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/Transforms/Utils/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 /// Maximum SIMD width.
52 const unsigned VectorizerParams::MaxVectorWidth = 64;
54 /// \brief We collect interesting dependences up to this threshold.
55 static cl::opt<unsigned> MaxInterestingDependence(
56 "max-interesting-dependences", cl::Hidden,
57 cl::desc("Maximum number of interesting dependences collected by "
58 "loop-access analysis (default = 100)"),
61 bool VectorizerParams::isInterleaveForced() {
62 return ::VectorizationInterleave.getNumOccurrences() > 0;
65 void LoopAccessReport::emitAnalysis(const LoopAccessReport &Message,
66 const Function *TheFunction,
68 const char *PassName) {
69 DebugLoc DL = TheLoop->getStartLoc();
70 if (const Instruction *I = Message.getInstr())
71 DL = I->getDebugLoc();
72 emitOptimizationRemarkAnalysis(TheFunction->getContext(), PassName,
73 *TheFunction, DL, Message.str());
76 Value *llvm::stripIntegerCast(Value *V) {
77 if (CastInst *CI = dyn_cast<CastInst>(V))
78 if (CI->getOperand(0)->getType()->isIntegerTy())
79 return CI->getOperand(0);
83 const SCEV *llvm::replaceSymbolicStrideSCEV(ScalarEvolution *SE,
84 const ValueToValueMap &PtrToStride,
85 Value *Ptr, Value *OrigPtr) {
87 const SCEV *OrigSCEV = SE->getSCEV(Ptr);
89 // If there is an entry in the map return the SCEV of the pointer with the
90 // symbolic stride replaced by one.
91 ValueToValueMap::const_iterator SI =
92 PtrToStride.find(OrigPtr ? OrigPtr : Ptr);
93 if (SI != PtrToStride.end()) {
94 Value *StrideVal = SI->second;
97 StrideVal = stripIntegerCast(StrideVal);
99 // Replace symbolic stride by one.
100 Value *One = ConstantInt::get(StrideVal->getType(), 1);
101 ValueToValueMap RewriteMap;
102 RewriteMap[StrideVal] = One;
105 SCEVParameterRewriter::rewrite(OrigSCEV, *SE, RewriteMap, true);
106 DEBUG(dbgs() << "LAA: Replacing SCEV: " << *OrigSCEV << " by: " << *ByOne
111 // Otherwise, just return the SCEV of the original pointer.
112 return SE->getSCEV(Ptr);
115 void LoopAccessInfo::RuntimePointerCheck::insert(
116 ScalarEvolution *SE, Loop *Lp, Value *Ptr, bool WritePtr, unsigned DepSetId,
117 unsigned ASId, const ValueToValueMap &Strides) {
118 // Get the stride replaced scev.
119 const SCEV *Sc = replaceSymbolicStrideSCEV(SE, Strides, Ptr);
120 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Sc);
121 assert(AR && "Invalid addrec expression");
122 const SCEV *Ex = SE->getBackedgeTakenCount(Lp);
123 const SCEV *ScEnd = AR->evaluateAtIteration(Ex, *SE);
124 Pointers.push_back(Ptr);
125 Starts.push_back(AR->getStart());
126 Ends.push_back(ScEnd);
127 IsWritePtr.push_back(WritePtr);
128 DependencySetId.push_back(DepSetId);
129 AliasSetId.push_back(ASId);
132 bool LoopAccessInfo::RuntimePointerCheck::needsChecking(
133 unsigned I, unsigned J, const SmallVectorImpl<int> *PtrPartition) const {
134 // No need to check if two readonly pointers intersect.
135 if (!IsWritePtr[I] && !IsWritePtr[J])
138 // Only need to check pointers between two different dependency sets.
139 if (DependencySetId[I] == DependencySetId[J])
142 // Only need to check pointers in the same alias set.
143 if (AliasSetId[I] != AliasSetId[J])
146 // If PtrPartition is set omit checks between pointers of the same partition.
147 // Partition number -1 means that the pointer is used in multiple partitions.
148 // In this case we can't omit the check.
149 if (PtrPartition && (*PtrPartition)[I] != -1 &&
150 (*PtrPartition)[I] == (*PtrPartition)[J])
156 void LoopAccessInfo::RuntimePointerCheck::print(
157 raw_ostream &OS, unsigned Depth,
158 const SmallVectorImpl<int> *PtrPartition) const {
159 unsigned NumPointers = Pointers.size();
160 if (NumPointers == 0)
163 OS.indent(Depth) << "Run-time memory checks:\n";
165 for (unsigned I = 0; I < NumPointers; ++I)
166 for (unsigned J = I + 1; J < NumPointers; ++J)
167 if (needsChecking(I, J, PtrPartition)) {
168 OS.indent(Depth) << N++ << ":\n";
169 OS.indent(Depth + 2) << *Pointers[I];
171 OS << " (Partition: " << (*PtrPartition)[I] << ")";
173 OS.indent(Depth + 2) << *Pointers[J];
175 OS << " (Partition: " << (*PtrPartition)[J] << ")";
180 bool LoopAccessInfo::RuntimePointerCheck::needsAnyChecking(
181 const SmallVectorImpl<int> *PtrPartition) const {
182 unsigned NumPointers = Pointers.size();
184 for (unsigned I = 0; I < NumPointers; ++I)
185 for (unsigned J = I + 1; J < NumPointers; ++J)
186 if (needsChecking(I, J, PtrPartition))
192 /// \brief Analyses memory accesses in a loop.
194 /// Checks whether run time pointer checks are needed and builds sets for data
195 /// dependence checking.
196 class AccessAnalysis {
198 /// \brief Read or write access location.
199 typedef PointerIntPair<Value *, 1, bool> MemAccessInfo;
200 typedef SmallPtrSet<MemAccessInfo, 8> MemAccessInfoSet;
202 AccessAnalysis(const DataLayout &Dl, AliasAnalysis *AA, LoopInfo *LI,
203 MemoryDepChecker::DepCandidates &DA)
204 : DL(Dl), AST(*AA), LI(LI), DepCands(DA), IsRTCheckNeeded(false) {}
206 /// \brief Register a load and whether it is only read from.
207 void addLoad(AliasAnalysis::Location &Loc, bool IsReadOnly) {
208 Value *Ptr = const_cast<Value*>(Loc.Ptr);
209 AST.add(Ptr, AliasAnalysis::UnknownSize, Loc.AATags);
210 Accesses.insert(MemAccessInfo(Ptr, false));
212 ReadOnlyPtr.insert(Ptr);
215 /// \brief Register a store.
216 void addStore(AliasAnalysis::Location &Loc) {
217 Value *Ptr = const_cast<Value*>(Loc.Ptr);
218 AST.add(Ptr, AliasAnalysis::UnknownSize, Loc.AATags);
219 Accesses.insert(MemAccessInfo(Ptr, true));
222 /// \brief Check whether we can check the pointers at runtime for
223 /// non-intersection.
224 bool canCheckPtrAtRT(LoopAccessInfo::RuntimePointerCheck &RtCheck,
225 unsigned &NumComparisons, ScalarEvolution *SE,
226 Loop *TheLoop, const ValueToValueMap &Strides,
227 bool ShouldCheckStride = false);
229 /// \brief Goes over all memory accesses, checks whether a RT check is needed
230 /// and builds sets of dependent accesses.
231 void buildDependenceSets() {
232 processMemAccesses();
235 bool isRTCheckNeeded() { return IsRTCheckNeeded; }
237 bool isDependencyCheckNeeded() { return !CheckDeps.empty(); }
238 void resetDepChecks() { CheckDeps.clear(); }
240 MemAccessInfoSet &getDependenciesToCheck() { return CheckDeps; }
243 typedef SetVector<MemAccessInfo> PtrAccessSet;
245 /// \brief Go over all memory access and check whether runtime pointer checks
246 /// are needed /// and build sets of dependency check candidates.
247 void processMemAccesses();
249 /// Set of all accesses.
250 PtrAccessSet Accesses;
252 const DataLayout &DL;
254 /// Set of accesses that need a further dependence check.
255 MemAccessInfoSet CheckDeps;
257 /// Set of pointers that are read only.
258 SmallPtrSet<Value*, 16> ReadOnlyPtr;
260 /// An alias set tracker to partition the access set by underlying object and
261 //intrinsic property (such as TBAA metadata).
266 /// Sets of potentially dependent accesses - members of one set share an
267 /// underlying pointer. The set "CheckDeps" identfies which sets really need a
268 /// dependence check.
269 MemoryDepChecker::DepCandidates &DepCands;
271 bool IsRTCheckNeeded;
274 } // end anonymous namespace
276 /// \brief Check whether a pointer can participate in a runtime bounds check.
277 static bool hasComputableBounds(ScalarEvolution *SE,
278 const ValueToValueMap &Strides, Value *Ptr) {
279 const SCEV *PtrScev = replaceSymbolicStrideSCEV(SE, Strides, Ptr);
280 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
284 return AR->isAffine();
287 /// \brief Check the stride of the pointer and ensure that it does not wrap in
288 /// the address space.
289 static int isStridedPtr(ScalarEvolution *SE, Value *Ptr, const Loop *Lp,
290 const ValueToValueMap &StridesMap);
292 bool AccessAnalysis::canCheckPtrAtRT(
293 LoopAccessInfo::RuntimePointerCheck &RtCheck, unsigned &NumComparisons,
294 ScalarEvolution *SE, Loop *TheLoop, const ValueToValueMap &StridesMap,
295 bool ShouldCheckStride) {
296 // Find pointers with computable bounds. We are going to use this information
297 // to place a runtime bound check.
300 bool IsDepCheckNeeded = isDependencyCheckNeeded();
303 // We assign a consecutive id to access from different alias sets.
304 // Accesses between different groups doesn't need to be checked.
306 for (auto &AS : AST) {
307 unsigned NumReadPtrChecks = 0;
308 unsigned NumWritePtrChecks = 0;
310 // We assign consecutive id to access from different dependence sets.
311 // Accesses within the same set don't need a runtime check.
312 unsigned RunningDepId = 1;
313 DenseMap<Value *, unsigned> DepSetId;
316 Value *Ptr = A.getValue();
317 bool IsWrite = Accesses.count(MemAccessInfo(Ptr, true));
318 MemAccessInfo Access(Ptr, IsWrite);
325 if (hasComputableBounds(SE, StridesMap, Ptr) &&
326 // When we run after a failing dependency check we have to make sure
327 // we don't have wrapping pointers.
328 (!ShouldCheckStride ||
329 isStridedPtr(SE, Ptr, TheLoop, StridesMap) == 1)) {
330 // The id of the dependence set.
333 if (IsDepCheckNeeded) {
334 Value *Leader = DepCands.getLeaderValue(Access).getPointer();
335 unsigned &LeaderId = DepSetId[Leader];
337 LeaderId = RunningDepId++;
340 // Each access has its own dependence set.
341 DepId = RunningDepId++;
343 RtCheck.insert(SE, TheLoop, Ptr, IsWrite, DepId, ASId, StridesMap);
345 DEBUG(dbgs() << "LAA: Found a runtime check ptr:" << *Ptr << '\n');
347 DEBUG(dbgs() << "LAA: Can't find bounds for ptr:" << *Ptr << '\n');
352 if (IsDepCheckNeeded && CanDoRT && RunningDepId == 2)
353 NumComparisons += 0; // Only one dependence set.
355 NumComparisons += (NumWritePtrChecks * (NumReadPtrChecks +
356 NumWritePtrChecks - 1));
362 // If the pointers that we would use for the bounds comparison have different
363 // address spaces, assume the values aren't directly comparable, so we can't
364 // use them for the runtime check. We also have to assume they could
365 // overlap. In the future there should be metadata for whether address spaces
367 unsigned NumPointers = RtCheck.Pointers.size();
368 for (unsigned i = 0; i < NumPointers; ++i) {
369 for (unsigned j = i + 1; j < NumPointers; ++j) {
370 // Only need to check pointers between two different dependency sets.
371 if (RtCheck.DependencySetId[i] == RtCheck.DependencySetId[j])
373 // Only need to check pointers in the same alias set.
374 if (RtCheck.AliasSetId[i] != RtCheck.AliasSetId[j])
377 Value *PtrI = RtCheck.Pointers[i];
378 Value *PtrJ = RtCheck.Pointers[j];
380 unsigned ASi = PtrI->getType()->getPointerAddressSpace();
381 unsigned ASj = PtrJ->getType()->getPointerAddressSpace();
383 DEBUG(dbgs() << "LAA: Runtime check would require comparison between"
384 " different address spaces\n");
393 void AccessAnalysis::processMemAccesses() {
394 // We process the set twice: first we process read-write pointers, last we
395 // process read-only pointers. This allows us to skip dependence tests for
396 // read-only pointers.
398 DEBUG(dbgs() << "LAA: Processing memory accesses...\n");
399 DEBUG(dbgs() << " AST: "; AST.dump());
400 DEBUG(dbgs() << "LAA: Accesses(" << Accesses.size() << "):\n");
402 for (auto A : Accesses)
403 dbgs() << "\t" << *A.getPointer() << " (" <<
404 (A.getInt() ? "write" : (ReadOnlyPtr.count(A.getPointer()) ?
405 "read-only" : "read")) << ")\n";
408 // The AliasSetTracker has nicely partitioned our pointers by metadata
409 // compatibility and potential for underlying-object overlap. As a result, we
410 // only need to check for potential pointer dependencies within each alias
412 for (auto &AS : AST) {
413 // Note that both the alias-set tracker and the alias sets themselves used
414 // linked lists internally and so the iteration order here is deterministic
415 // (matching the original instruction order within each set).
417 bool SetHasWrite = false;
419 // Map of pointers to last access encountered.
420 typedef DenseMap<Value*, MemAccessInfo> UnderlyingObjToAccessMap;
421 UnderlyingObjToAccessMap ObjToLastAccess;
423 // Set of access to check after all writes have been processed.
424 PtrAccessSet DeferredAccesses;
426 // Iterate over each alias set twice, once to process read/write pointers,
427 // and then to process read-only pointers.
428 for (int SetIteration = 0; SetIteration < 2; ++SetIteration) {
429 bool UseDeferred = SetIteration > 0;
430 PtrAccessSet &S = UseDeferred ? DeferredAccesses : Accesses;
433 Value *Ptr = AV.getValue();
435 // For a single memory access in AliasSetTracker, Accesses may contain
436 // both read and write, and they both need to be handled for CheckDeps.
438 if (AC.getPointer() != Ptr)
441 bool IsWrite = AC.getInt();
443 // If we're using the deferred access set, then it contains only
445 bool IsReadOnlyPtr = ReadOnlyPtr.count(Ptr) && !IsWrite;
446 if (UseDeferred && !IsReadOnlyPtr)
448 // Otherwise, the pointer must be in the PtrAccessSet, either as a
450 assert(((IsReadOnlyPtr && UseDeferred) || IsWrite ||
451 S.count(MemAccessInfo(Ptr, false))) &&
452 "Alias-set pointer not in the access set?");
454 MemAccessInfo Access(Ptr, IsWrite);
455 DepCands.insert(Access);
457 // Memorize read-only pointers for later processing and skip them in
458 // the first round (they need to be checked after we have seen all
459 // write pointers). Note: we also mark pointer that are not
460 // consecutive as "read-only" pointers (so that we check
461 // "a[b[i]] +="). Hence, we need the second check for "!IsWrite".
462 if (!UseDeferred && IsReadOnlyPtr) {
463 DeferredAccesses.insert(Access);
467 // If this is a write - check other reads and writes for conflicts. If
468 // this is a read only check other writes for conflicts (but only if
469 // there is no other write to the ptr - this is an optimization to
470 // catch "a[i] = a[i] + " without having to do a dependence check).
471 if ((IsWrite || IsReadOnlyPtr) && SetHasWrite) {
472 CheckDeps.insert(Access);
473 IsRTCheckNeeded = true;
479 // Create sets of pointers connected by a shared alias set and
480 // underlying object.
481 typedef SmallVector<Value *, 16> ValueVector;
482 ValueVector TempObjects;
484 GetUnderlyingObjects(Ptr, TempObjects, DL, LI);
485 DEBUG(dbgs() << "Underlying objects for pointer " << *Ptr << "\n");
486 for (Value *UnderlyingObj : TempObjects) {
487 UnderlyingObjToAccessMap::iterator Prev =
488 ObjToLastAccess.find(UnderlyingObj);
489 if (Prev != ObjToLastAccess.end())
490 DepCands.unionSets(Access, Prev->second);
492 ObjToLastAccess[UnderlyingObj] = Access;
493 DEBUG(dbgs() << " " << *UnderlyingObj << "\n");
501 static bool isInBoundsGep(Value *Ptr) {
502 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr))
503 return GEP->isInBounds();
507 /// \brief Check whether the access through \p Ptr has a constant stride.
508 static int isStridedPtr(ScalarEvolution *SE, Value *Ptr, const Loop *Lp,
509 const ValueToValueMap &StridesMap) {
510 const Type *Ty = Ptr->getType();
511 assert(Ty->isPointerTy() && "Unexpected non-ptr");
513 // Make sure that the pointer does not point to aggregate types.
514 const PointerType *PtrTy = cast<PointerType>(Ty);
515 if (PtrTy->getElementType()->isAggregateType()) {
516 DEBUG(dbgs() << "LAA: Bad stride - Not a pointer to a scalar type"
521 const SCEV *PtrScev = replaceSymbolicStrideSCEV(SE, StridesMap, Ptr);
523 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
525 DEBUG(dbgs() << "LAA: Bad stride - Not an AddRecExpr pointer "
526 << *Ptr << " SCEV: " << *PtrScev << "\n");
530 // The accesss function must stride over the innermost loop.
531 if (Lp != AR->getLoop()) {
532 DEBUG(dbgs() << "LAA: Bad stride - Not striding over innermost loop " <<
533 *Ptr << " SCEV: " << *PtrScev << "\n");
536 // The address calculation must not wrap. Otherwise, a dependence could be
538 // An inbounds getelementptr that is a AddRec with a unit stride
539 // cannot wrap per definition. The unit stride requirement is checked later.
540 // An getelementptr without an inbounds attribute and unit stride would have
541 // to access the pointer value "0" which is undefined behavior in address
542 // space 0, therefore we can also vectorize this case.
543 bool IsInBoundsGEP = isInBoundsGep(Ptr);
544 bool IsNoWrapAddRec = AR->getNoWrapFlags(SCEV::NoWrapMask);
545 bool IsInAddressSpaceZero = PtrTy->getAddressSpace() == 0;
546 if (!IsNoWrapAddRec && !IsInBoundsGEP && !IsInAddressSpaceZero) {
547 DEBUG(dbgs() << "LAA: Bad stride - Pointer may wrap in the address space "
548 << *Ptr << " SCEV: " << *PtrScev << "\n");
552 // Check the step is constant.
553 const SCEV *Step = AR->getStepRecurrence(*SE);
555 // Calculate the pointer stride and check if it is consecutive.
556 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
558 DEBUG(dbgs() << "LAA: Bad stride - Not a constant strided " << *Ptr <<
559 " SCEV: " << *PtrScev << "\n");
563 auto &DL = Lp->getHeader()->getModule()->getDataLayout();
564 int64_t Size = DL.getTypeAllocSize(PtrTy->getElementType());
565 const APInt &APStepVal = C->getValue()->getValue();
567 // Huge step value - give up.
568 if (APStepVal.getBitWidth() > 64)
571 int64_t StepVal = APStepVal.getSExtValue();
574 int64_t Stride = StepVal / Size;
575 int64_t Rem = StepVal % Size;
579 // If the SCEV could wrap but we have an inbounds gep with a unit stride we
580 // know we can't "wrap around the address space". In case of address space
581 // zero we know that this won't happen without triggering undefined behavior.
582 if (!IsNoWrapAddRec && (IsInBoundsGEP || IsInAddressSpaceZero) &&
583 Stride != 1 && Stride != -1)
589 bool MemoryDepChecker::Dependence::isSafeForVectorization(DepType Type) {
593 case BackwardVectorizable:
597 case ForwardButPreventsForwarding:
599 case BackwardVectorizableButPreventsForwarding:
602 llvm_unreachable("unexpected DepType!");
605 bool MemoryDepChecker::Dependence::isInterestingDependence(DepType Type) {
611 case BackwardVectorizable:
613 case ForwardButPreventsForwarding:
615 case BackwardVectorizableButPreventsForwarding:
618 llvm_unreachable("unexpected DepType!");
621 bool MemoryDepChecker::Dependence::isPossiblyBackward() const {
625 case ForwardButPreventsForwarding:
629 case BackwardVectorizable:
631 case BackwardVectorizableButPreventsForwarding:
634 llvm_unreachable("unexpected DepType!");
637 bool MemoryDepChecker::couldPreventStoreLoadForward(unsigned Distance,
638 unsigned TypeByteSize) {
639 // If loads occur at a distance that is not a multiple of a feasible vector
640 // factor store-load forwarding does not take place.
641 // Positive dependences might cause troubles because vectorizing them might
642 // prevent store-load forwarding making vectorized code run a lot slower.
643 // a[i] = a[i-3] ^ a[i-8];
644 // The stores to a[i:i+1] don't align with the stores to a[i-3:i-2] and
645 // hence on your typical architecture store-load forwarding does not take
646 // place. Vectorizing in such cases does not make sense.
647 // Store-load forwarding distance.
648 const unsigned NumCyclesForStoreLoadThroughMemory = 8*TypeByteSize;
649 // Maximum vector factor.
650 unsigned MaxVFWithoutSLForwardIssues =
651 VectorizerParams::MaxVectorWidth * TypeByteSize;
652 if(MaxSafeDepDistBytes < MaxVFWithoutSLForwardIssues)
653 MaxVFWithoutSLForwardIssues = MaxSafeDepDistBytes;
655 for (unsigned vf = 2*TypeByteSize; vf <= MaxVFWithoutSLForwardIssues;
657 if (Distance % vf && Distance / vf < NumCyclesForStoreLoadThroughMemory) {
658 MaxVFWithoutSLForwardIssues = (vf >>=1);
663 if (MaxVFWithoutSLForwardIssues< 2*TypeByteSize) {
664 DEBUG(dbgs() << "LAA: Distance " << Distance <<
665 " that could cause a store-load forwarding conflict\n");
669 if (MaxVFWithoutSLForwardIssues < MaxSafeDepDistBytes &&
670 MaxVFWithoutSLForwardIssues !=
671 VectorizerParams::MaxVectorWidth * TypeByteSize)
672 MaxSafeDepDistBytes = MaxVFWithoutSLForwardIssues;
676 MemoryDepChecker::Dependence::DepType
677 MemoryDepChecker::isDependent(const MemAccessInfo &A, unsigned AIdx,
678 const MemAccessInfo &B, unsigned BIdx,
679 const ValueToValueMap &Strides) {
680 assert (AIdx < BIdx && "Must pass arguments in program order");
682 Value *APtr = A.getPointer();
683 Value *BPtr = B.getPointer();
684 bool AIsWrite = A.getInt();
685 bool BIsWrite = B.getInt();
687 // Two reads are independent.
688 if (!AIsWrite && !BIsWrite)
689 return Dependence::NoDep;
691 // We cannot check pointers in different address spaces.
692 if (APtr->getType()->getPointerAddressSpace() !=
693 BPtr->getType()->getPointerAddressSpace())
694 return Dependence::Unknown;
696 const SCEV *AScev = replaceSymbolicStrideSCEV(SE, Strides, APtr);
697 const SCEV *BScev = replaceSymbolicStrideSCEV(SE, Strides, BPtr);
699 int StrideAPtr = isStridedPtr(SE, APtr, InnermostLoop, Strides);
700 int StrideBPtr = isStridedPtr(SE, BPtr, InnermostLoop, Strides);
702 const SCEV *Src = AScev;
703 const SCEV *Sink = BScev;
705 // If the induction step is negative we have to invert source and sink of the
707 if (StrideAPtr < 0) {
710 std::swap(APtr, BPtr);
711 std::swap(Src, Sink);
712 std::swap(AIsWrite, BIsWrite);
713 std::swap(AIdx, BIdx);
714 std::swap(StrideAPtr, StrideBPtr);
717 const SCEV *Dist = SE->getMinusSCEV(Sink, Src);
719 DEBUG(dbgs() << "LAA: Src Scev: " << *Src << "Sink Scev: " << *Sink
720 << "(Induction step: " << StrideAPtr << ")\n");
721 DEBUG(dbgs() << "LAA: Distance for " << *InstMap[AIdx] << " to "
722 << *InstMap[BIdx] << ": " << *Dist << "\n");
724 // Need consecutive accesses. We don't want to vectorize
725 // "A[B[i]] += ..." and similar code or pointer arithmetic that could wrap in
726 // the address space.
727 if (!StrideAPtr || !StrideBPtr || StrideAPtr != StrideBPtr){
728 DEBUG(dbgs() << "Non-consecutive pointer access\n");
729 return Dependence::Unknown;
732 const SCEVConstant *C = dyn_cast<SCEVConstant>(Dist);
734 DEBUG(dbgs() << "LAA: Dependence because of non-constant distance\n");
735 ShouldRetryWithRuntimeCheck = true;
736 return Dependence::Unknown;
739 Type *ATy = APtr->getType()->getPointerElementType();
740 Type *BTy = BPtr->getType()->getPointerElementType();
741 auto &DL = InnermostLoop->getHeader()->getModule()->getDataLayout();
742 unsigned TypeByteSize = DL.getTypeAllocSize(ATy);
744 // Negative distances are not plausible dependencies.
745 const APInt &Val = C->getValue()->getValue();
746 if (Val.isNegative()) {
747 bool IsTrueDataDependence = (AIsWrite && !BIsWrite);
748 if (IsTrueDataDependence &&
749 (couldPreventStoreLoadForward(Val.abs().getZExtValue(), TypeByteSize) ||
751 return Dependence::ForwardButPreventsForwarding;
753 DEBUG(dbgs() << "LAA: Dependence is negative: NoDep\n");
754 return Dependence::Forward;
757 // Write to the same location with the same size.
758 // Could be improved to assert type sizes are the same (i32 == float, etc).
761 return Dependence::NoDep;
762 DEBUG(dbgs() << "LAA: Zero dependence difference but different types\n");
763 return Dependence::Unknown;
766 assert(Val.isStrictlyPositive() && "Expect a positive value");
770 "LAA: ReadWrite-Write positive dependency with different types\n");
771 return Dependence::Unknown;
774 unsigned Distance = (unsigned) Val.getZExtValue();
776 // Bail out early if passed-in parameters make vectorization not feasible.
777 unsigned ForcedFactor = (VectorizerParams::VectorizationFactor ?
778 VectorizerParams::VectorizationFactor : 1);
779 unsigned ForcedUnroll = (VectorizerParams::VectorizationInterleave ?
780 VectorizerParams::VectorizationInterleave : 1);
782 // The distance must be bigger than the size needed for a vectorized version
783 // of the operation and the size of the vectorized operation must not be
784 // bigger than the currrent maximum size.
785 if (Distance < 2*TypeByteSize ||
786 2*TypeByteSize > MaxSafeDepDistBytes ||
787 Distance < TypeByteSize * ForcedUnroll * ForcedFactor) {
788 DEBUG(dbgs() << "LAA: Failure because of Positive distance "
789 << Val.getSExtValue() << '\n');
790 return Dependence::Backward;
793 // Positive distance bigger than max vectorization factor.
794 MaxSafeDepDistBytes = Distance < MaxSafeDepDistBytes ?
795 Distance : MaxSafeDepDistBytes;
797 bool IsTrueDataDependence = (!AIsWrite && BIsWrite);
798 if (IsTrueDataDependence &&
799 couldPreventStoreLoadForward(Distance, TypeByteSize))
800 return Dependence::BackwardVectorizableButPreventsForwarding;
802 DEBUG(dbgs() << "LAA: Positive distance " << Val.getSExtValue() <<
803 " with max VF = " << MaxSafeDepDistBytes / TypeByteSize << '\n');
805 return Dependence::BackwardVectorizable;
808 bool MemoryDepChecker::areDepsSafe(DepCandidates &AccessSets,
809 MemAccessInfoSet &CheckDeps,
810 const ValueToValueMap &Strides) {
812 MaxSafeDepDistBytes = -1U;
813 while (!CheckDeps.empty()) {
814 MemAccessInfo CurAccess = *CheckDeps.begin();
816 // Get the relevant memory access set.
817 EquivalenceClasses<MemAccessInfo>::iterator I =
818 AccessSets.findValue(AccessSets.getLeaderValue(CurAccess));
820 // Check accesses within this set.
821 EquivalenceClasses<MemAccessInfo>::member_iterator AI, AE;
822 AI = AccessSets.member_begin(I), AE = AccessSets.member_end();
824 // Check every access pair.
826 CheckDeps.erase(*AI);
827 EquivalenceClasses<MemAccessInfo>::member_iterator OI = std::next(AI);
829 // Check every accessing instruction pair in program order.
830 for (std::vector<unsigned>::iterator I1 = Accesses[*AI].begin(),
831 I1E = Accesses[*AI].end(); I1 != I1E; ++I1)
832 for (std::vector<unsigned>::iterator I2 = Accesses[*OI].begin(),
833 I2E = Accesses[*OI].end(); I2 != I2E; ++I2) {
834 auto A = std::make_pair(&*AI, *I1);
835 auto B = std::make_pair(&*OI, *I2);
841 Dependence::DepType Type =
842 isDependent(*A.first, A.second, *B.first, B.second, Strides);
843 SafeForVectorization &= Dependence::isSafeForVectorization(Type);
845 // Gather dependences unless we accumulated MaxInterestingDependence
846 // dependences. In that case return as soon as we find the first
847 // unsafe dependence. This puts a limit on this quadratic
849 if (RecordInterestingDependences) {
850 if (Dependence::isInterestingDependence(Type))
851 InterestingDependences.push_back(
852 Dependence(A.second, B.second, Type));
854 if (InterestingDependences.size() >= MaxInterestingDependence) {
855 RecordInterestingDependences = false;
856 InterestingDependences.clear();
857 DEBUG(dbgs() << "Too many dependences, stopped recording\n");
860 if (!RecordInterestingDependences && !SafeForVectorization)
869 DEBUG(dbgs() << "Total Interesting Dependences: "
870 << InterestingDependences.size() << "\n");
871 return SafeForVectorization;
874 SmallVector<Instruction *, 4>
875 MemoryDepChecker::getInstructionsForAccess(Value *Ptr, bool isWrite) const {
876 MemAccessInfo Access(Ptr, isWrite);
877 auto &IndexVector = Accesses.find(Access)->second;
879 SmallVector<Instruction *, 4> Insts;
880 std::transform(IndexVector.begin(), IndexVector.end(),
881 std::back_inserter(Insts),
882 [&](unsigned Idx) { return this->InstMap[Idx]; });
886 const char *MemoryDepChecker::Dependence::DepName[] = {
887 "NoDep", "Unknown", "Forward", "ForwardButPreventsForwarding", "Backward",
888 "BackwardVectorizable", "BackwardVectorizableButPreventsForwarding"};
890 void MemoryDepChecker::Dependence::print(
891 raw_ostream &OS, unsigned Depth,
892 const SmallVectorImpl<Instruction *> &Instrs) const {
893 OS.indent(Depth) << DepName[Type] << ":\n";
894 OS.indent(Depth + 2) << *Instrs[Source] << " -> \n";
895 OS.indent(Depth + 2) << *Instrs[Destination] << "\n";
898 bool LoopAccessInfo::canAnalyzeLoop() {
899 // We need to have a loop header.
900 DEBUG(dbgs() << "LAA: Found a loop: " <<
901 TheLoop->getHeader()->getName() << '\n');
903 // We can only analyze innermost loops.
904 if (!TheLoop->empty()) {
905 DEBUG(dbgs() << "LAA: loop is not the innermost loop\n");
906 emitAnalysis(LoopAccessReport() << "loop is not the innermost loop");
910 // We must have a single backedge.
911 if (TheLoop->getNumBackEdges() != 1) {
912 DEBUG(dbgs() << "LAA: loop control flow is not understood by analyzer\n");
914 LoopAccessReport() <<
915 "loop control flow is not understood by analyzer");
919 // We must have a single exiting block.
920 if (!TheLoop->getExitingBlock()) {
921 DEBUG(dbgs() << "LAA: loop control flow is not understood by analyzer\n");
923 LoopAccessReport() <<
924 "loop control flow is not understood by analyzer");
928 // We only handle bottom-tested loops, i.e. loop in which the condition is
929 // checked at the end of each iteration. With that we can assume that all
930 // instructions in the loop are executed the same number of times.
931 if (TheLoop->getExitingBlock() != TheLoop->getLoopLatch()) {
932 DEBUG(dbgs() << "LAA: loop control flow is not understood by analyzer\n");
934 LoopAccessReport() <<
935 "loop control flow is not understood by analyzer");
939 // ScalarEvolution needs to be able to find the exit count.
940 const SCEV *ExitCount = SE->getBackedgeTakenCount(TheLoop);
941 if (ExitCount == SE->getCouldNotCompute()) {
942 emitAnalysis(LoopAccessReport() <<
943 "could not determine number of loop iterations");
944 DEBUG(dbgs() << "LAA: SCEV could not compute the loop exit count.\n");
951 void LoopAccessInfo::analyzeLoop(const ValueToValueMap &Strides) {
953 typedef SmallVector<Value*, 16> ValueVector;
954 typedef SmallPtrSet<Value*, 16> ValueSet;
956 // Holds the Load and Store *instructions*.
960 // Holds all the different accesses in the loop.
961 unsigned NumReads = 0;
962 unsigned NumReadWrites = 0;
964 PtrRtCheck.Pointers.clear();
965 PtrRtCheck.Need = false;
967 const bool IsAnnotatedParallel = TheLoop->isAnnotatedParallel();
970 for (Loop::block_iterator bb = TheLoop->block_begin(),
971 be = TheLoop->block_end(); bb != be; ++bb) {
973 // Scan the BB and collect legal loads and stores.
974 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
977 // If this is a load, save it. If this instruction can read from memory
978 // but is not a load, then we quit. Notice that we don't handle function
979 // calls that read or write.
980 if (it->mayReadFromMemory()) {
981 // Many math library functions read the rounding mode. We will only
982 // vectorize a loop if it contains known function calls that don't set
983 // the flag. Therefore, it is safe to ignore this read from memory.
984 CallInst *Call = dyn_cast<CallInst>(it);
985 if (Call && getIntrinsicIDForCall(Call, TLI))
988 // If the function has an explicit vectorized counterpart, we can safely
989 // assume that it can be vectorized.
990 if (Call && !Call->isNoBuiltin() && Call->getCalledFunction() &&
991 TLI->isFunctionVectorizable(Call->getCalledFunction()->getName()))
994 LoadInst *Ld = dyn_cast<LoadInst>(it);
995 if (!Ld || (!Ld->isSimple() && !IsAnnotatedParallel)) {
996 emitAnalysis(LoopAccessReport(Ld)
997 << "read with atomic ordering or volatile read");
998 DEBUG(dbgs() << "LAA: Found a non-simple load.\n");
1003 Loads.push_back(Ld);
1004 DepChecker.addAccess(Ld);
1008 // Save 'store' instructions. Abort if other instructions write to memory.
1009 if (it->mayWriteToMemory()) {
1010 StoreInst *St = dyn_cast<StoreInst>(it);
1012 emitAnalysis(LoopAccessReport(it) <<
1013 "instruction cannot be vectorized");
1017 if (!St->isSimple() && !IsAnnotatedParallel) {
1018 emitAnalysis(LoopAccessReport(St)
1019 << "write with atomic ordering or volatile write");
1020 DEBUG(dbgs() << "LAA: Found a non-simple store.\n");
1025 Stores.push_back(St);
1026 DepChecker.addAccess(St);
1031 // Now we have two lists that hold the loads and the stores.
1032 // Next, we find the pointers that they use.
1034 // Check if we see any stores. If there are no stores, then we don't
1035 // care if the pointers are *restrict*.
1036 if (!Stores.size()) {
1037 DEBUG(dbgs() << "LAA: Found a read-only loop!\n");
1042 MemoryDepChecker::DepCandidates DependentAccesses;
1043 AccessAnalysis Accesses(TheLoop->getHeader()->getModule()->getDataLayout(),
1044 AA, LI, DependentAccesses);
1046 // Holds the analyzed pointers. We don't want to call GetUnderlyingObjects
1047 // multiple times on the same object. If the ptr is accessed twice, once
1048 // for read and once for write, it will only appear once (on the write
1049 // list). This is okay, since we are going to check for conflicts between
1050 // writes and between reads and writes, but not between reads and reads.
1053 ValueVector::iterator I, IE;
1054 for (I = Stores.begin(), IE = Stores.end(); I != IE; ++I) {
1055 StoreInst *ST = cast<StoreInst>(*I);
1056 Value* Ptr = ST->getPointerOperand();
1057 // Check for store to loop invariant address.
1058 StoreToLoopInvariantAddress |= isUniform(Ptr);
1059 // If we did *not* see this pointer before, insert it to the read-write
1060 // list. At this phase it is only a 'write' list.
1061 if (Seen.insert(Ptr).second) {
1064 AliasAnalysis::Location Loc = AA->getLocation(ST);
1065 // The TBAA metadata could have a control dependency on the predication
1066 // condition, so we cannot rely on it when determining whether or not we
1067 // need runtime pointer checks.
1068 if (blockNeedsPredication(ST->getParent(), TheLoop, DT))
1069 Loc.AATags.TBAA = nullptr;
1071 Accesses.addStore(Loc);
1075 if (IsAnnotatedParallel) {
1077 << "LAA: A loop annotated parallel, ignore memory dependency "
1083 for (I = Loads.begin(), IE = Loads.end(); I != IE; ++I) {
1084 LoadInst *LD = cast<LoadInst>(*I);
1085 Value* Ptr = LD->getPointerOperand();
1086 // If we did *not* see this pointer before, insert it to the
1087 // read list. If we *did* see it before, then it is already in
1088 // the read-write list. This allows us to vectorize expressions
1089 // such as A[i] += x; Because the address of A[i] is a read-write
1090 // pointer. This only works if the index of A[i] is consecutive.
1091 // If the address of i is unknown (for example A[B[i]]) then we may
1092 // read a few words, modify, and write a few words, and some of the
1093 // words may be written to the same address.
1094 bool IsReadOnlyPtr = false;
1095 if (Seen.insert(Ptr).second || !isStridedPtr(SE, Ptr, TheLoop, Strides)) {
1097 IsReadOnlyPtr = true;
1100 AliasAnalysis::Location Loc = AA->getLocation(LD);
1101 // The TBAA metadata could have a control dependency on the predication
1102 // condition, so we cannot rely on it when determining whether or not we
1103 // need runtime pointer checks.
1104 if (blockNeedsPredication(LD->getParent(), TheLoop, DT))
1105 Loc.AATags.TBAA = nullptr;
1107 Accesses.addLoad(Loc, IsReadOnlyPtr);
1110 // If we write (or read-write) to a single destination and there are no
1111 // other reads in this loop then is it safe to vectorize.
1112 if (NumReadWrites == 1 && NumReads == 0) {
1113 DEBUG(dbgs() << "LAA: Found a write-only loop!\n");
1118 // Build dependence sets and check whether we need a runtime pointer bounds
1120 Accesses.buildDependenceSets();
1121 bool NeedRTCheck = Accesses.isRTCheckNeeded();
1123 // Find pointers with computable bounds. We are going to use this information
1124 // to place a runtime bound check.
1125 bool CanDoRT = false;
1127 CanDoRT = Accesses.canCheckPtrAtRT(PtrRtCheck, NumComparisons, SE, TheLoop,
1130 DEBUG(dbgs() << "LAA: We need to do " << NumComparisons <<
1131 " pointer comparisons.\n");
1133 // If we only have one set of dependences to check pointers among we don't
1134 // need a runtime check.
1135 if (NumComparisons == 0 && NeedRTCheck)
1136 NeedRTCheck = false;
1138 // Check that we found the bounds for the pointer.
1140 DEBUG(dbgs() << "LAA: We can perform a memory runtime check if needed.\n");
1141 else if (NeedRTCheck) {
1142 emitAnalysis(LoopAccessReport() << "cannot identify array bounds");
1143 DEBUG(dbgs() << "LAA: We can't vectorize because we can't find " <<
1144 "the array bounds.\n");
1150 PtrRtCheck.Need = NeedRTCheck;
1153 if (Accesses.isDependencyCheckNeeded()) {
1154 DEBUG(dbgs() << "LAA: Checking memory dependencies\n");
1155 CanVecMem = DepChecker.areDepsSafe(
1156 DependentAccesses, Accesses.getDependenciesToCheck(), Strides);
1157 MaxSafeDepDistBytes = DepChecker.getMaxSafeDepDistBytes();
1159 if (!CanVecMem && DepChecker.shouldRetryWithRuntimeCheck()) {
1160 DEBUG(dbgs() << "LAA: Retrying with memory checks\n");
1163 // Clear the dependency checks. We assume they are not needed.
1164 Accesses.resetDepChecks();
1167 PtrRtCheck.Need = true;
1169 CanDoRT = Accesses.canCheckPtrAtRT(PtrRtCheck, NumComparisons, SE,
1170 TheLoop, Strides, true);
1171 // Check that we found the bounds for the pointer.
1172 if (!CanDoRT && NumComparisons > 0) {
1173 emitAnalysis(LoopAccessReport()
1174 << "cannot check memory dependencies at runtime");
1175 DEBUG(dbgs() << "LAA: Can't vectorize with memory checks\n");
1186 DEBUG(dbgs() << "LAA: No unsafe dependent memory operations in loop. We"
1187 << (NeedRTCheck ? "" : " don't")
1188 << " need a runtime memory check.\n");
1190 emitAnalysis(LoopAccessReport() <<
1191 "unsafe dependent memory operations in loop");
1192 DEBUG(dbgs() << "LAA: unsafe dependent memory operations in loop\n");
1196 bool LoopAccessInfo::blockNeedsPredication(BasicBlock *BB, Loop *TheLoop,
1197 DominatorTree *DT) {
1198 assert(TheLoop->contains(BB) && "Unknown block used");
1200 // Blocks that do not dominate the latch need predication.
1201 BasicBlock* Latch = TheLoop->getLoopLatch();
1202 return !DT->dominates(BB, Latch);
1205 void LoopAccessInfo::emitAnalysis(LoopAccessReport &Message) {
1206 assert(!Report && "Multiple reports generated");
1210 bool LoopAccessInfo::isUniform(Value *V) const {
1211 return (SE->isLoopInvariant(SE->getSCEV(V), TheLoop));
1214 // FIXME: this function is currently a duplicate of the one in
1215 // LoopVectorize.cpp.
1216 static Instruction *getFirstInst(Instruction *FirstInst, Value *V,
1220 if (Instruction *I = dyn_cast<Instruction>(V))
1221 return I->getParent() == Loc->getParent() ? I : nullptr;
1225 std::pair<Instruction *, Instruction *> LoopAccessInfo::addRuntimeCheck(
1226 Instruction *Loc, const SmallVectorImpl<int> *PtrPartition) const {
1227 if (!PtrRtCheck.Need)
1228 return std::make_pair(nullptr, nullptr);
1230 unsigned NumPointers = PtrRtCheck.Pointers.size();
1231 SmallVector<TrackingVH<Value> , 2> Starts;
1232 SmallVector<TrackingVH<Value> , 2> Ends;
1234 LLVMContext &Ctx = Loc->getContext();
1235 SCEVExpander Exp(*SE, DL, "induction");
1236 Instruction *FirstInst = nullptr;
1238 for (unsigned i = 0; i < NumPointers; ++i) {
1239 Value *Ptr = PtrRtCheck.Pointers[i];
1240 const SCEV *Sc = SE->getSCEV(Ptr);
1242 if (SE->isLoopInvariant(Sc, TheLoop)) {
1243 DEBUG(dbgs() << "LAA: Adding RT check for a loop invariant ptr:" <<
1245 Starts.push_back(Ptr);
1246 Ends.push_back(Ptr);
1248 DEBUG(dbgs() << "LAA: Adding RT check for range:" << *Ptr << '\n');
1249 unsigned AS = Ptr->getType()->getPointerAddressSpace();
1251 // Use this type for pointer arithmetic.
1252 Type *PtrArithTy = Type::getInt8PtrTy(Ctx, AS);
1254 Value *Start = Exp.expandCodeFor(PtrRtCheck.Starts[i], PtrArithTy, Loc);
1255 Value *End = Exp.expandCodeFor(PtrRtCheck.Ends[i], PtrArithTy, Loc);
1256 Starts.push_back(Start);
1257 Ends.push_back(End);
1261 IRBuilder<> ChkBuilder(Loc);
1262 // Our instructions might fold to a constant.
1263 Value *MemoryRuntimeCheck = nullptr;
1264 for (unsigned i = 0; i < NumPointers; ++i) {
1265 for (unsigned j = i+1; j < NumPointers; ++j) {
1266 if (!PtrRtCheck.needsChecking(i, j, PtrPartition))
1269 unsigned AS0 = Starts[i]->getType()->getPointerAddressSpace();
1270 unsigned AS1 = Starts[j]->getType()->getPointerAddressSpace();
1272 assert((AS0 == Ends[j]->getType()->getPointerAddressSpace()) &&
1273 (AS1 == Ends[i]->getType()->getPointerAddressSpace()) &&
1274 "Trying to bounds check pointers with different address spaces");
1276 Type *PtrArithTy0 = Type::getInt8PtrTy(Ctx, AS0);
1277 Type *PtrArithTy1 = Type::getInt8PtrTy(Ctx, AS1);
1279 Value *Start0 = ChkBuilder.CreateBitCast(Starts[i], PtrArithTy0, "bc");
1280 Value *Start1 = ChkBuilder.CreateBitCast(Starts[j], PtrArithTy1, "bc");
1281 Value *End0 = ChkBuilder.CreateBitCast(Ends[i], PtrArithTy1, "bc");
1282 Value *End1 = ChkBuilder.CreateBitCast(Ends[j], PtrArithTy0, "bc");
1284 Value *Cmp0 = ChkBuilder.CreateICmpULE(Start0, End1, "bound0");
1285 FirstInst = getFirstInst(FirstInst, Cmp0, Loc);
1286 Value *Cmp1 = ChkBuilder.CreateICmpULE(Start1, End0, "bound1");
1287 FirstInst = getFirstInst(FirstInst, Cmp1, Loc);
1288 Value *IsConflict = ChkBuilder.CreateAnd(Cmp0, Cmp1, "found.conflict");
1289 FirstInst = getFirstInst(FirstInst, IsConflict, Loc);
1290 if (MemoryRuntimeCheck) {
1291 IsConflict = ChkBuilder.CreateOr(MemoryRuntimeCheck, IsConflict,
1293 FirstInst = getFirstInst(FirstInst, IsConflict, Loc);
1295 MemoryRuntimeCheck = IsConflict;
1299 if (!MemoryRuntimeCheck)
1300 return std::make_pair(nullptr, nullptr);
1302 // We have to do this trickery because the IRBuilder might fold the check to a
1303 // constant expression in which case there is no Instruction anchored in a
1305 Instruction *Check = BinaryOperator::CreateAnd(MemoryRuntimeCheck,
1306 ConstantInt::getTrue(Ctx));
1307 ChkBuilder.Insert(Check, "memcheck.conflict");
1308 FirstInst = getFirstInst(FirstInst, Check, Loc);
1309 return std::make_pair(FirstInst, Check);
1312 LoopAccessInfo::LoopAccessInfo(Loop *L, ScalarEvolution *SE,
1313 const DataLayout &DL,
1314 const TargetLibraryInfo *TLI, AliasAnalysis *AA,
1315 DominatorTree *DT, LoopInfo *LI,
1316 const ValueToValueMap &Strides)
1317 : DepChecker(SE, L), NumComparisons(0), TheLoop(L), SE(SE), DL(DL),
1318 TLI(TLI), AA(AA), DT(DT), LI(LI), NumLoads(0), NumStores(0),
1319 MaxSafeDepDistBytes(-1U), CanVecMem(false),
1320 StoreToLoopInvariantAddress(false) {
1321 if (canAnalyzeLoop())
1322 analyzeLoop(Strides);
1325 void LoopAccessInfo::print(raw_ostream &OS, unsigned Depth) const {
1327 if (PtrRtCheck.Need)
1328 OS.indent(Depth) << "Memory dependences are safe with run-time checks\n";
1330 OS.indent(Depth) << "Memory dependences are safe\n";
1334 OS.indent(Depth) << "Report: " << Report->str() << "\n";
1336 if (auto *InterestingDependences = DepChecker.getInterestingDependences()) {
1337 OS.indent(Depth) << "Interesting Dependences:\n";
1338 for (auto &Dep : *InterestingDependences) {
1339 Dep.print(OS, Depth + 2, DepChecker.getMemoryInstructions());
1343 OS.indent(Depth) << "Too many interesting dependences, not recorded\n";
1345 // List the pair of accesses need run-time checks to prove independence.
1346 PtrRtCheck.print(OS, Depth);
1349 OS.indent(Depth) << "Store to invariant address was "
1350 << (StoreToLoopInvariantAddress ? "" : "not ")
1351 << "found in loop.\n";
1354 const LoopAccessInfo &
1355 LoopAccessAnalysis::getInfo(Loop *L, const ValueToValueMap &Strides) {
1356 auto &LAI = LoopAccessInfoMap[L];
1359 assert((!LAI || LAI->NumSymbolicStrides == Strides.size()) &&
1360 "Symbolic strides changed for loop");
1364 const DataLayout &DL = L->getHeader()->getModule()->getDataLayout();
1365 LAI = llvm::make_unique<LoopAccessInfo>(L, SE, DL, TLI, AA, DT, LI,
1368 LAI->NumSymbolicStrides = Strides.size();
1374 void LoopAccessAnalysis::print(raw_ostream &OS, const Module *M) const {
1375 LoopAccessAnalysis &LAA = *const_cast<LoopAccessAnalysis *>(this);
1377 ValueToValueMap NoSymbolicStrides;
1379 for (Loop *TopLevelLoop : *LI)
1380 for (Loop *L : depth_first(TopLevelLoop)) {
1381 OS.indent(2) << L->getHeader()->getName() << ":\n";
1382 auto &LAI = LAA.getInfo(L, NoSymbolicStrides);
1387 bool LoopAccessAnalysis::runOnFunction(Function &F) {
1388 SE = &getAnalysis<ScalarEvolution>();
1389 auto *TLIP = getAnalysisIfAvailable<TargetLibraryInfoWrapperPass>();
1390 TLI = TLIP ? &TLIP->getTLI() : nullptr;
1391 AA = &getAnalysis<AliasAnalysis>();
1392 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1393 LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
1398 void LoopAccessAnalysis::getAnalysisUsage(AnalysisUsage &AU) const {
1399 AU.addRequired<ScalarEvolution>();
1400 AU.addRequired<AliasAnalysis>();
1401 AU.addRequired<DominatorTreeWrapperPass>();
1402 AU.addRequired<LoopInfoWrapperPass>();
1404 AU.setPreservesAll();
1407 char LoopAccessAnalysis::ID = 0;
1408 static const char laa_name[] = "Loop Access Analysis";
1409 #define LAA_NAME "loop-accesses"
1411 INITIALIZE_PASS_BEGIN(LoopAccessAnalysis, LAA_NAME, laa_name, false, true)
1412 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
1413 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
1414 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
1415 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
1416 INITIALIZE_PASS_END(LoopAccessAnalysis, LAA_NAME, laa_name, false, true)
1419 Pass *createLAAPass() {
1420 return new LoopAccessAnalysis();