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/Analysis/LoopInfo.h"
17 #include "llvm/Analysis/ScalarEvolutionExpander.h"
18 #include "llvm/Analysis/TargetLibraryInfo.h"
19 #include "llvm/Analysis/ValueTracking.h"
20 #include "llvm/IR/DiagnosticInfo.h"
21 #include "llvm/IR/Dominators.h"
22 #include "llvm/IR/IRBuilder.h"
23 #include "llvm/Support/Debug.h"
24 #include "llvm/Support/raw_ostream.h"
25 #include "llvm/Analysis/VectorUtils.h"
28 #define DEBUG_TYPE "loop-accesses"
30 static cl::opt<unsigned, true>
31 VectorizationFactor("force-vector-width", cl::Hidden,
32 cl::desc("Sets the SIMD width. Zero is autoselect."),
33 cl::location(VectorizerParams::VectorizationFactor));
34 unsigned VectorizerParams::VectorizationFactor;
36 static cl::opt<unsigned, true>
37 VectorizationInterleave("force-vector-interleave", cl::Hidden,
38 cl::desc("Sets the vectorization interleave count. "
39 "Zero is autoselect."),
41 VectorizerParams::VectorizationInterleave));
42 unsigned VectorizerParams::VectorizationInterleave;
44 static cl::opt<unsigned, true> RuntimeMemoryCheckThreshold(
45 "runtime-memory-check-threshold", cl::Hidden,
46 cl::desc("When performing memory disambiguation checks at runtime do not "
47 "generate more than this number of comparisons (default = 8)."),
48 cl::location(VectorizerParams::RuntimeMemoryCheckThreshold), cl::init(8));
49 unsigned VectorizerParams::RuntimeMemoryCheckThreshold;
51 /// \brief The maximum iterations used to merge memory checks
52 static cl::opt<unsigned> MemoryCheckMergeThreshold(
53 "memory-check-merge-threshold", cl::Hidden,
54 cl::desc("Maximum number of comparisons done when trying to merge "
55 "runtime memory checks. (default = 100)"),
58 /// Maximum SIMD width.
59 const unsigned VectorizerParams::MaxVectorWidth = 64;
61 /// \brief We collect interesting dependences up to this threshold.
62 static cl::opt<unsigned> MaxInterestingDependence(
63 "max-interesting-dependences", cl::Hidden,
64 cl::desc("Maximum number of interesting dependences collected by "
65 "loop-access analysis (default = 100)"),
68 bool VectorizerParams::isInterleaveForced() {
69 return ::VectorizationInterleave.getNumOccurrences() > 0;
72 void LoopAccessReport::emitAnalysis(const LoopAccessReport &Message,
73 const Function *TheFunction,
75 const char *PassName) {
76 DebugLoc DL = TheLoop->getStartLoc();
77 if (const Instruction *I = Message.getInstr())
78 DL = I->getDebugLoc();
79 emitOptimizationRemarkAnalysis(TheFunction->getContext(), PassName,
80 *TheFunction, DL, Message.str());
83 Value *llvm::stripIntegerCast(Value *V) {
84 if (CastInst *CI = dyn_cast<CastInst>(V))
85 if (CI->getOperand(0)->getType()->isIntegerTy())
86 return CI->getOperand(0);
90 const SCEV *llvm::replaceSymbolicStrideSCEV(ScalarEvolution *SE,
91 const ValueToValueMap &PtrToStride,
92 Value *Ptr, Value *OrigPtr) {
94 const SCEV *OrigSCEV = SE->getSCEV(Ptr);
96 // If there is an entry in the map return the SCEV of the pointer with the
97 // symbolic stride replaced by one.
98 ValueToValueMap::const_iterator SI =
99 PtrToStride.find(OrigPtr ? OrigPtr : Ptr);
100 if (SI != PtrToStride.end()) {
101 Value *StrideVal = SI->second;
104 StrideVal = stripIntegerCast(StrideVal);
106 // Replace symbolic stride by one.
107 Value *One = ConstantInt::get(StrideVal->getType(), 1);
108 ValueToValueMap RewriteMap;
109 RewriteMap[StrideVal] = One;
112 SCEVParameterRewriter::rewrite(OrigSCEV, *SE, RewriteMap, true);
113 DEBUG(dbgs() << "LAA: Replacing SCEV: " << *OrigSCEV << " by: " << *ByOne
118 // Otherwise, just return the SCEV of the original pointer.
119 return SE->getSCEV(Ptr);
122 void LoopAccessInfo::RuntimePointerCheck::insert(
123 Loop *Lp, Value *Ptr, bool WritePtr, unsigned DepSetId, unsigned ASId,
124 const ValueToValueMap &Strides) {
125 // Get the stride replaced scev.
126 const SCEV *Sc = replaceSymbolicStrideSCEV(SE, Strides, Ptr);
127 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Sc);
128 assert(AR && "Invalid addrec expression");
129 const SCEV *Ex = SE->getBackedgeTakenCount(Lp);
130 const SCEV *ScEnd = AR->evaluateAtIteration(Ex, *SE);
131 Pointers.push_back(Ptr);
132 Starts.push_back(AR->getStart());
133 Ends.push_back(ScEnd);
134 IsWritePtr.push_back(WritePtr);
135 DependencySetId.push_back(DepSetId);
136 AliasSetId.push_back(ASId);
140 bool LoopAccessInfo::RuntimePointerCheck::needsChecking(
141 const CheckingPtrGroup &M, const CheckingPtrGroup &N,
142 const SmallVectorImpl<int> *PtrPartition) const {
143 for (unsigned I = 0, EI = M.Members.size(); EI != I; ++I)
144 for (unsigned J = 0, EJ = N.Members.size(); EJ != J; ++J)
145 if (needsChecking(M.Members[I], N.Members[J], PtrPartition))
150 /// Compare \p I and \p J and return the minimum.
151 /// Return nullptr in case we couldn't find an answer.
152 static const SCEV *getMinFromExprs(const SCEV *I, const SCEV *J,
153 ScalarEvolution *SE) {
154 const SCEV *Diff = SE->getMinusSCEV(J, I);
155 const SCEVConstant *C = dyn_cast<const SCEVConstant>(Diff);
159 if (C->getValue()->isNegative())
164 bool LoopAccessInfo::RuntimePointerCheck::CheckingPtrGroup::addPointer(
166 // Compare the starts and ends with the known minimum and maximum
167 // of this set. We need to know how we compare against the min/max
168 // of the set in order to be able to emit memchecks.
169 const SCEV *Min0 = getMinFromExprs(RtCheck.Starts[Index], Low, RtCheck.SE);
173 const SCEV *Min1 = getMinFromExprs(RtCheck.Ends[Index], High, RtCheck.SE);
177 // Update the low bound expression if we've found a new min value.
178 if (Min0 == RtCheck.Starts[Index])
179 Low = RtCheck.Starts[Index];
181 // Update the high bound expression if we've found a new max value.
182 if (Min1 != RtCheck.Ends[Index])
183 High = RtCheck.Ends[Index];
185 Members.push_back(Index);
189 void LoopAccessInfo::RuntimePointerCheck::groupChecks(
190 MemoryDepChecker::DepCandidates &DepCands,
191 bool UseDependencies) {
192 // We build the groups from dependency candidates equivalence classes
194 // - We know that pointers in the same equivalence class share
195 // the same underlying object and therefore there is a chance
196 // that we can compare pointers
197 // - We wouldn't be able to merge two pointers for which we need
198 // to emit a memcheck. The classes in DepCands are already
199 // conveniently built such that no two pointers in the same
200 // class need checking against each other.
202 // We use the following (greedy) algorithm to construct the groups
203 // For every pointer in the equivalence class:
204 // For each existing group:
205 // - if the difference between this pointer and the min/max bounds
206 // of the group is a constant, then make the pointer part of the
207 // group and update the min/max bounds of that group as required.
209 CheckingGroups.clear();
211 // If we don't have the dependency partitions, construct a new
212 // checking pointer group for each pointer.
213 if (!UseDependencies) {
214 for (unsigned I = 0; I < Pointers.size(); ++I)
215 CheckingGroups.push_back(CheckingPtrGroup(I, *this));
219 unsigned TotalComparisons = 0;
221 DenseMap<Value *, unsigned> PositionMap;
222 for (unsigned Pointer = 0; Pointer < Pointers.size(); ++Pointer)
223 PositionMap[Pointers[Pointer]] = Pointer;
225 // We need to keep track of what pointers we've already seen so we
226 // don't process them twice.
227 SmallSet<unsigned, 2> Seen;
229 // Go through all equivalence classes, get the the "pointer check groups"
230 // and add them to the overall solution. We use the order in which accesses
231 // appear in 'Pointers' to enforce determinism.
232 for (unsigned I = 0; I < Pointers.size(); ++I) {
233 // We've seen this pointer before, and therefore already processed
234 // its equivalence class.
238 MemoryDepChecker::MemAccessInfo Access(Pointers[I], IsWritePtr[I]);
240 SmallVector<CheckingPtrGroup, 2> Groups;
241 auto LeaderI = DepCands.findValue(DepCands.getLeaderValue(Access));
243 SmallVector<unsigned, 2> MemberIndices;
245 // Get all indeces of the members of this equivalence class and sort them.
246 // This will allow us to process all accesses in the order in which they
247 // were added to the RuntimePointerCheck.
248 for (auto MI = DepCands.member_begin(LeaderI), ME = DepCands.member_end();
250 unsigned Pointer = PositionMap[MI->getPointer()];
251 MemberIndices.push_back(Pointer);
253 std::sort(MemberIndices.begin(), MemberIndices.end());
255 for (unsigned Pointer : MemberIndices) {
257 // Mark this pointer as seen.
258 Seen.insert(Pointer);
260 // Go through all the existing sets and see if we can find one
261 // which can include this pointer.
262 for (CheckingPtrGroup &Group : Groups) {
263 // Don't perform more than a certain amount of comparisons.
264 // This should limit the cost of grouping the pointers to something
265 // reasonable. If we do end up hitting this threshold, the algorithm
266 // will create separate groups for all remaining pointers.
267 if (TotalComparisons > MemoryCheckMergeThreshold)
272 if (Group.addPointer(Pointer)) {
279 // We couldn't add this pointer to any existing set or the threshold
280 // for the number of comparisons has been reached. Create a new group
281 // to hold the current pointer.
282 Groups.push_back(CheckingPtrGroup(Pointer, *this));
285 // We've computed the grouped checks for this partition.
286 // Save the results and continue with the next one.
287 std::copy(Groups.begin(), Groups.end(), std::back_inserter(CheckingGroups));
291 bool LoopAccessInfo::RuntimePointerCheck::needsChecking(
292 unsigned I, unsigned J, const SmallVectorImpl<int> *PtrPartition) const {
293 // No need to check if two readonly pointers intersect.
294 if (!IsWritePtr[I] && !IsWritePtr[J])
297 // Only need to check pointers between two different dependency sets.
298 if (DependencySetId[I] == DependencySetId[J])
301 // Only need to check pointers in the same alias set.
302 if (AliasSetId[I] != AliasSetId[J])
305 // If PtrPartition is set omit checks between pointers of the same partition.
306 // Partition number -1 means that the pointer is used in multiple partitions.
307 // In this case we can't omit the check.
308 if (PtrPartition && (*PtrPartition)[I] != -1 &&
309 (*PtrPartition)[I] == (*PtrPartition)[J])
315 void LoopAccessInfo::RuntimePointerCheck::print(
316 raw_ostream &OS, unsigned Depth,
317 const SmallVectorImpl<int> *PtrPartition) const {
319 OS.indent(Depth) << "Run-time memory checks:\n";
322 for (unsigned I = 0; I < CheckingGroups.size(); ++I)
323 for (unsigned J = I + 1; J < CheckingGroups.size(); ++J)
324 if (needsChecking(CheckingGroups[I], CheckingGroups[J], PtrPartition)) {
325 OS.indent(Depth) << "Check " << N++ << ":\n";
326 OS.indent(Depth + 2) << "Comparing group " << I << ":\n";
328 for (unsigned K = 0; K < CheckingGroups[I].Members.size(); ++K) {
329 OS.indent(Depth + 2) << *Pointers[CheckingGroups[I].Members[K]]
332 OS << " (Partition: "
333 << (*PtrPartition)[CheckingGroups[I].Members[K]] << ")"
337 OS.indent(Depth + 2) << "Against group " << J << ":\n";
339 for (unsigned K = 0; K < CheckingGroups[J].Members.size(); ++K) {
340 OS.indent(Depth + 2) << *Pointers[CheckingGroups[J].Members[K]]
343 OS << " (Partition: "
344 << (*PtrPartition)[CheckingGroups[J].Members[K]] << ")"
349 OS.indent(Depth) << "Grouped accesses:\n";
350 for (unsigned I = 0; I < CheckingGroups.size(); ++I) {
351 OS.indent(Depth + 2) << "Group " << I << ":\n";
352 OS.indent(Depth + 4) << "(Low: " << *CheckingGroups[I].Low
353 << " High: " << *CheckingGroups[I].High << ")\n";
354 for (unsigned J = 0; J < CheckingGroups[I].Members.size(); ++J) {
355 OS.indent(Depth + 6) << "Member: " << *Exprs[CheckingGroups[I].Members[J]]
361 unsigned LoopAccessInfo::RuntimePointerCheck::getNumberOfChecks(
362 const SmallVectorImpl<int> *PtrPartition) const {
364 unsigned NumPartitions = CheckingGroups.size();
365 unsigned CheckCount = 0;
367 for (unsigned I = 0; I < NumPartitions; ++I)
368 for (unsigned J = I + 1; J < NumPartitions; ++J)
369 if (needsChecking(CheckingGroups[I], CheckingGroups[J], PtrPartition))
374 bool LoopAccessInfo::RuntimePointerCheck::needsAnyChecking(
375 const SmallVectorImpl<int> *PtrPartition) const {
376 unsigned NumPointers = Pointers.size();
378 for (unsigned I = 0; I < NumPointers; ++I)
379 for (unsigned J = I + 1; J < NumPointers; ++J)
380 if (needsChecking(I, J, PtrPartition))
386 /// \brief Analyses memory accesses in a loop.
388 /// Checks whether run time pointer checks are needed and builds sets for data
389 /// dependence checking.
390 class AccessAnalysis {
392 /// \brief Read or write access location.
393 typedef PointerIntPair<Value *, 1, bool> MemAccessInfo;
394 typedef SmallPtrSet<MemAccessInfo, 8> MemAccessInfoSet;
396 AccessAnalysis(const DataLayout &Dl, AliasAnalysis *AA, LoopInfo *LI,
397 MemoryDepChecker::DepCandidates &DA)
398 : DL(Dl), AST(*AA), LI(LI), DepCands(DA),
399 IsRTCheckAnalysisNeeded(false) {}
401 /// \brief Register a load and whether it is only read from.
402 void addLoad(MemoryLocation &Loc, bool IsReadOnly) {
403 Value *Ptr = const_cast<Value*>(Loc.Ptr);
404 AST.add(Ptr, MemoryLocation::UnknownSize, Loc.AATags);
405 Accesses.insert(MemAccessInfo(Ptr, false));
407 ReadOnlyPtr.insert(Ptr);
410 /// \brief Register a store.
411 void addStore(MemoryLocation &Loc) {
412 Value *Ptr = const_cast<Value*>(Loc.Ptr);
413 AST.add(Ptr, MemoryLocation::UnknownSize, Loc.AATags);
414 Accesses.insert(MemAccessInfo(Ptr, true));
417 /// \brief Check whether we can check the pointers at runtime for
418 /// non-intersection.
420 /// Returns true if we need no check or if we do and we can generate them
421 /// (i.e. the pointers have computable bounds).
422 bool canCheckPtrAtRT(LoopAccessInfo::RuntimePointerCheck &RtCheck,
423 ScalarEvolution *SE, Loop *TheLoop,
424 const ValueToValueMap &Strides,
425 bool ShouldCheckStride = false);
427 /// \brief Goes over all memory accesses, checks whether a RT check is needed
428 /// and builds sets of dependent accesses.
429 void buildDependenceSets() {
430 processMemAccesses();
433 /// \brief Initial processing of memory accesses determined that we need to
434 /// perform dependency checking.
436 /// Note that this can later be cleared if we retry memcheck analysis without
437 /// dependency checking (i.e. ShouldRetryWithRuntimeCheck).
438 bool isDependencyCheckNeeded() { return !CheckDeps.empty(); }
440 /// We decided that no dependence analysis would be used. Reset the state.
441 void resetDepChecks(MemoryDepChecker &DepChecker) {
443 DepChecker.clearInterestingDependences();
446 MemAccessInfoSet &getDependenciesToCheck() { return CheckDeps; }
449 typedef SetVector<MemAccessInfo> PtrAccessSet;
451 /// \brief Go over all memory access and check whether runtime pointer checks
452 /// are needed and build sets of dependency check candidates.
453 void processMemAccesses();
455 /// Set of all accesses.
456 PtrAccessSet Accesses;
458 const DataLayout &DL;
460 /// Set of accesses that need a further dependence check.
461 MemAccessInfoSet CheckDeps;
463 /// Set of pointers that are read only.
464 SmallPtrSet<Value*, 16> ReadOnlyPtr;
466 /// An alias set tracker to partition the access set by underlying object and
467 //intrinsic property (such as TBAA metadata).
472 /// Sets of potentially dependent accesses - members of one set share an
473 /// underlying pointer. The set "CheckDeps" identfies which sets really need a
474 /// dependence check.
475 MemoryDepChecker::DepCandidates &DepCands;
477 /// \brief Initial processing of memory accesses determined that we may need
478 /// to add memchecks. Perform the analysis to determine the necessary checks.
480 /// Note that, this is different from isDependencyCheckNeeded. When we retry
481 /// memcheck analysis without dependency checking
482 /// (i.e. ShouldRetryWithRuntimeCheck), isDependencyCheckNeeded is cleared
483 /// while this remains set if we have potentially dependent accesses.
484 bool IsRTCheckAnalysisNeeded;
487 } // end anonymous namespace
489 /// \brief Check whether a pointer can participate in a runtime bounds check.
490 static bool hasComputableBounds(ScalarEvolution *SE,
491 const ValueToValueMap &Strides, Value *Ptr) {
492 const SCEV *PtrScev = replaceSymbolicStrideSCEV(SE, Strides, Ptr);
493 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
497 return AR->isAffine();
500 bool AccessAnalysis::canCheckPtrAtRT(
501 LoopAccessInfo::RuntimePointerCheck &RtCheck, ScalarEvolution *SE,
502 Loop *TheLoop, const ValueToValueMap &StridesMap, bool ShouldCheckStride) {
503 // Find pointers with computable bounds. We are going to use this information
504 // to place a runtime bound check.
507 bool NeedRTCheck = false;
508 if (!IsRTCheckAnalysisNeeded) return true;
510 bool IsDepCheckNeeded = isDependencyCheckNeeded();
512 // We assign a consecutive id to access from different alias sets.
513 // Accesses between different groups doesn't need to be checked.
515 for (auto &AS : AST) {
516 int NumReadPtrChecks = 0;
517 int NumWritePtrChecks = 0;
519 // We assign consecutive id to access from different dependence sets.
520 // Accesses within the same set don't need a runtime check.
521 unsigned RunningDepId = 1;
522 DenseMap<Value *, unsigned> DepSetId;
525 Value *Ptr = A.getValue();
526 bool IsWrite = Accesses.count(MemAccessInfo(Ptr, true));
527 MemAccessInfo Access(Ptr, IsWrite);
534 if (hasComputableBounds(SE, StridesMap, Ptr) &&
535 // When we run after a failing dependency check we have to make sure
536 // we don't have wrapping pointers.
537 (!ShouldCheckStride ||
538 isStridedPtr(SE, Ptr, TheLoop, StridesMap) == 1)) {
539 // The id of the dependence set.
542 if (IsDepCheckNeeded) {
543 Value *Leader = DepCands.getLeaderValue(Access).getPointer();
544 unsigned &LeaderId = DepSetId[Leader];
546 LeaderId = RunningDepId++;
549 // Each access has its own dependence set.
550 DepId = RunningDepId++;
552 RtCheck.insert(TheLoop, Ptr, IsWrite, DepId, ASId, StridesMap);
554 DEBUG(dbgs() << "LAA: Found a runtime check ptr:" << *Ptr << '\n');
556 DEBUG(dbgs() << "LAA: Can't find bounds for ptr:" << *Ptr << '\n');
561 // If we have at least two writes or one write and a read then we need to
562 // check them. But there is no need to checks if there is only one
563 // dependence set for this alias set.
565 // Note that this function computes CanDoRT and NeedRTCheck independently.
566 // For example CanDoRT=false, NeedRTCheck=false means that we have a pointer
567 // for which we couldn't find the bounds but we don't actually need to emit
568 // any checks so it does not matter.
569 if (!(IsDepCheckNeeded && CanDoRT && RunningDepId == 2))
570 NeedRTCheck |= (NumWritePtrChecks >= 2 || (NumReadPtrChecks >= 1 &&
571 NumWritePtrChecks >= 1));
576 // If the pointers that we would use for the bounds comparison have different
577 // address spaces, assume the values aren't directly comparable, so we can't
578 // use them for the runtime check. We also have to assume they could
579 // overlap. In the future there should be metadata for whether address spaces
581 unsigned NumPointers = RtCheck.Pointers.size();
582 for (unsigned i = 0; i < NumPointers; ++i) {
583 for (unsigned j = i + 1; j < NumPointers; ++j) {
584 // Only need to check pointers between two different dependency sets.
585 if (RtCheck.DependencySetId[i] == RtCheck.DependencySetId[j])
587 // Only need to check pointers in the same alias set.
588 if (RtCheck.AliasSetId[i] != RtCheck.AliasSetId[j])
591 Value *PtrI = RtCheck.Pointers[i];
592 Value *PtrJ = RtCheck.Pointers[j];
594 unsigned ASi = PtrI->getType()->getPointerAddressSpace();
595 unsigned ASj = PtrJ->getType()->getPointerAddressSpace();
597 DEBUG(dbgs() << "LAA: Runtime check would require comparison between"
598 " different address spaces\n");
604 if (NeedRTCheck && CanDoRT)
605 RtCheck.groupChecks(DepCands, IsDepCheckNeeded);
607 DEBUG(dbgs() << "LAA: We need to do " << RtCheck.getNumberOfChecks(nullptr)
608 << " pointer comparisons.\n");
610 RtCheck.Need = NeedRTCheck;
612 bool CanDoRTIfNeeded = !NeedRTCheck || CanDoRT;
613 if (!CanDoRTIfNeeded)
615 return CanDoRTIfNeeded;
618 void AccessAnalysis::processMemAccesses() {
619 // We process the set twice: first we process read-write pointers, last we
620 // process read-only pointers. This allows us to skip dependence tests for
621 // read-only pointers.
623 DEBUG(dbgs() << "LAA: Processing memory accesses...\n");
624 DEBUG(dbgs() << " AST: "; AST.dump());
625 DEBUG(dbgs() << "LAA: Accesses(" << Accesses.size() << "):\n");
627 for (auto A : Accesses)
628 dbgs() << "\t" << *A.getPointer() << " (" <<
629 (A.getInt() ? "write" : (ReadOnlyPtr.count(A.getPointer()) ?
630 "read-only" : "read")) << ")\n";
633 // The AliasSetTracker has nicely partitioned our pointers by metadata
634 // compatibility and potential for underlying-object overlap. As a result, we
635 // only need to check for potential pointer dependencies within each alias
637 for (auto &AS : AST) {
638 // Note that both the alias-set tracker and the alias sets themselves used
639 // linked lists internally and so the iteration order here is deterministic
640 // (matching the original instruction order within each set).
642 bool SetHasWrite = false;
644 // Map of pointers to last access encountered.
645 typedef DenseMap<Value*, MemAccessInfo> UnderlyingObjToAccessMap;
646 UnderlyingObjToAccessMap ObjToLastAccess;
648 // Set of access to check after all writes have been processed.
649 PtrAccessSet DeferredAccesses;
651 // Iterate over each alias set twice, once to process read/write pointers,
652 // and then to process read-only pointers.
653 for (int SetIteration = 0; SetIteration < 2; ++SetIteration) {
654 bool UseDeferred = SetIteration > 0;
655 PtrAccessSet &S = UseDeferred ? DeferredAccesses : Accesses;
658 Value *Ptr = AV.getValue();
660 // For a single memory access in AliasSetTracker, Accesses may contain
661 // both read and write, and they both need to be handled for CheckDeps.
663 if (AC.getPointer() != Ptr)
666 bool IsWrite = AC.getInt();
668 // If we're using the deferred access set, then it contains only
670 bool IsReadOnlyPtr = ReadOnlyPtr.count(Ptr) && !IsWrite;
671 if (UseDeferred && !IsReadOnlyPtr)
673 // Otherwise, the pointer must be in the PtrAccessSet, either as a
675 assert(((IsReadOnlyPtr && UseDeferred) || IsWrite ||
676 S.count(MemAccessInfo(Ptr, false))) &&
677 "Alias-set pointer not in the access set?");
679 MemAccessInfo Access(Ptr, IsWrite);
680 DepCands.insert(Access);
682 // Memorize read-only pointers for later processing and skip them in
683 // the first round (they need to be checked after we have seen all
684 // write pointers). Note: we also mark pointer that are not
685 // consecutive as "read-only" pointers (so that we check
686 // "a[b[i]] +="). Hence, we need the second check for "!IsWrite".
687 if (!UseDeferred && IsReadOnlyPtr) {
688 DeferredAccesses.insert(Access);
692 // If this is a write - check other reads and writes for conflicts. If
693 // this is a read only check other writes for conflicts (but only if
694 // there is no other write to the ptr - this is an optimization to
695 // catch "a[i] = a[i] + " without having to do a dependence check).
696 if ((IsWrite || IsReadOnlyPtr) && SetHasWrite) {
697 CheckDeps.insert(Access);
698 IsRTCheckAnalysisNeeded = true;
704 // Create sets of pointers connected by a shared alias set and
705 // underlying object.
706 typedef SmallVector<Value *, 16> ValueVector;
707 ValueVector TempObjects;
709 GetUnderlyingObjects(Ptr, TempObjects, DL, LI);
710 DEBUG(dbgs() << "Underlying objects for pointer " << *Ptr << "\n");
711 for (Value *UnderlyingObj : TempObjects) {
712 UnderlyingObjToAccessMap::iterator Prev =
713 ObjToLastAccess.find(UnderlyingObj);
714 if (Prev != ObjToLastAccess.end())
715 DepCands.unionSets(Access, Prev->second);
717 ObjToLastAccess[UnderlyingObj] = Access;
718 DEBUG(dbgs() << " " << *UnderlyingObj << "\n");
726 static bool isInBoundsGep(Value *Ptr) {
727 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr))
728 return GEP->isInBounds();
732 /// \brief Return true if an AddRec pointer \p Ptr is unsigned non-wrapping,
733 /// i.e. monotonically increasing/decreasing.
734 static bool isNoWrapAddRec(Value *Ptr, const SCEVAddRecExpr *AR,
735 ScalarEvolution *SE, const Loop *L) {
736 // FIXME: This should probably only return true for NUW.
737 if (AR->getNoWrapFlags(SCEV::NoWrapMask))
740 // Scalar evolution does not propagate the non-wrapping flags to values that
741 // are derived from a non-wrapping induction variable because non-wrapping
742 // could be flow-sensitive.
744 // Look through the potentially overflowing instruction to try to prove
745 // non-wrapping for the *specific* value of Ptr.
747 // The arithmetic implied by an inbounds GEP can't overflow.
748 auto *GEP = dyn_cast<GetElementPtrInst>(Ptr);
749 if (!GEP || !GEP->isInBounds())
752 // Make sure there is only one non-const index and analyze that.
753 Value *NonConstIndex = nullptr;
754 for (auto Index = GEP->idx_begin(); Index != GEP->idx_end(); ++Index)
755 if (!isa<ConstantInt>(*Index)) {
758 NonConstIndex = *Index;
761 // The recurrence is on the pointer, ignore for now.
764 // The index in GEP is signed. It is non-wrapping if it's derived from a NSW
765 // AddRec using a NSW operation.
766 if (auto *OBO = dyn_cast<OverflowingBinaryOperator>(NonConstIndex))
767 if (OBO->hasNoSignedWrap() &&
768 // Assume constant for other the operand so that the AddRec can be
770 isa<ConstantInt>(OBO->getOperand(1))) {
771 auto *OpScev = SE->getSCEV(OBO->getOperand(0));
773 if (auto *OpAR = dyn_cast<SCEVAddRecExpr>(OpScev))
774 return OpAR->getLoop() == L && OpAR->getNoWrapFlags(SCEV::FlagNSW);
780 /// \brief Check whether the access through \p Ptr has a constant stride.
781 int llvm::isStridedPtr(ScalarEvolution *SE, Value *Ptr, const Loop *Lp,
782 const ValueToValueMap &StridesMap) {
783 const Type *Ty = Ptr->getType();
784 assert(Ty->isPointerTy() && "Unexpected non-ptr");
786 // Make sure that the pointer does not point to aggregate types.
787 const PointerType *PtrTy = cast<PointerType>(Ty);
788 if (PtrTy->getElementType()->isAggregateType()) {
789 DEBUG(dbgs() << "LAA: Bad stride - Not a pointer to a scalar type"
794 const SCEV *PtrScev = replaceSymbolicStrideSCEV(SE, StridesMap, Ptr);
796 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
798 DEBUG(dbgs() << "LAA: Bad stride - Not an AddRecExpr pointer "
799 << *Ptr << " SCEV: " << *PtrScev << "\n");
803 // The accesss function must stride over the innermost loop.
804 if (Lp != AR->getLoop()) {
805 DEBUG(dbgs() << "LAA: Bad stride - Not striding over innermost loop " <<
806 *Ptr << " SCEV: " << *PtrScev << "\n");
809 // The address calculation must not wrap. Otherwise, a dependence could be
811 // An inbounds getelementptr that is a AddRec with a unit stride
812 // cannot wrap per definition. The unit stride requirement is checked later.
813 // An getelementptr without an inbounds attribute and unit stride would have
814 // to access the pointer value "0" which is undefined behavior in address
815 // space 0, therefore we can also vectorize this case.
816 bool IsInBoundsGEP = isInBoundsGep(Ptr);
817 bool IsNoWrapAddRec = isNoWrapAddRec(Ptr, AR, SE, Lp);
818 bool IsInAddressSpaceZero = PtrTy->getAddressSpace() == 0;
819 if (!IsNoWrapAddRec && !IsInBoundsGEP && !IsInAddressSpaceZero) {
820 DEBUG(dbgs() << "LAA: Bad stride - Pointer may wrap in the address space "
821 << *Ptr << " SCEV: " << *PtrScev << "\n");
825 // Check the step is constant.
826 const SCEV *Step = AR->getStepRecurrence(*SE);
828 // Calculate the pointer stride and check if it is constant.
829 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
831 DEBUG(dbgs() << "LAA: Bad stride - Not a constant strided " << *Ptr <<
832 " SCEV: " << *PtrScev << "\n");
836 auto &DL = Lp->getHeader()->getModule()->getDataLayout();
837 int64_t Size = DL.getTypeAllocSize(PtrTy->getElementType());
838 const APInt &APStepVal = C->getValue()->getValue();
840 // Huge step value - give up.
841 if (APStepVal.getBitWidth() > 64)
844 int64_t StepVal = APStepVal.getSExtValue();
847 int64_t Stride = StepVal / Size;
848 int64_t Rem = StepVal % Size;
852 // If the SCEV could wrap but we have an inbounds gep with a unit stride we
853 // know we can't "wrap around the address space". In case of address space
854 // zero we know that this won't happen without triggering undefined behavior.
855 if (!IsNoWrapAddRec && (IsInBoundsGEP || IsInAddressSpaceZero) &&
856 Stride != 1 && Stride != -1)
862 bool MemoryDepChecker::Dependence::isSafeForVectorization(DepType Type) {
866 case BackwardVectorizable:
870 case ForwardButPreventsForwarding:
872 case BackwardVectorizableButPreventsForwarding:
875 llvm_unreachable("unexpected DepType!");
878 bool MemoryDepChecker::Dependence::isInterestingDependence(DepType Type) {
884 case BackwardVectorizable:
886 case ForwardButPreventsForwarding:
888 case BackwardVectorizableButPreventsForwarding:
891 llvm_unreachable("unexpected DepType!");
894 bool MemoryDepChecker::Dependence::isPossiblyBackward() const {
898 case ForwardButPreventsForwarding:
902 case BackwardVectorizable:
904 case BackwardVectorizableButPreventsForwarding:
907 llvm_unreachable("unexpected DepType!");
910 bool MemoryDepChecker::couldPreventStoreLoadForward(unsigned Distance,
911 unsigned TypeByteSize) {
912 // If loads occur at a distance that is not a multiple of a feasible vector
913 // factor store-load forwarding does not take place.
914 // Positive dependences might cause troubles because vectorizing them might
915 // prevent store-load forwarding making vectorized code run a lot slower.
916 // a[i] = a[i-3] ^ a[i-8];
917 // The stores to a[i:i+1] don't align with the stores to a[i-3:i-2] and
918 // hence on your typical architecture store-load forwarding does not take
919 // place. Vectorizing in such cases does not make sense.
920 // Store-load forwarding distance.
921 const unsigned NumCyclesForStoreLoadThroughMemory = 8*TypeByteSize;
922 // Maximum vector factor.
923 unsigned MaxVFWithoutSLForwardIssues =
924 VectorizerParams::MaxVectorWidth * TypeByteSize;
925 if(MaxSafeDepDistBytes < MaxVFWithoutSLForwardIssues)
926 MaxVFWithoutSLForwardIssues = MaxSafeDepDistBytes;
928 for (unsigned vf = 2*TypeByteSize; vf <= MaxVFWithoutSLForwardIssues;
930 if (Distance % vf && Distance / vf < NumCyclesForStoreLoadThroughMemory) {
931 MaxVFWithoutSLForwardIssues = (vf >>=1);
936 if (MaxVFWithoutSLForwardIssues< 2*TypeByteSize) {
937 DEBUG(dbgs() << "LAA: Distance " << Distance <<
938 " that could cause a store-load forwarding conflict\n");
942 if (MaxVFWithoutSLForwardIssues < MaxSafeDepDistBytes &&
943 MaxVFWithoutSLForwardIssues !=
944 VectorizerParams::MaxVectorWidth * TypeByteSize)
945 MaxSafeDepDistBytes = MaxVFWithoutSLForwardIssues;
949 /// \brief Check the dependence for two accesses with the same stride \p Stride.
950 /// \p Distance is the positive distance and \p TypeByteSize is type size in
953 /// \returns true if they are independent.
954 static bool areStridedAccessesIndependent(unsigned Distance, unsigned Stride,
955 unsigned TypeByteSize) {
956 assert(Stride > 1 && "The stride must be greater than 1");
957 assert(TypeByteSize > 0 && "The type size in byte must be non-zero");
958 assert(Distance > 0 && "The distance must be non-zero");
960 // Skip if the distance is not multiple of type byte size.
961 if (Distance % TypeByteSize)
964 unsigned ScaledDist = Distance / TypeByteSize;
966 // No dependence if the scaled distance is not multiple of the stride.
968 // for (i = 0; i < 1024 ; i += 4)
969 // A[i+2] = A[i] + 1;
971 // Two accesses in memory (scaled distance is 2, stride is 4):
972 // | A[0] | | | | A[4] | | | |
973 // | | | A[2] | | | | A[6] | |
976 // for (i = 0; i < 1024 ; i += 3)
977 // A[i+4] = A[i] + 1;
979 // Two accesses in memory (scaled distance is 4, stride is 3):
980 // | A[0] | | | A[3] | | | A[6] | | |
981 // | | | | | A[4] | | | A[7] | |
982 return ScaledDist % Stride;
985 MemoryDepChecker::Dependence::DepType
986 MemoryDepChecker::isDependent(const MemAccessInfo &A, unsigned AIdx,
987 const MemAccessInfo &B, unsigned BIdx,
988 const ValueToValueMap &Strides) {
989 assert (AIdx < BIdx && "Must pass arguments in program order");
991 Value *APtr = A.getPointer();
992 Value *BPtr = B.getPointer();
993 bool AIsWrite = A.getInt();
994 bool BIsWrite = B.getInt();
996 // Two reads are independent.
997 if (!AIsWrite && !BIsWrite)
998 return Dependence::NoDep;
1000 // We cannot check pointers in different address spaces.
1001 if (APtr->getType()->getPointerAddressSpace() !=
1002 BPtr->getType()->getPointerAddressSpace())
1003 return Dependence::Unknown;
1005 const SCEV *AScev = replaceSymbolicStrideSCEV(SE, Strides, APtr);
1006 const SCEV *BScev = replaceSymbolicStrideSCEV(SE, Strides, BPtr);
1008 int StrideAPtr = isStridedPtr(SE, APtr, InnermostLoop, Strides);
1009 int StrideBPtr = isStridedPtr(SE, BPtr, InnermostLoop, Strides);
1011 const SCEV *Src = AScev;
1012 const SCEV *Sink = BScev;
1014 // If the induction step is negative we have to invert source and sink of the
1016 if (StrideAPtr < 0) {
1019 std::swap(APtr, BPtr);
1020 std::swap(Src, Sink);
1021 std::swap(AIsWrite, BIsWrite);
1022 std::swap(AIdx, BIdx);
1023 std::swap(StrideAPtr, StrideBPtr);
1026 const SCEV *Dist = SE->getMinusSCEV(Sink, Src);
1028 DEBUG(dbgs() << "LAA: Src Scev: " << *Src << "Sink Scev: " << *Sink
1029 << "(Induction step: " << StrideAPtr << ")\n");
1030 DEBUG(dbgs() << "LAA: Distance for " << *InstMap[AIdx] << " to "
1031 << *InstMap[BIdx] << ": " << *Dist << "\n");
1033 // Need accesses with constant stride. We don't want to vectorize
1034 // "A[B[i]] += ..." and similar code or pointer arithmetic that could wrap in
1035 // the address space.
1036 if (!StrideAPtr || !StrideBPtr || StrideAPtr != StrideBPtr){
1037 DEBUG(dbgs() << "Pointer access with non-constant stride\n");
1038 return Dependence::Unknown;
1041 const SCEVConstant *C = dyn_cast<SCEVConstant>(Dist);
1043 DEBUG(dbgs() << "LAA: Dependence because of non-constant distance\n");
1044 ShouldRetryWithRuntimeCheck = true;
1045 return Dependence::Unknown;
1048 Type *ATy = APtr->getType()->getPointerElementType();
1049 Type *BTy = BPtr->getType()->getPointerElementType();
1050 auto &DL = InnermostLoop->getHeader()->getModule()->getDataLayout();
1051 unsigned TypeByteSize = DL.getTypeAllocSize(ATy);
1053 // Negative distances are not plausible dependencies.
1054 const APInt &Val = C->getValue()->getValue();
1055 if (Val.isNegative()) {
1056 bool IsTrueDataDependence = (AIsWrite && !BIsWrite);
1057 if (IsTrueDataDependence &&
1058 (couldPreventStoreLoadForward(Val.abs().getZExtValue(), TypeByteSize) ||
1060 return Dependence::ForwardButPreventsForwarding;
1062 DEBUG(dbgs() << "LAA: Dependence is negative: NoDep\n");
1063 return Dependence::Forward;
1066 // Write to the same location with the same size.
1067 // Could be improved to assert type sizes are the same (i32 == float, etc).
1070 return Dependence::NoDep;
1071 DEBUG(dbgs() << "LAA: Zero dependence difference but different types\n");
1072 return Dependence::Unknown;
1075 assert(Val.isStrictlyPositive() && "Expect a positive value");
1079 "LAA: ReadWrite-Write positive dependency with different types\n");
1080 return Dependence::Unknown;
1083 unsigned Distance = (unsigned) Val.getZExtValue();
1085 unsigned Stride = std::abs(StrideAPtr);
1087 areStridedAccessesIndependent(Distance, Stride, TypeByteSize)) {
1088 DEBUG(dbgs() << "LAA: Strided accesses are independent\n");
1089 return Dependence::NoDep;
1092 // Bail out early if passed-in parameters make vectorization not feasible.
1093 unsigned ForcedFactor = (VectorizerParams::VectorizationFactor ?
1094 VectorizerParams::VectorizationFactor : 1);
1095 unsigned ForcedUnroll = (VectorizerParams::VectorizationInterleave ?
1096 VectorizerParams::VectorizationInterleave : 1);
1097 // The minimum number of iterations for a vectorized/unrolled version.
1098 unsigned MinNumIter = std::max(ForcedFactor * ForcedUnroll, 2U);
1100 // It's not vectorizable if the distance is smaller than the minimum distance
1101 // needed for a vectroized/unrolled version. Vectorizing one iteration in
1102 // front needs TypeByteSize * Stride. Vectorizing the last iteration needs
1103 // TypeByteSize (No need to plus the last gap distance).
1105 // E.g. Assume one char is 1 byte in memory and one int is 4 bytes.
1107 // int *B = (int *)((char *)A + 14);
1108 // for (i = 0 ; i < 1024 ; i += 2)
1112 // Two accesses in memory (stride is 2):
1113 // | A[0] | | A[2] | | A[4] | | A[6] | |
1114 // | B[0] | | B[2] | | B[4] |
1116 // Distance needs for vectorizing iterations except the last iteration:
1117 // 4 * 2 * (MinNumIter - 1). Distance needs for the last iteration: 4.
1118 // So the minimum distance needed is: 4 * 2 * (MinNumIter - 1) + 4.
1120 // If MinNumIter is 2, it is vectorizable as the minimum distance needed is
1121 // 12, which is less than distance.
1123 // If MinNumIter is 4 (Say if a user forces the vectorization factor to be 4),
1124 // the minimum distance needed is 28, which is greater than distance. It is
1125 // not safe to do vectorization.
1126 unsigned MinDistanceNeeded =
1127 TypeByteSize * Stride * (MinNumIter - 1) + TypeByteSize;
1128 if (MinDistanceNeeded > Distance) {
1129 DEBUG(dbgs() << "LAA: Failure because of positive distance " << Distance
1131 return Dependence::Backward;
1134 // Unsafe if the minimum distance needed is greater than max safe distance.
1135 if (MinDistanceNeeded > MaxSafeDepDistBytes) {
1136 DEBUG(dbgs() << "LAA: Failure because it needs at least "
1137 << MinDistanceNeeded << " size in bytes");
1138 return Dependence::Backward;
1141 // Positive distance bigger than max vectorization factor.
1142 // FIXME: Should use max factor instead of max distance in bytes, which could
1143 // not handle different types.
1144 // E.g. Assume one char is 1 byte in memory and one int is 4 bytes.
1145 // void foo (int *A, char *B) {
1146 // for (unsigned i = 0; i < 1024; i++) {
1147 // A[i+2] = A[i] + 1;
1148 // B[i+2] = B[i] + 1;
1152 // This case is currently unsafe according to the max safe distance. If we
1153 // analyze the two accesses on array B, the max safe dependence distance
1154 // is 2. Then we analyze the accesses on array A, the minimum distance needed
1155 // is 8, which is less than 2 and forbidden vectorization, But actually
1156 // both A and B could be vectorized by 2 iterations.
1157 MaxSafeDepDistBytes =
1158 Distance < MaxSafeDepDistBytes ? Distance : MaxSafeDepDistBytes;
1160 bool IsTrueDataDependence = (!AIsWrite && BIsWrite);
1161 if (IsTrueDataDependence &&
1162 couldPreventStoreLoadForward(Distance, TypeByteSize))
1163 return Dependence::BackwardVectorizableButPreventsForwarding;
1165 DEBUG(dbgs() << "LAA: Positive distance " << Val.getSExtValue()
1166 << " with max VF = "
1167 << MaxSafeDepDistBytes / (TypeByteSize * Stride) << '\n');
1169 return Dependence::BackwardVectorizable;
1172 bool MemoryDepChecker::areDepsSafe(DepCandidates &AccessSets,
1173 MemAccessInfoSet &CheckDeps,
1174 const ValueToValueMap &Strides) {
1176 MaxSafeDepDistBytes = -1U;
1177 while (!CheckDeps.empty()) {
1178 MemAccessInfo CurAccess = *CheckDeps.begin();
1180 // Get the relevant memory access set.
1181 EquivalenceClasses<MemAccessInfo>::iterator I =
1182 AccessSets.findValue(AccessSets.getLeaderValue(CurAccess));
1184 // Check accesses within this set.
1185 EquivalenceClasses<MemAccessInfo>::member_iterator AI, AE;
1186 AI = AccessSets.member_begin(I), AE = AccessSets.member_end();
1188 // Check every access pair.
1190 CheckDeps.erase(*AI);
1191 EquivalenceClasses<MemAccessInfo>::member_iterator OI = std::next(AI);
1193 // Check every accessing instruction pair in program order.
1194 for (std::vector<unsigned>::iterator I1 = Accesses[*AI].begin(),
1195 I1E = Accesses[*AI].end(); I1 != I1E; ++I1)
1196 for (std::vector<unsigned>::iterator I2 = Accesses[*OI].begin(),
1197 I2E = Accesses[*OI].end(); I2 != I2E; ++I2) {
1198 auto A = std::make_pair(&*AI, *I1);
1199 auto B = std::make_pair(&*OI, *I2);
1205 Dependence::DepType Type =
1206 isDependent(*A.first, A.second, *B.first, B.second, Strides);
1207 SafeForVectorization &= Dependence::isSafeForVectorization(Type);
1209 // Gather dependences unless we accumulated MaxInterestingDependence
1210 // dependences. In that case return as soon as we find the first
1211 // unsafe dependence. This puts a limit on this quadratic
1213 if (RecordInterestingDependences) {
1214 if (Dependence::isInterestingDependence(Type))
1215 InterestingDependences.push_back(
1216 Dependence(A.second, B.second, Type));
1218 if (InterestingDependences.size() >= MaxInterestingDependence) {
1219 RecordInterestingDependences = false;
1220 InterestingDependences.clear();
1221 DEBUG(dbgs() << "Too many dependences, stopped recording\n");
1224 if (!RecordInterestingDependences && !SafeForVectorization)
1233 DEBUG(dbgs() << "Total Interesting Dependences: "
1234 << InterestingDependences.size() << "\n");
1235 return SafeForVectorization;
1238 SmallVector<Instruction *, 4>
1239 MemoryDepChecker::getInstructionsForAccess(Value *Ptr, bool isWrite) const {
1240 MemAccessInfo Access(Ptr, isWrite);
1241 auto &IndexVector = Accesses.find(Access)->second;
1243 SmallVector<Instruction *, 4> Insts;
1244 std::transform(IndexVector.begin(), IndexVector.end(),
1245 std::back_inserter(Insts),
1246 [&](unsigned Idx) { return this->InstMap[Idx]; });
1250 const char *MemoryDepChecker::Dependence::DepName[] = {
1251 "NoDep", "Unknown", "Forward", "ForwardButPreventsForwarding", "Backward",
1252 "BackwardVectorizable", "BackwardVectorizableButPreventsForwarding"};
1254 void MemoryDepChecker::Dependence::print(
1255 raw_ostream &OS, unsigned Depth,
1256 const SmallVectorImpl<Instruction *> &Instrs) const {
1257 OS.indent(Depth) << DepName[Type] << ":\n";
1258 OS.indent(Depth + 2) << *Instrs[Source] << " -> \n";
1259 OS.indent(Depth + 2) << *Instrs[Destination] << "\n";
1262 bool LoopAccessInfo::canAnalyzeLoop() {
1263 // We need to have a loop header.
1264 DEBUG(dbgs() << "LAA: Found a loop: " <<
1265 TheLoop->getHeader()->getName() << '\n');
1267 // We can only analyze innermost loops.
1268 if (!TheLoop->empty()) {
1269 DEBUG(dbgs() << "LAA: loop is not the innermost loop\n");
1270 emitAnalysis(LoopAccessReport() << "loop is not the innermost loop");
1274 // We must have a single backedge.
1275 if (TheLoop->getNumBackEdges() != 1) {
1276 DEBUG(dbgs() << "LAA: loop control flow is not understood by analyzer\n");
1278 LoopAccessReport() <<
1279 "loop control flow is not understood by analyzer");
1283 // We must have a single exiting block.
1284 if (!TheLoop->getExitingBlock()) {
1285 DEBUG(dbgs() << "LAA: loop control flow is not understood by analyzer\n");
1287 LoopAccessReport() <<
1288 "loop control flow is not understood by analyzer");
1292 // We only handle bottom-tested loops, i.e. loop in which the condition is
1293 // checked at the end of each iteration. With that we can assume that all
1294 // instructions in the loop are executed the same number of times.
1295 if (TheLoop->getExitingBlock() != TheLoop->getLoopLatch()) {
1296 DEBUG(dbgs() << "LAA: loop control flow is not understood by analyzer\n");
1298 LoopAccessReport() <<
1299 "loop control flow is not understood by analyzer");
1303 // ScalarEvolution needs to be able to find the exit count.
1304 const SCEV *ExitCount = SE->getBackedgeTakenCount(TheLoop);
1305 if (ExitCount == SE->getCouldNotCompute()) {
1306 emitAnalysis(LoopAccessReport() <<
1307 "could not determine number of loop iterations");
1308 DEBUG(dbgs() << "LAA: SCEV could not compute the loop exit count.\n");
1315 void LoopAccessInfo::analyzeLoop(const ValueToValueMap &Strides) {
1317 typedef SmallVector<Value*, 16> ValueVector;
1318 typedef SmallPtrSet<Value*, 16> ValueSet;
1320 // Holds the Load and Store *instructions*.
1324 // Holds all the different accesses in the loop.
1325 unsigned NumReads = 0;
1326 unsigned NumReadWrites = 0;
1328 PtrRtCheck.Pointers.clear();
1329 PtrRtCheck.Need = false;
1331 const bool IsAnnotatedParallel = TheLoop->isAnnotatedParallel();
1334 for (Loop::block_iterator bb = TheLoop->block_begin(),
1335 be = TheLoop->block_end(); bb != be; ++bb) {
1337 // Scan the BB and collect legal loads and stores.
1338 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
1341 // If this is a load, save it. If this instruction can read from memory
1342 // but is not a load, then we quit. Notice that we don't handle function
1343 // calls that read or write.
1344 if (it->mayReadFromMemory()) {
1345 // Many math library functions read the rounding mode. We will only
1346 // vectorize a loop if it contains known function calls that don't set
1347 // the flag. Therefore, it is safe to ignore this read from memory.
1348 CallInst *Call = dyn_cast<CallInst>(it);
1349 if (Call && getIntrinsicIDForCall(Call, TLI))
1352 // If the function has an explicit vectorized counterpart, we can safely
1353 // assume that it can be vectorized.
1354 if (Call && !Call->isNoBuiltin() && Call->getCalledFunction() &&
1355 TLI->isFunctionVectorizable(Call->getCalledFunction()->getName()))
1358 LoadInst *Ld = dyn_cast<LoadInst>(it);
1359 if (!Ld || (!Ld->isSimple() && !IsAnnotatedParallel)) {
1360 emitAnalysis(LoopAccessReport(Ld)
1361 << "read with atomic ordering or volatile read");
1362 DEBUG(dbgs() << "LAA: Found a non-simple load.\n");
1367 Loads.push_back(Ld);
1368 DepChecker.addAccess(Ld);
1372 // Save 'store' instructions. Abort if other instructions write to memory.
1373 if (it->mayWriteToMemory()) {
1374 StoreInst *St = dyn_cast<StoreInst>(it);
1376 emitAnalysis(LoopAccessReport(it) <<
1377 "instruction cannot be vectorized");
1381 if (!St->isSimple() && !IsAnnotatedParallel) {
1382 emitAnalysis(LoopAccessReport(St)
1383 << "write with atomic ordering or volatile write");
1384 DEBUG(dbgs() << "LAA: Found a non-simple store.\n");
1389 Stores.push_back(St);
1390 DepChecker.addAccess(St);
1395 // Now we have two lists that hold the loads and the stores.
1396 // Next, we find the pointers that they use.
1398 // Check if we see any stores. If there are no stores, then we don't
1399 // care if the pointers are *restrict*.
1400 if (!Stores.size()) {
1401 DEBUG(dbgs() << "LAA: Found a read-only loop!\n");
1406 MemoryDepChecker::DepCandidates DependentAccesses;
1407 AccessAnalysis Accesses(TheLoop->getHeader()->getModule()->getDataLayout(),
1408 AA, LI, DependentAccesses);
1410 // Holds the analyzed pointers. We don't want to call GetUnderlyingObjects
1411 // multiple times on the same object. If the ptr is accessed twice, once
1412 // for read and once for write, it will only appear once (on the write
1413 // list). This is okay, since we are going to check for conflicts between
1414 // writes and between reads and writes, but not between reads and reads.
1417 ValueVector::iterator I, IE;
1418 for (I = Stores.begin(), IE = Stores.end(); I != IE; ++I) {
1419 StoreInst *ST = cast<StoreInst>(*I);
1420 Value* Ptr = ST->getPointerOperand();
1421 // Check for store to loop invariant address.
1422 StoreToLoopInvariantAddress |= isUniform(Ptr);
1423 // If we did *not* see this pointer before, insert it to the read-write
1424 // list. At this phase it is only a 'write' list.
1425 if (Seen.insert(Ptr).second) {
1428 MemoryLocation Loc = MemoryLocation::get(ST);
1429 // The TBAA metadata could have a control dependency on the predication
1430 // condition, so we cannot rely on it when determining whether or not we
1431 // need runtime pointer checks.
1432 if (blockNeedsPredication(ST->getParent(), TheLoop, DT))
1433 Loc.AATags.TBAA = nullptr;
1435 Accesses.addStore(Loc);
1439 if (IsAnnotatedParallel) {
1441 << "LAA: A loop annotated parallel, ignore memory dependency "
1447 for (I = Loads.begin(), IE = Loads.end(); I != IE; ++I) {
1448 LoadInst *LD = cast<LoadInst>(*I);
1449 Value* Ptr = LD->getPointerOperand();
1450 // If we did *not* see this pointer before, insert it to the
1451 // read list. If we *did* see it before, then it is already in
1452 // the read-write list. This allows us to vectorize expressions
1453 // such as A[i] += x; Because the address of A[i] is a read-write
1454 // pointer. This only works if the index of A[i] is consecutive.
1455 // If the address of i is unknown (for example A[B[i]]) then we may
1456 // read a few words, modify, and write a few words, and some of the
1457 // words may be written to the same address.
1458 bool IsReadOnlyPtr = false;
1459 if (Seen.insert(Ptr).second || !isStridedPtr(SE, Ptr, TheLoop, Strides)) {
1461 IsReadOnlyPtr = true;
1464 MemoryLocation Loc = MemoryLocation::get(LD);
1465 // The TBAA metadata could have a control dependency on the predication
1466 // condition, so we cannot rely on it when determining whether or not we
1467 // need runtime pointer checks.
1468 if (blockNeedsPredication(LD->getParent(), TheLoop, DT))
1469 Loc.AATags.TBAA = nullptr;
1471 Accesses.addLoad(Loc, IsReadOnlyPtr);
1474 // If we write (or read-write) to a single destination and there are no
1475 // other reads in this loop then is it safe to vectorize.
1476 if (NumReadWrites == 1 && NumReads == 0) {
1477 DEBUG(dbgs() << "LAA: Found a write-only loop!\n");
1482 // Build dependence sets and check whether we need a runtime pointer bounds
1484 Accesses.buildDependenceSets();
1486 // Find pointers with computable bounds. We are going to use this information
1487 // to place a runtime bound check.
1488 bool CanDoRTIfNeeded =
1489 Accesses.canCheckPtrAtRT(PtrRtCheck, SE, TheLoop, Strides);
1490 if (!CanDoRTIfNeeded) {
1491 emitAnalysis(LoopAccessReport() << "cannot identify array bounds");
1492 DEBUG(dbgs() << "LAA: We can't vectorize because we can't find "
1493 << "the array bounds.\n");
1498 DEBUG(dbgs() << "LAA: We can perform a memory runtime check if needed.\n");
1501 if (Accesses.isDependencyCheckNeeded()) {
1502 DEBUG(dbgs() << "LAA: Checking memory dependencies\n");
1503 CanVecMem = DepChecker.areDepsSafe(
1504 DependentAccesses, Accesses.getDependenciesToCheck(), Strides);
1505 MaxSafeDepDistBytes = DepChecker.getMaxSafeDepDistBytes();
1507 if (!CanVecMem && DepChecker.shouldRetryWithRuntimeCheck()) {
1508 DEBUG(dbgs() << "LAA: Retrying with memory checks\n");
1510 // Clear the dependency checks. We assume they are not needed.
1511 Accesses.resetDepChecks(DepChecker);
1514 PtrRtCheck.Need = true;
1517 Accesses.canCheckPtrAtRT(PtrRtCheck, SE, TheLoop, Strides, true);
1519 // Check that we found the bounds for the pointer.
1520 if (!CanDoRTIfNeeded) {
1521 emitAnalysis(LoopAccessReport()
1522 << "cannot check memory dependencies at runtime");
1523 DEBUG(dbgs() << "LAA: Can't vectorize with memory checks\n");
1533 DEBUG(dbgs() << "LAA: No unsafe dependent memory operations in loop. We"
1534 << (PtrRtCheck.Need ? "" : " don't")
1535 << " need a runtime memory check.\n");
1537 emitAnalysis(LoopAccessReport() <<
1538 "unsafe dependent memory operations in loop");
1539 DEBUG(dbgs() << "LAA: unsafe dependent memory operations in loop\n");
1543 bool LoopAccessInfo::blockNeedsPredication(BasicBlock *BB, Loop *TheLoop,
1544 DominatorTree *DT) {
1545 assert(TheLoop->contains(BB) && "Unknown block used");
1547 // Blocks that do not dominate the latch need predication.
1548 BasicBlock* Latch = TheLoop->getLoopLatch();
1549 return !DT->dominates(BB, Latch);
1552 void LoopAccessInfo::emitAnalysis(LoopAccessReport &Message) {
1553 assert(!Report && "Multiple reports generated");
1557 bool LoopAccessInfo::isUniform(Value *V) const {
1558 return (SE->isLoopInvariant(SE->getSCEV(V), TheLoop));
1561 // FIXME: this function is currently a duplicate of the one in
1562 // LoopVectorize.cpp.
1563 static Instruction *getFirstInst(Instruction *FirstInst, Value *V,
1567 if (Instruction *I = dyn_cast<Instruction>(V))
1568 return I->getParent() == Loc->getParent() ? I : nullptr;
1572 std::pair<Instruction *, Instruction *> LoopAccessInfo::addRuntimeCheck(
1573 Instruction *Loc, const SmallVectorImpl<int> *PtrPartition) const {
1574 if (!PtrRtCheck.Need)
1575 return std::make_pair(nullptr, nullptr);
1577 SmallVector<TrackingVH<Value>, 2> Starts;
1578 SmallVector<TrackingVH<Value>, 2> Ends;
1580 LLVMContext &Ctx = Loc->getContext();
1581 SCEVExpander Exp(*SE, DL, "induction");
1582 Instruction *FirstInst = nullptr;
1584 for (unsigned i = 0; i < PtrRtCheck.CheckingGroups.size(); ++i) {
1585 const RuntimePointerCheck::CheckingPtrGroup &CG =
1586 PtrRtCheck.CheckingGroups[i];
1587 Value *Ptr = PtrRtCheck.Pointers[CG.Members[0]];
1588 const SCEV *Sc = SE->getSCEV(Ptr);
1590 if (SE->isLoopInvariant(Sc, TheLoop)) {
1591 DEBUG(dbgs() << "LAA: Adding RT check for a loop invariant ptr:" << *Ptr
1593 Starts.push_back(Ptr);
1594 Ends.push_back(Ptr);
1596 unsigned AS = Ptr->getType()->getPointerAddressSpace();
1598 // Use this type for pointer arithmetic.
1599 Type *PtrArithTy = Type::getInt8PtrTy(Ctx, AS);
1600 Value *Start = nullptr, *End = nullptr;
1602 DEBUG(dbgs() << "LAA: Adding RT check for range:\n");
1603 Start = Exp.expandCodeFor(CG.Low, PtrArithTy, Loc);
1604 End = Exp.expandCodeFor(CG.High, PtrArithTy, Loc);
1605 DEBUG(dbgs() << "Start: " << *CG.Low << " End: " << *CG.High << "\n");
1606 Starts.push_back(Start);
1607 Ends.push_back(End);
1611 IRBuilder<> ChkBuilder(Loc);
1612 // Our instructions might fold to a constant.
1613 Value *MemoryRuntimeCheck = nullptr;
1614 for (unsigned i = 0; i < PtrRtCheck.CheckingGroups.size(); ++i) {
1615 for (unsigned j = i + 1; j < PtrRtCheck.CheckingGroups.size(); ++j) {
1616 const RuntimePointerCheck::CheckingPtrGroup &CGI =
1617 PtrRtCheck.CheckingGroups[i];
1618 const RuntimePointerCheck::CheckingPtrGroup &CGJ =
1619 PtrRtCheck.CheckingGroups[j];
1621 if (!PtrRtCheck.needsChecking(CGI, CGJ, PtrPartition))
1624 unsigned AS0 = Starts[i]->getType()->getPointerAddressSpace();
1625 unsigned AS1 = Starts[j]->getType()->getPointerAddressSpace();
1627 assert((AS0 == Ends[j]->getType()->getPointerAddressSpace()) &&
1628 (AS1 == Ends[i]->getType()->getPointerAddressSpace()) &&
1629 "Trying to bounds check pointers with different address spaces");
1631 Type *PtrArithTy0 = Type::getInt8PtrTy(Ctx, AS0);
1632 Type *PtrArithTy1 = Type::getInt8PtrTy(Ctx, AS1);
1634 Value *Start0 = ChkBuilder.CreateBitCast(Starts[i], PtrArithTy0, "bc");
1635 Value *Start1 = ChkBuilder.CreateBitCast(Starts[j], PtrArithTy1, "bc");
1636 Value *End0 = ChkBuilder.CreateBitCast(Ends[i], PtrArithTy1, "bc");
1637 Value *End1 = ChkBuilder.CreateBitCast(Ends[j], PtrArithTy0, "bc");
1639 Value *Cmp0 = ChkBuilder.CreateICmpULE(Start0, End1, "bound0");
1640 FirstInst = getFirstInst(FirstInst, Cmp0, Loc);
1641 Value *Cmp1 = ChkBuilder.CreateICmpULE(Start1, End0, "bound1");
1642 FirstInst = getFirstInst(FirstInst, Cmp1, Loc);
1643 Value *IsConflict = ChkBuilder.CreateAnd(Cmp0, Cmp1, "found.conflict");
1644 FirstInst = getFirstInst(FirstInst, IsConflict, Loc);
1645 if (MemoryRuntimeCheck) {
1646 IsConflict = ChkBuilder.CreateOr(MemoryRuntimeCheck, IsConflict,
1648 FirstInst = getFirstInst(FirstInst, IsConflict, Loc);
1650 MemoryRuntimeCheck = IsConflict;
1654 if (!MemoryRuntimeCheck)
1655 return std::make_pair(nullptr, nullptr);
1657 // We have to do this trickery because the IRBuilder might fold the check to a
1658 // constant expression in which case there is no Instruction anchored in a
1660 Instruction *Check = BinaryOperator::CreateAnd(MemoryRuntimeCheck,
1661 ConstantInt::getTrue(Ctx));
1662 ChkBuilder.Insert(Check, "memcheck.conflict");
1663 FirstInst = getFirstInst(FirstInst, Check, Loc);
1664 return std::make_pair(FirstInst, Check);
1667 LoopAccessInfo::LoopAccessInfo(Loop *L, ScalarEvolution *SE,
1668 const DataLayout &DL,
1669 const TargetLibraryInfo *TLI, AliasAnalysis *AA,
1670 DominatorTree *DT, LoopInfo *LI,
1671 const ValueToValueMap &Strides)
1672 : PtrRtCheck(SE), DepChecker(SE, L), TheLoop(L), SE(SE), DL(DL), TLI(TLI),
1673 AA(AA), DT(DT), LI(LI), NumLoads(0), NumStores(0),
1674 MaxSafeDepDistBytes(-1U), CanVecMem(false),
1675 StoreToLoopInvariantAddress(false) {
1676 if (canAnalyzeLoop())
1677 analyzeLoop(Strides);
1680 void LoopAccessInfo::print(raw_ostream &OS, unsigned Depth) const {
1682 if (PtrRtCheck.Need)
1683 OS.indent(Depth) << "Memory dependences are safe with run-time checks\n";
1685 OS.indent(Depth) << "Memory dependences are safe\n";
1689 OS.indent(Depth) << "Report: " << Report->str() << "\n";
1691 if (auto *InterestingDependences = DepChecker.getInterestingDependences()) {
1692 OS.indent(Depth) << "Interesting Dependences:\n";
1693 for (auto &Dep : *InterestingDependences) {
1694 Dep.print(OS, Depth + 2, DepChecker.getMemoryInstructions());
1698 OS.indent(Depth) << "Too many interesting dependences, not recorded\n";
1700 // List the pair of accesses need run-time checks to prove independence.
1701 PtrRtCheck.print(OS, Depth);
1704 OS.indent(Depth) << "Store to invariant address was "
1705 << (StoreToLoopInvariantAddress ? "" : "not ")
1706 << "found in loop.\n";
1709 const LoopAccessInfo &
1710 LoopAccessAnalysis::getInfo(Loop *L, const ValueToValueMap &Strides) {
1711 auto &LAI = LoopAccessInfoMap[L];
1714 assert((!LAI || LAI->NumSymbolicStrides == Strides.size()) &&
1715 "Symbolic strides changed for loop");
1719 const DataLayout &DL = L->getHeader()->getModule()->getDataLayout();
1720 LAI = llvm::make_unique<LoopAccessInfo>(L, SE, DL, TLI, AA, DT, LI,
1723 LAI->NumSymbolicStrides = Strides.size();
1729 void LoopAccessAnalysis::print(raw_ostream &OS, const Module *M) const {
1730 LoopAccessAnalysis &LAA = *const_cast<LoopAccessAnalysis *>(this);
1732 ValueToValueMap NoSymbolicStrides;
1734 for (Loop *TopLevelLoop : *LI)
1735 for (Loop *L : depth_first(TopLevelLoop)) {
1736 OS.indent(2) << L->getHeader()->getName() << ":\n";
1737 auto &LAI = LAA.getInfo(L, NoSymbolicStrides);
1742 bool LoopAccessAnalysis::runOnFunction(Function &F) {
1743 SE = &getAnalysis<ScalarEvolution>();
1744 auto *TLIP = getAnalysisIfAvailable<TargetLibraryInfoWrapperPass>();
1745 TLI = TLIP ? &TLIP->getTLI() : nullptr;
1746 AA = &getAnalysis<AliasAnalysis>();
1747 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1748 LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
1753 void LoopAccessAnalysis::getAnalysisUsage(AnalysisUsage &AU) const {
1754 AU.addRequired<ScalarEvolution>();
1755 AU.addRequired<AliasAnalysis>();
1756 AU.addRequired<DominatorTreeWrapperPass>();
1757 AU.addRequired<LoopInfoWrapperPass>();
1759 AU.setPreservesAll();
1762 char LoopAccessAnalysis::ID = 0;
1763 static const char laa_name[] = "Loop Access Analysis";
1764 #define LAA_NAME "loop-accesses"
1766 INITIALIZE_PASS_BEGIN(LoopAccessAnalysis, LAA_NAME, laa_name, false, true)
1767 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
1768 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
1769 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
1770 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
1771 INITIALIZE_PASS_END(LoopAccessAnalysis, LAA_NAME, laa_name, false, true)
1774 Pass *createLAAPass() {
1775 return new LoopAccessAnalysis();