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 dependences up to this threshold.
62 static cl::opt<unsigned>
63 MaxDependences("max-dependences", cl::Hidden,
64 cl::desc("Maximum number of 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(PredicatedScalarEvolution &PSE,
91 const ValueToValueMap &PtrToStride,
92 Value *Ptr, Value *OrigPtr) {
93 const SCEV *OrigSCEV = PSE.getSCEV(Ptr);
95 // If there is an entry in the map return the SCEV of the pointer with the
96 // symbolic stride replaced by one.
97 ValueToValueMap::const_iterator SI =
98 PtrToStride.find(OrigPtr ? OrigPtr : Ptr);
99 if (SI != PtrToStride.end()) {
100 Value *StrideVal = SI->second;
103 StrideVal = stripIntegerCast(StrideVal);
105 // Replace symbolic stride by one.
106 Value *One = ConstantInt::get(StrideVal->getType(), 1);
107 ValueToValueMap RewriteMap;
108 RewriteMap[StrideVal] = One;
110 ScalarEvolution *SE = PSE.getSE();
111 const auto *U = cast<SCEVUnknown>(SE->getSCEV(StrideVal));
113 static_cast<const SCEVConstant *>(SE->getOne(StrideVal->getType()));
115 PSE.addPredicate(*SE->getEqualPredicate(U, CT));
116 auto *Expr = PSE.getSCEV(Ptr);
118 DEBUG(dbgs() << "LAA: Replacing SCEV: " << *OrigSCEV << " by: " << *Expr
123 // Otherwise, just return the SCEV of the original pointer.
127 void RuntimePointerChecking::insert(Loop *Lp, Value *Ptr, bool WritePtr,
128 unsigned DepSetId, unsigned ASId,
129 const ValueToValueMap &Strides,
130 PredicatedScalarEvolution &PSE) {
131 // Get the stride replaced scev.
132 const SCEV *Sc = replaceSymbolicStrideSCEV(PSE, Strides, Ptr);
133 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Sc);
134 assert(AR && "Invalid addrec expression");
135 ScalarEvolution *SE = PSE.getSE();
136 const SCEV *Ex = SE->getBackedgeTakenCount(Lp);
138 const SCEV *ScStart = AR->getStart();
139 const SCEV *ScEnd = AR->evaluateAtIteration(Ex, *SE);
140 const SCEV *Step = AR->getStepRecurrence(*SE);
142 // For expressions with negative step, the upper bound is ScStart and the
143 // lower bound is ScEnd.
144 if (const SCEVConstant *CStep = dyn_cast<const SCEVConstant>(Step)) {
145 if (CStep->getValue()->isNegative())
146 std::swap(ScStart, ScEnd);
148 // Fallback case: the step is not constant, but the we can still
149 // get the upper and lower bounds of the interval by using min/max
151 ScStart = SE->getUMinExpr(ScStart, ScEnd);
152 ScEnd = SE->getUMaxExpr(AR->getStart(), ScEnd);
155 Pointers.emplace_back(Ptr, ScStart, ScEnd, WritePtr, DepSetId, ASId, Sc);
158 SmallVector<RuntimePointerChecking::PointerCheck, 4>
159 RuntimePointerChecking::generateChecks() const {
160 SmallVector<PointerCheck, 4> Checks;
162 for (unsigned I = 0; I < CheckingGroups.size(); ++I) {
163 for (unsigned J = I + 1; J < CheckingGroups.size(); ++J) {
164 const RuntimePointerChecking::CheckingPtrGroup &CGI = CheckingGroups[I];
165 const RuntimePointerChecking::CheckingPtrGroup &CGJ = CheckingGroups[J];
167 if (needsChecking(CGI, CGJ))
168 Checks.push_back(std::make_pair(&CGI, &CGJ));
174 void RuntimePointerChecking::generateChecks(
175 MemoryDepChecker::DepCandidates &DepCands, bool UseDependencies) {
176 assert(Checks.empty() && "Checks is not empty");
177 groupChecks(DepCands, UseDependencies);
178 Checks = generateChecks();
181 bool RuntimePointerChecking::needsChecking(const CheckingPtrGroup &M,
182 const CheckingPtrGroup &N) const {
183 for (unsigned I = 0, EI = M.Members.size(); EI != I; ++I)
184 for (unsigned J = 0, EJ = N.Members.size(); EJ != J; ++J)
185 if (needsChecking(M.Members[I], N.Members[J]))
190 /// Compare \p I and \p J and return the minimum.
191 /// Return nullptr in case we couldn't find an answer.
192 static const SCEV *getMinFromExprs(const SCEV *I, const SCEV *J,
193 ScalarEvolution *SE) {
194 const SCEV *Diff = SE->getMinusSCEV(J, I);
195 const SCEVConstant *C = dyn_cast<const SCEVConstant>(Diff);
199 if (C->getValue()->isNegative())
204 bool RuntimePointerChecking::CheckingPtrGroup::addPointer(unsigned Index) {
205 const SCEV *Start = RtCheck.Pointers[Index].Start;
206 const SCEV *End = RtCheck.Pointers[Index].End;
208 // Compare the starts and ends with the known minimum and maximum
209 // of this set. We need to know how we compare against the min/max
210 // of the set in order to be able to emit memchecks.
211 const SCEV *Min0 = getMinFromExprs(Start, Low, RtCheck.SE);
215 const SCEV *Min1 = getMinFromExprs(End, High, RtCheck.SE);
219 // Update the low bound expression if we've found a new min value.
223 // Update the high bound expression if we've found a new max value.
227 Members.push_back(Index);
231 void RuntimePointerChecking::groupChecks(
232 MemoryDepChecker::DepCandidates &DepCands, bool UseDependencies) {
233 // We build the groups from dependency candidates equivalence classes
235 // - We know that pointers in the same equivalence class share
236 // the same underlying object and therefore there is a chance
237 // that we can compare pointers
238 // - We wouldn't be able to merge two pointers for which we need
239 // to emit a memcheck. The classes in DepCands are already
240 // conveniently built such that no two pointers in the same
241 // class need checking against each other.
243 // We use the following (greedy) algorithm to construct the groups
244 // For every pointer in the equivalence class:
245 // For each existing group:
246 // - if the difference between this pointer and the min/max bounds
247 // of the group is a constant, then make the pointer part of the
248 // group and update the min/max bounds of that group as required.
250 CheckingGroups.clear();
252 // If we need to check two pointers to the same underlying object
253 // with a non-constant difference, we shouldn't perform any pointer
254 // grouping with those pointers. This is because we can easily get
255 // into cases where the resulting check would return false, even when
256 // the accesses are safe.
258 // The following example shows this:
259 // for (i = 0; i < 1000; ++i)
260 // a[5000 + i * m] = a[i] + a[i + 9000]
262 // Here grouping gives a check of (5000, 5000 + 1000 * m) against
263 // (0, 10000) which is always false. However, if m is 1, there is no
264 // dependence. Not grouping the checks for a[i] and a[i + 9000] allows
265 // us to perform an accurate check in this case.
267 // The above case requires that we have an UnknownDependence between
268 // accesses to the same underlying object. This cannot happen unless
269 // ShouldRetryWithRuntimeCheck is set, and therefore UseDependencies
270 // is also false. In this case we will use the fallback path and create
271 // separate checking groups for all pointers.
273 // If we don't have the dependency partitions, construct a new
274 // checking pointer group for each pointer. This is also required
275 // for correctness, because in this case we can have checking between
276 // pointers to the same underlying object.
277 if (!UseDependencies) {
278 for (unsigned I = 0; I < Pointers.size(); ++I)
279 CheckingGroups.push_back(CheckingPtrGroup(I, *this));
283 unsigned TotalComparisons = 0;
285 DenseMap<Value *, unsigned> PositionMap;
286 for (unsigned Index = 0; Index < Pointers.size(); ++Index)
287 PositionMap[Pointers[Index].PointerValue] = Index;
289 // We need to keep track of what pointers we've already seen so we
290 // don't process them twice.
291 SmallSet<unsigned, 2> Seen;
293 // Go through all equivalence classes, get the "pointer check groups"
294 // and add them to the overall solution. We use the order in which accesses
295 // appear in 'Pointers' to enforce determinism.
296 for (unsigned I = 0; I < Pointers.size(); ++I) {
297 // We've seen this pointer before, and therefore already processed
298 // its equivalence class.
302 MemoryDepChecker::MemAccessInfo Access(Pointers[I].PointerValue,
303 Pointers[I].IsWritePtr);
305 SmallVector<CheckingPtrGroup, 2> Groups;
306 auto LeaderI = DepCands.findValue(DepCands.getLeaderValue(Access));
308 // Because DepCands is constructed by visiting accesses in the order in
309 // which they appear in alias sets (which is deterministic) and the
310 // iteration order within an equivalence class member is only dependent on
311 // the order in which unions and insertions are performed on the
312 // equivalence class, the iteration order is deterministic.
313 for (auto MI = DepCands.member_begin(LeaderI), ME = DepCands.member_end();
315 unsigned Pointer = PositionMap[MI->getPointer()];
317 // Mark this pointer as seen.
318 Seen.insert(Pointer);
320 // Go through all the existing sets and see if we can find one
321 // which can include this pointer.
322 for (CheckingPtrGroup &Group : Groups) {
323 // Don't perform more than a certain amount of comparisons.
324 // This should limit the cost of grouping the pointers to something
325 // reasonable. If we do end up hitting this threshold, the algorithm
326 // will create separate groups for all remaining pointers.
327 if (TotalComparisons > MemoryCheckMergeThreshold)
332 if (Group.addPointer(Pointer)) {
339 // We couldn't add this pointer to any existing set or the threshold
340 // for the number of comparisons has been reached. Create a new group
341 // to hold the current pointer.
342 Groups.push_back(CheckingPtrGroup(Pointer, *this));
345 // We've computed the grouped checks for this partition.
346 // Save the results and continue with the next one.
347 std::copy(Groups.begin(), Groups.end(), std::back_inserter(CheckingGroups));
351 bool RuntimePointerChecking::arePointersInSamePartition(
352 const SmallVectorImpl<int> &PtrToPartition, unsigned PtrIdx1,
354 return (PtrToPartition[PtrIdx1] != -1 &&
355 PtrToPartition[PtrIdx1] == PtrToPartition[PtrIdx2]);
358 bool RuntimePointerChecking::needsChecking(unsigned I, unsigned J) const {
359 const PointerInfo &PointerI = Pointers[I];
360 const PointerInfo &PointerJ = Pointers[J];
362 // No need to check if two readonly pointers intersect.
363 if (!PointerI.IsWritePtr && !PointerJ.IsWritePtr)
366 // Only need to check pointers between two different dependency sets.
367 if (PointerI.DependencySetId == PointerJ.DependencySetId)
370 // Only need to check pointers in the same alias set.
371 if (PointerI.AliasSetId != PointerJ.AliasSetId)
377 void RuntimePointerChecking::printChecks(
378 raw_ostream &OS, const SmallVectorImpl<PointerCheck> &Checks,
379 unsigned Depth) const {
381 for (const auto &Check : Checks) {
382 const auto &First = Check.first->Members, &Second = Check.second->Members;
384 OS.indent(Depth) << "Check " << N++ << ":\n";
386 OS.indent(Depth + 2) << "Comparing group (" << Check.first << "):\n";
387 for (unsigned K = 0; K < First.size(); ++K)
388 OS.indent(Depth + 2) << *Pointers[First[K]].PointerValue << "\n";
390 OS.indent(Depth + 2) << "Against group (" << Check.second << "):\n";
391 for (unsigned K = 0; K < Second.size(); ++K)
392 OS.indent(Depth + 2) << *Pointers[Second[K]].PointerValue << "\n";
396 void RuntimePointerChecking::print(raw_ostream &OS, unsigned Depth) const {
398 OS.indent(Depth) << "Run-time memory checks:\n";
399 printChecks(OS, Checks, Depth);
401 OS.indent(Depth) << "Grouped accesses:\n";
402 for (unsigned I = 0; I < CheckingGroups.size(); ++I) {
403 const auto &CG = CheckingGroups[I];
405 OS.indent(Depth + 2) << "Group " << &CG << ":\n";
406 OS.indent(Depth + 4) << "(Low: " << *CG.Low << " High: " << *CG.High
408 for (unsigned J = 0; J < CG.Members.size(); ++J) {
409 OS.indent(Depth + 6) << "Member: " << *Pointers[CG.Members[J]].Expr
416 /// \brief Analyses memory accesses in a loop.
418 /// Checks whether run time pointer checks are needed and builds sets for data
419 /// dependence checking.
420 class AccessAnalysis {
422 /// \brief Read or write access location.
423 typedef PointerIntPair<Value *, 1, bool> MemAccessInfo;
424 typedef SmallPtrSet<MemAccessInfo, 8> MemAccessInfoSet;
426 AccessAnalysis(const DataLayout &Dl, AliasAnalysis *AA, LoopInfo *LI,
427 MemoryDepChecker::DepCandidates &DA,
428 PredicatedScalarEvolution &PSE)
429 : DL(Dl), AST(*AA), LI(LI), DepCands(DA), IsRTCheckAnalysisNeeded(false),
432 /// \brief Register a load and whether it is only read from.
433 void addLoad(MemoryLocation &Loc, bool IsReadOnly) {
434 Value *Ptr = const_cast<Value*>(Loc.Ptr);
435 AST.add(Ptr, MemoryLocation::UnknownSize, Loc.AATags);
436 Accesses.insert(MemAccessInfo(Ptr, false));
438 ReadOnlyPtr.insert(Ptr);
441 /// \brief Register a store.
442 void addStore(MemoryLocation &Loc) {
443 Value *Ptr = const_cast<Value*>(Loc.Ptr);
444 AST.add(Ptr, MemoryLocation::UnknownSize, Loc.AATags);
445 Accesses.insert(MemAccessInfo(Ptr, true));
448 /// \brief Check whether we can check the pointers at runtime for
449 /// non-intersection.
451 /// Returns true if we need no check or if we do and we can generate them
452 /// (i.e. the pointers have computable bounds).
453 bool canCheckPtrAtRT(RuntimePointerChecking &RtCheck, ScalarEvolution *SE,
454 Loop *TheLoop, const ValueToValueMap &Strides,
455 bool ShouldCheckStride = false);
457 /// \brief Goes over all memory accesses, checks whether a RT check is needed
458 /// and builds sets of dependent accesses.
459 void buildDependenceSets() {
460 processMemAccesses();
463 /// \brief Initial processing of memory accesses determined that we need to
464 /// perform dependency checking.
466 /// Note that this can later be cleared if we retry memcheck analysis without
467 /// dependency checking (i.e. ShouldRetryWithRuntimeCheck).
468 bool isDependencyCheckNeeded() { return !CheckDeps.empty(); }
470 /// We decided that no dependence analysis would be used. Reset the state.
471 void resetDepChecks(MemoryDepChecker &DepChecker) {
473 DepChecker.clearDependences();
476 MemAccessInfoSet &getDependenciesToCheck() { return CheckDeps; }
479 typedef SetVector<MemAccessInfo> PtrAccessSet;
481 /// \brief Go over all memory access and check whether runtime pointer checks
482 /// are needed and build sets of dependency check candidates.
483 void processMemAccesses();
485 /// Set of all accesses.
486 PtrAccessSet Accesses;
488 const DataLayout &DL;
490 /// Set of accesses that need a further dependence check.
491 MemAccessInfoSet CheckDeps;
493 /// Set of pointers that are read only.
494 SmallPtrSet<Value*, 16> ReadOnlyPtr;
496 /// An alias set tracker to partition the access set by underlying object and
497 //intrinsic property (such as TBAA metadata).
502 /// Sets of potentially dependent accesses - members of one set share an
503 /// underlying pointer. The set "CheckDeps" identfies which sets really need a
504 /// dependence check.
505 MemoryDepChecker::DepCandidates &DepCands;
507 /// \brief Initial processing of memory accesses determined that we may need
508 /// to add memchecks. Perform the analysis to determine the necessary checks.
510 /// Note that, this is different from isDependencyCheckNeeded. When we retry
511 /// memcheck analysis without dependency checking
512 /// (i.e. ShouldRetryWithRuntimeCheck), isDependencyCheckNeeded is cleared
513 /// while this remains set if we have potentially dependent accesses.
514 bool IsRTCheckAnalysisNeeded;
516 /// The SCEV predicate containing all the SCEV-related assumptions.
517 PredicatedScalarEvolution &PSE;
520 } // end anonymous namespace
522 /// \brief Check whether a pointer can participate in a runtime bounds check.
523 static bool hasComputableBounds(PredicatedScalarEvolution &PSE,
524 const ValueToValueMap &Strides, Value *Ptr,
526 const SCEV *PtrScev = replaceSymbolicStrideSCEV(PSE, Strides, Ptr);
527 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
531 return AR->isAffine();
534 bool AccessAnalysis::canCheckPtrAtRT(RuntimePointerChecking &RtCheck,
535 ScalarEvolution *SE, Loop *TheLoop,
536 const ValueToValueMap &StridesMap,
537 bool ShouldCheckStride) {
538 // Find pointers with computable bounds. We are going to use this information
539 // to place a runtime bound check.
542 bool NeedRTCheck = false;
543 if (!IsRTCheckAnalysisNeeded) return true;
545 bool IsDepCheckNeeded = isDependencyCheckNeeded();
547 // We assign a consecutive id to access from different alias sets.
548 // Accesses between different groups doesn't need to be checked.
550 for (auto &AS : AST) {
551 int NumReadPtrChecks = 0;
552 int NumWritePtrChecks = 0;
554 // We assign consecutive id to access from different dependence sets.
555 // Accesses within the same set don't need a runtime check.
556 unsigned RunningDepId = 1;
557 DenseMap<Value *, unsigned> DepSetId;
560 Value *Ptr = A.getValue();
561 bool IsWrite = Accesses.count(MemAccessInfo(Ptr, true));
562 MemAccessInfo Access(Ptr, IsWrite);
569 if (hasComputableBounds(PSE, StridesMap, Ptr, TheLoop) &&
570 // When we run after a failing dependency check we have to make sure
571 // we don't have wrapping pointers.
572 (!ShouldCheckStride ||
573 isStridedPtr(PSE, Ptr, TheLoop, StridesMap) == 1)) {
574 // The id of the dependence set.
577 if (IsDepCheckNeeded) {
578 Value *Leader = DepCands.getLeaderValue(Access).getPointer();
579 unsigned &LeaderId = DepSetId[Leader];
581 LeaderId = RunningDepId++;
584 // Each access has its own dependence set.
585 DepId = RunningDepId++;
587 RtCheck.insert(TheLoop, Ptr, IsWrite, DepId, ASId, StridesMap, PSE);
589 DEBUG(dbgs() << "LAA: Found a runtime check ptr:" << *Ptr << '\n');
591 DEBUG(dbgs() << "LAA: Can't find bounds for ptr:" << *Ptr << '\n');
596 // If we have at least two writes or one write and a read then we need to
597 // check them. But there is no need to checks if there is only one
598 // dependence set for this alias set.
600 // Note that this function computes CanDoRT and NeedRTCheck independently.
601 // For example CanDoRT=false, NeedRTCheck=false means that we have a pointer
602 // for which we couldn't find the bounds but we don't actually need to emit
603 // any checks so it does not matter.
604 if (!(IsDepCheckNeeded && CanDoRT && RunningDepId == 2))
605 NeedRTCheck |= (NumWritePtrChecks >= 2 || (NumReadPtrChecks >= 1 &&
606 NumWritePtrChecks >= 1));
611 // If the pointers that we would use for the bounds comparison have different
612 // address spaces, assume the values aren't directly comparable, so we can't
613 // use them for the runtime check. We also have to assume they could
614 // overlap. In the future there should be metadata for whether address spaces
616 unsigned NumPointers = RtCheck.Pointers.size();
617 for (unsigned i = 0; i < NumPointers; ++i) {
618 for (unsigned j = i + 1; j < NumPointers; ++j) {
619 // Only need to check pointers between two different dependency sets.
620 if (RtCheck.Pointers[i].DependencySetId ==
621 RtCheck.Pointers[j].DependencySetId)
623 // Only need to check pointers in the same alias set.
624 if (RtCheck.Pointers[i].AliasSetId != RtCheck.Pointers[j].AliasSetId)
627 Value *PtrI = RtCheck.Pointers[i].PointerValue;
628 Value *PtrJ = RtCheck.Pointers[j].PointerValue;
630 unsigned ASi = PtrI->getType()->getPointerAddressSpace();
631 unsigned ASj = PtrJ->getType()->getPointerAddressSpace();
633 DEBUG(dbgs() << "LAA: Runtime check would require comparison between"
634 " different address spaces\n");
640 if (NeedRTCheck && CanDoRT)
641 RtCheck.generateChecks(DepCands, IsDepCheckNeeded);
643 DEBUG(dbgs() << "LAA: We need to do " << RtCheck.getNumberOfChecks()
644 << " pointer comparisons.\n");
646 RtCheck.Need = NeedRTCheck;
648 bool CanDoRTIfNeeded = !NeedRTCheck || CanDoRT;
649 if (!CanDoRTIfNeeded)
651 return CanDoRTIfNeeded;
654 void AccessAnalysis::processMemAccesses() {
655 // We process the set twice: first we process read-write pointers, last we
656 // process read-only pointers. This allows us to skip dependence tests for
657 // read-only pointers.
659 DEBUG(dbgs() << "LAA: Processing memory accesses...\n");
660 DEBUG(dbgs() << " AST: "; AST.dump());
661 DEBUG(dbgs() << "LAA: Accesses(" << Accesses.size() << "):\n");
663 for (auto A : Accesses)
664 dbgs() << "\t" << *A.getPointer() << " (" <<
665 (A.getInt() ? "write" : (ReadOnlyPtr.count(A.getPointer()) ?
666 "read-only" : "read")) << ")\n";
669 // The AliasSetTracker has nicely partitioned our pointers by metadata
670 // compatibility and potential for underlying-object overlap. As a result, we
671 // only need to check for potential pointer dependencies within each alias
673 for (auto &AS : AST) {
674 // Note that both the alias-set tracker and the alias sets themselves used
675 // linked lists internally and so the iteration order here is deterministic
676 // (matching the original instruction order within each set).
678 bool SetHasWrite = false;
680 // Map of pointers to last access encountered.
681 typedef DenseMap<Value*, MemAccessInfo> UnderlyingObjToAccessMap;
682 UnderlyingObjToAccessMap ObjToLastAccess;
684 // Set of access to check after all writes have been processed.
685 PtrAccessSet DeferredAccesses;
687 // Iterate over each alias set twice, once to process read/write pointers,
688 // and then to process read-only pointers.
689 for (int SetIteration = 0; SetIteration < 2; ++SetIteration) {
690 bool UseDeferred = SetIteration > 0;
691 PtrAccessSet &S = UseDeferred ? DeferredAccesses : Accesses;
694 Value *Ptr = AV.getValue();
696 // For a single memory access in AliasSetTracker, Accesses may contain
697 // both read and write, and they both need to be handled for CheckDeps.
699 if (AC.getPointer() != Ptr)
702 bool IsWrite = AC.getInt();
704 // If we're using the deferred access set, then it contains only
706 bool IsReadOnlyPtr = ReadOnlyPtr.count(Ptr) && !IsWrite;
707 if (UseDeferred && !IsReadOnlyPtr)
709 // Otherwise, the pointer must be in the PtrAccessSet, either as a
711 assert(((IsReadOnlyPtr && UseDeferred) || IsWrite ||
712 S.count(MemAccessInfo(Ptr, false))) &&
713 "Alias-set pointer not in the access set?");
715 MemAccessInfo Access(Ptr, IsWrite);
716 DepCands.insert(Access);
718 // Memorize read-only pointers for later processing and skip them in
719 // the first round (they need to be checked after we have seen all
720 // write pointers). Note: we also mark pointer that are not
721 // consecutive as "read-only" pointers (so that we check
722 // "a[b[i]] +="). Hence, we need the second check for "!IsWrite".
723 if (!UseDeferred && IsReadOnlyPtr) {
724 DeferredAccesses.insert(Access);
728 // If this is a write - check other reads and writes for conflicts. If
729 // this is a read only check other writes for conflicts (but only if
730 // there is no other write to the ptr - this is an optimization to
731 // catch "a[i] = a[i] + " without having to do a dependence check).
732 if ((IsWrite || IsReadOnlyPtr) && SetHasWrite) {
733 CheckDeps.insert(Access);
734 IsRTCheckAnalysisNeeded = true;
740 // Create sets of pointers connected by a shared alias set and
741 // underlying object.
742 typedef SmallVector<Value *, 16> ValueVector;
743 ValueVector TempObjects;
745 GetUnderlyingObjects(Ptr, TempObjects, DL, LI);
746 DEBUG(dbgs() << "Underlying objects for pointer " << *Ptr << "\n");
747 for (Value *UnderlyingObj : TempObjects) {
748 // nullptr never alias, don't join sets for pointer that have "null"
749 // in their UnderlyingObjects list.
750 if (isa<ConstantPointerNull>(UnderlyingObj))
753 UnderlyingObjToAccessMap::iterator Prev =
754 ObjToLastAccess.find(UnderlyingObj);
755 if (Prev != ObjToLastAccess.end())
756 DepCands.unionSets(Access, Prev->second);
758 ObjToLastAccess[UnderlyingObj] = Access;
759 DEBUG(dbgs() << " " << *UnderlyingObj << "\n");
767 static bool isInBoundsGep(Value *Ptr) {
768 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr))
769 return GEP->isInBounds();
773 /// \brief Return true if an AddRec pointer \p Ptr is unsigned non-wrapping,
774 /// i.e. monotonically increasing/decreasing.
775 static bool isNoWrapAddRec(Value *Ptr, const SCEVAddRecExpr *AR,
776 ScalarEvolution *SE, const Loop *L) {
777 // FIXME: This should probably only return true for NUW.
778 if (AR->getNoWrapFlags(SCEV::NoWrapMask))
781 // Scalar evolution does not propagate the non-wrapping flags to values that
782 // are derived from a non-wrapping induction variable because non-wrapping
783 // could be flow-sensitive.
785 // Look through the potentially overflowing instruction to try to prove
786 // non-wrapping for the *specific* value of Ptr.
788 // The arithmetic implied by an inbounds GEP can't overflow.
789 auto *GEP = dyn_cast<GetElementPtrInst>(Ptr);
790 if (!GEP || !GEP->isInBounds())
793 // Make sure there is only one non-const index and analyze that.
794 Value *NonConstIndex = nullptr;
795 for (auto Index = GEP->idx_begin(); Index != GEP->idx_end(); ++Index)
796 if (!isa<ConstantInt>(*Index)) {
799 NonConstIndex = *Index;
802 // The recurrence is on the pointer, ignore for now.
805 // The index in GEP is signed. It is non-wrapping if it's derived from a NSW
806 // AddRec using a NSW operation.
807 if (auto *OBO = dyn_cast<OverflowingBinaryOperator>(NonConstIndex))
808 if (OBO->hasNoSignedWrap() &&
809 // Assume constant for other the operand so that the AddRec can be
811 isa<ConstantInt>(OBO->getOperand(1))) {
812 auto *OpScev = SE->getSCEV(OBO->getOperand(0));
814 if (auto *OpAR = dyn_cast<SCEVAddRecExpr>(OpScev))
815 return OpAR->getLoop() == L && OpAR->getNoWrapFlags(SCEV::FlagNSW);
821 /// \brief Check whether the access through \p Ptr has a constant stride.
822 int llvm::isStridedPtr(PredicatedScalarEvolution &PSE, Value *Ptr,
823 const Loop *Lp, const ValueToValueMap &StridesMap) {
824 Type *Ty = Ptr->getType();
825 assert(Ty->isPointerTy() && "Unexpected non-ptr");
827 // Make sure that the pointer does not point to aggregate types.
828 auto *PtrTy = cast<PointerType>(Ty);
829 if (PtrTy->getElementType()->isAggregateType()) {
830 DEBUG(dbgs() << "LAA: Bad stride - Not a pointer to a scalar type"
835 const SCEV *PtrScev = replaceSymbolicStrideSCEV(PSE, StridesMap, Ptr);
837 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
839 DEBUG(dbgs() << "LAA: Bad stride - Not an AddRecExpr pointer "
840 << *Ptr << " SCEV: " << *PtrScev << "\n");
844 // The accesss function must stride over the innermost loop.
845 if (Lp != AR->getLoop()) {
846 DEBUG(dbgs() << "LAA: Bad stride - Not striding over innermost loop " <<
847 *Ptr << " SCEV: " << *PtrScev << "\n");
850 // The address calculation must not wrap. Otherwise, a dependence could be
852 // An inbounds getelementptr that is a AddRec with a unit stride
853 // cannot wrap per definition. The unit stride requirement is checked later.
854 // An getelementptr without an inbounds attribute and unit stride would have
855 // to access the pointer value "0" which is undefined behavior in address
856 // space 0, therefore we can also vectorize this case.
857 bool IsInBoundsGEP = isInBoundsGep(Ptr);
858 bool IsNoWrapAddRec = isNoWrapAddRec(Ptr, AR, PSE.getSE(), Lp);
859 bool IsInAddressSpaceZero = PtrTy->getAddressSpace() == 0;
860 if (!IsNoWrapAddRec && !IsInBoundsGEP && !IsInAddressSpaceZero) {
861 DEBUG(dbgs() << "LAA: Bad stride - Pointer may wrap in the address space "
862 << *Ptr << " SCEV: " << *PtrScev << "\n");
866 // Check the step is constant.
867 const SCEV *Step = AR->getStepRecurrence(*PSE.getSE());
869 // Calculate the pointer stride and check if it is constant.
870 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
872 DEBUG(dbgs() << "LAA: Bad stride - Not a constant strided " << *Ptr <<
873 " SCEV: " << *PtrScev << "\n");
877 auto &DL = Lp->getHeader()->getModule()->getDataLayout();
878 int64_t Size = DL.getTypeAllocSize(PtrTy->getElementType());
879 const APInt &APStepVal = C->getValue()->getValue();
881 // Huge step value - give up.
882 if (APStepVal.getBitWidth() > 64)
885 int64_t StepVal = APStepVal.getSExtValue();
888 int64_t Stride = StepVal / Size;
889 int64_t Rem = StepVal % Size;
893 // If the SCEV could wrap but we have an inbounds gep with a unit stride we
894 // know we can't "wrap around the address space". In case of address space
895 // zero we know that this won't happen without triggering undefined behavior.
896 if (!IsNoWrapAddRec && (IsInBoundsGEP || IsInAddressSpaceZero) &&
897 Stride != 1 && Stride != -1)
903 bool MemoryDepChecker::Dependence::isSafeForVectorization(DepType Type) {
907 case BackwardVectorizable:
911 case ForwardButPreventsForwarding:
913 case BackwardVectorizableButPreventsForwarding:
916 llvm_unreachable("unexpected DepType!");
919 bool MemoryDepChecker::Dependence::isBackward() const {
923 case ForwardButPreventsForwarding:
927 case BackwardVectorizable:
929 case BackwardVectorizableButPreventsForwarding:
932 llvm_unreachable("unexpected DepType!");
935 bool MemoryDepChecker::Dependence::isPossiblyBackward() const {
936 return isBackward() || Type == Unknown;
939 bool MemoryDepChecker::Dependence::isForward() const {
942 case ForwardButPreventsForwarding:
947 case BackwardVectorizable:
949 case BackwardVectorizableButPreventsForwarding:
952 llvm_unreachable("unexpected DepType!");
955 bool MemoryDepChecker::couldPreventStoreLoadForward(unsigned Distance,
956 unsigned TypeByteSize) {
957 // If loads occur at a distance that is not a multiple of a feasible vector
958 // factor store-load forwarding does not take place.
959 // Positive dependences might cause troubles because vectorizing them might
960 // prevent store-load forwarding making vectorized code run a lot slower.
961 // a[i] = a[i-3] ^ a[i-8];
962 // The stores to a[i:i+1] don't align with the stores to a[i-3:i-2] and
963 // hence on your typical architecture store-load forwarding does not take
964 // place. Vectorizing in such cases does not make sense.
965 // Store-load forwarding distance.
966 const unsigned NumCyclesForStoreLoadThroughMemory = 8*TypeByteSize;
967 // Maximum vector factor.
968 unsigned MaxVFWithoutSLForwardIssues =
969 VectorizerParams::MaxVectorWidth * TypeByteSize;
970 if(MaxSafeDepDistBytes < MaxVFWithoutSLForwardIssues)
971 MaxVFWithoutSLForwardIssues = MaxSafeDepDistBytes;
973 for (unsigned vf = 2*TypeByteSize; vf <= MaxVFWithoutSLForwardIssues;
975 if (Distance % vf && Distance / vf < NumCyclesForStoreLoadThroughMemory) {
976 MaxVFWithoutSLForwardIssues = (vf >>=1);
981 if (MaxVFWithoutSLForwardIssues< 2*TypeByteSize) {
982 DEBUG(dbgs() << "LAA: Distance " << Distance <<
983 " that could cause a store-load forwarding conflict\n");
987 if (MaxVFWithoutSLForwardIssues < MaxSafeDepDistBytes &&
988 MaxVFWithoutSLForwardIssues !=
989 VectorizerParams::MaxVectorWidth * TypeByteSize)
990 MaxSafeDepDistBytes = MaxVFWithoutSLForwardIssues;
994 /// \brief Check the dependence for two accesses with the same stride \p Stride.
995 /// \p Distance is the positive distance and \p TypeByteSize is type size in
998 /// \returns true if they are independent.
999 static bool areStridedAccessesIndependent(unsigned Distance, unsigned Stride,
1000 unsigned TypeByteSize) {
1001 assert(Stride > 1 && "The stride must be greater than 1");
1002 assert(TypeByteSize > 0 && "The type size in byte must be non-zero");
1003 assert(Distance > 0 && "The distance must be non-zero");
1005 // Skip if the distance is not multiple of type byte size.
1006 if (Distance % TypeByteSize)
1009 unsigned ScaledDist = Distance / TypeByteSize;
1011 // No dependence if the scaled distance is not multiple of the stride.
1013 // for (i = 0; i < 1024 ; i += 4)
1014 // A[i+2] = A[i] + 1;
1016 // Two accesses in memory (scaled distance is 2, stride is 4):
1017 // | A[0] | | | | A[4] | | | |
1018 // | | | A[2] | | | | A[6] | |
1021 // for (i = 0; i < 1024 ; i += 3)
1022 // A[i+4] = A[i] + 1;
1024 // Two accesses in memory (scaled distance is 4, stride is 3):
1025 // | A[0] | | | A[3] | | | A[6] | | |
1026 // | | | | | A[4] | | | A[7] | |
1027 return ScaledDist % Stride;
1030 MemoryDepChecker::Dependence::DepType
1031 MemoryDepChecker::isDependent(const MemAccessInfo &A, unsigned AIdx,
1032 const MemAccessInfo &B, unsigned BIdx,
1033 const ValueToValueMap &Strides) {
1034 assert (AIdx < BIdx && "Must pass arguments in program order");
1036 Value *APtr = A.getPointer();
1037 Value *BPtr = B.getPointer();
1038 bool AIsWrite = A.getInt();
1039 bool BIsWrite = B.getInt();
1041 // Two reads are independent.
1042 if (!AIsWrite && !BIsWrite)
1043 return Dependence::NoDep;
1045 // We cannot check pointers in different address spaces.
1046 if (APtr->getType()->getPointerAddressSpace() !=
1047 BPtr->getType()->getPointerAddressSpace())
1048 return Dependence::Unknown;
1050 const SCEV *AScev = replaceSymbolicStrideSCEV(PSE, Strides, APtr);
1051 const SCEV *BScev = replaceSymbolicStrideSCEV(PSE, Strides, BPtr);
1053 int StrideAPtr = isStridedPtr(PSE, APtr, InnermostLoop, Strides);
1054 int StrideBPtr = isStridedPtr(PSE, BPtr, InnermostLoop, Strides);
1056 const SCEV *Src = AScev;
1057 const SCEV *Sink = BScev;
1059 // If the induction step is negative we have to invert source and sink of the
1061 if (StrideAPtr < 0) {
1064 std::swap(APtr, BPtr);
1065 std::swap(Src, Sink);
1066 std::swap(AIsWrite, BIsWrite);
1067 std::swap(AIdx, BIdx);
1068 std::swap(StrideAPtr, StrideBPtr);
1071 const SCEV *Dist = PSE.getSE()->getMinusSCEV(Sink, Src);
1073 DEBUG(dbgs() << "LAA: Src Scev: " << *Src << "Sink Scev: " << *Sink
1074 << "(Induction step: " << StrideAPtr << ")\n");
1075 DEBUG(dbgs() << "LAA: Distance for " << *InstMap[AIdx] << " to "
1076 << *InstMap[BIdx] << ": " << *Dist << "\n");
1078 // Need accesses with constant stride. We don't want to vectorize
1079 // "A[B[i]] += ..." and similar code or pointer arithmetic that could wrap in
1080 // the address space.
1081 if (!StrideAPtr || !StrideBPtr || StrideAPtr != StrideBPtr){
1082 DEBUG(dbgs() << "Pointer access with non-constant stride\n");
1083 return Dependence::Unknown;
1086 const SCEVConstant *C = dyn_cast<SCEVConstant>(Dist);
1088 DEBUG(dbgs() << "LAA: Dependence because of non-constant distance\n");
1089 ShouldRetryWithRuntimeCheck = true;
1090 return Dependence::Unknown;
1093 Type *ATy = APtr->getType()->getPointerElementType();
1094 Type *BTy = BPtr->getType()->getPointerElementType();
1095 auto &DL = InnermostLoop->getHeader()->getModule()->getDataLayout();
1096 unsigned TypeByteSize = DL.getTypeAllocSize(ATy);
1098 // Negative distances are not plausible dependencies.
1099 const APInt &Val = C->getValue()->getValue();
1100 if (Val.isNegative()) {
1101 bool IsTrueDataDependence = (AIsWrite && !BIsWrite);
1102 if (IsTrueDataDependence &&
1103 (couldPreventStoreLoadForward(Val.abs().getZExtValue(), TypeByteSize) ||
1105 return Dependence::ForwardButPreventsForwarding;
1107 DEBUG(dbgs() << "LAA: Dependence is negative: NoDep\n");
1108 return Dependence::Forward;
1111 // Write to the same location with the same size.
1112 // Could be improved to assert type sizes are the same (i32 == float, etc).
1115 return Dependence::Forward;
1116 DEBUG(dbgs() << "LAA: Zero dependence difference but different types\n");
1117 return Dependence::Unknown;
1120 assert(Val.isStrictlyPositive() && "Expect a positive value");
1124 "LAA: ReadWrite-Write positive dependency with different types\n");
1125 return Dependence::Unknown;
1128 unsigned Distance = (unsigned) Val.getZExtValue();
1130 unsigned Stride = std::abs(StrideAPtr);
1132 areStridedAccessesIndependent(Distance, Stride, TypeByteSize)) {
1133 DEBUG(dbgs() << "LAA: Strided accesses are independent\n");
1134 return Dependence::NoDep;
1137 // Bail out early if passed-in parameters make vectorization not feasible.
1138 unsigned ForcedFactor = (VectorizerParams::VectorizationFactor ?
1139 VectorizerParams::VectorizationFactor : 1);
1140 unsigned ForcedUnroll = (VectorizerParams::VectorizationInterleave ?
1141 VectorizerParams::VectorizationInterleave : 1);
1142 // The minimum number of iterations for a vectorized/unrolled version.
1143 unsigned MinNumIter = std::max(ForcedFactor * ForcedUnroll, 2U);
1145 // It's not vectorizable if the distance is smaller than the minimum distance
1146 // needed for a vectroized/unrolled version. Vectorizing one iteration in
1147 // front needs TypeByteSize * Stride. Vectorizing the last iteration needs
1148 // TypeByteSize (No need to plus the last gap distance).
1150 // E.g. Assume one char is 1 byte in memory and one int is 4 bytes.
1152 // int *B = (int *)((char *)A + 14);
1153 // for (i = 0 ; i < 1024 ; i += 2)
1157 // Two accesses in memory (stride is 2):
1158 // | A[0] | | A[2] | | A[4] | | A[6] | |
1159 // | B[0] | | B[2] | | B[4] |
1161 // Distance needs for vectorizing iterations except the last iteration:
1162 // 4 * 2 * (MinNumIter - 1). Distance needs for the last iteration: 4.
1163 // So the minimum distance needed is: 4 * 2 * (MinNumIter - 1) + 4.
1165 // If MinNumIter is 2, it is vectorizable as the minimum distance needed is
1166 // 12, which is less than distance.
1168 // If MinNumIter is 4 (Say if a user forces the vectorization factor to be 4),
1169 // the minimum distance needed is 28, which is greater than distance. It is
1170 // not safe to do vectorization.
1171 unsigned MinDistanceNeeded =
1172 TypeByteSize * Stride * (MinNumIter - 1) + TypeByteSize;
1173 if (MinDistanceNeeded > Distance) {
1174 DEBUG(dbgs() << "LAA: Failure because of positive distance " << Distance
1176 return Dependence::Backward;
1179 // Unsafe if the minimum distance needed is greater than max safe distance.
1180 if (MinDistanceNeeded > MaxSafeDepDistBytes) {
1181 DEBUG(dbgs() << "LAA: Failure because it needs at least "
1182 << MinDistanceNeeded << " size in bytes");
1183 return Dependence::Backward;
1186 // Positive distance bigger than max vectorization factor.
1187 // FIXME: Should use max factor instead of max distance in bytes, which could
1188 // not handle different types.
1189 // E.g. Assume one char is 1 byte in memory and one int is 4 bytes.
1190 // void foo (int *A, char *B) {
1191 // for (unsigned i = 0; i < 1024; i++) {
1192 // A[i+2] = A[i] + 1;
1193 // B[i+2] = B[i] + 1;
1197 // This case is currently unsafe according to the max safe distance. If we
1198 // analyze the two accesses on array B, the max safe dependence distance
1199 // is 2. Then we analyze the accesses on array A, the minimum distance needed
1200 // is 8, which is less than 2 and forbidden vectorization, But actually
1201 // both A and B could be vectorized by 2 iterations.
1202 MaxSafeDepDistBytes =
1203 Distance < MaxSafeDepDistBytes ? Distance : MaxSafeDepDistBytes;
1205 bool IsTrueDataDependence = (!AIsWrite && BIsWrite);
1206 if (IsTrueDataDependence &&
1207 couldPreventStoreLoadForward(Distance, TypeByteSize))
1208 return Dependence::BackwardVectorizableButPreventsForwarding;
1210 DEBUG(dbgs() << "LAA: Positive distance " << Val.getSExtValue()
1211 << " with max VF = "
1212 << MaxSafeDepDistBytes / (TypeByteSize * Stride) << '\n');
1214 return Dependence::BackwardVectorizable;
1217 bool MemoryDepChecker::areDepsSafe(DepCandidates &AccessSets,
1218 MemAccessInfoSet &CheckDeps,
1219 const ValueToValueMap &Strides) {
1221 MaxSafeDepDistBytes = -1U;
1222 while (!CheckDeps.empty()) {
1223 MemAccessInfo CurAccess = *CheckDeps.begin();
1225 // Get the relevant memory access set.
1226 EquivalenceClasses<MemAccessInfo>::iterator I =
1227 AccessSets.findValue(AccessSets.getLeaderValue(CurAccess));
1229 // Check accesses within this set.
1230 EquivalenceClasses<MemAccessInfo>::member_iterator AI, AE;
1231 AI = AccessSets.member_begin(I), AE = AccessSets.member_end();
1233 // Check every access pair.
1235 CheckDeps.erase(*AI);
1236 EquivalenceClasses<MemAccessInfo>::member_iterator OI = std::next(AI);
1238 // Check every accessing instruction pair in program order.
1239 for (std::vector<unsigned>::iterator I1 = Accesses[*AI].begin(),
1240 I1E = Accesses[*AI].end(); I1 != I1E; ++I1)
1241 for (std::vector<unsigned>::iterator I2 = Accesses[*OI].begin(),
1242 I2E = Accesses[*OI].end(); I2 != I2E; ++I2) {
1243 auto A = std::make_pair(&*AI, *I1);
1244 auto B = std::make_pair(&*OI, *I2);
1250 Dependence::DepType Type =
1251 isDependent(*A.first, A.second, *B.first, B.second, Strides);
1252 SafeForVectorization &= Dependence::isSafeForVectorization(Type);
1254 // Gather dependences unless we accumulated MaxDependences
1255 // dependences. In that case return as soon as we find the first
1256 // unsafe dependence. This puts a limit on this quadratic
1258 if (RecordDependences) {
1259 if (Type != Dependence::NoDep)
1260 Dependences.push_back(Dependence(A.second, B.second, Type));
1262 if (Dependences.size() >= MaxDependences) {
1263 RecordDependences = false;
1264 Dependences.clear();
1265 DEBUG(dbgs() << "Too many dependences, stopped recording\n");
1268 if (!RecordDependences && !SafeForVectorization)
1277 DEBUG(dbgs() << "Total Dependences: " << Dependences.size() << "\n");
1278 return SafeForVectorization;
1281 SmallVector<Instruction *, 4>
1282 MemoryDepChecker::getInstructionsForAccess(Value *Ptr, bool isWrite) const {
1283 MemAccessInfo Access(Ptr, isWrite);
1284 auto &IndexVector = Accesses.find(Access)->second;
1286 SmallVector<Instruction *, 4> Insts;
1287 std::transform(IndexVector.begin(), IndexVector.end(),
1288 std::back_inserter(Insts),
1289 [&](unsigned Idx) { return this->InstMap[Idx]; });
1293 const char *MemoryDepChecker::Dependence::DepName[] = {
1294 "NoDep", "Unknown", "Forward", "ForwardButPreventsForwarding", "Backward",
1295 "BackwardVectorizable", "BackwardVectorizableButPreventsForwarding"};
1297 void MemoryDepChecker::Dependence::print(
1298 raw_ostream &OS, unsigned Depth,
1299 const SmallVectorImpl<Instruction *> &Instrs) const {
1300 OS.indent(Depth) << DepName[Type] << ":\n";
1301 OS.indent(Depth + 2) << *Instrs[Source] << " -> \n";
1302 OS.indent(Depth + 2) << *Instrs[Destination] << "\n";
1305 bool LoopAccessInfo::canAnalyzeLoop() {
1306 // We need to have a loop header.
1307 DEBUG(dbgs() << "LAA: Found a loop: " <<
1308 TheLoop->getHeader()->getName() << '\n');
1310 // We can only analyze innermost loops.
1311 if (!TheLoop->empty()) {
1312 DEBUG(dbgs() << "LAA: loop is not the innermost loop\n");
1313 emitAnalysis(LoopAccessReport() << "loop is not the innermost loop");
1317 // We must have a single backedge.
1318 if (TheLoop->getNumBackEdges() != 1) {
1319 DEBUG(dbgs() << "LAA: loop control flow is not understood by analyzer\n");
1321 LoopAccessReport() <<
1322 "loop control flow is not understood by analyzer");
1326 // We must have a single exiting block.
1327 if (!TheLoop->getExitingBlock()) {
1328 DEBUG(dbgs() << "LAA: loop control flow is not understood by analyzer\n");
1330 LoopAccessReport() <<
1331 "loop control flow is not understood by analyzer");
1335 // We only handle bottom-tested loops, i.e. loop in which the condition is
1336 // checked at the end of each iteration. With that we can assume that all
1337 // instructions in the loop are executed the same number of times.
1338 if (TheLoop->getExitingBlock() != TheLoop->getLoopLatch()) {
1339 DEBUG(dbgs() << "LAA: loop control flow is not understood by analyzer\n");
1341 LoopAccessReport() <<
1342 "loop control flow is not understood by analyzer");
1346 // ScalarEvolution needs to be able to find the exit count.
1347 const SCEV *ExitCount = PSE.getSE()->getBackedgeTakenCount(TheLoop);
1348 if (ExitCount == PSE.getSE()->getCouldNotCompute()) {
1349 emitAnalysis(LoopAccessReport()
1350 << "could not determine number of loop iterations");
1351 DEBUG(dbgs() << "LAA: SCEV could not compute the loop exit count.\n");
1358 void LoopAccessInfo::analyzeLoop(const ValueToValueMap &Strides) {
1360 typedef SmallVector<Value*, 16> ValueVector;
1361 typedef SmallPtrSet<Value*, 16> ValueSet;
1363 // Holds the Load and Store *instructions*.
1367 // Holds all the different accesses in the loop.
1368 unsigned NumReads = 0;
1369 unsigned NumReadWrites = 0;
1371 PtrRtChecking.Pointers.clear();
1372 PtrRtChecking.Need = false;
1374 const bool IsAnnotatedParallel = TheLoop->isAnnotatedParallel();
1377 for (Loop::block_iterator bb = TheLoop->block_begin(),
1378 be = TheLoop->block_end(); bb != be; ++bb) {
1380 // Scan the BB and collect legal loads and stores.
1381 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
1384 // If this is a load, save it. If this instruction can read from memory
1385 // but is not a load, then we quit. Notice that we don't handle function
1386 // calls that read or write.
1387 if (it->mayReadFromMemory()) {
1388 // Many math library functions read the rounding mode. We will only
1389 // vectorize a loop if it contains known function calls that don't set
1390 // the flag. Therefore, it is safe to ignore this read from memory.
1391 CallInst *Call = dyn_cast<CallInst>(it);
1392 if (Call && getIntrinsicIDForCall(Call, TLI))
1395 // If the function has an explicit vectorized counterpart, we can safely
1396 // assume that it can be vectorized.
1397 if (Call && !Call->isNoBuiltin() && Call->getCalledFunction() &&
1398 TLI->isFunctionVectorizable(Call->getCalledFunction()->getName()))
1401 LoadInst *Ld = dyn_cast<LoadInst>(it);
1402 if (!Ld || (!Ld->isSimple() && !IsAnnotatedParallel)) {
1403 emitAnalysis(LoopAccessReport(Ld)
1404 << "read with atomic ordering or volatile read");
1405 DEBUG(dbgs() << "LAA: Found a non-simple load.\n");
1410 Loads.push_back(Ld);
1411 DepChecker.addAccess(Ld);
1415 // Save 'store' instructions. Abort if other instructions write to memory.
1416 if (it->mayWriteToMemory()) {
1417 StoreInst *St = dyn_cast<StoreInst>(it);
1419 emitAnalysis(LoopAccessReport(&*it) <<
1420 "instruction cannot be vectorized");
1424 if (!St->isSimple() && !IsAnnotatedParallel) {
1425 emitAnalysis(LoopAccessReport(St)
1426 << "write with atomic ordering or volatile write");
1427 DEBUG(dbgs() << "LAA: Found a non-simple store.\n");
1432 Stores.push_back(St);
1433 DepChecker.addAccess(St);
1438 // Now we have two lists that hold the loads and the stores.
1439 // Next, we find the pointers that they use.
1441 // Check if we see any stores. If there are no stores, then we don't
1442 // care if the pointers are *restrict*.
1443 if (!Stores.size()) {
1444 DEBUG(dbgs() << "LAA: Found a read-only loop!\n");
1449 MemoryDepChecker::DepCandidates DependentAccesses;
1450 AccessAnalysis Accesses(TheLoop->getHeader()->getModule()->getDataLayout(),
1451 AA, LI, DependentAccesses, PSE);
1453 // Holds the analyzed pointers. We don't want to call GetUnderlyingObjects
1454 // multiple times on the same object. If the ptr is accessed twice, once
1455 // for read and once for write, it will only appear once (on the write
1456 // list). This is okay, since we are going to check for conflicts between
1457 // writes and between reads and writes, but not between reads and reads.
1460 ValueVector::iterator I, IE;
1461 for (I = Stores.begin(), IE = Stores.end(); I != IE; ++I) {
1462 StoreInst *ST = cast<StoreInst>(*I);
1463 Value* Ptr = ST->getPointerOperand();
1464 // Check for store to loop invariant address.
1465 StoreToLoopInvariantAddress |= isUniform(Ptr);
1466 // If we did *not* see this pointer before, insert it to the read-write
1467 // list. At this phase it is only a 'write' list.
1468 if (Seen.insert(Ptr).second) {
1471 MemoryLocation Loc = MemoryLocation::get(ST);
1472 // The TBAA metadata could have a control dependency on the predication
1473 // condition, so we cannot rely on it when determining whether or not we
1474 // need runtime pointer checks.
1475 if (blockNeedsPredication(ST->getParent(), TheLoop, DT))
1476 Loc.AATags.TBAA = nullptr;
1478 Accesses.addStore(Loc);
1482 if (IsAnnotatedParallel) {
1484 << "LAA: A loop annotated parallel, ignore memory dependency "
1490 for (I = Loads.begin(), IE = Loads.end(); I != IE; ++I) {
1491 LoadInst *LD = cast<LoadInst>(*I);
1492 Value* Ptr = LD->getPointerOperand();
1493 // If we did *not* see this pointer before, insert it to the
1494 // read list. If we *did* see it before, then it is already in
1495 // the read-write list. This allows us to vectorize expressions
1496 // such as A[i] += x; Because the address of A[i] is a read-write
1497 // pointer. This only works if the index of A[i] is consecutive.
1498 // If the address of i is unknown (for example A[B[i]]) then we may
1499 // read a few words, modify, and write a few words, and some of the
1500 // words may be written to the same address.
1501 bool IsReadOnlyPtr = false;
1502 if (Seen.insert(Ptr).second || !isStridedPtr(PSE, Ptr, TheLoop, Strides)) {
1504 IsReadOnlyPtr = true;
1507 MemoryLocation Loc = MemoryLocation::get(LD);
1508 // The TBAA metadata could have a control dependency on the predication
1509 // condition, so we cannot rely on it when determining whether or not we
1510 // need runtime pointer checks.
1511 if (blockNeedsPredication(LD->getParent(), TheLoop, DT))
1512 Loc.AATags.TBAA = nullptr;
1514 Accesses.addLoad(Loc, IsReadOnlyPtr);
1517 // If we write (or read-write) to a single destination and there are no
1518 // other reads in this loop then is it safe to vectorize.
1519 if (NumReadWrites == 1 && NumReads == 0) {
1520 DEBUG(dbgs() << "LAA: Found a write-only loop!\n");
1525 // Build dependence sets and check whether we need a runtime pointer bounds
1527 Accesses.buildDependenceSets();
1529 // Find pointers with computable bounds. We are going to use this information
1530 // to place a runtime bound check.
1531 bool CanDoRTIfNeeded =
1532 Accesses.canCheckPtrAtRT(PtrRtChecking, PSE.getSE(), TheLoop, Strides);
1533 if (!CanDoRTIfNeeded) {
1534 emitAnalysis(LoopAccessReport() << "cannot identify array bounds");
1535 DEBUG(dbgs() << "LAA: We can't vectorize because we can't find "
1536 << "the array bounds.\n");
1541 DEBUG(dbgs() << "LAA: We can perform a memory runtime check if needed.\n");
1544 if (Accesses.isDependencyCheckNeeded()) {
1545 DEBUG(dbgs() << "LAA: Checking memory dependencies\n");
1546 CanVecMem = DepChecker.areDepsSafe(
1547 DependentAccesses, Accesses.getDependenciesToCheck(), Strides);
1548 MaxSafeDepDistBytes = DepChecker.getMaxSafeDepDistBytes();
1550 if (!CanVecMem && DepChecker.shouldRetryWithRuntimeCheck()) {
1551 DEBUG(dbgs() << "LAA: Retrying with memory checks\n");
1553 // Clear the dependency checks. We assume they are not needed.
1554 Accesses.resetDepChecks(DepChecker);
1556 PtrRtChecking.reset();
1557 PtrRtChecking.Need = true;
1559 auto *SE = PSE.getSE();
1561 Accesses.canCheckPtrAtRT(PtrRtChecking, SE, TheLoop, Strides, true);
1563 // Check that we found the bounds for the pointer.
1564 if (!CanDoRTIfNeeded) {
1565 emitAnalysis(LoopAccessReport()
1566 << "cannot check memory dependencies at runtime");
1567 DEBUG(dbgs() << "LAA: Can't vectorize with memory checks\n");
1577 DEBUG(dbgs() << "LAA: No unsafe dependent memory operations in loop. We"
1578 << (PtrRtChecking.Need ? "" : " don't")
1579 << " need runtime memory checks.\n");
1581 emitAnalysis(LoopAccessReport() <<
1582 "unsafe dependent memory operations in loop");
1583 DEBUG(dbgs() << "LAA: unsafe dependent memory operations in loop\n");
1587 bool LoopAccessInfo::blockNeedsPredication(BasicBlock *BB, Loop *TheLoop,
1588 DominatorTree *DT) {
1589 assert(TheLoop->contains(BB) && "Unknown block used");
1591 // Blocks that do not dominate the latch need predication.
1592 BasicBlock* Latch = TheLoop->getLoopLatch();
1593 return !DT->dominates(BB, Latch);
1596 void LoopAccessInfo::emitAnalysis(LoopAccessReport &Message) {
1597 assert(!Report && "Multiple reports generated");
1601 bool LoopAccessInfo::isUniform(Value *V) const {
1602 return (PSE.getSE()->isLoopInvariant(PSE.getSE()->getSCEV(V), TheLoop));
1605 // FIXME: this function is currently a duplicate of the one in
1606 // LoopVectorize.cpp.
1607 static Instruction *getFirstInst(Instruction *FirstInst, Value *V,
1611 if (Instruction *I = dyn_cast<Instruction>(V))
1612 return I->getParent() == Loc->getParent() ? I : nullptr;
1617 /// \brief IR Values for the lower and upper bounds of a pointer evolution. We
1618 /// need to use value-handles because SCEV expansion can invalidate previously
1619 /// expanded values. Thus expansion of a pointer can invalidate the bounds for
1621 struct PointerBounds {
1622 TrackingVH<Value> Start;
1623 TrackingVH<Value> End;
1625 } // end anonymous namespace
1627 /// \brief Expand code for the lower and upper bound of the pointer group \p CG
1628 /// in \p TheLoop. \return the values for the bounds.
1629 static PointerBounds
1630 expandBounds(const RuntimePointerChecking::CheckingPtrGroup *CG, Loop *TheLoop,
1631 Instruction *Loc, SCEVExpander &Exp, ScalarEvolution *SE,
1632 const RuntimePointerChecking &PtrRtChecking) {
1633 Value *Ptr = PtrRtChecking.Pointers[CG->Members[0]].PointerValue;
1634 const SCEV *Sc = SE->getSCEV(Ptr);
1636 if (SE->isLoopInvariant(Sc, TheLoop)) {
1637 DEBUG(dbgs() << "LAA: Adding RT check for a loop invariant ptr:" << *Ptr
1641 unsigned AS = Ptr->getType()->getPointerAddressSpace();
1642 LLVMContext &Ctx = Loc->getContext();
1644 // Use this type for pointer arithmetic.
1645 Type *PtrArithTy = Type::getInt8PtrTy(Ctx, AS);
1646 Value *Start = nullptr, *End = nullptr;
1648 DEBUG(dbgs() << "LAA: Adding RT check for range:\n");
1649 Start = Exp.expandCodeFor(CG->Low, PtrArithTy, Loc);
1650 End = Exp.expandCodeFor(CG->High, PtrArithTy, Loc);
1651 DEBUG(dbgs() << "Start: " << *CG->Low << " End: " << *CG->High << "\n");
1652 return {Start, End};
1656 /// \brief Turns a collection of checks into a collection of expanded upper and
1657 /// lower bounds for both pointers in the check.
1658 static SmallVector<std::pair<PointerBounds, PointerBounds>, 4> expandBounds(
1659 const SmallVectorImpl<RuntimePointerChecking::PointerCheck> &PointerChecks,
1660 Loop *L, Instruction *Loc, ScalarEvolution *SE, SCEVExpander &Exp,
1661 const RuntimePointerChecking &PtrRtChecking) {
1662 SmallVector<std::pair<PointerBounds, PointerBounds>, 4> ChecksWithBounds;
1664 // Here we're relying on the SCEV Expander's cache to only emit code for the
1665 // same bounds once.
1667 PointerChecks.begin(), PointerChecks.end(),
1668 std::back_inserter(ChecksWithBounds),
1669 [&](const RuntimePointerChecking::PointerCheck &Check) {
1671 First = expandBounds(Check.first, L, Loc, Exp, SE, PtrRtChecking),
1672 Second = expandBounds(Check.second, L, Loc, Exp, SE, PtrRtChecking);
1673 return std::make_pair(First, Second);
1676 return ChecksWithBounds;
1679 std::pair<Instruction *, Instruction *> LoopAccessInfo::addRuntimeChecks(
1681 const SmallVectorImpl<RuntimePointerChecking::PointerCheck> &PointerChecks)
1683 auto *SE = PSE.getSE();
1684 SCEVExpander Exp(*SE, DL, "induction");
1685 auto ExpandedChecks =
1686 expandBounds(PointerChecks, TheLoop, Loc, SE, Exp, PtrRtChecking);
1688 LLVMContext &Ctx = Loc->getContext();
1689 Instruction *FirstInst = nullptr;
1690 IRBuilder<> ChkBuilder(Loc);
1691 // Our instructions might fold to a constant.
1692 Value *MemoryRuntimeCheck = nullptr;
1694 for (const auto &Check : ExpandedChecks) {
1695 const PointerBounds &A = Check.first, &B = Check.second;
1696 // Check if two pointers (A and B) conflict where conflict is computed as:
1697 // start(A) <= end(B) && start(B) <= end(A)
1698 unsigned AS0 = A.Start->getType()->getPointerAddressSpace();
1699 unsigned AS1 = B.Start->getType()->getPointerAddressSpace();
1701 assert((AS0 == B.End->getType()->getPointerAddressSpace()) &&
1702 (AS1 == A.End->getType()->getPointerAddressSpace()) &&
1703 "Trying to bounds check pointers with different address spaces");
1705 Type *PtrArithTy0 = Type::getInt8PtrTy(Ctx, AS0);
1706 Type *PtrArithTy1 = Type::getInt8PtrTy(Ctx, AS1);
1708 Value *Start0 = ChkBuilder.CreateBitCast(A.Start, PtrArithTy0, "bc");
1709 Value *Start1 = ChkBuilder.CreateBitCast(B.Start, PtrArithTy1, "bc");
1710 Value *End0 = ChkBuilder.CreateBitCast(A.End, PtrArithTy1, "bc");
1711 Value *End1 = ChkBuilder.CreateBitCast(B.End, PtrArithTy0, "bc");
1713 Value *Cmp0 = ChkBuilder.CreateICmpULE(Start0, End1, "bound0");
1714 FirstInst = getFirstInst(FirstInst, Cmp0, Loc);
1715 Value *Cmp1 = ChkBuilder.CreateICmpULE(Start1, End0, "bound1");
1716 FirstInst = getFirstInst(FirstInst, Cmp1, Loc);
1717 Value *IsConflict = ChkBuilder.CreateAnd(Cmp0, Cmp1, "found.conflict");
1718 FirstInst = getFirstInst(FirstInst, IsConflict, Loc);
1719 if (MemoryRuntimeCheck) {
1721 ChkBuilder.CreateOr(MemoryRuntimeCheck, IsConflict, "conflict.rdx");
1722 FirstInst = getFirstInst(FirstInst, IsConflict, Loc);
1724 MemoryRuntimeCheck = IsConflict;
1727 if (!MemoryRuntimeCheck)
1728 return std::make_pair(nullptr, nullptr);
1730 // We have to do this trickery because the IRBuilder might fold the check to a
1731 // constant expression in which case there is no Instruction anchored in a
1733 Instruction *Check = BinaryOperator::CreateAnd(MemoryRuntimeCheck,
1734 ConstantInt::getTrue(Ctx));
1735 ChkBuilder.Insert(Check, "memcheck.conflict");
1736 FirstInst = getFirstInst(FirstInst, Check, Loc);
1737 return std::make_pair(FirstInst, Check);
1740 std::pair<Instruction *, Instruction *>
1741 LoopAccessInfo::addRuntimeChecks(Instruction *Loc) const {
1742 if (!PtrRtChecking.Need)
1743 return std::make_pair(nullptr, nullptr);
1745 return addRuntimeChecks(Loc, PtrRtChecking.getChecks());
1748 LoopAccessInfo::LoopAccessInfo(Loop *L, ScalarEvolution *SE,
1749 const DataLayout &DL,
1750 const TargetLibraryInfo *TLI, AliasAnalysis *AA,
1751 DominatorTree *DT, LoopInfo *LI,
1752 const ValueToValueMap &Strides)
1753 : PSE(*SE), PtrRtChecking(SE), DepChecker(PSE, L), TheLoop(L), DL(DL),
1754 TLI(TLI), AA(AA), DT(DT), LI(LI), NumLoads(0), NumStores(0),
1755 MaxSafeDepDistBytes(-1U), CanVecMem(false),
1756 StoreToLoopInvariantAddress(false) {
1757 if (canAnalyzeLoop())
1758 analyzeLoop(Strides);
1761 void LoopAccessInfo::print(raw_ostream &OS, unsigned Depth) const {
1763 if (PtrRtChecking.Need)
1764 OS.indent(Depth) << "Memory dependences are safe with run-time checks\n";
1766 OS.indent(Depth) << "Memory dependences are safe\n";
1770 OS.indent(Depth) << "Report: " << Report->str() << "\n";
1772 if (auto *Dependences = DepChecker.getDependences()) {
1773 OS.indent(Depth) << "Dependences:\n";
1774 for (auto &Dep : *Dependences) {
1775 Dep.print(OS, Depth + 2, DepChecker.getMemoryInstructions());
1779 OS.indent(Depth) << "Too many dependences, not recorded\n";
1781 // List the pair of accesses need run-time checks to prove independence.
1782 PtrRtChecking.print(OS, Depth);
1785 OS.indent(Depth) << "Store to invariant address was "
1786 << (StoreToLoopInvariantAddress ? "" : "not ")
1787 << "found in loop.\n";
1789 OS.indent(Depth) << "SCEV assumptions:\n";
1790 PSE.getUnionPredicate().print(OS, Depth);
1793 const LoopAccessInfo &
1794 LoopAccessAnalysis::getInfo(Loop *L, const ValueToValueMap &Strides) {
1795 auto &LAI = LoopAccessInfoMap[L];
1798 assert((!LAI || LAI->NumSymbolicStrides == Strides.size()) &&
1799 "Symbolic strides changed for loop");
1803 const DataLayout &DL = L->getHeader()->getModule()->getDataLayout();
1805 llvm::make_unique<LoopAccessInfo>(L, SE, DL, TLI, AA, DT, LI, Strides);
1807 LAI->NumSymbolicStrides = Strides.size();
1813 void LoopAccessAnalysis::print(raw_ostream &OS, const Module *M) const {
1814 LoopAccessAnalysis &LAA = *const_cast<LoopAccessAnalysis *>(this);
1816 ValueToValueMap NoSymbolicStrides;
1818 for (Loop *TopLevelLoop : *LI)
1819 for (Loop *L : depth_first(TopLevelLoop)) {
1820 OS.indent(2) << L->getHeader()->getName() << ":\n";
1821 auto &LAI = LAA.getInfo(L, NoSymbolicStrides);
1826 bool LoopAccessAnalysis::runOnFunction(Function &F) {
1827 SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE();
1828 auto *TLIP = getAnalysisIfAvailable<TargetLibraryInfoWrapperPass>();
1829 TLI = TLIP ? &TLIP->getTLI() : nullptr;
1830 AA = &getAnalysis<AAResultsWrapperPass>().getAAResults();
1831 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1832 LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
1837 void LoopAccessAnalysis::getAnalysisUsage(AnalysisUsage &AU) const {
1838 AU.addRequired<ScalarEvolutionWrapperPass>();
1839 AU.addRequired<AAResultsWrapperPass>();
1840 AU.addRequired<DominatorTreeWrapperPass>();
1841 AU.addRequired<LoopInfoWrapperPass>();
1843 AU.setPreservesAll();
1846 char LoopAccessAnalysis::ID = 0;
1847 static const char laa_name[] = "Loop Access Analysis";
1848 #define LAA_NAME "loop-accesses"
1850 INITIALIZE_PASS_BEGIN(LoopAccessAnalysis, LAA_NAME, laa_name, false, true)
1851 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)
1852 INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass)
1853 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
1854 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
1855 INITIALIZE_PASS_END(LoopAccessAnalysis, LAA_NAME, laa_name, false, true)
1858 Pass *createLAAPass() {
1859 return new LoopAccessAnalysis();