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 need to check two pointers to the same underlying object
241 // with a non-constant difference, we shouldn't perform any pointer
242 // grouping with those pointers. This is because we can easily get
243 // into cases where the resulting check would return false, even when
244 // the accesses are safe.
246 // The following example shows this:
247 // for (i = 0; i < 1000; ++i)
248 // a[5000 + i * m] = a[i] + a[i + 9000]
250 // Here grouping gives a check of (5000, 5000 + 1000 * m) against
251 // (0, 10000) which is always false. However, if m is 1, there is no
252 // dependence. Not grouping the checks for a[i] and a[i + 9000] allows
253 // us to perform an accurate check in this case.
255 // The above case requires that we have an UnknownDependence between
256 // accesses to the same underlying object. This cannot happen unless
257 // ShouldRetryWithRuntimeCheck is set, and therefore UseDependencies
258 // is also false. In this case we will use the fallback path and create
259 // separate checking groups for all pointers.
261 // If we don't have the dependency partitions, construct a new
262 // checking pointer group for each pointer. This is also required
263 // for correctness, because in this case we can have checking between
264 // pointers to the same underlying object.
265 if (!UseDependencies) {
266 for (unsigned I = 0; I < Pointers.size(); ++I)
267 CheckingGroups.push_back(CheckingPtrGroup(I, *this));
271 unsigned TotalComparisons = 0;
273 DenseMap<Value *, unsigned> PositionMap;
274 for (unsigned Index = 0; Index < Pointers.size(); ++Index)
275 PositionMap[Pointers[Index].PointerValue] = Index;
277 // We need to keep track of what pointers we've already seen so we
278 // don't process them twice.
279 SmallSet<unsigned, 2> Seen;
281 // Go through all equivalence classes, get the the "pointer check groups"
282 // and add them to the overall solution. We use the order in which accesses
283 // appear in 'Pointers' to enforce determinism.
284 for (unsigned I = 0; I < Pointers.size(); ++I) {
285 // We've seen this pointer before, and therefore already processed
286 // its equivalence class.
290 MemoryDepChecker::MemAccessInfo Access(Pointers[I].PointerValue,
291 Pointers[I].IsWritePtr);
293 SmallVector<CheckingPtrGroup, 2> Groups;
294 auto LeaderI = DepCands.findValue(DepCands.getLeaderValue(Access));
296 // Because DepCands is constructed by visiting accesses in the order in
297 // which they appear in alias sets (which is deterministic) and the
298 // iteration order within an equivalence class member is only dependent on
299 // the order in which unions and insertions are performed on the
300 // equivalence class, the iteration order is deterministic.
301 for (auto MI = DepCands.member_begin(LeaderI), ME = DepCands.member_end();
303 unsigned Pointer = PositionMap[MI->getPointer()];
305 // Mark this pointer as seen.
306 Seen.insert(Pointer);
308 // Go through all the existing sets and see if we can find one
309 // which can include this pointer.
310 for (CheckingPtrGroup &Group : Groups) {
311 // Don't perform more than a certain amount of comparisons.
312 // This should limit the cost of grouping the pointers to something
313 // reasonable. If we do end up hitting this threshold, the algorithm
314 // will create separate groups for all remaining pointers.
315 if (TotalComparisons > MemoryCheckMergeThreshold)
320 if (Group.addPointer(Pointer)) {
327 // We couldn't add this pointer to any existing set or the threshold
328 // for the number of comparisons has been reached. Create a new group
329 // to hold the current pointer.
330 Groups.push_back(CheckingPtrGroup(Pointer, *this));
333 // We've computed the grouped checks for this partition.
334 // Save the results and continue with the next one.
335 std::copy(Groups.begin(), Groups.end(), std::back_inserter(CheckingGroups));
339 bool RuntimePointerChecking::arePointersInSamePartition(
340 const SmallVectorImpl<int> &PtrToPartition, unsigned PtrIdx1,
342 return (PtrToPartition[PtrIdx1] != -1 &&
343 PtrToPartition[PtrIdx1] == PtrToPartition[PtrIdx2]);
346 bool RuntimePointerChecking::needsChecking(
347 unsigned I, unsigned J, const SmallVectorImpl<int> *PtrPartition) const {
348 const PointerInfo &PointerI = Pointers[I];
349 const PointerInfo &PointerJ = Pointers[J];
351 // No need to check if two readonly pointers intersect.
352 if (!PointerI.IsWritePtr && !PointerJ.IsWritePtr)
355 // Only need to check pointers between two different dependency sets.
356 if (PointerI.DependencySetId == PointerJ.DependencySetId)
359 // Only need to check pointers in the same alias set.
360 if (PointerI.AliasSetId != PointerJ.AliasSetId)
363 // If PtrPartition is set omit checks between pointers of the same partition.
364 if (PtrPartition && arePointersInSamePartition(*PtrPartition, I, J))
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(
390 raw_ostream &OS, unsigned Depth,
391 const SmallVectorImpl<int> *PtrPartition) const {
393 OS.indent(Depth) << "Run-time memory checks:\n";
394 printChecks(OS, generateChecks(PtrPartition), Depth);
396 OS.indent(Depth) << "Grouped accesses:\n";
397 for (unsigned I = 0; I < CheckingGroups.size(); ++I) {
398 const auto &CG = CheckingGroups[I];
400 OS.indent(Depth + 2) << "Group " << &CG << ":\n";
401 OS.indent(Depth + 4) << "(Low: " << *CG.Low << " High: " << *CG.High
403 for (unsigned J = 0; J < CG.Members.size(); ++J) {
404 OS.indent(Depth + 6) << "Member: " << *Pointers[CG.Members[J]].Expr
410 unsigned RuntimePointerChecking::getNumberOfChecks(
411 const SmallVectorImpl<int> *PtrPartition) const {
413 unsigned NumPartitions = CheckingGroups.size();
414 unsigned CheckCount = 0;
416 for (unsigned I = 0; I < NumPartitions; ++I)
417 for (unsigned J = I + 1; J < NumPartitions; ++J)
418 if (needsChecking(CheckingGroups[I], CheckingGroups[J], PtrPartition))
424 /// \brief Analyses memory accesses in a loop.
426 /// Checks whether run time pointer checks are needed and builds sets for data
427 /// dependence checking.
428 class AccessAnalysis {
430 /// \brief Read or write access location.
431 typedef PointerIntPair<Value *, 1, bool> MemAccessInfo;
432 typedef SmallPtrSet<MemAccessInfo, 8> MemAccessInfoSet;
434 AccessAnalysis(const DataLayout &Dl, AliasAnalysis *AA, LoopInfo *LI,
435 MemoryDepChecker::DepCandidates &DA)
436 : DL(Dl), AST(*AA), LI(LI), DepCands(DA),
437 IsRTCheckAnalysisNeeded(false) {}
439 /// \brief Register a load and whether it is only read from.
440 void addLoad(MemoryLocation &Loc, bool IsReadOnly) {
441 Value *Ptr = const_cast<Value*>(Loc.Ptr);
442 AST.add(Ptr, MemoryLocation::UnknownSize, Loc.AATags);
443 Accesses.insert(MemAccessInfo(Ptr, false));
445 ReadOnlyPtr.insert(Ptr);
448 /// \brief Register a store.
449 void addStore(MemoryLocation &Loc) {
450 Value *Ptr = const_cast<Value*>(Loc.Ptr);
451 AST.add(Ptr, MemoryLocation::UnknownSize, Loc.AATags);
452 Accesses.insert(MemAccessInfo(Ptr, true));
455 /// \brief Check whether we can check the pointers at runtime for
456 /// non-intersection.
458 /// Returns true if we need no check or if we do and we can generate them
459 /// (i.e. the pointers have computable bounds).
460 bool canCheckPtrAtRT(RuntimePointerChecking &RtCheck, ScalarEvolution *SE,
461 Loop *TheLoop, const ValueToValueMap &Strides,
462 bool ShouldCheckStride = false);
464 /// \brief Goes over all memory accesses, checks whether a RT check is needed
465 /// and builds sets of dependent accesses.
466 void buildDependenceSets() {
467 processMemAccesses();
470 /// \brief Initial processing of memory accesses determined that we need to
471 /// perform dependency checking.
473 /// Note that this can later be cleared if we retry memcheck analysis without
474 /// dependency checking (i.e. ShouldRetryWithRuntimeCheck).
475 bool isDependencyCheckNeeded() { return !CheckDeps.empty(); }
477 /// We decided that no dependence analysis would be used. Reset the state.
478 void resetDepChecks(MemoryDepChecker &DepChecker) {
480 DepChecker.clearInterestingDependences();
483 MemAccessInfoSet &getDependenciesToCheck() { return CheckDeps; }
486 typedef SetVector<MemAccessInfo> PtrAccessSet;
488 /// \brief Go over all memory access and check whether runtime pointer checks
489 /// are needed and build sets of dependency check candidates.
490 void processMemAccesses();
492 /// Set of all accesses.
493 PtrAccessSet Accesses;
495 const DataLayout &DL;
497 /// Set of accesses that need a further dependence check.
498 MemAccessInfoSet CheckDeps;
500 /// Set of pointers that are read only.
501 SmallPtrSet<Value*, 16> ReadOnlyPtr;
503 /// An alias set tracker to partition the access set by underlying object and
504 //intrinsic property (such as TBAA metadata).
509 /// Sets of potentially dependent accesses - members of one set share an
510 /// underlying pointer. The set "CheckDeps" identfies which sets really need a
511 /// dependence check.
512 MemoryDepChecker::DepCandidates &DepCands;
514 /// \brief Initial processing of memory accesses determined that we may need
515 /// to add memchecks. Perform the analysis to determine the necessary checks.
517 /// Note that, this is different from isDependencyCheckNeeded. When we retry
518 /// memcheck analysis without dependency checking
519 /// (i.e. ShouldRetryWithRuntimeCheck), isDependencyCheckNeeded is cleared
520 /// while this remains set if we have potentially dependent accesses.
521 bool IsRTCheckAnalysisNeeded;
524 } // end anonymous namespace
526 /// \brief Check whether a pointer can participate in a runtime bounds check.
527 static bool hasComputableBounds(ScalarEvolution *SE,
528 const ValueToValueMap &Strides, Value *Ptr) {
529 const SCEV *PtrScev = replaceSymbolicStrideSCEV(SE, Strides, Ptr);
530 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
534 return AR->isAffine();
537 bool AccessAnalysis::canCheckPtrAtRT(RuntimePointerChecking &RtCheck,
538 ScalarEvolution *SE, Loop *TheLoop,
539 const ValueToValueMap &StridesMap,
540 bool ShouldCheckStride) {
541 // Find pointers with computable bounds. We are going to use this information
542 // to place a runtime bound check.
545 bool NeedRTCheck = false;
546 if (!IsRTCheckAnalysisNeeded) return true;
548 bool IsDepCheckNeeded = isDependencyCheckNeeded();
550 // We assign a consecutive id to access from different alias sets.
551 // Accesses between different groups doesn't need to be checked.
553 for (auto &AS : AST) {
554 int NumReadPtrChecks = 0;
555 int NumWritePtrChecks = 0;
557 // We assign consecutive id to access from different dependence sets.
558 // Accesses within the same set don't need a runtime check.
559 unsigned RunningDepId = 1;
560 DenseMap<Value *, unsigned> DepSetId;
563 Value *Ptr = A.getValue();
564 bool IsWrite = Accesses.count(MemAccessInfo(Ptr, true));
565 MemAccessInfo Access(Ptr, IsWrite);
572 if (hasComputableBounds(SE, StridesMap, Ptr) &&
573 // When we run after a failing dependency check we have to make sure
574 // we don't have wrapping pointers.
575 (!ShouldCheckStride ||
576 isStridedPtr(SE, Ptr, TheLoop, StridesMap) == 1)) {
577 // The id of the dependence set.
580 if (IsDepCheckNeeded) {
581 Value *Leader = DepCands.getLeaderValue(Access).getPointer();
582 unsigned &LeaderId = DepSetId[Leader];
584 LeaderId = RunningDepId++;
587 // Each access has its own dependence set.
588 DepId = RunningDepId++;
590 RtCheck.insert(TheLoop, Ptr, IsWrite, DepId, ASId, StridesMap);
592 DEBUG(dbgs() << "LAA: Found a runtime check ptr:" << *Ptr << '\n');
594 DEBUG(dbgs() << "LAA: Can't find bounds for ptr:" << *Ptr << '\n');
599 // If we have at least two writes or one write and a read then we need to
600 // check them. But there is no need to checks if there is only one
601 // dependence set for this alias set.
603 // Note that this function computes CanDoRT and NeedRTCheck independently.
604 // For example CanDoRT=false, NeedRTCheck=false means that we have a pointer
605 // for which we couldn't find the bounds but we don't actually need to emit
606 // any checks so it does not matter.
607 if (!(IsDepCheckNeeded && CanDoRT && RunningDepId == 2))
608 NeedRTCheck |= (NumWritePtrChecks >= 2 || (NumReadPtrChecks >= 1 &&
609 NumWritePtrChecks >= 1));
614 // If the pointers that we would use for the bounds comparison have different
615 // address spaces, assume the values aren't directly comparable, so we can't
616 // use them for the runtime check. We also have to assume they could
617 // overlap. In the future there should be metadata for whether address spaces
619 unsigned NumPointers = RtCheck.Pointers.size();
620 for (unsigned i = 0; i < NumPointers; ++i) {
621 for (unsigned j = i + 1; j < NumPointers; ++j) {
622 // Only need to check pointers between two different dependency sets.
623 if (RtCheck.Pointers[i].DependencySetId ==
624 RtCheck.Pointers[j].DependencySetId)
626 // Only need to check pointers in the same alias set.
627 if (RtCheck.Pointers[i].AliasSetId != RtCheck.Pointers[j].AliasSetId)
630 Value *PtrI = RtCheck.Pointers[i].PointerValue;
631 Value *PtrJ = RtCheck.Pointers[j].PointerValue;
633 unsigned ASi = PtrI->getType()->getPointerAddressSpace();
634 unsigned ASj = PtrJ->getType()->getPointerAddressSpace();
636 DEBUG(dbgs() << "LAA: Runtime check would require comparison between"
637 " different address spaces\n");
643 if (NeedRTCheck && CanDoRT)
644 RtCheck.groupChecks(DepCands, IsDepCheckNeeded);
646 DEBUG(dbgs() << "LAA: We need to do " << RtCheck.getNumberOfChecks(nullptr)
647 << " pointer comparisons.\n");
649 RtCheck.Need = NeedRTCheck;
651 bool CanDoRTIfNeeded = !NeedRTCheck || CanDoRT;
652 if (!CanDoRTIfNeeded)
654 return CanDoRTIfNeeded;
657 void AccessAnalysis::processMemAccesses() {
658 // We process the set twice: first we process read-write pointers, last we
659 // process read-only pointers. This allows us to skip dependence tests for
660 // read-only pointers.
662 DEBUG(dbgs() << "LAA: Processing memory accesses...\n");
663 DEBUG(dbgs() << " AST: "; AST.dump());
664 DEBUG(dbgs() << "LAA: Accesses(" << Accesses.size() << "):\n");
666 for (auto A : Accesses)
667 dbgs() << "\t" << *A.getPointer() << " (" <<
668 (A.getInt() ? "write" : (ReadOnlyPtr.count(A.getPointer()) ?
669 "read-only" : "read")) << ")\n";
672 // The AliasSetTracker has nicely partitioned our pointers by metadata
673 // compatibility and potential for underlying-object overlap. As a result, we
674 // only need to check for potential pointer dependencies within each alias
676 for (auto &AS : AST) {
677 // Note that both the alias-set tracker and the alias sets themselves used
678 // linked lists internally and so the iteration order here is deterministic
679 // (matching the original instruction order within each set).
681 bool SetHasWrite = false;
683 // Map of pointers to last access encountered.
684 typedef DenseMap<Value*, MemAccessInfo> UnderlyingObjToAccessMap;
685 UnderlyingObjToAccessMap ObjToLastAccess;
687 // Set of access to check after all writes have been processed.
688 PtrAccessSet DeferredAccesses;
690 // Iterate over each alias set twice, once to process read/write pointers,
691 // and then to process read-only pointers.
692 for (int SetIteration = 0; SetIteration < 2; ++SetIteration) {
693 bool UseDeferred = SetIteration > 0;
694 PtrAccessSet &S = UseDeferred ? DeferredAccesses : Accesses;
697 Value *Ptr = AV.getValue();
699 // For a single memory access in AliasSetTracker, Accesses may contain
700 // both read and write, and they both need to be handled for CheckDeps.
702 if (AC.getPointer() != Ptr)
705 bool IsWrite = AC.getInt();
707 // If we're using the deferred access set, then it contains only
709 bool IsReadOnlyPtr = ReadOnlyPtr.count(Ptr) && !IsWrite;
710 if (UseDeferred && !IsReadOnlyPtr)
712 // Otherwise, the pointer must be in the PtrAccessSet, either as a
714 assert(((IsReadOnlyPtr && UseDeferred) || IsWrite ||
715 S.count(MemAccessInfo(Ptr, false))) &&
716 "Alias-set pointer not in the access set?");
718 MemAccessInfo Access(Ptr, IsWrite);
719 DepCands.insert(Access);
721 // Memorize read-only pointers for later processing and skip them in
722 // the first round (they need to be checked after we have seen all
723 // write pointers). Note: we also mark pointer that are not
724 // consecutive as "read-only" pointers (so that we check
725 // "a[b[i]] +="). Hence, we need the second check for "!IsWrite".
726 if (!UseDeferred && IsReadOnlyPtr) {
727 DeferredAccesses.insert(Access);
731 // If this is a write - check other reads and writes for conflicts. If
732 // this is a read only check other writes for conflicts (but only if
733 // there is no other write to the ptr - this is an optimization to
734 // catch "a[i] = a[i] + " without having to do a dependence check).
735 if ((IsWrite || IsReadOnlyPtr) && SetHasWrite) {
736 CheckDeps.insert(Access);
737 IsRTCheckAnalysisNeeded = true;
743 // Create sets of pointers connected by a shared alias set and
744 // underlying object.
745 typedef SmallVector<Value *, 16> ValueVector;
746 ValueVector TempObjects;
748 GetUnderlyingObjects(Ptr, TempObjects, DL, LI);
749 DEBUG(dbgs() << "Underlying objects for pointer " << *Ptr << "\n");
750 for (Value *UnderlyingObj : TempObjects) {
751 UnderlyingObjToAccessMap::iterator Prev =
752 ObjToLastAccess.find(UnderlyingObj);
753 if (Prev != ObjToLastAccess.end())
754 DepCands.unionSets(Access, Prev->second);
756 ObjToLastAccess[UnderlyingObj] = Access;
757 DEBUG(dbgs() << " " << *UnderlyingObj << "\n");
765 static bool isInBoundsGep(Value *Ptr) {
766 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr))
767 return GEP->isInBounds();
771 /// \brief Return true if an AddRec pointer \p Ptr is unsigned non-wrapping,
772 /// i.e. monotonically increasing/decreasing.
773 static bool isNoWrapAddRec(Value *Ptr, const SCEVAddRecExpr *AR,
774 ScalarEvolution *SE, const Loop *L) {
775 // FIXME: This should probably only return true for NUW.
776 if (AR->getNoWrapFlags(SCEV::NoWrapMask))
779 // Scalar evolution does not propagate the non-wrapping flags to values that
780 // are derived from a non-wrapping induction variable because non-wrapping
781 // could be flow-sensitive.
783 // Look through the potentially overflowing instruction to try to prove
784 // non-wrapping for the *specific* value of Ptr.
786 // The arithmetic implied by an inbounds GEP can't overflow.
787 auto *GEP = dyn_cast<GetElementPtrInst>(Ptr);
788 if (!GEP || !GEP->isInBounds())
791 // Make sure there is only one non-const index and analyze that.
792 Value *NonConstIndex = nullptr;
793 for (auto Index = GEP->idx_begin(); Index != GEP->idx_end(); ++Index)
794 if (!isa<ConstantInt>(*Index)) {
797 NonConstIndex = *Index;
800 // The recurrence is on the pointer, ignore for now.
803 // The index in GEP is signed. It is non-wrapping if it's derived from a NSW
804 // AddRec using a NSW operation.
805 if (auto *OBO = dyn_cast<OverflowingBinaryOperator>(NonConstIndex))
806 if (OBO->hasNoSignedWrap() &&
807 // Assume constant for other the operand so that the AddRec can be
809 isa<ConstantInt>(OBO->getOperand(1))) {
810 auto *OpScev = SE->getSCEV(OBO->getOperand(0));
812 if (auto *OpAR = dyn_cast<SCEVAddRecExpr>(OpScev))
813 return OpAR->getLoop() == L && OpAR->getNoWrapFlags(SCEV::FlagNSW);
819 /// \brief Check whether the access through \p Ptr has a constant stride.
820 int llvm::isStridedPtr(ScalarEvolution *SE, Value *Ptr, const Loop *Lp,
821 const ValueToValueMap &StridesMap) {
822 Type *Ty = Ptr->getType();
823 assert(Ty->isPointerTy() && "Unexpected non-ptr");
825 // Make sure that the pointer does not point to aggregate types.
826 auto *PtrTy = cast<PointerType>(Ty);
827 if (PtrTy->getElementType()->isAggregateType()) {
828 DEBUG(dbgs() << "LAA: Bad stride - Not a pointer to a scalar type"
833 const SCEV *PtrScev = replaceSymbolicStrideSCEV(SE, StridesMap, Ptr);
835 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
837 DEBUG(dbgs() << "LAA: Bad stride - Not an AddRecExpr pointer "
838 << *Ptr << " SCEV: " << *PtrScev << "\n");
842 // The accesss function must stride over the innermost loop.
843 if (Lp != AR->getLoop()) {
844 DEBUG(dbgs() << "LAA: Bad stride - Not striding over innermost loop " <<
845 *Ptr << " SCEV: " << *PtrScev << "\n");
848 // The address calculation must not wrap. Otherwise, a dependence could be
850 // An inbounds getelementptr that is a AddRec with a unit stride
851 // cannot wrap per definition. The unit stride requirement is checked later.
852 // An getelementptr without an inbounds attribute and unit stride would have
853 // to access the pointer value "0" which is undefined behavior in address
854 // space 0, therefore we can also vectorize this case.
855 bool IsInBoundsGEP = isInBoundsGep(Ptr);
856 bool IsNoWrapAddRec = isNoWrapAddRec(Ptr, AR, SE, Lp);
857 bool IsInAddressSpaceZero = PtrTy->getAddressSpace() == 0;
858 if (!IsNoWrapAddRec && !IsInBoundsGEP && !IsInAddressSpaceZero) {
859 DEBUG(dbgs() << "LAA: Bad stride - Pointer may wrap in the address space "
860 << *Ptr << " SCEV: " << *PtrScev << "\n");
864 // Check the step is constant.
865 const SCEV *Step = AR->getStepRecurrence(*SE);
867 // Calculate the pointer stride and check if it is constant.
868 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
870 DEBUG(dbgs() << "LAA: Bad stride - Not a constant strided " << *Ptr <<
871 " SCEV: " << *PtrScev << "\n");
875 auto &DL = Lp->getHeader()->getModule()->getDataLayout();
876 int64_t Size = DL.getTypeAllocSize(PtrTy->getElementType());
877 const APInt &APStepVal = C->getValue()->getValue();
879 // Huge step value - give up.
880 if (APStepVal.getBitWidth() > 64)
883 int64_t StepVal = APStepVal.getSExtValue();
886 int64_t Stride = StepVal / Size;
887 int64_t Rem = StepVal % Size;
891 // If the SCEV could wrap but we have an inbounds gep with a unit stride we
892 // know we can't "wrap around the address space". In case of address space
893 // zero we know that this won't happen without triggering undefined behavior.
894 if (!IsNoWrapAddRec && (IsInBoundsGEP || IsInAddressSpaceZero) &&
895 Stride != 1 && Stride != -1)
901 bool MemoryDepChecker::Dependence::isSafeForVectorization(DepType Type) {
905 case BackwardVectorizable:
909 case ForwardButPreventsForwarding:
911 case BackwardVectorizableButPreventsForwarding:
914 llvm_unreachable("unexpected DepType!");
917 bool MemoryDepChecker::Dependence::isInterestingDependence(DepType Type) {
923 case BackwardVectorizable:
925 case ForwardButPreventsForwarding:
927 case BackwardVectorizableButPreventsForwarding:
930 llvm_unreachable("unexpected DepType!");
933 bool MemoryDepChecker::Dependence::isPossiblyBackward() const {
937 case ForwardButPreventsForwarding:
941 case BackwardVectorizable:
943 case BackwardVectorizableButPreventsForwarding:
946 llvm_unreachable("unexpected DepType!");
949 bool MemoryDepChecker::couldPreventStoreLoadForward(unsigned Distance,
950 unsigned TypeByteSize) {
951 // If loads occur at a distance that is not a multiple of a feasible vector
952 // factor store-load forwarding does not take place.
953 // Positive dependences might cause troubles because vectorizing them might
954 // prevent store-load forwarding making vectorized code run a lot slower.
955 // a[i] = a[i-3] ^ a[i-8];
956 // The stores to a[i:i+1] don't align with the stores to a[i-3:i-2] and
957 // hence on your typical architecture store-load forwarding does not take
958 // place. Vectorizing in such cases does not make sense.
959 // Store-load forwarding distance.
960 const unsigned NumCyclesForStoreLoadThroughMemory = 8*TypeByteSize;
961 // Maximum vector factor.
962 unsigned MaxVFWithoutSLForwardIssues =
963 VectorizerParams::MaxVectorWidth * TypeByteSize;
964 if(MaxSafeDepDistBytes < MaxVFWithoutSLForwardIssues)
965 MaxVFWithoutSLForwardIssues = MaxSafeDepDistBytes;
967 for (unsigned vf = 2*TypeByteSize; vf <= MaxVFWithoutSLForwardIssues;
969 if (Distance % vf && Distance / vf < NumCyclesForStoreLoadThroughMemory) {
970 MaxVFWithoutSLForwardIssues = (vf >>=1);
975 if (MaxVFWithoutSLForwardIssues< 2*TypeByteSize) {
976 DEBUG(dbgs() << "LAA: Distance " << Distance <<
977 " that could cause a store-load forwarding conflict\n");
981 if (MaxVFWithoutSLForwardIssues < MaxSafeDepDistBytes &&
982 MaxVFWithoutSLForwardIssues !=
983 VectorizerParams::MaxVectorWidth * TypeByteSize)
984 MaxSafeDepDistBytes = MaxVFWithoutSLForwardIssues;
988 /// \brief Check the dependence for two accesses with the same stride \p Stride.
989 /// \p Distance is the positive distance and \p TypeByteSize is type size in
992 /// \returns true if they are independent.
993 static bool areStridedAccessesIndependent(unsigned Distance, unsigned Stride,
994 unsigned TypeByteSize) {
995 assert(Stride > 1 && "The stride must be greater than 1");
996 assert(TypeByteSize > 0 && "The type size in byte must be non-zero");
997 assert(Distance > 0 && "The distance must be non-zero");
999 // Skip if the distance is not multiple of type byte size.
1000 if (Distance % TypeByteSize)
1003 unsigned ScaledDist = Distance / TypeByteSize;
1005 // No dependence if the scaled distance is not multiple of the stride.
1007 // for (i = 0; i < 1024 ; i += 4)
1008 // A[i+2] = A[i] + 1;
1010 // Two accesses in memory (scaled distance is 2, stride is 4):
1011 // | A[0] | | | | A[4] | | | |
1012 // | | | A[2] | | | | A[6] | |
1015 // for (i = 0; i < 1024 ; i += 3)
1016 // A[i+4] = A[i] + 1;
1018 // Two accesses in memory (scaled distance is 4, stride is 3):
1019 // | A[0] | | | A[3] | | | A[6] | | |
1020 // | | | | | A[4] | | | A[7] | |
1021 return ScaledDist % Stride;
1024 MemoryDepChecker::Dependence::DepType
1025 MemoryDepChecker::isDependent(const MemAccessInfo &A, unsigned AIdx,
1026 const MemAccessInfo &B, unsigned BIdx,
1027 const ValueToValueMap &Strides) {
1028 assert (AIdx < BIdx && "Must pass arguments in program order");
1030 Value *APtr = A.getPointer();
1031 Value *BPtr = B.getPointer();
1032 bool AIsWrite = A.getInt();
1033 bool BIsWrite = B.getInt();
1035 // Two reads are independent.
1036 if (!AIsWrite && !BIsWrite)
1037 return Dependence::NoDep;
1039 // We cannot check pointers in different address spaces.
1040 if (APtr->getType()->getPointerAddressSpace() !=
1041 BPtr->getType()->getPointerAddressSpace())
1042 return Dependence::Unknown;
1044 const SCEV *AScev = replaceSymbolicStrideSCEV(SE, Strides, APtr);
1045 const SCEV *BScev = replaceSymbolicStrideSCEV(SE, Strides, BPtr);
1047 int StrideAPtr = isStridedPtr(SE, APtr, InnermostLoop, Strides);
1048 int StrideBPtr = isStridedPtr(SE, BPtr, InnermostLoop, Strides);
1050 const SCEV *Src = AScev;
1051 const SCEV *Sink = BScev;
1053 // If the induction step is negative we have to invert source and sink of the
1055 if (StrideAPtr < 0) {
1058 std::swap(APtr, BPtr);
1059 std::swap(Src, Sink);
1060 std::swap(AIsWrite, BIsWrite);
1061 std::swap(AIdx, BIdx);
1062 std::swap(StrideAPtr, StrideBPtr);
1065 const SCEV *Dist = SE->getMinusSCEV(Sink, Src);
1067 DEBUG(dbgs() << "LAA: Src Scev: " << *Src << "Sink Scev: " << *Sink
1068 << "(Induction step: " << StrideAPtr << ")\n");
1069 DEBUG(dbgs() << "LAA: Distance for " << *InstMap[AIdx] << " to "
1070 << *InstMap[BIdx] << ": " << *Dist << "\n");
1072 // Need accesses with constant stride. We don't want to vectorize
1073 // "A[B[i]] += ..." and similar code or pointer arithmetic that could wrap in
1074 // the address space.
1075 if (!StrideAPtr || !StrideBPtr || StrideAPtr != StrideBPtr){
1076 DEBUG(dbgs() << "Pointer access with non-constant stride\n");
1077 return Dependence::Unknown;
1080 const SCEVConstant *C = dyn_cast<SCEVConstant>(Dist);
1082 DEBUG(dbgs() << "LAA: Dependence because of non-constant distance\n");
1083 ShouldRetryWithRuntimeCheck = true;
1084 return Dependence::Unknown;
1087 Type *ATy = APtr->getType()->getPointerElementType();
1088 Type *BTy = BPtr->getType()->getPointerElementType();
1089 auto &DL = InnermostLoop->getHeader()->getModule()->getDataLayout();
1090 unsigned TypeByteSize = DL.getTypeAllocSize(ATy);
1092 // Negative distances are not plausible dependencies.
1093 const APInt &Val = C->getValue()->getValue();
1094 if (Val.isNegative()) {
1095 bool IsTrueDataDependence = (AIsWrite && !BIsWrite);
1096 if (IsTrueDataDependence &&
1097 (couldPreventStoreLoadForward(Val.abs().getZExtValue(), TypeByteSize) ||
1099 return Dependence::ForwardButPreventsForwarding;
1101 DEBUG(dbgs() << "LAA: Dependence is negative: NoDep\n");
1102 return Dependence::Forward;
1105 // Write to the same location with the same size.
1106 // Could be improved to assert type sizes are the same (i32 == float, etc).
1109 return Dependence::NoDep;
1110 DEBUG(dbgs() << "LAA: Zero dependence difference but different types\n");
1111 return Dependence::Unknown;
1114 assert(Val.isStrictlyPositive() && "Expect a positive value");
1118 "LAA: ReadWrite-Write positive dependency with different types\n");
1119 return Dependence::Unknown;
1122 unsigned Distance = (unsigned) Val.getZExtValue();
1124 unsigned Stride = std::abs(StrideAPtr);
1126 areStridedAccessesIndependent(Distance, Stride, TypeByteSize)) {
1127 DEBUG(dbgs() << "LAA: Strided accesses are independent\n");
1128 return Dependence::NoDep;
1131 // Bail out early if passed-in parameters make vectorization not feasible.
1132 unsigned ForcedFactor = (VectorizerParams::VectorizationFactor ?
1133 VectorizerParams::VectorizationFactor : 1);
1134 unsigned ForcedUnroll = (VectorizerParams::VectorizationInterleave ?
1135 VectorizerParams::VectorizationInterleave : 1);
1136 // The minimum number of iterations for a vectorized/unrolled version.
1137 unsigned MinNumIter = std::max(ForcedFactor * ForcedUnroll, 2U);
1139 // It's not vectorizable if the distance is smaller than the minimum distance
1140 // needed for a vectroized/unrolled version. Vectorizing one iteration in
1141 // front needs TypeByteSize * Stride. Vectorizing the last iteration needs
1142 // TypeByteSize (No need to plus the last gap distance).
1144 // E.g. Assume one char is 1 byte in memory and one int is 4 bytes.
1146 // int *B = (int *)((char *)A + 14);
1147 // for (i = 0 ; i < 1024 ; i += 2)
1151 // Two accesses in memory (stride is 2):
1152 // | A[0] | | A[2] | | A[4] | | A[6] | |
1153 // | B[0] | | B[2] | | B[4] |
1155 // Distance needs for vectorizing iterations except the last iteration:
1156 // 4 * 2 * (MinNumIter - 1). Distance needs for the last iteration: 4.
1157 // So the minimum distance needed is: 4 * 2 * (MinNumIter - 1) + 4.
1159 // If MinNumIter is 2, it is vectorizable as the minimum distance needed is
1160 // 12, which is less than distance.
1162 // If MinNumIter is 4 (Say if a user forces the vectorization factor to be 4),
1163 // the minimum distance needed is 28, which is greater than distance. It is
1164 // not safe to do vectorization.
1165 unsigned MinDistanceNeeded =
1166 TypeByteSize * Stride * (MinNumIter - 1) + TypeByteSize;
1167 if (MinDistanceNeeded > Distance) {
1168 DEBUG(dbgs() << "LAA: Failure because of positive distance " << Distance
1170 return Dependence::Backward;
1173 // Unsafe if the minimum distance needed is greater than max safe distance.
1174 if (MinDistanceNeeded > MaxSafeDepDistBytes) {
1175 DEBUG(dbgs() << "LAA: Failure because it needs at least "
1176 << MinDistanceNeeded << " size in bytes");
1177 return Dependence::Backward;
1180 // Positive distance bigger than max vectorization factor.
1181 // FIXME: Should use max factor instead of max distance in bytes, which could
1182 // not handle different types.
1183 // E.g. Assume one char is 1 byte in memory and one int is 4 bytes.
1184 // void foo (int *A, char *B) {
1185 // for (unsigned i = 0; i < 1024; i++) {
1186 // A[i+2] = A[i] + 1;
1187 // B[i+2] = B[i] + 1;
1191 // This case is currently unsafe according to the max safe distance. If we
1192 // analyze the two accesses on array B, the max safe dependence distance
1193 // is 2. Then we analyze the accesses on array A, the minimum distance needed
1194 // is 8, which is less than 2 and forbidden vectorization, But actually
1195 // both A and B could be vectorized by 2 iterations.
1196 MaxSafeDepDistBytes =
1197 Distance < MaxSafeDepDistBytes ? Distance : MaxSafeDepDistBytes;
1199 bool IsTrueDataDependence = (!AIsWrite && BIsWrite);
1200 if (IsTrueDataDependence &&
1201 couldPreventStoreLoadForward(Distance, TypeByteSize))
1202 return Dependence::BackwardVectorizableButPreventsForwarding;
1204 DEBUG(dbgs() << "LAA: Positive distance " << Val.getSExtValue()
1205 << " with max VF = "
1206 << MaxSafeDepDistBytes / (TypeByteSize * Stride) << '\n');
1208 return Dependence::BackwardVectorizable;
1211 bool MemoryDepChecker::areDepsSafe(DepCandidates &AccessSets,
1212 MemAccessInfoSet &CheckDeps,
1213 const ValueToValueMap &Strides) {
1215 MaxSafeDepDistBytes = -1U;
1216 while (!CheckDeps.empty()) {
1217 MemAccessInfo CurAccess = *CheckDeps.begin();
1219 // Get the relevant memory access set.
1220 EquivalenceClasses<MemAccessInfo>::iterator I =
1221 AccessSets.findValue(AccessSets.getLeaderValue(CurAccess));
1223 // Check accesses within this set.
1224 EquivalenceClasses<MemAccessInfo>::member_iterator AI, AE;
1225 AI = AccessSets.member_begin(I), AE = AccessSets.member_end();
1227 // Check every access pair.
1229 CheckDeps.erase(*AI);
1230 EquivalenceClasses<MemAccessInfo>::member_iterator OI = std::next(AI);
1232 // Check every accessing instruction pair in program order.
1233 for (std::vector<unsigned>::iterator I1 = Accesses[*AI].begin(),
1234 I1E = Accesses[*AI].end(); I1 != I1E; ++I1)
1235 for (std::vector<unsigned>::iterator I2 = Accesses[*OI].begin(),
1236 I2E = Accesses[*OI].end(); I2 != I2E; ++I2) {
1237 auto A = std::make_pair(&*AI, *I1);
1238 auto B = std::make_pair(&*OI, *I2);
1244 Dependence::DepType Type =
1245 isDependent(*A.first, A.second, *B.first, B.second, Strides);
1246 SafeForVectorization &= Dependence::isSafeForVectorization(Type);
1248 // Gather dependences unless we accumulated MaxInterestingDependence
1249 // dependences. In that case return as soon as we find the first
1250 // unsafe dependence. This puts a limit on this quadratic
1252 if (RecordInterestingDependences) {
1253 if (Dependence::isInterestingDependence(Type))
1254 InterestingDependences.push_back(
1255 Dependence(A.second, B.second, Type));
1257 if (InterestingDependences.size() >= MaxInterestingDependence) {
1258 RecordInterestingDependences = false;
1259 InterestingDependences.clear();
1260 DEBUG(dbgs() << "Too many dependences, stopped recording\n");
1263 if (!RecordInterestingDependences && !SafeForVectorization)
1272 DEBUG(dbgs() << "Total Interesting Dependences: "
1273 << InterestingDependences.size() << "\n");
1274 return SafeForVectorization;
1277 SmallVector<Instruction *, 4>
1278 MemoryDepChecker::getInstructionsForAccess(Value *Ptr, bool isWrite) const {
1279 MemAccessInfo Access(Ptr, isWrite);
1280 auto &IndexVector = Accesses.find(Access)->second;
1282 SmallVector<Instruction *, 4> Insts;
1283 std::transform(IndexVector.begin(), IndexVector.end(),
1284 std::back_inserter(Insts),
1285 [&](unsigned Idx) { return this->InstMap[Idx]; });
1289 const char *MemoryDepChecker::Dependence::DepName[] = {
1290 "NoDep", "Unknown", "Forward", "ForwardButPreventsForwarding", "Backward",
1291 "BackwardVectorizable", "BackwardVectorizableButPreventsForwarding"};
1293 void MemoryDepChecker::Dependence::print(
1294 raw_ostream &OS, unsigned Depth,
1295 const SmallVectorImpl<Instruction *> &Instrs) const {
1296 OS.indent(Depth) << DepName[Type] << ":\n";
1297 OS.indent(Depth + 2) << *Instrs[Source] << " -> \n";
1298 OS.indent(Depth + 2) << *Instrs[Destination] << "\n";
1301 bool LoopAccessInfo::canAnalyzeLoop() {
1302 // We need to have a loop header.
1303 DEBUG(dbgs() << "LAA: Found a loop: " <<
1304 TheLoop->getHeader()->getName() << '\n');
1306 // We can only analyze innermost loops.
1307 if (!TheLoop->empty()) {
1308 DEBUG(dbgs() << "LAA: loop is not the innermost loop\n");
1309 emitAnalysis(LoopAccessReport() << "loop is not the innermost loop");
1313 // We must have a single backedge.
1314 if (TheLoop->getNumBackEdges() != 1) {
1315 DEBUG(dbgs() << "LAA: loop control flow is not understood by analyzer\n");
1317 LoopAccessReport() <<
1318 "loop control flow is not understood by analyzer");
1322 // We must have a single exiting block.
1323 if (!TheLoop->getExitingBlock()) {
1324 DEBUG(dbgs() << "LAA: loop control flow is not understood by analyzer\n");
1326 LoopAccessReport() <<
1327 "loop control flow is not understood by analyzer");
1331 // We only handle bottom-tested loops, i.e. loop in which the condition is
1332 // checked at the end of each iteration. With that we can assume that all
1333 // instructions in the loop are executed the same number of times.
1334 if (TheLoop->getExitingBlock() != TheLoop->getLoopLatch()) {
1335 DEBUG(dbgs() << "LAA: loop control flow is not understood by analyzer\n");
1337 LoopAccessReport() <<
1338 "loop control flow is not understood by analyzer");
1342 // ScalarEvolution needs to be able to find the exit count.
1343 const SCEV *ExitCount = SE->getBackedgeTakenCount(TheLoop);
1344 if (ExitCount == SE->getCouldNotCompute()) {
1345 emitAnalysis(LoopAccessReport() <<
1346 "could not determine number of loop iterations");
1347 DEBUG(dbgs() << "LAA: SCEV could not compute the loop exit count.\n");
1354 void LoopAccessInfo::analyzeLoop(const ValueToValueMap &Strides) {
1356 typedef SmallVector<Value*, 16> ValueVector;
1357 typedef SmallPtrSet<Value*, 16> ValueSet;
1359 // Holds the Load and Store *instructions*.
1363 // Holds all the different accesses in the loop.
1364 unsigned NumReads = 0;
1365 unsigned NumReadWrites = 0;
1367 PtrRtChecking.Pointers.clear();
1368 PtrRtChecking.Need = false;
1370 const bool IsAnnotatedParallel = TheLoop->isAnnotatedParallel();
1373 for (Loop::block_iterator bb = TheLoop->block_begin(),
1374 be = TheLoop->block_end(); bb != be; ++bb) {
1376 // Scan the BB and collect legal loads and stores.
1377 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
1380 // If this is a load, save it. If this instruction can read from memory
1381 // but is not a load, then we quit. Notice that we don't handle function
1382 // calls that read or write.
1383 if (it->mayReadFromMemory()) {
1384 // Many math library functions read the rounding mode. We will only
1385 // vectorize a loop if it contains known function calls that don't set
1386 // the flag. Therefore, it is safe to ignore this read from memory.
1387 CallInst *Call = dyn_cast<CallInst>(it);
1388 if (Call && getIntrinsicIDForCall(Call, TLI))
1391 // If the function has an explicit vectorized counterpart, we can safely
1392 // assume that it can be vectorized.
1393 if (Call && !Call->isNoBuiltin() && Call->getCalledFunction() &&
1394 TLI->isFunctionVectorizable(Call->getCalledFunction()->getName()))
1397 LoadInst *Ld = dyn_cast<LoadInst>(it);
1398 if (!Ld || (!Ld->isSimple() && !IsAnnotatedParallel)) {
1399 emitAnalysis(LoopAccessReport(Ld)
1400 << "read with atomic ordering or volatile read");
1401 DEBUG(dbgs() << "LAA: Found a non-simple load.\n");
1406 Loads.push_back(Ld);
1407 DepChecker.addAccess(Ld);
1411 // Save 'store' instructions. Abort if other instructions write to memory.
1412 if (it->mayWriteToMemory()) {
1413 StoreInst *St = dyn_cast<StoreInst>(it);
1415 emitAnalysis(LoopAccessReport(it) <<
1416 "instruction cannot be vectorized");
1420 if (!St->isSimple() && !IsAnnotatedParallel) {
1421 emitAnalysis(LoopAccessReport(St)
1422 << "write with atomic ordering or volatile write");
1423 DEBUG(dbgs() << "LAA: Found a non-simple store.\n");
1428 Stores.push_back(St);
1429 DepChecker.addAccess(St);
1434 // Now we have two lists that hold the loads and the stores.
1435 // Next, we find the pointers that they use.
1437 // Check if we see any stores. If there are no stores, then we don't
1438 // care if the pointers are *restrict*.
1439 if (!Stores.size()) {
1440 DEBUG(dbgs() << "LAA: Found a read-only loop!\n");
1445 MemoryDepChecker::DepCandidates DependentAccesses;
1446 AccessAnalysis Accesses(TheLoop->getHeader()->getModule()->getDataLayout(),
1447 AA, LI, DependentAccesses);
1449 // Holds the analyzed pointers. We don't want to call GetUnderlyingObjects
1450 // multiple times on the same object. If the ptr is accessed twice, once
1451 // for read and once for write, it will only appear once (on the write
1452 // list). This is okay, since we are going to check for conflicts between
1453 // writes and between reads and writes, but not between reads and reads.
1456 ValueVector::iterator I, IE;
1457 for (I = Stores.begin(), IE = Stores.end(); I != IE; ++I) {
1458 StoreInst *ST = cast<StoreInst>(*I);
1459 Value* Ptr = ST->getPointerOperand();
1460 // Check for store to loop invariant address.
1461 StoreToLoopInvariantAddress |= isUniform(Ptr);
1462 // If we did *not* see this pointer before, insert it to the read-write
1463 // list. At this phase it is only a 'write' list.
1464 if (Seen.insert(Ptr).second) {
1467 MemoryLocation Loc = MemoryLocation::get(ST);
1468 // The TBAA metadata could have a control dependency on the predication
1469 // condition, so we cannot rely on it when determining whether or not we
1470 // need runtime pointer checks.
1471 if (blockNeedsPredication(ST->getParent(), TheLoop, DT))
1472 Loc.AATags.TBAA = nullptr;
1474 Accesses.addStore(Loc);
1478 if (IsAnnotatedParallel) {
1480 << "LAA: A loop annotated parallel, ignore memory dependency "
1486 for (I = Loads.begin(), IE = Loads.end(); I != IE; ++I) {
1487 LoadInst *LD = cast<LoadInst>(*I);
1488 Value* Ptr = LD->getPointerOperand();
1489 // If we did *not* see this pointer before, insert it to the
1490 // read list. If we *did* see it before, then it is already in
1491 // the read-write list. This allows us to vectorize expressions
1492 // such as A[i] += x; Because the address of A[i] is a read-write
1493 // pointer. This only works if the index of A[i] is consecutive.
1494 // If the address of i is unknown (for example A[B[i]]) then we may
1495 // read a few words, modify, and write a few words, and some of the
1496 // words may be written to the same address.
1497 bool IsReadOnlyPtr = false;
1498 if (Seen.insert(Ptr).second || !isStridedPtr(SE, Ptr, TheLoop, Strides)) {
1500 IsReadOnlyPtr = true;
1503 MemoryLocation Loc = MemoryLocation::get(LD);
1504 // The TBAA metadata could have a control dependency on the predication
1505 // condition, so we cannot rely on it when determining whether or not we
1506 // need runtime pointer checks.
1507 if (blockNeedsPredication(LD->getParent(), TheLoop, DT))
1508 Loc.AATags.TBAA = nullptr;
1510 Accesses.addLoad(Loc, IsReadOnlyPtr);
1513 // If we write (or read-write) to a single destination and there are no
1514 // other reads in this loop then is it safe to vectorize.
1515 if (NumReadWrites == 1 && NumReads == 0) {
1516 DEBUG(dbgs() << "LAA: Found a write-only loop!\n");
1521 // Build dependence sets and check whether we need a runtime pointer bounds
1523 Accesses.buildDependenceSets();
1525 // Find pointers with computable bounds. We are going to use this information
1526 // to place a runtime bound check.
1527 bool CanDoRTIfNeeded =
1528 Accesses.canCheckPtrAtRT(PtrRtChecking, SE, TheLoop, Strides);
1529 if (!CanDoRTIfNeeded) {
1530 emitAnalysis(LoopAccessReport() << "cannot identify array bounds");
1531 DEBUG(dbgs() << "LAA: We can't vectorize because we can't find "
1532 << "the array bounds.\n");
1537 DEBUG(dbgs() << "LAA: We can perform a memory runtime check if needed.\n");
1540 if (Accesses.isDependencyCheckNeeded()) {
1541 DEBUG(dbgs() << "LAA: Checking memory dependencies\n");
1542 CanVecMem = DepChecker.areDepsSafe(
1543 DependentAccesses, Accesses.getDependenciesToCheck(), Strides);
1544 MaxSafeDepDistBytes = DepChecker.getMaxSafeDepDistBytes();
1546 if (!CanVecMem && DepChecker.shouldRetryWithRuntimeCheck()) {
1547 DEBUG(dbgs() << "LAA: Retrying with memory checks\n");
1549 // Clear the dependency checks. We assume they are not needed.
1550 Accesses.resetDepChecks(DepChecker);
1552 PtrRtChecking.reset();
1553 PtrRtChecking.Need = true;
1556 Accesses.canCheckPtrAtRT(PtrRtChecking, SE, TheLoop, Strides, true);
1558 // Check that we found the bounds for the pointer.
1559 if (!CanDoRTIfNeeded) {
1560 emitAnalysis(LoopAccessReport()
1561 << "cannot check memory dependencies at runtime");
1562 DEBUG(dbgs() << "LAA: Can't vectorize with memory checks\n");
1572 DEBUG(dbgs() << "LAA: No unsafe dependent memory operations in loop. We"
1573 << (PtrRtChecking.Need ? "" : " don't")
1574 << " need runtime memory checks.\n");
1576 emitAnalysis(LoopAccessReport() <<
1577 "unsafe dependent memory operations in loop");
1578 DEBUG(dbgs() << "LAA: unsafe dependent memory operations in loop\n");
1582 bool LoopAccessInfo::blockNeedsPredication(BasicBlock *BB, Loop *TheLoop,
1583 DominatorTree *DT) {
1584 assert(TheLoop->contains(BB) && "Unknown block used");
1586 // Blocks that do not dominate the latch need predication.
1587 BasicBlock* Latch = TheLoop->getLoopLatch();
1588 return !DT->dominates(BB, Latch);
1591 void LoopAccessInfo::emitAnalysis(LoopAccessReport &Message) {
1592 assert(!Report && "Multiple reports generated");
1596 bool LoopAccessInfo::isUniform(Value *V) const {
1597 return (SE->isLoopInvariant(SE->getSCEV(V), TheLoop));
1600 // FIXME: this function is currently a duplicate of the one in
1601 // LoopVectorize.cpp.
1602 static Instruction *getFirstInst(Instruction *FirstInst, Value *V,
1606 if (Instruction *I = dyn_cast<Instruction>(V))
1607 return I->getParent() == Loc->getParent() ? I : nullptr;
1611 /// \brief IR Values for the lower and upper bounds of a pointer evolution.
1612 struct PointerBounds {
1617 /// \brief Expand code for the lower and upper bound of the pointer group \p CG
1618 /// in \p TheLoop. \return the values for the bounds.
1619 static PointerBounds
1620 expandBounds(const RuntimePointerChecking::CheckingPtrGroup *CG, Loop *TheLoop,
1621 Instruction *Loc, SCEVExpander &Exp, ScalarEvolution *SE,
1622 const RuntimePointerChecking &PtrRtChecking) {
1623 Value *Ptr = PtrRtChecking.Pointers[CG->Members[0]].PointerValue;
1624 const SCEV *Sc = SE->getSCEV(Ptr);
1626 if (SE->isLoopInvariant(Sc, TheLoop)) {
1627 DEBUG(dbgs() << "LAA: Adding RT check for a loop invariant ptr:" << *Ptr
1631 unsigned AS = Ptr->getType()->getPointerAddressSpace();
1632 LLVMContext &Ctx = Loc->getContext();
1634 // Use this type for pointer arithmetic.
1635 Type *PtrArithTy = Type::getInt8PtrTy(Ctx, AS);
1636 Value *Start = nullptr, *End = nullptr;
1638 DEBUG(dbgs() << "LAA: Adding RT check for range:\n");
1639 Start = Exp.expandCodeFor(CG->Low, PtrArithTy, Loc);
1640 End = Exp.expandCodeFor(CG->High, PtrArithTy, Loc);
1641 DEBUG(dbgs() << "Start: " << *CG->Low << " End: " << *CG->High << "\n");
1642 return {Start, End};
1646 /// \brief Turns a collection of checks into a collection of expanded upper and
1647 /// lower bounds for both pointers in the check.
1648 static SmallVector<std::pair<PointerBounds, PointerBounds>, 4> expandBounds(
1649 const SmallVectorImpl<RuntimePointerChecking::PointerCheck> &PointerChecks,
1650 Loop *L, Instruction *Loc, ScalarEvolution *SE, SCEVExpander &Exp,
1651 const RuntimePointerChecking &PtrRtChecking) {
1652 SmallVector<std::pair<PointerBounds, PointerBounds>, 4> ChecksWithBounds;
1654 // Here we're relying on the SCEV Expander's cache to only emit code for the
1655 // same bounds once.
1657 PointerChecks.begin(), PointerChecks.end(),
1658 std::back_inserter(ChecksWithBounds),
1659 [&](const RuntimePointerChecking::PointerCheck &Check) {
1661 First = expandBounds(Check.first, L, Loc, Exp, SE, PtrRtChecking),
1662 Second = expandBounds(Check.second, L, Loc, Exp, SE, PtrRtChecking);
1663 return std::make_pair(First, Second);
1666 return ChecksWithBounds;
1669 std::pair<Instruction *, Instruction *> LoopAccessInfo::addRuntimeCheck(
1671 const SmallVectorImpl<RuntimePointerChecking::PointerCheck> &PointerChecks)
1674 SCEVExpander Exp(*SE, DL, "induction");
1675 auto ExpandedChecks =
1676 expandBounds(PointerChecks, TheLoop, Loc, SE, Exp, PtrRtChecking);
1678 LLVMContext &Ctx = Loc->getContext();
1679 Instruction *FirstInst = nullptr;
1680 IRBuilder<> ChkBuilder(Loc);
1681 // Our instructions might fold to a constant.
1682 Value *MemoryRuntimeCheck = nullptr;
1684 for (const auto &Check : ExpandedChecks) {
1685 const PointerBounds &A = Check.first, &B = Check.second;
1686 unsigned AS0 = A.Start->getType()->getPointerAddressSpace();
1687 unsigned AS1 = B.Start->getType()->getPointerAddressSpace();
1689 assert((AS0 == B.End->getType()->getPointerAddressSpace()) &&
1690 (AS1 == A.End->getType()->getPointerAddressSpace()) &&
1691 "Trying to bounds check pointers with different address spaces");
1693 Type *PtrArithTy0 = Type::getInt8PtrTy(Ctx, AS0);
1694 Type *PtrArithTy1 = Type::getInt8PtrTy(Ctx, AS1);
1696 Value *Start0 = ChkBuilder.CreateBitCast(A.Start, PtrArithTy0, "bc");
1697 Value *Start1 = ChkBuilder.CreateBitCast(B.Start, PtrArithTy1, "bc");
1698 Value *End0 = ChkBuilder.CreateBitCast(A.End, PtrArithTy1, "bc");
1699 Value *End1 = ChkBuilder.CreateBitCast(B.End, PtrArithTy0, "bc");
1701 Value *Cmp0 = ChkBuilder.CreateICmpULE(Start0, End1, "bound0");
1702 FirstInst = getFirstInst(FirstInst, Cmp0, Loc);
1703 Value *Cmp1 = ChkBuilder.CreateICmpULE(Start1, End0, "bound1");
1704 FirstInst = getFirstInst(FirstInst, Cmp1, Loc);
1705 Value *IsConflict = ChkBuilder.CreateAnd(Cmp0, Cmp1, "found.conflict");
1706 FirstInst = getFirstInst(FirstInst, IsConflict, Loc);
1707 if (MemoryRuntimeCheck) {
1709 ChkBuilder.CreateOr(MemoryRuntimeCheck, IsConflict, "conflict.rdx");
1710 FirstInst = getFirstInst(FirstInst, IsConflict, Loc);
1712 MemoryRuntimeCheck = IsConflict;
1715 if (!MemoryRuntimeCheck)
1716 return std::make_pair(nullptr, nullptr);
1718 // We have to do this trickery because the IRBuilder might fold the check to a
1719 // constant expression in which case there is no Instruction anchored in a
1721 Instruction *Check = BinaryOperator::CreateAnd(MemoryRuntimeCheck,
1722 ConstantInt::getTrue(Ctx));
1723 ChkBuilder.Insert(Check, "memcheck.conflict");
1724 FirstInst = getFirstInst(FirstInst, Check, Loc);
1725 return std::make_pair(FirstInst, Check);
1728 std::pair<Instruction *, Instruction *> LoopAccessInfo::addRuntimeCheck(
1729 Instruction *Loc, const SmallVectorImpl<int> *PtrPartition) const {
1730 if (!PtrRtChecking.Need)
1731 return std::make_pair(nullptr, nullptr);
1733 return addRuntimeCheck(Loc, PtrRtChecking.generateChecks(PtrPartition));
1736 LoopAccessInfo::LoopAccessInfo(Loop *L, ScalarEvolution *SE,
1737 const DataLayout &DL,
1738 const TargetLibraryInfo *TLI, AliasAnalysis *AA,
1739 DominatorTree *DT, LoopInfo *LI,
1740 const ValueToValueMap &Strides)
1741 : PtrRtChecking(SE), DepChecker(SE, L), TheLoop(L), SE(SE), DL(DL),
1742 TLI(TLI), AA(AA), DT(DT), LI(LI), NumLoads(0), NumStores(0),
1743 MaxSafeDepDistBytes(-1U), CanVecMem(false),
1744 StoreToLoopInvariantAddress(false) {
1745 if (canAnalyzeLoop())
1746 analyzeLoop(Strides);
1749 void LoopAccessInfo::print(raw_ostream &OS, unsigned Depth) const {
1751 if (PtrRtChecking.Need)
1752 OS.indent(Depth) << "Memory dependences are safe with run-time checks\n";
1754 OS.indent(Depth) << "Memory dependences are safe\n";
1758 OS.indent(Depth) << "Report: " << Report->str() << "\n";
1760 if (auto *InterestingDependences = DepChecker.getInterestingDependences()) {
1761 OS.indent(Depth) << "Interesting Dependences:\n";
1762 for (auto &Dep : *InterestingDependences) {
1763 Dep.print(OS, Depth + 2, DepChecker.getMemoryInstructions());
1767 OS.indent(Depth) << "Too many interesting dependences, not recorded\n";
1769 // List the pair of accesses need run-time checks to prove independence.
1770 PtrRtChecking.print(OS, Depth);
1773 OS.indent(Depth) << "Store to invariant address was "
1774 << (StoreToLoopInvariantAddress ? "" : "not ")
1775 << "found in loop.\n";
1778 const LoopAccessInfo &
1779 LoopAccessAnalysis::getInfo(Loop *L, const ValueToValueMap &Strides) {
1780 auto &LAI = LoopAccessInfoMap[L];
1783 assert((!LAI || LAI->NumSymbolicStrides == Strides.size()) &&
1784 "Symbolic strides changed for loop");
1788 const DataLayout &DL = L->getHeader()->getModule()->getDataLayout();
1789 LAI = llvm::make_unique<LoopAccessInfo>(L, SE, DL, TLI, AA, DT, LI,
1792 LAI->NumSymbolicStrides = Strides.size();
1798 void LoopAccessAnalysis::print(raw_ostream &OS, const Module *M) const {
1799 LoopAccessAnalysis &LAA = *const_cast<LoopAccessAnalysis *>(this);
1801 ValueToValueMap NoSymbolicStrides;
1803 for (Loop *TopLevelLoop : *LI)
1804 for (Loop *L : depth_first(TopLevelLoop)) {
1805 OS.indent(2) << L->getHeader()->getName() << ":\n";
1806 auto &LAI = LAA.getInfo(L, NoSymbolicStrides);
1811 bool LoopAccessAnalysis::runOnFunction(Function &F) {
1812 SE = &getAnalysis<ScalarEvolution>();
1813 auto *TLIP = getAnalysisIfAvailable<TargetLibraryInfoWrapperPass>();
1814 TLI = TLIP ? &TLIP->getTLI() : nullptr;
1815 AA = &getAnalysis<AliasAnalysis>();
1816 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1817 LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
1822 void LoopAccessAnalysis::getAnalysisUsage(AnalysisUsage &AU) const {
1823 AU.addRequired<ScalarEvolution>();
1824 AU.addRequired<AliasAnalysis>();
1825 AU.addRequired<DominatorTreeWrapperPass>();
1826 AU.addRequired<LoopInfoWrapperPass>();
1828 AU.setPreservesAll();
1831 char LoopAccessAnalysis::ID = 0;
1832 static const char laa_name[] = "Loop Access Analysis";
1833 #define LAA_NAME "loop-accesses"
1835 INITIALIZE_PASS_BEGIN(LoopAccessAnalysis, LAA_NAME, laa_name, false, true)
1836 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
1837 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
1838 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
1839 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
1840 INITIALIZE_PASS_END(LoopAccessAnalysis, LAA_NAME, laa_name, false, true)
1843 Pass *createLAAPass() {
1844 return new LoopAccessAnalysis();