1 //===- LoopAccessAnalysis.cpp - Loop Access Analysis Implementation --------==//
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 // The implementation for the loop memory dependence that was originally
11 // developed for the loop vectorizer.
13 //===----------------------------------------------------------------------===//
15 #include "llvm/Analysis/LoopAccessAnalysis.h"
16 #include "llvm/ADT/APInt.h"
17 #include "llvm/ADT/DenseMap.h"
18 #include "llvm/ADT/DepthFirstIterator.h"
19 #include "llvm/ADT/EquivalenceClasses.h"
20 #include "llvm/ADT/PointerIntPair.h"
21 #include "llvm/ADT/STLExtras.h"
22 #include "llvm/ADT/SetVector.h"
23 #include "llvm/ADT/SmallPtrSet.h"
24 #include "llvm/ADT/SmallSet.h"
25 #include "llvm/ADT/SmallVector.h"
26 #include "llvm/ADT/iterator_range.h"
27 #include "llvm/Analysis/AliasAnalysis.h"
28 #include "llvm/Analysis/AliasSetTracker.h"
29 #include "llvm/Analysis/LoopAnalysisManager.h"
30 #include "llvm/Analysis/LoopInfo.h"
31 #include "llvm/Analysis/MemoryLocation.h"
32 #include "llvm/Analysis/OptimizationDiagnosticInfo.h"
33 #include "llvm/Analysis/ScalarEvolution.h"
34 #include "llvm/Analysis/ScalarEvolutionExpander.h"
35 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
36 #include "llvm/Analysis/TargetLibraryInfo.h"
37 #include "llvm/Analysis/ValueTracking.h"
38 #include "llvm/Analysis/VectorUtils.h"
39 #include "llvm/IR/BasicBlock.h"
40 #include "llvm/IR/Constants.h"
41 #include "llvm/IR/DataLayout.h"
42 #include "llvm/IR/DebugLoc.h"
43 #include "llvm/IR/DerivedTypes.h"
44 #include "llvm/IR/DiagnosticInfo.h"
45 #include "llvm/IR/Dominators.h"
46 #include "llvm/IR/Function.h"
47 #include "llvm/IR/IRBuilder.h"
48 #include "llvm/IR/InstrTypes.h"
49 #include "llvm/IR/Instruction.h"
50 #include "llvm/IR/Instructions.h"
51 #include "llvm/IR/Operator.h"
52 #include "llvm/IR/PassManager.h"
53 #include "llvm/IR/Type.h"
54 #include "llvm/IR/Value.h"
55 #include "llvm/IR/ValueHandle.h"
56 #include "llvm/Pass.h"
57 #include "llvm/Support/Casting.h"
58 #include "llvm/Support/CommandLine.h"
59 #include "llvm/Support/Debug.h"
60 #include "llvm/Support/ErrorHandling.h"
61 #include "llvm/Support/raw_ostream.h"
72 #define DEBUG_TYPE "loop-accesses"
74 static cl::opt<unsigned, true>
75 VectorizationFactor("force-vector-width", cl::Hidden,
76 cl::desc("Sets the SIMD width. Zero is autoselect."),
77 cl::location(VectorizerParams::VectorizationFactor));
78 unsigned VectorizerParams::VectorizationFactor;
80 static cl::opt<unsigned, true>
81 VectorizationInterleave("force-vector-interleave", cl::Hidden,
82 cl::desc("Sets the vectorization interleave count. "
83 "Zero is autoselect."),
85 VectorizerParams::VectorizationInterleave));
86 unsigned VectorizerParams::VectorizationInterleave;
88 static cl::opt<unsigned, true> RuntimeMemoryCheckThreshold(
89 "runtime-memory-check-threshold", cl::Hidden,
90 cl::desc("When performing memory disambiguation checks at runtime do not "
91 "generate more than this number of comparisons (default = 8)."),
92 cl::location(VectorizerParams::RuntimeMemoryCheckThreshold), cl::init(8));
93 unsigned VectorizerParams::RuntimeMemoryCheckThreshold;
95 /// \brief The maximum iterations used to merge memory checks
96 static cl::opt<unsigned> MemoryCheckMergeThreshold(
97 "memory-check-merge-threshold", cl::Hidden,
98 cl::desc("Maximum number of comparisons done when trying to merge "
99 "runtime memory checks. (default = 100)"),
102 /// Maximum SIMD width.
103 const unsigned VectorizerParams::MaxVectorWidth = 64;
105 /// \brief We collect dependences up to this threshold.
106 static cl::opt<unsigned>
107 MaxDependences("max-dependences", cl::Hidden,
108 cl::desc("Maximum number of dependences collected by "
109 "loop-access analysis (default = 100)"),
112 /// This enables versioning on the strides of symbolically striding memory
113 /// accesses in code like the following.
114 /// for (i = 0; i < N; ++i)
115 /// A[i * Stride1] += B[i * Stride2] ...
117 /// Will be roughly translated to
118 /// if (Stride1 == 1 && Stride2 == 1) {
119 /// for (i = 0; i < N; i+=4)
123 static cl::opt<bool> EnableMemAccessVersioning(
124 "enable-mem-access-versioning", cl::init(true), cl::Hidden,
125 cl::desc("Enable symbolic stride memory access versioning"));
127 /// \brief Enable store-to-load forwarding conflict detection. This option can
128 /// be disabled for correctness testing.
129 static cl::opt<bool> EnableForwardingConflictDetection(
130 "store-to-load-forwarding-conflict-detection", cl::Hidden,
131 cl::desc("Enable conflict detection in loop-access analysis"),
134 bool VectorizerParams::isInterleaveForced() {
135 return ::VectorizationInterleave.getNumOccurrences() > 0;
138 void LoopAccessReport::emitAnalysis(const LoopAccessReport &Message,
139 const Loop *TheLoop, const char *PassName,
140 OptimizationRemarkEmitter &ORE) {
141 DebugLoc DL = TheLoop->getStartLoc();
142 const Value *V = TheLoop->getHeader();
143 if (const Instruction *I = Message.getInstr()) {
144 // If there is no debug location attached to the instruction, revert back to
146 if (I->getDebugLoc())
147 DL = I->getDebugLoc();
150 ORE.emitOptimizationRemarkAnalysis(PassName, DL, V, Message.str());
153 Value *llvm::stripIntegerCast(Value *V) {
154 if (auto *CI = dyn_cast<CastInst>(V))
155 if (CI->getOperand(0)->getType()->isIntegerTy())
156 return CI->getOperand(0);
160 const SCEV *llvm::replaceSymbolicStrideSCEV(PredicatedScalarEvolution &PSE,
161 const ValueToValueMap &PtrToStride,
162 Value *Ptr, Value *OrigPtr) {
163 const SCEV *OrigSCEV = PSE.getSCEV(Ptr);
165 // If there is an entry in the map return the SCEV of the pointer with the
166 // symbolic stride replaced by one.
167 ValueToValueMap::const_iterator SI =
168 PtrToStride.find(OrigPtr ? OrigPtr : Ptr);
169 if (SI != PtrToStride.end()) {
170 Value *StrideVal = SI->second;
173 StrideVal = stripIntegerCast(StrideVal);
175 // Replace symbolic stride by one.
176 Value *One = ConstantInt::get(StrideVal->getType(), 1);
177 ValueToValueMap RewriteMap;
178 RewriteMap[StrideVal] = One;
180 ScalarEvolution *SE = PSE.getSE();
181 const auto *U = cast<SCEVUnknown>(SE->getSCEV(StrideVal));
183 static_cast<const SCEVConstant *>(SE->getOne(StrideVal->getType()));
185 PSE.addPredicate(*SE->getEqualPredicate(U, CT));
186 auto *Expr = PSE.getSCEV(Ptr);
188 DEBUG(dbgs() << "LAA: Replacing SCEV: " << *OrigSCEV << " by: " << *Expr
193 // Otherwise, just return the SCEV of the original pointer.
197 /// Calculate Start and End points of memory access.
198 /// Let's assume A is the first access and B is a memory access on N-th loop
199 /// iteration. Then B is calculated as:
201 /// Step value may be positive or negative.
202 /// N is a calculated back-edge taken count:
203 /// N = (TripCount > 0) ? RoundDown(TripCount -1 , VF) : 0
204 /// Start and End points are calculated in the following way:
205 /// Start = UMIN(A, B) ; End = UMAX(A, B) + SizeOfElt,
206 /// where SizeOfElt is the size of single memory access in bytes.
208 /// There is no conflict when the intervals are disjoint:
209 /// NoConflict = (P2.Start >= P1.End) || (P1.Start >= P2.End)
210 void RuntimePointerChecking::insert(Loop *Lp, Value *Ptr, bool WritePtr,
211 unsigned DepSetId, unsigned ASId,
212 const ValueToValueMap &Strides,
213 PredicatedScalarEvolution &PSE) {
214 // Get the stride replaced scev.
215 const SCEV *Sc = replaceSymbolicStrideSCEV(PSE, Strides, Ptr);
216 ScalarEvolution *SE = PSE.getSE();
221 if (SE->isLoopInvariant(Sc, Lp))
222 ScStart = ScEnd = Sc;
224 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Sc);
225 assert(AR && "Invalid addrec expression");
226 const SCEV *Ex = PSE.getBackedgeTakenCount();
228 ScStart = AR->getStart();
229 ScEnd = AR->evaluateAtIteration(Ex, *SE);
230 const SCEV *Step = AR->getStepRecurrence(*SE);
232 // For expressions with negative step, the upper bound is ScStart and the
233 // lower bound is ScEnd.
234 if (const auto *CStep = dyn_cast<SCEVConstant>(Step)) {
235 if (CStep->getValue()->isNegative())
236 std::swap(ScStart, ScEnd);
238 // Fallback case: the step is not constant, but we can still
239 // get the upper and lower bounds of the interval by using min/max
241 ScStart = SE->getUMinExpr(ScStart, ScEnd);
242 ScEnd = SE->getUMaxExpr(AR->getStart(), ScEnd);
244 // Add the size of the pointed element to ScEnd.
246 Ptr->getType()->getPointerElementType()->getScalarSizeInBits() / 8;
247 const SCEV *EltSizeSCEV = SE->getConstant(ScEnd->getType(), EltSize);
248 ScEnd = SE->getAddExpr(ScEnd, EltSizeSCEV);
251 Pointers.emplace_back(Ptr, ScStart, ScEnd, WritePtr, DepSetId, ASId, Sc);
254 SmallVector<RuntimePointerChecking::PointerCheck, 4>
255 RuntimePointerChecking::generateChecks() const {
256 SmallVector<PointerCheck, 4> Checks;
258 for (unsigned I = 0; I < CheckingGroups.size(); ++I) {
259 for (unsigned J = I + 1; J < CheckingGroups.size(); ++J) {
260 const RuntimePointerChecking::CheckingPtrGroup &CGI = CheckingGroups[I];
261 const RuntimePointerChecking::CheckingPtrGroup &CGJ = CheckingGroups[J];
263 if (needsChecking(CGI, CGJ))
264 Checks.push_back(std::make_pair(&CGI, &CGJ));
270 void RuntimePointerChecking::generateChecks(
271 MemoryDepChecker::DepCandidates &DepCands, bool UseDependencies) {
272 assert(Checks.empty() && "Checks is not empty");
273 groupChecks(DepCands, UseDependencies);
274 Checks = generateChecks();
277 bool RuntimePointerChecking::needsChecking(const CheckingPtrGroup &M,
278 const CheckingPtrGroup &N) const {
279 for (unsigned I = 0, EI = M.Members.size(); EI != I; ++I)
280 for (unsigned J = 0, EJ = N.Members.size(); EJ != J; ++J)
281 if (needsChecking(M.Members[I], N.Members[J]))
286 /// Compare \p I and \p J and return the minimum.
287 /// Return nullptr in case we couldn't find an answer.
288 static const SCEV *getMinFromExprs(const SCEV *I, const SCEV *J,
289 ScalarEvolution *SE) {
290 const SCEV *Diff = SE->getMinusSCEV(J, I);
291 const SCEVConstant *C = dyn_cast<const SCEVConstant>(Diff);
295 if (C->getValue()->isNegative())
300 bool RuntimePointerChecking::CheckingPtrGroup::addPointer(unsigned Index) {
301 const SCEV *Start = RtCheck.Pointers[Index].Start;
302 const SCEV *End = RtCheck.Pointers[Index].End;
304 // Compare the starts and ends with the known minimum and maximum
305 // of this set. We need to know how we compare against the min/max
306 // of the set in order to be able to emit memchecks.
307 const SCEV *Min0 = getMinFromExprs(Start, Low, RtCheck.SE);
311 const SCEV *Min1 = getMinFromExprs(End, High, RtCheck.SE);
315 // Update the low bound expression if we've found a new min value.
319 // Update the high bound expression if we've found a new max value.
323 Members.push_back(Index);
327 void RuntimePointerChecking::groupChecks(
328 MemoryDepChecker::DepCandidates &DepCands, bool UseDependencies) {
329 // We build the groups from dependency candidates equivalence classes
331 // - We know that pointers in the same equivalence class share
332 // the same underlying object and therefore there is a chance
333 // that we can compare pointers
334 // - We wouldn't be able to merge two pointers for which we need
335 // to emit a memcheck. The classes in DepCands are already
336 // conveniently built such that no two pointers in the same
337 // class need checking against each other.
339 // We use the following (greedy) algorithm to construct the groups
340 // For every pointer in the equivalence class:
341 // For each existing group:
342 // - if the difference between this pointer and the min/max bounds
343 // of the group is a constant, then make the pointer part of the
344 // group and update the min/max bounds of that group as required.
346 CheckingGroups.clear();
348 // If we need to check two pointers to the same underlying object
349 // with a non-constant difference, we shouldn't perform any pointer
350 // grouping with those pointers. This is because we can easily get
351 // into cases where the resulting check would return false, even when
352 // the accesses are safe.
354 // The following example shows this:
355 // for (i = 0; i < 1000; ++i)
356 // a[5000 + i * m] = a[i] + a[i + 9000]
358 // Here grouping gives a check of (5000, 5000 + 1000 * m) against
359 // (0, 10000) which is always false. However, if m is 1, there is no
360 // dependence. Not grouping the checks for a[i] and a[i + 9000] allows
361 // us to perform an accurate check in this case.
363 // The above case requires that we have an UnknownDependence between
364 // accesses to the same underlying object. This cannot happen unless
365 // ShouldRetryWithRuntimeCheck is set, and therefore UseDependencies
366 // is also false. In this case we will use the fallback path and create
367 // separate checking groups for all pointers.
369 // If we don't have the dependency partitions, construct a new
370 // checking pointer group for each pointer. This is also required
371 // for correctness, because in this case we can have checking between
372 // pointers to the same underlying object.
373 if (!UseDependencies) {
374 for (unsigned I = 0; I < Pointers.size(); ++I)
375 CheckingGroups.push_back(CheckingPtrGroup(I, *this));
379 unsigned TotalComparisons = 0;
381 DenseMap<Value *, unsigned> PositionMap;
382 for (unsigned Index = 0; Index < Pointers.size(); ++Index)
383 PositionMap[Pointers[Index].PointerValue] = Index;
385 // We need to keep track of what pointers we've already seen so we
386 // don't process them twice.
387 SmallSet<unsigned, 2> Seen;
389 // Go through all equivalence classes, get the "pointer check groups"
390 // and add them to the overall solution. We use the order in which accesses
391 // appear in 'Pointers' to enforce determinism.
392 for (unsigned I = 0; I < Pointers.size(); ++I) {
393 // We've seen this pointer before, and therefore already processed
394 // its equivalence class.
398 MemoryDepChecker::MemAccessInfo Access(Pointers[I].PointerValue,
399 Pointers[I].IsWritePtr);
401 SmallVector<CheckingPtrGroup, 2> Groups;
402 auto LeaderI = DepCands.findValue(DepCands.getLeaderValue(Access));
404 // Because DepCands is constructed by visiting accesses in the order in
405 // which they appear in alias sets (which is deterministic) and the
406 // iteration order within an equivalence class member is only dependent on
407 // the order in which unions and insertions are performed on the
408 // equivalence class, the iteration order is deterministic.
409 for (auto MI = DepCands.member_begin(LeaderI), ME = DepCands.member_end();
411 unsigned Pointer = PositionMap[MI->getPointer()];
413 // Mark this pointer as seen.
414 Seen.insert(Pointer);
416 // Go through all the existing sets and see if we can find one
417 // which can include this pointer.
418 for (CheckingPtrGroup &Group : Groups) {
419 // Don't perform more than a certain amount of comparisons.
420 // This should limit the cost of grouping the pointers to something
421 // reasonable. If we do end up hitting this threshold, the algorithm
422 // will create separate groups for all remaining pointers.
423 if (TotalComparisons > MemoryCheckMergeThreshold)
428 if (Group.addPointer(Pointer)) {
435 // We couldn't add this pointer to any existing set or the threshold
436 // for the number of comparisons has been reached. Create a new group
437 // to hold the current pointer.
438 Groups.push_back(CheckingPtrGroup(Pointer, *this));
441 // We've computed the grouped checks for this partition.
442 // Save the results and continue with the next one.
443 std::copy(Groups.begin(), Groups.end(), std::back_inserter(CheckingGroups));
447 bool RuntimePointerChecking::arePointersInSamePartition(
448 const SmallVectorImpl<int> &PtrToPartition, unsigned PtrIdx1,
450 return (PtrToPartition[PtrIdx1] != -1 &&
451 PtrToPartition[PtrIdx1] == PtrToPartition[PtrIdx2]);
454 bool RuntimePointerChecking::needsChecking(unsigned I, unsigned J) const {
455 const PointerInfo &PointerI = Pointers[I];
456 const PointerInfo &PointerJ = Pointers[J];
458 // No need to check if two readonly pointers intersect.
459 if (!PointerI.IsWritePtr && !PointerJ.IsWritePtr)
462 // Only need to check pointers between two different dependency sets.
463 if (PointerI.DependencySetId == PointerJ.DependencySetId)
466 // Only need to check pointers in the same alias set.
467 if (PointerI.AliasSetId != PointerJ.AliasSetId)
473 void RuntimePointerChecking::printChecks(
474 raw_ostream &OS, const SmallVectorImpl<PointerCheck> &Checks,
475 unsigned Depth) const {
477 for (const auto &Check : Checks) {
478 const auto &First = Check.first->Members, &Second = Check.second->Members;
480 OS.indent(Depth) << "Check " << N++ << ":\n";
482 OS.indent(Depth + 2) << "Comparing group (" << Check.first << "):\n";
483 for (unsigned K = 0; K < First.size(); ++K)
484 OS.indent(Depth + 2) << *Pointers[First[K]].PointerValue << "\n";
486 OS.indent(Depth + 2) << "Against group (" << Check.second << "):\n";
487 for (unsigned K = 0; K < Second.size(); ++K)
488 OS.indent(Depth + 2) << *Pointers[Second[K]].PointerValue << "\n";
492 void RuntimePointerChecking::print(raw_ostream &OS, unsigned Depth) const {
494 OS.indent(Depth) << "Run-time memory checks:\n";
495 printChecks(OS, Checks, Depth);
497 OS.indent(Depth) << "Grouped accesses:\n";
498 for (unsigned I = 0; I < CheckingGroups.size(); ++I) {
499 const auto &CG = CheckingGroups[I];
501 OS.indent(Depth + 2) << "Group " << &CG << ":\n";
502 OS.indent(Depth + 4) << "(Low: " << *CG.Low << " High: " << *CG.High
504 for (unsigned J = 0; J < CG.Members.size(); ++J) {
505 OS.indent(Depth + 6) << "Member: " << *Pointers[CG.Members[J]].Expr
513 /// \brief Analyses memory accesses in a loop.
515 /// Checks whether run time pointer checks are needed and builds sets for data
516 /// dependence checking.
517 class AccessAnalysis {
519 /// \brief Read or write access location.
520 typedef PointerIntPair<Value *, 1, bool> MemAccessInfo;
521 typedef SmallPtrSet<MemAccessInfo, 8> MemAccessInfoSet;
523 AccessAnalysis(const DataLayout &Dl, AliasAnalysis *AA, LoopInfo *LI,
524 MemoryDepChecker::DepCandidates &DA,
525 PredicatedScalarEvolution &PSE)
526 : DL(Dl), AST(*AA), LI(LI), DepCands(DA), IsRTCheckAnalysisNeeded(false),
529 /// \brief Register a load and whether it is only read from.
530 void addLoad(MemoryLocation &Loc, bool IsReadOnly) {
531 Value *Ptr = const_cast<Value*>(Loc.Ptr);
532 AST.add(Ptr, MemoryLocation::UnknownSize, Loc.AATags);
533 Accesses.insert(MemAccessInfo(Ptr, false));
535 ReadOnlyPtr.insert(Ptr);
538 /// \brief Register a store.
539 void addStore(MemoryLocation &Loc) {
540 Value *Ptr = const_cast<Value*>(Loc.Ptr);
541 AST.add(Ptr, MemoryLocation::UnknownSize, Loc.AATags);
542 Accesses.insert(MemAccessInfo(Ptr, true));
545 /// \brief Check whether we can check the pointers at runtime for
546 /// non-intersection.
548 /// Returns true if we need no check or if we do and we can generate them
549 /// (i.e. the pointers have computable bounds).
550 bool canCheckPtrAtRT(RuntimePointerChecking &RtCheck, ScalarEvolution *SE,
551 Loop *TheLoop, const ValueToValueMap &Strides,
552 bool ShouldCheckWrap = false);
554 /// \brief Goes over all memory accesses, checks whether a RT check is needed
555 /// and builds sets of dependent accesses.
556 void buildDependenceSets() {
557 processMemAccesses();
560 /// \brief Initial processing of memory accesses determined that we need to
561 /// perform dependency checking.
563 /// Note that this can later be cleared if we retry memcheck analysis without
564 /// dependency checking (i.e. ShouldRetryWithRuntimeCheck).
565 bool isDependencyCheckNeeded() { return !CheckDeps.empty(); }
567 /// We decided that no dependence analysis would be used. Reset the state.
568 void resetDepChecks(MemoryDepChecker &DepChecker) {
570 DepChecker.clearDependences();
573 MemAccessInfoSet &getDependenciesToCheck() { return CheckDeps; }
576 typedef SetVector<MemAccessInfo> PtrAccessSet;
578 /// \brief Go over all memory access and check whether runtime pointer checks
579 /// are needed and build sets of dependency check candidates.
580 void processMemAccesses();
582 /// Set of all accesses.
583 PtrAccessSet Accesses;
585 const DataLayout &DL;
587 /// Set of accesses that need a further dependence check.
588 MemAccessInfoSet CheckDeps;
590 /// Set of pointers that are read only.
591 SmallPtrSet<Value*, 16> ReadOnlyPtr;
593 /// An alias set tracker to partition the access set by underlying object and
594 //intrinsic property (such as TBAA metadata).
599 /// Sets of potentially dependent accesses - members of one set share an
600 /// underlying pointer. The set "CheckDeps" identfies which sets really need a
601 /// dependence check.
602 MemoryDepChecker::DepCandidates &DepCands;
604 /// \brief Initial processing of memory accesses determined that we may need
605 /// to add memchecks. Perform the analysis to determine the necessary checks.
607 /// Note that, this is different from isDependencyCheckNeeded. When we retry
608 /// memcheck analysis without dependency checking
609 /// (i.e. ShouldRetryWithRuntimeCheck), isDependencyCheckNeeded is cleared
610 /// while this remains set if we have potentially dependent accesses.
611 bool IsRTCheckAnalysisNeeded;
613 /// The SCEV predicate containing all the SCEV-related assumptions.
614 PredicatedScalarEvolution &PSE;
617 } // end anonymous namespace
619 /// \brief Check whether a pointer can participate in a runtime bounds check.
620 static bool hasComputableBounds(PredicatedScalarEvolution &PSE,
621 const ValueToValueMap &Strides, Value *Ptr,
623 const SCEV *PtrScev = replaceSymbolicStrideSCEV(PSE, Strides, Ptr);
625 // The bounds for loop-invariant pointer is trivial.
626 if (PSE.getSE()->isLoopInvariant(PtrScev, L))
629 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
633 return AR->isAffine();
636 /// \brief Check whether a pointer address cannot wrap.
637 static bool isNoWrap(PredicatedScalarEvolution &PSE,
638 const ValueToValueMap &Strides, Value *Ptr, Loop *L) {
639 const SCEV *PtrScev = PSE.getSCEV(Ptr);
640 if (PSE.getSE()->isLoopInvariant(PtrScev, L))
643 int64_t Stride = getPtrStride(PSE, Ptr, L, Strides);
647 bool AccessAnalysis::canCheckPtrAtRT(RuntimePointerChecking &RtCheck,
648 ScalarEvolution *SE, Loop *TheLoop,
649 const ValueToValueMap &StridesMap,
650 bool ShouldCheckWrap) {
651 // Find pointers with computable bounds. We are going to use this information
652 // to place a runtime bound check.
655 bool NeedRTCheck = false;
656 if (!IsRTCheckAnalysisNeeded) return true;
658 bool IsDepCheckNeeded = isDependencyCheckNeeded();
660 // We assign a consecutive id to access from different alias sets.
661 // Accesses between different groups doesn't need to be checked.
663 for (auto &AS : AST) {
664 int NumReadPtrChecks = 0;
665 int NumWritePtrChecks = 0;
667 // We assign consecutive id to access from different dependence sets.
668 // Accesses within the same set don't need a runtime check.
669 unsigned RunningDepId = 1;
670 DenseMap<Value *, unsigned> DepSetId;
673 Value *Ptr = A.getValue();
674 bool IsWrite = Accesses.count(MemAccessInfo(Ptr, true));
675 MemAccessInfo Access(Ptr, IsWrite);
682 if (hasComputableBounds(PSE, StridesMap, Ptr, TheLoop) &&
683 // When we run after a failing dependency check we have to make sure
684 // we don't have wrapping pointers.
685 (!ShouldCheckWrap || isNoWrap(PSE, StridesMap, Ptr, TheLoop))) {
686 // The id of the dependence set.
689 if (IsDepCheckNeeded) {
690 Value *Leader = DepCands.getLeaderValue(Access).getPointer();
691 unsigned &LeaderId = DepSetId[Leader];
693 LeaderId = RunningDepId++;
696 // Each access has its own dependence set.
697 DepId = RunningDepId++;
699 RtCheck.insert(TheLoop, Ptr, IsWrite, DepId, ASId, StridesMap, PSE);
701 DEBUG(dbgs() << "LAA: Found a runtime check ptr:" << *Ptr << '\n');
703 DEBUG(dbgs() << "LAA: Can't find bounds for ptr:" << *Ptr << '\n');
708 // If we have at least two writes or one write and a read then we need to
709 // check them. But there is no need to checks if there is only one
710 // dependence set for this alias set.
712 // Note that this function computes CanDoRT and NeedRTCheck independently.
713 // For example CanDoRT=false, NeedRTCheck=false means that we have a pointer
714 // for which we couldn't find the bounds but we don't actually need to emit
715 // any checks so it does not matter.
716 if (!(IsDepCheckNeeded && CanDoRT && RunningDepId == 2))
717 NeedRTCheck |= (NumWritePtrChecks >= 2 || (NumReadPtrChecks >= 1 &&
718 NumWritePtrChecks >= 1));
723 // If the pointers that we would use for the bounds comparison have different
724 // address spaces, assume the values aren't directly comparable, so we can't
725 // use them for the runtime check. We also have to assume they could
726 // overlap. In the future there should be metadata for whether address spaces
728 unsigned NumPointers = RtCheck.Pointers.size();
729 for (unsigned i = 0; i < NumPointers; ++i) {
730 for (unsigned j = i + 1; j < NumPointers; ++j) {
731 // Only need to check pointers between two different dependency sets.
732 if (RtCheck.Pointers[i].DependencySetId ==
733 RtCheck.Pointers[j].DependencySetId)
735 // Only need to check pointers in the same alias set.
736 if (RtCheck.Pointers[i].AliasSetId != RtCheck.Pointers[j].AliasSetId)
739 Value *PtrI = RtCheck.Pointers[i].PointerValue;
740 Value *PtrJ = RtCheck.Pointers[j].PointerValue;
742 unsigned ASi = PtrI->getType()->getPointerAddressSpace();
743 unsigned ASj = PtrJ->getType()->getPointerAddressSpace();
745 DEBUG(dbgs() << "LAA: Runtime check would require comparison between"
746 " different address spaces\n");
752 if (NeedRTCheck && CanDoRT)
753 RtCheck.generateChecks(DepCands, IsDepCheckNeeded);
755 DEBUG(dbgs() << "LAA: We need to do " << RtCheck.getNumberOfChecks()
756 << " pointer comparisons.\n");
758 RtCheck.Need = NeedRTCheck;
760 bool CanDoRTIfNeeded = !NeedRTCheck || CanDoRT;
761 if (!CanDoRTIfNeeded)
763 return CanDoRTIfNeeded;
766 void AccessAnalysis::processMemAccesses() {
767 // We process the set twice: first we process read-write pointers, last we
768 // process read-only pointers. This allows us to skip dependence tests for
769 // read-only pointers.
771 DEBUG(dbgs() << "LAA: Processing memory accesses...\n");
772 DEBUG(dbgs() << " AST: "; AST.dump());
773 DEBUG(dbgs() << "LAA: Accesses(" << Accesses.size() << "):\n");
775 for (auto A : Accesses)
776 dbgs() << "\t" << *A.getPointer() << " (" <<
777 (A.getInt() ? "write" : (ReadOnlyPtr.count(A.getPointer()) ?
778 "read-only" : "read")) << ")\n";
781 // The AliasSetTracker has nicely partitioned our pointers by metadata
782 // compatibility and potential for underlying-object overlap. As a result, we
783 // only need to check for potential pointer dependencies within each alias
785 for (auto &AS : AST) {
786 // Note that both the alias-set tracker and the alias sets themselves used
787 // linked lists internally and so the iteration order here is deterministic
788 // (matching the original instruction order within each set).
790 bool SetHasWrite = false;
792 // Map of pointers to last access encountered.
793 typedef DenseMap<Value*, MemAccessInfo> UnderlyingObjToAccessMap;
794 UnderlyingObjToAccessMap ObjToLastAccess;
796 // Set of access to check after all writes have been processed.
797 PtrAccessSet DeferredAccesses;
799 // Iterate over each alias set twice, once to process read/write pointers,
800 // and then to process read-only pointers.
801 for (int SetIteration = 0; SetIteration < 2; ++SetIteration) {
802 bool UseDeferred = SetIteration > 0;
803 PtrAccessSet &S = UseDeferred ? DeferredAccesses : Accesses;
806 Value *Ptr = AV.getValue();
808 // For a single memory access in AliasSetTracker, Accesses may contain
809 // both read and write, and they both need to be handled for CheckDeps.
811 if (AC.getPointer() != Ptr)
814 bool IsWrite = AC.getInt();
816 // If we're using the deferred access set, then it contains only
818 bool IsReadOnlyPtr = ReadOnlyPtr.count(Ptr) && !IsWrite;
819 if (UseDeferred && !IsReadOnlyPtr)
821 // Otherwise, the pointer must be in the PtrAccessSet, either as a
823 assert(((IsReadOnlyPtr && UseDeferred) || IsWrite ||
824 S.count(MemAccessInfo(Ptr, false))) &&
825 "Alias-set pointer not in the access set?");
827 MemAccessInfo Access(Ptr, IsWrite);
828 DepCands.insert(Access);
830 // Memorize read-only pointers for later processing and skip them in
831 // the first round (they need to be checked after we have seen all
832 // write pointers). Note: we also mark pointer that are not
833 // consecutive as "read-only" pointers (so that we check
834 // "a[b[i]] +="). Hence, we need the second check for "!IsWrite".
835 if (!UseDeferred && IsReadOnlyPtr) {
836 DeferredAccesses.insert(Access);
840 // If this is a write - check other reads and writes for conflicts. If
841 // this is a read only check other writes for conflicts (but only if
842 // there is no other write to the ptr - this is an optimization to
843 // catch "a[i] = a[i] + " without having to do a dependence check).
844 if ((IsWrite || IsReadOnlyPtr) && SetHasWrite) {
845 CheckDeps.insert(Access);
846 IsRTCheckAnalysisNeeded = true;
852 // Create sets of pointers connected by a shared alias set and
853 // underlying object.
854 typedef SmallVector<Value *, 16> ValueVector;
855 ValueVector TempObjects;
857 GetUnderlyingObjects(Ptr, TempObjects, DL, LI);
858 DEBUG(dbgs() << "Underlying objects for pointer " << *Ptr << "\n");
859 for (Value *UnderlyingObj : TempObjects) {
860 // nullptr never alias, don't join sets for pointer that have "null"
861 // in their UnderlyingObjects list.
862 if (isa<ConstantPointerNull>(UnderlyingObj))
865 UnderlyingObjToAccessMap::iterator Prev =
866 ObjToLastAccess.find(UnderlyingObj);
867 if (Prev != ObjToLastAccess.end())
868 DepCands.unionSets(Access, Prev->second);
870 ObjToLastAccess[UnderlyingObj] = Access;
871 DEBUG(dbgs() << " " << *UnderlyingObj << "\n");
879 static bool isInBoundsGep(Value *Ptr) {
880 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr))
881 return GEP->isInBounds();
885 /// \brief Return true if an AddRec pointer \p Ptr is unsigned non-wrapping,
886 /// i.e. monotonically increasing/decreasing.
887 static bool isNoWrapAddRec(Value *Ptr, const SCEVAddRecExpr *AR,
888 PredicatedScalarEvolution &PSE, const Loop *L) {
889 // FIXME: This should probably only return true for NUW.
890 if (AR->getNoWrapFlags(SCEV::NoWrapMask))
893 // Scalar evolution does not propagate the non-wrapping flags to values that
894 // are derived from a non-wrapping induction variable because non-wrapping
895 // could be flow-sensitive.
897 // Look through the potentially overflowing instruction to try to prove
898 // non-wrapping for the *specific* value of Ptr.
900 // The arithmetic implied by an inbounds GEP can't overflow.
901 auto *GEP = dyn_cast<GetElementPtrInst>(Ptr);
902 if (!GEP || !GEP->isInBounds())
905 // Make sure there is only one non-const index and analyze that.
906 Value *NonConstIndex = nullptr;
907 for (Value *Index : make_range(GEP->idx_begin(), GEP->idx_end()))
908 if (!isa<ConstantInt>(Index)) {
911 NonConstIndex = Index;
914 // The recurrence is on the pointer, ignore for now.
917 // The index in GEP is signed. It is non-wrapping if it's derived from a NSW
918 // AddRec using a NSW operation.
919 if (auto *OBO = dyn_cast<OverflowingBinaryOperator>(NonConstIndex))
920 if (OBO->hasNoSignedWrap() &&
921 // Assume constant for other the operand so that the AddRec can be
923 isa<ConstantInt>(OBO->getOperand(1))) {
924 auto *OpScev = PSE.getSCEV(OBO->getOperand(0));
926 if (auto *OpAR = dyn_cast<SCEVAddRecExpr>(OpScev))
927 return OpAR->getLoop() == L && OpAR->getNoWrapFlags(SCEV::FlagNSW);
933 /// \brief Check whether the access through \p Ptr has a constant stride.
934 int64_t llvm::getPtrStride(PredicatedScalarEvolution &PSE, Value *Ptr,
935 const Loop *Lp, const ValueToValueMap &StridesMap,
936 bool Assume, bool ShouldCheckWrap) {
937 Type *Ty = Ptr->getType();
938 assert(Ty->isPointerTy() && "Unexpected non-ptr");
940 // Make sure that the pointer does not point to aggregate types.
941 auto *PtrTy = cast<PointerType>(Ty);
942 if (PtrTy->getElementType()->isAggregateType()) {
943 DEBUG(dbgs() << "LAA: Bad stride - Not a pointer to a scalar type" << *Ptr
948 const SCEV *PtrScev = replaceSymbolicStrideSCEV(PSE, StridesMap, Ptr);
950 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
952 AR = PSE.getAsAddRec(Ptr);
955 DEBUG(dbgs() << "LAA: Bad stride - Not an AddRecExpr pointer " << *Ptr
956 << " SCEV: " << *PtrScev << "\n");
960 // The accesss function must stride over the innermost loop.
961 if (Lp != AR->getLoop()) {
962 DEBUG(dbgs() << "LAA: Bad stride - Not striding over innermost loop " <<
963 *Ptr << " SCEV: " << *AR << "\n");
967 // The address calculation must not wrap. Otherwise, a dependence could be
969 // An inbounds getelementptr that is a AddRec with a unit stride
970 // cannot wrap per definition. The unit stride requirement is checked later.
971 // An getelementptr without an inbounds attribute and unit stride would have
972 // to access the pointer value "0" which is undefined behavior in address
973 // space 0, therefore we can also vectorize this case.
974 bool IsInBoundsGEP = isInBoundsGep(Ptr);
975 bool IsNoWrapAddRec = !ShouldCheckWrap ||
976 PSE.hasNoOverflow(Ptr, SCEVWrapPredicate::IncrementNUSW) ||
977 isNoWrapAddRec(Ptr, AR, PSE, Lp);
978 bool IsInAddressSpaceZero = PtrTy->getAddressSpace() == 0;
979 if (!IsNoWrapAddRec && !IsInBoundsGEP && !IsInAddressSpaceZero) {
981 PSE.setNoOverflow(Ptr, SCEVWrapPredicate::IncrementNUSW);
982 IsNoWrapAddRec = true;
983 DEBUG(dbgs() << "LAA: Pointer may wrap in the address space:\n"
984 << "LAA: Pointer: " << *Ptr << "\n"
985 << "LAA: SCEV: " << *AR << "\n"
986 << "LAA: Added an overflow assumption\n");
988 DEBUG(dbgs() << "LAA: Bad stride - Pointer may wrap in the address space "
989 << *Ptr << " SCEV: " << *AR << "\n");
994 // Check the step is constant.
995 const SCEV *Step = AR->getStepRecurrence(*PSE.getSE());
997 // Calculate the pointer stride and check if it is constant.
998 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
1000 DEBUG(dbgs() << "LAA: Bad stride - Not a constant strided " << *Ptr <<
1001 " SCEV: " << *AR << "\n");
1005 auto &DL = Lp->getHeader()->getModule()->getDataLayout();
1006 int64_t Size = DL.getTypeAllocSize(PtrTy->getElementType());
1007 const APInt &APStepVal = C->getAPInt();
1009 // Huge step value - give up.
1010 if (APStepVal.getBitWidth() > 64)
1013 int64_t StepVal = APStepVal.getSExtValue();
1016 int64_t Stride = StepVal / Size;
1017 int64_t Rem = StepVal % Size;
1021 // If the SCEV could wrap but we have an inbounds gep with a unit stride we
1022 // know we can't "wrap around the address space". In case of address space
1023 // zero we know that this won't happen without triggering undefined behavior.
1024 if (!IsNoWrapAddRec && (IsInBoundsGEP || IsInAddressSpaceZero) &&
1025 Stride != 1 && Stride != -1) {
1027 // We can avoid this case by adding a run-time check.
1028 DEBUG(dbgs() << "LAA: Non unit strided pointer which is not either "
1029 << "inbouds or in address space 0 may wrap:\n"
1030 << "LAA: Pointer: " << *Ptr << "\n"
1031 << "LAA: SCEV: " << *AR << "\n"
1032 << "LAA: Added an overflow assumption\n");
1033 PSE.setNoOverflow(Ptr, SCEVWrapPredicate::IncrementNUSW);
1041 /// Take the pointer operand from the Load/Store instruction.
1042 /// Returns NULL if this is not a valid Load/Store instruction.
1043 static Value *getPointerOperand(Value *I) {
1044 if (auto *LI = dyn_cast<LoadInst>(I))
1045 return LI->getPointerOperand();
1046 if (auto *SI = dyn_cast<StoreInst>(I))
1047 return SI->getPointerOperand();
1051 /// Take the address space operand from the Load/Store instruction.
1052 /// Returns -1 if this is not a valid Load/Store instruction.
1053 static unsigned getAddressSpaceOperand(Value *I) {
1054 if (LoadInst *L = dyn_cast<LoadInst>(I))
1055 return L->getPointerAddressSpace();
1056 if (StoreInst *S = dyn_cast<StoreInst>(I))
1057 return S->getPointerAddressSpace();
1061 /// Returns true if the memory operations \p A and \p B are consecutive.
1062 bool llvm::isConsecutiveAccess(Value *A, Value *B, const DataLayout &DL,
1063 ScalarEvolution &SE, bool CheckType) {
1064 Value *PtrA = getPointerOperand(A);
1065 Value *PtrB = getPointerOperand(B);
1066 unsigned ASA = getAddressSpaceOperand(A);
1067 unsigned ASB = getAddressSpaceOperand(B);
1069 // Check that the address spaces match and that the pointers are valid.
1070 if (!PtrA || !PtrB || (ASA != ASB))
1073 // Make sure that A and B are different pointers.
1077 // Make sure that A and B have the same type if required.
1078 if (CheckType && PtrA->getType() != PtrB->getType())
1081 unsigned PtrBitWidth = DL.getPointerSizeInBits(ASA);
1082 Type *Ty = cast<PointerType>(PtrA->getType())->getElementType();
1083 APInt Size(PtrBitWidth, DL.getTypeStoreSize(Ty));
1085 APInt OffsetA(PtrBitWidth, 0), OffsetB(PtrBitWidth, 0);
1086 PtrA = PtrA->stripAndAccumulateInBoundsConstantOffsets(DL, OffsetA);
1087 PtrB = PtrB->stripAndAccumulateInBoundsConstantOffsets(DL, OffsetB);
1089 // OffsetDelta = OffsetB - OffsetA;
1090 const SCEV *OffsetSCEVA = SE.getConstant(OffsetA);
1091 const SCEV *OffsetSCEVB = SE.getConstant(OffsetB);
1092 const SCEV *OffsetDeltaSCEV = SE.getMinusSCEV(OffsetSCEVB, OffsetSCEVA);
1093 const SCEVConstant *OffsetDeltaC = dyn_cast<SCEVConstant>(OffsetDeltaSCEV);
1094 const APInt &OffsetDelta = OffsetDeltaC->getAPInt();
1095 // Check if they are based on the same pointer. That makes the offsets
1098 return OffsetDelta == Size;
1100 // Compute the necessary base pointer delta to have the necessary final delta
1101 // equal to the size.
1102 // BaseDelta = Size - OffsetDelta;
1103 const SCEV *SizeSCEV = SE.getConstant(Size);
1104 const SCEV *BaseDelta = SE.getMinusSCEV(SizeSCEV, OffsetDeltaSCEV);
1106 // Otherwise compute the distance with SCEV between the base pointers.
1107 const SCEV *PtrSCEVA = SE.getSCEV(PtrA);
1108 const SCEV *PtrSCEVB = SE.getSCEV(PtrB);
1109 const SCEV *X = SE.getAddExpr(PtrSCEVA, BaseDelta);
1110 return X == PtrSCEVB;
1113 bool MemoryDepChecker::Dependence::isSafeForVectorization(DepType Type) {
1117 case BackwardVectorizable:
1121 case ForwardButPreventsForwarding:
1123 case BackwardVectorizableButPreventsForwarding:
1126 llvm_unreachable("unexpected DepType!");
1129 bool MemoryDepChecker::Dependence::isBackward() const {
1133 case ForwardButPreventsForwarding:
1137 case BackwardVectorizable:
1139 case BackwardVectorizableButPreventsForwarding:
1142 llvm_unreachable("unexpected DepType!");
1145 bool MemoryDepChecker::Dependence::isPossiblyBackward() const {
1146 return isBackward() || Type == Unknown;
1149 bool MemoryDepChecker::Dependence::isForward() const {
1152 case ForwardButPreventsForwarding:
1157 case BackwardVectorizable:
1159 case BackwardVectorizableButPreventsForwarding:
1162 llvm_unreachable("unexpected DepType!");
1165 bool MemoryDepChecker::couldPreventStoreLoadForward(uint64_t Distance,
1166 uint64_t TypeByteSize) {
1167 // If loads occur at a distance that is not a multiple of a feasible vector
1168 // factor store-load forwarding does not take place.
1169 // Positive dependences might cause troubles because vectorizing them might
1170 // prevent store-load forwarding making vectorized code run a lot slower.
1171 // a[i] = a[i-3] ^ a[i-8];
1172 // The stores to a[i:i+1] don't align with the stores to a[i-3:i-2] and
1173 // hence on your typical architecture store-load forwarding does not take
1174 // place. Vectorizing in such cases does not make sense.
1175 // Store-load forwarding distance.
1177 // After this many iterations store-to-load forwarding conflicts should not
1178 // cause any slowdowns.
1179 const uint64_t NumItersForStoreLoadThroughMemory = 8 * TypeByteSize;
1180 // Maximum vector factor.
1181 uint64_t MaxVFWithoutSLForwardIssues = std::min(
1182 VectorizerParams::MaxVectorWidth * TypeByteSize, MaxSafeDepDistBytes);
1184 // Compute the smallest VF at which the store and load would be misaligned.
1185 for (uint64_t VF = 2 * TypeByteSize; VF <= MaxVFWithoutSLForwardIssues;
1187 // If the number of vector iteration between the store and the load are
1188 // small we could incur conflicts.
1189 if (Distance % VF && Distance / VF < NumItersForStoreLoadThroughMemory) {
1190 MaxVFWithoutSLForwardIssues = (VF >>= 1);
1195 if (MaxVFWithoutSLForwardIssues < 2 * TypeByteSize) {
1196 DEBUG(dbgs() << "LAA: Distance " << Distance
1197 << " that could cause a store-load forwarding conflict\n");
1201 if (MaxVFWithoutSLForwardIssues < MaxSafeDepDistBytes &&
1202 MaxVFWithoutSLForwardIssues !=
1203 VectorizerParams::MaxVectorWidth * TypeByteSize)
1204 MaxSafeDepDistBytes = MaxVFWithoutSLForwardIssues;
1208 /// \brief Check the dependence for two accesses with the same stride \p Stride.
1209 /// \p Distance is the positive distance and \p TypeByteSize is type size in
1212 /// \returns true if they are independent.
1213 static bool areStridedAccessesIndependent(uint64_t Distance, uint64_t Stride,
1214 uint64_t TypeByteSize) {
1215 assert(Stride > 1 && "The stride must be greater than 1");
1216 assert(TypeByteSize > 0 && "The type size in byte must be non-zero");
1217 assert(Distance > 0 && "The distance must be non-zero");
1219 // Skip if the distance is not multiple of type byte size.
1220 if (Distance % TypeByteSize)
1223 uint64_t ScaledDist = Distance / TypeByteSize;
1225 // No dependence if the scaled distance is not multiple of the stride.
1227 // for (i = 0; i < 1024 ; i += 4)
1228 // A[i+2] = A[i] + 1;
1230 // Two accesses in memory (scaled distance is 2, stride is 4):
1231 // | A[0] | | | | A[4] | | | |
1232 // | | | A[2] | | | | A[6] | |
1235 // for (i = 0; i < 1024 ; i += 3)
1236 // A[i+4] = A[i] + 1;
1238 // Two accesses in memory (scaled distance is 4, stride is 3):
1239 // | A[0] | | | A[3] | | | A[6] | | |
1240 // | | | | | A[4] | | | A[7] | |
1241 return ScaledDist % Stride;
1244 MemoryDepChecker::Dependence::DepType
1245 MemoryDepChecker::isDependent(const MemAccessInfo &A, unsigned AIdx,
1246 const MemAccessInfo &B, unsigned BIdx,
1247 const ValueToValueMap &Strides) {
1248 assert (AIdx < BIdx && "Must pass arguments in program order");
1250 Value *APtr = A.getPointer();
1251 Value *BPtr = B.getPointer();
1252 bool AIsWrite = A.getInt();
1253 bool BIsWrite = B.getInt();
1255 // Two reads are independent.
1256 if (!AIsWrite && !BIsWrite)
1257 return Dependence::NoDep;
1259 // We cannot check pointers in different address spaces.
1260 if (APtr->getType()->getPointerAddressSpace() !=
1261 BPtr->getType()->getPointerAddressSpace())
1262 return Dependence::Unknown;
1264 int64_t StrideAPtr = getPtrStride(PSE, APtr, InnermostLoop, Strides, true);
1265 int64_t StrideBPtr = getPtrStride(PSE, BPtr, InnermostLoop, Strides, true);
1267 const SCEV *Src = PSE.getSCEV(APtr);
1268 const SCEV *Sink = PSE.getSCEV(BPtr);
1270 // If the induction step is negative we have to invert source and sink of the
1272 if (StrideAPtr < 0) {
1273 std::swap(APtr, BPtr);
1274 std::swap(Src, Sink);
1275 std::swap(AIsWrite, BIsWrite);
1276 std::swap(AIdx, BIdx);
1277 std::swap(StrideAPtr, StrideBPtr);
1280 const SCEV *Dist = PSE.getSE()->getMinusSCEV(Sink, Src);
1282 DEBUG(dbgs() << "LAA: Src Scev: " << *Src << "Sink Scev: " << *Sink
1283 << "(Induction step: " << StrideAPtr << ")\n");
1284 DEBUG(dbgs() << "LAA: Distance for " << *InstMap[AIdx] << " to "
1285 << *InstMap[BIdx] << ": " << *Dist << "\n");
1287 // Need accesses with constant stride. We don't want to vectorize
1288 // "A[B[i]] += ..." and similar code or pointer arithmetic that could wrap in
1289 // the address space.
1290 if (!StrideAPtr || !StrideBPtr || StrideAPtr != StrideBPtr){
1291 DEBUG(dbgs() << "Pointer access with non-constant stride\n");
1292 return Dependence::Unknown;
1295 const SCEVConstant *C = dyn_cast<SCEVConstant>(Dist);
1297 DEBUG(dbgs() << "LAA: Dependence because of non-constant distance\n");
1298 ShouldRetryWithRuntimeCheck = true;
1299 return Dependence::Unknown;
1302 Type *ATy = APtr->getType()->getPointerElementType();
1303 Type *BTy = BPtr->getType()->getPointerElementType();
1304 auto &DL = InnermostLoop->getHeader()->getModule()->getDataLayout();
1305 uint64_t TypeByteSize = DL.getTypeAllocSize(ATy);
1307 const APInt &Val = C->getAPInt();
1308 int64_t Distance = Val.getSExtValue();
1309 uint64_t Stride = std::abs(StrideAPtr);
1311 // Attempt to prove strided accesses independent.
1312 if (std::abs(Distance) > 0 && Stride > 1 && ATy == BTy &&
1313 areStridedAccessesIndependent(std::abs(Distance), Stride, TypeByteSize)) {
1314 DEBUG(dbgs() << "LAA: Strided accesses are independent\n");
1315 return Dependence::NoDep;
1318 // Negative distances are not plausible dependencies.
1319 if (Val.isNegative()) {
1320 bool IsTrueDataDependence = (AIsWrite && !BIsWrite);
1321 if (IsTrueDataDependence && EnableForwardingConflictDetection &&
1322 (couldPreventStoreLoadForward(Val.abs().getZExtValue(), TypeByteSize) ||
1324 DEBUG(dbgs() << "LAA: Forward but may prevent st->ld forwarding\n");
1325 return Dependence::ForwardButPreventsForwarding;
1328 DEBUG(dbgs() << "LAA: Dependence is negative\n");
1329 return Dependence::Forward;
1332 // Write to the same location with the same size.
1333 // Could be improved to assert type sizes are the same (i32 == float, etc).
1336 return Dependence::Forward;
1337 DEBUG(dbgs() << "LAA: Zero dependence difference but different types\n");
1338 return Dependence::Unknown;
1341 assert(Val.isStrictlyPositive() && "Expect a positive value");
1345 "LAA: ReadWrite-Write positive dependency with different types\n");
1346 return Dependence::Unknown;
1349 // Bail out early if passed-in parameters make vectorization not feasible.
1350 unsigned ForcedFactor = (VectorizerParams::VectorizationFactor ?
1351 VectorizerParams::VectorizationFactor : 1);
1352 unsigned ForcedUnroll = (VectorizerParams::VectorizationInterleave ?
1353 VectorizerParams::VectorizationInterleave : 1);
1354 // The minimum number of iterations for a vectorized/unrolled version.
1355 unsigned MinNumIter = std::max(ForcedFactor * ForcedUnroll, 2U);
1357 // It's not vectorizable if the distance is smaller than the minimum distance
1358 // needed for a vectroized/unrolled version. Vectorizing one iteration in
1359 // front needs TypeByteSize * Stride. Vectorizing the last iteration needs
1360 // TypeByteSize (No need to plus the last gap distance).
1362 // E.g. Assume one char is 1 byte in memory and one int is 4 bytes.
1364 // int *B = (int *)((char *)A + 14);
1365 // for (i = 0 ; i < 1024 ; i += 2)
1369 // Two accesses in memory (stride is 2):
1370 // | A[0] | | A[2] | | A[4] | | A[6] | |
1371 // | B[0] | | B[2] | | B[4] |
1373 // Distance needs for vectorizing iterations except the last iteration:
1374 // 4 * 2 * (MinNumIter - 1). Distance needs for the last iteration: 4.
1375 // So the minimum distance needed is: 4 * 2 * (MinNumIter - 1) + 4.
1377 // If MinNumIter is 2, it is vectorizable as the minimum distance needed is
1378 // 12, which is less than distance.
1380 // If MinNumIter is 4 (Say if a user forces the vectorization factor to be 4),
1381 // the minimum distance needed is 28, which is greater than distance. It is
1382 // not safe to do vectorization.
1383 uint64_t MinDistanceNeeded =
1384 TypeByteSize * Stride * (MinNumIter - 1) + TypeByteSize;
1385 if (MinDistanceNeeded > static_cast<uint64_t>(Distance)) {
1386 DEBUG(dbgs() << "LAA: Failure because of positive distance " << Distance
1388 return Dependence::Backward;
1391 // Unsafe if the minimum distance needed is greater than max safe distance.
1392 if (MinDistanceNeeded > MaxSafeDepDistBytes) {
1393 DEBUG(dbgs() << "LAA: Failure because it needs at least "
1394 << MinDistanceNeeded << " size in bytes");
1395 return Dependence::Backward;
1398 // Positive distance bigger than max vectorization factor.
1399 // FIXME: Should use max factor instead of max distance in bytes, which could
1400 // not handle different types.
1401 // E.g. Assume one char is 1 byte in memory and one int is 4 bytes.
1402 // void foo (int *A, char *B) {
1403 // for (unsigned i = 0; i < 1024; i++) {
1404 // A[i+2] = A[i] + 1;
1405 // B[i+2] = B[i] + 1;
1409 // This case is currently unsafe according to the max safe distance. If we
1410 // analyze the two accesses on array B, the max safe dependence distance
1411 // is 2. Then we analyze the accesses on array A, the minimum distance needed
1412 // is 8, which is less than 2 and forbidden vectorization, But actually
1413 // both A and B could be vectorized by 2 iterations.
1414 MaxSafeDepDistBytes =
1415 std::min(static_cast<uint64_t>(Distance), MaxSafeDepDistBytes);
1417 bool IsTrueDataDependence = (!AIsWrite && BIsWrite);
1418 if (IsTrueDataDependence && EnableForwardingConflictDetection &&
1419 couldPreventStoreLoadForward(Distance, TypeByteSize))
1420 return Dependence::BackwardVectorizableButPreventsForwarding;
1422 DEBUG(dbgs() << "LAA: Positive distance " << Val.getSExtValue()
1423 << " with max VF = "
1424 << MaxSafeDepDistBytes / (TypeByteSize * Stride) << '\n');
1426 return Dependence::BackwardVectorizable;
1429 bool MemoryDepChecker::areDepsSafe(DepCandidates &AccessSets,
1430 MemAccessInfoSet &CheckDeps,
1431 const ValueToValueMap &Strides) {
1433 MaxSafeDepDistBytes = -1;
1434 while (!CheckDeps.empty()) {
1435 MemAccessInfo CurAccess = *CheckDeps.begin();
1437 // Get the relevant memory access set.
1438 EquivalenceClasses<MemAccessInfo>::iterator I =
1439 AccessSets.findValue(AccessSets.getLeaderValue(CurAccess));
1441 // Check accesses within this set.
1442 EquivalenceClasses<MemAccessInfo>::member_iterator AI =
1443 AccessSets.member_begin(I);
1444 EquivalenceClasses<MemAccessInfo>::member_iterator AE =
1445 AccessSets.member_end();
1447 // Check every access pair.
1449 CheckDeps.erase(*AI);
1450 EquivalenceClasses<MemAccessInfo>::member_iterator OI = std::next(AI);
1452 // Check every accessing instruction pair in program order.
1453 for (std::vector<unsigned>::iterator I1 = Accesses[*AI].begin(),
1454 I1E = Accesses[*AI].end(); I1 != I1E; ++I1)
1455 for (std::vector<unsigned>::iterator I2 = Accesses[*OI].begin(),
1456 I2E = Accesses[*OI].end(); I2 != I2E; ++I2) {
1457 auto A = std::make_pair(&*AI, *I1);
1458 auto B = std::make_pair(&*OI, *I2);
1464 Dependence::DepType Type =
1465 isDependent(*A.first, A.second, *B.first, B.second, Strides);
1466 SafeForVectorization &= Dependence::isSafeForVectorization(Type);
1468 // Gather dependences unless we accumulated MaxDependences
1469 // dependences. In that case return as soon as we find the first
1470 // unsafe dependence. This puts a limit on this quadratic
1472 if (RecordDependences) {
1473 if (Type != Dependence::NoDep)
1474 Dependences.push_back(Dependence(A.second, B.second, Type));
1476 if (Dependences.size() >= MaxDependences) {
1477 RecordDependences = false;
1478 Dependences.clear();
1479 DEBUG(dbgs() << "Too many dependences, stopped recording\n");
1482 if (!RecordDependences && !SafeForVectorization)
1491 DEBUG(dbgs() << "Total Dependences: " << Dependences.size() << "\n");
1492 return SafeForVectorization;
1495 SmallVector<Instruction *, 4>
1496 MemoryDepChecker::getInstructionsForAccess(Value *Ptr, bool isWrite) const {
1497 MemAccessInfo Access(Ptr, isWrite);
1498 auto &IndexVector = Accesses.find(Access)->second;
1500 SmallVector<Instruction *, 4> Insts;
1501 transform(IndexVector,
1502 std::back_inserter(Insts),
1503 [&](unsigned Idx) { return this->InstMap[Idx]; });
1507 const char *MemoryDepChecker::Dependence::DepName[] = {
1508 "NoDep", "Unknown", "Forward", "ForwardButPreventsForwarding", "Backward",
1509 "BackwardVectorizable", "BackwardVectorizableButPreventsForwarding"};
1511 void MemoryDepChecker::Dependence::print(
1512 raw_ostream &OS, unsigned Depth,
1513 const SmallVectorImpl<Instruction *> &Instrs) const {
1514 OS.indent(Depth) << DepName[Type] << ":\n";
1515 OS.indent(Depth + 2) << *Instrs[Source] << " -> \n";
1516 OS.indent(Depth + 2) << *Instrs[Destination] << "\n";
1519 bool LoopAccessInfo::canAnalyzeLoop() {
1520 // We need to have a loop header.
1521 DEBUG(dbgs() << "LAA: Found a loop in "
1522 << TheLoop->getHeader()->getParent()->getName() << ": "
1523 << TheLoop->getHeader()->getName() << '\n');
1525 // We can only analyze innermost loops.
1526 if (!TheLoop->empty()) {
1527 DEBUG(dbgs() << "LAA: loop is not the innermost loop\n");
1528 recordAnalysis("NotInnerMostLoop") << "loop is not the innermost loop";
1532 // We must have a single backedge.
1533 if (TheLoop->getNumBackEdges() != 1) {
1534 DEBUG(dbgs() << "LAA: loop control flow is not understood by analyzer\n");
1535 recordAnalysis("CFGNotUnderstood")
1536 << "loop control flow is not understood by analyzer";
1540 // We must have a single exiting block.
1541 if (!TheLoop->getExitingBlock()) {
1542 DEBUG(dbgs() << "LAA: loop control flow is not understood by analyzer\n");
1543 recordAnalysis("CFGNotUnderstood")
1544 << "loop control flow is not understood by analyzer";
1548 // We only handle bottom-tested loops, i.e. loop in which the condition is
1549 // checked at the end of each iteration. With that we can assume that all
1550 // instructions in the loop are executed the same number of times.
1551 if (TheLoop->getExitingBlock() != TheLoop->getLoopLatch()) {
1552 DEBUG(dbgs() << "LAA: loop control flow is not understood by analyzer\n");
1553 recordAnalysis("CFGNotUnderstood")
1554 << "loop control flow is not understood by analyzer";
1558 // ScalarEvolution needs to be able to find the exit count.
1559 const SCEV *ExitCount = PSE->getBackedgeTakenCount();
1560 if (ExitCount == PSE->getSE()->getCouldNotCompute()) {
1561 recordAnalysis("CantComputeNumberOfIterations")
1562 << "could not determine number of loop iterations";
1563 DEBUG(dbgs() << "LAA: SCEV could not compute the loop exit count.\n");
1570 void LoopAccessInfo::analyzeLoop(AliasAnalysis *AA, LoopInfo *LI,
1571 const TargetLibraryInfo *TLI,
1572 DominatorTree *DT) {
1573 typedef SmallPtrSet<Value*, 16> ValueSet;
1575 // Holds the Load and Store instructions.
1576 SmallVector<LoadInst *, 16> Loads;
1577 SmallVector<StoreInst *, 16> Stores;
1579 // Holds all the different accesses in the loop.
1580 unsigned NumReads = 0;
1581 unsigned NumReadWrites = 0;
1583 PtrRtChecking->Pointers.clear();
1584 PtrRtChecking->Need = false;
1586 const bool IsAnnotatedParallel = TheLoop->isAnnotatedParallel();
1589 for (BasicBlock *BB : TheLoop->blocks()) {
1590 // Scan the BB and collect legal loads and stores.
1591 for (Instruction &I : *BB) {
1592 // If this is a load, save it. If this instruction can read from memory
1593 // but is not a load, then we quit. Notice that we don't handle function
1594 // calls that read or write.
1595 if (I.mayReadFromMemory()) {
1596 // Many math library functions read the rounding mode. We will only
1597 // vectorize a loop if it contains known function calls that don't set
1598 // the flag. Therefore, it is safe to ignore this read from memory.
1599 auto *Call = dyn_cast<CallInst>(&I);
1600 if (Call && getVectorIntrinsicIDForCall(Call, TLI))
1603 // If the function has an explicit vectorized counterpart, we can safely
1604 // assume that it can be vectorized.
1605 if (Call && !Call->isNoBuiltin() && Call->getCalledFunction() &&
1606 TLI->isFunctionVectorizable(Call->getCalledFunction()->getName()))
1609 auto *Ld = dyn_cast<LoadInst>(&I);
1610 if (!Ld || (!Ld->isSimple() && !IsAnnotatedParallel)) {
1611 recordAnalysis("NonSimpleLoad", Ld)
1612 << "read with atomic ordering or volatile read";
1613 DEBUG(dbgs() << "LAA: Found a non-simple load.\n");
1618 Loads.push_back(Ld);
1619 DepChecker->addAccess(Ld);
1620 if (EnableMemAccessVersioning)
1621 collectStridedAccess(Ld);
1625 // Save 'store' instructions. Abort if other instructions write to memory.
1626 if (I.mayWriteToMemory()) {
1627 auto *St = dyn_cast<StoreInst>(&I);
1629 recordAnalysis("CantVectorizeInstruction", St)
1630 << "instruction cannot be vectorized";
1634 if (!St->isSimple() && !IsAnnotatedParallel) {
1635 recordAnalysis("NonSimpleStore", St)
1636 << "write with atomic ordering or volatile write";
1637 DEBUG(dbgs() << "LAA: Found a non-simple store.\n");
1642 Stores.push_back(St);
1643 DepChecker->addAccess(St);
1644 if (EnableMemAccessVersioning)
1645 collectStridedAccess(St);
1650 // Now we have two lists that hold the loads and the stores.
1651 // Next, we find the pointers that they use.
1653 // Check if we see any stores. If there are no stores, then we don't
1654 // care if the pointers are *restrict*.
1655 if (!Stores.size()) {
1656 DEBUG(dbgs() << "LAA: Found a read-only loop!\n");
1661 MemoryDepChecker::DepCandidates DependentAccesses;
1662 AccessAnalysis Accesses(TheLoop->getHeader()->getModule()->getDataLayout(),
1663 AA, LI, DependentAccesses, *PSE);
1665 // Holds the analyzed pointers. We don't want to call GetUnderlyingObjects
1666 // multiple times on the same object. If the ptr is accessed twice, once
1667 // for read and once for write, it will only appear once (on the write
1668 // list). This is okay, since we are going to check for conflicts between
1669 // writes and between reads and writes, but not between reads and reads.
1672 for (StoreInst *ST : Stores) {
1673 Value *Ptr = ST->getPointerOperand();
1674 // Check for store to loop invariant address.
1675 StoreToLoopInvariantAddress |= isUniform(Ptr);
1676 // If we did *not* see this pointer before, insert it to the read-write
1677 // list. At this phase it is only a 'write' list.
1678 if (Seen.insert(Ptr).second) {
1681 MemoryLocation Loc = MemoryLocation::get(ST);
1682 // The TBAA metadata could have a control dependency on the predication
1683 // condition, so we cannot rely on it when determining whether or not we
1684 // need runtime pointer checks.
1685 if (blockNeedsPredication(ST->getParent(), TheLoop, DT))
1686 Loc.AATags.TBAA = nullptr;
1688 Accesses.addStore(Loc);
1692 if (IsAnnotatedParallel) {
1694 << "LAA: A loop annotated parallel, ignore memory dependency "
1700 for (LoadInst *LD : Loads) {
1701 Value *Ptr = LD->getPointerOperand();
1702 // If we did *not* see this pointer before, insert it to the
1703 // read list. If we *did* see it before, then it is already in
1704 // the read-write list. This allows us to vectorize expressions
1705 // such as A[i] += x; Because the address of A[i] is a read-write
1706 // pointer. This only works if the index of A[i] is consecutive.
1707 // If the address of i is unknown (for example A[B[i]]) then we may
1708 // read a few words, modify, and write a few words, and some of the
1709 // words may be written to the same address.
1710 bool IsReadOnlyPtr = false;
1711 if (Seen.insert(Ptr).second ||
1712 !getPtrStride(*PSE, Ptr, TheLoop, SymbolicStrides)) {
1714 IsReadOnlyPtr = true;
1717 MemoryLocation Loc = MemoryLocation::get(LD);
1718 // The TBAA metadata could have a control dependency on the predication
1719 // condition, so we cannot rely on it when determining whether or not we
1720 // need runtime pointer checks.
1721 if (blockNeedsPredication(LD->getParent(), TheLoop, DT))
1722 Loc.AATags.TBAA = nullptr;
1724 Accesses.addLoad(Loc, IsReadOnlyPtr);
1727 // If we write (or read-write) to a single destination and there are no
1728 // other reads in this loop then is it safe to vectorize.
1729 if (NumReadWrites == 1 && NumReads == 0) {
1730 DEBUG(dbgs() << "LAA: Found a write-only loop!\n");
1735 // Build dependence sets and check whether we need a runtime pointer bounds
1737 Accesses.buildDependenceSets();
1739 // Find pointers with computable bounds. We are going to use this information
1740 // to place a runtime bound check.
1741 bool CanDoRTIfNeeded = Accesses.canCheckPtrAtRT(*PtrRtChecking, PSE->getSE(),
1742 TheLoop, SymbolicStrides);
1743 if (!CanDoRTIfNeeded) {
1744 recordAnalysis("CantIdentifyArrayBounds") << "cannot identify array bounds";
1745 DEBUG(dbgs() << "LAA: We can't vectorize because we can't find "
1746 << "the array bounds.\n");
1751 DEBUG(dbgs() << "LAA: We can perform a memory runtime check if needed.\n");
1754 if (Accesses.isDependencyCheckNeeded()) {
1755 DEBUG(dbgs() << "LAA: Checking memory dependencies\n");
1756 CanVecMem = DepChecker->areDepsSafe(
1757 DependentAccesses, Accesses.getDependenciesToCheck(), SymbolicStrides);
1758 MaxSafeDepDistBytes = DepChecker->getMaxSafeDepDistBytes();
1760 if (!CanVecMem && DepChecker->shouldRetryWithRuntimeCheck()) {
1761 DEBUG(dbgs() << "LAA: Retrying with memory checks\n");
1763 // Clear the dependency checks. We assume they are not needed.
1764 Accesses.resetDepChecks(*DepChecker);
1766 PtrRtChecking->reset();
1767 PtrRtChecking->Need = true;
1769 auto *SE = PSE->getSE();
1770 CanDoRTIfNeeded = Accesses.canCheckPtrAtRT(*PtrRtChecking, SE, TheLoop,
1771 SymbolicStrides, true);
1773 // Check that we found the bounds for the pointer.
1774 if (!CanDoRTIfNeeded) {
1775 recordAnalysis("CantCheckMemDepsAtRunTime")
1776 << "cannot check memory dependencies at runtime";
1777 DEBUG(dbgs() << "LAA: Can't vectorize with memory checks\n");
1787 DEBUG(dbgs() << "LAA: No unsafe dependent memory operations in loop. We"
1788 << (PtrRtChecking->Need ? "" : " don't")
1789 << " need runtime memory checks.\n");
1791 recordAnalysis("UnsafeMemDep")
1792 << "unsafe dependent memory operations in loop. Use "
1793 "#pragma loop distribute(enable) to allow loop distribution "
1794 "to attempt to isolate the offending operations into a separate "
1796 DEBUG(dbgs() << "LAA: unsafe dependent memory operations in loop\n");
1800 bool LoopAccessInfo::blockNeedsPredication(BasicBlock *BB, Loop *TheLoop,
1801 DominatorTree *DT) {
1802 assert(TheLoop->contains(BB) && "Unknown block used");
1804 // Blocks that do not dominate the latch need predication.
1805 BasicBlock* Latch = TheLoop->getLoopLatch();
1806 return !DT->dominates(BB, Latch);
1809 OptimizationRemarkAnalysis &LoopAccessInfo::recordAnalysis(StringRef RemarkName,
1811 assert(!Report && "Multiple reports generated");
1813 Value *CodeRegion = TheLoop->getHeader();
1814 DebugLoc DL = TheLoop->getStartLoc();
1817 CodeRegion = I->getParent();
1818 // If there is no debug location attached to the instruction, revert back to
1819 // using the loop's.
1820 if (I->getDebugLoc())
1821 DL = I->getDebugLoc();
1824 Report = make_unique<OptimizationRemarkAnalysis>(DEBUG_TYPE, RemarkName, DL,
1829 bool LoopAccessInfo::isUniform(Value *V) const {
1830 auto *SE = PSE->getSE();
1831 // Since we rely on SCEV for uniformity, if the type is not SCEVable, it is
1832 // never considered uniform.
1833 // TODO: Is this really what we want? Even without FP SCEV, we may want some
1834 // trivially loop-invariant FP values to be considered uniform.
1835 if (!SE->isSCEVable(V->getType()))
1837 return (SE->isLoopInvariant(SE->getSCEV(V), TheLoop));
1840 // FIXME: this function is currently a duplicate of the one in
1841 // LoopVectorize.cpp.
1842 static Instruction *getFirstInst(Instruction *FirstInst, Value *V,
1846 if (Instruction *I = dyn_cast<Instruction>(V))
1847 return I->getParent() == Loc->getParent() ? I : nullptr;
1853 /// \brief IR Values for the lower and upper bounds of a pointer evolution. We
1854 /// need to use value-handles because SCEV expansion can invalidate previously
1855 /// expanded values. Thus expansion of a pointer can invalidate the bounds for
1857 struct PointerBounds {
1858 TrackingVH<Value> Start;
1859 TrackingVH<Value> End;
1862 } // end anonymous namespace
1864 /// \brief Expand code for the lower and upper bound of the pointer group \p CG
1865 /// in \p TheLoop. \return the values for the bounds.
1866 static PointerBounds
1867 expandBounds(const RuntimePointerChecking::CheckingPtrGroup *CG, Loop *TheLoop,
1868 Instruction *Loc, SCEVExpander &Exp, ScalarEvolution *SE,
1869 const RuntimePointerChecking &PtrRtChecking) {
1870 Value *Ptr = PtrRtChecking.Pointers[CG->Members[0]].PointerValue;
1871 const SCEV *Sc = SE->getSCEV(Ptr);
1873 unsigned AS = Ptr->getType()->getPointerAddressSpace();
1874 LLVMContext &Ctx = Loc->getContext();
1876 // Use this type for pointer arithmetic.
1877 Type *PtrArithTy = Type::getInt8PtrTy(Ctx, AS);
1879 if (SE->isLoopInvariant(Sc, TheLoop)) {
1880 DEBUG(dbgs() << "LAA: Adding RT check for a loop invariant ptr:" << *Ptr
1882 // Ptr could be in the loop body. If so, expand a new one at the correct
1884 Instruction *Inst = dyn_cast<Instruction>(Ptr);
1885 Value *NewPtr = (Inst && TheLoop->contains(Inst))
1886 ? Exp.expandCodeFor(Sc, PtrArithTy, Loc)
1888 return {NewPtr, NewPtr};
1890 Value *Start = nullptr, *End = nullptr;
1891 DEBUG(dbgs() << "LAA: Adding RT check for range:\n");
1892 Start = Exp.expandCodeFor(CG->Low, PtrArithTy, Loc);
1893 End = Exp.expandCodeFor(CG->High, PtrArithTy, Loc);
1894 DEBUG(dbgs() << "Start: " << *CG->Low << " End: " << *CG->High << "\n");
1895 return {Start, End};
1899 /// \brief Turns a collection of checks into a collection of expanded upper and
1900 /// lower bounds for both pointers in the check.
1901 static SmallVector<std::pair<PointerBounds, PointerBounds>, 4> expandBounds(
1902 const SmallVectorImpl<RuntimePointerChecking::PointerCheck> &PointerChecks,
1903 Loop *L, Instruction *Loc, ScalarEvolution *SE, SCEVExpander &Exp,
1904 const RuntimePointerChecking &PtrRtChecking) {
1905 SmallVector<std::pair<PointerBounds, PointerBounds>, 4> ChecksWithBounds;
1907 // Here we're relying on the SCEV Expander's cache to only emit code for the
1908 // same bounds once.
1910 PointerChecks, std::back_inserter(ChecksWithBounds),
1911 [&](const RuntimePointerChecking::PointerCheck &Check) {
1913 First = expandBounds(Check.first, L, Loc, Exp, SE, PtrRtChecking),
1914 Second = expandBounds(Check.second, L, Loc, Exp, SE, PtrRtChecking);
1915 return std::make_pair(First, Second);
1918 return ChecksWithBounds;
1921 std::pair<Instruction *, Instruction *> LoopAccessInfo::addRuntimeChecks(
1923 const SmallVectorImpl<RuntimePointerChecking::PointerCheck> &PointerChecks)
1925 const DataLayout &DL = TheLoop->getHeader()->getModule()->getDataLayout();
1926 auto *SE = PSE->getSE();
1927 SCEVExpander Exp(*SE, DL, "induction");
1928 auto ExpandedChecks =
1929 expandBounds(PointerChecks, TheLoop, Loc, SE, Exp, *PtrRtChecking);
1931 LLVMContext &Ctx = Loc->getContext();
1932 Instruction *FirstInst = nullptr;
1933 IRBuilder<> ChkBuilder(Loc);
1934 // Our instructions might fold to a constant.
1935 Value *MemoryRuntimeCheck = nullptr;
1937 for (const auto &Check : ExpandedChecks) {
1938 const PointerBounds &A = Check.first, &B = Check.second;
1939 // Check if two pointers (A and B) conflict where conflict is computed as:
1940 // start(A) <= end(B) && start(B) <= end(A)
1941 unsigned AS0 = A.Start->getType()->getPointerAddressSpace();
1942 unsigned AS1 = B.Start->getType()->getPointerAddressSpace();
1944 assert((AS0 == B.End->getType()->getPointerAddressSpace()) &&
1945 (AS1 == A.End->getType()->getPointerAddressSpace()) &&
1946 "Trying to bounds check pointers with different address spaces");
1948 Type *PtrArithTy0 = Type::getInt8PtrTy(Ctx, AS0);
1949 Type *PtrArithTy1 = Type::getInt8PtrTy(Ctx, AS1);
1951 Value *Start0 = ChkBuilder.CreateBitCast(A.Start, PtrArithTy0, "bc");
1952 Value *Start1 = ChkBuilder.CreateBitCast(B.Start, PtrArithTy1, "bc");
1953 Value *End0 = ChkBuilder.CreateBitCast(A.End, PtrArithTy1, "bc");
1954 Value *End1 = ChkBuilder.CreateBitCast(B.End, PtrArithTy0, "bc");
1956 // [A|B].Start points to the first accessed byte under base [A|B].
1957 // [A|B].End points to the last accessed byte, plus one.
1958 // There is no conflict when the intervals are disjoint:
1959 // NoConflict = (B.Start >= A.End) || (A.Start >= B.End)
1961 // bound0 = (B.Start < A.End)
1962 // bound1 = (A.Start < B.End)
1963 // IsConflict = bound0 & bound1
1964 Value *Cmp0 = ChkBuilder.CreateICmpULT(Start0, End1, "bound0");
1965 FirstInst = getFirstInst(FirstInst, Cmp0, Loc);
1966 Value *Cmp1 = ChkBuilder.CreateICmpULT(Start1, End0, "bound1");
1967 FirstInst = getFirstInst(FirstInst, Cmp1, Loc);
1968 Value *IsConflict = ChkBuilder.CreateAnd(Cmp0, Cmp1, "found.conflict");
1969 FirstInst = getFirstInst(FirstInst, IsConflict, Loc);
1970 if (MemoryRuntimeCheck) {
1972 ChkBuilder.CreateOr(MemoryRuntimeCheck, IsConflict, "conflict.rdx");
1973 FirstInst = getFirstInst(FirstInst, IsConflict, Loc);
1975 MemoryRuntimeCheck = IsConflict;
1978 if (!MemoryRuntimeCheck)
1979 return std::make_pair(nullptr, nullptr);
1981 // We have to do this trickery because the IRBuilder might fold the check to a
1982 // constant expression in which case there is no Instruction anchored in a
1984 Instruction *Check = BinaryOperator::CreateAnd(MemoryRuntimeCheck,
1985 ConstantInt::getTrue(Ctx));
1986 ChkBuilder.Insert(Check, "memcheck.conflict");
1987 FirstInst = getFirstInst(FirstInst, Check, Loc);
1988 return std::make_pair(FirstInst, Check);
1991 std::pair<Instruction *, Instruction *>
1992 LoopAccessInfo::addRuntimeChecks(Instruction *Loc) const {
1993 if (!PtrRtChecking->Need)
1994 return std::make_pair(nullptr, nullptr);
1996 return addRuntimeChecks(Loc, PtrRtChecking->getChecks());
1999 void LoopAccessInfo::collectStridedAccess(Value *MemAccess) {
2000 Value *Ptr = nullptr;
2001 if (LoadInst *LI = dyn_cast<LoadInst>(MemAccess))
2002 Ptr = LI->getPointerOperand();
2003 else if (StoreInst *SI = dyn_cast<StoreInst>(MemAccess))
2004 Ptr = SI->getPointerOperand();
2008 Value *Stride = getStrideFromPointer(Ptr, PSE->getSE(), TheLoop);
2012 DEBUG(dbgs() << "LAA: Found a strided access that we can version");
2013 DEBUG(dbgs() << " Ptr: " << *Ptr << " Stride: " << *Stride << "\n");
2014 SymbolicStrides[Ptr] = Stride;
2015 StrideSet.insert(Stride);
2018 LoopAccessInfo::LoopAccessInfo(Loop *L, ScalarEvolution *SE,
2019 const TargetLibraryInfo *TLI, AliasAnalysis *AA,
2020 DominatorTree *DT, LoopInfo *LI)
2021 : PSE(llvm::make_unique<PredicatedScalarEvolution>(*SE, *L)),
2022 PtrRtChecking(llvm::make_unique<RuntimePointerChecking>(SE)),
2023 DepChecker(llvm::make_unique<MemoryDepChecker>(*PSE, L)), TheLoop(L),
2024 NumLoads(0), NumStores(0), MaxSafeDepDistBytes(-1), CanVecMem(false),
2025 StoreToLoopInvariantAddress(false) {
2026 if (canAnalyzeLoop())
2027 analyzeLoop(AA, LI, TLI, DT);
2030 void LoopAccessInfo::print(raw_ostream &OS, unsigned Depth) const {
2032 OS.indent(Depth) << "Memory dependences are safe";
2033 if (MaxSafeDepDistBytes != -1ULL)
2034 OS << " with a maximum dependence distance of " << MaxSafeDepDistBytes
2036 if (PtrRtChecking->Need)
2037 OS << " with run-time checks";
2042 OS.indent(Depth) << "Report: " << Report->getMsg() << "\n";
2044 if (auto *Dependences = DepChecker->getDependences()) {
2045 OS.indent(Depth) << "Dependences:\n";
2046 for (auto &Dep : *Dependences) {
2047 Dep.print(OS, Depth + 2, DepChecker->getMemoryInstructions());
2051 OS.indent(Depth) << "Too many dependences, not recorded\n";
2053 // List the pair of accesses need run-time checks to prove independence.
2054 PtrRtChecking->print(OS, Depth);
2057 OS.indent(Depth) << "Store to invariant address was "
2058 << (StoreToLoopInvariantAddress ? "" : "not ")
2059 << "found in loop.\n";
2061 OS.indent(Depth) << "SCEV assumptions:\n";
2062 PSE->getUnionPredicate().print(OS, Depth);
2066 OS.indent(Depth) << "Expressions re-written:\n";
2067 PSE->print(OS, Depth);
2070 const LoopAccessInfo &LoopAccessLegacyAnalysis::getInfo(Loop *L) {
2071 auto &LAI = LoopAccessInfoMap[L];
2074 LAI = llvm::make_unique<LoopAccessInfo>(L, SE, TLI, AA, DT, LI);
2079 void LoopAccessLegacyAnalysis::print(raw_ostream &OS, const Module *M) const {
2080 LoopAccessLegacyAnalysis &LAA = *const_cast<LoopAccessLegacyAnalysis *>(this);
2082 for (Loop *TopLevelLoop : *LI)
2083 for (Loop *L : depth_first(TopLevelLoop)) {
2084 OS.indent(2) << L->getHeader()->getName() << ":\n";
2085 auto &LAI = LAA.getInfo(L);
2090 bool LoopAccessLegacyAnalysis::runOnFunction(Function &F) {
2091 SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE();
2092 auto *TLIP = getAnalysisIfAvailable<TargetLibraryInfoWrapperPass>();
2093 TLI = TLIP ? &TLIP->getTLI() : nullptr;
2094 AA = &getAnalysis<AAResultsWrapperPass>().getAAResults();
2095 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
2096 LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
2101 void LoopAccessLegacyAnalysis::getAnalysisUsage(AnalysisUsage &AU) const {
2102 AU.addRequired<ScalarEvolutionWrapperPass>();
2103 AU.addRequired<AAResultsWrapperPass>();
2104 AU.addRequired<DominatorTreeWrapperPass>();
2105 AU.addRequired<LoopInfoWrapperPass>();
2107 AU.setPreservesAll();
2110 char LoopAccessLegacyAnalysis::ID = 0;
2111 static const char laa_name[] = "Loop Access Analysis";
2112 #define LAA_NAME "loop-accesses"
2114 INITIALIZE_PASS_BEGIN(LoopAccessLegacyAnalysis, LAA_NAME, laa_name, false, true)
2115 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)
2116 INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass)
2117 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
2118 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
2119 INITIALIZE_PASS_END(LoopAccessLegacyAnalysis, LAA_NAME, laa_name, false, true)
2121 AnalysisKey LoopAccessAnalysis::Key;
2123 LoopAccessInfo LoopAccessAnalysis::run(Loop &L, LoopAnalysisManager &AM,
2124 LoopStandardAnalysisResults &AR) {
2125 return LoopAccessInfo(&L, &AR.SE, &AR.TLI, &AR.AA, &AR.DT, &AR.LI);
2130 Pass *createLAAPass() {
2131 return new LoopAccessLegacyAnalysis();
2134 } // end namespace llvm