1 //===- LoopAccessAnalysis.cpp - Loop Access Analysis Implementation --------==//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // The implementation for the loop memory dependence that was originally
11 // developed for the loop vectorizer.
13 //===----------------------------------------------------------------------===//
15 #include "llvm/Analysis/LoopAccessAnalysis.h"
16 #include "llvm/Analysis/LoopInfo.h"
17 #include "llvm/Analysis/ScalarEvolutionExpander.h"
18 #include "llvm/Analysis/TargetLibraryInfo.h"
19 #include "llvm/Analysis/ValueTracking.h"
20 #include "llvm/IR/DiagnosticInfo.h"
21 #include "llvm/IR/Dominators.h"
22 #include "llvm/IR/IRBuilder.h"
23 #include "llvm/Support/Debug.h"
24 #include "llvm/Support/raw_ostream.h"
25 #include "llvm/Analysis/VectorUtils.h"
28 #define DEBUG_TYPE "loop-accesses"
30 static cl::opt<unsigned, true>
31 VectorizationFactor("force-vector-width", cl::Hidden,
32 cl::desc("Sets the SIMD width. Zero is autoselect."),
33 cl::location(VectorizerParams::VectorizationFactor));
34 unsigned VectorizerParams::VectorizationFactor;
36 static cl::opt<unsigned, true>
37 VectorizationInterleave("force-vector-interleave", cl::Hidden,
38 cl::desc("Sets the vectorization interleave count. "
39 "Zero is autoselect."),
41 VectorizerParams::VectorizationInterleave));
42 unsigned VectorizerParams::VectorizationInterleave;
44 static cl::opt<unsigned, true> RuntimeMemoryCheckThreshold(
45 "runtime-memory-check-threshold", cl::Hidden,
46 cl::desc("When performing memory disambiguation checks at runtime do not "
47 "generate more than this number of comparisons (default = 8)."),
48 cl::location(VectorizerParams::RuntimeMemoryCheckThreshold), cl::init(8));
49 unsigned VectorizerParams::RuntimeMemoryCheckThreshold;
51 /// \brief The maximum iterations used to merge memory checks
52 static cl::opt<unsigned> MemoryCheckMergeThreshold(
53 "memory-check-merge-threshold", cl::Hidden,
54 cl::desc("Maximum number of comparisons done when trying to merge "
55 "runtime memory checks. (default = 100)"),
58 /// Maximum SIMD width.
59 const unsigned VectorizerParams::MaxVectorWidth = 64;
61 /// \brief We collect interesting dependences up to this threshold.
62 static cl::opt<unsigned> MaxInterestingDependence(
63 "max-interesting-dependences", cl::Hidden,
64 cl::desc("Maximum number of interesting dependences collected by "
65 "loop-access analysis (default = 100)"),
68 bool VectorizerParams::isInterleaveForced() {
69 return ::VectorizationInterleave.getNumOccurrences() > 0;
72 void LoopAccessReport::emitAnalysis(const LoopAccessReport &Message,
73 const Function *TheFunction,
75 const char *PassName) {
76 DebugLoc DL = TheLoop->getStartLoc();
77 if (const Instruction *I = Message.getInstr())
78 DL = I->getDebugLoc();
79 emitOptimizationRemarkAnalysis(TheFunction->getContext(), PassName,
80 *TheFunction, DL, Message.str());
83 Value *llvm::stripIntegerCast(Value *V) {
84 if (CastInst *CI = dyn_cast<CastInst>(V))
85 if (CI->getOperand(0)->getType()->isIntegerTy())
86 return CI->getOperand(0);
90 const SCEV *llvm::replaceSymbolicStrideSCEV(ScalarEvolution *SE,
91 const ValueToValueMap &PtrToStride,
92 Value *Ptr, Value *OrigPtr) {
94 const SCEV *OrigSCEV = SE->getSCEV(Ptr);
96 // If there is an entry in the map return the SCEV of the pointer with the
97 // symbolic stride replaced by one.
98 ValueToValueMap::const_iterator SI =
99 PtrToStride.find(OrigPtr ? OrigPtr : Ptr);
100 if (SI != PtrToStride.end()) {
101 Value *StrideVal = SI->second;
104 StrideVal = stripIntegerCast(StrideVal);
106 // Replace symbolic stride by one.
107 Value *One = ConstantInt::get(StrideVal->getType(), 1);
108 ValueToValueMap RewriteMap;
109 RewriteMap[StrideVal] = One;
112 SCEVParameterRewriter::rewrite(OrigSCEV, *SE, RewriteMap, true);
113 DEBUG(dbgs() << "LAA: Replacing SCEV: " << *OrigSCEV << " by: " << *ByOne
118 // Otherwise, just return the SCEV of the original pointer.
119 return SE->getSCEV(Ptr);
122 void RuntimePointerChecking::insert(Loop *Lp, Value *Ptr, bool WritePtr,
123 unsigned DepSetId, unsigned ASId,
124 const ValueToValueMap &Strides) {
125 // Get the stride replaced scev.
126 const SCEV *Sc = replaceSymbolicStrideSCEV(SE, Strides, Ptr);
127 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Sc);
128 assert(AR && "Invalid addrec expression");
129 const SCEV *Ex = SE->getBackedgeTakenCount(Lp);
131 const SCEV *ScStart = AR->getStart();
132 const SCEV *ScEnd = AR->evaluateAtIteration(Ex, *SE);
133 const SCEV *Step = AR->getStepRecurrence(*SE);
135 // For expressions with negative step, the upper bound is ScStart and the
136 // lower bound is ScEnd.
137 if (const SCEVConstant *CStep = dyn_cast<const SCEVConstant>(Step)) {
138 if (CStep->getValue()->isNegative())
139 std::swap(ScStart, ScEnd);
141 // Fallback case: the step is not constant, but the we can still
142 // get the upper and lower bounds of the interval by using min/max
144 ScStart = SE->getUMinExpr(ScStart, ScEnd);
145 ScEnd = SE->getUMaxExpr(AR->getStart(), ScEnd);
148 Pointers.emplace_back(Ptr, ScStart, ScEnd, WritePtr, DepSetId, ASId, Sc);
151 SmallVector<RuntimePointerChecking::PointerCheck, 4>
152 RuntimePointerChecking::generateChecks(
153 const SmallVectorImpl<int> *PtrPartition) const {
154 SmallVector<PointerCheck, 4> Checks;
156 for (unsigned I = 0; I < CheckingGroups.size(); ++I) {
157 for (unsigned J = I + 1; J < CheckingGroups.size(); ++J) {
158 const RuntimePointerChecking::CheckingPtrGroup &CGI = CheckingGroups[I];
159 const RuntimePointerChecking::CheckingPtrGroup &CGJ = CheckingGroups[J];
161 if (needsChecking(CGI, CGJ, PtrPartition))
162 Checks.push_back(std::make_pair(&CGI, &CGJ));
168 bool RuntimePointerChecking::needsChecking(
169 const CheckingPtrGroup &M, const CheckingPtrGroup &N,
170 const SmallVectorImpl<int> *PtrPartition) const {
171 for (unsigned I = 0, EI = M.Members.size(); EI != I; ++I)
172 for (unsigned J = 0, EJ = N.Members.size(); EJ != J; ++J)
173 if (needsChecking(M.Members[I], N.Members[J], PtrPartition))
178 /// Compare \p I and \p J and return the minimum.
179 /// Return nullptr in case we couldn't find an answer.
180 static const SCEV *getMinFromExprs(const SCEV *I, const SCEV *J,
181 ScalarEvolution *SE) {
182 const SCEV *Diff = SE->getMinusSCEV(J, I);
183 const SCEVConstant *C = dyn_cast<const SCEVConstant>(Diff);
187 if (C->getValue()->isNegative())
192 bool RuntimePointerChecking::CheckingPtrGroup::addPointer(unsigned Index) {
193 const SCEV *Start = RtCheck.Pointers[Index].Start;
194 const SCEV *End = RtCheck.Pointers[Index].End;
196 // Compare the starts and ends with the known minimum and maximum
197 // of this set. We need to know how we compare against the min/max
198 // of the set in order to be able to emit memchecks.
199 const SCEV *Min0 = getMinFromExprs(Start, Low, RtCheck.SE);
203 const SCEV *Min1 = getMinFromExprs(End, High, RtCheck.SE);
207 // Update the low bound expression if we've found a new min value.
211 // Update the high bound expression if we've found a new max value.
215 Members.push_back(Index);
219 void RuntimePointerChecking::groupChecks(
220 MemoryDepChecker::DepCandidates &DepCands, bool UseDependencies) {
221 // We build the groups from dependency candidates equivalence classes
223 // - We know that pointers in the same equivalence class share
224 // the same underlying object and therefore there is a chance
225 // that we can compare pointers
226 // - We wouldn't be able to merge two pointers for which we need
227 // to emit a memcheck. The classes in DepCands are already
228 // conveniently built such that no two pointers in the same
229 // class need checking against each other.
231 // We use the following (greedy) algorithm to construct the groups
232 // For every pointer in the equivalence class:
233 // For each existing group:
234 // - if the difference between this pointer and the min/max bounds
235 // of the group is a constant, then make the pointer part of the
236 // group and update the min/max bounds of that group as required.
238 CheckingGroups.clear();
240 // If we don't have the dependency partitions, construct a new
241 // checking pointer group for each pointer.
242 if (!UseDependencies) {
243 for (unsigned I = 0; I < Pointers.size(); ++I)
244 CheckingGroups.push_back(CheckingPtrGroup(I, *this));
248 unsigned TotalComparisons = 0;
250 DenseMap<Value *, unsigned> PositionMap;
251 for (unsigned Index = 0; Index < Pointers.size(); ++Index)
252 PositionMap[Pointers[Index].PointerValue] = Index;
254 // We need to keep track of what pointers we've already seen so we
255 // don't process them twice.
256 SmallSet<unsigned, 2> Seen;
258 // Go through all equivalence classes, get the the "pointer check groups"
259 // and add them to the overall solution. We use the order in which accesses
260 // appear in 'Pointers' to enforce determinism.
261 for (unsigned I = 0; I < Pointers.size(); ++I) {
262 // We've seen this pointer before, and therefore already processed
263 // its equivalence class.
267 MemoryDepChecker::MemAccessInfo Access(Pointers[I].PointerValue,
268 Pointers[I].IsWritePtr);
270 SmallVector<CheckingPtrGroup, 2> Groups;
271 auto LeaderI = DepCands.findValue(DepCands.getLeaderValue(Access));
273 // Because DepCands is constructed by visiting accesses in the order in
274 // which they appear in alias sets (which is deterministic) and the
275 // iteration order within an equivalence class member is only dependent on
276 // the order in which unions and insertions are performed on the
277 // equivalence class, the iteration order is deterministic.
278 for (auto MI = DepCands.member_begin(LeaderI), ME = DepCands.member_end();
280 unsigned Pointer = PositionMap[MI->getPointer()];
282 // Mark this pointer as seen.
283 Seen.insert(Pointer);
285 // Go through all the existing sets and see if we can find one
286 // which can include this pointer.
287 for (CheckingPtrGroup &Group : Groups) {
288 // Don't perform more than a certain amount of comparisons.
289 // This should limit the cost of grouping the pointers to something
290 // reasonable. If we do end up hitting this threshold, the algorithm
291 // will create separate groups for all remaining pointers.
292 if (TotalComparisons > MemoryCheckMergeThreshold)
297 if (Group.addPointer(Pointer)) {
304 // We couldn't add this pointer to any existing set or the threshold
305 // for the number of comparisons has been reached. Create a new group
306 // to hold the current pointer.
307 Groups.push_back(CheckingPtrGroup(Pointer, *this));
310 // We've computed the grouped checks for this partition.
311 // Save the results and continue with the next one.
312 std::copy(Groups.begin(), Groups.end(), std::back_inserter(CheckingGroups));
316 bool RuntimePointerChecking::arePointersInSamePartition(
317 const SmallVectorImpl<int> &PtrToPartition, unsigned PtrIdx1,
319 return (PtrToPartition[PtrIdx1] != -1 &&
320 PtrToPartition[PtrIdx1] == PtrToPartition[PtrIdx2]);
323 bool RuntimePointerChecking::needsChecking(
324 unsigned I, unsigned J, const SmallVectorImpl<int> *PtrPartition) const {
325 const PointerInfo &PointerI = Pointers[I];
326 const PointerInfo &PointerJ = Pointers[J];
328 // No need to check if two readonly pointers intersect.
329 if (!PointerI.IsWritePtr && !PointerJ.IsWritePtr)
332 // Only need to check pointers between two different dependency sets.
333 if (PointerI.DependencySetId == PointerJ.DependencySetId)
336 // Only need to check pointers in the same alias set.
337 if (PointerI.AliasSetId != PointerJ.AliasSetId)
340 // If PtrPartition is set omit checks between pointers of the same partition.
341 if (PtrPartition && arePointersInSamePartition(*PtrPartition, I, J))
347 void RuntimePointerChecking::printChecks(
348 raw_ostream &OS, const SmallVectorImpl<PointerCheck> &Checks,
349 unsigned Depth) const {
351 for (const auto &Check : Checks) {
352 const auto &First = Check.first->Members, &Second = Check.second->Members;
354 OS.indent(Depth) << "Check " << N++ << ":\n";
356 OS.indent(Depth + 2) << "Comparing group (" << Check.first << "):\n";
357 for (unsigned K = 0; K < First.size(); ++K)
358 OS.indent(Depth + 2) << *Pointers[First[K]].PointerValue << "\n";
360 OS.indent(Depth + 2) << "Against group (" << Check.second << "):\n";
361 for (unsigned K = 0; K < Second.size(); ++K)
362 OS.indent(Depth + 2) << *Pointers[Second[K]].PointerValue << "\n";
366 void RuntimePointerChecking::print(
367 raw_ostream &OS, unsigned Depth,
368 const SmallVectorImpl<int> *PtrPartition) const {
370 OS.indent(Depth) << "Run-time memory checks:\n";
371 printChecks(OS, generateChecks(PtrPartition), Depth);
373 OS.indent(Depth) << "Grouped accesses:\n";
374 for (unsigned I = 0; I < CheckingGroups.size(); ++I) {
375 const auto &CG = CheckingGroups[I];
377 OS.indent(Depth + 2) << "Group " << &CG << ":\n";
378 OS.indent(Depth + 4) << "(Low: " << *CG.Low << " High: " << *CG.High
380 for (unsigned J = 0; J < CG.Members.size(); ++J) {
381 OS.indent(Depth + 6) << "Member: " << *Pointers[CG.Members[J]].Expr
387 unsigned RuntimePointerChecking::getNumberOfChecks(
388 const SmallVectorImpl<int> *PtrPartition) const {
390 unsigned NumPartitions = CheckingGroups.size();
391 unsigned CheckCount = 0;
393 for (unsigned I = 0; I < NumPartitions; ++I)
394 for (unsigned J = I + 1; J < NumPartitions; ++J)
395 if (needsChecking(CheckingGroups[I], CheckingGroups[J], PtrPartition))
400 bool RuntimePointerChecking::needsAnyChecking(
401 const SmallVectorImpl<int> *PtrPartition) const {
402 unsigned NumPointers = Pointers.size();
404 for (unsigned I = 0; I < NumPointers; ++I)
405 for (unsigned J = I + 1; J < NumPointers; ++J)
406 if (needsChecking(I, J, PtrPartition))
412 /// \brief Analyses memory accesses in a loop.
414 /// Checks whether run time pointer checks are needed and builds sets for data
415 /// dependence checking.
416 class AccessAnalysis {
418 /// \brief Read or write access location.
419 typedef PointerIntPair<Value *, 1, bool> MemAccessInfo;
420 typedef SmallPtrSet<MemAccessInfo, 8> MemAccessInfoSet;
422 AccessAnalysis(const DataLayout &Dl, AliasAnalysis *AA, LoopInfo *LI,
423 MemoryDepChecker::DepCandidates &DA)
424 : DL(Dl), AST(*AA), LI(LI), DepCands(DA),
425 IsRTCheckAnalysisNeeded(false) {}
427 /// \brief Register a load and whether it is only read from.
428 void addLoad(MemoryLocation &Loc, bool IsReadOnly) {
429 Value *Ptr = const_cast<Value*>(Loc.Ptr);
430 AST.add(Ptr, MemoryLocation::UnknownSize, Loc.AATags);
431 Accesses.insert(MemAccessInfo(Ptr, false));
433 ReadOnlyPtr.insert(Ptr);
436 /// \brief Register a store.
437 void addStore(MemoryLocation &Loc) {
438 Value *Ptr = const_cast<Value*>(Loc.Ptr);
439 AST.add(Ptr, MemoryLocation::UnknownSize, Loc.AATags);
440 Accesses.insert(MemAccessInfo(Ptr, true));
443 /// \brief Check whether we can check the pointers at runtime for
444 /// non-intersection.
446 /// Returns true if we need no check or if we do and we can generate them
447 /// (i.e. the pointers have computable bounds).
448 bool canCheckPtrAtRT(RuntimePointerChecking &RtCheck, ScalarEvolution *SE,
449 Loop *TheLoop, const ValueToValueMap &Strides,
450 bool ShouldCheckStride = false);
452 /// \brief Goes over all memory accesses, checks whether a RT check is needed
453 /// and builds sets of dependent accesses.
454 void buildDependenceSets() {
455 processMemAccesses();
458 /// \brief Initial processing of memory accesses determined that we need to
459 /// perform dependency checking.
461 /// Note that this can later be cleared if we retry memcheck analysis without
462 /// dependency checking (i.e. ShouldRetryWithRuntimeCheck).
463 bool isDependencyCheckNeeded() { return !CheckDeps.empty(); }
465 /// We decided that no dependence analysis would be used. Reset the state.
466 void resetDepChecks(MemoryDepChecker &DepChecker) {
468 DepChecker.clearInterestingDependences();
471 MemAccessInfoSet &getDependenciesToCheck() { return CheckDeps; }
474 typedef SetVector<MemAccessInfo> PtrAccessSet;
476 /// \brief Go over all memory access and check whether runtime pointer checks
477 /// are needed and build sets of dependency check candidates.
478 void processMemAccesses();
480 /// Set of all accesses.
481 PtrAccessSet Accesses;
483 const DataLayout &DL;
485 /// Set of accesses that need a further dependence check.
486 MemAccessInfoSet CheckDeps;
488 /// Set of pointers that are read only.
489 SmallPtrSet<Value*, 16> ReadOnlyPtr;
491 /// An alias set tracker to partition the access set by underlying object and
492 //intrinsic property (such as TBAA metadata).
497 /// Sets of potentially dependent accesses - members of one set share an
498 /// underlying pointer. The set "CheckDeps" identfies which sets really need a
499 /// dependence check.
500 MemoryDepChecker::DepCandidates &DepCands;
502 /// \brief Initial processing of memory accesses determined that we may need
503 /// to add memchecks. Perform the analysis to determine the necessary checks.
505 /// Note that, this is different from isDependencyCheckNeeded. When we retry
506 /// memcheck analysis without dependency checking
507 /// (i.e. ShouldRetryWithRuntimeCheck), isDependencyCheckNeeded is cleared
508 /// while this remains set if we have potentially dependent accesses.
509 bool IsRTCheckAnalysisNeeded;
512 } // end anonymous namespace
514 /// \brief Check whether a pointer can participate in a runtime bounds check.
515 static bool hasComputableBounds(ScalarEvolution *SE,
516 const ValueToValueMap &Strides, Value *Ptr) {
517 const SCEV *PtrScev = replaceSymbolicStrideSCEV(SE, Strides, Ptr);
518 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
522 return AR->isAffine();
525 bool AccessAnalysis::canCheckPtrAtRT(RuntimePointerChecking &RtCheck,
526 ScalarEvolution *SE, Loop *TheLoop,
527 const ValueToValueMap &StridesMap,
528 bool ShouldCheckStride) {
529 // Find pointers with computable bounds. We are going to use this information
530 // to place a runtime bound check.
533 bool NeedRTCheck = false;
534 if (!IsRTCheckAnalysisNeeded) return true;
536 bool IsDepCheckNeeded = isDependencyCheckNeeded();
538 // We assign a consecutive id to access from different alias sets.
539 // Accesses between different groups doesn't need to be checked.
541 for (auto &AS : AST) {
542 int NumReadPtrChecks = 0;
543 int NumWritePtrChecks = 0;
545 // We assign consecutive id to access from different dependence sets.
546 // Accesses within the same set don't need a runtime check.
547 unsigned RunningDepId = 1;
548 DenseMap<Value *, unsigned> DepSetId;
551 Value *Ptr = A.getValue();
552 bool IsWrite = Accesses.count(MemAccessInfo(Ptr, true));
553 MemAccessInfo Access(Ptr, IsWrite);
560 if (hasComputableBounds(SE, StridesMap, Ptr) &&
561 // When we run after a failing dependency check we have to make sure
562 // we don't have wrapping pointers.
563 (!ShouldCheckStride ||
564 isStridedPtr(SE, Ptr, TheLoop, StridesMap) == 1)) {
565 // The id of the dependence set.
568 if (IsDepCheckNeeded) {
569 Value *Leader = DepCands.getLeaderValue(Access).getPointer();
570 unsigned &LeaderId = DepSetId[Leader];
572 LeaderId = RunningDepId++;
575 // Each access has its own dependence set.
576 DepId = RunningDepId++;
578 RtCheck.insert(TheLoop, Ptr, IsWrite, DepId, ASId, StridesMap);
580 DEBUG(dbgs() << "LAA: Found a runtime check ptr:" << *Ptr << '\n');
582 DEBUG(dbgs() << "LAA: Can't find bounds for ptr:" << *Ptr << '\n');
587 // If we have at least two writes or one write and a read then we need to
588 // check them. But there is no need to checks if there is only one
589 // dependence set for this alias set.
591 // Note that this function computes CanDoRT and NeedRTCheck independently.
592 // For example CanDoRT=false, NeedRTCheck=false means that we have a pointer
593 // for which we couldn't find the bounds but we don't actually need to emit
594 // any checks so it does not matter.
595 if (!(IsDepCheckNeeded && CanDoRT && RunningDepId == 2))
596 NeedRTCheck |= (NumWritePtrChecks >= 2 || (NumReadPtrChecks >= 1 &&
597 NumWritePtrChecks >= 1));
602 // If the pointers that we would use for the bounds comparison have different
603 // address spaces, assume the values aren't directly comparable, so we can't
604 // use them for the runtime check. We also have to assume they could
605 // overlap. In the future there should be metadata for whether address spaces
607 unsigned NumPointers = RtCheck.Pointers.size();
608 for (unsigned i = 0; i < NumPointers; ++i) {
609 for (unsigned j = i + 1; j < NumPointers; ++j) {
610 // Only need to check pointers between two different dependency sets.
611 if (RtCheck.Pointers[i].DependencySetId ==
612 RtCheck.Pointers[j].DependencySetId)
614 // Only need to check pointers in the same alias set.
615 if (RtCheck.Pointers[i].AliasSetId != RtCheck.Pointers[j].AliasSetId)
618 Value *PtrI = RtCheck.Pointers[i].PointerValue;
619 Value *PtrJ = RtCheck.Pointers[j].PointerValue;
621 unsigned ASi = PtrI->getType()->getPointerAddressSpace();
622 unsigned ASj = PtrJ->getType()->getPointerAddressSpace();
624 DEBUG(dbgs() << "LAA: Runtime check would require comparison between"
625 " different address spaces\n");
631 if (NeedRTCheck && CanDoRT)
632 RtCheck.groupChecks(DepCands, IsDepCheckNeeded);
634 DEBUG(dbgs() << "LAA: We need to do " << RtCheck.getNumberOfChecks(nullptr)
635 << " pointer comparisons.\n");
637 RtCheck.Need = NeedRTCheck;
639 bool CanDoRTIfNeeded = !NeedRTCheck || CanDoRT;
640 if (!CanDoRTIfNeeded)
642 return CanDoRTIfNeeded;
645 void AccessAnalysis::processMemAccesses() {
646 // We process the set twice: first we process read-write pointers, last we
647 // process read-only pointers. This allows us to skip dependence tests for
648 // read-only pointers.
650 DEBUG(dbgs() << "LAA: Processing memory accesses...\n");
651 DEBUG(dbgs() << " AST: "; AST.dump());
652 DEBUG(dbgs() << "LAA: Accesses(" << Accesses.size() << "):\n");
654 for (auto A : Accesses)
655 dbgs() << "\t" << *A.getPointer() << " (" <<
656 (A.getInt() ? "write" : (ReadOnlyPtr.count(A.getPointer()) ?
657 "read-only" : "read")) << ")\n";
660 // The AliasSetTracker has nicely partitioned our pointers by metadata
661 // compatibility and potential for underlying-object overlap. As a result, we
662 // only need to check for potential pointer dependencies within each alias
664 for (auto &AS : AST) {
665 // Note that both the alias-set tracker and the alias sets themselves used
666 // linked lists internally and so the iteration order here is deterministic
667 // (matching the original instruction order within each set).
669 bool SetHasWrite = false;
671 // Map of pointers to last access encountered.
672 typedef DenseMap<Value*, MemAccessInfo> UnderlyingObjToAccessMap;
673 UnderlyingObjToAccessMap ObjToLastAccess;
675 // Set of access to check after all writes have been processed.
676 PtrAccessSet DeferredAccesses;
678 // Iterate over each alias set twice, once to process read/write pointers,
679 // and then to process read-only pointers.
680 for (int SetIteration = 0; SetIteration < 2; ++SetIteration) {
681 bool UseDeferred = SetIteration > 0;
682 PtrAccessSet &S = UseDeferred ? DeferredAccesses : Accesses;
685 Value *Ptr = AV.getValue();
687 // For a single memory access in AliasSetTracker, Accesses may contain
688 // both read and write, and they both need to be handled for CheckDeps.
690 if (AC.getPointer() != Ptr)
693 bool IsWrite = AC.getInt();
695 // If we're using the deferred access set, then it contains only
697 bool IsReadOnlyPtr = ReadOnlyPtr.count(Ptr) && !IsWrite;
698 if (UseDeferred && !IsReadOnlyPtr)
700 // Otherwise, the pointer must be in the PtrAccessSet, either as a
702 assert(((IsReadOnlyPtr && UseDeferred) || IsWrite ||
703 S.count(MemAccessInfo(Ptr, false))) &&
704 "Alias-set pointer not in the access set?");
706 MemAccessInfo Access(Ptr, IsWrite);
707 DepCands.insert(Access);
709 // Memorize read-only pointers for later processing and skip them in
710 // the first round (they need to be checked after we have seen all
711 // write pointers). Note: we also mark pointer that are not
712 // consecutive as "read-only" pointers (so that we check
713 // "a[b[i]] +="). Hence, we need the second check for "!IsWrite".
714 if (!UseDeferred && IsReadOnlyPtr) {
715 DeferredAccesses.insert(Access);
719 // If this is a write - check other reads and writes for conflicts. If
720 // this is a read only check other writes for conflicts (but only if
721 // there is no other write to the ptr - this is an optimization to
722 // catch "a[i] = a[i] + " without having to do a dependence check).
723 if ((IsWrite || IsReadOnlyPtr) && SetHasWrite) {
724 CheckDeps.insert(Access);
725 IsRTCheckAnalysisNeeded = true;
731 // Create sets of pointers connected by a shared alias set and
732 // underlying object.
733 typedef SmallVector<Value *, 16> ValueVector;
734 ValueVector TempObjects;
736 GetUnderlyingObjects(Ptr, TempObjects, DL, LI);
737 DEBUG(dbgs() << "Underlying objects for pointer " << *Ptr << "\n");
738 for (Value *UnderlyingObj : TempObjects) {
739 UnderlyingObjToAccessMap::iterator Prev =
740 ObjToLastAccess.find(UnderlyingObj);
741 if (Prev != ObjToLastAccess.end())
742 DepCands.unionSets(Access, Prev->second);
744 ObjToLastAccess[UnderlyingObj] = Access;
745 DEBUG(dbgs() << " " << *UnderlyingObj << "\n");
753 static bool isInBoundsGep(Value *Ptr) {
754 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr))
755 return GEP->isInBounds();
759 /// \brief Return true if an AddRec pointer \p Ptr is unsigned non-wrapping,
760 /// i.e. monotonically increasing/decreasing.
761 static bool isNoWrapAddRec(Value *Ptr, const SCEVAddRecExpr *AR,
762 ScalarEvolution *SE, const Loop *L) {
763 // FIXME: This should probably only return true for NUW.
764 if (AR->getNoWrapFlags(SCEV::NoWrapMask))
767 // Scalar evolution does not propagate the non-wrapping flags to values that
768 // are derived from a non-wrapping induction variable because non-wrapping
769 // could be flow-sensitive.
771 // Look through the potentially overflowing instruction to try to prove
772 // non-wrapping for the *specific* value of Ptr.
774 // The arithmetic implied by an inbounds GEP can't overflow.
775 auto *GEP = dyn_cast<GetElementPtrInst>(Ptr);
776 if (!GEP || !GEP->isInBounds())
779 // Make sure there is only one non-const index and analyze that.
780 Value *NonConstIndex = nullptr;
781 for (auto Index = GEP->idx_begin(); Index != GEP->idx_end(); ++Index)
782 if (!isa<ConstantInt>(*Index)) {
785 NonConstIndex = *Index;
788 // The recurrence is on the pointer, ignore for now.
791 // The index in GEP is signed. It is non-wrapping if it's derived from a NSW
792 // AddRec using a NSW operation.
793 if (auto *OBO = dyn_cast<OverflowingBinaryOperator>(NonConstIndex))
794 if (OBO->hasNoSignedWrap() &&
795 // Assume constant for other the operand so that the AddRec can be
797 isa<ConstantInt>(OBO->getOperand(1))) {
798 auto *OpScev = SE->getSCEV(OBO->getOperand(0));
800 if (auto *OpAR = dyn_cast<SCEVAddRecExpr>(OpScev))
801 return OpAR->getLoop() == L && OpAR->getNoWrapFlags(SCEV::FlagNSW);
807 /// \brief Check whether the access through \p Ptr has a constant stride.
808 int llvm::isStridedPtr(ScalarEvolution *SE, Value *Ptr, const Loop *Lp,
809 const ValueToValueMap &StridesMap) {
810 const Type *Ty = Ptr->getType();
811 assert(Ty->isPointerTy() && "Unexpected non-ptr");
813 // Make sure that the pointer does not point to aggregate types.
814 const PointerType *PtrTy = cast<PointerType>(Ty);
815 if (PtrTy->getElementType()->isAggregateType()) {
816 DEBUG(dbgs() << "LAA: Bad stride - Not a pointer to a scalar type"
821 const SCEV *PtrScev = replaceSymbolicStrideSCEV(SE, StridesMap, Ptr);
823 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
825 DEBUG(dbgs() << "LAA: Bad stride - Not an AddRecExpr pointer "
826 << *Ptr << " SCEV: " << *PtrScev << "\n");
830 // The accesss function must stride over the innermost loop.
831 if (Lp != AR->getLoop()) {
832 DEBUG(dbgs() << "LAA: Bad stride - Not striding over innermost loop " <<
833 *Ptr << " SCEV: " << *PtrScev << "\n");
836 // The address calculation must not wrap. Otherwise, a dependence could be
838 // An inbounds getelementptr that is a AddRec with a unit stride
839 // cannot wrap per definition. The unit stride requirement is checked later.
840 // An getelementptr without an inbounds attribute and unit stride would have
841 // to access the pointer value "0" which is undefined behavior in address
842 // space 0, therefore we can also vectorize this case.
843 bool IsInBoundsGEP = isInBoundsGep(Ptr);
844 bool IsNoWrapAddRec = isNoWrapAddRec(Ptr, AR, SE, Lp);
845 bool IsInAddressSpaceZero = PtrTy->getAddressSpace() == 0;
846 if (!IsNoWrapAddRec && !IsInBoundsGEP && !IsInAddressSpaceZero) {
847 DEBUG(dbgs() << "LAA: Bad stride - Pointer may wrap in the address space "
848 << *Ptr << " SCEV: " << *PtrScev << "\n");
852 // Check the step is constant.
853 const SCEV *Step = AR->getStepRecurrence(*SE);
855 // Calculate the pointer stride and check if it is constant.
856 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
858 DEBUG(dbgs() << "LAA: Bad stride - Not a constant strided " << *Ptr <<
859 " SCEV: " << *PtrScev << "\n");
863 auto &DL = Lp->getHeader()->getModule()->getDataLayout();
864 int64_t Size = DL.getTypeAllocSize(PtrTy->getElementType());
865 const APInt &APStepVal = C->getValue()->getValue();
867 // Huge step value - give up.
868 if (APStepVal.getBitWidth() > 64)
871 int64_t StepVal = APStepVal.getSExtValue();
874 int64_t Stride = StepVal / Size;
875 int64_t Rem = StepVal % Size;
879 // If the SCEV could wrap but we have an inbounds gep with a unit stride we
880 // know we can't "wrap around the address space". In case of address space
881 // zero we know that this won't happen without triggering undefined behavior.
882 if (!IsNoWrapAddRec && (IsInBoundsGEP || IsInAddressSpaceZero) &&
883 Stride != 1 && Stride != -1)
889 bool MemoryDepChecker::Dependence::isSafeForVectorization(DepType Type) {
893 case BackwardVectorizable:
897 case ForwardButPreventsForwarding:
899 case BackwardVectorizableButPreventsForwarding:
902 llvm_unreachable("unexpected DepType!");
905 bool MemoryDepChecker::Dependence::isInterestingDependence(DepType Type) {
911 case BackwardVectorizable:
913 case ForwardButPreventsForwarding:
915 case BackwardVectorizableButPreventsForwarding:
918 llvm_unreachable("unexpected DepType!");
921 bool MemoryDepChecker::Dependence::isPossiblyBackward() const {
925 case ForwardButPreventsForwarding:
929 case BackwardVectorizable:
931 case BackwardVectorizableButPreventsForwarding:
934 llvm_unreachable("unexpected DepType!");
937 bool MemoryDepChecker::couldPreventStoreLoadForward(unsigned Distance,
938 unsigned TypeByteSize) {
939 // If loads occur at a distance that is not a multiple of a feasible vector
940 // factor store-load forwarding does not take place.
941 // Positive dependences might cause troubles because vectorizing them might
942 // prevent store-load forwarding making vectorized code run a lot slower.
943 // a[i] = a[i-3] ^ a[i-8];
944 // The stores to a[i:i+1] don't align with the stores to a[i-3:i-2] and
945 // hence on your typical architecture store-load forwarding does not take
946 // place. Vectorizing in such cases does not make sense.
947 // Store-load forwarding distance.
948 const unsigned NumCyclesForStoreLoadThroughMemory = 8*TypeByteSize;
949 // Maximum vector factor.
950 unsigned MaxVFWithoutSLForwardIssues =
951 VectorizerParams::MaxVectorWidth * TypeByteSize;
952 if(MaxSafeDepDistBytes < MaxVFWithoutSLForwardIssues)
953 MaxVFWithoutSLForwardIssues = MaxSafeDepDistBytes;
955 for (unsigned vf = 2*TypeByteSize; vf <= MaxVFWithoutSLForwardIssues;
957 if (Distance % vf && Distance / vf < NumCyclesForStoreLoadThroughMemory) {
958 MaxVFWithoutSLForwardIssues = (vf >>=1);
963 if (MaxVFWithoutSLForwardIssues< 2*TypeByteSize) {
964 DEBUG(dbgs() << "LAA: Distance " << Distance <<
965 " that could cause a store-load forwarding conflict\n");
969 if (MaxVFWithoutSLForwardIssues < MaxSafeDepDistBytes &&
970 MaxVFWithoutSLForwardIssues !=
971 VectorizerParams::MaxVectorWidth * TypeByteSize)
972 MaxSafeDepDistBytes = MaxVFWithoutSLForwardIssues;
976 /// \brief Check the dependence for two accesses with the same stride \p Stride.
977 /// \p Distance is the positive distance and \p TypeByteSize is type size in
980 /// \returns true if they are independent.
981 static bool areStridedAccessesIndependent(unsigned Distance, unsigned Stride,
982 unsigned TypeByteSize) {
983 assert(Stride > 1 && "The stride must be greater than 1");
984 assert(TypeByteSize > 0 && "The type size in byte must be non-zero");
985 assert(Distance > 0 && "The distance must be non-zero");
987 // Skip if the distance is not multiple of type byte size.
988 if (Distance % TypeByteSize)
991 unsigned ScaledDist = Distance / TypeByteSize;
993 // No dependence if the scaled distance is not multiple of the stride.
995 // for (i = 0; i < 1024 ; i += 4)
996 // A[i+2] = A[i] + 1;
998 // Two accesses in memory (scaled distance is 2, stride is 4):
999 // | A[0] | | | | A[4] | | | |
1000 // | | | A[2] | | | | A[6] | |
1003 // for (i = 0; i < 1024 ; i += 3)
1004 // A[i+4] = A[i] + 1;
1006 // Two accesses in memory (scaled distance is 4, stride is 3):
1007 // | A[0] | | | A[3] | | | A[6] | | |
1008 // | | | | | A[4] | | | A[7] | |
1009 return ScaledDist % Stride;
1012 MemoryDepChecker::Dependence::DepType
1013 MemoryDepChecker::isDependent(const MemAccessInfo &A, unsigned AIdx,
1014 const MemAccessInfo &B, unsigned BIdx,
1015 const ValueToValueMap &Strides) {
1016 assert (AIdx < BIdx && "Must pass arguments in program order");
1018 Value *APtr = A.getPointer();
1019 Value *BPtr = B.getPointer();
1020 bool AIsWrite = A.getInt();
1021 bool BIsWrite = B.getInt();
1023 // Two reads are independent.
1024 if (!AIsWrite && !BIsWrite)
1025 return Dependence::NoDep;
1027 // We cannot check pointers in different address spaces.
1028 if (APtr->getType()->getPointerAddressSpace() !=
1029 BPtr->getType()->getPointerAddressSpace())
1030 return Dependence::Unknown;
1032 const SCEV *AScev = replaceSymbolicStrideSCEV(SE, Strides, APtr);
1033 const SCEV *BScev = replaceSymbolicStrideSCEV(SE, Strides, BPtr);
1035 int StrideAPtr = isStridedPtr(SE, APtr, InnermostLoop, Strides);
1036 int StrideBPtr = isStridedPtr(SE, BPtr, InnermostLoop, Strides);
1038 const SCEV *Src = AScev;
1039 const SCEV *Sink = BScev;
1041 // If the induction step is negative we have to invert source and sink of the
1043 if (StrideAPtr < 0) {
1046 std::swap(APtr, BPtr);
1047 std::swap(Src, Sink);
1048 std::swap(AIsWrite, BIsWrite);
1049 std::swap(AIdx, BIdx);
1050 std::swap(StrideAPtr, StrideBPtr);
1053 const SCEV *Dist = SE->getMinusSCEV(Sink, Src);
1055 DEBUG(dbgs() << "LAA: Src Scev: " << *Src << "Sink Scev: " << *Sink
1056 << "(Induction step: " << StrideAPtr << ")\n");
1057 DEBUG(dbgs() << "LAA: Distance for " << *InstMap[AIdx] << " to "
1058 << *InstMap[BIdx] << ": " << *Dist << "\n");
1060 // Need accesses with constant stride. We don't want to vectorize
1061 // "A[B[i]] += ..." and similar code or pointer arithmetic that could wrap in
1062 // the address space.
1063 if (!StrideAPtr || !StrideBPtr || StrideAPtr != StrideBPtr){
1064 DEBUG(dbgs() << "Pointer access with non-constant stride\n");
1065 return Dependence::Unknown;
1068 const SCEVConstant *C = dyn_cast<SCEVConstant>(Dist);
1070 DEBUG(dbgs() << "LAA: Dependence because of non-constant distance\n");
1071 ShouldRetryWithRuntimeCheck = true;
1072 return Dependence::Unknown;
1075 Type *ATy = APtr->getType()->getPointerElementType();
1076 Type *BTy = BPtr->getType()->getPointerElementType();
1077 auto &DL = InnermostLoop->getHeader()->getModule()->getDataLayout();
1078 unsigned TypeByteSize = DL.getTypeAllocSize(ATy);
1080 // Negative distances are not plausible dependencies.
1081 const APInt &Val = C->getValue()->getValue();
1082 if (Val.isNegative()) {
1083 bool IsTrueDataDependence = (AIsWrite && !BIsWrite);
1084 if (IsTrueDataDependence &&
1085 (couldPreventStoreLoadForward(Val.abs().getZExtValue(), TypeByteSize) ||
1087 return Dependence::ForwardButPreventsForwarding;
1089 DEBUG(dbgs() << "LAA: Dependence is negative: NoDep\n");
1090 return Dependence::Forward;
1093 // Write to the same location with the same size.
1094 // Could be improved to assert type sizes are the same (i32 == float, etc).
1097 return Dependence::NoDep;
1098 DEBUG(dbgs() << "LAA: Zero dependence difference but different types\n");
1099 return Dependence::Unknown;
1102 assert(Val.isStrictlyPositive() && "Expect a positive value");
1106 "LAA: ReadWrite-Write positive dependency with different types\n");
1107 return Dependence::Unknown;
1110 unsigned Distance = (unsigned) Val.getZExtValue();
1112 unsigned Stride = std::abs(StrideAPtr);
1114 areStridedAccessesIndependent(Distance, Stride, TypeByteSize)) {
1115 DEBUG(dbgs() << "LAA: Strided accesses are independent\n");
1116 return Dependence::NoDep;
1119 // Bail out early if passed-in parameters make vectorization not feasible.
1120 unsigned ForcedFactor = (VectorizerParams::VectorizationFactor ?
1121 VectorizerParams::VectorizationFactor : 1);
1122 unsigned ForcedUnroll = (VectorizerParams::VectorizationInterleave ?
1123 VectorizerParams::VectorizationInterleave : 1);
1124 // The minimum number of iterations for a vectorized/unrolled version.
1125 unsigned MinNumIter = std::max(ForcedFactor * ForcedUnroll, 2U);
1127 // It's not vectorizable if the distance is smaller than the minimum distance
1128 // needed for a vectroized/unrolled version. Vectorizing one iteration in
1129 // front needs TypeByteSize * Stride. Vectorizing the last iteration needs
1130 // TypeByteSize (No need to plus the last gap distance).
1132 // E.g. Assume one char is 1 byte in memory and one int is 4 bytes.
1134 // int *B = (int *)((char *)A + 14);
1135 // for (i = 0 ; i < 1024 ; i += 2)
1139 // Two accesses in memory (stride is 2):
1140 // | A[0] | | A[2] | | A[4] | | A[6] | |
1141 // | B[0] | | B[2] | | B[4] |
1143 // Distance needs for vectorizing iterations except the last iteration:
1144 // 4 * 2 * (MinNumIter - 1). Distance needs for the last iteration: 4.
1145 // So the minimum distance needed is: 4 * 2 * (MinNumIter - 1) + 4.
1147 // If MinNumIter is 2, it is vectorizable as the minimum distance needed is
1148 // 12, which is less than distance.
1150 // If MinNumIter is 4 (Say if a user forces the vectorization factor to be 4),
1151 // the minimum distance needed is 28, which is greater than distance. It is
1152 // not safe to do vectorization.
1153 unsigned MinDistanceNeeded =
1154 TypeByteSize * Stride * (MinNumIter - 1) + TypeByteSize;
1155 if (MinDistanceNeeded > Distance) {
1156 DEBUG(dbgs() << "LAA: Failure because of positive distance " << Distance
1158 return Dependence::Backward;
1161 // Unsafe if the minimum distance needed is greater than max safe distance.
1162 if (MinDistanceNeeded > MaxSafeDepDistBytes) {
1163 DEBUG(dbgs() << "LAA: Failure because it needs at least "
1164 << MinDistanceNeeded << " size in bytes");
1165 return Dependence::Backward;
1168 // Positive distance bigger than max vectorization factor.
1169 // FIXME: Should use max factor instead of max distance in bytes, which could
1170 // not handle different types.
1171 // E.g. Assume one char is 1 byte in memory and one int is 4 bytes.
1172 // void foo (int *A, char *B) {
1173 // for (unsigned i = 0; i < 1024; i++) {
1174 // A[i+2] = A[i] + 1;
1175 // B[i+2] = B[i] + 1;
1179 // This case is currently unsafe according to the max safe distance. If we
1180 // analyze the two accesses on array B, the max safe dependence distance
1181 // is 2. Then we analyze the accesses on array A, the minimum distance needed
1182 // is 8, which is less than 2 and forbidden vectorization, But actually
1183 // both A and B could be vectorized by 2 iterations.
1184 MaxSafeDepDistBytes =
1185 Distance < MaxSafeDepDistBytes ? Distance : MaxSafeDepDistBytes;
1187 bool IsTrueDataDependence = (!AIsWrite && BIsWrite);
1188 if (IsTrueDataDependence &&
1189 couldPreventStoreLoadForward(Distance, TypeByteSize))
1190 return Dependence::BackwardVectorizableButPreventsForwarding;
1192 DEBUG(dbgs() << "LAA: Positive distance " << Val.getSExtValue()
1193 << " with max VF = "
1194 << MaxSafeDepDistBytes / (TypeByteSize * Stride) << '\n');
1196 return Dependence::BackwardVectorizable;
1199 bool MemoryDepChecker::areDepsSafe(DepCandidates &AccessSets,
1200 MemAccessInfoSet &CheckDeps,
1201 const ValueToValueMap &Strides) {
1203 MaxSafeDepDistBytes = -1U;
1204 while (!CheckDeps.empty()) {
1205 MemAccessInfo CurAccess = *CheckDeps.begin();
1207 // Get the relevant memory access set.
1208 EquivalenceClasses<MemAccessInfo>::iterator I =
1209 AccessSets.findValue(AccessSets.getLeaderValue(CurAccess));
1211 // Check accesses within this set.
1212 EquivalenceClasses<MemAccessInfo>::member_iterator AI, AE;
1213 AI = AccessSets.member_begin(I), AE = AccessSets.member_end();
1215 // Check every access pair.
1217 CheckDeps.erase(*AI);
1218 EquivalenceClasses<MemAccessInfo>::member_iterator OI = std::next(AI);
1220 // Check every accessing instruction pair in program order.
1221 for (std::vector<unsigned>::iterator I1 = Accesses[*AI].begin(),
1222 I1E = Accesses[*AI].end(); I1 != I1E; ++I1)
1223 for (std::vector<unsigned>::iterator I2 = Accesses[*OI].begin(),
1224 I2E = Accesses[*OI].end(); I2 != I2E; ++I2) {
1225 auto A = std::make_pair(&*AI, *I1);
1226 auto B = std::make_pair(&*OI, *I2);
1232 Dependence::DepType Type =
1233 isDependent(*A.first, A.second, *B.first, B.second, Strides);
1234 SafeForVectorization &= Dependence::isSafeForVectorization(Type);
1236 // Gather dependences unless we accumulated MaxInterestingDependence
1237 // dependences. In that case return as soon as we find the first
1238 // unsafe dependence. This puts a limit on this quadratic
1240 if (RecordInterestingDependences) {
1241 if (Dependence::isInterestingDependence(Type))
1242 InterestingDependences.push_back(
1243 Dependence(A.second, B.second, Type));
1245 if (InterestingDependences.size() >= MaxInterestingDependence) {
1246 RecordInterestingDependences = false;
1247 InterestingDependences.clear();
1248 DEBUG(dbgs() << "Too many dependences, stopped recording\n");
1251 if (!RecordInterestingDependences && !SafeForVectorization)
1260 DEBUG(dbgs() << "Total Interesting Dependences: "
1261 << InterestingDependences.size() << "\n");
1262 return SafeForVectorization;
1265 SmallVector<Instruction *, 4>
1266 MemoryDepChecker::getInstructionsForAccess(Value *Ptr, bool isWrite) const {
1267 MemAccessInfo Access(Ptr, isWrite);
1268 auto &IndexVector = Accesses.find(Access)->second;
1270 SmallVector<Instruction *, 4> Insts;
1271 std::transform(IndexVector.begin(), IndexVector.end(),
1272 std::back_inserter(Insts),
1273 [&](unsigned Idx) { return this->InstMap[Idx]; });
1277 const char *MemoryDepChecker::Dependence::DepName[] = {
1278 "NoDep", "Unknown", "Forward", "ForwardButPreventsForwarding", "Backward",
1279 "BackwardVectorizable", "BackwardVectorizableButPreventsForwarding"};
1281 void MemoryDepChecker::Dependence::print(
1282 raw_ostream &OS, unsigned Depth,
1283 const SmallVectorImpl<Instruction *> &Instrs) const {
1284 OS.indent(Depth) << DepName[Type] << ":\n";
1285 OS.indent(Depth + 2) << *Instrs[Source] << " -> \n";
1286 OS.indent(Depth + 2) << *Instrs[Destination] << "\n";
1289 bool LoopAccessInfo::canAnalyzeLoop() {
1290 // We need to have a loop header.
1291 DEBUG(dbgs() << "LAA: Found a loop: " <<
1292 TheLoop->getHeader()->getName() << '\n');
1294 // We can only analyze innermost loops.
1295 if (!TheLoop->empty()) {
1296 DEBUG(dbgs() << "LAA: loop is not the innermost loop\n");
1297 emitAnalysis(LoopAccessReport() << "loop is not the innermost loop");
1301 // We must have a single backedge.
1302 if (TheLoop->getNumBackEdges() != 1) {
1303 DEBUG(dbgs() << "LAA: loop control flow is not understood by analyzer\n");
1305 LoopAccessReport() <<
1306 "loop control flow is not understood by analyzer");
1310 // We must have a single exiting block.
1311 if (!TheLoop->getExitingBlock()) {
1312 DEBUG(dbgs() << "LAA: loop control flow is not understood by analyzer\n");
1314 LoopAccessReport() <<
1315 "loop control flow is not understood by analyzer");
1319 // We only handle bottom-tested loops, i.e. loop in which the condition is
1320 // checked at the end of each iteration. With that we can assume that all
1321 // instructions in the loop are executed the same number of times.
1322 if (TheLoop->getExitingBlock() != TheLoop->getLoopLatch()) {
1323 DEBUG(dbgs() << "LAA: loop control flow is not understood by analyzer\n");
1325 LoopAccessReport() <<
1326 "loop control flow is not understood by analyzer");
1330 // ScalarEvolution needs to be able to find the exit count.
1331 const SCEV *ExitCount = SE->getBackedgeTakenCount(TheLoop);
1332 if (ExitCount == SE->getCouldNotCompute()) {
1333 emitAnalysis(LoopAccessReport() <<
1334 "could not determine number of loop iterations");
1335 DEBUG(dbgs() << "LAA: SCEV could not compute the loop exit count.\n");
1342 void LoopAccessInfo::analyzeLoop(const ValueToValueMap &Strides) {
1344 typedef SmallVector<Value*, 16> ValueVector;
1345 typedef SmallPtrSet<Value*, 16> ValueSet;
1347 // Holds the Load and Store *instructions*.
1351 // Holds all the different accesses in the loop.
1352 unsigned NumReads = 0;
1353 unsigned NumReadWrites = 0;
1355 PtrRtChecking.Pointers.clear();
1356 PtrRtChecking.Need = false;
1358 const bool IsAnnotatedParallel = TheLoop->isAnnotatedParallel();
1361 for (Loop::block_iterator bb = TheLoop->block_begin(),
1362 be = TheLoop->block_end(); bb != be; ++bb) {
1364 // Scan the BB and collect legal loads and stores.
1365 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
1368 // If this is a load, save it. If this instruction can read from memory
1369 // but is not a load, then we quit. Notice that we don't handle function
1370 // calls that read or write.
1371 if (it->mayReadFromMemory()) {
1372 // Many math library functions read the rounding mode. We will only
1373 // vectorize a loop if it contains known function calls that don't set
1374 // the flag. Therefore, it is safe to ignore this read from memory.
1375 CallInst *Call = dyn_cast<CallInst>(it);
1376 if (Call && getIntrinsicIDForCall(Call, TLI))
1379 // If the function has an explicit vectorized counterpart, we can safely
1380 // assume that it can be vectorized.
1381 if (Call && !Call->isNoBuiltin() && Call->getCalledFunction() &&
1382 TLI->isFunctionVectorizable(Call->getCalledFunction()->getName()))
1385 LoadInst *Ld = dyn_cast<LoadInst>(it);
1386 if (!Ld || (!Ld->isSimple() && !IsAnnotatedParallel)) {
1387 emitAnalysis(LoopAccessReport(Ld)
1388 << "read with atomic ordering or volatile read");
1389 DEBUG(dbgs() << "LAA: Found a non-simple load.\n");
1394 Loads.push_back(Ld);
1395 DepChecker.addAccess(Ld);
1399 // Save 'store' instructions. Abort if other instructions write to memory.
1400 if (it->mayWriteToMemory()) {
1401 StoreInst *St = dyn_cast<StoreInst>(it);
1403 emitAnalysis(LoopAccessReport(it) <<
1404 "instruction cannot be vectorized");
1408 if (!St->isSimple() && !IsAnnotatedParallel) {
1409 emitAnalysis(LoopAccessReport(St)
1410 << "write with atomic ordering or volatile write");
1411 DEBUG(dbgs() << "LAA: Found a non-simple store.\n");
1416 Stores.push_back(St);
1417 DepChecker.addAccess(St);
1422 // Now we have two lists that hold the loads and the stores.
1423 // Next, we find the pointers that they use.
1425 // Check if we see any stores. If there are no stores, then we don't
1426 // care if the pointers are *restrict*.
1427 if (!Stores.size()) {
1428 DEBUG(dbgs() << "LAA: Found a read-only loop!\n");
1433 MemoryDepChecker::DepCandidates DependentAccesses;
1434 AccessAnalysis Accesses(TheLoop->getHeader()->getModule()->getDataLayout(),
1435 AA, LI, DependentAccesses);
1437 // Holds the analyzed pointers. We don't want to call GetUnderlyingObjects
1438 // multiple times on the same object. If the ptr is accessed twice, once
1439 // for read and once for write, it will only appear once (on the write
1440 // list). This is okay, since we are going to check for conflicts between
1441 // writes and between reads and writes, but not between reads and reads.
1444 ValueVector::iterator I, IE;
1445 for (I = Stores.begin(), IE = Stores.end(); I != IE; ++I) {
1446 StoreInst *ST = cast<StoreInst>(*I);
1447 Value* Ptr = ST->getPointerOperand();
1448 // Check for store to loop invariant address.
1449 StoreToLoopInvariantAddress |= isUniform(Ptr);
1450 // If we did *not* see this pointer before, insert it to the read-write
1451 // list. At this phase it is only a 'write' list.
1452 if (Seen.insert(Ptr).second) {
1455 MemoryLocation Loc = MemoryLocation::get(ST);
1456 // The TBAA metadata could have a control dependency on the predication
1457 // condition, so we cannot rely on it when determining whether or not we
1458 // need runtime pointer checks.
1459 if (blockNeedsPredication(ST->getParent(), TheLoop, DT))
1460 Loc.AATags.TBAA = nullptr;
1462 Accesses.addStore(Loc);
1466 if (IsAnnotatedParallel) {
1468 << "LAA: A loop annotated parallel, ignore memory dependency "
1474 for (I = Loads.begin(), IE = Loads.end(); I != IE; ++I) {
1475 LoadInst *LD = cast<LoadInst>(*I);
1476 Value* Ptr = LD->getPointerOperand();
1477 // If we did *not* see this pointer before, insert it to the
1478 // read list. If we *did* see it before, then it is already in
1479 // the read-write list. This allows us to vectorize expressions
1480 // such as A[i] += x; Because the address of A[i] is a read-write
1481 // pointer. This only works if the index of A[i] is consecutive.
1482 // If the address of i is unknown (for example A[B[i]]) then we may
1483 // read a few words, modify, and write a few words, and some of the
1484 // words may be written to the same address.
1485 bool IsReadOnlyPtr = false;
1486 if (Seen.insert(Ptr).second || !isStridedPtr(SE, Ptr, TheLoop, Strides)) {
1488 IsReadOnlyPtr = true;
1491 MemoryLocation Loc = MemoryLocation::get(LD);
1492 // The TBAA metadata could have a control dependency on the predication
1493 // condition, so we cannot rely on it when determining whether or not we
1494 // need runtime pointer checks.
1495 if (blockNeedsPredication(LD->getParent(), TheLoop, DT))
1496 Loc.AATags.TBAA = nullptr;
1498 Accesses.addLoad(Loc, IsReadOnlyPtr);
1501 // If we write (or read-write) to a single destination and there are no
1502 // other reads in this loop then is it safe to vectorize.
1503 if (NumReadWrites == 1 && NumReads == 0) {
1504 DEBUG(dbgs() << "LAA: Found a write-only loop!\n");
1509 // Build dependence sets and check whether we need a runtime pointer bounds
1511 Accesses.buildDependenceSets();
1513 // Find pointers with computable bounds. We are going to use this information
1514 // to place a runtime bound check.
1515 bool CanDoRTIfNeeded =
1516 Accesses.canCheckPtrAtRT(PtrRtChecking, SE, TheLoop, Strides);
1517 if (!CanDoRTIfNeeded) {
1518 emitAnalysis(LoopAccessReport() << "cannot identify array bounds");
1519 DEBUG(dbgs() << "LAA: We can't vectorize because we can't find "
1520 << "the array bounds.\n");
1525 DEBUG(dbgs() << "LAA: We can perform a memory runtime check if needed.\n");
1528 if (Accesses.isDependencyCheckNeeded()) {
1529 DEBUG(dbgs() << "LAA: Checking memory dependencies\n");
1530 CanVecMem = DepChecker.areDepsSafe(
1531 DependentAccesses, Accesses.getDependenciesToCheck(), Strides);
1532 MaxSafeDepDistBytes = DepChecker.getMaxSafeDepDistBytes();
1534 if (!CanVecMem && DepChecker.shouldRetryWithRuntimeCheck()) {
1535 DEBUG(dbgs() << "LAA: Retrying with memory checks\n");
1537 // Clear the dependency checks. We assume they are not needed.
1538 Accesses.resetDepChecks(DepChecker);
1540 PtrRtChecking.reset();
1541 PtrRtChecking.Need = true;
1544 Accesses.canCheckPtrAtRT(PtrRtChecking, SE, TheLoop, Strides, true);
1546 // Check that we found the bounds for the pointer.
1547 if (!CanDoRTIfNeeded) {
1548 emitAnalysis(LoopAccessReport()
1549 << "cannot check memory dependencies at runtime");
1550 DEBUG(dbgs() << "LAA: Can't vectorize with memory checks\n");
1560 DEBUG(dbgs() << "LAA: No unsafe dependent memory operations in loop. We"
1561 << (PtrRtChecking.Need ? "" : " don't")
1562 << " need runtime memory checks.\n");
1564 emitAnalysis(LoopAccessReport() <<
1565 "unsafe dependent memory operations in loop");
1566 DEBUG(dbgs() << "LAA: unsafe dependent memory operations in loop\n");
1570 bool LoopAccessInfo::blockNeedsPredication(BasicBlock *BB, Loop *TheLoop,
1571 DominatorTree *DT) {
1572 assert(TheLoop->contains(BB) && "Unknown block used");
1574 // Blocks that do not dominate the latch need predication.
1575 BasicBlock* Latch = TheLoop->getLoopLatch();
1576 return !DT->dominates(BB, Latch);
1579 void LoopAccessInfo::emitAnalysis(LoopAccessReport &Message) {
1580 assert(!Report && "Multiple reports generated");
1584 bool LoopAccessInfo::isUniform(Value *V) const {
1585 return (SE->isLoopInvariant(SE->getSCEV(V), TheLoop));
1588 // FIXME: this function is currently a duplicate of the one in
1589 // LoopVectorize.cpp.
1590 static Instruction *getFirstInst(Instruction *FirstInst, Value *V,
1594 if (Instruction *I = dyn_cast<Instruction>(V))
1595 return I->getParent() == Loc->getParent() ? I : nullptr;
1599 /// \brief IR Values for the lower and upper bounds of a pointer evolution.
1600 struct PointerBounds {
1605 /// \brief Expand code for the lower and upper bound of the pointer group \p CG
1606 /// in \p TheLoop. \return the values for the bounds.
1607 static PointerBounds
1608 expandBounds(const RuntimePointerChecking::CheckingPtrGroup *CG, Loop *TheLoop,
1609 Instruction *Loc, SCEVExpander &Exp, ScalarEvolution *SE,
1610 const RuntimePointerChecking &PtrRtChecking) {
1611 Value *Ptr = PtrRtChecking.Pointers[CG->Members[0]].PointerValue;
1612 const SCEV *Sc = SE->getSCEV(Ptr);
1614 if (SE->isLoopInvariant(Sc, TheLoop)) {
1615 DEBUG(dbgs() << "LAA: Adding RT check for a loop invariant ptr:" << *Ptr
1619 unsigned AS = Ptr->getType()->getPointerAddressSpace();
1620 LLVMContext &Ctx = Loc->getContext();
1622 // Use this type for pointer arithmetic.
1623 Type *PtrArithTy = Type::getInt8PtrTy(Ctx, AS);
1624 Value *Start = nullptr, *End = nullptr;
1626 DEBUG(dbgs() << "LAA: Adding RT check for range:\n");
1627 Start = Exp.expandCodeFor(CG->Low, PtrArithTy, Loc);
1628 End = Exp.expandCodeFor(CG->High, PtrArithTy, Loc);
1629 DEBUG(dbgs() << "Start: " << *CG->Low << " End: " << *CG->High << "\n");
1630 return {Start, End};
1634 /// \brief Turns a collection of checks into a collection of expanded upper and
1635 /// lower bounds for both pointers in the check.
1636 static SmallVector<std::pair<PointerBounds, PointerBounds>, 4> expandBounds(
1637 const SmallVectorImpl<RuntimePointerChecking::PointerCheck> &PointerChecks,
1638 Loop *L, Instruction *Loc, ScalarEvolution *SE, SCEVExpander &Exp,
1639 const RuntimePointerChecking &PtrRtChecking) {
1640 SmallVector<std::pair<PointerBounds, PointerBounds>, 4> ChecksWithBounds;
1642 // Here we're relying on the SCEV Expander's cache to only emit code for the
1643 // same bounds once.
1645 PointerChecks.begin(), PointerChecks.end(),
1646 std::back_inserter(ChecksWithBounds),
1647 [&](const RuntimePointerChecking::PointerCheck &Check) {
1649 First = expandBounds(Check.first, L, Loc, Exp, SE, PtrRtChecking),
1650 Second = expandBounds(Check.second, L, Loc, Exp, SE, PtrRtChecking);
1651 return std::make_pair(First, Second);
1654 return ChecksWithBounds;
1657 std::pair<Instruction *, Instruction *> LoopAccessInfo::addRuntimeCheck(
1659 const SmallVectorImpl<RuntimePointerChecking::PointerCheck> &PointerChecks)
1662 SCEVExpander Exp(*SE, DL, "induction");
1663 auto ExpandedChecks =
1664 expandBounds(PointerChecks, TheLoop, Loc, SE, Exp, PtrRtChecking);
1666 LLVMContext &Ctx = Loc->getContext();
1667 Instruction *FirstInst = nullptr;
1668 IRBuilder<> ChkBuilder(Loc);
1669 // Our instructions might fold to a constant.
1670 Value *MemoryRuntimeCheck = nullptr;
1672 for (const auto &Check : ExpandedChecks) {
1673 const PointerBounds &A = Check.first, &B = Check.second;
1674 unsigned AS0 = A.Start->getType()->getPointerAddressSpace();
1675 unsigned AS1 = B.Start->getType()->getPointerAddressSpace();
1677 assert((AS0 == B.End->getType()->getPointerAddressSpace()) &&
1678 (AS1 == A.End->getType()->getPointerAddressSpace()) &&
1679 "Trying to bounds check pointers with different address spaces");
1681 Type *PtrArithTy0 = Type::getInt8PtrTy(Ctx, AS0);
1682 Type *PtrArithTy1 = Type::getInt8PtrTy(Ctx, AS1);
1684 Value *Start0 = ChkBuilder.CreateBitCast(A.Start, PtrArithTy0, "bc");
1685 Value *Start1 = ChkBuilder.CreateBitCast(B.Start, PtrArithTy1, "bc");
1686 Value *End0 = ChkBuilder.CreateBitCast(A.End, PtrArithTy1, "bc");
1687 Value *End1 = ChkBuilder.CreateBitCast(B.End, PtrArithTy0, "bc");
1689 Value *Cmp0 = ChkBuilder.CreateICmpULE(Start0, End1, "bound0");
1690 FirstInst = getFirstInst(FirstInst, Cmp0, Loc);
1691 Value *Cmp1 = ChkBuilder.CreateICmpULE(Start1, End0, "bound1");
1692 FirstInst = getFirstInst(FirstInst, Cmp1, Loc);
1693 Value *IsConflict = ChkBuilder.CreateAnd(Cmp0, Cmp1, "found.conflict");
1694 FirstInst = getFirstInst(FirstInst, IsConflict, Loc);
1695 if (MemoryRuntimeCheck) {
1697 ChkBuilder.CreateOr(MemoryRuntimeCheck, IsConflict, "conflict.rdx");
1698 FirstInst = getFirstInst(FirstInst, IsConflict, Loc);
1700 MemoryRuntimeCheck = IsConflict;
1703 if (!MemoryRuntimeCheck)
1704 return std::make_pair(nullptr, nullptr);
1706 // We have to do this trickery because the IRBuilder might fold the check to a
1707 // constant expression in which case there is no Instruction anchored in a
1709 Instruction *Check = BinaryOperator::CreateAnd(MemoryRuntimeCheck,
1710 ConstantInt::getTrue(Ctx));
1711 ChkBuilder.Insert(Check, "memcheck.conflict");
1712 FirstInst = getFirstInst(FirstInst, Check, Loc);
1713 return std::make_pair(FirstInst, Check);
1716 std::pair<Instruction *, Instruction *> LoopAccessInfo::addRuntimeCheck(
1717 Instruction *Loc, const SmallVectorImpl<int> *PtrPartition) const {
1718 if (!PtrRtChecking.Need)
1719 return std::make_pair(nullptr, nullptr);
1721 return addRuntimeCheck(Loc, PtrRtChecking.generateChecks(PtrPartition));
1724 LoopAccessInfo::LoopAccessInfo(Loop *L, ScalarEvolution *SE,
1725 const DataLayout &DL,
1726 const TargetLibraryInfo *TLI, AliasAnalysis *AA,
1727 DominatorTree *DT, LoopInfo *LI,
1728 const ValueToValueMap &Strides)
1729 : PtrRtChecking(SE), DepChecker(SE, L), TheLoop(L), SE(SE), DL(DL),
1730 TLI(TLI), AA(AA), DT(DT), LI(LI), NumLoads(0), NumStores(0),
1731 MaxSafeDepDistBytes(-1U), CanVecMem(false),
1732 StoreToLoopInvariantAddress(false) {
1733 if (canAnalyzeLoop())
1734 analyzeLoop(Strides);
1737 void LoopAccessInfo::print(raw_ostream &OS, unsigned Depth) const {
1739 if (PtrRtChecking.Need)
1740 OS.indent(Depth) << "Memory dependences are safe with run-time checks\n";
1742 OS.indent(Depth) << "Memory dependences are safe\n";
1746 OS.indent(Depth) << "Report: " << Report->str() << "\n";
1748 if (auto *InterestingDependences = DepChecker.getInterestingDependences()) {
1749 OS.indent(Depth) << "Interesting Dependences:\n";
1750 for (auto &Dep : *InterestingDependences) {
1751 Dep.print(OS, Depth + 2, DepChecker.getMemoryInstructions());
1755 OS.indent(Depth) << "Too many interesting dependences, not recorded\n";
1757 // List the pair of accesses need run-time checks to prove independence.
1758 PtrRtChecking.print(OS, Depth);
1761 OS.indent(Depth) << "Store to invariant address was "
1762 << (StoreToLoopInvariantAddress ? "" : "not ")
1763 << "found in loop.\n";
1766 const LoopAccessInfo &
1767 LoopAccessAnalysis::getInfo(Loop *L, const ValueToValueMap &Strides) {
1768 auto &LAI = LoopAccessInfoMap[L];
1771 assert((!LAI || LAI->NumSymbolicStrides == Strides.size()) &&
1772 "Symbolic strides changed for loop");
1776 const DataLayout &DL = L->getHeader()->getModule()->getDataLayout();
1777 LAI = llvm::make_unique<LoopAccessInfo>(L, SE, DL, TLI, AA, DT, LI,
1780 LAI->NumSymbolicStrides = Strides.size();
1786 void LoopAccessAnalysis::print(raw_ostream &OS, const Module *M) const {
1787 LoopAccessAnalysis &LAA = *const_cast<LoopAccessAnalysis *>(this);
1789 ValueToValueMap NoSymbolicStrides;
1791 for (Loop *TopLevelLoop : *LI)
1792 for (Loop *L : depth_first(TopLevelLoop)) {
1793 OS.indent(2) << L->getHeader()->getName() << ":\n";
1794 auto &LAI = LAA.getInfo(L, NoSymbolicStrides);
1799 bool LoopAccessAnalysis::runOnFunction(Function &F) {
1800 SE = &getAnalysis<ScalarEvolution>();
1801 auto *TLIP = getAnalysisIfAvailable<TargetLibraryInfoWrapperPass>();
1802 TLI = TLIP ? &TLIP->getTLI() : nullptr;
1803 AA = &getAnalysis<AliasAnalysis>();
1804 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1805 LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
1810 void LoopAccessAnalysis::getAnalysisUsage(AnalysisUsage &AU) const {
1811 AU.addRequired<ScalarEvolution>();
1812 AU.addRequired<AliasAnalysis>();
1813 AU.addRequired<DominatorTreeWrapperPass>();
1814 AU.addRequired<LoopInfoWrapperPass>();
1816 AU.setPreservesAll();
1819 char LoopAccessAnalysis::ID = 0;
1820 static const char laa_name[] = "Loop Access Analysis";
1821 #define LAA_NAME "loop-accesses"
1823 INITIALIZE_PASS_BEGIN(LoopAccessAnalysis, LAA_NAME, laa_name, false, true)
1824 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
1825 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
1826 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
1827 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
1828 INITIALIZE_PASS_END(LoopAccessAnalysis, LAA_NAME, laa_name, false, true)
1831 Pass *createLAAPass() {
1832 return new LoopAccessAnalysis();