lib/Transforms/Scalar/LoopLoadElimination.cpp

   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/AliasAnalysis.h"
  31 #include "llvm/Analysis/AssumptionCache.h"
  32 #include "llvm/Analysis/BlockFrequencyInfo.h"
  33 #include "llvm/Analysis/GlobalsModRef.h"
  34 #include "llvm/Analysis/LazyBlockFrequencyInfo.h"
  35 #include "llvm/Analysis/LoopAccessAnalysis.h"
  36 #include "llvm/Analysis/LoopAnalysisManager.h"
  37 #include "llvm/Analysis/LoopInfo.h"
  38 #include "llvm/Analysis/MemorySSA.h"
  39 #include "llvm/Analysis/ProfileSummaryInfo.h"
  40 #include "llvm/Analysis/ScalarEvolution.h"
  41 #include "llvm/Analysis/ScalarEvolutionExpander.h"
  42 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
  43 #include "llvm/Analysis/TargetLibraryInfo.h"
  44 #include "llvm/Analysis/TargetTransformInfo.h"
  45 #include "llvm/IR/DataLayout.h"
  46 #include "llvm/IR/Dominators.h"
  47 #include "llvm/IR/Instructions.h"
  48 #include "llvm/IR/Module.h"
  49 #include "llvm/IR/PassManager.h"
  50 #include "llvm/IR/Type.h"
  51 #include "llvm/IR/Value.h"
  52 #include "llvm/Pass.h"
  53 #include "llvm/Support/Casting.h"
  54 #include "llvm/Support/CommandLine.h"
  55 #include "llvm/Support/Debug.h"
  56 #include "llvm/Support/raw_ostream.h"
  57 #include "llvm/Transforms/Scalar.h"
  58 #include "llvm/Transforms/Utils.h"
  59 #include "llvm/Transforms/Utils/LoopVersioning.h"
  60 #include "llvm/Transforms/Utils/SizeOpts.h"
  61 #include <algorithm>
  62 #include <cassert>
  63 #include <forward_list>
  64 #include <set>
  65 #include <tuple>
  66 #include <utility>
  67
  68 using namespace llvm;
  69
  70 #define LLE_OPTION "loop-load-elim"
  71 #define DEBUG_TYPE LLE_OPTION
  72
  73 static cl::opt<unsigned> CheckPerElim(
  74     "runtime-check-per-loop-load-elim", cl::Hidden,
  75     cl::desc("Max number of memchecks allowed per eliminated load on average"),
  76     cl::init(1));
  77
  78 static cl::opt<unsigned> LoadElimSCEVCheckThreshold(
  79     "loop-load-elimination-scev-check-threshold", cl::init(8), cl::Hidden,
  80     cl::desc("The maximum number of SCEV checks allowed for Loop "
  81              "Load Elimination"));
  82
  83 STATISTIC(NumLoopLoadEliminted, "Number of loads eliminated by LLE");
  84
  85 namespace {
  86
  87 /// Represent a store-to-forwarding candidate.
  88 struct StoreToLoadForwardingCandidate {
  89   LoadInst *Load;
  90   StoreInst *Store;
  91
  92   StoreToLoadForwardingCandidate(LoadInst *Load, StoreInst *Store)
  93       : Load(Load), Store(Store) {}
  94
  95   /// Return true if the dependence from the store to the load has a
  96   /// distance of one.  E.g. A[i+1] = A[i]
  97   bool isDependenceDistanceOfOne(PredicatedScalarEvolution &PSE,
  98                                  Loop *L) const {
  99     Value *LoadPtr = Load->getPointerOperand();
 100     Value *StorePtr = Store->getPointerOperand();
 101     Type *LoadPtrType = LoadPtr->getType();
 102     Type *LoadType = LoadPtrType->getPointerElementType();
 103
 104     assert(LoadPtrType->getPointerAddressSpace() ==
 105                StorePtr->getType()->getPointerAddressSpace() &&
 106            LoadType == StorePtr->getType()->getPointerElementType() &&
 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 SmallPtrSet<Value *, 4> &PtrsWrittenOnFwdingPath,
 311                      const std::set<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<RuntimePointerChecking::PointerCheck, 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     // FIXME: SmallPtrSet does not work with std::inserter.
 387     std::set<Value *> CandLoadPtrs;
 388     transform(Candidates,
 389                    std::inserter(CandLoadPtrs, CandLoadPtrs.begin()),
 390                    std::mem_fn(&StoreToLoadForwardingCandidate::getLoadPtr));
 391
 392     const auto &AllChecks = LAI.getRuntimePointerChecking()->getChecks();
 393     SmallVector<RuntimePointerChecking::PointerCheck, 4> Checks;
 394
 395     copy_if(AllChecks, std::back_inserter(Checks),
 396             [&](const RuntimePointerChecking::PointerCheck &Check) {
 397               for (auto PtrIdx1 : Check.first->Members)
 398                 for (auto PtrIdx2 : Check.second->Members)
 399                   if (needsChecking(PtrIdx1, PtrIdx2, PtrsWrittenOnFwdingPath,
 400                                     CandLoadPtrs))
 401                     return true;
 402               return false;
 403             });
 404
 405     LLVM_DEBUG(dbgs() << "\nPointer Checks (count: " << Checks.size()
 406                       << "):\n");
 407     LLVM_DEBUG(LAI.getRuntimePointerChecking()->printChecks(dbgs(), Checks));
 408
 409     return Checks;
 410   }
 411
 412   /// Perform the transformation for a candidate.
 413   void
 414   propagateStoredValueToLoadUsers(const StoreToLoadForwardingCandidate &Cand,
 415                                   SCEVExpander &SEE) {
 416     // loop:
 417     //      %x = load %gep_i
 418     //         = ... %x
 419     //      store %y, %gep_i_plus_1
 420     //
 421     // =>
 422     //
 423     // ph:
 424     //      %x.initial = load %gep_0
 425     // loop:
 426     //      %x.storeforward = phi [%x.initial, %ph] [%y, %loop]
 427     //      %x = load %gep_i            <---- now dead
 428     //         = ... %x.storeforward
 429     //      store %y, %gep_i_plus_1
 430
 431     Value *Ptr = Cand.Load->getPointerOperand();
 432     auto *PtrSCEV = cast<SCEVAddRecExpr>(PSE.getSCEV(Ptr));
 433     auto *PH = L->getLoopPreheader();
 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, MaybeAlign(Cand.Load->getAlignment()),
 439         PH->getTerminator());
 440
 441     PHINode *PHI = PHINode::Create(Initial->getType(), 2, "store_forwarded",
 442                                    &L->getHeader()->front());
 443     PHI->addIncoming(Initial, PH);
 444     PHI->addIncoming(Cand.Store->getOperand(0), L->getLoopLatch());
 445
 446     Cand.Load->replaceAllUsesWith(PHI);
 447   }
 448
 449   /// Top-level driver for each loop: find store->load forwarding
 450   /// candidates, add run-time checks and perform transformation.
 451   bool processLoop() {
 452     LLVM_DEBUG(dbgs() << "\nIn \"" << L->getHeader()->getParent()->getName()
 453                       << "\" checking " << *L << "\n");
 454
 455     // Look for store-to-load forwarding cases across the
 456     // backedge. E.g.:
 457     //
 458     // loop:
 459     //      %x = load %gep_i
 460     //         = ... %x
 461     //      store %y, %gep_i_plus_1
 462     //
 463     // =>
 464     //
 465     // ph:
 466     //      %x.initial = load %gep_0
 467     // loop:
 468     //      %x.storeforward = phi [%x.initial, %ph] [%y, %loop]
 469     //      %x = load %gep_i            <---- now dead
 470     //         = ... %x.storeforward
 471     //      store %y, %gep_i_plus_1
 472
 473     // First start with store->load dependences.
 474     auto StoreToLoadDependences = findStoreToLoadDependences(LAI);
 475     if (StoreToLoadDependences.empty())
 476       return false;
 477
 478     // Generate an index for each load and store according to the original
 479     // program order.  This will be used later.
 480     InstOrder = LAI.getDepChecker().generateInstructionOrderMap();
 481
 482     // To keep things simple for now, remove those where the load is potentially
 483     // fed by multiple stores.
 484     removeDependencesFromMultipleStores(StoreToLoadDependences);
 485     if (StoreToLoadDependences.empty())
 486       return false;
 487
 488     // Filter the candidates further.
 489     SmallVector<StoreToLoadForwardingCandidate, 4> Candidates;
 490     unsigned NumForwarding = 0;
 491     for (const StoreToLoadForwardingCandidate Cand : StoreToLoadDependences) {
 492       LLVM_DEBUG(dbgs() << "Candidate " << Cand);
 493
 494       // Make sure that the stored values is available everywhere in the loop in
 495       // the next iteration.
 496       if (!doesStoreDominatesAllLatches(Cand.Store->getParent(), L, DT))
 497         continue;
 498
 499       // If the load is conditional we can't hoist its 0-iteration instance to
 500       // the preheader because that would make it unconditional.  Thus we would
 501       // access a memory location that the original loop did not access.
 502       if (isLoadConditional(Cand.Load, L))
 503         continue;
 504
 505       // Check whether the SCEV difference is the same as the induction step,
 506       // thus we load the value in the next iteration.
 507       if (!Cand.isDependenceDistanceOfOne(PSE, L))
 508         continue;
 509
 510       ++NumForwarding;
 511       LLVM_DEBUG(
 512           dbgs()
 513           << NumForwarding
 514           << ". Valid store-to-load forwarding across the loop backedge\n");
 515       Candidates.push_back(Cand);
 516     }
 517     if (Candidates.empty())
 518       return false;
 519
 520     // Check intervening may-alias stores.  These need runtime checks for alias
 521     // disambiguation.
 522     SmallVector<RuntimePointerChecking::PointerCheck, 4> Checks =
 523         collectMemchecks(Candidates);
 524
 525     // Too many checks are likely to outweigh the benefits of forwarding.
 526     if (Checks.size() > Candidates.size() * CheckPerElim) {
 527       LLVM_DEBUG(dbgs() << "Too many run-time checks needed.\n");
 528       return false;
 529     }
 530
 531     if (LAI.getPSE().getUnionPredicate().getComplexity() >
 532         LoadElimSCEVCheckThreshold) {
 533       LLVM_DEBUG(dbgs() << "Too many SCEV run-time checks needed.\n");
 534       return false;
 535     }
 536
 537     if (!Checks.empty() || !LAI.getPSE().getUnionPredicate().isAlwaysTrue()) {
 538       if (LAI.hasConvergentOp()) {
 539         LLVM_DEBUG(dbgs() << "Versioning is needed but not allowed with "
 540                              "convergent calls\n");
 541         return false;
 542       }
 543
 544       auto *HeaderBB = L->getHeader();
 545       auto *F = HeaderBB->getParent();
 546       bool OptForSize = F->hasOptSize() ||
 547                         llvm::shouldOptimizeForSize(HeaderBB, PSI, BFI);
 548       if (OptForSize) {
 549         LLVM_DEBUG(
 550             dbgs() << "Versioning is needed but not allowed when optimizing "
 551                       "for size.\n");
 552         return false;
 553       }
 554
 555       if (!L->isLoopSimplifyForm()) {
 556         LLVM_DEBUG(dbgs() << "Loop is not is loop-simplify form");
 557         return false;
 558       }
 559
 560       // Point of no-return, start the transformation.  First, version the loop
 561       // if necessary.
 562
 563       LoopVersioning LV(LAI, L, LI, DT, PSE.getSE(), false);
 564       LV.setAliasChecks(std::move(Checks));
 565       LV.setSCEVChecks(LAI.getPSE().getUnionPredicate());
 566       LV.versionLoop();
 567     }
 568
 569     // Next, propagate the value stored by the store to the users of the load.
 570     // Also for the first iteration, generate the initial value of the load.
 571     SCEVExpander SEE(*PSE.getSE(), L->getHeader()->getModule()->getDataLayout(),
 572                      "storeforward");
 573     for (const auto &Cand : Candidates)
 574       propagateStoredValueToLoadUsers(Cand, SEE);
 575     NumLoopLoadEliminted += NumForwarding;
 576
 577     return true;
 578   }
 579
 580 private:
 581   Loop *L;
 582
 583   /// Maps the load/store instructions to their index according to
 584   /// program order.
 585   DenseMap<Instruction *, unsigned> InstOrder;
 586
 587   // Analyses used.
 588   LoopInfo *LI;
 589   const LoopAccessInfo &LAI;
 590   DominatorTree *DT;
 591   BlockFrequencyInfo *BFI;
 592   ProfileSummaryInfo *PSI;
 593   PredicatedScalarEvolution PSE;
 594 };
 595
 596 } // end anonymous namespace
 597
 598 static bool
 599 eliminateLoadsAcrossLoops(Function &F, LoopInfo &LI, DominatorTree &DT,
 600                           BlockFrequencyInfo *BFI, ProfileSummaryInfo *PSI,
 601                           function_ref<const LoopAccessInfo &(Loop &)> GetLAI) {
 602   // Build up a worklist of inner-loops to transform to avoid iterator
 603   // invalidation.
 604   // FIXME: This logic comes from other passes that actually change the loop
 605   // nest structure. It isn't clear this is necessary (or useful) for a pass
 606   // which merely optimizes the use of loads in a loop.
 607   SmallVector<Loop *, 8> Worklist;
 608
 609   for (Loop *TopLevelLoop : LI)
 610     for (Loop *L : depth_first(TopLevelLoop))
 611       // We only handle inner-most loops.
 612       if (L->empty())
 613         Worklist.push_back(L);
 614
 615   // Now walk the identified inner loops.
 616   bool Changed = false;
 617   for (Loop *L : Worklist) {
 618     // The actual work is performed by LoadEliminationForLoop.
 619     LoadEliminationForLoop LEL(L, &LI, GetLAI(*L), &DT, BFI, PSI);
 620     Changed |= LEL.processLoop();
 621   }
 622   return Changed;
 623 }
 624
 625 namespace {
 626
 627 /// The pass.  Most of the work is delegated to the per-loop
 628 /// LoadEliminationForLoop class.
 629 class LoopLoadElimination : public FunctionPass {
 630 public:
 631   static char ID;
 632
 633   LoopLoadElimination() : FunctionPass(ID) {
 634     initializeLoopLoadEliminationPass(*PassRegistry::getPassRegistry());
 635   }
 636
 637   bool runOnFunction(Function &F) override {
 638     if (skipFunction(F))
 639       return false;
 640
 641     auto &LI = getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
 642     auto &LAA = getAnalysis<LoopAccessLegacyAnalysis>();
 643     auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
 644     auto *PSI = &getAnalysis<ProfileSummaryInfoWrapperPass>().getPSI();
 645     auto *BFI = (PSI && PSI->hasProfileSummary()) ?
 646                 &getAnalysis<LazyBlockFrequencyInfoPass>().getBFI() :
 647                 nullptr;
 648
 649     // Process each loop nest in the function.
 650     return eliminateLoadsAcrossLoops(
 651         F, LI, DT, BFI, PSI,
 652         [&LAA](Loop &L) -> const LoopAccessInfo & { return LAA.getInfo(&L); });
 653   }
 654
 655   void getAnalysisUsage(AnalysisUsage &AU) const override {
 656     AU.addRequiredID(LoopSimplifyID);
 657     AU.addRequired<LoopInfoWrapperPass>();
 658     AU.addPreserved<LoopInfoWrapperPass>();
 659     AU.addRequired<LoopAccessLegacyAnalysis>();
 660     AU.addRequired<ScalarEvolutionWrapperPass>();
 661     AU.addRequired<DominatorTreeWrapperPass>();
 662     AU.addPreserved<DominatorTreeWrapperPass>();
 663     AU.addPreserved<GlobalsAAWrapperPass>();
 664     AU.addRequired<ProfileSummaryInfoWrapperPass>();
 665     LazyBlockFrequencyInfoPass::getLazyBFIAnalysisUsage(AU);
 666   }
 667 };
 668
 669 } // end anonymous namespace
 670
 671 char LoopLoadElimination::ID;
 672
 673 static const char LLE_name[] = "Loop Load Elimination";
 674
 675 INITIALIZE_PASS_BEGIN(LoopLoadElimination, LLE_OPTION, LLE_name, false, false)
 676 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
 677 INITIALIZE_PASS_DEPENDENCY(LoopAccessLegacyAnalysis)
 678 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
 679 INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass)
 680 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
 681 INITIALIZE_PASS_DEPENDENCY(ProfileSummaryInfoWrapperPass)
 682 INITIALIZE_PASS_DEPENDENCY(LazyBlockFrequencyInfoPass)
 683 INITIALIZE_PASS_END(LoopLoadElimination, LLE_OPTION, LLE_name, false, false)
 684
 685 FunctionPass *llvm::createLoopLoadEliminationPass() {
 686   return new LoopLoadElimination();
 687 }
 688
 689 PreservedAnalyses LoopLoadEliminationPass::run(Function &F,
 690                                                FunctionAnalysisManager &AM) {
 691   auto &SE = AM.getResult<ScalarEvolutionAnalysis>(F);
 692   auto &LI = AM.getResult<LoopAnalysis>(F);
 693   auto &TTI = AM.getResult<TargetIRAnalysis>(F);
 694   auto &DT = AM.getResult<DominatorTreeAnalysis>(F);
 695   auto &TLI = AM.getResult<TargetLibraryAnalysis>(F);
 696   auto &AA = AM.getResult<AAManager>(F);
 697   auto &AC = AM.getResult<AssumptionAnalysis>(F);
 698   auto &MAM = AM.getResult<ModuleAnalysisManagerFunctionProxy>(F).getManager();
 699   auto *PSI = MAM.getCachedResult<ProfileSummaryAnalysis>(*F.getParent());
 700   auto *BFI = (PSI && PSI->hasProfileSummary()) ?
 701       &AM.getResult<BlockFrequencyAnalysis>(F) : nullptr;
 702   MemorySSA *MSSA = EnableMSSALoopDependency
 703                         ? &AM.getResult<MemorySSAAnalysis>(F).getMSSA()
 704                         : nullptr;
 705
 706   auto &LAM = AM.getResult<LoopAnalysisManagerFunctionProxy>(F).getManager();
 707   bool Changed = eliminateLoadsAcrossLoops(
 708       F, LI, DT, BFI, PSI, [&](Loop &L) -> const LoopAccessInfo & {
 709         LoopStandardAnalysisResults AR = {AA, AC, DT, LI, SE, TLI, TTI, MSSA};
 710         return LAM.getResult<LoopAccessAnalysis>(L, AR);
 711       });
 712
 713   if (!Changed)
 714     return PreservedAnalyses::all();
 715
 716   PreservedAnalyses PA;
 717   return PA;
 718 }