xref: /freebsd/contrib/llvm-project/llvm/lib/Transforms/Scalar/LoopLoadElimination.cpp (revision 0fca6ea1d4eea4c934cfff25ac9ee8ad6fe95583)
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/ProfileSummaryInfo.h"
38 #include "llvm/Analysis/ScalarEvolution.h"
39 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
40 #include "llvm/Analysis/TargetLibraryInfo.h"
41 #include "llvm/Analysis/TargetTransformInfo.h"
42 #include "llvm/IR/DataLayout.h"
43 #include "llvm/IR/Dominators.h"
44 #include "llvm/IR/Instructions.h"
45 #include "llvm/IR/Module.h"
46 #include "llvm/IR/PassManager.h"
47 #include "llvm/IR/Type.h"
48 #include "llvm/IR/Value.h"
49 #include "llvm/Support/Casting.h"
50 #include "llvm/Support/CommandLine.h"
51 #include "llvm/Support/Debug.h"
52 #include "llvm/Support/raw_ostream.h"
53 #include "llvm/Transforms/Utils.h"
54 #include "llvm/Transforms/Utils/LoopSimplify.h"
55 #include "llvm/Transforms/Utils/LoopVersioning.h"
56 #include "llvm/Transforms/Utils/ScalarEvolutionExpander.h"
57 #include "llvm/Transforms/Utils/SizeOpts.h"
58 #include <algorithm>
59 #include <cassert>
60 #include <forward_list>
61 #include <tuple>
62 #include <utility>
63 
64 using namespace llvm;
65 
66 #define LLE_OPTION "loop-load-elim"
67 #define DEBUG_TYPE LLE_OPTION
68 
69 static cl::opt<unsigned> CheckPerElim(
70     "runtime-check-per-loop-load-elim", cl::Hidden,
71     cl::desc("Max number of memchecks allowed per eliminated load on average"),
72     cl::init(1));
73 
74 static cl::opt<unsigned> LoadElimSCEVCheckThreshold(
75     "loop-load-elimination-scev-check-threshold", cl::init(8), cl::Hidden,
76     cl::desc("The maximum number of SCEV checks allowed for Loop "
77              "Load Elimination"));
78 
79 STATISTIC(NumLoopLoadEliminted, "Number of loads eliminated by LLE");
80 
81 namespace {
82 
83 /// Represent a store-to-forwarding candidate.
84 struct StoreToLoadForwardingCandidate {
85   LoadInst *Load;
86   StoreInst *Store;
87 
StoreToLoadForwardingCandidate__anon087f49260111::StoreToLoadForwardingCandidate88   StoreToLoadForwardingCandidate(LoadInst *Load, StoreInst *Store)
89       : Load(Load), Store(Store) {}
90 
91   /// Return true if the dependence from the store to the load has an
92   /// absolute distance of one.
93   /// E.g. A[i+1] = A[i] (or A[i-1] = A[i] for descending loop)
isDependenceDistanceOfOne__anon087f49260111::StoreToLoadForwardingCandidate94   bool isDependenceDistanceOfOne(PredicatedScalarEvolution &PSE,
95                                  Loop *L) const {
96     Value *LoadPtr = Load->getPointerOperand();
97     Value *StorePtr = Store->getPointerOperand();
98     Type *LoadType = getLoadStoreType(Load);
99     auto &DL = Load->getDataLayout();
100 
101     assert(LoadPtr->getType()->getPointerAddressSpace() ==
102                StorePtr->getType()->getPointerAddressSpace() &&
103            DL.getTypeSizeInBits(LoadType) ==
104                DL.getTypeSizeInBits(getLoadStoreType(Store)) &&
105            "Should be a known dependence");
106 
107     int64_t StrideLoad = getPtrStride(PSE, LoadType, LoadPtr, L).value_or(0);
108     int64_t StrideStore = getPtrStride(PSE, LoadType, StorePtr, L).value_or(0);
109     if (!StrideLoad || !StrideStore || StrideLoad != StrideStore)
110       return false;
111 
112     // TODO: This check for stride values other than 1 and -1 can be eliminated.
113     // However, doing so may cause the LoopAccessAnalysis to overcompensate,
114     // generating numerous non-wrap runtime checks that may undermine the
115     // benefits of load elimination. To safely implement support for non-unit
116     // strides, we would need to ensure either that the processed case does not
117     // require these additional checks, or improve the LAA to handle them more
118     // efficiently, or potentially both.
119     if (std::abs(StrideLoad) != 1)
120       return false;
121 
122     unsigned TypeByteSize = DL.getTypeAllocSize(const_cast<Type *>(LoadType));
123 
124     auto *LoadPtrSCEV = cast<SCEVAddRecExpr>(PSE.getSCEV(LoadPtr));
125     auto *StorePtrSCEV = cast<SCEVAddRecExpr>(PSE.getSCEV(StorePtr));
126 
127     // We don't need to check non-wrapping here because forward/backward
128     // dependence wouldn't be valid if these weren't monotonic accesses.
129     auto *Dist = dyn_cast<SCEVConstant>(
130         PSE.getSE()->getMinusSCEV(StorePtrSCEV, LoadPtrSCEV));
131     if (!Dist)
132       return false;
133     const APInt &Val = Dist->getAPInt();
134     return Val == TypeByteSize * StrideLoad;
135   }
136 
getLoadPtr__anon087f49260111::StoreToLoadForwardingCandidate137   Value *getLoadPtr() const { return Load->getPointerOperand(); }
138 
139 #ifndef NDEBUG
operator <<(raw_ostream & OS,const StoreToLoadForwardingCandidate & Cand)140   friend raw_ostream &operator<<(raw_ostream &OS,
141                                  const StoreToLoadForwardingCandidate &Cand) {
142     OS << *Cand.Store << " -->\n";
143     OS.indent(2) << *Cand.Load << "\n";
144     return OS;
145   }
146 #endif
147 };
148 
149 } // end anonymous namespace
150 
151 /// Check if the store dominates all latches, so as long as there is no
152 /// intervening store this value will be loaded in the next iteration.
doesStoreDominatesAllLatches(BasicBlock * StoreBlock,Loop * L,DominatorTree * DT)153 static bool doesStoreDominatesAllLatches(BasicBlock *StoreBlock, Loop *L,
154                                          DominatorTree *DT) {
155   SmallVector<BasicBlock *, 8> Latches;
156   L->getLoopLatches(Latches);
157   return llvm::all_of(Latches, [&](const BasicBlock *Latch) {
158     return DT->dominates(StoreBlock, Latch);
159   });
160 }
161 
162 /// Return true if the load is not executed on all paths in the loop.
isLoadConditional(LoadInst * Load,Loop * L)163 static bool isLoadConditional(LoadInst *Load, Loop *L) {
164   return Load->getParent() != L->getHeader();
165 }
166 
167 namespace {
168 
169 /// The per-loop class that does most of the work.
170 class LoadEliminationForLoop {
171 public:
LoadEliminationForLoop(Loop * L,LoopInfo * LI,const LoopAccessInfo & LAI,DominatorTree * DT,BlockFrequencyInfo * BFI,ProfileSummaryInfo * PSI)172   LoadEliminationForLoop(Loop *L, LoopInfo *LI, const LoopAccessInfo &LAI,
173                          DominatorTree *DT, BlockFrequencyInfo *BFI,
174                          ProfileSummaryInfo* PSI)
175       : L(L), LI(LI), LAI(LAI), DT(DT), BFI(BFI), PSI(PSI), PSE(LAI.getPSE()) {}
176 
177   /// Look through the loop-carried and loop-independent dependences in
178   /// this loop and find store->load dependences.
179   ///
180   /// Note that no candidate is returned if LAA has failed to analyze the loop
181   /// (e.g. if it's not bottom-tested, contains volatile memops, etc.)
182   std::forward_list<StoreToLoadForwardingCandidate>
findStoreToLoadDependences(const LoopAccessInfo & LAI)183   findStoreToLoadDependences(const LoopAccessInfo &LAI) {
184     std::forward_list<StoreToLoadForwardingCandidate> Candidates;
185 
186     const auto &DepChecker = LAI.getDepChecker();
187     const auto *Deps = DepChecker.getDependences();
188     if (!Deps)
189       return Candidates;
190 
191     // Find store->load dependences (consequently true dep).  Both lexically
192     // forward and backward dependences qualify.  Disqualify loads that have
193     // other unknown dependences.
194 
195     SmallPtrSet<Instruction *, 4> LoadsWithUnknownDepedence;
196 
197     for (const auto &Dep : *Deps) {
198       Instruction *Source = Dep.getSource(DepChecker);
199       Instruction *Destination = Dep.getDestination(DepChecker);
200 
201       if (Dep.Type == MemoryDepChecker::Dependence::Unknown ||
202           Dep.Type == MemoryDepChecker::Dependence::IndirectUnsafe) {
203         if (isa<LoadInst>(Source))
204           LoadsWithUnknownDepedence.insert(Source);
205         if (isa<LoadInst>(Destination))
206           LoadsWithUnknownDepedence.insert(Destination);
207         continue;
208       }
209 
210       if (Dep.isBackward())
211         // Note that the designations source and destination follow the program
212         // order, i.e. source is always first.  (The direction is given by the
213         // DepType.)
214         std::swap(Source, Destination);
215       else
216         assert(Dep.isForward() && "Needs to be a forward dependence");
217 
218       auto *Store = dyn_cast<StoreInst>(Source);
219       if (!Store)
220         continue;
221       auto *Load = dyn_cast<LoadInst>(Destination);
222       if (!Load)
223         continue;
224 
225       // Only propagate if the stored values are bit/pointer castable.
226       if (!CastInst::isBitOrNoopPointerCastable(
227               getLoadStoreType(Store), getLoadStoreType(Load),
228               Store->getDataLayout()))
229         continue;
230 
231       Candidates.emplace_front(Load, Store);
232     }
233 
234     if (!LoadsWithUnknownDepedence.empty())
235       Candidates.remove_if([&](const StoreToLoadForwardingCandidate &C) {
236         return LoadsWithUnknownDepedence.count(C.Load);
237       });
238 
239     return Candidates;
240   }
241 
242   /// Return the index of the instruction according to program order.
getInstrIndex(Instruction * Inst)243   unsigned getInstrIndex(Instruction *Inst) {
244     auto I = InstOrder.find(Inst);
245     assert(I != InstOrder.end() && "No index for instruction");
246     return I->second;
247   }
248 
249   /// If a load has multiple candidates associated (i.e. different
250   /// stores), it means that it could be forwarding from multiple stores
251   /// depending on control flow.  Remove these candidates.
252   ///
253   /// Here, we rely on LAA to include the relevant loop-independent dependences.
254   /// LAA is known to omit these in the very simple case when the read and the
255   /// write within an alias set always takes place using the *same* pointer.
256   ///
257   /// However, we know that this is not the case here, i.e. we can rely on LAA
258   /// to provide us with loop-independent dependences for the cases we're
259   /// interested.  Consider the case for example where a loop-independent
260   /// dependece S1->S2 invalidates the forwarding S3->S2.
261   ///
262   ///         A[i]   = ...   (S1)
263   ///         ...    = A[i]  (S2)
264   ///         A[i+1] = ...   (S3)
265   ///
266   /// LAA will perform dependence analysis here because there are two
267   /// *different* pointers involved in the same alias set (&A[i] and &A[i+1]).
removeDependencesFromMultipleStores(std::forward_list<StoreToLoadForwardingCandidate> & Candidates)268   void removeDependencesFromMultipleStores(
269       std::forward_list<StoreToLoadForwardingCandidate> &Candidates) {
270     // If Store is nullptr it means that we have multiple stores forwarding to
271     // this store.
272     using LoadToSingleCandT =
273         DenseMap<LoadInst *, const StoreToLoadForwardingCandidate *>;
274     LoadToSingleCandT LoadToSingleCand;
275 
276     for (const auto &Cand : Candidates) {
277       bool NewElt;
278       LoadToSingleCandT::iterator Iter;
279 
280       std::tie(Iter, NewElt) =
281           LoadToSingleCand.insert(std::make_pair(Cand.Load, &Cand));
282       if (!NewElt) {
283         const StoreToLoadForwardingCandidate *&OtherCand = Iter->second;
284         // Already multiple stores forward to this load.
285         if (OtherCand == nullptr)
286           continue;
287 
288         // Handle the very basic case when the two stores are in the same block
289         // so deciding which one forwards is easy.  The later one forwards as
290         // long as they both have a dependence distance of one to the load.
291         if (Cand.Store->getParent() == OtherCand->Store->getParent() &&
292             Cand.isDependenceDistanceOfOne(PSE, L) &&
293             OtherCand->isDependenceDistanceOfOne(PSE, L)) {
294           // They are in the same block, the later one will forward to the load.
295           if (getInstrIndex(OtherCand->Store) < getInstrIndex(Cand.Store))
296             OtherCand = &Cand;
297         } else
298           OtherCand = nullptr;
299       }
300     }
301 
302     Candidates.remove_if([&](const StoreToLoadForwardingCandidate &Cand) {
303       if (LoadToSingleCand[Cand.Load] != &Cand) {
304         LLVM_DEBUG(
305             dbgs() << "Removing from candidates: \n"
306                    << Cand
307                    << "  The load may have multiple stores forwarding to "
308                    << "it\n");
309         return true;
310       }
311       return false;
312     });
313   }
314 
315   /// Given two pointers operations by their RuntimePointerChecking
316   /// indices, return true if they require an alias check.
317   ///
318   /// We need a check if one is a pointer for a candidate load and the other is
319   /// a pointer for a possibly intervening store.
needsChecking(unsigned PtrIdx1,unsigned PtrIdx2,const SmallPtrSetImpl<Value * > & PtrsWrittenOnFwdingPath,const SmallPtrSetImpl<Value * > & CandLoadPtrs)320   bool needsChecking(unsigned PtrIdx1, unsigned PtrIdx2,
321                      const SmallPtrSetImpl<Value *> &PtrsWrittenOnFwdingPath,
322                      const SmallPtrSetImpl<Value *> &CandLoadPtrs) {
323     Value *Ptr1 =
324         LAI.getRuntimePointerChecking()->getPointerInfo(PtrIdx1).PointerValue;
325     Value *Ptr2 =
326         LAI.getRuntimePointerChecking()->getPointerInfo(PtrIdx2).PointerValue;
327     return ((PtrsWrittenOnFwdingPath.count(Ptr1) && CandLoadPtrs.count(Ptr2)) ||
328             (PtrsWrittenOnFwdingPath.count(Ptr2) && CandLoadPtrs.count(Ptr1)));
329   }
330 
331   /// Return pointers that are possibly written to on the path from a
332   /// forwarding store to a load.
333   ///
334   /// These pointers need to be alias-checked against the forwarding candidates.
findPointersWrittenOnForwardingPath(const SmallVectorImpl<StoreToLoadForwardingCandidate> & Candidates)335   SmallPtrSet<Value *, 4> findPointersWrittenOnForwardingPath(
336       const SmallVectorImpl<StoreToLoadForwardingCandidate> &Candidates) {
337     // From FirstStore to LastLoad neither of the elimination candidate loads
338     // should overlap with any of the stores.
339     //
340     // E.g.:
341     //
342     // st1 C[i]
343     // ld1 B[i] <-------,
344     // ld0 A[i] <----,  |              * LastLoad
345     // ...           |  |
346     // st2 E[i]      |  |
347     // st3 B[i+1] -- | -'              * FirstStore
348     // st0 A[i+1] ---'
349     // st4 D[i]
350     //
351     // st0 forwards to ld0 if the accesses in st4 and st1 don't overlap with
352     // ld0.
353 
354     LoadInst *LastLoad =
355         llvm::max_element(Candidates,
356                           [&](const StoreToLoadForwardingCandidate &A,
357                               const StoreToLoadForwardingCandidate &B) {
358                             return getInstrIndex(A.Load) <
359                                    getInstrIndex(B.Load);
360                           })
361             ->Load;
362     StoreInst *FirstStore =
363         llvm::min_element(Candidates,
364                           [&](const StoreToLoadForwardingCandidate &A,
365                               const StoreToLoadForwardingCandidate &B) {
366                             return getInstrIndex(A.Store) <
367                                    getInstrIndex(B.Store);
368                           })
369             ->Store;
370 
371     // We're looking for stores after the first forwarding store until the end
372     // of the loop, then from the beginning of the loop until the last
373     // forwarded-to load.  Collect the pointer for the stores.
374     SmallPtrSet<Value *, 4> PtrsWrittenOnFwdingPath;
375 
376     auto InsertStorePtr = [&](Instruction *I) {
377       if (auto *S = dyn_cast<StoreInst>(I))
378         PtrsWrittenOnFwdingPath.insert(S->getPointerOperand());
379     };
380     const auto &MemInstrs = LAI.getDepChecker().getMemoryInstructions();
381     std::for_each(MemInstrs.begin() + getInstrIndex(FirstStore) + 1,
382                   MemInstrs.end(), InsertStorePtr);
383     std::for_each(MemInstrs.begin(), &MemInstrs[getInstrIndex(LastLoad)],
384                   InsertStorePtr);
385 
386     return PtrsWrittenOnFwdingPath;
387   }
388 
389   /// Determine the pointer alias checks to prove that there are no
390   /// intervening stores.
collectMemchecks(const SmallVectorImpl<StoreToLoadForwardingCandidate> & Candidates)391   SmallVector<RuntimePointerCheck, 4> collectMemchecks(
392       const SmallVectorImpl<StoreToLoadForwardingCandidate> &Candidates) {
393 
394     SmallPtrSet<Value *, 4> PtrsWrittenOnFwdingPath =
395         findPointersWrittenOnForwardingPath(Candidates);
396 
397     // Collect the pointers of the candidate loads.
398     SmallPtrSet<Value *, 4> CandLoadPtrs;
399     for (const auto &Candidate : Candidates)
400       CandLoadPtrs.insert(Candidate.getLoadPtr());
401 
402     const auto &AllChecks = LAI.getRuntimePointerChecking()->getChecks();
403     SmallVector<RuntimePointerCheck, 4> Checks;
404 
405     copy_if(AllChecks, std::back_inserter(Checks),
406             [&](const RuntimePointerCheck &Check) {
407               for (auto PtrIdx1 : Check.first->Members)
408                 for (auto PtrIdx2 : Check.second->Members)
409                   if (needsChecking(PtrIdx1, PtrIdx2, PtrsWrittenOnFwdingPath,
410                                     CandLoadPtrs))
411                     return true;
412               return false;
413             });
414 
415     LLVM_DEBUG(dbgs() << "\nPointer Checks (count: " << Checks.size()
416                       << "):\n");
417     LLVM_DEBUG(LAI.getRuntimePointerChecking()->printChecks(dbgs(), Checks));
418 
419     return Checks;
420   }
421 
422   /// Perform the transformation for a candidate.
423   void
propagateStoredValueToLoadUsers(const StoreToLoadForwardingCandidate & Cand,SCEVExpander & SEE)424   propagateStoredValueToLoadUsers(const StoreToLoadForwardingCandidate &Cand,
425                                   SCEVExpander &SEE) {
426     // loop:
427     //      %x = load %gep_i
428     //         = ... %x
429     //      store %y, %gep_i_plus_1
430     //
431     // =>
432     //
433     // ph:
434     //      %x.initial = load %gep_0
435     // loop:
436     //      %x.storeforward = phi [%x.initial, %ph] [%y, %loop]
437     //      %x = load %gep_i            <---- now dead
438     //         = ... %x.storeforward
439     //      store %y, %gep_i_plus_1
440 
441     Value *Ptr = Cand.Load->getPointerOperand();
442     auto *PtrSCEV = cast<SCEVAddRecExpr>(PSE.getSCEV(Ptr));
443     auto *PH = L->getLoopPreheader();
444     assert(PH && "Preheader should exist!");
445     Value *InitialPtr = SEE.expandCodeFor(PtrSCEV->getStart(), Ptr->getType(),
446                                           PH->getTerminator());
447     Value *Initial =
448         new LoadInst(Cand.Load->getType(), InitialPtr, "load_initial",
449                      /* isVolatile */ false, Cand.Load->getAlign(),
450                      PH->getTerminator()->getIterator());
451     // We don't give any debug location to Initial, because it is inserted
452     // into the loop's preheader. A debug location inside the loop will cause
453     // a misleading stepping when debugging. The test update-debugloc-store
454     // -forwarded.ll checks this.
455 
456     PHINode *PHI = PHINode::Create(Initial->getType(), 2, "store_forwarded");
457     PHI->insertBefore(L->getHeader()->begin());
458     PHI->addIncoming(Initial, PH);
459 
460     Type *LoadType = Initial->getType();
461     Type *StoreType = Cand.Store->getValueOperand()->getType();
462     auto &DL = Cand.Load->getDataLayout();
463     (void)DL;
464 
465     assert(DL.getTypeSizeInBits(LoadType) == DL.getTypeSizeInBits(StoreType) &&
466            "The type sizes should match!");
467 
468     Value *StoreValue = Cand.Store->getValueOperand();
469     if (LoadType != StoreType) {
470       StoreValue = CastInst::CreateBitOrPointerCast(StoreValue, LoadType,
471                                                     "store_forward_cast",
472                                                     Cand.Store->getIterator());
473       // Because it casts the old `load` value and is used by the new `phi`
474       // which replaces the old `load`, we give the `load`'s debug location
475       // to it.
476       cast<Instruction>(StoreValue)->setDebugLoc(Cand.Load->getDebugLoc());
477     }
478 
479     PHI->addIncoming(StoreValue, L->getLoopLatch());
480 
481     Cand.Load->replaceAllUsesWith(PHI);
482     PHI->setDebugLoc(Cand.Load->getDebugLoc());
483   }
484 
485   /// Top-level driver for each loop: find store->load forwarding
486   /// candidates, add run-time checks and perform transformation.
processLoop()487   bool processLoop() {
488     LLVM_DEBUG(dbgs() << "\nIn \"" << L->getHeader()->getParent()->getName()
489                       << "\" checking " << *L << "\n");
490 
491     // Look for store-to-load forwarding cases across the
492     // backedge. E.g.:
493     //
494     // loop:
495     //      %x = load %gep_i
496     //         = ... %x
497     //      store %y, %gep_i_plus_1
498     //
499     // =>
500     //
501     // ph:
502     //      %x.initial = load %gep_0
503     // loop:
504     //      %x.storeforward = phi [%x.initial, %ph] [%y, %loop]
505     //      %x = load %gep_i            <---- now dead
506     //         = ... %x.storeforward
507     //      store %y, %gep_i_plus_1
508 
509     // First start with store->load dependences.
510     auto StoreToLoadDependences = findStoreToLoadDependences(LAI);
511     if (StoreToLoadDependences.empty())
512       return false;
513 
514     // Generate an index for each load and store according to the original
515     // program order.  This will be used later.
516     InstOrder = LAI.getDepChecker().generateInstructionOrderMap();
517 
518     // To keep things simple for now, remove those where the load is potentially
519     // fed by multiple stores.
520     removeDependencesFromMultipleStores(StoreToLoadDependences);
521     if (StoreToLoadDependences.empty())
522       return false;
523 
524     // Filter the candidates further.
525     SmallVector<StoreToLoadForwardingCandidate, 4> Candidates;
526     for (const StoreToLoadForwardingCandidate &Cand : StoreToLoadDependences) {
527       LLVM_DEBUG(dbgs() << "Candidate " << Cand);
528 
529       // Make sure that the stored values is available everywhere in the loop in
530       // the next iteration.
531       if (!doesStoreDominatesAllLatches(Cand.Store->getParent(), L, DT))
532         continue;
533 
534       // If the load is conditional we can't hoist its 0-iteration instance to
535       // the preheader because that would make it unconditional.  Thus we would
536       // access a memory location that the original loop did not access.
537       if (isLoadConditional(Cand.Load, L))
538         continue;
539 
540       // Check whether the SCEV difference is the same as the induction step,
541       // thus we load the value in the next iteration.
542       if (!Cand.isDependenceDistanceOfOne(PSE, L))
543         continue;
544 
545       assert(isa<SCEVAddRecExpr>(PSE.getSCEV(Cand.Load->getPointerOperand())) &&
546              "Loading from something other than indvar?");
547       assert(
548           isa<SCEVAddRecExpr>(PSE.getSCEV(Cand.Store->getPointerOperand())) &&
549           "Storing to something other than indvar?");
550 
551       Candidates.push_back(Cand);
552       LLVM_DEBUG(
553           dbgs()
554           << Candidates.size()
555           << ". Valid store-to-load forwarding across the loop backedge\n");
556     }
557     if (Candidates.empty())
558       return false;
559 
560     // Check intervening may-alias stores.  These need runtime checks for alias
561     // disambiguation.
562     SmallVector<RuntimePointerCheck, 4> Checks = collectMemchecks(Candidates);
563 
564     // Too many checks are likely to outweigh the benefits of forwarding.
565     if (Checks.size() > Candidates.size() * CheckPerElim) {
566       LLVM_DEBUG(dbgs() << "Too many run-time checks needed.\n");
567       return false;
568     }
569 
570     if (LAI.getPSE().getPredicate().getComplexity() >
571         LoadElimSCEVCheckThreshold) {
572       LLVM_DEBUG(dbgs() << "Too many SCEV run-time checks needed.\n");
573       return false;
574     }
575 
576     if (!L->isLoopSimplifyForm()) {
577       LLVM_DEBUG(dbgs() << "Loop is not is loop-simplify form");
578       return false;
579     }
580 
581     if (!Checks.empty() || !LAI.getPSE().getPredicate().isAlwaysTrue()) {
582       if (LAI.hasConvergentOp()) {
583         LLVM_DEBUG(dbgs() << "Versioning is needed but not allowed with "
584                              "convergent calls\n");
585         return false;
586       }
587 
588       auto *HeaderBB = L->getHeader();
589       auto *F = HeaderBB->getParent();
590       bool OptForSize = F->hasOptSize() ||
591                         llvm::shouldOptimizeForSize(HeaderBB, PSI, BFI,
592                                                     PGSOQueryType::IRPass);
593       if (OptForSize) {
594         LLVM_DEBUG(
595             dbgs() << "Versioning is needed but not allowed when optimizing "
596                       "for size.\n");
597         return false;
598       }
599 
600       // Point of no-return, start the transformation.  First, version the loop
601       // if necessary.
602 
603       LoopVersioning LV(LAI, Checks, L, LI, DT, PSE.getSE());
604       LV.versionLoop();
605 
606       // After versioning, some of the candidates' pointers could stop being
607       // SCEVAddRecs. We need to filter them out.
608       auto NoLongerGoodCandidate = [this](
609           const StoreToLoadForwardingCandidate &Cand) {
610         return !isa<SCEVAddRecExpr>(
611                     PSE.getSCEV(Cand.Load->getPointerOperand())) ||
612                !isa<SCEVAddRecExpr>(
613                     PSE.getSCEV(Cand.Store->getPointerOperand()));
614       };
615       llvm::erase_if(Candidates, NoLongerGoodCandidate);
616     }
617 
618     // Next, propagate the value stored by the store to the users of the load.
619     // Also for the first iteration, generate the initial value of the load.
620     SCEVExpander SEE(*PSE.getSE(), L->getHeader()->getDataLayout(),
621                      "storeforward");
622     for (const auto &Cand : Candidates)
623       propagateStoredValueToLoadUsers(Cand, SEE);
624     NumLoopLoadEliminted += Candidates.size();
625 
626     return true;
627   }
628 
629 private:
630   Loop *L;
631 
632   /// Maps the load/store instructions to their index according to
633   /// program order.
634   DenseMap<Instruction *, unsigned> InstOrder;
635 
636   // Analyses used.
637   LoopInfo *LI;
638   const LoopAccessInfo &LAI;
639   DominatorTree *DT;
640   BlockFrequencyInfo *BFI;
641   ProfileSummaryInfo *PSI;
642   PredicatedScalarEvolution PSE;
643 };
644 
645 } // end anonymous namespace
646 
eliminateLoadsAcrossLoops(Function & F,LoopInfo & LI,DominatorTree & DT,BlockFrequencyInfo * BFI,ProfileSummaryInfo * PSI,ScalarEvolution * SE,AssumptionCache * AC,LoopAccessInfoManager & LAIs)647 static bool eliminateLoadsAcrossLoops(Function &F, LoopInfo &LI,
648                                       DominatorTree &DT,
649                                       BlockFrequencyInfo *BFI,
650                                       ProfileSummaryInfo *PSI,
651                                       ScalarEvolution *SE, AssumptionCache *AC,
652                                       LoopAccessInfoManager &LAIs) {
653   // Build up a worklist of inner-loops to transform to avoid iterator
654   // invalidation.
655   // FIXME: This logic comes from other passes that actually change the loop
656   // nest structure. It isn't clear this is necessary (or useful) for a pass
657   // which merely optimizes the use of loads in a loop.
658   SmallVector<Loop *, 8> Worklist;
659 
660   bool Changed = false;
661 
662   for (Loop *TopLevelLoop : LI)
663     for (Loop *L : depth_first(TopLevelLoop)) {
664       Changed |= simplifyLoop(L, &DT, &LI, SE, AC, /*MSSAU*/ nullptr, false);
665       // We only handle inner-most loops.
666       if (L->isInnermost())
667         Worklist.push_back(L);
668     }
669 
670   // Now walk the identified inner loops.
671   for (Loop *L : Worklist) {
672     // Match historical behavior
673     if (!L->isRotatedForm() || !L->getExitingBlock())
674       continue;
675     // The actual work is performed by LoadEliminationForLoop.
676     LoadEliminationForLoop LEL(L, &LI, LAIs.getInfo(*L), &DT, BFI, PSI);
677     Changed |= LEL.processLoop();
678     if (Changed)
679       LAIs.clear();
680   }
681   return Changed;
682 }
683 
run(Function & F,FunctionAnalysisManager & AM)684 PreservedAnalyses LoopLoadEliminationPass::run(Function &F,
685                                                FunctionAnalysisManager &AM) {
686   auto &LI = AM.getResult<LoopAnalysis>(F);
687   // There are no loops in the function. Return before computing other expensive
688   // analyses.
689   if (LI.empty())
690     return PreservedAnalyses::all();
691   auto &SE = AM.getResult<ScalarEvolutionAnalysis>(F);
692   auto &DT = AM.getResult<DominatorTreeAnalysis>(F);
693   auto &AC = AM.getResult<AssumptionAnalysis>(F);
694   auto &MAMProxy = AM.getResult<ModuleAnalysisManagerFunctionProxy>(F);
695   auto *PSI = MAMProxy.getCachedResult<ProfileSummaryAnalysis>(*F.getParent());
696   auto *BFI = (PSI && PSI->hasProfileSummary()) ?
697       &AM.getResult<BlockFrequencyAnalysis>(F) : nullptr;
698   LoopAccessInfoManager &LAIs = AM.getResult<LoopAccessAnalysis>(F);
699 
700   bool Changed = eliminateLoadsAcrossLoops(F, LI, DT, BFI, PSI, &SE, &AC, LAIs);
701 
702   if (!Changed)
703     return PreservedAnalyses::all();
704 
705   PreservedAnalyses PA;
706   PA.preserve<DominatorTreeAnalysis>();
707   PA.preserve<LoopAnalysis>();
708   return PA;
709 }
710