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/LoopPassManager.h"
18 #include "llvm/Analysis/ScalarEvolutionExpander.h"
19 #include "llvm/Analysis/TargetLibraryInfo.h"
20 #include "llvm/Analysis/ValueTracking.h"
21 #include "llvm/Analysis/VectorUtils.h"
22 #include "llvm/IR/DiagnosticInfo.h"
23 #include "llvm/IR/Dominators.h"
24 #include "llvm/IR/IRBuilder.h"
25 #include "llvm/IR/PassManager.h"
26 #include "llvm/Support/Debug.h"
27 #include "llvm/Support/raw_ostream.h"
30 #define DEBUG_TYPE "loop-accesses"
32 static cl::opt<unsigned, true>
33 VectorizationFactor("force-vector-width", cl::Hidden,
34 cl::desc("Sets the SIMD width. Zero is autoselect."),
35 cl::location(VectorizerParams::VectorizationFactor));
36 unsigned VectorizerParams::VectorizationFactor;
38 static cl::opt<unsigned, true>
39 VectorizationInterleave("force-vector-interleave", cl::Hidden,
40 cl::desc("Sets the vectorization interleave count. "
41 "Zero is autoselect."),
43 VectorizerParams::VectorizationInterleave));
44 unsigned VectorizerParams::VectorizationInterleave;
46 static cl::opt<unsigned, true> RuntimeMemoryCheckThreshold(
47 "runtime-memory-check-threshold", cl::Hidden,
48 cl::desc("When performing memory disambiguation checks at runtime do not "
49 "generate more than this number of comparisons (default = 8)."),
50 cl::location(VectorizerParams::RuntimeMemoryCheckThreshold), cl::init(8));
51 unsigned VectorizerParams::RuntimeMemoryCheckThreshold;
53 /// \brief The maximum iterations used to merge memory checks
54 static cl::opt<unsigned> MemoryCheckMergeThreshold(
55 "memory-check-merge-threshold", cl::Hidden,
56 cl::desc("Maximum number of comparisons done when trying to merge "
57 "runtime memory checks. (default = 100)"),
60 /// Maximum SIMD width.
61 const unsigned VectorizerParams::MaxVectorWidth = 64;
63 /// \brief We collect dependences up to this threshold.
64 static cl::opt<unsigned>
65 MaxDependences("max-dependences", cl::Hidden,
66 cl::desc("Maximum number of dependences collected by "
67 "loop-access analysis (default = 100)"),
70 /// This enables versioning on the strides of symbolically striding memory
71 /// accesses in code like the following.
72 /// for (i = 0; i < N; ++i)
73 /// A[i * Stride1] += B[i * Stride2] ...
75 /// Will be roughly translated to
76 /// if (Stride1 == 1 && Stride2 == 1) {
77 /// for (i = 0; i < N; i+=4)
81 static cl::opt<bool> EnableMemAccessVersioning(
82 "enable-mem-access-versioning", cl::init(true), cl::Hidden,
83 cl::desc("Enable symbolic stride memory access versioning"));
85 /// \brief Enable store-to-load forwarding conflict detection. This option can
86 /// be disabled for correctness testing.
87 static cl::opt<bool> EnableForwardingConflictDetection(
88 "store-to-load-forwarding-conflict-detection", cl::Hidden,
89 cl::desc("Enable conflict detection in loop-access analysis"),
92 bool VectorizerParams::isInterleaveForced() {
93 return ::VectorizationInterleave.getNumOccurrences() > 0;
96 void LoopAccessReport::emitAnalysis(const LoopAccessReport &Message,
97 const Function *TheFunction,
99 const char *PassName) {
100 DebugLoc DL = TheLoop->getStartLoc();
101 if (const Instruction *I = Message.getInstr())
102 DL = I->getDebugLoc();
103 emitOptimizationRemarkAnalysis(TheFunction->getContext(), PassName,
104 *TheFunction, DL, Message.str());
107 Value *llvm::stripIntegerCast(Value *V) {
108 if (auto *CI = dyn_cast<CastInst>(V))
109 if (CI->getOperand(0)->getType()->isIntegerTy())
110 return CI->getOperand(0);
114 const SCEV *llvm::replaceSymbolicStrideSCEV(PredicatedScalarEvolution &PSE,
115 const ValueToValueMap &PtrToStride,
116 Value *Ptr, Value *OrigPtr) {
117 const SCEV *OrigSCEV = PSE.getSCEV(Ptr);
119 // If there is an entry in the map return the SCEV of the pointer with the
120 // symbolic stride replaced by one.
121 ValueToValueMap::const_iterator SI =
122 PtrToStride.find(OrigPtr ? OrigPtr : Ptr);
123 if (SI != PtrToStride.end()) {
124 Value *StrideVal = SI->second;
127 StrideVal = stripIntegerCast(StrideVal);
129 // Replace symbolic stride by one.
130 Value *One = ConstantInt::get(StrideVal->getType(), 1);
131 ValueToValueMap RewriteMap;
132 RewriteMap[StrideVal] = One;
134 ScalarEvolution *SE = PSE.getSE();
135 const auto *U = cast<SCEVUnknown>(SE->getSCEV(StrideVal));
137 static_cast<const SCEVConstant *>(SE->getOne(StrideVal->getType()));
139 PSE.addPredicate(*SE->getEqualPredicate(U, CT));
140 auto *Expr = PSE.getSCEV(Ptr);
142 DEBUG(dbgs() << "LAA: Replacing SCEV: " << *OrigSCEV << " by: " << *Expr
147 // Otherwise, just return the SCEV of the original pointer.
151 /// Calculate Start and End points of memory access.
152 /// Let's assume A is the first access and B is a memory access on N-th loop
153 /// iteration. Then B is calculated as:
155 /// Step value may be positive or negative.
156 /// N is a calculated back-edge taken count:
157 /// N = (TripCount > 0) ? RoundDown(TripCount -1 , VF) : 0
158 /// Start and End points are calculated in the following way:
159 /// Start = UMIN(A, B) ; End = UMAX(A, B) + SizeOfElt,
160 /// where SizeOfElt is the size of single memory access in bytes.
162 /// There is no conflict when the intervals are disjoint:
163 /// NoConflict = (P2.Start >= P1.End) || (P1.Start >= P2.End)
164 void RuntimePointerChecking::insert(Loop *Lp, Value *Ptr, bool WritePtr,
165 unsigned DepSetId, unsigned ASId,
166 const ValueToValueMap &Strides,
167 PredicatedScalarEvolution &PSE) {
168 // Get the stride replaced scev.
169 const SCEV *Sc = replaceSymbolicStrideSCEV(PSE, Strides, Ptr);
170 ScalarEvolution *SE = PSE.getSE();
175 if (SE->isLoopInvariant(Sc, Lp))
176 ScStart = ScEnd = Sc;
178 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Sc);
179 assert(AR && "Invalid addrec expression");
180 const SCEV *Ex = PSE.getBackedgeTakenCount();
182 ScStart = AR->getStart();
183 ScEnd = AR->evaluateAtIteration(Ex, *SE);
184 const SCEV *Step = AR->getStepRecurrence(*SE);
186 // For expressions with negative step, the upper bound is ScStart and the
187 // lower bound is ScEnd.
188 if (const auto *CStep = dyn_cast<SCEVConstant>(Step)) {
189 if (CStep->getValue()->isNegative())
190 std::swap(ScStart, ScEnd);
192 // Fallback case: the step is not constant, but we can still
193 // get the upper and lower bounds of the interval by using min/max
195 ScStart = SE->getUMinExpr(ScStart, ScEnd);
196 ScEnd = SE->getUMaxExpr(AR->getStart(), ScEnd);
198 // Add the size of the pointed element to ScEnd.
200 Ptr->getType()->getPointerElementType()->getScalarSizeInBits() / 8;
201 const SCEV *EltSizeSCEV = SE->getConstant(ScEnd->getType(), EltSize);
202 ScEnd = SE->getAddExpr(ScEnd, EltSizeSCEV);
205 Pointers.emplace_back(Ptr, ScStart, ScEnd, WritePtr, DepSetId, ASId, Sc);
208 SmallVector<RuntimePointerChecking::PointerCheck, 4>
209 RuntimePointerChecking::generateChecks() const {
210 SmallVector<PointerCheck, 4> Checks;
212 for (unsigned I = 0; I < CheckingGroups.size(); ++I) {
213 for (unsigned J = I + 1; J < CheckingGroups.size(); ++J) {
214 const RuntimePointerChecking::CheckingPtrGroup &CGI = CheckingGroups[I];
215 const RuntimePointerChecking::CheckingPtrGroup &CGJ = CheckingGroups[J];
217 if (needsChecking(CGI, CGJ))
218 Checks.push_back(std::make_pair(&CGI, &CGJ));
224 void RuntimePointerChecking::generateChecks(
225 MemoryDepChecker::DepCandidates &DepCands, bool UseDependencies) {
226 assert(Checks.empty() && "Checks is not empty");
227 groupChecks(DepCands, UseDependencies);
228 Checks = generateChecks();
231 bool RuntimePointerChecking::needsChecking(const CheckingPtrGroup &M,
232 const CheckingPtrGroup &N) const {
233 for (unsigned I = 0, EI = M.Members.size(); EI != I; ++I)
234 for (unsigned J = 0, EJ = N.Members.size(); EJ != J; ++J)
235 if (needsChecking(M.Members[I], N.Members[J]))
240 /// Compare \p I and \p J and return the minimum.
241 /// Return nullptr in case we couldn't find an answer.
242 static const SCEV *getMinFromExprs(const SCEV *I, const SCEV *J,
243 ScalarEvolution *SE) {
244 const SCEV *Diff = SE->getMinusSCEV(J, I);
245 const SCEVConstant *C = dyn_cast<const SCEVConstant>(Diff);
249 if (C->getValue()->isNegative())
254 bool RuntimePointerChecking::CheckingPtrGroup::addPointer(unsigned Index) {
255 const SCEV *Start = RtCheck.Pointers[Index].Start;
256 const SCEV *End = RtCheck.Pointers[Index].End;
258 // Compare the starts and ends with the known minimum and maximum
259 // of this set. We need to know how we compare against the min/max
260 // of the set in order to be able to emit memchecks.
261 const SCEV *Min0 = getMinFromExprs(Start, Low, RtCheck.SE);
265 const SCEV *Min1 = getMinFromExprs(End, High, RtCheck.SE);
269 // Update the low bound expression if we've found a new min value.
273 // Update the high bound expression if we've found a new max value.
277 Members.push_back(Index);
281 void RuntimePointerChecking::groupChecks(
282 MemoryDepChecker::DepCandidates &DepCands, bool UseDependencies) {
283 // We build the groups from dependency candidates equivalence classes
285 // - We know that pointers in the same equivalence class share
286 // the same underlying object and therefore there is a chance
287 // that we can compare pointers
288 // - We wouldn't be able to merge two pointers for which we need
289 // to emit a memcheck. The classes in DepCands are already
290 // conveniently built such that no two pointers in the same
291 // class need checking against each other.
293 // We use the following (greedy) algorithm to construct the groups
294 // For every pointer in the equivalence class:
295 // For each existing group:
296 // - if the difference between this pointer and the min/max bounds
297 // of the group is a constant, then make the pointer part of the
298 // group and update the min/max bounds of that group as required.
300 CheckingGroups.clear();
302 // If we need to check two pointers to the same underlying object
303 // with a non-constant difference, we shouldn't perform any pointer
304 // grouping with those pointers. This is because we can easily get
305 // into cases where the resulting check would return false, even when
306 // the accesses are safe.
308 // The following example shows this:
309 // for (i = 0; i < 1000; ++i)
310 // a[5000 + i * m] = a[i] + a[i + 9000]
312 // Here grouping gives a check of (5000, 5000 + 1000 * m) against
313 // (0, 10000) which is always false. However, if m is 1, there is no
314 // dependence. Not grouping the checks for a[i] and a[i + 9000] allows
315 // us to perform an accurate check in this case.
317 // The above case requires that we have an UnknownDependence between
318 // accesses to the same underlying object. This cannot happen unless
319 // ShouldRetryWithRuntimeCheck is set, and therefore UseDependencies
320 // is also false. In this case we will use the fallback path and create
321 // separate checking groups for all pointers.
323 // If we don't have the dependency partitions, construct a new
324 // checking pointer group for each pointer. This is also required
325 // for correctness, because in this case we can have checking between
326 // pointers to the same underlying object.
327 if (!UseDependencies) {
328 for (unsigned I = 0; I < Pointers.size(); ++I)
329 CheckingGroups.push_back(CheckingPtrGroup(I, *this));
333 unsigned TotalComparisons = 0;
335 DenseMap<Value *, unsigned> PositionMap;
336 for (unsigned Index = 0; Index < Pointers.size(); ++Index)
337 PositionMap[Pointers[Index].PointerValue] = Index;
339 // We need to keep track of what pointers we've already seen so we
340 // don't process them twice.
341 SmallSet<unsigned, 2> Seen;
343 // Go through all equivalence classes, get the "pointer check groups"
344 // and add them to the overall solution. We use the order in which accesses
345 // appear in 'Pointers' to enforce determinism.
346 for (unsigned I = 0; I < Pointers.size(); ++I) {
347 // We've seen this pointer before, and therefore already processed
348 // its equivalence class.
352 MemoryDepChecker::MemAccessInfo Access(Pointers[I].PointerValue,
353 Pointers[I].IsWritePtr);
355 SmallVector<CheckingPtrGroup, 2> Groups;
356 auto LeaderI = DepCands.findValue(DepCands.getLeaderValue(Access));
358 // Because DepCands is constructed by visiting accesses in the order in
359 // which they appear in alias sets (which is deterministic) and the
360 // iteration order within an equivalence class member is only dependent on
361 // the order in which unions and insertions are performed on the
362 // equivalence class, the iteration order is deterministic.
363 for (auto MI = DepCands.member_begin(LeaderI), ME = DepCands.member_end();
365 unsigned Pointer = PositionMap[MI->getPointer()];
367 // Mark this pointer as seen.
368 Seen.insert(Pointer);
370 // Go through all the existing sets and see if we can find one
371 // which can include this pointer.
372 for (CheckingPtrGroup &Group : Groups) {
373 // Don't perform more than a certain amount of comparisons.
374 // This should limit the cost of grouping the pointers to something
375 // reasonable. If we do end up hitting this threshold, the algorithm
376 // will create separate groups for all remaining pointers.
377 if (TotalComparisons > MemoryCheckMergeThreshold)
382 if (Group.addPointer(Pointer)) {
389 // We couldn't add this pointer to any existing set or the threshold
390 // for the number of comparisons has been reached. Create a new group
391 // to hold the current pointer.
392 Groups.push_back(CheckingPtrGroup(Pointer, *this));
395 // We've computed the grouped checks for this partition.
396 // Save the results and continue with the next one.
397 std::copy(Groups.begin(), Groups.end(), std::back_inserter(CheckingGroups));
401 bool RuntimePointerChecking::arePointersInSamePartition(
402 const SmallVectorImpl<int> &PtrToPartition, unsigned PtrIdx1,
404 return (PtrToPartition[PtrIdx1] != -1 &&
405 PtrToPartition[PtrIdx1] == PtrToPartition[PtrIdx2]);
408 bool RuntimePointerChecking::needsChecking(unsigned I, unsigned J) const {
409 const PointerInfo &PointerI = Pointers[I];
410 const PointerInfo &PointerJ = Pointers[J];
412 // No need to check if two readonly pointers intersect.
413 if (!PointerI.IsWritePtr && !PointerJ.IsWritePtr)
416 // Only need to check pointers between two different dependency sets.
417 if (PointerI.DependencySetId == PointerJ.DependencySetId)
420 // Only need to check pointers in the same alias set.
421 if (PointerI.AliasSetId != PointerJ.AliasSetId)
427 void RuntimePointerChecking::printChecks(
428 raw_ostream &OS, const SmallVectorImpl<PointerCheck> &Checks,
429 unsigned Depth) const {
431 for (const auto &Check : Checks) {
432 const auto &First = Check.first->Members, &Second = Check.second->Members;
434 OS.indent(Depth) << "Check " << N++ << ":\n";
436 OS.indent(Depth + 2) << "Comparing group (" << Check.first << "):\n";
437 for (unsigned K = 0; K < First.size(); ++K)
438 OS.indent(Depth + 2) << *Pointers[First[K]].PointerValue << "\n";
440 OS.indent(Depth + 2) << "Against group (" << Check.second << "):\n";
441 for (unsigned K = 0; K < Second.size(); ++K)
442 OS.indent(Depth + 2) << *Pointers[Second[K]].PointerValue << "\n";
446 void RuntimePointerChecking::print(raw_ostream &OS, unsigned Depth) const {
448 OS.indent(Depth) << "Run-time memory checks:\n";
449 printChecks(OS, Checks, Depth);
451 OS.indent(Depth) << "Grouped accesses:\n";
452 for (unsigned I = 0; I < CheckingGroups.size(); ++I) {
453 const auto &CG = CheckingGroups[I];
455 OS.indent(Depth + 2) << "Group " << &CG << ":\n";
456 OS.indent(Depth + 4) << "(Low: " << *CG.Low << " High: " << *CG.High
458 for (unsigned J = 0; J < CG.Members.size(); ++J) {
459 OS.indent(Depth + 6) << "Member: " << *Pointers[CG.Members[J]].Expr
466 /// \brief Analyses memory accesses in a loop.
468 /// Checks whether run time pointer checks are needed and builds sets for data
469 /// dependence checking.
470 class AccessAnalysis {
472 /// \brief Read or write access location.
473 typedef PointerIntPair<Value *, 1, bool> MemAccessInfo;
474 typedef SmallPtrSet<MemAccessInfo, 8> MemAccessInfoSet;
476 AccessAnalysis(const DataLayout &Dl, AliasAnalysis *AA, LoopInfo *LI,
477 MemoryDepChecker::DepCandidates &DA,
478 PredicatedScalarEvolution &PSE)
479 : DL(Dl), AST(*AA), LI(LI), DepCands(DA), IsRTCheckAnalysisNeeded(false),
482 /// \brief Register a load and whether it is only read from.
483 void addLoad(MemoryLocation &Loc, bool IsReadOnly) {
484 Value *Ptr = const_cast<Value*>(Loc.Ptr);
485 AST.add(Ptr, MemoryLocation::UnknownSize, Loc.AATags);
486 Accesses.insert(MemAccessInfo(Ptr, false));
488 ReadOnlyPtr.insert(Ptr);
491 /// \brief Register a store.
492 void addStore(MemoryLocation &Loc) {
493 Value *Ptr = const_cast<Value*>(Loc.Ptr);
494 AST.add(Ptr, MemoryLocation::UnknownSize, Loc.AATags);
495 Accesses.insert(MemAccessInfo(Ptr, true));
498 /// \brief Check whether we can check the pointers at runtime for
499 /// non-intersection.
501 /// Returns true if we need no check or if we do and we can generate them
502 /// (i.e. the pointers have computable bounds).
503 bool canCheckPtrAtRT(RuntimePointerChecking &RtCheck, ScalarEvolution *SE,
504 Loop *TheLoop, const ValueToValueMap &Strides,
505 bool ShouldCheckWrap = false);
507 /// \brief Goes over all memory accesses, checks whether a RT check is needed
508 /// and builds sets of dependent accesses.
509 void buildDependenceSets() {
510 processMemAccesses();
513 /// \brief Initial processing of memory accesses determined that we need to
514 /// perform dependency checking.
516 /// Note that this can later be cleared if we retry memcheck analysis without
517 /// dependency checking (i.e. ShouldRetryWithRuntimeCheck).
518 bool isDependencyCheckNeeded() { return !CheckDeps.empty(); }
520 /// We decided that no dependence analysis would be used. Reset the state.
521 void resetDepChecks(MemoryDepChecker &DepChecker) {
523 DepChecker.clearDependences();
526 MemAccessInfoSet &getDependenciesToCheck() { return CheckDeps; }
529 typedef SetVector<MemAccessInfo> PtrAccessSet;
531 /// \brief Go over all memory access and check whether runtime pointer checks
532 /// are needed and build sets of dependency check candidates.
533 void processMemAccesses();
535 /// Set of all accesses.
536 PtrAccessSet Accesses;
538 const DataLayout &DL;
540 /// Set of accesses that need a further dependence check.
541 MemAccessInfoSet CheckDeps;
543 /// Set of pointers that are read only.
544 SmallPtrSet<Value*, 16> ReadOnlyPtr;
546 /// An alias set tracker to partition the access set by underlying object and
547 //intrinsic property (such as TBAA metadata).
552 /// Sets of potentially dependent accesses - members of one set share an
553 /// underlying pointer. The set "CheckDeps" identfies which sets really need a
554 /// dependence check.
555 MemoryDepChecker::DepCandidates &DepCands;
557 /// \brief Initial processing of memory accesses determined that we may need
558 /// to add memchecks. Perform the analysis to determine the necessary checks.
560 /// Note that, this is different from isDependencyCheckNeeded. When we retry
561 /// memcheck analysis without dependency checking
562 /// (i.e. ShouldRetryWithRuntimeCheck), isDependencyCheckNeeded is cleared
563 /// while this remains set if we have potentially dependent accesses.
564 bool IsRTCheckAnalysisNeeded;
566 /// The SCEV predicate containing all the SCEV-related assumptions.
567 PredicatedScalarEvolution &PSE;
570 } // end anonymous namespace
572 /// \brief Check whether a pointer can participate in a runtime bounds check.
573 static bool hasComputableBounds(PredicatedScalarEvolution &PSE,
574 const ValueToValueMap &Strides, Value *Ptr,
576 const SCEV *PtrScev = replaceSymbolicStrideSCEV(PSE, Strides, Ptr);
578 // The bounds for loop-invariant pointer is trivial.
579 if (PSE.getSE()->isLoopInvariant(PtrScev, L))
582 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
586 return AR->isAffine();
589 /// \brief Check whether a pointer address cannot wrap.
590 static bool isNoWrap(PredicatedScalarEvolution &PSE,
591 const ValueToValueMap &Strides, Value *Ptr, Loop *L) {
592 const SCEV *PtrScev = PSE.getSCEV(Ptr);
593 if (PSE.getSE()->isLoopInvariant(PtrScev, L))
596 int64_t Stride = getPtrStride(PSE, Ptr, L, Strides);
600 bool AccessAnalysis::canCheckPtrAtRT(RuntimePointerChecking &RtCheck,
601 ScalarEvolution *SE, Loop *TheLoop,
602 const ValueToValueMap &StridesMap,
603 bool ShouldCheckWrap) {
604 // Find pointers with computable bounds. We are going to use this information
605 // to place a runtime bound check.
608 bool NeedRTCheck = false;
609 if (!IsRTCheckAnalysisNeeded) return true;
611 bool IsDepCheckNeeded = isDependencyCheckNeeded();
613 // We assign a consecutive id to access from different alias sets.
614 // Accesses between different groups doesn't need to be checked.
616 for (auto &AS : AST) {
617 int NumReadPtrChecks = 0;
618 int NumWritePtrChecks = 0;
620 // We assign consecutive id to access from different dependence sets.
621 // Accesses within the same set don't need a runtime check.
622 unsigned RunningDepId = 1;
623 DenseMap<Value *, unsigned> DepSetId;
626 Value *Ptr = A.getValue();
627 bool IsWrite = Accesses.count(MemAccessInfo(Ptr, true));
628 MemAccessInfo Access(Ptr, IsWrite);
635 if (hasComputableBounds(PSE, StridesMap, Ptr, TheLoop) &&
636 // When we run after a failing dependency check we have to make sure
637 // we don't have wrapping pointers.
638 (!ShouldCheckWrap || isNoWrap(PSE, StridesMap, Ptr, TheLoop))) {
639 // The id of the dependence set.
642 if (IsDepCheckNeeded) {
643 Value *Leader = DepCands.getLeaderValue(Access).getPointer();
644 unsigned &LeaderId = DepSetId[Leader];
646 LeaderId = RunningDepId++;
649 // Each access has its own dependence set.
650 DepId = RunningDepId++;
652 RtCheck.insert(TheLoop, Ptr, IsWrite, DepId, ASId, StridesMap, PSE);
654 DEBUG(dbgs() << "LAA: Found a runtime check ptr:" << *Ptr << '\n');
656 DEBUG(dbgs() << "LAA: Can't find bounds for ptr:" << *Ptr << '\n');
661 // If we have at least two writes or one write and a read then we need to
662 // check them. But there is no need to checks if there is only one
663 // dependence set for this alias set.
665 // Note that this function computes CanDoRT and NeedRTCheck independently.
666 // For example CanDoRT=false, NeedRTCheck=false means that we have a pointer
667 // for which we couldn't find the bounds but we don't actually need to emit
668 // any checks so it does not matter.
669 if (!(IsDepCheckNeeded && CanDoRT && RunningDepId == 2))
670 NeedRTCheck |= (NumWritePtrChecks >= 2 || (NumReadPtrChecks >= 1 &&
671 NumWritePtrChecks >= 1));
676 // If the pointers that we would use for the bounds comparison have different
677 // address spaces, assume the values aren't directly comparable, so we can't
678 // use them for the runtime check. We also have to assume they could
679 // overlap. In the future there should be metadata for whether address spaces
681 unsigned NumPointers = RtCheck.Pointers.size();
682 for (unsigned i = 0; i < NumPointers; ++i) {
683 for (unsigned j = i + 1; j < NumPointers; ++j) {
684 // Only need to check pointers between two different dependency sets.
685 if (RtCheck.Pointers[i].DependencySetId ==
686 RtCheck.Pointers[j].DependencySetId)
688 // Only need to check pointers in the same alias set.
689 if (RtCheck.Pointers[i].AliasSetId != RtCheck.Pointers[j].AliasSetId)
692 Value *PtrI = RtCheck.Pointers[i].PointerValue;
693 Value *PtrJ = RtCheck.Pointers[j].PointerValue;
695 unsigned ASi = PtrI->getType()->getPointerAddressSpace();
696 unsigned ASj = PtrJ->getType()->getPointerAddressSpace();
698 DEBUG(dbgs() << "LAA: Runtime check would require comparison between"
699 " different address spaces\n");
705 if (NeedRTCheck && CanDoRT)
706 RtCheck.generateChecks(DepCands, IsDepCheckNeeded);
708 DEBUG(dbgs() << "LAA: We need to do " << RtCheck.getNumberOfChecks()
709 << " pointer comparisons.\n");
711 RtCheck.Need = NeedRTCheck;
713 bool CanDoRTIfNeeded = !NeedRTCheck || CanDoRT;
714 if (!CanDoRTIfNeeded)
716 return CanDoRTIfNeeded;
719 void AccessAnalysis::processMemAccesses() {
720 // We process the set twice: first we process read-write pointers, last we
721 // process read-only pointers. This allows us to skip dependence tests for
722 // read-only pointers.
724 DEBUG(dbgs() << "LAA: Processing memory accesses...\n");
725 DEBUG(dbgs() << " AST: "; AST.dump());
726 DEBUG(dbgs() << "LAA: Accesses(" << Accesses.size() << "):\n");
728 for (auto A : Accesses)
729 dbgs() << "\t" << *A.getPointer() << " (" <<
730 (A.getInt() ? "write" : (ReadOnlyPtr.count(A.getPointer()) ?
731 "read-only" : "read")) << ")\n";
734 // The AliasSetTracker has nicely partitioned our pointers by metadata
735 // compatibility and potential for underlying-object overlap. As a result, we
736 // only need to check for potential pointer dependencies within each alias
738 for (auto &AS : AST) {
739 // Note that both the alias-set tracker and the alias sets themselves used
740 // linked lists internally and so the iteration order here is deterministic
741 // (matching the original instruction order within each set).
743 bool SetHasWrite = false;
745 // Map of pointers to last access encountered.
746 typedef DenseMap<Value*, MemAccessInfo> UnderlyingObjToAccessMap;
747 UnderlyingObjToAccessMap ObjToLastAccess;
749 // Set of access to check after all writes have been processed.
750 PtrAccessSet DeferredAccesses;
752 // Iterate over each alias set twice, once to process read/write pointers,
753 // and then to process read-only pointers.
754 for (int SetIteration = 0; SetIteration < 2; ++SetIteration) {
755 bool UseDeferred = SetIteration > 0;
756 PtrAccessSet &S = UseDeferred ? DeferredAccesses : Accesses;
759 Value *Ptr = AV.getValue();
761 // For a single memory access in AliasSetTracker, Accesses may contain
762 // both read and write, and they both need to be handled for CheckDeps.
764 if (AC.getPointer() != Ptr)
767 bool IsWrite = AC.getInt();
769 // If we're using the deferred access set, then it contains only
771 bool IsReadOnlyPtr = ReadOnlyPtr.count(Ptr) && !IsWrite;
772 if (UseDeferred && !IsReadOnlyPtr)
774 // Otherwise, the pointer must be in the PtrAccessSet, either as a
776 assert(((IsReadOnlyPtr && UseDeferred) || IsWrite ||
777 S.count(MemAccessInfo(Ptr, false))) &&
778 "Alias-set pointer not in the access set?");
780 MemAccessInfo Access(Ptr, IsWrite);
781 DepCands.insert(Access);
783 // Memorize read-only pointers for later processing and skip them in
784 // the first round (they need to be checked after we have seen all
785 // write pointers). Note: we also mark pointer that are not
786 // consecutive as "read-only" pointers (so that we check
787 // "a[b[i]] +="). Hence, we need the second check for "!IsWrite".
788 if (!UseDeferred && IsReadOnlyPtr) {
789 DeferredAccesses.insert(Access);
793 // If this is a write - check other reads and writes for conflicts. If
794 // this is a read only check other writes for conflicts (but only if
795 // there is no other write to the ptr - this is an optimization to
796 // catch "a[i] = a[i] + " without having to do a dependence check).
797 if ((IsWrite || IsReadOnlyPtr) && SetHasWrite) {
798 CheckDeps.insert(Access);
799 IsRTCheckAnalysisNeeded = true;
805 // Create sets of pointers connected by a shared alias set and
806 // underlying object.
807 typedef SmallVector<Value *, 16> ValueVector;
808 ValueVector TempObjects;
810 GetUnderlyingObjects(Ptr, TempObjects, DL, LI);
811 DEBUG(dbgs() << "Underlying objects for pointer " << *Ptr << "\n");
812 for (Value *UnderlyingObj : TempObjects) {
813 // nullptr never alias, don't join sets for pointer that have "null"
814 // in their UnderlyingObjects list.
815 if (isa<ConstantPointerNull>(UnderlyingObj))
818 UnderlyingObjToAccessMap::iterator Prev =
819 ObjToLastAccess.find(UnderlyingObj);
820 if (Prev != ObjToLastAccess.end())
821 DepCands.unionSets(Access, Prev->second);
823 ObjToLastAccess[UnderlyingObj] = Access;
824 DEBUG(dbgs() << " " << *UnderlyingObj << "\n");
832 static bool isInBoundsGep(Value *Ptr) {
833 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr))
834 return GEP->isInBounds();
838 /// \brief Return true if an AddRec pointer \p Ptr is unsigned non-wrapping,
839 /// i.e. monotonically increasing/decreasing.
840 static bool isNoWrapAddRec(Value *Ptr, const SCEVAddRecExpr *AR,
841 PredicatedScalarEvolution &PSE, const Loop *L) {
842 // FIXME: This should probably only return true for NUW.
843 if (AR->getNoWrapFlags(SCEV::NoWrapMask))
846 // Scalar evolution does not propagate the non-wrapping flags to values that
847 // are derived from a non-wrapping induction variable because non-wrapping
848 // could be flow-sensitive.
850 // Look through the potentially overflowing instruction to try to prove
851 // non-wrapping for the *specific* value of Ptr.
853 // The arithmetic implied by an inbounds GEP can't overflow.
854 auto *GEP = dyn_cast<GetElementPtrInst>(Ptr);
855 if (!GEP || !GEP->isInBounds())
858 // Make sure there is only one non-const index and analyze that.
859 Value *NonConstIndex = nullptr;
860 for (Value *Index : make_range(GEP->idx_begin(), GEP->idx_end()))
861 if (!isa<ConstantInt>(Index)) {
864 NonConstIndex = Index;
867 // The recurrence is on the pointer, ignore for now.
870 // The index in GEP is signed. It is non-wrapping if it's derived from a NSW
871 // AddRec using a NSW operation.
872 if (auto *OBO = dyn_cast<OverflowingBinaryOperator>(NonConstIndex))
873 if (OBO->hasNoSignedWrap() &&
874 // Assume constant for other the operand so that the AddRec can be
876 isa<ConstantInt>(OBO->getOperand(1))) {
877 auto *OpScev = PSE.getSCEV(OBO->getOperand(0));
879 if (auto *OpAR = dyn_cast<SCEVAddRecExpr>(OpScev))
880 return OpAR->getLoop() == L && OpAR->getNoWrapFlags(SCEV::FlagNSW);
886 /// \brief Check whether the access through \p Ptr has a constant stride.
887 int64_t llvm::getPtrStride(PredicatedScalarEvolution &PSE, Value *Ptr,
888 const Loop *Lp, const ValueToValueMap &StridesMap,
890 Type *Ty = Ptr->getType();
891 assert(Ty->isPointerTy() && "Unexpected non-ptr");
893 // Make sure that the pointer does not point to aggregate types.
894 auto *PtrTy = cast<PointerType>(Ty);
895 if (PtrTy->getElementType()->isAggregateType()) {
896 DEBUG(dbgs() << "LAA: Bad stride - Not a pointer to a scalar type" << *Ptr
901 const SCEV *PtrScev = replaceSymbolicStrideSCEV(PSE, StridesMap, Ptr);
903 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
905 AR = PSE.getAsAddRec(Ptr);
908 DEBUG(dbgs() << "LAA: Bad stride - Not an AddRecExpr pointer " << *Ptr
909 << " SCEV: " << *PtrScev << "\n");
913 // The accesss function must stride over the innermost loop.
914 if (Lp != AR->getLoop()) {
915 DEBUG(dbgs() << "LAA: Bad stride - Not striding over innermost loop " <<
916 *Ptr << " SCEV: " << *AR << "\n");
920 // The address calculation must not wrap. Otherwise, a dependence could be
922 // An inbounds getelementptr that is a AddRec with a unit stride
923 // cannot wrap per definition. The unit stride requirement is checked later.
924 // An getelementptr without an inbounds attribute and unit stride would have
925 // to access the pointer value "0" which is undefined behavior in address
926 // space 0, therefore we can also vectorize this case.
927 bool IsInBoundsGEP = isInBoundsGep(Ptr);
928 bool IsNoWrapAddRec =
929 PSE.hasNoOverflow(Ptr, SCEVWrapPredicate::IncrementNUSW) ||
930 isNoWrapAddRec(Ptr, AR, PSE, Lp);
931 bool IsInAddressSpaceZero = PtrTy->getAddressSpace() == 0;
932 if (!IsNoWrapAddRec && !IsInBoundsGEP && !IsInAddressSpaceZero) {
934 PSE.setNoOverflow(Ptr, SCEVWrapPredicate::IncrementNUSW);
935 IsNoWrapAddRec = true;
936 DEBUG(dbgs() << "LAA: Pointer may wrap in the address space:\n"
937 << "LAA: Pointer: " << *Ptr << "\n"
938 << "LAA: SCEV: " << *AR << "\n"
939 << "LAA: Added an overflow assumption\n");
941 DEBUG(dbgs() << "LAA: Bad stride - Pointer may wrap in the address space "
942 << *Ptr << " SCEV: " << *AR << "\n");
947 // Check the step is constant.
948 const SCEV *Step = AR->getStepRecurrence(*PSE.getSE());
950 // Calculate the pointer stride and check if it is constant.
951 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
953 DEBUG(dbgs() << "LAA: Bad stride - Not a constant strided " << *Ptr <<
954 " SCEV: " << *AR << "\n");
958 auto &DL = Lp->getHeader()->getModule()->getDataLayout();
959 int64_t Size = DL.getTypeAllocSize(PtrTy->getElementType());
960 const APInt &APStepVal = C->getAPInt();
962 // Huge step value - give up.
963 if (APStepVal.getBitWidth() > 64)
966 int64_t StepVal = APStepVal.getSExtValue();
969 int64_t Stride = StepVal / Size;
970 int64_t Rem = StepVal % Size;
974 // If the SCEV could wrap but we have an inbounds gep with a unit stride we
975 // know we can't "wrap around the address space". In case of address space
976 // zero we know that this won't happen without triggering undefined behavior.
977 if (!IsNoWrapAddRec && (IsInBoundsGEP || IsInAddressSpaceZero) &&
978 Stride != 1 && Stride != -1) {
980 // We can avoid this case by adding a run-time check.
981 DEBUG(dbgs() << "LAA: Non unit strided pointer which is not either "
982 << "inbouds or in address space 0 may wrap:\n"
983 << "LAA: Pointer: " << *Ptr << "\n"
984 << "LAA: SCEV: " << *AR << "\n"
985 << "LAA: Added an overflow assumption\n");
986 PSE.setNoOverflow(Ptr, SCEVWrapPredicate::IncrementNUSW);
994 /// Take the pointer operand from the Load/Store instruction.
995 /// Returns NULL if this is not a valid Load/Store instruction.
996 static Value *getPointerOperand(Value *I) {
997 if (auto *LI = dyn_cast<LoadInst>(I))
998 return LI->getPointerOperand();
999 if (auto *SI = dyn_cast<StoreInst>(I))
1000 return SI->getPointerOperand();
1004 /// Take the address space operand from the Load/Store instruction.
1005 /// Returns -1 if this is not a valid Load/Store instruction.
1006 static unsigned getAddressSpaceOperand(Value *I) {
1007 if (LoadInst *L = dyn_cast<LoadInst>(I))
1008 return L->getPointerAddressSpace();
1009 if (StoreInst *S = dyn_cast<StoreInst>(I))
1010 return S->getPointerAddressSpace();
1014 /// Returns true if the memory operations \p A and \p B are consecutive.
1015 bool llvm::isConsecutiveAccess(Value *A, Value *B, const DataLayout &DL,
1016 ScalarEvolution &SE, bool CheckType) {
1017 Value *PtrA = getPointerOperand(A);
1018 Value *PtrB = getPointerOperand(B);
1019 unsigned ASA = getAddressSpaceOperand(A);
1020 unsigned ASB = getAddressSpaceOperand(B);
1022 // Check that the address spaces match and that the pointers are valid.
1023 if (!PtrA || !PtrB || (ASA != ASB))
1026 // Make sure that A and B are different pointers.
1030 // Make sure that A and B have the same type if required.
1031 if(CheckType && PtrA->getType() != PtrB->getType())
1034 unsigned PtrBitWidth = DL.getPointerSizeInBits(ASA);
1035 Type *Ty = cast<PointerType>(PtrA->getType())->getElementType();
1036 APInt Size(PtrBitWidth, DL.getTypeStoreSize(Ty));
1038 APInt OffsetA(PtrBitWidth, 0), OffsetB(PtrBitWidth, 0);
1039 PtrA = PtrA->stripAndAccumulateInBoundsConstantOffsets(DL, OffsetA);
1040 PtrB = PtrB->stripAndAccumulateInBoundsConstantOffsets(DL, OffsetB);
1042 // OffsetDelta = OffsetB - OffsetA;
1043 const SCEV *OffsetSCEVA = SE.getConstant(OffsetA);
1044 const SCEV *OffsetSCEVB = SE.getConstant(OffsetB);
1045 const SCEV *OffsetDeltaSCEV = SE.getMinusSCEV(OffsetSCEVB, OffsetSCEVA);
1046 const SCEVConstant *OffsetDeltaC = dyn_cast<SCEVConstant>(OffsetDeltaSCEV);
1047 const APInt &OffsetDelta = OffsetDeltaC->getAPInt();
1048 // Check if they are based on the same pointer. That makes the offsets
1051 return OffsetDelta == Size;
1053 // Compute the necessary base pointer delta to have the necessary final delta
1054 // equal to the size.
1055 // BaseDelta = Size - OffsetDelta;
1056 const SCEV *SizeSCEV = SE.getConstant(Size);
1057 const SCEV *BaseDelta = SE.getMinusSCEV(SizeSCEV, OffsetDeltaSCEV);
1059 // Otherwise compute the distance with SCEV between the base pointers.
1060 const SCEV *PtrSCEVA = SE.getSCEV(PtrA);
1061 const SCEV *PtrSCEVB = SE.getSCEV(PtrB);
1062 const SCEV *X = SE.getAddExpr(PtrSCEVA, BaseDelta);
1063 return X == PtrSCEVB;
1066 bool MemoryDepChecker::Dependence::isSafeForVectorization(DepType Type) {
1070 case BackwardVectorizable:
1074 case ForwardButPreventsForwarding:
1076 case BackwardVectorizableButPreventsForwarding:
1079 llvm_unreachable("unexpected DepType!");
1082 bool MemoryDepChecker::Dependence::isBackward() const {
1086 case ForwardButPreventsForwarding:
1090 case BackwardVectorizable:
1092 case BackwardVectorizableButPreventsForwarding:
1095 llvm_unreachable("unexpected DepType!");
1098 bool MemoryDepChecker::Dependence::isPossiblyBackward() const {
1099 return isBackward() || Type == Unknown;
1102 bool MemoryDepChecker::Dependence::isForward() const {
1105 case ForwardButPreventsForwarding:
1110 case BackwardVectorizable:
1112 case BackwardVectorizableButPreventsForwarding:
1115 llvm_unreachable("unexpected DepType!");
1118 bool MemoryDepChecker::couldPreventStoreLoadForward(uint64_t Distance,
1119 uint64_t TypeByteSize) {
1120 // If loads occur at a distance that is not a multiple of a feasible vector
1121 // factor store-load forwarding does not take place.
1122 // Positive dependences might cause troubles because vectorizing them might
1123 // prevent store-load forwarding making vectorized code run a lot slower.
1124 // a[i] = a[i-3] ^ a[i-8];
1125 // The stores to a[i:i+1] don't align with the stores to a[i-3:i-2] and
1126 // hence on your typical architecture store-load forwarding does not take
1127 // place. Vectorizing in such cases does not make sense.
1128 // Store-load forwarding distance.
1130 // After this many iterations store-to-load forwarding conflicts should not
1131 // cause any slowdowns.
1132 const uint64_t NumItersForStoreLoadThroughMemory = 8 * TypeByteSize;
1133 // Maximum vector factor.
1134 uint64_t MaxVFWithoutSLForwardIssues = std::min(
1135 VectorizerParams::MaxVectorWidth * TypeByteSize, MaxSafeDepDistBytes);
1137 // Compute the smallest VF at which the store and load would be misaligned.
1138 for (uint64_t VF = 2 * TypeByteSize; VF <= MaxVFWithoutSLForwardIssues;
1140 // If the number of vector iteration between the store and the load are
1141 // small we could incur conflicts.
1142 if (Distance % VF && Distance / VF < NumItersForStoreLoadThroughMemory) {
1143 MaxVFWithoutSLForwardIssues = (VF >>= 1);
1148 if (MaxVFWithoutSLForwardIssues < 2 * TypeByteSize) {
1149 DEBUG(dbgs() << "LAA: Distance " << Distance
1150 << " that could cause a store-load forwarding conflict\n");
1154 if (MaxVFWithoutSLForwardIssues < MaxSafeDepDistBytes &&
1155 MaxVFWithoutSLForwardIssues !=
1156 VectorizerParams::MaxVectorWidth * TypeByteSize)
1157 MaxSafeDepDistBytes = MaxVFWithoutSLForwardIssues;
1161 /// \brief Check the dependence for two accesses with the same stride \p Stride.
1162 /// \p Distance is the positive distance and \p TypeByteSize is type size in
1165 /// \returns true if they are independent.
1166 static bool areStridedAccessesIndependent(uint64_t Distance, uint64_t Stride,
1167 uint64_t TypeByteSize) {
1168 assert(Stride > 1 && "The stride must be greater than 1");
1169 assert(TypeByteSize > 0 && "The type size in byte must be non-zero");
1170 assert(Distance > 0 && "The distance must be non-zero");
1172 // Skip if the distance is not multiple of type byte size.
1173 if (Distance % TypeByteSize)
1176 uint64_t ScaledDist = Distance / TypeByteSize;
1178 // No dependence if the scaled distance is not multiple of the stride.
1180 // for (i = 0; i < 1024 ; i += 4)
1181 // A[i+2] = A[i] + 1;
1183 // Two accesses in memory (scaled distance is 2, stride is 4):
1184 // | A[0] | | | | A[4] | | | |
1185 // | | | A[2] | | | | A[6] | |
1188 // for (i = 0; i < 1024 ; i += 3)
1189 // A[i+4] = A[i] + 1;
1191 // Two accesses in memory (scaled distance is 4, stride is 3):
1192 // | A[0] | | | A[3] | | | A[6] | | |
1193 // | | | | | A[4] | | | A[7] | |
1194 return ScaledDist % Stride;
1197 MemoryDepChecker::Dependence::DepType
1198 MemoryDepChecker::isDependent(const MemAccessInfo &A, unsigned AIdx,
1199 const MemAccessInfo &B, unsigned BIdx,
1200 const ValueToValueMap &Strides) {
1201 assert (AIdx < BIdx && "Must pass arguments in program order");
1203 Value *APtr = A.getPointer();
1204 Value *BPtr = B.getPointer();
1205 bool AIsWrite = A.getInt();
1206 bool BIsWrite = B.getInt();
1208 // Two reads are independent.
1209 if (!AIsWrite && !BIsWrite)
1210 return Dependence::NoDep;
1212 // We cannot check pointers in different address spaces.
1213 if (APtr->getType()->getPointerAddressSpace() !=
1214 BPtr->getType()->getPointerAddressSpace())
1215 return Dependence::Unknown;
1217 int64_t StrideAPtr = getPtrStride(PSE, APtr, InnermostLoop, Strides, true);
1218 int64_t StrideBPtr = getPtrStride(PSE, BPtr, InnermostLoop, Strides, true);
1220 const SCEV *Src = PSE.getSCEV(APtr);
1221 const SCEV *Sink = PSE.getSCEV(BPtr);
1223 // If the induction step is negative we have to invert source and sink of the
1225 if (StrideAPtr < 0) {
1226 std::swap(APtr, BPtr);
1227 std::swap(Src, Sink);
1228 std::swap(AIsWrite, BIsWrite);
1229 std::swap(AIdx, BIdx);
1230 std::swap(StrideAPtr, StrideBPtr);
1233 const SCEV *Dist = PSE.getSE()->getMinusSCEV(Sink, Src);
1235 DEBUG(dbgs() << "LAA: Src Scev: " << *Src << "Sink Scev: " << *Sink
1236 << "(Induction step: " << StrideAPtr << ")\n");
1237 DEBUG(dbgs() << "LAA: Distance for " << *InstMap[AIdx] << " to "
1238 << *InstMap[BIdx] << ": " << *Dist << "\n");
1240 // Need accesses with constant stride. We don't want to vectorize
1241 // "A[B[i]] += ..." and similar code or pointer arithmetic that could wrap in
1242 // the address space.
1243 if (!StrideAPtr || !StrideBPtr || StrideAPtr != StrideBPtr){
1244 DEBUG(dbgs() << "Pointer access with non-constant stride\n");
1245 return Dependence::Unknown;
1248 const SCEVConstant *C = dyn_cast<SCEVConstant>(Dist);
1250 DEBUG(dbgs() << "LAA: Dependence because of non-constant distance\n");
1251 ShouldRetryWithRuntimeCheck = true;
1252 return Dependence::Unknown;
1255 Type *ATy = APtr->getType()->getPointerElementType();
1256 Type *BTy = BPtr->getType()->getPointerElementType();
1257 auto &DL = InnermostLoop->getHeader()->getModule()->getDataLayout();
1258 uint64_t TypeByteSize = DL.getTypeAllocSize(ATy);
1260 const APInt &Val = C->getAPInt();
1261 int64_t Distance = Val.getSExtValue();
1262 uint64_t Stride = std::abs(StrideAPtr);
1264 // Attempt to prove strided accesses independent.
1265 if (std::abs(Distance) > 0 && Stride > 1 && ATy == BTy &&
1266 areStridedAccessesIndependent(std::abs(Distance), Stride, TypeByteSize)) {
1267 DEBUG(dbgs() << "LAA: Strided accesses are independent\n");
1268 return Dependence::NoDep;
1271 // Negative distances are not plausible dependencies.
1272 if (Val.isNegative()) {
1273 bool IsTrueDataDependence = (AIsWrite && !BIsWrite);
1274 if (IsTrueDataDependence && EnableForwardingConflictDetection &&
1275 (couldPreventStoreLoadForward(Val.abs().getZExtValue(), TypeByteSize) ||
1277 DEBUG(dbgs() << "LAA: Forward but may prevent st->ld forwarding\n");
1278 return Dependence::ForwardButPreventsForwarding;
1281 DEBUG(dbgs() << "LAA: Dependence is negative\n");
1282 return Dependence::Forward;
1285 // Write to the same location with the same size.
1286 // Could be improved to assert type sizes are the same (i32 == float, etc).
1289 return Dependence::Forward;
1290 DEBUG(dbgs() << "LAA: Zero dependence difference but different types\n");
1291 return Dependence::Unknown;
1294 assert(Val.isStrictlyPositive() && "Expect a positive value");
1298 "LAA: ReadWrite-Write positive dependency with different types\n");
1299 return Dependence::Unknown;
1302 // Bail out early if passed-in parameters make vectorization not feasible.
1303 unsigned ForcedFactor = (VectorizerParams::VectorizationFactor ?
1304 VectorizerParams::VectorizationFactor : 1);
1305 unsigned ForcedUnroll = (VectorizerParams::VectorizationInterleave ?
1306 VectorizerParams::VectorizationInterleave : 1);
1307 // The minimum number of iterations for a vectorized/unrolled version.
1308 unsigned MinNumIter = std::max(ForcedFactor * ForcedUnroll, 2U);
1310 // It's not vectorizable if the distance is smaller than the minimum distance
1311 // needed for a vectroized/unrolled version. Vectorizing one iteration in
1312 // front needs TypeByteSize * Stride. Vectorizing the last iteration needs
1313 // TypeByteSize (No need to plus the last gap distance).
1315 // E.g. Assume one char is 1 byte in memory and one int is 4 bytes.
1317 // int *B = (int *)((char *)A + 14);
1318 // for (i = 0 ; i < 1024 ; i += 2)
1322 // Two accesses in memory (stride is 2):
1323 // | A[0] | | A[2] | | A[4] | | A[6] | |
1324 // | B[0] | | B[2] | | B[4] |
1326 // Distance needs for vectorizing iterations except the last iteration:
1327 // 4 * 2 * (MinNumIter - 1). Distance needs for the last iteration: 4.
1328 // So the minimum distance needed is: 4 * 2 * (MinNumIter - 1) + 4.
1330 // If MinNumIter is 2, it is vectorizable as the minimum distance needed is
1331 // 12, which is less than distance.
1333 // If MinNumIter is 4 (Say if a user forces the vectorization factor to be 4),
1334 // the minimum distance needed is 28, which is greater than distance. It is
1335 // not safe to do vectorization.
1336 uint64_t MinDistanceNeeded =
1337 TypeByteSize * Stride * (MinNumIter - 1) + TypeByteSize;
1338 if (MinDistanceNeeded > static_cast<uint64_t>(Distance)) {
1339 DEBUG(dbgs() << "LAA: Failure because of positive distance " << Distance
1341 return Dependence::Backward;
1344 // Unsafe if the minimum distance needed is greater than max safe distance.
1345 if (MinDistanceNeeded > MaxSafeDepDistBytes) {
1346 DEBUG(dbgs() << "LAA: Failure because it needs at least "
1347 << MinDistanceNeeded << " size in bytes");
1348 return Dependence::Backward;
1351 // Positive distance bigger than max vectorization factor.
1352 // FIXME: Should use max factor instead of max distance in bytes, which could
1353 // not handle different types.
1354 // E.g. Assume one char is 1 byte in memory and one int is 4 bytes.
1355 // void foo (int *A, char *B) {
1356 // for (unsigned i = 0; i < 1024; i++) {
1357 // A[i+2] = A[i] + 1;
1358 // B[i+2] = B[i] + 1;
1362 // This case is currently unsafe according to the max safe distance. If we
1363 // analyze the two accesses on array B, the max safe dependence distance
1364 // is 2. Then we analyze the accesses on array A, the minimum distance needed
1365 // is 8, which is less than 2 and forbidden vectorization, But actually
1366 // both A and B could be vectorized by 2 iterations.
1367 MaxSafeDepDistBytes =
1368 std::min(static_cast<uint64_t>(Distance), MaxSafeDepDistBytes);
1370 bool IsTrueDataDependence = (!AIsWrite && BIsWrite);
1371 if (IsTrueDataDependence && EnableForwardingConflictDetection &&
1372 couldPreventStoreLoadForward(Distance, TypeByteSize))
1373 return Dependence::BackwardVectorizableButPreventsForwarding;
1375 DEBUG(dbgs() << "LAA: Positive distance " << Val.getSExtValue()
1376 << " with max VF = "
1377 << MaxSafeDepDistBytes / (TypeByteSize * Stride) << '\n');
1379 return Dependence::BackwardVectorizable;
1382 bool MemoryDepChecker::areDepsSafe(DepCandidates &AccessSets,
1383 MemAccessInfoSet &CheckDeps,
1384 const ValueToValueMap &Strides) {
1386 MaxSafeDepDistBytes = -1;
1387 while (!CheckDeps.empty()) {
1388 MemAccessInfo CurAccess = *CheckDeps.begin();
1390 // Get the relevant memory access set.
1391 EquivalenceClasses<MemAccessInfo>::iterator I =
1392 AccessSets.findValue(AccessSets.getLeaderValue(CurAccess));
1394 // Check accesses within this set.
1395 EquivalenceClasses<MemAccessInfo>::member_iterator AI =
1396 AccessSets.member_begin(I);
1397 EquivalenceClasses<MemAccessInfo>::member_iterator AE =
1398 AccessSets.member_end();
1400 // Check every access pair.
1402 CheckDeps.erase(*AI);
1403 EquivalenceClasses<MemAccessInfo>::member_iterator OI = std::next(AI);
1405 // Check every accessing instruction pair in program order.
1406 for (std::vector<unsigned>::iterator I1 = Accesses[*AI].begin(),
1407 I1E = Accesses[*AI].end(); I1 != I1E; ++I1)
1408 for (std::vector<unsigned>::iterator I2 = Accesses[*OI].begin(),
1409 I2E = Accesses[*OI].end(); I2 != I2E; ++I2) {
1410 auto A = std::make_pair(&*AI, *I1);
1411 auto B = std::make_pair(&*OI, *I2);
1417 Dependence::DepType Type =
1418 isDependent(*A.first, A.second, *B.first, B.second, Strides);
1419 SafeForVectorization &= Dependence::isSafeForVectorization(Type);
1421 // Gather dependences unless we accumulated MaxDependences
1422 // dependences. In that case return as soon as we find the first
1423 // unsafe dependence. This puts a limit on this quadratic
1425 if (RecordDependences) {
1426 if (Type != Dependence::NoDep)
1427 Dependences.push_back(Dependence(A.second, B.second, Type));
1429 if (Dependences.size() >= MaxDependences) {
1430 RecordDependences = false;
1431 Dependences.clear();
1432 DEBUG(dbgs() << "Too many dependences, stopped recording\n");
1435 if (!RecordDependences && !SafeForVectorization)
1444 DEBUG(dbgs() << "Total Dependences: " << Dependences.size() << "\n");
1445 return SafeForVectorization;
1448 SmallVector<Instruction *, 4>
1449 MemoryDepChecker::getInstructionsForAccess(Value *Ptr, bool isWrite) const {
1450 MemAccessInfo Access(Ptr, isWrite);
1451 auto &IndexVector = Accesses.find(Access)->second;
1453 SmallVector<Instruction *, 4> Insts;
1454 std::transform(IndexVector.begin(), IndexVector.end(),
1455 std::back_inserter(Insts),
1456 [&](unsigned Idx) { return this->InstMap[Idx]; });
1460 const char *MemoryDepChecker::Dependence::DepName[] = {
1461 "NoDep", "Unknown", "Forward", "ForwardButPreventsForwarding", "Backward",
1462 "BackwardVectorizable", "BackwardVectorizableButPreventsForwarding"};
1464 void MemoryDepChecker::Dependence::print(
1465 raw_ostream &OS, unsigned Depth,
1466 const SmallVectorImpl<Instruction *> &Instrs) const {
1467 OS.indent(Depth) << DepName[Type] << ":\n";
1468 OS.indent(Depth + 2) << *Instrs[Source] << " -> \n";
1469 OS.indent(Depth + 2) << *Instrs[Destination] << "\n";
1472 bool LoopAccessInfo::canAnalyzeLoop() {
1473 // We need to have a loop header.
1474 DEBUG(dbgs() << "LAA: Found a loop in "
1475 << TheLoop->getHeader()->getParent()->getName() << ": "
1476 << TheLoop->getHeader()->getName() << '\n');
1478 // We can only analyze innermost loops.
1479 if (!TheLoop->empty()) {
1480 DEBUG(dbgs() << "LAA: loop is not the innermost loop\n");
1481 emitAnalysis(LoopAccessReport() << "loop is not the innermost loop");
1485 // We must have a single backedge.
1486 if (TheLoop->getNumBackEdges() != 1) {
1487 DEBUG(dbgs() << "LAA: loop control flow is not understood by analyzer\n");
1489 LoopAccessReport() <<
1490 "loop control flow is not understood by analyzer");
1494 // We must have a single exiting block.
1495 if (!TheLoop->getExitingBlock()) {
1496 DEBUG(dbgs() << "LAA: loop control flow is not understood by analyzer\n");
1498 LoopAccessReport() <<
1499 "loop control flow is not understood by analyzer");
1503 // We only handle bottom-tested loops, i.e. loop in which the condition is
1504 // checked at the end of each iteration. With that we can assume that all
1505 // instructions in the loop are executed the same number of times.
1506 if (TheLoop->getExitingBlock() != TheLoop->getLoopLatch()) {
1507 DEBUG(dbgs() << "LAA: loop control flow is not understood by analyzer\n");
1509 LoopAccessReport() <<
1510 "loop control flow is not understood by analyzer");
1514 // ScalarEvolution needs to be able to find the exit count.
1515 const SCEV *ExitCount = PSE->getBackedgeTakenCount();
1516 if (ExitCount == PSE->getSE()->getCouldNotCompute()) {
1517 emitAnalysis(LoopAccessReport()
1518 << "could not determine number of loop iterations");
1519 DEBUG(dbgs() << "LAA: SCEV could not compute the loop exit count.\n");
1526 void LoopAccessInfo::analyzeLoop(AliasAnalysis *AA, LoopInfo *LI,
1527 const TargetLibraryInfo *TLI,
1528 DominatorTree *DT) {
1529 typedef SmallPtrSet<Value*, 16> ValueSet;
1531 // Holds the Load and Store instructions.
1532 SmallVector<LoadInst *, 16> Loads;
1533 SmallVector<StoreInst *, 16> Stores;
1535 // Holds all the different accesses in the loop.
1536 unsigned NumReads = 0;
1537 unsigned NumReadWrites = 0;
1539 PtrRtChecking->Pointers.clear();
1540 PtrRtChecking->Need = false;
1542 const bool IsAnnotatedParallel = TheLoop->isAnnotatedParallel();
1545 for (BasicBlock *BB : TheLoop->blocks()) {
1546 // Scan the BB and collect legal loads and stores.
1547 for (Instruction &I : *BB) {
1548 // If this is a load, save it. If this instruction can read from memory
1549 // but is not a load, then we quit. Notice that we don't handle function
1550 // calls that read or write.
1551 if (I.mayReadFromMemory()) {
1552 // Many math library functions read the rounding mode. We will only
1553 // vectorize a loop if it contains known function calls that don't set
1554 // the flag. Therefore, it is safe to ignore this read from memory.
1555 auto *Call = dyn_cast<CallInst>(&I);
1556 if (Call && getVectorIntrinsicIDForCall(Call, TLI))
1559 // If the function has an explicit vectorized counterpart, we can safely
1560 // assume that it can be vectorized.
1561 if (Call && !Call->isNoBuiltin() && Call->getCalledFunction() &&
1562 TLI->isFunctionVectorizable(Call->getCalledFunction()->getName()))
1565 auto *Ld = dyn_cast<LoadInst>(&I);
1566 if (!Ld || (!Ld->isSimple() && !IsAnnotatedParallel)) {
1567 emitAnalysis(LoopAccessReport(Ld)
1568 << "read with atomic ordering or volatile read");
1569 DEBUG(dbgs() << "LAA: Found a non-simple load.\n");
1574 Loads.push_back(Ld);
1575 DepChecker->addAccess(Ld);
1576 if (EnableMemAccessVersioning)
1577 collectStridedAccess(Ld);
1581 // Save 'store' instructions. Abort if other instructions write to memory.
1582 if (I.mayWriteToMemory()) {
1583 auto *St = dyn_cast<StoreInst>(&I);
1585 emitAnalysis(LoopAccessReport(St)
1586 << "instruction cannot be vectorized");
1590 if (!St->isSimple() && !IsAnnotatedParallel) {
1591 emitAnalysis(LoopAccessReport(St)
1592 << "write with atomic ordering or volatile write");
1593 DEBUG(dbgs() << "LAA: Found a non-simple store.\n");
1598 Stores.push_back(St);
1599 DepChecker->addAccess(St);
1600 if (EnableMemAccessVersioning)
1601 collectStridedAccess(St);
1606 // Now we have two lists that hold the loads and the stores.
1607 // Next, we find the pointers that they use.
1609 // Check if we see any stores. If there are no stores, then we don't
1610 // care if the pointers are *restrict*.
1611 if (!Stores.size()) {
1612 DEBUG(dbgs() << "LAA: Found a read-only loop!\n");
1617 MemoryDepChecker::DepCandidates DependentAccesses;
1618 AccessAnalysis Accesses(TheLoop->getHeader()->getModule()->getDataLayout(),
1619 AA, LI, DependentAccesses, *PSE);
1621 // Holds the analyzed pointers. We don't want to call GetUnderlyingObjects
1622 // multiple times on the same object. If the ptr is accessed twice, once
1623 // for read and once for write, it will only appear once (on the write
1624 // list). This is okay, since we are going to check for conflicts between
1625 // writes and between reads and writes, but not between reads and reads.
1628 for (StoreInst *ST : Stores) {
1629 Value *Ptr = ST->getPointerOperand();
1630 // Check for store to loop invariant address.
1631 StoreToLoopInvariantAddress |= isUniform(Ptr);
1632 // If we did *not* see this pointer before, insert it to the read-write
1633 // list. At this phase it is only a 'write' list.
1634 if (Seen.insert(Ptr).second) {
1637 MemoryLocation Loc = MemoryLocation::get(ST);
1638 // The TBAA metadata could have a control dependency on the predication
1639 // condition, so we cannot rely on it when determining whether or not we
1640 // need runtime pointer checks.
1641 if (blockNeedsPredication(ST->getParent(), TheLoop, DT))
1642 Loc.AATags.TBAA = nullptr;
1644 Accesses.addStore(Loc);
1648 if (IsAnnotatedParallel) {
1650 << "LAA: A loop annotated parallel, ignore memory dependency "
1656 for (LoadInst *LD : Loads) {
1657 Value *Ptr = LD->getPointerOperand();
1658 // If we did *not* see this pointer before, insert it to the
1659 // read list. If we *did* see it before, then it is already in
1660 // the read-write list. This allows us to vectorize expressions
1661 // such as A[i] += x; Because the address of A[i] is a read-write
1662 // pointer. This only works if the index of A[i] is consecutive.
1663 // If the address of i is unknown (for example A[B[i]]) then we may
1664 // read a few words, modify, and write a few words, and some of the
1665 // words may be written to the same address.
1666 bool IsReadOnlyPtr = false;
1667 if (Seen.insert(Ptr).second ||
1668 !getPtrStride(*PSE, Ptr, TheLoop, SymbolicStrides)) {
1670 IsReadOnlyPtr = true;
1673 MemoryLocation Loc = MemoryLocation::get(LD);
1674 // The TBAA metadata could have a control dependency on the predication
1675 // condition, so we cannot rely on it when determining whether or not we
1676 // need runtime pointer checks.
1677 if (blockNeedsPredication(LD->getParent(), TheLoop, DT))
1678 Loc.AATags.TBAA = nullptr;
1680 Accesses.addLoad(Loc, IsReadOnlyPtr);
1683 // If we write (or read-write) to a single destination and there are no
1684 // other reads in this loop then is it safe to vectorize.
1685 if (NumReadWrites == 1 && NumReads == 0) {
1686 DEBUG(dbgs() << "LAA: Found a write-only loop!\n");
1691 // Build dependence sets and check whether we need a runtime pointer bounds
1693 Accesses.buildDependenceSets();
1695 // Find pointers with computable bounds. We are going to use this information
1696 // to place a runtime bound check.
1697 bool CanDoRTIfNeeded = Accesses.canCheckPtrAtRT(*PtrRtChecking, PSE->getSE(),
1698 TheLoop, SymbolicStrides);
1699 if (!CanDoRTIfNeeded) {
1700 emitAnalysis(LoopAccessReport() << "cannot identify array bounds");
1701 DEBUG(dbgs() << "LAA: We can't vectorize because we can't find "
1702 << "the array bounds.\n");
1707 DEBUG(dbgs() << "LAA: We can perform a memory runtime check if needed.\n");
1710 if (Accesses.isDependencyCheckNeeded()) {
1711 DEBUG(dbgs() << "LAA: Checking memory dependencies\n");
1712 CanVecMem = DepChecker->areDepsSafe(
1713 DependentAccesses, Accesses.getDependenciesToCheck(), SymbolicStrides);
1714 MaxSafeDepDistBytes = DepChecker->getMaxSafeDepDistBytes();
1716 if (!CanVecMem && DepChecker->shouldRetryWithRuntimeCheck()) {
1717 DEBUG(dbgs() << "LAA: Retrying with memory checks\n");
1719 // Clear the dependency checks. We assume they are not needed.
1720 Accesses.resetDepChecks(*DepChecker);
1722 PtrRtChecking->reset();
1723 PtrRtChecking->Need = true;
1725 auto *SE = PSE->getSE();
1726 CanDoRTIfNeeded = Accesses.canCheckPtrAtRT(*PtrRtChecking, SE, TheLoop,
1727 SymbolicStrides, true);
1729 // Check that we found the bounds for the pointer.
1730 if (!CanDoRTIfNeeded) {
1731 emitAnalysis(LoopAccessReport()
1732 << "cannot check memory dependencies at runtime");
1733 DEBUG(dbgs() << "LAA: Can't vectorize with memory checks\n");
1743 DEBUG(dbgs() << "LAA: No unsafe dependent memory operations in loop. We"
1744 << (PtrRtChecking->Need ? "" : " don't")
1745 << " need runtime memory checks.\n");
1749 << "unsafe dependent memory operations in loop. Use "
1750 "#pragma loop distribute(enable) to allow loop distribution "
1751 "to attempt to isolate the offending operations into a separate "
1753 DEBUG(dbgs() << "LAA: unsafe dependent memory operations in loop\n");
1757 bool LoopAccessInfo::blockNeedsPredication(BasicBlock *BB, Loop *TheLoop,
1758 DominatorTree *DT) {
1759 assert(TheLoop->contains(BB) && "Unknown block used");
1761 // Blocks that do not dominate the latch need predication.
1762 BasicBlock* Latch = TheLoop->getLoopLatch();
1763 return !DT->dominates(BB, Latch);
1766 void LoopAccessInfo::emitAnalysis(LoopAccessReport &Message) {
1767 assert(!Report && "Multiple reports generated");
1771 bool LoopAccessInfo::isUniform(Value *V) const {
1772 return (PSE->getSE()->isLoopInvariant(PSE->getSE()->getSCEV(V), TheLoop));
1775 // FIXME: this function is currently a duplicate of the one in
1776 // LoopVectorize.cpp.
1777 static Instruction *getFirstInst(Instruction *FirstInst, Value *V,
1781 if (Instruction *I = dyn_cast<Instruction>(V))
1782 return I->getParent() == Loc->getParent() ? I : nullptr;
1787 /// \brief IR Values for the lower and upper bounds of a pointer evolution. We
1788 /// need to use value-handles because SCEV expansion can invalidate previously
1789 /// expanded values. Thus expansion of a pointer can invalidate the bounds for
1791 struct PointerBounds {
1792 TrackingVH<Value> Start;
1793 TrackingVH<Value> End;
1795 } // end anonymous namespace
1797 /// \brief Expand code for the lower and upper bound of the pointer group \p CG
1798 /// in \p TheLoop. \return the values for the bounds.
1799 static PointerBounds
1800 expandBounds(const RuntimePointerChecking::CheckingPtrGroup *CG, Loop *TheLoop,
1801 Instruction *Loc, SCEVExpander &Exp, ScalarEvolution *SE,
1802 const RuntimePointerChecking &PtrRtChecking) {
1803 Value *Ptr = PtrRtChecking.Pointers[CG->Members[0]].PointerValue;
1804 const SCEV *Sc = SE->getSCEV(Ptr);
1806 if (SE->isLoopInvariant(Sc, TheLoop)) {
1807 DEBUG(dbgs() << "LAA: Adding RT check for a loop invariant ptr:" << *Ptr
1811 unsigned AS = Ptr->getType()->getPointerAddressSpace();
1812 LLVMContext &Ctx = Loc->getContext();
1814 // Use this type for pointer arithmetic.
1815 Type *PtrArithTy = Type::getInt8PtrTy(Ctx, AS);
1816 Value *Start = nullptr, *End = nullptr;
1818 DEBUG(dbgs() << "LAA: Adding RT check for range:\n");
1819 Start = Exp.expandCodeFor(CG->Low, PtrArithTy, Loc);
1820 End = Exp.expandCodeFor(CG->High, PtrArithTy, Loc);
1821 DEBUG(dbgs() << "Start: " << *CG->Low << " End: " << *CG->High << "\n");
1822 return {Start, End};
1826 /// \brief Turns a collection of checks into a collection of expanded upper and
1827 /// lower bounds for both pointers in the check.
1828 static SmallVector<std::pair<PointerBounds, PointerBounds>, 4> expandBounds(
1829 const SmallVectorImpl<RuntimePointerChecking::PointerCheck> &PointerChecks,
1830 Loop *L, Instruction *Loc, ScalarEvolution *SE, SCEVExpander &Exp,
1831 const RuntimePointerChecking &PtrRtChecking) {
1832 SmallVector<std::pair<PointerBounds, PointerBounds>, 4> ChecksWithBounds;
1834 // Here we're relying on the SCEV Expander's cache to only emit code for the
1835 // same bounds once.
1837 PointerChecks.begin(), PointerChecks.end(),
1838 std::back_inserter(ChecksWithBounds),
1839 [&](const RuntimePointerChecking::PointerCheck &Check) {
1841 First = expandBounds(Check.first, L, Loc, Exp, SE, PtrRtChecking),
1842 Second = expandBounds(Check.second, L, Loc, Exp, SE, PtrRtChecking);
1843 return std::make_pair(First, Second);
1846 return ChecksWithBounds;
1849 std::pair<Instruction *, Instruction *> LoopAccessInfo::addRuntimeChecks(
1851 const SmallVectorImpl<RuntimePointerChecking::PointerCheck> &PointerChecks)
1853 const DataLayout &DL = TheLoop->getHeader()->getModule()->getDataLayout();
1854 auto *SE = PSE->getSE();
1855 SCEVExpander Exp(*SE, DL, "induction");
1856 auto ExpandedChecks =
1857 expandBounds(PointerChecks, TheLoop, Loc, SE, Exp, *PtrRtChecking);
1859 LLVMContext &Ctx = Loc->getContext();
1860 Instruction *FirstInst = nullptr;
1861 IRBuilder<> ChkBuilder(Loc);
1862 // Our instructions might fold to a constant.
1863 Value *MemoryRuntimeCheck = nullptr;
1865 for (const auto &Check : ExpandedChecks) {
1866 const PointerBounds &A = Check.first, &B = Check.second;
1867 // Check if two pointers (A and B) conflict where conflict is computed as:
1868 // start(A) <= end(B) && start(B) <= end(A)
1869 unsigned AS0 = A.Start->getType()->getPointerAddressSpace();
1870 unsigned AS1 = B.Start->getType()->getPointerAddressSpace();
1872 assert((AS0 == B.End->getType()->getPointerAddressSpace()) &&
1873 (AS1 == A.End->getType()->getPointerAddressSpace()) &&
1874 "Trying to bounds check pointers with different address spaces");
1876 Type *PtrArithTy0 = Type::getInt8PtrTy(Ctx, AS0);
1877 Type *PtrArithTy1 = Type::getInt8PtrTy(Ctx, AS1);
1879 Value *Start0 = ChkBuilder.CreateBitCast(A.Start, PtrArithTy0, "bc");
1880 Value *Start1 = ChkBuilder.CreateBitCast(B.Start, PtrArithTy1, "bc");
1881 Value *End0 = ChkBuilder.CreateBitCast(A.End, PtrArithTy1, "bc");
1882 Value *End1 = ChkBuilder.CreateBitCast(B.End, PtrArithTy0, "bc");
1884 // [A|B].Start points to the first accessed byte under base [A|B].
1885 // [A|B].End points to the last accessed byte, plus one.
1886 // There is no conflict when the intervals are disjoint:
1887 // NoConflict = (B.Start >= A.End) || (A.Start >= B.End)
1889 // bound0 = (B.Start < A.End)
1890 // bound1 = (A.Start < B.End)
1891 // IsConflict = bound0 & bound1
1892 Value *Cmp0 = ChkBuilder.CreateICmpULT(Start0, End1, "bound0");
1893 FirstInst = getFirstInst(FirstInst, Cmp0, Loc);
1894 Value *Cmp1 = ChkBuilder.CreateICmpULT(Start1, End0, "bound1");
1895 FirstInst = getFirstInst(FirstInst, Cmp1, Loc);
1896 Value *IsConflict = ChkBuilder.CreateAnd(Cmp0, Cmp1, "found.conflict");
1897 FirstInst = getFirstInst(FirstInst, IsConflict, Loc);
1898 if (MemoryRuntimeCheck) {
1900 ChkBuilder.CreateOr(MemoryRuntimeCheck, IsConflict, "conflict.rdx");
1901 FirstInst = getFirstInst(FirstInst, IsConflict, Loc);
1903 MemoryRuntimeCheck = IsConflict;
1906 if (!MemoryRuntimeCheck)
1907 return std::make_pair(nullptr, nullptr);
1909 // We have to do this trickery because the IRBuilder might fold the check to a
1910 // constant expression in which case there is no Instruction anchored in a
1912 Instruction *Check = BinaryOperator::CreateAnd(MemoryRuntimeCheck,
1913 ConstantInt::getTrue(Ctx));
1914 ChkBuilder.Insert(Check, "memcheck.conflict");
1915 FirstInst = getFirstInst(FirstInst, Check, Loc);
1916 return std::make_pair(FirstInst, Check);
1919 std::pair<Instruction *, Instruction *>
1920 LoopAccessInfo::addRuntimeChecks(Instruction *Loc) const {
1921 if (!PtrRtChecking->Need)
1922 return std::make_pair(nullptr, nullptr);
1924 return addRuntimeChecks(Loc, PtrRtChecking->getChecks());
1927 void LoopAccessInfo::collectStridedAccess(Value *MemAccess) {
1928 Value *Ptr = nullptr;
1929 if (LoadInst *LI = dyn_cast<LoadInst>(MemAccess))
1930 Ptr = LI->getPointerOperand();
1931 else if (StoreInst *SI = dyn_cast<StoreInst>(MemAccess))
1932 Ptr = SI->getPointerOperand();
1936 Value *Stride = getStrideFromPointer(Ptr, PSE->getSE(), TheLoop);
1940 DEBUG(dbgs() << "LAA: Found a strided access that we can version");
1941 DEBUG(dbgs() << " Ptr: " << *Ptr << " Stride: " << *Stride << "\n");
1942 SymbolicStrides[Ptr] = Stride;
1943 StrideSet.insert(Stride);
1946 LoopAccessInfo::LoopAccessInfo(Loop *L, ScalarEvolution *SE,
1947 const TargetLibraryInfo *TLI, AliasAnalysis *AA,
1948 DominatorTree *DT, LoopInfo *LI)
1949 : PSE(llvm::make_unique<PredicatedScalarEvolution>(*SE, *L)),
1950 PtrRtChecking(llvm::make_unique<RuntimePointerChecking>(SE)),
1951 DepChecker(llvm::make_unique<MemoryDepChecker>(*PSE, L)), TheLoop(L),
1952 NumLoads(0), NumStores(0), MaxSafeDepDistBytes(-1), CanVecMem(false),
1953 StoreToLoopInvariantAddress(false) {
1954 if (canAnalyzeLoop())
1955 analyzeLoop(AA, LI, TLI, DT);
1958 void LoopAccessInfo::print(raw_ostream &OS, unsigned Depth) const {
1960 OS.indent(Depth) << "Memory dependences are safe";
1961 if (MaxSafeDepDistBytes != -1ULL)
1962 OS << " with a maximum dependence distance of " << MaxSafeDepDistBytes
1964 if (PtrRtChecking->Need)
1965 OS << " with run-time checks";
1970 OS.indent(Depth) << "Report: " << Report->str() << "\n";
1972 if (auto *Dependences = DepChecker->getDependences()) {
1973 OS.indent(Depth) << "Dependences:\n";
1974 for (auto &Dep : *Dependences) {
1975 Dep.print(OS, Depth + 2, DepChecker->getMemoryInstructions());
1979 OS.indent(Depth) << "Too many dependences, not recorded\n";
1981 // List the pair of accesses need run-time checks to prove independence.
1982 PtrRtChecking->print(OS, Depth);
1985 OS.indent(Depth) << "Store to invariant address was "
1986 << (StoreToLoopInvariantAddress ? "" : "not ")
1987 << "found in loop.\n";
1989 OS.indent(Depth) << "SCEV assumptions:\n";
1990 PSE->getUnionPredicate().print(OS, Depth);
1994 OS.indent(Depth) << "Expressions re-written:\n";
1995 PSE->print(OS, Depth);
1998 const LoopAccessInfo &LoopAccessLegacyAnalysis::getInfo(Loop *L) {
1999 auto &LAI = LoopAccessInfoMap[L];
2002 LAI = llvm::make_unique<LoopAccessInfo>(L, SE, TLI, AA, DT, LI);
2007 void LoopAccessLegacyAnalysis::print(raw_ostream &OS, const Module *M) const {
2008 LoopAccessLegacyAnalysis &LAA = *const_cast<LoopAccessLegacyAnalysis *>(this);
2010 for (Loop *TopLevelLoop : *LI)
2011 for (Loop *L : depth_first(TopLevelLoop)) {
2012 OS.indent(2) << L->getHeader()->getName() << ":\n";
2013 auto &LAI = LAA.getInfo(L);
2018 bool LoopAccessLegacyAnalysis::runOnFunction(Function &F) {
2019 SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE();
2020 auto *TLIP = getAnalysisIfAvailable<TargetLibraryInfoWrapperPass>();
2021 TLI = TLIP ? &TLIP->getTLI() : nullptr;
2022 AA = &getAnalysis<AAResultsWrapperPass>().getAAResults();
2023 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
2024 LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
2029 void LoopAccessLegacyAnalysis::getAnalysisUsage(AnalysisUsage &AU) const {
2030 AU.addRequired<ScalarEvolutionWrapperPass>();
2031 AU.addRequired<AAResultsWrapperPass>();
2032 AU.addRequired<DominatorTreeWrapperPass>();
2033 AU.addRequired<LoopInfoWrapperPass>();
2035 AU.setPreservesAll();
2038 char LoopAccessLegacyAnalysis::ID = 0;
2039 static const char laa_name[] = "Loop Access Analysis";
2040 #define LAA_NAME "loop-accesses"
2042 INITIALIZE_PASS_BEGIN(LoopAccessLegacyAnalysis, LAA_NAME, laa_name, false, true)
2043 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)
2044 INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass)
2045 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
2046 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
2047 INITIALIZE_PASS_END(LoopAccessLegacyAnalysis, LAA_NAME, laa_name, false, true)
2049 char LoopAccessAnalysis::PassID;
2051 LoopAccessInfo LoopAccessAnalysis::run(Loop &L, AnalysisManager<Loop> &AM) {
2052 const AnalysisManager<Function> &FAM =
2053 AM.getResult<FunctionAnalysisManagerLoopProxy>(L).getManager();
2054 Function &F = *L.getHeader()->getParent();
2055 auto *SE = FAM.getCachedResult<ScalarEvolutionAnalysis>(F);
2056 auto *TLI = FAM.getCachedResult<TargetLibraryAnalysis>(F);
2057 auto *AA = FAM.getCachedResult<AAManager>(F);
2058 auto *DT = FAM.getCachedResult<DominatorTreeAnalysis>(F);
2059 auto *LI = FAM.getCachedResult<LoopAnalysis>(F);
2062 "ScalarEvolution must have been cached at a higher level");
2064 report_fatal_error("AliasAnalysis must have been cached at a higher level");
2066 report_fatal_error("DominatorTree must have been cached at a higher level");
2068 report_fatal_error("LoopInfo must have been cached at a higher level");
2069 return LoopAccessInfo(&L, SE, TLI, AA, DT, LI);
2072 PreservedAnalyses LoopAccessInfoPrinterPass::run(Loop &L,
2073 AnalysisManager<Loop> &AM) {
2074 Function &F = *L.getHeader()->getParent();
2075 auto &LAI = AM.getResult<LoopAccessAnalysis>(L);
2076 OS << "Loop access info in function '" << F.getName() << "':\n";
2077 OS.indent(2) << L.getHeader()->getName() << ":\n";
2079 return PreservedAnalyses::all();
2083 Pass *createLAAPass() {
2084 return new LoopAccessLegacyAnalysis();