1 //===- LoopAccessAnalysis.cpp - Loop Access Analysis Implementation --------==//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // The implementation for the loop memory dependence that was originally
11 // developed for the loop vectorizer.
13 //===----------------------------------------------------------------------===//
15 #include "llvm/Analysis/LoopAccessAnalysis.h"
16 #include "llvm/Analysis/LoopInfo.h"
17 #include "llvm/Analysis/ScalarEvolutionExpander.h"
18 #include "llvm/Analysis/TargetLibraryInfo.h"
19 #include "llvm/Analysis/ValueTracking.h"
20 #include "llvm/IR/DiagnosticInfo.h"
21 #include "llvm/IR/Dominators.h"
22 #include "llvm/IR/IRBuilder.h"
23 #include "llvm/Support/Debug.h"
24 #include "llvm/Support/raw_ostream.h"
25 #include "llvm/Analysis/VectorUtils.h"
28 #define DEBUG_TYPE "loop-accesses"
30 static cl::opt<unsigned, true>
31 VectorizationFactor("force-vector-width", cl::Hidden,
32 cl::desc("Sets the SIMD width. Zero is autoselect."),
33 cl::location(VectorizerParams::VectorizationFactor));
34 unsigned VectorizerParams::VectorizationFactor;
36 static cl::opt<unsigned, true>
37 VectorizationInterleave("force-vector-interleave", cl::Hidden,
38 cl::desc("Sets the vectorization interleave count. "
39 "Zero is autoselect."),
41 VectorizerParams::VectorizationInterleave));
42 unsigned VectorizerParams::VectorizationInterleave;
44 static cl::opt<unsigned, true> RuntimeMemoryCheckThreshold(
45 "runtime-memory-check-threshold", cl::Hidden,
46 cl::desc("When performing memory disambiguation checks at runtime do not "
47 "generate more than this number of comparisons (default = 8)."),
48 cl::location(VectorizerParams::RuntimeMemoryCheckThreshold), cl::init(8));
49 unsigned VectorizerParams::RuntimeMemoryCheckThreshold;
51 /// \brief The maximum iterations used to merge memory checks
52 static cl::opt<unsigned> MemoryCheckMergeThreshold(
53 "memory-check-merge-threshold", cl::Hidden,
54 cl::desc("Maximum number of comparisons done when trying to merge "
55 "runtime memory checks. (default = 100)"),
58 /// Maximum SIMD width.
59 const unsigned VectorizerParams::MaxVectorWidth = 64;
61 /// \brief We collect interesting dependences up to this threshold.
62 static cl::opt<unsigned> MaxInterestingDependence(
63 "max-interesting-dependences", cl::Hidden,
64 cl::desc("Maximum number of interesting dependences collected by "
65 "loop-access analysis (default = 100)"),
68 bool VectorizerParams::isInterleaveForced() {
69 return ::VectorizationInterleave.getNumOccurrences() > 0;
72 void LoopAccessReport::emitAnalysis(const LoopAccessReport &Message,
73 const Function *TheFunction,
75 const char *PassName) {
76 DebugLoc DL = TheLoop->getStartLoc();
77 if (const Instruction *I = Message.getInstr())
78 DL = I->getDebugLoc();
79 emitOptimizationRemarkAnalysis(TheFunction->getContext(), PassName,
80 *TheFunction, DL, Message.str());
83 Value *llvm::stripIntegerCast(Value *V) {
84 if (CastInst *CI = dyn_cast<CastInst>(V))
85 if (CI->getOperand(0)->getType()->isIntegerTy())
86 return CI->getOperand(0);
90 const SCEV *llvm::replaceSymbolicStrideSCEV(ScalarEvolution *SE,
91 const ValueToValueMap &PtrToStride,
92 Value *Ptr, Value *OrigPtr) {
94 const SCEV *OrigSCEV = SE->getSCEV(Ptr);
96 // If there is an entry in the map return the SCEV of the pointer with the
97 // symbolic stride replaced by one.
98 ValueToValueMap::const_iterator SI =
99 PtrToStride.find(OrigPtr ? OrigPtr : Ptr);
100 if (SI != PtrToStride.end()) {
101 Value *StrideVal = SI->second;
104 StrideVal = stripIntegerCast(StrideVal);
106 // Replace symbolic stride by one.
107 Value *One = ConstantInt::get(StrideVal->getType(), 1);
108 ValueToValueMap RewriteMap;
109 RewriteMap[StrideVal] = One;
112 SCEVParameterRewriter::rewrite(OrigSCEV, *SE, RewriteMap, true);
113 DEBUG(dbgs() << "LAA: Replacing SCEV: " << *OrigSCEV << " by: " << *ByOne
118 // Otherwise, just return the SCEV of the original pointer.
119 return SE->getSCEV(Ptr);
122 void LoopAccessInfo::RuntimePointerCheck::insert(
123 Loop *Lp, Value *Ptr, bool WritePtr, unsigned DepSetId, unsigned ASId,
124 const ValueToValueMap &Strides) {
125 // Get the stride replaced scev.
126 const SCEV *Sc = replaceSymbolicStrideSCEV(SE, Strides, Ptr);
127 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Sc);
128 assert(AR && "Invalid addrec expression");
129 const SCEV *Ex = SE->getBackedgeTakenCount(Lp);
130 const SCEV *ScEnd = AR->evaluateAtIteration(Ex, *SE);
131 Pointers.push_back(Ptr);
132 Starts.push_back(AR->getStart());
133 Ends.push_back(ScEnd);
134 IsWritePtr.push_back(WritePtr);
135 DependencySetId.push_back(DepSetId);
136 AliasSetId.push_back(ASId);
140 bool LoopAccessInfo::RuntimePointerCheck::needsChecking(
141 const CheckingPtrGroup &M, const CheckingPtrGroup &N,
142 const SmallVectorImpl<int> *PtrPartition) const {
143 for (unsigned I = 0, EI = M.Members.size(); EI != I; ++I)
144 for (unsigned J = 0, EJ = N.Members.size(); EJ != J; ++J)
145 if (needsChecking(M.Members[I], N.Members[J], PtrPartition))
150 /// Compare \p I and \p J and return the minimum.
151 /// Return nullptr in case we couldn't find an answer.
152 static const SCEV *getMinFromExprs(const SCEV *I, const SCEV *J,
153 ScalarEvolution *SE) {
154 const SCEV *Diff = SE->getMinusSCEV(J, I);
155 const SCEVConstant *C = dyn_cast<const SCEVConstant>(Diff);
159 if (C->getValue()->isNegative())
164 bool LoopAccessInfo::RuntimePointerCheck::CheckingPtrGroup::addPointer(
166 // Compare the starts and ends with the known minimum and maximum
167 // of this set. We need to know how we compare against the min/max
168 // of the set in order to be able to emit memchecks.
169 const SCEV *Min0 = getMinFromExprs(RtCheck.Starts[Index], Low, RtCheck.SE);
173 const SCEV *Min1 = getMinFromExprs(RtCheck.Ends[Index], High, RtCheck.SE);
177 // Update the low bound expression if we've found a new min value.
178 if (Min0 == RtCheck.Starts[Index])
179 Low = RtCheck.Starts[Index];
181 // Update the high bound expression if we've found a new max value.
182 if (Min1 != RtCheck.Ends[Index])
183 High = RtCheck.Ends[Index];
185 Members.push_back(Index);
189 void LoopAccessInfo::RuntimePointerCheck::groupChecks(
190 MemoryDepChecker::DepCandidates &DepCands,
191 bool UseDependencies) {
192 // We build the groups from dependency candidates equivalence classes
194 // - We know that pointers in the same equivalence class share
195 // the same underlying object and therefore there is a chance
196 // that we can compare pointers
197 // - We wouldn't be able to merge two pointers for which we need
198 // to emit a memcheck. The classes in DepCands are already
199 // conveniently built such that no two pointers in the same
200 // class need checking against each other.
202 // We use the following (greedy) algorithm to construct the groups
203 // For every pointer in the equivalence class:
204 // For each existing group:
205 // - if the difference between this pointer and the min/max bounds
206 // of the group is a constant, then make the pointer part of the
207 // group and update the min/max bounds of that group as required.
209 CheckingGroups.clear();
211 // If we don't have the dependency partitions, construct a new
212 // checking pointer group for each pointer.
213 if (!UseDependencies) {
214 for (unsigned I = 0; I < Pointers.size(); ++I)
215 CheckingGroups.push_back(CheckingPtrGroup(I, *this));
219 unsigned TotalComparisons = 0;
221 DenseMap<Value *, unsigned> PositionMap;
222 for (unsigned Pointer = 0; Pointer < Pointers.size(); ++Pointer)
223 PositionMap[Pointers[Pointer]] = Pointer;
225 // Go through all equivalence classes, get the the "pointer check groups"
226 // and add them to the overall solution.
227 for (auto DI = DepCands.begin(), DE = DepCands.end(); DI != DE; ++DI) {
231 SmallVector<CheckingPtrGroup, 2> Groups;
233 for (auto MI = DepCands.member_begin(DI), ME = DepCands.member_end();
235 unsigned Pointer = PositionMap[MI->getPointer()];
238 // Go through all the existing sets and see if we can find one
239 // which can include this pointer.
240 for (CheckingPtrGroup &Group : Groups) {
241 // Don't perform more than a certain amount of comparisons.
242 // This should limit the cost of grouping the pointers to something
243 // reasonable. If we do end up hitting this threshold, the algorithm
244 // will create separate groups for all remaining pointers.
245 if (TotalComparisons > MemoryCheckMergeThreshold)
250 if (Group.addPointer(Pointer)) {
257 // We couldn't add this pointer to any existing set or the threshold
258 // for the number of comparisons has been reached. Create a new group
259 // to hold the current pointer.
260 Groups.push_back(CheckingPtrGroup(Pointer, *this));
263 // We've computed the grouped checks for this partition.
264 // Save the results and continue with the next one.
265 std::copy(Groups.begin(), Groups.end(), std::back_inserter(CheckingGroups));
269 bool LoopAccessInfo::RuntimePointerCheck::needsChecking(
270 unsigned I, unsigned J, const SmallVectorImpl<int> *PtrPartition) const {
271 // No need to check if two readonly pointers intersect.
272 if (!IsWritePtr[I] && !IsWritePtr[J])
275 // Only need to check pointers between two different dependency sets.
276 if (DependencySetId[I] == DependencySetId[J])
279 // Only need to check pointers in the same alias set.
280 if (AliasSetId[I] != AliasSetId[J])
283 // If PtrPartition is set omit checks between pointers of the same partition.
284 // Partition number -1 means that the pointer is used in multiple partitions.
285 // In this case we can't omit the check.
286 if (PtrPartition && (*PtrPartition)[I] != -1 &&
287 (*PtrPartition)[I] == (*PtrPartition)[J])
293 void LoopAccessInfo::RuntimePointerCheck::print(
294 raw_ostream &OS, unsigned Depth,
295 const SmallVectorImpl<int> *PtrPartition) const {
297 OS.indent(Depth) << "Run-time memory checks:\n";
300 for (unsigned I = 0; I < CheckingGroups.size(); ++I)
301 for (unsigned J = I + 1; J < CheckingGroups.size(); ++J)
302 if (needsChecking(CheckingGroups[I], CheckingGroups[J], PtrPartition)) {
303 OS.indent(Depth) << "Check " << N++ << ":\n";
304 OS.indent(Depth + 2) << "Comparing group " << I << ":\n";
306 for (unsigned K = 0; K < CheckingGroups[I].Members.size(); ++K) {
307 OS.indent(Depth + 2) << *Pointers[CheckingGroups[I].Members[K]]
310 OS << " (Partition: "
311 << (*PtrPartition)[CheckingGroups[I].Members[K]] << ")"
315 OS.indent(Depth + 2) << "Against group " << J << ":\n";
317 for (unsigned K = 0; K < CheckingGroups[J].Members.size(); ++K) {
318 OS.indent(Depth + 2) << *Pointers[CheckingGroups[J].Members[K]]
321 OS << " (Partition: "
322 << (*PtrPartition)[CheckingGroups[J].Members[K]] << ")"
327 OS.indent(Depth) << "Grouped accesses:\n";
328 for (unsigned I = 0; I < CheckingGroups.size(); ++I) {
329 OS.indent(Depth + 2) << "Group " << I << ":\n";
330 OS.indent(Depth + 4) << "(Low: " << *CheckingGroups[I].Low
331 << " High: " << *CheckingGroups[I].High << ")\n";
332 for (unsigned J = 0; J < CheckingGroups[I].Members.size(); ++J) {
333 OS.indent(Depth + 6) << "Member: " << *Exprs[CheckingGroups[I].Members[J]]
339 unsigned LoopAccessInfo::RuntimePointerCheck::getNumberOfChecks(
340 const SmallVectorImpl<int> *PtrPartition) const {
342 unsigned NumPartitions = CheckingGroups.size();
343 unsigned CheckCount = 0;
345 for (unsigned I = 0; I < NumPartitions; ++I)
346 for (unsigned J = I + 1; J < NumPartitions; ++J)
347 if (needsChecking(CheckingGroups[I], CheckingGroups[J], PtrPartition))
352 bool LoopAccessInfo::RuntimePointerCheck::needsAnyChecking(
353 const SmallVectorImpl<int> *PtrPartition) const {
354 unsigned NumPointers = Pointers.size();
356 for (unsigned I = 0; I < NumPointers; ++I)
357 for (unsigned J = I + 1; J < NumPointers; ++J)
358 if (needsChecking(I, J, PtrPartition))
364 /// \brief Analyses memory accesses in a loop.
366 /// Checks whether run time pointer checks are needed and builds sets for data
367 /// dependence checking.
368 class AccessAnalysis {
370 /// \brief Read or write access location.
371 typedef PointerIntPair<Value *, 1, bool> MemAccessInfo;
372 typedef SmallPtrSet<MemAccessInfo, 8> MemAccessInfoSet;
374 AccessAnalysis(const DataLayout &Dl, AliasAnalysis *AA, LoopInfo *LI,
375 MemoryDepChecker::DepCandidates &DA)
376 : DL(Dl), AST(*AA), LI(LI), DepCands(DA), IsRTCheckNeeded(false) {}
378 /// \brief Register a load and whether it is only read from.
379 void addLoad(MemoryLocation &Loc, bool IsReadOnly) {
380 Value *Ptr = const_cast<Value*>(Loc.Ptr);
381 AST.add(Ptr, MemoryLocation::UnknownSize, Loc.AATags);
382 Accesses.insert(MemAccessInfo(Ptr, false));
384 ReadOnlyPtr.insert(Ptr);
387 /// \brief Register a store.
388 void addStore(MemoryLocation &Loc) {
389 Value *Ptr = const_cast<Value*>(Loc.Ptr);
390 AST.add(Ptr, MemoryLocation::UnknownSize, Loc.AATags);
391 Accesses.insert(MemAccessInfo(Ptr, true));
394 /// \brief Check whether we can check the pointers at runtime for
395 /// non-intersection. Returns true when we have 0 pointers
396 /// (a check on 0 pointers for non-intersection will always return true).
397 bool canCheckPtrAtRT(LoopAccessInfo::RuntimePointerCheck &RtCheck,
398 bool &NeedRTCheck, ScalarEvolution *SE, Loop *TheLoop,
399 const ValueToValueMap &Strides,
400 bool ShouldCheckStride = false);
402 /// \brief Goes over all memory accesses, checks whether a RT check is needed
403 /// and builds sets of dependent accesses.
404 void buildDependenceSets() {
405 processMemAccesses();
408 bool isRTCheckNeeded() { return IsRTCheckNeeded; }
410 bool isDependencyCheckNeeded() { return !CheckDeps.empty(); }
412 /// We decided that no dependence analysis would be used. Reset the state.
413 void resetDepChecks(MemoryDepChecker &DepChecker) {
415 DepChecker.clearInterestingDependences();
418 MemAccessInfoSet &getDependenciesToCheck() { return CheckDeps; }
421 typedef SetVector<MemAccessInfo> PtrAccessSet;
423 /// \brief Go over all memory access and check whether runtime pointer checks
424 /// are needed /// and build sets of dependency check candidates.
425 void processMemAccesses();
427 /// Set of all accesses.
428 PtrAccessSet Accesses;
430 const DataLayout &DL;
432 /// Set of accesses that need a further dependence check.
433 MemAccessInfoSet CheckDeps;
435 /// Set of pointers that are read only.
436 SmallPtrSet<Value*, 16> ReadOnlyPtr;
438 /// An alias set tracker to partition the access set by underlying object and
439 //intrinsic property (such as TBAA metadata).
444 /// Sets of potentially dependent accesses - members of one set share an
445 /// underlying pointer. The set "CheckDeps" identfies which sets really need a
446 /// dependence check.
447 MemoryDepChecker::DepCandidates &DepCands;
449 bool IsRTCheckNeeded;
452 } // end anonymous namespace
454 /// \brief Check whether a pointer can participate in a runtime bounds check.
455 static bool hasComputableBounds(ScalarEvolution *SE,
456 const ValueToValueMap &Strides, Value *Ptr) {
457 const SCEV *PtrScev = replaceSymbolicStrideSCEV(SE, Strides, Ptr);
458 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
462 return AR->isAffine();
465 bool AccessAnalysis::canCheckPtrAtRT(
466 LoopAccessInfo::RuntimePointerCheck &RtCheck, bool &NeedRTCheck,
467 ScalarEvolution *SE, Loop *TheLoop, const ValueToValueMap &StridesMap,
468 bool ShouldCheckStride) {
469 // Find pointers with computable bounds. We are going to use this information
470 // to place a runtime bound check.
474 if (!IsRTCheckNeeded) return true;
476 bool IsDepCheckNeeded = isDependencyCheckNeeded();
478 // We assign a consecutive id to access from different alias sets.
479 // Accesses between different groups doesn't need to be checked.
481 for (auto &AS : AST) {
482 // We assign consecutive id to access from different dependence sets.
483 // Accesses within the same set don't need a runtime check.
484 unsigned RunningDepId = 1;
485 DenseMap<Value *, unsigned> DepSetId;
488 Value *Ptr = A.getValue();
489 bool IsWrite = Accesses.count(MemAccessInfo(Ptr, true));
490 MemAccessInfo Access(Ptr, IsWrite);
492 if (hasComputableBounds(SE, StridesMap, Ptr) &&
493 // When we run after a failing dependency check we have to make sure
494 // we don't have wrapping pointers.
495 (!ShouldCheckStride ||
496 isStridedPtr(SE, Ptr, TheLoop, StridesMap) == 1)) {
497 // The id of the dependence set.
500 if (IsDepCheckNeeded) {
501 Value *Leader = DepCands.getLeaderValue(Access).getPointer();
502 unsigned &LeaderId = DepSetId[Leader];
504 LeaderId = RunningDepId++;
507 // Each access has its own dependence set.
508 DepId = RunningDepId++;
510 RtCheck.insert(TheLoop, Ptr, IsWrite, DepId, ASId, StridesMap);
512 DEBUG(dbgs() << "LAA: Found a runtime check ptr:" << *Ptr << '\n');
514 DEBUG(dbgs() << "LAA: Can't find bounds for ptr:" << *Ptr << '\n');
522 // We need a runtime check if there are any accesses that need checking.
523 // However, some accesses cannot be checked (for example because we
524 // can't determine their bounds). In these cases we would need a check
525 // but wouldn't be able to add it.
526 NeedRTCheck = !CanDoRT || RtCheck.needsAnyChecking(nullptr);
528 // If the pointers that we would use for the bounds comparison have different
529 // address spaces, assume the values aren't directly comparable, so we can't
530 // use them for the runtime check. We also have to assume they could
531 // overlap. In the future there should be metadata for whether address spaces
533 unsigned NumPointers = RtCheck.Pointers.size();
534 for (unsigned i = 0; i < NumPointers; ++i) {
535 for (unsigned j = i + 1; j < NumPointers; ++j) {
536 // Only need to check pointers between two different dependency sets.
537 if (RtCheck.DependencySetId[i] == RtCheck.DependencySetId[j])
539 // Only need to check pointers in the same alias set.
540 if (RtCheck.AliasSetId[i] != RtCheck.AliasSetId[j])
543 Value *PtrI = RtCheck.Pointers[i];
544 Value *PtrJ = RtCheck.Pointers[j];
546 unsigned ASi = PtrI->getType()->getPointerAddressSpace();
547 unsigned ASj = PtrJ->getType()->getPointerAddressSpace();
549 DEBUG(dbgs() << "LAA: Runtime check would require comparison between"
550 " different address spaces\n");
556 if (NeedRTCheck && CanDoRT)
557 RtCheck.groupChecks(DepCands, IsDepCheckNeeded);
562 void AccessAnalysis::processMemAccesses() {
563 // We process the set twice: first we process read-write pointers, last we
564 // process read-only pointers. This allows us to skip dependence tests for
565 // read-only pointers.
567 DEBUG(dbgs() << "LAA: Processing memory accesses...\n");
568 DEBUG(dbgs() << " AST: "; AST.dump());
569 DEBUG(dbgs() << "LAA: Accesses(" << Accesses.size() << "):\n");
571 for (auto A : Accesses)
572 dbgs() << "\t" << *A.getPointer() << " (" <<
573 (A.getInt() ? "write" : (ReadOnlyPtr.count(A.getPointer()) ?
574 "read-only" : "read")) << ")\n";
577 // The AliasSetTracker has nicely partitioned our pointers by metadata
578 // compatibility and potential for underlying-object overlap. As a result, we
579 // only need to check for potential pointer dependencies within each alias
581 for (auto &AS : AST) {
582 // Note that both the alias-set tracker and the alias sets themselves used
583 // linked lists internally and so the iteration order here is deterministic
584 // (matching the original instruction order within each set).
586 bool SetHasWrite = false;
588 // Map of pointers to last access encountered.
589 typedef DenseMap<Value*, MemAccessInfo> UnderlyingObjToAccessMap;
590 UnderlyingObjToAccessMap ObjToLastAccess;
592 // Set of access to check after all writes have been processed.
593 PtrAccessSet DeferredAccesses;
595 // Iterate over each alias set twice, once to process read/write pointers,
596 // and then to process read-only pointers.
597 for (int SetIteration = 0; SetIteration < 2; ++SetIteration) {
598 bool UseDeferred = SetIteration > 0;
599 PtrAccessSet &S = UseDeferred ? DeferredAccesses : Accesses;
602 Value *Ptr = AV.getValue();
604 // For a single memory access in AliasSetTracker, Accesses may contain
605 // both read and write, and they both need to be handled for CheckDeps.
607 if (AC.getPointer() != Ptr)
610 bool IsWrite = AC.getInt();
612 // If we're using the deferred access set, then it contains only
614 bool IsReadOnlyPtr = ReadOnlyPtr.count(Ptr) && !IsWrite;
615 if (UseDeferred && !IsReadOnlyPtr)
617 // Otherwise, the pointer must be in the PtrAccessSet, either as a
619 assert(((IsReadOnlyPtr && UseDeferred) || IsWrite ||
620 S.count(MemAccessInfo(Ptr, false))) &&
621 "Alias-set pointer not in the access set?");
623 MemAccessInfo Access(Ptr, IsWrite);
624 DepCands.insert(Access);
626 // Memorize read-only pointers for later processing and skip them in
627 // the first round (they need to be checked after we have seen all
628 // write pointers). Note: we also mark pointer that are not
629 // consecutive as "read-only" pointers (so that we check
630 // "a[b[i]] +="). Hence, we need the second check for "!IsWrite".
631 if (!UseDeferred && IsReadOnlyPtr) {
632 DeferredAccesses.insert(Access);
636 // If this is a write - check other reads and writes for conflicts. If
637 // this is a read only check other writes for conflicts (but only if
638 // there is no other write to the ptr - this is an optimization to
639 // catch "a[i] = a[i] + " without having to do a dependence check).
640 if ((IsWrite || IsReadOnlyPtr) && SetHasWrite) {
641 CheckDeps.insert(Access);
642 IsRTCheckNeeded = true;
648 // Create sets of pointers connected by a shared alias set and
649 // underlying object.
650 typedef SmallVector<Value *, 16> ValueVector;
651 ValueVector TempObjects;
653 GetUnderlyingObjects(Ptr, TempObjects, DL, LI);
654 DEBUG(dbgs() << "Underlying objects for pointer " << *Ptr << "\n");
655 for (Value *UnderlyingObj : TempObjects) {
656 UnderlyingObjToAccessMap::iterator Prev =
657 ObjToLastAccess.find(UnderlyingObj);
658 if (Prev != ObjToLastAccess.end())
659 DepCands.unionSets(Access, Prev->second);
661 ObjToLastAccess[UnderlyingObj] = Access;
662 DEBUG(dbgs() << " " << *UnderlyingObj << "\n");
670 static bool isInBoundsGep(Value *Ptr) {
671 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr))
672 return GEP->isInBounds();
676 /// \brief Return true if an AddRec pointer \p Ptr is unsigned non-wrapping,
677 /// i.e. monotonically increasing/decreasing.
678 static bool isNoWrapAddRec(Value *Ptr, const SCEVAddRecExpr *AR,
679 ScalarEvolution *SE, const Loop *L) {
680 // FIXME: This should probably only return true for NUW.
681 if (AR->getNoWrapFlags(SCEV::NoWrapMask))
684 // Scalar evolution does not propagate the non-wrapping flags to values that
685 // are derived from a non-wrapping induction variable because non-wrapping
686 // could be flow-sensitive.
688 // Look through the potentially overflowing instruction to try to prove
689 // non-wrapping for the *specific* value of Ptr.
691 // The arithmetic implied by an inbounds GEP can't overflow.
692 auto *GEP = dyn_cast<GetElementPtrInst>(Ptr);
693 if (!GEP || !GEP->isInBounds())
696 // Make sure there is only one non-const index and analyze that.
697 Value *NonConstIndex = nullptr;
698 for (auto Index = GEP->idx_begin(); Index != GEP->idx_end(); ++Index)
699 if (!isa<ConstantInt>(*Index)) {
702 NonConstIndex = *Index;
705 // The recurrence is on the pointer, ignore for now.
708 // The index in GEP is signed. It is non-wrapping if it's derived from a NSW
709 // AddRec using a NSW operation.
710 if (auto *OBO = dyn_cast<OverflowingBinaryOperator>(NonConstIndex))
711 if (OBO->hasNoSignedWrap() &&
712 // Assume constant for other the operand so that the AddRec can be
714 isa<ConstantInt>(OBO->getOperand(1))) {
715 auto *OpScev = SE->getSCEV(OBO->getOperand(0));
717 if (auto *OpAR = dyn_cast<SCEVAddRecExpr>(OpScev))
718 return OpAR->getLoop() == L && OpAR->getNoWrapFlags(SCEV::FlagNSW);
724 /// \brief Check whether the access through \p Ptr has a constant stride.
725 int llvm::isStridedPtr(ScalarEvolution *SE, Value *Ptr, const Loop *Lp,
726 const ValueToValueMap &StridesMap) {
727 const Type *Ty = Ptr->getType();
728 assert(Ty->isPointerTy() && "Unexpected non-ptr");
730 // Make sure that the pointer does not point to aggregate types.
731 const PointerType *PtrTy = cast<PointerType>(Ty);
732 if (PtrTy->getElementType()->isAggregateType()) {
733 DEBUG(dbgs() << "LAA: Bad stride - Not a pointer to a scalar type"
738 const SCEV *PtrScev = replaceSymbolicStrideSCEV(SE, StridesMap, Ptr);
740 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
742 DEBUG(dbgs() << "LAA: Bad stride - Not an AddRecExpr pointer "
743 << *Ptr << " SCEV: " << *PtrScev << "\n");
747 // The accesss function must stride over the innermost loop.
748 if (Lp != AR->getLoop()) {
749 DEBUG(dbgs() << "LAA: Bad stride - Not striding over innermost loop " <<
750 *Ptr << " SCEV: " << *PtrScev << "\n");
753 // The address calculation must not wrap. Otherwise, a dependence could be
755 // An inbounds getelementptr that is a AddRec with a unit stride
756 // cannot wrap per definition. The unit stride requirement is checked later.
757 // An getelementptr without an inbounds attribute and unit stride would have
758 // to access the pointer value "0" which is undefined behavior in address
759 // space 0, therefore we can also vectorize this case.
760 bool IsInBoundsGEP = isInBoundsGep(Ptr);
761 bool IsNoWrapAddRec = isNoWrapAddRec(Ptr, AR, SE, Lp);
762 bool IsInAddressSpaceZero = PtrTy->getAddressSpace() == 0;
763 if (!IsNoWrapAddRec && !IsInBoundsGEP && !IsInAddressSpaceZero) {
764 DEBUG(dbgs() << "LAA: Bad stride - Pointer may wrap in the address space "
765 << *Ptr << " SCEV: " << *PtrScev << "\n");
769 // Check the step is constant.
770 const SCEV *Step = AR->getStepRecurrence(*SE);
772 // Calculate the pointer stride and check if it is consecutive.
773 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
775 DEBUG(dbgs() << "LAA: Bad stride - Not a constant strided " << *Ptr <<
776 " SCEV: " << *PtrScev << "\n");
780 auto &DL = Lp->getHeader()->getModule()->getDataLayout();
781 int64_t Size = DL.getTypeAllocSize(PtrTy->getElementType());
782 const APInt &APStepVal = C->getValue()->getValue();
784 // Huge step value - give up.
785 if (APStepVal.getBitWidth() > 64)
788 int64_t StepVal = APStepVal.getSExtValue();
791 int64_t Stride = StepVal / Size;
792 int64_t Rem = StepVal % Size;
796 // If the SCEV could wrap but we have an inbounds gep with a unit stride we
797 // know we can't "wrap around the address space". In case of address space
798 // zero we know that this won't happen without triggering undefined behavior.
799 if (!IsNoWrapAddRec && (IsInBoundsGEP || IsInAddressSpaceZero) &&
800 Stride != 1 && Stride != -1)
806 bool MemoryDepChecker::Dependence::isSafeForVectorization(DepType Type) {
810 case BackwardVectorizable:
814 case ForwardButPreventsForwarding:
816 case BackwardVectorizableButPreventsForwarding:
819 llvm_unreachable("unexpected DepType!");
822 bool MemoryDepChecker::Dependence::isInterestingDependence(DepType Type) {
828 case BackwardVectorizable:
830 case ForwardButPreventsForwarding:
832 case BackwardVectorizableButPreventsForwarding:
835 llvm_unreachable("unexpected DepType!");
838 bool MemoryDepChecker::Dependence::isPossiblyBackward() const {
842 case ForwardButPreventsForwarding:
846 case BackwardVectorizable:
848 case BackwardVectorizableButPreventsForwarding:
851 llvm_unreachable("unexpected DepType!");
854 bool MemoryDepChecker::couldPreventStoreLoadForward(unsigned Distance,
855 unsigned TypeByteSize) {
856 // If loads occur at a distance that is not a multiple of a feasible vector
857 // factor store-load forwarding does not take place.
858 // Positive dependences might cause troubles because vectorizing them might
859 // prevent store-load forwarding making vectorized code run a lot slower.
860 // a[i] = a[i-3] ^ a[i-8];
861 // The stores to a[i:i+1] don't align with the stores to a[i-3:i-2] and
862 // hence on your typical architecture store-load forwarding does not take
863 // place. Vectorizing in such cases does not make sense.
864 // Store-load forwarding distance.
865 const unsigned NumCyclesForStoreLoadThroughMemory = 8*TypeByteSize;
866 // Maximum vector factor.
867 unsigned MaxVFWithoutSLForwardIssues =
868 VectorizerParams::MaxVectorWidth * TypeByteSize;
869 if(MaxSafeDepDistBytes < MaxVFWithoutSLForwardIssues)
870 MaxVFWithoutSLForwardIssues = MaxSafeDepDistBytes;
872 for (unsigned vf = 2*TypeByteSize; vf <= MaxVFWithoutSLForwardIssues;
874 if (Distance % vf && Distance / vf < NumCyclesForStoreLoadThroughMemory) {
875 MaxVFWithoutSLForwardIssues = (vf >>=1);
880 if (MaxVFWithoutSLForwardIssues< 2*TypeByteSize) {
881 DEBUG(dbgs() << "LAA: Distance " << Distance <<
882 " that could cause a store-load forwarding conflict\n");
886 if (MaxVFWithoutSLForwardIssues < MaxSafeDepDistBytes &&
887 MaxVFWithoutSLForwardIssues !=
888 VectorizerParams::MaxVectorWidth * TypeByteSize)
889 MaxSafeDepDistBytes = MaxVFWithoutSLForwardIssues;
893 /// \brief Check the dependence for two accesses with the same stride \p Stride.
894 /// \p Distance is the positive distance and \p TypeByteSize is type size in
897 /// \returns true if they are independent.
898 static bool areStridedAccessesIndependent(unsigned Distance, unsigned Stride,
899 unsigned TypeByteSize) {
900 assert(Stride > 1 && "The stride must be greater than 1");
901 assert(TypeByteSize > 0 && "The type size in byte must be non-zero");
902 assert(Distance > 0 && "The distance must be non-zero");
904 // Skip if the distance is not multiple of type byte size.
905 if (Distance % TypeByteSize)
908 unsigned ScaledDist = Distance / TypeByteSize;
910 // No dependence if the scaled distance is not multiple of the stride.
912 // for (i = 0; i < 1024 ; i += 4)
913 // A[i+2] = A[i] + 1;
915 // Two accesses in memory (scaled distance is 2, stride is 4):
916 // | A[0] | | | | A[4] | | | |
917 // | | | A[2] | | | | A[6] | |
920 // for (i = 0; i < 1024 ; i += 3)
921 // A[i+4] = A[i] + 1;
923 // Two accesses in memory (scaled distance is 4, stride is 3):
924 // | A[0] | | | A[3] | | | A[6] | | |
925 // | | | | | A[4] | | | A[7] | |
926 return ScaledDist % Stride;
929 MemoryDepChecker::Dependence::DepType
930 MemoryDepChecker::isDependent(const MemAccessInfo &A, unsigned AIdx,
931 const MemAccessInfo &B, unsigned BIdx,
932 const ValueToValueMap &Strides) {
933 assert (AIdx < BIdx && "Must pass arguments in program order");
935 Value *APtr = A.getPointer();
936 Value *BPtr = B.getPointer();
937 bool AIsWrite = A.getInt();
938 bool BIsWrite = B.getInt();
940 // Two reads are independent.
941 if (!AIsWrite && !BIsWrite)
942 return Dependence::NoDep;
944 // We cannot check pointers in different address spaces.
945 if (APtr->getType()->getPointerAddressSpace() !=
946 BPtr->getType()->getPointerAddressSpace())
947 return Dependence::Unknown;
949 const SCEV *AScev = replaceSymbolicStrideSCEV(SE, Strides, APtr);
950 const SCEV *BScev = replaceSymbolicStrideSCEV(SE, Strides, BPtr);
952 int StrideAPtr = isStridedPtr(SE, APtr, InnermostLoop, Strides);
953 int StrideBPtr = isStridedPtr(SE, BPtr, InnermostLoop, Strides);
955 const SCEV *Src = AScev;
956 const SCEV *Sink = BScev;
958 // If the induction step is negative we have to invert source and sink of the
960 if (StrideAPtr < 0) {
963 std::swap(APtr, BPtr);
964 std::swap(Src, Sink);
965 std::swap(AIsWrite, BIsWrite);
966 std::swap(AIdx, BIdx);
967 std::swap(StrideAPtr, StrideBPtr);
970 const SCEV *Dist = SE->getMinusSCEV(Sink, Src);
972 DEBUG(dbgs() << "LAA: Src Scev: " << *Src << "Sink Scev: " << *Sink
973 << "(Induction step: " << StrideAPtr << ")\n");
974 DEBUG(dbgs() << "LAA: Distance for " << *InstMap[AIdx] << " to "
975 << *InstMap[BIdx] << ": " << *Dist << "\n");
977 // Need consecutive accesses. We don't want to vectorize
978 // "A[B[i]] += ..." and similar code or pointer arithmetic that could wrap in
979 // the address space.
980 if (!StrideAPtr || !StrideBPtr || StrideAPtr != StrideBPtr){
981 DEBUG(dbgs() << "Non-consecutive pointer access\n");
982 return Dependence::Unknown;
985 const SCEVConstant *C = dyn_cast<SCEVConstant>(Dist);
987 DEBUG(dbgs() << "LAA: Dependence because of non-constant distance\n");
988 ShouldRetryWithRuntimeCheck = true;
989 return Dependence::Unknown;
992 Type *ATy = APtr->getType()->getPointerElementType();
993 Type *BTy = BPtr->getType()->getPointerElementType();
994 auto &DL = InnermostLoop->getHeader()->getModule()->getDataLayout();
995 unsigned TypeByteSize = DL.getTypeAllocSize(ATy);
997 // Negative distances are not plausible dependencies.
998 const APInt &Val = C->getValue()->getValue();
999 if (Val.isNegative()) {
1000 bool IsTrueDataDependence = (AIsWrite && !BIsWrite);
1001 if (IsTrueDataDependence &&
1002 (couldPreventStoreLoadForward(Val.abs().getZExtValue(), TypeByteSize) ||
1004 return Dependence::ForwardButPreventsForwarding;
1006 DEBUG(dbgs() << "LAA: Dependence is negative: NoDep\n");
1007 return Dependence::Forward;
1010 // Write to the same location with the same size.
1011 // Could be improved to assert type sizes are the same (i32 == float, etc).
1014 return Dependence::NoDep;
1015 DEBUG(dbgs() << "LAA: Zero dependence difference but different types\n");
1016 return Dependence::Unknown;
1019 assert(Val.isStrictlyPositive() && "Expect a positive value");
1023 "LAA: ReadWrite-Write positive dependency with different types\n");
1024 return Dependence::Unknown;
1027 unsigned Distance = (unsigned) Val.getZExtValue();
1029 unsigned Stride = std::abs(StrideAPtr);
1031 areStridedAccessesIndependent(Distance, Stride, TypeByteSize)) {
1032 DEBUG(dbgs() << "LAA: Strided accesses are independent\n");
1033 return Dependence::NoDep;
1036 // Bail out early if passed-in parameters make vectorization not feasible.
1037 unsigned ForcedFactor = (VectorizerParams::VectorizationFactor ?
1038 VectorizerParams::VectorizationFactor : 1);
1039 unsigned ForcedUnroll = (VectorizerParams::VectorizationInterleave ?
1040 VectorizerParams::VectorizationInterleave : 1);
1041 // The minimum number of iterations for a vectorized/unrolled version.
1042 unsigned MinNumIter = std::max(ForcedFactor * ForcedUnroll, 2U);
1044 // It's not vectorizable if the distance is smaller than the minimum distance
1045 // needed for a vectroized/unrolled version. Vectorizing one iteration in
1046 // front needs TypeByteSize * Stride. Vectorizing the last iteration needs
1047 // TypeByteSize (No need to plus the last gap distance).
1049 // E.g. Assume one char is 1 byte in memory and one int is 4 bytes.
1051 // int *B = (int *)((char *)A + 14);
1052 // for (i = 0 ; i < 1024 ; i += 2)
1056 // Two accesses in memory (stride is 2):
1057 // | A[0] | | A[2] | | A[4] | | A[6] | |
1058 // | B[0] | | B[2] | | B[4] |
1060 // Distance needs for vectorizing iterations except the last iteration:
1061 // 4 * 2 * (MinNumIter - 1). Distance needs for the last iteration: 4.
1062 // So the minimum distance needed is: 4 * 2 * (MinNumIter - 1) + 4.
1064 // If MinNumIter is 2, it is vectorizable as the minimum distance needed is
1065 // 12, which is less than distance.
1067 // If MinNumIter is 4 (Say if a user forces the vectorization factor to be 4),
1068 // the minimum distance needed is 28, which is greater than distance. It is
1069 // not safe to do vectorization.
1070 unsigned MinDistanceNeeded =
1071 TypeByteSize * Stride * (MinNumIter - 1) + TypeByteSize;
1072 if (MinDistanceNeeded > Distance) {
1073 DEBUG(dbgs() << "LAA: Failure because of positive distance " << Distance
1075 return Dependence::Backward;
1078 // Unsafe if the minimum distance needed is greater than max safe distance.
1079 if (MinDistanceNeeded > MaxSafeDepDistBytes) {
1080 DEBUG(dbgs() << "LAA: Failure because it needs at least "
1081 << MinDistanceNeeded << " size in bytes");
1082 return Dependence::Backward;
1085 // Positive distance bigger than max vectorization factor.
1086 // FIXME: Should use max factor instead of max distance in bytes, which could
1087 // not handle different types.
1088 // E.g. Assume one char is 1 byte in memory and one int is 4 bytes.
1089 // void foo (int *A, char *B) {
1090 // for (unsigned i = 0; i < 1024; i++) {
1091 // A[i+2] = A[i] + 1;
1092 // B[i+2] = B[i] + 1;
1096 // This case is currently unsafe according to the max safe distance. If we
1097 // analyze the two accesses on array B, the max safe dependence distance
1098 // is 2. Then we analyze the accesses on array A, the minimum distance needed
1099 // is 8, which is less than 2 and forbidden vectorization, But actually
1100 // both A and B could be vectorized by 2 iterations.
1101 MaxSafeDepDistBytes =
1102 Distance < MaxSafeDepDistBytes ? Distance : MaxSafeDepDistBytes;
1104 bool IsTrueDataDependence = (!AIsWrite && BIsWrite);
1105 if (IsTrueDataDependence &&
1106 couldPreventStoreLoadForward(Distance, TypeByteSize))
1107 return Dependence::BackwardVectorizableButPreventsForwarding;
1109 DEBUG(dbgs() << "LAA: Positive distance " << Val.getSExtValue()
1110 << " with max VF = "
1111 << MaxSafeDepDistBytes / (TypeByteSize * Stride) << '\n');
1113 return Dependence::BackwardVectorizable;
1116 bool MemoryDepChecker::areDepsSafe(DepCandidates &AccessSets,
1117 MemAccessInfoSet &CheckDeps,
1118 const ValueToValueMap &Strides) {
1120 MaxSafeDepDistBytes = -1U;
1121 while (!CheckDeps.empty()) {
1122 MemAccessInfo CurAccess = *CheckDeps.begin();
1124 // Get the relevant memory access set.
1125 EquivalenceClasses<MemAccessInfo>::iterator I =
1126 AccessSets.findValue(AccessSets.getLeaderValue(CurAccess));
1128 // Check accesses within this set.
1129 EquivalenceClasses<MemAccessInfo>::member_iterator AI, AE;
1130 AI = AccessSets.member_begin(I), AE = AccessSets.member_end();
1132 // Check every access pair.
1134 CheckDeps.erase(*AI);
1135 EquivalenceClasses<MemAccessInfo>::member_iterator OI = std::next(AI);
1137 // Check every accessing instruction pair in program order.
1138 for (std::vector<unsigned>::iterator I1 = Accesses[*AI].begin(),
1139 I1E = Accesses[*AI].end(); I1 != I1E; ++I1)
1140 for (std::vector<unsigned>::iterator I2 = Accesses[*OI].begin(),
1141 I2E = Accesses[*OI].end(); I2 != I2E; ++I2) {
1142 auto A = std::make_pair(&*AI, *I1);
1143 auto B = std::make_pair(&*OI, *I2);
1149 Dependence::DepType Type =
1150 isDependent(*A.first, A.second, *B.first, B.second, Strides);
1151 SafeForVectorization &= Dependence::isSafeForVectorization(Type);
1153 // Gather dependences unless we accumulated MaxInterestingDependence
1154 // dependences. In that case return as soon as we find the first
1155 // unsafe dependence. This puts a limit on this quadratic
1157 if (RecordInterestingDependences) {
1158 if (Dependence::isInterestingDependence(Type))
1159 InterestingDependences.push_back(
1160 Dependence(A.second, B.second, Type));
1162 if (InterestingDependences.size() >= MaxInterestingDependence) {
1163 RecordInterestingDependences = false;
1164 InterestingDependences.clear();
1165 DEBUG(dbgs() << "Too many dependences, stopped recording\n");
1168 if (!RecordInterestingDependences && !SafeForVectorization)
1177 DEBUG(dbgs() << "Total Interesting Dependences: "
1178 << InterestingDependences.size() << "\n");
1179 return SafeForVectorization;
1182 SmallVector<Instruction *, 4>
1183 MemoryDepChecker::getInstructionsForAccess(Value *Ptr, bool isWrite) const {
1184 MemAccessInfo Access(Ptr, isWrite);
1185 auto &IndexVector = Accesses.find(Access)->second;
1187 SmallVector<Instruction *, 4> Insts;
1188 std::transform(IndexVector.begin(), IndexVector.end(),
1189 std::back_inserter(Insts),
1190 [&](unsigned Idx) { return this->InstMap[Idx]; });
1194 const char *MemoryDepChecker::Dependence::DepName[] = {
1195 "NoDep", "Unknown", "Forward", "ForwardButPreventsForwarding", "Backward",
1196 "BackwardVectorizable", "BackwardVectorizableButPreventsForwarding"};
1198 void MemoryDepChecker::Dependence::print(
1199 raw_ostream &OS, unsigned Depth,
1200 const SmallVectorImpl<Instruction *> &Instrs) const {
1201 OS.indent(Depth) << DepName[Type] << ":\n";
1202 OS.indent(Depth + 2) << *Instrs[Source] << " -> \n";
1203 OS.indent(Depth + 2) << *Instrs[Destination] << "\n";
1206 bool LoopAccessInfo::canAnalyzeLoop() {
1207 // We need to have a loop header.
1208 DEBUG(dbgs() << "LAA: Found a loop: " <<
1209 TheLoop->getHeader()->getName() << '\n');
1211 // We can only analyze innermost loops.
1212 if (!TheLoop->empty()) {
1213 DEBUG(dbgs() << "LAA: loop is not the innermost loop\n");
1214 emitAnalysis(LoopAccessReport() << "loop is not the innermost loop");
1218 // We must have a single backedge.
1219 if (TheLoop->getNumBackEdges() != 1) {
1220 DEBUG(dbgs() << "LAA: loop control flow is not understood by analyzer\n");
1222 LoopAccessReport() <<
1223 "loop control flow is not understood by analyzer");
1227 // We must have a single exiting block.
1228 if (!TheLoop->getExitingBlock()) {
1229 DEBUG(dbgs() << "LAA: loop control flow is not understood by analyzer\n");
1231 LoopAccessReport() <<
1232 "loop control flow is not understood by analyzer");
1236 // We only handle bottom-tested loops, i.e. loop in which the condition is
1237 // checked at the end of each iteration. With that we can assume that all
1238 // instructions in the loop are executed the same number of times.
1239 if (TheLoop->getExitingBlock() != TheLoop->getLoopLatch()) {
1240 DEBUG(dbgs() << "LAA: loop control flow is not understood by analyzer\n");
1242 LoopAccessReport() <<
1243 "loop control flow is not understood by analyzer");
1247 // ScalarEvolution needs to be able to find the exit count.
1248 const SCEV *ExitCount = SE->getBackedgeTakenCount(TheLoop);
1249 if (ExitCount == SE->getCouldNotCompute()) {
1250 emitAnalysis(LoopAccessReport() <<
1251 "could not determine number of loop iterations");
1252 DEBUG(dbgs() << "LAA: SCEV could not compute the loop exit count.\n");
1259 void LoopAccessInfo::analyzeLoop(const ValueToValueMap &Strides) {
1261 typedef SmallVector<Value*, 16> ValueVector;
1262 typedef SmallPtrSet<Value*, 16> ValueSet;
1264 // Holds the Load and Store *instructions*.
1268 // Holds all the different accesses in the loop.
1269 unsigned NumReads = 0;
1270 unsigned NumReadWrites = 0;
1272 PtrRtCheck.Pointers.clear();
1273 PtrRtCheck.Need = false;
1275 const bool IsAnnotatedParallel = TheLoop->isAnnotatedParallel();
1278 for (Loop::block_iterator bb = TheLoop->block_begin(),
1279 be = TheLoop->block_end(); bb != be; ++bb) {
1281 // Scan the BB and collect legal loads and stores.
1282 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
1285 // If this is a load, save it. If this instruction can read from memory
1286 // but is not a load, then we quit. Notice that we don't handle function
1287 // calls that read or write.
1288 if (it->mayReadFromMemory()) {
1289 // Many math library functions read the rounding mode. We will only
1290 // vectorize a loop if it contains known function calls that don't set
1291 // the flag. Therefore, it is safe to ignore this read from memory.
1292 CallInst *Call = dyn_cast<CallInst>(it);
1293 if (Call && getIntrinsicIDForCall(Call, TLI))
1296 // If the function has an explicit vectorized counterpart, we can safely
1297 // assume that it can be vectorized.
1298 if (Call && !Call->isNoBuiltin() && Call->getCalledFunction() &&
1299 TLI->isFunctionVectorizable(Call->getCalledFunction()->getName()))
1302 LoadInst *Ld = dyn_cast<LoadInst>(it);
1303 if (!Ld || (!Ld->isSimple() && !IsAnnotatedParallel)) {
1304 emitAnalysis(LoopAccessReport(Ld)
1305 << "read with atomic ordering or volatile read");
1306 DEBUG(dbgs() << "LAA: Found a non-simple load.\n");
1311 Loads.push_back(Ld);
1312 DepChecker.addAccess(Ld);
1316 // Save 'store' instructions. Abort if other instructions write to memory.
1317 if (it->mayWriteToMemory()) {
1318 StoreInst *St = dyn_cast<StoreInst>(it);
1320 emitAnalysis(LoopAccessReport(it) <<
1321 "instruction cannot be vectorized");
1325 if (!St->isSimple() && !IsAnnotatedParallel) {
1326 emitAnalysis(LoopAccessReport(St)
1327 << "write with atomic ordering or volatile write");
1328 DEBUG(dbgs() << "LAA: Found a non-simple store.\n");
1333 Stores.push_back(St);
1334 DepChecker.addAccess(St);
1339 // Now we have two lists that hold the loads and the stores.
1340 // Next, we find the pointers that they use.
1342 // Check if we see any stores. If there are no stores, then we don't
1343 // care if the pointers are *restrict*.
1344 if (!Stores.size()) {
1345 DEBUG(dbgs() << "LAA: Found a read-only loop!\n");
1350 MemoryDepChecker::DepCandidates DependentAccesses;
1351 AccessAnalysis Accesses(TheLoop->getHeader()->getModule()->getDataLayout(),
1352 AA, LI, DependentAccesses);
1354 // Holds the analyzed pointers. We don't want to call GetUnderlyingObjects
1355 // multiple times on the same object. If the ptr is accessed twice, once
1356 // for read and once for write, it will only appear once (on the write
1357 // list). This is okay, since we are going to check for conflicts between
1358 // writes and between reads and writes, but not between reads and reads.
1361 ValueVector::iterator I, IE;
1362 for (I = Stores.begin(), IE = Stores.end(); I != IE; ++I) {
1363 StoreInst *ST = cast<StoreInst>(*I);
1364 Value* Ptr = ST->getPointerOperand();
1365 // Check for store to loop invariant address.
1366 StoreToLoopInvariantAddress |= isUniform(Ptr);
1367 // If we did *not* see this pointer before, insert it to the read-write
1368 // list. At this phase it is only a 'write' list.
1369 if (Seen.insert(Ptr).second) {
1372 MemoryLocation Loc = MemoryLocation::get(ST);
1373 // The TBAA metadata could have a control dependency on the predication
1374 // condition, so we cannot rely on it when determining whether or not we
1375 // need runtime pointer checks.
1376 if (blockNeedsPredication(ST->getParent(), TheLoop, DT))
1377 Loc.AATags.TBAA = nullptr;
1379 Accesses.addStore(Loc);
1383 if (IsAnnotatedParallel) {
1385 << "LAA: A loop annotated parallel, ignore memory dependency "
1391 for (I = Loads.begin(), IE = Loads.end(); I != IE; ++I) {
1392 LoadInst *LD = cast<LoadInst>(*I);
1393 Value* Ptr = LD->getPointerOperand();
1394 // If we did *not* see this pointer before, insert it to the
1395 // read list. If we *did* see it before, then it is already in
1396 // the read-write list. This allows us to vectorize expressions
1397 // such as A[i] += x; Because the address of A[i] is a read-write
1398 // pointer. This only works if the index of A[i] is consecutive.
1399 // If the address of i is unknown (for example A[B[i]]) then we may
1400 // read a few words, modify, and write a few words, and some of the
1401 // words may be written to the same address.
1402 bool IsReadOnlyPtr = false;
1403 if (Seen.insert(Ptr).second || !isStridedPtr(SE, Ptr, TheLoop, Strides)) {
1405 IsReadOnlyPtr = true;
1408 MemoryLocation Loc = MemoryLocation::get(LD);
1409 // The TBAA metadata could have a control dependency on the predication
1410 // condition, so we cannot rely on it when determining whether or not we
1411 // need runtime pointer checks.
1412 if (blockNeedsPredication(LD->getParent(), TheLoop, DT))
1413 Loc.AATags.TBAA = nullptr;
1415 Accesses.addLoad(Loc, IsReadOnlyPtr);
1418 // If we write (or read-write) to a single destination and there are no
1419 // other reads in this loop then is it safe to vectorize.
1420 if (NumReadWrites == 1 && NumReads == 0) {
1421 DEBUG(dbgs() << "LAA: Found a write-only loop!\n");
1426 // Build dependence sets and check whether we need a runtime pointer bounds
1428 Accesses.buildDependenceSets();
1430 // Find pointers with computable bounds. We are going to use this information
1431 // to place a runtime bound check.
1433 bool CanDoRT = Accesses.canCheckPtrAtRT(PtrRtCheck,
1437 DEBUG(dbgs() << "LAA: We need to do "
1438 << PtrRtCheck.getNumberOfChecks(nullptr)
1439 << " pointer comparisons.\n");
1441 // Check that we found the bounds for the pointer.
1443 DEBUG(dbgs() << "LAA: We can perform a memory runtime check if needed.\n");
1444 else if (NeedRTCheck) {
1445 emitAnalysis(LoopAccessReport() << "cannot identify array bounds");
1446 DEBUG(dbgs() << "LAA: We can't vectorize because we can't find " <<
1447 "the array bounds.\n");
1453 PtrRtCheck.Need = NeedRTCheck;
1456 if (Accesses.isDependencyCheckNeeded()) {
1457 DEBUG(dbgs() << "LAA: Checking memory dependencies\n");
1458 CanVecMem = DepChecker.areDepsSafe(
1459 DependentAccesses, Accesses.getDependenciesToCheck(), Strides);
1460 MaxSafeDepDistBytes = DepChecker.getMaxSafeDepDistBytes();
1462 if (!CanVecMem && DepChecker.shouldRetryWithRuntimeCheck()) {
1463 DEBUG(dbgs() << "LAA: Retrying with memory checks\n");
1466 // Clear the dependency checks. We assume they are not needed.
1467 Accesses.resetDepChecks(DepChecker);
1470 PtrRtCheck.Need = true;
1472 CanDoRT = Accesses.canCheckPtrAtRT(PtrRtCheck, NeedRTCheck, SE,
1473 TheLoop, Strides, true);
1475 // Check that we found the bounds for the pointer.
1476 if (NeedRTCheck && !CanDoRT) {
1477 emitAnalysis(LoopAccessReport()
1478 << "cannot check memory dependencies at runtime");
1479 DEBUG(dbgs() << "LAA: Can't vectorize with memory checks\n");
1490 DEBUG(dbgs() << "LAA: No unsafe dependent memory operations in loop. We"
1491 << (NeedRTCheck ? "" : " don't")
1492 << " need a runtime memory check.\n");
1494 emitAnalysis(LoopAccessReport() <<
1495 "unsafe dependent memory operations in loop");
1496 DEBUG(dbgs() << "LAA: unsafe dependent memory operations in loop\n");
1500 bool LoopAccessInfo::blockNeedsPredication(BasicBlock *BB, Loop *TheLoop,
1501 DominatorTree *DT) {
1502 assert(TheLoop->contains(BB) && "Unknown block used");
1504 // Blocks that do not dominate the latch need predication.
1505 BasicBlock* Latch = TheLoop->getLoopLatch();
1506 return !DT->dominates(BB, Latch);
1509 void LoopAccessInfo::emitAnalysis(LoopAccessReport &Message) {
1510 assert(!Report && "Multiple reports generated");
1514 bool LoopAccessInfo::isUniform(Value *V) const {
1515 return (SE->isLoopInvariant(SE->getSCEV(V), TheLoop));
1518 // FIXME: this function is currently a duplicate of the one in
1519 // LoopVectorize.cpp.
1520 static Instruction *getFirstInst(Instruction *FirstInst, Value *V,
1524 if (Instruction *I = dyn_cast<Instruction>(V))
1525 return I->getParent() == Loc->getParent() ? I : nullptr;
1529 std::pair<Instruction *, Instruction *> LoopAccessInfo::addRuntimeCheck(
1530 Instruction *Loc, const SmallVectorImpl<int> *PtrPartition) const {
1531 if (!PtrRtCheck.Need)
1532 return std::make_pair(nullptr, nullptr);
1534 SmallVector<TrackingVH<Value>, 2> Starts;
1535 SmallVector<TrackingVH<Value>, 2> Ends;
1537 LLVMContext &Ctx = Loc->getContext();
1538 SCEVExpander Exp(*SE, DL, "induction");
1539 Instruction *FirstInst = nullptr;
1541 for (unsigned i = 0; i < PtrRtCheck.CheckingGroups.size(); ++i) {
1542 const RuntimePointerCheck::CheckingPtrGroup &CG =
1543 PtrRtCheck.CheckingGroups[i];
1544 Value *Ptr = PtrRtCheck.Pointers[CG.Members[0]];
1545 const SCEV *Sc = SE->getSCEV(Ptr);
1547 if (SE->isLoopInvariant(Sc, TheLoop)) {
1548 DEBUG(dbgs() << "LAA: Adding RT check for a loop invariant ptr:" << *Ptr
1550 Starts.push_back(Ptr);
1551 Ends.push_back(Ptr);
1553 unsigned AS = Ptr->getType()->getPointerAddressSpace();
1555 // Use this type for pointer arithmetic.
1556 Type *PtrArithTy = Type::getInt8PtrTy(Ctx, AS);
1557 Value *Start = nullptr, *End = nullptr;
1559 DEBUG(dbgs() << "LAA: Adding RT check for range:\n");
1560 Start = Exp.expandCodeFor(CG.Low, PtrArithTy, Loc);
1561 End = Exp.expandCodeFor(CG.High, PtrArithTy, Loc);
1562 DEBUG(dbgs() << "Start: " << *CG.Low << " End: " << *CG.High << "\n");
1563 Starts.push_back(Start);
1564 Ends.push_back(End);
1568 IRBuilder<> ChkBuilder(Loc);
1569 // Our instructions might fold to a constant.
1570 Value *MemoryRuntimeCheck = nullptr;
1571 for (unsigned i = 0; i < PtrRtCheck.CheckingGroups.size(); ++i) {
1572 for (unsigned j = i + 1; j < PtrRtCheck.CheckingGroups.size(); ++j) {
1573 const RuntimePointerCheck::CheckingPtrGroup &CGI =
1574 PtrRtCheck.CheckingGroups[i];
1575 const RuntimePointerCheck::CheckingPtrGroup &CGJ =
1576 PtrRtCheck.CheckingGroups[j];
1578 if (!PtrRtCheck.needsChecking(CGI, CGJ, PtrPartition))
1581 unsigned AS0 = Starts[i]->getType()->getPointerAddressSpace();
1582 unsigned AS1 = Starts[j]->getType()->getPointerAddressSpace();
1584 assert((AS0 == Ends[j]->getType()->getPointerAddressSpace()) &&
1585 (AS1 == Ends[i]->getType()->getPointerAddressSpace()) &&
1586 "Trying to bounds check pointers with different address spaces");
1588 Type *PtrArithTy0 = Type::getInt8PtrTy(Ctx, AS0);
1589 Type *PtrArithTy1 = Type::getInt8PtrTy(Ctx, AS1);
1591 Value *Start0 = ChkBuilder.CreateBitCast(Starts[i], PtrArithTy0, "bc");
1592 Value *Start1 = ChkBuilder.CreateBitCast(Starts[j], PtrArithTy1, "bc");
1593 Value *End0 = ChkBuilder.CreateBitCast(Ends[i], PtrArithTy1, "bc");
1594 Value *End1 = ChkBuilder.CreateBitCast(Ends[j], PtrArithTy0, "bc");
1596 Value *Cmp0 = ChkBuilder.CreateICmpULE(Start0, End1, "bound0");
1597 FirstInst = getFirstInst(FirstInst, Cmp0, Loc);
1598 Value *Cmp1 = ChkBuilder.CreateICmpULE(Start1, End0, "bound1");
1599 FirstInst = getFirstInst(FirstInst, Cmp1, Loc);
1600 Value *IsConflict = ChkBuilder.CreateAnd(Cmp0, Cmp1, "found.conflict");
1601 FirstInst = getFirstInst(FirstInst, IsConflict, Loc);
1602 if (MemoryRuntimeCheck) {
1603 IsConflict = ChkBuilder.CreateOr(MemoryRuntimeCheck, IsConflict,
1605 FirstInst = getFirstInst(FirstInst, IsConflict, Loc);
1607 MemoryRuntimeCheck = IsConflict;
1611 if (!MemoryRuntimeCheck)
1612 return std::make_pair(nullptr, nullptr);
1614 // We have to do this trickery because the IRBuilder might fold the check to a
1615 // constant expression in which case there is no Instruction anchored in a
1617 Instruction *Check = BinaryOperator::CreateAnd(MemoryRuntimeCheck,
1618 ConstantInt::getTrue(Ctx));
1619 ChkBuilder.Insert(Check, "memcheck.conflict");
1620 FirstInst = getFirstInst(FirstInst, Check, Loc);
1621 return std::make_pair(FirstInst, Check);
1624 LoopAccessInfo::LoopAccessInfo(Loop *L, ScalarEvolution *SE,
1625 const DataLayout &DL,
1626 const TargetLibraryInfo *TLI, AliasAnalysis *AA,
1627 DominatorTree *DT, LoopInfo *LI,
1628 const ValueToValueMap &Strides)
1629 : PtrRtCheck(SE), DepChecker(SE, L), TheLoop(L), SE(SE), DL(DL), TLI(TLI),
1630 AA(AA), DT(DT), LI(LI), NumLoads(0), NumStores(0),
1631 MaxSafeDepDistBytes(-1U), CanVecMem(false),
1632 StoreToLoopInvariantAddress(false) {
1633 if (canAnalyzeLoop())
1634 analyzeLoop(Strides);
1637 void LoopAccessInfo::print(raw_ostream &OS, unsigned Depth) const {
1639 if (PtrRtCheck.Need)
1640 OS.indent(Depth) << "Memory dependences are safe with run-time checks\n";
1642 OS.indent(Depth) << "Memory dependences are safe\n";
1646 OS.indent(Depth) << "Report: " << Report->str() << "\n";
1648 if (auto *InterestingDependences = DepChecker.getInterestingDependences()) {
1649 OS.indent(Depth) << "Interesting Dependences:\n";
1650 for (auto &Dep : *InterestingDependences) {
1651 Dep.print(OS, Depth + 2, DepChecker.getMemoryInstructions());
1655 OS.indent(Depth) << "Too many interesting dependences, not recorded\n";
1657 // List the pair of accesses need run-time checks to prove independence.
1658 PtrRtCheck.print(OS, Depth);
1661 OS.indent(Depth) << "Store to invariant address was "
1662 << (StoreToLoopInvariantAddress ? "" : "not ")
1663 << "found in loop.\n";
1666 const LoopAccessInfo &
1667 LoopAccessAnalysis::getInfo(Loop *L, const ValueToValueMap &Strides) {
1668 auto &LAI = LoopAccessInfoMap[L];
1671 assert((!LAI || LAI->NumSymbolicStrides == Strides.size()) &&
1672 "Symbolic strides changed for loop");
1676 const DataLayout &DL = L->getHeader()->getModule()->getDataLayout();
1677 LAI = llvm::make_unique<LoopAccessInfo>(L, SE, DL, TLI, AA, DT, LI,
1680 LAI->NumSymbolicStrides = Strides.size();
1686 void LoopAccessAnalysis::print(raw_ostream &OS, const Module *M) const {
1687 LoopAccessAnalysis &LAA = *const_cast<LoopAccessAnalysis *>(this);
1689 ValueToValueMap NoSymbolicStrides;
1691 for (Loop *TopLevelLoop : *LI)
1692 for (Loop *L : depth_first(TopLevelLoop)) {
1693 OS.indent(2) << L->getHeader()->getName() << ":\n";
1694 auto &LAI = LAA.getInfo(L, NoSymbolicStrides);
1699 bool LoopAccessAnalysis::runOnFunction(Function &F) {
1700 SE = &getAnalysis<ScalarEvolution>();
1701 auto *TLIP = getAnalysisIfAvailable<TargetLibraryInfoWrapperPass>();
1702 TLI = TLIP ? &TLIP->getTLI() : nullptr;
1703 AA = &getAnalysis<AliasAnalysis>();
1704 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1705 LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
1710 void LoopAccessAnalysis::getAnalysisUsage(AnalysisUsage &AU) const {
1711 AU.addRequired<ScalarEvolution>();
1712 AU.addRequired<AliasAnalysis>();
1713 AU.addRequired<DominatorTreeWrapperPass>();
1714 AU.addRequired<LoopInfoWrapperPass>();
1716 AU.setPreservesAll();
1719 char LoopAccessAnalysis::ID = 0;
1720 static const char laa_name[] = "Loop Access Analysis";
1721 #define LAA_NAME "loop-accesses"
1723 INITIALIZE_PASS_BEGIN(LoopAccessAnalysis, LAA_NAME, laa_name, false, true)
1724 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
1725 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
1726 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
1727 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
1728 INITIALIZE_PASS_END(LoopAccessAnalysis, LAA_NAME, laa_name, false, true)
1731 Pass *createLAAPass() {
1732 return new LoopAccessAnalysis();