• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 //===- BasicAliasAnalysis.cpp - Stateless Alias Analysis Impl -------------===//
2 //
3 //                     The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This file defines the primary stateless implementation of the
11 // Alias Analysis interface that implements identities (two different
12 // globals cannot alias, etc), but does no stateful analysis.
13 //
14 //===----------------------------------------------------------------------===//
15 
16 #include "llvm/Analysis/BasicAliasAnalysis.h"
17 #include "llvm/ADT/SmallVector.h"
18 #include "llvm/ADT/Statistic.h"
19 #include "llvm/Analysis/AliasAnalysis.h"
20 #include "llvm/Analysis/CFG.h"
21 #include "llvm/Analysis/CaptureTracking.h"
22 #include "llvm/Analysis/InstructionSimplify.h"
23 #include "llvm/Analysis/LoopInfo.h"
24 #include "llvm/Analysis/MemoryBuiltins.h"
25 #include "llvm/Analysis/ValueTracking.h"
26 #include "llvm/Analysis/AssumptionCache.h"
27 #include "llvm/IR/Constants.h"
28 #include "llvm/IR/DataLayout.h"
29 #include "llvm/IR/DerivedTypes.h"
30 #include "llvm/IR/Dominators.h"
31 #include "llvm/IR/GlobalAlias.h"
32 #include "llvm/IR/GlobalVariable.h"
33 #include "llvm/IR/Instructions.h"
34 #include "llvm/IR/IntrinsicInst.h"
35 #include "llvm/IR/LLVMContext.h"
36 #include "llvm/IR/Operator.h"
37 #include "llvm/Pass.h"
38 #include "llvm/Support/ErrorHandling.h"
39 #include <algorithm>
40 using namespace llvm;
41 
42 /// Enable analysis of recursive PHI nodes.
43 static cl::opt<bool> EnableRecPhiAnalysis("basicaa-recphi", cl::Hidden,
44                                           cl::init(false));
45 
46 /// SearchLimitReached / SearchTimes shows how often the limit of
47 /// to decompose GEPs is reached. It will affect the precision
48 /// of basic alias analysis.
49 #define DEBUG_TYPE "basicaa"
50 STATISTIC(SearchLimitReached, "Number of times the limit to "
51                               "decompose GEPs is reached");
52 STATISTIC(SearchTimes, "Number of times a GEP is decomposed");
53 
54 /// Cutoff after which to stop analysing a set of phi nodes potentially involved
55 /// in a cycle. Because we are analysing 'through' phi nodes we need to be
56 /// careful with value equivalence. We use reachability to make sure a value
57 /// cannot be involved in a cycle.
58 const unsigned MaxNumPhiBBsValueReachabilityCheck = 20;
59 
60 // The max limit of the search depth in DecomposeGEPExpression() and
61 // GetUnderlyingObject(), both functions need to use the same search
62 // depth otherwise the algorithm in aliasGEP will assert.
63 static const unsigned MaxLookupSearchDepth = 6;
64 
65 //===----------------------------------------------------------------------===//
66 // Useful predicates
67 //===----------------------------------------------------------------------===//
68 
69 /// Returns true if the pointer is to a function-local object that never
70 /// escapes from the function.
isNonEscapingLocalObject(const Value * V)71 static bool isNonEscapingLocalObject(const Value *V) {
72   // If this is a local allocation, check to see if it escapes.
73   if (isa<AllocaInst>(V) || isNoAliasCall(V))
74     // Set StoreCaptures to True so that we can assume in our callers that the
75     // pointer is not the result of a load instruction. Currently
76     // PointerMayBeCaptured doesn't have any special analysis for the
77     // StoreCaptures=false case; if it did, our callers could be refined to be
78     // more precise.
79     return !PointerMayBeCaptured(V, false, /*StoreCaptures=*/true);
80 
81   // If this is an argument that corresponds to a byval or noalias argument,
82   // then it has not escaped before entering the function.  Check if it escapes
83   // inside the function.
84   if (const Argument *A = dyn_cast<Argument>(V))
85     if (A->hasByValAttr() || A->hasNoAliasAttr())
86       // Note even if the argument is marked nocapture we still need to check
87       // for copies made inside the function. The nocapture attribute only
88       // specifies that there are no copies made that outlive the function.
89       return !PointerMayBeCaptured(V, false, /*StoreCaptures=*/true);
90 
91   return false;
92 }
93 
94 /// Returns true if the pointer is one which would have been considered an
95 /// escape by isNonEscapingLocalObject.
isEscapeSource(const Value * V)96 static bool isEscapeSource(const Value *V) {
97   if (isa<CallInst>(V) || isa<InvokeInst>(V) || isa<Argument>(V))
98     return true;
99 
100   // The load case works because isNonEscapingLocalObject considers all
101   // stores to be escapes (it passes true for the StoreCaptures argument
102   // to PointerMayBeCaptured).
103   if (isa<LoadInst>(V))
104     return true;
105 
106   return false;
107 }
108 
109 /// Returns the size of the object specified by V, or UnknownSize if unknown.
getObjectSize(const Value * V,const DataLayout & DL,const TargetLibraryInfo & TLI,bool RoundToAlign=false)110 static uint64_t getObjectSize(const Value *V, const DataLayout &DL,
111                               const TargetLibraryInfo &TLI,
112                               bool RoundToAlign = false) {
113   uint64_t Size;
114   if (getObjectSize(V, Size, DL, &TLI, RoundToAlign))
115     return Size;
116   return MemoryLocation::UnknownSize;
117 }
118 
119 /// Returns true if we can prove that the object specified by V is smaller than
120 /// Size.
isObjectSmallerThan(const Value * V,uint64_t Size,const DataLayout & DL,const TargetLibraryInfo & TLI)121 static bool isObjectSmallerThan(const Value *V, uint64_t Size,
122                                 const DataLayout &DL,
123                                 const TargetLibraryInfo &TLI) {
124   // Note that the meanings of the "object" are slightly different in the
125   // following contexts:
126   //    c1: llvm::getObjectSize()
127   //    c2: llvm.objectsize() intrinsic
128   //    c3: isObjectSmallerThan()
129   // c1 and c2 share the same meaning; however, the meaning of "object" in c3
130   // refers to the "entire object".
131   //
132   //  Consider this example:
133   //     char *p = (char*)malloc(100)
134   //     char *q = p+80;
135   //
136   //  In the context of c1 and c2, the "object" pointed by q refers to the
137   // stretch of memory of q[0:19]. So, getObjectSize(q) should return 20.
138   //
139   //  However, in the context of c3, the "object" refers to the chunk of memory
140   // being allocated. So, the "object" has 100 bytes, and q points to the middle
141   // the "object". In case q is passed to isObjectSmallerThan() as the 1st
142   // parameter, before the llvm::getObjectSize() is called to get the size of
143   // entire object, we should:
144   //    - either rewind the pointer q to the base-address of the object in
145   //      question (in this case rewind to p), or
146   //    - just give up. It is up to caller to make sure the pointer is pointing
147   //      to the base address the object.
148   //
149   // We go for 2nd option for simplicity.
150   if (!isIdentifiedObject(V))
151     return false;
152 
153   // This function needs to use the aligned object size because we allow
154   // reads a bit past the end given sufficient alignment.
155   uint64_t ObjectSize = getObjectSize(V, DL, TLI, /*RoundToAlign*/ true);
156 
157   return ObjectSize != MemoryLocation::UnknownSize && ObjectSize < Size;
158 }
159 
160 /// Returns true if we can prove that the object specified by V has size Size.
isObjectSize(const Value * V,uint64_t Size,const DataLayout & DL,const TargetLibraryInfo & TLI)161 static bool isObjectSize(const Value *V, uint64_t Size, const DataLayout &DL,
162                          const TargetLibraryInfo &TLI) {
163   uint64_t ObjectSize = getObjectSize(V, DL, TLI);
164   return ObjectSize != MemoryLocation::UnknownSize && ObjectSize == Size;
165 }
166 
167 //===----------------------------------------------------------------------===//
168 // GetElementPtr Instruction Decomposition and Analysis
169 //===----------------------------------------------------------------------===//
170 
171 /// Analyzes the specified value as a linear expression: "A*V + B", where A and
172 /// B are constant integers.
173 ///
174 /// Returns the scale and offset values as APInts and return V as a Value*, and
175 /// return whether we looked through any sign or zero extends.  The incoming
176 /// Value is known to have IntegerType and it may already be sign or zero
177 /// extended.
178 ///
179 /// Note that this looks through extends, so the high bits may not be
180 /// represented in the result.
GetLinearExpression(const Value * V,APInt & Scale,APInt & Offset,unsigned & ZExtBits,unsigned & SExtBits,const DataLayout & DL,unsigned Depth,AssumptionCache * AC,DominatorTree * DT,bool & NSW,bool & NUW)181 /*static*/ const Value *BasicAAResult::GetLinearExpression(
182     const Value *V, APInt &Scale, APInt &Offset, unsigned &ZExtBits,
183     unsigned &SExtBits, const DataLayout &DL, unsigned Depth,
184     AssumptionCache *AC, DominatorTree *DT, bool &NSW, bool &NUW) {
185   assert(V->getType()->isIntegerTy() && "Not an integer value");
186 
187   // Limit our recursion depth.
188   if (Depth == 6) {
189     Scale = 1;
190     Offset = 0;
191     return V;
192   }
193 
194   if (const ConstantInt *Const = dyn_cast<ConstantInt>(V)) {
195     // if it's a constant, just convert it to an offset and remove the variable.
196     // If we've been called recursively the Offset bit width will be greater
197     // than the constant's (the Offset's always as wide as the outermost call),
198     // so we'll zext here and process any extension in the isa<SExtInst> &
199     // isa<ZExtInst> cases below.
200     Offset += Const->getValue().zextOrSelf(Offset.getBitWidth());
201     assert(Scale == 0 && "Constant values don't have a scale");
202     return V;
203   }
204 
205   if (const BinaryOperator *BOp = dyn_cast<BinaryOperator>(V)) {
206     if (ConstantInt *RHSC = dyn_cast<ConstantInt>(BOp->getOperand(1))) {
207 
208       // If we've been called recursively then Offset and Scale will be wider
209       // that the BOp operands. We'll always zext it here as we'll process sign
210       // extensions below (see the isa<SExtInst> / isa<ZExtInst> cases).
211       APInt RHS = RHSC->getValue().zextOrSelf(Offset.getBitWidth());
212 
213       switch (BOp->getOpcode()) {
214       default:
215         // We don't understand this instruction, so we can't decompose it any
216         // further.
217         Scale = 1;
218         Offset = 0;
219         return V;
220       case Instruction::Or:
221         // X|C == X+C if all the bits in C are unset in X.  Otherwise we can't
222         // analyze it.
223         if (!MaskedValueIsZero(BOp->getOperand(0), RHSC->getValue(), DL, 0, AC,
224                                BOp, DT)) {
225           Scale = 1;
226           Offset = 0;
227           return V;
228         }
229       // FALL THROUGH.
230       case Instruction::Add:
231         V = GetLinearExpression(BOp->getOperand(0), Scale, Offset, ZExtBits,
232                                 SExtBits, DL, Depth + 1, AC, DT, NSW, NUW);
233         Offset += RHS;
234         break;
235       case Instruction::Sub:
236         V = GetLinearExpression(BOp->getOperand(0), Scale, Offset, ZExtBits,
237                                 SExtBits, DL, Depth + 1, AC, DT, NSW, NUW);
238         Offset -= RHS;
239         break;
240       case Instruction::Mul:
241         V = GetLinearExpression(BOp->getOperand(0), Scale, Offset, ZExtBits,
242                                 SExtBits, DL, Depth + 1, AC, DT, NSW, NUW);
243         Offset *= RHS;
244         Scale *= RHS;
245         break;
246       case Instruction::Shl:
247         V = GetLinearExpression(BOp->getOperand(0), Scale, Offset, ZExtBits,
248                                 SExtBits, DL, Depth + 1, AC, DT, NSW, NUW);
249         Offset <<= RHS.getLimitedValue();
250         Scale <<= RHS.getLimitedValue();
251         // the semantics of nsw and nuw for left shifts don't match those of
252         // multiplications, so we won't propagate them.
253         NSW = NUW = false;
254         return V;
255       }
256 
257       if (isa<OverflowingBinaryOperator>(BOp)) {
258         NUW &= BOp->hasNoUnsignedWrap();
259         NSW &= BOp->hasNoSignedWrap();
260       }
261       return V;
262     }
263   }
264 
265   // Since GEP indices are sign extended anyway, we don't care about the high
266   // bits of a sign or zero extended value - just scales and offsets.  The
267   // extensions have to be consistent though.
268   if (isa<SExtInst>(V) || isa<ZExtInst>(V)) {
269     Value *CastOp = cast<CastInst>(V)->getOperand(0);
270     unsigned NewWidth = V->getType()->getPrimitiveSizeInBits();
271     unsigned SmallWidth = CastOp->getType()->getPrimitiveSizeInBits();
272     unsigned OldZExtBits = ZExtBits, OldSExtBits = SExtBits;
273     const Value *Result =
274         GetLinearExpression(CastOp, Scale, Offset, ZExtBits, SExtBits, DL,
275                             Depth + 1, AC, DT, NSW, NUW);
276 
277     // zext(zext(%x)) == zext(%x), and similiarly for sext; we'll handle this
278     // by just incrementing the number of bits we've extended by.
279     unsigned ExtendedBy = NewWidth - SmallWidth;
280 
281     if (isa<SExtInst>(V) && ZExtBits == 0) {
282       // sext(sext(%x, a), b) == sext(%x, a + b)
283 
284       if (NSW) {
285         // We haven't sign-wrapped, so it's valid to decompose sext(%x + c)
286         // into sext(%x) + sext(c). We'll sext the Offset ourselves:
287         unsigned OldWidth = Offset.getBitWidth();
288         Offset = Offset.trunc(SmallWidth).sext(NewWidth).zextOrSelf(OldWidth);
289       } else {
290         // We may have signed-wrapped, so don't decompose sext(%x + c) into
291         // sext(%x) + sext(c)
292         Scale = 1;
293         Offset = 0;
294         Result = CastOp;
295         ZExtBits = OldZExtBits;
296         SExtBits = OldSExtBits;
297       }
298       SExtBits += ExtendedBy;
299     } else {
300       // sext(zext(%x, a), b) = zext(zext(%x, a), b) = zext(%x, a + b)
301 
302       if (!NUW) {
303         // We may have unsigned-wrapped, so don't decompose zext(%x + c) into
304         // zext(%x) + zext(c)
305         Scale = 1;
306         Offset = 0;
307         Result = CastOp;
308         ZExtBits = OldZExtBits;
309         SExtBits = OldSExtBits;
310       }
311       ZExtBits += ExtendedBy;
312     }
313 
314     return Result;
315   }
316 
317   Scale = 1;
318   Offset = 0;
319   return V;
320 }
321 
322 /// If V is a symbolic pointer expression, decompose it into a base pointer
323 /// with a constant offset and a number of scaled symbolic offsets.
324 ///
325 /// The scaled symbolic offsets (represented by pairs of a Value* and a scale
326 /// in the VarIndices vector) are Value*'s that are known to be scaled by the
327 /// specified amount, but which may have other unrepresented high bits. As
328 /// such, the gep cannot necessarily be reconstructed from its decomposed form.
329 ///
330 /// When DataLayout is around, this function is capable of analyzing everything
331 /// that GetUnderlyingObject can look through. To be able to do that
332 /// GetUnderlyingObject and DecomposeGEPExpression must use the same search
333 /// depth (MaxLookupSearchDepth). When DataLayout not is around, it just looks
334 /// through pointer casts.
DecomposeGEPExpression(const Value * V,int64_t & BaseOffs,SmallVectorImpl<VariableGEPIndex> & VarIndices,bool & MaxLookupReached,const DataLayout & DL,AssumptionCache * AC,DominatorTree * DT)335 /*static*/ const Value *BasicAAResult::DecomposeGEPExpression(
336     const Value *V, int64_t &BaseOffs,
337     SmallVectorImpl<VariableGEPIndex> &VarIndices, bool &MaxLookupReached,
338     const DataLayout &DL, AssumptionCache *AC, DominatorTree *DT) {
339   // Limit recursion depth to limit compile time in crazy cases.
340   unsigned MaxLookup = MaxLookupSearchDepth;
341   MaxLookupReached = false;
342   SearchTimes++;
343 
344   BaseOffs = 0;
345   do {
346     // See if this is a bitcast or GEP.
347     const Operator *Op = dyn_cast<Operator>(V);
348     if (!Op) {
349       // The only non-operator case we can handle are GlobalAliases.
350       if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
351         if (!GA->mayBeOverridden()) {
352           V = GA->getAliasee();
353           continue;
354         }
355       }
356       return V;
357     }
358 
359     if (Op->getOpcode() == Instruction::BitCast ||
360         Op->getOpcode() == Instruction::AddrSpaceCast) {
361       V = Op->getOperand(0);
362       continue;
363     }
364 
365     const GEPOperator *GEPOp = dyn_cast<GEPOperator>(Op);
366     if (!GEPOp) {
367       // If it's not a GEP, hand it off to SimplifyInstruction to see if it
368       // can come up with something. This matches what GetUnderlyingObject does.
369       if (const Instruction *I = dyn_cast<Instruction>(V))
370         // TODO: Get a DominatorTree and AssumptionCache and use them here
371         // (these are both now available in this function, but this should be
372         // updated when GetUnderlyingObject is updated). TLI should be
373         // provided also.
374         if (const Value *Simplified =
375                 SimplifyInstruction(const_cast<Instruction *>(I), DL)) {
376           V = Simplified;
377           continue;
378         }
379 
380       return V;
381     }
382 
383     // Don't attempt to analyze GEPs over unsized objects.
384     if (!GEPOp->getOperand(0)->getType()->getPointerElementType()->isSized())
385       return V;
386 
387     unsigned AS = GEPOp->getPointerAddressSpace();
388     // Walk the indices of the GEP, accumulating them into BaseOff/VarIndices.
389     gep_type_iterator GTI = gep_type_begin(GEPOp);
390     for (User::const_op_iterator I = GEPOp->op_begin() + 1, E = GEPOp->op_end();
391          I != E; ++I) {
392       const Value *Index = *I;
393       // Compute the (potentially symbolic) offset in bytes for this index.
394       if (StructType *STy = dyn_cast<StructType>(*GTI++)) {
395         // For a struct, add the member offset.
396         unsigned FieldNo = cast<ConstantInt>(Index)->getZExtValue();
397         if (FieldNo == 0)
398           continue;
399 
400         BaseOffs += DL.getStructLayout(STy)->getElementOffset(FieldNo);
401         continue;
402       }
403 
404       // For an array/pointer, add the element offset, explicitly scaled.
405       if (const ConstantInt *CIdx = dyn_cast<ConstantInt>(Index)) {
406         if (CIdx->isZero())
407           continue;
408         BaseOffs += DL.getTypeAllocSize(*GTI) * CIdx->getSExtValue();
409         continue;
410       }
411 
412       uint64_t Scale = DL.getTypeAllocSize(*GTI);
413       unsigned ZExtBits = 0, SExtBits = 0;
414 
415       // If the integer type is smaller than the pointer size, it is implicitly
416       // sign extended to pointer size.
417       unsigned Width = Index->getType()->getIntegerBitWidth();
418       unsigned PointerSize = DL.getPointerSizeInBits(AS);
419       if (PointerSize > Width)
420         SExtBits += PointerSize - Width;
421 
422       // Use GetLinearExpression to decompose the index into a C1*V+C2 form.
423       APInt IndexScale(Width, 0), IndexOffset(Width, 0);
424       bool NSW = true, NUW = true;
425       Index = GetLinearExpression(Index, IndexScale, IndexOffset, ZExtBits,
426                                   SExtBits, DL, 0, AC, DT, NSW, NUW);
427 
428       // The GEP index scale ("Scale") scales C1*V+C2, yielding (C1*V+C2)*Scale.
429       // This gives us an aggregate computation of (C1*Scale)*V + C2*Scale.
430       BaseOffs += IndexOffset.getSExtValue() * Scale;
431       Scale *= IndexScale.getSExtValue();
432 
433       // If we already had an occurrence of this index variable, merge this
434       // scale into it.  For example, we want to handle:
435       //   A[x][x] -> x*16 + x*4 -> x*20
436       // This also ensures that 'x' only appears in the index list once.
437       for (unsigned i = 0, e = VarIndices.size(); i != e; ++i) {
438         if (VarIndices[i].V == Index && VarIndices[i].ZExtBits == ZExtBits &&
439             VarIndices[i].SExtBits == SExtBits) {
440           Scale += VarIndices[i].Scale;
441           VarIndices.erase(VarIndices.begin() + i);
442           break;
443         }
444       }
445 
446       // Make sure that we have a scale that makes sense for this target's
447       // pointer size.
448       if (unsigned ShiftBits = 64 - PointerSize) {
449         Scale <<= ShiftBits;
450         Scale = (int64_t)Scale >> ShiftBits;
451       }
452 
453       if (Scale) {
454         VariableGEPIndex Entry = {Index, ZExtBits, SExtBits,
455                                   static_cast<int64_t>(Scale)};
456         VarIndices.push_back(Entry);
457       }
458     }
459 
460     // Analyze the base pointer next.
461     V = GEPOp->getOperand(0);
462   } while (--MaxLookup);
463 
464   // If the chain of expressions is too deep, just return early.
465   MaxLookupReached = true;
466   SearchLimitReached++;
467   return V;
468 }
469 
470 /// Returns whether the given pointer value points to memory that is local to
471 /// the function, with global constants being considered local to all
472 /// functions.
pointsToConstantMemory(const MemoryLocation & Loc,bool OrLocal)473 bool BasicAAResult::pointsToConstantMemory(const MemoryLocation &Loc,
474                                            bool OrLocal) {
475   assert(Visited.empty() && "Visited must be cleared after use!");
476 
477   unsigned MaxLookup = 8;
478   SmallVector<const Value *, 16> Worklist;
479   Worklist.push_back(Loc.Ptr);
480   do {
481     const Value *V = GetUnderlyingObject(Worklist.pop_back_val(), DL);
482     if (!Visited.insert(V).second) {
483       Visited.clear();
484       return AAResultBase::pointsToConstantMemory(Loc, OrLocal);
485     }
486 
487     // An alloca instruction defines local memory.
488     if (OrLocal && isa<AllocaInst>(V))
489       continue;
490 
491     // A global constant counts as local memory for our purposes.
492     if (const GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) {
493       // Note: this doesn't require GV to be "ODR" because it isn't legal for a
494       // global to be marked constant in some modules and non-constant in
495       // others.  GV may even be a declaration, not a definition.
496       if (!GV->isConstant()) {
497         Visited.clear();
498         return AAResultBase::pointsToConstantMemory(Loc, OrLocal);
499       }
500       continue;
501     }
502 
503     // If both select values point to local memory, then so does the select.
504     if (const SelectInst *SI = dyn_cast<SelectInst>(V)) {
505       Worklist.push_back(SI->getTrueValue());
506       Worklist.push_back(SI->getFalseValue());
507       continue;
508     }
509 
510     // If all values incoming to a phi node point to local memory, then so does
511     // the phi.
512     if (const PHINode *PN = dyn_cast<PHINode>(V)) {
513       // Don't bother inspecting phi nodes with many operands.
514       if (PN->getNumIncomingValues() > MaxLookup) {
515         Visited.clear();
516         return AAResultBase::pointsToConstantMemory(Loc, OrLocal);
517       }
518       for (Value *IncValue : PN->incoming_values())
519         Worklist.push_back(IncValue);
520       continue;
521     }
522 
523     // Otherwise be conservative.
524     Visited.clear();
525     return AAResultBase::pointsToConstantMemory(Loc, OrLocal);
526 
527   } while (!Worklist.empty() && --MaxLookup);
528 
529   Visited.clear();
530   return Worklist.empty();
531 }
532 
533 // FIXME: This code is duplicated with MemoryLocation and should be hoisted to
534 // some common utility location.
isMemsetPattern16(const Function * MS,const TargetLibraryInfo & TLI)535 static bool isMemsetPattern16(const Function *MS,
536                               const TargetLibraryInfo &TLI) {
537   if (TLI.has(LibFunc::memset_pattern16) &&
538       MS->getName() == "memset_pattern16") {
539     FunctionType *MemsetType = MS->getFunctionType();
540     if (!MemsetType->isVarArg() && MemsetType->getNumParams() == 3 &&
541         isa<PointerType>(MemsetType->getParamType(0)) &&
542         isa<PointerType>(MemsetType->getParamType(1)) &&
543         isa<IntegerType>(MemsetType->getParamType(2)))
544       return true;
545   }
546 
547   return false;
548 }
549 
550 /// Returns the behavior when calling the given call site.
getModRefBehavior(ImmutableCallSite CS)551 FunctionModRefBehavior BasicAAResult::getModRefBehavior(ImmutableCallSite CS) {
552   if (CS.doesNotAccessMemory())
553     // Can't do better than this.
554     return FMRB_DoesNotAccessMemory;
555 
556   FunctionModRefBehavior Min = FMRB_UnknownModRefBehavior;
557 
558   // If the callsite knows it only reads memory, don't return worse
559   // than that.
560   if (CS.onlyReadsMemory())
561     Min = FMRB_OnlyReadsMemory;
562 
563   if (CS.onlyAccessesArgMemory())
564     Min = FunctionModRefBehavior(Min & FMRB_OnlyAccessesArgumentPointees);
565 
566   // The AAResultBase base class has some smarts, lets use them.
567   return FunctionModRefBehavior(AAResultBase::getModRefBehavior(CS) & Min);
568 }
569 
570 /// Returns the behavior when calling the given function. For use when the call
571 /// site is not known.
getModRefBehavior(const Function * F)572 FunctionModRefBehavior BasicAAResult::getModRefBehavior(const Function *F) {
573   // If the function declares it doesn't access memory, we can't do better.
574   if (F->doesNotAccessMemory())
575     return FMRB_DoesNotAccessMemory;
576 
577   FunctionModRefBehavior Min = FMRB_UnknownModRefBehavior;
578 
579   // If the function declares it only reads memory, go with that.
580   if (F->onlyReadsMemory())
581     Min = FMRB_OnlyReadsMemory;
582 
583   if (F->onlyAccessesArgMemory())
584     Min = FunctionModRefBehavior(Min & FMRB_OnlyAccessesArgumentPointees);
585 
586   if (isMemsetPattern16(F, TLI))
587     Min = FMRB_OnlyAccessesArgumentPointees;
588 
589   // Otherwise be conservative.
590   return FunctionModRefBehavior(AAResultBase::getModRefBehavior(F) & Min);
591 }
592 
getArgModRefInfo(ImmutableCallSite CS,unsigned ArgIdx)593 ModRefInfo BasicAAResult::getArgModRefInfo(ImmutableCallSite CS,
594                                            unsigned ArgIdx) {
595   if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(CS.getInstruction()))
596     switch (II->getIntrinsicID()) {
597     default:
598       break;
599     case Intrinsic::memset:
600     case Intrinsic::memcpy:
601     case Intrinsic::memmove:
602       assert((ArgIdx == 0 || ArgIdx == 1) &&
603              "Invalid argument index for memory intrinsic");
604       return ArgIdx ? MRI_Ref : MRI_Mod;
605     }
606 
607   // We can bound the aliasing properties of memset_pattern16 just as we can
608   // for memcpy/memset.  This is particularly important because the
609   // LoopIdiomRecognizer likes to turn loops into calls to memset_pattern16
610   // whenever possible.
611   if (CS.getCalledFunction() &&
612       isMemsetPattern16(CS.getCalledFunction(), TLI)) {
613     assert((ArgIdx == 0 || ArgIdx == 1) &&
614            "Invalid argument index for memset_pattern16");
615     return ArgIdx ? MRI_Ref : MRI_Mod;
616   }
617   // FIXME: Handle memset_pattern4 and memset_pattern8 also.
618 
619   if (CS.paramHasAttr(ArgIdx + 1, Attribute::ReadOnly))
620     return MRI_Ref;
621 
622   if (CS.paramHasAttr(ArgIdx + 1, Attribute::ReadNone))
623     return MRI_NoModRef;
624 
625   return AAResultBase::getArgModRefInfo(CS, ArgIdx);
626 }
627 
isAssumeIntrinsic(ImmutableCallSite CS)628 static bool isAssumeIntrinsic(ImmutableCallSite CS) {
629   const IntrinsicInst *II = dyn_cast<IntrinsicInst>(CS.getInstruction());
630   return II && II->getIntrinsicID() == Intrinsic::assume;
631 }
632 
633 #ifndef NDEBUG
getParent(const Value * V)634 static const Function *getParent(const Value *V) {
635   if (const Instruction *inst = dyn_cast<Instruction>(V))
636     return inst->getParent()->getParent();
637 
638   if (const Argument *arg = dyn_cast<Argument>(V))
639     return arg->getParent();
640 
641   return nullptr;
642 }
643 
notDifferentParent(const Value * O1,const Value * O2)644 static bool notDifferentParent(const Value *O1, const Value *O2) {
645 
646   const Function *F1 = getParent(O1);
647   const Function *F2 = getParent(O2);
648 
649   return !F1 || !F2 || F1 == F2;
650 }
651 #endif
652 
alias(const MemoryLocation & LocA,const MemoryLocation & LocB)653 AliasResult BasicAAResult::alias(const MemoryLocation &LocA,
654                                  const MemoryLocation &LocB) {
655   assert(notDifferentParent(LocA.Ptr, LocB.Ptr) &&
656          "BasicAliasAnalysis doesn't support interprocedural queries.");
657 
658   // If we have a directly cached entry for these locations, we have recursed
659   // through this once, so just return the cached results. Notably, when this
660   // happens, we don't clear the cache.
661   auto CacheIt = AliasCache.find(LocPair(LocA, LocB));
662   if (CacheIt != AliasCache.end())
663     return CacheIt->second;
664 
665   AliasResult Alias = aliasCheck(LocA.Ptr, LocA.Size, LocA.AATags, LocB.Ptr,
666                                  LocB.Size, LocB.AATags);
667   // AliasCache rarely has more than 1 or 2 elements, always use
668   // shrink_and_clear so it quickly returns to the inline capacity of the
669   // SmallDenseMap if it ever grows larger.
670   // FIXME: This should really be shrink_to_inline_capacity_and_clear().
671   AliasCache.shrink_and_clear();
672   VisitedPhiBBs.clear();
673   return Alias;
674 }
675 
676 /// Checks to see if the specified callsite can clobber the specified memory
677 /// object.
678 ///
679 /// Since we only look at local properties of this function, we really can't
680 /// say much about this query.  We do, however, use simple "address taken"
681 /// analysis on local objects.
getModRefInfo(ImmutableCallSite CS,const MemoryLocation & Loc)682 ModRefInfo BasicAAResult::getModRefInfo(ImmutableCallSite CS,
683                                         const MemoryLocation &Loc) {
684   assert(notDifferentParent(CS.getInstruction(), Loc.Ptr) &&
685          "AliasAnalysis query involving multiple functions!");
686 
687   const Value *Object = GetUnderlyingObject(Loc.Ptr, DL);
688 
689   // If this is a tail call and Loc.Ptr points to a stack location, we know that
690   // the tail call cannot access or modify the local stack.
691   // We cannot exclude byval arguments here; these belong to the caller of
692   // the current function not to the current function, and a tail callee
693   // may reference them.
694   if (isa<AllocaInst>(Object))
695     if (const CallInst *CI = dyn_cast<CallInst>(CS.getInstruction()))
696       if (CI->isTailCall())
697         return MRI_NoModRef;
698 
699   // If the pointer is to a locally allocated object that does not escape,
700   // then the call can not mod/ref the pointer unless the call takes the pointer
701   // as an argument, and itself doesn't capture it.
702   if (!isa<Constant>(Object) && CS.getInstruction() != Object &&
703       isNonEscapingLocalObject(Object)) {
704     bool PassedAsArg = false;
705     unsigned ArgNo = 0;
706     for (ImmutableCallSite::arg_iterator CI = CS.arg_begin(), CE = CS.arg_end();
707          CI != CE; ++CI, ++ArgNo) {
708       // Only look at the no-capture or byval pointer arguments.  If this
709       // pointer were passed to arguments that were neither of these, then it
710       // couldn't be no-capture.
711       if (!(*CI)->getType()->isPointerTy() ||
712           (!CS.doesNotCapture(ArgNo) && !CS.isByValArgument(ArgNo)))
713         continue;
714 
715       // If this is a no-capture pointer argument, see if we can tell that it
716       // is impossible to alias the pointer we're checking.  If not, we have to
717       // assume that the call could touch the pointer, even though it doesn't
718       // escape.
719       AliasResult AR =
720           getBestAAResults().alias(MemoryLocation(*CI), MemoryLocation(Object));
721       if (AR) {
722         PassedAsArg = true;
723         break;
724       }
725     }
726 
727     if (!PassedAsArg)
728       return MRI_NoModRef;
729   }
730 
731   // While the assume intrinsic is marked as arbitrarily writing so that
732   // proper control dependencies will be maintained, it never aliases any
733   // particular memory location.
734   if (isAssumeIntrinsic(CS))
735     return MRI_NoModRef;
736 
737   // The AAResultBase base class has some smarts, lets use them.
738   return AAResultBase::getModRefInfo(CS, Loc);
739 }
740 
getModRefInfo(ImmutableCallSite CS1,ImmutableCallSite CS2)741 ModRefInfo BasicAAResult::getModRefInfo(ImmutableCallSite CS1,
742                                         ImmutableCallSite CS2) {
743   // While the assume intrinsic is marked as arbitrarily writing so that
744   // proper control dependencies will be maintained, it never aliases any
745   // particular memory location.
746   if (isAssumeIntrinsic(CS1) || isAssumeIntrinsic(CS2))
747     return MRI_NoModRef;
748 
749   // The AAResultBase base class has some smarts, lets use them.
750   return AAResultBase::getModRefInfo(CS1, CS2);
751 }
752 
753 /// Provide ad-hoc rules to disambiguate accesses through two GEP operators,
754 /// both having the exact same pointer operand.
aliasSameBasePointerGEPs(const GEPOperator * GEP1,uint64_t V1Size,const GEPOperator * GEP2,uint64_t V2Size,const DataLayout & DL)755 static AliasResult aliasSameBasePointerGEPs(const GEPOperator *GEP1,
756                                             uint64_t V1Size,
757                                             const GEPOperator *GEP2,
758                                             uint64_t V2Size,
759                                             const DataLayout &DL) {
760 
761   assert(GEP1->getPointerOperand() == GEP2->getPointerOperand() &&
762          "Expected GEPs with the same pointer operand");
763 
764   // Try to determine whether GEP1 and GEP2 index through arrays, into structs,
765   // such that the struct field accesses provably cannot alias.
766   // We also need at least two indices (the pointer, and the struct field).
767   if (GEP1->getNumIndices() != GEP2->getNumIndices() ||
768       GEP1->getNumIndices() < 2)
769     return MayAlias;
770 
771   // If we don't know the size of the accesses through both GEPs, we can't
772   // determine whether the struct fields accessed can't alias.
773   if (V1Size == MemoryLocation::UnknownSize ||
774       V2Size == MemoryLocation::UnknownSize)
775     return MayAlias;
776 
777   ConstantInt *C1 =
778       dyn_cast<ConstantInt>(GEP1->getOperand(GEP1->getNumOperands() - 1));
779   ConstantInt *C2 =
780       dyn_cast<ConstantInt>(GEP2->getOperand(GEP2->getNumOperands() - 1));
781 
782   // If the last (struct) indices are constants and are equal, the other indices
783   // might be also be dynamically equal, so the GEPs can alias.
784   if (C1 && C2 && C1 == C2)
785     return MayAlias;
786 
787   // Find the last-indexed type of the GEP, i.e., the type you'd get if
788   // you stripped the last index.
789   // On the way, look at each indexed type.  If there's something other
790   // than an array, different indices can lead to different final types.
791   SmallVector<Value *, 8> IntermediateIndices;
792 
793   // Insert the first index; we don't need to check the type indexed
794   // through it as it only drops the pointer indirection.
795   assert(GEP1->getNumIndices() > 1 && "Not enough GEP indices to examine");
796   IntermediateIndices.push_back(GEP1->getOperand(1));
797 
798   // Insert all the remaining indices but the last one.
799   // Also, check that they all index through arrays.
800   for (unsigned i = 1, e = GEP1->getNumIndices() - 1; i != e; ++i) {
801     if (!isa<ArrayType>(GetElementPtrInst::getIndexedType(
802             GEP1->getSourceElementType(), IntermediateIndices)))
803       return MayAlias;
804     IntermediateIndices.push_back(GEP1->getOperand(i + 1));
805   }
806 
807   auto *Ty = GetElementPtrInst::getIndexedType(
808     GEP1->getSourceElementType(), IntermediateIndices);
809   StructType *LastIndexedStruct = dyn_cast<StructType>(Ty);
810 
811   if (isa<SequentialType>(Ty)) {
812     // We know that:
813     // - both GEPs begin indexing from the exact same pointer;
814     // - the last indices in both GEPs are constants, indexing into a sequential
815     //   type (array or pointer);
816     // - both GEPs only index through arrays prior to that.
817     //
818     // Because array indices greater than the number of elements are valid in
819     // GEPs, unless we know the intermediate indices are identical between
820     // GEP1 and GEP2 we cannot guarantee that the last indexed arrays don't
821     // partially overlap. We also need to check that the loaded size matches
822     // the element size, otherwise we could still have overlap.
823     const uint64_t ElementSize =
824         DL.getTypeStoreSize(cast<SequentialType>(Ty)->getElementType());
825     if (V1Size != ElementSize || V2Size != ElementSize)
826       return MayAlias;
827 
828     for (unsigned i = 0, e = GEP1->getNumIndices() - 1; i != e; ++i)
829       if (GEP1->getOperand(i + 1) != GEP2->getOperand(i + 1))
830         return MayAlias;
831 
832     // Now we know that the array/pointer that GEP1 indexes into and that
833     // that GEP2 indexes into must either precisely overlap or be disjoint.
834     // Because they cannot partially overlap and because fields in an array
835     // cannot overlap, if we can prove the final indices are different between
836     // GEP1 and GEP2, we can conclude GEP1 and GEP2 don't alias.
837 
838     // If the last indices are constants, we've already checked they don't
839     // equal each other so we can exit early.
840     if (C1 && C2)
841       return NoAlias;
842     if (isKnownNonEqual(GEP1->getOperand(GEP1->getNumOperands() - 1),
843                         GEP2->getOperand(GEP2->getNumOperands() - 1),
844                         DL))
845       return NoAlias;
846     return MayAlias;
847   } else if (!LastIndexedStruct || !C1 || !C2) {
848     return MayAlias;
849   }
850 
851   // We know that:
852   // - both GEPs begin indexing from the exact same pointer;
853   // - the last indices in both GEPs are constants, indexing into a struct;
854   // - said indices are different, hence, the pointed-to fields are different;
855   // - both GEPs only index through arrays prior to that.
856   //
857   // This lets us determine that the struct that GEP1 indexes into and the
858   // struct that GEP2 indexes into must either precisely overlap or be
859   // completely disjoint.  Because they cannot partially overlap, indexing into
860   // different non-overlapping fields of the struct will never alias.
861 
862   // Therefore, the only remaining thing needed to show that both GEPs can't
863   // alias is that the fields are not overlapping.
864   const StructLayout *SL = DL.getStructLayout(LastIndexedStruct);
865   const uint64_t StructSize = SL->getSizeInBytes();
866   const uint64_t V1Off = SL->getElementOffset(C1->getZExtValue());
867   const uint64_t V2Off = SL->getElementOffset(C2->getZExtValue());
868 
869   auto EltsDontOverlap = [StructSize](uint64_t V1Off, uint64_t V1Size,
870                                       uint64_t V2Off, uint64_t V2Size) {
871     return V1Off < V2Off && V1Off + V1Size <= V2Off &&
872            ((V2Off + V2Size <= StructSize) ||
873             (V2Off + V2Size - StructSize <= V1Off));
874   };
875 
876   if (EltsDontOverlap(V1Off, V1Size, V2Off, V2Size) ||
877       EltsDontOverlap(V2Off, V2Size, V1Off, V1Size))
878     return NoAlias;
879 
880   return MayAlias;
881 }
882 
883 /// Provides a bunch of ad-hoc rules to disambiguate a GEP instruction against
884 /// another pointer.
885 ///
886 /// We know that V1 is a GEP, but we don't know anything about V2.
887 /// UnderlyingV1 is GetUnderlyingObject(GEP1, DL), UnderlyingV2 is the same for
888 /// V2.
aliasGEP(const GEPOperator * GEP1,uint64_t V1Size,const AAMDNodes & V1AAInfo,const Value * V2,uint64_t V2Size,const AAMDNodes & V2AAInfo,const Value * UnderlyingV1,const Value * UnderlyingV2)889 AliasResult BasicAAResult::aliasGEP(const GEPOperator *GEP1, uint64_t V1Size,
890                                     const AAMDNodes &V1AAInfo, const Value *V2,
891                                     uint64_t V2Size, const AAMDNodes &V2AAInfo,
892                                     const Value *UnderlyingV1,
893                                     const Value *UnderlyingV2) {
894   int64_t GEP1BaseOffset;
895   bool GEP1MaxLookupReached;
896   SmallVector<VariableGEPIndex, 4> GEP1VariableIndices;
897 
898   // If we have two gep instructions with must-alias or not-alias'ing base
899   // pointers, figure out if the indexes to the GEP tell us anything about the
900   // derived pointer.
901   if (const GEPOperator *GEP2 = dyn_cast<GEPOperator>(V2)) {
902     // Do the base pointers alias?
903     AliasResult BaseAlias =
904         aliasCheck(UnderlyingV1, MemoryLocation::UnknownSize, AAMDNodes(),
905                    UnderlyingV2, MemoryLocation::UnknownSize, AAMDNodes());
906 
907     // Check for geps of non-aliasing underlying pointers where the offsets are
908     // identical.
909     if ((BaseAlias == MayAlias) && V1Size == V2Size) {
910       // Do the base pointers alias assuming type and size.
911       AliasResult PreciseBaseAlias = aliasCheck(UnderlyingV1, V1Size, V1AAInfo,
912                                                 UnderlyingV2, V2Size, V2AAInfo);
913       if (PreciseBaseAlias == NoAlias) {
914         // See if the computed offset from the common pointer tells us about the
915         // relation of the resulting pointer.
916         int64_t GEP2BaseOffset;
917         bool GEP2MaxLookupReached;
918         SmallVector<VariableGEPIndex, 4> GEP2VariableIndices;
919         const Value *GEP2BasePtr =
920             DecomposeGEPExpression(GEP2, GEP2BaseOffset, GEP2VariableIndices,
921                                    GEP2MaxLookupReached, DL, &AC, DT);
922         const Value *GEP1BasePtr =
923             DecomposeGEPExpression(GEP1, GEP1BaseOffset, GEP1VariableIndices,
924                                    GEP1MaxLookupReached, DL, &AC, DT);
925         // DecomposeGEPExpression and GetUnderlyingObject should return the
926         // same result except when DecomposeGEPExpression has no DataLayout.
927         // FIXME: They always have a DataLayout so this should become an
928         // assert.
929         if (GEP1BasePtr != UnderlyingV1 || GEP2BasePtr != UnderlyingV2) {
930           return MayAlias;
931         }
932         // If the max search depth is reached the result is undefined
933         if (GEP2MaxLookupReached || GEP1MaxLookupReached)
934           return MayAlias;
935 
936         // Same offsets.
937         if (GEP1BaseOffset == GEP2BaseOffset &&
938             GEP1VariableIndices == GEP2VariableIndices)
939           return NoAlias;
940         GEP1VariableIndices.clear();
941       }
942     }
943 
944     // If we get a No or May, then return it immediately, no amount of analysis
945     // will improve this situation.
946     if (BaseAlias != MustAlias)
947       return BaseAlias;
948 
949     // Otherwise, we have a MustAlias.  Since the base pointers alias each other
950     // exactly, see if the computed offset from the common pointer tells us
951     // about the relation of the resulting pointer.
952     const Value *GEP1BasePtr =
953         DecomposeGEPExpression(GEP1, GEP1BaseOffset, GEP1VariableIndices,
954                                GEP1MaxLookupReached, DL, &AC, DT);
955 
956     int64_t GEP2BaseOffset;
957     bool GEP2MaxLookupReached;
958     SmallVector<VariableGEPIndex, 4> GEP2VariableIndices;
959     const Value *GEP2BasePtr =
960         DecomposeGEPExpression(GEP2, GEP2BaseOffset, GEP2VariableIndices,
961                                GEP2MaxLookupReached, DL, &AC, DT);
962 
963     // DecomposeGEPExpression and GetUnderlyingObject should return the
964     // same result except when DecomposeGEPExpression has no DataLayout.
965     // FIXME: They always have a DataLayout so this should become an assert.
966     if (GEP1BasePtr != UnderlyingV1 || GEP2BasePtr != UnderlyingV2) {
967       return MayAlias;
968     }
969 
970     // If we know the two GEPs are based off of the exact same pointer (and not
971     // just the same underlying object), see if that tells us anything about
972     // the resulting pointers.
973     if (GEP1->getPointerOperand() == GEP2->getPointerOperand()) {
974       AliasResult R = aliasSameBasePointerGEPs(GEP1, V1Size, GEP2, V2Size, DL);
975       // If we couldn't find anything interesting, don't abandon just yet.
976       if (R != MayAlias)
977         return R;
978     }
979 
980     // If the max search depth is reached the result is undefined
981     if (GEP2MaxLookupReached || GEP1MaxLookupReached)
982       return MayAlias;
983 
984     // Subtract the GEP2 pointer from the GEP1 pointer to find out their
985     // symbolic difference.
986     GEP1BaseOffset -= GEP2BaseOffset;
987     GetIndexDifference(GEP1VariableIndices, GEP2VariableIndices);
988 
989   } else {
990     // Check to see if these two pointers are related by the getelementptr
991     // instruction.  If one pointer is a GEP with a non-zero index of the other
992     // pointer, we know they cannot alias.
993 
994     // If both accesses are unknown size, we can't do anything useful here.
995     if (V1Size == MemoryLocation::UnknownSize &&
996         V2Size == MemoryLocation::UnknownSize)
997       return MayAlias;
998 
999     AliasResult R = aliasCheck(UnderlyingV1, MemoryLocation::UnknownSize,
1000                                AAMDNodes(), V2, V2Size, V2AAInfo);
1001     if (R != MustAlias)
1002       // If V2 may alias GEP base pointer, conservatively returns MayAlias.
1003       // If V2 is known not to alias GEP base pointer, then the two values
1004       // cannot alias per GEP semantics: "A pointer value formed from a
1005       // getelementptr instruction is associated with the addresses associated
1006       // with the first operand of the getelementptr".
1007       return R;
1008 
1009     const Value *GEP1BasePtr =
1010         DecomposeGEPExpression(GEP1, GEP1BaseOffset, GEP1VariableIndices,
1011                                GEP1MaxLookupReached, DL, &AC, DT);
1012 
1013     // DecomposeGEPExpression and GetUnderlyingObject should return the
1014     // same result except when DecomposeGEPExpression has no DataLayout.
1015     // FIXME: They always have a DataLayout so this should become an assert.
1016     if (GEP1BasePtr != UnderlyingV1) {
1017       return MayAlias;
1018     }
1019     // If the max search depth is reached the result is undefined
1020     if (GEP1MaxLookupReached)
1021       return MayAlias;
1022   }
1023 
1024   // In the two GEP Case, if there is no difference in the offsets of the
1025   // computed pointers, the resultant pointers are a must alias.  This
1026   // hapens when we have two lexically identical GEP's (for example).
1027   //
1028   // In the other case, if we have getelementptr <ptr>, 0, 0, 0, 0, ... and V2
1029   // must aliases the GEP, the end result is a must alias also.
1030   if (GEP1BaseOffset == 0 && GEP1VariableIndices.empty())
1031     return MustAlias;
1032 
1033   // If there is a constant difference between the pointers, but the difference
1034   // is less than the size of the associated memory object, then we know
1035   // that the objects are partially overlapping.  If the difference is
1036   // greater, we know they do not overlap.
1037   if (GEP1BaseOffset != 0 && GEP1VariableIndices.empty()) {
1038     if (GEP1BaseOffset >= 0) {
1039       if (V2Size != MemoryLocation::UnknownSize) {
1040         if ((uint64_t)GEP1BaseOffset < V2Size)
1041           return PartialAlias;
1042         return NoAlias;
1043       }
1044     } else {
1045       // We have the situation where:
1046       // +                +
1047       // | BaseOffset     |
1048       // ---------------->|
1049       // |-->V1Size       |-------> V2Size
1050       // GEP1             V2
1051       // We need to know that V2Size is not unknown, otherwise we might have
1052       // stripped a gep with negative index ('gep <ptr>, -1, ...).
1053       if (V1Size != MemoryLocation::UnknownSize &&
1054           V2Size != MemoryLocation::UnknownSize) {
1055         if (-(uint64_t)GEP1BaseOffset < V1Size)
1056           return PartialAlias;
1057         return NoAlias;
1058       }
1059     }
1060   }
1061 
1062   if (!GEP1VariableIndices.empty()) {
1063     uint64_t Modulo = 0;
1064     bool AllPositive = true;
1065     for (unsigned i = 0, e = GEP1VariableIndices.size(); i != e; ++i) {
1066 
1067       // Try to distinguish something like &A[i][1] against &A[42][0].
1068       // Grab the least significant bit set in any of the scales. We
1069       // don't need std::abs here (even if the scale's negative) as we'll
1070       // be ^'ing Modulo with itself later.
1071       Modulo |= (uint64_t)GEP1VariableIndices[i].Scale;
1072 
1073       if (AllPositive) {
1074         // If the Value could change between cycles, then any reasoning about
1075         // the Value this cycle may not hold in the next cycle. We'll just
1076         // give up if we can't determine conditions that hold for every cycle:
1077         const Value *V = GEP1VariableIndices[i].V;
1078 
1079         bool SignKnownZero, SignKnownOne;
1080         ComputeSignBit(const_cast<Value *>(V), SignKnownZero, SignKnownOne, DL,
1081                        0, &AC, nullptr, DT);
1082 
1083         // Zero-extension widens the variable, and so forces the sign
1084         // bit to zero.
1085         bool IsZExt = GEP1VariableIndices[i].ZExtBits > 0 || isa<ZExtInst>(V);
1086         SignKnownZero |= IsZExt;
1087         SignKnownOne &= !IsZExt;
1088 
1089         // If the variable begins with a zero then we know it's
1090         // positive, regardless of whether the value is signed or
1091         // unsigned.
1092         int64_t Scale = GEP1VariableIndices[i].Scale;
1093         AllPositive =
1094             (SignKnownZero && Scale >= 0) || (SignKnownOne && Scale < 0);
1095       }
1096     }
1097 
1098     Modulo = Modulo ^ (Modulo & (Modulo - 1));
1099 
1100     // We can compute the difference between the two addresses
1101     // mod Modulo. Check whether that difference guarantees that the
1102     // two locations do not alias.
1103     uint64_t ModOffset = (uint64_t)GEP1BaseOffset & (Modulo - 1);
1104     if (V1Size != MemoryLocation::UnknownSize &&
1105         V2Size != MemoryLocation::UnknownSize && ModOffset >= V2Size &&
1106         V1Size <= Modulo - ModOffset)
1107       return NoAlias;
1108 
1109     // If we know all the variables are positive, then GEP1 >= GEP1BasePtr.
1110     // If GEP1BasePtr > V2 (GEP1BaseOffset > 0) then we know the pointers
1111     // don't alias if V2Size can fit in the gap between V2 and GEP1BasePtr.
1112     if (AllPositive && GEP1BaseOffset > 0 && V2Size <= (uint64_t)GEP1BaseOffset)
1113       return NoAlias;
1114 
1115     if (constantOffsetHeuristic(GEP1VariableIndices, V1Size, V2Size,
1116                                 GEP1BaseOffset, &AC, DT))
1117       return NoAlias;
1118   }
1119 
1120   // Statically, we can see that the base objects are the same, but the
1121   // pointers have dynamic offsets which we can't resolve. And none of our
1122   // little tricks above worked.
1123   //
1124   // TODO: Returning PartialAlias instead of MayAlias is a mild hack; the
1125   // practical effect of this is protecting TBAA in the case of dynamic
1126   // indices into arrays of unions or malloc'd memory.
1127   return PartialAlias;
1128 }
1129 
MergeAliasResults(AliasResult A,AliasResult B)1130 static AliasResult MergeAliasResults(AliasResult A, AliasResult B) {
1131   // If the results agree, take it.
1132   if (A == B)
1133     return A;
1134   // A mix of PartialAlias and MustAlias is PartialAlias.
1135   if ((A == PartialAlias && B == MustAlias) ||
1136       (B == PartialAlias && A == MustAlias))
1137     return PartialAlias;
1138   // Otherwise, we don't know anything.
1139   return MayAlias;
1140 }
1141 
1142 /// Provides a bunch of ad-hoc rules to disambiguate a Select instruction
1143 /// against another.
aliasSelect(const SelectInst * SI,uint64_t SISize,const AAMDNodes & SIAAInfo,const Value * V2,uint64_t V2Size,const AAMDNodes & V2AAInfo)1144 AliasResult BasicAAResult::aliasSelect(const SelectInst *SI, uint64_t SISize,
1145                                        const AAMDNodes &SIAAInfo,
1146                                        const Value *V2, uint64_t V2Size,
1147                                        const AAMDNodes &V2AAInfo) {
1148   // If the values are Selects with the same condition, we can do a more precise
1149   // check: just check for aliases between the values on corresponding arms.
1150   if (const SelectInst *SI2 = dyn_cast<SelectInst>(V2))
1151     if (SI->getCondition() == SI2->getCondition()) {
1152       AliasResult Alias = aliasCheck(SI->getTrueValue(), SISize, SIAAInfo,
1153                                      SI2->getTrueValue(), V2Size, V2AAInfo);
1154       if (Alias == MayAlias)
1155         return MayAlias;
1156       AliasResult ThisAlias =
1157           aliasCheck(SI->getFalseValue(), SISize, SIAAInfo,
1158                      SI2->getFalseValue(), V2Size, V2AAInfo);
1159       return MergeAliasResults(ThisAlias, Alias);
1160     }
1161 
1162   // If both arms of the Select node NoAlias or MustAlias V2, then returns
1163   // NoAlias / MustAlias. Otherwise, returns MayAlias.
1164   AliasResult Alias =
1165       aliasCheck(V2, V2Size, V2AAInfo, SI->getTrueValue(), SISize, SIAAInfo);
1166   if (Alias == MayAlias)
1167     return MayAlias;
1168 
1169   AliasResult ThisAlias =
1170       aliasCheck(V2, V2Size, V2AAInfo, SI->getFalseValue(), SISize, SIAAInfo);
1171   return MergeAliasResults(ThisAlias, Alias);
1172 }
1173 
1174 /// Provide a bunch of ad-hoc rules to disambiguate a PHI instruction against
1175 /// another.
aliasPHI(const PHINode * PN,uint64_t PNSize,const AAMDNodes & PNAAInfo,const Value * V2,uint64_t V2Size,const AAMDNodes & V2AAInfo)1176 AliasResult BasicAAResult::aliasPHI(const PHINode *PN, uint64_t PNSize,
1177                                     const AAMDNodes &PNAAInfo, const Value *V2,
1178                                     uint64_t V2Size,
1179                                     const AAMDNodes &V2AAInfo) {
1180   // Track phi nodes we have visited. We use this information when we determine
1181   // value equivalence.
1182   VisitedPhiBBs.insert(PN->getParent());
1183 
1184   // If the values are PHIs in the same block, we can do a more precise
1185   // as well as efficient check: just check for aliases between the values
1186   // on corresponding edges.
1187   if (const PHINode *PN2 = dyn_cast<PHINode>(V2))
1188     if (PN2->getParent() == PN->getParent()) {
1189       LocPair Locs(MemoryLocation(PN, PNSize, PNAAInfo),
1190                    MemoryLocation(V2, V2Size, V2AAInfo));
1191       if (PN > V2)
1192         std::swap(Locs.first, Locs.second);
1193       // Analyse the PHIs' inputs under the assumption that the PHIs are
1194       // NoAlias.
1195       // If the PHIs are May/MustAlias there must be (recursively) an input
1196       // operand from outside the PHIs' cycle that is MayAlias/MustAlias or
1197       // there must be an operation on the PHIs within the PHIs' value cycle
1198       // that causes a MayAlias.
1199       // Pretend the phis do not alias.
1200       AliasResult Alias = NoAlias;
1201       assert(AliasCache.count(Locs) &&
1202              "There must exist an entry for the phi node");
1203       AliasResult OrigAliasResult = AliasCache[Locs];
1204       AliasCache[Locs] = NoAlias;
1205 
1206       for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1207         AliasResult ThisAlias =
1208             aliasCheck(PN->getIncomingValue(i), PNSize, PNAAInfo,
1209                        PN2->getIncomingValueForBlock(PN->getIncomingBlock(i)),
1210                        V2Size, V2AAInfo);
1211         Alias = MergeAliasResults(ThisAlias, Alias);
1212         if (Alias == MayAlias)
1213           break;
1214       }
1215 
1216       // Reset if speculation failed.
1217       if (Alias != NoAlias)
1218         AliasCache[Locs] = OrigAliasResult;
1219 
1220       return Alias;
1221     }
1222 
1223   SmallPtrSet<Value *, 4> UniqueSrc;
1224   SmallVector<Value *, 4> V1Srcs;
1225   bool isRecursive = false;
1226   for (Value *PV1 : PN->incoming_values()) {
1227     if (isa<PHINode>(PV1))
1228       // If any of the source itself is a PHI, return MayAlias conservatively
1229       // to avoid compile time explosion. The worst possible case is if both
1230       // sides are PHI nodes. In which case, this is O(m x n) time where 'm'
1231       // and 'n' are the number of PHI sources.
1232       return MayAlias;
1233 
1234     if (EnableRecPhiAnalysis)
1235       if (GEPOperator *PV1GEP = dyn_cast<GEPOperator>(PV1)) {
1236         // Check whether the incoming value is a GEP that advances the pointer
1237         // result of this PHI node (e.g. in a loop). If this is the case, we
1238         // would recurse and always get a MayAlias. Handle this case specially
1239         // below.
1240         if (PV1GEP->getPointerOperand() == PN && PV1GEP->getNumIndices() == 1 &&
1241             isa<ConstantInt>(PV1GEP->idx_begin())) {
1242           isRecursive = true;
1243           continue;
1244         }
1245       }
1246 
1247     if (UniqueSrc.insert(PV1).second)
1248       V1Srcs.push_back(PV1);
1249   }
1250 
1251   // If this PHI node is recursive, set the size of the accessed memory to
1252   // unknown to represent all the possible values the GEP could advance the
1253   // pointer to.
1254   if (isRecursive)
1255     PNSize = MemoryLocation::UnknownSize;
1256 
1257   AliasResult Alias =
1258       aliasCheck(V2, V2Size, V2AAInfo, V1Srcs[0], PNSize, PNAAInfo);
1259 
1260   // Early exit if the check of the first PHI source against V2 is MayAlias.
1261   // Other results are not possible.
1262   if (Alias == MayAlias)
1263     return MayAlias;
1264 
1265   // If all sources of the PHI node NoAlias or MustAlias V2, then returns
1266   // NoAlias / MustAlias. Otherwise, returns MayAlias.
1267   for (unsigned i = 1, e = V1Srcs.size(); i != e; ++i) {
1268     Value *V = V1Srcs[i];
1269 
1270     AliasResult ThisAlias =
1271         aliasCheck(V2, V2Size, V2AAInfo, V, PNSize, PNAAInfo);
1272     Alias = MergeAliasResults(ThisAlias, Alias);
1273     if (Alias == MayAlias)
1274       break;
1275   }
1276 
1277   return Alias;
1278 }
1279 
1280 /// Provides a bunch of ad-hoc rules to disambiguate in common cases, such as
1281 /// array references.
aliasCheck(const Value * V1,uint64_t V1Size,AAMDNodes V1AAInfo,const Value * V2,uint64_t V2Size,AAMDNodes V2AAInfo)1282 AliasResult BasicAAResult::aliasCheck(const Value *V1, uint64_t V1Size,
1283                                       AAMDNodes V1AAInfo, const Value *V2,
1284                                       uint64_t V2Size, AAMDNodes V2AAInfo) {
1285   // If either of the memory references is empty, it doesn't matter what the
1286   // pointer values are.
1287   if (V1Size == 0 || V2Size == 0)
1288     return NoAlias;
1289 
1290   // Strip off any casts if they exist.
1291   V1 = V1->stripPointerCasts();
1292   V2 = V2->stripPointerCasts();
1293 
1294   // If V1 or V2 is undef, the result is NoAlias because we can always pick a
1295   // value for undef that aliases nothing in the program.
1296   if (isa<UndefValue>(V1) || isa<UndefValue>(V2))
1297     return NoAlias;
1298 
1299   // Are we checking for alias of the same value?
1300   // Because we look 'through' phi nodes we could look at "Value" pointers from
1301   // different iterations. We must therefore make sure that this is not the
1302   // case. The function isValueEqualInPotentialCycles ensures that this cannot
1303   // happen by looking at the visited phi nodes and making sure they cannot
1304   // reach the value.
1305   if (isValueEqualInPotentialCycles(V1, V2))
1306     return MustAlias;
1307 
1308   if (!V1->getType()->isPointerTy() || !V2->getType()->isPointerTy())
1309     return NoAlias; // Scalars cannot alias each other
1310 
1311   // Figure out what objects these things are pointing to if we can.
1312   const Value *O1 = GetUnderlyingObject(V1, DL, MaxLookupSearchDepth);
1313   const Value *O2 = GetUnderlyingObject(V2, DL, MaxLookupSearchDepth);
1314 
1315   // Null values in the default address space don't point to any object, so they
1316   // don't alias any other pointer.
1317   if (const ConstantPointerNull *CPN = dyn_cast<ConstantPointerNull>(O1))
1318     if (CPN->getType()->getAddressSpace() == 0)
1319       return NoAlias;
1320   if (const ConstantPointerNull *CPN = dyn_cast<ConstantPointerNull>(O2))
1321     if (CPN->getType()->getAddressSpace() == 0)
1322       return NoAlias;
1323 
1324   if (O1 != O2) {
1325     // If V1/V2 point to two different objects we know that we have no alias.
1326     if (isIdentifiedObject(O1) && isIdentifiedObject(O2))
1327       return NoAlias;
1328 
1329     // Constant pointers can't alias with non-const isIdentifiedObject objects.
1330     if ((isa<Constant>(O1) && isIdentifiedObject(O2) && !isa<Constant>(O2)) ||
1331         (isa<Constant>(O2) && isIdentifiedObject(O1) && !isa<Constant>(O1)))
1332       return NoAlias;
1333 
1334     // Function arguments can't alias with things that are known to be
1335     // unambigously identified at the function level.
1336     if ((isa<Argument>(O1) && isIdentifiedFunctionLocal(O2)) ||
1337         (isa<Argument>(O2) && isIdentifiedFunctionLocal(O1)))
1338       return NoAlias;
1339 
1340     // Most objects can't alias null.
1341     if ((isa<ConstantPointerNull>(O2) && isKnownNonNull(O1)) ||
1342         (isa<ConstantPointerNull>(O1) && isKnownNonNull(O2)))
1343       return NoAlias;
1344 
1345     // If one pointer is the result of a call/invoke or load and the other is a
1346     // non-escaping local object within the same function, then we know the
1347     // object couldn't escape to a point where the call could return it.
1348     //
1349     // Note that if the pointers are in different functions, there are a
1350     // variety of complications. A call with a nocapture argument may still
1351     // temporary store the nocapture argument's value in a temporary memory
1352     // location if that memory location doesn't escape. Or it may pass a
1353     // nocapture value to other functions as long as they don't capture it.
1354     if (isEscapeSource(O1) && isNonEscapingLocalObject(O2))
1355       return NoAlias;
1356     if (isEscapeSource(O2) && isNonEscapingLocalObject(O1))
1357       return NoAlias;
1358   }
1359 
1360   // If the size of one access is larger than the entire object on the other
1361   // side, then we know such behavior is undefined and can assume no alias.
1362   if ((V1Size != MemoryLocation::UnknownSize &&
1363        isObjectSmallerThan(O2, V1Size, DL, TLI)) ||
1364       (V2Size != MemoryLocation::UnknownSize &&
1365        isObjectSmallerThan(O1, V2Size, DL, TLI)))
1366     return NoAlias;
1367 
1368   // Check the cache before climbing up use-def chains. This also terminates
1369   // otherwise infinitely recursive queries.
1370   LocPair Locs(MemoryLocation(V1, V1Size, V1AAInfo),
1371                MemoryLocation(V2, V2Size, V2AAInfo));
1372   if (V1 > V2)
1373     std::swap(Locs.first, Locs.second);
1374   std::pair<AliasCacheTy::iterator, bool> Pair =
1375       AliasCache.insert(std::make_pair(Locs, MayAlias));
1376   if (!Pair.second)
1377     return Pair.first->second;
1378 
1379   // FIXME: This isn't aggressively handling alias(GEP, PHI) for example: if the
1380   // GEP can't simplify, we don't even look at the PHI cases.
1381   if (!isa<GEPOperator>(V1) && isa<GEPOperator>(V2)) {
1382     std::swap(V1, V2);
1383     std::swap(V1Size, V2Size);
1384     std::swap(O1, O2);
1385     std::swap(V1AAInfo, V2AAInfo);
1386   }
1387   if (const GEPOperator *GV1 = dyn_cast<GEPOperator>(V1)) {
1388     AliasResult Result =
1389         aliasGEP(GV1, V1Size, V1AAInfo, V2, V2Size, V2AAInfo, O1, O2);
1390     if (Result != MayAlias)
1391       return AliasCache[Locs] = Result;
1392   }
1393 
1394   if (isa<PHINode>(V2) && !isa<PHINode>(V1)) {
1395     std::swap(V1, V2);
1396     std::swap(V1Size, V2Size);
1397     std::swap(V1AAInfo, V2AAInfo);
1398   }
1399   if (const PHINode *PN = dyn_cast<PHINode>(V1)) {
1400     AliasResult Result = aliasPHI(PN, V1Size, V1AAInfo, V2, V2Size, V2AAInfo);
1401     if (Result != MayAlias)
1402       return AliasCache[Locs] = Result;
1403   }
1404 
1405   if (isa<SelectInst>(V2) && !isa<SelectInst>(V1)) {
1406     std::swap(V1, V2);
1407     std::swap(V1Size, V2Size);
1408     std::swap(V1AAInfo, V2AAInfo);
1409   }
1410   if (const SelectInst *S1 = dyn_cast<SelectInst>(V1)) {
1411     AliasResult Result =
1412         aliasSelect(S1, V1Size, V1AAInfo, V2, V2Size, V2AAInfo);
1413     if (Result != MayAlias)
1414       return AliasCache[Locs] = Result;
1415   }
1416 
1417   // If both pointers are pointing into the same object and one of them
1418   // accesses is accessing the entire object, then the accesses must
1419   // overlap in some way.
1420   if (O1 == O2)
1421     if ((V1Size != MemoryLocation::UnknownSize &&
1422          isObjectSize(O1, V1Size, DL, TLI)) ||
1423         (V2Size != MemoryLocation::UnknownSize &&
1424          isObjectSize(O2, V2Size, DL, TLI)))
1425       return AliasCache[Locs] = PartialAlias;
1426 
1427   // Recurse back into the best AA results we have, potentially with refined
1428   // memory locations. We have already ensured that BasicAA has a MayAlias
1429   // cache result for these, so any recursion back into BasicAA won't loop.
1430   AliasResult Result = getBestAAResults().alias(Locs.first, Locs.second);
1431   return AliasCache[Locs] = Result;
1432 }
1433 
1434 /// Check whether two Values can be considered equivalent.
1435 ///
1436 /// In addition to pointer equivalence of \p V1 and \p V2 this checks whether
1437 /// they can not be part of a cycle in the value graph by looking at all
1438 /// visited phi nodes an making sure that the phis cannot reach the value. We
1439 /// have to do this because we are looking through phi nodes (That is we say
1440 /// noalias(V, phi(VA, VB)) if noalias(V, VA) and noalias(V, VB).
isValueEqualInPotentialCycles(const Value * V,const Value * V2)1441 bool BasicAAResult::isValueEqualInPotentialCycles(const Value *V,
1442                                                   const Value *V2) {
1443   if (V != V2)
1444     return false;
1445 
1446   const Instruction *Inst = dyn_cast<Instruction>(V);
1447   if (!Inst)
1448     return true;
1449 
1450   if (VisitedPhiBBs.empty())
1451     return true;
1452 
1453   if (VisitedPhiBBs.size() > MaxNumPhiBBsValueReachabilityCheck)
1454     return false;
1455 
1456   // Make sure that the visited phis cannot reach the Value. This ensures that
1457   // the Values cannot come from different iterations of a potential cycle the
1458   // phi nodes could be involved in.
1459   for (auto *P : VisitedPhiBBs)
1460     if (isPotentiallyReachable(&P->front(), Inst, DT, LI))
1461       return false;
1462 
1463   return true;
1464 }
1465 
1466 /// Computes the symbolic difference between two de-composed GEPs.
1467 ///
1468 /// Dest and Src are the variable indices from two decomposed GetElementPtr
1469 /// instructions GEP1 and GEP2 which have common base pointers.
GetIndexDifference(SmallVectorImpl<VariableGEPIndex> & Dest,const SmallVectorImpl<VariableGEPIndex> & Src)1470 void BasicAAResult::GetIndexDifference(
1471     SmallVectorImpl<VariableGEPIndex> &Dest,
1472     const SmallVectorImpl<VariableGEPIndex> &Src) {
1473   if (Src.empty())
1474     return;
1475 
1476   for (unsigned i = 0, e = Src.size(); i != e; ++i) {
1477     const Value *V = Src[i].V;
1478     unsigned ZExtBits = Src[i].ZExtBits, SExtBits = Src[i].SExtBits;
1479     int64_t Scale = Src[i].Scale;
1480 
1481     // Find V in Dest.  This is N^2, but pointer indices almost never have more
1482     // than a few variable indexes.
1483     for (unsigned j = 0, e = Dest.size(); j != e; ++j) {
1484       if (!isValueEqualInPotentialCycles(Dest[j].V, V) ||
1485           Dest[j].ZExtBits != ZExtBits || Dest[j].SExtBits != SExtBits)
1486         continue;
1487 
1488       // If we found it, subtract off Scale V's from the entry in Dest.  If it
1489       // goes to zero, remove the entry.
1490       if (Dest[j].Scale != Scale)
1491         Dest[j].Scale -= Scale;
1492       else
1493         Dest.erase(Dest.begin() + j);
1494       Scale = 0;
1495       break;
1496     }
1497 
1498     // If we didn't consume this entry, add it to the end of the Dest list.
1499     if (Scale) {
1500       VariableGEPIndex Entry = {V, ZExtBits, SExtBits, -Scale};
1501       Dest.push_back(Entry);
1502     }
1503   }
1504 }
1505 
constantOffsetHeuristic(const SmallVectorImpl<VariableGEPIndex> & VarIndices,uint64_t V1Size,uint64_t V2Size,int64_t BaseOffset,AssumptionCache * AC,DominatorTree * DT)1506 bool BasicAAResult::constantOffsetHeuristic(
1507     const SmallVectorImpl<VariableGEPIndex> &VarIndices, uint64_t V1Size,
1508     uint64_t V2Size, int64_t BaseOffset, AssumptionCache *AC,
1509     DominatorTree *DT) {
1510   if (VarIndices.size() != 2 || V1Size == MemoryLocation::UnknownSize ||
1511       V2Size == MemoryLocation::UnknownSize)
1512     return false;
1513 
1514   const VariableGEPIndex &Var0 = VarIndices[0], &Var1 = VarIndices[1];
1515 
1516   if (Var0.ZExtBits != Var1.ZExtBits || Var0.SExtBits != Var1.SExtBits ||
1517       Var0.Scale != -Var1.Scale)
1518     return false;
1519 
1520   unsigned Width = Var1.V->getType()->getIntegerBitWidth();
1521 
1522   // We'll strip off the Extensions of Var0 and Var1 and do another round
1523   // of GetLinearExpression decomposition. In the example above, if Var0
1524   // is zext(%x + 1) we should get V1 == %x and V1Offset == 1.
1525 
1526   APInt V0Scale(Width, 0), V0Offset(Width, 0), V1Scale(Width, 0),
1527       V1Offset(Width, 0);
1528   bool NSW = true, NUW = true;
1529   unsigned V0ZExtBits = 0, V0SExtBits = 0, V1ZExtBits = 0, V1SExtBits = 0;
1530   const Value *V0 = GetLinearExpression(Var0.V, V0Scale, V0Offset, V0ZExtBits,
1531                                         V0SExtBits, DL, 0, AC, DT, NSW, NUW);
1532   NSW = true, NUW = true;
1533   const Value *V1 = GetLinearExpression(Var1.V, V1Scale, V1Offset, V1ZExtBits,
1534                                         V1SExtBits, DL, 0, AC, DT, NSW, NUW);
1535 
1536   if (V0Scale != V1Scale || V0ZExtBits != V1ZExtBits ||
1537       V0SExtBits != V1SExtBits || !isValueEqualInPotentialCycles(V0, V1))
1538     return false;
1539 
1540   // We have a hit - Var0 and Var1 only differ by a constant offset!
1541 
1542   // If we've been sext'ed then zext'd the maximum difference between Var0 and
1543   // Var1 is possible to calculate, but we're just interested in the absolute
1544   // minimum difference between the two. The minimum distance may occur due to
1545   // wrapping; consider "add i3 %i, 5": if %i == 7 then 7 + 5 mod 8 == 4, and so
1546   // the minimum distance between %i and %i + 5 is 3.
1547   APInt MinDiff = V0Offset - V1Offset, Wrapped = -MinDiff;
1548   MinDiff = APIntOps::umin(MinDiff, Wrapped);
1549   uint64_t MinDiffBytes = MinDiff.getZExtValue() * std::abs(Var0.Scale);
1550 
1551   // We can't definitely say whether GEP1 is before or after V2 due to wrapping
1552   // arithmetic (i.e. for some values of GEP1 and V2 GEP1 < V2, and for other
1553   // values GEP1 > V2). We'll therefore only declare NoAlias if both V1Size and
1554   // V2Size can fit in the MinDiffBytes gap.
1555   return V1Size + std::abs(BaseOffset) <= MinDiffBytes &&
1556          V2Size + std::abs(BaseOffset) <= MinDiffBytes;
1557 }
1558 
1559 //===----------------------------------------------------------------------===//
1560 // BasicAliasAnalysis Pass
1561 //===----------------------------------------------------------------------===//
1562 
1563 char BasicAA::PassID;
1564 
run(Function & F,AnalysisManager<Function> * AM)1565 BasicAAResult BasicAA::run(Function &F, AnalysisManager<Function> *AM) {
1566   return BasicAAResult(F.getParent()->getDataLayout(),
1567                        AM->getResult<TargetLibraryAnalysis>(F),
1568                        AM->getResult<AssumptionAnalysis>(F),
1569                        AM->getCachedResult<DominatorTreeAnalysis>(F),
1570                        AM->getCachedResult<LoopAnalysis>(F));
1571 }
1572 
BasicAAWrapperPass()1573 BasicAAWrapperPass::BasicAAWrapperPass() : FunctionPass(ID) {
1574     initializeBasicAAWrapperPassPass(*PassRegistry::getPassRegistry());
1575 }
1576 
1577 char BasicAAWrapperPass::ID = 0;
anchor()1578 void BasicAAWrapperPass::anchor() {}
1579 
1580 INITIALIZE_PASS_BEGIN(BasicAAWrapperPass, "basicaa",
1581                       "Basic Alias Analysis (stateless AA impl)", true, true)
INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)1582 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
1583 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
1584 INITIALIZE_PASS_END(BasicAAWrapperPass, "basicaa",
1585                     "Basic Alias Analysis (stateless AA impl)", true, true)
1586 
1587 FunctionPass *llvm::createBasicAAWrapperPass() {
1588   return new BasicAAWrapperPass();
1589 }
1590 
runOnFunction(Function & F)1591 bool BasicAAWrapperPass::runOnFunction(Function &F) {
1592   auto &ACT = getAnalysis<AssumptionCacheTracker>();
1593   auto &TLIWP = getAnalysis<TargetLibraryInfoWrapperPass>();
1594   auto *DTWP = getAnalysisIfAvailable<DominatorTreeWrapperPass>();
1595   auto *LIWP = getAnalysisIfAvailable<LoopInfoWrapperPass>();
1596 
1597   Result.reset(new BasicAAResult(F.getParent()->getDataLayout(), TLIWP.getTLI(),
1598                                  ACT.getAssumptionCache(F),
1599                                  DTWP ? &DTWP->getDomTree() : nullptr,
1600                                  LIWP ? &LIWP->getLoopInfo() : nullptr));
1601 
1602   return false;
1603 }
1604 
getAnalysisUsage(AnalysisUsage & AU) const1605 void BasicAAWrapperPass::getAnalysisUsage(AnalysisUsage &AU) const {
1606   AU.setPreservesAll();
1607   AU.addRequired<AssumptionCacheTracker>();
1608   AU.addRequired<TargetLibraryInfoWrapperPass>();
1609 }
1610 
createLegacyPMBasicAAResult(Pass & P,Function & F)1611 BasicAAResult llvm::createLegacyPMBasicAAResult(Pass &P, Function &F) {
1612   return BasicAAResult(
1613       F.getParent()->getDataLayout(),
1614       P.getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(),
1615       P.getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F));
1616 }
1617