1 //===- llvm/Analysis/LoopAccessAnalysis.h -----------------------*- C++ -*-===//
2 //
3 //                     The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This file defines the interface for the loop memory dependence framework that
11 // was originally developed for the Loop Vectorizer.
12 //
13 //===----------------------------------------------------------------------===//
14 
15 #ifndef LLVM_ANALYSIS_LOOPACCESSANALYSIS_H
16 #define LLVM_ANALYSIS_LOOPACCESSANALYSIS_H
17 
18 #include "llvm/ADT/EquivalenceClasses.h"
19 #include "llvm/ADT/Optional.h"
20 #include "llvm/ADT/SetVector.h"
21 #include "llvm/Analysis/AliasAnalysis.h"
22 #include "llvm/Analysis/AliasSetTracker.h"
23 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
24 #include "llvm/IR/ValueHandle.h"
25 #include "llvm/Pass.h"
26 #include "llvm/Support/raw_ostream.h"
27 
28 namespace llvm {
29 
30 class Value;
31 class DataLayout;
32 class ScalarEvolution;
33 class Loop;
34 class SCEV;
35 class SCEVUnionPredicate;
36 class LoopAccessInfo;
37 
38 /// Optimization analysis message produced during vectorization. Messages inform
39 /// the user why vectorization did not occur.
40 class LoopAccessReport {
41   std::string Message;
42   const Instruction *Instr;
43 
44 protected:
LoopAccessReport(const Twine & Message,const Instruction * I)45   LoopAccessReport(const Twine &Message, const Instruction *I)
46       : Message(Message.str()), Instr(I) {}
47 
48 public:
Instr(I)49   LoopAccessReport(const Instruction *I = nullptr) : Instr(I) {}
50 
51   template <typename A> LoopAccessReport &operator<<(const A &Value) {
52     raw_string_ostream Out(Message);
53     Out << Value;
54     return *this;
55   }
56 
getInstr()57   const Instruction *getInstr() const { return Instr; }
58 
str()59   std::string &str() { return Message; }
str()60   const std::string &str() const { return Message; }
Twine()61   operator Twine() { return Message; }
62 
63   /// \brief Emit an analysis note for \p PassName with the debug location from
64   /// the instruction in \p Message if available.  Otherwise use the location of
65   /// \p TheLoop.
66   static void emitAnalysis(const LoopAccessReport &Message,
67                            const Function *TheFunction,
68                            const Loop *TheLoop,
69                            const char *PassName);
70 };
71 
72 /// \brief Collection of parameters shared beetween the Loop Vectorizer and the
73 /// Loop Access Analysis.
74 struct VectorizerParams {
75   /// \brief Maximum SIMD width.
76   static const unsigned MaxVectorWidth;
77 
78   /// \brief VF as overridden by the user.
79   static unsigned VectorizationFactor;
80   /// \brief Interleave factor as overridden by the user.
81   static unsigned VectorizationInterleave;
82   /// \brief True if force-vector-interleave was specified by the user.
83   static bool isInterleaveForced();
84 
85   /// \\brief When performing memory disambiguation checks at runtime do not
86   /// make more than this number of comparisons.
87   static unsigned RuntimeMemoryCheckThreshold;
88 };
89 
90 /// \brief Checks memory dependences among accesses to the same underlying
91 /// object to determine whether there vectorization is legal or not (and at
92 /// which vectorization factor).
93 ///
94 /// Note: This class will compute a conservative dependence for access to
95 /// different underlying pointers. Clients, such as the loop vectorizer, will
96 /// sometimes deal these potential dependencies by emitting runtime checks.
97 ///
98 /// We use the ScalarEvolution framework to symbolically evalutate access
99 /// functions pairs. Since we currently don't restructure the loop we can rely
100 /// on the program order of memory accesses to determine their safety.
101 /// At the moment we will only deem accesses as safe for:
102 ///  * A negative constant distance assuming program order.
103 ///
104 ///      Safe: tmp = a[i + 1];     OR     a[i + 1] = x;
105 ///            a[i] = tmp;                y = a[i];
106 ///
107 ///   The latter case is safe because later checks guarantuee that there can't
108 ///   be a cycle through a phi node (that is, we check that "x" and "y" is not
109 ///   the same variable: a header phi can only be an induction or a reduction, a
110 ///   reduction can't have a memory sink, an induction can't have a memory
111 ///   source). This is important and must not be violated (or we have to
112 ///   resort to checking for cycles through memory).
113 ///
114 ///  * A positive constant distance assuming program order that is bigger
115 ///    than the biggest memory access.
116 ///
117 ///     tmp = a[i]        OR              b[i] = x
118 ///     a[i+2] = tmp                      y = b[i+2];
119 ///
120 ///     Safe distance: 2 x sizeof(a[0]), and 2 x sizeof(b[0]), respectively.
121 ///
122 ///  * Zero distances and all accesses have the same size.
123 ///
124 class MemoryDepChecker {
125 public:
126   typedef PointerIntPair<Value *, 1, bool> MemAccessInfo;
127   typedef SmallPtrSet<MemAccessInfo, 8> MemAccessInfoSet;
128   /// \brief Set of potential dependent memory accesses.
129   typedef EquivalenceClasses<MemAccessInfo> DepCandidates;
130 
131   /// \brief Dependece between memory access instructions.
132   struct Dependence {
133     /// \brief The type of the dependence.
134     enum DepType {
135       // No dependence.
136       NoDep,
137       // We couldn't determine the direction or the distance.
138       Unknown,
139       // Lexically forward.
140       //
141       // FIXME: If we only have loop-independent forward dependences (e.g. a
142       // read and write of A[i]), LAA will locally deem the dependence "safe"
143       // without querying the MemoryDepChecker.  Therefore we can miss
144       // enumerating loop-independent forward dependences in
145       // getDependences.  Note that as soon as there are different
146       // indices used to access the same array, the MemoryDepChecker *is*
147       // queried and the dependence list is complete.
148       Forward,
149       // Forward, but if vectorized, is likely to prevent store-to-load
150       // forwarding.
151       ForwardButPreventsForwarding,
152       // Lexically backward.
153       Backward,
154       // Backward, but the distance allows a vectorization factor of
155       // MaxSafeDepDistBytes.
156       BackwardVectorizable,
157       // Same, but may prevent store-to-load forwarding.
158       BackwardVectorizableButPreventsForwarding
159     };
160 
161     /// \brief String version of the types.
162     static const char *DepName[];
163 
164     /// \brief Index of the source of the dependence in the InstMap vector.
165     unsigned Source;
166     /// \brief Index of the destination of the dependence in the InstMap vector.
167     unsigned Destination;
168     /// \brief The type of the dependence.
169     DepType Type;
170 
DependenceDependence171     Dependence(unsigned Source, unsigned Destination, DepType Type)
172         : Source(Source), Destination(Destination), Type(Type) {}
173 
174     /// \brief Return the source instruction of the dependence.
175     Instruction *getSource(const LoopAccessInfo &LAI) const;
176     /// \brief Return the destination instruction of the dependence.
177     Instruction *getDestination(const LoopAccessInfo &LAI) const;
178 
179     /// \brief Dependence types that don't prevent vectorization.
180     static bool isSafeForVectorization(DepType Type);
181 
182     /// \brief Lexically forward dependence.
183     bool isForward() const;
184     /// \brief Lexically backward dependence.
185     bool isBackward() const;
186 
187     /// \brief May be a lexically backward dependence type (includes Unknown).
188     bool isPossiblyBackward() const;
189 
190     /// \brief Print the dependence.  \p Instr is used to map the instruction
191     /// indices to instructions.
192     void print(raw_ostream &OS, unsigned Depth,
193                const SmallVectorImpl<Instruction *> &Instrs) const;
194   };
195 
MemoryDepChecker(PredicatedScalarEvolution & PSE,const Loop * L)196   MemoryDepChecker(PredicatedScalarEvolution &PSE, const Loop *L)
197       : PSE(PSE), InnermostLoop(L), AccessIdx(0),
198         ShouldRetryWithRuntimeCheck(false), SafeForVectorization(true),
199         RecordDependences(true) {}
200 
201   /// \brief Register the location (instructions are given increasing numbers)
202   /// of a write access.
addAccess(StoreInst * SI)203   void addAccess(StoreInst *SI) {
204     Value *Ptr = SI->getPointerOperand();
205     Accesses[MemAccessInfo(Ptr, true)].push_back(AccessIdx);
206     InstMap.push_back(SI);
207     ++AccessIdx;
208   }
209 
210   /// \brief Register the location (instructions are given increasing numbers)
211   /// of a write access.
addAccess(LoadInst * LI)212   void addAccess(LoadInst *LI) {
213     Value *Ptr = LI->getPointerOperand();
214     Accesses[MemAccessInfo(Ptr, false)].push_back(AccessIdx);
215     InstMap.push_back(LI);
216     ++AccessIdx;
217   }
218 
219   /// \brief Check whether the dependencies between the accesses are safe.
220   ///
221   /// Only checks sets with elements in \p CheckDeps.
222   bool areDepsSafe(DepCandidates &AccessSets, MemAccessInfoSet &CheckDeps,
223                    const ValueToValueMap &Strides);
224 
225   /// \brief No memory dependence was encountered that would inhibit
226   /// vectorization.
isSafeForVectorization()227   bool isSafeForVectorization() const { return SafeForVectorization; }
228 
229   /// \brief The maximum number of bytes of a vector register we can vectorize
230   /// the accesses safely with.
getMaxSafeDepDistBytes()231   unsigned getMaxSafeDepDistBytes() { return MaxSafeDepDistBytes; }
232 
233   /// \brief In same cases when the dependency check fails we can still
234   /// vectorize the loop with a dynamic array access check.
shouldRetryWithRuntimeCheck()235   bool shouldRetryWithRuntimeCheck() { return ShouldRetryWithRuntimeCheck; }
236 
237   /// \brief Returns the memory dependences.  If null is returned we exceeded
238   /// the MaxDependences threshold and this information is not
239   /// available.
getDependences()240   const SmallVectorImpl<Dependence> *getDependences() const {
241     return RecordDependences ? &Dependences : nullptr;
242   }
243 
clearDependences()244   void clearDependences() { Dependences.clear(); }
245 
246   /// \brief The vector of memory access instructions.  The indices are used as
247   /// instruction identifiers in the Dependence class.
getMemoryInstructions()248   const SmallVectorImpl<Instruction *> &getMemoryInstructions() const {
249     return InstMap;
250   }
251 
252   /// \brief Generate a mapping between the memory instructions and their
253   /// indices according to program order.
generateInstructionOrderMap()254   DenseMap<Instruction *, unsigned> generateInstructionOrderMap() const {
255     DenseMap<Instruction *, unsigned> OrderMap;
256 
257     for (unsigned I = 0; I < InstMap.size(); ++I)
258       OrderMap[InstMap[I]] = I;
259 
260     return OrderMap;
261   }
262 
263   /// \brief Find the set of instructions that read or write via \p Ptr.
264   SmallVector<Instruction *, 4> getInstructionsForAccess(Value *Ptr,
265                                                          bool isWrite) const;
266 
267 private:
268   /// A wrapper around ScalarEvolution, used to add runtime SCEV checks, and
269   /// applies dynamic knowledge to simplify SCEV expressions and convert them
270   /// to a more usable form. We need this in case assumptions about SCEV
271   /// expressions need to be made in order to avoid unknown dependences. For
272   /// example we might assume a unit stride for a pointer in order to prove
273   /// that a memory access is strided and doesn't wrap.
274   PredicatedScalarEvolution &PSE;
275   const Loop *InnermostLoop;
276 
277   /// \brief Maps access locations (ptr, read/write) to program order.
278   DenseMap<MemAccessInfo, std::vector<unsigned> > Accesses;
279 
280   /// \brief Memory access instructions in program order.
281   SmallVector<Instruction *, 16> InstMap;
282 
283   /// \brief The program order index to be used for the next instruction.
284   unsigned AccessIdx;
285 
286   // We can access this many bytes in parallel safely.
287   unsigned MaxSafeDepDistBytes;
288 
289   /// \brief If we see a non-constant dependence distance we can still try to
290   /// vectorize this loop with runtime checks.
291   bool ShouldRetryWithRuntimeCheck;
292 
293   /// \brief No memory dependence was encountered that would inhibit
294   /// vectorization.
295   bool SafeForVectorization;
296 
297   //// \brief True if Dependences reflects the dependences in the
298   //// loop.  If false we exceeded MaxDependences and
299   //// Dependences is invalid.
300   bool RecordDependences;
301 
302   /// \brief Memory dependences collected during the analysis.  Only valid if
303   /// RecordDependences is true.
304   SmallVector<Dependence, 8> Dependences;
305 
306   /// \brief Check whether there is a plausible dependence between the two
307   /// accesses.
308   ///
309   /// Access \p A must happen before \p B in program order. The two indices
310   /// identify the index into the program order map.
311   ///
312   /// This function checks  whether there is a plausible dependence (or the
313   /// absence of such can't be proved) between the two accesses. If there is a
314   /// plausible dependence but the dependence distance is bigger than one
315   /// element access it records this distance in \p MaxSafeDepDistBytes (if this
316   /// distance is smaller than any other distance encountered so far).
317   /// Otherwise, this function returns true signaling a possible dependence.
318   Dependence::DepType isDependent(const MemAccessInfo &A, unsigned AIdx,
319                                   const MemAccessInfo &B, unsigned BIdx,
320                                   const ValueToValueMap &Strides);
321 
322   /// \brief Check whether the data dependence could prevent store-load
323   /// forwarding.
324   bool couldPreventStoreLoadForward(unsigned Distance, unsigned TypeByteSize);
325 };
326 
327 /// \brief Holds information about the memory runtime legality checks to verify
328 /// that a group of pointers do not overlap.
329 class RuntimePointerChecking {
330 public:
331   struct PointerInfo {
332     /// Holds the pointer value that we need to check.
333     TrackingVH<Value> PointerValue;
334     /// Holds the pointer value at the beginning of the loop.
335     const SCEV *Start;
336     /// Holds the pointer value at the end of the loop.
337     const SCEV *End;
338     /// Holds the information if this pointer is used for writing to memory.
339     bool IsWritePtr;
340     /// Holds the id of the set of pointers that could be dependent because of a
341     /// shared underlying object.
342     unsigned DependencySetId;
343     /// Holds the id of the disjoint alias set to which this pointer belongs.
344     unsigned AliasSetId;
345     /// SCEV for the access.
346     const SCEV *Expr;
347 
PointerInfoPointerInfo348     PointerInfo(Value *PointerValue, const SCEV *Start, const SCEV *End,
349                 bool IsWritePtr, unsigned DependencySetId, unsigned AliasSetId,
350                 const SCEV *Expr)
351         : PointerValue(PointerValue), Start(Start), End(End),
352           IsWritePtr(IsWritePtr), DependencySetId(DependencySetId),
353           AliasSetId(AliasSetId), Expr(Expr) {}
354   };
355 
RuntimePointerChecking(ScalarEvolution * SE)356   RuntimePointerChecking(ScalarEvolution *SE) : Need(false), SE(SE) {}
357 
358   /// Reset the state of the pointer runtime information.
reset()359   void reset() {
360     Need = false;
361     Pointers.clear();
362     Checks.clear();
363   }
364 
365   /// Insert a pointer and calculate the start and end SCEVs.
366   /// \p We need Preds in order to compute the SCEV expression of the pointer
367   /// according to the assumptions that we've made during the analysis.
368   /// The method might also version the pointer stride according to \p Strides,
369   /// and change \p Preds.
370   void insert(Loop *Lp, Value *Ptr, bool WritePtr, unsigned DepSetId,
371               unsigned ASId, const ValueToValueMap &Strides,
372               PredicatedScalarEvolution &PSE);
373 
374   /// \brief No run-time memory checking is necessary.
empty()375   bool empty() const { return Pointers.empty(); }
376 
377   /// A grouping of pointers. A single memcheck is required between
378   /// two groups.
379   struct CheckingPtrGroup {
380     /// \brief Create a new pointer checking group containing a single
381     /// pointer, with index \p Index in RtCheck.
CheckingPtrGroupCheckingPtrGroup382     CheckingPtrGroup(unsigned Index, RuntimePointerChecking &RtCheck)
383         : RtCheck(RtCheck), High(RtCheck.Pointers[Index].End),
384           Low(RtCheck.Pointers[Index].Start) {
385       Members.push_back(Index);
386     }
387 
388     /// \brief Tries to add the pointer recorded in RtCheck at index
389     /// \p Index to this pointer checking group. We can only add a pointer
390     /// to a checking group if we will still be able to get
391     /// the upper and lower bounds of the check. Returns true in case
392     /// of success, false otherwise.
393     bool addPointer(unsigned Index);
394 
395     /// Constitutes the context of this pointer checking group. For each
396     /// pointer that is a member of this group we will retain the index
397     /// at which it appears in RtCheck.
398     RuntimePointerChecking &RtCheck;
399     /// The SCEV expression which represents the upper bound of all the
400     /// pointers in this group.
401     const SCEV *High;
402     /// The SCEV expression which represents the lower bound of all the
403     /// pointers in this group.
404     const SCEV *Low;
405     /// Indices of all the pointers that constitute this grouping.
406     SmallVector<unsigned, 2> Members;
407   };
408 
409   /// \brief A memcheck which made up of a pair of grouped pointers.
410   ///
411   /// These *have* to be const for now, since checks are generated from
412   /// CheckingPtrGroups in LAI::addRuntimeChecks which is a const member
413   /// function.  FIXME: once check-generation is moved inside this class (after
414   /// the PtrPartition hack is removed), we could drop const.
415   typedef std::pair<const CheckingPtrGroup *, const CheckingPtrGroup *>
416       PointerCheck;
417 
418   /// \brief Generate the checks and store it.  This also performs the grouping
419   /// of pointers to reduce the number of memchecks necessary.
420   void generateChecks(MemoryDepChecker::DepCandidates &DepCands,
421                       bool UseDependencies);
422 
423   /// \brief Returns the checks that generateChecks created.
getChecks()424   const SmallVector<PointerCheck, 4> &getChecks() const { return Checks; }
425 
426   /// \brief Decide if we need to add a check between two groups of pointers,
427   /// according to needsChecking.
428   bool needsChecking(const CheckingPtrGroup &M,
429                      const CheckingPtrGroup &N) const;
430 
431   /// \brief Returns the number of run-time checks required according to
432   /// needsChecking.
getNumberOfChecks()433   unsigned getNumberOfChecks() const { return Checks.size(); }
434 
435   /// \brief Print the list run-time memory checks necessary.
436   void print(raw_ostream &OS, unsigned Depth = 0) const;
437 
438   /// Print \p Checks.
439   void printChecks(raw_ostream &OS, const SmallVectorImpl<PointerCheck> &Checks,
440                    unsigned Depth = 0) const;
441 
442   /// This flag indicates if we need to add the runtime check.
443   bool Need;
444 
445   /// Information about the pointers that may require checking.
446   SmallVector<PointerInfo, 2> Pointers;
447 
448   /// Holds a partitioning of pointers into "check groups".
449   SmallVector<CheckingPtrGroup, 2> CheckingGroups;
450 
451   /// \brief Check if pointers are in the same partition
452   ///
453   /// \p PtrToPartition contains the partition number for pointers (-1 if the
454   /// pointer belongs to multiple partitions).
455   static bool
456   arePointersInSamePartition(const SmallVectorImpl<int> &PtrToPartition,
457                              unsigned PtrIdx1, unsigned PtrIdx2);
458 
459   /// \brief Decide whether we need to issue a run-time check for pointer at
460   /// index \p I and \p J to prove their independence.
461   bool needsChecking(unsigned I, unsigned J) const;
462 
463   /// \brief Return PointerInfo for pointer at index \p PtrIdx.
getPointerInfo(unsigned PtrIdx)464   const PointerInfo &getPointerInfo(unsigned PtrIdx) const {
465     return Pointers[PtrIdx];
466   }
467 
468 private:
469   /// \brief Groups pointers such that a single memcheck is required
470   /// between two different groups. This will clear the CheckingGroups vector
471   /// and re-compute it. We will only group dependecies if \p UseDependencies
472   /// is true, otherwise we will create a separate group for each pointer.
473   void groupChecks(MemoryDepChecker::DepCandidates &DepCands,
474                    bool UseDependencies);
475 
476   /// Generate the checks and return them.
477   SmallVector<PointerCheck, 4>
478   generateChecks() const;
479 
480   /// Holds a pointer to the ScalarEvolution analysis.
481   ScalarEvolution *SE;
482 
483   /// \brief Set of run-time checks required to establish independence of
484   /// otherwise may-aliasing pointers in the loop.
485   SmallVector<PointerCheck, 4> Checks;
486 };
487 
488 /// \brief Drive the analysis of memory accesses in the loop
489 ///
490 /// This class is responsible for analyzing the memory accesses of a loop.  It
491 /// collects the accesses and then its main helper the AccessAnalysis class
492 /// finds and categorizes the dependences in buildDependenceSets.
493 ///
494 /// For memory dependences that can be analyzed at compile time, it determines
495 /// whether the dependence is part of cycle inhibiting vectorization.  This work
496 /// is delegated to the MemoryDepChecker class.
497 ///
498 /// For memory dependences that cannot be determined at compile time, it
499 /// generates run-time checks to prove independence.  This is done by
500 /// AccessAnalysis::canCheckPtrAtRT and the checks are maintained by the
501 /// RuntimePointerCheck class.
502 ///
503 /// If pointers can wrap or can't be expressed as affine AddRec expressions by
504 /// ScalarEvolution, we will generate run-time checks by emitting a
505 /// SCEVUnionPredicate.
506 ///
507 /// Checks for both memory dependences and the SCEV predicates contained in the
508 /// PSE must be emitted in order for the results of this analysis to be valid.
509 class LoopAccessInfo {
510 public:
511   LoopAccessInfo(Loop *L, ScalarEvolution *SE, const DataLayout &DL,
512                  const TargetLibraryInfo *TLI, AliasAnalysis *AA,
513                  DominatorTree *DT, LoopInfo *LI,
514                  const ValueToValueMap &Strides);
515 
516   /// Return true we can analyze the memory accesses in the loop and there are
517   /// no memory dependence cycles.
canVectorizeMemory()518   bool canVectorizeMemory() const { return CanVecMem; }
519 
getRuntimePointerChecking()520   const RuntimePointerChecking *getRuntimePointerChecking() const {
521     return &PtrRtChecking;
522   }
523 
524   /// \brief Number of memchecks required to prove independence of otherwise
525   /// may-alias pointers.
getNumRuntimePointerChecks()526   unsigned getNumRuntimePointerChecks() const {
527     return PtrRtChecking.getNumberOfChecks();
528   }
529 
530   /// Return true if the block BB needs to be predicated in order for the loop
531   /// to be vectorized.
532   static bool blockNeedsPredication(BasicBlock *BB, Loop *TheLoop,
533                                     DominatorTree *DT);
534 
535   /// Returns true if the value V is uniform within the loop.
536   bool isUniform(Value *V) const;
537 
getMaxSafeDepDistBytes()538   unsigned getMaxSafeDepDistBytes() const { return MaxSafeDepDistBytes; }
getNumStores()539   unsigned getNumStores() const { return NumStores; }
getNumLoads()540   unsigned getNumLoads() const { return NumLoads;}
541 
542   /// \brief Add code that checks at runtime if the accessed arrays overlap.
543   ///
544   /// Returns a pair of instructions where the first element is the first
545   /// instruction generated in possibly a sequence of instructions and the
546   /// second value is the final comparator value or NULL if no check is needed.
547   std::pair<Instruction *, Instruction *>
548   addRuntimeChecks(Instruction *Loc) const;
549 
550   /// \brief Generete the instructions for the checks in \p PointerChecks.
551   ///
552   /// Returns a pair of instructions where the first element is the first
553   /// instruction generated in possibly a sequence of instructions and the
554   /// second value is the final comparator value or NULL if no check is needed.
555   std::pair<Instruction *, Instruction *>
556   addRuntimeChecks(Instruction *Loc,
557                    const SmallVectorImpl<RuntimePointerChecking::PointerCheck>
558                        &PointerChecks) const;
559 
560   /// \brief The diagnostics report generated for the analysis.  E.g. why we
561   /// couldn't analyze the loop.
getReport()562   const Optional<LoopAccessReport> &getReport() const { return Report; }
563 
564   /// \brief the Memory Dependence Checker which can determine the
565   /// loop-independent and loop-carried dependences between memory accesses.
getDepChecker()566   const MemoryDepChecker &getDepChecker() const { return DepChecker; }
567 
568   /// \brief Return the list of instructions that use \p Ptr to read or write
569   /// memory.
getInstructionsForAccess(Value * Ptr,bool isWrite)570   SmallVector<Instruction *, 4> getInstructionsForAccess(Value *Ptr,
571                                                          bool isWrite) const {
572     return DepChecker.getInstructionsForAccess(Ptr, isWrite);
573   }
574 
575   /// \brief Print the information about the memory accesses in the loop.
576   void print(raw_ostream &OS, unsigned Depth = 0) const;
577 
578   /// \brief Used to ensure that if the analysis was run with speculating the
579   /// value of symbolic strides, the client queries it with the same assumption.
580   /// Only used in DEBUG build but we don't want NDEBUG-dependent ABI.
581   unsigned NumSymbolicStrides;
582 
583   /// \brief Checks existence of store to invariant address inside loop.
584   /// If the loop has any store to invariant address, then it returns true,
585   /// else returns false.
hasStoreToLoopInvariantAddress()586   bool hasStoreToLoopInvariantAddress() const {
587     return StoreToLoopInvariantAddress;
588   }
589 
590   /// Used to add runtime SCEV checks. Simplifies SCEV expressions and converts
591   /// them to a more usable form.  All SCEV expressions during the analysis
592   /// should be re-written (and therefore simplified) according to PSE.
593   /// A user of LoopAccessAnalysis will need to emit the runtime checks
594   /// associated with this predicate.
595   PredicatedScalarEvolution PSE;
596 
597 private:
598   /// \brief Analyze the loop.  Substitute symbolic strides using Strides.
599   void analyzeLoop(const ValueToValueMap &Strides);
600 
601   /// \brief Check if the structure of the loop allows it to be analyzed by this
602   /// pass.
603   bool canAnalyzeLoop();
604 
605   void emitAnalysis(LoopAccessReport &Message);
606 
607   /// We need to check that all of the pointers in this list are disjoint
608   /// at runtime.
609   RuntimePointerChecking PtrRtChecking;
610 
611   /// \brief the Memory Dependence Checker which can determine the
612   /// loop-independent and loop-carried dependences between memory accesses.
613   MemoryDepChecker DepChecker;
614 
615   Loop *TheLoop;
616   const DataLayout &DL;
617   const TargetLibraryInfo *TLI;
618   AliasAnalysis *AA;
619   DominatorTree *DT;
620   LoopInfo *LI;
621 
622   unsigned NumLoads;
623   unsigned NumStores;
624 
625   unsigned MaxSafeDepDistBytes;
626 
627   /// \brief Cache the result of analyzeLoop.
628   bool CanVecMem;
629 
630   /// \brief Indicator for storing to uniform addresses.
631   /// If a loop has write to a loop invariant address then it should be true.
632   bool StoreToLoopInvariantAddress;
633 
634   /// \brief The diagnostics report generated for the analysis.  E.g. why we
635   /// couldn't analyze the loop.
636   Optional<LoopAccessReport> Report;
637 };
638 
639 Value *stripIntegerCast(Value *V);
640 
641 ///\brief Return the SCEV corresponding to a pointer with the symbolic stride
642 /// replaced with constant one, assuming \p Preds is true.
643 ///
644 /// If necessary this method will version the stride of the pointer according
645 /// to \p PtrToStride and therefore add a new predicate to \p Preds.
646 ///
647 /// If \p OrigPtr is not null, use it to look up the stride value instead of \p
648 /// Ptr.  \p PtrToStride provides the mapping between the pointer value and its
649 /// stride as collected by LoopVectorizationLegality::collectStridedAccess.
650 const SCEV *replaceSymbolicStrideSCEV(PredicatedScalarEvolution &PSE,
651                                       const ValueToValueMap &PtrToStride,
652                                       Value *Ptr, Value *OrigPtr = nullptr);
653 
654 /// \brief Check the stride of the pointer and ensure that it does not wrap in
655 /// the address space, assuming \p Preds is true.
656 ///
657 /// If necessary this method will version the stride of the pointer according
658 /// to \p PtrToStride and therefore add a new predicate to \p Preds.
659 int isStridedPtr(PredicatedScalarEvolution &PSE, Value *Ptr, const Loop *Lp,
660                  const ValueToValueMap &StridesMap);
661 
662 /// \brief This analysis provides dependence information for the memory accesses
663 /// of a loop.
664 ///
665 /// It runs the analysis for a loop on demand.  This can be initiated by
666 /// querying the loop access info via LAA::getInfo.  getInfo return a
667 /// LoopAccessInfo object.  See this class for the specifics of what information
668 /// is provided.
669 class LoopAccessAnalysis : public FunctionPass {
670 public:
671   static char ID;
672 
LoopAccessAnalysis()673   LoopAccessAnalysis() : FunctionPass(ID) {
674     initializeLoopAccessAnalysisPass(*PassRegistry::getPassRegistry());
675   }
676 
677   bool runOnFunction(Function &F) override;
678 
679   void getAnalysisUsage(AnalysisUsage &AU) const override;
680 
681   /// \brief Query the result of the loop access information for the loop \p L.
682   ///
683   /// If the client speculates (and then issues run-time checks) for the values
684   /// of symbolic strides, \p Strides provides the mapping (see
685   /// replaceSymbolicStrideSCEV).  If there is no cached result available run
686   /// the analysis.
687   const LoopAccessInfo &getInfo(Loop *L, const ValueToValueMap &Strides);
688 
releaseMemory()689   void releaseMemory() override {
690     // Invalidate the cache when the pass is freed.
691     LoopAccessInfoMap.clear();
692   }
693 
694   /// \brief Print the result of the analysis when invoked with -analyze.
695   void print(raw_ostream &OS, const Module *M = nullptr) const override;
696 
697 private:
698   /// \brief The cache.
699   DenseMap<Loop *, std::unique_ptr<LoopAccessInfo>> LoopAccessInfoMap;
700 
701   // The used analysis passes.
702   ScalarEvolution *SE;
703   const TargetLibraryInfo *TLI;
704   AliasAnalysis *AA;
705   DominatorTree *DT;
706   LoopInfo *LI;
707 };
708 
getSource(const LoopAccessInfo & LAI)709 inline Instruction *MemoryDepChecker::Dependence::getSource(
710     const LoopAccessInfo &LAI) const {
711   return LAI.getDepChecker().getMemoryInstructions()[Source];
712 }
713 
getDestination(const LoopAccessInfo & LAI)714 inline Instruction *MemoryDepChecker::Dependence::getDestination(
715     const LoopAccessInfo &LAI) const {
716   return LAI.getDepChecker().getMemoryInstructions()[Destination];
717 }
718 
719 } // End llvm namespace
720 
721 #endif
722