1 //===- LoopAccessAnalysis.cpp - Loop Access Analysis Implementation --------==//
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
7 //===----------------------------------------------------------------------===//
9 // The implementation for the loop memory dependence that was originally
10 // developed for the loop vectorizer.
12 //===----------------------------------------------------------------------===//
14 #include "llvm/Analysis/LoopAccessAnalysis.h"
15 #include "llvm/ADT/APInt.h"
16 #include "llvm/ADT/DenseMap.h"
17 #include "llvm/ADT/DepthFirstIterator.h"
18 #include "llvm/ADT/EquivalenceClasses.h"
19 #include "llvm/ADT/PointerIntPair.h"
20 #include "llvm/ADT/STLExtras.h"
21 #include "llvm/ADT/SetVector.h"
22 #include "llvm/ADT/SmallPtrSet.h"
23 #include "llvm/ADT/SmallSet.h"
24 #include "llvm/ADT/SmallVector.h"
25 #include "llvm/ADT/iterator_range.h"
26 #include "llvm/Analysis/AliasAnalysis.h"
27 #include "llvm/Analysis/AliasSetTracker.h"
28 #include "llvm/Analysis/LoopAnalysisManager.h"
29 #include "llvm/Analysis/LoopInfo.h"
30 #include "llvm/Analysis/MemoryLocation.h"
31 #include "llvm/Analysis/OptimizationRemarkEmitter.h"
32 #include "llvm/Analysis/ScalarEvolution.h"
33 #include "llvm/Analysis/ScalarEvolutionExpander.h"
34 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
35 #include "llvm/Analysis/TargetLibraryInfo.h"
36 #include "llvm/Analysis/ValueTracking.h"
37 #include "llvm/Analysis/VectorUtils.h"
38 #include "llvm/IR/BasicBlock.h"
39 #include "llvm/IR/Constants.h"
40 #include "llvm/IR/DataLayout.h"
41 #include "llvm/IR/DebugLoc.h"
42 #include "llvm/IR/DerivedTypes.h"
43 #include "llvm/IR/DiagnosticInfo.h"
44 #include "llvm/IR/Dominators.h"
45 #include "llvm/IR/Function.h"
46 #include "llvm/IR/IRBuilder.h"
47 #include "llvm/IR/InstrTypes.h"
48 #include "llvm/IR/Instruction.h"
49 #include "llvm/IR/Instructions.h"
50 #include "llvm/IR/Operator.h"
51 #include "llvm/IR/PassManager.h"
52 #include "llvm/IR/Type.h"
53 #include "llvm/IR/Value.h"
54 #include "llvm/IR/ValueHandle.h"
55 #include "llvm/Pass.h"
56 #include "llvm/Support/Casting.h"
57 #include "llvm/Support/CommandLine.h"
58 #include "llvm/Support/Debug.h"
59 #include "llvm/Support/ErrorHandling.h"
60 #include "llvm/Support/raw_ostream.h"
71 #define DEBUG_TYPE "loop-accesses"
73 static cl::opt
<unsigned, true>
74 VectorizationFactor("force-vector-width", cl::Hidden
,
75 cl::desc("Sets the SIMD width. Zero is autoselect."),
76 cl::location(VectorizerParams::VectorizationFactor
));
77 unsigned VectorizerParams::VectorizationFactor
;
79 static cl::opt
<unsigned, true>
80 VectorizationInterleave("force-vector-interleave", cl::Hidden
,
81 cl::desc("Sets the vectorization interleave count. "
82 "Zero is autoselect."),
84 VectorizerParams::VectorizationInterleave
));
85 unsigned VectorizerParams::VectorizationInterleave
;
87 static cl::opt
<unsigned, true> RuntimeMemoryCheckThreshold(
88 "runtime-memory-check-threshold", cl::Hidden
,
89 cl::desc("When performing memory disambiguation checks at runtime do not "
90 "generate more than this number of comparisons (default = 8)."),
91 cl::location(VectorizerParams::RuntimeMemoryCheckThreshold
), cl::init(8));
92 unsigned VectorizerParams::RuntimeMemoryCheckThreshold
;
94 /// The maximum iterations used to merge memory checks
95 static cl::opt
<unsigned> MemoryCheckMergeThreshold(
96 "memory-check-merge-threshold", cl::Hidden
,
97 cl::desc("Maximum number of comparisons done when trying to merge "
98 "runtime memory checks. (default = 100)"),
101 /// Maximum SIMD width.
102 const unsigned VectorizerParams::MaxVectorWidth
= 64;
104 /// We collect dependences up to this threshold.
105 static cl::opt
<unsigned>
106 MaxDependences("max-dependences", cl::Hidden
,
107 cl::desc("Maximum number of dependences collected by "
108 "loop-access analysis (default = 100)"),
111 /// This enables versioning on the strides of symbolically striding memory
112 /// accesses in code like the following.
113 /// for (i = 0; i < N; ++i)
114 /// A[i * Stride1] += B[i * Stride2] ...
116 /// Will be roughly translated to
117 /// if (Stride1 == 1 && Stride2 == 1) {
118 /// for (i = 0; i < N; i+=4)
122 static cl::opt
<bool> EnableMemAccessVersioning(
123 "enable-mem-access-versioning", cl::init(true), cl::Hidden
,
124 cl::desc("Enable symbolic stride memory access versioning"));
126 /// Enable store-to-load forwarding conflict detection. This option can
127 /// be disabled for correctness testing.
128 static cl::opt
<bool> EnableForwardingConflictDetection(
129 "store-to-load-forwarding-conflict-detection", cl::Hidden
,
130 cl::desc("Enable conflict detection in loop-access analysis"),
133 bool VectorizerParams::isInterleaveForced() {
134 return ::VectorizationInterleave
.getNumOccurrences() > 0;
137 Value
*llvm::stripIntegerCast(Value
*V
) {
138 if (auto *CI
= dyn_cast
<CastInst
>(V
))
139 if (CI
->getOperand(0)->getType()->isIntegerTy())
140 return CI
->getOperand(0);
144 const SCEV
*llvm::replaceSymbolicStrideSCEV(PredicatedScalarEvolution
&PSE
,
145 const ValueToValueMap
&PtrToStride
,
146 Value
*Ptr
, Value
*OrigPtr
) {
147 const SCEV
*OrigSCEV
= PSE
.getSCEV(Ptr
);
149 // If there is an entry in the map return the SCEV of the pointer with the
150 // symbolic stride replaced by one.
151 ValueToValueMap::const_iterator SI
=
152 PtrToStride
.find(OrigPtr
? OrigPtr
: Ptr
);
153 if (SI
!= PtrToStride
.end()) {
154 Value
*StrideVal
= SI
->second
;
157 StrideVal
= stripIntegerCast(StrideVal
);
159 ScalarEvolution
*SE
= PSE
.getSE();
160 const auto *U
= cast
<SCEVUnknown
>(SE
->getSCEV(StrideVal
));
162 static_cast<const SCEVConstant
*>(SE
->getOne(StrideVal
->getType()));
164 PSE
.addPredicate(*SE
->getEqualPredicate(U
, CT
));
165 auto *Expr
= PSE
.getSCEV(Ptr
);
167 LLVM_DEBUG(dbgs() << "LAA: Replacing SCEV: " << *OrigSCEV
168 << " by: " << *Expr
<< "\n");
172 // Otherwise, just return the SCEV of the original pointer.
176 /// Calculate Start and End points of memory access.
177 /// Let's assume A is the first access and B is a memory access on N-th loop
178 /// iteration. Then B is calculated as:
180 /// Step value may be positive or negative.
181 /// N is a calculated back-edge taken count:
182 /// N = (TripCount > 0) ? RoundDown(TripCount -1 , VF) : 0
183 /// Start and End points are calculated in the following way:
184 /// Start = UMIN(A, B) ; End = UMAX(A, B) + SizeOfElt,
185 /// where SizeOfElt is the size of single memory access in bytes.
187 /// There is no conflict when the intervals are disjoint:
188 /// NoConflict = (P2.Start >= P1.End) || (P1.Start >= P2.End)
189 void RuntimePointerChecking::insert(Loop
*Lp
, Value
*Ptr
, bool WritePtr
,
190 unsigned DepSetId
, unsigned ASId
,
191 const ValueToValueMap
&Strides
,
192 PredicatedScalarEvolution
&PSE
) {
193 // Get the stride replaced scev.
194 const SCEV
*Sc
= replaceSymbolicStrideSCEV(PSE
, Strides
, Ptr
);
195 ScalarEvolution
*SE
= PSE
.getSE();
200 if (SE
->isLoopInvariant(Sc
, Lp
))
201 ScStart
= ScEnd
= Sc
;
203 const SCEVAddRecExpr
*AR
= dyn_cast
<SCEVAddRecExpr
>(Sc
);
204 assert(AR
&& "Invalid addrec expression");
205 const SCEV
*Ex
= PSE
.getBackedgeTakenCount();
207 ScStart
= AR
->getStart();
208 ScEnd
= AR
->evaluateAtIteration(Ex
, *SE
);
209 const SCEV
*Step
= AR
->getStepRecurrence(*SE
);
211 // For expressions with negative step, the upper bound is ScStart and the
212 // lower bound is ScEnd.
213 if (const auto *CStep
= dyn_cast
<SCEVConstant
>(Step
)) {
214 if (CStep
->getValue()->isNegative())
215 std::swap(ScStart
, ScEnd
);
217 // Fallback case: the step is not constant, but we can still
218 // get the upper and lower bounds of the interval by using min/max
220 ScStart
= SE
->getUMinExpr(ScStart
, ScEnd
);
221 ScEnd
= SE
->getUMaxExpr(AR
->getStart(), ScEnd
);
223 // Add the size of the pointed element to ScEnd.
225 Ptr
->getType()->getPointerElementType()->getScalarSizeInBits() / 8;
226 const SCEV
*EltSizeSCEV
= SE
->getConstant(ScEnd
->getType(), EltSize
);
227 ScEnd
= SE
->getAddExpr(ScEnd
, EltSizeSCEV
);
230 Pointers
.emplace_back(Ptr
, ScStart
, ScEnd
, WritePtr
, DepSetId
, ASId
, Sc
);
233 SmallVector
<RuntimePointerChecking::PointerCheck
, 4>
234 RuntimePointerChecking::generateChecks() const {
235 SmallVector
<PointerCheck
, 4> Checks
;
237 for (unsigned I
= 0; I
< CheckingGroups
.size(); ++I
) {
238 for (unsigned J
= I
+ 1; J
< CheckingGroups
.size(); ++J
) {
239 const RuntimePointerChecking::CheckingPtrGroup
&CGI
= CheckingGroups
[I
];
240 const RuntimePointerChecking::CheckingPtrGroup
&CGJ
= CheckingGroups
[J
];
242 if (needsChecking(CGI
, CGJ
))
243 Checks
.push_back(std::make_pair(&CGI
, &CGJ
));
249 void RuntimePointerChecking::generateChecks(
250 MemoryDepChecker::DepCandidates
&DepCands
, bool UseDependencies
) {
251 assert(Checks
.empty() && "Checks is not empty");
252 groupChecks(DepCands
, UseDependencies
);
253 Checks
= generateChecks();
256 bool RuntimePointerChecking::needsChecking(const CheckingPtrGroup
&M
,
257 const CheckingPtrGroup
&N
) const {
258 for (unsigned I
= 0, EI
= M
.Members
.size(); EI
!= I
; ++I
)
259 for (unsigned J
= 0, EJ
= N
.Members
.size(); EJ
!= J
; ++J
)
260 if (needsChecking(M
.Members
[I
], N
.Members
[J
]))
265 /// Compare \p I and \p J and return the minimum.
266 /// Return nullptr in case we couldn't find an answer.
267 static const SCEV
*getMinFromExprs(const SCEV
*I
, const SCEV
*J
,
268 ScalarEvolution
*SE
) {
269 const SCEV
*Diff
= SE
->getMinusSCEV(J
, I
);
270 const SCEVConstant
*C
= dyn_cast
<const SCEVConstant
>(Diff
);
274 if (C
->getValue()->isNegative())
279 bool RuntimePointerChecking::CheckingPtrGroup::addPointer(unsigned Index
) {
280 const SCEV
*Start
= RtCheck
.Pointers
[Index
].Start
;
281 const SCEV
*End
= RtCheck
.Pointers
[Index
].End
;
283 // Compare the starts and ends with the known minimum and maximum
284 // of this set. We need to know how we compare against the min/max
285 // of the set in order to be able to emit memchecks.
286 const SCEV
*Min0
= getMinFromExprs(Start
, Low
, RtCheck
.SE
);
290 const SCEV
*Min1
= getMinFromExprs(End
, High
, RtCheck
.SE
);
294 // Update the low bound expression if we've found a new min value.
298 // Update the high bound expression if we've found a new max value.
302 Members
.push_back(Index
);
306 void RuntimePointerChecking::groupChecks(
307 MemoryDepChecker::DepCandidates
&DepCands
, bool UseDependencies
) {
308 // We build the groups from dependency candidates equivalence classes
310 // - We know that pointers in the same equivalence class share
311 // the same underlying object and therefore there is a chance
312 // that we can compare pointers
313 // - We wouldn't be able to merge two pointers for which we need
314 // to emit a memcheck. The classes in DepCands are already
315 // conveniently built such that no two pointers in the same
316 // class need checking against each other.
318 // We use the following (greedy) algorithm to construct the groups
319 // For every pointer in the equivalence class:
320 // For each existing group:
321 // - if the difference between this pointer and the min/max bounds
322 // of the group is a constant, then make the pointer part of the
323 // group and update the min/max bounds of that group as required.
325 CheckingGroups
.clear();
327 // If we need to check two pointers to the same underlying object
328 // with a non-constant difference, we shouldn't perform any pointer
329 // grouping with those pointers. This is because we can easily get
330 // into cases where the resulting check would return false, even when
331 // the accesses are safe.
333 // The following example shows this:
334 // for (i = 0; i < 1000; ++i)
335 // a[5000 + i * m] = a[i] + a[i + 9000]
337 // Here grouping gives a check of (5000, 5000 + 1000 * m) against
338 // (0, 10000) which is always false. However, if m is 1, there is no
339 // dependence. Not grouping the checks for a[i] and a[i + 9000] allows
340 // us to perform an accurate check in this case.
342 // The above case requires that we have an UnknownDependence between
343 // accesses to the same underlying object. This cannot happen unless
344 // FoundNonConstantDistanceDependence is set, and therefore UseDependencies
345 // is also false. In this case we will use the fallback path and create
346 // separate checking groups for all pointers.
348 // If we don't have the dependency partitions, construct a new
349 // checking pointer group for each pointer. This is also required
350 // for correctness, because in this case we can have checking between
351 // pointers to the same underlying object.
352 if (!UseDependencies
) {
353 for (unsigned I
= 0; I
< Pointers
.size(); ++I
)
354 CheckingGroups
.push_back(CheckingPtrGroup(I
, *this));
358 unsigned TotalComparisons
= 0;
360 DenseMap
<Value
*, unsigned> PositionMap
;
361 for (unsigned Index
= 0; Index
< Pointers
.size(); ++Index
)
362 PositionMap
[Pointers
[Index
].PointerValue
] = Index
;
364 // We need to keep track of what pointers we've already seen so we
365 // don't process them twice.
366 SmallSet
<unsigned, 2> Seen
;
368 // Go through all equivalence classes, get the "pointer check groups"
369 // and add them to the overall solution. We use the order in which accesses
370 // appear in 'Pointers' to enforce determinism.
371 for (unsigned I
= 0; I
< Pointers
.size(); ++I
) {
372 // We've seen this pointer before, and therefore already processed
373 // its equivalence class.
377 MemoryDepChecker::MemAccessInfo
Access(Pointers
[I
].PointerValue
,
378 Pointers
[I
].IsWritePtr
);
380 SmallVector
<CheckingPtrGroup
, 2> Groups
;
381 auto LeaderI
= DepCands
.findValue(DepCands
.getLeaderValue(Access
));
383 // Because DepCands is constructed by visiting accesses in the order in
384 // which they appear in alias sets (which is deterministic) and the
385 // iteration order within an equivalence class member is only dependent on
386 // the order in which unions and insertions are performed on the
387 // equivalence class, the iteration order is deterministic.
388 for (auto MI
= DepCands
.member_begin(LeaderI
), ME
= DepCands
.member_end();
390 unsigned Pointer
= PositionMap
[MI
->getPointer()];
392 // Mark this pointer as seen.
393 Seen
.insert(Pointer
);
395 // Go through all the existing sets and see if we can find one
396 // which can include this pointer.
397 for (CheckingPtrGroup
&Group
: Groups
) {
398 // Don't perform more than a certain amount of comparisons.
399 // This should limit the cost of grouping the pointers to something
400 // reasonable. If we do end up hitting this threshold, the algorithm
401 // will create separate groups for all remaining pointers.
402 if (TotalComparisons
> MemoryCheckMergeThreshold
)
407 if (Group
.addPointer(Pointer
)) {
414 // We couldn't add this pointer to any existing set or the threshold
415 // for the number of comparisons has been reached. Create a new group
416 // to hold the current pointer.
417 Groups
.push_back(CheckingPtrGroup(Pointer
, *this));
420 // We've computed the grouped checks for this partition.
421 // Save the results and continue with the next one.
422 llvm::copy(Groups
, std::back_inserter(CheckingGroups
));
426 bool RuntimePointerChecking::arePointersInSamePartition(
427 const SmallVectorImpl
<int> &PtrToPartition
, unsigned PtrIdx1
,
429 return (PtrToPartition
[PtrIdx1
] != -1 &&
430 PtrToPartition
[PtrIdx1
] == PtrToPartition
[PtrIdx2
]);
433 bool RuntimePointerChecking::needsChecking(unsigned I
, unsigned J
) const {
434 const PointerInfo
&PointerI
= Pointers
[I
];
435 const PointerInfo
&PointerJ
= Pointers
[J
];
437 // No need to check if two readonly pointers intersect.
438 if (!PointerI
.IsWritePtr
&& !PointerJ
.IsWritePtr
)
441 // Only need to check pointers between two different dependency sets.
442 if (PointerI
.DependencySetId
== PointerJ
.DependencySetId
)
445 // Only need to check pointers in the same alias set.
446 if (PointerI
.AliasSetId
!= PointerJ
.AliasSetId
)
452 void RuntimePointerChecking::printChecks(
453 raw_ostream
&OS
, const SmallVectorImpl
<PointerCheck
> &Checks
,
454 unsigned Depth
) const {
456 for (const auto &Check
: Checks
) {
457 const auto &First
= Check
.first
->Members
, &Second
= Check
.second
->Members
;
459 OS
.indent(Depth
) << "Check " << N
++ << ":\n";
461 OS
.indent(Depth
+ 2) << "Comparing group (" << Check
.first
<< "):\n";
462 for (unsigned K
= 0; K
< First
.size(); ++K
)
463 OS
.indent(Depth
+ 2) << *Pointers
[First
[K
]].PointerValue
<< "\n";
465 OS
.indent(Depth
+ 2) << "Against group (" << Check
.second
<< "):\n";
466 for (unsigned K
= 0; K
< Second
.size(); ++K
)
467 OS
.indent(Depth
+ 2) << *Pointers
[Second
[K
]].PointerValue
<< "\n";
471 void RuntimePointerChecking::print(raw_ostream
&OS
, unsigned Depth
) const {
473 OS
.indent(Depth
) << "Run-time memory checks:\n";
474 printChecks(OS
, Checks
, Depth
);
476 OS
.indent(Depth
) << "Grouped accesses:\n";
477 for (unsigned I
= 0; I
< CheckingGroups
.size(); ++I
) {
478 const auto &CG
= CheckingGroups
[I
];
480 OS
.indent(Depth
+ 2) << "Group " << &CG
<< ":\n";
481 OS
.indent(Depth
+ 4) << "(Low: " << *CG
.Low
<< " High: " << *CG
.High
483 for (unsigned J
= 0; J
< CG
.Members
.size(); ++J
) {
484 OS
.indent(Depth
+ 6) << "Member: " << *Pointers
[CG
.Members
[J
]].Expr
492 /// Analyses memory accesses in a loop.
494 /// Checks whether run time pointer checks are needed and builds sets for data
495 /// dependence checking.
496 class AccessAnalysis
{
498 /// Read or write access location.
499 typedef PointerIntPair
<Value
*, 1, bool> MemAccessInfo
;
500 typedef SmallVector
<MemAccessInfo
, 8> MemAccessInfoList
;
502 AccessAnalysis(const DataLayout
&Dl
, Loop
*TheLoop
, AliasAnalysis
*AA
,
503 LoopInfo
*LI
, MemoryDepChecker::DepCandidates
&DA
,
504 PredicatedScalarEvolution
&PSE
)
505 : DL(Dl
), TheLoop(TheLoop
), AST(*AA
), LI(LI
), DepCands(DA
),
506 IsRTCheckAnalysisNeeded(false), PSE(PSE
) {}
508 /// Register a load and whether it is only read from.
509 void addLoad(MemoryLocation
&Loc
, bool IsReadOnly
) {
510 Value
*Ptr
= const_cast<Value
*>(Loc
.Ptr
);
511 AST
.add(Ptr
, LocationSize::unknown(), Loc
.AATags
);
512 Accesses
.insert(MemAccessInfo(Ptr
, false));
514 ReadOnlyPtr
.insert(Ptr
);
517 /// Register a store.
518 void addStore(MemoryLocation
&Loc
) {
519 Value
*Ptr
= const_cast<Value
*>(Loc
.Ptr
);
520 AST
.add(Ptr
, LocationSize::unknown(), Loc
.AATags
);
521 Accesses
.insert(MemAccessInfo(Ptr
, true));
524 /// Check if we can emit a run-time no-alias check for \p Access.
526 /// Returns true if we can emit a run-time no alias check for \p Access.
527 /// If we can check this access, this also adds it to a dependence set and
528 /// adds a run-time to check for it to \p RtCheck. If \p Assume is true,
529 /// we will attempt to use additional run-time checks in order to get
530 /// the bounds of the pointer.
531 bool createCheckForAccess(RuntimePointerChecking
&RtCheck
,
532 MemAccessInfo Access
,
533 const ValueToValueMap
&Strides
,
534 DenseMap
<Value
*, unsigned> &DepSetId
,
535 Loop
*TheLoop
, unsigned &RunningDepId
,
536 unsigned ASId
, bool ShouldCheckStride
,
539 /// Check whether we can check the pointers at runtime for
540 /// non-intersection.
542 /// Returns true if we need no check or if we do and we can generate them
543 /// (i.e. the pointers have computable bounds).
544 bool canCheckPtrAtRT(RuntimePointerChecking
&RtCheck
, ScalarEvolution
*SE
,
545 Loop
*TheLoop
, const ValueToValueMap
&Strides
,
546 bool ShouldCheckWrap
= false);
548 /// Goes over all memory accesses, checks whether a RT check is needed
549 /// and builds sets of dependent accesses.
550 void buildDependenceSets() {
551 processMemAccesses();
554 /// Initial processing of memory accesses determined that we need to
555 /// perform dependency checking.
557 /// Note that this can later be cleared if we retry memcheck analysis without
558 /// dependency checking (i.e. FoundNonConstantDistanceDependence).
559 bool isDependencyCheckNeeded() { return !CheckDeps
.empty(); }
561 /// We decided that no dependence analysis would be used. Reset the state.
562 void resetDepChecks(MemoryDepChecker
&DepChecker
) {
564 DepChecker
.clearDependences();
567 MemAccessInfoList
&getDependenciesToCheck() { return CheckDeps
; }
570 typedef SetVector
<MemAccessInfo
> PtrAccessSet
;
572 /// Go over all memory access and check whether runtime pointer checks
573 /// are needed and build sets of dependency check candidates.
574 void processMemAccesses();
576 /// Set of all accesses.
577 PtrAccessSet Accesses
;
579 const DataLayout
&DL
;
581 /// The loop being checked.
584 /// List of accesses that need a further dependence check.
585 MemAccessInfoList CheckDeps
;
587 /// Set of pointers that are read only.
588 SmallPtrSet
<Value
*, 16> ReadOnlyPtr
;
590 /// An alias set tracker to partition the access set by underlying object and
591 //intrinsic property (such as TBAA metadata).
596 /// Sets of potentially dependent accesses - members of one set share an
597 /// underlying pointer. The set "CheckDeps" identfies which sets really need a
598 /// dependence check.
599 MemoryDepChecker::DepCandidates
&DepCands
;
601 /// Initial processing of memory accesses determined that we may need
602 /// to add memchecks. Perform the analysis to determine the necessary checks.
604 /// Note that, this is different from isDependencyCheckNeeded. When we retry
605 /// memcheck analysis without dependency checking
606 /// (i.e. FoundNonConstantDistanceDependence), isDependencyCheckNeeded is
607 /// cleared while this remains set if we have potentially dependent accesses.
608 bool IsRTCheckAnalysisNeeded
;
610 /// The SCEV predicate containing all the SCEV-related assumptions.
611 PredicatedScalarEvolution
&PSE
;
614 } // end anonymous namespace
616 /// Check whether a pointer can participate in a runtime bounds check.
617 /// If \p Assume, try harder to prove that we can compute the bounds of \p Ptr
618 /// by adding run-time checks (overflow checks) if necessary.
619 static bool hasComputableBounds(PredicatedScalarEvolution
&PSE
,
620 const ValueToValueMap
&Strides
, Value
*Ptr
,
621 Loop
*L
, bool Assume
) {
622 const SCEV
*PtrScev
= replaceSymbolicStrideSCEV(PSE
, Strides
, Ptr
);
624 // The bounds for loop-invariant pointer is trivial.
625 if (PSE
.getSE()->isLoopInvariant(PtrScev
, L
))
628 const SCEVAddRecExpr
*AR
= dyn_cast
<SCEVAddRecExpr
>(PtrScev
);
631 AR
= PSE
.getAsAddRec(Ptr
);
636 return AR
->isAffine();
639 /// Check whether a pointer address cannot wrap.
640 static bool isNoWrap(PredicatedScalarEvolution
&PSE
,
641 const ValueToValueMap
&Strides
, Value
*Ptr
, Loop
*L
) {
642 const SCEV
*PtrScev
= PSE
.getSCEV(Ptr
);
643 if (PSE
.getSE()->isLoopInvariant(PtrScev
, L
))
646 int64_t Stride
= getPtrStride(PSE
, Ptr
, L
, Strides
);
647 if (Stride
== 1 || PSE
.hasNoOverflow(Ptr
, SCEVWrapPredicate::IncrementNUSW
))
653 bool AccessAnalysis::createCheckForAccess(RuntimePointerChecking
&RtCheck
,
654 MemAccessInfo Access
,
655 const ValueToValueMap
&StridesMap
,
656 DenseMap
<Value
*, unsigned> &DepSetId
,
657 Loop
*TheLoop
, unsigned &RunningDepId
,
658 unsigned ASId
, bool ShouldCheckWrap
,
660 Value
*Ptr
= Access
.getPointer();
662 if (!hasComputableBounds(PSE
, StridesMap
, Ptr
, TheLoop
, Assume
))
665 // When we run after a failing dependency check we have to make sure
666 // we don't have wrapping pointers.
667 if (ShouldCheckWrap
&& !isNoWrap(PSE
, StridesMap
, Ptr
, TheLoop
)) {
668 auto *Expr
= PSE
.getSCEV(Ptr
);
669 if (!Assume
|| !isa
<SCEVAddRecExpr
>(Expr
))
671 PSE
.setNoOverflow(Ptr
, SCEVWrapPredicate::IncrementNUSW
);
674 // The id of the dependence set.
677 if (isDependencyCheckNeeded()) {
678 Value
*Leader
= DepCands
.getLeaderValue(Access
).getPointer();
679 unsigned &LeaderId
= DepSetId
[Leader
];
681 LeaderId
= RunningDepId
++;
684 // Each access has its own dependence set.
685 DepId
= RunningDepId
++;
687 bool IsWrite
= Access
.getInt();
688 RtCheck
.insert(TheLoop
, Ptr
, IsWrite
, DepId
, ASId
, StridesMap
, PSE
);
689 LLVM_DEBUG(dbgs() << "LAA: Found a runtime check ptr:" << *Ptr
<< '\n');
694 bool AccessAnalysis::canCheckPtrAtRT(RuntimePointerChecking
&RtCheck
,
695 ScalarEvolution
*SE
, Loop
*TheLoop
,
696 const ValueToValueMap
&StridesMap
,
697 bool ShouldCheckWrap
) {
698 // Find pointers with computable bounds. We are going to use this information
699 // to place a runtime bound check.
702 bool NeedRTCheck
= false;
703 if (!IsRTCheckAnalysisNeeded
) return true;
705 bool IsDepCheckNeeded
= isDependencyCheckNeeded();
707 // We assign a consecutive id to access from different alias sets.
708 // Accesses between different groups doesn't need to be checked.
710 for (auto &AS
: AST
) {
711 int NumReadPtrChecks
= 0;
712 int NumWritePtrChecks
= 0;
713 bool CanDoAliasSetRT
= true;
715 // We assign consecutive id to access from different dependence sets.
716 // Accesses within the same set don't need a runtime check.
717 unsigned RunningDepId
= 1;
718 DenseMap
<Value
*, unsigned> DepSetId
;
720 SmallVector
<MemAccessInfo
, 4> Retries
;
723 Value
*Ptr
= A
.getValue();
724 bool IsWrite
= Accesses
.count(MemAccessInfo(Ptr
, true));
725 MemAccessInfo
Access(Ptr
, IsWrite
);
732 if (!createCheckForAccess(RtCheck
, Access
, StridesMap
, DepSetId
, TheLoop
,
733 RunningDepId
, ASId
, ShouldCheckWrap
, false)) {
734 LLVM_DEBUG(dbgs() << "LAA: Can't find bounds for ptr:" << *Ptr
<< '\n');
735 Retries
.push_back(Access
);
736 CanDoAliasSetRT
= false;
740 // If we have at least two writes or one write and a read then we need to
741 // check them. But there is no need to checks if there is only one
742 // dependence set for this alias set.
744 // Note that this function computes CanDoRT and NeedRTCheck independently.
745 // For example CanDoRT=false, NeedRTCheck=false means that we have a pointer
746 // for which we couldn't find the bounds but we don't actually need to emit
747 // any checks so it does not matter.
748 bool NeedsAliasSetRTCheck
= false;
749 if (!(IsDepCheckNeeded
&& CanDoAliasSetRT
&& RunningDepId
== 2))
750 NeedsAliasSetRTCheck
= (NumWritePtrChecks
>= 2 ||
751 (NumReadPtrChecks
>= 1 && NumWritePtrChecks
>= 1));
753 // We need to perform run-time alias checks, but some pointers had bounds
754 // that couldn't be checked.
755 if (NeedsAliasSetRTCheck
&& !CanDoAliasSetRT
) {
756 // Reset the CanDoSetRt flag and retry all accesses that have failed.
757 // We know that we need these checks, so we can now be more aggressive
758 // and add further checks if required (overflow checks).
759 CanDoAliasSetRT
= true;
760 for (auto Access
: Retries
)
761 if (!createCheckForAccess(RtCheck
, Access
, StridesMap
, DepSetId
,
762 TheLoop
, RunningDepId
, ASId
,
763 ShouldCheckWrap
, /*Assume=*/true)) {
764 CanDoAliasSetRT
= false;
769 CanDoRT
&= CanDoAliasSetRT
;
770 NeedRTCheck
|= NeedsAliasSetRTCheck
;
774 // If the pointers that we would use for the bounds comparison have different
775 // address spaces, assume the values aren't directly comparable, so we can't
776 // use them for the runtime check. We also have to assume they could
777 // overlap. In the future there should be metadata for whether address spaces
779 unsigned NumPointers
= RtCheck
.Pointers
.size();
780 for (unsigned i
= 0; i
< NumPointers
; ++i
) {
781 for (unsigned j
= i
+ 1; j
< NumPointers
; ++j
) {
782 // Only need to check pointers between two different dependency sets.
783 if (RtCheck
.Pointers
[i
].DependencySetId
==
784 RtCheck
.Pointers
[j
].DependencySetId
)
786 // Only need to check pointers in the same alias set.
787 if (RtCheck
.Pointers
[i
].AliasSetId
!= RtCheck
.Pointers
[j
].AliasSetId
)
790 Value
*PtrI
= RtCheck
.Pointers
[i
].PointerValue
;
791 Value
*PtrJ
= RtCheck
.Pointers
[j
].PointerValue
;
793 unsigned ASi
= PtrI
->getType()->getPointerAddressSpace();
794 unsigned ASj
= PtrJ
->getType()->getPointerAddressSpace();
797 dbgs() << "LAA: Runtime check would require comparison between"
798 " different address spaces\n");
804 if (NeedRTCheck
&& CanDoRT
)
805 RtCheck
.generateChecks(DepCands
, IsDepCheckNeeded
);
807 LLVM_DEBUG(dbgs() << "LAA: We need to do " << RtCheck
.getNumberOfChecks()
808 << " pointer comparisons.\n");
810 RtCheck
.Need
= NeedRTCheck
;
812 bool CanDoRTIfNeeded
= !NeedRTCheck
|| CanDoRT
;
813 if (!CanDoRTIfNeeded
)
815 return CanDoRTIfNeeded
;
818 void AccessAnalysis::processMemAccesses() {
819 // We process the set twice: first we process read-write pointers, last we
820 // process read-only pointers. This allows us to skip dependence tests for
821 // read-only pointers.
823 LLVM_DEBUG(dbgs() << "LAA: Processing memory accesses...\n");
824 LLVM_DEBUG(dbgs() << " AST: "; AST
.dump());
825 LLVM_DEBUG(dbgs() << "LAA: Accesses(" << Accesses
.size() << "):\n");
827 for (auto A
: Accesses
)
828 dbgs() << "\t" << *A
.getPointer() << " (" <<
829 (A
.getInt() ? "write" : (ReadOnlyPtr
.count(A
.getPointer()) ?
830 "read-only" : "read")) << ")\n";
833 // The AliasSetTracker has nicely partitioned our pointers by metadata
834 // compatibility and potential for underlying-object overlap. As a result, we
835 // only need to check for potential pointer dependencies within each alias
837 for (auto &AS
: AST
) {
838 // Note that both the alias-set tracker and the alias sets themselves used
839 // linked lists internally and so the iteration order here is deterministic
840 // (matching the original instruction order within each set).
842 bool SetHasWrite
= false;
844 // Map of pointers to last access encountered.
845 typedef DenseMap
<const Value
*, MemAccessInfo
> UnderlyingObjToAccessMap
;
846 UnderlyingObjToAccessMap ObjToLastAccess
;
848 // Set of access to check after all writes have been processed.
849 PtrAccessSet DeferredAccesses
;
851 // Iterate over each alias set twice, once to process read/write pointers,
852 // and then to process read-only pointers.
853 for (int SetIteration
= 0; SetIteration
< 2; ++SetIteration
) {
854 bool UseDeferred
= SetIteration
> 0;
855 PtrAccessSet
&S
= UseDeferred
? DeferredAccesses
: Accesses
;
858 Value
*Ptr
= AV
.getValue();
860 // For a single memory access in AliasSetTracker, Accesses may contain
861 // both read and write, and they both need to be handled for CheckDeps.
863 if (AC
.getPointer() != Ptr
)
866 bool IsWrite
= AC
.getInt();
868 // If we're using the deferred access set, then it contains only
870 bool IsReadOnlyPtr
= ReadOnlyPtr
.count(Ptr
) && !IsWrite
;
871 if (UseDeferred
&& !IsReadOnlyPtr
)
873 // Otherwise, the pointer must be in the PtrAccessSet, either as a
875 assert(((IsReadOnlyPtr
&& UseDeferred
) || IsWrite
||
876 S
.count(MemAccessInfo(Ptr
, false))) &&
877 "Alias-set pointer not in the access set?");
879 MemAccessInfo
Access(Ptr
, IsWrite
);
880 DepCands
.insert(Access
);
882 // Memorize read-only pointers for later processing and skip them in
883 // the first round (they need to be checked after we have seen all
884 // write pointers). Note: we also mark pointer that are not
885 // consecutive as "read-only" pointers (so that we check
886 // "a[b[i]] +="). Hence, we need the second check for "!IsWrite".
887 if (!UseDeferred
&& IsReadOnlyPtr
) {
888 DeferredAccesses
.insert(Access
);
892 // If this is a write - check other reads and writes for conflicts. If
893 // this is a read only check other writes for conflicts (but only if
894 // there is no other write to the ptr - this is an optimization to
895 // catch "a[i] = a[i] + " without having to do a dependence check).
896 if ((IsWrite
|| IsReadOnlyPtr
) && SetHasWrite
) {
897 CheckDeps
.push_back(Access
);
898 IsRTCheckAnalysisNeeded
= true;
904 // Create sets of pointers connected by a shared alias set and
905 // underlying object.
906 typedef SmallVector
<const Value
*, 16> ValueVector
;
907 ValueVector TempObjects
;
909 GetUnderlyingObjects(Ptr
, TempObjects
, DL
, LI
);
911 << "Underlying objects for pointer " << *Ptr
<< "\n");
912 for (const Value
*UnderlyingObj
: TempObjects
) {
913 // nullptr never alias, don't join sets for pointer that have "null"
914 // in their UnderlyingObjects list.
915 if (isa
<ConstantPointerNull
>(UnderlyingObj
) &&
916 !NullPointerIsDefined(
917 TheLoop
->getHeader()->getParent(),
918 UnderlyingObj
->getType()->getPointerAddressSpace()))
921 UnderlyingObjToAccessMap::iterator Prev
=
922 ObjToLastAccess
.find(UnderlyingObj
);
923 if (Prev
!= ObjToLastAccess
.end())
924 DepCands
.unionSets(Access
, Prev
->second
);
926 ObjToLastAccess
[UnderlyingObj
] = Access
;
927 LLVM_DEBUG(dbgs() << " " << *UnderlyingObj
<< "\n");
935 static bool isInBoundsGep(Value
*Ptr
) {
936 if (GetElementPtrInst
*GEP
= dyn_cast
<GetElementPtrInst
>(Ptr
))
937 return GEP
->isInBounds();
941 /// Return true if an AddRec pointer \p Ptr is unsigned non-wrapping,
942 /// i.e. monotonically increasing/decreasing.
943 static bool isNoWrapAddRec(Value
*Ptr
, const SCEVAddRecExpr
*AR
,
944 PredicatedScalarEvolution
&PSE
, const Loop
*L
) {
945 // FIXME: This should probably only return true for NUW.
946 if (AR
->getNoWrapFlags(SCEV::NoWrapMask
))
949 // Scalar evolution does not propagate the non-wrapping flags to values that
950 // are derived from a non-wrapping induction variable because non-wrapping
951 // could be flow-sensitive.
953 // Look through the potentially overflowing instruction to try to prove
954 // non-wrapping for the *specific* value of Ptr.
956 // The arithmetic implied by an inbounds GEP can't overflow.
957 auto *GEP
= dyn_cast
<GetElementPtrInst
>(Ptr
);
958 if (!GEP
|| !GEP
->isInBounds())
961 // Make sure there is only one non-const index and analyze that.
962 Value
*NonConstIndex
= nullptr;
963 for (Value
*Index
: make_range(GEP
->idx_begin(), GEP
->idx_end()))
964 if (!isa
<ConstantInt
>(Index
)) {
967 NonConstIndex
= Index
;
970 // The recurrence is on the pointer, ignore for now.
973 // The index in GEP is signed. It is non-wrapping if it's derived from a NSW
974 // AddRec using a NSW operation.
975 if (auto *OBO
= dyn_cast
<OverflowingBinaryOperator
>(NonConstIndex
))
976 if (OBO
->hasNoSignedWrap() &&
977 // Assume constant for other the operand so that the AddRec can be
979 isa
<ConstantInt
>(OBO
->getOperand(1))) {
980 auto *OpScev
= PSE
.getSCEV(OBO
->getOperand(0));
982 if (auto *OpAR
= dyn_cast
<SCEVAddRecExpr
>(OpScev
))
983 return OpAR
->getLoop() == L
&& OpAR
->getNoWrapFlags(SCEV::FlagNSW
);
989 /// Check whether the access through \p Ptr has a constant stride.
990 int64_t llvm::getPtrStride(PredicatedScalarEvolution
&PSE
, Value
*Ptr
,
991 const Loop
*Lp
, const ValueToValueMap
&StridesMap
,
992 bool Assume
, bool ShouldCheckWrap
) {
993 Type
*Ty
= Ptr
->getType();
994 assert(Ty
->isPointerTy() && "Unexpected non-ptr");
996 // Make sure that the pointer does not point to aggregate types.
997 auto *PtrTy
= cast
<PointerType
>(Ty
);
998 if (PtrTy
->getElementType()->isAggregateType()) {
999 LLVM_DEBUG(dbgs() << "LAA: Bad stride - Not a pointer to a scalar type"
1004 const SCEV
*PtrScev
= replaceSymbolicStrideSCEV(PSE
, StridesMap
, Ptr
);
1006 const SCEVAddRecExpr
*AR
= dyn_cast
<SCEVAddRecExpr
>(PtrScev
);
1008 AR
= PSE
.getAsAddRec(Ptr
);
1011 LLVM_DEBUG(dbgs() << "LAA: Bad stride - Not an AddRecExpr pointer " << *Ptr
1012 << " SCEV: " << *PtrScev
<< "\n");
1016 // The access function must stride over the innermost loop.
1017 if (Lp
!= AR
->getLoop()) {
1018 LLVM_DEBUG(dbgs() << "LAA: Bad stride - Not striding over innermost loop "
1019 << *Ptr
<< " SCEV: " << *AR
<< "\n");
1023 // The address calculation must not wrap. Otherwise, a dependence could be
1025 // An inbounds getelementptr that is a AddRec with a unit stride
1026 // cannot wrap per definition. The unit stride requirement is checked later.
1027 // An getelementptr without an inbounds attribute and unit stride would have
1028 // to access the pointer value "0" which is undefined behavior in address
1029 // space 0, therefore we can also vectorize this case.
1030 bool IsInBoundsGEP
= isInBoundsGep(Ptr
);
1031 bool IsNoWrapAddRec
= !ShouldCheckWrap
||
1032 PSE
.hasNoOverflow(Ptr
, SCEVWrapPredicate::IncrementNUSW
) ||
1033 isNoWrapAddRec(Ptr
, AR
, PSE
, Lp
);
1034 if (!IsNoWrapAddRec
&& !IsInBoundsGEP
&&
1035 NullPointerIsDefined(Lp
->getHeader()->getParent(),
1036 PtrTy
->getAddressSpace())) {
1038 PSE
.setNoOverflow(Ptr
, SCEVWrapPredicate::IncrementNUSW
);
1039 IsNoWrapAddRec
= true;
1040 LLVM_DEBUG(dbgs() << "LAA: Pointer may wrap in the address space:\n"
1041 << "LAA: Pointer: " << *Ptr
<< "\n"
1042 << "LAA: SCEV: " << *AR
<< "\n"
1043 << "LAA: Added an overflow assumption\n");
1046 dbgs() << "LAA: Bad stride - Pointer may wrap in the address space "
1047 << *Ptr
<< " SCEV: " << *AR
<< "\n");
1052 // Check the step is constant.
1053 const SCEV
*Step
= AR
->getStepRecurrence(*PSE
.getSE());
1055 // Calculate the pointer stride and check if it is constant.
1056 const SCEVConstant
*C
= dyn_cast
<SCEVConstant
>(Step
);
1058 LLVM_DEBUG(dbgs() << "LAA: Bad stride - Not a constant strided " << *Ptr
1059 << " SCEV: " << *AR
<< "\n");
1063 auto &DL
= Lp
->getHeader()->getModule()->getDataLayout();
1064 int64_t Size
= DL
.getTypeAllocSize(PtrTy
->getElementType());
1065 const APInt
&APStepVal
= C
->getAPInt();
1067 // Huge step value - give up.
1068 if (APStepVal
.getBitWidth() > 64)
1071 int64_t StepVal
= APStepVal
.getSExtValue();
1074 int64_t Stride
= StepVal
/ Size
;
1075 int64_t Rem
= StepVal
% Size
;
1079 // If the SCEV could wrap but we have an inbounds gep with a unit stride we
1080 // know we can't "wrap around the address space". In case of address space
1081 // zero we know that this won't happen without triggering undefined behavior.
1082 if (!IsNoWrapAddRec
&& Stride
!= 1 && Stride
!= -1 &&
1083 (IsInBoundsGEP
|| !NullPointerIsDefined(Lp
->getHeader()->getParent(),
1084 PtrTy
->getAddressSpace()))) {
1086 // We can avoid this case by adding a run-time check.
1087 LLVM_DEBUG(dbgs() << "LAA: Non unit strided pointer which is not either "
1088 << "inbounds or in address space 0 may wrap:\n"
1089 << "LAA: Pointer: " << *Ptr
<< "\n"
1090 << "LAA: SCEV: " << *AR
<< "\n"
1091 << "LAA: Added an overflow assumption\n");
1092 PSE
.setNoOverflow(Ptr
, SCEVWrapPredicate::IncrementNUSW
);
1100 bool llvm::sortPtrAccesses(ArrayRef
<Value
*> VL
, const DataLayout
&DL
,
1101 ScalarEvolution
&SE
,
1102 SmallVectorImpl
<unsigned> &SortedIndices
) {
1103 assert(llvm::all_of(
1104 VL
, [](const Value
*V
) { return V
->getType()->isPointerTy(); }) &&
1105 "Expected list of pointer operands.");
1106 SmallVector
<std::pair
<int64_t, Value
*>, 4> OffValPairs
;
1107 OffValPairs
.reserve(VL
.size());
1109 // Walk over the pointers, and map each of them to an offset relative to
1110 // first pointer in the array.
1111 Value
*Ptr0
= VL
[0];
1112 const SCEV
*Scev0
= SE
.getSCEV(Ptr0
);
1113 Value
*Obj0
= GetUnderlyingObject(Ptr0
, DL
);
1115 llvm::SmallSet
<int64_t, 4> Offsets
;
1116 for (auto *Ptr
: VL
) {
1117 // TODO: Outline this code as a special, more time consuming, version of
1118 // computeConstantDifference() function.
1119 if (Ptr
->getType()->getPointerAddressSpace() !=
1120 Ptr0
->getType()->getPointerAddressSpace())
1122 // If a pointer refers to a different underlying object, bail - the
1123 // pointers are by definition incomparable.
1124 Value
*CurrObj
= GetUnderlyingObject(Ptr
, DL
);
1125 if (CurrObj
!= Obj0
)
1128 const SCEV
*Scev
= SE
.getSCEV(Ptr
);
1129 const auto *Diff
= dyn_cast
<SCEVConstant
>(SE
.getMinusSCEV(Scev
, Scev0
));
1130 // The pointers may not have a constant offset from each other, or SCEV
1131 // may just not be smart enough to figure out they do. Regardless,
1132 // there's nothing we can do.
1136 // Check if the pointer with the same offset is found.
1137 int64_t Offset
= Diff
->getAPInt().getSExtValue();
1138 if (!Offsets
.insert(Offset
).second
)
1140 OffValPairs
.emplace_back(Offset
, Ptr
);
1142 SortedIndices
.clear();
1143 SortedIndices
.resize(VL
.size());
1144 std::iota(SortedIndices
.begin(), SortedIndices
.end(), 0);
1146 // Sort the memory accesses and keep the order of their uses in UseOrder.
1147 llvm::stable_sort(SortedIndices
, [&](unsigned Left
, unsigned Right
) {
1148 return OffValPairs
[Left
].first
< OffValPairs
[Right
].first
;
1151 // Check if the order is consecutive already.
1152 if (llvm::all_of(SortedIndices
, [&SortedIndices
](const unsigned I
) {
1153 return I
== SortedIndices
[I
];
1155 SortedIndices
.clear();
1160 /// Take the address space operand from the Load/Store instruction.
1161 /// Returns -1 if this is not a valid Load/Store instruction.
1162 static unsigned getAddressSpaceOperand(Value
*I
) {
1163 if (LoadInst
*L
= dyn_cast
<LoadInst
>(I
))
1164 return L
->getPointerAddressSpace();
1165 if (StoreInst
*S
= dyn_cast
<StoreInst
>(I
))
1166 return S
->getPointerAddressSpace();
1170 /// Returns true if the memory operations \p A and \p B are consecutive.
1171 bool llvm::isConsecutiveAccess(Value
*A
, Value
*B
, const DataLayout
&DL
,
1172 ScalarEvolution
&SE
, bool CheckType
) {
1173 Value
*PtrA
= getLoadStorePointerOperand(A
);
1174 Value
*PtrB
= getLoadStorePointerOperand(B
);
1175 unsigned ASA
= getAddressSpaceOperand(A
);
1176 unsigned ASB
= getAddressSpaceOperand(B
);
1178 // Check that the address spaces match and that the pointers are valid.
1179 if (!PtrA
|| !PtrB
|| (ASA
!= ASB
))
1182 // Make sure that A and B are different pointers.
1186 // Make sure that A and B have the same type if required.
1187 if (CheckType
&& PtrA
->getType() != PtrB
->getType())
1190 unsigned IdxWidth
= DL
.getIndexSizeInBits(ASA
);
1191 Type
*Ty
= cast
<PointerType
>(PtrA
->getType())->getElementType();
1193 APInt
OffsetA(IdxWidth
, 0), OffsetB(IdxWidth
, 0);
1194 PtrA
= PtrA
->stripAndAccumulateInBoundsConstantOffsets(DL
, OffsetA
);
1195 PtrB
= PtrB
->stripAndAccumulateInBoundsConstantOffsets(DL
, OffsetB
);
1197 // Retrieve the address space again as pointer stripping now tracks through
1199 ASA
= cast
<PointerType
>(PtrA
->getType())->getAddressSpace();
1200 ASB
= cast
<PointerType
>(PtrB
->getType())->getAddressSpace();
1201 // Check that the address spaces match and that the pointers are valid.
1205 IdxWidth
= DL
.getIndexSizeInBits(ASA
);
1206 OffsetA
= OffsetA
.sextOrTrunc(IdxWidth
);
1207 OffsetB
= OffsetB
.sextOrTrunc(IdxWidth
);
1209 APInt
Size(IdxWidth
, DL
.getTypeStoreSize(Ty
));
1211 // OffsetDelta = OffsetB - OffsetA;
1212 const SCEV
*OffsetSCEVA
= SE
.getConstant(OffsetA
);
1213 const SCEV
*OffsetSCEVB
= SE
.getConstant(OffsetB
);
1214 const SCEV
*OffsetDeltaSCEV
= SE
.getMinusSCEV(OffsetSCEVB
, OffsetSCEVA
);
1215 const SCEVConstant
*OffsetDeltaC
= dyn_cast
<SCEVConstant
>(OffsetDeltaSCEV
);
1216 const APInt
&OffsetDelta
= OffsetDeltaC
->getAPInt();
1217 // Check if they are based on the same pointer. That makes the offsets
1220 return OffsetDelta
== Size
;
1222 // Compute the necessary base pointer delta to have the necessary final delta
1223 // equal to the size.
1224 // BaseDelta = Size - OffsetDelta;
1225 const SCEV
*SizeSCEV
= SE
.getConstant(Size
);
1226 const SCEV
*BaseDelta
= SE
.getMinusSCEV(SizeSCEV
, OffsetDeltaSCEV
);
1228 // Otherwise compute the distance with SCEV between the base pointers.
1229 const SCEV
*PtrSCEVA
= SE
.getSCEV(PtrA
);
1230 const SCEV
*PtrSCEVB
= SE
.getSCEV(PtrB
);
1231 const SCEV
*X
= SE
.getAddExpr(PtrSCEVA
, BaseDelta
);
1232 return X
== PtrSCEVB
;
1235 MemoryDepChecker::VectorizationSafetyStatus
1236 MemoryDepChecker::Dependence::isSafeForVectorization(DepType Type
) {
1240 case BackwardVectorizable
:
1241 return VectorizationSafetyStatus::Safe
;
1244 return VectorizationSafetyStatus::PossiblySafeWithRtChecks
;
1245 case ForwardButPreventsForwarding
:
1247 case BackwardVectorizableButPreventsForwarding
:
1248 return VectorizationSafetyStatus::Unsafe
;
1250 llvm_unreachable("unexpected DepType!");
1253 bool MemoryDepChecker::Dependence::isBackward() const {
1257 case ForwardButPreventsForwarding
:
1261 case BackwardVectorizable
:
1263 case BackwardVectorizableButPreventsForwarding
:
1266 llvm_unreachable("unexpected DepType!");
1269 bool MemoryDepChecker::Dependence::isPossiblyBackward() const {
1270 return isBackward() || Type
== Unknown
;
1273 bool MemoryDepChecker::Dependence::isForward() const {
1276 case ForwardButPreventsForwarding
:
1281 case BackwardVectorizable
:
1283 case BackwardVectorizableButPreventsForwarding
:
1286 llvm_unreachable("unexpected DepType!");
1289 bool MemoryDepChecker::couldPreventStoreLoadForward(uint64_t Distance
,
1290 uint64_t TypeByteSize
) {
1291 // If loads occur at a distance that is not a multiple of a feasible vector
1292 // factor store-load forwarding does not take place.
1293 // Positive dependences might cause troubles because vectorizing them might
1294 // prevent store-load forwarding making vectorized code run a lot slower.
1295 // a[i] = a[i-3] ^ a[i-8];
1296 // The stores to a[i:i+1] don't align with the stores to a[i-3:i-2] and
1297 // hence on your typical architecture store-load forwarding does not take
1298 // place. Vectorizing in such cases does not make sense.
1299 // Store-load forwarding distance.
1301 // After this many iterations store-to-load forwarding conflicts should not
1302 // cause any slowdowns.
1303 const uint64_t NumItersForStoreLoadThroughMemory
= 8 * TypeByteSize
;
1304 // Maximum vector factor.
1305 uint64_t MaxVFWithoutSLForwardIssues
= std::min(
1306 VectorizerParams::MaxVectorWidth
* TypeByteSize
, MaxSafeDepDistBytes
);
1308 // Compute the smallest VF at which the store and load would be misaligned.
1309 for (uint64_t VF
= 2 * TypeByteSize
; VF
<= MaxVFWithoutSLForwardIssues
;
1311 // If the number of vector iteration between the store and the load are
1312 // small we could incur conflicts.
1313 if (Distance
% VF
&& Distance
/ VF
< NumItersForStoreLoadThroughMemory
) {
1314 MaxVFWithoutSLForwardIssues
= (VF
>>= 1);
1319 if (MaxVFWithoutSLForwardIssues
< 2 * TypeByteSize
) {
1321 dbgs() << "LAA: Distance " << Distance
1322 << " that could cause a store-load forwarding conflict\n");
1326 if (MaxVFWithoutSLForwardIssues
< MaxSafeDepDistBytes
&&
1327 MaxVFWithoutSLForwardIssues
!=
1328 VectorizerParams::MaxVectorWidth
* TypeByteSize
)
1329 MaxSafeDepDistBytes
= MaxVFWithoutSLForwardIssues
;
1333 void MemoryDepChecker::mergeInStatus(VectorizationSafetyStatus S
) {
1338 /// Given a non-constant (unknown) dependence-distance \p Dist between two
1339 /// memory accesses, that have the same stride whose absolute value is given
1340 /// in \p Stride, and that have the same type size \p TypeByteSize,
1341 /// in a loop whose takenCount is \p BackedgeTakenCount, check if it is
1342 /// possible to prove statically that the dependence distance is larger
1343 /// than the range that the accesses will travel through the execution of
1344 /// the loop. If so, return true; false otherwise. This is useful for
1345 /// example in loops such as the following (PR31098):
1346 /// for (i = 0; i < D; ++i) {
1350 static bool isSafeDependenceDistance(const DataLayout
&DL
, ScalarEvolution
&SE
,
1351 const SCEV
&BackedgeTakenCount
,
1352 const SCEV
&Dist
, uint64_t Stride
,
1353 uint64_t TypeByteSize
) {
1355 // If we can prove that
1356 // (**) |Dist| > BackedgeTakenCount * Step
1357 // where Step is the absolute stride of the memory accesses in bytes,
1358 // then there is no dependence.
1361 // We basically want to check if the absolute distance (|Dist/Step|)
1362 // is >= the loop iteration count (or > BackedgeTakenCount).
1363 // This is equivalent to the Strong SIV Test (Practical Dependence Testing,
1364 // Section 4.2.1); Note, that for vectorization it is sufficient to prove
1365 // that the dependence distance is >= VF; This is checked elsewhere.
1366 // But in some cases we can prune unknown dependence distances early, and
1367 // even before selecting the VF, and without a runtime test, by comparing
1368 // the distance against the loop iteration count. Since the vectorized code
1369 // will be executed only if LoopCount >= VF, proving distance >= LoopCount
1370 // also guarantees that distance >= VF.
1372 const uint64_t ByteStride
= Stride
* TypeByteSize
;
1373 const SCEV
*Step
= SE
.getConstant(BackedgeTakenCount
.getType(), ByteStride
);
1374 const SCEV
*Product
= SE
.getMulExpr(&BackedgeTakenCount
, Step
);
1376 const SCEV
*CastedDist
= &Dist
;
1377 const SCEV
*CastedProduct
= Product
;
1378 uint64_t DistTypeSize
= DL
.getTypeAllocSize(Dist
.getType());
1379 uint64_t ProductTypeSize
= DL
.getTypeAllocSize(Product
->getType());
1381 // The dependence distance can be positive/negative, so we sign extend Dist;
1382 // The multiplication of the absolute stride in bytes and the
1383 // backedgeTakenCount is non-negative, so we zero extend Product.
1384 if (DistTypeSize
> ProductTypeSize
)
1385 CastedProduct
= SE
.getZeroExtendExpr(Product
, Dist
.getType());
1387 CastedDist
= SE
.getNoopOrSignExtend(&Dist
, Product
->getType());
1389 // Is Dist - (BackedgeTakenCount * Step) > 0 ?
1390 // (If so, then we have proven (**) because |Dist| >= Dist)
1391 const SCEV
*Minus
= SE
.getMinusSCEV(CastedDist
, CastedProduct
);
1392 if (SE
.isKnownPositive(Minus
))
1395 // Second try: Is -Dist - (BackedgeTakenCount * Step) > 0 ?
1396 // (If so, then we have proven (**) because |Dist| >= -1*Dist)
1397 const SCEV
*NegDist
= SE
.getNegativeSCEV(CastedDist
);
1398 Minus
= SE
.getMinusSCEV(NegDist
, CastedProduct
);
1399 if (SE
.isKnownPositive(Minus
))
1405 /// Check the dependence for two accesses with the same stride \p Stride.
1406 /// \p Distance is the positive distance and \p TypeByteSize is type size in
1409 /// \returns true if they are independent.
1410 static bool areStridedAccessesIndependent(uint64_t Distance
, uint64_t Stride
,
1411 uint64_t TypeByteSize
) {
1412 assert(Stride
> 1 && "The stride must be greater than 1");
1413 assert(TypeByteSize
> 0 && "The type size in byte must be non-zero");
1414 assert(Distance
> 0 && "The distance must be non-zero");
1416 // Skip if the distance is not multiple of type byte size.
1417 if (Distance
% TypeByteSize
)
1420 uint64_t ScaledDist
= Distance
/ TypeByteSize
;
1422 // No dependence if the scaled distance is not multiple of the stride.
1424 // for (i = 0; i < 1024 ; i += 4)
1425 // A[i+2] = A[i] + 1;
1427 // Two accesses in memory (scaled distance is 2, stride is 4):
1428 // | A[0] | | | | A[4] | | | |
1429 // | | | A[2] | | | | A[6] | |
1432 // for (i = 0; i < 1024 ; i += 3)
1433 // A[i+4] = A[i] + 1;
1435 // Two accesses in memory (scaled distance is 4, stride is 3):
1436 // | A[0] | | | A[3] | | | A[6] | | |
1437 // | | | | | A[4] | | | A[7] | |
1438 return ScaledDist
% Stride
;
1441 MemoryDepChecker::Dependence::DepType
1442 MemoryDepChecker::isDependent(const MemAccessInfo
&A
, unsigned AIdx
,
1443 const MemAccessInfo
&B
, unsigned BIdx
,
1444 const ValueToValueMap
&Strides
) {
1445 assert (AIdx
< BIdx
&& "Must pass arguments in program order");
1447 Value
*APtr
= A
.getPointer();
1448 Value
*BPtr
= B
.getPointer();
1449 bool AIsWrite
= A
.getInt();
1450 bool BIsWrite
= B
.getInt();
1452 // Two reads are independent.
1453 if (!AIsWrite
&& !BIsWrite
)
1454 return Dependence::NoDep
;
1456 // We cannot check pointers in different address spaces.
1457 if (APtr
->getType()->getPointerAddressSpace() !=
1458 BPtr
->getType()->getPointerAddressSpace())
1459 return Dependence::Unknown
;
1461 int64_t StrideAPtr
= getPtrStride(PSE
, APtr
, InnermostLoop
, Strides
, true);
1462 int64_t StrideBPtr
= getPtrStride(PSE
, BPtr
, InnermostLoop
, Strides
, true);
1464 const SCEV
*Src
= PSE
.getSCEV(APtr
);
1465 const SCEV
*Sink
= PSE
.getSCEV(BPtr
);
1467 // If the induction step is negative we have to invert source and sink of the
1469 if (StrideAPtr
< 0) {
1470 std::swap(APtr
, BPtr
);
1471 std::swap(Src
, Sink
);
1472 std::swap(AIsWrite
, BIsWrite
);
1473 std::swap(AIdx
, BIdx
);
1474 std::swap(StrideAPtr
, StrideBPtr
);
1477 const SCEV
*Dist
= PSE
.getSE()->getMinusSCEV(Sink
, Src
);
1479 LLVM_DEBUG(dbgs() << "LAA: Src Scev: " << *Src
<< "Sink Scev: " << *Sink
1480 << "(Induction step: " << StrideAPtr
<< ")\n");
1481 LLVM_DEBUG(dbgs() << "LAA: Distance for " << *InstMap
[AIdx
] << " to "
1482 << *InstMap
[BIdx
] << ": " << *Dist
<< "\n");
1484 // Need accesses with constant stride. We don't want to vectorize
1485 // "A[B[i]] += ..." and similar code or pointer arithmetic that could wrap in
1486 // the address space.
1487 if (!StrideAPtr
|| !StrideBPtr
|| StrideAPtr
!= StrideBPtr
){
1488 LLVM_DEBUG(dbgs() << "Pointer access with non-constant stride\n");
1489 return Dependence::Unknown
;
1492 Type
*ATy
= APtr
->getType()->getPointerElementType();
1493 Type
*BTy
= BPtr
->getType()->getPointerElementType();
1494 auto &DL
= InnermostLoop
->getHeader()->getModule()->getDataLayout();
1495 uint64_t TypeByteSize
= DL
.getTypeAllocSize(ATy
);
1496 uint64_t Stride
= std::abs(StrideAPtr
);
1497 const SCEVConstant
*C
= dyn_cast
<SCEVConstant
>(Dist
);
1499 if (TypeByteSize
== DL
.getTypeAllocSize(BTy
) &&
1500 isSafeDependenceDistance(DL
, *(PSE
.getSE()),
1501 *(PSE
.getBackedgeTakenCount()), *Dist
, Stride
,
1503 return Dependence::NoDep
;
1505 LLVM_DEBUG(dbgs() << "LAA: Dependence because of non-constant distance\n");
1506 FoundNonConstantDistanceDependence
= true;
1507 return Dependence::Unknown
;
1510 const APInt
&Val
= C
->getAPInt();
1511 int64_t Distance
= Val
.getSExtValue();
1513 // Attempt to prove strided accesses independent.
1514 if (std::abs(Distance
) > 0 && Stride
> 1 && ATy
== BTy
&&
1515 areStridedAccessesIndependent(std::abs(Distance
), Stride
, TypeByteSize
)) {
1516 LLVM_DEBUG(dbgs() << "LAA: Strided accesses are independent\n");
1517 return Dependence::NoDep
;
1520 // Negative distances are not plausible dependencies.
1521 if (Val
.isNegative()) {
1522 bool IsTrueDataDependence
= (AIsWrite
&& !BIsWrite
);
1523 if (IsTrueDataDependence
&& EnableForwardingConflictDetection
&&
1524 (couldPreventStoreLoadForward(Val
.abs().getZExtValue(), TypeByteSize
) ||
1526 LLVM_DEBUG(dbgs() << "LAA: Forward but may prevent st->ld forwarding\n");
1527 return Dependence::ForwardButPreventsForwarding
;
1530 LLVM_DEBUG(dbgs() << "LAA: Dependence is negative\n");
1531 return Dependence::Forward
;
1534 // Write to the same location with the same size.
1535 // Could be improved to assert type sizes are the same (i32 == float, etc).
1538 return Dependence::Forward
;
1540 dbgs() << "LAA: Zero dependence difference but different types\n");
1541 return Dependence::Unknown
;
1544 assert(Val
.isStrictlyPositive() && "Expect a positive value");
1549 << "LAA: ReadWrite-Write positive dependency with different types\n");
1550 return Dependence::Unknown
;
1553 // Bail out early if passed-in parameters make vectorization not feasible.
1554 unsigned ForcedFactor
= (VectorizerParams::VectorizationFactor
?
1555 VectorizerParams::VectorizationFactor
: 1);
1556 unsigned ForcedUnroll
= (VectorizerParams::VectorizationInterleave
?
1557 VectorizerParams::VectorizationInterleave
: 1);
1558 // The minimum number of iterations for a vectorized/unrolled version.
1559 unsigned MinNumIter
= std::max(ForcedFactor
* ForcedUnroll
, 2U);
1561 // It's not vectorizable if the distance is smaller than the minimum distance
1562 // needed for a vectroized/unrolled version. Vectorizing one iteration in
1563 // front needs TypeByteSize * Stride. Vectorizing the last iteration needs
1564 // TypeByteSize (No need to plus the last gap distance).
1566 // E.g. Assume one char is 1 byte in memory and one int is 4 bytes.
1568 // int *B = (int *)((char *)A + 14);
1569 // for (i = 0 ; i < 1024 ; i += 2)
1573 // Two accesses in memory (stride is 2):
1574 // | A[0] | | A[2] | | A[4] | | A[6] | |
1575 // | B[0] | | B[2] | | B[4] |
1577 // Distance needs for vectorizing iterations except the last iteration:
1578 // 4 * 2 * (MinNumIter - 1). Distance needs for the last iteration: 4.
1579 // So the minimum distance needed is: 4 * 2 * (MinNumIter - 1) + 4.
1581 // If MinNumIter is 2, it is vectorizable as the minimum distance needed is
1582 // 12, which is less than distance.
1584 // If MinNumIter is 4 (Say if a user forces the vectorization factor to be 4),
1585 // the minimum distance needed is 28, which is greater than distance. It is
1586 // not safe to do vectorization.
1587 uint64_t MinDistanceNeeded
=
1588 TypeByteSize
* Stride
* (MinNumIter
- 1) + TypeByteSize
;
1589 if (MinDistanceNeeded
> static_cast<uint64_t>(Distance
)) {
1590 LLVM_DEBUG(dbgs() << "LAA: Failure because of positive distance "
1591 << Distance
<< '\n');
1592 return Dependence::Backward
;
1595 // Unsafe if the minimum distance needed is greater than max safe distance.
1596 if (MinDistanceNeeded
> MaxSafeDepDistBytes
) {
1597 LLVM_DEBUG(dbgs() << "LAA: Failure because it needs at least "
1598 << MinDistanceNeeded
<< " size in bytes");
1599 return Dependence::Backward
;
1602 // Positive distance bigger than max vectorization factor.
1603 // FIXME: Should use max factor instead of max distance in bytes, which could
1604 // not handle different types.
1605 // E.g. Assume one char is 1 byte in memory and one int is 4 bytes.
1606 // void foo (int *A, char *B) {
1607 // for (unsigned i = 0; i < 1024; i++) {
1608 // A[i+2] = A[i] + 1;
1609 // B[i+2] = B[i] + 1;
1613 // This case is currently unsafe according to the max safe distance. If we
1614 // analyze the two accesses on array B, the max safe dependence distance
1615 // is 2. Then we analyze the accesses on array A, the minimum distance needed
1616 // is 8, which is less than 2 and forbidden vectorization, But actually
1617 // both A and B could be vectorized by 2 iterations.
1618 MaxSafeDepDistBytes
=
1619 std::min(static_cast<uint64_t>(Distance
), MaxSafeDepDistBytes
);
1621 bool IsTrueDataDependence
= (!AIsWrite
&& BIsWrite
);
1622 if (IsTrueDataDependence
&& EnableForwardingConflictDetection
&&
1623 couldPreventStoreLoadForward(Distance
, TypeByteSize
))
1624 return Dependence::BackwardVectorizableButPreventsForwarding
;
1626 uint64_t MaxVF
= MaxSafeDepDistBytes
/ (TypeByteSize
* Stride
);
1627 LLVM_DEBUG(dbgs() << "LAA: Positive distance " << Val
.getSExtValue()
1628 << " with max VF = " << MaxVF
<< '\n');
1629 uint64_t MaxVFInBits
= MaxVF
* TypeByteSize
* 8;
1630 MaxSafeRegisterWidth
= std::min(MaxSafeRegisterWidth
, MaxVFInBits
);
1631 return Dependence::BackwardVectorizable
;
1634 bool MemoryDepChecker::areDepsSafe(DepCandidates
&AccessSets
,
1635 MemAccessInfoList
&CheckDeps
,
1636 const ValueToValueMap
&Strides
) {
1638 MaxSafeDepDistBytes
= -1;
1639 SmallPtrSet
<MemAccessInfo
, 8> Visited
;
1640 for (MemAccessInfo CurAccess
: CheckDeps
) {
1641 if (Visited
.count(CurAccess
))
1644 // Get the relevant memory access set.
1645 EquivalenceClasses
<MemAccessInfo
>::iterator I
=
1646 AccessSets
.findValue(AccessSets
.getLeaderValue(CurAccess
));
1648 // Check accesses within this set.
1649 EquivalenceClasses
<MemAccessInfo
>::member_iterator AI
=
1650 AccessSets
.member_begin(I
);
1651 EquivalenceClasses
<MemAccessInfo
>::member_iterator AE
=
1652 AccessSets
.member_end();
1654 // Check every access pair.
1656 Visited
.insert(*AI
);
1657 EquivalenceClasses
<MemAccessInfo
>::member_iterator OI
= std::next(AI
);
1659 // Check every accessing instruction pair in program order.
1660 for (std::vector
<unsigned>::iterator I1
= Accesses
[*AI
].begin(),
1661 I1E
= Accesses
[*AI
].end(); I1
!= I1E
; ++I1
)
1662 for (std::vector
<unsigned>::iterator I2
= Accesses
[*OI
].begin(),
1663 I2E
= Accesses
[*OI
].end(); I2
!= I2E
; ++I2
) {
1664 auto A
= std::make_pair(&*AI
, *I1
);
1665 auto B
= std::make_pair(&*OI
, *I2
);
1671 Dependence::DepType Type
=
1672 isDependent(*A
.first
, A
.second
, *B
.first
, B
.second
, Strides
);
1673 mergeInStatus(Dependence::isSafeForVectorization(Type
));
1675 // Gather dependences unless we accumulated MaxDependences
1676 // dependences. In that case return as soon as we find the first
1677 // unsafe dependence. This puts a limit on this quadratic
1679 if (RecordDependences
) {
1680 if (Type
!= Dependence::NoDep
)
1681 Dependences
.push_back(Dependence(A
.second
, B
.second
, Type
));
1683 if (Dependences
.size() >= MaxDependences
) {
1684 RecordDependences
= false;
1685 Dependences
.clear();
1687 << "Too many dependences, stopped recording\n");
1690 if (!RecordDependences
&& !isSafeForVectorization())
1699 LLVM_DEBUG(dbgs() << "Total Dependences: " << Dependences
.size() << "\n");
1700 return isSafeForVectorization();
1703 SmallVector
<Instruction
*, 4>
1704 MemoryDepChecker::getInstructionsForAccess(Value
*Ptr
, bool isWrite
) const {
1705 MemAccessInfo
Access(Ptr
, isWrite
);
1706 auto &IndexVector
= Accesses
.find(Access
)->second
;
1708 SmallVector
<Instruction
*, 4> Insts
;
1709 transform(IndexVector
,
1710 std::back_inserter(Insts
),
1711 [&](unsigned Idx
) { return this->InstMap
[Idx
]; });
1715 const char *MemoryDepChecker::Dependence::DepName
[] = {
1716 "NoDep", "Unknown", "Forward", "ForwardButPreventsForwarding", "Backward",
1717 "BackwardVectorizable", "BackwardVectorizableButPreventsForwarding"};
1719 void MemoryDepChecker::Dependence::print(
1720 raw_ostream
&OS
, unsigned Depth
,
1721 const SmallVectorImpl
<Instruction
*> &Instrs
) const {
1722 OS
.indent(Depth
) << DepName
[Type
] << ":\n";
1723 OS
.indent(Depth
+ 2) << *Instrs
[Source
] << " -> \n";
1724 OS
.indent(Depth
+ 2) << *Instrs
[Destination
] << "\n";
1727 bool LoopAccessInfo::canAnalyzeLoop() {
1728 // We need to have a loop header.
1729 LLVM_DEBUG(dbgs() << "LAA: Found a loop in "
1730 << TheLoop
->getHeader()->getParent()->getName() << ": "
1731 << TheLoop
->getHeader()->getName() << '\n');
1733 // We can only analyze innermost loops.
1734 if (!TheLoop
->empty()) {
1735 LLVM_DEBUG(dbgs() << "LAA: loop is not the innermost loop\n");
1736 recordAnalysis("NotInnerMostLoop") << "loop is not the innermost loop";
1740 // We must have a single backedge.
1741 if (TheLoop
->getNumBackEdges() != 1) {
1743 dbgs() << "LAA: loop control flow is not understood by analyzer\n");
1744 recordAnalysis("CFGNotUnderstood")
1745 << "loop control flow is not understood by analyzer";
1749 // We must have a single exiting block.
1750 if (!TheLoop
->getExitingBlock()) {
1752 dbgs() << "LAA: loop control flow is not understood by analyzer\n");
1753 recordAnalysis("CFGNotUnderstood")
1754 << "loop control flow is not understood by analyzer";
1758 // We only handle bottom-tested loops, i.e. loop in which the condition is
1759 // checked at the end of each iteration. With that we can assume that all
1760 // instructions in the loop are executed the same number of times.
1761 if (TheLoop
->getExitingBlock() != TheLoop
->getLoopLatch()) {
1763 dbgs() << "LAA: loop control flow is not understood by analyzer\n");
1764 recordAnalysis("CFGNotUnderstood")
1765 << "loop control flow is not understood by analyzer";
1769 // ScalarEvolution needs to be able to find the exit count.
1770 const SCEV
*ExitCount
= PSE
->getBackedgeTakenCount();
1771 if (ExitCount
== PSE
->getSE()->getCouldNotCompute()) {
1772 recordAnalysis("CantComputeNumberOfIterations")
1773 << "could not determine number of loop iterations";
1774 LLVM_DEBUG(dbgs() << "LAA: SCEV could not compute the loop exit count.\n");
1781 void LoopAccessInfo::analyzeLoop(AliasAnalysis
*AA
, LoopInfo
*LI
,
1782 const TargetLibraryInfo
*TLI
,
1783 DominatorTree
*DT
) {
1784 typedef SmallPtrSet
<Value
*, 16> ValueSet
;
1786 // Holds the Load and Store instructions.
1787 SmallVector
<LoadInst
*, 16> Loads
;
1788 SmallVector
<StoreInst
*, 16> Stores
;
1790 // Holds all the different accesses in the loop.
1791 unsigned NumReads
= 0;
1792 unsigned NumReadWrites
= 0;
1794 bool HasComplexMemInst
= false;
1796 // A runtime check is only legal to insert if there are no convergent calls.
1797 HasConvergentOp
= false;
1799 PtrRtChecking
->Pointers
.clear();
1800 PtrRtChecking
->Need
= false;
1802 const bool IsAnnotatedParallel
= TheLoop
->isAnnotatedParallel();
1805 for (BasicBlock
*BB
: TheLoop
->blocks()) {
1806 // Scan the BB and collect legal loads and stores. Also detect any
1807 // convergent instructions.
1808 for (Instruction
&I
: *BB
) {
1809 if (auto *Call
= dyn_cast
<CallBase
>(&I
)) {
1810 if (Call
->isConvergent())
1811 HasConvergentOp
= true;
1814 // With both a non-vectorizable memory instruction and a convergent
1815 // operation, found in this loop, no reason to continue the search.
1816 if (HasComplexMemInst
&& HasConvergentOp
) {
1821 // Avoid hitting recordAnalysis multiple times.
1822 if (HasComplexMemInst
)
1825 // If this is a load, save it. If this instruction can read from memory
1826 // but is not a load, then we quit. Notice that we don't handle function
1827 // calls that read or write.
1828 if (I
.mayReadFromMemory()) {
1829 // Many math library functions read the rounding mode. We will only
1830 // vectorize a loop if it contains known function calls that don't set
1831 // the flag. Therefore, it is safe to ignore this read from memory.
1832 auto *Call
= dyn_cast
<CallInst
>(&I
);
1833 if (Call
&& getVectorIntrinsicIDForCall(Call
, TLI
))
1836 // If the function has an explicit vectorized counterpart, we can safely
1837 // assume that it can be vectorized.
1838 if (Call
&& !Call
->isNoBuiltin() && Call
->getCalledFunction() &&
1839 TLI
->isFunctionVectorizable(Call
->getCalledFunction()->getName()))
1842 auto *Ld
= dyn_cast
<LoadInst
>(&I
);
1844 recordAnalysis("CantVectorizeInstruction", Ld
)
1845 << "instruction cannot be vectorized";
1846 HasComplexMemInst
= true;
1849 if (!Ld
->isSimple() && !IsAnnotatedParallel
) {
1850 recordAnalysis("NonSimpleLoad", Ld
)
1851 << "read with atomic ordering or volatile read";
1852 LLVM_DEBUG(dbgs() << "LAA: Found a non-simple load.\n");
1853 HasComplexMemInst
= true;
1857 Loads
.push_back(Ld
);
1858 DepChecker
->addAccess(Ld
);
1859 if (EnableMemAccessVersioning
)
1860 collectStridedAccess(Ld
);
1864 // Save 'store' instructions. Abort if other instructions write to memory.
1865 if (I
.mayWriteToMemory()) {
1866 auto *St
= dyn_cast
<StoreInst
>(&I
);
1868 recordAnalysis("CantVectorizeInstruction", St
)
1869 << "instruction cannot be vectorized";
1870 HasComplexMemInst
= true;
1873 if (!St
->isSimple() && !IsAnnotatedParallel
) {
1874 recordAnalysis("NonSimpleStore", St
)
1875 << "write with atomic ordering or volatile write";
1876 LLVM_DEBUG(dbgs() << "LAA: Found a non-simple store.\n");
1877 HasComplexMemInst
= true;
1881 Stores
.push_back(St
);
1882 DepChecker
->addAccess(St
);
1883 if (EnableMemAccessVersioning
)
1884 collectStridedAccess(St
);
1889 if (HasComplexMemInst
) {
1894 // Now we have two lists that hold the loads and the stores.
1895 // Next, we find the pointers that they use.
1897 // Check if we see any stores. If there are no stores, then we don't
1898 // care if the pointers are *restrict*.
1899 if (!Stores
.size()) {
1900 LLVM_DEBUG(dbgs() << "LAA: Found a read-only loop!\n");
1905 MemoryDepChecker::DepCandidates DependentAccesses
;
1906 AccessAnalysis
Accesses(TheLoop
->getHeader()->getModule()->getDataLayout(),
1907 TheLoop
, AA
, LI
, DependentAccesses
, *PSE
);
1909 // Holds the analyzed pointers. We don't want to call GetUnderlyingObjects
1910 // multiple times on the same object. If the ptr is accessed twice, once
1911 // for read and once for write, it will only appear once (on the write
1912 // list). This is okay, since we are going to check for conflicts between
1913 // writes and between reads and writes, but not between reads and reads.
1916 // Record uniform store addresses to identify if we have multiple stores
1917 // to the same address.
1918 ValueSet UniformStores
;
1920 for (StoreInst
*ST
: Stores
) {
1921 Value
*Ptr
= ST
->getPointerOperand();
1924 HasDependenceInvolvingLoopInvariantAddress
|=
1925 !UniformStores
.insert(Ptr
).second
;
1927 // If we did *not* see this pointer before, insert it to the read-write
1928 // list. At this phase it is only a 'write' list.
1929 if (Seen
.insert(Ptr
).second
) {
1932 MemoryLocation Loc
= MemoryLocation::get(ST
);
1933 // The TBAA metadata could have a control dependency on the predication
1934 // condition, so we cannot rely on it when determining whether or not we
1935 // need runtime pointer checks.
1936 if (blockNeedsPredication(ST
->getParent(), TheLoop
, DT
))
1937 Loc
.AATags
.TBAA
= nullptr;
1939 Accesses
.addStore(Loc
);
1943 if (IsAnnotatedParallel
) {
1945 dbgs() << "LAA: A loop annotated parallel, ignore memory dependency "
1951 for (LoadInst
*LD
: Loads
) {
1952 Value
*Ptr
= LD
->getPointerOperand();
1953 // If we did *not* see this pointer before, insert it to the
1954 // read list. If we *did* see it before, then it is already in
1955 // the read-write list. This allows us to vectorize expressions
1956 // such as A[i] += x; Because the address of A[i] is a read-write
1957 // pointer. This only works if the index of A[i] is consecutive.
1958 // If the address of i is unknown (for example A[B[i]]) then we may
1959 // read a few words, modify, and write a few words, and some of the
1960 // words may be written to the same address.
1961 bool IsReadOnlyPtr
= false;
1962 if (Seen
.insert(Ptr
).second
||
1963 !getPtrStride(*PSE
, Ptr
, TheLoop
, SymbolicStrides
)) {
1965 IsReadOnlyPtr
= true;
1968 // See if there is an unsafe dependency between a load to a uniform address and
1969 // store to the same uniform address.
1970 if (UniformStores
.count(Ptr
)) {
1971 LLVM_DEBUG(dbgs() << "LAA: Found an unsafe dependency between a uniform "
1972 "load and uniform store to the same address!\n");
1973 HasDependenceInvolvingLoopInvariantAddress
= true;
1976 MemoryLocation Loc
= MemoryLocation::get(LD
);
1977 // The TBAA metadata could have a control dependency on the predication
1978 // condition, so we cannot rely on it when determining whether or not we
1979 // need runtime pointer checks.
1980 if (blockNeedsPredication(LD
->getParent(), TheLoop
, DT
))
1981 Loc
.AATags
.TBAA
= nullptr;
1983 Accesses
.addLoad(Loc
, IsReadOnlyPtr
);
1986 // If we write (or read-write) to a single destination and there are no
1987 // other reads in this loop then is it safe to vectorize.
1988 if (NumReadWrites
== 1 && NumReads
== 0) {
1989 LLVM_DEBUG(dbgs() << "LAA: Found a write-only loop!\n");
1994 // Build dependence sets and check whether we need a runtime pointer bounds
1996 Accesses
.buildDependenceSets();
1998 // Find pointers with computable bounds. We are going to use this information
1999 // to place a runtime bound check.
2000 bool CanDoRTIfNeeded
= Accesses
.canCheckPtrAtRT(*PtrRtChecking
, PSE
->getSE(),
2001 TheLoop
, SymbolicStrides
);
2002 if (!CanDoRTIfNeeded
) {
2003 recordAnalysis("CantIdentifyArrayBounds") << "cannot identify array bounds";
2004 LLVM_DEBUG(dbgs() << "LAA: We can't vectorize because we can't find "
2005 << "the array bounds.\n");
2011 dbgs() << "LAA: May be able to perform a memory runtime check if needed.\n");
2014 if (Accesses
.isDependencyCheckNeeded()) {
2015 LLVM_DEBUG(dbgs() << "LAA: Checking memory dependencies\n");
2016 CanVecMem
= DepChecker
->areDepsSafe(
2017 DependentAccesses
, Accesses
.getDependenciesToCheck(), SymbolicStrides
);
2018 MaxSafeDepDistBytes
= DepChecker
->getMaxSafeDepDistBytes();
2020 if (!CanVecMem
&& DepChecker
->shouldRetryWithRuntimeCheck()) {
2021 LLVM_DEBUG(dbgs() << "LAA: Retrying with memory checks\n");
2023 // Clear the dependency checks. We assume they are not needed.
2024 Accesses
.resetDepChecks(*DepChecker
);
2026 PtrRtChecking
->reset();
2027 PtrRtChecking
->Need
= true;
2029 auto *SE
= PSE
->getSE();
2030 CanDoRTIfNeeded
= Accesses
.canCheckPtrAtRT(*PtrRtChecking
, SE
, TheLoop
,
2031 SymbolicStrides
, true);
2033 // Check that we found the bounds for the pointer.
2034 if (!CanDoRTIfNeeded
) {
2035 recordAnalysis("CantCheckMemDepsAtRunTime")
2036 << "cannot check memory dependencies at runtime";
2037 LLVM_DEBUG(dbgs() << "LAA: Can't vectorize with memory checks\n");
2046 if (HasConvergentOp
) {
2047 recordAnalysis("CantInsertRuntimeCheckWithConvergent")
2048 << "cannot add control dependency to convergent operation";
2049 LLVM_DEBUG(dbgs() << "LAA: We can't vectorize because a runtime check "
2050 "would be needed with a convergent operation\n");
2057 dbgs() << "LAA: No unsafe dependent memory operations in loop. We"
2058 << (PtrRtChecking
->Need
? "" : " don't")
2059 << " need runtime memory checks.\n");
2061 recordAnalysis("UnsafeMemDep")
2062 << "unsafe dependent memory operations in loop. Use "
2063 "#pragma loop distribute(enable) to allow loop distribution "
2064 "to attempt to isolate the offending operations into a separate "
2066 LLVM_DEBUG(dbgs() << "LAA: unsafe dependent memory operations in loop\n");
2070 bool LoopAccessInfo::blockNeedsPredication(BasicBlock
*BB
, Loop
*TheLoop
,
2071 DominatorTree
*DT
) {
2072 assert(TheLoop
->contains(BB
) && "Unknown block used");
2074 // Blocks that do not dominate the latch need predication.
2075 BasicBlock
* Latch
= TheLoop
->getLoopLatch();
2076 return !DT
->dominates(BB
, Latch
);
2079 OptimizationRemarkAnalysis
&LoopAccessInfo::recordAnalysis(StringRef RemarkName
,
2081 assert(!Report
&& "Multiple reports generated");
2083 Value
*CodeRegion
= TheLoop
->getHeader();
2084 DebugLoc DL
= TheLoop
->getStartLoc();
2087 CodeRegion
= I
->getParent();
2088 // If there is no debug location attached to the instruction, revert back to
2089 // using the loop's.
2090 if (I
->getDebugLoc())
2091 DL
= I
->getDebugLoc();
2094 Report
= make_unique
<OptimizationRemarkAnalysis
>(DEBUG_TYPE
, RemarkName
, DL
,
2099 bool LoopAccessInfo::isUniform(Value
*V
) const {
2100 auto *SE
= PSE
->getSE();
2101 // Since we rely on SCEV for uniformity, if the type is not SCEVable, it is
2102 // never considered uniform.
2103 // TODO: Is this really what we want? Even without FP SCEV, we may want some
2104 // trivially loop-invariant FP values to be considered uniform.
2105 if (!SE
->isSCEVable(V
->getType()))
2107 return (SE
->isLoopInvariant(SE
->getSCEV(V
), TheLoop
));
2110 // FIXME: this function is currently a duplicate of the one in
2111 // LoopVectorize.cpp.
2112 static Instruction
*getFirstInst(Instruction
*FirstInst
, Value
*V
,
2116 if (Instruction
*I
= dyn_cast
<Instruction
>(V
))
2117 return I
->getParent() == Loc
->getParent() ? I
: nullptr;
2123 /// IR Values for the lower and upper bounds of a pointer evolution. We
2124 /// need to use value-handles because SCEV expansion can invalidate previously
2125 /// expanded values. Thus expansion of a pointer can invalidate the bounds for
2127 struct PointerBounds
{
2128 TrackingVH
<Value
> Start
;
2129 TrackingVH
<Value
> End
;
2132 } // end anonymous namespace
2134 /// Expand code for the lower and upper bound of the pointer group \p CG
2135 /// in \p TheLoop. \return the values for the bounds.
2136 static PointerBounds
2137 expandBounds(const RuntimePointerChecking::CheckingPtrGroup
*CG
, Loop
*TheLoop
,
2138 Instruction
*Loc
, SCEVExpander
&Exp
, ScalarEvolution
*SE
,
2139 const RuntimePointerChecking
&PtrRtChecking
) {
2140 Value
*Ptr
= PtrRtChecking
.Pointers
[CG
->Members
[0]].PointerValue
;
2141 const SCEV
*Sc
= SE
->getSCEV(Ptr
);
2143 unsigned AS
= Ptr
->getType()->getPointerAddressSpace();
2144 LLVMContext
&Ctx
= Loc
->getContext();
2146 // Use this type for pointer arithmetic.
2147 Type
*PtrArithTy
= Type::getInt8PtrTy(Ctx
, AS
);
2149 if (SE
->isLoopInvariant(Sc
, TheLoop
)) {
2150 LLVM_DEBUG(dbgs() << "LAA: Adding RT check for a loop invariant ptr:"
2152 // Ptr could be in the loop body. If so, expand a new one at the correct
2154 Instruction
*Inst
= dyn_cast
<Instruction
>(Ptr
);
2155 Value
*NewPtr
= (Inst
&& TheLoop
->contains(Inst
))
2156 ? Exp
.expandCodeFor(Sc
, PtrArithTy
, Loc
)
2158 // We must return a half-open range, which means incrementing Sc.
2159 const SCEV
*ScPlusOne
= SE
->getAddExpr(Sc
, SE
->getOne(PtrArithTy
));
2160 Value
*NewPtrPlusOne
= Exp
.expandCodeFor(ScPlusOne
, PtrArithTy
, Loc
);
2161 return {NewPtr
, NewPtrPlusOne
};
2163 Value
*Start
= nullptr, *End
= nullptr;
2164 LLVM_DEBUG(dbgs() << "LAA: Adding RT check for range:\n");
2165 Start
= Exp
.expandCodeFor(CG
->Low
, PtrArithTy
, Loc
);
2166 End
= Exp
.expandCodeFor(CG
->High
, PtrArithTy
, Loc
);
2167 LLVM_DEBUG(dbgs() << "Start: " << *CG
->Low
<< " End: " << *CG
->High
2169 return {Start
, End
};
2173 /// Turns a collection of checks into a collection of expanded upper and
2174 /// lower bounds for both pointers in the check.
2175 static SmallVector
<std::pair
<PointerBounds
, PointerBounds
>, 4> expandBounds(
2176 const SmallVectorImpl
<RuntimePointerChecking::PointerCheck
> &PointerChecks
,
2177 Loop
*L
, Instruction
*Loc
, ScalarEvolution
*SE
, SCEVExpander
&Exp
,
2178 const RuntimePointerChecking
&PtrRtChecking
) {
2179 SmallVector
<std::pair
<PointerBounds
, PointerBounds
>, 4> ChecksWithBounds
;
2181 // Here we're relying on the SCEV Expander's cache to only emit code for the
2182 // same bounds once.
2184 PointerChecks
, std::back_inserter(ChecksWithBounds
),
2185 [&](const RuntimePointerChecking::PointerCheck
&Check
) {
2187 First
= expandBounds(Check
.first
, L
, Loc
, Exp
, SE
, PtrRtChecking
),
2188 Second
= expandBounds(Check
.second
, L
, Loc
, Exp
, SE
, PtrRtChecking
);
2189 return std::make_pair(First
, Second
);
2192 return ChecksWithBounds
;
2195 std::pair
<Instruction
*, Instruction
*> LoopAccessInfo::addRuntimeChecks(
2197 const SmallVectorImpl
<RuntimePointerChecking::PointerCheck
> &PointerChecks
)
2199 const DataLayout
&DL
= TheLoop
->getHeader()->getModule()->getDataLayout();
2200 auto *SE
= PSE
->getSE();
2201 SCEVExpander
Exp(*SE
, DL
, "induction");
2202 auto ExpandedChecks
=
2203 expandBounds(PointerChecks
, TheLoop
, Loc
, SE
, Exp
, *PtrRtChecking
);
2205 LLVMContext
&Ctx
= Loc
->getContext();
2206 Instruction
*FirstInst
= nullptr;
2207 IRBuilder
<> ChkBuilder(Loc
);
2208 // Our instructions might fold to a constant.
2209 Value
*MemoryRuntimeCheck
= nullptr;
2211 for (const auto &Check
: ExpandedChecks
) {
2212 const PointerBounds
&A
= Check
.first
, &B
= Check
.second
;
2213 // Check if two pointers (A and B) conflict where conflict is computed as:
2214 // start(A) <= end(B) && start(B) <= end(A)
2215 unsigned AS0
= A
.Start
->getType()->getPointerAddressSpace();
2216 unsigned AS1
= B
.Start
->getType()->getPointerAddressSpace();
2218 assert((AS0
== B
.End
->getType()->getPointerAddressSpace()) &&
2219 (AS1
== A
.End
->getType()->getPointerAddressSpace()) &&
2220 "Trying to bounds check pointers with different address spaces");
2222 Type
*PtrArithTy0
= Type::getInt8PtrTy(Ctx
, AS0
);
2223 Type
*PtrArithTy1
= Type::getInt8PtrTy(Ctx
, AS1
);
2225 Value
*Start0
= ChkBuilder
.CreateBitCast(A
.Start
, PtrArithTy0
, "bc");
2226 Value
*Start1
= ChkBuilder
.CreateBitCast(B
.Start
, PtrArithTy1
, "bc");
2227 Value
*End0
= ChkBuilder
.CreateBitCast(A
.End
, PtrArithTy1
, "bc");
2228 Value
*End1
= ChkBuilder
.CreateBitCast(B
.End
, PtrArithTy0
, "bc");
2230 // [A|B].Start points to the first accessed byte under base [A|B].
2231 // [A|B].End points to the last accessed byte, plus one.
2232 // There is no conflict when the intervals are disjoint:
2233 // NoConflict = (B.Start >= A.End) || (A.Start >= B.End)
2235 // bound0 = (B.Start < A.End)
2236 // bound1 = (A.Start < B.End)
2237 // IsConflict = bound0 & bound1
2238 Value
*Cmp0
= ChkBuilder
.CreateICmpULT(Start0
, End1
, "bound0");
2239 FirstInst
= getFirstInst(FirstInst
, Cmp0
, Loc
);
2240 Value
*Cmp1
= ChkBuilder
.CreateICmpULT(Start1
, End0
, "bound1");
2241 FirstInst
= getFirstInst(FirstInst
, Cmp1
, Loc
);
2242 Value
*IsConflict
= ChkBuilder
.CreateAnd(Cmp0
, Cmp1
, "found.conflict");
2243 FirstInst
= getFirstInst(FirstInst
, IsConflict
, Loc
);
2244 if (MemoryRuntimeCheck
) {
2246 ChkBuilder
.CreateOr(MemoryRuntimeCheck
, IsConflict
, "conflict.rdx");
2247 FirstInst
= getFirstInst(FirstInst
, IsConflict
, Loc
);
2249 MemoryRuntimeCheck
= IsConflict
;
2252 if (!MemoryRuntimeCheck
)
2253 return std::make_pair(nullptr, nullptr);
2255 // We have to do this trickery because the IRBuilder might fold the check to a
2256 // constant expression in which case there is no Instruction anchored in a
2258 Instruction
*Check
= BinaryOperator::CreateAnd(MemoryRuntimeCheck
,
2259 ConstantInt::getTrue(Ctx
));
2260 ChkBuilder
.Insert(Check
, "memcheck.conflict");
2261 FirstInst
= getFirstInst(FirstInst
, Check
, Loc
);
2262 return std::make_pair(FirstInst
, Check
);
2265 std::pair
<Instruction
*, Instruction
*>
2266 LoopAccessInfo::addRuntimeChecks(Instruction
*Loc
) const {
2267 if (!PtrRtChecking
->Need
)
2268 return std::make_pair(nullptr, nullptr);
2270 return addRuntimeChecks(Loc
, PtrRtChecking
->getChecks());
2273 void LoopAccessInfo::collectStridedAccess(Value
*MemAccess
) {
2274 Value
*Ptr
= nullptr;
2275 if (LoadInst
*LI
= dyn_cast
<LoadInst
>(MemAccess
))
2276 Ptr
= LI
->getPointerOperand();
2277 else if (StoreInst
*SI
= dyn_cast
<StoreInst
>(MemAccess
))
2278 Ptr
= SI
->getPointerOperand();
2282 Value
*Stride
= getStrideFromPointer(Ptr
, PSE
->getSE(), TheLoop
);
2286 LLVM_DEBUG(dbgs() << "LAA: Found a strided access that is a candidate for "
2288 LLVM_DEBUG(dbgs() << " Ptr: " << *Ptr
<< " Stride: " << *Stride
<< "\n");
2290 // Avoid adding the "Stride == 1" predicate when we know that
2291 // Stride >= Trip-Count. Such a predicate will effectively optimize a single
2292 // or zero iteration loop, as Trip-Count <= Stride == 1.
2294 // TODO: We are currently not making a very informed decision on when it is
2295 // beneficial to apply stride versioning. It might make more sense that the
2296 // users of this analysis (such as the vectorizer) will trigger it, based on
2297 // their specific cost considerations; For example, in cases where stride
2298 // versioning does not help resolving memory accesses/dependences, the
2299 // vectorizer should evaluate the cost of the runtime test, and the benefit
2300 // of various possible stride specializations, considering the alternatives
2301 // of using gather/scatters (if available).
2303 const SCEV
*StrideExpr
= PSE
->getSCEV(Stride
);
2304 const SCEV
*BETakenCount
= PSE
->getBackedgeTakenCount();
2306 // Match the types so we can compare the stride and the BETakenCount.
2307 // The Stride can be positive/negative, so we sign extend Stride;
2308 // The backedgeTakenCount is non-negative, so we zero extend BETakenCount.
2309 const DataLayout
&DL
= TheLoop
->getHeader()->getModule()->getDataLayout();
2310 uint64_t StrideTypeSize
= DL
.getTypeAllocSize(StrideExpr
->getType());
2311 uint64_t BETypeSize
= DL
.getTypeAllocSize(BETakenCount
->getType());
2312 const SCEV
*CastedStride
= StrideExpr
;
2313 const SCEV
*CastedBECount
= BETakenCount
;
2314 ScalarEvolution
*SE
= PSE
->getSE();
2315 if (BETypeSize
>= StrideTypeSize
)
2316 CastedStride
= SE
->getNoopOrSignExtend(StrideExpr
, BETakenCount
->getType());
2318 CastedBECount
= SE
->getZeroExtendExpr(BETakenCount
, StrideExpr
->getType());
2319 const SCEV
*StrideMinusBETaken
= SE
->getMinusSCEV(CastedStride
, CastedBECount
);
2320 // Since TripCount == BackEdgeTakenCount + 1, checking:
2321 // "Stride >= TripCount" is equivalent to checking:
2322 // Stride - BETakenCount > 0
2323 if (SE
->isKnownPositive(StrideMinusBETaken
)) {
2325 dbgs() << "LAA: Stride>=TripCount; No point in versioning as the "
2326 "Stride==1 predicate will imply that the loop executes "
2330 LLVM_DEBUG(dbgs() << "LAA: Found a strided access that we can version.");
2332 SymbolicStrides
[Ptr
] = Stride
;
2333 StrideSet
.insert(Stride
);
2336 LoopAccessInfo::LoopAccessInfo(Loop
*L
, ScalarEvolution
*SE
,
2337 const TargetLibraryInfo
*TLI
, AliasAnalysis
*AA
,
2338 DominatorTree
*DT
, LoopInfo
*LI
)
2339 : PSE(llvm::make_unique
<PredicatedScalarEvolution
>(*SE
, *L
)),
2340 PtrRtChecking(llvm::make_unique
<RuntimePointerChecking
>(SE
)),
2341 DepChecker(llvm::make_unique
<MemoryDepChecker
>(*PSE
, L
)), TheLoop(L
),
2342 NumLoads(0), NumStores(0), MaxSafeDepDistBytes(-1), CanVecMem(false),
2343 HasConvergentOp(false),
2344 HasDependenceInvolvingLoopInvariantAddress(false) {
2345 if (canAnalyzeLoop())
2346 analyzeLoop(AA
, LI
, TLI
, DT
);
2349 void LoopAccessInfo::print(raw_ostream
&OS
, unsigned Depth
) const {
2351 OS
.indent(Depth
) << "Memory dependences are safe";
2352 if (MaxSafeDepDistBytes
!= -1ULL)
2353 OS
<< " with a maximum dependence distance of " << MaxSafeDepDistBytes
2355 if (PtrRtChecking
->Need
)
2356 OS
<< " with run-time checks";
2360 if (HasConvergentOp
)
2361 OS
.indent(Depth
) << "Has convergent operation in loop\n";
2364 OS
.indent(Depth
) << "Report: " << Report
->getMsg() << "\n";
2366 if (auto *Dependences
= DepChecker
->getDependences()) {
2367 OS
.indent(Depth
) << "Dependences:\n";
2368 for (auto &Dep
: *Dependences
) {
2369 Dep
.print(OS
, Depth
+ 2, DepChecker
->getMemoryInstructions());
2373 OS
.indent(Depth
) << "Too many dependences, not recorded\n";
2375 // List the pair of accesses need run-time checks to prove independence.
2376 PtrRtChecking
->print(OS
, Depth
);
2379 OS
.indent(Depth
) << "Non vectorizable stores to invariant address were "
2380 << (HasDependenceInvolvingLoopInvariantAddress
? "" : "not ")
2381 << "found in loop.\n";
2383 OS
.indent(Depth
) << "SCEV assumptions:\n";
2384 PSE
->getUnionPredicate().print(OS
, Depth
);
2388 OS
.indent(Depth
) << "Expressions re-written:\n";
2389 PSE
->print(OS
, Depth
);
2392 const LoopAccessInfo
&LoopAccessLegacyAnalysis::getInfo(Loop
*L
) {
2393 auto &LAI
= LoopAccessInfoMap
[L
];
2396 LAI
= llvm::make_unique
<LoopAccessInfo
>(L
, SE
, TLI
, AA
, DT
, LI
);
2401 void LoopAccessLegacyAnalysis::print(raw_ostream
&OS
, const Module
*M
) const {
2402 LoopAccessLegacyAnalysis
&LAA
= *const_cast<LoopAccessLegacyAnalysis
*>(this);
2404 for (Loop
*TopLevelLoop
: *LI
)
2405 for (Loop
*L
: depth_first(TopLevelLoop
)) {
2406 OS
.indent(2) << L
->getHeader()->getName() << ":\n";
2407 auto &LAI
= LAA
.getInfo(L
);
2412 bool LoopAccessLegacyAnalysis::runOnFunction(Function
&F
) {
2413 SE
= &getAnalysis
<ScalarEvolutionWrapperPass
>().getSE();
2414 auto *TLIP
= getAnalysisIfAvailable
<TargetLibraryInfoWrapperPass
>();
2415 TLI
= TLIP
? &TLIP
->getTLI() : nullptr;
2416 AA
= &getAnalysis
<AAResultsWrapperPass
>().getAAResults();
2417 DT
= &getAnalysis
<DominatorTreeWrapperPass
>().getDomTree();
2418 LI
= &getAnalysis
<LoopInfoWrapperPass
>().getLoopInfo();
2423 void LoopAccessLegacyAnalysis::getAnalysisUsage(AnalysisUsage
&AU
) const {
2424 AU
.addRequired
<ScalarEvolutionWrapperPass
>();
2425 AU
.addRequired
<AAResultsWrapperPass
>();
2426 AU
.addRequired
<DominatorTreeWrapperPass
>();
2427 AU
.addRequired
<LoopInfoWrapperPass
>();
2429 AU
.setPreservesAll();
2432 char LoopAccessLegacyAnalysis::ID
= 0;
2433 static const char laa_name
[] = "Loop Access Analysis";
2434 #define LAA_NAME "loop-accesses"
2436 INITIALIZE_PASS_BEGIN(LoopAccessLegacyAnalysis
, LAA_NAME
, laa_name
, false, true)
2437 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass
)
2438 INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass
)
2439 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass
)
2440 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass
)
2441 INITIALIZE_PASS_END(LoopAccessLegacyAnalysis
, LAA_NAME
, laa_name
, false, true)
2443 AnalysisKey
LoopAccessAnalysis::Key
;
2445 LoopAccessInfo
LoopAccessAnalysis::run(Loop
&L
, LoopAnalysisManager
&AM
,
2446 LoopStandardAnalysisResults
&AR
) {
2447 return LoopAccessInfo(&L
, &AR
.SE
, &AR
.TLI
, &AR
.AA
, &AR
.DT
, &AR
.LI
);
2452 Pass
*createLAAPass() {
2453 return new LoopAccessLegacyAnalysis();
2456 } // end namespace llvm