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 RuntimePointerChecking::insert(Loop *Lp, Value *Ptr, bool WritePtr,
123 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);
131 const SCEV *ScStart = AR->getStart();
132 const SCEV *ScEnd = AR->evaluateAtIteration(Ex, *SE);
133 const SCEV *Step = AR->getStepRecurrence(*SE);
135 // For expressions with negative step, the upper bound is ScStart and the
136 // lower bound is ScEnd.
137 if (const SCEVConstant *CStep = dyn_cast<const SCEVConstant>(Step)) {
138 if (CStep->getValue()->isNegative())
139 std::swap(ScStart, ScEnd);
141 // Fallback case: the step is not constant, but the we can still
142 // get the upper and lower bounds of the interval by using min/max
144 ScStart = SE->getUMinExpr(ScStart, ScEnd);
145 ScEnd = SE->getUMaxExpr(AR->getStart(), ScEnd);
148 Pointers.emplace_back(Ptr, ScStart, ScEnd, WritePtr, DepSetId, ASId, Sc);
151 bool RuntimePointerChecking::needsChecking(
152 const CheckingPtrGroup &M, const CheckingPtrGroup &N,
153 const SmallVectorImpl<int> *PtrPartition) const {
154 for (unsigned I = 0, EI = M.Members.size(); EI != I; ++I)
155 for (unsigned J = 0, EJ = N.Members.size(); EJ != J; ++J)
156 if (needsChecking(M.Members[I], N.Members[J], PtrPartition))
161 /// Compare \p I and \p J and return the minimum.
162 /// Return nullptr in case we couldn't find an answer.
163 static const SCEV *getMinFromExprs(const SCEV *I, const SCEV *J,
164 ScalarEvolution *SE) {
165 const SCEV *Diff = SE->getMinusSCEV(J, I);
166 const SCEVConstant *C = dyn_cast<const SCEVConstant>(Diff);
170 if (C->getValue()->isNegative())
175 bool RuntimePointerChecking::CheckingPtrGroup::addPointer(unsigned Index) {
176 const SCEV *Start = RtCheck.Pointers[Index].Start;
177 const SCEV *End = RtCheck.Pointers[Index].End;
179 // Compare the starts and ends with the known minimum and maximum
180 // of this set. We need to know how we compare against the min/max
181 // of the set in order to be able to emit memchecks.
182 const SCEV *Min0 = getMinFromExprs(Start, Low, RtCheck.SE);
186 const SCEV *Min1 = getMinFromExprs(End, High, RtCheck.SE);
190 // Update the low bound expression if we've found a new min value.
194 // Update the high bound expression if we've found a new max value.
198 Members.push_back(Index);
202 void RuntimePointerChecking::groupChecks(
203 MemoryDepChecker::DepCandidates &DepCands, bool UseDependencies) {
204 // We build the groups from dependency candidates equivalence classes
206 // - We know that pointers in the same equivalence class share
207 // the same underlying object and therefore there is a chance
208 // that we can compare pointers
209 // - We wouldn't be able to merge two pointers for which we need
210 // to emit a memcheck. The classes in DepCands are already
211 // conveniently built such that no two pointers in the same
212 // class need checking against each other.
214 // We use the following (greedy) algorithm to construct the groups
215 // For every pointer in the equivalence class:
216 // For each existing group:
217 // - if the difference between this pointer and the min/max bounds
218 // of the group is a constant, then make the pointer part of the
219 // group and update the min/max bounds of that group as required.
221 CheckingGroups.clear();
223 // If we don't have the dependency partitions, construct a new
224 // checking pointer group for each pointer.
225 if (!UseDependencies) {
226 for (unsigned I = 0; I < Pointers.size(); ++I)
227 CheckingGroups.push_back(CheckingPtrGroup(I, *this));
231 unsigned TotalComparisons = 0;
233 DenseMap<Value *, unsigned> PositionMap;
234 for (unsigned Index = 0; Index < Pointers.size(); ++Index)
235 PositionMap[Pointers[Index].PointerValue] = Index;
237 // We need to keep track of what pointers we've already seen so we
238 // don't process them twice.
239 SmallSet<unsigned, 2> Seen;
241 // Go through all equivalence classes, get the the "pointer check groups"
242 // and add them to the overall solution. We use the order in which accesses
243 // appear in 'Pointers' to enforce determinism.
244 for (unsigned I = 0; I < Pointers.size(); ++I) {
245 // We've seen this pointer before, and therefore already processed
246 // its equivalence class.
250 MemoryDepChecker::MemAccessInfo Access(Pointers[I].PointerValue,
251 Pointers[I].IsWritePtr);
253 SmallVector<CheckingPtrGroup, 2> Groups;
254 auto LeaderI = DepCands.findValue(DepCands.getLeaderValue(Access));
256 // Because DepCands is constructed by visiting accesses in the order in
257 // which they appear in alias sets (which is deterministic) and the
258 // iteration order within an equivalence class member is only dependent on
259 // the order in which unions and insertions are performed on the
260 // equivalence class, the iteration order is deterministic.
261 for (auto MI = DepCands.member_begin(LeaderI), ME = DepCands.member_end();
263 unsigned Pointer = PositionMap[MI->getPointer()];
265 // Mark this pointer as seen.
266 Seen.insert(Pointer);
268 // Go through all the existing sets and see if we can find one
269 // which can include this pointer.
270 for (CheckingPtrGroup &Group : Groups) {
271 // Don't perform more than a certain amount of comparisons.
272 // This should limit the cost of grouping the pointers to something
273 // reasonable. If we do end up hitting this threshold, the algorithm
274 // will create separate groups for all remaining pointers.
275 if (TotalComparisons > MemoryCheckMergeThreshold)
280 if (Group.addPointer(Pointer)) {
287 // We couldn't add this pointer to any existing set or the threshold
288 // for the number of comparisons has been reached. Create a new group
289 // to hold the current pointer.
290 Groups.push_back(CheckingPtrGroup(Pointer, *this));
293 // We've computed the grouped checks for this partition.
294 // Save the results and continue with the next one.
295 std::copy(Groups.begin(), Groups.end(), std::back_inserter(CheckingGroups));
299 bool RuntimePointerChecking::arePointersInSamePartition(
300 const SmallVectorImpl<int> &PtrToPartition, unsigned PtrIdx1,
302 return (PtrToPartition[PtrIdx1] != -1 &&
303 PtrToPartition[PtrIdx1] == PtrToPartition[PtrIdx2]);
306 bool RuntimePointerChecking::needsChecking(
307 unsigned I, unsigned J, const SmallVectorImpl<int> *PtrPartition) const {
308 const PointerInfo &PointerI = Pointers[I];
309 const PointerInfo &PointerJ = Pointers[J];
311 // No need to check if two readonly pointers intersect.
312 if (!PointerI.IsWritePtr && !PointerJ.IsWritePtr)
315 // Only need to check pointers between two different dependency sets.
316 if (PointerI.DependencySetId == PointerJ.DependencySetId)
319 // Only need to check pointers in the same alias set.
320 if (PointerI.AliasSetId != PointerJ.AliasSetId)
323 // If PtrPartition is set omit checks between pointers of the same partition.
324 if (PtrPartition && arePointersInSamePartition(*PtrPartition, I, J))
330 void RuntimePointerChecking::print(
331 raw_ostream &OS, unsigned Depth,
332 const SmallVectorImpl<int> *PtrPartition) const {
334 OS.indent(Depth) << "Run-time memory checks:\n";
337 for (unsigned I = 0; I < CheckingGroups.size(); ++I)
338 for (unsigned J = I + 1; J < CheckingGroups.size(); ++J)
339 if (needsChecking(CheckingGroups[I], CheckingGroups[J], PtrPartition)) {
340 OS.indent(Depth) << "Check " << N++ << ":\n";
341 OS.indent(Depth + 2) << "Comparing group " << I << ":\n";
343 for (unsigned K = 0; K < CheckingGroups[I].Members.size(); ++K) {
345 << *Pointers[CheckingGroups[I].Members[K]].PointerValue << "\n";
347 OS << " (Partition: "
348 << (*PtrPartition)[CheckingGroups[I].Members[K]] << ")"
352 OS.indent(Depth + 2) << "Against group " << J << ":\n";
354 for (unsigned K = 0; K < CheckingGroups[J].Members.size(); ++K) {
356 << *Pointers[CheckingGroups[J].Members[K]].PointerValue << "\n";
358 OS << " (Partition: "
359 << (*PtrPartition)[CheckingGroups[J].Members[K]] << ")"
364 OS.indent(Depth) << "Grouped accesses:\n";
365 for (unsigned I = 0; I < CheckingGroups.size(); ++I) {
366 OS.indent(Depth + 2) << "Group " << I << ":\n";
367 OS.indent(Depth + 4) << "(Low: " << *CheckingGroups[I].Low
368 << " High: " << *CheckingGroups[I].High << ")\n";
369 for (unsigned J = 0; J < CheckingGroups[I].Members.size(); ++J) {
370 OS.indent(Depth + 6) << "Member: "
371 << *Pointers[CheckingGroups[I].Members[J]].Expr
377 unsigned RuntimePointerChecking::getNumberOfChecks(
378 const SmallVectorImpl<int> *PtrPartition) const {
380 unsigned NumPartitions = CheckingGroups.size();
381 unsigned CheckCount = 0;
383 for (unsigned I = 0; I < NumPartitions; ++I)
384 for (unsigned J = I + 1; J < NumPartitions; ++J)
385 if (needsChecking(CheckingGroups[I], CheckingGroups[J], PtrPartition))
390 bool RuntimePointerChecking::needsAnyChecking(
391 const SmallVectorImpl<int> *PtrPartition) const {
392 unsigned NumPointers = Pointers.size();
394 for (unsigned I = 0; I < NumPointers; ++I)
395 for (unsigned J = I + 1; J < NumPointers; ++J)
396 if (needsChecking(I, J, PtrPartition))
402 /// \brief Analyses memory accesses in a loop.
404 /// Checks whether run time pointer checks are needed and builds sets for data
405 /// dependence checking.
406 class AccessAnalysis {
408 /// \brief Read or write access location.
409 typedef PointerIntPair<Value *, 1, bool> MemAccessInfo;
410 typedef SmallPtrSet<MemAccessInfo, 8> MemAccessInfoSet;
412 AccessAnalysis(const DataLayout &Dl, AliasAnalysis *AA, LoopInfo *LI,
413 MemoryDepChecker::DepCandidates &DA)
414 : DL(Dl), AST(*AA), LI(LI), DepCands(DA),
415 IsRTCheckAnalysisNeeded(false) {}
417 /// \brief Register a load and whether it is only read from.
418 void addLoad(MemoryLocation &Loc, bool IsReadOnly) {
419 Value *Ptr = const_cast<Value*>(Loc.Ptr);
420 AST.add(Ptr, MemoryLocation::UnknownSize, Loc.AATags);
421 Accesses.insert(MemAccessInfo(Ptr, false));
423 ReadOnlyPtr.insert(Ptr);
426 /// \brief Register a store.
427 void addStore(MemoryLocation &Loc) {
428 Value *Ptr = const_cast<Value*>(Loc.Ptr);
429 AST.add(Ptr, MemoryLocation::UnknownSize, Loc.AATags);
430 Accesses.insert(MemAccessInfo(Ptr, true));
433 /// \brief Check whether we can check the pointers at runtime for
434 /// non-intersection.
436 /// Returns true if we need no check or if we do and we can generate them
437 /// (i.e. the pointers have computable bounds).
438 bool canCheckPtrAtRT(RuntimePointerChecking &RtCheck, ScalarEvolution *SE,
439 Loop *TheLoop, const ValueToValueMap &Strides,
440 bool ShouldCheckStride = false);
442 /// \brief Goes over all memory accesses, checks whether a RT check is needed
443 /// and builds sets of dependent accesses.
444 void buildDependenceSets() {
445 processMemAccesses();
448 /// \brief Initial processing of memory accesses determined that we need to
449 /// perform dependency checking.
451 /// Note that this can later be cleared if we retry memcheck analysis without
452 /// dependency checking (i.e. ShouldRetryWithRuntimeCheck).
453 bool isDependencyCheckNeeded() { return !CheckDeps.empty(); }
455 /// We decided that no dependence analysis would be used. Reset the state.
456 void resetDepChecks(MemoryDepChecker &DepChecker) {
458 DepChecker.clearInterestingDependences();
461 MemAccessInfoSet &getDependenciesToCheck() { return CheckDeps; }
464 typedef SetVector<MemAccessInfo> PtrAccessSet;
466 /// \brief Go over all memory access and check whether runtime pointer checks
467 /// are needed and build sets of dependency check candidates.
468 void processMemAccesses();
470 /// Set of all accesses.
471 PtrAccessSet Accesses;
473 const DataLayout &DL;
475 /// Set of accesses that need a further dependence check.
476 MemAccessInfoSet CheckDeps;
478 /// Set of pointers that are read only.
479 SmallPtrSet<Value*, 16> ReadOnlyPtr;
481 /// An alias set tracker to partition the access set by underlying object and
482 //intrinsic property (such as TBAA metadata).
487 /// Sets of potentially dependent accesses - members of one set share an
488 /// underlying pointer. The set "CheckDeps" identfies which sets really need a
489 /// dependence check.
490 MemoryDepChecker::DepCandidates &DepCands;
492 /// \brief Initial processing of memory accesses determined that we may need
493 /// to add memchecks. Perform the analysis to determine the necessary checks.
495 /// Note that, this is different from isDependencyCheckNeeded. When we retry
496 /// memcheck analysis without dependency checking
497 /// (i.e. ShouldRetryWithRuntimeCheck), isDependencyCheckNeeded is cleared
498 /// while this remains set if we have potentially dependent accesses.
499 bool IsRTCheckAnalysisNeeded;
502 } // end anonymous namespace
504 /// \brief Check whether a pointer can participate in a runtime bounds check.
505 static bool hasComputableBounds(ScalarEvolution *SE,
506 const ValueToValueMap &Strides, Value *Ptr) {
507 const SCEV *PtrScev = replaceSymbolicStrideSCEV(SE, Strides, Ptr);
508 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
512 return AR->isAffine();
515 bool AccessAnalysis::canCheckPtrAtRT(RuntimePointerChecking &RtCheck,
516 ScalarEvolution *SE, Loop *TheLoop,
517 const ValueToValueMap &StridesMap,
518 bool ShouldCheckStride) {
519 // Find pointers with computable bounds. We are going to use this information
520 // to place a runtime bound check.
523 bool NeedRTCheck = false;
524 if (!IsRTCheckAnalysisNeeded) return true;
526 bool IsDepCheckNeeded = isDependencyCheckNeeded();
528 // We assign a consecutive id to access from different alias sets.
529 // Accesses between different groups doesn't need to be checked.
531 for (auto &AS : AST) {
532 int NumReadPtrChecks = 0;
533 int NumWritePtrChecks = 0;
535 // We assign consecutive id to access from different dependence sets.
536 // Accesses within the same set don't need a runtime check.
537 unsigned RunningDepId = 1;
538 DenseMap<Value *, unsigned> DepSetId;
541 Value *Ptr = A.getValue();
542 bool IsWrite = Accesses.count(MemAccessInfo(Ptr, true));
543 MemAccessInfo Access(Ptr, IsWrite);
550 if (hasComputableBounds(SE, StridesMap, Ptr) &&
551 // When we run after a failing dependency check we have to make sure
552 // we don't have wrapping pointers.
553 (!ShouldCheckStride ||
554 isStridedPtr(SE, Ptr, TheLoop, StridesMap) == 1)) {
555 // The id of the dependence set.
558 if (IsDepCheckNeeded) {
559 Value *Leader = DepCands.getLeaderValue(Access).getPointer();
560 unsigned &LeaderId = DepSetId[Leader];
562 LeaderId = RunningDepId++;
565 // Each access has its own dependence set.
566 DepId = RunningDepId++;
568 RtCheck.insert(TheLoop, Ptr, IsWrite, DepId, ASId, StridesMap);
570 DEBUG(dbgs() << "LAA: Found a runtime check ptr:" << *Ptr << '\n');
572 DEBUG(dbgs() << "LAA: Can't find bounds for ptr:" << *Ptr << '\n');
577 // If we have at least two writes or one write and a read then we need to
578 // check them. But there is no need to checks if there is only one
579 // dependence set for this alias set.
581 // Note that this function computes CanDoRT and NeedRTCheck independently.
582 // For example CanDoRT=false, NeedRTCheck=false means that we have a pointer
583 // for which we couldn't find the bounds but we don't actually need to emit
584 // any checks so it does not matter.
585 if (!(IsDepCheckNeeded && CanDoRT && RunningDepId == 2))
586 NeedRTCheck |= (NumWritePtrChecks >= 2 || (NumReadPtrChecks >= 1 &&
587 NumWritePtrChecks >= 1));
592 // If the pointers that we would use for the bounds comparison have different
593 // address spaces, assume the values aren't directly comparable, so we can't
594 // use them for the runtime check. We also have to assume they could
595 // overlap. In the future there should be metadata for whether address spaces
597 unsigned NumPointers = RtCheck.Pointers.size();
598 for (unsigned i = 0; i < NumPointers; ++i) {
599 for (unsigned j = i + 1; j < NumPointers; ++j) {
600 // Only need to check pointers between two different dependency sets.
601 if (RtCheck.Pointers[i].DependencySetId ==
602 RtCheck.Pointers[j].DependencySetId)
604 // Only need to check pointers in the same alias set.
605 if (RtCheck.Pointers[i].AliasSetId != RtCheck.Pointers[j].AliasSetId)
608 Value *PtrI = RtCheck.Pointers[i].PointerValue;
609 Value *PtrJ = RtCheck.Pointers[j].PointerValue;
611 unsigned ASi = PtrI->getType()->getPointerAddressSpace();
612 unsigned ASj = PtrJ->getType()->getPointerAddressSpace();
614 DEBUG(dbgs() << "LAA: Runtime check would require comparison between"
615 " different address spaces\n");
621 if (NeedRTCheck && CanDoRT)
622 RtCheck.groupChecks(DepCands, IsDepCheckNeeded);
624 DEBUG(dbgs() << "LAA: We need to do " << RtCheck.getNumberOfChecks(nullptr)
625 << " pointer comparisons.\n");
627 RtCheck.Need = NeedRTCheck;
629 bool CanDoRTIfNeeded = !NeedRTCheck || CanDoRT;
630 if (!CanDoRTIfNeeded)
632 return CanDoRTIfNeeded;
635 void AccessAnalysis::processMemAccesses() {
636 // We process the set twice: first we process read-write pointers, last we
637 // process read-only pointers. This allows us to skip dependence tests for
638 // read-only pointers.
640 DEBUG(dbgs() << "LAA: Processing memory accesses...\n");
641 DEBUG(dbgs() << " AST: "; AST.dump());
642 DEBUG(dbgs() << "LAA: Accesses(" << Accesses.size() << "):\n");
644 for (auto A : Accesses)
645 dbgs() << "\t" << *A.getPointer() << " (" <<
646 (A.getInt() ? "write" : (ReadOnlyPtr.count(A.getPointer()) ?
647 "read-only" : "read")) << ")\n";
650 // The AliasSetTracker has nicely partitioned our pointers by metadata
651 // compatibility and potential for underlying-object overlap. As a result, we
652 // only need to check for potential pointer dependencies within each alias
654 for (auto &AS : AST) {
655 // Note that both the alias-set tracker and the alias sets themselves used
656 // linked lists internally and so the iteration order here is deterministic
657 // (matching the original instruction order within each set).
659 bool SetHasWrite = false;
661 // Map of pointers to last access encountered.
662 typedef DenseMap<Value*, MemAccessInfo> UnderlyingObjToAccessMap;
663 UnderlyingObjToAccessMap ObjToLastAccess;
665 // Set of access to check after all writes have been processed.
666 PtrAccessSet DeferredAccesses;
668 // Iterate over each alias set twice, once to process read/write pointers,
669 // and then to process read-only pointers.
670 for (int SetIteration = 0; SetIteration < 2; ++SetIteration) {
671 bool UseDeferred = SetIteration > 0;
672 PtrAccessSet &S = UseDeferred ? DeferredAccesses : Accesses;
675 Value *Ptr = AV.getValue();
677 // For a single memory access in AliasSetTracker, Accesses may contain
678 // both read and write, and they both need to be handled for CheckDeps.
680 if (AC.getPointer() != Ptr)
683 bool IsWrite = AC.getInt();
685 // If we're using the deferred access set, then it contains only
687 bool IsReadOnlyPtr = ReadOnlyPtr.count(Ptr) && !IsWrite;
688 if (UseDeferred && !IsReadOnlyPtr)
690 // Otherwise, the pointer must be in the PtrAccessSet, either as a
692 assert(((IsReadOnlyPtr && UseDeferred) || IsWrite ||
693 S.count(MemAccessInfo(Ptr, false))) &&
694 "Alias-set pointer not in the access set?");
696 MemAccessInfo Access(Ptr, IsWrite);
697 DepCands.insert(Access);
699 // Memorize read-only pointers for later processing and skip them in
700 // the first round (they need to be checked after we have seen all
701 // write pointers). Note: we also mark pointer that are not
702 // consecutive as "read-only" pointers (so that we check
703 // "a[b[i]] +="). Hence, we need the second check for "!IsWrite".
704 if (!UseDeferred && IsReadOnlyPtr) {
705 DeferredAccesses.insert(Access);
709 // If this is a write - check other reads and writes for conflicts. If
710 // this is a read only check other writes for conflicts (but only if
711 // there is no other write to the ptr - this is an optimization to
712 // catch "a[i] = a[i] + " without having to do a dependence check).
713 if ((IsWrite || IsReadOnlyPtr) && SetHasWrite) {
714 CheckDeps.insert(Access);
715 IsRTCheckAnalysisNeeded = true;
721 // Create sets of pointers connected by a shared alias set and
722 // underlying object.
723 typedef SmallVector<Value *, 16> ValueVector;
724 ValueVector TempObjects;
726 GetUnderlyingObjects(Ptr, TempObjects, DL, LI);
727 DEBUG(dbgs() << "Underlying objects for pointer " << *Ptr << "\n");
728 for (Value *UnderlyingObj : TempObjects) {
729 UnderlyingObjToAccessMap::iterator Prev =
730 ObjToLastAccess.find(UnderlyingObj);
731 if (Prev != ObjToLastAccess.end())
732 DepCands.unionSets(Access, Prev->second);
734 ObjToLastAccess[UnderlyingObj] = Access;
735 DEBUG(dbgs() << " " << *UnderlyingObj << "\n");
743 static bool isInBoundsGep(Value *Ptr) {
744 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr))
745 return GEP->isInBounds();
749 /// \brief Return true if an AddRec pointer \p Ptr is unsigned non-wrapping,
750 /// i.e. monotonically increasing/decreasing.
751 static bool isNoWrapAddRec(Value *Ptr, const SCEVAddRecExpr *AR,
752 ScalarEvolution *SE, const Loop *L) {
753 // FIXME: This should probably only return true for NUW.
754 if (AR->getNoWrapFlags(SCEV::NoWrapMask))
757 // Scalar evolution does not propagate the non-wrapping flags to values that
758 // are derived from a non-wrapping induction variable because non-wrapping
759 // could be flow-sensitive.
761 // Look through the potentially overflowing instruction to try to prove
762 // non-wrapping for the *specific* value of Ptr.
764 // The arithmetic implied by an inbounds GEP can't overflow.
765 auto *GEP = dyn_cast<GetElementPtrInst>(Ptr);
766 if (!GEP || !GEP->isInBounds())
769 // Make sure there is only one non-const index and analyze that.
770 Value *NonConstIndex = nullptr;
771 for (auto Index = GEP->idx_begin(); Index != GEP->idx_end(); ++Index)
772 if (!isa<ConstantInt>(*Index)) {
775 NonConstIndex = *Index;
778 // The recurrence is on the pointer, ignore for now.
781 // The index in GEP is signed. It is non-wrapping if it's derived from a NSW
782 // AddRec using a NSW operation.
783 if (auto *OBO = dyn_cast<OverflowingBinaryOperator>(NonConstIndex))
784 if (OBO->hasNoSignedWrap() &&
785 // Assume constant for other the operand so that the AddRec can be
787 isa<ConstantInt>(OBO->getOperand(1))) {
788 auto *OpScev = SE->getSCEV(OBO->getOperand(0));
790 if (auto *OpAR = dyn_cast<SCEVAddRecExpr>(OpScev))
791 return OpAR->getLoop() == L && OpAR->getNoWrapFlags(SCEV::FlagNSW);
797 /// \brief Check whether the access through \p Ptr has a constant stride.
798 int llvm::isStridedPtr(ScalarEvolution *SE, Value *Ptr, const Loop *Lp,
799 const ValueToValueMap &StridesMap) {
800 const Type *Ty = Ptr->getType();
801 assert(Ty->isPointerTy() && "Unexpected non-ptr");
803 // Make sure that the pointer does not point to aggregate types.
804 const PointerType *PtrTy = cast<PointerType>(Ty);
805 if (PtrTy->getElementType()->isAggregateType()) {
806 DEBUG(dbgs() << "LAA: Bad stride - Not a pointer to a scalar type"
811 const SCEV *PtrScev = replaceSymbolicStrideSCEV(SE, StridesMap, Ptr);
813 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
815 DEBUG(dbgs() << "LAA: Bad stride - Not an AddRecExpr pointer "
816 << *Ptr << " SCEV: " << *PtrScev << "\n");
820 // The accesss function must stride over the innermost loop.
821 if (Lp != AR->getLoop()) {
822 DEBUG(dbgs() << "LAA: Bad stride - Not striding over innermost loop " <<
823 *Ptr << " SCEV: " << *PtrScev << "\n");
826 // The address calculation must not wrap. Otherwise, a dependence could be
828 // An inbounds getelementptr that is a AddRec with a unit stride
829 // cannot wrap per definition. The unit stride requirement is checked later.
830 // An getelementptr without an inbounds attribute and unit stride would have
831 // to access the pointer value "0" which is undefined behavior in address
832 // space 0, therefore we can also vectorize this case.
833 bool IsInBoundsGEP = isInBoundsGep(Ptr);
834 bool IsNoWrapAddRec = isNoWrapAddRec(Ptr, AR, SE, Lp);
835 bool IsInAddressSpaceZero = PtrTy->getAddressSpace() == 0;
836 if (!IsNoWrapAddRec && !IsInBoundsGEP && !IsInAddressSpaceZero) {
837 DEBUG(dbgs() << "LAA: Bad stride - Pointer may wrap in the address space "
838 << *Ptr << " SCEV: " << *PtrScev << "\n");
842 // Check the step is constant.
843 const SCEV *Step = AR->getStepRecurrence(*SE);
845 // Calculate the pointer stride and check if it is constant.
846 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
848 DEBUG(dbgs() << "LAA: Bad stride - Not a constant strided " << *Ptr <<
849 " SCEV: " << *PtrScev << "\n");
853 auto &DL = Lp->getHeader()->getModule()->getDataLayout();
854 int64_t Size = DL.getTypeAllocSize(PtrTy->getElementType());
855 const APInt &APStepVal = C->getValue()->getValue();
857 // Huge step value - give up.
858 if (APStepVal.getBitWidth() > 64)
861 int64_t StepVal = APStepVal.getSExtValue();
864 int64_t Stride = StepVal / Size;
865 int64_t Rem = StepVal % Size;
869 // If the SCEV could wrap but we have an inbounds gep with a unit stride we
870 // know we can't "wrap around the address space". In case of address space
871 // zero we know that this won't happen without triggering undefined behavior.
872 if (!IsNoWrapAddRec && (IsInBoundsGEP || IsInAddressSpaceZero) &&
873 Stride != 1 && Stride != -1)
879 bool MemoryDepChecker::Dependence::isSafeForVectorization(DepType Type) {
883 case BackwardVectorizable:
887 case ForwardButPreventsForwarding:
889 case BackwardVectorizableButPreventsForwarding:
892 llvm_unreachable("unexpected DepType!");
895 bool MemoryDepChecker::Dependence::isInterestingDependence(DepType Type) {
901 case BackwardVectorizable:
903 case ForwardButPreventsForwarding:
905 case BackwardVectorizableButPreventsForwarding:
908 llvm_unreachable("unexpected DepType!");
911 bool MemoryDepChecker::Dependence::isPossiblyBackward() const {
915 case ForwardButPreventsForwarding:
919 case BackwardVectorizable:
921 case BackwardVectorizableButPreventsForwarding:
924 llvm_unreachable("unexpected DepType!");
927 bool MemoryDepChecker::couldPreventStoreLoadForward(unsigned Distance,
928 unsigned TypeByteSize) {
929 // If loads occur at a distance that is not a multiple of a feasible vector
930 // factor store-load forwarding does not take place.
931 // Positive dependences might cause troubles because vectorizing them might
932 // prevent store-load forwarding making vectorized code run a lot slower.
933 // a[i] = a[i-3] ^ a[i-8];
934 // The stores to a[i:i+1] don't align with the stores to a[i-3:i-2] and
935 // hence on your typical architecture store-load forwarding does not take
936 // place. Vectorizing in such cases does not make sense.
937 // Store-load forwarding distance.
938 const unsigned NumCyclesForStoreLoadThroughMemory = 8*TypeByteSize;
939 // Maximum vector factor.
940 unsigned MaxVFWithoutSLForwardIssues =
941 VectorizerParams::MaxVectorWidth * TypeByteSize;
942 if(MaxSafeDepDistBytes < MaxVFWithoutSLForwardIssues)
943 MaxVFWithoutSLForwardIssues = MaxSafeDepDistBytes;
945 for (unsigned vf = 2*TypeByteSize; vf <= MaxVFWithoutSLForwardIssues;
947 if (Distance % vf && Distance / vf < NumCyclesForStoreLoadThroughMemory) {
948 MaxVFWithoutSLForwardIssues = (vf >>=1);
953 if (MaxVFWithoutSLForwardIssues< 2*TypeByteSize) {
954 DEBUG(dbgs() << "LAA: Distance " << Distance <<
955 " that could cause a store-load forwarding conflict\n");
959 if (MaxVFWithoutSLForwardIssues < MaxSafeDepDistBytes &&
960 MaxVFWithoutSLForwardIssues !=
961 VectorizerParams::MaxVectorWidth * TypeByteSize)
962 MaxSafeDepDistBytes = MaxVFWithoutSLForwardIssues;
966 /// \brief Check the dependence for two accesses with the same stride \p Stride.
967 /// \p Distance is the positive distance and \p TypeByteSize is type size in
970 /// \returns true if they are independent.
971 static bool areStridedAccessesIndependent(unsigned Distance, unsigned Stride,
972 unsigned TypeByteSize) {
973 assert(Stride > 1 && "The stride must be greater than 1");
974 assert(TypeByteSize > 0 && "The type size in byte must be non-zero");
975 assert(Distance > 0 && "The distance must be non-zero");
977 // Skip if the distance is not multiple of type byte size.
978 if (Distance % TypeByteSize)
981 unsigned ScaledDist = Distance / TypeByteSize;
983 // No dependence if the scaled distance is not multiple of the stride.
985 // for (i = 0; i < 1024 ; i += 4)
986 // A[i+2] = A[i] + 1;
988 // Two accesses in memory (scaled distance is 2, stride is 4):
989 // | A[0] | | | | A[4] | | | |
990 // | | | A[2] | | | | A[6] | |
993 // for (i = 0; i < 1024 ; i += 3)
994 // A[i+4] = A[i] + 1;
996 // Two accesses in memory (scaled distance is 4, stride is 3):
997 // | A[0] | | | A[3] | | | A[6] | | |
998 // | | | | | A[4] | | | A[7] | |
999 return ScaledDist % Stride;
1002 MemoryDepChecker::Dependence::DepType
1003 MemoryDepChecker::isDependent(const MemAccessInfo &A, unsigned AIdx,
1004 const MemAccessInfo &B, unsigned BIdx,
1005 const ValueToValueMap &Strides) {
1006 assert (AIdx < BIdx && "Must pass arguments in program order");
1008 Value *APtr = A.getPointer();
1009 Value *BPtr = B.getPointer();
1010 bool AIsWrite = A.getInt();
1011 bool BIsWrite = B.getInt();
1013 // Two reads are independent.
1014 if (!AIsWrite && !BIsWrite)
1015 return Dependence::NoDep;
1017 // We cannot check pointers in different address spaces.
1018 if (APtr->getType()->getPointerAddressSpace() !=
1019 BPtr->getType()->getPointerAddressSpace())
1020 return Dependence::Unknown;
1022 const SCEV *AScev = replaceSymbolicStrideSCEV(SE, Strides, APtr);
1023 const SCEV *BScev = replaceSymbolicStrideSCEV(SE, Strides, BPtr);
1025 int StrideAPtr = isStridedPtr(SE, APtr, InnermostLoop, Strides);
1026 int StrideBPtr = isStridedPtr(SE, BPtr, InnermostLoop, Strides);
1028 const SCEV *Src = AScev;
1029 const SCEV *Sink = BScev;
1031 // If the induction step is negative we have to invert source and sink of the
1033 if (StrideAPtr < 0) {
1036 std::swap(APtr, BPtr);
1037 std::swap(Src, Sink);
1038 std::swap(AIsWrite, BIsWrite);
1039 std::swap(AIdx, BIdx);
1040 std::swap(StrideAPtr, StrideBPtr);
1043 const SCEV *Dist = SE->getMinusSCEV(Sink, Src);
1045 DEBUG(dbgs() << "LAA: Src Scev: " << *Src << "Sink Scev: " << *Sink
1046 << "(Induction step: " << StrideAPtr << ")\n");
1047 DEBUG(dbgs() << "LAA: Distance for " << *InstMap[AIdx] << " to "
1048 << *InstMap[BIdx] << ": " << *Dist << "\n");
1050 // Need accesses with constant stride. We don't want to vectorize
1051 // "A[B[i]] += ..." and similar code or pointer arithmetic that could wrap in
1052 // the address space.
1053 if (!StrideAPtr || !StrideBPtr || StrideAPtr != StrideBPtr){
1054 DEBUG(dbgs() << "Pointer access with non-constant stride\n");
1055 return Dependence::Unknown;
1058 const SCEVConstant *C = dyn_cast<SCEVConstant>(Dist);
1060 DEBUG(dbgs() << "LAA: Dependence because of non-constant distance\n");
1061 ShouldRetryWithRuntimeCheck = true;
1062 return Dependence::Unknown;
1065 Type *ATy = APtr->getType()->getPointerElementType();
1066 Type *BTy = BPtr->getType()->getPointerElementType();
1067 auto &DL = InnermostLoop->getHeader()->getModule()->getDataLayout();
1068 unsigned TypeByteSize = DL.getTypeAllocSize(ATy);
1070 // Negative distances are not plausible dependencies.
1071 const APInt &Val = C->getValue()->getValue();
1072 if (Val.isNegative()) {
1073 bool IsTrueDataDependence = (AIsWrite && !BIsWrite);
1074 if (IsTrueDataDependence &&
1075 (couldPreventStoreLoadForward(Val.abs().getZExtValue(), TypeByteSize) ||
1077 return Dependence::ForwardButPreventsForwarding;
1079 DEBUG(dbgs() << "LAA: Dependence is negative: NoDep\n");
1080 return Dependence::Forward;
1083 // Write to the same location with the same size.
1084 // Could be improved to assert type sizes are the same (i32 == float, etc).
1087 return Dependence::NoDep;
1088 DEBUG(dbgs() << "LAA: Zero dependence difference but different types\n");
1089 return Dependence::Unknown;
1092 assert(Val.isStrictlyPositive() && "Expect a positive value");
1096 "LAA: ReadWrite-Write positive dependency with different types\n");
1097 return Dependence::Unknown;
1100 unsigned Distance = (unsigned) Val.getZExtValue();
1102 unsigned Stride = std::abs(StrideAPtr);
1104 areStridedAccessesIndependent(Distance, Stride, TypeByteSize)) {
1105 DEBUG(dbgs() << "LAA: Strided accesses are independent\n");
1106 return Dependence::NoDep;
1109 // Bail out early if passed-in parameters make vectorization not feasible.
1110 unsigned ForcedFactor = (VectorizerParams::VectorizationFactor ?
1111 VectorizerParams::VectorizationFactor : 1);
1112 unsigned ForcedUnroll = (VectorizerParams::VectorizationInterleave ?
1113 VectorizerParams::VectorizationInterleave : 1);
1114 // The minimum number of iterations for a vectorized/unrolled version.
1115 unsigned MinNumIter = std::max(ForcedFactor * ForcedUnroll, 2U);
1117 // It's not vectorizable if the distance is smaller than the minimum distance
1118 // needed for a vectroized/unrolled version. Vectorizing one iteration in
1119 // front needs TypeByteSize * Stride. Vectorizing the last iteration needs
1120 // TypeByteSize (No need to plus the last gap distance).
1122 // E.g. Assume one char is 1 byte in memory and one int is 4 bytes.
1124 // int *B = (int *)((char *)A + 14);
1125 // for (i = 0 ; i < 1024 ; i += 2)
1129 // Two accesses in memory (stride is 2):
1130 // | A[0] | | A[2] | | A[4] | | A[6] | |
1131 // | B[0] | | B[2] | | B[4] |
1133 // Distance needs for vectorizing iterations except the last iteration:
1134 // 4 * 2 * (MinNumIter - 1). Distance needs for the last iteration: 4.
1135 // So the minimum distance needed is: 4 * 2 * (MinNumIter - 1) + 4.
1137 // If MinNumIter is 2, it is vectorizable as the minimum distance needed is
1138 // 12, which is less than distance.
1140 // If MinNumIter is 4 (Say if a user forces the vectorization factor to be 4),
1141 // the minimum distance needed is 28, which is greater than distance. It is
1142 // not safe to do vectorization.
1143 unsigned MinDistanceNeeded =
1144 TypeByteSize * Stride * (MinNumIter - 1) + TypeByteSize;
1145 if (MinDistanceNeeded > Distance) {
1146 DEBUG(dbgs() << "LAA: Failure because of positive distance " << Distance
1148 return Dependence::Backward;
1151 // Unsafe if the minimum distance needed is greater than max safe distance.
1152 if (MinDistanceNeeded > MaxSafeDepDistBytes) {
1153 DEBUG(dbgs() << "LAA: Failure because it needs at least "
1154 << MinDistanceNeeded << " size in bytes");
1155 return Dependence::Backward;
1158 // Positive distance bigger than max vectorization factor.
1159 // FIXME: Should use max factor instead of max distance in bytes, which could
1160 // not handle different types.
1161 // E.g. Assume one char is 1 byte in memory and one int is 4 bytes.
1162 // void foo (int *A, char *B) {
1163 // for (unsigned i = 0; i < 1024; i++) {
1164 // A[i+2] = A[i] + 1;
1165 // B[i+2] = B[i] + 1;
1169 // This case is currently unsafe according to the max safe distance. If we
1170 // analyze the two accesses on array B, the max safe dependence distance
1171 // is 2. Then we analyze the accesses on array A, the minimum distance needed
1172 // is 8, which is less than 2 and forbidden vectorization, But actually
1173 // both A and B could be vectorized by 2 iterations.
1174 MaxSafeDepDistBytes =
1175 Distance < MaxSafeDepDistBytes ? Distance : MaxSafeDepDistBytes;
1177 bool IsTrueDataDependence = (!AIsWrite && BIsWrite);
1178 if (IsTrueDataDependence &&
1179 couldPreventStoreLoadForward(Distance, TypeByteSize))
1180 return Dependence::BackwardVectorizableButPreventsForwarding;
1182 DEBUG(dbgs() << "LAA: Positive distance " << Val.getSExtValue()
1183 << " with max VF = "
1184 << MaxSafeDepDistBytes / (TypeByteSize * Stride) << '\n');
1186 return Dependence::BackwardVectorizable;
1189 bool MemoryDepChecker::areDepsSafe(DepCandidates &AccessSets,
1190 MemAccessInfoSet &CheckDeps,
1191 const ValueToValueMap &Strides) {
1193 MaxSafeDepDistBytes = -1U;
1194 while (!CheckDeps.empty()) {
1195 MemAccessInfo CurAccess = *CheckDeps.begin();
1197 // Get the relevant memory access set.
1198 EquivalenceClasses<MemAccessInfo>::iterator I =
1199 AccessSets.findValue(AccessSets.getLeaderValue(CurAccess));
1201 // Check accesses within this set.
1202 EquivalenceClasses<MemAccessInfo>::member_iterator AI, AE;
1203 AI = AccessSets.member_begin(I), AE = AccessSets.member_end();
1205 // Check every access pair.
1207 CheckDeps.erase(*AI);
1208 EquivalenceClasses<MemAccessInfo>::member_iterator OI = std::next(AI);
1210 // Check every accessing instruction pair in program order.
1211 for (std::vector<unsigned>::iterator I1 = Accesses[*AI].begin(),
1212 I1E = Accesses[*AI].end(); I1 != I1E; ++I1)
1213 for (std::vector<unsigned>::iterator I2 = Accesses[*OI].begin(),
1214 I2E = Accesses[*OI].end(); I2 != I2E; ++I2) {
1215 auto A = std::make_pair(&*AI, *I1);
1216 auto B = std::make_pair(&*OI, *I2);
1222 Dependence::DepType Type =
1223 isDependent(*A.first, A.second, *B.first, B.second, Strides);
1224 SafeForVectorization &= Dependence::isSafeForVectorization(Type);
1226 // Gather dependences unless we accumulated MaxInterestingDependence
1227 // dependences. In that case return as soon as we find the first
1228 // unsafe dependence. This puts a limit on this quadratic
1230 if (RecordInterestingDependences) {
1231 if (Dependence::isInterestingDependence(Type))
1232 InterestingDependences.push_back(
1233 Dependence(A.second, B.second, Type));
1235 if (InterestingDependences.size() >= MaxInterestingDependence) {
1236 RecordInterestingDependences = false;
1237 InterestingDependences.clear();
1238 DEBUG(dbgs() << "Too many dependences, stopped recording\n");
1241 if (!RecordInterestingDependences && !SafeForVectorization)
1250 DEBUG(dbgs() << "Total Interesting Dependences: "
1251 << InterestingDependences.size() << "\n");
1252 return SafeForVectorization;
1255 SmallVector<Instruction *, 4>
1256 MemoryDepChecker::getInstructionsForAccess(Value *Ptr, bool isWrite) const {
1257 MemAccessInfo Access(Ptr, isWrite);
1258 auto &IndexVector = Accesses.find(Access)->second;
1260 SmallVector<Instruction *, 4> Insts;
1261 std::transform(IndexVector.begin(), IndexVector.end(),
1262 std::back_inserter(Insts),
1263 [&](unsigned Idx) { return this->InstMap[Idx]; });
1267 const char *MemoryDepChecker::Dependence::DepName[] = {
1268 "NoDep", "Unknown", "Forward", "ForwardButPreventsForwarding", "Backward",
1269 "BackwardVectorizable", "BackwardVectorizableButPreventsForwarding"};
1271 void MemoryDepChecker::Dependence::print(
1272 raw_ostream &OS, unsigned Depth,
1273 const SmallVectorImpl<Instruction *> &Instrs) const {
1274 OS.indent(Depth) << DepName[Type] << ":\n";
1275 OS.indent(Depth + 2) << *Instrs[Source] << " -> \n";
1276 OS.indent(Depth + 2) << *Instrs[Destination] << "\n";
1279 bool LoopAccessInfo::canAnalyzeLoop() {
1280 // We need to have a loop header.
1281 DEBUG(dbgs() << "LAA: Found a loop: " <<
1282 TheLoop->getHeader()->getName() << '\n');
1284 // We can only analyze innermost loops.
1285 if (!TheLoop->empty()) {
1286 DEBUG(dbgs() << "LAA: loop is not the innermost loop\n");
1287 emitAnalysis(LoopAccessReport() << "loop is not the innermost loop");
1291 // We must have a single backedge.
1292 if (TheLoop->getNumBackEdges() != 1) {
1293 DEBUG(dbgs() << "LAA: loop control flow is not understood by analyzer\n");
1295 LoopAccessReport() <<
1296 "loop control flow is not understood by analyzer");
1300 // We must have a single exiting block.
1301 if (!TheLoop->getExitingBlock()) {
1302 DEBUG(dbgs() << "LAA: loop control flow is not understood by analyzer\n");
1304 LoopAccessReport() <<
1305 "loop control flow is not understood by analyzer");
1309 // We only handle bottom-tested loops, i.e. loop in which the condition is
1310 // checked at the end of each iteration. With that we can assume that all
1311 // instructions in the loop are executed the same number of times.
1312 if (TheLoop->getExitingBlock() != TheLoop->getLoopLatch()) {
1313 DEBUG(dbgs() << "LAA: loop control flow is not understood by analyzer\n");
1315 LoopAccessReport() <<
1316 "loop control flow is not understood by analyzer");
1320 // ScalarEvolution needs to be able to find the exit count.
1321 const SCEV *ExitCount = SE->getBackedgeTakenCount(TheLoop);
1322 if (ExitCount == SE->getCouldNotCompute()) {
1323 emitAnalysis(LoopAccessReport() <<
1324 "could not determine number of loop iterations");
1325 DEBUG(dbgs() << "LAA: SCEV could not compute the loop exit count.\n");
1332 void LoopAccessInfo::analyzeLoop(const ValueToValueMap &Strides) {
1334 typedef SmallVector<Value*, 16> ValueVector;
1335 typedef SmallPtrSet<Value*, 16> ValueSet;
1337 // Holds the Load and Store *instructions*.
1341 // Holds all the different accesses in the loop.
1342 unsigned NumReads = 0;
1343 unsigned NumReadWrites = 0;
1345 PtrRtChecking.Pointers.clear();
1346 PtrRtChecking.Need = false;
1348 const bool IsAnnotatedParallel = TheLoop->isAnnotatedParallel();
1351 for (Loop::block_iterator bb = TheLoop->block_begin(),
1352 be = TheLoop->block_end(); bb != be; ++bb) {
1354 // Scan the BB and collect legal loads and stores.
1355 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
1358 // If this is a load, save it. If this instruction can read from memory
1359 // but is not a load, then we quit. Notice that we don't handle function
1360 // calls that read or write.
1361 if (it->mayReadFromMemory()) {
1362 // Many math library functions read the rounding mode. We will only
1363 // vectorize a loop if it contains known function calls that don't set
1364 // the flag. Therefore, it is safe to ignore this read from memory.
1365 CallInst *Call = dyn_cast<CallInst>(it);
1366 if (Call && getIntrinsicIDForCall(Call, TLI))
1369 // If the function has an explicit vectorized counterpart, we can safely
1370 // assume that it can be vectorized.
1371 if (Call && !Call->isNoBuiltin() && Call->getCalledFunction() &&
1372 TLI->isFunctionVectorizable(Call->getCalledFunction()->getName()))
1375 LoadInst *Ld = dyn_cast<LoadInst>(it);
1376 if (!Ld || (!Ld->isSimple() && !IsAnnotatedParallel)) {
1377 emitAnalysis(LoopAccessReport(Ld)
1378 << "read with atomic ordering or volatile read");
1379 DEBUG(dbgs() << "LAA: Found a non-simple load.\n");
1384 Loads.push_back(Ld);
1385 DepChecker.addAccess(Ld);
1389 // Save 'store' instructions. Abort if other instructions write to memory.
1390 if (it->mayWriteToMemory()) {
1391 StoreInst *St = dyn_cast<StoreInst>(it);
1393 emitAnalysis(LoopAccessReport(it) <<
1394 "instruction cannot be vectorized");
1398 if (!St->isSimple() && !IsAnnotatedParallel) {
1399 emitAnalysis(LoopAccessReport(St)
1400 << "write with atomic ordering or volatile write");
1401 DEBUG(dbgs() << "LAA: Found a non-simple store.\n");
1406 Stores.push_back(St);
1407 DepChecker.addAccess(St);
1412 // Now we have two lists that hold the loads and the stores.
1413 // Next, we find the pointers that they use.
1415 // Check if we see any stores. If there are no stores, then we don't
1416 // care if the pointers are *restrict*.
1417 if (!Stores.size()) {
1418 DEBUG(dbgs() << "LAA: Found a read-only loop!\n");
1423 MemoryDepChecker::DepCandidates DependentAccesses;
1424 AccessAnalysis Accesses(TheLoop->getHeader()->getModule()->getDataLayout(),
1425 AA, LI, DependentAccesses);
1427 // Holds the analyzed pointers. We don't want to call GetUnderlyingObjects
1428 // multiple times on the same object. If the ptr is accessed twice, once
1429 // for read and once for write, it will only appear once (on the write
1430 // list). This is okay, since we are going to check for conflicts between
1431 // writes and between reads and writes, but not between reads and reads.
1434 ValueVector::iterator I, IE;
1435 for (I = Stores.begin(), IE = Stores.end(); I != IE; ++I) {
1436 StoreInst *ST = cast<StoreInst>(*I);
1437 Value* Ptr = ST->getPointerOperand();
1438 // Check for store to loop invariant address.
1439 StoreToLoopInvariantAddress |= isUniform(Ptr);
1440 // If we did *not* see this pointer before, insert it to the read-write
1441 // list. At this phase it is only a 'write' list.
1442 if (Seen.insert(Ptr).second) {
1445 MemoryLocation Loc = MemoryLocation::get(ST);
1446 // The TBAA metadata could have a control dependency on the predication
1447 // condition, so we cannot rely on it when determining whether or not we
1448 // need runtime pointer checks.
1449 if (blockNeedsPredication(ST->getParent(), TheLoop, DT))
1450 Loc.AATags.TBAA = nullptr;
1452 Accesses.addStore(Loc);
1456 if (IsAnnotatedParallel) {
1458 << "LAA: A loop annotated parallel, ignore memory dependency "
1464 for (I = Loads.begin(), IE = Loads.end(); I != IE; ++I) {
1465 LoadInst *LD = cast<LoadInst>(*I);
1466 Value* Ptr = LD->getPointerOperand();
1467 // If we did *not* see this pointer before, insert it to the
1468 // read list. If we *did* see it before, then it is already in
1469 // the read-write list. This allows us to vectorize expressions
1470 // such as A[i] += x; Because the address of A[i] is a read-write
1471 // pointer. This only works if the index of A[i] is consecutive.
1472 // If the address of i is unknown (for example A[B[i]]) then we may
1473 // read a few words, modify, and write a few words, and some of the
1474 // words may be written to the same address.
1475 bool IsReadOnlyPtr = false;
1476 if (Seen.insert(Ptr).second || !isStridedPtr(SE, Ptr, TheLoop, Strides)) {
1478 IsReadOnlyPtr = true;
1481 MemoryLocation Loc = MemoryLocation::get(LD);
1482 // The TBAA metadata could have a control dependency on the predication
1483 // condition, so we cannot rely on it when determining whether or not we
1484 // need runtime pointer checks.
1485 if (blockNeedsPredication(LD->getParent(), TheLoop, DT))
1486 Loc.AATags.TBAA = nullptr;
1488 Accesses.addLoad(Loc, IsReadOnlyPtr);
1491 // If we write (or read-write) to a single destination and there are no
1492 // other reads in this loop then is it safe to vectorize.
1493 if (NumReadWrites == 1 && NumReads == 0) {
1494 DEBUG(dbgs() << "LAA: Found a write-only loop!\n");
1499 // Build dependence sets and check whether we need a runtime pointer bounds
1501 Accesses.buildDependenceSets();
1503 // Find pointers with computable bounds. We are going to use this information
1504 // to place a runtime bound check.
1505 bool CanDoRTIfNeeded =
1506 Accesses.canCheckPtrAtRT(PtrRtChecking, SE, TheLoop, Strides);
1507 if (!CanDoRTIfNeeded) {
1508 emitAnalysis(LoopAccessReport() << "cannot identify array bounds");
1509 DEBUG(dbgs() << "LAA: We can't vectorize because we can't find "
1510 << "the array bounds.\n");
1515 DEBUG(dbgs() << "LAA: We can perform a memory runtime check if needed.\n");
1518 if (Accesses.isDependencyCheckNeeded()) {
1519 DEBUG(dbgs() << "LAA: Checking memory dependencies\n");
1520 CanVecMem = DepChecker.areDepsSafe(
1521 DependentAccesses, Accesses.getDependenciesToCheck(), Strides);
1522 MaxSafeDepDistBytes = DepChecker.getMaxSafeDepDistBytes();
1524 if (!CanVecMem && DepChecker.shouldRetryWithRuntimeCheck()) {
1525 DEBUG(dbgs() << "LAA: Retrying with memory checks\n");
1527 // Clear the dependency checks. We assume they are not needed.
1528 Accesses.resetDepChecks(DepChecker);
1530 PtrRtChecking.reset();
1531 PtrRtChecking.Need = true;
1534 Accesses.canCheckPtrAtRT(PtrRtChecking, SE, TheLoop, Strides, true);
1536 // Check that we found the bounds for the pointer.
1537 if (!CanDoRTIfNeeded) {
1538 emitAnalysis(LoopAccessReport()
1539 << "cannot check memory dependencies at runtime");
1540 DEBUG(dbgs() << "LAA: Can't vectorize with memory checks\n");
1550 DEBUG(dbgs() << "LAA: No unsafe dependent memory operations in loop. We"
1551 << (PtrRtChecking.Need ? "" : " don't")
1552 << " need runtime memory checks.\n");
1554 emitAnalysis(LoopAccessReport() <<
1555 "unsafe dependent memory operations in loop");
1556 DEBUG(dbgs() << "LAA: unsafe dependent memory operations in loop\n");
1560 bool LoopAccessInfo::blockNeedsPredication(BasicBlock *BB, Loop *TheLoop,
1561 DominatorTree *DT) {
1562 assert(TheLoop->contains(BB) && "Unknown block used");
1564 // Blocks that do not dominate the latch need predication.
1565 BasicBlock* Latch = TheLoop->getLoopLatch();
1566 return !DT->dominates(BB, Latch);
1569 void LoopAccessInfo::emitAnalysis(LoopAccessReport &Message) {
1570 assert(!Report && "Multiple reports generated");
1574 bool LoopAccessInfo::isUniform(Value *V) const {
1575 return (SE->isLoopInvariant(SE->getSCEV(V), TheLoop));
1578 // FIXME: this function is currently a duplicate of the one in
1579 // LoopVectorize.cpp.
1580 static Instruction *getFirstInst(Instruction *FirstInst, Value *V,
1584 if (Instruction *I = dyn_cast<Instruction>(V))
1585 return I->getParent() == Loc->getParent() ? I : nullptr;
1589 std::pair<Instruction *, Instruction *> LoopAccessInfo::addRuntimeCheck(
1590 Instruction *Loc, const SmallVectorImpl<int> *PtrPartition) const {
1591 if (!PtrRtChecking.Need)
1592 return std::make_pair(nullptr, nullptr);
1594 SmallVector<TrackingVH<Value>, 2> Starts;
1595 SmallVector<TrackingVH<Value>, 2> Ends;
1597 LLVMContext &Ctx = Loc->getContext();
1598 SCEVExpander Exp(*SE, DL, "induction");
1599 Instruction *FirstInst = nullptr;
1601 for (unsigned i = 0; i < PtrRtChecking.CheckingGroups.size(); ++i) {
1602 const RuntimePointerChecking::CheckingPtrGroup &CG =
1603 PtrRtChecking.CheckingGroups[i];
1604 Value *Ptr = PtrRtChecking.Pointers[CG.Members[0]].PointerValue;
1605 const SCEV *Sc = SE->getSCEV(Ptr);
1607 if (SE->isLoopInvariant(Sc, TheLoop)) {
1608 DEBUG(dbgs() << "LAA: Adding RT check for a loop invariant ptr:" << *Ptr
1610 Starts.push_back(Ptr);
1611 Ends.push_back(Ptr);
1613 unsigned AS = Ptr->getType()->getPointerAddressSpace();
1615 // Use this type for pointer arithmetic.
1616 Type *PtrArithTy = Type::getInt8PtrTy(Ctx, AS);
1617 Value *Start = nullptr, *End = nullptr;
1619 DEBUG(dbgs() << "LAA: Adding RT check for range:\n");
1620 Start = Exp.expandCodeFor(CG.Low, PtrArithTy, Loc);
1621 End = Exp.expandCodeFor(CG.High, PtrArithTy, Loc);
1622 DEBUG(dbgs() << "Start: " << *CG.Low << " End: " << *CG.High << "\n");
1623 Starts.push_back(Start);
1624 Ends.push_back(End);
1628 IRBuilder<> ChkBuilder(Loc);
1629 // Our instructions might fold to a constant.
1630 Value *MemoryRuntimeCheck = nullptr;
1631 for (unsigned i = 0; i < PtrRtChecking.CheckingGroups.size(); ++i) {
1632 for (unsigned j = i + 1; j < PtrRtChecking.CheckingGroups.size(); ++j) {
1633 const RuntimePointerChecking::CheckingPtrGroup &CGI =
1634 PtrRtChecking.CheckingGroups[i];
1635 const RuntimePointerChecking::CheckingPtrGroup &CGJ =
1636 PtrRtChecking.CheckingGroups[j];
1638 if (!PtrRtChecking.needsChecking(CGI, CGJ, PtrPartition))
1641 unsigned AS0 = Starts[i]->getType()->getPointerAddressSpace();
1642 unsigned AS1 = Starts[j]->getType()->getPointerAddressSpace();
1644 assert((AS0 == Ends[j]->getType()->getPointerAddressSpace()) &&
1645 (AS1 == Ends[i]->getType()->getPointerAddressSpace()) &&
1646 "Trying to bounds check pointers with different address spaces");
1648 Type *PtrArithTy0 = Type::getInt8PtrTy(Ctx, AS0);
1649 Type *PtrArithTy1 = Type::getInt8PtrTy(Ctx, AS1);
1651 Value *Start0 = ChkBuilder.CreateBitCast(Starts[i], PtrArithTy0, "bc");
1652 Value *Start1 = ChkBuilder.CreateBitCast(Starts[j], PtrArithTy1, "bc");
1653 Value *End0 = ChkBuilder.CreateBitCast(Ends[i], PtrArithTy1, "bc");
1654 Value *End1 = ChkBuilder.CreateBitCast(Ends[j], PtrArithTy0, "bc");
1656 Value *Cmp0 = ChkBuilder.CreateICmpULE(Start0, End1, "bound0");
1657 FirstInst = getFirstInst(FirstInst, Cmp0, Loc);
1658 Value *Cmp1 = ChkBuilder.CreateICmpULE(Start1, End0, "bound1");
1659 FirstInst = getFirstInst(FirstInst, Cmp1, Loc);
1660 Value *IsConflict = ChkBuilder.CreateAnd(Cmp0, Cmp1, "found.conflict");
1661 FirstInst = getFirstInst(FirstInst, IsConflict, Loc);
1662 if (MemoryRuntimeCheck) {
1663 IsConflict = ChkBuilder.CreateOr(MemoryRuntimeCheck, IsConflict,
1665 FirstInst = getFirstInst(FirstInst, IsConflict, Loc);
1667 MemoryRuntimeCheck = IsConflict;
1671 if (!MemoryRuntimeCheck)
1672 return std::make_pair(nullptr, nullptr);
1674 // We have to do this trickery because the IRBuilder might fold the check to a
1675 // constant expression in which case there is no Instruction anchored in a
1677 Instruction *Check = BinaryOperator::CreateAnd(MemoryRuntimeCheck,
1678 ConstantInt::getTrue(Ctx));
1679 ChkBuilder.Insert(Check, "memcheck.conflict");
1680 FirstInst = getFirstInst(FirstInst, Check, Loc);
1681 return std::make_pair(FirstInst, Check);
1684 LoopAccessInfo::LoopAccessInfo(Loop *L, ScalarEvolution *SE,
1685 const DataLayout &DL,
1686 const TargetLibraryInfo *TLI, AliasAnalysis *AA,
1687 DominatorTree *DT, LoopInfo *LI,
1688 const ValueToValueMap &Strides)
1689 : PtrRtChecking(SE), DepChecker(SE, L), TheLoop(L), SE(SE), DL(DL),
1690 TLI(TLI), AA(AA), DT(DT), LI(LI), NumLoads(0), NumStores(0),
1691 MaxSafeDepDistBytes(-1U), CanVecMem(false),
1692 StoreToLoopInvariantAddress(false) {
1693 if (canAnalyzeLoop())
1694 analyzeLoop(Strides);
1697 void LoopAccessInfo::print(raw_ostream &OS, unsigned Depth) const {
1699 if (PtrRtChecking.Need)
1700 OS.indent(Depth) << "Memory dependences are safe with run-time checks\n";
1702 OS.indent(Depth) << "Memory dependences are safe\n";
1706 OS.indent(Depth) << "Report: " << Report->str() << "\n";
1708 if (auto *InterestingDependences = DepChecker.getInterestingDependences()) {
1709 OS.indent(Depth) << "Interesting Dependences:\n";
1710 for (auto &Dep : *InterestingDependences) {
1711 Dep.print(OS, Depth + 2, DepChecker.getMemoryInstructions());
1715 OS.indent(Depth) << "Too many interesting dependences, not recorded\n";
1717 // List the pair of accesses need run-time checks to prove independence.
1718 PtrRtChecking.print(OS, Depth);
1721 OS.indent(Depth) << "Store to invariant address was "
1722 << (StoreToLoopInvariantAddress ? "" : "not ")
1723 << "found in loop.\n";
1726 const LoopAccessInfo &
1727 LoopAccessAnalysis::getInfo(Loop *L, const ValueToValueMap &Strides) {
1728 auto &LAI = LoopAccessInfoMap[L];
1731 assert((!LAI || LAI->NumSymbolicStrides == Strides.size()) &&
1732 "Symbolic strides changed for loop");
1736 const DataLayout &DL = L->getHeader()->getModule()->getDataLayout();
1737 LAI = llvm::make_unique<LoopAccessInfo>(L, SE, DL, TLI, AA, DT, LI,
1740 LAI->NumSymbolicStrides = Strides.size();
1746 void LoopAccessAnalysis::print(raw_ostream &OS, const Module *M) const {
1747 LoopAccessAnalysis &LAA = *const_cast<LoopAccessAnalysis *>(this);
1749 ValueToValueMap NoSymbolicStrides;
1751 for (Loop *TopLevelLoop : *LI)
1752 for (Loop *L : depth_first(TopLevelLoop)) {
1753 OS.indent(2) << L->getHeader()->getName() << ":\n";
1754 auto &LAI = LAA.getInfo(L, NoSymbolicStrides);
1759 bool LoopAccessAnalysis::runOnFunction(Function &F) {
1760 SE = &getAnalysis<ScalarEvolution>();
1761 auto *TLIP = getAnalysisIfAvailable<TargetLibraryInfoWrapperPass>();
1762 TLI = TLIP ? &TLIP->getTLI() : nullptr;
1763 AA = &getAnalysis<AliasAnalysis>();
1764 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1765 LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
1770 void LoopAccessAnalysis::getAnalysisUsage(AnalysisUsage &AU) const {
1771 AU.addRequired<ScalarEvolution>();
1772 AU.addRequired<AliasAnalysis>();
1773 AU.addRequired<DominatorTreeWrapperPass>();
1774 AU.addRequired<LoopInfoWrapperPass>();
1776 AU.setPreservesAll();
1779 char LoopAccessAnalysis::ID = 0;
1780 static const char laa_name[] = "Loop Access Analysis";
1781 #define LAA_NAME "loop-accesses"
1783 INITIALIZE_PASS_BEGIN(LoopAccessAnalysis, LAA_NAME, laa_name, false, true)
1784 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
1785 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
1786 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
1787 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
1788 INITIALIZE_PASS_END(LoopAccessAnalysis, LAA_NAME, laa_name, false, true)
1791 Pass *createLAAPass() {
1792 return new LoopAccessAnalysis();