1 //===- LoopLoadElimination.cpp - Loop Load Elimination Pass ---------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file implement a loop-aware load elimination pass.
12 // It uses LoopAccessAnalysis to identify loop-carried dependences with a
13 // distance of one between stores and loads. These form the candidates for the
14 // transformation. The source value of each store then propagated to the user
15 // of the corresponding load. This makes the load dead.
17 // The pass can also version the loop and add memchecks in order to prove that
18 // may-aliasing stores can't change the value in memory before it's read by the
21 //===----------------------------------------------------------------------===//
23 #include "llvm/ADT/APInt.h"
24 #include "llvm/ADT/DenseMap.h"
25 #include "llvm/ADT/DepthFirstIterator.h"
26 #include "llvm/ADT/SmallSet.h"
27 #include "llvm/ADT/SmallVector.h"
28 #include "llvm/ADT/Statistic.h"
29 #include "llvm/ADT/STLExtras.h"
30 #include "llvm/Analysis/GlobalsModRef.h"
31 #include "llvm/Analysis/LoopAccessAnalysis.h"
32 #include "llvm/Analysis/LoopInfo.h"
33 #include "llvm/Analysis/ScalarEvolution.h"
34 #include "llvm/Analysis/ScalarEvolutionExpander.h"
35 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
36 #include "llvm/IR/DataLayout.h"
37 #include "llvm/IR/Dominators.h"
38 #include "llvm/IR/Instructions.h"
39 #include "llvm/IR/Module.h"
40 #include "llvm/IR/Type.h"
41 #include "llvm/IR/Value.h"
42 #include "llvm/Pass.h"
43 #include "llvm/Support/Casting.h"
44 #include "llvm/Support/CommandLine.h"
45 #include "llvm/Support/Debug.h"
46 #include "llvm/Transforms/Scalar.h"
47 #include "llvm/Transforms/Utils/LoopVersioning.h"
48 #include <forward_list>
55 #define LLE_OPTION "loop-load-elim"
56 #define DEBUG_TYPE LLE_OPTION
60 static cl::opt<unsigned> CheckPerElim(
61 "runtime-check-per-loop-load-elim", cl::Hidden,
62 cl::desc("Max number of memchecks allowed per eliminated load on average"),
65 static cl::opt<unsigned> LoadElimSCEVCheckThreshold(
66 "loop-load-elimination-scev-check-threshold", cl::init(8), cl::Hidden,
67 cl::desc("The maximum number of SCEV checks allowed for Loop "
70 STATISTIC(NumLoopLoadEliminted, "Number of loads eliminated by LLE");
74 /// \brief Represent a store-to-forwarding candidate.
75 struct StoreToLoadForwardingCandidate {
79 StoreToLoadForwardingCandidate(LoadInst *Load, StoreInst *Store)
80 : Load(Load), Store(Store) {}
82 /// \brief Return true if the dependence from the store to the load has a
83 /// distance of one. E.g. A[i+1] = A[i]
84 bool isDependenceDistanceOfOne(PredicatedScalarEvolution &PSE,
86 Value *LoadPtr = Load->getPointerOperand();
87 Value *StorePtr = Store->getPointerOperand();
88 Type *LoadPtrType = LoadPtr->getType();
89 Type *LoadType = LoadPtrType->getPointerElementType();
91 assert(LoadPtrType->getPointerAddressSpace() ==
92 StorePtr->getType()->getPointerAddressSpace() &&
93 LoadType == StorePtr->getType()->getPointerElementType() &&
94 "Should be a known dependence");
96 // Currently we only support accesses with unit stride. FIXME: we should be
97 // able to handle non unit stirde as well as long as the stride is equal to
98 // the dependence distance.
99 if (getPtrStride(PSE, LoadPtr, L) != 1 ||
100 getPtrStride(PSE, StorePtr, L) != 1)
103 auto &DL = Load->getParent()->getModule()->getDataLayout();
104 unsigned TypeByteSize = DL.getTypeAllocSize(const_cast<Type *>(LoadType));
106 auto *LoadPtrSCEV = cast<SCEVAddRecExpr>(PSE.getSCEV(LoadPtr));
107 auto *StorePtrSCEV = cast<SCEVAddRecExpr>(PSE.getSCEV(StorePtr));
109 // We don't need to check non-wrapping here because forward/backward
110 // dependence wouldn't be valid if these weren't monotonic accesses.
111 auto *Dist = cast<SCEVConstant>(
112 PSE.getSE()->getMinusSCEV(StorePtrSCEV, LoadPtrSCEV));
113 const APInt &Val = Dist->getAPInt();
114 return Val == TypeByteSize;
117 Value *getLoadPtr() const { return Load->getPointerOperand(); }
120 friend raw_ostream &operator<<(raw_ostream &OS,
121 const StoreToLoadForwardingCandidate &Cand) {
122 OS << *Cand.Store << " -->\n";
123 OS.indent(2) << *Cand.Load << "\n";
129 /// \brief Check if the store dominates all latches, so as long as there is no
130 /// intervening store this value will be loaded in the next iteration.
131 bool doesStoreDominatesAllLatches(BasicBlock *StoreBlock, Loop *L,
133 SmallVector<BasicBlock *, 8> Latches;
134 L->getLoopLatches(Latches);
135 return llvm::all_of(Latches, [&](const BasicBlock *Latch) {
136 return DT->dominates(StoreBlock, Latch);
140 /// \brief Return true if the load is not executed on all paths in the loop.
141 static bool isLoadConditional(LoadInst *Load, Loop *L) {
142 return Load->getParent() != L->getHeader();
145 /// \brief The per-loop class that does most of the work.
146 class LoadEliminationForLoop {
148 LoadEliminationForLoop(Loop *L, LoopInfo *LI, const LoopAccessInfo &LAI,
150 : L(L), LI(LI), LAI(LAI), DT(DT), PSE(LAI.getPSE()) {}
152 /// \brief Look through the loop-carried and loop-independent dependences in
153 /// this loop and find store->load dependences.
155 /// Note that no candidate is returned if LAA has failed to analyze the loop
156 /// (e.g. if it's not bottom-tested, contains volatile memops, etc.)
157 std::forward_list<StoreToLoadForwardingCandidate>
158 findStoreToLoadDependences(const LoopAccessInfo &LAI) {
159 std::forward_list<StoreToLoadForwardingCandidate> Candidates;
161 const auto *Deps = LAI.getDepChecker().getDependences();
165 // Find store->load dependences (consequently true dep). Both lexically
166 // forward and backward dependences qualify. Disqualify loads that have
167 // other unknown dependences.
169 SmallSet<Instruction *, 4> LoadsWithUnknownDepedence;
171 for (const auto &Dep : *Deps) {
172 Instruction *Source = Dep.getSource(LAI);
173 Instruction *Destination = Dep.getDestination(LAI);
175 if (Dep.Type == MemoryDepChecker::Dependence::Unknown) {
176 if (isa<LoadInst>(Source))
177 LoadsWithUnknownDepedence.insert(Source);
178 if (isa<LoadInst>(Destination))
179 LoadsWithUnknownDepedence.insert(Destination);
183 if (Dep.isBackward())
184 // Note that the designations source and destination follow the program
185 // order, i.e. source is always first. (The direction is given by the
187 std::swap(Source, Destination);
189 assert(Dep.isForward() && "Needs to be a forward dependence");
191 auto *Store = dyn_cast<StoreInst>(Source);
194 auto *Load = dyn_cast<LoadInst>(Destination);
198 // Only progagate the value if they are of the same type.
199 if (Store->getPointerOperand()->getType() !=
200 Load->getPointerOperand()->getType())
203 Candidates.emplace_front(Load, Store);
206 if (!LoadsWithUnknownDepedence.empty())
207 Candidates.remove_if([&](const StoreToLoadForwardingCandidate &C) {
208 return LoadsWithUnknownDepedence.count(C.Load);
214 /// \brief Return the index of the instruction according to program order.
215 unsigned getInstrIndex(Instruction *Inst) {
216 auto I = InstOrder.find(Inst);
217 assert(I != InstOrder.end() && "No index for instruction");
221 /// \brief If a load has multiple candidates associated (i.e. different
222 /// stores), it means that it could be forwarding from multiple stores
223 /// depending on control flow. Remove these candidates.
225 /// Here, we rely on LAA to include the relevant loop-independent dependences.
226 /// LAA is known to omit these in the very simple case when the read and the
227 /// write within an alias set always takes place using the *same* pointer.
229 /// However, we know that this is not the case here, i.e. we can rely on LAA
230 /// to provide us with loop-independent dependences for the cases we're
231 /// interested. Consider the case for example where a loop-independent
232 /// dependece S1->S2 invalidates the forwarding S3->S2.
236 /// A[i+1] = ... (S3)
238 /// LAA will perform dependence analysis here because there are two
239 /// *different* pointers involved in the same alias set (&A[i] and &A[i+1]).
240 void removeDependencesFromMultipleStores(
241 std::forward_list<StoreToLoadForwardingCandidate> &Candidates) {
242 // If Store is nullptr it means that we have multiple stores forwarding to
244 typedef DenseMap<LoadInst *, const StoreToLoadForwardingCandidate *>
246 LoadToSingleCandT LoadToSingleCand;
248 for (const auto &Cand : Candidates) {
250 LoadToSingleCandT::iterator Iter;
252 std::tie(Iter, NewElt) =
253 LoadToSingleCand.insert(std::make_pair(Cand.Load, &Cand));
255 const StoreToLoadForwardingCandidate *&OtherCand = Iter->second;
256 // Already multiple stores forward to this load.
257 if (OtherCand == nullptr)
260 // Handle the very basic case when the two stores are in the same block
261 // so deciding which one forwards is easy. The later one forwards as
262 // long as they both have a dependence distance of one to the load.
263 if (Cand.Store->getParent() == OtherCand->Store->getParent() &&
264 Cand.isDependenceDistanceOfOne(PSE, L) &&
265 OtherCand->isDependenceDistanceOfOne(PSE, L)) {
266 // They are in the same block, the later one will forward to the load.
267 if (getInstrIndex(OtherCand->Store) < getInstrIndex(Cand.Store))
274 Candidates.remove_if([&](const StoreToLoadForwardingCandidate &Cand) {
275 if (LoadToSingleCand[Cand.Load] != &Cand) {
276 DEBUG(dbgs() << "Removing from candidates: \n" << Cand
277 << " The load may have multiple stores forwarding to "
285 /// \brief Given two pointers operations by their RuntimePointerChecking
286 /// indices, return true if they require an alias check.
288 /// We need a check if one is a pointer for a candidate load and the other is
289 /// a pointer for a possibly intervening store.
290 bool needsChecking(unsigned PtrIdx1, unsigned PtrIdx2,
291 const SmallSet<Value *, 4> &PtrsWrittenOnFwdingPath,
292 const std::set<Value *> &CandLoadPtrs) {
294 LAI.getRuntimePointerChecking()->getPointerInfo(PtrIdx1).PointerValue;
296 LAI.getRuntimePointerChecking()->getPointerInfo(PtrIdx2).PointerValue;
297 return ((PtrsWrittenOnFwdingPath.count(Ptr1) && CandLoadPtrs.count(Ptr2)) ||
298 (PtrsWrittenOnFwdingPath.count(Ptr2) && CandLoadPtrs.count(Ptr1)));
301 /// \brief Return pointers that are possibly written to on the path from a
302 /// forwarding store to a load.
304 /// These pointers need to be alias-checked against the forwarding candidates.
305 SmallSet<Value *, 4> findPointersWrittenOnForwardingPath(
306 const SmallVectorImpl<StoreToLoadForwardingCandidate> &Candidates) {
307 // From FirstStore to LastLoad neither of the elimination candidate loads
308 // should overlap with any of the stores.
313 // ld1 B[i] <-------,
314 // ld0 A[i] <----, | * LastLoad
317 // st3 B[i+1] -- | -' * FirstStore
321 // st0 forwards to ld0 if the accesses in st4 and st1 don't overlap with
325 std::max_element(Candidates.begin(), Candidates.end(),
326 [&](const StoreToLoadForwardingCandidate &A,
327 const StoreToLoadForwardingCandidate &B) {
328 return getInstrIndex(A.Load) < getInstrIndex(B.Load);
331 StoreInst *FirstStore =
332 std::min_element(Candidates.begin(), Candidates.end(),
333 [&](const StoreToLoadForwardingCandidate &A,
334 const StoreToLoadForwardingCandidate &B) {
335 return getInstrIndex(A.Store) <
336 getInstrIndex(B.Store);
340 // We're looking for stores after the first forwarding store until the end
341 // of the loop, then from the beginning of the loop until the last
342 // forwarded-to load. Collect the pointer for the stores.
343 SmallSet<Value *, 4> PtrsWrittenOnFwdingPath;
345 auto InsertStorePtr = [&](Instruction *I) {
346 if (auto *S = dyn_cast<StoreInst>(I))
347 PtrsWrittenOnFwdingPath.insert(S->getPointerOperand());
349 const auto &MemInstrs = LAI.getDepChecker().getMemoryInstructions();
350 std::for_each(MemInstrs.begin() + getInstrIndex(FirstStore) + 1,
351 MemInstrs.end(), InsertStorePtr);
352 std::for_each(MemInstrs.begin(), &MemInstrs[getInstrIndex(LastLoad)],
355 return PtrsWrittenOnFwdingPath;
358 /// \brief Determine the pointer alias checks to prove that there are no
359 /// intervening stores.
360 SmallVector<RuntimePointerChecking::PointerCheck, 4> collectMemchecks(
361 const SmallVectorImpl<StoreToLoadForwardingCandidate> &Candidates) {
363 SmallSet<Value *, 4> PtrsWrittenOnFwdingPath =
364 findPointersWrittenOnForwardingPath(Candidates);
366 // Collect the pointers of the candidate loads.
367 // FIXME: SmallSet does not work with std::inserter.
368 std::set<Value *> CandLoadPtrs;
369 transform(Candidates,
370 std::inserter(CandLoadPtrs, CandLoadPtrs.begin()),
371 std::mem_fn(&StoreToLoadForwardingCandidate::getLoadPtr));
373 const auto &AllChecks = LAI.getRuntimePointerChecking()->getChecks();
374 SmallVector<RuntimePointerChecking::PointerCheck, 4> Checks;
376 std::copy_if(AllChecks.begin(), AllChecks.end(), std::back_inserter(Checks),
377 [&](const RuntimePointerChecking::PointerCheck &Check) {
378 for (auto PtrIdx1 : Check.first->Members)
379 for (auto PtrIdx2 : Check.second->Members)
380 if (needsChecking(PtrIdx1, PtrIdx2,
381 PtrsWrittenOnFwdingPath, CandLoadPtrs))
386 DEBUG(dbgs() << "\nPointer Checks (count: " << Checks.size() << "):\n");
387 DEBUG(LAI.getRuntimePointerChecking()->printChecks(dbgs(), Checks));
392 /// \brief Perform the transformation for a candidate.
394 propagateStoredValueToLoadUsers(const StoreToLoadForwardingCandidate &Cand,
400 // store %y, %gep_i_plus_1
405 // %x.initial = load %gep_0
407 // %x.storeforward = phi [%x.initial, %ph] [%y, %loop]
408 // %x = load %gep_i <---- now dead
409 // = ... %x.storeforward
410 // store %y, %gep_i_plus_1
412 Value *Ptr = Cand.Load->getPointerOperand();
413 auto *PtrSCEV = cast<SCEVAddRecExpr>(PSE.getSCEV(Ptr));
414 auto *PH = L->getLoopPreheader();
415 Value *InitialPtr = SEE.expandCodeFor(PtrSCEV->getStart(), Ptr->getType(),
416 PH->getTerminator());
418 new LoadInst(InitialPtr, "load_initial", PH->getTerminator());
419 PHINode *PHI = PHINode::Create(Initial->getType(), 2, "store_forwarded",
420 &L->getHeader()->front());
421 PHI->addIncoming(Initial, PH);
422 PHI->addIncoming(Cand.Store->getOperand(0), L->getLoopLatch());
424 Cand.Load->replaceAllUsesWith(PHI);
427 /// \brief Top-level driver for each loop: find store->load forwarding
428 /// candidates, add run-time checks and perform transformation.
430 DEBUG(dbgs() << "\nIn \"" << L->getHeader()->getParent()->getName()
431 << "\" checking " << *L << "\n");
432 // Look for store-to-load forwarding cases across the
438 // store %y, %gep_i_plus_1
443 // %x.initial = load %gep_0
445 // %x.storeforward = phi [%x.initial, %ph] [%y, %loop]
446 // %x = load %gep_i <---- now dead
447 // = ... %x.storeforward
448 // store %y, %gep_i_plus_1
450 // First start with store->load dependences.
451 auto StoreToLoadDependences = findStoreToLoadDependences(LAI);
452 if (StoreToLoadDependences.empty())
455 // Generate an index for each load and store according to the original
456 // program order. This will be used later.
457 InstOrder = LAI.getDepChecker().generateInstructionOrderMap();
459 // To keep things simple for now, remove those where the load is potentially
460 // fed by multiple stores.
461 removeDependencesFromMultipleStores(StoreToLoadDependences);
462 if (StoreToLoadDependences.empty())
465 // Filter the candidates further.
466 SmallVector<StoreToLoadForwardingCandidate, 4> Candidates;
467 unsigned NumForwarding = 0;
468 for (const StoreToLoadForwardingCandidate Cand : StoreToLoadDependences) {
469 DEBUG(dbgs() << "Candidate " << Cand);
471 // Make sure that the stored values is available everywhere in the loop in
472 // the next iteration.
473 if (!doesStoreDominatesAllLatches(Cand.Store->getParent(), L, DT))
476 // If the load is conditional we can't hoist its 0-iteration instance to
477 // the preheader because that would make it unconditional. Thus we would
478 // access a memory location that the original loop did not access.
479 if (isLoadConditional(Cand.Load, L))
482 // Check whether the SCEV difference is the same as the induction step,
483 // thus we load the value in the next iteration.
484 if (!Cand.isDependenceDistanceOfOne(PSE, L))
490 << ". Valid store-to-load forwarding across the loop backedge\n");
491 Candidates.push_back(Cand);
493 if (Candidates.empty())
496 // Check intervening may-alias stores. These need runtime checks for alias
498 SmallVector<RuntimePointerChecking::PointerCheck, 4> Checks =
499 collectMemchecks(Candidates);
501 // Too many checks are likely to outweigh the benefits of forwarding.
502 if (Checks.size() > Candidates.size() * CheckPerElim) {
503 DEBUG(dbgs() << "Too many run-time checks needed.\n");
507 if (LAI.getPSE().getUnionPredicate().getComplexity() >
508 LoadElimSCEVCheckThreshold) {
509 DEBUG(dbgs() << "Too many SCEV run-time checks needed.\n");
513 if (!Checks.empty() || !LAI.getPSE().getUnionPredicate().isAlwaysTrue()) {
514 if (L->getHeader()->getParent()->optForSize()) {
515 DEBUG(dbgs() << "Versioning is needed but not allowed when optimizing "
520 if (!L->isLoopSimplifyForm()) {
521 DEBUG(dbgs() << "Loop is not is loop-simplify form");
525 // Point of no-return, start the transformation. First, version the loop
528 LoopVersioning LV(LAI, L, LI, DT, PSE.getSE(), false);
529 LV.setAliasChecks(std::move(Checks));
530 LV.setSCEVChecks(LAI.getPSE().getUnionPredicate());
534 // Next, propagate the value stored by the store to the users of the load.
535 // Also for the first iteration, generate the initial value of the load.
536 SCEVExpander SEE(*PSE.getSE(), L->getHeader()->getModule()->getDataLayout(),
538 for (const auto &Cand : Candidates)
539 propagateStoredValueToLoadUsers(Cand, SEE);
540 NumLoopLoadEliminted += NumForwarding;
548 /// \brief Maps the load/store instructions to their index according to
550 DenseMap<Instruction *, unsigned> InstOrder;
554 const LoopAccessInfo &LAI;
556 PredicatedScalarEvolution PSE;
559 /// \brief The pass. Most of the work is delegated to the per-loop
560 /// LoadEliminationForLoop class.
561 class LoopLoadElimination : public FunctionPass {
563 LoopLoadElimination() : FunctionPass(ID) {
564 initializeLoopLoadEliminationPass(*PassRegistry::getPassRegistry());
567 bool runOnFunction(Function &F) override {
571 auto *LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
572 auto *LAA = &getAnalysis<LoopAccessLegacyAnalysis>();
573 auto *DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
575 // Build up a worklist of inner-loops to vectorize. This is necessary as the
576 // act of distributing a loop creates new loops and can invalidate iterators
578 SmallVector<Loop *, 8> Worklist;
580 for (Loop *TopLevelLoop : *LI)
581 for (Loop *L : depth_first(TopLevelLoop))
582 // We only handle inner-most loops.
584 Worklist.push_back(L);
586 // Now walk the identified inner loops.
587 bool Changed = false;
588 for (Loop *L : Worklist) {
589 const LoopAccessInfo &LAI = LAA->getInfo(L);
590 // The actual work is performed by LoadEliminationForLoop.
591 LoadEliminationForLoop LEL(L, LI, LAI, DT);
592 Changed |= LEL.processLoop();
595 // Process each loop nest in the function.
599 void getAnalysisUsage(AnalysisUsage &AU) const override {
600 AU.addRequiredID(LoopSimplifyID);
601 AU.addRequired<LoopInfoWrapperPass>();
602 AU.addPreserved<LoopInfoWrapperPass>();
603 AU.addRequired<LoopAccessLegacyAnalysis>();
604 AU.addRequired<ScalarEvolutionWrapperPass>();
605 AU.addRequired<DominatorTreeWrapperPass>();
606 AU.addPreserved<DominatorTreeWrapperPass>();
607 AU.addPreserved<GlobalsAAWrapperPass>();
613 } // end anonymous namespace
615 char LoopLoadElimination::ID;
616 static const char LLE_name[] = "Loop Load Elimination";
618 INITIALIZE_PASS_BEGIN(LoopLoadElimination, LLE_OPTION, LLE_name, false, false)
619 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
620 INITIALIZE_PASS_DEPENDENCY(LoopAccessLegacyAnalysis)
621 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
622 INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass)
623 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
624 INITIALIZE_PASS_END(LoopLoadElimination, LLE_OPTION, LLE_name, false, false)
628 FunctionPass *createLoopLoadEliminationPass() {
629 return new LoopLoadElimination();
632 } // end namespace llvm