| 1 | //===-- SIMachineScheduler.h - SI Scheduler Interface -----------*- C++ -*-===// |
| 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 | #ifndef LLVM_LIB_TARGET_AMDGPU_SIMACHINESCHEDULER_H |
| 15 | #define LLVM_LIB_TARGET_AMDGPU_SIMACHINESCHEDULER_H |
| 16 | |
| 17 | #include "llvm/CodeGen/MachineScheduler.h" |
| 18 | #include "llvm/CodeGen/RegisterPressure.h" |
| 19 | #include "llvm/CodeGen/ScheduleDAG.h" |
| 20 | #include <cstdint> |
| 21 | #include <set> |
| 22 | #include <vector> |
| 23 | |
| 24 | namespace llvm { |
| 25 | |
| 26 | class SIInstrInfo; |
| 27 | class SIRegisterInfo; |
| 28 | class SIScheduleDAGMI; |
| 29 | class SIScheduleBlockCreator; |
| 30 | |
| 31 | enum SIScheduleCandReason { |
| 32 | NoCand, |
| 33 | RegUsage, |
| 34 | Latency, |
| 35 | Successor, |
| 36 | Depth, |
| 37 | NodeOrder |
| 38 | }; |
| 39 | |
| 40 | struct SISchedulerCandidate { |
| 41 | // The reason for this candidate. |
| 42 | SIScheduleCandReason Reason = NoCand; |
| 43 | |
| 44 | // Set of reasons that apply to multiple candidates. |
| 45 | uint32_t RepeatReasonSet = 0; |
| 46 | |
| 47 | SISchedulerCandidate() = default; |
| 48 | |
| 49 | bool isRepeat(SIScheduleCandReason R) { return RepeatReasonSet & (1 << R); } |
| 50 | void setRepeat(SIScheduleCandReason R) { RepeatReasonSet |= (1 << R); } |
| 51 | }; |
| 52 | |
| 53 | enum SIScheduleBlockLinkKind { |
| 54 | NoData, |
| 55 | Data |
| 56 | }; |
| 57 | |
| 58 | class SIScheduleBlock { |
| 59 | SIScheduleDAGMI *DAG; |
| 60 | SIScheduleBlockCreator *BC; |
| 61 | |
| 62 | std::vector<SUnit*> SUnits; |
| 63 | std::map<unsigned, unsigned> NodeNum2Index; |
| 64 | std::vector<SUnit*> TopReadySUs; |
| 65 | std::vector<SUnit*> ScheduledSUnits; |
| 66 | |
| 67 | /// The top of the unscheduled zone. |
| 68 | IntervalPressure TopPressure; |
| 69 | RegPressureTracker TopRPTracker; |
| 70 | |
| 71 | // Pressure: number of said class of registers needed to |
| 72 | // store the live virtual and real registers. |
| 73 | // We do care only of SGPR32 and VGPR32 and do track only virtual registers. |
| 74 | // Pressure of additional registers required inside the block. |
| 75 | std::vector<unsigned> InternalAdditionalPressure; |
| 76 | // Pressure of input and output registers |
| 77 | std::vector<unsigned> LiveInPressure; |
| 78 | std::vector<unsigned> LiveOutPressure; |
| 79 | // Registers required by the block, and outputs. |
| 80 | // We do track only virtual registers. |
| 81 | // Note that some registers are not 32 bits, |
| 82 | // and thus the pressure is not equal |
| 83 | // to the number of live registers. |
| 84 | std::set<Register> LiveInRegs; |
| 85 | std::set<Register> LiveOutRegs; |
| 86 | |
| 87 | bool Scheduled = false; |
| 88 | bool HighLatencyBlock = false; |
| 89 | |
| 90 | std::vector<unsigned> HasLowLatencyNonWaitedParent; |
| 91 | |
| 92 | // Unique ID, the index of the Block in the SIScheduleDAGMI Blocks table. |
| 93 | unsigned ID; |
| 94 | |
| 95 | std::vector<SIScheduleBlock*> Preds; // All blocks predecessors. |
| 96 | // All blocks successors, and the kind of link |
| 97 | std::vector<std::pair<SIScheduleBlock*, SIScheduleBlockLinkKind>> Succs; |
| 98 | unsigned NumHighLatencySuccessors = 0; |
| 99 | |
| 100 | public: |
| 101 | SIScheduleBlock(SIScheduleDAGMI *DAG, SIScheduleBlockCreator *BC, |
| 102 | unsigned ID): |
| 103 | DAG(DAG), BC(BC), TopRPTracker(TopPressure), ID(ID) {} |
| 104 | |
| 105 | ~SIScheduleBlock() = default; |
| 106 | |
| 107 | unsigned getID() const { return ID; } |
| 108 | |
| 109 | /// Functions for Block construction. |
| 110 | void addUnit(SUnit *SU); |
| 111 | |
| 112 | // When all SUs have been added. |
| 113 | void finalizeUnits(); |
| 114 | |
| 115 | // Add block pred, which has instruction predecessor of SU. |
| 116 | void addPred(SIScheduleBlock *Pred); |
| 117 | void addSucc(SIScheduleBlock *Succ, SIScheduleBlockLinkKind Kind); |
| 118 | |
| 119 | const std::vector<SIScheduleBlock*>& getPreds() const { return Preds; } |
| 120 | ArrayRef<std::pair<SIScheduleBlock*, SIScheduleBlockLinkKind>> |
| 121 | getSuccs() const { return Succs; } |
| 122 | |
| 123 | unsigned Height = 0; // Maximum topdown path length to block without outputs |
| 124 | unsigned Depth = 0; // Maximum bottomup path length to block without inputs |
| 125 | |
| 126 | unsigned getNumHighLatencySuccessors() const { |
| 127 | return NumHighLatencySuccessors; |
| 128 | } |
| 129 | |
| 130 | bool isHighLatencyBlock() { return HighLatencyBlock; } |
| 131 | |
| 132 | // This is approximative. |
| 133 | // Ideally should take into accounts some instructions (rcp, etc) |
| 134 | // are 4 times slower. |
| 135 | int getCost() { return SUnits.size(); } |
| 136 | |
| 137 | // The block Predecessors and Successors must be all registered |
| 138 | // before fastSchedule(). |
| 139 | // Fast schedule with no particular requirement. |
| 140 | void fastSchedule(); |
| 141 | |
| 142 | std::vector<SUnit*> getScheduledUnits() { return ScheduledSUnits; } |
| 143 | |
| 144 | // Complete schedule that will try to minimize reg pressure and |
| 145 | // low latencies, and will fill liveins and liveouts. |
| 146 | // Needs all MIs to be grouped between BeginBlock and EndBlock. |
| 147 | // The MIs can be moved after the scheduling, |
| 148 | // it is just used to allow correct track of live registers. |
| 149 | void schedule(MachineBasicBlock::iterator BeginBlock, |
| 150 | MachineBasicBlock::iterator EndBlock); |
| 151 | |
| 152 | bool isScheduled() { return Scheduled; } |
| 153 | |
| 154 | // Needs the block to be scheduled inside |
| 155 | // TODO: find a way to compute it. |
| 156 | std::vector<unsigned> &getInternalAdditionalRegUsage() { |
| 157 | return InternalAdditionalPressure; |
| 158 | } |
| 159 | |
| 160 | std::set<Register> &getInRegs() { return LiveInRegs; } |
| 161 | std::set<Register> &getOutRegs() { return LiveOutRegs; } |
| 162 | |
| 163 | void printDebug(bool Full); |
| 164 | |
| 165 | private: |
| 166 | struct SISchedCandidate : SISchedulerCandidate { |
| 167 | // The best SUnit candidate. |
| 168 | SUnit *SU = nullptr; |
| 169 | |
| 170 | unsigned SGPRUsage; |
| 171 | unsigned VGPRUsage; |
| 172 | bool IsLowLatency; |
| 173 | unsigned LowLatencyOffset; |
| 174 | bool HasLowLatencyNonWaitedParent; |
| 175 | |
| 176 | SISchedCandidate() = default; |
| 177 | |
| 178 | bool isValid() const { return SU; } |
| 179 | |
| 180 | // Copy the status of another candidate without changing policy. |
| 181 | void setBest(SISchedCandidate &Best) { |
| 182 | assert(Best.Reason != NoCand && "uninitialized Sched candidate" ); |
| 183 | SU = Best.SU; |
| 184 | Reason = Best.Reason; |
| 185 | SGPRUsage = Best.SGPRUsage; |
| 186 | VGPRUsage = Best.VGPRUsage; |
| 187 | IsLowLatency = Best.IsLowLatency; |
| 188 | LowLatencyOffset = Best.LowLatencyOffset; |
| 189 | HasLowLatencyNonWaitedParent = Best.HasLowLatencyNonWaitedParent; |
| 190 | } |
| 191 | }; |
| 192 | |
| 193 | void undoSchedule(); |
| 194 | |
| 195 | void undoReleaseSucc(SUnit *SU, SDep *SuccEdge); |
| 196 | void releaseSucc(SUnit *SU, SDep *SuccEdge); |
| 197 | // InOrOutBlock: restrict to links pointing inside the block (true), |
| 198 | // or restrict to links pointing outside the block (false). |
| 199 | void releaseSuccessors(SUnit *SU, bool InOrOutBlock); |
| 200 | |
| 201 | void nodeScheduled(SUnit *SU); |
| 202 | void tryCandidateTopDown(SISchedCandidate &Cand, SISchedCandidate &TryCand); |
| 203 | void tryCandidateBottomUp(SISchedCandidate &Cand, SISchedCandidate &TryCand); |
| 204 | SUnit* pickNode(); |
| 205 | void traceCandidate(const SISchedCandidate &Cand); |
| 206 | void initRegPressure(MachineBasicBlock::iterator BeginBlock, |
| 207 | MachineBasicBlock::iterator EndBlock); |
| 208 | }; |
| 209 | |
| 210 | struct SIScheduleBlocks { |
| 211 | std::vector<SIScheduleBlock*> Blocks; |
| 212 | std::vector<int> TopDownIndex2Block; |
| 213 | std::vector<int> TopDownBlock2Index; |
| 214 | }; |
| 215 | |
| 216 | enum SISchedulerBlockCreatorVariant { |
| 217 | LatenciesAlone, |
| 218 | LatenciesGrouped, |
| 219 | LatenciesAlonePlusConsecutive |
| 220 | }; |
| 221 | |
| 222 | class SIScheduleBlockCreator { |
| 223 | SIScheduleDAGMI *DAG; |
| 224 | // unique_ptr handles freeing memory for us. |
| 225 | std::vector<std::unique_ptr<SIScheduleBlock>> BlockPtrs; |
| 226 | std::map<SISchedulerBlockCreatorVariant, |
| 227 | SIScheduleBlocks> Blocks; |
| 228 | std::vector<SIScheduleBlock*> CurrentBlocks; |
| 229 | std::vector<int> Node2CurrentBlock; |
| 230 | |
| 231 | // Topological sort |
| 232 | // Maps topological index to the node number. |
| 233 | std::vector<int> TopDownIndex2Block; |
| 234 | std::vector<int> TopDownBlock2Index; |
| 235 | std::vector<int> BottomUpIndex2Block; |
| 236 | |
| 237 | // 0 -> Color not given. |
| 238 | // 1 to SUnits.size() -> Reserved group (you should only add elements to them). |
| 239 | // Above -> Other groups. |
| 240 | int NextReservedID; |
| 241 | int NextNonReservedID; |
| 242 | std::vector<int> CurrentColoring; |
| 243 | std::vector<int> CurrentTopDownReservedDependencyColoring; |
| 244 | std::vector<int> CurrentBottomUpReservedDependencyColoring; |
| 245 | |
| 246 | public: |
| 247 | SIScheduleBlockCreator(SIScheduleDAGMI *DAG); |
| 248 | |
| 249 | SIScheduleBlocks |
| 250 | getBlocks(SISchedulerBlockCreatorVariant BlockVariant); |
| 251 | |
| 252 | bool isSUInBlock(SUnit *SU, unsigned ID); |
| 253 | |
| 254 | private: |
| 255 | // Give a Reserved color to every high latency. |
| 256 | void colorHighLatenciesAlone(); |
| 257 | |
| 258 | // Create groups of high latencies with a Reserved color. |
| 259 | void colorHighLatenciesGroups(); |
| 260 | |
| 261 | // Compute coloring for topdown and bottom traversals with |
| 262 | // different colors depending on dependencies on Reserved colors. |
| 263 | void colorComputeReservedDependencies(); |
| 264 | |
| 265 | // Give color to all non-colored SUs according to Reserved groups dependencies. |
| 266 | void colorAccordingToReservedDependencies(); |
| 267 | |
| 268 | // Divides Blocks having no bottom up or top down dependencies on Reserved groups. |
| 269 | // The new colors are computed according to the dependencies on the other blocks |
| 270 | // formed with colorAccordingToReservedDependencies. |
| 271 | void colorEndsAccordingToDependencies(); |
| 272 | |
| 273 | // Cut groups into groups with SUs in consecutive order (except for Reserved groups). |
| 274 | void colorForceConsecutiveOrderInGroup(); |
| 275 | |
| 276 | // Merge Constant loads that have all their users into another group to the group. |
| 277 | // (TODO: else if all their users depend on the same group, put them there) |
| 278 | void colorMergeConstantLoadsNextGroup(); |
| 279 | |
| 280 | // Merge SUs that have all their users into another group to the group |
| 281 | void colorMergeIfPossibleNextGroup(); |
| 282 | |
| 283 | // Merge SUs that have all their users into another group to the group, |
| 284 | // but only for Reserved groups. |
| 285 | void colorMergeIfPossibleNextGroupOnlyForReserved(); |
| 286 | |
| 287 | // Merge SUs that have all their users into another group to the group, |
| 288 | // but only if the group is no more than a few SUs. |
| 289 | void colorMergeIfPossibleSmallGroupsToNextGroup(); |
| 290 | |
| 291 | // Divides Blocks with important size. |
| 292 | // Idea of implementation: attribute new colors depending on topdown and |
| 293 | // bottom up links to other blocks. |
| 294 | void cutHugeBlocks(); |
| 295 | |
| 296 | // Put in one group all instructions with no users in this scheduling region |
| 297 | // (we'd want these groups be at the end). |
| 298 | void regroupNoUserInstructions(); |
| 299 | |
| 300 | // Give Reserved color to export instructions |
| 301 | void colorExports(); |
| 302 | |
| 303 | void createBlocksForVariant(SISchedulerBlockCreatorVariant BlockVariant); |
| 304 | |
| 305 | void topologicalSort(); |
| 306 | |
| 307 | void scheduleInsideBlocks(); |
| 308 | |
| 309 | void fillStats(); |
| 310 | }; |
| 311 | |
| 312 | enum SISchedulerBlockSchedulerVariant { |
| 313 | BlockLatencyRegUsage, |
| 314 | BlockRegUsageLatency, |
| 315 | BlockRegUsage |
| 316 | }; |
| 317 | |
| 318 | class SIScheduleBlockScheduler { |
| 319 | SIScheduleDAGMI *DAG; |
| 320 | SISchedulerBlockSchedulerVariant Variant; |
| 321 | std::vector<SIScheduleBlock*> Blocks; |
| 322 | |
| 323 | std::vector<std::map<Register, unsigned>> LiveOutRegsNumUsages; |
| 324 | std::set<Register> LiveRegs; |
| 325 | // Num of schedulable unscheduled blocks reading the register. |
| 326 | std::map<Register, unsigned> LiveRegsConsumers; |
| 327 | |
| 328 | std::vector<unsigned> LastPosHighLatencyParentScheduled; |
| 329 | int LastPosWaitedHighLatency; |
| 330 | |
| 331 | std::vector<SIScheduleBlock*> BlocksScheduled; |
| 332 | unsigned NumBlockScheduled; |
| 333 | std::vector<SIScheduleBlock*> ReadyBlocks; |
| 334 | |
| 335 | unsigned VregCurrentUsage; |
| 336 | unsigned SregCurrentUsage; |
| 337 | |
| 338 | // Currently is only approximation. |
| 339 | unsigned maxVregUsage; |
| 340 | unsigned maxSregUsage; |
| 341 | |
| 342 | std::vector<unsigned> BlockNumPredsLeft; |
| 343 | std::vector<unsigned> BlockNumSuccsLeft; |
| 344 | |
| 345 | public: |
| 346 | SIScheduleBlockScheduler(SIScheduleDAGMI *DAG, |
| 347 | SISchedulerBlockSchedulerVariant Variant, |
| 348 | SIScheduleBlocks BlocksStruct); |
| 349 | ~SIScheduleBlockScheduler() = default; |
| 350 | |
| 351 | std::vector<SIScheduleBlock*> getBlocks() { return BlocksScheduled; } |
| 352 | |
| 353 | unsigned getVGPRUsage() { return maxVregUsage; } |
| 354 | unsigned getSGPRUsage() { return maxSregUsage; } |
| 355 | |
| 356 | private: |
| 357 | struct SIBlockSchedCandidate : SISchedulerCandidate { |
| 358 | // The best Block candidate. |
| 359 | SIScheduleBlock *Block = nullptr; |
| 360 | |
| 361 | bool IsHighLatency; |
| 362 | int VGPRUsageDiff; |
| 363 | unsigned NumSuccessors; |
| 364 | unsigned NumHighLatencySuccessors; |
| 365 | unsigned LastPosHighLatParentScheduled; |
| 366 | unsigned Height; |
| 367 | |
| 368 | SIBlockSchedCandidate() = default; |
| 369 | |
| 370 | bool isValid() const { return Block; } |
| 371 | |
| 372 | // Copy the status of another candidate without changing policy. |
| 373 | void setBest(SIBlockSchedCandidate &Best) { |
| 374 | assert(Best.Reason != NoCand && "uninitialized Sched candidate" ); |
| 375 | Block = Best.Block; |
| 376 | Reason = Best.Reason; |
| 377 | IsHighLatency = Best.IsHighLatency; |
| 378 | VGPRUsageDiff = Best.VGPRUsageDiff; |
| 379 | NumSuccessors = Best.NumSuccessors; |
| 380 | NumHighLatencySuccessors = Best.NumHighLatencySuccessors; |
| 381 | LastPosHighLatParentScheduled = Best.LastPosHighLatParentScheduled; |
| 382 | Height = Best.Height; |
| 383 | } |
| 384 | }; |
| 385 | |
| 386 | bool tryCandidateLatency(SIBlockSchedCandidate &Cand, |
| 387 | SIBlockSchedCandidate &TryCand); |
| 388 | bool tryCandidateRegUsage(SIBlockSchedCandidate &Cand, |
| 389 | SIBlockSchedCandidate &TryCand); |
| 390 | SIScheduleBlock *pickBlock(); |
| 391 | |
| 392 | void addLiveRegs(std::set<Register> &Regs); |
| 393 | void decreaseLiveRegs(SIScheduleBlock *Block, std::set<Register> &Regs); |
| 394 | void releaseBlockSuccs(SIScheduleBlock *Parent); |
| 395 | void blockScheduled(SIScheduleBlock *Block); |
| 396 | |
| 397 | // Check register pressure change |
| 398 | // by scheduling a block with these LiveIn and LiveOut. |
| 399 | std::vector<int> checkRegUsageImpact(std::set<Register> &InRegs, |
| 400 | std::set<Register> &OutRegs); |
| 401 | |
| 402 | void schedule(); |
| 403 | }; |
| 404 | |
| 405 | struct SIScheduleBlockResult { |
| 406 | std::vector<unsigned> SUs; |
| 407 | unsigned MaxSGPRUsage; |
| 408 | unsigned MaxVGPRUsage; |
| 409 | }; |
| 410 | |
| 411 | class SIScheduler { |
| 412 | SIScheduleDAGMI *DAG; |
| 413 | SIScheduleBlockCreator BlockCreator; |
| 414 | |
| 415 | public: |
| 416 | SIScheduler(SIScheduleDAGMI *DAG) : DAG(DAG), BlockCreator(DAG) {} |
| 417 | |
| 418 | ~SIScheduler() = default; |
| 419 | |
| 420 | struct SIScheduleBlockResult |
| 421 | scheduleVariant(SISchedulerBlockCreatorVariant BlockVariant, |
| 422 | SISchedulerBlockSchedulerVariant ScheduleVariant); |
| 423 | }; |
| 424 | |
| 425 | class SIScheduleDAGMI final : public ScheduleDAGMILive { |
| 426 | const SIInstrInfo *SITII; |
| 427 | const SIRegisterInfo *SITRI; |
| 428 | |
| 429 | std::vector<SUnit> SUnitsLinksBackup; |
| 430 | |
| 431 | // For moveLowLatencies. After all Scheduling variants are tested. |
| 432 | std::vector<unsigned> ScheduledSUnits; |
| 433 | std::vector<unsigned> ScheduledSUnitsInv; |
| 434 | |
| 435 | public: |
| 436 | SIScheduleDAGMI(MachineSchedContext *C); |
| 437 | |
| 438 | ~SIScheduleDAGMI() override; |
| 439 | |
| 440 | // Entry point for the schedule. |
| 441 | void schedule() override; |
| 442 | |
| 443 | // To init Block's RPTracker. |
| 444 | void initRPTracker(RegPressureTracker &RPTracker) { |
| 445 | RPTracker.init(mf: &MF, rci: RegClassInfo, lis: LIS, mbb: BB, pos: RegionBegin, TrackLaneMasks: false, TrackUntiedDefs: false); |
| 446 | } |
| 447 | |
| 448 | MachineBasicBlock *getBB() { return BB; } |
| 449 | MachineBasicBlock::iterator getCurrentTop() { return CurrentTop; } |
| 450 | MachineBasicBlock::iterator getCurrentBottom() { return CurrentBottom; } |
| 451 | LiveIntervals *getLIS() { return LIS; } |
| 452 | MachineRegisterInfo *getMRI() { return &MRI; } |
| 453 | const TargetRegisterInfo *getTRI() { return TRI; } |
| 454 | ScheduleDAGTopologicalSort *GetTopo() { return &Topo; } |
| 455 | SUnit &getEntrySU() { return EntrySU; } |
| 456 | SUnit& getExitSU() { return ExitSU; } |
| 457 | |
| 458 | void restoreSULinksLeft(); |
| 459 | |
| 460 | template<typename _Iterator> void fillVgprSgprCost(_Iterator First, |
| 461 | _Iterator End, |
| 462 | unsigned &VgprUsage, |
| 463 | unsigned &SgprUsage); |
| 464 | |
| 465 | std::set<Register> getInRegs() { |
| 466 | std::set<Register> InRegs; |
| 467 | for (const auto &RegMaskPair : RPTracker.getPressure().LiveInRegs) { |
| 468 | InRegs.insert(x: RegMaskPair.RegUnit); |
| 469 | } |
| 470 | return InRegs; |
| 471 | } |
| 472 | |
| 473 | std::set<unsigned> getOutRegs() { |
| 474 | std::set<unsigned> OutRegs; |
| 475 | for (const auto &RegMaskPair : RPTracker.getPressure().LiveOutRegs) { |
| 476 | OutRegs.insert(x: RegMaskPair.RegUnit); |
| 477 | } |
| 478 | return OutRegs; |
| 479 | }; |
| 480 | |
| 481 | private: |
| 482 | void topologicalSort(); |
| 483 | // After scheduling is done, improve low latency placements. |
| 484 | void moveLowLatencies(); |
| 485 | |
| 486 | public: |
| 487 | // Some stats for scheduling inside blocks. |
| 488 | std::vector<unsigned> IsLowLatencySU; |
| 489 | std::vector<unsigned> LowLatencyOffset; |
| 490 | std::vector<unsigned> IsHighLatencySU; |
| 491 | // Topological sort |
| 492 | // Maps topological index to the node number. |
| 493 | std::vector<int> TopDownIndex2SU; |
| 494 | std::vector<int> BottomUpIndex2SU; |
| 495 | }; |
| 496 | |
| 497 | } // end namespace llvm |
| 498 | |
| 499 | #endif // LLVM_LIB_TARGET_AMDGPU_SIMACHINESCHEDULER_H |
| 500 | |