1 //===- MachinePipeliner.cpp - Machine Software Pipeliner Pass -------------===// 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 // An implementation of the Swing Modulo Scheduling (SMS) software pipeliner. 10 // 11 // This SMS implementation is a target-independent back-end pass. When enabled, 12 // the pass runs just prior to the register allocation pass, while the machine 13 // IR is in SSA form. If software pipelining is successful, then the original 14 // loop is replaced by the optimized loop. The optimized loop contains one or 15 // more prolog blocks, the pipelined kernel, and one or more epilog blocks. If 16 // the instructions cannot be scheduled in a given MII, we increase the MII by 17 // one and try again. 18 // 19 // The SMS implementation is an extension of the ScheduleDAGInstrs class. We 20 // represent loop carried dependences in the DAG as order edges to the Phi 21 // nodes. We also perform several passes over the DAG to eliminate unnecessary 22 // edges that inhibit the ability to pipeline. The implementation uses the 23 // DFAPacketizer class to compute the minimum initiation interval and the check 24 // where an instruction may be inserted in the pipelined schedule. 25 // 26 // In order for the SMS pass to work, several target specific hooks need to be 27 // implemented to get information about the loop structure and to rewrite 28 // instructions. 29 // 30 //===----------------------------------------------------------------------===// 31 32 #include "llvm/CodeGen/MachinePipeliner.h" 33 #include "llvm/ADT/ArrayRef.h" 34 #include "llvm/ADT/BitVector.h" 35 #include "llvm/ADT/DenseMap.h" 36 #include "llvm/ADT/PriorityQueue.h" 37 #include "llvm/ADT/STLExtras.h" 38 #include "llvm/ADT/SetOperations.h" 39 #include "llvm/ADT/SetVector.h" 40 #include "llvm/ADT/SmallPtrSet.h" 41 #include "llvm/ADT/SmallSet.h" 42 #include "llvm/ADT/SmallVector.h" 43 #include "llvm/ADT/Statistic.h" 44 #include "llvm/ADT/iterator_range.h" 45 #include "llvm/Analysis/AliasAnalysis.h" 46 #include "llvm/Analysis/MemoryLocation.h" 47 #include "llvm/Analysis/OptimizationRemarkEmitter.h" 48 #include "llvm/Analysis/ValueTracking.h" 49 #include "llvm/CodeGen/DFAPacketizer.h" 50 #include "llvm/CodeGen/LiveIntervals.h" 51 #include "llvm/CodeGen/MachineBasicBlock.h" 52 #include "llvm/CodeGen/MachineDominators.h" 53 #include "llvm/CodeGen/MachineFunction.h" 54 #include "llvm/CodeGen/MachineFunctionPass.h" 55 #include "llvm/CodeGen/MachineInstr.h" 56 #include "llvm/CodeGen/MachineInstrBuilder.h" 57 #include "llvm/CodeGen/MachineLoopInfo.h" 58 #include "llvm/CodeGen/MachineMemOperand.h" 59 #include "llvm/CodeGen/MachineOperand.h" 60 #include "llvm/CodeGen/MachineRegisterInfo.h" 61 #include "llvm/CodeGen/ModuloSchedule.h" 62 #include "llvm/CodeGen/Register.h" 63 #include "llvm/CodeGen/RegisterClassInfo.h" 64 #include "llvm/CodeGen/RegisterPressure.h" 65 #include "llvm/CodeGen/ScheduleDAG.h" 66 #include "llvm/CodeGen/ScheduleDAGMutation.h" 67 #include "llvm/CodeGen/TargetInstrInfo.h" 68 #include "llvm/CodeGen/TargetOpcodes.h" 69 #include "llvm/CodeGen/TargetPassConfig.h" 70 #include "llvm/CodeGen/TargetRegisterInfo.h" 71 #include "llvm/CodeGen/TargetSubtargetInfo.h" 72 #include "llvm/Config/llvm-config.h" 73 #include "llvm/IR/Attributes.h" 74 #include "llvm/IR/Function.h" 75 #include "llvm/MC/LaneBitmask.h" 76 #include "llvm/MC/MCInstrDesc.h" 77 #include "llvm/MC/MCInstrItineraries.h" 78 #include "llvm/Pass.h" 79 #include "llvm/Support/CommandLine.h" 80 #include "llvm/Support/Compiler.h" 81 #include "llvm/Support/Debug.h" 82 #include "llvm/Support/raw_ostream.h" 83 #include <algorithm> 84 #include <cassert> 85 #include <climits> 86 #include <cstdint> 87 #include <deque> 88 #include <functional> 89 #include <iomanip> 90 #include <iterator> 91 #include <map> 92 #include <memory> 93 #include <sstream> 94 #include <tuple> 95 #include <utility> 96 #include <vector> 97 98 using namespace llvm; 99 100 #define DEBUG_TYPE "pipeliner" 101 102 STATISTIC(NumTrytoPipeline, "Number of loops that we attempt to pipeline"); 103 STATISTIC(NumPipelined, "Number of loops software pipelined"); 104 STATISTIC(NumNodeOrderIssues, "Number of node order issues found"); 105 STATISTIC(NumFailBranch, "Pipeliner abort due to unknown branch"); 106 STATISTIC(NumFailLoop, "Pipeliner abort due to unsupported loop"); 107 STATISTIC(NumFailPreheader, "Pipeliner abort due to missing preheader"); 108 STATISTIC(NumFailLargeMaxMII, "Pipeliner abort due to MaxMII too large"); 109 STATISTIC(NumFailZeroMII, "Pipeliner abort due to zero MII"); 110 STATISTIC(NumFailNoSchedule, "Pipeliner abort due to no schedule found"); 111 STATISTIC(NumFailZeroStage, "Pipeliner abort due to zero stage"); 112 STATISTIC(NumFailLargeMaxStage, "Pipeliner abort due to too many stages"); 113 STATISTIC(NumFailTooManyStores, "Pipeliner abort due to too many stores"); 114 115 /// A command line option to turn software pipelining on or off. 116 static cl::opt<bool> EnableSWP("enable-pipeliner", cl::Hidden, cl::init(true), 117 cl::desc("Enable Software Pipelining")); 118 119 /// A command line option to enable SWP at -Os. 120 static cl::opt<bool> EnableSWPOptSize("enable-pipeliner-opt-size", 121 cl::desc("Enable SWP at Os."), cl::Hidden, 122 cl::init(false)); 123 124 /// A command line argument to limit minimum initial interval for pipelining. 125 static cl::opt<int> SwpMaxMii("pipeliner-max-mii", 126 cl::desc("Size limit for the MII."), 127 cl::Hidden, cl::init(27)); 128 129 /// A command line argument to force pipeliner to use specified initial 130 /// interval. 131 static cl::opt<int> SwpForceII("pipeliner-force-ii", 132 cl::desc("Force pipeliner to use specified II."), 133 cl::Hidden, cl::init(-1)); 134 135 /// A command line argument to limit the number of stages in the pipeline. 136 static cl::opt<int> 137 SwpMaxStages("pipeliner-max-stages", 138 cl::desc("Maximum stages allowed in the generated scheduled."), 139 cl::Hidden, cl::init(3)); 140 141 /// A command line option to disable the pruning of chain dependences due to 142 /// an unrelated Phi. 143 static cl::opt<bool> 144 SwpPruneDeps("pipeliner-prune-deps", 145 cl::desc("Prune dependences between unrelated Phi nodes."), 146 cl::Hidden, cl::init(true)); 147 148 /// A command line option to disable the pruning of loop carried order 149 /// dependences. 150 static cl::opt<bool> 151 SwpPruneLoopCarried("pipeliner-prune-loop-carried", 152 cl::desc("Prune loop carried order dependences."), 153 cl::Hidden, cl::init(true)); 154 155 #ifndef NDEBUG 156 static cl::opt<int> SwpLoopLimit("pipeliner-max", cl::Hidden, cl::init(-1)); 157 #endif 158 159 static cl::opt<bool> SwpIgnoreRecMII("pipeliner-ignore-recmii", 160 cl::ReallyHidden, 161 cl::desc("Ignore RecMII")); 162 163 static cl::opt<bool> SwpShowResMask("pipeliner-show-mask", cl::Hidden, 164 cl::init(false)); 165 static cl::opt<bool> SwpDebugResource("pipeliner-dbg-res", cl::Hidden, 166 cl::init(false)); 167 168 static cl::opt<bool> EmitTestAnnotations( 169 "pipeliner-annotate-for-testing", cl::Hidden, cl::init(false), 170 cl::desc("Instead of emitting the pipelined code, annotate instructions " 171 "with the generated schedule for feeding into the " 172 "-modulo-schedule-test pass")); 173 174 static cl::opt<bool> ExperimentalCodeGen( 175 "pipeliner-experimental-cg", cl::Hidden, cl::init(false), 176 cl::desc( 177 "Use the experimental peeling code generator for software pipelining")); 178 179 static cl::opt<int> SwpIISearchRange("pipeliner-ii-search-range", 180 cl::desc("Range to search for II"), 181 cl::Hidden, cl::init(10)); 182 183 static cl::opt<bool> 184 LimitRegPressure("pipeliner-register-pressure", cl::Hidden, cl::init(false), 185 cl::desc("Limit register pressure of scheduled loop")); 186 187 static cl::opt<int> 188 RegPressureMargin("pipeliner-register-pressure-margin", cl::Hidden, 189 cl::init(5), 190 cl::desc("Margin representing the unused percentage of " 191 "the register pressure limit")); 192 193 static cl::opt<bool> 194 MVECodeGen("pipeliner-mve-cg", cl::Hidden, cl::init(false), 195 cl::desc("Use the MVE code generator for software pipelining")); 196 197 /// A command line argument to limit the number of store instructions in the 198 /// target basic block. 199 static cl::opt<unsigned> SwpMaxNumStores( 200 "pipeliner-max-num-stores", 201 cl::desc("Maximum number of stores allwed in the target loop."), cl::Hidden, 202 cl::init(200)); 203 204 namespace llvm { 205 206 // A command line option to enable the CopyToPhi DAG mutation. 207 cl::opt<bool> SwpEnableCopyToPhi("pipeliner-enable-copytophi", cl::ReallyHidden, 208 cl::init(true), 209 cl::desc("Enable CopyToPhi DAG Mutation")); 210 211 /// A command line argument to force pipeliner to use specified issue 212 /// width. 213 cl::opt<int> SwpForceIssueWidth( 214 "pipeliner-force-issue-width", 215 cl::desc("Force pipeliner to use specified issue width."), cl::Hidden, 216 cl::init(-1)); 217 218 /// A command line argument to set the window scheduling option. 219 static cl::opt<WindowSchedulingFlag> WindowSchedulingOption( 220 "window-sched", cl::Hidden, cl::init(WindowSchedulingFlag::WS_On), 221 cl::desc("Set how to use window scheduling algorithm."), 222 cl::values(clEnumValN(WindowSchedulingFlag::WS_Off, "off", 223 "Turn off window algorithm."), 224 clEnumValN(WindowSchedulingFlag::WS_On, "on", 225 "Use window algorithm after SMS algorithm fails."), 226 clEnumValN(WindowSchedulingFlag::WS_Force, "force", 227 "Use window algorithm instead of SMS algorithm."))); 228 229 } // end namespace llvm 230 231 unsigned SwingSchedulerDAG::Circuits::MaxPaths = 5; 232 char MachinePipeliner::ID = 0; 233 #ifndef NDEBUG 234 int MachinePipeliner::NumTries = 0; 235 #endif 236 char &llvm::MachinePipelinerID = MachinePipeliner::ID; 237 238 INITIALIZE_PASS_BEGIN(MachinePipeliner, DEBUG_TYPE, 239 "Modulo Software Pipelining", false, false) 240 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass) 241 INITIALIZE_PASS_DEPENDENCY(MachineLoopInfoWrapperPass) 242 INITIALIZE_PASS_DEPENDENCY(MachineDominatorTreeWrapperPass) 243 INITIALIZE_PASS_DEPENDENCY(LiveIntervalsWrapperPass) 244 INITIALIZE_PASS_END(MachinePipeliner, DEBUG_TYPE, 245 "Modulo Software Pipelining", false, false) 246 247 namespace { 248 249 /// This class holds an SUnit corresponding to a memory operation and other 250 /// information related to the instruction. 251 struct SUnitWithMemInfo { 252 SUnit *SU; 253 SmallVector<const Value *, 2> UnderlyingObjs; 254 255 /// The value of a memory operand. 256 const Value *MemOpValue = nullptr; 257 258 /// The offset of a memory operand. 259 int64_t MemOpOffset = 0; 260 261 AAMDNodes AATags; 262 263 /// True if all the underlying objects are identified. 264 bool IsAllIdentified = false; 265 266 SUnitWithMemInfo(SUnit *SU); 267 268 bool isTriviallyDisjoint(const SUnitWithMemInfo &Other) const; 269 270 bool isUnknown() const { return MemOpValue == nullptr; } 271 272 private: 273 bool getUnderlyingObjects(); 274 }; 275 276 /// Add loop-carried chain dependencies. This class handles the same type of 277 /// dependencies added by `ScheduleDAGInstrs::buildSchedGraph`, but takes into 278 /// account dependencies across iterations. 279 class LoopCarriedOrderDepsTracker { 280 // Type of instruction that is relevant to order-dependencies 281 enum class InstrTag { 282 Barrier = 0, ///< A barrier event instruction. 283 LoadOrStore = 1, ///< An instruction that may load or store memory, but is 284 ///< not a barrier event. 285 FPExceptions = 2, ///< An instruction that does not match above, but may 286 ///< raise floatin-point exceptions. 287 }; 288 289 struct TaggedSUnit : PointerIntPair<SUnit *, 2> { 290 TaggedSUnit(SUnit *SU, InstrTag Tag) 291 : PointerIntPair<SUnit *, 2>(SU, unsigned(Tag)) {} 292 293 InstrTag getTag() const { return InstrTag(getInt()); } 294 }; 295 296 /// Holds loads and stores with memory related information. 297 struct LoadStoreChunk { 298 SmallVector<SUnitWithMemInfo, 4> Loads; 299 SmallVector<SUnitWithMemInfo, 4> Stores; 300 301 void append(SUnit *SU); 302 }; 303 304 SwingSchedulerDAG *DAG; 305 BatchAAResults *BAA; 306 std::vector<SUnit> &SUnits; 307 308 /// The size of SUnits, for convenience. 309 const unsigned N; 310 311 /// Loop-carried Edges. 312 std::vector<BitVector> LoopCarried; 313 314 /// Instructions related to chain dependencies. They are one of the 315 /// following: 316 /// 317 /// 1. Barrier event. 318 /// 2. Load, but neither a barrier event, invariant load, nor may load trap 319 /// value. 320 /// 3. Store, but not a barrier event. 321 /// 4. None of them, but may raise floating-point exceptions. 322 /// 323 /// This is used when analyzing loop-carried dependencies that access global 324 /// barrier instructions. 325 std::vector<TaggedSUnit> TaggedSUnits; 326 327 const TargetInstrInfo *TII = nullptr; 328 const TargetRegisterInfo *TRI = nullptr; 329 330 public: 331 LoopCarriedOrderDepsTracker(SwingSchedulerDAG *SSD, BatchAAResults *BAA, 332 const TargetInstrInfo *TII, 333 const TargetRegisterInfo *TRI); 334 335 /// The main function to compute loop-carried order-dependencies. 336 void computeDependencies(); 337 338 const BitVector &getLoopCarried(unsigned Idx) const { 339 return LoopCarried[Idx]; 340 } 341 342 private: 343 /// Tags to \p SU if the instruction may affect the order-dependencies. 344 std::optional<InstrTag> getInstrTag(SUnit *SU) const; 345 346 void addLoopCarriedDepenenciesForChunks(const LoadStoreChunk &From, 347 const LoadStoreChunk &To); 348 349 /// Add a loop-carried order dependency between \p Src and \p Dst if we 350 /// cannot prove they are independent. When \p PerformCheapCheck is true, a 351 /// lightweight dependency test (referred to as "cheap check" below) is 352 /// performed at first. Note that the cheap check is retained to maintain the 353 /// existing behavior and not expected to be used anymore. 354 /// 355 /// TODO: Remove \p PerformCheapCheck and the corresponding cheap check. 356 void addDependenciesBetweenSUs(const SUnitWithMemInfo &Src, 357 const SUnitWithMemInfo &Dst, 358 bool PerformCheapCheck = false); 359 360 void computeDependenciesAux(); 361 }; 362 363 } // end anonymous namespace 364 365 /// The "main" function for implementing Swing Modulo Scheduling. 366 bool MachinePipeliner::runOnMachineFunction(MachineFunction &mf) { 367 if (skipFunction(mf.getFunction())) 368 return false; 369 370 if (!EnableSWP) 371 return false; 372 373 if (mf.getFunction().getAttributes().hasFnAttr(Attribute::OptimizeForSize) && 374 !EnableSWPOptSize.getPosition()) 375 return false; 376 377 if (!mf.getSubtarget().enableMachinePipeliner()) 378 return false; 379 380 // Cannot pipeline loops without instruction itineraries if we are using 381 // DFA for the pipeliner. 382 if (mf.getSubtarget().useDFAforSMS() && 383 (!mf.getSubtarget().getInstrItineraryData() || 384 mf.getSubtarget().getInstrItineraryData()->isEmpty())) 385 return false; 386 387 MF = &mf; 388 MLI = &getAnalysis<MachineLoopInfoWrapperPass>().getLI(); 389 MDT = &getAnalysis<MachineDominatorTreeWrapperPass>().getDomTree(); 390 ORE = &getAnalysis<MachineOptimizationRemarkEmitterPass>().getORE(); 391 TII = MF->getSubtarget().getInstrInfo(); 392 RegClassInfo.runOnMachineFunction(*MF); 393 394 for (const auto &L : *MLI) 395 scheduleLoop(*L); 396 397 return false; 398 } 399 400 /// Attempt to perform the SMS algorithm on the specified loop. This function is 401 /// the main entry point for the algorithm. The function identifies candidate 402 /// loops, calculates the minimum initiation interval, and attempts to schedule 403 /// the loop. 404 bool MachinePipeliner::scheduleLoop(MachineLoop &L) { 405 bool Changed = false; 406 for (const auto &InnerLoop : L) 407 Changed |= scheduleLoop(*InnerLoop); 408 409 #ifndef NDEBUG 410 // Stop trying after reaching the limit (if any). 411 int Limit = SwpLoopLimit; 412 if (Limit >= 0) { 413 if (NumTries >= SwpLoopLimit) 414 return Changed; 415 NumTries++; 416 } 417 #endif 418 419 setPragmaPipelineOptions(L); 420 if (!canPipelineLoop(L)) { 421 LLVM_DEBUG(dbgs() << "\n!!! Can not pipeline loop.\n"); 422 ORE->emit([&]() { 423 return MachineOptimizationRemarkMissed(DEBUG_TYPE, "canPipelineLoop", 424 L.getStartLoc(), L.getHeader()) 425 << "Failed to pipeline loop"; 426 }); 427 428 LI.LoopPipelinerInfo.reset(); 429 return Changed; 430 } 431 432 ++NumTrytoPipeline; 433 if (useSwingModuloScheduler()) 434 Changed = swingModuloScheduler(L); 435 436 if (useWindowScheduler(Changed)) 437 Changed = runWindowScheduler(L); 438 439 LI.LoopPipelinerInfo.reset(); 440 return Changed; 441 } 442 443 void MachinePipeliner::setPragmaPipelineOptions(MachineLoop &L) { 444 // Reset the pragma for the next loop in iteration. 445 disabledByPragma = false; 446 II_setByPragma = 0; 447 448 MachineBasicBlock *LBLK = L.getTopBlock(); 449 450 if (LBLK == nullptr) 451 return; 452 453 const BasicBlock *BBLK = LBLK->getBasicBlock(); 454 if (BBLK == nullptr) 455 return; 456 457 const Instruction *TI = BBLK->getTerminator(); 458 if (TI == nullptr) 459 return; 460 461 MDNode *LoopID = TI->getMetadata(LLVMContext::MD_loop); 462 if (LoopID == nullptr) 463 return; 464 465 assert(LoopID->getNumOperands() > 0 && "requires atleast one operand"); 466 assert(LoopID->getOperand(0) == LoopID && "invalid loop"); 467 468 for (const MDOperand &MDO : llvm::drop_begin(LoopID->operands())) { 469 MDNode *MD = dyn_cast<MDNode>(MDO); 470 471 if (MD == nullptr) 472 continue; 473 474 MDString *S = dyn_cast<MDString>(MD->getOperand(0)); 475 476 if (S == nullptr) 477 continue; 478 479 if (S->getString() == "llvm.loop.pipeline.initiationinterval") { 480 assert(MD->getNumOperands() == 2 && 481 "Pipeline initiation interval hint metadata should have two operands."); 482 II_setByPragma = 483 mdconst::extract<ConstantInt>(MD->getOperand(1))->getZExtValue(); 484 assert(II_setByPragma >= 1 && "Pipeline initiation interval must be positive."); 485 } else if (S->getString() == "llvm.loop.pipeline.disable") { 486 disabledByPragma = true; 487 } 488 } 489 } 490 491 /// Return true if the loop can be software pipelined. The algorithm is 492 /// restricted to loops with a single basic block. Make sure that the 493 /// branch in the loop can be analyzed. 494 bool MachinePipeliner::canPipelineLoop(MachineLoop &L) { 495 if (L.getNumBlocks() != 1) { 496 ORE->emit([&]() { 497 return MachineOptimizationRemarkAnalysis(DEBUG_TYPE, "canPipelineLoop", 498 L.getStartLoc(), L.getHeader()) 499 << "Not a single basic block: " 500 << ore::NV("NumBlocks", L.getNumBlocks()); 501 }); 502 return false; 503 } 504 505 if (disabledByPragma) { 506 ORE->emit([&]() { 507 return MachineOptimizationRemarkAnalysis(DEBUG_TYPE, "canPipelineLoop", 508 L.getStartLoc(), L.getHeader()) 509 << "Disabled by Pragma."; 510 }); 511 return false; 512 } 513 514 // Check if the branch can't be understood because we can't do pipelining 515 // if that's the case. 516 LI.TBB = nullptr; 517 LI.FBB = nullptr; 518 LI.BrCond.clear(); 519 if (TII->analyzeBranch(*L.getHeader(), LI.TBB, LI.FBB, LI.BrCond)) { 520 LLVM_DEBUG(dbgs() << "Unable to analyzeBranch, can NOT pipeline Loop\n"); 521 NumFailBranch++; 522 ORE->emit([&]() { 523 return MachineOptimizationRemarkAnalysis(DEBUG_TYPE, "canPipelineLoop", 524 L.getStartLoc(), L.getHeader()) 525 << "The branch can't be understood"; 526 }); 527 return false; 528 } 529 530 LI.LoopInductionVar = nullptr; 531 LI.LoopCompare = nullptr; 532 LI.LoopPipelinerInfo = TII->analyzeLoopForPipelining(L.getTopBlock()); 533 if (!LI.LoopPipelinerInfo) { 534 LLVM_DEBUG(dbgs() << "Unable to analyzeLoop, can NOT pipeline Loop\n"); 535 NumFailLoop++; 536 ORE->emit([&]() { 537 return MachineOptimizationRemarkAnalysis(DEBUG_TYPE, "canPipelineLoop", 538 L.getStartLoc(), L.getHeader()) 539 << "The loop structure is not supported"; 540 }); 541 return false; 542 } 543 544 if (!L.getLoopPreheader()) { 545 LLVM_DEBUG(dbgs() << "Preheader not found, can NOT pipeline Loop\n"); 546 NumFailPreheader++; 547 ORE->emit([&]() { 548 return MachineOptimizationRemarkAnalysis(DEBUG_TYPE, "canPipelineLoop", 549 L.getStartLoc(), L.getHeader()) 550 << "No loop preheader found"; 551 }); 552 return false; 553 } 554 555 unsigned NumStores = 0; 556 for (MachineInstr &MI : *L.getHeader()) 557 if (MI.mayStore()) 558 ++NumStores; 559 if (NumStores > SwpMaxNumStores) { 560 LLVM_DEBUG(dbgs() << "Too many stores\n"); 561 NumFailTooManyStores++; 562 ORE->emit([&]() { 563 return MachineOptimizationRemarkAnalysis(DEBUG_TYPE, "canPipelineLoop", 564 L.getStartLoc(), L.getHeader()) 565 << "Too many store instructions in the loop: " 566 << ore::NV("NumStores", NumStores) << " > " 567 << ore::NV("SwpMaxNumStores", SwpMaxNumStores) << "."; 568 }); 569 return false; 570 } 571 572 // Remove any subregisters from inputs to phi nodes. 573 preprocessPhiNodes(*L.getHeader()); 574 return true; 575 } 576 577 void MachinePipeliner::preprocessPhiNodes(MachineBasicBlock &B) { 578 MachineRegisterInfo &MRI = MF->getRegInfo(); 579 SlotIndexes &Slots = 580 *getAnalysis<LiveIntervalsWrapperPass>().getLIS().getSlotIndexes(); 581 582 for (MachineInstr &PI : B.phis()) { 583 MachineOperand &DefOp = PI.getOperand(0); 584 assert(DefOp.getSubReg() == 0); 585 auto *RC = MRI.getRegClass(DefOp.getReg()); 586 587 for (unsigned i = 1, n = PI.getNumOperands(); i != n; i += 2) { 588 MachineOperand &RegOp = PI.getOperand(i); 589 if (RegOp.getSubReg() == 0) 590 continue; 591 592 // If the operand uses a subregister, replace it with a new register 593 // without subregisters, and generate a copy to the new register. 594 Register NewReg = MRI.createVirtualRegister(RC); 595 MachineBasicBlock &PredB = *PI.getOperand(i+1).getMBB(); 596 MachineBasicBlock::iterator At = PredB.getFirstTerminator(); 597 const DebugLoc &DL = PredB.findDebugLoc(At); 598 auto Copy = BuildMI(PredB, At, DL, TII->get(TargetOpcode::COPY), NewReg) 599 .addReg(RegOp.getReg(), getRegState(RegOp), 600 RegOp.getSubReg()); 601 Slots.insertMachineInstrInMaps(*Copy); 602 RegOp.setReg(NewReg); 603 RegOp.setSubReg(0); 604 } 605 } 606 } 607 608 /// The SMS algorithm consists of the following main steps: 609 /// 1. Computation and analysis of the dependence graph. 610 /// 2. Ordering of the nodes (instructions). 611 /// 3. Attempt to Schedule the loop. 612 bool MachinePipeliner::swingModuloScheduler(MachineLoop &L) { 613 assert(L.getBlocks().size() == 1 && "SMS works on single blocks only."); 614 615 AliasAnalysis *AA = &getAnalysis<AAResultsWrapperPass>().getAAResults(); 616 SwingSchedulerDAG SMS( 617 *this, L, getAnalysis<LiveIntervalsWrapperPass>().getLIS(), RegClassInfo, 618 II_setByPragma, LI.LoopPipelinerInfo.get(), AA); 619 620 MachineBasicBlock *MBB = L.getHeader(); 621 // The kernel should not include any terminator instructions. These 622 // will be added back later. 623 SMS.startBlock(MBB); 624 625 // Compute the number of 'real' instructions in the basic block by 626 // ignoring terminators. 627 unsigned size = MBB->size(); 628 for (MachineBasicBlock::iterator I = MBB->getFirstTerminator(), 629 E = MBB->instr_end(); 630 I != E; ++I, --size) 631 ; 632 633 SMS.enterRegion(MBB, MBB->begin(), MBB->getFirstTerminator(), size); 634 SMS.schedule(); 635 SMS.exitRegion(); 636 637 SMS.finishBlock(); 638 return SMS.hasNewSchedule(); 639 } 640 641 void MachinePipeliner::getAnalysisUsage(AnalysisUsage &AU) const { 642 AU.addRequired<AAResultsWrapperPass>(); 643 AU.addPreserved<AAResultsWrapperPass>(); 644 AU.addRequired<MachineLoopInfoWrapperPass>(); 645 AU.addRequired<MachineDominatorTreeWrapperPass>(); 646 AU.addRequired<LiveIntervalsWrapperPass>(); 647 AU.addRequired<MachineOptimizationRemarkEmitterPass>(); 648 AU.addRequired<TargetPassConfig>(); 649 MachineFunctionPass::getAnalysisUsage(AU); 650 } 651 652 bool MachinePipeliner::runWindowScheduler(MachineLoop &L) { 653 MachineSchedContext Context; 654 Context.MF = MF; 655 Context.MLI = MLI; 656 Context.MDT = MDT; 657 Context.TM = &getAnalysis<TargetPassConfig>().getTM<TargetMachine>(); 658 Context.AA = &getAnalysis<AAResultsWrapperPass>().getAAResults(); 659 Context.LIS = &getAnalysis<LiveIntervalsWrapperPass>().getLIS(); 660 Context.RegClassInfo->runOnMachineFunction(*MF); 661 WindowScheduler WS(&Context, L); 662 return WS.run(); 663 } 664 665 bool MachinePipeliner::useSwingModuloScheduler() { 666 // SwingModuloScheduler does not work when WindowScheduler is forced. 667 return WindowSchedulingOption != WindowSchedulingFlag::WS_Force; 668 } 669 670 bool MachinePipeliner::useWindowScheduler(bool Changed) { 671 // WindowScheduler does not work for following cases: 672 // 1. when it is off. 673 // 2. when SwingModuloScheduler is successfully scheduled. 674 // 3. when pragma II is enabled. 675 if (II_setByPragma) { 676 LLVM_DEBUG(dbgs() << "Window scheduling is disabled when " 677 "llvm.loop.pipeline.initiationinterval is set.\n"); 678 return false; 679 } 680 681 return WindowSchedulingOption == WindowSchedulingFlag::WS_Force || 682 (WindowSchedulingOption == WindowSchedulingFlag::WS_On && !Changed); 683 } 684 685 void SwingSchedulerDAG::setMII(unsigned ResMII, unsigned RecMII) { 686 if (SwpForceII > 0) 687 MII = SwpForceII; 688 else if (II_setByPragma > 0) 689 MII = II_setByPragma; 690 else 691 MII = std::max(ResMII, RecMII); 692 } 693 694 void SwingSchedulerDAG::setMAX_II() { 695 if (SwpForceII > 0) 696 MAX_II = SwpForceII; 697 else if (II_setByPragma > 0) 698 MAX_II = II_setByPragma; 699 else 700 MAX_II = MII + SwpIISearchRange; 701 } 702 703 /// We override the schedule function in ScheduleDAGInstrs to implement the 704 /// scheduling part of the Swing Modulo Scheduling algorithm. 705 void SwingSchedulerDAG::schedule() { 706 buildSchedGraph(AA); 707 const LoopCarriedEdges LCE = addLoopCarriedDependences(); 708 updatePhiDependences(); 709 Topo.InitDAGTopologicalSorting(); 710 changeDependences(); 711 postProcessDAG(); 712 DDG = std::make_unique<SwingSchedulerDDG>(SUnits, &EntrySU, &ExitSU, LCE); 713 LLVM_DEBUG({ 714 dump(); 715 dbgs() << "===== Loop Carried Edges Begin =====\n"; 716 for (SUnit &SU : SUnits) 717 LCE.dump(&SU, TRI, &MRI); 718 dbgs() << "===== Loop Carried Edges End =====\n"; 719 }); 720 721 NodeSetType NodeSets; 722 findCircuits(NodeSets); 723 NodeSetType Circuits = NodeSets; 724 725 // Calculate the MII. 726 unsigned ResMII = calculateResMII(); 727 unsigned RecMII = calculateRecMII(NodeSets); 728 729 fuseRecs(NodeSets); 730 731 // This flag is used for testing and can cause correctness problems. 732 if (SwpIgnoreRecMII) 733 RecMII = 0; 734 735 setMII(ResMII, RecMII); 736 setMAX_II(); 737 738 LLVM_DEBUG(dbgs() << "MII = " << MII << " MAX_II = " << MAX_II 739 << " (rec=" << RecMII << ", res=" << ResMII << ")\n"); 740 741 // Can't schedule a loop without a valid MII. 742 if (MII == 0) { 743 LLVM_DEBUG(dbgs() << "Invalid Minimal Initiation Interval: 0\n"); 744 NumFailZeroMII++; 745 Pass.ORE->emit([&]() { 746 return MachineOptimizationRemarkAnalysis( 747 DEBUG_TYPE, "schedule", Loop.getStartLoc(), Loop.getHeader()) 748 << "Invalid Minimal Initiation Interval: 0"; 749 }); 750 return; 751 } 752 753 // Don't pipeline large loops. 754 if (SwpMaxMii != -1 && (int)MII > SwpMaxMii) { 755 LLVM_DEBUG(dbgs() << "MII > " << SwpMaxMii 756 << ", we don't pipeline large loops\n"); 757 NumFailLargeMaxMII++; 758 Pass.ORE->emit([&]() { 759 return MachineOptimizationRemarkAnalysis( 760 DEBUG_TYPE, "schedule", Loop.getStartLoc(), Loop.getHeader()) 761 << "Minimal Initiation Interval too large: " 762 << ore::NV("MII", (int)MII) << " > " 763 << ore::NV("SwpMaxMii", SwpMaxMii) << "." 764 << "Refer to -pipeliner-max-mii."; 765 }); 766 return; 767 } 768 769 computeNodeFunctions(NodeSets); 770 771 registerPressureFilter(NodeSets); 772 773 colocateNodeSets(NodeSets); 774 775 checkNodeSets(NodeSets); 776 777 LLVM_DEBUG({ 778 for (auto &I : NodeSets) { 779 dbgs() << " Rec NodeSet "; 780 I.dump(); 781 } 782 }); 783 784 llvm::stable_sort(NodeSets, std::greater<NodeSet>()); 785 786 groupRemainingNodes(NodeSets); 787 788 removeDuplicateNodes(NodeSets); 789 790 LLVM_DEBUG({ 791 for (auto &I : NodeSets) { 792 dbgs() << " NodeSet "; 793 I.dump(); 794 } 795 }); 796 797 computeNodeOrder(NodeSets); 798 799 // check for node order issues 800 checkValidNodeOrder(Circuits); 801 802 SMSchedule Schedule(Pass.MF, this); 803 Scheduled = schedulePipeline(Schedule); 804 805 if (!Scheduled){ 806 LLVM_DEBUG(dbgs() << "No schedule found, return\n"); 807 NumFailNoSchedule++; 808 Pass.ORE->emit([&]() { 809 return MachineOptimizationRemarkAnalysis( 810 DEBUG_TYPE, "schedule", Loop.getStartLoc(), Loop.getHeader()) 811 << "Unable to find schedule"; 812 }); 813 return; 814 } 815 816 unsigned numStages = Schedule.getMaxStageCount(); 817 // No need to generate pipeline if there are no overlapped iterations. 818 if (numStages == 0) { 819 LLVM_DEBUG(dbgs() << "No overlapped iterations, skip.\n"); 820 NumFailZeroStage++; 821 Pass.ORE->emit([&]() { 822 return MachineOptimizationRemarkAnalysis( 823 DEBUG_TYPE, "schedule", Loop.getStartLoc(), Loop.getHeader()) 824 << "No need to pipeline - no overlapped iterations in schedule."; 825 }); 826 return; 827 } 828 // Check that the maximum stage count is less than user-defined limit. 829 if (SwpMaxStages > -1 && (int)numStages > SwpMaxStages) { 830 LLVM_DEBUG(dbgs() << "numStages:" << numStages << ">" << SwpMaxStages 831 << " : too many stages, abort\n"); 832 NumFailLargeMaxStage++; 833 Pass.ORE->emit([&]() { 834 return MachineOptimizationRemarkAnalysis( 835 DEBUG_TYPE, "schedule", Loop.getStartLoc(), Loop.getHeader()) 836 << "Too many stages in schedule: " 837 << ore::NV("numStages", (int)numStages) << " > " 838 << ore::NV("SwpMaxStages", SwpMaxStages) 839 << ". Refer to -pipeliner-max-stages."; 840 }); 841 return; 842 } 843 844 Pass.ORE->emit([&]() { 845 return MachineOptimizationRemark(DEBUG_TYPE, "schedule", Loop.getStartLoc(), 846 Loop.getHeader()) 847 << "Pipelined succesfully!"; 848 }); 849 850 // Generate the schedule as a ModuloSchedule. 851 DenseMap<MachineInstr *, int> Cycles, Stages; 852 std::vector<MachineInstr *> OrderedInsts; 853 for (int Cycle = Schedule.getFirstCycle(); Cycle <= Schedule.getFinalCycle(); 854 ++Cycle) { 855 for (SUnit *SU : Schedule.getInstructions(Cycle)) { 856 OrderedInsts.push_back(SU->getInstr()); 857 Cycles[SU->getInstr()] = Cycle; 858 Stages[SU->getInstr()] = Schedule.stageScheduled(SU); 859 } 860 } 861 DenseMap<MachineInstr *, std::pair<Register, int64_t>> NewInstrChanges; 862 for (auto &KV : NewMIs) { 863 Cycles[KV.first] = Cycles[KV.second]; 864 Stages[KV.first] = Stages[KV.second]; 865 NewInstrChanges[KV.first] = InstrChanges[getSUnit(KV.first)]; 866 } 867 868 ModuloSchedule MS(MF, &Loop, std::move(OrderedInsts), std::move(Cycles), 869 std::move(Stages)); 870 if (EmitTestAnnotations) { 871 assert(NewInstrChanges.empty() && 872 "Cannot serialize a schedule with InstrChanges!"); 873 ModuloScheduleTestAnnotater MSTI(MF, MS); 874 MSTI.annotate(); 875 return; 876 } 877 // The experimental code generator can't work if there are InstChanges. 878 if (ExperimentalCodeGen && NewInstrChanges.empty()) { 879 PeelingModuloScheduleExpander MSE(MF, MS, &LIS); 880 MSE.expand(); 881 } else if (MVECodeGen && NewInstrChanges.empty() && 882 LoopPipelinerInfo->isMVEExpanderSupported() && 883 ModuloScheduleExpanderMVE::canApply(Loop)) { 884 ModuloScheduleExpanderMVE MSE(MF, MS, LIS); 885 MSE.expand(); 886 } else { 887 ModuloScheduleExpander MSE(MF, MS, LIS, std::move(NewInstrChanges)); 888 MSE.expand(); 889 MSE.cleanup(); 890 } 891 ++NumPipelined; 892 } 893 894 /// Clean up after the software pipeliner runs. 895 void SwingSchedulerDAG::finishBlock() { 896 for (auto &KV : NewMIs) 897 MF.deleteMachineInstr(KV.second); 898 NewMIs.clear(); 899 900 // Call the superclass. 901 ScheduleDAGInstrs::finishBlock(); 902 } 903 904 /// Return the register values for the operands of a Phi instruction. 905 /// This function assume the instruction is a Phi. 906 static void getPhiRegs(MachineInstr &Phi, MachineBasicBlock *Loop, 907 Register &InitVal, Register &LoopVal) { 908 assert(Phi.isPHI() && "Expecting a Phi."); 909 910 InitVal = Register(); 911 LoopVal = Register(); 912 for (unsigned i = 1, e = Phi.getNumOperands(); i != e; i += 2) 913 if (Phi.getOperand(i + 1).getMBB() != Loop) 914 InitVal = Phi.getOperand(i).getReg(); 915 else 916 LoopVal = Phi.getOperand(i).getReg(); 917 918 assert(InitVal && LoopVal && "Unexpected Phi structure."); 919 } 920 921 /// Return the Phi register value that comes the loop block. 922 static Register getLoopPhiReg(const MachineInstr &Phi, 923 const MachineBasicBlock *LoopBB) { 924 for (unsigned i = 1, e = Phi.getNumOperands(); i != e; i += 2) 925 if (Phi.getOperand(i + 1).getMBB() == LoopBB) 926 return Phi.getOperand(i).getReg(); 927 return Register(); 928 } 929 930 /// Return true if SUb can be reached from SUa following the chain edges. 931 static bool isSuccOrder(SUnit *SUa, SUnit *SUb) { 932 SmallPtrSet<SUnit *, 8> Visited; 933 SmallVector<SUnit *, 8> Worklist; 934 Worklist.push_back(SUa); 935 while (!Worklist.empty()) { 936 const SUnit *SU = Worklist.pop_back_val(); 937 for (const auto &SI : SU->Succs) { 938 SUnit *SuccSU = SI.getSUnit(); 939 if (SI.getKind() == SDep::Order) { 940 if (Visited.count(SuccSU)) 941 continue; 942 if (SuccSU == SUb) 943 return true; 944 Worklist.push_back(SuccSU); 945 Visited.insert(SuccSU); 946 } 947 } 948 } 949 return false; 950 } 951 952 SUnitWithMemInfo::SUnitWithMemInfo(SUnit *SU) : SU(SU) { 953 if (!getUnderlyingObjects()) 954 return; 955 for (const Value *Obj : UnderlyingObjs) 956 if (!isIdentifiedObject(Obj)) { 957 IsAllIdentified = false; 958 break; 959 } 960 } 961 962 bool SUnitWithMemInfo::isTriviallyDisjoint( 963 const SUnitWithMemInfo &Other) const { 964 // If all underlying objects are identified objects and there is no overlap 965 // between them, then these two instructions are disjoint. 966 if (!IsAllIdentified || !Other.IsAllIdentified) 967 return false; 968 for (const Value *Obj : UnderlyingObjs) 969 if (llvm::is_contained(Other.UnderlyingObjs, Obj)) 970 return false; 971 return true; 972 } 973 974 /// Collect the underlying objects for the memory references of an instruction. 975 /// This function calls the code in ValueTracking, but first checks that the 976 /// instruction has a memory operand. 977 /// Returns false if we cannot find the underlying objects. 978 bool SUnitWithMemInfo::getUnderlyingObjects() { 979 const MachineInstr *MI = SU->getInstr(); 980 if (!MI->hasOneMemOperand()) 981 return false; 982 MachineMemOperand *MM = *MI->memoperands_begin(); 983 if (!MM->getValue()) 984 return false; 985 MemOpValue = MM->getValue(); 986 MemOpOffset = MM->getOffset(); 987 llvm::getUnderlyingObjects(MemOpValue, UnderlyingObjs); 988 989 // TODO: A no alias scope may be valid only in a single iteration. In this 990 // case we need to peel off it like LoopAccessAnalysis does. 991 AATags = MM->getAAInfo(); 992 return true; 993 } 994 995 /// Returns true if there is a loop-carried order dependency from \p Src to \p 996 /// Dst. 997 static bool 998 hasLoopCarriedMemDep(const SUnitWithMemInfo &Src, const SUnitWithMemInfo &Dst, 999 BatchAAResults &BAA, const TargetInstrInfo *TII, 1000 const TargetRegisterInfo *TRI, 1001 const SwingSchedulerDAG *SSD, bool PerformCheapCheck) { 1002 if (Src.isTriviallyDisjoint(Dst)) 1003 return false; 1004 if (isSuccOrder(Src.SU, Dst.SU)) 1005 return false; 1006 1007 MachineInstr &SrcMI = *Src.SU->getInstr(); 1008 MachineInstr &DstMI = *Dst.SU->getInstr(); 1009 if (PerformCheapCheck) { 1010 // First, perform the cheaper check that compares the base register. 1011 // If they are the same and the load offset is less than the store 1012 // offset, then mark the dependence as loop carried potentially. 1013 // 1014 // TODO: This check will be removed. 1015 const MachineOperand *BaseOp1, *BaseOp2; 1016 int64_t Offset1, Offset2; 1017 bool Offset1IsScalable, Offset2IsScalable; 1018 if (TII->getMemOperandWithOffset(SrcMI, BaseOp1, Offset1, Offset1IsScalable, 1019 TRI) && 1020 TII->getMemOperandWithOffset(DstMI, BaseOp2, Offset2, Offset2IsScalable, 1021 TRI)) { 1022 if (BaseOp1->isIdenticalTo(*BaseOp2) && 1023 Offset1IsScalable == Offset2IsScalable && 1024 (int)Offset1 < (int)Offset2) { 1025 assert(TII->areMemAccessesTriviallyDisjoint(SrcMI, DstMI) && 1026 "What happened to the chain edge?"); 1027 return true; 1028 } 1029 } 1030 } 1031 1032 if (!SSD->mayOverlapInLaterIter(&SrcMI, &DstMI)) 1033 return false; 1034 1035 // Second, the more expensive check that uses alias analysis on the 1036 // base registers. If they alias, and the load offset is less than 1037 // the store offset, the mark the dependence as loop carried. 1038 if (Src.isUnknown() || Dst.isUnknown()) 1039 return true; 1040 if (Src.MemOpValue == Dst.MemOpValue && Src.MemOpOffset <= Dst.MemOpOffset) 1041 return true; 1042 1043 if (BAA.isNoAlias( 1044 MemoryLocation::getBeforeOrAfter(Src.MemOpValue, Src.AATags), 1045 MemoryLocation::getBeforeOrAfter(Dst.MemOpValue, Dst.AATags))) 1046 return false; 1047 1048 // AliasAnalysis sometimes gives up on following the underlying 1049 // object. In such a case, separate checks for underlying objects may 1050 // prove that there are no aliases between two accesses. 1051 for (const Value *SrcObj : Src.UnderlyingObjs) 1052 for (const Value *DstObj : Dst.UnderlyingObjs) 1053 if (!BAA.isNoAlias(MemoryLocation::getBeforeOrAfter(SrcObj, Src.AATags), 1054 MemoryLocation::getBeforeOrAfter(DstObj, Dst.AATags))) 1055 return true; 1056 1057 return false; 1058 } 1059 1060 void LoopCarriedOrderDepsTracker::LoadStoreChunk::append(SUnit *SU) { 1061 const MachineInstr *MI = SU->getInstr(); 1062 if (!MI->mayLoadOrStore()) 1063 return; 1064 (MI->mayStore() ? Stores : Loads).emplace_back(SU); 1065 } 1066 1067 LoopCarriedOrderDepsTracker::LoopCarriedOrderDepsTracker( 1068 SwingSchedulerDAG *SSD, BatchAAResults *BAA, const TargetInstrInfo *TII, 1069 const TargetRegisterInfo *TRI) 1070 : DAG(SSD), BAA(BAA), SUnits(DAG->SUnits), N(SUnits.size()), 1071 LoopCarried(N, BitVector(N)), TII(TII), TRI(TRI) {} 1072 1073 void LoopCarriedOrderDepsTracker::computeDependencies() { 1074 // Traverse all instructions and extract only what we are targetting. 1075 for (auto &SU : SUnits) { 1076 auto Tagged = getInstrTag(&SU); 1077 1078 // This instruction has no loop-carried order-dependencies. 1079 if (!Tagged) 1080 continue; 1081 TaggedSUnits.emplace_back(&SU, *Tagged); 1082 } 1083 1084 computeDependenciesAux(); 1085 } 1086 1087 std::optional<LoopCarriedOrderDepsTracker::InstrTag> 1088 LoopCarriedOrderDepsTracker::getInstrTag(SUnit *SU) const { 1089 MachineInstr *MI = SU->getInstr(); 1090 if (TII->isGlobalMemoryObject(MI)) 1091 return InstrTag::Barrier; 1092 1093 if (MI->mayStore() || 1094 (MI->mayLoad() && !MI->isDereferenceableInvariantLoad())) 1095 return InstrTag::LoadOrStore; 1096 1097 if (MI->mayRaiseFPException()) 1098 return InstrTag::FPExceptions; 1099 1100 return std::nullopt; 1101 } 1102 1103 void LoopCarriedOrderDepsTracker::addDependenciesBetweenSUs( 1104 const SUnitWithMemInfo &Src, const SUnitWithMemInfo &Dst, 1105 bool PerformCheapCheck) { 1106 // Avoid self-dependencies. 1107 if (Src.SU == Dst.SU) 1108 return; 1109 1110 if (hasLoopCarriedMemDep(Src, Dst, *BAA, TII, TRI, DAG, PerformCheapCheck)) 1111 LoopCarried[Src.SU->NodeNum].set(Dst.SU->NodeNum); 1112 } 1113 1114 void LoopCarriedOrderDepsTracker::addLoopCarriedDepenenciesForChunks( 1115 const LoadStoreChunk &From, const LoadStoreChunk &To) { 1116 // Add load-to-store dependencies (WAR). 1117 for (const SUnitWithMemInfo &Src : From.Loads) 1118 for (const SUnitWithMemInfo &Dst : To.Stores) 1119 // Perform a cheap check first if this is a forward dependency. 1120 addDependenciesBetweenSUs(Src, Dst, Src.SU->NodeNum < Dst.SU->NodeNum); 1121 1122 // Add store-to-load dependencies (RAW). 1123 for (const SUnitWithMemInfo &Src : From.Stores) 1124 for (const SUnitWithMemInfo &Dst : To.Loads) 1125 addDependenciesBetweenSUs(Src, Dst); 1126 1127 // Add store-to-store dependencies (WAW). 1128 for (const SUnitWithMemInfo &Src : From.Stores) 1129 for (const SUnitWithMemInfo &Dst : To.Stores) 1130 addDependenciesBetweenSUs(Src, Dst); 1131 } 1132 1133 void LoopCarriedOrderDepsTracker::computeDependenciesAux() { 1134 SmallVector<LoadStoreChunk, 2> Chunks(1); 1135 for (const auto &TSU : TaggedSUnits) { 1136 InstrTag Tag = TSU.getTag(); 1137 SUnit *SU = TSU.getPointer(); 1138 switch (Tag) { 1139 case InstrTag::Barrier: 1140 Chunks.emplace_back(); 1141 break; 1142 case InstrTag::LoadOrStore: 1143 Chunks.back().append(SU); 1144 break; 1145 case InstrTag::FPExceptions: 1146 // TODO: Handle this properly. 1147 break; 1148 } 1149 } 1150 1151 // Add dependencies between memory operations. If there are one or more 1152 // barrier events between two memory instructions, we don't add a 1153 // loop-carried dependence for them. 1154 for (const LoadStoreChunk &Chunk : Chunks) 1155 addLoopCarriedDepenenciesForChunks(Chunk, Chunk); 1156 1157 // TODO: If there are multiple barrier instructions, dependencies from the 1158 // last barrier instruction (or load/store below it) to the first barrier 1159 // instruction (or load/store above it). 1160 } 1161 1162 /// Add a chain edge between a load and store if the store can be an 1163 /// alias of the load on a subsequent iteration, i.e., a loop carried 1164 /// dependence. This code is very similar to the code in ScheduleDAGInstrs 1165 /// but that code doesn't create loop carried dependences. 1166 /// TODO: Also compute output-dependencies. 1167 LoopCarriedEdges SwingSchedulerDAG::addLoopCarriedDependences() { 1168 LoopCarriedEdges LCE; 1169 1170 // Add loop-carried order-dependencies 1171 LoopCarriedOrderDepsTracker LCODTracker(this, &BAA, TII, TRI); 1172 LCODTracker.computeDependencies(); 1173 for (unsigned I = 0; I != SUnits.size(); I++) 1174 for (const int Succ : LCODTracker.getLoopCarried(I).set_bits()) 1175 LCE.OrderDeps[&SUnits[I]].insert(&SUnits[Succ]); 1176 1177 LCE.modifySUnits(SUnits, TII); 1178 return LCE; 1179 } 1180 1181 /// Update the phi dependences to the DAG because ScheduleDAGInstrs no longer 1182 /// processes dependences for PHIs. This function adds true dependences 1183 /// from a PHI to a use, and a loop carried dependence from the use to the 1184 /// PHI. The loop carried dependence is represented as an anti dependence 1185 /// edge. This function also removes chain dependences between unrelated 1186 /// PHIs. 1187 void SwingSchedulerDAG::updatePhiDependences() { 1188 SmallVector<SDep, 4> RemoveDeps; 1189 const TargetSubtargetInfo &ST = MF.getSubtarget<TargetSubtargetInfo>(); 1190 1191 // Iterate over each DAG node. 1192 for (SUnit &I : SUnits) { 1193 RemoveDeps.clear(); 1194 // Set to true if the instruction has an operand defined by a Phi. 1195 Register HasPhiUse; 1196 Register HasPhiDef; 1197 MachineInstr *MI = I.getInstr(); 1198 // Iterate over each operand, and we process the definitions. 1199 for (const MachineOperand &MO : MI->operands()) { 1200 if (!MO.isReg()) 1201 continue; 1202 Register Reg = MO.getReg(); 1203 if (MO.isDef()) { 1204 // If the register is used by a Phi, then create an anti dependence. 1205 for (MachineRegisterInfo::use_instr_iterator 1206 UI = MRI.use_instr_begin(Reg), 1207 UE = MRI.use_instr_end(); 1208 UI != UE; ++UI) { 1209 MachineInstr *UseMI = &*UI; 1210 SUnit *SU = getSUnit(UseMI); 1211 if (SU != nullptr && UseMI->isPHI()) { 1212 if (!MI->isPHI()) { 1213 SDep Dep(SU, SDep::Anti, Reg); 1214 Dep.setLatency(1); 1215 I.addPred(Dep); 1216 } else { 1217 HasPhiDef = Reg; 1218 // Add a chain edge to a dependent Phi that isn't an existing 1219 // predecessor. 1220 if (SU->NodeNum < I.NodeNum && !I.isPred(SU)) 1221 I.addPred(SDep(SU, SDep::Barrier)); 1222 } 1223 } 1224 } 1225 } else if (MO.isUse()) { 1226 // If the register is defined by a Phi, then create a true dependence. 1227 MachineInstr *DefMI = MRI.getUniqueVRegDef(Reg); 1228 if (DefMI == nullptr) 1229 continue; 1230 SUnit *SU = getSUnit(DefMI); 1231 if (SU != nullptr && DefMI->isPHI()) { 1232 if (!MI->isPHI()) { 1233 SDep Dep(SU, SDep::Data, Reg); 1234 Dep.setLatency(0); 1235 ST.adjustSchedDependency(SU, 0, &I, MO.getOperandNo(), Dep, 1236 &SchedModel); 1237 I.addPred(Dep); 1238 } else { 1239 HasPhiUse = Reg; 1240 // Add a chain edge to a dependent Phi that isn't an existing 1241 // predecessor. 1242 if (SU->NodeNum < I.NodeNum && !I.isPred(SU)) 1243 I.addPred(SDep(SU, SDep::Barrier)); 1244 } 1245 } 1246 } 1247 } 1248 // Remove order dependences from an unrelated Phi. 1249 if (!SwpPruneDeps) 1250 continue; 1251 for (auto &PI : I.Preds) { 1252 MachineInstr *PMI = PI.getSUnit()->getInstr(); 1253 if (PMI->isPHI() && PI.getKind() == SDep::Order) { 1254 if (I.getInstr()->isPHI()) { 1255 if (PMI->getOperand(0).getReg() == HasPhiUse) 1256 continue; 1257 if (getLoopPhiReg(*PMI, PMI->getParent()) == HasPhiDef) 1258 continue; 1259 } 1260 RemoveDeps.push_back(PI); 1261 } 1262 } 1263 for (const SDep &D : RemoveDeps) 1264 I.removePred(D); 1265 } 1266 } 1267 1268 /// Iterate over each DAG node and see if we can change any dependences 1269 /// in order to reduce the recurrence MII. 1270 void SwingSchedulerDAG::changeDependences() { 1271 // See if an instruction can use a value from the previous iteration. 1272 // If so, we update the base and offset of the instruction and change 1273 // the dependences. 1274 for (SUnit &I : SUnits) { 1275 unsigned BasePos = 0, OffsetPos = 0; 1276 Register NewBase; 1277 int64_t NewOffset = 0; 1278 if (!canUseLastOffsetValue(I.getInstr(), BasePos, OffsetPos, NewBase, 1279 NewOffset)) 1280 continue; 1281 1282 // Get the MI and SUnit for the instruction that defines the original base. 1283 Register OrigBase = I.getInstr()->getOperand(BasePos).getReg(); 1284 MachineInstr *DefMI = MRI.getUniqueVRegDef(OrigBase); 1285 if (!DefMI) 1286 continue; 1287 SUnit *DefSU = getSUnit(DefMI); 1288 if (!DefSU) 1289 continue; 1290 // Get the MI and SUnit for the instruction that defins the new base. 1291 MachineInstr *LastMI = MRI.getUniqueVRegDef(NewBase); 1292 if (!LastMI) 1293 continue; 1294 SUnit *LastSU = getSUnit(LastMI); 1295 if (!LastSU) 1296 continue; 1297 1298 if (Topo.IsReachable(&I, LastSU)) 1299 continue; 1300 1301 // Remove the dependence. The value now depends on a prior iteration. 1302 SmallVector<SDep, 4> Deps; 1303 for (const SDep &P : I.Preds) 1304 if (P.getSUnit() == DefSU) 1305 Deps.push_back(P); 1306 for (const SDep &D : Deps) { 1307 Topo.RemovePred(&I, D.getSUnit()); 1308 I.removePred(D); 1309 } 1310 // Remove the chain dependence between the instructions. 1311 Deps.clear(); 1312 for (auto &P : LastSU->Preds) 1313 if (P.getSUnit() == &I && P.getKind() == SDep::Order) 1314 Deps.push_back(P); 1315 for (const SDep &D : Deps) { 1316 Topo.RemovePred(LastSU, D.getSUnit()); 1317 LastSU->removePred(D); 1318 } 1319 1320 // Add a dependence between the new instruction and the instruction 1321 // that defines the new base. 1322 SDep Dep(&I, SDep::Anti, NewBase); 1323 Topo.AddPred(LastSU, &I); 1324 LastSU->addPred(Dep); 1325 1326 // Remember the base and offset information so that we can update the 1327 // instruction during code generation. 1328 InstrChanges[&I] = std::make_pair(NewBase, NewOffset); 1329 } 1330 } 1331 1332 /// Create an instruction stream that represents a single iteration and stage of 1333 /// each instruction. This function differs from SMSchedule::finalizeSchedule in 1334 /// that this doesn't have any side-effect to SwingSchedulerDAG. That is, this 1335 /// function is an approximation of SMSchedule::finalizeSchedule with all 1336 /// non-const operations removed. 1337 static void computeScheduledInsts(const SwingSchedulerDAG *SSD, 1338 SMSchedule &Schedule, 1339 std::vector<MachineInstr *> &OrderedInsts, 1340 DenseMap<MachineInstr *, unsigned> &Stages) { 1341 DenseMap<int, std::deque<SUnit *>> Instrs; 1342 1343 // Move all instructions to the first stage from the later stages. 1344 for (int Cycle = Schedule.getFirstCycle(); Cycle <= Schedule.getFinalCycle(); 1345 ++Cycle) { 1346 for (int Stage = 0, LastStage = Schedule.getMaxStageCount(); 1347 Stage <= LastStage; ++Stage) { 1348 for (SUnit *SU : llvm::reverse(Schedule.getInstructions( 1349 Cycle + Stage * Schedule.getInitiationInterval()))) { 1350 Instrs[Cycle].push_front(SU); 1351 } 1352 } 1353 } 1354 1355 for (int Cycle = Schedule.getFirstCycle(); Cycle <= Schedule.getFinalCycle(); 1356 ++Cycle) { 1357 std::deque<SUnit *> &CycleInstrs = Instrs[Cycle]; 1358 CycleInstrs = Schedule.reorderInstructions(SSD, CycleInstrs); 1359 for (SUnit *SU : CycleInstrs) { 1360 MachineInstr *MI = SU->getInstr(); 1361 OrderedInsts.push_back(MI); 1362 Stages[MI] = Schedule.stageScheduled(SU); 1363 } 1364 } 1365 } 1366 1367 namespace { 1368 1369 // FuncUnitSorter - Comparison operator used to sort instructions by 1370 // the number of functional unit choices. 1371 struct FuncUnitSorter { 1372 const InstrItineraryData *InstrItins; 1373 const MCSubtargetInfo *STI; 1374 DenseMap<InstrStage::FuncUnits, unsigned> Resources; 1375 1376 FuncUnitSorter(const TargetSubtargetInfo &TSI) 1377 : InstrItins(TSI.getInstrItineraryData()), STI(&TSI) {} 1378 1379 // Compute the number of functional unit alternatives needed 1380 // at each stage, and take the minimum value. We prioritize the 1381 // instructions by the least number of choices first. 1382 unsigned minFuncUnits(const MachineInstr *Inst, 1383 InstrStage::FuncUnits &F) const { 1384 unsigned SchedClass = Inst->getDesc().getSchedClass(); 1385 unsigned min = UINT_MAX; 1386 if (InstrItins && !InstrItins->isEmpty()) { 1387 for (const InstrStage &IS : 1388 make_range(InstrItins->beginStage(SchedClass), 1389 InstrItins->endStage(SchedClass))) { 1390 InstrStage::FuncUnits funcUnits = IS.getUnits(); 1391 unsigned numAlternatives = llvm::popcount(funcUnits); 1392 if (numAlternatives < min) { 1393 min = numAlternatives; 1394 F = funcUnits; 1395 } 1396 } 1397 return min; 1398 } 1399 if (STI && STI->getSchedModel().hasInstrSchedModel()) { 1400 const MCSchedClassDesc *SCDesc = 1401 STI->getSchedModel().getSchedClassDesc(SchedClass); 1402 if (!SCDesc->isValid()) 1403 // No valid Schedule Class Desc for schedClass, should be 1404 // Pseudo/PostRAPseudo 1405 return min; 1406 1407 for (const MCWriteProcResEntry &PRE : 1408 make_range(STI->getWriteProcResBegin(SCDesc), 1409 STI->getWriteProcResEnd(SCDesc))) { 1410 if (!PRE.ReleaseAtCycle) 1411 continue; 1412 const MCProcResourceDesc *ProcResource = 1413 STI->getSchedModel().getProcResource(PRE.ProcResourceIdx); 1414 unsigned NumUnits = ProcResource->NumUnits; 1415 if (NumUnits < min) { 1416 min = NumUnits; 1417 F = PRE.ProcResourceIdx; 1418 } 1419 } 1420 return min; 1421 } 1422 llvm_unreachable("Should have non-empty InstrItins or hasInstrSchedModel!"); 1423 } 1424 1425 // Compute the critical resources needed by the instruction. This 1426 // function records the functional units needed by instructions that 1427 // must use only one functional unit. We use this as a tie breaker 1428 // for computing the resource MII. The instrutions that require 1429 // the same, highly used, functional unit have high priority. 1430 void calcCriticalResources(MachineInstr &MI) { 1431 unsigned SchedClass = MI.getDesc().getSchedClass(); 1432 if (InstrItins && !InstrItins->isEmpty()) { 1433 for (const InstrStage &IS : 1434 make_range(InstrItins->beginStage(SchedClass), 1435 InstrItins->endStage(SchedClass))) { 1436 InstrStage::FuncUnits FuncUnits = IS.getUnits(); 1437 if (llvm::popcount(FuncUnits) == 1) 1438 Resources[FuncUnits]++; 1439 } 1440 return; 1441 } 1442 if (STI && STI->getSchedModel().hasInstrSchedModel()) { 1443 const MCSchedClassDesc *SCDesc = 1444 STI->getSchedModel().getSchedClassDesc(SchedClass); 1445 if (!SCDesc->isValid()) 1446 // No valid Schedule Class Desc for schedClass, should be 1447 // Pseudo/PostRAPseudo 1448 return; 1449 1450 for (const MCWriteProcResEntry &PRE : 1451 make_range(STI->getWriteProcResBegin(SCDesc), 1452 STI->getWriteProcResEnd(SCDesc))) { 1453 if (!PRE.ReleaseAtCycle) 1454 continue; 1455 Resources[PRE.ProcResourceIdx]++; 1456 } 1457 return; 1458 } 1459 llvm_unreachable("Should have non-empty InstrItins or hasInstrSchedModel!"); 1460 } 1461 1462 /// Return true if IS1 has less priority than IS2. 1463 bool operator()(const MachineInstr *IS1, const MachineInstr *IS2) const { 1464 InstrStage::FuncUnits F1 = 0, F2 = 0; 1465 unsigned MFUs1 = minFuncUnits(IS1, F1); 1466 unsigned MFUs2 = minFuncUnits(IS2, F2); 1467 if (MFUs1 == MFUs2) 1468 return Resources.lookup(F1) < Resources.lookup(F2); 1469 return MFUs1 > MFUs2; 1470 } 1471 }; 1472 1473 /// Calculate the maximum register pressure of the scheduled instructions stream 1474 class HighRegisterPressureDetector { 1475 MachineBasicBlock *OrigMBB; 1476 const MachineRegisterInfo &MRI; 1477 const TargetRegisterInfo *TRI; 1478 1479 const unsigned PSetNum; 1480 1481 // Indexed by PSet ID 1482 // InitSetPressure takes into account the register pressure of live-in 1483 // registers. It's not depend on how the loop is scheduled, so it's enough to 1484 // calculate them once at the beginning. 1485 std::vector<unsigned> InitSetPressure; 1486 1487 // Indexed by PSet ID 1488 // Upper limit for each register pressure set 1489 std::vector<unsigned> PressureSetLimit; 1490 1491 DenseMap<MachineInstr *, RegisterOperands> ROMap; 1492 1493 using Instr2LastUsesTy = DenseMap<MachineInstr *, SmallDenseSet<Register, 4>>; 1494 1495 public: 1496 using OrderedInstsTy = std::vector<MachineInstr *>; 1497 using Instr2StageTy = DenseMap<MachineInstr *, unsigned>; 1498 1499 private: 1500 static void dumpRegisterPressures(const std::vector<unsigned> &Pressures) { 1501 if (Pressures.size() == 0) { 1502 dbgs() << "[]"; 1503 } else { 1504 char Prefix = '['; 1505 for (unsigned P : Pressures) { 1506 dbgs() << Prefix << P; 1507 Prefix = ' '; 1508 } 1509 dbgs() << ']'; 1510 } 1511 } 1512 1513 void dumpPSet(Register Reg) const { 1514 dbgs() << "Reg=" << printReg(Reg, TRI, 0, &MRI) << " PSet="; 1515 for (auto PSetIter = MRI.getPressureSets(Reg); PSetIter.isValid(); 1516 ++PSetIter) { 1517 dbgs() << *PSetIter << ' '; 1518 } 1519 dbgs() << '\n'; 1520 } 1521 1522 void increaseRegisterPressure(std::vector<unsigned> &Pressure, 1523 Register Reg) const { 1524 auto PSetIter = MRI.getPressureSets(Reg); 1525 unsigned Weight = PSetIter.getWeight(); 1526 for (; PSetIter.isValid(); ++PSetIter) 1527 Pressure[*PSetIter] += Weight; 1528 } 1529 1530 void decreaseRegisterPressure(std::vector<unsigned> &Pressure, 1531 Register Reg) const { 1532 auto PSetIter = MRI.getPressureSets(Reg); 1533 unsigned Weight = PSetIter.getWeight(); 1534 for (; PSetIter.isValid(); ++PSetIter) { 1535 auto &P = Pressure[*PSetIter]; 1536 assert(P >= Weight && 1537 "register pressure must be greater than or equal weight"); 1538 P -= Weight; 1539 } 1540 } 1541 1542 // Return true if Reg is reserved one, for example, stack pointer 1543 bool isReservedRegister(Register Reg) const { 1544 return Reg.isPhysical() && MRI.isReserved(Reg.asMCReg()); 1545 } 1546 1547 bool isDefinedInThisLoop(Register Reg) const { 1548 return Reg.isVirtual() && MRI.getVRegDef(Reg)->getParent() == OrigMBB; 1549 } 1550 1551 // Search for live-in variables. They are factored into the register pressure 1552 // from the begining. Live-in variables used by every iteration should be 1553 // considered as alive throughout the loop. For example, the variable `c` in 1554 // following code. \code 1555 // int c = ...; 1556 // for (int i = 0; i < n; i++) 1557 // a[i] += b[i] + c; 1558 // \endcode 1559 void computeLiveIn() { 1560 DenseSet<Register> Used; 1561 for (auto &MI : *OrigMBB) { 1562 if (MI.isDebugInstr()) 1563 continue; 1564 for (auto &Use : ROMap[&MI].Uses) { 1565 auto Reg = Use.RegUnit; 1566 // Ignore the variable that appears only on one side of phi instruction 1567 // because it's used only at the first iteration. 1568 if (MI.isPHI() && Reg != getLoopPhiReg(MI, OrigMBB)) 1569 continue; 1570 if (isReservedRegister(Reg)) 1571 continue; 1572 if (isDefinedInThisLoop(Reg)) 1573 continue; 1574 Used.insert(Reg); 1575 } 1576 } 1577 1578 for (auto LiveIn : Used) 1579 increaseRegisterPressure(InitSetPressure, LiveIn); 1580 } 1581 1582 // Calculate the upper limit of each pressure set 1583 void computePressureSetLimit(const RegisterClassInfo &RCI) { 1584 for (unsigned PSet = 0; PSet < PSetNum; PSet++) 1585 PressureSetLimit[PSet] = RCI.getRegPressureSetLimit(PSet); 1586 } 1587 1588 // There are two patterns of last-use. 1589 // - by an instruction of the current iteration 1590 // - by a phi instruction of the next iteration (loop carried value) 1591 // 1592 // Furthermore, following two groups of instructions are executed 1593 // simultaneously 1594 // - next iteration's phi instructions in i-th stage 1595 // - current iteration's instructions in i+1-th stage 1596 // 1597 // This function calculates the last-use of each register while taking into 1598 // account the above two patterns. 1599 Instr2LastUsesTy computeLastUses(const OrderedInstsTy &OrderedInsts, 1600 Instr2StageTy &Stages) const { 1601 // We treat virtual registers that are defined and used in this loop. 1602 // Following virtual register will be ignored 1603 // - live-in one 1604 // - defined but not used in the loop (potentially live-out) 1605 DenseSet<Register> TargetRegs; 1606 const auto UpdateTargetRegs = [this, &TargetRegs](Register Reg) { 1607 if (isDefinedInThisLoop(Reg)) 1608 TargetRegs.insert(Reg); 1609 }; 1610 for (MachineInstr *MI : OrderedInsts) { 1611 if (MI->isPHI()) { 1612 Register Reg = getLoopPhiReg(*MI, OrigMBB); 1613 UpdateTargetRegs(Reg); 1614 } else { 1615 for (auto &Use : ROMap.find(MI)->getSecond().Uses) 1616 UpdateTargetRegs(Use.RegUnit); 1617 } 1618 } 1619 1620 const auto InstrScore = [&Stages](MachineInstr *MI) { 1621 return Stages[MI] + MI->isPHI(); 1622 }; 1623 1624 DenseMap<Register, MachineInstr *> LastUseMI; 1625 for (MachineInstr *MI : llvm::reverse(OrderedInsts)) { 1626 for (auto &Use : ROMap.find(MI)->getSecond().Uses) { 1627 auto Reg = Use.RegUnit; 1628 if (!TargetRegs.contains(Reg)) 1629 continue; 1630 auto [Ite, Inserted] = LastUseMI.try_emplace(Reg, MI); 1631 if (!Inserted) { 1632 MachineInstr *Orig = Ite->second; 1633 MachineInstr *New = MI; 1634 if (InstrScore(Orig) < InstrScore(New)) 1635 Ite->second = New; 1636 } 1637 } 1638 } 1639 1640 Instr2LastUsesTy LastUses; 1641 for (auto &Entry : LastUseMI) 1642 LastUses[Entry.second].insert(Entry.first); 1643 return LastUses; 1644 } 1645 1646 // Compute the maximum register pressure of the kernel. We'll simulate #Stage 1647 // iterations and check the register pressure at the point where all stages 1648 // overlapping. 1649 // 1650 // An example of unrolled loop where #Stage is 4.. 1651 // Iter i+0 i+1 i+2 i+3 1652 // ------------------------ 1653 // Stage 0 1654 // Stage 1 0 1655 // Stage 2 1 0 1656 // Stage 3 2 1 0 <- All stages overlap 1657 // 1658 std::vector<unsigned> 1659 computeMaxSetPressure(const OrderedInstsTy &OrderedInsts, 1660 Instr2StageTy &Stages, 1661 const unsigned StageCount) const { 1662 using RegSetTy = SmallDenseSet<Register, 16>; 1663 1664 // Indexed by #Iter. To treat "local" variables of each stage separately, we 1665 // manage the liveness of the registers independently by iterations. 1666 SmallVector<RegSetTy> LiveRegSets(StageCount); 1667 1668 auto CurSetPressure = InitSetPressure; 1669 auto MaxSetPressure = InitSetPressure; 1670 auto LastUses = computeLastUses(OrderedInsts, Stages); 1671 1672 LLVM_DEBUG({ 1673 dbgs() << "Ordered instructions:\n"; 1674 for (MachineInstr *MI : OrderedInsts) { 1675 dbgs() << "Stage " << Stages[MI] << ": "; 1676 MI->dump(); 1677 } 1678 }); 1679 1680 const auto InsertReg = [this, &CurSetPressure](RegSetTy &RegSet, 1681 Register Reg) { 1682 if (!Reg.isValid() || isReservedRegister(Reg)) 1683 return; 1684 1685 bool Inserted = RegSet.insert(Reg).second; 1686 if (!Inserted) 1687 return; 1688 1689 LLVM_DEBUG(dbgs() << "insert " << printReg(Reg, TRI, 0, &MRI) << "\n"); 1690 increaseRegisterPressure(CurSetPressure, Reg); 1691 LLVM_DEBUG(dumpPSet(Reg)); 1692 }; 1693 1694 const auto EraseReg = [this, &CurSetPressure](RegSetTy &RegSet, 1695 Register Reg) { 1696 if (!Reg.isValid() || isReservedRegister(Reg)) 1697 return; 1698 1699 // live-in register 1700 if (!RegSet.contains(Reg)) 1701 return; 1702 1703 LLVM_DEBUG(dbgs() << "erase " << printReg(Reg, TRI, 0, &MRI) << "\n"); 1704 RegSet.erase(Reg); 1705 decreaseRegisterPressure(CurSetPressure, Reg); 1706 LLVM_DEBUG(dumpPSet(Reg)); 1707 }; 1708 1709 for (unsigned I = 0; I < StageCount; I++) { 1710 for (MachineInstr *MI : OrderedInsts) { 1711 const auto Stage = Stages[MI]; 1712 if (I < Stage) 1713 continue; 1714 1715 const unsigned Iter = I - Stage; 1716 1717 for (auto &Def : ROMap.find(MI)->getSecond().Defs) 1718 InsertReg(LiveRegSets[Iter], Def.RegUnit); 1719 1720 for (auto LastUse : LastUses[MI]) { 1721 if (MI->isPHI()) { 1722 if (Iter != 0) 1723 EraseReg(LiveRegSets[Iter - 1], LastUse); 1724 } else { 1725 EraseReg(LiveRegSets[Iter], LastUse); 1726 } 1727 } 1728 1729 for (unsigned PSet = 0; PSet < PSetNum; PSet++) 1730 MaxSetPressure[PSet] = 1731 std::max(MaxSetPressure[PSet], CurSetPressure[PSet]); 1732 1733 LLVM_DEBUG({ 1734 dbgs() << "CurSetPressure="; 1735 dumpRegisterPressures(CurSetPressure); 1736 dbgs() << " iter=" << Iter << " stage=" << Stage << ":"; 1737 MI->dump(); 1738 }); 1739 } 1740 } 1741 1742 return MaxSetPressure; 1743 } 1744 1745 public: 1746 HighRegisterPressureDetector(MachineBasicBlock *OrigMBB, 1747 const MachineFunction &MF) 1748 : OrigMBB(OrigMBB), MRI(MF.getRegInfo()), 1749 TRI(MF.getSubtarget().getRegisterInfo()), 1750 PSetNum(TRI->getNumRegPressureSets()), InitSetPressure(PSetNum, 0), 1751 PressureSetLimit(PSetNum, 0) {} 1752 1753 // Used to calculate register pressure, which is independent of loop 1754 // scheduling. 1755 void init(const RegisterClassInfo &RCI) { 1756 for (MachineInstr &MI : *OrigMBB) { 1757 if (MI.isDebugInstr()) 1758 continue; 1759 ROMap[&MI].collect(MI, *TRI, MRI, false, true); 1760 } 1761 1762 computeLiveIn(); 1763 computePressureSetLimit(RCI); 1764 } 1765 1766 // Calculate the maximum register pressures of the loop and check if they 1767 // exceed the limit 1768 bool detect(const SwingSchedulerDAG *SSD, SMSchedule &Schedule, 1769 const unsigned MaxStage) const { 1770 assert(0 <= RegPressureMargin && RegPressureMargin <= 100 && 1771 "the percentage of the margin must be between 0 to 100"); 1772 1773 OrderedInstsTy OrderedInsts; 1774 Instr2StageTy Stages; 1775 computeScheduledInsts(SSD, Schedule, OrderedInsts, Stages); 1776 const auto MaxSetPressure = 1777 computeMaxSetPressure(OrderedInsts, Stages, MaxStage + 1); 1778 1779 LLVM_DEBUG({ 1780 dbgs() << "Dump MaxSetPressure:\n"; 1781 for (unsigned I = 0; I < MaxSetPressure.size(); I++) { 1782 dbgs() << format("MaxSetPressure[%d]=%d\n", I, MaxSetPressure[I]); 1783 } 1784 dbgs() << '\n'; 1785 }); 1786 1787 for (unsigned PSet = 0; PSet < PSetNum; PSet++) { 1788 unsigned Limit = PressureSetLimit[PSet]; 1789 unsigned Margin = Limit * RegPressureMargin / 100; 1790 LLVM_DEBUG(dbgs() << "PSet=" << PSet << " Limit=" << Limit 1791 << " Margin=" << Margin << "\n"); 1792 if (Limit < MaxSetPressure[PSet] + Margin) { 1793 LLVM_DEBUG( 1794 dbgs() 1795 << "Rejected the schedule because of too high register pressure\n"); 1796 return true; 1797 } 1798 } 1799 return false; 1800 } 1801 }; 1802 1803 } // end anonymous namespace 1804 1805 /// Calculate the resource constrained minimum initiation interval for the 1806 /// specified loop. We use the DFA to model the resources needed for 1807 /// each instruction, and we ignore dependences. A different DFA is created 1808 /// for each cycle that is required. When adding a new instruction, we attempt 1809 /// to add it to each existing DFA, until a legal space is found. If the 1810 /// instruction cannot be reserved in an existing DFA, we create a new one. 1811 unsigned SwingSchedulerDAG::calculateResMII() { 1812 LLVM_DEBUG(dbgs() << "calculateResMII:\n"); 1813 ResourceManager RM(&MF.getSubtarget(), this); 1814 return RM.calculateResMII(); 1815 } 1816 1817 /// Calculate the recurrence-constrainted minimum initiation interval. 1818 /// Iterate over each circuit. Compute the delay(c) and distance(c) 1819 /// for each circuit. The II needs to satisfy the inequality 1820 /// delay(c) - II*distance(c) <= 0. For each circuit, choose the smallest 1821 /// II that satisfies the inequality, and the RecMII is the maximum 1822 /// of those values. 1823 unsigned SwingSchedulerDAG::calculateRecMII(NodeSetType &NodeSets) { 1824 unsigned RecMII = 0; 1825 1826 for (NodeSet &Nodes : NodeSets) { 1827 if (Nodes.empty()) 1828 continue; 1829 1830 unsigned Delay = Nodes.getLatency(); 1831 unsigned Distance = 1; 1832 1833 // ii = ceil(delay / distance) 1834 unsigned CurMII = (Delay + Distance - 1) / Distance; 1835 Nodes.setRecMII(CurMII); 1836 if (CurMII > RecMII) 1837 RecMII = CurMII; 1838 } 1839 1840 return RecMII; 1841 } 1842 1843 /// Create the adjacency structure of the nodes in the graph. 1844 void SwingSchedulerDAG::Circuits::createAdjacencyStructure( 1845 SwingSchedulerDAG *DAG) { 1846 BitVector Added(SUnits.size()); 1847 DenseMap<int, int> OutputDeps; 1848 for (int i = 0, e = SUnits.size(); i != e; ++i) { 1849 Added.reset(); 1850 // Add any successor to the adjacency matrix and exclude duplicates. 1851 for (auto &OE : DAG->DDG->getOutEdges(&SUnits[i])) { 1852 // Only create a back-edge on the first and last nodes of a dependence 1853 // chain. This records any chains and adds them later. 1854 if (OE.isOutputDep()) { 1855 int N = OE.getDst()->NodeNum; 1856 int BackEdge = i; 1857 auto Dep = OutputDeps.find(BackEdge); 1858 if (Dep != OutputDeps.end()) { 1859 BackEdge = Dep->second; 1860 OutputDeps.erase(Dep); 1861 } 1862 OutputDeps[N] = BackEdge; 1863 } 1864 // Do not process a boundary node, an artificial node. 1865 if (OE.getDst()->isBoundaryNode() || OE.isArtificial()) 1866 continue; 1867 1868 // This code is retained o preserve previous behavior and prevent 1869 // regression. This condition means that anti-dependnecies within an 1870 // iteration are ignored when searching circuits. Therefore it's natural 1871 // to consider this dependence as well. 1872 // FIXME: Remove this code if it doesn't have significant impact on 1873 // performance. 1874 if (OE.isAntiDep()) 1875 continue; 1876 1877 int N = OE.getDst()->NodeNum; 1878 if (!Added.test(N)) { 1879 AdjK[i].push_back(N); 1880 Added.set(N); 1881 } 1882 } 1883 // A chain edge between a store and a load is treated as a back-edge in the 1884 // adjacency matrix. 1885 for (auto &IE : DAG->DDG->getInEdges(&SUnits[i])) { 1886 SUnit *Src = IE.getSrc(); 1887 SUnit *Dst = IE.getDst(); 1888 if (!Dst->getInstr()->mayStore() || !DAG->isLoopCarriedDep(IE)) 1889 continue; 1890 if (IE.isOrderDep() && Src->getInstr()->mayLoad()) { 1891 int N = Src->NodeNum; 1892 if (!Added.test(N)) { 1893 AdjK[i].push_back(N); 1894 Added.set(N); 1895 } 1896 } 1897 } 1898 } 1899 // Add back-edges in the adjacency matrix for the output dependences. 1900 for (auto &OD : OutputDeps) 1901 if (!Added.test(OD.second)) { 1902 AdjK[OD.first].push_back(OD.second); 1903 Added.set(OD.second); 1904 } 1905 } 1906 1907 /// Identify an elementary circuit in the dependence graph starting at the 1908 /// specified node. 1909 bool SwingSchedulerDAG::Circuits::circuit(int V, int S, NodeSetType &NodeSets, 1910 const SwingSchedulerDAG *DAG, 1911 bool HasBackedge) { 1912 SUnit *SV = &SUnits[V]; 1913 bool F = false; 1914 Stack.insert(SV); 1915 Blocked.set(V); 1916 1917 for (auto W : AdjK[V]) { 1918 if (NumPaths > MaxPaths) 1919 break; 1920 if (W < S) 1921 continue; 1922 if (W == S) { 1923 if (!HasBackedge) 1924 NodeSets.push_back(NodeSet(Stack.begin(), Stack.end(), DAG)); 1925 F = true; 1926 ++NumPaths; 1927 break; 1928 } 1929 if (!Blocked.test(W)) { 1930 if (circuit(W, S, NodeSets, DAG, 1931 Node2Idx->at(W) < Node2Idx->at(V) ? true : HasBackedge)) 1932 F = true; 1933 } 1934 } 1935 1936 if (F) 1937 unblock(V); 1938 else { 1939 for (auto W : AdjK[V]) { 1940 if (W < S) 1941 continue; 1942 B[W].insert(SV); 1943 } 1944 } 1945 Stack.pop_back(); 1946 return F; 1947 } 1948 1949 /// Unblock a node in the circuit finding algorithm. 1950 void SwingSchedulerDAG::Circuits::unblock(int U) { 1951 Blocked.reset(U); 1952 SmallPtrSet<SUnit *, 4> &BU = B[U]; 1953 while (!BU.empty()) { 1954 SmallPtrSet<SUnit *, 4>::iterator SI = BU.begin(); 1955 assert(SI != BU.end() && "Invalid B set."); 1956 SUnit *W = *SI; 1957 BU.erase(W); 1958 if (Blocked.test(W->NodeNum)) 1959 unblock(W->NodeNum); 1960 } 1961 } 1962 1963 /// Identify all the elementary circuits in the dependence graph using 1964 /// Johnson's circuit algorithm. 1965 void SwingSchedulerDAG::findCircuits(NodeSetType &NodeSets) { 1966 Circuits Cir(SUnits, Topo); 1967 // Create the adjacency structure. 1968 Cir.createAdjacencyStructure(this); 1969 for (int I = 0, E = SUnits.size(); I != E; ++I) { 1970 Cir.reset(); 1971 Cir.circuit(I, I, NodeSets, this); 1972 } 1973 } 1974 1975 // Create artificial dependencies between the source of COPY/REG_SEQUENCE that 1976 // is loop-carried to the USE in next iteration. This will help pipeliner avoid 1977 // additional copies that are needed across iterations. An artificial dependence 1978 // edge is added from USE to SOURCE of COPY/REG_SEQUENCE. 1979 1980 // PHI-------Anti-Dep-----> COPY/REG_SEQUENCE (loop-carried) 1981 // SRCOfCopY------True-Dep---> COPY/REG_SEQUENCE 1982 // PHI-------True-Dep------> USEOfPhi 1983 1984 // The mutation creates 1985 // USEOfPHI -------Artificial-Dep---> SRCOfCopy 1986 1987 // This overall will ensure, the USEOfPHI is scheduled before SRCOfCopy 1988 // (since USE is a predecessor), implies, the COPY/ REG_SEQUENCE is scheduled 1989 // late to avoid additional copies across iterations. The possible scheduling 1990 // order would be 1991 // USEOfPHI --- SRCOfCopy--- COPY/REG_SEQUENCE. 1992 1993 void SwingSchedulerDAG::CopyToPhiMutation::apply(ScheduleDAGInstrs *DAG) { 1994 for (SUnit &SU : DAG->SUnits) { 1995 // Find the COPY/REG_SEQUENCE instruction. 1996 if (!SU.getInstr()->isCopy() && !SU.getInstr()->isRegSequence()) 1997 continue; 1998 1999 // Record the loop carried PHIs. 2000 SmallVector<SUnit *, 4> PHISUs; 2001 // Record the SrcSUs that feed the COPY/REG_SEQUENCE instructions. 2002 SmallVector<SUnit *, 4> SrcSUs; 2003 2004 for (auto &Dep : SU.Preds) { 2005 SUnit *TmpSU = Dep.getSUnit(); 2006 MachineInstr *TmpMI = TmpSU->getInstr(); 2007 SDep::Kind DepKind = Dep.getKind(); 2008 // Save the loop carried PHI. 2009 if (DepKind == SDep::Anti && TmpMI->isPHI()) 2010 PHISUs.push_back(TmpSU); 2011 // Save the source of COPY/REG_SEQUENCE. 2012 // If the source has no pre-decessors, we will end up creating cycles. 2013 else if (DepKind == SDep::Data && !TmpMI->isPHI() && TmpSU->NumPreds > 0) 2014 SrcSUs.push_back(TmpSU); 2015 } 2016 2017 if (PHISUs.size() == 0 || SrcSUs.size() == 0) 2018 continue; 2019 2020 // Find the USEs of PHI. If the use is a PHI or REG_SEQUENCE, push back this 2021 // SUnit to the container. 2022 SmallVector<SUnit *, 8> UseSUs; 2023 // Do not use iterator based loop here as we are updating the container. 2024 for (size_t Index = 0; Index < PHISUs.size(); ++Index) { 2025 for (auto &Dep : PHISUs[Index]->Succs) { 2026 if (Dep.getKind() != SDep::Data) 2027 continue; 2028 2029 SUnit *TmpSU = Dep.getSUnit(); 2030 MachineInstr *TmpMI = TmpSU->getInstr(); 2031 if (TmpMI->isPHI() || TmpMI->isRegSequence()) { 2032 PHISUs.push_back(TmpSU); 2033 continue; 2034 } 2035 UseSUs.push_back(TmpSU); 2036 } 2037 } 2038 2039 if (UseSUs.size() == 0) 2040 continue; 2041 2042 SwingSchedulerDAG *SDAG = cast<SwingSchedulerDAG>(DAG); 2043 // Add the artificial dependencies if it does not form a cycle. 2044 for (auto *I : UseSUs) { 2045 for (auto *Src : SrcSUs) { 2046 if (!SDAG->Topo.IsReachable(I, Src) && Src != I) { 2047 Src->addPred(SDep(I, SDep::Artificial)); 2048 SDAG->Topo.AddPred(Src, I); 2049 } 2050 } 2051 } 2052 } 2053 } 2054 2055 /// Compute several functions need to order the nodes for scheduling. 2056 /// ASAP - Earliest time to schedule a node. 2057 /// ALAP - Latest time to schedule a node. 2058 /// MOV - Mobility function, difference between ALAP and ASAP. 2059 /// D - Depth of each node. 2060 /// H - Height of each node. 2061 void SwingSchedulerDAG::computeNodeFunctions(NodeSetType &NodeSets) { 2062 ScheduleInfo.resize(SUnits.size()); 2063 2064 LLVM_DEBUG({ 2065 for (int I : Topo) { 2066 const SUnit &SU = SUnits[I]; 2067 dumpNode(SU); 2068 } 2069 }); 2070 2071 int maxASAP = 0; 2072 // Compute ASAP and ZeroLatencyDepth. 2073 for (int I : Topo) { 2074 int asap = 0; 2075 int zeroLatencyDepth = 0; 2076 SUnit *SU = &SUnits[I]; 2077 for (const auto &IE : DDG->getInEdges(SU)) { 2078 SUnit *Pred = IE.getSrc(); 2079 if (IE.getLatency() == 0) 2080 zeroLatencyDepth = 2081 std::max(zeroLatencyDepth, getZeroLatencyDepth(Pred) + 1); 2082 if (IE.ignoreDependence(true)) 2083 continue; 2084 asap = std::max(asap, (int)(getASAP(Pred) + IE.getLatency() - 2085 IE.getDistance() * MII)); 2086 } 2087 maxASAP = std::max(maxASAP, asap); 2088 ScheduleInfo[I].ASAP = asap; 2089 ScheduleInfo[I].ZeroLatencyDepth = zeroLatencyDepth; 2090 } 2091 2092 // Compute ALAP, ZeroLatencyHeight, and MOV. 2093 for (int I : llvm::reverse(Topo)) { 2094 int alap = maxASAP; 2095 int zeroLatencyHeight = 0; 2096 SUnit *SU = &SUnits[I]; 2097 for (const auto &OE : DDG->getOutEdges(SU)) { 2098 SUnit *Succ = OE.getDst(); 2099 if (Succ->isBoundaryNode()) 2100 continue; 2101 if (OE.getLatency() == 0) 2102 zeroLatencyHeight = 2103 std::max(zeroLatencyHeight, getZeroLatencyHeight(Succ) + 1); 2104 if (OE.ignoreDependence(true)) 2105 continue; 2106 alap = std::min(alap, (int)(getALAP(Succ) - OE.getLatency() + 2107 OE.getDistance() * MII)); 2108 } 2109 2110 ScheduleInfo[I].ALAP = alap; 2111 ScheduleInfo[I].ZeroLatencyHeight = zeroLatencyHeight; 2112 } 2113 2114 // After computing the node functions, compute the summary for each node set. 2115 for (NodeSet &I : NodeSets) 2116 I.computeNodeSetInfo(this); 2117 2118 LLVM_DEBUG({ 2119 for (unsigned i = 0; i < SUnits.size(); i++) { 2120 dbgs() << "\tNode " << i << ":\n"; 2121 dbgs() << "\t ASAP = " << getASAP(&SUnits[i]) << "\n"; 2122 dbgs() << "\t ALAP = " << getALAP(&SUnits[i]) << "\n"; 2123 dbgs() << "\t MOV = " << getMOV(&SUnits[i]) << "\n"; 2124 dbgs() << "\t D = " << getDepth(&SUnits[i]) << "\n"; 2125 dbgs() << "\t H = " << getHeight(&SUnits[i]) << "\n"; 2126 dbgs() << "\t ZLD = " << getZeroLatencyDepth(&SUnits[i]) << "\n"; 2127 dbgs() << "\t ZLH = " << getZeroLatencyHeight(&SUnits[i]) << "\n"; 2128 } 2129 }); 2130 } 2131 2132 /// Compute the Pred_L(O) set, as defined in the paper. The set is defined 2133 /// as the predecessors of the elements of NodeOrder that are not also in 2134 /// NodeOrder. 2135 static bool pred_L(SetVector<SUnit *> &NodeOrder, 2136 SmallSetVector<SUnit *, 8> &Preds, SwingSchedulerDDG *DDG, 2137 const NodeSet *S = nullptr) { 2138 Preds.clear(); 2139 2140 for (SUnit *SU : NodeOrder) { 2141 for (const auto &IE : DDG->getInEdges(SU)) { 2142 SUnit *PredSU = IE.getSrc(); 2143 if (S && S->count(PredSU) == 0) 2144 continue; 2145 if (IE.ignoreDependence(true)) 2146 continue; 2147 if (NodeOrder.count(PredSU) == 0) 2148 Preds.insert(PredSU); 2149 } 2150 2151 // FIXME: The following loop-carried dependencies may also need to be 2152 // considered. 2153 // - Physical register dependencies (true-dependence and WAW). 2154 // - Memory dependencies. 2155 for (const auto &OE : DDG->getOutEdges(SU)) { 2156 SUnit *SuccSU = OE.getDst(); 2157 if (!OE.isAntiDep()) 2158 continue; 2159 if (S && S->count(SuccSU) == 0) 2160 continue; 2161 if (NodeOrder.count(SuccSU) == 0) 2162 Preds.insert(SuccSU); 2163 } 2164 } 2165 return !Preds.empty(); 2166 } 2167 2168 /// Compute the Succ_L(O) set, as defined in the paper. The set is defined 2169 /// as the successors of the elements of NodeOrder that are not also in 2170 /// NodeOrder. 2171 static bool succ_L(SetVector<SUnit *> &NodeOrder, 2172 SmallSetVector<SUnit *, 8> &Succs, SwingSchedulerDDG *DDG, 2173 const NodeSet *S = nullptr) { 2174 Succs.clear(); 2175 2176 for (SUnit *SU : NodeOrder) { 2177 for (const auto &OE : DDG->getOutEdges(SU)) { 2178 SUnit *SuccSU = OE.getDst(); 2179 if (S && S->count(SuccSU) == 0) 2180 continue; 2181 if (OE.ignoreDependence(false)) 2182 continue; 2183 if (NodeOrder.count(SuccSU) == 0) 2184 Succs.insert(SuccSU); 2185 } 2186 2187 // FIXME: The following loop-carried dependencies may also need to be 2188 // considered. 2189 // - Physical register dependnecies (true-dependnece and WAW). 2190 // - Memory dependencies. 2191 for (const auto &IE : DDG->getInEdges(SU)) { 2192 SUnit *PredSU = IE.getSrc(); 2193 if (!IE.isAntiDep()) 2194 continue; 2195 if (S && S->count(PredSU) == 0) 2196 continue; 2197 if (NodeOrder.count(PredSU) == 0) 2198 Succs.insert(PredSU); 2199 } 2200 } 2201 return !Succs.empty(); 2202 } 2203 2204 /// Return true if there is a path from the specified node to any of the nodes 2205 /// in DestNodes. Keep track and return the nodes in any path. 2206 static bool computePath(SUnit *Cur, SetVector<SUnit *> &Path, 2207 SetVector<SUnit *> &DestNodes, 2208 SetVector<SUnit *> &Exclude, 2209 SmallPtrSet<SUnit *, 8> &Visited, 2210 SwingSchedulerDDG *DDG) { 2211 if (Cur->isBoundaryNode()) 2212 return false; 2213 if (Exclude.contains(Cur)) 2214 return false; 2215 if (DestNodes.contains(Cur)) 2216 return true; 2217 if (!Visited.insert(Cur).second) 2218 return Path.contains(Cur); 2219 bool FoundPath = false; 2220 for (const auto &OE : DDG->getOutEdges(Cur)) 2221 if (!OE.ignoreDependence(false)) 2222 FoundPath |= 2223 computePath(OE.getDst(), Path, DestNodes, Exclude, Visited, DDG); 2224 for (const auto &IE : DDG->getInEdges(Cur)) 2225 if (IE.isAntiDep() && IE.getDistance() == 0) 2226 FoundPath |= 2227 computePath(IE.getSrc(), Path, DestNodes, Exclude, Visited, DDG); 2228 if (FoundPath) 2229 Path.insert(Cur); 2230 return FoundPath; 2231 } 2232 2233 /// Compute the live-out registers for the instructions in a node-set. 2234 /// The live-out registers are those that are defined in the node-set, 2235 /// but not used. Except for use operands of Phis. 2236 static void computeLiveOuts(MachineFunction &MF, RegPressureTracker &RPTracker, 2237 NodeSet &NS) { 2238 const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo(); 2239 MachineRegisterInfo &MRI = MF.getRegInfo(); 2240 SmallVector<VRegMaskOrUnit, 8> LiveOutRegs; 2241 SmallSet<Register, 4> Uses; 2242 for (SUnit *SU : NS) { 2243 const MachineInstr *MI = SU->getInstr(); 2244 if (MI->isPHI()) 2245 continue; 2246 for (const MachineOperand &MO : MI->all_uses()) { 2247 Register Reg = MO.getReg(); 2248 if (Reg.isVirtual()) 2249 Uses.insert(Reg); 2250 else if (MRI.isAllocatable(Reg)) 2251 Uses.insert_range(TRI->regunits(Reg.asMCReg())); 2252 } 2253 } 2254 for (SUnit *SU : NS) 2255 for (const MachineOperand &MO : SU->getInstr()->all_defs()) 2256 if (!MO.isDead()) { 2257 Register Reg = MO.getReg(); 2258 if (Reg.isVirtual()) { 2259 if (!Uses.count(Reg)) 2260 LiveOutRegs.emplace_back(Reg, LaneBitmask::getNone()); 2261 } else if (MRI.isAllocatable(Reg)) { 2262 for (MCRegUnit Unit : TRI->regunits(Reg.asMCReg())) 2263 if (!Uses.count(Unit)) 2264 LiveOutRegs.emplace_back(Unit, LaneBitmask::getNone()); 2265 } 2266 } 2267 RPTracker.addLiveRegs(LiveOutRegs); 2268 } 2269 2270 /// A heuristic to filter nodes in recurrent node-sets if the register 2271 /// pressure of a set is too high. 2272 void SwingSchedulerDAG::registerPressureFilter(NodeSetType &NodeSets) { 2273 for (auto &NS : NodeSets) { 2274 // Skip small node-sets since they won't cause register pressure problems. 2275 if (NS.size() <= 2) 2276 continue; 2277 IntervalPressure RecRegPressure; 2278 RegPressureTracker RecRPTracker(RecRegPressure); 2279 RecRPTracker.init(&MF, &RegClassInfo, &LIS, BB, BB->end(), false, true); 2280 computeLiveOuts(MF, RecRPTracker, NS); 2281 RecRPTracker.closeBottom(); 2282 2283 std::vector<SUnit *> SUnits(NS.begin(), NS.end()); 2284 llvm::sort(SUnits, [](const SUnit *A, const SUnit *B) { 2285 return A->NodeNum > B->NodeNum; 2286 }); 2287 2288 for (auto &SU : SUnits) { 2289 // Since we're computing the register pressure for a subset of the 2290 // instructions in a block, we need to set the tracker for each 2291 // instruction in the node-set. The tracker is set to the instruction 2292 // just after the one we're interested in. 2293 MachineBasicBlock::const_iterator CurInstI = SU->getInstr(); 2294 RecRPTracker.setPos(std::next(CurInstI)); 2295 2296 RegPressureDelta RPDelta; 2297 ArrayRef<PressureChange> CriticalPSets; 2298 RecRPTracker.getMaxUpwardPressureDelta(SU->getInstr(), nullptr, RPDelta, 2299 CriticalPSets, 2300 RecRegPressure.MaxSetPressure); 2301 if (RPDelta.Excess.isValid()) { 2302 LLVM_DEBUG( 2303 dbgs() << "Excess register pressure: SU(" << SU->NodeNum << ") " 2304 << TRI->getRegPressureSetName(RPDelta.Excess.getPSet()) 2305 << ":" << RPDelta.Excess.getUnitInc() << "\n"); 2306 NS.setExceedPressure(SU); 2307 break; 2308 } 2309 RecRPTracker.recede(); 2310 } 2311 } 2312 } 2313 2314 /// A heuristic to colocate node sets that have the same set of 2315 /// successors. 2316 void SwingSchedulerDAG::colocateNodeSets(NodeSetType &NodeSets) { 2317 unsigned Colocate = 0; 2318 for (int i = 0, e = NodeSets.size(); i < e; ++i) { 2319 NodeSet &N1 = NodeSets[i]; 2320 SmallSetVector<SUnit *, 8> S1; 2321 if (N1.empty() || !succ_L(N1, S1, DDG.get())) 2322 continue; 2323 for (int j = i + 1; j < e; ++j) { 2324 NodeSet &N2 = NodeSets[j]; 2325 if (N1.compareRecMII(N2) != 0) 2326 continue; 2327 SmallSetVector<SUnit *, 8> S2; 2328 if (N2.empty() || !succ_L(N2, S2, DDG.get())) 2329 continue; 2330 if (llvm::set_is_subset(S1, S2) && S1.size() == S2.size()) { 2331 N1.setColocate(++Colocate); 2332 N2.setColocate(Colocate); 2333 break; 2334 } 2335 } 2336 } 2337 } 2338 2339 /// Check if the existing node-sets are profitable. If not, then ignore the 2340 /// recurrent node-sets, and attempt to schedule all nodes together. This is 2341 /// a heuristic. If the MII is large and all the recurrent node-sets are small, 2342 /// then it's best to try to schedule all instructions together instead of 2343 /// starting with the recurrent node-sets. 2344 void SwingSchedulerDAG::checkNodeSets(NodeSetType &NodeSets) { 2345 // Look for loops with a large MII. 2346 if (MII < 17) 2347 return; 2348 // Check if the node-set contains only a simple add recurrence. 2349 for (auto &NS : NodeSets) { 2350 if (NS.getRecMII() > 2) 2351 return; 2352 if (NS.getMaxDepth() > MII) 2353 return; 2354 } 2355 NodeSets.clear(); 2356 LLVM_DEBUG(dbgs() << "Clear recurrence node-sets\n"); 2357 } 2358 2359 /// Add the nodes that do not belong to a recurrence set into groups 2360 /// based upon connected components. 2361 void SwingSchedulerDAG::groupRemainingNodes(NodeSetType &NodeSets) { 2362 SetVector<SUnit *> NodesAdded; 2363 SmallPtrSet<SUnit *, 8> Visited; 2364 // Add the nodes that are on a path between the previous node sets and 2365 // the current node set. 2366 for (NodeSet &I : NodeSets) { 2367 SmallSetVector<SUnit *, 8> N; 2368 // Add the nodes from the current node set to the previous node set. 2369 if (succ_L(I, N, DDG.get())) { 2370 SetVector<SUnit *> Path; 2371 for (SUnit *NI : N) { 2372 Visited.clear(); 2373 computePath(NI, Path, NodesAdded, I, Visited, DDG.get()); 2374 } 2375 if (!Path.empty()) 2376 I.insert(Path.begin(), Path.end()); 2377 } 2378 // Add the nodes from the previous node set to the current node set. 2379 N.clear(); 2380 if (succ_L(NodesAdded, N, DDG.get())) { 2381 SetVector<SUnit *> Path; 2382 for (SUnit *NI : N) { 2383 Visited.clear(); 2384 computePath(NI, Path, I, NodesAdded, Visited, DDG.get()); 2385 } 2386 if (!Path.empty()) 2387 I.insert(Path.begin(), Path.end()); 2388 } 2389 NodesAdded.insert_range(I); 2390 } 2391 2392 // Create a new node set with the connected nodes of any successor of a node 2393 // in a recurrent set. 2394 NodeSet NewSet; 2395 SmallSetVector<SUnit *, 8> N; 2396 if (succ_L(NodesAdded, N, DDG.get())) 2397 for (SUnit *I : N) 2398 addConnectedNodes(I, NewSet, NodesAdded); 2399 if (!NewSet.empty()) 2400 NodeSets.push_back(NewSet); 2401 2402 // Create a new node set with the connected nodes of any predecessor of a node 2403 // in a recurrent set. 2404 NewSet.clear(); 2405 if (pred_L(NodesAdded, N, DDG.get())) 2406 for (SUnit *I : N) 2407 addConnectedNodes(I, NewSet, NodesAdded); 2408 if (!NewSet.empty()) 2409 NodeSets.push_back(NewSet); 2410 2411 // Create new nodes sets with the connected nodes any remaining node that 2412 // has no predecessor. 2413 for (SUnit &SU : SUnits) { 2414 if (NodesAdded.count(&SU) == 0) { 2415 NewSet.clear(); 2416 addConnectedNodes(&SU, NewSet, NodesAdded); 2417 if (!NewSet.empty()) 2418 NodeSets.push_back(NewSet); 2419 } 2420 } 2421 } 2422 2423 /// Add the node to the set, and add all of its connected nodes to the set. 2424 void SwingSchedulerDAG::addConnectedNodes(SUnit *SU, NodeSet &NewSet, 2425 SetVector<SUnit *> &NodesAdded) { 2426 NewSet.insert(SU); 2427 NodesAdded.insert(SU); 2428 for (auto &OE : DDG->getOutEdges(SU)) { 2429 SUnit *Successor = OE.getDst(); 2430 if (!OE.isArtificial() && !Successor->isBoundaryNode() && 2431 NodesAdded.count(Successor) == 0) 2432 addConnectedNodes(Successor, NewSet, NodesAdded); 2433 } 2434 for (auto &IE : DDG->getInEdges(SU)) { 2435 SUnit *Predecessor = IE.getSrc(); 2436 if (!IE.isArtificial() && NodesAdded.count(Predecessor) == 0) 2437 addConnectedNodes(Predecessor, NewSet, NodesAdded); 2438 } 2439 } 2440 2441 /// Return true if Set1 contains elements in Set2. The elements in common 2442 /// are returned in a different container. 2443 static bool isIntersect(SmallSetVector<SUnit *, 8> &Set1, const NodeSet &Set2, 2444 SmallSetVector<SUnit *, 8> &Result) { 2445 Result.clear(); 2446 for (SUnit *SU : Set1) { 2447 if (Set2.count(SU) != 0) 2448 Result.insert(SU); 2449 } 2450 return !Result.empty(); 2451 } 2452 2453 /// Merge the recurrence node sets that have the same initial node. 2454 void SwingSchedulerDAG::fuseRecs(NodeSetType &NodeSets) { 2455 for (NodeSetType::iterator I = NodeSets.begin(), E = NodeSets.end(); I != E; 2456 ++I) { 2457 NodeSet &NI = *I; 2458 for (NodeSetType::iterator J = I + 1; J != E;) { 2459 NodeSet &NJ = *J; 2460 if (NI.getNode(0)->NodeNum == NJ.getNode(0)->NodeNum) { 2461 if (NJ.compareRecMII(NI) > 0) 2462 NI.setRecMII(NJ.getRecMII()); 2463 for (SUnit *SU : *J) 2464 I->insert(SU); 2465 NodeSets.erase(J); 2466 E = NodeSets.end(); 2467 } else { 2468 ++J; 2469 } 2470 } 2471 } 2472 } 2473 2474 /// Remove nodes that have been scheduled in previous NodeSets. 2475 void SwingSchedulerDAG::removeDuplicateNodes(NodeSetType &NodeSets) { 2476 for (NodeSetType::iterator I = NodeSets.begin(), E = NodeSets.end(); I != E; 2477 ++I) 2478 for (NodeSetType::iterator J = I + 1; J != E;) { 2479 J->remove_if([&](SUnit *SUJ) { return I->count(SUJ); }); 2480 2481 if (J->empty()) { 2482 NodeSets.erase(J); 2483 E = NodeSets.end(); 2484 } else { 2485 ++J; 2486 } 2487 } 2488 } 2489 2490 /// Compute an ordered list of the dependence graph nodes, which 2491 /// indicates the order that the nodes will be scheduled. This is a 2492 /// two-level algorithm. First, a partial order is created, which 2493 /// consists of a list of sets ordered from highest to lowest priority. 2494 void SwingSchedulerDAG::computeNodeOrder(NodeSetType &NodeSets) { 2495 SmallSetVector<SUnit *, 8> R; 2496 NodeOrder.clear(); 2497 2498 for (auto &Nodes : NodeSets) { 2499 LLVM_DEBUG(dbgs() << "NodeSet size " << Nodes.size() << "\n"); 2500 OrderKind Order; 2501 SmallSetVector<SUnit *, 8> N; 2502 if (pred_L(NodeOrder, N, DDG.get()) && llvm::set_is_subset(N, Nodes)) { 2503 R.insert_range(N); 2504 Order = BottomUp; 2505 LLVM_DEBUG(dbgs() << " Bottom up (preds) "); 2506 } else if (succ_L(NodeOrder, N, DDG.get()) && 2507 llvm::set_is_subset(N, Nodes)) { 2508 R.insert_range(N); 2509 Order = TopDown; 2510 LLVM_DEBUG(dbgs() << " Top down (succs) "); 2511 } else if (isIntersect(N, Nodes, R)) { 2512 // If some of the successors are in the existing node-set, then use the 2513 // top-down ordering. 2514 Order = TopDown; 2515 LLVM_DEBUG(dbgs() << " Top down (intersect) "); 2516 } else if (NodeSets.size() == 1) { 2517 for (const auto &N : Nodes) 2518 if (N->Succs.size() == 0) 2519 R.insert(N); 2520 Order = BottomUp; 2521 LLVM_DEBUG(dbgs() << " Bottom up (all) "); 2522 } else { 2523 // Find the node with the highest ASAP. 2524 SUnit *maxASAP = nullptr; 2525 for (SUnit *SU : Nodes) { 2526 if (maxASAP == nullptr || getASAP(SU) > getASAP(maxASAP) || 2527 (getASAP(SU) == getASAP(maxASAP) && SU->NodeNum > maxASAP->NodeNum)) 2528 maxASAP = SU; 2529 } 2530 R.insert(maxASAP); 2531 Order = BottomUp; 2532 LLVM_DEBUG(dbgs() << " Bottom up (default) "); 2533 } 2534 2535 while (!R.empty()) { 2536 if (Order == TopDown) { 2537 // Choose the node with the maximum height. If more than one, choose 2538 // the node wiTH the maximum ZeroLatencyHeight. If still more than one, 2539 // choose the node with the lowest MOV. 2540 while (!R.empty()) { 2541 SUnit *maxHeight = nullptr; 2542 for (SUnit *I : R) { 2543 if (maxHeight == nullptr || getHeight(I) > getHeight(maxHeight)) 2544 maxHeight = I; 2545 else if (getHeight(I) == getHeight(maxHeight) && 2546 getZeroLatencyHeight(I) > getZeroLatencyHeight(maxHeight)) 2547 maxHeight = I; 2548 else if (getHeight(I) == getHeight(maxHeight) && 2549 getZeroLatencyHeight(I) == 2550 getZeroLatencyHeight(maxHeight) && 2551 getMOV(I) < getMOV(maxHeight)) 2552 maxHeight = I; 2553 } 2554 NodeOrder.insert(maxHeight); 2555 LLVM_DEBUG(dbgs() << maxHeight->NodeNum << " "); 2556 R.remove(maxHeight); 2557 for (const auto &OE : DDG->getOutEdges(maxHeight)) { 2558 SUnit *SU = OE.getDst(); 2559 if (Nodes.count(SU) == 0) 2560 continue; 2561 if (NodeOrder.contains(SU)) 2562 continue; 2563 if (OE.ignoreDependence(false)) 2564 continue; 2565 R.insert(SU); 2566 } 2567 2568 // FIXME: The following loop-carried dependencies may also need to be 2569 // considered. 2570 // - Physical register dependnecies (true-dependnece and WAW). 2571 // - Memory dependencies. 2572 for (const auto &IE : DDG->getInEdges(maxHeight)) { 2573 SUnit *SU = IE.getSrc(); 2574 if (!IE.isAntiDep()) 2575 continue; 2576 if (Nodes.count(SU) == 0) 2577 continue; 2578 if (NodeOrder.contains(SU)) 2579 continue; 2580 R.insert(SU); 2581 } 2582 } 2583 Order = BottomUp; 2584 LLVM_DEBUG(dbgs() << "\n Switching order to bottom up "); 2585 SmallSetVector<SUnit *, 8> N; 2586 if (pred_L(NodeOrder, N, DDG.get(), &Nodes)) 2587 R.insert_range(N); 2588 } else { 2589 // Choose the node with the maximum depth. If more than one, choose 2590 // the node with the maximum ZeroLatencyDepth. If still more than one, 2591 // choose the node with the lowest MOV. 2592 while (!R.empty()) { 2593 SUnit *maxDepth = nullptr; 2594 for (SUnit *I : R) { 2595 if (maxDepth == nullptr || getDepth(I) > getDepth(maxDepth)) 2596 maxDepth = I; 2597 else if (getDepth(I) == getDepth(maxDepth) && 2598 getZeroLatencyDepth(I) > getZeroLatencyDepth(maxDepth)) 2599 maxDepth = I; 2600 else if (getDepth(I) == getDepth(maxDepth) && 2601 getZeroLatencyDepth(I) == getZeroLatencyDepth(maxDepth) && 2602 getMOV(I) < getMOV(maxDepth)) 2603 maxDepth = I; 2604 } 2605 NodeOrder.insert(maxDepth); 2606 LLVM_DEBUG(dbgs() << maxDepth->NodeNum << " "); 2607 R.remove(maxDepth); 2608 if (Nodes.isExceedSU(maxDepth)) { 2609 Order = TopDown; 2610 R.clear(); 2611 R.insert(Nodes.getNode(0)); 2612 break; 2613 } 2614 for (const auto &IE : DDG->getInEdges(maxDepth)) { 2615 SUnit *SU = IE.getSrc(); 2616 if (Nodes.count(SU) == 0) 2617 continue; 2618 if (NodeOrder.contains(SU)) 2619 continue; 2620 R.insert(SU); 2621 } 2622 2623 // FIXME: The following loop-carried dependencies may also need to be 2624 // considered. 2625 // - Physical register dependnecies (true-dependnece and WAW). 2626 // - Memory dependencies. 2627 for (const auto &OE : DDG->getOutEdges(maxDepth)) { 2628 SUnit *SU = OE.getDst(); 2629 if (!OE.isAntiDep()) 2630 continue; 2631 if (Nodes.count(SU) == 0) 2632 continue; 2633 if (NodeOrder.contains(SU)) 2634 continue; 2635 R.insert(SU); 2636 } 2637 } 2638 Order = TopDown; 2639 LLVM_DEBUG(dbgs() << "\n Switching order to top down "); 2640 SmallSetVector<SUnit *, 8> N; 2641 if (succ_L(NodeOrder, N, DDG.get(), &Nodes)) 2642 R.insert_range(N); 2643 } 2644 } 2645 LLVM_DEBUG(dbgs() << "\nDone with Nodeset\n"); 2646 } 2647 2648 LLVM_DEBUG({ 2649 dbgs() << "Node order: "; 2650 for (SUnit *I : NodeOrder) 2651 dbgs() << " " << I->NodeNum << " "; 2652 dbgs() << "\n"; 2653 }); 2654 } 2655 2656 /// Process the nodes in the computed order and create the pipelined schedule 2657 /// of the instructions, if possible. Return true if a schedule is found. 2658 bool SwingSchedulerDAG::schedulePipeline(SMSchedule &Schedule) { 2659 2660 if (NodeOrder.empty()){ 2661 LLVM_DEBUG(dbgs() << "NodeOrder is empty! abort scheduling\n" ); 2662 return false; 2663 } 2664 2665 bool scheduleFound = false; 2666 std::unique_ptr<HighRegisterPressureDetector> HRPDetector; 2667 if (LimitRegPressure) { 2668 HRPDetector = 2669 std::make_unique<HighRegisterPressureDetector>(Loop.getHeader(), MF); 2670 HRPDetector->init(RegClassInfo); 2671 } 2672 // Keep increasing II until a valid schedule is found. 2673 for (unsigned II = MII; II <= MAX_II && !scheduleFound; ++II) { 2674 Schedule.reset(); 2675 Schedule.setInitiationInterval(II); 2676 LLVM_DEBUG(dbgs() << "Try to schedule with " << II << "\n"); 2677 2678 SetVector<SUnit *>::iterator NI = NodeOrder.begin(); 2679 SetVector<SUnit *>::iterator NE = NodeOrder.end(); 2680 do { 2681 SUnit *SU = *NI; 2682 2683 // Compute the schedule time for the instruction, which is based 2684 // upon the scheduled time for any predecessors/successors. 2685 int EarlyStart = INT_MIN; 2686 int LateStart = INT_MAX; 2687 Schedule.computeStart(SU, &EarlyStart, &LateStart, II, this); 2688 LLVM_DEBUG({ 2689 dbgs() << "\n"; 2690 dbgs() << "Inst (" << SU->NodeNum << ") "; 2691 SU->getInstr()->dump(); 2692 dbgs() << "\n"; 2693 }); 2694 LLVM_DEBUG( 2695 dbgs() << format("\tes: %8x ls: %8x\n", EarlyStart, LateStart)); 2696 2697 if (EarlyStart > LateStart) 2698 scheduleFound = false; 2699 else if (EarlyStart != INT_MIN && LateStart == INT_MAX) 2700 scheduleFound = 2701 Schedule.insert(SU, EarlyStart, EarlyStart + (int)II - 1, II); 2702 else if (EarlyStart == INT_MIN && LateStart != INT_MAX) 2703 scheduleFound = 2704 Schedule.insert(SU, LateStart, LateStart - (int)II + 1, II); 2705 else if (EarlyStart != INT_MIN && LateStart != INT_MAX) { 2706 LateStart = std::min(LateStart, EarlyStart + (int)II - 1); 2707 // When scheduling a Phi it is better to start at the late cycle and 2708 // go backwards. The default order may insert the Phi too far away 2709 // from its first dependence. 2710 // Also, do backward search when all scheduled predecessors are 2711 // loop-carried output/order dependencies. Empirically, there are also 2712 // cases where scheduling becomes possible with backward search. 2713 if (SU->getInstr()->isPHI() || 2714 Schedule.onlyHasLoopCarriedOutputOrOrderPreds(SU, this->getDDG())) 2715 scheduleFound = Schedule.insert(SU, LateStart, EarlyStart, II); 2716 else 2717 scheduleFound = Schedule.insert(SU, EarlyStart, LateStart, II); 2718 } else { 2719 int FirstCycle = Schedule.getFirstCycle(); 2720 scheduleFound = Schedule.insert(SU, FirstCycle + getASAP(SU), 2721 FirstCycle + getASAP(SU) + II - 1, II); 2722 } 2723 2724 // Even if we find a schedule, make sure the schedule doesn't exceed the 2725 // allowable number of stages. We keep trying if this happens. 2726 if (scheduleFound) 2727 if (SwpMaxStages > -1 && 2728 Schedule.getMaxStageCount() > (unsigned)SwpMaxStages) 2729 scheduleFound = false; 2730 2731 LLVM_DEBUG({ 2732 if (!scheduleFound) 2733 dbgs() << "\tCan't schedule\n"; 2734 }); 2735 } while (++NI != NE && scheduleFound); 2736 2737 // If a schedule is found, validate it against the validation-only 2738 // dependencies. 2739 if (scheduleFound) 2740 scheduleFound = DDG->isValidSchedule(Schedule); 2741 2742 // If a schedule is found, ensure non-pipelined instructions are in stage 0 2743 if (scheduleFound) 2744 scheduleFound = 2745 Schedule.normalizeNonPipelinedInstructions(this, LoopPipelinerInfo); 2746 2747 // If a schedule is found, check if it is a valid schedule too. 2748 if (scheduleFound) 2749 scheduleFound = Schedule.isValidSchedule(this); 2750 2751 // If a schedule was found and the option is enabled, check if the schedule 2752 // might generate additional register spills/fills. 2753 if (scheduleFound && LimitRegPressure) 2754 scheduleFound = 2755 !HRPDetector->detect(this, Schedule, Schedule.getMaxStageCount()); 2756 } 2757 2758 LLVM_DEBUG(dbgs() << "Schedule Found? " << scheduleFound 2759 << " (II=" << Schedule.getInitiationInterval() 2760 << ")\n"); 2761 2762 if (scheduleFound) { 2763 scheduleFound = LoopPipelinerInfo->shouldUseSchedule(*this, Schedule); 2764 if (!scheduleFound) 2765 LLVM_DEBUG(dbgs() << "Target rejected schedule\n"); 2766 } 2767 2768 if (scheduleFound) { 2769 Schedule.finalizeSchedule(this); 2770 Pass.ORE->emit([&]() { 2771 return MachineOptimizationRemarkAnalysis( 2772 DEBUG_TYPE, "schedule", Loop.getStartLoc(), Loop.getHeader()) 2773 << "Schedule found with Initiation Interval: " 2774 << ore::NV("II", Schedule.getInitiationInterval()) 2775 << ", MaxStageCount: " 2776 << ore::NV("MaxStageCount", Schedule.getMaxStageCount()); 2777 }); 2778 } else 2779 Schedule.reset(); 2780 2781 return scheduleFound && Schedule.getMaxStageCount() > 0; 2782 } 2783 2784 static Register findUniqueOperandDefinedInLoop(const MachineInstr &MI) { 2785 const MachineRegisterInfo &MRI = MI.getParent()->getParent()->getRegInfo(); 2786 Register Result; 2787 for (const MachineOperand &Use : MI.all_uses()) { 2788 Register Reg = Use.getReg(); 2789 if (!Reg.isVirtual()) 2790 return Register(); 2791 if (MRI.getVRegDef(Reg)->getParent() != MI.getParent()) 2792 continue; 2793 if (Result) 2794 return Register(); 2795 Result = Reg; 2796 } 2797 return Result; 2798 } 2799 2800 /// When Op is a value that is incremented recursively in a loop and there is a 2801 /// unique instruction that increments it, returns true and sets Value. 2802 static bool findLoopIncrementValue(const MachineOperand &Op, int &Value) { 2803 if (!Op.isReg() || !Op.getReg().isVirtual()) 2804 return false; 2805 2806 Register OrgReg = Op.getReg(); 2807 Register CurReg = OrgReg; 2808 const MachineBasicBlock *LoopBB = Op.getParent()->getParent(); 2809 const MachineRegisterInfo &MRI = LoopBB->getParent()->getRegInfo(); 2810 2811 const TargetInstrInfo *TII = 2812 LoopBB->getParent()->getSubtarget().getInstrInfo(); 2813 const TargetRegisterInfo *TRI = 2814 LoopBB->getParent()->getSubtarget().getRegisterInfo(); 2815 2816 MachineInstr *Phi = nullptr; 2817 MachineInstr *Increment = nullptr; 2818 2819 // Traverse definitions until it reaches Op or an instruction that does not 2820 // satisfy the condition. 2821 // Acceptable example: 2822 // bb.0: 2823 // %0 = PHI %3, %bb.0, ... 2824 // %2 = ADD %0, Value 2825 // ... = LOAD %2(Op) 2826 // %3 = COPY %2 2827 while (true) { 2828 if (!CurReg.isValid() || !CurReg.isVirtual()) 2829 return false; 2830 MachineInstr *Def = MRI.getVRegDef(CurReg); 2831 if (Def->getParent() != LoopBB) 2832 return false; 2833 2834 if (Def->isCopy()) { 2835 // Ignore copy instructions unless they contain subregisters 2836 if (Def->getOperand(0).getSubReg() || Def->getOperand(1).getSubReg()) 2837 return false; 2838 CurReg = Def->getOperand(1).getReg(); 2839 } else if (Def->isPHI()) { 2840 // There must be just one Phi 2841 if (Phi) 2842 return false; 2843 Phi = Def; 2844 CurReg = getLoopPhiReg(*Def, LoopBB); 2845 } else if (TII->getIncrementValue(*Def, Value)) { 2846 // Potentially a unique increment 2847 if (Increment) 2848 // Multiple increments exist 2849 return false; 2850 2851 const MachineOperand *BaseOp; 2852 int64_t Offset; 2853 bool OffsetIsScalable; 2854 if (TII->getMemOperandWithOffset(*Def, BaseOp, Offset, OffsetIsScalable, 2855 TRI)) { 2856 // Pre/post increment instruction 2857 CurReg = BaseOp->getReg(); 2858 } else { 2859 // If only one of the operands is defined within the loop, it is assumed 2860 // to be an incremented value. 2861 CurReg = findUniqueOperandDefinedInLoop(*Def); 2862 if (!CurReg.isValid()) 2863 return false; 2864 } 2865 Increment = Def; 2866 } else { 2867 return false; 2868 } 2869 if (CurReg == OrgReg) 2870 break; 2871 } 2872 2873 if (!Phi || !Increment) 2874 return false; 2875 2876 return true; 2877 } 2878 2879 /// Return true if we can compute the amount the instruction changes 2880 /// during each iteration. Set Delta to the amount of the change. 2881 bool SwingSchedulerDAG::computeDelta(const MachineInstr &MI, int &Delta) const { 2882 const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo(); 2883 const MachineOperand *BaseOp; 2884 int64_t Offset; 2885 bool OffsetIsScalable; 2886 if (!TII->getMemOperandWithOffset(MI, BaseOp, Offset, OffsetIsScalable, TRI)) 2887 return false; 2888 2889 // FIXME: This algorithm assumes instructions have fixed-size offsets. 2890 if (OffsetIsScalable) 2891 return false; 2892 2893 if (!BaseOp->isReg()) 2894 return false; 2895 2896 return findLoopIncrementValue(*BaseOp, Delta); 2897 } 2898 2899 /// Check if we can change the instruction to use an offset value from the 2900 /// previous iteration. If so, return true and set the base and offset values 2901 /// so that we can rewrite the load, if necessary. 2902 /// v1 = Phi(v0, v3) 2903 /// v2 = load v1, 0 2904 /// v3 = post_store v1, 4, x 2905 /// This function enables the load to be rewritten as v2 = load v3, 4. 2906 bool SwingSchedulerDAG::canUseLastOffsetValue(MachineInstr *MI, 2907 unsigned &BasePos, 2908 unsigned &OffsetPos, 2909 Register &NewBase, 2910 int64_t &Offset) { 2911 // Get the load instruction. 2912 if (TII->isPostIncrement(*MI)) 2913 return false; 2914 unsigned BasePosLd, OffsetPosLd; 2915 if (!TII->getBaseAndOffsetPosition(*MI, BasePosLd, OffsetPosLd)) 2916 return false; 2917 Register BaseReg = MI->getOperand(BasePosLd).getReg(); 2918 2919 // Look for the Phi instruction. 2920 MachineRegisterInfo &MRI = MI->getMF()->getRegInfo(); 2921 MachineInstr *Phi = MRI.getVRegDef(BaseReg); 2922 if (!Phi || !Phi->isPHI()) 2923 return false; 2924 // Get the register defined in the loop block. 2925 Register PrevReg = getLoopPhiReg(*Phi, MI->getParent()); 2926 if (!PrevReg) 2927 return false; 2928 2929 // Check for the post-increment load/store instruction. 2930 MachineInstr *PrevDef = MRI.getVRegDef(PrevReg); 2931 if (!PrevDef || PrevDef == MI) 2932 return false; 2933 2934 if (!TII->isPostIncrement(*PrevDef)) 2935 return false; 2936 2937 unsigned BasePos1 = 0, OffsetPos1 = 0; 2938 if (!TII->getBaseAndOffsetPosition(*PrevDef, BasePos1, OffsetPos1)) 2939 return false; 2940 2941 // Make sure that the instructions do not access the same memory location in 2942 // the next iteration. 2943 int64_t LoadOffset = MI->getOperand(OffsetPosLd).getImm(); 2944 int64_t StoreOffset = PrevDef->getOperand(OffsetPos1).getImm(); 2945 MachineInstr *NewMI = MF.CloneMachineInstr(MI); 2946 NewMI->getOperand(OffsetPosLd).setImm(LoadOffset + StoreOffset); 2947 bool Disjoint = TII->areMemAccessesTriviallyDisjoint(*NewMI, *PrevDef); 2948 MF.deleteMachineInstr(NewMI); 2949 if (!Disjoint) 2950 return false; 2951 2952 // Set the return value once we determine that we return true. 2953 BasePos = BasePosLd; 2954 OffsetPos = OffsetPosLd; 2955 NewBase = PrevReg; 2956 Offset = StoreOffset; 2957 return true; 2958 } 2959 2960 /// Apply changes to the instruction if needed. The changes are need 2961 /// to improve the scheduling and depend up on the final schedule. 2962 void SwingSchedulerDAG::applyInstrChange(MachineInstr *MI, 2963 SMSchedule &Schedule) { 2964 SUnit *SU = getSUnit(MI); 2965 DenseMap<SUnit *, std::pair<Register, int64_t>>::iterator It = 2966 InstrChanges.find(SU); 2967 if (It != InstrChanges.end()) { 2968 std::pair<Register, int64_t> RegAndOffset = It->second; 2969 unsigned BasePos, OffsetPos; 2970 if (!TII->getBaseAndOffsetPosition(*MI, BasePos, OffsetPos)) 2971 return; 2972 Register BaseReg = MI->getOperand(BasePos).getReg(); 2973 MachineInstr *LoopDef = findDefInLoop(BaseReg); 2974 int DefStageNum = Schedule.stageScheduled(getSUnit(LoopDef)); 2975 int DefCycleNum = Schedule.cycleScheduled(getSUnit(LoopDef)); 2976 int BaseStageNum = Schedule.stageScheduled(SU); 2977 int BaseCycleNum = Schedule.cycleScheduled(SU); 2978 if (BaseStageNum < DefStageNum) { 2979 MachineInstr *NewMI = MF.CloneMachineInstr(MI); 2980 int OffsetDiff = DefStageNum - BaseStageNum; 2981 if (DefCycleNum < BaseCycleNum) { 2982 NewMI->getOperand(BasePos).setReg(RegAndOffset.first); 2983 if (OffsetDiff > 0) 2984 --OffsetDiff; 2985 } 2986 int64_t NewOffset = 2987 MI->getOperand(OffsetPos).getImm() + RegAndOffset.second * OffsetDiff; 2988 NewMI->getOperand(OffsetPos).setImm(NewOffset); 2989 SU->setInstr(NewMI); 2990 MISUnitMap[NewMI] = SU; 2991 NewMIs[MI] = NewMI; 2992 } 2993 } 2994 } 2995 2996 /// Return the instruction in the loop that defines the register. 2997 /// If the definition is a Phi, then follow the Phi operand to 2998 /// the instruction in the loop. 2999 MachineInstr *SwingSchedulerDAG::findDefInLoop(Register Reg) { 3000 SmallPtrSet<MachineInstr *, 8> Visited; 3001 MachineInstr *Def = MRI.getVRegDef(Reg); 3002 while (Def->isPHI()) { 3003 if (!Visited.insert(Def).second) 3004 break; 3005 for (unsigned i = 1, e = Def->getNumOperands(); i < e; i += 2) 3006 if (Def->getOperand(i + 1).getMBB() == BB) { 3007 Def = MRI.getVRegDef(Def->getOperand(i).getReg()); 3008 break; 3009 } 3010 } 3011 return Def; 3012 } 3013 3014 /// Return false if there is no overlap between the region accessed by BaseMI in 3015 /// an iteration and the region accessed by OtherMI in subsequent iterations. 3016 bool SwingSchedulerDAG::mayOverlapInLaterIter( 3017 const MachineInstr *BaseMI, const MachineInstr *OtherMI) const { 3018 int DeltaB, DeltaO, Delta; 3019 if (!computeDelta(*BaseMI, DeltaB) || !computeDelta(*OtherMI, DeltaO) || 3020 DeltaB != DeltaO) 3021 return true; 3022 Delta = DeltaB; 3023 3024 const MachineOperand *BaseOpB, *BaseOpO; 3025 int64_t OffsetB, OffsetO; 3026 bool OffsetBIsScalable, OffsetOIsScalable; 3027 const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo(); 3028 if (!TII->getMemOperandWithOffset(*BaseMI, BaseOpB, OffsetB, 3029 OffsetBIsScalable, TRI) || 3030 !TII->getMemOperandWithOffset(*OtherMI, BaseOpO, OffsetO, 3031 OffsetOIsScalable, TRI)) 3032 return true; 3033 3034 if (OffsetBIsScalable || OffsetOIsScalable) 3035 return true; 3036 3037 if (!BaseOpB->isIdenticalTo(*BaseOpO)) { 3038 // Pass cases with different base operands but same initial values. 3039 // Typically for when pre/post increment is used. 3040 3041 if (!BaseOpB->isReg() || !BaseOpO->isReg()) 3042 return true; 3043 Register RegB = BaseOpB->getReg(), RegO = BaseOpO->getReg(); 3044 if (!RegB.isVirtual() || !RegO.isVirtual()) 3045 return true; 3046 3047 MachineInstr *DefB = MRI.getVRegDef(BaseOpB->getReg()); 3048 MachineInstr *DefO = MRI.getVRegDef(BaseOpO->getReg()); 3049 if (!DefB || !DefO || !DefB->isPHI() || !DefO->isPHI()) 3050 return true; 3051 3052 Register InitValB; 3053 Register LoopValB; 3054 Register InitValO; 3055 Register LoopValO; 3056 getPhiRegs(*DefB, BB, InitValB, LoopValB); 3057 getPhiRegs(*DefO, BB, InitValO, LoopValO); 3058 MachineInstr *InitDefB = MRI.getVRegDef(InitValB); 3059 MachineInstr *InitDefO = MRI.getVRegDef(InitValO); 3060 3061 if (!InitDefB->isIdenticalTo(*InitDefO)) 3062 return true; 3063 } 3064 3065 LocationSize AccessSizeB = (*BaseMI->memoperands_begin())->getSize(); 3066 LocationSize AccessSizeO = (*OtherMI->memoperands_begin())->getSize(); 3067 3068 // This is the main test, which checks the offset values and the loop 3069 // increment value to determine if the accesses may be loop carried. 3070 if (!AccessSizeB.hasValue() || !AccessSizeO.hasValue()) 3071 return true; 3072 3073 LLVM_DEBUG({ 3074 dbgs() << "Overlap check:\n"; 3075 dbgs() << " BaseMI: "; 3076 BaseMI->dump(); 3077 dbgs() << " Base + " << OffsetB << " + I * " << Delta 3078 << ", Len: " << AccessSizeB.getValue() << "\n"; 3079 dbgs() << " OtherMI: "; 3080 OtherMI->dump(); 3081 dbgs() << " Base + " << OffsetO << " + I * " << Delta 3082 << ", Len: " << AccessSizeO.getValue() << "\n"; 3083 }); 3084 3085 // Excessive overlap may be detected in strided patterns. 3086 // For example, the memory addresses of the store and the load in 3087 // for (i=0; i<n; i+=2) a[i+1] = a[i]; 3088 // are assumed to overlap. 3089 if (Delta < 0) { 3090 int64_t BaseMinAddr = OffsetB; 3091 int64_t OhterNextIterMaxAddr = OffsetO + Delta + AccessSizeO.getValue() - 1; 3092 if (BaseMinAddr > OhterNextIterMaxAddr) { 3093 LLVM_DEBUG(dbgs() << " Result: No overlap\n"); 3094 return false; 3095 } 3096 } else { 3097 int64_t BaseMaxAddr = OffsetB + AccessSizeB.getValue() - 1; 3098 int64_t OtherNextIterMinAddr = OffsetO + Delta; 3099 if (BaseMaxAddr < OtherNextIterMinAddr) { 3100 LLVM_DEBUG(dbgs() << " Result: No overlap\n"); 3101 return false; 3102 } 3103 } 3104 LLVM_DEBUG(dbgs() << " Result: Overlap\n"); 3105 return true; 3106 } 3107 3108 /// Return true for an order or output dependence that is loop carried 3109 /// potentially. A dependence is loop carried if the destination defines a value 3110 /// that may be used or defined by the source in a subsequent iteration. 3111 bool SwingSchedulerDAG::isLoopCarriedDep( 3112 const SwingSchedulerDDGEdge &Edge) const { 3113 if ((!Edge.isOrderDep() && !Edge.isOutputDep()) || Edge.isArtificial() || 3114 Edge.getDst()->isBoundaryNode()) 3115 return false; 3116 3117 if (!SwpPruneLoopCarried) 3118 return true; 3119 3120 if (Edge.isOutputDep()) 3121 return true; 3122 3123 MachineInstr *SI = Edge.getSrc()->getInstr(); 3124 MachineInstr *DI = Edge.getDst()->getInstr(); 3125 assert(SI != nullptr && DI != nullptr && "Expecting SUnit with an MI."); 3126 3127 // Assume ordered loads and stores may have a loop carried dependence. 3128 if (SI->hasUnmodeledSideEffects() || DI->hasUnmodeledSideEffects() || 3129 SI->mayRaiseFPException() || DI->mayRaiseFPException() || 3130 SI->hasOrderedMemoryRef() || DI->hasOrderedMemoryRef()) 3131 return true; 3132 3133 if (!DI->mayLoadOrStore() || !SI->mayLoadOrStore()) 3134 return false; 3135 3136 // The conservative assumption is that a dependence between memory operations 3137 // may be loop carried. The following code checks when it can be proved that 3138 // there is no loop carried dependence. 3139 return mayOverlapInLaterIter(DI, SI); 3140 } 3141 3142 void SwingSchedulerDAG::postProcessDAG() { 3143 for (auto &M : Mutations) 3144 M->apply(this); 3145 } 3146 3147 /// Try to schedule the node at the specified StartCycle and continue 3148 /// until the node is schedule or the EndCycle is reached. This function 3149 /// returns true if the node is scheduled. This routine may search either 3150 /// forward or backward for a place to insert the instruction based upon 3151 /// the relative values of StartCycle and EndCycle. 3152 bool SMSchedule::insert(SUnit *SU, int StartCycle, int EndCycle, int II) { 3153 bool forward = true; 3154 LLVM_DEBUG({ 3155 dbgs() << "Trying to insert node between " << StartCycle << " and " 3156 << EndCycle << " II: " << II << "\n"; 3157 }); 3158 if (StartCycle > EndCycle) 3159 forward = false; 3160 3161 // The terminating condition depends on the direction. 3162 int termCycle = forward ? EndCycle + 1 : EndCycle - 1; 3163 for (int curCycle = StartCycle; curCycle != termCycle; 3164 forward ? ++curCycle : --curCycle) { 3165 3166 if (ST.getInstrInfo()->isZeroCost(SU->getInstr()->getOpcode()) || 3167 ProcItinResources.canReserveResources(*SU, curCycle)) { 3168 LLVM_DEBUG({ 3169 dbgs() << "\tinsert at cycle " << curCycle << " "; 3170 SU->getInstr()->dump(); 3171 }); 3172 3173 if (!ST.getInstrInfo()->isZeroCost(SU->getInstr()->getOpcode())) 3174 ProcItinResources.reserveResources(*SU, curCycle); 3175 ScheduledInstrs[curCycle].push_back(SU); 3176 InstrToCycle.insert(std::make_pair(SU, curCycle)); 3177 if (curCycle > LastCycle) 3178 LastCycle = curCycle; 3179 if (curCycle < FirstCycle) 3180 FirstCycle = curCycle; 3181 return true; 3182 } 3183 LLVM_DEBUG({ 3184 dbgs() << "\tfailed to insert at cycle " << curCycle << " "; 3185 SU->getInstr()->dump(); 3186 }); 3187 } 3188 return false; 3189 } 3190 3191 // Return the cycle of the earliest scheduled instruction in the chain. 3192 int SMSchedule::earliestCycleInChain(const SwingSchedulerDDGEdge &Dep, 3193 const SwingSchedulerDDG *DDG) { 3194 SmallPtrSet<SUnit *, 8> Visited; 3195 SmallVector<SwingSchedulerDDGEdge, 8> Worklist; 3196 Worklist.push_back(Dep); 3197 int EarlyCycle = INT_MAX; 3198 while (!Worklist.empty()) { 3199 const SwingSchedulerDDGEdge &Cur = Worklist.pop_back_val(); 3200 SUnit *PrevSU = Cur.getSrc(); 3201 if (Visited.count(PrevSU)) 3202 continue; 3203 std::map<SUnit *, int>::const_iterator it = InstrToCycle.find(PrevSU); 3204 if (it == InstrToCycle.end()) 3205 continue; 3206 EarlyCycle = std::min(EarlyCycle, it->second); 3207 for (const auto &IE : DDG->getInEdges(PrevSU)) 3208 if (IE.isOrderDep() || IE.isOutputDep()) 3209 Worklist.push_back(IE); 3210 Visited.insert(PrevSU); 3211 } 3212 return EarlyCycle; 3213 } 3214 3215 // Return the cycle of the latest scheduled instruction in the chain. 3216 int SMSchedule::latestCycleInChain(const SwingSchedulerDDGEdge &Dep, 3217 const SwingSchedulerDDG *DDG) { 3218 SmallPtrSet<SUnit *, 8> Visited; 3219 SmallVector<SwingSchedulerDDGEdge, 8> Worklist; 3220 Worklist.push_back(Dep); 3221 int LateCycle = INT_MIN; 3222 while (!Worklist.empty()) { 3223 const SwingSchedulerDDGEdge &Cur = Worklist.pop_back_val(); 3224 SUnit *SuccSU = Cur.getDst(); 3225 if (Visited.count(SuccSU) || SuccSU->isBoundaryNode()) 3226 continue; 3227 std::map<SUnit *, int>::const_iterator it = InstrToCycle.find(SuccSU); 3228 if (it == InstrToCycle.end()) 3229 continue; 3230 LateCycle = std::max(LateCycle, it->second); 3231 for (const auto &OE : DDG->getOutEdges(SuccSU)) 3232 if (OE.isOrderDep() || OE.isOutputDep()) 3233 Worklist.push_back(OE); 3234 Visited.insert(SuccSU); 3235 } 3236 return LateCycle; 3237 } 3238 3239 /// If an instruction has a use that spans multiple iterations, then 3240 /// return true. These instructions are characterized by having a back-ege 3241 /// to a Phi, which contains a reference to another Phi. 3242 static SUnit *multipleIterations(SUnit *SU, SwingSchedulerDAG *DAG) { 3243 for (auto &P : SU->Preds) 3244 if (P.getKind() == SDep::Anti && P.getSUnit()->getInstr()->isPHI()) 3245 for (auto &S : P.getSUnit()->Succs) 3246 if (S.getKind() == SDep::Data && S.getSUnit()->getInstr()->isPHI()) 3247 return P.getSUnit(); 3248 return nullptr; 3249 } 3250 3251 /// Compute the scheduling start slot for the instruction. The start slot 3252 /// depends on any predecessor or successor nodes scheduled already. 3253 void SMSchedule::computeStart(SUnit *SU, int *MaxEarlyStart, int *MinLateStart, 3254 int II, SwingSchedulerDAG *DAG) { 3255 const SwingSchedulerDDG *DDG = DAG->getDDG(); 3256 3257 // Iterate over each instruction that has been scheduled already. The start 3258 // slot computation depends on whether the previously scheduled instruction 3259 // is a predecessor or successor of the specified instruction. 3260 for (int cycle = getFirstCycle(); cycle <= LastCycle; ++cycle) { 3261 for (SUnit *I : getInstructions(cycle)) { 3262 for (const auto &IE : DDG->getInEdges(SU)) { 3263 if (IE.getSrc() == I) { 3264 // FIXME: Add reverse edge to `DDG` instead of calling 3265 // `isLoopCarriedDep` 3266 if (DAG->isLoopCarriedDep(IE)) { 3267 int End = earliestCycleInChain(IE, DDG) + (II - 1); 3268 *MinLateStart = std::min(*MinLateStart, End); 3269 } 3270 int EarlyStart = cycle + IE.getLatency() - IE.getDistance() * II; 3271 *MaxEarlyStart = std::max(*MaxEarlyStart, EarlyStart); 3272 } 3273 } 3274 3275 for (const auto &OE : DDG->getOutEdges(SU)) { 3276 if (OE.getDst() == I) { 3277 // FIXME: Add reverse edge to `DDG` instead of calling 3278 // `isLoopCarriedDep` 3279 if (DAG->isLoopCarriedDep(OE)) { 3280 int Start = latestCycleInChain(OE, DDG) + 1 - II; 3281 *MaxEarlyStart = std::max(*MaxEarlyStart, Start); 3282 } 3283 int LateStart = cycle - OE.getLatency() + OE.getDistance() * II; 3284 *MinLateStart = std::min(*MinLateStart, LateStart); 3285 } 3286 } 3287 3288 SUnit *BE = multipleIterations(I, DAG); 3289 for (const auto &Dep : SU->Preds) { 3290 // For instruction that requires multiple iterations, make sure that 3291 // the dependent instruction is not scheduled past the definition. 3292 if (BE && Dep.getSUnit() == BE && !SU->getInstr()->isPHI() && 3293 !SU->isPred(I)) 3294 *MinLateStart = std::min(*MinLateStart, cycle); 3295 } 3296 } 3297 } 3298 } 3299 3300 /// Order the instructions within a cycle so that the definitions occur 3301 /// before the uses. Returns true if the instruction is added to the start 3302 /// of the list, or false if added to the end. 3303 void SMSchedule::orderDependence(const SwingSchedulerDAG *SSD, SUnit *SU, 3304 std::deque<SUnit *> &Insts) const { 3305 MachineInstr *MI = SU->getInstr(); 3306 bool OrderBeforeUse = false; 3307 bool OrderAfterDef = false; 3308 bool OrderBeforeDef = false; 3309 unsigned MoveDef = 0; 3310 unsigned MoveUse = 0; 3311 int StageInst1 = stageScheduled(SU); 3312 const SwingSchedulerDDG *DDG = SSD->getDDG(); 3313 3314 unsigned Pos = 0; 3315 for (std::deque<SUnit *>::iterator I = Insts.begin(), E = Insts.end(); I != E; 3316 ++I, ++Pos) { 3317 for (MachineOperand &MO : MI->operands()) { 3318 if (!MO.isReg() || !MO.getReg().isVirtual()) 3319 continue; 3320 3321 Register Reg = MO.getReg(); 3322 unsigned BasePos, OffsetPos; 3323 if (ST.getInstrInfo()->getBaseAndOffsetPosition(*MI, BasePos, OffsetPos)) 3324 if (MI->getOperand(BasePos).getReg() == Reg) 3325 if (Register NewReg = SSD->getInstrBaseReg(SU)) 3326 Reg = NewReg; 3327 bool Reads, Writes; 3328 std::tie(Reads, Writes) = 3329 (*I)->getInstr()->readsWritesVirtualRegister(Reg); 3330 if (MO.isDef() && Reads && stageScheduled(*I) <= StageInst1) { 3331 OrderBeforeUse = true; 3332 if (MoveUse == 0) 3333 MoveUse = Pos; 3334 } else if (MO.isDef() && Reads && stageScheduled(*I) > StageInst1) { 3335 // Add the instruction after the scheduled instruction. 3336 OrderAfterDef = true; 3337 MoveDef = Pos; 3338 } else if (MO.isUse() && Writes && stageScheduled(*I) == StageInst1) { 3339 if (cycleScheduled(*I) == cycleScheduled(SU) && !(*I)->isSucc(SU)) { 3340 OrderBeforeUse = true; 3341 if (MoveUse == 0) 3342 MoveUse = Pos; 3343 } else { 3344 OrderAfterDef = true; 3345 MoveDef = Pos; 3346 } 3347 } else if (MO.isUse() && Writes && stageScheduled(*I) > StageInst1) { 3348 OrderBeforeUse = true; 3349 if (MoveUse == 0) 3350 MoveUse = Pos; 3351 if (MoveUse != 0) { 3352 OrderAfterDef = true; 3353 MoveDef = Pos - 1; 3354 } 3355 } else if (MO.isUse() && Writes && stageScheduled(*I) < StageInst1) { 3356 // Add the instruction before the scheduled instruction. 3357 OrderBeforeUse = true; 3358 if (MoveUse == 0) 3359 MoveUse = Pos; 3360 } else if (MO.isUse() && stageScheduled(*I) == StageInst1 && 3361 isLoopCarriedDefOfUse(SSD, (*I)->getInstr(), MO)) { 3362 if (MoveUse == 0) { 3363 OrderBeforeDef = true; 3364 MoveUse = Pos; 3365 } 3366 } 3367 } 3368 // Check for order dependences between instructions. Make sure the source 3369 // is ordered before the destination. 3370 for (auto &OE : DDG->getOutEdges(SU)) { 3371 if (OE.getDst() != *I) 3372 continue; 3373 if (OE.isOrderDep() && stageScheduled(*I) == StageInst1) { 3374 OrderBeforeUse = true; 3375 if (Pos < MoveUse) 3376 MoveUse = Pos; 3377 } 3378 // We did not handle HW dependences in previous for loop, 3379 // and we normally set Latency = 0 for Anti/Output deps, 3380 // so may have nodes in same cycle with Anti/Output dependent on HW regs. 3381 else if ((OE.isAntiDep() || OE.isOutputDep()) && 3382 stageScheduled(*I) == StageInst1) { 3383 OrderBeforeUse = true; 3384 if ((MoveUse == 0) || (Pos < MoveUse)) 3385 MoveUse = Pos; 3386 } 3387 } 3388 for (auto &IE : DDG->getInEdges(SU)) { 3389 if (IE.getSrc() != *I) 3390 continue; 3391 if ((IE.isAntiDep() || IE.isOutputDep() || IE.isOrderDep()) && 3392 stageScheduled(*I) == StageInst1) { 3393 OrderAfterDef = true; 3394 MoveDef = Pos; 3395 } 3396 } 3397 } 3398 3399 // A circular dependence. 3400 if (OrderAfterDef && OrderBeforeUse && MoveUse == MoveDef) 3401 OrderBeforeUse = false; 3402 3403 // OrderAfterDef takes precedences over OrderBeforeDef. The latter is due 3404 // to a loop-carried dependence. 3405 if (OrderBeforeDef) 3406 OrderBeforeUse = !OrderAfterDef || (MoveUse > MoveDef); 3407 3408 // The uncommon case when the instruction order needs to be updated because 3409 // there is both a use and def. 3410 if (OrderBeforeUse && OrderAfterDef) { 3411 SUnit *UseSU = Insts.at(MoveUse); 3412 SUnit *DefSU = Insts.at(MoveDef); 3413 if (MoveUse > MoveDef) { 3414 Insts.erase(Insts.begin() + MoveUse); 3415 Insts.erase(Insts.begin() + MoveDef); 3416 } else { 3417 Insts.erase(Insts.begin() + MoveDef); 3418 Insts.erase(Insts.begin() + MoveUse); 3419 } 3420 orderDependence(SSD, UseSU, Insts); 3421 orderDependence(SSD, SU, Insts); 3422 orderDependence(SSD, DefSU, Insts); 3423 return; 3424 } 3425 // Put the new instruction first if there is a use in the list. Otherwise, 3426 // put it at the end of the list. 3427 if (OrderBeforeUse) 3428 Insts.push_front(SU); 3429 else 3430 Insts.push_back(SU); 3431 } 3432 3433 /// Return true if the scheduled Phi has a loop carried operand. 3434 bool SMSchedule::isLoopCarried(const SwingSchedulerDAG *SSD, 3435 MachineInstr &Phi) const { 3436 if (!Phi.isPHI()) 3437 return false; 3438 assert(Phi.isPHI() && "Expecting a Phi."); 3439 SUnit *DefSU = SSD->getSUnit(&Phi); 3440 unsigned DefCycle = cycleScheduled(DefSU); 3441 int DefStage = stageScheduled(DefSU); 3442 3443 Register InitVal; 3444 Register LoopVal; 3445 getPhiRegs(Phi, Phi.getParent(), InitVal, LoopVal); 3446 SUnit *UseSU = SSD->getSUnit(MRI.getVRegDef(LoopVal)); 3447 if (!UseSU) 3448 return true; 3449 if (UseSU->getInstr()->isPHI()) 3450 return true; 3451 unsigned LoopCycle = cycleScheduled(UseSU); 3452 int LoopStage = stageScheduled(UseSU); 3453 return (LoopCycle > DefCycle) || (LoopStage <= DefStage); 3454 } 3455 3456 /// Return true if the instruction is a definition that is loop carried 3457 /// and defines the use on the next iteration. 3458 /// v1 = phi(v2, v3) 3459 /// (Def) v3 = op v1 3460 /// (MO) = v1 3461 /// If MO appears before Def, then v1 and v3 may get assigned to the same 3462 /// register. 3463 bool SMSchedule::isLoopCarriedDefOfUse(const SwingSchedulerDAG *SSD, 3464 MachineInstr *Def, 3465 MachineOperand &MO) const { 3466 if (!MO.isReg()) 3467 return false; 3468 if (Def->isPHI()) 3469 return false; 3470 MachineInstr *Phi = MRI.getVRegDef(MO.getReg()); 3471 if (!Phi || !Phi->isPHI() || Phi->getParent() != Def->getParent()) 3472 return false; 3473 if (!isLoopCarried(SSD, *Phi)) 3474 return false; 3475 Register LoopReg = getLoopPhiReg(*Phi, Phi->getParent()); 3476 for (MachineOperand &DMO : Def->all_defs()) { 3477 if (DMO.getReg() == LoopReg) 3478 return true; 3479 } 3480 return false; 3481 } 3482 3483 /// Return true if all scheduled predecessors are loop-carried output/order 3484 /// dependencies. 3485 bool SMSchedule::onlyHasLoopCarriedOutputOrOrderPreds( 3486 SUnit *SU, const SwingSchedulerDDG *DDG) const { 3487 for (const auto &IE : DDG->getInEdges(SU)) 3488 if (InstrToCycle.count(IE.getSrc())) 3489 return false; 3490 return true; 3491 } 3492 3493 /// Determine transitive dependences of unpipelineable instructions 3494 SmallSet<SUnit *, 8> SMSchedule::computeUnpipelineableNodes( 3495 SwingSchedulerDAG *SSD, TargetInstrInfo::PipelinerLoopInfo *PLI) { 3496 SmallSet<SUnit *, 8> DoNotPipeline; 3497 SmallVector<SUnit *, 8> Worklist; 3498 3499 for (auto &SU : SSD->SUnits) 3500 if (SU.isInstr() && PLI->shouldIgnoreForPipelining(SU.getInstr())) 3501 Worklist.push_back(&SU); 3502 3503 const SwingSchedulerDDG *DDG = SSD->getDDG(); 3504 while (!Worklist.empty()) { 3505 auto SU = Worklist.pop_back_val(); 3506 if (DoNotPipeline.count(SU)) 3507 continue; 3508 LLVM_DEBUG(dbgs() << "Do not pipeline SU(" << SU->NodeNum << ")\n"); 3509 DoNotPipeline.insert(SU); 3510 for (const auto &IE : DDG->getInEdges(SU)) 3511 Worklist.push_back(IE.getSrc()); 3512 3513 // To preserve previous behavior and prevent regression 3514 // FIXME: Remove if this doesn't have significant impact on 3515 for (const auto &OE : DDG->getOutEdges(SU)) 3516 if (OE.getDistance() == 1) 3517 Worklist.push_back(OE.getDst()); 3518 } 3519 return DoNotPipeline; 3520 } 3521 3522 // Determine all instructions upon which any unpipelineable instruction depends 3523 // and ensure that they are in stage 0. If unable to do so, return false. 3524 bool SMSchedule::normalizeNonPipelinedInstructions( 3525 SwingSchedulerDAG *SSD, TargetInstrInfo::PipelinerLoopInfo *PLI) { 3526 SmallSet<SUnit *, 8> DNP = computeUnpipelineableNodes(SSD, PLI); 3527 3528 int NewLastCycle = INT_MIN; 3529 for (SUnit &SU : SSD->SUnits) { 3530 if (!SU.isInstr()) 3531 continue; 3532 if (!DNP.contains(&SU) || stageScheduled(&SU) == 0) { 3533 NewLastCycle = std::max(NewLastCycle, InstrToCycle[&SU]); 3534 continue; 3535 } 3536 3537 // Put the non-pipelined instruction as early as possible in the schedule 3538 int NewCycle = getFirstCycle(); 3539 for (const auto &IE : SSD->getDDG()->getInEdges(&SU)) 3540 if (IE.getDistance() == 0) 3541 NewCycle = std::max(InstrToCycle[IE.getSrc()], NewCycle); 3542 3543 // To preserve previous behavior and prevent regression 3544 // FIXME: Remove if this doesn't have significant impact on performance 3545 for (auto &OE : SSD->getDDG()->getOutEdges(&SU)) 3546 if (OE.getDistance() == 1) 3547 NewCycle = std::max(InstrToCycle[OE.getDst()], NewCycle); 3548 3549 int OldCycle = InstrToCycle[&SU]; 3550 if (OldCycle != NewCycle) { 3551 InstrToCycle[&SU] = NewCycle; 3552 auto &OldS = getInstructions(OldCycle); 3553 llvm::erase(OldS, &SU); 3554 getInstructions(NewCycle).emplace_back(&SU); 3555 LLVM_DEBUG(dbgs() << "SU(" << SU.NodeNum 3556 << ") is not pipelined; moving from cycle " << OldCycle 3557 << " to " << NewCycle << " Instr:" << *SU.getInstr()); 3558 } 3559 3560 // We traverse the SUs in the order of the original basic block. Computing 3561 // NewCycle in this order normally works fine because all dependencies 3562 // (except for loop-carried dependencies) don't violate the original order. 3563 // However, an artificial dependency (e.g., added by CopyToPhiMutation) can 3564 // break it. That is, there may be exist an artificial dependency from 3565 // bottom to top. In such a case, NewCycle may become too large to be 3566 // scheduled in Stage 0. For example, assume that Inst0 is in DNP in the 3567 // following case: 3568 // 3569 // | Inst0 <-+ 3570 // SU order | | artificial dep 3571 // | Inst1 --+ 3572 // v 3573 // 3574 // If Inst1 is scheduled at cycle N and is not at Stage 0, then NewCycle of 3575 // Inst0 must be greater than or equal to N so that Inst0 is not be 3576 // scheduled at Stage 0. In such cases, we reject this schedule at this 3577 // time. 3578 // FIXME: The reason for this is the existence of artificial dependencies 3579 // that are contradict to the original SU order. If ignoring artificial 3580 // dependencies does not affect correctness, then it is better to ignore 3581 // them. 3582 if (FirstCycle + InitiationInterval <= NewCycle) 3583 return false; 3584 3585 NewLastCycle = std::max(NewLastCycle, NewCycle); 3586 } 3587 LastCycle = NewLastCycle; 3588 return true; 3589 } 3590 3591 // Check if the generated schedule is valid. This function checks if 3592 // an instruction that uses a physical register is scheduled in a 3593 // different stage than the definition. The pipeliner does not handle 3594 // physical register values that may cross a basic block boundary. 3595 // Furthermore, if a physical def/use pair is assigned to the same 3596 // cycle, orderDependence does not guarantee def/use ordering, so that 3597 // case should be considered invalid. (The test checks for both 3598 // earlier and same-cycle use to be more robust.) 3599 bool SMSchedule::isValidSchedule(SwingSchedulerDAG *SSD) { 3600 for (SUnit &SU : SSD->SUnits) { 3601 if (!SU.hasPhysRegDefs) 3602 continue; 3603 int StageDef = stageScheduled(&SU); 3604 int CycleDef = InstrToCycle[&SU]; 3605 assert(StageDef != -1 && "Instruction should have been scheduled."); 3606 for (auto &OE : SSD->getDDG()->getOutEdges(&SU)) { 3607 SUnit *Dst = OE.getDst(); 3608 if (OE.isAssignedRegDep() && !Dst->isBoundaryNode()) 3609 if (OE.getReg().isPhysical()) { 3610 if (stageScheduled(Dst) != StageDef) 3611 return false; 3612 if (InstrToCycle[Dst] <= CycleDef) 3613 return false; 3614 } 3615 } 3616 } 3617 return true; 3618 } 3619 3620 /// A property of the node order in swing-modulo-scheduling is 3621 /// that for nodes outside circuits the following holds: 3622 /// none of them is scheduled after both a successor and a 3623 /// predecessor. 3624 /// The method below checks whether the property is met. 3625 /// If not, debug information is printed and statistics information updated. 3626 /// Note that we do not use an assert statement. 3627 /// The reason is that although an invalid node order may prevent 3628 /// the pipeliner from finding a pipelined schedule for arbitrary II, 3629 /// it does not lead to the generation of incorrect code. 3630 void SwingSchedulerDAG::checkValidNodeOrder(const NodeSetType &Circuits) const { 3631 3632 // a sorted vector that maps each SUnit to its index in the NodeOrder 3633 typedef std::pair<SUnit *, unsigned> UnitIndex; 3634 std::vector<UnitIndex> Indices(NodeOrder.size(), std::make_pair(nullptr, 0)); 3635 3636 for (unsigned i = 0, s = NodeOrder.size(); i < s; ++i) 3637 Indices.push_back(std::make_pair(NodeOrder[i], i)); 3638 3639 auto CompareKey = [](UnitIndex i1, UnitIndex i2) { 3640 return std::get<0>(i1) < std::get<0>(i2); 3641 }; 3642 3643 // sort, so that we can perform a binary search 3644 llvm::sort(Indices, CompareKey); 3645 3646 bool Valid = true; 3647 (void)Valid; 3648 // for each SUnit in the NodeOrder, check whether 3649 // it appears after both a successor and a predecessor 3650 // of the SUnit. If this is the case, and the SUnit 3651 // is not part of circuit, then the NodeOrder is not 3652 // valid. 3653 for (unsigned i = 0, s = NodeOrder.size(); i < s; ++i) { 3654 SUnit *SU = NodeOrder[i]; 3655 unsigned Index = i; 3656 3657 bool PredBefore = false; 3658 bool SuccBefore = false; 3659 3660 SUnit *Succ; 3661 SUnit *Pred; 3662 (void)Succ; 3663 (void)Pred; 3664 3665 for (const auto &IE : DDG->getInEdges(SU)) { 3666 SUnit *PredSU = IE.getSrc(); 3667 unsigned PredIndex = std::get<1>( 3668 *llvm::lower_bound(Indices, std::make_pair(PredSU, 0), CompareKey)); 3669 if (!PredSU->getInstr()->isPHI() && PredIndex < Index) { 3670 PredBefore = true; 3671 Pred = PredSU; 3672 break; 3673 } 3674 } 3675 3676 for (const auto &OE : DDG->getOutEdges(SU)) { 3677 SUnit *SuccSU = OE.getDst(); 3678 // Do not process a boundary node, it was not included in NodeOrder, 3679 // hence not in Indices either, call to std::lower_bound() below will 3680 // return Indices.end(). 3681 if (SuccSU->isBoundaryNode()) 3682 continue; 3683 unsigned SuccIndex = std::get<1>( 3684 *llvm::lower_bound(Indices, std::make_pair(SuccSU, 0), CompareKey)); 3685 if (!SuccSU->getInstr()->isPHI() && SuccIndex < Index) { 3686 SuccBefore = true; 3687 Succ = SuccSU; 3688 break; 3689 } 3690 } 3691 3692 if (PredBefore && SuccBefore && !SU->getInstr()->isPHI()) { 3693 // instructions in circuits are allowed to be scheduled 3694 // after both a successor and predecessor. 3695 bool InCircuit = llvm::any_of( 3696 Circuits, [SU](const NodeSet &Circuit) { return Circuit.count(SU); }); 3697 if (InCircuit) 3698 LLVM_DEBUG(dbgs() << "In a circuit, predecessor "); 3699 else { 3700 Valid = false; 3701 NumNodeOrderIssues++; 3702 LLVM_DEBUG(dbgs() << "Predecessor "); 3703 } 3704 LLVM_DEBUG(dbgs() << Pred->NodeNum << " and successor " << Succ->NodeNum 3705 << " are scheduled before node " << SU->NodeNum 3706 << "\n"); 3707 } 3708 } 3709 3710 LLVM_DEBUG({ 3711 if (!Valid) 3712 dbgs() << "Invalid node order found!\n"; 3713 }); 3714 } 3715 3716 /// Attempt to fix the degenerate cases when the instruction serialization 3717 /// causes the register lifetimes to overlap. For example, 3718 /// p' = store_pi(p, b) 3719 /// = load p, offset 3720 /// In this case p and p' overlap, which means that two registers are needed. 3721 /// Instead, this function changes the load to use p' and updates the offset. 3722 void SwingSchedulerDAG::fixupRegisterOverlaps(std::deque<SUnit *> &Instrs) { 3723 Register OverlapReg; 3724 Register NewBaseReg; 3725 for (SUnit *SU : Instrs) { 3726 MachineInstr *MI = SU->getInstr(); 3727 for (unsigned i = 0, e = MI->getNumOperands(); i < e; ++i) { 3728 const MachineOperand &MO = MI->getOperand(i); 3729 // Look for an instruction that uses p. The instruction occurs in the 3730 // same cycle but occurs later in the serialized order. 3731 if (MO.isReg() && MO.isUse() && MO.getReg() == OverlapReg) { 3732 // Check that the instruction appears in the InstrChanges structure, 3733 // which contains instructions that can have the offset updated. 3734 DenseMap<SUnit *, std::pair<Register, int64_t>>::iterator It = 3735 InstrChanges.find(SU); 3736 if (It != InstrChanges.end()) { 3737 unsigned BasePos, OffsetPos; 3738 // Update the base register and adjust the offset. 3739 if (TII->getBaseAndOffsetPosition(*MI, BasePos, OffsetPos)) { 3740 MachineInstr *NewMI = MF.CloneMachineInstr(MI); 3741 NewMI->getOperand(BasePos).setReg(NewBaseReg); 3742 int64_t NewOffset = 3743 MI->getOperand(OffsetPos).getImm() - It->second.second; 3744 NewMI->getOperand(OffsetPos).setImm(NewOffset); 3745 SU->setInstr(NewMI); 3746 MISUnitMap[NewMI] = SU; 3747 NewMIs[MI] = NewMI; 3748 } 3749 } 3750 OverlapReg = Register(); 3751 NewBaseReg = Register(); 3752 break; 3753 } 3754 // Look for an instruction of the form p' = op(p), which uses and defines 3755 // two virtual registers that get allocated to the same physical register. 3756 unsigned TiedUseIdx = 0; 3757 if (MI->isRegTiedToUseOperand(i, &TiedUseIdx)) { 3758 // OverlapReg is p in the example above. 3759 OverlapReg = MI->getOperand(TiedUseIdx).getReg(); 3760 // NewBaseReg is p' in the example above. 3761 NewBaseReg = MI->getOperand(i).getReg(); 3762 break; 3763 } 3764 } 3765 } 3766 } 3767 3768 std::deque<SUnit *> 3769 SMSchedule::reorderInstructions(const SwingSchedulerDAG *SSD, 3770 const std::deque<SUnit *> &Instrs) const { 3771 std::deque<SUnit *> NewOrderPhi; 3772 for (SUnit *SU : Instrs) { 3773 if (SU->getInstr()->isPHI()) 3774 NewOrderPhi.push_back(SU); 3775 } 3776 std::deque<SUnit *> NewOrderI; 3777 for (SUnit *SU : Instrs) { 3778 if (!SU->getInstr()->isPHI()) 3779 orderDependence(SSD, SU, NewOrderI); 3780 } 3781 llvm::append_range(NewOrderPhi, NewOrderI); 3782 return NewOrderPhi; 3783 } 3784 3785 /// After the schedule has been formed, call this function to combine 3786 /// the instructions from the different stages/cycles. That is, this 3787 /// function creates a schedule that represents a single iteration. 3788 void SMSchedule::finalizeSchedule(SwingSchedulerDAG *SSD) { 3789 // Move all instructions to the first stage from later stages. 3790 for (int cycle = getFirstCycle(); cycle <= getFinalCycle(); ++cycle) { 3791 for (int stage = 1, lastStage = getMaxStageCount(); stage <= lastStage; 3792 ++stage) { 3793 std::deque<SUnit *> &cycleInstrs = 3794 ScheduledInstrs[cycle + (stage * InitiationInterval)]; 3795 for (SUnit *SU : llvm::reverse(cycleInstrs)) 3796 ScheduledInstrs[cycle].push_front(SU); 3797 } 3798 } 3799 3800 // Erase all the elements in the later stages. Only one iteration should 3801 // remain in the scheduled list, and it contains all the instructions. 3802 for (int cycle = getFinalCycle() + 1; cycle <= LastCycle; ++cycle) 3803 ScheduledInstrs.erase(cycle); 3804 3805 // Change the registers in instruction as specified in the InstrChanges 3806 // map. We need to use the new registers to create the correct order. 3807 for (const SUnit &SU : SSD->SUnits) 3808 SSD->applyInstrChange(SU.getInstr(), *this); 3809 3810 // Reorder the instructions in each cycle to fix and improve the 3811 // generated code. 3812 for (int Cycle = getFirstCycle(), E = getFinalCycle(); Cycle <= E; ++Cycle) { 3813 std::deque<SUnit *> &cycleInstrs = ScheduledInstrs[Cycle]; 3814 cycleInstrs = reorderInstructions(SSD, cycleInstrs); 3815 SSD->fixupRegisterOverlaps(cycleInstrs); 3816 } 3817 3818 LLVM_DEBUG(dump();); 3819 } 3820 3821 void NodeSet::print(raw_ostream &os) const { 3822 os << "Num nodes " << size() << " rec " << RecMII << " mov " << MaxMOV 3823 << " depth " << MaxDepth << " col " << Colocate << "\n"; 3824 for (const auto &I : Nodes) 3825 os << " SU(" << I->NodeNum << ") " << *(I->getInstr()); 3826 os << "\n"; 3827 } 3828 3829 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) 3830 /// Print the schedule information to the given output. 3831 void SMSchedule::print(raw_ostream &os) const { 3832 // Iterate over each cycle. 3833 for (int cycle = getFirstCycle(); cycle <= getFinalCycle(); ++cycle) { 3834 // Iterate over each instruction in the cycle. 3835 const_sched_iterator cycleInstrs = ScheduledInstrs.find(cycle); 3836 for (SUnit *CI : cycleInstrs->second) { 3837 os << "cycle " << cycle << " (" << stageScheduled(CI) << ") "; 3838 os << "(" << CI->NodeNum << ") "; 3839 CI->getInstr()->print(os); 3840 os << "\n"; 3841 } 3842 } 3843 } 3844 3845 /// Utility function used for debugging to print the schedule. 3846 LLVM_DUMP_METHOD void SMSchedule::dump() const { print(dbgs()); } 3847 LLVM_DUMP_METHOD void NodeSet::dump() const { print(dbgs()); } 3848 3849 void ResourceManager::dumpMRT() const { 3850 LLVM_DEBUG({ 3851 if (UseDFA) 3852 return; 3853 std::stringstream SS; 3854 SS << "MRT:\n"; 3855 SS << std::setw(4) << "Slot"; 3856 for (unsigned I = 1, E = SM.getNumProcResourceKinds(); I < E; ++I) 3857 SS << std::setw(3) << I; 3858 SS << std::setw(7) << "#Mops" 3859 << "\n"; 3860 for (int Slot = 0; Slot < InitiationInterval; ++Slot) { 3861 SS << std::setw(4) << Slot; 3862 for (unsigned I = 1, E = SM.getNumProcResourceKinds(); I < E; ++I) 3863 SS << std::setw(3) << MRT[Slot][I]; 3864 SS << std::setw(7) << NumScheduledMops[Slot] << "\n"; 3865 } 3866 dbgs() << SS.str(); 3867 }); 3868 } 3869 #endif 3870 3871 void ResourceManager::initProcResourceVectors( 3872 const MCSchedModel &SM, SmallVectorImpl<uint64_t> &Masks) { 3873 unsigned ProcResourceID = 0; 3874 3875 // We currently limit the resource kinds to 64 and below so that we can use 3876 // uint64_t for Masks 3877 assert(SM.getNumProcResourceKinds() < 64 && 3878 "Too many kinds of resources, unsupported"); 3879 // Create a unique bitmask for every processor resource unit. 3880 // Skip resource at index 0, since it always references 'InvalidUnit'. 3881 Masks.resize(SM.getNumProcResourceKinds()); 3882 for (unsigned I = 1, E = SM.getNumProcResourceKinds(); I < E; ++I) { 3883 const MCProcResourceDesc &Desc = *SM.getProcResource(I); 3884 if (Desc.SubUnitsIdxBegin) 3885 continue; 3886 Masks[I] = 1ULL << ProcResourceID; 3887 ProcResourceID++; 3888 } 3889 // Create a unique bitmask for every processor resource group. 3890 for (unsigned I = 1, E = SM.getNumProcResourceKinds(); I < E; ++I) { 3891 const MCProcResourceDesc &Desc = *SM.getProcResource(I); 3892 if (!Desc.SubUnitsIdxBegin) 3893 continue; 3894 Masks[I] = 1ULL << ProcResourceID; 3895 for (unsigned U = 0; U < Desc.NumUnits; ++U) 3896 Masks[I] |= Masks[Desc.SubUnitsIdxBegin[U]]; 3897 ProcResourceID++; 3898 } 3899 LLVM_DEBUG({ 3900 if (SwpShowResMask) { 3901 dbgs() << "ProcResourceDesc:\n"; 3902 for (unsigned I = 1, E = SM.getNumProcResourceKinds(); I < E; ++I) { 3903 const MCProcResourceDesc *ProcResource = SM.getProcResource(I); 3904 dbgs() << format(" %16s(%2d): Mask: 0x%08x, NumUnits:%2d\n", 3905 ProcResource->Name, I, Masks[I], 3906 ProcResource->NumUnits); 3907 } 3908 dbgs() << " -----------------\n"; 3909 } 3910 }); 3911 } 3912 3913 bool ResourceManager::canReserveResources(SUnit &SU, int Cycle) { 3914 LLVM_DEBUG({ 3915 if (SwpDebugResource) 3916 dbgs() << "canReserveResources:\n"; 3917 }); 3918 if (UseDFA) 3919 return DFAResources[positiveModulo(Cycle, InitiationInterval)] 3920 ->canReserveResources(&SU.getInstr()->getDesc()); 3921 3922 const MCSchedClassDesc *SCDesc = DAG->getSchedClass(&SU); 3923 if (!SCDesc->isValid()) { 3924 LLVM_DEBUG({ 3925 dbgs() << "No valid Schedule Class Desc for schedClass!\n"; 3926 dbgs() << "isPseudo:" << SU.getInstr()->isPseudo() << "\n"; 3927 }); 3928 return true; 3929 } 3930 3931 reserveResources(SCDesc, Cycle); 3932 bool Result = !isOverbooked(); 3933 unreserveResources(SCDesc, Cycle); 3934 3935 LLVM_DEBUG(if (SwpDebugResource) dbgs() << "return " << Result << "\n\n"); 3936 return Result; 3937 } 3938 3939 void ResourceManager::reserveResources(SUnit &SU, int Cycle) { 3940 LLVM_DEBUG({ 3941 if (SwpDebugResource) 3942 dbgs() << "reserveResources:\n"; 3943 }); 3944 if (UseDFA) 3945 return DFAResources[positiveModulo(Cycle, InitiationInterval)] 3946 ->reserveResources(&SU.getInstr()->getDesc()); 3947 3948 const MCSchedClassDesc *SCDesc = DAG->getSchedClass(&SU); 3949 if (!SCDesc->isValid()) { 3950 LLVM_DEBUG({ 3951 dbgs() << "No valid Schedule Class Desc for schedClass!\n"; 3952 dbgs() << "isPseudo:" << SU.getInstr()->isPseudo() << "\n"; 3953 }); 3954 return; 3955 } 3956 3957 reserveResources(SCDesc, Cycle); 3958 3959 LLVM_DEBUG({ 3960 if (SwpDebugResource) { 3961 dumpMRT(); 3962 dbgs() << "reserveResources: done!\n\n"; 3963 } 3964 }); 3965 } 3966 3967 void ResourceManager::reserveResources(const MCSchedClassDesc *SCDesc, 3968 int Cycle) { 3969 assert(!UseDFA); 3970 for (const MCWriteProcResEntry &PRE : make_range( 3971 STI->getWriteProcResBegin(SCDesc), STI->getWriteProcResEnd(SCDesc))) 3972 for (int C = Cycle; C < Cycle + PRE.ReleaseAtCycle; ++C) 3973 ++MRT[positiveModulo(C, InitiationInterval)][PRE.ProcResourceIdx]; 3974 3975 for (int C = Cycle; C < Cycle + SCDesc->NumMicroOps; ++C) 3976 ++NumScheduledMops[positiveModulo(C, InitiationInterval)]; 3977 } 3978 3979 void ResourceManager::unreserveResources(const MCSchedClassDesc *SCDesc, 3980 int Cycle) { 3981 assert(!UseDFA); 3982 for (const MCWriteProcResEntry &PRE : make_range( 3983 STI->getWriteProcResBegin(SCDesc), STI->getWriteProcResEnd(SCDesc))) 3984 for (int C = Cycle; C < Cycle + PRE.ReleaseAtCycle; ++C) 3985 --MRT[positiveModulo(C, InitiationInterval)][PRE.ProcResourceIdx]; 3986 3987 for (int C = Cycle; C < Cycle + SCDesc->NumMicroOps; ++C) 3988 --NumScheduledMops[positiveModulo(C, InitiationInterval)]; 3989 } 3990 3991 bool ResourceManager::isOverbooked() const { 3992 assert(!UseDFA); 3993 for (int Slot = 0; Slot < InitiationInterval; ++Slot) { 3994 for (unsigned I = 1, E = SM.getNumProcResourceKinds(); I < E; ++I) { 3995 const MCProcResourceDesc *Desc = SM.getProcResource(I); 3996 if (MRT[Slot][I] > Desc->NumUnits) 3997 return true; 3998 } 3999 if (NumScheduledMops[Slot] > IssueWidth) 4000 return true; 4001 } 4002 return false; 4003 } 4004 4005 int ResourceManager::calculateResMIIDFA() const { 4006 assert(UseDFA); 4007 4008 // Sort the instructions by the number of available choices for scheduling, 4009 // least to most. Use the number of critical resources as the tie breaker. 4010 FuncUnitSorter FUS = FuncUnitSorter(*ST); 4011 for (SUnit &SU : DAG->SUnits) 4012 FUS.calcCriticalResources(*SU.getInstr()); 4013 PriorityQueue<MachineInstr *, std::vector<MachineInstr *>, FuncUnitSorter> 4014 FuncUnitOrder(FUS); 4015 4016 for (SUnit &SU : DAG->SUnits) 4017 FuncUnitOrder.push(SU.getInstr()); 4018 4019 SmallVector<std::unique_ptr<DFAPacketizer>, 8> Resources; 4020 Resources.push_back( 4021 std::unique_ptr<DFAPacketizer>(TII->CreateTargetScheduleState(*ST))); 4022 4023 while (!FuncUnitOrder.empty()) { 4024 MachineInstr *MI = FuncUnitOrder.top(); 4025 FuncUnitOrder.pop(); 4026 if (TII->isZeroCost(MI->getOpcode())) 4027 continue; 4028 4029 // Attempt to reserve the instruction in an existing DFA. At least one 4030 // DFA is needed for each cycle. 4031 unsigned NumCycles = DAG->getSUnit(MI)->Latency; 4032 unsigned ReservedCycles = 0; 4033 auto *RI = Resources.begin(); 4034 auto *RE = Resources.end(); 4035 LLVM_DEBUG({ 4036 dbgs() << "Trying to reserve resource for " << NumCycles 4037 << " cycles for \n"; 4038 MI->dump(); 4039 }); 4040 for (unsigned C = 0; C < NumCycles; ++C) 4041 while (RI != RE) { 4042 if ((*RI)->canReserveResources(*MI)) { 4043 (*RI)->reserveResources(*MI); 4044 ++ReservedCycles; 4045 break; 4046 } 4047 RI++; 4048 } 4049 LLVM_DEBUG(dbgs() << "ReservedCycles:" << ReservedCycles 4050 << ", NumCycles:" << NumCycles << "\n"); 4051 // Add new DFAs, if needed, to reserve resources. 4052 for (unsigned C = ReservedCycles; C < NumCycles; ++C) { 4053 LLVM_DEBUG(if (SwpDebugResource) dbgs() 4054 << "NewResource created to reserve resources" 4055 << "\n"); 4056 auto *NewResource = TII->CreateTargetScheduleState(*ST); 4057 assert(NewResource->canReserveResources(*MI) && "Reserve error."); 4058 NewResource->reserveResources(*MI); 4059 Resources.push_back(std::unique_ptr<DFAPacketizer>(NewResource)); 4060 } 4061 } 4062 4063 int Resmii = Resources.size(); 4064 LLVM_DEBUG(dbgs() << "Return Res MII:" << Resmii << "\n"); 4065 return Resmii; 4066 } 4067 4068 int ResourceManager::calculateResMII() const { 4069 if (UseDFA) 4070 return calculateResMIIDFA(); 4071 4072 // Count each resource consumption and divide it by the number of units. 4073 // ResMII is the max value among them. 4074 4075 int NumMops = 0; 4076 SmallVector<uint64_t> ResourceCount(SM.getNumProcResourceKinds()); 4077 for (SUnit &SU : DAG->SUnits) { 4078 if (TII->isZeroCost(SU.getInstr()->getOpcode())) 4079 continue; 4080 4081 const MCSchedClassDesc *SCDesc = DAG->getSchedClass(&SU); 4082 if (!SCDesc->isValid()) 4083 continue; 4084 4085 LLVM_DEBUG({ 4086 if (SwpDebugResource) { 4087 DAG->dumpNode(SU); 4088 dbgs() << " #Mops: " << SCDesc->NumMicroOps << "\n" 4089 << " WriteProcRes: "; 4090 } 4091 }); 4092 NumMops += SCDesc->NumMicroOps; 4093 for (const MCWriteProcResEntry &PRE : 4094 make_range(STI->getWriteProcResBegin(SCDesc), 4095 STI->getWriteProcResEnd(SCDesc))) { 4096 LLVM_DEBUG({ 4097 if (SwpDebugResource) { 4098 const MCProcResourceDesc *Desc = 4099 SM.getProcResource(PRE.ProcResourceIdx); 4100 dbgs() << Desc->Name << ": " << PRE.ReleaseAtCycle << ", "; 4101 } 4102 }); 4103 ResourceCount[PRE.ProcResourceIdx] += PRE.ReleaseAtCycle; 4104 } 4105 LLVM_DEBUG(if (SwpDebugResource) dbgs() << "\n"); 4106 } 4107 4108 int Result = (NumMops + IssueWidth - 1) / IssueWidth; 4109 LLVM_DEBUG({ 4110 if (SwpDebugResource) 4111 dbgs() << "#Mops: " << NumMops << ", " 4112 << "IssueWidth: " << IssueWidth << ", " 4113 << "Cycles: " << Result << "\n"; 4114 }); 4115 4116 LLVM_DEBUG({ 4117 if (SwpDebugResource) { 4118 std::stringstream SS; 4119 SS << std::setw(2) << "ID" << std::setw(16) << "Name" << std::setw(10) 4120 << "Units" << std::setw(10) << "Consumed" << std::setw(10) << "Cycles" 4121 << "\n"; 4122 dbgs() << SS.str(); 4123 } 4124 }); 4125 for (unsigned I = 1, E = SM.getNumProcResourceKinds(); I < E; ++I) { 4126 const MCProcResourceDesc *Desc = SM.getProcResource(I); 4127 int Cycles = (ResourceCount[I] + Desc->NumUnits - 1) / Desc->NumUnits; 4128 LLVM_DEBUG({ 4129 if (SwpDebugResource) { 4130 std::stringstream SS; 4131 SS << std::setw(2) << I << std::setw(16) << Desc->Name << std::setw(10) 4132 << Desc->NumUnits << std::setw(10) << ResourceCount[I] 4133 << std::setw(10) << Cycles << "\n"; 4134 dbgs() << SS.str(); 4135 } 4136 }); 4137 if (Cycles > Result) 4138 Result = Cycles; 4139 } 4140 return Result; 4141 } 4142 4143 void ResourceManager::init(int II) { 4144 InitiationInterval = II; 4145 DFAResources.clear(); 4146 DFAResources.resize(II); 4147 for (auto &I : DFAResources) 4148 I.reset(ST->getInstrInfo()->CreateTargetScheduleState(*ST)); 4149 MRT.clear(); 4150 MRT.resize(II, SmallVector<uint64_t>(SM.getNumProcResourceKinds())); 4151 NumScheduledMops.clear(); 4152 NumScheduledMops.resize(II); 4153 } 4154 4155 bool SwingSchedulerDDGEdge::ignoreDependence(bool IgnoreAnti) const { 4156 if (Pred.isArtificial() || Dst->isBoundaryNode()) 4157 return true; 4158 // Currently, dependence that is an anti-dependences but not a loop-carried is 4159 // also ignored. This behavior is preserved to prevent regression. 4160 // FIXME: Remove if this doesn't have significant impact on performance 4161 return IgnoreAnti && (Pred.getKind() == SDep::Kind::Anti || Distance != 0); 4162 } 4163 4164 SwingSchedulerDDG::SwingSchedulerDDGEdges & 4165 SwingSchedulerDDG::getEdges(const SUnit *SU) { 4166 if (SU == EntrySU) 4167 return EntrySUEdges; 4168 if (SU == ExitSU) 4169 return ExitSUEdges; 4170 return EdgesVec[SU->NodeNum]; 4171 } 4172 4173 const SwingSchedulerDDG::SwingSchedulerDDGEdges & 4174 SwingSchedulerDDG::getEdges(const SUnit *SU) const { 4175 if (SU == EntrySU) 4176 return EntrySUEdges; 4177 if (SU == ExitSU) 4178 return ExitSUEdges; 4179 return EdgesVec[SU->NodeNum]; 4180 } 4181 4182 void SwingSchedulerDDG::addEdge(const SUnit *SU, 4183 const SwingSchedulerDDGEdge &Edge) { 4184 assert(!Edge.isValidationOnly() && 4185 "Validation-only edges are not expected here."); 4186 auto &Edges = getEdges(SU); 4187 if (Edge.getSrc() == SU) 4188 Edges.Succs.push_back(Edge); 4189 else 4190 Edges.Preds.push_back(Edge); 4191 } 4192 4193 void SwingSchedulerDDG::initEdges(SUnit *SU) { 4194 for (const auto &PI : SU->Preds) { 4195 SwingSchedulerDDGEdge Edge(SU, PI, /*IsSucc=*/false, 4196 /*IsValidationOnly=*/false); 4197 addEdge(SU, Edge); 4198 } 4199 4200 for (const auto &SI : SU->Succs) { 4201 SwingSchedulerDDGEdge Edge(SU, SI, /*IsSucc=*/true, 4202 /*IsValidationOnly=*/false); 4203 addEdge(SU, Edge); 4204 } 4205 } 4206 4207 SwingSchedulerDDG::SwingSchedulerDDG(std::vector<SUnit> &SUnits, SUnit *EntrySU, 4208 SUnit *ExitSU, const LoopCarriedEdges &LCE) 4209 : EntrySU(EntrySU), ExitSU(ExitSU) { 4210 EdgesVec.resize(SUnits.size()); 4211 4212 // Add non-loop-carried edges based on the DAG. 4213 initEdges(EntrySU); 4214 initEdges(ExitSU); 4215 for (auto &SU : SUnits) 4216 initEdges(&SU); 4217 4218 // Add loop-carried edges, which are not represented in the DAG. 4219 for (SUnit &SU : SUnits) { 4220 SUnit *Src = &SU; 4221 if (const LoopCarriedEdges::OrderDep *OD = LCE.getOrderDepOrNull(Src)) { 4222 SDep Base(Src, SDep::Barrier); 4223 Base.setLatency(1); 4224 for (SUnit *Dst : *OD) { 4225 SwingSchedulerDDGEdge Edge(Dst, Base, /*IsSucc=*/false, 4226 /*IsValidationOnly=*/true); 4227 Edge.setDistance(1); 4228 ValidationOnlyEdges.push_back(Edge); 4229 } 4230 } 4231 } 4232 } 4233 4234 const SwingSchedulerDDG::EdgesType & 4235 SwingSchedulerDDG::getInEdges(const SUnit *SU) const { 4236 return getEdges(SU).Preds; 4237 } 4238 4239 const SwingSchedulerDDG::EdgesType & 4240 SwingSchedulerDDG::getOutEdges(const SUnit *SU) const { 4241 return getEdges(SU).Succs; 4242 } 4243 4244 /// Check if \p Schedule doesn't violate the validation-only dependencies. 4245 bool SwingSchedulerDDG::isValidSchedule(const SMSchedule &Schedule) const { 4246 unsigned II = Schedule.getInitiationInterval(); 4247 4248 auto ExpandCycle = [&](SUnit *SU) { 4249 int Stage = Schedule.stageScheduled(SU); 4250 int Cycle = Schedule.cycleScheduled(SU); 4251 return Cycle + (Stage * II); 4252 }; 4253 4254 for (const SwingSchedulerDDGEdge &Edge : ValidationOnlyEdges) { 4255 SUnit *Src = Edge.getSrc(); 4256 SUnit *Dst = Edge.getDst(); 4257 if (!Src->isInstr() || !Dst->isInstr()) 4258 continue; 4259 int CycleSrc = ExpandCycle(Src); 4260 int CycleDst = ExpandCycle(Dst); 4261 int MaxLateStart = CycleDst + Edge.getDistance() * II - Edge.getLatency(); 4262 if (CycleSrc > MaxLateStart) { 4263 LLVM_DEBUG({ 4264 dbgs() << "Validation failed for edge from " << Src->NodeNum << " to " 4265 << Dst->NodeNum << "\n"; 4266 }); 4267 return false; 4268 } 4269 } 4270 return true; 4271 } 4272 4273 void LoopCarriedEdges::modifySUnits(std::vector<SUnit> &SUnits, 4274 const TargetInstrInfo *TII) { 4275 for (SUnit &SU : SUnits) { 4276 SUnit *Src = &SU; 4277 if (auto *OrderDep = getOrderDepOrNull(Src)) { 4278 SDep Dep(Src, SDep::Barrier); 4279 Dep.setLatency(1); 4280 for (SUnit *Dst : *OrderDep) { 4281 SUnit *From = Src; 4282 SUnit *To = Dst; 4283 if (From->NodeNum > To->NodeNum) 4284 std::swap(From, To); 4285 4286 // Add a forward edge if the following conditions are met: 4287 // 4288 // - The instruction of the source node (FromMI) may read memory. 4289 // - The instruction of the target node (ToMI) may modify memory, but 4290 // does not read it. 4291 // - Neither instruction is a global barrier. 4292 // - The load appears before the store in the original basic block. 4293 // - There are no barrier or store instructions between the two nodes. 4294 // - The target node is unreachable from the source node in the current 4295 // DAG. 4296 // 4297 // TODO: These conditions are inherited from a previous implementation, 4298 // and some may no longer be necessary. For now, we conservatively 4299 // retain all of them to avoid regressions, but the logic could 4300 // potentially be simplified 4301 MachineInstr *FromMI = From->getInstr(); 4302 MachineInstr *ToMI = To->getInstr(); 4303 if (FromMI->mayLoad() && !ToMI->mayLoad() && ToMI->mayStore() && 4304 !TII->isGlobalMemoryObject(FromMI) && 4305 !TII->isGlobalMemoryObject(ToMI) && !isSuccOrder(From, To)) { 4306 SDep Pred = Dep; 4307 Pred.setSUnit(From); 4308 To->addPred(Pred); 4309 } 4310 } 4311 } 4312 } 4313 } 4314 4315 void LoopCarriedEdges::dump(SUnit *SU, const TargetRegisterInfo *TRI, 4316 const MachineRegisterInfo *MRI) const { 4317 const auto *Order = getOrderDepOrNull(SU); 4318 4319 if (!Order) 4320 return; 4321 4322 const auto DumpSU = [](const SUnit *SU) { 4323 std::ostringstream OSS; 4324 OSS << "SU(" << SU->NodeNum << ")"; 4325 return OSS.str(); 4326 }; 4327 4328 dbgs() << " Loop carried edges from " << DumpSU(SU) << "\n" 4329 << " Order\n"; 4330 for (SUnit *Dst : *Order) 4331 dbgs() << " " << DumpSU(Dst) << "\n"; 4332 } 4333