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/MapVector.h" 37 #include "llvm/ADT/PriorityQueue.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/CycleAnalysis.h" 47 #include "llvm/Analysis/MemoryLocation.h" 48 #include "llvm/Analysis/OptimizationRemarkEmitter.h" 49 #include "llvm/Analysis/ValueTracking.h" 50 #include "llvm/CodeGen/DFAPacketizer.h" 51 #include "llvm/CodeGen/LiveIntervals.h" 52 #include "llvm/CodeGen/MachineBasicBlock.h" 53 #include "llvm/CodeGen/MachineDominators.h" 54 #include "llvm/CodeGen/MachineFunction.h" 55 #include "llvm/CodeGen/MachineFunctionPass.h" 56 #include "llvm/CodeGen/MachineInstr.h" 57 #include "llvm/CodeGen/MachineInstrBuilder.h" 58 #include "llvm/CodeGen/MachineLoopInfo.h" 59 #include "llvm/CodeGen/MachineMemOperand.h" 60 #include "llvm/CodeGen/MachineOperand.h" 61 #include "llvm/CodeGen/MachineRegisterInfo.h" 62 #include "llvm/CodeGen/ModuloSchedule.h" 63 #include "llvm/CodeGen/RegisterPressure.h" 64 #include "llvm/CodeGen/ScheduleDAG.h" 65 #include "llvm/CodeGen/ScheduleDAGMutation.h" 66 #include "llvm/CodeGen/TargetOpcodes.h" 67 #include "llvm/CodeGen/TargetRegisterInfo.h" 68 #include "llvm/CodeGen/TargetSubtargetInfo.h" 69 #include "llvm/Config/llvm-config.h" 70 #include "llvm/IR/Attributes.h" 71 #include "llvm/IR/Function.h" 72 #include "llvm/MC/LaneBitmask.h" 73 #include "llvm/MC/MCInstrDesc.h" 74 #include "llvm/MC/MCInstrItineraries.h" 75 #include "llvm/MC/MCRegisterInfo.h" 76 #include "llvm/Pass.h" 77 #include "llvm/Support/CommandLine.h" 78 #include "llvm/Support/Compiler.h" 79 #include "llvm/Support/Debug.h" 80 #include "llvm/Support/MathExtras.h" 81 #include "llvm/Support/raw_ostream.h" 82 #include <algorithm> 83 #include <cassert> 84 #include <climits> 85 #include <cstdint> 86 #include <deque> 87 #include <functional> 88 #include <iomanip> 89 #include <iterator> 90 #include <map> 91 #include <memory> 92 #include <sstream> 93 #include <tuple> 94 #include <utility> 95 #include <vector> 96 97 using namespace llvm; 98 99 #define DEBUG_TYPE "pipeliner" 100 101 STATISTIC(NumTrytoPipeline, "Number of loops that we attempt to pipeline"); 102 STATISTIC(NumPipelined, "Number of loops software pipelined"); 103 STATISTIC(NumNodeOrderIssues, "Number of node order issues found"); 104 STATISTIC(NumFailBranch, "Pipeliner abort due to unknown branch"); 105 STATISTIC(NumFailLoop, "Pipeliner abort due to unsupported loop"); 106 STATISTIC(NumFailPreheader, "Pipeliner abort due to missing preheader"); 107 STATISTIC(NumFailLargeMaxMII, "Pipeliner abort due to MaxMII too large"); 108 STATISTIC(NumFailZeroMII, "Pipeliner abort due to zero MII"); 109 STATISTIC(NumFailNoSchedule, "Pipeliner abort due to no schedule found"); 110 STATISTIC(NumFailZeroStage, "Pipeliner abort due to zero stage"); 111 STATISTIC(NumFailLargeMaxStage, "Pipeliner abort due to too many stages"); 112 113 /// A command line option to turn software pipelining on or off. 114 static cl::opt<bool> EnableSWP("enable-pipeliner", cl::Hidden, cl::init(true), 115 cl::desc("Enable Software Pipelining")); 116 117 /// A command line option to enable SWP at -Os. 118 static cl::opt<bool> EnableSWPOptSize("enable-pipeliner-opt-size", 119 cl::desc("Enable SWP at Os."), cl::Hidden, 120 cl::init(false)); 121 122 /// A command line argument to limit minimum initial interval for pipelining. 123 static cl::opt<int> SwpMaxMii("pipeliner-max-mii", 124 cl::desc("Size limit for the MII."), 125 cl::Hidden, cl::init(27)); 126 127 /// A command line argument to force pipeliner to use specified initial 128 /// interval. 129 static cl::opt<int> SwpForceII("pipeliner-force-ii", 130 cl::desc("Force pipeliner to use specified II."), 131 cl::Hidden, cl::init(-1)); 132 133 /// A command line argument to limit the number of stages in the pipeline. 134 static cl::opt<int> 135 SwpMaxStages("pipeliner-max-stages", 136 cl::desc("Maximum stages allowed in the generated scheduled."), 137 cl::Hidden, cl::init(3)); 138 139 /// A command line option to disable the pruning of chain dependences due to 140 /// an unrelated Phi. 141 static cl::opt<bool> 142 SwpPruneDeps("pipeliner-prune-deps", 143 cl::desc("Prune dependences between unrelated Phi nodes."), 144 cl::Hidden, cl::init(true)); 145 146 /// A command line option to disable the pruning of loop carried order 147 /// dependences. 148 static cl::opt<bool> 149 SwpPruneLoopCarried("pipeliner-prune-loop-carried", 150 cl::desc("Prune loop carried order dependences."), 151 cl::Hidden, cl::init(true)); 152 153 #ifndef NDEBUG 154 static cl::opt<int> SwpLoopLimit("pipeliner-max", cl::Hidden, cl::init(-1)); 155 #endif 156 157 static cl::opt<bool> SwpIgnoreRecMII("pipeliner-ignore-recmii", 158 cl::ReallyHidden, 159 cl::desc("Ignore RecMII")); 160 161 static cl::opt<bool> SwpShowResMask("pipeliner-show-mask", cl::Hidden, 162 cl::init(false)); 163 static cl::opt<bool> SwpDebugResource("pipeliner-dbg-res", cl::Hidden, 164 cl::init(false)); 165 166 static cl::opt<bool> EmitTestAnnotations( 167 "pipeliner-annotate-for-testing", cl::Hidden, cl::init(false), 168 cl::desc("Instead of emitting the pipelined code, annotate instructions " 169 "with the generated schedule for feeding into the " 170 "-modulo-schedule-test pass")); 171 172 static cl::opt<bool> ExperimentalCodeGen( 173 "pipeliner-experimental-cg", cl::Hidden, cl::init(false), 174 cl::desc( 175 "Use the experimental peeling code generator for software pipelining")); 176 177 namespace llvm { 178 179 // A command line option to enable the CopyToPhi DAG mutation. 180 cl::opt<bool> SwpEnableCopyToPhi("pipeliner-enable-copytophi", cl::ReallyHidden, 181 cl::init(true), 182 cl::desc("Enable CopyToPhi DAG Mutation")); 183 184 /// A command line argument to force pipeliner to use specified issue 185 /// width. 186 cl::opt<int> SwpForceIssueWidth( 187 "pipeliner-force-issue-width", 188 cl::desc("Force pipeliner to use specified issue width."), cl::Hidden, 189 cl::init(-1)); 190 191 } // end namespace llvm 192 193 unsigned SwingSchedulerDAG::Circuits::MaxPaths = 5; 194 char MachinePipeliner::ID = 0; 195 #ifndef NDEBUG 196 int MachinePipeliner::NumTries = 0; 197 #endif 198 char &llvm::MachinePipelinerID = MachinePipeliner::ID; 199 200 INITIALIZE_PASS_BEGIN(MachinePipeliner, DEBUG_TYPE, 201 "Modulo Software Pipelining", false, false) 202 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass) 203 INITIALIZE_PASS_DEPENDENCY(MachineLoopInfo) 204 INITIALIZE_PASS_DEPENDENCY(MachineDominatorTree) 205 INITIALIZE_PASS_DEPENDENCY(LiveIntervals) 206 INITIALIZE_PASS_END(MachinePipeliner, DEBUG_TYPE, 207 "Modulo Software Pipelining", false, false) 208 209 /// The "main" function for implementing Swing Modulo Scheduling. 210 bool MachinePipeliner::runOnMachineFunction(MachineFunction &mf) { 211 if (skipFunction(mf.getFunction())) 212 return false; 213 214 if (!EnableSWP) 215 return false; 216 217 if (mf.getFunction().getAttributes().hasFnAttr(Attribute::OptimizeForSize) && 218 !EnableSWPOptSize.getPosition()) 219 return false; 220 221 if (!mf.getSubtarget().enableMachinePipeliner()) 222 return false; 223 224 // Cannot pipeline loops without instruction itineraries if we are using 225 // DFA for the pipeliner. 226 if (mf.getSubtarget().useDFAforSMS() && 227 (!mf.getSubtarget().getInstrItineraryData() || 228 mf.getSubtarget().getInstrItineraryData()->isEmpty())) 229 return false; 230 231 MF = &mf; 232 MLI = &getAnalysis<MachineLoopInfo>(); 233 MDT = &getAnalysis<MachineDominatorTree>(); 234 ORE = &getAnalysis<MachineOptimizationRemarkEmitterPass>().getORE(); 235 TII = MF->getSubtarget().getInstrInfo(); 236 RegClassInfo.runOnMachineFunction(*MF); 237 238 for (const auto &L : *MLI) 239 scheduleLoop(*L); 240 241 return false; 242 } 243 244 /// Attempt to perform the SMS algorithm on the specified loop. This function is 245 /// the main entry point for the algorithm. The function identifies candidate 246 /// loops, calculates the minimum initiation interval, and attempts to schedule 247 /// the loop. 248 bool MachinePipeliner::scheduleLoop(MachineLoop &L) { 249 bool Changed = false; 250 for (const auto &InnerLoop : L) 251 Changed |= scheduleLoop(*InnerLoop); 252 253 #ifndef NDEBUG 254 // Stop trying after reaching the limit (if any). 255 int Limit = SwpLoopLimit; 256 if (Limit >= 0) { 257 if (NumTries >= SwpLoopLimit) 258 return Changed; 259 NumTries++; 260 } 261 #endif 262 263 setPragmaPipelineOptions(L); 264 if (!canPipelineLoop(L)) { 265 LLVM_DEBUG(dbgs() << "\n!!! Can not pipeline loop.\n"); 266 ORE->emit([&]() { 267 return MachineOptimizationRemarkMissed(DEBUG_TYPE, "canPipelineLoop", 268 L.getStartLoc(), L.getHeader()) 269 << "Failed to pipeline loop"; 270 }); 271 272 LI.LoopPipelinerInfo.reset(); 273 return Changed; 274 } 275 276 ++NumTrytoPipeline; 277 278 Changed = swingModuloScheduler(L); 279 280 LI.LoopPipelinerInfo.reset(); 281 return Changed; 282 } 283 284 void MachinePipeliner::setPragmaPipelineOptions(MachineLoop &L) { 285 // Reset the pragma for the next loop in iteration. 286 disabledByPragma = false; 287 II_setByPragma = 0; 288 289 MachineBasicBlock *LBLK = L.getTopBlock(); 290 291 if (LBLK == nullptr) 292 return; 293 294 const BasicBlock *BBLK = LBLK->getBasicBlock(); 295 if (BBLK == nullptr) 296 return; 297 298 const Instruction *TI = BBLK->getTerminator(); 299 if (TI == nullptr) 300 return; 301 302 MDNode *LoopID = TI->getMetadata(LLVMContext::MD_loop); 303 if (LoopID == nullptr) 304 return; 305 306 assert(LoopID->getNumOperands() > 0 && "requires atleast one operand"); 307 assert(LoopID->getOperand(0) == LoopID && "invalid loop"); 308 309 for (unsigned i = 1, e = LoopID->getNumOperands(); i < e; ++i) { 310 MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i)); 311 312 if (MD == nullptr) 313 continue; 314 315 MDString *S = dyn_cast<MDString>(MD->getOperand(0)); 316 317 if (S == nullptr) 318 continue; 319 320 if (S->getString() == "llvm.loop.pipeline.initiationinterval") { 321 assert(MD->getNumOperands() == 2 && 322 "Pipeline initiation interval hint metadata should have two operands."); 323 II_setByPragma = 324 mdconst::extract<ConstantInt>(MD->getOperand(1))->getZExtValue(); 325 assert(II_setByPragma >= 1 && "Pipeline initiation interval must be positive."); 326 } else if (S->getString() == "llvm.loop.pipeline.disable") { 327 disabledByPragma = true; 328 } 329 } 330 } 331 332 /// Return true if the loop can be software pipelined. The algorithm is 333 /// restricted to loops with a single basic block. Make sure that the 334 /// branch in the loop can be analyzed. 335 bool MachinePipeliner::canPipelineLoop(MachineLoop &L) { 336 if (L.getNumBlocks() != 1) { 337 ORE->emit([&]() { 338 return MachineOptimizationRemarkAnalysis(DEBUG_TYPE, "canPipelineLoop", 339 L.getStartLoc(), L.getHeader()) 340 << "Not a single basic block: " 341 << ore::NV("NumBlocks", L.getNumBlocks()); 342 }); 343 return false; 344 } 345 346 if (disabledByPragma) { 347 ORE->emit([&]() { 348 return MachineOptimizationRemarkAnalysis(DEBUG_TYPE, "canPipelineLoop", 349 L.getStartLoc(), L.getHeader()) 350 << "Disabled by Pragma."; 351 }); 352 return false; 353 } 354 355 // Check if the branch can't be understood because we can't do pipelining 356 // if that's the case. 357 LI.TBB = nullptr; 358 LI.FBB = nullptr; 359 LI.BrCond.clear(); 360 if (TII->analyzeBranch(*L.getHeader(), LI.TBB, LI.FBB, LI.BrCond)) { 361 LLVM_DEBUG(dbgs() << "Unable to analyzeBranch, can NOT pipeline Loop\n"); 362 NumFailBranch++; 363 ORE->emit([&]() { 364 return MachineOptimizationRemarkAnalysis(DEBUG_TYPE, "canPipelineLoop", 365 L.getStartLoc(), L.getHeader()) 366 << "The branch can't be understood"; 367 }); 368 return false; 369 } 370 371 LI.LoopInductionVar = nullptr; 372 LI.LoopCompare = nullptr; 373 LI.LoopPipelinerInfo = TII->analyzeLoopForPipelining(L.getTopBlock()); 374 if (!LI.LoopPipelinerInfo) { 375 LLVM_DEBUG(dbgs() << "Unable to analyzeLoop, can NOT pipeline Loop\n"); 376 NumFailLoop++; 377 ORE->emit([&]() { 378 return MachineOptimizationRemarkAnalysis(DEBUG_TYPE, "canPipelineLoop", 379 L.getStartLoc(), L.getHeader()) 380 << "The loop structure is not supported"; 381 }); 382 return false; 383 } 384 385 if (!L.getLoopPreheader()) { 386 LLVM_DEBUG(dbgs() << "Preheader not found, can NOT pipeline Loop\n"); 387 NumFailPreheader++; 388 ORE->emit([&]() { 389 return MachineOptimizationRemarkAnalysis(DEBUG_TYPE, "canPipelineLoop", 390 L.getStartLoc(), L.getHeader()) 391 << "No loop preheader found"; 392 }); 393 return false; 394 } 395 396 // Remove any subregisters from inputs to phi nodes. 397 preprocessPhiNodes(*L.getHeader()); 398 return true; 399 } 400 401 void MachinePipeliner::preprocessPhiNodes(MachineBasicBlock &B) { 402 MachineRegisterInfo &MRI = MF->getRegInfo(); 403 SlotIndexes &Slots = *getAnalysis<LiveIntervals>().getSlotIndexes(); 404 405 for (MachineInstr &PI : B.phis()) { 406 MachineOperand &DefOp = PI.getOperand(0); 407 assert(DefOp.getSubReg() == 0); 408 auto *RC = MRI.getRegClass(DefOp.getReg()); 409 410 for (unsigned i = 1, n = PI.getNumOperands(); i != n; i += 2) { 411 MachineOperand &RegOp = PI.getOperand(i); 412 if (RegOp.getSubReg() == 0) 413 continue; 414 415 // If the operand uses a subregister, replace it with a new register 416 // without subregisters, and generate a copy to the new register. 417 Register NewReg = MRI.createVirtualRegister(RC); 418 MachineBasicBlock &PredB = *PI.getOperand(i+1).getMBB(); 419 MachineBasicBlock::iterator At = PredB.getFirstTerminator(); 420 const DebugLoc &DL = PredB.findDebugLoc(At); 421 auto Copy = BuildMI(PredB, At, DL, TII->get(TargetOpcode::COPY), NewReg) 422 .addReg(RegOp.getReg(), getRegState(RegOp), 423 RegOp.getSubReg()); 424 Slots.insertMachineInstrInMaps(*Copy); 425 RegOp.setReg(NewReg); 426 RegOp.setSubReg(0); 427 } 428 } 429 } 430 431 /// The SMS algorithm consists of the following main steps: 432 /// 1. Computation and analysis of the dependence graph. 433 /// 2. Ordering of the nodes (instructions). 434 /// 3. Attempt to Schedule the loop. 435 bool MachinePipeliner::swingModuloScheduler(MachineLoop &L) { 436 assert(L.getBlocks().size() == 1 && "SMS works on single blocks only."); 437 438 SwingSchedulerDAG SMS(*this, L, getAnalysis<LiveIntervals>(), RegClassInfo, 439 II_setByPragma, LI.LoopPipelinerInfo.get()); 440 441 MachineBasicBlock *MBB = L.getHeader(); 442 // The kernel should not include any terminator instructions. These 443 // will be added back later. 444 SMS.startBlock(MBB); 445 446 // Compute the number of 'real' instructions in the basic block by 447 // ignoring terminators. 448 unsigned size = MBB->size(); 449 for (MachineBasicBlock::iterator I = MBB->getFirstTerminator(), 450 E = MBB->instr_end(); 451 I != E; ++I, --size) 452 ; 453 454 SMS.enterRegion(MBB, MBB->begin(), MBB->getFirstTerminator(), size); 455 SMS.schedule(); 456 SMS.exitRegion(); 457 458 SMS.finishBlock(); 459 return SMS.hasNewSchedule(); 460 } 461 462 void MachinePipeliner::getAnalysisUsage(AnalysisUsage &AU) const { 463 AU.addRequired<AAResultsWrapperPass>(); 464 AU.addPreserved<AAResultsWrapperPass>(); 465 AU.addRequired<MachineLoopInfo>(); 466 AU.addRequired<MachineDominatorTree>(); 467 AU.addRequired<LiveIntervals>(); 468 AU.addRequired<MachineOptimizationRemarkEmitterPass>(); 469 MachineFunctionPass::getAnalysisUsage(AU); 470 } 471 472 void SwingSchedulerDAG::setMII(unsigned ResMII, unsigned RecMII) { 473 if (SwpForceII > 0) 474 MII = SwpForceII; 475 else if (II_setByPragma > 0) 476 MII = II_setByPragma; 477 else 478 MII = std::max(ResMII, RecMII); 479 } 480 481 void SwingSchedulerDAG::setMAX_II() { 482 if (SwpForceII > 0) 483 MAX_II = SwpForceII; 484 else if (II_setByPragma > 0) 485 MAX_II = II_setByPragma; 486 else 487 MAX_II = MII + 10; 488 } 489 490 /// We override the schedule function in ScheduleDAGInstrs to implement the 491 /// scheduling part of the Swing Modulo Scheduling algorithm. 492 void SwingSchedulerDAG::schedule() { 493 AliasAnalysis *AA = &Pass.getAnalysis<AAResultsWrapperPass>().getAAResults(); 494 buildSchedGraph(AA); 495 addLoopCarriedDependences(AA); 496 updatePhiDependences(); 497 Topo.InitDAGTopologicalSorting(); 498 changeDependences(); 499 postprocessDAG(); 500 LLVM_DEBUG(dump()); 501 502 NodeSetType NodeSets; 503 findCircuits(NodeSets); 504 NodeSetType Circuits = NodeSets; 505 506 // Calculate the MII. 507 unsigned ResMII = calculateResMII(); 508 unsigned RecMII = calculateRecMII(NodeSets); 509 510 fuseRecs(NodeSets); 511 512 // This flag is used for testing and can cause correctness problems. 513 if (SwpIgnoreRecMII) 514 RecMII = 0; 515 516 setMII(ResMII, RecMII); 517 setMAX_II(); 518 519 LLVM_DEBUG(dbgs() << "MII = " << MII << " MAX_II = " << MAX_II 520 << " (rec=" << RecMII << ", res=" << ResMII << ")\n"); 521 522 // Can't schedule a loop without a valid MII. 523 if (MII == 0) { 524 LLVM_DEBUG(dbgs() << "Invalid Minimal Initiation Interval: 0\n"); 525 NumFailZeroMII++; 526 Pass.ORE->emit([&]() { 527 return MachineOptimizationRemarkAnalysis( 528 DEBUG_TYPE, "schedule", Loop.getStartLoc(), Loop.getHeader()) 529 << "Invalid Minimal Initiation Interval: 0"; 530 }); 531 return; 532 } 533 534 // Don't pipeline large loops. 535 if (SwpMaxMii != -1 && (int)MII > SwpMaxMii) { 536 LLVM_DEBUG(dbgs() << "MII > " << SwpMaxMii 537 << ", we don't pipeline large loops\n"); 538 NumFailLargeMaxMII++; 539 Pass.ORE->emit([&]() { 540 return MachineOptimizationRemarkAnalysis( 541 DEBUG_TYPE, "schedule", Loop.getStartLoc(), Loop.getHeader()) 542 << "Minimal Initiation Interval too large: " 543 << ore::NV("MII", (int)MII) << " > " 544 << ore::NV("SwpMaxMii", SwpMaxMii) << "." 545 << "Refer to -pipeliner-max-mii."; 546 }); 547 return; 548 } 549 550 computeNodeFunctions(NodeSets); 551 552 registerPressureFilter(NodeSets); 553 554 colocateNodeSets(NodeSets); 555 556 checkNodeSets(NodeSets); 557 558 LLVM_DEBUG({ 559 for (auto &I : NodeSets) { 560 dbgs() << " Rec NodeSet "; 561 I.dump(); 562 } 563 }); 564 565 llvm::stable_sort(NodeSets, std::greater<NodeSet>()); 566 567 groupRemainingNodes(NodeSets); 568 569 removeDuplicateNodes(NodeSets); 570 571 LLVM_DEBUG({ 572 for (auto &I : NodeSets) { 573 dbgs() << " NodeSet "; 574 I.dump(); 575 } 576 }); 577 578 computeNodeOrder(NodeSets); 579 580 // check for node order issues 581 checkValidNodeOrder(Circuits); 582 583 SMSchedule Schedule(Pass.MF, this); 584 Scheduled = schedulePipeline(Schedule); 585 586 if (!Scheduled){ 587 LLVM_DEBUG(dbgs() << "No schedule found, return\n"); 588 NumFailNoSchedule++; 589 Pass.ORE->emit([&]() { 590 return MachineOptimizationRemarkAnalysis( 591 DEBUG_TYPE, "schedule", Loop.getStartLoc(), Loop.getHeader()) 592 << "Unable to find schedule"; 593 }); 594 return; 595 } 596 597 unsigned numStages = Schedule.getMaxStageCount(); 598 // No need to generate pipeline if there are no overlapped iterations. 599 if (numStages == 0) { 600 LLVM_DEBUG(dbgs() << "No overlapped iterations, skip.\n"); 601 NumFailZeroStage++; 602 Pass.ORE->emit([&]() { 603 return MachineOptimizationRemarkAnalysis( 604 DEBUG_TYPE, "schedule", Loop.getStartLoc(), Loop.getHeader()) 605 << "No need to pipeline - no overlapped iterations in schedule."; 606 }); 607 return; 608 } 609 // Check that the maximum stage count is less than user-defined limit. 610 if (SwpMaxStages > -1 && (int)numStages > SwpMaxStages) { 611 LLVM_DEBUG(dbgs() << "numStages:" << numStages << ">" << SwpMaxStages 612 << " : too many stages, abort\n"); 613 NumFailLargeMaxStage++; 614 Pass.ORE->emit([&]() { 615 return MachineOptimizationRemarkAnalysis( 616 DEBUG_TYPE, "schedule", Loop.getStartLoc(), Loop.getHeader()) 617 << "Too many stages in schedule: " 618 << ore::NV("numStages", (int)numStages) << " > " 619 << ore::NV("SwpMaxStages", SwpMaxStages) 620 << ". Refer to -pipeliner-max-stages."; 621 }); 622 return; 623 } 624 625 Pass.ORE->emit([&]() { 626 return MachineOptimizationRemark(DEBUG_TYPE, "schedule", Loop.getStartLoc(), 627 Loop.getHeader()) 628 << "Pipelined succesfully!"; 629 }); 630 631 // Generate the schedule as a ModuloSchedule. 632 DenseMap<MachineInstr *, int> Cycles, Stages; 633 std::vector<MachineInstr *> OrderedInsts; 634 for (int Cycle = Schedule.getFirstCycle(); Cycle <= Schedule.getFinalCycle(); 635 ++Cycle) { 636 for (SUnit *SU : Schedule.getInstructions(Cycle)) { 637 OrderedInsts.push_back(SU->getInstr()); 638 Cycles[SU->getInstr()] = Cycle; 639 Stages[SU->getInstr()] = Schedule.stageScheduled(SU); 640 } 641 } 642 DenseMap<MachineInstr *, std::pair<unsigned, int64_t>> NewInstrChanges; 643 for (auto &KV : NewMIs) { 644 Cycles[KV.first] = Cycles[KV.second]; 645 Stages[KV.first] = Stages[KV.second]; 646 NewInstrChanges[KV.first] = InstrChanges[getSUnit(KV.first)]; 647 } 648 649 ModuloSchedule MS(MF, &Loop, std::move(OrderedInsts), std::move(Cycles), 650 std::move(Stages)); 651 if (EmitTestAnnotations) { 652 assert(NewInstrChanges.empty() && 653 "Cannot serialize a schedule with InstrChanges!"); 654 ModuloScheduleTestAnnotater MSTI(MF, MS); 655 MSTI.annotate(); 656 return; 657 } 658 // The experimental code generator can't work if there are InstChanges. 659 if (ExperimentalCodeGen && NewInstrChanges.empty()) { 660 PeelingModuloScheduleExpander MSE(MF, MS, &LIS); 661 MSE.expand(); 662 } else { 663 ModuloScheduleExpander MSE(MF, MS, LIS, std::move(NewInstrChanges)); 664 MSE.expand(); 665 MSE.cleanup(); 666 } 667 ++NumPipelined; 668 } 669 670 /// Clean up after the software pipeliner runs. 671 void SwingSchedulerDAG::finishBlock() { 672 for (auto &KV : NewMIs) 673 MF.deleteMachineInstr(KV.second); 674 NewMIs.clear(); 675 676 // Call the superclass. 677 ScheduleDAGInstrs::finishBlock(); 678 } 679 680 /// Return the register values for the operands of a Phi instruction. 681 /// This function assume the instruction is a Phi. 682 static void getPhiRegs(MachineInstr &Phi, MachineBasicBlock *Loop, 683 unsigned &InitVal, unsigned &LoopVal) { 684 assert(Phi.isPHI() && "Expecting a Phi."); 685 686 InitVal = 0; 687 LoopVal = 0; 688 for (unsigned i = 1, e = Phi.getNumOperands(); i != e; i += 2) 689 if (Phi.getOperand(i + 1).getMBB() != Loop) 690 InitVal = Phi.getOperand(i).getReg(); 691 else 692 LoopVal = Phi.getOperand(i).getReg(); 693 694 assert(InitVal != 0 && LoopVal != 0 && "Unexpected Phi structure."); 695 } 696 697 /// Return the Phi register value that comes the loop block. 698 static unsigned getLoopPhiReg(MachineInstr &Phi, MachineBasicBlock *LoopBB) { 699 for (unsigned i = 1, e = Phi.getNumOperands(); i != e; i += 2) 700 if (Phi.getOperand(i + 1).getMBB() == LoopBB) 701 return Phi.getOperand(i).getReg(); 702 return 0; 703 } 704 705 /// Return true if SUb can be reached from SUa following the chain edges. 706 static bool isSuccOrder(SUnit *SUa, SUnit *SUb) { 707 SmallPtrSet<SUnit *, 8> Visited; 708 SmallVector<SUnit *, 8> Worklist; 709 Worklist.push_back(SUa); 710 while (!Worklist.empty()) { 711 const SUnit *SU = Worklist.pop_back_val(); 712 for (const auto &SI : SU->Succs) { 713 SUnit *SuccSU = SI.getSUnit(); 714 if (SI.getKind() == SDep::Order) { 715 if (Visited.count(SuccSU)) 716 continue; 717 if (SuccSU == SUb) 718 return true; 719 Worklist.push_back(SuccSU); 720 Visited.insert(SuccSU); 721 } 722 } 723 } 724 return false; 725 } 726 727 /// Return true if the instruction causes a chain between memory 728 /// references before and after it. 729 static bool isDependenceBarrier(MachineInstr &MI) { 730 return MI.isCall() || MI.mayRaiseFPException() || 731 MI.hasUnmodeledSideEffects() || 732 (MI.hasOrderedMemoryRef() && 733 (!MI.mayLoad() || !MI.isDereferenceableInvariantLoad())); 734 } 735 736 /// Return the underlying objects for the memory references of an instruction. 737 /// This function calls the code in ValueTracking, but first checks that the 738 /// instruction has a memory operand. 739 static void getUnderlyingObjects(const MachineInstr *MI, 740 SmallVectorImpl<const Value *> &Objs) { 741 if (!MI->hasOneMemOperand()) 742 return; 743 MachineMemOperand *MM = *MI->memoperands_begin(); 744 if (!MM->getValue()) 745 return; 746 getUnderlyingObjects(MM->getValue(), Objs); 747 for (const Value *V : Objs) { 748 if (!isIdentifiedObject(V)) { 749 Objs.clear(); 750 return; 751 } 752 Objs.push_back(V); 753 } 754 } 755 756 /// Add a chain edge between a load and store if the store can be an 757 /// alias of the load on a subsequent iteration, i.e., a loop carried 758 /// dependence. This code is very similar to the code in ScheduleDAGInstrs 759 /// but that code doesn't create loop carried dependences. 760 void SwingSchedulerDAG::addLoopCarriedDependences(AliasAnalysis *AA) { 761 MapVector<const Value *, SmallVector<SUnit *, 4>> PendingLoads; 762 Value *UnknownValue = 763 UndefValue::get(Type::getVoidTy(MF.getFunction().getContext())); 764 for (auto &SU : SUnits) { 765 MachineInstr &MI = *SU.getInstr(); 766 if (isDependenceBarrier(MI)) 767 PendingLoads.clear(); 768 else if (MI.mayLoad()) { 769 SmallVector<const Value *, 4> Objs; 770 ::getUnderlyingObjects(&MI, Objs); 771 if (Objs.empty()) 772 Objs.push_back(UnknownValue); 773 for (const auto *V : Objs) { 774 SmallVector<SUnit *, 4> &SUs = PendingLoads[V]; 775 SUs.push_back(&SU); 776 } 777 } else if (MI.mayStore()) { 778 SmallVector<const Value *, 4> Objs; 779 ::getUnderlyingObjects(&MI, Objs); 780 if (Objs.empty()) 781 Objs.push_back(UnknownValue); 782 for (const auto *V : Objs) { 783 MapVector<const Value *, SmallVector<SUnit *, 4>>::iterator I = 784 PendingLoads.find(V); 785 if (I == PendingLoads.end()) 786 continue; 787 for (auto *Load : I->second) { 788 if (isSuccOrder(Load, &SU)) 789 continue; 790 MachineInstr &LdMI = *Load->getInstr(); 791 // First, perform the cheaper check that compares the base register. 792 // If they are the same and the load offset is less than the store 793 // offset, then mark the dependence as loop carried potentially. 794 const MachineOperand *BaseOp1, *BaseOp2; 795 int64_t Offset1, Offset2; 796 bool Offset1IsScalable, Offset2IsScalable; 797 if (TII->getMemOperandWithOffset(LdMI, BaseOp1, Offset1, 798 Offset1IsScalable, TRI) && 799 TII->getMemOperandWithOffset(MI, BaseOp2, Offset2, 800 Offset2IsScalable, TRI)) { 801 if (BaseOp1->isIdenticalTo(*BaseOp2) && 802 Offset1IsScalable == Offset2IsScalable && 803 (int)Offset1 < (int)Offset2) { 804 assert(TII->areMemAccessesTriviallyDisjoint(LdMI, MI) && 805 "What happened to the chain edge?"); 806 SDep Dep(Load, SDep::Barrier); 807 Dep.setLatency(1); 808 SU.addPred(Dep); 809 continue; 810 } 811 } 812 // Second, the more expensive check that uses alias analysis on the 813 // base registers. If they alias, and the load offset is less than 814 // the store offset, the mark the dependence as loop carried. 815 if (!AA) { 816 SDep Dep(Load, SDep::Barrier); 817 Dep.setLatency(1); 818 SU.addPred(Dep); 819 continue; 820 } 821 MachineMemOperand *MMO1 = *LdMI.memoperands_begin(); 822 MachineMemOperand *MMO2 = *MI.memoperands_begin(); 823 if (!MMO1->getValue() || !MMO2->getValue()) { 824 SDep Dep(Load, SDep::Barrier); 825 Dep.setLatency(1); 826 SU.addPred(Dep); 827 continue; 828 } 829 if (MMO1->getValue() == MMO2->getValue() && 830 MMO1->getOffset() <= MMO2->getOffset()) { 831 SDep Dep(Load, SDep::Barrier); 832 Dep.setLatency(1); 833 SU.addPred(Dep); 834 continue; 835 } 836 if (!AA->isNoAlias( 837 MemoryLocation::getAfter(MMO1->getValue(), MMO1->getAAInfo()), 838 MemoryLocation::getAfter(MMO2->getValue(), 839 MMO2->getAAInfo()))) { 840 SDep Dep(Load, SDep::Barrier); 841 Dep.setLatency(1); 842 SU.addPred(Dep); 843 } 844 } 845 } 846 } 847 } 848 } 849 850 /// Update the phi dependences to the DAG because ScheduleDAGInstrs no longer 851 /// processes dependences for PHIs. This function adds true dependences 852 /// from a PHI to a use, and a loop carried dependence from the use to the 853 /// PHI. The loop carried dependence is represented as an anti dependence 854 /// edge. This function also removes chain dependences between unrelated 855 /// PHIs. 856 void SwingSchedulerDAG::updatePhiDependences() { 857 SmallVector<SDep, 4> RemoveDeps; 858 const TargetSubtargetInfo &ST = MF.getSubtarget<TargetSubtargetInfo>(); 859 860 // Iterate over each DAG node. 861 for (SUnit &I : SUnits) { 862 RemoveDeps.clear(); 863 // Set to true if the instruction has an operand defined by a Phi. 864 unsigned HasPhiUse = 0; 865 unsigned HasPhiDef = 0; 866 MachineInstr *MI = I.getInstr(); 867 // Iterate over each operand, and we process the definitions. 868 for (MachineInstr::mop_iterator MOI = MI->operands_begin(), 869 MOE = MI->operands_end(); 870 MOI != MOE; ++MOI) { 871 if (!MOI->isReg()) 872 continue; 873 Register Reg = MOI->getReg(); 874 if (MOI->isDef()) { 875 // If the register is used by a Phi, then create an anti dependence. 876 for (MachineRegisterInfo::use_instr_iterator 877 UI = MRI.use_instr_begin(Reg), 878 UE = MRI.use_instr_end(); 879 UI != UE; ++UI) { 880 MachineInstr *UseMI = &*UI; 881 SUnit *SU = getSUnit(UseMI); 882 if (SU != nullptr && UseMI->isPHI()) { 883 if (!MI->isPHI()) { 884 SDep Dep(SU, SDep::Anti, Reg); 885 Dep.setLatency(1); 886 I.addPred(Dep); 887 } else { 888 HasPhiDef = Reg; 889 // Add a chain edge to a dependent Phi that isn't an existing 890 // predecessor. 891 if (SU->NodeNum < I.NodeNum && !I.isPred(SU)) 892 I.addPred(SDep(SU, SDep::Barrier)); 893 } 894 } 895 } 896 } else if (MOI->isUse()) { 897 // If the register is defined by a Phi, then create a true dependence. 898 MachineInstr *DefMI = MRI.getUniqueVRegDef(Reg); 899 if (DefMI == nullptr) 900 continue; 901 SUnit *SU = getSUnit(DefMI); 902 if (SU != nullptr && DefMI->isPHI()) { 903 if (!MI->isPHI()) { 904 SDep Dep(SU, SDep::Data, Reg); 905 Dep.setLatency(0); 906 ST.adjustSchedDependency(SU, 0, &I, MI->getOperandNo(MOI), Dep); 907 I.addPred(Dep); 908 } else { 909 HasPhiUse = Reg; 910 // Add a chain edge to a dependent Phi that isn't an existing 911 // predecessor. 912 if (SU->NodeNum < I.NodeNum && !I.isPred(SU)) 913 I.addPred(SDep(SU, SDep::Barrier)); 914 } 915 } 916 } 917 } 918 // Remove order dependences from an unrelated Phi. 919 if (!SwpPruneDeps) 920 continue; 921 for (auto &PI : I.Preds) { 922 MachineInstr *PMI = PI.getSUnit()->getInstr(); 923 if (PMI->isPHI() && PI.getKind() == SDep::Order) { 924 if (I.getInstr()->isPHI()) { 925 if (PMI->getOperand(0).getReg() == HasPhiUse) 926 continue; 927 if (getLoopPhiReg(*PMI, PMI->getParent()) == HasPhiDef) 928 continue; 929 } 930 RemoveDeps.push_back(PI); 931 } 932 } 933 for (int i = 0, e = RemoveDeps.size(); i != e; ++i) 934 I.removePred(RemoveDeps[i]); 935 } 936 } 937 938 /// Iterate over each DAG node and see if we can change any dependences 939 /// in order to reduce the recurrence MII. 940 void SwingSchedulerDAG::changeDependences() { 941 // See if an instruction can use a value from the previous iteration. 942 // If so, we update the base and offset of the instruction and change 943 // the dependences. 944 for (SUnit &I : SUnits) { 945 unsigned BasePos = 0, OffsetPos = 0, NewBase = 0; 946 int64_t NewOffset = 0; 947 if (!canUseLastOffsetValue(I.getInstr(), BasePos, OffsetPos, NewBase, 948 NewOffset)) 949 continue; 950 951 // Get the MI and SUnit for the instruction that defines the original base. 952 Register OrigBase = I.getInstr()->getOperand(BasePos).getReg(); 953 MachineInstr *DefMI = MRI.getUniqueVRegDef(OrigBase); 954 if (!DefMI) 955 continue; 956 SUnit *DefSU = getSUnit(DefMI); 957 if (!DefSU) 958 continue; 959 // Get the MI and SUnit for the instruction that defins the new base. 960 MachineInstr *LastMI = MRI.getUniqueVRegDef(NewBase); 961 if (!LastMI) 962 continue; 963 SUnit *LastSU = getSUnit(LastMI); 964 if (!LastSU) 965 continue; 966 967 if (Topo.IsReachable(&I, LastSU)) 968 continue; 969 970 // Remove the dependence. The value now depends on a prior iteration. 971 SmallVector<SDep, 4> Deps; 972 for (const SDep &P : I.Preds) 973 if (P.getSUnit() == DefSU) 974 Deps.push_back(P); 975 for (int i = 0, e = Deps.size(); i != e; i++) { 976 Topo.RemovePred(&I, Deps[i].getSUnit()); 977 I.removePred(Deps[i]); 978 } 979 // Remove the chain dependence between the instructions. 980 Deps.clear(); 981 for (auto &P : LastSU->Preds) 982 if (P.getSUnit() == &I && P.getKind() == SDep::Order) 983 Deps.push_back(P); 984 for (int i = 0, e = Deps.size(); i != e; i++) { 985 Topo.RemovePred(LastSU, Deps[i].getSUnit()); 986 LastSU->removePred(Deps[i]); 987 } 988 989 // Add a dependence between the new instruction and the instruction 990 // that defines the new base. 991 SDep Dep(&I, SDep::Anti, NewBase); 992 Topo.AddPred(LastSU, &I); 993 LastSU->addPred(Dep); 994 995 // Remember the base and offset information so that we can update the 996 // instruction during code generation. 997 InstrChanges[&I] = std::make_pair(NewBase, NewOffset); 998 } 999 } 1000 1001 namespace { 1002 1003 // FuncUnitSorter - Comparison operator used to sort instructions by 1004 // the number of functional unit choices. 1005 struct FuncUnitSorter { 1006 const InstrItineraryData *InstrItins; 1007 const MCSubtargetInfo *STI; 1008 DenseMap<InstrStage::FuncUnits, unsigned> Resources; 1009 1010 FuncUnitSorter(const TargetSubtargetInfo &TSI) 1011 : InstrItins(TSI.getInstrItineraryData()), STI(&TSI) {} 1012 1013 // Compute the number of functional unit alternatives needed 1014 // at each stage, and take the minimum value. We prioritize the 1015 // instructions by the least number of choices first. 1016 unsigned minFuncUnits(const MachineInstr *Inst, 1017 InstrStage::FuncUnits &F) const { 1018 unsigned SchedClass = Inst->getDesc().getSchedClass(); 1019 unsigned min = UINT_MAX; 1020 if (InstrItins && !InstrItins->isEmpty()) { 1021 for (const InstrStage &IS : 1022 make_range(InstrItins->beginStage(SchedClass), 1023 InstrItins->endStage(SchedClass))) { 1024 InstrStage::FuncUnits funcUnits = IS.getUnits(); 1025 unsigned numAlternatives = llvm::popcount(funcUnits); 1026 if (numAlternatives < min) { 1027 min = numAlternatives; 1028 F = funcUnits; 1029 } 1030 } 1031 return min; 1032 } 1033 if (STI && STI->getSchedModel().hasInstrSchedModel()) { 1034 const MCSchedClassDesc *SCDesc = 1035 STI->getSchedModel().getSchedClassDesc(SchedClass); 1036 if (!SCDesc->isValid()) 1037 // No valid Schedule Class Desc for schedClass, should be 1038 // Pseudo/PostRAPseudo 1039 return min; 1040 1041 for (const MCWriteProcResEntry &PRE : 1042 make_range(STI->getWriteProcResBegin(SCDesc), 1043 STI->getWriteProcResEnd(SCDesc))) { 1044 if (!PRE.Cycles) 1045 continue; 1046 const MCProcResourceDesc *ProcResource = 1047 STI->getSchedModel().getProcResource(PRE.ProcResourceIdx); 1048 unsigned NumUnits = ProcResource->NumUnits; 1049 if (NumUnits < min) { 1050 min = NumUnits; 1051 F = PRE.ProcResourceIdx; 1052 } 1053 } 1054 return min; 1055 } 1056 llvm_unreachable("Should have non-empty InstrItins or hasInstrSchedModel!"); 1057 } 1058 1059 // Compute the critical resources needed by the instruction. This 1060 // function records the functional units needed by instructions that 1061 // must use only one functional unit. We use this as a tie breaker 1062 // for computing the resource MII. The instrutions that require 1063 // the same, highly used, functional unit have high priority. 1064 void calcCriticalResources(MachineInstr &MI) { 1065 unsigned SchedClass = MI.getDesc().getSchedClass(); 1066 if (InstrItins && !InstrItins->isEmpty()) { 1067 for (const InstrStage &IS : 1068 make_range(InstrItins->beginStage(SchedClass), 1069 InstrItins->endStage(SchedClass))) { 1070 InstrStage::FuncUnits FuncUnits = IS.getUnits(); 1071 if (llvm::popcount(FuncUnits) == 1) 1072 Resources[FuncUnits]++; 1073 } 1074 return; 1075 } 1076 if (STI && STI->getSchedModel().hasInstrSchedModel()) { 1077 const MCSchedClassDesc *SCDesc = 1078 STI->getSchedModel().getSchedClassDesc(SchedClass); 1079 if (!SCDesc->isValid()) 1080 // No valid Schedule Class Desc for schedClass, should be 1081 // Pseudo/PostRAPseudo 1082 return; 1083 1084 for (const MCWriteProcResEntry &PRE : 1085 make_range(STI->getWriteProcResBegin(SCDesc), 1086 STI->getWriteProcResEnd(SCDesc))) { 1087 if (!PRE.Cycles) 1088 continue; 1089 Resources[PRE.ProcResourceIdx]++; 1090 } 1091 return; 1092 } 1093 llvm_unreachable("Should have non-empty InstrItins or hasInstrSchedModel!"); 1094 } 1095 1096 /// Return true if IS1 has less priority than IS2. 1097 bool operator()(const MachineInstr *IS1, const MachineInstr *IS2) const { 1098 InstrStage::FuncUnits F1 = 0, F2 = 0; 1099 unsigned MFUs1 = minFuncUnits(IS1, F1); 1100 unsigned MFUs2 = minFuncUnits(IS2, F2); 1101 if (MFUs1 == MFUs2) 1102 return Resources.lookup(F1) < Resources.lookup(F2); 1103 return MFUs1 > MFUs2; 1104 } 1105 }; 1106 1107 } // end anonymous namespace 1108 1109 /// Calculate the resource constrained minimum initiation interval for the 1110 /// specified loop. We use the DFA to model the resources needed for 1111 /// each instruction, and we ignore dependences. A different DFA is created 1112 /// for each cycle that is required. When adding a new instruction, we attempt 1113 /// to add it to each existing DFA, until a legal space is found. If the 1114 /// instruction cannot be reserved in an existing DFA, we create a new one. 1115 unsigned SwingSchedulerDAG::calculateResMII() { 1116 LLVM_DEBUG(dbgs() << "calculateResMII:\n"); 1117 ResourceManager RM(&MF.getSubtarget(), this); 1118 return RM.calculateResMII(); 1119 } 1120 1121 /// Calculate the recurrence-constrainted minimum initiation interval. 1122 /// Iterate over each circuit. Compute the delay(c) and distance(c) 1123 /// for each circuit. The II needs to satisfy the inequality 1124 /// delay(c) - II*distance(c) <= 0. For each circuit, choose the smallest 1125 /// II that satisfies the inequality, and the RecMII is the maximum 1126 /// of those values. 1127 unsigned SwingSchedulerDAG::calculateRecMII(NodeSetType &NodeSets) { 1128 unsigned RecMII = 0; 1129 1130 for (NodeSet &Nodes : NodeSets) { 1131 if (Nodes.empty()) 1132 continue; 1133 1134 unsigned Delay = Nodes.getLatency(); 1135 unsigned Distance = 1; 1136 1137 // ii = ceil(delay / distance) 1138 unsigned CurMII = (Delay + Distance - 1) / Distance; 1139 Nodes.setRecMII(CurMII); 1140 if (CurMII > RecMII) 1141 RecMII = CurMII; 1142 } 1143 1144 return RecMII; 1145 } 1146 1147 /// Swap all the anti dependences in the DAG. That means it is no longer a DAG, 1148 /// but we do this to find the circuits, and then change them back. 1149 static void swapAntiDependences(std::vector<SUnit> &SUnits) { 1150 SmallVector<std::pair<SUnit *, SDep>, 8> DepsAdded; 1151 for (SUnit &SU : SUnits) { 1152 for (SDep &Pred : SU.Preds) 1153 if (Pred.getKind() == SDep::Anti) 1154 DepsAdded.push_back(std::make_pair(&SU, Pred)); 1155 } 1156 for (std::pair<SUnit *, SDep> &P : DepsAdded) { 1157 // Remove this anti dependency and add one in the reverse direction. 1158 SUnit *SU = P.first; 1159 SDep &D = P.second; 1160 SUnit *TargetSU = D.getSUnit(); 1161 unsigned Reg = D.getReg(); 1162 unsigned Lat = D.getLatency(); 1163 SU->removePred(D); 1164 SDep Dep(SU, SDep::Anti, Reg); 1165 Dep.setLatency(Lat); 1166 TargetSU->addPred(Dep); 1167 } 1168 } 1169 1170 /// Create the adjacency structure of the nodes in the graph. 1171 void SwingSchedulerDAG::Circuits::createAdjacencyStructure( 1172 SwingSchedulerDAG *DAG) { 1173 BitVector Added(SUnits.size()); 1174 DenseMap<int, int> OutputDeps; 1175 for (int i = 0, e = SUnits.size(); i != e; ++i) { 1176 Added.reset(); 1177 // Add any successor to the adjacency matrix and exclude duplicates. 1178 for (auto &SI : SUnits[i].Succs) { 1179 // Only create a back-edge on the first and last nodes of a dependence 1180 // chain. This records any chains and adds them later. 1181 if (SI.getKind() == SDep::Output) { 1182 int N = SI.getSUnit()->NodeNum; 1183 int BackEdge = i; 1184 auto Dep = OutputDeps.find(BackEdge); 1185 if (Dep != OutputDeps.end()) { 1186 BackEdge = Dep->second; 1187 OutputDeps.erase(Dep); 1188 } 1189 OutputDeps[N] = BackEdge; 1190 } 1191 // Do not process a boundary node, an artificial node. 1192 // A back-edge is processed only if it goes to a Phi. 1193 if (SI.getSUnit()->isBoundaryNode() || SI.isArtificial() || 1194 (SI.getKind() == SDep::Anti && !SI.getSUnit()->getInstr()->isPHI())) 1195 continue; 1196 int N = SI.getSUnit()->NodeNum; 1197 if (!Added.test(N)) { 1198 AdjK[i].push_back(N); 1199 Added.set(N); 1200 } 1201 } 1202 // A chain edge between a store and a load is treated as a back-edge in the 1203 // adjacency matrix. 1204 for (auto &PI : SUnits[i].Preds) { 1205 if (!SUnits[i].getInstr()->mayStore() || 1206 !DAG->isLoopCarriedDep(&SUnits[i], PI, false)) 1207 continue; 1208 if (PI.getKind() == SDep::Order && PI.getSUnit()->getInstr()->mayLoad()) { 1209 int N = PI.getSUnit()->NodeNum; 1210 if (!Added.test(N)) { 1211 AdjK[i].push_back(N); 1212 Added.set(N); 1213 } 1214 } 1215 } 1216 } 1217 // Add back-edges in the adjacency matrix for the output dependences. 1218 for (auto &OD : OutputDeps) 1219 if (!Added.test(OD.second)) { 1220 AdjK[OD.first].push_back(OD.second); 1221 Added.set(OD.second); 1222 } 1223 } 1224 1225 /// Identify an elementary circuit in the dependence graph starting at the 1226 /// specified node. 1227 bool SwingSchedulerDAG::Circuits::circuit(int V, int S, NodeSetType &NodeSets, 1228 bool HasBackedge) { 1229 SUnit *SV = &SUnits[V]; 1230 bool F = false; 1231 Stack.insert(SV); 1232 Blocked.set(V); 1233 1234 for (auto W : AdjK[V]) { 1235 if (NumPaths > MaxPaths) 1236 break; 1237 if (W < S) 1238 continue; 1239 if (W == S) { 1240 if (!HasBackedge) 1241 NodeSets.push_back(NodeSet(Stack.begin(), Stack.end())); 1242 F = true; 1243 ++NumPaths; 1244 break; 1245 } else if (!Blocked.test(W)) { 1246 if (circuit(W, S, NodeSets, 1247 Node2Idx->at(W) < Node2Idx->at(V) ? true : HasBackedge)) 1248 F = true; 1249 } 1250 } 1251 1252 if (F) 1253 unblock(V); 1254 else { 1255 for (auto W : AdjK[V]) { 1256 if (W < S) 1257 continue; 1258 B[W].insert(SV); 1259 } 1260 } 1261 Stack.pop_back(); 1262 return F; 1263 } 1264 1265 /// Unblock a node in the circuit finding algorithm. 1266 void SwingSchedulerDAG::Circuits::unblock(int U) { 1267 Blocked.reset(U); 1268 SmallPtrSet<SUnit *, 4> &BU = B[U]; 1269 while (!BU.empty()) { 1270 SmallPtrSet<SUnit *, 4>::iterator SI = BU.begin(); 1271 assert(SI != BU.end() && "Invalid B set."); 1272 SUnit *W = *SI; 1273 BU.erase(W); 1274 if (Blocked.test(W->NodeNum)) 1275 unblock(W->NodeNum); 1276 } 1277 } 1278 1279 /// Identify all the elementary circuits in the dependence graph using 1280 /// Johnson's circuit algorithm. 1281 void SwingSchedulerDAG::findCircuits(NodeSetType &NodeSets) { 1282 // Swap all the anti dependences in the DAG. That means it is no longer a DAG, 1283 // but we do this to find the circuits, and then change them back. 1284 swapAntiDependences(SUnits); 1285 1286 Circuits Cir(SUnits, Topo); 1287 // Create the adjacency structure. 1288 Cir.createAdjacencyStructure(this); 1289 for (int i = 0, e = SUnits.size(); i != e; ++i) { 1290 Cir.reset(); 1291 Cir.circuit(i, i, NodeSets); 1292 } 1293 1294 // Change the dependences back so that we've created a DAG again. 1295 swapAntiDependences(SUnits); 1296 } 1297 1298 // Create artificial dependencies between the source of COPY/REG_SEQUENCE that 1299 // is loop-carried to the USE in next iteration. This will help pipeliner avoid 1300 // additional copies that are needed across iterations. An artificial dependence 1301 // edge is added from USE to SOURCE of COPY/REG_SEQUENCE. 1302 1303 // PHI-------Anti-Dep-----> COPY/REG_SEQUENCE (loop-carried) 1304 // SRCOfCopY------True-Dep---> COPY/REG_SEQUENCE 1305 // PHI-------True-Dep------> USEOfPhi 1306 1307 // The mutation creates 1308 // USEOfPHI -------Artificial-Dep---> SRCOfCopy 1309 1310 // This overall will ensure, the USEOfPHI is scheduled before SRCOfCopy 1311 // (since USE is a predecessor), implies, the COPY/ REG_SEQUENCE is scheduled 1312 // late to avoid additional copies across iterations. The possible scheduling 1313 // order would be 1314 // USEOfPHI --- SRCOfCopy--- COPY/REG_SEQUENCE. 1315 1316 void SwingSchedulerDAG::CopyToPhiMutation::apply(ScheduleDAGInstrs *DAG) { 1317 for (SUnit &SU : DAG->SUnits) { 1318 // Find the COPY/REG_SEQUENCE instruction. 1319 if (!SU.getInstr()->isCopy() && !SU.getInstr()->isRegSequence()) 1320 continue; 1321 1322 // Record the loop carried PHIs. 1323 SmallVector<SUnit *, 4> PHISUs; 1324 // Record the SrcSUs that feed the COPY/REG_SEQUENCE instructions. 1325 SmallVector<SUnit *, 4> SrcSUs; 1326 1327 for (auto &Dep : SU.Preds) { 1328 SUnit *TmpSU = Dep.getSUnit(); 1329 MachineInstr *TmpMI = TmpSU->getInstr(); 1330 SDep::Kind DepKind = Dep.getKind(); 1331 // Save the loop carried PHI. 1332 if (DepKind == SDep::Anti && TmpMI->isPHI()) 1333 PHISUs.push_back(TmpSU); 1334 // Save the source of COPY/REG_SEQUENCE. 1335 // If the source has no pre-decessors, we will end up creating cycles. 1336 else if (DepKind == SDep::Data && !TmpMI->isPHI() && TmpSU->NumPreds > 0) 1337 SrcSUs.push_back(TmpSU); 1338 } 1339 1340 if (PHISUs.size() == 0 || SrcSUs.size() == 0) 1341 continue; 1342 1343 // Find the USEs of PHI. If the use is a PHI or REG_SEQUENCE, push back this 1344 // SUnit to the container. 1345 SmallVector<SUnit *, 8> UseSUs; 1346 // Do not use iterator based loop here as we are updating the container. 1347 for (size_t Index = 0; Index < PHISUs.size(); ++Index) { 1348 for (auto &Dep : PHISUs[Index]->Succs) { 1349 if (Dep.getKind() != SDep::Data) 1350 continue; 1351 1352 SUnit *TmpSU = Dep.getSUnit(); 1353 MachineInstr *TmpMI = TmpSU->getInstr(); 1354 if (TmpMI->isPHI() || TmpMI->isRegSequence()) { 1355 PHISUs.push_back(TmpSU); 1356 continue; 1357 } 1358 UseSUs.push_back(TmpSU); 1359 } 1360 } 1361 1362 if (UseSUs.size() == 0) 1363 continue; 1364 1365 SwingSchedulerDAG *SDAG = cast<SwingSchedulerDAG>(DAG); 1366 // Add the artificial dependencies if it does not form a cycle. 1367 for (auto *I : UseSUs) { 1368 for (auto *Src : SrcSUs) { 1369 if (!SDAG->Topo.IsReachable(I, Src) && Src != I) { 1370 Src->addPred(SDep(I, SDep::Artificial)); 1371 SDAG->Topo.AddPred(Src, I); 1372 } 1373 } 1374 } 1375 } 1376 } 1377 1378 /// Return true for DAG nodes that we ignore when computing the cost functions. 1379 /// We ignore the back-edge recurrence in order to avoid unbounded recursion 1380 /// in the calculation of the ASAP, ALAP, etc functions. 1381 static bool ignoreDependence(const SDep &D, bool isPred) { 1382 if (D.isArtificial() || D.getSUnit()->isBoundaryNode()) 1383 return true; 1384 return D.getKind() == SDep::Anti && isPred; 1385 } 1386 1387 /// Compute several functions need to order the nodes for scheduling. 1388 /// ASAP - Earliest time to schedule a node. 1389 /// ALAP - Latest time to schedule a node. 1390 /// MOV - Mobility function, difference between ALAP and ASAP. 1391 /// D - Depth of each node. 1392 /// H - Height of each node. 1393 void SwingSchedulerDAG::computeNodeFunctions(NodeSetType &NodeSets) { 1394 ScheduleInfo.resize(SUnits.size()); 1395 1396 LLVM_DEBUG({ 1397 for (int I : Topo) { 1398 const SUnit &SU = SUnits[I]; 1399 dumpNode(SU); 1400 } 1401 }); 1402 1403 int maxASAP = 0; 1404 // Compute ASAP and ZeroLatencyDepth. 1405 for (int I : Topo) { 1406 int asap = 0; 1407 int zeroLatencyDepth = 0; 1408 SUnit *SU = &SUnits[I]; 1409 for (const SDep &P : SU->Preds) { 1410 SUnit *pred = P.getSUnit(); 1411 if (P.getLatency() == 0) 1412 zeroLatencyDepth = 1413 std::max(zeroLatencyDepth, getZeroLatencyDepth(pred) + 1); 1414 if (ignoreDependence(P, true)) 1415 continue; 1416 asap = std::max(asap, (int)(getASAP(pred) + P.getLatency() - 1417 getDistance(pred, SU, P) * MII)); 1418 } 1419 maxASAP = std::max(maxASAP, asap); 1420 ScheduleInfo[I].ASAP = asap; 1421 ScheduleInfo[I].ZeroLatencyDepth = zeroLatencyDepth; 1422 } 1423 1424 // Compute ALAP, ZeroLatencyHeight, and MOV. 1425 for (int I : llvm::reverse(Topo)) { 1426 int alap = maxASAP; 1427 int zeroLatencyHeight = 0; 1428 SUnit *SU = &SUnits[I]; 1429 for (const SDep &S : SU->Succs) { 1430 SUnit *succ = S.getSUnit(); 1431 if (succ->isBoundaryNode()) 1432 continue; 1433 if (S.getLatency() == 0) 1434 zeroLatencyHeight = 1435 std::max(zeroLatencyHeight, getZeroLatencyHeight(succ) + 1); 1436 if (ignoreDependence(S, true)) 1437 continue; 1438 alap = std::min(alap, (int)(getALAP(succ) - S.getLatency() + 1439 getDistance(SU, succ, S) * MII)); 1440 } 1441 1442 ScheduleInfo[I].ALAP = alap; 1443 ScheduleInfo[I].ZeroLatencyHeight = zeroLatencyHeight; 1444 } 1445 1446 // After computing the node functions, compute the summary for each node set. 1447 for (NodeSet &I : NodeSets) 1448 I.computeNodeSetInfo(this); 1449 1450 LLVM_DEBUG({ 1451 for (unsigned i = 0; i < SUnits.size(); i++) { 1452 dbgs() << "\tNode " << i << ":\n"; 1453 dbgs() << "\t ASAP = " << getASAP(&SUnits[i]) << "\n"; 1454 dbgs() << "\t ALAP = " << getALAP(&SUnits[i]) << "\n"; 1455 dbgs() << "\t MOV = " << getMOV(&SUnits[i]) << "\n"; 1456 dbgs() << "\t D = " << getDepth(&SUnits[i]) << "\n"; 1457 dbgs() << "\t H = " << getHeight(&SUnits[i]) << "\n"; 1458 dbgs() << "\t ZLD = " << getZeroLatencyDepth(&SUnits[i]) << "\n"; 1459 dbgs() << "\t ZLH = " << getZeroLatencyHeight(&SUnits[i]) << "\n"; 1460 } 1461 }); 1462 } 1463 1464 /// Compute the Pred_L(O) set, as defined in the paper. The set is defined 1465 /// as the predecessors of the elements of NodeOrder that are not also in 1466 /// NodeOrder. 1467 static bool pred_L(SetVector<SUnit *> &NodeOrder, 1468 SmallSetVector<SUnit *, 8> &Preds, 1469 const NodeSet *S = nullptr) { 1470 Preds.clear(); 1471 for (const SUnit *SU : NodeOrder) { 1472 for (const SDep &Pred : SU->Preds) { 1473 if (S && S->count(Pred.getSUnit()) == 0) 1474 continue; 1475 if (ignoreDependence(Pred, true)) 1476 continue; 1477 if (NodeOrder.count(Pred.getSUnit()) == 0) 1478 Preds.insert(Pred.getSUnit()); 1479 } 1480 // Back-edges are predecessors with an anti-dependence. 1481 for (const SDep &Succ : SU->Succs) { 1482 if (Succ.getKind() != SDep::Anti) 1483 continue; 1484 if (S && S->count(Succ.getSUnit()) == 0) 1485 continue; 1486 if (NodeOrder.count(Succ.getSUnit()) == 0) 1487 Preds.insert(Succ.getSUnit()); 1488 } 1489 } 1490 return !Preds.empty(); 1491 } 1492 1493 /// Compute the Succ_L(O) set, as defined in the paper. The set is defined 1494 /// as the successors of the elements of NodeOrder that are not also in 1495 /// NodeOrder. 1496 static bool succ_L(SetVector<SUnit *> &NodeOrder, 1497 SmallSetVector<SUnit *, 8> &Succs, 1498 const NodeSet *S = nullptr) { 1499 Succs.clear(); 1500 for (const SUnit *SU : NodeOrder) { 1501 for (const SDep &Succ : SU->Succs) { 1502 if (S && S->count(Succ.getSUnit()) == 0) 1503 continue; 1504 if (ignoreDependence(Succ, false)) 1505 continue; 1506 if (NodeOrder.count(Succ.getSUnit()) == 0) 1507 Succs.insert(Succ.getSUnit()); 1508 } 1509 for (const SDep &Pred : SU->Preds) { 1510 if (Pred.getKind() != SDep::Anti) 1511 continue; 1512 if (S && S->count(Pred.getSUnit()) == 0) 1513 continue; 1514 if (NodeOrder.count(Pred.getSUnit()) == 0) 1515 Succs.insert(Pred.getSUnit()); 1516 } 1517 } 1518 return !Succs.empty(); 1519 } 1520 1521 /// Return true if there is a path from the specified node to any of the nodes 1522 /// in DestNodes. Keep track and return the nodes in any path. 1523 static bool computePath(SUnit *Cur, SetVector<SUnit *> &Path, 1524 SetVector<SUnit *> &DestNodes, 1525 SetVector<SUnit *> &Exclude, 1526 SmallPtrSet<SUnit *, 8> &Visited) { 1527 if (Cur->isBoundaryNode()) 1528 return false; 1529 if (Exclude.contains(Cur)) 1530 return false; 1531 if (DestNodes.contains(Cur)) 1532 return true; 1533 if (!Visited.insert(Cur).second) 1534 return Path.contains(Cur); 1535 bool FoundPath = false; 1536 for (auto &SI : Cur->Succs) 1537 if (!ignoreDependence(SI, false)) 1538 FoundPath |= 1539 computePath(SI.getSUnit(), Path, DestNodes, Exclude, Visited); 1540 for (auto &PI : Cur->Preds) 1541 if (PI.getKind() == SDep::Anti) 1542 FoundPath |= 1543 computePath(PI.getSUnit(), Path, DestNodes, Exclude, Visited); 1544 if (FoundPath) 1545 Path.insert(Cur); 1546 return FoundPath; 1547 } 1548 1549 /// Compute the live-out registers for the instructions in a node-set. 1550 /// The live-out registers are those that are defined in the node-set, 1551 /// but not used. Except for use operands of Phis. 1552 static void computeLiveOuts(MachineFunction &MF, RegPressureTracker &RPTracker, 1553 NodeSet &NS) { 1554 const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo(); 1555 MachineRegisterInfo &MRI = MF.getRegInfo(); 1556 SmallVector<RegisterMaskPair, 8> LiveOutRegs; 1557 SmallSet<unsigned, 4> Uses; 1558 for (SUnit *SU : NS) { 1559 const MachineInstr *MI = SU->getInstr(); 1560 if (MI->isPHI()) 1561 continue; 1562 for (const MachineOperand &MO : MI->operands()) 1563 if (MO.isReg() && MO.isUse()) { 1564 Register Reg = MO.getReg(); 1565 if (Reg.isVirtual()) 1566 Uses.insert(Reg); 1567 else if (MRI.isAllocatable(Reg)) 1568 for (MCRegUnitIterator Units(Reg.asMCReg(), TRI); Units.isValid(); 1569 ++Units) 1570 Uses.insert(*Units); 1571 } 1572 } 1573 for (SUnit *SU : NS) 1574 for (const MachineOperand &MO : SU->getInstr()->operands()) 1575 if (MO.isReg() && MO.isDef() && !MO.isDead()) { 1576 Register Reg = MO.getReg(); 1577 if (Reg.isVirtual()) { 1578 if (!Uses.count(Reg)) 1579 LiveOutRegs.push_back(RegisterMaskPair(Reg, 1580 LaneBitmask::getNone())); 1581 } else if (MRI.isAllocatable(Reg)) { 1582 for (MCRegUnitIterator Units(Reg.asMCReg(), TRI); Units.isValid(); 1583 ++Units) 1584 if (!Uses.count(*Units)) 1585 LiveOutRegs.push_back(RegisterMaskPair(*Units, 1586 LaneBitmask::getNone())); 1587 } 1588 } 1589 RPTracker.addLiveRegs(LiveOutRegs); 1590 } 1591 1592 /// A heuristic to filter nodes in recurrent node-sets if the register 1593 /// pressure of a set is too high. 1594 void SwingSchedulerDAG::registerPressureFilter(NodeSetType &NodeSets) { 1595 for (auto &NS : NodeSets) { 1596 // Skip small node-sets since they won't cause register pressure problems. 1597 if (NS.size() <= 2) 1598 continue; 1599 IntervalPressure RecRegPressure; 1600 RegPressureTracker RecRPTracker(RecRegPressure); 1601 RecRPTracker.init(&MF, &RegClassInfo, &LIS, BB, BB->end(), false, true); 1602 computeLiveOuts(MF, RecRPTracker, NS); 1603 RecRPTracker.closeBottom(); 1604 1605 std::vector<SUnit *> SUnits(NS.begin(), NS.end()); 1606 llvm::sort(SUnits, [](const SUnit *A, const SUnit *B) { 1607 return A->NodeNum > B->NodeNum; 1608 }); 1609 1610 for (auto &SU : SUnits) { 1611 // Since we're computing the register pressure for a subset of the 1612 // instructions in a block, we need to set the tracker for each 1613 // instruction in the node-set. The tracker is set to the instruction 1614 // just after the one we're interested in. 1615 MachineBasicBlock::const_iterator CurInstI = SU->getInstr(); 1616 RecRPTracker.setPos(std::next(CurInstI)); 1617 1618 RegPressureDelta RPDelta; 1619 ArrayRef<PressureChange> CriticalPSets; 1620 RecRPTracker.getMaxUpwardPressureDelta(SU->getInstr(), nullptr, RPDelta, 1621 CriticalPSets, 1622 RecRegPressure.MaxSetPressure); 1623 if (RPDelta.Excess.isValid()) { 1624 LLVM_DEBUG( 1625 dbgs() << "Excess register pressure: SU(" << SU->NodeNum << ") " 1626 << TRI->getRegPressureSetName(RPDelta.Excess.getPSet()) 1627 << ":" << RPDelta.Excess.getUnitInc() << "\n"); 1628 NS.setExceedPressure(SU); 1629 break; 1630 } 1631 RecRPTracker.recede(); 1632 } 1633 } 1634 } 1635 1636 /// A heuristic to colocate node sets that have the same set of 1637 /// successors. 1638 void SwingSchedulerDAG::colocateNodeSets(NodeSetType &NodeSets) { 1639 unsigned Colocate = 0; 1640 for (int i = 0, e = NodeSets.size(); i < e; ++i) { 1641 NodeSet &N1 = NodeSets[i]; 1642 SmallSetVector<SUnit *, 8> S1; 1643 if (N1.empty() || !succ_L(N1, S1)) 1644 continue; 1645 for (int j = i + 1; j < e; ++j) { 1646 NodeSet &N2 = NodeSets[j]; 1647 if (N1.compareRecMII(N2) != 0) 1648 continue; 1649 SmallSetVector<SUnit *, 8> S2; 1650 if (N2.empty() || !succ_L(N2, S2)) 1651 continue; 1652 if (llvm::set_is_subset(S1, S2) && S1.size() == S2.size()) { 1653 N1.setColocate(++Colocate); 1654 N2.setColocate(Colocate); 1655 break; 1656 } 1657 } 1658 } 1659 } 1660 1661 /// Check if the existing node-sets are profitable. If not, then ignore the 1662 /// recurrent node-sets, and attempt to schedule all nodes together. This is 1663 /// a heuristic. If the MII is large and all the recurrent node-sets are small, 1664 /// then it's best to try to schedule all instructions together instead of 1665 /// starting with the recurrent node-sets. 1666 void SwingSchedulerDAG::checkNodeSets(NodeSetType &NodeSets) { 1667 // Look for loops with a large MII. 1668 if (MII < 17) 1669 return; 1670 // Check if the node-set contains only a simple add recurrence. 1671 for (auto &NS : NodeSets) { 1672 if (NS.getRecMII() > 2) 1673 return; 1674 if (NS.getMaxDepth() > MII) 1675 return; 1676 } 1677 NodeSets.clear(); 1678 LLVM_DEBUG(dbgs() << "Clear recurrence node-sets\n"); 1679 } 1680 1681 /// Add the nodes that do not belong to a recurrence set into groups 1682 /// based upon connected components. 1683 void SwingSchedulerDAG::groupRemainingNodes(NodeSetType &NodeSets) { 1684 SetVector<SUnit *> NodesAdded; 1685 SmallPtrSet<SUnit *, 8> Visited; 1686 // Add the nodes that are on a path between the previous node sets and 1687 // the current node set. 1688 for (NodeSet &I : NodeSets) { 1689 SmallSetVector<SUnit *, 8> N; 1690 // Add the nodes from the current node set to the previous node set. 1691 if (succ_L(I, N)) { 1692 SetVector<SUnit *> Path; 1693 for (SUnit *NI : N) { 1694 Visited.clear(); 1695 computePath(NI, Path, NodesAdded, I, Visited); 1696 } 1697 if (!Path.empty()) 1698 I.insert(Path.begin(), Path.end()); 1699 } 1700 // Add the nodes from the previous node set to the current node set. 1701 N.clear(); 1702 if (succ_L(NodesAdded, N)) { 1703 SetVector<SUnit *> Path; 1704 for (SUnit *NI : N) { 1705 Visited.clear(); 1706 computePath(NI, Path, I, NodesAdded, Visited); 1707 } 1708 if (!Path.empty()) 1709 I.insert(Path.begin(), Path.end()); 1710 } 1711 NodesAdded.insert(I.begin(), I.end()); 1712 } 1713 1714 // Create a new node set with the connected nodes of any successor of a node 1715 // in a recurrent set. 1716 NodeSet NewSet; 1717 SmallSetVector<SUnit *, 8> N; 1718 if (succ_L(NodesAdded, N)) 1719 for (SUnit *I : N) 1720 addConnectedNodes(I, NewSet, NodesAdded); 1721 if (!NewSet.empty()) 1722 NodeSets.push_back(NewSet); 1723 1724 // Create a new node set with the connected nodes of any predecessor of a node 1725 // in a recurrent set. 1726 NewSet.clear(); 1727 if (pred_L(NodesAdded, N)) 1728 for (SUnit *I : N) 1729 addConnectedNodes(I, NewSet, NodesAdded); 1730 if (!NewSet.empty()) 1731 NodeSets.push_back(NewSet); 1732 1733 // Create new nodes sets with the connected nodes any remaining node that 1734 // has no predecessor. 1735 for (SUnit &SU : SUnits) { 1736 if (NodesAdded.count(&SU) == 0) { 1737 NewSet.clear(); 1738 addConnectedNodes(&SU, NewSet, NodesAdded); 1739 if (!NewSet.empty()) 1740 NodeSets.push_back(NewSet); 1741 } 1742 } 1743 } 1744 1745 /// Add the node to the set, and add all of its connected nodes to the set. 1746 void SwingSchedulerDAG::addConnectedNodes(SUnit *SU, NodeSet &NewSet, 1747 SetVector<SUnit *> &NodesAdded) { 1748 NewSet.insert(SU); 1749 NodesAdded.insert(SU); 1750 for (auto &SI : SU->Succs) { 1751 SUnit *Successor = SI.getSUnit(); 1752 if (!SI.isArtificial() && !Successor->isBoundaryNode() && 1753 NodesAdded.count(Successor) == 0) 1754 addConnectedNodes(Successor, NewSet, NodesAdded); 1755 } 1756 for (auto &PI : SU->Preds) { 1757 SUnit *Predecessor = PI.getSUnit(); 1758 if (!PI.isArtificial() && NodesAdded.count(Predecessor) == 0) 1759 addConnectedNodes(Predecessor, NewSet, NodesAdded); 1760 } 1761 } 1762 1763 /// Return true if Set1 contains elements in Set2. The elements in common 1764 /// are returned in a different container. 1765 static bool isIntersect(SmallSetVector<SUnit *, 8> &Set1, const NodeSet &Set2, 1766 SmallSetVector<SUnit *, 8> &Result) { 1767 Result.clear(); 1768 for (SUnit *SU : Set1) { 1769 if (Set2.count(SU) != 0) 1770 Result.insert(SU); 1771 } 1772 return !Result.empty(); 1773 } 1774 1775 /// Merge the recurrence node sets that have the same initial node. 1776 void SwingSchedulerDAG::fuseRecs(NodeSetType &NodeSets) { 1777 for (NodeSetType::iterator I = NodeSets.begin(), E = NodeSets.end(); I != E; 1778 ++I) { 1779 NodeSet &NI = *I; 1780 for (NodeSetType::iterator J = I + 1; J != E;) { 1781 NodeSet &NJ = *J; 1782 if (NI.getNode(0)->NodeNum == NJ.getNode(0)->NodeNum) { 1783 if (NJ.compareRecMII(NI) > 0) 1784 NI.setRecMII(NJ.getRecMII()); 1785 for (SUnit *SU : *J) 1786 I->insert(SU); 1787 NodeSets.erase(J); 1788 E = NodeSets.end(); 1789 } else { 1790 ++J; 1791 } 1792 } 1793 } 1794 } 1795 1796 /// Remove nodes that have been scheduled in previous NodeSets. 1797 void SwingSchedulerDAG::removeDuplicateNodes(NodeSetType &NodeSets) { 1798 for (NodeSetType::iterator I = NodeSets.begin(), E = NodeSets.end(); I != E; 1799 ++I) 1800 for (NodeSetType::iterator J = I + 1; J != E;) { 1801 J->remove_if([&](SUnit *SUJ) { return I->count(SUJ); }); 1802 1803 if (J->empty()) { 1804 NodeSets.erase(J); 1805 E = NodeSets.end(); 1806 } else { 1807 ++J; 1808 } 1809 } 1810 } 1811 1812 /// Compute an ordered list of the dependence graph nodes, which 1813 /// indicates the order that the nodes will be scheduled. This is a 1814 /// two-level algorithm. First, a partial order is created, which 1815 /// consists of a list of sets ordered from highest to lowest priority. 1816 void SwingSchedulerDAG::computeNodeOrder(NodeSetType &NodeSets) { 1817 SmallSetVector<SUnit *, 8> R; 1818 NodeOrder.clear(); 1819 1820 for (auto &Nodes : NodeSets) { 1821 LLVM_DEBUG(dbgs() << "NodeSet size " << Nodes.size() << "\n"); 1822 OrderKind Order; 1823 SmallSetVector<SUnit *, 8> N; 1824 if (pred_L(NodeOrder, N) && llvm::set_is_subset(N, Nodes)) { 1825 R.insert(N.begin(), N.end()); 1826 Order = BottomUp; 1827 LLVM_DEBUG(dbgs() << " Bottom up (preds) "); 1828 } else if (succ_L(NodeOrder, N) && llvm::set_is_subset(N, Nodes)) { 1829 R.insert(N.begin(), N.end()); 1830 Order = TopDown; 1831 LLVM_DEBUG(dbgs() << " Top down (succs) "); 1832 } else if (isIntersect(N, Nodes, R)) { 1833 // If some of the successors are in the existing node-set, then use the 1834 // top-down ordering. 1835 Order = TopDown; 1836 LLVM_DEBUG(dbgs() << " Top down (intersect) "); 1837 } else if (NodeSets.size() == 1) { 1838 for (const auto &N : Nodes) 1839 if (N->Succs.size() == 0) 1840 R.insert(N); 1841 Order = BottomUp; 1842 LLVM_DEBUG(dbgs() << " Bottom up (all) "); 1843 } else { 1844 // Find the node with the highest ASAP. 1845 SUnit *maxASAP = nullptr; 1846 for (SUnit *SU : Nodes) { 1847 if (maxASAP == nullptr || getASAP(SU) > getASAP(maxASAP) || 1848 (getASAP(SU) == getASAP(maxASAP) && SU->NodeNum > maxASAP->NodeNum)) 1849 maxASAP = SU; 1850 } 1851 R.insert(maxASAP); 1852 Order = BottomUp; 1853 LLVM_DEBUG(dbgs() << " Bottom up (default) "); 1854 } 1855 1856 while (!R.empty()) { 1857 if (Order == TopDown) { 1858 // Choose the node with the maximum height. If more than one, choose 1859 // the node wiTH the maximum ZeroLatencyHeight. If still more than one, 1860 // choose the node with the lowest MOV. 1861 while (!R.empty()) { 1862 SUnit *maxHeight = nullptr; 1863 for (SUnit *I : R) { 1864 if (maxHeight == nullptr || getHeight(I) > getHeight(maxHeight)) 1865 maxHeight = I; 1866 else if (getHeight(I) == getHeight(maxHeight) && 1867 getZeroLatencyHeight(I) > getZeroLatencyHeight(maxHeight)) 1868 maxHeight = I; 1869 else if (getHeight(I) == getHeight(maxHeight) && 1870 getZeroLatencyHeight(I) == 1871 getZeroLatencyHeight(maxHeight) && 1872 getMOV(I) < getMOV(maxHeight)) 1873 maxHeight = I; 1874 } 1875 NodeOrder.insert(maxHeight); 1876 LLVM_DEBUG(dbgs() << maxHeight->NodeNum << " "); 1877 R.remove(maxHeight); 1878 for (const auto &I : maxHeight->Succs) { 1879 if (Nodes.count(I.getSUnit()) == 0) 1880 continue; 1881 if (NodeOrder.contains(I.getSUnit())) 1882 continue; 1883 if (ignoreDependence(I, false)) 1884 continue; 1885 R.insert(I.getSUnit()); 1886 } 1887 // Back-edges are predecessors with an anti-dependence. 1888 for (const auto &I : maxHeight->Preds) { 1889 if (I.getKind() != SDep::Anti) 1890 continue; 1891 if (Nodes.count(I.getSUnit()) == 0) 1892 continue; 1893 if (NodeOrder.contains(I.getSUnit())) 1894 continue; 1895 R.insert(I.getSUnit()); 1896 } 1897 } 1898 Order = BottomUp; 1899 LLVM_DEBUG(dbgs() << "\n Switching order to bottom up "); 1900 SmallSetVector<SUnit *, 8> N; 1901 if (pred_L(NodeOrder, N, &Nodes)) 1902 R.insert(N.begin(), N.end()); 1903 } else { 1904 // Choose the node with the maximum depth. If more than one, choose 1905 // the node with the maximum ZeroLatencyDepth. If still more than one, 1906 // choose the node with the lowest MOV. 1907 while (!R.empty()) { 1908 SUnit *maxDepth = nullptr; 1909 for (SUnit *I : R) { 1910 if (maxDepth == nullptr || getDepth(I) > getDepth(maxDepth)) 1911 maxDepth = I; 1912 else if (getDepth(I) == getDepth(maxDepth) && 1913 getZeroLatencyDepth(I) > getZeroLatencyDepth(maxDepth)) 1914 maxDepth = I; 1915 else if (getDepth(I) == getDepth(maxDepth) && 1916 getZeroLatencyDepth(I) == getZeroLatencyDepth(maxDepth) && 1917 getMOV(I) < getMOV(maxDepth)) 1918 maxDepth = I; 1919 } 1920 NodeOrder.insert(maxDepth); 1921 LLVM_DEBUG(dbgs() << maxDepth->NodeNum << " "); 1922 R.remove(maxDepth); 1923 if (Nodes.isExceedSU(maxDepth)) { 1924 Order = TopDown; 1925 R.clear(); 1926 R.insert(Nodes.getNode(0)); 1927 break; 1928 } 1929 for (const auto &I : maxDepth->Preds) { 1930 if (Nodes.count(I.getSUnit()) == 0) 1931 continue; 1932 if (NodeOrder.contains(I.getSUnit())) 1933 continue; 1934 R.insert(I.getSUnit()); 1935 } 1936 // Back-edges are predecessors with an anti-dependence. 1937 for (const auto &I : maxDepth->Succs) { 1938 if (I.getKind() != SDep::Anti) 1939 continue; 1940 if (Nodes.count(I.getSUnit()) == 0) 1941 continue; 1942 if (NodeOrder.contains(I.getSUnit())) 1943 continue; 1944 R.insert(I.getSUnit()); 1945 } 1946 } 1947 Order = TopDown; 1948 LLVM_DEBUG(dbgs() << "\n Switching order to top down "); 1949 SmallSetVector<SUnit *, 8> N; 1950 if (succ_L(NodeOrder, N, &Nodes)) 1951 R.insert(N.begin(), N.end()); 1952 } 1953 } 1954 LLVM_DEBUG(dbgs() << "\nDone with Nodeset\n"); 1955 } 1956 1957 LLVM_DEBUG({ 1958 dbgs() << "Node order: "; 1959 for (SUnit *I : NodeOrder) 1960 dbgs() << " " << I->NodeNum << " "; 1961 dbgs() << "\n"; 1962 }); 1963 } 1964 1965 /// Process the nodes in the computed order and create the pipelined schedule 1966 /// of the instructions, if possible. Return true if a schedule is found. 1967 bool SwingSchedulerDAG::schedulePipeline(SMSchedule &Schedule) { 1968 1969 if (NodeOrder.empty()){ 1970 LLVM_DEBUG(dbgs() << "NodeOrder is empty! abort scheduling\n" ); 1971 return false; 1972 } 1973 1974 bool scheduleFound = false; 1975 // Keep increasing II until a valid schedule is found. 1976 for (unsigned II = MII; II <= MAX_II && !scheduleFound; ++II) { 1977 Schedule.reset(); 1978 Schedule.setInitiationInterval(II); 1979 LLVM_DEBUG(dbgs() << "Try to schedule with " << II << "\n"); 1980 1981 SetVector<SUnit *>::iterator NI = NodeOrder.begin(); 1982 SetVector<SUnit *>::iterator NE = NodeOrder.end(); 1983 do { 1984 SUnit *SU = *NI; 1985 1986 // Compute the schedule time for the instruction, which is based 1987 // upon the scheduled time for any predecessors/successors. 1988 int EarlyStart = INT_MIN; 1989 int LateStart = INT_MAX; 1990 // These values are set when the size of the schedule window is limited 1991 // due to chain dependences. 1992 int SchedEnd = INT_MAX; 1993 int SchedStart = INT_MIN; 1994 Schedule.computeStart(SU, &EarlyStart, &LateStart, &SchedEnd, &SchedStart, 1995 II, this); 1996 LLVM_DEBUG({ 1997 dbgs() << "\n"; 1998 dbgs() << "Inst (" << SU->NodeNum << ") "; 1999 SU->getInstr()->dump(); 2000 dbgs() << "\n"; 2001 }); 2002 LLVM_DEBUG({ 2003 dbgs() << format("\tes: %8x ls: %8x me: %8x ms: %8x\n", EarlyStart, 2004 LateStart, SchedEnd, SchedStart); 2005 }); 2006 2007 if (EarlyStart > LateStart || SchedEnd < EarlyStart || 2008 SchedStart > LateStart) 2009 scheduleFound = false; 2010 else if (EarlyStart != INT_MIN && LateStart == INT_MAX) { 2011 SchedEnd = std::min(SchedEnd, EarlyStart + (int)II - 1); 2012 scheduleFound = Schedule.insert(SU, EarlyStart, SchedEnd, II); 2013 } else if (EarlyStart == INT_MIN && LateStart != INT_MAX) { 2014 SchedStart = std::max(SchedStart, LateStart - (int)II + 1); 2015 scheduleFound = Schedule.insert(SU, LateStart, SchedStart, II); 2016 } else if (EarlyStart != INT_MIN && LateStart != INT_MAX) { 2017 SchedEnd = 2018 std::min(SchedEnd, std::min(LateStart, EarlyStart + (int)II - 1)); 2019 // When scheduling a Phi it is better to start at the late cycle and go 2020 // backwards. The default order may insert the Phi too far away from 2021 // its first dependence. 2022 if (SU->getInstr()->isPHI()) 2023 scheduleFound = Schedule.insert(SU, SchedEnd, EarlyStart, II); 2024 else 2025 scheduleFound = Schedule.insert(SU, EarlyStart, SchedEnd, II); 2026 } else { 2027 int FirstCycle = Schedule.getFirstCycle(); 2028 scheduleFound = Schedule.insert(SU, FirstCycle + getASAP(SU), 2029 FirstCycle + getASAP(SU) + II - 1, II); 2030 } 2031 // Even if we find a schedule, make sure the schedule doesn't exceed the 2032 // allowable number of stages. We keep trying if this happens. 2033 if (scheduleFound) 2034 if (SwpMaxStages > -1 && 2035 Schedule.getMaxStageCount() > (unsigned)SwpMaxStages) 2036 scheduleFound = false; 2037 2038 LLVM_DEBUG({ 2039 if (!scheduleFound) 2040 dbgs() << "\tCan't schedule\n"; 2041 }); 2042 } while (++NI != NE && scheduleFound); 2043 2044 // If a schedule is found, ensure non-pipelined instructions are in stage 0 2045 if (scheduleFound) 2046 scheduleFound = 2047 Schedule.normalizeNonPipelinedInstructions(this, LoopPipelinerInfo); 2048 2049 // If a schedule is found, check if it is a valid schedule too. 2050 if (scheduleFound) 2051 scheduleFound = Schedule.isValidSchedule(this); 2052 } 2053 2054 LLVM_DEBUG(dbgs() << "Schedule Found? " << scheduleFound 2055 << " (II=" << Schedule.getInitiationInterval() 2056 << ")\n"); 2057 2058 if (scheduleFound) { 2059 scheduleFound = LoopPipelinerInfo->shouldUseSchedule(*this, Schedule); 2060 if (!scheduleFound) 2061 LLVM_DEBUG(dbgs() << "Target rejected schedule\n"); 2062 } 2063 2064 if (scheduleFound) { 2065 Schedule.finalizeSchedule(this); 2066 Pass.ORE->emit([&]() { 2067 return MachineOptimizationRemarkAnalysis( 2068 DEBUG_TYPE, "schedule", Loop.getStartLoc(), Loop.getHeader()) 2069 << "Schedule found with Initiation Interval: " 2070 << ore::NV("II", Schedule.getInitiationInterval()) 2071 << ", MaxStageCount: " 2072 << ore::NV("MaxStageCount", Schedule.getMaxStageCount()); 2073 }); 2074 } else 2075 Schedule.reset(); 2076 2077 return scheduleFound && Schedule.getMaxStageCount() > 0; 2078 } 2079 2080 /// Return true if we can compute the amount the instruction changes 2081 /// during each iteration. Set Delta to the amount of the change. 2082 bool SwingSchedulerDAG::computeDelta(MachineInstr &MI, unsigned &Delta) { 2083 const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo(); 2084 const MachineOperand *BaseOp; 2085 int64_t Offset; 2086 bool OffsetIsScalable; 2087 if (!TII->getMemOperandWithOffset(MI, BaseOp, Offset, OffsetIsScalable, TRI)) 2088 return false; 2089 2090 // FIXME: This algorithm assumes instructions have fixed-size offsets. 2091 if (OffsetIsScalable) 2092 return false; 2093 2094 if (!BaseOp->isReg()) 2095 return false; 2096 2097 Register BaseReg = BaseOp->getReg(); 2098 2099 MachineRegisterInfo &MRI = MF.getRegInfo(); 2100 // Check if there is a Phi. If so, get the definition in the loop. 2101 MachineInstr *BaseDef = MRI.getVRegDef(BaseReg); 2102 if (BaseDef && BaseDef->isPHI()) { 2103 BaseReg = getLoopPhiReg(*BaseDef, MI.getParent()); 2104 BaseDef = MRI.getVRegDef(BaseReg); 2105 } 2106 if (!BaseDef) 2107 return false; 2108 2109 int D = 0; 2110 if (!TII->getIncrementValue(*BaseDef, D) && D >= 0) 2111 return false; 2112 2113 Delta = D; 2114 return true; 2115 } 2116 2117 /// Check if we can change the instruction to use an offset value from the 2118 /// previous iteration. If so, return true and set the base and offset values 2119 /// so that we can rewrite the load, if necessary. 2120 /// v1 = Phi(v0, v3) 2121 /// v2 = load v1, 0 2122 /// v3 = post_store v1, 4, x 2123 /// This function enables the load to be rewritten as v2 = load v3, 4. 2124 bool SwingSchedulerDAG::canUseLastOffsetValue(MachineInstr *MI, 2125 unsigned &BasePos, 2126 unsigned &OffsetPos, 2127 unsigned &NewBase, 2128 int64_t &Offset) { 2129 // Get the load instruction. 2130 if (TII->isPostIncrement(*MI)) 2131 return false; 2132 unsigned BasePosLd, OffsetPosLd; 2133 if (!TII->getBaseAndOffsetPosition(*MI, BasePosLd, OffsetPosLd)) 2134 return false; 2135 Register BaseReg = MI->getOperand(BasePosLd).getReg(); 2136 2137 // Look for the Phi instruction. 2138 MachineRegisterInfo &MRI = MI->getMF()->getRegInfo(); 2139 MachineInstr *Phi = MRI.getVRegDef(BaseReg); 2140 if (!Phi || !Phi->isPHI()) 2141 return false; 2142 // Get the register defined in the loop block. 2143 unsigned PrevReg = getLoopPhiReg(*Phi, MI->getParent()); 2144 if (!PrevReg) 2145 return false; 2146 2147 // Check for the post-increment load/store instruction. 2148 MachineInstr *PrevDef = MRI.getVRegDef(PrevReg); 2149 if (!PrevDef || PrevDef == MI) 2150 return false; 2151 2152 if (!TII->isPostIncrement(*PrevDef)) 2153 return false; 2154 2155 unsigned BasePos1 = 0, OffsetPos1 = 0; 2156 if (!TII->getBaseAndOffsetPosition(*PrevDef, BasePos1, OffsetPos1)) 2157 return false; 2158 2159 // Make sure that the instructions do not access the same memory location in 2160 // the next iteration. 2161 int64_t LoadOffset = MI->getOperand(OffsetPosLd).getImm(); 2162 int64_t StoreOffset = PrevDef->getOperand(OffsetPos1).getImm(); 2163 MachineInstr *NewMI = MF.CloneMachineInstr(MI); 2164 NewMI->getOperand(OffsetPosLd).setImm(LoadOffset + StoreOffset); 2165 bool Disjoint = TII->areMemAccessesTriviallyDisjoint(*NewMI, *PrevDef); 2166 MF.deleteMachineInstr(NewMI); 2167 if (!Disjoint) 2168 return false; 2169 2170 // Set the return value once we determine that we return true. 2171 BasePos = BasePosLd; 2172 OffsetPos = OffsetPosLd; 2173 NewBase = PrevReg; 2174 Offset = StoreOffset; 2175 return true; 2176 } 2177 2178 /// Apply changes to the instruction if needed. The changes are need 2179 /// to improve the scheduling and depend up on the final schedule. 2180 void SwingSchedulerDAG::applyInstrChange(MachineInstr *MI, 2181 SMSchedule &Schedule) { 2182 SUnit *SU = getSUnit(MI); 2183 DenseMap<SUnit *, std::pair<unsigned, int64_t>>::iterator It = 2184 InstrChanges.find(SU); 2185 if (It != InstrChanges.end()) { 2186 std::pair<unsigned, int64_t> RegAndOffset = It->second; 2187 unsigned BasePos, OffsetPos; 2188 if (!TII->getBaseAndOffsetPosition(*MI, BasePos, OffsetPos)) 2189 return; 2190 Register BaseReg = MI->getOperand(BasePos).getReg(); 2191 MachineInstr *LoopDef = findDefInLoop(BaseReg); 2192 int DefStageNum = Schedule.stageScheduled(getSUnit(LoopDef)); 2193 int DefCycleNum = Schedule.cycleScheduled(getSUnit(LoopDef)); 2194 int BaseStageNum = Schedule.stageScheduled(SU); 2195 int BaseCycleNum = Schedule.cycleScheduled(SU); 2196 if (BaseStageNum < DefStageNum) { 2197 MachineInstr *NewMI = MF.CloneMachineInstr(MI); 2198 int OffsetDiff = DefStageNum - BaseStageNum; 2199 if (DefCycleNum < BaseCycleNum) { 2200 NewMI->getOperand(BasePos).setReg(RegAndOffset.first); 2201 if (OffsetDiff > 0) 2202 --OffsetDiff; 2203 } 2204 int64_t NewOffset = 2205 MI->getOperand(OffsetPos).getImm() + RegAndOffset.second * OffsetDiff; 2206 NewMI->getOperand(OffsetPos).setImm(NewOffset); 2207 SU->setInstr(NewMI); 2208 MISUnitMap[NewMI] = SU; 2209 NewMIs[MI] = NewMI; 2210 } 2211 } 2212 } 2213 2214 /// Return the instruction in the loop that defines the register. 2215 /// If the definition is a Phi, then follow the Phi operand to 2216 /// the instruction in the loop. 2217 MachineInstr *SwingSchedulerDAG::findDefInLoop(Register Reg) { 2218 SmallPtrSet<MachineInstr *, 8> Visited; 2219 MachineInstr *Def = MRI.getVRegDef(Reg); 2220 while (Def->isPHI()) { 2221 if (!Visited.insert(Def).second) 2222 break; 2223 for (unsigned i = 1, e = Def->getNumOperands(); i < e; i += 2) 2224 if (Def->getOperand(i + 1).getMBB() == BB) { 2225 Def = MRI.getVRegDef(Def->getOperand(i).getReg()); 2226 break; 2227 } 2228 } 2229 return Def; 2230 } 2231 2232 /// Return true for an order or output dependence that is loop carried 2233 /// potentially. A dependence is loop carried if the destination defines a valu 2234 /// that may be used or defined by the source in a subsequent iteration. 2235 bool SwingSchedulerDAG::isLoopCarriedDep(SUnit *Source, const SDep &Dep, 2236 bool isSucc) { 2237 if ((Dep.getKind() != SDep::Order && Dep.getKind() != SDep::Output) || 2238 Dep.isArtificial() || Dep.getSUnit()->isBoundaryNode()) 2239 return false; 2240 2241 if (!SwpPruneLoopCarried) 2242 return true; 2243 2244 if (Dep.getKind() == SDep::Output) 2245 return true; 2246 2247 MachineInstr *SI = Source->getInstr(); 2248 MachineInstr *DI = Dep.getSUnit()->getInstr(); 2249 if (!isSucc) 2250 std::swap(SI, DI); 2251 assert(SI != nullptr && DI != nullptr && "Expecting SUnit with an MI."); 2252 2253 // Assume ordered loads and stores may have a loop carried dependence. 2254 if (SI->hasUnmodeledSideEffects() || DI->hasUnmodeledSideEffects() || 2255 SI->mayRaiseFPException() || DI->mayRaiseFPException() || 2256 SI->hasOrderedMemoryRef() || DI->hasOrderedMemoryRef()) 2257 return true; 2258 2259 // Only chain dependences between a load and store can be loop carried. 2260 if (!DI->mayStore() || !SI->mayLoad()) 2261 return false; 2262 2263 unsigned DeltaS, DeltaD; 2264 if (!computeDelta(*SI, DeltaS) || !computeDelta(*DI, DeltaD)) 2265 return true; 2266 2267 const MachineOperand *BaseOpS, *BaseOpD; 2268 int64_t OffsetS, OffsetD; 2269 bool OffsetSIsScalable, OffsetDIsScalable; 2270 const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo(); 2271 if (!TII->getMemOperandWithOffset(*SI, BaseOpS, OffsetS, OffsetSIsScalable, 2272 TRI) || 2273 !TII->getMemOperandWithOffset(*DI, BaseOpD, OffsetD, OffsetDIsScalable, 2274 TRI)) 2275 return true; 2276 2277 assert(!OffsetSIsScalable && !OffsetDIsScalable && 2278 "Expected offsets to be byte offsets"); 2279 2280 MachineInstr *DefS = MRI.getVRegDef(BaseOpS->getReg()); 2281 MachineInstr *DefD = MRI.getVRegDef(BaseOpD->getReg()); 2282 if (!DefS || !DefD || !DefS->isPHI() || !DefD->isPHI()) 2283 return true; 2284 2285 unsigned InitValS = 0; 2286 unsigned LoopValS = 0; 2287 unsigned InitValD = 0; 2288 unsigned LoopValD = 0; 2289 getPhiRegs(*DefS, BB, InitValS, LoopValS); 2290 getPhiRegs(*DefD, BB, InitValD, LoopValD); 2291 MachineInstr *InitDefS = MRI.getVRegDef(InitValS); 2292 MachineInstr *InitDefD = MRI.getVRegDef(InitValD); 2293 2294 if (!InitDefS->isIdenticalTo(*InitDefD)) 2295 return true; 2296 2297 // Check that the base register is incremented by a constant value for each 2298 // iteration. 2299 MachineInstr *LoopDefS = MRI.getVRegDef(LoopValS); 2300 int D = 0; 2301 if (!LoopDefS || !TII->getIncrementValue(*LoopDefS, D)) 2302 return true; 2303 2304 uint64_t AccessSizeS = (*SI->memoperands_begin())->getSize(); 2305 uint64_t AccessSizeD = (*DI->memoperands_begin())->getSize(); 2306 2307 // This is the main test, which checks the offset values and the loop 2308 // increment value to determine if the accesses may be loop carried. 2309 if (AccessSizeS == MemoryLocation::UnknownSize || 2310 AccessSizeD == MemoryLocation::UnknownSize) 2311 return true; 2312 2313 if (DeltaS != DeltaD || DeltaS < AccessSizeS || DeltaD < AccessSizeD) 2314 return true; 2315 2316 return (OffsetS + (int64_t)AccessSizeS < OffsetD + (int64_t)AccessSizeD); 2317 } 2318 2319 void SwingSchedulerDAG::postprocessDAG() { 2320 for (auto &M : Mutations) 2321 M->apply(this); 2322 } 2323 2324 /// Try to schedule the node at the specified StartCycle and continue 2325 /// until the node is schedule or the EndCycle is reached. This function 2326 /// returns true if the node is scheduled. This routine may search either 2327 /// forward or backward for a place to insert the instruction based upon 2328 /// the relative values of StartCycle and EndCycle. 2329 bool SMSchedule::insert(SUnit *SU, int StartCycle, int EndCycle, int II) { 2330 bool forward = true; 2331 LLVM_DEBUG({ 2332 dbgs() << "Trying to insert node between " << StartCycle << " and " 2333 << EndCycle << " II: " << II << "\n"; 2334 }); 2335 if (StartCycle > EndCycle) 2336 forward = false; 2337 2338 // The terminating condition depends on the direction. 2339 int termCycle = forward ? EndCycle + 1 : EndCycle - 1; 2340 for (int curCycle = StartCycle; curCycle != termCycle; 2341 forward ? ++curCycle : --curCycle) { 2342 2343 if (ST.getInstrInfo()->isZeroCost(SU->getInstr()->getOpcode()) || 2344 ProcItinResources.canReserveResources(*SU, curCycle)) { 2345 LLVM_DEBUG({ 2346 dbgs() << "\tinsert at cycle " << curCycle << " "; 2347 SU->getInstr()->dump(); 2348 }); 2349 2350 if (!ST.getInstrInfo()->isZeroCost(SU->getInstr()->getOpcode())) 2351 ProcItinResources.reserveResources(*SU, curCycle); 2352 ScheduledInstrs[curCycle].push_back(SU); 2353 InstrToCycle.insert(std::make_pair(SU, curCycle)); 2354 if (curCycle > LastCycle) 2355 LastCycle = curCycle; 2356 if (curCycle < FirstCycle) 2357 FirstCycle = curCycle; 2358 return true; 2359 } 2360 LLVM_DEBUG({ 2361 dbgs() << "\tfailed to insert at cycle " << curCycle << " "; 2362 SU->getInstr()->dump(); 2363 }); 2364 } 2365 return false; 2366 } 2367 2368 // Return the cycle of the earliest scheduled instruction in the chain. 2369 int SMSchedule::earliestCycleInChain(const SDep &Dep) { 2370 SmallPtrSet<SUnit *, 8> Visited; 2371 SmallVector<SDep, 8> Worklist; 2372 Worklist.push_back(Dep); 2373 int EarlyCycle = INT_MAX; 2374 while (!Worklist.empty()) { 2375 const SDep &Cur = Worklist.pop_back_val(); 2376 SUnit *PrevSU = Cur.getSUnit(); 2377 if (Visited.count(PrevSU)) 2378 continue; 2379 std::map<SUnit *, int>::const_iterator it = InstrToCycle.find(PrevSU); 2380 if (it == InstrToCycle.end()) 2381 continue; 2382 EarlyCycle = std::min(EarlyCycle, it->second); 2383 for (const auto &PI : PrevSU->Preds) 2384 if (PI.getKind() == SDep::Order || PI.getKind() == SDep::Output) 2385 Worklist.push_back(PI); 2386 Visited.insert(PrevSU); 2387 } 2388 return EarlyCycle; 2389 } 2390 2391 // Return the cycle of the latest scheduled instruction in the chain. 2392 int SMSchedule::latestCycleInChain(const SDep &Dep) { 2393 SmallPtrSet<SUnit *, 8> Visited; 2394 SmallVector<SDep, 8> Worklist; 2395 Worklist.push_back(Dep); 2396 int LateCycle = INT_MIN; 2397 while (!Worklist.empty()) { 2398 const SDep &Cur = Worklist.pop_back_val(); 2399 SUnit *SuccSU = Cur.getSUnit(); 2400 if (Visited.count(SuccSU) || SuccSU->isBoundaryNode()) 2401 continue; 2402 std::map<SUnit *, int>::const_iterator it = InstrToCycle.find(SuccSU); 2403 if (it == InstrToCycle.end()) 2404 continue; 2405 LateCycle = std::max(LateCycle, it->second); 2406 for (const auto &SI : SuccSU->Succs) 2407 if (SI.getKind() == SDep::Order || SI.getKind() == SDep::Output) 2408 Worklist.push_back(SI); 2409 Visited.insert(SuccSU); 2410 } 2411 return LateCycle; 2412 } 2413 2414 /// If an instruction has a use that spans multiple iterations, then 2415 /// return true. These instructions are characterized by having a back-ege 2416 /// to a Phi, which contains a reference to another Phi. 2417 static SUnit *multipleIterations(SUnit *SU, SwingSchedulerDAG *DAG) { 2418 for (auto &P : SU->Preds) 2419 if (DAG->isBackedge(SU, P) && P.getSUnit()->getInstr()->isPHI()) 2420 for (auto &S : P.getSUnit()->Succs) 2421 if (S.getKind() == SDep::Data && S.getSUnit()->getInstr()->isPHI()) 2422 return P.getSUnit(); 2423 return nullptr; 2424 } 2425 2426 /// Compute the scheduling start slot for the instruction. The start slot 2427 /// depends on any predecessor or successor nodes scheduled already. 2428 void SMSchedule::computeStart(SUnit *SU, int *MaxEarlyStart, int *MinLateStart, 2429 int *MinEnd, int *MaxStart, int II, 2430 SwingSchedulerDAG *DAG) { 2431 // Iterate over each instruction that has been scheduled already. The start 2432 // slot computation depends on whether the previously scheduled instruction 2433 // is a predecessor or successor of the specified instruction. 2434 for (int cycle = getFirstCycle(); cycle <= LastCycle; ++cycle) { 2435 2436 // Iterate over each instruction in the current cycle. 2437 for (SUnit *I : getInstructions(cycle)) { 2438 // Because we're processing a DAG for the dependences, we recognize 2439 // the back-edge in recurrences by anti dependences. 2440 for (unsigned i = 0, e = (unsigned)SU->Preds.size(); i != e; ++i) { 2441 const SDep &Dep = SU->Preds[i]; 2442 if (Dep.getSUnit() == I) { 2443 if (!DAG->isBackedge(SU, Dep)) { 2444 int EarlyStart = cycle + Dep.getLatency() - 2445 DAG->getDistance(Dep.getSUnit(), SU, Dep) * II; 2446 *MaxEarlyStart = std::max(*MaxEarlyStart, EarlyStart); 2447 if (DAG->isLoopCarriedDep(SU, Dep, false)) { 2448 int End = earliestCycleInChain(Dep) + (II - 1); 2449 *MinEnd = std::min(*MinEnd, End); 2450 } 2451 } else { 2452 int LateStart = cycle - Dep.getLatency() + 2453 DAG->getDistance(SU, Dep.getSUnit(), Dep) * II; 2454 *MinLateStart = std::min(*MinLateStart, LateStart); 2455 } 2456 } 2457 // For instruction that requires multiple iterations, make sure that 2458 // the dependent instruction is not scheduled past the definition. 2459 SUnit *BE = multipleIterations(I, DAG); 2460 if (BE && Dep.getSUnit() == BE && !SU->getInstr()->isPHI() && 2461 !SU->isPred(I)) 2462 *MinLateStart = std::min(*MinLateStart, cycle); 2463 } 2464 for (unsigned i = 0, e = (unsigned)SU->Succs.size(); i != e; ++i) { 2465 if (SU->Succs[i].getSUnit() == I) { 2466 const SDep &Dep = SU->Succs[i]; 2467 if (!DAG->isBackedge(SU, Dep)) { 2468 int LateStart = cycle - Dep.getLatency() + 2469 DAG->getDistance(SU, Dep.getSUnit(), Dep) * II; 2470 *MinLateStart = std::min(*MinLateStart, LateStart); 2471 if (DAG->isLoopCarriedDep(SU, Dep)) { 2472 int Start = latestCycleInChain(Dep) + 1 - II; 2473 *MaxStart = std::max(*MaxStart, Start); 2474 } 2475 } else { 2476 int EarlyStart = cycle + Dep.getLatency() - 2477 DAG->getDistance(Dep.getSUnit(), SU, Dep) * II; 2478 *MaxEarlyStart = std::max(*MaxEarlyStart, EarlyStart); 2479 } 2480 } 2481 } 2482 } 2483 } 2484 } 2485 2486 /// Order the instructions within a cycle so that the definitions occur 2487 /// before the uses. Returns true if the instruction is added to the start 2488 /// of the list, or false if added to the end. 2489 void SMSchedule::orderDependence(SwingSchedulerDAG *SSD, SUnit *SU, 2490 std::deque<SUnit *> &Insts) { 2491 MachineInstr *MI = SU->getInstr(); 2492 bool OrderBeforeUse = false; 2493 bool OrderAfterDef = false; 2494 bool OrderBeforeDef = false; 2495 unsigned MoveDef = 0; 2496 unsigned MoveUse = 0; 2497 int StageInst1 = stageScheduled(SU); 2498 2499 unsigned Pos = 0; 2500 for (std::deque<SUnit *>::iterator I = Insts.begin(), E = Insts.end(); I != E; 2501 ++I, ++Pos) { 2502 for (MachineOperand &MO : MI->operands()) { 2503 if (!MO.isReg() || !MO.getReg().isVirtual()) 2504 continue; 2505 2506 Register Reg = MO.getReg(); 2507 unsigned BasePos, OffsetPos; 2508 if (ST.getInstrInfo()->getBaseAndOffsetPosition(*MI, BasePos, OffsetPos)) 2509 if (MI->getOperand(BasePos).getReg() == Reg) 2510 if (unsigned NewReg = SSD->getInstrBaseReg(SU)) 2511 Reg = NewReg; 2512 bool Reads, Writes; 2513 std::tie(Reads, Writes) = 2514 (*I)->getInstr()->readsWritesVirtualRegister(Reg); 2515 if (MO.isDef() && Reads && stageScheduled(*I) <= StageInst1) { 2516 OrderBeforeUse = true; 2517 if (MoveUse == 0) 2518 MoveUse = Pos; 2519 } else if (MO.isDef() && Reads && stageScheduled(*I) > StageInst1) { 2520 // Add the instruction after the scheduled instruction. 2521 OrderAfterDef = true; 2522 MoveDef = Pos; 2523 } else if (MO.isUse() && Writes && stageScheduled(*I) == StageInst1) { 2524 if (cycleScheduled(*I) == cycleScheduled(SU) && !(*I)->isSucc(SU)) { 2525 OrderBeforeUse = true; 2526 if (MoveUse == 0) 2527 MoveUse = Pos; 2528 } else { 2529 OrderAfterDef = true; 2530 MoveDef = Pos; 2531 } 2532 } else if (MO.isUse() && Writes && stageScheduled(*I) > StageInst1) { 2533 OrderBeforeUse = true; 2534 if (MoveUse == 0) 2535 MoveUse = Pos; 2536 if (MoveUse != 0) { 2537 OrderAfterDef = true; 2538 MoveDef = Pos - 1; 2539 } 2540 } else if (MO.isUse() && Writes && stageScheduled(*I) < StageInst1) { 2541 // Add the instruction before the scheduled instruction. 2542 OrderBeforeUse = true; 2543 if (MoveUse == 0) 2544 MoveUse = Pos; 2545 } else if (MO.isUse() && stageScheduled(*I) == StageInst1 && 2546 isLoopCarriedDefOfUse(SSD, (*I)->getInstr(), MO)) { 2547 if (MoveUse == 0) { 2548 OrderBeforeDef = true; 2549 MoveUse = Pos; 2550 } 2551 } 2552 } 2553 // Check for order dependences between instructions. Make sure the source 2554 // is ordered before the destination. 2555 for (auto &S : SU->Succs) { 2556 if (S.getSUnit() != *I) 2557 continue; 2558 if (S.getKind() == SDep::Order && stageScheduled(*I) == StageInst1) { 2559 OrderBeforeUse = true; 2560 if (Pos < MoveUse) 2561 MoveUse = Pos; 2562 } 2563 // We did not handle HW dependences in previous for loop, 2564 // and we normally set Latency = 0 for Anti deps, 2565 // so may have nodes in same cycle with Anti denpendent on HW regs. 2566 else if (S.getKind() == SDep::Anti && stageScheduled(*I) == StageInst1) { 2567 OrderBeforeUse = true; 2568 if ((MoveUse == 0) || (Pos < MoveUse)) 2569 MoveUse = Pos; 2570 } 2571 } 2572 for (auto &P : SU->Preds) { 2573 if (P.getSUnit() != *I) 2574 continue; 2575 if (P.getKind() == SDep::Order && stageScheduled(*I) == StageInst1) { 2576 OrderAfterDef = true; 2577 MoveDef = Pos; 2578 } 2579 } 2580 } 2581 2582 // A circular dependence. 2583 if (OrderAfterDef && OrderBeforeUse && MoveUse == MoveDef) 2584 OrderBeforeUse = false; 2585 2586 // OrderAfterDef takes precedences over OrderBeforeDef. The latter is due 2587 // to a loop-carried dependence. 2588 if (OrderBeforeDef) 2589 OrderBeforeUse = !OrderAfterDef || (MoveUse > MoveDef); 2590 2591 // The uncommon case when the instruction order needs to be updated because 2592 // there is both a use and def. 2593 if (OrderBeforeUse && OrderAfterDef) { 2594 SUnit *UseSU = Insts.at(MoveUse); 2595 SUnit *DefSU = Insts.at(MoveDef); 2596 if (MoveUse > MoveDef) { 2597 Insts.erase(Insts.begin() + MoveUse); 2598 Insts.erase(Insts.begin() + MoveDef); 2599 } else { 2600 Insts.erase(Insts.begin() + MoveDef); 2601 Insts.erase(Insts.begin() + MoveUse); 2602 } 2603 orderDependence(SSD, UseSU, Insts); 2604 orderDependence(SSD, SU, Insts); 2605 orderDependence(SSD, DefSU, Insts); 2606 return; 2607 } 2608 // Put the new instruction first if there is a use in the list. Otherwise, 2609 // put it at the end of the list. 2610 if (OrderBeforeUse) 2611 Insts.push_front(SU); 2612 else 2613 Insts.push_back(SU); 2614 } 2615 2616 /// Return true if the scheduled Phi has a loop carried operand. 2617 bool SMSchedule::isLoopCarried(SwingSchedulerDAG *SSD, MachineInstr &Phi) { 2618 if (!Phi.isPHI()) 2619 return false; 2620 assert(Phi.isPHI() && "Expecting a Phi."); 2621 SUnit *DefSU = SSD->getSUnit(&Phi); 2622 unsigned DefCycle = cycleScheduled(DefSU); 2623 int DefStage = stageScheduled(DefSU); 2624 2625 unsigned InitVal = 0; 2626 unsigned LoopVal = 0; 2627 getPhiRegs(Phi, Phi.getParent(), InitVal, LoopVal); 2628 SUnit *UseSU = SSD->getSUnit(MRI.getVRegDef(LoopVal)); 2629 if (!UseSU) 2630 return true; 2631 if (UseSU->getInstr()->isPHI()) 2632 return true; 2633 unsigned LoopCycle = cycleScheduled(UseSU); 2634 int LoopStage = stageScheduled(UseSU); 2635 return (LoopCycle > DefCycle) || (LoopStage <= DefStage); 2636 } 2637 2638 /// Return true if the instruction is a definition that is loop carried 2639 /// and defines the use on the next iteration. 2640 /// v1 = phi(v2, v3) 2641 /// (Def) v3 = op v1 2642 /// (MO) = v1 2643 /// If MO appears before Def, then then v1 and v3 may get assigned to the same 2644 /// register. 2645 bool SMSchedule::isLoopCarriedDefOfUse(SwingSchedulerDAG *SSD, 2646 MachineInstr *Def, MachineOperand &MO) { 2647 if (!MO.isReg()) 2648 return false; 2649 if (Def->isPHI()) 2650 return false; 2651 MachineInstr *Phi = MRI.getVRegDef(MO.getReg()); 2652 if (!Phi || !Phi->isPHI() || Phi->getParent() != Def->getParent()) 2653 return false; 2654 if (!isLoopCarried(SSD, *Phi)) 2655 return false; 2656 unsigned LoopReg = getLoopPhiReg(*Phi, Phi->getParent()); 2657 for (unsigned i = 0, e = Def->getNumOperands(); i != e; ++i) { 2658 MachineOperand &DMO = Def->getOperand(i); 2659 if (!DMO.isReg() || !DMO.isDef()) 2660 continue; 2661 if (DMO.getReg() == LoopReg) 2662 return true; 2663 } 2664 return false; 2665 } 2666 2667 /// Determine transitive dependences of unpipelineable instructions 2668 SmallSet<SUnit *, 8> SMSchedule::computeUnpipelineableNodes( 2669 SwingSchedulerDAG *SSD, TargetInstrInfo::PipelinerLoopInfo *PLI) { 2670 SmallSet<SUnit *, 8> DoNotPipeline; 2671 SmallVector<SUnit *, 8> Worklist; 2672 2673 for (auto &SU : SSD->SUnits) 2674 if (SU.isInstr() && PLI->shouldIgnoreForPipelining(SU.getInstr())) 2675 Worklist.push_back(&SU); 2676 2677 while (!Worklist.empty()) { 2678 auto SU = Worklist.pop_back_val(); 2679 if (DoNotPipeline.count(SU)) 2680 continue; 2681 LLVM_DEBUG(dbgs() << "Do not pipeline SU(" << SU->NodeNum << ")\n"); 2682 DoNotPipeline.insert(SU); 2683 for (auto &Dep : SU->Preds) 2684 Worklist.push_back(Dep.getSUnit()); 2685 if (SU->getInstr()->isPHI()) 2686 for (auto &Dep : SU->Succs) 2687 if (Dep.getKind() == SDep::Anti) 2688 Worklist.push_back(Dep.getSUnit()); 2689 } 2690 return DoNotPipeline; 2691 } 2692 2693 // Determine all instructions upon which any unpipelineable instruction depends 2694 // and ensure that they are in stage 0. If unable to do so, return false. 2695 bool SMSchedule::normalizeNonPipelinedInstructions( 2696 SwingSchedulerDAG *SSD, TargetInstrInfo::PipelinerLoopInfo *PLI) { 2697 SmallSet<SUnit *, 8> DNP = computeUnpipelineableNodes(SSD, PLI); 2698 2699 int NewLastCycle = INT_MIN; 2700 for (SUnit &SU : SSD->SUnits) { 2701 if (!SU.isInstr()) 2702 continue; 2703 if (!DNP.contains(&SU) || stageScheduled(&SU) == 0) { 2704 NewLastCycle = std::max(NewLastCycle, InstrToCycle[&SU]); 2705 continue; 2706 } 2707 2708 // Put the non-pipelined instruction as early as possible in the schedule 2709 int NewCycle = getFirstCycle(); 2710 for (auto &Dep : SU.Preds) 2711 NewCycle = std::max(InstrToCycle[Dep.getSUnit()], NewCycle); 2712 2713 int OldCycle = InstrToCycle[&SU]; 2714 if (OldCycle != NewCycle) { 2715 InstrToCycle[&SU] = NewCycle; 2716 auto &OldS = getInstructions(OldCycle); 2717 llvm::erase_value(OldS, &SU); 2718 getInstructions(NewCycle).emplace_back(&SU); 2719 LLVM_DEBUG(dbgs() << "SU(" << SU.NodeNum 2720 << ") is not pipelined; moving from cycle " << OldCycle 2721 << " to " << NewCycle << " Instr:" << *SU.getInstr()); 2722 } 2723 NewLastCycle = std::max(NewLastCycle, NewCycle); 2724 } 2725 LastCycle = NewLastCycle; 2726 return true; 2727 } 2728 2729 // Check if the generated schedule is valid. This function checks if 2730 // an instruction that uses a physical register is scheduled in a 2731 // different stage than the definition. The pipeliner does not handle 2732 // physical register values that may cross a basic block boundary. 2733 // Furthermore, if a physical def/use pair is assigned to the same 2734 // cycle, orderDependence does not guarantee def/use ordering, so that 2735 // case should be considered invalid. (The test checks for both 2736 // earlier and same-cycle use to be more robust.) 2737 bool SMSchedule::isValidSchedule(SwingSchedulerDAG *SSD) { 2738 for (SUnit &SU : SSD->SUnits) { 2739 if (!SU.hasPhysRegDefs) 2740 continue; 2741 int StageDef = stageScheduled(&SU); 2742 int CycleDef = InstrToCycle[&SU]; 2743 assert(StageDef != -1 && "Instruction should have been scheduled."); 2744 for (auto &SI : SU.Succs) 2745 if (SI.isAssignedRegDep() && !SI.getSUnit()->isBoundaryNode()) 2746 if (Register::isPhysicalRegister(SI.getReg())) { 2747 if (stageScheduled(SI.getSUnit()) != StageDef) 2748 return false; 2749 if (InstrToCycle[SI.getSUnit()] <= CycleDef) 2750 return false; 2751 } 2752 } 2753 return true; 2754 } 2755 2756 /// A property of the node order in swing-modulo-scheduling is 2757 /// that for nodes outside circuits the following holds: 2758 /// none of them is scheduled after both a successor and a 2759 /// predecessor. 2760 /// The method below checks whether the property is met. 2761 /// If not, debug information is printed and statistics information updated. 2762 /// Note that we do not use an assert statement. 2763 /// The reason is that although an invalid node oder may prevent 2764 /// the pipeliner from finding a pipelined schedule for arbitrary II, 2765 /// it does not lead to the generation of incorrect code. 2766 void SwingSchedulerDAG::checkValidNodeOrder(const NodeSetType &Circuits) const { 2767 2768 // a sorted vector that maps each SUnit to its index in the NodeOrder 2769 typedef std::pair<SUnit *, unsigned> UnitIndex; 2770 std::vector<UnitIndex> Indices(NodeOrder.size(), std::make_pair(nullptr, 0)); 2771 2772 for (unsigned i = 0, s = NodeOrder.size(); i < s; ++i) 2773 Indices.push_back(std::make_pair(NodeOrder[i], i)); 2774 2775 auto CompareKey = [](UnitIndex i1, UnitIndex i2) { 2776 return std::get<0>(i1) < std::get<0>(i2); 2777 }; 2778 2779 // sort, so that we can perform a binary search 2780 llvm::sort(Indices, CompareKey); 2781 2782 bool Valid = true; 2783 (void)Valid; 2784 // for each SUnit in the NodeOrder, check whether 2785 // it appears after both a successor and a predecessor 2786 // of the SUnit. If this is the case, and the SUnit 2787 // is not part of circuit, then the NodeOrder is not 2788 // valid. 2789 for (unsigned i = 0, s = NodeOrder.size(); i < s; ++i) { 2790 SUnit *SU = NodeOrder[i]; 2791 unsigned Index = i; 2792 2793 bool PredBefore = false; 2794 bool SuccBefore = false; 2795 2796 SUnit *Succ; 2797 SUnit *Pred; 2798 (void)Succ; 2799 (void)Pred; 2800 2801 for (SDep &PredEdge : SU->Preds) { 2802 SUnit *PredSU = PredEdge.getSUnit(); 2803 unsigned PredIndex = std::get<1>( 2804 *llvm::lower_bound(Indices, std::make_pair(PredSU, 0), CompareKey)); 2805 if (!PredSU->getInstr()->isPHI() && PredIndex < Index) { 2806 PredBefore = true; 2807 Pred = PredSU; 2808 break; 2809 } 2810 } 2811 2812 for (SDep &SuccEdge : SU->Succs) { 2813 SUnit *SuccSU = SuccEdge.getSUnit(); 2814 // Do not process a boundary node, it was not included in NodeOrder, 2815 // hence not in Indices either, call to std::lower_bound() below will 2816 // return Indices.end(). 2817 if (SuccSU->isBoundaryNode()) 2818 continue; 2819 unsigned SuccIndex = std::get<1>( 2820 *llvm::lower_bound(Indices, std::make_pair(SuccSU, 0), CompareKey)); 2821 if (!SuccSU->getInstr()->isPHI() && SuccIndex < Index) { 2822 SuccBefore = true; 2823 Succ = SuccSU; 2824 break; 2825 } 2826 } 2827 2828 if (PredBefore && SuccBefore && !SU->getInstr()->isPHI()) { 2829 // instructions in circuits are allowed to be scheduled 2830 // after both a successor and predecessor. 2831 bool InCircuit = llvm::any_of( 2832 Circuits, [SU](const NodeSet &Circuit) { return Circuit.count(SU); }); 2833 if (InCircuit) 2834 LLVM_DEBUG(dbgs() << "In a circuit, predecessor ";); 2835 else { 2836 Valid = false; 2837 NumNodeOrderIssues++; 2838 LLVM_DEBUG(dbgs() << "Predecessor ";); 2839 } 2840 LLVM_DEBUG(dbgs() << Pred->NodeNum << " and successor " << Succ->NodeNum 2841 << " are scheduled before node " << SU->NodeNum 2842 << "\n";); 2843 } 2844 } 2845 2846 LLVM_DEBUG({ 2847 if (!Valid) 2848 dbgs() << "Invalid node order found!\n"; 2849 }); 2850 } 2851 2852 /// Attempt to fix the degenerate cases when the instruction serialization 2853 /// causes the register lifetimes to overlap. For example, 2854 /// p' = store_pi(p, b) 2855 /// = load p, offset 2856 /// In this case p and p' overlap, which means that two registers are needed. 2857 /// Instead, this function changes the load to use p' and updates the offset. 2858 void SwingSchedulerDAG::fixupRegisterOverlaps(std::deque<SUnit *> &Instrs) { 2859 unsigned OverlapReg = 0; 2860 unsigned NewBaseReg = 0; 2861 for (SUnit *SU : Instrs) { 2862 MachineInstr *MI = SU->getInstr(); 2863 for (unsigned i = 0, e = MI->getNumOperands(); i < e; ++i) { 2864 const MachineOperand &MO = MI->getOperand(i); 2865 // Look for an instruction that uses p. The instruction occurs in the 2866 // same cycle but occurs later in the serialized order. 2867 if (MO.isReg() && MO.isUse() && MO.getReg() == OverlapReg) { 2868 // Check that the instruction appears in the InstrChanges structure, 2869 // which contains instructions that can have the offset updated. 2870 DenseMap<SUnit *, std::pair<unsigned, int64_t>>::iterator It = 2871 InstrChanges.find(SU); 2872 if (It != InstrChanges.end()) { 2873 unsigned BasePos, OffsetPos; 2874 // Update the base register and adjust the offset. 2875 if (TII->getBaseAndOffsetPosition(*MI, BasePos, OffsetPos)) { 2876 MachineInstr *NewMI = MF.CloneMachineInstr(MI); 2877 NewMI->getOperand(BasePos).setReg(NewBaseReg); 2878 int64_t NewOffset = 2879 MI->getOperand(OffsetPos).getImm() - It->second.second; 2880 NewMI->getOperand(OffsetPos).setImm(NewOffset); 2881 SU->setInstr(NewMI); 2882 MISUnitMap[NewMI] = SU; 2883 NewMIs[MI] = NewMI; 2884 } 2885 } 2886 OverlapReg = 0; 2887 NewBaseReg = 0; 2888 break; 2889 } 2890 // Look for an instruction of the form p' = op(p), which uses and defines 2891 // two virtual registers that get allocated to the same physical register. 2892 unsigned TiedUseIdx = 0; 2893 if (MI->isRegTiedToUseOperand(i, &TiedUseIdx)) { 2894 // OverlapReg is p in the example above. 2895 OverlapReg = MI->getOperand(TiedUseIdx).getReg(); 2896 // NewBaseReg is p' in the example above. 2897 NewBaseReg = MI->getOperand(i).getReg(); 2898 break; 2899 } 2900 } 2901 } 2902 } 2903 2904 /// After the schedule has been formed, call this function to combine 2905 /// the instructions from the different stages/cycles. That is, this 2906 /// function creates a schedule that represents a single iteration. 2907 void SMSchedule::finalizeSchedule(SwingSchedulerDAG *SSD) { 2908 // Move all instructions to the first stage from later stages. 2909 for (int cycle = getFirstCycle(); cycle <= getFinalCycle(); ++cycle) { 2910 for (int stage = 1, lastStage = getMaxStageCount(); stage <= lastStage; 2911 ++stage) { 2912 std::deque<SUnit *> &cycleInstrs = 2913 ScheduledInstrs[cycle + (stage * InitiationInterval)]; 2914 for (SUnit *SU : llvm::reverse(cycleInstrs)) 2915 ScheduledInstrs[cycle].push_front(SU); 2916 } 2917 } 2918 2919 // Erase all the elements in the later stages. Only one iteration should 2920 // remain in the scheduled list, and it contains all the instructions. 2921 for (int cycle = getFinalCycle() + 1; cycle <= LastCycle; ++cycle) 2922 ScheduledInstrs.erase(cycle); 2923 2924 // Change the registers in instruction as specified in the InstrChanges 2925 // map. We need to use the new registers to create the correct order. 2926 for (const SUnit &SU : SSD->SUnits) 2927 SSD->applyInstrChange(SU.getInstr(), *this); 2928 2929 // Reorder the instructions in each cycle to fix and improve the 2930 // generated code. 2931 for (int Cycle = getFirstCycle(), E = getFinalCycle(); Cycle <= E; ++Cycle) { 2932 std::deque<SUnit *> &cycleInstrs = ScheduledInstrs[Cycle]; 2933 std::deque<SUnit *> newOrderPhi; 2934 for (SUnit *SU : cycleInstrs) { 2935 if (SU->getInstr()->isPHI()) 2936 newOrderPhi.push_back(SU); 2937 } 2938 std::deque<SUnit *> newOrderI; 2939 for (SUnit *SU : cycleInstrs) { 2940 if (!SU->getInstr()->isPHI()) 2941 orderDependence(SSD, SU, newOrderI); 2942 } 2943 // Replace the old order with the new order. 2944 cycleInstrs.swap(newOrderPhi); 2945 llvm::append_range(cycleInstrs, newOrderI); 2946 SSD->fixupRegisterOverlaps(cycleInstrs); 2947 } 2948 2949 LLVM_DEBUG(dump();); 2950 } 2951 2952 void NodeSet::print(raw_ostream &os) const { 2953 os << "Num nodes " << size() << " rec " << RecMII << " mov " << MaxMOV 2954 << " depth " << MaxDepth << " col " << Colocate << "\n"; 2955 for (const auto &I : Nodes) 2956 os << " SU(" << I->NodeNum << ") " << *(I->getInstr()); 2957 os << "\n"; 2958 } 2959 2960 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) 2961 /// Print the schedule information to the given output. 2962 void SMSchedule::print(raw_ostream &os) const { 2963 // Iterate over each cycle. 2964 for (int cycle = getFirstCycle(); cycle <= getFinalCycle(); ++cycle) { 2965 // Iterate over each instruction in the cycle. 2966 const_sched_iterator cycleInstrs = ScheduledInstrs.find(cycle); 2967 for (SUnit *CI : cycleInstrs->second) { 2968 os << "cycle " << cycle << " (" << stageScheduled(CI) << ") "; 2969 os << "(" << CI->NodeNum << ") "; 2970 CI->getInstr()->print(os); 2971 os << "\n"; 2972 } 2973 } 2974 } 2975 2976 /// Utility function used for debugging to print the schedule. 2977 LLVM_DUMP_METHOD void SMSchedule::dump() const { print(dbgs()); } 2978 LLVM_DUMP_METHOD void NodeSet::dump() const { print(dbgs()); } 2979 2980 void ResourceManager::dumpMRT() const { 2981 LLVM_DEBUG({ 2982 if (UseDFA) 2983 return; 2984 std::stringstream SS; 2985 SS << "MRT:\n"; 2986 SS << std::setw(4) << "Slot"; 2987 for (unsigned I = 1, E = SM.getNumProcResourceKinds(); I < E; ++I) 2988 SS << std::setw(3) << I; 2989 SS << std::setw(7) << "#Mops" 2990 << "\n"; 2991 for (int Slot = 0; Slot < InitiationInterval; ++Slot) { 2992 SS << std::setw(4) << Slot; 2993 for (unsigned I = 1, E = SM.getNumProcResourceKinds(); I < E; ++I) 2994 SS << std::setw(3) << MRT[Slot][I]; 2995 SS << std::setw(7) << NumScheduledMops[Slot] << "\n"; 2996 } 2997 dbgs() << SS.str(); 2998 }); 2999 } 3000 #endif 3001 3002 void ResourceManager::initProcResourceVectors( 3003 const MCSchedModel &SM, SmallVectorImpl<uint64_t> &Masks) { 3004 unsigned ProcResourceID = 0; 3005 3006 // We currently limit the resource kinds to 64 and below so that we can use 3007 // uint64_t for Masks 3008 assert(SM.getNumProcResourceKinds() < 64 && 3009 "Too many kinds of resources, unsupported"); 3010 // Create a unique bitmask for every processor resource unit. 3011 // Skip resource at index 0, since it always references 'InvalidUnit'. 3012 Masks.resize(SM.getNumProcResourceKinds()); 3013 for (unsigned I = 1, E = SM.getNumProcResourceKinds(); I < E; ++I) { 3014 const MCProcResourceDesc &Desc = *SM.getProcResource(I); 3015 if (Desc.SubUnitsIdxBegin) 3016 continue; 3017 Masks[I] = 1ULL << ProcResourceID; 3018 ProcResourceID++; 3019 } 3020 // Create a unique bitmask for every processor resource group. 3021 for (unsigned I = 1, E = SM.getNumProcResourceKinds(); I < E; ++I) { 3022 const MCProcResourceDesc &Desc = *SM.getProcResource(I); 3023 if (!Desc.SubUnitsIdxBegin) 3024 continue; 3025 Masks[I] = 1ULL << ProcResourceID; 3026 for (unsigned U = 0; U < Desc.NumUnits; ++U) 3027 Masks[I] |= Masks[Desc.SubUnitsIdxBegin[U]]; 3028 ProcResourceID++; 3029 } 3030 LLVM_DEBUG({ 3031 if (SwpShowResMask) { 3032 dbgs() << "ProcResourceDesc:\n"; 3033 for (unsigned I = 1, E = SM.getNumProcResourceKinds(); I < E; ++I) { 3034 const MCProcResourceDesc *ProcResource = SM.getProcResource(I); 3035 dbgs() << format(" %16s(%2d): Mask: 0x%08x, NumUnits:%2d\n", 3036 ProcResource->Name, I, Masks[I], 3037 ProcResource->NumUnits); 3038 } 3039 dbgs() << " -----------------\n"; 3040 } 3041 }); 3042 } 3043 3044 bool ResourceManager::canReserveResources(SUnit &SU, int Cycle) { 3045 LLVM_DEBUG({ 3046 if (SwpDebugResource) 3047 dbgs() << "canReserveResources:\n"; 3048 }); 3049 if (UseDFA) 3050 return DFAResources[positiveModulo(Cycle, InitiationInterval)] 3051 ->canReserveResources(&SU.getInstr()->getDesc()); 3052 3053 const MCSchedClassDesc *SCDesc = DAG->getSchedClass(&SU); 3054 if (!SCDesc->isValid()) { 3055 LLVM_DEBUG({ 3056 dbgs() << "No valid Schedule Class Desc for schedClass!\n"; 3057 dbgs() << "isPseudo:" << SU.getInstr()->isPseudo() << "\n"; 3058 }); 3059 return true; 3060 } 3061 3062 reserveResources(SCDesc, Cycle); 3063 bool Result = !isOverbooked(); 3064 unreserveResources(SCDesc, Cycle); 3065 3066 LLVM_DEBUG(if (SwpDebugResource) dbgs() << "return " << Result << "\n\n";); 3067 return Result; 3068 } 3069 3070 void ResourceManager::reserveResources(SUnit &SU, int Cycle) { 3071 LLVM_DEBUG({ 3072 if (SwpDebugResource) 3073 dbgs() << "reserveResources:\n"; 3074 }); 3075 if (UseDFA) 3076 return DFAResources[positiveModulo(Cycle, InitiationInterval)] 3077 ->reserveResources(&SU.getInstr()->getDesc()); 3078 3079 const MCSchedClassDesc *SCDesc = DAG->getSchedClass(&SU); 3080 if (!SCDesc->isValid()) { 3081 LLVM_DEBUG({ 3082 dbgs() << "No valid Schedule Class Desc for schedClass!\n"; 3083 dbgs() << "isPseudo:" << SU.getInstr()->isPseudo() << "\n"; 3084 }); 3085 return; 3086 } 3087 3088 reserveResources(SCDesc, Cycle); 3089 3090 LLVM_DEBUG({ 3091 if (SwpDebugResource) { 3092 dumpMRT(); 3093 dbgs() << "reserveResources: done!\n\n"; 3094 } 3095 }); 3096 } 3097 3098 void ResourceManager::reserveResources(const MCSchedClassDesc *SCDesc, 3099 int Cycle) { 3100 assert(!UseDFA); 3101 for (const MCWriteProcResEntry &PRE : make_range( 3102 STI->getWriteProcResBegin(SCDesc), STI->getWriteProcResEnd(SCDesc))) 3103 for (int C = Cycle; C < Cycle + PRE.Cycles; ++C) 3104 ++MRT[positiveModulo(C, InitiationInterval)][PRE.ProcResourceIdx]; 3105 3106 for (int C = Cycle; C < Cycle + SCDesc->NumMicroOps; ++C) 3107 ++NumScheduledMops[positiveModulo(C, InitiationInterval)]; 3108 } 3109 3110 void ResourceManager::unreserveResources(const MCSchedClassDesc *SCDesc, 3111 int Cycle) { 3112 assert(!UseDFA); 3113 for (const MCWriteProcResEntry &PRE : make_range( 3114 STI->getWriteProcResBegin(SCDesc), STI->getWriteProcResEnd(SCDesc))) 3115 for (int C = Cycle; C < Cycle + PRE.Cycles; ++C) 3116 --MRT[positiveModulo(C, InitiationInterval)][PRE.ProcResourceIdx]; 3117 3118 for (int C = Cycle; C < Cycle + SCDesc->NumMicroOps; ++C) 3119 --NumScheduledMops[positiveModulo(C, InitiationInterval)]; 3120 } 3121 3122 bool ResourceManager::isOverbooked() const { 3123 assert(!UseDFA); 3124 for (int Slot = 0; Slot < InitiationInterval; ++Slot) { 3125 for (unsigned I = 1, E = SM.getNumProcResourceKinds(); I < E; ++I) { 3126 const MCProcResourceDesc *Desc = SM.getProcResource(I); 3127 if (MRT[Slot][I] > Desc->NumUnits) 3128 return true; 3129 } 3130 if (NumScheduledMops[Slot] > IssueWidth) 3131 return true; 3132 } 3133 return false; 3134 } 3135 3136 int ResourceManager::calculateResMIIDFA() const { 3137 assert(UseDFA); 3138 3139 // Sort the instructions by the number of available choices for scheduling, 3140 // least to most. Use the number of critical resources as the tie breaker. 3141 FuncUnitSorter FUS = FuncUnitSorter(*ST); 3142 for (SUnit &SU : DAG->SUnits) 3143 FUS.calcCriticalResources(*SU.getInstr()); 3144 PriorityQueue<MachineInstr *, std::vector<MachineInstr *>, FuncUnitSorter> 3145 FuncUnitOrder(FUS); 3146 3147 for (SUnit &SU : DAG->SUnits) 3148 FuncUnitOrder.push(SU.getInstr()); 3149 3150 SmallVector<std::unique_ptr<DFAPacketizer>, 8> Resources; 3151 Resources.push_back( 3152 std::unique_ptr<DFAPacketizer>(TII->CreateTargetScheduleState(*ST))); 3153 3154 while (!FuncUnitOrder.empty()) { 3155 MachineInstr *MI = FuncUnitOrder.top(); 3156 FuncUnitOrder.pop(); 3157 if (TII->isZeroCost(MI->getOpcode())) 3158 continue; 3159 3160 // Attempt to reserve the instruction in an existing DFA. At least one 3161 // DFA is needed for each cycle. 3162 unsigned NumCycles = DAG->getSUnit(MI)->Latency; 3163 unsigned ReservedCycles = 0; 3164 auto *RI = Resources.begin(); 3165 auto *RE = Resources.end(); 3166 LLVM_DEBUG({ 3167 dbgs() << "Trying to reserve resource for " << NumCycles 3168 << " cycles for \n"; 3169 MI->dump(); 3170 }); 3171 for (unsigned C = 0; C < NumCycles; ++C) 3172 while (RI != RE) { 3173 if ((*RI)->canReserveResources(*MI)) { 3174 (*RI)->reserveResources(*MI); 3175 ++ReservedCycles; 3176 break; 3177 } 3178 RI++; 3179 } 3180 LLVM_DEBUG(dbgs() << "ReservedCycles:" << ReservedCycles 3181 << ", NumCycles:" << NumCycles << "\n"); 3182 // Add new DFAs, if needed, to reserve resources. 3183 for (unsigned C = ReservedCycles; C < NumCycles; ++C) { 3184 LLVM_DEBUG(if (SwpDebugResource) dbgs() 3185 << "NewResource created to reserve resources" 3186 << "\n"); 3187 auto *NewResource = TII->CreateTargetScheduleState(*ST); 3188 assert(NewResource->canReserveResources(*MI) && "Reserve error."); 3189 NewResource->reserveResources(*MI); 3190 Resources.push_back(std::unique_ptr<DFAPacketizer>(NewResource)); 3191 } 3192 } 3193 3194 int Resmii = Resources.size(); 3195 LLVM_DEBUG(dbgs() << "Return Res MII:" << Resmii << "\n"); 3196 return Resmii; 3197 } 3198 3199 int ResourceManager::calculateResMII() const { 3200 if (UseDFA) 3201 return calculateResMIIDFA(); 3202 3203 // Count each resource consumption and divide it by the number of units. 3204 // ResMII is the max value among them. 3205 3206 int NumMops = 0; 3207 SmallVector<uint64_t> ResourceCount(SM.getNumProcResourceKinds()); 3208 for (SUnit &SU : DAG->SUnits) { 3209 if (TII->isZeroCost(SU.getInstr()->getOpcode())) 3210 continue; 3211 3212 const MCSchedClassDesc *SCDesc = DAG->getSchedClass(&SU); 3213 if (!SCDesc->isValid()) 3214 continue; 3215 3216 LLVM_DEBUG({ 3217 if (SwpDebugResource) { 3218 DAG->dumpNode(SU); 3219 dbgs() << " #Mops: " << SCDesc->NumMicroOps << "\n" 3220 << " WriteProcRes: "; 3221 } 3222 }); 3223 NumMops += SCDesc->NumMicroOps; 3224 for (const MCWriteProcResEntry &PRE : 3225 make_range(STI->getWriteProcResBegin(SCDesc), 3226 STI->getWriteProcResEnd(SCDesc))) { 3227 LLVM_DEBUG({ 3228 if (SwpDebugResource) { 3229 const MCProcResourceDesc *Desc = 3230 SM.getProcResource(PRE.ProcResourceIdx); 3231 dbgs() << Desc->Name << ": " << PRE.Cycles << ", "; 3232 } 3233 }); 3234 ResourceCount[PRE.ProcResourceIdx] += PRE.Cycles; 3235 } 3236 LLVM_DEBUG(if (SwpDebugResource) dbgs() << "\n"); 3237 } 3238 3239 int Result = (NumMops + IssueWidth - 1) / IssueWidth; 3240 LLVM_DEBUG({ 3241 if (SwpDebugResource) 3242 dbgs() << "#Mops: " << NumMops << ", " 3243 << "IssueWidth: " << IssueWidth << ", " 3244 << "Cycles: " << Result << "\n"; 3245 }); 3246 3247 LLVM_DEBUG({ 3248 if (SwpDebugResource) { 3249 std::stringstream SS; 3250 SS << std::setw(2) << "ID" << std::setw(16) << "Name" << std::setw(10) 3251 << "Units" << std::setw(10) << "Consumed" << std::setw(10) << "Cycles" 3252 << "\n"; 3253 dbgs() << SS.str(); 3254 } 3255 }); 3256 for (unsigned I = 1, E = SM.getNumProcResourceKinds(); I < E; ++I) { 3257 const MCProcResourceDesc *Desc = SM.getProcResource(I); 3258 int Cycles = (ResourceCount[I] + Desc->NumUnits - 1) / Desc->NumUnits; 3259 LLVM_DEBUG({ 3260 if (SwpDebugResource) { 3261 std::stringstream SS; 3262 SS << std::setw(2) << I << std::setw(16) << Desc->Name << std::setw(10) 3263 << Desc->NumUnits << std::setw(10) << ResourceCount[I] 3264 << std::setw(10) << Cycles << "\n"; 3265 dbgs() << SS.str(); 3266 } 3267 }); 3268 if (Cycles > Result) 3269 Result = Cycles; 3270 } 3271 return Result; 3272 } 3273 3274 void ResourceManager::init(int II) { 3275 InitiationInterval = II; 3276 DFAResources.clear(); 3277 DFAResources.resize(II); 3278 for (auto &I : DFAResources) 3279 I.reset(ST->getInstrInfo()->CreateTargetScheduleState(*ST)); 3280 MRT.clear(); 3281 MRT.resize(II, SmallVector<uint64_t>(SM.getNumProcResourceKinds())); 3282 NumScheduledMops.clear(); 3283 NumScheduledMops.resize(II); 3284 } 3285