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/DepthFirstIterator.h"
18 #include "llvm/ADT/EquivalenceClasses.h"
19 #include "llvm/ADT/PointerIntPair.h"
20 #include "llvm/ADT/STLExtras.h"
21 #include "llvm/ADT/SetVector.h"
22 #include "llvm/ADT/SmallPtrSet.h"
23 #include "llvm/ADT/SmallSet.h"
24 #include "llvm/ADT/SmallVector.h"
25 #include "llvm/ADT/iterator_range.h"
26 #include "llvm/Analysis/AliasAnalysis.h"
27 #include "llvm/Analysis/AliasSetTracker.h"
28 #include "llvm/Analysis/LoopAnalysisManager.h"
29 #include "llvm/Analysis/LoopInfo.h"
30 #include "llvm/Analysis/MemoryLocation.h"
31 #include "llvm/Analysis/OptimizationRemarkEmitter.h"
32 #include "llvm/Analysis/ScalarEvolution.h"
33 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
34 #include "llvm/Analysis/TargetLibraryInfo.h"
35 #include "llvm/Analysis/ValueTracking.h"
36 #include "llvm/Analysis/VectorUtils.h"
37 #include "llvm/IR/BasicBlock.h"
38 #include "llvm/IR/Constants.h"
39 #include "llvm/IR/DataLayout.h"
40 #include "llvm/IR/DebugLoc.h"
41 #include "llvm/IR/DerivedTypes.h"
42 #include "llvm/IR/DiagnosticInfo.h"
43 #include "llvm/IR/Dominators.h"
44 #include "llvm/IR/Function.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/Type.h"
51 #include "llvm/IR/Value.h"
52 #include "llvm/IR/ValueHandle.h"
53 #include "llvm/InitializePasses.h"
54 #include "llvm/Pass.h"
55 #include "llvm/Support/Casting.h"
56 #include "llvm/Support/CommandLine.h"
57 #include "llvm/Support/Debug.h"
58 #include "llvm/Support/ErrorHandling.h"
59 #include "llvm/Support/raw_ostream.h"
60 #include <algorithm>
61 #include <cassert>
62 #include <cstdint>
63 #include <cstdlib>
64 #include <iterator>
65 #include <utility>
66 #include <vector>
67 
68 using namespace llvm;
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 
isInterleaveForced()132 bool VectorizerParams::isInterleaveForced() {
133   return ::VectorizationInterleave.getNumOccurrences() > 0;
134 }
135 
stripIntegerCast(Value * V)136 Value *llvm::stripIntegerCast(Value *V) {
137   if (auto *CI = dyn_cast<CastInst>(V))
138     if (CI->getOperand(0)->getType()->isIntegerTy())
139       return CI->getOperand(0);
140   return V;
141 }
142 
replaceSymbolicStrideSCEV(PredicatedScalarEvolution & PSE,const ValueToValueMap & PtrToStride,Value * Ptr)143 const SCEV *llvm::replaceSymbolicStrideSCEV(PredicatedScalarEvolution &PSE,
144                                             const ValueToValueMap &PtrToStride,
145                                             Value *Ptr) {
146   const SCEV *OrigSCEV = PSE.getSCEV(Ptr);
147 
148   // If there is an entry in the map return the SCEV of the pointer with the
149   // symbolic stride replaced by one.
150   ValueToValueMap::const_iterator SI = PtrToStride.find(Ptr);
151   if (SI == PtrToStride.end())
152     // For a non-symbolic stride, just return the original expression.
153     return OrigSCEV;
154 
155   Value *StrideVal = stripIntegerCast(SI->second);
156 
157   ScalarEvolution *SE = PSE.getSE();
158   const auto *U = cast<SCEVUnknown>(SE->getSCEV(StrideVal));
159   const auto *CT =
160     static_cast<const SCEVConstant *>(SE->getOne(StrideVal->getType()));
161 
162   PSE.addPredicate(*SE->getEqualPredicate(U, CT));
163   auto *Expr = PSE.getSCEV(Ptr);
164 
165   LLVM_DEBUG(dbgs() << "LAA: Replacing SCEV: " << *OrigSCEV
166 	     << " by: " << *Expr << "\n");
167   return Expr;
168 }
169 
RuntimeCheckingPtrGroup(unsigned Index,RuntimePointerChecking & RtCheck)170 RuntimeCheckingPtrGroup::RuntimeCheckingPtrGroup(
171     unsigned Index, RuntimePointerChecking &RtCheck)
172     : High(RtCheck.Pointers[Index].End), Low(RtCheck.Pointers[Index].Start),
173       AddressSpace(RtCheck.Pointers[Index]
174                        .PointerValue->getType()
175                        ->getPointerAddressSpace()) {
176   Members.push_back(Index);
177 }
178 
179 /// Calculate Start and End points of memory access.
180 /// Let's assume A is the first access and B is a memory access on N-th loop
181 /// iteration. Then B is calculated as:
182 ///   B = A + Step*N .
183 /// Step value may be positive or negative.
184 /// N is a calculated back-edge taken count:
185 ///     N = (TripCount > 0) ? RoundDown(TripCount -1 , VF) : 0
186 /// Start and End points are calculated in the following way:
187 /// Start = UMIN(A, B) ; End = UMAX(A, B) + SizeOfElt,
188 /// where SizeOfElt is the size of single memory access in bytes.
189 ///
190 /// There is no conflict when the intervals are disjoint:
191 /// NoConflict = (P2.Start >= P1.End) || (P1.Start >= P2.End)
insert(Loop * Lp,Value * Ptr,bool WritePtr,unsigned DepSetId,unsigned ASId,const ValueToValueMap & Strides,PredicatedScalarEvolution & PSE)192 void RuntimePointerChecking::insert(Loop *Lp, Value *Ptr, bool WritePtr,
193                                     unsigned DepSetId, unsigned ASId,
194                                     const ValueToValueMap &Strides,
195                                     PredicatedScalarEvolution &PSE) {
196   // Get the stride replaced scev.
197   const SCEV *Sc = replaceSymbolicStrideSCEV(PSE, Strides, Ptr);
198   ScalarEvolution *SE = PSE.getSE();
199 
200   const SCEV *ScStart;
201   const SCEV *ScEnd;
202 
203   if (SE->isLoopInvariant(Sc, Lp)) {
204     ScStart = ScEnd = Sc;
205   } else {
206     const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Sc);
207     assert(AR && "Invalid addrec expression");
208     const SCEV *Ex = PSE.getBackedgeTakenCount();
209 
210     ScStart = AR->getStart();
211     ScEnd = AR->evaluateAtIteration(Ex, *SE);
212     const SCEV *Step = AR->getStepRecurrence(*SE);
213 
214     // For expressions with negative step, the upper bound is ScStart and the
215     // lower bound is ScEnd.
216     if (const auto *CStep = dyn_cast<SCEVConstant>(Step)) {
217       if (CStep->getValue()->isNegative())
218         std::swap(ScStart, ScEnd);
219     } else {
220       // Fallback case: the step is not constant, but we can still
221       // get the upper and lower bounds of the interval by using min/max
222       // expressions.
223       ScStart = SE->getUMinExpr(ScStart, ScEnd);
224       ScEnd = SE->getUMaxExpr(AR->getStart(), ScEnd);
225     }
226   }
227   // Add the size of the pointed element to ScEnd.
228   auto &DL = Lp->getHeader()->getModule()->getDataLayout();
229   Type *IdxTy = DL.getIndexType(Ptr->getType());
230   const SCEV *EltSizeSCEV =
231       SE->getStoreSizeOfExpr(IdxTy, Ptr->getType()->getPointerElementType());
232   ScEnd = SE->getAddExpr(ScEnd, EltSizeSCEV);
233 
234   Pointers.emplace_back(Ptr, ScStart, ScEnd, WritePtr, DepSetId, ASId, Sc);
235 }
236 
237 SmallVector<RuntimePointerCheck, 4>
generateChecks() const238 RuntimePointerChecking::generateChecks() const {
239   SmallVector<RuntimePointerCheck, 4> Checks;
240 
241   for (unsigned I = 0; I < CheckingGroups.size(); ++I) {
242     for (unsigned J = I + 1; J < CheckingGroups.size(); ++J) {
243       const RuntimeCheckingPtrGroup &CGI = CheckingGroups[I];
244       const RuntimeCheckingPtrGroup &CGJ = CheckingGroups[J];
245 
246       if (needsChecking(CGI, CGJ))
247         Checks.push_back(std::make_pair(&CGI, &CGJ));
248     }
249   }
250   return Checks;
251 }
252 
generateChecks(MemoryDepChecker::DepCandidates & DepCands,bool UseDependencies)253 void RuntimePointerChecking::generateChecks(
254     MemoryDepChecker::DepCandidates &DepCands, bool UseDependencies) {
255   assert(Checks.empty() && "Checks is not empty");
256   groupChecks(DepCands, UseDependencies);
257   Checks = generateChecks();
258 }
259 
needsChecking(const RuntimeCheckingPtrGroup & M,const RuntimeCheckingPtrGroup & N) const260 bool RuntimePointerChecking::needsChecking(
261     const RuntimeCheckingPtrGroup &M, const RuntimeCheckingPtrGroup &N) const {
262   for (unsigned I = 0, EI = M.Members.size(); EI != I; ++I)
263     for (unsigned J = 0, EJ = N.Members.size(); EJ != J; ++J)
264       if (needsChecking(M.Members[I], N.Members[J]))
265         return true;
266   return false;
267 }
268 
269 /// Compare \p I and \p J and return the minimum.
270 /// Return nullptr in case we couldn't find an answer.
getMinFromExprs(const SCEV * I,const SCEV * J,ScalarEvolution * SE)271 static const SCEV *getMinFromExprs(const SCEV *I, const SCEV *J,
272                                    ScalarEvolution *SE) {
273   const SCEV *Diff = SE->getMinusSCEV(J, I);
274   const SCEVConstant *C = dyn_cast<const SCEVConstant>(Diff);
275 
276   if (!C)
277     return nullptr;
278   if (C->getValue()->isNegative())
279     return J;
280   return I;
281 }
282 
addPointer(unsigned Index,RuntimePointerChecking & RtCheck)283 bool RuntimeCheckingPtrGroup::addPointer(unsigned Index,
284                                          RuntimePointerChecking &RtCheck) {
285   return addPointer(
286       Index, RtCheck.Pointers[Index].Start, RtCheck.Pointers[Index].End,
287       RtCheck.Pointers[Index].PointerValue->getType()->getPointerAddressSpace(),
288       *RtCheck.SE);
289 }
290 
addPointer(unsigned Index,const SCEV * Start,const SCEV * End,unsigned AS,ScalarEvolution & SE)291 bool RuntimeCheckingPtrGroup::addPointer(unsigned Index, const SCEV *Start,
292                                          const SCEV *End, unsigned AS,
293                                          ScalarEvolution &SE) {
294   assert(AddressSpace == AS &&
295          "all pointers in a checking group must be in the same address space");
296 
297   // Compare the starts and ends with the known minimum and maximum
298   // of this set. We need to know how we compare against the min/max
299   // of the set in order to be able to emit memchecks.
300   const SCEV *Min0 = getMinFromExprs(Start, Low, &SE);
301   if (!Min0)
302     return false;
303 
304   const SCEV *Min1 = getMinFromExprs(End, High, &SE);
305   if (!Min1)
306     return false;
307 
308   // Update the low bound  expression if we've found a new min value.
309   if (Min0 == Start)
310     Low = Start;
311 
312   // Update the high bound expression if we've found a new max value.
313   if (Min1 != End)
314     High = End;
315 
316   Members.push_back(Index);
317   return true;
318 }
319 
groupChecks(MemoryDepChecker::DepCandidates & DepCands,bool UseDependencies)320 void RuntimePointerChecking::groupChecks(
321     MemoryDepChecker::DepCandidates &DepCands, bool UseDependencies) {
322   // We build the groups from dependency candidates equivalence classes
323   // because:
324   //    - We know that pointers in the same equivalence class share
325   //      the same underlying object and therefore there is a chance
326   //      that we can compare pointers
327   //    - We wouldn't be able to merge two pointers for which we need
328   //      to emit a memcheck. The classes in DepCands are already
329   //      conveniently built such that no two pointers in the same
330   //      class need checking against each other.
331 
332   // We use the following (greedy) algorithm to construct the groups
333   // For every pointer in the equivalence class:
334   //   For each existing group:
335   //   - if the difference between this pointer and the min/max bounds
336   //     of the group is a constant, then make the pointer part of the
337   //     group and update the min/max bounds of that group as required.
338 
339   CheckingGroups.clear();
340 
341   // If we need to check two pointers to the same underlying object
342   // with a non-constant difference, we shouldn't perform any pointer
343   // grouping with those pointers. This is because we can easily get
344   // into cases where the resulting check would return false, even when
345   // the accesses are safe.
346   //
347   // The following example shows this:
348   // for (i = 0; i < 1000; ++i)
349   //   a[5000 + i * m] = a[i] + a[i + 9000]
350   //
351   // Here grouping gives a check of (5000, 5000 + 1000 * m) against
352   // (0, 10000) which is always false. However, if m is 1, there is no
353   // dependence. Not grouping the checks for a[i] and a[i + 9000] allows
354   // us to perform an accurate check in this case.
355   //
356   // The above case requires that we have an UnknownDependence between
357   // accesses to the same underlying object. This cannot happen unless
358   // FoundNonConstantDistanceDependence is set, and therefore UseDependencies
359   // is also false. In this case we will use the fallback path and create
360   // separate checking groups for all pointers.
361 
362   // If we don't have the dependency partitions, construct a new
363   // checking pointer group for each pointer. This is also required
364   // for correctness, because in this case we can have checking between
365   // pointers to the same underlying object.
366   if (!UseDependencies) {
367     for (unsigned I = 0; I < Pointers.size(); ++I)
368       CheckingGroups.push_back(RuntimeCheckingPtrGroup(I, *this));
369     return;
370   }
371 
372   unsigned TotalComparisons = 0;
373 
374   DenseMap<Value *, unsigned> PositionMap;
375   for (unsigned Index = 0; Index < Pointers.size(); ++Index)
376     PositionMap[Pointers[Index].PointerValue] = Index;
377 
378   // We need to keep track of what pointers we've already seen so we
379   // don't process them twice.
380   SmallSet<unsigned, 2> Seen;
381 
382   // Go through all equivalence classes, get the "pointer check groups"
383   // and add them to the overall solution. We use the order in which accesses
384   // appear in 'Pointers' to enforce determinism.
385   for (unsigned I = 0; I < Pointers.size(); ++I) {
386     // We've seen this pointer before, and therefore already processed
387     // its equivalence class.
388     if (Seen.count(I))
389       continue;
390 
391     MemoryDepChecker::MemAccessInfo Access(Pointers[I].PointerValue,
392                                            Pointers[I].IsWritePtr);
393 
394     SmallVector<RuntimeCheckingPtrGroup, 2> Groups;
395     auto LeaderI = DepCands.findValue(DepCands.getLeaderValue(Access));
396 
397     // Because DepCands is constructed by visiting accesses in the order in
398     // which they appear in alias sets (which is deterministic) and the
399     // iteration order within an equivalence class member is only dependent on
400     // the order in which unions and insertions are performed on the
401     // equivalence class, the iteration order is deterministic.
402     for (auto MI = DepCands.member_begin(LeaderI), ME = DepCands.member_end();
403          MI != ME; ++MI) {
404       auto PointerI = PositionMap.find(MI->getPointer());
405       assert(PointerI != PositionMap.end() &&
406              "pointer in equivalence class not found in PositionMap");
407       unsigned Pointer = PointerI->second;
408       bool Merged = false;
409       // Mark this pointer as seen.
410       Seen.insert(Pointer);
411 
412       // Go through all the existing sets and see if we can find one
413       // which can include this pointer.
414       for (RuntimeCheckingPtrGroup &Group : Groups) {
415         // Don't perform more than a certain amount of comparisons.
416         // This should limit the cost of grouping the pointers to something
417         // reasonable.  If we do end up hitting this threshold, the algorithm
418         // will create separate groups for all remaining pointers.
419         if (TotalComparisons > MemoryCheckMergeThreshold)
420           break;
421 
422         TotalComparisons++;
423 
424         if (Group.addPointer(Pointer, *this)) {
425           Merged = true;
426           break;
427         }
428       }
429 
430       if (!Merged)
431         // We couldn't add this pointer to any existing set or the threshold
432         // for the number of comparisons has been reached. Create a new group
433         // to hold the current pointer.
434         Groups.push_back(RuntimeCheckingPtrGroup(Pointer, *this));
435     }
436 
437     // We've computed the grouped checks for this partition.
438     // Save the results and continue with the next one.
439     llvm::copy(Groups, std::back_inserter(CheckingGroups));
440   }
441 }
442 
arePointersInSamePartition(const SmallVectorImpl<int> & PtrToPartition,unsigned PtrIdx1,unsigned PtrIdx2)443 bool RuntimePointerChecking::arePointersInSamePartition(
444     const SmallVectorImpl<int> &PtrToPartition, unsigned PtrIdx1,
445     unsigned PtrIdx2) {
446   return (PtrToPartition[PtrIdx1] != -1 &&
447           PtrToPartition[PtrIdx1] == PtrToPartition[PtrIdx2]);
448 }
449 
needsChecking(unsigned I,unsigned J) const450 bool RuntimePointerChecking::needsChecking(unsigned I, unsigned J) const {
451   const PointerInfo &PointerI = Pointers[I];
452   const PointerInfo &PointerJ = Pointers[J];
453 
454   // No need to check if two readonly pointers intersect.
455   if (!PointerI.IsWritePtr && !PointerJ.IsWritePtr)
456     return false;
457 
458   // Only need to check pointers between two different dependency sets.
459   if (PointerI.DependencySetId == PointerJ.DependencySetId)
460     return false;
461 
462   // Only need to check pointers in the same alias set.
463   if (PointerI.AliasSetId != PointerJ.AliasSetId)
464     return false;
465 
466   return true;
467 }
468 
printChecks(raw_ostream & OS,const SmallVectorImpl<RuntimePointerCheck> & Checks,unsigned Depth) const469 void RuntimePointerChecking::printChecks(
470     raw_ostream &OS, const SmallVectorImpl<RuntimePointerCheck> &Checks,
471     unsigned Depth) const {
472   unsigned N = 0;
473   for (const auto &Check : Checks) {
474     const auto &First = Check.first->Members, &Second = Check.second->Members;
475 
476     OS.indent(Depth) << "Check " << N++ << ":\n";
477 
478     OS.indent(Depth + 2) << "Comparing group (" << Check.first << "):\n";
479     for (unsigned K = 0; K < First.size(); ++K)
480       OS.indent(Depth + 2) << *Pointers[First[K]].PointerValue << "\n";
481 
482     OS.indent(Depth + 2) << "Against group (" << Check.second << "):\n";
483     for (unsigned K = 0; K < Second.size(); ++K)
484       OS.indent(Depth + 2) << *Pointers[Second[K]].PointerValue << "\n";
485   }
486 }
487 
print(raw_ostream & OS,unsigned Depth) const488 void RuntimePointerChecking::print(raw_ostream &OS, unsigned Depth) const {
489 
490   OS.indent(Depth) << "Run-time memory checks:\n";
491   printChecks(OS, Checks, Depth);
492 
493   OS.indent(Depth) << "Grouped accesses:\n";
494   for (unsigned I = 0; I < CheckingGroups.size(); ++I) {
495     const auto &CG = CheckingGroups[I];
496 
497     OS.indent(Depth + 2) << "Group " << &CG << ":\n";
498     OS.indent(Depth + 4) << "(Low: " << *CG.Low << " High: " << *CG.High
499                          << ")\n";
500     for (unsigned J = 0; J < CG.Members.size(); ++J) {
501       OS.indent(Depth + 6) << "Member: " << *Pointers[CG.Members[J]].Expr
502                            << "\n";
503     }
504   }
505 }
506 
507 namespace {
508 
509 /// Analyses memory accesses in a loop.
510 ///
511 /// Checks whether run time pointer checks are needed and builds sets for data
512 /// dependence checking.
513 class AccessAnalysis {
514 public:
515   /// Read or write access location.
516   typedef PointerIntPair<Value *, 1, bool> MemAccessInfo;
517   typedef SmallVector<MemAccessInfo, 8> MemAccessInfoList;
518 
AccessAnalysis(Loop * TheLoop,AAResults * AA,LoopInfo * LI,MemoryDepChecker::DepCandidates & DA,PredicatedScalarEvolution & PSE)519   AccessAnalysis(Loop *TheLoop, AAResults *AA, LoopInfo *LI,
520                  MemoryDepChecker::DepCandidates &DA,
521                  PredicatedScalarEvolution &PSE)
522       : TheLoop(TheLoop), AST(*AA), LI(LI), DepCands(DA),
523         IsRTCheckAnalysisNeeded(false), PSE(PSE) {}
524 
525   /// Register a load  and whether it is only read from.
addLoad(MemoryLocation & Loc,bool IsReadOnly)526   void addLoad(MemoryLocation &Loc, bool IsReadOnly) {
527     Value *Ptr = const_cast<Value*>(Loc.Ptr);
528     AST.add(Ptr, LocationSize::beforeOrAfterPointer(), Loc.AATags);
529     Accesses.insert(MemAccessInfo(Ptr, false));
530     if (IsReadOnly)
531       ReadOnlyPtr.insert(Ptr);
532   }
533 
534   /// Register a store.
addStore(MemoryLocation & Loc)535   void addStore(MemoryLocation &Loc) {
536     Value *Ptr = const_cast<Value*>(Loc.Ptr);
537     AST.add(Ptr, LocationSize::beforeOrAfterPointer(), Loc.AATags);
538     Accesses.insert(MemAccessInfo(Ptr, true));
539   }
540 
541   /// Check if we can emit a run-time no-alias check for \p Access.
542   ///
543   /// Returns true if we can emit a run-time no alias check for \p Access.
544   /// If we can check this access, this also adds it to a dependence set and
545   /// adds a run-time to check for it to \p RtCheck. If \p Assume is true,
546   /// we will attempt to use additional run-time checks in order to get
547   /// the bounds of the pointer.
548   bool createCheckForAccess(RuntimePointerChecking &RtCheck,
549                             MemAccessInfo Access,
550                             const ValueToValueMap &Strides,
551                             DenseMap<Value *, unsigned> &DepSetId,
552                             Loop *TheLoop, unsigned &RunningDepId,
553                             unsigned ASId, bool ShouldCheckStride,
554                             bool Assume);
555 
556   /// Check whether we can check the pointers at runtime for
557   /// non-intersection.
558   ///
559   /// Returns true if we need no check or if we do and we can generate them
560   /// (i.e. the pointers have computable bounds).
561   bool canCheckPtrAtRT(RuntimePointerChecking &RtCheck, ScalarEvolution *SE,
562                        Loop *TheLoop, const ValueToValueMap &Strides,
563                        bool ShouldCheckWrap = false);
564 
565   /// Goes over all memory accesses, checks whether a RT check is needed
566   /// and builds sets of dependent accesses.
buildDependenceSets()567   void buildDependenceSets() {
568     processMemAccesses();
569   }
570 
571   /// Initial processing of memory accesses determined that we need to
572   /// perform dependency checking.
573   ///
574   /// Note that this can later be cleared if we retry memcheck analysis without
575   /// dependency checking (i.e. FoundNonConstantDistanceDependence).
isDependencyCheckNeeded()576   bool isDependencyCheckNeeded() { return !CheckDeps.empty(); }
577 
578   /// We decided that no dependence analysis would be used.  Reset the state.
resetDepChecks(MemoryDepChecker & DepChecker)579   void resetDepChecks(MemoryDepChecker &DepChecker) {
580     CheckDeps.clear();
581     DepChecker.clearDependences();
582   }
583 
getDependenciesToCheck()584   MemAccessInfoList &getDependenciesToCheck() { return CheckDeps; }
585 
586 private:
587   typedef SetVector<MemAccessInfo> PtrAccessSet;
588 
589   /// Go over all memory access and check whether runtime pointer checks
590   /// are needed and build sets of dependency check candidates.
591   void processMemAccesses();
592 
593   /// Set of all accesses.
594   PtrAccessSet Accesses;
595 
596   /// The loop being checked.
597   const Loop *TheLoop;
598 
599   /// List of accesses that need a further dependence check.
600   MemAccessInfoList CheckDeps;
601 
602   /// Set of pointers that are read only.
603   SmallPtrSet<Value*, 16> ReadOnlyPtr;
604 
605   /// An alias set tracker to partition the access set by underlying object and
606   //intrinsic property (such as TBAA metadata).
607   AliasSetTracker AST;
608 
609   LoopInfo *LI;
610 
611   /// Sets of potentially dependent accesses - members of one set share an
612   /// underlying pointer. The set "CheckDeps" identfies which sets really need a
613   /// dependence check.
614   MemoryDepChecker::DepCandidates &DepCands;
615 
616   /// Initial processing of memory accesses determined that we may need
617   /// to add memchecks.  Perform the analysis to determine the necessary checks.
618   ///
619   /// Note that, this is different from isDependencyCheckNeeded.  When we retry
620   /// memcheck analysis without dependency checking
621   /// (i.e. FoundNonConstantDistanceDependence), isDependencyCheckNeeded is
622   /// cleared while this remains set if we have potentially dependent accesses.
623   bool IsRTCheckAnalysisNeeded;
624 
625   /// The SCEV predicate containing all the SCEV-related assumptions.
626   PredicatedScalarEvolution &PSE;
627 };
628 
629 } // end anonymous namespace
630 
631 /// Check whether a pointer can participate in a runtime bounds check.
632 /// If \p Assume, try harder to prove that we can compute the bounds of \p Ptr
633 /// by adding run-time checks (overflow checks) if necessary.
hasComputableBounds(PredicatedScalarEvolution & PSE,const ValueToValueMap & Strides,Value * Ptr,Loop * L,bool Assume)634 static bool hasComputableBounds(PredicatedScalarEvolution &PSE,
635                                 const ValueToValueMap &Strides, Value *Ptr,
636                                 Loop *L, bool Assume) {
637   const SCEV *PtrScev = replaceSymbolicStrideSCEV(PSE, Strides, Ptr);
638 
639   // The bounds for loop-invariant pointer is trivial.
640   if (PSE.getSE()->isLoopInvariant(PtrScev, L))
641     return true;
642 
643   const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
644 
645   if (!AR && Assume)
646     AR = PSE.getAsAddRec(Ptr);
647 
648   if (!AR)
649     return false;
650 
651   return AR->isAffine();
652 }
653 
654 /// Check whether a pointer address cannot wrap.
isNoWrap(PredicatedScalarEvolution & PSE,const ValueToValueMap & Strides,Value * Ptr,Loop * L)655 static bool isNoWrap(PredicatedScalarEvolution &PSE,
656                      const ValueToValueMap &Strides, Value *Ptr, Loop *L) {
657   const SCEV *PtrScev = PSE.getSCEV(Ptr);
658   if (PSE.getSE()->isLoopInvariant(PtrScev, L))
659     return true;
660 
661   Type *AccessTy = Ptr->getType()->getPointerElementType();
662   int64_t Stride = getPtrStride(PSE, AccessTy, Ptr, L, Strides);
663   if (Stride == 1 || PSE.hasNoOverflow(Ptr, SCEVWrapPredicate::IncrementNUSW))
664     return true;
665 
666   return false;
667 }
668 
createCheckForAccess(RuntimePointerChecking & RtCheck,MemAccessInfo Access,const ValueToValueMap & StridesMap,DenseMap<Value *,unsigned> & DepSetId,Loop * TheLoop,unsigned & RunningDepId,unsigned ASId,bool ShouldCheckWrap,bool Assume)669 bool AccessAnalysis::createCheckForAccess(RuntimePointerChecking &RtCheck,
670                                           MemAccessInfo Access,
671                                           const ValueToValueMap &StridesMap,
672                                           DenseMap<Value *, unsigned> &DepSetId,
673                                           Loop *TheLoop, unsigned &RunningDepId,
674                                           unsigned ASId, bool ShouldCheckWrap,
675                                           bool Assume) {
676   Value *Ptr = Access.getPointer();
677 
678   if (!hasComputableBounds(PSE, StridesMap, Ptr, TheLoop, Assume))
679     return false;
680 
681   // When we run after a failing dependency check we have to make sure
682   // we don't have wrapping pointers.
683   if (ShouldCheckWrap && !isNoWrap(PSE, StridesMap, Ptr, TheLoop)) {
684     auto *Expr = PSE.getSCEV(Ptr);
685     if (!Assume || !isa<SCEVAddRecExpr>(Expr))
686       return false;
687     PSE.setNoOverflow(Ptr, SCEVWrapPredicate::IncrementNUSW);
688   }
689 
690   // The id of the dependence set.
691   unsigned DepId;
692 
693   if (isDependencyCheckNeeded()) {
694     Value *Leader = DepCands.getLeaderValue(Access).getPointer();
695     unsigned &LeaderId = DepSetId[Leader];
696     if (!LeaderId)
697       LeaderId = RunningDepId++;
698     DepId = LeaderId;
699   } else
700     // Each access has its own dependence set.
701     DepId = RunningDepId++;
702 
703   bool IsWrite = Access.getInt();
704   RtCheck.insert(TheLoop, Ptr, IsWrite, DepId, ASId, StridesMap, PSE);
705   LLVM_DEBUG(dbgs() << "LAA: Found a runtime check ptr:" << *Ptr << '\n');
706 
707   return true;
708  }
709 
canCheckPtrAtRT(RuntimePointerChecking & RtCheck,ScalarEvolution * SE,Loop * TheLoop,const ValueToValueMap & StridesMap,bool ShouldCheckWrap)710 bool AccessAnalysis::canCheckPtrAtRT(RuntimePointerChecking &RtCheck,
711                                      ScalarEvolution *SE, Loop *TheLoop,
712                                      const ValueToValueMap &StridesMap,
713                                      bool ShouldCheckWrap) {
714   // Find pointers with computable bounds. We are going to use this information
715   // to place a runtime bound check.
716   bool CanDoRT = true;
717 
718   bool MayNeedRTCheck = false;
719   if (!IsRTCheckAnalysisNeeded) return true;
720 
721   bool IsDepCheckNeeded = isDependencyCheckNeeded();
722 
723   // We assign a consecutive id to access from different alias sets.
724   // Accesses between different groups doesn't need to be checked.
725   unsigned ASId = 0;
726   for (auto &AS : AST) {
727     int NumReadPtrChecks = 0;
728     int NumWritePtrChecks = 0;
729     bool CanDoAliasSetRT = true;
730     ++ASId;
731 
732     // We assign consecutive id to access from different dependence sets.
733     // Accesses within the same set don't need a runtime check.
734     unsigned RunningDepId = 1;
735     DenseMap<Value *, unsigned> DepSetId;
736 
737     SmallVector<MemAccessInfo, 4> Retries;
738 
739     // First, count how many write and read accesses are in the alias set. Also
740     // collect MemAccessInfos for later.
741     SmallVector<MemAccessInfo, 4> AccessInfos;
742     for (const auto &A : AS) {
743       Value *Ptr = A.getValue();
744       bool IsWrite = Accesses.count(MemAccessInfo(Ptr, true));
745 
746       if (IsWrite)
747         ++NumWritePtrChecks;
748       else
749         ++NumReadPtrChecks;
750       AccessInfos.emplace_back(Ptr, IsWrite);
751     }
752 
753     // We do not need runtime checks for this alias set, if there are no writes
754     // or a single write and no reads.
755     if (NumWritePtrChecks == 0 ||
756         (NumWritePtrChecks == 1 && NumReadPtrChecks == 0)) {
757       assert((AS.size() <= 1 ||
758               all_of(AS,
759                      [this](auto AC) {
760                        MemAccessInfo AccessWrite(AC.getValue(), true);
761                        return DepCands.findValue(AccessWrite) == DepCands.end();
762                      })) &&
763              "Can only skip updating CanDoRT below, if all entries in AS "
764              "are reads or there is at most 1 entry");
765       continue;
766     }
767 
768     for (auto &Access : AccessInfos) {
769       if (!createCheckForAccess(RtCheck, Access, StridesMap, DepSetId, TheLoop,
770                                 RunningDepId, ASId, ShouldCheckWrap, false)) {
771         LLVM_DEBUG(dbgs() << "LAA: Can't find bounds for ptr:"
772                           << *Access.getPointer() << '\n');
773         Retries.push_back(Access);
774         CanDoAliasSetRT = false;
775       }
776     }
777 
778     // Note that this function computes CanDoRT and MayNeedRTCheck
779     // independently. For example CanDoRT=false, MayNeedRTCheck=false means that
780     // we have a pointer for which we couldn't find the bounds but we don't
781     // actually need to emit any checks so it does not matter.
782     //
783     // We need runtime checks for this alias set, if there are at least 2
784     // dependence sets (in which case RunningDepId > 2) or if we need to re-try
785     // any bound checks (because in that case the number of dependence sets is
786     // incomplete).
787     bool NeedsAliasSetRTCheck = RunningDepId > 2 || !Retries.empty();
788 
789     // We need to perform run-time alias checks, but some pointers had bounds
790     // that couldn't be checked.
791     if (NeedsAliasSetRTCheck && !CanDoAliasSetRT) {
792       // Reset the CanDoSetRt flag and retry all accesses that have failed.
793       // We know that we need these checks, so we can now be more aggressive
794       // and add further checks if required (overflow checks).
795       CanDoAliasSetRT = true;
796       for (auto Access : Retries)
797         if (!createCheckForAccess(RtCheck, Access, StridesMap, DepSetId,
798                                   TheLoop, RunningDepId, ASId,
799                                   ShouldCheckWrap, /*Assume=*/true)) {
800           CanDoAliasSetRT = false;
801           break;
802         }
803     }
804 
805     CanDoRT &= CanDoAliasSetRT;
806     MayNeedRTCheck |= NeedsAliasSetRTCheck;
807     ++ASId;
808   }
809 
810   // If the pointers that we would use for the bounds comparison have different
811   // address spaces, assume the values aren't directly comparable, so we can't
812   // use them for the runtime check. We also have to assume they could
813   // overlap. In the future there should be metadata for whether address spaces
814   // are disjoint.
815   unsigned NumPointers = RtCheck.Pointers.size();
816   for (unsigned i = 0; i < NumPointers; ++i) {
817     for (unsigned j = i + 1; j < NumPointers; ++j) {
818       // Only need to check pointers between two different dependency sets.
819       if (RtCheck.Pointers[i].DependencySetId ==
820           RtCheck.Pointers[j].DependencySetId)
821        continue;
822       // Only need to check pointers in the same alias set.
823       if (RtCheck.Pointers[i].AliasSetId != RtCheck.Pointers[j].AliasSetId)
824         continue;
825 
826       Value *PtrI = RtCheck.Pointers[i].PointerValue;
827       Value *PtrJ = RtCheck.Pointers[j].PointerValue;
828 
829       unsigned ASi = PtrI->getType()->getPointerAddressSpace();
830       unsigned ASj = PtrJ->getType()->getPointerAddressSpace();
831       if (ASi != ASj) {
832         LLVM_DEBUG(
833             dbgs() << "LAA: Runtime check would require comparison between"
834                       " different address spaces\n");
835         return false;
836       }
837     }
838   }
839 
840   if (MayNeedRTCheck && CanDoRT)
841     RtCheck.generateChecks(DepCands, IsDepCheckNeeded);
842 
843   LLVM_DEBUG(dbgs() << "LAA: We need to do " << RtCheck.getNumberOfChecks()
844                     << " pointer comparisons.\n");
845 
846   // If we can do run-time checks, but there are no checks, no runtime checks
847   // are needed. This can happen when all pointers point to the same underlying
848   // object for example.
849   RtCheck.Need = CanDoRT ? RtCheck.getNumberOfChecks() != 0 : MayNeedRTCheck;
850 
851   bool CanDoRTIfNeeded = !RtCheck.Need || CanDoRT;
852   if (!CanDoRTIfNeeded)
853     RtCheck.reset();
854   return CanDoRTIfNeeded;
855 }
856 
processMemAccesses()857 void AccessAnalysis::processMemAccesses() {
858   // We process the set twice: first we process read-write pointers, last we
859   // process read-only pointers. This allows us to skip dependence tests for
860   // read-only pointers.
861 
862   LLVM_DEBUG(dbgs() << "LAA: Processing memory accesses...\n");
863   LLVM_DEBUG(dbgs() << "  AST: "; AST.dump());
864   LLVM_DEBUG(dbgs() << "LAA:   Accesses(" << Accesses.size() << "):\n");
865   LLVM_DEBUG({
866     for (auto A : Accesses)
867       dbgs() << "\t" << *A.getPointer() << " (" <<
868                 (A.getInt() ? "write" : (ReadOnlyPtr.count(A.getPointer()) ?
869                                          "read-only" : "read")) << ")\n";
870   });
871 
872   // The AliasSetTracker has nicely partitioned our pointers by metadata
873   // compatibility and potential for underlying-object overlap. As a result, we
874   // only need to check for potential pointer dependencies within each alias
875   // set.
876   for (const auto &AS : AST) {
877     // Note that both the alias-set tracker and the alias sets themselves used
878     // linked lists internally and so the iteration order here is deterministic
879     // (matching the original instruction order within each set).
880 
881     bool SetHasWrite = false;
882 
883     // Map of pointers to last access encountered.
884     typedef DenseMap<const Value*, MemAccessInfo> UnderlyingObjToAccessMap;
885     UnderlyingObjToAccessMap ObjToLastAccess;
886 
887     // Set of access to check after all writes have been processed.
888     PtrAccessSet DeferredAccesses;
889 
890     // Iterate over each alias set twice, once to process read/write pointers,
891     // and then to process read-only pointers.
892     for (int SetIteration = 0; SetIteration < 2; ++SetIteration) {
893       bool UseDeferred = SetIteration > 0;
894       PtrAccessSet &S = UseDeferred ? DeferredAccesses : Accesses;
895 
896       for (const auto &AV : AS) {
897         Value *Ptr = AV.getValue();
898 
899         // For a single memory access in AliasSetTracker, Accesses may contain
900         // both read and write, and they both need to be handled for CheckDeps.
901         for (const auto &AC : S) {
902           if (AC.getPointer() != Ptr)
903             continue;
904 
905           bool IsWrite = AC.getInt();
906 
907           // If we're using the deferred access set, then it contains only
908           // reads.
909           bool IsReadOnlyPtr = ReadOnlyPtr.count(Ptr) && !IsWrite;
910           if (UseDeferred && !IsReadOnlyPtr)
911             continue;
912           // Otherwise, the pointer must be in the PtrAccessSet, either as a
913           // read or a write.
914           assert(((IsReadOnlyPtr && UseDeferred) || IsWrite ||
915                   S.count(MemAccessInfo(Ptr, false))) &&
916                  "Alias-set pointer not in the access set?");
917 
918           MemAccessInfo Access(Ptr, IsWrite);
919           DepCands.insert(Access);
920 
921           // Memorize read-only pointers for later processing and skip them in
922           // the first round (they need to be checked after we have seen all
923           // write pointers). Note: we also mark pointer that are not
924           // consecutive as "read-only" pointers (so that we check
925           // "a[b[i]] +="). Hence, we need the second check for "!IsWrite".
926           if (!UseDeferred && IsReadOnlyPtr) {
927             DeferredAccesses.insert(Access);
928             continue;
929           }
930 
931           // If this is a write - check other reads and writes for conflicts. If
932           // this is a read only check other writes for conflicts (but only if
933           // there is no other write to the ptr - this is an optimization to
934           // catch "a[i] = a[i] + " without having to do a dependence check).
935           if ((IsWrite || IsReadOnlyPtr) && SetHasWrite) {
936             CheckDeps.push_back(Access);
937             IsRTCheckAnalysisNeeded = true;
938           }
939 
940           if (IsWrite)
941             SetHasWrite = true;
942 
943           // Create sets of pointers connected by a shared alias set and
944           // underlying object.
945           typedef SmallVector<const Value *, 16> ValueVector;
946           ValueVector TempObjects;
947 
948           getUnderlyingObjects(Ptr, TempObjects, LI);
949           LLVM_DEBUG(dbgs()
950                      << "Underlying objects for pointer " << *Ptr << "\n");
951           for (const Value *UnderlyingObj : TempObjects) {
952             // nullptr never alias, don't join sets for pointer that have "null"
953             // in their UnderlyingObjects list.
954             if (isa<ConstantPointerNull>(UnderlyingObj) &&
955                 !NullPointerIsDefined(
956                     TheLoop->getHeader()->getParent(),
957                     UnderlyingObj->getType()->getPointerAddressSpace()))
958               continue;
959 
960             UnderlyingObjToAccessMap::iterator Prev =
961                 ObjToLastAccess.find(UnderlyingObj);
962             if (Prev != ObjToLastAccess.end())
963               DepCands.unionSets(Access, Prev->second);
964 
965             ObjToLastAccess[UnderlyingObj] = Access;
966             LLVM_DEBUG(dbgs() << "  " << *UnderlyingObj << "\n");
967           }
968         }
969       }
970     }
971   }
972 }
973 
isInBoundsGep(Value * Ptr)974 static bool isInBoundsGep(Value *Ptr) {
975   if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr))
976     return GEP->isInBounds();
977   return false;
978 }
979 
980 /// Return true if an AddRec pointer \p Ptr is unsigned non-wrapping,
981 /// i.e. monotonically increasing/decreasing.
isNoWrapAddRec(Value * Ptr,const SCEVAddRecExpr * AR,PredicatedScalarEvolution & PSE,const Loop * L)982 static bool isNoWrapAddRec(Value *Ptr, const SCEVAddRecExpr *AR,
983                            PredicatedScalarEvolution &PSE, const Loop *L) {
984   // FIXME: This should probably only return true for NUW.
985   if (AR->getNoWrapFlags(SCEV::NoWrapMask))
986     return true;
987 
988   // Scalar evolution does not propagate the non-wrapping flags to values that
989   // are derived from a non-wrapping induction variable because non-wrapping
990   // could be flow-sensitive.
991   //
992   // Look through the potentially overflowing instruction to try to prove
993   // non-wrapping for the *specific* value of Ptr.
994 
995   // The arithmetic implied by an inbounds GEP can't overflow.
996   auto *GEP = dyn_cast<GetElementPtrInst>(Ptr);
997   if (!GEP || !GEP->isInBounds())
998     return false;
999 
1000   // Make sure there is only one non-const index and analyze that.
1001   Value *NonConstIndex = nullptr;
1002   for (Value *Index : GEP->indices())
1003     if (!isa<ConstantInt>(Index)) {
1004       if (NonConstIndex)
1005         return false;
1006       NonConstIndex = Index;
1007     }
1008   if (!NonConstIndex)
1009     // The recurrence is on the pointer, ignore for now.
1010     return false;
1011 
1012   // The index in GEP is signed.  It is non-wrapping if it's derived from a NSW
1013   // AddRec using a NSW operation.
1014   if (auto *OBO = dyn_cast<OverflowingBinaryOperator>(NonConstIndex))
1015     if (OBO->hasNoSignedWrap() &&
1016         // Assume constant for other the operand so that the AddRec can be
1017         // easily found.
1018         isa<ConstantInt>(OBO->getOperand(1))) {
1019       auto *OpScev = PSE.getSCEV(OBO->getOperand(0));
1020 
1021       if (auto *OpAR = dyn_cast<SCEVAddRecExpr>(OpScev))
1022         return OpAR->getLoop() == L && OpAR->getNoWrapFlags(SCEV::FlagNSW);
1023     }
1024 
1025   return false;
1026 }
1027 
1028 /// Check whether the access through \p Ptr has a constant stride.
getPtrStride(PredicatedScalarEvolution & PSE,Type * AccessTy,Value * Ptr,const Loop * Lp,const ValueToValueMap & StridesMap,bool Assume,bool ShouldCheckWrap)1029 int64_t llvm::getPtrStride(PredicatedScalarEvolution &PSE, Type *AccessTy,
1030                            Value *Ptr, const Loop *Lp,
1031                            const ValueToValueMap &StridesMap, bool Assume,
1032                            bool ShouldCheckWrap) {
1033   Type *Ty = Ptr->getType();
1034   assert(Ty->isPointerTy() && "Unexpected non-ptr");
1035   unsigned AddrSpace = Ty->getPointerAddressSpace();
1036 
1037   // Make sure we're not accessing an aggregate type.
1038   // TODO: Why? This doesn't make any sense.
1039   if (AccessTy->isAggregateType()) {
1040     LLVM_DEBUG(dbgs() << "LAA: Bad stride - Not a pointer to a scalar type"
1041                       << *Ptr << "\n");
1042     return 0;
1043   }
1044 
1045   const SCEV *PtrScev = replaceSymbolicStrideSCEV(PSE, StridesMap, Ptr);
1046 
1047   const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
1048   if (Assume && !AR)
1049     AR = PSE.getAsAddRec(Ptr);
1050 
1051   if (!AR) {
1052     LLVM_DEBUG(dbgs() << "LAA: Bad stride - Not an AddRecExpr pointer " << *Ptr
1053                       << " SCEV: " << *PtrScev << "\n");
1054     return 0;
1055   }
1056 
1057   // The access function must stride over the innermost loop.
1058   if (Lp != AR->getLoop()) {
1059     LLVM_DEBUG(dbgs() << "LAA: Bad stride - Not striding over innermost loop "
1060                       << *Ptr << " SCEV: " << *AR << "\n");
1061     return 0;
1062   }
1063 
1064   // The address calculation must not wrap. Otherwise, a dependence could be
1065   // inverted.
1066   // An inbounds getelementptr that is a AddRec with a unit stride
1067   // cannot wrap per definition. The unit stride requirement is checked later.
1068   // An getelementptr without an inbounds attribute and unit stride would have
1069   // to access the pointer value "0" which is undefined behavior in address
1070   // space 0, therefore we can also vectorize this case.
1071   bool IsInBoundsGEP = isInBoundsGep(Ptr);
1072   bool IsNoWrapAddRec = !ShouldCheckWrap ||
1073     PSE.hasNoOverflow(Ptr, SCEVWrapPredicate::IncrementNUSW) ||
1074     isNoWrapAddRec(Ptr, AR, PSE, Lp);
1075   if (!IsNoWrapAddRec && !IsInBoundsGEP &&
1076       NullPointerIsDefined(Lp->getHeader()->getParent(), AddrSpace)) {
1077     if (Assume) {
1078       PSE.setNoOverflow(Ptr, SCEVWrapPredicate::IncrementNUSW);
1079       IsNoWrapAddRec = true;
1080       LLVM_DEBUG(dbgs() << "LAA: Pointer may wrap in the address space:\n"
1081                         << "LAA:   Pointer: " << *Ptr << "\n"
1082                         << "LAA:   SCEV: " << *AR << "\n"
1083                         << "LAA:   Added an overflow assumption\n");
1084     } else {
1085       LLVM_DEBUG(
1086           dbgs() << "LAA: Bad stride - Pointer may wrap in the address space "
1087                  << *Ptr << " SCEV: " << *AR << "\n");
1088       return 0;
1089     }
1090   }
1091 
1092   // Check the step is constant.
1093   const SCEV *Step = AR->getStepRecurrence(*PSE.getSE());
1094 
1095   // Calculate the pointer stride and check if it is constant.
1096   const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
1097   if (!C) {
1098     LLVM_DEBUG(dbgs() << "LAA: Bad stride - Not a constant strided " << *Ptr
1099                       << " SCEV: " << *AR << "\n");
1100     return 0;
1101   }
1102 
1103   auto &DL = Lp->getHeader()->getModule()->getDataLayout();
1104   int64_t Size = DL.getTypeAllocSize(AccessTy);
1105   const APInt &APStepVal = C->getAPInt();
1106 
1107   // Huge step value - give up.
1108   if (APStepVal.getBitWidth() > 64)
1109     return 0;
1110 
1111   int64_t StepVal = APStepVal.getSExtValue();
1112 
1113   // Strided access.
1114   int64_t Stride = StepVal / Size;
1115   int64_t Rem = StepVal % Size;
1116   if (Rem)
1117     return 0;
1118 
1119   // If the SCEV could wrap but we have an inbounds gep with a unit stride we
1120   // know we can't "wrap around the address space". In case of address space
1121   // zero we know that this won't happen without triggering undefined behavior.
1122   if (!IsNoWrapAddRec && Stride != 1 && Stride != -1 &&
1123       (IsInBoundsGEP || !NullPointerIsDefined(Lp->getHeader()->getParent(),
1124                                               AddrSpace))) {
1125     if (Assume) {
1126       // We can avoid this case by adding a run-time check.
1127       LLVM_DEBUG(dbgs() << "LAA: Non unit strided pointer which is not either "
1128                         << "inbounds or in address space 0 may wrap:\n"
1129                         << "LAA:   Pointer: " << *Ptr << "\n"
1130                         << "LAA:   SCEV: " << *AR << "\n"
1131                         << "LAA:   Added an overflow assumption\n");
1132       PSE.setNoOverflow(Ptr, SCEVWrapPredicate::IncrementNUSW);
1133     } else
1134       return 0;
1135   }
1136 
1137   return Stride;
1138 }
1139 
getPointersDiff(Type * ElemTyA,Value * PtrA,Type * ElemTyB,Value * PtrB,const DataLayout & DL,ScalarEvolution & SE,bool StrictCheck,bool CheckType)1140 Optional<int> llvm::getPointersDiff(Type *ElemTyA, Value *PtrA, Type *ElemTyB,
1141                                     Value *PtrB, const DataLayout &DL,
1142                                     ScalarEvolution &SE, bool StrictCheck,
1143                                     bool CheckType) {
1144   assert(PtrA && PtrB && "Expected non-nullptr pointers.");
1145   assert(cast<PointerType>(PtrA->getType())
1146              ->isOpaqueOrPointeeTypeMatches(ElemTyA) && "Wrong PtrA type");
1147   assert(cast<PointerType>(PtrB->getType())
1148              ->isOpaqueOrPointeeTypeMatches(ElemTyB) && "Wrong PtrB type");
1149 
1150   // Make sure that A and B are different pointers.
1151   if (PtrA == PtrB)
1152     return 0;
1153 
1154   // Make sure that the element types are the same if required.
1155   if (CheckType && ElemTyA != ElemTyB)
1156     return None;
1157 
1158   unsigned ASA = PtrA->getType()->getPointerAddressSpace();
1159   unsigned ASB = PtrB->getType()->getPointerAddressSpace();
1160 
1161   // Check that the address spaces match.
1162   if (ASA != ASB)
1163     return None;
1164   unsigned IdxWidth = DL.getIndexSizeInBits(ASA);
1165 
1166   APInt OffsetA(IdxWidth, 0), OffsetB(IdxWidth, 0);
1167   Value *PtrA1 = PtrA->stripAndAccumulateInBoundsConstantOffsets(DL, OffsetA);
1168   Value *PtrB1 = PtrB->stripAndAccumulateInBoundsConstantOffsets(DL, OffsetB);
1169 
1170   int Val;
1171   if (PtrA1 == PtrB1) {
1172     // Retrieve the address space again as pointer stripping now tracks through
1173     // `addrspacecast`.
1174     ASA = cast<PointerType>(PtrA1->getType())->getAddressSpace();
1175     ASB = cast<PointerType>(PtrB1->getType())->getAddressSpace();
1176     // Check that the address spaces match and that the pointers are valid.
1177     if (ASA != ASB)
1178       return None;
1179 
1180     IdxWidth = DL.getIndexSizeInBits(ASA);
1181     OffsetA = OffsetA.sextOrTrunc(IdxWidth);
1182     OffsetB = OffsetB.sextOrTrunc(IdxWidth);
1183 
1184     OffsetB -= OffsetA;
1185     Val = OffsetB.getSExtValue();
1186   } else {
1187     // Otherwise compute the distance with SCEV between the base pointers.
1188     const SCEV *PtrSCEVA = SE.getSCEV(PtrA);
1189     const SCEV *PtrSCEVB = SE.getSCEV(PtrB);
1190     const auto *Diff =
1191         dyn_cast<SCEVConstant>(SE.getMinusSCEV(PtrSCEVB, PtrSCEVA));
1192     if (!Diff)
1193       return None;
1194     Val = Diff->getAPInt().getSExtValue();
1195   }
1196   int Size = DL.getTypeStoreSize(ElemTyA);
1197   int Dist = Val / Size;
1198 
1199   // Ensure that the calculated distance matches the type-based one after all
1200   // the bitcasts removal in the provided pointers.
1201   if (!StrictCheck || Dist * Size == Val)
1202     return Dist;
1203   return None;
1204 }
1205 
sortPtrAccesses(ArrayRef<Value * > VL,Type * ElemTy,const DataLayout & DL,ScalarEvolution & SE,SmallVectorImpl<unsigned> & SortedIndices)1206 bool llvm::sortPtrAccesses(ArrayRef<Value *> VL, Type *ElemTy,
1207                            const DataLayout &DL, ScalarEvolution &SE,
1208                            SmallVectorImpl<unsigned> &SortedIndices) {
1209   assert(llvm::all_of(
1210              VL, [](const Value *V) { return V->getType()->isPointerTy(); }) &&
1211          "Expected list of pointer operands.");
1212   // Walk over the pointers, and map each of them to an offset relative to
1213   // first pointer in the array.
1214   Value *Ptr0 = VL[0];
1215 
1216   using DistOrdPair = std::pair<int64_t, int>;
1217   auto Compare = [](const DistOrdPair &L, const DistOrdPair &R) {
1218     return L.first < R.first;
1219   };
1220   std::set<DistOrdPair, decltype(Compare)> Offsets(Compare);
1221   Offsets.emplace(0, 0);
1222   int Cnt = 1;
1223   bool IsConsecutive = true;
1224   for (auto *Ptr : VL.drop_front()) {
1225     Optional<int> Diff = getPointersDiff(ElemTy, Ptr0, ElemTy, Ptr, DL, SE,
1226                                          /*StrictCheck=*/true);
1227     if (!Diff)
1228       return false;
1229 
1230     // Check if the pointer with the same offset is found.
1231     int64_t Offset = *Diff;
1232     auto Res = Offsets.emplace(Offset, Cnt);
1233     if (!Res.second)
1234       return false;
1235     // Consecutive order if the inserted element is the last one.
1236     IsConsecutive = IsConsecutive && std::next(Res.first) == Offsets.end();
1237     ++Cnt;
1238   }
1239   SortedIndices.clear();
1240   if (!IsConsecutive) {
1241     // Fill SortedIndices array only if it is non-consecutive.
1242     SortedIndices.resize(VL.size());
1243     Cnt = 0;
1244     for (const std::pair<int64_t, int> &Pair : Offsets) {
1245       SortedIndices[Cnt] = Pair.second;
1246       ++Cnt;
1247     }
1248   }
1249   return true;
1250 }
1251 
1252 /// Returns true if the memory operations \p A and \p B are consecutive.
isConsecutiveAccess(Value * A,Value * B,const DataLayout & DL,ScalarEvolution & SE,bool CheckType)1253 bool llvm::isConsecutiveAccess(Value *A, Value *B, const DataLayout &DL,
1254                                ScalarEvolution &SE, bool CheckType) {
1255   Value *PtrA = getLoadStorePointerOperand(A);
1256   Value *PtrB = getLoadStorePointerOperand(B);
1257   if (!PtrA || !PtrB)
1258     return false;
1259   Type *ElemTyA = getLoadStoreType(A);
1260   Type *ElemTyB = getLoadStoreType(B);
1261   Optional<int> Diff = getPointersDiff(ElemTyA, PtrA, ElemTyB, PtrB, DL, SE,
1262                                        /*StrictCheck=*/true, CheckType);
1263   return Diff && *Diff == 1;
1264 }
1265 
visitPointers(Value * StartPtr,const Loop & InnermostLoop,function_ref<void (Value *)> AddPointer)1266 static void visitPointers(Value *StartPtr, const Loop &InnermostLoop,
1267                           function_ref<void(Value *)> AddPointer) {
1268   SmallPtrSet<Value *, 8> Visited;
1269   SmallVector<Value *> WorkList;
1270   WorkList.push_back(StartPtr);
1271 
1272   while (!WorkList.empty()) {
1273     Value *Ptr = WorkList.pop_back_val();
1274     if (!Visited.insert(Ptr).second)
1275       continue;
1276     auto *PN = dyn_cast<PHINode>(Ptr);
1277     // SCEV does not look through non-header PHIs inside the loop. Such phis
1278     // can be analyzed by adding separate accesses for each incoming pointer
1279     // value.
1280     if (PN && InnermostLoop.contains(PN->getParent()) &&
1281         PN->getParent() != InnermostLoop.getHeader()) {
1282       for (const Use &Inc : PN->incoming_values())
1283         WorkList.push_back(Inc);
1284     } else
1285       AddPointer(Ptr);
1286   }
1287 }
1288 
addAccess(StoreInst * SI)1289 void MemoryDepChecker::addAccess(StoreInst *SI) {
1290   visitPointers(SI->getPointerOperand(), *InnermostLoop,
1291                 [this, SI](Value *Ptr) {
1292                   Accesses[MemAccessInfo(Ptr, true)].push_back(AccessIdx);
1293                   InstMap.push_back(SI);
1294                   ++AccessIdx;
1295                 });
1296 }
1297 
addAccess(LoadInst * LI)1298 void MemoryDepChecker::addAccess(LoadInst *LI) {
1299   visitPointers(LI->getPointerOperand(), *InnermostLoop,
1300                 [this, LI](Value *Ptr) {
1301                   Accesses[MemAccessInfo(Ptr, false)].push_back(AccessIdx);
1302                   InstMap.push_back(LI);
1303                   ++AccessIdx;
1304                 });
1305 }
1306 
1307 MemoryDepChecker::VectorizationSafetyStatus
isSafeForVectorization(DepType Type)1308 MemoryDepChecker::Dependence::isSafeForVectorization(DepType Type) {
1309   switch (Type) {
1310   case NoDep:
1311   case Forward:
1312   case BackwardVectorizable:
1313     return VectorizationSafetyStatus::Safe;
1314 
1315   case Unknown:
1316     return VectorizationSafetyStatus::PossiblySafeWithRtChecks;
1317   case ForwardButPreventsForwarding:
1318   case Backward:
1319   case BackwardVectorizableButPreventsForwarding:
1320     return VectorizationSafetyStatus::Unsafe;
1321   }
1322   llvm_unreachable("unexpected DepType!");
1323 }
1324 
isBackward() const1325 bool MemoryDepChecker::Dependence::isBackward() const {
1326   switch (Type) {
1327   case NoDep:
1328   case Forward:
1329   case ForwardButPreventsForwarding:
1330   case Unknown:
1331     return false;
1332 
1333   case BackwardVectorizable:
1334   case Backward:
1335   case BackwardVectorizableButPreventsForwarding:
1336     return true;
1337   }
1338   llvm_unreachable("unexpected DepType!");
1339 }
1340 
isPossiblyBackward() const1341 bool MemoryDepChecker::Dependence::isPossiblyBackward() const {
1342   return isBackward() || Type == Unknown;
1343 }
1344 
isForward() const1345 bool MemoryDepChecker::Dependence::isForward() const {
1346   switch (Type) {
1347   case Forward:
1348   case ForwardButPreventsForwarding:
1349     return true;
1350 
1351   case NoDep:
1352   case Unknown:
1353   case BackwardVectorizable:
1354   case Backward:
1355   case BackwardVectorizableButPreventsForwarding:
1356     return false;
1357   }
1358   llvm_unreachable("unexpected DepType!");
1359 }
1360 
couldPreventStoreLoadForward(uint64_t Distance,uint64_t TypeByteSize)1361 bool MemoryDepChecker::couldPreventStoreLoadForward(uint64_t Distance,
1362                                                     uint64_t TypeByteSize) {
1363   // If loads occur at a distance that is not a multiple of a feasible vector
1364   // factor store-load forwarding does not take place.
1365   // Positive dependences might cause troubles because vectorizing them might
1366   // prevent store-load forwarding making vectorized code run a lot slower.
1367   //   a[i] = a[i-3] ^ a[i-8];
1368   //   The stores to a[i:i+1] don't align with the stores to a[i-3:i-2] and
1369   //   hence on your typical architecture store-load forwarding does not take
1370   //   place. Vectorizing in such cases does not make sense.
1371   // Store-load forwarding distance.
1372 
1373   // After this many iterations store-to-load forwarding conflicts should not
1374   // cause any slowdowns.
1375   const uint64_t NumItersForStoreLoadThroughMemory = 8 * TypeByteSize;
1376   // Maximum vector factor.
1377   uint64_t MaxVFWithoutSLForwardIssues = std::min(
1378       VectorizerParams::MaxVectorWidth * TypeByteSize, MaxSafeDepDistBytes);
1379 
1380   // Compute the smallest VF at which the store and load would be misaligned.
1381   for (uint64_t VF = 2 * TypeByteSize; VF <= MaxVFWithoutSLForwardIssues;
1382        VF *= 2) {
1383     // If the number of vector iteration between the store and the load are
1384     // small we could incur conflicts.
1385     if (Distance % VF && Distance / VF < NumItersForStoreLoadThroughMemory) {
1386       MaxVFWithoutSLForwardIssues = (VF >> 1);
1387       break;
1388     }
1389   }
1390 
1391   if (MaxVFWithoutSLForwardIssues < 2 * TypeByteSize) {
1392     LLVM_DEBUG(
1393         dbgs() << "LAA: Distance " << Distance
1394                << " that could cause a store-load forwarding conflict\n");
1395     return true;
1396   }
1397 
1398   if (MaxVFWithoutSLForwardIssues < MaxSafeDepDistBytes &&
1399       MaxVFWithoutSLForwardIssues !=
1400           VectorizerParams::MaxVectorWidth * TypeByteSize)
1401     MaxSafeDepDistBytes = MaxVFWithoutSLForwardIssues;
1402   return false;
1403 }
1404 
mergeInStatus(VectorizationSafetyStatus S)1405 void MemoryDepChecker::mergeInStatus(VectorizationSafetyStatus S) {
1406   if (Status < S)
1407     Status = S;
1408 }
1409 
1410 /// Given a non-constant (unknown) dependence-distance \p Dist between two
1411 /// memory accesses, that have the same stride whose absolute value is given
1412 /// in \p Stride, and that have the same type size \p TypeByteSize,
1413 /// in a loop whose takenCount is \p BackedgeTakenCount, check if it is
1414 /// possible to prove statically that the dependence distance is larger
1415 /// than the range that the accesses will travel through the execution of
1416 /// the loop. If so, return true; false otherwise. This is useful for
1417 /// example in loops such as the following (PR31098):
1418 ///     for (i = 0; i < D; ++i) {
1419 ///                = out[i];
1420 ///       out[i+D] =
1421 ///     }
isSafeDependenceDistance(const DataLayout & DL,ScalarEvolution & SE,const SCEV & BackedgeTakenCount,const SCEV & Dist,uint64_t Stride,uint64_t TypeByteSize)1422 static bool isSafeDependenceDistance(const DataLayout &DL, ScalarEvolution &SE,
1423                                      const SCEV &BackedgeTakenCount,
1424                                      const SCEV &Dist, uint64_t Stride,
1425                                      uint64_t TypeByteSize) {
1426 
1427   // If we can prove that
1428   //      (**) |Dist| > BackedgeTakenCount * Step
1429   // where Step is the absolute stride of the memory accesses in bytes,
1430   // then there is no dependence.
1431   //
1432   // Rationale:
1433   // We basically want to check if the absolute distance (|Dist/Step|)
1434   // is >= the loop iteration count (or > BackedgeTakenCount).
1435   // This is equivalent to the Strong SIV Test (Practical Dependence Testing,
1436   // Section 4.2.1); Note, that for vectorization it is sufficient to prove
1437   // that the dependence distance is >= VF; This is checked elsewhere.
1438   // But in some cases we can prune unknown dependence distances early, and
1439   // even before selecting the VF, and without a runtime test, by comparing
1440   // the distance against the loop iteration count. Since the vectorized code
1441   // will be executed only if LoopCount >= VF, proving distance >= LoopCount
1442   // also guarantees that distance >= VF.
1443   //
1444   const uint64_t ByteStride = Stride * TypeByteSize;
1445   const SCEV *Step = SE.getConstant(BackedgeTakenCount.getType(), ByteStride);
1446   const SCEV *Product = SE.getMulExpr(&BackedgeTakenCount, Step);
1447 
1448   const SCEV *CastedDist = &Dist;
1449   const SCEV *CastedProduct = Product;
1450   uint64_t DistTypeSize = DL.getTypeAllocSize(Dist.getType());
1451   uint64_t ProductTypeSize = DL.getTypeAllocSize(Product->getType());
1452 
1453   // The dependence distance can be positive/negative, so we sign extend Dist;
1454   // The multiplication of the absolute stride in bytes and the
1455   // backedgeTakenCount is non-negative, so we zero extend Product.
1456   if (DistTypeSize > ProductTypeSize)
1457     CastedProduct = SE.getZeroExtendExpr(Product, Dist.getType());
1458   else
1459     CastedDist = SE.getNoopOrSignExtend(&Dist, Product->getType());
1460 
1461   // Is  Dist - (BackedgeTakenCount * Step) > 0 ?
1462   // (If so, then we have proven (**) because |Dist| >= Dist)
1463   const SCEV *Minus = SE.getMinusSCEV(CastedDist, CastedProduct);
1464   if (SE.isKnownPositive(Minus))
1465     return true;
1466 
1467   // Second try: Is  -Dist - (BackedgeTakenCount * Step) > 0 ?
1468   // (If so, then we have proven (**) because |Dist| >= -1*Dist)
1469   const SCEV *NegDist = SE.getNegativeSCEV(CastedDist);
1470   Minus = SE.getMinusSCEV(NegDist, CastedProduct);
1471   if (SE.isKnownPositive(Minus))
1472     return true;
1473 
1474   return false;
1475 }
1476 
1477 /// Check the dependence for two accesses with the same stride \p Stride.
1478 /// \p Distance is the positive distance and \p TypeByteSize is type size in
1479 /// bytes.
1480 ///
1481 /// \returns true if they are independent.
areStridedAccessesIndependent(uint64_t Distance,uint64_t Stride,uint64_t TypeByteSize)1482 static bool areStridedAccessesIndependent(uint64_t Distance, uint64_t Stride,
1483                                           uint64_t TypeByteSize) {
1484   assert(Stride > 1 && "The stride must be greater than 1");
1485   assert(TypeByteSize > 0 && "The type size in byte must be non-zero");
1486   assert(Distance > 0 && "The distance must be non-zero");
1487 
1488   // Skip if the distance is not multiple of type byte size.
1489   if (Distance % TypeByteSize)
1490     return false;
1491 
1492   uint64_t ScaledDist = Distance / TypeByteSize;
1493 
1494   // No dependence if the scaled distance is not multiple of the stride.
1495   // E.g.
1496   //      for (i = 0; i < 1024 ; i += 4)
1497   //        A[i+2] = A[i] + 1;
1498   //
1499   // Two accesses in memory (scaled distance is 2, stride is 4):
1500   //     | A[0] |      |      |      | A[4] |      |      |      |
1501   //     |      |      | A[2] |      |      |      | A[6] |      |
1502   //
1503   // E.g.
1504   //      for (i = 0; i < 1024 ; i += 3)
1505   //        A[i+4] = A[i] + 1;
1506   //
1507   // Two accesses in memory (scaled distance is 4, stride is 3):
1508   //     | A[0] |      |      | A[3] |      |      | A[6] |      |      |
1509   //     |      |      |      |      | A[4] |      |      | A[7] |      |
1510   return ScaledDist % Stride;
1511 }
1512 
1513 MemoryDepChecker::Dependence::DepType
isDependent(const MemAccessInfo & A,unsigned AIdx,const MemAccessInfo & B,unsigned BIdx,const ValueToValueMap & Strides)1514 MemoryDepChecker::isDependent(const MemAccessInfo &A, unsigned AIdx,
1515                               const MemAccessInfo &B, unsigned BIdx,
1516                               const ValueToValueMap &Strides) {
1517   assert (AIdx < BIdx && "Must pass arguments in program order");
1518 
1519   Value *APtr = A.getPointer();
1520   Value *BPtr = B.getPointer();
1521   bool AIsWrite = A.getInt();
1522   bool BIsWrite = B.getInt();
1523   Type *ATy = APtr->getType()->getPointerElementType();
1524   Type *BTy = BPtr->getType()->getPointerElementType();
1525 
1526   // Two reads are independent.
1527   if (!AIsWrite && !BIsWrite)
1528     return Dependence::NoDep;
1529 
1530   // We cannot check pointers in different address spaces.
1531   if (APtr->getType()->getPointerAddressSpace() !=
1532       BPtr->getType()->getPointerAddressSpace())
1533     return Dependence::Unknown;
1534 
1535   int64_t StrideAPtr =
1536       getPtrStride(PSE, ATy, APtr, InnermostLoop, Strides, true);
1537   int64_t StrideBPtr =
1538       getPtrStride(PSE, BTy, BPtr, InnermostLoop, Strides, true);
1539 
1540   const SCEV *Src = PSE.getSCEV(APtr);
1541   const SCEV *Sink = PSE.getSCEV(BPtr);
1542 
1543   // If the induction step is negative we have to invert source and sink of the
1544   // dependence.
1545   if (StrideAPtr < 0) {
1546     std::swap(APtr, BPtr);
1547     std::swap(ATy, BTy);
1548     std::swap(Src, Sink);
1549     std::swap(AIsWrite, BIsWrite);
1550     std::swap(AIdx, BIdx);
1551     std::swap(StrideAPtr, StrideBPtr);
1552   }
1553 
1554   const SCEV *Dist = PSE.getSE()->getMinusSCEV(Sink, Src);
1555 
1556   LLVM_DEBUG(dbgs() << "LAA: Src Scev: " << *Src << "Sink Scev: " << *Sink
1557                     << "(Induction step: " << StrideAPtr << ")\n");
1558   LLVM_DEBUG(dbgs() << "LAA: Distance for " << *InstMap[AIdx] << " to "
1559                     << *InstMap[BIdx] << ": " << *Dist << "\n");
1560 
1561   // Need accesses with constant stride. We don't want to vectorize
1562   // "A[B[i]] += ..." and similar code or pointer arithmetic that could wrap in
1563   // the address space.
1564   if (!StrideAPtr || !StrideBPtr || StrideAPtr != StrideBPtr){
1565     LLVM_DEBUG(dbgs() << "Pointer access with non-constant stride\n");
1566     return Dependence::Unknown;
1567   }
1568 
1569   auto &DL = InnermostLoop->getHeader()->getModule()->getDataLayout();
1570   uint64_t TypeByteSize = DL.getTypeAllocSize(ATy);
1571   uint64_t Stride = std::abs(StrideAPtr);
1572   const SCEVConstant *C = dyn_cast<SCEVConstant>(Dist);
1573   if (!C) {
1574     if (!isa<SCEVCouldNotCompute>(Dist) &&
1575         TypeByteSize == DL.getTypeAllocSize(BTy) &&
1576         isSafeDependenceDistance(DL, *(PSE.getSE()),
1577                                  *(PSE.getBackedgeTakenCount()), *Dist, Stride,
1578                                  TypeByteSize))
1579       return Dependence::NoDep;
1580 
1581     LLVM_DEBUG(dbgs() << "LAA: Dependence because of non-constant distance\n");
1582     FoundNonConstantDistanceDependence = true;
1583     return Dependence::Unknown;
1584   }
1585 
1586   const APInt &Val = C->getAPInt();
1587   int64_t Distance = Val.getSExtValue();
1588 
1589   // Attempt to prove strided accesses independent.
1590   if (std::abs(Distance) > 0 && Stride > 1 && ATy == BTy &&
1591       areStridedAccessesIndependent(std::abs(Distance), Stride, TypeByteSize)) {
1592     LLVM_DEBUG(dbgs() << "LAA: Strided accesses are independent\n");
1593     return Dependence::NoDep;
1594   }
1595 
1596   // Negative distances are not plausible dependencies.
1597   if (Val.isNegative()) {
1598     bool IsTrueDataDependence = (AIsWrite && !BIsWrite);
1599     if (IsTrueDataDependence && EnableForwardingConflictDetection &&
1600         (couldPreventStoreLoadForward(Val.abs().getZExtValue(), TypeByteSize) ||
1601          ATy != BTy)) {
1602       LLVM_DEBUG(dbgs() << "LAA: Forward but may prevent st->ld forwarding\n");
1603       return Dependence::ForwardButPreventsForwarding;
1604     }
1605 
1606     LLVM_DEBUG(dbgs() << "LAA: Dependence is negative\n");
1607     return Dependence::Forward;
1608   }
1609 
1610   // Write to the same location with the same size.
1611   // Could be improved to assert type sizes are the same (i32 == float, etc).
1612   if (Val == 0) {
1613     if (ATy == BTy)
1614       return Dependence::Forward;
1615     LLVM_DEBUG(
1616         dbgs() << "LAA: Zero dependence difference but different types\n");
1617     return Dependence::Unknown;
1618   }
1619 
1620   assert(Val.isStrictlyPositive() && "Expect a positive value");
1621 
1622   if (ATy != BTy) {
1623     LLVM_DEBUG(
1624         dbgs()
1625         << "LAA: ReadWrite-Write positive dependency with different types\n");
1626     return Dependence::Unknown;
1627   }
1628 
1629   // Bail out early if passed-in parameters make vectorization not feasible.
1630   unsigned ForcedFactor = (VectorizerParams::VectorizationFactor ?
1631                            VectorizerParams::VectorizationFactor : 1);
1632   unsigned ForcedUnroll = (VectorizerParams::VectorizationInterleave ?
1633                            VectorizerParams::VectorizationInterleave : 1);
1634   // The minimum number of iterations for a vectorized/unrolled version.
1635   unsigned MinNumIter = std::max(ForcedFactor * ForcedUnroll, 2U);
1636 
1637   // It's not vectorizable if the distance is smaller than the minimum distance
1638   // needed for a vectroized/unrolled version. Vectorizing one iteration in
1639   // front needs TypeByteSize * Stride. Vectorizing the last iteration needs
1640   // TypeByteSize (No need to plus the last gap distance).
1641   //
1642   // E.g. Assume one char is 1 byte in memory and one int is 4 bytes.
1643   //      foo(int *A) {
1644   //        int *B = (int *)((char *)A + 14);
1645   //        for (i = 0 ; i < 1024 ; i += 2)
1646   //          B[i] = A[i] + 1;
1647   //      }
1648   //
1649   // Two accesses in memory (stride is 2):
1650   //     | A[0] |      | A[2] |      | A[4] |      | A[6] |      |
1651   //                              | B[0] |      | B[2] |      | B[4] |
1652   //
1653   // Distance needs for vectorizing iterations except the last iteration:
1654   // 4 * 2 * (MinNumIter - 1). Distance needs for the last iteration: 4.
1655   // So the minimum distance needed is: 4 * 2 * (MinNumIter - 1) + 4.
1656   //
1657   // If MinNumIter is 2, it is vectorizable as the minimum distance needed is
1658   // 12, which is less than distance.
1659   //
1660   // If MinNumIter is 4 (Say if a user forces the vectorization factor to be 4),
1661   // the minimum distance needed is 28, which is greater than distance. It is
1662   // not safe to do vectorization.
1663   uint64_t MinDistanceNeeded =
1664       TypeByteSize * Stride * (MinNumIter - 1) + TypeByteSize;
1665   if (MinDistanceNeeded > static_cast<uint64_t>(Distance)) {
1666     LLVM_DEBUG(dbgs() << "LAA: Failure because of positive distance "
1667                       << Distance << '\n');
1668     return Dependence::Backward;
1669   }
1670 
1671   // Unsafe if the minimum distance needed is greater than max safe distance.
1672   if (MinDistanceNeeded > MaxSafeDepDistBytes) {
1673     LLVM_DEBUG(dbgs() << "LAA: Failure because it needs at least "
1674                       << MinDistanceNeeded << " size in bytes");
1675     return Dependence::Backward;
1676   }
1677 
1678   // Positive distance bigger than max vectorization factor.
1679   // FIXME: Should use max factor instead of max distance in bytes, which could
1680   // not handle different types.
1681   // E.g. Assume one char is 1 byte in memory and one int is 4 bytes.
1682   //      void foo (int *A, char *B) {
1683   //        for (unsigned i = 0; i < 1024; i++) {
1684   //          A[i+2] = A[i] + 1;
1685   //          B[i+2] = B[i] + 1;
1686   //        }
1687   //      }
1688   //
1689   // This case is currently unsafe according to the max safe distance. If we
1690   // analyze the two accesses on array B, the max safe dependence distance
1691   // is 2. Then we analyze the accesses on array A, the minimum distance needed
1692   // is 8, which is less than 2 and forbidden vectorization, But actually
1693   // both A and B could be vectorized by 2 iterations.
1694   MaxSafeDepDistBytes =
1695       std::min(static_cast<uint64_t>(Distance), MaxSafeDepDistBytes);
1696 
1697   bool IsTrueDataDependence = (!AIsWrite && BIsWrite);
1698   if (IsTrueDataDependence && EnableForwardingConflictDetection &&
1699       couldPreventStoreLoadForward(Distance, TypeByteSize))
1700     return Dependence::BackwardVectorizableButPreventsForwarding;
1701 
1702   uint64_t MaxVF = MaxSafeDepDistBytes / (TypeByteSize * Stride);
1703   LLVM_DEBUG(dbgs() << "LAA: Positive distance " << Val.getSExtValue()
1704                     << " with max VF = " << MaxVF << '\n');
1705   uint64_t MaxVFInBits = MaxVF * TypeByteSize * 8;
1706   MaxSafeVectorWidthInBits = std::min(MaxSafeVectorWidthInBits, MaxVFInBits);
1707   return Dependence::BackwardVectorizable;
1708 }
1709 
areDepsSafe(DepCandidates & AccessSets,MemAccessInfoList & CheckDeps,const ValueToValueMap & Strides)1710 bool MemoryDepChecker::areDepsSafe(DepCandidates &AccessSets,
1711                                    MemAccessInfoList &CheckDeps,
1712                                    const ValueToValueMap &Strides) {
1713 
1714   MaxSafeDepDistBytes = -1;
1715   SmallPtrSet<MemAccessInfo, 8> Visited;
1716   for (MemAccessInfo CurAccess : CheckDeps) {
1717     if (Visited.count(CurAccess))
1718       continue;
1719 
1720     // Get the relevant memory access set.
1721     EquivalenceClasses<MemAccessInfo>::iterator I =
1722       AccessSets.findValue(AccessSets.getLeaderValue(CurAccess));
1723 
1724     // Check accesses within this set.
1725     EquivalenceClasses<MemAccessInfo>::member_iterator AI =
1726         AccessSets.member_begin(I);
1727     EquivalenceClasses<MemAccessInfo>::member_iterator AE =
1728         AccessSets.member_end();
1729 
1730     // Check every access pair.
1731     while (AI != AE) {
1732       Visited.insert(*AI);
1733       bool AIIsWrite = AI->getInt();
1734       // Check loads only against next equivalent class, but stores also against
1735       // other stores in the same equivalence class - to the same address.
1736       EquivalenceClasses<MemAccessInfo>::member_iterator OI =
1737           (AIIsWrite ? AI : std::next(AI));
1738       while (OI != AE) {
1739         // Check every accessing instruction pair in program order.
1740         for (std::vector<unsigned>::iterator I1 = Accesses[*AI].begin(),
1741              I1E = Accesses[*AI].end(); I1 != I1E; ++I1)
1742           // Scan all accesses of another equivalence class, but only the next
1743           // accesses of the same equivalent class.
1744           for (std::vector<unsigned>::iterator
1745                    I2 = (OI == AI ? std::next(I1) : Accesses[*OI].begin()),
1746                    I2E = (OI == AI ? I1E : Accesses[*OI].end());
1747                I2 != I2E; ++I2) {
1748             auto A = std::make_pair(&*AI, *I1);
1749             auto B = std::make_pair(&*OI, *I2);
1750 
1751             assert(*I1 != *I2);
1752             if (*I1 > *I2)
1753               std::swap(A, B);
1754 
1755             Dependence::DepType Type =
1756                 isDependent(*A.first, A.second, *B.first, B.second, Strides);
1757             mergeInStatus(Dependence::isSafeForVectorization(Type));
1758 
1759             // Gather dependences unless we accumulated MaxDependences
1760             // dependences.  In that case return as soon as we find the first
1761             // unsafe dependence.  This puts a limit on this quadratic
1762             // algorithm.
1763             if (RecordDependences) {
1764               if (Type != Dependence::NoDep)
1765                 Dependences.push_back(Dependence(A.second, B.second, Type));
1766 
1767               if (Dependences.size() >= MaxDependences) {
1768                 RecordDependences = false;
1769                 Dependences.clear();
1770                 LLVM_DEBUG(dbgs()
1771                            << "Too many dependences, stopped recording\n");
1772               }
1773             }
1774             if (!RecordDependences && !isSafeForVectorization())
1775               return false;
1776           }
1777         ++OI;
1778       }
1779       AI++;
1780     }
1781   }
1782 
1783   LLVM_DEBUG(dbgs() << "Total Dependences: " << Dependences.size() << "\n");
1784   return isSafeForVectorization();
1785 }
1786 
1787 SmallVector<Instruction *, 4>
getInstructionsForAccess(Value * Ptr,bool isWrite) const1788 MemoryDepChecker::getInstructionsForAccess(Value *Ptr, bool isWrite) const {
1789   MemAccessInfo Access(Ptr, isWrite);
1790   auto &IndexVector = Accesses.find(Access)->second;
1791 
1792   SmallVector<Instruction *, 4> Insts;
1793   transform(IndexVector,
1794                  std::back_inserter(Insts),
1795                  [&](unsigned Idx) { return this->InstMap[Idx]; });
1796   return Insts;
1797 }
1798 
1799 const char *MemoryDepChecker::Dependence::DepName[] = {
1800     "NoDep", "Unknown", "Forward", "ForwardButPreventsForwarding", "Backward",
1801     "BackwardVectorizable", "BackwardVectorizableButPreventsForwarding"};
1802 
print(raw_ostream & OS,unsigned Depth,const SmallVectorImpl<Instruction * > & Instrs) const1803 void MemoryDepChecker::Dependence::print(
1804     raw_ostream &OS, unsigned Depth,
1805     const SmallVectorImpl<Instruction *> &Instrs) const {
1806   OS.indent(Depth) << DepName[Type] << ":\n";
1807   OS.indent(Depth + 2) << *Instrs[Source] << " -> \n";
1808   OS.indent(Depth + 2) << *Instrs[Destination] << "\n";
1809 }
1810 
canAnalyzeLoop()1811 bool LoopAccessInfo::canAnalyzeLoop() {
1812   // We need to have a loop header.
1813   LLVM_DEBUG(dbgs() << "LAA: Found a loop in "
1814                     << TheLoop->getHeader()->getParent()->getName() << ": "
1815                     << TheLoop->getHeader()->getName() << '\n');
1816 
1817   // We can only analyze innermost loops.
1818   if (!TheLoop->isInnermost()) {
1819     LLVM_DEBUG(dbgs() << "LAA: loop is not the innermost loop\n");
1820     recordAnalysis("NotInnerMostLoop") << "loop is not the innermost loop";
1821     return false;
1822   }
1823 
1824   // We must have a single backedge.
1825   if (TheLoop->getNumBackEdges() != 1) {
1826     LLVM_DEBUG(
1827         dbgs() << "LAA: loop control flow is not understood by analyzer\n");
1828     recordAnalysis("CFGNotUnderstood")
1829         << "loop control flow is not understood by analyzer";
1830     return false;
1831   }
1832 
1833   // ScalarEvolution needs to be able to find the exit count.
1834   const SCEV *ExitCount = PSE->getBackedgeTakenCount();
1835   if (isa<SCEVCouldNotCompute>(ExitCount)) {
1836     recordAnalysis("CantComputeNumberOfIterations")
1837         << "could not determine number of loop iterations";
1838     LLVM_DEBUG(dbgs() << "LAA: SCEV could not compute the loop exit count.\n");
1839     return false;
1840   }
1841 
1842   return true;
1843 }
1844 
analyzeLoop(AAResults * AA,LoopInfo * LI,const TargetLibraryInfo * TLI,DominatorTree * DT)1845 void LoopAccessInfo::analyzeLoop(AAResults *AA, LoopInfo *LI,
1846                                  const TargetLibraryInfo *TLI,
1847                                  DominatorTree *DT) {
1848   typedef SmallPtrSet<Value*, 16> ValueSet;
1849 
1850   // Holds the Load and Store instructions.
1851   SmallVector<LoadInst *, 16> Loads;
1852   SmallVector<StoreInst *, 16> Stores;
1853 
1854   // Holds all the different accesses in the loop.
1855   unsigned NumReads = 0;
1856   unsigned NumReadWrites = 0;
1857 
1858   bool HasComplexMemInst = false;
1859 
1860   // A runtime check is only legal to insert if there are no convergent calls.
1861   HasConvergentOp = false;
1862 
1863   PtrRtChecking->Pointers.clear();
1864   PtrRtChecking->Need = false;
1865 
1866   const bool IsAnnotatedParallel = TheLoop->isAnnotatedParallel();
1867 
1868   const bool EnableMemAccessVersioningOfLoop =
1869       EnableMemAccessVersioning &&
1870       !TheLoop->getHeader()->getParent()->hasOptSize();
1871 
1872   // For each block.
1873   for (BasicBlock *BB : TheLoop->blocks()) {
1874     // Scan the BB and collect legal loads and stores. Also detect any
1875     // convergent instructions.
1876     for (Instruction &I : *BB) {
1877       if (auto *Call = dyn_cast<CallBase>(&I)) {
1878         if (Call->isConvergent())
1879           HasConvergentOp = true;
1880       }
1881 
1882       // With both a non-vectorizable memory instruction and a convergent
1883       // operation, found in this loop, no reason to continue the search.
1884       if (HasComplexMemInst && HasConvergentOp) {
1885         CanVecMem = false;
1886         return;
1887       }
1888 
1889       // Avoid hitting recordAnalysis multiple times.
1890       if (HasComplexMemInst)
1891         continue;
1892 
1893       // If this is a load, save it. If this instruction can read from memory
1894       // but is not a load, then we quit. Notice that we don't handle function
1895       // calls that read or write.
1896       if (I.mayReadFromMemory()) {
1897         // Many math library functions read the rounding mode. We will only
1898         // vectorize a loop if it contains known function calls that don't set
1899         // the flag. Therefore, it is safe to ignore this read from memory.
1900         auto *Call = dyn_cast<CallInst>(&I);
1901         if (Call && getVectorIntrinsicIDForCall(Call, TLI))
1902           continue;
1903 
1904         // If the function has an explicit vectorized counterpart, we can safely
1905         // assume that it can be vectorized.
1906         if (Call && !Call->isNoBuiltin() && Call->getCalledFunction() &&
1907             !VFDatabase::getMappings(*Call).empty())
1908           continue;
1909 
1910         auto *Ld = dyn_cast<LoadInst>(&I);
1911         if (!Ld) {
1912           recordAnalysis("CantVectorizeInstruction", Ld)
1913             << "instruction cannot be vectorized";
1914           HasComplexMemInst = true;
1915           continue;
1916         }
1917         if (!Ld->isSimple() && !IsAnnotatedParallel) {
1918           recordAnalysis("NonSimpleLoad", Ld)
1919               << "read with atomic ordering or volatile read";
1920           LLVM_DEBUG(dbgs() << "LAA: Found a non-simple load.\n");
1921           HasComplexMemInst = true;
1922           continue;
1923         }
1924         NumLoads++;
1925         Loads.push_back(Ld);
1926         DepChecker->addAccess(Ld);
1927         if (EnableMemAccessVersioningOfLoop)
1928           collectStridedAccess(Ld);
1929         continue;
1930       }
1931 
1932       // Save 'store' instructions. Abort if other instructions write to memory.
1933       if (I.mayWriteToMemory()) {
1934         auto *St = dyn_cast<StoreInst>(&I);
1935         if (!St) {
1936           recordAnalysis("CantVectorizeInstruction", St)
1937               << "instruction cannot be vectorized";
1938           HasComplexMemInst = true;
1939           continue;
1940         }
1941         if (!St->isSimple() && !IsAnnotatedParallel) {
1942           recordAnalysis("NonSimpleStore", St)
1943               << "write with atomic ordering or volatile write";
1944           LLVM_DEBUG(dbgs() << "LAA: Found a non-simple store.\n");
1945           HasComplexMemInst = true;
1946           continue;
1947         }
1948         NumStores++;
1949         Stores.push_back(St);
1950         DepChecker->addAccess(St);
1951         if (EnableMemAccessVersioningOfLoop)
1952           collectStridedAccess(St);
1953       }
1954     } // Next instr.
1955   } // Next block.
1956 
1957   if (HasComplexMemInst) {
1958     CanVecMem = false;
1959     return;
1960   }
1961 
1962   // Now we have two lists that hold the loads and the stores.
1963   // Next, we find the pointers that they use.
1964 
1965   // Check if we see any stores. If there are no stores, then we don't
1966   // care if the pointers are *restrict*.
1967   if (!Stores.size()) {
1968     LLVM_DEBUG(dbgs() << "LAA: Found a read-only loop!\n");
1969     CanVecMem = true;
1970     return;
1971   }
1972 
1973   MemoryDepChecker::DepCandidates DependentAccesses;
1974   AccessAnalysis Accesses(TheLoop, AA, LI, DependentAccesses, *PSE);
1975 
1976   // Holds the analyzed pointers. We don't want to call getUnderlyingObjects
1977   // multiple times on the same object. If the ptr is accessed twice, once
1978   // for read and once for write, it will only appear once (on the write
1979   // list). This is okay, since we are going to check for conflicts between
1980   // writes and between reads and writes, but not between reads and reads.
1981   ValueSet Seen;
1982 
1983   // Record uniform store addresses to identify if we have multiple stores
1984   // to the same address.
1985   ValueSet UniformStores;
1986 
1987   for (StoreInst *ST : Stores) {
1988     Value *Ptr = ST->getPointerOperand();
1989 
1990     if (isUniform(Ptr))
1991       HasDependenceInvolvingLoopInvariantAddress |=
1992           !UniformStores.insert(Ptr).second;
1993 
1994     // If we did *not* see this pointer before, insert it to  the read-write
1995     // list. At this phase it is only a 'write' list.
1996     if (Seen.insert(Ptr).second) {
1997       ++NumReadWrites;
1998 
1999       MemoryLocation Loc = MemoryLocation::get(ST);
2000       // The TBAA metadata could have a control dependency on the predication
2001       // condition, so we cannot rely on it when determining whether or not we
2002       // need runtime pointer checks.
2003       if (blockNeedsPredication(ST->getParent(), TheLoop, DT))
2004         Loc.AATags.TBAA = nullptr;
2005 
2006       visitPointers(const_cast<Value *>(Loc.Ptr), *TheLoop,
2007                     [&Accesses, Loc](Value *Ptr) {
2008                       MemoryLocation NewLoc = Loc.getWithNewPtr(Ptr);
2009                       Accesses.addStore(NewLoc);
2010                     });
2011     }
2012   }
2013 
2014   if (IsAnnotatedParallel) {
2015     LLVM_DEBUG(
2016         dbgs() << "LAA: A loop annotated parallel, ignore memory dependency "
2017                << "checks.\n");
2018     CanVecMem = true;
2019     return;
2020   }
2021 
2022   for (LoadInst *LD : Loads) {
2023     Value *Ptr = LD->getPointerOperand();
2024     // If we did *not* see this pointer before, insert it to the
2025     // read list. If we *did* see it before, then it is already in
2026     // the read-write list. This allows us to vectorize expressions
2027     // such as A[i] += x;  Because the address of A[i] is a read-write
2028     // pointer. This only works if the index of A[i] is consecutive.
2029     // If the address of i is unknown (for example A[B[i]]) then we may
2030     // read a few words, modify, and write a few words, and some of the
2031     // words may be written to the same address.
2032     bool IsReadOnlyPtr = false;
2033     if (Seen.insert(Ptr).second ||
2034         !getPtrStride(*PSE, LD->getType(), Ptr, TheLoop, SymbolicStrides)) {
2035       ++NumReads;
2036       IsReadOnlyPtr = true;
2037     }
2038 
2039     // See if there is an unsafe dependency between a load to a uniform address and
2040     // store to the same uniform address.
2041     if (UniformStores.count(Ptr)) {
2042       LLVM_DEBUG(dbgs() << "LAA: Found an unsafe dependency between a uniform "
2043                            "load and uniform store to the same address!\n");
2044       HasDependenceInvolvingLoopInvariantAddress = true;
2045     }
2046 
2047     MemoryLocation Loc = MemoryLocation::get(LD);
2048     // The TBAA metadata could have a control dependency on the predication
2049     // condition, so we cannot rely on it when determining whether or not we
2050     // need runtime pointer checks.
2051     if (blockNeedsPredication(LD->getParent(), TheLoop, DT))
2052       Loc.AATags.TBAA = nullptr;
2053 
2054     visitPointers(const_cast<Value *>(Loc.Ptr), *TheLoop,
2055                   [&Accesses, Loc, IsReadOnlyPtr](Value *Ptr) {
2056                     MemoryLocation NewLoc = Loc.getWithNewPtr(Ptr);
2057                     Accesses.addLoad(NewLoc, IsReadOnlyPtr);
2058                   });
2059   }
2060 
2061   // If we write (or read-write) to a single destination and there are no
2062   // other reads in this loop then is it safe to vectorize.
2063   if (NumReadWrites == 1 && NumReads == 0) {
2064     LLVM_DEBUG(dbgs() << "LAA: Found a write-only loop!\n");
2065     CanVecMem = true;
2066     return;
2067   }
2068 
2069   // Build dependence sets and check whether we need a runtime pointer bounds
2070   // check.
2071   Accesses.buildDependenceSets();
2072 
2073   // Find pointers with computable bounds. We are going to use this information
2074   // to place a runtime bound check.
2075   bool CanDoRTIfNeeded = Accesses.canCheckPtrAtRT(*PtrRtChecking, PSE->getSE(),
2076                                                   TheLoop, SymbolicStrides);
2077   if (!CanDoRTIfNeeded) {
2078     recordAnalysis("CantIdentifyArrayBounds") << "cannot identify array bounds";
2079     LLVM_DEBUG(dbgs() << "LAA: We can't vectorize because we can't find "
2080                       << "the array bounds.\n");
2081     CanVecMem = false;
2082     return;
2083   }
2084 
2085   LLVM_DEBUG(
2086     dbgs() << "LAA: May be able to perform a memory runtime check if needed.\n");
2087 
2088   CanVecMem = true;
2089   if (Accesses.isDependencyCheckNeeded()) {
2090     LLVM_DEBUG(dbgs() << "LAA: Checking memory dependencies\n");
2091     CanVecMem = DepChecker->areDepsSafe(
2092         DependentAccesses, Accesses.getDependenciesToCheck(), SymbolicStrides);
2093     MaxSafeDepDistBytes = DepChecker->getMaxSafeDepDistBytes();
2094 
2095     if (!CanVecMem && DepChecker->shouldRetryWithRuntimeCheck()) {
2096       LLVM_DEBUG(dbgs() << "LAA: Retrying with memory checks\n");
2097 
2098       // Clear the dependency checks. We assume they are not needed.
2099       Accesses.resetDepChecks(*DepChecker);
2100 
2101       PtrRtChecking->reset();
2102       PtrRtChecking->Need = true;
2103 
2104       auto *SE = PSE->getSE();
2105       CanDoRTIfNeeded = Accesses.canCheckPtrAtRT(*PtrRtChecking, SE, TheLoop,
2106                                                  SymbolicStrides, true);
2107 
2108       // Check that we found the bounds for the pointer.
2109       if (!CanDoRTIfNeeded) {
2110         recordAnalysis("CantCheckMemDepsAtRunTime")
2111             << "cannot check memory dependencies at runtime";
2112         LLVM_DEBUG(dbgs() << "LAA: Can't vectorize with memory checks\n");
2113         CanVecMem = false;
2114         return;
2115       }
2116 
2117       CanVecMem = true;
2118     }
2119   }
2120 
2121   if (HasConvergentOp) {
2122     recordAnalysis("CantInsertRuntimeCheckWithConvergent")
2123       << "cannot add control dependency to convergent operation";
2124     LLVM_DEBUG(dbgs() << "LAA: We can't vectorize because a runtime check "
2125                          "would be needed with a convergent operation\n");
2126     CanVecMem = false;
2127     return;
2128   }
2129 
2130   if (CanVecMem)
2131     LLVM_DEBUG(
2132         dbgs() << "LAA: No unsafe dependent memory operations in loop.  We"
2133                << (PtrRtChecking->Need ? "" : " don't")
2134                << " need runtime memory checks.\n");
2135   else {
2136     recordAnalysis("UnsafeMemDep")
2137         << "unsafe dependent memory operations in loop. Use "
2138            "#pragma loop distribute(enable) to allow loop distribution "
2139            "to attempt to isolate the offending operations into a separate "
2140            "loop";
2141     LLVM_DEBUG(dbgs() << "LAA: unsafe dependent memory operations in loop\n");
2142   }
2143 }
2144 
blockNeedsPredication(BasicBlock * BB,Loop * TheLoop,DominatorTree * DT)2145 bool LoopAccessInfo::blockNeedsPredication(BasicBlock *BB, Loop *TheLoop,
2146                                            DominatorTree *DT)  {
2147   assert(TheLoop->contains(BB) && "Unknown block used");
2148 
2149   // Blocks that do not dominate the latch need predication.
2150   BasicBlock* Latch = TheLoop->getLoopLatch();
2151   return !DT->dominates(BB, Latch);
2152 }
2153 
recordAnalysis(StringRef RemarkName,Instruction * I)2154 OptimizationRemarkAnalysis &LoopAccessInfo::recordAnalysis(StringRef RemarkName,
2155                                                            Instruction *I) {
2156   assert(!Report && "Multiple reports generated");
2157 
2158   Value *CodeRegion = TheLoop->getHeader();
2159   DebugLoc DL = TheLoop->getStartLoc();
2160 
2161   if (I) {
2162     CodeRegion = I->getParent();
2163     // If there is no debug location attached to the instruction, revert back to
2164     // using the loop's.
2165     if (I->getDebugLoc())
2166       DL = I->getDebugLoc();
2167   }
2168 
2169   Report = std::make_unique<OptimizationRemarkAnalysis>(DEBUG_TYPE, RemarkName, DL,
2170                                                    CodeRegion);
2171   return *Report;
2172 }
2173 
isUniform(Value * V) const2174 bool LoopAccessInfo::isUniform(Value *V) const {
2175   auto *SE = PSE->getSE();
2176   // Since we rely on SCEV for uniformity, if the type is not SCEVable, it is
2177   // never considered uniform.
2178   // TODO: Is this really what we want? Even without FP SCEV, we may want some
2179   // trivially loop-invariant FP values to be considered uniform.
2180   if (!SE->isSCEVable(V->getType()))
2181     return false;
2182   return (SE->isLoopInvariant(SE->getSCEV(V), TheLoop));
2183 }
2184 
collectStridedAccess(Value * MemAccess)2185 void LoopAccessInfo::collectStridedAccess(Value *MemAccess) {
2186   Value *Ptr = getLoadStorePointerOperand(MemAccess);
2187   if (!Ptr)
2188     return;
2189 
2190   Value *Stride = getStrideFromPointer(Ptr, PSE->getSE(), TheLoop);
2191   if (!Stride)
2192     return;
2193 
2194   LLVM_DEBUG(dbgs() << "LAA: Found a strided access that is a candidate for "
2195                        "versioning:");
2196   LLVM_DEBUG(dbgs() << "  Ptr: " << *Ptr << " Stride: " << *Stride << "\n");
2197 
2198   // Avoid adding the "Stride == 1" predicate when we know that
2199   // Stride >= Trip-Count. Such a predicate will effectively optimize a single
2200   // or zero iteration loop, as Trip-Count <= Stride == 1.
2201   //
2202   // TODO: We are currently not making a very informed decision on when it is
2203   // beneficial to apply stride versioning. It might make more sense that the
2204   // users of this analysis (such as the vectorizer) will trigger it, based on
2205   // their specific cost considerations; For example, in cases where stride
2206   // versioning does  not help resolving memory accesses/dependences, the
2207   // vectorizer should evaluate the cost of the runtime test, and the benefit
2208   // of various possible stride specializations, considering the alternatives
2209   // of using gather/scatters (if available).
2210 
2211   const SCEV *StrideExpr = PSE->getSCEV(Stride);
2212   const SCEV *BETakenCount = PSE->getBackedgeTakenCount();
2213 
2214   // Match the types so we can compare the stride and the BETakenCount.
2215   // The Stride can be positive/negative, so we sign extend Stride;
2216   // The backedgeTakenCount is non-negative, so we zero extend BETakenCount.
2217   const DataLayout &DL = TheLoop->getHeader()->getModule()->getDataLayout();
2218   uint64_t StrideTypeSize = DL.getTypeAllocSize(StrideExpr->getType());
2219   uint64_t BETypeSize = DL.getTypeAllocSize(BETakenCount->getType());
2220   const SCEV *CastedStride = StrideExpr;
2221   const SCEV *CastedBECount = BETakenCount;
2222   ScalarEvolution *SE = PSE->getSE();
2223   if (BETypeSize >= StrideTypeSize)
2224     CastedStride = SE->getNoopOrSignExtend(StrideExpr, BETakenCount->getType());
2225   else
2226     CastedBECount = SE->getZeroExtendExpr(BETakenCount, StrideExpr->getType());
2227   const SCEV *StrideMinusBETaken = SE->getMinusSCEV(CastedStride, CastedBECount);
2228   // Since TripCount == BackEdgeTakenCount + 1, checking:
2229   // "Stride >= TripCount" is equivalent to checking:
2230   // Stride - BETakenCount > 0
2231   if (SE->isKnownPositive(StrideMinusBETaken)) {
2232     LLVM_DEBUG(
2233         dbgs() << "LAA: Stride>=TripCount; No point in versioning as the "
2234                   "Stride==1 predicate will imply that the loop executes "
2235                   "at most once.\n");
2236     return;
2237   }
2238   LLVM_DEBUG(dbgs() << "LAA: Found a strided access that we can version.");
2239 
2240   SymbolicStrides[Ptr] = Stride;
2241   StrideSet.insert(Stride);
2242 }
2243 
LoopAccessInfo(Loop * L,ScalarEvolution * SE,const TargetLibraryInfo * TLI,AAResults * AA,DominatorTree * DT,LoopInfo * LI)2244 LoopAccessInfo::LoopAccessInfo(Loop *L, ScalarEvolution *SE,
2245                                const TargetLibraryInfo *TLI, AAResults *AA,
2246                                DominatorTree *DT, LoopInfo *LI)
2247     : PSE(std::make_unique<PredicatedScalarEvolution>(*SE, *L)),
2248       PtrRtChecking(std::make_unique<RuntimePointerChecking>(SE)),
2249       DepChecker(std::make_unique<MemoryDepChecker>(*PSE, L)), TheLoop(L),
2250       NumLoads(0), NumStores(0), MaxSafeDepDistBytes(-1), CanVecMem(false),
2251       HasConvergentOp(false),
2252       HasDependenceInvolvingLoopInvariantAddress(false) {
2253   if (canAnalyzeLoop())
2254     analyzeLoop(AA, LI, TLI, DT);
2255 }
2256 
print(raw_ostream & OS,unsigned Depth) const2257 void LoopAccessInfo::print(raw_ostream &OS, unsigned Depth) const {
2258   if (CanVecMem) {
2259     OS.indent(Depth) << "Memory dependences are safe";
2260     if (MaxSafeDepDistBytes != -1ULL)
2261       OS << " with a maximum dependence distance of " << MaxSafeDepDistBytes
2262          << " bytes";
2263     if (PtrRtChecking->Need)
2264       OS << " with run-time checks";
2265     OS << "\n";
2266   }
2267 
2268   if (HasConvergentOp)
2269     OS.indent(Depth) << "Has convergent operation in loop\n";
2270 
2271   if (Report)
2272     OS.indent(Depth) << "Report: " << Report->getMsg() << "\n";
2273 
2274   if (auto *Dependences = DepChecker->getDependences()) {
2275     OS.indent(Depth) << "Dependences:\n";
2276     for (auto &Dep : *Dependences) {
2277       Dep.print(OS, Depth + 2, DepChecker->getMemoryInstructions());
2278       OS << "\n";
2279     }
2280   } else
2281     OS.indent(Depth) << "Too many dependences, not recorded\n";
2282 
2283   // List the pair of accesses need run-time checks to prove independence.
2284   PtrRtChecking->print(OS, Depth);
2285   OS << "\n";
2286 
2287   OS.indent(Depth) << "Non vectorizable stores to invariant address were "
2288                    << (HasDependenceInvolvingLoopInvariantAddress ? "" : "not ")
2289                    << "found in loop.\n";
2290 
2291   OS.indent(Depth) << "SCEV assumptions:\n";
2292   PSE->getUnionPredicate().print(OS, Depth);
2293 
2294   OS << "\n";
2295 
2296   OS.indent(Depth) << "Expressions re-written:\n";
2297   PSE->print(OS, Depth);
2298 }
2299 
LoopAccessLegacyAnalysis()2300 LoopAccessLegacyAnalysis::LoopAccessLegacyAnalysis() : FunctionPass(ID) {
2301   initializeLoopAccessLegacyAnalysisPass(*PassRegistry::getPassRegistry());
2302 }
2303 
getInfo(Loop * L)2304 const LoopAccessInfo &LoopAccessLegacyAnalysis::getInfo(Loop *L) {
2305   auto &LAI = LoopAccessInfoMap[L];
2306 
2307   if (!LAI)
2308     LAI = std::make_unique<LoopAccessInfo>(L, SE, TLI, AA, DT, LI);
2309 
2310   return *LAI.get();
2311 }
2312 
print(raw_ostream & OS,const Module * M) const2313 void LoopAccessLegacyAnalysis::print(raw_ostream &OS, const Module *M) const {
2314   LoopAccessLegacyAnalysis &LAA = *const_cast<LoopAccessLegacyAnalysis *>(this);
2315 
2316   for (Loop *TopLevelLoop : *LI)
2317     for (Loop *L : depth_first(TopLevelLoop)) {
2318       OS.indent(2) << L->getHeader()->getName() << ":\n";
2319       auto &LAI = LAA.getInfo(L);
2320       LAI.print(OS, 4);
2321     }
2322 }
2323 
runOnFunction(Function & F)2324 bool LoopAccessLegacyAnalysis::runOnFunction(Function &F) {
2325   SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE();
2326   auto *TLIP = getAnalysisIfAvailable<TargetLibraryInfoWrapperPass>();
2327   TLI = TLIP ? &TLIP->getTLI(F) : nullptr;
2328   AA = &getAnalysis<AAResultsWrapperPass>().getAAResults();
2329   DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
2330   LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
2331 
2332   return false;
2333 }
2334 
getAnalysisUsage(AnalysisUsage & AU) const2335 void LoopAccessLegacyAnalysis::getAnalysisUsage(AnalysisUsage &AU) const {
2336   AU.addRequiredTransitive<ScalarEvolutionWrapperPass>();
2337   AU.addRequiredTransitive<AAResultsWrapperPass>();
2338   AU.addRequiredTransitive<DominatorTreeWrapperPass>();
2339   AU.addRequiredTransitive<LoopInfoWrapperPass>();
2340 
2341   AU.setPreservesAll();
2342 }
2343 
2344 char LoopAccessLegacyAnalysis::ID = 0;
2345 static const char laa_name[] = "Loop Access Analysis";
2346 #define LAA_NAME "loop-accesses"
2347 
2348 INITIALIZE_PASS_BEGIN(LoopAccessLegacyAnalysis, LAA_NAME, laa_name, false, true)
2349 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)
2350 INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass)
2351 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
2352 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
2353 INITIALIZE_PASS_END(LoopAccessLegacyAnalysis, LAA_NAME, laa_name, false, true)
2354 
2355 AnalysisKey LoopAccessAnalysis::Key;
2356 
run(Loop & L,LoopAnalysisManager & AM,LoopStandardAnalysisResults & AR)2357 LoopAccessInfo LoopAccessAnalysis::run(Loop &L, LoopAnalysisManager &AM,
2358                                        LoopStandardAnalysisResults &AR) {
2359   return LoopAccessInfo(&L, &AR.SE, &AR.TLI, &AR.AA, &AR.DT, &AR.LI);
2360 }
2361 
2362 namespace llvm {
2363 
createLAAPass()2364   Pass *createLAAPass() {
2365     return new LoopAccessLegacyAnalysis();
2366   }
2367 
2368 } // end namespace llvm
2369