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 // Go through all equivalence classes, get the the "pointer check groups"
226 // and add them to the overall solution.
227 for (auto DI = DepCands.begin(), DE = DepCands.end(); DI != DE; ++DI) {
231 SmallVector<CheckingPtrGroup, 2> Groups;
233 for (auto MI = DepCands.member_begin(DI), ME = DepCands.member_end();
235 unsigned Pointer = PositionMap[MI->getPointer()];
238 // Go through all the existing sets and see if we can find one
239 // which can include this pointer.
240 for (CheckingPtrGroup &Group : Groups) {
241 // Don't perform more than a certain amount of comparisons.
242 // This should limit the cost of grouping the pointers to something
243 // reasonable. If we do end up hitting this threshold, the algorithm
244 // will create separate groups for all remaining pointers.
245 if (TotalComparisons > MemoryCheckMergeThreshold)
250 if (Group.addPointer(Pointer)) {
257 // We couldn't add this pointer to any existing set or the threshold
258 // for the number of comparisons has been reached. Create a new group
259 // to hold the current pointer.
260 Groups.push_back(CheckingPtrGroup(Pointer, *this));
263 // We've computed the grouped checks for this partition.
264 // Save the results and continue with the next one.
265 std::copy(Groups.begin(), Groups.end(), std::back_inserter(CheckingGroups));
269 bool LoopAccessInfo::RuntimePointerCheck::needsChecking(
270 unsigned I, unsigned J, const SmallVectorImpl<int> *PtrPartition) const {
271 // No need to check if two readonly pointers intersect.
272 if (!IsWritePtr[I] && !IsWritePtr[J])
275 // Only need to check pointers between two different dependency sets.
276 if (DependencySetId[I] == DependencySetId[J])
279 // Only need to check pointers in the same alias set.
280 if (AliasSetId[I] != AliasSetId[J])
283 // If PtrPartition is set omit checks between pointers of the same partition.
284 // Partition number -1 means that the pointer is used in multiple partitions.
285 // In this case we can't omit the check.
286 if (PtrPartition && (*PtrPartition)[I] != -1 &&
287 (*PtrPartition)[I] == (*PtrPartition)[J])
293 void LoopAccessInfo::RuntimePointerCheck::print(
294 raw_ostream &OS, unsigned Depth,
295 const SmallVectorImpl<int> *PtrPartition) const {
297 OS.indent(Depth) << "Run-time memory checks:\n";
300 for (unsigned I = 0; I < CheckingGroups.size(); ++I)
301 for (unsigned J = I + 1; J < CheckingGroups.size(); ++J)
302 if (needsChecking(CheckingGroups[I], CheckingGroups[J], PtrPartition)) {
303 OS.indent(Depth) << "Check " << N++ << ":\n";
304 OS.indent(Depth + 2) << "Comparing group " << I << ":\n";
306 for (unsigned K = 0; K < CheckingGroups[I].Members.size(); ++K) {
307 OS.indent(Depth + 2) << *Pointers[CheckingGroups[I].Members[K]]
310 OS << " (Partition: "
311 << (*PtrPartition)[CheckingGroups[I].Members[K]] << ")"
315 OS.indent(Depth + 2) << "Against group " << J << ":\n";
317 for (unsigned K = 0; K < CheckingGroups[J].Members.size(); ++K) {
318 OS.indent(Depth + 2) << *Pointers[CheckingGroups[J].Members[K]]
321 OS << " (Partition: "
322 << (*PtrPartition)[CheckingGroups[J].Members[K]] << ")"
327 OS.indent(Depth) << "Grouped accesses:\n";
328 for (unsigned I = 0; I < CheckingGroups.size(); ++I) {
329 OS.indent(Depth + 2) << "Group " << I << ":\n";
330 OS.indent(Depth + 4) << "(Low: " << *CheckingGroups[I].Low
331 << " High: " << *CheckingGroups[I].High << ")\n";
332 for (unsigned J = 0; J < CheckingGroups[I].Members.size(); ++J) {
333 OS.indent(Depth + 6) << "Member: " << *Exprs[CheckingGroups[I].Members[J]]
339 unsigned LoopAccessInfo::RuntimePointerCheck::getNumberOfChecks(
340 const SmallVectorImpl<int> *PtrPartition) const {
342 unsigned NumPartitions = CheckingGroups.size();
343 unsigned CheckCount = 0;
345 for (unsigned I = 0; I < NumPartitions; ++I)
346 for (unsigned J = I + 1; J < NumPartitions; ++J)
347 if (needsChecking(CheckingGroups[I], CheckingGroups[J], PtrPartition))
352 bool LoopAccessInfo::RuntimePointerCheck::needsAnyChecking(
353 const SmallVectorImpl<int> *PtrPartition) const {
354 unsigned NumPointers = Pointers.size();
356 for (unsigned I = 0; I < NumPointers; ++I)
357 for (unsigned J = I + 1; J < NumPointers; ++J)
358 if (needsChecking(I, J, PtrPartition))
364 /// \brief Analyses memory accesses in a loop.
366 /// Checks whether run time pointer checks are needed and builds sets for data
367 /// dependence checking.
368 class AccessAnalysis {
370 /// \brief Read or write access location.
371 typedef PointerIntPair<Value *, 1, bool> MemAccessInfo;
372 typedef SmallPtrSet<MemAccessInfo, 8> MemAccessInfoSet;
374 AccessAnalysis(const DataLayout &Dl, AliasAnalysis *AA, LoopInfo *LI,
375 MemoryDepChecker::DepCandidates &DA)
376 : DL(Dl), AST(*AA), LI(LI), DepCands(DA),
377 IsRTCheckAnalysisNeeded(false) {}
379 /// \brief Register a load and whether it is only read from.
380 void addLoad(MemoryLocation &Loc, bool IsReadOnly) {
381 Value *Ptr = const_cast<Value*>(Loc.Ptr);
382 AST.add(Ptr, MemoryLocation::UnknownSize, Loc.AATags);
383 Accesses.insert(MemAccessInfo(Ptr, false));
385 ReadOnlyPtr.insert(Ptr);
388 /// \brief Register a store.
389 void addStore(MemoryLocation &Loc) {
390 Value *Ptr = const_cast<Value*>(Loc.Ptr);
391 AST.add(Ptr, MemoryLocation::UnknownSize, Loc.AATags);
392 Accesses.insert(MemAccessInfo(Ptr, true));
395 /// \brief Check whether we can check the pointers at runtime for
396 /// non-intersection. Returns true when we have 0 pointers
397 /// (a check on 0 pointers for non-intersection will always return true).
398 bool canCheckPtrAtRT(LoopAccessInfo::RuntimePointerCheck &RtCheck,
399 bool &NeedRTCheck, ScalarEvolution *SE, Loop *TheLoop,
400 const ValueToValueMap &Strides,
401 bool ShouldCheckStride = false);
403 /// \brief Goes over all memory accesses, checks whether a RT check is needed
404 /// and builds sets of dependent accesses.
405 void buildDependenceSets() {
406 processMemAccesses();
409 /// \brief Initial processing of memory accesses determined that we need to
410 /// perform dependency checking.
412 /// Note that this can later be cleared if we retry memcheck analysis without
413 /// dependency checking (i.e. ShouldRetryWithRuntimeCheck).
414 bool isDependencyCheckNeeded() { return !CheckDeps.empty(); }
416 /// We decided that no dependence analysis would be used. Reset the state.
417 void resetDepChecks(MemoryDepChecker &DepChecker) {
419 DepChecker.clearInterestingDependences();
422 MemAccessInfoSet &getDependenciesToCheck() { return CheckDeps; }
425 typedef SetVector<MemAccessInfo> PtrAccessSet;
427 /// \brief Go over all memory access and check whether runtime pointer checks
428 /// are needed and build sets of dependency check candidates.
429 void processMemAccesses();
431 /// Set of all accesses.
432 PtrAccessSet Accesses;
434 const DataLayout &DL;
436 /// Set of accesses that need a further dependence check.
437 MemAccessInfoSet CheckDeps;
439 /// Set of pointers that are read only.
440 SmallPtrSet<Value*, 16> ReadOnlyPtr;
442 /// An alias set tracker to partition the access set by underlying object and
443 //intrinsic property (such as TBAA metadata).
448 /// Sets of potentially dependent accesses - members of one set share an
449 /// underlying pointer. The set "CheckDeps" identfies which sets really need a
450 /// dependence check.
451 MemoryDepChecker::DepCandidates &DepCands;
453 /// \brief Initial processing of memory accesses determined that we may need
454 /// to add memchecks. Perform the analysis to determine the necessary checks.
456 /// Note that, this is different from isDependencyCheckNeeded. When we retry
457 /// memcheck analysis without dependency checking
458 /// (i.e. ShouldRetryWithRuntimeCheck), isDependencyCheckNeeded is cleared
459 /// while this remains set if we have potentially dependent accesses.
460 bool IsRTCheckAnalysisNeeded;
463 } // end anonymous namespace
465 /// \brief Check whether a pointer can participate in a runtime bounds check.
466 static bool hasComputableBounds(ScalarEvolution *SE,
467 const ValueToValueMap &Strides, Value *Ptr) {
468 const SCEV *PtrScev = replaceSymbolicStrideSCEV(SE, Strides, Ptr);
469 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
473 return AR->isAffine();
476 bool AccessAnalysis::canCheckPtrAtRT(
477 LoopAccessInfo::RuntimePointerCheck &RtCheck, bool &NeedRTCheck,
478 ScalarEvolution *SE, Loop *TheLoop, const ValueToValueMap &StridesMap,
479 bool ShouldCheckStride) {
480 // Find pointers with computable bounds. We are going to use this information
481 // to place a runtime bound check.
485 if (!IsRTCheckAnalysisNeeded) return true;
487 bool IsDepCheckNeeded = isDependencyCheckNeeded();
489 // We assign a consecutive id to access from different alias sets.
490 // Accesses between different groups doesn't need to be checked.
492 for (auto &AS : AST) {
493 int NumReadPtrChecks = 0;
494 int NumWritePtrChecks = 0;
496 // We assign consecutive id to access from different dependence sets.
497 // Accesses within the same set don't need a runtime check.
498 unsigned RunningDepId = 1;
499 DenseMap<Value *, unsigned> DepSetId;
502 Value *Ptr = A.getValue();
503 bool IsWrite = Accesses.count(MemAccessInfo(Ptr, true));
504 MemAccessInfo Access(Ptr, IsWrite);
511 if (hasComputableBounds(SE, StridesMap, Ptr) &&
512 // When we run after a failing dependency check we have to make sure
513 // we don't have wrapping pointers.
514 (!ShouldCheckStride ||
515 isStridedPtr(SE, Ptr, TheLoop, StridesMap) == 1)) {
516 // The id of the dependence set.
519 if (IsDepCheckNeeded) {
520 Value *Leader = DepCands.getLeaderValue(Access).getPointer();
521 unsigned &LeaderId = DepSetId[Leader];
523 LeaderId = RunningDepId++;
526 // Each access has its own dependence set.
527 DepId = RunningDepId++;
529 RtCheck.insert(TheLoop, Ptr, IsWrite, DepId, ASId, StridesMap);
531 DEBUG(dbgs() << "LAA: Found a runtime check ptr:" << *Ptr << '\n');
533 DEBUG(dbgs() << "LAA: Can't find bounds for ptr:" << *Ptr << '\n');
538 // If we have at least two writes or one write and a read then we need to
539 // check them. But there is no need to checks if there is only one
540 // dependence set for this alias set.
542 // Note that this function computes CanDoRT and NeedRTCheck independently.
543 // For example CanDoRT=false, NeedRTCheck=false means that we have a pointer
544 // for which we couldn't find the bounds but we don't actually need to emit
545 // any checks so it does not matter.
546 if (!(IsDepCheckNeeded && CanDoRT && RunningDepId == 2))
547 NeedRTCheck |= (NumWritePtrChecks >= 2 || (NumReadPtrChecks >= 1 &&
548 NumWritePtrChecks >= 1));
553 // If the pointers that we would use for the bounds comparison have different
554 // address spaces, assume the values aren't directly comparable, so we can't
555 // use them for the runtime check. We also have to assume they could
556 // overlap. In the future there should be metadata for whether address spaces
558 unsigned NumPointers = RtCheck.Pointers.size();
559 for (unsigned i = 0; i < NumPointers; ++i) {
560 for (unsigned j = i + 1; j < NumPointers; ++j) {
561 // Only need to check pointers between two different dependency sets.
562 if (RtCheck.DependencySetId[i] == RtCheck.DependencySetId[j])
564 // Only need to check pointers in the same alias set.
565 if (RtCheck.AliasSetId[i] != RtCheck.AliasSetId[j])
568 Value *PtrI = RtCheck.Pointers[i];
569 Value *PtrJ = RtCheck.Pointers[j];
571 unsigned ASi = PtrI->getType()->getPointerAddressSpace();
572 unsigned ASj = PtrJ->getType()->getPointerAddressSpace();
574 DEBUG(dbgs() << "LAA: Runtime check would require comparison between"
575 " different address spaces\n");
581 if (NeedRTCheck && CanDoRT)
582 RtCheck.groupChecks(DepCands, IsDepCheckNeeded);
587 void AccessAnalysis::processMemAccesses() {
588 // We process the set twice: first we process read-write pointers, last we
589 // process read-only pointers. This allows us to skip dependence tests for
590 // read-only pointers.
592 DEBUG(dbgs() << "LAA: Processing memory accesses...\n");
593 DEBUG(dbgs() << " AST: "; AST.dump());
594 DEBUG(dbgs() << "LAA: Accesses(" << Accesses.size() << "):\n");
596 for (auto A : Accesses)
597 dbgs() << "\t" << *A.getPointer() << " (" <<
598 (A.getInt() ? "write" : (ReadOnlyPtr.count(A.getPointer()) ?
599 "read-only" : "read")) << ")\n";
602 // The AliasSetTracker has nicely partitioned our pointers by metadata
603 // compatibility and potential for underlying-object overlap. As a result, we
604 // only need to check for potential pointer dependencies within each alias
606 for (auto &AS : AST) {
607 // Note that both the alias-set tracker and the alias sets themselves used
608 // linked lists internally and so the iteration order here is deterministic
609 // (matching the original instruction order within each set).
611 bool SetHasWrite = false;
613 // Map of pointers to last access encountered.
614 typedef DenseMap<Value*, MemAccessInfo> UnderlyingObjToAccessMap;
615 UnderlyingObjToAccessMap ObjToLastAccess;
617 // Set of access to check after all writes have been processed.
618 PtrAccessSet DeferredAccesses;
620 // Iterate over each alias set twice, once to process read/write pointers,
621 // and then to process read-only pointers.
622 for (int SetIteration = 0; SetIteration < 2; ++SetIteration) {
623 bool UseDeferred = SetIteration > 0;
624 PtrAccessSet &S = UseDeferred ? DeferredAccesses : Accesses;
627 Value *Ptr = AV.getValue();
629 // For a single memory access in AliasSetTracker, Accesses may contain
630 // both read and write, and they both need to be handled for CheckDeps.
632 if (AC.getPointer() != Ptr)
635 bool IsWrite = AC.getInt();
637 // If we're using the deferred access set, then it contains only
639 bool IsReadOnlyPtr = ReadOnlyPtr.count(Ptr) && !IsWrite;
640 if (UseDeferred && !IsReadOnlyPtr)
642 // Otherwise, the pointer must be in the PtrAccessSet, either as a
644 assert(((IsReadOnlyPtr && UseDeferred) || IsWrite ||
645 S.count(MemAccessInfo(Ptr, false))) &&
646 "Alias-set pointer not in the access set?");
648 MemAccessInfo Access(Ptr, IsWrite);
649 DepCands.insert(Access);
651 // Memorize read-only pointers for later processing and skip them in
652 // the first round (they need to be checked after we have seen all
653 // write pointers). Note: we also mark pointer that are not
654 // consecutive as "read-only" pointers (so that we check
655 // "a[b[i]] +="). Hence, we need the second check for "!IsWrite".
656 if (!UseDeferred && IsReadOnlyPtr) {
657 DeferredAccesses.insert(Access);
661 // If this is a write - check other reads and writes for conflicts. If
662 // this is a read only check other writes for conflicts (but only if
663 // there is no other write to the ptr - this is an optimization to
664 // catch "a[i] = a[i] + " without having to do a dependence check).
665 if ((IsWrite || IsReadOnlyPtr) && SetHasWrite) {
666 CheckDeps.insert(Access);
667 IsRTCheckAnalysisNeeded = true;
673 // Create sets of pointers connected by a shared alias set and
674 // underlying object.
675 typedef SmallVector<Value *, 16> ValueVector;
676 ValueVector TempObjects;
678 GetUnderlyingObjects(Ptr, TempObjects, DL, LI);
679 DEBUG(dbgs() << "Underlying objects for pointer " << *Ptr << "\n");
680 for (Value *UnderlyingObj : TempObjects) {
681 UnderlyingObjToAccessMap::iterator Prev =
682 ObjToLastAccess.find(UnderlyingObj);
683 if (Prev != ObjToLastAccess.end())
684 DepCands.unionSets(Access, Prev->second);
686 ObjToLastAccess[UnderlyingObj] = Access;
687 DEBUG(dbgs() << " " << *UnderlyingObj << "\n");
695 static bool isInBoundsGep(Value *Ptr) {
696 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr))
697 return GEP->isInBounds();
701 /// \brief Return true if an AddRec pointer \p Ptr is unsigned non-wrapping,
702 /// i.e. monotonically increasing/decreasing.
703 static bool isNoWrapAddRec(Value *Ptr, const SCEVAddRecExpr *AR,
704 ScalarEvolution *SE, const Loop *L) {
705 // FIXME: This should probably only return true for NUW.
706 if (AR->getNoWrapFlags(SCEV::NoWrapMask))
709 // Scalar evolution does not propagate the non-wrapping flags to values that
710 // are derived from a non-wrapping induction variable because non-wrapping
711 // could be flow-sensitive.
713 // Look through the potentially overflowing instruction to try to prove
714 // non-wrapping for the *specific* value of Ptr.
716 // The arithmetic implied by an inbounds GEP can't overflow.
717 auto *GEP = dyn_cast<GetElementPtrInst>(Ptr);
718 if (!GEP || !GEP->isInBounds())
721 // Make sure there is only one non-const index and analyze that.
722 Value *NonConstIndex = nullptr;
723 for (auto Index = GEP->idx_begin(); Index != GEP->idx_end(); ++Index)
724 if (!isa<ConstantInt>(*Index)) {
727 NonConstIndex = *Index;
730 // The recurrence is on the pointer, ignore for now.
733 // The index in GEP is signed. It is non-wrapping if it's derived from a NSW
734 // AddRec using a NSW operation.
735 if (auto *OBO = dyn_cast<OverflowingBinaryOperator>(NonConstIndex))
736 if (OBO->hasNoSignedWrap() &&
737 // Assume constant for other the operand so that the AddRec can be
739 isa<ConstantInt>(OBO->getOperand(1))) {
740 auto *OpScev = SE->getSCEV(OBO->getOperand(0));
742 if (auto *OpAR = dyn_cast<SCEVAddRecExpr>(OpScev))
743 return OpAR->getLoop() == L && OpAR->getNoWrapFlags(SCEV::FlagNSW);
749 /// \brief Check whether the access through \p Ptr has a constant stride.
750 int llvm::isStridedPtr(ScalarEvolution *SE, Value *Ptr, const Loop *Lp,
751 const ValueToValueMap &StridesMap) {
752 const Type *Ty = Ptr->getType();
753 assert(Ty->isPointerTy() && "Unexpected non-ptr");
755 // Make sure that the pointer does not point to aggregate types.
756 const PointerType *PtrTy = cast<PointerType>(Ty);
757 if (PtrTy->getElementType()->isAggregateType()) {
758 DEBUG(dbgs() << "LAA: Bad stride - Not a pointer to a scalar type"
763 const SCEV *PtrScev = replaceSymbolicStrideSCEV(SE, StridesMap, Ptr);
765 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
767 DEBUG(dbgs() << "LAA: Bad stride - Not an AddRecExpr pointer "
768 << *Ptr << " SCEV: " << *PtrScev << "\n");
772 // The accesss function must stride over the innermost loop.
773 if (Lp != AR->getLoop()) {
774 DEBUG(dbgs() << "LAA: Bad stride - Not striding over innermost loop " <<
775 *Ptr << " SCEV: " << *PtrScev << "\n");
778 // The address calculation must not wrap. Otherwise, a dependence could be
780 // An inbounds getelementptr that is a AddRec with a unit stride
781 // cannot wrap per definition. The unit stride requirement is checked later.
782 // An getelementptr without an inbounds attribute and unit stride would have
783 // to access the pointer value "0" which is undefined behavior in address
784 // space 0, therefore we can also vectorize this case.
785 bool IsInBoundsGEP = isInBoundsGep(Ptr);
786 bool IsNoWrapAddRec = isNoWrapAddRec(Ptr, AR, SE, Lp);
787 bool IsInAddressSpaceZero = PtrTy->getAddressSpace() == 0;
788 if (!IsNoWrapAddRec && !IsInBoundsGEP && !IsInAddressSpaceZero) {
789 DEBUG(dbgs() << "LAA: Bad stride - Pointer may wrap in the address space "
790 << *Ptr << " SCEV: " << *PtrScev << "\n");
794 // Check the step is constant.
795 const SCEV *Step = AR->getStepRecurrence(*SE);
797 // Calculate the pointer stride and check if it is constant.
798 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
800 DEBUG(dbgs() << "LAA: Bad stride - Not a constant strided " << *Ptr <<
801 " SCEV: " << *PtrScev << "\n");
805 auto &DL = Lp->getHeader()->getModule()->getDataLayout();
806 int64_t Size = DL.getTypeAllocSize(PtrTy->getElementType());
807 const APInt &APStepVal = C->getValue()->getValue();
809 // Huge step value - give up.
810 if (APStepVal.getBitWidth() > 64)
813 int64_t StepVal = APStepVal.getSExtValue();
816 int64_t Stride = StepVal / Size;
817 int64_t Rem = StepVal % Size;
821 // If the SCEV could wrap but we have an inbounds gep with a unit stride we
822 // know we can't "wrap around the address space". In case of address space
823 // zero we know that this won't happen without triggering undefined behavior.
824 if (!IsNoWrapAddRec && (IsInBoundsGEP || IsInAddressSpaceZero) &&
825 Stride != 1 && Stride != -1)
831 bool MemoryDepChecker::Dependence::isSafeForVectorization(DepType Type) {
835 case BackwardVectorizable:
839 case ForwardButPreventsForwarding:
841 case BackwardVectorizableButPreventsForwarding:
844 llvm_unreachable("unexpected DepType!");
847 bool MemoryDepChecker::Dependence::isInterestingDependence(DepType Type) {
853 case BackwardVectorizable:
855 case ForwardButPreventsForwarding:
857 case BackwardVectorizableButPreventsForwarding:
860 llvm_unreachable("unexpected DepType!");
863 bool MemoryDepChecker::Dependence::isPossiblyBackward() const {
867 case ForwardButPreventsForwarding:
871 case BackwardVectorizable:
873 case BackwardVectorizableButPreventsForwarding:
876 llvm_unreachable("unexpected DepType!");
879 bool MemoryDepChecker::couldPreventStoreLoadForward(unsigned Distance,
880 unsigned TypeByteSize) {
881 // If loads occur at a distance that is not a multiple of a feasible vector
882 // factor store-load forwarding does not take place.
883 // Positive dependences might cause troubles because vectorizing them might
884 // prevent store-load forwarding making vectorized code run a lot slower.
885 // a[i] = a[i-3] ^ a[i-8];
886 // The stores to a[i:i+1] don't align with the stores to a[i-3:i-2] and
887 // hence on your typical architecture store-load forwarding does not take
888 // place. Vectorizing in such cases does not make sense.
889 // Store-load forwarding distance.
890 const unsigned NumCyclesForStoreLoadThroughMemory = 8*TypeByteSize;
891 // Maximum vector factor.
892 unsigned MaxVFWithoutSLForwardIssues =
893 VectorizerParams::MaxVectorWidth * TypeByteSize;
894 if(MaxSafeDepDistBytes < MaxVFWithoutSLForwardIssues)
895 MaxVFWithoutSLForwardIssues = MaxSafeDepDistBytes;
897 for (unsigned vf = 2*TypeByteSize; vf <= MaxVFWithoutSLForwardIssues;
899 if (Distance % vf && Distance / vf < NumCyclesForStoreLoadThroughMemory) {
900 MaxVFWithoutSLForwardIssues = (vf >>=1);
905 if (MaxVFWithoutSLForwardIssues< 2*TypeByteSize) {
906 DEBUG(dbgs() << "LAA: Distance " << Distance <<
907 " that could cause a store-load forwarding conflict\n");
911 if (MaxVFWithoutSLForwardIssues < MaxSafeDepDistBytes &&
912 MaxVFWithoutSLForwardIssues !=
913 VectorizerParams::MaxVectorWidth * TypeByteSize)
914 MaxSafeDepDistBytes = MaxVFWithoutSLForwardIssues;
918 /// \brief Check the dependence for two accesses with the same stride \p Stride.
919 /// \p Distance is the positive distance and \p TypeByteSize is type size in
922 /// \returns true if they are independent.
923 static bool areStridedAccessesIndependent(unsigned Distance, unsigned Stride,
924 unsigned TypeByteSize) {
925 assert(Stride > 1 && "The stride must be greater than 1");
926 assert(TypeByteSize > 0 && "The type size in byte must be non-zero");
927 assert(Distance > 0 && "The distance must be non-zero");
929 // Skip if the distance is not multiple of type byte size.
930 if (Distance % TypeByteSize)
933 unsigned ScaledDist = Distance / TypeByteSize;
935 // No dependence if the scaled distance is not multiple of the stride.
937 // for (i = 0; i < 1024 ; i += 4)
938 // A[i+2] = A[i] + 1;
940 // Two accesses in memory (scaled distance is 2, stride is 4):
941 // | A[0] | | | | A[4] | | | |
942 // | | | A[2] | | | | A[6] | |
945 // for (i = 0; i < 1024 ; i += 3)
946 // A[i+4] = A[i] + 1;
948 // Two accesses in memory (scaled distance is 4, stride is 3):
949 // | A[0] | | | A[3] | | | A[6] | | |
950 // | | | | | A[4] | | | A[7] | |
951 return ScaledDist % Stride;
954 MemoryDepChecker::Dependence::DepType
955 MemoryDepChecker::isDependent(const MemAccessInfo &A, unsigned AIdx,
956 const MemAccessInfo &B, unsigned BIdx,
957 const ValueToValueMap &Strides) {
958 assert (AIdx < BIdx && "Must pass arguments in program order");
960 Value *APtr = A.getPointer();
961 Value *BPtr = B.getPointer();
962 bool AIsWrite = A.getInt();
963 bool BIsWrite = B.getInt();
965 // Two reads are independent.
966 if (!AIsWrite && !BIsWrite)
967 return Dependence::NoDep;
969 // We cannot check pointers in different address spaces.
970 if (APtr->getType()->getPointerAddressSpace() !=
971 BPtr->getType()->getPointerAddressSpace())
972 return Dependence::Unknown;
974 const SCEV *AScev = replaceSymbolicStrideSCEV(SE, Strides, APtr);
975 const SCEV *BScev = replaceSymbolicStrideSCEV(SE, Strides, BPtr);
977 int StrideAPtr = isStridedPtr(SE, APtr, InnermostLoop, Strides);
978 int StrideBPtr = isStridedPtr(SE, BPtr, InnermostLoop, Strides);
980 const SCEV *Src = AScev;
981 const SCEV *Sink = BScev;
983 // If the induction step is negative we have to invert source and sink of the
985 if (StrideAPtr < 0) {
988 std::swap(APtr, BPtr);
989 std::swap(Src, Sink);
990 std::swap(AIsWrite, BIsWrite);
991 std::swap(AIdx, BIdx);
992 std::swap(StrideAPtr, StrideBPtr);
995 const SCEV *Dist = SE->getMinusSCEV(Sink, Src);
997 DEBUG(dbgs() << "LAA: Src Scev: " << *Src << "Sink Scev: " << *Sink
998 << "(Induction step: " << StrideAPtr << ")\n");
999 DEBUG(dbgs() << "LAA: Distance for " << *InstMap[AIdx] << " to "
1000 << *InstMap[BIdx] << ": " << *Dist << "\n");
1002 // Need accesses with constant stride. We don't want to vectorize
1003 // "A[B[i]] += ..." and similar code or pointer arithmetic that could wrap in
1004 // the address space.
1005 if (!StrideAPtr || !StrideBPtr || StrideAPtr != StrideBPtr){
1006 DEBUG(dbgs() << "Pointer access with non-constant stride\n");
1007 return Dependence::Unknown;
1010 const SCEVConstant *C = dyn_cast<SCEVConstant>(Dist);
1012 DEBUG(dbgs() << "LAA: Dependence because of non-constant distance\n");
1013 ShouldRetryWithRuntimeCheck = true;
1014 return Dependence::Unknown;
1017 Type *ATy = APtr->getType()->getPointerElementType();
1018 Type *BTy = BPtr->getType()->getPointerElementType();
1019 auto &DL = InnermostLoop->getHeader()->getModule()->getDataLayout();
1020 unsigned TypeByteSize = DL.getTypeAllocSize(ATy);
1022 // Negative distances are not plausible dependencies.
1023 const APInt &Val = C->getValue()->getValue();
1024 if (Val.isNegative()) {
1025 bool IsTrueDataDependence = (AIsWrite && !BIsWrite);
1026 if (IsTrueDataDependence &&
1027 (couldPreventStoreLoadForward(Val.abs().getZExtValue(), TypeByteSize) ||
1029 return Dependence::ForwardButPreventsForwarding;
1031 DEBUG(dbgs() << "LAA: Dependence is negative: NoDep\n");
1032 return Dependence::Forward;
1035 // Write to the same location with the same size.
1036 // Could be improved to assert type sizes are the same (i32 == float, etc).
1039 return Dependence::NoDep;
1040 DEBUG(dbgs() << "LAA: Zero dependence difference but different types\n");
1041 return Dependence::Unknown;
1044 assert(Val.isStrictlyPositive() && "Expect a positive value");
1048 "LAA: ReadWrite-Write positive dependency with different types\n");
1049 return Dependence::Unknown;
1052 unsigned Distance = (unsigned) Val.getZExtValue();
1054 unsigned Stride = std::abs(StrideAPtr);
1056 areStridedAccessesIndependent(Distance, Stride, TypeByteSize)) {
1057 DEBUG(dbgs() << "LAA: Strided accesses are independent\n");
1058 return Dependence::NoDep;
1061 // Bail out early if passed-in parameters make vectorization not feasible.
1062 unsigned ForcedFactor = (VectorizerParams::VectorizationFactor ?
1063 VectorizerParams::VectorizationFactor : 1);
1064 unsigned ForcedUnroll = (VectorizerParams::VectorizationInterleave ?
1065 VectorizerParams::VectorizationInterleave : 1);
1066 // The minimum number of iterations for a vectorized/unrolled version.
1067 unsigned MinNumIter = std::max(ForcedFactor * ForcedUnroll, 2U);
1069 // It's not vectorizable if the distance is smaller than the minimum distance
1070 // needed for a vectroized/unrolled version. Vectorizing one iteration in
1071 // front needs TypeByteSize * Stride. Vectorizing the last iteration needs
1072 // TypeByteSize (No need to plus the last gap distance).
1074 // E.g. Assume one char is 1 byte in memory and one int is 4 bytes.
1076 // int *B = (int *)((char *)A + 14);
1077 // for (i = 0 ; i < 1024 ; i += 2)
1081 // Two accesses in memory (stride is 2):
1082 // | A[0] | | A[2] | | A[4] | | A[6] | |
1083 // | B[0] | | B[2] | | B[4] |
1085 // Distance needs for vectorizing iterations except the last iteration:
1086 // 4 * 2 * (MinNumIter - 1). Distance needs for the last iteration: 4.
1087 // So the minimum distance needed is: 4 * 2 * (MinNumIter - 1) + 4.
1089 // If MinNumIter is 2, it is vectorizable as the minimum distance needed is
1090 // 12, which is less than distance.
1092 // If MinNumIter is 4 (Say if a user forces the vectorization factor to be 4),
1093 // the minimum distance needed is 28, which is greater than distance. It is
1094 // not safe to do vectorization.
1095 unsigned MinDistanceNeeded =
1096 TypeByteSize * Stride * (MinNumIter - 1) + TypeByteSize;
1097 if (MinDistanceNeeded > Distance) {
1098 DEBUG(dbgs() << "LAA: Failure because of positive distance " << Distance
1100 return Dependence::Backward;
1103 // Unsafe if the minimum distance needed is greater than max safe distance.
1104 if (MinDistanceNeeded > MaxSafeDepDistBytes) {
1105 DEBUG(dbgs() << "LAA: Failure because it needs at least "
1106 << MinDistanceNeeded << " size in bytes");
1107 return Dependence::Backward;
1110 // Positive distance bigger than max vectorization factor.
1111 // FIXME: Should use max factor instead of max distance in bytes, which could
1112 // not handle different types.
1113 // E.g. Assume one char is 1 byte in memory and one int is 4 bytes.
1114 // void foo (int *A, char *B) {
1115 // for (unsigned i = 0; i < 1024; i++) {
1116 // A[i+2] = A[i] + 1;
1117 // B[i+2] = B[i] + 1;
1121 // This case is currently unsafe according to the max safe distance. If we
1122 // analyze the two accesses on array B, the max safe dependence distance
1123 // is 2. Then we analyze the accesses on array A, the minimum distance needed
1124 // is 8, which is less than 2 and forbidden vectorization, But actually
1125 // both A and B could be vectorized by 2 iterations.
1126 MaxSafeDepDistBytes =
1127 Distance < MaxSafeDepDistBytes ? Distance : MaxSafeDepDistBytes;
1129 bool IsTrueDataDependence = (!AIsWrite && BIsWrite);
1130 if (IsTrueDataDependence &&
1131 couldPreventStoreLoadForward(Distance, TypeByteSize))
1132 return Dependence::BackwardVectorizableButPreventsForwarding;
1134 DEBUG(dbgs() << "LAA: Positive distance " << Val.getSExtValue()
1135 << " with max VF = "
1136 << MaxSafeDepDistBytes / (TypeByteSize * Stride) << '\n');
1138 return Dependence::BackwardVectorizable;
1141 bool MemoryDepChecker::areDepsSafe(DepCandidates &AccessSets,
1142 MemAccessInfoSet &CheckDeps,
1143 const ValueToValueMap &Strides) {
1145 MaxSafeDepDistBytes = -1U;
1146 while (!CheckDeps.empty()) {
1147 MemAccessInfo CurAccess = *CheckDeps.begin();
1149 // Get the relevant memory access set.
1150 EquivalenceClasses<MemAccessInfo>::iterator I =
1151 AccessSets.findValue(AccessSets.getLeaderValue(CurAccess));
1153 // Check accesses within this set.
1154 EquivalenceClasses<MemAccessInfo>::member_iterator AI, AE;
1155 AI = AccessSets.member_begin(I), AE = AccessSets.member_end();
1157 // Check every access pair.
1159 CheckDeps.erase(*AI);
1160 EquivalenceClasses<MemAccessInfo>::member_iterator OI = std::next(AI);
1162 // Check every accessing instruction pair in program order.
1163 for (std::vector<unsigned>::iterator I1 = Accesses[*AI].begin(),
1164 I1E = Accesses[*AI].end(); I1 != I1E; ++I1)
1165 for (std::vector<unsigned>::iterator I2 = Accesses[*OI].begin(),
1166 I2E = Accesses[*OI].end(); I2 != I2E; ++I2) {
1167 auto A = std::make_pair(&*AI, *I1);
1168 auto B = std::make_pair(&*OI, *I2);
1174 Dependence::DepType Type =
1175 isDependent(*A.first, A.second, *B.first, B.second, Strides);
1176 SafeForVectorization &= Dependence::isSafeForVectorization(Type);
1178 // Gather dependences unless we accumulated MaxInterestingDependence
1179 // dependences. In that case return as soon as we find the first
1180 // unsafe dependence. This puts a limit on this quadratic
1182 if (RecordInterestingDependences) {
1183 if (Dependence::isInterestingDependence(Type))
1184 InterestingDependences.push_back(
1185 Dependence(A.second, B.second, Type));
1187 if (InterestingDependences.size() >= MaxInterestingDependence) {
1188 RecordInterestingDependences = false;
1189 InterestingDependences.clear();
1190 DEBUG(dbgs() << "Too many dependences, stopped recording\n");
1193 if (!RecordInterestingDependences && !SafeForVectorization)
1202 DEBUG(dbgs() << "Total Interesting Dependences: "
1203 << InterestingDependences.size() << "\n");
1204 return SafeForVectorization;
1207 SmallVector<Instruction *, 4>
1208 MemoryDepChecker::getInstructionsForAccess(Value *Ptr, bool isWrite) const {
1209 MemAccessInfo Access(Ptr, isWrite);
1210 auto &IndexVector = Accesses.find(Access)->second;
1212 SmallVector<Instruction *, 4> Insts;
1213 std::transform(IndexVector.begin(), IndexVector.end(),
1214 std::back_inserter(Insts),
1215 [&](unsigned Idx) { return this->InstMap[Idx]; });
1219 const char *MemoryDepChecker::Dependence::DepName[] = {
1220 "NoDep", "Unknown", "Forward", "ForwardButPreventsForwarding", "Backward",
1221 "BackwardVectorizable", "BackwardVectorizableButPreventsForwarding"};
1223 void MemoryDepChecker::Dependence::print(
1224 raw_ostream &OS, unsigned Depth,
1225 const SmallVectorImpl<Instruction *> &Instrs) const {
1226 OS.indent(Depth) << DepName[Type] << ":\n";
1227 OS.indent(Depth + 2) << *Instrs[Source] << " -> \n";
1228 OS.indent(Depth + 2) << *Instrs[Destination] << "\n";
1231 bool LoopAccessInfo::canAnalyzeLoop() {
1232 // We need to have a loop header.
1233 DEBUG(dbgs() << "LAA: Found a loop: " <<
1234 TheLoop->getHeader()->getName() << '\n');
1236 // We can only analyze innermost loops.
1237 if (!TheLoop->empty()) {
1238 DEBUG(dbgs() << "LAA: loop is not the innermost loop\n");
1239 emitAnalysis(LoopAccessReport() << "loop is not the innermost loop");
1243 // We must have a single backedge.
1244 if (TheLoop->getNumBackEdges() != 1) {
1245 DEBUG(dbgs() << "LAA: loop control flow is not understood by analyzer\n");
1247 LoopAccessReport() <<
1248 "loop control flow is not understood by analyzer");
1252 // We must have a single exiting block.
1253 if (!TheLoop->getExitingBlock()) {
1254 DEBUG(dbgs() << "LAA: loop control flow is not understood by analyzer\n");
1256 LoopAccessReport() <<
1257 "loop control flow is not understood by analyzer");
1261 // We only handle bottom-tested loops, i.e. loop in which the condition is
1262 // checked at the end of each iteration. With that we can assume that all
1263 // instructions in the loop are executed the same number of times.
1264 if (TheLoop->getExitingBlock() != TheLoop->getLoopLatch()) {
1265 DEBUG(dbgs() << "LAA: loop control flow is not understood by analyzer\n");
1267 LoopAccessReport() <<
1268 "loop control flow is not understood by analyzer");
1272 // ScalarEvolution needs to be able to find the exit count.
1273 const SCEV *ExitCount = SE->getBackedgeTakenCount(TheLoop);
1274 if (ExitCount == SE->getCouldNotCompute()) {
1275 emitAnalysis(LoopAccessReport() <<
1276 "could not determine number of loop iterations");
1277 DEBUG(dbgs() << "LAA: SCEV could not compute the loop exit count.\n");
1284 void LoopAccessInfo::analyzeLoop(const ValueToValueMap &Strides) {
1286 typedef SmallVector<Value*, 16> ValueVector;
1287 typedef SmallPtrSet<Value*, 16> ValueSet;
1289 // Holds the Load and Store *instructions*.
1293 // Holds all the different accesses in the loop.
1294 unsigned NumReads = 0;
1295 unsigned NumReadWrites = 0;
1297 PtrRtCheck.Pointers.clear();
1298 PtrRtCheck.Need = false;
1300 const bool IsAnnotatedParallel = TheLoop->isAnnotatedParallel();
1303 for (Loop::block_iterator bb = TheLoop->block_begin(),
1304 be = TheLoop->block_end(); bb != be; ++bb) {
1306 // Scan the BB and collect legal loads and stores.
1307 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
1310 // If this is a load, save it. If this instruction can read from memory
1311 // but is not a load, then we quit. Notice that we don't handle function
1312 // calls that read or write.
1313 if (it->mayReadFromMemory()) {
1314 // Many math library functions read the rounding mode. We will only
1315 // vectorize a loop if it contains known function calls that don't set
1316 // the flag. Therefore, it is safe to ignore this read from memory.
1317 CallInst *Call = dyn_cast<CallInst>(it);
1318 if (Call && getIntrinsicIDForCall(Call, TLI))
1321 // If the function has an explicit vectorized counterpart, we can safely
1322 // assume that it can be vectorized.
1323 if (Call && !Call->isNoBuiltin() && Call->getCalledFunction() &&
1324 TLI->isFunctionVectorizable(Call->getCalledFunction()->getName()))
1327 LoadInst *Ld = dyn_cast<LoadInst>(it);
1328 if (!Ld || (!Ld->isSimple() && !IsAnnotatedParallel)) {
1329 emitAnalysis(LoopAccessReport(Ld)
1330 << "read with atomic ordering or volatile read");
1331 DEBUG(dbgs() << "LAA: Found a non-simple load.\n");
1336 Loads.push_back(Ld);
1337 DepChecker.addAccess(Ld);
1341 // Save 'store' instructions. Abort if other instructions write to memory.
1342 if (it->mayWriteToMemory()) {
1343 StoreInst *St = dyn_cast<StoreInst>(it);
1345 emitAnalysis(LoopAccessReport(it) <<
1346 "instruction cannot be vectorized");
1350 if (!St->isSimple() && !IsAnnotatedParallel) {
1351 emitAnalysis(LoopAccessReport(St)
1352 << "write with atomic ordering or volatile write");
1353 DEBUG(dbgs() << "LAA: Found a non-simple store.\n");
1358 Stores.push_back(St);
1359 DepChecker.addAccess(St);
1364 // Now we have two lists that hold the loads and the stores.
1365 // Next, we find the pointers that they use.
1367 // Check if we see any stores. If there are no stores, then we don't
1368 // care if the pointers are *restrict*.
1369 if (!Stores.size()) {
1370 DEBUG(dbgs() << "LAA: Found a read-only loop!\n");
1375 MemoryDepChecker::DepCandidates DependentAccesses;
1376 AccessAnalysis Accesses(TheLoop->getHeader()->getModule()->getDataLayout(),
1377 AA, LI, DependentAccesses);
1379 // Holds the analyzed pointers. We don't want to call GetUnderlyingObjects
1380 // multiple times on the same object. If the ptr is accessed twice, once
1381 // for read and once for write, it will only appear once (on the write
1382 // list). This is okay, since we are going to check for conflicts between
1383 // writes and between reads and writes, but not between reads and reads.
1386 ValueVector::iterator I, IE;
1387 for (I = Stores.begin(), IE = Stores.end(); I != IE; ++I) {
1388 StoreInst *ST = cast<StoreInst>(*I);
1389 Value* Ptr = ST->getPointerOperand();
1390 // Check for store to loop invariant address.
1391 StoreToLoopInvariantAddress |= isUniform(Ptr);
1392 // If we did *not* see this pointer before, insert it to the read-write
1393 // list. At this phase it is only a 'write' list.
1394 if (Seen.insert(Ptr).second) {
1397 MemoryLocation Loc = MemoryLocation::get(ST);
1398 // The TBAA metadata could have a control dependency on the predication
1399 // condition, so we cannot rely on it when determining whether or not we
1400 // need runtime pointer checks.
1401 if (blockNeedsPredication(ST->getParent(), TheLoop, DT))
1402 Loc.AATags.TBAA = nullptr;
1404 Accesses.addStore(Loc);
1408 if (IsAnnotatedParallel) {
1410 << "LAA: A loop annotated parallel, ignore memory dependency "
1416 for (I = Loads.begin(), IE = Loads.end(); I != IE; ++I) {
1417 LoadInst *LD = cast<LoadInst>(*I);
1418 Value* Ptr = LD->getPointerOperand();
1419 // If we did *not* see this pointer before, insert it to the
1420 // read list. If we *did* see it before, then it is already in
1421 // the read-write list. This allows us to vectorize expressions
1422 // such as A[i] += x; Because the address of A[i] is a read-write
1423 // pointer. This only works if the index of A[i] is consecutive.
1424 // If the address of i is unknown (for example A[B[i]]) then we may
1425 // read a few words, modify, and write a few words, and some of the
1426 // words may be written to the same address.
1427 bool IsReadOnlyPtr = false;
1428 if (Seen.insert(Ptr).second || !isStridedPtr(SE, Ptr, TheLoop, Strides)) {
1430 IsReadOnlyPtr = true;
1433 MemoryLocation Loc = MemoryLocation::get(LD);
1434 // The TBAA metadata could have a control dependency on the predication
1435 // condition, so we cannot rely on it when determining whether or not we
1436 // need runtime pointer checks.
1437 if (blockNeedsPredication(LD->getParent(), TheLoop, DT))
1438 Loc.AATags.TBAA = nullptr;
1440 Accesses.addLoad(Loc, IsReadOnlyPtr);
1443 // If we write (or read-write) to a single destination and there are no
1444 // other reads in this loop then is it safe to vectorize.
1445 if (NumReadWrites == 1 && NumReads == 0) {
1446 DEBUG(dbgs() << "LAA: Found a write-only loop!\n");
1451 // Build dependence sets and check whether we need a runtime pointer bounds
1453 Accesses.buildDependenceSets();
1455 // Find pointers with computable bounds. We are going to use this information
1456 // to place a runtime bound check.
1458 bool CanDoRT = Accesses.canCheckPtrAtRT(PtrRtCheck,
1462 DEBUG(dbgs() << "LAA: We need to do "
1463 << PtrRtCheck.getNumberOfChecks(nullptr)
1464 << " pointer comparisons.\n");
1466 // Check that we found the bounds for the pointer.
1468 DEBUG(dbgs() << "LAA: We can perform a memory runtime check if needed.\n");
1469 else if (NeedRTCheck) {
1470 emitAnalysis(LoopAccessReport() << "cannot identify array bounds");
1471 DEBUG(dbgs() << "LAA: We can't vectorize because we can't find " <<
1472 "the array bounds.\n");
1478 PtrRtCheck.Need = NeedRTCheck;
1481 if (Accesses.isDependencyCheckNeeded()) {
1482 DEBUG(dbgs() << "LAA: Checking memory dependencies\n");
1483 CanVecMem = DepChecker.areDepsSafe(
1484 DependentAccesses, Accesses.getDependenciesToCheck(), Strides);
1485 MaxSafeDepDistBytes = DepChecker.getMaxSafeDepDistBytes();
1487 if (!CanVecMem && DepChecker.shouldRetryWithRuntimeCheck()) {
1488 DEBUG(dbgs() << "LAA: Retrying with memory checks\n");
1491 // Clear the dependency checks. We assume they are not needed.
1492 Accesses.resetDepChecks(DepChecker);
1495 PtrRtCheck.Need = true;
1497 CanDoRT = Accesses.canCheckPtrAtRT(PtrRtCheck, NeedRTCheck, SE,
1498 TheLoop, Strides, true);
1500 // Check that we found the bounds for the pointer.
1501 if (NeedRTCheck && !CanDoRT) {
1502 emitAnalysis(LoopAccessReport()
1503 << "cannot check memory dependencies at runtime");
1504 DEBUG(dbgs() << "LAA: Can't vectorize with memory checks\n");
1515 DEBUG(dbgs() << "LAA: No unsafe dependent memory operations in loop. We"
1516 << (NeedRTCheck ? "" : " don't")
1517 << " need a runtime memory check.\n");
1519 emitAnalysis(LoopAccessReport() <<
1520 "unsafe dependent memory operations in loop");
1521 DEBUG(dbgs() << "LAA: unsafe dependent memory operations in loop\n");
1525 bool LoopAccessInfo::blockNeedsPredication(BasicBlock *BB, Loop *TheLoop,
1526 DominatorTree *DT) {
1527 assert(TheLoop->contains(BB) && "Unknown block used");
1529 // Blocks that do not dominate the latch need predication.
1530 BasicBlock* Latch = TheLoop->getLoopLatch();
1531 return !DT->dominates(BB, Latch);
1534 void LoopAccessInfo::emitAnalysis(LoopAccessReport &Message) {
1535 assert(!Report && "Multiple reports generated");
1539 bool LoopAccessInfo::isUniform(Value *V) const {
1540 return (SE->isLoopInvariant(SE->getSCEV(V), TheLoop));
1543 // FIXME: this function is currently a duplicate of the one in
1544 // LoopVectorize.cpp.
1545 static Instruction *getFirstInst(Instruction *FirstInst, Value *V,
1549 if (Instruction *I = dyn_cast<Instruction>(V))
1550 return I->getParent() == Loc->getParent() ? I : nullptr;
1554 std::pair<Instruction *, Instruction *> LoopAccessInfo::addRuntimeCheck(
1555 Instruction *Loc, const SmallVectorImpl<int> *PtrPartition) const {
1556 if (!PtrRtCheck.Need)
1557 return std::make_pair(nullptr, nullptr);
1559 SmallVector<TrackingVH<Value>, 2> Starts;
1560 SmallVector<TrackingVH<Value>, 2> Ends;
1562 LLVMContext &Ctx = Loc->getContext();
1563 SCEVExpander Exp(*SE, DL, "induction");
1564 Instruction *FirstInst = nullptr;
1566 for (unsigned i = 0; i < PtrRtCheck.CheckingGroups.size(); ++i) {
1567 const RuntimePointerCheck::CheckingPtrGroup &CG =
1568 PtrRtCheck.CheckingGroups[i];
1569 Value *Ptr = PtrRtCheck.Pointers[CG.Members[0]];
1570 const SCEV *Sc = SE->getSCEV(Ptr);
1572 if (SE->isLoopInvariant(Sc, TheLoop)) {
1573 DEBUG(dbgs() << "LAA: Adding RT check for a loop invariant ptr:" << *Ptr
1575 Starts.push_back(Ptr);
1576 Ends.push_back(Ptr);
1578 unsigned AS = Ptr->getType()->getPointerAddressSpace();
1580 // Use this type for pointer arithmetic.
1581 Type *PtrArithTy = Type::getInt8PtrTy(Ctx, AS);
1582 Value *Start = nullptr, *End = nullptr;
1584 DEBUG(dbgs() << "LAA: Adding RT check for range:\n");
1585 Start = Exp.expandCodeFor(CG.Low, PtrArithTy, Loc);
1586 End = Exp.expandCodeFor(CG.High, PtrArithTy, Loc);
1587 DEBUG(dbgs() << "Start: " << *CG.Low << " End: " << *CG.High << "\n");
1588 Starts.push_back(Start);
1589 Ends.push_back(End);
1593 IRBuilder<> ChkBuilder(Loc);
1594 // Our instructions might fold to a constant.
1595 Value *MemoryRuntimeCheck = nullptr;
1596 for (unsigned i = 0; i < PtrRtCheck.CheckingGroups.size(); ++i) {
1597 for (unsigned j = i + 1; j < PtrRtCheck.CheckingGroups.size(); ++j) {
1598 const RuntimePointerCheck::CheckingPtrGroup &CGI =
1599 PtrRtCheck.CheckingGroups[i];
1600 const RuntimePointerCheck::CheckingPtrGroup &CGJ =
1601 PtrRtCheck.CheckingGroups[j];
1603 if (!PtrRtCheck.needsChecking(CGI, CGJ, PtrPartition))
1606 unsigned AS0 = Starts[i]->getType()->getPointerAddressSpace();
1607 unsigned AS1 = Starts[j]->getType()->getPointerAddressSpace();
1609 assert((AS0 == Ends[j]->getType()->getPointerAddressSpace()) &&
1610 (AS1 == Ends[i]->getType()->getPointerAddressSpace()) &&
1611 "Trying to bounds check pointers with different address spaces");
1613 Type *PtrArithTy0 = Type::getInt8PtrTy(Ctx, AS0);
1614 Type *PtrArithTy1 = Type::getInt8PtrTy(Ctx, AS1);
1616 Value *Start0 = ChkBuilder.CreateBitCast(Starts[i], PtrArithTy0, "bc");
1617 Value *Start1 = ChkBuilder.CreateBitCast(Starts[j], PtrArithTy1, "bc");
1618 Value *End0 = ChkBuilder.CreateBitCast(Ends[i], PtrArithTy1, "bc");
1619 Value *End1 = ChkBuilder.CreateBitCast(Ends[j], PtrArithTy0, "bc");
1621 Value *Cmp0 = ChkBuilder.CreateICmpULE(Start0, End1, "bound0");
1622 FirstInst = getFirstInst(FirstInst, Cmp0, Loc);
1623 Value *Cmp1 = ChkBuilder.CreateICmpULE(Start1, End0, "bound1");
1624 FirstInst = getFirstInst(FirstInst, Cmp1, Loc);
1625 Value *IsConflict = ChkBuilder.CreateAnd(Cmp0, Cmp1, "found.conflict");
1626 FirstInst = getFirstInst(FirstInst, IsConflict, Loc);
1627 if (MemoryRuntimeCheck) {
1628 IsConflict = ChkBuilder.CreateOr(MemoryRuntimeCheck, IsConflict,
1630 FirstInst = getFirstInst(FirstInst, IsConflict, Loc);
1632 MemoryRuntimeCheck = IsConflict;
1636 if (!MemoryRuntimeCheck)
1637 return std::make_pair(nullptr, nullptr);
1639 // We have to do this trickery because the IRBuilder might fold the check to a
1640 // constant expression in which case there is no Instruction anchored in a
1642 Instruction *Check = BinaryOperator::CreateAnd(MemoryRuntimeCheck,
1643 ConstantInt::getTrue(Ctx));
1644 ChkBuilder.Insert(Check, "memcheck.conflict");
1645 FirstInst = getFirstInst(FirstInst, Check, Loc);
1646 return std::make_pair(FirstInst, Check);
1649 LoopAccessInfo::LoopAccessInfo(Loop *L, ScalarEvolution *SE,
1650 const DataLayout &DL,
1651 const TargetLibraryInfo *TLI, AliasAnalysis *AA,
1652 DominatorTree *DT, LoopInfo *LI,
1653 const ValueToValueMap &Strides)
1654 : PtrRtCheck(SE), DepChecker(SE, L), TheLoop(L), SE(SE), DL(DL), TLI(TLI),
1655 AA(AA), DT(DT), LI(LI), NumLoads(0), NumStores(0),
1656 MaxSafeDepDistBytes(-1U), CanVecMem(false),
1657 StoreToLoopInvariantAddress(false) {
1658 if (canAnalyzeLoop())
1659 analyzeLoop(Strides);
1662 void LoopAccessInfo::print(raw_ostream &OS, unsigned Depth) const {
1664 if (PtrRtCheck.Need)
1665 OS.indent(Depth) << "Memory dependences are safe with run-time checks\n";
1667 OS.indent(Depth) << "Memory dependences are safe\n";
1671 OS.indent(Depth) << "Report: " << Report->str() << "\n";
1673 if (auto *InterestingDependences = DepChecker.getInterestingDependences()) {
1674 OS.indent(Depth) << "Interesting Dependences:\n";
1675 for (auto &Dep : *InterestingDependences) {
1676 Dep.print(OS, Depth + 2, DepChecker.getMemoryInstructions());
1680 OS.indent(Depth) << "Too many interesting dependences, not recorded\n";
1682 // List the pair of accesses need run-time checks to prove independence.
1683 PtrRtCheck.print(OS, Depth);
1686 OS.indent(Depth) << "Store to invariant address was "
1687 << (StoreToLoopInvariantAddress ? "" : "not ")
1688 << "found in loop.\n";
1691 const LoopAccessInfo &
1692 LoopAccessAnalysis::getInfo(Loop *L, const ValueToValueMap &Strides) {
1693 auto &LAI = LoopAccessInfoMap[L];
1696 assert((!LAI || LAI->NumSymbolicStrides == Strides.size()) &&
1697 "Symbolic strides changed for loop");
1701 const DataLayout &DL = L->getHeader()->getModule()->getDataLayout();
1702 LAI = llvm::make_unique<LoopAccessInfo>(L, SE, DL, TLI, AA, DT, LI,
1705 LAI->NumSymbolicStrides = Strides.size();
1711 void LoopAccessAnalysis::print(raw_ostream &OS, const Module *M) const {
1712 LoopAccessAnalysis &LAA = *const_cast<LoopAccessAnalysis *>(this);
1714 ValueToValueMap NoSymbolicStrides;
1716 for (Loop *TopLevelLoop : *LI)
1717 for (Loop *L : depth_first(TopLevelLoop)) {
1718 OS.indent(2) << L->getHeader()->getName() << ":\n";
1719 auto &LAI = LAA.getInfo(L, NoSymbolicStrides);
1724 bool LoopAccessAnalysis::runOnFunction(Function &F) {
1725 SE = &getAnalysis<ScalarEvolution>();
1726 auto *TLIP = getAnalysisIfAvailable<TargetLibraryInfoWrapperPass>();
1727 TLI = TLIP ? &TLIP->getTLI() : nullptr;
1728 AA = &getAnalysis<AliasAnalysis>();
1729 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1730 LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
1735 void LoopAccessAnalysis::getAnalysisUsage(AnalysisUsage &AU) const {
1736 AU.addRequired<ScalarEvolution>();
1737 AU.addRequired<AliasAnalysis>();
1738 AU.addRequired<DominatorTreeWrapperPass>();
1739 AU.addRequired<LoopInfoWrapperPass>();
1741 AU.setPreservesAll();
1744 char LoopAccessAnalysis::ID = 0;
1745 static const char laa_name[] = "Loop Access Analysis";
1746 #define LAA_NAME "loop-accesses"
1748 INITIALIZE_PASS_BEGIN(LoopAccessAnalysis, LAA_NAME, laa_name, false, true)
1749 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
1750 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
1751 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
1752 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
1753 INITIALIZE_PASS_END(LoopAccessAnalysis, LAA_NAME, laa_name, false, true)
1756 Pass *createLAAPass() {
1757 return new LoopAccessAnalysis();