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() const {
153 SmallVector<PointerCheck, 4> Checks;
155 for (unsigned I = 0; I < CheckingGroups.size(); ++I) {
156 for (unsigned J = I + 1; J < CheckingGroups.size(); ++J) {
157 const RuntimePointerChecking::CheckingPtrGroup &CGI = CheckingGroups[I];
158 const RuntimePointerChecking::CheckingPtrGroup &CGJ = CheckingGroups[J];
160 if (needsChecking(CGI, CGJ))
161 Checks.push_back(std::make_pair(&CGI, &CGJ));
167 void RuntimePointerChecking::generateChecks(
168 MemoryDepChecker::DepCandidates &DepCands, bool UseDependencies) {
169 assert(Checks.empty() && "Checks is not empty");
170 groupChecks(DepCands, UseDependencies);
171 Checks = generateChecks();
174 bool RuntimePointerChecking::needsChecking(const CheckingPtrGroup &M,
175 const CheckingPtrGroup &N) const {
176 for (unsigned I = 0, EI = M.Members.size(); EI != I; ++I)
177 for (unsigned J = 0, EJ = N.Members.size(); EJ != J; ++J)
178 if (needsChecking(M.Members[I], N.Members[J]))
183 /// Compare \p I and \p J and return the minimum.
184 /// Return nullptr in case we couldn't find an answer.
185 static const SCEV *getMinFromExprs(const SCEV *I, const SCEV *J,
186 ScalarEvolution *SE) {
187 const SCEV *Diff = SE->getMinusSCEV(J, I);
188 const SCEVConstant *C = dyn_cast<const SCEVConstant>(Diff);
192 if (C->getValue()->isNegative())
197 bool RuntimePointerChecking::CheckingPtrGroup::addPointer(unsigned Index) {
198 const SCEV *Start = RtCheck.Pointers[Index].Start;
199 const SCEV *End = RtCheck.Pointers[Index].End;
201 // Compare the starts and ends with the known minimum and maximum
202 // of this set. We need to know how we compare against the min/max
203 // of the set in order to be able to emit memchecks.
204 const SCEV *Min0 = getMinFromExprs(Start, Low, RtCheck.SE);
208 const SCEV *Min1 = getMinFromExprs(End, High, RtCheck.SE);
212 // Update the low bound expression if we've found a new min value.
216 // Update the high bound expression if we've found a new max value.
220 Members.push_back(Index);
224 void RuntimePointerChecking::groupChecks(
225 MemoryDepChecker::DepCandidates &DepCands, bool UseDependencies) {
226 // We build the groups from dependency candidates equivalence classes
228 // - We know that pointers in the same equivalence class share
229 // the same underlying object and therefore there is a chance
230 // that we can compare pointers
231 // - We wouldn't be able to merge two pointers for which we need
232 // to emit a memcheck. The classes in DepCands are already
233 // conveniently built such that no two pointers in the same
234 // class need checking against each other.
236 // We use the following (greedy) algorithm to construct the groups
237 // For every pointer in the equivalence class:
238 // For each existing group:
239 // - if the difference between this pointer and the min/max bounds
240 // of the group is a constant, then make the pointer part of the
241 // group and update the min/max bounds of that group as required.
243 CheckingGroups.clear();
245 // If we need to check two pointers to the same underlying object
246 // with a non-constant difference, we shouldn't perform any pointer
247 // grouping with those pointers. This is because we can easily get
248 // into cases where the resulting check would return false, even when
249 // the accesses are safe.
251 // The following example shows this:
252 // for (i = 0; i < 1000; ++i)
253 // a[5000 + i * m] = a[i] + a[i + 9000]
255 // Here grouping gives a check of (5000, 5000 + 1000 * m) against
256 // (0, 10000) which is always false. However, if m is 1, there is no
257 // dependence. Not grouping the checks for a[i] and a[i + 9000] allows
258 // us to perform an accurate check in this case.
260 // The above case requires that we have an UnknownDependence between
261 // accesses to the same underlying object. This cannot happen unless
262 // ShouldRetryWithRuntimeCheck is set, and therefore UseDependencies
263 // is also false. In this case we will use the fallback path and create
264 // separate checking groups for all pointers.
266 // If we don't have the dependency partitions, construct a new
267 // checking pointer group for each pointer. This is also required
268 // for correctness, because in this case we can have checking between
269 // pointers to the same underlying object.
270 if (!UseDependencies) {
271 for (unsigned I = 0; I < Pointers.size(); ++I)
272 CheckingGroups.push_back(CheckingPtrGroup(I, *this));
276 unsigned TotalComparisons = 0;
278 DenseMap<Value *, unsigned> PositionMap;
279 for (unsigned Index = 0; Index < Pointers.size(); ++Index)
280 PositionMap[Pointers[Index].PointerValue] = Index;
282 // We need to keep track of what pointers we've already seen so we
283 // don't process them twice.
284 SmallSet<unsigned, 2> Seen;
286 // Go through all equivalence classes, get the the "pointer check groups"
287 // and add them to the overall solution. We use the order in which accesses
288 // appear in 'Pointers' to enforce determinism.
289 for (unsigned I = 0; I < Pointers.size(); ++I) {
290 // We've seen this pointer before, and therefore already processed
291 // its equivalence class.
295 MemoryDepChecker::MemAccessInfo Access(Pointers[I].PointerValue,
296 Pointers[I].IsWritePtr);
298 SmallVector<CheckingPtrGroup, 2> Groups;
299 auto LeaderI = DepCands.findValue(DepCands.getLeaderValue(Access));
301 // Because DepCands is constructed by visiting accesses in the order in
302 // which they appear in alias sets (which is deterministic) and the
303 // iteration order within an equivalence class member is only dependent on
304 // the order in which unions and insertions are performed on the
305 // equivalence class, the iteration order is deterministic.
306 for (auto MI = DepCands.member_begin(LeaderI), ME = DepCands.member_end();
308 unsigned Pointer = PositionMap[MI->getPointer()];
310 // Mark this pointer as seen.
311 Seen.insert(Pointer);
313 // Go through all the existing sets and see if we can find one
314 // which can include this pointer.
315 for (CheckingPtrGroup &Group : Groups) {
316 // Don't perform more than a certain amount of comparisons.
317 // This should limit the cost of grouping the pointers to something
318 // reasonable. If we do end up hitting this threshold, the algorithm
319 // will create separate groups for all remaining pointers.
320 if (TotalComparisons > MemoryCheckMergeThreshold)
325 if (Group.addPointer(Pointer)) {
332 // We couldn't add this pointer to any existing set or the threshold
333 // for the number of comparisons has been reached. Create a new group
334 // to hold the current pointer.
335 Groups.push_back(CheckingPtrGroup(Pointer, *this));
338 // We've computed the grouped checks for this partition.
339 // Save the results and continue with the next one.
340 std::copy(Groups.begin(), Groups.end(), std::back_inserter(CheckingGroups));
344 bool RuntimePointerChecking::arePointersInSamePartition(
345 const SmallVectorImpl<int> &PtrToPartition, unsigned PtrIdx1,
347 return (PtrToPartition[PtrIdx1] != -1 &&
348 PtrToPartition[PtrIdx1] == PtrToPartition[PtrIdx2]);
351 bool RuntimePointerChecking::needsChecking(unsigned I, unsigned J) const {
352 const PointerInfo &PointerI = Pointers[I];
353 const PointerInfo &PointerJ = Pointers[J];
355 // No need to check if two readonly pointers intersect.
356 if (!PointerI.IsWritePtr && !PointerJ.IsWritePtr)
359 // Only need to check pointers between two different dependency sets.
360 if (PointerI.DependencySetId == PointerJ.DependencySetId)
363 // Only need to check pointers in the same alias set.
364 if (PointerI.AliasSetId != PointerJ.AliasSetId)
370 void RuntimePointerChecking::printChecks(
371 raw_ostream &OS, const SmallVectorImpl<PointerCheck> &Checks,
372 unsigned Depth) const {
374 for (const auto &Check : Checks) {
375 const auto &First = Check.first->Members, &Second = Check.second->Members;
377 OS.indent(Depth) << "Check " << N++ << ":\n";
379 OS.indent(Depth + 2) << "Comparing group (" << Check.first << "):\n";
380 for (unsigned K = 0; K < First.size(); ++K)
381 OS.indent(Depth + 2) << *Pointers[First[K]].PointerValue << "\n";
383 OS.indent(Depth + 2) << "Against group (" << Check.second << "):\n";
384 for (unsigned K = 0; K < Second.size(); ++K)
385 OS.indent(Depth + 2) << *Pointers[Second[K]].PointerValue << "\n";
389 void RuntimePointerChecking::print(raw_ostream &OS, unsigned Depth) const {
391 OS.indent(Depth) << "Run-time memory checks:\n";
392 printChecks(OS, Checks, Depth);
394 OS.indent(Depth) << "Grouped accesses:\n";
395 for (unsigned I = 0; I < CheckingGroups.size(); ++I) {
396 const auto &CG = CheckingGroups[I];
398 OS.indent(Depth + 2) << "Group " << &CG << ":\n";
399 OS.indent(Depth + 4) << "(Low: " << *CG.Low << " High: " << *CG.High
401 for (unsigned J = 0; J < CG.Members.size(); ++J) {
402 OS.indent(Depth + 6) << "Member: " << *Pointers[CG.Members[J]].Expr
409 /// \brief Analyses memory accesses in a loop.
411 /// Checks whether run time pointer checks are needed and builds sets for data
412 /// dependence checking.
413 class AccessAnalysis {
415 /// \brief Read or write access location.
416 typedef PointerIntPair<Value *, 1, bool> MemAccessInfo;
417 typedef SmallPtrSet<MemAccessInfo, 8> MemAccessInfoSet;
419 AccessAnalysis(const DataLayout &Dl, AliasAnalysis *AA, LoopInfo *LI,
420 MemoryDepChecker::DepCandidates &DA)
421 : DL(Dl), AST(*AA), LI(LI), DepCands(DA),
422 IsRTCheckAnalysisNeeded(false) {}
424 /// \brief Register a load and whether it is only read from.
425 void addLoad(MemoryLocation &Loc, bool IsReadOnly) {
426 Value *Ptr = const_cast<Value*>(Loc.Ptr);
427 AST.add(Ptr, MemoryLocation::UnknownSize, Loc.AATags);
428 Accesses.insert(MemAccessInfo(Ptr, false));
430 ReadOnlyPtr.insert(Ptr);
433 /// \brief Register a store.
434 void addStore(MemoryLocation &Loc) {
435 Value *Ptr = const_cast<Value*>(Loc.Ptr);
436 AST.add(Ptr, MemoryLocation::UnknownSize, Loc.AATags);
437 Accesses.insert(MemAccessInfo(Ptr, true));
440 /// \brief Check whether we can check the pointers at runtime for
441 /// non-intersection.
443 /// Returns true if we need no check or if we do and we can generate them
444 /// (i.e. the pointers have computable bounds).
445 bool canCheckPtrAtRT(RuntimePointerChecking &RtCheck, ScalarEvolution *SE,
446 Loop *TheLoop, const ValueToValueMap &Strides,
447 bool ShouldCheckStride = false);
449 /// \brief Goes over all memory accesses, checks whether a RT check is needed
450 /// and builds sets of dependent accesses.
451 void buildDependenceSets() {
452 processMemAccesses();
455 /// \brief Initial processing of memory accesses determined that we need to
456 /// perform dependency checking.
458 /// Note that this can later be cleared if we retry memcheck analysis without
459 /// dependency checking (i.e. ShouldRetryWithRuntimeCheck).
460 bool isDependencyCheckNeeded() { return !CheckDeps.empty(); }
462 /// We decided that no dependence analysis would be used. Reset the state.
463 void resetDepChecks(MemoryDepChecker &DepChecker) {
465 DepChecker.clearInterestingDependences();
468 MemAccessInfoSet &getDependenciesToCheck() { return CheckDeps; }
471 typedef SetVector<MemAccessInfo> PtrAccessSet;
473 /// \brief Go over all memory access and check whether runtime pointer checks
474 /// are needed and build sets of dependency check candidates.
475 void processMemAccesses();
477 /// Set of all accesses.
478 PtrAccessSet Accesses;
480 const DataLayout &DL;
482 /// Set of accesses that need a further dependence check.
483 MemAccessInfoSet CheckDeps;
485 /// Set of pointers that are read only.
486 SmallPtrSet<Value*, 16> ReadOnlyPtr;
488 /// An alias set tracker to partition the access set by underlying object and
489 //intrinsic property (such as TBAA metadata).
494 /// Sets of potentially dependent accesses - members of one set share an
495 /// underlying pointer. The set "CheckDeps" identfies which sets really need a
496 /// dependence check.
497 MemoryDepChecker::DepCandidates &DepCands;
499 /// \brief Initial processing of memory accesses determined that we may need
500 /// to add memchecks. Perform the analysis to determine the necessary checks.
502 /// Note that, this is different from isDependencyCheckNeeded. When we retry
503 /// memcheck analysis without dependency checking
504 /// (i.e. ShouldRetryWithRuntimeCheck), isDependencyCheckNeeded is cleared
505 /// while this remains set if we have potentially dependent accesses.
506 bool IsRTCheckAnalysisNeeded;
509 } // end anonymous namespace
511 /// \brief Check whether a pointer can participate in a runtime bounds check.
512 static bool hasComputableBounds(ScalarEvolution *SE,
513 const ValueToValueMap &Strides, Value *Ptr) {
514 const SCEV *PtrScev = replaceSymbolicStrideSCEV(SE, Strides, Ptr);
515 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
519 return AR->isAffine();
522 bool AccessAnalysis::canCheckPtrAtRT(RuntimePointerChecking &RtCheck,
523 ScalarEvolution *SE, Loop *TheLoop,
524 const ValueToValueMap &StridesMap,
525 bool ShouldCheckStride) {
526 // Find pointers with computable bounds. We are going to use this information
527 // to place a runtime bound check.
530 bool NeedRTCheck = false;
531 if (!IsRTCheckAnalysisNeeded) return true;
533 bool IsDepCheckNeeded = isDependencyCheckNeeded();
535 // We assign a consecutive id to access from different alias sets.
536 // Accesses between different groups doesn't need to be checked.
538 for (auto &AS : AST) {
539 int NumReadPtrChecks = 0;
540 int NumWritePtrChecks = 0;
542 // We assign consecutive id to access from different dependence sets.
543 // Accesses within the same set don't need a runtime check.
544 unsigned RunningDepId = 1;
545 DenseMap<Value *, unsigned> DepSetId;
548 Value *Ptr = A.getValue();
549 bool IsWrite = Accesses.count(MemAccessInfo(Ptr, true));
550 MemAccessInfo Access(Ptr, IsWrite);
557 if (hasComputableBounds(SE, StridesMap, Ptr) &&
558 // When we run after a failing dependency check we have to make sure
559 // we don't have wrapping pointers.
560 (!ShouldCheckStride ||
561 isStridedPtr(SE, Ptr, TheLoop, StridesMap) == 1)) {
562 // The id of the dependence set.
565 if (IsDepCheckNeeded) {
566 Value *Leader = DepCands.getLeaderValue(Access).getPointer();
567 unsigned &LeaderId = DepSetId[Leader];
569 LeaderId = RunningDepId++;
572 // Each access has its own dependence set.
573 DepId = RunningDepId++;
575 RtCheck.insert(TheLoop, Ptr, IsWrite, DepId, ASId, StridesMap);
577 DEBUG(dbgs() << "LAA: Found a runtime check ptr:" << *Ptr << '\n');
579 DEBUG(dbgs() << "LAA: Can't find bounds for ptr:" << *Ptr << '\n');
584 // If we have at least two writes or one write and a read then we need to
585 // check them. But there is no need to checks if there is only one
586 // dependence set for this alias set.
588 // Note that this function computes CanDoRT and NeedRTCheck independently.
589 // For example CanDoRT=false, NeedRTCheck=false means that we have a pointer
590 // for which we couldn't find the bounds but we don't actually need to emit
591 // any checks so it does not matter.
592 if (!(IsDepCheckNeeded && CanDoRT && RunningDepId == 2))
593 NeedRTCheck |= (NumWritePtrChecks >= 2 || (NumReadPtrChecks >= 1 &&
594 NumWritePtrChecks >= 1));
599 // If the pointers that we would use for the bounds comparison have different
600 // address spaces, assume the values aren't directly comparable, so we can't
601 // use them for the runtime check. We also have to assume they could
602 // overlap. In the future there should be metadata for whether address spaces
604 unsigned NumPointers = RtCheck.Pointers.size();
605 for (unsigned i = 0; i < NumPointers; ++i) {
606 for (unsigned j = i + 1; j < NumPointers; ++j) {
607 // Only need to check pointers between two different dependency sets.
608 if (RtCheck.Pointers[i].DependencySetId ==
609 RtCheck.Pointers[j].DependencySetId)
611 // Only need to check pointers in the same alias set.
612 if (RtCheck.Pointers[i].AliasSetId != RtCheck.Pointers[j].AliasSetId)
615 Value *PtrI = RtCheck.Pointers[i].PointerValue;
616 Value *PtrJ = RtCheck.Pointers[j].PointerValue;
618 unsigned ASi = PtrI->getType()->getPointerAddressSpace();
619 unsigned ASj = PtrJ->getType()->getPointerAddressSpace();
621 DEBUG(dbgs() << "LAA: Runtime check would require comparison between"
622 " different address spaces\n");
628 if (NeedRTCheck && CanDoRT)
629 RtCheck.generateChecks(DepCands, IsDepCheckNeeded);
631 DEBUG(dbgs() << "LAA: We need to do " << RtCheck.getNumberOfChecks()
632 << " pointer comparisons.\n");
634 RtCheck.Need = NeedRTCheck;
636 bool CanDoRTIfNeeded = !NeedRTCheck || CanDoRT;
637 if (!CanDoRTIfNeeded)
639 return CanDoRTIfNeeded;
642 void AccessAnalysis::processMemAccesses() {
643 // We process the set twice: first we process read-write pointers, last we
644 // process read-only pointers. This allows us to skip dependence tests for
645 // read-only pointers.
647 DEBUG(dbgs() << "LAA: Processing memory accesses...\n");
648 DEBUG(dbgs() << " AST: "; AST.dump());
649 DEBUG(dbgs() << "LAA: Accesses(" << Accesses.size() << "):\n");
651 for (auto A : Accesses)
652 dbgs() << "\t" << *A.getPointer() << " (" <<
653 (A.getInt() ? "write" : (ReadOnlyPtr.count(A.getPointer()) ?
654 "read-only" : "read")) << ")\n";
657 // The AliasSetTracker has nicely partitioned our pointers by metadata
658 // compatibility and potential for underlying-object overlap. As a result, we
659 // only need to check for potential pointer dependencies within each alias
661 for (auto &AS : AST) {
662 // Note that both the alias-set tracker and the alias sets themselves used
663 // linked lists internally and so the iteration order here is deterministic
664 // (matching the original instruction order within each set).
666 bool SetHasWrite = false;
668 // Map of pointers to last access encountered.
669 typedef DenseMap<Value*, MemAccessInfo> UnderlyingObjToAccessMap;
670 UnderlyingObjToAccessMap ObjToLastAccess;
672 // Set of access to check after all writes have been processed.
673 PtrAccessSet DeferredAccesses;
675 // Iterate over each alias set twice, once to process read/write pointers,
676 // and then to process read-only pointers.
677 for (int SetIteration = 0; SetIteration < 2; ++SetIteration) {
678 bool UseDeferred = SetIteration > 0;
679 PtrAccessSet &S = UseDeferred ? DeferredAccesses : Accesses;
682 Value *Ptr = AV.getValue();
684 // For a single memory access in AliasSetTracker, Accesses may contain
685 // both read and write, and they both need to be handled for CheckDeps.
687 if (AC.getPointer() != Ptr)
690 bool IsWrite = AC.getInt();
692 // If we're using the deferred access set, then it contains only
694 bool IsReadOnlyPtr = ReadOnlyPtr.count(Ptr) && !IsWrite;
695 if (UseDeferred && !IsReadOnlyPtr)
697 // Otherwise, the pointer must be in the PtrAccessSet, either as a
699 assert(((IsReadOnlyPtr && UseDeferred) || IsWrite ||
700 S.count(MemAccessInfo(Ptr, false))) &&
701 "Alias-set pointer not in the access set?");
703 MemAccessInfo Access(Ptr, IsWrite);
704 DepCands.insert(Access);
706 // Memorize read-only pointers for later processing and skip them in
707 // the first round (they need to be checked after we have seen all
708 // write pointers). Note: we also mark pointer that are not
709 // consecutive as "read-only" pointers (so that we check
710 // "a[b[i]] +="). Hence, we need the second check for "!IsWrite".
711 if (!UseDeferred && IsReadOnlyPtr) {
712 DeferredAccesses.insert(Access);
716 // If this is a write - check other reads and writes for conflicts. If
717 // this is a read only check other writes for conflicts (but only if
718 // there is no other write to the ptr - this is an optimization to
719 // catch "a[i] = a[i] + " without having to do a dependence check).
720 if ((IsWrite || IsReadOnlyPtr) && SetHasWrite) {
721 CheckDeps.insert(Access);
722 IsRTCheckAnalysisNeeded = true;
728 // Create sets of pointers connected by a shared alias set and
729 // underlying object.
730 typedef SmallVector<Value *, 16> ValueVector;
731 ValueVector TempObjects;
733 GetUnderlyingObjects(Ptr, TempObjects, DL, LI);
734 DEBUG(dbgs() << "Underlying objects for pointer " << *Ptr << "\n");
735 for (Value *UnderlyingObj : TempObjects) {
736 UnderlyingObjToAccessMap::iterator Prev =
737 ObjToLastAccess.find(UnderlyingObj);
738 if (Prev != ObjToLastAccess.end())
739 DepCands.unionSets(Access, Prev->second);
741 ObjToLastAccess[UnderlyingObj] = Access;
742 DEBUG(dbgs() << " " << *UnderlyingObj << "\n");
750 static bool isInBoundsGep(Value *Ptr) {
751 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr))
752 return GEP->isInBounds();
756 /// \brief Return true if an AddRec pointer \p Ptr is unsigned non-wrapping,
757 /// i.e. monotonically increasing/decreasing.
758 static bool isNoWrapAddRec(Value *Ptr, const SCEVAddRecExpr *AR,
759 ScalarEvolution *SE, const Loop *L) {
760 // FIXME: This should probably only return true for NUW.
761 if (AR->getNoWrapFlags(SCEV::NoWrapMask))
764 // Scalar evolution does not propagate the non-wrapping flags to values that
765 // are derived from a non-wrapping induction variable because non-wrapping
766 // could be flow-sensitive.
768 // Look through the potentially overflowing instruction to try to prove
769 // non-wrapping for the *specific* value of Ptr.
771 // The arithmetic implied by an inbounds GEP can't overflow.
772 auto *GEP = dyn_cast<GetElementPtrInst>(Ptr);
773 if (!GEP || !GEP->isInBounds())
776 // Make sure there is only one non-const index and analyze that.
777 Value *NonConstIndex = nullptr;
778 for (auto Index = GEP->idx_begin(); Index != GEP->idx_end(); ++Index)
779 if (!isa<ConstantInt>(*Index)) {
782 NonConstIndex = *Index;
785 // The recurrence is on the pointer, ignore for now.
788 // The index in GEP is signed. It is non-wrapping if it's derived from a NSW
789 // AddRec using a NSW operation.
790 if (auto *OBO = dyn_cast<OverflowingBinaryOperator>(NonConstIndex))
791 if (OBO->hasNoSignedWrap() &&
792 // Assume constant for other the operand so that the AddRec can be
794 isa<ConstantInt>(OBO->getOperand(1))) {
795 auto *OpScev = SE->getSCEV(OBO->getOperand(0));
797 if (auto *OpAR = dyn_cast<SCEVAddRecExpr>(OpScev))
798 return OpAR->getLoop() == L && OpAR->getNoWrapFlags(SCEV::FlagNSW);
804 /// \brief Check whether the access through \p Ptr has a constant stride.
805 int llvm::isStridedPtr(ScalarEvolution *SE, Value *Ptr, const Loop *Lp,
806 const ValueToValueMap &StridesMap) {
807 Type *Ty = Ptr->getType();
808 assert(Ty->isPointerTy() && "Unexpected non-ptr");
810 // Make sure that the pointer does not point to aggregate types.
811 auto *PtrTy = cast<PointerType>(Ty);
812 if (PtrTy->getElementType()->isAggregateType()) {
813 DEBUG(dbgs() << "LAA: Bad stride - Not a pointer to a scalar type"
818 const SCEV *PtrScev = replaceSymbolicStrideSCEV(SE, StridesMap, Ptr);
820 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
822 DEBUG(dbgs() << "LAA: Bad stride - Not an AddRecExpr pointer "
823 << *Ptr << " SCEV: " << *PtrScev << "\n");
827 // The accesss function must stride over the innermost loop.
828 if (Lp != AR->getLoop()) {
829 DEBUG(dbgs() << "LAA: Bad stride - Not striding over innermost loop " <<
830 *Ptr << " SCEV: " << *PtrScev << "\n");
833 // The address calculation must not wrap. Otherwise, a dependence could be
835 // An inbounds getelementptr that is a AddRec with a unit stride
836 // cannot wrap per definition. The unit stride requirement is checked later.
837 // An getelementptr without an inbounds attribute and unit stride would have
838 // to access the pointer value "0" which is undefined behavior in address
839 // space 0, therefore we can also vectorize this case.
840 bool IsInBoundsGEP = isInBoundsGep(Ptr);
841 bool IsNoWrapAddRec = isNoWrapAddRec(Ptr, AR, SE, Lp);
842 bool IsInAddressSpaceZero = PtrTy->getAddressSpace() == 0;
843 if (!IsNoWrapAddRec && !IsInBoundsGEP && !IsInAddressSpaceZero) {
844 DEBUG(dbgs() << "LAA: Bad stride - Pointer may wrap in the address space "
845 << *Ptr << " SCEV: " << *PtrScev << "\n");
849 // Check the step is constant.
850 const SCEV *Step = AR->getStepRecurrence(*SE);
852 // Calculate the pointer stride and check if it is constant.
853 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
855 DEBUG(dbgs() << "LAA: Bad stride - Not a constant strided " << *Ptr <<
856 " SCEV: " << *PtrScev << "\n");
860 auto &DL = Lp->getHeader()->getModule()->getDataLayout();
861 int64_t Size = DL.getTypeAllocSize(PtrTy->getElementType());
862 const APInt &APStepVal = C->getValue()->getValue();
864 // Huge step value - give up.
865 if (APStepVal.getBitWidth() > 64)
868 int64_t StepVal = APStepVal.getSExtValue();
871 int64_t Stride = StepVal / Size;
872 int64_t Rem = StepVal % Size;
876 // If the SCEV could wrap but we have an inbounds gep with a unit stride we
877 // know we can't "wrap around the address space". In case of address space
878 // zero we know that this won't happen without triggering undefined behavior.
879 if (!IsNoWrapAddRec && (IsInBoundsGEP || IsInAddressSpaceZero) &&
880 Stride != 1 && Stride != -1)
886 bool MemoryDepChecker::Dependence::isSafeForVectorization(DepType Type) {
890 case BackwardVectorizable:
894 case ForwardButPreventsForwarding:
896 case BackwardVectorizableButPreventsForwarding:
899 llvm_unreachable("unexpected DepType!");
902 bool MemoryDepChecker::Dependence::isInterestingDependence(DepType Type) {
908 case BackwardVectorizable:
910 case ForwardButPreventsForwarding:
912 case BackwardVectorizableButPreventsForwarding:
915 llvm_unreachable("unexpected DepType!");
918 bool MemoryDepChecker::Dependence::isPossiblyBackward() const {
922 case ForwardButPreventsForwarding:
926 case BackwardVectorizable:
928 case BackwardVectorizableButPreventsForwarding:
931 llvm_unreachable("unexpected DepType!");
934 bool MemoryDepChecker::couldPreventStoreLoadForward(unsigned Distance,
935 unsigned TypeByteSize) {
936 // If loads occur at a distance that is not a multiple of a feasible vector
937 // factor store-load forwarding does not take place.
938 // Positive dependences might cause troubles because vectorizing them might
939 // prevent store-load forwarding making vectorized code run a lot slower.
940 // a[i] = a[i-3] ^ a[i-8];
941 // The stores to a[i:i+1] don't align with the stores to a[i-3:i-2] and
942 // hence on your typical architecture store-load forwarding does not take
943 // place. Vectorizing in such cases does not make sense.
944 // Store-load forwarding distance.
945 const unsigned NumCyclesForStoreLoadThroughMemory = 8*TypeByteSize;
946 // Maximum vector factor.
947 unsigned MaxVFWithoutSLForwardIssues =
948 VectorizerParams::MaxVectorWidth * TypeByteSize;
949 if(MaxSafeDepDistBytes < MaxVFWithoutSLForwardIssues)
950 MaxVFWithoutSLForwardIssues = MaxSafeDepDistBytes;
952 for (unsigned vf = 2*TypeByteSize; vf <= MaxVFWithoutSLForwardIssues;
954 if (Distance % vf && Distance / vf < NumCyclesForStoreLoadThroughMemory) {
955 MaxVFWithoutSLForwardIssues = (vf >>=1);
960 if (MaxVFWithoutSLForwardIssues< 2*TypeByteSize) {
961 DEBUG(dbgs() << "LAA: Distance " << Distance <<
962 " that could cause a store-load forwarding conflict\n");
966 if (MaxVFWithoutSLForwardIssues < MaxSafeDepDistBytes &&
967 MaxVFWithoutSLForwardIssues !=
968 VectorizerParams::MaxVectorWidth * TypeByteSize)
969 MaxSafeDepDistBytes = MaxVFWithoutSLForwardIssues;
973 /// \brief Check the dependence for two accesses with the same stride \p Stride.
974 /// \p Distance is the positive distance and \p TypeByteSize is type size in
977 /// \returns true if they are independent.
978 static bool areStridedAccessesIndependent(unsigned Distance, unsigned Stride,
979 unsigned TypeByteSize) {
980 assert(Stride > 1 && "The stride must be greater than 1");
981 assert(TypeByteSize > 0 && "The type size in byte must be non-zero");
982 assert(Distance > 0 && "The distance must be non-zero");
984 // Skip if the distance is not multiple of type byte size.
985 if (Distance % TypeByteSize)
988 unsigned ScaledDist = Distance / TypeByteSize;
990 // No dependence if the scaled distance is not multiple of the stride.
992 // for (i = 0; i < 1024 ; i += 4)
993 // A[i+2] = A[i] + 1;
995 // Two accesses in memory (scaled distance is 2, stride is 4):
996 // | A[0] | | | | A[4] | | | |
997 // | | | A[2] | | | | A[6] | |
1000 // for (i = 0; i < 1024 ; i += 3)
1001 // A[i+4] = A[i] + 1;
1003 // Two accesses in memory (scaled distance is 4, stride is 3):
1004 // | A[0] | | | A[3] | | | A[6] | | |
1005 // | | | | | A[4] | | | A[7] | |
1006 return ScaledDist % Stride;
1009 MemoryDepChecker::Dependence::DepType
1010 MemoryDepChecker::isDependent(const MemAccessInfo &A, unsigned AIdx,
1011 const MemAccessInfo &B, unsigned BIdx,
1012 const ValueToValueMap &Strides) {
1013 assert (AIdx < BIdx && "Must pass arguments in program order");
1015 Value *APtr = A.getPointer();
1016 Value *BPtr = B.getPointer();
1017 bool AIsWrite = A.getInt();
1018 bool BIsWrite = B.getInt();
1020 // Two reads are independent.
1021 if (!AIsWrite && !BIsWrite)
1022 return Dependence::NoDep;
1024 // We cannot check pointers in different address spaces.
1025 if (APtr->getType()->getPointerAddressSpace() !=
1026 BPtr->getType()->getPointerAddressSpace())
1027 return Dependence::Unknown;
1029 const SCEV *AScev = replaceSymbolicStrideSCEV(SE, Strides, APtr);
1030 const SCEV *BScev = replaceSymbolicStrideSCEV(SE, Strides, BPtr);
1032 int StrideAPtr = isStridedPtr(SE, APtr, InnermostLoop, Strides);
1033 int StrideBPtr = isStridedPtr(SE, BPtr, InnermostLoop, Strides);
1035 const SCEV *Src = AScev;
1036 const SCEV *Sink = BScev;
1038 // If the induction step is negative we have to invert source and sink of the
1040 if (StrideAPtr < 0) {
1043 std::swap(APtr, BPtr);
1044 std::swap(Src, Sink);
1045 std::swap(AIsWrite, BIsWrite);
1046 std::swap(AIdx, BIdx);
1047 std::swap(StrideAPtr, StrideBPtr);
1050 const SCEV *Dist = SE->getMinusSCEV(Sink, Src);
1052 DEBUG(dbgs() << "LAA: Src Scev: " << *Src << "Sink Scev: " << *Sink
1053 << "(Induction step: " << StrideAPtr << ")\n");
1054 DEBUG(dbgs() << "LAA: Distance for " << *InstMap[AIdx] << " to "
1055 << *InstMap[BIdx] << ": " << *Dist << "\n");
1057 // Need accesses with constant stride. We don't want to vectorize
1058 // "A[B[i]] += ..." and similar code or pointer arithmetic that could wrap in
1059 // the address space.
1060 if (!StrideAPtr || !StrideBPtr || StrideAPtr != StrideBPtr){
1061 DEBUG(dbgs() << "Pointer access with non-constant stride\n");
1062 return Dependence::Unknown;
1065 const SCEVConstant *C = dyn_cast<SCEVConstant>(Dist);
1067 DEBUG(dbgs() << "LAA: Dependence because of non-constant distance\n");
1068 ShouldRetryWithRuntimeCheck = true;
1069 return Dependence::Unknown;
1072 Type *ATy = APtr->getType()->getPointerElementType();
1073 Type *BTy = BPtr->getType()->getPointerElementType();
1074 auto &DL = InnermostLoop->getHeader()->getModule()->getDataLayout();
1075 unsigned TypeByteSize = DL.getTypeAllocSize(ATy);
1077 // Negative distances are not plausible dependencies.
1078 const APInt &Val = C->getValue()->getValue();
1079 if (Val.isNegative()) {
1080 bool IsTrueDataDependence = (AIsWrite && !BIsWrite);
1081 if (IsTrueDataDependence &&
1082 (couldPreventStoreLoadForward(Val.abs().getZExtValue(), TypeByteSize) ||
1084 return Dependence::ForwardButPreventsForwarding;
1086 DEBUG(dbgs() << "LAA: Dependence is negative: NoDep\n");
1087 return Dependence::Forward;
1090 // Write to the same location with the same size.
1091 // Could be improved to assert type sizes are the same (i32 == float, etc).
1094 return Dependence::NoDep;
1095 DEBUG(dbgs() << "LAA: Zero dependence difference but different types\n");
1096 return Dependence::Unknown;
1099 assert(Val.isStrictlyPositive() && "Expect a positive value");
1103 "LAA: ReadWrite-Write positive dependency with different types\n");
1104 return Dependence::Unknown;
1107 unsigned Distance = (unsigned) Val.getZExtValue();
1109 unsigned Stride = std::abs(StrideAPtr);
1111 areStridedAccessesIndependent(Distance, Stride, TypeByteSize)) {
1112 DEBUG(dbgs() << "LAA: Strided accesses are independent\n");
1113 return Dependence::NoDep;
1116 // Bail out early if passed-in parameters make vectorization not feasible.
1117 unsigned ForcedFactor = (VectorizerParams::VectorizationFactor ?
1118 VectorizerParams::VectorizationFactor : 1);
1119 unsigned ForcedUnroll = (VectorizerParams::VectorizationInterleave ?
1120 VectorizerParams::VectorizationInterleave : 1);
1121 // The minimum number of iterations for a vectorized/unrolled version.
1122 unsigned MinNumIter = std::max(ForcedFactor * ForcedUnroll, 2U);
1124 // It's not vectorizable if the distance is smaller than the minimum distance
1125 // needed for a vectroized/unrolled version. Vectorizing one iteration in
1126 // front needs TypeByteSize * Stride. Vectorizing the last iteration needs
1127 // TypeByteSize (No need to plus the last gap distance).
1129 // E.g. Assume one char is 1 byte in memory and one int is 4 bytes.
1131 // int *B = (int *)((char *)A + 14);
1132 // for (i = 0 ; i < 1024 ; i += 2)
1136 // Two accesses in memory (stride is 2):
1137 // | A[0] | | A[2] | | A[4] | | A[6] | |
1138 // | B[0] | | B[2] | | B[4] |
1140 // Distance needs for vectorizing iterations except the last iteration:
1141 // 4 * 2 * (MinNumIter - 1). Distance needs for the last iteration: 4.
1142 // So the minimum distance needed is: 4 * 2 * (MinNumIter - 1) + 4.
1144 // If MinNumIter is 2, it is vectorizable as the minimum distance needed is
1145 // 12, which is less than distance.
1147 // If MinNumIter is 4 (Say if a user forces the vectorization factor to be 4),
1148 // the minimum distance needed is 28, which is greater than distance. It is
1149 // not safe to do vectorization.
1150 unsigned MinDistanceNeeded =
1151 TypeByteSize * Stride * (MinNumIter - 1) + TypeByteSize;
1152 if (MinDistanceNeeded > Distance) {
1153 DEBUG(dbgs() << "LAA: Failure because of positive distance " << Distance
1155 return Dependence::Backward;
1158 // Unsafe if the minimum distance needed is greater than max safe distance.
1159 if (MinDistanceNeeded > MaxSafeDepDistBytes) {
1160 DEBUG(dbgs() << "LAA: Failure because it needs at least "
1161 << MinDistanceNeeded << " size in bytes");
1162 return Dependence::Backward;
1165 // Positive distance bigger than max vectorization factor.
1166 // FIXME: Should use max factor instead of max distance in bytes, which could
1167 // not handle different types.
1168 // E.g. Assume one char is 1 byte in memory and one int is 4 bytes.
1169 // void foo (int *A, char *B) {
1170 // for (unsigned i = 0; i < 1024; i++) {
1171 // A[i+2] = A[i] + 1;
1172 // B[i+2] = B[i] + 1;
1176 // This case is currently unsafe according to the max safe distance. If we
1177 // analyze the two accesses on array B, the max safe dependence distance
1178 // is 2. Then we analyze the accesses on array A, the minimum distance needed
1179 // is 8, which is less than 2 and forbidden vectorization, But actually
1180 // both A and B could be vectorized by 2 iterations.
1181 MaxSafeDepDistBytes =
1182 Distance < MaxSafeDepDistBytes ? Distance : MaxSafeDepDistBytes;
1184 bool IsTrueDataDependence = (!AIsWrite && BIsWrite);
1185 if (IsTrueDataDependence &&
1186 couldPreventStoreLoadForward(Distance, TypeByteSize))
1187 return Dependence::BackwardVectorizableButPreventsForwarding;
1189 DEBUG(dbgs() << "LAA: Positive distance " << Val.getSExtValue()
1190 << " with max VF = "
1191 << MaxSafeDepDistBytes / (TypeByteSize * Stride) << '\n');
1193 return Dependence::BackwardVectorizable;
1196 bool MemoryDepChecker::areDepsSafe(DepCandidates &AccessSets,
1197 MemAccessInfoSet &CheckDeps,
1198 const ValueToValueMap &Strides) {
1200 MaxSafeDepDistBytes = -1U;
1201 while (!CheckDeps.empty()) {
1202 MemAccessInfo CurAccess = *CheckDeps.begin();
1204 // Get the relevant memory access set.
1205 EquivalenceClasses<MemAccessInfo>::iterator I =
1206 AccessSets.findValue(AccessSets.getLeaderValue(CurAccess));
1208 // Check accesses within this set.
1209 EquivalenceClasses<MemAccessInfo>::member_iterator AI, AE;
1210 AI = AccessSets.member_begin(I), AE = AccessSets.member_end();
1212 // Check every access pair.
1214 CheckDeps.erase(*AI);
1215 EquivalenceClasses<MemAccessInfo>::member_iterator OI = std::next(AI);
1217 // Check every accessing instruction pair in program order.
1218 for (std::vector<unsigned>::iterator I1 = Accesses[*AI].begin(),
1219 I1E = Accesses[*AI].end(); I1 != I1E; ++I1)
1220 for (std::vector<unsigned>::iterator I2 = Accesses[*OI].begin(),
1221 I2E = Accesses[*OI].end(); I2 != I2E; ++I2) {
1222 auto A = std::make_pair(&*AI, *I1);
1223 auto B = std::make_pair(&*OI, *I2);
1229 Dependence::DepType Type =
1230 isDependent(*A.first, A.second, *B.first, B.second, Strides);
1231 SafeForVectorization &= Dependence::isSafeForVectorization(Type);
1233 // Gather dependences unless we accumulated MaxInterestingDependence
1234 // dependences. In that case return as soon as we find the first
1235 // unsafe dependence. This puts a limit on this quadratic
1237 if (RecordInterestingDependences) {
1238 if (Dependence::isInterestingDependence(Type))
1239 InterestingDependences.push_back(
1240 Dependence(A.second, B.second, Type));
1242 if (InterestingDependences.size() >= MaxInterestingDependence) {
1243 RecordInterestingDependences = false;
1244 InterestingDependences.clear();
1245 DEBUG(dbgs() << "Too many dependences, stopped recording\n");
1248 if (!RecordInterestingDependences && !SafeForVectorization)
1257 DEBUG(dbgs() << "Total Interesting Dependences: "
1258 << InterestingDependences.size() << "\n");
1259 return SafeForVectorization;
1262 SmallVector<Instruction *, 4>
1263 MemoryDepChecker::getInstructionsForAccess(Value *Ptr, bool isWrite) const {
1264 MemAccessInfo Access(Ptr, isWrite);
1265 auto &IndexVector = Accesses.find(Access)->second;
1267 SmallVector<Instruction *, 4> Insts;
1268 std::transform(IndexVector.begin(), IndexVector.end(),
1269 std::back_inserter(Insts),
1270 [&](unsigned Idx) { return this->InstMap[Idx]; });
1274 const char *MemoryDepChecker::Dependence::DepName[] = {
1275 "NoDep", "Unknown", "Forward", "ForwardButPreventsForwarding", "Backward",
1276 "BackwardVectorizable", "BackwardVectorizableButPreventsForwarding"};
1278 void MemoryDepChecker::Dependence::print(
1279 raw_ostream &OS, unsigned Depth,
1280 const SmallVectorImpl<Instruction *> &Instrs) const {
1281 OS.indent(Depth) << DepName[Type] << ":\n";
1282 OS.indent(Depth + 2) << *Instrs[Source] << " -> \n";
1283 OS.indent(Depth + 2) << *Instrs[Destination] << "\n";
1286 bool LoopAccessInfo::canAnalyzeLoop() {
1287 // We need to have a loop header.
1288 DEBUG(dbgs() << "LAA: Found a loop: " <<
1289 TheLoop->getHeader()->getName() << '\n');
1291 // We can only analyze innermost loops.
1292 if (!TheLoop->empty()) {
1293 DEBUG(dbgs() << "LAA: loop is not the innermost loop\n");
1294 emitAnalysis(LoopAccessReport() << "loop is not the innermost loop");
1298 // We must have a single backedge.
1299 if (TheLoop->getNumBackEdges() != 1) {
1300 DEBUG(dbgs() << "LAA: loop control flow is not understood by analyzer\n");
1302 LoopAccessReport() <<
1303 "loop control flow is not understood by analyzer");
1307 // We must have a single exiting block.
1308 if (!TheLoop->getExitingBlock()) {
1309 DEBUG(dbgs() << "LAA: loop control flow is not understood by analyzer\n");
1311 LoopAccessReport() <<
1312 "loop control flow is not understood by analyzer");
1316 // We only handle bottom-tested loops, i.e. loop in which the condition is
1317 // checked at the end of each iteration. With that we can assume that all
1318 // instructions in the loop are executed the same number of times.
1319 if (TheLoop->getExitingBlock() != TheLoop->getLoopLatch()) {
1320 DEBUG(dbgs() << "LAA: loop control flow is not understood by analyzer\n");
1322 LoopAccessReport() <<
1323 "loop control flow is not understood by analyzer");
1327 // ScalarEvolution needs to be able to find the exit count.
1328 const SCEV *ExitCount = SE->getBackedgeTakenCount(TheLoop);
1329 if (ExitCount == SE->getCouldNotCompute()) {
1330 emitAnalysis(LoopAccessReport() <<
1331 "could not determine number of loop iterations");
1332 DEBUG(dbgs() << "LAA: SCEV could not compute the loop exit count.\n");
1339 void LoopAccessInfo::analyzeLoop(const ValueToValueMap &Strides) {
1341 typedef SmallVector<Value*, 16> ValueVector;
1342 typedef SmallPtrSet<Value*, 16> ValueSet;
1344 // Holds the Load and Store *instructions*.
1348 // Holds all the different accesses in the loop.
1349 unsigned NumReads = 0;
1350 unsigned NumReadWrites = 0;
1352 PtrRtChecking.Pointers.clear();
1353 PtrRtChecking.Need = false;
1355 const bool IsAnnotatedParallel = TheLoop->isAnnotatedParallel();
1358 for (Loop::block_iterator bb = TheLoop->block_begin(),
1359 be = TheLoop->block_end(); bb != be; ++bb) {
1361 // Scan the BB and collect legal loads and stores.
1362 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
1365 // If this is a load, save it. If this instruction can read from memory
1366 // but is not a load, then we quit. Notice that we don't handle function
1367 // calls that read or write.
1368 if (it->mayReadFromMemory()) {
1369 // Many math library functions read the rounding mode. We will only
1370 // vectorize a loop if it contains known function calls that don't set
1371 // the flag. Therefore, it is safe to ignore this read from memory.
1372 CallInst *Call = dyn_cast<CallInst>(it);
1373 if (Call && getIntrinsicIDForCall(Call, TLI))
1376 // If the function has an explicit vectorized counterpart, we can safely
1377 // assume that it can be vectorized.
1378 if (Call && !Call->isNoBuiltin() && Call->getCalledFunction() &&
1379 TLI->isFunctionVectorizable(Call->getCalledFunction()->getName()))
1382 LoadInst *Ld = dyn_cast<LoadInst>(it);
1383 if (!Ld || (!Ld->isSimple() && !IsAnnotatedParallel)) {
1384 emitAnalysis(LoopAccessReport(Ld)
1385 << "read with atomic ordering or volatile read");
1386 DEBUG(dbgs() << "LAA: Found a non-simple load.\n");
1391 Loads.push_back(Ld);
1392 DepChecker.addAccess(Ld);
1396 // Save 'store' instructions. Abort if other instructions write to memory.
1397 if (it->mayWriteToMemory()) {
1398 StoreInst *St = dyn_cast<StoreInst>(it);
1400 emitAnalysis(LoopAccessReport(it) <<
1401 "instruction cannot be vectorized");
1405 if (!St->isSimple() && !IsAnnotatedParallel) {
1406 emitAnalysis(LoopAccessReport(St)
1407 << "write with atomic ordering or volatile write");
1408 DEBUG(dbgs() << "LAA: Found a non-simple store.\n");
1413 Stores.push_back(St);
1414 DepChecker.addAccess(St);
1419 // Now we have two lists that hold the loads and the stores.
1420 // Next, we find the pointers that they use.
1422 // Check if we see any stores. If there are no stores, then we don't
1423 // care if the pointers are *restrict*.
1424 if (!Stores.size()) {
1425 DEBUG(dbgs() << "LAA: Found a read-only loop!\n");
1430 MemoryDepChecker::DepCandidates DependentAccesses;
1431 AccessAnalysis Accesses(TheLoop->getHeader()->getModule()->getDataLayout(),
1432 AA, LI, DependentAccesses);
1434 // Holds the analyzed pointers. We don't want to call GetUnderlyingObjects
1435 // multiple times on the same object. If the ptr is accessed twice, once
1436 // for read and once for write, it will only appear once (on the write
1437 // list). This is okay, since we are going to check for conflicts between
1438 // writes and between reads and writes, but not between reads and reads.
1441 ValueVector::iterator I, IE;
1442 for (I = Stores.begin(), IE = Stores.end(); I != IE; ++I) {
1443 StoreInst *ST = cast<StoreInst>(*I);
1444 Value* Ptr = ST->getPointerOperand();
1445 // Check for store to loop invariant address.
1446 StoreToLoopInvariantAddress |= isUniform(Ptr);
1447 // If we did *not* see this pointer before, insert it to the read-write
1448 // list. At this phase it is only a 'write' list.
1449 if (Seen.insert(Ptr).second) {
1452 MemoryLocation Loc = MemoryLocation::get(ST);
1453 // The TBAA metadata could have a control dependency on the predication
1454 // condition, so we cannot rely on it when determining whether or not we
1455 // need runtime pointer checks.
1456 if (blockNeedsPredication(ST->getParent(), TheLoop, DT))
1457 Loc.AATags.TBAA = nullptr;
1459 Accesses.addStore(Loc);
1463 if (IsAnnotatedParallel) {
1465 << "LAA: A loop annotated parallel, ignore memory dependency "
1471 for (I = Loads.begin(), IE = Loads.end(); I != IE; ++I) {
1472 LoadInst *LD = cast<LoadInst>(*I);
1473 Value* Ptr = LD->getPointerOperand();
1474 // If we did *not* see this pointer before, insert it to the
1475 // read list. If we *did* see it before, then it is already in
1476 // the read-write list. This allows us to vectorize expressions
1477 // such as A[i] += x; Because the address of A[i] is a read-write
1478 // pointer. This only works if the index of A[i] is consecutive.
1479 // If the address of i is unknown (for example A[B[i]]) then we may
1480 // read a few words, modify, and write a few words, and some of the
1481 // words may be written to the same address.
1482 bool IsReadOnlyPtr = false;
1483 if (Seen.insert(Ptr).second || !isStridedPtr(SE, Ptr, TheLoop, Strides)) {
1485 IsReadOnlyPtr = true;
1488 MemoryLocation Loc = MemoryLocation::get(LD);
1489 // The TBAA metadata could have a control dependency on the predication
1490 // condition, so we cannot rely on it when determining whether or not we
1491 // need runtime pointer checks.
1492 if (blockNeedsPredication(LD->getParent(), TheLoop, DT))
1493 Loc.AATags.TBAA = nullptr;
1495 Accesses.addLoad(Loc, IsReadOnlyPtr);
1498 // If we write (or read-write) to a single destination and there are no
1499 // other reads in this loop then is it safe to vectorize.
1500 if (NumReadWrites == 1 && NumReads == 0) {
1501 DEBUG(dbgs() << "LAA: Found a write-only loop!\n");
1506 // Build dependence sets and check whether we need a runtime pointer bounds
1508 Accesses.buildDependenceSets();
1510 // Find pointers with computable bounds. We are going to use this information
1511 // to place a runtime bound check.
1512 bool CanDoRTIfNeeded =
1513 Accesses.canCheckPtrAtRT(PtrRtChecking, SE, TheLoop, Strides);
1514 if (!CanDoRTIfNeeded) {
1515 emitAnalysis(LoopAccessReport() << "cannot identify array bounds");
1516 DEBUG(dbgs() << "LAA: We can't vectorize because we can't find "
1517 << "the array bounds.\n");
1522 DEBUG(dbgs() << "LAA: We can perform a memory runtime check if needed.\n");
1525 if (Accesses.isDependencyCheckNeeded()) {
1526 DEBUG(dbgs() << "LAA: Checking memory dependencies\n");
1527 CanVecMem = DepChecker.areDepsSafe(
1528 DependentAccesses, Accesses.getDependenciesToCheck(), Strides);
1529 MaxSafeDepDistBytes = DepChecker.getMaxSafeDepDistBytes();
1531 if (!CanVecMem && DepChecker.shouldRetryWithRuntimeCheck()) {
1532 DEBUG(dbgs() << "LAA: Retrying with memory checks\n");
1534 // Clear the dependency checks. We assume they are not needed.
1535 Accesses.resetDepChecks(DepChecker);
1537 PtrRtChecking.reset();
1538 PtrRtChecking.Need = true;
1541 Accesses.canCheckPtrAtRT(PtrRtChecking, SE, TheLoop, Strides, true);
1543 // Check that we found the bounds for the pointer.
1544 if (!CanDoRTIfNeeded) {
1545 emitAnalysis(LoopAccessReport()
1546 << "cannot check memory dependencies at runtime");
1547 DEBUG(dbgs() << "LAA: Can't vectorize with memory checks\n");
1557 DEBUG(dbgs() << "LAA: No unsafe dependent memory operations in loop. We"
1558 << (PtrRtChecking.Need ? "" : " don't")
1559 << " need runtime memory checks.\n");
1561 emitAnalysis(LoopAccessReport() <<
1562 "unsafe dependent memory operations in loop");
1563 DEBUG(dbgs() << "LAA: unsafe dependent memory operations in loop\n");
1567 bool LoopAccessInfo::blockNeedsPredication(BasicBlock *BB, Loop *TheLoop,
1568 DominatorTree *DT) {
1569 assert(TheLoop->contains(BB) && "Unknown block used");
1571 // Blocks that do not dominate the latch need predication.
1572 BasicBlock* Latch = TheLoop->getLoopLatch();
1573 return !DT->dominates(BB, Latch);
1576 void LoopAccessInfo::emitAnalysis(LoopAccessReport &Message) {
1577 assert(!Report && "Multiple reports generated");
1581 bool LoopAccessInfo::isUniform(Value *V) const {
1582 return (SE->isLoopInvariant(SE->getSCEV(V), TheLoop));
1585 // FIXME: this function is currently a duplicate of the one in
1586 // LoopVectorize.cpp.
1587 static Instruction *getFirstInst(Instruction *FirstInst, Value *V,
1591 if (Instruction *I = dyn_cast<Instruction>(V))
1592 return I->getParent() == Loc->getParent() ? I : nullptr;
1596 /// \brief IR Values for the lower and upper bounds of a pointer evolution. We
1597 /// need to use value-handles because SCEV expansion can invalidate previously
1598 /// expanded values. Thus expansion of a pointer can invalidate the bounds for
1600 struct PointerBounds {
1601 TrackingVH<Value> Start;
1602 TrackingVH<Value> End;
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::addRuntimeChecks(
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 // Check if two pointers (A and B) conflict where conflict is computed as:
1675 // start(A) <= end(B) && start(B) <= end(A)
1676 unsigned AS0 = A.Start->getType()->getPointerAddressSpace();
1677 unsigned AS1 = B.Start->getType()->getPointerAddressSpace();
1679 assert((AS0 == B.End->getType()->getPointerAddressSpace()) &&
1680 (AS1 == A.End->getType()->getPointerAddressSpace()) &&
1681 "Trying to bounds check pointers with different address spaces");
1683 Type *PtrArithTy0 = Type::getInt8PtrTy(Ctx, AS0);
1684 Type *PtrArithTy1 = Type::getInt8PtrTy(Ctx, AS1);
1686 Value *Start0 = ChkBuilder.CreateBitCast(A.Start, PtrArithTy0, "bc");
1687 Value *Start1 = ChkBuilder.CreateBitCast(B.Start, PtrArithTy1, "bc");
1688 Value *End0 = ChkBuilder.CreateBitCast(A.End, PtrArithTy1, "bc");
1689 Value *End1 = ChkBuilder.CreateBitCast(B.End, PtrArithTy0, "bc");
1691 Value *Cmp0 = ChkBuilder.CreateICmpULE(Start0, End1, "bound0");
1692 FirstInst = getFirstInst(FirstInst, Cmp0, Loc);
1693 Value *Cmp1 = ChkBuilder.CreateICmpULE(Start1, End0, "bound1");
1694 FirstInst = getFirstInst(FirstInst, Cmp1, Loc);
1695 Value *IsConflict = ChkBuilder.CreateAnd(Cmp0, Cmp1, "found.conflict");
1696 FirstInst = getFirstInst(FirstInst, IsConflict, Loc);
1697 if (MemoryRuntimeCheck) {
1699 ChkBuilder.CreateOr(MemoryRuntimeCheck, IsConflict, "conflict.rdx");
1700 FirstInst = getFirstInst(FirstInst, IsConflict, Loc);
1702 MemoryRuntimeCheck = IsConflict;
1705 if (!MemoryRuntimeCheck)
1706 return std::make_pair(nullptr, nullptr);
1708 // We have to do this trickery because the IRBuilder might fold the check to a
1709 // constant expression in which case there is no Instruction anchored in a
1711 Instruction *Check = BinaryOperator::CreateAnd(MemoryRuntimeCheck,
1712 ConstantInt::getTrue(Ctx));
1713 ChkBuilder.Insert(Check, "memcheck.conflict");
1714 FirstInst = getFirstInst(FirstInst, Check, Loc);
1715 return std::make_pair(FirstInst, Check);
1718 std::pair<Instruction *, Instruction *>
1719 LoopAccessInfo::addRuntimeChecks(Instruction *Loc) const {
1720 if (!PtrRtChecking.Need)
1721 return std::make_pair(nullptr, nullptr);
1723 return addRuntimeChecks(Loc, PtrRtChecking.getChecks());
1726 LoopAccessInfo::LoopAccessInfo(Loop *L, ScalarEvolution *SE,
1727 const DataLayout &DL,
1728 const TargetLibraryInfo *TLI, AliasAnalysis *AA,
1729 DominatorTree *DT, LoopInfo *LI,
1730 const ValueToValueMap &Strides)
1731 : PtrRtChecking(SE), DepChecker(SE, L), TheLoop(L), SE(SE), DL(DL),
1732 TLI(TLI), AA(AA), DT(DT), LI(LI), NumLoads(0), NumStores(0),
1733 MaxSafeDepDistBytes(-1U), CanVecMem(false),
1734 StoreToLoopInvariantAddress(false) {
1735 if (canAnalyzeLoop())
1736 analyzeLoop(Strides);
1739 void LoopAccessInfo::print(raw_ostream &OS, unsigned Depth) const {
1741 if (PtrRtChecking.Need)
1742 OS.indent(Depth) << "Memory dependences are safe with run-time checks\n";
1744 OS.indent(Depth) << "Memory dependences are safe\n";
1748 OS.indent(Depth) << "Report: " << Report->str() << "\n";
1750 if (auto *InterestingDependences = DepChecker.getInterestingDependences()) {
1751 OS.indent(Depth) << "Interesting Dependences:\n";
1752 for (auto &Dep : *InterestingDependences) {
1753 Dep.print(OS, Depth + 2, DepChecker.getMemoryInstructions());
1757 OS.indent(Depth) << "Too many interesting dependences, not recorded\n";
1759 // List the pair of accesses need run-time checks to prove independence.
1760 PtrRtChecking.print(OS, Depth);
1763 OS.indent(Depth) << "Store to invariant address was "
1764 << (StoreToLoopInvariantAddress ? "" : "not ")
1765 << "found in loop.\n";
1768 const LoopAccessInfo &
1769 LoopAccessAnalysis::getInfo(Loop *L, const ValueToValueMap &Strides) {
1770 auto &LAI = LoopAccessInfoMap[L];
1773 assert((!LAI || LAI->NumSymbolicStrides == Strides.size()) &&
1774 "Symbolic strides changed for loop");
1778 const DataLayout &DL = L->getHeader()->getModule()->getDataLayout();
1779 LAI = llvm::make_unique<LoopAccessInfo>(L, SE, DL, TLI, AA, DT, LI,
1782 LAI->NumSymbolicStrides = Strides.size();
1788 void LoopAccessAnalysis::print(raw_ostream &OS, const Module *M) const {
1789 LoopAccessAnalysis &LAA = *const_cast<LoopAccessAnalysis *>(this);
1791 ValueToValueMap NoSymbolicStrides;
1793 for (Loop *TopLevelLoop : *LI)
1794 for (Loop *L : depth_first(TopLevelLoop)) {
1795 OS.indent(2) << L->getHeader()->getName() << ":\n";
1796 auto &LAI = LAA.getInfo(L, NoSymbolicStrides);
1801 bool LoopAccessAnalysis::runOnFunction(Function &F) {
1802 SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE();
1803 auto *TLIP = getAnalysisIfAvailable<TargetLibraryInfoWrapperPass>();
1804 TLI = TLIP ? &TLIP->getTLI() : nullptr;
1805 AA = &getAnalysis<AAResultsWrapperPass>().getAAResults();
1806 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1807 LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
1812 void LoopAccessAnalysis::getAnalysisUsage(AnalysisUsage &AU) const {
1813 AU.addRequired<ScalarEvolutionWrapperPass>();
1814 AU.addRequired<AAResultsWrapperPass>();
1815 AU.addRequired<DominatorTreeWrapperPass>();
1816 AU.addRequired<LoopInfoWrapperPass>();
1818 AU.setPreservesAll();
1821 char LoopAccessAnalysis::ID = 0;
1822 static const char laa_name[] = "Loop Access Analysis";
1823 #define LAA_NAME "loop-accesses"
1825 INITIALIZE_PASS_BEGIN(LoopAccessAnalysis, LAA_NAME, laa_name, false, true)
1826 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)
1827 INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass)
1828 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
1829 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
1830 INITIALIZE_PASS_END(LoopAccessAnalysis, LAA_NAME, laa_name, false, true)
1833 Pass *createLAAPass() {
1834 return new LoopAccessAnalysis();