1 //===- llvm/Analysis/LoopAccessAnalysis.h -----------------------*- C++ -*-===//
2 //
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
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This file defines the interface for the loop memory dependence framework that
10 // was originally developed for the Loop Vectorizer.
11 //
12 //===----------------------------------------------------------------------===//
13
14 #ifndef LLVM_ANALYSIS_LOOPACCESSANALYSIS_H
15 #define LLVM_ANALYSIS_LOOPACCESSANALYSIS_H
16
17 #include "llvm/ADT/EquivalenceClasses.h"
18 #include "llvm/ADT/Optional.h"
19 #include "llvm/ADT/SetVector.h"
20 #include "llvm/Analysis/AliasAnalysis.h"
21 #include "llvm/Analysis/AliasSetTracker.h"
22 #include "llvm/Analysis/LoopAnalysisManager.h"
23 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
24 #include "llvm/IR/DiagnosticInfo.h"
25 #include "llvm/IR/ValueHandle.h"
26 #include "llvm/Pass.h"
27 #include "llvm/Support/raw_ostream.h"
28
29 namespace llvm {
30
31 class Value;
32 class DataLayout;
33 class ScalarEvolution;
34 class Loop;
35 class SCEV;
36 class SCEVUnionPredicate;
37 class LoopAccessInfo;
38 class OptimizationRemarkEmitter;
39
40 /// Collection of parameters shared beetween the Loop Vectorizer and the
41 /// Loop Access Analysis.
42 struct VectorizerParams {
43 /// Maximum SIMD width.
44 static const unsigned MaxVectorWidth;
45
46 /// VF as overridden by the user.
47 static unsigned VectorizationFactor;
48 /// Interleave factor as overridden by the user.
49 static unsigned VectorizationInterleave;
50 /// True if force-vector-interleave was specified by the user.
51 static bool isInterleaveForced();
52
53 /// \When performing memory disambiguation checks at runtime do not
54 /// make more than this number of comparisons.
55 static unsigned RuntimeMemoryCheckThreshold;
56 };
57
58 /// Checks memory dependences among accesses to the same underlying
59 /// object to determine whether there vectorization is legal or not (and at
60 /// which vectorization factor).
61 ///
62 /// Note: This class will compute a conservative dependence for access to
63 /// different underlying pointers. Clients, such as the loop vectorizer, will
64 /// sometimes deal these potential dependencies by emitting runtime checks.
65 ///
66 /// We use the ScalarEvolution framework to symbolically evalutate access
67 /// functions pairs. Since we currently don't restructure the loop we can rely
68 /// on the program order of memory accesses to determine their safety.
69 /// At the moment we will only deem accesses as safe for:
70 /// * A negative constant distance assuming program order.
71 ///
72 /// Safe: tmp = a[i + 1]; OR a[i + 1] = x;
73 /// a[i] = tmp; y = a[i];
74 ///
75 /// The latter case is safe because later checks guarantuee that there can't
76 /// be a cycle through a phi node (that is, we check that "x" and "y" is not
77 /// the same variable: a header phi can only be an induction or a reduction, a
78 /// reduction can't have a memory sink, an induction can't have a memory
79 /// source). This is important and must not be violated (or we have to
80 /// resort to checking for cycles through memory).
81 ///
82 /// * A positive constant distance assuming program order that is bigger
83 /// than the biggest memory access.
84 ///
85 /// tmp = a[i] OR b[i] = x
86 /// a[i+2] = tmp y = b[i+2];
87 ///
88 /// Safe distance: 2 x sizeof(a[0]), and 2 x sizeof(b[0]), respectively.
89 ///
90 /// * Zero distances and all accesses have the same size.
91 ///
92 class MemoryDepChecker {
93 public:
94 typedef PointerIntPair<Value *, 1, bool> MemAccessInfo;
95 typedef SmallVector<MemAccessInfo, 8> MemAccessInfoList;
96 /// Set of potential dependent memory accesses.
97 typedef EquivalenceClasses<MemAccessInfo> DepCandidates;
98
99 /// Type to keep track of the status of the dependence check. The order of
100 /// the elements is important and has to be from most permissive to least
101 /// permissive.
102 enum class VectorizationSafetyStatus {
103 // Can vectorize safely without RT checks. All dependences are known to be
104 // safe.
105 Safe,
106 // Can possibly vectorize with RT checks to overcome unknown dependencies.
107 PossiblySafeWithRtChecks,
108 // Cannot vectorize due to known unsafe dependencies.
109 Unsafe,
110 };
111
112 /// Dependece between memory access instructions.
113 struct Dependence {
114 /// The type of the dependence.
115 enum DepType {
116 // No dependence.
117 NoDep,
118 // We couldn't determine the direction or the distance.
119 Unknown,
120 // Lexically forward.
121 //
122 // FIXME: If we only have loop-independent forward dependences (e.g. a
123 // read and write of A[i]), LAA will locally deem the dependence "safe"
124 // without querying the MemoryDepChecker. Therefore we can miss
125 // enumerating loop-independent forward dependences in
126 // getDependences. Note that as soon as there are different
127 // indices used to access the same array, the MemoryDepChecker *is*
128 // queried and the dependence list is complete.
129 Forward,
130 // Forward, but if vectorized, is likely to prevent store-to-load
131 // forwarding.
132 ForwardButPreventsForwarding,
133 // Lexically backward.
134 Backward,
135 // Backward, but the distance allows a vectorization factor of
136 // MaxSafeDepDistBytes.
137 BackwardVectorizable,
138 // Same, but may prevent store-to-load forwarding.
139 BackwardVectorizableButPreventsForwarding
140 };
141
142 /// String version of the types.
143 static const char *DepName[];
144
145 /// Index of the source of the dependence in the InstMap vector.
146 unsigned Source;
147 /// Index of the destination of the dependence in the InstMap vector.
148 unsigned Destination;
149 /// The type of the dependence.
150 DepType Type;
151
DependenceDependence152 Dependence(unsigned Source, unsigned Destination, DepType Type)
153 : Source(Source), Destination(Destination), Type(Type) {}
154
155 /// Return the source instruction of the dependence.
156 Instruction *getSource(const LoopAccessInfo &LAI) const;
157 /// Return the destination instruction of the dependence.
158 Instruction *getDestination(const LoopAccessInfo &LAI) const;
159
160 /// Dependence types that don't prevent vectorization.
161 static VectorizationSafetyStatus isSafeForVectorization(DepType Type);
162
163 /// Lexically forward dependence.
164 bool isForward() const;
165 /// Lexically backward dependence.
166 bool isBackward() const;
167
168 /// May be a lexically backward dependence type (includes Unknown).
169 bool isPossiblyBackward() const;
170
171 /// Print the dependence. \p Instr is used to map the instruction
172 /// indices to instructions.
173 void print(raw_ostream &OS, unsigned Depth,
174 const SmallVectorImpl<Instruction *> &Instrs) const;
175 };
176
MemoryDepChecker(PredicatedScalarEvolution & PSE,const Loop * L)177 MemoryDepChecker(PredicatedScalarEvolution &PSE, const Loop *L)
178 : PSE(PSE), InnermostLoop(L), AccessIdx(0), MaxSafeDepDistBytes(0),
179 MaxSafeRegisterWidth(-1U), FoundNonConstantDistanceDependence(false),
180 Status(VectorizationSafetyStatus::Safe), RecordDependences(true) {}
181
182 /// Register the location (instructions are given increasing numbers)
183 /// of a write access.
addAccess(StoreInst * SI)184 void addAccess(StoreInst *SI) {
185 Value *Ptr = SI->getPointerOperand();
186 Accesses[MemAccessInfo(Ptr, true)].push_back(AccessIdx);
187 InstMap.push_back(SI);
188 ++AccessIdx;
189 }
190
191 /// Register the location (instructions are given increasing numbers)
192 /// of a write access.
addAccess(LoadInst * LI)193 void addAccess(LoadInst *LI) {
194 Value *Ptr = LI->getPointerOperand();
195 Accesses[MemAccessInfo(Ptr, false)].push_back(AccessIdx);
196 InstMap.push_back(LI);
197 ++AccessIdx;
198 }
199
200 /// Check whether the dependencies between the accesses are safe.
201 ///
202 /// Only checks sets with elements in \p CheckDeps.
203 bool areDepsSafe(DepCandidates &AccessSets, MemAccessInfoList &CheckDeps,
204 const ValueToValueMap &Strides);
205
206 /// No memory dependence was encountered that would inhibit
207 /// vectorization.
isSafeForVectorization()208 bool isSafeForVectorization() const {
209 return Status == VectorizationSafetyStatus::Safe;
210 }
211
212 /// The maximum number of bytes of a vector register we can vectorize
213 /// the accesses safely with.
getMaxSafeDepDistBytes()214 uint64_t getMaxSafeDepDistBytes() { return MaxSafeDepDistBytes; }
215
216 /// Return the number of elements that are safe to operate on
217 /// simultaneously, multiplied by the size of the element in bits.
getMaxSafeRegisterWidth()218 uint64_t getMaxSafeRegisterWidth() const { return MaxSafeRegisterWidth; }
219
220 /// In same cases when the dependency check fails we can still
221 /// vectorize the loop with a dynamic array access check.
shouldRetryWithRuntimeCheck()222 bool shouldRetryWithRuntimeCheck() const {
223 return FoundNonConstantDistanceDependence &&
224 Status == VectorizationSafetyStatus::PossiblySafeWithRtChecks;
225 }
226
227 /// Returns the memory dependences. If null is returned we exceeded
228 /// the MaxDependences threshold and this information is not
229 /// available.
getDependences()230 const SmallVectorImpl<Dependence> *getDependences() const {
231 return RecordDependences ? &Dependences : nullptr;
232 }
233
clearDependences()234 void clearDependences() { Dependences.clear(); }
235
236 /// The vector of memory access instructions. The indices are used as
237 /// instruction identifiers in the Dependence class.
getMemoryInstructions()238 const SmallVectorImpl<Instruction *> &getMemoryInstructions() const {
239 return InstMap;
240 }
241
242 /// Generate a mapping between the memory instructions and their
243 /// indices according to program order.
generateInstructionOrderMap()244 DenseMap<Instruction *, unsigned> generateInstructionOrderMap() const {
245 DenseMap<Instruction *, unsigned> OrderMap;
246
247 for (unsigned I = 0; I < InstMap.size(); ++I)
248 OrderMap[InstMap[I]] = I;
249
250 return OrderMap;
251 }
252
253 /// Find the set of instructions that read or write via \p Ptr.
254 SmallVector<Instruction *, 4> getInstructionsForAccess(Value *Ptr,
255 bool isWrite) const;
256
257 private:
258 /// A wrapper around ScalarEvolution, used to add runtime SCEV checks, and
259 /// applies dynamic knowledge to simplify SCEV expressions and convert them
260 /// to a more usable form. We need this in case assumptions about SCEV
261 /// expressions need to be made in order to avoid unknown dependences. For
262 /// example we might assume a unit stride for a pointer in order to prove
263 /// that a memory access is strided and doesn't wrap.
264 PredicatedScalarEvolution &PSE;
265 const Loop *InnermostLoop;
266
267 /// Maps access locations (ptr, read/write) to program order.
268 DenseMap<MemAccessInfo, std::vector<unsigned> > Accesses;
269
270 /// Memory access instructions in program order.
271 SmallVector<Instruction *, 16> InstMap;
272
273 /// The program order index to be used for the next instruction.
274 unsigned AccessIdx;
275
276 // We can access this many bytes in parallel safely.
277 uint64_t MaxSafeDepDistBytes;
278
279 /// Number of elements (from consecutive iterations) that are safe to
280 /// operate on simultaneously, multiplied by the size of the element in bits.
281 /// The size of the element is taken from the memory access that is most
282 /// restrictive.
283 uint64_t MaxSafeRegisterWidth;
284
285 /// If we see a non-constant dependence distance we can still try to
286 /// vectorize this loop with runtime checks.
287 bool FoundNonConstantDistanceDependence;
288
289 /// Result of the dependence checks, indicating whether the checked
290 /// dependences are safe for vectorization, require RT checks or are known to
291 /// be unsafe.
292 VectorizationSafetyStatus Status;
293
294 //// True if Dependences reflects the dependences in the
295 //// loop. If false we exceeded MaxDependences and
296 //// Dependences is invalid.
297 bool RecordDependences;
298
299 /// Memory dependences collected during the analysis. Only valid if
300 /// RecordDependences is true.
301 SmallVector<Dependence, 8> Dependences;
302
303 /// Check whether there is a plausible dependence between the two
304 /// accesses.
305 ///
306 /// Access \p A must happen before \p B in program order. The two indices
307 /// identify the index into the program order map.
308 ///
309 /// This function checks whether there is a plausible dependence (or the
310 /// absence of such can't be proved) between the two accesses. If there is a
311 /// plausible dependence but the dependence distance is bigger than one
312 /// element access it records this distance in \p MaxSafeDepDistBytes (if this
313 /// distance is smaller than any other distance encountered so far).
314 /// Otherwise, this function returns true signaling a possible dependence.
315 Dependence::DepType isDependent(const MemAccessInfo &A, unsigned AIdx,
316 const MemAccessInfo &B, unsigned BIdx,
317 const ValueToValueMap &Strides);
318
319 /// Check whether the data dependence could prevent store-load
320 /// forwarding.
321 ///
322 /// \return false if we shouldn't vectorize at all or avoid larger
323 /// vectorization factors by limiting MaxSafeDepDistBytes.
324 bool couldPreventStoreLoadForward(uint64_t Distance, uint64_t TypeByteSize);
325
326 /// Updates the current safety status with \p S. We can go from Safe to
327 /// either PossiblySafeWithRtChecks or Unsafe and from
328 /// PossiblySafeWithRtChecks to Unsafe.
329 void mergeInStatus(VectorizationSafetyStatus S);
330 };
331
332 /// Holds information about the memory runtime legality checks to verify
333 /// that a group of pointers do not overlap.
334 class RuntimePointerChecking {
335 public:
336 struct PointerInfo {
337 /// Holds the pointer value that we need to check.
338 TrackingVH<Value> PointerValue;
339 /// Holds the smallest byte address accessed by the pointer throughout all
340 /// iterations of the loop.
341 const SCEV *Start;
342 /// Holds the largest byte address accessed by the pointer throughout all
343 /// iterations of the loop, plus 1.
344 const SCEV *End;
345 /// Holds the information if this pointer is used for writing to memory.
346 bool IsWritePtr;
347 /// Holds the id of the set of pointers that could be dependent because of a
348 /// shared underlying object.
349 unsigned DependencySetId;
350 /// Holds the id of the disjoint alias set to which this pointer belongs.
351 unsigned AliasSetId;
352 /// SCEV for the access.
353 const SCEV *Expr;
354
PointerInfoPointerInfo355 PointerInfo(Value *PointerValue, const SCEV *Start, const SCEV *End,
356 bool IsWritePtr, unsigned DependencySetId, unsigned AliasSetId,
357 const SCEV *Expr)
358 : PointerValue(PointerValue), Start(Start), End(End),
359 IsWritePtr(IsWritePtr), DependencySetId(DependencySetId),
360 AliasSetId(AliasSetId), Expr(Expr) {}
361 };
362
RuntimePointerChecking(ScalarEvolution * SE)363 RuntimePointerChecking(ScalarEvolution *SE) : Need(false), SE(SE) {}
364
365 /// Reset the state of the pointer runtime information.
reset()366 void reset() {
367 Need = false;
368 Pointers.clear();
369 Checks.clear();
370 }
371
372 /// Insert a pointer and calculate the start and end SCEVs.
373 /// We need \p PSE in order to compute the SCEV expression of the pointer
374 /// according to the assumptions that we've made during the analysis.
375 /// The method might also version the pointer stride according to \p Strides,
376 /// and add new predicates to \p PSE.
377 void insert(Loop *Lp, Value *Ptr, bool WritePtr, unsigned DepSetId,
378 unsigned ASId, const ValueToValueMap &Strides,
379 PredicatedScalarEvolution &PSE);
380
381 /// No run-time memory checking is necessary.
empty()382 bool empty() const { return Pointers.empty(); }
383
384 /// A grouping of pointers. A single memcheck is required between
385 /// two groups.
386 struct CheckingPtrGroup {
387 /// Create a new pointer checking group containing a single
388 /// pointer, with index \p Index in RtCheck.
CheckingPtrGroupCheckingPtrGroup389 CheckingPtrGroup(unsigned Index, RuntimePointerChecking &RtCheck)
390 : RtCheck(RtCheck), High(RtCheck.Pointers[Index].End),
391 Low(RtCheck.Pointers[Index].Start) {
392 Members.push_back(Index);
393 }
394
395 /// Tries to add the pointer recorded in RtCheck at index
396 /// \p Index to this pointer checking group. We can only add a pointer
397 /// to a checking group if we will still be able to get
398 /// the upper and lower bounds of the check. Returns true in case
399 /// of success, false otherwise.
400 bool addPointer(unsigned Index);
401
402 /// Constitutes the context of this pointer checking group. For each
403 /// pointer that is a member of this group we will retain the index
404 /// at which it appears in RtCheck.
405 RuntimePointerChecking &RtCheck;
406 /// The SCEV expression which represents the upper bound of all the
407 /// pointers in this group.
408 const SCEV *High;
409 /// The SCEV expression which represents the lower bound of all the
410 /// pointers in this group.
411 const SCEV *Low;
412 /// Indices of all the pointers that constitute this grouping.
413 SmallVector<unsigned, 2> Members;
414 };
415
416 /// A memcheck which made up of a pair of grouped pointers.
417 ///
418 /// These *have* to be const for now, since checks are generated from
419 /// CheckingPtrGroups in LAI::addRuntimeChecks which is a const member
420 /// function. FIXME: once check-generation is moved inside this class (after
421 /// the PtrPartition hack is removed), we could drop const.
422 typedef std::pair<const CheckingPtrGroup *, const CheckingPtrGroup *>
423 PointerCheck;
424
425 /// Generate the checks and store it. This also performs the grouping
426 /// of pointers to reduce the number of memchecks necessary.
427 void generateChecks(MemoryDepChecker::DepCandidates &DepCands,
428 bool UseDependencies);
429
430 /// Returns the checks that generateChecks created.
getChecks()431 const SmallVector<PointerCheck, 4> &getChecks() const { return Checks; }
432
433 /// Decide if we need to add a check between two groups of pointers,
434 /// according to needsChecking.
435 bool needsChecking(const CheckingPtrGroup &M,
436 const CheckingPtrGroup &N) const;
437
438 /// Returns the number of run-time checks required according to
439 /// needsChecking.
getNumberOfChecks()440 unsigned getNumberOfChecks() const { return Checks.size(); }
441
442 /// Print the list run-time memory checks necessary.
443 void print(raw_ostream &OS, unsigned Depth = 0) const;
444
445 /// Print \p Checks.
446 void printChecks(raw_ostream &OS, const SmallVectorImpl<PointerCheck> &Checks,
447 unsigned Depth = 0) const;
448
449 /// This flag indicates if we need to add the runtime check.
450 bool Need;
451
452 /// Information about the pointers that may require checking.
453 SmallVector<PointerInfo, 2> Pointers;
454
455 /// Holds a partitioning of pointers into "check groups".
456 SmallVector<CheckingPtrGroup, 2> CheckingGroups;
457
458 /// Check if pointers are in the same partition
459 ///
460 /// \p PtrToPartition contains the partition number for pointers (-1 if the
461 /// pointer belongs to multiple partitions).
462 static bool
463 arePointersInSamePartition(const SmallVectorImpl<int> &PtrToPartition,
464 unsigned PtrIdx1, unsigned PtrIdx2);
465
466 /// Decide whether we need to issue a run-time check for pointer at
467 /// index \p I and \p J to prove their independence.
468 bool needsChecking(unsigned I, unsigned J) const;
469
470 /// Return PointerInfo for pointer at index \p PtrIdx.
getPointerInfo(unsigned PtrIdx)471 const PointerInfo &getPointerInfo(unsigned PtrIdx) const {
472 return Pointers[PtrIdx];
473 }
474
475 private:
476 /// Groups pointers such that a single memcheck is required
477 /// between two different groups. This will clear the CheckingGroups vector
478 /// and re-compute it. We will only group dependecies if \p UseDependencies
479 /// is true, otherwise we will create a separate group for each pointer.
480 void groupChecks(MemoryDepChecker::DepCandidates &DepCands,
481 bool UseDependencies);
482
483 /// Generate the checks and return them.
484 SmallVector<PointerCheck, 4>
485 generateChecks() const;
486
487 /// Holds a pointer to the ScalarEvolution analysis.
488 ScalarEvolution *SE;
489
490 /// Set of run-time checks required to establish independence of
491 /// otherwise may-aliasing pointers in the loop.
492 SmallVector<PointerCheck, 4> Checks;
493 };
494
495 /// Drive the analysis of memory accesses in the loop
496 ///
497 /// This class is responsible for analyzing the memory accesses of a loop. It
498 /// collects the accesses and then its main helper the AccessAnalysis class
499 /// finds and categorizes the dependences in buildDependenceSets.
500 ///
501 /// For memory dependences that can be analyzed at compile time, it determines
502 /// whether the dependence is part of cycle inhibiting vectorization. This work
503 /// is delegated to the MemoryDepChecker class.
504 ///
505 /// For memory dependences that cannot be determined at compile time, it
506 /// generates run-time checks to prove independence. This is done by
507 /// AccessAnalysis::canCheckPtrAtRT and the checks are maintained by the
508 /// RuntimePointerCheck class.
509 ///
510 /// If pointers can wrap or can't be expressed as affine AddRec expressions by
511 /// ScalarEvolution, we will generate run-time checks by emitting a
512 /// SCEVUnionPredicate.
513 ///
514 /// Checks for both memory dependences and the SCEV predicates contained in the
515 /// PSE must be emitted in order for the results of this analysis to be valid.
516 class LoopAccessInfo {
517 public:
518 LoopAccessInfo(Loop *L, ScalarEvolution *SE, const TargetLibraryInfo *TLI,
519 AliasAnalysis *AA, DominatorTree *DT, LoopInfo *LI);
520
521 /// Return true we can analyze the memory accesses in the loop and there are
522 /// no memory dependence cycles.
canVectorizeMemory()523 bool canVectorizeMemory() const { return CanVecMem; }
524
525 /// Return true if there is a convergent operation in the loop. There may
526 /// still be reported runtime pointer checks that would be required, but it is
527 /// not legal to insert them.
hasConvergentOp()528 bool hasConvergentOp() const { return HasConvergentOp; }
529
getRuntimePointerChecking()530 const RuntimePointerChecking *getRuntimePointerChecking() const {
531 return PtrRtChecking.get();
532 }
533
534 /// Number of memchecks required to prove independence of otherwise
535 /// may-alias pointers.
getNumRuntimePointerChecks()536 unsigned getNumRuntimePointerChecks() const {
537 return PtrRtChecking->getNumberOfChecks();
538 }
539
540 /// Return true if the block BB needs to be predicated in order for the loop
541 /// to be vectorized.
542 static bool blockNeedsPredication(BasicBlock *BB, Loop *TheLoop,
543 DominatorTree *DT);
544
545 /// Returns true if the value V is uniform within the loop.
546 bool isUniform(Value *V) const;
547
getMaxSafeDepDistBytes()548 uint64_t getMaxSafeDepDistBytes() const { return MaxSafeDepDistBytes; }
getNumStores()549 unsigned getNumStores() const { return NumStores; }
getNumLoads()550 unsigned getNumLoads() const { return NumLoads;}
551
552 /// Add code that checks at runtime if the accessed arrays overlap.
553 ///
554 /// Returns a pair of instructions where the first element is the first
555 /// instruction generated in possibly a sequence of instructions and the
556 /// second value is the final comparator value or NULL if no check is needed.
557 std::pair<Instruction *, Instruction *>
558 addRuntimeChecks(Instruction *Loc) const;
559
560 /// Generete the instructions for the checks in \p PointerChecks.
561 ///
562 /// Returns a pair of instructions where the first element is the first
563 /// instruction generated in possibly a sequence of instructions and the
564 /// second value is the final comparator value or NULL if no check is needed.
565 std::pair<Instruction *, Instruction *>
566 addRuntimeChecks(Instruction *Loc,
567 const SmallVectorImpl<RuntimePointerChecking::PointerCheck>
568 &PointerChecks) const;
569
570 /// The diagnostics report generated for the analysis. E.g. why we
571 /// couldn't analyze the loop.
getReport()572 const OptimizationRemarkAnalysis *getReport() const { return Report.get(); }
573
574 /// the Memory Dependence Checker which can determine the
575 /// loop-independent and loop-carried dependences between memory accesses.
getDepChecker()576 const MemoryDepChecker &getDepChecker() const { return *DepChecker; }
577
578 /// Return the list of instructions that use \p Ptr to read or write
579 /// memory.
getInstructionsForAccess(Value * Ptr,bool isWrite)580 SmallVector<Instruction *, 4> getInstructionsForAccess(Value *Ptr,
581 bool isWrite) const {
582 return DepChecker->getInstructionsForAccess(Ptr, isWrite);
583 }
584
585 /// If an access has a symbolic strides, this maps the pointer value to
586 /// the stride symbol.
getSymbolicStrides()587 const ValueToValueMap &getSymbolicStrides() const { return SymbolicStrides; }
588
589 /// Pointer has a symbolic stride.
hasStride(Value * V)590 bool hasStride(Value *V) const { return StrideSet.count(V); }
591
592 /// Print the information about the memory accesses in the loop.
593 void print(raw_ostream &OS, unsigned Depth = 0) const;
594
595 /// If the loop has memory dependence involving an invariant address, i.e. two
596 /// stores or a store and a load, then return true, else return false.
hasDependenceInvolvingLoopInvariantAddress()597 bool hasDependenceInvolvingLoopInvariantAddress() const {
598 return HasDependenceInvolvingLoopInvariantAddress;
599 }
600
601 /// Used to add runtime SCEV checks. Simplifies SCEV expressions and converts
602 /// them to a more usable form. All SCEV expressions during the analysis
603 /// should be re-written (and therefore simplified) according to PSE.
604 /// A user of LoopAccessAnalysis will need to emit the runtime checks
605 /// associated with this predicate.
getPSE()606 const PredicatedScalarEvolution &getPSE() const { return *PSE; }
607
608 private:
609 /// Analyze the loop.
610 void analyzeLoop(AliasAnalysis *AA, LoopInfo *LI,
611 const TargetLibraryInfo *TLI, DominatorTree *DT);
612
613 /// Check if the structure of the loop allows it to be analyzed by this
614 /// pass.
615 bool canAnalyzeLoop();
616
617 /// Save the analysis remark.
618 ///
619 /// LAA does not directly emits the remarks. Instead it stores it which the
620 /// client can retrieve and presents as its own analysis
621 /// (e.g. -Rpass-analysis=loop-vectorize).
622 OptimizationRemarkAnalysis &recordAnalysis(StringRef RemarkName,
623 Instruction *Instr = nullptr);
624
625 /// Collect memory access with loop invariant strides.
626 ///
627 /// Looks for accesses like "a[i * StrideA]" where "StrideA" is loop
628 /// invariant.
629 void collectStridedAccess(Value *LoadOrStoreInst);
630
631 std::unique_ptr<PredicatedScalarEvolution> PSE;
632
633 /// We need to check that all of the pointers in this list are disjoint
634 /// at runtime. Using std::unique_ptr to make using move ctor simpler.
635 std::unique_ptr<RuntimePointerChecking> PtrRtChecking;
636
637 /// the Memory Dependence Checker which can determine the
638 /// loop-independent and loop-carried dependences between memory accesses.
639 std::unique_ptr<MemoryDepChecker> DepChecker;
640
641 Loop *TheLoop;
642
643 unsigned NumLoads;
644 unsigned NumStores;
645
646 uint64_t MaxSafeDepDistBytes;
647
648 /// Cache the result of analyzeLoop.
649 bool CanVecMem;
650 bool HasConvergentOp;
651
652 /// Indicator that there are non vectorizable stores to a uniform address.
653 bool HasDependenceInvolvingLoopInvariantAddress;
654
655 /// The diagnostics report generated for the analysis. E.g. why we
656 /// couldn't analyze the loop.
657 std::unique_ptr<OptimizationRemarkAnalysis> Report;
658
659 /// If an access has a symbolic strides, this maps the pointer value to
660 /// the stride symbol.
661 ValueToValueMap SymbolicStrides;
662
663 /// Set of symbolic strides values.
664 SmallPtrSet<Value *, 8> StrideSet;
665 };
666
667 Value *stripIntegerCast(Value *V);
668
669 /// Return the SCEV corresponding to a pointer with the symbolic stride
670 /// replaced with constant one, assuming the SCEV predicate associated with
671 /// \p PSE is true.
672 ///
673 /// If necessary this method will version the stride of the pointer according
674 /// to \p PtrToStride and therefore add further predicates to \p PSE.
675 ///
676 /// If \p OrigPtr is not null, use it to look up the stride value instead of \p
677 /// Ptr. \p PtrToStride provides the mapping between the pointer value and its
678 /// stride as collected by LoopVectorizationLegality::collectStridedAccess.
679 const SCEV *replaceSymbolicStrideSCEV(PredicatedScalarEvolution &PSE,
680 const ValueToValueMap &PtrToStride,
681 Value *Ptr, Value *OrigPtr = nullptr);
682
683 /// If the pointer has a constant stride return it in units of its
684 /// element size. Otherwise return zero.
685 ///
686 /// Ensure that it does not wrap in the address space, assuming the predicate
687 /// associated with \p PSE is true.
688 ///
689 /// If necessary this method will version the stride of the pointer according
690 /// to \p PtrToStride and therefore add further predicates to \p PSE.
691 /// The \p Assume parameter indicates if we are allowed to make additional
692 /// run-time assumptions.
693 int64_t getPtrStride(PredicatedScalarEvolution &PSE, Value *Ptr, const Loop *Lp,
694 const ValueToValueMap &StridesMap = ValueToValueMap(),
695 bool Assume = false, bool ShouldCheckWrap = true);
696
697 /// Attempt to sort the pointers in \p VL and return the sorted indices
698 /// in \p SortedIndices, if reordering is required.
699 ///
700 /// Returns 'true' if sorting is legal, otherwise returns 'false'.
701 ///
702 /// For example, for a given \p VL of memory accesses in program order, a[i+4],
703 /// a[i+0], a[i+1] and a[i+7], this function will sort the \p VL and save the
704 /// sorted indices in \p SortedIndices as a[i+0], a[i+1], a[i+4], a[i+7] and
705 /// saves the mask for actual memory accesses in program order in
706 /// \p SortedIndices as <1,2,0,3>
707 bool sortPtrAccesses(ArrayRef<Value *> VL, const DataLayout &DL,
708 ScalarEvolution &SE,
709 SmallVectorImpl<unsigned> &SortedIndices);
710
711 /// Returns true if the memory operations \p A and \p B are consecutive.
712 /// This is a simple API that does not depend on the analysis pass.
713 bool isConsecutiveAccess(Value *A, Value *B, const DataLayout &DL,
714 ScalarEvolution &SE, bool CheckType = true);
715
716 /// This analysis provides dependence information for the memory accesses
717 /// of a loop.
718 ///
719 /// It runs the analysis for a loop on demand. This can be initiated by
720 /// querying the loop access info via LAA::getInfo. getInfo return a
721 /// LoopAccessInfo object. See this class for the specifics of what information
722 /// is provided.
723 class LoopAccessLegacyAnalysis : public FunctionPass {
724 public:
725 static char ID;
726
727 LoopAccessLegacyAnalysis();
728
729 bool runOnFunction(Function &F) override;
730
731 void getAnalysisUsage(AnalysisUsage &AU) const override;
732
733 /// Query the result of the loop access information for the loop \p L.
734 ///
735 /// If there is no cached result available run the analysis.
736 const LoopAccessInfo &getInfo(Loop *L);
737
releaseMemory()738 void releaseMemory() override {
739 // Invalidate the cache when the pass is freed.
740 LoopAccessInfoMap.clear();
741 }
742
743 /// Print the result of the analysis when invoked with -analyze.
744 void print(raw_ostream &OS, const Module *M = nullptr) const override;
745
746 private:
747 /// The cache.
748 DenseMap<Loop *, std::unique_ptr<LoopAccessInfo>> LoopAccessInfoMap;
749
750 // The used analysis passes.
751 ScalarEvolution *SE = nullptr;
752 const TargetLibraryInfo *TLI = nullptr;
753 AliasAnalysis *AA = nullptr;
754 DominatorTree *DT = nullptr;
755 LoopInfo *LI = nullptr;
756 };
757
758 /// This analysis provides dependence information for the memory
759 /// accesses of a loop.
760 ///
761 /// It runs the analysis for a loop on demand. This can be initiated by
762 /// querying the loop access info via AM.getResult<LoopAccessAnalysis>.
763 /// getResult return a LoopAccessInfo object. See this class for the
764 /// specifics of what information is provided.
765 class LoopAccessAnalysis
766 : public AnalysisInfoMixin<LoopAccessAnalysis> {
767 friend AnalysisInfoMixin<LoopAccessAnalysis>;
768 static AnalysisKey Key;
769
770 public:
771 typedef LoopAccessInfo Result;
772
773 Result run(Loop &L, LoopAnalysisManager &AM, LoopStandardAnalysisResults &AR);
774 };
775
getSource(const LoopAccessInfo & LAI)776 inline Instruction *MemoryDepChecker::Dependence::getSource(
777 const LoopAccessInfo &LAI) const {
778 return LAI.getDepChecker().getMemoryInstructions()[Source];
779 }
780
getDestination(const LoopAccessInfo & LAI)781 inline Instruction *MemoryDepChecker::Dependence::getDestination(
782 const LoopAccessInfo &LAI) const {
783 return LAI.getDepChecker().getMemoryInstructions()[Destination];
784 }
785
786 } // End llvm namespace
787
788 #endif
789