1 //===- BasicAliasAnalysis.cpp - Local 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 default implementation of the Alias Analysis interface
11 // that simply implements a few identities (two different globals cannot alias,
12 // etc), but otherwise does no analysis.
14 //===----------------------------------------------------------------------===//
16 #include "llvm/Analysis/AliasAnalysis.h"
17 #include "llvm/Analysis/CaptureTracking.h"
18 #include "llvm/Analysis/MemoryBuiltins.h"
19 #include "llvm/Analysis/Passes.h"
20 #include "llvm/Constants.h"
21 #include "llvm/DerivedTypes.h"
22 #include "llvm/Function.h"
23 #include "llvm/GlobalAlias.h"
24 #include "llvm/GlobalVariable.h"
25 #include "llvm/Instructions.h"
26 #include "llvm/IntrinsicInst.h"
27 #include "llvm/Operator.h"
28 #include "llvm/Pass.h"
29 #include "llvm/Target/TargetData.h"
30 #include "llvm/ADT/SmallSet.h"
31 #include "llvm/ADT/SmallVector.h"
32 #include "llvm/ADT/STLExtras.h"
33 #include "llvm/Support/ErrorHandling.h"
34 #include "llvm/Support/GetElementPtrTypeIterator.h"
38 //===----------------------------------------------------------------------===//
40 //===----------------------------------------------------------------------===//
42 static const Value *GetGEPOperands(const GEPOperator *V,
43 SmallVector<Value*, 16> &GEPOps) {
44 assert(GEPOps.empty() && "Expect empty list to populate!");
45 GEPOps.insert(GEPOps.end(), V->op_begin()+1, V->op_end());
47 // Accumulate all of the chained indexes into the operand array.
48 Value *BasePtr = V->getOperand(0);
50 V = dyn_cast<GEPOperator>(BasePtr);
51 if (V == 0) return BasePtr;
53 // Don't handle folding arbitrary pointer offsets yet.
54 if (!isa<Constant>(GEPOps[0]) || !cast<Constant>(GEPOps[0])->isNullValue())
57 GEPOps.erase(GEPOps.begin()); // Drop the zero index
58 GEPOps.insert(GEPOps.begin(), V->op_begin()+1, V->op_end());
62 /// isKnownNonNull - Return true if we know that the specified value is never
64 static bool isKnownNonNull(const Value *V) {
65 // Alloca never returns null, malloc might.
66 if (isa<AllocaInst>(V)) return true;
68 // A byval argument is never null.
69 if (const Argument *A = dyn_cast<Argument>(V))
70 return A->hasByValAttr();
72 // Global values are not null unless extern weak.
73 if (const GlobalValue *GV = dyn_cast<GlobalValue>(V))
74 return !GV->hasExternalWeakLinkage();
78 /// isNonEscapingLocalObject - Return true if the pointer is to a function-local
79 /// object that never escapes from the function.
80 static bool isNonEscapingLocalObject(const Value *V) {
81 // If this is a local allocation, check to see if it escapes.
82 if (isa<AllocaInst>(V) || isNoAliasCall(V))
83 // Set StoreCaptures to True so that we can assume in our callers that the
84 // pointer is not the result of a load instruction. Currently
85 // PointerMayBeCaptured doesn't have any special analysis for the
86 // StoreCaptures=false case; if it did, our callers could be refined to be
88 return !PointerMayBeCaptured(V, false, /*StoreCaptures=*/true);
90 // If this is an argument that corresponds to a byval or noalias argument,
91 // then it has not escaped before entering the function. Check if it escapes
92 // inside the function.
93 if (const Argument *A = dyn_cast<Argument>(V))
94 if (A->hasByValAttr() || A->hasNoAliasAttr()) {
95 // Don't bother analyzing arguments already known not to escape.
96 if (A->hasNoCaptureAttr())
98 return !PointerMayBeCaptured(V, false, /*StoreCaptures=*/true);
104 /// isObjectSmallerThan - Return true if we can prove that the object specified
105 /// by V is smaller than Size.
106 static bool isObjectSmallerThan(const Value *V, unsigned Size,
107 const TargetData &TD) {
108 const Type *AccessTy;
109 if (const GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) {
110 AccessTy = GV->getType()->getElementType();
111 } else if (const AllocaInst *AI = dyn_cast<AllocaInst>(V)) {
112 if (!AI->isArrayAllocation())
113 AccessTy = AI->getType()->getElementType();
116 } else if (const CallInst* CI = extractMallocCall(V)) {
117 if (!isArrayMalloc(V, &TD))
118 // The size is the argument to the malloc call.
119 if (const ConstantInt* C = dyn_cast<ConstantInt>(CI->getOperand(1)))
120 return (C->getZExtValue() < Size);
122 } else if (const Argument *A = dyn_cast<Argument>(V)) {
123 if (A->hasByValAttr())
124 AccessTy = cast<PointerType>(A->getType())->getElementType();
131 if (AccessTy->isSized())
132 return TD.getTypeAllocSize(AccessTy) < Size;
136 //===----------------------------------------------------------------------===//
138 //===----------------------------------------------------------------------===//
141 /// NoAA - This class implements the -no-aa pass, which always returns "I
142 /// don't know" for alias queries. NoAA is unlike other alias analysis
143 /// implementations, in that it does not chain to a previous analysis. As
144 /// such it doesn't follow many of the rules that other alias analyses must.
146 struct NoAA : public ImmutablePass, public AliasAnalysis {
147 static char ID; // Class identification, replacement for typeinfo
148 NoAA() : ImmutablePass(&ID) {}
149 explicit NoAA(void *PID) : ImmutablePass(PID) { }
151 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
154 virtual void initializePass() {
155 TD = getAnalysisIfAvailable<TargetData>();
158 virtual AliasResult alias(const Value *V1, unsigned V1Size,
159 const Value *V2, unsigned V2Size) {
163 virtual void getArgumentAccesses(Function *F, CallSite CS,
164 std::vector<PointerAccessInfo> &Info) {
165 llvm_unreachable("This method may not be called on this function!");
168 virtual bool pointsToConstantMemory(const Value *P) { return false; }
169 virtual ModRefResult getModRefInfo(CallSite CS, Value *P, unsigned Size) {
172 virtual ModRefResult getModRefInfo(CallSite CS1, CallSite CS2) {
176 virtual void deleteValue(Value *V) {}
177 virtual void copyValue(Value *From, Value *To) {}
179 } // End of anonymous namespace
181 // Register this pass...
183 static RegisterPass<NoAA>
184 U("no-aa", "No Alias Analysis (always returns 'may' alias)", true, true);
186 // Declare that we implement the AliasAnalysis interface
187 static RegisterAnalysisGroup<AliasAnalysis> V(U);
189 ImmutablePass *llvm::createNoAAPass() { return new NoAA(); }
191 //===----------------------------------------------------------------------===//
193 //===----------------------------------------------------------------------===//
196 /// BasicAliasAnalysis - This is the default alias analysis implementation.
197 /// Because it doesn't chain to a previous alias analysis (like -no-aa), it
198 /// derives from the NoAA class.
199 struct BasicAliasAnalysis : public NoAA {
200 static char ID; // Class identification, replacement for typeinfo
201 BasicAliasAnalysis() : NoAA(&ID) {}
202 AliasResult alias(const Value *V1, unsigned V1Size,
203 const Value *V2, unsigned V2Size) {
204 assert(VisitedPHIs.empty() && "VisitedPHIs must be cleared after use!");
205 AliasResult Alias = aliasCheck(V1, V1Size, V2, V2Size);
210 ModRefResult getModRefInfo(CallSite CS, Value *P, unsigned Size);
211 ModRefResult getModRefInfo(CallSite CS1, CallSite CS2);
213 /// pointsToConstantMemory - Chase pointers until we find a (constant
215 bool pointsToConstantMemory(const Value *P);
218 // VisitedPHIs - Track PHI nodes visited by a aliasCheck() call.
219 SmallPtrSet<const Value*, 16> VisitedPHIs;
221 // aliasGEP - Provide a bunch of ad-hoc rules to disambiguate a GEP
222 // instruction against another.
223 AliasResult aliasGEP(const GEPOperator *V1, unsigned V1Size,
224 const Value *V2, unsigned V2Size,
225 const Value *UnderlyingV1, const Value *UnderlyingV2);
227 // aliasPHI - Provide a bunch of ad-hoc rules to disambiguate a PHI
228 // instruction against another.
229 AliasResult aliasPHI(const PHINode *PN, unsigned PNSize,
230 const Value *V2, unsigned V2Size);
232 /// aliasSelect - Disambiguate a Select instruction against another value.
233 AliasResult aliasSelect(const SelectInst *SI, unsigned SISize,
234 const Value *V2, unsigned V2Size);
236 AliasResult aliasCheck(const Value *V1, unsigned V1Size,
237 const Value *V2, unsigned V2Size);
239 // CheckGEPInstructions - Check two GEP instructions with known
240 // must-aliasing base pointers. This checks to see if the index expressions
241 // preclude the pointers from aliasing.
243 CheckGEPInstructions(const Type* BasePtr1Ty,
244 Value **GEP1Ops, unsigned NumGEP1Ops, unsigned G1Size,
245 const Type *BasePtr2Ty,
246 Value **GEP2Ops, unsigned NumGEP2Ops, unsigned G2Size);
248 } // End of anonymous namespace
250 // Register this pass...
251 char BasicAliasAnalysis::ID = 0;
252 static RegisterPass<BasicAliasAnalysis>
253 X("basicaa", "Basic Alias Analysis (default AA impl)", false, true);
255 // Declare that we implement the AliasAnalysis interface
256 static RegisterAnalysisGroup<AliasAnalysis, true> Y(X);
258 ImmutablePass *llvm::createBasicAliasAnalysisPass() {
259 return new BasicAliasAnalysis();
263 /// pointsToConstantMemory - Chase pointers until we find a (constant
265 bool BasicAliasAnalysis::pointsToConstantMemory(const Value *P) {
266 if (const GlobalVariable *GV =
267 dyn_cast<GlobalVariable>(P->getUnderlyingObject()))
268 // Note: this doesn't require GV to be "ODR" because it isn't legal for a
269 // global to be marked constant in some modules and non-constant in others.
270 // GV may even be a declaration, not a definition.
271 return GV->isConstant();
276 /// getModRefInfo - Check to see if the specified callsite can clobber the
277 /// specified memory object. Since we only look at local properties of this
278 /// function, we really can't say much about this query. We do, however, use
279 /// simple "address taken" analysis on local objects.
280 AliasAnalysis::ModRefResult
281 BasicAliasAnalysis::getModRefInfo(CallSite CS, Value *P, unsigned Size) {
282 const Value *Object = P->getUnderlyingObject();
284 // If this is a tail call and P points to a stack location, we know that
285 // the tail call cannot access or modify the local stack.
286 // We cannot exclude byval arguments here; these belong to the caller of
287 // the current function not to the current function, and a tail callee
288 // may reference them.
289 if (isa<AllocaInst>(Object))
290 if (CallInst *CI = dyn_cast<CallInst>(CS.getInstruction()))
291 if (CI->isTailCall())
294 // If the pointer is to a locally allocated object that does not escape,
295 // then the call can not mod/ref the pointer unless the call takes the pointer
296 // as an argument, and itself doesn't capture it.
297 if (!isa<Constant>(Object) && CS.getInstruction() != Object &&
298 isNonEscapingLocalObject(Object)) {
299 bool PassedAsArg = false;
301 for (CallSite::arg_iterator CI = CS.arg_begin(), CE = CS.arg_end();
302 CI != CE; ++CI, ++ArgNo) {
303 // Only look at the no-capture pointer arguments.
304 if (!isa<PointerType>((*CI)->getType()) ||
305 !CS.paramHasAttr(ArgNo+1, Attribute::NoCapture))
308 // If this is a no-capture pointer argument, see if we can tell that it
309 // is impossible to alias the pointer we're checking. If not, we have to
310 // assume that the call could touch the pointer, even though it doesn't
312 if (!isNoAlias(cast<Value>(CI), ~0U, P, ~0U)) {
322 // Finally, handle specific knowledge of intrinsics.
323 IntrinsicInst *II = dyn_cast<IntrinsicInst>(CS.getInstruction());
325 return AliasAnalysis::getModRefInfo(CS, P, Size);
327 switch (II->getIntrinsicID()) {
329 case Intrinsic::memcpy:
330 case Intrinsic::memmove: {
332 if (ConstantInt *LenCI = dyn_cast<ConstantInt>(II->getOperand(3)))
333 Len = LenCI->getZExtValue();
334 Value *Dest = II->getOperand(1);
335 Value *Src = II->getOperand(2);
336 if (isNoAlias(Dest, Len, P, Size)) {
337 if (isNoAlias(Src, Len, P, Size))
343 case Intrinsic::memset:
344 // Since memset is 'accesses arguments' only, the AliasAnalysis base class
345 // will handle it for the variable length case.
346 if (ConstantInt *LenCI = dyn_cast<ConstantInt>(II->getOperand(3))) {
347 unsigned Len = LenCI->getZExtValue();
348 Value *Dest = II->getOperand(1);
349 if (isNoAlias(Dest, Len, P, Size))
353 case Intrinsic::atomic_cmp_swap:
354 case Intrinsic::atomic_swap:
355 case Intrinsic::atomic_load_add:
356 case Intrinsic::atomic_load_sub:
357 case Intrinsic::atomic_load_and:
358 case Intrinsic::atomic_load_nand:
359 case Intrinsic::atomic_load_or:
360 case Intrinsic::atomic_load_xor:
361 case Intrinsic::atomic_load_max:
362 case Intrinsic::atomic_load_min:
363 case Intrinsic::atomic_load_umax:
364 case Intrinsic::atomic_load_umin:
366 Value *Op1 = II->getOperand(1);
367 unsigned Op1Size = TD->getTypeStoreSize(Op1->getType());
368 if (isNoAlias(Op1, Op1Size, P, Size))
372 case Intrinsic::lifetime_start:
373 case Intrinsic::lifetime_end:
374 case Intrinsic::invariant_start: {
375 unsigned PtrSize = cast<ConstantInt>(II->getOperand(1))->getZExtValue();
376 if (isNoAlias(II->getOperand(2), PtrSize, P, Size))
380 case Intrinsic::invariant_end: {
381 unsigned PtrSize = cast<ConstantInt>(II->getOperand(2))->getZExtValue();
382 if (isNoAlias(II->getOperand(3), PtrSize, P, Size))
388 // The AliasAnalysis base class has some smarts, lets use them.
389 return AliasAnalysis::getModRefInfo(CS, P, Size);
393 AliasAnalysis::ModRefResult
394 BasicAliasAnalysis::getModRefInfo(CallSite CS1, CallSite CS2) {
395 // If CS1 or CS2 are readnone, they don't interact.
396 ModRefBehavior CS1B = AliasAnalysis::getModRefBehavior(CS1);
397 if (CS1B == DoesNotAccessMemory) return NoModRef;
399 ModRefBehavior CS2B = AliasAnalysis::getModRefBehavior(CS2);
400 if (CS2B == DoesNotAccessMemory) return NoModRef;
402 // If they both only read from memory, just return ref.
403 if (CS1B == OnlyReadsMemory && CS2B == OnlyReadsMemory)
406 // Otherwise, fall back to NoAA (mod+ref).
407 return NoAA::getModRefInfo(CS1, CS2);
410 /// DecomposeGEPExpression - If V is a symbolic pointer expression, decompose it
411 /// into a base pointer with a constant offset and a number of scaled symbolic
414 /// When TargetData is around, this function is capable of analyzing everything
415 /// that Value::getUnderlyingObject() can look through. When not, it just looks
416 /// through pointer casts.
418 /// FIXME: Move this out to ValueTracking.cpp
420 static const Value *DecomposeGEPExpression(const Value *V, int64_t &BaseOffs,
421 SmallVectorImpl<std::pair<const Value*, uint64_t> > &VarIndices,
422 const TargetData *TD) {
423 // FIXME: Should limit depth like getUnderlyingObject?
426 // See if this is a bitcast or GEP.
427 const Operator *Op = dyn_cast<Operator>(V);
429 // The only non-operator case we can handle are GlobalAliases.
430 if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
431 if (!GA->mayBeOverridden()) {
432 V = GA->getAliasee();
439 if (Op->getOpcode() == Instruction::BitCast) {
440 V = Op->getOperand(0);
444 const GEPOperator *GEPOp = dyn_cast<GEPOperator>(Op);
448 // Don't attempt to analyze GEPs over unsized objects.
449 if (!cast<PointerType>(GEPOp->getOperand(0)->getType())
450 ->getElementType()->isSized())
453 // If we are lacking TargetData information, we can't compute the offets of
454 // elements computed by GEPs. However, we can handle bitcast equivalent
457 if (!GEPOp->hasAllZeroIndices())
459 V = GEPOp->getOperand(0);
463 // Walk the indices of the GEP, accumulating them into BaseOff/VarIndices.
464 gep_type_iterator GTI = gep_type_begin(GEPOp);
465 for (User::const_op_iterator I = next(GEPOp->op_begin()),
466 E = GEPOp->op_end(); I != E; ++I) {
468 // Compute the (potentially symbolic) offset in bytes for this index.
469 if (const StructType *STy = dyn_cast<StructType>(*GTI++)) {
470 // For a struct, add the member offset.
471 unsigned FieldNo = cast<ConstantInt>(Index)->getZExtValue();
472 if (FieldNo == 0) continue;
474 BaseOffs += TD->getStructLayout(STy)->getElementOffset(FieldNo);
478 // For an array/pointer, add the element offset, explicitly scaled.
479 if (ConstantInt *CIdx = dyn_cast<ConstantInt>(Index)) {
480 if (CIdx->isZero()) continue;
481 BaseOffs += TD->getTypeAllocSize(*GTI)*CIdx->getSExtValue();
485 // TODO: Could handle linear expressions here like A[X+1], also A[X*4|1].
486 uint64_t Scale = TD->getTypeAllocSize(*GTI);
488 // If we already had an occurrance of this index variable, merge this
489 // scale into it. For example, we want to handle:
490 // A[x][x] -> x*16 + x*4 -> x*20
491 for (unsigned i = 0, e = VarIndices.size(); i != e; ++i) {
492 if (VarIndices[i].first == Index) {
493 Scale += VarIndices[i].second;
494 VarIndices.erase(VarIndices.begin()+i);
499 // Make sure that we have a scale that makes sense for this target's
501 if (unsigned ShiftBits = 64-TD->getPointerSizeInBits()) {
507 VarIndices.push_back(std::make_pair(Index, Scale));
510 // Analyze the base pointer next.
511 V = GEPOp->getOperand(0);
516 /// aliasGEP - Provide a bunch of ad-hoc rules to disambiguate a GEP instruction
517 /// against another pointer. We know that V1 is a GEP, but we don't know
518 /// anything about V2. UnderlyingV1 is GEP1->getUnderlyingObject(),
519 /// UnderlyingV2 is the same for V2.
521 AliasAnalysis::AliasResult
522 BasicAliasAnalysis::aliasGEP(const GEPOperator *GEP1, unsigned V1Size,
523 const Value *V2, unsigned V2Size,
524 const Value *UnderlyingV1,
525 const Value *UnderlyingV2) {
526 // If we have two gep instructions with must-alias'ing base pointers, figure
527 // out if the indexes to the GEP tell us anything about the derived pointer.
528 // Note that we also handle chains of getelementptr instructions as well as
529 // constant expression getelementptrs here.
531 if (const GEPOperator *GEP2 = dyn_cast<GEPOperator>(V2)) {
532 // If V1 and V2 are identical GEPs, just recurse down on both of them.
533 // This allows us to analyze things like:
534 // P = gep A, 0, i, 1
535 // Q = gep B, 0, i, 1
536 // by just analyzing A and B. This is even safe for variable indices.
537 if (GEP1->getType() == GEP2->getType() &&
538 GEP1->getNumOperands() == GEP2->getNumOperands() &&
539 GEP1->getOperand(0)->getType() == GEP2->getOperand(0)->getType() &&
540 // All operands are the same, ignoring the base.
541 std::equal(GEP1->op_begin()+1, GEP1->op_end(), GEP2->op_begin()+1))
542 return aliasCheck(GEP1->getOperand(0), V1Size,
543 GEP2->getOperand(0), V2Size);
545 // Drill down into the first non-gep value, to test for must-aliasing of
546 // the base pointers.
547 while (isa<GEPOperator>(GEP1->getOperand(0)) &&
548 GEP1->getOperand(1) ==
549 Constant::getNullValue(GEP1->getOperand(1)->getType()))
550 GEP1 = cast<GEPOperator>(GEP1->getOperand(0));
551 const Value *BasePtr1 = GEP1->getOperand(0);
553 while (isa<GEPOperator>(GEP2->getOperand(0)) &&
554 GEP2->getOperand(1) ==
555 Constant::getNullValue(GEP2->getOperand(1)->getType()))
556 GEP2 = cast<GEPOperator>(GEP2->getOperand(0));
557 const Value *BasePtr2 = GEP2->getOperand(0);
559 // Do the base pointers alias?
560 AliasResult BaseAlias = aliasCheck(BasePtr1, ~0U, BasePtr2, ~0U);
561 if (BaseAlias == NoAlias) return NoAlias;
562 if (BaseAlias == MustAlias) {
563 // If the base pointers alias each other exactly, check to see if we can
564 // figure out anything about the resultant pointers, to try to prove
567 // Collect all of the chained GEP operands together into one simple place
568 SmallVector<Value*, 16> GEP1Ops, GEP2Ops;
569 BasePtr1 = GetGEPOperands(GEP1, GEP1Ops);
570 BasePtr2 = GetGEPOperands(GEP2, GEP2Ops);
572 // If GetGEPOperands were able to fold to the same must-aliased pointer,
573 // do the comparison.
574 if (BasePtr1 == BasePtr2) {
576 CheckGEPInstructions(BasePtr1->getType(),
577 &GEP1Ops[0], GEP1Ops.size(), V1Size,
579 &GEP2Ops[0], GEP2Ops.size(), V2Size);
580 if (GAlias != MayAlias)
586 // Check to see if these two pointers are related by a getelementptr
587 // instruction. If one pointer is a GEP with a non-zero index of the other
588 // pointer, we know they cannot alias.
590 // FIXME: The check below only looks at the size of one of the pointers, not
591 // both, this may cause us to miss things.
592 if (V1Size == ~0U || V2Size == ~0U)
595 AliasResult R = aliasCheck(UnderlyingV1, ~0U, V2, V2Size);
597 // If V2 may alias GEP base pointer, conservatively returns MayAlias.
598 // If V2 is known not to alias GEP base pointer, then the two values
599 // cannot alias per GEP semantics: "A pointer value formed from a
600 // getelementptr instruction is associated with the addresses associated
601 // with the first operand of the getelementptr".
604 int64_t GEP1BaseOffset;
605 SmallVector<std::pair<const Value*, uint64_t>, 4> VariableIndices;
606 const Value *GEP1BasePtr =
607 DecomposeGEPExpression(GEP1, GEP1BaseOffset, VariableIndices, TD);
609 // If DecomposeGEPExpression isn't able to look all the way through the
610 // addressing operation, we must not have TD and this is too complex for us
611 // to handle without it.
612 if (GEP1BasePtr != UnderlyingV1) {
614 "DecomposeGEPExpression and getUnderlyingObject disagree!");
618 // If we have getelementptr <ptr>, 0, 0, 0, 0, ... and V2 must aliases
619 // the ptr, the end result is a must alias also.
620 if (GEP1BaseOffset == 0 && VariableIndices.empty())
623 // If we have a known constant offset, see if this offset is larger than the
624 // access size being queried. If so, and if no variable indices can remove
625 // pieces of this constant, then we know we have a no-alias. For example,
628 // In order to handle cases like &A[100][i] where i is an out of range
629 // subscript, we have to ignore all constant offset pieces that are a multiple
630 // of a scaled index. Do this by removing constant offsets that are a
631 // multiple of any of our variable indices. This allows us to transform
632 // things like &A[i][1] because i has a stride of (e.g.) 8 bytes but the 1
633 // provides an offset of 4 bytes (assuming a <= 4 byte access).
634 for (unsigned i = 0, e = VariableIndices.size(); i != e && GEP1BaseOffset;++i)
635 if (int64_t RemovedOffset = GEP1BaseOffset/VariableIndices[i].second)
636 GEP1BaseOffset -= RemovedOffset*VariableIndices[i].second;
638 // If our known offset is bigger than the access size, we know we don't have
640 if (GEP1BaseOffset) {
641 if (GEP1BaseOffset >= (int64_t)V2Size ||
642 GEP1BaseOffset <= -(int64_t)V1Size)
649 /// aliasSelect - Provide a bunch of ad-hoc rules to disambiguate a Select
650 /// instruction against another.
651 AliasAnalysis::AliasResult
652 BasicAliasAnalysis::aliasSelect(const SelectInst *SI, unsigned SISize,
653 const Value *V2, unsigned V2Size) {
654 // If the values are Selects with the same condition, we can do a more precise
655 // check: just check for aliases between the values on corresponding arms.
656 if (const SelectInst *SI2 = dyn_cast<SelectInst>(V2))
657 if (SI->getCondition() == SI2->getCondition()) {
659 aliasCheck(SI->getTrueValue(), SISize,
660 SI2->getTrueValue(), V2Size);
661 if (Alias == MayAlias)
663 AliasResult ThisAlias =
664 aliasCheck(SI->getFalseValue(), SISize,
665 SI2->getFalseValue(), V2Size);
666 if (ThisAlias != Alias)
671 // If both arms of the Select node NoAlias or MustAlias V2, then returns
672 // NoAlias / MustAlias. Otherwise, returns MayAlias.
674 aliasCheck(SI->getTrueValue(), SISize, V2, V2Size);
675 if (Alias == MayAlias)
677 AliasResult ThisAlias =
678 aliasCheck(SI->getFalseValue(), SISize, V2, V2Size);
679 if (ThisAlias != Alias)
684 // aliasPHI - Provide a bunch of ad-hoc rules to disambiguate a PHI instruction
686 AliasAnalysis::AliasResult
687 BasicAliasAnalysis::aliasPHI(const PHINode *PN, unsigned PNSize,
688 const Value *V2, unsigned V2Size) {
689 // The PHI node has already been visited, avoid recursion any further.
690 if (!VisitedPHIs.insert(PN))
693 // If the values are PHIs in the same block, we can do a more precise
694 // as well as efficient check: just check for aliases between the values
695 // on corresponding edges.
696 if (const PHINode *PN2 = dyn_cast<PHINode>(V2))
697 if (PN2->getParent() == PN->getParent()) {
699 aliasCheck(PN->getIncomingValue(0), PNSize,
700 PN2->getIncomingValueForBlock(PN->getIncomingBlock(0)),
702 if (Alias == MayAlias)
704 for (unsigned i = 1, e = PN->getNumIncomingValues(); i != e; ++i) {
705 AliasResult ThisAlias =
706 aliasCheck(PN->getIncomingValue(i), PNSize,
707 PN2->getIncomingValueForBlock(PN->getIncomingBlock(i)),
709 if (ThisAlias != Alias)
715 SmallPtrSet<Value*, 4> UniqueSrc;
716 SmallVector<Value*, 4> V1Srcs;
717 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
718 Value *PV1 = PN->getIncomingValue(i);
719 if (isa<PHINode>(PV1))
720 // If any of the source itself is a PHI, return MayAlias conservatively
721 // to avoid compile time explosion. The worst possible case is if both
722 // sides are PHI nodes. In which case, this is O(m x n) time where 'm'
723 // and 'n' are the number of PHI sources.
725 if (UniqueSrc.insert(PV1))
726 V1Srcs.push_back(PV1);
729 AliasResult Alias = aliasCheck(V2, V2Size, V1Srcs[0], PNSize);
730 // Early exit if the check of the first PHI source against V2 is MayAlias.
731 // Other results are not possible.
732 if (Alias == MayAlias)
735 // If all sources of the PHI node NoAlias or MustAlias V2, then returns
736 // NoAlias / MustAlias. Otherwise, returns MayAlias.
737 for (unsigned i = 1, e = V1Srcs.size(); i != e; ++i) {
738 Value *V = V1Srcs[i];
740 // If V2 is a PHI, the recursive case will have been caught in the
741 // above aliasCheck call, so these subsequent calls to aliasCheck
742 // don't need to assume that V2 is being visited recursively.
743 VisitedPHIs.erase(V2);
745 AliasResult ThisAlias = aliasCheck(V2, V2Size, V, PNSize);
746 if (ThisAlias != Alias || ThisAlias == MayAlias)
753 // aliasCheck - Provide a bunch of ad-hoc rules to disambiguate in common cases,
754 // such as array references.
756 AliasAnalysis::AliasResult
757 BasicAliasAnalysis::aliasCheck(const Value *V1, unsigned V1Size,
758 const Value *V2, unsigned V2Size) {
759 // Strip off any casts if they exist.
760 V1 = V1->stripPointerCasts();
761 V2 = V2->stripPointerCasts();
763 // Are we checking for alias of the same value?
764 if (V1 == V2) return MustAlias;
766 if (!isa<PointerType>(V1->getType()) || !isa<PointerType>(V2->getType()))
767 return NoAlias; // Scalars cannot alias each other
769 // Figure out what objects these things are pointing to if we can.
770 const Value *O1 = V1->getUnderlyingObject();
771 const Value *O2 = V2->getUnderlyingObject();
773 // Null values in the default address space don't point to any object, so they
774 // don't alias any other pointer.
775 if (const ConstantPointerNull *CPN = dyn_cast<ConstantPointerNull>(O1))
776 if (CPN->getType()->getAddressSpace() == 0)
778 if (const ConstantPointerNull *CPN = dyn_cast<ConstantPointerNull>(O2))
779 if (CPN->getType()->getAddressSpace() == 0)
783 // If V1/V2 point to two different objects we know that we have no alias.
784 if (isIdentifiedObject(O1) && isIdentifiedObject(O2))
787 // Constant pointers can't alias with non-const isIdentifiedObject objects.
788 if ((isa<Constant>(O1) && isIdentifiedObject(O2) && !isa<Constant>(O2)) ||
789 (isa<Constant>(O2) && isIdentifiedObject(O1) && !isa<Constant>(O1)))
792 // Arguments can't alias with local allocations or noalias calls.
793 if ((isa<Argument>(O1) && (isa<AllocaInst>(O2) || isNoAliasCall(O2))) ||
794 (isa<Argument>(O2) && (isa<AllocaInst>(O1) || isNoAliasCall(O1))))
797 // Most objects can't alias null.
798 if ((isa<ConstantPointerNull>(V2) && isKnownNonNull(O1)) ||
799 (isa<ConstantPointerNull>(V1) && isKnownNonNull(O2)))
803 // If the size of one access is larger than the entire object on the other
804 // side, then we know such behavior is undefined and can assume no alias.
806 if ((V1Size != ~0U && isObjectSmallerThan(O2, V1Size, *TD)) ||
807 (V2Size != ~0U && isObjectSmallerThan(O1, V2Size, *TD)))
810 // If one pointer is the result of a call/invoke or load and the other is a
811 // non-escaping local object, then we know the object couldn't escape to a
812 // point where the call could return it. The load case works because
813 // isNonEscapingLocalObject considers all stores to be escapes (it
814 // passes true for the StoreCaptures argument to PointerMayBeCaptured).
816 if ((isa<CallInst>(O1) || isa<InvokeInst>(O1) || isa<LoadInst>(O1) ||
817 isa<Argument>(O1)) &&
818 isNonEscapingLocalObject(O2))
820 if ((isa<CallInst>(O2) || isa<InvokeInst>(O2) || isa<LoadInst>(O2) ||
821 isa<Argument>(O2)) &&
822 isNonEscapingLocalObject(O1))
826 // FIXME: This isn't aggressively handling alias(GEP, PHI) for example: if the
827 // GEP can't simplify, we don't even look at the PHI cases.
828 if (!isa<GEPOperator>(V1) && isa<GEPOperator>(V2)) {
830 std::swap(V1Size, V2Size);
833 if (const GEPOperator *GV1 = dyn_cast<GEPOperator>(V1))
834 return aliasGEP(GV1, V1Size, V2, V2Size, O1, O2);
836 if (isa<PHINode>(V2) && !isa<PHINode>(V1)) {
838 std::swap(V1Size, V2Size);
840 if (const PHINode *PN = dyn_cast<PHINode>(V1))
841 return aliasPHI(PN, V1Size, V2, V2Size);
843 if (isa<SelectInst>(V2) && !isa<SelectInst>(V1)) {
845 std::swap(V1Size, V2Size);
847 if (const SelectInst *S1 = dyn_cast<SelectInst>(V1))
848 return aliasSelect(S1, V1Size, V2, V2Size);
853 // This function is used to determine if the indices of two GEP instructions are
854 // equal. V1 and V2 are the indices.
855 static bool IndexOperandsEqual(Value *V1, Value *V2) {
856 if (V1->getType() == V2->getType())
858 if (Constant *C1 = dyn_cast<Constant>(V1))
859 if (Constant *C2 = dyn_cast<Constant>(V2)) {
860 // Sign extend the constants to long types, if necessary
861 if (C1->getType() != Type::getInt64Ty(C1->getContext()))
862 C1 = ConstantExpr::getSExt(C1, Type::getInt64Ty(C1->getContext()));
863 if (C2->getType() != Type::getInt64Ty(C1->getContext()))
864 C2 = ConstantExpr::getSExt(C2, Type::getInt64Ty(C1->getContext()));
870 /// CheckGEPInstructions - Check two GEP instructions with known must-aliasing
871 /// base pointers. This checks to see if the index expressions preclude the
872 /// pointers from aliasing.
873 AliasAnalysis::AliasResult
874 BasicAliasAnalysis::CheckGEPInstructions(
875 const Type* BasePtr1Ty, Value **GEP1Ops, unsigned NumGEP1Ops, unsigned G1S,
876 const Type *BasePtr2Ty, Value **GEP2Ops, unsigned NumGEP2Ops, unsigned G2S) {
877 // We currently can't handle the case when the base pointers have different
878 // primitive types. Since this is uncommon anyway, we are happy being
879 // extremely conservative.
880 if (BasePtr1Ty != BasePtr2Ty)
883 const PointerType *GEPPointerTy = cast<PointerType>(BasePtr1Ty);
885 // Find the (possibly empty) initial sequence of equal values... which are not
886 // necessarily constants.
887 unsigned NumGEP1Operands = NumGEP1Ops, NumGEP2Operands = NumGEP2Ops;
888 unsigned MinOperands = std::min(NumGEP1Operands, NumGEP2Operands);
889 unsigned MaxOperands = std::max(NumGEP1Operands, NumGEP2Operands);
890 unsigned UnequalOper = 0;
891 while (UnequalOper != MinOperands &&
892 IndexOperandsEqual(GEP1Ops[UnequalOper], GEP2Ops[UnequalOper])) {
893 // Advance through the type as we go...
895 if (const CompositeType *CT = dyn_cast<CompositeType>(BasePtr1Ty))
896 BasePtr1Ty = CT->getTypeAtIndex(GEP1Ops[UnequalOper-1]);
898 // If all operands equal each other, then the derived pointers must
899 // alias each other...
901 assert(UnequalOper == NumGEP1Operands && UnequalOper == NumGEP2Operands &&
902 "Ran out of type nesting, but not out of operands?");
907 // If we have seen all constant operands, and run out of indexes on one of the
908 // getelementptrs, check to see if the tail of the leftover one is all zeros.
909 // If so, return mustalias.
910 if (UnequalOper == MinOperands) {
911 if (NumGEP1Ops < NumGEP2Ops) {
912 std::swap(GEP1Ops, GEP2Ops);
913 std::swap(NumGEP1Ops, NumGEP2Ops);
916 bool AllAreZeros = true;
917 for (unsigned i = UnequalOper; i != MaxOperands; ++i)
918 if (!isa<Constant>(GEP1Ops[i]) ||
919 !cast<Constant>(GEP1Ops[i])->isNullValue()) {
923 if (AllAreZeros) return MustAlias;
927 // So now we know that the indexes derived from the base pointers,
928 // which are known to alias, are different. We can still determine a
929 // no-alias result if there are differing constant pairs in the index
930 // chain. For example:
931 // A[i][0] != A[j][1] iff (&A[0][1]-&A[0][0] >= std::max(G1S, G2S))
933 // We have to be careful here about array accesses. In particular, consider:
934 // A[1][0] vs A[0][i]
935 // In this case, we don't *know* that the array will be accessed in bounds:
936 // the index could even be negative. Because of this, we have to
937 // conservatively *give up* and return may alias. We disregard differing
938 // array subscripts that are followed by a variable index without going
941 unsigned SizeMax = std::max(G1S, G2S);
942 if (SizeMax == ~0U) return MayAlias; // Avoid frivolous work.
944 // Scan for the first operand that is constant and unequal in the
945 // two getelementptrs...
946 unsigned FirstConstantOper = UnequalOper;
947 for (; FirstConstantOper != MinOperands; ++FirstConstantOper) {
948 const Value *G1Oper = GEP1Ops[FirstConstantOper];
949 const Value *G2Oper = GEP2Ops[FirstConstantOper];
951 if (G1Oper != G2Oper) // Found non-equal constant indexes...
952 if (Constant *G1OC = dyn_cast<ConstantInt>(const_cast<Value*>(G1Oper)))
953 if (Constant *G2OC = dyn_cast<ConstantInt>(const_cast<Value*>(G2Oper))){
954 if (G1OC->getType() != G2OC->getType()) {
955 // Sign extend both operands to long.
956 const Type *Int64Ty = Type::getInt64Ty(G1OC->getContext());
957 if (G1OC->getType() != Int64Ty)
958 G1OC = ConstantExpr::getSExt(G1OC, Int64Ty);
959 if (G2OC->getType() != Int64Ty)
960 G2OC = ConstantExpr::getSExt(G2OC, Int64Ty);
961 GEP1Ops[FirstConstantOper] = G1OC;
962 GEP2Ops[FirstConstantOper] = G2OC;
966 // Handle the "be careful" case above: if this is an array/vector
967 // subscript, scan for a subsequent variable array index.
968 if (const SequentialType *STy =
969 dyn_cast<SequentialType>(BasePtr1Ty)) {
970 const Type *NextTy = STy;
971 bool isBadCase = false;
973 for (unsigned Idx = FirstConstantOper;
974 Idx != MinOperands && isa<SequentialType>(NextTy); ++Idx) {
975 const Value *V1 = GEP1Ops[Idx], *V2 = GEP2Ops[Idx];
976 if (!isa<Constant>(V1) || !isa<Constant>(V2)) {
980 // If the array is indexed beyond the bounds of the static type
981 // at this level, it will also fall into the "be careful" case.
982 // It would theoretically be possible to analyze these cases,
983 // but for now just be conservatively correct.
984 if (const ArrayType *ATy = dyn_cast<ArrayType>(STy))
985 if (cast<ConstantInt>(G1OC)->getZExtValue() >=
986 ATy->getNumElements() ||
987 cast<ConstantInt>(G2OC)->getZExtValue() >=
988 ATy->getNumElements()) {
992 if (const VectorType *VTy = dyn_cast<VectorType>(STy))
993 if (cast<ConstantInt>(G1OC)->getZExtValue() >=
994 VTy->getNumElements() ||
995 cast<ConstantInt>(G2OC)->getZExtValue() >=
996 VTy->getNumElements()) {
1000 STy = cast<SequentialType>(NextTy);
1001 NextTy = cast<SequentialType>(NextTy)->getElementType();
1004 if (isBadCase) G1OC = 0;
1007 // Make sure they are comparable (ie, not constant expressions), and
1008 // make sure the GEP with the smaller leading constant is GEP1.
1010 Constant *Compare = ConstantExpr::getICmp(ICmpInst::ICMP_SGT,
1012 if (ConstantInt *CV = dyn_cast<ConstantInt>(Compare)) {
1013 if (CV->getZExtValue()) { // If they are comparable and G2 > G1
1014 std::swap(GEP1Ops, GEP2Ops); // Make GEP1 < GEP2
1015 std::swap(NumGEP1Ops, NumGEP2Ops);
1022 BasePtr1Ty = cast<CompositeType>(BasePtr1Ty)->getTypeAtIndex(G1Oper);
1025 // No shared constant operands, and we ran out of common operands. At this
1026 // point, the GEP instructions have run through all of their operands, and we
1027 // haven't found evidence that there are any deltas between the GEP's.
1028 // However, one GEP may have more operands than the other. If this is the
1029 // case, there may still be hope. Check this now.
1030 if (FirstConstantOper == MinOperands) {
1031 // Without TargetData, we won't know what the offsets are.
1035 // Make GEP1Ops be the longer one if there is a longer one.
1036 if (NumGEP1Ops < NumGEP2Ops) {
1037 std::swap(GEP1Ops, GEP2Ops);
1038 std::swap(NumGEP1Ops, NumGEP2Ops);
1041 // Is there anything to check?
1042 if (NumGEP1Ops > MinOperands) {
1043 for (unsigned i = FirstConstantOper; i != MaxOperands; ++i)
1044 if (isa<ConstantInt>(GEP1Ops[i]) &&
1045 !cast<ConstantInt>(GEP1Ops[i])->isZero()) {
1046 // Yup, there's a constant in the tail. Set all variables to
1047 // constants in the GEP instruction to make it suitable for
1048 // TargetData::getIndexedOffset.
1049 for (i = 0; i != MaxOperands; ++i)
1050 if (!isa<ConstantInt>(GEP1Ops[i]))
1051 GEP1Ops[i] = Constant::getNullValue(GEP1Ops[i]->getType());
1052 // Okay, now get the offset. This is the relative offset for the full
1054 int64_t Offset1 = TD->getIndexedOffset(GEPPointerTy, GEP1Ops,
1057 // Now check without any constants at the end.
1058 int64_t Offset2 = TD->getIndexedOffset(GEPPointerTy, GEP1Ops,
1061 // Make sure we compare the absolute difference.
1062 if (Offset1 > Offset2)
1063 std::swap(Offset1, Offset2);
1065 // If the tail provided a bit enough offset, return noalias!
1066 if ((uint64_t)(Offset2-Offset1) >= SizeMax)
1068 // Otherwise break - we don't look for another constant in the tail.
1073 // Couldn't find anything useful.
1077 // If there are non-equal constants arguments, then we can figure
1078 // out a minimum known delta between the two index expressions... at
1079 // this point we know that the first constant index of GEP1 is less
1080 // than the first constant index of GEP2.
1082 // Advance BasePtr[12]Ty over this first differing constant operand.
1083 BasePtr2Ty = cast<CompositeType>(BasePtr1Ty)->
1084 getTypeAtIndex(GEP2Ops[FirstConstantOper]);
1085 BasePtr1Ty = cast<CompositeType>(BasePtr1Ty)->
1086 getTypeAtIndex(GEP1Ops[FirstConstantOper]);
1088 // We are going to be using TargetData::getIndexedOffset to determine the
1089 // offset that each of the GEP's is reaching. To do this, we have to convert
1090 // all variable references to constant references. To do this, we convert the
1091 // initial sequence of array subscripts into constant zeros to start with.
1092 const Type *ZeroIdxTy = GEPPointerTy;
1093 for (unsigned i = 0; i != FirstConstantOper; ++i) {
1094 if (!isa<StructType>(ZeroIdxTy))
1095 GEP1Ops[i] = GEP2Ops[i] =
1096 Constant::getNullValue(Type::getInt32Ty(ZeroIdxTy->getContext()));
1098 if (const CompositeType *CT = dyn_cast<CompositeType>(ZeroIdxTy))
1099 ZeroIdxTy = CT->getTypeAtIndex(GEP1Ops[i]);
1102 // We know that GEP1Ops[FirstConstantOper] & GEP2Ops[FirstConstantOper] are ok
1104 // Loop over the rest of the operands...
1105 for (unsigned i = FirstConstantOper+1; i != MaxOperands; ++i) {
1106 const Value *Op1 = i < NumGEP1Ops ? GEP1Ops[i] : 0;
1107 const Value *Op2 = i < NumGEP2Ops ? GEP2Ops[i] : 0;
1108 // If they are equal, use a zero index...
1109 if (Op1 == Op2 && BasePtr1Ty == BasePtr2Ty) {
1110 if (!isa<ConstantInt>(Op1))
1111 GEP1Ops[i] = GEP2Ops[i] = Constant::getNullValue(Op1->getType());
1112 // Otherwise, just keep the constants we have.
1115 if (const ConstantInt *Op1C = dyn_cast<ConstantInt>(Op1)) {
1116 // If this is an array index, make sure the array element is in range.
1117 if (const ArrayType *AT = dyn_cast<ArrayType>(BasePtr1Ty)) {
1118 if (Op1C->getZExtValue() >= AT->getNumElements())
1119 return MayAlias; // Be conservative with out-of-range accesses
1120 } else if (const VectorType *VT = dyn_cast<VectorType>(BasePtr1Ty)) {
1121 if (Op1C->getZExtValue() >= VT->getNumElements())
1122 return MayAlias; // Be conservative with out-of-range accesses
1126 // GEP1 is known to produce a value less than GEP2. To be
1127 // conservatively correct, we must assume the largest possible
1128 // constant is used in this position. This cannot be the initial
1129 // index to the GEP instructions (because we know we have at least one
1130 // element before this one with the different constant arguments), so
1131 // we know that the current index must be into either a struct or
1132 // array. Because we know it's not constant, this cannot be a
1133 // structure index. Because of this, we can calculate the maximum
1136 if (const ArrayType *AT = dyn_cast<ArrayType>(BasePtr1Ty))
1138 ConstantInt::get(Type::getInt64Ty(AT->getContext()),
1139 AT->getNumElements()-1);
1140 else if (const VectorType *VT = dyn_cast<VectorType>(BasePtr1Ty))
1142 ConstantInt::get(Type::getInt64Ty(VT->getContext()),
1143 VT->getNumElements()-1);
1148 if (const ConstantInt *Op2C = dyn_cast<ConstantInt>(Op2)) {
1149 // If this is an array index, make sure the array element is in range.
1150 if (const ArrayType *AT = dyn_cast<ArrayType>(BasePtr2Ty)) {
1151 if (Op2C->getZExtValue() >= AT->getNumElements())
1152 return MayAlias; // Be conservative with out-of-range accesses
1153 } else if (const VectorType *VT = dyn_cast<VectorType>(BasePtr2Ty)) {
1154 if (Op2C->getZExtValue() >= VT->getNumElements())
1155 return MayAlias; // Be conservative with out-of-range accesses
1157 } else { // Conservatively assume the minimum value for this index
1158 GEP2Ops[i] = Constant::getNullValue(Op2->getType());
1163 if (BasePtr1Ty && Op1) {
1164 if (const CompositeType *CT = dyn_cast<CompositeType>(BasePtr1Ty))
1165 BasePtr1Ty = CT->getTypeAtIndex(GEP1Ops[i]);
1170 if (BasePtr2Ty && Op2) {
1171 if (const CompositeType *CT = dyn_cast<CompositeType>(BasePtr2Ty))
1172 BasePtr2Ty = CT->getTypeAtIndex(GEP2Ops[i]);
1178 if (TD && GEPPointerTy->getElementType()->isSized()) {
1180 TD->getIndexedOffset(GEPPointerTy, GEP1Ops, NumGEP1Ops);
1182 TD->getIndexedOffset(GEPPointerTy, GEP2Ops, NumGEP2Ops);
1183 assert(Offset1 != Offset2 &&
1184 "There is at least one different constant here!");
1186 // Make sure we compare the absolute difference.
1187 if (Offset1 > Offset2)
1188 std::swap(Offset1, Offset2);
1190 if ((uint64_t)(Offset2-Offset1) >= SizeMax) {
1191 //cerr << "Determined that these two GEP's don't alias ["
1192 // << SizeMax << " bytes]: \n" << *GEP1 << *GEP2;
1199 // Make sure that anything that uses AliasAnalysis pulls in this file.
1200 DEFINING_FILE_FOR(BasicAliasAnalysis)