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

source code of llvm/lib/Transforms/Scalar/LoopLoadElimination.cpp