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[DAG->getSGPRSetID()]; 273 TryCand.VGPRUsage = pressure[DAG->getVGPRSetID()]; 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 " << LiveInPressure[DAG->getSGPRSetID()] << ' ' 599 << LiveInPressure[DAG->getVGPRSetID()] << '\n'; 600 dbgs() << "LiveOutPressure " << LiveOutPressure[DAG->getSGPRSetID()] << ' ' 601 << LiveOutPressure[DAG->getVGPRSetID()] << "\n\n"; 602 dbgs() << "LiveIns:\n"; 603 for (unsigned Reg : LiveInRegs) 604 dbgs() << printVRegOrUnit(Reg, DAG->getTRI()) << ' '; 605 606 dbgs() << "\nLiveOuts:\n"; 607 for (unsigned Reg : LiveOutRegs) 608 dbgs() << printVRegOrUnit(Reg, DAG->getTRI()) << ' '; 609 } 610 611 dbgs() << "\nInstructions:\n"; 612 for (const SUnit* SU : SUnits) 613 DAG->dumpNode(*SU); 614 615 dbgs() << "///////////////////////\n"; 616 } 617 #endif 618 619 // SIScheduleBlockCreator // 620 621 SIScheduleBlockCreator::SIScheduleBlockCreator(SIScheduleDAGMI *DAG) 622 : DAG(DAG) {} 623 624 SIScheduleBlocks 625 SIScheduleBlockCreator::getBlocks(SISchedulerBlockCreatorVariant BlockVariant) { 626 std::map<SISchedulerBlockCreatorVariant, SIScheduleBlocks>::iterator B = 627 Blocks.find(BlockVariant); 628 if (B == Blocks.end()) { 629 SIScheduleBlocks Res; 630 createBlocksForVariant(BlockVariant); 631 topologicalSort(); 632 scheduleInsideBlocks(); 633 fillStats(); 634 Res.Blocks = CurrentBlocks; 635 Res.TopDownIndex2Block = TopDownIndex2Block; 636 Res.TopDownBlock2Index = TopDownBlock2Index; 637 Blocks[BlockVariant] = Res; 638 return Res; 639 } else { 640 return B->second; 641 } 642 } 643 644 bool SIScheduleBlockCreator::isSUInBlock(SUnit *SU, unsigned ID) { 645 if (SU->NodeNum >= DAG->SUnits.size()) 646 return false; 647 return CurrentBlocks[Node2CurrentBlock[SU->NodeNum]]->getID() == ID; 648 } 649 650 void SIScheduleBlockCreator::colorHighLatenciesAlone() { 651 unsigned DAGSize = DAG->SUnits.size(); 652 653 for (unsigned i = 0, e = DAGSize; i != e; ++i) { 654 SUnit *SU = &DAG->SUnits[i]; 655 if (DAG->IsHighLatencySU[SU->NodeNum]) { 656 CurrentColoring[SU->NodeNum] = NextReservedID++; 657 } 658 } 659 } 660 661 static bool 662 hasDataDependencyPred(const SUnit &SU, const SUnit &FromSU) { 663 for (const auto &PredDep : SU.Preds) { 664 if (PredDep.getSUnit() == &FromSU && 665 PredDep.getKind() == llvm::SDep::Data) 666 return true; 667 } 668 return false; 669 } 670 671 void SIScheduleBlockCreator::colorHighLatenciesGroups() { 672 unsigned DAGSize = DAG->SUnits.size(); 673 unsigned NumHighLatencies = 0; 674 unsigned GroupSize; 675 int Color = NextReservedID; 676 unsigned Count = 0; 677 std::set<unsigned> FormingGroup; 678 679 for (unsigned i = 0, e = DAGSize; i != e; ++i) { 680 SUnit *SU = &DAG->SUnits[i]; 681 if (DAG->IsHighLatencySU[SU->NodeNum]) 682 ++NumHighLatencies; 683 } 684 685 if (NumHighLatencies == 0) 686 return; 687 688 if (NumHighLatencies <= 6) 689 GroupSize = 2; 690 else if (NumHighLatencies <= 12) 691 GroupSize = 3; 692 else 693 GroupSize = 4; 694 695 for (unsigned SUNum : DAG->TopDownIndex2SU) { 696 const SUnit &SU = DAG->SUnits[SUNum]; 697 if (DAG->IsHighLatencySU[SU.NodeNum]) { 698 unsigned CompatibleGroup = true; 699 int ProposedColor = Color; 700 std::vector<int> AdditionalElements; 701 702 // We don't want to put in the same block 703 // two high latency instructions that depend 704 // on each other. 705 // One way would be to check canAddEdge 706 // in both directions, but that currently is not 707 // enough because there the high latency order is 708 // enforced (via links). 709 // Instead, look at the dependencies between the 710 // high latency instructions and deduce if it is 711 // a data dependency or not. 712 for (unsigned j : FormingGroup) { 713 bool HasSubGraph; 714 std::vector<int> SubGraph; 715 // By construction (topological order), if SU and 716 // DAG->SUnits[j] are linked, DAG->SUnits[j] is neccessary 717 // in the parent graph of SU. 718 #ifndef NDEBUG 719 SubGraph = DAG->GetTopo()->GetSubGraph(SU, DAG->SUnits[j], 720 HasSubGraph); 721 assert(!HasSubGraph); 722 #endif 723 SubGraph = DAG->GetTopo()->GetSubGraph(DAG->SUnits[j], SU, 724 HasSubGraph); 725 if (!HasSubGraph) 726 continue; // No dependencies between each other 727 else if (SubGraph.size() > 5) { 728 // Too many elements would be required to be added to the block. 729 CompatibleGroup = false; 730 break; 731 } 732 else { 733 // Check the type of dependency 734 for (unsigned k : SubGraph) { 735 // If in the path to join the two instructions, 736 // there is another high latency instruction, 737 // or instructions colored for another block 738 // abort the merge. 739 if (DAG->IsHighLatencySU[k] || 740 (CurrentColoring[k] != ProposedColor && 741 CurrentColoring[k] != 0)) { 742 CompatibleGroup = false; 743 break; 744 } 745 // If one of the SU in the subgraph depends on the result of SU j, 746 // there'll be a data dependency. 747 if (hasDataDependencyPred(DAG->SUnits[k], DAG->SUnits[j])) { 748 CompatibleGroup = false; 749 break; 750 } 751 } 752 if (!CompatibleGroup) 753 break; 754 // Same check for the SU 755 if (hasDataDependencyPred(SU, DAG->SUnits[j])) { 756 CompatibleGroup = false; 757 break; 758 } 759 // Add all the required instructions to the block 760 // These cannot live in another block (because they 761 // depend (order dependency) on one of the 762 // instruction in the block, and are required for the 763 // high latency instruction we add. 764 AdditionalElements.insert(AdditionalElements.end(), 765 SubGraph.begin(), SubGraph.end()); 766 } 767 } 768 if (CompatibleGroup) { 769 FormingGroup.insert(SU.NodeNum); 770 for (unsigned j : AdditionalElements) 771 CurrentColoring[j] = ProposedColor; 772 CurrentColoring[SU.NodeNum] = ProposedColor; 773 ++Count; 774 } 775 // Found one incompatible instruction, 776 // or has filled a big enough group. 777 // -> start a new one. 778 if (!CompatibleGroup) { 779 FormingGroup.clear(); 780 Color = ++NextReservedID; 781 ProposedColor = Color; 782 FormingGroup.insert(SU.NodeNum); 783 CurrentColoring[SU.NodeNum] = ProposedColor; 784 Count = 0; 785 } else if (Count == GroupSize) { 786 FormingGroup.clear(); 787 Color = ++NextReservedID; 788 ProposedColor = Color; 789 Count = 0; 790 } 791 } 792 } 793 } 794 795 void SIScheduleBlockCreator::colorComputeReservedDependencies() { 796 unsigned DAGSize = DAG->SUnits.size(); 797 std::map<std::set<unsigned>, unsigned> ColorCombinations; 798 799 CurrentTopDownReservedDependencyColoring.clear(); 800 CurrentBottomUpReservedDependencyColoring.clear(); 801 802 CurrentTopDownReservedDependencyColoring.resize(DAGSize, 0); 803 CurrentBottomUpReservedDependencyColoring.resize(DAGSize, 0); 804 805 // Traverse TopDown, and give different colors to SUs depending 806 // on which combination of High Latencies they depend on. 807 808 for (unsigned SUNum : DAG->TopDownIndex2SU) { 809 SUnit *SU = &DAG->SUnits[SUNum]; 810 std::set<unsigned> SUColors; 811 812 // Already given. 813 if (CurrentColoring[SU->NodeNum]) { 814 CurrentTopDownReservedDependencyColoring[SU->NodeNum] = 815 CurrentColoring[SU->NodeNum]; 816 continue; 817 } 818 819 for (SDep& PredDep : SU->Preds) { 820 SUnit *Pred = PredDep.getSUnit(); 821 if (PredDep.isWeak() || Pred->NodeNum >= DAGSize) 822 continue; 823 if (CurrentTopDownReservedDependencyColoring[Pred->NodeNum] > 0) 824 SUColors.insert(CurrentTopDownReservedDependencyColoring[Pred->NodeNum]); 825 } 826 // Color 0 by default. 827 if (SUColors.empty()) 828 continue; 829 // Same color than parents. 830 if (SUColors.size() == 1 && *SUColors.begin() > DAGSize) 831 CurrentTopDownReservedDependencyColoring[SU->NodeNum] = 832 *SUColors.begin(); 833 else { 834 std::map<std::set<unsigned>, unsigned>::iterator Pos = 835 ColorCombinations.find(SUColors); 836 if (Pos != ColorCombinations.end()) { 837 CurrentTopDownReservedDependencyColoring[SU->NodeNum] = Pos->second; 838 } else { 839 CurrentTopDownReservedDependencyColoring[SU->NodeNum] = 840 NextNonReservedID; 841 ColorCombinations[SUColors] = NextNonReservedID++; 842 } 843 } 844 } 845 846 ColorCombinations.clear(); 847 848 // Same as before, but BottomUp. 849 850 for (unsigned SUNum : DAG->BottomUpIndex2SU) { 851 SUnit *SU = &DAG->SUnits[SUNum]; 852 std::set<unsigned> SUColors; 853 854 // Already given. 855 if (CurrentColoring[SU->NodeNum]) { 856 CurrentBottomUpReservedDependencyColoring[SU->NodeNum] = 857 CurrentColoring[SU->NodeNum]; 858 continue; 859 } 860 861 for (SDep& SuccDep : SU->Succs) { 862 SUnit *Succ = SuccDep.getSUnit(); 863 if (SuccDep.isWeak() || Succ->NodeNum >= DAGSize) 864 continue; 865 if (CurrentBottomUpReservedDependencyColoring[Succ->NodeNum] > 0) 866 SUColors.insert(CurrentBottomUpReservedDependencyColoring[Succ->NodeNum]); 867 } 868 // Keep color 0. 869 if (SUColors.empty()) 870 continue; 871 // Same color than parents. 872 if (SUColors.size() == 1 && *SUColors.begin() > DAGSize) 873 CurrentBottomUpReservedDependencyColoring[SU->NodeNum] = 874 *SUColors.begin(); 875 else { 876 std::map<std::set<unsigned>, unsigned>::iterator Pos = 877 ColorCombinations.find(SUColors); 878 if (Pos != ColorCombinations.end()) { 879 CurrentBottomUpReservedDependencyColoring[SU->NodeNum] = Pos->second; 880 } else { 881 CurrentBottomUpReservedDependencyColoring[SU->NodeNum] = 882 NextNonReservedID; 883 ColorCombinations[SUColors] = NextNonReservedID++; 884 } 885 } 886 } 887 } 888 889 void SIScheduleBlockCreator::colorAccordingToReservedDependencies() { 890 unsigned DAGSize = DAG->SUnits.size(); 891 std::map<std::pair<unsigned, unsigned>, unsigned> ColorCombinations; 892 893 // Every combination of colors given by the top down 894 // and bottom up Reserved node dependency 895 896 for (unsigned i = 0, e = DAGSize; i != e; ++i) { 897 SUnit *SU = &DAG->SUnits[i]; 898 std::pair<unsigned, unsigned> SUColors; 899 900 // High latency instructions: already given. 901 if (CurrentColoring[SU->NodeNum]) 902 continue; 903 904 SUColors.first = CurrentTopDownReservedDependencyColoring[SU->NodeNum]; 905 SUColors.second = CurrentBottomUpReservedDependencyColoring[SU->NodeNum]; 906 907 std::map<std::pair<unsigned, unsigned>, unsigned>::iterator Pos = 908 ColorCombinations.find(SUColors); 909 if (Pos != ColorCombinations.end()) { 910 CurrentColoring[SU->NodeNum] = Pos->second; 911 } else { 912 CurrentColoring[SU->NodeNum] = NextNonReservedID; 913 ColorCombinations[SUColors] = NextNonReservedID++; 914 } 915 } 916 } 917 918 void SIScheduleBlockCreator::colorEndsAccordingToDependencies() { 919 unsigned DAGSize = DAG->SUnits.size(); 920 std::vector<int> PendingColoring = CurrentColoring; 921 922 assert(DAGSize >= 1 && 923 CurrentBottomUpReservedDependencyColoring.size() == DAGSize && 924 CurrentTopDownReservedDependencyColoring.size() == DAGSize); 925 // If there is no reserved block at all, do nothing. We don't want 926 // everything in one block. 927 if (*std::max_element(CurrentBottomUpReservedDependencyColoring.begin(), 928 CurrentBottomUpReservedDependencyColoring.end()) == 0 && 929 *std::max_element(CurrentTopDownReservedDependencyColoring.begin(), 930 CurrentTopDownReservedDependencyColoring.end()) == 0) 931 return; 932 933 for (unsigned SUNum : DAG->BottomUpIndex2SU) { 934 SUnit *SU = &DAG->SUnits[SUNum]; 935 std::set<unsigned> SUColors; 936 std::set<unsigned> SUColorsPending; 937 938 if (CurrentColoring[SU->NodeNum] <= (int)DAGSize) 939 continue; 940 941 if (CurrentBottomUpReservedDependencyColoring[SU->NodeNum] > 0 || 942 CurrentTopDownReservedDependencyColoring[SU->NodeNum] > 0) 943 continue; 944 945 for (SDep& SuccDep : SU->Succs) { 946 SUnit *Succ = SuccDep.getSUnit(); 947 if (SuccDep.isWeak() || Succ->NodeNum >= DAGSize) 948 continue; 949 if (CurrentBottomUpReservedDependencyColoring[Succ->NodeNum] > 0 || 950 CurrentTopDownReservedDependencyColoring[Succ->NodeNum] > 0) 951 SUColors.insert(CurrentColoring[Succ->NodeNum]); 952 SUColorsPending.insert(PendingColoring[Succ->NodeNum]); 953 } 954 // If there is only one child/parent block, and that block 955 // is not among the ones we are removing in this path, then 956 // merge the instruction to that block 957 if (SUColors.size() == 1 && SUColorsPending.size() == 1) 958 PendingColoring[SU->NodeNum] = *SUColors.begin(); 959 else // TODO: Attribute new colors depending on color 960 // combination of children. 961 PendingColoring[SU->NodeNum] = NextNonReservedID++; 962 } 963 CurrentColoring = PendingColoring; 964 } 965 966 967 void SIScheduleBlockCreator::colorForceConsecutiveOrderInGroup() { 968 unsigned DAGSize = DAG->SUnits.size(); 969 unsigned PreviousColor; 970 std::set<unsigned> SeenColors; 971 972 if (DAGSize <= 1) 973 return; 974 975 PreviousColor = CurrentColoring[0]; 976 977 for (unsigned i = 1, e = DAGSize; i != e; ++i) { 978 SUnit *SU = &DAG->SUnits[i]; 979 unsigned CurrentColor = CurrentColoring[i]; 980 unsigned PreviousColorSave = PreviousColor; 981 assert(i == SU->NodeNum); 982 983 if (CurrentColor != PreviousColor) 984 SeenColors.insert(PreviousColor); 985 PreviousColor = CurrentColor; 986 987 if (CurrentColoring[SU->NodeNum] <= (int)DAGSize) 988 continue; 989 990 if (SeenColors.find(CurrentColor) == SeenColors.end()) 991 continue; 992 993 if (PreviousColorSave != CurrentColor) 994 CurrentColoring[i] = NextNonReservedID++; 995 else 996 CurrentColoring[i] = CurrentColoring[i-1]; 997 } 998 } 999 1000 void SIScheduleBlockCreator::colorMergeConstantLoadsNextGroup() { 1001 unsigned DAGSize = DAG->SUnits.size(); 1002 1003 for (unsigned SUNum : DAG->BottomUpIndex2SU) { 1004 SUnit *SU = &DAG->SUnits[SUNum]; 1005 std::set<unsigned> SUColors; 1006 1007 if (CurrentColoring[SU->NodeNum] <= (int)DAGSize) 1008 continue; 1009 1010 // No predecessor: Vgpr constant loading. 1011 // Low latency instructions usually have a predecessor (the address) 1012 if (SU->Preds.size() > 0 && !DAG->IsLowLatencySU[SU->NodeNum]) 1013 continue; 1014 1015 for (SDep& SuccDep : SU->Succs) { 1016 SUnit *Succ = SuccDep.getSUnit(); 1017 if (SuccDep.isWeak() || Succ->NodeNum >= DAGSize) 1018 continue; 1019 SUColors.insert(CurrentColoring[Succ->NodeNum]); 1020 } 1021 if (SUColors.size() == 1) 1022 CurrentColoring[SU->NodeNum] = *SUColors.begin(); 1023 } 1024 } 1025 1026 void SIScheduleBlockCreator::colorMergeIfPossibleNextGroup() { 1027 unsigned DAGSize = DAG->SUnits.size(); 1028 1029 for (unsigned SUNum : DAG->BottomUpIndex2SU) { 1030 SUnit *SU = &DAG->SUnits[SUNum]; 1031 std::set<unsigned> SUColors; 1032 1033 if (CurrentColoring[SU->NodeNum] <= (int)DAGSize) 1034 continue; 1035 1036 for (SDep& SuccDep : SU->Succs) { 1037 SUnit *Succ = SuccDep.getSUnit(); 1038 if (SuccDep.isWeak() || Succ->NodeNum >= DAGSize) 1039 continue; 1040 SUColors.insert(CurrentColoring[Succ->NodeNum]); 1041 } 1042 if (SUColors.size() == 1) 1043 CurrentColoring[SU->NodeNum] = *SUColors.begin(); 1044 } 1045 } 1046 1047 void SIScheduleBlockCreator::colorMergeIfPossibleNextGroupOnlyForReserved() { 1048 unsigned DAGSize = DAG->SUnits.size(); 1049 1050 for (unsigned SUNum : DAG->BottomUpIndex2SU) { 1051 SUnit *SU = &DAG->SUnits[SUNum]; 1052 std::set<unsigned> SUColors; 1053 1054 if (CurrentColoring[SU->NodeNum] <= (int)DAGSize) 1055 continue; 1056 1057 for (SDep& SuccDep : SU->Succs) { 1058 SUnit *Succ = SuccDep.getSUnit(); 1059 if (SuccDep.isWeak() || Succ->NodeNum >= DAGSize) 1060 continue; 1061 SUColors.insert(CurrentColoring[Succ->NodeNum]); 1062 } 1063 if (SUColors.size() == 1 && *SUColors.begin() <= DAGSize) 1064 CurrentColoring[SU->NodeNum] = *SUColors.begin(); 1065 } 1066 } 1067 1068 void SIScheduleBlockCreator::colorMergeIfPossibleSmallGroupsToNextGroup() { 1069 unsigned DAGSize = DAG->SUnits.size(); 1070 std::map<unsigned, unsigned> ColorCount; 1071 1072 for (unsigned SUNum : DAG->BottomUpIndex2SU) { 1073 SUnit *SU = &DAG->SUnits[SUNum]; 1074 unsigned color = CurrentColoring[SU->NodeNum]; 1075 ++ColorCount[color]; 1076 } 1077 1078 for (unsigned SUNum : DAG->BottomUpIndex2SU) { 1079 SUnit *SU = &DAG->SUnits[SUNum]; 1080 unsigned color = CurrentColoring[SU->NodeNum]; 1081 std::set<unsigned> SUColors; 1082 1083 if (CurrentColoring[SU->NodeNum] <= (int)DAGSize) 1084 continue; 1085 1086 if (ColorCount[color] > 1) 1087 continue; 1088 1089 for (SDep& SuccDep : SU->Succs) { 1090 SUnit *Succ = SuccDep.getSUnit(); 1091 if (SuccDep.isWeak() || Succ->NodeNum >= DAGSize) 1092 continue; 1093 SUColors.insert(CurrentColoring[Succ->NodeNum]); 1094 } 1095 if (SUColors.size() == 1 && *SUColors.begin() != color) { 1096 --ColorCount[color]; 1097 CurrentColoring[SU->NodeNum] = *SUColors.begin(); 1098 ++ColorCount[*SUColors.begin()]; 1099 } 1100 } 1101 } 1102 1103 void SIScheduleBlockCreator::cutHugeBlocks() { 1104 // TODO 1105 } 1106 1107 void SIScheduleBlockCreator::regroupNoUserInstructions() { 1108 unsigned DAGSize = DAG->SUnits.size(); 1109 int GroupID = NextNonReservedID++; 1110 1111 for (unsigned SUNum : DAG->BottomUpIndex2SU) { 1112 SUnit *SU = &DAG->SUnits[SUNum]; 1113 bool hasSuccessor = false; 1114 1115 if (CurrentColoring[SU->NodeNum] <= (int)DAGSize) 1116 continue; 1117 1118 for (SDep& SuccDep : SU->Succs) { 1119 SUnit *Succ = SuccDep.getSUnit(); 1120 if (SuccDep.isWeak() || Succ->NodeNum >= DAGSize) 1121 continue; 1122 hasSuccessor = true; 1123 } 1124 if (!hasSuccessor) 1125 CurrentColoring[SU->NodeNum] = GroupID; 1126 } 1127 } 1128 1129 void SIScheduleBlockCreator::colorExports() { 1130 unsigned ExportColor = NextNonReservedID++; 1131 SmallVector<unsigned, 8> ExpGroup; 1132 1133 // Put all exports together in a block. 1134 // The block will naturally end up being scheduled last, 1135 // thus putting exports at the end of the schedule, which 1136 // is better for performance. 1137 // However we must ensure, for safety, the exports can be put 1138 // together in the same block without any other instruction. 1139 // This could happen, for example, when scheduling after regalloc 1140 // if reloading a spilled register from memory using the same 1141 // register than used in a previous export. 1142 // If that happens, do not regroup the exports. 1143 for (unsigned SUNum : DAG->TopDownIndex2SU) { 1144 const SUnit &SU = DAG->SUnits[SUNum]; 1145 if (SIInstrInfo::isEXP(*SU.getInstr())) { 1146 // Check the EXP can be added to the group safely, 1147 // ie without needing any other instruction. 1148 // The EXP is allowed to depend on other EXP 1149 // (they will be in the same group). 1150 for (unsigned j : ExpGroup) { 1151 bool HasSubGraph; 1152 std::vector<int> SubGraph; 1153 // By construction (topological order), if SU and 1154 // DAG->SUnits[j] are linked, DAG->SUnits[j] is neccessary 1155 // in the parent graph of SU. 1156 #ifndef NDEBUG 1157 SubGraph = DAG->GetTopo()->GetSubGraph(SU, DAG->SUnits[j], 1158 HasSubGraph); 1159 assert(!HasSubGraph); 1160 #endif 1161 SubGraph = DAG->GetTopo()->GetSubGraph(DAG->SUnits[j], SU, 1162 HasSubGraph); 1163 if (!HasSubGraph) 1164 continue; // No dependencies between each other 1165 1166 // SubGraph contains all the instructions required 1167 // between EXP SUnits[j] and EXP SU. 1168 for (unsigned k : SubGraph) { 1169 if (!SIInstrInfo::isEXP(*DAG->SUnits[k].getInstr())) 1170 // Other instructions than EXP would be required in the group. 1171 // Abort the groupping. 1172 return; 1173 } 1174 } 1175 1176 ExpGroup.push_back(SUNum); 1177 } 1178 } 1179 1180 // The group can be formed. Give the color. 1181 for (unsigned j : ExpGroup) 1182 CurrentColoring[j] = ExportColor; 1183 } 1184 1185 void SIScheduleBlockCreator::createBlocksForVariant(SISchedulerBlockCreatorVariant BlockVariant) { 1186 unsigned DAGSize = DAG->SUnits.size(); 1187 std::map<unsigned,unsigned> RealID; 1188 1189 CurrentBlocks.clear(); 1190 CurrentColoring.clear(); 1191 CurrentColoring.resize(DAGSize, 0); 1192 Node2CurrentBlock.clear(); 1193 1194 // Restore links previous scheduling variant has overridden. 1195 DAG->restoreSULinksLeft(); 1196 1197 NextReservedID = 1; 1198 NextNonReservedID = DAGSize + 1; 1199 1200 LLVM_DEBUG(dbgs() << "Coloring the graph\n"); 1201 1202 if (BlockVariant == SISchedulerBlockCreatorVariant::LatenciesGrouped) 1203 colorHighLatenciesGroups(); 1204 else 1205 colorHighLatenciesAlone(); 1206 colorComputeReservedDependencies(); 1207 colorAccordingToReservedDependencies(); 1208 colorEndsAccordingToDependencies(); 1209 if (BlockVariant == SISchedulerBlockCreatorVariant::LatenciesAlonePlusConsecutive) 1210 colorForceConsecutiveOrderInGroup(); 1211 regroupNoUserInstructions(); 1212 colorMergeConstantLoadsNextGroup(); 1213 colorMergeIfPossibleNextGroupOnlyForReserved(); 1214 colorExports(); 1215 1216 // Put SUs of same color into same block 1217 Node2CurrentBlock.resize(DAGSize, -1); 1218 for (unsigned i = 0, e = DAGSize; i != e; ++i) { 1219 SUnit *SU = &DAG->SUnits[i]; 1220 unsigned Color = CurrentColoring[SU->NodeNum]; 1221 if (RealID.find(Color) == RealID.end()) { 1222 int ID = CurrentBlocks.size(); 1223 BlockPtrs.push_back(std::make_unique<SIScheduleBlock>(DAG, this, ID)); 1224 CurrentBlocks.push_back(BlockPtrs.rbegin()->get()); 1225 RealID[Color] = ID; 1226 } 1227 CurrentBlocks[RealID[Color]]->addUnit(SU); 1228 Node2CurrentBlock[SU->NodeNum] = RealID[Color]; 1229 } 1230 1231 // Build dependencies between blocks. 1232 for (unsigned i = 0, e = DAGSize; i != e; ++i) { 1233 SUnit *SU = &DAG->SUnits[i]; 1234 int SUID = Node2CurrentBlock[i]; 1235 for (SDep& SuccDep : SU->Succs) { 1236 SUnit *Succ = SuccDep.getSUnit(); 1237 if (SuccDep.isWeak() || Succ->NodeNum >= DAGSize) 1238 continue; 1239 if (Node2CurrentBlock[Succ->NodeNum] != SUID) 1240 CurrentBlocks[SUID]->addSucc(CurrentBlocks[Node2CurrentBlock[Succ->NodeNum]], 1241 SuccDep.isCtrl() ? NoData : Data); 1242 } 1243 for (SDep& PredDep : SU->Preds) { 1244 SUnit *Pred = PredDep.getSUnit(); 1245 if (PredDep.isWeak() || Pred->NodeNum >= DAGSize) 1246 continue; 1247 if (Node2CurrentBlock[Pred->NodeNum] != SUID) 1248 CurrentBlocks[SUID]->addPred(CurrentBlocks[Node2CurrentBlock[Pred->NodeNum]]); 1249 } 1250 } 1251 1252 // Free root and leafs of all blocks to enable scheduling inside them. 1253 for (unsigned i = 0, e = CurrentBlocks.size(); i != e; ++i) { 1254 SIScheduleBlock *Block = CurrentBlocks[i]; 1255 Block->finalizeUnits(); 1256 } 1257 LLVM_DEBUG(dbgs() << "Blocks created:\n\n"; 1258 for (unsigned i = 0, e = CurrentBlocks.size(); i != e; ++i) { 1259 SIScheduleBlock *Block = CurrentBlocks[i]; 1260 Block->printDebug(true); 1261 }); 1262 } 1263 1264 // Two functions taken from Codegen/MachineScheduler.cpp 1265 1266 /// Non-const version. 1267 static MachineBasicBlock::iterator 1268 nextIfDebug(MachineBasicBlock::iterator I, 1269 MachineBasicBlock::const_iterator End) { 1270 for (; I != End; ++I) { 1271 if (!I->isDebugInstr()) 1272 break; 1273 } 1274 return I; 1275 } 1276 1277 void SIScheduleBlockCreator::topologicalSort() { 1278 unsigned DAGSize = CurrentBlocks.size(); 1279 std::vector<int> WorkList; 1280 1281 LLVM_DEBUG(dbgs() << "Topological Sort\n"); 1282 1283 WorkList.reserve(DAGSize); 1284 TopDownIndex2Block.resize(DAGSize); 1285 TopDownBlock2Index.resize(DAGSize); 1286 BottomUpIndex2Block.resize(DAGSize); 1287 1288 for (unsigned i = 0, e = DAGSize; i != e; ++i) { 1289 SIScheduleBlock *Block = CurrentBlocks[i]; 1290 unsigned Degree = Block->getSuccs().size(); 1291 TopDownBlock2Index[i] = Degree; 1292 if (Degree == 0) { 1293 WorkList.push_back(i); 1294 } 1295 } 1296 1297 int Id = DAGSize; 1298 while (!WorkList.empty()) { 1299 int i = WorkList.back(); 1300 SIScheduleBlock *Block = CurrentBlocks[i]; 1301 WorkList.pop_back(); 1302 TopDownBlock2Index[i] = --Id; 1303 TopDownIndex2Block[Id] = i; 1304 for (SIScheduleBlock* Pred : Block->getPreds()) { 1305 if (!--TopDownBlock2Index[Pred->getID()]) 1306 WorkList.push_back(Pred->getID()); 1307 } 1308 } 1309 1310 #ifndef NDEBUG 1311 // Check correctness of the ordering. 1312 for (unsigned i = 0, e = DAGSize; i != e; ++i) { 1313 SIScheduleBlock *Block = CurrentBlocks[i]; 1314 for (SIScheduleBlock* Pred : Block->getPreds()) { 1315 assert(TopDownBlock2Index[i] > TopDownBlock2Index[Pred->getID()] && 1316 "Wrong Top Down topological sorting"); 1317 } 1318 } 1319 #endif 1320 1321 BottomUpIndex2Block = std::vector<int>(TopDownIndex2Block.rbegin(), 1322 TopDownIndex2Block.rend()); 1323 } 1324 1325 void SIScheduleBlockCreator::scheduleInsideBlocks() { 1326 unsigned DAGSize = CurrentBlocks.size(); 1327 1328 LLVM_DEBUG(dbgs() << "\nScheduling Blocks\n\n"); 1329 1330 // We do schedule a valid scheduling such that a Block corresponds 1331 // to a range of instructions. 1332 LLVM_DEBUG(dbgs() << "First phase: Fast scheduling for Reg Liveness\n"); 1333 for (unsigned i = 0, e = DAGSize; i != e; ++i) { 1334 SIScheduleBlock *Block = CurrentBlocks[i]; 1335 Block->fastSchedule(); 1336 } 1337 1338 // Note: the following code, and the part restoring previous position 1339 // is by far the most expensive operation of the Scheduler. 1340 1341 // Do not update CurrentTop. 1342 MachineBasicBlock::iterator CurrentTopFastSched = DAG->getCurrentTop(); 1343 std::vector<MachineBasicBlock::iterator> PosOld; 1344 std::vector<MachineBasicBlock::iterator> PosNew; 1345 PosOld.reserve(DAG->SUnits.size()); 1346 PosNew.reserve(DAG->SUnits.size()); 1347 1348 for (unsigned i = 0, e = DAGSize; i != e; ++i) { 1349 int BlockIndice = TopDownIndex2Block[i]; 1350 SIScheduleBlock *Block = CurrentBlocks[BlockIndice]; 1351 std::vector<SUnit*> SUs = Block->getScheduledUnits(); 1352 1353 for (SUnit* SU : SUs) { 1354 MachineInstr *MI = SU->getInstr(); 1355 MachineBasicBlock::iterator Pos = MI; 1356 PosOld.push_back(Pos); 1357 if (&*CurrentTopFastSched == MI) { 1358 PosNew.push_back(Pos); 1359 CurrentTopFastSched = nextIfDebug(++CurrentTopFastSched, 1360 DAG->getCurrentBottom()); 1361 } else { 1362 // Update the instruction stream. 1363 DAG->getBB()->splice(CurrentTopFastSched, DAG->getBB(), MI); 1364 1365 // Update LiveIntervals. 1366 // Note: Moving all instructions and calling handleMove every time 1367 // is the most cpu intensive operation of the scheduler. 1368 // It would gain a lot if there was a way to recompute the 1369 // LiveIntervals for the entire scheduling region. 1370 DAG->getLIS()->handleMove(*MI, /*UpdateFlags=*/true); 1371 PosNew.push_back(CurrentTopFastSched); 1372 } 1373 } 1374 } 1375 1376 // Now we have Block of SUs == Block of MI. 1377 // We do the final schedule for the instructions inside the block. 1378 // The property that all the SUs of the Block are grouped together as MI 1379 // is used for correct reg usage tracking. 1380 for (unsigned i = 0, e = DAGSize; i != e; ++i) { 1381 SIScheduleBlock *Block = CurrentBlocks[i]; 1382 std::vector<SUnit*> SUs = Block->getScheduledUnits(); 1383 Block->schedule((*SUs.begin())->getInstr(), (*SUs.rbegin())->getInstr()); 1384 } 1385 1386 LLVM_DEBUG(dbgs() << "Restoring MI Pos\n"); 1387 // Restore old ordering (which prevents a LIS->handleMove bug). 1388 for (unsigned i = PosOld.size(), e = 0; i != e; --i) { 1389 MachineBasicBlock::iterator POld = PosOld[i-1]; 1390 MachineBasicBlock::iterator PNew = PosNew[i-1]; 1391 if (PNew != POld) { 1392 // Update the instruction stream. 1393 DAG->getBB()->splice(POld, DAG->getBB(), PNew); 1394 1395 // Update LiveIntervals. 1396 DAG->getLIS()->handleMove(*POld, /*UpdateFlags=*/true); 1397 } 1398 } 1399 1400 LLVM_DEBUG(for (unsigned i = 0, e = CurrentBlocks.size(); i != e; ++i) { 1401 SIScheduleBlock *Block = CurrentBlocks[i]; 1402 Block->printDebug(true); 1403 }); 1404 } 1405 1406 void SIScheduleBlockCreator::fillStats() { 1407 unsigned DAGSize = CurrentBlocks.size(); 1408 1409 for (unsigned i = 0, e = DAGSize; i != e; ++i) { 1410 int BlockIndice = TopDownIndex2Block[i]; 1411 SIScheduleBlock *Block = CurrentBlocks[BlockIndice]; 1412 if (Block->getPreds().empty()) 1413 Block->Depth = 0; 1414 else { 1415 unsigned Depth = 0; 1416 for (SIScheduleBlock *Pred : Block->getPreds()) { 1417 if (Depth < Pred->Depth + Pred->getCost()) 1418 Depth = Pred->Depth + Pred->getCost(); 1419 } 1420 Block->Depth = Depth; 1421 } 1422 } 1423 1424 for (unsigned i = 0, e = DAGSize; i != e; ++i) { 1425 int BlockIndice = BottomUpIndex2Block[i]; 1426 SIScheduleBlock *Block = CurrentBlocks[BlockIndice]; 1427 if (Block->getSuccs().empty()) 1428 Block->Height = 0; 1429 else { 1430 unsigned Height = 0; 1431 for (const auto &Succ : Block->getSuccs()) 1432 Height = std::max(Height, Succ.first->Height + Succ.first->getCost()); 1433 Block->Height = Height; 1434 } 1435 } 1436 } 1437 1438 // SIScheduleBlockScheduler // 1439 1440 SIScheduleBlockScheduler::SIScheduleBlockScheduler(SIScheduleDAGMI *DAG, 1441 SISchedulerBlockSchedulerVariant Variant, 1442 SIScheduleBlocks BlocksStruct) : 1443 DAG(DAG), Variant(Variant), Blocks(BlocksStruct.Blocks), 1444 LastPosWaitedHighLatency(0), NumBlockScheduled(0), VregCurrentUsage(0), 1445 SregCurrentUsage(0), maxVregUsage(0), maxSregUsage(0) { 1446 1447 // Fill the usage of every output 1448 // Warning: while by construction we always have a link between two blocks 1449 // when one needs a result from the other, the number of users of an output 1450 // is not the sum of child blocks having as input the same virtual register. 1451 // Here is an example. A produces x and y. B eats x and produces x'. 1452 // C eats x' and y. The register coalescer may have attributed the same 1453 // virtual register to x and x'. 1454 // To count accurately, we do a topological sort. In case the register is 1455 // found for several parents, we increment the usage of the one with the 1456 // highest topological index. 1457 LiveOutRegsNumUsages.resize(Blocks.size()); 1458 for (unsigned i = 0, e = Blocks.size(); i != e; ++i) { 1459 SIScheduleBlock *Block = Blocks[i]; 1460 for (unsigned Reg : Block->getInRegs()) { 1461 bool Found = false; 1462 int topoInd = -1; 1463 for (SIScheduleBlock* Pred: Block->getPreds()) { 1464 std::set<unsigned> PredOutRegs = Pred->getOutRegs(); 1465 std::set<unsigned>::iterator RegPos = PredOutRegs.find(Reg); 1466 1467 if (RegPos != PredOutRegs.end()) { 1468 Found = true; 1469 if (topoInd < BlocksStruct.TopDownBlock2Index[Pred->getID()]) { 1470 topoInd = BlocksStruct.TopDownBlock2Index[Pred->getID()]; 1471 } 1472 } 1473 } 1474 1475 if (!Found) 1476 continue; 1477 1478 int PredID = BlocksStruct.TopDownIndex2Block[topoInd]; 1479 ++LiveOutRegsNumUsages[PredID][Reg]; 1480 } 1481 } 1482 1483 LastPosHighLatencyParentScheduled.resize(Blocks.size(), 0); 1484 BlockNumPredsLeft.resize(Blocks.size()); 1485 BlockNumSuccsLeft.resize(Blocks.size()); 1486 1487 for (unsigned i = 0, e = Blocks.size(); i != e; ++i) { 1488 SIScheduleBlock *Block = Blocks[i]; 1489 BlockNumPredsLeft[i] = Block->getPreds().size(); 1490 BlockNumSuccsLeft[i] = Block->getSuccs().size(); 1491 } 1492 1493 #ifndef NDEBUG 1494 for (unsigned i = 0, e = Blocks.size(); i != e; ++i) { 1495 SIScheduleBlock *Block = Blocks[i]; 1496 assert(Block->getID() == i); 1497 } 1498 #endif 1499 1500 std::set<unsigned> InRegs = DAG->getInRegs(); 1501 addLiveRegs(InRegs); 1502 1503 // Increase LiveOutRegsNumUsages for blocks 1504 // producing registers consumed in another 1505 // scheduling region. 1506 for (unsigned Reg : DAG->getOutRegs()) { 1507 for (unsigned i = 0, e = Blocks.size(); i != e; ++i) { 1508 // Do reverse traversal 1509 int ID = BlocksStruct.TopDownIndex2Block[Blocks.size()-1-i]; 1510 SIScheduleBlock *Block = Blocks[ID]; 1511 const std::set<unsigned> &OutRegs = Block->getOutRegs(); 1512 1513 if (OutRegs.find(Reg) == OutRegs.end()) 1514 continue; 1515 1516 ++LiveOutRegsNumUsages[ID][Reg]; 1517 break; 1518 } 1519 } 1520 1521 // Fill LiveRegsConsumers for regs that were already 1522 // defined before scheduling. 1523 for (unsigned i = 0, e = Blocks.size(); i != e; ++i) { 1524 SIScheduleBlock *Block = Blocks[i]; 1525 for (unsigned Reg : Block->getInRegs()) { 1526 bool Found = false; 1527 for (SIScheduleBlock* Pred: Block->getPreds()) { 1528 std::set<unsigned> PredOutRegs = Pred->getOutRegs(); 1529 std::set<unsigned>::iterator RegPos = PredOutRegs.find(Reg); 1530 1531 if (RegPos != PredOutRegs.end()) { 1532 Found = true; 1533 break; 1534 } 1535 } 1536 1537 if (!Found) 1538 ++LiveRegsConsumers[Reg]; 1539 } 1540 } 1541 1542 for (unsigned i = 0, e = Blocks.size(); i != e; ++i) { 1543 SIScheduleBlock *Block = Blocks[i]; 1544 if (BlockNumPredsLeft[i] == 0) { 1545 ReadyBlocks.push_back(Block); 1546 } 1547 } 1548 1549 while (SIScheduleBlock *Block = pickBlock()) { 1550 BlocksScheduled.push_back(Block); 1551 blockScheduled(Block); 1552 } 1553 1554 LLVM_DEBUG(dbgs() << "Block Order:"; for (SIScheduleBlock *Block 1555 : BlocksScheduled) { 1556 dbgs() << ' ' << Block->getID(); 1557 } dbgs() << '\n';); 1558 } 1559 1560 bool SIScheduleBlockScheduler::tryCandidateLatency(SIBlockSchedCandidate &Cand, 1561 SIBlockSchedCandidate &TryCand) { 1562 if (!Cand.isValid()) { 1563 TryCand.Reason = NodeOrder; 1564 return true; 1565 } 1566 1567 // Try to hide high latencies. 1568 if (SISched::tryLess(TryCand.LastPosHighLatParentScheduled, 1569 Cand.LastPosHighLatParentScheduled, TryCand, Cand, Latency)) 1570 return true; 1571 // Schedule high latencies early so you can hide them better. 1572 if (SISched::tryGreater(TryCand.IsHighLatency, Cand.IsHighLatency, 1573 TryCand, Cand, Latency)) 1574 return true; 1575 if (TryCand.IsHighLatency && SISched::tryGreater(TryCand.Height, Cand.Height, 1576 TryCand, Cand, Depth)) 1577 return true; 1578 if (SISched::tryGreater(TryCand.NumHighLatencySuccessors, 1579 Cand.NumHighLatencySuccessors, 1580 TryCand, Cand, Successor)) 1581 return true; 1582 return false; 1583 } 1584 1585 bool SIScheduleBlockScheduler::tryCandidateRegUsage(SIBlockSchedCandidate &Cand, 1586 SIBlockSchedCandidate &TryCand) { 1587 if (!Cand.isValid()) { 1588 TryCand.Reason = NodeOrder; 1589 return true; 1590 } 1591 1592 if (SISched::tryLess(TryCand.VGPRUsageDiff > 0, Cand.VGPRUsageDiff > 0, 1593 TryCand, Cand, RegUsage)) 1594 return true; 1595 if (SISched::tryGreater(TryCand.NumSuccessors > 0, 1596 Cand.NumSuccessors > 0, 1597 TryCand, Cand, Successor)) 1598 return true; 1599 if (SISched::tryGreater(TryCand.Height, Cand.Height, TryCand, Cand, Depth)) 1600 return true; 1601 if (SISched::tryLess(TryCand.VGPRUsageDiff, Cand.VGPRUsageDiff, 1602 TryCand, Cand, RegUsage)) 1603 return true; 1604 return false; 1605 } 1606 1607 SIScheduleBlock *SIScheduleBlockScheduler::pickBlock() { 1608 SIBlockSchedCandidate Cand; 1609 std::vector<SIScheduleBlock*>::iterator Best; 1610 SIScheduleBlock *Block; 1611 if (ReadyBlocks.empty()) 1612 return nullptr; 1613 1614 DAG->fillVgprSgprCost(LiveRegs.begin(), LiveRegs.end(), 1615 VregCurrentUsage, SregCurrentUsage); 1616 if (VregCurrentUsage > maxVregUsage) 1617 maxVregUsage = VregCurrentUsage; 1618 if (SregCurrentUsage > maxSregUsage) 1619 maxSregUsage = SregCurrentUsage; 1620 LLVM_DEBUG(dbgs() << "Picking New Blocks\n"; dbgs() << "Available: "; 1621 for (SIScheduleBlock *Block 1622 : ReadyBlocks) dbgs() 1623 << Block->getID() << ' '; 1624 dbgs() << "\nCurrent Live:\n"; 1625 for (unsigned Reg 1626 : LiveRegs) dbgs() 1627 << printVRegOrUnit(Reg, DAG->getTRI()) << ' '; 1628 dbgs() << '\n'; 1629 dbgs() << "Current VGPRs: " << VregCurrentUsage << '\n'; 1630 dbgs() << "Current SGPRs: " << SregCurrentUsage << '\n';); 1631 1632 Cand.Block = nullptr; 1633 for (std::vector<SIScheduleBlock*>::iterator I = ReadyBlocks.begin(), 1634 E = ReadyBlocks.end(); I != E; ++I) { 1635 SIBlockSchedCandidate TryCand; 1636 TryCand.Block = *I; 1637 TryCand.IsHighLatency = TryCand.Block->isHighLatencyBlock(); 1638 TryCand.VGPRUsageDiff = 1639 checkRegUsageImpact(TryCand.Block->getInRegs(), 1640 TryCand.Block->getOutRegs())[DAG->getVGPRSetID()]; 1641 TryCand.NumSuccessors = TryCand.Block->getSuccs().size(); 1642 TryCand.NumHighLatencySuccessors = 1643 TryCand.Block->getNumHighLatencySuccessors(); 1644 TryCand.LastPosHighLatParentScheduled = 1645 (unsigned int) std::max<int> (0, 1646 LastPosHighLatencyParentScheduled[TryCand.Block->getID()] - 1647 LastPosWaitedHighLatency); 1648 TryCand.Height = TryCand.Block->Height; 1649 // Try not to increase VGPR usage too much, else we may spill. 1650 if (VregCurrentUsage > 120 || 1651 Variant != SISchedulerBlockSchedulerVariant::BlockLatencyRegUsage) { 1652 if (!tryCandidateRegUsage(Cand, TryCand) && 1653 Variant != SISchedulerBlockSchedulerVariant::BlockRegUsage) 1654 tryCandidateLatency(Cand, TryCand); 1655 } else { 1656 if (!tryCandidateLatency(Cand, TryCand)) 1657 tryCandidateRegUsage(Cand, TryCand); 1658 } 1659 if (TryCand.Reason != NoCand) { 1660 Cand.setBest(TryCand); 1661 Best = I; 1662 LLVM_DEBUG(dbgs() << "Best Current Choice: " << Cand.Block->getID() << ' ' 1663 << getReasonStr(Cand.Reason) << '\n'); 1664 } 1665 } 1666 1667 LLVM_DEBUG(dbgs() << "Picking: " << Cand.Block->getID() << '\n'; 1668 dbgs() << "Is a block with high latency instruction: " 1669 << (Cand.IsHighLatency ? "yes\n" : "no\n"); 1670 dbgs() << "Position of last high latency dependency: " 1671 << Cand.LastPosHighLatParentScheduled << '\n'; 1672 dbgs() << "VGPRUsageDiff: " << Cand.VGPRUsageDiff << '\n'; 1673 dbgs() << '\n';); 1674 1675 Block = Cand.Block; 1676 ReadyBlocks.erase(Best); 1677 return Block; 1678 } 1679 1680 // Tracking of currently alive registers to determine VGPR Usage. 1681 1682 void SIScheduleBlockScheduler::addLiveRegs(std::set<unsigned> &Regs) { 1683 for (unsigned Reg : Regs) { 1684 // For now only track virtual registers. 1685 if (!Register::isVirtualRegister(Reg)) 1686 continue; 1687 // If not already in the live set, then add it. 1688 (void) LiveRegs.insert(Reg); 1689 } 1690 } 1691 1692 void SIScheduleBlockScheduler::decreaseLiveRegs(SIScheduleBlock *Block, 1693 std::set<unsigned> &Regs) { 1694 for (unsigned Reg : Regs) { 1695 // For now only track virtual registers. 1696 std::set<unsigned>::iterator Pos = LiveRegs.find(Reg); 1697 assert (Pos != LiveRegs.end() && // Reg must be live. 1698 LiveRegsConsumers.find(Reg) != LiveRegsConsumers.end() && 1699 LiveRegsConsumers[Reg] >= 1); 1700 --LiveRegsConsumers[Reg]; 1701 if (LiveRegsConsumers[Reg] == 0) 1702 LiveRegs.erase(Pos); 1703 } 1704 } 1705 1706 void SIScheduleBlockScheduler::releaseBlockSuccs(SIScheduleBlock *Parent) { 1707 for (const auto &Block : Parent->getSuccs()) { 1708 if (--BlockNumPredsLeft[Block.first->getID()] == 0) 1709 ReadyBlocks.push_back(Block.first); 1710 1711 if (Parent->isHighLatencyBlock() && 1712 Block.second == SIScheduleBlockLinkKind::Data) 1713 LastPosHighLatencyParentScheduled[Block.first->getID()] = NumBlockScheduled; 1714 } 1715 } 1716 1717 void SIScheduleBlockScheduler::blockScheduled(SIScheduleBlock *Block) { 1718 decreaseLiveRegs(Block, Block->getInRegs()); 1719 addLiveRegs(Block->getOutRegs()); 1720 releaseBlockSuccs(Block); 1721 for (std::map<unsigned, unsigned>::iterator RegI = 1722 LiveOutRegsNumUsages[Block->getID()].begin(), 1723 E = LiveOutRegsNumUsages[Block->getID()].end(); RegI != E; ++RegI) { 1724 std::pair<unsigned, unsigned> RegP = *RegI; 1725 // We produce this register, thus it must not be previously alive. 1726 assert(LiveRegsConsumers.find(RegP.first) == LiveRegsConsumers.end() || 1727 LiveRegsConsumers[RegP.first] == 0); 1728 LiveRegsConsumers[RegP.first] += RegP.second; 1729 } 1730 if (LastPosHighLatencyParentScheduled[Block->getID()] > 1731 (unsigned)LastPosWaitedHighLatency) 1732 LastPosWaitedHighLatency = 1733 LastPosHighLatencyParentScheduled[Block->getID()]; 1734 ++NumBlockScheduled; 1735 } 1736 1737 std::vector<int> 1738 SIScheduleBlockScheduler::checkRegUsageImpact(std::set<unsigned> &InRegs, 1739 std::set<unsigned> &OutRegs) { 1740 std::vector<int> DiffSetPressure; 1741 DiffSetPressure.assign(DAG->getTRI()->getNumRegPressureSets(), 0); 1742 1743 for (unsigned Reg : InRegs) { 1744 // For now only track virtual registers. 1745 if (!Register::isVirtualRegister(Reg)) 1746 continue; 1747 if (LiveRegsConsumers[Reg] > 1) 1748 continue; 1749 PSetIterator PSetI = DAG->getMRI()->getPressureSets(Reg); 1750 for (; PSetI.isValid(); ++PSetI) { 1751 DiffSetPressure[*PSetI] -= PSetI.getWeight(); 1752 } 1753 } 1754 1755 for (unsigned Reg : OutRegs) { 1756 // For now only track virtual registers. 1757 if (!Register::isVirtualRegister(Reg)) 1758 continue; 1759 PSetIterator PSetI = DAG->getMRI()->getPressureSets(Reg); 1760 for (; PSetI.isValid(); ++PSetI) { 1761 DiffSetPressure[*PSetI] += PSetI.getWeight(); 1762 } 1763 } 1764 1765 return DiffSetPressure; 1766 } 1767 1768 // SIScheduler // 1769 1770 struct SIScheduleBlockResult 1771 SIScheduler::scheduleVariant(SISchedulerBlockCreatorVariant BlockVariant, 1772 SISchedulerBlockSchedulerVariant ScheduleVariant) { 1773 SIScheduleBlocks Blocks = BlockCreator.getBlocks(BlockVariant); 1774 SIScheduleBlockScheduler Scheduler(DAG, ScheduleVariant, Blocks); 1775 std::vector<SIScheduleBlock*> ScheduledBlocks; 1776 struct SIScheduleBlockResult Res; 1777 1778 ScheduledBlocks = Scheduler.getBlocks(); 1779 1780 for (unsigned b = 0; b < ScheduledBlocks.size(); ++b) { 1781 SIScheduleBlock *Block = ScheduledBlocks[b]; 1782 std::vector<SUnit*> SUs = Block->getScheduledUnits(); 1783 1784 for (SUnit* SU : SUs) 1785 Res.SUs.push_back(SU->NodeNum); 1786 } 1787 1788 Res.MaxSGPRUsage = Scheduler.getSGPRUsage(); 1789 Res.MaxVGPRUsage = Scheduler.getVGPRUsage(); 1790 return Res; 1791 } 1792 1793 // SIScheduleDAGMI // 1794 1795 SIScheduleDAGMI::SIScheduleDAGMI(MachineSchedContext *C) : 1796 ScheduleDAGMILive(C, std::make_unique<GenericScheduler>(C)) { 1797 SITII = static_cast<const SIInstrInfo*>(TII); 1798 SITRI = static_cast<const SIRegisterInfo*>(TRI); 1799 1800 VGPRSetID = SITRI->getVGPRPressureSet(); 1801 SGPRSetID = SITRI->getSGPRPressureSet(); 1802 } 1803 1804 SIScheduleDAGMI::~SIScheduleDAGMI() = default; 1805 1806 // Code adapted from scheduleDAG.cpp 1807 // Does a topological sort over the SUs. 1808 // Both TopDown and BottomUp 1809 void SIScheduleDAGMI::topologicalSort() { 1810 Topo.InitDAGTopologicalSorting(); 1811 1812 TopDownIndex2SU = std::vector<int>(Topo.begin(), Topo.end()); 1813 BottomUpIndex2SU = std::vector<int>(Topo.rbegin(), Topo.rend()); 1814 } 1815 1816 // Move low latencies further from their user without 1817 // increasing SGPR usage (in general) 1818 // This is to be replaced by a better pass that would 1819 // take into account SGPR usage (based on VGPR Usage 1820 // and the corresponding wavefront count), that would 1821 // try to merge groups of loads if it make sense, etc 1822 void SIScheduleDAGMI::moveLowLatencies() { 1823 unsigned DAGSize = SUnits.size(); 1824 int LastLowLatencyUser = -1; 1825 int LastLowLatencyPos = -1; 1826 1827 for (unsigned i = 0, e = ScheduledSUnits.size(); i != e; ++i) { 1828 SUnit *SU = &SUnits[ScheduledSUnits[i]]; 1829 bool IsLowLatencyUser = false; 1830 unsigned MinPos = 0; 1831 1832 for (SDep& PredDep : SU->Preds) { 1833 SUnit *Pred = PredDep.getSUnit(); 1834 if (SITII->isLowLatencyInstruction(*Pred->getInstr())) { 1835 IsLowLatencyUser = true; 1836 } 1837 if (Pred->NodeNum >= DAGSize) 1838 continue; 1839 unsigned PredPos = ScheduledSUnitsInv[Pred->NodeNum]; 1840 if (PredPos >= MinPos) 1841 MinPos = PredPos + 1; 1842 } 1843 1844 if (SITII->isLowLatencyInstruction(*SU->getInstr())) { 1845 unsigned BestPos = LastLowLatencyUser + 1; 1846 if ((int)BestPos <= LastLowLatencyPos) 1847 BestPos = LastLowLatencyPos + 1; 1848 if (BestPos < MinPos) 1849 BestPos = MinPos; 1850 if (BestPos < i) { 1851 for (unsigned u = i; u > BestPos; --u) { 1852 ++ScheduledSUnitsInv[ScheduledSUnits[u-1]]; 1853 ScheduledSUnits[u] = ScheduledSUnits[u-1]; 1854 } 1855 ScheduledSUnits[BestPos] = SU->NodeNum; 1856 ScheduledSUnitsInv[SU->NodeNum] = BestPos; 1857 } 1858 LastLowLatencyPos = BestPos; 1859 if (IsLowLatencyUser) 1860 LastLowLatencyUser = BestPos; 1861 } else if (IsLowLatencyUser) { 1862 LastLowLatencyUser = i; 1863 // Moves COPY instructions on which depends 1864 // the low latency instructions too. 1865 } else if (SU->getInstr()->getOpcode() == AMDGPU::COPY) { 1866 bool CopyForLowLat = false; 1867 for (SDep& SuccDep : SU->Succs) { 1868 SUnit *Succ = SuccDep.getSUnit(); 1869 if (SuccDep.isWeak() || Succ->NodeNum >= DAGSize) 1870 continue; 1871 if (SITII->isLowLatencyInstruction(*Succ->getInstr())) { 1872 CopyForLowLat = true; 1873 } 1874 } 1875 if (!CopyForLowLat) 1876 continue; 1877 if (MinPos < i) { 1878 for (unsigned u = i; u > MinPos; --u) { 1879 ++ScheduledSUnitsInv[ScheduledSUnits[u-1]]; 1880 ScheduledSUnits[u] = ScheduledSUnits[u-1]; 1881 } 1882 ScheduledSUnits[MinPos] = SU->NodeNum; 1883 ScheduledSUnitsInv[SU->NodeNum] = MinPos; 1884 } 1885 } 1886 } 1887 } 1888 1889 void SIScheduleDAGMI::restoreSULinksLeft() { 1890 for (unsigned i = 0, e = SUnits.size(); i != e; ++i) { 1891 SUnits[i].isScheduled = false; 1892 SUnits[i].WeakPredsLeft = SUnitsLinksBackup[i].WeakPredsLeft; 1893 SUnits[i].NumPredsLeft = SUnitsLinksBackup[i].NumPredsLeft; 1894 SUnits[i].WeakSuccsLeft = SUnitsLinksBackup[i].WeakSuccsLeft; 1895 SUnits[i].NumSuccsLeft = SUnitsLinksBackup[i].NumSuccsLeft; 1896 } 1897 } 1898 1899 // Return the Vgpr and Sgpr usage corresponding to some virtual registers. 1900 template<typename _Iterator> void 1901 SIScheduleDAGMI::fillVgprSgprCost(_Iterator First, _Iterator End, 1902 unsigned &VgprUsage, unsigned &SgprUsage) { 1903 VgprUsage = 0; 1904 SgprUsage = 0; 1905 for (_Iterator RegI = First; RegI != End; ++RegI) { 1906 unsigned Reg = *RegI; 1907 // For now only track virtual registers 1908 if (!Register::isVirtualRegister(Reg)) 1909 continue; 1910 PSetIterator PSetI = MRI.getPressureSets(Reg); 1911 for (; PSetI.isValid(); ++PSetI) { 1912 if (*PSetI == VGPRSetID) 1913 VgprUsage += PSetI.getWeight(); 1914 else if (*PSetI == SGPRSetID) 1915 SgprUsage += PSetI.getWeight(); 1916 } 1917 } 1918 } 1919 1920 void SIScheduleDAGMI::schedule() 1921 { 1922 SmallVector<SUnit*, 8> TopRoots, BotRoots; 1923 SIScheduleBlockResult Best, Temp; 1924 LLVM_DEBUG(dbgs() << "Preparing Scheduling\n"); 1925 1926 buildDAGWithRegPressure(); 1927 LLVM_DEBUG(dump()); 1928 1929 topologicalSort(); 1930 findRootsAndBiasEdges(TopRoots, BotRoots); 1931 // We reuse several ScheduleDAGMI and ScheduleDAGMILive 1932 // functions, but to make them happy we must initialize 1933 // the default Scheduler implementation (even if we do not 1934 // run it) 1935 SchedImpl->initialize(this); 1936 initQueues(TopRoots, BotRoots); 1937 1938 // Fill some stats to help scheduling. 1939 1940 SUnitsLinksBackup = SUnits; 1941 IsLowLatencySU.clear(); 1942 LowLatencyOffset.clear(); 1943 IsHighLatencySU.clear(); 1944 1945 IsLowLatencySU.resize(SUnits.size(), 0); 1946 LowLatencyOffset.resize(SUnits.size(), 0); 1947 IsHighLatencySU.resize(SUnits.size(), 0); 1948 1949 for (unsigned i = 0, e = (unsigned)SUnits.size(); i != e; ++i) { 1950 SUnit *SU = &SUnits[i]; 1951 const MachineOperand *BaseLatOp; 1952 int64_t OffLatReg; 1953 if (SITII->isLowLatencyInstruction(*SU->getInstr())) { 1954 IsLowLatencySU[i] = 1; 1955 if (SITII->getMemOperandWithOffset(*SU->getInstr(), BaseLatOp, OffLatReg, 1956 TRI)) 1957 LowLatencyOffset[i] = OffLatReg; 1958 } else if (SITII->isHighLatencyInstruction(*SU->getInstr())) 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