1 //===- MachineScheduler.cpp - Machine Instruction Scheduler ---------------===//
2 //
3 // The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // MachineScheduler schedules machine instructions after phi elimination. It
11 // preserves LiveIntervals so it can be invoked before register allocation.
12 //
13 //===----------------------------------------------------------------------===//
14
15 #include "llvm/CodeGen/MachineScheduler.h"
16 #include "llvm/ADT/PriorityQueue.h"
17 #include "llvm/Analysis/AliasAnalysis.h"
18 #include "llvm/CodeGen/LiveIntervalAnalysis.h"
19 #include "llvm/CodeGen/MachineDominators.h"
20 #include "llvm/CodeGen/MachineLoopInfo.h"
21 #include "llvm/CodeGen/MachineRegisterInfo.h"
22 #include "llvm/CodeGen/Passes.h"
23 #include "llvm/CodeGen/RegisterClassInfo.h"
24 #include "llvm/CodeGen/ScheduleDFS.h"
25 #include "llvm/CodeGen/ScheduleHazardRecognizer.h"
26 #include "llvm/Support/CommandLine.h"
27 #include "llvm/Support/Debug.h"
28 #include "llvm/Support/ErrorHandling.h"
29 #include "llvm/Support/GraphWriter.h"
30 #include "llvm/Support/raw_ostream.h"
31 #include "llvm/Target/TargetInstrInfo.h"
32 #include <queue>
33
34 using namespace llvm;
35
36 #define DEBUG_TYPE "misched"
37
38 namespace llvm {
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"));
43 cl::opt<bool>
44 DumpCriticalPathLength("misched-dcpl", cl::Hidden,
45 cl::desc("Print critical path length to stdout"));
46 }
47
48 #ifndef NDEBUG
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"));
51
52 static cl::opt<unsigned> MISchedCutoff("misched-cutoff", cl::Hidden,
53 cl::desc("Stop scheduling after N instructions"), cl::init(~0U));
54
55 static cl::opt<std::string> SchedOnlyFunc("misched-only-func", cl::Hidden,
56 cl::desc("Only schedule this function"));
57 static cl::opt<unsigned> SchedOnlyBlock("misched-only-block", cl::Hidden,
58 cl::desc("Only schedule this MBB#"));
59 #else
60 static bool ViewMISchedDAGs = false;
61 #endif // NDEBUG
62
63 static cl::opt<bool> EnableRegPressure("misched-regpressure", cl::Hidden,
64 cl::desc("Enable register pressure scheduling."), cl::init(true));
65
66 static cl::opt<bool> EnableCyclicPath("misched-cyclicpath", cl::Hidden,
67 cl::desc("Enable cyclic critical path analysis."), cl::init(true));
68
69 static cl::opt<bool> EnableLoadCluster("misched-cluster", cl::Hidden,
70 cl::desc("Enable load clustering."), cl::init(true));
71
72 // Experimental heuristics
73 static cl::opt<bool> EnableMacroFusion("misched-fusion", cl::Hidden,
74 cl::desc("Enable scheduling for macro fusion."), cl::init(true));
75
76 static cl::opt<bool> VerifyScheduling("verify-misched", cl::Hidden,
77 cl::desc("Verify machine instrs before and after machine scheduling"));
78
79 // DAG subtrees must have at least this many nodes.
80 static const unsigned MinSubtreeSize = 8;
81
82 // Pin the vtables to this file.
anchor()83 void MachineSchedStrategy::anchor() {}
anchor()84 void ScheduleDAGMutation::anchor() {}
85
86 //===----------------------------------------------------------------------===//
87 // Machine Instruction Scheduling Pass and Registry
88 //===----------------------------------------------------------------------===//
89
MachineSchedContext()90 MachineSchedContext::MachineSchedContext():
91 MF(nullptr), MLI(nullptr), MDT(nullptr), PassConfig(nullptr), AA(nullptr), LIS(nullptr) {
92 RegClassInfo = new RegisterClassInfo();
93 }
94
~MachineSchedContext()95 MachineSchedContext::~MachineSchedContext() {
96 delete RegClassInfo;
97 }
98
99 namespace {
100 /// Base class for a machine scheduler class that can run at any point.
101 class MachineSchedulerBase : public MachineSchedContext,
102 public MachineFunctionPass {
103 public:
MachineSchedulerBase(char & ID)104 MachineSchedulerBase(char &ID): MachineFunctionPass(ID) {}
105
106 void print(raw_ostream &O, const Module* = nullptr) const override;
107
108 protected:
109 void scheduleRegions(ScheduleDAGInstrs &Scheduler);
110 };
111
112 /// MachineScheduler runs after coalescing and before register allocation.
113 class MachineScheduler : public MachineSchedulerBase {
114 public:
115 MachineScheduler();
116
117 void getAnalysisUsage(AnalysisUsage &AU) const override;
118
119 bool runOnMachineFunction(MachineFunction&) override;
120
121 static char ID; // Class identification, replacement for typeinfo
122
123 protected:
124 ScheduleDAGInstrs *createMachineScheduler();
125 };
126
127 /// PostMachineScheduler runs after shortly before code emission.
128 class PostMachineScheduler : public MachineSchedulerBase {
129 public:
130 PostMachineScheduler();
131
132 void getAnalysisUsage(AnalysisUsage &AU) const override;
133
134 bool runOnMachineFunction(MachineFunction&) override;
135
136 static char ID; // Class identification, replacement for typeinfo
137
138 protected:
139 ScheduleDAGInstrs *createPostMachineScheduler();
140 };
141 } // namespace
142
143 char MachineScheduler::ID = 0;
144
145 char &llvm::MachineSchedulerID = MachineScheduler::ID;
146
147 INITIALIZE_PASS_BEGIN(MachineScheduler, "machine-scheduler",
148 "Machine Instruction Scheduler", false, false)
INITIALIZE_AG_DEPENDENCY(AliasAnalysis)149 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
150 INITIALIZE_PASS_DEPENDENCY(SlotIndexes)
151 INITIALIZE_PASS_DEPENDENCY(LiveIntervals)
152 INITIALIZE_PASS_END(MachineScheduler, "machine-scheduler",
153 "Machine Instruction Scheduler", false, false)
154
155 MachineScheduler::MachineScheduler()
156 : MachineSchedulerBase(ID) {
157 initializeMachineSchedulerPass(*PassRegistry::getPassRegistry());
158 }
159
getAnalysisUsage(AnalysisUsage & AU) const160 void MachineScheduler::getAnalysisUsage(AnalysisUsage &AU) const {
161 AU.setPreservesCFG();
162 AU.addRequiredID(MachineDominatorsID);
163 AU.addRequired<MachineLoopInfo>();
164 AU.addRequired<AliasAnalysis>();
165 AU.addRequired<TargetPassConfig>();
166 AU.addRequired<SlotIndexes>();
167 AU.addPreserved<SlotIndexes>();
168 AU.addRequired<LiveIntervals>();
169 AU.addPreserved<LiveIntervals>();
170 MachineFunctionPass::getAnalysisUsage(AU);
171 }
172
173 char PostMachineScheduler::ID = 0;
174
175 char &llvm::PostMachineSchedulerID = PostMachineScheduler::ID;
176
177 INITIALIZE_PASS(PostMachineScheduler, "postmisched",
178 "PostRA Machine Instruction Scheduler", false, false)
179
PostMachineScheduler()180 PostMachineScheduler::PostMachineScheduler()
181 : MachineSchedulerBase(ID) {
182 initializePostMachineSchedulerPass(*PassRegistry::getPassRegistry());
183 }
184
getAnalysisUsage(AnalysisUsage & AU) const185 void PostMachineScheduler::getAnalysisUsage(AnalysisUsage &AU) const {
186 AU.setPreservesCFG();
187 AU.addRequiredID(MachineDominatorsID);
188 AU.addRequired<MachineLoopInfo>();
189 AU.addRequired<TargetPassConfig>();
190 MachineFunctionPass::getAnalysisUsage(AU);
191 }
192
193 MachinePassRegistry MachineSchedRegistry::Registry;
194
195 /// A dummy default scheduler factory indicates whether the scheduler
196 /// is overridden on the command line.
useDefaultMachineSched(MachineSchedContext * C)197 static ScheduleDAGInstrs *useDefaultMachineSched(MachineSchedContext *C) {
198 return nullptr;
199 }
200
201 /// MachineSchedOpt allows command line selection of the scheduler.
202 static cl::opt<MachineSchedRegistry::ScheduleDAGCtor, false,
203 RegisterPassParser<MachineSchedRegistry> >
204 MachineSchedOpt("misched",
205 cl::init(&useDefaultMachineSched), cl::Hidden,
206 cl::desc("Machine instruction scheduler to use"));
207
208 static MachineSchedRegistry
209 DefaultSchedRegistry("default", "Use the target's default scheduler choice.",
210 useDefaultMachineSched);
211
212 /// Forward declare the standard machine scheduler. This will be used as the
213 /// default scheduler if the target does not set a default.
214 static ScheduleDAGInstrs *createGenericSchedLive(MachineSchedContext *C);
215 static ScheduleDAGInstrs *createGenericSchedPostRA(MachineSchedContext *C);
216
217 /// Decrement this iterator until reaching the top or a non-debug instr.
218 static MachineBasicBlock::const_iterator
priorNonDebug(MachineBasicBlock::const_iterator I,MachineBasicBlock::const_iterator Beg)219 priorNonDebug(MachineBasicBlock::const_iterator I,
220 MachineBasicBlock::const_iterator Beg) {
221 assert(I != Beg && "reached the top of the region, cannot decrement");
222 while (--I != Beg) {
223 if (!I->isDebugValue())
224 break;
225 }
226 return I;
227 }
228
229 /// Non-const version.
230 static MachineBasicBlock::iterator
priorNonDebug(MachineBasicBlock::iterator I,MachineBasicBlock::const_iterator Beg)231 priorNonDebug(MachineBasicBlock::iterator I,
232 MachineBasicBlock::const_iterator Beg) {
233 return const_cast<MachineInstr*>(
234 &*priorNonDebug(MachineBasicBlock::const_iterator(I), Beg));
235 }
236
237 /// If this iterator is a debug value, increment until reaching the End or a
238 /// non-debug instruction.
239 static MachineBasicBlock::const_iterator
nextIfDebug(MachineBasicBlock::const_iterator I,MachineBasicBlock::const_iterator End)240 nextIfDebug(MachineBasicBlock::const_iterator I,
241 MachineBasicBlock::const_iterator End) {
242 for(; I != End; ++I) {
243 if (!I->isDebugValue())
244 break;
245 }
246 return I;
247 }
248
249 /// Non-const version.
250 static MachineBasicBlock::iterator
nextIfDebug(MachineBasicBlock::iterator I,MachineBasicBlock::const_iterator End)251 nextIfDebug(MachineBasicBlock::iterator I,
252 MachineBasicBlock::const_iterator End) {
253 // Cast the return value to nonconst MachineInstr, then cast to an
254 // instr_iterator, which does not check for null, finally return a
255 // bundle_iterator.
256 return MachineBasicBlock::instr_iterator(
257 const_cast<MachineInstr*>(
258 &*nextIfDebug(MachineBasicBlock::const_iterator(I), End)));
259 }
260
261 /// Instantiate a ScheduleDAGInstrs that will be owned by the caller.
createMachineScheduler()262 ScheduleDAGInstrs *MachineScheduler::createMachineScheduler() {
263 // Select the scheduler, or set the default.
264 MachineSchedRegistry::ScheduleDAGCtor Ctor = MachineSchedOpt;
265 if (Ctor != useDefaultMachineSched)
266 return Ctor(this);
267
268 // Get the default scheduler set by the target for this function.
269 ScheduleDAGInstrs *Scheduler = PassConfig->createMachineScheduler(this);
270 if (Scheduler)
271 return Scheduler;
272
273 // Default to GenericScheduler.
274 return createGenericSchedLive(this);
275 }
276
277 /// Instantiate a ScheduleDAGInstrs for PostRA scheduling that will be owned by
278 /// the caller. We don't have a command line option to override the postRA
279 /// scheduler. The Target must configure it.
createPostMachineScheduler()280 ScheduleDAGInstrs *PostMachineScheduler::createPostMachineScheduler() {
281 // Get the postRA scheduler set by the target for this function.
282 ScheduleDAGInstrs *Scheduler = PassConfig->createPostMachineScheduler(this);
283 if (Scheduler)
284 return Scheduler;
285
286 // Default to GenericScheduler.
287 return createGenericSchedPostRA(this);
288 }
289
290 /// Top-level MachineScheduler pass driver.
291 ///
292 /// Visit blocks in function order. Divide each block into scheduling regions
293 /// and visit them bottom-up. Visiting regions bottom-up is not required, but is
294 /// consistent with the DAG builder, which traverses the interior of the
295 /// scheduling regions bottom-up.
296 ///
297 /// This design avoids exposing scheduling boundaries to the DAG builder,
298 /// simplifying the DAG builder's support for "special" target instructions.
299 /// At the same time the design allows target schedulers to operate across
300 /// scheduling boundaries, for example to bundle the boudary instructions
301 /// without reordering them. This creates complexity, because the target
302 /// scheduler must update the RegionBegin and RegionEnd positions cached by
303 /// ScheduleDAGInstrs whenever adding or removing instructions. A much simpler
304 /// design would be to split blocks at scheduling boundaries, but LLVM has a
305 /// general bias against block splitting purely for implementation simplicity.
runOnMachineFunction(MachineFunction & mf)306 bool MachineScheduler::runOnMachineFunction(MachineFunction &mf) {
307 DEBUG(dbgs() << "Before MISsched:\n"; mf.print(dbgs()));
308
309 // Initialize the context of the pass.
310 MF = &mf;
311 MLI = &getAnalysis<MachineLoopInfo>();
312 MDT = &getAnalysis<MachineDominatorTree>();
313 PassConfig = &getAnalysis<TargetPassConfig>();
314 AA = &getAnalysis<AliasAnalysis>();
315
316 LIS = &getAnalysis<LiveIntervals>();
317
318 if (VerifyScheduling) {
319 DEBUG(LIS->dump());
320 MF->verify(this, "Before machine scheduling.");
321 }
322 RegClassInfo->runOnMachineFunction(*MF);
323
324 // Instantiate the selected scheduler for this target, function, and
325 // optimization level.
326 std::unique_ptr<ScheduleDAGInstrs> Scheduler(createMachineScheduler());
327 scheduleRegions(*Scheduler);
328
329 DEBUG(LIS->dump());
330 if (VerifyScheduling)
331 MF->verify(this, "After machine scheduling.");
332 return true;
333 }
334
runOnMachineFunction(MachineFunction & mf)335 bool PostMachineScheduler::runOnMachineFunction(MachineFunction &mf) {
336 if (skipOptnoneFunction(*mf.getFunction()))
337 return false;
338
339 const TargetSubtargetInfo &ST =
340 mf.getTarget().getSubtarget<TargetSubtargetInfo>();
341 if (!ST.enablePostMachineScheduler()) {
342 DEBUG(dbgs() << "Subtarget disables post-MI-sched.\n");
343 return false;
344 }
345 DEBUG(dbgs() << "Before post-MI-sched:\n"; mf.print(dbgs()));
346
347 // Initialize the context of the pass.
348 MF = &mf;
349 PassConfig = &getAnalysis<TargetPassConfig>();
350
351 if (VerifyScheduling)
352 MF->verify(this, "Before post machine scheduling.");
353
354 // Instantiate the selected scheduler for this target, function, and
355 // optimization level.
356 std::unique_ptr<ScheduleDAGInstrs> Scheduler(createPostMachineScheduler());
357 scheduleRegions(*Scheduler);
358
359 if (VerifyScheduling)
360 MF->verify(this, "After post machine scheduling.");
361 return true;
362 }
363
364 /// Return true of the given instruction should not be included in a scheduling
365 /// region.
366 ///
367 /// MachineScheduler does not currently support scheduling across calls. To
368 /// handle calls, the DAG builder needs to be modified to create register
369 /// anti/output dependencies on the registers clobbered by the call's regmask
370 /// operand. In PreRA scheduling, the stack pointer adjustment already prevents
371 /// scheduling across calls. In PostRA scheduling, we need the isCall to enforce
372 /// the boundary, but there would be no benefit to postRA scheduling across
373 /// calls this late anyway.
isSchedBoundary(MachineBasicBlock::iterator MI,MachineBasicBlock * MBB,MachineFunction * MF,const TargetInstrInfo * TII,bool IsPostRA)374 static bool isSchedBoundary(MachineBasicBlock::iterator MI,
375 MachineBasicBlock *MBB,
376 MachineFunction *MF,
377 const TargetInstrInfo *TII,
378 bool IsPostRA) {
379 return MI->isCall() || TII->isSchedulingBoundary(MI, MBB, *MF);
380 }
381
382 /// Main driver for both MachineScheduler and PostMachineScheduler.
scheduleRegions(ScheduleDAGInstrs & Scheduler)383 void MachineSchedulerBase::scheduleRegions(ScheduleDAGInstrs &Scheduler) {
384 const TargetInstrInfo *TII = MF->getSubtarget().getInstrInfo();
385 bool IsPostRA = Scheduler.isPostRA();
386
387 // Visit all machine basic blocks.
388 //
389 // TODO: Visit blocks in global postorder or postorder within the bottom-up
390 // loop tree. Then we can optionally compute global RegPressure.
391 for (MachineFunction::iterator MBB = MF->begin(), MBBEnd = MF->end();
392 MBB != MBBEnd; ++MBB) {
393
394 Scheduler.startBlock(MBB);
395
396 #ifndef NDEBUG
397 if (SchedOnlyFunc.getNumOccurrences() && SchedOnlyFunc != MF->getName())
398 continue;
399 if (SchedOnlyBlock.getNumOccurrences()
400 && (int)SchedOnlyBlock != MBB->getNumber())
401 continue;
402 #endif
403
404 // Break the block into scheduling regions [I, RegionEnd), and schedule each
405 // region as soon as it is discovered. RegionEnd points the scheduling
406 // boundary at the bottom of the region. The DAG does not include RegionEnd,
407 // but the region does (i.e. the next RegionEnd is above the previous
408 // RegionBegin). If the current block has no terminator then RegionEnd ==
409 // MBB->end() for the bottom region.
410 //
411 // The Scheduler may insert instructions during either schedule() or
412 // exitRegion(), even for empty regions. So the local iterators 'I' and
413 // 'RegionEnd' are invalid across these calls.
414 //
415 // MBB::size() uses instr_iterator to count. Here we need a bundle to count
416 // as a single instruction.
417 unsigned RemainingInstrs = std::distance(MBB->begin(), MBB->end());
418 for(MachineBasicBlock::iterator RegionEnd = MBB->end();
419 RegionEnd != MBB->begin(); RegionEnd = Scheduler.begin()) {
420
421 // Avoid decrementing RegionEnd for blocks with no terminator.
422 if (RegionEnd != MBB->end() ||
423 isSchedBoundary(std::prev(RegionEnd), MBB, MF, TII, IsPostRA)) {
424 --RegionEnd;
425 // Count the boundary instruction.
426 --RemainingInstrs;
427 }
428
429 // The next region starts above the previous region. Look backward in the
430 // instruction stream until we find the nearest boundary.
431 unsigned NumRegionInstrs = 0;
432 MachineBasicBlock::iterator I = RegionEnd;
433 for(;I != MBB->begin(); --I, --RemainingInstrs) {
434 if (isSchedBoundary(std::prev(I), MBB, MF, TII, IsPostRA))
435 break;
436 if (!I->isDebugValue())
437 ++NumRegionInstrs;
438 }
439 // Notify the scheduler of the region, even if we may skip scheduling
440 // it. Perhaps it still needs to be bundled.
441 Scheduler.enterRegion(MBB, I, RegionEnd, NumRegionInstrs);
442
443 // Skip empty scheduling regions (0 or 1 schedulable instructions).
444 if (I == RegionEnd || I == std::prev(RegionEnd)) {
445 // Close the current region. Bundle the terminator if needed.
446 // This invalidates 'RegionEnd' and 'I'.
447 Scheduler.exitRegion();
448 continue;
449 }
450 DEBUG(dbgs() << "********** " << ((Scheduler.isPostRA()) ? "PostRA " : "")
451 << "MI Scheduling **********\n");
452 DEBUG(dbgs() << MF->getName()
453 << ":BB#" << MBB->getNumber() << " " << MBB->getName()
454 << "\n From: " << *I << " To: ";
455 if (RegionEnd != MBB->end()) dbgs() << *RegionEnd;
456 else dbgs() << "End";
457 dbgs() << " RegionInstrs: " << NumRegionInstrs
458 << " Remaining: " << RemainingInstrs << "\n");
459 if (DumpCriticalPathLength) {
460 errs() << MF->getName();
461 errs() << ":BB# " << MBB->getNumber();
462 errs() << " " << MBB->getName() << " \n";
463 }
464
465 // Schedule a region: possibly reorder instructions.
466 // This invalidates 'RegionEnd' and 'I'.
467 Scheduler.schedule();
468
469 // Close the current region.
470 Scheduler.exitRegion();
471
472 // Scheduling has invalidated the current iterator 'I'. Ask the
473 // scheduler for the top of it's scheduled region.
474 RegionEnd = Scheduler.begin();
475 }
476 assert(RemainingInstrs == 0 && "Instruction count mismatch!");
477 Scheduler.finishBlock();
478 if (Scheduler.isPostRA()) {
479 // FIXME: Ideally, no further passes should rely on kill flags. However,
480 // thumb2 size reduction is currently an exception.
481 Scheduler.fixupKills(MBB);
482 }
483 }
484 Scheduler.finalizeSchedule();
485 }
486
print(raw_ostream & O,const Module * m) const487 void MachineSchedulerBase::print(raw_ostream &O, const Module* m) const {
488 // unimplemented
489 }
490
491 LLVM_DUMP_METHOD
dump()492 void ReadyQueue::dump() {
493 dbgs() << Name << ": ";
494 for (unsigned i = 0, e = Queue.size(); i < e; ++i)
495 dbgs() << Queue[i]->NodeNum << " ";
496 dbgs() << "\n";
497 }
498
499 //===----------------------------------------------------------------------===//
500 // ScheduleDAGMI - Basic machine instruction scheduling. This is
501 // independent of PreRA/PostRA scheduling and involves no extra book-keeping for
502 // virtual registers.
503 // ===----------------------------------------------------------------------===/
504
505 // Provide a vtable anchor.
~ScheduleDAGMI()506 ScheduleDAGMI::~ScheduleDAGMI() {
507 }
508
canAddEdge(SUnit * SuccSU,SUnit * PredSU)509 bool ScheduleDAGMI::canAddEdge(SUnit *SuccSU, SUnit *PredSU) {
510 return SuccSU == &ExitSU || !Topo.IsReachable(PredSU, SuccSU);
511 }
512
addEdge(SUnit * SuccSU,const SDep & PredDep)513 bool ScheduleDAGMI::addEdge(SUnit *SuccSU, const SDep &PredDep) {
514 if (SuccSU != &ExitSU) {
515 // Do not use WillCreateCycle, it assumes SD scheduling.
516 // If Pred is reachable from Succ, then the edge creates a cycle.
517 if (Topo.IsReachable(PredDep.getSUnit(), SuccSU))
518 return false;
519 Topo.AddPred(SuccSU, PredDep.getSUnit());
520 }
521 SuccSU->addPred(PredDep, /*Required=*/!PredDep.isArtificial());
522 // Return true regardless of whether a new edge needed to be inserted.
523 return true;
524 }
525
526 /// ReleaseSucc - Decrement the NumPredsLeft count of a successor. When
527 /// NumPredsLeft reaches zero, release the successor node.
528 ///
529 /// FIXME: Adjust SuccSU height based on MinLatency.
releaseSucc(SUnit * SU,SDep * SuccEdge)530 void ScheduleDAGMI::releaseSucc(SUnit *SU, SDep *SuccEdge) {
531 SUnit *SuccSU = SuccEdge->getSUnit();
532
533 if (SuccEdge->isWeak()) {
534 --SuccSU->WeakPredsLeft;
535 if (SuccEdge->isCluster())
536 NextClusterSucc = SuccSU;
537 return;
538 }
539 #ifndef NDEBUG
540 if (SuccSU->NumPredsLeft == 0) {
541 dbgs() << "*** Scheduling failed! ***\n";
542 SuccSU->dump(this);
543 dbgs() << " has been released too many times!\n";
544 llvm_unreachable(nullptr);
545 }
546 #endif
547 // SU->TopReadyCycle was set to CurrCycle when it was scheduled. However,
548 // CurrCycle may have advanced since then.
549 if (SuccSU->TopReadyCycle < SU->TopReadyCycle + SuccEdge->getLatency())
550 SuccSU->TopReadyCycle = SU->TopReadyCycle + SuccEdge->getLatency();
551
552 --SuccSU->NumPredsLeft;
553 if (SuccSU->NumPredsLeft == 0 && SuccSU != &ExitSU)
554 SchedImpl->releaseTopNode(SuccSU);
555 }
556
557 /// releaseSuccessors - Call releaseSucc on each of SU's successors.
releaseSuccessors(SUnit * SU)558 void ScheduleDAGMI::releaseSuccessors(SUnit *SU) {
559 for (SUnit::succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
560 I != E; ++I) {
561 releaseSucc(SU, &*I);
562 }
563 }
564
565 /// ReleasePred - Decrement the NumSuccsLeft count of a predecessor. When
566 /// NumSuccsLeft reaches zero, release the predecessor node.
567 ///
568 /// FIXME: Adjust PredSU height based on MinLatency.
releasePred(SUnit * SU,SDep * PredEdge)569 void ScheduleDAGMI::releasePred(SUnit *SU, SDep *PredEdge) {
570 SUnit *PredSU = PredEdge->getSUnit();
571
572 if (PredEdge->isWeak()) {
573 --PredSU->WeakSuccsLeft;
574 if (PredEdge->isCluster())
575 NextClusterPred = PredSU;
576 return;
577 }
578 #ifndef NDEBUG
579 if (PredSU->NumSuccsLeft == 0) {
580 dbgs() << "*** Scheduling failed! ***\n";
581 PredSU->dump(this);
582 dbgs() << " has been released too many times!\n";
583 llvm_unreachable(nullptr);
584 }
585 #endif
586 // SU->BotReadyCycle was set to CurrCycle when it was scheduled. However,
587 // CurrCycle may have advanced since then.
588 if (PredSU->BotReadyCycle < SU->BotReadyCycle + PredEdge->getLatency())
589 PredSU->BotReadyCycle = SU->BotReadyCycle + PredEdge->getLatency();
590
591 --PredSU->NumSuccsLeft;
592 if (PredSU->NumSuccsLeft == 0 && PredSU != &EntrySU)
593 SchedImpl->releaseBottomNode(PredSU);
594 }
595
596 /// releasePredecessors - Call releasePred on each of SU's predecessors.
releasePredecessors(SUnit * SU)597 void ScheduleDAGMI::releasePredecessors(SUnit *SU) {
598 for (SUnit::pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
599 I != E; ++I) {
600 releasePred(SU, &*I);
601 }
602 }
603
604 /// enterRegion - Called back from MachineScheduler::runOnMachineFunction after
605 /// crossing a scheduling boundary. [begin, end) includes all instructions in
606 /// the region, including the boundary itself and single-instruction regions
607 /// that don't get scheduled.
enterRegion(MachineBasicBlock * bb,MachineBasicBlock::iterator begin,MachineBasicBlock::iterator end,unsigned regioninstrs)608 void ScheduleDAGMI::enterRegion(MachineBasicBlock *bb,
609 MachineBasicBlock::iterator begin,
610 MachineBasicBlock::iterator end,
611 unsigned regioninstrs)
612 {
613 ScheduleDAGInstrs::enterRegion(bb, begin, end, regioninstrs);
614
615 SchedImpl->initPolicy(begin, end, regioninstrs);
616 }
617
618 /// This is normally called from the main scheduler loop but may also be invoked
619 /// by the scheduling strategy to perform additional code motion.
moveInstruction(MachineInstr * MI,MachineBasicBlock::iterator InsertPos)620 void ScheduleDAGMI::moveInstruction(
621 MachineInstr *MI, MachineBasicBlock::iterator InsertPos) {
622 // Advance RegionBegin if the first instruction moves down.
623 if (&*RegionBegin == MI)
624 ++RegionBegin;
625
626 // Update the instruction stream.
627 BB->splice(InsertPos, BB, MI);
628
629 // Update LiveIntervals
630 if (LIS)
631 LIS->handleMove(MI, /*UpdateFlags=*/true);
632
633 // Recede RegionBegin if an instruction moves above the first.
634 if (RegionBegin == InsertPos)
635 RegionBegin = MI;
636 }
637
checkSchedLimit()638 bool ScheduleDAGMI::checkSchedLimit() {
639 #ifndef NDEBUG
640 if (NumInstrsScheduled == MISchedCutoff && MISchedCutoff != ~0U) {
641 CurrentTop = CurrentBottom;
642 return false;
643 }
644 ++NumInstrsScheduled;
645 #endif
646 return true;
647 }
648
649 /// Per-region scheduling driver, called back from
650 /// MachineScheduler::runOnMachineFunction. This is a simplified driver that
651 /// does not consider liveness or register pressure. It is useful for PostRA
652 /// scheduling and potentially other custom schedulers.
schedule()653 void ScheduleDAGMI::schedule() {
654 // Build the DAG.
655 buildSchedGraph(AA);
656
657 Topo.InitDAGTopologicalSorting();
658
659 postprocessDAG();
660
661 SmallVector<SUnit*, 8> TopRoots, BotRoots;
662 findRootsAndBiasEdges(TopRoots, BotRoots);
663
664 // Initialize the strategy before modifying the DAG.
665 // This may initialize a DFSResult to be used for queue priority.
666 SchedImpl->initialize(this);
667
668 DEBUG(for (unsigned su = 0, e = SUnits.size(); su != e; ++su)
669 SUnits[su].dumpAll(this));
670 if (ViewMISchedDAGs) viewGraph();
671
672 // Initialize ready queues now that the DAG and priority data are finalized.
673 initQueues(TopRoots, BotRoots);
674
675 bool IsTopNode = false;
676 while (SUnit *SU = SchedImpl->pickNode(IsTopNode)) {
677 assert(!SU->isScheduled && "Node already scheduled");
678 if (!checkSchedLimit())
679 break;
680
681 MachineInstr *MI = SU->getInstr();
682 if (IsTopNode) {
683 assert(SU->isTopReady() && "node still has unscheduled dependencies");
684 if (&*CurrentTop == MI)
685 CurrentTop = nextIfDebug(++CurrentTop, CurrentBottom);
686 else
687 moveInstruction(MI, CurrentTop);
688 }
689 else {
690 assert(SU->isBottomReady() && "node still has unscheduled dependencies");
691 MachineBasicBlock::iterator priorII =
692 priorNonDebug(CurrentBottom, CurrentTop);
693 if (&*priorII == MI)
694 CurrentBottom = priorII;
695 else {
696 if (&*CurrentTop == MI)
697 CurrentTop = nextIfDebug(++CurrentTop, priorII);
698 moveInstruction(MI, CurrentBottom);
699 CurrentBottom = MI;
700 }
701 }
702 // Notify the scheduling strategy before updating the DAG.
703 // This sets the scheduled node's ReadyCycle to CurrCycle. When updateQueues
704 // runs, it can then use the accurate ReadyCycle time to determine whether
705 // newly released nodes can move to the readyQ.
706 SchedImpl->schedNode(SU, IsTopNode);
707
708 updateQueues(SU, IsTopNode);
709 }
710 assert(CurrentTop == CurrentBottom && "Nonempty unscheduled zone.");
711
712 placeDebugValues();
713
714 DEBUG({
715 unsigned BBNum = begin()->getParent()->getNumber();
716 dbgs() << "*** Final schedule for BB#" << BBNum << " ***\n";
717 dumpSchedule();
718 dbgs() << '\n';
719 });
720 }
721
722 /// Apply each ScheduleDAGMutation step in order.
postprocessDAG()723 void ScheduleDAGMI::postprocessDAG() {
724 for (unsigned i = 0, e = Mutations.size(); i < e; ++i) {
725 Mutations[i]->apply(this);
726 }
727 }
728
729 void ScheduleDAGMI::
findRootsAndBiasEdges(SmallVectorImpl<SUnit * > & TopRoots,SmallVectorImpl<SUnit * > & BotRoots)730 findRootsAndBiasEdges(SmallVectorImpl<SUnit*> &TopRoots,
731 SmallVectorImpl<SUnit*> &BotRoots) {
732 for (std::vector<SUnit>::iterator
733 I = SUnits.begin(), E = SUnits.end(); I != E; ++I) {
734 SUnit *SU = &(*I);
735 assert(!SU->isBoundaryNode() && "Boundary node should not be in SUnits");
736
737 // Order predecessors so DFSResult follows the critical path.
738 SU->biasCriticalPath();
739
740 // A SUnit is ready to top schedule if it has no predecessors.
741 if (!I->NumPredsLeft)
742 TopRoots.push_back(SU);
743 // A SUnit is ready to bottom schedule if it has no successors.
744 if (!I->NumSuccsLeft)
745 BotRoots.push_back(SU);
746 }
747 ExitSU.biasCriticalPath();
748 }
749
750 /// Identify DAG roots and setup scheduler queues.
initQueues(ArrayRef<SUnit * > TopRoots,ArrayRef<SUnit * > BotRoots)751 void ScheduleDAGMI::initQueues(ArrayRef<SUnit*> TopRoots,
752 ArrayRef<SUnit*> BotRoots) {
753 NextClusterSucc = nullptr;
754 NextClusterPred = nullptr;
755
756 // Release all DAG roots for scheduling, not including EntrySU/ExitSU.
757 //
758 // Nodes with unreleased weak edges can still be roots.
759 // Release top roots in forward order.
760 for (SmallVectorImpl<SUnit*>::const_iterator
761 I = TopRoots.begin(), E = TopRoots.end(); I != E; ++I) {
762 SchedImpl->releaseTopNode(*I);
763 }
764 // Release bottom roots in reverse order so the higher priority nodes appear
765 // first. This is more natural and slightly more efficient.
766 for (SmallVectorImpl<SUnit*>::const_reverse_iterator
767 I = BotRoots.rbegin(), E = BotRoots.rend(); I != E; ++I) {
768 SchedImpl->releaseBottomNode(*I);
769 }
770
771 releaseSuccessors(&EntrySU);
772 releasePredecessors(&ExitSU);
773
774 SchedImpl->registerRoots();
775
776 // Advance past initial DebugValues.
777 CurrentTop = nextIfDebug(RegionBegin, RegionEnd);
778 CurrentBottom = RegionEnd;
779 }
780
781 /// Update scheduler queues after scheduling an instruction.
updateQueues(SUnit * SU,bool IsTopNode)782 void ScheduleDAGMI::updateQueues(SUnit *SU, bool IsTopNode) {
783 // Release dependent instructions for scheduling.
784 if (IsTopNode)
785 releaseSuccessors(SU);
786 else
787 releasePredecessors(SU);
788
789 SU->isScheduled = true;
790 }
791
792 /// Reinsert any remaining debug_values, just like the PostRA scheduler.
placeDebugValues()793 void ScheduleDAGMI::placeDebugValues() {
794 // If first instruction was a DBG_VALUE then put it back.
795 if (FirstDbgValue) {
796 BB->splice(RegionBegin, BB, FirstDbgValue);
797 RegionBegin = FirstDbgValue;
798 }
799
800 for (std::vector<std::pair<MachineInstr *, MachineInstr *> >::iterator
801 DI = DbgValues.end(), DE = DbgValues.begin(); DI != DE; --DI) {
802 std::pair<MachineInstr *, MachineInstr *> P = *std::prev(DI);
803 MachineInstr *DbgValue = P.first;
804 MachineBasicBlock::iterator OrigPrevMI = P.second;
805 if (&*RegionBegin == DbgValue)
806 ++RegionBegin;
807 BB->splice(++OrigPrevMI, BB, DbgValue);
808 if (OrigPrevMI == std::prev(RegionEnd))
809 RegionEnd = DbgValue;
810 }
811 DbgValues.clear();
812 FirstDbgValue = nullptr;
813 }
814
815 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
dumpSchedule() const816 void ScheduleDAGMI::dumpSchedule() const {
817 for (MachineBasicBlock::iterator MI = begin(), ME = end(); MI != ME; ++MI) {
818 if (SUnit *SU = getSUnit(&(*MI)))
819 SU->dump(this);
820 else
821 dbgs() << "Missing SUnit\n";
822 }
823 }
824 #endif
825
826 //===----------------------------------------------------------------------===//
827 // ScheduleDAGMILive - Base class for MachineInstr scheduling with LiveIntervals
828 // preservation.
829 //===----------------------------------------------------------------------===//
830
~ScheduleDAGMILive()831 ScheduleDAGMILive::~ScheduleDAGMILive() {
832 delete DFSResult;
833 }
834
835 /// enterRegion - Called back from MachineScheduler::runOnMachineFunction after
836 /// crossing a scheduling boundary. [begin, end) includes all instructions in
837 /// the region, including the boundary itself and single-instruction regions
838 /// that don't get scheduled.
enterRegion(MachineBasicBlock * bb,MachineBasicBlock::iterator begin,MachineBasicBlock::iterator end,unsigned regioninstrs)839 void ScheduleDAGMILive::enterRegion(MachineBasicBlock *bb,
840 MachineBasicBlock::iterator begin,
841 MachineBasicBlock::iterator end,
842 unsigned regioninstrs)
843 {
844 // ScheduleDAGMI initializes SchedImpl's per-region policy.
845 ScheduleDAGMI::enterRegion(bb, begin, end, regioninstrs);
846
847 // For convenience remember the end of the liveness region.
848 LiveRegionEnd = (RegionEnd == bb->end()) ? RegionEnd : std::next(RegionEnd);
849
850 SUPressureDiffs.clear();
851
852 ShouldTrackPressure = SchedImpl->shouldTrackPressure();
853 }
854
855 // Setup the register pressure trackers for the top scheduled top and bottom
856 // scheduled regions.
initRegPressure()857 void ScheduleDAGMILive::initRegPressure() {
858 TopRPTracker.init(&MF, RegClassInfo, LIS, BB, RegionBegin);
859 BotRPTracker.init(&MF, RegClassInfo, LIS, BB, LiveRegionEnd);
860
861 // Close the RPTracker to finalize live ins.
862 RPTracker.closeRegion();
863
864 DEBUG(RPTracker.dump());
865
866 // Initialize the live ins and live outs.
867 TopRPTracker.addLiveRegs(RPTracker.getPressure().LiveInRegs);
868 BotRPTracker.addLiveRegs(RPTracker.getPressure().LiveOutRegs);
869
870 // Close one end of the tracker so we can call
871 // getMaxUpward/DownwardPressureDelta before advancing across any
872 // instructions. This converts currently live regs into live ins/outs.
873 TopRPTracker.closeTop();
874 BotRPTracker.closeBottom();
875
876 BotRPTracker.initLiveThru(RPTracker);
877 if (!BotRPTracker.getLiveThru().empty()) {
878 TopRPTracker.initLiveThru(BotRPTracker.getLiveThru());
879 DEBUG(dbgs() << "Live Thru: ";
880 dumpRegSetPressure(BotRPTracker.getLiveThru(), TRI));
881 };
882
883 // For each live out vreg reduce the pressure change associated with other
884 // uses of the same vreg below the live-out reaching def.
885 updatePressureDiffs(RPTracker.getPressure().LiveOutRegs);
886
887 // Account for liveness generated by the region boundary.
888 if (LiveRegionEnd != RegionEnd) {
889 SmallVector<unsigned, 8> LiveUses;
890 BotRPTracker.recede(&LiveUses);
891 updatePressureDiffs(LiveUses);
892 }
893
894 assert(BotRPTracker.getPos() == RegionEnd && "Can't find the region bottom");
895
896 // Cache the list of excess pressure sets in this region. This will also track
897 // the max pressure in the scheduled code for these sets.
898 RegionCriticalPSets.clear();
899 const std::vector<unsigned> &RegionPressure =
900 RPTracker.getPressure().MaxSetPressure;
901 for (unsigned i = 0, e = RegionPressure.size(); i < e; ++i) {
902 unsigned Limit = RegClassInfo->getRegPressureSetLimit(i);
903 if (RegionPressure[i] > Limit) {
904 DEBUG(dbgs() << TRI->getRegPressureSetName(i)
905 << " Limit " << Limit
906 << " Actual " << RegionPressure[i] << "\n");
907 RegionCriticalPSets.push_back(PressureChange(i));
908 }
909 }
910 DEBUG(dbgs() << "Excess PSets: ";
911 for (unsigned i = 0, e = RegionCriticalPSets.size(); i != e; ++i)
912 dbgs() << TRI->getRegPressureSetName(
913 RegionCriticalPSets[i].getPSet()) << " ";
914 dbgs() << "\n");
915 }
916
917 void ScheduleDAGMILive::
updateScheduledPressure(const SUnit * SU,const std::vector<unsigned> & NewMaxPressure)918 updateScheduledPressure(const SUnit *SU,
919 const std::vector<unsigned> &NewMaxPressure) {
920 const PressureDiff &PDiff = getPressureDiff(SU);
921 unsigned CritIdx = 0, CritEnd = RegionCriticalPSets.size();
922 for (PressureDiff::const_iterator I = PDiff.begin(), E = PDiff.end();
923 I != E; ++I) {
924 if (!I->isValid())
925 break;
926 unsigned ID = I->getPSet();
927 while (CritIdx != CritEnd && RegionCriticalPSets[CritIdx].getPSet() < ID)
928 ++CritIdx;
929 if (CritIdx != CritEnd && RegionCriticalPSets[CritIdx].getPSet() == ID) {
930 if ((int)NewMaxPressure[ID] > RegionCriticalPSets[CritIdx].getUnitInc()
931 && NewMaxPressure[ID] <= INT16_MAX)
932 RegionCriticalPSets[CritIdx].setUnitInc(NewMaxPressure[ID]);
933 }
934 unsigned Limit = RegClassInfo->getRegPressureSetLimit(ID);
935 if (NewMaxPressure[ID] >= Limit - 2) {
936 DEBUG(dbgs() << " " << TRI->getRegPressureSetName(ID) << ": "
937 << NewMaxPressure[ID] << " > " << Limit << "(+ "
938 << BotRPTracker.getLiveThru()[ID] << " livethru)\n");
939 }
940 }
941 }
942
943 /// Update the PressureDiff array for liveness after scheduling this
944 /// instruction.
updatePressureDiffs(ArrayRef<unsigned> LiveUses)945 void ScheduleDAGMILive::updatePressureDiffs(ArrayRef<unsigned> LiveUses) {
946 for (unsigned LUIdx = 0, LUEnd = LiveUses.size(); LUIdx != LUEnd; ++LUIdx) {
947 /// FIXME: Currently assuming single-use physregs.
948 unsigned Reg = LiveUses[LUIdx];
949 DEBUG(dbgs() << " LiveReg: " << PrintVRegOrUnit(Reg, TRI) << "\n");
950 if (!TRI->isVirtualRegister(Reg))
951 continue;
952
953 // This may be called before CurrentBottom has been initialized. However,
954 // BotRPTracker must have a valid position. We want the value live into the
955 // instruction or live out of the block, so ask for the previous
956 // instruction's live-out.
957 const LiveInterval &LI = LIS->getInterval(Reg);
958 VNInfo *VNI;
959 MachineBasicBlock::const_iterator I =
960 nextIfDebug(BotRPTracker.getPos(), BB->end());
961 if (I == BB->end())
962 VNI = LI.getVNInfoBefore(LIS->getMBBEndIdx(BB));
963 else {
964 LiveQueryResult LRQ = LI.Query(LIS->getInstructionIndex(I));
965 VNI = LRQ.valueIn();
966 }
967 // RegisterPressureTracker guarantees that readsReg is true for LiveUses.
968 assert(VNI && "No live value at use.");
969 for (VReg2UseMap::iterator
970 UI = VRegUses.find(Reg); UI != VRegUses.end(); ++UI) {
971 SUnit *SU = UI->SU;
972 DEBUG(dbgs() << " UpdateRegP: SU(" << SU->NodeNum << ") "
973 << *SU->getInstr());
974 // If this use comes before the reaching def, it cannot be a last use, so
975 // descrease its pressure change.
976 if (!SU->isScheduled && SU != &ExitSU) {
977 LiveQueryResult LRQ
978 = LI.Query(LIS->getInstructionIndex(SU->getInstr()));
979 if (LRQ.valueIn() == VNI)
980 getPressureDiff(SU).addPressureChange(Reg, true, &MRI);
981 }
982 }
983 }
984 }
985
986 /// schedule - Called back from MachineScheduler::runOnMachineFunction
987 /// after setting up the current scheduling region. [RegionBegin, RegionEnd)
988 /// only includes instructions that have DAG nodes, not scheduling boundaries.
989 ///
990 /// This is a skeletal driver, with all the functionality pushed into helpers,
991 /// so that it can be easilly extended by experimental schedulers. Generally,
992 /// implementing MachineSchedStrategy should be sufficient to implement a new
993 /// scheduling algorithm. However, if a scheduler further subclasses
994 /// ScheduleDAGMILive then it will want to override this virtual method in order
995 /// to update any specialized state.
schedule()996 void ScheduleDAGMILive::schedule() {
997 buildDAGWithRegPressure();
998
999 Topo.InitDAGTopologicalSorting();
1000
1001 postprocessDAG();
1002
1003 SmallVector<SUnit*, 8> TopRoots, BotRoots;
1004 findRootsAndBiasEdges(TopRoots, BotRoots);
1005
1006 // Initialize the strategy before modifying the DAG.
1007 // This may initialize a DFSResult to be used for queue priority.
1008 SchedImpl->initialize(this);
1009
1010 DEBUG(for (unsigned su = 0, e = SUnits.size(); su != e; ++su)
1011 SUnits[su].dumpAll(this));
1012 if (ViewMISchedDAGs) viewGraph();
1013
1014 // Initialize ready queues now that the DAG and priority data are finalized.
1015 initQueues(TopRoots, BotRoots);
1016
1017 if (ShouldTrackPressure) {
1018 assert(TopRPTracker.getPos() == RegionBegin && "bad initial Top tracker");
1019 TopRPTracker.setPos(CurrentTop);
1020 }
1021
1022 bool IsTopNode = false;
1023 while (SUnit *SU = SchedImpl->pickNode(IsTopNode)) {
1024 assert(!SU->isScheduled && "Node already scheduled");
1025 if (!checkSchedLimit())
1026 break;
1027
1028 scheduleMI(SU, IsTopNode);
1029
1030 updateQueues(SU, IsTopNode);
1031
1032 if (DFSResult) {
1033 unsigned SubtreeID = DFSResult->getSubtreeID(SU);
1034 if (!ScheduledTrees.test(SubtreeID)) {
1035 ScheduledTrees.set(SubtreeID);
1036 DFSResult->scheduleTree(SubtreeID);
1037 SchedImpl->scheduleTree(SubtreeID);
1038 }
1039 }
1040
1041 // Notify the scheduling strategy after updating the DAG.
1042 SchedImpl->schedNode(SU, IsTopNode);
1043 }
1044 assert(CurrentTop == CurrentBottom && "Nonempty unscheduled zone.");
1045
1046 placeDebugValues();
1047
1048 DEBUG({
1049 unsigned BBNum = begin()->getParent()->getNumber();
1050 dbgs() << "*** Final schedule for BB#" << BBNum << " ***\n";
1051 dumpSchedule();
1052 dbgs() << '\n';
1053 });
1054 }
1055
1056 /// Build the DAG and setup three register pressure trackers.
buildDAGWithRegPressure()1057 void ScheduleDAGMILive::buildDAGWithRegPressure() {
1058 if (!ShouldTrackPressure) {
1059 RPTracker.reset();
1060 RegionCriticalPSets.clear();
1061 buildSchedGraph(AA);
1062 return;
1063 }
1064
1065 // Initialize the register pressure tracker used by buildSchedGraph.
1066 RPTracker.init(&MF, RegClassInfo, LIS, BB, LiveRegionEnd,
1067 /*TrackUntiedDefs=*/true);
1068
1069 // Account for liveness generate by the region boundary.
1070 if (LiveRegionEnd != RegionEnd)
1071 RPTracker.recede();
1072
1073 // Build the DAG, and compute current register pressure.
1074 buildSchedGraph(AA, &RPTracker, &SUPressureDiffs);
1075
1076 // Initialize top/bottom trackers after computing region pressure.
1077 initRegPressure();
1078 }
1079
computeDFSResult()1080 void ScheduleDAGMILive::computeDFSResult() {
1081 if (!DFSResult)
1082 DFSResult = new SchedDFSResult(/*BottomU*/true, MinSubtreeSize);
1083 DFSResult->clear();
1084 ScheduledTrees.clear();
1085 DFSResult->resize(SUnits.size());
1086 DFSResult->compute(SUnits);
1087 ScheduledTrees.resize(DFSResult->getNumSubtrees());
1088 }
1089
1090 /// Compute the max cyclic critical path through the DAG. The scheduling DAG
1091 /// only provides the critical path for single block loops. To handle loops that
1092 /// span blocks, we could use the vreg path latencies provided by
1093 /// MachineTraceMetrics instead. However, MachineTraceMetrics is not currently
1094 /// available for use in the scheduler.
1095 ///
1096 /// The cyclic path estimation identifies a def-use pair that crosses the back
1097 /// edge and considers the depth and height of the nodes. For example, consider
1098 /// the following instruction sequence where each instruction has unit latency
1099 /// and defines an epomymous virtual register:
1100 ///
1101 /// a->b(a,c)->c(b)->d(c)->exit
1102 ///
1103 /// The cyclic critical path is a two cycles: b->c->b
1104 /// The acyclic critical path is four cycles: a->b->c->d->exit
1105 /// LiveOutHeight = height(c) = len(c->d->exit) = 2
1106 /// LiveOutDepth = depth(c) + 1 = len(a->b->c) + 1 = 3
1107 /// LiveInHeight = height(b) + 1 = len(b->c->d->exit) + 1 = 4
1108 /// LiveInDepth = depth(b) = len(a->b) = 1
1109 ///
1110 /// LiveOutDepth - LiveInDepth = 3 - 1 = 2
1111 /// LiveInHeight - LiveOutHeight = 4 - 2 = 2
1112 /// CyclicCriticalPath = min(2, 2) = 2
1113 ///
1114 /// This could be relevant to PostRA scheduling, but is currently implemented
1115 /// assuming LiveIntervals.
computeCyclicCriticalPath()1116 unsigned ScheduleDAGMILive::computeCyclicCriticalPath() {
1117 // This only applies to single block loop.
1118 if (!BB->isSuccessor(BB))
1119 return 0;
1120
1121 unsigned MaxCyclicLatency = 0;
1122 // Visit each live out vreg def to find def/use pairs that cross iterations.
1123 ArrayRef<unsigned> LiveOuts = RPTracker.getPressure().LiveOutRegs;
1124 for (ArrayRef<unsigned>::iterator RI = LiveOuts.begin(), RE = LiveOuts.end();
1125 RI != RE; ++RI) {
1126 unsigned Reg = *RI;
1127 if (!TRI->isVirtualRegister(Reg))
1128 continue;
1129 const LiveInterval &LI = LIS->getInterval(Reg);
1130 const VNInfo *DefVNI = LI.getVNInfoBefore(LIS->getMBBEndIdx(BB));
1131 if (!DefVNI)
1132 continue;
1133
1134 MachineInstr *DefMI = LIS->getInstructionFromIndex(DefVNI->def);
1135 const SUnit *DefSU = getSUnit(DefMI);
1136 if (!DefSU)
1137 continue;
1138
1139 unsigned LiveOutHeight = DefSU->getHeight();
1140 unsigned LiveOutDepth = DefSU->getDepth() + DefSU->Latency;
1141 // Visit all local users of the vreg def.
1142 for (VReg2UseMap::iterator
1143 UI = VRegUses.find(Reg); UI != VRegUses.end(); ++UI) {
1144 if (UI->SU == &ExitSU)
1145 continue;
1146
1147 // Only consider uses of the phi.
1148 LiveQueryResult LRQ =
1149 LI.Query(LIS->getInstructionIndex(UI->SU->getInstr()));
1150 if (!LRQ.valueIn()->isPHIDef())
1151 continue;
1152
1153 // Assume that a path spanning two iterations is a cycle, which could
1154 // overestimate in strange cases. This allows cyclic latency to be
1155 // estimated as the minimum slack of the vreg's depth or height.
1156 unsigned CyclicLatency = 0;
1157 if (LiveOutDepth > UI->SU->getDepth())
1158 CyclicLatency = LiveOutDepth - UI->SU->getDepth();
1159
1160 unsigned LiveInHeight = UI->SU->getHeight() + DefSU->Latency;
1161 if (LiveInHeight > LiveOutHeight) {
1162 if (LiveInHeight - LiveOutHeight < CyclicLatency)
1163 CyclicLatency = LiveInHeight - LiveOutHeight;
1164 }
1165 else
1166 CyclicLatency = 0;
1167
1168 DEBUG(dbgs() << "Cyclic Path: SU(" << DefSU->NodeNum << ") -> SU("
1169 << UI->SU->NodeNum << ") = " << CyclicLatency << "c\n");
1170 if (CyclicLatency > MaxCyclicLatency)
1171 MaxCyclicLatency = CyclicLatency;
1172 }
1173 }
1174 DEBUG(dbgs() << "Cyclic Critical Path: " << MaxCyclicLatency << "c\n");
1175 return MaxCyclicLatency;
1176 }
1177
1178 /// Move an instruction and update register pressure.
scheduleMI(SUnit * SU,bool IsTopNode)1179 void ScheduleDAGMILive::scheduleMI(SUnit *SU, bool IsTopNode) {
1180 // Move the instruction to its new location in the instruction stream.
1181 MachineInstr *MI = SU->getInstr();
1182
1183 if (IsTopNode) {
1184 assert(SU->isTopReady() && "node still has unscheduled dependencies");
1185 if (&*CurrentTop == MI)
1186 CurrentTop = nextIfDebug(++CurrentTop, CurrentBottom);
1187 else {
1188 moveInstruction(MI, CurrentTop);
1189 TopRPTracker.setPos(MI);
1190 }
1191
1192 if (ShouldTrackPressure) {
1193 // Update top scheduled pressure.
1194 TopRPTracker.advance();
1195 assert(TopRPTracker.getPos() == CurrentTop && "out of sync");
1196 updateScheduledPressure(SU, TopRPTracker.getPressure().MaxSetPressure);
1197 }
1198 }
1199 else {
1200 assert(SU->isBottomReady() && "node still has unscheduled dependencies");
1201 MachineBasicBlock::iterator priorII =
1202 priorNonDebug(CurrentBottom, CurrentTop);
1203 if (&*priorII == MI)
1204 CurrentBottom = priorII;
1205 else {
1206 if (&*CurrentTop == MI) {
1207 CurrentTop = nextIfDebug(++CurrentTop, priorII);
1208 TopRPTracker.setPos(CurrentTop);
1209 }
1210 moveInstruction(MI, CurrentBottom);
1211 CurrentBottom = MI;
1212 }
1213 if (ShouldTrackPressure) {
1214 // Update bottom scheduled pressure.
1215 SmallVector<unsigned, 8> LiveUses;
1216 BotRPTracker.recede(&LiveUses);
1217 assert(BotRPTracker.getPos() == CurrentBottom && "out of sync");
1218 updateScheduledPressure(SU, BotRPTracker.getPressure().MaxSetPressure);
1219 updatePressureDiffs(LiveUses);
1220 }
1221 }
1222 }
1223
1224 //===----------------------------------------------------------------------===//
1225 // LoadClusterMutation - DAG post-processing to cluster loads.
1226 //===----------------------------------------------------------------------===//
1227
1228 namespace {
1229 /// \brief Post-process the DAG to create cluster edges between neighboring
1230 /// loads.
1231 class LoadClusterMutation : public ScheduleDAGMutation {
1232 struct LoadInfo {
1233 SUnit *SU;
1234 unsigned BaseReg;
1235 unsigned Offset;
LoadInfo__anonbcb081730211::LoadClusterMutation::LoadInfo1236 LoadInfo(SUnit *su, unsigned reg, unsigned ofs)
1237 : SU(su), BaseReg(reg), Offset(ofs) {}
1238
operator <__anonbcb081730211::LoadClusterMutation::LoadInfo1239 bool operator<(const LoadInfo &RHS) const {
1240 return std::tie(BaseReg, Offset) < std::tie(RHS.BaseReg, RHS.Offset);
1241 }
1242 };
1243
1244 const TargetInstrInfo *TII;
1245 const TargetRegisterInfo *TRI;
1246 public:
LoadClusterMutation(const TargetInstrInfo * tii,const TargetRegisterInfo * tri)1247 LoadClusterMutation(const TargetInstrInfo *tii,
1248 const TargetRegisterInfo *tri)
1249 : TII(tii), TRI(tri) {}
1250
1251 void apply(ScheduleDAGMI *DAG) override;
1252 protected:
1253 void clusterNeighboringLoads(ArrayRef<SUnit*> Loads, ScheduleDAGMI *DAG);
1254 };
1255 } // anonymous
1256
clusterNeighboringLoads(ArrayRef<SUnit * > Loads,ScheduleDAGMI * DAG)1257 void LoadClusterMutation::clusterNeighboringLoads(ArrayRef<SUnit*> Loads,
1258 ScheduleDAGMI *DAG) {
1259 SmallVector<LoadClusterMutation::LoadInfo,32> LoadRecords;
1260 for (unsigned Idx = 0, End = Loads.size(); Idx != End; ++Idx) {
1261 SUnit *SU = Loads[Idx];
1262 unsigned BaseReg;
1263 unsigned Offset;
1264 if (TII->getLdStBaseRegImmOfs(SU->getInstr(), BaseReg, Offset, TRI))
1265 LoadRecords.push_back(LoadInfo(SU, BaseReg, Offset));
1266 }
1267 if (LoadRecords.size() < 2)
1268 return;
1269 std::sort(LoadRecords.begin(), LoadRecords.end());
1270 unsigned ClusterLength = 1;
1271 for (unsigned Idx = 0, End = LoadRecords.size(); Idx < (End - 1); ++Idx) {
1272 if (LoadRecords[Idx].BaseReg != LoadRecords[Idx+1].BaseReg) {
1273 ClusterLength = 1;
1274 continue;
1275 }
1276
1277 SUnit *SUa = LoadRecords[Idx].SU;
1278 SUnit *SUb = LoadRecords[Idx+1].SU;
1279 if (TII->shouldClusterLoads(SUa->getInstr(), SUb->getInstr(), ClusterLength)
1280 && DAG->addEdge(SUb, SDep(SUa, SDep::Cluster))) {
1281
1282 DEBUG(dbgs() << "Cluster loads SU(" << SUa->NodeNum << ") - SU("
1283 << SUb->NodeNum << ")\n");
1284 // Copy successor edges from SUa to SUb. Interleaving computation
1285 // dependent on SUa can prevent load combining due to register reuse.
1286 // Predecessor edges do not need to be copied from SUb to SUa since nearby
1287 // loads should have effectively the same inputs.
1288 for (SUnit::const_succ_iterator
1289 SI = SUa->Succs.begin(), SE = SUa->Succs.end(); SI != SE; ++SI) {
1290 if (SI->getSUnit() == SUb)
1291 continue;
1292 DEBUG(dbgs() << " Copy Succ SU(" << SI->getSUnit()->NodeNum << ")\n");
1293 DAG->addEdge(SI->getSUnit(), SDep(SUb, SDep::Artificial));
1294 }
1295 ++ClusterLength;
1296 }
1297 else
1298 ClusterLength = 1;
1299 }
1300 }
1301
1302 /// \brief Callback from DAG postProcessing to create cluster edges for loads.
apply(ScheduleDAGMI * DAG)1303 void LoadClusterMutation::apply(ScheduleDAGMI *DAG) {
1304 // Map DAG NodeNum to store chain ID.
1305 DenseMap<unsigned, unsigned> StoreChainIDs;
1306 // Map each store chain to a set of dependent loads.
1307 SmallVector<SmallVector<SUnit*,4>, 32> StoreChainDependents;
1308 for (unsigned Idx = 0, End = DAG->SUnits.size(); Idx != End; ++Idx) {
1309 SUnit *SU = &DAG->SUnits[Idx];
1310 if (!SU->getInstr()->mayLoad())
1311 continue;
1312 unsigned ChainPredID = DAG->SUnits.size();
1313 for (SUnit::const_pred_iterator
1314 PI = SU->Preds.begin(), PE = SU->Preds.end(); PI != PE; ++PI) {
1315 if (PI->isCtrl()) {
1316 ChainPredID = PI->getSUnit()->NodeNum;
1317 break;
1318 }
1319 }
1320 // Check if this chain-like pred has been seen
1321 // before. ChainPredID==MaxNodeID for loads at the top of the schedule.
1322 unsigned NumChains = StoreChainDependents.size();
1323 std::pair<DenseMap<unsigned, unsigned>::iterator, bool> Result =
1324 StoreChainIDs.insert(std::make_pair(ChainPredID, NumChains));
1325 if (Result.second)
1326 StoreChainDependents.resize(NumChains + 1);
1327 StoreChainDependents[Result.first->second].push_back(SU);
1328 }
1329 // Iterate over the store chains.
1330 for (unsigned Idx = 0, End = StoreChainDependents.size(); Idx != End; ++Idx)
1331 clusterNeighboringLoads(StoreChainDependents[Idx], DAG);
1332 }
1333
1334 //===----------------------------------------------------------------------===//
1335 // MacroFusion - DAG post-processing to encourage fusion of macro ops.
1336 //===----------------------------------------------------------------------===//
1337
1338 namespace {
1339 /// \brief Post-process the DAG to create cluster edges between instructions
1340 /// that may be fused by the processor into a single operation.
1341 class MacroFusion : public ScheduleDAGMutation {
1342 const TargetInstrInfo *TII;
1343 public:
MacroFusion(const TargetInstrInfo * tii)1344 MacroFusion(const TargetInstrInfo *tii): TII(tii) {}
1345
1346 void apply(ScheduleDAGMI *DAG) override;
1347 };
1348 } // anonymous
1349
1350 /// \brief Callback from DAG postProcessing to create cluster edges to encourage
1351 /// fused operations.
apply(ScheduleDAGMI * DAG)1352 void MacroFusion::apply(ScheduleDAGMI *DAG) {
1353 // For now, assume targets can only fuse with the branch.
1354 MachineInstr *Branch = DAG->ExitSU.getInstr();
1355 if (!Branch)
1356 return;
1357
1358 for (unsigned Idx = DAG->SUnits.size(); Idx > 0;) {
1359 SUnit *SU = &DAG->SUnits[--Idx];
1360 if (!TII->shouldScheduleAdjacent(SU->getInstr(), Branch))
1361 continue;
1362
1363 // Create a single weak edge from SU to ExitSU. The only effect is to cause
1364 // bottom-up scheduling to heavily prioritize the clustered SU. There is no
1365 // need to copy predecessor edges from ExitSU to SU, since top-down
1366 // scheduling cannot prioritize ExitSU anyway. To defer top-down scheduling
1367 // of SU, we could create an artificial edge from the deepest root, but it
1368 // hasn't been needed yet.
1369 bool Success = DAG->addEdge(&DAG->ExitSU, SDep(SU, SDep::Cluster));
1370 (void)Success;
1371 assert(Success && "No DAG nodes should be reachable from ExitSU");
1372
1373 DEBUG(dbgs() << "Macro Fuse SU(" << SU->NodeNum << ")\n");
1374 break;
1375 }
1376 }
1377
1378 //===----------------------------------------------------------------------===//
1379 // CopyConstrain - DAG post-processing to encourage copy elimination.
1380 //===----------------------------------------------------------------------===//
1381
1382 namespace {
1383 /// \brief Post-process the DAG to create weak edges from all uses of a copy to
1384 /// the one use that defines the copy's source vreg, most likely an induction
1385 /// variable increment.
1386 class CopyConstrain : public ScheduleDAGMutation {
1387 // Transient state.
1388 SlotIndex RegionBeginIdx;
1389 // RegionEndIdx is the slot index of the last non-debug instruction in the
1390 // scheduling region. So we may have RegionBeginIdx == RegionEndIdx.
1391 SlotIndex RegionEndIdx;
1392 public:
CopyConstrain(const TargetInstrInfo *,const TargetRegisterInfo *)1393 CopyConstrain(const TargetInstrInfo *, const TargetRegisterInfo *) {}
1394
1395 void apply(ScheduleDAGMI *DAG) override;
1396
1397 protected:
1398 void constrainLocalCopy(SUnit *CopySU, ScheduleDAGMILive *DAG);
1399 };
1400 } // anonymous
1401
1402 /// constrainLocalCopy handles two possibilities:
1403 /// 1) Local src:
1404 /// I0: = dst
1405 /// I1: src = ...
1406 /// I2: = dst
1407 /// I3: dst = src (copy)
1408 /// (create pred->succ edges I0->I1, I2->I1)
1409 ///
1410 /// 2) Local copy:
1411 /// I0: dst = src (copy)
1412 /// I1: = dst
1413 /// I2: src = ...
1414 /// I3: = dst
1415 /// (create pred->succ edges I1->I2, I3->I2)
1416 ///
1417 /// Although the MachineScheduler is currently constrained to single blocks,
1418 /// this algorithm should handle extended blocks. An EBB is a set of
1419 /// contiguously numbered blocks such that the previous block in the EBB is
1420 /// always the single predecessor.
constrainLocalCopy(SUnit * CopySU,ScheduleDAGMILive * DAG)1421 void CopyConstrain::constrainLocalCopy(SUnit *CopySU, ScheduleDAGMILive *DAG) {
1422 LiveIntervals *LIS = DAG->getLIS();
1423 MachineInstr *Copy = CopySU->getInstr();
1424
1425 // Check for pure vreg copies.
1426 unsigned SrcReg = Copy->getOperand(1).getReg();
1427 if (!TargetRegisterInfo::isVirtualRegister(SrcReg))
1428 return;
1429
1430 unsigned DstReg = Copy->getOperand(0).getReg();
1431 if (!TargetRegisterInfo::isVirtualRegister(DstReg))
1432 return;
1433
1434 // Check if either the dest or source is local. If it's live across a back
1435 // edge, it's not local. Note that if both vregs are live across the back
1436 // edge, we cannot successfully contrain the copy without cyclic scheduling.
1437 unsigned LocalReg = DstReg;
1438 unsigned GlobalReg = SrcReg;
1439 LiveInterval *LocalLI = &LIS->getInterval(LocalReg);
1440 if (!LocalLI->isLocal(RegionBeginIdx, RegionEndIdx)) {
1441 LocalReg = SrcReg;
1442 GlobalReg = DstReg;
1443 LocalLI = &LIS->getInterval(LocalReg);
1444 if (!LocalLI->isLocal(RegionBeginIdx, RegionEndIdx))
1445 return;
1446 }
1447 LiveInterval *GlobalLI = &LIS->getInterval(GlobalReg);
1448
1449 // Find the global segment after the start of the local LI.
1450 LiveInterval::iterator GlobalSegment = GlobalLI->find(LocalLI->beginIndex());
1451 // If GlobalLI does not overlap LocalLI->start, then a copy directly feeds a
1452 // local live range. We could create edges from other global uses to the local
1453 // start, but the coalescer should have already eliminated these cases, so
1454 // don't bother dealing with it.
1455 if (GlobalSegment == GlobalLI->end())
1456 return;
1457
1458 // If GlobalSegment is killed at the LocalLI->start, the call to find()
1459 // returned the next global segment. But if GlobalSegment overlaps with
1460 // LocalLI->start, then advance to the next segement. If a hole in GlobalLI
1461 // exists in LocalLI's vicinity, GlobalSegment will be the end of the hole.
1462 if (GlobalSegment->contains(LocalLI->beginIndex()))
1463 ++GlobalSegment;
1464
1465 if (GlobalSegment == GlobalLI->end())
1466 return;
1467
1468 // Check if GlobalLI contains a hole in the vicinity of LocalLI.
1469 if (GlobalSegment != GlobalLI->begin()) {
1470 // Two address defs have no hole.
1471 if (SlotIndex::isSameInstr(std::prev(GlobalSegment)->end,
1472 GlobalSegment->start)) {
1473 return;
1474 }
1475 // If the prior global segment may be defined by the same two-address
1476 // instruction that also defines LocalLI, then can't make a hole here.
1477 if (SlotIndex::isSameInstr(std::prev(GlobalSegment)->start,
1478 LocalLI->beginIndex())) {
1479 return;
1480 }
1481 // If GlobalLI has a prior segment, it must be live into the EBB. Otherwise
1482 // it would be a disconnected component in the live range.
1483 assert(std::prev(GlobalSegment)->start < LocalLI->beginIndex() &&
1484 "Disconnected LRG within the scheduling region.");
1485 }
1486 MachineInstr *GlobalDef = LIS->getInstructionFromIndex(GlobalSegment->start);
1487 if (!GlobalDef)
1488 return;
1489
1490 SUnit *GlobalSU = DAG->getSUnit(GlobalDef);
1491 if (!GlobalSU)
1492 return;
1493
1494 // GlobalDef is the bottom of the GlobalLI hole. Open the hole by
1495 // constraining the uses of the last local def to precede GlobalDef.
1496 SmallVector<SUnit*,8> LocalUses;
1497 const VNInfo *LastLocalVN = LocalLI->getVNInfoBefore(LocalLI->endIndex());
1498 MachineInstr *LastLocalDef = LIS->getInstructionFromIndex(LastLocalVN->def);
1499 SUnit *LastLocalSU = DAG->getSUnit(LastLocalDef);
1500 for (SUnit::const_succ_iterator
1501 I = LastLocalSU->Succs.begin(), E = LastLocalSU->Succs.end();
1502 I != E; ++I) {
1503 if (I->getKind() != SDep::Data || I->getReg() != LocalReg)
1504 continue;
1505 if (I->getSUnit() == GlobalSU)
1506 continue;
1507 if (!DAG->canAddEdge(GlobalSU, I->getSUnit()))
1508 return;
1509 LocalUses.push_back(I->getSUnit());
1510 }
1511 // Open the top of the GlobalLI hole by constraining any earlier global uses
1512 // to precede the start of LocalLI.
1513 SmallVector<SUnit*,8> GlobalUses;
1514 MachineInstr *FirstLocalDef =
1515 LIS->getInstructionFromIndex(LocalLI->beginIndex());
1516 SUnit *FirstLocalSU = DAG->getSUnit(FirstLocalDef);
1517 for (SUnit::const_pred_iterator
1518 I = GlobalSU->Preds.begin(), E = GlobalSU->Preds.end(); I != E; ++I) {
1519 if (I->getKind() != SDep::Anti || I->getReg() != GlobalReg)
1520 continue;
1521 if (I->getSUnit() == FirstLocalSU)
1522 continue;
1523 if (!DAG->canAddEdge(FirstLocalSU, I->getSUnit()))
1524 return;
1525 GlobalUses.push_back(I->getSUnit());
1526 }
1527 DEBUG(dbgs() << "Constraining copy SU(" << CopySU->NodeNum << ")\n");
1528 // Add the weak edges.
1529 for (SmallVectorImpl<SUnit*>::const_iterator
1530 I = LocalUses.begin(), E = LocalUses.end(); I != E; ++I) {
1531 DEBUG(dbgs() << " Local use SU(" << (*I)->NodeNum << ") -> SU("
1532 << GlobalSU->NodeNum << ")\n");
1533 DAG->addEdge(GlobalSU, SDep(*I, SDep::Weak));
1534 }
1535 for (SmallVectorImpl<SUnit*>::const_iterator
1536 I = GlobalUses.begin(), E = GlobalUses.end(); I != E; ++I) {
1537 DEBUG(dbgs() << " Global use SU(" << (*I)->NodeNum << ") -> SU("
1538 << FirstLocalSU->NodeNum << ")\n");
1539 DAG->addEdge(FirstLocalSU, SDep(*I, SDep::Weak));
1540 }
1541 }
1542
1543 /// \brief Callback from DAG postProcessing to create weak edges to encourage
1544 /// copy elimination.
apply(ScheduleDAGMI * DAG)1545 void CopyConstrain::apply(ScheduleDAGMI *DAG) {
1546 assert(DAG->hasVRegLiveness() && "Expect VRegs with LiveIntervals");
1547
1548 MachineBasicBlock::iterator FirstPos = nextIfDebug(DAG->begin(), DAG->end());
1549 if (FirstPos == DAG->end())
1550 return;
1551 RegionBeginIdx = DAG->getLIS()->getInstructionIndex(&*FirstPos);
1552 RegionEndIdx = DAG->getLIS()->getInstructionIndex(
1553 &*priorNonDebug(DAG->end(), DAG->begin()));
1554
1555 for (unsigned Idx = 0, End = DAG->SUnits.size(); Idx != End; ++Idx) {
1556 SUnit *SU = &DAG->SUnits[Idx];
1557 if (!SU->getInstr()->isCopy())
1558 continue;
1559
1560 constrainLocalCopy(SU, static_cast<ScheduleDAGMILive*>(DAG));
1561 }
1562 }
1563
1564 //===----------------------------------------------------------------------===//
1565 // MachineSchedStrategy helpers used by GenericScheduler, GenericPostScheduler
1566 // and possibly other custom schedulers.
1567 //===----------------------------------------------------------------------===//
1568
1569 static const unsigned InvalidCycle = ~0U;
1570
~SchedBoundary()1571 SchedBoundary::~SchedBoundary() { delete HazardRec; }
1572
reset()1573 void SchedBoundary::reset() {
1574 // A new HazardRec is created for each DAG and owned by SchedBoundary.
1575 // Destroying and reconstructing it is very expensive though. So keep
1576 // invalid, placeholder HazardRecs.
1577 if (HazardRec && HazardRec->isEnabled()) {
1578 delete HazardRec;
1579 HazardRec = nullptr;
1580 }
1581 Available.clear();
1582 Pending.clear();
1583 CheckPending = false;
1584 NextSUs.clear();
1585 CurrCycle = 0;
1586 CurrMOps = 0;
1587 MinReadyCycle = UINT_MAX;
1588 ExpectedLatency = 0;
1589 DependentLatency = 0;
1590 RetiredMOps = 0;
1591 MaxExecutedResCount = 0;
1592 ZoneCritResIdx = 0;
1593 IsResourceLimited = false;
1594 ReservedCycles.clear();
1595 #ifndef NDEBUG
1596 // Track the maximum number of stall cycles that could arise either from the
1597 // latency of a DAG edge or the number of cycles that a processor resource is
1598 // reserved (SchedBoundary::ReservedCycles).
1599 MaxObservedStall = 0;
1600 #endif
1601 // Reserve a zero-count for invalid CritResIdx.
1602 ExecutedResCounts.resize(1);
1603 assert(!ExecutedResCounts[0] && "nonzero count for bad resource");
1604 }
1605
1606 void SchedRemainder::
init(ScheduleDAGMI * DAG,const TargetSchedModel * SchedModel)1607 init(ScheduleDAGMI *DAG, const TargetSchedModel *SchedModel) {
1608 reset();
1609 if (!SchedModel->hasInstrSchedModel())
1610 return;
1611 RemainingCounts.resize(SchedModel->getNumProcResourceKinds());
1612 for (std::vector<SUnit>::iterator
1613 I = DAG->SUnits.begin(), E = DAG->SUnits.end(); I != E; ++I) {
1614 const MCSchedClassDesc *SC = DAG->getSchedClass(&*I);
1615 RemIssueCount += SchedModel->getNumMicroOps(I->getInstr(), SC)
1616 * SchedModel->getMicroOpFactor();
1617 for (TargetSchedModel::ProcResIter
1618 PI = SchedModel->getWriteProcResBegin(SC),
1619 PE = SchedModel->getWriteProcResEnd(SC); PI != PE; ++PI) {
1620 unsigned PIdx = PI->ProcResourceIdx;
1621 unsigned Factor = SchedModel->getResourceFactor(PIdx);
1622 RemainingCounts[PIdx] += (Factor * PI->Cycles);
1623 }
1624 }
1625 }
1626
1627 void SchedBoundary::
init(ScheduleDAGMI * dag,const TargetSchedModel * smodel,SchedRemainder * rem)1628 init(ScheduleDAGMI *dag, const TargetSchedModel *smodel, SchedRemainder *rem) {
1629 reset();
1630 DAG = dag;
1631 SchedModel = smodel;
1632 Rem = rem;
1633 if (SchedModel->hasInstrSchedModel()) {
1634 ExecutedResCounts.resize(SchedModel->getNumProcResourceKinds());
1635 ReservedCycles.resize(SchedModel->getNumProcResourceKinds(), InvalidCycle);
1636 }
1637 }
1638
1639 /// Compute the stall cycles based on this SUnit's ready time. Heuristics treat
1640 /// these "soft stalls" differently than the hard stall cycles based on CPU
1641 /// resources and computed by checkHazard(). A fully in-order model
1642 /// (MicroOpBufferSize==0) will not make use of this since instructions are not
1643 /// available for scheduling until they are ready. However, a weaker in-order
1644 /// model may use this for heuristics. For example, if a processor has in-order
1645 /// behavior when reading certain resources, this may come into play.
getLatencyStallCycles(SUnit * SU)1646 unsigned SchedBoundary::getLatencyStallCycles(SUnit *SU) {
1647 if (!SU->isUnbuffered)
1648 return 0;
1649
1650 unsigned ReadyCycle = (isTop() ? SU->TopReadyCycle : SU->BotReadyCycle);
1651 if (ReadyCycle > CurrCycle)
1652 return ReadyCycle - CurrCycle;
1653 return 0;
1654 }
1655
1656 /// Compute the next cycle at which the given processor resource can be
1657 /// scheduled.
1658 unsigned SchedBoundary::
getNextResourceCycle(unsigned PIdx,unsigned Cycles)1659 getNextResourceCycle(unsigned PIdx, unsigned Cycles) {
1660 unsigned NextUnreserved = ReservedCycles[PIdx];
1661 // If this resource has never been used, always return cycle zero.
1662 if (NextUnreserved == InvalidCycle)
1663 return 0;
1664 // For bottom-up scheduling add the cycles needed for the current operation.
1665 if (!isTop())
1666 NextUnreserved += Cycles;
1667 return NextUnreserved;
1668 }
1669
1670 /// Does this SU have a hazard within the current instruction group.
1671 ///
1672 /// The scheduler supports two modes of hazard recognition. The first is the
1673 /// ScheduleHazardRecognizer API. It is a fully general hazard recognizer that
1674 /// supports highly complicated in-order reservation tables
1675 /// (ScoreboardHazardRecognizer) and arbitraty target-specific logic.
1676 ///
1677 /// The second is a streamlined mechanism that checks for hazards based on
1678 /// simple counters that the scheduler itself maintains. It explicitly checks
1679 /// for instruction dispatch limitations, including the number of micro-ops that
1680 /// can dispatch per cycle.
1681 ///
1682 /// TODO: Also check whether the SU must start a new group.
checkHazard(SUnit * SU)1683 bool SchedBoundary::checkHazard(SUnit *SU) {
1684 if (HazardRec->isEnabled()
1685 && HazardRec->getHazardType(SU) != ScheduleHazardRecognizer::NoHazard) {
1686 return true;
1687 }
1688 unsigned uops = SchedModel->getNumMicroOps(SU->getInstr());
1689 if ((CurrMOps > 0) && (CurrMOps + uops > SchedModel->getIssueWidth())) {
1690 DEBUG(dbgs() << " SU(" << SU->NodeNum << ") uops="
1691 << SchedModel->getNumMicroOps(SU->getInstr()) << '\n');
1692 return true;
1693 }
1694 if (SchedModel->hasInstrSchedModel() && SU->hasReservedResource) {
1695 const MCSchedClassDesc *SC = DAG->getSchedClass(SU);
1696 for (TargetSchedModel::ProcResIter
1697 PI = SchedModel->getWriteProcResBegin(SC),
1698 PE = SchedModel->getWriteProcResEnd(SC); PI != PE; ++PI) {
1699 unsigned NRCycle = getNextResourceCycle(PI->ProcResourceIdx, PI->Cycles);
1700 if (NRCycle > CurrCycle) {
1701 #ifndef NDEBUG
1702 MaxObservedStall = std::max(PI->Cycles, MaxObservedStall);
1703 #endif
1704 DEBUG(dbgs() << " SU(" << SU->NodeNum << ") "
1705 << SchedModel->getResourceName(PI->ProcResourceIdx)
1706 << "=" << NRCycle << "c\n");
1707 return true;
1708 }
1709 }
1710 }
1711 return false;
1712 }
1713
1714 // Find the unscheduled node in ReadySUs with the highest latency.
1715 unsigned SchedBoundary::
findMaxLatency(ArrayRef<SUnit * > ReadySUs)1716 findMaxLatency(ArrayRef<SUnit*> ReadySUs) {
1717 SUnit *LateSU = nullptr;
1718 unsigned RemLatency = 0;
1719 for (ArrayRef<SUnit*>::iterator I = ReadySUs.begin(), E = ReadySUs.end();
1720 I != E; ++I) {
1721 unsigned L = getUnscheduledLatency(*I);
1722 if (L > RemLatency) {
1723 RemLatency = L;
1724 LateSU = *I;
1725 }
1726 }
1727 if (LateSU) {
1728 DEBUG(dbgs() << Available.getName() << " RemLatency SU("
1729 << LateSU->NodeNum << ") " << RemLatency << "c\n");
1730 }
1731 return RemLatency;
1732 }
1733
1734 // Count resources in this zone and the remaining unscheduled
1735 // instruction. Return the max count, scaled. Set OtherCritIdx to the critical
1736 // resource index, or zero if the zone is issue limited.
1737 unsigned SchedBoundary::
getOtherResourceCount(unsigned & OtherCritIdx)1738 getOtherResourceCount(unsigned &OtherCritIdx) {
1739 OtherCritIdx = 0;
1740 if (!SchedModel->hasInstrSchedModel())
1741 return 0;
1742
1743 unsigned OtherCritCount = Rem->RemIssueCount
1744 + (RetiredMOps * SchedModel->getMicroOpFactor());
1745 DEBUG(dbgs() << " " << Available.getName() << " + Remain MOps: "
1746 << OtherCritCount / SchedModel->getMicroOpFactor() << '\n');
1747 for (unsigned PIdx = 1, PEnd = SchedModel->getNumProcResourceKinds();
1748 PIdx != PEnd; ++PIdx) {
1749 unsigned OtherCount = getResourceCount(PIdx) + Rem->RemainingCounts[PIdx];
1750 if (OtherCount > OtherCritCount) {
1751 OtherCritCount = OtherCount;
1752 OtherCritIdx = PIdx;
1753 }
1754 }
1755 if (OtherCritIdx) {
1756 DEBUG(dbgs() << " " << Available.getName() << " + Remain CritRes: "
1757 << OtherCritCount / SchedModel->getResourceFactor(OtherCritIdx)
1758 << " " << SchedModel->getResourceName(OtherCritIdx) << "\n");
1759 }
1760 return OtherCritCount;
1761 }
1762
releaseNode(SUnit * SU,unsigned ReadyCycle)1763 void SchedBoundary::releaseNode(SUnit *SU, unsigned ReadyCycle) {
1764 assert(SU->getInstr() && "Scheduled SUnit must have instr");
1765
1766 #ifndef NDEBUG
1767 // ReadyCycle was been bumped up to the CurrCycle when this node was
1768 // scheduled, but CurrCycle may have been eagerly advanced immediately after
1769 // scheduling, so may now be greater than ReadyCycle.
1770 if (ReadyCycle > CurrCycle)
1771 MaxObservedStall = std::max(ReadyCycle - CurrCycle, MaxObservedStall);
1772 #endif
1773
1774 if (ReadyCycle < MinReadyCycle)
1775 MinReadyCycle = ReadyCycle;
1776
1777 // Check for interlocks first. For the purpose of other heuristics, an
1778 // instruction that cannot issue appears as if it's not in the ReadyQueue.
1779 bool IsBuffered = SchedModel->getMicroOpBufferSize() != 0;
1780 if ((!IsBuffered && ReadyCycle > CurrCycle) || checkHazard(SU))
1781 Pending.push(SU);
1782 else
1783 Available.push(SU);
1784
1785 // Record this node as an immediate dependent of the scheduled node.
1786 NextSUs.insert(SU);
1787 }
1788
releaseTopNode(SUnit * SU)1789 void SchedBoundary::releaseTopNode(SUnit *SU) {
1790 if (SU->isScheduled)
1791 return;
1792
1793 releaseNode(SU, SU->TopReadyCycle);
1794 }
1795
releaseBottomNode(SUnit * SU)1796 void SchedBoundary::releaseBottomNode(SUnit *SU) {
1797 if (SU->isScheduled)
1798 return;
1799
1800 releaseNode(SU, SU->BotReadyCycle);
1801 }
1802
1803 /// Move the boundary of scheduled code by one cycle.
bumpCycle(unsigned NextCycle)1804 void SchedBoundary::bumpCycle(unsigned NextCycle) {
1805 if (SchedModel->getMicroOpBufferSize() == 0) {
1806 assert(MinReadyCycle < UINT_MAX && "MinReadyCycle uninitialized");
1807 if (MinReadyCycle > NextCycle)
1808 NextCycle = MinReadyCycle;
1809 }
1810 // Update the current micro-ops, which will issue in the next cycle.
1811 unsigned DecMOps = SchedModel->getIssueWidth() * (NextCycle - CurrCycle);
1812 CurrMOps = (CurrMOps <= DecMOps) ? 0 : CurrMOps - DecMOps;
1813
1814 // Decrement DependentLatency based on the next cycle.
1815 if ((NextCycle - CurrCycle) > DependentLatency)
1816 DependentLatency = 0;
1817 else
1818 DependentLatency -= (NextCycle - CurrCycle);
1819
1820 if (!HazardRec->isEnabled()) {
1821 // Bypass HazardRec virtual calls.
1822 CurrCycle = NextCycle;
1823 }
1824 else {
1825 // Bypass getHazardType calls in case of long latency.
1826 for (; CurrCycle != NextCycle; ++CurrCycle) {
1827 if (isTop())
1828 HazardRec->AdvanceCycle();
1829 else
1830 HazardRec->RecedeCycle();
1831 }
1832 }
1833 CheckPending = true;
1834 unsigned LFactor = SchedModel->getLatencyFactor();
1835 IsResourceLimited =
1836 (int)(getCriticalCount() - (getScheduledLatency() * LFactor))
1837 > (int)LFactor;
1838
1839 DEBUG(dbgs() << "Cycle: " << CurrCycle << ' ' << Available.getName() << '\n');
1840 }
1841
incExecutedResources(unsigned PIdx,unsigned Count)1842 void SchedBoundary::incExecutedResources(unsigned PIdx, unsigned Count) {
1843 ExecutedResCounts[PIdx] += Count;
1844 if (ExecutedResCounts[PIdx] > MaxExecutedResCount)
1845 MaxExecutedResCount = ExecutedResCounts[PIdx];
1846 }
1847
1848 /// Add the given processor resource to this scheduled zone.
1849 ///
1850 /// \param Cycles indicates the number of consecutive (non-pipelined) cycles
1851 /// during which this resource is consumed.
1852 ///
1853 /// \return the next cycle at which the instruction may execute without
1854 /// oversubscribing resources.
1855 unsigned SchedBoundary::
countResource(unsigned PIdx,unsigned Cycles,unsigned NextCycle)1856 countResource(unsigned PIdx, unsigned Cycles, unsigned NextCycle) {
1857 unsigned Factor = SchedModel->getResourceFactor(PIdx);
1858 unsigned Count = Factor * Cycles;
1859 DEBUG(dbgs() << " " << SchedModel->getResourceName(PIdx)
1860 << " +" << Cycles << "x" << Factor << "u\n");
1861
1862 // Update Executed resources counts.
1863 incExecutedResources(PIdx, Count);
1864 assert(Rem->RemainingCounts[PIdx] >= Count && "resource double counted");
1865 Rem->RemainingCounts[PIdx] -= Count;
1866
1867 // Check if this resource exceeds the current critical resource. If so, it
1868 // becomes the critical resource.
1869 if (ZoneCritResIdx != PIdx && (getResourceCount(PIdx) > getCriticalCount())) {
1870 ZoneCritResIdx = PIdx;
1871 DEBUG(dbgs() << " *** Critical resource "
1872 << SchedModel->getResourceName(PIdx) << ": "
1873 << getResourceCount(PIdx) / SchedModel->getLatencyFactor() << "c\n");
1874 }
1875 // For reserved resources, record the highest cycle using the resource.
1876 unsigned NextAvailable = getNextResourceCycle(PIdx, Cycles);
1877 if (NextAvailable > CurrCycle) {
1878 DEBUG(dbgs() << " Resource conflict: "
1879 << SchedModel->getProcResource(PIdx)->Name << " reserved until @"
1880 << NextAvailable << "\n");
1881 }
1882 return NextAvailable;
1883 }
1884
1885 /// Move the boundary of scheduled code by one SUnit.
bumpNode(SUnit * SU)1886 void SchedBoundary::bumpNode(SUnit *SU) {
1887 // Update the reservation table.
1888 if (HazardRec->isEnabled()) {
1889 if (!isTop() && SU->isCall) {
1890 // Calls are scheduled with their preceding instructions. For bottom-up
1891 // scheduling, clear the pipeline state before emitting.
1892 HazardRec->Reset();
1893 }
1894 HazardRec->EmitInstruction(SU);
1895 }
1896 // checkHazard should prevent scheduling multiple instructions per cycle that
1897 // exceed the issue width.
1898 const MCSchedClassDesc *SC = DAG->getSchedClass(SU);
1899 unsigned IncMOps = SchedModel->getNumMicroOps(SU->getInstr());
1900 assert(
1901 (CurrMOps == 0 || (CurrMOps + IncMOps) <= SchedModel->getIssueWidth()) &&
1902 "Cannot schedule this instruction's MicroOps in the current cycle.");
1903
1904 unsigned ReadyCycle = (isTop() ? SU->TopReadyCycle : SU->BotReadyCycle);
1905 DEBUG(dbgs() << " Ready @" << ReadyCycle << "c\n");
1906
1907 unsigned NextCycle = CurrCycle;
1908 switch (SchedModel->getMicroOpBufferSize()) {
1909 case 0:
1910 assert(ReadyCycle <= CurrCycle && "Broken PendingQueue");
1911 break;
1912 case 1:
1913 if (ReadyCycle > NextCycle) {
1914 NextCycle = ReadyCycle;
1915 DEBUG(dbgs() << " *** Stall until: " << ReadyCycle << "\n");
1916 }
1917 break;
1918 default:
1919 // We don't currently model the OOO reorder buffer, so consider all
1920 // scheduled MOps to be "retired". We do loosely model in-order resource
1921 // latency. If this instruction uses an in-order resource, account for any
1922 // likely stall cycles.
1923 if (SU->isUnbuffered && ReadyCycle > NextCycle)
1924 NextCycle = ReadyCycle;
1925 break;
1926 }
1927 RetiredMOps += IncMOps;
1928
1929 // Update resource counts and critical resource.
1930 if (SchedModel->hasInstrSchedModel()) {
1931 unsigned DecRemIssue = IncMOps * SchedModel->getMicroOpFactor();
1932 assert(Rem->RemIssueCount >= DecRemIssue && "MOps double counted");
1933 Rem->RemIssueCount -= DecRemIssue;
1934 if (ZoneCritResIdx) {
1935 // Scale scheduled micro-ops for comparing with the critical resource.
1936 unsigned ScaledMOps =
1937 RetiredMOps * SchedModel->getMicroOpFactor();
1938
1939 // If scaled micro-ops are now more than the previous critical resource by
1940 // a full cycle, then micro-ops issue becomes critical.
1941 if ((int)(ScaledMOps - getResourceCount(ZoneCritResIdx))
1942 >= (int)SchedModel->getLatencyFactor()) {
1943 ZoneCritResIdx = 0;
1944 DEBUG(dbgs() << " *** Critical resource NumMicroOps: "
1945 << ScaledMOps / SchedModel->getLatencyFactor() << "c\n");
1946 }
1947 }
1948 for (TargetSchedModel::ProcResIter
1949 PI = SchedModel->getWriteProcResBegin(SC),
1950 PE = SchedModel->getWriteProcResEnd(SC); PI != PE; ++PI) {
1951 unsigned RCycle =
1952 countResource(PI->ProcResourceIdx, PI->Cycles, NextCycle);
1953 if (RCycle > NextCycle)
1954 NextCycle = RCycle;
1955 }
1956 if (SU->hasReservedResource) {
1957 // For reserved resources, record the highest cycle using the resource.
1958 // For top-down scheduling, this is the cycle in which we schedule this
1959 // instruction plus the number of cycles the operations reserves the
1960 // resource. For bottom-up is it simply the instruction's cycle.
1961 for (TargetSchedModel::ProcResIter
1962 PI = SchedModel->getWriteProcResBegin(SC),
1963 PE = SchedModel->getWriteProcResEnd(SC); PI != PE; ++PI) {
1964 unsigned PIdx = PI->ProcResourceIdx;
1965 if (SchedModel->getProcResource(PIdx)->BufferSize == 0) {
1966 if (isTop()) {
1967 ReservedCycles[PIdx] =
1968 std::max(getNextResourceCycle(PIdx, 0), NextCycle + PI->Cycles);
1969 }
1970 else
1971 ReservedCycles[PIdx] = NextCycle;
1972 }
1973 }
1974 }
1975 }
1976 // Update ExpectedLatency and DependentLatency.
1977 unsigned &TopLatency = isTop() ? ExpectedLatency : DependentLatency;
1978 unsigned &BotLatency = isTop() ? DependentLatency : ExpectedLatency;
1979 if (SU->getDepth() > TopLatency) {
1980 TopLatency = SU->getDepth();
1981 DEBUG(dbgs() << " " << Available.getName()
1982 << " TopLatency SU(" << SU->NodeNum << ") " << TopLatency << "c\n");
1983 }
1984 if (SU->getHeight() > BotLatency) {
1985 BotLatency = SU->getHeight();
1986 DEBUG(dbgs() << " " << Available.getName()
1987 << " BotLatency SU(" << SU->NodeNum << ") " << BotLatency << "c\n");
1988 }
1989 // If we stall for any reason, bump the cycle.
1990 if (NextCycle > CurrCycle) {
1991 bumpCycle(NextCycle);
1992 }
1993 else {
1994 // After updating ZoneCritResIdx and ExpectedLatency, check if we're
1995 // resource limited. If a stall occurred, bumpCycle does this.
1996 unsigned LFactor = SchedModel->getLatencyFactor();
1997 IsResourceLimited =
1998 (int)(getCriticalCount() - (getScheduledLatency() * LFactor))
1999 > (int)LFactor;
2000 }
2001 // Update CurrMOps after calling bumpCycle to handle stalls, since bumpCycle
2002 // resets CurrMOps. Loop to handle instructions with more MOps than issue in
2003 // one cycle. Since we commonly reach the max MOps here, opportunistically
2004 // bump the cycle to avoid uselessly checking everything in the readyQ.
2005 CurrMOps += IncMOps;
2006 while (CurrMOps >= SchedModel->getIssueWidth()) {
2007 DEBUG(dbgs() << " *** Max MOps " << CurrMOps
2008 << " at cycle " << CurrCycle << '\n');
2009 bumpCycle(++NextCycle);
2010 }
2011 DEBUG(dumpScheduledState());
2012 }
2013
2014 /// Release pending ready nodes in to the available queue. This makes them
2015 /// visible to heuristics.
releasePending()2016 void SchedBoundary::releasePending() {
2017 // If the available queue is empty, it is safe to reset MinReadyCycle.
2018 if (Available.empty())
2019 MinReadyCycle = UINT_MAX;
2020
2021 // Check to see if any of the pending instructions are ready to issue. If
2022 // so, add them to the available queue.
2023 bool IsBuffered = SchedModel->getMicroOpBufferSize() != 0;
2024 for (unsigned i = 0, e = Pending.size(); i != e; ++i) {
2025 SUnit *SU = *(Pending.begin()+i);
2026 unsigned ReadyCycle = isTop() ? SU->TopReadyCycle : SU->BotReadyCycle;
2027
2028 if (ReadyCycle < MinReadyCycle)
2029 MinReadyCycle = ReadyCycle;
2030
2031 if (!IsBuffered && ReadyCycle > CurrCycle)
2032 continue;
2033
2034 if (checkHazard(SU))
2035 continue;
2036
2037 Available.push(SU);
2038 Pending.remove(Pending.begin()+i);
2039 --i; --e;
2040 }
2041 DEBUG(if (!Pending.empty()) Pending.dump());
2042 CheckPending = false;
2043 }
2044
2045 /// Remove SU from the ready set for this boundary.
removeReady(SUnit * SU)2046 void SchedBoundary::removeReady(SUnit *SU) {
2047 if (Available.isInQueue(SU))
2048 Available.remove(Available.find(SU));
2049 else {
2050 assert(Pending.isInQueue(SU) && "bad ready count");
2051 Pending.remove(Pending.find(SU));
2052 }
2053 }
2054
2055 /// If this queue only has one ready candidate, return it. As a side effect,
2056 /// defer any nodes that now hit a hazard, and advance the cycle until at least
2057 /// one node is ready. If multiple instructions are ready, return NULL.
pickOnlyChoice()2058 SUnit *SchedBoundary::pickOnlyChoice() {
2059 if (CheckPending)
2060 releasePending();
2061
2062 if (CurrMOps > 0) {
2063 // Defer any ready instrs that now have a hazard.
2064 for (ReadyQueue::iterator I = Available.begin(); I != Available.end();) {
2065 if (checkHazard(*I)) {
2066 Pending.push(*I);
2067 I = Available.remove(I);
2068 continue;
2069 }
2070 ++I;
2071 }
2072 }
2073 for (unsigned i = 0; Available.empty(); ++i) {
2074 // FIXME: Re-enable assert once PR20057 is resolved.
2075 // assert(i <= (HazardRec->getMaxLookAhead() + MaxObservedStall) &&
2076 // "permanent hazard");
2077 (void)i;
2078 bumpCycle(CurrCycle + 1);
2079 releasePending();
2080 }
2081 if (Available.size() == 1)
2082 return *Available.begin();
2083 return nullptr;
2084 }
2085
2086 #ifndef NDEBUG
2087 // This is useful information to dump after bumpNode.
2088 // Note that the Queue contents are more useful before pickNodeFromQueue.
dumpScheduledState()2089 void SchedBoundary::dumpScheduledState() {
2090 unsigned ResFactor;
2091 unsigned ResCount;
2092 if (ZoneCritResIdx) {
2093 ResFactor = SchedModel->getResourceFactor(ZoneCritResIdx);
2094 ResCount = getResourceCount(ZoneCritResIdx);
2095 }
2096 else {
2097 ResFactor = SchedModel->getMicroOpFactor();
2098 ResCount = RetiredMOps * SchedModel->getMicroOpFactor();
2099 }
2100 unsigned LFactor = SchedModel->getLatencyFactor();
2101 dbgs() << Available.getName() << " @" << CurrCycle << "c\n"
2102 << " Retired: " << RetiredMOps;
2103 dbgs() << "\n Executed: " << getExecutedCount() / LFactor << "c";
2104 dbgs() << "\n Critical: " << ResCount / LFactor << "c, "
2105 << ResCount / ResFactor << " "
2106 << SchedModel->getResourceName(ZoneCritResIdx)
2107 << "\n ExpectedLatency: " << ExpectedLatency << "c\n"
2108 << (IsResourceLimited ? " - Resource" : " - Latency")
2109 << " limited.\n";
2110 }
2111 #endif
2112
2113 //===----------------------------------------------------------------------===//
2114 // GenericScheduler - Generic implementation of MachineSchedStrategy.
2115 //===----------------------------------------------------------------------===//
2116
2117 void GenericSchedulerBase::SchedCandidate::
initResourceDelta(const ScheduleDAGMI * DAG,const TargetSchedModel * SchedModel)2118 initResourceDelta(const ScheduleDAGMI *DAG,
2119 const TargetSchedModel *SchedModel) {
2120 if (!Policy.ReduceResIdx && !Policy.DemandResIdx)
2121 return;
2122
2123 const MCSchedClassDesc *SC = DAG->getSchedClass(SU);
2124 for (TargetSchedModel::ProcResIter
2125 PI = SchedModel->getWriteProcResBegin(SC),
2126 PE = SchedModel->getWriteProcResEnd(SC); PI != PE; ++PI) {
2127 if (PI->ProcResourceIdx == Policy.ReduceResIdx)
2128 ResDelta.CritResources += PI->Cycles;
2129 if (PI->ProcResourceIdx == Policy.DemandResIdx)
2130 ResDelta.DemandedResources += PI->Cycles;
2131 }
2132 }
2133
2134 /// Set the CandPolicy given a scheduling zone given the current resources and
2135 /// latencies inside and outside the zone.
setPolicy(CandPolicy & Policy,bool IsPostRA,SchedBoundary & CurrZone,SchedBoundary * OtherZone)2136 void GenericSchedulerBase::setPolicy(CandPolicy &Policy,
2137 bool IsPostRA,
2138 SchedBoundary &CurrZone,
2139 SchedBoundary *OtherZone) {
2140 // Apply preemptive heuristics based on the the total latency and resources
2141 // inside and outside this zone. Potential stalls should be considered before
2142 // following this policy.
2143
2144 // Compute remaining latency. We need this both to determine whether the
2145 // overall schedule has become latency-limited and whether the instructions
2146 // outside this zone are resource or latency limited.
2147 //
2148 // The "dependent" latency is updated incrementally during scheduling as the
2149 // max height/depth of scheduled nodes minus the cycles since it was
2150 // scheduled:
2151 // DLat = max (N.depth - (CurrCycle - N.ReadyCycle) for N in Zone
2152 //
2153 // The "independent" latency is the max ready queue depth:
2154 // ILat = max N.depth for N in Available|Pending
2155 //
2156 // RemainingLatency is the greater of independent and dependent latency.
2157 unsigned RemLatency = CurrZone.getDependentLatency();
2158 RemLatency = std::max(RemLatency,
2159 CurrZone.findMaxLatency(CurrZone.Available.elements()));
2160 RemLatency = std::max(RemLatency,
2161 CurrZone.findMaxLatency(CurrZone.Pending.elements()));
2162
2163 // Compute the critical resource outside the zone.
2164 unsigned OtherCritIdx = 0;
2165 unsigned OtherCount =
2166 OtherZone ? OtherZone->getOtherResourceCount(OtherCritIdx) : 0;
2167
2168 bool OtherResLimited = false;
2169 if (SchedModel->hasInstrSchedModel()) {
2170 unsigned LFactor = SchedModel->getLatencyFactor();
2171 OtherResLimited = (int)(OtherCount - (RemLatency * LFactor)) > (int)LFactor;
2172 }
2173 // Schedule aggressively for latency in PostRA mode. We don't check for
2174 // acyclic latency during PostRA, and highly out-of-order processors will
2175 // skip PostRA scheduling.
2176 if (!OtherResLimited) {
2177 if (IsPostRA || (RemLatency + CurrZone.getCurrCycle() > Rem.CriticalPath)) {
2178 Policy.ReduceLatency |= true;
2179 DEBUG(dbgs() << " " << CurrZone.Available.getName()
2180 << " RemainingLatency " << RemLatency << " + "
2181 << CurrZone.getCurrCycle() << "c > CritPath "
2182 << Rem.CriticalPath << "\n");
2183 }
2184 }
2185 // If the same resource is limiting inside and outside the zone, do nothing.
2186 if (CurrZone.getZoneCritResIdx() == OtherCritIdx)
2187 return;
2188
2189 DEBUG(
2190 if (CurrZone.isResourceLimited()) {
2191 dbgs() << " " << CurrZone.Available.getName() << " ResourceLimited: "
2192 << SchedModel->getResourceName(CurrZone.getZoneCritResIdx())
2193 << "\n";
2194 }
2195 if (OtherResLimited)
2196 dbgs() << " RemainingLimit: "
2197 << SchedModel->getResourceName(OtherCritIdx) << "\n";
2198 if (!CurrZone.isResourceLimited() && !OtherResLimited)
2199 dbgs() << " Latency limited both directions.\n");
2200
2201 if (CurrZone.isResourceLimited() && !Policy.ReduceResIdx)
2202 Policy.ReduceResIdx = CurrZone.getZoneCritResIdx();
2203
2204 if (OtherResLimited)
2205 Policy.DemandResIdx = OtherCritIdx;
2206 }
2207
2208 #ifndef NDEBUG
getReasonStr(GenericSchedulerBase::CandReason Reason)2209 const char *GenericSchedulerBase::getReasonStr(
2210 GenericSchedulerBase::CandReason Reason) {
2211 switch (Reason) {
2212 case NoCand: return "NOCAND ";
2213 case PhysRegCopy: return "PREG-COPY";
2214 case RegExcess: return "REG-EXCESS";
2215 case RegCritical: return "REG-CRIT ";
2216 case Stall: return "STALL ";
2217 case Cluster: return "CLUSTER ";
2218 case Weak: return "WEAK ";
2219 case RegMax: return "REG-MAX ";
2220 case ResourceReduce: return "RES-REDUCE";
2221 case ResourceDemand: return "RES-DEMAND";
2222 case TopDepthReduce: return "TOP-DEPTH ";
2223 case TopPathReduce: return "TOP-PATH ";
2224 case BotHeightReduce:return "BOT-HEIGHT";
2225 case BotPathReduce: return "BOT-PATH ";
2226 case NextDefUse: return "DEF-USE ";
2227 case NodeOrder: return "ORDER ";
2228 };
2229 llvm_unreachable("Unknown reason!");
2230 }
2231
traceCandidate(const SchedCandidate & Cand)2232 void GenericSchedulerBase::traceCandidate(const SchedCandidate &Cand) {
2233 PressureChange P;
2234 unsigned ResIdx = 0;
2235 unsigned Latency = 0;
2236 switch (Cand.Reason) {
2237 default:
2238 break;
2239 case RegExcess:
2240 P = Cand.RPDelta.Excess;
2241 break;
2242 case RegCritical:
2243 P = Cand.RPDelta.CriticalMax;
2244 break;
2245 case RegMax:
2246 P = Cand.RPDelta.CurrentMax;
2247 break;
2248 case ResourceReduce:
2249 ResIdx = Cand.Policy.ReduceResIdx;
2250 break;
2251 case ResourceDemand:
2252 ResIdx = Cand.Policy.DemandResIdx;
2253 break;
2254 case TopDepthReduce:
2255 Latency = Cand.SU->getDepth();
2256 break;
2257 case TopPathReduce:
2258 Latency = Cand.SU->getHeight();
2259 break;
2260 case BotHeightReduce:
2261 Latency = Cand.SU->getHeight();
2262 break;
2263 case BotPathReduce:
2264 Latency = Cand.SU->getDepth();
2265 break;
2266 }
2267 dbgs() << " SU(" << Cand.SU->NodeNum << ") " << getReasonStr(Cand.Reason);
2268 if (P.isValid())
2269 dbgs() << " " << TRI->getRegPressureSetName(P.getPSet())
2270 << ":" << P.getUnitInc() << " ";
2271 else
2272 dbgs() << " ";
2273 if (ResIdx)
2274 dbgs() << " " << SchedModel->getProcResource(ResIdx)->Name << " ";
2275 else
2276 dbgs() << " ";
2277 if (Latency)
2278 dbgs() << " " << Latency << " cycles ";
2279 else
2280 dbgs() << " ";
2281 dbgs() << '\n';
2282 }
2283 #endif
2284
2285 /// Return true if this heuristic determines order.
tryLess(int TryVal,int CandVal,GenericSchedulerBase::SchedCandidate & TryCand,GenericSchedulerBase::SchedCandidate & Cand,GenericSchedulerBase::CandReason Reason)2286 static bool tryLess(int TryVal, int CandVal,
2287 GenericSchedulerBase::SchedCandidate &TryCand,
2288 GenericSchedulerBase::SchedCandidate &Cand,
2289 GenericSchedulerBase::CandReason Reason) {
2290 if (TryVal < CandVal) {
2291 TryCand.Reason = Reason;
2292 return true;
2293 }
2294 if (TryVal > CandVal) {
2295 if (Cand.Reason > Reason)
2296 Cand.Reason = Reason;
2297 return true;
2298 }
2299 Cand.setRepeat(Reason);
2300 return false;
2301 }
2302
tryGreater(int TryVal,int CandVal,GenericSchedulerBase::SchedCandidate & TryCand,GenericSchedulerBase::SchedCandidate & Cand,GenericSchedulerBase::CandReason Reason)2303 static bool tryGreater(int TryVal, int CandVal,
2304 GenericSchedulerBase::SchedCandidate &TryCand,
2305 GenericSchedulerBase::SchedCandidate &Cand,
2306 GenericSchedulerBase::CandReason Reason) {
2307 if (TryVal > CandVal) {
2308 TryCand.Reason = Reason;
2309 return true;
2310 }
2311 if (TryVal < CandVal) {
2312 if (Cand.Reason > Reason)
2313 Cand.Reason = Reason;
2314 return true;
2315 }
2316 Cand.setRepeat(Reason);
2317 return false;
2318 }
2319
tryLatency(GenericSchedulerBase::SchedCandidate & TryCand,GenericSchedulerBase::SchedCandidate & Cand,SchedBoundary & Zone)2320 static bool tryLatency(GenericSchedulerBase::SchedCandidate &TryCand,
2321 GenericSchedulerBase::SchedCandidate &Cand,
2322 SchedBoundary &Zone) {
2323 if (Zone.isTop()) {
2324 if (Cand.SU->getDepth() > Zone.getScheduledLatency()) {
2325 if (tryLess(TryCand.SU->getDepth(), Cand.SU->getDepth(),
2326 TryCand, Cand, GenericSchedulerBase::TopDepthReduce))
2327 return true;
2328 }
2329 if (tryGreater(TryCand.SU->getHeight(), Cand.SU->getHeight(),
2330 TryCand, Cand, GenericSchedulerBase::TopPathReduce))
2331 return true;
2332 }
2333 else {
2334 if (Cand.SU->getHeight() > Zone.getScheduledLatency()) {
2335 if (tryLess(TryCand.SU->getHeight(), Cand.SU->getHeight(),
2336 TryCand, Cand, GenericSchedulerBase::BotHeightReduce))
2337 return true;
2338 }
2339 if (tryGreater(TryCand.SU->getDepth(), Cand.SU->getDepth(),
2340 TryCand, Cand, GenericSchedulerBase::BotPathReduce))
2341 return true;
2342 }
2343 return false;
2344 }
2345
tracePick(const GenericSchedulerBase::SchedCandidate & Cand,bool IsTop)2346 static void tracePick(const GenericSchedulerBase::SchedCandidate &Cand,
2347 bool IsTop) {
2348 DEBUG(dbgs() << "Pick " << (IsTop ? "Top " : "Bot ")
2349 << GenericSchedulerBase::getReasonStr(Cand.Reason) << '\n');
2350 }
2351
initialize(ScheduleDAGMI * dag)2352 void GenericScheduler::initialize(ScheduleDAGMI *dag) {
2353 assert(dag->hasVRegLiveness() &&
2354 "(PreRA)GenericScheduler needs vreg liveness");
2355 DAG = static_cast<ScheduleDAGMILive*>(dag);
2356 SchedModel = DAG->getSchedModel();
2357 TRI = DAG->TRI;
2358
2359 Rem.init(DAG, SchedModel);
2360 Top.init(DAG, SchedModel, &Rem);
2361 Bot.init(DAG, SchedModel, &Rem);
2362
2363 // Initialize resource counts.
2364
2365 // Initialize the HazardRecognizers. If itineraries don't exist, are empty, or
2366 // are disabled, then these HazardRecs will be disabled.
2367 const InstrItineraryData *Itin = SchedModel->getInstrItineraries();
2368 if (!Top.HazardRec) {
2369 Top.HazardRec =
2370 DAG->MF.getSubtarget().getInstrInfo()->CreateTargetMIHazardRecognizer(
2371 Itin, DAG);
2372 }
2373 if (!Bot.HazardRec) {
2374 Bot.HazardRec =
2375 DAG->MF.getSubtarget().getInstrInfo()->CreateTargetMIHazardRecognizer(
2376 Itin, DAG);
2377 }
2378 }
2379
2380 /// Initialize the per-region scheduling policy.
initPolicy(MachineBasicBlock::iterator Begin,MachineBasicBlock::iterator End,unsigned NumRegionInstrs)2381 void GenericScheduler::initPolicy(MachineBasicBlock::iterator Begin,
2382 MachineBasicBlock::iterator End,
2383 unsigned NumRegionInstrs) {
2384 const MachineFunction &MF = *Begin->getParent()->getParent();
2385 const TargetLowering *TLI = MF.getSubtarget().getTargetLowering();
2386
2387 // Avoid setting up the register pressure tracker for small regions to save
2388 // compile time. As a rough heuristic, only track pressure when the number of
2389 // schedulable instructions exceeds half the integer register file.
2390 RegionPolicy.ShouldTrackPressure = true;
2391 for (unsigned VT = MVT::i32; VT > (unsigned)MVT::i1; --VT) {
2392 MVT::SimpleValueType LegalIntVT = (MVT::SimpleValueType)VT;
2393 if (TLI->isTypeLegal(LegalIntVT)) {
2394 unsigned NIntRegs = Context->RegClassInfo->getNumAllocatableRegs(
2395 TLI->getRegClassFor(LegalIntVT));
2396 RegionPolicy.ShouldTrackPressure = NumRegionInstrs > (NIntRegs / 2);
2397 }
2398 }
2399
2400 // For generic targets, we default to bottom-up, because it's simpler and more
2401 // compile-time optimizations have been implemented in that direction.
2402 RegionPolicy.OnlyBottomUp = true;
2403
2404 // Allow the subtarget to override default policy.
2405 MF.getSubtarget().overrideSchedPolicy(RegionPolicy, Begin, End,
2406 NumRegionInstrs);
2407
2408 // After subtarget overrides, apply command line options.
2409 if (!EnableRegPressure)
2410 RegionPolicy.ShouldTrackPressure = false;
2411
2412 // Check -misched-topdown/bottomup can force or unforce scheduling direction.
2413 // e.g. -misched-bottomup=false allows scheduling in both directions.
2414 assert((!ForceTopDown || !ForceBottomUp) &&
2415 "-misched-topdown incompatible with -misched-bottomup");
2416 if (ForceBottomUp.getNumOccurrences() > 0) {
2417 RegionPolicy.OnlyBottomUp = ForceBottomUp;
2418 if (RegionPolicy.OnlyBottomUp)
2419 RegionPolicy.OnlyTopDown = false;
2420 }
2421 if (ForceTopDown.getNumOccurrences() > 0) {
2422 RegionPolicy.OnlyTopDown = ForceTopDown;
2423 if (RegionPolicy.OnlyTopDown)
2424 RegionPolicy.OnlyBottomUp = false;
2425 }
2426 }
2427
2428 /// Set IsAcyclicLatencyLimited if the acyclic path is longer than the cyclic
2429 /// critical path by more cycles than it takes to drain the instruction buffer.
2430 /// We estimate an upper bounds on in-flight instructions as:
2431 ///
2432 /// CyclesPerIteration = max( CyclicPath, Loop-Resource-Height )
2433 /// InFlightIterations = AcyclicPath / CyclesPerIteration
2434 /// InFlightResources = InFlightIterations * LoopResources
2435 ///
2436 /// TODO: Check execution resources in addition to IssueCount.
checkAcyclicLatency()2437 void GenericScheduler::checkAcyclicLatency() {
2438 if (Rem.CyclicCritPath == 0 || Rem.CyclicCritPath >= Rem.CriticalPath)
2439 return;
2440
2441 // Scaled number of cycles per loop iteration.
2442 unsigned IterCount =
2443 std::max(Rem.CyclicCritPath * SchedModel->getLatencyFactor(),
2444 Rem.RemIssueCount);
2445 // Scaled acyclic critical path.
2446 unsigned AcyclicCount = Rem.CriticalPath * SchedModel->getLatencyFactor();
2447 // InFlightCount = (AcyclicPath / IterCycles) * InstrPerLoop
2448 unsigned InFlightCount =
2449 (AcyclicCount * Rem.RemIssueCount + IterCount-1) / IterCount;
2450 unsigned BufferLimit =
2451 SchedModel->getMicroOpBufferSize() * SchedModel->getMicroOpFactor();
2452
2453 Rem.IsAcyclicLatencyLimited = InFlightCount > BufferLimit;
2454
2455 DEBUG(dbgs() << "IssueCycles="
2456 << Rem.RemIssueCount / SchedModel->getLatencyFactor() << "c "
2457 << "IterCycles=" << IterCount / SchedModel->getLatencyFactor()
2458 << "c NumIters=" << (AcyclicCount + IterCount-1) / IterCount
2459 << " InFlight=" << InFlightCount / SchedModel->getMicroOpFactor()
2460 << "m BufferLim=" << SchedModel->getMicroOpBufferSize() << "m\n";
2461 if (Rem.IsAcyclicLatencyLimited)
2462 dbgs() << " ACYCLIC LATENCY LIMIT\n");
2463 }
2464
registerRoots()2465 void GenericScheduler::registerRoots() {
2466 Rem.CriticalPath = DAG->ExitSU.getDepth();
2467
2468 // Some roots may not feed into ExitSU. Check all of them in case.
2469 for (std::vector<SUnit*>::const_iterator
2470 I = Bot.Available.begin(), E = Bot.Available.end(); I != E; ++I) {
2471 if ((*I)->getDepth() > Rem.CriticalPath)
2472 Rem.CriticalPath = (*I)->getDepth();
2473 }
2474 DEBUG(dbgs() << "Critical Path(GS-RR ): " << Rem.CriticalPath << '\n');
2475 if (DumpCriticalPathLength) {
2476 errs() << "Critical Path(GS-RR ): " << Rem.CriticalPath << " \n";
2477 }
2478
2479 if (EnableCyclicPath) {
2480 Rem.CyclicCritPath = DAG->computeCyclicCriticalPath();
2481 checkAcyclicLatency();
2482 }
2483 }
2484
tryPressure(const PressureChange & TryP,const PressureChange & CandP,GenericSchedulerBase::SchedCandidate & TryCand,GenericSchedulerBase::SchedCandidate & Cand,GenericSchedulerBase::CandReason Reason)2485 static bool tryPressure(const PressureChange &TryP,
2486 const PressureChange &CandP,
2487 GenericSchedulerBase::SchedCandidate &TryCand,
2488 GenericSchedulerBase::SchedCandidate &Cand,
2489 GenericSchedulerBase::CandReason Reason) {
2490 int TryRank = TryP.getPSetOrMax();
2491 int CandRank = CandP.getPSetOrMax();
2492 // If both candidates affect the same set, go with the smallest increase.
2493 if (TryRank == CandRank) {
2494 return tryLess(TryP.getUnitInc(), CandP.getUnitInc(), TryCand, Cand,
2495 Reason);
2496 }
2497 // If one candidate decreases and the other increases, go with it.
2498 // Invalid candidates have UnitInc==0.
2499 if (tryGreater(TryP.getUnitInc() < 0, CandP.getUnitInc() < 0, TryCand, Cand,
2500 Reason)) {
2501 return true;
2502 }
2503 // If the candidates are decreasing pressure, reverse priority.
2504 if (TryP.getUnitInc() < 0)
2505 std::swap(TryRank, CandRank);
2506 return tryGreater(TryRank, CandRank, TryCand, Cand, Reason);
2507 }
2508
getWeakLeft(const SUnit * SU,bool isTop)2509 static unsigned getWeakLeft(const SUnit *SU, bool isTop) {
2510 return (isTop) ? SU->WeakPredsLeft : SU->WeakSuccsLeft;
2511 }
2512
2513 /// Minimize physical register live ranges. Regalloc wants them adjacent to
2514 /// their physreg def/use.
2515 ///
2516 /// FIXME: This is an unnecessary check on the critical path. Most are root/leaf
2517 /// copies which can be prescheduled. The rest (e.g. x86 MUL) could be bundled
2518 /// with the operation that produces or consumes the physreg. We'll do this when
2519 /// regalloc has support for parallel copies.
biasPhysRegCopy(const SUnit * SU,bool isTop)2520 static int biasPhysRegCopy(const SUnit *SU, bool isTop) {
2521 const MachineInstr *MI = SU->getInstr();
2522 if (!MI->isCopy())
2523 return 0;
2524
2525 unsigned ScheduledOper = isTop ? 1 : 0;
2526 unsigned UnscheduledOper = isTop ? 0 : 1;
2527 // If we have already scheduled the physreg produce/consumer, immediately
2528 // schedule the copy.
2529 if (TargetRegisterInfo::isPhysicalRegister(
2530 MI->getOperand(ScheduledOper).getReg()))
2531 return 1;
2532 // If the physreg is at the boundary, defer it. Otherwise schedule it
2533 // immediately to free the dependent. We can hoist the copy later.
2534 bool AtBoundary = isTop ? !SU->NumSuccsLeft : !SU->NumPredsLeft;
2535 if (TargetRegisterInfo::isPhysicalRegister(
2536 MI->getOperand(UnscheduledOper).getReg()))
2537 return AtBoundary ? -1 : 1;
2538 return 0;
2539 }
2540
2541 /// Apply a set of heursitics to a new candidate. Heuristics are currently
2542 /// hierarchical. This may be more efficient than a graduated cost model because
2543 /// we don't need to evaluate all aspects of the model for each node in the
2544 /// queue. But it's really done to make the heuristics easier to debug and
2545 /// statistically analyze.
2546 ///
2547 /// \param Cand provides the policy and current best candidate.
2548 /// \param TryCand refers to the next SUnit candidate, otherwise uninitialized.
2549 /// \param Zone describes the scheduled zone that we are extending.
2550 /// \param RPTracker describes reg pressure within the scheduled zone.
2551 /// \param TempTracker is a scratch pressure tracker to reuse in queries.
tryCandidate(SchedCandidate & Cand,SchedCandidate & TryCand,SchedBoundary & Zone,const RegPressureTracker & RPTracker,RegPressureTracker & TempTracker)2552 void GenericScheduler::tryCandidate(SchedCandidate &Cand,
2553 SchedCandidate &TryCand,
2554 SchedBoundary &Zone,
2555 const RegPressureTracker &RPTracker,
2556 RegPressureTracker &TempTracker) {
2557
2558 if (DAG->isTrackingPressure()) {
2559 // Always initialize TryCand's RPDelta.
2560 if (Zone.isTop()) {
2561 TempTracker.getMaxDownwardPressureDelta(
2562 TryCand.SU->getInstr(),
2563 TryCand.RPDelta,
2564 DAG->getRegionCriticalPSets(),
2565 DAG->getRegPressure().MaxSetPressure);
2566 }
2567 else {
2568 if (VerifyScheduling) {
2569 TempTracker.getMaxUpwardPressureDelta(
2570 TryCand.SU->getInstr(),
2571 &DAG->getPressureDiff(TryCand.SU),
2572 TryCand.RPDelta,
2573 DAG->getRegionCriticalPSets(),
2574 DAG->getRegPressure().MaxSetPressure);
2575 }
2576 else {
2577 RPTracker.getUpwardPressureDelta(
2578 TryCand.SU->getInstr(),
2579 DAG->getPressureDiff(TryCand.SU),
2580 TryCand.RPDelta,
2581 DAG->getRegionCriticalPSets(),
2582 DAG->getRegPressure().MaxSetPressure);
2583 }
2584 }
2585 }
2586 DEBUG(if (TryCand.RPDelta.Excess.isValid())
2587 dbgs() << " SU(" << TryCand.SU->NodeNum << ") "
2588 << TRI->getRegPressureSetName(TryCand.RPDelta.Excess.getPSet())
2589 << ":" << TryCand.RPDelta.Excess.getUnitInc() << "\n");
2590
2591 // Initialize the candidate if needed.
2592 if (!Cand.isValid()) {
2593 TryCand.Reason = NodeOrder;
2594 return;
2595 }
2596
2597 if (tryGreater(biasPhysRegCopy(TryCand.SU, Zone.isTop()),
2598 biasPhysRegCopy(Cand.SU, Zone.isTop()),
2599 TryCand, Cand, PhysRegCopy))
2600 return;
2601
2602 // Avoid exceeding the target's limit. If signed PSetID is negative, it is
2603 // invalid; convert it to INT_MAX to give it lowest priority.
2604 if (DAG->isTrackingPressure() && tryPressure(TryCand.RPDelta.Excess,
2605 Cand.RPDelta.Excess,
2606 TryCand, Cand, RegExcess))
2607 return;
2608
2609 // Avoid increasing the max critical pressure in the scheduled region.
2610 if (DAG->isTrackingPressure() && tryPressure(TryCand.RPDelta.CriticalMax,
2611 Cand.RPDelta.CriticalMax,
2612 TryCand, Cand, RegCritical))
2613 return;
2614
2615 // For loops that are acyclic path limited, aggressively schedule for latency.
2616 // This can result in very long dependence chains scheduled in sequence, so
2617 // once every cycle (when CurrMOps == 0), switch to normal heuristics.
2618 if (Rem.IsAcyclicLatencyLimited && !Zone.getCurrMOps()
2619 && tryLatency(TryCand, Cand, Zone))
2620 return;
2621
2622 // Prioritize instructions that read unbuffered resources by stall cycles.
2623 if (tryLess(Zone.getLatencyStallCycles(TryCand.SU),
2624 Zone.getLatencyStallCycles(Cand.SU), TryCand, Cand, Stall))
2625 return;
2626
2627 // Keep clustered nodes together to encourage downstream peephole
2628 // optimizations which may reduce resource requirements.
2629 //
2630 // This is a best effort to set things up for a post-RA pass. Optimizations
2631 // like generating loads of multiple registers should ideally be done within
2632 // the scheduler pass by combining the loads during DAG postprocessing.
2633 const SUnit *NextClusterSU =
2634 Zone.isTop() ? DAG->getNextClusterSucc() : DAG->getNextClusterPred();
2635 if (tryGreater(TryCand.SU == NextClusterSU, Cand.SU == NextClusterSU,
2636 TryCand, Cand, Cluster))
2637 return;
2638
2639 // Weak edges are for clustering and other constraints.
2640 if (tryLess(getWeakLeft(TryCand.SU, Zone.isTop()),
2641 getWeakLeft(Cand.SU, Zone.isTop()),
2642 TryCand, Cand, Weak)) {
2643 return;
2644 }
2645 // Avoid increasing the max pressure of the entire region.
2646 if (DAG->isTrackingPressure() && tryPressure(TryCand.RPDelta.CurrentMax,
2647 Cand.RPDelta.CurrentMax,
2648 TryCand, Cand, RegMax))
2649 return;
2650
2651 // Avoid critical resource consumption and balance the schedule.
2652 TryCand.initResourceDelta(DAG, SchedModel);
2653 if (tryLess(TryCand.ResDelta.CritResources, Cand.ResDelta.CritResources,
2654 TryCand, Cand, ResourceReduce))
2655 return;
2656 if (tryGreater(TryCand.ResDelta.DemandedResources,
2657 Cand.ResDelta.DemandedResources,
2658 TryCand, Cand, ResourceDemand))
2659 return;
2660
2661 // Avoid serializing long latency dependence chains.
2662 // For acyclic path limited loops, latency was already checked above.
2663 if (Cand.Policy.ReduceLatency && !Rem.IsAcyclicLatencyLimited
2664 && tryLatency(TryCand, Cand, Zone)) {
2665 return;
2666 }
2667
2668 // Prefer immediate defs/users of the last scheduled instruction. This is a
2669 // local pressure avoidance strategy that also makes the machine code
2670 // readable.
2671 if (tryGreater(Zone.isNextSU(TryCand.SU), Zone.isNextSU(Cand.SU),
2672 TryCand, Cand, NextDefUse))
2673 return;
2674
2675 // Fall through to original instruction order.
2676 if ((Zone.isTop() && TryCand.SU->NodeNum < Cand.SU->NodeNum)
2677 || (!Zone.isTop() && TryCand.SU->NodeNum > Cand.SU->NodeNum)) {
2678 TryCand.Reason = NodeOrder;
2679 }
2680 }
2681
2682 /// Pick the best candidate from the queue.
2683 ///
2684 /// TODO: getMaxPressureDelta results can be mostly cached for each SUnit during
2685 /// DAG building. To adjust for the current scheduling location we need to
2686 /// maintain the number of vreg uses remaining to be top-scheduled.
pickNodeFromQueue(SchedBoundary & Zone,const RegPressureTracker & RPTracker,SchedCandidate & Cand)2687 void GenericScheduler::pickNodeFromQueue(SchedBoundary &Zone,
2688 const RegPressureTracker &RPTracker,
2689 SchedCandidate &Cand) {
2690 ReadyQueue &Q = Zone.Available;
2691
2692 DEBUG(Q.dump());
2693
2694 // getMaxPressureDelta temporarily modifies the tracker.
2695 RegPressureTracker &TempTracker = const_cast<RegPressureTracker&>(RPTracker);
2696
2697 for (ReadyQueue::iterator I = Q.begin(), E = Q.end(); I != E; ++I) {
2698
2699 SchedCandidate TryCand(Cand.Policy);
2700 TryCand.SU = *I;
2701 tryCandidate(Cand, TryCand, Zone, RPTracker, TempTracker);
2702 if (TryCand.Reason != NoCand) {
2703 // Initialize resource delta if needed in case future heuristics query it.
2704 if (TryCand.ResDelta == SchedResourceDelta())
2705 TryCand.initResourceDelta(DAG, SchedModel);
2706 Cand.setBest(TryCand);
2707 DEBUG(traceCandidate(Cand));
2708 }
2709 }
2710 }
2711
2712 /// Pick the best candidate node from either the top or bottom queue.
pickNodeBidirectional(bool & IsTopNode)2713 SUnit *GenericScheduler::pickNodeBidirectional(bool &IsTopNode) {
2714 // Schedule as far as possible in the direction of no choice. This is most
2715 // efficient, but also provides the best heuristics for CriticalPSets.
2716 if (SUnit *SU = Bot.pickOnlyChoice()) {
2717 IsTopNode = false;
2718 DEBUG(dbgs() << "Pick Bot NOCAND\n");
2719 return SU;
2720 }
2721 if (SUnit *SU = Top.pickOnlyChoice()) {
2722 IsTopNode = true;
2723 DEBUG(dbgs() << "Pick Top NOCAND\n");
2724 return SU;
2725 }
2726 CandPolicy NoPolicy;
2727 SchedCandidate BotCand(NoPolicy);
2728 SchedCandidate TopCand(NoPolicy);
2729 // Set the bottom-up policy based on the state of the current bottom zone and
2730 // the instructions outside the zone, including the top zone.
2731 setPolicy(BotCand.Policy, /*IsPostRA=*/false, Bot, &Top);
2732 // Set the top-down policy based on the state of the current top zone and
2733 // the instructions outside the zone, including the bottom zone.
2734 setPolicy(TopCand.Policy, /*IsPostRA=*/false, Top, &Bot);
2735
2736 // Prefer bottom scheduling when heuristics are silent.
2737 pickNodeFromQueue(Bot, DAG->getBotRPTracker(), BotCand);
2738 assert(BotCand.Reason != NoCand && "failed to find the first candidate");
2739
2740 // If either Q has a single candidate that provides the least increase in
2741 // Excess pressure, we can immediately schedule from that Q.
2742 //
2743 // RegionCriticalPSets summarizes the pressure within the scheduled region and
2744 // affects picking from either Q. If scheduling in one direction must
2745 // increase pressure for one of the excess PSets, then schedule in that
2746 // direction first to provide more freedom in the other direction.
2747 if ((BotCand.Reason == RegExcess && !BotCand.isRepeat(RegExcess))
2748 || (BotCand.Reason == RegCritical
2749 && !BotCand.isRepeat(RegCritical)))
2750 {
2751 IsTopNode = false;
2752 tracePick(BotCand, IsTopNode);
2753 return BotCand.SU;
2754 }
2755 // Check if the top Q has a better candidate.
2756 pickNodeFromQueue(Top, DAG->getTopRPTracker(), TopCand);
2757 assert(TopCand.Reason != NoCand && "failed to find the first candidate");
2758
2759 // Choose the queue with the most important (lowest enum) reason.
2760 if (TopCand.Reason < BotCand.Reason) {
2761 IsTopNode = true;
2762 tracePick(TopCand, IsTopNode);
2763 return TopCand.SU;
2764 }
2765 // Otherwise prefer the bottom candidate, in node order if all else failed.
2766 IsTopNode = false;
2767 tracePick(BotCand, IsTopNode);
2768 return BotCand.SU;
2769 }
2770
2771 /// Pick the best node to balance the schedule. Implements MachineSchedStrategy.
pickNode(bool & IsTopNode)2772 SUnit *GenericScheduler::pickNode(bool &IsTopNode) {
2773 if (DAG->top() == DAG->bottom()) {
2774 assert(Top.Available.empty() && Top.Pending.empty() &&
2775 Bot.Available.empty() && Bot.Pending.empty() && "ReadyQ garbage");
2776 return nullptr;
2777 }
2778 SUnit *SU;
2779 do {
2780 if (RegionPolicy.OnlyTopDown) {
2781 SU = Top.pickOnlyChoice();
2782 if (!SU) {
2783 CandPolicy NoPolicy;
2784 SchedCandidate TopCand(NoPolicy);
2785 pickNodeFromQueue(Top, DAG->getTopRPTracker(), TopCand);
2786 assert(TopCand.Reason != NoCand && "failed to find a candidate");
2787 tracePick(TopCand, true);
2788 SU = TopCand.SU;
2789 }
2790 IsTopNode = true;
2791 }
2792 else if (RegionPolicy.OnlyBottomUp) {
2793 SU = Bot.pickOnlyChoice();
2794 if (!SU) {
2795 CandPolicy NoPolicy;
2796 SchedCandidate BotCand(NoPolicy);
2797 pickNodeFromQueue(Bot, DAG->getBotRPTracker(), BotCand);
2798 assert(BotCand.Reason != NoCand && "failed to find a candidate");
2799 tracePick(BotCand, false);
2800 SU = BotCand.SU;
2801 }
2802 IsTopNode = false;
2803 }
2804 else {
2805 SU = pickNodeBidirectional(IsTopNode);
2806 }
2807 } while (SU->isScheduled);
2808
2809 if (SU->isTopReady())
2810 Top.removeReady(SU);
2811 if (SU->isBottomReady())
2812 Bot.removeReady(SU);
2813
2814 DEBUG(dbgs() << "Scheduling SU(" << SU->NodeNum << ") " << *SU->getInstr());
2815 return SU;
2816 }
2817
reschedulePhysRegCopies(SUnit * SU,bool isTop)2818 void GenericScheduler::reschedulePhysRegCopies(SUnit *SU, bool isTop) {
2819
2820 MachineBasicBlock::iterator InsertPos = SU->getInstr();
2821 if (!isTop)
2822 ++InsertPos;
2823 SmallVectorImpl<SDep> &Deps = isTop ? SU->Preds : SU->Succs;
2824
2825 // Find already scheduled copies with a single physreg dependence and move
2826 // them just above the scheduled instruction.
2827 for (SmallVectorImpl<SDep>::iterator I = Deps.begin(), E = Deps.end();
2828 I != E; ++I) {
2829 if (I->getKind() != SDep::Data || !TRI->isPhysicalRegister(I->getReg()))
2830 continue;
2831 SUnit *DepSU = I->getSUnit();
2832 if (isTop ? DepSU->Succs.size() > 1 : DepSU->Preds.size() > 1)
2833 continue;
2834 MachineInstr *Copy = DepSU->getInstr();
2835 if (!Copy->isCopy())
2836 continue;
2837 DEBUG(dbgs() << " Rescheduling physreg copy ";
2838 I->getSUnit()->dump(DAG));
2839 DAG->moveInstruction(Copy, InsertPos);
2840 }
2841 }
2842
2843 /// Update the scheduler's state after scheduling a node. This is the same node
2844 /// that was just returned by pickNode(). However, ScheduleDAGMILive needs to
2845 /// update it's state based on the current cycle before MachineSchedStrategy
2846 /// does.
2847 ///
2848 /// FIXME: Eventually, we may bundle physreg copies rather than rescheduling
2849 /// them here. See comments in biasPhysRegCopy.
schedNode(SUnit * SU,bool IsTopNode)2850 void GenericScheduler::schedNode(SUnit *SU, bool IsTopNode) {
2851 if (IsTopNode) {
2852 SU->TopReadyCycle = std::max(SU->TopReadyCycle, Top.getCurrCycle());
2853 Top.bumpNode(SU);
2854 if (SU->hasPhysRegUses)
2855 reschedulePhysRegCopies(SU, true);
2856 }
2857 else {
2858 SU->BotReadyCycle = std::max(SU->BotReadyCycle, Bot.getCurrCycle());
2859 Bot.bumpNode(SU);
2860 if (SU->hasPhysRegDefs)
2861 reschedulePhysRegCopies(SU, false);
2862 }
2863 }
2864
2865 /// Create the standard converging machine scheduler. This will be used as the
2866 /// default scheduler if the target does not set a default.
createGenericSchedLive(MachineSchedContext * C)2867 static ScheduleDAGInstrs *createGenericSchedLive(MachineSchedContext *C) {
2868 ScheduleDAGMILive *DAG = new ScheduleDAGMILive(C, make_unique<GenericScheduler>(C));
2869 // Register DAG post-processors.
2870 //
2871 // FIXME: extend the mutation API to allow earlier mutations to instantiate
2872 // data and pass it to later mutations. Have a single mutation that gathers
2873 // the interesting nodes in one pass.
2874 DAG->addMutation(make_unique<CopyConstrain>(DAG->TII, DAG->TRI));
2875 if (EnableLoadCluster && DAG->TII->enableClusterLoads())
2876 DAG->addMutation(make_unique<LoadClusterMutation>(DAG->TII, DAG->TRI));
2877 if (EnableMacroFusion)
2878 DAG->addMutation(make_unique<MacroFusion>(DAG->TII));
2879 return DAG;
2880 }
2881
2882 static MachineSchedRegistry
2883 GenericSchedRegistry("converge", "Standard converging scheduler.",
2884 createGenericSchedLive);
2885
2886 //===----------------------------------------------------------------------===//
2887 // PostGenericScheduler - Generic PostRA implementation of MachineSchedStrategy.
2888 //===----------------------------------------------------------------------===//
2889
initialize(ScheduleDAGMI * Dag)2890 void PostGenericScheduler::initialize(ScheduleDAGMI *Dag) {
2891 DAG = Dag;
2892 SchedModel = DAG->getSchedModel();
2893 TRI = DAG->TRI;
2894
2895 Rem.init(DAG, SchedModel);
2896 Top.init(DAG, SchedModel, &Rem);
2897 BotRoots.clear();
2898
2899 // Initialize the HazardRecognizers. If itineraries don't exist, are empty,
2900 // or are disabled, then these HazardRecs will be disabled.
2901 const InstrItineraryData *Itin = SchedModel->getInstrItineraries();
2902 if (!Top.HazardRec) {
2903 Top.HazardRec =
2904 DAG->MF.getSubtarget().getInstrInfo()->CreateTargetMIHazardRecognizer(
2905 Itin, DAG);
2906 }
2907 }
2908
2909
registerRoots()2910 void PostGenericScheduler::registerRoots() {
2911 Rem.CriticalPath = DAG->ExitSU.getDepth();
2912
2913 // Some roots may not feed into ExitSU. Check all of them in case.
2914 for (SmallVectorImpl<SUnit*>::const_iterator
2915 I = BotRoots.begin(), E = BotRoots.end(); I != E; ++I) {
2916 if ((*I)->getDepth() > Rem.CriticalPath)
2917 Rem.CriticalPath = (*I)->getDepth();
2918 }
2919 DEBUG(dbgs() << "Critical Path: (PGS-RR) " << Rem.CriticalPath << '\n');
2920 if (DumpCriticalPathLength) {
2921 errs() << "Critical Path(PGS-RR ): " << Rem.CriticalPath << " \n";
2922 }
2923 }
2924
2925 /// Apply a set of heursitics to a new candidate for PostRA scheduling.
2926 ///
2927 /// \param Cand provides the policy and current best candidate.
2928 /// \param TryCand refers to the next SUnit candidate, otherwise uninitialized.
tryCandidate(SchedCandidate & Cand,SchedCandidate & TryCand)2929 void PostGenericScheduler::tryCandidate(SchedCandidate &Cand,
2930 SchedCandidate &TryCand) {
2931
2932 // Initialize the candidate if needed.
2933 if (!Cand.isValid()) {
2934 TryCand.Reason = NodeOrder;
2935 return;
2936 }
2937
2938 // Prioritize instructions that read unbuffered resources by stall cycles.
2939 if (tryLess(Top.getLatencyStallCycles(TryCand.SU),
2940 Top.getLatencyStallCycles(Cand.SU), TryCand, Cand, Stall))
2941 return;
2942
2943 // Avoid critical resource consumption and balance the schedule.
2944 if (tryLess(TryCand.ResDelta.CritResources, Cand.ResDelta.CritResources,
2945 TryCand, Cand, ResourceReduce))
2946 return;
2947 if (tryGreater(TryCand.ResDelta.DemandedResources,
2948 Cand.ResDelta.DemandedResources,
2949 TryCand, Cand, ResourceDemand))
2950 return;
2951
2952 // Avoid serializing long latency dependence chains.
2953 if (Cand.Policy.ReduceLatency && tryLatency(TryCand, Cand, Top)) {
2954 return;
2955 }
2956
2957 // Fall through to original instruction order.
2958 if (TryCand.SU->NodeNum < Cand.SU->NodeNum)
2959 TryCand.Reason = NodeOrder;
2960 }
2961
pickNodeFromQueue(SchedCandidate & Cand)2962 void PostGenericScheduler::pickNodeFromQueue(SchedCandidate &Cand) {
2963 ReadyQueue &Q = Top.Available;
2964
2965 DEBUG(Q.dump());
2966
2967 for (ReadyQueue::iterator I = Q.begin(), E = Q.end(); I != E; ++I) {
2968 SchedCandidate TryCand(Cand.Policy);
2969 TryCand.SU = *I;
2970 TryCand.initResourceDelta(DAG, SchedModel);
2971 tryCandidate(Cand, TryCand);
2972 if (TryCand.Reason != NoCand) {
2973 Cand.setBest(TryCand);
2974 DEBUG(traceCandidate(Cand));
2975 }
2976 }
2977 }
2978
2979 /// Pick the next node to schedule.
pickNode(bool & IsTopNode)2980 SUnit *PostGenericScheduler::pickNode(bool &IsTopNode) {
2981 if (DAG->top() == DAG->bottom()) {
2982 assert(Top.Available.empty() && Top.Pending.empty() && "ReadyQ garbage");
2983 return nullptr;
2984 }
2985 SUnit *SU;
2986 do {
2987 SU = Top.pickOnlyChoice();
2988 if (!SU) {
2989 CandPolicy NoPolicy;
2990 SchedCandidate TopCand(NoPolicy);
2991 // Set the top-down policy based on the state of the current top zone and
2992 // the instructions outside the zone, including the bottom zone.
2993 setPolicy(TopCand.Policy, /*IsPostRA=*/true, Top, nullptr);
2994 pickNodeFromQueue(TopCand);
2995 assert(TopCand.Reason != NoCand && "failed to find a candidate");
2996 tracePick(TopCand, true);
2997 SU = TopCand.SU;
2998 }
2999 } while (SU->isScheduled);
3000
3001 IsTopNode = true;
3002 Top.removeReady(SU);
3003
3004 DEBUG(dbgs() << "Scheduling SU(" << SU->NodeNum << ") " << *SU->getInstr());
3005 return SU;
3006 }
3007
3008 /// Called after ScheduleDAGMI has scheduled an instruction and updated
3009 /// scheduled/remaining flags in the DAG nodes.
schedNode(SUnit * SU,bool IsTopNode)3010 void PostGenericScheduler::schedNode(SUnit *SU, bool IsTopNode) {
3011 SU->TopReadyCycle = std::max(SU->TopReadyCycle, Top.getCurrCycle());
3012 Top.bumpNode(SU);
3013 }
3014
3015 /// Create a generic scheduler with no vreg liveness or DAG mutation passes.
createGenericSchedPostRA(MachineSchedContext * C)3016 static ScheduleDAGInstrs *createGenericSchedPostRA(MachineSchedContext *C) {
3017 return new ScheduleDAGMI(C, make_unique<PostGenericScheduler>(C), /*IsPostRA=*/true);
3018 }
3019
3020 //===----------------------------------------------------------------------===//
3021 // ILP Scheduler. Currently for experimental analysis of heuristics.
3022 //===----------------------------------------------------------------------===//
3023
3024 namespace {
3025 /// \brief Order nodes by the ILP metric.
3026 struct ILPOrder {
3027 const SchedDFSResult *DFSResult;
3028 const BitVector *ScheduledTrees;
3029 bool MaximizeILP;
3030
ILPOrder__anonbcb081730511::ILPOrder3031 ILPOrder(bool MaxILP)
3032 : DFSResult(nullptr), ScheduledTrees(nullptr), MaximizeILP(MaxILP) {}
3033
3034 /// \brief Apply a less-than relation on node priority.
3035 ///
3036 /// (Return true if A comes after B in the Q.)
operator ()__anonbcb081730511::ILPOrder3037 bool operator()(const SUnit *A, const SUnit *B) const {
3038 unsigned SchedTreeA = DFSResult->getSubtreeID(A);
3039 unsigned SchedTreeB = DFSResult->getSubtreeID(B);
3040 if (SchedTreeA != SchedTreeB) {
3041 // Unscheduled trees have lower priority.
3042 if (ScheduledTrees->test(SchedTreeA) != ScheduledTrees->test(SchedTreeB))
3043 return ScheduledTrees->test(SchedTreeB);
3044
3045 // Trees with shallower connections have have lower priority.
3046 if (DFSResult->getSubtreeLevel(SchedTreeA)
3047 != DFSResult->getSubtreeLevel(SchedTreeB)) {
3048 return DFSResult->getSubtreeLevel(SchedTreeA)
3049 < DFSResult->getSubtreeLevel(SchedTreeB);
3050 }
3051 }
3052 if (MaximizeILP)
3053 return DFSResult->getILP(A) < DFSResult->getILP(B);
3054 else
3055 return DFSResult->getILP(A) > DFSResult->getILP(B);
3056 }
3057 };
3058
3059 /// \brief Schedule based on the ILP metric.
3060 class ILPScheduler : public MachineSchedStrategy {
3061 ScheduleDAGMILive *DAG;
3062 ILPOrder Cmp;
3063
3064 std::vector<SUnit*> ReadyQ;
3065 public:
ILPScheduler(bool MaximizeILP)3066 ILPScheduler(bool MaximizeILP): DAG(nullptr), Cmp(MaximizeILP) {}
3067
initialize(ScheduleDAGMI * dag)3068 void initialize(ScheduleDAGMI *dag) override {
3069 assert(dag->hasVRegLiveness() && "ILPScheduler needs vreg liveness");
3070 DAG = static_cast<ScheduleDAGMILive*>(dag);
3071 DAG->computeDFSResult();
3072 Cmp.DFSResult = DAG->getDFSResult();
3073 Cmp.ScheduledTrees = &DAG->getScheduledTrees();
3074 ReadyQ.clear();
3075 }
3076
registerRoots()3077 void registerRoots() override {
3078 // Restore the heap in ReadyQ with the updated DFS results.
3079 std::make_heap(ReadyQ.begin(), ReadyQ.end(), Cmp);
3080 }
3081
3082 /// Implement MachineSchedStrategy interface.
3083 /// -----------------------------------------
3084
3085 /// Callback to select the highest priority node from the ready Q.
pickNode(bool & IsTopNode)3086 SUnit *pickNode(bool &IsTopNode) override {
3087 if (ReadyQ.empty()) return nullptr;
3088 std::pop_heap(ReadyQ.begin(), ReadyQ.end(), Cmp);
3089 SUnit *SU = ReadyQ.back();
3090 ReadyQ.pop_back();
3091 IsTopNode = false;
3092 DEBUG(dbgs() << "Pick node " << "SU(" << SU->NodeNum << ") "
3093 << " ILP: " << DAG->getDFSResult()->getILP(SU)
3094 << " Tree: " << DAG->getDFSResult()->getSubtreeID(SU) << " @"
3095 << DAG->getDFSResult()->getSubtreeLevel(
3096 DAG->getDFSResult()->getSubtreeID(SU)) << '\n'
3097 << "Scheduling " << *SU->getInstr());
3098 return SU;
3099 }
3100
3101 /// \brief Scheduler callback to notify that a new subtree is scheduled.
scheduleTree(unsigned SubtreeID)3102 void scheduleTree(unsigned SubtreeID) override {
3103 std::make_heap(ReadyQ.begin(), ReadyQ.end(), Cmp);
3104 }
3105
3106 /// Callback after a node is scheduled. Mark a newly scheduled tree, notify
3107 /// DFSResults, and resort the priority Q.
schedNode(SUnit * SU,bool IsTopNode)3108 void schedNode(SUnit *SU, bool IsTopNode) override {
3109 assert(!IsTopNode && "SchedDFSResult needs bottom-up");
3110 }
3111
releaseTopNode(SUnit *)3112 void releaseTopNode(SUnit *) override { /*only called for top roots*/ }
3113
releaseBottomNode(SUnit * SU)3114 void releaseBottomNode(SUnit *SU) override {
3115 ReadyQ.push_back(SU);
3116 std::push_heap(ReadyQ.begin(), ReadyQ.end(), Cmp);
3117 }
3118 };
3119 } // namespace
3120
createILPMaxScheduler(MachineSchedContext * C)3121 static ScheduleDAGInstrs *createILPMaxScheduler(MachineSchedContext *C) {
3122 return new ScheduleDAGMILive(C, make_unique<ILPScheduler>(true));
3123 }
createILPMinScheduler(MachineSchedContext * C)3124 static ScheduleDAGInstrs *createILPMinScheduler(MachineSchedContext *C) {
3125 return new ScheduleDAGMILive(C, make_unique<ILPScheduler>(false));
3126 }
3127 static MachineSchedRegistry ILPMaxRegistry(
3128 "ilpmax", "Schedule bottom-up for max ILP", createILPMaxScheduler);
3129 static MachineSchedRegistry ILPMinRegistry(
3130 "ilpmin", "Schedule bottom-up for min ILP", createILPMinScheduler);
3131
3132 //===----------------------------------------------------------------------===//
3133 // Machine Instruction Shuffler for Correctness Testing
3134 //===----------------------------------------------------------------------===//
3135
3136 #ifndef NDEBUG
3137 namespace {
3138 /// Apply a less-than relation on the node order, which corresponds to the
3139 /// instruction order prior to scheduling. IsReverse implements greater-than.
3140 template<bool IsReverse>
3141 struct SUnitOrder {
operator ()__anonbcb081730611::SUnitOrder3142 bool operator()(SUnit *A, SUnit *B) const {
3143 if (IsReverse)
3144 return A->NodeNum > B->NodeNum;
3145 else
3146 return A->NodeNum < B->NodeNum;
3147 }
3148 };
3149
3150 /// Reorder instructions as much as possible.
3151 class InstructionShuffler : public MachineSchedStrategy {
3152 bool IsAlternating;
3153 bool IsTopDown;
3154
3155 // Using a less-than relation (SUnitOrder<false>) for the TopQ priority
3156 // gives nodes with a higher number higher priority causing the latest
3157 // instructions to be scheduled first.
3158 PriorityQueue<SUnit*, std::vector<SUnit*>, SUnitOrder<false> >
3159 TopQ;
3160 // When scheduling bottom-up, use greater-than as the queue priority.
3161 PriorityQueue<SUnit*, std::vector<SUnit*>, SUnitOrder<true> >
3162 BottomQ;
3163 public:
InstructionShuffler(bool alternate,bool topdown)3164 InstructionShuffler(bool alternate, bool topdown)
3165 : IsAlternating(alternate), IsTopDown(topdown) {}
3166
initialize(ScheduleDAGMI *)3167 void initialize(ScheduleDAGMI*) override {
3168 TopQ.clear();
3169 BottomQ.clear();
3170 }
3171
3172 /// Implement MachineSchedStrategy interface.
3173 /// -----------------------------------------
3174
pickNode(bool & IsTopNode)3175 SUnit *pickNode(bool &IsTopNode) override {
3176 SUnit *SU;
3177 if (IsTopDown) {
3178 do {
3179 if (TopQ.empty()) return nullptr;
3180 SU = TopQ.top();
3181 TopQ.pop();
3182 } while (SU->isScheduled);
3183 IsTopNode = true;
3184 }
3185 else {
3186 do {
3187 if (BottomQ.empty()) return nullptr;
3188 SU = BottomQ.top();
3189 BottomQ.pop();
3190 } while (SU->isScheduled);
3191 IsTopNode = false;
3192 }
3193 if (IsAlternating)
3194 IsTopDown = !IsTopDown;
3195 return SU;
3196 }
3197
schedNode(SUnit * SU,bool IsTopNode)3198 void schedNode(SUnit *SU, bool IsTopNode) override {}
3199
releaseTopNode(SUnit * SU)3200 void releaseTopNode(SUnit *SU) override {
3201 TopQ.push(SU);
3202 }
releaseBottomNode(SUnit * SU)3203 void releaseBottomNode(SUnit *SU) override {
3204 BottomQ.push(SU);
3205 }
3206 };
3207 } // namespace
3208
createInstructionShuffler(MachineSchedContext * C)3209 static ScheduleDAGInstrs *createInstructionShuffler(MachineSchedContext *C) {
3210 bool Alternate = !ForceTopDown && !ForceBottomUp;
3211 bool TopDown = !ForceBottomUp;
3212 assert((TopDown || !ForceTopDown) &&
3213 "-misched-topdown incompatible with -misched-bottomup");
3214 return new ScheduleDAGMILive(C, make_unique<InstructionShuffler>(Alternate, TopDown));
3215 }
3216 static MachineSchedRegistry ShufflerRegistry(
3217 "shuffle", "Shuffle machine instructions alternating directions",
3218 createInstructionShuffler);
3219 #endif // !NDEBUG
3220
3221 //===----------------------------------------------------------------------===//
3222 // GraphWriter support for ScheduleDAGMILive.
3223 //===----------------------------------------------------------------------===//
3224
3225 #ifndef NDEBUG
3226 namespace llvm {
3227
3228 template<> struct GraphTraits<
3229 ScheduleDAGMI*> : public GraphTraits<ScheduleDAG*> {};
3230
3231 template<>
3232 struct DOTGraphTraits<ScheduleDAGMI*> : public DefaultDOTGraphTraits {
3233
DOTGraphTraitsllvm::DOTGraphTraits3234 DOTGraphTraits (bool isSimple=false) : DefaultDOTGraphTraits(isSimple) {}
3235
getGraphNamellvm::DOTGraphTraits3236 static std::string getGraphName(const ScheduleDAG *G) {
3237 return G->MF.getName();
3238 }
3239
renderGraphFromBottomUpllvm::DOTGraphTraits3240 static bool renderGraphFromBottomUp() {
3241 return true;
3242 }
3243
isNodeHiddenllvm::DOTGraphTraits3244 static bool isNodeHidden(const SUnit *Node) {
3245 return (Node->Preds.size() > 10 || Node->Succs.size() > 10);
3246 }
3247
hasNodeAddressLabelllvm::DOTGraphTraits3248 static bool hasNodeAddressLabel(const SUnit *Node,
3249 const ScheduleDAG *Graph) {
3250 return false;
3251 }
3252
3253 /// If you want to override the dot attributes printed for a particular
3254 /// edge, override this method.
getEdgeAttributesllvm::DOTGraphTraits3255 static std::string getEdgeAttributes(const SUnit *Node,
3256 SUnitIterator EI,
3257 const ScheduleDAG *Graph) {
3258 if (EI.isArtificialDep())
3259 return "color=cyan,style=dashed";
3260 if (EI.isCtrlDep())
3261 return "color=blue,style=dashed";
3262 return "";
3263 }
3264
getNodeLabelllvm::DOTGraphTraits3265 static std::string getNodeLabel(const SUnit *SU, const ScheduleDAG *G) {
3266 std::string Str;
3267 raw_string_ostream SS(Str);
3268 const ScheduleDAGMI *DAG = static_cast<const ScheduleDAGMI*>(G);
3269 const SchedDFSResult *DFS = DAG->hasVRegLiveness() ?
3270 static_cast<const ScheduleDAGMILive*>(G)->getDFSResult() : nullptr;
3271 SS << "SU:" << SU->NodeNum;
3272 if (DFS)
3273 SS << " I:" << DFS->getNumInstrs(SU);
3274 return SS.str();
3275 }
getNodeDescriptionllvm::DOTGraphTraits3276 static std::string getNodeDescription(const SUnit *SU, const ScheduleDAG *G) {
3277 return G->getGraphNodeLabel(SU);
3278 }
3279
getNodeAttributesllvm::DOTGraphTraits3280 static std::string getNodeAttributes(const SUnit *N, const ScheduleDAG *G) {
3281 std::string Str("shape=Mrecord");
3282 const ScheduleDAGMI *DAG = static_cast<const ScheduleDAGMI*>(G);
3283 const SchedDFSResult *DFS = DAG->hasVRegLiveness() ?
3284 static_cast<const ScheduleDAGMILive*>(G)->getDFSResult() : nullptr;
3285 if (DFS) {
3286 Str += ",style=filled,fillcolor=\"#";
3287 Str += DOT::getColorString(DFS->getSubtreeID(N));
3288 Str += '"';
3289 }
3290 return Str;
3291 }
3292 };
3293 } // namespace llvm
3294 #endif // NDEBUG
3295
3296 /// viewGraph - Pop up a ghostview window with the reachable parts of the DAG
3297 /// rendered using 'dot'.
3298 ///
viewGraph(const Twine & Name,const Twine & Title)3299 void ScheduleDAGMI::viewGraph(const Twine &Name, const Twine &Title) {
3300 #ifndef NDEBUG
3301 ViewGraph(this, Name, false, Title);
3302 #else
3303 errs() << "ScheduleDAGMI::viewGraph is only available in debug builds on "
3304 << "systems with Graphviz or gv!\n";
3305 #endif // NDEBUG
3306 }
3307
3308 /// Out-of-line implementation with no arguments is handy for gdb.
viewGraph()3309 void ScheduleDAGMI::viewGraph() {
3310 viewGraph(getDAGName(), "Scheduling-Units Graph for " + getDAGName());
3311 }
3312