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/PassManager.h"
46#include "llvm/IR/Type.h"
47#include "llvm/IR/Value.h"
48#include "llvm/Support/Casting.h"
49#include "llvm/Support/CommandLine.h"
50#include "llvm/Support/Debug.h"
51#include "llvm/Support/raw_ostream.h"
52#include "llvm/Transforms/Utils/LoopSimplify.h"
53#include "llvm/Transforms/Utils/LoopVersioning.h"
54#include "llvm/Transforms/Utils/ScalarEvolutionExpander.h"
55#include "llvm/Transforms/Utils/SizeOpts.h"
56#include <algorithm>
57#include <cassert>
58#include <forward_list>
59#include <tuple>
60#include <utility>
61
62using namespace llvm;
63
64#define LLE_OPTION "loop-load-elim"
65#define DEBUG_TYPE LLE_OPTION
66
67static cl::opt<unsigned> CheckPerElim(
68 "runtime-check-per-loop-load-elim", cl::Hidden,
69 cl::desc("Max number of memchecks allowed per eliminated load on average"),
70 cl::init(Val: 1));
71
72static cl::opt<unsigned> LoadElimSCEVCheckThreshold(
73 "loop-load-elimination-scev-check-threshold", cl::init(Val: 8), cl::Hidden,
74 cl::desc("The maximum number of SCEV checks allowed for Loop "
75 "Load Elimination"));
76
77STATISTIC(NumLoopLoadEliminted, "Number of loads eliminated by LLE");
78
79namespace {
80
81/// Represent a store-to-forwarding candidate.
82struct StoreToLoadForwardingCandidate {
83 LoadInst *Load;
84 StoreInst *Store;
85
86 StoreToLoadForwardingCandidate(LoadInst *Load, StoreInst *Store)
87 : Load(Load), Store(Store) {}
88
89 /// Return true if the dependence from the store to the load has an
90 /// absolute distance of one.
91 /// E.g. A[i+1] = A[i] (or A[i-1] = A[i] for descending loop)
92 bool isDependenceDistanceOfOne(PredicatedScalarEvolution &PSE,
93 Loop *L) const {
94 Value *LoadPtr = Load->getPointerOperand();
95 Value *StorePtr = Store->getPointerOperand();
96 Type *LoadType = getLoadStoreType(I: Load);
97 auto &DL = Load->getDataLayout();
98
99 assert(LoadPtr->getType()->getPointerAddressSpace() ==
100 StorePtr->getType()->getPointerAddressSpace() &&
101 DL.getTypeSizeInBits(LoadType) ==
102 DL.getTypeSizeInBits(getLoadStoreType(Store)) &&
103 "Should be a known dependence");
104
105 int64_t StrideLoad = getPtrStride(PSE, AccessTy: LoadType, Ptr: LoadPtr, Lp: L).value_or(u: 0);
106 int64_t StrideStore = getPtrStride(PSE, AccessTy: LoadType, Ptr: StorePtr, Lp: L).value_or(u: 0);
107 if (!StrideLoad || !StrideStore || StrideLoad != StrideStore)
108 return false;
109
110 // TODO: This check for stride values other than 1 and -1 can be eliminated.
111 // However, doing so may cause the LoopAccessAnalysis to overcompensate,
112 // generating numerous non-wrap runtime checks that may undermine the
113 // benefits of load elimination. To safely implement support for non-unit
114 // strides, we would need to ensure either that the processed case does not
115 // require these additional checks, or improve the LAA to handle them more
116 // efficiently, or potentially both.
117 if (std::abs(i: StrideLoad) != 1)
118 return false;
119
120 unsigned TypeByteSize = DL.getTypeAllocSize(Ty: const_cast<Type *>(LoadType));
121
122 auto *LoadPtrSCEV = cast<SCEVAddRecExpr>(Val: PSE.getSCEV(V: LoadPtr));
123 auto *StorePtrSCEV = cast<SCEVAddRecExpr>(Val: PSE.getSCEV(V: StorePtr));
124
125 // We don't need to check non-wrapping here because forward/backward
126 // dependence wouldn't be valid if these weren't monotonic accesses.
127 auto *Dist = dyn_cast<SCEVConstant>(
128 Val: PSE.getSE()->getMinusSCEV(LHS: StorePtrSCEV, RHS: LoadPtrSCEV));
129 if (!Dist)
130 return false;
131 const APInt &Val = Dist->getAPInt();
132 return Val == TypeByteSize * StrideLoad;
133 }
134
135 Value *getLoadPtr() const { return Load->getPointerOperand(); }
136
137#ifndef NDEBUG
138 friend raw_ostream &operator<<(raw_ostream &OS,
139 const StoreToLoadForwardingCandidate &Cand) {
140 OS << *Cand.Store << " -->\n";
141 OS.indent(2) << *Cand.Load << "\n";
142 return OS;
143 }
144#endif
145};
146
147} // end anonymous namespace
148
149/// Check if the store dominates all latches, so as long as there is no
150/// intervening store this value will be loaded in the next iteration.
151static bool doesStoreDominatesAllLatches(BasicBlock *StoreBlock, Loop *L,
152 DominatorTree *DT) {
153 SmallVector<BasicBlock *, 8> Latches;
154 L->getLoopLatches(LoopLatches&: Latches);
155 return llvm::all_of(Range&: Latches, P: [&](const BasicBlock *Latch) {
156 return DT->dominates(A: StoreBlock, B: Latch);
157 });
158}
159
160/// Return true if the load is not executed on all paths in the loop.
161static bool isLoadConditional(LoadInst *Load, Loop *L) {
162 return Load->getParent() != L->getHeader();
163}
164
165namespace {
166
167/// The per-loop class that does most of the work.
168class LoadEliminationForLoop {
169public:
170 LoadEliminationForLoop(Loop *L, LoopInfo *LI, const LoopAccessInfo &LAI,
171 DominatorTree *DT, BlockFrequencyInfo *BFI,
172 ProfileSummaryInfo* PSI)
173 : L(L), LI(LI), LAI(LAI), DT(DT), BFI(BFI), PSI(PSI), PSE(LAI.getPSE()) {}
174
175 /// Look through the loop-carried and loop-independent dependences in
176 /// this loop and find store->load dependences.
177 ///
178 /// Note that no candidate is returned if LAA has failed to analyze the loop
179 /// (e.g. if it's not bottom-tested, contains volatile memops, etc.)
180 std::forward_list<StoreToLoadForwardingCandidate>
181 findStoreToLoadDependences(const LoopAccessInfo &LAI) {
182 std::forward_list<StoreToLoadForwardingCandidate> Candidates;
183
184 const auto &DepChecker = LAI.getDepChecker();
185 const auto *Deps = DepChecker.getDependences();
186 if (!Deps)
187 return Candidates;
188
189 // Find store->load dependences (consequently true dep). Both lexically
190 // forward and backward dependences qualify. Disqualify loads that have
191 // other unknown dependences.
192
193 SmallPtrSet<Instruction *, 4> LoadsWithUnknownDependence;
194
195 for (const auto &Dep : *Deps) {
196 Instruction *Source = Dep.getSource(DepChecker);
197 Instruction *Destination = Dep.getDestination(DepChecker);
198
199 if (Dep.Type == MemoryDepChecker::Dependence::Unknown ||
200 Dep.Type == MemoryDepChecker::Dependence::IndirectUnsafe) {
201 if (isa<LoadInst>(Val: Source))
202 LoadsWithUnknownDependence.insert(Ptr: Source);
203 if (isa<LoadInst>(Val: Destination))
204 LoadsWithUnknownDependence.insert(Ptr: Destination);
205 continue;
206 }
207
208 if (Dep.isBackward())
209 // Note that the designations source and destination follow the program
210 // order, i.e. source is always first. (The direction is given by the
211 // DepType.)
212 std::swap(a&: Source, b&: Destination);
213 else
214 assert(Dep.isForward() && "Needs to be a forward dependence");
215
216 auto *Store = dyn_cast<StoreInst>(Val: Source);
217 if (!Store)
218 continue;
219 auto *Load = dyn_cast<LoadInst>(Val: Destination);
220 if (!Load)
221 continue;
222
223 // Only propagate if the stored values are bit/pointer castable.
224 if (!CastInst::isBitOrNoopPointerCastable(
225 SrcTy: getLoadStoreType(I: Store), DestTy: getLoadStoreType(I: Load),
226 DL: Store->getDataLayout()))
227 continue;
228
229 Candidates.emplace_front(args&: Load, args&: Store);
230 }
231
232 if (!LoadsWithUnknownDependence.empty())
233 Candidates.remove_if(pred: [&](const StoreToLoadForwardingCandidate &C) {
234 return LoadsWithUnknownDependence.count(Ptr: C.Load);
235 });
236
237 return Candidates;
238 }
239
240 /// Return the index of the instruction according to program order.
241 unsigned getInstrIndex(Instruction *Inst) {
242 auto I = InstOrder.find(Val: Inst);
243 assert(I != InstOrder.end() && "No index for instruction");
244 return I->second;
245 }
246
247 /// If a load has multiple candidates associated (i.e. different
248 /// stores), it means that it could be forwarding from multiple stores
249 /// depending on control flow. Remove these candidates.
250 ///
251 /// Here, we rely on LAA to include the relevant loop-independent dependences.
252 /// LAA is known to omit these in the very simple case when the read and the
253 /// write within an alias set always takes place using the *same* pointer.
254 ///
255 /// However, we know that this is not the case here, i.e. we can rely on LAA
256 /// to provide us with loop-independent dependences for the cases we're
257 /// interested. Consider the case for example where a loop-independent
258 /// dependece S1->S2 invalidates the forwarding S3->S2.
259 ///
260 /// A[i] = ... (S1)
261 /// ... = A[i] (S2)
262 /// A[i+1] = ... (S3)
263 ///
264 /// LAA will perform dependence analysis here because there are two
265 /// *different* pointers involved in the same alias set (&A[i] and &A[i+1]).
266 void removeDependencesFromMultipleStores(
267 std::forward_list<StoreToLoadForwardingCandidate> &Candidates) {
268 // If Store is nullptr it means that we have multiple stores forwarding to
269 // this store.
270 using LoadToSingleCandT =
271 DenseMap<LoadInst *, const StoreToLoadForwardingCandidate *>;
272 LoadToSingleCandT LoadToSingleCand;
273
274 for (const auto &Cand : Candidates) {
275 bool NewElt;
276 LoadToSingleCandT::iterator Iter;
277
278 std::tie(args&: Iter, args&: NewElt) =
279 LoadToSingleCand.insert(KV: std::make_pair(x: Cand.Load, y: &Cand));
280 if (!NewElt) {
281 const StoreToLoadForwardingCandidate *&OtherCand = Iter->second;
282 // Already multiple stores forward to this load.
283 if (OtherCand == nullptr)
284 continue;
285
286 // Handle the very basic case when the two stores are in the same block
287 // so deciding which one forwards is easy. The later one forwards as
288 // long as they both have a dependence distance of one to the load.
289 if (Cand.Store->getParent() == OtherCand->Store->getParent() &&
290 Cand.isDependenceDistanceOfOne(PSE, L) &&
291 OtherCand->isDependenceDistanceOfOne(PSE, L)) {
292 // They are in the same block, the later one will forward to the load.
293 if (getInstrIndex(Inst: OtherCand->Store) < getInstrIndex(Inst: Cand.Store))
294 OtherCand = &Cand;
295 } else
296 OtherCand = nullptr;
297 }
298 }
299
300 Candidates.remove_if(pred: [&](const StoreToLoadForwardingCandidate &Cand) {
301 if (LoadToSingleCand[Cand.Load] != &Cand) {
302 LLVM_DEBUG(
303 dbgs() << "Removing from candidates: \n"
304 << Cand
305 << " The load may have multiple stores forwarding to "
306 << "it\n");
307 return true;
308 }
309 return false;
310 });
311 }
312
313 /// Given two pointers operations by their RuntimePointerChecking
314 /// indices, return true if they require an alias check.
315 ///
316 /// We need a check if one is a pointer for a candidate load and the other is
317 /// a pointer for a possibly intervening store.
318 bool needsChecking(unsigned PtrIdx1, unsigned PtrIdx2,
319 const SmallPtrSetImpl<Value *> &PtrsWrittenOnFwdingPath,
320 const SmallPtrSetImpl<Value *> &CandLoadPtrs) {
321 Value *Ptr1 =
322 LAI.getRuntimePointerChecking()->getPointerInfo(PtrIdx: PtrIdx1).PointerValue;
323 Value *Ptr2 =
324 LAI.getRuntimePointerChecking()->getPointerInfo(PtrIdx: PtrIdx2).PointerValue;
325 return ((PtrsWrittenOnFwdingPath.count(Ptr: Ptr1) && CandLoadPtrs.count(Ptr: Ptr2)) ||
326 (PtrsWrittenOnFwdingPath.count(Ptr: Ptr2) && CandLoadPtrs.count(Ptr: Ptr1)));
327 }
328
329 /// Return pointers that are possibly written to on the path from a
330 /// forwarding store to a load.
331 ///
332 /// These pointers need to be alias-checked against the forwarding candidates.
333 SmallPtrSet<Value *, 4> findPointersWrittenOnForwardingPath(
334 const SmallVectorImpl<StoreToLoadForwardingCandidate> &Candidates) {
335 // From FirstStore to LastLoad neither of the elimination candidate loads
336 // should overlap with any of the stores.
337 //
338 // E.g.:
339 //
340 // st1 C[i]
341 // ld1 B[i] <-------,
342 // ld0 A[i] <----, | * LastLoad
343 // ... | |
344 // st2 E[i] | |
345 // st3 B[i+1] -- | -' * FirstStore
346 // st0 A[i+1] ---'
347 // st4 D[i]
348 //
349 // st0 forwards to ld0 if the accesses in st4 and st1 don't overlap with
350 // ld0.
351
352 LoadInst *LastLoad =
353 llvm::max_element(Range: Candidates,
354 C: [&](const StoreToLoadForwardingCandidate &A,
355 const StoreToLoadForwardingCandidate &B) {
356 return getInstrIndex(Inst: A.Load) <
357 getInstrIndex(Inst: B.Load);
358 })
359 ->Load;
360 StoreInst *FirstStore =
361 llvm::min_element(Range: Candidates,
362 C: [&](const StoreToLoadForwardingCandidate &A,
363 const StoreToLoadForwardingCandidate &B) {
364 return getInstrIndex(Inst: A.Store) <
365 getInstrIndex(Inst: B.Store);
366 })
367 ->Store;
368
369 // We're looking for stores after the first forwarding store until the end
370 // of the loop, then from the beginning of the loop until the last
371 // forwarded-to load. Collect the pointer for the stores.
372 SmallPtrSet<Value *, 4> PtrsWrittenOnFwdingPath;
373
374 auto InsertStorePtr = [&](Instruction *I) {
375 if (auto *S = dyn_cast<StoreInst>(Val: I))
376 PtrsWrittenOnFwdingPath.insert(Ptr: S->getPointerOperand());
377 };
378 const auto &MemInstrs = LAI.getDepChecker().getMemoryInstructions();
379 std::for_each(first: MemInstrs.begin() + getInstrIndex(Inst: FirstStore) + 1,
380 last: MemInstrs.end(), f: InsertStorePtr);
381 std::for_each(first: MemInstrs.begin(), last: &MemInstrs[getInstrIndex(Inst: LastLoad)],
382 f: InsertStorePtr);
383
384 return PtrsWrittenOnFwdingPath;
385 }
386
387 /// Determine the pointer alias checks to prove that there are no
388 /// intervening stores.
389 SmallVector<RuntimePointerCheck, 4> collectMemchecks(
390 const SmallVectorImpl<StoreToLoadForwardingCandidate> &Candidates) {
391
392 SmallPtrSet<Value *, 4> PtrsWrittenOnFwdingPath =
393 findPointersWrittenOnForwardingPath(Candidates);
394
395 // Collect the pointers of the candidate loads.
396 SmallPtrSet<Value *, 4> CandLoadPtrs;
397 for (const auto &Candidate : Candidates)
398 CandLoadPtrs.insert(Ptr: Candidate.getLoadPtr());
399
400 const auto &AllChecks = LAI.getRuntimePointerChecking()->getChecks();
401 SmallVector<RuntimePointerCheck, 4> Checks;
402
403 copy_if(Range: AllChecks, Out: std::back_inserter(x&: Checks),
404 P: [&](const RuntimePointerCheck &Check) {
405 for (auto PtrIdx1 : Check.first->Members)
406 for (auto PtrIdx2 : Check.second->Members)
407 if (needsChecking(PtrIdx1, PtrIdx2, PtrsWrittenOnFwdingPath,
408 CandLoadPtrs))
409 return true;
410 return false;
411 });
412
413 LLVM_DEBUG(dbgs() << "\nPointer Checks (count: " << Checks.size()
414 << "):\n");
415 LLVM_DEBUG(LAI.getRuntimePointerChecking()->printChecks(dbgs(), Checks));
416
417 return Checks;
418 }
419
420 /// Perform the transformation for a candidate.
421 void
422 propagateStoredValueToLoadUsers(const StoreToLoadForwardingCandidate &Cand,
423 SCEVExpander &SEE) {
424 // loop:
425 // %x = load %gep_i
426 // = ... %x
427 // store %y, %gep_i_plus_1
428 //
429 // =>
430 //
431 // ph:
432 // %x.initial = load %gep_0
433 // loop:
434 // %x.storeforward = phi [%x.initial, %ph] [%y, %loop]
435 // %x = load %gep_i <---- now dead
436 // = ... %x.storeforward
437 // store %y, %gep_i_plus_1
438
439 Value *Ptr = Cand.Load->getPointerOperand();
440 auto *PtrSCEV = cast<SCEVAddRecExpr>(Val: PSE.getSCEV(V: Ptr));
441 auto *PH = L->getLoopPreheader();
442 assert(PH && "Preheader should exist!");
443 Value *InitialPtr = SEE.expandCodeFor(SH: PtrSCEV->getStart(), Ty: Ptr->getType(),
444 I: PH->getTerminator());
445 Instruction *Initial =
446 new LoadInst(Cand.Load->getType(), InitialPtr, "load_initial",
447 /* isVolatile */ false, Cand.Load->getAlign(),
448 PH->getTerminator()->getIterator());
449 // We don't give any debug location to Initial, because it is inserted
450 // into the loop's preheader. A debug location inside the loop will cause
451 // a misleading stepping when debugging. The test update-debugloc-store
452 // -forwarded.ll checks this.
453 Initial->setDebugLoc(DebugLoc::getDropped());
454
455 PHINode *PHI = PHINode::Create(Ty: Initial->getType(), NumReservedValues: 2, NameStr: "store_forwarded");
456 PHI->insertBefore(InsertPos: L->getHeader()->begin());
457 PHI->addIncoming(V: Initial, BB: PH);
458
459 Type *LoadType = Initial->getType();
460 Type *StoreType = Cand.Store->getValueOperand()->getType();
461 auto &DL = Cand.Load->getDataLayout();
462 (void)DL;
463
464 assert(DL.getTypeSizeInBits(LoadType) == DL.getTypeSizeInBits(StoreType) &&
465 "The type sizes should match!");
466
467 Value *StoreValue = Cand.Store->getValueOperand();
468 if (LoadType != StoreType) {
469 StoreValue = CastInst::CreateBitOrPointerCast(S: StoreValue, Ty: LoadType,
470 Name: "store_forward_cast",
471 InsertBefore: Cand.Store->getIterator());
472 // Because it casts the old `load` value and is used by the new `phi`
473 // which replaces the old `load`, we give the `load`'s debug location
474 // to it.
475 cast<Instruction>(Val: StoreValue)->setDebugLoc(Cand.Load->getDebugLoc());
476 }
477
478 PHI->addIncoming(V: StoreValue, BB: L->getLoopLatch());
479
480 Cand.Load->replaceAllUsesWith(V: PHI);
481 PHI->setDebugLoc(Cand.Load->getDebugLoc());
482 }
483
484 /// Top-level driver for each loop: find store->load forwarding
485 /// candidates, add run-time checks and perform transformation.
486 bool processLoop() {
487 LLVM_DEBUG(dbgs() << "\nIn \"" << L->getHeader()->getParent()->getName()
488 << "\" checking " << *L << "\n");
489
490 // Look for store-to-load forwarding cases across the
491 // backedge. E.g.:
492 //
493 // loop:
494 // %x = load %gep_i
495 // = ... %x
496 // store %y, %gep_i_plus_1
497 //
498 // =>
499 //
500 // ph:
501 // %x.initial = load %gep_0
502 // loop:
503 // %x.storeforward = phi [%x.initial, %ph] [%y, %loop]
504 // %x = load %gep_i <---- now dead
505 // = ... %x.storeforward
506 // store %y, %gep_i_plus_1
507
508 // First start with store->load dependences.
509 auto StoreToLoadDependences = findStoreToLoadDependences(LAI);
510 if (StoreToLoadDependences.empty())
511 return false;
512
513 // Generate an index for each load and store according to the original
514 // program order. This will be used later.
515 InstOrder = LAI.getDepChecker().generateInstructionOrderMap();
516
517 // To keep things simple for now, remove those where the load is potentially
518 // fed by multiple stores.
519 removeDependencesFromMultipleStores(Candidates&: StoreToLoadDependences);
520 if (StoreToLoadDependences.empty())
521 return false;
522
523 // Filter the candidates further.
524 SmallVector<StoreToLoadForwardingCandidate, 4> Candidates;
525 for (const StoreToLoadForwardingCandidate &Cand : StoreToLoadDependences) {
526 LLVM_DEBUG(dbgs() << "Candidate " << Cand);
527
528 // Make sure that the stored values is available everywhere in the loop in
529 // the next iteration.
530 if (!doesStoreDominatesAllLatches(StoreBlock: Cand.Store->getParent(), L, DT))
531 continue;
532
533 // If the load is conditional we can't hoist its 0-iteration instance to
534 // the preheader because that would make it unconditional. Thus we would
535 // access a memory location that the original loop did not access.
536 if (isLoadConditional(Load: Cand.Load, L))
537 continue;
538
539 // Check whether the SCEV difference is the same as the induction step,
540 // thus we load the value in the next iteration.
541 if (!Cand.isDependenceDistanceOfOne(PSE, L))
542 continue;
543
544 assert(isa<SCEVAddRecExpr>(PSE.getSCEV(Cand.Load->getPointerOperand())) &&
545 "Loading from something other than indvar?");
546 assert(
547 isa<SCEVAddRecExpr>(PSE.getSCEV(Cand.Store->getPointerOperand())) &&
548 "Storing to something other than indvar?");
549
550 Candidates.push_back(Elt: Cand);
551 LLVM_DEBUG(
552 dbgs()
553 << Candidates.size()
554 << ". Valid store-to-load forwarding across the loop backedge\n");
555 }
556 if (Candidates.empty())
557 return false;
558
559 // Check intervening may-alias stores. These need runtime checks for alias
560 // disambiguation.
561 SmallVector<RuntimePointerCheck, 4> Checks = collectMemchecks(Candidates);
562
563 // Too many checks are likely to outweigh the benefits of forwarding.
564 if (Checks.size() > Candidates.size() * CheckPerElim) {
565 LLVM_DEBUG(dbgs() << "Too many run-time checks needed.\n");
566 return false;
567 }
568
569 if (LAI.getPSE().getPredicate().getComplexity() >
570 LoadElimSCEVCheckThreshold) {
571 LLVM_DEBUG(dbgs() << "Too many SCEV run-time checks needed.\n");
572 return false;
573 }
574
575 if (!L->isLoopSimplifyForm()) {
576 LLVM_DEBUG(dbgs() << "Loop is not is loop-simplify form");
577 return false;
578 }
579
580 if (!Checks.empty() || !LAI.getPSE().getPredicate().isAlwaysTrue()) {
581 if (LAI.hasConvergentOp()) {
582 LLVM_DEBUG(dbgs() << "Versioning is needed but not allowed with "
583 "convergent calls\n");
584 return false;
585 }
586
587 auto *HeaderBB = L->getHeader();
588 if (llvm::shouldOptimizeForSize(BB: HeaderBB, PSI, BFI,
589 QueryType: PGSOQueryType::IRPass)) {
590 LLVM_DEBUG(
591 dbgs() << "Versioning is needed but not allowed when optimizing "
592 "for size.\n");
593 return false;
594 }
595
596 // Point of no-return, start the transformation. First, version the loop
597 // if necessary.
598
599 LoopVersioning LV(LAI, Checks, L, LI, DT, PSE.getSE());
600 LV.versionLoop();
601
602 // After versioning, some of the candidates' pointers could stop being
603 // SCEVAddRecs. We need to filter them out.
604 auto NoLongerGoodCandidate = [this](
605 const StoreToLoadForwardingCandidate &Cand) {
606 return !isa<SCEVAddRecExpr>(
607 Val: PSE.getSCEV(V: Cand.Load->getPointerOperand())) ||
608 !isa<SCEVAddRecExpr>(
609 Val: PSE.getSCEV(V: Cand.Store->getPointerOperand()));
610 };
611 llvm::erase_if(C&: Candidates, P: NoLongerGoodCandidate);
612 }
613
614 // Next, propagate the value stored by the store to the users of the load.
615 // Also for the first iteration, generate the initial value of the load.
616 SCEVExpander SEE(*PSE.getSE(), L->getHeader()->getDataLayout(),
617 "storeforward");
618 for (const auto &Cand : Candidates)
619 propagateStoredValueToLoadUsers(Cand, SEE);
620 NumLoopLoadEliminted += Candidates.size();
621
622 return true;
623 }
624
625private:
626 Loop *L;
627
628 /// Maps the load/store instructions to their index according to
629 /// program order.
630 DenseMap<Instruction *, unsigned> InstOrder;
631
632 // Analyses used.
633 LoopInfo *LI;
634 const LoopAccessInfo &LAI;
635 DominatorTree *DT;
636 BlockFrequencyInfo *BFI;
637 ProfileSummaryInfo *PSI;
638 PredicatedScalarEvolution PSE;
639};
640
641} // end anonymous namespace
642
643static bool eliminateLoadsAcrossLoops(Function &F, LoopInfo &LI,
644 DominatorTree &DT,
645 BlockFrequencyInfo *BFI,
646 ProfileSummaryInfo *PSI,
647 ScalarEvolution *SE, AssumptionCache *AC,
648 LoopAccessInfoManager &LAIs) {
649 // Build up a worklist of inner-loops to transform to avoid iterator
650 // invalidation.
651 // FIXME: This logic comes from other passes that actually change the loop
652 // nest structure. It isn't clear this is necessary (or useful) for a pass
653 // which merely optimizes the use of loads in a loop.
654 SmallVector<Loop *, 8> Worklist;
655
656 bool Changed = false;
657
658 for (Loop *TopLevelLoop : LI)
659 for (Loop *L : depth_first(G: TopLevelLoop)) {
660 Changed |= simplifyLoop(L, DT: &DT, LI: &LI, SE, AC, /*MSSAU*/ nullptr, PreserveLCSSA: false);
661 // We only handle inner-most loops.
662 if (L->isInnermost())
663 Worklist.push_back(Elt: L);
664 }
665
666 // Now walk the identified inner loops.
667 for (Loop *L : Worklist) {
668 // Match historical behavior
669 if (!L->isRotatedForm() || !L->getExitingBlock())
670 continue;
671 // The actual work is performed by LoadEliminationForLoop.
672 LoadEliminationForLoop LEL(L, &LI, LAIs.getInfo(L&: *L), &DT, BFI, PSI);
673 Changed |= LEL.processLoop();
674 if (Changed)
675 LAIs.clear();
676 }
677 return Changed;
678}
679
680PreservedAnalyses LoopLoadEliminationPass::run(Function &F,
681 FunctionAnalysisManager &AM) {
682 auto &LI = AM.getResult<LoopAnalysis>(IR&: F);
683 // There are no loops in the function. Return before computing other expensive
684 // analyses.
685 if (LI.empty())
686 return PreservedAnalyses::all();
687 auto &SE = AM.getResult<ScalarEvolutionAnalysis>(IR&: F);
688 auto &DT = AM.getResult<DominatorTreeAnalysis>(IR&: F);
689 auto &AC = AM.getResult<AssumptionAnalysis>(IR&: F);
690 auto &MAMProxy = AM.getResult<ModuleAnalysisManagerFunctionProxy>(IR&: F);
691 auto *PSI = MAMProxy.getCachedResult<ProfileSummaryAnalysis>(IR&: *F.getParent());
692 auto *BFI = (PSI && PSI->hasProfileSummary()) ?
693 &AM.getResult<BlockFrequencyAnalysis>(IR&: F) : nullptr;
694 LoopAccessInfoManager &LAIs = AM.getResult<LoopAccessAnalysis>(IR&: F);
695
696 bool Changed = eliminateLoadsAcrossLoops(F, LI, DT, BFI, PSI, SE: &SE, AC: &AC, LAIs);
697
698 if (!Changed)
699 return PreservedAnalyses::all();
700
701 PreservedAnalyses PA;
702 PA.preserve<DominatorTreeAnalysis>();
703 PA.preserve<LoopAnalysis>();
704 return PA;
705}
706