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-vectorize"
28 void VectorizationReport::emitAnalysis(VectorizationReport &Message,
29 const Function *TheFunction,
30 const Loop *TheLoop) {
31 DebugLoc DL = TheLoop->getStartLoc();
32 if (Instruction *I = Message.getInstr())
33 DL = I->getDebugLoc();
34 emitOptimizationRemarkAnalysis(TheFunction->getContext(), DEBUG_TYPE,
35 *TheFunction, DL, Message.str());
38 Value *llvm::stripIntegerCast(Value *V) {
39 if (CastInst *CI = dyn_cast<CastInst>(V))
40 if (CI->getOperand(0)->getType()->isIntegerTy())
41 return CI->getOperand(0);
45 const SCEV *llvm::replaceSymbolicStrideSCEV(ScalarEvolution *SE,
46 ValueToValueMap &PtrToStride,
47 Value *Ptr, Value *OrigPtr) {
49 const SCEV *OrigSCEV = SE->getSCEV(Ptr);
51 // If there is an entry in the map return the SCEV of the pointer with the
52 // symbolic stride replaced by one.
53 ValueToValueMap::iterator SI = PtrToStride.find(OrigPtr ? OrigPtr : Ptr);
54 if (SI != PtrToStride.end()) {
55 Value *StrideVal = SI->second;
58 StrideVal = stripIntegerCast(StrideVal);
60 // Replace symbolic stride by one.
61 Value *One = ConstantInt::get(StrideVal->getType(), 1);
62 ValueToValueMap RewriteMap;
63 RewriteMap[StrideVal] = One;
66 SCEVParameterRewriter::rewrite(OrigSCEV, *SE, RewriteMap, true);
67 DEBUG(dbgs() << "LV: Replacing SCEV: " << *OrigSCEV << " by: " << *ByOne
72 // Otherwise, just return the SCEV of the original pointer.
73 return SE->getSCEV(Ptr);
76 void LoopAccessAnalysis::RuntimePointerCheck::insert(ScalarEvolution *SE,
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(LoopAccessAnalysis::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 LoopAccessAnalysis::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() << "LV: 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() << "LV: 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() << "LV: Processing memory accesses...\n");
304 DEBUG(dbgs() << " AST: "; AST.dump());
305 DEBUG(dbgs() << "LV: 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 const LoopAccessAnalysis::VectorizerParams &VectParams)
443 : SE(Se), DL(Dl), InnermostLoop(L), AccessIdx(0),
444 ShouldRetryWithRuntimeCheck(false), VectParams(VectParams) {}
446 /// \brief Register the location (instructions are given increasing numbers)
447 /// of a write access.
448 void addAccess(StoreInst *SI) {
449 Value *Ptr = SI->getPointerOperand();
450 Accesses[MemAccessInfo(Ptr, true)].push_back(AccessIdx);
451 InstMap.push_back(SI);
455 /// \brief Register the location (instructions are given increasing numbers)
456 /// of a write access.
457 void addAccess(LoadInst *LI) {
458 Value *Ptr = LI->getPointerOperand();
459 Accesses[MemAccessInfo(Ptr, false)].push_back(AccessIdx);
460 InstMap.push_back(LI);
464 /// \brief Check whether the dependencies between the accesses are safe.
466 /// Only checks sets with elements in \p CheckDeps.
467 bool areDepsSafe(AccessAnalysis::DepCandidates &AccessSets,
468 MemAccessInfoSet &CheckDeps, ValueToValueMap &Strides);
470 /// \brief The maximum number of bytes of a vector register we can vectorize
471 /// the accesses safely with.
472 unsigned getMaxSafeDepDistBytes() { return MaxSafeDepDistBytes; }
474 /// \brief In same cases when the dependency check fails we can still
475 /// vectorize the loop with a dynamic array access check.
476 bool shouldRetryWithRuntimeCheck() { return ShouldRetryWithRuntimeCheck; }
480 const DataLayout *DL;
481 const Loop *InnermostLoop;
483 /// \brief Maps access locations (ptr, read/write) to program order.
484 DenseMap<MemAccessInfo, std::vector<unsigned> > Accesses;
486 /// \brief Memory access instructions in program order.
487 SmallVector<Instruction *, 16> InstMap;
489 /// \brief The program order index to be used for the next instruction.
492 // We can access this many bytes in parallel safely.
493 unsigned MaxSafeDepDistBytes;
495 /// \brief If we see a non-constant dependence distance we can still try to
496 /// vectorize this loop with runtime checks.
497 bool ShouldRetryWithRuntimeCheck;
499 /// \brief Vectorizer parameters used by the analysis.
500 LoopAccessAnalysis::VectorizerParams VectParams;
502 /// \brief Check whether there is a plausible dependence between the two
505 /// Access \p A must happen before \p B in program order. The two indices
506 /// identify the index into the program order map.
508 /// This function checks whether there is a plausible dependence (or the
509 /// absence of such can't be proved) between the two accesses. If there is a
510 /// plausible dependence but the dependence distance is bigger than one
511 /// element access it records this distance in \p MaxSafeDepDistBytes (if this
512 /// distance is smaller than any other distance encountered so far).
513 /// Otherwise, this function returns true signaling a possible dependence.
514 bool isDependent(const MemAccessInfo &A, unsigned AIdx,
515 const MemAccessInfo &B, unsigned BIdx,
516 ValueToValueMap &Strides);
518 /// \brief Check whether the data dependence could prevent store-load
520 bool couldPreventStoreLoadForward(unsigned Distance, unsigned TypeByteSize);
523 } // end anonymous namespace
525 static bool isInBoundsGep(Value *Ptr) {
526 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr))
527 return GEP->isInBounds();
531 /// \brief Check whether the access through \p Ptr has a constant stride.
532 static int isStridedPtr(ScalarEvolution *SE, const DataLayout *DL, Value *Ptr,
533 const Loop *Lp, ValueToValueMap &StridesMap) {
534 const Type *Ty = Ptr->getType();
535 assert(Ty->isPointerTy() && "Unexpected non-ptr");
537 // Make sure that the pointer does not point to aggregate types.
538 const PointerType *PtrTy = cast<PointerType>(Ty);
539 if (PtrTy->getElementType()->isAggregateType()) {
540 DEBUG(dbgs() << "LV: Bad stride - Not a pointer to a scalar type" << *Ptr <<
545 const SCEV *PtrScev = replaceSymbolicStrideSCEV(SE, StridesMap, Ptr);
547 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
549 DEBUG(dbgs() << "LV: Bad stride - Not an AddRecExpr pointer "
550 << *Ptr << " SCEV: " << *PtrScev << "\n");
554 // The accesss function must stride over the innermost loop.
555 if (Lp != AR->getLoop()) {
556 DEBUG(dbgs() << "LV: Bad stride - Not striding over innermost loop " <<
557 *Ptr << " SCEV: " << *PtrScev << "\n");
560 // The address calculation must not wrap. Otherwise, a dependence could be
562 // An inbounds getelementptr that is a AddRec with a unit stride
563 // cannot wrap per definition. The unit stride requirement is checked later.
564 // An getelementptr without an inbounds attribute and unit stride would have
565 // to access the pointer value "0" which is undefined behavior in address
566 // space 0, therefore we can also vectorize this case.
567 bool IsInBoundsGEP = isInBoundsGep(Ptr);
568 bool IsNoWrapAddRec = AR->getNoWrapFlags(SCEV::NoWrapMask);
569 bool IsInAddressSpaceZero = PtrTy->getAddressSpace() == 0;
570 if (!IsNoWrapAddRec && !IsInBoundsGEP && !IsInAddressSpaceZero) {
571 DEBUG(dbgs() << "LV: Bad stride - Pointer may wrap in the address space "
572 << *Ptr << " SCEV: " << *PtrScev << "\n");
576 // Check the step is constant.
577 const SCEV *Step = AR->getStepRecurrence(*SE);
579 // Calculate the pointer stride and check if it is consecutive.
580 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
582 DEBUG(dbgs() << "LV: Bad stride - Not a constant strided " << *Ptr <<
583 " SCEV: " << *PtrScev << "\n");
587 int64_t Size = DL->getTypeAllocSize(PtrTy->getElementType());
588 const APInt &APStepVal = C->getValue()->getValue();
590 // Huge step value - give up.
591 if (APStepVal.getBitWidth() > 64)
594 int64_t StepVal = APStepVal.getSExtValue();
597 int64_t Stride = StepVal / Size;
598 int64_t Rem = StepVal % Size;
602 // If the SCEV could wrap but we have an inbounds gep with a unit stride we
603 // know we can't "wrap around the address space". In case of address space
604 // zero we know that this won't happen without triggering undefined behavior.
605 if (!IsNoWrapAddRec && (IsInBoundsGEP || IsInAddressSpaceZero) &&
606 Stride != 1 && Stride != -1)
612 bool MemoryDepChecker::couldPreventStoreLoadForward(unsigned Distance,
613 unsigned TypeByteSize) {
614 // If loads occur at a distance that is not a multiple of a feasible vector
615 // factor store-load forwarding does not take place.
616 // Positive dependences might cause troubles because vectorizing them might
617 // prevent store-load forwarding making vectorized code run a lot slower.
618 // a[i] = a[i-3] ^ a[i-8];
619 // The stores to a[i:i+1] don't align with the stores to a[i-3:i-2] and
620 // hence on your typical architecture store-load forwarding does not take
621 // place. Vectorizing in such cases does not make sense.
622 // Store-load forwarding distance.
623 const unsigned NumCyclesForStoreLoadThroughMemory = 8*TypeByteSize;
624 // Maximum vector factor.
625 unsigned MaxVFWithoutSLForwardIssues = VectParams.MaxVectorWidth*TypeByteSize;
626 if(MaxSafeDepDistBytes < MaxVFWithoutSLForwardIssues)
627 MaxVFWithoutSLForwardIssues = MaxSafeDepDistBytes;
629 for (unsigned vf = 2*TypeByteSize; vf <= MaxVFWithoutSLForwardIssues;
631 if (Distance % vf && Distance / vf < NumCyclesForStoreLoadThroughMemory) {
632 MaxVFWithoutSLForwardIssues = (vf >>=1);
637 if (MaxVFWithoutSLForwardIssues< 2*TypeByteSize) {
638 DEBUG(dbgs() << "LV: Distance " << Distance <<
639 " that could cause a store-load forwarding conflict\n");
643 if (MaxVFWithoutSLForwardIssues < MaxSafeDepDistBytes &&
644 MaxVFWithoutSLForwardIssues != VectParams.MaxVectorWidth*TypeByteSize)
645 MaxSafeDepDistBytes = MaxVFWithoutSLForwardIssues;
649 bool MemoryDepChecker::isDependent(const MemAccessInfo &A, unsigned AIdx,
650 const MemAccessInfo &B, unsigned BIdx,
651 ValueToValueMap &Strides) {
652 assert (AIdx < BIdx && "Must pass arguments in program order");
654 Value *APtr = A.getPointer();
655 Value *BPtr = B.getPointer();
656 bool AIsWrite = A.getInt();
657 bool BIsWrite = B.getInt();
659 // Two reads are independent.
660 if (!AIsWrite && !BIsWrite)
663 // We cannot check pointers in different address spaces.
664 if (APtr->getType()->getPointerAddressSpace() !=
665 BPtr->getType()->getPointerAddressSpace())
668 const SCEV *AScev = replaceSymbolicStrideSCEV(SE, Strides, APtr);
669 const SCEV *BScev = replaceSymbolicStrideSCEV(SE, Strides, BPtr);
671 int StrideAPtr = isStridedPtr(SE, DL, APtr, InnermostLoop, Strides);
672 int StrideBPtr = isStridedPtr(SE, DL, BPtr, InnermostLoop, Strides);
674 const SCEV *Src = AScev;
675 const SCEV *Sink = BScev;
677 // If the induction step is negative we have to invert source and sink of the
679 if (StrideAPtr < 0) {
682 std::swap(APtr, BPtr);
683 std::swap(Src, Sink);
684 std::swap(AIsWrite, BIsWrite);
685 std::swap(AIdx, BIdx);
686 std::swap(StrideAPtr, StrideBPtr);
689 const SCEV *Dist = SE->getMinusSCEV(Sink, Src);
691 DEBUG(dbgs() << "LV: Src Scev: " << *Src << "Sink Scev: " << *Sink
692 << "(Induction step: " << StrideAPtr << ")\n");
693 DEBUG(dbgs() << "LV: Distance for " << *InstMap[AIdx] << " to "
694 << *InstMap[BIdx] << ": " << *Dist << "\n");
696 // Need consecutive accesses. We don't want to vectorize
697 // "A[B[i]] += ..." and similar code or pointer arithmetic that could wrap in
698 // the address space.
699 if (!StrideAPtr || !StrideBPtr || StrideAPtr != StrideBPtr){
700 DEBUG(dbgs() << "Non-consecutive pointer access\n");
704 const SCEVConstant *C = dyn_cast<SCEVConstant>(Dist);
706 DEBUG(dbgs() << "LV: Dependence because of non-constant distance\n");
707 ShouldRetryWithRuntimeCheck = true;
711 Type *ATy = APtr->getType()->getPointerElementType();
712 Type *BTy = BPtr->getType()->getPointerElementType();
713 unsigned TypeByteSize = DL->getTypeAllocSize(ATy);
715 // Negative distances are not plausible dependencies.
716 const APInt &Val = C->getValue()->getValue();
717 if (Val.isNegative()) {
718 bool IsTrueDataDependence = (AIsWrite && !BIsWrite);
719 if (IsTrueDataDependence &&
720 (couldPreventStoreLoadForward(Val.abs().getZExtValue(), TypeByteSize) ||
724 DEBUG(dbgs() << "LV: Dependence is negative: NoDep\n");
728 // Write to the same location with the same size.
729 // Could be improved to assert type sizes are the same (i32 == float, etc).
733 DEBUG(dbgs() << "LV: Zero dependence difference but different types\n");
737 assert(Val.isStrictlyPositive() && "Expect a positive value");
739 // Positive distance bigger than max vectorization factor.
742 "LV: ReadWrite-Write positive dependency with different types\n");
746 unsigned Distance = (unsigned) Val.getZExtValue();
748 // Bail out early if passed-in parameters make vectorization not feasible.
749 unsigned ForcedFactor = (VectParams.VectorizationFactor ?
750 VectParams.VectorizationFactor : 1);
751 unsigned ForcedUnroll = (VectParams.VectorizationInterleave ?
752 VectParams.VectorizationInterleave : 1);
754 // The distance must be bigger than the size needed for a vectorized version
755 // of the operation and the size of the vectorized operation must not be
756 // bigger than the currrent maximum size.
757 if (Distance < 2*TypeByteSize ||
758 2*TypeByteSize > MaxSafeDepDistBytes ||
759 Distance < TypeByteSize * ForcedUnroll * ForcedFactor) {
760 DEBUG(dbgs() << "LV: Failure because of Positive distance "
761 << Val.getSExtValue() << '\n');
765 MaxSafeDepDistBytes = Distance < MaxSafeDepDistBytes ?
766 Distance : MaxSafeDepDistBytes;
768 bool IsTrueDataDependence = (!AIsWrite && BIsWrite);
769 if (IsTrueDataDependence &&
770 couldPreventStoreLoadForward(Distance, TypeByteSize))
773 DEBUG(dbgs() << "LV: Positive distance " << Val.getSExtValue() <<
774 " with max VF = " << MaxSafeDepDistBytes / TypeByteSize << '\n');
779 bool MemoryDepChecker::areDepsSafe(AccessAnalysis::DepCandidates &AccessSets,
780 MemAccessInfoSet &CheckDeps,
781 ValueToValueMap &Strides) {
783 MaxSafeDepDistBytes = -1U;
784 while (!CheckDeps.empty()) {
785 MemAccessInfo CurAccess = *CheckDeps.begin();
787 // Get the relevant memory access set.
788 EquivalenceClasses<MemAccessInfo>::iterator I =
789 AccessSets.findValue(AccessSets.getLeaderValue(CurAccess));
791 // Check accesses within this set.
792 EquivalenceClasses<MemAccessInfo>::member_iterator AI, AE;
793 AI = AccessSets.member_begin(I), AE = AccessSets.member_end();
795 // Check every access pair.
797 CheckDeps.erase(*AI);
798 EquivalenceClasses<MemAccessInfo>::member_iterator OI = std::next(AI);
800 // Check every accessing instruction pair in program order.
801 for (std::vector<unsigned>::iterator I1 = Accesses[*AI].begin(),
802 I1E = Accesses[*AI].end(); I1 != I1E; ++I1)
803 for (std::vector<unsigned>::iterator I2 = Accesses[*OI].begin(),
804 I2E = Accesses[*OI].end(); I2 != I2E; ++I2) {
805 if (*I1 < *I2 && isDependent(*AI, *I1, *OI, *I2, Strides))
807 if (*I2 < *I1 && isDependent(*OI, *I2, *AI, *I1, Strides))
818 bool LoopAccessAnalysis::canVectorizeMemory(ValueToValueMap &Strides) {
820 typedef SmallVector<Value*, 16> ValueVector;
821 typedef SmallPtrSet<Value*, 16> ValueSet;
823 // Holds the Load and Store *instructions*.
827 // Holds all the different accesses in the loop.
828 unsigned NumReads = 0;
829 unsigned NumReadWrites = 0;
831 PtrRtCheck.Pointers.clear();
832 PtrRtCheck.Need = false;
834 const bool IsAnnotatedParallel = TheLoop->isAnnotatedParallel();
835 MemoryDepChecker DepChecker(SE, DL, TheLoop, VectParams);
838 for (Loop::block_iterator bb = TheLoop->block_begin(),
839 be = TheLoop->block_end(); bb != be; ++bb) {
841 // Scan the BB and collect legal loads and stores.
842 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
845 // If this is a load, save it. If this instruction can read from memory
846 // but is not a load, then we quit. Notice that we don't handle function
847 // calls that read or write.
848 if (it->mayReadFromMemory()) {
849 // Many math library functions read the rounding mode. We will only
850 // vectorize a loop if it contains known function calls that don't set
851 // the flag. Therefore, it is safe to ignore this read from memory.
852 CallInst *Call = dyn_cast<CallInst>(it);
853 if (Call && getIntrinsicIDForCall(Call, TLI))
856 LoadInst *Ld = dyn_cast<LoadInst>(it);
857 if (!Ld || (!Ld->isSimple() && !IsAnnotatedParallel)) {
858 emitAnalysis(VectorizationReport(Ld)
859 << "read with atomic ordering or volatile read");
860 DEBUG(dbgs() << "LV: Found a non-simple load.\n");
865 DepChecker.addAccess(Ld);
869 // Save 'store' instructions. Abort if other instructions write to memory.
870 if (it->mayWriteToMemory()) {
871 StoreInst *St = dyn_cast<StoreInst>(it);
873 emitAnalysis(VectorizationReport(it) <<
874 "instruction cannot be vectorized");
877 if (!St->isSimple() && !IsAnnotatedParallel) {
878 emitAnalysis(VectorizationReport(St)
879 << "write with atomic ordering or volatile write");
880 DEBUG(dbgs() << "LV: Found a non-simple store.\n");
884 Stores.push_back(St);
885 DepChecker.addAccess(St);
890 // Now we have two lists that hold the loads and the stores.
891 // Next, we find the pointers that they use.
893 // Check if we see any stores. If there are no stores, then we don't
894 // care if the pointers are *restrict*.
895 if (!Stores.size()) {
896 DEBUG(dbgs() << "LV: Found a read-only loop!\n");
900 AccessAnalysis::DepCandidates DependentAccesses;
901 AccessAnalysis Accesses(DL, AA, DependentAccesses);
903 // Holds the analyzed pointers. We don't want to call GetUnderlyingObjects
904 // multiple times on the same object. If the ptr is accessed twice, once
905 // for read and once for write, it will only appear once (on the write
906 // list). This is okay, since we are going to check for conflicts between
907 // writes and between reads and writes, but not between reads and reads.
910 ValueVector::iterator I, IE;
911 for (I = Stores.begin(), IE = Stores.end(); I != IE; ++I) {
912 StoreInst *ST = cast<StoreInst>(*I);
913 Value* Ptr = ST->getPointerOperand();
915 if (isUniform(Ptr)) {
917 VectorizationReport(ST)
918 << "write to a loop invariant address could not be vectorized");
919 DEBUG(dbgs() << "LV: We don't allow storing to uniform addresses\n");
923 // If we did *not* see this pointer before, insert it to the read-write
924 // list. At this phase it is only a 'write' list.
925 if (Seen.insert(Ptr).second) {
928 AliasAnalysis::Location Loc = AA->getLocation(ST);
929 // The TBAA metadata could have a control dependency on the predication
930 // condition, so we cannot rely on it when determining whether or not we
931 // need runtime pointer checks.
932 if (blockNeedsPredication(ST->getParent()))
933 Loc.AATags.TBAA = nullptr;
935 Accesses.addStore(Loc);
939 if (IsAnnotatedParallel) {
941 << "LV: A loop annotated parallel, ignore memory dependency "
946 for (I = Loads.begin(), IE = Loads.end(); I != IE; ++I) {
947 LoadInst *LD = cast<LoadInst>(*I);
948 Value* Ptr = LD->getPointerOperand();
949 // If we did *not* see this pointer before, insert it to the
950 // read list. If we *did* see it before, then it is already in
951 // the read-write list. This allows us to vectorize expressions
952 // such as A[i] += x; Because the address of A[i] is a read-write
953 // pointer. This only works if the index of A[i] is consecutive.
954 // If the address of i is unknown (for example A[B[i]]) then we may
955 // read a few words, modify, and write a few words, and some of the
956 // words may be written to the same address.
957 bool IsReadOnlyPtr = false;
958 if (Seen.insert(Ptr).second ||
959 !isStridedPtr(SE, DL, Ptr, TheLoop, Strides)) {
961 IsReadOnlyPtr = true;
964 AliasAnalysis::Location Loc = AA->getLocation(LD);
965 // The TBAA metadata could have a control dependency on the predication
966 // condition, so we cannot rely on it when determining whether or not we
967 // need runtime pointer checks.
968 if (blockNeedsPredication(LD->getParent()))
969 Loc.AATags.TBAA = nullptr;
971 Accesses.addLoad(Loc, IsReadOnlyPtr);
974 // If we write (or read-write) to a single destination and there are no
975 // other reads in this loop then is it safe to vectorize.
976 if (NumReadWrites == 1 && NumReads == 0) {
977 DEBUG(dbgs() << "LV: Found a write-only loop!\n");
981 // Build dependence sets and check whether we need a runtime pointer bounds
983 Accesses.buildDependenceSets();
984 bool NeedRTCheck = Accesses.isRTCheckNeeded();
986 // Find pointers with computable bounds. We are going to use this information
987 // to place a runtime bound check.
988 unsigned NumComparisons = 0;
989 bool CanDoRT = false;
991 CanDoRT = Accesses.canCheckPtrAtRT(PtrRtCheck, NumComparisons, SE, TheLoop,
994 DEBUG(dbgs() << "LV: We need to do " << NumComparisons <<
995 " pointer comparisons.\n");
997 // If we only have one set of dependences to check pointers among we don't
998 // need a runtime check.
999 if (NumComparisons == 0 && NeedRTCheck)
1000 NeedRTCheck = false;
1002 // Check that we did not collect too many pointers or found an unsizeable
1004 if (!CanDoRT || NumComparisons > VectParams.RuntimeMemoryCheckThreshold) {
1010 DEBUG(dbgs() << "LV: We can perform a memory runtime check if needed.\n");
1013 if (NeedRTCheck && !CanDoRT) {
1014 emitAnalysis(VectorizationReport() << "cannot identify array bounds");
1015 DEBUG(dbgs() << "LV: We can't vectorize because we can't find " <<
1016 "the array bounds.\n");
1021 PtrRtCheck.Need = NeedRTCheck;
1023 bool CanVecMem = true;
1024 if (Accesses.isDependencyCheckNeeded()) {
1025 DEBUG(dbgs() << "LV: Checking memory dependencies\n");
1026 CanVecMem = DepChecker.areDepsSafe(
1027 DependentAccesses, Accesses.getDependenciesToCheck(), Strides);
1028 MaxSafeDepDistBytes = DepChecker.getMaxSafeDepDistBytes();
1030 if (!CanVecMem && DepChecker.shouldRetryWithRuntimeCheck()) {
1031 DEBUG(dbgs() << "LV: Retrying with memory checks\n");
1034 // Clear the dependency checks. We assume they are not needed.
1035 Accesses.resetDepChecks();
1038 PtrRtCheck.Need = true;
1040 CanDoRT = Accesses.canCheckPtrAtRT(PtrRtCheck, NumComparisons, SE,
1041 TheLoop, Strides, true);
1042 // Check that we did not collect too many pointers or found an unsizeable
1044 if (!CanDoRT || NumComparisons > VectParams.RuntimeMemoryCheckThreshold) {
1045 if (!CanDoRT && NumComparisons > 0)
1046 emitAnalysis(VectorizationReport()
1047 << "cannot check memory dependencies at runtime");
1049 emitAnalysis(VectorizationReport()
1050 << NumComparisons << " exceeds limit of "
1051 << VectParams.RuntimeMemoryCheckThreshold
1052 << " dependent memory operations checked at runtime");
1053 DEBUG(dbgs() << "LV: Can't vectorize with memory checks\n");
1063 emitAnalysis(VectorizationReport() <<
1064 "unsafe dependent memory operations in loop");
1066 DEBUG(dbgs() << "LV: We" << (NeedRTCheck ? "" : " don't") <<
1067 " need a runtime memory check.\n");
1072 bool LoopAccessAnalysis::blockNeedsPredication(BasicBlock *BB) {
1073 assert(TheLoop->contains(BB) && "Unknown block used");
1075 // Blocks that do not dominate the latch need predication.
1076 BasicBlock* Latch = TheLoop->getLoopLatch();
1077 return !DT->dominates(BB, Latch);
1080 void LoopAccessAnalysis::emitAnalysis(VectorizationReport &Message) {
1081 VectorizationReport::emitAnalysis(Message, TheFunction, TheLoop);
1084 bool LoopAccessAnalysis::isUniform(Value *V) {
1085 return (SE->isLoopInvariant(SE->getSCEV(V), TheLoop));
1088 // FIXME: this function is currently a duplicate of the one in
1089 // LoopVectorize.cpp.
1090 static Instruction *getFirstInst(Instruction *FirstInst, Value *V,
1094 if (Instruction *I = dyn_cast<Instruction>(V))
1095 return I->getParent() == Loc->getParent() ? I : nullptr;
1099 std::pair<Instruction *, Instruction *>
1100 LoopAccessAnalysis::addRuntimeCheck(Instruction *Loc) {
1101 Instruction *tnullptr = nullptr;
1102 if (!PtrRtCheck.Need)
1103 return std::pair<Instruction *, Instruction *>(tnullptr, tnullptr);
1105 unsigned NumPointers = PtrRtCheck.Pointers.size();
1106 SmallVector<TrackingVH<Value> , 2> Starts;
1107 SmallVector<TrackingVH<Value> , 2> Ends;
1109 LLVMContext &Ctx = Loc->getContext();
1110 SCEVExpander Exp(*SE, "induction");
1111 Instruction *FirstInst = nullptr;
1113 for (unsigned i = 0; i < NumPointers; ++i) {
1114 Value *Ptr = PtrRtCheck.Pointers[i];
1115 const SCEV *Sc = SE->getSCEV(Ptr);
1117 if (SE->isLoopInvariant(Sc, TheLoop)) {
1118 DEBUG(dbgs() << "LV: Adding RT check for a loop invariant ptr:" <<
1120 Starts.push_back(Ptr);
1121 Ends.push_back(Ptr);
1123 DEBUG(dbgs() << "LV: Adding RT check for range:" << *Ptr << '\n');
1124 unsigned AS = Ptr->getType()->getPointerAddressSpace();
1126 // Use this type for pointer arithmetic.
1127 Type *PtrArithTy = Type::getInt8PtrTy(Ctx, AS);
1129 Value *Start = Exp.expandCodeFor(PtrRtCheck.Starts[i], PtrArithTy, Loc);
1130 Value *End = Exp.expandCodeFor(PtrRtCheck.Ends[i], PtrArithTy, Loc);
1131 Starts.push_back(Start);
1132 Ends.push_back(End);
1136 IRBuilder<> ChkBuilder(Loc);
1137 // Our instructions might fold to a constant.
1138 Value *MemoryRuntimeCheck = nullptr;
1139 for (unsigned i = 0; i < NumPointers; ++i) {
1140 for (unsigned j = i+1; j < NumPointers; ++j) {
1141 // No need to check if two readonly pointers intersect.
1142 if (!PtrRtCheck.IsWritePtr[i] && !PtrRtCheck.IsWritePtr[j])
1145 // Only need to check pointers between two different dependency sets.
1146 if (PtrRtCheck.DependencySetId[i] == PtrRtCheck.DependencySetId[j])
1148 // Only need to check pointers in the same alias set.
1149 if (PtrRtCheck.AliasSetId[i] != PtrRtCheck.AliasSetId[j])
1152 unsigned AS0 = Starts[i]->getType()->getPointerAddressSpace();
1153 unsigned AS1 = Starts[j]->getType()->getPointerAddressSpace();
1155 assert((AS0 == Ends[j]->getType()->getPointerAddressSpace()) &&
1156 (AS1 == Ends[i]->getType()->getPointerAddressSpace()) &&
1157 "Trying to bounds check pointers with different address spaces");
1159 Type *PtrArithTy0 = Type::getInt8PtrTy(Ctx, AS0);
1160 Type *PtrArithTy1 = Type::getInt8PtrTy(Ctx, AS1);
1162 Value *Start0 = ChkBuilder.CreateBitCast(Starts[i], PtrArithTy0, "bc");
1163 Value *Start1 = ChkBuilder.CreateBitCast(Starts[j], PtrArithTy1, "bc");
1164 Value *End0 = ChkBuilder.CreateBitCast(Ends[i], PtrArithTy1, "bc");
1165 Value *End1 = ChkBuilder.CreateBitCast(Ends[j], PtrArithTy0, "bc");
1167 Value *Cmp0 = ChkBuilder.CreateICmpULE(Start0, End1, "bound0");
1168 FirstInst = getFirstInst(FirstInst, Cmp0, Loc);
1169 Value *Cmp1 = ChkBuilder.CreateICmpULE(Start1, End0, "bound1");
1170 FirstInst = getFirstInst(FirstInst, Cmp1, Loc);
1171 Value *IsConflict = ChkBuilder.CreateAnd(Cmp0, Cmp1, "found.conflict");
1172 FirstInst = getFirstInst(FirstInst, IsConflict, Loc);
1173 if (MemoryRuntimeCheck) {
1174 IsConflict = ChkBuilder.CreateOr(MemoryRuntimeCheck, IsConflict,
1176 FirstInst = getFirstInst(FirstInst, IsConflict, Loc);
1178 MemoryRuntimeCheck = IsConflict;
1182 // We have to do this trickery because the IRBuilder might fold the check to a
1183 // constant expression in which case there is no Instruction anchored in a
1185 Instruction *Check = BinaryOperator::CreateAnd(MemoryRuntimeCheck,
1186 ConstantInt::getTrue(Ctx));
1187 ChkBuilder.Insert(Check, "memcheck.conflict");
1188 FirstInst = getFirstInst(FirstInst, Check, Loc);
1189 return std::make_pair(FirstInst, Check);