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);
130 const SCEV *ScEnd = AR->evaluateAtIteration(Ex, *SE);
131 Pointers.emplace_back(Ptr, AR->getStart(), ScEnd, WritePtr, DepSetId, ASId,
135 bool RuntimePointerChecking::needsChecking(
136 const CheckingPtrGroup &M, const CheckingPtrGroup &N,
137 const SmallVectorImpl<int> *PtrPartition) const {
138 for (unsigned I = 0, EI = M.Members.size(); EI != I; ++I)
139 for (unsigned J = 0, EJ = N.Members.size(); EJ != J; ++J)
140 if (needsChecking(M.Members[I], N.Members[J], PtrPartition))
145 /// Compare \p I and \p J and return the minimum.
146 /// Return nullptr in case we couldn't find an answer.
147 static const SCEV *getMinFromExprs(const SCEV *I, const SCEV *J,
148 ScalarEvolution *SE) {
149 const SCEV *Diff = SE->getMinusSCEV(J, I);
150 const SCEVConstant *C = dyn_cast<const SCEVConstant>(Diff);
154 if (C->getValue()->isNegative())
159 bool RuntimePointerChecking::CheckingPtrGroup::addPointer(unsigned Index) {
160 const SCEV *Start = RtCheck.Pointers[Index].Start;
161 const SCEV *End = RtCheck.Pointers[Index].End;
163 // Compare the starts and ends with the known minimum and maximum
164 // of this set. We need to know how we compare against the min/max
165 // of the set in order to be able to emit memchecks.
166 const SCEV *Min0 = getMinFromExprs(Start, Low, RtCheck.SE);
170 const SCEV *Min1 = getMinFromExprs(End, High, RtCheck.SE);
174 // Update the low bound expression if we've found a new min value.
178 // Update the high bound expression if we've found a new max value.
182 Members.push_back(Index);
186 void RuntimePointerChecking::groupChecks(
187 MemoryDepChecker::DepCandidates &DepCands, bool UseDependencies) {
188 // We build the groups from dependency candidates equivalence classes
190 // - We know that pointers in the same equivalence class share
191 // the same underlying object and therefore there is a chance
192 // that we can compare pointers
193 // - We wouldn't be able to merge two pointers for which we need
194 // to emit a memcheck. The classes in DepCands are already
195 // conveniently built such that no two pointers in the same
196 // class need checking against each other.
198 // We use the following (greedy) algorithm to construct the groups
199 // For every pointer in the equivalence class:
200 // For each existing group:
201 // - if the difference between this pointer and the min/max bounds
202 // of the group is a constant, then make the pointer part of the
203 // group and update the min/max bounds of that group as required.
205 CheckingGroups.clear();
207 // If we don't have the dependency partitions, construct a new
208 // checking pointer group for each pointer.
209 if (!UseDependencies) {
210 for (unsigned I = 0; I < Pointers.size(); ++I)
211 CheckingGroups.push_back(CheckingPtrGroup(I, *this));
215 unsigned TotalComparisons = 0;
217 DenseMap<Value *, unsigned> PositionMap;
218 for (unsigned Index = 0; Index < Pointers.size(); ++Index)
219 PositionMap[Pointers[Index].PointerValue] = Index;
221 // We need to keep track of what pointers we've already seen so we
222 // don't process them twice.
223 SmallSet<unsigned, 2> Seen;
225 // Go through all equivalence classes, get the the "pointer check groups"
226 // and add them to the overall solution. We use the order in which accesses
227 // appear in 'Pointers' to enforce determinism.
228 for (unsigned I = 0; I < Pointers.size(); ++I) {
229 // We've seen this pointer before, and therefore already processed
230 // its equivalence class.
234 MemoryDepChecker::MemAccessInfo Access(Pointers[I].PointerValue,
235 Pointers[I].IsWritePtr);
237 SmallVector<CheckingPtrGroup, 2> Groups;
238 auto LeaderI = DepCands.findValue(DepCands.getLeaderValue(Access));
240 // Because DepCands is constructed by visiting accesses in the order in
241 // which they appear in alias sets (which is deterministic) and the
242 // iteration order within an equivalence class member is only dependent on
243 // the order in which unions and insertions are performed on the
244 // equivalence class, the iteration order is deterministic.
245 for (auto MI = DepCands.member_begin(LeaderI), ME = DepCands.member_end();
247 unsigned Pointer = PositionMap[MI->getPointer()];
249 // Mark this pointer as seen.
250 Seen.insert(Pointer);
252 // Go through all the existing sets and see if we can find one
253 // which can include this pointer.
254 for (CheckingPtrGroup &Group : Groups) {
255 // Don't perform more than a certain amount of comparisons.
256 // This should limit the cost of grouping the pointers to something
257 // reasonable. If we do end up hitting this threshold, the algorithm
258 // will create separate groups for all remaining pointers.
259 if (TotalComparisons > MemoryCheckMergeThreshold)
264 if (Group.addPointer(Pointer)) {
271 // We couldn't add this pointer to any existing set or the threshold
272 // for the number of comparisons has been reached. Create a new group
273 // to hold the current pointer.
274 Groups.push_back(CheckingPtrGroup(Pointer, *this));
277 // We've computed the grouped checks for this partition.
278 // Save the results and continue with the next one.
279 std::copy(Groups.begin(), Groups.end(), std::back_inserter(CheckingGroups));
283 bool RuntimePointerChecking::needsChecking(
284 unsigned I, unsigned J, const SmallVectorImpl<int> *PtrPartition) const {
285 const PointerInfo &PointerI = Pointers[I];
286 const PointerInfo &PointerJ = Pointers[J];
288 // No need to check if two readonly pointers intersect.
289 if (!PointerI.IsWritePtr && !PointerJ.IsWritePtr)
292 // Only need to check pointers between two different dependency sets.
293 if (PointerI.DependencySetId == PointerJ.DependencySetId)
296 // Only need to check pointers in the same alias set.
297 if (PointerI.AliasSetId != PointerJ.AliasSetId)
300 // If PtrPartition is set omit checks between pointers of the same partition.
301 // Partition number -1 means that the pointer is used in multiple partitions.
302 // In this case we can't omit the check.
303 if (PtrPartition && (*PtrPartition)[I] != -1 &&
304 (*PtrPartition)[I] == (*PtrPartition)[J])
310 void RuntimePointerChecking::print(
311 raw_ostream &OS, unsigned Depth,
312 const SmallVectorImpl<int> *PtrPartition) const {
314 OS.indent(Depth) << "Run-time memory checks:\n";
317 for (unsigned I = 0; I < CheckingGroups.size(); ++I)
318 for (unsigned J = I + 1; J < CheckingGroups.size(); ++J)
319 if (needsChecking(CheckingGroups[I], CheckingGroups[J], PtrPartition)) {
320 OS.indent(Depth) << "Check " << N++ << ":\n";
321 OS.indent(Depth + 2) << "Comparing group " << I << ":\n";
323 for (unsigned K = 0; K < CheckingGroups[I].Members.size(); ++K) {
325 << *Pointers[CheckingGroups[I].Members[K]].PointerValue << "\n";
327 OS << " (Partition: "
328 << (*PtrPartition)[CheckingGroups[I].Members[K]] << ")"
332 OS.indent(Depth + 2) << "Against group " << J << ":\n";
334 for (unsigned K = 0; K < CheckingGroups[J].Members.size(); ++K) {
336 << *Pointers[CheckingGroups[J].Members[K]].PointerValue << "\n";
338 OS << " (Partition: "
339 << (*PtrPartition)[CheckingGroups[J].Members[K]] << ")"
344 OS.indent(Depth) << "Grouped accesses:\n";
345 for (unsigned I = 0; I < CheckingGroups.size(); ++I) {
346 OS.indent(Depth + 2) << "Group " << I << ":\n";
347 OS.indent(Depth + 4) << "(Low: " << *CheckingGroups[I].Low
348 << " High: " << *CheckingGroups[I].High << ")\n";
349 for (unsigned J = 0; J < CheckingGroups[I].Members.size(); ++J) {
350 OS.indent(Depth + 6) << "Member: "
351 << *Pointers[CheckingGroups[I].Members[J]].Expr
357 unsigned RuntimePointerChecking::getNumberOfChecks(
358 const SmallVectorImpl<int> *PtrPartition) const {
360 unsigned NumPartitions = CheckingGroups.size();
361 unsigned CheckCount = 0;
363 for (unsigned I = 0; I < NumPartitions; ++I)
364 for (unsigned J = I + 1; J < NumPartitions; ++J)
365 if (needsChecking(CheckingGroups[I], CheckingGroups[J], PtrPartition))
370 bool RuntimePointerChecking::needsAnyChecking(
371 const SmallVectorImpl<int> *PtrPartition) const {
372 unsigned NumPointers = Pointers.size();
374 for (unsigned I = 0; I < NumPointers; ++I)
375 for (unsigned J = I + 1; J < NumPointers; ++J)
376 if (needsChecking(I, J, PtrPartition))
382 /// \brief Analyses memory accesses in a loop.
384 /// Checks whether run time pointer checks are needed and builds sets for data
385 /// dependence checking.
386 class AccessAnalysis {
388 /// \brief Read or write access location.
389 typedef PointerIntPair<Value *, 1, bool> MemAccessInfo;
390 typedef SmallPtrSet<MemAccessInfo, 8> MemAccessInfoSet;
392 AccessAnalysis(const DataLayout &Dl, AliasAnalysis *AA, LoopInfo *LI,
393 MemoryDepChecker::DepCandidates &DA)
394 : DL(Dl), AST(*AA), LI(LI), DepCands(DA),
395 IsRTCheckAnalysisNeeded(false) {}
397 /// \brief Register a load and whether it is only read from.
398 void addLoad(MemoryLocation &Loc, bool IsReadOnly) {
399 Value *Ptr = const_cast<Value*>(Loc.Ptr);
400 AST.add(Ptr, MemoryLocation::UnknownSize, Loc.AATags);
401 Accesses.insert(MemAccessInfo(Ptr, false));
403 ReadOnlyPtr.insert(Ptr);
406 /// \brief Register a store.
407 void addStore(MemoryLocation &Loc) {
408 Value *Ptr = const_cast<Value*>(Loc.Ptr);
409 AST.add(Ptr, MemoryLocation::UnknownSize, Loc.AATags);
410 Accesses.insert(MemAccessInfo(Ptr, true));
413 /// \brief Check whether we can check the pointers at runtime for
414 /// non-intersection.
416 /// Returns true if we need no check or if we do and we can generate them
417 /// (i.e. the pointers have computable bounds).
418 bool canCheckPtrAtRT(RuntimePointerChecking &RtCheck, ScalarEvolution *SE,
419 Loop *TheLoop, const ValueToValueMap &Strides,
420 bool ShouldCheckStride = false);
422 /// \brief Goes over all memory accesses, checks whether a RT check is needed
423 /// and builds sets of dependent accesses.
424 void buildDependenceSets() {
425 processMemAccesses();
428 /// \brief Initial processing of memory accesses determined that we need to
429 /// perform dependency checking.
431 /// Note that this can later be cleared if we retry memcheck analysis without
432 /// dependency checking (i.e. ShouldRetryWithRuntimeCheck).
433 bool isDependencyCheckNeeded() { return !CheckDeps.empty(); }
435 /// We decided that no dependence analysis would be used. Reset the state.
436 void resetDepChecks(MemoryDepChecker &DepChecker) {
438 DepChecker.clearInterestingDependences();
441 MemAccessInfoSet &getDependenciesToCheck() { return CheckDeps; }
444 typedef SetVector<MemAccessInfo> PtrAccessSet;
446 /// \brief Go over all memory access and check whether runtime pointer checks
447 /// are needed and build sets of dependency check candidates.
448 void processMemAccesses();
450 /// Set of all accesses.
451 PtrAccessSet Accesses;
453 const DataLayout &DL;
455 /// Set of accesses that need a further dependence check.
456 MemAccessInfoSet CheckDeps;
458 /// Set of pointers that are read only.
459 SmallPtrSet<Value*, 16> ReadOnlyPtr;
461 /// An alias set tracker to partition the access set by underlying object and
462 //intrinsic property (such as TBAA metadata).
467 /// Sets of potentially dependent accesses - members of one set share an
468 /// underlying pointer. The set "CheckDeps" identfies which sets really need a
469 /// dependence check.
470 MemoryDepChecker::DepCandidates &DepCands;
472 /// \brief Initial processing of memory accesses determined that we may need
473 /// to add memchecks. Perform the analysis to determine the necessary checks.
475 /// Note that, this is different from isDependencyCheckNeeded. When we retry
476 /// memcheck analysis without dependency checking
477 /// (i.e. ShouldRetryWithRuntimeCheck), isDependencyCheckNeeded is cleared
478 /// while this remains set if we have potentially dependent accesses.
479 bool IsRTCheckAnalysisNeeded;
482 } // end anonymous namespace
484 /// \brief Check whether a pointer can participate in a runtime bounds check.
485 static bool hasComputableBounds(ScalarEvolution *SE,
486 const ValueToValueMap &Strides, Value *Ptr) {
487 const SCEV *PtrScev = replaceSymbolicStrideSCEV(SE, Strides, Ptr);
488 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
492 return AR->isAffine();
495 bool AccessAnalysis::canCheckPtrAtRT(RuntimePointerChecking &RtCheck,
496 ScalarEvolution *SE, Loop *TheLoop,
497 const ValueToValueMap &StridesMap,
498 bool ShouldCheckStride) {
499 // Find pointers with computable bounds. We are going to use this information
500 // to place a runtime bound check.
503 bool NeedRTCheck = false;
504 if (!IsRTCheckAnalysisNeeded) return true;
506 bool IsDepCheckNeeded = isDependencyCheckNeeded();
508 // We assign a consecutive id to access from different alias sets.
509 // Accesses between different groups doesn't need to be checked.
511 for (auto &AS : AST) {
512 int NumReadPtrChecks = 0;
513 int NumWritePtrChecks = 0;
515 // We assign consecutive id to access from different dependence sets.
516 // Accesses within the same set don't need a runtime check.
517 unsigned RunningDepId = 1;
518 DenseMap<Value *, unsigned> DepSetId;
521 Value *Ptr = A.getValue();
522 bool IsWrite = Accesses.count(MemAccessInfo(Ptr, true));
523 MemAccessInfo Access(Ptr, IsWrite);
530 if (hasComputableBounds(SE, StridesMap, Ptr) &&
531 // When we run after a failing dependency check we have to make sure
532 // we don't have wrapping pointers.
533 (!ShouldCheckStride ||
534 isStridedPtr(SE, Ptr, TheLoop, StridesMap) == 1)) {
535 // The id of the dependence set.
538 if (IsDepCheckNeeded) {
539 Value *Leader = DepCands.getLeaderValue(Access).getPointer();
540 unsigned &LeaderId = DepSetId[Leader];
542 LeaderId = RunningDepId++;
545 // Each access has its own dependence set.
546 DepId = RunningDepId++;
548 RtCheck.insert(TheLoop, Ptr, IsWrite, DepId, ASId, StridesMap);
550 DEBUG(dbgs() << "LAA: Found a runtime check ptr:" << *Ptr << '\n');
552 DEBUG(dbgs() << "LAA: Can't find bounds for ptr:" << *Ptr << '\n');
557 // If we have at least two writes or one write and a read then we need to
558 // check them. But there is no need to checks if there is only one
559 // dependence set for this alias set.
561 // Note that this function computes CanDoRT and NeedRTCheck independently.
562 // For example CanDoRT=false, NeedRTCheck=false means that we have a pointer
563 // for which we couldn't find the bounds but we don't actually need to emit
564 // any checks so it does not matter.
565 if (!(IsDepCheckNeeded && CanDoRT && RunningDepId == 2))
566 NeedRTCheck |= (NumWritePtrChecks >= 2 || (NumReadPtrChecks >= 1 &&
567 NumWritePtrChecks >= 1));
572 // If the pointers that we would use for the bounds comparison have different
573 // address spaces, assume the values aren't directly comparable, so we can't
574 // use them for the runtime check. We also have to assume they could
575 // overlap. In the future there should be metadata for whether address spaces
577 unsigned NumPointers = RtCheck.Pointers.size();
578 for (unsigned i = 0; i < NumPointers; ++i) {
579 for (unsigned j = i + 1; j < NumPointers; ++j) {
580 // Only need to check pointers between two different dependency sets.
581 if (RtCheck.Pointers[i].DependencySetId ==
582 RtCheck.Pointers[j].DependencySetId)
584 // Only need to check pointers in the same alias set.
585 if (RtCheck.Pointers[i].AliasSetId != RtCheck.Pointers[j].AliasSetId)
588 Value *PtrI = RtCheck.Pointers[i].PointerValue;
589 Value *PtrJ = RtCheck.Pointers[j].PointerValue;
591 unsigned ASi = PtrI->getType()->getPointerAddressSpace();
592 unsigned ASj = PtrJ->getType()->getPointerAddressSpace();
594 DEBUG(dbgs() << "LAA: Runtime check would require comparison between"
595 " different address spaces\n");
601 if (NeedRTCheck && CanDoRT)
602 RtCheck.groupChecks(DepCands, IsDepCheckNeeded);
604 DEBUG(dbgs() << "LAA: We need to do " << RtCheck.getNumberOfChecks(nullptr)
605 << " pointer comparisons.\n");
607 RtCheck.Need = NeedRTCheck;
609 bool CanDoRTIfNeeded = !NeedRTCheck || CanDoRT;
610 if (!CanDoRTIfNeeded)
612 return CanDoRTIfNeeded;
615 void AccessAnalysis::processMemAccesses() {
616 // We process the set twice: first we process read-write pointers, last we
617 // process read-only pointers. This allows us to skip dependence tests for
618 // read-only pointers.
620 DEBUG(dbgs() << "LAA: Processing memory accesses...\n");
621 DEBUG(dbgs() << " AST: "; AST.dump());
622 DEBUG(dbgs() << "LAA: Accesses(" << Accesses.size() << "):\n");
624 for (auto A : Accesses)
625 dbgs() << "\t" << *A.getPointer() << " (" <<
626 (A.getInt() ? "write" : (ReadOnlyPtr.count(A.getPointer()) ?
627 "read-only" : "read")) << ")\n";
630 // The AliasSetTracker has nicely partitioned our pointers by metadata
631 // compatibility and potential for underlying-object overlap. As a result, we
632 // only need to check for potential pointer dependencies within each alias
634 for (auto &AS : AST) {
635 // Note that both the alias-set tracker and the alias sets themselves used
636 // linked lists internally and so the iteration order here is deterministic
637 // (matching the original instruction order within each set).
639 bool SetHasWrite = false;
641 // Map of pointers to last access encountered.
642 typedef DenseMap<Value*, MemAccessInfo> UnderlyingObjToAccessMap;
643 UnderlyingObjToAccessMap ObjToLastAccess;
645 // Set of access to check after all writes have been processed.
646 PtrAccessSet DeferredAccesses;
648 // Iterate over each alias set twice, once to process read/write pointers,
649 // and then to process read-only pointers.
650 for (int SetIteration = 0; SetIteration < 2; ++SetIteration) {
651 bool UseDeferred = SetIteration > 0;
652 PtrAccessSet &S = UseDeferred ? DeferredAccesses : Accesses;
655 Value *Ptr = AV.getValue();
657 // For a single memory access in AliasSetTracker, Accesses may contain
658 // both read and write, and they both need to be handled for CheckDeps.
660 if (AC.getPointer() != Ptr)
663 bool IsWrite = AC.getInt();
665 // If we're using the deferred access set, then it contains only
667 bool IsReadOnlyPtr = ReadOnlyPtr.count(Ptr) && !IsWrite;
668 if (UseDeferred && !IsReadOnlyPtr)
670 // Otherwise, the pointer must be in the PtrAccessSet, either as a
672 assert(((IsReadOnlyPtr && UseDeferred) || IsWrite ||
673 S.count(MemAccessInfo(Ptr, false))) &&
674 "Alias-set pointer not in the access set?");
676 MemAccessInfo Access(Ptr, IsWrite);
677 DepCands.insert(Access);
679 // Memorize read-only pointers for later processing and skip them in
680 // the first round (they need to be checked after we have seen all
681 // write pointers). Note: we also mark pointer that are not
682 // consecutive as "read-only" pointers (so that we check
683 // "a[b[i]] +="). Hence, we need the second check for "!IsWrite".
684 if (!UseDeferred && IsReadOnlyPtr) {
685 DeferredAccesses.insert(Access);
689 // If this is a write - check other reads and writes for conflicts. If
690 // this is a read only check other writes for conflicts (but only if
691 // there is no other write to the ptr - this is an optimization to
692 // catch "a[i] = a[i] + " without having to do a dependence check).
693 if ((IsWrite || IsReadOnlyPtr) && SetHasWrite) {
694 CheckDeps.insert(Access);
695 IsRTCheckAnalysisNeeded = true;
701 // Create sets of pointers connected by a shared alias set and
702 // underlying object.
703 typedef SmallVector<Value *, 16> ValueVector;
704 ValueVector TempObjects;
706 GetUnderlyingObjects(Ptr, TempObjects, DL, LI);
707 DEBUG(dbgs() << "Underlying objects for pointer " << *Ptr << "\n");
708 for (Value *UnderlyingObj : TempObjects) {
709 UnderlyingObjToAccessMap::iterator Prev =
710 ObjToLastAccess.find(UnderlyingObj);
711 if (Prev != ObjToLastAccess.end())
712 DepCands.unionSets(Access, Prev->second);
714 ObjToLastAccess[UnderlyingObj] = Access;
715 DEBUG(dbgs() << " " << *UnderlyingObj << "\n");
723 static bool isInBoundsGep(Value *Ptr) {
724 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr))
725 return GEP->isInBounds();
729 /// \brief Return true if an AddRec pointer \p Ptr is unsigned non-wrapping,
730 /// i.e. monotonically increasing/decreasing.
731 static bool isNoWrapAddRec(Value *Ptr, const SCEVAddRecExpr *AR,
732 ScalarEvolution *SE, const Loop *L) {
733 // FIXME: This should probably only return true for NUW.
734 if (AR->getNoWrapFlags(SCEV::NoWrapMask))
737 // Scalar evolution does not propagate the non-wrapping flags to values that
738 // are derived from a non-wrapping induction variable because non-wrapping
739 // could be flow-sensitive.
741 // Look through the potentially overflowing instruction to try to prove
742 // non-wrapping for the *specific* value of Ptr.
744 // The arithmetic implied by an inbounds GEP can't overflow.
745 auto *GEP = dyn_cast<GetElementPtrInst>(Ptr);
746 if (!GEP || !GEP->isInBounds())
749 // Make sure there is only one non-const index and analyze that.
750 Value *NonConstIndex = nullptr;
751 for (auto Index = GEP->idx_begin(); Index != GEP->idx_end(); ++Index)
752 if (!isa<ConstantInt>(*Index)) {
755 NonConstIndex = *Index;
758 // The recurrence is on the pointer, ignore for now.
761 // The index in GEP is signed. It is non-wrapping if it's derived from a NSW
762 // AddRec using a NSW operation.
763 if (auto *OBO = dyn_cast<OverflowingBinaryOperator>(NonConstIndex))
764 if (OBO->hasNoSignedWrap() &&
765 // Assume constant for other the operand so that the AddRec can be
767 isa<ConstantInt>(OBO->getOperand(1))) {
768 auto *OpScev = SE->getSCEV(OBO->getOperand(0));
770 if (auto *OpAR = dyn_cast<SCEVAddRecExpr>(OpScev))
771 return OpAR->getLoop() == L && OpAR->getNoWrapFlags(SCEV::FlagNSW);
777 /// \brief Check whether the access through \p Ptr has a constant stride.
778 int llvm::isStridedPtr(ScalarEvolution *SE, Value *Ptr, const Loop *Lp,
779 const ValueToValueMap &StridesMap) {
780 const Type *Ty = Ptr->getType();
781 assert(Ty->isPointerTy() && "Unexpected non-ptr");
783 // Make sure that the pointer does not point to aggregate types.
784 const PointerType *PtrTy = cast<PointerType>(Ty);
785 if (PtrTy->getElementType()->isAggregateType()) {
786 DEBUG(dbgs() << "LAA: Bad stride - Not a pointer to a scalar type"
791 const SCEV *PtrScev = replaceSymbolicStrideSCEV(SE, StridesMap, Ptr);
793 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
795 DEBUG(dbgs() << "LAA: Bad stride - Not an AddRecExpr pointer "
796 << *Ptr << " SCEV: " << *PtrScev << "\n");
800 // The accesss function must stride over the innermost loop.
801 if (Lp != AR->getLoop()) {
802 DEBUG(dbgs() << "LAA: Bad stride - Not striding over innermost loop " <<
803 *Ptr << " SCEV: " << *PtrScev << "\n");
806 // The address calculation must not wrap. Otherwise, a dependence could be
808 // An inbounds getelementptr that is a AddRec with a unit stride
809 // cannot wrap per definition. The unit stride requirement is checked later.
810 // An getelementptr without an inbounds attribute and unit stride would have
811 // to access the pointer value "0" which is undefined behavior in address
812 // space 0, therefore we can also vectorize this case.
813 bool IsInBoundsGEP = isInBoundsGep(Ptr);
814 bool IsNoWrapAddRec = isNoWrapAddRec(Ptr, AR, SE, Lp);
815 bool IsInAddressSpaceZero = PtrTy->getAddressSpace() == 0;
816 if (!IsNoWrapAddRec && !IsInBoundsGEP && !IsInAddressSpaceZero) {
817 DEBUG(dbgs() << "LAA: Bad stride - Pointer may wrap in the address space "
818 << *Ptr << " SCEV: " << *PtrScev << "\n");
822 // Check the step is constant.
823 const SCEV *Step = AR->getStepRecurrence(*SE);
825 // Calculate the pointer stride and check if it is constant.
826 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
828 DEBUG(dbgs() << "LAA: Bad stride - Not a constant strided " << *Ptr <<
829 " SCEV: " << *PtrScev << "\n");
833 auto &DL = Lp->getHeader()->getModule()->getDataLayout();
834 int64_t Size = DL.getTypeAllocSize(PtrTy->getElementType());
835 const APInt &APStepVal = C->getValue()->getValue();
837 // Huge step value - give up.
838 if (APStepVal.getBitWidth() > 64)
841 int64_t StepVal = APStepVal.getSExtValue();
844 int64_t Stride = StepVal / Size;
845 int64_t Rem = StepVal % Size;
849 // If the SCEV could wrap but we have an inbounds gep with a unit stride we
850 // know we can't "wrap around the address space". In case of address space
851 // zero we know that this won't happen without triggering undefined behavior.
852 if (!IsNoWrapAddRec && (IsInBoundsGEP || IsInAddressSpaceZero) &&
853 Stride != 1 && Stride != -1)
859 bool MemoryDepChecker::Dependence::isSafeForVectorization(DepType Type) {
863 case BackwardVectorizable:
867 case ForwardButPreventsForwarding:
869 case BackwardVectorizableButPreventsForwarding:
872 llvm_unreachable("unexpected DepType!");
875 bool MemoryDepChecker::Dependence::isInterestingDependence(DepType Type) {
881 case BackwardVectorizable:
883 case ForwardButPreventsForwarding:
885 case BackwardVectorizableButPreventsForwarding:
888 llvm_unreachable("unexpected DepType!");
891 bool MemoryDepChecker::Dependence::isPossiblyBackward() const {
895 case ForwardButPreventsForwarding:
899 case BackwardVectorizable:
901 case BackwardVectorizableButPreventsForwarding:
904 llvm_unreachable("unexpected DepType!");
907 bool MemoryDepChecker::couldPreventStoreLoadForward(unsigned Distance,
908 unsigned TypeByteSize) {
909 // If loads occur at a distance that is not a multiple of a feasible vector
910 // factor store-load forwarding does not take place.
911 // Positive dependences might cause troubles because vectorizing them might
912 // prevent store-load forwarding making vectorized code run a lot slower.
913 // a[i] = a[i-3] ^ a[i-8];
914 // The stores to a[i:i+1] don't align with the stores to a[i-3:i-2] and
915 // hence on your typical architecture store-load forwarding does not take
916 // place. Vectorizing in such cases does not make sense.
917 // Store-load forwarding distance.
918 const unsigned NumCyclesForStoreLoadThroughMemory = 8*TypeByteSize;
919 // Maximum vector factor.
920 unsigned MaxVFWithoutSLForwardIssues =
921 VectorizerParams::MaxVectorWidth * TypeByteSize;
922 if(MaxSafeDepDistBytes < MaxVFWithoutSLForwardIssues)
923 MaxVFWithoutSLForwardIssues = MaxSafeDepDistBytes;
925 for (unsigned vf = 2*TypeByteSize; vf <= MaxVFWithoutSLForwardIssues;
927 if (Distance % vf && Distance / vf < NumCyclesForStoreLoadThroughMemory) {
928 MaxVFWithoutSLForwardIssues = (vf >>=1);
933 if (MaxVFWithoutSLForwardIssues< 2*TypeByteSize) {
934 DEBUG(dbgs() << "LAA: Distance " << Distance <<
935 " that could cause a store-load forwarding conflict\n");
939 if (MaxVFWithoutSLForwardIssues < MaxSafeDepDistBytes &&
940 MaxVFWithoutSLForwardIssues !=
941 VectorizerParams::MaxVectorWidth * TypeByteSize)
942 MaxSafeDepDistBytes = MaxVFWithoutSLForwardIssues;
946 /// \brief Check the dependence for two accesses with the same stride \p Stride.
947 /// \p Distance is the positive distance and \p TypeByteSize is type size in
950 /// \returns true if they are independent.
951 static bool areStridedAccessesIndependent(unsigned Distance, unsigned Stride,
952 unsigned TypeByteSize) {
953 assert(Stride > 1 && "The stride must be greater than 1");
954 assert(TypeByteSize > 0 && "The type size in byte must be non-zero");
955 assert(Distance > 0 && "The distance must be non-zero");
957 // Skip if the distance is not multiple of type byte size.
958 if (Distance % TypeByteSize)
961 unsigned ScaledDist = Distance / TypeByteSize;
963 // No dependence if the scaled distance is not multiple of the stride.
965 // for (i = 0; i < 1024 ; i += 4)
966 // A[i+2] = A[i] + 1;
968 // Two accesses in memory (scaled distance is 2, stride is 4):
969 // | A[0] | | | | A[4] | | | |
970 // | | | A[2] | | | | A[6] | |
973 // for (i = 0; i < 1024 ; i += 3)
974 // A[i+4] = A[i] + 1;
976 // Two accesses in memory (scaled distance is 4, stride is 3):
977 // | A[0] | | | A[3] | | | A[6] | | |
978 // | | | | | A[4] | | | A[7] | |
979 return ScaledDist % Stride;
982 MemoryDepChecker::Dependence::DepType
983 MemoryDepChecker::isDependent(const MemAccessInfo &A, unsigned AIdx,
984 const MemAccessInfo &B, unsigned BIdx,
985 const ValueToValueMap &Strides) {
986 assert (AIdx < BIdx && "Must pass arguments in program order");
988 Value *APtr = A.getPointer();
989 Value *BPtr = B.getPointer();
990 bool AIsWrite = A.getInt();
991 bool BIsWrite = B.getInt();
993 // Two reads are independent.
994 if (!AIsWrite && !BIsWrite)
995 return Dependence::NoDep;
997 // We cannot check pointers in different address spaces.
998 if (APtr->getType()->getPointerAddressSpace() !=
999 BPtr->getType()->getPointerAddressSpace())
1000 return Dependence::Unknown;
1002 const SCEV *AScev = replaceSymbolicStrideSCEV(SE, Strides, APtr);
1003 const SCEV *BScev = replaceSymbolicStrideSCEV(SE, Strides, BPtr);
1005 int StrideAPtr = isStridedPtr(SE, APtr, InnermostLoop, Strides);
1006 int StrideBPtr = isStridedPtr(SE, BPtr, InnermostLoop, Strides);
1008 const SCEV *Src = AScev;
1009 const SCEV *Sink = BScev;
1011 // If the induction step is negative we have to invert source and sink of the
1013 if (StrideAPtr < 0) {
1016 std::swap(APtr, BPtr);
1017 std::swap(Src, Sink);
1018 std::swap(AIsWrite, BIsWrite);
1019 std::swap(AIdx, BIdx);
1020 std::swap(StrideAPtr, StrideBPtr);
1023 const SCEV *Dist = SE->getMinusSCEV(Sink, Src);
1025 DEBUG(dbgs() << "LAA: Src Scev: " << *Src << "Sink Scev: " << *Sink
1026 << "(Induction step: " << StrideAPtr << ")\n");
1027 DEBUG(dbgs() << "LAA: Distance for " << *InstMap[AIdx] << " to "
1028 << *InstMap[BIdx] << ": " << *Dist << "\n");
1030 // Need accesses with constant stride. We don't want to vectorize
1031 // "A[B[i]] += ..." and similar code or pointer arithmetic that could wrap in
1032 // the address space.
1033 if (!StrideAPtr || !StrideBPtr || StrideAPtr != StrideBPtr){
1034 DEBUG(dbgs() << "Pointer access with non-constant stride\n");
1035 return Dependence::Unknown;
1038 const SCEVConstant *C = dyn_cast<SCEVConstant>(Dist);
1040 DEBUG(dbgs() << "LAA: Dependence because of non-constant distance\n");
1041 ShouldRetryWithRuntimeCheck = true;
1042 return Dependence::Unknown;
1045 Type *ATy = APtr->getType()->getPointerElementType();
1046 Type *BTy = BPtr->getType()->getPointerElementType();
1047 auto &DL = InnermostLoop->getHeader()->getModule()->getDataLayout();
1048 unsigned TypeByteSize = DL.getTypeAllocSize(ATy);
1050 // Negative distances are not plausible dependencies.
1051 const APInt &Val = C->getValue()->getValue();
1052 if (Val.isNegative()) {
1053 bool IsTrueDataDependence = (AIsWrite && !BIsWrite);
1054 if (IsTrueDataDependence &&
1055 (couldPreventStoreLoadForward(Val.abs().getZExtValue(), TypeByteSize) ||
1057 return Dependence::ForwardButPreventsForwarding;
1059 DEBUG(dbgs() << "LAA: Dependence is negative: NoDep\n");
1060 return Dependence::Forward;
1063 // Write to the same location with the same size.
1064 // Could be improved to assert type sizes are the same (i32 == float, etc).
1067 return Dependence::NoDep;
1068 DEBUG(dbgs() << "LAA: Zero dependence difference but different types\n");
1069 return Dependence::Unknown;
1072 assert(Val.isStrictlyPositive() && "Expect a positive value");
1076 "LAA: ReadWrite-Write positive dependency with different types\n");
1077 return Dependence::Unknown;
1080 unsigned Distance = (unsigned) Val.getZExtValue();
1082 unsigned Stride = std::abs(StrideAPtr);
1084 areStridedAccessesIndependent(Distance, Stride, TypeByteSize)) {
1085 DEBUG(dbgs() << "LAA: Strided accesses are independent\n");
1086 return Dependence::NoDep;
1089 // Bail out early if passed-in parameters make vectorization not feasible.
1090 unsigned ForcedFactor = (VectorizerParams::VectorizationFactor ?
1091 VectorizerParams::VectorizationFactor : 1);
1092 unsigned ForcedUnroll = (VectorizerParams::VectorizationInterleave ?
1093 VectorizerParams::VectorizationInterleave : 1);
1094 // The minimum number of iterations for a vectorized/unrolled version.
1095 unsigned MinNumIter = std::max(ForcedFactor * ForcedUnroll, 2U);
1097 // It's not vectorizable if the distance is smaller than the minimum distance
1098 // needed for a vectroized/unrolled version. Vectorizing one iteration in
1099 // front needs TypeByteSize * Stride. Vectorizing the last iteration needs
1100 // TypeByteSize (No need to plus the last gap distance).
1102 // E.g. Assume one char is 1 byte in memory and one int is 4 bytes.
1104 // int *B = (int *)((char *)A + 14);
1105 // for (i = 0 ; i < 1024 ; i += 2)
1109 // Two accesses in memory (stride is 2):
1110 // | A[0] | | A[2] | | A[4] | | A[6] | |
1111 // | B[0] | | B[2] | | B[4] |
1113 // Distance needs for vectorizing iterations except the last iteration:
1114 // 4 * 2 * (MinNumIter - 1). Distance needs for the last iteration: 4.
1115 // So the minimum distance needed is: 4 * 2 * (MinNumIter - 1) + 4.
1117 // If MinNumIter is 2, it is vectorizable as the minimum distance needed is
1118 // 12, which is less than distance.
1120 // If MinNumIter is 4 (Say if a user forces the vectorization factor to be 4),
1121 // the minimum distance needed is 28, which is greater than distance. It is
1122 // not safe to do vectorization.
1123 unsigned MinDistanceNeeded =
1124 TypeByteSize * Stride * (MinNumIter - 1) + TypeByteSize;
1125 if (MinDistanceNeeded > Distance) {
1126 DEBUG(dbgs() << "LAA: Failure because of positive distance " << Distance
1128 return Dependence::Backward;
1131 // Unsafe if the minimum distance needed is greater than max safe distance.
1132 if (MinDistanceNeeded > MaxSafeDepDistBytes) {
1133 DEBUG(dbgs() << "LAA: Failure because it needs at least "
1134 << MinDistanceNeeded << " size in bytes");
1135 return Dependence::Backward;
1138 // Positive distance bigger than max vectorization factor.
1139 // FIXME: Should use max factor instead of max distance in bytes, which could
1140 // not handle different types.
1141 // E.g. Assume one char is 1 byte in memory and one int is 4 bytes.
1142 // void foo (int *A, char *B) {
1143 // for (unsigned i = 0; i < 1024; i++) {
1144 // A[i+2] = A[i] + 1;
1145 // B[i+2] = B[i] + 1;
1149 // This case is currently unsafe according to the max safe distance. If we
1150 // analyze the two accesses on array B, the max safe dependence distance
1151 // is 2. Then we analyze the accesses on array A, the minimum distance needed
1152 // is 8, which is less than 2 and forbidden vectorization, But actually
1153 // both A and B could be vectorized by 2 iterations.
1154 MaxSafeDepDistBytes =
1155 Distance < MaxSafeDepDistBytes ? Distance : MaxSafeDepDistBytes;
1157 bool IsTrueDataDependence = (!AIsWrite && BIsWrite);
1158 if (IsTrueDataDependence &&
1159 couldPreventStoreLoadForward(Distance, TypeByteSize))
1160 return Dependence::BackwardVectorizableButPreventsForwarding;
1162 DEBUG(dbgs() << "LAA: Positive distance " << Val.getSExtValue()
1163 << " with max VF = "
1164 << MaxSafeDepDistBytes / (TypeByteSize * Stride) << '\n');
1166 return Dependence::BackwardVectorizable;
1169 bool MemoryDepChecker::areDepsSafe(DepCandidates &AccessSets,
1170 MemAccessInfoSet &CheckDeps,
1171 const ValueToValueMap &Strides) {
1173 MaxSafeDepDistBytes = -1U;
1174 while (!CheckDeps.empty()) {
1175 MemAccessInfo CurAccess = *CheckDeps.begin();
1177 // Get the relevant memory access set.
1178 EquivalenceClasses<MemAccessInfo>::iterator I =
1179 AccessSets.findValue(AccessSets.getLeaderValue(CurAccess));
1181 // Check accesses within this set.
1182 EquivalenceClasses<MemAccessInfo>::member_iterator AI, AE;
1183 AI = AccessSets.member_begin(I), AE = AccessSets.member_end();
1185 // Check every access pair.
1187 CheckDeps.erase(*AI);
1188 EquivalenceClasses<MemAccessInfo>::member_iterator OI = std::next(AI);
1190 // Check every accessing instruction pair in program order.
1191 for (std::vector<unsigned>::iterator I1 = Accesses[*AI].begin(),
1192 I1E = Accesses[*AI].end(); I1 != I1E; ++I1)
1193 for (std::vector<unsigned>::iterator I2 = Accesses[*OI].begin(),
1194 I2E = Accesses[*OI].end(); I2 != I2E; ++I2) {
1195 auto A = std::make_pair(&*AI, *I1);
1196 auto B = std::make_pair(&*OI, *I2);
1202 Dependence::DepType Type =
1203 isDependent(*A.first, A.second, *B.first, B.second, Strides);
1204 SafeForVectorization &= Dependence::isSafeForVectorization(Type);
1206 // Gather dependences unless we accumulated MaxInterestingDependence
1207 // dependences. In that case return as soon as we find the first
1208 // unsafe dependence. This puts a limit on this quadratic
1210 if (RecordInterestingDependences) {
1211 if (Dependence::isInterestingDependence(Type))
1212 InterestingDependences.push_back(
1213 Dependence(A.second, B.second, Type));
1215 if (InterestingDependences.size() >= MaxInterestingDependence) {
1216 RecordInterestingDependences = false;
1217 InterestingDependences.clear();
1218 DEBUG(dbgs() << "Too many dependences, stopped recording\n");
1221 if (!RecordInterestingDependences && !SafeForVectorization)
1230 DEBUG(dbgs() << "Total Interesting Dependences: "
1231 << InterestingDependences.size() << "\n");
1232 return SafeForVectorization;
1235 SmallVector<Instruction *, 4>
1236 MemoryDepChecker::getInstructionsForAccess(Value *Ptr, bool isWrite) const {
1237 MemAccessInfo Access(Ptr, isWrite);
1238 auto &IndexVector = Accesses.find(Access)->second;
1240 SmallVector<Instruction *, 4> Insts;
1241 std::transform(IndexVector.begin(), IndexVector.end(),
1242 std::back_inserter(Insts),
1243 [&](unsigned Idx) { return this->InstMap[Idx]; });
1247 const char *MemoryDepChecker::Dependence::DepName[] = {
1248 "NoDep", "Unknown", "Forward", "ForwardButPreventsForwarding", "Backward",
1249 "BackwardVectorizable", "BackwardVectorizableButPreventsForwarding"};
1251 void MemoryDepChecker::Dependence::print(
1252 raw_ostream &OS, unsigned Depth,
1253 const SmallVectorImpl<Instruction *> &Instrs) const {
1254 OS.indent(Depth) << DepName[Type] << ":\n";
1255 OS.indent(Depth + 2) << *Instrs[Source] << " -> \n";
1256 OS.indent(Depth + 2) << *Instrs[Destination] << "\n";
1259 bool LoopAccessInfo::canAnalyzeLoop() {
1260 // We need to have a loop header.
1261 DEBUG(dbgs() << "LAA: Found a loop: " <<
1262 TheLoop->getHeader()->getName() << '\n');
1264 // We can only analyze innermost loops.
1265 if (!TheLoop->empty()) {
1266 DEBUG(dbgs() << "LAA: loop is not the innermost loop\n");
1267 emitAnalysis(LoopAccessReport() << "loop is not the innermost loop");
1271 // We must have a single backedge.
1272 if (TheLoop->getNumBackEdges() != 1) {
1273 DEBUG(dbgs() << "LAA: loop control flow is not understood by analyzer\n");
1275 LoopAccessReport() <<
1276 "loop control flow is not understood by analyzer");
1280 // We must have a single exiting block.
1281 if (!TheLoop->getExitingBlock()) {
1282 DEBUG(dbgs() << "LAA: loop control flow is not understood by analyzer\n");
1284 LoopAccessReport() <<
1285 "loop control flow is not understood by analyzer");
1289 // We only handle bottom-tested loops, i.e. loop in which the condition is
1290 // checked at the end of each iteration. With that we can assume that all
1291 // instructions in the loop are executed the same number of times.
1292 if (TheLoop->getExitingBlock() != TheLoop->getLoopLatch()) {
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 // ScalarEvolution needs to be able to find the exit count.
1301 const SCEV *ExitCount = SE->getBackedgeTakenCount(TheLoop);
1302 if (ExitCount == SE->getCouldNotCompute()) {
1303 emitAnalysis(LoopAccessReport() <<
1304 "could not determine number of loop iterations");
1305 DEBUG(dbgs() << "LAA: SCEV could not compute the loop exit count.\n");
1312 void LoopAccessInfo::analyzeLoop(const ValueToValueMap &Strides) {
1314 typedef SmallVector<Value*, 16> ValueVector;
1315 typedef SmallPtrSet<Value*, 16> ValueSet;
1317 // Holds the Load and Store *instructions*.
1321 // Holds all the different accesses in the loop.
1322 unsigned NumReads = 0;
1323 unsigned NumReadWrites = 0;
1325 PtrRtChecking.Pointers.clear();
1326 PtrRtChecking.Need = false;
1328 const bool IsAnnotatedParallel = TheLoop->isAnnotatedParallel();
1331 for (Loop::block_iterator bb = TheLoop->block_begin(),
1332 be = TheLoop->block_end(); bb != be; ++bb) {
1334 // Scan the BB and collect legal loads and stores.
1335 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
1338 // If this is a load, save it. If this instruction can read from memory
1339 // but is not a load, then we quit. Notice that we don't handle function
1340 // calls that read or write.
1341 if (it->mayReadFromMemory()) {
1342 // Many math library functions read the rounding mode. We will only
1343 // vectorize a loop if it contains known function calls that don't set
1344 // the flag. Therefore, it is safe to ignore this read from memory.
1345 CallInst *Call = dyn_cast<CallInst>(it);
1346 if (Call && getIntrinsicIDForCall(Call, TLI))
1349 // If the function has an explicit vectorized counterpart, we can safely
1350 // assume that it can be vectorized.
1351 if (Call && !Call->isNoBuiltin() && Call->getCalledFunction() &&
1352 TLI->isFunctionVectorizable(Call->getCalledFunction()->getName()))
1355 LoadInst *Ld = dyn_cast<LoadInst>(it);
1356 if (!Ld || (!Ld->isSimple() && !IsAnnotatedParallel)) {
1357 emitAnalysis(LoopAccessReport(Ld)
1358 << "read with atomic ordering or volatile read");
1359 DEBUG(dbgs() << "LAA: Found a non-simple load.\n");
1364 Loads.push_back(Ld);
1365 DepChecker.addAccess(Ld);
1369 // Save 'store' instructions. Abort if other instructions write to memory.
1370 if (it->mayWriteToMemory()) {
1371 StoreInst *St = dyn_cast<StoreInst>(it);
1373 emitAnalysis(LoopAccessReport(it) <<
1374 "instruction cannot be vectorized");
1378 if (!St->isSimple() && !IsAnnotatedParallel) {
1379 emitAnalysis(LoopAccessReport(St)
1380 << "write with atomic ordering or volatile write");
1381 DEBUG(dbgs() << "LAA: Found a non-simple store.\n");
1386 Stores.push_back(St);
1387 DepChecker.addAccess(St);
1392 // Now we have two lists that hold the loads and the stores.
1393 // Next, we find the pointers that they use.
1395 // Check if we see any stores. If there are no stores, then we don't
1396 // care if the pointers are *restrict*.
1397 if (!Stores.size()) {
1398 DEBUG(dbgs() << "LAA: Found a read-only loop!\n");
1403 MemoryDepChecker::DepCandidates DependentAccesses;
1404 AccessAnalysis Accesses(TheLoop->getHeader()->getModule()->getDataLayout(),
1405 AA, LI, DependentAccesses);
1407 // Holds the analyzed pointers. We don't want to call GetUnderlyingObjects
1408 // multiple times on the same object. If the ptr is accessed twice, once
1409 // for read and once for write, it will only appear once (on the write
1410 // list). This is okay, since we are going to check for conflicts between
1411 // writes and between reads and writes, but not between reads and reads.
1414 ValueVector::iterator I, IE;
1415 for (I = Stores.begin(), IE = Stores.end(); I != IE; ++I) {
1416 StoreInst *ST = cast<StoreInst>(*I);
1417 Value* Ptr = ST->getPointerOperand();
1418 // Check for store to loop invariant address.
1419 StoreToLoopInvariantAddress |= isUniform(Ptr);
1420 // If we did *not* see this pointer before, insert it to the read-write
1421 // list. At this phase it is only a 'write' list.
1422 if (Seen.insert(Ptr).second) {
1425 MemoryLocation Loc = MemoryLocation::get(ST);
1426 // The TBAA metadata could have a control dependency on the predication
1427 // condition, so we cannot rely on it when determining whether or not we
1428 // need runtime pointer checks.
1429 if (blockNeedsPredication(ST->getParent(), TheLoop, DT))
1430 Loc.AATags.TBAA = nullptr;
1432 Accesses.addStore(Loc);
1436 if (IsAnnotatedParallel) {
1438 << "LAA: A loop annotated parallel, ignore memory dependency "
1444 for (I = Loads.begin(), IE = Loads.end(); I != IE; ++I) {
1445 LoadInst *LD = cast<LoadInst>(*I);
1446 Value* Ptr = LD->getPointerOperand();
1447 // If we did *not* see this pointer before, insert it to the
1448 // read list. If we *did* see it before, then it is already in
1449 // the read-write list. This allows us to vectorize expressions
1450 // such as A[i] += x; Because the address of A[i] is a read-write
1451 // pointer. This only works if the index of A[i] is consecutive.
1452 // If the address of i is unknown (for example A[B[i]]) then we may
1453 // read a few words, modify, and write a few words, and some of the
1454 // words may be written to the same address.
1455 bool IsReadOnlyPtr = false;
1456 if (Seen.insert(Ptr).second || !isStridedPtr(SE, Ptr, TheLoop, Strides)) {
1458 IsReadOnlyPtr = true;
1461 MemoryLocation Loc = MemoryLocation::get(LD);
1462 // The TBAA metadata could have a control dependency on the predication
1463 // condition, so we cannot rely on it when determining whether or not we
1464 // need runtime pointer checks.
1465 if (blockNeedsPredication(LD->getParent(), TheLoop, DT))
1466 Loc.AATags.TBAA = nullptr;
1468 Accesses.addLoad(Loc, IsReadOnlyPtr);
1471 // If we write (or read-write) to a single destination and there are no
1472 // other reads in this loop then is it safe to vectorize.
1473 if (NumReadWrites == 1 && NumReads == 0) {
1474 DEBUG(dbgs() << "LAA: Found a write-only loop!\n");
1479 // Build dependence sets and check whether we need a runtime pointer bounds
1481 Accesses.buildDependenceSets();
1483 // Find pointers with computable bounds. We are going to use this information
1484 // to place a runtime bound check.
1485 bool CanDoRTIfNeeded =
1486 Accesses.canCheckPtrAtRT(PtrRtChecking, SE, TheLoop, Strides);
1487 if (!CanDoRTIfNeeded) {
1488 emitAnalysis(LoopAccessReport() << "cannot identify array bounds");
1489 DEBUG(dbgs() << "LAA: We can't vectorize because we can't find "
1490 << "the array bounds.\n");
1495 DEBUG(dbgs() << "LAA: We can perform a memory runtime check if needed.\n");
1498 if (Accesses.isDependencyCheckNeeded()) {
1499 DEBUG(dbgs() << "LAA: Checking memory dependencies\n");
1500 CanVecMem = DepChecker.areDepsSafe(
1501 DependentAccesses, Accesses.getDependenciesToCheck(), Strides);
1502 MaxSafeDepDistBytes = DepChecker.getMaxSafeDepDistBytes();
1504 if (!CanVecMem && DepChecker.shouldRetryWithRuntimeCheck()) {
1505 DEBUG(dbgs() << "LAA: Retrying with memory checks\n");
1507 // Clear the dependency checks. We assume they are not needed.
1508 Accesses.resetDepChecks(DepChecker);
1510 PtrRtChecking.reset();
1511 PtrRtChecking.Need = true;
1514 Accesses.canCheckPtrAtRT(PtrRtChecking, SE, TheLoop, Strides, true);
1516 // Check that we found the bounds for the pointer.
1517 if (!CanDoRTIfNeeded) {
1518 emitAnalysis(LoopAccessReport()
1519 << "cannot check memory dependencies at runtime");
1520 DEBUG(dbgs() << "LAA: Can't vectorize with memory checks\n");
1530 DEBUG(dbgs() << "LAA: No unsafe dependent memory operations in loop. We"
1531 << (PtrRtChecking.Need ? "" : " don't")
1532 << " need runtime memory checks.\n");
1534 emitAnalysis(LoopAccessReport() <<
1535 "unsafe dependent memory operations in loop");
1536 DEBUG(dbgs() << "LAA: unsafe dependent memory operations in loop\n");
1540 bool LoopAccessInfo::blockNeedsPredication(BasicBlock *BB, Loop *TheLoop,
1541 DominatorTree *DT) {
1542 assert(TheLoop->contains(BB) && "Unknown block used");
1544 // Blocks that do not dominate the latch need predication.
1545 BasicBlock* Latch = TheLoop->getLoopLatch();
1546 return !DT->dominates(BB, Latch);
1549 void LoopAccessInfo::emitAnalysis(LoopAccessReport &Message) {
1550 assert(!Report && "Multiple reports generated");
1554 bool LoopAccessInfo::isUniform(Value *V) const {
1555 return (SE->isLoopInvariant(SE->getSCEV(V), TheLoop));
1558 // FIXME: this function is currently a duplicate of the one in
1559 // LoopVectorize.cpp.
1560 static Instruction *getFirstInst(Instruction *FirstInst, Value *V,
1564 if (Instruction *I = dyn_cast<Instruction>(V))
1565 return I->getParent() == Loc->getParent() ? I : nullptr;
1569 std::pair<Instruction *, Instruction *> LoopAccessInfo::addRuntimeCheck(
1570 Instruction *Loc, const SmallVectorImpl<int> *PtrPartition) const {
1571 if (!PtrRtChecking.Need)
1572 return std::make_pair(nullptr, nullptr);
1574 SmallVector<TrackingVH<Value>, 2> Starts;
1575 SmallVector<TrackingVH<Value>, 2> Ends;
1577 LLVMContext &Ctx = Loc->getContext();
1578 SCEVExpander Exp(*SE, DL, "induction");
1579 Instruction *FirstInst = nullptr;
1581 for (unsigned i = 0; i < PtrRtChecking.CheckingGroups.size(); ++i) {
1582 const RuntimePointerChecking::CheckingPtrGroup &CG =
1583 PtrRtChecking.CheckingGroups[i];
1584 Value *Ptr = PtrRtChecking.Pointers[CG.Members[0]].PointerValue;
1585 const SCEV *Sc = SE->getSCEV(Ptr);
1587 if (SE->isLoopInvariant(Sc, TheLoop)) {
1588 DEBUG(dbgs() << "LAA: Adding RT check for a loop invariant ptr:" << *Ptr
1590 Starts.push_back(Ptr);
1591 Ends.push_back(Ptr);
1593 unsigned AS = Ptr->getType()->getPointerAddressSpace();
1595 // Use this type for pointer arithmetic.
1596 Type *PtrArithTy = Type::getInt8PtrTy(Ctx, AS);
1597 Value *Start = nullptr, *End = nullptr;
1599 DEBUG(dbgs() << "LAA: Adding RT check for range:\n");
1600 Start = Exp.expandCodeFor(CG.Low, PtrArithTy, Loc);
1601 End = Exp.expandCodeFor(CG.High, PtrArithTy, Loc);
1602 DEBUG(dbgs() << "Start: " << *CG.Low << " End: " << *CG.High << "\n");
1603 Starts.push_back(Start);
1604 Ends.push_back(End);
1608 IRBuilder<> ChkBuilder(Loc);
1609 // Our instructions might fold to a constant.
1610 Value *MemoryRuntimeCheck = nullptr;
1611 for (unsigned i = 0; i < PtrRtChecking.CheckingGroups.size(); ++i) {
1612 for (unsigned j = i + 1; j < PtrRtChecking.CheckingGroups.size(); ++j) {
1613 const RuntimePointerChecking::CheckingPtrGroup &CGI =
1614 PtrRtChecking.CheckingGroups[i];
1615 const RuntimePointerChecking::CheckingPtrGroup &CGJ =
1616 PtrRtChecking.CheckingGroups[j];
1618 if (!PtrRtChecking.needsChecking(CGI, CGJ, PtrPartition))
1621 unsigned AS0 = Starts[i]->getType()->getPointerAddressSpace();
1622 unsigned AS1 = Starts[j]->getType()->getPointerAddressSpace();
1624 assert((AS0 == Ends[j]->getType()->getPointerAddressSpace()) &&
1625 (AS1 == Ends[i]->getType()->getPointerAddressSpace()) &&
1626 "Trying to bounds check pointers with different address spaces");
1628 Type *PtrArithTy0 = Type::getInt8PtrTy(Ctx, AS0);
1629 Type *PtrArithTy1 = Type::getInt8PtrTy(Ctx, AS1);
1631 Value *Start0 = ChkBuilder.CreateBitCast(Starts[i], PtrArithTy0, "bc");
1632 Value *Start1 = ChkBuilder.CreateBitCast(Starts[j], PtrArithTy1, "bc");
1633 Value *End0 = ChkBuilder.CreateBitCast(Ends[i], PtrArithTy1, "bc");
1634 Value *End1 = ChkBuilder.CreateBitCast(Ends[j], PtrArithTy0, "bc");
1636 Value *Cmp0 = ChkBuilder.CreateICmpULE(Start0, End1, "bound0");
1637 FirstInst = getFirstInst(FirstInst, Cmp0, Loc);
1638 Value *Cmp1 = ChkBuilder.CreateICmpULE(Start1, End0, "bound1");
1639 FirstInst = getFirstInst(FirstInst, Cmp1, Loc);
1640 Value *IsConflict = ChkBuilder.CreateAnd(Cmp0, Cmp1, "found.conflict");
1641 FirstInst = getFirstInst(FirstInst, IsConflict, Loc);
1642 if (MemoryRuntimeCheck) {
1643 IsConflict = ChkBuilder.CreateOr(MemoryRuntimeCheck, IsConflict,
1645 FirstInst = getFirstInst(FirstInst, IsConflict, Loc);
1647 MemoryRuntimeCheck = IsConflict;
1651 if (!MemoryRuntimeCheck)
1652 return std::make_pair(nullptr, nullptr);
1654 // We have to do this trickery because the IRBuilder might fold the check to a
1655 // constant expression in which case there is no Instruction anchored in a
1657 Instruction *Check = BinaryOperator::CreateAnd(MemoryRuntimeCheck,
1658 ConstantInt::getTrue(Ctx));
1659 ChkBuilder.Insert(Check, "memcheck.conflict");
1660 FirstInst = getFirstInst(FirstInst, Check, Loc);
1661 return std::make_pair(FirstInst, Check);
1664 LoopAccessInfo::LoopAccessInfo(Loop *L, ScalarEvolution *SE,
1665 const DataLayout &DL,
1666 const TargetLibraryInfo *TLI, AliasAnalysis *AA,
1667 DominatorTree *DT, LoopInfo *LI,
1668 const ValueToValueMap &Strides)
1669 : PtrRtChecking(SE), DepChecker(SE, L), TheLoop(L), SE(SE), DL(DL),
1670 TLI(TLI), AA(AA), DT(DT), LI(LI), NumLoads(0), NumStores(0),
1671 MaxSafeDepDistBytes(-1U), CanVecMem(false),
1672 StoreToLoopInvariantAddress(false) {
1673 if (canAnalyzeLoop())
1674 analyzeLoop(Strides);
1677 void LoopAccessInfo::print(raw_ostream &OS, unsigned Depth) const {
1679 if (PtrRtChecking.Need)
1680 OS.indent(Depth) << "Memory dependences are safe with run-time checks\n";
1682 OS.indent(Depth) << "Memory dependences are safe\n";
1686 OS.indent(Depth) << "Report: " << Report->str() << "\n";
1688 if (auto *InterestingDependences = DepChecker.getInterestingDependences()) {
1689 OS.indent(Depth) << "Interesting Dependences:\n";
1690 for (auto &Dep : *InterestingDependences) {
1691 Dep.print(OS, Depth + 2, DepChecker.getMemoryInstructions());
1695 OS.indent(Depth) << "Too many interesting dependences, not recorded\n";
1697 // List the pair of accesses need run-time checks to prove independence.
1698 PtrRtChecking.print(OS, Depth);
1701 OS.indent(Depth) << "Store to invariant address was "
1702 << (StoreToLoopInvariantAddress ? "" : "not ")
1703 << "found in loop.\n";
1706 const LoopAccessInfo &
1707 LoopAccessAnalysis::getInfo(Loop *L, const ValueToValueMap &Strides) {
1708 auto &LAI = LoopAccessInfoMap[L];
1711 assert((!LAI || LAI->NumSymbolicStrides == Strides.size()) &&
1712 "Symbolic strides changed for loop");
1716 const DataLayout &DL = L->getHeader()->getModule()->getDataLayout();
1717 LAI = llvm::make_unique<LoopAccessInfo>(L, SE, DL, TLI, AA, DT, LI,
1720 LAI->NumSymbolicStrides = Strides.size();
1726 void LoopAccessAnalysis::print(raw_ostream &OS, const Module *M) const {
1727 LoopAccessAnalysis &LAA = *const_cast<LoopAccessAnalysis *>(this);
1729 ValueToValueMap NoSymbolicStrides;
1731 for (Loop *TopLevelLoop : *LI)
1732 for (Loop *L : depth_first(TopLevelLoop)) {
1733 OS.indent(2) << L->getHeader()->getName() << ":\n";
1734 auto &LAI = LAA.getInfo(L, NoSymbolicStrides);
1739 bool LoopAccessAnalysis::runOnFunction(Function &F) {
1740 SE = &getAnalysis<ScalarEvolution>();
1741 auto *TLIP = getAnalysisIfAvailable<TargetLibraryInfoWrapperPass>();
1742 TLI = TLIP ? &TLIP->getTLI() : nullptr;
1743 AA = &getAnalysis<AliasAnalysis>();
1744 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1745 LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
1750 void LoopAccessAnalysis::getAnalysisUsage(AnalysisUsage &AU) const {
1751 AU.addRequired<ScalarEvolution>();
1752 AU.addRequired<AliasAnalysis>();
1753 AU.addRequired<DominatorTreeWrapperPass>();
1754 AU.addRequired<LoopInfoWrapperPass>();
1756 AU.setPreservesAll();
1759 char LoopAccessAnalysis::ID = 0;
1760 static const char laa_name[] = "Loop Access Analysis";
1761 #define LAA_NAME "loop-accesses"
1763 INITIALIZE_PASS_BEGIN(LoopAccessAnalysis, LAA_NAME, laa_name, false, true)
1764 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
1765 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
1766 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
1767 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
1768 INITIALIZE_PASS_END(LoopAccessAnalysis, LAA_NAME, laa_name, false, true)
1771 Pass *createLAAPass() {
1772 return new LoopAccessAnalysis();