1 //===- LoopAccessAnalysis.cpp - Loop Access Analysis Implementation --------==//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // The implementation for the loop memory dependence that was originally
11 // developed for the loop vectorizer.
13 //===----------------------------------------------------------------------===//
15 #include "llvm/Analysis/LoopAccessAnalysis.h"
16 #include "llvm/Analysis/LoopInfo.h"
17 #include "llvm/Analysis/ScalarEvolutionExpander.h"
18 #include "llvm/Analysis/TargetLibraryInfo.h"
19 #include "llvm/Analysis/ValueTracking.h"
20 #include "llvm/IR/DiagnosticInfo.h"
21 #include "llvm/IR/Dominators.h"
22 #include "llvm/IR/IRBuilder.h"
23 #include "llvm/Support/Debug.h"
24 #include "llvm/Support/raw_ostream.h"
25 #include "llvm/Analysis/VectorUtils.h"
28 #define DEBUG_TYPE "loop-accesses"
30 static cl::opt<unsigned, true>
31 VectorizationFactor("force-vector-width", cl::Hidden,
32 cl::desc("Sets the SIMD width. Zero is autoselect."),
33 cl::location(VectorizerParams::VectorizationFactor));
34 unsigned VectorizerParams::VectorizationFactor;
36 static cl::opt<unsigned, true>
37 VectorizationInterleave("force-vector-interleave", cl::Hidden,
38 cl::desc("Sets the vectorization interleave count. "
39 "Zero is autoselect."),
41 VectorizerParams::VectorizationInterleave));
42 unsigned VectorizerParams::VectorizationInterleave;
44 static cl::opt<unsigned, true> RuntimeMemoryCheckThreshold(
45 "runtime-memory-check-threshold", cl::Hidden,
46 cl::desc("When performing memory disambiguation checks at runtime do not "
47 "generate more than this number of comparisons (default = 8)."),
48 cl::location(VectorizerParams::RuntimeMemoryCheckThreshold), cl::init(8));
49 unsigned VectorizerParams::RuntimeMemoryCheckThreshold;
51 /// \brief The maximum iterations used to merge memory checks
52 static cl::opt<unsigned> MemoryCheckMergeThreshold(
53 "memory-check-merge-threshold", cl::Hidden,
54 cl::desc("Maximum number of comparisons done when trying to merge "
55 "runtime memory checks. (default = 100)"),
58 /// Maximum SIMD width.
59 const unsigned VectorizerParams::MaxVectorWidth = 64;
61 /// \brief We collect interesting dependences up to this threshold.
62 static cl::opt<unsigned> MaxInterestingDependence(
63 "max-interesting-dependences", cl::Hidden,
64 cl::desc("Maximum number of interesting dependences collected by "
65 "loop-access analysis (default = 100)"),
68 bool VectorizerParams::isInterleaveForced() {
69 return ::VectorizationInterleave.getNumOccurrences() > 0;
72 void LoopAccessReport::emitAnalysis(const LoopAccessReport &Message,
73 const Function *TheFunction,
75 const char *PassName) {
76 DebugLoc DL = TheLoop->getStartLoc();
77 if (const Instruction *I = Message.getInstr())
78 DL = I->getDebugLoc();
79 emitOptimizationRemarkAnalysis(TheFunction->getContext(), PassName,
80 *TheFunction, DL, Message.str());
83 Value *llvm::stripIntegerCast(Value *V) {
84 if (CastInst *CI = dyn_cast<CastInst>(V))
85 if (CI->getOperand(0)->getType()->isIntegerTy())
86 return CI->getOperand(0);
90 const SCEV *llvm::replaceSymbolicStrideSCEV(ScalarEvolution *SE,
91 const ValueToValueMap &PtrToStride,
92 Value *Ptr, Value *OrigPtr) {
94 const SCEV *OrigSCEV = SE->getSCEV(Ptr);
96 // If there is an entry in the map return the SCEV of the pointer with the
97 // symbolic stride replaced by one.
98 ValueToValueMap::const_iterator SI =
99 PtrToStride.find(OrigPtr ? OrigPtr : Ptr);
100 if (SI != PtrToStride.end()) {
101 Value *StrideVal = SI->second;
104 StrideVal = stripIntegerCast(StrideVal);
106 // Replace symbolic stride by one.
107 Value *One = ConstantInt::get(StrideVal->getType(), 1);
108 ValueToValueMap RewriteMap;
109 RewriteMap[StrideVal] = One;
112 SCEVParameterRewriter::rewrite(OrigSCEV, *SE, RewriteMap, true);
113 DEBUG(dbgs() << "LAA: Replacing SCEV: " << *OrigSCEV << " by: " << *ByOne
118 // Otherwise, just return the SCEV of the original pointer.
119 return SE->getSCEV(Ptr);
122 void RuntimePointerChecking::insert(Loop *Lp, Value *Ptr, bool WritePtr,
123 unsigned DepSetId, unsigned ASId,
124 const ValueToValueMap &Strides) {
125 // Get the stride replaced scev.
126 const SCEV *Sc = replaceSymbolicStrideSCEV(SE, Strides, Ptr);
127 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Sc);
128 assert(AR && "Invalid addrec expression");
129 const SCEV *Ex = SE->getBackedgeTakenCount(Lp);
131 const SCEV *ScStart = AR->getStart();
132 const SCEV *ScEnd = AR->evaluateAtIteration(Ex, *SE);
133 const SCEV *Step = AR->getStepRecurrence(*SE);
135 // For expressions with negative step, the upper bound is ScStart and the
136 // lower bound is ScEnd.
137 if (const SCEVConstant *CStep = dyn_cast<const SCEVConstant>(Step)) {
138 if (CStep->getValue()->isNegative())
139 std::swap(ScStart, ScEnd);
141 // Fallback case: the step is not constant, but the we can still
142 // get the upper and lower bounds of the interval by using min/max
144 ScStart = SE->getUMinExpr(ScStart, ScEnd);
145 ScEnd = SE->getUMaxExpr(AR->getStart(), ScEnd);
148 Pointers.emplace_back(Ptr, ScStart, ScEnd, WritePtr, DepSetId, ASId, Sc);
151 SmallVector<RuntimePointerChecking::PointerCheck, 4>
152 RuntimePointerChecking::generateChecks(
153 const SmallVectorImpl<int> *PtrPartition) const {
154 SmallVector<PointerCheck, 4> Checks;
156 for (unsigned i = 0; i < CheckingGroups.size(); ++i) {
157 for (unsigned j = i + 1; j < CheckingGroups.size(); ++j) {
158 const RuntimePointerChecking::CheckingPtrGroup &CGI = CheckingGroups[i];
159 const RuntimePointerChecking::CheckingPtrGroup &CGJ = CheckingGroups[j];
161 if (needsChecking(CGI, CGJ, PtrPartition))
162 Checks.push_back(std::make_pair(&CGI, &CGJ));
168 bool RuntimePointerChecking::needsChecking(
169 const CheckingPtrGroup &M, const CheckingPtrGroup &N,
170 const SmallVectorImpl<int> *PtrPartition) const {
171 for (unsigned I = 0, EI = M.Members.size(); EI != I; ++I)
172 for (unsigned J = 0, EJ = N.Members.size(); EJ != J; ++J)
173 if (needsChecking(M.Members[I], N.Members[J], PtrPartition))
178 /// Compare \p I and \p J and return the minimum.
179 /// Return nullptr in case we couldn't find an answer.
180 static const SCEV *getMinFromExprs(const SCEV *I, const SCEV *J,
181 ScalarEvolution *SE) {
182 const SCEV *Diff = SE->getMinusSCEV(J, I);
183 const SCEVConstant *C = dyn_cast<const SCEVConstant>(Diff);
187 if (C->getValue()->isNegative())
192 bool RuntimePointerChecking::CheckingPtrGroup::addPointer(unsigned Index) {
193 const SCEV *Start = RtCheck.Pointers[Index].Start;
194 const SCEV *End = RtCheck.Pointers[Index].End;
196 // Compare the starts and ends with the known minimum and maximum
197 // of this set. We need to know how we compare against the min/max
198 // of the set in order to be able to emit memchecks.
199 const SCEV *Min0 = getMinFromExprs(Start, Low, RtCheck.SE);
203 const SCEV *Min1 = getMinFromExprs(End, High, RtCheck.SE);
207 // Update the low bound expression if we've found a new min value.
211 // Update the high bound expression if we've found a new max value.
215 Members.push_back(Index);
219 void RuntimePointerChecking::groupChecks(
220 MemoryDepChecker::DepCandidates &DepCands, bool UseDependencies) {
221 // We build the groups from dependency candidates equivalence classes
223 // - We know that pointers in the same equivalence class share
224 // the same underlying object and therefore there is a chance
225 // that we can compare pointers
226 // - We wouldn't be able to merge two pointers for which we need
227 // to emit a memcheck. The classes in DepCands are already
228 // conveniently built such that no two pointers in the same
229 // class need checking against each other.
231 // We use the following (greedy) algorithm to construct the groups
232 // For every pointer in the equivalence class:
233 // For each existing group:
234 // - if the difference between this pointer and the min/max bounds
235 // of the group is a constant, then make the pointer part of the
236 // group and update the min/max bounds of that group as required.
238 CheckingGroups.clear();
240 // If we don't have the dependency partitions, construct a new
241 // checking pointer group for each pointer.
242 if (!UseDependencies) {
243 for (unsigned I = 0; I < Pointers.size(); ++I)
244 CheckingGroups.push_back(CheckingPtrGroup(I, *this));
248 unsigned TotalComparisons = 0;
250 DenseMap<Value *, unsigned> PositionMap;
251 for (unsigned Index = 0; Index < Pointers.size(); ++Index)
252 PositionMap[Pointers[Index].PointerValue] = Index;
254 // We need to keep track of what pointers we've already seen so we
255 // don't process them twice.
256 SmallSet<unsigned, 2> Seen;
258 // Go through all equivalence classes, get the the "pointer check groups"
259 // and add them to the overall solution. We use the order in which accesses
260 // appear in 'Pointers' to enforce determinism.
261 for (unsigned I = 0; I < Pointers.size(); ++I) {
262 // We've seen this pointer before, and therefore already processed
263 // its equivalence class.
267 MemoryDepChecker::MemAccessInfo Access(Pointers[I].PointerValue,
268 Pointers[I].IsWritePtr);
270 SmallVector<CheckingPtrGroup, 2> Groups;
271 auto LeaderI = DepCands.findValue(DepCands.getLeaderValue(Access));
273 // Because DepCands is constructed by visiting accesses in the order in
274 // which they appear in alias sets (which is deterministic) and the
275 // iteration order within an equivalence class member is only dependent on
276 // the order in which unions and insertions are performed on the
277 // equivalence class, the iteration order is deterministic.
278 for (auto MI = DepCands.member_begin(LeaderI), ME = DepCands.member_end();
280 unsigned Pointer = PositionMap[MI->getPointer()];
282 // Mark this pointer as seen.
283 Seen.insert(Pointer);
285 // Go through all the existing sets and see if we can find one
286 // which can include this pointer.
287 for (CheckingPtrGroup &Group : Groups) {
288 // Don't perform more than a certain amount of comparisons.
289 // This should limit the cost of grouping the pointers to something
290 // reasonable. If we do end up hitting this threshold, the algorithm
291 // will create separate groups for all remaining pointers.
292 if (TotalComparisons > MemoryCheckMergeThreshold)
297 if (Group.addPointer(Pointer)) {
304 // We couldn't add this pointer to any existing set or the threshold
305 // for the number of comparisons has been reached. Create a new group
306 // to hold the current pointer.
307 Groups.push_back(CheckingPtrGroup(Pointer, *this));
310 // We've computed the grouped checks for this partition.
311 // Save the results and continue with the next one.
312 std::copy(Groups.begin(), Groups.end(), std::back_inserter(CheckingGroups));
316 bool RuntimePointerChecking::arePointersInSamePartition(
317 const SmallVectorImpl<int> &PtrToPartition, unsigned PtrIdx1,
319 return (PtrToPartition[PtrIdx1] != -1 &&
320 PtrToPartition[PtrIdx1] == PtrToPartition[PtrIdx2]);
323 bool RuntimePointerChecking::needsChecking(
324 unsigned I, unsigned J, const SmallVectorImpl<int> *PtrPartition) const {
325 const PointerInfo &PointerI = Pointers[I];
326 const PointerInfo &PointerJ = Pointers[J];
328 // No need to check if two readonly pointers intersect.
329 if (!PointerI.IsWritePtr && !PointerJ.IsWritePtr)
332 // Only need to check pointers between two different dependency sets.
333 if (PointerI.DependencySetId == PointerJ.DependencySetId)
336 // Only need to check pointers in the same alias set.
337 if (PointerI.AliasSetId != PointerJ.AliasSetId)
340 // If PtrPartition is set omit checks between pointers of the same partition.
341 if (PtrPartition && arePointersInSamePartition(*PtrPartition, I, J))
347 void RuntimePointerChecking::print(
348 raw_ostream &OS, unsigned Depth,
349 const SmallVectorImpl<int> *PtrPartition) const {
351 OS.indent(Depth) << "Run-time memory checks:\n";
354 for (unsigned I = 0; I < CheckingGroups.size(); ++I)
355 for (unsigned J = I + 1; J < CheckingGroups.size(); ++J)
356 if (needsChecking(CheckingGroups[I], CheckingGroups[J], PtrPartition)) {
357 OS.indent(Depth) << "Check " << N++ << ":\n";
358 OS.indent(Depth + 2) << "Comparing group " << I << ":\n";
360 for (unsigned K = 0; K < CheckingGroups[I].Members.size(); ++K) {
362 << *Pointers[CheckingGroups[I].Members[K]].PointerValue << "\n";
364 OS << " (Partition: "
365 << (*PtrPartition)[CheckingGroups[I].Members[K]] << ")"
369 OS.indent(Depth + 2) << "Against group " << J << ":\n";
371 for (unsigned K = 0; K < CheckingGroups[J].Members.size(); ++K) {
373 << *Pointers[CheckingGroups[J].Members[K]].PointerValue << "\n";
375 OS << " (Partition: "
376 << (*PtrPartition)[CheckingGroups[J].Members[K]] << ")"
381 OS.indent(Depth) << "Grouped accesses:\n";
382 for (unsigned I = 0; I < CheckingGroups.size(); ++I) {
383 OS.indent(Depth + 2) << "Group " << I << ":\n";
384 OS.indent(Depth + 4) << "(Low: " << *CheckingGroups[I].Low
385 << " High: " << *CheckingGroups[I].High << ")\n";
386 for (unsigned J = 0; J < CheckingGroups[I].Members.size(); ++J) {
387 OS.indent(Depth + 6) << "Member: "
388 << *Pointers[CheckingGroups[I].Members[J]].Expr
394 unsigned RuntimePointerChecking::getNumberOfChecks(
395 const SmallVectorImpl<int> *PtrPartition) const {
397 unsigned NumPartitions = CheckingGroups.size();
398 unsigned CheckCount = 0;
400 for (unsigned I = 0; I < NumPartitions; ++I)
401 for (unsigned J = I + 1; J < NumPartitions; ++J)
402 if (needsChecking(CheckingGroups[I], CheckingGroups[J], PtrPartition))
407 bool RuntimePointerChecking::needsAnyChecking(
408 const SmallVectorImpl<int> *PtrPartition) const {
409 unsigned NumPointers = Pointers.size();
411 for (unsigned I = 0; I < NumPointers; ++I)
412 for (unsigned J = I + 1; J < NumPointers; ++J)
413 if (needsChecking(I, J, PtrPartition))
419 /// \brief Analyses memory accesses in a loop.
421 /// Checks whether run time pointer checks are needed and builds sets for data
422 /// dependence checking.
423 class AccessAnalysis {
425 /// \brief Read or write access location.
426 typedef PointerIntPair<Value *, 1, bool> MemAccessInfo;
427 typedef SmallPtrSet<MemAccessInfo, 8> MemAccessInfoSet;
429 AccessAnalysis(const DataLayout &Dl, AliasAnalysis *AA, LoopInfo *LI,
430 MemoryDepChecker::DepCandidates &DA)
431 : DL(Dl), AST(*AA), LI(LI), DepCands(DA),
432 IsRTCheckAnalysisNeeded(false) {}
434 /// \brief Register a load and whether it is only read from.
435 void addLoad(MemoryLocation &Loc, bool IsReadOnly) {
436 Value *Ptr = const_cast<Value*>(Loc.Ptr);
437 AST.add(Ptr, MemoryLocation::UnknownSize, Loc.AATags);
438 Accesses.insert(MemAccessInfo(Ptr, false));
440 ReadOnlyPtr.insert(Ptr);
443 /// \brief Register a store.
444 void addStore(MemoryLocation &Loc) {
445 Value *Ptr = const_cast<Value*>(Loc.Ptr);
446 AST.add(Ptr, MemoryLocation::UnknownSize, Loc.AATags);
447 Accesses.insert(MemAccessInfo(Ptr, true));
450 /// \brief Check whether we can check the pointers at runtime for
451 /// non-intersection.
453 /// Returns true if we need no check or if we do and we can generate them
454 /// (i.e. the pointers have computable bounds).
455 bool canCheckPtrAtRT(RuntimePointerChecking &RtCheck, ScalarEvolution *SE,
456 Loop *TheLoop, const ValueToValueMap &Strides,
457 bool ShouldCheckStride = false);
459 /// \brief Goes over all memory accesses, checks whether a RT check is needed
460 /// and builds sets of dependent accesses.
461 void buildDependenceSets() {
462 processMemAccesses();
465 /// \brief Initial processing of memory accesses determined that we need to
466 /// perform dependency checking.
468 /// Note that this can later be cleared if we retry memcheck analysis without
469 /// dependency checking (i.e. ShouldRetryWithRuntimeCheck).
470 bool isDependencyCheckNeeded() { return !CheckDeps.empty(); }
472 /// We decided that no dependence analysis would be used. Reset the state.
473 void resetDepChecks(MemoryDepChecker &DepChecker) {
475 DepChecker.clearInterestingDependences();
478 MemAccessInfoSet &getDependenciesToCheck() { return CheckDeps; }
481 typedef SetVector<MemAccessInfo> PtrAccessSet;
483 /// \brief Go over all memory access and check whether runtime pointer checks
484 /// are needed and build sets of dependency check candidates.
485 void processMemAccesses();
487 /// Set of all accesses.
488 PtrAccessSet Accesses;
490 const DataLayout &DL;
492 /// Set of accesses that need a further dependence check.
493 MemAccessInfoSet CheckDeps;
495 /// Set of pointers that are read only.
496 SmallPtrSet<Value*, 16> ReadOnlyPtr;
498 /// An alias set tracker to partition the access set by underlying object and
499 //intrinsic property (such as TBAA metadata).
504 /// Sets of potentially dependent accesses - members of one set share an
505 /// underlying pointer. The set "CheckDeps" identfies which sets really need a
506 /// dependence check.
507 MemoryDepChecker::DepCandidates &DepCands;
509 /// \brief Initial processing of memory accesses determined that we may need
510 /// to add memchecks. Perform the analysis to determine the necessary checks.
512 /// Note that, this is different from isDependencyCheckNeeded. When we retry
513 /// memcheck analysis without dependency checking
514 /// (i.e. ShouldRetryWithRuntimeCheck), isDependencyCheckNeeded is cleared
515 /// while this remains set if we have potentially dependent accesses.
516 bool IsRTCheckAnalysisNeeded;
519 } // end anonymous namespace
521 /// \brief Check whether a pointer can participate in a runtime bounds check.
522 static bool hasComputableBounds(ScalarEvolution *SE,
523 const ValueToValueMap &Strides, Value *Ptr) {
524 const SCEV *PtrScev = replaceSymbolicStrideSCEV(SE, Strides, Ptr);
525 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
529 return AR->isAffine();
532 bool AccessAnalysis::canCheckPtrAtRT(RuntimePointerChecking &RtCheck,
533 ScalarEvolution *SE, Loop *TheLoop,
534 const ValueToValueMap &StridesMap,
535 bool ShouldCheckStride) {
536 // Find pointers with computable bounds. We are going to use this information
537 // to place a runtime bound check.
540 bool NeedRTCheck = false;
541 if (!IsRTCheckAnalysisNeeded) return true;
543 bool IsDepCheckNeeded = isDependencyCheckNeeded();
545 // We assign a consecutive id to access from different alias sets.
546 // Accesses between different groups doesn't need to be checked.
548 for (auto &AS : AST) {
549 int NumReadPtrChecks = 0;
550 int NumWritePtrChecks = 0;
552 // We assign consecutive id to access from different dependence sets.
553 // Accesses within the same set don't need a runtime check.
554 unsigned RunningDepId = 1;
555 DenseMap<Value *, unsigned> DepSetId;
558 Value *Ptr = A.getValue();
559 bool IsWrite = Accesses.count(MemAccessInfo(Ptr, true));
560 MemAccessInfo Access(Ptr, IsWrite);
567 if (hasComputableBounds(SE, StridesMap, Ptr) &&
568 // When we run after a failing dependency check we have to make sure
569 // we don't have wrapping pointers.
570 (!ShouldCheckStride ||
571 isStridedPtr(SE, Ptr, TheLoop, StridesMap) == 1)) {
572 // The id of the dependence set.
575 if (IsDepCheckNeeded) {
576 Value *Leader = DepCands.getLeaderValue(Access).getPointer();
577 unsigned &LeaderId = DepSetId[Leader];
579 LeaderId = RunningDepId++;
582 // Each access has its own dependence set.
583 DepId = RunningDepId++;
585 RtCheck.insert(TheLoop, Ptr, IsWrite, DepId, ASId, StridesMap);
587 DEBUG(dbgs() << "LAA: Found a runtime check ptr:" << *Ptr << '\n');
589 DEBUG(dbgs() << "LAA: Can't find bounds for ptr:" << *Ptr << '\n');
594 // If we have at least two writes or one write and a read then we need to
595 // check them. But there is no need to checks if there is only one
596 // dependence set for this alias set.
598 // Note that this function computes CanDoRT and NeedRTCheck independently.
599 // For example CanDoRT=false, NeedRTCheck=false means that we have a pointer
600 // for which we couldn't find the bounds but we don't actually need to emit
601 // any checks so it does not matter.
602 if (!(IsDepCheckNeeded && CanDoRT && RunningDepId == 2))
603 NeedRTCheck |= (NumWritePtrChecks >= 2 || (NumReadPtrChecks >= 1 &&
604 NumWritePtrChecks >= 1));
609 // If the pointers that we would use for the bounds comparison have different
610 // address spaces, assume the values aren't directly comparable, so we can't
611 // use them for the runtime check. We also have to assume they could
612 // overlap. In the future there should be metadata for whether address spaces
614 unsigned NumPointers = RtCheck.Pointers.size();
615 for (unsigned i = 0; i < NumPointers; ++i) {
616 for (unsigned j = i + 1; j < NumPointers; ++j) {
617 // Only need to check pointers between two different dependency sets.
618 if (RtCheck.Pointers[i].DependencySetId ==
619 RtCheck.Pointers[j].DependencySetId)
621 // Only need to check pointers in the same alias set.
622 if (RtCheck.Pointers[i].AliasSetId != RtCheck.Pointers[j].AliasSetId)
625 Value *PtrI = RtCheck.Pointers[i].PointerValue;
626 Value *PtrJ = RtCheck.Pointers[j].PointerValue;
628 unsigned ASi = PtrI->getType()->getPointerAddressSpace();
629 unsigned ASj = PtrJ->getType()->getPointerAddressSpace();
631 DEBUG(dbgs() << "LAA: Runtime check would require comparison between"
632 " different address spaces\n");
638 if (NeedRTCheck && CanDoRT)
639 RtCheck.groupChecks(DepCands, IsDepCheckNeeded);
641 DEBUG(dbgs() << "LAA: We need to do " << RtCheck.getNumberOfChecks(nullptr)
642 << " pointer comparisons.\n");
644 RtCheck.Need = NeedRTCheck;
646 bool CanDoRTIfNeeded = !NeedRTCheck || CanDoRT;
647 if (!CanDoRTIfNeeded)
649 return CanDoRTIfNeeded;
652 void AccessAnalysis::processMemAccesses() {
653 // We process the set twice: first we process read-write pointers, last we
654 // process read-only pointers. This allows us to skip dependence tests for
655 // read-only pointers.
657 DEBUG(dbgs() << "LAA: Processing memory accesses...\n");
658 DEBUG(dbgs() << " AST: "; AST.dump());
659 DEBUG(dbgs() << "LAA: Accesses(" << Accesses.size() << "):\n");
661 for (auto A : Accesses)
662 dbgs() << "\t" << *A.getPointer() << " (" <<
663 (A.getInt() ? "write" : (ReadOnlyPtr.count(A.getPointer()) ?
664 "read-only" : "read")) << ")\n";
667 // The AliasSetTracker has nicely partitioned our pointers by metadata
668 // compatibility and potential for underlying-object overlap. As a result, we
669 // only need to check for potential pointer dependencies within each alias
671 for (auto &AS : AST) {
672 // Note that both the alias-set tracker and the alias sets themselves used
673 // linked lists internally and so the iteration order here is deterministic
674 // (matching the original instruction order within each set).
676 bool SetHasWrite = false;
678 // Map of pointers to last access encountered.
679 typedef DenseMap<Value*, MemAccessInfo> UnderlyingObjToAccessMap;
680 UnderlyingObjToAccessMap ObjToLastAccess;
682 // Set of access to check after all writes have been processed.
683 PtrAccessSet DeferredAccesses;
685 // Iterate over each alias set twice, once to process read/write pointers,
686 // and then to process read-only pointers.
687 for (int SetIteration = 0; SetIteration < 2; ++SetIteration) {
688 bool UseDeferred = SetIteration > 0;
689 PtrAccessSet &S = UseDeferred ? DeferredAccesses : Accesses;
692 Value *Ptr = AV.getValue();
694 // For a single memory access in AliasSetTracker, Accesses may contain
695 // both read and write, and they both need to be handled for CheckDeps.
697 if (AC.getPointer() != Ptr)
700 bool IsWrite = AC.getInt();
702 // If we're using the deferred access set, then it contains only
704 bool IsReadOnlyPtr = ReadOnlyPtr.count(Ptr) && !IsWrite;
705 if (UseDeferred && !IsReadOnlyPtr)
707 // Otherwise, the pointer must be in the PtrAccessSet, either as a
709 assert(((IsReadOnlyPtr && UseDeferred) || IsWrite ||
710 S.count(MemAccessInfo(Ptr, false))) &&
711 "Alias-set pointer not in the access set?");
713 MemAccessInfo Access(Ptr, IsWrite);
714 DepCands.insert(Access);
716 // Memorize read-only pointers for later processing and skip them in
717 // the first round (they need to be checked after we have seen all
718 // write pointers). Note: we also mark pointer that are not
719 // consecutive as "read-only" pointers (so that we check
720 // "a[b[i]] +="). Hence, we need the second check for "!IsWrite".
721 if (!UseDeferred && IsReadOnlyPtr) {
722 DeferredAccesses.insert(Access);
726 // If this is a write - check other reads and writes for conflicts. If
727 // this is a read only check other writes for conflicts (but only if
728 // there is no other write to the ptr - this is an optimization to
729 // catch "a[i] = a[i] + " without having to do a dependence check).
730 if ((IsWrite || IsReadOnlyPtr) && SetHasWrite) {
731 CheckDeps.insert(Access);
732 IsRTCheckAnalysisNeeded = true;
738 // Create sets of pointers connected by a shared alias set and
739 // underlying object.
740 typedef SmallVector<Value *, 16> ValueVector;
741 ValueVector TempObjects;
743 GetUnderlyingObjects(Ptr, TempObjects, DL, LI);
744 DEBUG(dbgs() << "Underlying objects for pointer " << *Ptr << "\n");
745 for (Value *UnderlyingObj : TempObjects) {
746 UnderlyingObjToAccessMap::iterator Prev =
747 ObjToLastAccess.find(UnderlyingObj);
748 if (Prev != ObjToLastAccess.end())
749 DepCands.unionSets(Access, Prev->second);
751 ObjToLastAccess[UnderlyingObj] = Access;
752 DEBUG(dbgs() << " " << *UnderlyingObj << "\n");
760 static bool isInBoundsGep(Value *Ptr) {
761 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr))
762 return GEP->isInBounds();
766 /// \brief Return true if an AddRec pointer \p Ptr is unsigned non-wrapping,
767 /// i.e. monotonically increasing/decreasing.
768 static bool isNoWrapAddRec(Value *Ptr, const SCEVAddRecExpr *AR,
769 ScalarEvolution *SE, const Loop *L) {
770 // FIXME: This should probably only return true for NUW.
771 if (AR->getNoWrapFlags(SCEV::NoWrapMask))
774 // Scalar evolution does not propagate the non-wrapping flags to values that
775 // are derived from a non-wrapping induction variable because non-wrapping
776 // could be flow-sensitive.
778 // Look through the potentially overflowing instruction to try to prove
779 // non-wrapping for the *specific* value of Ptr.
781 // The arithmetic implied by an inbounds GEP can't overflow.
782 auto *GEP = dyn_cast<GetElementPtrInst>(Ptr);
783 if (!GEP || !GEP->isInBounds())
786 // Make sure there is only one non-const index and analyze that.
787 Value *NonConstIndex = nullptr;
788 for (auto Index = GEP->idx_begin(); Index != GEP->idx_end(); ++Index)
789 if (!isa<ConstantInt>(*Index)) {
792 NonConstIndex = *Index;
795 // The recurrence is on the pointer, ignore for now.
798 // The index in GEP is signed. It is non-wrapping if it's derived from a NSW
799 // AddRec using a NSW operation.
800 if (auto *OBO = dyn_cast<OverflowingBinaryOperator>(NonConstIndex))
801 if (OBO->hasNoSignedWrap() &&
802 // Assume constant for other the operand so that the AddRec can be
804 isa<ConstantInt>(OBO->getOperand(1))) {
805 auto *OpScev = SE->getSCEV(OBO->getOperand(0));
807 if (auto *OpAR = dyn_cast<SCEVAddRecExpr>(OpScev))
808 return OpAR->getLoop() == L && OpAR->getNoWrapFlags(SCEV::FlagNSW);
814 /// \brief Check whether the access through \p Ptr has a constant stride.
815 int llvm::isStridedPtr(ScalarEvolution *SE, Value *Ptr, const Loop *Lp,
816 const ValueToValueMap &StridesMap) {
817 const Type *Ty = Ptr->getType();
818 assert(Ty->isPointerTy() && "Unexpected non-ptr");
820 // Make sure that the pointer does not point to aggregate types.
821 const PointerType *PtrTy = cast<PointerType>(Ty);
822 if (PtrTy->getElementType()->isAggregateType()) {
823 DEBUG(dbgs() << "LAA: Bad stride - Not a pointer to a scalar type"
828 const SCEV *PtrScev = replaceSymbolicStrideSCEV(SE, StridesMap, Ptr);
830 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
832 DEBUG(dbgs() << "LAA: Bad stride - Not an AddRecExpr pointer "
833 << *Ptr << " SCEV: " << *PtrScev << "\n");
837 // The accesss function must stride over the innermost loop.
838 if (Lp != AR->getLoop()) {
839 DEBUG(dbgs() << "LAA: Bad stride - Not striding over innermost loop " <<
840 *Ptr << " SCEV: " << *PtrScev << "\n");
843 // The address calculation must not wrap. Otherwise, a dependence could be
845 // An inbounds getelementptr that is a AddRec with a unit stride
846 // cannot wrap per definition. The unit stride requirement is checked later.
847 // An getelementptr without an inbounds attribute and unit stride would have
848 // to access the pointer value "0" which is undefined behavior in address
849 // space 0, therefore we can also vectorize this case.
850 bool IsInBoundsGEP = isInBoundsGep(Ptr);
851 bool IsNoWrapAddRec = isNoWrapAddRec(Ptr, AR, SE, Lp);
852 bool IsInAddressSpaceZero = PtrTy->getAddressSpace() == 0;
853 if (!IsNoWrapAddRec && !IsInBoundsGEP && !IsInAddressSpaceZero) {
854 DEBUG(dbgs() << "LAA: Bad stride - Pointer may wrap in the address space "
855 << *Ptr << " SCEV: " << *PtrScev << "\n");
859 // Check the step is constant.
860 const SCEV *Step = AR->getStepRecurrence(*SE);
862 // Calculate the pointer stride and check if it is constant.
863 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
865 DEBUG(dbgs() << "LAA: Bad stride - Not a constant strided " << *Ptr <<
866 " SCEV: " << *PtrScev << "\n");
870 auto &DL = Lp->getHeader()->getModule()->getDataLayout();
871 int64_t Size = DL.getTypeAllocSize(PtrTy->getElementType());
872 const APInt &APStepVal = C->getValue()->getValue();
874 // Huge step value - give up.
875 if (APStepVal.getBitWidth() > 64)
878 int64_t StepVal = APStepVal.getSExtValue();
881 int64_t Stride = StepVal / Size;
882 int64_t Rem = StepVal % Size;
886 // If the SCEV could wrap but we have an inbounds gep with a unit stride we
887 // know we can't "wrap around the address space". In case of address space
888 // zero we know that this won't happen without triggering undefined behavior.
889 if (!IsNoWrapAddRec && (IsInBoundsGEP || IsInAddressSpaceZero) &&
890 Stride != 1 && Stride != -1)
896 bool MemoryDepChecker::Dependence::isSafeForVectorization(DepType Type) {
900 case BackwardVectorizable:
904 case ForwardButPreventsForwarding:
906 case BackwardVectorizableButPreventsForwarding:
909 llvm_unreachable("unexpected DepType!");
912 bool MemoryDepChecker::Dependence::isInterestingDependence(DepType Type) {
918 case BackwardVectorizable:
920 case ForwardButPreventsForwarding:
922 case BackwardVectorizableButPreventsForwarding:
925 llvm_unreachable("unexpected DepType!");
928 bool MemoryDepChecker::Dependence::isPossiblyBackward() const {
932 case ForwardButPreventsForwarding:
936 case BackwardVectorizable:
938 case BackwardVectorizableButPreventsForwarding:
941 llvm_unreachable("unexpected DepType!");
944 bool MemoryDepChecker::couldPreventStoreLoadForward(unsigned Distance,
945 unsigned TypeByteSize) {
946 // If loads occur at a distance that is not a multiple of a feasible vector
947 // factor store-load forwarding does not take place.
948 // Positive dependences might cause troubles because vectorizing them might
949 // prevent store-load forwarding making vectorized code run a lot slower.
950 // a[i] = a[i-3] ^ a[i-8];
951 // The stores to a[i:i+1] don't align with the stores to a[i-3:i-2] and
952 // hence on your typical architecture store-load forwarding does not take
953 // place. Vectorizing in such cases does not make sense.
954 // Store-load forwarding distance.
955 const unsigned NumCyclesForStoreLoadThroughMemory = 8*TypeByteSize;
956 // Maximum vector factor.
957 unsigned MaxVFWithoutSLForwardIssues =
958 VectorizerParams::MaxVectorWidth * TypeByteSize;
959 if(MaxSafeDepDistBytes < MaxVFWithoutSLForwardIssues)
960 MaxVFWithoutSLForwardIssues = MaxSafeDepDistBytes;
962 for (unsigned vf = 2*TypeByteSize; vf <= MaxVFWithoutSLForwardIssues;
964 if (Distance % vf && Distance / vf < NumCyclesForStoreLoadThroughMemory) {
965 MaxVFWithoutSLForwardIssues = (vf >>=1);
970 if (MaxVFWithoutSLForwardIssues< 2*TypeByteSize) {
971 DEBUG(dbgs() << "LAA: Distance " << Distance <<
972 " that could cause a store-load forwarding conflict\n");
976 if (MaxVFWithoutSLForwardIssues < MaxSafeDepDistBytes &&
977 MaxVFWithoutSLForwardIssues !=
978 VectorizerParams::MaxVectorWidth * TypeByteSize)
979 MaxSafeDepDistBytes = MaxVFWithoutSLForwardIssues;
983 /// \brief Check the dependence for two accesses with the same stride \p Stride.
984 /// \p Distance is the positive distance and \p TypeByteSize is type size in
987 /// \returns true if they are independent.
988 static bool areStridedAccessesIndependent(unsigned Distance, unsigned Stride,
989 unsigned TypeByteSize) {
990 assert(Stride > 1 && "The stride must be greater than 1");
991 assert(TypeByteSize > 0 && "The type size in byte must be non-zero");
992 assert(Distance > 0 && "The distance must be non-zero");
994 // Skip if the distance is not multiple of type byte size.
995 if (Distance % TypeByteSize)
998 unsigned ScaledDist = Distance / TypeByteSize;
1000 // No dependence if the scaled distance is not multiple of the stride.
1002 // for (i = 0; i < 1024 ; i += 4)
1003 // A[i+2] = A[i] + 1;
1005 // Two accesses in memory (scaled distance is 2, stride is 4):
1006 // | A[0] | | | | A[4] | | | |
1007 // | | | A[2] | | | | A[6] | |
1010 // for (i = 0; i < 1024 ; i += 3)
1011 // A[i+4] = A[i] + 1;
1013 // Two accesses in memory (scaled distance is 4, stride is 3):
1014 // | A[0] | | | A[3] | | | A[6] | | |
1015 // | | | | | A[4] | | | A[7] | |
1016 return ScaledDist % Stride;
1019 MemoryDepChecker::Dependence::DepType
1020 MemoryDepChecker::isDependent(const MemAccessInfo &A, unsigned AIdx,
1021 const MemAccessInfo &B, unsigned BIdx,
1022 const ValueToValueMap &Strides) {
1023 assert (AIdx < BIdx && "Must pass arguments in program order");
1025 Value *APtr = A.getPointer();
1026 Value *BPtr = B.getPointer();
1027 bool AIsWrite = A.getInt();
1028 bool BIsWrite = B.getInt();
1030 // Two reads are independent.
1031 if (!AIsWrite && !BIsWrite)
1032 return Dependence::NoDep;
1034 // We cannot check pointers in different address spaces.
1035 if (APtr->getType()->getPointerAddressSpace() !=
1036 BPtr->getType()->getPointerAddressSpace())
1037 return Dependence::Unknown;
1039 const SCEV *AScev = replaceSymbolicStrideSCEV(SE, Strides, APtr);
1040 const SCEV *BScev = replaceSymbolicStrideSCEV(SE, Strides, BPtr);
1042 int StrideAPtr = isStridedPtr(SE, APtr, InnermostLoop, Strides);
1043 int StrideBPtr = isStridedPtr(SE, BPtr, InnermostLoop, Strides);
1045 const SCEV *Src = AScev;
1046 const SCEV *Sink = BScev;
1048 // If the induction step is negative we have to invert source and sink of the
1050 if (StrideAPtr < 0) {
1053 std::swap(APtr, BPtr);
1054 std::swap(Src, Sink);
1055 std::swap(AIsWrite, BIsWrite);
1056 std::swap(AIdx, BIdx);
1057 std::swap(StrideAPtr, StrideBPtr);
1060 const SCEV *Dist = SE->getMinusSCEV(Sink, Src);
1062 DEBUG(dbgs() << "LAA: Src Scev: " << *Src << "Sink Scev: " << *Sink
1063 << "(Induction step: " << StrideAPtr << ")\n");
1064 DEBUG(dbgs() << "LAA: Distance for " << *InstMap[AIdx] << " to "
1065 << *InstMap[BIdx] << ": " << *Dist << "\n");
1067 // Need accesses with constant stride. We don't want to vectorize
1068 // "A[B[i]] += ..." and similar code or pointer arithmetic that could wrap in
1069 // the address space.
1070 if (!StrideAPtr || !StrideBPtr || StrideAPtr != StrideBPtr){
1071 DEBUG(dbgs() << "Pointer access with non-constant stride\n");
1072 return Dependence::Unknown;
1075 const SCEVConstant *C = dyn_cast<SCEVConstant>(Dist);
1077 DEBUG(dbgs() << "LAA: Dependence because of non-constant distance\n");
1078 ShouldRetryWithRuntimeCheck = true;
1079 return Dependence::Unknown;
1082 Type *ATy = APtr->getType()->getPointerElementType();
1083 Type *BTy = BPtr->getType()->getPointerElementType();
1084 auto &DL = InnermostLoop->getHeader()->getModule()->getDataLayout();
1085 unsigned TypeByteSize = DL.getTypeAllocSize(ATy);
1087 // Negative distances are not plausible dependencies.
1088 const APInt &Val = C->getValue()->getValue();
1089 if (Val.isNegative()) {
1090 bool IsTrueDataDependence = (AIsWrite && !BIsWrite);
1091 if (IsTrueDataDependence &&
1092 (couldPreventStoreLoadForward(Val.abs().getZExtValue(), TypeByteSize) ||
1094 return Dependence::ForwardButPreventsForwarding;
1096 DEBUG(dbgs() << "LAA: Dependence is negative: NoDep\n");
1097 return Dependence::Forward;
1100 // Write to the same location with the same size.
1101 // Could be improved to assert type sizes are the same (i32 == float, etc).
1104 return Dependence::NoDep;
1105 DEBUG(dbgs() << "LAA: Zero dependence difference but different types\n");
1106 return Dependence::Unknown;
1109 assert(Val.isStrictlyPositive() && "Expect a positive value");
1113 "LAA: ReadWrite-Write positive dependency with different types\n");
1114 return Dependence::Unknown;
1117 unsigned Distance = (unsigned) Val.getZExtValue();
1119 unsigned Stride = std::abs(StrideAPtr);
1121 areStridedAccessesIndependent(Distance, Stride, TypeByteSize)) {
1122 DEBUG(dbgs() << "LAA: Strided accesses are independent\n");
1123 return Dependence::NoDep;
1126 // Bail out early if passed-in parameters make vectorization not feasible.
1127 unsigned ForcedFactor = (VectorizerParams::VectorizationFactor ?
1128 VectorizerParams::VectorizationFactor : 1);
1129 unsigned ForcedUnroll = (VectorizerParams::VectorizationInterleave ?
1130 VectorizerParams::VectorizationInterleave : 1);
1131 // The minimum number of iterations for a vectorized/unrolled version.
1132 unsigned MinNumIter = std::max(ForcedFactor * ForcedUnroll, 2U);
1134 // It's not vectorizable if the distance is smaller than the minimum distance
1135 // needed for a vectroized/unrolled version. Vectorizing one iteration in
1136 // front needs TypeByteSize * Stride. Vectorizing the last iteration needs
1137 // TypeByteSize (No need to plus the last gap distance).
1139 // E.g. Assume one char is 1 byte in memory and one int is 4 bytes.
1141 // int *B = (int *)((char *)A + 14);
1142 // for (i = 0 ; i < 1024 ; i += 2)
1146 // Two accesses in memory (stride is 2):
1147 // | A[0] | | A[2] | | A[4] | | A[6] | |
1148 // | B[0] | | B[2] | | B[4] |
1150 // Distance needs for vectorizing iterations except the last iteration:
1151 // 4 * 2 * (MinNumIter - 1). Distance needs for the last iteration: 4.
1152 // So the minimum distance needed is: 4 * 2 * (MinNumIter - 1) + 4.
1154 // If MinNumIter is 2, it is vectorizable as the minimum distance needed is
1155 // 12, which is less than distance.
1157 // If MinNumIter is 4 (Say if a user forces the vectorization factor to be 4),
1158 // the minimum distance needed is 28, which is greater than distance. It is
1159 // not safe to do vectorization.
1160 unsigned MinDistanceNeeded =
1161 TypeByteSize * Stride * (MinNumIter - 1) + TypeByteSize;
1162 if (MinDistanceNeeded > Distance) {
1163 DEBUG(dbgs() << "LAA: Failure because of positive distance " << Distance
1165 return Dependence::Backward;
1168 // Unsafe if the minimum distance needed is greater than max safe distance.
1169 if (MinDistanceNeeded > MaxSafeDepDistBytes) {
1170 DEBUG(dbgs() << "LAA: Failure because it needs at least "
1171 << MinDistanceNeeded << " size in bytes");
1172 return Dependence::Backward;
1175 // Positive distance bigger than max vectorization factor.
1176 // FIXME: Should use max factor instead of max distance in bytes, which could
1177 // not handle different types.
1178 // E.g. Assume one char is 1 byte in memory and one int is 4 bytes.
1179 // void foo (int *A, char *B) {
1180 // for (unsigned i = 0; i < 1024; i++) {
1181 // A[i+2] = A[i] + 1;
1182 // B[i+2] = B[i] + 1;
1186 // This case is currently unsafe according to the max safe distance. If we
1187 // analyze the two accesses on array B, the max safe dependence distance
1188 // is 2. Then we analyze the accesses on array A, the minimum distance needed
1189 // is 8, which is less than 2 and forbidden vectorization, But actually
1190 // both A and B could be vectorized by 2 iterations.
1191 MaxSafeDepDistBytes =
1192 Distance < MaxSafeDepDistBytes ? Distance : MaxSafeDepDistBytes;
1194 bool IsTrueDataDependence = (!AIsWrite && BIsWrite);
1195 if (IsTrueDataDependence &&
1196 couldPreventStoreLoadForward(Distance, TypeByteSize))
1197 return Dependence::BackwardVectorizableButPreventsForwarding;
1199 DEBUG(dbgs() << "LAA: Positive distance " << Val.getSExtValue()
1200 << " with max VF = "
1201 << MaxSafeDepDistBytes / (TypeByteSize * Stride) << '\n');
1203 return Dependence::BackwardVectorizable;
1206 bool MemoryDepChecker::areDepsSafe(DepCandidates &AccessSets,
1207 MemAccessInfoSet &CheckDeps,
1208 const ValueToValueMap &Strides) {
1210 MaxSafeDepDistBytes = -1U;
1211 while (!CheckDeps.empty()) {
1212 MemAccessInfo CurAccess = *CheckDeps.begin();
1214 // Get the relevant memory access set.
1215 EquivalenceClasses<MemAccessInfo>::iterator I =
1216 AccessSets.findValue(AccessSets.getLeaderValue(CurAccess));
1218 // Check accesses within this set.
1219 EquivalenceClasses<MemAccessInfo>::member_iterator AI, AE;
1220 AI = AccessSets.member_begin(I), AE = AccessSets.member_end();
1222 // Check every access pair.
1224 CheckDeps.erase(*AI);
1225 EquivalenceClasses<MemAccessInfo>::member_iterator OI = std::next(AI);
1227 // Check every accessing instruction pair in program order.
1228 for (std::vector<unsigned>::iterator I1 = Accesses[*AI].begin(),
1229 I1E = Accesses[*AI].end(); I1 != I1E; ++I1)
1230 for (std::vector<unsigned>::iterator I2 = Accesses[*OI].begin(),
1231 I2E = Accesses[*OI].end(); I2 != I2E; ++I2) {
1232 auto A = std::make_pair(&*AI, *I1);
1233 auto B = std::make_pair(&*OI, *I2);
1239 Dependence::DepType Type =
1240 isDependent(*A.first, A.second, *B.first, B.second, Strides);
1241 SafeForVectorization &= Dependence::isSafeForVectorization(Type);
1243 // Gather dependences unless we accumulated MaxInterestingDependence
1244 // dependences. In that case return as soon as we find the first
1245 // unsafe dependence. This puts a limit on this quadratic
1247 if (RecordInterestingDependences) {
1248 if (Dependence::isInterestingDependence(Type))
1249 InterestingDependences.push_back(
1250 Dependence(A.second, B.second, Type));
1252 if (InterestingDependences.size() >= MaxInterestingDependence) {
1253 RecordInterestingDependences = false;
1254 InterestingDependences.clear();
1255 DEBUG(dbgs() << "Too many dependences, stopped recording\n");
1258 if (!RecordInterestingDependences && !SafeForVectorization)
1267 DEBUG(dbgs() << "Total Interesting Dependences: "
1268 << InterestingDependences.size() << "\n");
1269 return SafeForVectorization;
1272 SmallVector<Instruction *, 4>
1273 MemoryDepChecker::getInstructionsForAccess(Value *Ptr, bool isWrite) const {
1274 MemAccessInfo Access(Ptr, isWrite);
1275 auto &IndexVector = Accesses.find(Access)->second;
1277 SmallVector<Instruction *, 4> Insts;
1278 std::transform(IndexVector.begin(), IndexVector.end(),
1279 std::back_inserter(Insts),
1280 [&](unsigned Idx) { return this->InstMap[Idx]; });
1284 const char *MemoryDepChecker::Dependence::DepName[] = {
1285 "NoDep", "Unknown", "Forward", "ForwardButPreventsForwarding", "Backward",
1286 "BackwardVectorizable", "BackwardVectorizableButPreventsForwarding"};
1288 void MemoryDepChecker::Dependence::print(
1289 raw_ostream &OS, unsigned Depth,
1290 const SmallVectorImpl<Instruction *> &Instrs) const {
1291 OS.indent(Depth) << DepName[Type] << ":\n";
1292 OS.indent(Depth + 2) << *Instrs[Source] << " -> \n";
1293 OS.indent(Depth + 2) << *Instrs[Destination] << "\n";
1296 bool LoopAccessInfo::canAnalyzeLoop() {
1297 // We need to have a loop header.
1298 DEBUG(dbgs() << "LAA: Found a loop: " <<
1299 TheLoop->getHeader()->getName() << '\n');
1301 // We can only analyze innermost loops.
1302 if (!TheLoop->empty()) {
1303 DEBUG(dbgs() << "LAA: loop is not the innermost loop\n");
1304 emitAnalysis(LoopAccessReport() << "loop is not the innermost loop");
1308 // We must have a single backedge.
1309 if (TheLoop->getNumBackEdges() != 1) {
1310 DEBUG(dbgs() << "LAA: loop control flow is not understood by analyzer\n");
1312 LoopAccessReport() <<
1313 "loop control flow is not understood by analyzer");
1317 // We must have a single exiting block.
1318 if (!TheLoop->getExitingBlock()) {
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 only handle bottom-tested loops, i.e. loop in which the condition is
1327 // checked at the end of each iteration. With that we can assume that all
1328 // instructions in the loop are executed the same number of times.
1329 if (TheLoop->getExitingBlock() != TheLoop->getLoopLatch()) {
1330 DEBUG(dbgs() << "LAA: loop control flow is not understood by analyzer\n");
1332 LoopAccessReport() <<
1333 "loop control flow is not understood by analyzer");
1337 // ScalarEvolution needs to be able to find the exit count.
1338 const SCEV *ExitCount = SE->getBackedgeTakenCount(TheLoop);
1339 if (ExitCount == SE->getCouldNotCompute()) {
1340 emitAnalysis(LoopAccessReport() <<
1341 "could not determine number of loop iterations");
1342 DEBUG(dbgs() << "LAA: SCEV could not compute the loop exit count.\n");
1349 void LoopAccessInfo::analyzeLoop(const ValueToValueMap &Strides) {
1351 typedef SmallVector<Value*, 16> ValueVector;
1352 typedef SmallPtrSet<Value*, 16> ValueSet;
1354 // Holds the Load and Store *instructions*.
1358 // Holds all the different accesses in the loop.
1359 unsigned NumReads = 0;
1360 unsigned NumReadWrites = 0;
1362 PtrRtChecking.Pointers.clear();
1363 PtrRtChecking.Need = false;
1365 const bool IsAnnotatedParallel = TheLoop->isAnnotatedParallel();
1368 for (Loop::block_iterator bb = TheLoop->block_begin(),
1369 be = TheLoop->block_end(); bb != be; ++bb) {
1371 // Scan the BB and collect legal loads and stores.
1372 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
1375 // If this is a load, save it. If this instruction can read from memory
1376 // but is not a load, then we quit. Notice that we don't handle function
1377 // calls that read or write.
1378 if (it->mayReadFromMemory()) {
1379 // Many math library functions read the rounding mode. We will only
1380 // vectorize a loop if it contains known function calls that don't set
1381 // the flag. Therefore, it is safe to ignore this read from memory.
1382 CallInst *Call = dyn_cast<CallInst>(it);
1383 if (Call && getIntrinsicIDForCall(Call, TLI))
1386 // If the function has an explicit vectorized counterpart, we can safely
1387 // assume that it can be vectorized.
1388 if (Call && !Call->isNoBuiltin() && Call->getCalledFunction() &&
1389 TLI->isFunctionVectorizable(Call->getCalledFunction()->getName()))
1392 LoadInst *Ld = dyn_cast<LoadInst>(it);
1393 if (!Ld || (!Ld->isSimple() && !IsAnnotatedParallel)) {
1394 emitAnalysis(LoopAccessReport(Ld)
1395 << "read with atomic ordering or volatile read");
1396 DEBUG(dbgs() << "LAA: Found a non-simple load.\n");
1401 Loads.push_back(Ld);
1402 DepChecker.addAccess(Ld);
1406 // Save 'store' instructions. Abort if other instructions write to memory.
1407 if (it->mayWriteToMemory()) {
1408 StoreInst *St = dyn_cast<StoreInst>(it);
1410 emitAnalysis(LoopAccessReport(it) <<
1411 "instruction cannot be vectorized");
1415 if (!St->isSimple() && !IsAnnotatedParallel) {
1416 emitAnalysis(LoopAccessReport(St)
1417 << "write with atomic ordering or volatile write");
1418 DEBUG(dbgs() << "LAA: Found a non-simple store.\n");
1423 Stores.push_back(St);
1424 DepChecker.addAccess(St);
1429 // Now we have two lists that hold the loads and the stores.
1430 // Next, we find the pointers that they use.
1432 // Check if we see any stores. If there are no stores, then we don't
1433 // care if the pointers are *restrict*.
1434 if (!Stores.size()) {
1435 DEBUG(dbgs() << "LAA: Found a read-only loop!\n");
1440 MemoryDepChecker::DepCandidates DependentAccesses;
1441 AccessAnalysis Accesses(TheLoop->getHeader()->getModule()->getDataLayout(),
1442 AA, LI, DependentAccesses);
1444 // Holds the analyzed pointers. We don't want to call GetUnderlyingObjects
1445 // multiple times on the same object. If the ptr is accessed twice, once
1446 // for read and once for write, it will only appear once (on the write
1447 // list). This is okay, since we are going to check for conflicts between
1448 // writes and between reads and writes, but not between reads and reads.
1451 ValueVector::iterator I, IE;
1452 for (I = Stores.begin(), IE = Stores.end(); I != IE; ++I) {
1453 StoreInst *ST = cast<StoreInst>(*I);
1454 Value* Ptr = ST->getPointerOperand();
1455 // Check for store to loop invariant address.
1456 StoreToLoopInvariantAddress |= isUniform(Ptr);
1457 // If we did *not* see this pointer before, insert it to the read-write
1458 // list. At this phase it is only a 'write' list.
1459 if (Seen.insert(Ptr).second) {
1462 MemoryLocation Loc = MemoryLocation::get(ST);
1463 // The TBAA metadata could have a control dependency on the predication
1464 // condition, so we cannot rely on it when determining whether or not we
1465 // need runtime pointer checks.
1466 if (blockNeedsPredication(ST->getParent(), TheLoop, DT))
1467 Loc.AATags.TBAA = nullptr;
1469 Accesses.addStore(Loc);
1473 if (IsAnnotatedParallel) {
1475 << "LAA: A loop annotated parallel, ignore memory dependency "
1481 for (I = Loads.begin(), IE = Loads.end(); I != IE; ++I) {
1482 LoadInst *LD = cast<LoadInst>(*I);
1483 Value* Ptr = LD->getPointerOperand();
1484 // If we did *not* see this pointer before, insert it to the
1485 // read list. If we *did* see it before, then it is already in
1486 // the read-write list. This allows us to vectorize expressions
1487 // such as A[i] += x; Because the address of A[i] is a read-write
1488 // pointer. This only works if the index of A[i] is consecutive.
1489 // If the address of i is unknown (for example A[B[i]]) then we may
1490 // read a few words, modify, and write a few words, and some of the
1491 // words may be written to the same address.
1492 bool IsReadOnlyPtr = false;
1493 if (Seen.insert(Ptr).second || !isStridedPtr(SE, Ptr, TheLoop, Strides)) {
1495 IsReadOnlyPtr = true;
1498 MemoryLocation Loc = MemoryLocation::get(LD);
1499 // The TBAA metadata could have a control dependency on the predication
1500 // condition, so we cannot rely on it when determining whether or not we
1501 // need runtime pointer checks.
1502 if (blockNeedsPredication(LD->getParent(), TheLoop, DT))
1503 Loc.AATags.TBAA = nullptr;
1505 Accesses.addLoad(Loc, IsReadOnlyPtr);
1508 // If we write (or read-write) to a single destination and there are no
1509 // other reads in this loop then is it safe to vectorize.
1510 if (NumReadWrites == 1 && NumReads == 0) {
1511 DEBUG(dbgs() << "LAA: Found a write-only loop!\n");
1516 // Build dependence sets and check whether we need a runtime pointer bounds
1518 Accesses.buildDependenceSets();
1520 // Find pointers with computable bounds. We are going to use this information
1521 // to place a runtime bound check.
1522 bool CanDoRTIfNeeded =
1523 Accesses.canCheckPtrAtRT(PtrRtChecking, SE, TheLoop, Strides);
1524 if (!CanDoRTIfNeeded) {
1525 emitAnalysis(LoopAccessReport() << "cannot identify array bounds");
1526 DEBUG(dbgs() << "LAA: We can't vectorize because we can't find "
1527 << "the array bounds.\n");
1532 DEBUG(dbgs() << "LAA: We can perform a memory runtime check if needed.\n");
1535 if (Accesses.isDependencyCheckNeeded()) {
1536 DEBUG(dbgs() << "LAA: Checking memory dependencies\n");
1537 CanVecMem = DepChecker.areDepsSafe(
1538 DependentAccesses, Accesses.getDependenciesToCheck(), Strides);
1539 MaxSafeDepDistBytes = DepChecker.getMaxSafeDepDistBytes();
1541 if (!CanVecMem && DepChecker.shouldRetryWithRuntimeCheck()) {
1542 DEBUG(dbgs() << "LAA: Retrying with memory checks\n");
1544 // Clear the dependency checks. We assume they are not needed.
1545 Accesses.resetDepChecks(DepChecker);
1547 PtrRtChecking.reset();
1548 PtrRtChecking.Need = true;
1551 Accesses.canCheckPtrAtRT(PtrRtChecking, SE, TheLoop, Strides, true);
1553 // Check that we found the bounds for the pointer.
1554 if (!CanDoRTIfNeeded) {
1555 emitAnalysis(LoopAccessReport()
1556 << "cannot check memory dependencies at runtime");
1557 DEBUG(dbgs() << "LAA: Can't vectorize with memory checks\n");
1567 DEBUG(dbgs() << "LAA: No unsafe dependent memory operations in loop. We"
1568 << (PtrRtChecking.Need ? "" : " don't")
1569 << " need runtime memory checks.\n");
1571 emitAnalysis(LoopAccessReport() <<
1572 "unsafe dependent memory operations in loop");
1573 DEBUG(dbgs() << "LAA: unsafe dependent memory operations in loop\n");
1577 bool LoopAccessInfo::blockNeedsPredication(BasicBlock *BB, Loop *TheLoop,
1578 DominatorTree *DT) {
1579 assert(TheLoop->contains(BB) && "Unknown block used");
1581 // Blocks that do not dominate the latch need predication.
1582 BasicBlock* Latch = TheLoop->getLoopLatch();
1583 return !DT->dominates(BB, Latch);
1586 void LoopAccessInfo::emitAnalysis(LoopAccessReport &Message) {
1587 assert(!Report && "Multiple reports generated");
1591 bool LoopAccessInfo::isUniform(Value *V) const {
1592 return (SE->isLoopInvariant(SE->getSCEV(V), TheLoop));
1595 // FIXME: this function is currently a duplicate of the one in
1596 // LoopVectorize.cpp.
1597 static Instruction *getFirstInst(Instruction *FirstInst, Value *V,
1601 if (Instruction *I = dyn_cast<Instruction>(V))
1602 return I->getParent() == Loc->getParent() ? I : nullptr;
1606 /// \brief IR Values for the lower and upper bounds of a pointer evolution.
1607 struct PointerBounds {
1612 /// \brief Expand code for the lower and upper bound of the pointer group \p CG
1613 /// in \p TheLoop. \return the values for the bounds.
1614 static PointerBounds
1615 expandBounds(const RuntimePointerChecking::CheckingPtrGroup *CG, Loop *TheLoop,
1616 Instruction *Loc, SCEVExpander &Exp, ScalarEvolution *SE,
1617 const RuntimePointerChecking &PtrRtChecking) {
1618 Value *Ptr = PtrRtChecking.Pointers[CG->Members[0]].PointerValue;
1619 const SCEV *Sc = SE->getSCEV(Ptr);
1621 if (SE->isLoopInvariant(Sc, TheLoop)) {
1622 DEBUG(dbgs() << "LAA: Adding RT check for a loop invariant ptr:" << *Ptr
1626 unsigned AS = Ptr->getType()->getPointerAddressSpace();
1627 LLVMContext &Ctx = Loc->getContext();
1629 // Use this type for pointer arithmetic.
1630 Type *PtrArithTy = Type::getInt8PtrTy(Ctx, AS);
1631 Value *Start = nullptr, *End = nullptr;
1633 DEBUG(dbgs() << "LAA: Adding RT check for range:\n");
1634 Start = Exp.expandCodeFor(CG->Low, PtrArithTy, Loc);
1635 End = Exp.expandCodeFor(CG->High, PtrArithTy, Loc);
1636 DEBUG(dbgs() << "Start: " << *CG->Low << " End: " << *CG->High << "\n");
1637 return {Start, End};
1641 /// \brief Turns a collection of checks into a collection of expanded upper and
1642 /// lower bounds for both pointers in the check.
1643 static SmallVector<std::pair<PointerBounds, PointerBounds>, 4> expandBounds(
1644 const SmallVectorImpl<RuntimePointerChecking::PointerCheck> &PointerChecks,
1645 Loop *L, Instruction *Loc, ScalarEvolution *SE, SCEVExpander &Exp,
1646 const RuntimePointerChecking &PtrRtChecking) {
1647 SmallVector<std::pair<PointerBounds, PointerBounds>, 4> ChecksWithBounds;
1649 // Here we're relying on the SCEV Expander's cache to only emit code for the
1650 // same bounds once.
1652 PointerChecks.begin(), PointerChecks.end(),
1653 std::back_inserter(ChecksWithBounds),
1654 [&](const RuntimePointerChecking::PointerCheck &Check) {
1656 First = expandBounds(Check.first, L, Loc, Exp, SE, PtrRtChecking),
1657 Second = expandBounds(Check.second, L, Loc, Exp, SE, PtrRtChecking);
1658 return std::make_pair(First, Second);
1661 return ChecksWithBounds;
1664 std::pair<Instruction *, Instruction *> LoopAccessInfo::addRuntimeCheck(
1666 const SmallVectorImpl<RuntimePointerChecking::PointerCheck> &PointerChecks)
1669 SCEVExpander Exp(*SE, DL, "induction");
1670 auto ExpandedChecks =
1671 expandBounds(PointerChecks, TheLoop, Loc, SE, Exp, PtrRtChecking);
1673 LLVMContext &Ctx = Loc->getContext();
1674 Instruction *FirstInst = nullptr;
1675 IRBuilder<> ChkBuilder(Loc);
1676 // Our instructions might fold to a constant.
1677 Value *MemoryRuntimeCheck = nullptr;
1679 for (const auto &Check : ExpandedChecks) {
1680 const PointerBounds &A = Check.first, &B = Check.second;
1681 unsigned AS0 = A.Start->getType()->getPointerAddressSpace();
1682 unsigned AS1 = B.Start->getType()->getPointerAddressSpace();
1684 assert((AS0 == B.End->getType()->getPointerAddressSpace()) &&
1685 (AS1 == A.End->getType()->getPointerAddressSpace()) &&
1686 "Trying to bounds check pointers with different address spaces");
1688 Type *PtrArithTy0 = Type::getInt8PtrTy(Ctx, AS0);
1689 Type *PtrArithTy1 = Type::getInt8PtrTy(Ctx, AS1);
1691 Value *Start0 = ChkBuilder.CreateBitCast(A.Start, PtrArithTy0, "bc");
1692 Value *Start1 = ChkBuilder.CreateBitCast(B.Start, PtrArithTy1, "bc");
1693 Value *End0 = ChkBuilder.CreateBitCast(A.End, PtrArithTy1, "bc");
1694 Value *End1 = ChkBuilder.CreateBitCast(B.End, PtrArithTy0, "bc");
1696 Value *Cmp0 = ChkBuilder.CreateICmpULE(Start0, End1, "bound0");
1697 FirstInst = getFirstInst(FirstInst, Cmp0, Loc);
1698 Value *Cmp1 = ChkBuilder.CreateICmpULE(Start1, End0, "bound1");
1699 FirstInst = getFirstInst(FirstInst, Cmp1, Loc);
1700 Value *IsConflict = ChkBuilder.CreateAnd(Cmp0, Cmp1, "found.conflict");
1701 FirstInst = getFirstInst(FirstInst, IsConflict, Loc);
1702 if (MemoryRuntimeCheck) {
1704 ChkBuilder.CreateOr(MemoryRuntimeCheck, IsConflict, "conflict.rdx");
1705 FirstInst = getFirstInst(FirstInst, IsConflict, Loc);
1707 MemoryRuntimeCheck = IsConflict;
1710 if (!MemoryRuntimeCheck)
1711 return std::make_pair(nullptr, nullptr);
1713 // We have to do this trickery because the IRBuilder might fold the check to a
1714 // constant expression in which case there is no Instruction anchored in a
1716 Instruction *Check = BinaryOperator::CreateAnd(MemoryRuntimeCheck,
1717 ConstantInt::getTrue(Ctx));
1718 ChkBuilder.Insert(Check, "memcheck.conflict");
1719 FirstInst = getFirstInst(FirstInst, Check, Loc);
1720 return std::make_pair(FirstInst, Check);
1723 std::pair<Instruction *, Instruction *> LoopAccessInfo::addRuntimeCheck(
1724 Instruction *Loc, const SmallVectorImpl<int> *PtrPartition) const {
1725 if (!PtrRtChecking.Need)
1726 return std::make_pair(nullptr, nullptr);
1728 return addRuntimeCheck(Loc, PtrRtChecking.generateChecks(PtrPartition));
1731 LoopAccessInfo::LoopAccessInfo(Loop *L, ScalarEvolution *SE,
1732 const DataLayout &DL,
1733 const TargetLibraryInfo *TLI, AliasAnalysis *AA,
1734 DominatorTree *DT, LoopInfo *LI,
1735 const ValueToValueMap &Strides)
1736 : PtrRtChecking(SE), DepChecker(SE, L), TheLoop(L), SE(SE), DL(DL),
1737 TLI(TLI), AA(AA), DT(DT), LI(LI), NumLoads(0), NumStores(0),
1738 MaxSafeDepDistBytes(-1U), CanVecMem(false),
1739 StoreToLoopInvariantAddress(false) {
1740 if (canAnalyzeLoop())
1741 analyzeLoop(Strides);
1744 void LoopAccessInfo::print(raw_ostream &OS, unsigned Depth) const {
1746 if (PtrRtChecking.Need)
1747 OS.indent(Depth) << "Memory dependences are safe with run-time checks\n";
1749 OS.indent(Depth) << "Memory dependences are safe\n";
1753 OS.indent(Depth) << "Report: " << Report->str() << "\n";
1755 if (auto *InterestingDependences = DepChecker.getInterestingDependences()) {
1756 OS.indent(Depth) << "Interesting Dependences:\n";
1757 for (auto &Dep : *InterestingDependences) {
1758 Dep.print(OS, Depth + 2, DepChecker.getMemoryInstructions());
1762 OS.indent(Depth) << "Too many interesting dependences, not recorded\n";
1764 // List the pair of accesses need run-time checks to prove independence.
1765 PtrRtChecking.print(OS, Depth);
1768 OS.indent(Depth) << "Store to invariant address was "
1769 << (StoreToLoopInvariantAddress ? "" : "not ")
1770 << "found in loop.\n";
1773 const LoopAccessInfo &
1774 LoopAccessAnalysis::getInfo(Loop *L, const ValueToValueMap &Strides) {
1775 auto &LAI = LoopAccessInfoMap[L];
1778 assert((!LAI || LAI->NumSymbolicStrides == Strides.size()) &&
1779 "Symbolic strides changed for loop");
1783 const DataLayout &DL = L->getHeader()->getModule()->getDataLayout();
1784 LAI = llvm::make_unique<LoopAccessInfo>(L, SE, DL, TLI, AA, DT, LI,
1787 LAI->NumSymbolicStrides = Strides.size();
1793 void LoopAccessAnalysis::print(raw_ostream &OS, const Module *M) const {
1794 LoopAccessAnalysis &LAA = *const_cast<LoopAccessAnalysis *>(this);
1796 ValueToValueMap NoSymbolicStrides;
1798 for (Loop *TopLevelLoop : *LI)
1799 for (Loop *L : depth_first(TopLevelLoop)) {
1800 OS.indent(2) << L->getHeader()->getName() << ":\n";
1801 auto &LAI = LAA.getInfo(L, NoSymbolicStrides);
1806 bool LoopAccessAnalysis::runOnFunction(Function &F) {
1807 SE = &getAnalysis<ScalarEvolution>();
1808 auto *TLIP = getAnalysisIfAvailable<TargetLibraryInfoWrapperPass>();
1809 TLI = TLIP ? &TLIP->getTLI() : nullptr;
1810 AA = &getAnalysis<AliasAnalysis>();
1811 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1812 LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
1817 void LoopAccessAnalysis::getAnalysisUsage(AnalysisUsage &AU) const {
1818 AU.addRequired<ScalarEvolution>();
1819 AU.addRequired<AliasAnalysis>();
1820 AU.addRequired<DominatorTreeWrapperPass>();
1821 AU.addRequired<LoopInfoWrapperPass>();
1823 AU.setPreservesAll();
1826 char LoopAccessAnalysis::ID = 0;
1827 static const char laa_name[] = "Loop Access Analysis";
1828 #define LAA_NAME "loop-accesses"
1830 INITIALIZE_PASS_BEGIN(LoopAccessAnalysis, LAA_NAME, laa_name, false, true)
1831 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
1832 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
1833 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
1834 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
1835 INITIALIZE_PASS_END(LoopAccessAnalysis, LAA_NAME, laa_name, false, true)
1838 Pass *createLAAPass() {
1839 return new LoopAccessAnalysis();