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 static cl::opt<unsigned, true>
29 VectorizationFactor("force-vector-width", cl::Hidden,
30 cl::desc("Sets the SIMD width. Zero is autoselect."),
31 cl::location(VectorizerParams::VectorizationFactor));
32 unsigned VectorizerParams::VectorizationFactor = 0;
34 static cl::opt<unsigned, true>
35 VectorizationInterleave("force-vector-interleave", cl::Hidden,
36 cl::desc("Sets the vectorization interleave count. "
37 "Zero is autoselect."),
39 VectorizerParams::VectorizationInterleave));
40 unsigned VectorizerParams::VectorizationInterleave = 0;
42 /// When performing memory disambiguation checks at runtime do not make more
43 /// than this number of comparisons.
44 const unsigned VectorizerParams::RuntimeMemoryCheckThreshold = 8;
46 /// Maximum SIMD width.
47 const unsigned VectorizerParams::MaxVectorWidth = 64;
49 bool VectorizerParams::isInterleaveForced() {
50 return ::VectorizationInterleave.getNumOccurrences() > 0;
53 void VectorizationReport::emitAnalysis(VectorizationReport &Message,
54 const Function *TheFunction,
55 const Loop *TheLoop) {
56 DebugLoc DL = TheLoop->getStartLoc();
57 if (Instruction *I = Message.getInstr())
58 DL = I->getDebugLoc();
59 emitOptimizationRemarkAnalysis(TheFunction->getContext(), DEBUG_TYPE,
60 *TheFunction, DL, Message.str());
63 Value *llvm::stripIntegerCast(Value *V) {
64 if (CastInst *CI = dyn_cast<CastInst>(V))
65 if (CI->getOperand(0)->getType()->isIntegerTy())
66 return CI->getOperand(0);
70 const SCEV *llvm::replaceSymbolicStrideSCEV(ScalarEvolution *SE,
71 ValueToValueMap &PtrToStride,
72 Value *Ptr, Value *OrigPtr) {
74 const SCEV *OrigSCEV = SE->getSCEV(Ptr);
76 // If there is an entry in the map return the SCEV of the pointer with the
77 // symbolic stride replaced by one.
78 ValueToValueMap::iterator SI = PtrToStride.find(OrigPtr ? OrigPtr : Ptr);
79 if (SI != PtrToStride.end()) {
80 Value *StrideVal = SI->second;
83 StrideVal = stripIntegerCast(StrideVal);
85 // Replace symbolic stride by one.
86 Value *One = ConstantInt::get(StrideVal->getType(), 1);
87 ValueToValueMap RewriteMap;
88 RewriteMap[StrideVal] = One;
91 SCEVParameterRewriter::rewrite(OrigSCEV, *SE, RewriteMap, true);
92 DEBUG(dbgs() << "LV: Replacing SCEV: " << *OrigSCEV << " by: " << *ByOne
97 // Otherwise, just return the SCEV of the original pointer.
98 return SE->getSCEV(Ptr);
101 void LoopAccessInfo::RuntimePointerCheck::insert(ScalarEvolution *SE, Loop *Lp,
102 Value *Ptr, bool WritePtr,
105 ValueToValueMap &Strides) {
106 // Get the stride replaced scev.
107 const SCEV *Sc = replaceSymbolicStrideSCEV(SE, Strides, Ptr);
108 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Sc);
109 assert(AR && "Invalid addrec expression");
110 const SCEV *Ex = SE->getBackedgeTakenCount(Lp);
111 const SCEV *ScEnd = AR->evaluateAtIteration(Ex, *SE);
112 Pointers.push_back(Ptr);
113 Starts.push_back(AR->getStart());
114 Ends.push_back(ScEnd);
115 IsWritePtr.push_back(WritePtr);
116 DependencySetId.push_back(DepSetId);
117 AliasSetId.push_back(ASId);
120 bool LoopAccessInfo::RuntimePointerCheck::needsChecking(unsigned I,
122 // No need to check if two readonly pointers intersect.
123 if (!IsWritePtr[I] && !IsWritePtr[J])
126 // Only need to check pointers between two different dependency sets.
127 if (DependencySetId[I] == DependencySetId[J])
130 // Only need to check pointers in the same alias set.
131 if (AliasSetId[I] != AliasSetId[J])
138 /// \brief Analyses memory accesses in a loop.
140 /// Checks whether run time pointer checks are needed and builds sets for data
141 /// dependence checking.
142 class AccessAnalysis {
144 /// \brief Read or write access location.
145 typedef PointerIntPair<Value *, 1, bool> MemAccessInfo;
146 typedef SmallPtrSet<MemAccessInfo, 8> MemAccessInfoSet;
148 /// \brief Set of potential dependent memory accesses.
149 typedef EquivalenceClasses<MemAccessInfo> DepCandidates;
151 AccessAnalysis(const DataLayout *Dl, AliasAnalysis *AA, DepCandidates &DA) :
152 DL(Dl), AST(*AA), DepCands(DA), IsRTCheckNeeded(false) {}
154 /// \brief Register a load and whether it is only read from.
155 void addLoad(AliasAnalysis::Location &Loc, bool IsReadOnly) {
156 Value *Ptr = const_cast<Value*>(Loc.Ptr);
157 AST.add(Ptr, AliasAnalysis::UnknownSize, Loc.AATags);
158 Accesses.insert(MemAccessInfo(Ptr, false));
160 ReadOnlyPtr.insert(Ptr);
163 /// \brief Register a store.
164 void addStore(AliasAnalysis::Location &Loc) {
165 Value *Ptr = const_cast<Value*>(Loc.Ptr);
166 AST.add(Ptr, AliasAnalysis::UnknownSize, Loc.AATags);
167 Accesses.insert(MemAccessInfo(Ptr, true));
170 /// \brief Check whether we can check the pointers at runtime for
171 /// non-intersection.
172 bool canCheckPtrAtRT(LoopAccessInfo::RuntimePointerCheck &RtCheck,
173 unsigned &NumComparisons,
174 ScalarEvolution *SE, Loop *TheLoop,
175 ValueToValueMap &Strides,
176 bool ShouldCheckStride = false);
178 /// \brief Goes over all memory accesses, checks whether a RT check is needed
179 /// and builds sets of dependent accesses.
180 void buildDependenceSets() {
181 processMemAccesses();
184 bool isRTCheckNeeded() { return IsRTCheckNeeded; }
186 bool isDependencyCheckNeeded() { return !CheckDeps.empty(); }
187 void resetDepChecks() { CheckDeps.clear(); }
189 MemAccessInfoSet &getDependenciesToCheck() { return CheckDeps; }
192 typedef SetVector<MemAccessInfo> PtrAccessSet;
194 /// \brief Go over all memory access and check whether runtime pointer checks
195 /// are needed /// and build sets of dependency check candidates.
196 void processMemAccesses();
198 /// Set of all accesses.
199 PtrAccessSet Accesses;
201 /// Set of accesses that need a further dependence check.
202 MemAccessInfoSet CheckDeps;
204 /// Set of pointers that are read only.
205 SmallPtrSet<Value*, 16> ReadOnlyPtr;
207 const DataLayout *DL;
209 /// An alias set tracker to partition the access set by underlying object and
210 //intrinsic property (such as TBAA metadata).
213 /// Sets of potentially dependent accesses - members of one set share an
214 /// underlying pointer. The set "CheckDeps" identfies which sets really need a
215 /// dependence check.
216 DepCandidates &DepCands;
218 bool IsRTCheckNeeded;
221 } // end anonymous namespace
223 /// \brief Check whether a pointer can participate in a runtime bounds check.
224 static bool hasComputableBounds(ScalarEvolution *SE, ValueToValueMap &Strides,
226 const SCEV *PtrScev = replaceSymbolicStrideSCEV(SE, Strides, Ptr);
227 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
231 return AR->isAffine();
234 /// \brief Check the stride of the pointer and ensure that it does not wrap in
235 /// the address space.
236 static int isStridedPtr(ScalarEvolution *SE, const DataLayout *DL, Value *Ptr,
237 const Loop *Lp, ValueToValueMap &StridesMap);
239 bool AccessAnalysis::canCheckPtrAtRT(
240 LoopAccessInfo::RuntimePointerCheck &RtCheck,
241 unsigned &NumComparisons, ScalarEvolution *SE, Loop *TheLoop,
242 ValueToValueMap &StridesMap, bool ShouldCheckStride) {
243 // Find pointers with computable bounds. We are going to use this information
244 // to place a runtime bound check.
247 bool IsDepCheckNeeded = isDependencyCheckNeeded();
250 // We assign a consecutive id to access from different alias sets.
251 // Accesses between different groups doesn't need to be checked.
253 for (auto &AS : AST) {
254 unsigned NumReadPtrChecks = 0;
255 unsigned NumWritePtrChecks = 0;
257 // We assign consecutive id to access from different dependence sets.
258 // Accesses within the same set don't need a runtime check.
259 unsigned RunningDepId = 1;
260 DenseMap<Value *, unsigned> DepSetId;
263 Value *Ptr = A.getValue();
264 bool IsWrite = Accesses.count(MemAccessInfo(Ptr, true));
265 MemAccessInfo Access(Ptr, IsWrite);
272 if (hasComputableBounds(SE, StridesMap, Ptr) &&
273 // When we run after a failing dependency check we have to make sure we
274 // don't have wrapping pointers.
275 (!ShouldCheckStride ||
276 isStridedPtr(SE, DL, Ptr, TheLoop, StridesMap) == 1)) {
277 // The id of the dependence set.
280 if (IsDepCheckNeeded) {
281 Value *Leader = DepCands.getLeaderValue(Access).getPointer();
282 unsigned &LeaderId = DepSetId[Leader];
284 LeaderId = RunningDepId++;
287 // Each access has its own dependence set.
288 DepId = RunningDepId++;
290 RtCheck.insert(SE, TheLoop, Ptr, IsWrite, DepId, ASId, StridesMap);
292 DEBUG(dbgs() << "LV: Found a runtime check ptr:" << *Ptr << '\n');
298 if (IsDepCheckNeeded && CanDoRT && RunningDepId == 2)
299 NumComparisons += 0; // Only one dependence set.
301 NumComparisons += (NumWritePtrChecks * (NumReadPtrChecks +
302 NumWritePtrChecks - 1));
308 // If the pointers that we would use for the bounds comparison have different
309 // address spaces, assume the values aren't directly comparable, so we can't
310 // use them for the runtime check. We also have to assume they could
311 // overlap. In the future there should be metadata for whether address spaces
313 unsigned NumPointers = RtCheck.Pointers.size();
314 for (unsigned i = 0; i < NumPointers; ++i) {
315 for (unsigned j = i + 1; j < NumPointers; ++j) {
316 // Only need to check pointers between two different dependency sets.
317 if (RtCheck.DependencySetId[i] == RtCheck.DependencySetId[j])
319 // Only need to check pointers in the same alias set.
320 if (RtCheck.AliasSetId[i] != RtCheck.AliasSetId[j])
323 Value *PtrI = RtCheck.Pointers[i];
324 Value *PtrJ = RtCheck.Pointers[j];
326 unsigned ASi = PtrI->getType()->getPointerAddressSpace();
327 unsigned ASj = PtrJ->getType()->getPointerAddressSpace();
329 DEBUG(dbgs() << "LV: Runtime check would require comparison between"
330 " different address spaces\n");
339 void AccessAnalysis::processMemAccesses() {
340 // We process the set twice: first we process read-write pointers, last we
341 // process read-only pointers. This allows us to skip dependence tests for
342 // read-only pointers.
344 DEBUG(dbgs() << "LV: Processing memory accesses...\n");
345 DEBUG(dbgs() << " AST: "; AST.dump());
346 DEBUG(dbgs() << "LV: Accesses:\n");
348 for (auto A : Accesses)
349 dbgs() << "\t" << *A.getPointer() << " (" <<
350 (A.getInt() ? "write" : (ReadOnlyPtr.count(A.getPointer()) ?
351 "read-only" : "read")) << ")\n";
354 // The AliasSetTracker has nicely partitioned our pointers by metadata
355 // compatibility and potential for underlying-object overlap. As a result, we
356 // only need to check for potential pointer dependencies within each alias
358 for (auto &AS : AST) {
359 // Note that both the alias-set tracker and the alias sets themselves used
360 // linked lists internally and so the iteration order here is deterministic
361 // (matching the original instruction order within each set).
363 bool SetHasWrite = false;
365 // Map of pointers to last access encountered.
366 typedef DenseMap<Value*, MemAccessInfo> UnderlyingObjToAccessMap;
367 UnderlyingObjToAccessMap ObjToLastAccess;
369 // Set of access to check after all writes have been processed.
370 PtrAccessSet DeferredAccesses;
372 // Iterate over each alias set twice, once to process read/write pointers,
373 // and then to process read-only pointers.
374 for (int SetIteration = 0; SetIteration < 2; ++SetIteration) {
375 bool UseDeferred = SetIteration > 0;
376 PtrAccessSet &S = UseDeferred ? DeferredAccesses : Accesses;
379 Value *Ptr = AV.getValue();
381 // For a single memory access in AliasSetTracker, Accesses may contain
382 // both read and write, and they both need to be handled for CheckDeps.
384 if (AC.getPointer() != Ptr)
387 bool IsWrite = AC.getInt();
389 // If we're using the deferred access set, then it contains only
391 bool IsReadOnlyPtr = ReadOnlyPtr.count(Ptr) && !IsWrite;
392 if (UseDeferred && !IsReadOnlyPtr)
394 // Otherwise, the pointer must be in the PtrAccessSet, either as a
396 assert(((IsReadOnlyPtr && UseDeferred) || IsWrite ||
397 S.count(MemAccessInfo(Ptr, false))) &&
398 "Alias-set pointer not in the access set?");
400 MemAccessInfo Access(Ptr, IsWrite);
401 DepCands.insert(Access);
403 // Memorize read-only pointers for later processing and skip them in
404 // the first round (they need to be checked after we have seen all
405 // write pointers). Note: we also mark pointer that are not
406 // consecutive as "read-only" pointers (so that we check
407 // "a[b[i]] +="). Hence, we need the second check for "!IsWrite".
408 if (!UseDeferred && IsReadOnlyPtr) {
409 DeferredAccesses.insert(Access);
413 // If this is a write - check other reads and writes for conflicts. If
414 // this is a read only check other writes for conflicts (but only if
415 // there is no other write to the ptr - this is an optimization to
416 // catch "a[i] = a[i] + " without having to do a dependence check).
417 if ((IsWrite || IsReadOnlyPtr) && SetHasWrite) {
418 CheckDeps.insert(Access);
419 IsRTCheckNeeded = true;
425 // Create sets of pointers connected by a shared alias set and
426 // underlying object.
427 typedef SmallVector<Value *, 16> ValueVector;
428 ValueVector TempObjects;
429 GetUnderlyingObjects(Ptr, TempObjects, DL);
430 for (Value *UnderlyingObj : TempObjects) {
431 UnderlyingObjToAccessMap::iterator Prev =
432 ObjToLastAccess.find(UnderlyingObj);
433 if (Prev != ObjToLastAccess.end())
434 DepCands.unionSets(Access, Prev->second);
436 ObjToLastAccess[UnderlyingObj] = Access;
445 /// \brief Checks memory dependences among accesses to the same underlying
446 /// object to determine whether there vectorization is legal or not (and at
447 /// which vectorization factor).
449 /// This class works under the assumption that we already checked that memory
450 /// locations with different underlying pointers are "must-not alias".
451 /// We use the ScalarEvolution framework to symbolically evalutate access
452 /// functions pairs. Since we currently don't restructure the loop we can rely
453 /// on the program order of memory accesses to determine their safety.
454 /// At the moment we will only deem accesses as safe for:
455 /// * A negative constant distance assuming program order.
457 /// Safe: tmp = a[i + 1]; OR a[i + 1] = x;
458 /// a[i] = tmp; y = a[i];
460 /// The latter case is safe because later checks guarantuee that there can't
461 /// be a cycle through a phi node (that is, we check that "x" and "y" is not
462 /// the same variable: a header phi can only be an induction or a reduction, a
463 /// reduction can't have a memory sink, an induction can't have a memory
464 /// source). This is important and must not be violated (or we have to
465 /// resort to checking for cycles through memory).
467 /// * A positive constant distance assuming program order that is bigger
468 /// than the biggest memory access.
470 /// tmp = a[i] OR b[i] = x
471 /// a[i+2] = tmp y = b[i+2];
473 /// Safe distance: 2 x sizeof(a[0]), and 2 x sizeof(b[0]), respectively.
475 /// * Zero distances and all accesses have the same size.
477 class MemoryDepChecker {
479 typedef PointerIntPair<Value *, 1, bool> MemAccessInfo;
480 typedef SmallPtrSet<MemAccessInfo, 8> MemAccessInfoSet;
482 MemoryDepChecker(ScalarEvolution *Se, const DataLayout *Dl, const Loop *L)
483 : SE(Se), DL(Dl), InnermostLoop(L), AccessIdx(0),
484 ShouldRetryWithRuntimeCheck(false) {}
486 /// \brief Register the location (instructions are given increasing numbers)
487 /// of a write access.
488 void addAccess(StoreInst *SI) {
489 Value *Ptr = SI->getPointerOperand();
490 Accesses[MemAccessInfo(Ptr, true)].push_back(AccessIdx);
491 InstMap.push_back(SI);
495 /// \brief Register the location (instructions are given increasing numbers)
496 /// of a write access.
497 void addAccess(LoadInst *LI) {
498 Value *Ptr = LI->getPointerOperand();
499 Accesses[MemAccessInfo(Ptr, false)].push_back(AccessIdx);
500 InstMap.push_back(LI);
504 /// \brief Check whether the dependencies between the accesses are safe.
506 /// Only checks sets with elements in \p CheckDeps.
507 bool areDepsSafe(AccessAnalysis::DepCandidates &AccessSets,
508 MemAccessInfoSet &CheckDeps, ValueToValueMap &Strides);
510 /// \brief The maximum number of bytes of a vector register we can vectorize
511 /// the accesses safely with.
512 unsigned getMaxSafeDepDistBytes() { return MaxSafeDepDistBytes; }
514 /// \brief In same cases when the dependency check fails we can still
515 /// vectorize the loop with a dynamic array access check.
516 bool shouldRetryWithRuntimeCheck() { return ShouldRetryWithRuntimeCheck; }
520 const DataLayout *DL;
521 const Loop *InnermostLoop;
523 /// \brief Maps access locations (ptr, read/write) to program order.
524 DenseMap<MemAccessInfo, std::vector<unsigned> > Accesses;
526 /// \brief Memory access instructions in program order.
527 SmallVector<Instruction *, 16> InstMap;
529 /// \brief The program order index to be used for the next instruction.
532 // We can access this many bytes in parallel safely.
533 unsigned MaxSafeDepDistBytes;
535 /// \brief If we see a non-constant dependence distance we can still try to
536 /// vectorize this loop with runtime checks.
537 bool ShouldRetryWithRuntimeCheck;
539 /// \brief Check whether there is a plausible dependence between the two
542 /// Access \p A must happen before \p B in program order. The two indices
543 /// identify the index into the program order map.
545 /// This function checks whether there is a plausible dependence (or the
546 /// absence of such can't be proved) between the two accesses. If there is a
547 /// plausible dependence but the dependence distance is bigger than one
548 /// element access it records this distance in \p MaxSafeDepDistBytes (if this
549 /// distance is smaller than any other distance encountered so far).
550 /// Otherwise, this function returns true signaling a possible dependence.
551 bool isDependent(const MemAccessInfo &A, unsigned AIdx,
552 const MemAccessInfo &B, unsigned BIdx,
553 ValueToValueMap &Strides);
555 /// \brief Check whether the data dependence could prevent store-load
557 bool couldPreventStoreLoadForward(unsigned Distance, unsigned TypeByteSize);
560 } // end anonymous namespace
562 static bool isInBoundsGep(Value *Ptr) {
563 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr))
564 return GEP->isInBounds();
568 /// \brief Check whether the access through \p Ptr has a constant stride.
569 static int isStridedPtr(ScalarEvolution *SE, const DataLayout *DL, Value *Ptr,
570 const Loop *Lp, ValueToValueMap &StridesMap) {
571 const Type *Ty = Ptr->getType();
572 assert(Ty->isPointerTy() && "Unexpected non-ptr");
574 // Make sure that the pointer does not point to aggregate types.
575 const PointerType *PtrTy = cast<PointerType>(Ty);
576 if (PtrTy->getElementType()->isAggregateType()) {
577 DEBUG(dbgs() << "LV: Bad stride - Not a pointer to a scalar type" << *Ptr <<
582 const SCEV *PtrScev = replaceSymbolicStrideSCEV(SE, StridesMap, Ptr);
584 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
586 DEBUG(dbgs() << "LV: Bad stride - Not an AddRecExpr pointer "
587 << *Ptr << " SCEV: " << *PtrScev << "\n");
591 // The accesss function must stride over the innermost loop.
592 if (Lp != AR->getLoop()) {
593 DEBUG(dbgs() << "LV: Bad stride - Not striding over innermost loop " <<
594 *Ptr << " SCEV: " << *PtrScev << "\n");
597 // The address calculation must not wrap. Otherwise, a dependence could be
599 // An inbounds getelementptr that is a AddRec with a unit stride
600 // cannot wrap per definition. The unit stride requirement is checked later.
601 // An getelementptr without an inbounds attribute and unit stride would have
602 // to access the pointer value "0" which is undefined behavior in address
603 // space 0, therefore we can also vectorize this case.
604 bool IsInBoundsGEP = isInBoundsGep(Ptr);
605 bool IsNoWrapAddRec = AR->getNoWrapFlags(SCEV::NoWrapMask);
606 bool IsInAddressSpaceZero = PtrTy->getAddressSpace() == 0;
607 if (!IsNoWrapAddRec && !IsInBoundsGEP && !IsInAddressSpaceZero) {
608 DEBUG(dbgs() << "LV: Bad stride - Pointer may wrap in the address space "
609 << *Ptr << " SCEV: " << *PtrScev << "\n");
613 // Check the step is constant.
614 const SCEV *Step = AR->getStepRecurrence(*SE);
616 // Calculate the pointer stride and check if it is consecutive.
617 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
619 DEBUG(dbgs() << "LV: Bad stride - Not a constant strided " << *Ptr <<
620 " SCEV: " << *PtrScev << "\n");
624 int64_t Size = DL->getTypeAllocSize(PtrTy->getElementType());
625 const APInt &APStepVal = C->getValue()->getValue();
627 // Huge step value - give up.
628 if (APStepVal.getBitWidth() > 64)
631 int64_t StepVal = APStepVal.getSExtValue();
634 int64_t Stride = StepVal / Size;
635 int64_t Rem = StepVal % Size;
639 // If the SCEV could wrap but we have an inbounds gep with a unit stride we
640 // know we can't "wrap around the address space". In case of address space
641 // zero we know that this won't happen without triggering undefined behavior.
642 if (!IsNoWrapAddRec && (IsInBoundsGEP || IsInAddressSpaceZero) &&
643 Stride != 1 && Stride != -1)
649 bool MemoryDepChecker::couldPreventStoreLoadForward(unsigned Distance,
650 unsigned TypeByteSize) {
651 // If loads occur at a distance that is not a multiple of a feasible vector
652 // factor store-load forwarding does not take place.
653 // Positive dependences might cause troubles because vectorizing them might
654 // prevent store-load forwarding making vectorized code run a lot slower.
655 // a[i] = a[i-3] ^ a[i-8];
656 // The stores to a[i:i+1] don't align with the stores to a[i-3:i-2] and
657 // hence on your typical architecture store-load forwarding does not take
658 // place. Vectorizing in such cases does not make sense.
659 // Store-load forwarding distance.
660 const unsigned NumCyclesForStoreLoadThroughMemory = 8*TypeByteSize;
661 // Maximum vector factor.
662 unsigned MaxVFWithoutSLForwardIssues =
663 VectorizerParams::MaxVectorWidth * TypeByteSize;
664 if(MaxSafeDepDistBytes < MaxVFWithoutSLForwardIssues)
665 MaxVFWithoutSLForwardIssues = MaxSafeDepDistBytes;
667 for (unsigned vf = 2*TypeByteSize; vf <= MaxVFWithoutSLForwardIssues;
669 if (Distance % vf && Distance / vf < NumCyclesForStoreLoadThroughMemory) {
670 MaxVFWithoutSLForwardIssues = (vf >>=1);
675 if (MaxVFWithoutSLForwardIssues< 2*TypeByteSize) {
676 DEBUG(dbgs() << "LV: Distance " << Distance <<
677 " that could cause a store-load forwarding conflict\n");
681 if (MaxVFWithoutSLForwardIssues < MaxSafeDepDistBytes &&
682 MaxVFWithoutSLForwardIssues !=
683 VectorizerParams::MaxVectorWidth * TypeByteSize)
684 MaxSafeDepDistBytes = MaxVFWithoutSLForwardIssues;
688 bool MemoryDepChecker::isDependent(const MemAccessInfo &A, unsigned AIdx,
689 const MemAccessInfo &B, unsigned BIdx,
690 ValueToValueMap &Strides) {
691 assert (AIdx < BIdx && "Must pass arguments in program order");
693 Value *APtr = A.getPointer();
694 Value *BPtr = B.getPointer();
695 bool AIsWrite = A.getInt();
696 bool BIsWrite = B.getInt();
698 // Two reads are independent.
699 if (!AIsWrite && !BIsWrite)
702 // We cannot check pointers in different address spaces.
703 if (APtr->getType()->getPointerAddressSpace() !=
704 BPtr->getType()->getPointerAddressSpace())
707 const SCEV *AScev = replaceSymbolicStrideSCEV(SE, Strides, APtr);
708 const SCEV *BScev = replaceSymbolicStrideSCEV(SE, Strides, BPtr);
710 int StrideAPtr = isStridedPtr(SE, DL, APtr, InnermostLoop, Strides);
711 int StrideBPtr = isStridedPtr(SE, DL, BPtr, InnermostLoop, Strides);
713 const SCEV *Src = AScev;
714 const SCEV *Sink = BScev;
716 // If the induction step is negative we have to invert source and sink of the
718 if (StrideAPtr < 0) {
721 std::swap(APtr, BPtr);
722 std::swap(Src, Sink);
723 std::swap(AIsWrite, BIsWrite);
724 std::swap(AIdx, BIdx);
725 std::swap(StrideAPtr, StrideBPtr);
728 const SCEV *Dist = SE->getMinusSCEV(Sink, Src);
730 DEBUG(dbgs() << "LV: Src Scev: " << *Src << "Sink Scev: " << *Sink
731 << "(Induction step: " << StrideAPtr << ")\n");
732 DEBUG(dbgs() << "LV: Distance for " << *InstMap[AIdx] << " to "
733 << *InstMap[BIdx] << ": " << *Dist << "\n");
735 // Need consecutive accesses. We don't want to vectorize
736 // "A[B[i]] += ..." and similar code or pointer arithmetic that could wrap in
737 // the address space.
738 if (!StrideAPtr || !StrideBPtr || StrideAPtr != StrideBPtr){
739 DEBUG(dbgs() << "Non-consecutive pointer access\n");
743 const SCEVConstant *C = dyn_cast<SCEVConstant>(Dist);
745 DEBUG(dbgs() << "LV: Dependence because of non-constant distance\n");
746 ShouldRetryWithRuntimeCheck = true;
750 Type *ATy = APtr->getType()->getPointerElementType();
751 Type *BTy = BPtr->getType()->getPointerElementType();
752 unsigned TypeByteSize = DL->getTypeAllocSize(ATy);
754 // Negative distances are not plausible dependencies.
755 const APInt &Val = C->getValue()->getValue();
756 if (Val.isNegative()) {
757 bool IsTrueDataDependence = (AIsWrite && !BIsWrite);
758 if (IsTrueDataDependence &&
759 (couldPreventStoreLoadForward(Val.abs().getZExtValue(), TypeByteSize) ||
763 DEBUG(dbgs() << "LV: Dependence is negative: NoDep\n");
767 // Write to the same location with the same size.
768 // Could be improved to assert type sizes are the same (i32 == float, etc).
772 DEBUG(dbgs() << "LV: Zero dependence difference but different types\n");
776 assert(Val.isStrictlyPositive() && "Expect a positive value");
778 // Positive distance bigger than max vectorization factor.
781 "LV: ReadWrite-Write positive dependency with different types\n");
785 unsigned Distance = (unsigned) Val.getZExtValue();
787 // Bail out early if passed-in parameters make vectorization not feasible.
788 unsigned ForcedFactor = (VectorizerParams::VectorizationFactor ?
789 VectorizerParams::VectorizationFactor : 1);
790 unsigned ForcedUnroll = (VectorizerParams::VectorizationInterleave ?
791 VectorizerParams::VectorizationInterleave : 1);
793 // The distance must be bigger than the size needed for a vectorized version
794 // of the operation and the size of the vectorized operation must not be
795 // bigger than the currrent maximum size.
796 if (Distance < 2*TypeByteSize ||
797 2*TypeByteSize > MaxSafeDepDistBytes ||
798 Distance < TypeByteSize * ForcedUnroll * ForcedFactor) {
799 DEBUG(dbgs() << "LV: Failure because of Positive distance "
800 << Val.getSExtValue() << '\n');
804 MaxSafeDepDistBytes = Distance < MaxSafeDepDistBytes ?
805 Distance : MaxSafeDepDistBytes;
807 bool IsTrueDataDependence = (!AIsWrite && BIsWrite);
808 if (IsTrueDataDependence &&
809 couldPreventStoreLoadForward(Distance, TypeByteSize))
812 DEBUG(dbgs() << "LV: Positive distance " << Val.getSExtValue() <<
813 " with max VF = " << MaxSafeDepDistBytes / TypeByteSize << '\n');
818 bool MemoryDepChecker::areDepsSafe(AccessAnalysis::DepCandidates &AccessSets,
819 MemAccessInfoSet &CheckDeps,
820 ValueToValueMap &Strides) {
822 MaxSafeDepDistBytes = -1U;
823 while (!CheckDeps.empty()) {
824 MemAccessInfo CurAccess = *CheckDeps.begin();
826 // Get the relevant memory access set.
827 EquivalenceClasses<MemAccessInfo>::iterator I =
828 AccessSets.findValue(AccessSets.getLeaderValue(CurAccess));
830 // Check accesses within this set.
831 EquivalenceClasses<MemAccessInfo>::member_iterator AI, AE;
832 AI = AccessSets.member_begin(I), AE = AccessSets.member_end();
834 // Check every access pair.
836 CheckDeps.erase(*AI);
837 EquivalenceClasses<MemAccessInfo>::member_iterator OI = std::next(AI);
839 // Check every accessing instruction pair in program order.
840 for (std::vector<unsigned>::iterator I1 = Accesses[*AI].begin(),
841 I1E = Accesses[*AI].end(); I1 != I1E; ++I1)
842 for (std::vector<unsigned>::iterator I2 = Accesses[*OI].begin(),
843 I2E = Accesses[*OI].end(); I2 != I2E; ++I2) {
844 if (*I1 < *I2 && isDependent(*AI, *I1, *OI, *I2, Strides))
846 if (*I2 < *I1 && isDependent(*OI, *I2, *AI, *I1, Strides))
857 bool LoopAccessInfo::canVectorizeMemory(ValueToValueMap &Strides) {
859 typedef SmallVector<Value*, 16> ValueVector;
860 typedef SmallPtrSet<Value*, 16> ValueSet;
862 // Holds the Load and Store *instructions*.
866 // Holds all the different accesses in the loop.
867 unsigned NumReads = 0;
868 unsigned NumReadWrites = 0;
870 PtrRtCheck.Pointers.clear();
871 PtrRtCheck.Need = false;
873 const bool IsAnnotatedParallel = TheLoop->isAnnotatedParallel();
874 MemoryDepChecker DepChecker(SE, DL, TheLoop);
877 for (Loop::block_iterator bb = TheLoop->block_begin(),
878 be = TheLoop->block_end(); bb != be; ++bb) {
880 // Scan the BB and collect legal loads and stores.
881 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
884 // If this is a load, save it. If this instruction can read from memory
885 // but is not a load, then we quit. Notice that we don't handle function
886 // calls that read or write.
887 if (it->mayReadFromMemory()) {
888 // Many math library functions read the rounding mode. We will only
889 // vectorize a loop if it contains known function calls that don't set
890 // the flag. Therefore, it is safe to ignore this read from memory.
891 CallInst *Call = dyn_cast<CallInst>(it);
892 if (Call && getIntrinsicIDForCall(Call, TLI))
895 LoadInst *Ld = dyn_cast<LoadInst>(it);
896 if (!Ld || (!Ld->isSimple() && !IsAnnotatedParallel)) {
897 emitAnalysis(VectorizationReport(Ld)
898 << "read with atomic ordering or volatile read");
899 DEBUG(dbgs() << "LV: Found a non-simple load.\n");
904 DepChecker.addAccess(Ld);
908 // Save 'store' instructions. Abort if other instructions write to memory.
909 if (it->mayWriteToMemory()) {
910 StoreInst *St = dyn_cast<StoreInst>(it);
912 emitAnalysis(VectorizationReport(it) <<
913 "instruction cannot be vectorized");
916 if (!St->isSimple() && !IsAnnotatedParallel) {
917 emitAnalysis(VectorizationReport(St)
918 << "write with atomic ordering or volatile write");
919 DEBUG(dbgs() << "LV: Found a non-simple store.\n");
923 Stores.push_back(St);
924 DepChecker.addAccess(St);
929 // Now we have two lists that hold the loads and the stores.
930 // Next, we find the pointers that they use.
932 // Check if we see any stores. If there are no stores, then we don't
933 // care if the pointers are *restrict*.
934 if (!Stores.size()) {
935 DEBUG(dbgs() << "LV: Found a read-only loop!\n");
939 AccessAnalysis::DepCandidates DependentAccesses;
940 AccessAnalysis Accesses(DL, AA, DependentAccesses);
942 // Holds the analyzed pointers. We don't want to call GetUnderlyingObjects
943 // multiple times on the same object. If the ptr is accessed twice, once
944 // for read and once for write, it will only appear once (on the write
945 // list). This is okay, since we are going to check for conflicts between
946 // writes and between reads and writes, but not between reads and reads.
949 ValueVector::iterator I, IE;
950 for (I = Stores.begin(), IE = Stores.end(); I != IE; ++I) {
951 StoreInst *ST = cast<StoreInst>(*I);
952 Value* Ptr = ST->getPointerOperand();
954 if (isUniform(Ptr)) {
956 VectorizationReport(ST)
957 << "write to a loop invariant address could not be vectorized");
958 DEBUG(dbgs() << "LV: We don't allow storing to uniform addresses\n");
962 // If we did *not* see this pointer before, insert it to the read-write
963 // list. At this phase it is only a 'write' list.
964 if (Seen.insert(Ptr).second) {
967 AliasAnalysis::Location Loc = AA->getLocation(ST);
968 // The TBAA metadata could have a control dependency on the predication
969 // condition, so we cannot rely on it when determining whether or not we
970 // need runtime pointer checks.
971 if (blockNeedsPredication(ST->getParent(), TheLoop, DT))
972 Loc.AATags.TBAA = nullptr;
974 Accesses.addStore(Loc);
978 if (IsAnnotatedParallel) {
980 << "LV: A loop annotated parallel, ignore memory dependency "
985 for (I = Loads.begin(), IE = Loads.end(); I != IE; ++I) {
986 LoadInst *LD = cast<LoadInst>(*I);
987 Value* Ptr = LD->getPointerOperand();
988 // If we did *not* see this pointer before, insert it to the
989 // read list. If we *did* see it before, then it is already in
990 // the read-write list. This allows us to vectorize expressions
991 // such as A[i] += x; Because the address of A[i] is a read-write
992 // pointer. This only works if the index of A[i] is consecutive.
993 // If the address of i is unknown (for example A[B[i]]) then we may
994 // read a few words, modify, and write a few words, and some of the
995 // words may be written to the same address.
996 bool IsReadOnlyPtr = false;
997 if (Seen.insert(Ptr).second ||
998 !isStridedPtr(SE, DL, Ptr, TheLoop, Strides)) {
1000 IsReadOnlyPtr = true;
1003 AliasAnalysis::Location Loc = AA->getLocation(LD);
1004 // The TBAA metadata could have a control dependency on the predication
1005 // condition, so we cannot rely on it when determining whether or not we
1006 // need runtime pointer checks.
1007 if (blockNeedsPredication(LD->getParent(), TheLoop, DT))
1008 Loc.AATags.TBAA = nullptr;
1010 Accesses.addLoad(Loc, IsReadOnlyPtr);
1013 // If we write (or read-write) to a single destination and there are no
1014 // other reads in this loop then is it safe to vectorize.
1015 if (NumReadWrites == 1 && NumReads == 0) {
1016 DEBUG(dbgs() << "LV: Found a write-only loop!\n");
1020 // Build dependence sets and check whether we need a runtime pointer bounds
1022 Accesses.buildDependenceSets();
1023 bool NeedRTCheck = Accesses.isRTCheckNeeded();
1025 // Find pointers with computable bounds. We are going to use this information
1026 // to place a runtime bound check.
1027 unsigned NumComparisons = 0;
1028 bool CanDoRT = false;
1030 CanDoRT = Accesses.canCheckPtrAtRT(PtrRtCheck, NumComparisons, SE, TheLoop,
1033 DEBUG(dbgs() << "LV: We need to do " << NumComparisons <<
1034 " pointer comparisons.\n");
1036 // If we only have one set of dependences to check pointers among we don't
1037 // need a runtime check.
1038 if (NumComparisons == 0 && NeedRTCheck)
1039 NeedRTCheck = false;
1041 // Check that we did not collect too many pointers or found an unsizeable
1044 NumComparisons > VectorizerParams::RuntimeMemoryCheckThreshold) {
1050 DEBUG(dbgs() << "LV: We can perform a memory runtime check if needed.\n");
1053 if (NeedRTCheck && !CanDoRT) {
1054 emitAnalysis(VectorizationReport() << "cannot identify array bounds");
1055 DEBUG(dbgs() << "LV: We can't vectorize because we can't find " <<
1056 "the array bounds.\n");
1061 PtrRtCheck.Need = NeedRTCheck;
1063 bool CanVecMem = true;
1064 if (Accesses.isDependencyCheckNeeded()) {
1065 DEBUG(dbgs() << "LV: Checking memory dependencies\n");
1066 CanVecMem = DepChecker.areDepsSafe(
1067 DependentAccesses, Accesses.getDependenciesToCheck(), Strides);
1068 MaxSafeDepDistBytes = DepChecker.getMaxSafeDepDistBytes();
1070 if (!CanVecMem && DepChecker.shouldRetryWithRuntimeCheck()) {
1071 DEBUG(dbgs() << "LV: Retrying with memory checks\n");
1074 // Clear the dependency checks. We assume they are not needed.
1075 Accesses.resetDepChecks();
1078 PtrRtCheck.Need = true;
1080 CanDoRT = Accesses.canCheckPtrAtRT(PtrRtCheck, NumComparisons, SE,
1081 TheLoop, Strides, true);
1082 // Check that we did not collect too many pointers or found an unsizeable
1085 NumComparisons > VectorizerParams::RuntimeMemoryCheckThreshold) {
1086 if (!CanDoRT && NumComparisons > 0)
1087 emitAnalysis(VectorizationReport()
1088 << "cannot check memory dependencies at runtime");
1090 emitAnalysis(VectorizationReport()
1091 << NumComparisons << " exceeds limit of "
1092 << VectorizerParams::RuntimeMemoryCheckThreshold
1093 << " dependent memory operations checked at runtime");
1094 DEBUG(dbgs() << "LV: Can't vectorize with memory checks\n");
1104 emitAnalysis(VectorizationReport() <<
1105 "unsafe dependent memory operations in loop");
1107 DEBUG(dbgs() << "LV: We" << (NeedRTCheck ? "" : " don't") <<
1108 " need a runtime memory check.\n");
1113 bool LoopAccessInfo::blockNeedsPredication(BasicBlock *BB, Loop *TheLoop,
1114 DominatorTree *DT) {
1115 assert(TheLoop->contains(BB) && "Unknown block used");
1117 // Blocks that do not dominate the latch need predication.
1118 BasicBlock* Latch = TheLoop->getLoopLatch();
1119 return !DT->dominates(BB, Latch);
1122 void LoopAccessInfo::emitAnalysis(VectorizationReport &Message) {
1123 assert(!Report && "Multiple reports generated");
1127 bool LoopAccessInfo::isUniform(Value *V) {
1128 return (SE->isLoopInvariant(SE->getSCEV(V), TheLoop));
1131 // FIXME: this function is currently a duplicate of the one in
1132 // LoopVectorize.cpp.
1133 static Instruction *getFirstInst(Instruction *FirstInst, Value *V,
1137 if (Instruction *I = dyn_cast<Instruction>(V))
1138 return I->getParent() == Loc->getParent() ? I : nullptr;
1142 std::pair<Instruction *, Instruction *>
1143 LoopAccessInfo::addRuntimeCheck(Instruction *Loc) {
1144 Instruction *tnullptr = nullptr;
1145 if (!PtrRtCheck.Need)
1146 return std::pair<Instruction *, Instruction *>(tnullptr, tnullptr);
1148 unsigned NumPointers = PtrRtCheck.Pointers.size();
1149 SmallVector<TrackingVH<Value> , 2> Starts;
1150 SmallVector<TrackingVH<Value> , 2> Ends;
1152 LLVMContext &Ctx = Loc->getContext();
1153 SCEVExpander Exp(*SE, "induction");
1154 Instruction *FirstInst = nullptr;
1156 for (unsigned i = 0; i < NumPointers; ++i) {
1157 Value *Ptr = PtrRtCheck.Pointers[i];
1158 const SCEV *Sc = SE->getSCEV(Ptr);
1160 if (SE->isLoopInvariant(Sc, TheLoop)) {
1161 DEBUG(dbgs() << "LV: Adding RT check for a loop invariant ptr:" <<
1163 Starts.push_back(Ptr);
1164 Ends.push_back(Ptr);
1166 DEBUG(dbgs() << "LV: Adding RT check for range:" << *Ptr << '\n');
1167 unsigned AS = Ptr->getType()->getPointerAddressSpace();
1169 // Use this type for pointer arithmetic.
1170 Type *PtrArithTy = Type::getInt8PtrTy(Ctx, AS);
1172 Value *Start = Exp.expandCodeFor(PtrRtCheck.Starts[i], PtrArithTy, Loc);
1173 Value *End = Exp.expandCodeFor(PtrRtCheck.Ends[i], PtrArithTy, Loc);
1174 Starts.push_back(Start);
1175 Ends.push_back(End);
1179 IRBuilder<> ChkBuilder(Loc);
1180 // Our instructions might fold to a constant.
1181 Value *MemoryRuntimeCheck = nullptr;
1182 for (unsigned i = 0; i < NumPointers; ++i) {
1183 for (unsigned j = i+1; j < NumPointers; ++j) {
1184 if (!PtrRtCheck.needsChecking(i, j))
1187 unsigned AS0 = Starts[i]->getType()->getPointerAddressSpace();
1188 unsigned AS1 = Starts[j]->getType()->getPointerAddressSpace();
1190 assert((AS0 == Ends[j]->getType()->getPointerAddressSpace()) &&
1191 (AS1 == Ends[i]->getType()->getPointerAddressSpace()) &&
1192 "Trying to bounds check pointers with different address spaces");
1194 Type *PtrArithTy0 = Type::getInt8PtrTy(Ctx, AS0);
1195 Type *PtrArithTy1 = Type::getInt8PtrTy(Ctx, AS1);
1197 Value *Start0 = ChkBuilder.CreateBitCast(Starts[i], PtrArithTy0, "bc");
1198 Value *Start1 = ChkBuilder.CreateBitCast(Starts[j], PtrArithTy1, "bc");
1199 Value *End0 = ChkBuilder.CreateBitCast(Ends[i], PtrArithTy1, "bc");
1200 Value *End1 = ChkBuilder.CreateBitCast(Ends[j], PtrArithTy0, "bc");
1202 Value *Cmp0 = ChkBuilder.CreateICmpULE(Start0, End1, "bound0");
1203 FirstInst = getFirstInst(FirstInst, Cmp0, Loc);
1204 Value *Cmp1 = ChkBuilder.CreateICmpULE(Start1, End0, "bound1");
1205 FirstInst = getFirstInst(FirstInst, Cmp1, Loc);
1206 Value *IsConflict = ChkBuilder.CreateAnd(Cmp0, Cmp1, "found.conflict");
1207 FirstInst = getFirstInst(FirstInst, IsConflict, Loc);
1208 if (MemoryRuntimeCheck) {
1209 IsConflict = ChkBuilder.CreateOr(MemoryRuntimeCheck, IsConflict,
1211 FirstInst = getFirstInst(FirstInst, IsConflict, Loc);
1213 MemoryRuntimeCheck = IsConflict;
1217 // We have to do this trickery because the IRBuilder might fold the check to a
1218 // constant expression in which case there is no Instruction anchored in a
1220 Instruction *Check = BinaryOperator::CreateAnd(MemoryRuntimeCheck,
1221 ConstantInt::getTrue(Ctx));
1222 ChkBuilder.Insert(Check, "memcheck.conflict");
1223 FirstInst = getFirstInst(FirstInst, Check, Loc);
1224 return std::make_pair(FirstInst, Check);