xref: /freebsd/contrib/llvm-project/llvm/lib/Analysis/LoopAccessAnalysis.cpp (revision 8aac90f18aef7c9eea906c3ff9a001ca7b94f375)
1 //===- LoopAccessAnalysis.cpp - Loop Access Analysis Implementation --------==//
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 // The implementation for the loop memory dependence that was originally
10 // developed for the loop vectorizer.
11 //
12 //===----------------------------------------------------------------------===//
13 
14 #include "llvm/Analysis/LoopAccessAnalysis.h"
15 #include "llvm/ADT/APInt.h"
16 #include "llvm/ADT/DenseMap.h"
17 #include "llvm/ADT/EquivalenceClasses.h"
18 #include "llvm/ADT/PointerIntPair.h"
19 #include "llvm/ADT/STLExtras.h"
20 #include "llvm/ADT/SetVector.h"
21 #include "llvm/ADT/SmallPtrSet.h"
22 #include "llvm/ADT/SmallSet.h"
23 #include "llvm/ADT/SmallVector.h"
24 #include "llvm/Analysis/AliasAnalysis.h"
25 #include "llvm/Analysis/AliasSetTracker.h"
26 #include "llvm/Analysis/LoopAnalysisManager.h"
27 #include "llvm/Analysis/LoopInfo.h"
28 #include "llvm/Analysis/LoopIterator.h"
29 #include "llvm/Analysis/MemoryLocation.h"
30 #include "llvm/Analysis/OptimizationRemarkEmitter.h"
31 #include "llvm/Analysis/ScalarEvolution.h"
32 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
33 #include "llvm/Analysis/TargetLibraryInfo.h"
34 #include "llvm/Analysis/ValueTracking.h"
35 #include "llvm/Analysis/VectorUtils.h"
36 #include "llvm/IR/BasicBlock.h"
37 #include "llvm/IR/Constants.h"
38 #include "llvm/IR/DataLayout.h"
39 #include "llvm/IR/DebugLoc.h"
40 #include "llvm/IR/DerivedTypes.h"
41 #include "llvm/IR/DiagnosticInfo.h"
42 #include "llvm/IR/Dominators.h"
43 #include "llvm/IR/Function.h"
44 #include "llvm/IR/GetElementPtrTypeIterator.h"
45 #include "llvm/IR/InstrTypes.h"
46 #include "llvm/IR/Instruction.h"
47 #include "llvm/IR/Instructions.h"
48 #include "llvm/IR/Operator.h"
49 #include "llvm/IR/PassManager.h"
50 #include "llvm/IR/PatternMatch.h"
51 #include "llvm/IR/Type.h"
52 #include "llvm/IR/Value.h"
53 #include "llvm/IR/ValueHandle.h"
54 #include "llvm/Support/Casting.h"
55 #include "llvm/Support/CommandLine.h"
56 #include "llvm/Support/Debug.h"
57 #include "llvm/Support/ErrorHandling.h"
58 #include "llvm/Support/raw_ostream.h"
59 #include <algorithm>
60 #include <cassert>
61 #include <cstdint>
62 #include <iterator>
63 #include <utility>
64 #include <variant>
65 #include <vector>
66 
67 using namespace llvm;
68 using namespace llvm::PatternMatch;
69 
70 #define DEBUG_TYPE "loop-accesses"
71 
72 static cl::opt<unsigned, true>
73 VectorizationFactor("force-vector-width", cl::Hidden,
74                     cl::desc("Sets the SIMD width. Zero is autoselect."),
75                     cl::location(VectorizerParams::VectorizationFactor));
76 unsigned VectorizerParams::VectorizationFactor;
77 
78 static cl::opt<unsigned, true>
79 VectorizationInterleave("force-vector-interleave", cl::Hidden,
80                         cl::desc("Sets the vectorization interleave count. "
81                                  "Zero is autoselect."),
82                         cl::location(
83                             VectorizerParams::VectorizationInterleave));
84 unsigned VectorizerParams::VectorizationInterleave;
85 
86 static cl::opt<unsigned, true> RuntimeMemoryCheckThreshold(
87     "runtime-memory-check-threshold", cl::Hidden,
88     cl::desc("When performing memory disambiguation checks at runtime do not "
89              "generate more than this number of comparisons (default = 8)."),
90     cl::location(VectorizerParams::RuntimeMemoryCheckThreshold), cl::init(8));
91 unsigned VectorizerParams::RuntimeMemoryCheckThreshold;
92 
93 /// The maximum iterations used to merge memory checks
94 static cl::opt<unsigned> MemoryCheckMergeThreshold(
95     "memory-check-merge-threshold", cl::Hidden,
96     cl::desc("Maximum number of comparisons done when trying to merge "
97              "runtime memory checks. (default = 100)"),
98     cl::init(100));
99 
100 /// Maximum SIMD width.
101 const unsigned VectorizerParams::MaxVectorWidth = 64;
102 
103 /// We collect dependences up to this threshold.
104 static cl::opt<unsigned>
105     MaxDependences("max-dependences", cl::Hidden,
106                    cl::desc("Maximum number of dependences collected by "
107                             "loop-access analysis (default = 100)"),
108                    cl::init(100));
109 
110 /// This enables versioning on the strides of symbolically striding memory
111 /// accesses in code like the following.
112 ///   for (i = 0; i < N; ++i)
113 ///     A[i * Stride1] += B[i * Stride2] ...
114 ///
115 /// Will be roughly translated to
116 ///    if (Stride1 == 1 && Stride2 == 1) {
117 ///      for (i = 0; i < N; i+=4)
118 ///       A[i:i+3] += ...
119 ///    } else
120 ///      ...
121 static cl::opt<bool> EnableMemAccessVersioning(
122     "enable-mem-access-versioning", cl::init(true), cl::Hidden,
123     cl::desc("Enable symbolic stride memory access versioning"));
124 
125 /// Enable store-to-load forwarding conflict detection. This option can
126 /// be disabled for correctness testing.
127 static cl::opt<bool> EnableForwardingConflictDetection(
128     "store-to-load-forwarding-conflict-detection", cl::Hidden,
129     cl::desc("Enable conflict detection in loop-access analysis"),
130     cl::init(true));
131 
132 static cl::opt<unsigned> MaxForkedSCEVDepth(
133     "max-forked-scev-depth", cl::Hidden,
134     cl::desc("Maximum recursion depth when finding forked SCEVs (default = 5)"),
135     cl::init(5));
136 
137 static cl::opt<bool> SpeculateUnitStride(
138     "laa-speculate-unit-stride", cl::Hidden,
139     cl::desc("Speculate that non-constant strides are unit in LAA"),
140     cl::init(true));
141 
142 static cl::opt<bool, true> HoistRuntimeChecks(
143     "hoist-runtime-checks", cl::Hidden,
144     cl::desc(
145         "Hoist inner loop runtime memory checks to outer loop if possible"),
146     cl::location(VectorizerParams::HoistRuntimeChecks), cl::init(true));
147 bool VectorizerParams::HoistRuntimeChecks;
148 
149 bool VectorizerParams::isInterleaveForced() {
150   return ::VectorizationInterleave.getNumOccurrences() > 0;
151 }
152 
153 const SCEV *llvm::replaceSymbolicStrideSCEV(PredicatedScalarEvolution &PSE,
154                                             const DenseMap<Value *, const SCEV *> &PtrToStride,
155                                             Value *Ptr) {
156   const SCEV *OrigSCEV = PSE.getSCEV(Ptr);
157 
158   // If there is an entry in the map return the SCEV of the pointer with the
159   // symbolic stride replaced by one.
160   DenseMap<Value *, const SCEV *>::const_iterator SI = PtrToStride.find(Ptr);
161   if (SI == PtrToStride.end())
162     // For a non-symbolic stride, just return the original expression.
163     return OrigSCEV;
164 
165   const SCEV *StrideSCEV = SI->second;
166   // Note: This assert is both overly strong and overly weak.  The actual
167   // invariant here is that StrideSCEV should be loop invariant.  The only
168   // such invariant strides we happen to speculate right now are unknowns
169   // and thus this is a reasonable proxy of the actual invariant.
170   assert(isa<SCEVUnknown>(StrideSCEV) && "shouldn't be in map");
171 
172   ScalarEvolution *SE = PSE.getSE();
173   const auto *CT = SE->getOne(StrideSCEV->getType());
174   PSE.addPredicate(*SE->getEqualPredicate(StrideSCEV, CT));
175   auto *Expr = PSE.getSCEV(Ptr);
176 
177   LLVM_DEBUG(dbgs() << "LAA: Replacing SCEV: " << *OrigSCEV
178 	     << " by: " << *Expr << "\n");
179   return Expr;
180 }
181 
182 RuntimeCheckingPtrGroup::RuntimeCheckingPtrGroup(
183     unsigned Index, RuntimePointerChecking &RtCheck)
184     : High(RtCheck.Pointers[Index].End), Low(RtCheck.Pointers[Index].Start),
185       AddressSpace(RtCheck.Pointers[Index]
186                        .PointerValue->getType()
187                        ->getPointerAddressSpace()),
188       NeedsFreeze(RtCheck.Pointers[Index].NeedsFreeze) {
189   Members.push_back(Index);
190 }
191 
192 /// Calculate Start and End points of memory access.
193 /// Let's assume A is the first access and B is a memory access on N-th loop
194 /// iteration. Then B is calculated as:
195 ///   B = A + Step*N .
196 /// Step value may be positive or negative.
197 /// N is a calculated back-edge taken count:
198 ///     N = (TripCount > 0) ? RoundDown(TripCount -1 , VF) : 0
199 /// Start and End points are calculated in the following way:
200 /// Start = UMIN(A, B) ; End = UMAX(A, B) + SizeOfElt,
201 /// where SizeOfElt is the size of single memory access in bytes.
202 ///
203 /// There is no conflict when the intervals are disjoint:
204 /// NoConflict = (P2.Start >= P1.End) || (P1.Start >= P2.End)
205 void RuntimePointerChecking::insert(Loop *Lp, Value *Ptr, const SCEV *PtrExpr,
206                                     Type *AccessTy, bool WritePtr,
207                                     unsigned DepSetId, unsigned ASId,
208                                     PredicatedScalarEvolution &PSE,
209                                     bool NeedsFreeze) {
210   ScalarEvolution *SE = PSE.getSE();
211 
212   const SCEV *ScStart;
213   const SCEV *ScEnd;
214 
215   if (SE->isLoopInvariant(PtrExpr, Lp)) {
216     ScStart = ScEnd = PtrExpr;
217   } else {
218     const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrExpr);
219     assert(AR && "Invalid addrec expression");
220     const SCEV *Ex = PSE.getBackedgeTakenCount();
221 
222     ScStart = AR->getStart();
223     ScEnd = AR->evaluateAtIteration(Ex, *SE);
224     const SCEV *Step = AR->getStepRecurrence(*SE);
225 
226     // For expressions with negative step, the upper bound is ScStart and the
227     // lower bound is ScEnd.
228     if (const auto *CStep = dyn_cast<SCEVConstant>(Step)) {
229       if (CStep->getValue()->isNegative())
230         std::swap(ScStart, ScEnd);
231     } else {
232       // Fallback case: the step is not constant, but we can still
233       // get the upper and lower bounds of the interval by using min/max
234       // expressions.
235       ScStart = SE->getUMinExpr(ScStart, ScEnd);
236       ScEnd = SE->getUMaxExpr(AR->getStart(), ScEnd);
237     }
238   }
239   assert(SE->isLoopInvariant(ScStart, Lp) && "ScStart needs to be invariant");
240   assert(SE->isLoopInvariant(ScEnd, Lp)&& "ScEnd needs to be invariant");
241 
242   // Add the size of the pointed element to ScEnd.
243   auto &DL = Lp->getHeader()->getModule()->getDataLayout();
244   Type *IdxTy = DL.getIndexType(Ptr->getType());
245   const SCEV *EltSizeSCEV = SE->getStoreSizeOfExpr(IdxTy, AccessTy);
246   ScEnd = SE->getAddExpr(ScEnd, EltSizeSCEV);
247 
248   Pointers.emplace_back(Ptr, ScStart, ScEnd, WritePtr, DepSetId, ASId, PtrExpr,
249                         NeedsFreeze);
250 }
251 
252 void RuntimePointerChecking::tryToCreateDiffCheck(
253     const RuntimeCheckingPtrGroup &CGI, const RuntimeCheckingPtrGroup &CGJ) {
254   if (!CanUseDiffCheck)
255     return;
256 
257   // If either group contains multiple different pointers, bail out.
258   // TODO: Support multiple pointers by using the minimum or maximum pointer,
259   // depending on src & sink.
260   if (CGI.Members.size() != 1 || CGJ.Members.size() != 1) {
261     CanUseDiffCheck = false;
262     return;
263   }
264 
265   PointerInfo *Src = &Pointers[CGI.Members[0]];
266   PointerInfo *Sink = &Pointers[CGJ.Members[0]];
267 
268   // If either pointer is read and written, multiple checks may be needed. Bail
269   // out.
270   if (!DC.getOrderForAccess(Src->PointerValue, !Src->IsWritePtr).empty() ||
271       !DC.getOrderForAccess(Sink->PointerValue, !Sink->IsWritePtr).empty()) {
272     CanUseDiffCheck = false;
273     return;
274   }
275 
276   ArrayRef<unsigned> AccSrc =
277       DC.getOrderForAccess(Src->PointerValue, Src->IsWritePtr);
278   ArrayRef<unsigned> AccSink =
279       DC.getOrderForAccess(Sink->PointerValue, Sink->IsWritePtr);
280   // If either pointer is accessed multiple times, there may not be a clear
281   // src/sink relation. Bail out for now.
282   if (AccSrc.size() != 1 || AccSink.size() != 1) {
283     CanUseDiffCheck = false;
284     return;
285   }
286   // If the sink is accessed before src, swap src/sink.
287   if (AccSink[0] < AccSrc[0])
288     std::swap(Src, Sink);
289 
290   auto *SrcAR = dyn_cast<SCEVAddRecExpr>(Src->Expr);
291   auto *SinkAR = dyn_cast<SCEVAddRecExpr>(Sink->Expr);
292   if (!SrcAR || !SinkAR || SrcAR->getLoop() != DC.getInnermostLoop() ||
293       SinkAR->getLoop() != DC.getInnermostLoop()) {
294     CanUseDiffCheck = false;
295     return;
296   }
297 
298   SmallVector<Instruction *, 4> SrcInsts =
299       DC.getInstructionsForAccess(Src->PointerValue, Src->IsWritePtr);
300   SmallVector<Instruction *, 4> SinkInsts =
301       DC.getInstructionsForAccess(Sink->PointerValue, Sink->IsWritePtr);
302   Type *SrcTy = getLoadStoreType(SrcInsts[0]);
303   Type *DstTy = getLoadStoreType(SinkInsts[0]);
304   if (isa<ScalableVectorType>(SrcTy) || isa<ScalableVectorType>(DstTy)) {
305     CanUseDiffCheck = false;
306     return;
307   }
308   const DataLayout &DL =
309       SinkAR->getLoop()->getHeader()->getModule()->getDataLayout();
310   unsigned AllocSize =
311       std::max(DL.getTypeAllocSize(SrcTy), DL.getTypeAllocSize(DstTy));
312 
313   // Only matching constant steps matching the AllocSize are supported at the
314   // moment. This simplifies the difference computation. Can be extended in the
315   // future.
316   auto *Step = dyn_cast<SCEVConstant>(SinkAR->getStepRecurrence(*SE));
317   if (!Step || Step != SrcAR->getStepRecurrence(*SE) ||
318       Step->getAPInt().abs() != AllocSize) {
319     CanUseDiffCheck = false;
320     return;
321   }
322 
323   IntegerType *IntTy =
324       IntegerType::get(Src->PointerValue->getContext(),
325                        DL.getPointerSizeInBits(CGI.AddressSpace));
326 
327   // When counting down, the dependence distance needs to be swapped.
328   if (Step->getValue()->isNegative())
329     std::swap(SinkAR, SrcAR);
330 
331   const SCEV *SinkStartInt = SE->getPtrToIntExpr(SinkAR->getStart(), IntTy);
332   const SCEV *SrcStartInt = SE->getPtrToIntExpr(SrcAR->getStart(), IntTy);
333   if (isa<SCEVCouldNotCompute>(SinkStartInt) ||
334       isa<SCEVCouldNotCompute>(SrcStartInt)) {
335     CanUseDiffCheck = false;
336     return;
337   }
338 
339   const Loop *InnerLoop = SrcAR->getLoop();
340   // If the start values for both Src and Sink also vary according to an outer
341   // loop, then it's probably better to avoid creating diff checks because
342   // they may not be hoisted. We should instead let llvm::addRuntimeChecks
343   // do the expanded full range overlap checks, which can be hoisted.
344   if (HoistRuntimeChecks && InnerLoop->getParentLoop() &&
345       isa<SCEVAddRecExpr>(SinkStartInt) && isa<SCEVAddRecExpr>(SrcStartInt)) {
346     auto *SrcStartAR = cast<SCEVAddRecExpr>(SrcStartInt);
347     auto *SinkStartAR = cast<SCEVAddRecExpr>(SinkStartInt);
348     const Loop *StartARLoop = SrcStartAR->getLoop();
349     if (StartARLoop == SinkStartAR->getLoop() &&
350         StartARLoop == InnerLoop->getParentLoop() &&
351         // If the diff check would already be loop invariant (due to the
352         // recurrences being the same), then we prefer to keep the diff checks
353         // because they are cheaper.
354         SrcStartAR->getStepRecurrence(*SE) !=
355             SinkStartAR->getStepRecurrence(*SE)) {
356       LLVM_DEBUG(dbgs() << "LAA: Not creating diff runtime check, since these "
357                            "cannot be hoisted out of the outer loop\n");
358       CanUseDiffCheck = false;
359       return;
360     }
361   }
362 
363   LLVM_DEBUG(dbgs() << "LAA: Creating diff runtime check for:\n"
364                     << "SrcStart: " << *SrcStartInt << '\n'
365                     << "SinkStartInt: " << *SinkStartInt << '\n');
366   DiffChecks.emplace_back(SrcStartInt, SinkStartInt, AllocSize,
367                           Src->NeedsFreeze || Sink->NeedsFreeze);
368 }
369 
370 SmallVector<RuntimePointerCheck, 4> RuntimePointerChecking::generateChecks() {
371   SmallVector<RuntimePointerCheck, 4> Checks;
372 
373   for (unsigned I = 0; I < CheckingGroups.size(); ++I) {
374     for (unsigned J = I + 1; J < CheckingGroups.size(); ++J) {
375       const RuntimeCheckingPtrGroup &CGI = CheckingGroups[I];
376       const RuntimeCheckingPtrGroup &CGJ = CheckingGroups[J];
377 
378       if (needsChecking(CGI, CGJ)) {
379         tryToCreateDiffCheck(CGI, CGJ);
380         Checks.push_back(std::make_pair(&CGI, &CGJ));
381       }
382     }
383   }
384   return Checks;
385 }
386 
387 void RuntimePointerChecking::generateChecks(
388     MemoryDepChecker::DepCandidates &DepCands, bool UseDependencies) {
389   assert(Checks.empty() && "Checks is not empty");
390   groupChecks(DepCands, UseDependencies);
391   Checks = generateChecks();
392 }
393 
394 bool RuntimePointerChecking::needsChecking(
395     const RuntimeCheckingPtrGroup &M, const RuntimeCheckingPtrGroup &N) const {
396   for (unsigned I = 0, EI = M.Members.size(); EI != I; ++I)
397     for (unsigned J = 0, EJ = N.Members.size(); EJ != J; ++J)
398       if (needsChecking(M.Members[I], N.Members[J]))
399         return true;
400   return false;
401 }
402 
403 /// Compare \p I and \p J and return the minimum.
404 /// Return nullptr in case we couldn't find an answer.
405 static const SCEV *getMinFromExprs(const SCEV *I, const SCEV *J,
406                                    ScalarEvolution *SE) {
407   const SCEV *Diff = SE->getMinusSCEV(J, I);
408   const SCEVConstant *C = dyn_cast<const SCEVConstant>(Diff);
409 
410   if (!C)
411     return nullptr;
412   if (C->getValue()->isNegative())
413     return J;
414   return I;
415 }
416 
417 bool RuntimeCheckingPtrGroup::addPointer(unsigned Index,
418                                          RuntimePointerChecking &RtCheck) {
419   return addPointer(
420       Index, RtCheck.Pointers[Index].Start, RtCheck.Pointers[Index].End,
421       RtCheck.Pointers[Index].PointerValue->getType()->getPointerAddressSpace(),
422       RtCheck.Pointers[Index].NeedsFreeze, *RtCheck.SE);
423 }
424 
425 bool RuntimeCheckingPtrGroup::addPointer(unsigned Index, const SCEV *Start,
426                                          const SCEV *End, unsigned AS,
427                                          bool NeedsFreeze,
428                                          ScalarEvolution &SE) {
429   assert(AddressSpace == AS &&
430          "all pointers in a checking group must be in the same address space");
431 
432   // Compare the starts and ends with the known minimum and maximum
433   // of this set. We need to know how we compare against the min/max
434   // of the set in order to be able to emit memchecks.
435   const SCEV *Min0 = getMinFromExprs(Start, Low, &SE);
436   if (!Min0)
437     return false;
438 
439   const SCEV *Min1 = getMinFromExprs(End, High, &SE);
440   if (!Min1)
441     return false;
442 
443   // Update the low bound  expression if we've found a new min value.
444   if (Min0 == Start)
445     Low = Start;
446 
447   // Update the high bound expression if we've found a new max value.
448   if (Min1 != End)
449     High = End;
450 
451   Members.push_back(Index);
452   this->NeedsFreeze |= NeedsFreeze;
453   return true;
454 }
455 
456 void RuntimePointerChecking::groupChecks(
457     MemoryDepChecker::DepCandidates &DepCands, bool UseDependencies) {
458   // We build the groups from dependency candidates equivalence classes
459   // because:
460   //    - We know that pointers in the same equivalence class share
461   //      the same underlying object and therefore there is a chance
462   //      that we can compare pointers
463   //    - We wouldn't be able to merge two pointers for which we need
464   //      to emit a memcheck. The classes in DepCands are already
465   //      conveniently built such that no two pointers in the same
466   //      class need checking against each other.
467 
468   // We use the following (greedy) algorithm to construct the groups
469   // For every pointer in the equivalence class:
470   //   For each existing group:
471   //   - if the difference between this pointer and the min/max bounds
472   //     of the group is a constant, then make the pointer part of the
473   //     group and update the min/max bounds of that group as required.
474 
475   CheckingGroups.clear();
476 
477   // If we need to check two pointers to the same underlying object
478   // with a non-constant difference, we shouldn't perform any pointer
479   // grouping with those pointers. This is because we can easily get
480   // into cases where the resulting check would return false, even when
481   // the accesses are safe.
482   //
483   // The following example shows this:
484   // for (i = 0; i < 1000; ++i)
485   //   a[5000 + i * m] = a[i] + a[i + 9000]
486   //
487   // Here grouping gives a check of (5000, 5000 + 1000 * m) against
488   // (0, 10000) which is always false. However, if m is 1, there is no
489   // dependence. Not grouping the checks for a[i] and a[i + 9000] allows
490   // us to perform an accurate check in this case.
491   //
492   // The above case requires that we have an UnknownDependence between
493   // accesses to the same underlying object. This cannot happen unless
494   // FoundNonConstantDistanceDependence is set, and therefore UseDependencies
495   // is also false. In this case we will use the fallback path and create
496   // separate checking groups for all pointers.
497 
498   // If we don't have the dependency partitions, construct a new
499   // checking pointer group for each pointer. This is also required
500   // for correctness, because in this case we can have checking between
501   // pointers to the same underlying object.
502   if (!UseDependencies) {
503     for (unsigned I = 0; I < Pointers.size(); ++I)
504       CheckingGroups.push_back(RuntimeCheckingPtrGroup(I, *this));
505     return;
506   }
507 
508   unsigned TotalComparisons = 0;
509 
510   DenseMap<Value *, SmallVector<unsigned>> PositionMap;
511   for (unsigned Index = 0; Index < Pointers.size(); ++Index) {
512     auto Iter = PositionMap.insert({Pointers[Index].PointerValue, {}});
513     Iter.first->second.push_back(Index);
514   }
515 
516   // We need to keep track of what pointers we've already seen so we
517   // don't process them twice.
518   SmallSet<unsigned, 2> Seen;
519 
520   // Go through all equivalence classes, get the "pointer check groups"
521   // and add them to the overall solution. We use the order in which accesses
522   // appear in 'Pointers' to enforce determinism.
523   for (unsigned I = 0; I < Pointers.size(); ++I) {
524     // We've seen this pointer before, and therefore already processed
525     // its equivalence class.
526     if (Seen.count(I))
527       continue;
528 
529     MemoryDepChecker::MemAccessInfo Access(Pointers[I].PointerValue,
530                                            Pointers[I].IsWritePtr);
531 
532     SmallVector<RuntimeCheckingPtrGroup, 2> Groups;
533     auto LeaderI = DepCands.findValue(DepCands.getLeaderValue(Access));
534 
535     // Because DepCands is constructed by visiting accesses in the order in
536     // which they appear in alias sets (which is deterministic) and the
537     // iteration order within an equivalence class member is only dependent on
538     // the order in which unions and insertions are performed on the
539     // equivalence class, the iteration order is deterministic.
540     for (auto MI = DepCands.member_begin(LeaderI), ME = DepCands.member_end();
541          MI != ME; ++MI) {
542       auto PointerI = PositionMap.find(MI->getPointer());
543       assert(PointerI != PositionMap.end() &&
544              "pointer in equivalence class not found in PositionMap");
545       for (unsigned Pointer : PointerI->second) {
546         bool Merged = false;
547         // Mark this pointer as seen.
548         Seen.insert(Pointer);
549 
550         // Go through all the existing sets and see if we can find one
551         // which can include this pointer.
552         for (RuntimeCheckingPtrGroup &Group : Groups) {
553           // Don't perform more than a certain amount of comparisons.
554           // This should limit the cost of grouping the pointers to something
555           // reasonable.  If we do end up hitting this threshold, the algorithm
556           // will create separate groups for all remaining pointers.
557           if (TotalComparisons > MemoryCheckMergeThreshold)
558             break;
559 
560           TotalComparisons++;
561 
562           if (Group.addPointer(Pointer, *this)) {
563             Merged = true;
564             break;
565           }
566         }
567 
568         if (!Merged)
569           // We couldn't add this pointer to any existing set or the threshold
570           // for the number of comparisons has been reached. Create a new group
571           // to hold the current pointer.
572           Groups.push_back(RuntimeCheckingPtrGroup(Pointer, *this));
573       }
574     }
575 
576     // We've computed the grouped checks for this partition.
577     // Save the results and continue with the next one.
578     llvm::copy(Groups, std::back_inserter(CheckingGroups));
579   }
580 }
581 
582 bool RuntimePointerChecking::arePointersInSamePartition(
583     const SmallVectorImpl<int> &PtrToPartition, unsigned PtrIdx1,
584     unsigned PtrIdx2) {
585   return (PtrToPartition[PtrIdx1] != -1 &&
586           PtrToPartition[PtrIdx1] == PtrToPartition[PtrIdx2]);
587 }
588 
589 bool RuntimePointerChecking::needsChecking(unsigned I, unsigned J) const {
590   const PointerInfo &PointerI = Pointers[I];
591   const PointerInfo &PointerJ = Pointers[J];
592 
593   // No need to check if two readonly pointers intersect.
594   if (!PointerI.IsWritePtr && !PointerJ.IsWritePtr)
595     return false;
596 
597   // Only need to check pointers between two different dependency sets.
598   if (PointerI.DependencySetId == PointerJ.DependencySetId)
599     return false;
600 
601   // Only need to check pointers in the same alias set.
602   if (PointerI.AliasSetId != PointerJ.AliasSetId)
603     return false;
604 
605   return true;
606 }
607 
608 void RuntimePointerChecking::printChecks(
609     raw_ostream &OS, const SmallVectorImpl<RuntimePointerCheck> &Checks,
610     unsigned Depth) const {
611   unsigned N = 0;
612   for (const auto &Check : Checks) {
613     const auto &First = Check.first->Members, &Second = Check.second->Members;
614 
615     OS.indent(Depth) << "Check " << N++ << ":\n";
616 
617     OS.indent(Depth + 2) << "Comparing group (" << Check.first << "):\n";
618     for (unsigned K = 0; K < First.size(); ++K)
619       OS.indent(Depth + 2) << *Pointers[First[K]].PointerValue << "\n";
620 
621     OS.indent(Depth + 2) << "Against group (" << Check.second << "):\n";
622     for (unsigned K = 0; K < Second.size(); ++K)
623       OS.indent(Depth + 2) << *Pointers[Second[K]].PointerValue << "\n";
624   }
625 }
626 
627 void RuntimePointerChecking::print(raw_ostream &OS, unsigned Depth) const {
628 
629   OS.indent(Depth) << "Run-time memory checks:\n";
630   printChecks(OS, Checks, Depth);
631 
632   OS.indent(Depth) << "Grouped accesses:\n";
633   for (unsigned I = 0; I < CheckingGroups.size(); ++I) {
634     const auto &CG = CheckingGroups[I];
635 
636     OS.indent(Depth + 2) << "Group " << &CG << ":\n";
637     OS.indent(Depth + 4) << "(Low: " << *CG.Low << " High: " << *CG.High
638                          << ")\n";
639     for (unsigned J = 0; J < CG.Members.size(); ++J) {
640       OS.indent(Depth + 6) << "Member: " << *Pointers[CG.Members[J]].Expr
641                            << "\n";
642     }
643   }
644 }
645 
646 namespace {
647 
648 /// Analyses memory accesses in a loop.
649 ///
650 /// Checks whether run time pointer checks are needed and builds sets for data
651 /// dependence checking.
652 class AccessAnalysis {
653 public:
654   /// Read or write access location.
655   typedef PointerIntPair<Value *, 1, bool> MemAccessInfo;
656   typedef SmallVector<MemAccessInfo, 8> MemAccessInfoList;
657 
658   AccessAnalysis(Loop *TheLoop, AAResults *AA, LoopInfo *LI,
659                  MemoryDepChecker::DepCandidates &DA,
660                  PredicatedScalarEvolution &PSE,
661                  SmallPtrSetImpl<MDNode *> &LoopAliasScopes)
662       : TheLoop(TheLoop), BAA(*AA), AST(BAA), LI(LI), DepCands(DA), PSE(PSE),
663         LoopAliasScopes(LoopAliasScopes) {
664     // We're analyzing dependences across loop iterations.
665     BAA.enableCrossIterationMode();
666   }
667 
668   /// Register a load  and whether it is only read from.
669   void addLoad(MemoryLocation &Loc, Type *AccessTy, bool IsReadOnly) {
670     Value *Ptr = const_cast<Value *>(Loc.Ptr);
671     AST.add(adjustLoc(Loc));
672     Accesses[MemAccessInfo(Ptr, false)].insert(AccessTy);
673     if (IsReadOnly)
674       ReadOnlyPtr.insert(Ptr);
675   }
676 
677   /// Register a store.
678   void addStore(MemoryLocation &Loc, Type *AccessTy) {
679     Value *Ptr = const_cast<Value *>(Loc.Ptr);
680     AST.add(adjustLoc(Loc));
681     Accesses[MemAccessInfo(Ptr, true)].insert(AccessTy);
682   }
683 
684   /// Check if we can emit a run-time no-alias check for \p Access.
685   ///
686   /// Returns true if we can emit a run-time no alias check for \p Access.
687   /// If we can check this access, this also adds it to a dependence set and
688   /// adds a run-time to check for it to \p RtCheck. If \p Assume is true,
689   /// we will attempt to use additional run-time checks in order to get
690   /// the bounds of the pointer.
691   bool createCheckForAccess(RuntimePointerChecking &RtCheck,
692                             MemAccessInfo Access, Type *AccessTy,
693                             const DenseMap<Value *, const SCEV *> &Strides,
694                             DenseMap<Value *, unsigned> &DepSetId,
695                             Loop *TheLoop, unsigned &RunningDepId,
696                             unsigned ASId, bool ShouldCheckStride, bool Assume);
697 
698   /// Check whether we can check the pointers at runtime for
699   /// non-intersection.
700   ///
701   /// Returns true if we need no check or if we do and we can generate them
702   /// (i.e. the pointers have computable bounds).
703   bool canCheckPtrAtRT(RuntimePointerChecking &RtCheck, ScalarEvolution *SE,
704                        Loop *TheLoop, const DenseMap<Value *, const SCEV *> &Strides,
705                        Value *&UncomputablePtr, bool ShouldCheckWrap = false);
706 
707   /// Goes over all memory accesses, checks whether a RT check is needed
708   /// and builds sets of dependent accesses.
709   void buildDependenceSets() {
710     processMemAccesses();
711   }
712 
713   /// Initial processing of memory accesses determined that we need to
714   /// perform dependency checking.
715   ///
716   /// Note that this can later be cleared if we retry memcheck analysis without
717   /// dependency checking (i.e. FoundNonConstantDistanceDependence).
718   bool isDependencyCheckNeeded() { return !CheckDeps.empty(); }
719 
720   /// We decided that no dependence analysis would be used.  Reset the state.
721   void resetDepChecks(MemoryDepChecker &DepChecker) {
722     CheckDeps.clear();
723     DepChecker.clearDependences();
724   }
725 
726   MemAccessInfoList &getDependenciesToCheck() { return CheckDeps; }
727 
728   const DenseMap<Value *, SmallVector<const Value *, 16>> &
729   getUnderlyingObjects() {
730     return UnderlyingObjects;
731   }
732 
733 private:
734   typedef MapVector<MemAccessInfo, SmallSetVector<Type *, 1>> PtrAccessMap;
735 
736   /// Adjust the MemoryLocation so that it represents accesses to this
737   /// location across all iterations, rather than a single one.
738   MemoryLocation adjustLoc(MemoryLocation Loc) const {
739     // The accessed location varies within the loop, but remains within the
740     // underlying object.
741     Loc.Size = LocationSize::beforeOrAfterPointer();
742     Loc.AATags.Scope = adjustAliasScopeList(Loc.AATags.Scope);
743     Loc.AATags.NoAlias = adjustAliasScopeList(Loc.AATags.NoAlias);
744     return Loc;
745   }
746 
747   /// Drop alias scopes that are only valid within a single loop iteration.
748   MDNode *adjustAliasScopeList(MDNode *ScopeList) const {
749     if (!ScopeList)
750       return nullptr;
751 
752     // For the sake of simplicity, drop the whole scope list if any scope is
753     // iteration-local.
754     if (any_of(ScopeList->operands(), [&](Metadata *Scope) {
755           return LoopAliasScopes.contains(cast<MDNode>(Scope));
756         }))
757       return nullptr;
758 
759     return ScopeList;
760   }
761 
762   /// Go over all memory access and check whether runtime pointer checks
763   /// are needed and build sets of dependency check candidates.
764   void processMemAccesses();
765 
766   /// Map of all accesses. Values are the types used to access memory pointed to
767   /// by the pointer.
768   PtrAccessMap Accesses;
769 
770   /// The loop being checked.
771   const Loop *TheLoop;
772 
773   /// List of accesses that need a further dependence check.
774   MemAccessInfoList CheckDeps;
775 
776   /// Set of pointers that are read only.
777   SmallPtrSet<Value*, 16> ReadOnlyPtr;
778 
779   /// Batched alias analysis results.
780   BatchAAResults BAA;
781 
782   /// An alias set tracker to partition the access set by underlying object and
783   //intrinsic property (such as TBAA metadata).
784   AliasSetTracker AST;
785 
786   LoopInfo *LI;
787 
788   /// Sets of potentially dependent accesses - members of one set share an
789   /// underlying pointer. The set "CheckDeps" identfies which sets really need a
790   /// dependence check.
791   MemoryDepChecker::DepCandidates &DepCands;
792 
793   /// Initial processing of memory accesses determined that we may need
794   /// to add memchecks.  Perform the analysis to determine the necessary checks.
795   ///
796   /// Note that, this is different from isDependencyCheckNeeded.  When we retry
797   /// memcheck analysis without dependency checking
798   /// (i.e. FoundNonConstantDistanceDependence), isDependencyCheckNeeded is
799   /// cleared while this remains set if we have potentially dependent accesses.
800   bool IsRTCheckAnalysisNeeded = false;
801 
802   /// The SCEV predicate containing all the SCEV-related assumptions.
803   PredicatedScalarEvolution &PSE;
804 
805   DenseMap<Value *, SmallVector<const Value *, 16>> UnderlyingObjects;
806 
807   /// Alias scopes that are declared inside the loop, and as such not valid
808   /// across iterations.
809   SmallPtrSetImpl<MDNode *> &LoopAliasScopes;
810 };
811 
812 } // end anonymous namespace
813 
814 /// Check whether a pointer can participate in a runtime bounds check.
815 /// If \p Assume, try harder to prove that we can compute the bounds of \p Ptr
816 /// by adding run-time checks (overflow checks) if necessary.
817 static bool hasComputableBounds(PredicatedScalarEvolution &PSE, Value *Ptr,
818                                 const SCEV *PtrScev, Loop *L, bool Assume) {
819   // The bounds for loop-invariant pointer is trivial.
820   if (PSE.getSE()->isLoopInvariant(PtrScev, L))
821     return true;
822 
823   const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
824 
825   if (!AR && Assume)
826     AR = PSE.getAsAddRec(Ptr);
827 
828   if (!AR)
829     return false;
830 
831   return AR->isAffine();
832 }
833 
834 /// Check whether a pointer address cannot wrap.
835 static bool isNoWrap(PredicatedScalarEvolution &PSE,
836                      const DenseMap<Value *, const SCEV *> &Strides, Value *Ptr, Type *AccessTy,
837                      Loop *L) {
838   const SCEV *PtrScev = PSE.getSCEV(Ptr);
839   if (PSE.getSE()->isLoopInvariant(PtrScev, L))
840     return true;
841 
842   int64_t Stride = getPtrStride(PSE, AccessTy, Ptr, L, Strides).value_or(0);
843   if (Stride == 1 || PSE.hasNoOverflow(Ptr, SCEVWrapPredicate::IncrementNUSW))
844     return true;
845 
846   return false;
847 }
848 
849 static void visitPointers(Value *StartPtr, const Loop &InnermostLoop,
850                           function_ref<void(Value *)> AddPointer) {
851   SmallPtrSet<Value *, 8> Visited;
852   SmallVector<Value *> WorkList;
853   WorkList.push_back(StartPtr);
854 
855   while (!WorkList.empty()) {
856     Value *Ptr = WorkList.pop_back_val();
857     if (!Visited.insert(Ptr).second)
858       continue;
859     auto *PN = dyn_cast<PHINode>(Ptr);
860     // SCEV does not look through non-header PHIs inside the loop. Such phis
861     // can be analyzed by adding separate accesses for each incoming pointer
862     // value.
863     if (PN && InnermostLoop.contains(PN->getParent()) &&
864         PN->getParent() != InnermostLoop.getHeader()) {
865       for (const Use &Inc : PN->incoming_values())
866         WorkList.push_back(Inc);
867     } else
868       AddPointer(Ptr);
869   }
870 }
871 
872 // Walk back through the IR for a pointer, looking for a select like the
873 // following:
874 //
875 //  %offset = select i1 %cmp, i64 %a, i64 %b
876 //  %addr = getelementptr double, double* %base, i64 %offset
877 //  %ld = load double, double* %addr, align 8
878 //
879 // We won't be able to form a single SCEVAddRecExpr from this since the
880 // address for each loop iteration depends on %cmp. We could potentially
881 // produce multiple valid SCEVAddRecExprs, though, and check all of them for
882 // memory safety/aliasing if needed.
883 //
884 // If we encounter some IR we don't yet handle, or something obviously fine
885 // like a constant, then we just add the SCEV for that term to the list passed
886 // in by the caller. If we have a node that may potentially yield a valid
887 // SCEVAddRecExpr then we decompose it into parts and build the SCEV terms
888 // ourselves before adding to the list.
889 static void findForkedSCEVs(
890     ScalarEvolution *SE, const Loop *L, Value *Ptr,
891     SmallVectorImpl<PointerIntPair<const SCEV *, 1, bool>> &ScevList,
892     unsigned Depth) {
893   // If our Value is a SCEVAddRecExpr, loop invariant, not an instruction, or
894   // we've exceeded our limit on recursion, just return whatever we have
895   // regardless of whether it can be used for a forked pointer or not, along
896   // with an indication of whether it might be a poison or undef value.
897   const SCEV *Scev = SE->getSCEV(Ptr);
898   if (isa<SCEVAddRecExpr>(Scev) || L->isLoopInvariant(Ptr) ||
899       !isa<Instruction>(Ptr) || Depth == 0) {
900     ScevList.emplace_back(Scev, !isGuaranteedNotToBeUndefOrPoison(Ptr));
901     return;
902   }
903 
904   Depth--;
905 
906   auto UndefPoisonCheck = [](PointerIntPair<const SCEV *, 1, bool> S) {
907     return get<1>(S);
908   };
909 
910   auto GetBinOpExpr = [&SE](unsigned Opcode, const SCEV *L, const SCEV *R) {
911     switch (Opcode) {
912     case Instruction::Add:
913       return SE->getAddExpr(L, R);
914     case Instruction::Sub:
915       return SE->getMinusSCEV(L, R);
916     default:
917       llvm_unreachable("Unexpected binary operator when walking ForkedPtrs");
918     }
919   };
920 
921   Instruction *I = cast<Instruction>(Ptr);
922   unsigned Opcode = I->getOpcode();
923   switch (Opcode) {
924   case Instruction::GetElementPtr: {
925     GetElementPtrInst *GEP = cast<GetElementPtrInst>(I);
926     Type *SourceTy = GEP->getSourceElementType();
927     // We only handle base + single offset GEPs here for now.
928     // Not dealing with preexisting gathers yet, so no vectors.
929     if (I->getNumOperands() != 2 || SourceTy->isVectorTy()) {
930       ScevList.emplace_back(Scev, !isGuaranteedNotToBeUndefOrPoison(GEP));
931       break;
932     }
933     SmallVector<PointerIntPair<const SCEV *, 1, bool>, 2> BaseScevs;
934     SmallVector<PointerIntPair<const SCEV *, 1, bool>, 2> OffsetScevs;
935     findForkedSCEVs(SE, L, I->getOperand(0), BaseScevs, Depth);
936     findForkedSCEVs(SE, L, I->getOperand(1), OffsetScevs, Depth);
937 
938     // See if we need to freeze our fork...
939     bool NeedsFreeze = any_of(BaseScevs, UndefPoisonCheck) ||
940                        any_of(OffsetScevs, UndefPoisonCheck);
941 
942     // Check that we only have a single fork, on either the base or the offset.
943     // Copy the SCEV across for the one without a fork in order to generate
944     // the full SCEV for both sides of the GEP.
945     if (OffsetScevs.size() == 2 && BaseScevs.size() == 1)
946       BaseScevs.push_back(BaseScevs[0]);
947     else if (BaseScevs.size() == 2 && OffsetScevs.size() == 1)
948       OffsetScevs.push_back(OffsetScevs[0]);
949     else {
950       ScevList.emplace_back(Scev, NeedsFreeze);
951       break;
952     }
953 
954     // Find the pointer type we need to extend to.
955     Type *IntPtrTy = SE->getEffectiveSCEVType(
956         SE->getSCEV(GEP->getPointerOperand())->getType());
957 
958     // Find the size of the type being pointed to. We only have a single
959     // index term (guarded above) so we don't need to index into arrays or
960     // structures, just get the size of the scalar value.
961     const SCEV *Size = SE->getSizeOfExpr(IntPtrTy, SourceTy);
962 
963     // Scale up the offsets by the size of the type, then add to the bases.
964     const SCEV *Scaled1 = SE->getMulExpr(
965         Size, SE->getTruncateOrSignExtend(get<0>(OffsetScevs[0]), IntPtrTy));
966     const SCEV *Scaled2 = SE->getMulExpr(
967         Size, SE->getTruncateOrSignExtend(get<0>(OffsetScevs[1]), IntPtrTy));
968     ScevList.emplace_back(SE->getAddExpr(get<0>(BaseScevs[0]), Scaled1),
969                           NeedsFreeze);
970     ScevList.emplace_back(SE->getAddExpr(get<0>(BaseScevs[1]), Scaled2),
971                           NeedsFreeze);
972     break;
973   }
974   case Instruction::Select: {
975     SmallVector<PointerIntPair<const SCEV *, 1, bool>, 2> ChildScevs;
976     // A select means we've found a forked pointer, but we currently only
977     // support a single select per pointer so if there's another behind this
978     // then we just bail out and return the generic SCEV.
979     findForkedSCEVs(SE, L, I->getOperand(1), ChildScevs, Depth);
980     findForkedSCEVs(SE, L, I->getOperand(2), ChildScevs, Depth);
981     if (ChildScevs.size() == 2) {
982       ScevList.push_back(ChildScevs[0]);
983       ScevList.push_back(ChildScevs[1]);
984     } else
985       ScevList.emplace_back(Scev, !isGuaranteedNotToBeUndefOrPoison(Ptr));
986     break;
987   }
988   case Instruction::PHI: {
989     SmallVector<PointerIntPair<const SCEV *, 1, bool>, 2> ChildScevs;
990     // A phi means we've found a forked pointer, but we currently only
991     // support a single phi per pointer so if there's another behind this
992     // then we just bail out and return the generic SCEV.
993     if (I->getNumOperands() == 2) {
994       findForkedSCEVs(SE, L, I->getOperand(0), ChildScevs, Depth);
995       findForkedSCEVs(SE, L, I->getOperand(1), ChildScevs, Depth);
996     }
997     if (ChildScevs.size() == 2) {
998       ScevList.push_back(ChildScevs[0]);
999       ScevList.push_back(ChildScevs[1]);
1000     } else
1001       ScevList.emplace_back(Scev, !isGuaranteedNotToBeUndefOrPoison(Ptr));
1002     break;
1003   }
1004   case Instruction::Add:
1005   case Instruction::Sub: {
1006     SmallVector<PointerIntPair<const SCEV *, 1, bool>> LScevs;
1007     SmallVector<PointerIntPair<const SCEV *, 1, bool>> RScevs;
1008     findForkedSCEVs(SE, L, I->getOperand(0), LScevs, Depth);
1009     findForkedSCEVs(SE, L, I->getOperand(1), RScevs, Depth);
1010 
1011     // See if we need to freeze our fork...
1012     bool NeedsFreeze =
1013         any_of(LScevs, UndefPoisonCheck) || any_of(RScevs, UndefPoisonCheck);
1014 
1015     // Check that we only have a single fork, on either the left or right side.
1016     // Copy the SCEV across for the one without a fork in order to generate
1017     // the full SCEV for both sides of the BinOp.
1018     if (LScevs.size() == 2 && RScevs.size() == 1)
1019       RScevs.push_back(RScevs[0]);
1020     else if (RScevs.size() == 2 && LScevs.size() == 1)
1021       LScevs.push_back(LScevs[0]);
1022     else {
1023       ScevList.emplace_back(Scev, NeedsFreeze);
1024       break;
1025     }
1026 
1027     ScevList.emplace_back(
1028         GetBinOpExpr(Opcode, get<0>(LScevs[0]), get<0>(RScevs[0])),
1029         NeedsFreeze);
1030     ScevList.emplace_back(
1031         GetBinOpExpr(Opcode, get<0>(LScevs[1]), get<0>(RScevs[1])),
1032         NeedsFreeze);
1033     break;
1034   }
1035   default:
1036     // Just return the current SCEV if we haven't handled the instruction yet.
1037     LLVM_DEBUG(dbgs() << "ForkedPtr unhandled instruction: " << *I << "\n");
1038     ScevList.emplace_back(Scev, !isGuaranteedNotToBeUndefOrPoison(Ptr));
1039     break;
1040   }
1041 }
1042 
1043 static SmallVector<PointerIntPair<const SCEV *, 1, bool>>
1044 findForkedPointer(PredicatedScalarEvolution &PSE,
1045                   const DenseMap<Value *, const SCEV *> &StridesMap, Value *Ptr,
1046                   const Loop *L) {
1047   ScalarEvolution *SE = PSE.getSE();
1048   assert(SE->isSCEVable(Ptr->getType()) && "Value is not SCEVable!");
1049   SmallVector<PointerIntPair<const SCEV *, 1, bool>> Scevs;
1050   findForkedSCEVs(SE, L, Ptr, Scevs, MaxForkedSCEVDepth);
1051 
1052   // For now, we will only accept a forked pointer with two possible SCEVs
1053   // that are either SCEVAddRecExprs or loop invariant.
1054   if (Scevs.size() == 2 &&
1055       (isa<SCEVAddRecExpr>(get<0>(Scevs[0])) ||
1056        SE->isLoopInvariant(get<0>(Scevs[0]), L)) &&
1057       (isa<SCEVAddRecExpr>(get<0>(Scevs[1])) ||
1058        SE->isLoopInvariant(get<0>(Scevs[1]), L))) {
1059     LLVM_DEBUG(dbgs() << "LAA: Found forked pointer: " << *Ptr << "\n");
1060     LLVM_DEBUG(dbgs() << "\t(1) " << *get<0>(Scevs[0]) << "\n");
1061     LLVM_DEBUG(dbgs() << "\t(2) " << *get<0>(Scevs[1]) << "\n");
1062     return Scevs;
1063   }
1064 
1065   return {{replaceSymbolicStrideSCEV(PSE, StridesMap, Ptr), false}};
1066 }
1067 
1068 bool AccessAnalysis::createCheckForAccess(RuntimePointerChecking &RtCheck,
1069                                           MemAccessInfo Access, Type *AccessTy,
1070                                           const DenseMap<Value *, const SCEV *> &StridesMap,
1071                                           DenseMap<Value *, unsigned> &DepSetId,
1072                                           Loop *TheLoop, unsigned &RunningDepId,
1073                                           unsigned ASId, bool ShouldCheckWrap,
1074                                           bool Assume) {
1075   Value *Ptr = Access.getPointer();
1076 
1077   SmallVector<PointerIntPair<const SCEV *, 1, bool>> TranslatedPtrs =
1078       findForkedPointer(PSE, StridesMap, Ptr, TheLoop);
1079 
1080   for (auto &P : TranslatedPtrs) {
1081     const SCEV *PtrExpr = get<0>(P);
1082     if (!hasComputableBounds(PSE, Ptr, PtrExpr, TheLoop, Assume))
1083       return false;
1084 
1085     // When we run after a failing dependency check we have to make sure
1086     // we don't have wrapping pointers.
1087     if (ShouldCheckWrap) {
1088       // Skip wrap checking when translating pointers.
1089       if (TranslatedPtrs.size() > 1)
1090         return false;
1091 
1092       if (!isNoWrap(PSE, StridesMap, Ptr, AccessTy, TheLoop)) {
1093         auto *Expr = PSE.getSCEV(Ptr);
1094         if (!Assume || !isa<SCEVAddRecExpr>(Expr))
1095           return false;
1096         PSE.setNoOverflow(Ptr, SCEVWrapPredicate::IncrementNUSW);
1097       }
1098     }
1099     // If there's only one option for Ptr, look it up after bounds and wrap
1100     // checking, because assumptions might have been added to PSE.
1101     if (TranslatedPtrs.size() == 1)
1102       TranslatedPtrs[0] = {replaceSymbolicStrideSCEV(PSE, StridesMap, Ptr),
1103                            false};
1104   }
1105 
1106   for (auto [PtrExpr, NeedsFreeze] : TranslatedPtrs) {
1107     // The id of the dependence set.
1108     unsigned DepId;
1109 
1110     if (isDependencyCheckNeeded()) {
1111       Value *Leader = DepCands.getLeaderValue(Access).getPointer();
1112       unsigned &LeaderId = DepSetId[Leader];
1113       if (!LeaderId)
1114         LeaderId = RunningDepId++;
1115       DepId = LeaderId;
1116     } else
1117       // Each access has its own dependence set.
1118       DepId = RunningDepId++;
1119 
1120     bool IsWrite = Access.getInt();
1121     RtCheck.insert(TheLoop, Ptr, PtrExpr, AccessTy, IsWrite, DepId, ASId, PSE,
1122                    NeedsFreeze);
1123     LLVM_DEBUG(dbgs() << "LAA: Found a runtime check ptr:" << *Ptr << '\n');
1124   }
1125 
1126   return true;
1127 }
1128 
1129 bool AccessAnalysis::canCheckPtrAtRT(RuntimePointerChecking &RtCheck,
1130                                      ScalarEvolution *SE, Loop *TheLoop,
1131                                      const DenseMap<Value *, const SCEV *> &StridesMap,
1132                                      Value *&UncomputablePtr, bool ShouldCheckWrap) {
1133   // Find pointers with computable bounds. We are going to use this information
1134   // to place a runtime bound check.
1135   bool CanDoRT = true;
1136 
1137   bool MayNeedRTCheck = false;
1138   if (!IsRTCheckAnalysisNeeded) return true;
1139 
1140   bool IsDepCheckNeeded = isDependencyCheckNeeded();
1141 
1142   // We assign a consecutive id to access from different alias sets.
1143   // Accesses between different groups doesn't need to be checked.
1144   unsigned ASId = 0;
1145   for (auto &AS : AST) {
1146     int NumReadPtrChecks = 0;
1147     int NumWritePtrChecks = 0;
1148     bool CanDoAliasSetRT = true;
1149     ++ASId;
1150     auto ASPointers = AS.getPointers();
1151 
1152     // We assign consecutive id to access from different dependence sets.
1153     // Accesses within the same set don't need a runtime check.
1154     unsigned RunningDepId = 1;
1155     DenseMap<Value *, unsigned> DepSetId;
1156 
1157     SmallVector<std::pair<MemAccessInfo, Type *>, 4> Retries;
1158 
1159     // First, count how many write and read accesses are in the alias set. Also
1160     // collect MemAccessInfos for later.
1161     SmallVector<MemAccessInfo, 4> AccessInfos;
1162     for (const Value *Ptr_ : ASPointers) {
1163       Value *Ptr = const_cast<Value *>(Ptr_);
1164       bool IsWrite = Accesses.count(MemAccessInfo(Ptr, true));
1165       if (IsWrite)
1166         ++NumWritePtrChecks;
1167       else
1168         ++NumReadPtrChecks;
1169       AccessInfos.emplace_back(Ptr, IsWrite);
1170     }
1171 
1172     // We do not need runtime checks for this alias set, if there are no writes
1173     // or a single write and no reads.
1174     if (NumWritePtrChecks == 0 ||
1175         (NumWritePtrChecks == 1 && NumReadPtrChecks == 0)) {
1176       assert((ASPointers.size() <= 1 ||
1177               all_of(ASPointers,
1178                      [this](const Value *Ptr) {
1179                        MemAccessInfo AccessWrite(const_cast<Value *>(Ptr),
1180                                                  true);
1181                        return DepCands.findValue(AccessWrite) == DepCands.end();
1182                      })) &&
1183              "Can only skip updating CanDoRT below, if all entries in AS "
1184              "are reads or there is at most 1 entry");
1185       continue;
1186     }
1187 
1188     for (auto &Access : AccessInfos) {
1189       for (const auto &AccessTy : Accesses[Access]) {
1190         if (!createCheckForAccess(RtCheck, Access, AccessTy, StridesMap,
1191                                   DepSetId, TheLoop, RunningDepId, ASId,
1192                                   ShouldCheckWrap, false)) {
1193           LLVM_DEBUG(dbgs() << "LAA: Can't find bounds for ptr:"
1194                             << *Access.getPointer() << '\n');
1195           Retries.push_back({Access, AccessTy});
1196           CanDoAliasSetRT = false;
1197         }
1198       }
1199     }
1200 
1201     // Note that this function computes CanDoRT and MayNeedRTCheck
1202     // independently. For example CanDoRT=false, MayNeedRTCheck=false means that
1203     // we have a pointer for which we couldn't find the bounds but we don't
1204     // actually need to emit any checks so it does not matter.
1205     //
1206     // We need runtime checks for this alias set, if there are at least 2
1207     // dependence sets (in which case RunningDepId > 2) or if we need to re-try
1208     // any bound checks (because in that case the number of dependence sets is
1209     // incomplete).
1210     bool NeedsAliasSetRTCheck = RunningDepId > 2 || !Retries.empty();
1211 
1212     // We need to perform run-time alias checks, but some pointers had bounds
1213     // that couldn't be checked.
1214     if (NeedsAliasSetRTCheck && !CanDoAliasSetRT) {
1215       // Reset the CanDoSetRt flag and retry all accesses that have failed.
1216       // We know that we need these checks, so we can now be more aggressive
1217       // and add further checks if required (overflow checks).
1218       CanDoAliasSetRT = true;
1219       for (auto Retry : Retries) {
1220         MemAccessInfo Access = Retry.first;
1221         Type *AccessTy = Retry.second;
1222         if (!createCheckForAccess(RtCheck, Access, AccessTy, StridesMap,
1223                                   DepSetId, TheLoop, RunningDepId, ASId,
1224                                   ShouldCheckWrap, /*Assume=*/true)) {
1225           CanDoAliasSetRT = false;
1226           UncomputablePtr = Access.getPointer();
1227           break;
1228         }
1229       }
1230     }
1231 
1232     CanDoRT &= CanDoAliasSetRT;
1233     MayNeedRTCheck |= NeedsAliasSetRTCheck;
1234     ++ASId;
1235   }
1236 
1237   // If the pointers that we would use for the bounds comparison have different
1238   // address spaces, assume the values aren't directly comparable, so we can't
1239   // use them for the runtime check. We also have to assume they could
1240   // overlap. In the future there should be metadata for whether address spaces
1241   // are disjoint.
1242   unsigned NumPointers = RtCheck.Pointers.size();
1243   for (unsigned i = 0; i < NumPointers; ++i) {
1244     for (unsigned j = i + 1; j < NumPointers; ++j) {
1245       // Only need to check pointers between two different dependency sets.
1246       if (RtCheck.Pointers[i].DependencySetId ==
1247           RtCheck.Pointers[j].DependencySetId)
1248        continue;
1249       // Only need to check pointers in the same alias set.
1250       if (RtCheck.Pointers[i].AliasSetId != RtCheck.Pointers[j].AliasSetId)
1251         continue;
1252 
1253       Value *PtrI = RtCheck.Pointers[i].PointerValue;
1254       Value *PtrJ = RtCheck.Pointers[j].PointerValue;
1255 
1256       unsigned ASi = PtrI->getType()->getPointerAddressSpace();
1257       unsigned ASj = PtrJ->getType()->getPointerAddressSpace();
1258       if (ASi != ASj) {
1259         LLVM_DEBUG(
1260             dbgs() << "LAA: Runtime check would require comparison between"
1261                       " different address spaces\n");
1262         return false;
1263       }
1264     }
1265   }
1266 
1267   if (MayNeedRTCheck && CanDoRT)
1268     RtCheck.generateChecks(DepCands, IsDepCheckNeeded);
1269 
1270   LLVM_DEBUG(dbgs() << "LAA: We need to do " << RtCheck.getNumberOfChecks()
1271                     << " pointer comparisons.\n");
1272 
1273   // If we can do run-time checks, but there are no checks, no runtime checks
1274   // are needed. This can happen when all pointers point to the same underlying
1275   // object for example.
1276   RtCheck.Need = CanDoRT ? RtCheck.getNumberOfChecks() != 0 : MayNeedRTCheck;
1277 
1278   bool CanDoRTIfNeeded = !RtCheck.Need || CanDoRT;
1279   if (!CanDoRTIfNeeded)
1280     RtCheck.reset();
1281   return CanDoRTIfNeeded;
1282 }
1283 
1284 void AccessAnalysis::processMemAccesses() {
1285   // We process the set twice: first we process read-write pointers, last we
1286   // process read-only pointers. This allows us to skip dependence tests for
1287   // read-only pointers.
1288 
1289   LLVM_DEBUG(dbgs() << "LAA: Processing memory accesses...\n");
1290   LLVM_DEBUG(dbgs() << "  AST: "; AST.dump());
1291   LLVM_DEBUG(dbgs() << "LAA:   Accesses(" << Accesses.size() << "):\n");
1292   LLVM_DEBUG({
1293     for (auto A : Accesses)
1294       dbgs() << "\t" << *A.first.getPointer() << " ("
1295              << (A.first.getInt()
1296                      ? "write"
1297                      : (ReadOnlyPtr.count(A.first.getPointer()) ? "read-only"
1298                                                                 : "read"))
1299              << ")\n";
1300   });
1301 
1302   // The AliasSetTracker has nicely partitioned our pointers by metadata
1303   // compatibility and potential for underlying-object overlap. As a result, we
1304   // only need to check for potential pointer dependencies within each alias
1305   // set.
1306   for (const auto &AS : AST) {
1307     // Note that both the alias-set tracker and the alias sets themselves used
1308     // ordered collections internally and so the iteration order here is
1309     // deterministic.
1310     auto ASPointers = AS.getPointers();
1311 
1312     bool SetHasWrite = false;
1313 
1314     // Map of pointers to last access encountered.
1315     typedef DenseMap<const Value*, MemAccessInfo> UnderlyingObjToAccessMap;
1316     UnderlyingObjToAccessMap ObjToLastAccess;
1317 
1318     // Set of access to check after all writes have been processed.
1319     PtrAccessMap DeferredAccesses;
1320 
1321     // Iterate over each alias set twice, once to process read/write pointers,
1322     // and then to process read-only pointers.
1323     for (int SetIteration = 0; SetIteration < 2; ++SetIteration) {
1324       bool UseDeferred = SetIteration > 0;
1325       PtrAccessMap &S = UseDeferred ? DeferredAccesses : Accesses;
1326 
1327       for (const Value *Ptr_ : ASPointers) {
1328         Value *Ptr = const_cast<Value *>(Ptr_);
1329 
1330         // For a single memory access in AliasSetTracker, Accesses may contain
1331         // both read and write, and they both need to be handled for CheckDeps.
1332         for (const auto &AC : S) {
1333           if (AC.first.getPointer() != Ptr)
1334             continue;
1335 
1336           bool IsWrite = AC.first.getInt();
1337 
1338           // If we're using the deferred access set, then it contains only
1339           // reads.
1340           bool IsReadOnlyPtr = ReadOnlyPtr.count(Ptr) && !IsWrite;
1341           if (UseDeferred && !IsReadOnlyPtr)
1342             continue;
1343           // Otherwise, the pointer must be in the PtrAccessSet, either as a
1344           // read or a write.
1345           assert(((IsReadOnlyPtr && UseDeferred) || IsWrite ||
1346                   S.count(MemAccessInfo(Ptr, false))) &&
1347                  "Alias-set pointer not in the access set?");
1348 
1349           MemAccessInfo Access(Ptr, IsWrite);
1350           DepCands.insert(Access);
1351 
1352           // Memorize read-only pointers for later processing and skip them in
1353           // the first round (they need to be checked after we have seen all
1354           // write pointers). Note: we also mark pointer that are not
1355           // consecutive as "read-only" pointers (so that we check
1356           // "a[b[i]] +="). Hence, we need the second check for "!IsWrite".
1357           if (!UseDeferred && IsReadOnlyPtr) {
1358             // We only use the pointer keys, the types vector values don't
1359             // matter.
1360             DeferredAccesses.insert({Access, {}});
1361             continue;
1362           }
1363 
1364           // If this is a write - check other reads and writes for conflicts. If
1365           // this is a read only check other writes for conflicts (but only if
1366           // there is no other write to the ptr - this is an optimization to
1367           // catch "a[i] = a[i] + " without having to do a dependence check).
1368           if ((IsWrite || IsReadOnlyPtr) && SetHasWrite) {
1369             CheckDeps.push_back(Access);
1370             IsRTCheckAnalysisNeeded = true;
1371           }
1372 
1373           if (IsWrite)
1374             SetHasWrite = true;
1375 
1376           // Create sets of pointers connected by a shared alias set and
1377           // underlying object.
1378           typedef SmallVector<const Value *, 16> ValueVector;
1379           ValueVector TempObjects;
1380 
1381           UnderlyingObjects[Ptr] = {};
1382           SmallVector<const Value *, 16> &UOs = UnderlyingObjects[Ptr];
1383           ::getUnderlyingObjects(Ptr, UOs, LI);
1384           LLVM_DEBUG(dbgs()
1385                      << "Underlying objects for pointer " << *Ptr << "\n");
1386           for (const Value *UnderlyingObj : UOs) {
1387             // nullptr never alias, don't join sets for pointer that have "null"
1388             // in their UnderlyingObjects list.
1389             if (isa<ConstantPointerNull>(UnderlyingObj) &&
1390                 !NullPointerIsDefined(
1391                     TheLoop->getHeader()->getParent(),
1392                     UnderlyingObj->getType()->getPointerAddressSpace()))
1393               continue;
1394 
1395             UnderlyingObjToAccessMap::iterator Prev =
1396                 ObjToLastAccess.find(UnderlyingObj);
1397             if (Prev != ObjToLastAccess.end())
1398               DepCands.unionSets(Access, Prev->second);
1399 
1400             ObjToLastAccess[UnderlyingObj] = Access;
1401             LLVM_DEBUG(dbgs() << "  " << *UnderlyingObj << "\n");
1402           }
1403         }
1404       }
1405     }
1406   }
1407 }
1408 
1409 /// Return true if an AddRec pointer \p Ptr is unsigned non-wrapping,
1410 /// i.e. monotonically increasing/decreasing.
1411 static bool isNoWrapAddRec(Value *Ptr, const SCEVAddRecExpr *AR,
1412                            PredicatedScalarEvolution &PSE, const Loop *L) {
1413 
1414   // FIXME: This should probably only return true for NUW.
1415   if (AR->getNoWrapFlags(SCEV::NoWrapMask))
1416     return true;
1417 
1418   if (PSE.hasNoOverflow(Ptr, SCEVWrapPredicate::IncrementNUSW))
1419     return true;
1420 
1421   // Scalar evolution does not propagate the non-wrapping flags to values that
1422   // are derived from a non-wrapping induction variable because non-wrapping
1423   // could be flow-sensitive.
1424   //
1425   // Look through the potentially overflowing instruction to try to prove
1426   // non-wrapping for the *specific* value of Ptr.
1427 
1428   // The arithmetic implied by an inbounds GEP can't overflow.
1429   auto *GEP = dyn_cast<GetElementPtrInst>(Ptr);
1430   if (!GEP || !GEP->isInBounds())
1431     return false;
1432 
1433   // Make sure there is only one non-const index and analyze that.
1434   Value *NonConstIndex = nullptr;
1435   for (Value *Index : GEP->indices())
1436     if (!isa<ConstantInt>(Index)) {
1437       if (NonConstIndex)
1438         return false;
1439       NonConstIndex = Index;
1440     }
1441   if (!NonConstIndex)
1442     // The recurrence is on the pointer, ignore for now.
1443     return false;
1444 
1445   // The index in GEP is signed.  It is non-wrapping if it's derived from a NSW
1446   // AddRec using a NSW operation.
1447   if (auto *OBO = dyn_cast<OverflowingBinaryOperator>(NonConstIndex))
1448     if (OBO->hasNoSignedWrap() &&
1449         // Assume constant for other the operand so that the AddRec can be
1450         // easily found.
1451         isa<ConstantInt>(OBO->getOperand(1))) {
1452       auto *OpScev = PSE.getSCEV(OBO->getOperand(0));
1453 
1454       if (auto *OpAR = dyn_cast<SCEVAddRecExpr>(OpScev))
1455         return OpAR->getLoop() == L && OpAR->getNoWrapFlags(SCEV::FlagNSW);
1456     }
1457 
1458   return false;
1459 }
1460 
1461 /// Check whether the access through \p Ptr has a constant stride.
1462 std::optional<int64_t> llvm::getPtrStride(PredicatedScalarEvolution &PSE,
1463                                           Type *AccessTy, Value *Ptr,
1464                                           const Loop *Lp,
1465                                           const DenseMap<Value *, const SCEV *> &StridesMap,
1466                                           bool Assume, bool ShouldCheckWrap) {
1467   Type *Ty = Ptr->getType();
1468   assert(Ty->isPointerTy() && "Unexpected non-ptr");
1469 
1470   if (isa<ScalableVectorType>(AccessTy)) {
1471     LLVM_DEBUG(dbgs() << "LAA: Bad stride - Scalable object: " << *AccessTy
1472                       << "\n");
1473     return std::nullopt;
1474   }
1475 
1476   const SCEV *PtrScev = replaceSymbolicStrideSCEV(PSE, StridesMap, Ptr);
1477 
1478   const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
1479   if (Assume && !AR)
1480     AR = PSE.getAsAddRec(Ptr);
1481 
1482   if (!AR) {
1483     LLVM_DEBUG(dbgs() << "LAA: Bad stride - Not an AddRecExpr pointer " << *Ptr
1484                       << " SCEV: " << *PtrScev << "\n");
1485     return std::nullopt;
1486   }
1487 
1488   // The access function must stride over the innermost loop.
1489   if (Lp != AR->getLoop()) {
1490     LLVM_DEBUG(dbgs() << "LAA: Bad stride - Not striding over innermost loop "
1491                       << *Ptr << " SCEV: " << *AR << "\n");
1492     return std::nullopt;
1493   }
1494 
1495   // Check the step is constant.
1496   const SCEV *Step = AR->getStepRecurrence(*PSE.getSE());
1497 
1498   // Calculate the pointer stride and check if it is constant.
1499   const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
1500   if (!C) {
1501     LLVM_DEBUG(dbgs() << "LAA: Bad stride - Not a constant strided " << *Ptr
1502                       << " SCEV: " << *AR << "\n");
1503     return std::nullopt;
1504   }
1505 
1506   auto &DL = Lp->getHeader()->getModule()->getDataLayout();
1507   TypeSize AllocSize = DL.getTypeAllocSize(AccessTy);
1508   int64_t Size = AllocSize.getFixedValue();
1509   const APInt &APStepVal = C->getAPInt();
1510 
1511   // Huge step value - give up.
1512   if (APStepVal.getBitWidth() > 64)
1513     return std::nullopt;
1514 
1515   int64_t StepVal = APStepVal.getSExtValue();
1516 
1517   // Strided access.
1518   int64_t Stride = StepVal / Size;
1519   int64_t Rem = StepVal % Size;
1520   if (Rem)
1521     return std::nullopt;
1522 
1523   if (!ShouldCheckWrap)
1524     return Stride;
1525 
1526   // The address calculation must not wrap. Otherwise, a dependence could be
1527   // inverted.
1528   if (isNoWrapAddRec(Ptr, AR, PSE, Lp))
1529     return Stride;
1530 
1531   // An inbounds getelementptr that is a AddRec with a unit stride
1532   // cannot wrap per definition.  If it did, the result would be poison
1533   // and any memory access dependent on it would be immediate UB
1534   // when executed.
1535   if (auto *GEP = dyn_cast<GetElementPtrInst>(Ptr);
1536       GEP && GEP->isInBounds() && (Stride == 1 || Stride == -1))
1537     return Stride;
1538 
1539   // If the null pointer is undefined, then a access sequence which would
1540   // otherwise access it can be assumed not to unsigned wrap.  Note that this
1541   // assumes the object in memory is aligned to the natural alignment.
1542   unsigned AddrSpace = Ty->getPointerAddressSpace();
1543   if (!NullPointerIsDefined(Lp->getHeader()->getParent(), AddrSpace) &&
1544       (Stride == 1 || Stride == -1))
1545     return Stride;
1546 
1547   if (Assume) {
1548     PSE.setNoOverflow(Ptr, SCEVWrapPredicate::IncrementNUSW);
1549     LLVM_DEBUG(dbgs() << "LAA: Pointer may wrap:\n"
1550                       << "LAA:   Pointer: " << *Ptr << "\n"
1551                       << "LAA:   SCEV: " << *AR << "\n"
1552                       << "LAA:   Added an overflow assumption\n");
1553     return Stride;
1554   }
1555   LLVM_DEBUG(
1556       dbgs() << "LAA: Bad stride - Pointer may wrap in the address space "
1557              << *Ptr << " SCEV: " << *AR << "\n");
1558   return std::nullopt;
1559 }
1560 
1561 std::optional<int> llvm::getPointersDiff(Type *ElemTyA, Value *PtrA,
1562                                          Type *ElemTyB, Value *PtrB,
1563                                          const DataLayout &DL,
1564                                          ScalarEvolution &SE, bool StrictCheck,
1565                                          bool CheckType) {
1566   assert(PtrA && PtrB && "Expected non-nullptr pointers.");
1567 
1568   // Make sure that A and B are different pointers.
1569   if (PtrA == PtrB)
1570     return 0;
1571 
1572   // Make sure that the element types are the same if required.
1573   if (CheckType && ElemTyA != ElemTyB)
1574     return std::nullopt;
1575 
1576   unsigned ASA = PtrA->getType()->getPointerAddressSpace();
1577   unsigned ASB = PtrB->getType()->getPointerAddressSpace();
1578 
1579   // Check that the address spaces match.
1580   if (ASA != ASB)
1581     return std::nullopt;
1582   unsigned IdxWidth = DL.getIndexSizeInBits(ASA);
1583 
1584   APInt OffsetA(IdxWidth, 0), OffsetB(IdxWidth, 0);
1585   Value *PtrA1 = PtrA->stripAndAccumulateInBoundsConstantOffsets(DL, OffsetA);
1586   Value *PtrB1 = PtrB->stripAndAccumulateInBoundsConstantOffsets(DL, OffsetB);
1587 
1588   int Val;
1589   if (PtrA1 == PtrB1) {
1590     // Retrieve the address space again as pointer stripping now tracks through
1591     // `addrspacecast`.
1592     ASA = cast<PointerType>(PtrA1->getType())->getAddressSpace();
1593     ASB = cast<PointerType>(PtrB1->getType())->getAddressSpace();
1594     // Check that the address spaces match and that the pointers are valid.
1595     if (ASA != ASB)
1596       return std::nullopt;
1597 
1598     IdxWidth = DL.getIndexSizeInBits(ASA);
1599     OffsetA = OffsetA.sextOrTrunc(IdxWidth);
1600     OffsetB = OffsetB.sextOrTrunc(IdxWidth);
1601 
1602     OffsetB -= OffsetA;
1603     Val = OffsetB.getSExtValue();
1604   } else {
1605     // Otherwise compute the distance with SCEV between the base pointers.
1606     const SCEV *PtrSCEVA = SE.getSCEV(PtrA);
1607     const SCEV *PtrSCEVB = SE.getSCEV(PtrB);
1608     const auto *Diff =
1609         dyn_cast<SCEVConstant>(SE.getMinusSCEV(PtrSCEVB, PtrSCEVA));
1610     if (!Diff)
1611       return std::nullopt;
1612     Val = Diff->getAPInt().getSExtValue();
1613   }
1614   int Size = DL.getTypeStoreSize(ElemTyA);
1615   int Dist = Val / Size;
1616 
1617   // Ensure that the calculated distance matches the type-based one after all
1618   // the bitcasts removal in the provided pointers.
1619   if (!StrictCheck || Dist * Size == Val)
1620     return Dist;
1621   return std::nullopt;
1622 }
1623 
1624 bool llvm::sortPtrAccesses(ArrayRef<Value *> VL, Type *ElemTy,
1625                            const DataLayout &DL, ScalarEvolution &SE,
1626                            SmallVectorImpl<unsigned> &SortedIndices) {
1627   assert(llvm::all_of(
1628              VL, [](const Value *V) { return V->getType()->isPointerTy(); }) &&
1629          "Expected list of pointer operands.");
1630   // Walk over the pointers, and map each of them to an offset relative to
1631   // first pointer in the array.
1632   Value *Ptr0 = VL[0];
1633 
1634   using DistOrdPair = std::pair<int64_t, int>;
1635   auto Compare = llvm::less_first();
1636   std::set<DistOrdPair, decltype(Compare)> Offsets(Compare);
1637   Offsets.emplace(0, 0);
1638   int Cnt = 1;
1639   bool IsConsecutive = true;
1640   for (auto *Ptr : VL.drop_front()) {
1641     std::optional<int> Diff = getPointersDiff(ElemTy, Ptr0, ElemTy, Ptr, DL, SE,
1642                                               /*StrictCheck=*/true);
1643     if (!Diff)
1644       return false;
1645 
1646     // Check if the pointer with the same offset is found.
1647     int64_t Offset = *Diff;
1648     auto Res = Offsets.emplace(Offset, Cnt);
1649     if (!Res.second)
1650       return false;
1651     // Consecutive order if the inserted element is the last one.
1652     IsConsecutive = IsConsecutive && std::next(Res.first) == Offsets.end();
1653     ++Cnt;
1654   }
1655   SortedIndices.clear();
1656   if (!IsConsecutive) {
1657     // Fill SortedIndices array only if it is non-consecutive.
1658     SortedIndices.resize(VL.size());
1659     Cnt = 0;
1660     for (const std::pair<int64_t, int> &Pair : Offsets) {
1661       SortedIndices[Cnt] = Pair.second;
1662       ++Cnt;
1663     }
1664   }
1665   return true;
1666 }
1667 
1668 /// Returns true if the memory operations \p A and \p B are consecutive.
1669 bool llvm::isConsecutiveAccess(Value *A, Value *B, const DataLayout &DL,
1670                                ScalarEvolution &SE, bool CheckType) {
1671   Value *PtrA = getLoadStorePointerOperand(A);
1672   Value *PtrB = getLoadStorePointerOperand(B);
1673   if (!PtrA || !PtrB)
1674     return false;
1675   Type *ElemTyA = getLoadStoreType(A);
1676   Type *ElemTyB = getLoadStoreType(B);
1677   std::optional<int> Diff =
1678       getPointersDiff(ElemTyA, PtrA, ElemTyB, PtrB, DL, SE,
1679                       /*StrictCheck=*/true, CheckType);
1680   return Diff && *Diff == 1;
1681 }
1682 
1683 void MemoryDepChecker::addAccess(StoreInst *SI) {
1684   visitPointers(SI->getPointerOperand(), *InnermostLoop,
1685                 [this, SI](Value *Ptr) {
1686                   Accesses[MemAccessInfo(Ptr, true)].push_back(AccessIdx);
1687                   InstMap.push_back(SI);
1688                   ++AccessIdx;
1689                 });
1690 }
1691 
1692 void MemoryDepChecker::addAccess(LoadInst *LI) {
1693   visitPointers(LI->getPointerOperand(), *InnermostLoop,
1694                 [this, LI](Value *Ptr) {
1695                   Accesses[MemAccessInfo(Ptr, false)].push_back(AccessIdx);
1696                   InstMap.push_back(LI);
1697                   ++AccessIdx;
1698                 });
1699 }
1700 
1701 MemoryDepChecker::VectorizationSafetyStatus
1702 MemoryDepChecker::Dependence::isSafeForVectorization(DepType Type) {
1703   switch (Type) {
1704   case NoDep:
1705   case Forward:
1706   case BackwardVectorizable:
1707     return VectorizationSafetyStatus::Safe;
1708 
1709   case Unknown:
1710     return VectorizationSafetyStatus::PossiblySafeWithRtChecks;
1711   case ForwardButPreventsForwarding:
1712   case Backward:
1713   case BackwardVectorizableButPreventsForwarding:
1714   case IndirectUnsafe:
1715     return VectorizationSafetyStatus::Unsafe;
1716   }
1717   llvm_unreachable("unexpected DepType!");
1718 }
1719 
1720 bool MemoryDepChecker::Dependence::isBackward() const {
1721   switch (Type) {
1722   case NoDep:
1723   case Forward:
1724   case ForwardButPreventsForwarding:
1725   case Unknown:
1726   case IndirectUnsafe:
1727     return false;
1728 
1729   case BackwardVectorizable:
1730   case Backward:
1731   case BackwardVectorizableButPreventsForwarding:
1732     return true;
1733   }
1734   llvm_unreachable("unexpected DepType!");
1735 }
1736 
1737 bool MemoryDepChecker::Dependence::isPossiblyBackward() const {
1738   return isBackward() || Type == Unknown;
1739 }
1740 
1741 bool MemoryDepChecker::Dependence::isForward() const {
1742   switch (Type) {
1743   case Forward:
1744   case ForwardButPreventsForwarding:
1745     return true;
1746 
1747   case NoDep:
1748   case Unknown:
1749   case BackwardVectorizable:
1750   case Backward:
1751   case BackwardVectorizableButPreventsForwarding:
1752   case IndirectUnsafe:
1753     return false;
1754   }
1755   llvm_unreachable("unexpected DepType!");
1756 }
1757 
1758 bool MemoryDepChecker::couldPreventStoreLoadForward(uint64_t Distance,
1759                                                     uint64_t TypeByteSize) {
1760   // If loads occur at a distance that is not a multiple of a feasible vector
1761   // factor store-load forwarding does not take place.
1762   // Positive dependences might cause troubles because vectorizing them might
1763   // prevent store-load forwarding making vectorized code run a lot slower.
1764   //   a[i] = a[i-3] ^ a[i-8];
1765   //   The stores to a[i:i+1] don't align with the stores to a[i-3:i-2] and
1766   //   hence on your typical architecture store-load forwarding does not take
1767   //   place. Vectorizing in such cases does not make sense.
1768   // Store-load forwarding distance.
1769 
1770   // After this many iterations store-to-load forwarding conflicts should not
1771   // cause any slowdowns.
1772   const uint64_t NumItersForStoreLoadThroughMemory = 8 * TypeByteSize;
1773   // Maximum vector factor.
1774   uint64_t MaxVFWithoutSLForwardIssues = std::min(
1775       VectorizerParams::MaxVectorWidth * TypeByteSize, MinDepDistBytes);
1776 
1777   // Compute the smallest VF at which the store and load would be misaligned.
1778   for (uint64_t VF = 2 * TypeByteSize; VF <= MaxVFWithoutSLForwardIssues;
1779        VF *= 2) {
1780     // If the number of vector iteration between the store and the load are
1781     // small we could incur conflicts.
1782     if (Distance % VF && Distance / VF < NumItersForStoreLoadThroughMemory) {
1783       MaxVFWithoutSLForwardIssues = (VF >> 1);
1784       break;
1785     }
1786   }
1787 
1788   if (MaxVFWithoutSLForwardIssues < 2 * TypeByteSize) {
1789     LLVM_DEBUG(
1790         dbgs() << "LAA: Distance " << Distance
1791                << " that could cause a store-load forwarding conflict\n");
1792     return true;
1793   }
1794 
1795   if (MaxVFWithoutSLForwardIssues < MinDepDistBytes &&
1796       MaxVFWithoutSLForwardIssues !=
1797           VectorizerParams::MaxVectorWidth * TypeByteSize)
1798     MinDepDistBytes = MaxVFWithoutSLForwardIssues;
1799   return false;
1800 }
1801 
1802 void MemoryDepChecker::mergeInStatus(VectorizationSafetyStatus S) {
1803   if (Status < S)
1804     Status = S;
1805 }
1806 
1807 /// Given a dependence-distance \p Dist between two
1808 /// memory accesses, that have the same stride whose absolute value is given
1809 /// in \p Stride, and that have the same type size \p TypeByteSize,
1810 /// in a loop whose takenCount is \p BackedgeTakenCount, check if it is
1811 /// possible to prove statically that the dependence distance is larger
1812 /// than the range that the accesses will travel through the execution of
1813 /// the loop. If so, return true; false otherwise. This is useful for
1814 /// example in loops such as the following (PR31098):
1815 ///     for (i = 0; i < D; ++i) {
1816 ///                = out[i];
1817 ///       out[i+D] =
1818 ///     }
1819 static bool isSafeDependenceDistance(const DataLayout &DL, ScalarEvolution &SE,
1820                                      const SCEV &BackedgeTakenCount,
1821                                      const SCEV &Dist, uint64_t Stride,
1822                                      uint64_t TypeByteSize) {
1823 
1824   // If we can prove that
1825   //      (**) |Dist| > BackedgeTakenCount * Step
1826   // where Step is the absolute stride of the memory accesses in bytes,
1827   // then there is no dependence.
1828   //
1829   // Rationale:
1830   // We basically want to check if the absolute distance (|Dist/Step|)
1831   // is >= the loop iteration count (or > BackedgeTakenCount).
1832   // This is equivalent to the Strong SIV Test (Practical Dependence Testing,
1833   // Section 4.2.1); Note, that for vectorization it is sufficient to prove
1834   // that the dependence distance is >= VF; This is checked elsewhere.
1835   // But in some cases we can prune dependence distances early, and
1836   // even before selecting the VF, and without a runtime test, by comparing
1837   // the distance against the loop iteration count. Since the vectorized code
1838   // will be executed only if LoopCount >= VF, proving distance >= LoopCount
1839   // also guarantees that distance >= VF.
1840   //
1841   const uint64_t ByteStride = Stride * TypeByteSize;
1842   const SCEV *Step = SE.getConstant(BackedgeTakenCount.getType(), ByteStride);
1843   const SCEV *Product = SE.getMulExpr(&BackedgeTakenCount, Step);
1844 
1845   const SCEV *CastedDist = &Dist;
1846   const SCEV *CastedProduct = Product;
1847   uint64_t DistTypeSizeBits = DL.getTypeSizeInBits(Dist.getType());
1848   uint64_t ProductTypeSizeBits = DL.getTypeSizeInBits(Product->getType());
1849 
1850   // The dependence distance can be positive/negative, so we sign extend Dist;
1851   // The multiplication of the absolute stride in bytes and the
1852   // backedgeTakenCount is non-negative, so we zero extend Product.
1853   if (DistTypeSizeBits > ProductTypeSizeBits)
1854     CastedProduct = SE.getZeroExtendExpr(Product, Dist.getType());
1855   else
1856     CastedDist = SE.getNoopOrSignExtend(&Dist, Product->getType());
1857 
1858   // Is  Dist - (BackedgeTakenCount * Step) > 0 ?
1859   // (If so, then we have proven (**) because |Dist| >= Dist)
1860   const SCEV *Minus = SE.getMinusSCEV(CastedDist, CastedProduct);
1861   if (SE.isKnownPositive(Minus))
1862     return true;
1863 
1864   // Second try: Is  -Dist - (BackedgeTakenCount * Step) > 0 ?
1865   // (If so, then we have proven (**) because |Dist| >= -1*Dist)
1866   const SCEV *NegDist = SE.getNegativeSCEV(CastedDist);
1867   Minus = SE.getMinusSCEV(NegDist, CastedProduct);
1868   if (SE.isKnownPositive(Minus))
1869     return true;
1870 
1871   return false;
1872 }
1873 
1874 /// Check the dependence for two accesses with the same stride \p Stride.
1875 /// \p Distance is the positive distance and \p TypeByteSize is type size in
1876 /// bytes.
1877 ///
1878 /// \returns true if they are independent.
1879 static bool areStridedAccessesIndependent(uint64_t Distance, uint64_t Stride,
1880                                           uint64_t TypeByteSize) {
1881   assert(Stride > 1 && "The stride must be greater than 1");
1882   assert(TypeByteSize > 0 && "The type size in byte must be non-zero");
1883   assert(Distance > 0 && "The distance must be non-zero");
1884 
1885   // Skip if the distance is not multiple of type byte size.
1886   if (Distance % TypeByteSize)
1887     return false;
1888 
1889   uint64_t ScaledDist = Distance / TypeByteSize;
1890 
1891   // No dependence if the scaled distance is not multiple of the stride.
1892   // E.g.
1893   //      for (i = 0; i < 1024 ; i += 4)
1894   //        A[i+2] = A[i] + 1;
1895   //
1896   // Two accesses in memory (scaled distance is 2, stride is 4):
1897   //     | A[0] |      |      |      | A[4] |      |      |      |
1898   //     |      |      | A[2] |      |      |      | A[6] |      |
1899   //
1900   // E.g.
1901   //      for (i = 0; i < 1024 ; i += 3)
1902   //        A[i+4] = A[i] + 1;
1903   //
1904   // Two accesses in memory (scaled distance is 4, stride is 3):
1905   //     | A[0] |      |      | A[3] |      |      | A[6] |      |      |
1906   //     |      |      |      |      | A[4] |      |      | A[7] |      |
1907   return ScaledDist % Stride;
1908 }
1909 
1910 /// Returns true if any of the underlying objects has a loop varying address,
1911 /// i.e. may change in \p L.
1912 static bool
1913 isLoopVariantIndirectAddress(ArrayRef<const Value *> UnderlyingObjects,
1914                              ScalarEvolution &SE, const Loop *L) {
1915   return any_of(UnderlyingObjects, [&SE, L](const Value *UO) {
1916     return !SE.isLoopInvariant(SE.getSCEV(const_cast<Value *>(UO)), L);
1917   });
1918 }
1919 
1920 // Get the dependence distance, stride, type size in whether i is a write for
1921 // the dependence between A and B. Returns a DepType, if we can prove there's
1922 // no dependence or the analysis fails. Outlined to lambda to limit he scope
1923 // of various temporary variables, like A/BPtr, StrideA/BPtr and others.
1924 // Returns either the dependence result, if it could already be determined, or a
1925 // tuple with (Distance, Stride, TypeSize, AIsWrite, BIsWrite).
1926 static std::variant<MemoryDepChecker::Dependence::DepType,
1927                     std::tuple<const SCEV *, uint64_t, uint64_t, bool, bool>>
1928 getDependenceDistanceStrideAndSize(
1929     const AccessAnalysis::MemAccessInfo &A, Instruction *AInst,
1930     const AccessAnalysis::MemAccessInfo &B, Instruction *BInst,
1931     const DenseMap<Value *, const SCEV *> &Strides,
1932     const DenseMap<Value *, SmallVector<const Value *, 16>> &UnderlyingObjects,
1933     PredicatedScalarEvolution &PSE, const Loop *InnermostLoop) {
1934   auto &DL = InnermostLoop->getHeader()->getModule()->getDataLayout();
1935   auto &SE = *PSE.getSE();
1936   auto [APtr, AIsWrite] = A;
1937   auto [BPtr, BIsWrite] = B;
1938 
1939   // Two reads are independent.
1940   if (!AIsWrite && !BIsWrite)
1941     return MemoryDepChecker::Dependence::NoDep;
1942 
1943   Type *ATy = getLoadStoreType(AInst);
1944   Type *BTy = getLoadStoreType(BInst);
1945 
1946   // We cannot check pointers in different address spaces.
1947   if (APtr->getType()->getPointerAddressSpace() !=
1948       BPtr->getType()->getPointerAddressSpace())
1949     return MemoryDepChecker::Dependence::Unknown;
1950 
1951   int64_t StrideAPtr =
1952       getPtrStride(PSE, ATy, APtr, InnermostLoop, Strides, true).value_or(0);
1953   int64_t StrideBPtr =
1954       getPtrStride(PSE, BTy, BPtr, InnermostLoop, Strides, true).value_or(0);
1955 
1956   const SCEV *Src = PSE.getSCEV(APtr);
1957   const SCEV *Sink = PSE.getSCEV(BPtr);
1958 
1959   // If the induction step is negative we have to invert source and sink of the
1960   // dependence when measuring the distance between them. We should not swap
1961   // AIsWrite with BIsWrite, as their uses expect them in program order.
1962   if (StrideAPtr < 0) {
1963     std::swap(Src, Sink);
1964     std::swap(AInst, BInst);
1965   }
1966 
1967   const SCEV *Dist = SE.getMinusSCEV(Sink, Src);
1968 
1969   LLVM_DEBUG(dbgs() << "LAA: Src Scev: " << *Src << "Sink Scev: " << *Sink
1970                     << "(Induction step: " << StrideAPtr << ")\n");
1971   LLVM_DEBUG(dbgs() << "LAA: Distance for " << *AInst << " to " << *BInst
1972                     << ": " << *Dist << "\n");
1973 
1974   // Needs accesses where the addresses of the accessed underlying objects do
1975   // not change within the loop.
1976   if (isLoopVariantIndirectAddress(UnderlyingObjects.find(APtr)->second, SE,
1977                                    InnermostLoop) ||
1978       isLoopVariantIndirectAddress(UnderlyingObjects.find(BPtr)->second, SE,
1979                                    InnermostLoop))
1980     return MemoryDepChecker::Dependence::IndirectUnsafe;
1981 
1982   // Need accesses with constant stride. We don't want to vectorize
1983   // "A[B[i]] += ..." and similar code or pointer arithmetic that could wrap
1984   // in the address space.
1985   if (!StrideAPtr || !StrideBPtr || StrideAPtr != StrideBPtr) {
1986     LLVM_DEBUG(dbgs() << "Pointer access with non-constant stride\n");
1987     return MemoryDepChecker::Dependence::Unknown;
1988   }
1989 
1990   uint64_t TypeByteSize = DL.getTypeAllocSize(ATy);
1991   bool HasSameSize =
1992       DL.getTypeStoreSizeInBits(ATy) == DL.getTypeStoreSizeInBits(BTy);
1993   if (!HasSameSize)
1994     TypeByteSize = 0;
1995   uint64_t Stride = std::abs(StrideAPtr);
1996   return std::make_tuple(Dist, Stride, TypeByteSize, AIsWrite, BIsWrite);
1997 }
1998 
1999 MemoryDepChecker::Dependence::DepType MemoryDepChecker::isDependent(
2000     const MemAccessInfo &A, unsigned AIdx, const MemAccessInfo &B,
2001     unsigned BIdx, const DenseMap<Value *, const SCEV *> &Strides,
2002     const DenseMap<Value *, SmallVector<const Value *, 16>>
2003         &UnderlyingObjects) {
2004   assert(AIdx < BIdx && "Must pass arguments in program order");
2005 
2006   // Get the dependence distance, stride, type size and what access writes for
2007   // the dependence between A and B.
2008   auto Res = getDependenceDistanceStrideAndSize(
2009       A, InstMap[AIdx], B, InstMap[BIdx], Strides, UnderlyingObjects, PSE,
2010       InnermostLoop);
2011   if (std::holds_alternative<Dependence::DepType>(Res))
2012     return std::get<Dependence::DepType>(Res);
2013 
2014   const auto &[Dist, Stride, TypeByteSize, AIsWrite, BIsWrite] =
2015       std::get<std::tuple<const SCEV *, uint64_t, uint64_t, bool, bool>>(Res);
2016   bool HasSameSize = TypeByteSize > 0;
2017 
2018   ScalarEvolution &SE = *PSE.getSE();
2019   auto &DL = InnermostLoop->getHeader()->getModule()->getDataLayout();
2020   if (!isa<SCEVCouldNotCompute>(Dist) && HasSameSize &&
2021       isSafeDependenceDistance(DL, SE, *(PSE.getBackedgeTakenCount()), *Dist,
2022                                Stride, TypeByteSize))
2023     return Dependence::NoDep;
2024 
2025   const SCEVConstant *C = dyn_cast<SCEVConstant>(Dist);
2026   if (!C) {
2027     LLVM_DEBUG(dbgs() << "LAA: Dependence because of non-constant distance\n");
2028     FoundNonConstantDistanceDependence = true;
2029     return Dependence::Unknown;
2030   }
2031 
2032   const APInt &Val = C->getAPInt();
2033   int64_t Distance = Val.getSExtValue();
2034 
2035   // Attempt to prove strided accesses independent.
2036   if (std::abs(Distance) > 0 && Stride > 1 && HasSameSize &&
2037       areStridedAccessesIndependent(std::abs(Distance), Stride, TypeByteSize)) {
2038     LLVM_DEBUG(dbgs() << "LAA: Strided accesses are independent\n");
2039     return Dependence::NoDep;
2040   }
2041 
2042   // Negative distances are not plausible dependencies.
2043   if (Val.isNegative()) {
2044     bool IsTrueDataDependence = (AIsWrite && !BIsWrite);
2045     // There is no need to update MaxSafeVectorWidthInBits after call to
2046     // couldPreventStoreLoadForward, even if it changed MinDepDistBytes,
2047     // since a forward dependency will allow vectorization using any width.
2048     if (IsTrueDataDependence && EnableForwardingConflictDetection &&
2049         (!HasSameSize || couldPreventStoreLoadForward(Val.abs().getZExtValue(),
2050                                                       TypeByteSize))) {
2051       LLVM_DEBUG(dbgs() << "LAA: Forward but may prevent st->ld forwarding\n");
2052       return Dependence::ForwardButPreventsForwarding;
2053     }
2054 
2055     LLVM_DEBUG(dbgs() << "LAA: Dependence is negative\n");
2056     return Dependence::Forward;
2057   }
2058 
2059   // Write to the same location with the same size.
2060   if (Val == 0) {
2061     if (HasSameSize)
2062       return Dependence::Forward;
2063     LLVM_DEBUG(
2064         dbgs() << "LAA: Zero dependence difference but different type sizes\n");
2065     return Dependence::Unknown;
2066   }
2067 
2068   assert(Val.isStrictlyPositive() && "Expect a positive value");
2069 
2070   if (!HasSameSize) {
2071     LLVM_DEBUG(dbgs() << "LAA: ReadWrite-Write positive dependency with "
2072                          "different type sizes\n");
2073     return Dependence::Unknown;
2074   }
2075 
2076   // Bail out early if passed-in parameters make vectorization not feasible.
2077   unsigned ForcedFactor = (VectorizerParams::VectorizationFactor ?
2078                            VectorizerParams::VectorizationFactor : 1);
2079   unsigned ForcedUnroll = (VectorizerParams::VectorizationInterleave ?
2080                            VectorizerParams::VectorizationInterleave : 1);
2081   // The minimum number of iterations for a vectorized/unrolled version.
2082   unsigned MinNumIter = std::max(ForcedFactor * ForcedUnroll, 2U);
2083 
2084   // It's not vectorizable if the distance is smaller than the minimum distance
2085   // needed for a vectroized/unrolled version. Vectorizing one iteration in
2086   // front needs TypeByteSize * Stride. Vectorizing the last iteration needs
2087   // TypeByteSize (No need to plus the last gap distance).
2088   //
2089   // E.g. Assume one char is 1 byte in memory and one int is 4 bytes.
2090   //      foo(int *A) {
2091   //        int *B = (int *)((char *)A + 14);
2092   //        for (i = 0 ; i < 1024 ; i += 2)
2093   //          B[i] = A[i] + 1;
2094   //      }
2095   //
2096   // Two accesses in memory (stride is 2):
2097   //     | A[0] |      | A[2] |      | A[4] |      | A[6] |      |
2098   //                              | B[0] |      | B[2] |      | B[4] |
2099   //
2100   // Distance needs for vectorizing iterations except the last iteration:
2101   // 4 * 2 * (MinNumIter - 1). Distance needs for the last iteration: 4.
2102   // So the minimum distance needed is: 4 * 2 * (MinNumIter - 1) + 4.
2103   //
2104   // If MinNumIter is 2, it is vectorizable as the minimum distance needed is
2105   // 12, which is less than distance.
2106   //
2107   // If MinNumIter is 4 (Say if a user forces the vectorization factor to be 4),
2108   // the minimum distance needed is 28, which is greater than distance. It is
2109   // not safe to do vectorization.
2110   uint64_t MinDistanceNeeded =
2111       TypeByteSize * Stride * (MinNumIter - 1) + TypeByteSize;
2112   if (MinDistanceNeeded > static_cast<uint64_t>(Distance)) {
2113     LLVM_DEBUG(dbgs() << "LAA: Failure because of positive distance "
2114                       << Distance << '\n');
2115     return Dependence::Backward;
2116   }
2117 
2118   // Unsafe if the minimum distance needed is greater than smallest dependence
2119   // distance distance.
2120   if (MinDistanceNeeded > MinDepDistBytes) {
2121     LLVM_DEBUG(dbgs() << "LAA: Failure because it needs at least "
2122                       << MinDistanceNeeded << " size in bytes\n");
2123     return Dependence::Backward;
2124   }
2125 
2126   // Positive distance bigger than max vectorization factor.
2127   // FIXME: Should use max factor instead of max distance in bytes, which could
2128   // not handle different types.
2129   // E.g. Assume one char is 1 byte in memory and one int is 4 bytes.
2130   //      void foo (int *A, char *B) {
2131   //        for (unsigned i = 0; i < 1024; i++) {
2132   //          A[i+2] = A[i] + 1;
2133   //          B[i+2] = B[i] + 1;
2134   //        }
2135   //      }
2136   //
2137   // This case is currently unsafe according to the max safe distance. If we
2138   // analyze the two accesses on array B, the max safe dependence distance
2139   // is 2. Then we analyze the accesses on array A, the minimum distance needed
2140   // is 8, which is less than 2 and forbidden vectorization, But actually
2141   // both A and B could be vectorized by 2 iterations.
2142   MinDepDistBytes =
2143       std::min(static_cast<uint64_t>(Distance), MinDepDistBytes);
2144 
2145   bool IsTrueDataDependence = (!AIsWrite && BIsWrite);
2146   uint64_t MinDepDistBytesOld = MinDepDistBytes;
2147   if (IsTrueDataDependence && EnableForwardingConflictDetection &&
2148       couldPreventStoreLoadForward(Distance, TypeByteSize)) {
2149     // Sanity check that we didn't update MinDepDistBytes when calling
2150     // couldPreventStoreLoadForward
2151     assert(MinDepDistBytes == MinDepDistBytesOld &&
2152            "An update to MinDepDistBytes requires an update to "
2153            "MaxSafeVectorWidthInBits");
2154     (void)MinDepDistBytesOld;
2155     return Dependence::BackwardVectorizableButPreventsForwarding;
2156   }
2157 
2158   // An update to MinDepDistBytes requires an update to MaxSafeVectorWidthInBits
2159   // since there is a backwards dependency.
2160   uint64_t MaxVF = MinDepDistBytes / (TypeByteSize * Stride);
2161   LLVM_DEBUG(dbgs() << "LAA: Positive distance " << Val.getSExtValue()
2162                     << " with max VF = " << MaxVF << '\n');
2163   uint64_t MaxVFInBits = MaxVF * TypeByteSize * 8;
2164   MaxSafeVectorWidthInBits = std::min(MaxSafeVectorWidthInBits, MaxVFInBits);
2165   return Dependence::BackwardVectorizable;
2166 }
2167 
2168 bool MemoryDepChecker::areDepsSafe(
2169     DepCandidates &AccessSets, MemAccessInfoList &CheckDeps,
2170     const DenseMap<Value *, const SCEV *> &Strides,
2171     const DenseMap<Value *, SmallVector<const Value *, 16>>
2172         &UnderlyingObjects) {
2173 
2174   MinDepDistBytes = -1;
2175   SmallPtrSet<MemAccessInfo, 8> Visited;
2176   for (MemAccessInfo CurAccess : CheckDeps) {
2177     if (Visited.count(CurAccess))
2178       continue;
2179 
2180     // Get the relevant memory access set.
2181     EquivalenceClasses<MemAccessInfo>::iterator I =
2182       AccessSets.findValue(AccessSets.getLeaderValue(CurAccess));
2183 
2184     // Check accesses within this set.
2185     EquivalenceClasses<MemAccessInfo>::member_iterator AI =
2186         AccessSets.member_begin(I);
2187     EquivalenceClasses<MemAccessInfo>::member_iterator AE =
2188         AccessSets.member_end();
2189 
2190     // Check every access pair.
2191     while (AI != AE) {
2192       Visited.insert(*AI);
2193       bool AIIsWrite = AI->getInt();
2194       // Check loads only against next equivalent class, but stores also against
2195       // other stores in the same equivalence class - to the same address.
2196       EquivalenceClasses<MemAccessInfo>::member_iterator OI =
2197           (AIIsWrite ? AI : std::next(AI));
2198       while (OI != AE) {
2199         // Check every accessing instruction pair in program order.
2200         for (std::vector<unsigned>::iterator I1 = Accesses[*AI].begin(),
2201              I1E = Accesses[*AI].end(); I1 != I1E; ++I1)
2202           // Scan all accesses of another equivalence class, but only the next
2203           // accesses of the same equivalent class.
2204           for (std::vector<unsigned>::iterator
2205                    I2 = (OI == AI ? std::next(I1) : Accesses[*OI].begin()),
2206                    I2E = (OI == AI ? I1E : Accesses[*OI].end());
2207                I2 != I2E; ++I2) {
2208             auto A = std::make_pair(&*AI, *I1);
2209             auto B = std::make_pair(&*OI, *I2);
2210 
2211             assert(*I1 != *I2);
2212             if (*I1 > *I2)
2213               std::swap(A, B);
2214 
2215             Dependence::DepType Type =
2216                 isDependent(*A.first, A.second, *B.first, B.second, Strides,
2217                             UnderlyingObjects);
2218             mergeInStatus(Dependence::isSafeForVectorization(Type));
2219 
2220             // Gather dependences unless we accumulated MaxDependences
2221             // dependences.  In that case return as soon as we find the first
2222             // unsafe dependence.  This puts a limit on this quadratic
2223             // algorithm.
2224             if (RecordDependences) {
2225               if (Type != Dependence::NoDep)
2226                 Dependences.push_back(Dependence(A.second, B.second, Type));
2227 
2228               if (Dependences.size() >= MaxDependences) {
2229                 RecordDependences = false;
2230                 Dependences.clear();
2231                 LLVM_DEBUG(dbgs()
2232                            << "Too many dependences, stopped recording\n");
2233               }
2234             }
2235             if (!RecordDependences && !isSafeForVectorization())
2236               return false;
2237           }
2238         ++OI;
2239       }
2240       AI++;
2241     }
2242   }
2243 
2244   LLVM_DEBUG(dbgs() << "Total Dependences: " << Dependences.size() << "\n");
2245   return isSafeForVectorization();
2246 }
2247 
2248 SmallVector<Instruction *, 4>
2249 MemoryDepChecker::getInstructionsForAccess(Value *Ptr, bool isWrite) const {
2250   MemAccessInfo Access(Ptr, isWrite);
2251   auto &IndexVector = Accesses.find(Access)->second;
2252 
2253   SmallVector<Instruction *, 4> Insts;
2254   transform(IndexVector,
2255                  std::back_inserter(Insts),
2256                  [&](unsigned Idx) { return this->InstMap[Idx]; });
2257   return Insts;
2258 }
2259 
2260 const char *MemoryDepChecker::Dependence::DepName[] = {
2261     "NoDep",
2262     "Unknown",
2263     "IndidrectUnsafe",
2264     "Forward",
2265     "ForwardButPreventsForwarding",
2266     "Backward",
2267     "BackwardVectorizable",
2268     "BackwardVectorizableButPreventsForwarding"};
2269 
2270 void MemoryDepChecker::Dependence::print(
2271     raw_ostream &OS, unsigned Depth,
2272     const SmallVectorImpl<Instruction *> &Instrs) const {
2273   OS.indent(Depth) << DepName[Type] << ":\n";
2274   OS.indent(Depth + 2) << *Instrs[Source] << " -> \n";
2275   OS.indent(Depth + 2) << *Instrs[Destination] << "\n";
2276 }
2277 
2278 bool LoopAccessInfo::canAnalyzeLoop() {
2279   // We need to have a loop header.
2280   LLVM_DEBUG(dbgs() << "LAA: Found a loop in "
2281                     << TheLoop->getHeader()->getParent()->getName() << ": "
2282                     << TheLoop->getHeader()->getName() << '\n');
2283 
2284   // We can only analyze innermost loops.
2285   if (!TheLoop->isInnermost()) {
2286     LLVM_DEBUG(dbgs() << "LAA: loop is not the innermost loop\n");
2287     recordAnalysis("NotInnerMostLoop") << "loop is not the innermost loop";
2288     return false;
2289   }
2290 
2291   // We must have a single backedge.
2292   if (TheLoop->getNumBackEdges() != 1) {
2293     LLVM_DEBUG(
2294         dbgs() << "LAA: loop control flow is not understood by analyzer\n");
2295     recordAnalysis("CFGNotUnderstood")
2296         << "loop control flow is not understood by analyzer";
2297     return false;
2298   }
2299 
2300   // ScalarEvolution needs to be able to find the exit count.
2301   const SCEV *ExitCount = PSE->getBackedgeTakenCount();
2302   if (isa<SCEVCouldNotCompute>(ExitCount)) {
2303     recordAnalysis("CantComputeNumberOfIterations")
2304         << "could not determine number of loop iterations";
2305     LLVM_DEBUG(dbgs() << "LAA: SCEV could not compute the loop exit count.\n");
2306     return false;
2307   }
2308 
2309   return true;
2310 }
2311 
2312 void LoopAccessInfo::analyzeLoop(AAResults *AA, LoopInfo *LI,
2313                                  const TargetLibraryInfo *TLI,
2314                                  DominatorTree *DT) {
2315   // Holds the Load and Store instructions.
2316   SmallVector<LoadInst *, 16> Loads;
2317   SmallVector<StoreInst *, 16> Stores;
2318   SmallPtrSet<MDNode *, 8> LoopAliasScopes;
2319 
2320   // Holds all the different accesses in the loop.
2321   unsigned NumReads = 0;
2322   unsigned NumReadWrites = 0;
2323 
2324   bool HasComplexMemInst = false;
2325 
2326   // A runtime check is only legal to insert if there are no convergent calls.
2327   HasConvergentOp = false;
2328 
2329   PtrRtChecking->Pointers.clear();
2330   PtrRtChecking->Need = false;
2331 
2332   const bool IsAnnotatedParallel = TheLoop->isAnnotatedParallel();
2333 
2334   const bool EnableMemAccessVersioningOfLoop =
2335       EnableMemAccessVersioning &&
2336       !TheLoop->getHeader()->getParent()->hasOptSize();
2337 
2338   // Traverse blocks in fixed RPOT order, regardless of their storage in the
2339   // loop info, as it may be arbitrary.
2340   LoopBlocksRPO RPOT(TheLoop);
2341   RPOT.perform(LI);
2342   for (BasicBlock *BB : RPOT) {
2343     // Scan the BB and collect legal loads and stores. Also detect any
2344     // convergent instructions.
2345     for (Instruction &I : *BB) {
2346       if (auto *Call = dyn_cast<CallBase>(&I)) {
2347         if (Call->isConvergent())
2348           HasConvergentOp = true;
2349       }
2350 
2351       // With both a non-vectorizable memory instruction and a convergent
2352       // operation, found in this loop, no reason to continue the search.
2353       if (HasComplexMemInst && HasConvergentOp) {
2354         CanVecMem = false;
2355         return;
2356       }
2357 
2358       // Avoid hitting recordAnalysis multiple times.
2359       if (HasComplexMemInst)
2360         continue;
2361 
2362       // Record alias scopes defined inside the loop.
2363       if (auto *Decl = dyn_cast<NoAliasScopeDeclInst>(&I))
2364         for (Metadata *Op : Decl->getScopeList()->operands())
2365           LoopAliasScopes.insert(cast<MDNode>(Op));
2366 
2367       // Many math library functions read the rounding mode. We will only
2368       // vectorize a loop if it contains known function calls that don't set
2369       // the flag. Therefore, it is safe to ignore this read from memory.
2370       auto *Call = dyn_cast<CallInst>(&I);
2371       if (Call && getVectorIntrinsicIDForCall(Call, TLI))
2372         continue;
2373 
2374       // If this is a load, save it. If this instruction can read from memory
2375       // but is not a load, then we quit. Notice that we don't handle function
2376       // calls that read or write.
2377       if (I.mayReadFromMemory()) {
2378         // If the function has an explicit vectorized counterpart, we can safely
2379         // assume that it can be vectorized.
2380         if (Call && !Call->isNoBuiltin() && Call->getCalledFunction() &&
2381             !VFDatabase::getMappings(*Call).empty())
2382           continue;
2383 
2384         auto *Ld = dyn_cast<LoadInst>(&I);
2385         if (!Ld) {
2386           recordAnalysis("CantVectorizeInstruction", Ld)
2387             << "instruction cannot be vectorized";
2388           HasComplexMemInst = true;
2389           continue;
2390         }
2391         if (!Ld->isSimple() && !IsAnnotatedParallel) {
2392           recordAnalysis("NonSimpleLoad", Ld)
2393               << "read with atomic ordering or volatile read";
2394           LLVM_DEBUG(dbgs() << "LAA: Found a non-simple load.\n");
2395           HasComplexMemInst = true;
2396           continue;
2397         }
2398         NumLoads++;
2399         Loads.push_back(Ld);
2400         DepChecker->addAccess(Ld);
2401         if (EnableMemAccessVersioningOfLoop)
2402           collectStridedAccess(Ld);
2403         continue;
2404       }
2405 
2406       // Save 'store' instructions. Abort if other instructions write to memory.
2407       if (I.mayWriteToMemory()) {
2408         auto *St = dyn_cast<StoreInst>(&I);
2409         if (!St) {
2410           recordAnalysis("CantVectorizeInstruction", St)
2411               << "instruction cannot be vectorized";
2412           HasComplexMemInst = true;
2413           continue;
2414         }
2415         if (!St->isSimple() && !IsAnnotatedParallel) {
2416           recordAnalysis("NonSimpleStore", St)
2417               << "write with atomic ordering or volatile write";
2418           LLVM_DEBUG(dbgs() << "LAA: Found a non-simple store.\n");
2419           HasComplexMemInst = true;
2420           continue;
2421         }
2422         NumStores++;
2423         Stores.push_back(St);
2424         DepChecker->addAccess(St);
2425         if (EnableMemAccessVersioningOfLoop)
2426           collectStridedAccess(St);
2427       }
2428     } // Next instr.
2429   } // Next block.
2430 
2431   if (HasComplexMemInst) {
2432     CanVecMem = false;
2433     return;
2434   }
2435 
2436   // Now we have two lists that hold the loads and the stores.
2437   // Next, we find the pointers that they use.
2438 
2439   // Check if we see any stores. If there are no stores, then we don't
2440   // care if the pointers are *restrict*.
2441   if (!Stores.size()) {
2442     LLVM_DEBUG(dbgs() << "LAA: Found a read-only loop!\n");
2443     CanVecMem = true;
2444     return;
2445   }
2446 
2447   MemoryDepChecker::DepCandidates DependentAccesses;
2448   AccessAnalysis Accesses(TheLoop, AA, LI, DependentAccesses, *PSE,
2449                           LoopAliasScopes);
2450 
2451   // Holds the analyzed pointers. We don't want to call getUnderlyingObjects
2452   // multiple times on the same object. If the ptr is accessed twice, once
2453   // for read and once for write, it will only appear once (on the write
2454   // list). This is okay, since we are going to check for conflicts between
2455   // writes and between reads and writes, but not between reads and reads.
2456   SmallSet<std::pair<Value *, Type *>, 16> Seen;
2457 
2458   // Record uniform store addresses to identify if we have multiple stores
2459   // to the same address.
2460   SmallPtrSet<Value *, 16> UniformStores;
2461 
2462   for (StoreInst *ST : Stores) {
2463     Value *Ptr = ST->getPointerOperand();
2464 
2465     if (isInvariant(Ptr)) {
2466       // Record store instructions to loop invariant addresses
2467       StoresToInvariantAddresses.push_back(ST);
2468       HasDependenceInvolvingLoopInvariantAddress |=
2469           !UniformStores.insert(Ptr).second;
2470     }
2471 
2472     // If we did *not* see this pointer before, insert it to  the read-write
2473     // list. At this phase it is only a 'write' list.
2474     Type *AccessTy = getLoadStoreType(ST);
2475     if (Seen.insert({Ptr, AccessTy}).second) {
2476       ++NumReadWrites;
2477 
2478       MemoryLocation Loc = MemoryLocation::get(ST);
2479       // The TBAA metadata could have a control dependency on the predication
2480       // condition, so we cannot rely on it when determining whether or not we
2481       // need runtime pointer checks.
2482       if (blockNeedsPredication(ST->getParent(), TheLoop, DT))
2483         Loc.AATags.TBAA = nullptr;
2484 
2485       visitPointers(const_cast<Value *>(Loc.Ptr), *TheLoop,
2486                     [&Accesses, AccessTy, Loc](Value *Ptr) {
2487                       MemoryLocation NewLoc = Loc.getWithNewPtr(Ptr);
2488                       Accesses.addStore(NewLoc, AccessTy);
2489                     });
2490     }
2491   }
2492 
2493   if (IsAnnotatedParallel) {
2494     LLVM_DEBUG(
2495         dbgs() << "LAA: A loop annotated parallel, ignore memory dependency "
2496                << "checks.\n");
2497     CanVecMem = true;
2498     return;
2499   }
2500 
2501   for (LoadInst *LD : Loads) {
2502     Value *Ptr = LD->getPointerOperand();
2503     // If we did *not* see this pointer before, insert it to the
2504     // read list. If we *did* see it before, then it is already in
2505     // the read-write list. This allows us to vectorize expressions
2506     // such as A[i] += x;  Because the address of A[i] is a read-write
2507     // pointer. This only works if the index of A[i] is consecutive.
2508     // If the address of i is unknown (for example A[B[i]]) then we may
2509     // read a few words, modify, and write a few words, and some of the
2510     // words may be written to the same address.
2511     bool IsReadOnlyPtr = false;
2512     Type *AccessTy = getLoadStoreType(LD);
2513     if (Seen.insert({Ptr, AccessTy}).second ||
2514         !getPtrStride(*PSE, LD->getType(), Ptr, TheLoop, SymbolicStrides).value_or(0)) {
2515       ++NumReads;
2516       IsReadOnlyPtr = true;
2517     }
2518 
2519     // See if there is an unsafe dependency between a load to a uniform address and
2520     // store to the same uniform address.
2521     if (UniformStores.count(Ptr)) {
2522       LLVM_DEBUG(dbgs() << "LAA: Found an unsafe dependency between a uniform "
2523                            "load and uniform store to the same address!\n");
2524       HasDependenceInvolvingLoopInvariantAddress = true;
2525     }
2526 
2527     MemoryLocation Loc = MemoryLocation::get(LD);
2528     // The TBAA metadata could have a control dependency on the predication
2529     // condition, so we cannot rely on it when determining whether or not we
2530     // need runtime pointer checks.
2531     if (blockNeedsPredication(LD->getParent(), TheLoop, DT))
2532       Loc.AATags.TBAA = nullptr;
2533 
2534     visitPointers(const_cast<Value *>(Loc.Ptr), *TheLoop,
2535                   [&Accesses, AccessTy, Loc, IsReadOnlyPtr](Value *Ptr) {
2536                     MemoryLocation NewLoc = Loc.getWithNewPtr(Ptr);
2537                     Accesses.addLoad(NewLoc, AccessTy, IsReadOnlyPtr);
2538                   });
2539   }
2540 
2541   // If we write (or read-write) to a single destination and there are no
2542   // other reads in this loop then is it safe to vectorize.
2543   if (NumReadWrites == 1 && NumReads == 0) {
2544     LLVM_DEBUG(dbgs() << "LAA: Found a write-only loop!\n");
2545     CanVecMem = true;
2546     return;
2547   }
2548 
2549   // Build dependence sets and check whether we need a runtime pointer bounds
2550   // check.
2551   Accesses.buildDependenceSets();
2552 
2553   // Find pointers with computable bounds. We are going to use this information
2554   // to place a runtime bound check.
2555   Value *UncomputablePtr = nullptr;
2556   bool CanDoRTIfNeeded =
2557       Accesses.canCheckPtrAtRT(*PtrRtChecking, PSE->getSE(), TheLoop,
2558                                SymbolicStrides, UncomputablePtr, false);
2559   if (!CanDoRTIfNeeded) {
2560     auto *I = dyn_cast_or_null<Instruction>(UncomputablePtr);
2561     recordAnalysis("CantIdentifyArrayBounds", I)
2562         << "cannot identify array bounds";
2563     LLVM_DEBUG(dbgs() << "LAA: We can't vectorize because we can't find "
2564                       << "the array bounds.\n");
2565     CanVecMem = false;
2566     return;
2567   }
2568 
2569   LLVM_DEBUG(
2570     dbgs() << "LAA: May be able to perform a memory runtime check if needed.\n");
2571 
2572   CanVecMem = true;
2573   if (Accesses.isDependencyCheckNeeded()) {
2574     LLVM_DEBUG(dbgs() << "LAA: Checking memory dependencies\n");
2575     CanVecMem = DepChecker->areDepsSafe(
2576         DependentAccesses, Accesses.getDependenciesToCheck(), SymbolicStrides,
2577         Accesses.getUnderlyingObjects());
2578 
2579     if (!CanVecMem && DepChecker->shouldRetryWithRuntimeCheck()) {
2580       LLVM_DEBUG(dbgs() << "LAA: Retrying with memory checks\n");
2581 
2582       // Clear the dependency checks. We assume they are not needed.
2583       Accesses.resetDepChecks(*DepChecker);
2584 
2585       PtrRtChecking->reset();
2586       PtrRtChecking->Need = true;
2587 
2588       auto *SE = PSE->getSE();
2589       UncomputablePtr = nullptr;
2590       CanDoRTIfNeeded = Accesses.canCheckPtrAtRT(
2591           *PtrRtChecking, SE, TheLoop, SymbolicStrides, UncomputablePtr, true);
2592 
2593       // Check that we found the bounds for the pointer.
2594       if (!CanDoRTIfNeeded) {
2595         auto *I = dyn_cast_or_null<Instruction>(UncomputablePtr);
2596         recordAnalysis("CantCheckMemDepsAtRunTime", I)
2597             << "cannot check memory dependencies at runtime";
2598         LLVM_DEBUG(dbgs() << "LAA: Can't vectorize with memory checks\n");
2599         CanVecMem = false;
2600         return;
2601       }
2602 
2603       CanVecMem = true;
2604     }
2605   }
2606 
2607   if (HasConvergentOp) {
2608     recordAnalysis("CantInsertRuntimeCheckWithConvergent")
2609       << "cannot add control dependency to convergent operation";
2610     LLVM_DEBUG(dbgs() << "LAA: We can't vectorize because a runtime check "
2611                          "would be needed with a convergent operation\n");
2612     CanVecMem = false;
2613     return;
2614   }
2615 
2616   if (CanVecMem)
2617     LLVM_DEBUG(
2618         dbgs() << "LAA: No unsafe dependent memory operations in loop.  We"
2619                << (PtrRtChecking->Need ? "" : " don't")
2620                << " need runtime memory checks.\n");
2621   else
2622     emitUnsafeDependenceRemark();
2623 }
2624 
2625 void LoopAccessInfo::emitUnsafeDependenceRemark() {
2626   auto Deps = getDepChecker().getDependences();
2627   if (!Deps)
2628     return;
2629   auto Found = llvm::find_if(*Deps, [](const MemoryDepChecker::Dependence &D) {
2630     return MemoryDepChecker::Dependence::isSafeForVectorization(D.Type) !=
2631            MemoryDepChecker::VectorizationSafetyStatus::Safe;
2632   });
2633   if (Found == Deps->end())
2634     return;
2635   MemoryDepChecker::Dependence Dep = *Found;
2636 
2637   LLVM_DEBUG(dbgs() << "LAA: unsafe dependent memory operations in loop\n");
2638 
2639   // Emit remark for first unsafe dependence
2640   bool HasForcedDistribution = false;
2641   std::optional<const MDOperand *> Value =
2642       findStringMetadataForLoop(TheLoop, "llvm.loop.distribute.enable");
2643   if (Value) {
2644     const MDOperand *Op = *Value;
2645     assert(Op && mdconst::hasa<ConstantInt>(*Op) && "invalid metadata");
2646     HasForcedDistribution = mdconst::extract<ConstantInt>(*Op)->getZExtValue();
2647   }
2648 
2649   const std::string Info =
2650       HasForcedDistribution
2651           ? "unsafe dependent memory operations in loop."
2652           : "unsafe dependent memory operations in loop. Use "
2653             "#pragma clang loop distribute(enable) to allow loop distribution "
2654             "to attempt to isolate the offending operations into a separate "
2655             "loop";
2656   OptimizationRemarkAnalysis &R =
2657       recordAnalysis("UnsafeDep", Dep.getDestination(*this)) << Info;
2658 
2659   switch (Dep.Type) {
2660   case MemoryDepChecker::Dependence::NoDep:
2661   case MemoryDepChecker::Dependence::Forward:
2662   case MemoryDepChecker::Dependence::BackwardVectorizable:
2663     llvm_unreachable("Unexpected dependence");
2664   case MemoryDepChecker::Dependence::Backward:
2665     R << "\nBackward loop carried data dependence.";
2666     break;
2667   case MemoryDepChecker::Dependence::ForwardButPreventsForwarding:
2668     R << "\nForward loop carried data dependence that prevents "
2669          "store-to-load forwarding.";
2670     break;
2671   case MemoryDepChecker::Dependence::BackwardVectorizableButPreventsForwarding:
2672     R << "\nBackward loop carried data dependence that prevents "
2673          "store-to-load forwarding.";
2674     break;
2675   case MemoryDepChecker::Dependence::IndirectUnsafe:
2676     R << "\nUnsafe indirect dependence.";
2677     break;
2678   case MemoryDepChecker::Dependence::Unknown:
2679     R << "\nUnknown data dependence.";
2680     break;
2681   }
2682 
2683   if (Instruction *I = Dep.getSource(*this)) {
2684     DebugLoc SourceLoc = I->getDebugLoc();
2685     if (auto *DD = dyn_cast_or_null<Instruction>(getPointerOperand(I)))
2686       SourceLoc = DD->getDebugLoc();
2687     if (SourceLoc)
2688       R << " Memory location is the same as accessed at "
2689         << ore::NV("Location", SourceLoc);
2690   }
2691 }
2692 
2693 bool LoopAccessInfo::blockNeedsPredication(BasicBlock *BB, Loop *TheLoop,
2694                                            DominatorTree *DT)  {
2695   assert(TheLoop->contains(BB) && "Unknown block used");
2696 
2697   // Blocks that do not dominate the latch need predication.
2698   BasicBlock* Latch = TheLoop->getLoopLatch();
2699   return !DT->dominates(BB, Latch);
2700 }
2701 
2702 OptimizationRemarkAnalysis &LoopAccessInfo::recordAnalysis(StringRef RemarkName,
2703                                                            Instruction *I) {
2704   assert(!Report && "Multiple reports generated");
2705 
2706   Value *CodeRegion = TheLoop->getHeader();
2707   DebugLoc DL = TheLoop->getStartLoc();
2708 
2709   if (I) {
2710     CodeRegion = I->getParent();
2711     // If there is no debug location attached to the instruction, revert back to
2712     // using the loop's.
2713     if (I->getDebugLoc())
2714       DL = I->getDebugLoc();
2715   }
2716 
2717   Report = std::make_unique<OptimizationRemarkAnalysis>(DEBUG_TYPE, RemarkName, DL,
2718                                                    CodeRegion);
2719   return *Report;
2720 }
2721 
2722 bool LoopAccessInfo::isInvariant(Value *V) const {
2723   auto *SE = PSE->getSE();
2724   // TODO: Is this really what we want? Even without FP SCEV, we may want some
2725   // trivially loop-invariant FP values to be considered invariant.
2726   if (!SE->isSCEVable(V->getType()))
2727     return false;
2728   const SCEV *S = SE->getSCEV(V);
2729   return SE->isLoopInvariant(S, TheLoop);
2730 }
2731 
2732 /// Find the operand of the GEP that should be checked for consecutive
2733 /// stores. This ignores trailing indices that have no effect on the final
2734 /// pointer.
2735 static unsigned getGEPInductionOperand(const GetElementPtrInst *Gep) {
2736   const DataLayout &DL = Gep->getModule()->getDataLayout();
2737   unsigned LastOperand = Gep->getNumOperands() - 1;
2738   TypeSize GEPAllocSize = DL.getTypeAllocSize(Gep->getResultElementType());
2739 
2740   // Walk backwards and try to peel off zeros.
2741   while (LastOperand > 1 && match(Gep->getOperand(LastOperand), m_Zero())) {
2742     // Find the type we're currently indexing into.
2743     gep_type_iterator GEPTI = gep_type_begin(Gep);
2744     std::advance(GEPTI, LastOperand - 2);
2745 
2746     // If it's a type with the same allocation size as the result of the GEP we
2747     // can peel off the zero index.
2748     TypeSize ElemSize = GEPTI.isStruct()
2749                             ? DL.getTypeAllocSize(GEPTI.getIndexedType())
2750                             : GEPTI.getSequentialElementStride(DL);
2751     if (ElemSize != GEPAllocSize)
2752       break;
2753     --LastOperand;
2754   }
2755 
2756   return LastOperand;
2757 }
2758 
2759 /// If the argument is a GEP, then returns the operand identified by
2760 /// getGEPInductionOperand. However, if there is some other non-loop-invariant
2761 /// operand, it returns that instead.
2762 static Value *stripGetElementPtr(Value *Ptr, ScalarEvolution *SE, Loop *Lp) {
2763   GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr);
2764   if (!GEP)
2765     return Ptr;
2766 
2767   unsigned InductionOperand = getGEPInductionOperand(GEP);
2768 
2769   // Check that all of the gep indices are uniform except for our induction
2770   // operand.
2771   for (unsigned i = 0, e = GEP->getNumOperands(); i != e; ++i)
2772     if (i != InductionOperand &&
2773         !SE->isLoopInvariant(SE->getSCEV(GEP->getOperand(i)), Lp))
2774       return Ptr;
2775   return GEP->getOperand(InductionOperand);
2776 }
2777 
2778 /// If a value has only one user that is a CastInst, return it.
2779 static Value *getUniqueCastUse(Value *Ptr, Loop *Lp, Type *Ty) {
2780   Value *UniqueCast = nullptr;
2781   for (User *U : Ptr->users()) {
2782     CastInst *CI = dyn_cast<CastInst>(U);
2783     if (CI && CI->getType() == Ty) {
2784       if (!UniqueCast)
2785         UniqueCast = CI;
2786       else
2787         return nullptr;
2788     }
2789   }
2790   return UniqueCast;
2791 }
2792 
2793 /// Get the stride of a pointer access in a loop. Looks for symbolic
2794 /// strides "a[i*stride]". Returns the symbolic stride, or null otherwise.
2795 static const SCEV *getStrideFromPointer(Value *Ptr, ScalarEvolution *SE, Loop *Lp) {
2796   auto *PtrTy = dyn_cast<PointerType>(Ptr->getType());
2797   if (!PtrTy || PtrTy->isAggregateType())
2798     return nullptr;
2799 
2800   // Try to remove a gep instruction to make the pointer (actually index at this
2801   // point) easier analyzable. If OrigPtr is equal to Ptr we are analyzing the
2802   // pointer, otherwise, we are analyzing the index.
2803   Value *OrigPtr = Ptr;
2804 
2805   // The size of the pointer access.
2806   int64_t PtrAccessSize = 1;
2807 
2808   Ptr = stripGetElementPtr(Ptr, SE, Lp);
2809   const SCEV *V = SE->getSCEV(Ptr);
2810 
2811   if (Ptr != OrigPtr)
2812     // Strip off casts.
2813     while (const SCEVIntegralCastExpr *C = dyn_cast<SCEVIntegralCastExpr>(V))
2814       V = C->getOperand();
2815 
2816   const SCEVAddRecExpr *S = dyn_cast<SCEVAddRecExpr>(V);
2817   if (!S)
2818     return nullptr;
2819 
2820   // If the pointer is invariant then there is no stride and it makes no
2821   // sense to add it here.
2822   if (Lp != S->getLoop())
2823     return nullptr;
2824 
2825   V = S->getStepRecurrence(*SE);
2826   if (!V)
2827     return nullptr;
2828 
2829   // Strip off the size of access multiplication if we are still analyzing the
2830   // pointer.
2831   if (OrigPtr == Ptr) {
2832     if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(V)) {
2833       if (M->getOperand(0)->getSCEVType() != scConstant)
2834         return nullptr;
2835 
2836       const APInt &APStepVal = cast<SCEVConstant>(M->getOperand(0))->getAPInt();
2837 
2838       // Huge step value - give up.
2839       if (APStepVal.getBitWidth() > 64)
2840         return nullptr;
2841 
2842       int64_t StepVal = APStepVal.getSExtValue();
2843       if (PtrAccessSize != StepVal)
2844         return nullptr;
2845       V = M->getOperand(1);
2846     }
2847   }
2848 
2849   // Note that the restriction after this loop invariant check are only
2850   // profitability restrictions.
2851   if (!SE->isLoopInvariant(V, Lp))
2852     return nullptr;
2853 
2854   // Look for the loop invariant symbolic value.
2855   const SCEVUnknown *U = dyn_cast<SCEVUnknown>(V);
2856   if (!U) {
2857     const auto *C = dyn_cast<SCEVIntegralCastExpr>(V);
2858     if (!C)
2859       return nullptr;
2860     U = dyn_cast<SCEVUnknown>(C->getOperand());
2861     if (!U)
2862       return nullptr;
2863 
2864     // Match legacy behavior - this is not needed for correctness
2865     if (!getUniqueCastUse(U->getValue(), Lp, V->getType()))
2866       return nullptr;
2867   }
2868 
2869   return V;
2870 }
2871 
2872 void LoopAccessInfo::collectStridedAccess(Value *MemAccess) {
2873   Value *Ptr = getLoadStorePointerOperand(MemAccess);
2874   if (!Ptr)
2875     return;
2876 
2877   // Note: getStrideFromPointer is a *profitability* heuristic.  We
2878   // could broaden the scope of values returned here - to anything
2879   // which happens to be loop invariant and contributes to the
2880   // computation of an interesting IV - but we chose not to as we
2881   // don't have a cost model here, and broadening the scope exposes
2882   // far too many unprofitable cases.
2883   const SCEV *StrideExpr = getStrideFromPointer(Ptr, PSE->getSE(), TheLoop);
2884   if (!StrideExpr)
2885     return;
2886 
2887   LLVM_DEBUG(dbgs() << "LAA: Found a strided access that is a candidate for "
2888                        "versioning:");
2889   LLVM_DEBUG(dbgs() << "  Ptr: " << *Ptr << " Stride: " << *StrideExpr << "\n");
2890 
2891   if (!SpeculateUnitStride) {
2892     LLVM_DEBUG(dbgs() << "  Chose not to due to -laa-speculate-unit-stride\n");
2893     return;
2894   }
2895 
2896   // Avoid adding the "Stride == 1" predicate when we know that
2897   // Stride >= Trip-Count. Such a predicate will effectively optimize a single
2898   // or zero iteration loop, as Trip-Count <= Stride == 1.
2899   //
2900   // TODO: We are currently not making a very informed decision on when it is
2901   // beneficial to apply stride versioning. It might make more sense that the
2902   // users of this analysis (such as the vectorizer) will trigger it, based on
2903   // their specific cost considerations; For example, in cases where stride
2904   // versioning does  not help resolving memory accesses/dependences, the
2905   // vectorizer should evaluate the cost of the runtime test, and the benefit
2906   // of various possible stride specializations, considering the alternatives
2907   // of using gather/scatters (if available).
2908 
2909   const SCEV *BETakenCount = PSE->getBackedgeTakenCount();
2910 
2911   // Match the types so we can compare the stride and the BETakenCount.
2912   // The Stride can be positive/negative, so we sign extend Stride;
2913   // The backedgeTakenCount is non-negative, so we zero extend BETakenCount.
2914   const DataLayout &DL = TheLoop->getHeader()->getModule()->getDataLayout();
2915   uint64_t StrideTypeSizeBits = DL.getTypeSizeInBits(StrideExpr->getType());
2916   uint64_t BETypeSizeBits = DL.getTypeSizeInBits(BETakenCount->getType());
2917   const SCEV *CastedStride = StrideExpr;
2918   const SCEV *CastedBECount = BETakenCount;
2919   ScalarEvolution *SE = PSE->getSE();
2920   if (BETypeSizeBits >= StrideTypeSizeBits)
2921     CastedStride = SE->getNoopOrSignExtend(StrideExpr, BETakenCount->getType());
2922   else
2923     CastedBECount = SE->getZeroExtendExpr(BETakenCount, StrideExpr->getType());
2924   const SCEV *StrideMinusBETaken = SE->getMinusSCEV(CastedStride, CastedBECount);
2925   // Since TripCount == BackEdgeTakenCount + 1, checking:
2926   // "Stride >= TripCount" is equivalent to checking:
2927   // Stride - BETakenCount > 0
2928   if (SE->isKnownPositive(StrideMinusBETaken)) {
2929     LLVM_DEBUG(
2930         dbgs() << "LAA: Stride>=TripCount; No point in versioning as the "
2931                   "Stride==1 predicate will imply that the loop executes "
2932                   "at most once.\n");
2933     return;
2934   }
2935   LLVM_DEBUG(dbgs() << "LAA: Found a strided access that we can version.\n");
2936 
2937   // Strip back off the integer cast, and check that our result is a
2938   // SCEVUnknown as we expect.
2939   const SCEV *StrideBase = StrideExpr;
2940   if (const auto *C = dyn_cast<SCEVIntegralCastExpr>(StrideBase))
2941     StrideBase = C->getOperand();
2942   SymbolicStrides[Ptr] = cast<SCEVUnknown>(StrideBase);
2943 }
2944 
2945 LoopAccessInfo::LoopAccessInfo(Loop *L, ScalarEvolution *SE,
2946                                const TargetLibraryInfo *TLI, AAResults *AA,
2947                                DominatorTree *DT, LoopInfo *LI)
2948     : PSE(std::make_unique<PredicatedScalarEvolution>(*SE, *L)),
2949       PtrRtChecking(nullptr),
2950       DepChecker(std::make_unique<MemoryDepChecker>(*PSE, L)), TheLoop(L) {
2951   PtrRtChecking = std::make_unique<RuntimePointerChecking>(*DepChecker, SE);
2952   if (canAnalyzeLoop()) {
2953     analyzeLoop(AA, LI, TLI, DT);
2954   }
2955 }
2956 
2957 void LoopAccessInfo::print(raw_ostream &OS, unsigned Depth) const {
2958   if (CanVecMem) {
2959     OS.indent(Depth) << "Memory dependences are safe";
2960     const MemoryDepChecker &DC = getDepChecker();
2961     if (!DC.isSafeForAnyVectorWidth())
2962       OS << " with a maximum safe vector width of "
2963          << DC.getMaxSafeVectorWidthInBits() << " bits";
2964     if (PtrRtChecking->Need)
2965       OS << " with run-time checks";
2966     OS << "\n";
2967   }
2968 
2969   if (HasConvergentOp)
2970     OS.indent(Depth) << "Has convergent operation in loop\n";
2971 
2972   if (Report)
2973     OS.indent(Depth) << "Report: " << Report->getMsg() << "\n";
2974 
2975   if (auto *Dependences = DepChecker->getDependences()) {
2976     OS.indent(Depth) << "Dependences:\n";
2977     for (const auto &Dep : *Dependences) {
2978       Dep.print(OS, Depth + 2, DepChecker->getMemoryInstructions());
2979       OS << "\n";
2980     }
2981   } else
2982     OS.indent(Depth) << "Too many dependences, not recorded\n";
2983 
2984   // List the pair of accesses need run-time checks to prove independence.
2985   PtrRtChecking->print(OS, Depth);
2986   OS << "\n";
2987 
2988   OS.indent(Depth) << "Non vectorizable stores to invariant address were "
2989                    << (HasDependenceInvolvingLoopInvariantAddress ? "" : "not ")
2990                    << "found in loop.\n";
2991 
2992   OS.indent(Depth) << "SCEV assumptions:\n";
2993   PSE->getPredicate().print(OS, Depth);
2994 
2995   OS << "\n";
2996 
2997   OS.indent(Depth) << "Expressions re-written:\n";
2998   PSE->print(OS, Depth);
2999 }
3000 
3001 const LoopAccessInfo &LoopAccessInfoManager::getInfo(Loop &L) {
3002   auto I = LoopAccessInfoMap.insert({&L, nullptr});
3003 
3004   if (I.second)
3005     I.first->second =
3006         std::make_unique<LoopAccessInfo>(&L, &SE, TLI, &AA, &DT, &LI);
3007 
3008   return *I.first->second;
3009 }
3010 
3011 bool LoopAccessInfoManager::invalidate(
3012     Function &F, const PreservedAnalyses &PA,
3013     FunctionAnalysisManager::Invalidator &Inv) {
3014   // Check whether our analysis is preserved.
3015   auto PAC = PA.getChecker<LoopAccessAnalysis>();
3016   if (!PAC.preserved() && !PAC.preservedSet<AllAnalysesOn<Function>>())
3017     // If not, give up now.
3018     return true;
3019 
3020   // Check whether the analyses we depend on became invalid for any reason.
3021   // Skip checking TargetLibraryAnalysis as it is immutable and can't become
3022   // invalid.
3023   return Inv.invalidate<AAManager>(F, PA) ||
3024          Inv.invalidate<ScalarEvolutionAnalysis>(F, PA) ||
3025          Inv.invalidate<LoopAnalysis>(F, PA) ||
3026          Inv.invalidate<DominatorTreeAnalysis>(F, PA);
3027 }
3028 
3029 LoopAccessInfoManager LoopAccessAnalysis::run(Function &F,
3030                                               FunctionAnalysisManager &FAM) {
3031   auto &SE = FAM.getResult<ScalarEvolutionAnalysis>(F);
3032   auto &AA = FAM.getResult<AAManager>(F);
3033   auto &DT = FAM.getResult<DominatorTreeAnalysis>(F);
3034   auto &LI = FAM.getResult<LoopAnalysis>(F);
3035   auto &TLI = FAM.getResult<TargetLibraryAnalysis>(F);
3036   return LoopAccessInfoManager(SE, AA, DT, LI, &TLI);
3037 }
3038 
3039 AnalysisKey LoopAccessAnalysis::Key;
3040