blob: 1db6f5ee61db4d78a1a7a48c7c709d56ae0023a1 [file] [log] [blame]
Sergei Larin3e590402012-09-04 14:49:56 +00001//===- HexagonMachineScheduler.cpp - MI Scheduler for Hexagon -------------===//
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#define DEBUG_TYPE "misched"
16
17#include "HexagonMachineScheduler.h"
18
19#include <queue>
20
21using namespace llvm;
22
Sergei Larinc6a66602012-09-14 15:07:59 +000023/// Platform specific modifications to DAG.
24void VLIWMachineScheduler::postprocessDAG() {
25 SUnit* LastSequentialCall = NULL;
26 // Currently we only catch the situation when compare gets scheduled
27 // before preceding call.
28 for (unsigned su = 0, e = SUnits.size(); su != e; ++su) {
29 // Remember the call.
30 if (SUnits[su].getInstr()->isCall())
31 LastSequentialCall = &(SUnits[su]);
32 // Look for a compare that defines a predicate.
33 else if (SUnits[su].getInstr()->isCompare() && LastSequentialCall)
34 SUnits[su].addPred(SDep(LastSequentialCall, SDep::Order, 0, /*Reg=*/0,
35 false));
36 }
37}
38
Sergei Larin3e590402012-09-04 14:49:56 +000039/// Check if scheduling of this SU is possible
40/// in the current packet.
41/// It is _not_ precise (statefull), it is more like
42/// another heuristic. Many corner cases are figured
43/// empirically.
44bool VLIWResourceModel::isResourceAvailable(SUnit *SU) {
45 if (!SU || !SU->getInstr())
46 return false;
47
48 // First see if the pipeline could receive this instruction
49 // in the current cycle.
50 switch (SU->getInstr()->getOpcode()) {
51 default:
52 if (!ResourcesModel->canReserveResources(SU->getInstr()))
53 return false;
54 case TargetOpcode::EXTRACT_SUBREG:
55 case TargetOpcode::INSERT_SUBREG:
56 case TargetOpcode::SUBREG_TO_REG:
57 case TargetOpcode::REG_SEQUENCE:
58 case TargetOpcode::IMPLICIT_DEF:
59 case TargetOpcode::COPY:
60 case TargetOpcode::INLINEASM:
61 break;
62 }
63
64 // Now see if there are no other dependencies to instructions already
65 // in the packet.
66 for (unsigned i = 0, e = Packet.size(); i != e; ++i) {
67 if (Packet[i]->Succs.size() == 0)
68 continue;
69 for (SUnit::const_succ_iterator I = Packet[i]->Succs.begin(),
70 E = Packet[i]->Succs.end(); I != E; ++I) {
71 // Since we do not add pseudos to packets, might as well
72 // ignore order dependencies.
73 if (I->isCtrl())
74 continue;
75
76 if (I->getSUnit() == SU)
77 return false;
78 }
79 }
80 return true;
81}
82
83/// Keep track of available resources.
Sergei Larin7ae51be2012-09-10 17:31:34 +000084bool VLIWResourceModel::reserveResources(SUnit *SU) {
85 bool startNewCycle = false;
Sergei Larinc6a66602012-09-14 15:07:59 +000086 // Artificially reset state.
87 if (!SU) {
88 ResourcesModel->clearResources();
89 Packet.clear();
90 TotalPackets++;
91 return false;
92 }
Sergei Larin3e590402012-09-04 14:49:56 +000093 // If this SU does not fit in the packet
94 // start a new one.
95 if (!isResourceAvailable(SU)) {
96 ResourcesModel->clearResources();
97 Packet.clear();
98 TotalPackets++;
Sergei Larin7ae51be2012-09-10 17:31:34 +000099 startNewCycle = true;
Sergei Larin3e590402012-09-04 14:49:56 +0000100 }
101
102 switch (SU->getInstr()->getOpcode()) {
103 default:
104 ResourcesModel->reserveResources(SU->getInstr());
105 break;
106 case TargetOpcode::EXTRACT_SUBREG:
107 case TargetOpcode::INSERT_SUBREG:
108 case TargetOpcode::SUBREG_TO_REG:
109 case TargetOpcode::REG_SEQUENCE:
110 case TargetOpcode::IMPLICIT_DEF:
111 case TargetOpcode::KILL:
112 case TargetOpcode::PROLOG_LABEL:
113 case TargetOpcode::EH_LABEL:
114 case TargetOpcode::COPY:
115 case TargetOpcode::INLINEASM:
116 break;
117 }
118 Packet.push_back(SU);
119
120#ifndef NDEBUG
121 DEBUG(dbgs() << "Packet[" << TotalPackets << "]:\n");
122 for (unsigned i = 0, e = Packet.size(); i != e; ++i) {
123 DEBUG(dbgs() << "\t[" << i << "] SU(");
Sergei Larin7ae51be2012-09-10 17:31:34 +0000124 DEBUG(dbgs() << Packet[i]->NodeNum << ")\t");
125 DEBUG(Packet[i]->getInstr()->dump());
Sergei Larin3e590402012-09-04 14:49:56 +0000126 }
127#endif
128
129 // If packet is now full, reset the state so in the next cycle
130 // we start fresh.
131 if (Packet.size() >= InstrItins->SchedModel->IssueWidth) {
132 ResourcesModel->clearResources();
133 Packet.clear();
134 TotalPackets++;
Sergei Larin7ae51be2012-09-10 17:31:34 +0000135 startNewCycle = true;
Sergei Larin3e590402012-09-04 14:49:56 +0000136 }
Sergei Larin7ae51be2012-09-10 17:31:34 +0000137
138 return startNewCycle;
Sergei Larin3e590402012-09-04 14:49:56 +0000139}
140
Sergei Larin3e590402012-09-04 14:49:56 +0000141/// schedule - Called back from MachineScheduler::runOnMachineFunction
142/// after setting up the current scheduling region. [RegionBegin, RegionEnd)
143/// only includes instructions that have DAG nodes, not scheduling boundaries.
144void VLIWMachineScheduler::schedule() {
145 DEBUG(dbgs()
146 << "********** MI Converging Scheduling VLIW BB#" << BB->getNumber()
147 << " " << BB->getName()
148 << " in_func " << BB->getParent()->getFunction()->getName()
Benjamin Kramere5c4fe52012-09-14 12:19:58 +0000149 << " at loop depth " << MLI.getLoopDepth(BB)
Sergei Larin3e590402012-09-04 14:49:56 +0000150 << " \n");
151
Andrew Trick78e5efe2012-09-11 00:39:15 +0000152 buildDAGWithRegPressure();
Sergei Larin3e590402012-09-04 14:49:56 +0000153
Sergei Larinc6a66602012-09-14 15:07:59 +0000154 // Postprocess the DAG to add platform specific artificial dependencies.
155 postprocessDAG();
156
Sergei Larin7ae51be2012-09-10 17:31:34 +0000157 // To view Height/Depth correctly, they should be accessed at least once.
158 DEBUG(unsigned maxH = 0;
159 for (unsigned su = 0, e = SUnits.size(); su != e; ++su)
160 if (SUnits[su].getHeight() > maxH)
161 maxH = SUnits[su].getHeight();
162 dbgs() << "Max Height " << maxH << "\n";);
163 DEBUG(unsigned maxD = 0;
164 for (unsigned su = 0, e = SUnits.size(); su != e; ++su)
165 if (SUnits[su].getDepth() > maxD)
166 maxD = SUnits[su].getDepth();
167 dbgs() << "Max Depth " << maxD << "\n";);
Sergei Larin3e590402012-09-04 14:49:56 +0000168 DEBUG(for (unsigned su = 0, e = SUnits.size(); su != e; ++su)
169 SUnits[su].dumpAll(this));
170
Andrew Trick78e5efe2012-09-11 00:39:15 +0000171 initQueues();
Sergei Larin3e590402012-09-04 14:49:56 +0000172
Sergei Larin3e590402012-09-04 14:49:56 +0000173 bool IsTopNode = false;
174 while (SUnit *SU = SchedImpl->pickNode(IsTopNode)) {
175 if (!checkSchedLimit())
176 break;
177
Andrew Trick78e5efe2012-09-11 00:39:15 +0000178 scheduleMI(SU, IsTopNode);
Sergei Larin3e590402012-09-04 14:49:56 +0000179
Andrew Trick78e5efe2012-09-11 00:39:15 +0000180 updateQueues(SU, IsTopNode);
Sergei Larin3e590402012-09-04 14:49:56 +0000181 }
182 assert(CurrentTop == CurrentBottom && "Nonempty unscheduled zone.");
183
Sergei Larin3e590402012-09-04 14:49:56 +0000184 placeDebugValues();
185}
186
Andrew Trick78e5efe2012-09-11 00:39:15 +0000187void ConvergingVLIWScheduler::initialize(ScheduleDAGMI *dag) {
188 DAG = static_cast<VLIWMachineScheduler*>(dag);
Sergei Larin3e590402012-09-04 14:49:56 +0000189 TRI = DAG->TRI;
Andrew Trick78e5efe2012-09-11 00:39:15 +0000190 Top.DAG = DAG;
191 Bot.DAG = DAG;
Sergei Larin3e590402012-09-04 14:49:56 +0000192
193 // Initialize the HazardRecognizers.
194 const TargetMachine &TM = DAG->MF.getTarget();
195 const InstrItineraryData *Itin = TM.getInstrItineraryData();
196 Top.HazardRec = TM.getInstrInfo()->CreateTargetMIHazardRecognizer(Itin, DAG);
197 Bot.HazardRec = TM.getInstrInfo()->CreateTargetMIHazardRecognizer(Itin, DAG);
198
Sergei Larin7ae51be2012-09-10 17:31:34 +0000199 Top.ResourceModel = new VLIWResourceModel(TM);
200 Bot.ResourceModel = new VLIWResourceModel(TM);
201
Andrew Trick78e5efe2012-09-11 00:39:15 +0000202 assert((!llvm::ForceTopDown || !llvm::ForceBottomUp) &&
Sergei Larin3e590402012-09-04 14:49:56 +0000203 "-misched-topdown incompatible with -misched-bottomup");
204}
205
206void ConvergingVLIWScheduler::releaseTopNode(SUnit *SU) {
207 if (SU->isScheduled)
208 return;
209
210 for (SUnit::succ_iterator I = SU->Preds.begin(), E = SU->Preds.end();
211 I != E; ++I) {
212 unsigned PredReadyCycle = I->getSUnit()->TopReadyCycle;
213 unsigned MinLatency = I->getMinLatency();
214#ifndef NDEBUG
215 Top.MaxMinLatency = std::max(MinLatency, Top.MaxMinLatency);
216#endif
217 if (SU->TopReadyCycle < PredReadyCycle + MinLatency)
218 SU->TopReadyCycle = PredReadyCycle + MinLatency;
219 }
220 Top.releaseNode(SU, SU->TopReadyCycle);
221}
222
223void ConvergingVLIWScheduler::releaseBottomNode(SUnit *SU) {
224 if (SU->isScheduled)
225 return;
226
227 assert(SU->getInstr() && "Scheduled SUnit must have instr");
228
229 for (SUnit::succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
230 I != E; ++I) {
231 unsigned SuccReadyCycle = I->getSUnit()->BotReadyCycle;
232 unsigned MinLatency = I->getMinLatency();
233#ifndef NDEBUG
234 Bot.MaxMinLatency = std::max(MinLatency, Bot.MaxMinLatency);
235#endif
236 if (SU->BotReadyCycle < SuccReadyCycle + MinLatency)
237 SU->BotReadyCycle = SuccReadyCycle + MinLatency;
238 }
239 Bot.releaseNode(SU, SU->BotReadyCycle);
240}
241
242/// Does this SU have a hazard within the current instruction group.
243///
244/// The scheduler supports two modes of hazard recognition. The first is the
245/// ScheduleHazardRecognizer API. It is a fully general hazard recognizer that
246/// supports highly complicated in-order reservation tables
247/// (ScoreboardHazardRecognizer) and arbitrary target-specific logic.
248///
249/// The second is a streamlined mechanism that checks for hazards based on
250/// simple counters that the scheduler itself maintains. It explicitly checks
251/// for instruction dispatch limitations, including the number of micro-ops that
252/// can dispatch per cycle.
253///
254/// TODO: Also check whether the SU must start a new group.
255bool ConvergingVLIWScheduler::SchedBoundary::checkHazard(SUnit *SU) {
256 if (HazardRec->isEnabled())
257 return HazardRec->getHazardType(SU) != ScheduleHazardRecognizer::NoHazard;
258
259 if (IssueCount + DAG->getNumMicroOps(SU->getInstr()) > DAG->getIssueWidth())
260 return true;
261
262 return false;
263}
264
265void ConvergingVLIWScheduler::SchedBoundary::releaseNode(SUnit *SU,
266 unsigned ReadyCycle) {
267 if (ReadyCycle < MinReadyCycle)
268 MinReadyCycle = ReadyCycle;
269
270 // Check for interlocks first. For the purpose of other heuristics, an
271 // instruction that cannot issue appears as if it's not in the ReadyQueue.
272 if (ReadyCycle > CurrCycle || checkHazard(SU))
273
274 Pending.push(SU);
275 else
276 Available.push(SU);
277}
278
279/// Move the boundary of scheduled code by one cycle.
280void ConvergingVLIWScheduler::SchedBoundary::bumpCycle() {
281 unsigned Width = DAG->getIssueWidth();
282 IssueCount = (IssueCount <= Width) ? 0 : IssueCount - Width;
283
284 assert(MinReadyCycle < UINT_MAX && "MinReadyCycle uninitialized");
285 unsigned NextCycle = std::max(CurrCycle + 1, MinReadyCycle);
286
287 if (!HazardRec->isEnabled()) {
288 // Bypass HazardRec virtual calls.
289 CurrCycle = NextCycle;
Sergei Larin7ae51be2012-09-10 17:31:34 +0000290 } else {
Sergei Larin3e590402012-09-04 14:49:56 +0000291 // Bypass getHazardType calls in case of long latency.
292 for (; CurrCycle != NextCycle; ++CurrCycle) {
293 if (isTop())
294 HazardRec->AdvanceCycle();
295 else
296 HazardRec->RecedeCycle();
297 }
298 }
299 CheckPending = true;
300
301 DEBUG(dbgs() << "*** " << Available.getName() << " cycle "
302 << CurrCycle << '\n');
303}
304
305/// Move the boundary of scheduled code by one SUnit.
306void ConvergingVLIWScheduler::SchedBoundary::bumpNode(SUnit *SU) {
Sergei Larin7ae51be2012-09-10 17:31:34 +0000307 bool startNewCycle = false;
Sergei Larin3e590402012-09-04 14:49:56 +0000308
309 // Update the reservation table.
310 if (HazardRec->isEnabled()) {
311 if (!isTop() && SU->isCall) {
312 // Calls are scheduled with their preceding instructions. For bottom-up
313 // scheduling, clear the pipeline state before emitting.
314 HazardRec->Reset();
315 }
316 HazardRec->EmitInstruction(SU);
317 }
Sergei Larin7ae51be2012-09-10 17:31:34 +0000318
319 // Update DFA model.
320 startNewCycle = ResourceModel->reserveResources(SU);
321
Sergei Larin3e590402012-09-04 14:49:56 +0000322 // Check the instruction group dispatch limit.
323 // TODO: Check if this SU must end a dispatch group.
324 IssueCount += DAG->getNumMicroOps(SU->getInstr());
Sergei Larin7ae51be2012-09-10 17:31:34 +0000325 if (startNewCycle) {
Sergei Larin3e590402012-09-04 14:49:56 +0000326 DEBUG(dbgs() << "*** Max instrs at cycle " << CurrCycle << '\n');
327 bumpCycle();
328 }
Sergei Larin7ae51be2012-09-10 17:31:34 +0000329 else
330 DEBUG(dbgs() << "*** IssueCount " << IssueCount
331 << " at cycle " << CurrCycle << '\n');
Sergei Larin3e590402012-09-04 14:49:56 +0000332}
333
334/// Release pending ready nodes in to the available queue. This makes them
335/// visible to heuristics.
336void ConvergingVLIWScheduler::SchedBoundary::releasePending() {
337 // If the available queue is empty, it is safe to reset MinReadyCycle.
338 if (Available.empty())
339 MinReadyCycle = UINT_MAX;
340
341 // Check to see if any of the pending instructions are ready to issue. If
342 // so, add them to the available queue.
343 for (unsigned i = 0, e = Pending.size(); i != e; ++i) {
344 SUnit *SU = *(Pending.begin()+i);
345 unsigned ReadyCycle = isTop() ? SU->TopReadyCycle : SU->BotReadyCycle;
346
347 if (ReadyCycle < MinReadyCycle)
348 MinReadyCycle = ReadyCycle;
349
350 if (ReadyCycle > CurrCycle)
351 continue;
352
353 if (checkHazard(SU))
354 continue;
355
356 Available.push(SU);
357 Pending.remove(Pending.begin()+i);
358 --i; --e;
359 }
360 CheckPending = false;
361}
362
363/// Remove SU from the ready set for this boundary.
364void ConvergingVLIWScheduler::SchedBoundary::removeReady(SUnit *SU) {
365 if (Available.isInQueue(SU))
366 Available.remove(Available.find(SU));
367 else {
368 assert(Pending.isInQueue(SU) && "bad ready count");
369 Pending.remove(Pending.find(SU));
370 }
371}
372
373/// If this queue only has one ready candidate, return it. As a side effect,
374/// advance the cycle until at least one node is ready. If multiple instructions
375/// are ready, return NULL.
376SUnit *ConvergingVLIWScheduler::SchedBoundary::pickOnlyChoice() {
377 if (CheckPending)
378 releasePending();
379
380 for (unsigned i = 0; Available.empty(); ++i) {
381 assert(i <= (HazardRec->getMaxLookAhead() + MaxMinLatency) &&
382 "permanent hazard"); (void)i;
Sergei Larinc6a66602012-09-14 15:07:59 +0000383 ResourceModel->reserveResources(0);
Sergei Larin3e590402012-09-04 14:49:56 +0000384 bumpCycle();
385 releasePending();
386 }
387 if (Available.size() == 1)
388 return *Available.begin();
389 return NULL;
390}
391
392#ifndef NDEBUG
Sergei Larin7ae51be2012-09-10 17:31:34 +0000393void ConvergingVLIWScheduler::traceCandidate(const char *Label,
394 const ReadyQueue &Q,
395 SUnit *SU, PressureElement P) {
Sergei Larin3e590402012-09-04 14:49:56 +0000396 dbgs() << Label << " " << Q.getName() << " ";
397 if (P.isValid())
398 dbgs() << TRI->getRegPressureSetName(P.PSetID) << ":" << P.UnitIncrease
399 << " ";
400 else
401 dbgs() << " ";
402 SU->dump(DAG);
403}
404#endif
405
Sergei Larin7ae51be2012-09-10 17:31:34 +0000406/// getSingleUnscheduledPred - If there is exactly one unscheduled predecessor
407/// of SU, return it, otherwise return null.
408static SUnit *getSingleUnscheduledPred(SUnit *SU) {
409 SUnit *OnlyAvailablePred = 0;
410 for (SUnit::const_pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
411 I != E; ++I) {
412 SUnit &Pred = *I->getSUnit();
413 if (!Pred.isScheduled) {
414 // We found an available, but not scheduled, predecessor. If it's the
415 // only one we have found, keep track of it... otherwise give up.
416 if (OnlyAvailablePred && OnlyAvailablePred != &Pred)
417 return 0;
418 OnlyAvailablePred = &Pred;
419 }
420 }
421 return OnlyAvailablePred;
422}
423
424/// getSingleUnscheduledSucc - If there is exactly one unscheduled successor
425/// of SU, return it, otherwise return null.
426static SUnit *getSingleUnscheduledSucc(SUnit *SU) {
427 SUnit *OnlyAvailableSucc = 0;
428 for (SUnit::const_succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
429 I != E; ++I) {
430 SUnit &Succ = *I->getSUnit();
431 if (!Succ.isScheduled) {
432 // We found an available, but not scheduled, successor. If it's the
433 // only one we have found, keep track of it... otherwise give up.
434 if (OnlyAvailableSucc && OnlyAvailableSucc != &Succ)
435 return 0;
436 OnlyAvailableSucc = &Succ;
437 }
438 }
439 return OnlyAvailableSucc;
440}
441
Sergei Larin3e590402012-09-04 14:49:56 +0000442// Constants used to denote relative importance of
443// heuristic components for cost computation.
444static const unsigned PriorityOne = 200;
Sergei Larin7ae51be2012-09-10 17:31:34 +0000445static const unsigned PriorityTwo = 100;
Sergei Larin3e590402012-09-04 14:49:56 +0000446static const unsigned PriorityThree = 50;
Sergei Larin7ae51be2012-09-10 17:31:34 +0000447static const unsigned PriorityFour = 20;
Sergei Larin3e590402012-09-04 14:49:56 +0000448static const unsigned ScaleTwo = 10;
449static const unsigned FactorOne = 2;
450
451/// Single point to compute overall scheduling cost.
452/// TODO: More heuristics will be used soon.
453int ConvergingVLIWScheduler::SchedulingCost(ReadyQueue &Q, SUnit *SU,
454 SchedCandidate &Candidate,
455 RegPressureDelta &Delta,
456 bool verbose) {
457 // Initial trivial priority.
458 int ResCount = 1;
459
460 // Do not waste time on a node that is already scheduled.
461 if (!SU || SU->isScheduled)
462 return ResCount;
463
464 // Forced priority is high.
465 if (SU->isScheduleHigh)
466 ResCount += PriorityOne;
467
468 // Critical path first.
Sergei Larin7ae51be2012-09-10 17:31:34 +0000469 if (Q.getID() == TopQID) {
Sergei Larin3e590402012-09-04 14:49:56 +0000470 ResCount += (SU->getHeight() * ScaleTwo);
Sergei Larin7ae51be2012-09-10 17:31:34 +0000471
472 // If resources are available for it, multiply the
473 // chance of scheduling.
474 if (Top.ResourceModel->isResourceAvailable(SU))
475 ResCount <<= FactorOne;
476 } else {
Sergei Larin3e590402012-09-04 14:49:56 +0000477 ResCount += (SU->getDepth() * ScaleTwo);
478
Sergei Larin7ae51be2012-09-10 17:31:34 +0000479 // If resources are available for it, multiply the
480 // chance of scheduling.
481 if (Bot.ResourceModel->isResourceAvailable(SU))
482 ResCount <<= FactorOne;
483 }
484
485 unsigned NumNodesBlocking = 0;
486 if (Q.getID() == TopQID) {
487 // How many SUs does it block from scheduling?
488 // Look at all of the successors of this node.
489 // Count the number of nodes that
490 // this node is the sole unscheduled node for.
491 for (SUnit::const_succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
492 I != E; ++I)
493 if (getSingleUnscheduledPred(I->getSUnit()) == SU)
494 ++NumNodesBlocking;
495 } else {
496 // How many unscheduled predecessors block this node?
497 for (SUnit::const_pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
498 I != E; ++I)
499 if (getSingleUnscheduledSucc(I->getSUnit()) == SU)
500 ++NumNodesBlocking;
501 }
502 ResCount += (NumNodesBlocking * ScaleTwo);
Sergei Larin3e590402012-09-04 14:49:56 +0000503
504 // Factor in reg pressure as a heuristic.
Sergei Larin7ae51be2012-09-10 17:31:34 +0000505 ResCount -= (Delta.Excess.UnitIncrease*PriorityThree);
506 ResCount -= (Delta.CriticalMax.UnitIncrease*PriorityThree);
Sergei Larin3e590402012-09-04 14:49:56 +0000507
508 DEBUG(if (verbose) dbgs() << " Total(" << ResCount << ")");
509
510 return ResCount;
511}
512
513/// Pick the best candidate from the top queue.
514///
515/// TODO: getMaxPressureDelta results can be mostly cached for each SUnit during
516/// DAG building. To adjust for the current scheduling location we need to
517/// maintain the number of vreg uses remaining to be top-scheduled.
518ConvergingVLIWScheduler::CandResult ConvergingVLIWScheduler::
519pickNodeFromQueue(ReadyQueue &Q, const RegPressureTracker &RPTracker,
520 SchedCandidate &Candidate) {
521 DEBUG(Q.dump());
522
523 // getMaxPressureDelta temporarily modifies the tracker.
524 RegPressureTracker &TempTracker = const_cast<RegPressureTracker&>(RPTracker);
525
526 // BestSU remains NULL if no top candidates beat the best existing candidate.
527 CandResult FoundCandidate = NoCand;
528 for (ReadyQueue::iterator I = Q.begin(), E = Q.end(); I != E; ++I) {
529 RegPressureDelta RPDelta;
530 TempTracker.getMaxPressureDelta((*I)->getInstr(), RPDelta,
531 DAG->getRegionCriticalPSets(),
532 DAG->getRegPressure().MaxSetPressure);
533
534 int CurrentCost = SchedulingCost(Q, *I, Candidate, RPDelta, false);
535
536 // Initialize the candidate if needed.
537 if (!Candidate.SU) {
538 Candidate.SU = *I;
539 Candidate.RPDelta = RPDelta;
540 Candidate.SCost = CurrentCost;
541 FoundCandidate = NodeOrder;
542 continue;
543 }
544
Sergei Larin3e590402012-09-04 14:49:56 +0000545 // Best cost.
546 if (CurrentCost > Candidate.SCost) {
547 DEBUG(traceCandidate("CCAND", Q, *I));
548 Candidate.SU = *I;
549 Candidate.RPDelta = RPDelta;
550 Candidate.SCost = CurrentCost;
551 FoundCandidate = BestCost;
552 continue;
553 }
554
555 // Fall through to original instruction order.
556 // Only consider node order if Candidate was chosen from this Q.
557 if (FoundCandidate == NoCand)
558 continue;
559 }
560 return FoundCandidate;
561}
562
563/// Pick the best candidate node from either the top or bottom queue.
564SUnit *ConvergingVLIWScheduler::pickNodeBidrectional(bool &IsTopNode) {
565 // Schedule as far as possible in the direction of no choice. This is most
566 // efficient, but also provides the best heuristics for CriticalPSets.
567 if (SUnit *SU = Bot.pickOnlyChoice()) {
568 IsTopNode = false;
569 return SU;
570 }
571 if (SUnit *SU = Top.pickOnlyChoice()) {
572 IsTopNode = true;
573 return SU;
574 }
575 SchedCandidate BotCand;
576 // Prefer bottom scheduling when heuristics are silent.
577 CandResult BotResult = pickNodeFromQueue(Bot.Available,
578 DAG->getBotRPTracker(), BotCand);
579 assert(BotResult != NoCand && "failed to find the first candidate");
580
581 // If either Q has a single candidate that provides the least increase in
582 // Excess pressure, we can immediately schedule from that Q.
583 //
584 // RegionCriticalPSets summarizes the pressure within the scheduled region and
585 // affects picking from either Q. If scheduling in one direction must
586 // increase pressure for one of the excess PSets, then schedule in that
587 // direction first to provide more freedom in the other direction.
588 if (BotResult == SingleExcess || BotResult == SingleCritical) {
589 IsTopNode = false;
590 return BotCand.SU;
591 }
592 // Check if the top Q has a better candidate.
593 SchedCandidate TopCand;
594 CandResult TopResult = pickNodeFromQueue(Top.Available,
595 DAG->getTopRPTracker(), TopCand);
596 assert(TopResult != NoCand && "failed to find the first candidate");
597
598 if (TopResult == SingleExcess || TopResult == SingleCritical) {
599 IsTopNode = true;
600 return TopCand.SU;
601 }
602 // If either Q has a single candidate that minimizes pressure above the
603 // original region's pressure pick it.
604 if (BotResult == SingleMax) {
605 IsTopNode = false;
606 return BotCand.SU;
607 }
608 if (TopResult == SingleMax) {
609 IsTopNode = true;
610 return TopCand.SU;
611 }
612 if (TopCand.SCost > BotCand.SCost) {
613 IsTopNode = true;
614 return TopCand.SU;
615 }
616 // Otherwise prefer the bottom candidate in node order.
617 IsTopNode = false;
618 return BotCand.SU;
619}
620
621/// Pick the best node to balance the schedule. Implements MachineSchedStrategy.
622SUnit *ConvergingVLIWScheduler::pickNode(bool &IsTopNode) {
623 if (DAG->top() == DAG->bottom()) {
624 assert(Top.Available.empty() && Top.Pending.empty() &&
625 Bot.Available.empty() && Bot.Pending.empty() && "ReadyQ garbage");
626 return NULL;
627 }
628 SUnit *SU;
Andrew Trick78e5efe2012-09-11 00:39:15 +0000629 if (llvm::ForceTopDown) {
Sergei Larin3e590402012-09-04 14:49:56 +0000630 SU = Top.pickOnlyChoice();
631 if (!SU) {
632 SchedCandidate TopCand;
633 CandResult TopResult =
634 pickNodeFromQueue(Top.Available, DAG->getTopRPTracker(), TopCand);
635 assert(TopResult != NoCand && "failed to find the first candidate");
636 (void)TopResult;
637 SU = TopCand.SU;
638 }
639 IsTopNode = true;
Andrew Trick78e5efe2012-09-11 00:39:15 +0000640 } else if (llvm::ForceBottomUp) {
Sergei Larin3e590402012-09-04 14:49:56 +0000641 SU = Bot.pickOnlyChoice();
642 if (!SU) {
643 SchedCandidate BotCand;
644 CandResult BotResult =
645 pickNodeFromQueue(Bot.Available, DAG->getBotRPTracker(), BotCand);
646 assert(BotResult != NoCand && "failed to find the first candidate");
647 (void)BotResult;
648 SU = BotCand.SU;
649 }
650 IsTopNode = false;
651 } else {
652 SU = pickNodeBidrectional(IsTopNode);
653 }
654 if (SU->isTopReady())
655 Top.removeReady(SU);
656 if (SU->isBottomReady())
657 Bot.removeReady(SU);
658
659 DEBUG(dbgs() << "*** " << (IsTopNode ? "Top" : "Bottom")
660 << " Scheduling Instruction in cycle "
661 << (IsTopNode ? Top.CurrCycle : Bot.CurrCycle) << '\n';
662 SU->dump(DAG));
663 return SU;
664}
665
666/// Update the scheduler's state after scheduling a node. This is the same node
Sergei Larin7ae51be2012-09-10 17:31:34 +0000667/// that was just returned by pickNode(). However, VLIWMachineScheduler needs
668/// to update it's state based on the current cycle before MachineSchedStrategy
669/// does.
Sergei Larin3e590402012-09-04 14:49:56 +0000670void ConvergingVLIWScheduler::schedNode(SUnit *SU, bool IsTopNode) {
671 if (IsTopNode) {
672 SU->TopReadyCycle = Top.CurrCycle;
673 Top.bumpNode(SU);
Sergei Larin7ae51be2012-09-10 17:31:34 +0000674 } else {
Sergei Larin3e590402012-09-04 14:49:56 +0000675 SU->BotReadyCycle = Bot.CurrCycle;
676 Bot.bumpNode(SU);
677 }
678}
679