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/ValueTracking.h"
19 #include "llvm/IR/DiagnosticInfo.h"
20 #include "llvm/IR/Dominators.h"
21 #include "llvm/IR/IRBuilder.h"
22 #include "llvm/Support/Debug.h"
23 #include "llvm/Transforms/Utils/VectorUtils.h"
26 #define DEBUG_TYPE "loop-accesses"
28 void VectorizationReport::emitAnalysis(VectorizationReport &Message,
29 const Function *TheFunction,
31 const char *PassName) {
32 DebugLoc DL = TheLoop->getStartLoc();
33 if (Instruction *I = Message.getInstr())
34 DL = I->getDebugLoc();
35 emitOptimizationRemarkAnalysis(TheFunction->getContext(), PassName,
36 *TheFunction, DL, Message.str());
39 Value *llvm::stripIntegerCast(Value *V) {
40 if (CastInst *CI = dyn_cast<CastInst>(V))
41 if (CI->getOperand(0)->getType()->isIntegerTy())
42 return CI->getOperand(0);
46 const SCEV *llvm::replaceSymbolicStrideSCEV(ScalarEvolution *SE,
47 ValueToValueMap &PtrToStride,
48 Value *Ptr, Value *OrigPtr) {
50 const SCEV *OrigSCEV = SE->getSCEV(Ptr);
52 // If there is an entry in the map return the SCEV of the pointer with the
53 // symbolic stride replaced by one.
54 ValueToValueMap::iterator SI = PtrToStride.find(OrigPtr ? OrigPtr : Ptr);
55 if (SI != PtrToStride.end()) {
56 Value *StrideVal = SI->second;
59 StrideVal = stripIntegerCast(StrideVal);
61 // Replace symbolic stride by one.
62 Value *One = ConstantInt::get(StrideVal->getType(), 1);
63 ValueToValueMap RewriteMap;
64 RewriteMap[StrideVal] = One;
67 SCEVParameterRewriter::rewrite(OrigSCEV, *SE, RewriteMap, true);
68 DEBUG(dbgs() << "LAA: Replacing SCEV: " << *OrigSCEV << " by: " << *ByOne
73 // Otherwise, just return the SCEV of the original pointer.
74 return SE->getSCEV(Ptr);
77 void LoopAccessInfo::RuntimePointerCheck::insert(ScalarEvolution *SE, Loop *Lp,
78 Value *Ptr, bool WritePtr,
81 ValueToValueMap &Strides) {
82 // Get the stride replaced scev.
83 const SCEV *Sc = replaceSymbolicStrideSCEV(SE, Strides, Ptr);
84 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Sc);
85 assert(AR && "Invalid addrec expression");
86 const SCEV *Ex = SE->getBackedgeTakenCount(Lp);
87 const SCEV *ScEnd = AR->evaluateAtIteration(Ex, *SE);
88 Pointers.push_back(Ptr);
89 Starts.push_back(AR->getStart());
90 Ends.push_back(ScEnd);
91 IsWritePtr.push_back(WritePtr);
92 DependencySetId.push_back(DepSetId);
93 AliasSetId.push_back(ASId);
97 /// \brief Analyses memory accesses in a loop.
99 /// Checks whether run time pointer checks are needed and builds sets for data
100 /// dependence checking.
101 class AccessAnalysis {
103 /// \brief Read or write access location.
104 typedef PointerIntPair<Value *, 1, bool> MemAccessInfo;
105 typedef SmallPtrSet<MemAccessInfo, 8> MemAccessInfoSet;
107 /// \brief Set of potential dependent memory accesses.
108 typedef EquivalenceClasses<MemAccessInfo> DepCandidates;
110 AccessAnalysis(const DataLayout *Dl, AliasAnalysis *AA, DepCandidates &DA) :
111 DL(Dl), AST(*AA), DepCands(DA), IsRTCheckNeeded(false) {}
113 /// \brief Register a load and whether it is only read from.
114 void addLoad(AliasAnalysis::Location &Loc, bool IsReadOnly) {
115 Value *Ptr = const_cast<Value*>(Loc.Ptr);
116 AST.add(Ptr, AliasAnalysis::UnknownSize, Loc.AATags);
117 Accesses.insert(MemAccessInfo(Ptr, false));
119 ReadOnlyPtr.insert(Ptr);
122 /// \brief Register a store.
123 void addStore(AliasAnalysis::Location &Loc) {
124 Value *Ptr = const_cast<Value*>(Loc.Ptr);
125 AST.add(Ptr, AliasAnalysis::UnknownSize, Loc.AATags);
126 Accesses.insert(MemAccessInfo(Ptr, true));
129 /// \brief Check whether we can check the pointers at runtime for
130 /// non-intersection.
131 bool canCheckPtrAtRT(LoopAccessInfo::RuntimePointerCheck &RtCheck,
132 unsigned &NumComparisons,
133 ScalarEvolution *SE, Loop *TheLoop,
134 ValueToValueMap &Strides,
135 bool ShouldCheckStride = false);
137 /// \brief Goes over all memory accesses, checks whether a RT check is needed
138 /// and builds sets of dependent accesses.
139 void buildDependenceSets() {
140 processMemAccesses();
143 bool isRTCheckNeeded() { return IsRTCheckNeeded; }
145 bool isDependencyCheckNeeded() { return !CheckDeps.empty(); }
146 void resetDepChecks() { CheckDeps.clear(); }
148 MemAccessInfoSet &getDependenciesToCheck() { return CheckDeps; }
151 typedef SetVector<MemAccessInfo> PtrAccessSet;
153 /// \brief Go over all memory access and check whether runtime pointer checks
154 /// are needed /// and build sets of dependency check candidates.
155 void processMemAccesses();
157 /// Set of all accesses.
158 PtrAccessSet Accesses;
160 /// Set of accesses that need a further dependence check.
161 MemAccessInfoSet CheckDeps;
163 /// Set of pointers that are read only.
164 SmallPtrSet<Value*, 16> ReadOnlyPtr;
166 const DataLayout *DL;
168 /// An alias set tracker to partition the access set by underlying object and
169 //intrinsic property (such as TBAA metadata).
172 /// Sets of potentially dependent accesses - members of one set share an
173 /// underlying pointer. The set "CheckDeps" identfies which sets really need a
174 /// dependence check.
175 DepCandidates &DepCands;
177 bool IsRTCheckNeeded;
180 } // end anonymous namespace
182 /// \brief Check whether a pointer can participate in a runtime bounds check.
183 static bool hasComputableBounds(ScalarEvolution *SE, ValueToValueMap &Strides,
185 const SCEV *PtrScev = replaceSymbolicStrideSCEV(SE, Strides, Ptr);
186 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
190 return AR->isAffine();
193 /// \brief Check the stride of the pointer and ensure that it does not wrap in
194 /// the address space.
195 static int isStridedPtr(ScalarEvolution *SE, const DataLayout *DL, Value *Ptr,
196 const Loop *Lp, ValueToValueMap &StridesMap);
198 bool AccessAnalysis::canCheckPtrAtRT(
199 LoopAccessInfo::RuntimePointerCheck &RtCheck,
200 unsigned &NumComparisons, ScalarEvolution *SE, Loop *TheLoop,
201 ValueToValueMap &StridesMap, bool ShouldCheckStride) {
202 // Find pointers with computable bounds. We are going to use this information
203 // to place a runtime bound check.
206 bool IsDepCheckNeeded = isDependencyCheckNeeded();
209 // We assign a consecutive id to access from different alias sets.
210 // Accesses between different groups doesn't need to be checked.
212 for (auto &AS : AST) {
213 unsigned NumReadPtrChecks = 0;
214 unsigned NumWritePtrChecks = 0;
216 // We assign consecutive id to access from different dependence sets.
217 // Accesses within the same set don't need a runtime check.
218 unsigned RunningDepId = 1;
219 DenseMap<Value *, unsigned> DepSetId;
222 Value *Ptr = A.getValue();
223 bool IsWrite = Accesses.count(MemAccessInfo(Ptr, true));
224 MemAccessInfo Access(Ptr, IsWrite);
231 if (hasComputableBounds(SE, StridesMap, Ptr) &&
232 // When we run after a failing dependency check we have to make sure we
233 // don't have wrapping pointers.
234 (!ShouldCheckStride ||
235 isStridedPtr(SE, DL, Ptr, TheLoop, StridesMap) == 1)) {
236 // The id of the dependence set.
239 if (IsDepCheckNeeded) {
240 Value *Leader = DepCands.getLeaderValue(Access).getPointer();
241 unsigned &LeaderId = DepSetId[Leader];
243 LeaderId = RunningDepId++;
246 // Each access has its own dependence set.
247 DepId = RunningDepId++;
249 RtCheck.insert(SE, TheLoop, Ptr, IsWrite, DepId, ASId, StridesMap);
251 DEBUG(dbgs() << "LAA: Found a runtime check ptr:" << *Ptr << '\n');
257 if (IsDepCheckNeeded && CanDoRT && RunningDepId == 2)
258 NumComparisons += 0; // Only one dependence set.
260 NumComparisons += (NumWritePtrChecks * (NumReadPtrChecks +
261 NumWritePtrChecks - 1));
267 // If the pointers that we would use for the bounds comparison have different
268 // address spaces, assume the values aren't directly comparable, so we can't
269 // use them for the runtime check. We also have to assume they could
270 // overlap. In the future there should be metadata for whether address spaces
272 unsigned NumPointers = RtCheck.Pointers.size();
273 for (unsigned i = 0; i < NumPointers; ++i) {
274 for (unsigned j = i + 1; j < NumPointers; ++j) {
275 // Only need to check pointers between two different dependency sets.
276 if (RtCheck.DependencySetId[i] == RtCheck.DependencySetId[j])
278 // Only need to check pointers in the same alias set.
279 if (RtCheck.AliasSetId[i] != RtCheck.AliasSetId[j])
282 Value *PtrI = RtCheck.Pointers[i];
283 Value *PtrJ = RtCheck.Pointers[j];
285 unsigned ASi = PtrI->getType()->getPointerAddressSpace();
286 unsigned ASj = PtrJ->getType()->getPointerAddressSpace();
288 DEBUG(dbgs() << "LAA: Runtime check would require comparison between"
289 " different address spaces\n");
298 void AccessAnalysis::processMemAccesses() {
299 // We process the set twice: first we process read-write pointers, last we
300 // process read-only pointers. This allows us to skip dependence tests for
301 // read-only pointers.
303 DEBUG(dbgs() << "LAA: Processing memory accesses...\n");
304 DEBUG(dbgs() << " AST: "; AST.dump());
305 DEBUG(dbgs() << "LAA: Accesses:\n");
307 for (auto A : Accesses)
308 dbgs() << "\t" << *A.getPointer() << " (" <<
309 (A.getInt() ? "write" : (ReadOnlyPtr.count(A.getPointer()) ?
310 "read-only" : "read")) << ")\n";
313 // The AliasSetTracker has nicely partitioned our pointers by metadata
314 // compatibility and potential for underlying-object overlap. As a result, we
315 // only need to check for potential pointer dependencies within each alias
317 for (auto &AS : AST) {
318 // Note that both the alias-set tracker and the alias sets themselves used
319 // linked lists internally and so the iteration order here is deterministic
320 // (matching the original instruction order within each set).
322 bool SetHasWrite = false;
324 // Map of pointers to last access encountered.
325 typedef DenseMap<Value*, MemAccessInfo> UnderlyingObjToAccessMap;
326 UnderlyingObjToAccessMap ObjToLastAccess;
328 // Set of access to check after all writes have been processed.
329 PtrAccessSet DeferredAccesses;
331 // Iterate over each alias set twice, once to process read/write pointers,
332 // and then to process read-only pointers.
333 for (int SetIteration = 0; SetIteration < 2; ++SetIteration) {
334 bool UseDeferred = SetIteration > 0;
335 PtrAccessSet &S = UseDeferred ? DeferredAccesses : Accesses;
338 Value *Ptr = AV.getValue();
340 // For a single memory access in AliasSetTracker, Accesses may contain
341 // both read and write, and they both need to be handled for CheckDeps.
343 if (AC.getPointer() != Ptr)
346 bool IsWrite = AC.getInt();
348 // If we're using the deferred access set, then it contains only
350 bool IsReadOnlyPtr = ReadOnlyPtr.count(Ptr) && !IsWrite;
351 if (UseDeferred && !IsReadOnlyPtr)
353 // Otherwise, the pointer must be in the PtrAccessSet, either as a
355 assert(((IsReadOnlyPtr && UseDeferred) || IsWrite ||
356 S.count(MemAccessInfo(Ptr, false))) &&
357 "Alias-set pointer not in the access set?");
359 MemAccessInfo Access(Ptr, IsWrite);
360 DepCands.insert(Access);
362 // Memorize read-only pointers for later processing and skip them in
363 // the first round (they need to be checked after we have seen all
364 // write pointers). Note: we also mark pointer that are not
365 // consecutive as "read-only" pointers (so that we check
366 // "a[b[i]] +="). Hence, we need the second check for "!IsWrite".
367 if (!UseDeferred && IsReadOnlyPtr) {
368 DeferredAccesses.insert(Access);
372 // If this is a write - check other reads and writes for conflicts. If
373 // this is a read only check other writes for conflicts (but only if
374 // there is no other write to the ptr - this is an optimization to
375 // catch "a[i] = a[i] + " without having to do a dependence check).
376 if ((IsWrite || IsReadOnlyPtr) && SetHasWrite) {
377 CheckDeps.insert(Access);
378 IsRTCheckNeeded = true;
384 // Create sets of pointers connected by a shared alias set and
385 // underlying object.
386 typedef SmallVector<Value *, 16> ValueVector;
387 ValueVector TempObjects;
388 GetUnderlyingObjects(Ptr, TempObjects, DL);
389 for (Value *UnderlyingObj : TempObjects) {
390 UnderlyingObjToAccessMap::iterator Prev =
391 ObjToLastAccess.find(UnderlyingObj);
392 if (Prev != ObjToLastAccess.end())
393 DepCands.unionSets(Access, Prev->second);
395 ObjToLastAccess[UnderlyingObj] = Access;
404 /// \brief Checks memory dependences among accesses to the same underlying
405 /// object to determine whether there vectorization is legal or not (and at
406 /// which vectorization factor).
408 /// This class works under the assumption that we already checked that memory
409 /// locations with different underlying pointers are "must-not alias".
410 /// We use the ScalarEvolution framework to symbolically evalutate access
411 /// functions pairs. Since we currently don't restructure the loop we can rely
412 /// on the program order of memory accesses to determine their safety.
413 /// At the moment we will only deem accesses as safe for:
414 /// * A negative constant distance assuming program order.
416 /// Safe: tmp = a[i + 1]; OR a[i + 1] = x;
417 /// a[i] = tmp; y = a[i];
419 /// The latter case is safe because later checks guarantuee that there can't
420 /// be a cycle through a phi node (that is, we check that "x" and "y" is not
421 /// the same variable: a header phi can only be an induction or a reduction, a
422 /// reduction can't have a memory sink, an induction can't have a memory
423 /// source). This is important and must not be violated (or we have to
424 /// resort to checking for cycles through memory).
426 /// * A positive constant distance assuming program order that is bigger
427 /// than the biggest memory access.
429 /// tmp = a[i] OR b[i] = x
430 /// a[i+2] = tmp y = b[i+2];
432 /// Safe distance: 2 x sizeof(a[0]), and 2 x sizeof(b[0]), respectively.
434 /// * Zero distances and all accesses have the same size.
436 class MemoryDepChecker {
438 typedef PointerIntPair<Value *, 1, bool> MemAccessInfo;
439 typedef SmallPtrSet<MemAccessInfo, 8> MemAccessInfoSet;
441 MemoryDepChecker(ScalarEvolution *Se, const DataLayout *Dl, const Loop *L)
442 : SE(Se), DL(Dl), InnermostLoop(L), AccessIdx(0),
443 ShouldRetryWithRuntimeCheck(false) {}
445 /// \brief Register the location (instructions are given increasing numbers)
446 /// of a write access.
447 void addAccess(StoreInst *SI) {
448 Value *Ptr = SI->getPointerOperand();
449 Accesses[MemAccessInfo(Ptr, true)].push_back(AccessIdx);
450 InstMap.push_back(SI);
454 /// \brief Register the location (instructions are given increasing numbers)
455 /// of a write access.
456 void addAccess(LoadInst *LI) {
457 Value *Ptr = LI->getPointerOperand();
458 Accesses[MemAccessInfo(Ptr, false)].push_back(AccessIdx);
459 InstMap.push_back(LI);
463 /// \brief Check whether the dependencies between the accesses are safe.
465 /// Only checks sets with elements in \p CheckDeps.
466 bool areDepsSafe(AccessAnalysis::DepCandidates &AccessSets,
467 MemAccessInfoSet &CheckDeps, ValueToValueMap &Strides);
469 /// \brief The maximum number of bytes of a vector register we can vectorize
470 /// the accesses safely with.
471 unsigned getMaxSafeDepDistBytes() { return MaxSafeDepDistBytes; }
473 /// \brief In same cases when the dependency check fails we can still
474 /// vectorize the loop with a dynamic array access check.
475 bool shouldRetryWithRuntimeCheck() { return ShouldRetryWithRuntimeCheck; }
479 const DataLayout *DL;
480 const Loop *InnermostLoop;
482 /// \brief Maps access locations (ptr, read/write) to program order.
483 DenseMap<MemAccessInfo, std::vector<unsigned> > Accesses;
485 /// \brief Memory access instructions in program order.
486 SmallVector<Instruction *, 16> InstMap;
488 /// \brief The program order index to be used for the next instruction.
491 // We can access this many bytes in parallel safely.
492 unsigned MaxSafeDepDistBytes;
494 /// \brief If we see a non-constant dependence distance we can still try to
495 /// vectorize this loop with runtime checks.
496 bool ShouldRetryWithRuntimeCheck;
498 /// \brief Check whether there is a plausible dependence between the two
501 /// Access \p A must happen before \p B in program order. The two indices
502 /// identify the index into the program order map.
504 /// This function checks whether there is a plausible dependence (or the
505 /// absence of such can't be proved) between the two accesses. If there is a
506 /// plausible dependence but the dependence distance is bigger than one
507 /// element access it records this distance in \p MaxSafeDepDistBytes (if this
508 /// distance is smaller than any other distance encountered so far).
509 /// Otherwise, this function returns true signaling a possible dependence.
510 bool isDependent(const MemAccessInfo &A, unsigned AIdx,
511 const MemAccessInfo &B, unsigned BIdx,
512 ValueToValueMap &Strides);
514 /// \brief Check whether the data dependence could prevent store-load
516 bool couldPreventStoreLoadForward(unsigned Distance, unsigned TypeByteSize);
519 } // end anonymous namespace
521 static bool isInBoundsGep(Value *Ptr) {
522 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr))
523 return GEP->isInBounds();
527 /// \brief Check whether the access through \p Ptr has a constant stride.
528 static int isStridedPtr(ScalarEvolution *SE, const DataLayout *DL, Value *Ptr,
529 const Loop *Lp, ValueToValueMap &StridesMap) {
530 const Type *Ty = Ptr->getType();
531 assert(Ty->isPointerTy() && "Unexpected non-ptr");
533 // Make sure that the pointer does not point to aggregate types.
534 const PointerType *PtrTy = cast<PointerType>(Ty);
535 if (PtrTy->getElementType()->isAggregateType()) {
536 DEBUG(dbgs() << "LAA: Bad stride - Not a pointer to a scalar type"
541 const SCEV *PtrScev = replaceSymbolicStrideSCEV(SE, StridesMap, Ptr);
543 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
545 DEBUG(dbgs() << "LAA: Bad stride - Not an AddRecExpr pointer "
546 << *Ptr << " SCEV: " << *PtrScev << "\n");
550 // The accesss function must stride over the innermost loop.
551 if (Lp != AR->getLoop()) {
552 DEBUG(dbgs() << "LAA: Bad stride - Not striding over innermost loop " <<
553 *Ptr << " SCEV: " << *PtrScev << "\n");
556 // The address calculation must not wrap. Otherwise, a dependence could be
558 // An inbounds getelementptr that is a AddRec with a unit stride
559 // cannot wrap per definition. The unit stride requirement is checked later.
560 // An getelementptr without an inbounds attribute and unit stride would have
561 // to access the pointer value "0" which is undefined behavior in address
562 // space 0, therefore we can also vectorize this case.
563 bool IsInBoundsGEP = isInBoundsGep(Ptr);
564 bool IsNoWrapAddRec = AR->getNoWrapFlags(SCEV::NoWrapMask);
565 bool IsInAddressSpaceZero = PtrTy->getAddressSpace() == 0;
566 if (!IsNoWrapAddRec && !IsInBoundsGEP && !IsInAddressSpaceZero) {
567 DEBUG(dbgs() << "LAA: Bad stride - Pointer may wrap in the address space "
568 << *Ptr << " SCEV: " << *PtrScev << "\n");
572 // Check the step is constant.
573 const SCEV *Step = AR->getStepRecurrence(*SE);
575 // Calculate the pointer stride and check if it is consecutive.
576 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
578 DEBUG(dbgs() << "LAA: Bad stride - Not a constant strided " << *Ptr <<
579 " SCEV: " << *PtrScev << "\n");
583 int64_t Size = DL->getTypeAllocSize(PtrTy->getElementType());
584 const APInt &APStepVal = C->getValue()->getValue();
586 // Huge step value - give up.
587 if (APStepVal.getBitWidth() > 64)
590 int64_t StepVal = APStepVal.getSExtValue();
593 int64_t Stride = StepVal / Size;
594 int64_t Rem = StepVal % Size;
598 // If the SCEV could wrap but we have an inbounds gep with a unit stride we
599 // know we can't "wrap around the address space". In case of address space
600 // zero we know that this won't happen without triggering undefined behavior.
601 if (!IsNoWrapAddRec && (IsInBoundsGEP || IsInAddressSpaceZero) &&
602 Stride != 1 && Stride != -1)
608 bool MemoryDepChecker::couldPreventStoreLoadForward(unsigned Distance,
609 unsigned TypeByteSize) {
610 // If loads occur at a distance that is not a multiple of a feasible vector
611 // factor store-load forwarding does not take place.
612 // Positive dependences might cause troubles because vectorizing them might
613 // prevent store-load forwarding making vectorized code run a lot slower.
614 // a[i] = a[i-3] ^ a[i-8];
615 // The stores to a[i:i+1] don't align with the stores to a[i-3:i-2] and
616 // hence on your typical architecture store-load forwarding does not take
617 // place. Vectorizing in such cases does not make sense.
618 // Store-load forwarding distance.
619 const unsigned NumCyclesForStoreLoadThroughMemory = 8*TypeByteSize;
620 // Maximum vector factor.
621 unsigned MaxVFWithoutSLForwardIssues =
622 VectorizerParams::MaxVectorWidth * TypeByteSize;
623 if(MaxSafeDepDistBytes < MaxVFWithoutSLForwardIssues)
624 MaxVFWithoutSLForwardIssues = MaxSafeDepDistBytes;
626 for (unsigned vf = 2*TypeByteSize; vf <= MaxVFWithoutSLForwardIssues;
628 if (Distance % vf && Distance / vf < NumCyclesForStoreLoadThroughMemory) {
629 MaxVFWithoutSLForwardIssues = (vf >>=1);
634 if (MaxVFWithoutSLForwardIssues< 2*TypeByteSize) {
635 DEBUG(dbgs() << "LAA: Distance " << Distance <<
636 " that could cause a store-load forwarding conflict\n");
640 if (MaxVFWithoutSLForwardIssues < MaxSafeDepDistBytes &&
641 MaxVFWithoutSLForwardIssues !=
642 VectorizerParams::MaxVectorWidth * TypeByteSize)
643 MaxSafeDepDistBytes = MaxVFWithoutSLForwardIssues;
647 bool MemoryDepChecker::isDependent(const MemAccessInfo &A, unsigned AIdx,
648 const MemAccessInfo &B, unsigned BIdx,
649 ValueToValueMap &Strides) {
650 assert (AIdx < BIdx && "Must pass arguments in program order");
652 Value *APtr = A.getPointer();
653 Value *BPtr = B.getPointer();
654 bool AIsWrite = A.getInt();
655 bool BIsWrite = B.getInt();
657 // Two reads are independent.
658 if (!AIsWrite && !BIsWrite)
661 // We cannot check pointers in different address spaces.
662 if (APtr->getType()->getPointerAddressSpace() !=
663 BPtr->getType()->getPointerAddressSpace())
666 const SCEV *AScev = replaceSymbolicStrideSCEV(SE, Strides, APtr);
667 const SCEV *BScev = replaceSymbolicStrideSCEV(SE, Strides, BPtr);
669 int StrideAPtr = isStridedPtr(SE, DL, APtr, InnermostLoop, Strides);
670 int StrideBPtr = isStridedPtr(SE, DL, BPtr, InnermostLoop, Strides);
672 const SCEV *Src = AScev;
673 const SCEV *Sink = BScev;
675 // If the induction step is negative we have to invert source and sink of the
677 if (StrideAPtr < 0) {
680 std::swap(APtr, BPtr);
681 std::swap(Src, Sink);
682 std::swap(AIsWrite, BIsWrite);
683 std::swap(AIdx, BIdx);
684 std::swap(StrideAPtr, StrideBPtr);
687 const SCEV *Dist = SE->getMinusSCEV(Sink, Src);
689 DEBUG(dbgs() << "LAA: Src Scev: " << *Src << "Sink Scev: " << *Sink
690 << "(Induction step: " << StrideAPtr << ")\n");
691 DEBUG(dbgs() << "LAA: Distance for " << *InstMap[AIdx] << " to "
692 << *InstMap[BIdx] << ": " << *Dist << "\n");
694 // Need consecutive accesses. We don't want to vectorize
695 // "A[B[i]] += ..." and similar code or pointer arithmetic that could wrap in
696 // the address space.
697 if (!StrideAPtr || !StrideBPtr || StrideAPtr != StrideBPtr){
698 DEBUG(dbgs() << "Non-consecutive pointer access\n");
702 const SCEVConstant *C = dyn_cast<SCEVConstant>(Dist);
704 DEBUG(dbgs() << "LAA: Dependence because of non-constant distance\n");
705 ShouldRetryWithRuntimeCheck = true;
709 Type *ATy = APtr->getType()->getPointerElementType();
710 Type *BTy = BPtr->getType()->getPointerElementType();
711 unsigned TypeByteSize = DL->getTypeAllocSize(ATy);
713 // Negative distances are not plausible dependencies.
714 const APInt &Val = C->getValue()->getValue();
715 if (Val.isNegative()) {
716 bool IsTrueDataDependence = (AIsWrite && !BIsWrite);
717 if (IsTrueDataDependence &&
718 (couldPreventStoreLoadForward(Val.abs().getZExtValue(), TypeByteSize) ||
722 DEBUG(dbgs() << "LAA: Dependence is negative: NoDep\n");
726 // Write to the same location with the same size.
727 // Could be improved to assert type sizes are the same (i32 == float, etc).
731 DEBUG(dbgs() << "LAA: Zero dependence difference but different types\n");
735 assert(Val.isStrictlyPositive() && "Expect a positive value");
737 // Positive distance bigger than max vectorization factor.
740 "LAA: ReadWrite-Write positive dependency with different types\n");
744 unsigned Distance = (unsigned) Val.getZExtValue();
746 // Bail out early if passed-in parameters make vectorization not feasible.
747 unsigned ForcedFactor = (VectorizerParams::VectorizationFactor ?
748 VectorizerParams::VectorizationFactor : 1);
749 unsigned ForcedUnroll = (VectorizerParams::VectorizationInterleave ?
750 VectorizerParams::VectorizationInterleave : 1);
752 // The distance must be bigger than the size needed for a vectorized version
753 // of the operation and the size of the vectorized operation must not be
754 // bigger than the currrent maximum size.
755 if (Distance < 2*TypeByteSize ||
756 2*TypeByteSize > MaxSafeDepDistBytes ||
757 Distance < TypeByteSize * ForcedUnroll * ForcedFactor) {
758 DEBUG(dbgs() << "LAA: Failure because of Positive distance "
759 << Val.getSExtValue() << '\n');
763 MaxSafeDepDistBytes = Distance < MaxSafeDepDistBytes ?
764 Distance : MaxSafeDepDistBytes;
766 bool IsTrueDataDependence = (!AIsWrite && BIsWrite);
767 if (IsTrueDataDependence &&
768 couldPreventStoreLoadForward(Distance, TypeByteSize))
771 DEBUG(dbgs() << "LAA: Positive distance " << Val.getSExtValue() <<
772 " with max VF = " << MaxSafeDepDistBytes / TypeByteSize << '\n');
777 bool MemoryDepChecker::areDepsSafe(AccessAnalysis::DepCandidates &AccessSets,
778 MemAccessInfoSet &CheckDeps,
779 ValueToValueMap &Strides) {
781 MaxSafeDepDistBytes = -1U;
782 while (!CheckDeps.empty()) {
783 MemAccessInfo CurAccess = *CheckDeps.begin();
785 // Get the relevant memory access set.
786 EquivalenceClasses<MemAccessInfo>::iterator I =
787 AccessSets.findValue(AccessSets.getLeaderValue(CurAccess));
789 // Check accesses within this set.
790 EquivalenceClasses<MemAccessInfo>::member_iterator AI, AE;
791 AI = AccessSets.member_begin(I), AE = AccessSets.member_end();
793 // Check every access pair.
795 CheckDeps.erase(*AI);
796 EquivalenceClasses<MemAccessInfo>::member_iterator OI = std::next(AI);
798 // Check every accessing instruction pair in program order.
799 for (std::vector<unsigned>::iterator I1 = Accesses[*AI].begin(),
800 I1E = Accesses[*AI].end(); I1 != I1E; ++I1)
801 for (std::vector<unsigned>::iterator I2 = Accesses[*OI].begin(),
802 I2E = Accesses[*OI].end(); I2 != I2E; ++I2) {
803 if (*I1 < *I2 && isDependent(*AI, *I1, *OI, *I2, Strides))
805 if (*I2 < *I1 && isDependent(*OI, *I2, *AI, *I1, Strides))
816 void LoopAccessInfo::analyzeLoop(ValueToValueMap &Strides) {
818 typedef SmallVector<Value*, 16> ValueVector;
819 typedef SmallPtrSet<Value*, 16> ValueSet;
821 // Holds the Load and Store *instructions*.
825 // Holds all the different accesses in the loop.
826 unsigned NumReads = 0;
827 unsigned NumReadWrites = 0;
829 PtrRtCheck.Pointers.clear();
830 PtrRtCheck.Need = false;
832 const bool IsAnnotatedParallel = TheLoop->isAnnotatedParallel();
833 MemoryDepChecker DepChecker(SE, DL, TheLoop);
836 for (Loop::block_iterator bb = TheLoop->block_begin(),
837 be = TheLoop->block_end(); bb != be; ++bb) {
839 // Scan the BB and collect legal loads and stores.
840 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
843 // If this is a load, save it. If this instruction can read from memory
844 // but is not a load, then we quit. Notice that we don't handle function
845 // calls that read or write.
846 if (it->mayReadFromMemory()) {
847 // Many math library functions read the rounding mode. We will only
848 // vectorize a loop if it contains known function calls that don't set
849 // the flag. Therefore, it is safe to ignore this read from memory.
850 CallInst *Call = dyn_cast<CallInst>(it);
851 if (Call && getIntrinsicIDForCall(Call, TLI))
854 LoadInst *Ld = dyn_cast<LoadInst>(it);
855 if (!Ld || (!Ld->isSimple() && !IsAnnotatedParallel)) {
856 emitAnalysis(VectorizationReport(Ld)
857 << "read with atomic ordering or volatile read");
858 DEBUG(dbgs() << "LAA: Found a non-simple load.\n");
864 DepChecker.addAccess(Ld);
868 // Save 'store' instructions. Abort if other instructions write to memory.
869 if (it->mayWriteToMemory()) {
870 StoreInst *St = dyn_cast<StoreInst>(it);
872 emitAnalysis(VectorizationReport(it) <<
873 "instruction cannot be vectorized");
877 if (!St->isSimple() && !IsAnnotatedParallel) {
878 emitAnalysis(VectorizationReport(St)
879 << "write with atomic ordering or volatile write");
880 DEBUG(dbgs() << "LAA: Found a non-simple store.\n");
885 Stores.push_back(St);
886 DepChecker.addAccess(St);
891 // Now we have two lists that hold the loads and the stores.
892 // Next, we find the pointers that they use.
894 // Check if we see any stores. If there are no stores, then we don't
895 // care if the pointers are *restrict*.
896 if (!Stores.size()) {
897 DEBUG(dbgs() << "LAA: Found a read-only loop!\n");
902 AccessAnalysis::DepCandidates DependentAccesses;
903 AccessAnalysis Accesses(DL, AA, DependentAccesses);
905 // Holds the analyzed pointers. We don't want to call GetUnderlyingObjects
906 // multiple times on the same object. If the ptr is accessed twice, once
907 // for read and once for write, it will only appear once (on the write
908 // list). This is okay, since we are going to check for conflicts between
909 // writes and between reads and writes, but not between reads and reads.
912 ValueVector::iterator I, IE;
913 for (I = Stores.begin(), IE = Stores.end(); I != IE; ++I) {
914 StoreInst *ST = cast<StoreInst>(*I);
915 Value* Ptr = ST->getPointerOperand();
917 if (isUniform(Ptr)) {
919 VectorizationReport(ST)
920 << "write to a loop invariant address could not be vectorized");
921 DEBUG(dbgs() << "LAA: We don't allow storing to uniform addresses\n");
926 // If we did *not* see this pointer before, insert it to the read-write
927 // list. At this phase it is only a 'write' list.
928 if (Seen.insert(Ptr).second) {
931 AliasAnalysis::Location Loc = AA->getLocation(ST);
932 // The TBAA metadata could have a control dependency on the predication
933 // condition, so we cannot rely on it when determining whether or not we
934 // need runtime pointer checks.
935 if (blockNeedsPredication(ST->getParent(), TheLoop, DT))
936 Loc.AATags.TBAA = nullptr;
938 Accesses.addStore(Loc);
942 if (IsAnnotatedParallel) {
944 << "LAA: A loop annotated parallel, ignore memory dependency "
950 for (I = Loads.begin(), IE = Loads.end(); I != IE; ++I) {
951 LoadInst *LD = cast<LoadInst>(*I);
952 Value* Ptr = LD->getPointerOperand();
953 // If we did *not* see this pointer before, insert it to the
954 // read list. If we *did* see it before, then it is already in
955 // the read-write list. This allows us to vectorize expressions
956 // such as A[i] += x; Because the address of A[i] is a read-write
957 // pointer. This only works if the index of A[i] is consecutive.
958 // If the address of i is unknown (for example A[B[i]]) then we may
959 // read a few words, modify, and write a few words, and some of the
960 // words may be written to the same address.
961 bool IsReadOnlyPtr = false;
962 if (Seen.insert(Ptr).second ||
963 !isStridedPtr(SE, DL, Ptr, TheLoop, Strides)) {
965 IsReadOnlyPtr = true;
968 AliasAnalysis::Location Loc = AA->getLocation(LD);
969 // The TBAA metadata could have a control dependency on the predication
970 // condition, so we cannot rely on it when determining whether or not we
971 // need runtime pointer checks.
972 if (blockNeedsPredication(LD->getParent(), TheLoop, DT))
973 Loc.AATags.TBAA = nullptr;
975 Accesses.addLoad(Loc, IsReadOnlyPtr);
978 // If we write (or read-write) to a single destination and there are no
979 // other reads in this loop then is it safe to vectorize.
980 if (NumReadWrites == 1 && NumReads == 0) {
981 DEBUG(dbgs() << "LAA: Found a write-only loop!\n");
986 // Build dependence sets and check whether we need a runtime pointer bounds
988 Accesses.buildDependenceSets();
989 bool NeedRTCheck = Accesses.isRTCheckNeeded();
991 // Find pointers with computable bounds. We are going to use this information
992 // to place a runtime bound check.
993 unsigned NumComparisons = 0;
994 bool CanDoRT = false;
996 CanDoRT = Accesses.canCheckPtrAtRT(PtrRtCheck, NumComparisons, SE, TheLoop,
999 DEBUG(dbgs() << "LAA: We need to do " << NumComparisons <<
1000 " pointer comparisons.\n");
1002 // If we only have one set of dependences to check pointers among we don't
1003 // need a runtime check.
1004 if (NumComparisons == 0 && NeedRTCheck)
1005 NeedRTCheck = false;
1007 // Check that we did not collect too many pointers or found an unsizeable
1010 NumComparisons > VectorizerParams::RuntimeMemoryCheckThreshold) {
1016 DEBUG(dbgs() << "LAA: We can perform a memory runtime check if needed.\n");
1019 if (NeedRTCheck && !CanDoRT) {
1020 emitAnalysis(VectorizationReport() << "cannot identify array bounds");
1021 DEBUG(dbgs() << "LAA: We can't vectorize because we can't find " <<
1022 "the array bounds.\n");
1028 PtrRtCheck.Need = NeedRTCheck;
1031 if (Accesses.isDependencyCheckNeeded()) {
1032 DEBUG(dbgs() << "LAA: Checking memory dependencies\n");
1033 CanVecMem = DepChecker.areDepsSafe(
1034 DependentAccesses, Accesses.getDependenciesToCheck(), Strides);
1035 MaxSafeDepDistBytes = DepChecker.getMaxSafeDepDistBytes();
1037 if (!CanVecMem && DepChecker.shouldRetryWithRuntimeCheck()) {
1038 DEBUG(dbgs() << "LAA: Retrying with memory checks\n");
1041 // Clear the dependency checks. We assume they are not needed.
1042 Accesses.resetDepChecks();
1045 PtrRtCheck.Need = true;
1047 CanDoRT = Accesses.canCheckPtrAtRT(PtrRtCheck, NumComparisons, SE,
1048 TheLoop, Strides, true);
1049 // Check that we did not collect too many pointers or found an unsizeable
1052 NumComparisons > VectorizerParams::RuntimeMemoryCheckThreshold) {
1053 if (!CanDoRT && NumComparisons > 0)
1054 emitAnalysis(VectorizationReport()
1055 << "cannot check memory dependencies at runtime");
1057 emitAnalysis(VectorizationReport()
1058 << NumComparisons << " exceeds limit of "
1059 << VectorizerParams::RuntimeMemoryCheckThreshold
1060 << " dependent memory operations checked at runtime");
1061 DEBUG(dbgs() << "LAA: Can't vectorize with memory checks\n");
1072 emitAnalysis(VectorizationReport() <<
1073 "unsafe dependent memory operations in loop");
1075 DEBUG(dbgs() << "LAA: We" << (NeedRTCheck ? "" : " don't") <<
1076 " need a runtime memory check.\n");
1079 bool LoopAccessInfo::blockNeedsPredication(BasicBlock *BB, Loop *TheLoop,
1080 DominatorTree *DT) {
1081 assert(TheLoop->contains(BB) && "Unknown block used");
1083 // Blocks that do not dominate the latch need predication.
1084 BasicBlock* Latch = TheLoop->getLoopLatch();
1085 return !DT->dominates(BB, Latch);
1088 void LoopAccessInfo::emitAnalysis(VectorizationReport &Message) {
1089 assert(!Report && "Multiple report generated");
1093 bool LoopAccessInfo::isUniform(Value *V) {
1094 return (SE->isLoopInvariant(SE->getSCEV(V), TheLoop));
1097 // FIXME: this function is currently a duplicate of the one in
1098 // LoopVectorize.cpp.
1099 static Instruction *getFirstInst(Instruction *FirstInst, Value *V,
1103 if (Instruction *I = dyn_cast<Instruction>(V))
1104 return I->getParent() == Loc->getParent() ? I : nullptr;
1108 std::pair<Instruction *, Instruction *>
1109 LoopAccessInfo::addRuntimeCheck(Instruction *Loc) {
1110 Instruction *tnullptr = nullptr;
1111 if (!PtrRtCheck.Need)
1112 return std::pair<Instruction *, Instruction *>(tnullptr, tnullptr);
1114 unsigned NumPointers = PtrRtCheck.Pointers.size();
1115 SmallVector<TrackingVH<Value> , 2> Starts;
1116 SmallVector<TrackingVH<Value> , 2> Ends;
1118 LLVMContext &Ctx = Loc->getContext();
1119 SCEVExpander Exp(*SE, "induction");
1120 Instruction *FirstInst = nullptr;
1122 for (unsigned i = 0; i < NumPointers; ++i) {
1123 Value *Ptr = PtrRtCheck.Pointers[i];
1124 const SCEV *Sc = SE->getSCEV(Ptr);
1126 if (SE->isLoopInvariant(Sc, TheLoop)) {
1127 DEBUG(dbgs() << "LAA: Adding RT check for a loop invariant ptr:" <<
1129 Starts.push_back(Ptr);
1130 Ends.push_back(Ptr);
1132 DEBUG(dbgs() << "LAA: Adding RT check for range:" << *Ptr << '\n');
1133 unsigned AS = Ptr->getType()->getPointerAddressSpace();
1135 // Use this type for pointer arithmetic.
1136 Type *PtrArithTy = Type::getInt8PtrTy(Ctx, AS);
1138 Value *Start = Exp.expandCodeFor(PtrRtCheck.Starts[i], PtrArithTy, Loc);
1139 Value *End = Exp.expandCodeFor(PtrRtCheck.Ends[i], PtrArithTy, Loc);
1140 Starts.push_back(Start);
1141 Ends.push_back(End);
1145 IRBuilder<> ChkBuilder(Loc);
1146 // Our instructions might fold to a constant.
1147 Value *MemoryRuntimeCheck = nullptr;
1148 for (unsigned i = 0; i < NumPointers; ++i) {
1149 for (unsigned j = i+1; j < NumPointers; ++j) {
1150 // No need to check if two readonly pointers intersect.
1151 if (!PtrRtCheck.IsWritePtr[i] && !PtrRtCheck.IsWritePtr[j])
1154 // Only need to check pointers between two different dependency sets.
1155 if (PtrRtCheck.DependencySetId[i] == PtrRtCheck.DependencySetId[j])
1157 // Only need to check pointers in the same alias set.
1158 if (PtrRtCheck.AliasSetId[i] != PtrRtCheck.AliasSetId[j])
1161 unsigned AS0 = Starts[i]->getType()->getPointerAddressSpace();
1162 unsigned AS1 = Starts[j]->getType()->getPointerAddressSpace();
1164 assert((AS0 == Ends[j]->getType()->getPointerAddressSpace()) &&
1165 (AS1 == Ends[i]->getType()->getPointerAddressSpace()) &&
1166 "Trying to bounds check pointers with different address spaces");
1168 Type *PtrArithTy0 = Type::getInt8PtrTy(Ctx, AS0);
1169 Type *PtrArithTy1 = Type::getInt8PtrTy(Ctx, AS1);
1171 Value *Start0 = ChkBuilder.CreateBitCast(Starts[i], PtrArithTy0, "bc");
1172 Value *Start1 = ChkBuilder.CreateBitCast(Starts[j], PtrArithTy1, "bc");
1173 Value *End0 = ChkBuilder.CreateBitCast(Ends[i], PtrArithTy1, "bc");
1174 Value *End1 = ChkBuilder.CreateBitCast(Ends[j], PtrArithTy0, "bc");
1176 Value *Cmp0 = ChkBuilder.CreateICmpULE(Start0, End1, "bound0");
1177 FirstInst = getFirstInst(FirstInst, Cmp0, Loc);
1178 Value *Cmp1 = ChkBuilder.CreateICmpULE(Start1, End0, "bound1");
1179 FirstInst = getFirstInst(FirstInst, Cmp1, Loc);
1180 Value *IsConflict = ChkBuilder.CreateAnd(Cmp0, Cmp1, "found.conflict");
1181 FirstInst = getFirstInst(FirstInst, IsConflict, Loc);
1182 if (MemoryRuntimeCheck) {
1183 IsConflict = ChkBuilder.CreateOr(MemoryRuntimeCheck, IsConflict,
1185 FirstInst = getFirstInst(FirstInst, IsConflict, Loc);
1187 MemoryRuntimeCheck = IsConflict;
1191 // We have to do this trickery because the IRBuilder might fold the check to a
1192 // constant expression in which case there is no Instruction anchored in a
1194 Instruction *Check = BinaryOperator::CreateAnd(MemoryRuntimeCheck,
1195 ConstantInt::getTrue(Ctx));
1196 ChkBuilder.Insert(Check, "memcheck.conflict");
1197 FirstInst = getFirstInst(FirstInst, Check, Loc);
1198 return std::make_pair(FirstInst, Check);
1201 LoopAccessInfo::LoopAccessInfo(Loop *L, ScalarEvolution *SE,
1202 const DataLayout *DL,
1203 const TargetLibraryInfo *TLI, AliasAnalysis *AA,
1204 DominatorTree *DT, ValueToValueMap &Strides)
1205 : TheLoop(L), SE(SE), DL(DL), TLI(TLI), AA(AA), DT(DT), NumLoads(0),
1206 NumStores(0), MaxSafeDepDistBytes(-1U), CanVecMem(false) {
1207 analyzeLoop(Strides);
1210 LoopAccessInfo &LoopAccessAnalysis::getInfo(Loop *L, ValueToValueMap &Strides) {
1211 auto &LAI = LoopAccessInfoMap[L];
1214 assert((!LAI || LAI->NumSymbolicStrides == Strides.size()) &&
1215 "Symbolic strides changed for loop");
1219 LAI = make_unique<LoopAccessInfo>(L, SE, DL, TLI, AA, DT, Strides);
1221 LAI->NumSymbolicStrides = Strides.size();
1227 bool LoopAccessAnalysis::runOnFunction(Function &F) {
1228 SE = &getAnalysis<ScalarEvolution>();
1229 DL = F.getParent()->getDataLayout();
1230 auto *TLIP = getAnalysisIfAvailable<TargetLibraryInfoWrapperPass>();
1231 TLI = TLIP ? &TLIP->getTLI() : nullptr;
1232 AA = &getAnalysis<AliasAnalysis>();
1233 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1238 void LoopAccessAnalysis::getAnalysisUsage(AnalysisUsage &AU) const {
1239 AU.addRequired<ScalarEvolution>();
1240 AU.addRequired<AliasAnalysis>();
1241 AU.addRequired<DominatorTreeWrapperPass>();
1243 AU.setPreservesAll();
1246 char LoopAccessAnalysis::ID = 0;
1247 static const char laa_name[] = "Loop Access Analysis";
1248 #define LAA_NAME "loop-accesses"
1250 INITIALIZE_PASS_BEGIN(LoopAccessAnalysis, LAA_NAME, laa_name, false, true)
1251 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
1252 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
1253 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
1254 INITIALIZE_PASS_END(LoopAccessAnalysis, LAA_NAME, laa_name, false, true)
1257 Pass *createLAAPass() {
1258 return new LoopAccessAnalysis();