1 //===- BasicAliasAnalysis.cpp - Stateless Alias Analysis Impl -------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
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.
14 //===----------------------------------------------------------------------===//
16 #include "llvm/Analysis/BasicAliasAnalysis.h"
17 #include "llvm/ADT/APInt.h"
18 #include "llvm/ADT/SmallPtrSet.h"
19 #include "llvm/ADT/SmallVector.h"
20 #include "llvm/ADT/Statistic.h"
21 #include "llvm/Analysis/AliasAnalysis.h"
22 #include "llvm/Analysis/AssumptionCache.h"
23 #include "llvm/Analysis/CFG.h"
24 #include "llvm/Analysis/CaptureTracking.h"
25 #include "llvm/Analysis/InstructionSimplify.h"
26 #include "llvm/Analysis/LoopInfo.h"
27 #include "llvm/Analysis/MemoryBuiltins.h"
28 #include "llvm/Analysis/MemoryLocation.h"
29 #include "llvm/Analysis/TargetLibraryInfo.h"
30 #include "llvm/Analysis/ValueTracking.h"
31 #include "llvm/IR/Argument.h"
32 #include "llvm/IR/Attributes.h"
33 #include "llvm/IR/CallSite.h"
34 #include "llvm/IR/Constant.h"
35 #include "llvm/IR/Constants.h"
36 #include "llvm/IR/DataLayout.h"
37 #include "llvm/IR/DerivedTypes.h"
38 #include "llvm/IR/Dominators.h"
39 #include "llvm/IR/Function.h"
40 #include "llvm/IR/GetElementPtrTypeIterator.h"
41 #include "llvm/IR/GlobalAlias.h"
42 #include "llvm/IR/GlobalVariable.h"
43 #include "llvm/IR/InstrTypes.h"
44 #include "llvm/IR/Instruction.h"
45 #include "llvm/IR/Instructions.h"
46 #include "llvm/IR/IntrinsicInst.h"
47 #include "llvm/IR/Intrinsics.h"
48 #include "llvm/IR/Metadata.h"
49 #include "llvm/IR/Operator.h"
50 #include "llvm/IR/Type.h"
51 #include "llvm/IR/User.h"
52 #include "llvm/IR/Value.h"
53 #include "llvm/Pass.h"
54 #include "llvm/Support/Casting.h"
55 #include "llvm/Support/CommandLine.h"
56 #include "llvm/Support/Compiler.h"
57 #include "llvm/Support/KnownBits.h"
63 #define DEBUG_TYPE "basicaa"
67 /// Enable analysis of recursive PHI nodes.
68 static cl::opt<bool> EnableRecPhiAnalysis("basicaa-recphi", cl::Hidden,
70 /// SearchLimitReached / SearchTimes shows how often the limit of
71 /// to decompose GEPs is reached. It will affect the precision
72 /// of basic alias analysis.
73 STATISTIC(SearchLimitReached, "Number of times the limit to "
74 "decompose GEPs is reached");
75 STATISTIC(SearchTimes, "Number of times a GEP is decomposed");
77 /// Cutoff after which to stop analysing a set of phi nodes potentially involved
78 /// in a cycle. Because we are analysing 'through' phi nodes, we need to be
79 /// careful with value equivalence. We use reachability to make sure a value
80 /// cannot be involved in a cycle.
81 const unsigned MaxNumPhiBBsValueReachabilityCheck = 20;
83 // The max limit of the search depth in DecomposeGEPExpression() and
84 // GetUnderlyingObject(), both functions need to use the same search
85 // depth otherwise the algorithm in aliasGEP will assert.
86 static const unsigned MaxLookupSearchDepth = 6;
88 bool BasicAAResult::invalidate(Function &F, const PreservedAnalyses &PA,
89 FunctionAnalysisManager::Invalidator &Inv) {
90 // We don't care if this analysis itself is preserved, it has no state. But
91 // we need to check that the analyses it depends on have been. Note that we
92 // may be created without handles to some analyses and in that case don't
94 if (Inv.invalidate<AssumptionAnalysis>(F, PA) ||
95 (DT && Inv.invalidate<DominatorTreeAnalysis>(F, PA)) ||
96 (LI && Inv.invalidate<LoopAnalysis>(F, PA)))
99 // Otherwise this analysis result remains valid.
103 //===----------------------------------------------------------------------===//
105 //===----------------------------------------------------------------------===//
107 /// Returns true if the pointer is to a function-local object that never
108 /// escapes from the function.
109 static bool isNonEscapingLocalObject(const Value *V) {
110 // If this is a local allocation, check to see if it escapes.
111 if (isa<AllocaInst>(V) || isNoAliasCall(V))
112 // Set StoreCaptures to True so that we can assume in our callers that the
113 // pointer is not the result of a load instruction. Currently
114 // PointerMayBeCaptured doesn't have any special analysis for the
115 // StoreCaptures=false case; if it did, our callers could be refined to be
117 return !PointerMayBeCaptured(V, false, /*StoreCaptures=*/true);
119 // If this is an argument that corresponds to a byval or noalias argument,
120 // then it has not escaped before entering the function. Check if it escapes
121 // inside the function.
122 if (const Argument *A = dyn_cast<Argument>(V))
123 if (A->hasByValAttr() || A->hasNoAliasAttr())
124 // Note even if the argument is marked nocapture, we still need to check
125 // for copies made inside the function. The nocapture attribute only
126 // specifies that there are no copies made that outlive the function.
127 return !PointerMayBeCaptured(V, false, /*StoreCaptures=*/true);
132 /// Returns true if the pointer is one which would have been considered an
133 /// escape by isNonEscapingLocalObject.
134 static bool isEscapeSource(const Value *V) {
135 if (isa<CallInst>(V) || isa<InvokeInst>(V) || isa<Argument>(V))
138 // The load case works because isNonEscapingLocalObject considers all
139 // stores to be escapes (it passes true for the StoreCaptures argument
140 // to PointerMayBeCaptured).
141 if (isa<LoadInst>(V))
147 /// Returns the size of the object specified by V or UnknownSize if unknown.
148 static uint64_t getObjectSize(const Value *V, const DataLayout &DL,
149 const TargetLibraryInfo &TLI,
150 bool RoundToAlign = false) {
153 Opts.RoundToAlign = RoundToAlign;
154 if (getObjectSize(V, Size, DL, &TLI, Opts))
156 return MemoryLocation::UnknownSize;
159 /// Returns true if we can prove that the object specified by V is smaller than
161 static bool isObjectSmallerThan(const Value *V, uint64_t Size,
162 const DataLayout &DL,
163 const TargetLibraryInfo &TLI) {
164 // Note that the meanings of the "object" are slightly different in the
165 // following contexts:
166 // c1: llvm::getObjectSize()
167 // c2: llvm.objectsize() intrinsic
168 // c3: isObjectSmallerThan()
169 // c1 and c2 share the same meaning; however, the meaning of "object" in c3
170 // refers to the "entire object".
172 // Consider this example:
173 // char *p = (char*)malloc(100)
176 // In the context of c1 and c2, the "object" pointed by q refers to the
177 // stretch of memory of q[0:19]. So, getObjectSize(q) should return 20.
179 // However, in the context of c3, the "object" refers to the chunk of memory
180 // being allocated. So, the "object" has 100 bytes, and q points to the middle
181 // the "object". In case q is passed to isObjectSmallerThan() as the 1st
182 // parameter, before the llvm::getObjectSize() is called to get the size of
183 // entire object, we should:
184 // - either rewind the pointer q to the base-address of the object in
185 // question (in this case rewind to p), or
186 // - just give up. It is up to caller to make sure the pointer is pointing
187 // to the base address the object.
189 // We go for 2nd option for simplicity.
190 if (!isIdentifiedObject(V))
193 // This function needs to use the aligned object size because we allow
194 // reads a bit past the end given sufficient alignment.
195 uint64_t ObjectSize = getObjectSize(V, DL, TLI, /*RoundToAlign*/ true);
197 return ObjectSize != MemoryLocation::UnknownSize && ObjectSize < Size;
200 /// Returns true if we can prove that the object specified by V has size Size.
201 static bool isObjectSize(const Value *V, uint64_t Size, const DataLayout &DL,
202 const TargetLibraryInfo &TLI) {
203 uint64_t ObjectSize = getObjectSize(V, DL, TLI);
204 return ObjectSize != MemoryLocation::UnknownSize && ObjectSize == Size;
207 //===----------------------------------------------------------------------===//
208 // GetElementPtr Instruction Decomposition and Analysis
209 //===----------------------------------------------------------------------===//
211 /// Analyzes the specified value as a linear expression: "A*V + B", where A and
212 /// B are constant integers.
214 /// Returns the scale and offset values as APInts and return V as a Value*, and
215 /// return whether we looked through any sign or zero extends. The incoming
216 /// Value is known to have IntegerType, and it may already be sign or zero
219 /// Note that this looks through extends, so the high bits may not be
220 /// represented in the result.
221 /*static*/ const Value *BasicAAResult::GetLinearExpression(
222 const Value *V, APInt &Scale, APInt &Offset, unsigned &ZExtBits,
223 unsigned &SExtBits, const DataLayout &DL, unsigned Depth,
224 AssumptionCache *AC, DominatorTree *DT, bool &NSW, bool &NUW) {
225 assert(V->getType()->isIntegerTy() && "Not an integer value");
227 // Limit our recursion depth.
234 if (const ConstantInt *Const = dyn_cast<ConstantInt>(V)) {
235 // If it's a constant, just convert it to an offset and remove the variable.
236 // If we've been called recursively, the Offset bit width will be greater
237 // than the constant's (the Offset's always as wide as the outermost call),
238 // so we'll zext here and process any extension in the isa<SExtInst> &
239 // isa<ZExtInst> cases below.
240 Offset += Const->getValue().zextOrSelf(Offset.getBitWidth());
241 assert(Scale == 0 && "Constant values don't have a scale");
245 if (const BinaryOperator *BOp = dyn_cast<BinaryOperator>(V)) {
246 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(BOp->getOperand(1))) {
247 // If we've been called recursively, then Offset and Scale will be wider
248 // than the BOp operands. We'll always zext it here as we'll process sign
249 // extensions below (see the isa<SExtInst> / isa<ZExtInst> cases).
250 APInt RHS = RHSC->getValue().zextOrSelf(Offset.getBitWidth());
252 switch (BOp->getOpcode()) {
254 // We don't understand this instruction, so we can't decompose it any
259 case Instruction::Or:
260 // X|C == X+C if all the bits in C are unset in X. Otherwise we can't
262 if (!MaskedValueIsZero(BOp->getOperand(0), RHSC->getValue(), DL, 0, AC,
269 case Instruction::Add:
270 V = GetLinearExpression(BOp->getOperand(0), Scale, Offset, ZExtBits,
271 SExtBits, DL, Depth + 1, AC, DT, NSW, NUW);
274 case Instruction::Sub:
275 V = GetLinearExpression(BOp->getOperand(0), Scale, Offset, ZExtBits,
276 SExtBits, DL, Depth + 1, AC, DT, NSW, NUW);
279 case Instruction::Mul:
280 V = GetLinearExpression(BOp->getOperand(0), Scale, Offset, ZExtBits,
281 SExtBits, DL, Depth + 1, AC, DT, NSW, NUW);
285 case Instruction::Shl:
286 V = GetLinearExpression(BOp->getOperand(0), Scale, Offset, ZExtBits,
287 SExtBits, DL, Depth + 1, AC, DT, NSW, NUW);
288 Offset <<= RHS.getLimitedValue();
289 Scale <<= RHS.getLimitedValue();
290 // the semantics of nsw and nuw for left shifts don't match those of
291 // multiplications, so we won't propagate them.
296 if (isa<OverflowingBinaryOperator>(BOp)) {
297 NUW &= BOp->hasNoUnsignedWrap();
298 NSW &= BOp->hasNoSignedWrap();
304 // Since GEP indices are sign extended anyway, we don't care about the high
305 // bits of a sign or zero extended value - just scales and offsets. The
306 // extensions have to be consistent though.
307 if (isa<SExtInst>(V) || isa<ZExtInst>(V)) {
308 Value *CastOp = cast<CastInst>(V)->getOperand(0);
309 unsigned NewWidth = V->getType()->getPrimitiveSizeInBits();
310 unsigned SmallWidth = CastOp->getType()->getPrimitiveSizeInBits();
311 unsigned OldZExtBits = ZExtBits, OldSExtBits = SExtBits;
312 const Value *Result =
313 GetLinearExpression(CastOp, Scale, Offset, ZExtBits, SExtBits, DL,
314 Depth + 1, AC, DT, NSW, NUW);
316 // zext(zext(%x)) == zext(%x), and similarly for sext; we'll handle this
317 // by just incrementing the number of bits we've extended by.
318 unsigned ExtendedBy = NewWidth - SmallWidth;
320 if (isa<SExtInst>(V) && ZExtBits == 0) {
321 // sext(sext(%x, a), b) == sext(%x, a + b)
324 // We haven't sign-wrapped, so it's valid to decompose sext(%x + c)
325 // into sext(%x) + sext(c). We'll sext the Offset ourselves:
326 unsigned OldWidth = Offset.getBitWidth();
327 Offset = Offset.trunc(SmallWidth).sext(NewWidth).zextOrSelf(OldWidth);
329 // We may have signed-wrapped, so don't decompose sext(%x + c) into
330 // sext(%x) + sext(c)
334 ZExtBits = OldZExtBits;
335 SExtBits = OldSExtBits;
337 SExtBits += ExtendedBy;
339 // sext(zext(%x, a), b) = zext(zext(%x, a), b) = zext(%x, a + b)
342 // We may have unsigned-wrapped, so don't decompose zext(%x + c) into
343 // zext(%x) + zext(c)
347 ZExtBits = OldZExtBits;
348 SExtBits = OldSExtBits;
350 ZExtBits += ExtendedBy;
361 /// To ensure a pointer offset fits in an integer of size PointerSize
362 /// (in bits) when that size is smaller than 64. This is an issue in
363 /// particular for 32b programs with negative indices that rely on two's
364 /// complement wrap-arounds for precise alias information.
365 static int64_t adjustToPointerSize(int64_t Offset, unsigned PointerSize) {
366 assert(PointerSize <= 64 && "Invalid PointerSize!");
367 unsigned ShiftBits = 64 - PointerSize;
368 return (int64_t)((uint64_t)Offset << ShiftBits) >> ShiftBits;
371 /// If V is a symbolic pointer expression, decompose it into a base pointer
372 /// with a constant offset and a number of scaled symbolic offsets.
374 /// The scaled symbolic offsets (represented by pairs of a Value* and a scale
375 /// in the VarIndices vector) are Value*'s that are known to be scaled by the
376 /// specified amount, but which may have other unrepresented high bits. As
377 /// such, the gep cannot necessarily be reconstructed from its decomposed form.
379 /// When DataLayout is around, this function is capable of analyzing everything
380 /// that GetUnderlyingObject can look through. To be able to do that
381 /// GetUnderlyingObject and DecomposeGEPExpression must use the same search
382 /// depth (MaxLookupSearchDepth). When DataLayout not is around, it just looks
383 /// through pointer casts.
384 bool BasicAAResult::DecomposeGEPExpression(const Value *V,
385 DecomposedGEP &Decomposed, const DataLayout &DL, AssumptionCache *AC,
387 // Limit recursion depth to limit compile time in crazy cases.
388 unsigned MaxLookup = MaxLookupSearchDepth;
391 Decomposed.StructOffset = 0;
392 Decomposed.OtherOffset = 0;
393 Decomposed.VarIndices.clear();
395 // See if this is a bitcast or GEP.
396 const Operator *Op = dyn_cast<Operator>(V);
398 // The only non-operator case we can handle are GlobalAliases.
399 if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
400 if (!GA->isInterposable()) {
401 V = GA->getAliasee();
409 if (Op->getOpcode() == Instruction::BitCast ||
410 Op->getOpcode() == Instruction::AddrSpaceCast) {
411 V = Op->getOperand(0);
415 const GEPOperator *GEPOp = dyn_cast<GEPOperator>(Op);
417 if (auto CS = ImmutableCallSite(V))
418 if (const Value *RV = CS.getReturnedArgOperand()) {
423 // If it's not a GEP, hand it off to SimplifyInstruction to see if it
424 // can come up with something. This matches what GetUnderlyingObject does.
425 if (const Instruction *I = dyn_cast<Instruction>(V))
426 // TODO: Get a DominatorTree and AssumptionCache and use them here
427 // (these are both now available in this function, but this should be
428 // updated when GetUnderlyingObject is updated). TLI should be
430 if (const Value *Simplified =
431 SimplifyInstruction(const_cast<Instruction *>(I), DL)) {
440 // Don't attempt to analyze GEPs over unsized objects.
441 if (!GEPOp->getSourceElementType()->isSized()) {
446 unsigned AS = GEPOp->getPointerAddressSpace();
447 // Walk the indices of the GEP, accumulating them into BaseOff/VarIndices.
448 gep_type_iterator GTI = gep_type_begin(GEPOp);
449 unsigned PointerSize = DL.getPointerSizeInBits(AS);
450 // Assume all GEP operands are constants until proven otherwise.
451 bool GepHasConstantOffset = true;
452 for (User::const_op_iterator I = GEPOp->op_begin() + 1, E = GEPOp->op_end();
453 I != E; ++I, ++GTI) {
454 const Value *Index = *I;
455 // Compute the (potentially symbolic) offset in bytes for this index.
456 if (StructType *STy = GTI.getStructTypeOrNull()) {
457 // For a struct, add the member offset.
458 unsigned FieldNo = cast<ConstantInt>(Index)->getZExtValue();
462 Decomposed.StructOffset +=
463 DL.getStructLayout(STy)->getElementOffset(FieldNo);
467 // For an array/pointer, add the element offset, explicitly scaled.
468 if (const ConstantInt *CIdx = dyn_cast<ConstantInt>(Index)) {
471 Decomposed.OtherOffset +=
472 DL.getTypeAllocSize(GTI.getIndexedType()) * CIdx->getSExtValue();
476 GepHasConstantOffset = false;
478 uint64_t Scale = DL.getTypeAllocSize(GTI.getIndexedType());
479 unsigned ZExtBits = 0, SExtBits = 0;
481 // If the integer type is smaller than the pointer size, it is implicitly
482 // sign extended to pointer size.
483 unsigned Width = Index->getType()->getIntegerBitWidth();
484 if (PointerSize > Width)
485 SExtBits += PointerSize - Width;
487 // Use GetLinearExpression to decompose the index into a C1*V+C2 form.
488 APInt IndexScale(Width, 0), IndexOffset(Width, 0);
489 bool NSW = true, NUW = true;
490 Index = GetLinearExpression(Index, IndexScale, IndexOffset, ZExtBits,
491 SExtBits, DL, 0, AC, DT, NSW, NUW);
493 // The GEP index scale ("Scale") scales C1*V+C2, yielding (C1*V+C2)*Scale.
494 // This gives us an aggregate computation of (C1*Scale)*V + C2*Scale.
495 Decomposed.OtherOffset += IndexOffset.getSExtValue() * Scale;
496 Scale *= IndexScale.getSExtValue();
498 // If we already had an occurrence of this index variable, merge this
499 // scale into it. For example, we want to handle:
500 // A[x][x] -> x*16 + x*4 -> x*20
501 // This also ensures that 'x' only appears in the index list once.
502 for (unsigned i = 0, e = Decomposed.VarIndices.size(); i != e; ++i) {
503 if (Decomposed.VarIndices[i].V == Index &&
504 Decomposed.VarIndices[i].ZExtBits == ZExtBits &&
505 Decomposed.VarIndices[i].SExtBits == SExtBits) {
506 Scale += Decomposed.VarIndices[i].Scale;
507 Decomposed.VarIndices.erase(Decomposed.VarIndices.begin() + i);
512 // Make sure that we have a scale that makes sense for this target's
514 Scale = adjustToPointerSize(Scale, PointerSize);
517 VariableGEPIndex Entry = {Index, ZExtBits, SExtBits,
518 static_cast<int64_t>(Scale)};
519 Decomposed.VarIndices.push_back(Entry);
523 // Take care of wrap-arounds
524 if (GepHasConstantOffset) {
525 Decomposed.StructOffset =
526 adjustToPointerSize(Decomposed.StructOffset, PointerSize);
527 Decomposed.OtherOffset =
528 adjustToPointerSize(Decomposed.OtherOffset, PointerSize);
531 // Analyze the base pointer next.
532 V = GEPOp->getOperand(0);
533 } while (--MaxLookup);
535 // If the chain of expressions is too deep, just return early.
537 SearchLimitReached++;
541 /// Returns whether the given pointer value points to memory that is local to
542 /// the function, with global constants being considered local to all
544 bool BasicAAResult::pointsToConstantMemory(const MemoryLocation &Loc,
546 assert(Visited.empty() && "Visited must be cleared after use!");
548 unsigned MaxLookup = 8;
549 SmallVector<const Value *, 16> Worklist;
550 Worklist.push_back(Loc.Ptr);
552 const Value *V = GetUnderlyingObject(Worklist.pop_back_val(), DL);
553 if (!Visited.insert(V).second) {
555 return AAResultBase::pointsToConstantMemory(Loc, OrLocal);
558 // An alloca instruction defines local memory.
559 if (OrLocal && isa<AllocaInst>(V))
562 // A global constant counts as local memory for our purposes.
563 if (const GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) {
564 // Note: this doesn't require GV to be "ODR" because it isn't legal for a
565 // global to be marked constant in some modules and non-constant in
566 // others. GV may even be a declaration, not a definition.
567 if (!GV->isConstant()) {
569 return AAResultBase::pointsToConstantMemory(Loc, OrLocal);
574 // If both select values point to local memory, then so does the select.
575 if (const SelectInst *SI = dyn_cast<SelectInst>(V)) {
576 Worklist.push_back(SI->getTrueValue());
577 Worklist.push_back(SI->getFalseValue());
581 // If all values incoming to a phi node point to local memory, then so does
583 if (const PHINode *PN = dyn_cast<PHINode>(V)) {
584 // Don't bother inspecting phi nodes with many operands.
585 if (PN->getNumIncomingValues() > MaxLookup) {
587 return AAResultBase::pointsToConstantMemory(Loc, OrLocal);
589 for (Value *IncValue : PN->incoming_values())
590 Worklist.push_back(IncValue);
594 // Otherwise be conservative.
596 return AAResultBase::pointsToConstantMemory(Loc, OrLocal);
597 } while (!Worklist.empty() && --MaxLookup);
600 return Worklist.empty();
603 /// Returns the behavior when calling the given call site.
604 FunctionModRefBehavior BasicAAResult::getModRefBehavior(ImmutableCallSite CS) {
605 if (CS.doesNotAccessMemory())
606 // Can't do better than this.
607 return FMRB_DoesNotAccessMemory;
609 FunctionModRefBehavior Min = FMRB_UnknownModRefBehavior;
611 // If the callsite knows it only reads memory, don't return worse
613 if (CS.onlyReadsMemory())
614 Min = FMRB_OnlyReadsMemory;
615 else if (CS.doesNotReadMemory())
616 Min = FMRB_DoesNotReadMemory;
618 if (CS.onlyAccessesArgMemory())
619 Min = FunctionModRefBehavior(Min & FMRB_OnlyAccessesArgumentPointees);
620 else if (CS.onlyAccessesInaccessibleMemory())
621 Min = FunctionModRefBehavior(Min & FMRB_OnlyAccessesInaccessibleMem);
622 else if (CS.onlyAccessesInaccessibleMemOrArgMem())
623 Min = FunctionModRefBehavior(Min & FMRB_OnlyAccessesInaccessibleOrArgMem);
625 // If CS has operand bundles then aliasing attributes from the function it
626 // calls do not directly apply to the CallSite. This can be made more
627 // precise in the future.
628 if (!CS.hasOperandBundles())
629 if (const Function *F = CS.getCalledFunction())
631 FunctionModRefBehavior(Min & getBestAAResults().getModRefBehavior(F));
636 /// Returns the behavior when calling the given function. For use when the call
637 /// site is not known.
638 FunctionModRefBehavior BasicAAResult::getModRefBehavior(const Function *F) {
639 // If the function declares it doesn't access memory, we can't do better.
640 if (F->doesNotAccessMemory())
641 return FMRB_DoesNotAccessMemory;
643 FunctionModRefBehavior Min = FMRB_UnknownModRefBehavior;
645 // If the function declares it only reads memory, go with that.
646 if (F->onlyReadsMemory())
647 Min = FMRB_OnlyReadsMemory;
648 else if (F->doesNotReadMemory())
649 Min = FMRB_DoesNotReadMemory;
651 if (F->onlyAccessesArgMemory())
652 Min = FunctionModRefBehavior(Min & FMRB_OnlyAccessesArgumentPointees);
653 else if (F->onlyAccessesInaccessibleMemory())
654 Min = FunctionModRefBehavior(Min & FMRB_OnlyAccessesInaccessibleMem);
655 else if (F->onlyAccessesInaccessibleMemOrArgMem())
656 Min = FunctionModRefBehavior(Min & FMRB_OnlyAccessesInaccessibleOrArgMem);
661 /// Returns true if this is a writeonly (i.e Mod only) parameter.
662 static bool isWriteOnlyParam(ImmutableCallSite CS, unsigned ArgIdx,
663 const TargetLibraryInfo &TLI) {
664 if (CS.paramHasAttr(ArgIdx, Attribute::WriteOnly))
667 // We can bound the aliasing properties of memset_pattern16 just as we can
668 // for memcpy/memset. This is particularly important because the
669 // LoopIdiomRecognizer likes to turn loops into calls to memset_pattern16
670 // whenever possible.
671 // FIXME Consider handling this in InferFunctionAttr.cpp together with other
674 if (CS.getCalledFunction() && TLI.getLibFunc(*CS.getCalledFunction(), F) &&
675 F == LibFunc_memset_pattern16 && TLI.has(F))
679 // TODO: memset_pattern4, memset_pattern8
680 // TODO: _chk variants
681 // TODO: strcmp, strcpy
686 ModRefInfo BasicAAResult::getArgModRefInfo(ImmutableCallSite CS,
688 // Checking for known builtin intrinsics and target library functions.
689 if (isWriteOnlyParam(CS, ArgIdx, TLI))
690 return ModRefInfo::Mod;
692 if (CS.paramHasAttr(ArgIdx, Attribute::ReadOnly))
693 return ModRefInfo::Ref;
695 if (CS.paramHasAttr(ArgIdx, Attribute::ReadNone))
696 return ModRefInfo::NoModRef;
698 return AAResultBase::getArgModRefInfo(CS, ArgIdx);
701 static bool isIntrinsicCall(ImmutableCallSite CS, Intrinsic::ID IID) {
702 const IntrinsicInst *II = dyn_cast<IntrinsicInst>(CS.getInstruction());
703 return II && II->getIntrinsicID() == IID;
707 static const Function *getParent(const Value *V) {
708 if (const Instruction *inst = dyn_cast<Instruction>(V)) {
709 if (!inst->getParent())
711 return inst->getParent()->getParent();
714 if (const Argument *arg = dyn_cast<Argument>(V))
715 return arg->getParent();
720 static bool notDifferentParent(const Value *O1, const Value *O2) {
722 const Function *F1 = getParent(O1);
723 const Function *F2 = getParent(O2);
725 return !F1 || !F2 || F1 == F2;
729 AliasResult BasicAAResult::alias(const MemoryLocation &LocA,
730 const MemoryLocation &LocB) {
731 assert(notDifferentParent(LocA.Ptr, LocB.Ptr) &&
732 "BasicAliasAnalysis doesn't support interprocedural queries.");
734 // If we have a directly cached entry for these locations, we have recursed
735 // through this once, so just return the cached results. Notably, when this
736 // happens, we don't clear the cache.
737 auto CacheIt = AliasCache.find(LocPair(LocA, LocB));
738 if (CacheIt != AliasCache.end())
739 return CacheIt->second;
741 AliasResult Alias = aliasCheck(LocA.Ptr, LocA.Size, LocA.AATags, LocB.Ptr,
742 LocB.Size, LocB.AATags);
743 // AliasCache rarely has more than 1 or 2 elements, always use
744 // shrink_and_clear so it quickly returns to the inline capacity of the
745 // SmallDenseMap if it ever grows larger.
746 // FIXME: This should really be shrink_to_inline_capacity_and_clear().
747 AliasCache.shrink_and_clear();
748 VisitedPhiBBs.clear();
752 /// Checks to see if the specified callsite can clobber the specified memory
755 /// Since we only look at local properties of this function, we really can't
756 /// say much about this query. We do, however, use simple "address taken"
757 /// analysis on local objects.
758 ModRefInfo BasicAAResult::getModRefInfo(ImmutableCallSite CS,
759 const MemoryLocation &Loc) {
760 assert(notDifferentParent(CS.getInstruction(), Loc.Ptr) &&
761 "AliasAnalysis query involving multiple functions!");
763 const Value *Object = GetUnderlyingObject(Loc.Ptr, DL);
765 // If this is a tail call and Loc.Ptr points to a stack location, we know that
766 // the tail call cannot access or modify the local stack.
767 // We cannot exclude byval arguments here; these belong to the caller of
768 // the current function not to the current function, and a tail callee
769 // may reference them.
770 if (isa<AllocaInst>(Object))
771 if (const CallInst *CI = dyn_cast<CallInst>(CS.getInstruction()))
772 if (CI->isTailCall())
773 return ModRefInfo::NoModRef;
775 // If the pointer is to a locally allocated object that does not escape,
776 // then the call can not mod/ref the pointer unless the call takes the pointer
777 // as an argument, and itself doesn't capture it.
778 if (!isa<Constant>(Object) && CS.getInstruction() != Object &&
779 isNonEscapingLocalObject(Object)) {
781 // Optimistically assume that call doesn't touch Object and check this
782 // assumption in the following loop.
783 ModRefInfo Result = ModRefInfo::NoModRef;
784 bool IsMustAlias = true;
786 unsigned OperandNo = 0;
787 for (auto CI = CS.data_operands_begin(), CE = CS.data_operands_end();
788 CI != CE; ++CI, ++OperandNo) {
789 // Only look at the no-capture or byval pointer arguments. If this
790 // pointer were passed to arguments that were neither of these, then it
791 // couldn't be no-capture.
792 if (!(*CI)->getType()->isPointerTy() ||
793 (!CS.doesNotCapture(OperandNo) &&
794 OperandNo < CS.getNumArgOperands() && !CS.isByValArgument(OperandNo)))
797 // Call doesn't access memory through this operand, so we don't care
798 // if it aliases with Object.
799 if (CS.doesNotAccessMemory(OperandNo))
802 // If this is a no-capture pointer argument, see if we can tell that it
803 // is impossible to alias the pointer we're checking.
805 getBestAAResults().alias(MemoryLocation(*CI), MemoryLocation(Object));
808 // Operand doesnt alias 'Object', continue looking for other aliases
811 // Operand aliases 'Object', but call doesn't modify it. Strengthen
812 // initial assumption and keep looking in case if there are more aliases.
813 if (CS.onlyReadsMemory(OperandNo)) {
814 Result = setRef(Result);
817 // Operand aliases 'Object' but call only writes into it.
818 if (CS.doesNotReadMemory(OperandNo)) {
819 Result = setMod(Result);
822 // This operand aliases 'Object' and call reads and writes into it.
823 // Setting ModRef will not yield an early return below, MustAlias is not
825 Result = ModRefInfo::ModRef;
829 // No operand aliases, reset Must bit. Add below if at least one aliases
830 // and all aliases found are MustAlias.
831 if (isNoModRef(Result))
834 // Early return if we improved mod ref information
835 if (!isModAndRefSet(Result))
836 return IsMustAlias ? setMust(Result) : clearMust(Result);
839 // If the CallSite is to malloc or calloc, we can assume that it doesn't
840 // modify any IR visible value. This is only valid because we assume these
841 // routines do not read values visible in the IR. TODO: Consider special
842 // casing realloc and strdup routines which access only their arguments as
843 // well. Or alternatively, replace all of this with inaccessiblememonly once
844 // that's implemented fully.
845 auto *Inst = CS.getInstruction();
846 if (isMallocOrCallocLikeFn(Inst, &TLI)) {
847 // Be conservative if the accessed pointer may alias the allocation -
848 // fallback to the generic handling below.
849 if (getBestAAResults().alias(MemoryLocation(Inst), Loc) == NoAlias)
850 return ModRefInfo::NoModRef;
853 // The semantics of memcpy intrinsics forbid overlap between their respective
854 // operands, i.e., source and destination of any given memcpy must no-alias.
855 // If Loc must-aliases either one of these two locations, then it necessarily
856 // no-aliases the other.
857 if (auto *Inst = dyn_cast<MemCpyInst>(CS.getInstruction())) {
858 AliasResult SrcAA, DestAA;
860 if ((SrcAA = getBestAAResults().alias(MemoryLocation::getForSource(Inst),
862 // Loc is exactly the memcpy source thus disjoint from memcpy dest.
863 return ModRefInfo::Ref;
864 if ((DestAA = getBestAAResults().alias(MemoryLocation::getForDest(Inst),
866 // The converse case.
867 return ModRefInfo::Mod;
869 // It's also possible for Loc to alias both src and dest, or neither.
870 ModRefInfo rv = ModRefInfo::NoModRef;
871 if (SrcAA != NoAlias)
873 if (DestAA != NoAlias)
878 // While the assume intrinsic is marked as arbitrarily writing so that
879 // proper control dependencies will be maintained, it never aliases any
880 // particular memory location.
881 if (isIntrinsicCall(CS, Intrinsic::assume))
882 return ModRefInfo::NoModRef;
884 // Like assumes, guard intrinsics are also marked as arbitrarily writing so
885 // that proper control dependencies are maintained but they never mods any
886 // particular memory location.
888 // *Unlike* assumes, guard intrinsics are modeled as reading memory since the
889 // heap state at the point the guard is issued needs to be consistent in case
890 // the guard invokes the "deopt" continuation.
891 if (isIntrinsicCall(CS, Intrinsic::experimental_guard))
892 return ModRefInfo::Ref;
894 // Like assumes, invariant.start intrinsics were also marked as arbitrarily
895 // writing so that proper control dependencies are maintained but they never
896 // mod any particular memory location visible to the IR.
897 // *Unlike* assumes (which are now modeled as NoModRef), invariant.start
898 // intrinsic is now modeled as reading memory. This prevents hoisting the
899 // invariant.start intrinsic over stores. Consider:
902 // invariant_start(ptr)
906 // This cannot be transformed to:
909 // invariant_start(ptr)
914 // The transformation will cause the second store to be ignored (based on
915 // rules of invariant.start) and print 40, while the first program always
917 if (isIntrinsicCall(CS, Intrinsic::invariant_start))
918 return ModRefInfo::Ref;
920 // The AAResultBase base class has some smarts, lets use them.
921 return AAResultBase::getModRefInfo(CS, Loc);
924 ModRefInfo BasicAAResult::getModRefInfo(ImmutableCallSite CS1,
925 ImmutableCallSite CS2) {
926 // While the assume intrinsic is marked as arbitrarily writing so that
927 // proper control dependencies will be maintained, it never aliases any
928 // particular memory location.
929 if (isIntrinsicCall(CS1, Intrinsic::assume) ||
930 isIntrinsicCall(CS2, Intrinsic::assume))
931 return ModRefInfo::NoModRef;
933 // Like assumes, guard intrinsics are also marked as arbitrarily writing so
934 // that proper control dependencies are maintained but they never mod any
935 // particular memory location.
937 // *Unlike* assumes, guard intrinsics are modeled as reading memory since the
938 // heap state at the point the guard is issued needs to be consistent in case
939 // the guard invokes the "deopt" continuation.
941 // NB! This function is *not* commutative, so we specical case two
942 // possibilities for guard intrinsics.
944 if (isIntrinsicCall(CS1, Intrinsic::experimental_guard))
945 return isModSet(createModRefInfo(getModRefBehavior(CS2)))
947 : ModRefInfo::NoModRef;
949 if (isIntrinsicCall(CS2, Intrinsic::experimental_guard))
950 return isModSet(createModRefInfo(getModRefBehavior(CS1)))
952 : ModRefInfo::NoModRef;
954 // The AAResultBase base class has some smarts, lets use them.
955 return AAResultBase::getModRefInfo(CS1, CS2);
958 /// Provide ad-hoc rules to disambiguate accesses through two GEP operators,
959 /// both having the exact same pointer operand.
960 static AliasResult aliasSameBasePointerGEPs(const GEPOperator *GEP1,
962 const GEPOperator *GEP2,
964 const DataLayout &DL) {
965 assert(GEP1->getPointerOperand()->stripPointerCastsAndBarriers() ==
966 GEP2->getPointerOperand()->stripPointerCastsAndBarriers() &&
967 GEP1->getPointerOperandType() == GEP2->getPointerOperandType() &&
968 "Expected GEPs with the same pointer operand");
970 // Try to determine whether GEP1 and GEP2 index through arrays, into structs,
971 // such that the struct field accesses provably cannot alias.
972 // We also need at least two indices (the pointer, and the struct field).
973 if (GEP1->getNumIndices() != GEP2->getNumIndices() ||
974 GEP1->getNumIndices() < 2)
977 // If we don't know the size of the accesses through both GEPs, we can't
978 // determine whether the struct fields accessed can't alias.
979 if (V1Size == MemoryLocation::UnknownSize ||
980 V2Size == MemoryLocation::UnknownSize)
984 dyn_cast<ConstantInt>(GEP1->getOperand(GEP1->getNumOperands() - 1));
986 dyn_cast<ConstantInt>(GEP2->getOperand(GEP2->getNumOperands() - 1));
988 // If the last (struct) indices are constants and are equal, the other indices
989 // might be also be dynamically equal, so the GEPs can alias.
990 if (C1 && C2 && C1->getSExtValue() == C2->getSExtValue())
993 // Find the last-indexed type of the GEP, i.e., the type you'd get if
994 // you stripped the last index.
995 // On the way, look at each indexed type. If there's something other
996 // than an array, different indices can lead to different final types.
997 SmallVector<Value *, 8> IntermediateIndices;
999 // Insert the first index; we don't need to check the type indexed
1000 // through it as it only drops the pointer indirection.
1001 assert(GEP1->getNumIndices() > 1 && "Not enough GEP indices to examine");
1002 IntermediateIndices.push_back(GEP1->getOperand(1));
1004 // Insert all the remaining indices but the last one.
1005 // Also, check that they all index through arrays.
1006 for (unsigned i = 1, e = GEP1->getNumIndices() - 1; i != e; ++i) {
1007 if (!isa<ArrayType>(GetElementPtrInst::getIndexedType(
1008 GEP1->getSourceElementType(), IntermediateIndices)))
1010 IntermediateIndices.push_back(GEP1->getOperand(i + 1));
1013 auto *Ty = GetElementPtrInst::getIndexedType(
1014 GEP1->getSourceElementType(), IntermediateIndices);
1015 StructType *LastIndexedStruct = dyn_cast<StructType>(Ty);
1017 if (isa<SequentialType>(Ty)) {
1019 // - both GEPs begin indexing from the exact same pointer;
1020 // - the last indices in both GEPs are constants, indexing into a sequential
1021 // type (array or pointer);
1022 // - both GEPs only index through arrays prior to that.
1024 // Because array indices greater than the number of elements are valid in
1025 // GEPs, unless we know the intermediate indices are identical between
1026 // GEP1 and GEP2 we cannot guarantee that the last indexed arrays don't
1027 // partially overlap. We also need to check that the loaded size matches
1028 // the element size, otherwise we could still have overlap.
1029 const uint64_t ElementSize =
1030 DL.getTypeStoreSize(cast<SequentialType>(Ty)->getElementType());
1031 if (V1Size != ElementSize || V2Size != ElementSize)
1034 for (unsigned i = 0, e = GEP1->getNumIndices() - 1; i != e; ++i)
1035 if (GEP1->getOperand(i + 1) != GEP2->getOperand(i + 1))
1038 // Now we know that the array/pointer that GEP1 indexes into and that
1039 // that GEP2 indexes into must either precisely overlap or be disjoint.
1040 // Because they cannot partially overlap and because fields in an array
1041 // cannot overlap, if we can prove the final indices are different between
1042 // GEP1 and GEP2, we can conclude GEP1 and GEP2 don't alias.
1044 // If the last indices are constants, we've already checked they don't
1045 // equal each other so we can exit early.
1049 Value *GEP1LastIdx = GEP1->getOperand(GEP1->getNumOperands() - 1);
1050 Value *GEP2LastIdx = GEP2->getOperand(GEP2->getNumOperands() - 1);
1051 if (isa<PHINode>(GEP1LastIdx) || isa<PHINode>(GEP2LastIdx)) {
1052 // If one of the indices is a PHI node, be safe and only use
1053 // computeKnownBits so we don't make any assumptions about the
1054 // relationships between the two indices. This is important if we're
1055 // asking about values from different loop iterations. See PR32314.
1056 // TODO: We may be able to change the check so we only do this when
1057 // we definitely looked through a PHINode.
1058 if (GEP1LastIdx != GEP2LastIdx &&
1059 GEP1LastIdx->getType() == GEP2LastIdx->getType()) {
1060 KnownBits Known1 = computeKnownBits(GEP1LastIdx, DL);
1061 KnownBits Known2 = computeKnownBits(GEP2LastIdx, DL);
1062 if (Known1.Zero.intersects(Known2.One) ||
1063 Known1.One.intersects(Known2.Zero))
1066 } else if (isKnownNonEqual(GEP1LastIdx, GEP2LastIdx, DL))
1070 } else if (!LastIndexedStruct || !C1 || !C2) {
1075 // - both GEPs begin indexing from the exact same pointer;
1076 // - the last indices in both GEPs are constants, indexing into a struct;
1077 // - said indices are different, hence, the pointed-to fields are different;
1078 // - both GEPs only index through arrays prior to that.
1080 // This lets us determine that the struct that GEP1 indexes into and the
1081 // struct that GEP2 indexes into must either precisely overlap or be
1082 // completely disjoint. Because they cannot partially overlap, indexing into
1083 // different non-overlapping fields of the struct will never alias.
1085 // Therefore, the only remaining thing needed to show that both GEPs can't
1086 // alias is that the fields are not overlapping.
1087 const StructLayout *SL = DL.getStructLayout(LastIndexedStruct);
1088 const uint64_t StructSize = SL->getSizeInBytes();
1089 const uint64_t V1Off = SL->getElementOffset(C1->getZExtValue());
1090 const uint64_t V2Off = SL->getElementOffset(C2->getZExtValue());
1092 auto EltsDontOverlap = [StructSize](uint64_t V1Off, uint64_t V1Size,
1093 uint64_t V2Off, uint64_t V2Size) {
1094 return V1Off < V2Off && V1Off + V1Size <= V2Off &&
1095 ((V2Off + V2Size <= StructSize) ||
1096 (V2Off + V2Size - StructSize <= V1Off));
1099 if (EltsDontOverlap(V1Off, V1Size, V2Off, V2Size) ||
1100 EltsDontOverlap(V2Off, V2Size, V1Off, V1Size))
1106 // If a we have (a) a GEP and (b) a pointer based on an alloca, and the
1107 // beginning of the object the GEP points would have a negative offset with
1108 // repsect to the alloca, that means the GEP can not alias pointer (b).
1109 // Note that the pointer based on the alloca may not be a GEP. For
1110 // example, it may be the alloca itself.
1111 // The same applies if (b) is based on a GlobalVariable. Note that just being
1112 // based on isIdentifiedObject() is not enough - we need an identified object
1113 // that does not permit access to negative offsets. For example, a negative
1114 // offset from a noalias argument or call can be inbounds w.r.t the actual
1115 // underlying object.
1117 // For example, consider:
1119 // struct { int f0, int f1, ...} foo;
1121 // foo* random = bar(alloca);
1122 // int *f0 = &alloca.f0
1123 // int *f1 = &random->f1;
1125 // Which is lowered, approximately, to:
1127 // %alloca = alloca %struct.foo
1128 // %random = call %struct.foo* @random(%struct.foo* %alloca)
1129 // %f0 = getelementptr inbounds %struct, %struct.foo* %alloca, i32 0, i32 0
1130 // %f1 = getelementptr inbounds %struct, %struct.foo* %random, i32 0, i32 1
1132 // Assume %f1 and %f0 alias. Then %f1 would point into the object allocated
1133 // by %alloca. Since the %f1 GEP is inbounds, that means %random must also
1134 // point into the same object. But since %f0 points to the beginning of %alloca,
1135 // the highest %f1 can be is (%alloca + 3). This means %random can not be higher
1136 // than (%alloca - 1), and so is not inbounds, a contradiction.
1137 bool BasicAAResult::isGEPBaseAtNegativeOffset(const GEPOperator *GEPOp,
1138 const DecomposedGEP &DecompGEP, const DecomposedGEP &DecompObject,
1139 uint64_t ObjectAccessSize) {
1140 // If the object access size is unknown, or the GEP isn't inbounds, bail.
1141 if (ObjectAccessSize == MemoryLocation::UnknownSize || !GEPOp->isInBounds())
1144 // We need the object to be an alloca or a globalvariable, and want to know
1145 // the offset of the pointer from the object precisely, so no variable
1146 // indices are allowed.
1147 if (!(isa<AllocaInst>(DecompObject.Base) ||
1148 isa<GlobalVariable>(DecompObject.Base)) ||
1149 !DecompObject.VarIndices.empty())
1152 int64_t ObjectBaseOffset = DecompObject.StructOffset +
1153 DecompObject.OtherOffset;
1155 // If the GEP has no variable indices, we know the precise offset
1156 // from the base, then use it. If the GEP has variable indices, we're in
1157 // a bit more trouble: we can't count on the constant offsets that come
1158 // from non-struct sources, since these can be "rewound" by a negative
1159 // variable offset. So use only offsets that came from structs.
1160 int64_t GEPBaseOffset = DecompGEP.StructOffset;
1161 if (DecompGEP.VarIndices.empty())
1162 GEPBaseOffset += DecompGEP.OtherOffset;
1164 return (GEPBaseOffset >= ObjectBaseOffset + (int64_t)ObjectAccessSize);
1167 /// Provides a bunch of ad-hoc rules to disambiguate a GEP instruction against
1168 /// another pointer.
1170 /// We know that V1 is a GEP, but we don't know anything about V2.
1171 /// UnderlyingV1 is GetUnderlyingObject(GEP1, DL), UnderlyingV2 is the same for
1173 AliasResult BasicAAResult::aliasGEP(const GEPOperator *GEP1, uint64_t V1Size,
1174 const AAMDNodes &V1AAInfo, const Value *V2,
1175 uint64_t V2Size, const AAMDNodes &V2AAInfo,
1176 const Value *UnderlyingV1,
1177 const Value *UnderlyingV2) {
1178 DecomposedGEP DecompGEP1, DecompGEP2;
1179 bool GEP1MaxLookupReached =
1180 DecomposeGEPExpression(GEP1, DecompGEP1, DL, &AC, DT);
1181 bool GEP2MaxLookupReached =
1182 DecomposeGEPExpression(V2, DecompGEP2, DL, &AC, DT);
1184 int64_t GEP1BaseOffset = DecompGEP1.StructOffset + DecompGEP1.OtherOffset;
1185 int64_t GEP2BaseOffset = DecompGEP2.StructOffset + DecompGEP2.OtherOffset;
1187 assert(DecompGEP1.Base == UnderlyingV1 && DecompGEP2.Base == UnderlyingV2 &&
1188 "DecomposeGEPExpression returned a result different from "
1189 "GetUnderlyingObject");
1191 // If the GEP's offset relative to its base is such that the base would
1192 // fall below the start of the object underlying V2, then the GEP and V2
1194 if (!GEP1MaxLookupReached && !GEP2MaxLookupReached &&
1195 isGEPBaseAtNegativeOffset(GEP1, DecompGEP1, DecompGEP2, V2Size))
1197 // If we have two gep instructions with must-alias or not-alias'ing base
1198 // pointers, figure out if the indexes to the GEP tell us anything about the
1200 if (const GEPOperator *GEP2 = dyn_cast<GEPOperator>(V2)) {
1201 // Check for the GEP base being at a negative offset, this time in the other
1203 if (!GEP1MaxLookupReached && !GEP2MaxLookupReached &&
1204 isGEPBaseAtNegativeOffset(GEP2, DecompGEP2, DecompGEP1, V1Size))
1206 // Do the base pointers alias?
1207 AliasResult BaseAlias =
1208 aliasCheck(UnderlyingV1, MemoryLocation::UnknownSize, AAMDNodes(),
1209 UnderlyingV2, MemoryLocation::UnknownSize, AAMDNodes());
1211 // Check for geps of non-aliasing underlying pointers where the offsets are
1213 if ((BaseAlias == MayAlias) && V1Size == V2Size) {
1214 // Do the base pointers alias assuming type and size.
1215 AliasResult PreciseBaseAlias = aliasCheck(UnderlyingV1, V1Size, V1AAInfo,
1216 UnderlyingV2, V2Size, V2AAInfo);
1217 if (PreciseBaseAlias == NoAlias) {
1218 // See if the computed offset from the common pointer tells us about the
1219 // relation of the resulting pointer.
1220 // If the max search depth is reached the result is undefined
1221 if (GEP2MaxLookupReached || GEP1MaxLookupReached)
1225 if (GEP1BaseOffset == GEP2BaseOffset &&
1226 DecompGEP1.VarIndices == DecompGEP2.VarIndices)
1231 // If we get a No or May, then return it immediately, no amount of analysis
1232 // will improve this situation.
1233 if (BaseAlias != MustAlias) {
1234 assert(BaseAlias == NoAlias || BaseAlias == MayAlias);
1238 // Otherwise, we have a MustAlias. Since the base pointers alias each other
1239 // exactly, see if the computed offset from the common pointer tells us
1240 // about the relation of the resulting pointer.
1241 // If we know the two GEPs are based off of the exact same pointer (and not
1242 // just the same underlying object), see if that tells us anything about
1243 // the resulting pointers.
1244 if (GEP1->getPointerOperand()->stripPointerCastsAndBarriers() ==
1245 GEP2->getPointerOperand()->stripPointerCastsAndBarriers() &&
1246 GEP1->getPointerOperandType() == GEP2->getPointerOperandType()) {
1247 AliasResult R = aliasSameBasePointerGEPs(GEP1, V1Size, GEP2, V2Size, DL);
1248 // If we couldn't find anything interesting, don't abandon just yet.
1253 // If the max search depth is reached, the result is undefined
1254 if (GEP2MaxLookupReached || GEP1MaxLookupReached)
1257 // Subtract the GEP2 pointer from the GEP1 pointer to find out their
1258 // symbolic difference.
1259 GEP1BaseOffset -= GEP2BaseOffset;
1260 GetIndexDifference(DecompGEP1.VarIndices, DecompGEP2.VarIndices);
1263 // Check to see if these two pointers are related by the getelementptr
1264 // instruction. If one pointer is a GEP with a non-zero index of the other
1265 // pointer, we know they cannot alias.
1267 // If both accesses are unknown size, we can't do anything useful here.
1268 if (V1Size == MemoryLocation::UnknownSize &&
1269 V2Size == MemoryLocation::UnknownSize)
1272 AliasResult R = aliasCheck(UnderlyingV1, MemoryLocation::UnknownSize,
1273 AAMDNodes(), V2, MemoryLocation::UnknownSize,
1274 V2AAInfo, nullptr, UnderlyingV2);
1275 if (R != MustAlias) {
1276 // If V2 may alias GEP base pointer, conservatively returns MayAlias.
1277 // If V2 is known not to alias GEP base pointer, then the two values
1278 // cannot alias per GEP semantics: "Any memory access must be done through
1279 // a pointer value associated with an address range of the memory access,
1280 // otherwise the behavior is undefined.".
1281 assert(R == NoAlias || R == MayAlias);
1285 // If the max search depth is reached the result is undefined
1286 if (GEP1MaxLookupReached)
1290 // In the two GEP Case, if there is no difference in the offsets of the
1291 // computed pointers, the resultant pointers are a must alias. This
1292 // happens when we have two lexically identical GEP's (for example).
1294 // In the other case, if we have getelementptr <ptr>, 0, 0, 0, 0, ... and V2
1295 // must aliases the GEP, the end result is a must alias also.
1296 if (GEP1BaseOffset == 0 && DecompGEP1.VarIndices.empty())
1299 // If there is a constant difference between the pointers, but the difference
1300 // is less than the size of the associated memory object, then we know
1301 // that the objects are partially overlapping. If the difference is
1302 // greater, we know they do not overlap.
1303 if (GEP1BaseOffset != 0 && DecompGEP1.VarIndices.empty()) {
1304 if (GEP1BaseOffset >= 0) {
1305 if (V2Size != MemoryLocation::UnknownSize) {
1306 if ((uint64_t)GEP1BaseOffset < V2Size)
1307 return PartialAlias;
1311 // We have the situation where:
1314 // ---------------->|
1315 // |-->V1Size |-------> V2Size
1317 // We need to know that V2Size is not unknown, otherwise we might have
1318 // stripped a gep with negative index ('gep <ptr>, -1, ...).
1319 if (V1Size != MemoryLocation::UnknownSize &&
1320 V2Size != MemoryLocation::UnknownSize) {
1321 if (-(uint64_t)GEP1BaseOffset < V1Size)
1322 return PartialAlias;
1328 if (!DecompGEP1.VarIndices.empty()) {
1329 uint64_t Modulo = 0;
1330 bool AllPositive = true;
1331 for (unsigned i = 0, e = DecompGEP1.VarIndices.size(); i != e; ++i) {
1333 // Try to distinguish something like &A[i][1] against &A[42][0].
1334 // Grab the least significant bit set in any of the scales. We
1335 // don't need std::abs here (even if the scale's negative) as we'll
1336 // be ^'ing Modulo with itself later.
1337 Modulo |= (uint64_t)DecompGEP1.VarIndices[i].Scale;
1340 // If the Value could change between cycles, then any reasoning about
1341 // the Value this cycle may not hold in the next cycle. We'll just
1342 // give up if we can't determine conditions that hold for every cycle:
1343 const Value *V = DecompGEP1.VarIndices[i].V;
1345 KnownBits Known = computeKnownBits(V, DL, 0, &AC, nullptr, DT);
1346 bool SignKnownZero = Known.isNonNegative();
1347 bool SignKnownOne = Known.isNegative();
1349 // Zero-extension widens the variable, and so forces the sign
1351 bool IsZExt = DecompGEP1.VarIndices[i].ZExtBits > 0 || isa<ZExtInst>(V);
1352 SignKnownZero |= IsZExt;
1353 SignKnownOne &= !IsZExt;
1355 // If the variable begins with a zero then we know it's
1356 // positive, regardless of whether the value is signed or
1358 int64_t Scale = DecompGEP1.VarIndices[i].Scale;
1360 (SignKnownZero && Scale >= 0) || (SignKnownOne && Scale < 0);
1364 Modulo = Modulo ^ (Modulo & (Modulo - 1));
1366 // We can compute the difference between the two addresses
1367 // mod Modulo. Check whether that difference guarantees that the
1368 // two locations do not alias.
1369 uint64_t ModOffset = (uint64_t)GEP1BaseOffset & (Modulo - 1);
1370 if (V1Size != MemoryLocation::UnknownSize &&
1371 V2Size != MemoryLocation::UnknownSize && ModOffset >= V2Size &&
1372 V1Size <= Modulo - ModOffset)
1375 // If we know all the variables are positive, then GEP1 >= GEP1BasePtr.
1376 // If GEP1BasePtr > V2 (GEP1BaseOffset > 0) then we know the pointers
1377 // don't alias if V2Size can fit in the gap between V2 and GEP1BasePtr.
1378 if (AllPositive && GEP1BaseOffset > 0 && V2Size <= (uint64_t)GEP1BaseOffset)
1381 if (constantOffsetHeuristic(DecompGEP1.VarIndices, V1Size, V2Size,
1382 GEP1BaseOffset, &AC, DT))
1386 // Statically, we can see that the base objects are the same, but the
1387 // pointers have dynamic offsets which we can't resolve. And none of our
1388 // little tricks above worked.
1392 static AliasResult MergeAliasResults(AliasResult A, AliasResult B) {
1393 // If the results agree, take it.
1396 // A mix of PartialAlias and MustAlias is PartialAlias.
1397 if ((A == PartialAlias && B == MustAlias) ||
1398 (B == PartialAlias && A == MustAlias))
1399 return PartialAlias;
1400 // Otherwise, we don't know anything.
1404 /// Provides a bunch of ad-hoc rules to disambiguate a Select instruction
1405 /// against another.
1406 AliasResult BasicAAResult::aliasSelect(const SelectInst *SI, uint64_t SISize,
1407 const AAMDNodes &SIAAInfo,
1408 const Value *V2, uint64_t V2Size,
1409 const AAMDNodes &V2AAInfo,
1410 const Value *UnderV2) {
1411 // If the values are Selects with the same condition, we can do a more precise
1412 // check: just check for aliases between the values on corresponding arms.
1413 if (const SelectInst *SI2 = dyn_cast<SelectInst>(V2))
1414 if (SI->getCondition() == SI2->getCondition()) {
1415 AliasResult Alias = aliasCheck(SI->getTrueValue(), SISize, SIAAInfo,
1416 SI2->getTrueValue(), V2Size, V2AAInfo);
1417 if (Alias == MayAlias)
1419 AliasResult ThisAlias =
1420 aliasCheck(SI->getFalseValue(), SISize, SIAAInfo,
1421 SI2->getFalseValue(), V2Size, V2AAInfo);
1422 return MergeAliasResults(ThisAlias, Alias);
1425 // If both arms of the Select node NoAlias or MustAlias V2, then returns
1426 // NoAlias / MustAlias. Otherwise, returns MayAlias.
1428 aliasCheck(V2, V2Size, V2AAInfo, SI->getTrueValue(),
1429 SISize, SIAAInfo, UnderV2);
1430 if (Alias == MayAlias)
1433 AliasResult ThisAlias =
1434 aliasCheck(V2, V2Size, V2AAInfo, SI->getFalseValue(), SISize, SIAAInfo,
1436 return MergeAliasResults(ThisAlias, Alias);
1439 /// Provide a bunch of ad-hoc rules to disambiguate a PHI instruction against
1441 AliasResult BasicAAResult::aliasPHI(const PHINode *PN, uint64_t PNSize,
1442 const AAMDNodes &PNAAInfo, const Value *V2,
1443 uint64_t V2Size, const AAMDNodes &V2AAInfo,
1444 const Value *UnderV2) {
1445 // Track phi nodes we have visited. We use this information when we determine
1446 // value equivalence.
1447 VisitedPhiBBs.insert(PN->getParent());
1449 // If the values are PHIs in the same block, we can do a more precise
1450 // as well as efficient check: just check for aliases between the values
1451 // on corresponding edges.
1452 if (const PHINode *PN2 = dyn_cast<PHINode>(V2))
1453 if (PN2->getParent() == PN->getParent()) {
1454 LocPair Locs(MemoryLocation(PN, PNSize, PNAAInfo),
1455 MemoryLocation(V2, V2Size, V2AAInfo));
1457 std::swap(Locs.first, Locs.second);
1458 // Analyse the PHIs' inputs under the assumption that the PHIs are
1460 // If the PHIs are May/MustAlias there must be (recursively) an input
1461 // operand from outside the PHIs' cycle that is MayAlias/MustAlias or
1462 // there must be an operation on the PHIs within the PHIs' value cycle
1463 // that causes a MayAlias.
1464 // Pretend the phis do not alias.
1465 AliasResult Alias = NoAlias;
1466 assert(AliasCache.count(Locs) &&
1467 "There must exist an entry for the phi node");
1468 AliasResult OrigAliasResult = AliasCache[Locs];
1469 AliasCache[Locs] = NoAlias;
1471 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1472 AliasResult ThisAlias =
1473 aliasCheck(PN->getIncomingValue(i), PNSize, PNAAInfo,
1474 PN2->getIncomingValueForBlock(PN->getIncomingBlock(i)),
1476 Alias = MergeAliasResults(ThisAlias, Alias);
1477 if (Alias == MayAlias)
1481 // Reset if speculation failed.
1482 if (Alias != NoAlias)
1483 AliasCache[Locs] = OrigAliasResult;
1488 SmallPtrSet<Value *, 4> UniqueSrc;
1489 SmallVector<Value *, 4> V1Srcs;
1490 bool isRecursive = false;
1491 for (Value *PV1 : PN->incoming_values()) {
1492 if (isa<PHINode>(PV1))
1493 // If any of the source itself is a PHI, return MayAlias conservatively
1494 // to avoid compile time explosion. The worst possible case is if both
1495 // sides are PHI nodes. In which case, this is O(m x n) time where 'm'
1496 // and 'n' are the number of PHI sources.
1499 if (EnableRecPhiAnalysis)
1500 if (GEPOperator *PV1GEP = dyn_cast<GEPOperator>(PV1)) {
1501 // Check whether the incoming value is a GEP that advances the pointer
1502 // result of this PHI node (e.g. in a loop). If this is the case, we
1503 // would recurse and always get a MayAlias. Handle this case specially
1505 if (PV1GEP->getPointerOperand() == PN && PV1GEP->getNumIndices() == 1 &&
1506 isa<ConstantInt>(PV1GEP->idx_begin())) {
1512 if (UniqueSrc.insert(PV1).second)
1513 V1Srcs.push_back(PV1);
1516 // If this PHI node is recursive, set the size of the accessed memory to
1517 // unknown to represent all the possible values the GEP could advance the
1520 PNSize = MemoryLocation::UnknownSize;
1523 aliasCheck(V2, V2Size, V2AAInfo, V1Srcs[0],
1524 PNSize, PNAAInfo, UnderV2);
1526 // Early exit if the check of the first PHI source against V2 is MayAlias.
1527 // Other results are not possible.
1528 if (Alias == MayAlias)
1531 // If all sources of the PHI node NoAlias or MustAlias V2, then returns
1532 // NoAlias / MustAlias. Otherwise, returns MayAlias.
1533 for (unsigned i = 1, e = V1Srcs.size(); i != e; ++i) {
1534 Value *V = V1Srcs[i];
1536 AliasResult ThisAlias =
1537 aliasCheck(V2, V2Size, V2AAInfo, V, PNSize, PNAAInfo, UnderV2);
1538 Alias = MergeAliasResults(ThisAlias, Alias);
1539 if (Alias == MayAlias)
1546 /// Provides a bunch of ad-hoc rules to disambiguate in common cases, such as
1547 /// array references.
1548 AliasResult BasicAAResult::aliasCheck(const Value *V1, uint64_t V1Size,
1549 AAMDNodes V1AAInfo, const Value *V2,
1550 uint64_t V2Size, AAMDNodes V2AAInfo,
1551 const Value *O1, const Value *O2) {
1552 // If either of the memory references is empty, it doesn't matter what the
1553 // pointer values are.
1554 if (V1Size == 0 || V2Size == 0)
1557 // Strip off any casts if they exist.
1558 V1 = V1->stripPointerCastsAndBarriers();
1559 V2 = V2->stripPointerCastsAndBarriers();
1561 // If V1 or V2 is undef, the result is NoAlias because we can always pick a
1562 // value for undef that aliases nothing in the program.
1563 if (isa<UndefValue>(V1) || isa<UndefValue>(V2))
1566 // Are we checking for alias of the same value?
1567 // Because we look 'through' phi nodes, we could look at "Value" pointers from
1568 // different iterations. We must therefore make sure that this is not the
1569 // case. The function isValueEqualInPotentialCycles ensures that this cannot
1570 // happen by looking at the visited phi nodes and making sure they cannot
1572 if (isValueEqualInPotentialCycles(V1, V2))
1575 if (!V1->getType()->isPointerTy() || !V2->getType()->isPointerTy())
1576 return NoAlias; // Scalars cannot alias each other
1578 // Figure out what objects these things are pointing to if we can.
1580 O1 = GetUnderlyingObject(V1, DL, MaxLookupSearchDepth);
1583 O2 = GetUnderlyingObject(V2, DL, MaxLookupSearchDepth);
1585 // Null values in the default address space don't point to any object, so they
1586 // don't alias any other pointer.
1587 if (const ConstantPointerNull *CPN = dyn_cast<ConstantPointerNull>(O1))
1588 if (CPN->getType()->getAddressSpace() == 0)
1590 if (const ConstantPointerNull *CPN = dyn_cast<ConstantPointerNull>(O2))
1591 if (CPN->getType()->getAddressSpace() == 0)
1595 // If V1/V2 point to two different objects, we know that we have no alias.
1596 if (isIdentifiedObject(O1) && isIdentifiedObject(O2))
1599 // Constant pointers can't alias with non-const isIdentifiedObject objects.
1600 if ((isa<Constant>(O1) && isIdentifiedObject(O2) && !isa<Constant>(O2)) ||
1601 (isa<Constant>(O2) && isIdentifiedObject(O1) && !isa<Constant>(O1)))
1604 // Function arguments can't alias with things that are known to be
1605 // unambigously identified at the function level.
1606 if ((isa<Argument>(O1) && isIdentifiedFunctionLocal(O2)) ||
1607 (isa<Argument>(O2) && isIdentifiedFunctionLocal(O1)))
1610 // If one pointer is the result of a call/invoke or load and the other is a
1611 // non-escaping local object within the same function, then we know the
1612 // object couldn't escape to a point where the call could return it.
1614 // Note that if the pointers are in different functions, there are a
1615 // variety of complications. A call with a nocapture argument may still
1616 // temporary store the nocapture argument's value in a temporary memory
1617 // location if that memory location doesn't escape. Or it may pass a
1618 // nocapture value to other functions as long as they don't capture it.
1619 if (isEscapeSource(O1) && isNonEscapingLocalObject(O2))
1621 if (isEscapeSource(O2) && isNonEscapingLocalObject(O1))
1625 // If the size of one access is larger than the entire object on the other
1626 // side, then we know such behavior is undefined and can assume no alias.
1627 if ((V1Size != MemoryLocation::UnknownSize &&
1628 isObjectSmallerThan(O2, V1Size, DL, TLI)) ||
1629 (V2Size != MemoryLocation::UnknownSize &&
1630 isObjectSmallerThan(O1, V2Size, DL, TLI)))
1633 // Check the cache before climbing up use-def chains. This also terminates
1634 // otherwise infinitely recursive queries.
1635 LocPair Locs(MemoryLocation(V1, V1Size, V1AAInfo),
1636 MemoryLocation(V2, V2Size, V2AAInfo));
1638 std::swap(Locs.first, Locs.second);
1639 std::pair<AliasCacheTy::iterator, bool> Pair =
1640 AliasCache.insert(std::make_pair(Locs, MayAlias));
1642 return Pair.first->second;
1644 // FIXME: This isn't aggressively handling alias(GEP, PHI) for example: if the
1645 // GEP can't simplify, we don't even look at the PHI cases.
1646 if (!isa<GEPOperator>(V1) && isa<GEPOperator>(V2)) {
1648 std::swap(V1Size, V2Size);
1650 std::swap(V1AAInfo, V2AAInfo);
1652 if (const GEPOperator *GV1 = dyn_cast<GEPOperator>(V1)) {
1653 AliasResult Result =
1654 aliasGEP(GV1, V1Size, V1AAInfo, V2, V2Size, V2AAInfo, O1, O2);
1655 if (Result != MayAlias)
1656 return AliasCache[Locs] = Result;
1659 if (isa<PHINode>(V2) && !isa<PHINode>(V1)) {
1662 std::swap(V1Size, V2Size);
1663 std::swap(V1AAInfo, V2AAInfo);
1665 if (const PHINode *PN = dyn_cast<PHINode>(V1)) {
1666 AliasResult Result = aliasPHI(PN, V1Size, V1AAInfo,
1667 V2, V2Size, V2AAInfo, O2);
1668 if (Result != MayAlias)
1669 return AliasCache[Locs] = Result;
1672 if (isa<SelectInst>(V2) && !isa<SelectInst>(V1)) {
1675 std::swap(V1Size, V2Size);
1676 std::swap(V1AAInfo, V2AAInfo);
1678 if (const SelectInst *S1 = dyn_cast<SelectInst>(V1)) {
1679 AliasResult Result =
1680 aliasSelect(S1, V1Size, V1AAInfo, V2, V2Size, V2AAInfo, O2);
1681 if (Result != MayAlias)
1682 return AliasCache[Locs] = Result;
1685 // If both pointers are pointing into the same object and one of them
1686 // accesses the entire object, then the accesses must overlap in some way.
1688 if (V1Size != MemoryLocation::UnknownSize &&
1689 V2Size != MemoryLocation::UnknownSize &&
1690 (isObjectSize(O1, V1Size, DL, TLI) ||
1691 isObjectSize(O2, V2Size, DL, TLI)))
1692 return AliasCache[Locs] = PartialAlias;
1694 // Recurse back into the best AA results we have, potentially with refined
1695 // memory locations. We have already ensured that BasicAA has a MayAlias
1696 // cache result for these, so any recursion back into BasicAA won't loop.
1697 AliasResult Result = getBestAAResults().alias(Locs.first, Locs.second);
1698 return AliasCache[Locs] = Result;
1701 /// Check whether two Values can be considered equivalent.
1703 /// In addition to pointer equivalence of \p V1 and \p V2 this checks whether
1704 /// they can not be part of a cycle in the value graph by looking at all
1705 /// visited phi nodes an making sure that the phis cannot reach the value. We
1706 /// have to do this because we are looking through phi nodes (That is we say
1707 /// noalias(V, phi(VA, VB)) if noalias(V, VA) and noalias(V, VB).
1708 bool BasicAAResult::isValueEqualInPotentialCycles(const Value *V,
1713 const Instruction *Inst = dyn_cast<Instruction>(V);
1717 if (VisitedPhiBBs.empty())
1720 if (VisitedPhiBBs.size() > MaxNumPhiBBsValueReachabilityCheck)
1723 // Make sure that the visited phis cannot reach the Value. This ensures that
1724 // the Values cannot come from different iterations of a potential cycle the
1725 // phi nodes could be involved in.
1726 for (auto *P : VisitedPhiBBs)
1727 if (isPotentiallyReachable(&P->front(), Inst, DT, LI))
1733 /// Computes the symbolic difference between two de-composed GEPs.
1735 /// Dest and Src are the variable indices from two decomposed GetElementPtr
1736 /// instructions GEP1 and GEP2 which have common base pointers.
1737 void BasicAAResult::GetIndexDifference(
1738 SmallVectorImpl<VariableGEPIndex> &Dest,
1739 const SmallVectorImpl<VariableGEPIndex> &Src) {
1743 for (unsigned i = 0, e = Src.size(); i != e; ++i) {
1744 const Value *V = Src[i].V;
1745 unsigned ZExtBits = Src[i].ZExtBits, SExtBits = Src[i].SExtBits;
1746 int64_t Scale = Src[i].Scale;
1748 // Find V in Dest. This is N^2, but pointer indices almost never have more
1749 // than a few variable indexes.
1750 for (unsigned j = 0, e = Dest.size(); j != e; ++j) {
1751 if (!isValueEqualInPotentialCycles(Dest[j].V, V) ||
1752 Dest[j].ZExtBits != ZExtBits || Dest[j].SExtBits != SExtBits)
1755 // If we found it, subtract off Scale V's from the entry in Dest. If it
1756 // goes to zero, remove the entry.
1757 if (Dest[j].Scale != Scale)
1758 Dest[j].Scale -= Scale;
1760 Dest.erase(Dest.begin() + j);
1765 // If we didn't consume this entry, add it to the end of the Dest list.
1767 VariableGEPIndex Entry = {V, ZExtBits, SExtBits, -Scale};
1768 Dest.push_back(Entry);
1773 bool BasicAAResult::constantOffsetHeuristic(
1774 const SmallVectorImpl<VariableGEPIndex> &VarIndices, uint64_t V1Size,
1775 uint64_t V2Size, int64_t BaseOffset, AssumptionCache *AC,
1776 DominatorTree *DT) {
1777 if (VarIndices.size() != 2 || V1Size == MemoryLocation::UnknownSize ||
1778 V2Size == MemoryLocation::UnknownSize)
1781 const VariableGEPIndex &Var0 = VarIndices[0], &Var1 = VarIndices[1];
1783 if (Var0.ZExtBits != Var1.ZExtBits || Var0.SExtBits != Var1.SExtBits ||
1784 Var0.Scale != -Var1.Scale)
1787 unsigned Width = Var1.V->getType()->getIntegerBitWidth();
1789 // We'll strip off the Extensions of Var0 and Var1 and do another round
1790 // of GetLinearExpression decomposition. In the example above, if Var0
1791 // is zext(%x + 1) we should get V1 == %x and V1Offset == 1.
1793 APInt V0Scale(Width, 0), V0Offset(Width, 0), V1Scale(Width, 0),
1795 bool NSW = true, NUW = true;
1796 unsigned V0ZExtBits = 0, V0SExtBits = 0, V1ZExtBits = 0, V1SExtBits = 0;
1797 const Value *V0 = GetLinearExpression(Var0.V, V0Scale, V0Offset, V0ZExtBits,
1798 V0SExtBits, DL, 0, AC, DT, NSW, NUW);
1801 const Value *V1 = GetLinearExpression(Var1.V, V1Scale, V1Offset, V1ZExtBits,
1802 V1SExtBits, DL, 0, AC, DT, NSW, NUW);
1804 if (V0Scale != V1Scale || V0ZExtBits != V1ZExtBits ||
1805 V0SExtBits != V1SExtBits || !isValueEqualInPotentialCycles(V0, V1))
1808 // We have a hit - Var0 and Var1 only differ by a constant offset!
1810 // If we've been sext'ed then zext'd the maximum difference between Var0 and
1811 // Var1 is possible to calculate, but we're just interested in the absolute
1812 // minimum difference between the two. The minimum distance may occur due to
1813 // wrapping; consider "add i3 %i, 5": if %i == 7 then 7 + 5 mod 8 == 4, and so
1814 // the minimum distance between %i and %i + 5 is 3.
1815 APInt MinDiff = V0Offset - V1Offset, Wrapped = -MinDiff;
1816 MinDiff = APIntOps::umin(MinDiff, Wrapped);
1817 uint64_t MinDiffBytes = MinDiff.getZExtValue() * std::abs(Var0.Scale);
1819 // We can't definitely say whether GEP1 is before or after V2 due to wrapping
1820 // arithmetic (i.e. for some values of GEP1 and V2 GEP1 < V2, and for other
1821 // values GEP1 > V2). We'll therefore only declare NoAlias if both V1Size and
1822 // V2Size can fit in the MinDiffBytes gap.
1823 return V1Size + std::abs(BaseOffset) <= MinDiffBytes &&
1824 V2Size + std::abs(BaseOffset) <= MinDiffBytes;
1827 //===----------------------------------------------------------------------===//
1828 // BasicAliasAnalysis Pass
1829 //===----------------------------------------------------------------------===//
1831 AnalysisKey BasicAA::Key;
1833 BasicAAResult BasicAA::run(Function &F, FunctionAnalysisManager &AM) {
1834 return BasicAAResult(F.getParent()->getDataLayout(),
1835 AM.getResult<TargetLibraryAnalysis>(F),
1836 AM.getResult<AssumptionAnalysis>(F),
1837 &AM.getResult<DominatorTreeAnalysis>(F),
1838 AM.getCachedResult<LoopAnalysis>(F));
1841 BasicAAWrapperPass::BasicAAWrapperPass() : FunctionPass(ID) {
1842 initializeBasicAAWrapperPassPass(*PassRegistry::getPassRegistry());
1845 char BasicAAWrapperPass::ID = 0;
1847 void BasicAAWrapperPass::anchor() {}
1849 INITIALIZE_PASS_BEGIN(BasicAAWrapperPass, "basicaa",
1850 "Basic Alias Analysis (stateless AA impl)", true, true)
1851 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
1852 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
1853 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
1854 INITIALIZE_PASS_END(BasicAAWrapperPass, "basicaa",
1855 "Basic Alias Analysis (stateless AA impl)", true, true)
1857 FunctionPass *llvm::createBasicAAWrapperPass() {
1858 return new BasicAAWrapperPass();
1861 bool BasicAAWrapperPass::runOnFunction(Function &F) {
1862 auto &ACT = getAnalysis<AssumptionCacheTracker>();
1863 auto &TLIWP = getAnalysis<TargetLibraryInfoWrapperPass>();
1864 auto &DTWP = getAnalysis<DominatorTreeWrapperPass>();
1865 auto *LIWP = getAnalysisIfAvailable<LoopInfoWrapperPass>();
1867 Result.reset(new BasicAAResult(F.getParent()->getDataLayout(), TLIWP.getTLI(),
1868 ACT.getAssumptionCache(F), &DTWP.getDomTree(),
1869 LIWP ? &LIWP->getLoopInfo() : nullptr));
1874 void BasicAAWrapperPass::getAnalysisUsage(AnalysisUsage &AU) const {
1875 AU.setPreservesAll();
1876 AU.addRequired<AssumptionCacheTracker>();
1877 AU.addRequired<DominatorTreeWrapperPass>();
1878 AU.addRequired<TargetLibraryInfoWrapperPass>();
1881 BasicAAResult llvm::createLegacyPMBasicAAResult(Pass &P, Function &F) {
1882 return BasicAAResult(
1883 F.getParent()->getDataLayout(),
1884 P.getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(),
1885 P.getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F));