1 //===-- GCNSchedStrategy.cpp - GCN Scheduler Strategy ---------------------===// 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 contains a MachineSchedStrategy implementation for maximizing wave 11 /// occupancy on GCN hardware. 12 /// 13 /// This pass will apply multiple scheduling stages to the same function. 14 /// Regions are first recorded in GCNScheduleDAGMILive::schedule. The actual 15 /// entry point for the scheduling of those regions is 16 /// GCNScheduleDAGMILive::runSchedStages. 17 18 /// Generally, the reason for having multiple scheduling stages is to account 19 /// for the kernel-wide effect of register usage on occupancy. Usually, only a 20 /// few scheduling regions will have register pressure high enough to limit 21 /// occupancy for the kernel, so constraints can be relaxed to improve ILP in 22 /// other regions. 23 /// 24 //===----------------------------------------------------------------------===// 25 26 #include "GCNSchedStrategy.h" 27 #include "AMDGPUIGroupLP.h" 28 #include "SIMachineFunctionInfo.h" 29 #include "llvm/CodeGen/RegisterClassInfo.h" 30 31 #define DEBUG_TYPE "machine-scheduler" 32 33 using namespace llvm; 34 35 static cl::opt<bool> DisableUnclusterHighRP( 36 "amdgpu-disable-unclustered-high-rp-reschedule", cl::Hidden, 37 cl::desc("Disable unclustered high register pressure " 38 "reduction scheduling stage."), 39 cl::init(false)); 40 41 static cl::opt<bool> DisableClusteredLowOccupancy( 42 "amdgpu-disable-clustered-low-occupancy-reschedule", cl::Hidden, 43 cl::desc("Disable clustered low occupancy " 44 "rescheduling for ILP scheduling stage."), 45 cl::init(false)); 46 47 static cl::opt<unsigned> ScheduleMetricBias( 48 "amdgpu-schedule-metric-bias", cl::Hidden, 49 cl::desc( 50 "Sets the bias which adds weight to occupancy vs latency. Set it to " 51 "100 to chase the occupancy only."), 52 cl::init(10)); 53 54 static cl::opt<bool> 55 RelaxedOcc("amdgpu-schedule-relaxed-occupancy", cl::Hidden, 56 cl::desc("Relax occupancy targets for kernels which are memory " 57 "bound (amdgpu-membound-threshold), or " 58 "Wave Limited (amdgpu-limit-wave-threshold)."), 59 cl::init(false)); 60 61 const unsigned ScheduleMetrics::ScaleFactor = 100; 62 63 GCNSchedStrategy::GCNSchedStrategy(const MachineSchedContext *C) 64 : GenericScheduler(C), TargetOccupancy(0), MF(nullptr), 65 HasHighPressure(false) {} 66 67 void GCNSchedStrategy::initialize(ScheduleDAGMI *DAG) { 68 GenericScheduler::initialize(DAG); 69 70 MF = &DAG->MF; 71 72 const GCNSubtarget &ST = MF->getSubtarget<GCNSubtarget>(); 73 74 SGPRExcessLimit = 75 Context->RegClassInfo->getNumAllocatableRegs(&AMDGPU::SGPR_32RegClass); 76 VGPRExcessLimit = 77 Context->RegClassInfo->getNumAllocatableRegs(&AMDGPU::VGPR_32RegClass); 78 79 SIMachineFunctionInfo &MFI = *MF->getInfo<SIMachineFunctionInfo>(); 80 // Set the initial TargetOccupnacy to the maximum occupancy that we can 81 // achieve for this function. This effectively sets a lower bound on the 82 // 'Critical' register limits in the scheduler. 83 // Allow for lower occupancy targets if kernel is wave limited or memory 84 // bound, and using the relaxed occupancy feature. 85 TargetOccupancy = 86 RelaxedOcc ? MFI.getMinAllowedOccupancy() : MFI.getOccupancy(); 87 SGPRCriticalLimit = 88 std::min(ST.getMaxNumSGPRs(TargetOccupancy, true), SGPRExcessLimit); 89 90 if (!KnownExcessRP) { 91 VGPRCriticalLimit = 92 std::min(ST.getMaxNumVGPRs(TargetOccupancy), VGPRExcessLimit); 93 } else { 94 // This is similar to ST.getMaxNumVGPRs(TargetOccupancy) result except 95 // returns a reasonably small number for targets with lots of VGPRs, such 96 // as GFX10 and GFX11. 97 LLVM_DEBUG(dbgs() << "Region is known to spill, use alternative " 98 "VGPRCriticalLimit calculation method.\n"); 99 100 unsigned Granule = AMDGPU::IsaInfo::getVGPRAllocGranule(&ST); 101 unsigned Addressable = AMDGPU::IsaInfo::getAddressableNumVGPRs(&ST); 102 unsigned VGPRBudget = alignDown(Addressable / TargetOccupancy, Granule); 103 VGPRBudget = std::max(VGPRBudget, Granule); 104 VGPRCriticalLimit = std::min(VGPRBudget, VGPRExcessLimit); 105 } 106 107 // Subtract error margin and bias from register limits and avoid overflow. 108 SGPRCriticalLimit -= std::min(SGPRLimitBias + ErrorMargin, SGPRCriticalLimit); 109 VGPRCriticalLimit -= std::min(VGPRLimitBias + ErrorMargin, VGPRCriticalLimit); 110 SGPRExcessLimit -= std::min(SGPRLimitBias + ErrorMargin, SGPRExcessLimit); 111 VGPRExcessLimit -= std::min(VGPRLimitBias + ErrorMargin, VGPRExcessLimit); 112 113 LLVM_DEBUG(dbgs() << "VGPRCriticalLimit = " << VGPRCriticalLimit 114 << ", VGPRExcessLimit = " << VGPRExcessLimit 115 << ", SGPRCriticalLimit = " << SGPRCriticalLimit 116 << ", SGPRExcessLimit = " << SGPRExcessLimit << "\n\n"); 117 } 118 119 void GCNSchedStrategy::initCandidate(SchedCandidate &Cand, SUnit *SU, 120 bool AtTop, 121 const RegPressureTracker &RPTracker, 122 const SIRegisterInfo *SRI, 123 unsigned SGPRPressure, 124 unsigned VGPRPressure) { 125 Cand.SU = SU; 126 Cand.AtTop = AtTop; 127 128 if (!DAG->isTrackingPressure()) 129 return; 130 131 // getDownwardPressure() and getUpwardPressure() make temporary changes to 132 // the tracker, so we need to pass those function a non-const copy. 133 RegPressureTracker &TempTracker = const_cast<RegPressureTracker&>(RPTracker); 134 135 Pressure.clear(); 136 MaxPressure.clear(); 137 138 if (AtTop) 139 TempTracker.getDownwardPressure(SU->getInstr(), Pressure, MaxPressure); 140 else { 141 // FIXME: I think for bottom up scheduling, the register pressure is cached 142 // and can be retrieved by DAG->getPressureDif(SU). 143 TempTracker.getUpwardPressure(SU->getInstr(), Pressure, MaxPressure); 144 } 145 146 unsigned NewSGPRPressure = Pressure[AMDGPU::RegisterPressureSets::SReg_32]; 147 unsigned NewVGPRPressure = Pressure[AMDGPU::RegisterPressureSets::VGPR_32]; 148 149 // If two instructions increase the pressure of different register sets 150 // by the same amount, the generic scheduler will prefer to schedule the 151 // instruction that increases the set with the least amount of registers, 152 // which in our case would be SGPRs. This is rarely what we want, so 153 // when we report excess/critical register pressure, we do it either 154 // only for VGPRs or only for SGPRs. 155 156 // FIXME: Better heuristics to determine whether to prefer SGPRs or VGPRs. 157 const unsigned MaxVGPRPressureInc = 16; 158 bool ShouldTrackVGPRs = VGPRPressure + MaxVGPRPressureInc >= VGPRExcessLimit; 159 bool ShouldTrackSGPRs = !ShouldTrackVGPRs && SGPRPressure >= SGPRExcessLimit; 160 161 162 // FIXME: We have to enter REG-EXCESS before we reach the actual threshold 163 // to increase the likelihood we don't go over the limits. We should improve 164 // the analysis to look through dependencies to find the path with the least 165 // register pressure. 166 167 // We only need to update the RPDelta for instructions that increase register 168 // pressure. Instructions that decrease or keep reg pressure the same will be 169 // marked as RegExcess in tryCandidate() when they are compared with 170 // instructions that increase the register pressure. 171 if (ShouldTrackVGPRs && NewVGPRPressure >= VGPRExcessLimit) { 172 HasHighPressure = true; 173 Cand.RPDelta.Excess = PressureChange(AMDGPU::RegisterPressureSets::VGPR_32); 174 Cand.RPDelta.Excess.setUnitInc(NewVGPRPressure - VGPRExcessLimit); 175 } 176 177 if (ShouldTrackSGPRs && NewSGPRPressure >= SGPRExcessLimit) { 178 HasHighPressure = true; 179 Cand.RPDelta.Excess = PressureChange(AMDGPU::RegisterPressureSets::SReg_32); 180 Cand.RPDelta.Excess.setUnitInc(NewSGPRPressure - SGPRExcessLimit); 181 } 182 183 // Register pressure is considered 'CRITICAL' if it is approaching a value 184 // that would reduce the wave occupancy for the execution unit. When 185 // register pressure is 'CRITICAL', increasing SGPR and VGPR pressure both 186 // has the same cost, so we don't need to prefer one over the other. 187 188 int SGPRDelta = NewSGPRPressure - SGPRCriticalLimit; 189 int VGPRDelta = NewVGPRPressure - VGPRCriticalLimit; 190 191 if (SGPRDelta >= 0 || VGPRDelta >= 0) { 192 HasHighPressure = true; 193 if (SGPRDelta > VGPRDelta) { 194 Cand.RPDelta.CriticalMax = 195 PressureChange(AMDGPU::RegisterPressureSets::SReg_32); 196 Cand.RPDelta.CriticalMax.setUnitInc(SGPRDelta); 197 } else { 198 Cand.RPDelta.CriticalMax = 199 PressureChange(AMDGPU::RegisterPressureSets::VGPR_32); 200 Cand.RPDelta.CriticalMax.setUnitInc(VGPRDelta); 201 } 202 } 203 } 204 205 // This function is mostly cut and pasted from 206 // GenericScheduler::pickNodeFromQueue() 207 void GCNSchedStrategy::pickNodeFromQueue(SchedBoundary &Zone, 208 const CandPolicy &ZonePolicy, 209 const RegPressureTracker &RPTracker, 210 SchedCandidate &Cand) { 211 const SIRegisterInfo *SRI = static_cast<const SIRegisterInfo*>(TRI); 212 ArrayRef<unsigned> Pressure = RPTracker.getRegSetPressureAtPos(); 213 unsigned SGPRPressure = 0; 214 unsigned VGPRPressure = 0; 215 if (DAG->isTrackingPressure()) { 216 SGPRPressure = Pressure[AMDGPU::RegisterPressureSets::SReg_32]; 217 VGPRPressure = Pressure[AMDGPU::RegisterPressureSets::VGPR_32]; 218 } 219 ReadyQueue &Q = Zone.Available; 220 for (SUnit *SU : Q) { 221 222 SchedCandidate TryCand(ZonePolicy); 223 initCandidate(TryCand, SU, Zone.isTop(), RPTracker, SRI, 224 SGPRPressure, VGPRPressure); 225 // Pass SchedBoundary only when comparing nodes from the same boundary. 226 SchedBoundary *ZoneArg = Cand.AtTop == TryCand.AtTop ? &Zone : nullptr; 227 tryCandidate(Cand, TryCand, ZoneArg); 228 if (TryCand.Reason != NoCand) { 229 // Initialize resource delta if needed in case future heuristics query it. 230 if (TryCand.ResDelta == SchedResourceDelta()) 231 TryCand.initResourceDelta(Zone.DAG, SchedModel); 232 Cand.setBest(TryCand); 233 LLVM_DEBUG(traceCandidate(Cand)); 234 } 235 } 236 } 237 238 // This function is mostly cut and pasted from 239 // GenericScheduler::pickNodeBidirectional() 240 SUnit *GCNSchedStrategy::pickNodeBidirectional(bool &IsTopNode) { 241 // Schedule as far as possible in the direction of no choice. This is most 242 // efficient, but also provides the best heuristics for CriticalPSets. 243 if (SUnit *SU = Bot.pickOnlyChoice()) { 244 IsTopNode = false; 245 return SU; 246 } 247 if (SUnit *SU = Top.pickOnlyChoice()) { 248 IsTopNode = true; 249 return SU; 250 } 251 // Set the bottom-up policy based on the state of the current bottom zone and 252 // the instructions outside the zone, including the top zone. 253 CandPolicy BotPolicy; 254 setPolicy(BotPolicy, /*IsPostRA=*/false, Bot, &Top); 255 // Set the top-down policy based on the state of the current top zone and 256 // the instructions outside the zone, including the bottom zone. 257 CandPolicy TopPolicy; 258 setPolicy(TopPolicy, /*IsPostRA=*/false, Top, &Bot); 259 260 // See if BotCand is still valid (because we previously scheduled from Top). 261 LLVM_DEBUG(dbgs() << "Picking from Bot:\n"); 262 if (!BotCand.isValid() || BotCand.SU->isScheduled || 263 BotCand.Policy != BotPolicy) { 264 BotCand.reset(CandPolicy()); 265 pickNodeFromQueue(Bot, BotPolicy, DAG->getBotRPTracker(), BotCand); 266 assert(BotCand.Reason != NoCand && "failed to find the first candidate"); 267 } else { 268 LLVM_DEBUG(traceCandidate(BotCand)); 269 #ifndef NDEBUG 270 if (VerifyScheduling) { 271 SchedCandidate TCand; 272 TCand.reset(CandPolicy()); 273 pickNodeFromQueue(Bot, BotPolicy, DAG->getBotRPTracker(), TCand); 274 assert(TCand.SU == BotCand.SU && 275 "Last pick result should correspond to re-picking right now"); 276 } 277 #endif 278 } 279 280 // Check if the top Q has a better candidate. 281 LLVM_DEBUG(dbgs() << "Picking from Top:\n"); 282 if (!TopCand.isValid() || TopCand.SU->isScheduled || 283 TopCand.Policy != TopPolicy) { 284 TopCand.reset(CandPolicy()); 285 pickNodeFromQueue(Top, TopPolicy, DAG->getTopRPTracker(), TopCand); 286 assert(TopCand.Reason != NoCand && "failed to find the first candidate"); 287 } else { 288 LLVM_DEBUG(traceCandidate(TopCand)); 289 #ifndef NDEBUG 290 if (VerifyScheduling) { 291 SchedCandidate TCand; 292 TCand.reset(CandPolicy()); 293 pickNodeFromQueue(Top, TopPolicy, DAG->getTopRPTracker(), TCand); 294 assert(TCand.SU == TopCand.SU && 295 "Last pick result should correspond to re-picking right now"); 296 } 297 #endif 298 } 299 300 // Pick best from BotCand and TopCand. 301 LLVM_DEBUG(dbgs() << "Top Cand: "; traceCandidate(TopCand); 302 dbgs() << "Bot Cand: "; traceCandidate(BotCand);); 303 SchedCandidate Cand = BotCand; 304 TopCand.Reason = NoCand; 305 tryCandidate(Cand, TopCand, nullptr); 306 if (TopCand.Reason != NoCand) { 307 Cand.setBest(TopCand); 308 } 309 LLVM_DEBUG(dbgs() << "Picking: "; traceCandidate(Cand);); 310 311 IsTopNode = Cand.AtTop; 312 return Cand.SU; 313 } 314 315 // This function is mostly cut and pasted from 316 // GenericScheduler::pickNode() 317 SUnit *GCNSchedStrategy::pickNode(bool &IsTopNode) { 318 if (DAG->top() == DAG->bottom()) { 319 assert(Top.Available.empty() && Top.Pending.empty() && 320 Bot.Available.empty() && Bot.Pending.empty() && "ReadyQ garbage"); 321 return nullptr; 322 } 323 SUnit *SU; 324 do { 325 if (RegionPolicy.OnlyTopDown) { 326 SU = Top.pickOnlyChoice(); 327 if (!SU) { 328 CandPolicy NoPolicy; 329 TopCand.reset(NoPolicy); 330 pickNodeFromQueue(Top, NoPolicy, DAG->getTopRPTracker(), TopCand); 331 assert(TopCand.Reason != NoCand && "failed to find a candidate"); 332 SU = TopCand.SU; 333 } 334 IsTopNode = true; 335 } else if (RegionPolicy.OnlyBottomUp) { 336 SU = Bot.pickOnlyChoice(); 337 if (!SU) { 338 CandPolicy NoPolicy; 339 BotCand.reset(NoPolicy); 340 pickNodeFromQueue(Bot, NoPolicy, DAG->getBotRPTracker(), BotCand); 341 assert(BotCand.Reason != NoCand && "failed to find a candidate"); 342 SU = BotCand.SU; 343 } 344 IsTopNode = false; 345 } else { 346 SU = pickNodeBidirectional(IsTopNode); 347 } 348 } while (SU->isScheduled); 349 350 if (SU->isTopReady()) 351 Top.removeReady(SU); 352 if (SU->isBottomReady()) 353 Bot.removeReady(SU); 354 355 LLVM_DEBUG(dbgs() << "Scheduling SU(" << SU->NodeNum << ") " 356 << *SU->getInstr()); 357 return SU; 358 } 359 360 GCNSchedStageID GCNSchedStrategy::getCurrentStage() { 361 assert(CurrentStage && CurrentStage != SchedStages.end()); 362 return *CurrentStage; 363 } 364 365 bool GCNSchedStrategy::advanceStage() { 366 assert(CurrentStage != SchedStages.end()); 367 if (!CurrentStage) 368 CurrentStage = SchedStages.begin(); 369 else 370 CurrentStage++; 371 372 return CurrentStage != SchedStages.end(); 373 } 374 375 bool GCNSchedStrategy::hasNextStage() const { 376 assert(CurrentStage); 377 return std::next(CurrentStage) != SchedStages.end(); 378 } 379 380 GCNSchedStageID GCNSchedStrategy::getNextStage() const { 381 assert(CurrentStage && std::next(CurrentStage) != SchedStages.end()); 382 return *std::next(CurrentStage); 383 } 384 385 GCNMaxOccupancySchedStrategy::GCNMaxOccupancySchedStrategy( 386 const MachineSchedContext *C) 387 : GCNSchedStrategy(C) { 388 SchedStages.push_back(GCNSchedStageID::OccInitialSchedule); 389 SchedStages.push_back(GCNSchedStageID::UnclusteredHighRPReschedule); 390 SchedStages.push_back(GCNSchedStageID::ClusteredLowOccupancyReschedule); 391 SchedStages.push_back(GCNSchedStageID::PreRARematerialize); 392 } 393 394 GCNMaxILPSchedStrategy::GCNMaxILPSchedStrategy(const MachineSchedContext *C) 395 : GCNSchedStrategy(C) { 396 SchedStages.push_back(GCNSchedStageID::ILPInitialSchedule); 397 } 398 399 bool GCNMaxILPSchedStrategy::tryCandidate(SchedCandidate &Cand, 400 SchedCandidate &TryCand, 401 SchedBoundary *Zone) const { 402 // Initialize the candidate if needed. 403 if (!Cand.isValid()) { 404 TryCand.Reason = NodeOrder; 405 return true; 406 } 407 408 // Avoid spilling by exceeding the register limit. 409 if (DAG->isTrackingPressure() && 410 tryPressure(TryCand.RPDelta.Excess, Cand.RPDelta.Excess, TryCand, Cand, 411 RegExcess, TRI, DAG->MF)) 412 return TryCand.Reason != NoCand; 413 414 // Bias PhysReg Defs and copies to their uses and defined respectively. 415 if (tryGreater(biasPhysReg(TryCand.SU, TryCand.AtTop), 416 biasPhysReg(Cand.SU, Cand.AtTop), TryCand, Cand, PhysReg)) 417 return TryCand.Reason != NoCand; 418 419 bool SameBoundary = Zone != nullptr; 420 if (SameBoundary) { 421 // Prioritize instructions that read unbuffered resources by stall cycles. 422 if (tryLess(Zone->getLatencyStallCycles(TryCand.SU), 423 Zone->getLatencyStallCycles(Cand.SU), TryCand, Cand, Stall)) 424 return TryCand.Reason != NoCand; 425 426 // Avoid critical resource consumption and balance the schedule. 427 TryCand.initResourceDelta(DAG, SchedModel); 428 if (tryLess(TryCand.ResDelta.CritResources, Cand.ResDelta.CritResources, 429 TryCand, Cand, ResourceReduce)) 430 return TryCand.Reason != NoCand; 431 if (tryGreater(TryCand.ResDelta.DemandedResources, 432 Cand.ResDelta.DemandedResources, TryCand, Cand, 433 ResourceDemand)) 434 return TryCand.Reason != NoCand; 435 436 // Unconditionally try to reduce latency. 437 if (tryLatency(TryCand, Cand, *Zone)) 438 return TryCand.Reason != NoCand; 439 440 // Weak edges are for clustering and other constraints. 441 if (tryLess(getWeakLeft(TryCand.SU, TryCand.AtTop), 442 getWeakLeft(Cand.SU, Cand.AtTop), TryCand, Cand, Weak)) 443 return TryCand.Reason != NoCand; 444 } 445 446 // Keep clustered nodes together to encourage downstream peephole 447 // optimizations which may reduce resource requirements. 448 // 449 // This is a best effort to set things up for a post-RA pass. Optimizations 450 // like generating loads of multiple registers should ideally be done within 451 // the scheduler pass by combining the loads during DAG postprocessing. 452 const SUnit *CandNextClusterSU = 453 Cand.AtTop ? DAG->getNextClusterSucc() : DAG->getNextClusterPred(); 454 const SUnit *TryCandNextClusterSU = 455 TryCand.AtTop ? DAG->getNextClusterSucc() : DAG->getNextClusterPred(); 456 if (tryGreater(TryCand.SU == TryCandNextClusterSU, 457 Cand.SU == CandNextClusterSU, TryCand, Cand, Cluster)) 458 return TryCand.Reason != NoCand; 459 460 // Avoid increasing the max critical pressure in the scheduled region. 461 if (DAG->isTrackingPressure() && 462 tryPressure(TryCand.RPDelta.CriticalMax, Cand.RPDelta.CriticalMax, 463 TryCand, Cand, RegCritical, TRI, DAG->MF)) 464 return TryCand.Reason != NoCand; 465 466 // Avoid increasing the max pressure of the entire region. 467 if (DAG->isTrackingPressure() && 468 tryPressure(TryCand.RPDelta.CurrentMax, Cand.RPDelta.CurrentMax, TryCand, 469 Cand, RegMax, TRI, DAG->MF)) 470 return TryCand.Reason != NoCand; 471 472 if (SameBoundary) { 473 // Fall through to original instruction order. 474 if ((Zone->isTop() && TryCand.SU->NodeNum < Cand.SU->NodeNum) || 475 (!Zone->isTop() && TryCand.SU->NodeNum > Cand.SU->NodeNum)) { 476 TryCand.Reason = NodeOrder; 477 return true; 478 } 479 } 480 return false; 481 } 482 483 GCNScheduleDAGMILive::GCNScheduleDAGMILive( 484 MachineSchedContext *C, std::unique_ptr<MachineSchedStrategy> S) 485 : ScheduleDAGMILive(C, std::move(S)), ST(MF.getSubtarget<GCNSubtarget>()), 486 MFI(*MF.getInfo<SIMachineFunctionInfo>()), 487 StartingOccupancy(MFI.getOccupancy()), MinOccupancy(StartingOccupancy) { 488 489 LLVM_DEBUG(dbgs() << "Starting occupancy is " << StartingOccupancy << ".\n"); 490 if (RelaxedOcc) { 491 MinOccupancy = std::min(MFI.getMinAllowedOccupancy(), StartingOccupancy); 492 if (MinOccupancy != StartingOccupancy) 493 LLVM_DEBUG(dbgs() << "Allowing Occupancy drops to " << MinOccupancy 494 << ".\n"); 495 } 496 } 497 498 std::unique_ptr<GCNSchedStage> 499 GCNScheduleDAGMILive::createSchedStage(GCNSchedStageID SchedStageID) { 500 switch (SchedStageID) { 501 case GCNSchedStageID::OccInitialSchedule: 502 return std::make_unique<OccInitialScheduleStage>(SchedStageID, *this); 503 case GCNSchedStageID::UnclusteredHighRPReschedule: 504 return std::make_unique<UnclusteredHighRPStage>(SchedStageID, *this); 505 case GCNSchedStageID::ClusteredLowOccupancyReschedule: 506 return std::make_unique<ClusteredLowOccStage>(SchedStageID, *this); 507 case GCNSchedStageID::PreRARematerialize: 508 return std::make_unique<PreRARematStage>(SchedStageID, *this); 509 case GCNSchedStageID::ILPInitialSchedule: 510 return std::make_unique<ILPInitialScheduleStage>(SchedStageID, *this); 511 } 512 513 llvm_unreachable("Unknown SchedStageID."); 514 } 515 516 void GCNScheduleDAGMILive::schedule() { 517 // Collect all scheduling regions. The actual scheduling is performed in 518 // GCNScheduleDAGMILive::finalizeSchedule. 519 Regions.push_back(std::pair(RegionBegin, RegionEnd)); 520 } 521 522 GCNRegPressure 523 GCNScheduleDAGMILive::getRealRegPressure(unsigned RegionIdx) const { 524 GCNDownwardRPTracker RPTracker(*LIS); 525 RPTracker.advance(begin(), end(), &LiveIns[RegionIdx]); 526 return RPTracker.moveMaxPressure(); 527 } 528 529 void GCNScheduleDAGMILive::computeBlockPressure(unsigned RegionIdx, 530 const MachineBasicBlock *MBB) { 531 GCNDownwardRPTracker RPTracker(*LIS); 532 533 // If the block has the only successor then live-ins of that successor are 534 // live-outs of the current block. We can reuse calculated live set if the 535 // successor will be sent to scheduling past current block. 536 537 // However, due to the bug in LiveInterval analysis it may happen that two 538 // predecessors of the same successor block have different lane bitmasks for 539 // a live-out register. Workaround that by sticking to one-to-one relationship 540 // i.e. one predecessor with one successor block. 541 const MachineBasicBlock *OnlySucc = nullptr; 542 if (MBB->succ_size() == 1) { 543 auto *Candidate = *MBB->succ_begin(); 544 if (!Candidate->empty() && Candidate->pred_size() == 1) { 545 SlotIndexes *Ind = LIS->getSlotIndexes(); 546 if (Ind->getMBBStartIdx(MBB) < Ind->getMBBStartIdx(Candidate)) 547 OnlySucc = Candidate; 548 } 549 } 550 551 // Scheduler sends regions from the end of the block upwards. 552 size_t CurRegion = RegionIdx; 553 for (size_t E = Regions.size(); CurRegion != E; ++CurRegion) 554 if (Regions[CurRegion].first->getParent() != MBB) 555 break; 556 --CurRegion; 557 558 auto I = MBB->begin(); 559 auto LiveInIt = MBBLiveIns.find(MBB); 560 auto &Rgn = Regions[CurRegion]; 561 auto *NonDbgMI = &*skipDebugInstructionsForward(Rgn.first, Rgn.second); 562 if (LiveInIt != MBBLiveIns.end()) { 563 auto LiveIn = std::move(LiveInIt->second); 564 RPTracker.reset(*MBB->begin(), &LiveIn); 565 MBBLiveIns.erase(LiveInIt); 566 } else { 567 I = Rgn.first; 568 auto LRS = BBLiveInMap.lookup(NonDbgMI); 569 #ifdef EXPENSIVE_CHECKS 570 assert(isEqual(getLiveRegsBefore(*NonDbgMI, *LIS), LRS)); 571 #endif 572 RPTracker.reset(*I, &LRS); 573 } 574 575 for (;;) { 576 I = RPTracker.getNext(); 577 578 if (Regions[CurRegion].first == I || NonDbgMI == I) { 579 LiveIns[CurRegion] = RPTracker.getLiveRegs(); 580 RPTracker.clearMaxPressure(); 581 } 582 583 if (Regions[CurRegion].second == I) { 584 Pressure[CurRegion] = RPTracker.moveMaxPressure(); 585 if (CurRegion-- == RegionIdx) 586 break; 587 } 588 RPTracker.advanceToNext(); 589 RPTracker.advanceBeforeNext(); 590 } 591 592 if (OnlySucc) { 593 if (I != MBB->end()) { 594 RPTracker.advanceToNext(); 595 RPTracker.advance(MBB->end()); 596 } 597 RPTracker.advanceBeforeNext(); 598 MBBLiveIns[OnlySucc] = RPTracker.moveLiveRegs(); 599 } 600 } 601 602 DenseMap<MachineInstr *, GCNRPTracker::LiveRegSet> 603 GCNScheduleDAGMILive::getBBLiveInMap() const { 604 assert(!Regions.empty()); 605 std::vector<MachineInstr *> BBStarters; 606 BBStarters.reserve(Regions.size()); 607 auto I = Regions.rbegin(), E = Regions.rend(); 608 auto *BB = I->first->getParent(); 609 do { 610 auto *MI = &*skipDebugInstructionsForward(I->first, I->second); 611 BBStarters.push_back(MI); 612 do { 613 ++I; 614 } while (I != E && I->first->getParent() == BB); 615 } while (I != E); 616 return getLiveRegMap(BBStarters, false /*After*/, *LIS); 617 } 618 619 void GCNScheduleDAGMILive::finalizeSchedule() { 620 // Start actual scheduling here. This function is called by the base 621 // MachineScheduler after all regions have been recorded by 622 // GCNScheduleDAGMILive::schedule(). 623 LiveIns.resize(Regions.size()); 624 Pressure.resize(Regions.size()); 625 RescheduleRegions.resize(Regions.size()); 626 RegionsWithHighRP.resize(Regions.size()); 627 RegionsWithExcessRP.resize(Regions.size()); 628 RegionsWithMinOcc.resize(Regions.size()); 629 RegionsWithIGLPInstrs.resize(Regions.size()); 630 RescheduleRegions.set(); 631 RegionsWithHighRP.reset(); 632 RegionsWithExcessRP.reset(); 633 RegionsWithMinOcc.reset(); 634 RegionsWithIGLPInstrs.reset(); 635 636 runSchedStages(); 637 } 638 639 void GCNScheduleDAGMILive::runSchedStages() { 640 LLVM_DEBUG(dbgs() << "All regions recorded, starting actual scheduling.\n"); 641 642 if (!Regions.empty()) 643 BBLiveInMap = getBBLiveInMap(); 644 645 GCNSchedStrategy &S = static_cast<GCNSchedStrategy &>(*SchedImpl); 646 while (S.advanceStage()) { 647 auto Stage = createSchedStage(S.getCurrentStage()); 648 if (!Stage->initGCNSchedStage()) 649 continue; 650 651 for (auto Region : Regions) { 652 RegionBegin = Region.first; 653 RegionEnd = Region.second; 654 // Setup for scheduling the region and check whether it should be skipped. 655 if (!Stage->initGCNRegion()) { 656 Stage->advanceRegion(); 657 exitRegion(); 658 continue; 659 } 660 661 ScheduleDAGMILive::schedule(); 662 Stage->finalizeGCNRegion(); 663 } 664 665 Stage->finalizeGCNSchedStage(); 666 } 667 } 668 669 #ifndef NDEBUG 670 raw_ostream &llvm::operator<<(raw_ostream &OS, const GCNSchedStageID &StageID) { 671 switch (StageID) { 672 case GCNSchedStageID::OccInitialSchedule: 673 OS << "Max Occupancy Initial Schedule"; 674 break; 675 case GCNSchedStageID::UnclusteredHighRPReschedule: 676 OS << "Unclustered High Register Pressure Reschedule"; 677 break; 678 case GCNSchedStageID::ClusteredLowOccupancyReschedule: 679 OS << "Clustered Low Occupancy Reschedule"; 680 break; 681 case GCNSchedStageID::PreRARematerialize: 682 OS << "Pre-RA Rematerialize"; 683 break; 684 case GCNSchedStageID::ILPInitialSchedule: 685 OS << "Max ILP Initial Schedule"; 686 break; 687 } 688 689 return OS; 690 } 691 #endif 692 693 GCNSchedStage::GCNSchedStage(GCNSchedStageID StageID, GCNScheduleDAGMILive &DAG) 694 : DAG(DAG), S(static_cast<GCNSchedStrategy &>(*DAG.SchedImpl)), MF(DAG.MF), 695 MFI(DAG.MFI), ST(DAG.ST), StageID(StageID) {} 696 697 bool GCNSchedStage::initGCNSchedStage() { 698 if (!DAG.LIS) 699 return false; 700 701 LLVM_DEBUG(dbgs() << "Starting scheduling stage: " << StageID << "\n"); 702 return true; 703 } 704 705 bool UnclusteredHighRPStage::initGCNSchedStage() { 706 if (DisableUnclusterHighRP) 707 return false; 708 709 if (!GCNSchedStage::initGCNSchedStage()) 710 return false; 711 712 if (DAG.RegionsWithHighRP.none() && DAG.RegionsWithExcessRP.none()) 713 return false; 714 715 SavedMutations.swap(DAG.Mutations); 716 DAG.addMutation(createIGroupLPDAGMutation(/*IsPostRA=*/false)); 717 718 InitialOccupancy = DAG.MinOccupancy; 719 // Aggressivly try to reduce register pressure in the unclustered high RP 720 // stage. Temporarily increase occupancy target in the region. 721 S.SGPRLimitBias = S.HighRPSGPRBias; 722 S.VGPRLimitBias = S.HighRPVGPRBias; 723 if (MFI.getMaxWavesPerEU() > DAG.MinOccupancy) 724 MFI.increaseOccupancy(MF, ++DAG.MinOccupancy); 725 726 LLVM_DEBUG( 727 dbgs() 728 << "Retrying function scheduling without clustering. " 729 "Aggressivly try to reduce register pressure to achieve occupancy " 730 << DAG.MinOccupancy << ".\n"); 731 732 return true; 733 } 734 735 bool ClusteredLowOccStage::initGCNSchedStage() { 736 if (DisableClusteredLowOccupancy) 737 return false; 738 739 if (!GCNSchedStage::initGCNSchedStage()) 740 return false; 741 742 // Don't bother trying to improve ILP in lower RP regions if occupancy has not 743 // been dropped. All regions will have already been scheduled with the ideal 744 // occupancy targets. 745 if (DAG.StartingOccupancy <= DAG.MinOccupancy) 746 return false; 747 748 LLVM_DEBUG( 749 dbgs() << "Retrying function scheduling with lowest recorded occupancy " 750 << DAG.MinOccupancy << ".\n"); 751 return true; 752 } 753 754 bool PreRARematStage::initGCNSchedStage() { 755 if (!GCNSchedStage::initGCNSchedStage()) 756 return false; 757 758 if (DAG.RegionsWithMinOcc.none() || DAG.Regions.size() == 1) 759 return false; 760 761 const TargetInstrInfo *TII = MF.getSubtarget().getInstrInfo(); 762 // Check maximum occupancy 763 if (ST.computeOccupancy(MF.getFunction(), MFI.getLDSSize()) == 764 DAG.MinOccupancy) 765 return false; 766 767 // FIXME: This pass will invalidate cached MBBLiveIns for regions 768 // inbetween the defs and region we sinked the def to. Cached pressure 769 // for regions where a def is sinked from will also be invalidated. Will 770 // need to be fixed if there is another pass after this pass. 771 assert(!S.hasNextStage()); 772 773 collectRematerializableInstructions(); 774 if (RematerializableInsts.empty() || !sinkTriviallyRematInsts(ST, TII)) 775 return false; 776 777 LLVM_DEBUG( 778 dbgs() << "Retrying function scheduling with improved occupancy of " 779 << DAG.MinOccupancy << " from rematerializing\n"); 780 return true; 781 } 782 783 void GCNSchedStage::finalizeGCNSchedStage() { 784 DAG.finishBlock(); 785 LLVM_DEBUG(dbgs() << "Ending scheduling stage: " << StageID << "\n"); 786 } 787 788 void UnclusteredHighRPStage::finalizeGCNSchedStage() { 789 SavedMutations.swap(DAG.Mutations); 790 S.SGPRLimitBias = S.VGPRLimitBias = 0; 791 if (DAG.MinOccupancy > InitialOccupancy) { 792 for (unsigned IDX = 0; IDX < DAG.Pressure.size(); ++IDX) 793 DAG.RegionsWithMinOcc[IDX] = 794 DAG.Pressure[IDX].getOccupancy(DAG.ST) == DAG.MinOccupancy; 795 796 LLVM_DEBUG(dbgs() << StageID 797 << " stage successfully increased occupancy to " 798 << DAG.MinOccupancy << '\n'); 799 } 800 801 GCNSchedStage::finalizeGCNSchedStage(); 802 } 803 804 bool GCNSchedStage::initGCNRegion() { 805 // Check whether this new region is also a new block. 806 if (DAG.RegionBegin->getParent() != CurrentMBB) 807 setupNewBlock(); 808 809 unsigned NumRegionInstrs = std::distance(DAG.begin(), DAG.end()); 810 DAG.enterRegion(CurrentMBB, DAG.begin(), DAG.end(), NumRegionInstrs); 811 812 // Skip empty scheduling regions (0 or 1 schedulable instructions). 813 if (DAG.begin() == DAG.end() || DAG.begin() == std::prev(DAG.end())) 814 return false; 815 816 LLVM_DEBUG(dbgs() << "********** MI Scheduling **********\n"); 817 LLVM_DEBUG(dbgs() << MF.getName() << ":" << printMBBReference(*CurrentMBB) 818 << " " << CurrentMBB->getName() 819 << "\n From: " << *DAG.begin() << " To: "; 820 if (DAG.RegionEnd != CurrentMBB->end()) dbgs() << *DAG.RegionEnd; 821 else dbgs() << "End"; 822 dbgs() << " RegionInstrs: " << NumRegionInstrs << '\n'); 823 824 // Save original instruction order before scheduling for possible revert. 825 Unsched.clear(); 826 Unsched.reserve(DAG.NumRegionInstrs); 827 if (StageID == GCNSchedStageID::OccInitialSchedule || 828 StageID == GCNSchedStageID::ILPInitialSchedule) { 829 for (auto &I : DAG) { 830 Unsched.push_back(&I); 831 if (I.getOpcode() == AMDGPU::SCHED_GROUP_BARRIER || 832 I.getOpcode() == AMDGPU::IGLP_OPT) 833 DAG.RegionsWithIGLPInstrs[RegionIdx] = true; 834 } 835 } else { 836 for (auto &I : DAG) 837 Unsched.push_back(&I); 838 } 839 840 PressureBefore = DAG.Pressure[RegionIdx]; 841 842 LLVM_DEBUG( 843 dbgs() << "Pressure before scheduling:\nRegion live-ins:" 844 << print(DAG.LiveIns[RegionIdx], DAG.MRI) 845 << "Region live-in pressure: " 846 << print(llvm::getRegPressure(DAG.MRI, DAG.LiveIns[RegionIdx])) 847 << "Region register pressure: " << print(PressureBefore)); 848 849 S.HasHighPressure = false; 850 S.KnownExcessRP = isRegionWithExcessRP(); 851 852 if (DAG.RegionsWithIGLPInstrs[RegionIdx] && 853 StageID != GCNSchedStageID::UnclusteredHighRPReschedule) { 854 SavedMutations.clear(); 855 SavedMutations.swap(DAG.Mutations); 856 bool IsInitialStage = StageID == GCNSchedStageID::OccInitialSchedule || 857 StageID == GCNSchedStageID::ILPInitialSchedule; 858 DAG.addMutation(createIGroupLPDAGMutation(/*IsReentry=*/!IsInitialStage)); 859 } 860 861 return true; 862 } 863 864 bool UnclusteredHighRPStage::initGCNRegion() { 865 // Only reschedule regions with the minimum occupancy or regions that may have 866 // spilling (excess register pressure). 867 if ((!DAG.RegionsWithMinOcc[RegionIdx] || 868 DAG.MinOccupancy <= InitialOccupancy) && 869 !DAG.RegionsWithExcessRP[RegionIdx]) 870 return false; 871 872 return GCNSchedStage::initGCNRegion(); 873 } 874 875 bool ClusteredLowOccStage::initGCNRegion() { 876 // We may need to reschedule this region if it wasn't rescheduled in the last 877 // stage, or if we found it was testing critical register pressure limits in 878 // the unclustered reschedule stage. The later is because we may not have been 879 // able to raise the min occupancy in the previous stage so the region may be 880 // overly constrained even if it was already rescheduled. 881 if (!DAG.RegionsWithHighRP[RegionIdx]) 882 return false; 883 884 return GCNSchedStage::initGCNRegion(); 885 } 886 887 bool PreRARematStage::initGCNRegion() { 888 if (!DAG.RescheduleRegions[RegionIdx]) 889 return false; 890 891 return GCNSchedStage::initGCNRegion(); 892 } 893 894 void GCNSchedStage::setupNewBlock() { 895 if (CurrentMBB) 896 DAG.finishBlock(); 897 898 CurrentMBB = DAG.RegionBegin->getParent(); 899 DAG.startBlock(CurrentMBB); 900 // Get real RP for the region if it hasn't be calculated before. After the 901 // initial schedule stage real RP will be collected after scheduling. 902 if (StageID == GCNSchedStageID::OccInitialSchedule || 903 StageID == GCNSchedStageID::ILPInitialSchedule) 904 DAG.computeBlockPressure(RegionIdx, CurrentMBB); 905 } 906 907 void GCNSchedStage::finalizeGCNRegion() { 908 DAG.Regions[RegionIdx] = std::pair(DAG.RegionBegin, DAG.RegionEnd); 909 DAG.RescheduleRegions[RegionIdx] = false; 910 if (S.HasHighPressure) 911 DAG.RegionsWithHighRP[RegionIdx] = true; 912 913 // Revert scheduling if we have dropped occupancy or there is some other 914 // reason that the original schedule is better. 915 checkScheduling(); 916 917 if (DAG.RegionsWithIGLPInstrs[RegionIdx] && 918 StageID != GCNSchedStageID::UnclusteredHighRPReschedule) 919 SavedMutations.swap(DAG.Mutations); 920 921 DAG.exitRegion(); 922 RegionIdx++; 923 } 924 925 void GCNSchedStage::checkScheduling() { 926 // Check the results of scheduling. 927 PressureAfter = DAG.getRealRegPressure(RegionIdx); 928 LLVM_DEBUG(dbgs() << "Pressure after scheduling: " << print(PressureAfter)); 929 LLVM_DEBUG(dbgs() << "Region: " << RegionIdx << ".\n"); 930 931 if (PressureAfter.getSGPRNum() <= S.SGPRCriticalLimit && 932 PressureAfter.getVGPRNum(ST.hasGFX90AInsts()) <= S.VGPRCriticalLimit) { 933 DAG.Pressure[RegionIdx] = PressureAfter; 934 DAG.RegionsWithMinOcc[RegionIdx] = 935 PressureAfter.getOccupancy(ST) == DAG.MinOccupancy; 936 937 // Early out if we have achieve the occupancy target. 938 LLVM_DEBUG(dbgs() << "Pressure in desired limits, done.\n"); 939 return; 940 } 941 942 unsigned TargetOccupancy = 943 std::min(S.getTargetOccupancy(), ST.getOccupancyWithLocalMemSize(MF)); 944 unsigned WavesAfter = 945 std::min(TargetOccupancy, PressureAfter.getOccupancy(ST)); 946 unsigned WavesBefore = 947 std::min(TargetOccupancy, PressureBefore.getOccupancy(ST)); 948 LLVM_DEBUG(dbgs() << "Occupancy before scheduling: " << WavesBefore 949 << ", after " << WavesAfter << ".\n"); 950 951 // We may not be able to keep the current target occupancy because of the just 952 // scheduled region. We might still be able to revert scheduling if the 953 // occupancy before was higher, or if the current schedule has register 954 // pressure higher than the excess limits which could lead to more spilling. 955 unsigned NewOccupancy = std::max(WavesAfter, WavesBefore); 956 957 // Allow memory bound functions to drop to 4 waves if not limited by an 958 // attribute. 959 if (WavesAfter < WavesBefore && WavesAfter < DAG.MinOccupancy && 960 WavesAfter >= MFI.getMinAllowedOccupancy()) { 961 LLVM_DEBUG(dbgs() << "Function is memory bound, allow occupancy drop up to " 962 << MFI.getMinAllowedOccupancy() << " waves\n"); 963 NewOccupancy = WavesAfter; 964 } 965 966 if (NewOccupancy < DAG.MinOccupancy) { 967 DAG.MinOccupancy = NewOccupancy; 968 MFI.limitOccupancy(DAG.MinOccupancy); 969 DAG.RegionsWithMinOcc.reset(); 970 LLVM_DEBUG(dbgs() << "Occupancy lowered for the function to " 971 << DAG.MinOccupancy << ".\n"); 972 } 973 974 unsigned MaxVGPRs = ST.getMaxNumVGPRs(MF); 975 unsigned MaxSGPRs = ST.getMaxNumSGPRs(MF); 976 if (PressureAfter.getVGPRNum(false) > MaxVGPRs || 977 PressureAfter.getAGPRNum() > MaxVGPRs || 978 PressureAfter.getSGPRNum() > MaxSGPRs) { 979 DAG.RescheduleRegions[RegionIdx] = true; 980 DAG.RegionsWithHighRP[RegionIdx] = true; 981 DAG.RegionsWithExcessRP[RegionIdx] = true; 982 } 983 984 // Revert if this region's schedule would cause a drop in occupancy or 985 // spilling. 986 if (shouldRevertScheduling(WavesAfter)) { 987 revertScheduling(); 988 } else { 989 DAG.Pressure[RegionIdx] = PressureAfter; 990 DAG.RegionsWithMinOcc[RegionIdx] = 991 PressureAfter.getOccupancy(ST) == DAG.MinOccupancy; 992 } 993 } 994 995 unsigned 996 GCNSchedStage::computeSUnitReadyCycle(const SUnit &SU, unsigned CurrCycle, 997 DenseMap<unsigned, unsigned> &ReadyCycles, 998 const TargetSchedModel &SM) { 999 unsigned ReadyCycle = CurrCycle; 1000 for (auto &D : SU.Preds) { 1001 if (D.isAssignedRegDep()) { 1002 MachineInstr *DefMI = D.getSUnit()->getInstr(); 1003 unsigned Latency = SM.computeInstrLatency(DefMI); 1004 unsigned DefReady = ReadyCycles[DAG.getSUnit(DefMI)->NodeNum]; 1005 ReadyCycle = std::max(ReadyCycle, DefReady + Latency); 1006 } 1007 } 1008 ReadyCycles[SU.NodeNum] = ReadyCycle; 1009 return ReadyCycle; 1010 } 1011 1012 #ifndef NDEBUG 1013 struct EarlierIssuingCycle { 1014 bool operator()(std::pair<MachineInstr *, unsigned> A, 1015 std::pair<MachineInstr *, unsigned> B) const { 1016 return A.second < B.second; 1017 } 1018 }; 1019 1020 static void printScheduleModel(std::set<std::pair<MachineInstr *, unsigned>, 1021 EarlierIssuingCycle> &ReadyCycles) { 1022 if (ReadyCycles.empty()) 1023 return; 1024 unsigned BBNum = ReadyCycles.begin()->first->getParent()->getNumber(); 1025 dbgs() << "\n################## Schedule time ReadyCycles for MBB : " << BBNum 1026 << " ##################\n# Cycle #\t\t\tInstruction " 1027 " " 1028 " \n"; 1029 unsigned IPrev = 1; 1030 for (auto &I : ReadyCycles) { 1031 if (I.second > IPrev + 1) 1032 dbgs() << "****************************** BUBBLE OF " << I.second - IPrev 1033 << " CYCLES DETECTED ******************************\n\n"; 1034 dbgs() << "[ " << I.second << " ] : " << *I.first << "\n"; 1035 IPrev = I.second; 1036 } 1037 } 1038 #endif 1039 1040 ScheduleMetrics 1041 GCNSchedStage::getScheduleMetrics(const std::vector<SUnit> &InputSchedule) { 1042 #ifndef NDEBUG 1043 std::set<std::pair<MachineInstr *, unsigned>, EarlierIssuingCycle> 1044 ReadyCyclesSorted; 1045 #endif 1046 const TargetSchedModel &SM = ST.getInstrInfo()->getSchedModel(); 1047 unsigned SumBubbles = 0; 1048 DenseMap<unsigned, unsigned> ReadyCycles; 1049 unsigned CurrCycle = 0; 1050 for (auto &SU : InputSchedule) { 1051 unsigned ReadyCycle = 1052 computeSUnitReadyCycle(SU, CurrCycle, ReadyCycles, SM); 1053 SumBubbles += ReadyCycle - CurrCycle; 1054 #ifndef NDEBUG 1055 ReadyCyclesSorted.insert(std::make_pair(SU.getInstr(), ReadyCycle)); 1056 #endif 1057 CurrCycle = ++ReadyCycle; 1058 } 1059 #ifndef NDEBUG 1060 LLVM_DEBUG( 1061 printScheduleModel(ReadyCyclesSorted); 1062 dbgs() << "\n\t" 1063 << "Metric: " 1064 << (SumBubbles 1065 ? (SumBubbles * ScheduleMetrics::ScaleFactor) / CurrCycle 1066 : 1) 1067 << "\n\n"); 1068 #endif 1069 1070 return ScheduleMetrics(CurrCycle, SumBubbles); 1071 } 1072 1073 ScheduleMetrics 1074 GCNSchedStage::getScheduleMetrics(const GCNScheduleDAGMILive &DAG) { 1075 #ifndef NDEBUG 1076 std::set<std::pair<MachineInstr *, unsigned>, EarlierIssuingCycle> 1077 ReadyCyclesSorted; 1078 #endif 1079 const TargetSchedModel &SM = ST.getInstrInfo()->getSchedModel(); 1080 unsigned SumBubbles = 0; 1081 DenseMap<unsigned, unsigned> ReadyCycles; 1082 unsigned CurrCycle = 0; 1083 for (auto &MI : DAG) { 1084 SUnit *SU = DAG.getSUnit(&MI); 1085 if (!SU) 1086 continue; 1087 unsigned ReadyCycle = 1088 computeSUnitReadyCycle(*SU, CurrCycle, ReadyCycles, SM); 1089 SumBubbles += ReadyCycle - CurrCycle; 1090 #ifndef NDEBUG 1091 ReadyCyclesSorted.insert(std::make_pair(SU->getInstr(), ReadyCycle)); 1092 #endif 1093 CurrCycle = ++ReadyCycle; 1094 } 1095 #ifndef NDEBUG 1096 LLVM_DEBUG( 1097 printScheduleModel(ReadyCyclesSorted); 1098 dbgs() << "\n\t" 1099 << "Metric: " 1100 << (SumBubbles 1101 ? (SumBubbles * ScheduleMetrics::ScaleFactor) / CurrCycle 1102 : 1) 1103 << "\n\n"); 1104 #endif 1105 1106 return ScheduleMetrics(CurrCycle, SumBubbles); 1107 } 1108 1109 bool GCNSchedStage::shouldRevertScheduling(unsigned WavesAfter) { 1110 if (WavesAfter < DAG.MinOccupancy) 1111 return true; 1112 1113 return false; 1114 } 1115 1116 bool OccInitialScheduleStage::shouldRevertScheduling(unsigned WavesAfter) { 1117 if (PressureAfter == PressureBefore) 1118 return false; 1119 1120 if (GCNSchedStage::shouldRevertScheduling(WavesAfter)) 1121 return true; 1122 1123 if (mayCauseSpilling(WavesAfter)) 1124 return true; 1125 1126 return false; 1127 } 1128 1129 bool UnclusteredHighRPStage::shouldRevertScheduling(unsigned WavesAfter) { 1130 // If RP is not reduced in the unclustered reschedule stage, revert to the 1131 // old schedule. 1132 if ((WavesAfter <= PressureBefore.getOccupancy(ST) && 1133 mayCauseSpilling(WavesAfter)) || 1134 GCNSchedStage::shouldRevertScheduling(WavesAfter)) { 1135 LLVM_DEBUG(dbgs() << "Unclustered reschedule did not help.\n"); 1136 return true; 1137 } 1138 1139 // Do not attempt to relax schedule even more if we are already spilling. 1140 if (isRegionWithExcessRP()) 1141 return false; 1142 1143 LLVM_DEBUG( 1144 dbgs() 1145 << "\n\t *** In shouldRevertScheduling ***\n" 1146 << " *********** BEFORE UnclusteredHighRPStage ***********\n"); 1147 ScheduleMetrics MBefore = 1148 getScheduleMetrics(DAG.SUnits); 1149 LLVM_DEBUG( 1150 dbgs() 1151 << "\n *********** AFTER UnclusteredHighRPStage ***********\n"); 1152 ScheduleMetrics MAfter = getScheduleMetrics(DAG); 1153 unsigned OldMetric = MBefore.getMetric(); 1154 unsigned NewMetric = MAfter.getMetric(); 1155 unsigned WavesBefore = 1156 std::min(S.getTargetOccupancy(), PressureBefore.getOccupancy(ST)); 1157 unsigned Profit = 1158 ((WavesAfter * ScheduleMetrics::ScaleFactor) / WavesBefore * 1159 ((OldMetric + ScheduleMetricBias) * ScheduleMetrics::ScaleFactor) / 1160 NewMetric) / 1161 ScheduleMetrics::ScaleFactor; 1162 LLVM_DEBUG(dbgs() << "\tMetric before " << MBefore << "\tMetric after " 1163 << MAfter << "Profit: " << Profit << "\n"); 1164 return Profit < ScheduleMetrics::ScaleFactor; 1165 } 1166 1167 bool ClusteredLowOccStage::shouldRevertScheduling(unsigned WavesAfter) { 1168 if (PressureAfter == PressureBefore) 1169 return false; 1170 1171 if (GCNSchedStage::shouldRevertScheduling(WavesAfter)) 1172 return true; 1173 1174 if (mayCauseSpilling(WavesAfter)) 1175 return true; 1176 1177 return false; 1178 } 1179 1180 bool PreRARematStage::shouldRevertScheduling(unsigned WavesAfter) { 1181 if (GCNSchedStage::shouldRevertScheduling(WavesAfter)) 1182 return true; 1183 1184 if (mayCauseSpilling(WavesAfter)) 1185 return true; 1186 1187 return false; 1188 } 1189 1190 bool ILPInitialScheduleStage::shouldRevertScheduling(unsigned WavesAfter) { 1191 if (mayCauseSpilling(WavesAfter)) 1192 return true; 1193 1194 return false; 1195 } 1196 1197 bool GCNSchedStage::mayCauseSpilling(unsigned WavesAfter) { 1198 if (WavesAfter <= MFI.getMinWavesPerEU() && 1199 !PressureAfter.less(ST, PressureBefore) && 1200 isRegionWithExcessRP()) { 1201 LLVM_DEBUG(dbgs() << "New pressure will result in more spilling.\n"); 1202 return true; 1203 } 1204 1205 return false; 1206 } 1207 1208 void GCNSchedStage::revertScheduling() { 1209 DAG.RegionsWithMinOcc[RegionIdx] = 1210 PressureBefore.getOccupancy(ST) == DAG.MinOccupancy; 1211 LLVM_DEBUG(dbgs() << "Attempting to revert scheduling.\n"); 1212 DAG.RescheduleRegions[RegionIdx] = 1213 S.hasNextStage() && 1214 S.getNextStage() != GCNSchedStageID::UnclusteredHighRPReschedule; 1215 DAG.RegionEnd = DAG.RegionBegin; 1216 int SkippedDebugInstr = 0; 1217 for (MachineInstr *MI : Unsched) { 1218 if (MI->isDebugInstr()) { 1219 ++SkippedDebugInstr; 1220 continue; 1221 } 1222 1223 if (MI->getIterator() != DAG.RegionEnd) { 1224 DAG.BB->remove(MI); 1225 DAG.BB->insert(DAG.RegionEnd, MI); 1226 if (!MI->isDebugInstr()) 1227 DAG.LIS->handleMove(*MI, true); 1228 } 1229 1230 // Reset read-undef flags and update them later. 1231 for (auto &Op : MI->all_defs()) 1232 Op.setIsUndef(false); 1233 RegisterOperands RegOpers; 1234 RegOpers.collect(*MI, *DAG.TRI, DAG.MRI, DAG.ShouldTrackLaneMasks, false); 1235 if (!MI->isDebugInstr()) { 1236 if (DAG.ShouldTrackLaneMasks) { 1237 // Adjust liveness and add missing dead+read-undef flags. 1238 SlotIndex SlotIdx = DAG.LIS->getInstructionIndex(*MI).getRegSlot(); 1239 RegOpers.adjustLaneLiveness(*DAG.LIS, DAG.MRI, SlotIdx, MI); 1240 } else { 1241 // Adjust for missing dead-def flags. 1242 RegOpers.detectDeadDefs(*MI, *DAG.LIS); 1243 } 1244 } 1245 DAG.RegionEnd = MI->getIterator(); 1246 ++DAG.RegionEnd; 1247 LLVM_DEBUG(dbgs() << "Scheduling " << *MI); 1248 } 1249 1250 // After reverting schedule, debug instrs will now be at the end of the block 1251 // and RegionEnd will point to the first debug instr. Increment RegionEnd 1252 // pass debug instrs to the actual end of the scheduling region. 1253 while (SkippedDebugInstr-- > 0) 1254 ++DAG.RegionEnd; 1255 1256 // If Unsched.front() instruction is a debug instruction, this will actually 1257 // shrink the region since we moved all debug instructions to the end of the 1258 // block. Find the first instruction that is not a debug instruction. 1259 DAG.RegionBegin = Unsched.front()->getIterator(); 1260 if (DAG.RegionBegin->isDebugInstr()) { 1261 for (MachineInstr *MI : Unsched) { 1262 if (MI->isDebugInstr()) 1263 continue; 1264 DAG.RegionBegin = MI->getIterator(); 1265 break; 1266 } 1267 } 1268 1269 // Then move the debug instructions back into their correct place and set 1270 // RegionBegin and RegionEnd if needed. 1271 DAG.placeDebugValues(); 1272 1273 DAG.Regions[RegionIdx] = std::pair(DAG.RegionBegin, DAG.RegionEnd); 1274 } 1275 1276 void PreRARematStage::collectRematerializableInstructions() { 1277 const SIRegisterInfo *SRI = static_cast<const SIRegisterInfo *>(DAG.TRI); 1278 for (unsigned I = 0, E = DAG.MRI.getNumVirtRegs(); I != E; ++I) { 1279 Register Reg = Register::index2VirtReg(I); 1280 if (!DAG.LIS->hasInterval(Reg)) 1281 continue; 1282 1283 // TODO: Handle AGPR and SGPR rematerialization 1284 if (!SRI->isVGPRClass(DAG.MRI.getRegClass(Reg)) || 1285 !DAG.MRI.hasOneDef(Reg) || !DAG.MRI.hasOneNonDBGUse(Reg)) 1286 continue; 1287 1288 MachineOperand *Op = DAG.MRI.getOneDef(Reg); 1289 MachineInstr *Def = Op->getParent(); 1290 if (Op->getSubReg() != 0 || !isTriviallyReMaterializable(*Def)) 1291 continue; 1292 1293 MachineInstr *UseI = &*DAG.MRI.use_instr_nodbg_begin(Reg); 1294 if (Def->getParent() == UseI->getParent()) 1295 continue; 1296 1297 // We are only collecting defs that are defined in another block and are 1298 // live-through or used inside regions at MinOccupancy. This means that the 1299 // register must be in the live-in set for the region. 1300 bool AddedToRematList = false; 1301 for (unsigned I = 0, E = DAG.Regions.size(); I != E; ++I) { 1302 auto It = DAG.LiveIns[I].find(Reg); 1303 if (It != DAG.LiveIns[I].end() && !It->second.none()) { 1304 if (DAG.RegionsWithMinOcc[I]) { 1305 RematerializableInsts[I][Def] = UseI; 1306 AddedToRematList = true; 1307 } 1308 1309 // Collect regions with rematerializable reg as live-in to avoid 1310 // searching later when updating RP. 1311 RematDefToLiveInRegions[Def].push_back(I); 1312 } 1313 } 1314 if (!AddedToRematList) 1315 RematDefToLiveInRegions.erase(Def); 1316 } 1317 } 1318 1319 bool PreRARematStage::sinkTriviallyRematInsts(const GCNSubtarget &ST, 1320 const TargetInstrInfo *TII) { 1321 // Temporary copies of cached variables we will be modifying and replacing if 1322 // sinking succeeds. 1323 SmallVector< 1324 std::pair<MachineBasicBlock::iterator, MachineBasicBlock::iterator>, 32> 1325 NewRegions; 1326 DenseMap<unsigned, GCNRPTracker::LiveRegSet> NewLiveIns; 1327 DenseMap<unsigned, GCNRegPressure> NewPressure; 1328 BitVector NewRescheduleRegions; 1329 LiveIntervals *LIS = DAG.LIS; 1330 1331 NewRegions.resize(DAG.Regions.size()); 1332 NewRescheduleRegions.resize(DAG.Regions.size()); 1333 1334 // Collect only regions that has a rematerializable def as a live-in. 1335 SmallSet<unsigned, 16> ImpactedRegions; 1336 for (const auto &It : RematDefToLiveInRegions) 1337 ImpactedRegions.insert(It.second.begin(), It.second.end()); 1338 1339 // Make copies of register pressure and live-ins cache that will be updated 1340 // as we rematerialize. 1341 for (auto Idx : ImpactedRegions) { 1342 NewPressure[Idx] = DAG.Pressure[Idx]; 1343 NewLiveIns[Idx] = DAG.LiveIns[Idx]; 1344 } 1345 NewRegions = DAG.Regions; 1346 NewRescheduleRegions.reset(); 1347 1348 DenseMap<MachineInstr *, MachineInstr *> InsertedMIToOldDef; 1349 bool Improved = false; 1350 for (auto I : ImpactedRegions) { 1351 if (!DAG.RegionsWithMinOcc[I]) 1352 continue; 1353 1354 Improved = false; 1355 int VGPRUsage = NewPressure[I].getVGPRNum(ST.hasGFX90AInsts()); 1356 int SGPRUsage = NewPressure[I].getSGPRNum(); 1357 1358 // TODO: Handle occupancy drop due to AGPR and SGPR. 1359 // Check if cause of occupancy drop is due to VGPR usage and not SGPR. 1360 if (ST.getOccupancyWithNumSGPRs(SGPRUsage) == DAG.MinOccupancy) 1361 break; 1362 1363 // The occupancy of this region could have been improved by a previous 1364 // iteration's sinking of defs. 1365 if (NewPressure[I].getOccupancy(ST) > DAG.MinOccupancy) { 1366 NewRescheduleRegions[I] = true; 1367 Improved = true; 1368 continue; 1369 } 1370 1371 // First check if we have enough trivially rematerializable instructions to 1372 // improve occupancy. Optimistically assume all instructions we are able to 1373 // sink decreased RP. 1374 int TotalSinkableRegs = 0; 1375 for (const auto &It : RematerializableInsts[I]) { 1376 MachineInstr *Def = It.first; 1377 Register DefReg = Def->getOperand(0).getReg(); 1378 TotalSinkableRegs += 1379 SIRegisterInfo::getNumCoveredRegs(NewLiveIns[I][DefReg]); 1380 } 1381 int VGPRsAfterSink = VGPRUsage - TotalSinkableRegs; 1382 unsigned OptimisticOccupancy = ST.getOccupancyWithNumVGPRs(VGPRsAfterSink); 1383 // If in the most optimistic scenario, we cannot improve occupancy, then do 1384 // not attempt to sink any instructions. 1385 if (OptimisticOccupancy <= DAG.MinOccupancy) 1386 break; 1387 1388 unsigned ImproveOccupancy = 0; 1389 SmallVector<MachineInstr *, 4> SinkedDefs; 1390 for (auto &It : RematerializableInsts[I]) { 1391 MachineInstr *Def = It.first; 1392 MachineBasicBlock::iterator InsertPos = 1393 MachineBasicBlock::iterator(It.second); 1394 Register Reg = Def->getOperand(0).getReg(); 1395 // Rematerialize MI to its use block. Since we are only rematerializing 1396 // instructions that do not have any virtual reg uses, we do not need to 1397 // call LiveRangeEdit::allUsesAvailableAt() and 1398 // LiveRangeEdit::canRematerializeAt(). 1399 TII->reMaterialize(*InsertPos->getParent(), InsertPos, Reg, 1400 Def->getOperand(0).getSubReg(), *Def, *DAG.TRI); 1401 MachineInstr *NewMI = &*std::prev(InsertPos); 1402 LIS->InsertMachineInstrInMaps(*NewMI); 1403 LIS->removeInterval(Reg); 1404 LIS->createAndComputeVirtRegInterval(Reg); 1405 InsertedMIToOldDef[NewMI] = Def; 1406 1407 // Update region boundaries in scheduling region we sinked from since we 1408 // may sink an instruction that was at the beginning or end of its region 1409 DAG.updateRegionBoundaries(NewRegions, Def, /*NewMI =*/nullptr, 1410 /*Removing =*/true); 1411 1412 // Update region boundaries in region we sinked to. 1413 DAG.updateRegionBoundaries(NewRegions, InsertPos, NewMI); 1414 1415 LaneBitmask PrevMask = NewLiveIns[I][Reg]; 1416 // FIXME: Also update cached pressure for where the def was sinked from. 1417 // Update RP for all regions that has this reg as a live-in and remove 1418 // the reg from all regions as a live-in. 1419 for (auto Idx : RematDefToLiveInRegions[Def]) { 1420 NewLiveIns[Idx].erase(Reg); 1421 if (InsertPos->getParent() != DAG.Regions[Idx].first->getParent()) { 1422 // Def is live-through and not used in this block. 1423 NewPressure[Idx].inc(Reg, PrevMask, LaneBitmask::getNone(), DAG.MRI); 1424 } else { 1425 // Def is used and rematerialized into this block. 1426 GCNDownwardRPTracker RPT(*LIS); 1427 auto *NonDbgMI = &*skipDebugInstructionsForward( 1428 NewRegions[Idx].first, NewRegions[Idx].second); 1429 RPT.reset(*NonDbgMI, &NewLiveIns[Idx]); 1430 RPT.advance(NewRegions[Idx].second); 1431 NewPressure[Idx] = RPT.moveMaxPressure(); 1432 } 1433 } 1434 1435 SinkedDefs.push_back(Def); 1436 ImproveOccupancy = NewPressure[I].getOccupancy(ST); 1437 if (ImproveOccupancy > DAG.MinOccupancy) 1438 break; 1439 } 1440 1441 // Remove defs we just sinked from all regions' list of sinkable defs 1442 for (auto &Def : SinkedDefs) 1443 for (auto TrackedIdx : RematDefToLiveInRegions[Def]) 1444 RematerializableInsts[TrackedIdx].erase(Def); 1445 1446 if (ImproveOccupancy <= DAG.MinOccupancy) 1447 break; 1448 1449 NewRescheduleRegions[I] = true; 1450 Improved = true; 1451 } 1452 1453 if (!Improved) { 1454 // Occupancy was not improved for all regions that were at MinOccupancy. 1455 // Undo sinking and remove newly rematerialized instructions. 1456 for (auto &Entry : InsertedMIToOldDef) { 1457 MachineInstr *MI = Entry.first; 1458 MachineInstr *OldMI = Entry.second; 1459 Register Reg = MI->getOperand(0).getReg(); 1460 LIS->RemoveMachineInstrFromMaps(*MI); 1461 MI->eraseFromParent(); 1462 OldMI->clearRegisterDeads(Reg); 1463 LIS->removeInterval(Reg); 1464 LIS->createAndComputeVirtRegInterval(Reg); 1465 } 1466 return false; 1467 } 1468 1469 // Occupancy was improved for all regions. 1470 for (auto &Entry : InsertedMIToOldDef) { 1471 MachineInstr *MI = Entry.first; 1472 MachineInstr *OldMI = Entry.second; 1473 1474 // Remove OldMI from BBLiveInMap since we are sinking it from its MBB. 1475 DAG.BBLiveInMap.erase(OldMI); 1476 1477 // Remove OldMI and update LIS 1478 Register Reg = MI->getOperand(0).getReg(); 1479 LIS->RemoveMachineInstrFromMaps(*OldMI); 1480 OldMI->eraseFromParent(); 1481 LIS->removeInterval(Reg); 1482 LIS->createAndComputeVirtRegInterval(Reg); 1483 } 1484 1485 // Update live-ins, register pressure, and regions caches. 1486 for (auto Idx : ImpactedRegions) { 1487 DAG.LiveIns[Idx] = NewLiveIns[Idx]; 1488 DAG.Pressure[Idx] = NewPressure[Idx]; 1489 DAG.MBBLiveIns.erase(DAG.Regions[Idx].first->getParent()); 1490 } 1491 DAG.Regions = NewRegions; 1492 DAG.RescheduleRegions = NewRescheduleRegions; 1493 1494 SIMachineFunctionInfo &MFI = *MF.getInfo<SIMachineFunctionInfo>(); 1495 MFI.increaseOccupancy(MF, ++DAG.MinOccupancy); 1496 1497 return true; 1498 } 1499 1500 // Copied from MachineLICM 1501 bool PreRARematStage::isTriviallyReMaterializable(const MachineInstr &MI) { 1502 if (!DAG.TII->isTriviallyReMaterializable(MI)) 1503 return false; 1504 1505 for (const MachineOperand &MO : MI.all_uses()) 1506 if (MO.getReg().isVirtual()) 1507 return false; 1508 1509 return true; 1510 } 1511 1512 // When removing, we will have to check both beginning and ending of the region. 1513 // When inserting, we will only have to check if we are inserting NewMI in front 1514 // of a scheduling region and do not need to check the ending since we will only 1515 // ever be inserting before an already existing MI. 1516 void GCNScheduleDAGMILive::updateRegionBoundaries( 1517 SmallVectorImpl<std::pair<MachineBasicBlock::iterator, 1518 MachineBasicBlock::iterator>> &RegionBoundaries, 1519 MachineBasicBlock::iterator MI, MachineInstr *NewMI, bool Removing) { 1520 unsigned I = 0, E = RegionBoundaries.size(); 1521 // Search for first region of the block where MI is located 1522 while (I != E && MI->getParent() != RegionBoundaries[I].first->getParent()) 1523 ++I; 1524 1525 for (; I != E; ++I) { 1526 if (MI->getParent() != RegionBoundaries[I].first->getParent()) 1527 return; 1528 1529 if (Removing && MI == RegionBoundaries[I].first && 1530 MI == RegionBoundaries[I].second) { 1531 // MI is in a region with size 1, after removing, the region will be 1532 // size 0, set RegionBegin and RegionEnd to pass end of block iterator. 1533 RegionBoundaries[I] = 1534 std::pair(MI->getParent()->end(), MI->getParent()->end()); 1535 return; 1536 } 1537 if (MI == RegionBoundaries[I].first) { 1538 if (Removing) 1539 RegionBoundaries[I] = 1540 std::pair(std::next(MI), RegionBoundaries[I].second); 1541 else 1542 // Inserted NewMI in front of region, set new RegionBegin to NewMI 1543 RegionBoundaries[I] = std::pair(MachineBasicBlock::iterator(NewMI), 1544 RegionBoundaries[I].second); 1545 return; 1546 } 1547 if (Removing && MI == RegionBoundaries[I].second) { 1548 RegionBoundaries[I] = std::pair(RegionBoundaries[I].first, std::prev(MI)); 1549 return; 1550 } 1551 } 1552 } 1553 1554 static bool hasIGLPInstrs(ScheduleDAGInstrs *DAG) { 1555 return std::any_of( 1556 DAG->begin(), DAG->end(), [](MachineBasicBlock::iterator MI) { 1557 unsigned Opc = MI->getOpcode(); 1558 return Opc == AMDGPU::SCHED_GROUP_BARRIER || Opc == AMDGPU::IGLP_OPT; 1559 }); 1560 } 1561 1562 GCNPostScheduleDAGMILive::GCNPostScheduleDAGMILive( 1563 MachineSchedContext *C, std::unique_ptr<MachineSchedStrategy> S, 1564 bool RemoveKillFlags) 1565 : ScheduleDAGMI(C, std::move(S), RemoveKillFlags) {} 1566 1567 void GCNPostScheduleDAGMILive::schedule() { 1568 HasIGLPInstrs = hasIGLPInstrs(this); 1569 if (HasIGLPInstrs) { 1570 SavedMutations.clear(); 1571 SavedMutations.swap(Mutations); 1572 addMutation(createIGroupLPDAGMutation(/*IsReentry=*/true)); 1573 } 1574 1575 ScheduleDAGMI::schedule(); 1576 } 1577 1578 void GCNPostScheduleDAGMILive::finalizeSchedule() { 1579 if (HasIGLPInstrs) 1580 SavedMutations.swap(Mutations); 1581 1582 ScheduleDAGMI::finalizeSchedule(); 1583 } 1584