1 //===- MachineLICM.cpp - Machine Loop Invariant Code Motion Pass ----------===// 2 // 3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4 // See https://llvm.org/LICENSE.txt for license information. 5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6 // 7 //===----------------------------------------------------------------------===// 8 // 9 // This pass performs loop invariant code motion on machine instructions. We 10 // attempt to remove as much code from the body of a loop as possible. 11 // 12 // This pass is not intended to be a replacement or a complete alternative 13 // for the LLVM-IR-level LICM pass. It is only designed to hoist simple 14 // constructs that are not exposed before lowering and instruction selection. 15 // 16 //===----------------------------------------------------------------------===// 17 18 #include "llvm/ADT/BitVector.h" 19 #include "llvm/ADT/DenseMap.h" 20 #include "llvm/ADT/STLExtras.h" 21 #include "llvm/ADT/SmallSet.h" 22 #include "llvm/ADT/SmallVector.h" 23 #include "llvm/ADT/Statistic.h" 24 #include "llvm/Analysis/AliasAnalysis.h" 25 #include "llvm/CodeGen/MachineBasicBlock.h" 26 #include "llvm/CodeGen/MachineBlockFrequencyInfo.h" 27 #include "llvm/CodeGen/MachineDominators.h" 28 #include "llvm/CodeGen/MachineFrameInfo.h" 29 #include "llvm/CodeGen/MachineFunction.h" 30 #include "llvm/CodeGen/MachineFunctionPass.h" 31 #include "llvm/CodeGen/MachineInstr.h" 32 #include "llvm/CodeGen/MachineLoopInfo.h" 33 #include "llvm/CodeGen/MachineMemOperand.h" 34 #include "llvm/CodeGen/MachineOperand.h" 35 #include "llvm/CodeGen/MachineRegisterInfo.h" 36 #include "llvm/CodeGen/PseudoSourceValue.h" 37 #include "llvm/CodeGen/TargetInstrInfo.h" 38 #include "llvm/CodeGen/TargetLowering.h" 39 #include "llvm/CodeGen/TargetRegisterInfo.h" 40 #include "llvm/CodeGen/TargetSchedule.h" 41 #include "llvm/CodeGen/TargetSubtargetInfo.h" 42 #include "llvm/IR/DebugLoc.h" 43 #include "llvm/InitializePasses.h" 44 #include "llvm/MC/MCInstrDesc.h" 45 #include "llvm/MC/MCRegister.h" 46 #include "llvm/MC/MCRegisterInfo.h" 47 #include "llvm/Pass.h" 48 #include "llvm/Support/Casting.h" 49 #include "llvm/Support/CommandLine.h" 50 #include "llvm/Support/Debug.h" 51 #include "llvm/Support/raw_ostream.h" 52 #include <algorithm> 53 #include <cassert> 54 #include <limits> 55 #include <vector> 56 57 using namespace llvm; 58 59 #define DEBUG_TYPE "machinelicm" 60 61 static cl::opt<bool> 62 AvoidSpeculation("avoid-speculation", 63 cl::desc("MachineLICM should avoid speculation"), 64 cl::init(true), cl::Hidden); 65 66 static cl::opt<bool> 67 HoistCheapInsts("hoist-cheap-insts", 68 cl::desc("MachineLICM should hoist even cheap instructions"), 69 cl::init(false), cl::Hidden); 70 71 static cl::opt<bool> 72 HoistConstStores("hoist-const-stores", 73 cl::desc("Hoist invariant stores"), 74 cl::init(true), cl::Hidden); 75 // The default threshold of 100 (i.e. if target block is 100 times hotter) 76 // is based on empirical data on a single target and is subject to tuning. 77 static cl::opt<unsigned> 78 BlockFrequencyRatioThreshold("block-freq-ratio-threshold", 79 cl::desc("Do not hoist instructions if target" 80 "block is N times hotter than the source."), 81 cl::init(100), cl::Hidden); 82 83 enum class UseBFI { None, PGO, All }; 84 85 static cl::opt<UseBFI> 86 DisableHoistingToHotterBlocks("disable-hoisting-to-hotter-blocks", 87 cl::desc("Disable hoisting instructions to" 88 " hotter blocks"), 89 cl::init(UseBFI::PGO), cl::Hidden, 90 cl::values(clEnumValN(UseBFI::None, "none", 91 "disable the feature"), 92 clEnumValN(UseBFI::PGO, "pgo", 93 "enable the feature when using profile data"), 94 clEnumValN(UseBFI::All, "all", 95 "enable the feature with/wo profile data"))); 96 97 STATISTIC(NumHoisted, 98 "Number of machine instructions hoisted out of loops"); 99 STATISTIC(NumLowRP, 100 "Number of instructions hoisted in low reg pressure situation"); 101 STATISTIC(NumHighLatency, 102 "Number of high latency instructions hoisted"); 103 STATISTIC(NumCSEed, 104 "Number of hoisted machine instructions CSEed"); 105 STATISTIC(NumPostRAHoisted, 106 "Number of machine instructions hoisted out of loops post regalloc"); 107 STATISTIC(NumStoreConst, 108 "Number of stores of const phys reg hoisted out of loops"); 109 STATISTIC(NumNotHoistedDueToHotness, 110 "Number of instructions not hoisted due to block frequency"); 111 112 namespace { 113 114 class MachineLICMBase : public MachineFunctionPass { 115 const TargetInstrInfo *TII; 116 const TargetLoweringBase *TLI; 117 const TargetRegisterInfo *TRI; 118 const MachineFrameInfo *MFI; 119 MachineRegisterInfo *MRI; 120 TargetSchedModel SchedModel; 121 bool PreRegAlloc; 122 bool HasProfileData; 123 124 // Various analyses that we use... 125 AliasAnalysis *AA; // Alias analysis info. 126 MachineBlockFrequencyInfo *MBFI; // Machine block frequncy info 127 MachineLoopInfo *MLI; // Current MachineLoopInfo 128 MachineDominatorTree *DT; // Machine dominator tree for the cur loop 129 130 // State that is updated as we process loops 131 bool Changed; // True if a loop is changed. 132 bool FirstInLoop; // True if it's the first LICM in the loop. 133 MachineLoop *CurLoop; // The current loop we are working on. 134 MachineBasicBlock *CurPreheader; // The preheader for CurLoop. 135 136 // Exit blocks for CurLoop. 137 SmallVector<MachineBasicBlock *, 8> ExitBlocks; 138 139 bool isExitBlock(const MachineBasicBlock *MBB) const { 140 return is_contained(ExitBlocks, MBB); 141 } 142 143 // Track 'estimated' register pressure. 144 SmallSet<Register, 32> RegSeen; 145 SmallVector<unsigned, 8> RegPressure; 146 147 // Register pressure "limit" per register pressure set. If the pressure 148 // is higher than the limit, then it's considered high. 149 SmallVector<unsigned, 8> RegLimit; 150 151 // Register pressure on path leading from loop preheader to current BB. 152 SmallVector<SmallVector<unsigned, 8>, 16> BackTrace; 153 154 // For each opcode, keep a list of potential CSE instructions. 155 DenseMap<unsigned, std::vector<MachineInstr *>> CSEMap; 156 157 enum { 158 SpeculateFalse = 0, 159 SpeculateTrue = 1, 160 SpeculateUnknown = 2 161 }; 162 163 // If a MBB does not dominate loop exiting blocks then it may not safe 164 // to hoist loads from this block. 165 // Tri-state: 0 - false, 1 - true, 2 - unknown 166 unsigned SpeculationState; 167 168 public: 169 MachineLICMBase(char &PassID, bool PreRegAlloc) 170 : MachineFunctionPass(PassID), PreRegAlloc(PreRegAlloc) {} 171 172 bool runOnMachineFunction(MachineFunction &MF) override; 173 174 void getAnalysisUsage(AnalysisUsage &AU) const override { 175 AU.addRequired<MachineLoopInfo>(); 176 if (DisableHoistingToHotterBlocks != UseBFI::None) 177 AU.addRequired<MachineBlockFrequencyInfo>(); 178 AU.addRequired<MachineDominatorTree>(); 179 AU.addRequired<AAResultsWrapperPass>(); 180 AU.addPreserved<MachineLoopInfo>(); 181 MachineFunctionPass::getAnalysisUsage(AU); 182 } 183 184 void releaseMemory() override { 185 RegSeen.clear(); 186 RegPressure.clear(); 187 RegLimit.clear(); 188 BackTrace.clear(); 189 CSEMap.clear(); 190 } 191 192 private: 193 /// Keep track of information about hoisting candidates. 194 struct CandidateInfo { 195 MachineInstr *MI; 196 unsigned Def; 197 int FI; 198 199 CandidateInfo(MachineInstr *mi, unsigned def, int fi) 200 : MI(mi), Def(def), FI(fi) {} 201 }; 202 203 void HoistRegionPostRA(); 204 205 void HoistPostRA(MachineInstr *MI, unsigned Def); 206 207 void ProcessMI(MachineInstr *MI, BitVector &PhysRegDefs, 208 BitVector &PhysRegClobbers, SmallSet<int, 32> &StoredFIs, 209 SmallVectorImpl<CandidateInfo> &Candidates); 210 211 void AddToLiveIns(MCRegister Reg); 212 213 bool IsLICMCandidate(MachineInstr &I); 214 215 bool IsLoopInvariantInst(MachineInstr &I); 216 217 bool HasLoopPHIUse(const MachineInstr *MI) const; 218 219 bool HasHighOperandLatency(MachineInstr &MI, unsigned DefIdx, 220 Register Reg) const; 221 222 bool IsCheapInstruction(MachineInstr &MI) const; 223 224 bool CanCauseHighRegPressure(const DenseMap<unsigned, int> &Cost, 225 bool Cheap); 226 227 void UpdateBackTraceRegPressure(const MachineInstr *MI); 228 229 bool IsProfitableToHoist(MachineInstr &MI); 230 231 bool IsGuaranteedToExecute(MachineBasicBlock *BB); 232 233 bool isTriviallyReMaterializable(const MachineInstr &MI) const; 234 235 void EnterScope(MachineBasicBlock *MBB); 236 237 void ExitScope(MachineBasicBlock *MBB); 238 239 void ExitScopeIfDone( 240 MachineDomTreeNode *Node, 241 DenseMap<MachineDomTreeNode *, unsigned> &OpenChildren, 242 const DenseMap<MachineDomTreeNode *, MachineDomTreeNode *> &ParentMap); 243 244 void HoistOutOfLoop(MachineDomTreeNode *HeaderN); 245 246 void InitRegPressure(MachineBasicBlock *BB); 247 248 DenseMap<unsigned, int> calcRegisterCost(const MachineInstr *MI, 249 bool ConsiderSeen, 250 bool ConsiderUnseenAsDef); 251 252 void UpdateRegPressure(const MachineInstr *MI, 253 bool ConsiderUnseenAsDef = false); 254 255 MachineInstr *ExtractHoistableLoad(MachineInstr *MI); 256 257 MachineInstr *LookForDuplicate(const MachineInstr *MI, 258 std::vector<MachineInstr *> &PrevMIs); 259 260 bool 261 EliminateCSE(MachineInstr *MI, 262 DenseMap<unsigned, std::vector<MachineInstr *>>::iterator &CI); 263 264 bool MayCSE(MachineInstr *MI); 265 266 bool Hoist(MachineInstr *MI, MachineBasicBlock *Preheader); 267 268 void InitCSEMap(MachineBasicBlock *BB); 269 270 bool isTgtHotterThanSrc(MachineBasicBlock *SrcBlock, 271 MachineBasicBlock *TgtBlock); 272 MachineBasicBlock *getCurPreheader(); 273 }; 274 275 class MachineLICM : public MachineLICMBase { 276 public: 277 static char ID; 278 MachineLICM() : MachineLICMBase(ID, false) { 279 initializeMachineLICMPass(*PassRegistry::getPassRegistry()); 280 } 281 }; 282 283 class EarlyMachineLICM : public MachineLICMBase { 284 public: 285 static char ID; 286 EarlyMachineLICM() : MachineLICMBase(ID, true) { 287 initializeEarlyMachineLICMPass(*PassRegistry::getPassRegistry()); 288 } 289 }; 290 291 } // end anonymous namespace 292 293 char MachineLICM::ID; 294 char EarlyMachineLICM::ID; 295 296 char &llvm::MachineLICMID = MachineLICM::ID; 297 char &llvm::EarlyMachineLICMID = EarlyMachineLICM::ID; 298 299 INITIALIZE_PASS_BEGIN(MachineLICM, DEBUG_TYPE, 300 "Machine Loop Invariant Code Motion", false, false) 301 INITIALIZE_PASS_DEPENDENCY(MachineLoopInfo) 302 INITIALIZE_PASS_DEPENDENCY(MachineBlockFrequencyInfo) 303 INITIALIZE_PASS_DEPENDENCY(MachineDominatorTree) 304 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass) 305 INITIALIZE_PASS_END(MachineLICM, DEBUG_TYPE, 306 "Machine Loop Invariant Code Motion", false, false) 307 308 INITIALIZE_PASS_BEGIN(EarlyMachineLICM, "early-machinelicm", 309 "Early Machine Loop Invariant Code Motion", false, false) 310 INITIALIZE_PASS_DEPENDENCY(MachineLoopInfo) 311 INITIALIZE_PASS_DEPENDENCY(MachineBlockFrequencyInfo) 312 INITIALIZE_PASS_DEPENDENCY(MachineDominatorTree) 313 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass) 314 INITIALIZE_PASS_END(EarlyMachineLICM, "early-machinelicm", 315 "Early Machine Loop Invariant Code Motion", false, false) 316 317 /// Test if the given loop is the outer-most loop that has a unique predecessor. 318 static bool LoopIsOuterMostWithPredecessor(MachineLoop *CurLoop) { 319 // Check whether this loop even has a unique predecessor. 320 if (!CurLoop->getLoopPredecessor()) 321 return false; 322 // Ok, now check to see if any of its outer loops do. 323 for (MachineLoop *L = CurLoop->getParentLoop(); L; L = L->getParentLoop()) 324 if (L->getLoopPredecessor()) 325 return false; 326 // None of them did, so this is the outermost with a unique predecessor. 327 return true; 328 } 329 330 bool MachineLICMBase::runOnMachineFunction(MachineFunction &MF) { 331 if (skipFunction(MF.getFunction())) 332 return false; 333 334 Changed = FirstInLoop = false; 335 const TargetSubtargetInfo &ST = MF.getSubtarget(); 336 TII = ST.getInstrInfo(); 337 TLI = ST.getTargetLowering(); 338 TRI = ST.getRegisterInfo(); 339 MFI = &MF.getFrameInfo(); 340 MRI = &MF.getRegInfo(); 341 SchedModel.init(&ST); 342 343 PreRegAlloc = MRI->isSSA(); 344 HasProfileData = MF.getFunction().hasProfileData(); 345 346 if (PreRegAlloc) 347 LLVM_DEBUG(dbgs() << "******** Pre-regalloc Machine LICM: "); 348 else 349 LLVM_DEBUG(dbgs() << "******** Post-regalloc Machine LICM: "); 350 LLVM_DEBUG(dbgs() << MF.getName() << " ********\n"); 351 352 if (PreRegAlloc) { 353 // Estimate register pressure during pre-regalloc pass. 354 unsigned NumRPS = TRI->getNumRegPressureSets(); 355 RegPressure.resize(NumRPS); 356 std::fill(RegPressure.begin(), RegPressure.end(), 0); 357 RegLimit.resize(NumRPS); 358 for (unsigned i = 0, e = NumRPS; i != e; ++i) 359 RegLimit[i] = TRI->getRegPressureSetLimit(MF, i); 360 } 361 362 // Get our Loop information... 363 if (DisableHoistingToHotterBlocks != UseBFI::None) 364 MBFI = &getAnalysis<MachineBlockFrequencyInfo>(); 365 MLI = &getAnalysis<MachineLoopInfo>(); 366 DT = &getAnalysis<MachineDominatorTree>(); 367 AA = &getAnalysis<AAResultsWrapperPass>().getAAResults(); 368 369 SmallVector<MachineLoop *, 8> Worklist(MLI->begin(), MLI->end()); 370 while (!Worklist.empty()) { 371 CurLoop = Worklist.pop_back_val(); 372 CurPreheader = nullptr; 373 ExitBlocks.clear(); 374 375 // If this is done before regalloc, only visit outer-most preheader-sporting 376 // loops. 377 if (PreRegAlloc && !LoopIsOuterMostWithPredecessor(CurLoop)) { 378 Worklist.append(CurLoop->begin(), CurLoop->end()); 379 continue; 380 } 381 382 CurLoop->getExitBlocks(ExitBlocks); 383 384 if (!PreRegAlloc) 385 HoistRegionPostRA(); 386 else { 387 // CSEMap is initialized for loop header when the first instruction is 388 // being hoisted. 389 MachineDomTreeNode *N = DT->getNode(CurLoop->getHeader()); 390 FirstInLoop = true; 391 HoistOutOfLoop(N); 392 CSEMap.clear(); 393 } 394 } 395 396 return Changed; 397 } 398 399 /// Return true if instruction stores to the specified frame. 400 static bool InstructionStoresToFI(const MachineInstr *MI, int FI) { 401 // Check mayStore before memory operands so that e.g. DBG_VALUEs will return 402 // true since they have no memory operands. 403 if (!MI->mayStore()) 404 return false; 405 // If we lost memory operands, conservatively assume that the instruction 406 // writes to all slots. 407 if (MI->memoperands_empty()) 408 return true; 409 for (const MachineMemOperand *MemOp : MI->memoperands()) { 410 if (!MemOp->isStore() || !MemOp->getPseudoValue()) 411 continue; 412 if (const FixedStackPseudoSourceValue *Value = 413 dyn_cast<FixedStackPseudoSourceValue>(MemOp->getPseudoValue())) { 414 if (Value->getFrameIndex() == FI) 415 return true; 416 } 417 } 418 return false; 419 } 420 421 /// Examine the instruction for potentai LICM candidate. Also 422 /// gather register def and frame object update information. 423 void MachineLICMBase::ProcessMI(MachineInstr *MI, 424 BitVector &PhysRegDefs, 425 BitVector &PhysRegClobbers, 426 SmallSet<int, 32> &StoredFIs, 427 SmallVectorImpl<CandidateInfo> &Candidates) { 428 bool RuledOut = false; 429 bool HasNonInvariantUse = false; 430 unsigned Def = 0; 431 for (const MachineOperand &MO : MI->operands()) { 432 if (MO.isFI()) { 433 // Remember if the instruction stores to the frame index. 434 int FI = MO.getIndex(); 435 if (!StoredFIs.count(FI) && 436 MFI->isSpillSlotObjectIndex(FI) && 437 InstructionStoresToFI(MI, FI)) 438 StoredFIs.insert(FI); 439 HasNonInvariantUse = true; 440 continue; 441 } 442 443 // We can't hoist an instruction defining a physreg that is clobbered in 444 // the loop. 445 if (MO.isRegMask()) { 446 PhysRegClobbers.setBitsNotInMask(MO.getRegMask()); 447 continue; 448 } 449 450 if (!MO.isReg()) 451 continue; 452 Register Reg = MO.getReg(); 453 if (!Reg) 454 continue; 455 assert(Register::isPhysicalRegister(Reg) && 456 "Not expecting virtual register!"); 457 458 if (!MO.isDef()) { 459 if (Reg && (PhysRegDefs.test(Reg) || PhysRegClobbers.test(Reg))) 460 // If it's using a non-loop-invariant register, then it's obviously not 461 // safe to hoist. 462 HasNonInvariantUse = true; 463 continue; 464 } 465 466 if (MO.isImplicit()) { 467 for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI) 468 PhysRegClobbers.set(*AI); 469 if (!MO.isDead()) 470 // Non-dead implicit def? This cannot be hoisted. 471 RuledOut = true; 472 // No need to check if a dead implicit def is also defined by 473 // another instruction. 474 continue; 475 } 476 477 // FIXME: For now, avoid instructions with multiple defs, unless 478 // it's a dead implicit def. 479 if (Def) 480 RuledOut = true; 481 else 482 Def = Reg; 483 484 // If we have already seen another instruction that defines the same 485 // register, then this is not safe. Two defs is indicated by setting a 486 // PhysRegClobbers bit. 487 for (MCRegAliasIterator AS(Reg, TRI, true); AS.isValid(); ++AS) { 488 if (PhysRegDefs.test(*AS)) 489 PhysRegClobbers.set(*AS); 490 } 491 // Need a second loop because MCRegAliasIterator can visit the same 492 // register twice. 493 for (MCRegAliasIterator AS(Reg, TRI, true); AS.isValid(); ++AS) 494 PhysRegDefs.set(*AS); 495 496 if (PhysRegClobbers.test(Reg)) 497 // MI defined register is seen defined by another instruction in 498 // the loop, it cannot be a LICM candidate. 499 RuledOut = true; 500 } 501 502 // Only consider reloads for now and remats which do not have register 503 // operands. FIXME: Consider unfold load folding instructions. 504 if (Def && !RuledOut) { 505 int FI = std::numeric_limits<int>::min(); 506 if ((!HasNonInvariantUse && IsLICMCandidate(*MI)) || 507 (TII->isLoadFromStackSlot(*MI, FI) && MFI->isSpillSlotObjectIndex(FI))) 508 Candidates.push_back(CandidateInfo(MI, Def, FI)); 509 } 510 } 511 512 /// Walk the specified region of the CFG and hoist loop invariants out to the 513 /// preheader. 514 void MachineLICMBase::HoistRegionPostRA() { 515 MachineBasicBlock *Preheader = getCurPreheader(); 516 if (!Preheader) 517 return; 518 519 unsigned NumRegs = TRI->getNumRegs(); 520 BitVector PhysRegDefs(NumRegs); // Regs defined once in the loop. 521 BitVector PhysRegClobbers(NumRegs); // Regs defined more than once. 522 523 SmallVector<CandidateInfo, 32> Candidates; 524 SmallSet<int, 32> StoredFIs; 525 526 // Walk the entire region, count number of defs for each register, and 527 // collect potential LICM candidates. 528 for (MachineBasicBlock *BB : CurLoop->getBlocks()) { 529 // If the header of the loop containing this basic block is a landing pad, 530 // then don't try to hoist instructions out of this loop. 531 const MachineLoop *ML = MLI->getLoopFor(BB); 532 if (ML && ML->getHeader()->isEHPad()) continue; 533 534 // Conservatively treat live-in's as an external def. 535 // FIXME: That means a reload that're reused in successor block(s) will not 536 // be LICM'ed. 537 for (const auto &LI : BB->liveins()) { 538 for (MCRegAliasIterator AI(LI.PhysReg, TRI, true); AI.isValid(); ++AI) 539 PhysRegDefs.set(*AI); 540 } 541 542 SpeculationState = SpeculateUnknown; 543 for (MachineInstr &MI : *BB) 544 ProcessMI(&MI, PhysRegDefs, PhysRegClobbers, StoredFIs, Candidates); 545 } 546 547 // Gather the registers read / clobbered by the terminator. 548 BitVector TermRegs(NumRegs); 549 MachineBasicBlock::iterator TI = Preheader->getFirstTerminator(); 550 if (TI != Preheader->end()) { 551 for (const MachineOperand &MO : TI->operands()) { 552 if (!MO.isReg()) 553 continue; 554 Register Reg = MO.getReg(); 555 if (!Reg) 556 continue; 557 for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI) 558 TermRegs.set(*AI); 559 } 560 } 561 562 // Now evaluate whether the potential candidates qualify. 563 // 1. Check if the candidate defined register is defined by another 564 // instruction in the loop. 565 // 2. If the candidate is a load from stack slot (always true for now), 566 // check if the slot is stored anywhere in the loop. 567 // 3. Make sure candidate def should not clobber 568 // registers read by the terminator. Similarly its def should not be 569 // clobbered by the terminator. 570 for (CandidateInfo &Candidate : Candidates) { 571 if (Candidate.FI != std::numeric_limits<int>::min() && 572 StoredFIs.count(Candidate.FI)) 573 continue; 574 575 unsigned Def = Candidate.Def; 576 if (!PhysRegClobbers.test(Def) && !TermRegs.test(Def)) { 577 bool Safe = true; 578 MachineInstr *MI = Candidate.MI; 579 for (const MachineOperand &MO : MI->operands()) { 580 if (!MO.isReg() || MO.isDef() || !MO.getReg()) 581 continue; 582 Register Reg = MO.getReg(); 583 if (PhysRegDefs.test(Reg) || 584 PhysRegClobbers.test(Reg)) { 585 // If it's using a non-loop-invariant register, then it's obviously 586 // not safe to hoist. 587 Safe = false; 588 break; 589 } 590 } 591 if (Safe) 592 HoistPostRA(MI, Candidate.Def); 593 } 594 } 595 } 596 597 /// Add register 'Reg' to the livein sets of BBs in the current loop, and make 598 /// sure it is not killed by any instructions in the loop. 599 void MachineLICMBase::AddToLiveIns(MCRegister Reg) { 600 for (MachineBasicBlock *BB : CurLoop->getBlocks()) { 601 if (!BB->isLiveIn(Reg)) 602 BB->addLiveIn(Reg); 603 for (MachineInstr &MI : *BB) { 604 for (MachineOperand &MO : MI.operands()) { 605 if (!MO.isReg() || !MO.getReg() || MO.isDef()) continue; 606 if (MO.getReg() == Reg || TRI->isSuperRegister(Reg, MO.getReg())) 607 MO.setIsKill(false); 608 } 609 } 610 } 611 } 612 613 /// When an instruction is found to only use loop invariant operands that is 614 /// safe to hoist, this instruction is called to do the dirty work. 615 void MachineLICMBase::HoistPostRA(MachineInstr *MI, unsigned Def) { 616 MachineBasicBlock *Preheader = getCurPreheader(); 617 618 // Now move the instructions to the predecessor, inserting it before any 619 // terminator instructions. 620 LLVM_DEBUG(dbgs() << "Hoisting to " << printMBBReference(*Preheader) 621 << " from " << printMBBReference(*MI->getParent()) << ": " 622 << *MI); 623 624 // Splice the instruction to the preheader. 625 MachineBasicBlock *MBB = MI->getParent(); 626 Preheader->splice(Preheader->getFirstTerminator(), MBB, MI); 627 628 // Since we are moving the instruction out of its basic block, we do not 629 // retain its debug location. Doing so would degrade the debugging 630 // experience and adversely affect the accuracy of profiling information. 631 assert(!MI->isDebugInstr() && "Should not hoist debug inst"); 632 MI->setDebugLoc(DebugLoc()); 633 634 // Add register to livein list to all the BBs in the current loop since a 635 // loop invariant must be kept live throughout the whole loop. This is 636 // important to ensure later passes do not scavenge the def register. 637 AddToLiveIns(Def); 638 639 ++NumPostRAHoisted; 640 Changed = true; 641 } 642 643 /// Check if this mbb is guaranteed to execute. If not then a load from this mbb 644 /// may not be safe to hoist. 645 bool MachineLICMBase::IsGuaranteedToExecute(MachineBasicBlock *BB) { 646 if (SpeculationState != SpeculateUnknown) 647 return SpeculationState == SpeculateFalse; 648 649 if (BB != CurLoop->getHeader()) { 650 // Check loop exiting blocks. 651 SmallVector<MachineBasicBlock*, 8> CurrentLoopExitingBlocks; 652 CurLoop->getExitingBlocks(CurrentLoopExitingBlocks); 653 for (MachineBasicBlock *CurrentLoopExitingBlock : CurrentLoopExitingBlocks) 654 if (!DT->dominates(BB, CurrentLoopExitingBlock)) { 655 SpeculationState = SpeculateTrue; 656 return false; 657 } 658 } 659 660 SpeculationState = SpeculateFalse; 661 return true; 662 } 663 664 /// Check if \p MI is trivially remateralizable and if it does not have any 665 /// virtual register uses. Even though rematerializable RA might not actually 666 /// rematerialize it in this scenario. In that case we do not want to hoist such 667 /// instruction out of the loop in a belief RA will sink it back if needed. 668 bool MachineLICMBase::isTriviallyReMaterializable( 669 const MachineInstr &MI) const { 670 if (!TII->isTriviallyReMaterializable(MI)) 671 return false; 672 673 for (const MachineOperand &MO : MI.operands()) { 674 if (MO.isReg() && MO.isUse() && MO.getReg().isVirtual()) 675 return false; 676 } 677 678 return true; 679 } 680 681 void MachineLICMBase::EnterScope(MachineBasicBlock *MBB) { 682 LLVM_DEBUG(dbgs() << "Entering " << printMBBReference(*MBB) << '\n'); 683 684 // Remember livein register pressure. 685 BackTrace.push_back(RegPressure); 686 } 687 688 void MachineLICMBase::ExitScope(MachineBasicBlock *MBB) { 689 LLVM_DEBUG(dbgs() << "Exiting " << printMBBReference(*MBB) << '\n'); 690 BackTrace.pop_back(); 691 } 692 693 /// Destroy scope for the MBB that corresponds to the given dominator tree node 694 /// if its a leaf or all of its children are done. Walk up the dominator tree to 695 /// destroy ancestors which are now done. 696 void MachineLICMBase::ExitScopeIfDone(MachineDomTreeNode *Node, 697 DenseMap<MachineDomTreeNode*, unsigned> &OpenChildren, 698 const DenseMap<MachineDomTreeNode*, MachineDomTreeNode*> &ParentMap) { 699 if (OpenChildren[Node]) 700 return; 701 702 for(;;) { 703 ExitScope(Node->getBlock()); 704 // Now traverse upwards to pop ancestors whose offsprings are all done. 705 MachineDomTreeNode *Parent = ParentMap.lookup(Node); 706 if (!Parent || --OpenChildren[Parent] != 0) 707 break; 708 Node = Parent; 709 } 710 } 711 712 /// Walk the specified loop in the CFG (defined by all blocks dominated by the 713 /// specified header block, and that are in the current loop) in depth first 714 /// order w.r.t the DominatorTree. This allows us to visit definitions before 715 /// uses, allowing us to hoist a loop body in one pass without iteration. 716 void MachineLICMBase::HoistOutOfLoop(MachineDomTreeNode *HeaderN) { 717 MachineBasicBlock *Preheader = getCurPreheader(); 718 if (!Preheader) 719 return; 720 721 SmallVector<MachineDomTreeNode*, 32> Scopes; 722 SmallVector<MachineDomTreeNode*, 8> WorkList; 723 DenseMap<MachineDomTreeNode*, MachineDomTreeNode*> ParentMap; 724 DenseMap<MachineDomTreeNode*, unsigned> OpenChildren; 725 726 // Perform a DFS walk to determine the order of visit. 727 WorkList.push_back(HeaderN); 728 while (!WorkList.empty()) { 729 MachineDomTreeNode *Node = WorkList.pop_back_val(); 730 assert(Node && "Null dominator tree node?"); 731 MachineBasicBlock *BB = Node->getBlock(); 732 733 // If the header of the loop containing this basic block is a landing pad, 734 // then don't try to hoist instructions out of this loop. 735 const MachineLoop *ML = MLI->getLoopFor(BB); 736 if (ML && ML->getHeader()->isEHPad()) 737 continue; 738 739 // If this subregion is not in the top level loop at all, exit. 740 if (!CurLoop->contains(BB)) 741 continue; 742 743 Scopes.push_back(Node); 744 unsigned NumChildren = Node->getNumChildren(); 745 746 // Don't hoist things out of a large switch statement. This often causes 747 // code to be hoisted that wasn't going to be executed, and increases 748 // register pressure in a situation where it's likely to matter. 749 if (BB->succ_size() >= 25) 750 NumChildren = 0; 751 752 OpenChildren[Node] = NumChildren; 753 if (NumChildren) { 754 // Add children in reverse order as then the next popped worklist node is 755 // the first child of this node. This means we ultimately traverse the 756 // DOM tree in exactly the same order as if we'd recursed. 757 for (MachineDomTreeNode *Child : reverse(Node->children())) { 758 ParentMap[Child] = Node; 759 WorkList.push_back(Child); 760 } 761 } 762 } 763 764 if (Scopes.size() == 0) 765 return; 766 767 // Compute registers which are livein into the loop headers. 768 RegSeen.clear(); 769 BackTrace.clear(); 770 InitRegPressure(Preheader); 771 772 // Now perform LICM. 773 for (MachineDomTreeNode *Node : Scopes) { 774 MachineBasicBlock *MBB = Node->getBlock(); 775 776 EnterScope(MBB); 777 778 // Process the block 779 SpeculationState = SpeculateUnknown; 780 for (MachineInstr &MI : llvm::make_early_inc_range(*MBB)) { 781 if (!Hoist(&MI, Preheader)) 782 UpdateRegPressure(&MI); 783 // If we have hoisted an instruction that may store, it can only be a 784 // constant store. 785 } 786 787 // If it's a leaf node, it's done. Traverse upwards to pop ancestors. 788 ExitScopeIfDone(Node, OpenChildren, ParentMap); 789 } 790 } 791 792 static bool isOperandKill(const MachineOperand &MO, MachineRegisterInfo *MRI) { 793 return MO.isKill() || MRI->hasOneNonDBGUse(MO.getReg()); 794 } 795 796 /// Find all virtual register references that are liveout of the preheader to 797 /// initialize the starting "register pressure". Note this does not count live 798 /// through (livein but not used) registers. 799 void MachineLICMBase::InitRegPressure(MachineBasicBlock *BB) { 800 std::fill(RegPressure.begin(), RegPressure.end(), 0); 801 802 // If the preheader has only a single predecessor and it ends with a 803 // fallthrough or an unconditional branch, then scan its predecessor for live 804 // defs as well. This happens whenever the preheader is created by splitting 805 // the critical edge from the loop predecessor to the loop header. 806 if (BB->pred_size() == 1) { 807 MachineBasicBlock *TBB = nullptr, *FBB = nullptr; 808 SmallVector<MachineOperand, 4> Cond; 809 if (!TII->analyzeBranch(*BB, TBB, FBB, Cond, false) && Cond.empty()) 810 InitRegPressure(*BB->pred_begin()); 811 } 812 813 for (const MachineInstr &MI : *BB) 814 UpdateRegPressure(&MI, /*ConsiderUnseenAsDef=*/true); 815 } 816 817 /// Update estimate of register pressure after the specified instruction. 818 void MachineLICMBase::UpdateRegPressure(const MachineInstr *MI, 819 bool ConsiderUnseenAsDef) { 820 auto Cost = calcRegisterCost(MI, /*ConsiderSeen=*/true, ConsiderUnseenAsDef); 821 for (const auto &RPIdAndCost : Cost) { 822 unsigned Class = RPIdAndCost.first; 823 if (static_cast<int>(RegPressure[Class]) < -RPIdAndCost.second) 824 RegPressure[Class] = 0; 825 else 826 RegPressure[Class] += RPIdAndCost.second; 827 } 828 } 829 830 /// Calculate the additional register pressure that the registers used in MI 831 /// cause. 832 /// 833 /// If 'ConsiderSeen' is true, updates 'RegSeen' and uses the information to 834 /// figure out which usages are live-ins. 835 /// FIXME: Figure out a way to consider 'RegSeen' from all code paths. 836 DenseMap<unsigned, int> 837 MachineLICMBase::calcRegisterCost(const MachineInstr *MI, bool ConsiderSeen, 838 bool ConsiderUnseenAsDef) { 839 DenseMap<unsigned, int> Cost; 840 if (MI->isImplicitDef()) 841 return Cost; 842 for (unsigned i = 0, e = MI->getDesc().getNumOperands(); i != e; ++i) { 843 const MachineOperand &MO = MI->getOperand(i); 844 if (!MO.isReg() || MO.isImplicit()) 845 continue; 846 Register Reg = MO.getReg(); 847 if (!Register::isVirtualRegister(Reg)) 848 continue; 849 850 // FIXME: It seems bad to use RegSeen only for some of these calculations. 851 bool isNew = ConsiderSeen ? RegSeen.insert(Reg).second : false; 852 const TargetRegisterClass *RC = MRI->getRegClass(Reg); 853 854 RegClassWeight W = TRI->getRegClassWeight(RC); 855 int RCCost = 0; 856 if (MO.isDef()) 857 RCCost = W.RegWeight; 858 else { 859 bool isKill = isOperandKill(MO, MRI); 860 if (isNew && !isKill && ConsiderUnseenAsDef) 861 // Haven't seen this, it must be a livein. 862 RCCost = W.RegWeight; 863 else if (!isNew && isKill) 864 RCCost = -W.RegWeight; 865 } 866 if (RCCost == 0) 867 continue; 868 const int *PS = TRI->getRegClassPressureSets(RC); 869 for (; *PS != -1; ++PS) { 870 if (Cost.find(*PS) == Cost.end()) 871 Cost[*PS] = RCCost; 872 else 873 Cost[*PS] += RCCost; 874 } 875 } 876 return Cost; 877 } 878 879 /// Return true if this machine instruction loads from global offset table or 880 /// constant pool. 881 static bool mayLoadFromGOTOrConstantPool(MachineInstr &MI) { 882 assert(MI.mayLoad() && "Expected MI that loads!"); 883 884 // If we lost memory operands, conservatively assume that the instruction 885 // reads from everything.. 886 if (MI.memoperands_empty()) 887 return true; 888 889 for (MachineMemOperand *MemOp : MI.memoperands()) 890 if (const PseudoSourceValue *PSV = MemOp->getPseudoValue()) 891 if (PSV->isGOT() || PSV->isConstantPool()) 892 return true; 893 894 return false; 895 } 896 897 // This function iterates through all the operands of the input store MI and 898 // checks that each register operand statisfies isCallerPreservedPhysReg. 899 // This means, the value being stored and the address where it is being stored 900 // is constant throughout the body of the function (not including prologue and 901 // epilogue). When called with an MI that isn't a store, it returns false. 902 // A future improvement can be to check if the store registers are constant 903 // throughout the loop rather than throughout the funtion. 904 static bool isInvariantStore(const MachineInstr &MI, 905 const TargetRegisterInfo *TRI, 906 const MachineRegisterInfo *MRI) { 907 908 bool FoundCallerPresReg = false; 909 if (!MI.mayStore() || MI.hasUnmodeledSideEffects() || 910 (MI.getNumOperands() == 0)) 911 return false; 912 913 // Check that all register operands are caller-preserved physical registers. 914 for (const MachineOperand &MO : MI.operands()) { 915 if (MO.isReg()) { 916 Register Reg = MO.getReg(); 917 // If operand is a virtual register, check if it comes from a copy of a 918 // physical register. 919 if (Register::isVirtualRegister(Reg)) 920 Reg = TRI->lookThruCopyLike(MO.getReg(), MRI); 921 if (Register::isVirtualRegister(Reg)) 922 return false; 923 if (!TRI->isCallerPreservedPhysReg(Reg.asMCReg(), *MI.getMF())) 924 return false; 925 else 926 FoundCallerPresReg = true; 927 } else if (!MO.isImm()) { 928 return false; 929 } 930 } 931 return FoundCallerPresReg; 932 } 933 934 // Return true if the input MI is a copy instruction that feeds an invariant 935 // store instruction. This means that the src of the copy has to satisfy 936 // isCallerPreservedPhysReg and atleast one of it's users should satisfy 937 // isInvariantStore. 938 static bool isCopyFeedingInvariantStore(const MachineInstr &MI, 939 const MachineRegisterInfo *MRI, 940 const TargetRegisterInfo *TRI) { 941 942 // FIXME: If targets would like to look through instructions that aren't 943 // pure copies, this can be updated to a query. 944 if (!MI.isCopy()) 945 return false; 946 947 const MachineFunction *MF = MI.getMF(); 948 // Check that we are copying a constant physical register. 949 Register CopySrcReg = MI.getOperand(1).getReg(); 950 if (Register::isVirtualRegister(CopySrcReg)) 951 return false; 952 953 if (!TRI->isCallerPreservedPhysReg(CopySrcReg.asMCReg(), *MF)) 954 return false; 955 956 Register CopyDstReg = MI.getOperand(0).getReg(); 957 // Check if any of the uses of the copy are invariant stores. 958 assert(Register::isVirtualRegister(CopyDstReg) && 959 "copy dst is not a virtual reg"); 960 961 for (MachineInstr &UseMI : MRI->use_instructions(CopyDstReg)) { 962 if (UseMI.mayStore() && isInvariantStore(UseMI, TRI, MRI)) 963 return true; 964 } 965 return false; 966 } 967 968 /// Returns true if the instruction may be a suitable candidate for LICM. 969 /// e.g. If the instruction is a call, then it's obviously not safe to hoist it. 970 bool MachineLICMBase::IsLICMCandidate(MachineInstr &I) { 971 // Check if it's safe to move the instruction. 972 bool DontMoveAcrossStore = true; 973 if ((!I.isSafeToMove(AA, DontMoveAcrossStore)) && 974 !(HoistConstStores && isInvariantStore(I, TRI, MRI))) { 975 LLVM_DEBUG(dbgs() << "LICM: Instruction not safe to move.\n"); 976 return false; 977 } 978 979 // If it is a load then check if it is guaranteed to execute by making sure 980 // that it dominates all exiting blocks. If it doesn't, then there is a path 981 // out of the loop which does not execute this load, so we can't hoist it. 982 // Loads from constant memory are safe to speculate, for example indexed load 983 // from a jump table. 984 // Stores and side effects are already checked by isSafeToMove. 985 if (I.mayLoad() && !mayLoadFromGOTOrConstantPool(I) && 986 !IsGuaranteedToExecute(I.getParent())) { 987 LLVM_DEBUG(dbgs() << "LICM: Load not guaranteed to execute.\n"); 988 return false; 989 } 990 991 // Convergent attribute has been used on operations that involve inter-thread 992 // communication which results are implicitly affected by the enclosing 993 // control flows. It is not safe to hoist or sink such operations across 994 // control flow. 995 if (I.isConvergent()) 996 return false; 997 998 if (!TII->shouldHoist(I, CurLoop)) 999 return false; 1000 1001 return true; 1002 } 1003 1004 /// Returns true if the instruction is loop invariant. 1005 bool MachineLICMBase::IsLoopInvariantInst(MachineInstr &I) { 1006 if (!IsLICMCandidate(I)) { 1007 LLVM_DEBUG(dbgs() << "LICM: Instruction not a LICM candidate\n"); 1008 return false; 1009 } 1010 return CurLoop->isLoopInvariant(I); 1011 } 1012 1013 /// Return true if the specified instruction is used by a phi node and hoisting 1014 /// it could cause a copy to be inserted. 1015 bool MachineLICMBase::HasLoopPHIUse(const MachineInstr *MI) const { 1016 SmallVector<const MachineInstr*, 8> Work(1, MI); 1017 do { 1018 MI = Work.pop_back_val(); 1019 for (const MachineOperand &MO : MI->operands()) { 1020 if (!MO.isReg() || !MO.isDef()) 1021 continue; 1022 Register Reg = MO.getReg(); 1023 if (!Register::isVirtualRegister(Reg)) 1024 continue; 1025 for (MachineInstr &UseMI : MRI->use_instructions(Reg)) { 1026 // A PHI may cause a copy to be inserted. 1027 if (UseMI.isPHI()) { 1028 // A PHI inside the loop causes a copy because the live range of Reg is 1029 // extended across the PHI. 1030 if (CurLoop->contains(&UseMI)) 1031 return true; 1032 // A PHI in an exit block can cause a copy to be inserted if the PHI 1033 // has multiple predecessors in the loop with different values. 1034 // For now, approximate by rejecting all exit blocks. 1035 if (isExitBlock(UseMI.getParent())) 1036 return true; 1037 continue; 1038 } 1039 // Look past copies as well. 1040 if (UseMI.isCopy() && CurLoop->contains(&UseMI)) 1041 Work.push_back(&UseMI); 1042 } 1043 } 1044 } while (!Work.empty()); 1045 return false; 1046 } 1047 1048 /// Compute operand latency between a def of 'Reg' and an use in the current 1049 /// loop, return true if the target considered it high. 1050 bool MachineLICMBase::HasHighOperandLatency(MachineInstr &MI, unsigned DefIdx, 1051 Register Reg) const { 1052 if (MRI->use_nodbg_empty(Reg)) 1053 return false; 1054 1055 for (MachineInstr &UseMI : MRI->use_nodbg_instructions(Reg)) { 1056 if (UseMI.isCopyLike()) 1057 continue; 1058 if (!CurLoop->contains(UseMI.getParent())) 1059 continue; 1060 for (unsigned i = 0, e = UseMI.getNumOperands(); i != e; ++i) { 1061 const MachineOperand &MO = UseMI.getOperand(i); 1062 if (!MO.isReg() || !MO.isUse()) 1063 continue; 1064 Register MOReg = MO.getReg(); 1065 if (MOReg != Reg) 1066 continue; 1067 1068 if (TII->hasHighOperandLatency(SchedModel, MRI, MI, DefIdx, UseMI, i)) 1069 return true; 1070 } 1071 1072 // Only look at the first in loop use. 1073 break; 1074 } 1075 1076 return false; 1077 } 1078 1079 /// Return true if the instruction is marked "cheap" or the operand latency 1080 /// between its def and a use is one or less. 1081 bool MachineLICMBase::IsCheapInstruction(MachineInstr &MI) const { 1082 if (TII->isAsCheapAsAMove(MI) || MI.isCopyLike()) 1083 return true; 1084 1085 bool isCheap = false; 1086 unsigned NumDefs = MI.getDesc().getNumDefs(); 1087 for (unsigned i = 0, e = MI.getNumOperands(); NumDefs && i != e; ++i) { 1088 MachineOperand &DefMO = MI.getOperand(i); 1089 if (!DefMO.isReg() || !DefMO.isDef()) 1090 continue; 1091 --NumDefs; 1092 Register Reg = DefMO.getReg(); 1093 if (Register::isPhysicalRegister(Reg)) 1094 continue; 1095 1096 if (!TII->hasLowDefLatency(SchedModel, MI, i)) 1097 return false; 1098 isCheap = true; 1099 } 1100 1101 return isCheap; 1102 } 1103 1104 /// Visit BBs from header to current BB, check if hoisting an instruction of the 1105 /// given cost matrix can cause high register pressure. 1106 bool 1107 MachineLICMBase::CanCauseHighRegPressure(const DenseMap<unsigned, int>& Cost, 1108 bool CheapInstr) { 1109 for (const auto &RPIdAndCost : Cost) { 1110 if (RPIdAndCost.second <= 0) 1111 continue; 1112 1113 unsigned Class = RPIdAndCost.first; 1114 int Limit = RegLimit[Class]; 1115 1116 // Don't hoist cheap instructions if they would increase register pressure, 1117 // even if we're under the limit. 1118 if (CheapInstr && !HoistCheapInsts) 1119 return true; 1120 1121 for (const auto &RP : BackTrace) 1122 if (static_cast<int>(RP[Class]) + RPIdAndCost.second >= Limit) 1123 return true; 1124 } 1125 1126 return false; 1127 } 1128 1129 /// Traverse the back trace from header to the current block and update their 1130 /// register pressures to reflect the effect of hoisting MI from the current 1131 /// block to the preheader. 1132 void MachineLICMBase::UpdateBackTraceRegPressure(const MachineInstr *MI) { 1133 // First compute the 'cost' of the instruction, i.e. its contribution 1134 // to register pressure. 1135 auto Cost = calcRegisterCost(MI, /*ConsiderSeen=*/false, 1136 /*ConsiderUnseenAsDef=*/false); 1137 1138 // Update register pressure of blocks from loop header to current block. 1139 for (auto &RP : BackTrace) 1140 for (const auto &RPIdAndCost : Cost) 1141 RP[RPIdAndCost.first] += RPIdAndCost.second; 1142 } 1143 1144 /// Return true if it is potentially profitable to hoist the given loop 1145 /// invariant. 1146 bool MachineLICMBase::IsProfitableToHoist(MachineInstr &MI) { 1147 if (MI.isImplicitDef()) 1148 return true; 1149 1150 // Besides removing computation from the loop, hoisting an instruction has 1151 // these effects: 1152 // 1153 // - The value defined by the instruction becomes live across the entire 1154 // loop. This increases register pressure in the loop. 1155 // 1156 // - If the value is used by a PHI in the loop, a copy will be required for 1157 // lowering the PHI after extending the live range. 1158 // 1159 // - When hoisting the last use of a value in the loop, that value no longer 1160 // needs to be live in the loop. This lowers register pressure in the loop. 1161 1162 if (HoistConstStores && isCopyFeedingInvariantStore(MI, MRI, TRI)) 1163 return true; 1164 1165 bool CheapInstr = IsCheapInstruction(MI); 1166 bool CreatesCopy = HasLoopPHIUse(&MI); 1167 1168 // Don't hoist a cheap instruction if it would create a copy in the loop. 1169 if (CheapInstr && CreatesCopy) { 1170 LLVM_DEBUG(dbgs() << "Won't hoist cheap instr with loop PHI use: " << MI); 1171 return false; 1172 } 1173 1174 // Rematerializable instructions should always be hoisted providing the 1175 // register allocator can just pull them down again when needed. 1176 if (isTriviallyReMaterializable(MI)) 1177 return true; 1178 1179 // FIXME: If there are long latency loop-invariant instructions inside the 1180 // loop at this point, why didn't the optimizer's LICM hoist them? 1181 for (unsigned i = 0, e = MI.getDesc().getNumOperands(); i != e; ++i) { 1182 const MachineOperand &MO = MI.getOperand(i); 1183 if (!MO.isReg() || MO.isImplicit()) 1184 continue; 1185 Register Reg = MO.getReg(); 1186 if (!Register::isVirtualRegister(Reg)) 1187 continue; 1188 if (MO.isDef() && HasHighOperandLatency(MI, i, Reg)) { 1189 LLVM_DEBUG(dbgs() << "Hoist High Latency: " << MI); 1190 ++NumHighLatency; 1191 return true; 1192 } 1193 } 1194 1195 // Estimate register pressure to determine whether to LICM the instruction. 1196 // In low register pressure situation, we can be more aggressive about 1197 // hoisting. Also, favors hoisting long latency instructions even in 1198 // moderately high pressure situation. 1199 // Cheap instructions will only be hoisted if they don't increase register 1200 // pressure at all. 1201 auto Cost = calcRegisterCost(&MI, /*ConsiderSeen=*/false, 1202 /*ConsiderUnseenAsDef=*/false); 1203 1204 // Visit BBs from header to current BB, if hoisting this doesn't cause 1205 // high register pressure, then it's safe to proceed. 1206 if (!CanCauseHighRegPressure(Cost, CheapInstr)) { 1207 LLVM_DEBUG(dbgs() << "Hoist non-reg-pressure: " << MI); 1208 ++NumLowRP; 1209 return true; 1210 } 1211 1212 // Don't risk increasing register pressure if it would create copies. 1213 if (CreatesCopy) { 1214 LLVM_DEBUG(dbgs() << "Won't hoist instr with loop PHI use: " << MI); 1215 return false; 1216 } 1217 1218 // Do not "speculate" in high register pressure situation. If an 1219 // instruction is not guaranteed to be executed in the loop, it's best to be 1220 // conservative. 1221 if (AvoidSpeculation && 1222 (!IsGuaranteedToExecute(MI.getParent()) && !MayCSE(&MI))) { 1223 LLVM_DEBUG(dbgs() << "Won't speculate: " << MI); 1224 return false; 1225 } 1226 1227 // High register pressure situation, only hoist if the instruction is going 1228 // to be remat'ed. 1229 if (!isTriviallyReMaterializable(MI) && 1230 !MI.isDereferenceableInvariantLoad()) { 1231 LLVM_DEBUG(dbgs() << "Can't remat / high reg-pressure: " << MI); 1232 return false; 1233 } 1234 1235 return true; 1236 } 1237 1238 /// Unfold a load from the given machineinstr if the load itself could be 1239 /// hoisted. Return the unfolded and hoistable load, or null if the load 1240 /// couldn't be unfolded or if it wouldn't be hoistable. 1241 MachineInstr *MachineLICMBase::ExtractHoistableLoad(MachineInstr *MI) { 1242 // Don't unfold simple loads. 1243 if (MI->canFoldAsLoad()) 1244 return nullptr; 1245 1246 // If not, we may be able to unfold a load and hoist that. 1247 // First test whether the instruction is loading from an amenable 1248 // memory location. 1249 if (!MI->isDereferenceableInvariantLoad()) 1250 return nullptr; 1251 1252 // Next determine the register class for a temporary register. 1253 unsigned LoadRegIndex; 1254 unsigned NewOpc = 1255 TII->getOpcodeAfterMemoryUnfold(MI->getOpcode(), 1256 /*UnfoldLoad=*/true, 1257 /*UnfoldStore=*/false, 1258 &LoadRegIndex); 1259 if (NewOpc == 0) return nullptr; 1260 const MCInstrDesc &MID = TII->get(NewOpc); 1261 MachineFunction &MF = *MI->getMF(); 1262 const TargetRegisterClass *RC = TII->getRegClass(MID, LoadRegIndex, TRI, MF); 1263 // Ok, we're unfolding. Create a temporary register and do the unfold. 1264 Register Reg = MRI->createVirtualRegister(RC); 1265 1266 SmallVector<MachineInstr *, 2> NewMIs; 1267 bool Success = TII->unfoldMemoryOperand(MF, *MI, Reg, 1268 /*UnfoldLoad=*/true, 1269 /*UnfoldStore=*/false, NewMIs); 1270 (void)Success; 1271 assert(Success && 1272 "unfoldMemoryOperand failed when getOpcodeAfterMemoryUnfold " 1273 "succeeded!"); 1274 assert(NewMIs.size() == 2 && 1275 "Unfolded a load into multiple instructions!"); 1276 MachineBasicBlock *MBB = MI->getParent(); 1277 MachineBasicBlock::iterator Pos = MI; 1278 MBB->insert(Pos, NewMIs[0]); 1279 MBB->insert(Pos, NewMIs[1]); 1280 // If unfolding produced a load that wasn't loop-invariant or profitable to 1281 // hoist, discard the new instructions and bail. 1282 if (!IsLoopInvariantInst(*NewMIs[0]) || !IsProfitableToHoist(*NewMIs[0])) { 1283 NewMIs[0]->eraseFromParent(); 1284 NewMIs[1]->eraseFromParent(); 1285 return nullptr; 1286 } 1287 1288 // Update register pressure for the unfolded instruction. 1289 UpdateRegPressure(NewMIs[1]); 1290 1291 // Otherwise we successfully unfolded a load that we can hoist. 1292 1293 // Update the call site info. 1294 if (MI->shouldUpdateCallSiteInfo()) 1295 MF.eraseCallSiteInfo(MI); 1296 1297 MI->eraseFromParent(); 1298 return NewMIs[0]; 1299 } 1300 1301 /// Initialize the CSE map with instructions that are in the current loop 1302 /// preheader that may become duplicates of instructions that are hoisted 1303 /// out of the loop. 1304 void MachineLICMBase::InitCSEMap(MachineBasicBlock *BB) { 1305 for (MachineInstr &MI : *BB) 1306 CSEMap[MI.getOpcode()].push_back(&MI); 1307 } 1308 1309 /// Find an instruction amount PrevMIs that is a duplicate of MI. 1310 /// Return this instruction if it's found. 1311 MachineInstr * 1312 MachineLICMBase::LookForDuplicate(const MachineInstr *MI, 1313 std::vector<MachineInstr *> &PrevMIs) { 1314 for (MachineInstr *PrevMI : PrevMIs) 1315 if (TII->produceSameValue(*MI, *PrevMI, (PreRegAlloc ? MRI : nullptr))) 1316 return PrevMI; 1317 1318 return nullptr; 1319 } 1320 1321 /// Given a LICM'ed instruction, look for an instruction on the preheader that 1322 /// computes the same value. If it's found, do a RAU on with the definition of 1323 /// the existing instruction rather than hoisting the instruction to the 1324 /// preheader. 1325 bool MachineLICMBase::EliminateCSE( 1326 MachineInstr *MI, 1327 DenseMap<unsigned, std::vector<MachineInstr *>>::iterator &CI) { 1328 // Do not CSE implicit_def so ProcessImplicitDefs can properly propagate 1329 // the undef property onto uses. 1330 if (CI == CSEMap.end() || MI->isImplicitDef()) 1331 return false; 1332 1333 if (MachineInstr *Dup = LookForDuplicate(MI, CI->second)) { 1334 LLVM_DEBUG(dbgs() << "CSEing " << *MI << " with " << *Dup); 1335 1336 // Replace virtual registers defined by MI by their counterparts defined 1337 // by Dup. 1338 SmallVector<unsigned, 2> Defs; 1339 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) { 1340 const MachineOperand &MO = MI->getOperand(i); 1341 1342 // Physical registers may not differ here. 1343 assert((!MO.isReg() || MO.getReg() == 0 || 1344 !Register::isPhysicalRegister(MO.getReg()) || 1345 MO.getReg() == Dup->getOperand(i).getReg()) && 1346 "Instructions with different phys regs are not identical!"); 1347 1348 if (MO.isReg() && MO.isDef() && 1349 !Register::isPhysicalRegister(MO.getReg())) 1350 Defs.push_back(i); 1351 } 1352 1353 SmallVector<const TargetRegisterClass*, 2> OrigRCs; 1354 for (unsigned i = 0, e = Defs.size(); i != e; ++i) { 1355 unsigned Idx = Defs[i]; 1356 Register Reg = MI->getOperand(Idx).getReg(); 1357 Register DupReg = Dup->getOperand(Idx).getReg(); 1358 OrigRCs.push_back(MRI->getRegClass(DupReg)); 1359 1360 if (!MRI->constrainRegClass(DupReg, MRI->getRegClass(Reg))) { 1361 // Restore old RCs if more than one defs. 1362 for (unsigned j = 0; j != i; ++j) 1363 MRI->setRegClass(Dup->getOperand(Defs[j]).getReg(), OrigRCs[j]); 1364 return false; 1365 } 1366 } 1367 1368 for (unsigned Idx : Defs) { 1369 Register Reg = MI->getOperand(Idx).getReg(); 1370 Register DupReg = Dup->getOperand(Idx).getReg(); 1371 MRI->replaceRegWith(Reg, DupReg); 1372 MRI->clearKillFlags(DupReg); 1373 // Clear Dup dead flag if any, we reuse it for Reg. 1374 if (!MRI->use_nodbg_empty(DupReg)) 1375 Dup->getOperand(Idx).setIsDead(false); 1376 } 1377 1378 MI->eraseFromParent(); 1379 ++NumCSEed; 1380 return true; 1381 } 1382 return false; 1383 } 1384 1385 /// Return true if the given instruction will be CSE'd if it's hoisted out of 1386 /// the loop. 1387 bool MachineLICMBase::MayCSE(MachineInstr *MI) { 1388 unsigned Opcode = MI->getOpcode(); 1389 DenseMap<unsigned, std::vector<MachineInstr *>>::iterator CI = 1390 CSEMap.find(Opcode); 1391 // Do not CSE implicit_def so ProcessImplicitDefs can properly propagate 1392 // the undef property onto uses. 1393 if (CI == CSEMap.end() || MI->isImplicitDef()) 1394 return false; 1395 1396 return LookForDuplicate(MI, CI->second) != nullptr; 1397 } 1398 1399 /// When an instruction is found to use only loop invariant operands 1400 /// that are safe to hoist, this instruction is called to do the dirty work. 1401 /// It returns true if the instruction is hoisted. 1402 bool MachineLICMBase::Hoist(MachineInstr *MI, MachineBasicBlock *Preheader) { 1403 MachineBasicBlock *SrcBlock = MI->getParent(); 1404 1405 // Disable the instruction hoisting due to block hotness 1406 if ((DisableHoistingToHotterBlocks == UseBFI::All || 1407 (DisableHoistingToHotterBlocks == UseBFI::PGO && HasProfileData)) && 1408 isTgtHotterThanSrc(SrcBlock, Preheader)) { 1409 ++NumNotHoistedDueToHotness; 1410 return false; 1411 } 1412 // First check whether we should hoist this instruction. 1413 if (!IsLoopInvariantInst(*MI) || !IsProfitableToHoist(*MI)) { 1414 // If not, try unfolding a hoistable load. 1415 MI = ExtractHoistableLoad(MI); 1416 if (!MI) return false; 1417 } 1418 1419 // If we have hoisted an instruction that may store, it can only be a constant 1420 // store. 1421 if (MI->mayStore()) 1422 NumStoreConst++; 1423 1424 // Now move the instructions to the predecessor, inserting it before any 1425 // terminator instructions. 1426 LLVM_DEBUG({ 1427 dbgs() << "Hoisting " << *MI; 1428 if (MI->getParent()->getBasicBlock()) 1429 dbgs() << " from " << printMBBReference(*MI->getParent()); 1430 if (Preheader->getBasicBlock()) 1431 dbgs() << " to " << printMBBReference(*Preheader); 1432 dbgs() << "\n"; 1433 }); 1434 1435 // If this is the first instruction being hoisted to the preheader, 1436 // initialize the CSE map with potential common expressions. 1437 if (FirstInLoop) { 1438 InitCSEMap(Preheader); 1439 FirstInLoop = false; 1440 } 1441 1442 // Look for opportunity to CSE the hoisted instruction. 1443 unsigned Opcode = MI->getOpcode(); 1444 DenseMap<unsigned, std::vector<MachineInstr *>>::iterator CI = 1445 CSEMap.find(Opcode); 1446 if (!EliminateCSE(MI, CI)) { 1447 // Otherwise, splice the instruction to the preheader. 1448 Preheader->splice(Preheader->getFirstTerminator(),MI->getParent(),MI); 1449 1450 // Since we are moving the instruction out of its basic block, we do not 1451 // retain its debug location. Doing so would degrade the debugging 1452 // experience and adversely affect the accuracy of profiling information. 1453 assert(!MI->isDebugInstr() && "Should not hoist debug inst"); 1454 MI->setDebugLoc(DebugLoc()); 1455 1456 // Update register pressure for BBs from header to this block. 1457 UpdateBackTraceRegPressure(MI); 1458 1459 // Clear the kill flags of any register this instruction defines, 1460 // since they may need to be live throughout the entire loop 1461 // rather than just live for part of it. 1462 for (MachineOperand &MO : MI->operands()) 1463 if (MO.isReg() && MO.isDef() && !MO.isDead()) 1464 MRI->clearKillFlags(MO.getReg()); 1465 1466 // Add to the CSE map. 1467 if (CI != CSEMap.end()) 1468 CI->second.push_back(MI); 1469 else 1470 CSEMap[Opcode].push_back(MI); 1471 } 1472 1473 ++NumHoisted; 1474 Changed = true; 1475 1476 return true; 1477 } 1478 1479 /// Get the preheader for the current loop, splitting a critical edge if needed. 1480 MachineBasicBlock *MachineLICMBase::getCurPreheader() { 1481 // Determine the block to which to hoist instructions. If we can't find a 1482 // suitable loop predecessor, we can't do any hoisting. 1483 1484 // If we've tried to get a preheader and failed, don't try again. 1485 if (CurPreheader == reinterpret_cast<MachineBasicBlock *>(-1)) 1486 return nullptr; 1487 1488 if (!CurPreheader) { 1489 CurPreheader = CurLoop->getLoopPreheader(); 1490 if (!CurPreheader) { 1491 MachineBasicBlock *Pred = CurLoop->getLoopPredecessor(); 1492 if (!Pred) { 1493 CurPreheader = reinterpret_cast<MachineBasicBlock *>(-1); 1494 return nullptr; 1495 } 1496 1497 CurPreheader = Pred->SplitCriticalEdge(CurLoop->getHeader(), *this); 1498 if (!CurPreheader) { 1499 CurPreheader = reinterpret_cast<MachineBasicBlock *>(-1); 1500 return nullptr; 1501 } 1502 } 1503 } 1504 return CurPreheader; 1505 } 1506 1507 /// Is the target basic block at least "BlockFrequencyRatioThreshold" 1508 /// times hotter than the source basic block. 1509 bool MachineLICMBase::isTgtHotterThanSrc(MachineBasicBlock *SrcBlock, 1510 MachineBasicBlock *TgtBlock) { 1511 // Parse source and target basic block frequency from MBFI 1512 uint64_t SrcBF = MBFI->getBlockFreq(SrcBlock).getFrequency(); 1513 uint64_t DstBF = MBFI->getBlockFreq(TgtBlock).getFrequency(); 1514 1515 // Disable the hoisting if source block frequency is zero 1516 if (!SrcBF) 1517 return true; 1518 1519 double Ratio = (double)DstBF / SrcBF; 1520 1521 // Compare the block frequency ratio with the threshold 1522 return Ratio > BlockFrequencyRatioThreshold; 1523 } 1524