1 //===- LoopAccessAnalysis.cpp - Loop Access Analysis Implementation --------==//
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
7 //===----------------------------------------------------------------------===//
9 // The implementation for the loop memory dependence that was originally
10 // developed for the loop vectorizer.
12 //===----------------------------------------------------------------------===//
14 #include "llvm/Analysis/LoopAccessAnalysis.h"
15 #include "llvm/ADT/APInt.h"
16 #include "llvm/ADT/DenseMap.h"
17 #include "llvm/ADT/DepthFirstIterator.h"
18 #include "llvm/ADT/EquivalenceClasses.h"
19 #include "llvm/ADT/PointerIntPair.h"
20 #include "llvm/ADT/STLExtras.h"
21 #include "llvm/ADT/SetVector.h"
22 #include "llvm/ADT/SmallPtrSet.h"
23 #include "llvm/ADT/SmallSet.h"
24 #include "llvm/ADT/SmallVector.h"
25 #include "llvm/ADT/iterator_range.h"
26 #include "llvm/Analysis/AliasAnalysis.h"
27 #include "llvm/Analysis/AliasSetTracker.h"
28 #include "llvm/Analysis/LoopAnalysisManager.h"
29 #include "llvm/Analysis/LoopInfo.h"
30 #include "llvm/Analysis/MemoryLocation.h"
31 #include "llvm/Analysis/OptimizationRemarkEmitter.h"
32 #include "llvm/Analysis/ScalarEvolution.h"
33 #include "llvm/Analysis/ScalarEvolutionExpander.h"
34 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
35 #include "llvm/Analysis/TargetLibraryInfo.h"
36 #include "llvm/Analysis/ValueTracking.h"
37 #include "llvm/Analysis/VectorUtils.h"
38 #include "llvm/IR/BasicBlock.h"
39 #include "llvm/IR/Constants.h"
40 #include "llvm/IR/DataLayout.h"
41 #include "llvm/IR/DebugLoc.h"
42 #include "llvm/IR/DerivedTypes.h"
43 #include "llvm/IR/DiagnosticInfo.h"
44 #include "llvm/IR/Dominators.h"
45 #include "llvm/IR/Function.h"
46 #include "llvm/IR/IRBuilder.h"
47 #include "llvm/IR/InstrTypes.h"
48 #include "llvm/IR/Instruction.h"
49 #include "llvm/IR/Instructions.h"
50 #include "llvm/IR/Operator.h"
51 #include "llvm/IR/PassManager.h"
52 #include "llvm/IR/Type.h"
53 #include "llvm/IR/Value.h"
54 #include "llvm/IR/ValueHandle.h"
55 #include "llvm/Pass.h"
56 #include "llvm/Support/Casting.h"
57 #include "llvm/Support/CommandLine.h"
58 #include "llvm/Support/Debug.h"
59 #include "llvm/Support/ErrorHandling.h"
60 #include "llvm/Support/raw_ostream.h"
71 #define DEBUG_TYPE "loop-accesses"
73 static cl::opt<unsigned, true>
74 VectorizationFactor("force-vector-width", cl::Hidden,
75 cl::desc("Sets the SIMD width. Zero is autoselect."),
76 cl::location(VectorizerParams::VectorizationFactor));
77 unsigned VectorizerParams::VectorizationFactor;
79 static cl::opt<unsigned, true>
80 VectorizationInterleave("force-vector-interleave", cl::Hidden,
81 cl::desc("Sets the vectorization interleave count. "
82 "Zero is autoselect."),
84 VectorizerParams::VectorizationInterleave));
85 unsigned VectorizerParams::VectorizationInterleave;
87 static cl::opt<unsigned, true> RuntimeMemoryCheckThreshold(
88 "runtime-memory-check-threshold", cl::Hidden,
89 cl::desc("When performing memory disambiguation checks at runtime do not "
90 "generate more than this number of comparisons (default = 8)."),
91 cl::location(VectorizerParams::RuntimeMemoryCheckThreshold), cl::init(8));
92 unsigned VectorizerParams::RuntimeMemoryCheckThreshold;
94 /// The maximum iterations used to merge memory checks
95 static cl::opt<unsigned> MemoryCheckMergeThreshold(
96 "memory-check-merge-threshold", cl::Hidden,
97 cl::desc("Maximum number of comparisons done when trying to merge "
98 "runtime memory checks. (default = 100)"),
101 /// Maximum SIMD width.
102 const unsigned VectorizerParams::MaxVectorWidth = 64;
104 /// We collect dependences up to this threshold.
105 static cl::opt<unsigned>
106 MaxDependences("max-dependences", cl::Hidden,
107 cl::desc("Maximum number of dependences collected by "
108 "loop-access analysis (default = 100)"),
111 /// This enables versioning on the strides of symbolically striding memory
112 /// accesses in code like the following.
113 /// for (i = 0; i < N; ++i)
114 /// A[i * Stride1] += B[i * Stride2] ...
116 /// Will be roughly translated to
117 /// if (Stride1 == 1 && Stride2 == 1) {
118 /// for (i = 0; i < N; i+=4)
122 static cl::opt<bool> EnableMemAccessVersioning(
123 "enable-mem-access-versioning", cl::init(true), cl::Hidden,
124 cl::desc("Enable symbolic stride memory access versioning"));
126 /// Enable store-to-load forwarding conflict detection. This option can
127 /// be disabled for correctness testing.
128 static cl::opt<bool> EnableForwardingConflictDetection(
129 "store-to-load-forwarding-conflict-detection", cl::Hidden,
130 cl::desc("Enable conflict detection in loop-access analysis"),
133 bool VectorizerParams::isInterleaveForced() {
134 return ::VectorizationInterleave.getNumOccurrences() > 0;
137 Value *llvm::stripIntegerCast(Value *V) {
138 if (auto *CI = dyn_cast<CastInst>(V))
139 if (CI->getOperand(0)->getType()->isIntegerTy())
140 return CI->getOperand(0);
144 const SCEV *llvm::replaceSymbolicStrideSCEV(PredicatedScalarEvolution &PSE,
145 const ValueToValueMap &PtrToStride,
146 Value *Ptr, Value *OrigPtr) {
147 const SCEV *OrigSCEV = PSE.getSCEV(Ptr);
149 // If there is an entry in the map return the SCEV of the pointer with the
150 // symbolic stride replaced by one.
151 ValueToValueMap::const_iterator SI =
152 PtrToStride.find(OrigPtr ? OrigPtr : Ptr);
153 if (SI != PtrToStride.end()) {
154 Value *StrideVal = SI->second;
157 StrideVal = stripIntegerCast(StrideVal);
159 ScalarEvolution *SE = PSE.getSE();
160 const auto *U = cast<SCEVUnknown>(SE->getSCEV(StrideVal));
162 static_cast<const SCEVConstant *>(SE->getOne(StrideVal->getType()));
164 PSE.addPredicate(*SE->getEqualPredicate(U, CT));
165 auto *Expr = PSE.getSCEV(Ptr);
167 LLVM_DEBUG(dbgs() << "LAA: Replacing SCEV: " << *OrigSCEV
168 << " by: " << *Expr << "\n");
172 // Otherwise, just return the SCEV of the original pointer.
176 /// Calculate Start and End points of memory access.
177 /// Let's assume A is the first access and B is a memory access on N-th loop
178 /// iteration. Then B is calculated as:
180 /// Step value may be positive or negative.
181 /// N is a calculated back-edge taken count:
182 /// N = (TripCount > 0) ? RoundDown(TripCount -1 , VF) : 0
183 /// Start and End points are calculated in the following way:
184 /// Start = UMIN(A, B) ; End = UMAX(A, B) + SizeOfElt,
185 /// where SizeOfElt is the size of single memory access in bytes.
187 /// There is no conflict when the intervals are disjoint:
188 /// NoConflict = (P2.Start >= P1.End) || (P1.Start >= P2.End)
189 void RuntimePointerChecking::insert(Loop *Lp, Value *Ptr, bool WritePtr,
190 unsigned DepSetId, unsigned ASId,
191 const ValueToValueMap &Strides,
192 PredicatedScalarEvolution &PSE) {
193 // Get the stride replaced scev.
194 const SCEV *Sc = replaceSymbolicStrideSCEV(PSE, Strides, Ptr);
195 ScalarEvolution *SE = PSE.getSE();
200 if (SE->isLoopInvariant(Sc, Lp))
201 ScStart = ScEnd = Sc;
203 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Sc);
204 assert(AR && "Invalid addrec expression");
205 const SCEV *Ex = PSE.getBackedgeTakenCount();
207 ScStart = AR->getStart();
208 ScEnd = AR->evaluateAtIteration(Ex, *SE);
209 const SCEV *Step = AR->getStepRecurrence(*SE);
211 // For expressions with negative step, the upper bound is ScStart and the
212 // lower bound is ScEnd.
213 if (const auto *CStep = dyn_cast<SCEVConstant>(Step)) {
214 if (CStep->getValue()->isNegative())
215 std::swap(ScStart, ScEnd);
217 // Fallback case: the step is not constant, but we can still
218 // get the upper and lower bounds of the interval by using min/max
220 ScStart = SE->getUMinExpr(ScStart, ScEnd);
221 ScEnd = SE->getUMaxExpr(AR->getStart(), ScEnd);
223 // Add the size of the pointed element to ScEnd.
225 Ptr->getType()->getPointerElementType()->getScalarSizeInBits() / 8;
226 const SCEV *EltSizeSCEV = SE->getConstant(ScEnd->getType(), EltSize);
227 ScEnd = SE->getAddExpr(ScEnd, EltSizeSCEV);
230 Pointers.emplace_back(Ptr, ScStart, ScEnd, WritePtr, DepSetId, ASId, Sc);
233 SmallVector<RuntimePointerChecking::PointerCheck, 4>
234 RuntimePointerChecking::generateChecks() const {
235 SmallVector<PointerCheck, 4> Checks;
237 for (unsigned I = 0; I < CheckingGroups.size(); ++I) {
238 for (unsigned J = I + 1; J < CheckingGroups.size(); ++J) {
239 const RuntimePointerChecking::CheckingPtrGroup &CGI = CheckingGroups[I];
240 const RuntimePointerChecking::CheckingPtrGroup &CGJ = CheckingGroups[J];
242 if (needsChecking(CGI, CGJ))
243 Checks.push_back(std::make_pair(&CGI, &CGJ));
249 void RuntimePointerChecking::generateChecks(
250 MemoryDepChecker::DepCandidates &DepCands, bool UseDependencies) {
251 assert(Checks.empty() && "Checks is not empty");
252 groupChecks(DepCands, UseDependencies);
253 Checks = generateChecks();
256 bool RuntimePointerChecking::needsChecking(const CheckingPtrGroup &M,
257 const CheckingPtrGroup &N) const {
258 for (unsigned I = 0, EI = M.Members.size(); EI != I; ++I)
259 for (unsigned J = 0, EJ = N.Members.size(); EJ != J; ++J)
260 if (needsChecking(M.Members[I], N.Members[J]))
265 /// Compare \p I and \p J and return the minimum.
266 /// Return nullptr in case we couldn't find an answer.
267 static const SCEV *getMinFromExprs(const SCEV *I, const SCEV *J,
268 ScalarEvolution *SE) {
269 const SCEV *Diff = SE->getMinusSCEV(J, I);
270 const SCEVConstant *C = dyn_cast<const SCEVConstant>(Diff);
274 if (C->getValue()->isNegative())
279 bool RuntimePointerChecking::CheckingPtrGroup::addPointer(unsigned Index) {
280 const SCEV *Start = RtCheck.Pointers[Index].Start;
281 const SCEV *End = RtCheck.Pointers[Index].End;
283 // Compare the starts and ends with the known minimum and maximum
284 // of this set. We need to know how we compare against the min/max
285 // of the set in order to be able to emit memchecks.
286 const SCEV *Min0 = getMinFromExprs(Start, Low, RtCheck.SE);
290 const SCEV *Min1 = getMinFromExprs(End, High, RtCheck.SE);
294 // Update the low bound expression if we've found a new min value.
298 // Update the high bound expression if we've found a new max value.
302 Members.push_back(Index);
306 void RuntimePointerChecking::groupChecks(
307 MemoryDepChecker::DepCandidates &DepCands, bool UseDependencies) {
308 // We build the groups from dependency candidates equivalence classes
310 // - We know that pointers in the same equivalence class share
311 // the same underlying object and therefore there is a chance
312 // that we can compare pointers
313 // - We wouldn't be able to merge two pointers for which we need
314 // to emit a memcheck. The classes in DepCands are already
315 // conveniently built such that no two pointers in the same
316 // class need checking against each other.
318 // We use the following (greedy) algorithm to construct the groups
319 // For every pointer in the equivalence class:
320 // For each existing group:
321 // - if the difference between this pointer and the min/max bounds
322 // of the group is a constant, then make the pointer part of the
323 // group and update the min/max bounds of that group as required.
325 CheckingGroups.clear();
327 // If we need to check two pointers to the same underlying object
328 // with a non-constant difference, we shouldn't perform any pointer
329 // grouping with those pointers. This is because we can easily get
330 // into cases where the resulting check would return false, even when
331 // the accesses are safe.
333 // The following example shows this:
334 // for (i = 0; i < 1000; ++i)
335 // a[5000 + i * m] = a[i] + a[i + 9000]
337 // Here grouping gives a check of (5000, 5000 + 1000 * m) against
338 // (0, 10000) which is always false. However, if m is 1, there is no
339 // dependence. Not grouping the checks for a[i] and a[i + 9000] allows
340 // us to perform an accurate check in this case.
342 // The above case requires that we have an UnknownDependence between
343 // accesses to the same underlying object. This cannot happen unless
344 // FoundNonConstantDistanceDependence is set, and therefore UseDependencies
345 // is also false. In this case we will use the fallback path and create
346 // separate checking groups for all pointers.
348 // If we don't have the dependency partitions, construct a new
349 // checking pointer group for each pointer. This is also required
350 // for correctness, because in this case we can have checking between
351 // pointers to the same underlying object.
352 if (!UseDependencies) {
353 for (unsigned I = 0; I < Pointers.size(); ++I)
354 CheckingGroups.push_back(CheckingPtrGroup(I, *this));
358 unsigned TotalComparisons = 0;
360 DenseMap<Value *, unsigned> PositionMap;
361 for (unsigned Index = 0; Index < Pointers.size(); ++Index)
362 PositionMap[Pointers[Index].PointerValue] = Index;
364 // We need to keep track of what pointers we've already seen so we
365 // don't process them twice.
366 SmallSet<unsigned, 2> Seen;
368 // Go through all equivalence classes, get the "pointer check groups"
369 // and add them to the overall solution. We use the order in which accesses
370 // appear in 'Pointers' to enforce determinism.
371 for (unsigned I = 0; I < Pointers.size(); ++I) {
372 // We've seen this pointer before, and therefore already processed
373 // its equivalence class.
377 MemoryDepChecker::MemAccessInfo Access(Pointers[I].PointerValue,
378 Pointers[I].IsWritePtr);
380 SmallVector<CheckingPtrGroup, 2> Groups;
381 auto LeaderI = DepCands.findValue(DepCands.getLeaderValue(Access));
383 // Because DepCands is constructed by visiting accesses in the order in
384 // which they appear in alias sets (which is deterministic) and the
385 // iteration order within an equivalence class member is only dependent on
386 // the order in which unions and insertions are performed on the
387 // equivalence class, the iteration order is deterministic.
388 for (auto MI = DepCands.member_begin(LeaderI), ME = DepCands.member_end();
390 unsigned Pointer = PositionMap[MI->getPointer()];
392 // Mark this pointer as seen.
393 Seen.insert(Pointer);
395 // Go through all the existing sets and see if we can find one
396 // which can include this pointer.
397 for (CheckingPtrGroup &Group : Groups) {
398 // Don't perform more than a certain amount of comparisons.
399 // This should limit the cost of grouping the pointers to something
400 // reasonable. If we do end up hitting this threshold, the algorithm
401 // will create separate groups for all remaining pointers.
402 if (TotalComparisons > MemoryCheckMergeThreshold)
407 if (Group.addPointer(Pointer)) {
414 // We couldn't add this pointer to any existing set or the threshold
415 // for the number of comparisons has been reached. Create a new group
416 // to hold the current pointer.
417 Groups.push_back(CheckingPtrGroup(Pointer, *this));
420 // We've computed the grouped checks for this partition.
421 // Save the results and continue with the next one.
422 llvm::copy(Groups, std::back_inserter(CheckingGroups));
426 bool RuntimePointerChecking::arePointersInSamePartition(
427 const SmallVectorImpl<int> &PtrToPartition, unsigned PtrIdx1,
429 return (PtrToPartition[PtrIdx1] != -1 &&
430 PtrToPartition[PtrIdx1] == PtrToPartition[PtrIdx2]);
433 bool RuntimePointerChecking::needsChecking(unsigned I, unsigned J) const {
434 const PointerInfo &PointerI = Pointers[I];
435 const PointerInfo &PointerJ = Pointers[J];
437 // No need to check if two readonly pointers intersect.
438 if (!PointerI.IsWritePtr && !PointerJ.IsWritePtr)
441 // Only need to check pointers between two different dependency sets.
442 if (PointerI.DependencySetId == PointerJ.DependencySetId)
445 // Only need to check pointers in the same alias set.
446 if (PointerI.AliasSetId != PointerJ.AliasSetId)
452 void RuntimePointerChecking::printChecks(
453 raw_ostream &OS, const SmallVectorImpl<PointerCheck> &Checks,
454 unsigned Depth) const {
456 for (const auto &Check : Checks) {
457 const auto &First = Check.first->Members, &Second = Check.second->Members;
459 OS.indent(Depth) << "Check " << N++ << ":\n";
461 OS.indent(Depth + 2) << "Comparing group (" << Check.first << "):\n";
462 for (unsigned K = 0; K < First.size(); ++K)
463 OS.indent(Depth + 2) << *Pointers[First[K]].PointerValue << "\n";
465 OS.indent(Depth + 2) << "Against group (" << Check.second << "):\n";
466 for (unsigned K = 0; K < Second.size(); ++K)
467 OS.indent(Depth + 2) << *Pointers[Second[K]].PointerValue << "\n";
471 void RuntimePointerChecking::print(raw_ostream &OS, unsigned Depth) const {
473 OS.indent(Depth) << "Run-time memory checks:\n";
474 printChecks(OS, Checks, Depth);
476 OS.indent(Depth) << "Grouped accesses:\n";
477 for (unsigned I = 0; I < CheckingGroups.size(); ++I) {
478 const auto &CG = CheckingGroups[I];
480 OS.indent(Depth + 2) << "Group " << &CG << ":\n";
481 OS.indent(Depth + 4) << "(Low: " << *CG.Low << " High: " << *CG.High
483 for (unsigned J = 0; J < CG.Members.size(); ++J) {
484 OS.indent(Depth + 6) << "Member: " << *Pointers[CG.Members[J]].Expr
492 /// Analyses memory accesses in a loop.
494 /// Checks whether run time pointer checks are needed and builds sets for data
495 /// dependence checking.
496 class AccessAnalysis {
498 /// Read or write access location.
499 typedef PointerIntPair<Value *, 1, bool> MemAccessInfo;
500 typedef SmallVector<MemAccessInfo, 8> MemAccessInfoList;
502 AccessAnalysis(const DataLayout &Dl, Loop *TheLoop, AliasAnalysis *AA,
503 LoopInfo *LI, MemoryDepChecker::DepCandidates &DA,
504 PredicatedScalarEvolution &PSE)
505 : DL(Dl), TheLoop(TheLoop), AST(*AA), LI(LI), DepCands(DA),
506 IsRTCheckAnalysisNeeded(false), PSE(PSE) {}
508 /// Register a load and whether it is only read from.
509 void addLoad(MemoryLocation &Loc, bool IsReadOnly) {
510 Value *Ptr = const_cast<Value*>(Loc.Ptr);
511 AST.add(Ptr, LocationSize::unknown(), Loc.AATags);
512 Accesses.insert(MemAccessInfo(Ptr, false));
514 ReadOnlyPtr.insert(Ptr);
517 /// Register a store.
518 void addStore(MemoryLocation &Loc) {
519 Value *Ptr = const_cast<Value*>(Loc.Ptr);
520 AST.add(Ptr, LocationSize::unknown(), Loc.AATags);
521 Accesses.insert(MemAccessInfo(Ptr, true));
524 /// Check if we can emit a run-time no-alias check for \p Access.
526 /// Returns true if we can emit a run-time no alias check for \p Access.
527 /// If we can check this access, this also adds it to a dependence set and
528 /// adds a run-time to check for it to \p RtCheck. If \p Assume is true,
529 /// we will attempt to use additional run-time checks in order to get
530 /// the bounds of the pointer.
531 bool createCheckForAccess(RuntimePointerChecking &RtCheck,
532 MemAccessInfo Access,
533 const ValueToValueMap &Strides,
534 DenseMap<Value *, unsigned> &DepSetId,
535 Loop *TheLoop, unsigned &RunningDepId,
536 unsigned ASId, bool ShouldCheckStride,
539 /// Check whether we can check the pointers at runtime for
540 /// non-intersection.
542 /// Returns true if we need no check or if we do and we can generate them
543 /// (i.e. the pointers have computable bounds).
544 bool canCheckPtrAtRT(RuntimePointerChecking &RtCheck, ScalarEvolution *SE,
545 Loop *TheLoop, const ValueToValueMap &Strides,
546 bool ShouldCheckWrap = false);
548 /// Goes over all memory accesses, checks whether a RT check is needed
549 /// and builds sets of dependent accesses.
550 void buildDependenceSets() {
551 processMemAccesses();
554 /// Initial processing of memory accesses determined that we need to
555 /// perform dependency checking.
557 /// Note that this can later be cleared if we retry memcheck analysis without
558 /// dependency checking (i.e. FoundNonConstantDistanceDependence).
559 bool isDependencyCheckNeeded() { return !CheckDeps.empty(); }
561 /// We decided that no dependence analysis would be used. Reset the state.
562 void resetDepChecks(MemoryDepChecker &DepChecker) {
564 DepChecker.clearDependences();
567 MemAccessInfoList &getDependenciesToCheck() { return CheckDeps; }
570 typedef SetVector<MemAccessInfo> PtrAccessSet;
572 /// Go over all memory access and check whether runtime pointer checks
573 /// are needed and build sets of dependency check candidates.
574 void processMemAccesses();
576 /// Set of all accesses.
577 PtrAccessSet Accesses;
579 const DataLayout &DL;
581 /// The loop being checked.
584 /// List of accesses that need a further dependence check.
585 MemAccessInfoList CheckDeps;
587 /// Set of pointers that are read only.
588 SmallPtrSet<Value*, 16> ReadOnlyPtr;
590 /// An alias set tracker to partition the access set by underlying object and
591 //intrinsic property (such as TBAA metadata).
596 /// Sets of potentially dependent accesses - members of one set share an
597 /// underlying pointer. The set "CheckDeps" identfies which sets really need a
598 /// dependence check.
599 MemoryDepChecker::DepCandidates &DepCands;
601 /// Initial processing of memory accesses determined that we may need
602 /// to add memchecks. Perform the analysis to determine the necessary checks.
604 /// Note that, this is different from isDependencyCheckNeeded. When we retry
605 /// memcheck analysis without dependency checking
606 /// (i.e. FoundNonConstantDistanceDependence), isDependencyCheckNeeded is
607 /// cleared while this remains set if we have potentially dependent accesses.
608 bool IsRTCheckAnalysisNeeded;
610 /// The SCEV predicate containing all the SCEV-related assumptions.
611 PredicatedScalarEvolution &PSE;
614 } // end anonymous namespace
616 /// Check whether a pointer can participate in a runtime bounds check.
617 /// If \p Assume, try harder to prove that we can compute the bounds of \p Ptr
618 /// by adding run-time checks (overflow checks) if necessary.
619 static bool hasComputableBounds(PredicatedScalarEvolution &PSE,
620 const ValueToValueMap &Strides, Value *Ptr,
621 Loop *L, bool Assume) {
622 const SCEV *PtrScev = replaceSymbolicStrideSCEV(PSE, Strides, Ptr);
624 // The bounds for loop-invariant pointer is trivial.
625 if (PSE.getSE()->isLoopInvariant(PtrScev, L))
628 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
631 AR = PSE.getAsAddRec(Ptr);
636 return AR->isAffine();
639 /// Check whether a pointer address cannot wrap.
640 static bool isNoWrap(PredicatedScalarEvolution &PSE,
641 const ValueToValueMap &Strides, Value *Ptr, Loop *L) {
642 const SCEV *PtrScev = PSE.getSCEV(Ptr);
643 if (PSE.getSE()->isLoopInvariant(PtrScev, L))
646 int64_t Stride = getPtrStride(PSE, Ptr, L, Strides);
647 if (Stride == 1 || PSE.hasNoOverflow(Ptr, SCEVWrapPredicate::IncrementNUSW))
653 bool AccessAnalysis::createCheckForAccess(RuntimePointerChecking &RtCheck,
654 MemAccessInfo Access,
655 const ValueToValueMap &StridesMap,
656 DenseMap<Value *, unsigned> &DepSetId,
657 Loop *TheLoop, unsigned &RunningDepId,
658 unsigned ASId, bool ShouldCheckWrap,
660 Value *Ptr = Access.getPointer();
662 if (!hasComputableBounds(PSE, StridesMap, Ptr, TheLoop, Assume))
665 // When we run after a failing dependency check we have to make sure
666 // we don't have wrapping pointers.
667 if (ShouldCheckWrap && !isNoWrap(PSE, StridesMap, Ptr, TheLoop)) {
668 auto *Expr = PSE.getSCEV(Ptr);
669 if (!Assume || !isa<SCEVAddRecExpr>(Expr))
671 PSE.setNoOverflow(Ptr, SCEVWrapPredicate::IncrementNUSW);
674 // The id of the dependence set.
677 if (isDependencyCheckNeeded()) {
678 Value *Leader = DepCands.getLeaderValue(Access).getPointer();
679 unsigned &LeaderId = DepSetId[Leader];
681 LeaderId = RunningDepId++;
684 // Each access has its own dependence set.
685 DepId = RunningDepId++;
687 bool IsWrite = Access.getInt();
688 RtCheck.insert(TheLoop, Ptr, IsWrite, DepId, ASId, StridesMap, PSE);
689 LLVM_DEBUG(dbgs() << "LAA: Found a runtime check ptr:" << *Ptr << '\n');
694 bool AccessAnalysis::canCheckPtrAtRT(RuntimePointerChecking &RtCheck,
695 ScalarEvolution *SE, Loop *TheLoop,
696 const ValueToValueMap &StridesMap,
697 bool ShouldCheckWrap) {
698 // Find pointers with computable bounds. We are going to use this information
699 // to place a runtime bound check.
702 bool NeedRTCheck = false;
703 if (!IsRTCheckAnalysisNeeded) return true;
705 bool IsDepCheckNeeded = isDependencyCheckNeeded();
707 // We assign a consecutive id to access from different alias sets.
708 // Accesses between different groups doesn't need to be checked.
710 for (auto &AS : AST) {
711 int NumReadPtrChecks = 0;
712 int NumWritePtrChecks = 0;
713 bool CanDoAliasSetRT = true;
715 // We assign consecutive id to access from different dependence sets.
716 // Accesses within the same set don't need a runtime check.
717 unsigned RunningDepId = 1;
718 DenseMap<Value *, unsigned> DepSetId;
720 SmallVector<MemAccessInfo, 4> Retries;
723 Value *Ptr = A.getValue();
724 bool IsWrite = Accesses.count(MemAccessInfo(Ptr, true));
725 MemAccessInfo Access(Ptr, IsWrite);
732 if (!createCheckForAccess(RtCheck, Access, StridesMap, DepSetId, TheLoop,
733 RunningDepId, ASId, ShouldCheckWrap, false)) {
734 LLVM_DEBUG(dbgs() << "LAA: Can't find bounds for ptr:" << *Ptr << '\n');
735 Retries.push_back(Access);
736 CanDoAliasSetRT = false;
740 // If we have at least two writes or one write and a read then we need to
741 // check them. But there is no need to checks if there is only one
742 // dependence set for this alias set.
744 // Note that this function computes CanDoRT and NeedRTCheck independently.
745 // For example CanDoRT=false, NeedRTCheck=false means that we have a pointer
746 // for which we couldn't find the bounds but we don't actually need to emit
747 // any checks so it does not matter.
748 bool NeedsAliasSetRTCheck = false;
749 if (!(IsDepCheckNeeded && CanDoAliasSetRT && RunningDepId == 2))
750 NeedsAliasSetRTCheck = (NumWritePtrChecks >= 2 ||
751 (NumReadPtrChecks >= 1 && NumWritePtrChecks >= 1));
753 // We need to perform run-time alias checks, but some pointers had bounds
754 // that couldn't be checked.
755 if (NeedsAliasSetRTCheck && !CanDoAliasSetRT) {
756 // Reset the CanDoSetRt flag and retry all accesses that have failed.
757 // We know that we need these checks, so we can now be more aggressive
758 // and add further checks if required (overflow checks).
759 CanDoAliasSetRT = true;
760 for (auto Access : Retries)
761 if (!createCheckForAccess(RtCheck, Access, StridesMap, DepSetId,
762 TheLoop, RunningDepId, ASId,
763 ShouldCheckWrap, /*Assume=*/true)) {
764 CanDoAliasSetRT = false;
769 CanDoRT &= CanDoAliasSetRT;
770 NeedRTCheck |= NeedsAliasSetRTCheck;
774 // If the pointers that we would use for the bounds comparison have different
775 // address spaces, assume the values aren't directly comparable, so we can't
776 // use them for the runtime check. We also have to assume they could
777 // overlap. In the future there should be metadata for whether address spaces
779 unsigned NumPointers = RtCheck.Pointers.size();
780 for (unsigned i = 0; i < NumPointers; ++i) {
781 for (unsigned j = i + 1; j < NumPointers; ++j) {
782 // Only need to check pointers between two different dependency sets.
783 if (RtCheck.Pointers[i].DependencySetId ==
784 RtCheck.Pointers[j].DependencySetId)
786 // Only need to check pointers in the same alias set.
787 if (RtCheck.Pointers[i].AliasSetId != RtCheck.Pointers[j].AliasSetId)
790 Value *PtrI = RtCheck.Pointers[i].PointerValue;
791 Value *PtrJ = RtCheck.Pointers[j].PointerValue;
793 unsigned ASi = PtrI->getType()->getPointerAddressSpace();
794 unsigned ASj = PtrJ->getType()->getPointerAddressSpace();
797 dbgs() << "LAA: Runtime check would require comparison between"
798 " different address spaces\n");
804 if (NeedRTCheck && CanDoRT)
805 RtCheck.generateChecks(DepCands, IsDepCheckNeeded);
807 LLVM_DEBUG(dbgs() << "LAA: We need to do " << RtCheck.getNumberOfChecks()
808 << " pointer comparisons.\n");
810 RtCheck.Need = NeedRTCheck;
812 bool CanDoRTIfNeeded = !NeedRTCheck || CanDoRT;
813 if (!CanDoRTIfNeeded)
815 return CanDoRTIfNeeded;
818 void AccessAnalysis::processMemAccesses() {
819 // We process the set twice: first we process read-write pointers, last we
820 // process read-only pointers. This allows us to skip dependence tests for
821 // read-only pointers.
823 LLVM_DEBUG(dbgs() << "LAA: Processing memory accesses...\n");
824 LLVM_DEBUG(dbgs() << " AST: "; AST.dump());
825 LLVM_DEBUG(dbgs() << "LAA: Accesses(" << Accesses.size() << "):\n");
827 for (auto A : Accesses)
828 dbgs() << "\t" << *A.getPointer() << " (" <<
829 (A.getInt() ? "write" : (ReadOnlyPtr.count(A.getPointer()) ?
830 "read-only" : "read")) << ")\n";
833 // The AliasSetTracker has nicely partitioned our pointers by metadata
834 // compatibility and potential for underlying-object overlap. As a result, we
835 // only need to check for potential pointer dependencies within each alias
837 for (auto &AS : AST) {
838 // Note that both the alias-set tracker and the alias sets themselves used
839 // linked lists internally and so the iteration order here is deterministic
840 // (matching the original instruction order within each set).
842 bool SetHasWrite = false;
844 // Map of pointers to last access encountered.
845 typedef DenseMap<const Value*, MemAccessInfo> UnderlyingObjToAccessMap;
846 UnderlyingObjToAccessMap ObjToLastAccess;
848 // Set of access to check after all writes have been processed.
849 PtrAccessSet DeferredAccesses;
851 // Iterate over each alias set twice, once to process read/write pointers,
852 // and then to process read-only pointers.
853 for (int SetIteration = 0; SetIteration < 2; ++SetIteration) {
854 bool UseDeferred = SetIteration > 0;
855 PtrAccessSet &S = UseDeferred ? DeferredAccesses : Accesses;
858 Value *Ptr = AV.getValue();
860 // For a single memory access in AliasSetTracker, Accesses may contain
861 // both read and write, and they both need to be handled for CheckDeps.
863 if (AC.getPointer() != Ptr)
866 bool IsWrite = AC.getInt();
868 // If we're using the deferred access set, then it contains only
870 bool IsReadOnlyPtr = ReadOnlyPtr.count(Ptr) && !IsWrite;
871 if (UseDeferred && !IsReadOnlyPtr)
873 // Otherwise, the pointer must be in the PtrAccessSet, either as a
875 assert(((IsReadOnlyPtr && UseDeferred) || IsWrite ||
876 S.count(MemAccessInfo(Ptr, false))) &&
877 "Alias-set pointer not in the access set?");
879 MemAccessInfo Access(Ptr, IsWrite);
880 DepCands.insert(Access);
882 // Memorize read-only pointers for later processing and skip them in
883 // the first round (they need to be checked after we have seen all
884 // write pointers). Note: we also mark pointer that are not
885 // consecutive as "read-only" pointers (so that we check
886 // "a[b[i]] +="). Hence, we need the second check for "!IsWrite".
887 if (!UseDeferred && IsReadOnlyPtr) {
888 DeferredAccesses.insert(Access);
892 // If this is a write - check other reads and writes for conflicts. If
893 // this is a read only check other writes for conflicts (but only if
894 // there is no other write to the ptr - this is an optimization to
895 // catch "a[i] = a[i] + " without having to do a dependence check).
896 if ((IsWrite || IsReadOnlyPtr) && SetHasWrite) {
897 CheckDeps.push_back(Access);
898 IsRTCheckAnalysisNeeded = true;
904 // Create sets of pointers connected by a shared alias set and
905 // underlying object.
906 typedef SmallVector<const Value *, 16> ValueVector;
907 ValueVector TempObjects;
909 GetUnderlyingObjects(Ptr, TempObjects, DL, LI);
911 << "Underlying objects for pointer " << *Ptr << "\n");
912 for (const Value *UnderlyingObj : TempObjects) {
913 // nullptr never alias, don't join sets for pointer that have "null"
914 // in their UnderlyingObjects list.
915 if (isa<ConstantPointerNull>(UnderlyingObj) &&
916 !NullPointerIsDefined(
917 TheLoop->getHeader()->getParent(),
918 UnderlyingObj->getType()->getPointerAddressSpace()))
921 UnderlyingObjToAccessMap::iterator Prev =
922 ObjToLastAccess.find(UnderlyingObj);
923 if (Prev != ObjToLastAccess.end())
924 DepCands.unionSets(Access, Prev->second);
926 ObjToLastAccess[UnderlyingObj] = Access;
927 LLVM_DEBUG(dbgs() << " " << *UnderlyingObj << "\n");
935 static bool isInBoundsGep(Value *Ptr) {
936 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr))
937 return GEP->isInBounds();
941 /// Return true if an AddRec pointer \p Ptr is unsigned non-wrapping,
942 /// i.e. monotonically increasing/decreasing.
943 static bool isNoWrapAddRec(Value *Ptr, const SCEVAddRecExpr *AR,
944 PredicatedScalarEvolution &PSE, const Loop *L) {
945 // FIXME: This should probably only return true for NUW.
946 if (AR->getNoWrapFlags(SCEV::NoWrapMask))
949 // Scalar evolution does not propagate the non-wrapping flags to values that
950 // are derived from a non-wrapping induction variable because non-wrapping
951 // could be flow-sensitive.
953 // Look through the potentially overflowing instruction to try to prove
954 // non-wrapping for the *specific* value of Ptr.
956 // The arithmetic implied by an inbounds GEP can't overflow.
957 auto *GEP = dyn_cast<GetElementPtrInst>(Ptr);
958 if (!GEP || !GEP->isInBounds())
961 // Make sure there is only one non-const index and analyze that.
962 Value *NonConstIndex = nullptr;
963 for (Value *Index : make_range(GEP->idx_begin(), GEP->idx_end()))
964 if (!isa<ConstantInt>(Index)) {
967 NonConstIndex = Index;
970 // The recurrence is on the pointer, ignore for now.
973 // The index in GEP is signed. It is non-wrapping if it's derived from a NSW
974 // AddRec using a NSW operation.
975 if (auto *OBO = dyn_cast<OverflowingBinaryOperator>(NonConstIndex))
976 if (OBO->hasNoSignedWrap() &&
977 // Assume constant for other the operand so that the AddRec can be
979 isa<ConstantInt>(OBO->getOperand(1))) {
980 auto *OpScev = PSE.getSCEV(OBO->getOperand(0));
982 if (auto *OpAR = dyn_cast<SCEVAddRecExpr>(OpScev))
983 return OpAR->getLoop() == L && OpAR->getNoWrapFlags(SCEV::FlagNSW);
989 /// Check whether the access through \p Ptr has a constant stride.
990 int64_t llvm::getPtrStride(PredicatedScalarEvolution &PSE, Value *Ptr,
991 const Loop *Lp, const ValueToValueMap &StridesMap,
992 bool Assume, bool ShouldCheckWrap) {
993 Type *Ty = Ptr->getType();
994 assert(Ty->isPointerTy() && "Unexpected non-ptr");
996 // Make sure that the pointer does not point to aggregate types.
997 auto *PtrTy = cast<PointerType>(Ty);
998 if (PtrTy->getElementType()->isAggregateType()) {
999 LLVM_DEBUG(dbgs() << "LAA: Bad stride - Not a pointer to a scalar type"
1004 const SCEV *PtrScev = replaceSymbolicStrideSCEV(PSE, StridesMap, Ptr);
1006 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
1008 AR = PSE.getAsAddRec(Ptr);
1011 LLVM_DEBUG(dbgs() << "LAA: Bad stride - Not an AddRecExpr pointer " << *Ptr
1012 << " SCEV: " << *PtrScev << "\n");
1016 // The access function must stride over the innermost loop.
1017 if (Lp != AR->getLoop()) {
1018 LLVM_DEBUG(dbgs() << "LAA: Bad stride - Not striding over innermost loop "
1019 << *Ptr << " SCEV: " << *AR << "\n");
1023 // The address calculation must not wrap. Otherwise, a dependence could be
1025 // An inbounds getelementptr that is a AddRec with a unit stride
1026 // cannot wrap per definition. The unit stride requirement is checked later.
1027 // An getelementptr without an inbounds attribute and unit stride would have
1028 // to access the pointer value "0" which is undefined behavior in address
1029 // space 0, therefore we can also vectorize this case.
1030 bool IsInBoundsGEP = isInBoundsGep(Ptr);
1031 bool IsNoWrapAddRec = !ShouldCheckWrap ||
1032 PSE.hasNoOverflow(Ptr, SCEVWrapPredicate::IncrementNUSW) ||
1033 isNoWrapAddRec(Ptr, AR, PSE, Lp);
1034 if (!IsNoWrapAddRec && !IsInBoundsGEP &&
1035 NullPointerIsDefined(Lp->getHeader()->getParent(),
1036 PtrTy->getAddressSpace())) {
1038 PSE.setNoOverflow(Ptr, SCEVWrapPredicate::IncrementNUSW);
1039 IsNoWrapAddRec = true;
1040 LLVM_DEBUG(dbgs() << "LAA: Pointer may wrap in the address space:\n"
1041 << "LAA: Pointer: " << *Ptr << "\n"
1042 << "LAA: SCEV: " << *AR << "\n"
1043 << "LAA: Added an overflow assumption\n");
1046 dbgs() << "LAA: Bad stride - Pointer may wrap in the address space "
1047 << *Ptr << " SCEV: " << *AR << "\n");
1052 // Check the step is constant.
1053 const SCEV *Step = AR->getStepRecurrence(*PSE.getSE());
1055 // Calculate the pointer stride and check if it is constant.
1056 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
1058 LLVM_DEBUG(dbgs() << "LAA: Bad stride - Not a constant strided " << *Ptr
1059 << " SCEV: " << *AR << "\n");
1063 auto &DL = Lp->getHeader()->getModule()->getDataLayout();
1064 int64_t Size = DL.getTypeAllocSize(PtrTy->getElementType());
1065 const APInt &APStepVal = C->getAPInt();
1067 // Huge step value - give up.
1068 if (APStepVal.getBitWidth() > 64)
1071 int64_t StepVal = APStepVal.getSExtValue();
1074 int64_t Stride = StepVal / Size;
1075 int64_t Rem = StepVal % Size;
1079 // If the SCEV could wrap but we have an inbounds gep with a unit stride we
1080 // know we can't "wrap around the address space". In case of address space
1081 // zero we know that this won't happen without triggering undefined behavior.
1082 if (!IsNoWrapAddRec && Stride != 1 && Stride != -1 &&
1083 (IsInBoundsGEP || !NullPointerIsDefined(Lp->getHeader()->getParent(),
1084 PtrTy->getAddressSpace()))) {
1086 // We can avoid this case by adding a run-time check.
1087 LLVM_DEBUG(dbgs() << "LAA: Non unit strided pointer which is not either "
1088 << "inbounds or in address space 0 may wrap:\n"
1089 << "LAA: Pointer: " << *Ptr << "\n"
1090 << "LAA: SCEV: " << *AR << "\n"
1091 << "LAA: Added an overflow assumption\n");
1092 PSE.setNoOverflow(Ptr, SCEVWrapPredicate::IncrementNUSW);
1100 bool llvm::sortPtrAccesses(ArrayRef<Value *> VL, const DataLayout &DL,
1101 ScalarEvolution &SE,
1102 SmallVectorImpl<unsigned> &SortedIndices) {
1103 assert(llvm::all_of(
1104 VL, [](const Value *V) { return V->getType()->isPointerTy(); }) &&
1105 "Expected list of pointer operands.");
1106 SmallVector<std::pair<int64_t, Value *>, 4> OffValPairs;
1107 OffValPairs.reserve(VL.size());
1109 // Walk over the pointers, and map each of them to an offset relative to
1110 // first pointer in the array.
1111 Value *Ptr0 = VL[0];
1112 const SCEV *Scev0 = SE.getSCEV(Ptr0);
1113 Value *Obj0 = GetUnderlyingObject(Ptr0, DL);
1115 llvm::SmallSet<int64_t, 4> Offsets;
1116 for (auto *Ptr : VL) {
1117 // TODO: Outline this code as a special, more time consuming, version of
1118 // computeConstantDifference() function.
1119 if (Ptr->getType()->getPointerAddressSpace() !=
1120 Ptr0->getType()->getPointerAddressSpace())
1122 // If a pointer refers to a different underlying object, bail - the
1123 // pointers are by definition incomparable.
1124 Value *CurrObj = GetUnderlyingObject(Ptr, DL);
1125 if (CurrObj != Obj0)
1128 const SCEV *Scev = SE.getSCEV(Ptr);
1129 const auto *Diff = dyn_cast<SCEVConstant>(SE.getMinusSCEV(Scev, Scev0));
1130 // The pointers may not have a constant offset from each other, or SCEV
1131 // may just not be smart enough to figure out they do. Regardless,
1132 // there's nothing we can do.
1136 // Check if the pointer with the same offset is found.
1137 int64_t Offset = Diff->getAPInt().getSExtValue();
1138 if (!Offsets.insert(Offset).second)
1140 OffValPairs.emplace_back(Offset, Ptr);
1142 SortedIndices.clear();
1143 SortedIndices.resize(VL.size());
1144 std::iota(SortedIndices.begin(), SortedIndices.end(), 0);
1146 // Sort the memory accesses and keep the order of their uses in UseOrder.
1147 llvm::stable_sort(SortedIndices, [&](unsigned Left, unsigned Right) {
1148 return OffValPairs[Left].first < OffValPairs[Right].first;
1151 // Check if the order is consecutive already.
1152 if (llvm::all_of(SortedIndices, [&SortedIndices](const unsigned I) {
1153 return I == SortedIndices[I];
1155 SortedIndices.clear();
1160 /// Take the address space operand from the Load/Store instruction.
1161 /// Returns -1 if this is not a valid Load/Store instruction.
1162 static unsigned getAddressSpaceOperand(Value *I) {
1163 if (LoadInst *L = dyn_cast<LoadInst>(I))
1164 return L->getPointerAddressSpace();
1165 if (StoreInst *S = dyn_cast<StoreInst>(I))
1166 return S->getPointerAddressSpace();
1170 /// Returns true if the memory operations \p A and \p B are consecutive.
1171 bool llvm::isConsecutiveAccess(Value *A, Value *B, const DataLayout &DL,
1172 ScalarEvolution &SE, bool CheckType) {
1173 Value *PtrA = getLoadStorePointerOperand(A);
1174 Value *PtrB = getLoadStorePointerOperand(B);
1175 unsigned ASA = getAddressSpaceOperand(A);
1176 unsigned ASB = getAddressSpaceOperand(B);
1178 // Check that the address spaces match and that the pointers are valid.
1179 if (!PtrA || !PtrB || (ASA != ASB))
1182 // Make sure that A and B are different pointers.
1186 // Make sure that A and B have the same type if required.
1187 if (CheckType && PtrA->getType() != PtrB->getType())
1190 unsigned IdxWidth = DL.getIndexSizeInBits(ASA);
1191 Type *Ty = cast<PointerType>(PtrA->getType())->getElementType();
1192 APInt Size(IdxWidth, DL.getTypeStoreSize(Ty));
1194 APInt OffsetA(IdxWidth, 0), OffsetB(IdxWidth, 0);
1195 PtrA = PtrA->stripAndAccumulateInBoundsConstantOffsets(DL, OffsetA);
1196 PtrB = PtrB->stripAndAccumulateInBoundsConstantOffsets(DL, OffsetB);
1198 // OffsetDelta = OffsetB - OffsetA;
1199 const SCEV *OffsetSCEVA = SE.getConstant(OffsetA);
1200 const SCEV *OffsetSCEVB = SE.getConstant(OffsetB);
1201 const SCEV *OffsetDeltaSCEV = SE.getMinusSCEV(OffsetSCEVB, OffsetSCEVA);
1202 const SCEVConstant *OffsetDeltaC = dyn_cast<SCEVConstant>(OffsetDeltaSCEV);
1203 const APInt &OffsetDelta = OffsetDeltaC->getAPInt();
1204 // Check if they are based on the same pointer. That makes the offsets
1207 return OffsetDelta == Size;
1209 // Compute the necessary base pointer delta to have the necessary final delta
1210 // equal to the size.
1211 // BaseDelta = Size - OffsetDelta;
1212 const SCEV *SizeSCEV = SE.getConstant(Size);
1213 const SCEV *BaseDelta = SE.getMinusSCEV(SizeSCEV, OffsetDeltaSCEV);
1215 // Otherwise compute the distance with SCEV between the base pointers.
1216 const SCEV *PtrSCEVA = SE.getSCEV(PtrA);
1217 const SCEV *PtrSCEVB = SE.getSCEV(PtrB);
1218 const SCEV *X = SE.getAddExpr(PtrSCEVA, BaseDelta);
1219 return X == PtrSCEVB;
1222 MemoryDepChecker::VectorizationSafetyStatus
1223 MemoryDepChecker::Dependence::isSafeForVectorization(DepType Type) {
1227 case BackwardVectorizable:
1228 return VectorizationSafetyStatus::Safe;
1231 return VectorizationSafetyStatus::PossiblySafeWithRtChecks;
1232 case ForwardButPreventsForwarding:
1234 case BackwardVectorizableButPreventsForwarding:
1235 return VectorizationSafetyStatus::Unsafe;
1237 llvm_unreachable("unexpected DepType!");
1240 bool MemoryDepChecker::Dependence::isBackward() const {
1244 case ForwardButPreventsForwarding:
1248 case BackwardVectorizable:
1250 case BackwardVectorizableButPreventsForwarding:
1253 llvm_unreachable("unexpected DepType!");
1256 bool MemoryDepChecker::Dependence::isPossiblyBackward() const {
1257 return isBackward() || Type == Unknown;
1260 bool MemoryDepChecker::Dependence::isForward() const {
1263 case ForwardButPreventsForwarding:
1268 case BackwardVectorizable:
1270 case BackwardVectorizableButPreventsForwarding:
1273 llvm_unreachable("unexpected DepType!");
1276 bool MemoryDepChecker::couldPreventStoreLoadForward(uint64_t Distance,
1277 uint64_t TypeByteSize) {
1278 // If loads occur at a distance that is not a multiple of a feasible vector
1279 // factor store-load forwarding does not take place.
1280 // Positive dependences might cause troubles because vectorizing them might
1281 // prevent store-load forwarding making vectorized code run a lot slower.
1282 // a[i] = a[i-3] ^ a[i-8];
1283 // The stores to a[i:i+1] don't align with the stores to a[i-3:i-2] and
1284 // hence on your typical architecture store-load forwarding does not take
1285 // place. Vectorizing in such cases does not make sense.
1286 // Store-load forwarding distance.
1288 // After this many iterations store-to-load forwarding conflicts should not
1289 // cause any slowdowns.
1290 const uint64_t NumItersForStoreLoadThroughMemory = 8 * TypeByteSize;
1291 // Maximum vector factor.
1292 uint64_t MaxVFWithoutSLForwardIssues = std::min(
1293 VectorizerParams::MaxVectorWidth * TypeByteSize, MaxSafeDepDistBytes);
1295 // Compute the smallest VF at which the store and load would be misaligned.
1296 for (uint64_t VF = 2 * TypeByteSize; VF <= MaxVFWithoutSLForwardIssues;
1298 // If the number of vector iteration between the store and the load are
1299 // small we could incur conflicts.
1300 if (Distance % VF && Distance / VF < NumItersForStoreLoadThroughMemory) {
1301 MaxVFWithoutSLForwardIssues = (VF >>= 1);
1306 if (MaxVFWithoutSLForwardIssues < 2 * TypeByteSize) {
1308 dbgs() << "LAA: Distance " << Distance
1309 << " that could cause a store-load forwarding conflict\n");
1313 if (MaxVFWithoutSLForwardIssues < MaxSafeDepDistBytes &&
1314 MaxVFWithoutSLForwardIssues !=
1315 VectorizerParams::MaxVectorWidth * TypeByteSize)
1316 MaxSafeDepDistBytes = MaxVFWithoutSLForwardIssues;
1320 void MemoryDepChecker::mergeInStatus(VectorizationSafetyStatus S) {
1325 /// Given a non-constant (unknown) dependence-distance \p Dist between two
1326 /// memory accesses, that have the same stride whose absolute value is given
1327 /// in \p Stride, and that have the same type size \p TypeByteSize,
1328 /// in a loop whose takenCount is \p BackedgeTakenCount, check if it is
1329 /// possible to prove statically that the dependence distance is larger
1330 /// than the range that the accesses will travel through the execution of
1331 /// the loop. If so, return true; false otherwise. This is useful for
1332 /// example in loops such as the following (PR31098):
1333 /// for (i = 0; i < D; ++i) {
1337 static bool isSafeDependenceDistance(const DataLayout &DL, ScalarEvolution &SE,
1338 const SCEV &BackedgeTakenCount,
1339 const SCEV &Dist, uint64_t Stride,
1340 uint64_t TypeByteSize) {
1342 // If we can prove that
1343 // (**) |Dist| > BackedgeTakenCount * Step
1344 // where Step is the absolute stride of the memory accesses in bytes,
1345 // then there is no dependence.
1348 // We basically want to check if the absolute distance (|Dist/Step|)
1349 // is >= the loop iteration count (or > BackedgeTakenCount).
1350 // This is equivalent to the Strong SIV Test (Practical Dependence Testing,
1351 // Section 4.2.1); Note, that for vectorization it is sufficient to prove
1352 // that the dependence distance is >= VF; This is checked elsewhere.
1353 // But in some cases we can prune unknown dependence distances early, and
1354 // even before selecting the VF, and without a runtime test, by comparing
1355 // the distance against the loop iteration count. Since the vectorized code
1356 // will be executed only if LoopCount >= VF, proving distance >= LoopCount
1357 // also guarantees that distance >= VF.
1359 const uint64_t ByteStride = Stride * TypeByteSize;
1360 const SCEV *Step = SE.getConstant(BackedgeTakenCount.getType(), ByteStride);
1361 const SCEV *Product = SE.getMulExpr(&BackedgeTakenCount, Step);
1363 const SCEV *CastedDist = &Dist;
1364 const SCEV *CastedProduct = Product;
1365 uint64_t DistTypeSize = DL.getTypeAllocSize(Dist.getType());
1366 uint64_t ProductTypeSize = DL.getTypeAllocSize(Product->getType());
1368 // The dependence distance can be positive/negative, so we sign extend Dist;
1369 // The multiplication of the absolute stride in bytes and the
1370 // backedgeTakenCount is non-negative, so we zero extend Product.
1371 if (DistTypeSize > ProductTypeSize)
1372 CastedProduct = SE.getZeroExtendExpr(Product, Dist.getType());
1374 CastedDist = SE.getNoopOrSignExtend(&Dist, Product->getType());
1376 // Is Dist - (BackedgeTakenCount * Step) > 0 ?
1377 // (If so, then we have proven (**) because |Dist| >= Dist)
1378 const SCEV *Minus = SE.getMinusSCEV(CastedDist, CastedProduct);
1379 if (SE.isKnownPositive(Minus))
1382 // Second try: Is -Dist - (BackedgeTakenCount * Step) > 0 ?
1383 // (If so, then we have proven (**) because |Dist| >= -1*Dist)
1384 const SCEV *NegDist = SE.getNegativeSCEV(CastedDist);
1385 Minus = SE.getMinusSCEV(NegDist, CastedProduct);
1386 if (SE.isKnownPositive(Minus))
1392 /// Check the dependence for two accesses with the same stride \p Stride.
1393 /// \p Distance is the positive distance and \p TypeByteSize is type size in
1396 /// \returns true if they are independent.
1397 static bool areStridedAccessesIndependent(uint64_t Distance, uint64_t Stride,
1398 uint64_t TypeByteSize) {
1399 assert(Stride > 1 && "The stride must be greater than 1");
1400 assert(TypeByteSize > 0 && "The type size in byte must be non-zero");
1401 assert(Distance > 0 && "The distance must be non-zero");
1403 // Skip if the distance is not multiple of type byte size.
1404 if (Distance % TypeByteSize)
1407 uint64_t ScaledDist = Distance / TypeByteSize;
1409 // No dependence if the scaled distance is not multiple of the stride.
1411 // for (i = 0; i < 1024 ; i += 4)
1412 // A[i+2] = A[i] + 1;
1414 // Two accesses in memory (scaled distance is 2, stride is 4):
1415 // | A[0] | | | | A[4] | | | |
1416 // | | | A[2] | | | | A[6] | |
1419 // for (i = 0; i < 1024 ; i += 3)
1420 // A[i+4] = A[i] + 1;
1422 // Two accesses in memory (scaled distance is 4, stride is 3):
1423 // | A[0] | | | A[3] | | | A[6] | | |
1424 // | | | | | A[4] | | | A[7] | |
1425 return ScaledDist % Stride;
1428 MemoryDepChecker::Dependence::DepType
1429 MemoryDepChecker::isDependent(const MemAccessInfo &A, unsigned AIdx,
1430 const MemAccessInfo &B, unsigned BIdx,
1431 const ValueToValueMap &Strides) {
1432 assert (AIdx < BIdx && "Must pass arguments in program order");
1434 Value *APtr = A.getPointer();
1435 Value *BPtr = B.getPointer();
1436 bool AIsWrite = A.getInt();
1437 bool BIsWrite = B.getInt();
1439 // Two reads are independent.
1440 if (!AIsWrite && !BIsWrite)
1441 return Dependence::NoDep;
1443 // We cannot check pointers in different address spaces.
1444 if (APtr->getType()->getPointerAddressSpace() !=
1445 BPtr->getType()->getPointerAddressSpace())
1446 return Dependence::Unknown;
1448 int64_t StrideAPtr = getPtrStride(PSE, APtr, InnermostLoop, Strides, true);
1449 int64_t StrideBPtr = getPtrStride(PSE, BPtr, InnermostLoop, Strides, true);
1451 const SCEV *Src = PSE.getSCEV(APtr);
1452 const SCEV *Sink = PSE.getSCEV(BPtr);
1454 // If the induction step is negative we have to invert source and sink of the
1456 if (StrideAPtr < 0) {
1457 std::swap(APtr, BPtr);
1458 std::swap(Src, Sink);
1459 std::swap(AIsWrite, BIsWrite);
1460 std::swap(AIdx, BIdx);
1461 std::swap(StrideAPtr, StrideBPtr);
1464 const SCEV *Dist = PSE.getSE()->getMinusSCEV(Sink, Src);
1466 LLVM_DEBUG(dbgs() << "LAA: Src Scev: " << *Src << "Sink Scev: " << *Sink
1467 << "(Induction step: " << StrideAPtr << ")\n");
1468 LLVM_DEBUG(dbgs() << "LAA: Distance for " << *InstMap[AIdx] << " to "
1469 << *InstMap[BIdx] << ": " << *Dist << "\n");
1471 // Need accesses with constant stride. We don't want to vectorize
1472 // "A[B[i]] += ..." and similar code or pointer arithmetic that could wrap in
1473 // the address space.
1474 if (!StrideAPtr || !StrideBPtr || StrideAPtr != StrideBPtr){
1475 LLVM_DEBUG(dbgs() << "Pointer access with non-constant stride\n");
1476 return Dependence::Unknown;
1479 Type *ATy = APtr->getType()->getPointerElementType();
1480 Type *BTy = BPtr->getType()->getPointerElementType();
1481 auto &DL = InnermostLoop->getHeader()->getModule()->getDataLayout();
1482 uint64_t TypeByteSize = DL.getTypeAllocSize(ATy);
1483 uint64_t Stride = std::abs(StrideAPtr);
1484 const SCEVConstant *C = dyn_cast<SCEVConstant>(Dist);
1486 if (TypeByteSize == DL.getTypeAllocSize(BTy) &&
1487 isSafeDependenceDistance(DL, *(PSE.getSE()),
1488 *(PSE.getBackedgeTakenCount()), *Dist, Stride,
1490 return Dependence::NoDep;
1492 LLVM_DEBUG(dbgs() << "LAA: Dependence because of non-constant distance\n");
1493 FoundNonConstantDistanceDependence = true;
1494 return Dependence::Unknown;
1497 const APInt &Val = C->getAPInt();
1498 int64_t Distance = Val.getSExtValue();
1500 // Attempt to prove strided accesses independent.
1501 if (std::abs(Distance) > 0 && Stride > 1 && ATy == BTy &&
1502 areStridedAccessesIndependent(std::abs(Distance), Stride, TypeByteSize)) {
1503 LLVM_DEBUG(dbgs() << "LAA: Strided accesses are independent\n");
1504 return Dependence::NoDep;
1507 // Negative distances are not plausible dependencies.
1508 if (Val.isNegative()) {
1509 bool IsTrueDataDependence = (AIsWrite && !BIsWrite);
1510 if (IsTrueDataDependence && EnableForwardingConflictDetection &&
1511 (couldPreventStoreLoadForward(Val.abs().getZExtValue(), TypeByteSize) ||
1513 LLVM_DEBUG(dbgs() << "LAA: Forward but may prevent st->ld forwarding\n");
1514 return Dependence::ForwardButPreventsForwarding;
1517 LLVM_DEBUG(dbgs() << "LAA: Dependence is negative\n");
1518 return Dependence::Forward;
1521 // Write to the same location with the same size.
1522 // Could be improved to assert type sizes are the same (i32 == float, etc).
1525 return Dependence::Forward;
1527 dbgs() << "LAA: Zero dependence difference but different types\n");
1528 return Dependence::Unknown;
1531 assert(Val.isStrictlyPositive() && "Expect a positive value");
1536 << "LAA: ReadWrite-Write positive dependency with different types\n");
1537 return Dependence::Unknown;
1540 // Bail out early if passed-in parameters make vectorization not feasible.
1541 unsigned ForcedFactor = (VectorizerParams::VectorizationFactor ?
1542 VectorizerParams::VectorizationFactor : 1);
1543 unsigned ForcedUnroll = (VectorizerParams::VectorizationInterleave ?
1544 VectorizerParams::VectorizationInterleave : 1);
1545 // The minimum number of iterations for a vectorized/unrolled version.
1546 unsigned MinNumIter = std::max(ForcedFactor * ForcedUnroll, 2U);
1548 // It's not vectorizable if the distance is smaller than the minimum distance
1549 // needed for a vectroized/unrolled version. Vectorizing one iteration in
1550 // front needs TypeByteSize * Stride. Vectorizing the last iteration needs
1551 // TypeByteSize (No need to plus the last gap distance).
1553 // E.g. Assume one char is 1 byte in memory and one int is 4 bytes.
1555 // int *B = (int *)((char *)A + 14);
1556 // for (i = 0 ; i < 1024 ; i += 2)
1560 // Two accesses in memory (stride is 2):
1561 // | A[0] | | A[2] | | A[4] | | A[6] | |
1562 // | B[0] | | B[2] | | B[4] |
1564 // Distance needs for vectorizing iterations except the last iteration:
1565 // 4 * 2 * (MinNumIter - 1). Distance needs for the last iteration: 4.
1566 // So the minimum distance needed is: 4 * 2 * (MinNumIter - 1) + 4.
1568 // If MinNumIter is 2, it is vectorizable as the minimum distance needed is
1569 // 12, which is less than distance.
1571 // If MinNumIter is 4 (Say if a user forces the vectorization factor to be 4),
1572 // the minimum distance needed is 28, which is greater than distance. It is
1573 // not safe to do vectorization.
1574 uint64_t MinDistanceNeeded =
1575 TypeByteSize * Stride * (MinNumIter - 1) + TypeByteSize;
1576 if (MinDistanceNeeded > static_cast<uint64_t>(Distance)) {
1577 LLVM_DEBUG(dbgs() << "LAA: Failure because of positive distance "
1578 << Distance << '\n');
1579 return Dependence::Backward;
1582 // Unsafe if the minimum distance needed is greater than max safe distance.
1583 if (MinDistanceNeeded > MaxSafeDepDistBytes) {
1584 LLVM_DEBUG(dbgs() << "LAA: Failure because it needs at least "
1585 << MinDistanceNeeded << " size in bytes");
1586 return Dependence::Backward;
1589 // Positive distance bigger than max vectorization factor.
1590 // FIXME: Should use max factor instead of max distance in bytes, which could
1591 // not handle different types.
1592 // E.g. Assume one char is 1 byte in memory and one int is 4 bytes.
1593 // void foo (int *A, char *B) {
1594 // for (unsigned i = 0; i < 1024; i++) {
1595 // A[i+2] = A[i] + 1;
1596 // B[i+2] = B[i] + 1;
1600 // This case is currently unsafe according to the max safe distance. If we
1601 // analyze the two accesses on array B, the max safe dependence distance
1602 // is 2. Then we analyze the accesses on array A, the minimum distance needed
1603 // is 8, which is less than 2 and forbidden vectorization, But actually
1604 // both A and B could be vectorized by 2 iterations.
1605 MaxSafeDepDistBytes =
1606 std::min(static_cast<uint64_t>(Distance), MaxSafeDepDistBytes);
1608 bool IsTrueDataDependence = (!AIsWrite && BIsWrite);
1609 if (IsTrueDataDependence && EnableForwardingConflictDetection &&
1610 couldPreventStoreLoadForward(Distance, TypeByteSize))
1611 return Dependence::BackwardVectorizableButPreventsForwarding;
1613 uint64_t MaxVF = MaxSafeDepDistBytes / (TypeByteSize * Stride);
1614 LLVM_DEBUG(dbgs() << "LAA: Positive distance " << Val.getSExtValue()
1615 << " with max VF = " << MaxVF << '\n');
1616 uint64_t MaxVFInBits = MaxVF * TypeByteSize * 8;
1617 MaxSafeRegisterWidth = std::min(MaxSafeRegisterWidth, MaxVFInBits);
1618 return Dependence::BackwardVectorizable;
1621 bool MemoryDepChecker::areDepsSafe(DepCandidates &AccessSets,
1622 MemAccessInfoList &CheckDeps,
1623 const ValueToValueMap &Strides) {
1625 MaxSafeDepDistBytes = -1;
1626 SmallPtrSet<MemAccessInfo, 8> Visited;
1627 for (MemAccessInfo CurAccess : CheckDeps) {
1628 if (Visited.count(CurAccess))
1631 // Get the relevant memory access set.
1632 EquivalenceClasses<MemAccessInfo>::iterator I =
1633 AccessSets.findValue(AccessSets.getLeaderValue(CurAccess));
1635 // Check accesses within this set.
1636 EquivalenceClasses<MemAccessInfo>::member_iterator AI =
1637 AccessSets.member_begin(I);
1638 EquivalenceClasses<MemAccessInfo>::member_iterator AE =
1639 AccessSets.member_end();
1641 // Check every access pair.
1643 Visited.insert(*AI);
1644 EquivalenceClasses<MemAccessInfo>::member_iterator OI = std::next(AI);
1646 // Check every accessing instruction pair in program order.
1647 for (std::vector<unsigned>::iterator I1 = Accesses[*AI].begin(),
1648 I1E = Accesses[*AI].end(); I1 != I1E; ++I1)
1649 for (std::vector<unsigned>::iterator I2 = Accesses[*OI].begin(),
1650 I2E = Accesses[*OI].end(); I2 != I2E; ++I2) {
1651 auto A = std::make_pair(&*AI, *I1);
1652 auto B = std::make_pair(&*OI, *I2);
1658 Dependence::DepType Type =
1659 isDependent(*A.first, A.second, *B.first, B.second, Strides);
1660 mergeInStatus(Dependence::isSafeForVectorization(Type));
1662 // Gather dependences unless we accumulated MaxDependences
1663 // dependences. In that case return as soon as we find the first
1664 // unsafe dependence. This puts a limit on this quadratic
1666 if (RecordDependences) {
1667 if (Type != Dependence::NoDep)
1668 Dependences.push_back(Dependence(A.second, B.second, Type));
1670 if (Dependences.size() >= MaxDependences) {
1671 RecordDependences = false;
1672 Dependences.clear();
1674 << "Too many dependences, stopped recording\n");
1677 if (!RecordDependences && !isSafeForVectorization())
1686 LLVM_DEBUG(dbgs() << "Total Dependences: " << Dependences.size() << "\n");
1687 return isSafeForVectorization();
1690 SmallVector<Instruction *, 4>
1691 MemoryDepChecker::getInstructionsForAccess(Value *Ptr, bool isWrite) const {
1692 MemAccessInfo Access(Ptr, isWrite);
1693 auto &IndexVector = Accesses.find(Access)->second;
1695 SmallVector<Instruction *, 4> Insts;
1696 transform(IndexVector,
1697 std::back_inserter(Insts),
1698 [&](unsigned Idx) { return this->InstMap[Idx]; });
1702 const char *MemoryDepChecker::Dependence::DepName[] = {
1703 "NoDep", "Unknown", "Forward", "ForwardButPreventsForwarding", "Backward",
1704 "BackwardVectorizable", "BackwardVectorizableButPreventsForwarding"};
1706 void MemoryDepChecker::Dependence::print(
1707 raw_ostream &OS, unsigned Depth,
1708 const SmallVectorImpl<Instruction *> &Instrs) const {
1709 OS.indent(Depth) << DepName[Type] << ":\n";
1710 OS.indent(Depth + 2) << *Instrs[Source] << " -> \n";
1711 OS.indent(Depth + 2) << *Instrs[Destination] << "\n";
1714 bool LoopAccessInfo::canAnalyzeLoop() {
1715 // We need to have a loop header.
1716 LLVM_DEBUG(dbgs() << "LAA: Found a loop in "
1717 << TheLoop->getHeader()->getParent()->getName() << ": "
1718 << TheLoop->getHeader()->getName() << '\n');
1720 // We can only analyze innermost loops.
1721 if (!TheLoop->empty()) {
1722 LLVM_DEBUG(dbgs() << "LAA: loop is not the innermost loop\n");
1723 recordAnalysis("NotInnerMostLoop") << "loop is not the innermost loop";
1727 // We must have a single backedge.
1728 if (TheLoop->getNumBackEdges() != 1) {
1730 dbgs() << "LAA: loop control flow is not understood by analyzer\n");
1731 recordAnalysis("CFGNotUnderstood")
1732 << "loop control flow is not understood by analyzer";
1736 // We must have a single exiting block.
1737 if (!TheLoop->getExitingBlock()) {
1739 dbgs() << "LAA: loop control flow is not understood by analyzer\n");
1740 recordAnalysis("CFGNotUnderstood")
1741 << "loop control flow is not understood by analyzer";
1745 // We only handle bottom-tested loops, i.e. loop in which the condition is
1746 // checked at the end of each iteration. With that we can assume that all
1747 // instructions in the loop are executed the same number of times.
1748 if (TheLoop->getExitingBlock() != TheLoop->getLoopLatch()) {
1750 dbgs() << "LAA: loop control flow is not understood by analyzer\n");
1751 recordAnalysis("CFGNotUnderstood")
1752 << "loop control flow is not understood by analyzer";
1756 // ScalarEvolution needs to be able to find the exit count.
1757 const SCEV *ExitCount = PSE->getBackedgeTakenCount();
1758 if (ExitCount == PSE->getSE()->getCouldNotCompute()) {
1759 recordAnalysis("CantComputeNumberOfIterations")
1760 << "could not determine number of loop iterations";
1761 LLVM_DEBUG(dbgs() << "LAA: SCEV could not compute the loop exit count.\n");
1768 void LoopAccessInfo::analyzeLoop(AliasAnalysis *AA, LoopInfo *LI,
1769 const TargetLibraryInfo *TLI,
1770 DominatorTree *DT) {
1771 typedef SmallPtrSet<Value*, 16> ValueSet;
1773 // Holds the Load and Store instructions.
1774 SmallVector<LoadInst *, 16> Loads;
1775 SmallVector<StoreInst *, 16> Stores;
1777 // Holds all the different accesses in the loop.
1778 unsigned NumReads = 0;
1779 unsigned NumReadWrites = 0;
1781 bool HasComplexMemInst = false;
1783 // A runtime check is only legal to insert if there are no convergent calls.
1784 HasConvergentOp = false;
1786 PtrRtChecking->Pointers.clear();
1787 PtrRtChecking->Need = false;
1789 const bool IsAnnotatedParallel = TheLoop->isAnnotatedParallel();
1792 for (BasicBlock *BB : TheLoop->blocks()) {
1793 // Scan the BB and collect legal loads and stores. Also detect any
1794 // convergent instructions.
1795 for (Instruction &I : *BB) {
1796 if (auto *Call = dyn_cast<CallBase>(&I)) {
1797 if (Call->isConvergent())
1798 HasConvergentOp = true;
1801 // With both a non-vectorizable memory instruction and a convergent
1802 // operation, found in this loop, no reason to continue the search.
1803 if (HasComplexMemInst && HasConvergentOp) {
1808 // Avoid hitting recordAnalysis multiple times.
1809 if (HasComplexMemInst)
1812 // If this is a load, save it. If this instruction can read from memory
1813 // but is not a load, then we quit. Notice that we don't handle function
1814 // calls that read or write.
1815 if (I.mayReadFromMemory()) {
1816 // Many math library functions read the rounding mode. We will only
1817 // vectorize a loop if it contains known function calls that don't set
1818 // the flag. Therefore, it is safe to ignore this read from memory.
1819 auto *Call = dyn_cast<CallInst>(&I);
1820 if (Call && getVectorIntrinsicIDForCall(Call, TLI))
1823 // If the function has an explicit vectorized counterpart, we can safely
1824 // assume that it can be vectorized.
1825 if (Call && !Call->isNoBuiltin() && Call->getCalledFunction() &&
1826 TLI->isFunctionVectorizable(Call->getCalledFunction()->getName()))
1829 auto *Ld = dyn_cast<LoadInst>(&I);
1831 recordAnalysis("CantVectorizeInstruction", Ld)
1832 << "instruction cannot be vectorized";
1833 HasComplexMemInst = true;
1836 if (!Ld->isSimple() && !IsAnnotatedParallel) {
1837 recordAnalysis("NonSimpleLoad", Ld)
1838 << "read with atomic ordering or volatile read";
1839 LLVM_DEBUG(dbgs() << "LAA: Found a non-simple load.\n");
1840 HasComplexMemInst = true;
1844 Loads.push_back(Ld);
1845 DepChecker->addAccess(Ld);
1846 if (EnableMemAccessVersioning)
1847 collectStridedAccess(Ld);
1851 // Save 'store' instructions. Abort if other instructions write to memory.
1852 if (I.mayWriteToMemory()) {
1853 auto *St = dyn_cast<StoreInst>(&I);
1855 recordAnalysis("CantVectorizeInstruction", St)
1856 << "instruction cannot be vectorized";
1857 HasComplexMemInst = true;
1860 if (!St->isSimple() && !IsAnnotatedParallel) {
1861 recordAnalysis("NonSimpleStore", St)
1862 << "write with atomic ordering or volatile write";
1863 LLVM_DEBUG(dbgs() << "LAA: Found a non-simple store.\n");
1864 HasComplexMemInst = true;
1868 Stores.push_back(St);
1869 DepChecker->addAccess(St);
1870 if (EnableMemAccessVersioning)
1871 collectStridedAccess(St);
1876 if (HasComplexMemInst) {
1881 // Now we have two lists that hold the loads and the stores.
1882 // Next, we find the pointers that they use.
1884 // Check if we see any stores. If there are no stores, then we don't
1885 // care if the pointers are *restrict*.
1886 if (!Stores.size()) {
1887 LLVM_DEBUG(dbgs() << "LAA: Found a read-only loop!\n");
1892 MemoryDepChecker::DepCandidates DependentAccesses;
1893 AccessAnalysis Accesses(TheLoop->getHeader()->getModule()->getDataLayout(),
1894 TheLoop, AA, LI, DependentAccesses, *PSE);
1896 // Holds the analyzed pointers. We don't want to call GetUnderlyingObjects
1897 // multiple times on the same object. If the ptr is accessed twice, once
1898 // for read and once for write, it will only appear once (on the write
1899 // list). This is okay, since we are going to check for conflicts between
1900 // writes and between reads and writes, but not between reads and reads.
1903 // Record uniform store addresses to identify if we have multiple stores
1904 // to the same address.
1905 ValueSet UniformStores;
1907 for (StoreInst *ST : Stores) {
1908 Value *Ptr = ST->getPointerOperand();
1911 HasDependenceInvolvingLoopInvariantAddress |=
1912 !UniformStores.insert(Ptr).second;
1914 // If we did *not* see this pointer before, insert it to the read-write
1915 // list. At this phase it is only a 'write' list.
1916 if (Seen.insert(Ptr).second) {
1919 MemoryLocation Loc = MemoryLocation::get(ST);
1920 // The TBAA metadata could have a control dependency on the predication
1921 // condition, so we cannot rely on it when determining whether or not we
1922 // need runtime pointer checks.
1923 if (blockNeedsPredication(ST->getParent(), TheLoop, DT))
1924 Loc.AATags.TBAA = nullptr;
1926 Accesses.addStore(Loc);
1930 if (IsAnnotatedParallel) {
1932 dbgs() << "LAA: A loop annotated parallel, ignore memory dependency "
1938 for (LoadInst *LD : Loads) {
1939 Value *Ptr = LD->getPointerOperand();
1940 // If we did *not* see this pointer before, insert it to the
1941 // read list. If we *did* see it before, then it is already in
1942 // the read-write list. This allows us to vectorize expressions
1943 // such as A[i] += x; Because the address of A[i] is a read-write
1944 // pointer. This only works if the index of A[i] is consecutive.
1945 // If the address of i is unknown (for example A[B[i]]) then we may
1946 // read a few words, modify, and write a few words, and some of the
1947 // words may be written to the same address.
1948 bool IsReadOnlyPtr = false;
1949 if (Seen.insert(Ptr).second ||
1950 !getPtrStride(*PSE, Ptr, TheLoop, SymbolicStrides)) {
1952 IsReadOnlyPtr = true;
1955 // See if there is an unsafe dependency between a load to a uniform address and
1956 // store to the same uniform address.
1957 if (UniformStores.count(Ptr)) {
1958 LLVM_DEBUG(dbgs() << "LAA: Found an unsafe dependency between a uniform "
1959 "load and uniform store to the same address!\n");
1960 HasDependenceInvolvingLoopInvariantAddress = true;
1963 MemoryLocation Loc = MemoryLocation::get(LD);
1964 // The TBAA metadata could have a control dependency on the predication
1965 // condition, so we cannot rely on it when determining whether or not we
1966 // need runtime pointer checks.
1967 if (blockNeedsPredication(LD->getParent(), TheLoop, DT))
1968 Loc.AATags.TBAA = nullptr;
1970 Accesses.addLoad(Loc, IsReadOnlyPtr);
1973 // If we write (or read-write) to a single destination and there are no
1974 // other reads in this loop then is it safe to vectorize.
1975 if (NumReadWrites == 1 && NumReads == 0) {
1976 LLVM_DEBUG(dbgs() << "LAA: Found a write-only loop!\n");
1981 // Build dependence sets and check whether we need a runtime pointer bounds
1983 Accesses.buildDependenceSets();
1985 // Find pointers with computable bounds. We are going to use this information
1986 // to place a runtime bound check.
1987 bool CanDoRTIfNeeded = Accesses.canCheckPtrAtRT(*PtrRtChecking, PSE->getSE(),
1988 TheLoop, SymbolicStrides);
1989 if (!CanDoRTIfNeeded) {
1990 recordAnalysis("CantIdentifyArrayBounds") << "cannot identify array bounds";
1991 LLVM_DEBUG(dbgs() << "LAA: We can't vectorize because we can't find "
1992 << "the array bounds.\n");
1998 dbgs() << "LAA: May be able to perform a memory runtime check if needed.\n");
2001 if (Accesses.isDependencyCheckNeeded()) {
2002 LLVM_DEBUG(dbgs() << "LAA: Checking memory dependencies\n");
2003 CanVecMem = DepChecker->areDepsSafe(
2004 DependentAccesses, Accesses.getDependenciesToCheck(), SymbolicStrides);
2005 MaxSafeDepDistBytes = DepChecker->getMaxSafeDepDistBytes();
2007 if (!CanVecMem && DepChecker->shouldRetryWithRuntimeCheck()) {
2008 LLVM_DEBUG(dbgs() << "LAA: Retrying with memory checks\n");
2010 // Clear the dependency checks. We assume they are not needed.
2011 Accesses.resetDepChecks(*DepChecker);
2013 PtrRtChecking->reset();
2014 PtrRtChecking->Need = true;
2016 auto *SE = PSE->getSE();
2017 CanDoRTIfNeeded = Accesses.canCheckPtrAtRT(*PtrRtChecking, SE, TheLoop,
2018 SymbolicStrides, true);
2020 // Check that we found the bounds for the pointer.
2021 if (!CanDoRTIfNeeded) {
2022 recordAnalysis("CantCheckMemDepsAtRunTime")
2023 << "cannot check memory dependencies at runtime";
2024 LLVM_DEBUG(dbgs() << "LAA: Can't vectorize with memory checks\n");
2033 if (HasConvergentOp) {
2034 recordAnalysis("CantInsertRuntimeCheckWithConvergent")
2035 << "cannot add control dependency to convergent operation";
2036 LLVM_DEBUG(dbgs() << "LAA: We can't vectorize because a runtime check "
2037 "would be needed with a convergent operation\n");
2044 dbgs() << "LAA: No unsafe dependent memory operations in loop. We"
2045 << (PtrRtChecking->Need ? "" : " don't")
2046 << " need runtime memory checks.\n");
2048 recordAnalysis("UnsafeMemDep")
2049 << "unsafe dependent memory operations in loop. Use "
2050 "#pragma loop distribute(enable) to allow loop distribution "
2051 "to attempt to isolate the offending operations into a separate "
2053 LLVM_DEBUG(dbgs() << "LAA: unsafe dependent memory operations in loop\n");
2057 bool LoopAccessInfo::blockNeedsPredication(BasicBlock *BB, Loop *TheLoop,
2058 DominatorTree *DT) {
2059 assert(TheLoop->contains(BB) && "Unknown block used");
2061 // Blocks that do not dominate the latch need predication.
2062 BasicBlock* Latch = TheLoop->getLoopLatch();
2063 return !DT->dominates(BB, Latch);
2066 OptimizationRemarkAnalysis &LoopAccessInfo::recordAnalysis(StringRef RemarkName,
2068 assert(!Report && "Multiple reports generated");
2070 Value *CodeRegion = TheLoop->getHeader();
2071 DebugLoc DL = TheLoop->getStartLoc();
2074 CodeRegion = I->getParent();
2075 // If there is no debug location attached to the instruction, revert back to
2076 // using the loop's.
2077 if (I->getDebugLoc())
2078 DL = I->getDebugLoc();
2081 Report = make_unique<OptimizationRemarkAnalysis>(DEBUG_TYPE, RemarkName, DL,
2086 bool LoopAccessInfo::isUniform(Value *V) const {
2087 auto *SE = PSE->getSE();
2088 // Since we rely on SCEV for uniformity, if the type is not SCEVable, it is
2089 // never considered uniform.
2090 // TODO: Is this really what we want? Even without FP SCEV, we may want some
2091 // trivially loop-invariant FP values to be considered uniform.
2092 if (!SE->isSCEVable(V->getType()))
2094 return (SE->isLoopInvariant(SE->getSCEV(V), TheLoop));
2097 // FIXME: this function is currently a duplicate of the one in
2098 // LoopVectorize.cpp.
2099 static Instruction *getFirstInst(Instruction *FirstInst, Value *V,
2103 if (Instruction *I = dyn_cast<Instruction>(V))
2104 return I->getParent() == Loc->getParent() ? I : nullptr;
2110 /// IR Values for the lower and upper bounds of a pointer evolution. We
2111 /// need to use value-handles because SCEV expansion can invalidate previously
2112 /// expanded values. Thus expansion of a pointer can invalidate the bounds for
2114 struct PointerBounds {
2115 TrackingVH<Value> Start;
2116 TrackingVH<Value> End;
2119 } // end anonymous namespace
2121 /// Expand code for the lower and upper bound of the pointer group \p CG
2122 /// in \p TheLoop. \return the values for the bounds.
2123 static PointerBounds
2124 expandBounds(const RuntimePointerChecking::CheckingPtrGroup *CG, Loop *TheLoop,
2125 Instruction *Loc, SCEVExpander &Exp, ScalarEvolution *SE,
2126 const RuntimePointerChecking &PtrRtChecking) {
2127 Value *Ptr = PtrRtChecking.Pointers[CG->Members[0]].PointerValue;
2128 const SCEV *Sc = SE->getSCEV(Ptr);
2130 unsigned AS = Ptr->getType()->getPointerAddressSpace();
2131 LLVMContext &Ctx = Loc->getContext();
2133 // Use this type for pointer arithmetic.
2134 Type *PtrArithTy = Type::getInt8PtrTy(Ctx, AS);
2136 if (SE->isLoopInvariant(Sc, TheLoop)) {
2137 LLVM_DEBUG(dbgs() << "LAA: Adding RT check for a loop invariant ptr:"
2139 // Ptr could be in the loop body. If so, expand a new one at the correct
2141 Instruction *Inst = dyn_cast<Instruction>(Ptr);
2142 Value *NewPtr = (Inst && TheLoop->contains(Inst))
2143 ? Exp.expandCodeFor(Sc, PtrArithTy, Loc)
2145 // We must return a half-open range, which means incrementing Sc.
2146 const SCEV *ScPlusOne = SE->getAddExpr(Sc, SE->getOne(PtrArithTy));
2147 Value *NewPtrPlusOne = Exp.expandCodeFor(ScPlusOne, PtrArithTy, Loc);
2148 return {NewPtr, NewPtrPlusOne};
2150 Value *Start = nullptr, *End = nullptr;
2151 LLVM_DEBUG(dbgs() << "LAA: Adding RT check for range:\n");
2152 Start = Exp.expandCodeFor(CG->Low, PtrArithTy, Loc);
2153 End = Exp.expandCodeFor(CG->High, PtrArithTy, Loc);
2154 LLVM_DEBUG(dbgs() << "Start: " << *CG->Low << " End: " << *CG->High
2156 return {Start, End};
2160 /// Turns a collection of checks into a collection of expanded upper and
2161 /// lower bounds for both pointers in the check.
2162 static SmallVector<std::pair<PointerBounds, PointerBounds>, 4> expandBounds(
2163 const SmallVectorImpl<RuntimePointerChecking::PointerCheck> &PointerChecks,
2164 Loop *L, Instruction *Loc, ScalarEvolution *SE, SCEVExpander &Exp,
2165 const RuntimePointerChecking &PtrRtChecking) {
2166 SmallVector<std::pair<PointerBounds, PointerBounds>, 4> ChecksWithBounds;
2168 // Here we're relying on the SCEV Expander's cache to only emit code for the
2169 // same bounds once.
2171 PointerChecks, std::back_inserter(ChecksWithBounds),
2172 [&](const RuntimePointerChecking::PointerCheck &Check) {
2174 First = expandBounds(Check.first, L, Loc, Exp, SE, PtrRtChecking),
2175 Second = expandBounds(Check.second, L, Loc, Exp, SE, PtrRtChecking);
2176 return std::make_pair(First, Second);
2179 return ChecksWithBounds;
2182 std::pair<Instruction *, Instruction *> LoopAccessInfo::addRuntimeChecks(
2184 const SmallVectorImpl<RuntimePointerChecking::PointerCheck> &PointerChecks)
2186 const DataLayout &DL = TheLoop->getHeader()->getModule()->getDataLayout();
2187 auto *SE = PSE->getSE();
2188 SCEVExpander Exp(*SE, DL, "induction");
2189 auto ExpandedChecks =
2190 expandBounds(PointerChecks, TheLoop, Loc, SE, Exp, *PtrRtChecking);
2192 LLVMContext &Ctx = Loc->getContext();
2193 Instruction *FirstInst = nullptr;
2194 IRBuilder<> ChkBuilder(Loc);
2195 // Our instructions might fold to a constant.
2196 Value *MemoryRuntimeCheck = nullptr;
2198 for (const auto &Check : ExpandedChecks) {
2199 const PointerBounds &A = Check.first, &B = Check.second;
2200 // Check if two pointers (A and B) conflict where conflict is computed as:
2201 // start(A) <= end(B) && start(B) <= end(A)
2202 unsigned AS0 = A.Start->getType()->getPointerAddressSpace();
2203 unsigned AS1 = B.Start->getType()->getPointerAddressSpace();
2205 assert((AS0 == B.End->getType()->getPointerAddressSpace()) &&
2206 (AS1 == A.End->getType()->getPointerAddressSpace()) &&
2207 "Trying to bounds check pointers with different address spaces");
2209 Type *PtrArithTy0 = Type::getInt8PtrTy(Ctx, AS0);
2210 Type *PtrArithTy1 = Type::getInt8PtrTy(Ctx, AS1);
2212 Value *Start0 = ChkBuilder.CreateBitCast(A.Start, PtrArithTy0, "bc");
2213 Value *Start1 = ChkBuilder.CreateBitCast(B.Start, PtrArithTy1, "bc");
2214 Value *End0 = ChkBuilder.CreateBitCast(A.End, PtrArithTy1, "bc");
2215 Value *End1 = ChkBuilder.CreateBitCast(B.End, PtrArithTy0, "bc");
2217 // [A|B].Start points to the first accessed byte under base [A|B].
2218 // [A|B].End points to the last accessed byte, plus one.
2219 // There is no conflict when the intervals are disjoint:
2220 // NoConflict = (B.Start >= A.End) || (A.Start >= B.End)
2222 // bound0 = (B.Start < A.End)
2223 // bound1 = (A.Start < B.End)
2224 // IsConflict = bound0 & bound1
2225 Value *Cmp0 = ChkBuilder.CreateICmpULT(Start0, End1, "bound0");
2226 FirstInst = getFirstInst(FirstInst, Cmp0, Loc);
2227 Value *Cmp1 = ChkBuilder.CreateICmpULT(Start1, End0, "bound1");
2228 FirstInst = getFirstInst(FirstInst, Cmp1, Loc);
2229 Value *IsConflict = ChkBuilder.CreateAnd(Cmp0, Cmp1, "found.conflict");
2230 FirstInst = getFirstInst(FirstInst, IsConflict, Loc);
2231 if (MemoryRuntimeCheck) {
2233 ChkBuilder.CreateOr(MemoryRuntimeCheck, IsConflict, "conflict.rdx");
2234 FirstInst = getFirstInst(FirstInst, IsConflict, Loc);
2236 MemoryRuntimeCheck = IsConflict;
2239 if (!MemoryRuntimeCheck)
2240 return std::make_pair(nullptr, nullptr);
2242 // We have to do this trickery because the IRBuilder might fold the check to a
2243 // constant expression in which case there is no Instruction anchored in a
2245 Instruction *Check = BinaryOperator::CreateAnd(MemoryRuntimeCheck,
2246 ConstantInt::getTrue(Ctx));
2247 ChkBuilder.Insert(Check, "memcheck.conflict");
2248 FirstInst = getFirstInst(FirstInst, Check, Loc);
2249 return std::make_pair(FirstInst, Check);
2252 std::pair<Instruction *, Instruction *>
2253 LoopAccessInfo::addRuntimeChecks(Instruction *Loc) const {
2254 if (!PtrRtChecking->Need)
2255 return std::make_pair(nullptr, nullptr);
2257 return addRuntimeChecks(Loc, PtrRtChecking->getChecks());
2260 void LoopAccessInfo::collectStridedAccess(Value *MemAccess) {
2261 Value *Ptr = nullptr;
2262 if (LoadInst *LI = dyn_cast<LoadInst>(MemAccess))
2263 Ptr = LI->getPointerOperand();
2264 else if (StoreInst *SI = dyn_cast<StoreInst>(MemAccess))
2265 Ptr = SI->getPointerOperand();
2269 Value *Stride = getStrideFromPointer(Ptr, PSE->getSE(), TheLoop);
2273 LLVM_DEBUG(dbgs() << "LAA: Found a strided access that is a candidate for "
2275 LLVM_DEBUG(dbgs() << " Ptr: " << *Ptr << " Stride: " << *Stride << "\n");
2277 // Avoid adding the "Stride == 1" predicate when we know that
2278 // Stride >= Trip-Count. Such a predicate will effectively optimize a single
2279 // or zero iteration loop, as Trip-Count <= Stride == 1.
2281 // TODO: We are currently not making a very informed decision on when it is
2282 // beneficial to apply stride versioning. It might make more sense that the
2283 // users of this analysis (such as the vectorizer) will trigger it, based on
2284 // their specific cost considerations; For example, in cases where stride
2285 // versioning does not help resolving memory accesses/dependences, the
2286 // vectorizer should evaluate the cost of the runtime test, and the benefit
2287 // of various possible stride specializations, considering the alternatives
2288 // of using gather/scatters (if available).
2290 const SCEV *StrideExpr = PSE->getSCEV(Stride);
2291 const SCEV *BETakenCount = PSE->getBackedgeTakenCount();
2293 // Match the types so we can compare the stride and the BETakenCount.
2294 // The Stride can be positive/negative, so we sign extend Stride;
2295 // The backedgeTakenCount is non-negative, so we zero extend BETakenCount.
2296 const DataLayout &DL = TheLoop->getHeader()->getModule()->getDataLayout();
2297 uint64_t StrideTypeSize = DL.getTypeAllocSize(StrideExpr->getType());
2298 uint64_t BETypeSize = DL.getTypeAllocSize(BETakenCount->getType());
2299 const SCEV *CastedStride = StrideExpr;
2300 const SCEV *CastedBECount = BETakenCount;
2301 ScalarEvolution *SE = PSE->getSE();
2302 if (BETypeSize >= StrideTypeSize)
2303 CastedStride = SE->getNoopOrSignExtend(StrideExpr, BETakenCount->getType());
2305 CastedBECount = SE->getZeroExtendExpr(BETakenCount, StrideExpr->getType());
2306 const SCEV *StrideMinusBETaken = SE->getMinusSCEV(CastedStride, CastedBECount);
2307 // Since TripCount == BackEdgeTakenCount + 1, checking:
2308 // "Stride >= TripCount" is equivalent to checking:
2309 // Stride - BETakenCount > 0
2310 if (SE->isKnownPositive(StrideMinusBETaken)) {
2312 dbgs() << "LAA: Stride>=TripCount; No point in versioning as the "
2313 "Stride==1 predicate will imply that the loop executes "
2317 LLVM_DEBUG(dbgs() << "LAA: Found a strided access that we can version.");
2319 SymbolicStrides[Ptr] = Stride;
2320 StrideSet.insert(Stride);
2323 LoopAccessInfo::LoopAccessInfo(Loop *L, ScalarEvolution *SE,
2324 const TargetLibraryInfo *TLI, AliasAnalysis *AA,
2325 DominatorTree *DT, LoopInfo *LI)
2326 : PSE(llvm::make_unique<PredicatedScalarEvolution>(*SE, *L)),
2327 PtrRtChecking(llvm::make_unique<RuntimePointerChecking>(SE)),
2328 DepChecker(llvm::make_unique<MemoryDepChecker>(*PSE, L)), TheLoop(L),
2329 NumLoads(0), NumStores(0), MaxSafeDepDistBytes(-1), CanVecMem(false),
2330 HasConvergentOp(false),
2331 HasDependenceInvolvingLoopInvariantAddress(false) {
2332 if (canAnalyzeLoop())
2333 analyzeLoop(AA, LI, TLI, DT);
2336 void LoopAccessInfo::print(raw_ostream &OS, unsigned Depth) const {
2338 OS.indent(Depth) << "Memory dependences are safe";
2339 if (MaxSafeDepDistBytes != -1ULL)
2340 OS << " with a maximum dependence distance of " << MaxSafeDepDistBytes
2342 if (PtrRtChecking->Need)
2343 OS << " with run-time checks";
2347 if (HasConvergentOp)
2348 OS.indent(Depth) << "Has convergent operation in loop\n";
2351 OS.indent(Depth) << "Report: " << Report->getMsg() << "\n";
2353 if (auto *Dependences = DepChecker->getDependences()) {
2354 OS.indent(Depth) << "Dependences:\n";
2355 for (auto &Dep : *Dependences) {
2356 Dep.print(OS, Depth + 2, DepChecker->getMemoryInstructions());
2360 OS.indent(Depth) << "Too many dependences, not recorded\n";
2362 // List the pair of accesses need run-time checks to prove independence.
2363 PtrRtChecking->print(OS, Depth);
2366 OS.indent(Depth) << "Non vectorizable stores to invariant address were "
2367 << (HasDependenceInvolvingLoopInvariantAddress ? "" : "not ")
2368 << "found in loop.\n";
2370 OS.indent(Depth) << "SCEV assumptions:\n";
2371 PSE->getUnionPredicate().print(OS, Depth);
2375 OS.indent(Depth) << "Expressions re-written:\n";
2376 PSE->print(OS, Depth);
2379 const LoopAccessInfo &LoopAccessLegacyAnalysis::getInfo(Loop *L) {
2380 auto &LAI = LoopAccessInfoMap[L];
2383 LAI = llvm::make_unique<LoopAccessInfo>(L, SE, TLI, AA, DT, LI);
2388 void LoopAccessLegacyAnalysis::print(raw_ostream &OS, const Module *M) const {
2389 LoopAccessLegacyAnalysis &LAA = *const_cast<LoopAccessLegacyAnalysis *>(this);
2391 for (Loop *TopLevelLoop : *LI)
2392 for (Loop *L : depth_first(TopLevelLoop)) {
2393 OS.indent(2) << L->getHeader()->getName() << ":\n";
2394 auto &LAI = LAA.getInfo(L);
2399 bool LoopAccessLegacyAnalysis::runOnFunction(Function &F) {
2400 SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE();
2401 auto *TLIP = getAnalysisIfAvailable<TargetLibraryInfoWrapperPass>();
2402 TLI = TLIP ? &TLIP->getTLI() : nullptr;
2403 AA = &getAnalysis<AAResultsWrapperPass>().getAAResults();
2404 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
2405 LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
2410 void LoopAccessLegacyAnalysis::getAnalysisUsage(AnalysisUsage &AU) const {
2411 AU.addRequired<ScalarEvolutionWrapperPass>();
2412 AU.addRequired<AAResultsWrapperPass>();
2413 AU.addRequired<DominatorTreeWrapperPass>();
2414 AU.addRequired<LoopInfoWrapperPass>();
2416 AU.setPreservesAll();
2419 char LoopAccessLegacyAnalysis::ID = 0;
2420 static const char laa_name[] = "Loop Access Analysis";
2421 #define LAA_NAME "loop-accesses"
2423 INITIALIZE_PASS_BEGIN(LoopAccessLegacyAnalysis, LAA_NAME, laa_name, false, true)
2424 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)
2425 INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass)
2426 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
2427 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
2428 INITIALIZE_PASS_END(LoopAccessLegacyAnalysis, LAA_NAME, laa_name, false, true)
2430 AnalysisKey LoopAccessAnalysis::Key;
2432 LoopAccessInfo LoopAccessAnalysis::run(Loop &L, LoopAnalysisManager &AM,
2433 LoopStandardAnalysisResults &AR) {
2434 return LoopAccessInfo(&L, &AR.SE, &AR.TLI, &AR.AA, &AR.DT, &AR.LI);
2439 Pass *createLAAPass() {
2440 return new LoopAccessLegacyAnalysis();
2443 } // end namespace llvm