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