1 //===- LoopLoadElimination.cpp - Loop Load Elimination 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 file implement a loop-aware load elimination pass. 10 // 11 // It uses LoopAccessAnalysis to identify loop-carried dependences with a 12 // distance of one between stores and loads. These form the candidates for the 13 // transformation. The source value of each store then propagated to the user 14 // of the corresponding load. This makes the load dead. 15 // 16 // The pass can also version the loop and add memchecks in order to prove that 17 // may-aliasing stores can't change the value in memory before it's read by the 18 // load. 19 // 20 //===----------------------------------------------------------------------===// 21 22 #include "llvm/Transforms/Scalar/LoopLoadElimination.h" 23 #include "llvm/ADT/APInt.h" 24 #include "llvm/ADT/DenseMap.h" 25 #include "llvm/ADT/DepthFirstIterator.h" 26 #include "llvm/ADT/STLExtras.h" 27 #include "llvm/ADT/SmallPtrSet.h" 28 #include "llvm/ADT/SmallVector.h" 29 #include "llvm/ADT/Statistic.h" 30 #include "llvm/Analysis/AssumptionCache.h" 31 #include "llvm/Analysis/BlockFrequencyInfo.h" 32 #include "llvm/Analysis/GlobalsModRef.h" 33 #include "llvm/Analysis/LazyBlockFrequencyInfo.h" 34 #include "llvm/Analysis/LoopAccessAnalysis.h" 35 #include "llvm/Analysis/LoopAnalysisManager.h" 36 #include "llvm/Analysis/LoopInfo.h" 37 #include "llvm/Analysis/MemorySSA.h" 38 #include "llvm/Analysis/ProfileSummaryInfo.h" 39 #include "llvm/Analysis/ScalarEvolution.h" 40 #include "llvm/Analysis/ScalarEvolutionExpressions.h" 41 #include "llvm/Analysis/TargetLibraryInfo.h" 42 #include "llvm/Analysis/TargetTransformInfo.h" 43 #include "llvm/IR/DataLayout.h" 44 #include "llvm/IR/Dominators.h" 45 #include "llvm/IR/Instructions.h" 46 #include "llvm/IR/Module.h" 47 #include "llvm/IR/PassManager.h" 48 #include "llvm/IR/Type.h" 49 #include "llvm/IR/Value.h" 50 #include "llvm/InitializePasses.h" 51 #include "llvm/Pass.h" 52 #include "llvm/Support/Casting.h" 53 #include "llvm/Support/CommandLine.h" 54 #include "llvm/Support/Debug.h" 55 #include "llvm/Support/raw_ostream.h" 56 #include "llvm/Transforms/Scalar.h" 57 #include "llvm/Transforms/Utils.h" 58 #include "llvm/Transforms/Utils/LoopSimplify.h" 59 #include "llvm/Transforms/Utils/LoopVersioning.h" 60 #include "llvm/Transforms/Utils/ScalarEvolutionExpander.h" 61 #include "llvm/Transforms/Utils/SizeOpts.h" 62 #include <algorithm> 63 #include <cassert> 64 #include <forward_list> 65 #include <set> 66 #include <tuple> 67 #include <utility> 68 69 using namespace llvm; 70 71 #define LLE_OPTION "loop-load-elim" 72 #define DEBUG_TYPE LLE_OPTION 73 74 static cl::opt<unsigned> CheckPerElim( 75 "runtime-check-per-loop-load-elim", cl::Hidden, 76 cl::desc("Max number of memchecks allowed per eliminated load on average"), 77 cl::init(1)); 78 79 static cl::opt<unsigned> LoadElimSCEVCheckThreshold( 80 "loop-load-elimination-scev-check-threshold", cl::init(8), cl::Hidden, 81 cl::desc("The maximum number of SCEV checks allowed for Loop " 82 "Load Elimination")); 83 84 STATISTIC(NumLoopLoadEliminted, "Number of loads eliminated by LLE"); 85 86 namespace { 87 88 /// Represent a store-to-forwarding candidate. 89 struct StoreToLoadForwardingCandidate { 90 LoadInst *Load; 91 StoreInst *Store; 92 93 StoreToLoadForwardingCandidate(LoadInst *Load, StoreInst *Store) 94 : Load(Load), Store(Store) {} 95 96 /// Return true if the dependence from the store to the load has a 97 /// distance of one. E.g. A[i+1] = A[i] 98 bool isDependenceDistanceOfOne(PredicatedScalarEvolution &PSE, 99 Loop *L) const { 100 Value *LoadPtr = Load->getPointerOperand(); 101 Value *StorePtr = Store->getPointerOperand(); 102 Type *LoadType = getLoadStoreType(Load); 103 104 assert(LoadPtr->getType()->getPointerAddressSpace() == 105 StorePtr->getType()->getPointerAddressSpace() && 106 LoadType == getLoadStoreType(Store) && 107 "Should be a known dependence"); 108 109 // Currently we only support accesses with unit stride. FIXME: we should be 110 // able to handle non unit stirde as well as long as the stride is equal to 111 // the dependence distance. 112 if (getPtrStride(PSE, LoadPtr, L) != 1 || 113 getPtrStride(PSE, StorePtr, L) != 1) 114 return false; 115 116 auto &DL = Load->getParent()->getModule()->getDataLayout(); 117 unsigned TypeByteSize = DL.getTypeAllocSize(const_cast<Type *>(LoadType)); 118 119 auto *LoadPtrSCEV = cast<SCEVAddRecExpr>(PSE.getSCEV(LoadPtr)); 120 auto *StorePtrSCEV = cast<SCEVAddRecExpr>(PSE.getSCEV(StorePtr)); 121 122 // We don't need to check non-wrapping here because forward/backward 123 // dependence wouldn't be valid if these weren't monotonic accesses. 124 auto *Dist = cast<SCEVConstant>( 125 PSE.getSE()->getMinusSCEV(StorePtrSCEV, LoadPtrSCEV)); 126 const APInt &Val = Dist->getAPInt(); 127 return Val == TypeByteSize; 128 } 129 130 Value *getLoadPtr() const { return Load->getPointerOperand(); } 131 132 #ifndef NDEBUG 133 friend raw_ostream &operator<<(raw_ostream &OS, 134 const StoreToLoadForwardingCandidate &Cand) { 135 OS << *Cand.Store << " -->\n"; 136 OS.indent(2) << *Cand.Load << "\n"; 137 return OS; 138 } 139 #endif 140 }; 141 142 } // end anonymous namespace 143 144 /// Check if the store dominates all latches, so as long as there is no 145 /// intervening store this value will be loaded in the next iteration. 146 static bool doesStoreDominatesAllLatches(BasicBlock *StoreBlock, Loop *L, 147 DominatorTree *DT) { 148 SmallVector<BasicBlock *, 8> Latches; 149 L->getLoopLatches(Latches); 150 return llvm::all_of(Latches, [&](const BasicBlock *Latch) { 151 return DT->dominates(StoreBlock, Latch); 152 }); 153 } 154 155 /// Return true if the load is not executed on all paths in the loop. 156 static bool isLoadConditional(LoadInst *Load, Loop *L) { 157 return Load->getParent() != L->getHeader(); 158 } 159 160 namespace { 161 162 /// The per-loop class that does most of the work. 163 class LoadEliminationForLoop { 164 public: 165 LoadEliminationForLoop(Loop *L, LoopInfo *LI, const LoopAccessInfo &LAI, 166 DominatorTree *DT, BlockFrequencyInfo *BFI, 167 ProfileSummaryInfo* PSI) 168 : L(L), LI(LI), LAI(LAI), DT(DT), BFI(BFI), PSI(PSI), PSE(LAI.getPSE()) {} 169 170 /// Look through the loop-carried and loop-independent dependences in 171 /// this loop and find store->load dependences. 172 /// 173 /// Note that no candidate is returned if LAA has failed to analyze the loop 174 /// (e.g. if it's not bottom-tested, contains volatile memops, etc.) 175 std::forward_list<StoreToLoadForwardingCandidate> 176 findStoreToLoadDependences(const LoopAccessInfo &LAI) { 177 std::forward_list<StoreToLoadForwardingCandidate> Candidates; 178 179 const auto *Deps = LAI.getDepChecker().getDependences(); 180 if (!Deps) 181 return Candidates; 182 183 // Find store->load dependences (consequently true dep). Both lexically 184 // forward and backward dependences qualify. Disqualify loads that have 185 // other unknown dependences. 186 187 SmallPtrSet<Instruction *, 4> LoadsWithUnknownDepedence; 188 189 for (const auto &Dep : *Deps) { 190 Instruction *Source = Dep.getSource(LAI); 191 Instruction *Destination = Dep.getDestination(LAI); 192 193 if (Dep.Type == MemoryDepChecker::Dependence::Unknown) { 194 if (isa<LoadInst>(Source)) 195 LoadsWithUnknownDepedence.insert(Source); 196 if (isa<LoadInst>(Destination)) 197 LoadsWithUnknownDepedence.insert(Destination); 198 continue; 199 } 200 201 if (Dep.isBackward()) 202 // Note that the designations source and destination follow the program 203 // order, i.e. source is always first. (The direction is given by the 204 // DepType.) 205 std::swap(Source, Destination); 206 else 207 assert(Dep.isForward() && "Needs to be a forward dependence"); 208 209 auto *Store = dyn_cast<StoreInst>(Source); 210 if (!Store) 211 continue; 212 auto *Load = dyn_cast<LoadInst>(Destination); 213 if (!Load) 214 continue; 215 216 // Only progagate the value if they are of the same type. 217 if (Store->getPointerOperandType() != Load->getPointerOperandType()) 218 continue; 219 220 Candidates.emplace_front(Load, Store); 221 } 222 223 if (!LoadsWithUnknownDepedence.empty()) 224 Candidates.remove_if([&](const StoreToLoadForwardingCandidate &C) { 225 return LoadsWithUnknownDepedence.count(C.Load); 226 }); 227 228 return Candidates; 229 } 230 231 /// Return the index of the instruction according to program order. 232 unsigned getInstrIndex(Instruction *Inst) { 233 auto I = InstOrder.find(Inst); 234 assert(I != InstOrder.end() && "No index for instruction"); 235 return I->second; 236 } 237 238 /// If a load has multiple candidates associated (i.e. different 239 /// stores), it means that it could be forwarding from multiple stores 240 /// depending on control flow. Remove these candidates. 241 /// 242 /// Here, we rely on LAA to include the relevant loop-independent dependences. 243 /// LAA is known to omit these in the very simple case when the read and the 244 /// write within an alias set always takes place using the *same* pointer. 245 /// 246 /// However, we know that this is not the case here, i.e. we can rely on LAA 247 /// to provide us with loop-independent dependences for the cases we're 248 /// interested. Consider the case for example where a loop-independent 249 /// dependece S1->S2 invalidates the forwarding S3->S2. 250 /// 251 /// A[i] = ... (S1) 252 /// ... = A[i] (S2) 253 /// A[i+1] = ... (S3) 254 /// 255 /// LAA will perform dependence analysis here because there are two 256 /// *different* pointers involved in the same alias set (&A[i] and &A[i+1]). 257 void removeDependencesFromMultipleStores( 258 std::forward_list<StoreToLoadForwardingCandidate> &Candidates) { 259 // If Store is nullptr it means that we have multiple stores forwarding to 260 // this store. 261 using LoadToSingleCandT = 262 DenseMap<LoadInst *, const StoreToLoadForwardingCandidate *>; 263 LoadToSingleCandT LoadToSingleCand; 264 265 for (const auto &Cand : Candidates) { 266 bool NewElt; 267 LoadToSingleCandT::iterator Iter; 268 269 std::tie(Iter, NewElt) = 270 LoadToSingleCand.insert(std::make_pair(Cand.Load, &Cand)); 271 if (!NewElt) { 272 const StoreToLoadForwardingCandidate *&OtherCand = Iter->second; 273 // Already multiple stores forward to this load. 274 if (OtherCand == nullptr) 275 continue; 276 277 // Handle the very basic case when the two stores are in the same block 278 // so deciding which one forwards is easy. The later one forwards as 279 // long as they both have a dependence distance of one to the load. 280 if (Cand.Store->getParent() == OtherCand->Store->getParent() && 281 Cand.isDependenceDistanceOfOne(PSE, L) && 282 OtherCand->isDependenceDistanceOfOne(PSE, L)) { 283 // They are in the same block, the later one will forward to the load. 284 if (getInstrIndex(OtherCand->Store) < getInstrIndex(Cand.Store)) 285 OtherCand = &Cand; 286 } else 287 OtherCand = nullptr; 288 } 289 } 290 291 Candidates.remove_if([&](const StoreToLoadForwardingCandidate &Cand) { 292 if (LoadToSingleCand[Cand.Load] != &Cand) { 293 LLVM_DEBUG( 294 dbgs() << "Removing from candidates: \n" 295 << Cand 296 << " The load may have multiple stores forwarding to " 297 << "it\n"); 298 return true; 299 } 300 return false; 301 }); 302 } 303 304 /// Given two pointers operations by their RuntimePointerChecking 305 /// indices, return true if they require an alias check. 306 /// 307 /// We need a check if one is a pointer for a candidate load and the other is 308 /// a pointer for a possibly intervening store. 309 bool needsChecking(unsigned PtrIdx1, unsigned PtrIdx2, 310 const SmallPtrSetImpl<Value *> &PtrsWrittenOnFwdingPath, 311 const SmallPtrSetImpl<Value *> &CandLoadPtrs) { 312 Value *Ptr1 = 313 LAI.getRuntimePointerChecking()->getPointerInfo(PtrIdx1).PointerValue; 314 Value *Ptr2 = 315 LAI.getRuntimePointerChecking()->getPointerInfo(PtrIdx2).PointerValue; 316 return ((PtrsWrittenOnFwdingPath.count(Ptr1) && CandLoadPtrs.count(Ptr2)) || 317 (PtrsWrittenOnFwdingPath.count(Ptr2) && CandLoadPtrs.count(Ptr1))); 318 } 319 320 /// Return pointers that are possibly written to on the path from a 321 /// forwarding store to a load. 322 /// 323 /// These pointers need to be alias-checked against the forwarding candidates. 324 SmallPtrSet<Value *, 4> findPointersWrittenOnForwardingPath( 325 const SmallVectorImpl<StoreToLoadForwardingCandidate> &Candidates) { 326 // From FirstStore to LastLoad neither of the elimination candidate loads 327 // should overlap with any of the stores. 328 // 329 // E.g.: 330 // 331 // st1 C[i] 332 // ld1 B[i] <-------, 333 // ld0 A[i] <----, | * LastLoad 334 // ... | | 335 // st2 E[i] | | 336 // st3 B[i+1] -- | -' * FirstStore 337 // st0 A[i+1] ---' 338 // st4 D[i] 339 // 340 // st0 forwards to ld0 if the accesses in st4 and st1 don't overlap with 341 // ld0. 342 343 LoadInst *LastLoad = 344 std::max_element(Candidates.begin(), Candidates.end(), 345 [&](const StoreToLoadForwardingCandidate &A, 346 const StoreToLoadForwardingCandidate &B) { 347 return getInstrIndex(A.Load) < getInstrIndex(B.Load); 348 }) 349 ->Load; 350 StoreInst *FirstStore = 351 std::min_element(Candidates.begin(), Candidates.end(), 352 [&](const StoreToLoadForwardingCandidate &A, 353 const StoreToLoadForwardingCandidate &B) { 354 return getInstrIndex(A.Store) < 355 getInstrIndex(B.Store); 356 }) 357 ->Store; 358 359 // We're looking for stores after the first forwarding store until the end 360 // of the loop, then from the beginning of the loop until the last 361 // forwarded-to load. Collect the pointer for the stores. 362 SmallPtrSet<Value *, 4> PtrsWrittenOnFwdingPath; 363 364 auto InsertStorePtr = [&](Instruction *I) { 365 if (auto *S = dyn_cast<StoreInst>(I)) 366 PtrsWrittenOnFwdingPath.insert(S->getPointerOperand()); 367 }; 368 const auto &MemInstrs = LAI.getDepChecker().getMemoryInstructions(); 369 std::for_each(MemInstrs.begin() + getInstrIndex(FirstStore) + 1, 370 MemInstrs.end(), InsertStorePtr); 371 std::for_each(MemInstrs.begin(), &MemInstrs[getInstrIndex(LastLoad)], 372 InsertStorePtr); 373 374 return PtrsWrittenOnFwdingPath; 375 } 376 377 /// Determine the pointer alias checks to prove that there are no 378 /// intervening stores. 379 SmallVector<RuntimePointerCheck, 4> collectMemchecks( 380 const SmallVectorImpl<StoreToLoadForwardingCandidate> &Candidates) { 381 382 SmallPtrSet<Value *, 4> PtrsWrittenOnFwdingPath = 383 findPointersWrittenOnForwardingPath(Candidates); 384 385 // Collect the pointers of the candidate loads. 386 SmallPtrSet<Value *, 4> CandLoadPtrs; 387 for (const auto &Candidate : Candidates) 388 CandLoadPtrs.insert(Candidate.getLoadPtr()); 389 390 const auto &AllChecks = LAI.getRuntimePointerChecking()->getChecks(); 391 SmallVector<RuntimePointerCheck, 4> Checks; 392 393 copy_if(AllChecks, std::back_inserter(Checks), 394 [&](const RuntimePointerCheck &Check) { 395 for (auto PtrIdx1 : Check.first->Members) 396 for (auto PtrIdx2 : Check.second->Members) 397 if (needsChecking(PtrIdx1, PtrIdx2, PtrsWrittenOnFwdingPath, 398 CandLoadPtrs)) 399 return true; 400 return false; 401 }); 402 403 LLVM_DEBUG(dbgs() << "\nPointer Checks (count: " << Checks.size() 404 << "):\n"); 405 LLVM_DEBUG(LAI.getRuntimePointerChecking()->printChecks(dbgs(), Checks)); 406 407 return Checks; 408 } 409 410 /// Perform the transformation for a candidate. 411 void 412 propagateStoredValueToLoadUsers(const StoreToLoadForwardingCandidate &Cand, 413 SCEVExpander &SEE) { 414 // loop: 415 // %x = load %gep_i 416 // = ... %x 417 // store %y, %gep_i_plus_1 418 // 419 // => 420 // 421 // ph: 422 // %x.initial = load %gep_0 423 // loop: 424 // %x.storeforward = phi [%x.initial, %ph] [%y, %loop] 425 // %x = load %gep_i <---- now dead 426 // = ... %x.storeforward 427 // store %y, %gep_i_plus_1 428 429 Value *Ptr = Cand.Load->getPointerOperand(); 430 auto *PtrSCEV = cast<SCEVAddRecExpr>(PSE.getSCEV(Ptr)); 431 auto *PH = L->getLoopPreheader(); 432 assert(PH && "Preheader should exist!"); 433 Value *InitialPtr = SEE.expandCodeFor(PtrSCEV->getStart(), Ptr->getType(), 434 PH->getTerminator()); 435 Value *Initial = new LoadInst( 436 Cand.Load->getType(), InitialPtr, "load_initial", 437 /* isVolatile */ false, Cand.Load->getAlign(), PH->getTerminator()); 438 439 PHINode *PHI = PHINode::Create(Initial->getType(), 2, "store_forwarded", 440 &L->getHeader()->front()); 441 PHI->addIncoming(Initial, PH); 442 PHI->addIncoming(Cand.Store->getOperand(0), L->getLoopLatch()); 443 444 Cand.Load->replaceAllUsesWith(PHI); 445 } 446 447 /// Top-level driver for each loop: find store->load forwarding 448 /// candidates, add run-time checks and perform transformation. 449 bool processLoop() { 450 LLVM_DEBUG(dbgs() << "\nIn \"" << L->getHeader()->getParent()->getName() 451 << "\" checking " << *L << "\n"); 452 453 // Look for store-to-load forwarding cases across the 454 // backedge. E.g.: 455 // 456 // loop: 457 // %x = load %gep_i 458 // = ... %x 459 // store %y, %gep_i_plus_1 460 // 461 // => 462 // 463 // ph: 464 // %x.initial = load %gep_0 465 // loop: 466 // %x.storeforward = phi [%x.initial, %ph] [%y, %loop] 467 // %x = load %gep_i <---- now dead 468 // = ... %x.storeforward 469 // store %y, %gep_i_plus_1 470 471 // First start with store->load dependences. 472 auto StoreToLoadDependences = findStoreToLoadDependences(LAI); 473 if (StoreToLoadDependences.empty()) 474 return false; 475 476 // Generate an index for each load and store according to the original 477 // program order. This will be used later. 478 InstOrder = LAI.getDepChecker().generateInstructionOrderMap(); 479 480 // To keep things simple for now, remove those where the load is potentially 481 // fed by multiple stores. 482 removeDependencesFromMultipleStores(StoreToLoadDependences); 483 if (StoreToLoadDependences.empty()) 484 return false; 485 486 // Filter the candidates further. 487 SmallVector<StoreToLoadForwardingCandidate, 4> Candidates; 488 for (const StoreToLoadForwardingCandidate &Cand : StoreToLoadDependences) { 489 LLVM_DEBUG(dbgs() << "Candidate " << Cand); 490 491 // Make sure that the stored values is available everywhere in the loop in 492 // the next iteration. 493 if (!doesStoreDominatesAllLatches(Cand.Store->getParent(), L, DT)) 494 continue; 495 496 // If the load is conditional we can't hoist its 0-iteration instance to 497 // the preheader because that would make it unconditional. Thus we would 498 // access a memory location that the original loop did not access. 499 if (isLoadConditional(Cand.Load, L)) 500 continue; 501 502 // Check whether the SCEV difference is the same as the induction step, 503 // thus we load the value in the next iteration. 504 if (!Cand.isDependenceDistanceOfOne(PSE, L)) 505 continue; 506 507 assert(isa<SCEVAddRecExpr>(PSE.getSCEV(Cand.Load->getPointerOperand())) && 508 "Loading from something other than indvar?"); 509 assert( 510 isa<SCEVAddRecExpr>(PSE.getSCEV(Cand.Store->getPointerOperand())) && 511 "Storing to something other than indvar?"); 512 513 Candidates.push_back(Cand); 514 LLVM_DEBUG( 515 dbgs() 516 << Candidates.size() 517 << ". Valid store-to-load forwarding across the loop backedge\n"); 518 } 519 if (Candidates.empty()) 520 return false; 521 522 // Check intervening may-alias stores. These need runtime checks for alias 523 // disambiguation. 524 SmallVector<RuntimePointerCheck, 4> Checks = collectMemchecks(Candidates); 525 526 // Too many checks are likely to outweigh the benefits of forwarding. 527 if (Checks.size() > Candidates.size() * CheckPerElim) { 528 LLVM_DEBUG(dbgs() << "Too many run-time checks needed.\n"); 529 return false; 530 } 531 532 if (LAI.getPSE().getUnionPredicate().getComplexity() > 533 LoadElimSCEVCheckThreshold) { 534 LLVM_DEBUG(dbgs() << "Too many SCEV run-time checks needed.\n"); 535 return false; 536 } 537 538 if (!L->isLoopSimplifyForm()) { 539 LLVM_DEBUG(dbgs() << "Loop is not is loop-simplify form"); 540 return false; 541 } 542 543 if (!Checks.empty() || !LAI.getPSE().getUnionPredicate().isAlwaysTrue()) { 544 if (LAI.hasConvergentOp()) { 545 LLVM_DEBUG(dbgs() << "Versioning is needed but not allowed with " 546 "convergent calls\n"); 547 return false; 548 } 549 550 auto *HeaderBB = L->getHeader(); 551 auto *F = HeaderBB->getParent(); 552 bool OptForSize = F->hasOptSize() || 553 llvm::shouldOptimizeForSize(HeaderBB, PSI, BFI, 554 PGSOQueryType::IRPass); 555 if (OptForSize) { 556 LLVM_DEBUG( 557 dbgs() << "Versioning is needed but not allowed when optimizing " 558 "for size.\n"); 559 return false; 560 } 561 562 // Point of no-return, start the transformation. First, version the loop 563 // if necessary. 564 565 LoopVersioning LV(LAI, Checks, L, LI, DT, PSE.getSE()); 566 LV.versionLoop(); 567 568 // After versioning, some of the candidates' pointers could stop being 569 // SCEVAddRecs. We need to filter them out. 570 auto NoLongerGoodCandidate = [this]( 571 const StoreToLoadForwardingCandidate &Cand) { 572 return !isa<SCEVAddRecExpr>( 573 PSE.getSCEV(Cand.Load->getPointerOperand())) || 574 !isa<SCEVAddRecExpr>( 575 PSE.getSCEV(Cand.Store->getPointerOperand())); 576 }; 577 llvm::erase_if(Candidates, NoLongerGoodCandidate); 578 } 579 580 // Next, propagate the value stored by the store to the users of the load. 581 // Also for the first iteration, generate the initial value of the load. 582 SCEVExpander SEE(*PSE.getSE(), L->getHeader()->getModule()->getDataLayout(), 583 "storeforward"); 584 for (const auto &Cand : Candidates) 585 propagateStoredValueToLoadUsers(Cand, SEE); 586 NumLoopLoadEliminted += Candidates.size(); 587 588 return true; 589 } 590 591 private: 592 Loop *L; 593 594 /// Maps the load/store instructions to their index according to 595 /// program order. 596 DenseMap<Instruction *, unsigned> InstOrder; 597 598 // Analyses used. 599 LoopInfo *LI; 600 const LoopAccessInfo &LAI; 601 DominatorTree *DT; 602 BlockFrequencyInfo *BFI; 603 ProfileSummaryInfo *PSI; 604 PredicatedScalarEvolution PSE; 605 }; 606 607 } // end anonymous namespace 608 609 static bool 610 eliminateLoadsAcrossLoops(Function &F, LoopInfo &LI, DominatorTree &DT, 611 BlockFrequencyInfo *BFI, ProfileSummaryInfo *PSI, 612 ScalarEvolution *SE, AssumptionCache *AC, 613 function_ref<const LoopAccessInfo &(Loop &)> GetLAI) { 614 // Build up a worklist of inner-loops to transform to avoid iterator 615 // invalidation. 616 // FIXME: This logic comes from other passes that actually change the loop 617 // nest structure. It isn't clear this is necessary (or useful) for a pass 618 // which merely optimizes the use of loads in a loop. 619 SmallVector<Loop *, 8> Worklist; 620 621 bool Changed = false; 622 623 for (Loop *TopLevelLoop : LI) 624 for (Loop *L : depth_first(TopLevelLoop)) { 625 Changed |= simplifyLoop(L, &DT, &LI, SE, AC, /*MSSAU*/ nullptr, false); 626 // We only handle inner-most loops. 627 if (L->isInnermost()) 628 Worklist.push_back(L); 629 } 630 631 // Now walk the identified inner loops. 632 for (Loop *L : Worklist) { 633 // Match historical behavior 634 if (!L->isRotatedForm() || !L->getExitingBlock()) 635 continue; 636 // The actual work is performed by LoadEliminationForLoop. 637 LoadEliminationForLoop LEL(L, &LI, GetLAI(*L), &DT, BFI, PSI); 638 Changed |= LEL.processLoop(); 639 } 640 return Changed; 641 } 642 643 namespace { 644 645 /// The pass. Most of the work is delegated to the per-loop 646 /// LoadEliminationForLoop class. 647 class LoopLoadElimination : public FunctionPass { 648 public: 649 static char ID; 650 651 LoopLoadElimination() : FunctionPass(ID) { 652 initializeLoopLoadEliminationPass(*PassRegistry::getPassRegistry()); 653 } 654 655 bool runOnFunction(Function &F) override { 656 if (skipFunction(F)) 657 return false; 658 659 auto &LI = getAnalysis<LoopInfoWrapperPass>().getLoopInfo(); 660 auto &LAA = getAnalysis<LoopAccessLegacyAnalysis>(); 661 auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree(); 662 auto *PSI = &getAnalysis<ProfileSummaryInfoWrapperPass>().getPSI(); 663 auto *BFI = (PSI && PSI->hasProfileSummary()) ? 664 &getAnalysis<LazyBlockFrequencyInfoPass>().getBFI() : 665 nullptr; 666 auto *SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE(); 667 668 // Process each loop nest in the function. 669 return eliminateLoadsAcrossLoops( 670 F, LI, DT, BFI, PSI, SE, /*AC*/ nullptr, 671 [&LAA](Loop &L) -> const LoopAccessInfo & { return LAA.getInfo(&L); }); 672 } 673 674 void getAnalysisUsage(AnalysisUsage &AU) const override { 675 AU.addRequiredID(LoopSimplifyID); 676 AU.addRequired<LoopInfoWrapperPass>(); 677 AU.addPreserved<LoopInfoWrapperPass>(); 678 AU.addRequired<LoopAccessLegacyAnalysis>(); 679 AU.addRequired<ScalarEvolutionWrapperPass>(); 680 AU.addRequired<DominatorTreeWrapperPass>(); 681 AU.addPreserved<DominatorTreeWrapperPass>(); 682 AU.addPreserved<GlobalsAAWrapperPass>(); 683 AU.addRequired<ProfileSummaryInfoWrapperPass>(); 684 LazyBlockFrequencyInfoPass::getLazyBFIAnalysisUsage(AU); 685 } 686 }; 687 688 } // end anonymous namespace 689 690 char LoopLoadElimination::ID; 691 692 static const char LLE_name[] = "Loop Load Elimination"; 693 694 INITIALIZE_PASS_BEGIN(LoopLoadElimination, LLE_OPTION, LLE_name, false, false) 695 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass) 696 INITIALIZE_PASS_DEPENDENCY(LoopAccessLegacyAnalysis) 697 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass) 698 INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass) 699 INITIALIZE_PASS_DEPENDENCY(LoopSimplify) 700 INITIALIZE_PASS_DEPENDENCY(ProfileSummaryInfoWrapperPass) 701 INITIALIZE_PASS_DEPENDENCY(LazyBlockFrequencyInfoPass) 702 INITIALIZE_PASS_END(LoopLoadElimination, LLE_OPTION, LLE_name, false, false) 703 704 FunctionPass *llvm::createLoopLoadEliminationPass() { 705 return new LoopLoadElimination(); 706 } 707 708 PreservedAnalyses LoopLoadEliminationPass::run(Function &F, 709 FunctionAnalysisManager &AM) { 710 auto &SE = AM.getResult<ScalarEvolutionAnalysis>(F); 711 auto &LI = AM.getResult<LoopAnalysis>(F); 712 auto &TTI = AM.getResult<TargetIRAnalysis>(F); 713 auto &DT = AM.getResult<DominatorTreeAnalysis>(F); 714 auto &TLI = AM.getResult<TargetLibraryAnalysis>(F); 715 auto &AA = AM.getResult<AAManager>(F); 716 auto &AC = AM.getResult<AssumptionAnalysis>(F); 717 auto &MAMProxy = AM.getResult<ModuleAnalysisManagerFunctionProxy>(F); 718 auto *PSI = MAMProxy.getCachedResult<ProfileSummaryAnalysis>(*F.getParent()); 719 auto *BFI = (PSI && PSI->hasProfileSummary()) ? 720 &AM.getResult<BlockFrequencyAnalysis>(F) : nullptr; 721 MemorySSA *MSSA = EnableMSSALoopDependency 722 ? &AM.getResult<MemorySSAAnalysis>(F).getMSSA() 723 : nullptr; 724 725 auto &LAM = AM.getResult<LoopAnalysisManagerFunctionProxy>(F).getManager(); 726 bool Changed = eliminateLoadsAcrossLoops( 727 F, LI, DT, BFI, PSI, &SE, &AC, [&](Loop &L) -> const LoopAccessInfo & { 728 LoopStandardAnalysisResults AR = {AA, AC, DT, LI, SE, 729 TLI, TTI, nullptr, MSSA}; 730 return LAM.getResult<LoopAccessAnalysis>(L, AR); 731 }); 732 733 if (!Changed) 734 return PreservedAnalyses::all(); 735 736 PreservedAnalyses PA; 737 return PA; 738 } 739