1 //===- MachineScheduler.cpp - Machine Instruction Scheduler ---------------===// 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 // MachineScheduler schedules machine instructions after phi elimination. It 10 // preserves LiveIntervals so it can be invoked before register allocation. 11 // 12 //===----------------------------------------------------------------------===// 13 14 #include "llvm/CodeGen/MachineScheduler.h" 15 #include "llvm/ADT/ArrayRef.h" 16 #include "llvm/ADT/BitVector.h" 17 #include "llvm/ADT/DenseMap.h" 18 #include "llvm/ADT/PriorityQueue.h" 19 #include "llvm/ADT/STLExtras.h" 20 #include "llvm/ADT/SmallVector.h" 21 #include "llvm/ADT/Statistic.h" 22 #include "llvm/ADT/iterator_range.h" 23 #include "llvm/Analysis/AliasAnalysis.h" 24 #include "llvm/CodeGen/LiveInterval.h" 25 #include "llvm/CodeGen/LiveIntervals.h" 26 #include "llvm/CodeGen/MachineBasicBlock.h" 27 #include "llvm/CodeGen/MachineDominators.h" 28 #include "llvm/CodeGen/MachineFunction.h" 29 #include "llvm/CodeGen/MachineFunctionPass.h" 30 #include "llvm/CodeGen/MachineInstr.h" 31 #include "llvm/CodeGen/MachineLoopInfo.h" 32 #include "llvm/CodeGen/MachineOperand.h" 33 #include "llvm/CodeGen/MachinePassRegistry.h" 34 #include "llvm/CodeGen/MachineRegisterInfo.h" 35 #include "llvm/CodeGen/RegisterClassInfo.h" 36 #include "llvm/CodeGen/RegisterPressure.h" 37 #include "llvm/CodeGen/ScheduleDAG.h" 38 #include "llvm/CodeGen/ScheduleDAGInstrs.h" 39 #include "llvm/CodeGen/ScheduleDAGMutation.h" 40 #include "llvm/CodeGen/ScheduleDFS.h" 41 #include "llvm/CodeGen/ScheduleHazardRecognizer.h" 42 #include "llvm/CodeGen/SlotIndexes.h" 43 #include "llvm/CodeGen/TargetFrameLowering.h" 44 #include "llvm/CodeGen/TargetInstrInfo.h" 45 #include "llvm/CodeGen/TargetLowering.h" 46 #include "llvm/CodeGen/TargetPassConfig.h" 47 #include "llvm/CodeGen/TargetRegisterInfo.h" 48 #include "llvm/CodeGen/TargetSchedule.h" 49 #include "llvm/CodeGen/TargetSubtargetInfo.h" 50 #include "llvm/Config/llvm-config.h" 51 #include "llvm/InitializePasses.h" 52 #include "llvm/MC/LaneBitmask.h" 53 #include "llvm/Pass.h" 54 #include "llvm/Support/CommandLine.h" 55 #include "llvm/Support/Compiler.h" 56 #include "llvm/Support/Debug.h" 57 #include "llvm/Support/ErrorHandling.h" 58 #include "llvm/Support/GraphWriter.h" 59 #include "llvm/Support/MachineValueType.h" 60 #include "llvm/Support/raw_ostream.h" 61 #include <algorithm> 62 #include <cassert> 63 #include <cstdint> 64 #include <iterator> 65 #include <limits> 66 #include <memory> 67 #include <string> 68 #include <tuple> 69 #include <utility> 70 #include <vector> 71 72 using namespace llvm; 73 74 #define DEBUG_TYPE "machine-scheduler" 75 76 STATISTIC(NumClustered, "Number of load/store pairs clustered"); 77 78 namespace llvm { 79 80 cl::opt<bool> ForceTopDown("misched-topdown", cl::Hidden, 81 cl::desc("Force top-down list scheduling")); 82 cl::opt<bool> ForceBottomUp("misched-bottomup", cl::Hidden, 83 cl::desc("Force bottom-up list scheduling")); 84 cl::opt<bool> 85 DumpCriticalPathLength("misched-dcpl", cl::Hidden, 86 cl::desc("Print critical path length to stdout")); 87 88 cl::opt<bool> VerifyScheduling( 89 "verify-misched", cl::Hidden, 90 cl::desc("Verify machine instrs before and after machine scheduling")); 91 92 #ifndef NDEBUG 93 cl::opt<bool> ViewMISchedDAGs( 94 "view-misched-dags", cl::Hidden, 95 cl::desc("Pop up a window to show MISched dags after they are processed")); 96 cl::opt<bool> PrintDAGs("misched-print-dags", cl::Hidden, 97 cl::desc("Print schedule DAGs")); 98 #else 99 const bool ViewMISchedDAGs = false; 100 const bool PrintDAGs = false; 101 #endif // NDEBUG 102 103 } // end namespace llvm 104 105 #ifndef NDEBUG 106 /// In some situations a few uninteresting nodes depend on nearly all other 107 /// nodes in the graph, provide a cutoff to hide them. 108 static cl::opt<unsigned> ViewMISchedCutoff("view-misched-cutoff", cl::Hidden, 109 cl::desc("Hide nodes with more predecessor/successor than cutoff")); 110 111 static cl::opt<unsigned> MISchedCutoff("misched-cutoff", cl::Hidden, 112 cl::desc("Stop scheduling after N instructions"), cl::init(~0U)); 113 114 static cl::opt<std::string> SchedOnlyFunc("misched-only-func", cl::Hidden, 115 cl::desc("Only schedule this function")); 116 static cl::opt<unsigned> SchedOnlyBlock("misched-only-block", cl::Hidden, 117 cl::desc("Only schedule this MBB#")); 118 #endif // NDEBUG 119 120 /// Avoid quadratic complexity in unusually large basic blocks by limiting the 121 /// size of the ready lists. 122 static cl::opt<unsigned> ReadyListLimit("misched-limit", cl::Hidden, 123 cl::desc("Limit ready list to N instructions"), cl::init(256)); 124 125 static cl::opt<bool> EnableRegPressure("misched-regpressure", cl::Hidden, 126 cl::desc("Enable register pressure scheduling."), cl::init(true)); 127 128 static cl::opt<bool> EnableCyclicPath("misched-cyclicpath", cl::Hidden, 129 cl::desc("Enable cyclic critical path analysis."), cl::init(true)); 130 131 static cl::opt<bool> EnableMemOpCluster("misched-cluster", cl::Hidden, 132 cl::desc("Enable memop clustering."), 133 cl::init(true)); 134 static cl::opt<bool> 135 ForceFastCluster("force-fast-cluster", cl::Hidden, 136 cl::desc("Switch to fast cluster algorithm with the lost " 137 "of some fusion opportunities"), 138 cl::init(false)); 139 static cl::opt<unsigned> 140 FastClusterThreshold("fast-cluster-threshold", cl::Hidden, 141 cl::desc("The threshold for fast cluster"), 142 cl::init(1000)); 143 144 // DAG subtrees must have at least this many nodes. 145 static const unsigned MinSubtreeSize = 8; 146 147 // Pin the vtables to this file. 148 void MachineSchedStrategy::anchor() {} 149 150 void ScheduleDAGMutation::anchor() {} 151 152 //===----------------------------------------------------------------------===// 153 // Machine Instruction Scheduling Pass and Registry 154 //===----------------------------------------------------------------------===// 155 156 MachineSchedContext::MachineSchedContext() { 157 RegClassInfo = new RegisterClassInfo(); 158 } 159 160 MachineSchedContext::~MachineSchedContext() { 161 delete RegClassInfo; 162 } 163 164 namespace { 165 166 /// Base class for a machine scheduler class that can run at any point. 167 class MachineSchedulerBase : public MachineSchedContext, 168 public MachineFunctionPass { 169 public: 170 MachineSchedulerBase(char &ID): MachineFunctionPass(ID) {} 171 172 void print(raw_ostream &O, const Module* = nullptr) const override; 173 174 protected: 175 void scheduleRegions(ScheduleDAGInstrs &Scheduler, bool FixKillFlags); 176 }; 177 178 /// MachineScheduler runs after coalescing and before register allocation. 179 class MachineScheduler : public MachineSchedulerBase { 180 public: 181 MachineScheduler(); 182 183 void getAnalysisUsage(AnalysisUsage &AU) const override; 184 185 bool runOnMachineFunction(MachineFunction&) override; 186 187 static char ID; // Class identification, replacement for typeinfo 188 189 protected: 190 ScheduleDAGInstrs *createMachineScheduler(); 191 }; 192 193 /// PostMachineScheduler runs after shortly before code emission. 194 class PostMachineScheduler : public MachineSchedulerBase { 195 public: 196 PostMachineScheduler(); 197 198 void getAnalysisUsage(AnalysisUsage &AU) const override; 199 200 bool runOnMachineFunction(MachineFunction&) override; 201 202 static char ID; // Class identification, replacement for typeinfo 203 204 protected: 205 ScheduleDAGInstrs *createPostMachineScheduler(); 206 }; 207 208 } // end anonymous namespace 209 210 char MachineScheduler::ID = 0; 211 212 char &llvm::MachineSchedulerID = MachineScheduler::ID; 213 214 INITIALIZE_PASS_BEGIN(MachineScheduler, DEBUG_TYPE, 215 "Machine Instruction Scheduler", false, false) 216 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass) 217 INITIALIZE_PASS_DEPENDENCY(MachineDominatorTree) 218 INITIALIZE_PASS_DEPENDENCY(MachineLoopInfo) 219 INITIALIZE_PASS_DEPENDENCY(SlotIndexes) 220 INITIALIZE_PASS_DEPENDENCY(LiveIntervals) 221 INITIALIZE_PASS_END(MachineScheduler, DEBUG_TYPE, 222 "Machine Instruction Scheduler", false, false) 223 224 MachineScheduler::MachineScheduler() : MachineSchedulerBase(ID) { 225 initializeMachineSchedulerPass(*PassRegistry::getPassRegistry()); 226 } 227 228 void MachineScheduler::getAnalysisUsage(AnalysisUsage &AU) const { 229 AU.setPreservesCFG(); 230 AU.addRequired<MachineDominatorTree>(); 231 AU.addRequired<MachineLoopInfo>(); 232 AU.addRequired<AAResultsWrapperPass>(); 233 AU.addRequired<TargetPassConfig>(); 234 AU.addRequired<SlotIndexes>(); 235 AU.addPreserved<SlotIndexes>(); 236 AU.addRequired<LiveIntervals>(); 237 AU.addPreserved<LiveIntervals>(); 238 MachineFunctionPass::getAnalysisUsage(AU); 239 } 240 241 char PostMachineScheduler::ID = 0; 242 243 char &llvm::PostMachineSchedulerID = PostMachineScheduler::ID; 244 245 INITIALIZE_PASS_BEGIN(PostMachineScheduler, "postmisched", 246 "PostRA Machine Instruction Scheduler", false, false) 247 INITIALIZE_PASS_DEPENDENCY(MachineDominatorTree) 248 INITIALIZE_PASS_DEPENDENCY(MachineLoopInfo) 249 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass) 250 INITIALIZE_PASS_END(PostMachineScheduler, "postmisched", 251 "PostRA Machine Instruction Scheduler", false, false) 252 253 PostMachineScheduler::PostMachineScheduler() : MachineSchedulerBase(ID) { 254 initializePostMachineSchedulerPass(*PassRegistry::getPassRegistry()); 255 } 256 257 void PostMachineScheduler::getAnalysisUsage(AnalysisUsage &AU) const { 258 AU.setPreservesCFG(); 259 AU.addRequired<MachineDominatorTree>(); 260 AU.addRequired<MachineLoopInfo>(); 261 AU.addRequired<AAResultsWrapperPass>(); 262 AU.addRequired<TargetPassConfig>(); 263 MachineFunctionPass::getAnalysisUsage(AU); 264 } 265 266 MachinePassRegistry<MachineSchedRegistry::ScheduleDAGCtor> 267 MachineSchedRegistry::Registry; 268 269 /// A dummy default scheduler factory indicates whether the scheduler 270 /// is overridden on the command line. 271 static ScheduleDAGInstrs *useDefaultMachineSched(MachineSchedContext *C) { 272 return nullptr; 273 } 274 275 /// MachineSchedOpt allows command line selection of the scheduler. 276 static cl::opt<MachineSchedRegistry::ScheduleDAGCtor, false, 277 RegisterPassParser<MachineSchedRegistry>> 278 MachineSchedOpt("misched", 279 cl::init(&useDefaultMachineSched), cl::Hidden, 280 cl::desc("Machine instruction scheduler to use")); 281 282 static MachineSchedRegistry 283 DefaultSchedRegistry("default", "Use the target's default scheduler choice.", 284 useDefaultMachineSched); 285 286 static cl::opt<bool> EnableMachineSched( 287 "enable-misched", 288 cl::desc("Enable the machine instruction scheduling pass."), cl::init(true), 289 cl::Hidden); 290 291 static cl::opt<bool> EnablePostRAMachineSched( 292 "enable-post-misched", 293 cl::desc("Enable the post-ra machine instruction scheduling pass."), 294 cl::init(true), cl::Hidden); 295 296 /// Decrement this iterator until reaching the top or a non-debug instr. 297 static MachineBasicBlock::const_iterator 298 priorNonDebug(MachineBasicBlock::const_iterator I, 299 MachineBasicBlock::const_iterator Beg) { 300 assert(I != Beg && "reached the top of the region, cannot decrement"); 301 while (--I != Beg) { 302 if (!I->isDebugOrPseudoInstr()) 303 break; 304 } 305 return I; 306 } 307 308 /// Non-const version. 309 static MachineBasicBlock::iterator 310 priorNonDebug(MachineBasicBlock::iterator I, 311 MachineBasicBlock::const_iterator Beg) { 312 return priorNonDebug(MachineBasicBlock::const_iterator(I), Beg) 313 .getNonConstIterator(); 314 } 315 316 /// If this iterator is a debug value, increment until reaching the End or a 317 /// non-debug instruction. 318 static MachineBasicBlock::const_iterator 319 nextIfDebug(MachineBasicBlock::const_iterator I, 320 MachineBasicBlock::const_iterator End) { 321 for(; I != End; ++I) { 322 if (!I->isDebugOrPseudoInstr()) 323 break; 324 } 325 return I; 326 } 327 328 /// Non-const version. 329 static MachineBasicBlock::iterator 330 nextIfDebug(MachineBasicBlock::iterator I, 331 MachineBasicBlock::const_iterator End) { 332 return nextIfDebug(MachineBasicBlock::const_iterator(I), End) 333 .getNonConstIterator(); 334 } 335 336 /// Instantiate a ScheduleDAGInstrs that will be owned by the caller. 337 ScheduleDAGInstrs *MachineScheduler::createMachineScheduler() { 338 // Select the scheduler, or set the default. 339 MachineSchedRegistry::ScheduleDAGCtor Ctor = MachineSchedOpt; 340 if (Ctor != useDefaultMachineSched) 341 return Ctor(this); 342 343 // Get the default scheduler set by the target for this function. 344 ScheduleDAGInstrs *Scheduler = PassConfig->createMachineScheduler(this); 345 if (Scheduler) 346 return Scheduler; 347 348 // Default to GenericScheduler. 349 return createGenericSchedLive(this); 350 } 351 352 /// Instantiate a ScheduleDAGInstrs for PostRA scheduling that will be owned by 353 /// the caller. We don't have a command line option to override the postRA 354 /// scheduler. The Target must configure it. 355 ScheduleDAGInstrs *PostMachineScheduler::createPostMachineScheduler() { 356 // Get the postRA scheduler set by the target for this function. 357 ScheduleDAGInstrs *Scheduler = PassConfig->createPostMachineScheduler(this); 358 if (Scheduler) 359 return Scheduler; 360 361 // Default to GenericScheduler. 362 return createGenericSchedPostRA(this); 363 } 364 365 /// Top-level MachineScheduler pass driver. 366 /// 367 /// Visit blocks in function order. Divide each block into scheduling regions 368 /// and visit them bottom-up. Visiting regions bottom-up is not required, but is 369 /// consistent with the DAG builder, which traverses the interior of the 370 /// scheduling regions bottom-up. 371 /// 372 /// This design avoids exposing scheduling boundaries to the DAG builder, 373 /// simplifying the DAG builder's support for "special" target instructions. 374 /// At the same time the design allows target schedulers to operate across 375 /// scheduling boundaries, for example to bundle the boundary instructions 376 /// without reordering them. This creates complexity, because the target 377 /// scheduler must update the RegionBegin and RegionEnd positions cached by 378 /// ScheduleDAGInstrs whenever adding or removing instructions. A much simpler 379 /// design would be to split blocks at scheduling boundaries, but LLVM has a 380 /// general bias against block splitting purely for implementation simplicity. 381 bool MachineScheduler::runOnMachineFunction(MachineFunction &mf) { 382 if (skipFunction(mf.getFunction())) 383 return false; 384 385 if (EnableMachineSched.getNumOccurrences()) { 386 if (!EnableMachineSched) 387 return false; 388 } else if (!mf.getSubtarget().enableMachineScheduler()) 389 return false; 390 391 LLVM_DEBUG(dbgs() << "Before MISched:\n"; mf.print(dbgs())); 392 393 // Initialize the context of the pass. 394 MF = &mf; 395 MLI = &getAnalysis<MachineLoopInfo>(); 396 MDT = &getAnalysis<MachineDominatorTree>(); 397 PassConfig = &getAnalysis<TargetPassConfig>(); 398 AA = &getAnalysis<AAResultsWrapperPass>().getAAResults(); 399 400 LIS = &getAnalysis<LiveIntervals>(); 401 402 if (VerifyScheduling) { 403 LLVM_DEBUG(LIS->dump()); 404 MF->verify(this, "Before machine scheduling."); 405 } 406 RegClassInfo->runOnMachineFunction(*MF); 407 408 // Instantiate the selected scheduler for this target, function, and 409 // optimization level. 410 std::unique_ptr<ScheduleDAGInstrs> Scheduler(createMachineScheduler()); 411 scheduleRegions(*Scheduler, false); 412 413 LLVM_DEBUG(LIS->dump()); 414 if (VerifyScheduling) 415 MF->verify(this, "After machine scheduling."); 416 return true; 417 } 418 419 bool PostMachineScheduler::runOnMachineFunction(MachineFunction &mf) { 420 if (skipFunction(mf.getFunction())) 421 return false; 422 423 if (EnablePostRAMachineSched.getNumOccurrences()) { 424 if (!EnablePostRAMachineSched) 425 return false; 426 } else if (!mf.getSubtarget().enablePostRAMachineScheduler()) { 427 LLVM_DEBUG(dbgs() << "Subtarget disables post-MI-sched.\n"); 428 return false; 429 } 430 LLVM_DEBUG(dbgs() << "Before post-MI-sched:\n"; mf.print(dbgs())); 431 432 // Initialize the context of the pass. 433 MF = &mf; 434 MLI = &getAnalysis<MachineLoopInfo>(); 435 PassConfig = &getAnalysis<TargetPassConfig>(); 436 AA = &getAnalysis<AAResultsWrapperPass>().getAAResults(); 437 438 if (VerifyScheduling) 439 MF->verify(this, "Before post machine scheduling."); 440 441 // Instantiate the selected scheduler for this target, function, and 442 // optimization level. 443 std::unique_ptr<ScheduleDAGInstrs> Scheduler(createPostMachineScheduler()); 444 scheduleRegions(*Scheduler, true); 445 446 if (VerifyScheduling) 447 MF->verify(this, "After post machine scheduling."); 448 return true; 449 } 450 451 /// Return true of the given instruction should not be included in a scheduling 452 /// region. 453 /// 454 /// MachineScheduler does not currently support scheduling across calls. To 455 /// handle calls, the DAG builder needs to be modified to create register 456 /// anti/output dependencies on the registers clobbered by the call's regmask 457 /// operand. In PreRA scheduling, the stack pointer adjustment already prevents 458 /// scheduling across calls. In PostRA scheduling, we need the isCall to enforce 459 /// the boundary, but there would be no benefit to postRA scheduling across 460 /// calls this late anyway. 461 static bool isSchedBoundary(MachineBasicBlock::iterator MI, 462 MachineBasicBlock *MBB, 463 MachineFunction *MF, 464 const TargetInstrInfo *TII) { 465 return MI->isCall() || TII->isSchedulingBoundary(*MI, MBB, *MF); 466 } 467 468 /// A region of an MBB for scheduling. 469 namespace { 470 struct SchedRegion { 471 /// RegionBegin is the first instruction in the scheduling region, and 472 /// RegionEnd is either MBB->end() or the scheduling boundary after the 473 /// last instruction in the scheduling region. These iterators cannot refer 474 /// to instructions outside of the identified scheduling region because 475 /// those may be reordered before scheduling this region. 476 MachineBasicBlock::iterator RegionBegin; 477 MachineBasicBlock::iterator RegionEnd; 478 unsigned NumRegionInstrs; 479 480 SchedRegion(MachineBasicBlock::iterator B, MachineBasicBlock::iterator E, 481 unsigned N) : 482 RegionBegin(B), RegionEnd(E), NumRegionInstrs(N) {} 483 }; 484 } // end anonymous namespace 485 486 using MBBRegionsVector = SmallVector<SchedRegion, 16>; 487 488 static void 489 getSchedRegions(MachineBasicBlock *MBB, 490 MBBRegionsVector &Regions, 491 bool RegionsTopDown) { 492 MachineFunction *MF = MBB->getParent(); 493 const TargetInstrInfo *TII = MF->getSubtarget().getInstrInfo(); 494 495 MachineBasicBlock::iterator I = nullptr; 496 for(MachineBasicBlock::iterator RegionEnd = MBB->end(); 497 RegionEnd != MBB->begin(); RegionEnd = I) { 498 499 // Avoid decrementing RegionEnd for blocks with no terminator. 500 if (RegionEnd != MBB->end() || 501 isSchedBoundary(&*std::prev(RegionEnd), &*MBB, MF, TII)) { 502 --RegionEnd; 503 } 504 505 // The next region starts above the previous region. Look backward in the 506 // instruction stream until we find the nearest boundary. 507 unsigned NumRegionInstrs = 0; 508 I = RegionEnd; 509 for (;I != MBB->begin(); --I) { 510 MachineInstr &MI = *std::prev(I); 511 if (isSchedBoundary(&MI, &*MBB, MF, TII)) 512 break; 513 if (!MI.isDebugOrPseudoInstr()) { 514 // MBB::size() uses instr_iterator to count. Here we need a bundle to 515 // count as a single instruction. 516 ++NumRegionInstrs; 517 } 518 } 519 520 // It's possible we found a scheduling region that only has debug 521 // instructions. Don't bother scheduling these. 522 if (NumRegionInstrs != 0) 523 Regions.push_back(SchedRegion(I, RegionEnd, NumRegionInstrs)); 524 } 525 526 if (RegionsTopDown) 527 std::reverse(Regions.begin(), Regions.end()); 528 } 529 530 /// Main driver for both MachineScheduler and PostMachineScheduler. 531 void MachineSchedulerBase::scheduleRegions(ScheduleDAGInstrs &Scheduler, 532 bool FixKillFlags) { 533 // Visit all machine basic blocks. 534 // 535 // TODO: Visit blocks in global postorder or postorder within the bottom-up 536 // loop tree. Then we can optionally compute global RegPressure. 537 for (MachineFunction::iterator MBB = MF->begin(), MBBEnd = MF->end(); 538 MBB != MBBEnd; ++MBB) { 539 540 Scheduler.startBlock(&*MBB); 541 542 #ifndef NDEBUG 543 if (SchedOnlyFunc.getNumOccurrences() && SchedOnlyFunc != MF->getName()) 544 continue; 545 if (SchedOnlyBlock.getNumOccurrences() 546 && (int)SchedOnlyBlock != MBB->getNumber()) 547 continue; 548 #endif 549 550 // Break the block into scheduling regions [I, RegionEnd). RegionEnd 551 // points to the scheduling boundary at the bottom of the region. The DAG 552 // does not include RegionEnd, but the region does (i.e. the next 553 // RegionEnd is above the previous RegionBegin). If the current block has 554 // no terminator then RegionEnd == MBB->end() for the bottom region. 555 // 556 // All the regions of MBB are first found and stored in MBBRegions, which 557 // will be processed (MBB) top-down if initialized with true. 558 // 559 // The Scheduler may insert instructions during either schedule() or 560 // exitRegion(), even for empty regions. So the local iterators 'I' and 561 // 'RegionEnd' are invalid across these calls. Instructions must not be 562 // added to other regions than the current one without updating MBBRegions. 563 564 MBBRegionsVector MBBRegions; 565 getSchedRegions(&*MBB, MBBRegions, Scheduler.doMBBSchedRegionsTopDown()); 566 for (const SchedRegion &R : MBBRegions) { 567 MachineBasicBlock::iterator I = R.RegionBegin; 568 MachineBasicBlock::iterator RegionEnd = R.RegionEnd; 569 unsigned NumRegionInstrs = R.NumRegionInstrs; 570 571 // Notify the scheduler of the region, even if we may skip scheduling 572 // it. Perhaps it still needs to be bundled. 573 Scheduler.enterRegion(&*MBB, I, RegionEnd, NumRegionInstrs); 574 575 // Skip empty scheduling regions (0 or 1 schedulable instructions). 576 if (I == RegionEnd || I == std::prev(RegionEnd)) { 577 // Close the current region. Bundle the terminator if needed. 578 // This invalidates 'RegionEnd' and 'I'. 579 Scheduler.exitRegion(); 580 continue; 581 } 582 LLVM_DEBUG(dbgs() << "********** MI Scheduling **********\n"); 583 LLVM_DEBUG(dbgs() << MF->getName() << ":" << printMBBReference(*MBB) 584 << " " << MBB->getName() << "\n From: " << *I 585 << " To: "; 586 if (RegionEnd != MBB->end()) dbgs() << *RegionEnd; 587 else dbgs() << "End\n"; 588 dbgs() << " RegionInstrs: " << NumRegionInstrs << '\n'); 589 if (DumpCriticalPathLength) { 590 errs() << MF->getName(); 591 errs() << ":%bb. " << MBB->getNumber(); 592 errs() << " " << MBB->getName() << " \n"; 593 } 594 595 // Schedule a region: possibly reorder instructions. 596 // This invalidates the original region iterators. 597 Scheduler.schedule(); 598 599 // Close the current region. 600 Scheduler.exitRegion(); 601 } 602 Scheduler.finishBlock(); 603 // FIXME: Ideally, no further passes should rely on kill flags. However, 604 // thumb2 size reduction is currently an exception, so the PostMIScheduler 605 // needs to do this. 606 if (FixKillFlags) 607 Scheduler.fixupKills(*MBB); 608 } 609 Scheduler.finalizeSchedule(); 610 } 611 612 void MachineSchedulerBase::print(raw_ostream &O, const Module* m) const { 613 // unimplemented 614 } 615 616 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) 617 LLVM_DUMP_METHOD void ReadyQueue::dump() const { 618 dbgs() << "Queue " << Name << ": "; 619 for (const SUnit *SU : Queue) 620 dbgs() << SU->NodeNum << " "; 621 dbgs() << "\n"; 622 } 623 #endif 624 625 //===----------------------------------------------------------------------===// 626 // ScheduleDAGMI - Basic machine instruction scheduling. This is 627 // independent of PreRA/PostRA scheduling and involves no extra book-keeping for 628 // virtual registers. 629 // ===----------------------------------------------------------------------===/ 630 631 // Provide a vtable anchor. 632 ScheduleDAGMI::~ScheduleDAGMI() = default; 633 634 /// ReleaseSucc - Decrement the NumPredsLeft count of a successor. When 635 /// NumPredsLeft reaches zero, release the successor node. 636 /// 637 /// FIXME: Adjust SuccSU height based on MinLatency. 638 void ScheduleDAGMI::releaseSucc(SUnit *SU, SDep *SuccEdge) { 639 SUnit *SuccSU = SuccEdge->getSUnit(); 640 641 if (SuccEdge->isWeak()) { 642 --SuccSU->WeakPredsLeft; 643 if (SuccEdge->isCluster()) 644 NextClusterSucc = SuccSU; 645 return; 646 } 647 #ifndef NDEBUG 648 if (SuccSU->NumPredsLeft == 0) { 649 dbgs() << "*** Scheduling failed! ***\n"; 650 dumpNode(*SuccSU); 651 dbgs() << " has been released too many times!\n"; 652 llvm_unreachable(nullptr); 653 } 654 #endif 655 // SU->TopReadyCycle was set to CurrCycle when it was scheduled. However, 656 // CurrCycle may have advanced since then. 657 if (SuccSU->TopReadyCycle < SU->TopReadyCycle + SuccEdge->getLatency()) 658 SuccSU->TopReadyCycle = SU->TopReadyCycle + SuccEdge->getLatency(); 659 660 --SuccSU->NumPredsLeft; 661 if (SuccSU->NumPredsLeft == 0 && SuccSU != &ExitSU) 662 SchedImpl->releaseTopNode(SuccSU); 663 } 664 665 /// releaseSuccessors - Call releaseSucc on each of SU's successors. 666 void ScheduleDAGMI::releaseSuccessors(SUnit *SU) { 667 for (SDep &Succ : SU->Succs) 668 releaseSucc(SU, &Succ); 669 } 670 671 /// ReleasePred - Decrement the NumSuccsLeft count of a predecessor. When 672 /// NumSuccsLeft reaches zero, release the predecessor node. 673 /// 674 /// FIXME: Adjust PredSU height based on MinLatency. 675 void ScheduleDAGMI::releasePred(SUnit *SU, SDep *PredEdge) { 676 SUnit *PredSU = PredEdge->getSUnit(); 677 678 if (PredEdge->isWeak()) { 679 --PredSU->WeakSuccsLeft; 680 if (PredEdge->isCluster()) 681 NextClusterPred = PredSU; 682 return; 683 } 684 #ifndef NDEBUG 685 if (PredSU->NumSuccsLeft == 0) { 686 dbgs() << "*** Scheduling failed! ***\n"; 687 dumpNode(*PredSU); 688 dbgs() << " has been released too many times!\n"; 689 llvm_unreachable(nullptr); 690 } 691 #endif 692 // SU->BotReadyCycle was set to CurrCycle when it was scheduled. However, 693 // CurrCycle may have advanced since then. 694 if (PredSU->BotReadyCycle < SU->BotReadyCycle + PredEdge->getLatency()) 695 PredSU->BotReadyCycle = SU->BotReadyCycle + PredEdge->getLatency(); 696 697 --PredSU->NumSuccsLeft; 698 if (PredSU->NumSuccsLeft == 0 && PredSU != &EntrySU) 699 SchedImpl->releaseBottomNode(PredSU); 700 } 701 702 /// releasePredecessors - Call releasePred on each of SU's predecessors. 703 void ScheduleDAGMI::releasePredecessors(SUnit *SU) { 704 for (SDep &Pred : SU->Preds) 705 releasePred(SU, &Pred); 706 } 707 708 void ScheduleDAGMI::startBlock(MachineBasicBlock *bb) { 709 ScheduleDAGInstrs::startBlock(bb); 710 SchedImpl->enterMBB(bb); 711 } 712 713 void ScheduleDAGMI::finishBlock() { 714 SchedImpl->leaveMBB(); 715 ScheduleDAGInstrs::finishBlock(); 716 } 717 718 /// enterRegion - Called back from MachineScheduler::runOnMachineFunction after 719 /// crossing a scheduling boundary. [begin, end) includes all instructions in 720 /// the region, including the boundary itself and single-instruction regions 721 /// that don't get scheduled. 722 void ScheduleDAGMI::enterRegion(MachineBasicBlock *bb, 723 MachineBasicBlock::iterator begin, 724 MachineBasicBlock::iterator end, 725 unsigned regioninstrs) 726 { 727 ScheduleDAGInstrs::enterRegion(bb, begin, end, regioninstrs); 728 729 SchedImpl->initPolicy(begin, end, regioninstrs); 730 } 731 732 /// This is normally called from the main scheduler loop but may also be invoked 733 /// by the scheduling strategy to perform additional code motion. 734 void ScheduleDAGMI::moveInstruction( 735 MachineInstr *MI, MachineBasicBlock::iterator InsertPos) { 736 // Advance RegionBegin if the first instruction moves down. 737 if (&*RegionBegin == MI) 738 ++RegionBegin; 739 740 // Update the instruction stream. 741 BB->splice(InsertPos, BB, MI); 742 743 // Update LiveIntervals 744 if (LIS) 745 LIS->handleMove(*MI, /*UpdateFlags=*/true); 746 747 // Recede RegionBegin if an instruction moves above the first. 748 if (RegionBegin == InsertPos) 749 RegionBegin = MI; 750 } 751 752 bool ScheduleDAGMI::checkSchedLimit() { 753 #if LLVM_ENABLE_ABI_BREAKING_CHECKS && !defined(NDEBUG) 754 if (NumInstrsScheduled == MISchedCutoff && MISchedCutoff != ~0U) { 755 CurrentTop = CurrentBottom; 756 return false; 757 } 758 ++NumInstrsScheduled; 759 #endif 760 return true; 761 } 762 763 /// Per-region scheduling driver, called back from 764 /// MachineScheduler::runOnMachineFunction. This is a simplified driver that 765 /// does not consider liveness or register pressure. It is useful for PostRA 766 /// scheduling and potentially other custom schedulers. 767 void ScheduleDAGMI::schedule() { 768 LLVM_DEBUG(dbgs() << "ScheduleDAGMI::schedule starting\n"); 769 LLVM_DEBUG(SchedImpl->dumpPolicy()); 770 771 // Build the DAG. 772 buildSchedGraph(AA); 773 774 postprocessDAG(); 775 776 SmallVector<SUnit*, 8> TopRoots, BotRoots; 777 findRootsAndBiasEdges(TopRoots, BotRoots); 778 779 LLVM_DEBUG(dump()); 780 if (PrintDAGs) dump(); 781 if (ViewMISchedDAGs) viewGraph(); 782 783 // Initialize the strategy before modifying the DAG. 784 // This may initialize a DFSResult to be used for queue priority. 785 SchedImpl->initialize(this); 786 787 // Initialize ready queues now that the DAG and priority data are finalized. 788 initQueues(TopRoots, BotRoots); 789 790 bool IsTopNode = false; 791 while (true) { 792 LLVM_DEBUG(dbgs() << "** ScheduleDAGMI::schedule picking next node\n"); 793 SUnit *SU = SchedImpl->pickNode(IsTopNode); 794 if (!SU) break; 795 796 assert(!SU->isScheduled && "Node already scheduled"); 797 if (!checkSchedLimit()) 798 break; 799 800 MachineInstr *MI = SU->getInstr(); 801 if (IsTopNode) { 802 assert(SU->isTopReady() && "node still has unscheduled dependencies"); 803 if (&*CurrentTop == MI) 804 CurrentTop = nextIfDebug(++CurrentTop, CurrentBottom); 805 else 806 moveInstruction(MI, CurrentTop); 807 } else { 808 assert(SU->isBottomReady() && "node still has unscheduled dependencies"); 809 MachineBasicBlock::iterator priorII = 810 priorNonDebug(CurrentBottom, CurrentTop); 811 if (&*priorII == MI) 812 CurrentBottom = priorII; 813 else { 814 if (&*CurrentTop == MI) 815 CurrentTop = nextIfDebug(++CurrentTop, priorII); 816 moveInstruction(MI, CurrentBottom); 817 CurrentBottom = MI; 818 } 819 } 820 // Notify the scheduling strategy before updating the DAG. 821 // This sets the scheduled node's ReadyCycle to CurrCycle. When updateQueues 822 // runs, it can then use the accurate ReadyCycle time to determine whether 823 // newly released nodes can move to the readyQ. 824 SchedImpl->schedNode(SU, IsTopNode); 825 826 updateQueues(SU, IsTopNode); 827 } 828 assert(CurrentTop == CurrentBottom && "Nonempty unscheduled zone."); 829 830 placeDebugValues(); 831 832 LLVM_DEBUG({ 833 dbgs() << "*** Final schedule for " 834 << printMBBReference(*begin()->getParent()) << " ***\n"; 835 dumpSchedule(); 836 dbgs() << '\n'; 837 }); 838 } 839 840 /// Apply each ScheduleDAGMutation step in order. 841 void ScheduleDAGMI::postprocessDAG() { 842 for (auto &m : Mutations) 843 m->apply(this); 844 } 845 846 void ScheduleDAGMI:: 847 findRootsAndBiasEdges(SmallVectorImpl<SUnit*> &TopRoots, 848 SmallVectorImpl<SUnit*> &BotRoots) { 849 for (SUnit &SU : SUnits) { 850 assert(!SU.isBoundaryNode() && "Boundary node should not be in SUnits"); 851 852 // Order predecessors so DFSResult follows the critical path. 853 SU.biasCriticalPath(); 854 855 // A SUnit is ready to top schedule if it has no predecessors. 856 if (!SU.NumPredsLeft) 857 TopRoots.push_back(&SU); 858 // A SUnit is ready to bottom schedule if it has no successors. 859 if (!SU.NumSuccsLeft) 860 BotRoots.push_back(&SU); 861 } 862 ExitSU.biasCriticalPath(); 863 } 864 865 /// Identify DAG roots and setup scheduler queues. 866 void ScheduleDAGMI::initQueues(ArrayRef<SUnit*> TopRoots, 867 ArrayRef<SUnit*> BotRoots) { 868 NextClusterSucc = nullptr; 869 NextClusterPred = nullptr; 870 871 // Release all DAG roots for scheduling, not including EntrySU/ExitSU. 872 // 873 // Nodes with unreleased weak edges can still be roots. 874 // Release top roots in forward order. 875 for (SUnit *SU : TopRoots) 876 SchedImpl->releaseTopNode(SU); 877 878 // Release bottom roots in reverse order so the higher priority nodes appear 879 // first. This is more natural and slightly more efficient. 880 for (SmallVectorImpl<SUnit*>::const_reverse_iterator 881 I = BotRoots.rbegin(), E = BotRoots.rend(); I != E; ++I) { 882 SchedImpl->releaseBottomNode(*I); 883 } 884 885 releaseSuccessors(&EntrySU); 886 releasePredecessors(&ExitSU); 887 888 SchedImpl->registerRoots(); 889 890 // Advance past initial DebugValues. 891 CurrentTop = nextIfDebug(RegionBegin, RegionEnd); 892 CurrentBottom = RegionEnd; 893 } 894 895 /// Update scheduler queues after scheduling an instruction. 896 void ScheduleDAGMI::updateQueues(SUnit *SU, bool IsTopNode) { 897 // Release dependent instructions for scheduling. 898 if (IsTopNode) 899 releaseSuccessors(SU); 900 else 901 releasePredecessors(SU); 902 903 SU->isScheduled = true; 904 } 905 906 /// Reinsert any remaining debug_values, just like the PostRA scheduler. 907 void ScheduleDAGMI::placeDebugValues() { 908 // If first instruction was a DBG_VALUE then put it back. 909 if (FirstDbgValue) { 910 BB->splice(RegionBegin, BB, FirstDbgValue); 911 RegionBegin = FirstDbgValue; 912 } 913 914 for (std::vector<std::pair<MachineInstr *, MachineInstr *>>::iterator 915 DI = DbgValues.end(), DE = DbgValues.begin(); DI != DE; --DI) { 916 std::pair<MachineInstr *, MachineInstr *> P = *std::prev(DI); 917 MachineInstr *DbgValue = P.first; 918 MachineBasicBlock::iterator OrigPrevMI = P.second; 919 if (&*RegionBegin == DbgValue) 920 ++RegionBegin; 921 BB->splice(std::next(OrigPrevMI), BB, DbgValue); 922 if (RegionEnd != BB->end() && OrigPrevMI == &*RegionEnd) 923 RegionEnd = DbgValue; 924 } 925 } 926 927 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) 928 LLVM_DUMP_METHOD void ScheduleDAGMI::dumpSchedule() const { 929 for (MachineInstr &MI : *this) { 930 if (SUnit *SU = getSUnit(&MI)) 931 dumpNode(*SU); 932 else 933 dbgs() << "Missing SUnit\n"; 934 } 935 } 936 #endif 937 938 //===----------------------------------------------------------------------===// 939 // ScheduleDAGMILive - Base class for MachineInstr scheduling with LiveIntervals 940 // preservation. 941 //===----------------------------------------------------------------------===// 942 943 ScheduleDAGMILive::~ScheduleDAGMILive() { 944 delete DFSResult; 945 } 946 947 void ScheduleDAGMILive::collectVRegUses(SUnit &SU) { 948 const MachineInstr &MI = *SU.getInstr(); 949 for (const MachineOperand &MO : MI.operands()) { 950 if (!MO.isReg()) 951 continue; 952 if (!MO.readsReg()) 953 continue; 954 if (TrackLaneMasks && !MO.isUse()) 955 continue; 956 957 Register Reg = MO.getReg(); 958 if (!Register::isVirtualRegister(Reg)) 959 continue; 960 961 // Ignore re-defs. 962 if (TrackLaneMasks) { 963 bool FoundDef = false; 964 for (const MachineOperand &MO2 : MI.operands()) { 965 if (MO2.isReg() && MO2.isDef() && MO2.getReg() == Reg && !MO2.isDead()) { 966 FoundDef = true; 967 break; 968 } 969 } 970 if (FoundDef) 971 continue; 972 } 973 974 // Record this local VReg use. 975 VReg2SUnitMultiMap::iterator UI = VRegUses.find(Reg); 976 for (; UI != VRegUses.end(); ++UI) { 977 if (UI->SU == &SU) 978 break; 979 } 980 if (UI == VRegUses.end()) 981 VRegUses.insert(VReg2SUnit(Reg, LaneBitmask::getNone(), &SU)); 982 } 983 } 984 985 /// enterRegion - Called back from MachineScheduler::runOnMachineFunction after 986 /// crossing a scheduling boundary. [begin, end) includes all instructions in 987 /// the region, including the boundary itself and single-instruction regions 988 /// that don't get scheduled. 989 void ScheduleDAGMILive::enterRegion(MachineBasicBlock *bb, 990 MachineBasicBlock::iterator begin, 991 MachineBasicBlock::iterator end, 992 unsigned regioninstrs) 993 { 994 // ScheduleDAGMI initializes SchedImpl's per-region policy. 995 ScheduleDAGMI::enterRegion(bb, begin, end, regioninstrs); 996 997 // For convenience remember the end of the liveness region. 998 LiveRegionEnd = (RegionEnd == bb->end()) ? RegionEnd : std::next(RegionEnd); 999 1000 SUPressureDiffs.clear(); 1001 1002 ShouldTrackPressure = SchedImpl->shouldTrackPressure(); 1003 ShouldTrackLaneMasks = SchedImpl->shouldTrackLaneMasks(); 1004 1005 assert((!ShouldTrackLaneMasks || ShouldTrackPressure) && 1006 "ShouldTrackLaneMasks requires ShouldTrackPressure"); 1007 } 1008 1009 // Setup the register pressure trackers for the top scheduled and bottom 1010 // scheduled regions. 1011 void ScheduleDAGMILive::initRegPressure() { 1012 VRegUses.clear(); 1013 VRegUses.setUniverse(MRI.getNumVirtRegs()); 1014 for (SUnit &SU : SUnits) 1015 collectVRegUses(SU); 1016 1017 TopRPTracker.init(&MF, RegClassInfo, LIS, BB, RegionBegin, 1018 ShouldTrackLaneMasks, false); 1019 BotRPTracker.init(&MF, RegClassInfo, LIS, BB, LiveRegionEnd, 1020 ShouldTrackLaneMasks, false); 1021 1022 // Close the RPTracker to finalize live ins. 1023 RPTracker.closeRegion(); 1024 1025 LLVM_DEBUG(RPTracker.dump()); 1026 1027 // Initialize the live ins and live outs. 1028 TopRPTracker.addLiveRegs(RPTracker.getPressure().LiveInRegs); 1029 BotRPTracker.addLiveRegs(RPTracker.getPressure().LiveOutRegs); 1030 1031 // Close one end of the tracker so we can call 1032 // getMaxUpward/DownwardPressureDelta before advancing across any 1033 // instructions. This converts currently live regs into live ins/outs. 1034 TopRPTracker.closeTop(); 1035 BotRPTracker.closeBottom(); 1036 1037 BotRPTracker.initLiveThru(RPTracker); 1038 if (!BotRPTracker.getLiveThru().empty()) { 1039 TopRPTracker.initLiveThru(BotRPTracker.getLiveThru()); 1040 LLVM_DEBUG(dbgs() << "Live Thru: "; 1041 dumpRegSetPressure(BotRPTracker.getLiveThru(), TRI)); 1042 }; 1043 1044 // For each live out vreg reduce the pressure change associated with other 1045 // uses of the same vreg below the live-out reaching def. 1046 updatePressureDiffs(RPTracker.getPressure().LiveOutRegs); 1047 1048 // Account for liveness generated by the region boundary. 1049 if (LiveRegionEnd != RegionEnd) { 1050 SmallVector<RegisterMaskPair, 8> LiveUses; 1051 BotRPTracker.recede(&LiveUses); 1052 updatePressureDiffs(LiveUses); 1053 } 1054 1055 LLVM_DEBUG(dbgs() << "Top Pressure:\n"; 1056 dumpRegSetPressure(TopRPTracker.getRegSetPressureAtPos(), TRI); 1057 dbgs() << "Bottom Pressure:\n"; 1058 dumpRegSetPressure(BotRPTracker.getRegSetPressureAtPos(), TRI);); 1059 1060 assert((BotRPTracker.getPos() == RegionEnd || 1061 (RegionEnd->isDebugInstr() && 1062 BotRPTracker.getPos() == priorNonDebug(RegionEnd, RegionBegin))) && 1063 "Can't find the region bottom"); 1064 1065 // Cache the list of excess pressure sets in this region. This will also track 1066 // the max pressure in the scheduled code for these sets. 1067 RegionCriticalPSets.clear(); 1068 const std::vector<unsigned> &RegionPressure = 1069 RPTracker.getPressure().MaxSetPressure; 1070 for (unsigned i = 0, e = RegionPressure.size(); i < e; ++i) { 1071 unsigned Limit = RegClassInfo->getRegPressureSetLimit(i); 1072 if (RegionPressure[i] > Limit) { 1073 LLVM_DEBUG(dbgs() << TRI->getRegPressureSetName(i) << " Limit " << Limit 1074 << " Actual " << RegionPressure[i] << "\n"); 1075 RegionCriticalPSets.push_back(PressureChange(i)); 1076 } 1077 } 1078 LLVM_DEBUG(dbgs() << "Excess PSets: "; 1079 for (const PressureChange &RCPS 1080 : RegionCriticalPSets) dbgs() 1081 << TRI->getRegPressureSetName(RCPS.getPSet()) << " "; 1082 dbgs() << "\n"); 1083 } 1084 1085 void ScheduleDAGMILive:: 1086 updateScheduledPressure(const SUnit *SU, 1087 const std::vector<unsigned> &NewMaxPressure) { 1088 const PressureDiff &PDiff = getPressureDiff(SU); 1089 unsigned CritIdx = 0, CritEnd = RegionCriticalPSets.size(); 1090 for (const PressureChange &PC : PDiff) { 1091 if (!PC.isValid()) 1092 break; 1093 unsigned ID = PC.getPSet(); 1094 while (CritIdx != CritEnd && RegionCriticalPSets[CritIdx].getPSet() < ID) 1095 ++CritIdx; 1096 if (CritIdx != CritEnd && RegionCriticalPSets[CritIdx].getPSet() == ID) { 1097 if ((int)NewMaxPressure[ID] > RegionCriticalPSets[CritIdx].getUnitInc() 1098 && NewMaxPressure[ID] <= (unsigned)std::numeric_limits<int16_t>::max()) 1099 RegionCriticalPSets[CritIdx].setUnitInc(NewMaxPressure[ID]); 1100 } 1101 unsigned Limit = RegClassInfo->getRegPressureSetLimit(ID); 1102 if (NewMaxPressure[ID] >= Limit - 2) { 1103 LLVM_DEBUG(dbgs() << " " << TRI->getRegPressureSetName(ID) << ": " 1104 << NewMaxPressure[ID] 1105 << ((NewMaxPressure[ID] > Limit) ? " > " : " <= ") 1106 << Limit << "(+ " << BotRPTracker.getLiveThru()[ID] 1107 << " livethru)\n"); 1108 } 1109 } 1110 } 1111 1112 /// Update the PressureDiff array for liveness after scheduling this 1113 /// instruction. 1114 void ScheduleDAGMILive::updatePressureDiffs( 1115 ArrayRef<RegisterMaskPair> LiveUses) { 1116 for (const RegisterMaskPair &P : LiveUses) { 1117 Register Reg = P.RegUnit; 1118 /// FIXME: Currently assuming single-use physregs. 1119 if (!Register::isVirtualRegister(Reg)) 1120 continue; 1121 1122 if (ShouldTrackLaneMasks) { 1123 // If the register has just become live then other uses won't change 1124 // this fact anymore => decrement pressure. 1125 // If the register has just become dead then other uses make it come 1126 // back to life => increment pressure. 1127 bool Decrement = P.LaneMask.any(); 1128 1129 for (const VReg2SUnit &V2SU 1130 : make_range(VRegUses.find(Reg), VRegUses.end())) { 1131 SUnit &SU = *V2SU.SU; 1132 if (SU.isScheduled || &SU == &ExitSU) 1133 continue; 1134 1135 PressureDiff &PDiff = getPressureDiff(&SU); 1136 PDiff.addPressureChange(Reg, Decrement, &MRI); 1137 LLVM_DEBUG(dbgs() << " UpdateRegP: SU(" << SU.NodeNum << ") " 1138 << printReg(Reg, TRI) << ':' 1139 << PrintLaneMask(P.LaneMask) << ' ' << *SU.getInstr(); 1140 dbgs() << " to "; PDiff.dump(*TRI);); 1141 } 1142 } else { 1143 assert(P.LaneMask.any()); 1144 LLVM_DEBUG(dbgs() << " LiveReg: " << printVRegOrUnit(Reg, TRI) << "\n"); 1145 // This may be called before CurrentBottom has been initialized. However, 1146 // BotRPTracker must have a valid position. We want the value live into the 1147 // instruction or live out of the block, so ask for the previous 1148 // instruction's live-out. 1149 const LiveInterval &LI = LIS->getInterval(Reg); 1150 VNInfo *VNI; 1151 MachineBasicBlock::const_iterator I = 1152 nextIfDebug(BotRPTracker.getPos(), BB->end()); 1153 if (I == BB->end()) 1154 VNI = LI.getVNInfoBefore(LIS->getMBBEndIdx(BB)); 1155 else { 1156 LiveQueryResult LRQ = LI.Query(LIS->getInstructionIndex(*I)); 1157 VNI = LRQ.valueIn(); 1158 } 1159 // RegisterPressureTracker guarantees that readsReg is true for LiveUses. 1160 assert(VNI && "No live value at use."); 1161 for (const VReg2SUnit &V2SU 1162 : make_range(VRegUses.find(Reg), VRegUses.end())) { 1163 SUnit *SU = V2SU.SU; 1164 // If this use comes before the reaching def, it cannot be a last use, 1165 // so decrease its pressure change. 1166 if (!SU->isScheduled && SU != &ExitSU) { 1167 LiveQueryResult LRQ = 1168 LI.Query(LIS->getInstructionIndex(*SU->getInstr())); 1169 if (LRQ.valueIn() == VNI) { 1170 PressureDiff &PDiff = getPressureDiff(SU); 1171 PDiff.addPressureChange(Reg, true, &MRI); 1172 LLVM_DEBUG(dbgs() << " UpdateRegP: SU(" << SU->NodeNum << ") " 1173 << *SU->getInstr(); 1174 dbgs() << " to "; PDiff.dump(*TRI);); 1175 } 1176 } 1177 } 1178 } 1179 } 1180 } 1181 1182 void ScheduleDAGMILive::dump() const { 1183 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) 1184 if (EntrySU.getInstr() != nullptr) 1185 dumpNodeAll(EntrySU); 1186 for (const SUnit &SU : SUnits) { 1187 dumpNodeAll(SU); 1188 if (ShouldTrackPressure) { 1189 dbgs() << " Pressure Diff : "; 1190 getPressureDiff(&SU).dump(*TRI); 1191 } 1192 dbgs() << " Single Issue : "; 1193 if (SchedModel.mustBeginGroup(SU.getInstr()) && 1194 SchedModel.mustEndGroup(SU.getInstr())) 1195 dbgs() << "true;"; 1196 else 1197 dbgs() << "false;"; 1198 dbgs() << '\n'; 1199 } 1200 if (ExitSU.getInstr() != nullptr) 1201 dumpNodeAll(ExitSU); 1202 #endif 1203 } 1204 1205 /// schedule - Called back from MachineScheduler::runOnMachineFunction 1206 /// after setting up the current scheduling region. [RegionBegin, RegionEnd) 1207 /// only includes instructions that have DAG nodes, not scheduling boundaries. 1208 /// 1209 /// This is a skeletal driver, with all the functionality pushed into helpers, 1210 /// so that it can be easily extended by experimental schedulers. Generally, 1211 /// implementing MachineSchedStrategy should be sufficient to implement a new 1212 /// scheduling algorithm. However, if a scheduler further subclasses 1213 /// ScheduleDAGMILive then it will want to override this virtual method in order 1214 /// to update any specialized state. 1215 void ScheduleDAGMILive::schedule() { 1216 LLVM_DEBUG(dbgs() << "ScheduleDAGMILive::schedule starting\n"); 1217 LLVM_DEBUG(SchedImpl->dumpPolicy()); 1218 buildDAGWithRegPressure(); 1219 1220 postprocessDAG(); 1221 1222 SmallVector<SUnit*, 8> TopRoots, BotRoots; 1223 findRootsAndBiasEdges(TopRoots, BotRoots); 1224 1225 // Initialize the strategy before modifying the DAG. 1226 // This may initialize a DFSResult to be used for queue priority. 1227 SchedImpl->initialize(this); 1228 1229 LLVM_DEBUG(dump()); 1230 if (PrintDAGs) dump(); 1231 if (ViewMISchedDAGs) viewGraph(); 1232 1233 // Initialize ready queues now that the DAG and priority data are finalized. 1234 initQueues(TopRoots, BotRoots); 1235 1236 bool IsTopNode = false; 1237 while (true) { 1238 LLVM_DEBUG(dbgs() << "** ScheduleDAGMILive::schedule picking next node\n"); 1239 SUnit *SU = SchedImpl->pickNode(IsTopNode); 1240 if (!SU) break; 1241 1242 assert(!SU->isScheduled && "Node already scheduled"); 1243 if (!checkSchedLimit()) 1244 break; 1245 1246 scheduleMI(SU, IsTopNode); 1247 1248 if (DFSResult) { 1249 unsigned SubtreeID = DFSResult->getSubtreeID(SU); 1250 if (!ScheduledTrees.test(SubtreeID)) { 1251 ScheduledTrees.set(SubtreeID); 1252 DFSResult->scheduleTree(SubtreeID); 1253 SchedImpl->scheduleTree(SubtreeID); 1254 } 1255 } 1256 1257 // Notify the scheduling strategy after updating the DAG. 1258 SchedImpl->schedNode(SU, IsTopNode); 1259 1260 updateQueues(SU, IsTopNode); 1261 } 1262 assert(CurrentTop == CurrentBottom && "Nonempty unscheduled zone."); 1263 1264 placeDebugValues(); 1265 1266 LLVM_DEBUG({ 1267 dbgs() << "*** Final schedule for " 1268 << printMBBReference(*begin()->getParent()) << " ***\n"; 1269 dumpSchedule(); 1270 dbgs() << '\n'; 1271 }); 1272 } 1273 1274 /// Build the DAG and setup three register pressure trackers. 1275 void ScheduleDAGMILive::buildDAGWithRegPressure() { 1276 if (!ShouldTrackPressure) { 1277 RPTracker.reset(); 1278 RegionCriticalPSets.clear(); 1279 buildSchedGraph(AA); 1280 return; 1281 } 1282 1283 // Initialize the register pressure tracker used by buildSchedGraph. 1284 RPTracker.init(&MF, RegClassInfo, LIS, BB, LiveRegionEnd, 1285 ShouldTrackLaneMasks, /*TrackUntiedDefs=*/true); 1286 1287 // Account for liveness generate by the region boundary. 1288 if (LiveRegionEnd != RegionEnd) 1289 RPTracker.recede(); 1290 1291 // Build the DAG, and compute current register pressure. 1292 buildSchedGraph(AA, &RPTracker, &SUPressureDiffs, LIS, ShouldTrackLaneMasks); 1293 1294 // Initialize top/bottom trackers after computing region pressure. 1295 initRegPressure(); 1296 } 1297 1298 void ScheduleDAGMILive::computeDFSResult() { 1299 if (!DFSResult) 1300 DFSResult = new SchedDFSResult(/*BottomU*/true, MinSubtreeSize); 1301 DFSResult->clear(); 1302 ScheduledTrees.clear(); 1303 DFSResult->resize(SUnits.size()); 1304 DFSResult->compute(SUnits); 1305 ScheduledTrees.resize(DFSResult->getNumSubtrees()); 1306 } 1307 1308 /// Compute the max cyclic critical path through the DAG. The scheduling DAG 1309 /// only provides the critical path for single block loops. To handle loops that 1310 /// span blocks, we could use the vreg path latencies provided by 1311 /// MachineTraceMetrics instead. However, MachineTraceMetrics is not currently 1312 /// available for use in the scheduler. 1313 /// 1314 /// The cyclic path estimation identifies a def-use pair that crosses the back 1315 /// edge and considers the depth and height of the nodes. For example, consider 1316 /// the following instruction sequence where each instruction has unit latency 1317 /// and defines an eponymous virtual register: 1318 /// 1319 /// a->b(a,c)->c(b)->d(c)->exit 1320 /// 1321 /// The cyclic critical path is a two cycles: b->c->b 1322 /// The acyclic critical path is four cycles: a->b->c->d->exit 1323 /// LiveOutHeight = height(c) = len(c->d->exit) = 2 1324 /// LiveOutDepth = depth(c) + 1 = len(a->b->c) + 1 = 3 1325 /// LiveInHeight = height(b) + 1 = len(b->c->d->exit) + 1 = 4 1326 /// LiveInDepth = depth(b) = len(a->b) = 1 1327 /// 1328 /// LiveOutDepth - LiveInDepth = 3 - 1 = 2 1329 /// LiveInHeight - LiveOutHeight = 4 - 2 = 2 1330 /// CyclicCriticalPath = min(2, 2) = 2 1331 /// 1332 /// This could be relevant to PostRA scheduling, but is currently implemented 1333 /// assuming LiveIntervals. 1334 unsigned ScheduleDAGMILive::computeCyclicCriticalPath() { 1335 // This only applies to single block loop. 1336 if (!BB->isSuccessor(BB)) 1337 return 0; 1338 1339 unsigned MaxCyclicLatency = 0; 1340 // Visit each live out vreg def to find def/use pairs that cross iterations. 1341 for (const RegisterMaskPair &P : RPTracker.getPressure().LiveOutRegs) { 1342 Register Reg = P.RegUnit; 1343 if (!Register::isVirtualRegister(Reg)) 1344 continue; 1345 const LiveInterval &LI = LIS->getInterval(Reg); 1346 const VNInfo *DefVNI = LI.getVNInfoBefore(LIS->getMBBEndIdx(BB)); 1347 if (!DefVNI) 1348 continue; 1349 1350 MachineInstr *DefMI = LIS->getInstructionFromIndex(DefVNI->def); 1351 const SUnit *DefSU = getSUnit(DefMI); 1352 if (!DefSU) 1353 continue; 1354 1355 unsigned LiveOutHeight = DefSU->getHeight(); 1356 unsigned LiveOutDepth = DefSU->getDepth() + DefSU->Latency; 1357 // Visit all local users of the vreg def. 1358 for (const VReg2SUnit &V2SU 1359 : make_range(VRegUses.find(Reg), VRegUses.end())) { 1360 SUnit *SU = V2SU.SU; 1361 if (SU == &ExitSU) 1362 continue; 1363 1364 // Only consider uses of the phi. 1365 LiveQueryResult LRQ = LI.Query(LIS->getInstructionIndex(*SU->getInstr())); 1366 if (!LRQ.valueIn()->isPHIDef()) 1367 continue; 1368 1369 // Assume that a path spanning two iterations is a cycle, which could 1370 // overestimate in strange cases. This allows cyclic latency to be 1371 // estimated as the minimum slack of the vreg's depth or height. 1372 unsigned CyclicLatency = 0; 1373 if (LiveOutDepth > SU->getDepth()) 1374 CyclicLatency = LiveOutDepth - SU->getDepth(); 1375 1376 unsigned LiveInHeight = SU->getHeight() + DefSU->Latency; 1377 if (LiveInHeight > LiveOutHeight) { 1378 if (LiveInHeight - LiveOutHeight < CyclicLatency) 1379 CyclicLatency = LiveInHeight - LiveOutHeight; 1380 } else 1381 CyclicLatency = 0; 1382 1383 LLVM_DEBUG(dbgs() << "Cyclic Path: SU(" << DefSU->NodeNum << ") -> SU(" 1384 << SU->NodeNum << ") = " << CyclicLatency << "c\n"); 1385 if (CyclicLatency > MaxCyclicLatency) 1386 MaxCyclicLatency = CyclicLatency; 1387 } 1388 } 1389 LLVM_DEBUG(dbgs() << "Cyclic Critical Path: " << MaxCyclicLatency << "c\n"); 1390 return MaxCyclicLatency; 1391 } 1392 1393 /// Release ExitSU predecessors and setup scheduler queues. Re-position 1394 /// the Top RP tracker in case the region beginning has changed. 1395 void ScheduleDAGMILive::initQueues(ArrayRef<SUnit*> TopRoots, 1396 ArrayRef<SUnit*> BotRoots) { 1397 ScheduleDAGMI::initQueues(TopRoots, BotRoots); 1398 if (ShouldTrackPressure) { 1399 assert(TopRPTracker.getPos() == RegionBegin && "bad initial Top tracker"); 1400 TopRPTracker.setPos(CurrentTop); 1401 } 1402 } 1403 1404 /// Move an instruction and update register pressure. 1405 void ScheduleDAGMILive::scheduleMI(SUnit *SU, bool IsTopNode) { 1406 // Move the instruction to its new location in the instruction stream. 1407 MachineInstr *MI = SU->getInstr(); 1408 1409 if (IsTopNode) { 1410 assert(SU->isTopReady() && "node still has unscheduled dependencies"); 1411 if (&*CurrentTop == MI) 1412 CurrentTop = nextIfDebug(++CurrentTop, CurrentBottom); 1413 else { 1414 moveInstruction(MI, CurrentTop); 1415 TopRPTracker.setPos(MI); 1416 } 1417 1418 if (ShouldTrackPressure) { 1419 // Update top scheduled pressure. 1420 RegisterOperands RegOpers; 1421 RegOpers.collect(*MI, *TRI, MRI, ShouldTrackLaneMasks, false); 1422 if (ShouldTrackLaneMasks) { 1423 // Adjust liveness and add missing dead+read-undef flags. 1424 SlotIndex SlotIdx = LIS->getInstructionIndex(*MI).getRegSlot(); 1425 RegOpers.adjustLaneLiveness(*LIS, MRI, SlotIdx, MI); 1426 } else { 1427 // Adjust for missing dead-def flags. 1428 RegOpers.detectDeadDefs(*MI, *LIS); 1429 } 1430 1431 TopRPTracker.advance(RegOpers); 1432 assert(TopRPTracker.getPos() == CurrentTop && "out of sync"); 1433 LLVM_DEBUG(dbgs() << "Top Pressure:\n"; dumpRegSetPressure( 1434 TopRPTracker.getRegSetPressureAtPos(), TRI);); 1435 1436 updateScheduledPressure(SU, TopRPTracker.getPressure().MaxSetPressure); 1437 } 1438 } else { 1439 assert(SU->isBottomReady() && "node still has unscheduled dependencies"); 1440 MachineBasicBlock::iterator priorII = 1441 priorNonDebug(CurrentBottom, CurrentTop); 1442 if (&*priorII == MI) 1443 CurrentBottom = priorII; 1444 else { 1445 if (&*CurrentTop == MI) { 1446 CurrentTop = nextIfDebug(++CurrentTop, priorII); 1447 TopRPTracker.setPos(CurrentTop); 1448 } 1449 moveInstruction(MI, CurrentBottom); 1450 CurrentBottom = MI; 1451 BotRPTracker.setPos(CurrentBottom); 1452 } 1453 if (ShouldTrackPressure) { 1454 RegisterOperands RegOpers; 1455 RegOpers.collect(*MI, *TRI, MRI, ShouldTrackLaneMasks, false); 1456 if (ShouldTrackLaneMasks) { 1457 // Adjust liveness and add missing dead+read-undef flags. 1458 SlotIndex SlotIdx = LIS->getInstructionIndex(*MI).getRegSlot(); 1459 RegOpers.adjustLaneLiveness(*LIS, MRI, SlotIdx, MI); 1460 } else { 1461 // Adjust for missing dead-def flags. 1462 RegOpers.detectDeadDefs(*MI, *LIS); 1463 } 1464 1465 if (BotRPTracker.getPos() != CurrentBottom) 1466 BotRPTracker.recedeSkipDebugValues(); 1467 SmallVector<RegisterMaskPair, 8> LiveUses; 1468 BotRPTracker.recede(RegOpers, &LiveUses); 1469 assert(BotRPTracker.getPos() == CurrentBottom && "out of sync"); 1470 LLVM_DEBUG(dbgs() << "Bottom Pressure:\n"; dumpRegSetPressure( 1471 BotRPTracker.getRegSetPressureAtPos(), TRI);); 1472 1473 updateScheduledPressure(SU, BotRPTracker.getPressure().MaxSetPressure); 1474 updatePressureDiffs(LiveUses); 1475 } 1476 } 1477 } 1478 1479 //===----------------------------------------------------------------------===// 1480 // BaseMemOpClusterMutation - DAG post-processing to cluster loads or stores. 1481 //===----------------------------------------------------------------------===// 1482 1483 namespace { 1484 1485 /// Post-process the DAG to create cluster edges between neighboring 1486 /// loads or between neighboring stores. 1487 class BaseMemOpClusterMutation : public ScheduleDAGMutation { 1488 struct MemOpInfo { 1489 SUnit *SU; 1490 SmallVector<const MachineOperand *, 4> BaseOps; 1491 int64_t Offset; 1492 unsigned Width; 1493 1494 MemOpInfo(SUnit *SU, ArrayRef<const MachineOperand *> BaseOps, 1495 int64_t Offset, unsigned Width) 1496 : SU(SU), BaseOps(BaseOps.begin(), BaseOps.end()), Offset(Offset), 1497 Width(Width) {} 1498 1499 static bool Compare(const MachineOperand *const &A, 1500 const MachineOperand *const &B) { 1501 if (A->getType() != B->getType()) 1502 return A->getType() < B->getType(); 1503 if (A->isReg()) 1504 return A->getReg() < B->getReg(); 1505 if (A->isFI()) { 1506 const MachineFunction &MF = *A->getParent()->getParent()->getParent(); 1507 const TargetFrameLowering &TFI = *MF.getSubtarget().getFrameLowering(); 1508 bool StackGrowsDown = TFI.getStackGrowthDirection() == 1509 TargetFrameLowering::StackGrowsDown; 1510 return StackGrowsDown ? A->getIndex() > B->getIndex() 1511 : A->getIndex() < B->getIndex(); 1512 } 1513 1514 llvm_unreachable("MemOpClusterMutation only supports register or frame " 1515 "index bases."); 1516 } 1517 1518 bool operator<(const MemOpInfo &RHS) const { 1519 // FIXME: Don't compare everything twice. Maybe use C++20 three way 1520 // comparison instead when it's available. 1521 if (std::lexicographical_compare(BaseOps.begin(), BaseOps.end(), 1522 RHS.BaseOps.begin(), RHS.BaseOps.end(), 1523 Compare)) 1524 return true; 1525 if (std::lexicographical_compare(RHS.BaseOps.begin(), RHS.BaseOps.end(), 1526 BaseOps.begin(), BaseOps.end(), Compare)) 1527 return false; 1528 if (Offset != RHS.Offset) 1529 return Offset < RHS.Offset; 1530 return SU->NodeNum < RHS.SU->NodeNum; 1531 } 1532 }; 1533 1534 const TargetInstrInfo *TII; 1535 const TargetRegisterInfo *TRI; 1536 bool IsLoad; 1537 1538 public: 1539 BaseMemOpClusterMutation(const TargetInstrInfo *tii, 1540 const TargetRegisterInfo *tri, bool IsLoad) 1541 : TII(tii), TRI(tri), IsLoad(IsLoad) {} 1542 1543 void apply(ScheduleDAGInstrs *DAGInstrs) override; 1544 1545 protected: 1546 void clusterNeighboringMemOps(ArrayRef<MemOpInfo> MemOps, bool FastCluster, 1547 ScheduleDAGInstrs *DAG); 1548 void collectMemOpRecords(std::vector<SUnit> &SUnits, 1549 SmallVectorImpl<MemOpInfo> &MemOpRecords); 1550 bool groupMemOps(ArrayRef<MemOpInfo> MemOps, ScheduleDAGInstrs *DAG, 1551 DenseMap<unsigned, SmallVector<MemOpInfo, 32>> &Groups); 1552 }; 1553 1554 class StoreClusterMutation : public BaseMemOpClusterMutation { 1555 public: 1556 StoreClusterMutation(const TargetInstrInfo *tii, 1557 const TargetRegisterInfo *tri) 1558 : BaseMemOpClusterMutation(tii, tri, false) {} 1559 }; 1560 1561 class LoadClusterMutation : public BaseMemOpClusterMutation { 1562 public: 1563 LoadClusterMutation(const TargetInstrInfo *tii, const TargetRegisterInfo *tri) 1564 : BaseMemOpClusterMutation(tii, tri, true) {} 1565 }; 1566 1567 } // end anonymous namespace 1568 1569 namespace llvm { 1570 1571 std::unique_ptr<ScheduleDAGMutation> 1572 createLoadClusterDAGMutation(const TargetInstrInfo *TII, 1573 const TargetRegisterInfo *TRI) { 1574 return EnableMemOpCluster ? std::make_unique<LoadClusterMutation>(TII, TRI) 1575 : nullptr; 1576 } 1577 1578 std::unique_ptr<ScheduleDAGMutation> 1579 createStoreClusterDAGMutation(const TargetInstrInfo *TII, 1580 const TargetRegisterInfo *TRI) { 1581 return EnableMemOpCluster ? std::make_unique<StoreClusterMutation>(TII, TRI) 1582 : nullptr; 1583 } 1584 1585 } // end namespace llvm 1586 1587 // Sorting all the loads/stores first, then for each load/store, checking the 1588 // following load/store one by one, until reach the first non-dependent one and 1589 // call target hook to see if they can cluster. 1590 // If FastCluster is enabled, we assume that, all the loads/stores have been 1591 // preprocessed and now, they didn't have dependencies on each other. 1592 void BaseMemOpClusterMutation::clusterNeighboringMemOps( 1593 ArrayRef<MemOpInfo> MemOpRecords, bool FastCluster, 1594 ScheduleDAGInstrs *DAG) { 1595 // Keep track of the current cluster length and bytes for each SUnit. 1596 DenseMap<unsigned, std::pair<unsigned, unsigned>> SUnit2ClusterInfo; 1597 1598 // At this point, `MemOpRecords` array must hold atleast two mem ops. Try to 1599 // cluster mem ops collected within `MemOpRecords` array. 1600 for (unsigned Idx = 0, End = MemOpRecords.size(); Idx < (End - 1); ++Idx) { 1601 // Decision to cluster mem ops is taken based on target dependent logic 1602 auto MemOpa = MemOpRecords[Idx]; 1603 1604 // Seek for the next load/store to do the cluster. 1605 unsigned NextIdx = Idx + 1; 1606 for (; NextIdx < End; ++NextIdx) 1607 // Skip if MemOpb has been clustered already or has dependency with 1608 // MemOpa. 1609 if (!SUnit2ClusterInfo.count(MemOpRecords[NextIdx].SU->NodeNum) && 1610 (FastCluster || 1611 (!DAG->IsReachable(MemOpRecords[NextIdx].SU, MemOpa.SU) && 1612 !DAG->IsReachable(MemOpa.SU, MemOpRecords[NextIdx].SU)))) 1613 break; 1614 if (NextIdx == End) 1615 continue; 1616 1617 auto MemOpb = MemOpRecords[NextIdx]; 1618 unsigned ClusterLength = 2; 1619 unsigned CurrentClusterBytes = MemOpa.Width + MemOpb.Width; 1620 if (SUnit2ClusterInfo.count(MemOpa.SU->NodeNum)) { 1621 ClusterLength = SUnit2ClusterInfo[MemOpa.SU->NodeNum].first + 1; 1622 CurrentClusterBytes = 1623 SUnit2ClusterInfo[MemOpa.SU->NodeNum].second + MemOpb.Width; 1624 } 1625 1626 if (!TII->shouldClusterMemOps(MemOpa.BaseOps, MemOpb.BaseOps, ClusterLength, 1627 CurrentClusterBytes)) 1628 continue; 1629 1630 SUnit *SUa = MemOpa.SU; 1631 SUnit *SUb = MemOpb.SU; 1632 if (SUa->NodeNum > SUb->NodeNum) 1633 std::swap(SUa, SUb); 1634 1635 // FIXME: Is this check really required? 1636 if (!DAG->addEdge(SUb, SDep(SUa, SDep::Cluster))) 1637 continue; 1638 1639 LLVM_DEBUG(dbgs() << "Cluster ld/st SU(" << SUa->NodeNum << ") - SU(" 1640 << SUb->NodeNum << ")\n"); 1641 ++NumClustered; 1642 1643 if (IsLoad) { 1644 // Copy successor edges from SUa to SUb. Interleaving computation 1645 // dependent on SUa can prevent load combining due to register reuse. 1646 // Predecessor edges do not need to be copied from SUb to SUa since 1647 // nearby loads should have effectively the same inputs. 1648 for (const SDep &Succ : SUa->Succs) { 1649 if (Succ.getSUnit() == SUb) 1650 continue; 1651 LLVM_DEBUG(dbgs() << " Copy Succ SU(" << Succ.getSUnit()->NodeNum 1652 << ")\n"); 1653 DAG->addEdge(Succ.getSUnit(), SDep(SUb, SDep::Artificial)); 1654 } 1655 } else { 1656 // Copy predecessor edges from SUb to SUa to avoid the SUnits that 1657 // SUb dependent on scheduled in-between SUb and SUa. Successor edges 1658 // do not need to be copied from SUa to SUb since no one will depend 1659 // on stores. 1660 // Notice that, we don't need to care about the memory dependency as 1661 // we won't try to cluster them if they have any memory dependency. 1662 for (const SDep &Pred : SUb->Preds) { 1663 if (Pred.getSUnit() == SUa) 1664 continue; 1665 LLVM_DEBUG(dbgs() << " Copy Pred SU(" << Pred.getSUnit()->NodeNum 1666 << ")\n"); 1667 DAG->addEdge(SUa, SDep(Pred.getSUnit(), SDep::Artificial)); 1668 } 1669 } 1670 1671 SUnit2ClusterInfo[MemOpb.SU->NodeNum] = {ClusterLength, 1672 CurrentClusterBytes}; 1673 1674 LLVM_DEBUG(dbgs() << " Curr cluster length: " << ClusterLength 1675 << ", Curr cluster bytes: " << CurrentClusterBytes 1676 << "\n"); 1677 } 1678 } 1679 1680 void BaseMemOpClusterMutation::collectMemOpRecords( 1681 std::vector<SUnit> &SUnits, SmallVectorImpl<MemOpInfo> &MemOpRecords) { 1682 for (auto &SU : SUnits) { 1683 if ((IsLoad && !SU.getInstr()->mayLoad()) || 1684 (!IsLoad && !SU.getInstr()->mayStore())) 1685 continue; 1686 1687 const MachineInstr &MI = *SU.getInstr(); 1688 SmallVector<const MachineOperand *, 4> BaseOps; 1689 int64_t Offset; 1690 bool OffsetIsScalable; 1691 unsigned Width; 1692 if (TII->getMemOperandsWithOffsetWidth(MI, BaseOps, Offset, 1693 OffsetIsScalable, Width, TRI)) { 1694 MemOpRecords.push_back(MemOpInfo(&SU, BaseOps, Offset, Width)); 1695 1696 LLVM_DEBUG(dbgs() << "Num BaseOps: " << BaseOps.size() << ", Offset: " 1697 << Offset << ", OffsetIsScalable: " << OffsetIsScalable 1698 << ", Width: " << Width << "\n"); 1699 } 1700 #ifndef NDEBUG 1701 for (const auto *Op : BaseOps) 1702 assert(Op); 1703 #endif 1704 } 1705 } 1706 1707 bool BaseMemOpClusterMutation::groupMemOps( 1708 ArrayRef<MemOpInfo> MemOps, ScheduleDAGInstrs *DAG, 1709 DenseMap<unsigned, SmallVector<MemOpInfo, 32>> &Groups) { 1710 bool FastCluster = 1711 ForceFastCluster || 1712 MemOps.size() * DAG->SUnits.size() / 1000 > FastClusterThreshold; 1713 1714 for (const auto &MemOp : MemOps) { 1715 unsigned ChainPredID = DAG->SUnits.size(); 1716 if (FastCluster) { 1717 for (const SDep &Pred : MemOp.SU->Preds) { 1718 // We only want to cluster the mem ops that have the same ctrl(non-data) 1719 // pred so that they didn't have ctrl dependency for each other. But for 1720 // store instrs, we can still cluster them if the pred is load instr. 1721 if ((Pred.isCtrl() && 1722 (IsLoad || 1723 (Pred.getSUnit() && Pred.getSUnit()->getInstr()->mayStore()))) && 1724 !Pred.isArtificial()) { 1725 ChainPredID = Pred.getSUnit()->NodeNum; 1726 break; 1727 } 1728 } 1729 } else 1730 ChainPredID = 0; 1731 1732 Groups[ChainPredID].push_back(MemOp); 1733 } 1734 return FastCluster; 1735 } 1736 1737 /// Callback from DAG postProcessing to create cluster edges for loads/stores. 1738 void BaseMemOpClusterMutation::apply(ScheduleDAGInstrs *DAG) { 1739 // Collect all the clusterable loads/stores 1740 SmallVector<MemOpInfo, 32> MemOpRecords; 1741 collectMemOpRecords(DAG->SUnits, MemOpRecords); 1742 1743 if (MemOpRecords.size() < 2) 1744 return; 1745 1746 // Put the loads/stores without dependency into the same group with some 1747 // heuristic if the DAG is too complex to avoid compiling time blow up. 1748 // Notice that, some fusion pair could be lost with this. 1749 DenseMap<unsigned, SmallVector<MemOpInfo, 32>> Groups; 1750 bool FastCluster = groupMemOps(MemOpRecords, DAG, Groups); 1751 1752 for (auto &Group : Groups) { 1753 // Sorting the loads/stores, so that, we can stop the cluster as early as 1754 // possible. 1755 llvm::sort(Group.second); 1756 1757 // Trying to cluster all the neighboring loads/stores. 1758 clusterNeighboringMemOps(Group.second, FastCluster, DAG); 1759 } 1760 } 1761 1762 //===----------------------------------------------------------------------===// 1763 // CopyConstrain - DAG post-processing to encourage copy elimination. 1764 //===----------------------------------------------------------------------===// 1765 1766 namespace { 1767 1768 /// Post-process the DAG to create weak edges from all uses of a copy to 1769 /// the one use that defines the copy's source vreg, most likely an induction 1770 /// variable increment. 1771 class CopyConstrain : public ScheduleDAGMutation { 1772 // Transient state. 1773 SlotIndex RegionBeginIdx; 1774 1775 // RegionEndIdx is the slot index of the last non-debug instruction in the 1776 // scheduling region. So we may have RegionBeginIdx == RegionEndIdx. 1777 SlotIndex RegionEndIdx; 1778 1779 public: 1780 CopyConstrain(const TargetInstrInfo *, const TargetRegisterInfo *) {} 1781 1782 void apply(ScheduleDAGInstrs *DAGInstrs) override; 1783 1784 protected: 1785 void constrainLocalCopy(SUnit *CopySU, ScheduleDAGMILive *DAG); 1786 }; 1787 1788 } // end anonymous namespace 1789 1790 namespace llvm { 1791 1792 std::unique_ptr<ScheduleDAGMutation> 1793 createCopyConstrainDAGMutation(const TargetInstrInfo *TII, 1794 const TargetRegisterInfo *TRI) { 1795 return std::make_unique<CopyConstrain>(TII, TRI); 1796 } 1797 1798 } // end namespace llvm 1799 1800 /// constrainLocalCopy handles two possibilities: 1801 /// 1) Local src: 1802 /// I0: = dst 1803 /// I1: src = ... 1804 /// I2: = dst 1805 /// I3: dst = src (copy) 1806 /// (create pred->succ edges I0->I1, I2->I1) 1807 /// 1808 /// 2) Local copy: 1809 /// I0: dst = src (copy) 1810 /// I1: = dst 1811 /// I2: src = ... 1812 /// I3: = dst 1813 /// (create pred->succ edges I1->I2, I3->I2) 1814 /// 1815 /// Although the MachineScheduler is currently constrained to single blocks, 1816 /// this algorithm should handle extended blocks. An EBB is a set of 1817 /// contiguously numbered blocks such that the previous block in the EBB is 1818 /// always the single predecessor. 1819 void CopyConstrain::constrainLocalCopy(SUnit *CopySU, ScheduleDAGMILive *DAG) { 1820 LiveIntervals *LIS = DAG->getLIS(); 1821 MachineInstr *Copy = CopySU->getInstr(); 1822 1823 // Check for pure vreg copies. 1824 const MachineOperand &SrcOp = Copy->getOperand(1); 1825 Register SrcReg = SrcOp.getReg(); 1826 if (!Register::isVirtualRegister(SrcReg) || !SrcOp.readsReg()) 1827 return; 1828 1829 const MachineOperand &DstOp = Copy->getOperand(0); 1830 Register DstReg = DstOp.getReg(); 1831 if (!Register::isVirtualRegister(DstReg) || DstOp.isDead()) 1832 return; 1833 1834 // Check if either the dest or source is local. If it's live across a back 1835 // edge, it's not local. Note that if both vregs are live across the back 1836 // edge, we cannot successfully contrain the copy without cyclic scheduling. 1837 // If both the copy's source and dest are local live intervals, then we 1838 // should treat the dest as the global for the purpose of adding 1839 // constraints. This adds edges from source's other uses to the copy. 1840 unsigned LocalReg = SrcReg; 1841 unsigned GlobalReg = DstReg; 1842 LiveInterval *LocalLI = &LIS->getInterval(LocalReg); 1843 if (!LocalLI->isLocal(RegionBeginIdx, RegionEndIdx)) { 1844 LocalReg = DstReg; 1845 GlobalReg = SrcReg; 1846 LocalLI = &LIS->getInterval(LocalReg); 1847 if (!LocalLI->isLocal(RegionBeginIdx, RegionEndIdx)) 1848 return; 1849 } 1850 LiveInterval *GlobalLI = &LIS->getInterval(GlobalReg); 1851 1852 // Find the global segment after the start of the local LI. 1853 LiveInterval::iterator GlobalSegment = GlobalLI->find(LocalLI->beginIndex()); 1854 // If GlobalLI does not overlap LocalLI->start, then a copy directly feeds a 1855 // local live range. We could create edges from other global uses to the local 1856 // start, but the coalescer should have already eliminated these cases, so 1857 // don't bother dealing with it. 1858 if (GlobalSegment == GlobalLI->end()) 1859 return; 1860 1861 // If GlobalSegment is killed at the LocalLI->start, the call to find() 1862 // returned the next global segment. But if GlobalSegment overlaps with 1863 // LocalLI->start, then advance to the next segment. If a hole in GlobalLI 1864 // exists in LocalLI's vicinity, GlobalSegment will be the end of the hole. 1865 if (GlobalSegment->contains(LocalLI->beginIndex())) 1866 ++GlobalSegment; 1867 1868 if (GlobalSegment == GlobalLI->end()) 1869 return; 1870 1871 // Check if GlobalLI contains a hole in the vicinity of LocalLI. 1872 if (GlobalSegment != GlobalLI->begin()) { 1873 // Two address defs have no hole. 1874 if (SlotIndex::isSameInstr(std::prev(GlobalSegment)->end, 1875 GlobalSegment->start)) { 1876 return; 1877 } 1878 // If the prior global segment may be defined by the same two-address 1879 // instruction that also defines LocalLI, then can't make a hole here. 1880 if (SlotIndex::isSameInstr(std::prev(GlobalSegment)->start, 1881 LocalLI->beginIndex())) { 1882 return; 1883 } 1884 // If GlobalLI has a prior segment, it must be live into the EBB. Otherwise 1885 // it would be a disconnected component in the live range. 1886 assert(std::prev(GlobalSegment)->start < LocalLI->beginIndex() && 1887 "Disconnected LRG within the scheduling region."); 1888 } 1889 MachineInstr *GlobalDef = LIS->getInstructionFromIndex(GlobalSegment->start); 1890 if (!GlobalDef) 1891 return; 1892 1893 SUnit *GlobalSU = DAG->getSUnit(GlobalDef); 1894 if (!GlobalSU) 1895 return; 1896 1897 // GlobalDef is the bottom of the GlobalLI hole. Open the hole by 1898 // constraining the uses of the last local def to precede GlobalDef. 1899 SmallVector<SUnit*,8> LocalUses; 1900 const VNInfo *LastLocalVN = LocalLI->getVNInfoBefore(LocalLI->endIndex()); 1901 MachineInstr *LastLocalDef = LIS->getInstructionFromIndex(LastLocalVN->def); 1902 SUnit *LastLocalSU = DAG->getSUnit(LastLocalDef); 1903 for (const SDep &Succ : LastLocalSU->Succs) { 1904 if (Succ.getKind() != SDep::Data || Succ.getReg() != LocalReg) 1905 continue; 1906 if (Succ.getSUnit() == GlobalSU) 1907 continue; 1908 if (!DAG->canAddEdge(GlobalSU, Succ.getSUnit())) 1909 return; 1910 LocalUses.push_back(Succ.getSUnit()); 1911 } 1912 // Open the top of the GlobalLI hole by constraining any earlier global uses 1913 // to precede the start of LocalLI. 1914 SmallVector<SUnit*,8> GlobalUses; 1915 MachineInstr *FirstLocalDef = 1916 LIS->getInstructionFromIndex(LocalLI->beginIndex()); 1917 SUnit *FirstLocalSU = DAG->getSUnit(FirstLocalDef); 1918 for (const SDep &Pred : GlobalSU->Preds) { 1919 if (Pred.getKind() != SDep::Anti || Pred.getReg() != GlobalReg) 1920 continue; 1921 if (Pred.getSUnit() == FirstLocalSU) 1922 continue; 1923 if (!DAG->canAddEdge(FirstLocalSU, Pred.getSUnit())) 1924 return; 1925 GlobalUses.push_back(Pred.getSUnit()); 1926 } 1927 LLVM_DEBUG(dbgs() << "Constraining copy SU(" << CopySU->NodeNum << ")\n"); 1928 // Add the weak edges. 1929 for (SUnit *LU : LocalUses) { 1930 LLVM_DEBUG(dbgs() << " Local use SU(" << LU->NodeNum << ") -> SU(" 1931 << GlobalSU->NodeNum << ")\n"); 1932 DAG->addEdge(GlobalSU, SDep(LU, SDep::Weak)); 1933 } 1934 for (SUnit *GU : GlobalUses) { 1935 LLVM_DEBUG(dbgs() << " Global use SU(" << GU->NodeNum << ") -> SU(" 1936 << FirstLocalSU->NodeNum << ")\n"); 1937 DAG->addEdge(FirstLocalSU, SDep(GU, SDep::Weak)); 1938 } 1939 } 1940 1941 /// Callback from DAG postProcessing to create weak edges to encourage 1942 /// copy elimination. 1943 void CopyConstrain::apply(ScheduleDAGInstrs *DAGInstrs) { 1944 ScheduleDAGMI *DAG = static_cast<ScheduleDAGMI*>(DAGInstrs); 1945 assert(DAG->hasVRegLiveness() && "Expect VRegs with LiveIntervals"); 1946 1947 MachineBasicBlock::iterator FirstPos = nextIfDebug(DAG->begin(), DAG->end()); 1948 if (FirstPos == DAG->end()) 1949 return; 1950 RegionBeginIdx = DAG->getLIS()->getInstructionIndex(*FirstPos); 1951 RegionEndIdx = DAG->getLIS()->getInstructionIndex( 1952 *priorNonDebug(DAG->end(), DAG->begin())); 1953 1954 for (SUnit &SU : DAG->SUnits) { 1955 if (!SU.getInstr()->isCopy()) 1956 continue; 1957 1958 constrainLocalCopy(&SU, static_cast<ScheduleDAGMILive*>(DAG)); 1959 } 1960 } 1961 1962 //===----------------------------------------------------------------------===// 1963 // MachineSchedStrategy helpers used by GenericScheduler, GenericPostScheduler 1964 // and possibly other custom schedulers. 1965 //===----------------------------------------------------------------------===// 1966 1967 static const unsigned InvalidCycle = ~0U; 1968 1969 SchedBoundary::~SchedBoundary() { delete HazardRec; } 1970 1971 /// Given a Count of resource usage and a Latency value, return true if a 1972 /// SchedBoundary becomes resource limited. 1973 /// If we are checking after scheduling a node, we should return true when 1974 /// we just reach the resource limit. 1975 static bool checkResourceLimit(unsigned LFactor, unsigned Count, 1976 unsigned Latency, bool AfterSchedNode) { 1977 int ResCntFactor = (int)(Count - (Latency * LFactor)); 1978 if (AfterSchedNode) 1979 return ResCntFactor >= (int)LFactor; 1980 else 1981 return ResCntFactor > (int)LFactor; 1982 } 1983 1984 void SchedBoundary::reset() { 1985 // A new HazardRec is created for each DAG and owned by SchedBoundary. 1986 // Destroying and reconstructing it is very expensive though. So keep 1987 // invalid, placeholder HazardRecs. 1988 if (HazardRec && HazardRec->isEnabled()) { 1989 delete HazardRec; 1990 HazardRec = nullptr; 1991 } 1992 Available.clear(); 1993 Pending.clear(); 1994 CheckPending = false; 1995 CurrCycle = 0; 1996 CurrMOps = 0; 1997 MinReadyCycle = std::numeric_limits<unsigned>::max(); 1998 ExpectedLatency = 0; 1999 DependentLatency = 0; 2000 RetiredMOps = 0; 2001 MaxExecutedResCount = 0; 2002 ZoneCritResIdx = 0; 2003 IsResourceLimited = false; 2004 ReservedCycles.clear(); 2005 ReservedCyclesIndex.clear(); 2006 ResourceGroupSubUnitMasks.clear(); 2007 #if LLVM_ENABLE_ABI_BREAKING_CHECKS 2008 // Track the maximum number of stall cycles that could arise either from the 2009 // latency of a DAG edge or the number of cycles that a processor resource is 2010 // reserved (SchedBoundary::ReservedCycles). 2011 MaxObservedStall = 0; 2012 #endif 2013 // Reserve a zero-count for invalid CritResIdx. 2014 ExecutedResCounts.resize(1); 2015 assert(!ExecutedResCounts[0] && "nonzero count for bad resource"); 2016 } 2017 2018 void SchedRemainder:: 2019 init(ScheduleDAGMI *DAG, const TargetSchedModel *SchedModel) { 2020 reset(); 2021 if (!SchedModel->hasInstrSchedModel()) 2022 return; 2023 RemainingCounts.resize(SchedModel->getNumProcResourceKinds()); 2024 for (SUnit &SU : DAG->SUnits) { 2025 const MCSchedClassDesc *SC = DAG->getSchedClass(&SU); 2026 RemIssueCount += SchedModel->getNumMicroOps(SU.getInstr(), SC) 2027 * SchedModel->getMicroOpFactor(); 2028 for (TargetSchedModel::ProcResIter 2029 PI = SchedModel->getWriteProcResBegin(SC), 2030 PE = SchedModel->getWriteProcResEnd(SC); PI != PE; ++PI) { 2031 unsigned PIdx = PI->ProcResourceIdx; 2032 unsigned Factor = SchedModel->getResourceFactor(PIdx); 2033 RemainingCounts[PIdx] += (Factor * PI->Cycles); 2034 } 2035 } 2036 } 2037 2038 void SchedBoundary:: 2039 init(ScheduleDAGMI *dag, const TargetSchedModel *smodel, SchedRemainder *rem) { 2040 reset(); 2041 DAG = dag; 2042 SchedModel = smodel; 2043 Rem = rem; 2044 if (SchedModel->hasInstrSchedModel()) { 2045 unsigned ResourceCount = SchedModel->getNumProcResourceKinds(); 2046 ReservedCyclesIndex.resize(ResourceCount); 2047 ExecutedResCounts.resize(ResourceCount); 2048 ResourceGroupSubUnitMasks.resize(ResourceCount, APInt(ResourceCount, 0)); 2049 unsigned NumUnits = 0; 2050 2051 for (unsigned i = 0; i < ResourceCount; ++i) { 2052 ReservedCyclesIndex[i] = NumUnits; 2053 NumUnits += SchedModel->getProcResource(i)->NumUnits; 2054 if (isUnbufferedGroup(i)) { 2055 auto SubUnits = SchedModel->getProcResource(i)->SubUnitsIdxBegin; 2056 for (unsigned U = 0, UE = SchedModel->getProcResource(i)->NumUnits; 2057 U != UE; ++U) 2058 ResourceGroupSubUnitMasks[i].setBit(SubUnits[U]); 2059 } 2060 } 2061 2062 ReservedCycles.resize(NumUnits, InvalidCycle); 2063 } 2064 } 2065 2066 /// Compute the stall cycles based on this SUnit's ready time. Heuristics treat 2067 /// these "soft stalls" differently than the hard stall cycles based on CPU 2068 /// resources and computed by checkHazard(). A fully in-order model 2069 /// (MicroOpBufferSize==0) will not make use of this since instructions are not 2070 /// available for scheduling until they are ready. However, a weaker in-order 2071 /// model may use this for heuristics. For example, if a processor has in-order 2072 /// behavior when reading certain resources, this may come into play. 2073 unsigned SchedBoundary::getLatencyStallCycles(SUnit *SU) { 2074 if (!SU->isUnbuffered) 2075 return 0; 2076 2077 unsigned ReadyCycle = (isTop() ? SU->TopReadyCycle : SU->BotReadyCycle); 2078 if (ReadyCycle > CurrCycle) 2079 return ReadyCycle - CurrCycle; 2080 return 0; 2081 } 2082 2083 /// Compute the next cycle at which the given processor resource unit 2084 /// can be scheduled. 2085 unsigned SchedBoundary::getNextResourceCycleByInstance(unsigned InstanceIdx, 2086 unsigned Cycles) { 2087 unsigned NextUnreserved = ReservedCycles[InstanceIdx]; 2088 // If this resource has never been used, always return cycle zero. 2089 if (NextUnreserved == InvalidCycle) 2090 return 0; 2091 // For bottom-up scheduling add the cycles needed for the current operation. 2092 if (!isTop()) 2093 NextUnreserved += Cycles; 2094 return NextUnreserved; 2095 } 2096 2097 /// Compute the next cycle at which the given processor resource can be 2098 /// scheduled. Returns the next cycle and the index of the processor resource 2099 /// instance in the reserved cycles vector. 2100 std::pair<unsigned, unsigned> 2101 SchedBoundary::getNextResourceCycle(const MCSchedClassDesc *SC, unsigned PIdx, 2102 unsigned Cycles) { 2103 2104 unsigned MinNextUnreserved = InvalidCycle; 2105 unsigned InstanceIdx = 0; 2106 unsigned StartIndex = ReservedCyclesIndex[PIdx]; 2107 unsigned NumberOfInstances = SchedModel->getProcResource(PIdx)->NumUnits; 2108 assert(NumberOfInstances > 0 && 2109 "Cannot have zero instances of a ProcResource"); 2110 2111 if (isUnbufferedGroup(PIdx)) { 2112 // If any subunits are used by the instruction, report that the resource 2113 // group is available at 0, effectively removing the group record from 2114 // hazarding and basing the hazarding decisions on the subunit records. 2115 // Otherwise, choose the first available instance from among the subunits. 2116 // Specifications which assign cycles to both the subunits and the group or 2117 // which use an unbuffered group with buffered subunits will appear to 2118 // schedule strangely. In the first case, the additional cycles for the 2119 // group will be ignored. In the second, the group will be ignored 2120 // entirely. 2121 for (const MCWriteProcResEntry &PE : 2122 make_range(SchedModel->getWriteProcResBegin(SC), 2123 SchedModel->getWriteProcResEnd(SC))) 2124 if (ResourceGroupSubUnitMasks[PIdx][PE.ProcResourceIdx]) 2125 return std::make_pair(0u, StartIndex); 2126 2127 auto SubUnits = SchedModel->getProcResource(PIdx)->SubUnitsIdxBegin; 2128 for (unsigned I = 0, End = NumberOfInstances; I < End; ++I) { 2129 unsigned NextUnreserved, NextInstanceIdx; 2130 std::tie(NextUnreserved, NextInstanceIdx) = 2131 getNextResourceCycle(SC, SubUnits[I], Cycles); 2132 if (MinNextUnreserved > NextUnreserved) { 2133 InstanceIdx = NextInstanceIdx; 2134 MinNextUnreserved = NextUnreserved; 2135 } 2136 } 2137 return std::make_pair(MinNextUnreserved, InstanceIdx); 2138 } 2139 2140 for (unsigned I = StartIndex, End = StartIndex + NumberOfInstances; I < End; 2141 ++I) { 2142 unsigned NextUnreserved = getNextResourceCycleByInstance(I, Cycles); 2143 if (MinNextUnreserved > NextUnreserved) { 2144 InstanceIdx = I; 2145 MinNextUnreserved = NextUnreserved; 2146 } 2147 } 2148 return std::make_pair(MinNextUnreserved, InstanceIdx); 2149 } 2150 2151 /// Does this SU have a hazard within the current instruction group. 2152 /// 2153 /// The scheduler supports two modes of hazard recognition. The first is the 2154 /// ScheduleHazardRecognizer API. It is a fully general hazard recognizer that 2155 /// supports highly complicated in-order reservation tables 2156 /// (ScoreboardHazardRecognizer) and arbitrary target-specific logic. 2157 /// 2158 /// The second is a streamlined mechanism that checks for hazards based on 2159 /// simple counters that the scheduler itself maintains. It explicitly checks 2160 /// for instruction dispatch limitations, including the number of micro-ops that 2161 /// can dispatch per cycle. 2162 /// 2163 /// TODO: Also check whether the SU must start a new group. 2164 bool SchedBoundary::checkHazard(SUnit *SU) { 2165 if (HazardRec->isEnabled() 2166 && HazardRec->getHazardType(SU) != ScheduleHazardRecognizer::NoHazard) { 2167 return true; 2168 } 2169 2170 unsigned uops = SchedModel->getNumMicroOps(SU->getInstr()); 2171 if ((CurrMOps > 0) && (CurrMOps + uops > SchedModel->getIssueWidth())) { 2172 LLVM_DEBUG(dbgs() << " SU(" << SU->NodeNum << ") uops=" 2173 << SchedModel->getNumMicroOps(SU->getInstr()) << '\n'); 2174 return true; 2175 } 2176 2177 if (CurrMOps > 0 && 2178 ((isTop() && SchedModel->mustBeginGroup(SU->getInstr())) || 2179 (!isTop() && SchedModel->mustEndGroup(SU->getInstr())))) { 2180 LLVM_DEBUG(dbgs() << " hazard: SU(" << SU->NodeNum << ") must " 2181 << (isTop() ? "begin" : "end") << " group\n"); 2182 return true; 2183 } 2184 2185 if (SchedModel->hasInstrSchedModel() && SU->hasReservedResource) { 2186 const MCSchedClassDesc *SC = DAG->getSchedClass(SU); 2187 for (const MCWriteProcResEntry &PE : 2188 make_range(SchedModel->getWriteProcResBegin(SC), 2189 SchedModel->getWriteProcResEnd(SC))) { 2190 unsigned ResIdx = PE.ProcResourceIdx; 2191 unsigned Cycles = PE.Cycles; 2192 unsigned NRCycle, InstanceIdx; 2193 std::tie(NRCycle, InstanceIdx) = getNextResourceCycle(SC, ResIdx, Cycles); 2194 if (NRCycle > CurrCycle) { 2195 #if LLVM_ENABLE_ABI_BREAKING_CHECKS 2196 MaxObservedStall = std::max(Cycles, MaxObservedStall); 2197 #endif 2198 LLVM_DEBUG(dbgs() << " SU(" << SU->NodeNum << ") " 2199 << SchedModel->getResourceName(ResIdx) 2200 << '[' << InstanceIdx - ReservedCyclesIndex[ResIdx] << ']' 2201 << "=" << NRCycle << "c\n"); 2202 return true; 2203 } 2204 } 2205 } 2206 return false; 2207 } 2208 2209 // Find the unscheduled node in ReadySUs with the highest latency. 2210 unsigned SchedBoundary:: 2211 findMaxLatency(ArrayRef<SUnit*> ReadySUs) { 2212 SUnit *LateSU = nullptr; 2213 unsigned RemLatency = 0; 2214 for (SUnit *SU : ReadySUs) { 2215 unsigned L = getUnscheduledLatency(SU); 2216 if (L > RemLatency) { 2217 RemLatency = L; 2218 LateSU = SU; 2219 } 2220 } 2221 if (LateSU) { 2222 LLVM_DEBUG(dbgs() << Available.getName() << " RemLatency SU(" 2223 << LateSU->NodeNum << ") " << RemLatency << "c\n"); 2224 } 2225 return RemLatency; 2226 } 2227 2228 // Count resources in this zone and the remaining unscheduled 2229 // instruction. Return the max count, scaled. Set OtherCritIdx to the critical 2230 // resource index, or zero if the zone is issue limited. 2231 unsigned SchedBoundary:: 2232 getOtherResourceCount(unsigned &OtherCritIdx) { 2233 OtherCritIdx = 0; 2234 if (!SchedModel->hasInstrSchedModel()) 2235 return 0; 2236 2237 unsigned OtherCritCount = Rem->RemIssueCount 2238 + (RetiredMOps * SchedModel->getMicroOpFactor()); 2239 LLVM_DEBUG(dbgs() << " " << Available.getName() << " + Remain MOps: " 2240 << OtherCritCount / SchedModel->getMicroOpFactor() << '\n'); 2241 for (unsigned PIdx = 1, PEnd = SchedModel->getNumProcResourceKinds(); 2242 PIdx != PEnd; ++PIdx) { 2243 unsigned OtherCount = getResourceCount(PIdx) + Rem->RemainingCounts[PIdx]; 2244 if (OtherCount > OtherCritCount) { 2245 OtherCritCount = OtherCount; 2246 OtherCritIdx = PIdx; 2247 } 2248 } 2249 if (OtherCritIdx) { 2250 LLVM_DEBUG( 2251 dbgs() << " " << Available.getName() << " + Remain CritRes: " 2252 << OtherCritCount / SchedModel->getResourceFactor(OtherCritIdx) 2253 << " " << SchedModel->getResourceName(OtherCritIdx) << "\n"); 2254 } 2255 return OtherCritCount; 2256 } 2257 2258 void SchedBoundary::releaseNode(SUnit *SU, unsigned ReadyCycle, bool InPQueue, 2259 unsigned Idx) { 2260 assert(SU->getInstr() && "Scheduled SUnit must have instr"); 2261 2262 #if LLVM_ENABLE_ABI_BREAKING_CHECKS 2263 // ReadyCycle was been bumped up to the CurrCycle when this node was 2264 // scheduled, but CurrCycle may have been eagerly advanced immediately after 2265 // scheduling, so may now be greater than ReadyCycle. 2266 if (ReadyCycle > CurrCycle) 2267 MaxObservedStall = std::max(ReadyCycle - CurrCycle, MaxObservedStall); 2268 #endif 2269 2270 if (ReadyCycle < MinReadyCycle) 2271 MinReadyCycle = ReadyCycle; 2272 2273 // Check for interlocks first. For the purpose of other heuristics, an 2274 // instruction that cannot issue appears as if it's not in the ReadyQueue. 2275 bool IsBuffered = SchedModel->getMicroOpBufferSize() != 0; 2276 bool HazardDetected = (!IsBuffered && ReadyCycle > CurrCycle) || 2277 checkHazard(SU) || (Available.size() >= ReadyListLimit); 2278 2279 if (!HazardDetected) { 2280 Available.push(SU); 2281 2282 if (InPQueue) 2283 Pending.remove(Pending.begin() + Idx); 2284 return; 2285 } 2286 2287 if (!InPQueue) 2288 Pending.push(SU); 2289 } 2290 2291 /// Move the boundary of scheduled code by one cycle. 2292 void SchedBoundary::bumpCycle(unsigned NextCycle) { 2293 if (SchedModel->getMicroOpBufferSize() == 0) { 2294 assert(MinReadyCycle < std::numeric_limits<unsigned>::max() && 2295 "MinReadyCycle uninitialized"); 2296 if (MinReadyCycle > NextCycle) 2297 NextCycle = MinReadyCycle; 2298 } 2299 // Update the current micro-ops, which will issue in the next cycle. 2300 unsigned DecMOps = SchedModel->getIssueWidth() * (NextCycle - CurrCycle); 2301 CurrMOps = (CurrMOps <= DecMOps) ? 0 : CurrMOps - DecMOps; 2302 2303 // Decrement DependentLatency based on the next cycle. 2304 if ((NextCycle - CurrCycle) > DependentLatency) 2305 DependentLatency = 0; 2306 else 2307 DependentLatency -= (NextCycle - CurrCycle); 2308 2309 if (!HazardRec->isEnabled()) { 2310 // Bypass HazardRec virtual calls. 2311 CurrCycle = NextCycle; 2312 } else { 2313 // Bypass getHazardType calls in case of long latency. 2314 for (; CurrCycle != NextCycle; ++CurrCycle) { 2315 if (isTop()) 2316 HazardRec->AdvanceCycle(); 2317 else 2318 HazardRec->RecedeCycle(); 2319 } 2320 } 2321 CheckPending = true; 2322 IsResourceLimited = 2323 checkResourceLimit(SchedModel->getLatencyFactor(), getCriticalCount(), 2324 getScheduledLatency(), true); 2325 2326 LLVM_DEBUG(dbgs() << "Cycle: " << CurrCycle << ' ' << Available.getName() 2327 << '\n'); 2328 } 2329 2330 void SchedBoundary::incExecutedResources(unsigned PIdx, unsigned Count) { 2331 ExecutedResCounts[PIdx] += Count; 2332 if (ExecutedResCounts[PIdx] > MaxExecutedResCount) 2333 MaxExecutedResCount = ExecutedResCounts[PIdx]; 2334 } 2335 2336 /// Add the given processor resource to this scheduled zone. 2337 /// 2338 /// \param Cycles indicates the number of consecutive (non-pipelined) cycles 2339 /// during which this resource is consumed. 2340 /// 2341 /// \return the next cycle at which the instruction may execute without 2342 /// oversubscribing resources. 2343 unsigned SchedBoundary::countResource(const MCSchedClassDesc *SC, unsigned PIdx, 2344 unsigned Cycles, unsigned NextCycle) { 2345 unsigned Factor = SchedModel->getResourceFactor(PIdx); 2346 unsigned Count = Factor * Cycles; 2347 LLVM_DEBUG(dbgs() << " " << SchedModel->getResourceName(PIdx) << " +" 2348 << Cycles << "x" << Factor << "u\n"); 2349 2350 // Update Executed resources counts. 2351 incExecutedResources(PIdx, Count); 2352 assert(Rem->RemainingCounts[PIdx] >= Count && "resource double counted"); 2353 Rem->RemainingCounts[PIdx] -= Count; 2354 2355 // Check if this resource exceeds the current critical resource. If so, it 2356 // becomes the critical resource. 2357 if (ZoneCritResIdx != PIdx && (getResourceCount(PIdx) > getCriticalCount())) { 2358 ZoneCritResIdx = PIdx; 2359 LLVM_DEBUG(dbgs() << " *** Critical resource " 2360 << SchedModel->getResourceName(PIdx) << ": " 2361 << getResourceCount(PIdx) / SchedModel->getLatencyFactor() 2362 << "c\n"); 2363 } 2364 // For reserved resources, record the highest cycle using the resource. 2365 unsigned NextAvailable, InstanceIdx; 2366 std::tie(NextAvailable, InstanceIdx) = getNextResourceCycle(SC, PIdx, Cycles); 2367 if (NextAvailable > CurrCycle) { 2368 LLVM_DEBUG(dbgs() << " Resource conflict: " 2369 << SchedModel->getResourceName(PIdx) 2370 << '[' << InstanceIdx - ReservedCyclesIndex[PIdx] << ']' 2371 << " reserved until @" << NextAvailable << "\n"); 2372 } 2373 return NextAvailable; 2374 } 2375 2376 /// Move the boundary of scheduled code by one SUnit. 2377 void SchedBoundary::bumpNode(SUnit *SU) { 2378 // Update the reservation table. 2379 if (HazardRec->isEnabled()) { 2380 if (!isTop() && SU->isCall) { 2381 // Calls are scheduled with their preceding instructions. For bottom-up 2382 // scheduling, clear the pipeline state before emitting. 2383 HazardRec->Reset(); 2384 } 2385 HazardRec->EmitInstruction(SU); 2386 // Scheduling an instruction may have made pending instructions available. 2387 CheckPending = true; 2388 } 2389 // checkHazard should prevent scheduling multiple instructions per cycle that 2390 // exceed the issue width. 2391 const MCSchedClassDesc *SC = DAG->getSchedClass(SU); 2392 unsigned IncMOps = SchedModel->getNumMicroOps(SU->getInstr()); 2393 assert( 2394 (CurrMOps == 0 || (CurrMOps + IncMOps) <= SchedModel->getIssueWidth()) && 2395 "Cannot schedule this instruction's MicroOps in the current cycle."); 2396 2397 unsigned ReadyCycle = (isTop() ? SU->TopReadyCycle : SU->BotReadyCycle); 2398 LLVM_DEBUG(dbgs() << " Ready @" << ReadyCycle << "c\n"); 2399 2400 unsigned NextCycle = CurrCycle; 2401 switch (SchedModel->getMicroOpBufferSize()) { 2402 case 0: 2403 assert(ReadyCycle <= CurrCycle && "Broken PendingQueue"); 2404 break; 2405 case 1: 2406 if (ReadyCycle > NextCycle) { 2407 NextCycle = ReadyCycle; 2408 LLVM_DEBUG(dbgs() << " *** Stall until: " << ReadyCycle << "\n"); 2409 } 2410 break; 2411 default: 2412 // We don't currently model the OOO reorder buffer, so consider all 2413 // scheduled MOps to be "retired". We do loosely model in-order resource 2414 // latency. If this instruction uses an in-order resource, account for any 2415 // likely stall cycles. 2416 if (SU->isUnbuffered && ReadyCycle > NextCycle) 2417 NextCycle = ReadyCycle; 2418 break; 2419 } 2420 RetiredMOps += IncMOps; 2421 2422 // Update resource counts and critical resource. 2423 if (SchedModel->hasInstrSchedModel()) { 2424 unsigned DecRemIssue = IncMOps * SchedModel->getMicroOpFactor(); 2425 assert(Rem->RemIssueCount >= DecRemIssue && "MOps double counted"); 2426 Rem->RemIssueCount -= DecRemIssue; 2427 if (ZoneCritResIdx) { 2428 // Scale scheduled micro-ops for comparing with the critical resource. 2429 unsigned ScaledMOps = 2430 RetiredMOps * SchedModel->getMicroOpFactor(); 2431 2432 // If scaled micro-ops are now more than the previous critical resource by 2433 // a full cycle, then micro-ops issue becomes critical. 2434 if ((int)(ScaledMOps - getResourceCount(ZoneCritResIdx)) 2435 >= (int)SchedModel->getLatencyFactor()) { 2436 ZoneCritResIdx = 0; 2437 LLVM_DEBUG(dbgs() << " *** Critical resource NumMicroOps: " 2438 << ScaledMOps / SchedModel->getLatencyFactor() 2439 << "c\n"); 2440 } 2441 } 2442 for (TargetSchedModel::ProcResIter 2443 PI = SchedModel->getWriteProcResBegin(SC), 2444 PE = SchedModel->getWriteProcResEnd(SC); PI != PE; ++PI) { 2445 unsigned RCycle = 2446 countResource(SC, PI->ProcResourceIdx, PI->Cycles, NextCycle); 2447 if (RCycle > NextCycle) 2448 NextCycle = RCycle; 2449 } 2450 if (SU->hasReservedResource) { 2451 // For reserved resources, record the highest cycle using the resource. 2452 // For top-down scheduling, this is the cycle in which we schedule this 2453 // instruction plus the number of cycles the operations reserves the 2454 // resource. For bottom-up is it simply the instruction's cycle. 2455 for (TargetSchedModel::ProcResIter 2456 PI = SchedModel->getWriteProcResBegin(SC), 2457 PE = SchedModel->getWriteProcResEnd(SC); PI != PE; ++PI) { 2458 unsigned PIdx = PI->ProcResourceIdx; 2459 if (SchedModel->getProcResource(PIdx)->BufferSize == 0) { 2460 unsigned ReservedUntil, InstanceIdx; 2461 std::tie(ReservedUntil, InstanceIdx) = 2462 getNextResourceCycle(SC, PIdx, 0); 2463 if (isTop()) { 2464 ReservedCycles[InstanceIdx] = 2465 std::max(ReservedUntil, NextCycle + PI->Cycles); 2466 } else 2467 ReservedCycles[InstanceIdx] = NextCycle; 2468 } 2469 } 2470 } 2471 } 2472 // Update ExpectedLatency and DependentLatency. 2473 unsigned &TopLatency = isTop() ? ExpectedLatency : DependentLatency; 2474 unsigned &BotLatency = isTop() ? DependentLatency : ExpectedLatency; 2475 if (SU->getDepth() > TopLatency) { 2476 TopLatency = SU->getDepth(); 2477 LLVM_DEBUG(dbgs() << " " << Available.getName() << " TopLatency SU(" 2478 << SU->NodeNum << ") " << TopLatency << "c\n"); 2479 } 2480 if (SU->getHeight() > BotLatency) { 2481 BotLatency = SU->getHeight(); 2482 LLVM_DEBUG(dbgs() << " " << Available.getName() << " BotLatency SU(" 2483 << SU->NodeNum << ") " << BotLatency << "c\n"); 2484 } 2485 // If we stall for any reason, bump the cycle. 2486 if (NextCycle > CurrCycle) 2487 bumpCycle(NextCycle); 2488 else 2489 // After updating ZoneCritResIdx and ExpectedLatency, check if we're 2490 // resource limited. If a stall occurred, bumpCycle does this. 2491 IsResourceLimited = 2492 checkResourceLimit(SchedModel->getLatencyFactor(), getCriticalCount(), 2493 getScheduledLatency(), true); 2494 2495 // Update CurrMOps after calling bumpCycle to handle stalls, since bumpCycle 2496 // resets CurrMOps. Loop to handle instructions with more MOps than issue in 2497 // one cycle. Since we commonly reach the max MOps here, opportunistically 2498 // bump the cycle to avoid uselessly checking everything in the readyQ. 2499 CurrMOps += IncMOps; 2500 2501 // Bump the cycle count for issue group constraints. 2502 // This must be done after NextCycle has been adjust for all other stalls. 2503 // Calling bumpCycle(X) will reduce CurrMOps by one issue group and set 2504 // currCycle to X. 2505 if ((isTop() && SchedModel->mustEndGroup(SU->getInstr())) || 2506 (!isTop() && SchedModel->mustBeginGroup(SU->getInstr()))) { 2507 LLVM_DEBUG(dbgs() << " Bump cycle to " << (isTop() ? "end" : "begin") 2508 << " group\n"); 2509 bumpCycle(++NextCycle); 2510 } 2511 2512 while (CurrMOps >= SchedModel->getIssueWidth()) { 2513 LLVM_DEBUG(dbgs() << " *** Max MOps " << CurrMOps << " at cycle " 2514 << CurrCycle << '\n'); 2515 bumpCycle(++NextCycle); 2516 } 2517 LLVM_DEBUG(dumpScheduledState()); 2518 } 2519 2520 /// Release pending ready nodes in to the available queue. This makes them 2521 /// visible to heuristics. 2522 void SchedBoundary::releasePending() { 2523 // If the available queue is empty, it is safe to reset MinReadyCycle. 2524 if (Available.empty()) 2525 MinReadyCycle = std::numeric_limits<unsigned>::max(); 2526 2527 // Check to see if any of the pending instructions are ready to issue. If 2528 // so, add them to the available queue. 2529 for (unsigned I = 0, E = Pending.size(); I < E; ++I) { 2530 SUnit *SU = *(Pending.begin() + I); 2531 unsigned ReadyCycle = isTop() ? SU->TopReadyCycle : SU->BotReadyCycle; 2532 2533 if (ReadyCycle < MinReadyCycle) 2534 MinReadyCycle = ReadyCycle; 2535 2536 if (Available.size() >= ReadyListLimit) 2537 break; 2538 2539 releaseNode(SU, ReadyCycle, true, I); 2540 if (E != Pending.size()) { 2541 --I; 2542 --E; 2543 } 2544 } 2545 CheckPending = false; 2546 } 2547 2548 /// Remove SU from the ready set for this boundary. 2549 void SchedBoundary::removeReady(SUnit *SU) { 2550 if (Available.isInQueue(SU)) 2551 Available.remove(Available.find(SU)); 2552 else { 2553 assert(Pending.isInQueue(SU) && "bad ready count"); 2554 Pending.remove(Pending.find(SU)); 2555 } 2556 } 2557 2558 /// If this queue only has one ready candidate, return it. As a side effect, 2559 /// defer any nodes that now hit a hazard, and advance the cycle until at least 2560 /// one node is ready. If multiple instructions are ready, return NULL. 2561 SUnit *SchedBoundary::pickOnlyChoice() { 2562 if (CheckPending) 2563 releasePending(); 2564 2565 // Defer any ready instrs that now have a hazard. 2566 for (ReadyQueue::iterator I = Available.begin(); I != Available.end();) { 2567 if (checkHazard(*I)) { 2568 Pending.push(*I); 2569 I = Available.remove(I); 2570 continue; 2571 } 2572 ++I; 2573 } 2574 for (unsigned i = 0; Available.empty(); ++i) { 2575 // FIXME: Re-enable assert once PR20057 is resolved. 2576 // assert(i <= (HazardRec->getMaxLookAhead() + MaxObservedStall) && 2577 // "permanent hazard"); 2578 (void)i; 2579 bumpCycle(CurrCycle + 1); 2580 releasePending(); 2581 } 2582 2583 LLVM_DEBUG(Pending.dump()); 2584 LLVM_DEBUG(Available.dump()); 2585 2586 if (Available.size() == 1) 2587 return *Available.begin(); 2588 return nullptr; 2589 } 2590 2591 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) 2592 // This is useful information to dump after bumpNode. 2593 // Note that the Queue contents are more useful before pickNodeFromQueue. 2594 LLVM_DUMP_METHOD void SchedBoundary::dumpScheduledState() const { 2595 unsigned ResFactor; 2596 unsigned ResCount; 2597 if (ZoneCritResIdx) { 2598 ResFactor = SchedModel->getResourceFactor(ZoneCritResIdx); 2599 ResCount = getResourceCount(ZoneCritResIdx); 2600 } else { 2601 ResFactor = SchedModel->getMicroOpFactor(); 2602 ResCount = RetiredMOps * ResFactor; 2603 } 2604 unsigned LFactor = SchedModel->getLatencyFactor(); 2605 dbgs() << Available.getName() << " @" << CurrCycle << "c\n" 2606 << " Retired: " << RetiredMOps; 2607 dbgs() << "\n Executed: " << getExecutedCount() / LFactor << "c"; 2608 dbgs() << "\n Critical: " << ResCount / LFactor << "c, " 2609 << ResCount / ResFactor << " " 2610 << SchedModel->getResourceName(ZoneCritResIdx) 2611 << "\n ExpectedLatency: " << ExpectedLatency << "c\n" 2612 << (IsResourceLimited ? " - Resource" : " - Latency") 2613 << " limited.\n"; 2614 } 2615 #endif 2616 2617 //===----------------------------------------------------------------------===// 2618 // GenericScheduler - Generic implementation of MachineSchedStrategy. 2619 //===----------------------------------------------------------------------===// 2620 2621 void GenericSchedulerBase::SchedCandidate:: 2622 initResourceDelta(const ScheduleDAGMI *DAG, 2623 const TargetSchedModel *SchedModel) { 2624 if (!Policy.ReduceResIdx && !Policy.DemandResIdx) 2625 return; 2626 2627 const MCSchedClassDesc *SC = DAG->getSchedClass(SU); 2628 for (TargetSchedModel::ProcResIter 2629 PI = SchedModel->getWriteProcResBegin(SC), 2630 PE = SchedModel->getWriteProcResEnd(SC); PI != PE; ++PI) { 2631 if (PI->ProcResourceIdx == Policy.ReduceResIdx) 2632 ResDelta.CritResources += PI->Cycles; 2633 if (PI->ProcResourceIdx == Policy.DemandResIdx) 2634 ResDelta.DemandedResources += PI->Cycles; 2635 } 2636 } 2637 2638 /// Compute remaining latency. We need this both to determine whether the 2639 /// overall schedule has become latency-limited and whether the instructions 2640 /// outside this zone are resource or latency limited. 2641 /// 2642 /// The "dependent" latency is updated incrementally during scheduling as the 2643 /// max height/depth of scheduled nodes minus the cycles since it was 2644 /// scheduled: 2645 /// DLat = max (N.depth - (CurrCycle - N.ReadyCycle) for N in Zone 2646 /// 2647 /// The "independent" latency is the max ready queue depth: 2648 /// ILat = max N.depth for N in Available|Pending 2649 /// 2650 /// RemainingLatency is the greater of independent and dependent latency. 2651 /// 2652 /// These computations are expensive, especially in DAGs with many edges, so 2653 /// only do them if necessary. 2654 static unsigned computeRemLatency(SchedBoundary &CurrZone) { 2655 unsigned RemLatency = CurrZone.getDependentLatency(); 2656 RemLatency = std::max(RemLatency, 2657 CurrZone.findMaxLatency(CurrZone.Available.elements())); 2658 RemLatency = std::max(RemLatency, 2659 CurrZone.findMaxLatency(CurrZone.Pending.elements())); 2660 return RemLatency; 2661 } 2662 2663 /// Returns true if the current cycle plus remaning latency is greater than 2664 /// the critical path in the scheduling region. 2665 bool GenericSchedulerBase::shouldReduceLatency(const CandPolicy &Policy, 2666 SchedBoundary &CurrZone, 2667 bool ComputeRemLatency, 2668 unsigned &RemLatency) const { 2669 // The current cycle is already greater than the critical path, so we are 2670 // already latency limited and don't need to compute the remaining latency. 2671 if (CurrZone.getCurrCycle() > Rem.CriticalPath) 2672 return true; 2673 2674 // If we haven't scheduled anything yet, then we aren't latency limited. 2675 if (CurrZone.getCurrCycle() == 0) 2676 return false; 2677 2678 if (ComputeRemLatency) 2679 RemLatency = computeRemLatency(CurrZone); 2680 2681 return RemLatency + CurrZone.getCurrCycle() > Rem.CriticalPath; 2682 } 2683 2684 /// Set the CandPolicy given a scheduling zone given the current resources and 2685 /// latencies inside and outside the zone. 2686 void GenericSchedulerBase::setPolicy(CandPolicy &Policy, bool IsPostRA, 2687 SchedBoundary &CurrZone, 2688 SchedBoundary *OtherZone) { 2689 // Apply preemptive heuristics based on the total latency and resources 2690 // inside and outside this zone. Potential stalls should be considered before 2691 // following this policy. 2692 2693 // Compute the critical resource outside the zone. 2694 unsigned OtherCritIdx = 0; 2695 unsigned OtherCount = 2696 OtherZone ? OtherZone->getOtherResourceCount(OtherCritIdx) : 0; 2697 2698 bool OtherResLimited = false; 2699 unsigned RemLatency = 0; 2700 bool RemLatencyComputed = false; 2701 if (SchedModel->hasInstrSchedModel() && OtherCount != 0) { 2702 RemLatency = computeRemLatency(CurrZone); 2703 RemLatencyComputed = true; 2704 OtherResLimited = checkResourceLimit(SchedModel->getLatencyFactor(), 2705 OtherCount, RemLatency, false); 2706 } 2707 2708 // Schedule aggressively for latency in PostRA mode. We don't check for 2709 // acyclic latency during PostRA, and highly out-of-order processors will 2710 // skip PostRA scheduling. 2711 if (!OtherResLimited && 2712 (IsPostRA || shouldReduceLatency(Policy, CurrZone, !RemLatencyComputed, 2713 RemLatency))) { 2714 Policy.ReduceLatency |= true; 2715 LLVM_DEBUG(dbgs() << " " << CurrZone.Available.getName() 2716 << " RemainingLatency " << RemLatency << " + " 2717 << CurrZone.getCurrCycle() << "c > CritPath " 2718 << Rem.CriticalPath << "\n"); 2719 } 2720 // If the same resource is limiting inside and outside the zone, do nothing. 2721 if (CurrZone.getZoneCritResIdx() == OtherCritIdx) 2722 return; 2723 2724 LLVM_DEBUG(if (CurrZone.isResourceLimited()) { 2725 dbgs() << " " << CurrZone.Available.getName() << " ResourceLimited: " 2726 << SchedModel->getResourceName(CurrZone.getZoneCritResIdx()) << "\n"; 2727 } if (OtherResLimited) dbgs() 2728 << " RemainingLimit: " 2729 << SchedModel->getResourceName(OtherCritIdx) << "\n"; 2730 if (!CurrZone.isResourceLimited() && !OtherResLimited) dbgs() 2731 << " Latency limited both directions.\n"); 2732 2733 if (CurrZone.isResourceLimited() && !Policy.ReduceResIdx) 2734 Policy.ReduceResIdx = CurrZone.getZoneCritResIdx(); 2735 2736 if (OtherResLimited) 2737 Policy.DemandResIdx = OtherCritIdx; 2738 } 2739 2740 #ifndef NDEBUG 2741 const char *GenericSchedulerBase::getReasonStr( 2742 GenericSchedulerBase::CandReason Reason) { 2743 switch (Reason) { 2744 case NoCand: return "NOCAND "; 2745 case Only1: return "ONLY1 "; 2746 case PhysReg: return "PHYS-REG "; 2747 case RegExcess: return "REG-EXCESS"; 2748 case RegCritical: return "REG-CRIT "; 2749 case Stall: return "STALL "; 2750 case Cluster: return "CLUSTER "; 2751 case Weak: return "WEAK "; 2752 case RegMax: return "REG-MAX "; 2753 case ResourceReduce: return "RES-REDUCE"; 2754 case ResourceDemand: return "RES-DEMAND"; 2755 case TopDepthReduce: return "TOP-DEPTH "; 2756 case TopPathReduce: return "TOP-PATH "; 2757 case BotHeightReduce:return "BOT-HEIGHT"; 2758 case BotPathReduce: return "BOT-PATH "; 2759 case NextDefUse: return "DEF-USE "; 2760 case NodeOrder: return "ORDER "; 2761 }; 2762 llvm_unreachable("Unknown reason!"); 2763 } 2764 2765 void GenericSchedulerBase::traceCandidate(const SchedCandidate &Cand) { 2766 PressureChange P; 2767 unsigned ResIdx = 0; 2768 unsigned Latency = 0; 2769 switch (Cand.Reason) { 2770 default: 2771 break; 2772 case RegExcess: 2773 P = Cand.RPDelta.Excess; 2774 break; 2775 case RegCritical: 2776 P = Cand.RPDelta.CriticalMax; 2777 break; 2778 case RegMax: 2779 P = Cand.RPDelta.CurrentMax; 2780 break; 2781 case ResourceReduce: 2782 ResIdx = Cand.Policy.ReduceResIdx; 2783 break; 2784 case ResourceDemand: 2785 ResIdx = Cand.Policy.DemandResIdx; 2786 break; 2787 case TopDepthReduce: 2788 Latency = Cand.SU->getDepth(); 2789 break; 2790 case TopPathReduce: 2791 Latency = Cand.SU->getHeight(); 2792 break; 2793 case BotHeightReduce: 2794 Latency = Cand.SU->getHeight(); 2795 break; 2796 case BotPathReduce: 2797 Latency = Cand.SU->getDepth(); 2798 break; 2799 } 2800 dbgs() << " Cand SU(" << Cand.SU->NodeNum << ") " << getReasonStr(Cand.Reason); 2801 if (P.isValid()) 2802 dbgs() << " " << TRI->getRegPressureSetName(P.getPSet()) 2803 << ":" << P.getUnitInc() << " "; 2804 else 2805 dbgs() << " "; 2806 if (ResIdx) 2807 dbgs() << " " << SchedModel->getProcResource(ResIdx)->Name << " "; 2808 else 2809 dbgs() << " "; 2810 if (Latency) 2811 dbgs() << " " << Latency << " cycles "; 2812 else 2813 dbgs() << " "; 2814 dbgs() << '\n'; 2815 } 2816 #endif 2817 2818 namespace llvm { 2819 /// Return true if this heuristic determines order. 2820 /// TODO: Consider refactor return type of these functions as integer or enum, 2821 /// as we may need to differentiate whether TryCand is better than Cand. 2822 bool tryLess(int TryVal, int CandVal, 2823 GenericSchedulerBase::SchedCandidate &TryCand, 2824 GenericSchedulerBase::SchedCandidate &Cand, 2825 GenericSchedulerBase::CandReason Reason) { 2826 if (TryVal < CandVal) { 2827 TryCand.Reason = Reason; 2828 return true; 2829 } 2830 if (TryVal > CandVal) { 2831 if (Cand.Reason > Reason) 2832 Cand.Reason = Reason; 2833 return true; 2834 } 2835 return false; 2836 } 2837 2838 bool tryGreater(int TryVal, int CandVal, 2839 GenericSchedulerBase::SchedCandidate &TryCand, 2840 GenericSchedulerBase::SchedCandidate &Cand, 2841 GenericSchedulerBase::CandReason Reason) { 2842 if (TryVal > CandVal) { 2843 TryCand.Reason = Reason; 2844 return true; 2845 } 2846 if (TryVal < CandVal) { 2847 if (Cand.Reason > Reason) 2848 Cand.Reason = Reason; 2849 return true; 2850 } 2851 return false; 2852 } 2853 2854 bool tryLatency(GenericSchedulerBase::SchedCandidate &TryCand, 2855 GenericSchedulerBase::SchedCandidate &Cand, 2856 SchedBoundary &Zone) { 2857 if (Zone.isTop()) { 2858 // Prefer the candidate with the lesser depth, but only if one of them has 2859 // depth greater than the total latency scheduled so far, otherwise either 2860 // of them could be scheduled now with no stall. 2861 if (std::max(TryCand.SU->getDepth(), Cand.SU->getDepth()) > 2862 Zone.getScheduledLatency()) { 2863 if (tryLess(TryCand.SU->getDepth(), Cand.SU->getDepth(), 2864 TryCand, Cand, GenericSchedulerBase::TopDepthReduce)) 2865 return true; 2866 } 2867 if (tryGreater(TryCand.SU->getHeight(), Cand.SU->getHeight(), 2868 TryCand, Cand, GenericSchedulerBase::TopPathReduce)) 2869 return true; 2870 } else { 2871 // Prefer the candidate with the lesser height, but only if one of them has 2872 // height greater than the total latency scheduled so far, otherwise either 2873 // of them could be scheduled now with no stall. 2874 if (std::max(TryCand.SU->getHeight(), Cand.SU->getHeight()) > 2875 Zone.getScheduledLatency()) { 2876 if (tryLess(TryCand.SU->getHeight(), Cand.SU->getHeight(), 2877 TryCand, Cand, GenericSchedulerBase::BotHeightReduce)) 2878 return true; 2879 } 2880 if (tryGreater(TryCand.SU->getDepth(), Cand.SU->getDepth(), 2881 TryCand, Cand, GenericSchedulerBase::BotPathReduce)) 2882 return true; 2883 } 2884 return false; 2885 } 2886 } // end namespace llvm 2887 2888 static void tracePick(GenericSchedulerBase::CandReason Reason, bool IsTop) { 2889 LLVM_DEBUG(dbgs() << "Pick " << (IsTop ? "Top " : "Bot ") 2890 << GenericSchedulerBase::getReasonStr(Reason) << '\n'); 2891 } 2892 2893 static void tracePick(const GenericSchedulerBase::SchedCandidate &Cand) { 2894 tracePick(Cand.Reason, Cand.AtTop); 2895 } 2896 2897 void GenericScheduler::initialize(ScheduleDAGMI *dag) { 2898 assert(dag->hasVRegLiveness() && 2899 "(PreRA)GenericScheduler needs vreg liveness"); 2900 DAG = static_cast<ScheduleDAGMILive*>(dag); 2901 SchedModel = DAG->getSchedModel(); 2902 TRI = DAG->TRI; 2903 2904 if (RegionPolicy.ComputeDFSResult) 2905 DAG->computeDFSResult(); 2906 2907 Rem.init(DAG, SchedModel); 2908 Top.init(DAG, SchedModel, &Rem); 2909 Bot.init(DAG, SchedModel, &Rem); 2910 2911 // Initialize resource counts. 2912 2913 // Initialize the HazardRecognizers. If itineraries don't exist, are empty, or 2914 // are disabled, then these HazardRecs will be disabled. 2915 const InstrItineraryData *Itin = SchedModel->getInstrItineraries(); 2916 if (!Top.HazardRec) { 2917 Top.HazardRec = 2918 DAG->MF.getSubtarget().getInstrInfo()->CreateTargetMIHazardRecognizer( 2919 Itin, DAG); 2920 } 2921 if (!Bot.HazardRec) { 2922 Bot.HazardRec = 2923 DAG->MF.getSubtarget().getInstrInfo()->CreateTargetMIHazardRecognizer( 2924 Itin, DAG); 2925 } 2926 TopCand.SU = nullptr; 2927 BotCand.SU = nullptr; 2928 } 2929 2930 /// Initialize the per-region scheduling policy. 2931 void GenericScheduler::initPolicy(MachineBasicBlock::iterator Begin, 2932 MachineBasicBlock::iterator End, 2933 unsigned NumRegionInstrs) { 2934 const MachineFunction &MF = *Begin->getMF(); 2935 const TargetLowering *TLI = MF.getSubtarget().getTargetLowering(); 2936 2937 // Avoid setting up the register pressure tracker for small regions to save 2938 // compile time. As a rough heuristic, only track pressure when the number of 2939 // schedulable instructions exceeds half the integer register file. 2940 RegionPolicy.ShouldTrackPressure = true; 2941 for (unsigned VT = MVT::i32; VT > (unsigned)MVT::i1; --VT) { 2942 MVT::SimpleValueType LegalIntVT = (MVT::SimpleValueType)VT; 2943 if (TLI->isTypeLegal(LegalIntVT)) { 2944 unsigned NIntRegs = Context->RegClassInfo->getNumAllocatableRegs( 2945 TLI->getRegClassFor(LegalIntVT)); 2946 RegionPolicy.ShouldTrackPressure = NumRegionInstrs > (NIntRegs / 2); 2947 } 2948 } 2949 2950 // For generic targets, we default to bottom-up, because it's simpler and more 2951 // compile-time optimizations have been implemented in that direction. 2952 RegionPolicy.OnlyBottomUp = true; 2953 2954 // Allow the subtarget to override default policy. 2955 MF.getSubtarget().overrideSchedPolicy(RegionPolicy, NumRegionInstrs); 2956 2957 // After subtarget overrides, apply command line options. 2958 if (!EnableRegPressure) { 2959 RegionPolicy.ShouldTrackPressure = false; 2960 RegionPolicy.ShouldTrackLaneMasks = false; 2961 } 2962 2963 // Check -misched-topdown/bottomup can force or unforce scheduling direction. 2964 // e.g. -misched-bottomup=false allows scheduling in both directions. 2965 assert((!ForceTopDown || !ForceBottomUp) && 2966 "-misched-topdown incompatible with -misched-bottomup"); 2967 if (ForceBottomUp.getNumOccurrences() > 0) { 2968 RegionPolicy.OnlyBottomUp = ForceBottomUp; 2969 if (RegionPolicy.OnlyBottomUp) 2970 RegionPolicy.OnlyTopDown = false; 2971 } 2972 if (ForceTopDown.getNumOccurrences() > 0) { 2973 RegionPolicy.OnlyTopDown = ForceTopDown; 2974 if (RegionPolicy.OnlyTopDown) 2975 RegionPolicy.OnlyBottomUp = false; 2976 } 2977 } 2978 2979 void GenericScheduler::dumpPolicy() const { 2980 // Cannot completely remove virtual function even in release mode. 2981 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) 2982 dbgs() << "GenericScheduler RegionPolicy: " 2983 << " ShouldTrackPressure=" << RegionPolicy.ShouldTrackPressure 2984 << " OnlyTopDown=" << RegionPolicy.OnlyTopDown 2985 << " OnlyBottomUp=" << RegionPolicy.OnlyBottomUp 2986 << "\n"; 2987 #endif 2988 } 2989 2990 /// Set IsAcyclicLatencyLimited if the acyclic path is longer than the cyclic 2991 /// critical path by more cycles than it takes to drain the instruction buffer. 2992 /// We estimate an upper bounds on in-flight instructions as: 2993 /// 2994 /// CyclesPerIteration = max( CyclicPath, Loop-Resource-Height ) 2995 /// InFlightIterations = AcyclicPath / CyclesPerIteration 2996 /// InFlightResources = InFlightIterations * LoopResources 2997 /// 2998 /// TODO: Check execution resources in addition to IssueCount. 2999 void GenericScheduler::checkAcyclicLatency() { 3000 if (Rem.CyclicCritPath == 0 || Rem.CyclicCritPath >= Rem.CriticalPath) 3001 return; 3002 3003 // Scaled number of cycles per loop iteration. 3004 unsigned IterCount = 3005 std::max(Rem.CyclicCritPath * SchedModel->getLatencyFactor(), 3006 Rem.RemIssueCount); 3007 // Scaled acyclic critical path. 3008 unsigned AcyclicCount = Rem.CriticalPath * SchedModel->getLatencyFactor(); 3009 // InFlightCount = (AcyclicPath / IterCycles) * InstrPerLoop 3010 unsigned InFlightCount = 3011 (AcyclicCount * Rem.RemIssueCount + IterCount-1) / IterCount; 3012 unsigned BufferLimit = 3013 SchedModel->getMicroOpBufferSize() * SchedModel->getMicroOpFactor(); 3014 3015 Rem.IsAcyclicLatencyLimited = InFlightCount > BufferLimit; 3016 3017 LLVM_DEBUG( 3018 dbgs() << "IssueCycles=" 3019 << Rem.RemIssueCount / SchedModel->getLatencyFactor() << "c " 3020 << "IterCycles=" << IterCount / SchedModel->getLatencyFactor() 3021 << "c NumIters=" << (AcyclicCount + IterCount - 1) / IterCount 3022 << " InFlight=" << InFlightCount / SchedModel->getMicroOpFactor() 3023 << "m BufferLim=" << SchedModel->getMicroOpBufferSize() << "m\n"; 3024 if (Rem.IsAcyclicLatencyLimited) dbgs() << " ACYCLIC LATENCY LIMIT\n"); 3025 } 3026 3027 void GenericScheduler::registerRoots() { 3028 Rem.CriticalPath = DAG->ExitSU.getDepth(); 3029 3030 // Some roots may not feed into ExitSU. Check all of them in case. 3031 for (const SUnit *SU : Bot.Available) { 3032 if (SU->getDepth() > Rem.CriticalPath) 3033 Rem.CriticalPath = SU->getDepth(); 3034 } 3035 LLVM_DEBUG(dbgs() << "Critical Path(GS-RR ): " << Rem.CriticalPath << '\n'); 3036 if (DumpCriticalPathLength) { 3037 errs() << "Critical Path(GS-RR ): " << Rem.CriticalPath << " \n"; 3038 } 3039 3040 if (EnableCyclicPath && SchedModel->getMicroOpBufferSize() > 0) { 3041 Rem.CyclicCritPath = DAG->computeCyclicCriticalPath(); 3042 checkAcyclicLatency(); 3043 } 3044 } 3045 3046 namespace llvm { 3047 bool tryPressure(const PressureChange &TryP, 3048 const PressureChange &CandP, 3049 GenericSchedulerBase::SchedCandidate &TryCand, 3050 GenericSchedulerBase::SchedCandidate &Cand, 3051 GenericSchedulerBase::CandReason Reason, 3052 const TargetRegisterInfo *TRI, 3053 const MachineFunction &MF) { 3054 // If one candidate decreases and the other increases, go with it. 3055 // Invalid candidates have UnitInc==0. 3056 if (tryGreater(TryP.getUnitInc() < 0, CandP.getUnitInc() < 0, TryCand, Cand, 3057 Reason)) { 3058 return true; 3059 } 3060 // Do not compare the magnitude of pressure changes between top and bottom 3061 // boundary. 3062 if (Cand.AtTop != TryCand.AtTop) 3063 return false; 3064 3065 // If both candidates affect the same set in the same boundary, go with the 3066 // smallest increase. 3067 unsigned TryPSet = TryP.getPSetOrMax(); 3068 unsigned CandPSet = CandP.getPSetOrMax(); 3069 if (TryPSet == CandPSet) { 3070 return tryLess(TryP.getUnitInc(), CandP.getUnitInc(), TryCand, Cand, 3071 Reason); 3072 } 3073 3074 int TryRank = TryP.isValid() ? TRI->getRegPressureSetScore(MF, TryPSet) : 3075 std::numeric_limits<int>::max(); 3076 3077 int CandRank = CandP.isValid() ? TRI->getRegPressureSetScore(MF, CandPSet) : 3078 std::numeric_limits<int>::max(); 3079 3080 // If the candidates are decreasing pressure, reverse priority. 3081 if (TryP.getUnitInc() < 0) 3082 std::swap(TryRank, CandRank); 3083 return tryGreater(TryRank, CandRank, TryCand, Cand, Reason); 3084 } 3085 3086 unsigned getWeakLeft(const SUnit *SU, bool isTop) { 3087 return (isTop) ? SU->WeakPredsLeft : SU->WeakSuccsLeft; 3088 } 3089 3090 /// Minimize physical register live ranges. Regalloc wants them adjacent to 3091 /// their physreg def/use. 3092 /// 3093 /// FIXME: This is an unnecessary check on the critical path. Most are root/leaf 3094 /// copies which can be prescheduled. The rest (e.g. x86 MUL) could be bundled 3095 /// with the operation that produces or consumes the physreg. We'll do this when 3096 /// regalloc has support for parallel copies. 3097 int biasPhysReg(const SUnit *SU, bool isTop) { 3098 const MachineInstr *MI = SU->getInstr(); 3099 3100 if (MI->isCopy()) { 3101 unsigned ScheduledOper = isTop ? 1 : 0; 3102 unsigned UnscheduledOper = isTop ? 0 : 1; 3103 // If we have already scheduled the physreg produce/consumer, immediately 3104 // schedule the copy. 3105 if (Register::isPhysicalRegister(MI->getOperand(ScheduledOper).getReg())) 3106 return 1; 3107 // If the physreg is at the boundary, defer it. Otherwise schedule it 3108 // immediately to free the dependent. We can hoist the copy later. 3109 bool AtBoundary = isTop ? !SU->NumSuccsLeft : !SU->NumPredsLeft; 3110 if (Register::isPhysicalRegister(MI->getOperand(UnscheduledOper).getReg())) 3111 return AtBoundary ? -1 : 1; 3112 } 3113 3114 if (MI->isMoveImmediate()) { 3115 // If we have a move immediate and all successors have been assigned, bias 3116 // towards scheduling this later. Make sure all register defs are to 3117 // physical registers. 3118 bool DoBias = true; 3119 for (const MachineOperand &Op : MI->defs()) { 3120 if (Op.isReg() && !Register::isPhysicalRegister(Op.getReg())) { 3121 DoBias = false; 3122 break; 3123 } 3124 } 3125 3126 if (DoBias) 3127 return isTop ? -1 : 1; 3128 } 3129 3130 return 0; 3131 } 3132 } // end namespace llvm 3133 3134 void GenericScheduler::initCandidate(SchedCandidate &Cand, SUnit *SU, 3135 bool AtTop, 3136 const RegPressureTracker &RPTracker, 3137 RegPressureTracker &TempTracker) { 3138 Cand.SU = SU; 3139 Cand.AtTop = AtTop; 3140 if (DAG->isTrackingPressure()) { 3141 if (AtTop) { 3142 TempTracker.getMaxDownwardPressureDelta( 3143 Cand.SU->getInstr(), 3144 Cand.RPDelta, 3145 DAG->getRegionCriticalPSets(), 3146 DAG->getRegPressure().MaxSetPressure); 3147 } else { 3148 if (VerifyScheduling) { 3149 TempTracker.getMaxUpwardPressureDelta( 3150 Cand.SU->getInstr(), 3151 &DAG->getPressureDiff(Cand.SU), 3152 Cand.RPDelta, 3153 DAG->getRegionCriticalPSets(), 3154 DAG->getRegPressure().MaxSetPressure); 3155 } else { 3156 RPTracker.getUpwardPressureDelta( 3157 Cand.SU->getInstr(), 3158 DAG->getPressureDiff(Cand.SU), 3159 Cand.RPDelta, 3160 DAG->getRegionCriticalPSets(), 3161 DAG->getRegPressure().MaxSetPressure); 3162 } 3163 } 3164 } 3165 LLVM_DEBUG(if (Cand.RPDelta.Excess.isValid()) dbgs() 3166 << " Try SU(" << Cand.SU->NodeNum << ") " 3167 << TRI->getRegPressureSetName(Cand.RPDelta.Excess.getPSet()) << ":" 3168 << Cand.RPDelta.Excess.getUnitInc() << "\n"); 3169 } 3170 3171 /// Apply a set of heuristics to a new candidate. Heuristics are currently 3172 /// hierarchical. This may be more efficient than a graduated cost model because 3173 /// we don't need to evaluate all aspects of the model for each node in the 3174 /// queue. But it's really done to make the heuristics easier to debug and 3175 /// statistically analyze. 3176 /// 3177 /// \param Cand provides the policy and current best candidate. 3178 /// \param TryCand refers to the next SUnit candidate, otherwise uninitialized. 3179 /// \param Zone describes the scheduled zone that we are extending, or nullptr 3180 /// if Cand is from a different zone than TryCand. 3181 /// \return \c true if TryCand is better than Cand (Reason is NOT NoCand) 3182 bool GenericScheduler::tryCandidate(SchedCandidate &Cand, 3183 SchedCandidate &TryCand, 3184 SchedBoundary *Zone) const { 3185 // Initialize the candidate if needed. 3186 if (!Cand.isValid()) { 3187 TryCand.Reason = NodeOrder; 3188 return true; 3189 } 3190 3191 // Bias PhysReg Defs and copies to their uses and defined respectively. 3192 if (tryGreater(biasPhysReg(TryCand.SU, TryCand.AtTop), 3193 biasPhysReg(Cand.SU, Cand.AtTop), TryCand, Cand, PhysReg)) 3194 return TryCand.Reason != NoCand; 3195 3196 // Avoid exceeding the target's limit. 3197 if (DAG->isTrackingPressure() && tryPressure(TryCand.RPDelta.Excess, 3198 Cand.RPDelta.Excess, 3199 TryCand, Cand, RegExcess, TRI, 3200 DAG->MF)) 3201 return TryCand.Reason != NoCand; 3202 3203 // Avoid increasing the max critical pressure in the scheduled region. 3204 if (DAG->isTrackingPressure() && tryPressure(TryCand.RPDelta.CriticalMax, 3205 Cand.RPDelta.CriticalMax, 3206 TryCand, Cand, RegCritical, TRI, 3207 DAG->MF)) 3208 return TryCand.Reason != NoCand; 3209 3210 // We only compare a subset of features when comparing nodes between 3211 // Top and Bottom boundary. Some properties are simply incomparable, in many 3212 // other instances we should only override the other boundary if something 3213 // is a clear good pick on one boundary. Skip heuristics that are more 3214 // "tie-breaking" in nature. 3215 bool SameBoundary = Zone != nullptr; 3216 if (SameBoundary) { 3217 // For loops that are acyclic path limited, aggressively schedule for 3218 // latency. Within an single cycle, whenever CurrMOps > 0, allow normal 3219 // heuristics to take precedence. 3220 if (Rem.IsAcyclicLatencyLimited && !Zone->getCurrMOps() && 3221 tryLatency(TryCand, Cand, *Zone)) 3222 return TryCand.Reason != NoCand; 3223 3224 // Prioritize instructions that read unbuffered resources by stall cycles. 3225 if (tryLess(Zone->getLatencyStallCycles(TryCand.SU), 3226 Zone->getLatencyStallCycles(Cand.SU), TryCand, Cand, Stall)) 3227 return TryCand.Reason != NoCand; 3228 } 3229 3230 // Keep clustered nodes together to encourage downstream peephole 3231 // optimizations which may reduce resource requirements. 3232 // 3233 // This is a best effort to set things up for a post-RA pass. Optimizations 3234 // like generating loads of multiple registers should ideally be done within 3235 // the scheduler pass by combining the loads during DAG postprocessing. 3236 const SUnit *CandNextClusterSU = 3237 Cand.AtTop ? DAG->getNextClusterSucc() : DAG->getNextClusterPred(); 3238 const SUnit *TryCandNextClusterSU = 3239 TryCand.AtTop ? DAG->getNextClusterSucc() : DAG->getNextClusterPred(); 3240 if (tryGreater(TryCand.SU == TryCandNextClusterSU, 3241 Cand.SU == CandNextClusterSU, 3242 TryCand, Cand, Cluster)) 3243 return TryCand.Reason != NoCand; 3244 3245 if (SameBoundary) { 3246 // Weak edges are for clustering and other constraints. 3247 if (tryLess(getWeakLeft(TryCand.SU, TryCand.AtTop), 3248 getWeakLeft(Cand.SU, Cand.AtTop), 3249 TryCand, Cand, Weak)) 3250 return TryCand.Reason != NoCand; 3251 } 3252 3253 // Avoid increasing the max pressure of the entire region. 3254 if (DAG->isTrackingPressure() && tryPressure(TryCand.RPDelta.CurrentMax, 3255 Cand.RPDelta.CurrentMax, 3256 TryCand, Cand, RegMax, TRI, 3257 DAG->MF)) 3258 return TryCand.Reason != NoCand; 3259 3260 if (SameBoundary) { 3261 // Avoid critical resource consumption and balance the schedule. 3262 TryCand.initResourceDelta(DAG, SchedModel); 3263 if (tryLess(TryCand.ResDelta.CritResources, Cand.ResDelta.CritResources, 3264 TryCand, Cand, ResourceReduce)) 3265 return TryCand.Reason != NoCand; 3266 if (tryGreater(TryCand.ResDelta.DemandedResources, 3267 Cand.ResDelta.DemandedResources, 3268 TryCand, Cand, ResourceDemand)) 3269 return TryCand.Reason != NoCand; 3270 3271 // Avoid serializing long latency dependence chains. 3272 // For acyclic path limited loops, latency was already checked above. 3273 if (!RegionPolicy.DisableLatencyHeuristic && TryCand.Policy.ReduceLatency && 3274 !Rem.IsAcyclicLatencyLimited && tryLatency(TryCand, Cand, *Zone)) 3275 return TryCand.Reason != NoCand; 3276 3277 // Fall through to original instruction order. 3278 if ((Zone->isTop() && TryCand.SU->NodeNum < Cand.SU->NodeNum) 3279 || (!Zone->isTop() && TryCand.SU->NodeNum > Cand.SU->NodeNum)) { 3280 TryCand.Reason = NodeOrder; 3281 return true; 3282 } 3283 } 3284 3285 return false; 3286 } 3287 3288 /// Pick the best candidate from the queue. 3289 /// 3290 /// TODO: getMaxPressureDelta results can be mostly cached for each SUnit during 3291 /// DAG building. To adjust for the current scheduling location we need to 3292 /// maintain the number of vreg uses remaining to be top-scheduled. 3293 void GenericScheduler::pickNodeFromQueue(SchedBoundary &Zone, 3294 const CandPolicy &ZonePolicy, 3295 const RegPressureTracker &RPTracker, 3296 SchedCandidate &Cand) { 3297 // getMaxPressureDelta temporarily modifies the tracker. 3298 RegPressureTracker &TempTracker = const_cast<RegPressureTracker&>(RPTracker); 3299 3300 ReadyQueue &Q = Zone.Available; 3301 for (SUnit *SU : Q) { 3302 3303 SchedCandidate TryCand(ZonePolicy); 3304 initCandidate(TryCand, SU, Zone.isTop(), RPTracker, TempTracker); 3305 // Pass SchedBoundary only when comparing nodes from the same boundary. 3306 SchedBoundary *ZoneArg = Cand.AtTop == TryCand.AtTop ? &Zone : nullptr; 3307 if (tryCandidate(Cand, TryCand, ZoneArg)) { 3308 // Initialize resource delta if needed in case future heuristics query it. 3309 if (TryCand.ResDelta == SchedResourceDelta()) 3310 TryCand.initResourceDelta(DAG, SchedModel); 3311 Cand.setBest(TryCand); 3312 LLVM_DEBUG(traceCandidate(Cand)); 3313 } 3314 } 3315 } 3316 3317 /// Pick the best candidate node from either the top or bottom queue. 3318 SUnit *GenericScheduler::pickNodeBidirectional(bool &IsTopNode) { 3319 // Schedule as far as possible in the direction of no choice. This is most 3320 // efficient, but also provides the best heuristics for CriticalPSets. 3321 if (SUnit *SU = Bot.pickOnlyChoice()) { 3322 IsTopNode = false; 3323 tracePick(Only1, false); 3324 return SU; 3325 } 3326 if (SUnit *SU = Top.pickOnlyChoice()) { 3327 IsTopNode = true; 3328 tracePick(Only1, true); 3329 return SU; 3330 } 3331 // Set the bottom-up policy based on the state of the current bottom zone and 3332 // the instructions outside the zone, including the top zone. 3333 CandPolicy BotPolicy; 3334 setPolicy(BotPolicy, /*IsPostRA=*/false, Bot, &Top); 3335 // Set the top-down policy based on the state of the current top zone and 3336 // the instructions outside the zone, including the bottom zone. 3337 CandPolicy TopPolicy; 3338 setPolicy(TopPolicy, /*IsPostRA=*/false, Top, &Bot); 3339 3340 // See if BotCand is still valid (because we previously scheduled from Top). 3341 LLVM_DEBUG(dbgs() << "Picking from Bot:\n"); 3342 if (!BotCand.isValid() || BotCand.SU->isScheduled || 3343 BotCand.Policy != BotPolicy) { 3344 BotCand.reset(CandPolicy()); 3345 pickNodeFromQueue(Bot, BotPolicy, DAG->getBotRPTracker(), BotCand); 3346 assert(BotCand.Reason != NoCand && "failed to find the first candidate"); 3347 } else { 3348 LLVM_DEBUG(traceCandidate(BotCand)); 3349 #ifndef NDEBUG 3350 if (VerifyScheduling) { 3351 SchedCandidate TCand; 3352 TCand.reset(CandPolicy()); 3353 pickNodeFromQueue(Bot, BotPolicy, DAG->getBotRPTracker(), TCand); 3354 assert(TCand.SU == BotCand.SU && 3355 "Last pick result should correspond to re-picking right now"); 3356 } 3357 #endif 3358 } 3359 3360 // Check if the top Q has a better candidate. 3361 LLVM_DEBUG(dbgs() << "Picking from Top:\n"); 3362 if (!TopCand.isValid() || TopCand.SU->isScheduled || 3363 TopCand.Policy != TopPolicy) { 3364 TopCand.reset(CandPolicy()); 3365 pickNodeFromQueue(Top, TopPolicy, DAG->getTopRPTracker(), TopCand); 3366 assert(TopCand.Reason != NoCand && "failed to find the first candidate"); 3367 } else { 3368 LLVM_DEBUG(traceCandidate(TopCand)); 3369 #ifndef NDEBUG 3370 if (VerifyScheduling) { 3371 SchedCandidate TCand; 3372 TCand.reset(CandPolicy()); 3373 pickNodeFromQueue(Top, TopPolicy, DAG->getTopRPTracker(), TCand); 3374 assert(TCand.SU == TopCand.SU && 3375 "Last pick result should correspond to re-picking right now"); 3376 } 3377 #endif 3378 } 3379 3380 // Pick best from BotCand and TopCand. 3381 assert(BotCand.isValid()); 3382 assert(TopCand.isValid()); 3383 SchedCandidate Cand = BotCand; 3384 TopCand.Reason = NoCand; 3385 if (tryCandidate(Cand, TopCand, nullptr)) { 3386 Cand.setBest(TopCand); 3387 LLVM_DEBUG(traceCandidate(Cand)); 3388 } 3389 3390 IsTopNode = Cand.AtTop; 3391 tracePick(Cand); 3392 return Cand.SU; 3393 } 3394 3395 /// Pick the best node to balance the schedule. Implements MachineSchedStrategy. 3396 SUnit *GenericScheduler::pickNode(bool &IsTopNode) { 3397 if (DAG->top() == DAG->bottom()) { 3398 assert(Top.Available.empty() && Top.Pending.empty() && 3399 Bot.Available.empty() && Bot.Pending.empty() && "ReadyQ garbage"); 3400 return nullptr; 3401 } 3402 SUnit *SU; 3403 do { 3404 if (RegionPolicy.OnlyTopDown) { 3405 SU = Top.pickOnlyChoice(); 3406 if (!SU) { 3407 CandPolicy NoPolicy; 3408 TopCand.reset(NoPolicy); 3409 pickNodeFromQueue(Top, NoPolicy, DAG->getTopRPTracker(), TopCand); 3410 assert(TopCand.Reason != NoCand && "failed to find a candidate"); 3411 tracePick(TopCand); 3412 SU = TopCand.SU; 3413 } 3414 IsTopNode = true; 3415 } else if (RegionPolicy.OnlyBottomUp) { 3416 SU = Bot.pickOnlyChoice(); 3417 if (!SU) { 3418 CandPolicy NoPolicy; 3419 BotCand.reset(NoPolicy); 3420 pickNodeFromQueue(Bot, NoPolicy, DAG->getBotRPTracker(), BotCand); 3421 assert(BotCand.Reason != NoCand && "failed to find a candidate"); 3422 tracePick(BotCand); 3423 SU = BotCand.SU; 3424 } 3425 IsTopNode = false; 3426 } else { 3427 SU = pickNodeBidirectional(IsTopNode); 3428 } 3429 } while (SU->isScheduled); 3430 3431 if (SU->isTopReady()) 3432 Top.removeReady(SU); 3433 if (SU->isBottomReady()) 3434 Bot.removeReady(SU); 3435 3436 LLVM_DEBUG(dbgs() << "Scheduling SU(" << SU->NodeNum << ") " 3437 << *SU->getInstr()); 3438 return SU; 3439 } 3440 3441 void GenericScheduler::reschedulePhysReg(SUnit *SU, bool isTop) { 3442 MachineBasicBlock::iterator InsertPos = SU->getInstr(); 3443 if (!isTop) 3444 ++InsertPos; 3445 SmallVectorImpl<SDep> &Deps = isTop ? SU->Preds : SU->Succs; 3446 3447 // Find already scheduled copies with a single physreg dependence and move 3448 // them just above the scheduled instruction. 3449 for (SDep &Dep : Deps) { 3450 if (Dep.getKind() != SDep::Data || 3451 !Register::isPhysicalRegister(Dep.getReg())) 3452 continue; 3453 SUnit *DepSU = Dep.getSUnit(); 3454 if (isTop ? DepSU->Succs.size() > 1 : DepSU->Preds.size() > 1) 3455 continue; 3456 MachineInstr *Copy = DepSU->getInstr(); 3457 if (!Copy->isCopy() && !Copy->isMoveImmediate()) 3458 continue; 3459 LLVM_DEBUG(dbgs() << " Rescheduling physreg copy "; 3460 DAG->dumpNode(*Dep.getSUnit())); 3461 DAG->moveInstruction(Copy, InsertPos); 3462 } 3463 } 3464 3465 /// Update the scheduler's state after scheduling a node. This is the same node 3466 /// that was just returned by pickNode(). However, ScheduleDAGMILive needs to 3467 /// update it's state based on the current cycle before MachineSchedStrategy 3468 /// does. 3469 /// 3470 /// FIXME: Eventually, we may bundle physreg copies rather than rescheduling 3471 /// them here. See comments in biasPhysReg. 3472 void GenericScheduler::schedNode(SUnit *SU, bool IsTopNode) { 3473 if (IsTopNode) { 3474 SU->TopReadyCycle = std::max(SU->TopReadyCycle, Top.getCurrCycle()); 3475 Top.bumpNode(SU); 3476 if (SU->hasPhysRegUses) 3477 reschedulePhysReg(SU, true); 3478 } else { 3479 SU->BotReadyCycle = std::max(SU->BotReadyCycle, Bot.getCurrCycle()); 3480 Bot.bumpNode(SU); 3481 if (SU->hasPhysRegDefs) 3482 reschedulePhysReg(SU, false); 3483 } 3484 } 3485 3486 /// Create the standard converging machine scheduler. This will be used as the 3487 /// default scheduler if the target does not set a default. 3488 ScheduleDAGMILive *llvm::createGenericSchedLive(MachineSchedContext *C) { 3489 ScheduleDAGMILive *DAG = 3490 new ScheduleDAGMILive(C, std::make_unique<GenericScheduler>(C)); 3491 // Register DAG post-processors. 3492 // 3493 // FIXME: extend the mutation API to allow earlier mutations to instantiate 3494 // data and pass it to later mutations. Have a single mutation that gathers 3495 // the interesting nodes in one pass. 3496 DAG->addMutation(createCopyConstrainDAGMutation(DAG->TII, DAG->TRI)); 3497 return DAG; 3498 } 3499 3500 static ScheduleDAGInstrs *createConvergingSched(MachineSchedContext *C) { 3501 return createGenericSchedLive(C); 3502 } 3503 3504 static MachineSchedRegistry 3505 GenericSchedRegistry("converge", "Standard converging scheduler.", 3506 createConvergingSched); 3507 3508 //===----------------------------------------------------------------------===// 3509 // PostGenericScheduler - Generic PostRA implementation of MachineSchedStrategy. 3510 //===----------------------------------------------------------------------===// 3511 3512 void PostGenericScheduler::initialize(ScheduleDAGMI *Dag) { 3513 DAG = Dag; 3514 SchedModel = DAG->getSchedModel(); 3515 TRI = DAG->TRI; 3516 3517 Rem.init(DAG, SchedModel); 3518 Top.init(DAG, SchedModel, &Rem); 3519 BotRoots.clear(); 3520 3521 // Initialize the HazardRecognizers. If itineraries don't exist, are empty, 3522 // or are disabled, then these HazardRecs will be disabled. 3523 const InstrItineraryData *Itin = SchedModel->getInstrItineraries(); 3524 if (!Top.HazardRec) { 3525 Top.HazardRec = 3526 DAG->MF.getSubtarget().getInstrInfo()->CreateTargetMIHazardRecognizer( 3527 Itin, DAG); 3528 } 3529 } 3530 3531 void PostGenericScheduler::registerRoots() { 3532 Rem.CriticalPath = DAG->ExitSU.getDepth(); 3533 3534 // Some roots may not feed into ExitSU. Check all of them in case. 3535 for (const SUnit *SU : BotRoots) { 3536 if (SU->getDepth() > Rem.CriticalPath) 3537 Rem.CriticalPath = SU->getDepth(); 3538 } 3539 LLVM_DEBUG(dbgs() << "Critical Path: (PGS-RR) " << Rem.CriticalPath << '\n'); 3540 if (DumpCriticalPathLength) { 3541 errs() << "Critical Path(PGS-RR ): " << Rem.CriticalPath << " \n"; 3542 } 3543 } 3544 3545 /// Apply a set of heuristics to a new candidate for PostRA scheduling. 3546 /// 3547 /// \param Cand provides the policy and current best candidate. 3548 /// \param TryCand refers to the next SUnit candidate, otherwise uninitialized. 3549 /// \return \c true if TryCand is better than Cand (Reason is NOT NoCand) 3550 bool PostGenericScheduler::tryCandidate(SchedCandidate &Cand, 3551 SchedCandidate &TryCand) { 3552 // Initialize the candidate if needed. 3553 if (!Cand.isValid()) { 3554 TryCand.Reason = NodeOrder; 3555 return true; 3556 } 3557 3558 // Prioritize instructions that read unbuffered resources by stall cycles. 3559 if (tryLess(Top.getLatencyStallCycles(TryCand.SU), 3560 Top.getLatencyStallCycles(Cand.SU), TryCand, Cand, Stall)) 3561 return TryCand.Reason != NoCand; 3562 3563 // Keep clustered nodes together. 3564 if (tryGreater(TryCand.SU == DAG->getNextClusterSucc(), 3565 Cand.SU == DAG->getNextClusterSucc(), 3566 TryCand, Cand, Cluster)) 3567 return TryCand.Reason != NoCand; 3568 3569 // Avoid critical resource consumption and balance the schedule. 3570 if (tryLess(TryCand.ResDelta.CritResources, Cand.ResDelta.CritResources, 3571 TryCand, Cand, ResourceReduce)) 3572 return TryCand.Reason != NoCand; 3573 if (tryGreater(TryCand.ResDelta.DemandedResources, 3574 Cand.ResDelta.DemandedResources, 3575 TryCand, Cand, ResourceDemand)) 3576 return TryCand.Reason != NoCand; 3577 3578 // Avoid serializing long latency dependence chains. 3579 if (Cand.Policy.ReduceLatency && tryLatency(TryCand, Cand, Top)) { 3580 return TryCand.Reason != NoCand; 3581 } 3582 3583 // Fall through to original instruction order. 3584 if (TryCand.SU->NodeNum < Cand.SU->NodeNum) { 3585 TryCand.Reason = NodeOrder; 3586 return true; 3587 } 3588 3589 return false; 3590 } 3591 3592 void PostGenericScheduler::pickNodeFromQueue(SchedCandidate &Cand) { 3593 ReadyQueue &Q = Top.Available; 3594 for (SUnit *SU : Q) { 3595 SchedCandidate TryCand(Cand.Policy); 3596 TryCand.SU = SU; 3597 TryCand.AtTop = true; 3598 TryCand.initResourceDelta(DAG, SchedModel); 3599 if (tryCandidate(Cand, TryCand)) { 3600 Cand.setBest(TryCand); 3601 LLVM_DEBUG(traceCandidate(Cand)); 3602 } 3603 } 3604 } 3605 3606 /// Pick the next node to schedule. 3607 SUnit *PostGenericScheduler::pickNode(bool &IsTopNode) { 3608 if (DAG->top() == DAG->bottom()) { 3609 assert(Top.Available.empty() && Top.Pending.empty() && "ReadyQ garbage"); 3610 return nullptr; 3611 } 3612 SUnit *SU; 3613 do { 3614 SU = Top.pickOnlyChoice(); 3615 if (SU) { 3616 tracePick(Only1, true); 3617 } else { 3618 CandPolicy NoPolicy; 3619 SchedCandidate TopCand(NoPolicy); 3620 // Set the top-down policy based on the state of the current top zone and 3621 // the instructions outside the zone, including the bottom zone. 3622 setPolicy(TopCand.Policy, /*IsPostRA=*/true, Top, nullptr); 3623 pickNodeFromQueue(TopCand); 3624 assert(TopCand.Reason != NoCand && "failed to find a candidate"); 3625 tracePick(TopCand); 3626 SU = TopCand.SU; 3627 } 3628 } while (SU->isScheduled); 3629 3630 IsTopNode = true; 3631 Top.removeReady(SU); 3632 3633 LLVM_DEBUG(dbgs() << "Scheduling SU(" << SU->NodeNum << ") " 3634 << *SU->getInstr()); 3635 return SU; 3636 } 3637 3638 /// Called after ScheduleDAGMI has scheduled an instruction and updated 3639 /// scheduled/remaining flags in the DAG nodes. 3640 void PostGenericScheduler::schedNode(SUnit *SU, bool IsTopNode) { 3641 SU->TopReadyCycle = std::max(SU->TopReadyCycle, Top.getCurrCycle()); 3642 Top.bumpNode(SU); 3643 } 3644 3645 ScheduleDAGMI *llvm::createGenericSchedPostRA(MachineSchedContext *C) { 3646 return new ScheduleDAGMI(C, std::make_unique<PostGenericScheduler>(C), 3647 /*RemoveKillFlags=*/true); 3648 } 3649 3650 //===----------------------------------------------------------------------===// 3651 // ILP Scheduler. Currently for experimental analysis of heuristics. 3652 //===----------------------------------------------------------------------===// 3653 3654 namespace { 3655 3656 /// Order nodes by the ILP metric. 3657 struct ILPOrder { 3658 const SchedDFSResult *DFSResult = nullptr; 3659 const BitVector *ScheduledTrees = nullptr; 3660 bool MaximizeILP; 3661 3662 ILPOrder(bool MaxILP) : MaximizeILP(MaxILP) {} 3663 3664 /// Apply a less-than relation on node priority. 3665 /// 3666 /// (Return true if A comes after B in the Q.) 3667 bool operator()(const SUnit *A, const SUnit *B) const { 3668 unsigned SchedTreeA = DFSResult->getSubtreeID(A); 3669 unsigned SchedTreeB = DFSResult->getSubtreeID(B); 3670 if (SchedTreeA != SchedTreeB) { 3671 // Unscheduled trees have lower priority. 3672 if (ScheduledTrees->test(SchedTreeA) != ScheduledTrees->test(SchedTreeB)) 3673 return ScheduledTrees->test(SchedTreeB); 3674 3675 // Trees with shallower connections have have lower priority. 3676 if (DFSResult->getSubtreeLevel(SchedTreeA) 3677 != DFSResult->getSubtreeLevel(SchedTreeB)) { 3678 return DFSResult->getSubtreeLevel(SchedTreeA) 3679 < DFSResult->getSubtreeLevel(SchedTreeB); 3680 } 3681 } 3682 if (MaximizeILP) 3683 return DFSResult->getILP(A) < DFSResult->getILP(B); 3684 else 3685 return DFSResult->getILP(A) > DFSResult->getILP(B); 3686 } 3687 }; 3688 3689 /// Schedule based on the ILP metric. 3690 class ILPScheduler : public MachineSchedStrategy { 3691 ScheduleDAGMILive *DAG = nullptr; 3692 ILPOrder Cmp; 3693 3694 std::vector<SUnit*> ReadyQ; 3695 3696 public: 3697 ILPScheduler(bool MaximizeILP) : Cmp(MaximizeILP) {} 3698 3699 void initialize(ScheduleDAGMI *dag) override { 3700 assert(dag->hasVRegLiveness() && "ILPScheduler needs vreg liveness"); 3701 DAG = static_cast<ScheduleDAGMILive*>(dag); 3702 DAG->computeDFSResult(); 3703 Cmp.DFSResult = DAG->getDFSResult(); 3704 Cmp.ScheduledTrees = &DAG->getScheduledTrees(); 3705 ReadyQ.clear(); 3706 } 3707 3708 void registerRoots() override { 3709 // Restore the heap in ReadyQ with the updated DFS results. 3710 std::make_heap(ReadyQ.begin(), ReadyQ.end(), Cmp); 3711 } 3712 3713 /// Implement MachineSchedStrategy interface. 3714 /// ----------------------------------------- 3715 3716 /// Callback to select the highest priority node from the ready Q. 3717 SUnit *pickNode(bool &IsTopNode) override { 3718 if (ReadyQ.empty()) return nullptr; 3719 std::pop_heap(ReadyQ.begin(), ReadyQ.end(), Cmp); 3720 SUnit *SU = ReadyQ.back(); 3721 ReadyQ.pop_back(); 3722 IsTopNode = false; 3723 LLVM_DEBUG(dbgs() << "Pick node " 3724 << "SU(" << SU->NodeNum << ") " 3725 << " ILP: " << DAG->getDFSResult()->getILP(SU) 3726 << " Tree: " << DAG->getDFSResult()->getSubtreeID(SU) 3727 << " @" 3728 << DAG->getDFSResult()->getSubtreeLevel( 3729 DAG->getDFSResult()->getSubtreeID(SU)) 3730 << '\n' 3731 << "Scheduling " << *SU->getInstr()); 3732 return SU; 3733 } 3734 3735 /// Scheduler callback to notify that a new subtree is scheduled. 3736 void scheduleTree(unsigned SubtreeID) override { 3737 std::make_heap(ReadyQ.begin(), ReadyQ.end(), Cmp); 3738 } 3739 3740 /// Callback after a node is scheduled. Mark a newly scheduled tree, notify 3741 /// DFSResults, and resort the priority Q. 3742 void schedNode(SUnit *SU, bool IsTopNode) override { 3743 assert(!IsTopNode && "SchedDFSResult needs bottom-up"); 3744 } 3745 3746 void releaseTopNode(SUnit *) override { /*only called for top roots*/ } 3747 3748 void releaseBottomNode(SUnit *SU) override { 3749 ReadyQ.push_back(SU); 3750 std::push_heap(ReadyQ.begin(), ReadyQ.end(), Cmp); 3751 } 3752 }; 3753 3754 } // end anonymous namespace 3755 3756 static ScheduleDAGInstrs *createILPMaxScheduler(MachineSchedContext *C) { 3757 return new ScheduleDAGMILive(C, std::make_unique<ILPScheduler>(true)); 3758 } 3759 static ScheduleDAGInstrs *createILPMinScheduler(MachineSchedContext *C) { 3760 return new ScheduleDAGMILive(C, std::make_unique<ILPScheduler>(false)); 3761 } 3762 3763 static MachineSchedRegistry ILPMaxRegistry( 3764 "ilpmax", "Schedule bottom-up for max ILP", createILPMaxScheduler); 3765 static MachineSchedRegistry ILPMinRegistry( 3766 "ilpmin", "Schedule bottom-up for min ILP", createILPMinScheduler); 3767 3768 //===----------------------------------------------------------------------===// 3769 // Machine Instruction Shuffler for Correctness Testing 3770 //===----------------------------------------------------------------------===// 3771 3772 #ifndef NDEBUG 3773 namespace { 3774 3775 /// Apply a less-than relation on the node order, which corresponds to the 3776 /// instruction order prior to scheduling. IsReverse implements greater-than. 3777 template<bool IsReverse> 3778 struct SUnitOrder { 3779 bool operator()(SUnit *A, SUnit *B) const { 3780 if (IsReverse) 3781 return A->NodeNum > B->NodeNum; 3782 else 3783 return A->NodeNum < B->NodeNum; 3784 } 3785 }; 3786 3787 /// Reorder instructions as much as possible. 3788 class InstructionShuffler : public MachineSchedStrategy { 3789 bool IsAlternating; 3790 bool IsTopDown; 3791 3792 // Using a less-than relation (SUnitOrder<false>) for the TopQ priority 3793 // gives nodes with a higher number higher priority causing the latest 3794 // instructions to be scheduled first. 3795 PriorityQueue<SUnit*, std::vector<SUnit*>, SUnitOrder<false>> 3796 TopQ; 3797 3798 // When scheduling bottom-up, use greater-than as the queue priority. 3799 PriorityQueue<SUnit*, std::vector<SUnit*>, SUnitOrder<true>> 3800 BottomQ; 3801 3802 public: 3803 InstructionShuffler(bool alternate, bool topdown) 3804 : IsAlternating(alternate), IsTopDown(topdown) {} 3805 3806 void initialize(ScheduleDAGMI*) override { 3807 TopQ.clear(); 3808 BottomQ.clear(); 3809 } 3810 3811 /// Implement MachineSchedStrategy interface. 3812 /// ----------------------------------------- 3813 3814 SUnit *pickNode(bool &IsTopNode) override { 3815 SUnit *SU; 3816 if (IsTopDown) { 3817 do { 3818 if (TopQ.empty()) return nullptr; 3819 SU = TopQ.top(); 3820 TopQ.pop(); 3821 } while (SU->isScheduled); 3822 IsTopNode = true; 3823 } else { 3824 do { 3825 if (BottomQ.empty()) return nullptr; 3826 SU = BottomQ.top(); 3827 BottomQ.pop(); 3828 } while (SU->isScheduled); 3829 IsTopNode = false; 3830 } 3831 if (IsAlternating) 3832 IsTopDown = !IsTopDown; 3833 return SU; 3834 } 3835 3836 void schedNode(SUnit *SU, bool IsTopNode) override {} 3837 3838 void releaseTopNode(SUnit *SU) override { 3839 TopQ.push(SU); 3840 } 3841 void releaseBottomNode(SUnit *SU) override { 3842 BottomQ.push(SU); 3843 } 3844 }; 3845 3846 } // end anonymous namespace 3847 3848 static ScheduleDAGInstrs *createInstructionShuffler(MachineSchedContext *C) { 3849 bool Alternate = !ForceTopDown && !ForceBottomUp; 3850 bool TopDown = !ForceBottomUp; 3851 assert((TopDown || !ForceTopDown) && 3852 "-misched-topdown incompatible with -misched-bottomup"); 3853 return new ScheduleDAGMILive( 3854 C, std::make_unique<InstructionShuffler>(Alternate, TopDown)); 3855 } 3856 3857 static MachineSchedRegistry ShufflerRegistry( 3858 "shuffle", "Shuffle machine instructions alternating directions", 3859 createInstructionShuffler); 3860 #endif // !NDEBUG 3861 3862 //===----------------------------------------------------------------------===// 3863 // GraphWriter support for ScheduleDAGMILive. 3864 //===----------------------------------------------------------------------===// 3865 3866 #ifndef NDEBUG 3867 namespace llvm { 3868 3869 template<> struct GraphTraits< 3870 ScheduleDAGMI*> : public GraphTraits<ScheduleDAG*> {}; 3871 3872 template<> 3873 struct DOTGraphTraits<ScheduleDAGMI*> : public DefaultDOTGraphTraits { 3874 DOTGraphTraits(bool isSimple = false) : DefaultDOTGraphTraits(isSimple) {} 3875 3876 static std::string getGraphName(const ScheduleDAG *G) { 3877 return std::string(G->MF.getName()); 3878 } 3879 3880 static bool renderGraphFromBottomUp() { 3881 return true; 3882 } 3883 3884 static bool isNodeHidden(const SUnit *Node, const ScheduleDAG *G) { 3885 if (ViewMISchedCutoff == 0) 3886 return false; 3887 return (Node->Preds.size() > ViewMISchedCutoff 3888 || Node->Succs.size() > ViewMISchedCutoff); 3889 } 3890 3891 /// If you want to override the dot attributes printed for a particular 3892 /// edge, override this method. 3893 static std::string getEdgeAttributes(const SUnit *Node, 3894 SUnitIterator EI, 3895 const ScheduleDAG *Graph) { 3896 if (EI.isArtificialDep()) 3897 return "color=cyan,style=dashed"; 3898 if (EI.isCtrlDep()) 3899 return "color=blue,style=dashed"; 3900 return ""; 3901 } 3902 3903 static std::string getNodeLabel(const SUnit *SU, const ScheduleDAG *G) { 3904 std::string Str; 3905 raw_string_ostream SS(Str); 3906 const ScheduleDAGMI *DAG = static_cast<const ScheduleDAGMI*>(G); 3907 const SchedDFSResult *DFS = DAG->hasVRegLiveness() ? 3908 static_cast<const ScheduleDAGMILive*>(G)->getDFSResult() : nullptr; 3909 SS << "SU:" << SU->NodeNum; 3910 if (DFS) 3911 SS << " I:" << DFS->getNumInstrs(SU); 3912 return SS.str(); 3913 } 3914 3915 static std::string getNodeDescription(const SUnit *SU, const ScheduleDAG *G) { 3916 return G->getGraphNodeLabel(SU); 3917 } 3918 3919 static std::string getNodeAttributes(const SUnit *N, const ScheduleDAG *G) { 3920 std::string Str("shape=Mrecord"); 3921 const ScheduleDAGMI *DAG = static_cast<const ScheduleDAGMI*>(G); 3922 const SchedDFSResult *DFS = DAG->hasVRegLiveness() ? 3923 static_cast<const ScheduleDAGMILive*>(G)->getDFSResult() : nullptr; 3924 if (DFS) { 3925 Str += ",style=filled,fillcolor=\"#"; 3926 Str += DOT::getColorString(DFS->getSubtreeID(N)); 3927 Str += '"'; 3928 } 3929 return Str; 3930 } 3931 }; 3932 3933 } // end namespace llvm 3934 #endif // NDEBUG 3935 3936 /// viewGraph - Pop up a ghostview window with the reachable parts of the DAG 3937 /// rendered using 'dot'. 3938 void ScheduleDAGMI::viewGraph(const Twine &Name, const Twine &Title) { 3939 #ifndef NDEBUG 3940 ViewGraph(this, Name, false, Title); 3941 #else 3942 errs() << "ScheduleDAGMI::viewGraph is only available in debug builds on " 3943 << "systems with Graphviz or gv!\n"; 3944 #endif // NDEBUG 3945 } 3946 3947 /// Out-of-line implementation with no arguments is handy for gdb. 3948 void ScheduleDAGMI::viewGraph() { 3949 viewGraph(getDAGName(), "Scheduling-Units Graph for " + getDAGName()); 3950 } 3951