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 (ImmutableCallSite(V))
138 if (isa<Argument>(V))
141 // The load case works because isNonEscapingLocalObject considers all
142 // stores to be escapes (it passes true for the StoreCaptures argument
143 // to PointerMayBeCaptured).
144 if (isa<LoadInst>(V))
150 /// Returns the size of the object specified by V or UnknownSize if unknown.
151 static uint64_t getObjectSize(const Value *V, const DataLayout &DL,
152 const TargetLibraryInfo &TLI,
153 bool RoundToAlign = false) {
156 Opts.RoundToAlign = RoundToAlign;
157 if (getObjectSize(V, Size, DL, &TLI, Opts))
159 return MemoryLocation::UnknownSize;
162 /// Returns true if we can prove that the object specified by V is smaller than
164 static bool isObjectSmallerThan(const Value *V, uint64_t Size,
165 const DataLayout &DL,
166 const TargetLibraryInfo &TLI) {
167 // Note that the meanings of the "object" are slightly different in the
168 // following contexts:
169 // c1: llvm::getObjectSize()
170 // c2: llvm.objectsize() intrinsic
171 // c3: isObjectSmallerThan()
172 // c1 and c2 share the same meaning; however, the meaning of "object" in c3
173 // refers to the "entire object".
175 // Consider this example:
176 // char *p = (char*)malloc(100)
179 // In the context of c1 and c2, the "object" pointed by q refers to the
180 // stretch of memory of q[0:19]. So, getObjectSize(q) should return 20.
182 // However, in the context of c3, the "object" refers to the chunk of memory
183 // being allocated. So, the "object" has 100 bytes, and q points to the middle
184 // the "object". In case q is passed to isObjectSmallerThan() as the 1st
185 // parameter, before the llvm::getObjectSize() is called to get the size of
186 // entire object, we should:
187 // - either rewind the pointer q to the base-address of the object in
188 // question (in this case rewind to p), or
189 // - just give up. It is up to caller to make sure the pointer is pointing
190 // to the base address the object.
192 // We go for 2nd option for simplicity.
193 if (!isIdentifiedObject(V))
196 // This function needs to use the aligned object size because we allow
197 // reads a bit past the end given sufficient alignment.
198 uint64_t ObjectSize = getObjectSize(V, DL, TLI, /*RoundToAlign*/ true);
200 return ObjectSize != MemoryLocation::UnknownSize && ObjectSize < Size;
203 /// Returns true if we can prove that the object specified by V has size Size.
204 static bool isObjectSize(const Value *V, uint64_t Size, const DataLayout &DL,
205 const TargetLibraryInfo &TLI) {
206 uint64_t ObjectSize = getObjectSize(V, DL, TLI);
207 return ObjectSize != MemoryLocation::UnknownSize && ObjectSize == Size;
210 //===----------------------------------------------------------------------===//
211 // GetElementPtr Instruction Decomposition and Analysis
212 //===----------------------------------------------------------------------===//
214 /// Analyzes the specified value as a linear expression: "A*V + B", where A and
215 /// B are constant integers.
217 /// Returns the scale and offset values as APInts and return V as a Value*, and
218 /// return whether we looked through any sign or zero extends. The incoming
219 /// Value is known to have IntegerType, and it may already be sign or zero
222 /// Note that this looks through extends, so the high bits may not be
223 /// represented in the result.
224 /*static*/ const Value *BasicAAResult::GetLinearExpression(
225 const Value *V, APInt &Scale, APInt &Offset, unsigned &ZExtBits,
226 unsigned &SExtBits, const DataLayout &DL, unsigned Depth,
227 AssumptionCache *AC, DominatorTree *DT, bool &NSW, bool &NUW) {
228 assert(V->getType()->isIntegerTy() && "Not an integer value");
230 // Limit our recursion depth.
237 if (const ConstantInt *Const = dyn_cast<ConstantInt>(V)) {
238 // If it's a constant, just convert it to an offset and remove the variable.
239 // If we've been called recursively, the Offset bit width will be greater
240 // than the constant's (the Offset's always as wide as the outermost call),
241 // so we'll zext here and process any extension in the isa<SExtInst> &
242 // isa<ZExtInst> cases below.
243 Offset += Const->getValue().zextOrSelf(Offset.getBitWidth());
244 assert(Scale == 0 && "Constant values don't have a scale");
248 if (const BinaryOperator *BOp = dyn_cast<BinaryOperator>(V)) {
249 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(BOp->getOperand(1))) {
250 // If we've been called recursively, then Offset and Scale will be wider
251 // than the BOp operands. We'll always zext it here as we'll process sign
252 // extensions below (see the isa<SExtInst> / isa<ZExtInst> cases).
253 APInt RHS = RHSC->getValue().zextOrSelf(Offset.getBitWidth());
255 switch (BOp->getOpcode()) {
257 // We don't understand this instruction, so we can't decompose it any
262 case Instruction::Or:
263 // X|C == X+C if all the bits in C are unset in X. Otherwise we can't
265 if (!MaskedValueIsZero(BOp->getOperand(0), RHSC->getValue(), DL, 0, AC,
272 case Instruction::Add:
273 V = GetLinearExpression(BOp->getOperand(0), Scale, Offset, ZExtBits,
274 SExtBits, DL, Depth + 1, AC, DT, NSW, NUW);
277 case Instruction::Sub:
278 V = GetLinearExpression(BOp->getOperand(0), Scale, Offset, ZExtBits,
279 SExtBits, DL, Depth + 1, AC, DT, NSW, NUW);
282 case Instruction::Mul:
283 V = GetLinearExpression(BOp->getOperand(0), Scale, Offset, ZExtBits,
284 SExtBits, DL, Depth + 1, AC, DT, NSW, NUW);
288 case Instruction::Shl:
289 V = GetLinearExpression(BOp->getOperand(0), Scale, Offset, ZExtBits,
290 SExtBits, DL, Depth + 1, AC, DT, NSW, NUW);
292 // We're trying to linearize an expression of the kind:
294 // where the shift count exceeds the bitwidth of the type.
295 // We can't decompose this further (the expression would return
297 if (Offset.getBitWidth() < RHS.getLimitedValue() ||
298 Scale.getBitWidth() < RHS.getLimitedValue()) {
304 Offset <<= RHS.getLimitedValue();
305 Scale <<= RHS.getLimitedValue();
306 // the semantics of nsw and nuw for left shifts don't match those of
307 // multiplications, so we won't propagate them.
312 if (isa<OverflowingBinaryOperator>(BOp)) {
313 NUW &= BOp->hasNoUnsignedWrap();
314 NSW &= BOp->hasNoSignedWrap();
320 // Since GEP indices are sign extended anyway, we don't care about the high
321 // bits of a sign or zero extended value - just scales and offsets. The
322 // extensions have to be consistent though.
323 if (isa<SExtInst>(V) || isa<ZExtInst>(V)) {
324 Value *CastOp = cast<CastInst>(V)->getOperand(0);
325 unsigned NewWidth = V->getType()->getPrimitiveSizeInBits();
326 unsigned SmallWidth = CastOp->getType()->getPrimitiveSizeInBits();
327 unsigned OldZExtBits = ZExtBits, OldSExtBits = SExtBits;
328 const Value *Result =
329 GetLinearExpression(CastOp, Scale, Offset, ZExtBits, SExtBits, DL,
330 Depth + 1, AC, DT, NSW, NUW);
332 // zext(zext(%x)) == zext(%x), and similarly for sext; we'll handle this
333 // by just incrementing the number of bits we've extended by.
334 unsigned ExtendedBy = NewWidth - SmallWidth;
336 if (isa<SExtInst>(V) && ZExtBits == 0) {
337 // sext(sext(%x, a), b) == sext(%x, a + b)
340 // We haven't sign-wrapped, so it's valid to decompose sext(%x + c)
341 // into sext(%x) + sext(c). We'll sext the Offset ourselves:
342 unsigned OldWidth = Offset.getBitWidth();
343 Offset = Offset.trunc(SmallWidth).sext(NewWidth).zextOrSelf(OldWidth);
345 // We may have signed-wrapped, so don't decompose sext(%x + c) into
346 // sext(%x) + sext(c)
350 ZExtBits = OldZExtBits;
351 SExtBits = OldSExtBits;
353 SExtBits += ExtendedBy;
355 // sext(zext(%x, a), b) = zext(zext(%x, a), b) = zext(%x, a + b)
358 // We may have unsigned-wrapped, so don't decompose zext(%x + c) into
359 // zext(%x) + zext(c)
363 ZExtBits = OldZExtBits;
364 SExtBits = OldSExtBits;
366 ZExtBits += ExtendedBy;
377 /// To ensure a pointer offset fits in an integer of size PointerSize
378 /// (in bits) when that size is smaller than 64. This is an issue in
379 /// particular for 32b programs with negative indices that rely on two's
380 /// complement wrap-arounds for precise alias information.
381 static int64_t adjustToPointerSize(int64_t Offset, unsigned PointerSize) {
382 assert(PointerSize <= 64 && "Invalid PointerSize!");
383 unsigned ShiftBits = 64 - PointerSize;
384 return (int64_t)((uint64_t)Offset << ShiftBits) >> ShiftBits;
387 /// If V is a symbolic pointer expression, decompose it into a base pointer
388 /// with a constant offset and a number of scaled symbolic offsets.
390 /// The scaled symbolic offsets (represented by pairs of a Value* and a scale
391 /// in the VarIndices vector) are Value*'s that are known to be scaled by the
392 /// specified amount, but which may have other unrepresented high bits. As
393 /// such, the gep cannot necessarily be reconstructed from its decomposed form.
395 /// When DataLayout is around, this function is capable of analyzing everything
396 /// that GetUnderlyingObject can look through. To be able to do that
397 /// GetUnderlyingObject and DecomposeGEPExpression must use the same search
398 /// depth (MaxLookupSearchDepth). When DataLayout not is around, it just looks
399 /// through pointer casts.
400 bool BasicAAResult::DecomposeGEPExpression(const Value *V,
401 DecomposedGEP &Decomposed, const DataLayout &DL, AssumptionCache *AC,
403 // Limit recursion depth to limit compile time in crazy cases.
404 unsigned MaxLookup = MaxLookupSearchDepth;
407 Decomposed.StructOffset = 0;
408 Decomposed.OtherOffset = 0;
409 Decomposed.VarIndices.clear();
411 // See if this is a bitcast or GEP.
412 const Operator *Op = dyn_cast<Operator>(V);
414 // The only non-operator case we can handle are GlobalAliases.
415 if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
416 if (!GA->isInterposable()) {
417 V = GA->getAliasee();
425 if (Op->getOpcode() == Instruction::BitCast ||
426 Op->getOpcode() == Instruction::AddrSpaceCast) {
427 V = Op->getOperand(0);
431 const GEPOperator *GEPOp = dyn_cast<GEPOperator>(Op);
433 if (auto CS = ImmutableCallSite(V)) {
434 // Note: getArgumentAliasingToReturnedPointer keeps it in sync with
435 // CaptureTracking, which is needed for correctness. This is because
436 // some intrinsics like launder.invariant.group returns pointers that
437 // are aliasing it's argument, which is known to CaptureTracking.
438 // If AliasAnalysis does not use the same information, it could assume
439 // that pointer returned from launder does not alias it's argument
440 // because launder could not return it if the pointer was not captured.
441 if (auto *RP = getArgumentAliasingToReturnedPointer(CS)) {
447 // If it's not a GEP, hand it off to SimplifyInstruction to see if it
448 // can come up with something. This matches what GetUnderlyingObject does.
449 if (const Instruction *I = dyn_cast<Instruction>(V))
450 // TODO: Get a DominatorTree and AssumptionCache and use them here
451 // (these are both now available in this function, but this should be
452 // updated when GetUnderlyingObject is updated). TLI should be
454 if (const Value *Simplified =
455 SimplifyInstruction(const_cast<Instruction *>(I), DL)) {
464 // Don't attempt to analyze GEPs over unsized objects.
465 if (!GEPOp->getSourceElementType()->isSized()) {
470 unsigned AS = GEPOp->getPointerAddressSpace();
471 // Walk the indices of the GEP, accumulating them into BaseOff/VarIndices.
472 gep_type_iterator GTI = gep_type_begin(GEPOp);
473 unsigned PointerSize = DL.getPointerSizeInBits(AS);
474 // Assume all GEP operands are constants until proven otherwise.
475 bool GepHasConstantOffset = true;
476 for (User::const_op_iterator I = GEPOp->op_begin() + 1, E = GEPOp->op_end();
477 I != E; ++I, ++GTI) {
478 const Value *Index = *I;
479 // Compute the (potentially symbolic) offset in bytes for this index.
480 if (StructType *STy = GTI.getStructTypeOrNull()) {
481 // For a struct, add the member offset.
482 unsigned FieldNo = cast<ConstantInt>(Index)->getZExtValue();
486 Decomposed.StructOffset +=
487 DL.getStructLayout(STy)->getElementOffset(FieldNo);
491 // For an array/pointer, add the element offset, explicitly scaled.
492 if (const ConstantInt *CIdx = dyn_cast<ConstantInt>(Index)) {
495 Decomposed.OtherOffset +=
496 DL.getTypeAllocSize(GTI.getIndexedType()) * CIdx->getSExtValue();
500 GepHasConstantOffset = false;
502 uint64_t Scale = DL.getTypeAllocSize(GTI.getIndexedType());
503 unsigned ZExtBits = 0, SExtBits = 0;
505 // If the integer type is smaller than the pointer size, it is implicitly
506 // sign extended to pointer size.
507 unsigned Width = Index->getType()->getIntegerBitWidth();
508 if (PointerSize > Width)
509 SExtBits += PointerSize - Width;
511 // Use GetLinearExpression to decompose the index into a C1*V+C2 form.
512 APInt IndexScale(Width, 0), IndexOffset(Width, 0);
513 bool NSW = true, NUW = true;
514 Index = GetLinearExpression(Index, IndexScale, IndexOffset, ZExtBits,
515 SExtBits, DL, 0, AC, DT, NSW, NUW);
517 // All GEP math happens in the width of the pointer type,
518 // so we can truncate the value to 64-bits as we don't handle
519 // currently pointers larger than 64 bits and we would crash
520 // later. TODO: Make `Scale` an APInt to avoid this problem.
521 if (IndexScale.getBitWidth() > 64)
522 IndexScale = IndexScale.sextOrTrunc(64);
524 // The GEP index scale ("Scale") scales C1*V+C2, yielding (C1*V+C2)*Scale.
525 // This gives us an aggregate computation of (C1*Scale)*V + C2*Scale.
526 Decomposed.OtherOffset += IndexOffset.getSExtValue() * Scale;
527 Scale *= IndexScale.getSExtValue();
529 // If we already had an occurrence of this index variable, merge this
530 // scale into it. For example, we want to handle:
531 // A[x][x] -> x*16 + x*4 -> x*20
532 // This also ensures that 'x' only appears in the index list once.
533 for (unsigned i = 0, e = Decomposed.VarIndices.size(); i != e; ++i) {
534 if (Decomposed.VarIndices[i].V == Index &&
535 Decomposed.VarIndices[i].ZExtBits == ZExtBits &&
536 Decomposed.VarIndices[i].SExtBits == SExtBits) {
537 Scale += Decomposed.VarIndices[i].Scale;
538 Decomposed.VarIndices.erase(Decomposed.VarIndices.begin() + i);
543 // Make sure that we have a scale that makes sense for this target's
545 Scale = adjustToPointerSize(Scale, PointerSize);
548 VariableGEPIndex Entry = {Index, ZExtBits, SExtBits,
549 static_cast<int64_t>(Scale)};
550 Decomposed.VarIndices.push_back(Entry);
554 // Take care of wrap-arounds
555 if (GepHasConstantOffset) {
556 Decomposed.StructOffset =
557 adjustToPointerSize(Decomposed.StructOffset, PointerSize);
558 Decomposed.OtherOffset =
559 adjustToPointerSize(Decomposed.OtherOffset, PointerSize);
562 // Analyze the base pointer next.
563 V = GEPOp->getOperand(0);
564 } while (--MaxLookup);
566 // If the chain of expressions is too deep, just return early.
568 SearchLimitReached++;
572 /// Returns whether the given pointer value points to memory that is local to
573 /// the function, with global constants being considered local to all
575 bool BasicAAResult::pointsToConstantMemory(const MemoryLocation &Loc,
577 assert(Visited.empty() && "Visited must be cleared after use!");
579 unsigned MaxLookup = 8;
580 SmallVector<const Value *, 16> Worklist;
581 Worklist.push_back(Loc.Ptr);
583 const Value *V = GetUnderlyingObject(Worklist.pop_back_val(), DL);
584 if (!Visited.insert(V).second) {
586 return AAResultBase::pointsToConstantMemory(Loc, OrLocal);
589 // An alloca instruction defines local memory.
590 if (OrLocal && isa<AllocaInst>(V))
593 // A global constant counts as local memory for our purposes.
594 if (const GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) {
595 // Note: this doesn't require GV to be "ODR" because it isn't legal for a
596 // global to be marked constant in some modules and non-constant in
597 // others. GV may even be a declaration, not a definition.
598 if (!GV->isConstant()) {
600 return AAResultBase::pointsToConstantMemory(Loc, OrLocal);
605 // If both select values point to local memory, then so does the select.
606 if (const SelectInst *SI = dyn_cast<SelectInst>(V)) {
607 Worklist.push_back(SI->getTrueValue());
608 Worklist.push_back(SI->getFalseValue());
612 // If all values incoming to a phi node point to local memory, then so does
614 if (const PHINode *PN = dyn_cast<PHINode>(V)) {
615 // Don't bother inspecting phi nodes with many operands.
616 if (PN->getNumIncomingValues() > MaxLookup) {
618 return AAResultBase::pointsToConstantMemory(Loc, OrLocal);
620 for (Value *IncValue : PN->incoming_values())
621 Worklist.push_back(IncValue);
625 // Otherwise be conservative.
627 return AAResultBase::pointsToConstantMemory(Loc, OrLocal);
628 } while (!Worklist.empty() && --MaxLookup);
631 return Worklist.empty();
634 /// Returns the behavior when calling the given call site.
635 FunctionModRefBehavior BasicAAResult::getModRefBehavior(ImmutableCallSite CS) {
636 if (CS.doesNotAccessMemory())
637 // Can't do better than this.
638 return FMRB_DoesNotAccessMemory;
640 FunctionModRefBehavior Min = FMRB_UnknownModRefBehavior;
642 // If the callsite knows it only reads memory, don't return worse
644 if (CS.onlyReadsMemory())
645 Min = FMRB_OnlyReadsMemory;
646 else if (CS.doesNotReadMemory())
647 Min = FMRB_DoesNotReadMemory;
649 if (CS.onlyAccessesArgMemory())
650 Min = FunctionModRefBehavior(Min & FMRB_OnlyAccessesArgumentPointees);
651 else if (CS.onlyAccessesInaccessibleMemory())
652 Min = FunctionModRefBehavior(Min & FMRB_OnlyAccessesInaccessibleMem);
653 else if (CS.onlyAccessesInaccessibleMemOrArgMem())
654 Min = FunctionModRefBehavior(Min & FMRB_OnlyAccessesInaccessibleOrArgMem);
656 // If CS has operand bundles then aliasing attributes from the function it
657 // calls do not directly apply to the CallSite. This can be made more
658 // precise in the future.
659 if (!CS.hasOperandBundles())
660 if (const Function *F = CS.getCalledFunction())
662 FunctionModRefBehavior(Min & getBestAAResults().getModRefBehavior(F));
667 /// Returns the behavior when calling the given function. For use when the call
668 /// site is not known.
669 FunctionModRefBehavior BasicAAResult::getModRefBehavior(const Function *F) {
670 // If the function declares it doesn't access memory, we can't do better.
671 if (F->doesNotAccessMemory())
672 return FMRB_DoesNotAccessMemory;
674 FunctionModRefBehavior Min = FMRB_UnknownModRefBehavior;
676 // If the function declares it only reads memory, go with that.
677 if (F->onlyReadsMemory())
678 Min = FMRB_OnlyReadsMemory;
679 else if (F->doesNotReadMemory())
680 Min = FMRB_DoesNotReadMemory;
682 if (F->onlyAccessesArgMemory())
683 Min = FunctionModRefBehavior(Min & FMRB_OnlyAccessesArgumentPointees);
684 else if (F->onlyAccessesInaccessibleMemory())
685 Min = FunctionModRefBehavior(Min & FMRB_OnlyAccessesInaccessibleMem);
686 else if (F->onlyAccessesInaccessibleMemOrArgMem())
687 Min = FunctionModRefBehavior(Min & FMRB_OnlyAccessesInaccessibleOrArgMem);
692 /// Returns true if this is a writeonly (i.e Mod only) parameter.
693 static bool isWriteOnlyParam(ImmutableCallSite CS, unsigned ArgIdx,
694 const TargetLibraryInfo &TLI) {
695 if (CS.paramHasAttr(ArgIdx, Attribute::WriteOnly))
698 // We can bound the aliasing properties of memset_pattern16 just as we can
699 // for memcpy/memset. This is particularly important because the
700 // LoopIdiomRecognizer likes to turn loops into calls to memset_pattern16
701 // whenever possible.
702 // FIXME Consider handling this in InferFunctionAttr.cpp together with other
705 if (CS.getCalledFunction() && TLI.getLibFunc(*CS.getCalledFunction(), F) &&
706 F == LibFunc_memset_pattern16 && TLI.has(F))
710 // TODO: memset_pattern4, memset_pattern8
711 // TODO: _chk variants
712 // TODO: strcmp, strcpy
717 ModRefInfo BasicAAResult::getArgModRefInfo(ImmutableCallSite CS,
719 // Checking for known builtin intrinsics and target library functions.
720 if (isWriteOnlyParam(CS, ArgIdx, TLI))
721 return ModRefInfo::Mod;
723 if (CS.paramHasAttr(ArgIdx, Attribute::ReadOnly))
724 return ModRefInfo::Ref;
726 if (CS.paramHasAttr(ArgIdx, Attribute::ReadNone))
727 return ModRefInfo::NoModRef;
729 return AAResultBase::getArgModRefInfo(CS, ArgIdx);
732 static bool isIntrinsicCall(ImmutableCallSite CS, Intrinsic::ID IID) {
733 const IntrinsicInst *II = dyn_cast<IntrinsicInst>(CS.getInstruction());
734 return II && II->getIntrinsicID() == IID;
738 static const Function *getParent(const Value *V) {
739 if (const Instruction *inst = dyn_cast<Instruction>(V)) {
740 if (!inst->getParent())
742 return inst->getParent()->getParent();
745 if (const Argument *arg = dyn_cast<Argument>(V))
746 return arg->getParent();
751 static bool notDifferentParent(const Value *O1, const Value *O2) {
753 const Function *F1 = getParent(O1);
754 const Function *F2 = getParent(O2);
756 return !F1 || !F2 || F1 == F2;
760 AliasResult BasicAAResult::alias(const MemoryLocation &LocA,
761 const MemoryLocation &LocB) {
762 assert(notDifferentParent(LocA.Ptr, LocB.Ptr) &&
763 "BasicAliasAnalysis doesn't support interprocedural queries.");
765 // If we have a directly cached entry for these locations, we have recursed
766 // through this once, so just return the cached results. Notably, when this
767 // happens, we don't clear the cache.
768 auto CacheIt = AliasCache.find(LocPair(LocA, LocB));
769 if (CacheIt != AliasCache.end())
770 return CacheIt->second;
772 AliasResult Alias = aliasCheck(LocA.Ptr, LocA.Size, LocA.AATags, LocB.Ptr,
773 LocB.Size, LocB.AATags);
774 // AliasCache rarely has more than 1 or 2 elements, always use
775 // shrink_and_clear so it quickly returns to the inline capacity of the
776 // SmallDenseMap if it ever grows larger.
777 // FIXME: This should really be shrink_to_inline_capacity_and_clear().
778 AliasCache.shrink_and_clear();
779 VisitedPhiBBs.clear();
783 /// Checks to see if the specified callsite can clobber the specified memory
786 /// Since we only look at local properties of this function, we really can't
787 /// say much about this query. We do, however, use simple "address taken"
788 /// analysis on local objects.
789 ModRefInfo BasicAAResult::getModRefInfo(ImmutableCallSite CS,
790 const MemoryLocation &Loc) {
791 assert(notDifferentParent(CS.getInstruction(), Loc.Ptr) &&
792 "AliasAnalysis query involving multiple functions!");
794 const Value *Object = GetUnderlyingObject(Loc.Ptr, DL);
796 // If this is a tail call and Loc.Ptr points to a stack location, we know that
797 // the tail call cannot access or modify the local stack.
798 // We cannot exclude byval arguments here; these belong to the caller of
799 // the current function not to the current function, and a tail callee
800 // may reference them.
801 if (isa<AllocaInst>(Object))
802 if (const CallInst *CI = dyn_cast<CallInst>(CS.getInstruction()))
803 if (CI->isTailCall())
804 return ModRefInfo::NoModRef;
806 // If the pointer is to a locally allocated object that does not escape,
807 // then the call can not mod/ref the pointer unless the call takes the pointer
808 // as an argument, and itself doesn't capture it.
809 if (!isa<Constant>(Object) && CS.getInstruction() != Object &&
810 isNonEscapingLocalObject(Object)) {
812 // Optimistically assume that call doesn't touch Object and check this
813 // assumption in the following loop.
814 ModRefInfo Result = ModRefInfo::NoModRef;
815 bool IsMustAlias = true;
817 unsigned OperandNo = 0;
818 for (auto CI = CS.data_operands_begin(), CE = CS.data_operands_end();
819 CI != CE; ++CI, ++OperandNo) {
820 // Only look at the no-capture or byval pointer arguments. If this
821 // pointer were passed to arguments that were neither of these, then it
822 // couldn't be no-capture.
823 if (!(*CI)->getType()->isPointerTy() ||
824 (!CS.doesNotCapture(OperandNo) &&
825 OperandNo < CS.getNumArgOperands() && !CS.isByValArgument(OperandNo)))
828 // Call doesn't access memory through this operand, so we don't care
829 // if it aliases with Object.
830 if (CS.doesNotAccessMemory(OperandNo))
833 // If this is a no-capture pointer argument, see if we can tell that it
834 // is impossible to alias the pointer we're checking.
836 getBestAAResults().alias(MemoryLocation(*CI), MemoryLocation(Object));
839 // Operand doesnt alias 'Object', continue looking for other aliases
842 // Operand aliases 'Object', but call doesn't modify it. Strengthen
843 // initial assumption and keep looking in case if there are more aliases.
844 if (CS.onlyReadsMemory(OperandNo)) {
845 Result = setRef(Result);
848 // Operand aliases 'Object' but call only writes into it.
849 if (CS.doesNotReadMemory(OperandNo)) {
850 Result = setMod(Result);
853 // This operand aliases 'Object' and call reads and writes into it.
854 // Setting ModRef will not yield an early return below, MustAlias is not
856 Result = ModRefInfo::ModRef;
860 // No operand aliases, reset Must bit. Add below if at least one aliases
861 // and all aliases found are MustAlias.
862 if (isNoModRef(Result))
865 // Early return if we improved mod ref information
866 if (!isModAndRefSet(Result)) {
867 if (isNoModRef(Result))
868 return ModRefInfo::NoModRef;
869 return IsMustAlias ? setMust(Result) : clearMust(Result);
873 // If the CallSite is to malloc or calloc, we can assume that it doesn't
874 // modify any IR visible value. This is only valid because we assume these
875 // routines do not read values visible in the IR. TODO: Consider special
876 // casing realloc and strdup routines which access only their arguments as
877 // well. Or alternatively, replace all of this with inaccessiblememonly once
878 // that's implemented fully.
879 auto *Inst = CS.getInstruction();
880 if (isMallocOrCallocLikeFn(Inst, &TLI)) {
881 // Be conservative if the accessed pointer may alias the allocation -
882 // fallback to the generic handling below.
883 if (getBestAAResults().alias(MemoryLocation(Inst), Loc) == NoAlias)
884 return ModRefInfo::NoModRef;
887 // The semantics of memcpy intrinsics forbid overlap between their respective
888 // operands, i.e., source and destination of any given memcpy must no-alias.
889 // If Loc must-aliases either one of these two locations, then it necessarily
890 // no-aliases the other.
891 if (auto *Inst = dyn_cast<AnyMemCpyInst>(CS.getInstruction())) {
892 AliasResult SrcAA, DestAA;
894 if ((SrcAA = getBestAAResults().alias(MemoryLocation::getForSource(Inst),
896 // Loc is exactly the memcpy source thus disjoint from memcpy dest.
897 return ModRefInfo::Ref;
898 if ((DestAA = getBestAAResults().alias(MemoryLocation::getForDest(Inst),
900 // The converse case.
901 return ModRefInfo::Mod;
903 // It's also possible for Loc to alias both src and dest, or neither.
904 ModRefInfo rv = ModRefInfo::NoModRef;
905 if (SrcAA != NoAlias)
907 if (DestAA != NoAlias)
912 // While the assume intrinsic is marked as arbitrarily writing so that
913 // proper control dependencies will be maintained, it never aliases any
914 // particular memory location.
915 if (isIntrinsicCall(CS, Intrinsic::assume))
916 return ModRefInfo::NoModRef;
918 // Like assumes, guard intrinsics are also marked as arbitrarily writing so
919 // that proper control dependencies are maintained but they never mods any
920 // particular memory location.
922 // *Unlike* assumes, guard intrinsics are modeled as reading memory since the
923 // heap state at the point the guard is issued needs to be consistent in case
924 // the guard invokes the "deopt" continuation.
925 if (isIntrinsicCall(CS, Intrinsic::experimental_guard))
926 return ModRefInfo::Ref;
928 // Like assumes, invariant.start intrinsics were also marked as arbitrarily
929 // writing so that proper control dependencies are maintained but they never
930 // mod any particular memory location visible to the IR.
931 // *Unlike* assumes (which are now modeled as NoModRef), invariant.start
932 // intrinsic is now modeled as reading memory. This prevents hoisting the
933 // invariant.start intrinsic over stores. Consider:
936 // invariant_start(ptr)
940 // This cannot be transformed to:
943 // invariant_start(ptr)
948 // The transformation will cause the second store to be ignored (based on
949 // rules of invariant.start) and print 40, while the first program always
951 if (isIntrinsicCall(CS, Intrinsic::invariant_start))
952 return ModRefInfo::Ref;
954 // The AAResultBase base class has some smarts, lets use them.
955 return AAResultBase::getModRefInfo(CS, Loc);
958 ModRefInfo BasicAAResult::getModRefInfo(ImmutableCallSite CS1,
959 ImmutableCallSite CS2) {
960 // While the assume intrinsic is marked as arbitrarily writing so that
961 // proper control dependencies will be maintained, it never aliases any
962 // particular memory location.
963 if (isIntrinsicCall(CS1, Intrinsic::assume) ||
964 isIntrinsicCall(CS2, Intrinsic::assume))
965 return ModRefInfo::NoModRef;
967 // Like assumes, guard intrinsics are also marked as arbitrarily writing so
968 // that proper control dependencies are maintained but they never mod any
969 // particular memory location.
971 // *Unlike* assumes, guard intrinsics are modeled as reading memory since the
972 // heap state at the point the guard is issued needs to be consistent in case
973 // the guard invokes the "deopt" continuation.
975 // NB! This function is *not* commutative, so we specical case two
976 // possibilities for guard intrinsics.
978 if (isIntrinsicCall(CS1, Intrinsic::experimental_guard))
979 return isModSet(createModRefInfo(getModRefBehavior(CS2)))
981 : ModRefInfo::NoModRef;
983 if (isIntrinsicCall(CS2, Intrinsic::experimental_guard))
984 return isModSet(createModRefInfo(getModRefBehavior(CS1)))
986 : ModRefInfo::NoModRef;
988 // The AAResultBase base class has some smarts, lets use them.
989 return AAResultBase::getModRefInfo(CS1, CS2);
992 /// Provide ad-hoc rules to disambiguate accesses through two GEP operators,
993 /// both having the exact same pointer operand.
994 static AliasResult aliasSameBasePointerGEPs(const GEPOperator *GEP1,
996 const GEPOperator *GEP2,
998 const DataLayout &DL) {
999 assert(GEP1->getPointerOperand()->stripPointerCastsAndInvariantGroups() ==
1000 GEP2->getPointerOperand()->stripPointerCastsAndInvariantGroups() &&
1001 GEP1->getPointerOperandType() == GEP2->getPointerOperandType() &&
1002 "Expected GEPs with the same pointer operand");
1004 // Try to determine whether GEP1 and GEP2 index through arrays, into structs,
1005 // such that the struct field accesses provably cannot alias.
1006 // We also need at least two indices (the pointer, and the struct field).
1007 if (GEP1->getNumIndices() != GEP2->getNumIndices() ||
1008 GEP1->getNumIndices() < 2)
1011 // If we don't know the size of the accesses through both GEPs, we can't
1012 // determine whether the struct fields accessed can't alias.
1013 if (V1Size == MemoryLocation::UnknownSize ||
1014 V2Size == MemoryLocation::UnknownSize)
1018 dyn_cast<ConstantInt>(GEP1->getOperand(GEP1->getNumOperands() - 1));
1020 dyn_cast<ConstantInt>(GEP2->getOperand(GEP2->getNumOperands() - 1));
1022 // If the last (struct) indices are constants and are equal, the other indices
1023 // might be also be dynamically equal, so the GEPs can alias.
1024 if (C1 && C2 && C1->getSExtValue() == C2->getSExtValue())
1027 // Find the last-indexed type of the GEP, i.e., the type you'd get if
1028 // you stripped the last index.
1029 // On the way, look at each indexed type. If there's something other
1030 // than an array, different indices can lead to different final types.
1031 SmallVector<Value *, 8> IntermediateIndices;
1033 // Insert the first index; we don't need to check the type indexed
1034 // through it as it only drops the pointer indirection.
1035 assert(GEP1->getNumIndices() > 1 && "Not enough GEP indices to examine");
1036 IntermediateIndices.push_back(GEP1->getOperand(1));
1038 // Insert all the remaining indices but the last one.
1039 // Also, check that they all index through arrays.
1040 for (unsigned i = 1, e = GEP1->getNumIndices() - 1; i != e; ++i) {
1041 if (!isa<ArrayType>(GetElementPtrInst::getIndexedType(
1042 GEP1->getSourceElementType(), IntermediateIndices)))
1044 IntermediateIndices.push_back(GEP1->getOperand(i + 1));
1047 auto *Ty = GetElementPtrInst::getIndexedType(
1048 GEP1->getSourceElementType(), IntermediateIndices);
1049 StructType *LastIndexedStruct = dyn_cast<StructType>(Ty);
1051 if (isa<SequentialType>(Ty)) {
1053 // - both GEPs begin indexing from the exact same pointer;
1054 // - the last indices in both GEPs are constants, indexing into a sequential
1055 // type (array or pointer);
1056 // - both GEPs only index through arrays prior to that.
1058 // Because array indices greater than the number of elements are valid in
1059 // GEPs, unless we know the intermediate indices are identical between
1060 // GEP1 and GEP2 we cannot guarantee that the last indexed arrays don't
1061 // partially overlap. We also need to check that the loaded size matches
1062 // the element size, otherwise we could still have overlap.
1063 const uint64_t ElementSize =
1064 DL.getTypeStoreSize(cast<SequentialType>(Ty)->getElementType());
1065 if (V1Size != ElementSize || V2Size != ElementSize)
1068 for (unsigned i = 0, e = GEP1->getNumIndices() - 1; i != e; ++i)
1069 if (GEP1->getOperand(i + 1) != GEP2->getOperand(i + 1))
1072 // Now we know that the array/pointer that GEP1 indexes into and that
1073 // that GEP2 indexes into must either precisely overlap or be disjoint.
1074 // Because they cannot partially overlap and because fields in an array
1075 // cannot overlap, if we can prove the final indices are different between
1076 // GEP1 and GEP2, we can conclude GEP1 and GEP2 don't alias.
1078 // If the last indices are constants, we've already checked they don't
1079 // equal each other so we can exit early.
1083 Value *GEP1LastIdx = GEP1->getOperand(GEP1->getNumOperands() - 1);
1084 Value *GEP2LastIdx = GEP2->getOperand(GEP2->getNumOperands() - 1);
1085 if (isa<PHINode>(GEP1LastIdx) || isa<PHINode>(GEP2LastIdx)) {
1086 // If one of the indices is a PHI node, be safe and only use
1087 // computeKnownBits so we don't make any assumptions about the
1088 // relationships between the two indices. This is important if we're
1089 // asking about values from different loop iterations. See PR32314.
1090 // TODO: We may be able to change the check so we only do this when
1091 // we definitely looked through a PHINode.
1092 if (GEP1LastIdx != GEP2LastIdx &&
1093 GEP1LastIdx->getType() == GEP2LastIdx->getType()) {
1094 KnownBits Known1 = computeKnownBits(GEP1LastIdx, DL);
1095 KnownBits Known2 = computeKnownBits(GEP2LastIdx, DL);
1096 if (Known1.Zero.intersects(Known2.One) ||
1097 Known1.One.intersects(Known2.Zero))
1100 } else if (isKnownNonEqual(GEP1LastIdx, GEP2LastIdx, DL))
1104 } else if (!LastIndexedStruct || !C1 || !C2) {
1109 // - both GEPs begin indexing from the exact same pointer;
1110 // - the last indices in both GEPs are constants, indexing into a struct;
1111 // - said indices are different, hence, the pointed-to fields are different;
1112 // - both GEPs only index through arrays prior to that.
1114 // This lets us determine that the struct that GEP1 indexes into and the
1115 // struct that GEP2 indexes into must either precisely overlap or be
1116 // completely disjoint. Because they cannot partially overlap, indexing into
1117 // different non-overlapping fields of the struct will never alias.
1119 // Therefore, the only remaining thing needed to show that both GEPs can't
1120 // alias is that the fields are not overlapping.
1121 const StructLayout *SL = DL.getStructLayout(LastIndexedStruct);
1122 const uint64_t StructSize = SL->getSizeInBytes();
1123 const uint64_t V1Off = SL->getElementOffset(C1->getZExtValue());
1124 const uint64_t V2Off = SL->getElementOffset(C2->getZExtValue());
1126 auto EltsDontOverlap = [StructSize](uint64_t V1Off, uint64_t V1Size,
1127 uint64_t V2Off, uint64_t V2Size) {
1128 return V1Off < V2Off && V1Off + V1Size <= V2Off &&
1129 ((V2Off + V2Size <= StructSize) ||
1130 (V2Off + V2Size - StructSize <= V1Off));
1133 if (EltsDontOverlap(V1Off, V1Size, V2Off, V2Size) ||
1134 EltsDontOverlap(V2Off, V2Size, V1Off, V1Size))
1140 // If a we have (a) a GEP and (b) a pointer based on an alloca, and the
1141 // beginning of the object the GEP points would have a negative offset with
1142 // repsect to the alloca, that means the GEP can not alias pointer (b).
1143 // Note that the pointer based on the alloca may not be a GEP. For
1144 // example, it may be the alloca itself.
1145 // The same applies if (b) is based on a GlobalVariable. Note that just being
1146 // based on isIdentifiedObject() is not enough - we need an identified object
1147 // that does not permit access to negative offsets. For example, a negative
1148 // offset from a noalias argument or call can be inbounds w.r.t the actual
1149 // underlying object.
1151 // For example, consider:
1153 // struct { int f0, int f1, ...} foo;
1155 // foo* random = bar(alloca);
1156 // int *f0 = &alloca.f0
1157 // int *f1 = &random->f1;
1159 // Which is lowered, approximately, to:
1161 // %alloca = alloca %struct.foo
1162 // %random = call %struct.foo* @random(%struct.foo* %alloca)
1163 // %f0 = getelementptr inbounds %struct, %struct.foo* %alloca, i32 0, i32 0
1164 // %f1 = getelementptr inbounds %struct, %struct.foo* %random, i32 0, i32 1
1166 // Assume %f1 and %f0 alias. Then %f1 would point into the object allocated
1167 // by %alloca. Since the %f1 GEP is inbounds, that means %random must also
1168 // point into the same object. But since %f0 points to the beginning of %alloca,
1169 // the highest %f1 can be is (%alloca + 3). This means %random can not be higher
1170 // than (%alloca - 1), and so is not inbounds, a contradiction.
1171 bool BasicAAResult::isGEPBaseAtNegativeOffset(const GEPOperator *GEPOp,
1172 const DecomposedGEP &DecompGEP, const DecomposedGEP &DecompObject,
1173 LocationSize ObjectAccessSize) {
1174 // If the object access size is unknown, or the GEP isn't inbounds, bail.
1175 if (ObjectAccessSize == MemoryLocation::UnknownSize || !GEPOp->isInBounds())
1178 // We need the object to be an alloca or a globalvariable, and want to know
1179 // the offset of the pointer from the object precisely, so no variable
1180 // indices are allowed.
1181 if (!(isa<AllocaInst>(DecompObject.Base) ||
1182 isa<GlobalVariable>(DecompObject.Base)) ||
1183 !DecompObject.VarIndices.empty())
1186 int64_t ObjectBaseOffset = DecompObject.StructOffset +
1187 DecompObject.OtherOffset;
1189 // If the GEP has no variable indices, we know the precise offset
1190 // from the base, then use it. If the GEP has variable indices,
1191 // we can't get exact GEP offset to identify pointer alias. So return
1192 // false in that case.
1193 if (!DecompGEP.VarIndices.empty())
1195 int64_t GEPBaseOffset = DecompGEP.StructOffset;
1196 GEPBaseOffset += DecompGEP.OtherOffset;
1198 return (GEPBaseOffset >= ObjectBaseOffset + (int64_t)ObjectAccessSize);
1201 /// Provides a bunch of ad-hoc rules to disambiguate a GEP instruction against
1202 /// another pointer.
1204 /// We know that V1 is a GEP, but we don't know anything about V2.
1205 /// UnderlyingV1 is GetUnderlyingObject(GEP1, DL), UnderlyingV2 is the same for
1208 BasicAAResult::aliasGEP(const GEPOperator *GEP1, LocationSize V1Size,
1209 const AAMDNodes &V1AAInfo, const Value *V2,
1210 LocationSize V2Size, const AAMDNodes &V2AAInfo,
1211 const Value *UnderlyingV1, const Value *UnderlyingV2) {
1212 DecomposedGEP DecompGEP1, DecompGEP2;
1213 bool GEP1MaxLookupReached =
1214 DecomposeGEPExpression(GEP1, DecompGEP1, DL, &AC, DT);
1215 bool GEP2MaxLookupReached =
1216 DecomposeGEPExpression(V2, DecompGEP2, DL, &AC, DT);
1218 int64_t GEP1BaseOffset = DecompGEP1.StructOffset + DecompGEP1.OtherOffset;
1219 int64_t GEP2BaseOffset = DecompGEP2.StructOffset + DecompGEP2.OtherOffset;
1221 assert(DecompGEP1.Base == UnderlyingV1 && DecompGEP2.Base == UnderlyingV2 &&
1222 "DecomposeGEPExpression returned a result different from "
1223 "GetUnderlyingObject");
1225 // If the GEP's offset relative to its base is such that the base would
1226 // fall below the start of the object underlying V2, then the GEP and V2
1228 if (!GEP1MaxLookupReached && !GEP2MaxLookupReached &&
1229 isGEPBaseAtNegativeOffset(GEP1, DecompGEP1, DecompGEP2, V2Size))
1231 // If we have two gep instructions with must-alias or not-alias'ing base
1232 // pointers, figure out if the indexes to the GEP tell us anything about the
1234 if (const GEPOperator *GEP2 = dyn_cast<GEPOperator>(V2)) {
1235 // Check for the GEP base being at a negative offset, this time in the other
1237 if (!GEP1MaxLookupReached && !GEP2MaxLookupReached &&
1238 isGEPBaseAtNegativeOffset(GEP2, DecompGEP2, DecompGEP1, V1Size))
1240 // Do the base pointers alias?
1241 AliasResult BaseAlias =
1242 aliasCheck(UnderlyingV1, MemoryLocation::UnknownSize, AAMDNodes(),
1243 UnderlyingV2, MemoryLocation::UnknownSize, AAMDNodes());
1245 // Check for geps of non-aliasing underlying pointers where the offsets are
1247 if ((BaseAlias == MayAlias) && V1Size == V2Size) {
1248 // Do the base pointers alias assuming type and size.
1249 AliasResult PreciseBaseAlias = aliasCheck(UnderlyingV1, V1Size, V1AAInfo,
1250 UnderlyingV2, V2Size, V2AAInfo);
1251 if (PreciseBaseAlias == NoAlias) {
1252 // See if the computed offset from the common pointer tells us about the
1253 // relation of the resulting pointer.
1254 // If the max search depth is reached the result is undefined
1255 if (GEP2MaxLookupReached || GEP1MaxLookupReached)
1259 if (GEP1BaseOffset == GEP2BaseOffset &&
1260 DecompGEP1.VarIndices == DecompGEP2.VarIndices)
1265 // If we get a No or May, then return it immediately, no amount of analysis
1266 // will improve this situation.
1267 if (BaseAlias != MustAlias) {
1268 assert(BaseAlias == NoAlias || BaseAlias == MayAlias);
1272 // Otherwise, we have a MustAlias. Since the base pointers alias each other
1273 // exactly, see if the computed offset from the common pointer tells us
1274 // about the relation of the resulting pointer.
1275 // If we know the two GEPs are based off of the exact same pointer (and not
1276 // just the same underlying object), see if that tells us anything about
1277 // the resulting pointers.
1278 if (GEP1->getPointerOperand()->stripPointerCastsAndInvariantGroups() ==
1279 GEP2->getPointerOperand()->stripPointerCastsAndInvariantGroups() &&
1280 GEP1->getPointerOperandType() == GEP2->getPointerOperandType()) {
1281 AliasResult R = aliasSameBasePointerGEPs(GEP1, V1Size, GEP2, V2Size, DL);
1282 // If we couldn't find anything interesting, don't abandon just yet.
1287 // If the max search depth is reached, the result is undefined
1288 if (GEP2MaxLookupReached || GEP1MaxLookupReached)
1291 // Subtract the GEP2 pointer from the GEP1 pointer to find out their
1292 // symbolic difference.
1293 GEP1BaseOffset -= GEP2BaseOffset;
1294 GetIndexDifference(DecompGEP1.VarIndices, DecompGEP2.VarIndices);
1297 // Check to see if these two pointers are related by the getelementptr
1298 // instruction. If one pointer is a GEP with a non-zero index of the other
1299 // pointer, we know they cannot alias.
1301 // If both accesses are unknown size, we can't do anything useful here.
1302 if (V1Size == MemoryLocation::UnknownSize &&
1303 V2Size == MemoryLocation::UnknownSize)
1306 AliasResult R = aliasCheck(UnderlyingV1, MemoryLocation::UnknownSize,
1307 AAMDNodes(), V2, MemoryLocation::UnknownSize,
1308 V2AAInfo, nullptr, UnderlyingV2);
1309 if (R != MustAlias) {
1310 // If V2 may alias GEP base pointer, conservatively returns MayAlias.
1311 // If V2 is known not to alias GEP base pointer, then the two values
1312 // cannot alias per GEP semantics: "Any memory access must be done through
1313 // a pointer value associated with an address range of the memory access,
1314 // otherwise the behavior is undefined.".
1315 assert(R == NoAlias || R == MayAlias);
1319 // If the max search depth is reached the result is undefined
1320 if (GEP1MaxLookupReached)
1324 // In the two GEP Case, if there is no difference in the offsets of the
1325 // computed pointers, the resultant pointers are a must alias. This
1326 // happens when we have two lexically identical GEP's (for example).
1328 // In the other case, if we have getelementptr <ptr>, 0, 0, 0, 0, ... and V2
1329 // must aliases the GEP, the end result is a must alias also.
1330 if (GEP1BaseOffset == 0 && DecompGEP1.VarIndices.empty())
1333 // If there is a constant difference between the pointers, but the difference
1334 // is less than the size of the associated memory object, then we know
1335 // that the objects are partially overlapping. If the difference is
1336 // greater, we know they do not overlap.
1337 if (GEP1BaseOffset != 0 && DecompGEP1.VarIndices.empty()) {
1338 if (GEP1BaseOffset >= 0) {
1339 if (V2Size != MemoryLocation::UnknownSize) {
1340 if ((uint64_t)GEP1BaseOffset < V2Size)
1341 return PartialAlias;
1345 // We have the situation where:
1348 // ---------------->|
1349 // |-->V1Size |-------> V2Size
1351 // We need to know that V2Size is not unknown, otherwise we might have
1352 // stripped a gep with negative index ('gep <ptr>, -1, ...).
1353 if (V1Size != MemoryLocation::UnknownSize &&
1354 V2Size != MemoryLocation::UnknownSize) {
1355 if (-(uint64_t)GEP1BaseOffset < V1Size)
1356 return PartialAlias;
1362 if (!DecompGEP1.VarIndices.empty()) {
1363 uint64_t Modulo = 0;
1364 bool AllPositive = true;
1365 for (unsigned i = 0, e = DecompGEP1.VarIndices.size(); i != e; ++i) {
1367 // Try to distinguish something like &A[i][1] against &A[42][0].
1368 // Grab the least significant bit set in any of the scales. We
1369 // don't need std::abs here (even if the scale's negative) as we'll
1370 // be ^'ing Modulo with itself later.
1371 Modulo |= (uint64_t)DecompGEP1.VarIndices[i].Scale;
1374 // If the Value could change between cycles, then any reasoning about
1375 // the Value this cycle may not hold in the next cycle. We'll just
1376 // give up if we can't determine conditions that hold for every cycle:
1377 const Value *V = DecompGEP1.VarIndices[i].V;
1379 KnownBits Known = computeKnownBits(V, DL, 0, &AC, nullptr, DT);
1380 bool SignKnownZero = Known.isNonNegative();
1381 bool SignKnownOne = Known.isNegative();
1383 // Zero-extension widens the variable, and so forces the sign
1385 bool IsZExt = DecompGEP1.VarIndices[i].ZExtBits > 0 || isa<ZExtInst>(V);
1386 SignKnownZero |= IsZExt;
1387 SignKnownOne &= !IsZExt;
1389 // If the variable begins with a zero then we know it's
1390 // positive, regardless of whether the value is signed or
1392 int64_t Scale = DecompGEP1.VarIndices[i].Scale;
1394 (SignKnownZero && Scale >= 0) || (SignKnownOne && Scale < 0);
1398 Modulo = Modulo ^ (Modulo & (Modulo - 1));
1400 // We can compute the difference between the two addresses
1401 // mod Modulo. Check whether that difference guarantees that the
1402 // two locations do not alias.
1403 uint64_t ModOffset = (uint64_t)GEP1BaseOffset & (Modulo - 1);
1404 if (V1Size != MemoryLocation::UnknownSize &&
1405 V2Size != MemoryLocation::UnknownSize && ModOffset >= V2Size &&
1406 V1Size <= Modulo - ModOffset)
1409 // If we know all the variables are positive, then GEP1 >= GEP1BasePtr.
1410 // If GEP1BasePtr > V2 (GEP1BaseOffset > 0) then we know the pointers
1411 // don't alias if V2Size can fit in the gap between V2 and GEP1BasePtr.
1412 if (AllPositive && GEP1BaseOffset > 0 && V2Size <= (uint64_t)GEP1BaseOffset)
1415 if (constantOffsetHeuristic(DecompGEP1.VarIndices, V1Size, V2Size,
1416 GEP1BaseOffset, &AC, DT))
1420 // Statically, we can see that the base objects are the same, but the
1421 // pointers have dynamic offsets which we can't resolve. And none of our
1422 // little tricks above worked.
1426 static AliasResult MergeAliasResults(AliasResult A, AliasResult B) {
1427 // If the results agree, take it.
1430 // A mix of PartialAlias and MustAlias is PartialAlias.
1431 if ((A == PartialAlias && B == MustAlias) ||
1432 (B == PartialAlias && A == MustAlias))
1433 return PartialAlias;
1434 // Otherwise, we don't know anything.
1438 /// Provides a bunch of ad-hoc rules to disambiguate a Select instruction
1439 /// against another.
1440 AliasResult BasicAAResult::aliasSelect(const SelectInst *SI,
1441 LocationSize SISize,
1442 const AAMDNodes &SIAAInfo,
1443 const Value *V2, LocationSize V2Size,
1444 const AAMDNodes &V2AAInfo,
1445 const Value *UnderV2) {
1446 // If the values are Selects with the same condition, we can do a more precise
1447 // check: just check for aliases between the values on corresponding arms.
1448 if (const SelectInst *SI2 = dyn_cast<SelectInst>(V2))
1449 if (SI->getCondition() == SI2->getCondition()) {
1450 AliasResult Alias = aliasCheck(SI->getTrueValue(), SISize, SIAAInfo,
1451 SI2->getTrueValue(), V2Size, V2AAInfo);
1452 if (Alias == MayAlias)
1454 AliasResult ThisAlias =
1455 aliasCheck(SI->getFalseValue(), SISize, SIAAInfo,
1456 SI2->getFalseValue(), V2Size, V2AAInfo);
1457 return MergeAliasResults(ThisAlias, Alias);
1460 // If both arms of the Select node NoAlias or MustAlias V2, then returns
1461 // NoAlias / MustAlias. Otherwise, returns MayAlias.
1463 aliasCheck(V2, V2Size, V2AAInfo, SI->getTrueValue(),
1464 SISize, SIAAInfo, UnderV2);
1465 if (Alias == MayAlias)
1468 AliasResult ThisAlias =
1469 aliasCheck(V2, V2Size, V2AAInfo, SI->getFalseValue(), SISize, SIAAInfo,
1471 return MergeAliasResults(ThisAlias, Alias);
1474 /// Provide a bunch of ad-hoc rules to disambiguate a PHI instruction against
1476 AliasResult BasicAAResult::aliasPHI(const PHINode *PN, LocationSize PNSize,
1477 const AAMDNodes &PNAAInfo, const Value *V2,
1478 LocationSize V2Size,
1479 const AAMDNodes &V2AAInfo,
1480 const Value *UnderV2) {
1481 // Track phi nodes we have visited. We use this information when we determine
1482 // value equivalence.
1483 VisitedPhiBBs.insert(PN->getParent());
1485 // If the values are PHIs in the same block, we can do a more precise
1486 // as well as efficient check: just check for aliases between the values
1487 // on corresponding edges.
1488 if (const PHINode *PN2 = dyn_cast<PHINode>(V2))
1489 if (PN2->getParent() == PN->getParent()) {
1490 LocPair Locs(MemoryLocation(PN, PNSize, PNAAInfo),
1491 MemoryLocation(V2, V2Size, V2AAInfo));
1493 std::swap(Locs.first, Locs.second);
1494 // Analyse the PHIs' inputs under the assumption that the PHIs are
1496 // If the PHIs are May/MustAlias there must be (recursively) an input
1497 // operand from outside the PHIs' cycle that is MayAlias/MustAlias or
1498 // there must be an operation on the PHIs within the PHIs' value cycle
1499 // that causes a MayAlias.
1500 // Pretend the phis do not alias.
1501 AliasResult Alias = NoAlias;
1502 assert(AliasCache.count(Locs) &&
1503 "There must exist an entry for the phi node");
1504 AliasResult OrigAliasResult = AliasCache[Locs];
1505 AliasCache[Locs] = NoAlias;
1507 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1508 AliasResult ThisAlias =
1509 aliasCheck(PN->getIncomingValue(i), PNSize, PNAAInfo,
1510 PN2->getIncomingValueForBlock(PN->getIncomingBlock(i)),
1512 Alias = MergeAliasResults(ThisAlias, Alias);
1513 if (Alias == MayAlias)
1517 // Reset if speculation failed.
1518 if (Alias != NoAlias)
1519 AliasCache[Locs] = OrigAliasResult;
1524 SmallPtrSet<Value *, 4> UniqueSrc;
1525 SmallVector<Value *, 4> V1Srcs;
1526 bool isRecursive = false;
1527 for (Value *PV1 : PN->incoming_values()) {
1528 if (isa<PHINode>(PV1))
1529 // If any of the source itself is a PHI, return MayAlias conservatively
1530 // to avoid compile time explosion. The worst possible case is if both
1531 // sides are PHI nodes. In which case, this is O(m x n) time where 'm'
1532 // and 'n' are the number of PHI sources.
1535 if (EnableRecPhiAnalysis)
1536 if (GEPOperator *PV1GEP = dyn_cast<GEPOperator>(PV1)) {
1537 // Check whether the incoming value is a GEP that advances the pointer
1538 // result of this PHI node (e.g. in a loop). If this is the case, we
1539 // would recurse and always get a MayAlias. Handle this case specially
1541 if (PV1GEP->getPointerOperand() == PN && PV1GEP->getNumIndices() == 1 &&
1542 isa<ConstantInt>(PV1GEP->idx_begin())) {
1548 if (UniqueSrc.insert(PV1).second)
1549 V1Srcs.push_back(PV1);
1552 // If this PHI node is recursive, set the size of the accessed memory to
1553 // unknown to represent all the possible values the GEP could advance the
1556 PNSize = MemoryLocation::UnknownSize;
1559 aliasCheck(V2, V2Size, V2AAInfo, V1Srcs[0],
1560 PNSize, PNAAInfo, UnderV2);
1562 // Early exit if the check of the first PHI source against V2 is MayAlias.
1563 // Other results are not possible.
1564 if (Alias == MayAlias)
1567 // If all sources of the PHI node NoAlias or MustAlias V2, then returns
1568 // NoAlias / MustAlias. Otherwise, returns MayAlias.
1569 for (unsigned i = 1, e = V1Srcs.size(); i != e; ++i) {
1570 Value *V = V1Srcs[i];
1572 AliasResult ThisAlias =
1573 aliasCheck(V2, V2Size, V2AAInfo, V, PNSize, PNAAInfo, UnderV2);
1574 Alias = MergeAliasResults(ThisAlias, Alias);
1575 if (Alias == MayAlias)
1582 /// Provides a bunch of ad-hoc rules to disambiguate in common cases, such as
1583 /// array references.
1584 AliasResult BasicAAResult::aliasCheck(const Value *V1, LocationSize V1Size,
1585 AAMDNodes V1AAInfo, const Value *V2,
1586 LocationSize V2Size, AAMDNodes V2AAInfo,
1587 const Value *O1, const Value *O2) {
1588 // If either of the memory references is empty, it doesn't matter what the
1589 // pointer values are.
1590 if (V1Size == 0 || V2Size == 0)
1593 // Strip off any casts if they exist.
1594 V1 = V1->stripPointerCastsAndInvariantGroups();
1595 V2 = V2->stripPointerCastsAndInvariantGroups();
1597 // If V1 or V2 is undef, the result is NoAlias because we can always pick a
1598 // value for undef that aliases nothing in the program.
1599 if (isa<UndefValue>(V1) || isa<UndefValue>(V2))
1602 // Are we checking for alias of the same value?
1603 // Because we look 'through' phi nodes, we could look at "Value" pointers from
1604 // different iterations. We must therefore make sure that this is not the
1605 // case. The function isValueEqualInPotentialCycles ensures that this cannot
1606 // happen by looking at the visited phi nodes and making sure they cannot
1608 if (isValueEqualInPotentialCycles(V1, V2))
1611 if (!V1->getType()->isPointerTy() || !V2->getType()->isPointerTy())
1612 return NoAlias; // Scalars cannot alias each other
1614 // Figure out what objects these things are pointing to if we can.
1616 O1 = GetUnderlyingObject(V1, DL, MaxLookupSearchDepth);
1619 O2 = GetUnderlyingObject(V2, DL, MaxLookupSearchDepth);
1621 // Null values in the default address space don't point to any object, so they
1622 // don't alias any other pointer.
1623 if (const ConstantPointerNull *CPN = dyn_cast<ConstantPointerNull>(O1))
1624 if (CPN->getType()->getAddressSpace() == 0)
1626 if (const ConstantPointerNull *CPN = dyn_cast<ConstantPointerNull>(O2))
1627 if (CPN->getType()->getAddressSpace() == 0)
1631 // If V1/V2 point to two different objects, we know that we have no alias.
1632 if (isIdentifiedObject(O1) && isIdentifiedObject(O2))
1635 // Constant pointers can't alias with non-const isIdentifiedObject objects.
1636 if ((isa<Constant>(O1) && isIdentifiedObject(O2) && !isa<Constant>(O2)) ||
1637 (isa<Constant>(O2) && isIdentifiedObject(O1) && !isa<Constant>(O1)))
1640 // Function arguments can't alias with things that are known to be
1641 // unambigously identified at the function level.
1642 if ((isa<Argument>(O1) && isIdentifiedFunctionLocal(O2)) ||
1643 (isa<Argument>(O2) && isIdentifiedFunctionLocal(O1)))
1646 // If one pointer is the result of a call/invoke or load and the other is a
1647 // non-escaping local object within the same function, then we know the
1648 // object couldn't escape to a point where the call could return it.
1650 // Note that if the pointers are in different functions, there are a
1651 // variety of complications. A call with a nocapture argument may still
1652 // temporary store the nocapture argument's value in a temporary memory
1653 // location if that memory location doesn't escape. Or it may pass a
1654 // nocapture value to other functions as long as they don't capture it.
1655 if (isEscapeSource(O1) && isNonEscapingLocalObject(O2))
1657 if (isEscapeSource(O2) && isNonEscapingLocalObject(O1))
1661 // If the size of one access is larger than the entire object on the other
1662 // side, then we know such behavior is undefined and can assume no alias.
1663 if ((V1Size != MemoryLocation::UnknownSize &&
1664 isObjectSmallerThan(O2, V1Size, DL, TLI)) ||
1665 (V2Size != MemoryLocation::UnknownSize &&
1666 isObjectSmallerThan(O1, V2Size, DL, TLI)))
1669 // Check the cache before climbing up use-def chains. This also terminates
1670 // otherwise infinitely recursive queries.
1671 LocPair Locs(MemoryLocation(V1, V1Size, V1AAInfo),
1672 MemoryLocation(V2, V2Size, V2AAInfo));
1674 std::swap(Locs.first, Locs.second);
1675 std::pair<AliasCacheTy::iterator, bool> Pair =
1676 AliasCache.insert(std::make_pair(Locs, MayAlias));
1678 return Pair.first->second;
1680 // FIXME: This isn't aggressively handling alias(GEP, PHI) for example: if the
1681 // GEP can't simplify, we don't even look at the PHI cases.
1682 if (!isa<GEPOperator>(V1) && isa<GEPOperator>(V2)) {
1684 std::swap(V1Size, V2Size);
1686 std::swap(V1AAInfo, V2AAInfo);
1688 if (const GEPOperator *GV1 = dyn_cast<GEPOperator>(V1)) {
1689 AliasResult Result =
1690 aliasGEP(GV1, V1Size, V1AAInfo, V2, V2Size, V2AAInfo, O1, O2);
1691 if (Result != MayAlias)
1692 return AliasCache[Locs] = Result;
1695 if (isa<PHINode>(V2) && !isa<PHINode>(V1)) {
1698 std::swap(V1Size, V2Size);
1699 std::swap(V1AAInfo, V2AAInfo);
1701 if (const PHINode *PN = dyn_cast<PHINode>(V1)) {
1702 AliasResult Result = aliasPHI(PN, V1Size, V1AAInfo,
1703 V2, V2Size, V2AAInfo, O2);
1704 if (Result != MayAlias)
1705 return AliasCache[Locs] = Result;
1708 if (isa<SelectInst>(V2) && !isa<SelectInst>(V1)) {
1711 std::swap(V1Size, V2Size);
1712 std::swap(V1AAInfo, V2AAInfo);
1714 if (const SelectInst *S1 = dyn_cast<SelectInst>(V1)) {
1715 AliasResult Result =
1716 aliasSelect(S1, V1Size, V1AAInfo, V2, V2Size, V2AAInfo, O2);
1717 if (Result != MayAlias)
1718 return AliasCache[Locs] = Result;
1721 // If both pointers are pointing into the same object and one of them
1722 // accesses the entire object, then the accesses must overlap in some way.
1724 if (V1Size != MemoryLocation::UnknownSize &&
1725 V2Size != MemoryLocation::UnknownSize &&
1726 (isObjectSize(O1, V1Size, DL, TLI) ||
1727 isObjectSize(O2, V2Size, DL, TLI)))
1728 return AliasCache[Locs] = PartialAlias;
1730 // Recurse back into the best AA results we have, potentially with refined
1731 // memory locations. We have already ensured that BasicAA has a MayAlias
1732 // cache result for these, so any recursion back into BasicAA won't loop.
1733 AliasResult Result = getBestAAResults().alias(Locs.first, Locs.second);
1734 return AliasCache[Locs] = Result;
1737 /// Check whether two Values can be considered equivalent.
1739 /// In addition to pointer equivalence of \p V1 and \p V2 this checks whether
1740 /// they can not be part of a cycle in the value graph by looking at all
1741 /// visited phi nodes an making sure that the phis cannot reach the value. We
1742 /// have to do this because we are looking through phi nodes (That is we say
1743 /// noalias(V, phi(VA, VB)) if noalias(V, VA) and noalias(V, VB).
1744 bool BasicAAResult::isValueEqualInPotentialCycles(const Value *V,
1749 const Instruction *Inst = dyn_cast<Instruction>(V);
1753 if (VisitedPhiBBs.empty())
1756 if (VisitedPhiBBs.size() > MaxNumPhiBBsValueReachabilityCheck)
1759 // Make sure that the visited phis cannot reach the Value. This ensures that
1760 // the Values cannot come from different iterations of a potential cycle the
1761 // phi nodes could be involved in.
1762 for (auto *P : VisitedPhiBBs)
1763 if (isPotentiallyReachable(&P->front(), Inst, DT, LI))
1769 /// Computes the symbolic difference between two de-composed GEPs.
1771 /// Dest and Src are the variable indices from two decomposed GetElementPtr
1772 /// instructions GEP1 and GEP2 which have common base pointers.
1773 void BasicAAResult::GetIndexDifference(
1774 SmallVectorImpl<VariableGEPIndex> &Dest,
1775 const SmallVectorImpl<VariableGEPIndex> &Src) {
1779 for (unsigned i = 0, e = Src.size(); i != e; ++i) {
1780 const Value *V = Src[i].V;
1781 unsigned ZExtBits = Src[i].ZExtBits, SExtBits = Src[i].SExtBits;
1782 int64_t Scale = Src[i].Scale;
1784 // Find V in Dest. This is N^2, but pointer indices almost never have more
1785 // than a few variable indexes.
1786 for (unsigned j = 0, e = Dest.size(); j != e; ++j) {
1787 if (!isValueEqualInPotentialCycles(Dest[j].V, V) ||
1788 Dest[j].ZExtBits != ZExtBits || Dest[j].SExtBits != SExtBits)
1791 // If we found it, subtract off Scale V's from the entry in Dest. If it
1792 // goes to zero, remove the entry.
1793 if (Dest[j].Scale != Scale)
1794 Dest[j].Scale -= Scale;
1796 Dest.erase(Dest.begin() + j);
1801 // If we didn't consume this entry, add it to the end of the Dest list.
1803 VariableGEPIndex Entry = {V, ZExtBits, SExtBits, -Scale};
1804 Dest.push_back(Entry);
1809 bool BasicAAResult::constantOffsetHeuristic(
1810 const SmallVectorImpl<VariableGEPIndex> &VarIndices, LocationSize V1Size,
1811 LocationSize V2Size, int64_t BaseOffset, AssumptionCache *AC,
1812 DominatorTree *DT) {
1813 if (VarIndices.size() != 2 || V1Size == MemoryLocation::UnknownSize ||
1814 V2Size == MemoryLocation::UnknownSize)
1817 const VariableGEPIndex &Var0 = VarIndices[0], &Var1 = VarIndices[1];
1819 if (Var0.ZExtBits != Var1.ZExtBits || Var0.SExtBits != Var1.SExtBits ||
1820 Var0.Scale != -Var1.Scale)
1823 unsigned Width = Var1.V->getType()->getIntegerBitWidth();
1825 // We'll strip off the Extensions of Var0 and Var1 and do another round
1826 // of GetLinearExpression decomposition. In the example above, if Var0
1827 // is zext(%x + 1) we should get V1 == %x and V1Offset == 1.
1829 APInt V0Scale(Width, 0), V0Offset(Width, 0), V1Scale(Width, 0),
1831 bool NSW = true, NUW = true;
1832 unsigned V0ZExtBits = 0, V0SExtBits = 0, V1ZExtBits = 0, V1SExtBits = 0;
1833 const Value *V0 = GetLinearExpression(Var0.V, V0Scale, V0Offset, V0ZExtBits,
1834 V0SExtBits, DL, 0, AC, DT, NSW, NUW);
1837 const Value *V1 = GetLinearExpression(Var1.V, V1Scale, V1Offset, V1ZExtBits,
1838 V1SExtBits, DL, 0, AC, DT, NSW, NUW);
1840 if (V0Scale != V1Scale || V0ZExtBits != V1ZExtBits ||
1841 V0SExtBits != V1SExtBits || !isValueEqualInPotentialCycles(V0, V1))
1844 // We have a hit - Var0 and Var1 only differ by a constant offset!
1846 // If we've been sext'ed then zext'd the maximum difference between Var0 and
1847 // Var1 is possible to calculate, but we're just interested in the absolute
1848 // minimum difference between the two. The minimum distance may occur due to
1849 // wrapping; consider "add i3 %i, 5": if %i == 7 then 7 + 5 mod 8 == 4, and so
1850 // the minimum distance between %i and %i + 5 is 3.
1851 APInt MinDiff = V0Offset - V1Offset, Wrapped = -MinDiff;
1852 MinDiff = APIntOps::umin(MinDiff, Wrapped);
1853 uint64_t MinDiffBytes = MinDiff.getZExtValue() * std::abs(Var0.Scale);
1855 // We can't definitely say whether GEP1 is before or after V2 due to wrapping
1856 // arithmetic (i.e. for some values of GEP1 and V2 GEP1 < V2, and for other
1857 // values GEP1 > V2). We'll therefore only declare NoAlias if both V1Size and
1858 // V2Size can fit in the MinDiffBytes gap.
1859 return V1Size + std::abs(BaseOffset) <= MinDiffBytes &&
1860 V2Size + std::abs(BaseOffset) <= MinDiffBytes;
1863 //===----------------------------------------------------------------------===//
1864 // BasicAliasAnalysis Pass
1865 //===----------------------------------------------------------------------===//
1867 AnalysisKey BasicAA::Key;
1869 BasicAAResult BasicAA::run(Function &F, FunctionAnalysisManager &AM) {
1870 return BasicAAResult(F.getParent()->getDataLayout(),
1871 AM.getResult<TargetLibraryAnalysis>(F),
1872 AM.getResult<AssumptionAnalysis>(F),
1873 &AM.getResult<DominatorTreeAnalysis>(F),
1874 AM.getCachedResult<LoopAnalysis>(F));
1877 BasicAAWrapperPass::BasicAAWrapperPass() : FunctionPass(ID) {
1878 initializeBasicAAWrapperPassPass(*PassRegistry::getPassRegistry());
1881 char BasicAAWrapperPass::ID = 0;
1883 void BasicAAWrapperPass::anchor() {}
1885 INITIALIZE_PASS_BEGIN(BasicAAWrapperPass, "basicaa",
1886 "Basic Alias Analysis (stateless AA impl)", true, true)
1887 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
1888 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
1889 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
1890 INITIALIZE_PASS_END(BasicAAWrapperPass, "basicaa",
1891 "Basic Alias Analysis (stateless AA impl)", true, true)
1893 FunctionPass *llvm::createBasicAAWrapperPass() {
1894 return new BasicAAWrapperPass();
1897 bool BasicAAWrapperPass::runOnFunction(Function &F) {
1898 auto &ACT = getAnalysis<AssumptionCacheTracker>();
1899 auto &TLIWP = getAnalysis<TargetLibraryInfoWrapperPass>();
1900 auto &DTWP = getAnalysis<DominatorTreeWrapperPass>();
1901 auto *LIWP = getAnalysisIfAvailable<LoopInfoWrapperPass>();
1903 Result.reset(new BasicAAResult(F.getParent()->getDataLayout(), TLIWP.getTLI(),
1904 ACT.getAssumptionCache(F), &DTWP.getDomTree(),
1905 LIWP ? &LIWP->getLoopInfo() : nullptr));
1910 void BasicAAWrapperPass::getAnalysisUsage(AnalysisUsage &AU) const {
1911 AU.setPreservesAll();
1912 AU.addRequired<AssumptionCacheTracker>();
1913 AU.addRequired<DominatorTreeWrapperPass>();
1914 AU.addRequired<TargetLibraryInfoWrapperPass>();
1917 BasicAAResult llvm::createLegacyPMBasicAAResult(Pass &P, Function &F) {
1918 return BasicAAResult(
1919 F.getParent()->getDataLayout(),
1920 P.getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(),
1921 P.getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F));