• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 //===- LoopAccessAnalysis.cpp - Loop Access Analysis Implementation --------==//
2 //
3 //                     The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // The implementation for the loop memory dependence that was originally
11 // developed for the loop vectorizer.
12 //
13 //===----------------------------------------------------------------------===//
14 
15 #include "llvm/Analysis/LoopAccessAnalysis.h"
16 #include "llvm/Analysis/LoopInfo.h"
17 #include "llvm/Analysis/LoopPassManager.h"
18 #include "llvm/Analysis/ScalarEvolutionExpander.h"
19 #include "llvm/Analysis/TargetLibraryInfo.h"
20 #include "llvm/Analysis/ValueTracking.h"
21 #include "llvm/Analysis/VectorUtils.h"
22 #include "llvm/IR/DiagnosticInfo.h"
23 #include "llvm/IR/Dominators.h"
24 #include "llvm/IR/IRBuilder.h"
25 #include "llvm/IR/PassManager.h"
26 #include "llvm/Support/Debug.h"
27 #include "llvm/Support/raw_ostream.h"
28 using namespace llvm;
29 
30 #define DEBUG_TYPE "loop-accesses"
31 
32 static cl::opt<unsigned, true>
33 VectorizationFactor("force-vector-width", cl::Hidden,
34                     cl::desc("Sets the SIMD width. Zero is autoselect."),
35                     cl::location(VectorizerParams::VectorizationFactor));
36 unsigned VectorizerParams::VectorizationFactor;
37 
38 static cl::opt<unsigned, true>
39 VectorizationInterleave("force-vector-interleave", cl::Hidden,
40                         cl::desc("Sets the vectorization interleave count. "
41                                  "Zero is autoselect."),
42                         cl::location(
43                             VectorizerParams::VectorizationInterleave));
44 unsigned VectorizerParams::VectorizationInterleave;
45 
46 static cl::opt<unsigned, true> RuntimeMemoryCheckThreshold(
47     "runtime-memory-check-threshold", cl::Hidden,
48     cl::desc("When performing memory disambiguation checks at runtime do not "
49              "generate more than this number of comparisons (default = 8)."),
50     cl::location(VectorizerParams::RuntimeMemoryCheckThreshold), cl::init(8));
51 unsigned VectorizerParams::RuntimeMemoryCheckThreshold;
52 
53 /// \brief The maximum iterations used to merge memory checks
54 static cl::opt<unsigned> MemoryCheckMergeThreshold(
55     "memory-check-merge-threshold", cl::Hidden,
56     cl::desc("Maximum number of comparisons done when trying to merge "
57              "runtime memory checks. (default = 100)"),
58     cl::init(100));
59 
60 /// Maximum SIMD width.
61 const unsigned VectorizerParams::MaxVectorWidth = 64;
62 
63 /// \brief We collect dependences up to this threshold.
64 static cl::opt<unsigned>
65     MaxDependences("max-dependences", cl::Hidden,
66                    cl::desc("Maximum number of dependences collected by "
67                             "loop-access analysis (default = 100)"),
68                    cl::init(100));
69 
70 /// This enables versioning on the strides of symbolically striding memory
71 /// accesses in code like the following.
72 ///   for (i = 0; i < N; ++i)
73 ///     A[i * Stride1] += B[i * Stride2] ...
74 ///
75 /// Will be roughly translated to
76 ///    if (Stride1 == 1 && Stride2 == 1) {
77 ///      for (i = 0; i < N; i+=4)
78 ///       A[i:i+3] += ...
79 ///    } else
80 ///      ...
81 static cl::opt<bool> EnableMemAccessVersioning(
82     "enable-mem-access-versioning", cl::init(true), cl::Hidden,
83     cl::desc("Enable symbolic stride memory access versioning"));
84 
85 /// \brief Enable store-to-load forwarding conflict detection. This option can
86 /// be disabled for correctness testing.
87 static cl::opt<bool> EnableForwardingConflictDetection(
88     "store-to-load-forwarding-conflict-detection", cl::Hidden,
89     cl::desc("Enable conflict detection in loop-access analysis"),
90     cl::init(true));
91 
isInterleaveForced()92 bool VectorizerParams::isInterleaveForced() {
93   return ::VectorizationInterleave.getNumOccurrences() > 0;
94 }
95 
emitAnalysis(const LoopAccessReport & Message,const Function * TheFunction,const Loop * TheLoop,const char * PassName)96 void LoopAccessReport::emitAnalysis(const LoopAccessReport &Message,
97                                     const Function *TheFunction,
98                                     const Loop *TheLoop,
99                                     const char *PassName) {
100   DebugLoc DL = TheLoop->getStartLoc();
101   if (const Instruction *I = Message.getInstr())
102     DL = I->getDebugLoc();
103   emitOptimizationRemarkAnalysis(TheFunction->getContext(), PassName,
104                                  *TheFunction, DL, Message.str());
105 }
106 
stripIntegerCast(Value * V)107 Value *llvm::stripIntegerCast(Value *V) {
108   if (auto *CI = dyn_cast<CastInst>(V))
109     if (CI->getOperand(0)->getType()->isIntegerTy())
110       return CI->getOperand(0);
111   return V;
112 }
113 
replaceSymbolicStrideSCEV(PredicatedScalarEvolution & PSE,const ValueToValueMap & PtrToStride,Value * Ptr,Value * OrigPtr)114 const SCEV *llvm::replaceSymbolicStrideSCEV(PredicatedScalarEvolution &PSE,
115                                             const ValueToValueMap &PtrToStride,
116                                             Value *Ptr, Value *OrigPtr) {
117   const SCEV *OrigSCEV = PSE.getSCEV(Ptr);
118 
119   // If there is an entry in the map return the SCEV of the pointer with the
120   // symbolic stride replaced by one.
121   ValueToValueMap::const_iterator SI =
122       PtrToStride.find(OrigPtr ? OrigPtr : Ptr);
123   if (SI != PtrToStride.end()) {
124     Value *StrideVal = SI->second;
125 
126     // Strip casts.
127     StrideVal = stripIntegerCast(StrideVal);
128 
129     // Replace symbolic stride by one.
130     Value *One = ConstantInt::get(StrideVal->getType(), 1);
131     ValueToValueMap RewriteMap;
132     RewriteMap[StrideVal] = One;
133 
134     ScalarEvolution *SE = PSE.getSE();
135     const auto *U = cast<SCEVUnknown>(SE->getSCEV(StrideVal));
136     const auto *CT =
137         static_cast<const SCEVConstant *>(SE->getOne(StrideVal->getType()));
138 
139     PSE.addPredicate(*SE->getEqualPredicate(U, CT));
140     auto *Expr = PSE.getSCEV(Ptr);
141 
142     DEBUG(dbgs() << "LAA: Replacing SCEV: " << *OrigSCEV << " by: " << *Expr
143                  << "\n");
144     return Expr;
145   }
146 
147   // Otherwise, just return the SCEV of the original pointer.
148   return OrigSCEV;
149 }
150 
insert(Loop * Lp,Value * Ptr,bool WritePtr,unsigned DepSetId,unsigned ASId,const ValueToValueMap & Strides,PredicatedScalarEvolution & PSE)151 void RuntimePointerChecking::insert(Loop *Lp, Value *Ptr, bool WritePtr,
152                                     unsigned DepSetId, unsigned ASId,
153                                     const ValueToValueMap &Strides,
154                                     PredicatedScalarEvolution &PSE) {
155   // Get the stride replaced scev.
156   const SCEV *Sc = replaceSymbolicStrideSCEV(PSE, Strides, Ptr);
157   ScalarEvolution *SE = PSE.getSE();
158 
159   const SCEV *ScStart;
160   const SCEV *ScEnd;
161 
162   if (SE->isLoopInvariant(Sc, Lp))
163     ScStart = ScEnd = Sc;
164   else {
165     const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Sc);
166     assert(AR && "Invalid addrec expression");
167     const SCEV *Ex = PSE.getBackedgeTakenCount();
168 
169     ScStart = AR->getStart();
170     ScEnd = AR->evaluateAtIteration(Ex, *SE);
171     const SCEV *Step = AR->getStepRecurrence(*SE);
172 
173     // For expressions with negative step, the upper bound is ScStart and the
174     // lower bound is ScEnd.
175     if (const auto *CStep = dyn_cast<SCEVConstant>(Step)) {
176       if (CStep->getValue()->isNegative())
177         std::swap(ScStart, ScEnd);
178     } else {
179       // Fallback case: the step is not constant, but the we can still
180       // get the upper and lower bounds of the interval by using min/max
181       // expressions.
182       ScStart = SE->getUMinExpr(ScStart, ScEnd);
183       ScEnd = SE->getUMaxExpr(AR->getStart(), ScEnd);
184     }
185   }
186 
187   Pointers.emplace_back(Ptr, ScStart, ScEnd, WritePtr, DepSetId, ASId, Sc);
188 }
189 
190 SmallVector<RuntimePointerChecking::PointerCheck, 4>
generateChecks() const191 RuntimePointerChecking::generateChecks() const {
192   SmallVector<PointerCheck, 4> Checks;
193 
194   for (unsigned I = 0; I < CheckingGroups.size(); ++I) {
195     for (unsigned J = I + 1; J < CheckingGroups.size(); ++J) {
196       const RuntimePointerChecking::CheckingPtrGroup &CGI = CheckingGroups[I];
197       const RuntimePointerChecking::CheckingPtrGroup &CGJ = CheckingGroups[J];
198 
199       if (needsChecking(CGI, CGJ))
200         Checks.push_back(std::make_pair(&CGI, &CGJ));
201     }
202   }
203   return Checks;
204 }
205 
generateChecks(MemoryDepChecker::DepCandidates & DepCands,bool UseDependencies)206 void RuntimePointerChecking::generateChecks(
207     MemoryDepChecker::DepCandidates &DepCands, bool UseDependencies) {
208   assert(Checks.empty() && "Checks is not empty");
209   groupChecks(DepCands, UseDependencies);
210   Checks = generateChecks();
211 }
212 
needsChecking(const CheckingPtrGroup & M,const CheckingPtrGroup & N) const213 bool RuntimePointerChecking::needsChecking(const CheckingPtrGroup &M,
214                                            const CheckingPtrGroup &N) const {
215   for (unsigned I = 0, EI = M.Members.size(); EI != I; ++I)
216     for (unsigned J = 0, EJ = N.Members.size(); EJ != J; ++J)
217       if (needsChecking(M.Members[I], N.Members[J]))
218         return true;
219   return false;
220 }
221 
222 /// Compare \p I and \p J and return the minimum.
223 /// Return nullptr in case we couldn't find an answer.
getMinFromExprs(const SCEV * I,const SCEV * J,ScalarEvolution * SE)224 static const SCEV *getMinFromExprs(const SCEV *I, const SCEV *J,
225                                    ScalarEvolution *SE) {
226   const SCEV *Diff = SE->getMinusSCEV(J, I);
227   const SCEVConstant *C = dyn_cast<const SCEVConstant>(Diff);
228 
229   if (!C)
230     return nullptr;
231   if (C->getValue()->isNegative())
232     return J;
233   return I;
234 }
235 
addPointer(unsigned Index)236 bool RuntimePointerChecking::CheckingPtrGroup::addPointer(unsigned Index) {
237   const SCEV *Start = RtCheck.Pointers[Index].Start;
238   const SCEV *End = RtCheck.Pointers[Index].End;
239 
240   // Compare the starts and ends with the known minimum and maximum
241   // of this set. We need to know how we compare against the min/max
242   // of the set in order to be able to emit memchecks.
243   const SCEV *Min0 = getMinFromExprs(Start, Low, RtCheck.SE);
244   if (!Min0)
245     return false;
246 
247   const SCEV *Min1 = getMinFromExprs(End, High, RtCheck.SE);
248   if (!Min1)
249     return false;
250 
251   // Update the low bound  expression if we've found a new min value.
252   if (Min0 == Start)
253     Low = Start;
254 
255   // Update the high bound expression if we've found a new max value.
256   if (Min1 != End)
257     High = End;
258 
259   Members.push_back(Index);
260   return true;
261 }
262 
groupChecks(MemoryDepChecker::DepCandidates & DepCands,bool UseDependencies)263 void RuntimePointerChecking::groupChecks(
264     MemoryDepChecker::DepCandidates &DepCands, bool UseDependencies) {
265   // We build the groups from dependency candidates equivalence classes
266   // because:
267   //    - We know that pointers in the same equivalence class share
268   //      the same underlying object and therefore there is a chance
269   //      that we can compare pointers
270   //    - We wouldn't be able to merge two pointers for which we need
271   //      to emit a memcheck. The classes in DepCands are already
272   //      conveniently built such that no two pointers in the same
273   //      class need checking against each other.
274 
275   // We use the following (greedy) algorithm to construct the groups
276   // For every pointer in the equivalence class:
277   //   For each existing group:
278   //   - if the difference between this pointer and the min/max bounds
279   //     of the group is a constant, then make the pointer part of the
280   //     group and update the min/max bounds of that group as required.
281 
282   CheckingGroups.clear();
283 
284   // If we need to check two pointers to the same underlying object
285   // with a non-constant difference, we shouldn't perform any pointer
286   // grouping with those pointers. This is because we can easily get
287   // into cases where the resulting check would return false, even when
288   // the accesses are safe.
289   //
290   // The following example shows this:
291   // for (i = 0; i < 1000; ++i)
292   //   a[5000 + i * m] = a[i] + a[i + 9000]
293   //
294   // Here grouping gives a check of (5000, 5000 + 1000 * m) against
295   // (0, 10000) which is always false. However, if m is 1, there is no
296   // dependence. Not grouping the checks for a[i] and a[i + 9000] allows
297   // us to perform an accurate check in this case.
298   //
299   // The above case requires that we have an UnknownDependence between
300   // accesses to the same underlying object. This cannot happen unless
301   // ShouldRetryWithRuntimeCheck is set, and therefore UseDependencies
302   // is also false. In this case we will use the fallback path and create
303   // separate checking groups for all pointers.
304 
305   // If we don't have the dependency partitions, construct a new
306   // checking pointer group for each pointer. This is also required
307   // for correctness, because in this case we can have checking between
308   // pointers to the same underlying object.
309   if (!UseDependencies) {
310     for (unsigned I = 0; I < Pointers.size(); ++I)
311       CheckingGroups.push_back(CheckingPtrGroup(I, *this));
312     return;
313   }
314 
315   unsigned TotalComparisons = 0;
316 
317   DenseMap<Value *, unsigned> PositionMap;
318   for (unsigned Index = 0; Index < Pointers.size(); ++Index)
319     PositionMap[Pointers[Index].PointerValue] = Index;
320 
321   // We need to keep track of what pointers we've already seen so we
322   // don't process them twice.
323   SmallSet<unsigned, 2> Seen;
324 
325   // Go through all equivalence classes, get the "pointer check groups"
326   // and add them to the overall solution. We use the order in which accesses
327   // appear in 'Pointers' to enforce determinism.
328   for (unsigned I = 0; I < Pointers.size(); ++I) {
329     // We've seen this pointer before, and therefore already processed
330     // its equivalence class.
331     if (Seen.count(I))
332       continue;
333 
334     MemoryDepChecker::MemAccessInfo Access(Pointers[I].PointerValue,
335                                            Pointers[I].IsWritePtr);
336 
337     SmallVector<CheckingPtrGroup, 2> Groups;
338     auto LeaderI = DepCands.findValue(DepCands.getLeaderValue(Access));
339 
340     // Because DepCands is constructed by visiting accesses in the order in
341     // which they appear in alias sets (which is deterministic) and the
342     // iteration order within an equivalence class member is only dependent on
343     // the order in which unions and insertions are performed on the
344     // equivalence class, the iteration order is deterministic.
345     for (auto MI = DepCands.member_begin(LeaderI), ME = DepCands.member_end();
346          MI != ME; ++MI) {
347       unsigned Pointer = PositionMap[MI->getPointer()];
348       bool Merged = false;
349       // Mark this pointer as seen.
350       Seen.insert(Pointer);
351 
352       // Go through all the existing sets and see if we can find one
353       // which can include this pointer.
354       for (CheckingPtrGroup &Group : Groups) {
355         // Don't perform more than a certain amount of comparisons.
356         // This should limit the cost of grouping the pointers to something
357         // reasonable.  If we do end up hitting this threshold, the algorithm
358         // will create separate groups for all remaining pointers.
359         if (TotalComparisons > MemoryCheckMergeThreshold)
360           break;
361 
362         TotalComparisons++;
363 
364         if (Group.addPointer(Pointer)) {
365           Merged = true;
366           break;
367         }
368       }
369 
370       if (!Merged)
371         // We couldn't add this pointer to any existing set or the threshold
372         // for the number of comparisons has been reached. Create a new group
373         // to hold the current pointer.
374         Groups.push_back(CheckingPtrGroup(Pointer, *this));
375     }
376 
377     // We've computed the grouped checks for this partition.
378     // Save the results and continue with the next one.
379     std::copy(Groups.begin(), Groups.end(), std::back_inserter(CheckingGroups));
380   }
381 }
382 
arePointersInSamePartition(const SmallVectorImpl<int> & PtrToPartition,unsigned PtrIdx1,unsigned PtrIdx2)383 bool RuntimePointerChecking::arePointersInSamePartition(
384     const SmallVectorImpl<int> &PtrToPartition, unsigned PtrIdx1,
385     unsigned PtrIdx2) {
386   return (PtrToPartition[PtrIdx1] != -1 &&
387           PtrToPartition[PtrIdx1] == PtrToPartition[PtrIdx2]);
388 }
389 
needsChecking(unsigned I,unsigned J) const390 bool RuntimePointerChecking::needsChecking(unsigned I, unsigned J) const {
391   const PointerInfo &PointerI = Pointers[I];
392   const PointerInfo &PointerJ = Pointers[J];
393 
394   // No need to check if two readonly pointers intersect.
395   if (!PointerI.IsWritePtr && !PointerJ.IsWritePtr)
396     return false;
397 
398   // Only need to check pointers between two different dependency sets.
399   if (PointerI.DependencySetId == PointerJ.DependencySetId)
400     return false;
401 
402   // Only need to check pointers in the same alias set.
403   if (PointerI.AliasSetId != PointerJ.AliasSetId)
404     return false;
405 
406   return true;
407 }
408 
printChecks(raw_ostream & OS,const SmallVectorImpl<PointerCheck> & Checks,unsigned Depth) const409 void RuntimePointerChecking::printChecks(
410     raw_ostream &OS, const SmallVectorImpl<PointerCheck> &Checks,
411     unsigned Depth) const {
412   unsigned N = 0;
413   for (const auto &Check : Checks) {
414     const auto &First = Check.first->Members, &Second = Check.second->Members;
415 
416     OS.indent(Depth) << "Check " << N++ << ":\n";
417 
418     OS.indent(Depth + 2) << "Comparing group (" << Check.first << "):\n";
419     for (unsigned K = 0; K < First.size(); ++K)
420       OS.indent(Depth + 2) << *Pointers[First[K]].PointerValue << "\n";
421 
422     OS.indent(Depth + 2) << "Against group (" << Check.second << "):\n";
423     for (unsigned K = 0; K < Second.size(); ++K)
424       OS.indent(Depth + 2) << *Pointers[Second[K]].PointerValue << "\n";
425   }
426 }
427 
print(raw_ostream & OS,unsigned Depth) const428 void RuntimePointerChecking::print(raw_ostream &OS, unsigned Depth) const {
429 
430   OS.indent(Depth) << "Run-time memory checks:\n";
431   printChecks(OS, Checks, Depth);
432 
433   OS.indent(Depth) << "Grouped accesses:\n";
434   for (unsigned I = 0; I < CheckingGroups.size(); ++I) {
435     const auto &CG = CheckingGroups[I];
436 
437     OS.indent(Depth + 2) << "Group " << &CG << ":\n";
438     OS.indent(Depth + 4) << "(Low: " << *CG.Low << " High: " << *CG.High
439                          << ")\n";
440     for (unsigned J = 0; J < CG.Members.size(); ++J) {
441       OS.indent(Depth + 6) << "Member: " << *Pointers[CG.Members[J]].Expr
442                            << "\n";
443     }
444   }
445 }
446 
447 namespace {
448 /// \brief Analyses memory accesses in a loop.
449 ///
450 /// Checks whether run time pointer checks are needed and builds sets for data
451 /// dependence checking.
452 class AccessAnalysis {
453 public:
454   /// \brief Read or write access location.
455   typedef PointerIntPair<Value *, 1, bool> MemAccessInfo;
456   typedef SmallPtrSet<MemAccessInfo, 8> MemAccessInfoSet;
457 
AccessAnalysis(const DataLayout & Dl,AliasAnalysis * AA,LoopInfo * LI,MemoryDepChecker::DepCandidates & DA,PredicatedScalarEvolution & PSE)458   AccessAnalysis(const DataLayout &Dl, AliasAnalysis *AA, LoopInfo *LI,
459                  MemoryDepChecker::DepCandidates &DA,
460                  PredicatedScalarEvolution &PSE)
461       : DL(Dl), AST(*AA), LI(LI), DepCands(DA), IsRTCheckAnalysisNeeded(false),
462         PSE(PSE) {}
463 
464   /// \brief Register a load  and whether it is only read from.
addLoad(MemoryLocation & Loc,bool IsReadOnly)465   void addLoad(MemoryLocation &Loc, bool IsReadOnly) {
466     Value *Ptr = const_cast<Value*>(Loc.Ptr);
467     AST.add(Ptr, MemoryLocation::UnknownSize, Loc.AATags);
468     Accesses.insert(MemAccessInfo(Ptr, false));
469     if (IsReadOnly)
470       ReadOnlyPtr.insert(Ptr);
471   }
472 
473   /// \brief Register a store.
addStore(MemoryLocation & Loc)474   void addStore(MemoryLocation &Loc) {
475     Value *Ptr = const_cast<Value*>(Loc.Ptr);
476     AST.add(Ptr, MemoryLocation::UnknownSize, Loc.AATags);
477     Accesses.insert(MemAccessInfo(Ptr, true));
478   }
479 
480   /// \brief Check whether we can check the pointers at runtime for
481   /// non-intersection.
482   ///
483   /// Returns true if we need no check or if we do and we can generate them
484   /// (i.e. the pointers have computable bounds).
485   bool canCheckPtrAtRT(RuntimePointerChecking &RtCheck, ScalarEvolution *SE,
486                        Loop *TheLoop, const ValueToValueMap &Strides,
487                        bool ShouldCheckWrap = false);
488 
489   /// \brief Goes over all memory accesses, checks whether a RT check is needed
490   /// and builds sets of dependent accesses.
buildDependenceSets()491   void buildDependenceSets() {
492     processMemAccesses();
493   }
494 
495   /// \brief Initial processing of memory accesses determined that we need to
496   /// perform dependency checking.
497   ///
498   /// Note that this can later be cleared if we retry memcheck analysis without
499   /// dependency checking (i.e. ShouldRetryWithRuntimeCheck).
isDependencyCheckNeeded()500   bool isDependencyCheckNeeded() { return !CheckDeps.empty(); }
501 
502   /// We decided that no dependence analysis would be used.  Reset the state.
resetDepChecks(MemoryDepChecker & DepChecker)503   void resetDepChecks(MemoryDepChecker &DepChecker) {
504     CheckDeps.clear();
505     DepChecker.clearDependences();
506   }
507 
getDependenciesToCheck()508   MemAccessInfoSet &getDependenciesToCheck() { return CheckDeps; }
509 
510 private:
511   typedef SetVector<MemAccessInfo> PtrAccessSet;
512 
513   /// \brief Go over all memory access and check whether runtime pointer checks
514   /// are needed and build sets of dependency check candidates.
515   void processMemAccesses();
516 
517   /// Set of all accesses.
518   PtrAccessSet Accesses;
519 
520   const DataLayout &DL;
521 
522   /// Set of accesses that need a further dependence check.
523   MemAccessInfoSet CheckDeps;
524 
525   /// Set of pointers that are read only.
526   SmallPtrSet<Value*, 16> ReadOnlyPtr;
527 
528   /// An alias set tracker to partition the access set by underlying object and
529   //intrinsic property (such as TBAA metadata).
530   AliasSetTracker AST;
531 
532   LoopInfo *LI;
533 
534   /// Sets of potentially dependent accesses - members of one set share an
535   /// underlying pointer. The set "CheckDeps" identfies which sets really need a
536   /// dependence check.
537   MemoryDepChecker::DepCandidates &DepCands;
538 
539   /// \brief Initial processing of memory accesses determined that we may need
540   /// to add memchecks.  Perform the analysis to determine the necessary checks.
541   ///
542   /// Note that, this is different from isDependencyCheckNeeded.  When we retry
543   /// memcheck analysis without dependency checking
544   /// (i.e. ShouldRetryWithRuntimeCheck), isDependencyCheckNeeded is cleared
545   /// while this remains set if we have potentially dependent accesses.
546   bool IsRTCheckAnalysisNeeded;
547 
548   /// The SCEV predicate containing all the SCEV-related assumptions.
549   PredicatedScalarEvolution &PSE;
550 };
551 
552 } // end anonymous namespace
553 
554 /// \brief Check whether a pointer can participate in a runtime bounds check.
hasComputableBounds(PredicatedScalarEvolution & PSE,const ValueToValueMap & Strides,Value * Ptr,Loop * L)555 static bool hasComputableBounds(PredicatedScalarEvolution &PSE,
556                                 const ValueToValueMap &Strides, Value *Ptr,
557                                 Loop *L) {
558   const SCEV *PtrScev = replaceSymbolicStrideSCEV(PSE, Strides, Ptr);
559 
560   // The bounds for loop-invariant pointer is trivial.
561   if (PSE.getSE()->isLoopInvariant(PtrScev, L))
562     return true;
563 
564   const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
565   if (!AR)
566     return false;
567 
568   return AR->isAffine();
569 }
570 
571 /// \brief Check whether a pointer address cannot wrap.
isNoWrap(PredicatedScalarEvolution & PSE,const ValueToValueMap & Strides,Value * Ptr,Loop * L)572 static bool isNoWrap(PredicatedScalarEvolution &PSE,
573                      const ValueToValueMap &Strides, Value *Ptr, Loop *L) {
574   const SCEV *PtrScev = PSE.getSCEV(Ptr);
575   if (PSE.getSE()->isLoopInvariant(PtrScev, L))
576     return true;
577 
578   int64_t Stride = getPtrStride(PSE, Ptr, L, Strides);
579   return Stride == 1;
580 }
581 
canCheckPtrAtRT(RuntimePointerChecking & RtCheck,ScalarEvolution * SE,Loop * TheLoop,const ValueToValueMap & StridesMap,bool ShouldCheckWrap)582 bool AccessAnalysis::canCheckPtrAtRT(RuntimePointerChecking &RtCheck,
583                                      ScalarEvolution *SE, Loop *TheLoop,
584                                      const ValueToValueMap &StridesMap,
585                                      bool ShouldCheckWrap) {
586   // Find pointers with computable bounds. We are going to use this information
587   // to place a runtime bound check.
588   bool CanDoRT = true;
589 
590   bool NeedRTCheck = false;
591   if (!IsRTCheckAnalysisNeeded) return true;
592 
593   bool IsDepCheckNeeded = isDependencyCheckNeeded();
594 
595   // We assign a consecutive id to access from different alias sets.
596   // Accesses between different groups doesn't need to be checked.
597   unsigned ASId = 1;
598   for (auto &AS : AST) {
599     int NumReadPtrChecks = 0;
600     int NumWritePtrChecks = 0;
601 
602     // We assign consecutive id to access from different dependence sets.
603     // Accesses within the same set don't need a runtime check.
604     unsigned RunningDepId = 1;
605     DenseMap<Value *, unsigned> DepSetId;
606 
607     for (auto A : AS) {
608       Value *Ptr = A.getValue();
609       bool IsWrite = Accesses.count(MemAccessInfo(Ptr, true));
610       MemAccessInfo Access(Ptr, IsWrite);
611 
612       if (IsWrite)
613         ++NumWritePtrChecks;
614       else
615         ++NumReadPtrChecks;
616 
617       if (hasComputableBounds(PSE, StridesMap, Ptr, TheLoop) &&
618           // When we run after a failing dependency check we have to make sure
619           // we don't have wrapping pointers.
620           (!ShouldCheckWrap || isNoWrap(PSE, StridesMap, Ptr, TheLoop))) {
621         // The id of the dependence set.
622         unsigned DepId;
623 
624         if (IsDepCheckNeeded) {
625           Value *Leader = DepCands.getLeaderValue(Access).getPointer();
626           unsigned &LeaderId = DepSetId[Leader];
627           if (!LeaderId)
628             LeaderId = RunningDepId++;
629           DepId = LeaderId;
630         } else
631           // Each access has its own dependence set.
632           DepId = RunningDepId++;
633 
634         RtCheck.insert(TheLoop, Ptr, IsWrite, DepId, ASId, StridesMap, PSE);
635 
636         DEBUG(dbgs() << "LAA: Found a runtime check ptr:" << *Ptr << '\n');
637       } else {
638         DEBUG(dbgs() << "LAA: Can't find bounds for ptr:" << *Ptr << '\n');
639         CanDoRT = false;
640       }
641     }
642 
643     // If we have at least two writes or one write and a read then we need to
644     // check them.  But there is no need to checks if there is only one
645     // dependence set for this alias set.
646     //
647     // Note that this function computes CanDoRT and NeedRTCheck independently.
648     // For example CanDoRT=false, NeedRTCheck=false means that we have a pointer
649     // for which we couldn't find the bounds but we don't actually need to emit
650     // any checks so it does not matter.
651     if (!(IsDepCheckNeeded && CanDoRT && RunningDepId == 2))
652       NeedRTCheck |= (NumWritePtrChecks >= 2 || (NumReadPtrChecks >= 1 &&
653                                                  NumWritePtrChecks >= 1));
654 
655     ++ASId;
656   }
657 
658   // If the pointers that we would use for the bounds comparison have different
659   // address spaces, assume the values aren't directly comparable, so we can't
660   // use them for the runtime check. We also have to assume they could
661   // overlap. In the future there should be metadata for whether address spaces
662   // are disjoint.
663   unsigned NumPointers = RtCheck.Pointers.size();
664   for (unsigned i = 0; i < NumPointers; ++i) {
665     for (unsigned j = i + 1; j < NumPointers; ++j) {
666       // Only need to check pointers between two different dependency sets.
667       if (RtCheck.Pointers[i].DependencySetId ==
668           RtCheck.Pointers[j].DependencySetId)
669        continue;
670       // Only need to check pointers in the same alias set.
671       if (RtCheck.Pointers[i].AliasSetId != RtCheck.Pointers[j].AliasSetId)
672         continue;
673 
674       Value *PtrI = RtCheck.Pointers[i].PointerValue;
675       Value *PtrJ = RtCheck.Pointers[j].PointerValue;
676 
677       unsigned ASi = PtrI->getType()->getPointerAddressSpace();
678       unsigned ASj = PtrJ->getType()->getPointerAddressSpace();
679       if (ASi != ASj) {
680         DEBUG(dbgs() << "LAA: Runtime check would require comparison between"
681                        " different address spaces\n");
682         return false;
683       }
684     }
685   }
686 
687   if (NeedRTCheck && CanDoRT)
688     RtCheck.generateChecks(DepCands, IsDepCheckNeeded);
689 
690   DEBUG(dbgs() << "LAA: We need to do " << RtCheck.getNumberOfChecks()
691                << " pointer comparisons.\n");
692 
693   RtCheck.Need = NeedRTCheck;
694 
695   bool CanDoRTIfNeeded = !NeedRTCheck || CanDoRT;
696   if (!CanDoRTIfNeeded)
697     RtCheck.reset();
698   return CanDoRTIfNeeded;
699 }
700 
processMemAccesses()701 void AccessAnalysis::processMemAccesses() {
702   // We process the set twice: first we process read-write pointers, last we
703   // process read-only pointers. This allows us to skip dependence tests for
704   // read-only pointers.
705 
706   DEBUG(dbgs() << "LAA: Processing memory accesses...\n");
707   DEBUG(dbgs() << "  AST: "; AST.dump());
708   DEBUG(dbgs() << "LAA:   Accesses(" << Accesses.size() << "):\n");
709   DEBUG({
710     for (auto A : Accesses)
711       dbgs() << "\t" << *A.getPointer() << " (" <<
712                 (A.getInt() ? "write" : (ReadOnlyPtr.count(A.getPointer()) ?
713                                          "read-only" : "read")) << ")\n";
714   });
715 
716   // The AliasSetTracker has nicely partitioned our pointers by metadata
717   // compatibility and potential for underlying-object overlap. As a result, we
718   // only need to check for potential pointer dependencies within each alias
719   // set.
720   for (auto &AS : AST) {
721     // Note that both the alias-set tracker and the alias sets themselves used
722     // linked lists internally and so the iteration order here is deterministic
723     // (matching the original instruction order within each set).
724 
725     bool SetHasWrite = false;
726 
727     // Map of pointers to last access encountered.
728     typedef DenseMap<Value*, MemAccessInfo> UnderlyingObjToAccessMap;
729     UnderlyingObjToAccessMap ObjToLastAccess;
730 
731     // Set of access to check after all writes have been processed.
732     PtrAccessSet DeferredAccesses;
733 
734     // Iterate over each alias set twice, once to process read/write pointers,
735     // and then to process read-only pointers.
736     for (int SetIteration = 0; SetIteration < 2; ++SetIteration) {
737       bool UseDeferred = SetIteration > 0;
738       PtrAccessSet &S = UseDeferred ? DeferredAccesses : Accesses;
739 
740       for (auto AV : AS) {
741         Value *Ptr = AV.getValue();
742 
743         // For a single memory access in AliasSetTracker, Accesses may contain
744         // both read and write, and they both need to be handled for CheckDeps.
745         for (auto AC : S) {
746           if (AC.getPointer() != Ptr)
747             continue;
748 
749           bool IsWrite = AC.getInt();
750 
751           // If we're using the deferred access set, then it contains only
752           // reads.
753           bool IsReadOnlyPtr = ReadOnlyPtr.count(Ptr) && !IsWrite;
754           if (UseDeferred && !IsReadOnlyPtr)
755             continue;
756           // Otherwise, the pointer must be in the PtrAccessSet, either as a
757           // read or a write.
758           assert(((IsReadOnlyPtr && UseDeferred) || IsWrite ||
759                   S.count(MemAccessInfo(Ptr, false))) &&
760                  "Alias-set pointer not in the access set?");
761 
762           MemAccessInfo Access(Ptr, IsWrite);
763           DepCands.insert(Access);
764 
765           // Memorize read-only pointers for later processing and skip them in
766           // the first round (they need to be checked after we have seen all
767           // write pointers). Note: we also mark pointer that are not
768           // consecutive as "read-only" pointers (so that we check
769           // "a[b[i]] +="). Hence, we need the second check for "!IsWrite".
770           if (!UseDeferred && IsReadOnlyPtr) {
771             DeferredAccesses.insert(Access);
772             continue;
773           }
774 
775           // If this is a write - check other reads and writes for conflicts. If
776           // this is a read only check other writes for conflicts (but only if
777           // there is no other write to the ptr - this is an optimization to
778           // catch "a[i] = a[i] + " without having to do a dependence check).
779           if ((IsWrite || IsReadOnlyPtr) && SetHasWrite) {
780             CheckDeps.insert(Access);
781             IsRTCheckAnalysisNeeded = true;
782           }
783 
784           if (IsWrite)
785             SetHasWrite = true;
786 
787           // Create sets of pointers connected by a shared alias set and
788           // underlying object.
789           typedef SmallVector<Value *, 16> ValueVector;
790           ValueVector TempObjects;
791 
792           GetUnderlyingObjects(Ptr, TempObjects, DL, LI);
793           DEBUG(dbgs() << "Underlying objects for pointer " << *Ptr << "\n");
794           for (Value *UnderlyingObj : TempObjects) {
795             // nullptr never alias, don't join sets for pointer that have "null"
796             // in their UnderlyingObjects list.
797             if (isa<ConstantPointerNull>(UnderlyingObj))
798               continue;
799 
800             UnderlyingObjToAccessMap::iterator Prev =
801                 ObjToLastAccess.find(UnderlyingObj);
802             if (Prev != ObjToLastAccess.end())
803               DepCands.unionSets(Access, Prev->second);
804 
805             ObjToLastAccess[UnderlyingObj] = Access;
806             DEBUG(dbgs() << "  " << *UnderlyingObj << "\n");
807           }
808         }
809       }
810     }
811   }
812 }
813 
isInBoundsGep(Value * Ptr)814 static bool isInBoundsGep(Value *Ptr) {
815   if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr))
816     return GEP->isInBounds();
817   return false;
818 }
819 
820 /// \brief Return true if an AddRec pointer \p Ptr is unsigned non-wrapping,
821 /// i.e. monotonically increasing/decreasing.
isNoWrapAddRec(Value * Ptr,const SCEVAddRecExpr * AR,PredicatedScalarEvolution & PSE,const Loop * L)822 static bool isNoWrapAddRec(Value *Ptr, const SCEVAddRecExpr *AR,
823                            PredicatedScalarEvolution &PSE, const Loop *L) {
824   // FIXME: This should probably only return true for NUW.
825   if (AR->getNoWrapFlags(SCEV::NoWrapMask))
826     return true;
827 
828   // Scalar evolution does not propagate the non-wrapping flags to values that
829   // are derived from a non-wrapping induction variable because non-wrapping
830   // could be flow-sensitive.
831   //
832   // Look through the potentially overflowing instruction to try to prove
833   // non-wrapping for the *specific* value of Ptr.
834 
835   // The arithmetic implied by an inbounds GEP can't overflow.
836   auto *GEP = dyn_cast<GetElementPtrInst>(Ptr);
837   if (!GEP || !GEP->isInBounds())
838     return false;
839 
840   // Make sure there is only one non-const index and analyze that.
841   Value *NonConstIndex = nullptr;
842   for (Value *Index : make_range(GEP->idx_begin(), GEP->idx_end()))
843     if (!isa<ConstantInt>(Index)) {
844       if (NonConstIndex)
845         return false;
846       NonConstIndex = Index;
847     }
848   if (!NonConstIndex)
849     // The recurrence is on the pointer, ignore for now.
850     return false;
851 
852   // The index in GEP is signed.  It is non-wrapping if it's derived from a NSW
853   // AddRec using a NSW operation.
854   if (auto *OBO = dyn_cast<OverflowingBinaryOperator>(NonConstIndex))
855     if (OBO->hasNoSignedWrap() &&
856         // Assume constant for other the operand so that the AddRec can be
857         // easily found.
858         isa<ConstantInt>(OBO->getOperand(1))) {
859       auto *OpScev = PSE.getSCEV(OBO->getOperand(0));
860 
861       if (auto *OpAR = dyn_cast<SCEVAddRecExpr>(OpScev))
862         return OpAR->getLoop() == L && OpAR->getNoWrapFlags(SCEV::FlagNSW);
863     }
864 
865   return false;
866 }
867 
868 /// \brief Check whether the access through \p Ptr has a constant stride.
getPtrStride(PredicatedScalarEvolution & PSE,Value * Ptr,const Loop * Lp,const ValueToValueMap & StridesMap,bool Assume)869 int64_t llvm::getPtrStride(PredicatedScalarEvolution &PSE, Value *Ptr,
870                            const Loop *Lp, const ValueToValueMap &StridesMap,
871                            bool Assume) {
872   Type *Ty = Ptr->getType();
873   assert(Ty->isPointerTy() && "Unexpected non-ptr");
874 
875   // Make sure that the pointer does not point to aggregate types.
876   auto *PtrTy = cast<PointerType>(Ty);
877   if (PtrTy->getElementType()->isAggregateType()) {
878     DEBUG(dbgs() << "LAA: Bad stride - Not a pointer to a scalar type" << *Ptr
879                  << "\n");
880     return 0;
881   }
882 
883   const SCEV *PtrScev = replaceSymbolicStrideSCEV(PSE, StridesMap, Ptr);
884 
885   const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
886   if (Assume && !AR)
887     AR = PSE.getAsAddRec(Ptr);
888 
889   if (!AR) {
890     DEBUG(dbgs() << "LAA: Bad stride - Not an AddRecExpr pointer " << *Ptr
891                  << " SCEV: " << *PtrScev << "\n");
892     return 0;
893   }
894 
895   // The accesss function must stride over the innermost loop.
896   if (Lp != AR->getLoop()) {
897     DEBUG(dbgs() << "LAA: Bad stride - Not striding over innermost loop " <<
898           *Ptr << " SCEV: " << *AR << "\n");
899     return 0;
900   }
901 
902   // The address calculation must not wrap. Otherwise, a dependence could be
903   // inverted.
904   // An inbounds getelementptr that is a AddRec with a unit stride
905   // cannot wrap per definition. The unit stride requirement is checked later.
906   // An getelementptr without an inbounds attribute and unit stride would have
907   // to access the pointer value "0" which is undefined behavior in address
908   // space 0, therefore we can also vectorize this case.
909   bool IsInBoundsGEP = isInBoundsGep(Ptr);
910   bool IsNoWrapAddRec =
911       PSE.hasNoOverflow(Ptr, SCEVWrapPredicate::IncrementNUSW) ||
912       isNoWrapAddRec(Ptr, AR, PSE, Lp);
913   bool IsInAddressSpaceZero = PtrTy->getAddressSpace() == 0;
914   if (!IsNoWrapAddRec && !IsInBoundsGEP && !IsInAddressSpaceZero) {
915     if (Assume) {
916       PSE.setNoOverflow(Ptr, SCEVWrapPredicate::IncrementNUSW);
917       IsNoWrapAddRec = true;
918       DEBUG(dbgs() << "LAA: Pointer may wrap in the address space:\n"
919                    << "LAA:   Pointer: " << *Ptr << "\n"
920                    << "LAA:   SCEV: " << *AR << "\n"
921                    << "LAA:   Added an overflow assumption\n");
922     } else {
923       DEBUG(dbgs() << "LAA: Bad stride - Pointer may wrap in the address space "
924                    << *Ptr << " SCEV: " << *AR << "\n");
925       return 0;
926     }
927   }
928 
929   // Check the step is constant.
930   const SCEV *Step = AR->getStepRecurrence(*PSE.getSE());
931 
932   // Calculate the pointer stride and check if it is constant.
933   const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
934   if (!C) {
935     DEBUG(dbgs() << "LAA: Bad stride - Not a constant strided " << *Ptr <<
936           " SCEV: " << *AR << "\n");
937     return 0;
938   }
939 
940   auto &DL = Lp->getHeader()->getModule()->getDataLayout();
941   int64_t Size = DL.getTypeAllocSize(PtrTy->getElementType());
942   const APInt &APStepVal = C->getAPInt();
943 
944   // Huge step value - give up.
945   if (APStepVal.getBitWidth() > 64)
946     return 0;
947 
948   int64_t StepVal = APStepVal.getSExtValue();
949 
950   // Strided access.
951   int64_t Stride = StepVal / Size;
952   int64_t Rem = StepVal % Size;
953   if (Rem)
954     return 0;
955 
956   // If the SCEV could wrap but we have an inbounds gep with a unit stride we
957   // know we can't "wrap around the address space". In case of address space
958   // zero we know that this won't happen without triggering undefined behavior.
959   if (!IsNoWrapAddRec && (IsInBoundsGEP || IsInAddressSpaceZero) &&
960       Stride != 1 && Stride != -1) {
961     if (Assume) {
962       // We can avoid this case by adding a run-time check.
963       DEBUG(dbgs() << "LAA: Non unit strided pointer which is not either "
964                    << "inbouds or in address space 0 may wrap:\n"
965                    << "LAA:   Pointer: " << *Ptr << "\n"
966                    << "LAA:   SCEV: " << *AR << "\n"
967                    << "LAA:   Added an overflow assumption\n");
968       PSE.setNoOverflow(Ptr, SCEVWrapPredicate::IncrementNUSW);
969     } else
970       return 0;
971   }
972 
973   return Stride;
974 }
975 
976 /// Take the pointer operand from the Load/Store instruction.
977 /// Returns NULL if this is not a valid Load/Store instruction.
getPointerOperand(Value * I)978 static Value *getPointerOperand(Value *I) {
979   if (auto *LI = dyn_cast<LoadInst>(I))
980     return LI->getPointerOperand();
981   if (auto *SI = dyn_cast<StoreInst>(I))
982     return SI->getPointerOperand();
983   return nullptr;
984 }
985 
986 /// Take the address space operand from the Load/Store instruction.
987 /// Returns -1 if this is not a valid Load/Store instruction.
getAddressSpaceOperand(Value * I)988 static unsigned getAddressSpaceOperand(Value *I) {
989   if (LoadInst *L = dyn_cast<LoadInst>(I))
990     return L->getPointerAddressSpace();
991   if (StoreInst *S = dyn_cast<StoreInst>(I))
992     return S->getPointerAddressSpace();
993   return -1;
994 }
995 
996 /// 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)997 bool llvm::isConsecutiveAccess(Value *A, Value *B, const DataLayout &DL,
998                                ScalarEvolution &SE, bool CheckType) {
999   Value *PtrA = getPointerOperand(A);
1000   Value *PtrB = getPointerOperand(B);
1001   unsigned ASA = getAddressSpaceOperand(A);
1002   unsigned ASB = getAddressSpaceOperand(B);
1003 
1004   // Check that the address spaces match and that the pointers are valid.
1005   if (!PtrA || !PtrB || (ASA != ASB))
1006     return false;
1007 
1008   // Make sure that A and B are different pointers.
1009   if (PtrA == PtrB)
1010     return false;
1011 
1012   // Make sure that A and B have the same type if required.
1013   if(CheckType && PtrA->getType() != PtrB->getType())
1014       return false;
1015 
1016   unsigned PtrBitWidth = DL.getPointerSizeInBits(ASA);
1017   Type *Ty = cast<PointerType>(PtrA->getType())->getElementType();
1018   APInt Size(PtrBitWidth, DL.getTypeStoreSize(Ty));
1019 
1020   APInt OffsetA(PtrBitWidth, 0), OffsetB(PtrBitWidth, 0);
1021   PtrA = PtrA->stripAndAccumulateInBoundsConstantOffsets(DL, OffsetA);
1022   PtrB = PtrB->stripAndAccumulateInBoundsConstantOffsets(DL, OffsetB);
1023 
1024   //  OffsetDelta = OffsetB - OffsetA;
1025   const SCEV *OffsetSCEVA = SE.getConstant(OffsetA);
1026   const SCEV *OffsetSCEVB = SE.getConstant(OffsetB);
1027   const SCEV *OffsetDeltaSCEV = SE.getMinusSCEV(OffsetSCEVB, OffsetSCEVA);
1028   const SCEVConstant *OffsetDeltaC = dyn_cast<SCEVConstant>(OffsetDeltaSCEV);
1029   const APInt &OffsetDelta = OffsetDeltaC->getAPInt();
1030   // Check if they are based on the same pointer. That makes the offsets
1031   // sufficient.
1032   if (PtrA == PtrB)
1033     return OffsetDelta == Size;
1034 
1035   // Compute the necessary base pointer delta to have the necessary final delta
1036   // equal to the size.
1037   // BaseDelta = Size - OffsetDelta;
1038   const SCEV *SizeSCEV = SE.getConstant(Size);
1039   const SCEV *BaseDelta = SE.getMinusSCEV(SizeSCEV, OffsetDeltaSCEV);
1040 
1041   // Otherwise compute the distance with SCEV between the base pointers.
1042   const SCEV *PtrSCEVA = SE.getSCEV(PtrA);
1043   const SCEV *PtrSCEVB = SE.getSCEV(PtrB);
1044   const SCEV *X = SE.getAddExpr(PtrSCEVA, BaseDelta);
1045   return X == PtrSCEVB;
1046 }
1047 
isSafeForVectorization(DepType Type)1048 bool MemoryDepChecker::Dependence::isSafeForVectorization(DepType Type) {
1049   switch (Type) {
1050   case NoDep:
1051   case Forward:
1052   case BackwardVectorizable:
1053     return true;
1054 
1055   case Unknown:
1056   case ForwardButPreventsForwarding:
1057   case Backward:
1058   case BackwardVectorizableButPreventsForwarding:
1059     return false;
1060   }
1061   llvm_unreachable("unexpected DepType!");
1062 }
1063 
isBackward() const1064 bool MemoryDepChecker::Dependence::isBackward() const {
1065   switch (Type) {
1066   case NoDep:
1067   case Forward:
1068   case ForwardButPreventsForwarding:
1069   case Unknown:
1070     return false;
1071 
1072   case BackwardVectorizable:
1073   case Backward:
1074   case BackwardVectorizableButPreventsForwarding:
1075     return true;
1076   }
1077   llvm_unreachable("unexpected DepType!");
1078 }
1079 
isPossiblyBackward() const1080 bool MemoryDepChecker::Dependence::isPossiblyBackward() const {
1081   return isBackward() || Type == Unknown;
1082 }
1083 
isForward() const1084 bool MemoryDepChecker::Dependence::isForward() const {
1085   switch (Type) {
1086   case Forward:
1087   case ForwardButPreventsForwarding:
1088     return true;
1089 
1090   case NoDep:
1091   case Unknown:
1092   case BackwardVectorizable:
1093   case Backward:
1094   case BackwardVectorizableButPreventsForwarding:
1095     return false;
1096   }
1097   llvm_unreachable("unexpected DepType!");
1098 }
1099 
couldPreventStoreLoadForward(uint64_t Distance,uint64_t TypeByteSize)1100 bool MemoryDepChecker::couldPreventStoreLoadForward(uint64_t Distance,
1101                                                     uint64_t TypeByteSize) {
1102   // If loads occur at a distance that is not a multiple of a feasible vector
1103   // factor store-load forwarding does not take place.
1104   // Positive dependences might cause troubles because vectorizing them might
1105   // prevent store-load forwarding making vectorized code run a lot slower.
1106   //   a[i] = a[i-3] ^ a[i-8];
1107   //   The stores to a[i:i+1] don't align with the stores to a[i-3:i-2] and
1108   //   hence on your typical architecture store-load forwarding does not take
1109   //   place. Vectorizing in such cases does not make sense.
1110   // Store-load forwarding distance.
1111 
1112   // After this many iterations store-to-load forwarding conflicts should not
1113   // cause any slowdowns.
1114   const uint64_t NumItersForStoreLoadThroughMemory = 8 * TypeByteSize;
1115   // Maximum vector factor.
1116   uint64_t MaxVFWithoutSLForwardIssues = std::min(
1117       VectorizerParams::MaxVectorWidth * TypeByteSize, MaxSafeDepDistBytes);
1118 
1119   // Compute the smallest VF at which the store and load would be misaligned.
1120   for (uint64_t VF = 2 * TypeByteSize; VF <= MaxVFWithoutSLForwardIssues;
1121        VF *= 2) {
1122     // If the number of vector iteration between the store and the load are
1123     // small we could incur conflicts.
1124     if (Distance % VF && Distance / VF < NumItersForStoreLoadThroughMemory) {
1125       MaxVFWithoutSLForwardIssues = (VF >>= 1);
1126       break;
1127     }
1128   }
1129 
1130   if (MaxVFWithoutSLForwardIssues < 2 * TypeByteSize) {
1131     DEBUG(dbgs() << "LAA: Distance " << Distance
1132                  << " that could cause a store-load forwarding conflict\n");
1133     return true;
1134   }
1135 
1136   if (MaxVFWithoutSLForwardIssues < MaxSafeDepDistBytes &&
1137       MaxVFWithoutSLForwardIssues !=
1138           VectorizerParams::MaxVectorWidth * TypeByteSize)
1139     MaxSafeDepDistBytes = MaxVFWithoutSLForwardIssues;
1140   return false;
1141 }
1142 
1143 /// \brief Check the dependence for two accesses with the same stride \p Stride.
1144 /// \p Distance is the positive distance and \p TypeByteSize is type size in
1145 /// bytes.
1146 ///
1147 /// \returns true if they are independent.
areStridedAccessesIndependent(uint64_t Distance,uint64_t Stride,uint64_t TypeByteSize)1148 static bool areStridedAccessesIndependent(uint64_t Distance, uint64_t Stride,
1149                                           uint64_t TypeByteSize) {
1150   assert(Stride > 1 && "The stride must be greater than 1");
1151   assert(TypeByteSize > 0 && "The type size in byte must be non-zero");
1152   assert(Distance > 0 && "The distance must be non-zero");
1153 
1154   // Skip if the distance is not multiple of type byte size.
1155   if (Distance % TypeByteSize)
1156     return false;
1157 
1158   uint64_t ScaledDist = Distance / TypeByteSize;
1159 
1160   // No dependence if the scaled distance is not multiple of the stride.
1161   // E.g.
1162   //      for (i = 0; i < 1024 ; i += 4)
1163   //        A[i+2] = A[i] + 1;
1164   //
1165   // Two accesses in memory (scaled distance is 2, stride is 4):
1166   //     | A[0] |      |      |      | A[4] |      |      |      |
1167   //     |      |      | A[2] |      |      |      | A[6] |      |
1168   //
1169   // E.g.
1170   //      for (i = 0; i < 1024 ; i += 3)
1171   //        A[i+4] = A[i] + 1;
1172   //
1173   // Two accesses in memory (scaled distance is 4, stride is 3):
1174   //     | A[0] |      |      | A[3] |      |      | A[6] |      |      |
1175   //     |      |      |      |      | A[4] |      |      | A[7] |      |
1176   return ScaledDist % Stride;
1177 }
1178 
1179 MemoryDepChecker::Dependence::DepType
isDependent(const MemAccessInfo & A,unsigned AIdx,const MemAccessInfo & B,unsigned BIdx,const ValueToValueMap & Strides)1180 MemoryDepChecker::isDependent(const MemAccessInfo &A, unsigned AIdx,
1181                               const MemAccessInfo &B, unsigned BIdx,
1182                               const ValueToValueMap &Strides) {
1183   assert (AIdx < BIdx && "Must pass arguments in program order");
1184 
1185   Value *APtr = A.getPointer();
1186   Value *BPtr = B.getPointer();
1187   bool AIsWrite = A.getInt();
1188   bool BIsWrite = B.getInt();
1189 
1190   // Two reads are independent.
1191   if (!AIsWrite && !BIsWrite)
1192     return Dependence::NoDep;
1193 
1194   // We cannot check pointers in different address spaces.
1195   if (APtr->getType()->getPointerAddressSpace() !=
1196       BPtr->getType()->getPointerAddressSpace())
1197     return Dependence::Unknown;
1198 
1199   int64_t StrideAPtr = getPtrStride(PSE, APtr, InnermostLoop, Strides, true);
1200   int64_t StrideBPtr = getPtrStride(PSE, BPtr, InnermostLoop, Strides, true);
1201 
1202   const SCEV *Src = PSE.getSCEV(APtr);
1203   const SCEV *Sink = PSE.getSCEV(BPtr);
1204 
1205   // If the induction step is negative we have to invert source and sink of the
1206   // dependence.
1207   if (StrideAPtr < 0) {
1208     std::swap(APtr, BPtr);
1209     std::swap(Src, Sink);
1210     std::swap(AIsWrite, BIsWrite);
1211     std::swap(AIdx, BIdx);
1212     std::swap(StrideAPtr, StrideBPtr);
1213   }
1214 
1215   const SCEV *Dist = PSE.getSE()->getMinusSCEV(Sink, Src);
1216 
1217   DEBUG(dbgs() << "LAA: Src Scev: " << *Src << "Sink Scev: " << *Sink
1218                << "(Induction step: " << StrideAPtr << ")\n");
1219   DEBUG(dbgs() << "LAA: Distance for " << *InstMap[AIdx] << " to "
1220                << *InstMap[BIdx] << ": " << *Dist << "\n");
1221 
1222   // Need accesses with constant stride. We don't want to vectorize
1223   // "A[B[i]] += ..." and similar code or pointer arithmetic that could wrap in
1224   // the address space.
1225   if (!StrideAPtr || !StrideBPtr || StrideAPtr != StrideBPtr){
1226     DEBUG(dbgs() << "Pointer access with non-constant stride\n");
1227     return Dependence::Unknown;
1228   }
1229 
1230   const SCEVConstant *C = dyn_cast<SCEVConstant>(Dist);
1231   if (!C) {
1232     DEBUG(dbgs() << "LAA: Dependence because of non-constant distance\n");
1233     ShouldRetryWithRuntimeCheck = true;
1234     return Dependence::Unknown;
1235   }
1236 
1237   Type *ATy = APtr->getType()->getPointerElementType();
1238   Type *BTy = BPtr->getType()->getPointerElementType();
1239   auto &DL = InnermostLoop->getHeader()->getModule()->getDataLayout();
1240   uint64_t TypeByteSize = DL.getTypeAllocSize(ATy);
1241 
1242   const APInt &Val = C->getAPInt();
1243   int64_t Distance = Val.getSExtValue();
1244   uint64_t Stride = std::abs(StrideAPtr);
1245 
1246   // Attempt to prove strided accesses independent.
1247   if (std::abs(Distance) > 0 && Stride > 1 && ATy == BTy &&
1248       areStridedAccessesIndependent(std::abs(Distance), Stride, TypeByteSize)) {
1249     DEBUG(dbgs() << "LAA: Strided accesses are independent\n");
1250     return Dependence::NoDep;
1251   }
1252 
1253   // Negative distances are not plausible dependencies.
1254   if (Val.isNegative()) {
1255     bool IsTrueDataDependence = (AIsWrite && !BIsWrite);
1256     if (IsTrueDataDependence && EnableForwardingConflictDetection &&
1257         (couldPreventStoreLoadForward(Val.abs().getZExtValue(), TypeByteSize) ||
1258          ATy != BTy)) {
1259       DEBUG(dbgs() << "LAA: Forward but may prevent st->ld forwarding\n");
1260       return Dependence::ForwardButPreventsForwarding;
1261     }
1262 
1263     DEBUG(dbgs() << "LAA: Dependence is negative\n");
1264     return Dependence::Forward;
1265   }
1266 
1267   // Write to the same location with the same size.
1268   // Could be improved to assert type sizes are the same (i32 == float, etc).
1269   if (Val == 0) {
1270     if (ATy == BTy)
1271       return Dependence::Forward;
1272     DEBUG(dbgs() << "LAA: Zero dependence difference but different types\n");
1273     return Dependence::Unknown;
1274   }
1275 
1276   assert(Val.isStrictlyPositive() && "Expect a positive value");
1277 
1278   if (ATy != BTy) {
1279     DEBUG(dbgs() <<
1280           "LAA: ReadWrite-Write positive dependency with different types\n");
1281     return Dependence::Unknown;
1282   }
1283 
1284   // Bail out early if passed-in parameters make vectorization not feasible.
1285   unsigned ForcedFactor = (VectorizerParams::VectorizationFactor ?
1286                            VectorizerParams::VectorizationFactor : 1);
1287   unsigned ForcedUnroll = (VectorizerParams::VectorizationInterleave ?
1288                            VectorizerParams::VectorizationInterleave : 1);
1289   // The minimum number of iterations for a vectorized/unrolled version.
1290   unsigned MinNumIter = std::max(ForcedFactor * ForcedUnroll, 2U);
1291 
1292   // It's not vectorizable if the distance is smaller than the minimum distance
1293   // needed for a vectroized/unrolled version. Vectorizing one iteration in
1294   // front needs TypeByteSize * Stride. Vectorizing the last iteration needs
1295   // TypeByteSize (No need to plus the last gap distance).
1296   //
1297   // E.g. Assume one char is 1 byte in memory and one int is 4 bytes.
1298   //      foo(int *A) {
1299   //        int *B = (int *)((char *)A + 14);
1300   //        for (i = 0 ; i < 1024 ; i += 2)
1301   //          B[i] = A[i] + 1;
1302   //      }
1303   //
1304   // Two accesses in memory (stride is 2):
1305   //     | A[0] |      | A[2] |      | A[4] |      | A[6] |      |
1306   //                              | B[0] |      | B[2] |      | B[4] |
1307   //
1308   // Distance needs for vectorizing iterations except the last iteration:
1309   // 4 * 2 * (MinNumIter - 1). Distance needs for the last iteration: 4.
1310   // So the minimum distance needed is: 4 * 2 * (MinNumIter - 1) + 4.
1311   //
1312   // If MinNumIter is 2, it is vectorizable as the minimum distance needed is
1313   // 12, which is less than distance.
1314   //
1315   // If MinNumIter is 4 (Say if a user forces the vectorization factor to be 4),
1316   // the minimum distance needed is 28, which is greater than distance. It is
1317   // not safe to do vectorization.
1318   uint64_t MinDistanceNeeded =
1319       TypeByteSize * Stride * (MinNumIter - 1) + TypeByteSize;
1320   if (MinDistanceNeeded > static_cast<uint64_t>(Distance)) {
1321     DEBUG(dbgs() << "LAA: Failure because of positive distance " << Distance
1322                  << '\n');
1323     return Dependence::Backward;
1324   }
1325 
1326   // Unsafe if the minimum distance needed is greater than max safe distance.
1327   if (MinDistanceNeeded > MaxSafeDepDistBytes) {
1328     DEBUG(dbgs() << "LAA: Failure because it needs at least "
1329                  << MinDistanceNeeded << " size in bytes");
1330     return Dependence::Backward;
1331   }
1332 
1333   // Positive distance bigger than max vectorization factor.
1334   // FIXME: Should use max factor instead of max distance in bytes, which could
1335   // not handle different types.
1336   // E.g. Assume one char is 1 byte in memory and one int is 4 bytes.
1337   //      void foo (int *A, char *B) {
1338   //        for (unsigned i = 0; i < 1024; i++) {
1339   //          A[i+2] = A[i] + 1;
1340   //          B[i+2] = B[i] + 1;
1341   //        }
1342   //      }
1343   //
1344   // This case is currently unsafe according to the max safe distance. If we
1345   // analyze the two accesses on array B, the max safe dependence distance
1346   // is 2. Then we analyze the accesses on array A, the minimum distance needed
1347   // is 8, which is less than 2 and forbidden vectorization, But actually
1348   // both A and B could be vectorized by 2 iterations.
1349   MaxSafeDepDistBytes =
1350       std::min(static_cast<uint64_t>(Distance), MaxSafeDepDistBytes);
1351 
1352   bool IsTrueDataDependence = (!AIsWrite && BIsWrite);
1353   if (IsTrueDataDependence && EnableForwardingConflictDetection &&
1354       couldPreventStoreLoadForward(Distance, TypeByteSize))
1355     return Dependence::BackwardVectorizableButPreventsForwarding;
1356 
1357   DEBUG(dbgs() << "LAA: Positive distance " << Val.getSExtValue()
1358                << " with max VF = "
1359                << MaxSafeDepDistBytes / (TypeByteSize * Stride) << '\n');
1360 
1361   return Dependence::BackwardVectorizable;
1362 }
1363 
areDepsSafe(DepCandidates & AccessSets,MemAccessInfoSet & CheckDeps,const ValueToValueMap & Strides)1364 bool MemoryDepChecker::areDepsSafe(DepCandidates &AccessSets,
1365                                    MemAccessInfoSet &CheckDeps,
1366                                    const ValueToValueMap &Strides) {
1367 
1368   MaxSafeDepDistBytes = -1;
1369   while (!CheckDeps.empty()) {
1370     MemAccessInfo CurAccess = *CheckDeps.begin();
1371 
1372     // Get the relevant memory access set.
1373     EquivalenceClasses<MemAccessInfo>::iterator I =
1374       AccessSets.findValue(AccessSets.getLeaderValue(CurAccess));
1375 
1376     // Check accesses within this set.
1377     EquivalenceClasses<MemAccessInfo>::member_iterator AI =
1378         AccessSets.member_begin(I);
1379     EquivalenceClasses<MemAccessInfo>::member_iterator AE =
1380         AccessSets.member_end();
1381 
1382     // Check every access pair.
1383     while (AI != AE) {
1384       CheckDeps.erase(*AI);
1385       EquivalenceClasses<MemAccessInfo>::member_iterator OI = std::next(AI);
1386       while (OI != AE) {
1387         // Check every accessing instruction pair in program order.
1388         for (std::vector<unsigned>::iterator I1 = Accesses[*AI].begin(),
1389              I1E = Accesses[*AI].end(); I1 != I1E; ++I1)
1390           for (std::vector<unsigned>::iterator I2 = Accesses[*OI].begin(),
1391                I2E = Accesses[*OI].end(); I2 != I2E; ++I2) {
1392             auto A = std::make_pair(&*AI, *I1);
1393             auto B = std::make_pair(&*OI, *I2);
1394 
1395             assert(*I1 != *I2);
1396             if (*I1 > *I2)
1397               std::swap(A, B);
1398 
1399             Dependence::DepType Type =
1400                 isDependent(*A.first, A.second, *B.first, B.second, Strides);
1401             SafeForVectorization &= Dependence::isSafeForVectorization(Type);
1402 
1403             // Gather dependences unless we accumulated MaxDependences
1404             // dependences.  In that case return as soon as we find the first
1405             // unsafe dependence.  This puts a limit on this quadratic
1406             // algorithm.
1407             if (RecordDependences) {
1408               if (Type != Dependence::NoDep)
1409                 Dependences.push_back(Dependence(A.second, B.second, Type));
1410 
1411               if (Dependences.size() >= MaxDependences) {
1412                 RecordDependences = false;
1413                 Dependences.clear();
1414                 DEBUG(dbgs() << "Too many dependences, stopped recording\n");
1415               }
1416             }
1417             if (!RecordDependences && !SafeForVectorization)
1418               return false;
1419           }
1420         ++OI;
1421       }
1422       AI++;
1423     }
1424   }
1425 
1426   DEBUG(dbgs() << "Total Dependences: " << Dependences.size() << "\n");
1427   return SafeForVectorization;
1428 }
1429 
1430 SmallVector<Instruction *, 4>
getInstructionsForAccess(Value * Ptr,bool isWrite) const1431 MemoryDepChecker::getInstructionsForAccess(Value *Ptr, bool isWrite) const {
1432   MemAccessInfo Access(Ptr, isWrite);
1433   auto &IndexVector = Accesses.find(Access)->second;
1434 
1435   SmallVector<Instruction *, 4> Insts;
1436   std::transform(IndexVector.begin(), IndexVector.end(),
1437                  std::back_inserter(Insts),
1438                  [&](unsigned Idx) { return this->InstMap[Idx]; });
1439   return Insts;
1440 }
1441 
1442 const char *MemoryDepChecker::Dependence::DepName[] = {
1443     "NoDep", "Unknown", "Forward", "ForwardButPreventsForwarding", "Backward",
1444     "BackwardVectorizable", "BackwardVectorizableButPreventsForwarding"};
1445 
print(raw_ostream & OS,unsigned Depth,const SmallVectorImpl<Instruction * > & Instrs) const1446 void MemoryDepChecker::Dependence::print(
1447     raw_ostream &OS, unsigned Depth,
1448     const SmallVectorImpl<Instruction *> &Instrs) const {
1449   OS.indent(Depth) << DepName[Type] << ":\n";
1450   OS.indent(Depth + 2) << *Instrs[Source] << " -> \n";
1451   OS.indent(Depth + 2) << *Instrs[Destination] << "\n";
1452 }
1453 
canAnalyzeLoop()1454 bool LoopAccessInfo::canAnalyzeLoop() {
1455   // We need to have a loop header.
1456   DEBUG(dbgs() << "LAA: Found a loop in "
1457                << TheLoop->getHeader()->getParent()->getName() << ": "
1458                << TheLoop->getHeader()->getName() << '\n');
1459 
1460   // We can only analyze innermost loops.
1461   if (!TheLoop->empty()) {
1462     DEBUG(dbgs() << "LAA: loop is not the innermost loop\n");
1463     emitAnalysis(LoopAccessReport() << "loop is not the innermost loop");
1464     return false;
1465   }
1466 
1467   // We must have a single backedge.
1468   if (TheLoop->getNumBackEdges() != 1) {
1469     DEBUG(dbgs() << "LAA: loop control flow is not understood by analyzer\n");
1470     emitAnalysis(
1471         LoopAccessReport() <<
1472         "loop control flow is not understood by analyzer");
1473     return false;
1474   }
1475 
1476   // We must have a single exiting block.
1477   if (!TheLoop->getExitingBlock()) {
1478     DEBUG(dbgs() << "LAA: loop control flow is not understood by analyzer\n");
1479     emitAnalysis(
1480         LoopAccessReport() <<
1481         "loop control flow is not understood by analyzer");
1482     return false;
1483   }
1484 
1485   // We only handle bottom-tested loops, i.e. loop in which the condition is
1486   // checked at the end of each iteration. With that we can assume that all
1487   // instructions in the loop are executed the same number of times.
1488   if (TheLoop->getExitingBlock() != TheLoop->getLoopLatch()) {
1489     DEBUG(dbgs() << "LAA: loop control flow is not understood by analyzer\n");
1490     emitAnalysis(
1491         LoopAccessReport() <<
1492         "loop control flow is not understood by analyzer");
1493     return false;
1494   }
1495 
1496   // ScalarEvolution needs to be able to find the exit count.
1497   const SCEV *ExitCount = PSE->getBackedgeTakenCount();
1498   if (ExitCount == PSE->getSE()->getCouldNotCompute()) {
1499     emitAnalysis(LoopAccessReport()
1500                  << "could not determine number of loop iterations");
1501     DEBUG(dbgs() << "LAA: SCEV could not compute the loop exit count.\n");
1502     return false;
1503   }
1504 
1505   return true;
1506 }
1507 
analyzeLoop(AliasAnalysis * AA,LoopInfo * LI,const TargetLibraryInfo * TLI,DominatorTree * DT)1508 void LoopAccessInfo::analyzeLoop(AliasAnalysis *AA, LoopInfo *LI,
1509                                  const TargetLibraryInfo *TLI,
1510                                  DominatorTree *DT) {
1511   typedef SmallPtrSet<Value*, 16> ValueSet;
1512 
1513   // Holds the Load and Store instructions.
1514   SmallVector<LoadInst *, 16> Loads;
1515   SmallVector<StoreInst *, 16> Stores;
1516 
1517   // Holds all the different accesses in the loop.
1518   unsigned NumReads = 0;
1519   unsigned NumReadWrites = 0;
1520 
1521   PtrRtChecking->Pointers.clear();
1522   PtrRtChecking->Need = false;
1523 
1524   const bool IsAnnotatedParallel = TheLoop->isAnnotatedParallel();
1525 
1526   // For each block.
1527   for (BasicBlock *BB : TheLoop->blocks()) {
1528     // Scan the BB and collect legal loads and stores.
1529     for (Instruction &I : *BB) {
1530       // If this is a load, save it. If this instruction can read from memory
1531       // but is not a load, then we quit. Notice that we don't handle function
1532       // calls that read or write.
1533       if (I.mayReadFromMemory()) {
1534         // Many math library functions read the rounding mode. We will only
1535         // vectorize a loop if it contains known function calls that don't set
1536         // the flag. Therefore, it is safe to ignore this read from memory.
1537         auto *Call = dyn_cast<CallInst>(&I);
1538         if (Call && getVectorIntrinsicIDForCall(Call, TLI))
1539           continue;
1540 
1541         // If the function has an explicit vectorized counterpart, we can safely
1542         // assume that it can be vectorized.
1543         if (Call && !Call->isNoBuiltin() && Call->getCalledFunction() &&
1544             TLI->isFunctionVectorizable(Call->getCalledFunction()->getName()))
1545           continue;
1546 
1547         auto *Ld = dyn_cast<LoadInst>(&I);
1548         if (!Ld || (!Ld->isSimple() && !IsAnnotatedParallel)) {
1549           emitAnalysis(LoopAccessReport(Ld)
1550                        << "read with atomic ordering or volatile read");
1551           DEBUG(dbgs() << "LAA: Found a non-simple load.\n");
1552           CanVecMem = false;
1553           return;
1554         }
1555         NumLoads++;
1556         Loads.push_back(Ld);
1557         DepChecker->addAccess(Ld);
1558         if (EnableMemAccessVersioning)
1559           collectStridedAccess(Ld);
1560         continue;
1561       }
1562 
1563       // Save 'store' instructions. Abort if other instructions write to memory.
1564       if (I.mayWriteToMemory()) {
1565         auto *St = dyn_cast<StoreInst>(&I);
1566         if (!St) {
1567           emitAnalysis(LoopAccessReport(St)
1568                        << "instruction cannot be vectorized");
1569           CanVecMem = false;
1570           return;
1571         }
1572         if (!St->isSimple() && !IsAnnotatedParallel) {
1573           emitAnalysis(LoopAccessReport(St)
1574                        << "write with atomic ordering or volatile write");
1575           DEBUG(dbgs() << "LAA: Found a non-simple store.\n");
1576           CanVecMem = false;
1577           return;
1578         }
1579         NumStores++;
1580         Stores.push_back(St);
1581         DepChecker->addAccess(St);
1582         if (EnableMemAccessVersioning)
1583           collectStridedAccess(St);
1584       }
1585     } // Next instr.
1586   } // Next block.
1587 
1588   // Now we have two lists that hold the loads and the stores.
1589   // Next, we find the pointers that they use.
1590 
1591   // Check if we see any stores. If there are no stores, then we don't
1592   // care if the pointers are *restrict*.
1593   if (!Stores.size()) {
1594     DEBUG(dbgs() << "LAA: Found a read-only loop!\n");
1595     CanVecMem = true;
1596     return;
1597   }
1598 
1599   MemoryDepChecker::DepCandidates DependentAccesses;
1600   AccessAnalysis Accesses(TheLoop->getHeader()->getModule()->getDataLayout(),
1601                           AA, LI, DependentAccesses, *PSE);
1602 
1603   // Holds the analyzed pointers. We don't want to call GetUnderlyingObjects
1604   // multiple times on the same object. If the ptr is accessed twice, once
1605   // for read and once for write, it will only appear once (on the write
1606   // list). This is okay, since we are going to check for conflicts between
1607   // writes and between reads and writes, but not between reads and reads.
1608   ValueSet Seen;
1609 
1610   for (StoreInst *ST : Stores) {
1611     Value *Ptr = ST->getPointerOperand();
1612     // Check for store to loop invariant address.
1613     StoreToLoopInvariantAddress |= isUniform(Ptr);
1614     // If we did *not* see this pointer before, insert it to  the read-write
1615     // list. At this phase it is only a 'write' list.
1616     if (Seen.insert(Ptr).second) {
1617       ++NumReadWrites;
1618 
1619       MemoryLocation Loc = MemoryLocation::get(ST);
1620       // The TBAA metadata could have a control dependency on the predication
1621       // condition, so we cannot rely on it when determining whether or not we
1622       // need runtime pointer checks.
1623       if (blockNeedsPredication(ST->getParent(), TheLoop, DT))
1624         Loc.AATags.TBAA = nullptr;
1625 
1626       Accesses.addStore(Loc);
1627     }
1628   }
1629 
1630   if (IsAnnotatedParallel) {
1631     DEBUG(dbgs()
1632           << "LAA: A loop annotated parallel, ignore memory dependency "
1633           << "checks.\n");
1634     CanVecMem = true;
1635     return;
1636   }
1637 
1638   for (LoadInst *LD : Loads) {
1639     Value *Ptr = LD->getPointerOperand();
1640     // If we did *not* see this pointer before, insert it to the
1641     // read list. If we *did* see it before, then it is already in
1642     // the read-write list. This allows us to vectorize expressions
1643     // such as A[i] += x;  Because the address of A[i] is a read-write
1644     // pointer. This only works if the index of A[i] is consecutive.
1645     // If the address of i is unknown (for example A[B[i]]) then we may
1646     // read a few words, modify, and write a few words, and some of the
1647     // words may be written to the same address.
1648     bool IsReadOnlyPtr = false;
1649     if (Seen.insert(Ptr).second ||
1650         !getPtrStride(*PSE, Ptr, TheLoop, SymbolicStrides)) {
1651       ++NumReads;
1652       IsReadOnlyPtr = true;
1653     }
1654 
1655     MemoryLocation Loc = MemoryLocation::get(LD);
1656     // The TBAA metadata could have a control dependency on the predication
1657     // condition, so we cannot rely on it when determining whether or not we
1658     // need runtime pointer checks.
1659     if (blockNeedsPredication(LD->getParent(), TheLoop, DT))
1660       Loc.AATags.TBAA = nullptr;
1661 
1662     Accesses.addLoad(Loc, IsReadOnlyPtr);
1663   }
1664 
1665   // If we write (or read-write) to a single destination and there are no
1666   // other reads in this loop then is it safe to vectorize.
1667   if (NumReadWrites == 1 && NumReads == 0) {
1668     DEBUG(dbgs() << "LAA: Found a write-only loop!\n");
1669     CanVecMem = true;
1670     return;
1671   }
1672 
1673   // Build dependence sets and check whether we need a runtime pointer bounds
1674   // check.
1675   Accesses.buildDependenceSets();
1676 
1677   // Find pointers with computable bounds. We are going to use this information
1678   // to place a runtime bound check.
1679   bool CanDoRTIfNeeded = Accesses.canCheckPtrAtRT(*PtrRtChecking, PSE->getSE(),
1680                                                   TheLoop, SymbolicStrides);
1681   if (!CanDoRTIfNeeded) {
1682     emitAnalysis(LoopAccessReport() << "cannot identify array bounds");
1683     DEBUG(dbgs() << "LAA: We can't vectorize because we can't find "
1684                  << "the array bounds.\n");
1685     CanVecMem = false;
1686     return;
1687   }
1688 
1689   DEBUG(dbgs() << "LAA: We can perform a memory runtime check if needed.\n");
1690 
1691   CanVecMem = true;
1692   if (Accesses.isDependencyCheckNeeded()) {
1693     DEBUG(dbgs() << "LAA: Checking memory dependencies\n");
1694     CanVecMem = DepChecker->areDepsSafe(
1695         DependentAccesses, Accesses.getDependenciesToCheck(), SymbolicStrides);
1696     MaxSafeDepDistBytes = DepChecker->getMaxSafeDepDistBytes();
1697 
1698     if (!CanVecMem && DepChecker->shouldRetryWithRuntimeCheck()) {
1699       DEBUG(dbgs() << "LAA: Retrying with memory checks\n");
1700 
1701       // Clear the dependency checks. We assume they are not needed.
1702       Accesses.resetDepChecks(*DepChecker);
1703 
1704       PtrRtChecking->reset();
1705       PtrRtChecking->Need = true;
1706 
1707       auto *SE = PSE->getSE();
1708       CanDoRTIfNeeded = Accesses.canCheckPtrAtRT(*PtrRtChecking, SE, TheLoop,
1709                                                  SymbolicStrides, true);
1710 
1711       // Check that we found the bounds for the pointer.
1712       if (!CanDoRTIfNeeded) {
1713         emitAnalysis(LoopAccessReport()
1714                      << "cannot check memory dependencies at runtime");
1715         DEBUG(dbgs() << "LAA: Can't vectorize with memory checks\n");
1716         CanVecMem = false;
1717         return;
1718       }
1719 
1720       CanVecMem = true;
1721     }
1722   }
1723 
1724   if (CanVecMem)
1725     DEBUG(dbgs() << "LAA: No unsafe dependent memory operations in loop.  We"
1726                  << (PtrRtChecking->Need ? "" : " don't")
1727                  << " need runtime memory checks.\n");
1728   else {
1729     emitAnalysis(
1730         LoopAccessReport()
1731         << "unsafe dependent memory operations in loop. Use "
1732            "#pragma loop distribute(enable) to allow loop distribution "
1733            "to attempt to isolate the offending operations into a separate "
1734            "loop");
1735     DEBUG(dbgs() << "LAA: unsafe dependent memory operations in loop\n");
1736   }
1737 }
1738 
blockNeedsPredication(BasicBlock * BB,Loop * TheLoop,DominatorTree * DT)1739 bool LoopAccessInfo::blockNeedsPredication(BasicBlock *BB, Loop *TheLoop,
1740                                            DominatorTree *DT)  {
1741   assert(TheLoop->contains(BB) && "Unknown block used");
1742 
1743   // Blocks that do not dominate the latch need predication.
1744   BasicBlock* Latch = TheLoop->getLoopLatch();
1745   return !DT->dominates(BB, Latch);
1746 }
1747 
emitAnalysis(LoopAccessReport & Message)1748 void LoopAccessInfo::emitAnalysis(LoopAccessReport &Message) {
1749   assert(!Report && "Multiple reports generated");
1750   Report = Message;
1751 }
1752 
isUniform(Value * V) const1753 bool LoopAccessInfo::isUniform(Value *V) const {
1754   return (PSE->getSE()->isLoopInvariant(PSE->getSE()->getSCEV(V), TheLoop));
1755 }
1756 
1757 // FIXME: this function is currently a duplicate of the one in
1758 // LoopVectorize.cpp.
getFirstInst(Instruction * FirstInst,Value * V,Instruction * Loc)1759 static Instruction *getFirstInst(Instruction *FirstInst, Value *V,
1760                                  Instruction *Loc) {
1761   if (FirstInst)
1762     return FirstInst;
1763   if (Instruction *I = dyn_cast<Instruction>(V))
1764     return I->getParent() == Loc->getParent() ? I : nullptr;
1765   return nullptr;
1766 }
1767 
1768 namespace {
1769 /// \brief IR Values for the lower and upper bounds of a pointer evolution.  We
1770 /// need to use value-handles because SCEV expansion can invalidate previously
1771 /// expanded values.  Thus expansion of a pointer can invalidate the bounds for
1772 /// a previous one.
1773 struct PointerBounds {
1774   TrackingVH<Value> Start;
1775   TrackingVH<Value> End;
1776 };
1777 } // end anonymous namespace
1778 
1779 /// \brief Expand code for the lower and upper bound of the pointer group \p CG
1780 /// in \p TheLoop.  \return the values for the bounds.
1781 static PointerBounds
expandBounds(const RuntimePointerChecking::CheckingPtrGroup * CG,Loop * TheLoop,Instruction * Loc,SCEVExpander & Exp,ScalarEvolution * SE,const RuntimePointerChecking & PtrRtChecking)1782 expandBounds(const RuntimePointerChecking::CheckingPtrGroup *CG, Loop *TheLoop,
1783              Instruction *Loc, SCEVExpander &Exp, ScalarEvolution *SE,
1784              const RuntimePointerChecking &PtrRtChecking) {
1785   Value *Ptr = PtrRtChecking.Pointers[CG->Members[0]].PointerValue;
1786   const SCEV *Sc = SE->getSCEV(Ptr);
1787 
1788   if (SE->isLoopInvariant(Sc, TheLoop)) {
1789     DEBUG(dbgs() << "LAA: Adding RT check for a loop invariant ptr:" << *Ptr
1790                  << "\n");
1791     return {Ptr, Ptr};
1792   } else {
1793     unsigned AS = Ptr->getType()->getPointerAddressSpace();
1794     LLVMContext &Ctx = Loc->getContext();
1795 
1796     // Use this type for pointer arithmetic.
1797     Type *PtrArithTy = Type::getInt8PtrTy(Ctx, AS);
1798     Value *Start = nullptr, *End = nullptr;
1799 
1800     DEBUG(dbgs() << "LAA: Adding RT check for range:\n");
1801     Start = Exp.expandCodeFor(CG->Low, PtrArithTy, Loc);
1802     End = Exp.expandCodeFor(CG->High, PtrArithTy, Loc);
1803     DEBUG(dbgs() << "Start: " << *CG->Low << " End: " << *CG->High << "\n");
1804     return {Start, End};
1805   }
1806 }
1807 
1808 /// \brief Turns a collection of checks into a collection of expanded upper and
1809 /// lower bounds for both pointers in the check.
expandBounds(const SmallVectorImpl<RuntimePointerChecking::PointerCheck> & PointerChecks,Loop * L,Instruction * Loc,ScalarEvolution * SE,SCEVExpander & Exp,const RuntimePointerChecking & PtrRtChecking)1810 static SmallVector<std::pair<PointerBounds, PointerBounds>, 4> expandBounds(
1811     const SmallVectorImpl<RuntimePointerChecking::PointerCheck> &PointerChecks,
1812     Loop *L, Instruction *Loc, ScalarEvolution *SE, SCEVExpander &Exp,
1813     const RuntimePointerChecking &PtrRtChecking) {
1814   SmallVector<std::pair<PointerBounds, PointerBounds>, 4> ChecksWithBounds;
1815 
1816   // Here we're relying on the SCEV Expander's cache to only emit code for the
1817   // same bounds once.
1818   std::transform(
1819       PointerChecks.begin(), PointerChecks.end(),
1820       std::back_inserter(ChecksWithBounds),
1821       [&](const RuntimePointerChecking::PointerCheck &Check) {
1822         PointerBounds
1823           First = expandBounds(Check.first, L, Loc, Exp, SE, PtrRtChecking),
1824           Second = expandBounds(Check.second, L, Loc, Exp, SE, PtrRtChecking);
1825         return std::make_pair(First, Second);
1826       });
1827 
1828   return ChecksWithBounds;
1829 }
1830 
addRuntimeChecks(Instruction * Loc,const SmallVectorImpl<RuntimePointerChecking::PointerCheck> & PointerChecks) const1831 std::pair<Instruction *, Instruction *> LoopAccessInfo::addRuntimeChecks(
1832     Instruction *Loc,
1833     const SmallVectorImpl<RuntimePointerChecking::PointerCheck> &PointerChecks)
1834     const {
1835   const DataLayout &DL = TheLoop->getHeader()->getModule()->getDataLayout();
1836   auto *SE = PSE->getSE();
1837   SCEVExpander Exp(*SE, DL, "induction");
1838   auto ExpandedChecks =
1839       expandBounds(PointerChecks, TheLoop, Loc, SE, Exp, *PtrRtChecking);
1840 
1841   LLVMContext &Ctx = Loc->getContext();
1842   Instruction *FirstInst = nullptr;
1843   IRBuilder<> ChkBuilder(Loc);
1844   // Our instructions might fold to a constant.
1845   Value *MemoryRuntimeCheck = nullptr;
1846 
1847   for (const auto &Check : ExpandedChecks) {
1848     const PointerBounds &A = Check.first, &B = Check.second;
1849     // Check if two pointers (A and B) conflict where conflict is computed as:
1850     // start(A) <= end(B) && start(B) <= end(A)
1851     unsigned AS0 = A.Start->getType()->getPointerAddressSpace();
1852     unsigned AS1 = B.Start->getType()->getPointerAddressSpace();
1853 
1854     assert((AS0 == B.End->getType()->getPointerAddressSpace()) &&
1855            (AS1 == A.End->getType()->getPointerAddressSpace()) &&
1856            "Trying to bounds check pointers with different address spaces");
1857 
1858     Type *PtrArithTy0 = Type::getInt8PtrTy(Ctx, AS0);
1859     Type *PtrArithTy1 = Type::getInt8PtrTy(Ctx, AS1);
1860 
1861     Value *Start0 = ChkBuilder.CreateBitCast(A.Start, PtrArithTy0, "bc");
1862     Value *Start1 = ChkBuilder.CreateBitCast(B.Start, PtrArithTy1, "bc");
1863     Value *End0 =   ChkBuilder.CreateBitCast(A.End,   PtrArithTy1, "bc");
1864     Value *End1 =   ChkBuilder.CreateBitCast(B.End,   PtrArithTy0, "bc");
1865 
1866     Value *Cmp0 = ChkBuilder.CreateICmpULE(Start0, End1, "bound0");
1867     FirstInst = getFirstInst(FirstInst, Cmp0, Loc);
1868     Value *Cmp1 = ChkBuilder.CreateICmpULE(Start1, End0, "bound1");
1869     FirstInst = getFirstInst(FirstInst, Cmp1, Loc);
1870     Value *IsConflict = ChkBuilder.CreateAnd(Cmp0, Cmp1, "found.conflict");
1871     FirstInst = getFirstInst(FirstInst, IsConflict, Loc);
1872     if (MemoryRuntimeCheck) {
1873       IsConflict =
1874           ChkBuilder.CreateOr(MemoryRuntimeCheck, IsConflict, "conflict.rdx");
1875       FirstInst = getFirstInst(FirstInst, IsConflict, Loc);
1876     }
1877     MemoryRuntimeCheck = IsConflict;
1878   }
1879 
1880   if (!MemoryRuntimeCheck)
1881     return std::make_pair(nullptr, nullptr);
1882 
1883   // We have to do this trickery because the IRBuilder might fold the check to a
1884   // constant expression in which case there is no Instruction anchored in a
1885   // the block.
1886   Instruction *Check = BinaryOperator::CreateAnd(MemoryRuntimeCheck,
1887                                                  ConstantInt::getTrue(Ctx));
1888   ChkBuilder.Insert(Check, "memcheck.conflict");
1889   FirstInst = getFirstInst(FirstInst, Check, Loc);
1890   return std::make_pair(FirstInst, Check);
1891 }
1892 
1893 std::pair<Instruction *, Instruction *>
addRuntimeChecks(Instruction * Loc) const1894 LoopAccessInfo::addRuntimeChecks(Instruction *Loc) const {
1895   if (!PtrRtChecking->Need)
1896     return std::make_pair(nullptr, nullptr);
1897 
1898   return addRuntimeChecks(Loc, PtrRtChecking->getChecks());
1899 }
1900 
collectStridedAccess(Value * MemAccess)1901 void LoopAccessInfo::collectStridedAccess(Value *MemAccess) {
1902   Value *Ptr = nullptr;
1903   if (LoadInst *LI = dyn_cast<LoadInst>(MemAccess))
1904     Ptr = LI->getPointerOperand();
1905   else if (StoreInst *SI = dyn_cast<StoreInst>(MemAccess))
1906     Ptr = SI->getPointerOperand();
1907   else
1908     return;
1909 
1910   Value *Stride = getStrideFromPointer(Ptr, PSE->getSE(), TheLoop);
1911   if (!Stride)
1912     return;
1913 
1914   DEBUG(dbgs() << "LAA: Found a strided access that we can version");
1915   DEBUG(dbgs() << "  Ptr: " << *Ptr << " Stride: " << *Stride << "\n");
1916   SymbolicStrides[Ptr] = Stride;
1917   StrideSet.insert(Stride);
1918 }
1919 
LoopAccessInfo(Loop * L,ScalarEvolution * SE,const TargetLibraryInfo * TLI,AliasAnalysis * AA,DominatorTree * DT,LoopInfo * LI)1920 LoopAccessInfo::LoopAccessInfo(Loop *L, ScalarEvolution *SE,
1921                                const TargetLibraryInfo *TLI, AliasAnalysis *AA,
1922                                DominatorTree *DT, LoopInfo *LI)
1923     : PSE(llvm::make_unique<PredicatedScalarEvolution>(*SE, *L)),
1924       PtrRtChecking(llvm::make_unique<RuntimePointerChecking>(SE)),
1925       DepChecker(llvm::make_unique<MemoryDepChecker>(*PSE, L)), TheLoop(L),
1926       NumLoads(0), NumStores(0), MaxSafeDepDistBytes(-1), CanVecMem(false),
1927       StoreToLoopInvariantAddress(false) {
1928   if (canAnalyzeLoop())
1929     analyzeLoop(AA, LI, TLI, DT);
1930 }
1931 
print(raw_ostream & OS,unsigned Depth) const1932 void LoopAccessInfo::print(raw_ostream &OS, unsigned Depth) const {
1933   if (CanVecMem) {
1934     OS.indent(Depth) << "Memory dependences are safe";
1935     if (MaxSafeDepDistBytes != -1ULL)
1936       OS << " with a maximum dependence distance of " << MaxSafeDepDistBytes
1937          << " bytes";
1938     if (PtrRtChecking->Need)
1939       OS << " with run-time checks";
1940     OS << "\n";
1941   }
1942 
1943   if (Report)
1944     OS.indent(Depth) << "Report: " << Report->str() << "\n";
1945 
1946   if (auto *Dependences = DepChecker->getDependences()) {
1947     OS.indent(Depth) << "Dependences:\n";
1948     for (auto &Dep : *Dependences) {
1949       Dep.print(OS, Depth + 2, DepChecker->getMemoryInstructions());
1950       OS << "\n";
1951     }
1952   } else
1953     OS.indent(Depth) << "Too many dependences, not recorded\n";
1954 
1955   // List the pair of accesses need run-time checks to prove independence.
1956   PtrRtChecking->print(OS, Depth);
1957   OS << "\n";
1958 
1959   OS.indent(Depth) << "Store to invariant address was "
1960                    << (StoreToLoopInvariantAddress ? "" : "not ")
1961                    << "found in loop.\n";
1962 
1963   OS.indent(Depth) << "SCEV assumptions:\n";
1964   PSE->getUnionPredicate().print(OS, Depth);
1965 
1966   OS << "\n";
1967 
1968   OS.indent(Depth) << "Expressions re-written:\n";
1969   PSE->print(OS, Depth);
1970 }
1971 
getInfo(Loop * L)1972 const LoopAccessInfo &LoopAccessLegacyAnalysis::getInfo(Loop *L) {
1973   auto &LAI = LoopAccessInfoMap[L];
1974 
1975   if (!LAI)
1976     LAI = llvm::make_unique<LoopAccessInfo>(L, SE, TLI, AA, DT, LI);
1977 
1978   return *LAI.get();
1979 }
1980 
print(raw_ostream & OS,const Module * M) const1981 void LoopAccessLegacyAnalysis::print(raw_ostream &OS, const Module *M) const {
1982   LoopAccessLegacyAnalysis &LAA = *const_cast<LoopAccessLegacyAnalysis *>(this);
1983 
1984   for (Loop *TopLevelLoop : *LI)
1985     for (Loop *L : depth_first(TopLevelLoop)) {
1986       OS.indent(2) << L->getHeader()->getName() << ":\n";
1987       auto &LAI = LAA.getInfo(L);
1988       LAI.print(OS, 4);
1989     }
1990 }
1991 
runOnFunction(Function & F)1992 bool LoopAccessLegacyAnalysis::runOnFunction(Function &F) {
1993   SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE();
1994   auto *TLIP = getAnalysisIfAvailable<TargetLibraryInfoWrapperPass>();
1995   TLI = TLIP ? &TLIP->getTLI() : nullptr;
1996   AA = &getAnalysis<AAResultsWrapperPass>().getAAResults();
1997   DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1998   LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
1999 
2000   return false;
2001 }
2002 
getAnalysisUsage(AnalysisUsage & AU) const2003 void LoopAccessLegacyAnalysis::getAnalysisUsage(AnalysisUsage &AU) const {
2004     AU.addRequired<ScalarEvolutionWrapperPass>();
2005     AU.addRequired<AAResultsWrapperPass>();
2006     AU.addRequired<DominatorTreeWrapperPass>();
2007     AU.addRequired<LoopInfoWrapperPass>();
2008 
2009     AU.setPreservesAll();
2010 }
2011 
2012 char LoopAccessLegacyAnalysis::ID = 0;
2013 static const char laa_name[] = "Loop Access Analysis";
2014 #define LAA_NAME "loop-accesses"
2015 
2016 INITIALIZE_PASS_BEGIN(LoopAccessLegacyAnalysis, LAA_NAME, laa_name, false, true)
2017 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)
2018 INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass)
2019 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
2020 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
2021 INITIALIZE_PASS_END(LoopAccessLegacyAnalysis, LAA_NAME, laa_name, false, true)
2022 
2023 char LoopAccessAnalysis::PassID;
2024 
run(Loop & L,AnalysisManager<Loop> & AM)2025 LoopAccessInfo LoopAccessAnalysis::run(Loop &L, AnalysisManager<Loop> &AM) {
2026   const AnalysisManager<Function> &FAM =
2027       AM.getResult<FunctionAnalysisManagerLoopProxy>(L).getManager();
2028   Function &F = *L.getHeader()->getParent();
2029   auto *SE = FAM.getCachedResult<ScalarEvolutionAnalysis>(F);
2030   auto *TLI = FAM.getCachedResult<TargetLibraryAnalysis>(F);
2031   auto *AA = FAM.getCachedResult<AAManager>(F);
2032   auto *DT = FAM.getCachedResult<DominatorTreeAnalysis>(F);
2033   auto *LI = FAM.getCachedResult<LoopAnalysis>(F);
2034   if (!SE)
2035     report_fatal_error(
2036         "ScalarEvolution must have been cached at a higher level");
2037   if (!AA)
2038     report_fatal_error("AliasAnalysis must have been cached at a higher level");
2039   if (!DT)
2040     report_fatal_error("DominatorTree must have been cached at a higher level");
2041   if (!LI)
2042     report_fatal_error("LoopInfo must have been cached at a higher level");
2043   return LoopAccessInfo(&L, SE, TLI, AA, DT, LI);
2044 }
2045 
run(Loop & L,AnalysisManager<Loop> & AM)2046 PreservedAnalyses LoopAccessInfoPrinterPass::run(Loop &L,
2047                                                  AnalysisManager<Loop> &AM) {
2048   Function &F = *L.getHeader()->getParent();
2049   auto &LAI = AM.getResult<LoopAccessAnalysis>(L);
2050   OS << "Loop access info in function '" << F.getName() << "':\n";
2051   OS.indent(2) << L.getHeader()->getName() << ":\n";
2052   LAI.print(OS, 4);
2053   return PreservedAnalyses::all();
2054 }
2055 
2056 namespace llvm {
createLAAPass()2057   Pass *createLAAPass() {
2058     return new LoopAccessLegacyAnalysis();
2059   }
2060 }
2061