1 //===-- SIMachineScheduler.cpp - SI Scheduler Interface -------------------===// 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 /// SI Machine Scheduler interface 11 // 12 //===----------------------------------------------------------------------===// 13 14 #include "SIMachineScheduler.h" 15 #include "AMDGPU.h" 16 #include "SIInstrInfo.h" 17 #include "SIRegisterInfo.h" 18 #include "MCTargetDesc/AMDGPUMCTargetDesc.h" 19 #include "llvm/ADT/STLExtras.h" 20 #include "llvm/ADT/SmallVector.h" 21 #include "llvm/CodeGen/LiveInterval.h" 22 #include "llvm/CodeGen/LiveIntervals.h" 23 #include "llvm/CodeGen/MachineInstr.h" 24 #include "llvm/CodeGen/MachineRegisterInfo.h" 25 #include "llvm/CodeGen/MachineScheduler.h" 26 #include "llvm/CodeGen/RegisterPressure.h" 27 #include "llvm/CodeGen/SlotIndexes.h" 28 #include "llvm/CodeGen/TargetRegisterInfo.h" 29 #include "llvm/Support/Debug.h" 30 #include "llvm/Support/ErrorHandling.h" 31 #include "llvm/Support/raw_ostream.h" 32 #include <algorithm> 33 #include <cassert> 34 #include <map> 35 #include <set> 36 #include <utility> 37 #include <vector> 38 39 using namespace llvm; 40 41 #define DEBUG_TYPE "machine-scheduler" 42 43 // This scheduler implements a different scheduling algorithm than 44 // GenericScheduler. 45 // 46 // There are several specific architecture behaviours that can't be modelled 47 // for GenericScheduler: 48 // . When accessing the result of an SGPR load instruction, you have to wait 49 // for all the SGPR load instructions before your current instruction to 50 // have finished. 51 // . When accessing the result of an VGPR load instruction, you have to wait 52 // for all the VGPR load instructions previous to the VGPR load instruction 53 // you are interested in to finish. 54 // . The less the register pressure, the best load latencies are hidden 55 // 56 // Moreover some specifities (like the fact a lot of instructions in the shader 57 // have few dependencies) makes the generic scheduler have some unpredictable 58 // behaviours. For example when register pressure becomes high, it can either 59 // manage to prevent register pressure from going too high, or it can 60 // increase register pressure even more than if it hadn't taken register 61 // pressure into account. 62 // 63 // Also some other bad behaviours are generated, like loading at the beginning 64 // of the shader a constant in VGPR you won't need until the end of the shader. 65 // 66 // The scheduling problem for SI can distinguish three main parts: 67 // . Hiding high latencies (texture sampling, etc) 68 // . Hiding low latencies (SGPR constant loading, etc) 69 // . Keeping register usage low for better latency hiding and general 70 // performance 71 // 72 // Some other things can also affect performance, but are hard to predict 73 // (cache usage, the fact the HW can issue several instructions from different 74 // wavefronts if different types, etc) 75 // 76 // This scheduler tries to solve the scheduling problem by dividing it into 77 // simpler sub-problems. It divides the instructions into blocks, schedules 78 // locally inside the blocks where it takes care of low latencies, and then 79 // chooses the order of the blocks by taking care of high latencies. 80 // Dividing the instructions into blocks helps control keeping register 81 // usage low. 82 // 83 // First the instructions are put into blocks. 84 // We want the blocks help control register usage and hide high latencies 85 // later. To help control register usage, we typically want all local 86 // computations, when for example you create a result that can be comsummed 87 // right away, to be contained in a block. Block inputs and outputs would 88 // typically be important results that are needed in several locations of 89 // the shader. Since we do want blocks to help hide high latencies, we want 90 // the instructions inside the block to have a minimal set of dependencies 91 // on high latencies. It will make it easy to pick blocks to hide specific 92 // high latencies. 93 // The block creation algorithm is divided into several steps, and several 94 // variants can be tried during the scheduling process. 95 // 96 // Second the order of the instructions inside the blocks is chosen. 97 // At that step we do take into account only register usage and hiding 98 // low latency instructions 99 // 100 // Third the block order is chosen, there we try to hide high latencies 101 // and keep register usage low. 102 // 103 // After the third step, a pass is done to improve the hiding of low 104 // latencies. 105 // 106 // Actually when talking about 'low latency' or 'high latency' it includes 107 // both the latency to get the cache (or global mem) data go to the register, 108 // and the bandwidth limitations. 109 // Increasing the number of active wavefronts helps hide the former, but it 110 // doesn't solve the latter, thus why even if wavefront count is high, we have 111 // to try have as many instructions hiding high latencies as possible. 112 // The OpenCL doc says for example latency of 400 cycles for a global mem access, 113 // which is hidden by 10 instructions if the wavefront count is 10. 114 115 // Some figures taken from AMD docs: 116 // Both texture and constant L1 caches are 4-way associative with 64 bytes 117 // lines. 118 // Constant cache is shared with 4 CUs. 119 // For texture sampling, the address generation unit receives 4 texture 120 // addresses per cycle, thus we could expect texture sampling latency to be 121 // equivalent to 4 instructions in the very best case (a VGPR is 64 work items, 122 // instructions in a wavefront group are executed every 4 cycles), 123 // or 16 instructions if the other wavefronts associated to the 3 other VALUs 124 // of the CU do texture sampling too. (Don't take these figures too seriously, 125 // as I'm not 100% sure of the computation) 126 // Data exports should get similar latency. 127 // For constant loading, the cache is shader with 4 CUs. 128 // The doc says "a throughput of 16B/cycle for each of the 4 Compute Unit" 129 // I guess if the other CU don't read the cache, it can go up to 64B/cycle. 130 // It means a simple s_buffer_load should take one instruction to hide, as 131 // well as a s_buffer_loadx2 and potentially a s_buffer_loadx8 if on the same 132 // cache line. 133 // 134 // As of today the driver doesn't preload the constants in cache, thus the 135 // first loads get extra latency. The doc says global memory access can be 136 // 300-600 cycles. We do not specially take that into account when scheduling 137 // As we expect the driver to be able to preload the constants soon. 138 139 // common code // 140 141 #ifndef NDEBUG 142 143 static const char *getReasonStr(SIScheduleCandReason Reason) { 144 switch (Reason) { 145 case NoCand: return "NOCAND"; 146 case RegUsage: return "REGUSAGE"; 147 case Latency: return "LATENCY"; 148 case Successor: return "SUCCESSOR"; 149 case Depth: return "DEPTH"; 150 case NodeOrder: return "ORDER"; 151 } 152 llvm_unreachable("Unknown reason!"); 153 } 154 155 #endif 156 157 namespace llvm { 158 namespace SISched { 159 static bool tryLess(int TryVal, int CandVal, 160 SISchedulerCandidate &TryCand, 161 SISchedulerCandidate &Cand, 162 SIScheduleCandReason Reason) { 163 if (TryVal < CandVal) { 164 TryCand.Reason = Reason; 165 return true; 166 } 167 if (TryVal > CandVal) { 168 if (Cand.Reason > Reason) 169 Cand.Reason = Reason; 170 return true; 171 } 172 Cand.setRepeat(Reason); 173 return false; 174 } 175 176 static bool tryGreater(int TryVal, int CandVal, 177 SISchedulerCandidate &TryCand, 178 SISchedulerCandidate &Cand, 179 SIScheduleCandReason Reason) { 180 if (TryVal > CandVal) { 181 TryCand.Reason = Reason; 182 return true; 183 } 184 if (TryVal < CandVal) { 185 if (Cand.Reason > Reason) 186 Cand.Reason = Reason; 187 return true; 188 } 189 Cand.setRepeat(Reason); 190 return false; 191 } 192 } // end namespace SISched 193 } // end namespace llvm 194 195 // SIScheduleBlock // 196 197 void SIScheduleBlock::addUnit(SUnit *SU) { 198 NodeNum2Index[SU->NodeNum] = SUnits.size(); 199 SUnits.push_back(SU); 200 } 201 202 #ifndef NDEBUG 203 void SIScheduleBlock::traceCandidate(const SISchedCandidate &Cand) { 204 205 dbgs() << " SU(" << Cand.SU->NodeNum << ") " << getReasonStr(Cand.Reason); 206 dbgs() << '\n'; 207 } 208 #endif 209 210 void SIScheduleBlock::tryCandidateTopDown(SISchedCandidate &Cand, 211 SISchedCandidate &TryCand) { 212 // Initialize the candidate if needed. 213 if (!Cand.isValid()) { 214 TryCand.Reason = NodeOrder; 215 return; 216 } 217 218 if (Cand.SGPRUsage > 60 && 219 SISched::tryLess(TryCand.SGPRUsage, Cand.SGPRUsage, 220 TryCand, Cand, RegUsage)) 221 return; 222 223 // Schedule low latency instructions as top as possible. 224 // Order of priority is: 225 // . Low latency instructions which do not depend on other low latency 226 // instructions we haven't waited for 227 // . Other instructions which do not depend on low latency instructions 228 // we haven't waited for 229 // . Low latencies 230 // . All other instructions 231 // Goal is to get: low latency instructions - independent instructions 232 // - (eventually some more low latency instructions) 233 // - instructions that depend on the first low latency instructions. 234 // If in the block there is a lot of constant loads, the SGPR usage 235 // could go quite high, thus above the arbitrary limit of 60 will encourage 236 // use the already loaded constants (in order to release some SGPRs) before 237 // loading more. 238 if (SISched::tryLess(TryCand.HasLowLatencyNonWaitedParent, 239 Cand.HasLowLatencyNonWaitedParent, 240 TryCand, Cand, SIScheduleCandReason::Depth)) 241 return; 242 243 if (SISched::tryGreater(TryCand.IsLowLatency, Cand.IsLowLatency, 244 TryCand, Cand, SIScheduleCandReason::Depth)) 245 return; 246 247 if (TryCand.IsLowLatency && 248 SISched::tryLess(TryCand.LowLatencyOffset, Cand.LowLatencyOffset, 249 TryCand, Cand, SIScheduleCandReason::Depth)) 250 return; 251 252 if (SISched::tryLess(TryCand.VGPRUsage, Cand.VGPRUsage, 253 TryCand, Cand, RegUsage)) 254 return; 255 256 // Fall through to original instruction order. 257 if (TryCand.SU->NodeNum < Cand.SU->NodeNum) { 258 TryCand.Reason = NodeOrder; 259 } 260 } 261 262 SUnit* SIScheduleBlock::pickNode() { 263 SISchedCandidate TopCand; 264 265 for (SUnit* SU : TopReadySUs) { 266 SISchedCandidate TryCand; 267 std::vector<unsigned> pressure; 268 std::vector<unsigned> MaxPressure; 269 // Predict register usage after this instruction. 270 TryCand.SU = SU; 271 TopRPTracker.getDownwardPressure(SU->getInstr(), pressure, MaxPressure); 272 TryCand.SGPRUsage = pressure[AMDGPU::RegisterPressureSets::SReg_32]; 273 TryCand.VGPRUsage = pressure[AMDGPU::RegisterPressureSets::VGPR_32]; 274 TryCand.IsLowLatency = DAG->IsLowLatencySU[SU->NodeNum]; 275 TryCand.LowLatencyOffset = DAG->LowLatencyOffset[SU->NodeNum]; 276 TryCand.HasLowLatencyNonWaitedParent = 277 HasLowLatencyNonWaitedParent[NodeNum2Index[SU->NodeNum]]; 278 tryCandidateTopDown(TopCand, TryCand); 279 if (TryCand.Reason != NoCand) 280 TopCand.setBest(TryCand); 281 } 282 283 return TopCand.SU; 284 } 285 286 287 // Schedule something valid. 288 void SIScheduleBlock::fastSchedule() { 289 TopReadySUs.clear(); 290 if (Scheduled) 291 undoSchedule(); 292 293 for (SUnit* SU : SUnits) { 294 if (!SU->NumPredsLeft) 295 TopReadySUs.push_back(SU); 296 } 297 298 while (!TopReadySUs.empty()) { 299 SUnit *SU = TopReadySUs[0]; 300 ScheduledSUnits.push_back(SU); 301 nodeScheduled(SU); 302 } 303 304 Scheduled = true; 305 } 306 307 // Returns if the register was set between first and last. 308 static bool isDefBetween(unsigned Reg, 309 SlotIndex First, SlotIndex Last, 310 const MachineRegisterInfo *MRI, 311 const LiveIntervals *LIS) { 312 for (MachineRegisterInfo::def_instr_iterator 313 UI = MRI->def_instr_begin(Reg), 314 UE = MRI->def_instr_end(); UI != UE; ++UI) { 315 const MachineInstr* MI = &*UI; 316 if (MI->isDebugValue()) 317 continue; 318 SlotIndex InstSlot = LIS->getInstructionIndex(*MI).getRegSlot(); 319 if (InstSlot >= First && InstSlot <= Last) 320 return true; 321 } 322 return false; 323 } 324 325 void SIScheduleBlock::initRegPressure(MachineBasicBlock::iterator BeginBlock, 326 MachineBasicBlock::iterator EndBlock) { 327 IntervalPressure Pressure, BotPressure; 328 RegPressureTracker RPTracker(Pressure), BotRPTracker(BotPressure); 329 LiveIntervals *LIS = DAG->getLIS(); 330 MachineRegisterInfo *MRI = DAG->getMRI(); 331 DAG->initRPTracker(TopRPTracker); 332 DAG->initRPTracker(BotRPTracker); 333 DAG->initRPTracker(RPTracker); 334 335 // Goes though all SU. RPTracker captures what had to be alive for the SUs 336 // to execute, and what is still alive at the end. 337 for (SUnit* SU : ScheduledSUnits) { 338 RPTracker.setPos(SU->getInstr()); 339 RPTracker.advance(); 340 } 341 342 // Close the RPTracker to finalize live ins/outs. 343 RPTracker.closeRegion(); 344 345 // Initialize the live ins and live outs. 346 TopRPTracker.addLiveRegs(RPTracker.getPressure().LiveInRegs); 347 BotRPTracker.addLiveRegs(RPTracker.getPressure().LiveOutRegs); 348 349 // Do not Track Physical Registers, because it messes up. 350 for (const auto &RegMaskPair : RPTracker.getPressure().LiveInRegs) { 351 if (Register::isVirtualRegister(RegMaskPair.RegUnit)) 352 LiveInRegs.insert(RegMaskPair.RegUnit); 353 } 354 LiveOutRegs.clear(); 355 // There is several possibilities to distinguish: 356 // 1) Reg is not input to any instruction in the block, but is output of one 357 // 2) 1) + read in the block and not needed after it 358 // 3) 1) + read in the block but needed in another block 359 // 4) Reg is input of an instruction but another block will read it too 360 // 5) Reg is input of an instruction and then rewritten in the block. 361 // result is not read in the block (implies used in another block) 362 // 6) Reg is input of an instruction and then rewritten in the block. 363 // result is read in the block and not needed in another block 364 // 7) Reg is input of an instruction and then rewritten in the block. 365 // result is read in the block but also needed in another block 366 // LiveInRegs will contains all the regs in situation 4, 5, 6, 7 367 // We want LiveOutRegs to contain only Regs whose content will be read after 368 // in another block, and whose content was written in the current block, 369 // that is we want it to get 1, 3, 5, 7 370 // Since we made the MIs of a block to be packed all together before 371 // scheduling, then the LiveIntervals were correct, and the RPTracker was 372 // able to correctly handle 5 vs 6, 2 vs 3. 373 // (Note: This is not sufficient for RPTracker to not do mistakes for case 4) 374 // The RPTracker's LiveOutRegs has 1, 3, (some correct or incorrect)4, 5, 7 375 // Comparing to LiveInRegs is not sufficient to differenciate 4 vs 5, 7 376 // The use of findDefBetween removes the case 4. 377 for (const auto &RegMaskPair : RPTracker.getPressure().LiveOutRegs) { 378 unsigned Reg = RegMaskPair.RegUnit; 379 if (Register::isVirtualRegister(Reg) && 380 isDefBetween(Reg, LIS->getInstructionIndex(*BeginBlock).getRegSlot(), 381 LIS->getInstructionIndex(*EndBlock).getRegSlot(), MRI, 382 LIS)) { 383 LiveOutRegs.insert(Reg); 384 } 385 } 386 387 // Pressure = sum_alive_registers register size 388 // Internally llvm will represent some registers as big 128 bits registers 389 // for example, but they actually correspond to 4 actual 32 bits registers. 390 // Thus Pressure is not equal to num_alive_registers * constant. 391 LiveInPressure = TopPressure.MaxSetPressure; 392 LiveOutPressure = BotPressure.MaxSetPressure; 393 394 // Prepares TopRPTracker for top down scheduling. 395 TopRPTracker.closeTop(); 396 } 397 398 void SIScheduleBlock::schedule(MachineBasicBlock::iterator BeginBlock, 399 MachineBasicBlock::iterator EndBlock) { 400 if (!Scheduled) 401 fastSchedule(); 402 403 // PreScheduling phase to set LiveIn and LiveOut. 404 initRegPressure(BeginBlock, EndBlock); 405 undoSchedule(); 406 407 // Schedule for real now. 408 409 TopReadySUs.clear(); 410 411 for (SUnit* SU : SUnits) { 412 if (!SU->NumPredsLeft) 413 TopReadySUs.push_back(SU); 414 } 415 416 while (!TopReadySUs.empty()) { 417 SUnit *SU = pickNode(); 418 ScheduledSUnits.push_back(SU); 419 TopRPTracker.setPos(SU->getInstr()); 420 TopRPTracker.advance(); 421 nodeScheduled(SU); 422 } 423 424 // TODO: compute InternalAdditionnalPressure. 425 InternalAdditionnalPressure.resize(TopPressure.MaxSetPressure.size()); 426 427 // Check everything is right. 428 #ifndef NDEBUG 429 assert(SUnits.size() == ScheduledSUnits.size() && 430 TopReadySUs.empty()); 431 for (SUnit* SU : SUnits) { 432 assert(SU->isScheduled && 433 SU->NumPredsLeft == 0); 434 } 435 #endif 436 437 Scheduled = true; 438 } 439 440 void SIScheduleBlock::undoSchedule() { 441 for (SUnit* SU : SUnits) { 442 SU->isScheduled = false; 443 for (SDep& Succ : SU->Succs) { 444 if (BC->isSUInBlock(Succ.getSUnit(), ID)) 445 undoReleaseSucc(SU, &Succ); 446 } 447 } 448 HasLowLatencyNonWaitedParent.assign(SUnits.size(), 0); 449 ScheduledSUnits.clear(); 450 Scheduled = false; 451 } 452 453 void SIScheduleBlock::undoReleaseSucc(SUnit *SU, SDep *SuccEdge) { 454 SUnit *SuccSU = SuccEdge->getSUnit(); 455 456 if (SuccEdge->isWeak()) { 457 ++SuccSU->WeakPredsLeft; 458 return; 459 } 460 ++SuccSU->NumPredsLeft; 461 } 462 463 void SIScheduleBlock::releaseSucc(SUnit *SU, SDep *SuccEdge) { 464 SUnit *SuccSU = SuccEdge->getSUnit(); 465 466 if (SuccEdge->isWeak()) { 467 --SuccSU->WeakPredsLeft; 468 return; 469 } 470 #ifndef NDEBUG 471 if (SuccSU->NumPredsLeft == 0) { 472 dbgs() << "*** Scheduling failed! ***\n"; 473 DAG->dumpNode(*SuccSU); 474 dbgs() << " has been released too many times!\n"; 475 llvm_unreachable(nullptr); 476 } 477 #endif 478 479 --SuccSU->NumPredsLeft; 480 } 481 482 /// Release Successors of the SU that are in the block or not. 483 void SIScheduleBlock::releaseSuccessors(SUnit *SU, bool InOrOutBlock) { 484 for (SDep& Succ : SU->Succs) { 485 SUnit *SuccSU = Succ.getSUnit(); 486 487 if (SuccSU->NodeNum >= DAG->SUnits.size()) 488 continue; 489 490 if (BC->isSUInBlock(SuccSU, ID) != InOrOutBlock) 491 continue; 492 493 releaseSucc(SU, &Succ); 494 if (SuccSU->NumPredsLeft == 0 && InOrOutBlock) 495 TopReadySUs.push_back(SuccSU); 496 } 497 } 498 499 void SIScheduleBlock::nodeScheduled(SUnit *SU) { 500 // Is in TopReadySUs 501 assert (!SU->NumPredsLeft); 502 std::vector<SUnit *>::iterator I = llvm::find(TopReadySUs, SU); 503 if (I == TopReadySUs.end()) { 504 dbgs() << "Data Structure Bug in SI Scheduler\n"; 505 llvm_unreachable(nullptr); 506 } 507 TopReadySUs.erase(I); 508 509 releaseSuccessors(SU, true); 510 // Scheduling this node will trigger a wait, 511 // thus propagate to other instructions that they do not need to wait either. 512 if (HasLowLatencyNonWaitedParent[NodeNum2Index[SU->NodeNum]]) 513 HasLowLatencyNonWaitedParent.assign(SUnits.size(), 0); 514 515 if (DAG->IsLowLatencySU[SU->NodeNum]) { 516 for (SDep& Succ : SU->Succs) { 517 std::map<unsigned, unsigned>::iterator I = 518 NodeNum2Index.find(Succ.getSUnit()->NodeNum); 519 if (I != NodeNum2Index.end()) 520 HasLowLatencyNonWaitedParent[I->second] = 1; 521 } 522 } 523 SU->isScheduled = true; 524 } 525 526 void SIScheduleBlock::finalizeUnits() { 527 // We remove links from outside blocks to enable scheduling inside the block. 528 for (SUnit* SU : SUnits) { 529 releaseSuccessors(SU, false); 530 if (DAG->IsHighLatencySU[SU->NodeNum]) 531 HighLatencyBlock = true; 532 } 533 HasLowLatencyNonWaitedParent.resize(SUnits.size(), 0); 534 } 535 536 // we maintain ascending order of IDs 537 void SIScheduleBlock::addPred(SIScheduleBlock *Pred) { 538 unsigned PredID = Pred->getID(); 539 540 // Check if not already predecessor. 541 for (SIScheduleBlock* P : Preds) { 542 if (PredID == P->getID()) 543 return; 544 } 545 Preds.push_back(Pred); 546 547 assert(none_of(Succs, 548 [=](std::pair<SIScheduleBlock*, 549 SIScheduleBlockLinkKind> S) { 550 return PredID == S.first->getID(); 551 }) && 552 "Loop in the Block Graph!"); 553 } 554 555 void SIScheduleBlock::addSucc(SIScheduleBlock *Succ, 556 SIScheduleBlockLinkKind Kind) { 557 unsigned SuccID = Succ->getID(); 558 559 // Check if not already predecessor. 560 for (std::pair<SIScheduleBlock*, SIScheduleBlockLinkKind> &S : Succs) { 561 if (SuccID == S.first->getID()) { 562 if (S.second == SIScheduleBlockLinkKind::NoData && 563 Kind == SIScheduleBlockLinkKind::Data) 564 S.second = Kind; 565 return; 566 } 567 } 568 if (Succ->isHighLatencyBlock()) 569 ++NumHighLatencySuccessors; 570 Succs.push_back(std::make_pair(Succ, Kind)); 571 572 assert(none_of(Preds, 573 [=](SIScheduleBlock *P) { return SuccID == P->getID(); }) && 574 "Loop in the Block Graph!"); 575 } 576 577 #ifndef NDEBUG 578 void SIScheduleBlock::printDebug(bool full) { 579 dbgs() << "Block (" << ID << ")\n"; 580 if (!full) 581 return; 582 583 dbgs() << "\nContains High Latency Instruction: " 584 << HighLatencyBlock << '\n'; 585 dbgs() << "\nDepends On:\n"; 586 for (SIScheduleBlock* P : Preds) { 587 P->printDebug(false); 588 } 589 590 dbgs() << "\nSuccessors:\n"; 591 for (std::pair<SIScheduleBlock*, SIScheduleBlockLinkKind> S : Succs) { 592 if (S.second == SIScheduleBlockLinkKind::Data) 593 dbgs() << "(Data Dep) "; 594 S.first->printDebug(false); 595 } 596 597 if (Scheduled) { 598 dbgs() << "LiveInPressure " 599 << LiveInPressure[AMDGPU::RegisterPressureSets::SReg_32] << ' ' 600 << LiveInPressure[AMDGPU::RegisterPressureSets::VGPR_32] << '\n'; 601 dbgs() << "LiveOutPressure " 602 << LiveOutPressure[AMDGPU::RegisterPressureSets::SReg_32] << ' ' 603 << LiveOutPressure[AMDGPU::RegisterPressureSets::VGPR_32] << "\n\n"; 604 dbgs() << "LiveIns:\n"; 605 for (unsigned Reg : LiveInRegs) 606 dbgs() << printVRegOrUnit(Reg, DAG->getTRI()) << ' '; 607 608 dbgs() << "\nLiveOuts:\n"; 609 for (unsigned Reg : LiveOutRegs) 610 dbgs() << printVRegOrUnit(Reg, DAG->getTRI()) << ' '; 611 } 612 613 dbgs() << "\nInstructions:\n"; 614 for (const SUnit* SU : SUnits) 615 DAG->dumpNode(*SU); 616 617 dbgs() << "///////////////////////\n"; 618 } 619 #endif 620 621 // SIScheduleBlockCreator // 622 623 SIScheduleBlockCreator::SIScheduleBlockCreator(SIScheduleDAGMI *DAG) 624 : DAG(DAG) {} 625 626 SIScheduleBlocks 627 SIScheduleBlockCreator::getBlocks(SISchedulerBlockCreatorVariant BlockVariant) { 628 std::map<SISchedulerBlockCreatorVariant, SIScheduleBlocks>::iterator B = 629 Blocks.find(BlockVariant); 630 if (B == Blocks.end()) { 631 SIScheduleBlocks Res; 632 createBlocksForVariant(BlockVariant); 633 topologicalSort(); 634 scheduleInsideBlocks(); 635 fillStats(); 636 Res.Blocks = CurrentBlocks; 637 Res.TopDownIndex2Block = TopDownIndex2Block; 638 Res.TopDownBlock2Index = TopDownBlock2Index; 639 Blocks[BlockVariant] = Res; 640 return Res; 641 } else { 642 return B->second; 643 } 644 } 645 646 bool SIScheduleBlockCreator::isSUInBlock(SUnit *SU, unsigned ID) { 647 if (SU->NodeNum >= DAG->SUnits.size()) 648 return false; 649 return CurrentBlocks[Node2CurrentBlock[SU->NodeNum]]->getID() == ID; 650 } 651 652 void SIScheduleBlockCreator::colorHighLatenciesAlone() { 653 unsigned DAGSize = DAG->SUnits.size(); 654 655 for (unsigned i = 0, e = DAGSize; i != e; ++i) { 656 SUnit *SU = &DAG->SUnits[i]; 657 if (DAG->IsHighLatencySU[SU->NodeNum]) { 658 CurrentColoring[SU->NodeNum] = NextReservedID++; 659 } 660 } 661 } 662 663 static bool 664 hasDataDependencyPred(const SUnit &SU, const SUnit &FromSU) { 665 for (const auto &PredDep : SU.Preds) { 666 if (PredDep.getSUnit() == &FromSU && 667 PredDep.getKind() == llvm::SDep::Data) 668 return true; 669 } 670 return false; 671 } 672 673 void SIScheduleBlockCreator::colorHighLatenciesGroups() { 674 unsigned DAGSize = DAG->SUnits.size(); 675 unsigned NumHighLatencies = 0; 676 unsigned GroupSize; 677 int Color = NextReservedID; 678 unsigned Count = 0; 679 std::set<unsigned> FormingGroup; 680 681 for (unsigned i = 0, e = DAGSize; i != e; ++i) { 682 SUnit *SU = &DAG->SUnits[i]; 683 if (DAG->IsHighLatencySU[SU->NodeNum]) 684 ++NumHighLatencies; 685 } 686 687 if (NumHighLatencies == 0) 688 return; 689 690 if (NumHighLatencies <= 6) 691 GroupSize = 2; 692 else if (NumHighLatencies <= 12) 693 GroupSize = 3; 694 else 695 GroupSize = 4; 696 697 for (unsigned SUNum : DAG->TopDownIndex2SU) { 698 const SUnit &SU = DAG->SUnits[SUNum]; 699 if (DAG->IsHighLatencySU[SU.NodeNum]) { 700 unsigned CompatibleGroup = true; 701 int ProposedColor = Color; 702 std::vector<int> AdditionalElements; 703 704 // We don't want to put in the same block 705 // two high latency instructions that depend 706 // on each other. 707 // One way would be to check canAddEdge 708 // in both directions, but that currently is not 709 // enough because there the high latency order is 710 // enforced (via links). 711 // Instead, look at the dependencies between the 712 // high latency instructions and deduce if it is 713 // a data dependency or not. 714 for (unsigned j : FormingGroup) { 715 bool HasSubGraph; 716 std::vector<int> SubGraph; 717 // By construction (topological order), if SU and 718 // DAG->SUnits[j] are linked, DAG->SUnits[j] is neccessary 719 // in the parent graph of SU. 720 #ifndef NDEBUG 721 SubGraph = DAG->GetTopo()->GetSubGraph(SU, DAG->SUnits[j], 722 HasSubGraph); 723 assert(!HasSubGraph); 724 #endif 725 SubGraph = DAG->GetTopo()->GetSubGraph(DAG->SUnits[j], SU, 726 HasSubGraph); 727 if (!HasSubGraph) 728 continue; // No dependencies between each other 729 else if (SubGraph.size() > 5) { 730 // Too many elements would be required to be added to the block. 731 CompatibleGroup = false; 732 break; 733 } 734 else { 735 // Check the type of dependency 736 for (unsigned k : SubGraph) { 737 // If in the path to join the two instructions, 738 // there is another high latency instruction, 739 // or instructions colored for another block 740 // abort the merge. 741 if (DAG->IsHighLatencySU[k] || 742 (CurrentColoring[k] != ProposedColor && 743 CurrentColoring[k] != 0)) { 744 CompatibleGroup = false; 745 break; 746 } 747 // If one of the SU in the subgraph depends on the result of SU j, 748 // there'll be a data dependency. 749 if (hasDataDependencyPred(DAG->SUnits[k], DAG->SUnits[j])) { 750 CompatibleGroup = false; 751 break; 752 } 753 } 754 if (!CompatibleGroup) 755 break; 756 // Same check for the SU 757 if (hasDataDependencyPred(SU, DAG->SUnits[j])) { 758 CompatibleGroup = false; 759 break; 760 } 761 // Add all the required instructions to the block 762 // These cannot live in another block (because they 763 // depend (order dependency) on one of the 764 // instruction in the block, and are required for the 765 // high latency instruction we add. 766 AdditionalElements.insert(AdditionalElements.end(), 767 SubGraph.begin(), SubGraph.end()); 768 } 769 } 770 if (CompatibleGroup) { 771 FormingGroup.insert(SU.NodeNum); 772 for (unsigned j : AdditionalElements) 773 CurrentColoring[j] = ProposedColor; 774 CurrentColoring[SU.NodeNum] = ProposedColor; 775 ++Count; 776 } 777 // Found one incompatible instruction, 778 // or has filled a big enough group. 779 // -> start a new one. 780 if (!CompatibleGroup) { 781 FormingGroup.clear(); 782 Color = ++NextReservedID; 783 ProposedColor = Color; 784 FormingGroup.insert(SU.NodeNum); 785 CurrentColoring[SU.NodeNum] = ProposedColor; 786 Count = 0; 787 } else if (Count == GroupSize) { 788 FormingGroup.clear(); 789 Color = ++NextReservedID; 790 ProposedColor = Color; 791 Count = 0; 792 } 793 } 794 } 795 } 796 797 void SIScheduleBlockCreator::colorComputeReservedDependencies() { 798 unsigned DAGSize = DAG->SUnits.size(); 799 std::map<std::set<unsigned>, unsigned> ColorCombinations; 800 801 CurrentTopDownReservedDependencyColoring.clear(); 802 CurrentBottomUpReservedDependencyColoring.clear(); 803 804 CurrentTopDownReservedDependencyColoring.resize(DAGSize, 0); 805 CurrentBottomUpReservedDependencyColoring.resize(DAGSize, 0); 806 807 // Traverse TopDown, and give different colors to SUs depending 808 // on which combination of High Latencies they depend on. 809 810 for (unsigned SUNum : DAG->TopDownIndex2SU) { 811 SUnit *SU = &DAG->SUnits[SUNum]; 812 std::set<unsigned> SUColors; 813 814 // Already given. 815 if (CurrentColoring[SU->NodeNum]) { 816 CurrentTopDownReservedDependencyColoring[SU->NodeNum] = 817 CurrentColoring[SU->NodeNum]; 818 continue; 819 } 820 821 for (SDep& PredDep : SU->Preds) { 822 SUnit *Pred = PredDep.getSUnit(); 823 if (PredDep.isWeak() || Pred->NodeNum >= DAGSize) 824 continue; 825 if (CurrentTopDownReservedDependencyColoring[Pred->NodeNum] > 0) 826 SUColors.insert(CurrentTopDownReservedDependencyColoring[Pred->NodeNum]); 827 } 828 // Color 0 by default. 829 if (SUColors.empty()) 830 continue; 831 // Same color than parents. 832 if (SUColors.size() == 1 && *SUColors.begin() > DAGSize) 833 CurrentTopDownReservedDependencyColoring[SU->NodeNum] = 834 *SUColors.begin(); 835 else { 836 std::map<std::set<unsigned>, unsigned>::iterator Pos = 837 ColorCombinations.find(SUColors); 838 if (Pos != ColorCombinations.end()) { 839 CurrentTopDownReservedDependencyColoring[SU->NodeNum] = Pos->second; 840 } else { 841 CurrentTopDownReservedDependencyColoring[SU->NodeNum] = 842 NextNonReservedID; 843 ColorCombinations[SUColors] = NextNonReservedID++; 844 } 845 } 846 } 847 848 ColorCombinations.clear(); 849 850 // Same as before, but BottomUp. 851 852 for (unsigned SUNum : DAG->BottomUpIndex2SU) { 853 SUnit *SU = &DAG->SUnits[SUNum]; 854 std::set<unsigned> SUColors; 855 856 // Already given. 857 if (CurrentColoring[SU->NodeNum]) { 858 CurrentBottomUpReservedDependencyColoring[SU->NodeNum] = 859 CurrentColoring[SU->NodeNum]; 860 continue; 861 } 862 863 for (SDep& SuccDep : SU->Succs) { 864 SUnit *Succ = SuccDep.getSUnit(); 865 if (SuccDep.isWeak() || Succ->NodeNum >= DAGSize) 866 continue; 867 if (CurrentBottomUpReservedDependencyColoring[Succ->NodeNum] > 0) 868 SUColors.insert(CurrentBottomUpReservedDependencyColoring[Succ->NodeNum]); 869 } 870 // Keep color 0. 871 if (SUColors.empty()) 872 continue; 873 // Same color than parents. 874 if (SUColors.size() == 1 && *SUColors.begin() > DAGSize) 875 CurrentBottomUpReservedDependencyColoring[SU->NodeNum] = 876 *SUColors.begin(); 877 else { 878 std::map<std::set<unsigned>, unsigned>::iterator Pos = 879 ColorCombinations.find(SUColors); 880 if (Pos != ColorCombinations.end()) { 881 CurrentBottomUpReservedDependencyColoring[SU->NodeNum] = Pos->second; 882 } else { 883 CurrentBottomUpReservedDependencyColoring[SU->NodeNum] = 884 NextNonReservedID; 885 ColorCombinations[SUColors] = NextNonReservedID++; 886 } 887 } 888 } 889 } 890 891 void SIScheduleBlockCreator::colorAccordingToReservedDependencies() { 892 unsigned DAGSize = DAG->SUnits.size(); 893 std::map<std::pair<unsigned, unsigned>, unsigned> ColorCombinations; 894 895 // Every combination of colors given by the top down 896 // and bottom up Reserved node dependency 897 898 for (unsigned i = 0, e = DAGSize; i != e; ++i) { 899 SUnit *SU = &DAG->SUnits[i]; 900 std::pair<unsigned, unsigned> SUColors; 901 902 // High latency instructions: already given. 903 if (CurrentColoring[SU->NodeNum]) 904 continue; 905 906 SUColors.first = CurrentTopDownReservedDependencyColoring[SU->NodeNum]; 907 SUColors.second = CurrentBottomUpReservedDependencyColoring[SU->NodeNum]; 908 909 std::map<std::pair<unsigned, unsigned>, unsigned>::iterator Pos = 910 ColorCombinations.find(SUColors); 911 if (Pos != ColorCombinations.end()) { 912 CurrentColoring[SU->NodeNum] = Pos->second; 913 } else { 914 CurrentColoring[SU->NodeNum] = NextNonReservedID; 915 ColorCombinations[SUColors] = NextNonReservedID++; 916 } 917 } 918 } 919 920 void SIScheduleBlockCreator::colorEndsAccordingToDependencies() { 921 unsigned DAGSize = DAG->SUnits.size(); 922 std::vector<int> PendingColoring = CurrentColoring; 923 924 assert(DAGSize >= 1 && 925 CurrentBottomUpReservedDependencyColoring.size() == DAGSize && 926 CurrentTopDownReservedDependencyColoring.size() == DAGSize); 927 // If there is no reserved block at all, do nothing. We don't want 928 // everything in one block. 929 if (*std::max_element(CurrentBottomUpReservedDependencyColoring.begin(), 930 CurrentBottomUpReservedDependencyColoring.end()) == 0 && 931 *std::max_element(CurrentTopDownReservedDependencyColoring.begin(), 932 CurrentTopDownReservedDependencyColoring.end()) == 0) 933 return; 934 935 for (unsigned SUNum : DAG->BottomUpIndex2SU) { 936 SUnit *SU = &DAG->SUnits[SUNum]; 937 std::set<unsigned> SUColors; 938 std::set<unsigned> SUColorsPending; 939 940 if (CurrentColoring[SU->NodeNum] <= (int)DAGSize) 941 continue; 942 943 if (CurrentBottomUpReservedDependencyColoring[SU->NodeNum] > 0 || 944 CurrentTopDownReservedDependencyColoring[SU->NodeNum] > 0) 945 continue; 946 947 for (SDep& SuccDep : SU->Succs) { 948 SUnit *Succ = SuccDep.getSUnit(); 949 if (SuccDep.isWeak() || Succ->NodeNum >= DAGSize) 950 continue; 951 if (CurrentBottomUpReservedDependencyColoring[Succ->NodeNum] > 0 || 952 CurrentTopDownReservedDependencyColoring[Succ->NodeNum] > 0) 953 SUColors.insert(CurrentColoring[Succ->NodeNum]); 954 SUColorsPending.insert(PendingColoring[Succ->NodeNum]); 955 } 956 // If there is only one child/parent block, and that block 957 // is not among the ones we are removing in this path, then 958 // merge the instruction to that block 959 if (SUColors.size() == 1 && SUColorsPending.size() == 1) 960 PendingColoring[SU->NodeNum] = *SUColors.begin(); 961 else // TODO: Attribute new colors depending on color 962 // combination of children. 963 PendingColoring[SU->NodeNum] = NextNonReservedID++; 964 } 965 CurrentColoring = PendingColoring; 966 } 967 968 969 void SIScheduleBlockCreator::colorForceConsecutiveOrderInGroup() { 970 unsigned DAGSize = DAG->SUnits.size(); 971 unsigned PreviousColor; 972 std::set<unsigned> SeenColors; 973 974 if (DAGSize <= 1) 975 return; 976 977 PreviousColor = CurrentColoring[0]; 978 979 for (unsigned i = 1, e = DAGSize; i != e; ++i) { 980 SUnit *SU = &DAG->SUnits[i]; 981 unsigned CurrentColor = CurrentColoring[i]; 982 unsigned PreviousColorSave = PreviousColor; 983 assert(i == SU->NodeNum); 984 985 if (CurrentColor != PreviousColor) 986 SeenColors.insert(PreviousColor); 987 PreviousColor = CurrentColor; 988 989 if (CurrentColoring[SU->NodeNum] <= (int)DAGSize) 990 continue; 991 992 if (SeenColors.find(CurrentColor) == SeenColors.end()) 993 continue; 994 995 if (PreviousColorSave != CurrentColor) 996 CurrentColoring[i] = NextNonReservedID++; 997 else 998 CurrentColoring[i] = CurrentColoring[i-1]; 999 } 1000 } 1001 1002 void SIScheduleBlockCreator::colorMergeConstantLoadsNextGroup() { 1003 unsigned DAGSize = DAG->SUnits.size(); 1004 1005 for (unsigned SUNum : DAG->BottomUpIndex2SU) { 1006 SUnit *SU = &DAG->SUnits[SUNum]; 1007 std::set<unsigned> SUColors; 1008 1009 if (CurrentColoring[SU->NodeNum] <= (int)DAGSize) 1010 continue; 1011 1012 // No predecessor: Vgpr constant loading. 1013 // Low latency instructions usually have a predecessor (the address) 1014 if (SU->Preds.size() > 0 && !DAG->IsLowLatencySU[SU->NodeNum]) 1015 continue; 1016 1017 for (SDep& SuccDep : SU->Succs) { 1018 SUnit *Succ = SuccDep.getSUnit(); 1019 if (SuccDep.isWeak() || Succ->NodeNum >= DAGSize) 1020 continue; 1021 SUColors.insert(CurrentColoring[Succ->NodeNum]); 1022 } 1023 if (SUColors.size() == 1) 1024 CurrentColoring[SU->NodeNum] = *SUColors.begin(); 1025 } 1026 } 1027 1028 void SIScheduleBlockCreator::colorMergeIfPossibleNextGroup() { 1029 unsigned DAGSize = DAG->SUnits.size(); 1030 1031 for (unsigned SUNum : DAG->BottomUpIndex2SU) { 1032 SUnit *SU = &DAG->SUnits[SUNum]; 1033 std::set<unsigned> SUColors; 1034 1035 if (CurrentColoring[SU->NodeNum] <= (int)DAGSize) 1036 continue; 1037 1038 for (SDep& SuccDep : SU->Succs) { 1039 SUnit *Succ = SuccDep.getSUnit(); 1040 if (SuccDep.isWeak() || Succ->NodeNum >= DAGSize) 1041 continue; 1042 SUColors.insert(CurrentColoring[Succ->NodeNum]); 1043 } 1044 if (SUColors.size() == 1) 1045 CurrentColoring[SU->NodeNum] = *SUColors.begin(); 1046 } 1047 } 1048 1049 void SIScheduleBlockCreator::colorMergeIfPossibleNextGroupOnlyForReserved() { 1050 unsigned DAGSize = DAG->SUnits.size(); 1051 1052 for (unsigned SUNum : DAG->BottomUpIndex2SU) { 1053 SUnit *SU = &DAG->SUnits[SUNum]; 1054 std::set<unsigned> SUColors; 1055 1056 if (CurrentColoring[SU->NodeNum] <= (int)DAGSize) 1057 continue; 1058 1059 for (SDep& SuccDep : SU->Succs) { 1060 SUnit *Succ = SuccDep.getSUnit(); 1061 if (SuccDep.isWeak() || Succ->NodeNum >= DAGSize) 1062 continue; 1063 SUColors.insert(CurrentColoring[Succ->NodeNum]); 1064 } 1065 if (SUColors.size() == 1 && *SUColors.begin() <= DAGSize) 1066 CurrentColoring[SU->NodeNum] = *SUColors.begin(); 1067 } 1068 } 1069 1070 void SIScheduleBlockCreator::colorMergeIfPossibleSmallGroupsToNextGroup() { 1071 unsigned DAGSize = DAG->SUnits.size(); 1072 std::map<unsigned, unsigned> ColorCount; 1073 1074 for (unsigned SUNum : DAG->BottomUpIndex2SU) { 1075 SUnit *SU = &DAG->SUnits[SUNum]; 1076 unsigned color = CurrentColoring[SU->NodeNum]; 1077 ++ColorCount[color]; 1078 } 1079 1080 for (unsigned SUNum : DAG->BottomUpIndex2SU) { 1081 SUnit *SU = &DAG->SUnits[SUNum]; 1082 unsigned color = CurrentColoring[SU->NodeNum]; 1083 std::set<unsigned> SUColors; 1084 1085 if (CurrentColoring[SU->NodeNum] <= (int)DAGSize) 1086 continue; 1087 1088 if (ColorCount[color] > 1) 1089 continue; 1090 1091 for (SDep& SuccDep : SU->Succs) { 1092 SUnit *Succ = SuccDep.getSUnit(); 1093 if (SuccDep.isWeak() || Succ->NodeNum >= DAGSize) 1094 continue; 1095 SUColors.insert(CurrentColoring[Succ->NodeNum]); 1096 } 1097 if (SUColors.size() == 1 && *SUColors.begin() != color) { 1098 --ColorCount[color]; 1099 CurrentColoring[SU->NodeNum] = *SUColors.begin(); 1100 ++ColorCount[*SUColors.begin()]; 1101 } 1102 } 1103 } 1104 1105 void SIScheduleBlockCreator::cutHugeBlocks() { 1106 // TODO 1107 } 1108 1109 void SIScheduleBlockCreator::regroupNoUserInstructions() { 1110 unsigned DAGSize = DAG->SUnits.size(); 1111 int GroupID = NextNonReservedID++; 1112 1113 for (unsigned SUNum : DAG->BottomUpIndex2SU) { 1114 SUnit *SU = &DAG->SUnits[SUNum]; 1115 bool hasSuccessor = false; 1116 1117 if (CurrentColoring[SU->NodeNum] <= (int)DAGSize) 1118 continue; 1119 1120 for (SDep& SuccDep : SU->Succs) { 1121 SUnit *Succ = SuccDep.getSUnit(); 1122 if (SuccDep.isWeak() || Succ->NodeNum >= DAGSize) 1123 continue; 1124 hasSuccessor = true; 1125 } 1126 if (!hasSuccessor) 1127 CurrentColoring[SU->NodeNum] = GroupID; 1128 } 1129 } 1130 1131 void SIScheduleBlockCreator::colorExports() { 1132 unsigned ExportColor = NextNonReservedID++; 1133 SmallVector<unsigned, 8> ExpGroup; 1134 1135 // Put all exports together in a block. 1136 // The block will naturally end up being scheduled last, 1137 // thus putting exports at the end of the schedule, which 1138 // is better for performance. 1139 // However we must ensure, for safety, the exports can be put 1140 // together in the same block without any other instruction. 1141 // This could happen, for example, when scheduling after regalloc 1142 // if reloading a spilled register from memory using the same 1143 // register than used in a previous export. 1144 // If that happens, do not regroup the exports. 1145 for (unsigned SUNum : DAG->TopDownIndex2SU) { 1146 const SUnit &SU = DAG->SUnits[SUNum]; 1147 if (SIInstrInfo::isEXP(*SU.getInstr())) { 1148 // Check the EXP can be added to the group safely, 1149 // ie without needing any other instruction. 1150 // The EXP is allowed to depend on other EXP 1151 // (they will be in the same group). 1152 for (unsigned j : ExpGroup) { 1153 bool HasSubGraph; 1154 std::vector<int> SubGraph; 1155 // By construction (topological order), if SU and 1156 // DAG->SUnits[j] are linked, DAG->SUnits[j] is neccessary 1157 // in the parent graph of SU. 1158 #ifndef NDEBUG 1159 SubGraph = DAG->GetTopo()->GetSubGraph(SU, DAG->SUnits[j], 1160 HasSubGraph); 1161 assert(!HasSubGraph); 1162 #endif 1163 SubGraph = DAG->GetTopo()->GetSubGraph(DAG->SUnits[j], SU, 1164 HasSubGraph); 1165 if (!HasSubGraph) 1166 continue; // No dependencies between each other 1167 1168 // SubGraph contains all the instructions required 1169 // between EXP SUnits[j] and EXP SU. 1170 for (unsigned k : SubGraph) { 1171 if (!SIInstrInfo::isEXP(*DAG->SUnits[k].getInstr())) 1172 // Other instructions than EXP would be required in the group. 1173 // Abort the groupping. 1174 return; 1175 } 1176 } 1177 1178 ExpGroup.push_back(SUNum); 1179 } 1180 } 1181 1182 // The group can be formed. Give the color. 1183 for (unsigned j : ExpGroup) 1184 CurrentColoring[j] = ExportColor; 1185 } 1186 1187 void SIScheduleBlockCreator::createBlocksForVariant(SISchedulerBlockCreatorVariant BlockVariant) { 1188 unsigned DAGSize = DAG->SUnits.size(); 1189 std::map<unsigned,unsigned> RealID; 1190 1191 CurrentBlocks.clear(); 1192 CurrentColoring.clear(); 1193 CurrentColoring.resize(DAGSize, 0); 1194 Node2CurrentBlock.clear(); 1195 1196 // Restore links previous scheduling variant has overridden. 1197 DAG->restoreSULinksLeft(); 1198 1199 NextReservedID = 1; 1200 NextNonReservedID = DAGSize + 1; 1201 1202 LLVM_DEBUG(dbgs() << "Coloring the graph\n"); 1203 1204 if (BlockVariant == SISchedulerBlockCreatorVariant::LatenciesGrouped) 1205 colorHighLatenciesGroups(); 1206 else 1207 colorHighLatenciesAlone(); 1208 colorComputeReservedDependencies(); 1209 colorAccordingToReservedDependencies(); 1210 colorEndsAccordingToDependencies(); 1211 if (BlockVariant == SISchedulerBlockCreatorVariant::LatenciesAlonePlusConsecutive) 1212 colorForceConsecutiveOrderInGroup(); 1213 regroupNoUserInstructions(); 1214 colorMergeConstantLoadsNextGroup(); 1215 colorMergeIfPossibleNextGroupOnlyForReserved(); 1216 colorExports(); 1217 1218 // Put SUs of same color into same block 1219 Node2CurrentBlock.resize(DAGSize, -1); 1220 for (unsigned i = 0, e = DAGSize; i != e; ++i) { 1221 SUnit *SU = &DAG->SUnits[i]; 1222 unsigned Color = CurrentColoring[SU->NodeNum]; 1223 if (RealID.find(Color) == RealID.end()) { 1224 int ID = CurrentBlocks.size(); 1225 BlockPtrs.push_back(std::make_unique<SIScheduleBlock>(DAG, this, ID)); 1226 CurrentBlocks.push_back(BlockPtrs.rbegin()->get()); 1227 RealID[Color] = ID; 1228 } 1229 CurrentBlocks[RealID[Color]]->addUnit(SU); 1230 Node2CurrentBlock[SU->NodeNum] = RealID[Color]; 1231 } 1232 1233 // Build dependencies between blocks. 1234 for (unsigned i = 0, e = DAGSize; i != e; ++i) { 1235 SUnit *SU = &DAG->SUnits[i]; 1236 int SUID = Node2CurrentBlock[i]; 1237 for (SDep& SuccDep : SU->Succs) { 1238 SUnit *Succ = SuccDep.getSUnit(); 1239 if (SuccDep.isWeak() || Succ->NodeNum >= DAGSize) 1240 continue; 1241 if (Node2CurrentBlock[Succ->NodeNum] != SUID) 1242 CurrentBlocks[SUID]->addSucc(CurrentBlocks[Node2CurrentBlock[Succ->NodeNum]], 1243 SuccDep.isCtrl() ? NoData : Data); 1244 } 1245 for (SDep& PredDep : SU->Preds) { 1246 SUnit *Pred = PredDep.getSUnit(); 1247 if (PredDep.isWeak() || Pred->NodeNum >= DAGSize) 1248 continue; 1249 if (Node2CurrentBlock[Pred->NodeNum] != SUID) 1250 CurrentBlocks[SUID]->addPred(CurrentBlocks[Node2CurrentBlock[Pred->NodeNum]]); 1251 } 1252 } 1253 1254 // Free root and leafs of all blocks to enable scheduling inside them. 1255 for (unsigned i = 0, e = CurrentBlocks.size(); i != e; ++i) { 1256 SIScheduleBlock *Block = CurrentBlocks[i]; 1257 Block->finalizeUnits(); 1258 } 1259 LLVM_DEBUG(dbgs() << "Blocks created:\n\n"; 1260 for (unsigned i = 0, e = CurrentBlocks.size(); i != e; ++i) { 1261 SIScheduleBlock *Block = CurrentBlocks[i]; 1262 Block->printDebug(true); 1263 }); 1264 } 1265 1266 // Two functions taken from Codegen/MachineScheduler.cpp 1267 1268 /// Non-const version. 1269 static MachineBasicBlock::iterator 1270 nextIfDebug(MachineBasicBlock::iterator I, 1271 MachineBasicBlock::const_iterator End) { 1272 for (; I != End; ++I) { 1273 if (!I->isDebugInstr()) 1274 break; 1275 } 1276 return I; 1277 } 1278 1279 void SIScheduleBlockCreator::topologicalSort() { 1280 unsigned DAGSize = CurrentBlocks.size(); 1281 std::vector<int> WorkList; 1282 1283 LLVM_DEBUG(dbgs() << "Topological Sort\n"); 1284 1285 WorkList.reserve(DAGSize); 1286 TopDownIndex2Block.resize(DAGSize); 1287 TopDownBlock2Index.resize(DAGSize); 1288 BottomUpIndex2Block.resize(DAGSize); 1289 1290 for (unsigned i = 0, e = DAGSize; i != e; ++i) { 1291 SIScheduleBlock *Block = CurrentBlocks[i]; 1292 unsigned Degree = Block->getSuccs().size(); 1293 TopDownBlock2Index[i] = Degree; 1294 if (Degree == 0) { 1295 WorkList.push_back(i); 1296 } 1297 } 1298 1299 int Id = DAGSize; 1300 while (!WorkList.empty()) { 1301 int i = WorkList.back(); 1302 SIScheduleBlock *Block = CurrentBlocks[i]; 1303 WorkList.pop_back(); 1304 TopDownBlock2Index[i] = --Id; 1305 TopDownIndex2Block[Id] = i; 1306 for (SIScheduleBlock* Pred : Block->getPreds()) { 1307 if (!--TopDownBlock2Index[Pred->getID()]) 1308 WorkList.push_back(Pred->getID()); 1309 } 1310 } 1311 1312 #ifndef NDEBUG 1313 // Check correctness of the ordering. 1314 for (unsigned i = 0, e = DAGSize; i != e; ++i) { 1315 SIScheduleBlock *Block = CurrentBlocks[i]; 1316 for (SIScheduleBlock* Pred : Block->getPreds()) { 1317 assert(TopDownBlock2Index[i] > TopDownBlock2Index[Pred->getID()] && 1318 "Wrong Top Down topological sorting"); 1319 } 1320 } 1321 #endif 1322 1323 BottomUpIndex2Block = std::vector<int>(TopDownIndex2Block.rbegin(), 1324 TopDownIndex2Block.rend()); 1325 } 1326 1327 void SIScheduleBlockCreator::scheduleInsideBlocks() { 1328 unsigned DAGSize = CurrentBlocks.size(); 1329 1330 LLVM_DEBUG(dbgs() << "\nScheduling Blocks\n\n"); 1331 1332 // We do schedule a valid scheduling such that a Block corresponds 1333 // to a range of instructions. 1334 LLVM_DEBUG(dbgs() << "First phase: Fast scheduling for Reg Liveness\n"); 1335 for (unsigned i = 0, e = DAGSize; i != e; ++i) { 1336 SIScheduleBlock *Block = CurrentBlocks[i]; 1337 Block->fastSchedule(); 1338 } 1339 1340 // Note: the following code, and the part restoring previous position 1341 // is by far the most expensive operation of the Scheduler. 1342 1343 // Do not update CurrentTop. 1344 MachineBasicBlock::iterator CurrentTopFastSched = DAG->getCurrentTop(); 1345 std::vector<MachineBasicBlock::iterator> PosOld; 1346 std::vector<MachineBasicBlock::iterator> PosNew; 1347 PosOld.reserve(DAG->SUnits.size()); 1348 PosNew.reserve(DAG->SUnits.size()); 1349 1350 for (unsigned i = 0, e = DAGSize; i != e; ++i) { 1351 int BlockIndice = TopDownIndex2Block[i]; 1352 SIScheduleBlock *Block = CurrentBlocks[BlockIndice]; 1353 std::vector<SUnit*> SUs = Block->getScheduledUnits(); 1354 1355 for (SUnit* SU : SUs) { 1356 MachineInstr *MI = SU->getInstr(); 1357 MachineBasicBlock::iterator Pos = MI; 1358 PosOld.push_back(Pos); 1359 if (&*CurrentTopFastSched == MI) { 1360 PosNew.push_back(Pos); 1361 CurrentTopFastSched = nextIfDebug(++CurrentTopFastSched, 1362 DAG->getCurrentBottom()); 1363 } else { 1364 // Update the instruction stream. 1365 DAG->getBB()->splice(CurrentTopFastSched, DAG->getBB(), MI); 1366 1367 // Update LiveIntervals. 1368 // Note: Moving all instructions and calling handleMove every time 1369 // is the most cpu intensive operation of the scheduler. 1370 // It would gain a lot if there was a way to recompute the 1371 // LiveIntervals for the entire scheduling region. 1372 DAG->getLIS()->handleMove(*MI, /*UpdateFlags=*/true); 1373 PosNew.push_back(CurrentTopFastSched); 1374 } 1375 } 1376 } 1377 1378 // Now we have Block of SUs == Block of MI. 1379 // We do the final schedule for the instructions inside the block. 1380 // The property that all the SUs of the Block are grouped together as MI 1381 // is used for correct reg usage tracking. 1382 for (unsigned i = 0, e = DAGSize; i != e; ++i) { 1383 SIScheduleBlock *Block = CurrentBlocks[i]; 1384 std::vector<SUnit*> SUs = Block->getScheduledUnits(); 1385 Block->schedule((*SUs.begin())->getInstr(), (*SUs.rbegin())->getInstr()); 1386 } 1387 1388 LLVM_DEBUG(dbgs() << "Restoring MI Pos\n"); 1389 // Restore old ordering (which prevents a LIS->handleMove bug). 1390 for (unsigned i = PosOld.size(), e = 0; i != e; --i) { 1391 MachineBasicBlock::iterator POld = PosOld[i-1]; 1392 MachineBasicBlock::iterator PNew = PosNew[i-1]; 1393 if (PNew != POld) { 1394 // Update the instruction stream. 1395 DAG->getBB()->splice(POld, DAG->getBB(), PNew); 1396 1397 // Update LiveIntervals. 1398 DAG->getLIS()->handleMove(*POld, /*UpdateFlags=*/true); 1399 } 1400 } 1401 1402 LLVM_DEBUG(for (unsigned i = 0, e = CurrentBlocks.size(); i != e; ++i) { 1403 SIScheduleBlock *Block = CurrentBlocks[i]; 1404 Block->printDebug(true); 1405 }); 1406 } 1407 1408 void SIScheduleBlockCreator::fillStats() { 1409 unsigned DAGSize = CurrentBlocks.size(); 1410 1411 for (unsigned i = 0, e = DAGSize; i != e; ++i) { 1412 int BlockIndice = TopDownIndex2Block[i]; 1413 SIScheduleBlock *Block = CurrentBlocks[BlockIndice]; 1414 if (Block->getPreds().empty()) 1415 Block->Depth = 0; 1416 else { 1417 unsigned Depth = 0; 1418 for (SIScheduleBlock *Pred : Block->getPreds()) { 1419 if (Depth < Pred->Depth + Pred->getCost()) 1420 Depth = Pred->Depth + Pred->getCost(); 1421 } 1422 Block->Depth = Depth; 1423 } 1424 } 1425 1426 for (unsigned i = 0, e = DAGSize; i != e; ++i) { 1427 int BlockIndice = BottomUpIndex2Block[i]; 1428 SIScheduleBlock *Block = CurrentBlocks[BlockIndice]; 1429 if (Block->getSuccs().empty()) 1430 Block->Height = 0; 1431 else { 1432 unsigned Height = 0; 1433 for (const auto &Succ : Block->getSuccs()) 1434 Height = std::max(Height, Succ.first->Height + Succ.first->getCost()); 1435 Block->Height = Height; 1436 } 1437 } 1438 } 1439 1440 // SIScheduleBlockScheduler // 1441 1442 SIScheduleBlockScheduler::SIScheduleBlockScheduler(SIScheduleDAGMI *DAG, 1443 SISchedulerBlockSchedulerVariant Variant, 1444 SIScheduleBlocks BlocksStruct) : 1445 DAG(DAG), Variant(Variant), Blocks(BlocksStruct.Blocks), 1446 LastPosWaitedHighLatency(0), NumBlockScheduled(0), VregCurrentUsage(0), 1447 SregCurrentUsage(0), maxVregUsage(0), maxSregUsage(0) { 1448 1449 // Fill the usage of every output 1450 // Warning: while by construction we always have a link between two blocks 1451 // when one needs a result from the other, the number of users of an output 1452 // is not the sum of child blocks having as input the same virtual register. 1453 // Here is an example. A produces x and y. B eats x and produces x'. 1454 // C eats x' and y. The register coalescer may have attributed the same 1455 // virtual register to x and x'. 1456 // To count accurately, we do a topological sort. In case the register is 1457 // found for several parents, we increment the usage of the one with the 1458 // highest topological index. 1459 LiveOutRegsNumUsages.resize(Blocks.size()); 1460 for (unsigned i = 0, e = Blocks.size(); i != e; ++i) { 1461 SIScheduleBlock *Block = Blocks[i]; 1462 for (unsigned Reg : Block->getInRegs()) { 1463 bool Found = false; 1464 int topoInd = -1; 1465 for (SIScheduleBlock* Pred: Block->getPreds()) { 1466 std::set<unsigned> PredOutRegs = Pred->getOutRegs(); 1467 std::set<unsigned>::iterator RegPos = PredOutRegs.find(Reg); 1468 1469 if (RegPos != PredOutRegs.end()) { 1470 Found = true; 1471 if (topoInd < BlocksStruct.TopDownBlock2Index[Pred->getID()]) { 1472 topoInd = BlocksStruct.TopDownBlock2Index[Pred->getID()]; 1473 } 1474 } 1475 } 1476 1477 if (!Found) 1478 continue; 1479 1480 int PredID = BlocksStruct.TopDownIndex2Block[topoInd]; 1481 ++LiveOutRegsNumUsages[PredID][Reg]; 1482 } 1483 } 1484 1485 LastPosHighLatencyParentScheduled.resize(Blocks.size(), 0); 1486 BlockNumPredsLeft.resize(Blocks.size()); 1487 BlockNumSuccsLeft.resize(Blocks.size()); 1488 1489 for (unsigned i = 0, e = Blocks.size(); i != e; ++i) { 1490 SIScheduleBlock *Block = Blocks[i]; 1491 BlockNumPredsLeft[i] = Block->getPreds().size(); 1492 BlockNumSuccsLeft[i] = Block->getSuccs().size(); 1493 } 1494 1495 #ifndef NDEBUG 1496 for (unsigned i = 0, e = Blocks.size(); i != e; ++i) { 1497 SIScheduleBlock *Block = Blocks[i]; 1498 assert(Block->getID() == i); 1499 } 1500 #endif 1501 1502 std::set<unsigned> InRegs = DAG->getInRegs(); 1503 addLiveRegs(InRegs); 1504 1505 // Increase LiveOutRegsNumUsages for blocks 1506 // producing registers consumed in another 1507 // scheduling region. 1508 for (unsigned Reg : DAG->getOutRegs()) { 1509 for (unsigned i = 0, e = Blocks.size(); i != e; ++i) { 1510 // Do reverse traversal 1511 int ID = BlocksStruct.TopDownIndex2Block[Blocks.size()-1-i]; 1512 SIScheduleBlock *Block = Blocks[ID]; 1513 const std::set<unsigned> &OutRegs = Block->getOutRegs(); 1514 1515 if (OutRegs.find(Reg) == OutRegs.end()) 1516 continue; 1517 1518 ++LiveOutRegsNumUsages[ID][Reg]; 1519 break; 1520 } 1521 } 1522 1523 // Fill LiveRegsConsumers for regs that were already 1524 // defined before scheduling. 1525 for (unsigned i = 0, e = Blocks.size(); i != e; ++i) { 1526 SIScheduleBlock *Block = Blocks[i]; 1527 for (unsigned Reg : Block->getInRegs()) { 1528 bool Found = false; 1529 for (SIScheduleBlock* Pred: Block->getPreds()) { 1530 std::set<unsigned> PredOutRegs = Pred->getOutRegs(); 1531 std::set<unsigned>::iterator RegPos = PredOutRegs.find(Reg); 1532 1533 if (RegPos != PredOutRegs.end()) { 1534 Found = true; 1535 break; 1536 } 1537 } 1538 1539 if (!Found) 1540 ++LiveRegsConsumers[Reg]; 1541 } 1542 } 1543 1544 for (unsigned i = 0, e = Blocks.size(); i != e; ++i) { 1545 SIScheduleBlock *Block = Blocks[i]; 1546 if (BlockNumPredsLeft[i] == 0) { 1547 ReadyBlocks.push_back(Block); 1548 } 1549 } 1550 1551 while (SIScheduleBlock *Block = pickBlock()) { 1552 BlocksScheduled.push_back(Block); 1553 blockScheduled(Block); 1554 } 1555 1556 LLVM_DEBUG(dbgs() << "Block Order:"; for (SIScheduleBlock *Block 1557 : BlocksScheduled) { 1558 dbgs() << ' ' << Block->getID(); 1559 } dbgs() << '\n';); 1560 } 1561 1562 bool SIScheduleBlockScheduler::tryCandidateLatency(SIBlockSchedCandidate &Cand, 1563 SIBlockSchedCandidate &TryCand) { 1564 if (!Cand.isValid()) { 1565 TryCand.Reason = NodeOrder; 1566 return true; 1567 } 1568 1569 // Try to hide high latencies. 1570 if (SISched::tryLess(TryCand.LastPosHighLatParentScheduled, 1571 Cand.LastPosHighLatParentScheduled, TryCand, Cand, Latency)) 1572 return true; 1573 // Schedule high latencies early so you can hide them better. 1574 if (SISched::tryGreater(TryCand.IsHighLatency, Cand.IsHighLatency, 1575 TryCand, Cand, Latency)) 1576 return true; 1577 if (TryCand.IsHighLatency && SISched::tryGreater(TryCand.Height, Cand.Height, 1578 TryCand, Cand, Depth)) 1579 return true; 1580 if (SISched::tryGreater(TryCand.NumHighLatencySuccessors, 1581 Cand.NumHighLatencySuccessors, 1582 TryCand, Cand, Successor)) 1583 return true; 1584 return false; 1585 } 1586 1587 bool SIScheduleBlockScheduler::tryCandidateRegUsage(SIBlockSchedCandidate &Cand, 1588 SIBlockSchedCandidate &TryCand) { 1589 if (!Cand.isValid()) { 1590 TryCand.Reason = NodeOrder; 1591 return true; 1592 } 1593 1594 if (SISched::tryLess(TryCand.VGPRUsageDiff > 0, Cand.VGPRUsageDiff > 0, 1595 TryCand, Cand, RegUsage)) 1596 return true; 1597 if (SISched::tryGreater(TryCand.NumSuccessors > 0, 1598 Cand.NumSuccessors > 0, 1599 TryCand, Cand, Successor)) 1600 return true; 1601 if (SISched::tryGreater(TryCand.Height, Cand.Height, TryCand, Cand, Depth)) 1602 return true; 1603 if (SISched::tryLess(TryCand.VGPRUsageDiff, Cand.VGPRUsageDiff, 1604 TryCand, Cand, RegUsage)) 1605 return true; 1606 return false; 1607 } 1608 1609 SIScheduleBlock *SIScheduleBlockScheduler::pickBlock() { 1610 SIBlockSchedCandidate Cand; 1611 std::vector<SIScheduleBlock*>::iterator Best; 1612 SIScheduleBlock *Block; 1613 if (ReadyBlocks.empty()) 1614 return nullptr; 1615 1616 DAG->fillVgprSgprCost(LiveRegs.begin(), LiveRegs.end(), 1617 VregCurrentUsage, SregCurrentUsage); 1618 if (VregCurrentUsage > maxVregUsage) 1619 maxVregUsage = VregCurrentUsage; 1620 if (SregCurrentUsage > maxSregUsage) 1621 maxSregUsage = SregCurrentUsage; 1622 LLVM_DEBUG(dbgs() << "Picking New Blocks\n"; dbgs() << "Available: "; 1623 for (SIScheduleBlock *Block 1624 : ReadyBlocks) dbgs() 1625 << Block->getID() << ' '; 1626 dbgs() << "\nCurrent Live:\n"; 1627 for (unsigned Reg 1628 : LiveRegs) dbgs() 1629 << printVRegOrUnit(Reg, DAG->getTRI()) << ' '; 1630 dbgs() << '\n'; 1631 dbgs() << "Current VGPRs: " << VregCurrentUsage << '\n'; 1632 dbgs() << "Current SGPRs: " << SregCurrentUsage << '\n';); 1633 1634 Cand.Block = nullptr; 1635 for (std::vector<SIScheduleBlock*>::iterator I = ReadyBlocks.begin(), 1636 E = ReadyBlocks.end(); I != E; ++I) { 1637 SIBlockSchedCandidate TryCand; 1638 TryCand.Block = *I; 1639 TryCand.IsHighLatency = TryCand.Block->isHighLatencyBlock(); 1640 TryCand.VGPRUsageDiff = 1641 checkRegUsageImpact(TryCand.Block->getInRegs(), 1642 TryCand.Block->getOutRegs())[AMDGPU::RegisterPressureSets::VGPR_32]; 1643 TryCand.NumSuccessors = TryCand.Block->getSuccs().size(); 1644 TryCand.NumHighLatencySuccessors = 1645 TryCand.Block->getNumHighLatencySuccessors(); 1646 TryCand.LastPosHighLatParentScheduled = 1647 (unsigned int) std::max<int> (0, 1648 LastPosHighLatencyParentScheduled[TryCand.Block->getID()] - 1649 LastPosWaitedHighLatency); 1650 TryCand.Height = TryCand.Block->Height; 1651 // Try not to increase VGPR usage too much, else we may spill. 1652 if (VregCurrentUsage > 120 || 1653 Variant != SISchedulerBlockSchedulerVariant::BlockLatencyRegUsage) { 1654 if (!tryCandidateRegUsage(Cand, TryCand) && 1655 Variant != SISchedulerBlockSchedulerVariant::BlockRegUsage) 1656 tryCandidateLatency(Cand, TryCand); 1657 } else { 1658 if (!tryCandidateLatency(Cand, TryCand)) 1659 tryCandidateRegUsage(Cand, TryCand); 1660 } 1661 if (TryCand.Reason != NoCand) { 1662 Cand.setBest(TryCand); 1663 Best = I; 1664 LLVM_DEBUG(dbgs() << "Best Current Choice: " << Cand.Block->getID() << ' ' 1665 << getReasonStr(Cand.Reason) << '\n'); 1666 } 1667 } 1668 1669 LLVM_DEBUG(dbgs() << "Picking: " << Cand.Block->getID() << '\n'; 1670 dbgs() << "Is a block with high latency instruction: " 1671 << (Cand.IsHighLatency ? "yes\n" : "no\n"); 1672 dbgs() << "Position of last high latency dependency: " 1673 << Cand.LastPosHighLatParentScheduled << '\n'; 1674 dbgs() << "VGPRUsageDiff: " << Cand.VGPRUsageDiff << '\n'; 1675 dbgs() << '\n';); 1676 1677 Block = Cand.Block; 1678 ReadyBlocks.erase(Best); 1679 return Block; 1680 } 1681 1682 // Tracking of currently alive registers to determine VGPR Usage. 1683 1684 void SIScheduleBlockScheduler::addLiveRegs(std::set<unsigned> &Regs) { 1685 for (unsigned Reg : Regs) { 1686 // For now only track virtual registers. 1687 if (!Register::isVirtualRegister(Reg)) 1688 continue; 1689 // If not already in the live set, then add it. 1690 (void) LiveRegs.insert(Reg); 1691 } 1692 } 1693 1694 void SIScheduleBlockScheduler::decreaseLiveRegs(SIScheduleBlock *Block, 1695 std::set<unsigned> &Regs) { 1696 for (unsigned Reg : Regs) { 1697 // For now only track virtual registers. 1698 std::set<unsigned>::iterator Pos = LiveRegs.find(Reg); 1699 assert (Pos != LiveRegs.end() && // Reg must be live. 1700 LiveRegsConsumers.find(Reg) != LiveRegsConsumers.end() && 1701 LiveRegsConsumers[Reg] >= 1); 1702 --LiveRegsConsumers[Reg]; 1703 if (LiveRegsConsumers[Reg] == 0) 1704 LiveRegs.erase(Pos); 1705 } 1706 } 1707 1708 void SIScheduleBlockScheduler::releaseBlockSuccs(SIScheduleBlock *Parent) { 1709 for (const auto &Block : Parent->getSuccs()) { 1710 if (--BlockNumPredsLeft[Block.first->getID()] == 0) 1711 ReadyBlocks.push_back(Block.first); 1712 1713 if (Parent->isHighLatencyBlock() && 1714 Block.second == SIScheduleBlockLinkKind::Data) 1715 LastPosHighLatencyParentScheduled[Block.first->getID()] = NumBlockScheduled; 1716 } 1717 } 1718 1719 void SIScheduleBlockScheduler::blockScheduled(SIScheduleBlock *Block) { 1720 decreaseLiveRegs(Block, Block->getInRegs()); 1721 addLiveRegs(Block->getOutRegs()); 1722 releaseBlockSuccs(Block); 1723 for (std::map<unsigned, unsigned>::iterator RegI = 1724 LiveOutRegsNumUsages[Block->getID()].begin(), 1725 E = LiveOutRegsNumUsages[Block->getID()].end(); RegI != E; ++RegI) { 1726 std::pair<unsigned, unsigned> RegP = *RegI; 1727 // We produce this register, thus it must not be previously alive. 1728 assert(LiveRegsConsumers.find(RegP.first) == LiveRegsConsumers.end() || 1729 LiveRegsConsumers[RegP.first] == 0); 1730 LiveRegsConsumers[RegP.first] += RegP.second; 1731 } 1732 if (LastPosHighLatencyParentScheduled[Block->getID()] > 1733 (unsigned)LastPosWaitedHighLatency) 1734 LastPosWaitedHighLatency = 1735 LastPosHighLatencyParentScheduled[Block->getID()]; 1736 ++NumBlockScheduled; 1737 } 1738 1739 std::vector<int> 1740 SIScheduleBlockScheduler::checkRegUsageImpact(std::set<unsigned> &InRegs, 1741 std::set<unsigned> &OutRegs) { 1742 std::vector<int> DiffSetPressure; 1743 DiffSetPressure.assign(DAG->getTRI()->getNumRegPressureSets(), 0); 1744 1745 for (unsigned Reg : InRegs) { 1746 // For now only track virtual registers. 1747 if (!Register::isVirtualRegister(Reg)) 1748 continue; 1749 if (LiveRegsConsumers[Reg] > 1) 1750 continue; 1751 PSetIterator PSetI = DAG->getMRI()->getPressureSets(Reg); 1752 for (; PSetI.isValid(); ++PSetI) { 1753 DiffSetPressure[*PSetI] -= PSetI.getWeight(); 1754 } 1755 } 1756 1757 for (unsigned Reg : OutRegs) { 1758 // For now only track virtual registers. 1759 if (!Register::isVirtualRegister(Reg)) 1760 continue; 1761 PSetIterator PSetI = DAG->getMRI()->getPressureSets(Reg); 1762 for (; PSetI.isValid(); ++PSetI) { 1763 DiffSetPressure[*PSetI] += PSetI.getWeight(); 1764 } 1765 } 1766 1767 return DiffSetPressure; 1768 } 1769 1770 // SIScheduler // 1771 1772 struct SIScheduleBlockResult 1773 SIScheduler::scheduleVariant(SISchedulerBlockCreatorVariant BlockVariant, 1774 SISchedulerBlockSchedulerVariant ScheduleVariant) { 1775 SIScheduleBlocks Blocks = BlockCreator.getBlocks(BlockVariant); 1776 SIScheduleBlockScheduler Scheduler(DAG, ScheduleVariant, Blocks); 1777 std::vector<SIScheduleBlock*> ScheduledBlocks; 1778 struct SIScheduleBlockResult Res; 1779 1780 ScheduledBlocks = Scheduler.getBlocks(); 1781 1782 for (unsigned b = 0; b < ScheduledBlocks.size(); ++b) { 1783 SIScheduleBlock *Block = ScheduledBlocks[b]; 1784 std::vector<SUnit*> SUs = Block->getScheduledUnits(); 1785 1786 for (SUnit* SU : SUs) 1787 Res.SUs.push_back(SU->NodeNum); 1788 } 1789 1790 Res.MaxSGPRUsage = Scheduler.getSGPRUsage(); 1791 Res.MaxVGPRUsage = Scheduler.getVGPRUsage(); 1792 return Res; 1793 } 1794 1795 // SIScheduleDAGMI // 1796 1797 SIScheduleDAGMI::SIScheduleDAGMI(MachineSchedContext *C) : 1798 ScheduleDAGMILive(C, std::make_unique<GenericScheduler>(C)) { 1799 SITII = static_cast<const SIInstrInfo*>(TII); 1800 SITRI = static_cast<const SIRegisterInfo*>(TRI); 1801 } 1802 1803 SIScheduleDAGMI::~SIScheduleDAGMI() = default; 1804 1805 // Code adapted from scheduleDAG.cpp 1806 // Does a topological sort over the SUs. 1807 // Both TopDown and BottomUp 1808 void SIScheduleDAGMI::topologicalSort() { 1809 Topo.InitDAGTopologicalSorting(); 1810 1811 TopDownIndex2SU = std::vector<int>(Topo.begin(), Topo.end()); 1812 BottomUpIndex2SU = std::vector<int>(Topo.rbegin(), Topo.rend()); 1813 } 1814 1815 // Move low latencies further from their user without 1816 // increasing SGPR usage (in general) 1817 // This is to be replaced by a better pass that would 1818 // take into account SGPR usage (based on VGPR Usage 1819 // and the corresponding wavefront count), that would 1820 // try to merge groups of loads if it make sense, etc 1821 void SIScheduleDAGMI::moveLowLatencies() { 1822 unsigned DAGSize = SUnits.size(); 1823 int LastLowLatencyUser = -1; 1824 int LastLowLatencyPos = -1; 1825 1826 for (unsigned i = 0, e = ScheduledSUnits.size(); i != e; ++i) { 1827 SUnit *SU = &SUnits[ScheduledSUnits[i]]; 1828 bool IsLowLatencyUser = false; 1829 unsigned MinPos = 0; 1830 1831 for (SDep& PredDep : SU->Preds) { 1832 SUnit *Pred = PredDep.getSUnit(); 1833 if (SITII->isLowLatencyInstruction(*Pred->getInstr())) { 1834 IsLowLatencyUser = true; 1835 } 1836 if (Pred->NodeNum >= DAGSize) 1837 continue; 1838 unsigned PredPos = ScheduledSUnitsInv[Pred->NodeNum]; 1839 if (PredPos >= MinPos) 1840 MinPos = PredPos + 1; 1841 } 1842 1843 if (SITII->isLowLatencyInstruction(*SU->getInstr())) { 1844 unsigned BestPos = LastLowLatencyUser + 1; 1845 if ((int)BestPos <= LastLowLatencyPos) 1846 BestPos = LastLowLatencyPos + 1; 1847 if (BestPos < MinPos) 1848 BestPos = MinPos; 1849 if (BestPos < i) { 1850 for (unsigned u = i; u > BestPos; --u) { 1851 ++ScheduledSUnitsInv[ScheduledSUnits[u-1]]; 1852 ScheduledSUnits[u] = ScheduledSUnits[u-1]; 1853 } 1854 ScheduledSUnits[BestPos] = SU->NodeNum; 1855 ScheduledSUnitsInv[SU->NodeNum] = BestPos; 1856 } 1857 LastLowLatencyPos = BestPos; 1858 if (IsLowLatencyUser) 1859 LastLowLatencyUser = BestPos; 1860 } else if (IsLowLatencyUser) { 1861 LastLowLatencyUser = i; 1862 // Moves COPY instructions on which depends 1863 // the low latency instructions too. 1864 } else if (SU->getInstr()->getOpcode() == AMDGPU::COPY) { 1865 bool CopyForLowLat = false; 1866 for (SDep& SuccDep : SU->Succs) { 1867 SUnit *Succ = SuccDep.getSUnit(); 1868 if (SuccDep.isWeak() || Succ->NodeNum >= DAGSize) 1869 continue; 1870 if (SITII->isLowLatencyInstruction(*Succ->getInstr())) { 1871 CopyForLowLat = true; 1872 } 1873 } 1874 if (!CopyForLowLat) 1875 continue; 1876 if (MinPos < i) { 1877 for (unsigned u = i; u > MinPos; --u) { 1878 ++ScheduledSUnitsInv[ScheduledSUnits[u-1]]; 1879 ScheduledSUnits[u] = ScheduledSUnits[u-1]; 1880 } 1881 ScheduledSUnits[MinPos] = SU->NodeNum; 1882 ScheduledSUnitsInv[SU->NodeNum] = MinPos; 1883 } 1884 } 1885 } 1886 } 1887 1888 void SIScheduleDAGMI::restoreSULinksLeft() { 1889 for (unsigned i = 0, e = SUnits.size(); i != e; ++i) { 1890 SUnits[i].isScheduled = false; 1891 SUnits[i].WeakPredsLeft = SUnitsLinksBackup[i].WeakPredsLeft; 1892 SUnits[i].NumPredsLeft = SUnitsLinksBackup[i].NumPredsLeft; 1893 SUnits[i].WeakSuccsLeft = SUnitsLinksBackup[i].WeakSuccsLeft; 1894 SUnits[i].NumSuccsLeft = SUnitsLinksBackup[i].NumSuccsLeft; 1895 } 1896 } 1897 1898 // Return the Vgpr and Sgpr usage corresponding to some virtual registers. 1899 template<typename _Iterator> void 1900 SIScheduleDAGMI::fillVgprSgprCost(_Iterator First, _Iterator End, 1901 unsigned &VgprUsage, unsigned &SgprUsage) { 1902 VgprUsage = 0; 1903 SgprUsage = 0; 1904 for (_Iterator RegI = First; RegI != End; ++RegI) { 1905 unsigned Reg = *RegI; 1906 // For now only track virtual registers 1907 if (!Register::isVirtualRegister(Reg)) 1908 continue; 1909 PSetIterator PSetI = MRI.getPressureSets(Reg); 1910 for (; PSetI.isValid(); ++PSetI) { 1911 if (*PSetI == AMDGPU::RegisterPressureSets::VGPR_32) 1912 VgprUsage += PSetI.getWeight(); 1913 else if (*PSetI == AMDGPU::RegisterPressureSets::SReg_32) 1914 SgprUsage += PSetI.getWeight(); 1915 } 1916 } 1917 } 1918 1919 void SIScheduleDAGMI::schedule() 1920 { 1921 SmallVector<SUnit*, 8> TopRoots, BotRoots; 1922 SIScheduleBlockResult Best, Temp; 1923 LLVM_DEBUG(dbgs() << "Preparing Scheduling\n"); 1924 1925 buildDAGWithRegPressure(); 1926 LLVM_DEBUG(dump()); 1927 1928 topologicalSort(); 1929 findRootsAndBiasEdges(TopRoots, BotRoots); 1930 // We reuse several ScheduleDAGMI and ScheduleDAGMILive 1931 // functions, but to make them happy we must initialize 1932 // the default Scheduler implementation (even if we do not 1933 // run it) 1934 SchedImpl->initialize(this); 1935 initQueues(TopRoots, BotRoots); 1936 1937 // Fill some stats to help scheduling. 1938 1939 SUnitsLinksBackup = SUnits; 1940 IsLowLatencySU.clear(); 1941 LowLatencyOffset.clear(); 1942 IsHighLatencySU.clear(); 1943 1944 IsLowLatencySU.resize(SUnits.size(), 0); 1945 LowLatencyOffset.resize(SUnits.size(), 0); 1946 IsHighLatencySU.resize(SUnits.size(), 0); 1947 1948 for (unsigned i = 0, e = (unsigned)SUnits.size(); i != e; ++i) { 1949 SUnit *SU = &SUnits[i]; 1950 const MachineOperand *BaseLatOp; 1951 int64_t OffLatReg; 1952 if (SITII->isLowLatencyInstruction(*SU->getInstr())) { 1953 IsLowLatencySU[i] = 1; 1954 bool OffsetIsScalable; 1955 if (SITII->getMemOperandWithOffset(*SU->getInstr(), BaseLatOp, OffLatReg, 1956 OffsetIsScalable, TRI)) 1957 LowLatencyOffset[i] = OffLatReg; 1958 } else if (SITII->isHighLatencyDef(SU->getInstr()->getOpcode())) 1959 IsHighLatencySU[i] = 1; 1960 } 1961 1962 SIScheduler Scheduler(this); 1963 Best = Scheduler.scheduleVariant(SISchedulerBlockCreatorVariant::LatenciesAlone, 1964 SISchedulerBlockSchedulerVariant::BlockLatencyRegUsage); 1965 1966 // if VGPR usage is extremely high, try other good performing variants 1967 // which could lead to lower VGPR usage 1968 if (Best.MaxVGPRUsage > 180) { 1969 static const std::pair<SISchedulerBlockCreatorVariant, 1970 SISchedulerBlockSchedulerVariant> 1971 Variants[] = { 1972 { LatenciesAlone, BlockRegUsageLatency }, 1973 // { LatenciesAlone, BlockRegUsage }, 1974 { LatenciesGrouped, BlockLatencyRegUsage }, 1975 // { LatenciesGrouped, BlockRegUsageLatency }, 1976 // { LatenciesGrouped, BlockRegUsage }, 1977 { LatenciesAlonePlusConsecutive, BlockLatencyRegUsage }, 1978 // { LatenciesAlonePlusConsecutive, BlockRegUsageLatency }, 1979 // { LatenciesAlonePlusConsecutive, BlockRegUsage } 1980 }; 1981 for (std::pair<SISchedulerBlockCreatorVariant, SISchedulerBlockSchedulerVariant> v : Variants) { 1982 Temp = Scheduler.scheduleVariant(v.first, v.second); 1983 if (Temp.MaxVGPRUsage < Best.MaxVGPRUsage) 1984 Best = Temp; 1985 } 1986 } 1987 // if VGPR usage is still extremely high, we may spill. Try other variants 1988 // which are less performing, but that could lead to lower VGPR usage. 1989 if (Best.MaxVGPRUsage > 200) { 1990 static const std::pair<SISchedulerBlockCreatorVariant, 1991 SISchedulerBlockSchedulerVariant> 1992 Variants[] = { 1993 // { LatenciesAlone, BlockRegUsageLatency }, 1994 { LatenciesAlone, BlockRegUsage }, 1995 // { LatenciesGrouped, BlockLatencyRegUsage }, 1996 { LatenciesGrouped, BlockRegUsageLatency }, 1997 { LatenciesGrouped, BlockRegUsage }, 1998 // { LatenciesAlonePlusConsecutive, BlockLatencyRegUsage }, 1999 { LatenciesAlonePlusConsecutive, BlockRegUsageLatency }, 2000 { LatenciesAlonePlusConsecutive, BlockRegUsage } 2001 }; 2002 for (std::pair<SISchedulerBlockCreatorVariant, SISchedulerBlockSchedulerVariant> v : Variants) { 2003 Temp = Scheduler.scheduleVariant(v.first, v.second); 2004 if (Temp.MaxVGPRUsage < Best.MaxVGPRUsage) 2005 Best = Temp; 2006 } 2007 } 2008 2009 ScheduledSUnits = Best.SUs; 2010 ScheduledSUnitsInv.resize(SUnits.size()); 2011 2012 for (unsigned i = 0, e = (unsigned)SUnits.size(); i != e; ++i) { 2013 ScheduledSUnitsInv[ScheduledSUnits[i]] = i; 2014 } 2015 2016 moveLowLatencies(); 2017 2018 // Tell the outside world about the result of the scheduling. 2019 2020 assert(TopRPTracker.getPos() == RegionBegin && "bad initial Top tracker"); 2021 TopRPTracker.setPos(CurrentTop); 2022 2023 for (std::vector<unsigned>::iterator I = ScheduledSUnits.begin(), 2024 E = ScheduledSUnits.end(); I != E; ++I) { 2025 SUnit *SU = &SUnits[*I]; 2026 2027 scheduleMI(SU, true); 2028 2029 LLVM_DEBUG(dbgs() << "Scheduling SU(" << SU->NodeNum << ") " 2030 << *SU->getInstr()); 2031 } 2032 2033 assert(CurrentTop == CurrentBottom && "Nonempty unscheduled zone."); 2034 2035 placeDebugValues(); 2036 2037 LLVM_DEBUG({ 2038 dbgs() << "*** Final schedule for " 2039 << printMBBReference(*begin()->getParent()) << " ***\n"; 2040 dumpSchedule(); 2041 dbgs() << '\n'; 2042 }); 2043 } 2044