llvm.org GIT mirror llvm / 7716d65 lib / Transforms / Scalar / LoopLoadElimination.cpp
7716d65

Tree @7716d65 (Download .tar.gz)

LoopLoadElimination.cpp @7716d65raw · history · blame

  1
  2
  3
  4
  5
  6
  7
  8
  9
 10
 11
 12
 13
 14
 15
 16
 17
 18
 19
 20
 21
 22
 23
 24
 25
 26
 27
 28
 29
 30
 31
 32
 33
 34
 35
 36
 37
 38
 39
 40
 41
 42
 43
 44
 45
 46
 47
 48
 49
 50
 51
 52
 53
 54
 55
 56
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
//===- LoopLoadElimination.cpp - Loop Load Elimination Pass ---------------===//
//
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
// See https://llvm.org/LICENSE.txt for license information.
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
//
//===----------------------------------------------------------------------===//
//
// This file implement a loop-aware load elimination pass.
//
// It uses LoopAccessAnalysis to identify loop-carried dependences with a
// distance of one between stores and loads.  These form the candidates for the
// transformation.  The source value of each store then propagated to the user
// of the corresponding load.  This makes the load dead.
//
// The pass can also version the loop and add memchecks in order to prove that
// may-aliasing stores can't change the value in memory before it's read by the
// load.
//
//===----------------------------------------------------------------------===//

#include "llvm/Transforms/Scalar/LoopLoadElimination.h"
#include "llvm/ADT/APInt.h"
#include "llvm/ADT/DenseMap.h"
#include "llvm/ADT/DepthFirstIterator.h"
#include "llvm/ADT/STLExtras.h"
#include "llvm/ADT/SmallPtrSet.h"
#include "llvm/ADT/SmallVector.h"
#include "llvm/ADT/Statistic.h"
#include "llvm/Analysis/AliasAnalysis.h"
#include "llvm/Analysis/AssumptionCache.h"
#include "llvm/Analysis/BlockFrequencyInfo.h"
#include "llvm/Analysis/GlobalsModRef.h"
#include "llvm/Analysis/LazyBlockFrequencyInfo.h"
#include "llvm/Analysis/LoopAccessAnalysis.h"
#include "llvm/Analysis/LoopAnalysisManager.h"
#include "llvm/Analysis/LoopInfo.h"
#include "llvm/Analysis/MemorySSA.h"
#include "llvm/Analysis/ProfileSummaryInfo.h"
#include "llvm/Analysis/ScalarEvolution.h"
#include "llvm/Analysis/ScalarEvolutionExpander.h"
#include "llvm/Analysis/ScalarEvolutionExpressions.h"
#include "llvm/Analysis/TargetLibraryInfo.h"
#include "llvm/Analysis/TargetTransformInfo.h"
#include "llvm/IR/DataLayout.h"
#include "llvm/IR/Dominators.h"
#include "llvm/IR/Instructions.h"
#include "llvm/IR/Module.h"
#include "llvm/IR/PassManager.h"
#include "llvm/IR/Type.h"
#include "llvm/IR/Value.h"
#include "llvm/Pass.h"
#include "llvm/Support/Casting.h"
#include "llvm/Support/CommandLine.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/raw_ostream.h"
#include "llvm/Transforms/Scalar.h"
#include "llvm/Transforms/Utils.h"
#include "llvm/Transforms/Utils/LoopVersioning.h"
#include "llvm/Transforms/Utils/SizeOpts.h"
#include <algorithm>
#include <cassert>
#include <forward_list>
#include <set>
#include <tuple>
#include <utility>

using namespace llvm;

#define LLE_OPTION "loop-load-elim"
#define DEBUG_TYPE LLE_OPTION

static cl::opt<unsigned> CheckPerElim(
    "runtime-check-per-loop-load-elim", cl::Hidden,
    cl::desc("Max number of memchecks allowed per eliminated load on average"),
    cl::init(1));

static cl::opt<unsigned> LoadElimSCEVCheckThreshold(
    "loop-load-elimination-scev-check-threshold", cl::init(8), cl::Hidden,
    cl::desc("The maximum number of SCEV checks allowed for Loop "
             "Load Elimination"));

STATISTIC(NumLoopLoadEliminted, "Number of loads eliminated by LLE");

namespace {

/// Represent a store-to-forwarding candidate.
struct StoreToLoadForwardingCandidate {
  LoadInst *Load;
  StoreInst *Store;

  StoreToLoadForwardingCandidate(LoadInst *Load, StoreInst *Store)
      : Load(Load), Store(Store) {}

  /// Return true if the dependence from the store to the load has a
  /// distance of one.  E.g. A[i+1] = A[i]
  bool isDependenceDistanceOfOne(PredicatedScalarEvolution &PSE,
                                 Loop *L) const {
    Value *LoadPtr = Load->getPointerOperand();
    Value *StorePtr = Store->getPointerOperand();
    Type *LoadPtrType = LoadPtr->getType();
    Type *LoadType = LoadPtrType->getPointerElementType();

    assert(LoadPtrType->getPointerAddressSpace() ==
               StorePtr->getType()->getPointerAddressSpace() &&
           LoadType == StorePtr->getType()->getPointerElementType() &&
           "Should be a known dependence");

    // Currently we only support accesses with unit stride.  FIXME: we should be
    // able to handle non unit stirde as well as long as the stride is equal to
    // the dependence distance.
    if (getPtrStride(PSE, LoadPtr, L) != 1 ||
        getPtrStride(PSE, StorePtr, L) != 1)
      return false;

    auto &DL = Load->getParent()->getModule()->getDataLayout();
    unsigned TypeByteSize = DL.getTypeAllocSize(const_cast<Type *>(LoadType));

    auto *LoadPtrSCEV = cast<SCEVAddRecExpr>(PSE.getSCEV(LoadPtr));
    auto *StorePtrSCEV = cast<SCEVAddRecExpr>(PSE.getSCEV(StorePtr));

    // We don't need to check non-wrapping here because forward/backward
    // dependence wouldn't be valid if these weren't monotonic accesses.
    auto *Dist = cast<SCEVConstant>(
        PSE.getSE()->getMinusSCEV(StorePtrSCEV, LoadPtrSCEV));
    const APInt &Val = Dist->getAPInt();
    return Val == TypeByteSize;
  }

  Value *getLoadPtr() const { return Load->getPointerOperand(); }

#ifndef NDEBUG
  friend raw_ostream &operator<<(raw_ostream &OS,
                                 const StoreToLoadForwardingCandidate &Cand) {
    OS << *Cand.Store << " -->\n";
    OS.indent(2) << *Cand.Load << "\n";
    return OS;
  }
#endif
};

} // end anonymous namespace

/// Check if the store dominates all latches, so as long as there is no
/// intervening store this value will be loaded in the next iteration.
static bool doesStoreDominatesAllLatches(BasicBlock *StoreBlock, Loop *L,
                                         DominatorTree *DT) {
  SmallVector<BasicBlock *, 8> Latches;
  L->getLoopLatches(Latches);
  return llvm::all_of(Latches, [&](const BasicBlock *Latch) {
    return DT->dominates(StoreBlock, Latch);
  });
}

/// Return true if the load is not executed on all paths in the loop.
static bool isLoadConditional(LoadInst *Load, Loop *L) {
  return Load->getParent() != L->getHeader();
}

namespace {

/// The per-loop class that does most of the work.
class LoadEliminationForLoop {
public:
  LoadEliminationForLoop(Loop *L, LoopInfo *LI, const LoopAccessInfo &LAI,
                         DominatorTree *DT, BlockFrequencyInfo *BFI,
                         ProfileSummaryInfo* PSI)
      : L(L), LI(LI), LAI(LAI), DT(DT), BFI(BFI), PSI(PSI), PSE(LAI.getPSE()) {}

  /// Look through the loop-carried and loop-independent dependences in
  /// this loop and find store->load dependences.
  ///
  /// Note that no candidate is returned if LAA has failed to analyze the loop
  /// (e.g. if it's not bottom-tested, contains volatile memops, etc.)
  std::forward_list<StoreToLoadForwardingCandidate>
  findStoreToLoadDependences(const LoopAccessInfo &LAI) {
    std::forward_list<StoreToLoadForwardingCandidate> Candidates;

    const auto *Deps = LAI.getDepChecker().getDependences();
    if (!Deps)
      return Candidates;

    // Find store->load dependences (consequently true dep).  Both lexically
    // forward and backward dependences qualify.  Disqualify loads that have
    // other unknown dependences.

    SmallPtrSet<Instruction *, 4> LoadsWithUnknownDepedence;

    for (const auto &Dep : *Deps) {
      Instruction *Source = Dep.getSource(LAI);
      Instruction *Destination = Dep.getDestination(LAI);

      if (Dep.Type == MemoryDepChecker::Dependence::Unknown) {
        if (isa<LoadInst>(Source))
          LoadsWithUnknownDepedence.insert(Source);
        if (isa<LoadInst>(Destination))
          LoadsWithUnknownDepedence.insert(Destination);
        continue;
      }

      if (Dep.isBackward())
        // Note that the designations source and destination follow the program
        // order, i.e. source is always first.  (The direction is given by the
        // DepType.)
        std::swap(Source, Destination);
      else
        assert(Dep.isForward() && "Needs to be a forward dependence");

      auto *Store = dyn_cast<StoreInst>(Source);
      if (!Store)
        continue;
      auto *Load = dyn_cast<LoadInst>(Destination);
      if (!Load)
        continue;

      // Only progagate the value if they are of the same type.
      if (Store->getPointerOperandType() != Load->getPointerOperandType())
        continue;

      Candidates.emplace_front(Load, Store);
    }

    if (!LoadsWithUnknownDepedence.empty())
      Candidates.remove_if([&](const StoreToLoadForwardingCandidate &C) {
        return LoadsWithUnknownDepedence.count(C.Load);
      });

    return Candidates;
  }

  /// Return the index of the instruction according to program order.
  unsigned getInstrIndex(Instruction *Inst) {
    auto I = InstOrder.find(Inst);
    assert(I != InstOrder.end() && "No index for instruction");
    return I->second;
  }

  /// If a load has multiple candidates associated (i.e. different
  /// stores), it means that it could be forwarding from multiple stores
  /// depending on control flow.  Remove these candidates.
  ///
  /// Here, we rely on LAA to include the relevant loop-independent dependences.
  /// LAA is known to omit these in the very simple case when the read and the
  /// write within an alias set always takes place using the *same* pointer.
  ///
  /// However, we know that this is not the case here, i.e. we can rely on LAA
  /// to provide us with loop-independent dependences for the cases we're
  /// interested.  Consider the case for example where a loop-independent
  /// dependece S1->S2 invalidates the forwarding S3->S2.
  ///
  ///         A[i]   = ...   (S1)
  ///         ...    = A[i]  (S2)
  ///         A[i+1] = ...   (S3)
  ///
  /// LAA will perform dependence analysis here because there are two
  /// *different* pointers involved in the same alias set (&A[i] and &A[i+1]).
  void removeDependencesFromMultipleStores(
      std::forward_list<StoreToLoadForwardingCandidate> &Candidates) {
    // If Store is nullptr it means that we have multiple stores forwarding to
    // this store.
    using LoadToSingleCandT =
        DenseMap<LoadInst *, const StoreToLoadForwardingCandidate *>;
    LoadToSingleCandT LoadToSingleCand;

    for (const auto &Cand : Candidates) {
      bool NewElt;
      LoadToSingleCandT::iterator Iter;

      std::tie(Iter, NewElt) =
          LoadToSingleCand.insert(std::make_pair(Cand.Load, &Cand));
      if (!NewElt) {
        const StoreToLoadForwardingCandidate *&OtherCand = Iter->second;
        // Already multiple stores forward to this load.
        if (OtherCand == nullptr)
          continue;

        // Handle the very basic case when the two stores are in the same block
        // so deciding which one forwards is easy.  The later one forwards as
        // long as they both have a dependence distance of one to the load.
        if (Cand.Store->getParent() == OtherCand->Store->getParent() &&
            Cand.isDependenceDistanceOfOne(PSE, L) &&
            OtherCand->isDependenceDistanceOfOne(PSE, L)) {
          // They are in the same block, the later one will forward to the load.
          if (getInstrIndex(OtherCand->Store) < getInstrIndex(Cand.Store))
            OtherCand = &Cand;
        } else
          OtherCand = nullptr;
      }
    }

    Candidates.remove_if([&](const StoreToLoadForwardingCandidate &Cand) {
      if (LoadToSingleCand[Cand.Load] != &Cand) {
        LLVM_DEBUG(
            dbgs() << "Removing from candidates: \n"
                   << Cand
                   << "  The load may have multiple stores forwarding to "
                   << "it\n");
        return true;
      }
      return false;
    });
  }

  /// Given two pointers operations by their RuntimePointerChecking
  /// indices, return true if they require an alias check.
  ///
  /// We need a check if one is a pointer for a candidate load and the other is
  /// a pointer for a possibly intervening store.
  bool needsChecking(unsigned PtrIdx1, unsigned PtrIdx2,
                     const SmallPtrSet<Value *, 4> &PtrsWrittenOnFwdingPath,
                     const std::set<Value *> &CandLoadPtrs) {
    Value *Ptr1 =
        LAI.getRuntimePointerChecking()->getPointerInfo(PtrIdx1).PointerValue;
    Value *Ptr2 =
        LAI.getRuntimePointerChecking()->getPointerInfo(PtrIdx2).PointerValue;
    return ((PtrsWrittenOnFwdingPath.count(Ptr1) && CandLoadPtrs.count(Ptr2)) ||
            (PtrsWrittenOnFwdingPath.count(Ptr2) && CandLoadPtrs.count(Ptr1)));
  }

  /// Return pointers that are possibly written to on the path from a
  /// forwarding store to a load.
  ///
  /// These pointers need to be alias-checked against the forwarding candidates.
  SmallPtrSet<Value *, 4> findPointersWrittenOnForwardingPath(
      const SmallVectorImpl<StoreToLoadForwardingCandidate> &Candidates) {
    // From FirstStore to LastLoad neither of the elimination candidate loads
    // should overlap with any of the stores.
    //
    // E.g.:
    //
    // st1 C[i]
    // ld1 B[i] <-------,
    // ld0 A[i] <----,  |              * LastLoad
    // ...           |  |
    // st2 E[i]      |  |
    // st3 B[i+1] -- | -'              * FirstStore
    // st0 A[i+1] ---'
    // st4 D[i]
    //
    // st0 forwards to ld0 if the accesses in st4 and st1 don't overlap with
    // ld0.

    LoadInst *LastLoad =
        std::max_element(Candidates.begin(), Candidates.end(),
                         [&](const StoreToLoadForwardingCandidate &A,
                             const StoreToLoadForwardingCandidate &B) {
                           return getInstrIndex(A.Load) < getInstrIndex(B.Load);
                         })
            ->Load;
    StoreInst *FirstStore =
        std::min_element(Candidates.begin(), Candidates.end(),
                         [&](const StoreToLoadForwardingCandidate &A,
                             const StoreToLoadForwardingCandidate &B) {
                           return getInstrIndex(A.Store) <
                                  getInstrIndex(B.Store);
                         })
            ->Store;

    // We're looking for stores after the first forwarding store until the end
    // of the loop, then from the beginning of the loop until the last
    // forwarded-to load.  Collect the pointer for the stores.
    SmallPtrSet<Value *, 4> PtrsWrittenOnFwdingPath;

    auto InsertStorePtr = [&](Instruction *I) {
      if (auto *S = dyn_cast<StoreInst>(I))
        PtrsWrittenOnFwdingPath.insert(S->getPointerOperand());
    };
    const auto &MemInstrs = LAI.getDepChecker().getMemoryInstructions();
    std::for_each(MemInstrs.begin() + getInstrIndex(FirstStore) + 1,
                  MemInstrs.end(), InsertStorePtr);
    std::for_each(MemInstrs.begin(), &MemInstrs[getInstrIndex(LastLoad)],
                  InsertStorePtr);

    return PtrsWrittenOnFwdingPath;
  }

  /// Determine the pointer alias checks to prove that there are no
  /// intervening stores.
  SmallVector<RuntimePointerChecking::PointerCheck, 4> collectMemchecks(
      const SmallVectorImpl<StoreToLoadForwardingCandidate> &Candidates) {

    SmallPtrSet<Value *, 4> PtrsWrittenOnFwdingPath =
        findPointersWrittenOnForwardingPath(Candidates);

    // Collect the pointers of the candidate loads.
    // FIXME: SmallPtrSet does not work with std::inserter.
    std::set<Value *> CandLoadPtrs;
    transform(Candidates,
                   std::inserter(CandLoadPtrs, CandLoadPtrs.begin()),
                   std::mem_fn(&StoreToLoadForwardingCandidate::getLoadPtr));

    const auto &AllChecks = LAI.getRuntimePointerChecking()->getChecks();
    SmallVector<RuntimePointerChecking::PointerCheck, 4> Checks;

    copy_if(AllChecks, std::back_inserter(Checks),
            [&](const RuntimePointerChecking::PointerCheck &Check) {
              for (auto PtrIdx1 : Check.first->Members)
                for (auto PtrIdx2 : Check.second->Members)
                  if (needsChecking(PtrIdx1, PtrIdx2, PtrsWrittenOnFwdingPath,
                                    CandLoadPtrs))
                    return true;
              return false;
            });

    LLVM_DEBUG(dbgs() << "\nPointer Checks (count: " << Checks.size()
                      << "):\n");
    LLVM_DEBUG(LAI.getRuntimePointerChecking()->printChecks(dbgs(), Checks));

    return Checks;
  }

  /// Perform the transformation for a candidate.
  void
  propagateStoredValueToLoadUsers(const StoreToLoadForwardingCandidate &Cand,
                                  SCEVExpander &SEE) {
    // loop:
    //      %x = load %gep_i
    //         = ... %x
    //      store %y, %gep_i_plus_1
    //
    // =>
    //
    // ph:
    //      %x.initial = load %gep_0
    // loop:
    //      %x.storeforward = phi [%x.initial, %ph] [%y, %loop]
    //      %x = load %gep_i            <---- now dead
    //         = ... %x.storeforward
    //      store %y, %gep_i_plus_1

    Value *Ptr = Cand.Load->getPointerOperand();
    auto *PtrSCEV = cast<SCEVAddRecExpr>(PSE.getSCEV(Ptr));
    auto *PH = L->getLoopPreheader();
    Value *InitialPtr = SEE.expandCodeFor(PtrSCEV->getStart(), Ptr->getType(),
                                          PH->getTerminator());
    Value *Initial = new LoadInst(
        Cand.Load->getType(), InitialPtr, "load_initial",
        /* isVolatile */ false, Cand.Load->getAlignment(), PH->getTerminator());

    PHINode *PHI = PHINode::Create(Initial->getType(), 2, "store_forwarded",
                                   &L->getHeader()->front());
    PHI->addIncoming(Initial, PH);
    PHI->addIncoming(Cand.Store->getOperand(0), L->getLoopLatch());

    Cand.Load->replaceAllUsesWith(PHI);
  }

  /// Top-level driver for each loop: find store->load forwarding
  /// candidates, add run-time checks and perform transformation.
  bool processLoop() {
    LLVM_DEBUG(dbgs() << "\nIn \"" << L->getHeader()->getParent()->getName()
                      << "\" checking " << *L << "\n");

    // Look for store-to-load forwarding cases across the
    // backedge. E.g.:
    //
    // loop:
    //      %x = load %gep_i
    //         = ... %x
    //      store %y, %gep_i_plus_1
    //
    // =>
    //
    // ph:
    //      %x.initial = load %gep_0
    // loop:
    //      %x.storeforward = phi [%x.initial, %ph] [%y, %loop]
    //      %x = load %gep_i            <---- now dead
    //         = ... %x.storeforward
    //      store %y, %gep_i_plus_1

    // First start with store->load dependences.
    auto StoreToLoadDependences = findStoreToLoadDependences(LAI);
    if (StoreToLoadDependences.empty())
      return false;

    // Generate an index for each load and store according to the original
    // program order.  This will be used later.
    InstOrder = LAI.getDepChecker().generateInstructionOrderMap();

    // To keep things simple for now, remove those where the load is potentially
    // fed by multiple stores.
    removeDependencesFromMultipleStores(StoreToLoadDependences);
    if (StoreToLoadDependences.empty())
      return false;

    // Filter the candidates further.
    SmallVector<StoreToLoadForwardingCandidate, 4> Candidates;
    unsigned NumForwarding = 0;
    for (const StoreToLoadForwardingCandidate Cand : StoreToLoadDependences) {
      LLVM_DEBUG(dbgs() << "Candidate " << Cand);

      // Make sure that the stored values is available everywhere in the loop in
      // the next iteration.
      if (!doesStoreDominatesAllLatches(Cand.Store->getParent(), L, DT))
        continue;

      // If the load is conditional we can't hoist its 0-iteration instance to
      // the preheader because that would make it unconditional.  Thus we would
      // access a memory location that the original loop did not access.
      if (isLoadConditional(Cand.Load, L))
        continue;

      // Check whether the SCEV difference is the same as the induction step,
      // thus we load the value in the next iteration.
      if (!Cand.isDependenceDistanceOfOne(PSE, L))
        continue;

      ++NumForwarding;
      LLVM_DEBUG(
          dbgs()
          << NumForwarding
          << ". Valid store-to-load forwarding across the loop backedge\n");
      Candidates.push_back(Cand);
    }
    if (Candidates.empty())
      return false;

    // Check intervening may-alias stores.  These need runtime checks for alias
    // disambiguation.
    SmallVector<RuntimePointerChecking::PointerCheck, 4> Checks =
        collectMemchecks(Candidates);

    // Too many checks are likely to outweigh the benefits of forwarding.
    if (Checks.size() > Candidates.size() * CheckPerElim) {
      LLVM_DEBUG(dbgs() << "Too many run-time checks needed.\n");
      return false;
    }

    if (LAI.getPSE().getUnionPredicate().getComplexity() >
        LoadElimSCEVCheckThreshold) {
      LLVM_DEBUG(dbgs() << "Too many SCEV run-time checks needed.\n");
      return false;
    }

    if (!Checks.empty() || !LAI.getPSE().getUnionPredicate().isAlwaysTrue()) {
      auto *HeaderBB = L->getHeader();
      auto *F = HeaderBB->getParent();
      bool OptForSize = F->hasOptSize() ||
                        llvm::shouldOptimizeForSize(HeaderBB, PSI, BFI);
      if (OptForSize) {
        LLVM_DEBUG(
            dbgs() << "Versioning is needed but not allowed when optimizing "
                      "for size.\n");
        return false;
      }

      if (!L->isLoopSimplifyForm()) {
        LLVM_DEBUG(dbgs() << "Loop is not is loop-simplify form");
        return false;
      }

      // Point of no-return, start the transformation.  First, version the loop
      // if necessary.

      LoopVersioning LV(LAI, L, LI, DT, PSE.getSE(), false);
      LV.setAliasChecks(std::move(Checks));
      LV.setSCEVChecks(LAI.getPSE().getUnionPredicate());
      LV.versionLoop();
    }

    // Next, propagate the value stored by the store to the users of the load.
    // Also for the first iteration, generate the initial value of the load.
    SCEVExpander SEE(*PSE.getSE(), L->getHeader()->getModule()->getDataLayout(),
                     "storeforward");
    for (const auto &Cand : Candidates)
      propagateStoredValueToLoadUsers(Cand, SEE);
    NumLoopLoadEliminted += NumForwarding;

    return true;
  }

private:
  Loop *L;

  /// Maps the load/store instructions to their index according to
  /// program order.
  DenseMap<Instruction *, unsigned> InstOrder;

  // Analyses used.
  LoopInfo *LI;
  const LoopAccessInfo &LAI;
  DominatorTree *DT;
  BlockFrequencyInfo *BFI;
  ProfileSummaryInfo *PSI;
  PredicatedScalarEvolution PSE;
};

} // end anonymous namespace

static bool
eliminateLoadsAcrossLoops(Function &F, LoopInfo &LI, DominatorTree &DT,
                          BlockFrequencyInfo *BFI, ProfileSummaryInfo *PSI,
                          function_ref<const LoopAccessInfo &(Loop &)> GetLAI) {
  // Build up a worklist of inner-loops to transform to avoid iterator
  // invalidation.
  // FIXME: This logic comes from other passes that actually change the loop
  // nest structure. It isn't clear this is necessary (or useful) for a pass
  // which merely optimizes the use of loads in a loop.
  SmallVector<Loop *, 8> Worklist;

  for (Loop *TopLevelLoop : LI)
    for (Loop *L : depth_first(TopLevelLoop))
      // We only handle inner-most loops.
      if (L->empty())
        Worklist.push_back(L);

  // Now walk the identified inner loops.
  bool Changed = false;
  for (Loop *L : Worklist) {
    // The actual work is performed by LoadEliminationForLoop.
    LoadEliminationForLoop LEL(L, &LI, GetLAI(*L), &DT, BFI, PSI);
    Changed |= LEL.processLoop();
  }
  return Changed;
}

namespace {

/// The pass.  Most of the work is delegated to the per-loop
/// LoadEliminationForLoop class.
class LoopLoadElimination : public FunctionPass {
public:
  static char ID;

  LoopLoadElimination() : FunctionPass(ID) {
    initializeLoopLoadEliminationPass(*PassRegistry::getPassRegistry());
  }

  bool runOnFunction(Function &F) override {
    if (skipFunction(F))
      return false;

    auto &LI = getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
    auto &LAA = getAnalysis<LoopAccessLegacyAnalysis>();
    auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
    auto *PSI = &getAnalysis<ProfileSummaryInfoWrapperPass>().getPSI();
    auto *BFI = (PSI && PSI->hasProfileSummary()) ?
                &getAnalysis<LazyBlockFrequencyInfoPass>().getBFI() :
                nullptr;

    // Process each loop nest in the function.
    return eliminateLoadsAcrossLoops(
        F, LI, DT, BFI, PSI,
        [&LAA](Loop &L) -> const LoopAccessInfo & { return LAA.getInfo(&L); });
  }

  void getAnalysisUsage(AnalysisUsage &AU) const override {
    AU.addRequiredID(LoopSimplifyID);
    AU.addRequired<LoopInfoWrapperPass>();
    AU.addPreserved<LoopInfoWrapperPass>();
    AU.addRequired<LoopAccessLegacyAnalysis>();
    AU.addRequired<ScalarEvolutionWrapperPass>();
    AU.addRequired<DominatorTreeWrapperPass>();
    AU.addPreserved<DominatorTreeWrapperPass>();
    AU.addPreserved<GlobalsAAWrapperPass>();
    AU.addRequired<ProfileSummaryInfoWrapperPass>();
    LazyBlockFrequencyInfoPass::getLazyBFIAnalysisUsage(AU);
  }
};

} // end anonymous namespace

char LoopLoadElimination::ID;

static const char LLE_name[] = "Loop Load Elimination";

INITIALIZE_PASS_BEGIN(LoopLoadElimination, LLE_OPTION, LLE_name, false, false)
INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
INITIALIZE_PASS_DEPENDENCY(LoopAccessLegacyAnalysis)
INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass)
INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
INITIALIZE_PASS_DEPENDENCY(ProfileSummaryInfoWrapperPass)
INITIALIZE_PASS_DEPENDENCY(LazyBlockFrequencyInfoPass)
INITIALIZE_PASS_END(LoopLoadElimination, LLE_OPTION, LLE_name, false, false)

FunctionPass *llvm::createLoopLoadEliminationPass() {
  return new LoopLoadElimination();
}

PreservedAnalyses LoopLoadEliminationPass::run(Function &F,
                                               FunctionAnalysisManager &AM) {
  auto &SE = AM.getResult<ScalarEvolutionAnalysis>(F);
  auto &LI = AM.getResult<LoopAnalysis>(F);
  auto &TTI = AM.getResult<TargetIRAnalysis>(F);
  auto &DT = AM.getResult<DominatorTreeAnalysis>(F);
  auto &TLI = AM.getResult<TargetLibraryAnalysis>(F);
  auto &AA = AM.getResult<AAManager>(F);
  auto &AC = AM.getResult<AssumptionAnalysis>(F);
  auto &MAM = AM.getResult<ModuleAnalysisManagerFunctionProxy>(F).getManager();
  auto *PSI = MAM.getCachedResult<ProfileSummaryAnalysis>(*F.getParent());
  auto *BFI = (PSI && PSI->hasProfileSummary()) ?
      &AM.getResult<BlockFrequencyAnalysis>(F) : nullptr;
  MemorySSA *MSSA = EnableMSSALoopDependency
                        ? &AM.getResult<MemorySSAAnalysis>(F).getMSSA()
                        : nullptr;

  auto &LAM = AM.getResult<LoopAnalysisManagerFunctionProxy>(F).getManager();
  bool Changed = eliminateLoadsAcrossLoops(
      F, LI, DT, BFI, PSI, [&](Loop &L) -> const LoopAccessInfo & {
        LoopStandardAnalysisResults AR = {AA, AC, DT, LI, SE, TLI, TTI, MSSA};
        return LAM.getResult<LoopAccessAnalysis>(L, AR);
      });

  if (!Changed)
    return PreservedAnalyses::all();

  PreservedAnalyses PA;
  return PA;
}