1 //===- GCNIterativeScheduler.cpp ------------------------------------------===// 2 // 3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4 // See https://llvm.org/LICENSE.txt for license information. 5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6 // 7 //===----------------------------------------------------------------------===// 8 /// 9 /// \file 10 /// This file implements the class GCNIterativeScheduler. 11 /// 12 //===----------------------------------------------------------------------===// 13 14 #include "GCNIterativeScheduler.h" 15 #include "GCNSchedStrategy.h" 16 #include "SIMachineFunctionInfo.h" 17 18 using namespace llvm; 19 20 #define DEBUG_TYPE "machine-scheduler" 21 22 namespace llvm { 23 24 std::vector<const SUnit *> makeMinRegSchedule(ArrayRef<const SUnit *> TopRoots, 25 const ScheduleDAG &DAG); 26 27 std::vector<const SUnit*> makeGCNILPScheduler(ArrayRef<const SUnit*> BotRoots, 28 const ScheduleDAG &DAG); 29 } 30 31 // shim accessors for different order containers 32 static inline MachineInstr *getMachineInstr(MachineInstr *MI) { 33 return MI; 34 } 35 static inline MachineInstr *getMachineInstr(const SUnit *SU) { 36 return SU->getInstr(); 37 } 38 static inline MachineInstr *getMachineInstr(const SUnit &SU) { 39 return SU.getInstr(); 40 } 41 42 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) 43 LLVM_DUMP_METHOD 44 static void printRegion(raw_ostream &OS, 45 MachineBasicBlock::iterator Begin, 46 MachineBasicBlock::iterator End, 47 const LiveIntervals *LIS, 48 unsigned MaxInstNum = 49 std::numeric_limits<unsigned>::max()) { 50 auto BB = Begin->getParent(); 51 OS << BB->getParent()->getName() << ":" << printMBBReference(*BB) << ' ' 52 << BB->getName() << ":\n"; 53 auto I = Begin; 54 MaxInstNum = std::max(MaxInstNum, 1u); 55 for (; I != End && MaxInstNum; ++I, --MaxInstNum) { 56 if (!I->isDebugInstr() && LIS) 57 OS << LIS->getInstructionIndex(*I); 58 OS << '\t' << *I; 59 } 60 if (I != End) { 61 OS << "\t...\n"; 62 I = std::prev(End); 63 if (!I->isDebugInstr() && LIS) 64 OS << LIS->getInstructionIndex(*I); 65 OS << '\t' << *I; 66 } 67 if (End != BB->end()) { // print boundary inst if present 68 OS << "----\n"; 69 if (LIS) OS << LIS->getInstructionIndex(*End) << '\t'; 70 OS << *End; 71 } 72 } 73 74 LLVM_DUMP_METHOD 75 static void printLivenessInfo(raw_ostream &OS, 76 MachineBasicBlock::iterator Begin, 77 MachineBasicBlock::iterator End, 78 const LiveIntervals *LIS) { 79 const auto BB = Begin->getParent(); 80 const auto &MRI = BB->getParent()->getRegInfo(); 81 82 const auto LiveIns = getLiveRegsBefore(*Begin, *LIS); 83 OS << "LIn RP: "; 84 getRegPressure(MRI, LiveIns).print(OS); 85 86 const auto BottomMI = End == BB->end() ? std::prev(End) : End; 87 const auto LiveOuts = getLiveRegsAfter(*BottomMI, *LIS); 88 OS << "LOt RP: "; 89 getRegPressure(MRI, LiveOuts).print(OS); 90 } 91 92 LLVM_DUMP_METHOD 93 void GCNIterativeScheduler::printRegions(raw_ostream &OS) const { 94 const auto &ST = MF.getSubtarget<GCNSubtarget>(); 95 for (const auto R : Regions) { 96 OS << "Region to schedule "; 97 printRegion(OS, R->Begin, R->End, LIS, 1); 98 printLivenessInfo(OS, R->Begin, R->End, LIS); 99 OS << "Max RP: "; 100 R->MaxPressure.print(OS, &ST); 101 } 102 } 103 104 LLVM_DUMP_METHOD 105 void GCNIterativeScheduler::printSchedResult(raw_ostream &OS, 106 const Region *R, 107 const GCNRegPressure &RP) const { 108 OS << "\nAfter scheduling "; 109 printRegion(OS, R->Begin, R->End, LIS); 110 printSchedRP(OS, R->MaxPressure, RP); 111 OS << '\n'; 112 } 113 114 LLVM_DUMP_METHOD 115 void GCNIterativeScheduler::printSchedRP(raw_ostream &OS, 116 const GCNRegPressure &Before, 117 const GCNRegPressure &After) const { 118 const auto &ST = MF.getSubtarget<GCNSubtarget>(); 119 OS << "RP before: "; 120 Before.print(OS, &ST); 121 OS << "RP after: "; 122 After.print(OS, &ST); 123 } 124 #endif 125 126 // DAG builder helper 127 class GCNIterativeScheduler::BuildDAG { 128 GCNIterativeScheduler &Sch; 129 SmallVector<SUnit *, 8> TopRoots; 130 131 SmallVector<SUnit*, 8> BotRoots; 132 public: 133 BuildDAG(const Region &R, GCNIterativeScheduler &_Sch) 134 : Sch(_Sch) { 135 auto BB = R.Begin->getParent(); 136 Sch.BaseClass::startBlock(BB); 137 Sch.BaseClass::enterRegion(BB, R.Begin, R.End, R.NumRegionInstrs); 138 139 Sch.buildSchedGraph(Sch.AA, nullptr, nullptr, nullptr, 140 /*TrackLaneMask*/true); 141 Sch.Topo.InitDAGTopologicalSorting(); 142 Sch.findRootsAndBiasEdges(TopRoots, BotRoots); 143 } 144 145 ~BuildDAG() { 146 Sch.BaseClass::exitRegion(); 147 Sch.BaseClass::finishBlock(); 148 } 149 150 ArrayRef<const SUnit *> getTopRoots() const { 151 return TopRoots; 152 } 153 ArrayRef<SUnit*> getBottomRoots() const { 154 return BotRoots; 155 } 156 }; 157 158 class GCNIterativeScheduler::OverrideLegacyStrategy { 159 GCNIterativeScheduler &Sch; 160 Region &Rgn; 161 std::unique_ptr<MachineSchedStrategy> SaveSchedImpl; 162 GCNRegPressure SaveMaxRP; 163 164 public: 165 OverrideLegacyStrategy(Region &R, 166 MachineSchedStrategy &OverrideStrategy, 167 GCNIterativeScheduler &_Sch) 168 : Sch(_Sch) 169 , Rgn(R) 170 , SaveSchedImpl(std::move(_Sch.SchedImpl)) 171 , SaveMaxRP(R.MaxPressure) { 172 Sch.SchedImpl.reset(&OverrideStrategy); 173 auto BB = R.Begin->getParent(); 174 Sch.BaseClass::startBlock(BB); 175 Sch.BaseClass::enterRegion(BB, R.Begin, R.End, R.NumRegionInstrs); 176 } 177 178 ~OverrideLegacyStrategy() { 179 Sch.BaseClass::exitRegion(); 180 Sch.BaseClass::finishBlock(); 181 Sch.SchedImpl.release(); 182 Sch.SchedImpl = std::move(SaveSchedImpl); 183 } 184 185 void schedule() { 186 assert(Sch.RegionBegin == Rgn.Begin && Sch.RegionEnd == Rgn.End); 187 LLVM_DEBUG(dbgs() << "\nScheduling "; 188 printRegion(dbgs(), Rgn.Begin, Rgn.End, Sch.LIS, 2)); 189 Sch.BaseClass::schedule(); 190 191 // Unfortunately placeDebugValues incorrectly modifies RegionEnd, restore 192 Sch.RegionEnd = Rgn.End; 193 //assert(Rgn.End == Sch.RegionEnd); 194 Rgn.Begin = Sch.RegionBegin; 195 Rgn.MaxPressure.clear(); 196 } 197 198 void restoreOrder() { 199 assert(Sch.RegionBegin == Rgn.Begin && Sch.RegionEnd == Rgn.End); 200 // DAG SUnits are stored using original region's order 201 // so just use SUnits as the restoring schedule 202 Sch.scheduleRegion(Rgn, Sch.SUnits, SaveMaxRP); 203 } 204 }; 205 206 namespace { 207 208 // just a stub to make base class happy 209 class SchedStrategyStub : public MachineSchedStrategy { 210 public: 211 bool shouldTrackPressure() const override { return false; } 212 bool shouldTrackLaneMasks() const override { return false; } 213 void initialize(ScheduleDAGMI *DAG) override {} 214 SUnit *pickNode(bool &IsTopNode) override { return nullptr; } 215 void schedNode(SUnit *SU, bool IsTopNode) override {} 216 void releaseTopNode(SUnit *SU) override {} 217 void releaseBottomNode(SUnit *SU) override {} 218 }; 219 220 } // end anonymous namespace 221 222 GCNIterativeScheduler::GCNIterativeScheduler(MachineSchedContext *C, 223 StrategyKind S) 224 : BaseClass(C, std::make_unique<SchedStrategyStub>()) 225 , Context(C) 226 , Strategy(S) 227 , UPTracker(*LIS) { 228 } 229 230 // returns max pressure for a region 231 GCNRegPressure 232 GCNIterativeScheduler::getRegionPressure(MachineBasicBlock::iterator Begin, 233 MachineBasicBlock::iterator End) 234 const { 235 // For the purpose of pressure tracking bottom inst of the region should 236 // be also processed. End is either BB end, BB terminator inst or sched 237 // boundary inst. 238 auto const BBEnd = Begin->getParent()->end(); 239 auto const BottomMI = End == BBEnd ? std::prev(End) : End; 240 241 // scheduleRegions walks bottom to top, so its likely we just get next 242 // instruction to track 243 auto AfterBottomMI = std::next(BottomMI); 244 if (AfterBottomMI == BBEnd || 245 &*AfterBottomMI != UPTracker.getLastTrackedMI()) { 246 UPTracker.reset(*BottomMI); 247 } else { 248 assert(UPTracker.isValid()); 249 } 250 251 for (auto I = BottomMI; I != Begin; --I) 252 UPTracker.recede(*I); 253 254 UPTracker.recede(*Begin); 255 256 assert(UPTracker.isValid() || 257 (dbgs() << "Tracked region ", 258 printRegion(dbgs(), Begin, End, LIS), false)); 259 return UPTracker.moveMaxPressure(); 260 } 261 262 // returns max pressure for a tentative schedule 263 template <typename Range> GCNRegPressure 264 GCNIterativeScheduler::getSchedulePressure(const Region &R, 265 Range &&Schedule) const { 266 auto const BBEnd = R.Begin->getParent()->end(); 267 GCNUpwardRPTracker RPTracker(*LIS); 268 if (R.End != BBEnd) { 269 // R.End points to the boundary instruction but the 270 // schedule doesn't include it 271 RPTracker.reset(*R.End); 272 RPTracker.recede(*R.End); 273 } else { 274 // R.End doesn't point to the boundary instruction 275 RPTracker.reset(*std::prev(BBEnd)); 276 } 277 for (auto I = Schedule.end(), B = Schedule.begin(); I != B;) { 278 RPTracker.recede(*getMachineInstr(*--I)); 279 } 280 return RPTracker.moveMaxPressure(); 281 } 282 283 void GCNIterativeScheduler::enterRegion(MachineBasicBlock *BB, // overridden 284 MachineBasicBlock::iterator Begin, 285 MachineBasicBlock::iterator End, 286 unsigned NumRegionInstrs) { 287 BaseClass::enterRegion(BB, Begin, End, NumRegionInstrs); 288 if (NumRegionInstrs > 2) { 289 Regions.push_back( 290 new (Alloc.Allocate()) 291 Region { Begin, End, NumRegionInstrs, 292 getRegionPressure(Begin, End), nullptr }); 293 } 294 } 295 296 void GCNIterativeScheduler::schedule() { // overridden 297 // do nothing 298 LLVM_DEBUG(printLivenessInfo(dbgs(), RegionBegin, RegionEnd, LIS); 299 if (!Regions.empty() && Regions.back()->Begin == RegionBegin) { 300 dbgs() << "Max RP: "; 301 Regions.back()->MaxPressure.print( 302 dbgs(), &MF.getSubtarget<GCNSubtarget>()); 303 } dbgs() 304 << '\n';); 305 } 306 307 void GCNIterativeScheduler::finalizeSchedule() { // overridden 308 if (Regions.empty()) 309 return; 310 switch (Strategy) { 311 case SCHEDULE_MINREGONLY: scheduleMinReg(); break; 312 case SCHEDULE_MINREGFORCED: scheduleMinReg(true); break; 313 case SCHEDULE_LEGACYMAXOCCUPANCY: scheduleLegacyMaxOccupancy(); break; 314 case SCHEDULE_ILP: scheduleILP(false); break; 315 } 316 } 317 318 // Detach schedule from SUnits and interleave it with debug values. 319 // Returned schedule becomes independent of DAG state. 320 std::vector<MachineInstr*> 321 GCNIterativeScheduler::detachSchedule(ScheduleRef Schedule) const { 322 std::vector<MachineInstr*> Res; 323 Res.reserve(Schedule.size() * 2); 324 325 if (FirstDbgValue) 326 Res.push_back(FirstDbgValue); 327 328 const auto DbgB = DbgValues.begin(), DbgE = DbgValues.end(); 329 for (auto SU : Schedule) { 330 Res.push_back(SU->getInstr()); 331 const auto &D = std::find_if(DbgB, DbgE, [SU](decltype(*DbgB) &P) { 332 return P.second == SU->getInstr(); 333 }); 334 if (D != DbgE) 335 Res.push_back(D->first); 336 } 337 return Res; 338 } 339 340 void GCNIterativeScheduler::setBestSchedule(Region &R, 341 ScheduleRef Schedule, 342 const GCNRegPressure &MaxRP) { 343 R.BestSchedule.reset( 344 new TentativeSchedule{ detachSchedule(Schedule), MaxRP }); 345 } 346 347 void GCNIterativeScheduler::scheduleBest(Region &R) { 348 assert(R.BestSchedule.get() && "No schedule specified"); 349 scheduleRegion(R, R.BestSchedule->Schedule, R.BestSchedule->MaxPressure); 350 R.BestSchedule.reset(); 351 } 352 353 // minimal required region scheduler, works for ranges of SUnits*, 354 // SUnits or MachineIntrs* 355 template <typename Range> 356 void GCNIterativeScheduler::scheduleRegion(Region &R, Range &&Schedule, 357 const GCNRegPressure &MaxRP) { 358 assert(RegionBegin == R.Begin && RegionEnd == R.End); 359 assert(LIS != nullptr); 360 #ifndef NDEBUG 361 const auto SchedMaxRP = getSchedulePressure(R, Schedule); 362 #endif 363 auto BB = R.Begin->getParent(); 364 auto Top = R.Begin; 365 for (const auto &I : Schedule) { 366 auto MI = getMachineInstr(I); 367 if (MI != &*Top) { 368 BB->remove(MI); 369 BB->insert(Top, MI); 370 if (!MI->isDebugInstr()) 371 LIS->handleMove(*MI, true); 372 } 373 if (!MI->isDebugInstr()) { 374 // Reset read - undef flags and update them later. 375 for (auto &Op : MI->operands()) 376 if (Op.isReg() && Op.isDef()) 377 Op.setIsUndef(false); 378 379 RegisterOperands RegOpers; 380 RegOpers.collect(*MI, *TRI, MRI, /*ShouldTrackLaneMasks*/true, 381 /*IgnoreDead*/false); 382 // Adjust liveness and add missing dead+read-undef flags. 383 auto SlotIdx = LIS->getInstructionIndex(*MI).getRegSlot(); 384 RegOpers.adjustLaneLiveness(*LIS, MRI, SlotIdx, MI); 385 } 386 Top = std::next(MI->getIterator()); 387 } 388 RegionBegin = getMachineInstr(Schedule.front()); 389 390 // Schedule consisting of MachineInstr* is considered 'detached' 391 // and already interleaved with debug values 392 if (!std::is_same<decltype(*Schedule.begin()), MachineInstr*>::value) { 393 placeDebugValues(); 394 // Unfortunately placeDebugValues incorrectly modifies RegionEnd, restore 395 // assert(R.End == RegionEnd); 396 RegionEnd = R.End; 397 } 398 399 R.Begin = RegionBegin; 400 R.MaxPressure = MaxRP; 401 402 #ifndef NDEBUG 403 const auto RegionMaxRP = getRegionPressure(R); 404 const auto &ST = MF.getSubtarget<GCNSubtarget>(); 405 #endif 406 assert((SchedMaxRP == RegionMaxRP && (MaxRP.empty() || SchedMaxRP == MaxRP)) 407 || (dbgs() << "Max RP mismatch!!!\n" 408 "RP for schedule (calculated): ", 409 SchedMaxRP.print(dbgs(), &ST), 410 dbgs() << "RP for schedule (reported): ", 411 MaxRP.print(dbgs(), &ST), 412 dbgs() << "RP after scheduling: ", 413 RegionMaxRP.print(dbgs(), &ST), 414 false)); 415 } 416 417 // Sort recorded regions by pressure - highest at the front 418 void GCNIterativeScheduler::sortRegionsByPressure(unsigned TargetOcc) { 419 const auto &ST = MF.getSubtarget<GCNSubtarget>(); 420 llvm::sort(Regions, [&ST, TargetOcc](const Region *R1, const Region *R2) { 421 return R2->MaxPressure.less(ST, R1->MaxPressure, TargetOcc); 422 }); 423 } 424 425 /////////////////////////////////////////////////////////////////////////////// 426 // Legacy MaxOccupancy Strategy 427 428 // Tries to increase occupancy applying minreg scheduler for a sequence of 429 // most demanding regions. Obtained schedules are saved as BestSchedule for a 430 // region. 431 // TargetOcc is the best achievable occupancy for a kernel. 432 // Returns better occupancy on success or current occupancy on fail. 433 // BestSchedules aren't deleted on fail. 434 unsigned GCNIterativeScheduler::tryMaximizeOccupancy(unsigned TargetOcc) { 435 // TODO: assert Regions are sorted descending by pressure 436 const auto &ST = MF.getSubtarget<GCNSubtarget>(); 437 const auto Occ = Regions.front()->MaxPressure.getOccupancy(ST); 438 LLVM_DEBUG(dbgs() << "Trying to improve occupancy, target = " << TargetOcc 439 << ", current = " << Occ << '\n'); 440 441 auto NewOcc = TargetOcc; 442 for (auto R : Regions) { 443 if (R->MaxPressure.getOccupancy(ST) >= NewOcc) 444 break; 445 446 LLVM_DEBUG(printRegion(dbgs(), R->Begin, R->End, LIS, 3); 447 printLivenessInfo(dbgs(), R->Begin, R->End, LIS)); 448 449 BuildDAG DAG(*R, *this); 450 const auto MinSchedule = makeMinRegSchedule(DAG.getTopRoots(), *this); 451 const auto MaxRP = getSchedulePressure(*R, MinSchedule); 452 LLVM_DEBUG(dbgs() << "Occupancy improvement attempt:\n"; 453 printSchedRP(dbgs(), R->MaxPressure, MaxRP)); 454 455 NewOcc = std::min(NewOcc, MaxRP.getOccupancy(ST)); 456 if (NewOcc <= Occ) 457 break; 458 459 setBestSchedule(*R, MinSchedule, MaxRP); 460 } 461 LLVM_DEBUG(dbgs() << "New occupancy = " << NewOcc 462 << ", prev occupancy = " << Occ << '\n'); 463 if (NewOcc > Occ) { 464 SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>(); 465 MFI->increaseOccupancy(MF, NewOcc); 466 } 467 468 return std::max(NewOcc, Occ); 469 } 470 471 void GCNIterativeScheduler::scheduleLegacyMaxOccupancy( 472 bool TryMaximizeOccupancy) { 473 const auto &ST = MF.getSubtarget<GCNSubtarget>(); 474 SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>(); 475 auto TgtOcc = MFI->getMinAllowedOccupancy(); 476 477 sortRegionsByPressure(TgtOcc); 478 auto Occ = Regions.front()->MaxPressure.getOccupancy(ST); 479 480 if (TryMaximizeOccupancy && Occ < TgtOcc) 481 Occ = tryMaximizeOccupancy(TgtOcc); 482 483 // This is really weird but for some magic scheduling regions twice 484 // gives performance improvement 485 const int NumPasses = Occ < TgtOcc ? 2 : 1; 486 487 TgtOcc = std::min(Occ, TgtOcc); 488 LLVM_DEBUG(dbgs() << "Scheduling using default scheduler, " 489 "target occupancy = " 490 << TgtOcc << '\n'); 491 GCNMaxOccupancySchedStrategy LStrgy(Context); 492 unsigned FinalOccupancy = std::min(Occ, MFI->getOccupancy()); 493 494 for (int I = 0; I < NumPasses; ++I) { 495 // running first pass with TargetOccupancy = 0 mimics previous scheduling 496 // approach and is a performance magic 497 LStrgy.setTargetOccupancy(I == 0 ? 0 : TgtOcc); 498 for (auto R : Regions) { 499 OverrideLegacyStrategy Ovr(*R, LStrgy, *this); 500 501 Ovr.schedule(); 502 const auto RP = getRegionPressure(*R); 503 LLVM_DEBUG(printSchedRP(dbgs(), R->MaxPressure, RP)); 504 505 if (RP.getOccupancy(ST) < TgtOcc) { 506 LLVM_DEBUG(dbgs() << "Didn't fit into target occupancy O" << TgtOcc); 507 if (R->BestSchedule.get() && 508 R->BestSchedule->MaxPressure.getOccupancy(ST) >= TgtOcc) { 509 LLVM_DEBUG(dbgs() << ", scheduling minimal register\n"); 510 scheduleBest(*R); 511 } else { 512 LLVM_DEBUG(dbgs() << ", restoring\n"); 513 Ovr.restoreOrder(); 514 assert(R->MaxPressure.getOccupancy(ST) >= TgtOcc); 515 } 516 } 517 FinalOccupancy = std::min(FinalOccupancy, RP.getOccupancy(ST)); 518 } 519 } 520 MFI->limitOccupancy(FinalOccupancy); 521 } 522 523 /////////////////////////////////////////////////////////////////////////////// 524 // Minimal Register Strategy 525 526 void GCNIterativeScheduler::scheduleMinReg(bool force) { 527 const auto &ST = MF.getSubtarget<GCNSubtarget>(); 528 const SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>(); 529 const auto TgtOcc = MFI->getOccupancy(); 530 sortRegionsByPressure(TgtOcc); 531 532 auto MaxPressure = Regions.front()->MaxPressure; 533 for (auto R : Regions) { 534 if (!force && R->MaxPressure.less(ST, MaxPressure, TgtOcc)) 535 break; 536 537 BuildDAG DAG(*R, *this); 538 const auto MinSchedule = makeMinRegSchedule(DAG.getTopRoots(), *this); 539 540 const auto RP = getSchedulePressure(*R, MinSchedule); 541 LLVM_DEBUG(if (R->MaxPressure.less(ST, RP, TgtOcc)) { 542 dbgs() << "\nWarning: Pressure becomes worse after minreg!"; 543 printSchedRP(dbgs(), R->MaxPressure, RP); 544 }); 545 546 if (!force && MaxPressure.less(ST, RP, TgtOcc)) 547 break; 548 549 scheduleRegion(*R, MinSchedule, RP); 550 LLVM_DEBUG(printSchedResult(dbgs(), R, RP)); 551 552 MaxPressure = RP; 553 } 554 } 555 556 /////////////////////////////////////////////////////////////////////////////// 557 // ILP scheduler port 558 559 void GCNIterativeScheduler::scheduleILP( 560 bool TryMaximizeOccupancy) { 561 const auto &ST = MF.getSubtarget<GCNSubtarget>(); 562 SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>(); 563 auto TgtOcc = MFI->getMinAllowedOccupancy(); 564 565 sortRegionsByPressure(TgtOcc); 566 auto Occ = Regions.front()->MaxPressure.getOccupancy(ST); 567 568 if (TryMaximizeOccupancy && Occ < TgtOcc) 569 Occ = tryMaximizeOccupancy(TgtOcc); 570 571 TgtOcc = std::min(Occ, TgtOcc); 572 LLVM_DEBUG(dbgs() << "Scheduling using default scheduler, " 573 "target occupancy = " 574 << TgtOcc << '\n'); 575 576 unsigned FinalOccupancy = std::min(Occ, MFI->getOccupancy()); 577 for (auto R : Regions) { 578 BuildDAG DAG(*R, *this); 579 const auto ILPSchedule = makeGCNILPScheduler(DAG.getBottomRoots(), *this); 580 581 const auto RP = getSchedulePressure(*R, ILPSchedule); 582 LLVM_DEBUG(printSchedRP(dbgs(), R->MaxPressure, RP)); 583 584 if (RP.getOccupancy(ST) < TgtOcc) { 585 LLVM_DEBUG(dbgs() << "Didn't fit into target occupancy O" << TgtOcc); 586 if (R->BestSchedule.get() && 587 R->BestSchedule->MaxPressure.getOccupancy(ST) >= TgtOcc) { 588 LLVM_DEBUG(dbgs() << ", scheduling minimal register\n"); 589 scheduleBest(*R); 590 } 591 } else { 592 scheduleRegion(*R, ILPSchedule, RP); 593 LLVM_DEBUG(printSchedResult(dbgs(), R, RP)); 594 FinalOccupancy = std::min(FinalOccupancy, RP.getOccupancy(ST)); 595 } 596 } 597 MFI->limitOccupancy(FinalOccupancy); 598 } 599