG->getTopRPTracker(), TopCand);
3078 assert(TopCand.Reason != NoCand && "failed to find a candidate");
3083 } else if (RegionPolicy.OnlyBottomUp) {
3084 SU = Bot.pickOnlyChoice();
3086 CandPolicy NoPolicy;
3087 BotCand.reset(NoPolicy);
3088 pickNodeFromQueue(Bot, NoPolicy, DAG->getBotRPTracker(), BotCand);
3089 assert(BotCand.Reason != NoCand && "failed to find a candidate");
3095 SU = pickNodeBidirectional(IsTopNode);
3097 } while (SU->isScheduled);
3099 if (SU->isTopReady())
3100 Top.removeReady(SU);
3101 if (SU->isBottomReady())
3102 Bot.removeReady(SU);
3104 DEBUG(dbgs() << "Scheduling SU(" << SU->NodeNum << ") " << *SU->getInstr());
3108 void GenericScheduler::reschedulePhysRegCopies(SUnit *SU, bool isTop) {
3110 MachineBasicBlock::iterator InsertPos = SU->getInstr();
3113 SmallVectorImpl<SDep> &Deps = isTop ? SU->Preds : SU->Succs;
3115 // Find already scheduled copies with a single physreg dependence and move
3116 // them just above the scheduled instruction.
3117 for (SmallVectorImpl<SDep>::iterator I = Deps.begin(), E = Deps.end();
3119 if (I->getKind() != SDep::Data || !TRI->isPhysicalRegister(I->getReg()))
3121 SUnit *DepSU = I->getSUnit();
3122 if (isTop ? DepSU->Succs.size() > 1 : DepSU->Preds.size() > 1)
3124 MachineInstr *Copy = DepSU->getInstr();
3125 if (!Copy->isCopy())
3127 DEBUG(dbgs() << " Rescheduling physreg copy ";
3128 I->getSUnit()->dump(DAG));
3129 DAG->moveInstruction(Copy, InsertPos);
3133 /// Update the scheduler's state after scheduling a node. This is the same node
3134 /// that was just returned by pickNode(). However, ScheduleDAGMILive needs to
3135 /// update it's state based on the current cycle before MachineSchedStrategy
3138 /// FIXME: Eventually, we may bundle physreg copies rather than rescheduling
3139 /// them here. See comments in biasPhysRegCopy.
3140 void GenericScheduler::schedNode(SUnit *SU, bool IsTopNode) {
3142 SU->TopReadyCycle = std::max(SU->TopReadyCycle, Top.getCurrCycle());
3144 if (SU->hasPhysRegUses)
3145 reschedulePhysRegCopies(SU, true);
3147 SU->BotReadyCycle = std::max(SU->BotReadyCycle, Bot.getCurrCycle());
3149 if (SU->hasPhysRegDefs)
3150 reschedulePhysRegCopies(SU, false);
3154 /// Create the standard converging machine scheduler. This will be used as the
3155 /// default scheduler if the target does not set a default.
3156 ScheduleDAGMILive *llvm::createGenericSchedLive(MachineSchedContext *C) {
3157 ScheduleDAGMILive *DAG = new ScheduleDAGMILive(C, make_unique<GenericScheduler>(C));
3158 // Register DAG post-processors.
3160 // FIXME: extend the mutation API to allow earlier mutations to instantiate
3161 // data and pass it to later mutations. Have a single mutation that gathers
3162 // the interesting nodes in one pass.
3163 DAG->addMutation(createCopyConstrainDAGMutation(DAG->TII, DAG->TRI));
3167 static ScheduleDAGInstrs *createConveringSched(MachineSchedContext *C) {
3168 return createGenericSchedLive(C);
3171 static MachineSchedRegistry
3172 GenericSchedRegistry("converge", "Standard converging scheduler.",
3173 createConveringSched);
3175 //===----------------------------------------------------------------------===//
3176 // PostGenericScheduler - Generic PostRA implementation of MachineSchedStrategy.
3177 //===----------------------------------------------------------------------===//
3179 void PostGenericScheduler::initialize(ScheduleDAGMI *Dag) {
3181 SchedModel = DAG->getSchedModel();
3184 Rem.init(DAG, SchedModel);
3185 Top.init(DAG, SchedModel, &Rem);
3188 // Initialize the HazardRecognizers. If itineraries don't exist, are empty,
3189 // or are disabled, then these HazardRecs will be disabled.
3190 const InstrItineraryData *Itin = SchedModel->getInstrItineraries();
3191 if (!Top.HazardRec) {
3193 DAG->MF.getSubtarget().getInstrInfo()->CreateTargetMIHazardRecognizer(
3199 void PostGenericScheduler::registerRoots() {
3200 Rem.CriticalPath = DAG->ExitSU.getDepth();
3202 // Some roots may not feed into ExitSU. Check all of them in case.
3203 for (SmallVectorImpl<SUnit*>::const_iterator
3204 I = BotRoots.begin(), E = BotRoots.end(); I != E; ++I) {
3205 if ((*I)->getDepth() > Rem.CriticalPath)
3206 Rem.CriticalPath = (*I)->getDepth();
3208 DEBUG(dbgs() << "Critical Path: (PGS-RR) " << Rem.CriticalPath << '\n');
3209 if (DumpCriticalPathLength) {
3210 errs() << "Critical Path(PGS-RR ): " << Rem.CriticalPath << " \n";
3214 /// Apply a set of heursitics to a new candidate for PostRA scheduling.
3216 /// \param Cand provides the policy and current best candidate.
3217 /// \param TryCand refers to the next SUnit candidate, otherwise uninitialized.
3218 void PostGenericScheduler::tryCandidate(SchedCandidate &Cand,
3219 SchedCandidate &TryCand) {
3221 // Initialize the candidate if needed.
3222 if (!Cand.isValid()) {
3223 TryCand.Reason = NodeOrder;
3227 // Prioritize instructions that read unbuffered resources by stall cycles.
3228 if (tryLess(Top.getLatencyStallCycles(TryCand.SU),
3229 Top.getLatencyStallCycles(Cand.SU), TryCand, Cand, Stall))
3232 // Avoid critical resource consumption and balance the schedule.
3233 if (tryLess(TryCand.ResDelta.CritResources, Cand.ResDelta.CritResources,
3234 TryCand, Cand, ResourceReduce))
3236 if (tryGreater(TryCand.ResDelta.DemandedResources,
3237 Cand.ResDelta.DemandedResources,
3238 TryCand, Cand, ResourceDemand))
3241 // Avoid serializing long latency dependence chains.
3242 if (Cand.Policy.ReduceLatency && tryLatency(TryCand, Cand, Top)) {
3246 // Fall through to original instruction order.
3247 if (TryCand.SU->NodeNum < Cand.SU->NodeNum)
3248 TryCand.Reason = NodeOrder;
3251 void PostGenericScheduler::pickNodeFromQueue(SchedCandidate &Cand) {
3252 ReadyQueue &Q = Top.Available;
3253 for (ReadyQueue::iterator I = Q.begin(), E = Q.end(); I != E; ++I) {
3254 SchedCandidate TryCand(Cand.Policy);
3256 TryCand.AtTop = true;
3257 TryCand.initResourceDelta(DAG, SchedModel);
3258 tryCandidate(Cand, TryCand);
3259 if (TryCand.Reason != NoCand) {
3260 Cand.setBest(TryCand);
3261 DEBUG(traceCandidate(Cand));
3266 /// Pick the next node to schedule.
3267 SUnit *PostGenericScheduler::pickNode(bool &IsTopNode) {
3268 if (DAG->top() == DAG->bottom()) {
3269 assert(Top.Available.empty() && Top.Pending.empty() && "ReadyQ garbage");
3274 SU = Top.pickOnlyChoice();
3276 tracePick(Only1, true);
3278 CandPolicy NoPolicy;
3279 SchedCandidate TopCand(NoPolicy);
3280 // Set the top-down policy based on the state of the current top zone and
3281 // the instructions outside the zone, including the bottom zone.
3282 setPolicy(TopCand.Policy, /*IsPostRA=*/true, Top, nullptr);
3283 pickNodeFromQueue(TopCand);
3284 assert(TopCand.Reason != NoCand && "failed to find a candidate");
3288 } while (SU->isScheduled);
3291 Top.removeReady(SU);
3293 DEBUG(dbgs() << "Scheduling SU(" << SU->NodeNum << ") " << *SU->getInstr());
3297 /// Called after ScheduleDAGMI has scheduled an instruction and updated
3298 /// scheduled/remaining flags in the DAG nodes.
3299 void PostGenericScheduler::schedNode(SUnit *SU, bool IsTopNode) {
3300 SU->TopReadyCycle = std::max(SU->TopReadyCycle, Top.getCurrCycle());
3304 ScheduleDAGMI *llvm::createGenericSchedPostRA(MachineSchedContext *C) {
3305 return new ScheduleDAGMI(C, make_unique<PostGenericScheduler>(C),
3306 /*RemoveKillFlags=*/true);
3309 //===----------------------------------------------------------------------===//
3310 // ILP Scheduler. Currently for experimental analysis of heuristics.
3311 //===----------------------------------------------------------------------===//
3314 /// \brief Order nodes by the ILP metric.
3316 const SchedDFSResult *DFSResult;
3317 const BitVector *ScheduledTrees;
3320 ILPOrder(bool MaxILP)
3321 : DFSResult(nullptr), ScheduledTrees(nullptr), MaximizeILP(MaxILP) {}
3323 /// \brief Apply a less-than relation on node priority.
3325 /// (Return true if A comes after B in the Q.)
3326 bool operator()(const SUnit *A, const SUnit *B) const {
3327 unsigned SchedTreeA = DFSResult->getSubtreeID(A);
3328 unsigned SchedTreeB = DFSResult->getSubtreeID(B);
3329 if (SchedTreeA != SchedTreeB) {
3330 // Unscheduled trees have lower priority.
3331 if (ScheduledTrees->test(SchedTreeA) != ScheduledTrees->test(SchedTreeB))
3332 return ScheduledTrees->test(SchedTreeB);
3334 // Trees with shallower connections have have lower priority.
3335 if (DFSResult->getSubtreeLevel(SchedTreeA)
3336 != DFSResult->getSubtreeLevel(SchedTreeB)) {
3337 return DFSResult->getSubtreeLevel(SchedTreeA)
3338 < DFSResult->getSubtreeLevel(SchedTreeB);
3342 return DFSResult->getILP(A) < DFSResult->getILP(B);
3344 return DFSResult->getILP(A) > DFSResult->getILP(B);
3348 /// \brief Schedule based on the ILP metric.
3349 class ILPScheduler : public MachineSchedStrategy {
3350 ScheduleDAGMILive *DAG;
3353 std::vector<SUnit*> ReadyQ;
3355 ILPScheduler(bool MaximizeILP): DAG(nullptr), Cmp(MaximizeILP) {}
3357 void initialize(ScheduleDAGMI *dag) override {
3358 assert(dag->hasVRegLiveness() && "ILPScheduler needs vreg liveness");
3359 DAG = static_cast<ScheduleDAGMILive*>(dag);
3360 DAG->computeDFSResult();
3361 Cmp.DFSResult = DAG->getDFSResult();
3362 Cmp.ScheduledTrees = &DAG->getScheduledTrees();
3366 void registerRoots() override {
3367 // Restore the heap in ReadyQ with the updated DFS results.
3368 std::make_heap(ReadyQ.begin(), ReadyQ.end(), Cmp);
3371 /// Implement MachineSchedStrategy interface.
3372 /// -----------------------------------------
3374 /// Callback to select the highest priority node from the ready Q.
3375 SUnit *pickNode(bool &IsTopNode) override {
3376 if (ReadyQ.empty()) return nullptr;
3377 std::pop_heap(ReadyQ.begin(), ReadyQ.end(), Cmp);
3378 SUnit *SU = ReadyQ.back();
3381 DEBUG(dbgs() << "Pick node " << "SU(" << SU->NodeNum << ") "
3382 << " ILP: " << DAG->getDFSResult()->getILP(SU)
3383 << " Tree: " << DAG->getDFSResult()->getSubtreeID(SU) << " @"
3384 << DAG->getDFSResult()->getSubtreeLevel(
3385 DAG->getDFSResult()->getSubtreeID(SU)) << '\n'
3386 << "Scheduling " << *SU->getInstr());
3390 /// \brief Scheduler callback to notify that a new subtree is scheduled.
3391 void scheduleTree(unsigned SubtreeID) override {
3392 std::make_heap(ReadyQ.begin(), ReadyQ.end(), Cmp);
3395 /// Callback after a node is scheduled. Mark a newly scheduled tree, notify
3396 /// DFSResults, and resort the priority Q.
3397 void schedNode(SUnit *SU, bool IsTopNode) override {
3398 assert(!IsTopNode && "SchedDFSResult needs bottom-up");
3401 void releaseTopNode(SUnit *) override { /*only called for top roots*/ }
3403 void releaseBottomNode(SUnit *SU) override {
3404 ReadyQ.push_back(SU);
3405 std::push_heap(ReadyQ.begin(), ReadyQ.end(), Cmp);
3410 static ScheduleDAGInstrs *createILPMaxScheduler(MachineSchedContext *C) {
3411 return new ScheduleDAGMILive(C, make_unique<ILPScheduler>(true));
3413 static ScheduleDAGInstrs *createILPMinScheduler(MachineSchedContext *C) {
3414 return new ScheduleDAGMILive(C, make_unique<ILPScheduler>(false));
3416 static MachineSchedRegistry ILPMaxRegistry(
3417 "ilpmax", "Schedule bottom-up for max ILP", createILPMaxScheduler);
3418 static MachineSchedRegistry ILPMinRegistry(
3419 "ilpmin", "Schedule bottom-up for min ILP", createILPMinScheduler);
3421 //===----------------------------------------------------------------------===//
3422 // Machine Instruction Shuffler for Correctness Testing
3423 //===----------------------------------------------------------------------===//
3427 /// Apply a less-than relation on the node order, which corresponds to the
3428 /// instruction order prior to scheduling. IsReverse implements greater-than.
3429 template<bool IsReverse>
3431 bool operator()(SUnit *A, SUnit *B) const {
3433 return A->NodeNum > B->NodeNum;
3435 return A->NodeNum < B->NodeNum;
3439 /// Reorder instructions as much as possible.
3440 class InstructionShuffler : public MachineSchedStrategy {
3444 // Using a less-than relation (SUnitOrder<false>) for the TopQ priority
3445 // gives nodes with a higher number higher priority causing the latest
3446 // instructions to be scheduled first.
3447 PriorityQueue<SUnit*, std::vector<SUnit*>, SUnitOrder<false> >
3449 // When scheduling bottom-up, use greater-than as the queue priority.
3450 PriorityQueue<SUnit*, std::vector<SUnit*>, SUnitOrder<true> >
3453 InstructionShuffler(bool alternate, bool topdown)
3454 : IsAlternating(alternate), IsTopDown(topdown) {}
3456 void initialize(ScheduleDAGMI*) override {
3461 /// Implement MachineSchedStrategy interface.
3462 /// -----------------------------------------
3464 SUnit *pickNode(bool &IsTopNode) override {
3468 if (TopQ.empty()) return nullptr;
3471 } while (SU->isScheduled);
3475 if (BottomQ.empty()) return nullptr;
3478 } while (SU->isScheduled);
3482 IsTopDown = !IsTopDown;
3486 void schedNode(SUnit *SU, bool IsTopNode) override {}
3488 void releaseTopNode(SUnit *SU) override {
3491 void releaseBottomNode(SUnit *SU) override {
3497 static ScheduleDAGInstrs *createInstructionShuffler(MachineSchedContext *C) {
3498 bool Alternate = !ForceTopDown && !ForceBottomUp;
3499 bool TopDown = !ForceBottomUp;
3500 assert((TopDown || !ForceTopDown) &&
3501 "-misched-topdown incompatible with -misched-bottomup");
3502 return new ScheduleDAGMILive(C, make_unique<InstructionShuffler>(Alternate, TopDown));
3504 static MachineSchedRegistry ShufflerRegistry(
3505 "shuffle", "Shuffle machine instructions alternating directions",
3506 createInstructionShuffler);
3509 //===----------------------------------------------------------------------===//
3510 // GraphWriter support for ScheduleDAGMILive.
3511 //===----------------------------------------------------------------------===//
3516 template<> struct GraphTraits<
3517 ScheduleDAGMI*> : public GraphTraits<ScheduleDAG*> {};
3520 struct DOTGraphTraits<ScheduleDAGMI*> : public DefaultDOTGraphTraits {
3522 DOTGraphTraits (bool isSimple=false) : DefaultDOTGraphTraits(isSimple) {}
3524 static std::string getGraphName(const ScheduleDAG *G) {
3525 return G->MF.getName();
3528 static bool renderGraphFromBottomUp() {
3532 static bool isNodeHidden(const SUnit *Node) {
3533 if (ViewMISchedCutoff == 0)
3535 return (Node->Preds.size() > ViewMISchedCutoff
3536 || Node->Succs.size() > ViewMISchedCutoff);
3539 /// If you want to override the dot attributes printed for a particular
3540 /// edge, override this method.
3541 static std::string getEdgeAttributes(const SUnit *Node,
3543 const ScheduleDAG *Graph) {
3544 if (EI.isArtificialDep())
3545 return "color=cyan,style=dashed";
3547 return "color=blue,style=dashed";
3551 static std::string getNodeLabel(const SUnit *SU, const ScheduleDAG *G) {
3553 raw_string_ostream SS(Str);
3554 const ScheduleDAGMI *DAG = static_cast<const ScheduleDAGMI*>(G);
3555 const SchedDFSResult *DFS = DAG->hasVRegLiveness() ?
3556 static_cast<const ScheduleDAGMILive*>(G)->getDFSResult() : nullptr;
3557 SS << "SU:" << SU->NodeNum;
3559 SS << " I:" << DFS->getNumInstrs(SU);
3562 static std::string getNodeDescription(const SUnit *SU, const ScheduleDAG *G) {
3563 return G->getGraphNodeLabel(SU);
3566 static std::string getNodeAttributes(const SUnit *N, const ScheduleDAG *G) {
3567 std::string Str("shape=Mrecord");
3568 const ScheduleDAGMI *DAG = static_cast<const ScheduleDAGMI*>(G);
3569 const SchedDFSResult *DFS = DAG->hasVRegLiveness() ?
3570 static_cast<const ScheduleDAGMILive*>(G)->getDFSResult() : nullptr;
3572 Str += ",style=filled,fillcolor=\"#";
3573 Str += DOT::getColorString(DFS->getSubtreeID(N));
3582 /// viewGraph - Pop up a ghostview window with the reachable parts of the DAG
3583 /// rendered using 'dot'.
3585 void ScheduleDAGMI::viewGraph(const Twine &Name, const Twine &Title) {
3587 ViewGraph(this, Name, false, Title);
3589 errs() << "ScheduleDAGMI::viewGraph is only available in debug builds on "
3590 << "systems with Graphviz or gv!\n";
3594 /// Out-of-line implementation with no arguments is handy for gdb.
3595 void ScheduleDAGMI::viewGraph() {
3596 viewGraph(getDAGName(), "Scheduling-Units Graph for " + getDAGName());