1 //===- BasicAliasAnalysis.cpp - Local Alias Analysis Impl -----------------===//
3 // The LLVM Compiler Infrastructure
5 // This file was developed by the LLVM research group and is distributed under
6 // the University of Illinois Open Source 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 // FIXME: This could be extended for a very simple form of mod/ref information.
15 // If a pointer is locally allocated (either malloc or alloca) and never passed
16 // into a call or stored to memory, then we know that calls will not mod/ref the
17 // memory. This can be important for tailcallelim.
19 //===----------------------------------------------------------------------===//
21 #include "llvm/Analysis/AliasAnalysis.h"
22 #include "llvm/Pass.h"
23 #include "llvm/Argument.h"
24 #include "llvm/iOther.h"
25 #include "llvm/iMemory.h"
26 #include "llvm/Constants.h"
27 #include "llvm/GlobalVariable.h"
28 #include "llvm/DerivedTypes.h"
29 #include "llvm/Target/TargetData.h"
30 #include "llvm/Support/GetElementPtrTypeIterator.h"
33 // Make sure that anything that uses AliasAnalysis pulls in this file...
34 void llvm::BasicAAStub() {}
37 struct BasicAliasAnalysis : public ImmutablePass, public AliasAnalysis {
39 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
40 AliasAnalysis::getAnalysisUsage(AU);
43 virtual void initializePass();
45 AliasResult alias(const Value *V1, unsigned V1Size,
46 const Value *V2, unsigned V2Size);
48 /// pointsToConstantMemory - Chase pointers until we find a (constant
50 bool pointsToConstantMemory(const Value *P);
53 // CheckGEPInstructions - Check two GEP instructions with known
54 // must-aliasing base pointers. This checks to see if the index expressions
55 // preclude the pointers from aliasing...
57 CheckGEPInstructions(const Type* BasePtr1Ty, std::vector<Value*> &GEP1Ops,
59 const Type *BasePtr2Ty, std::vector<Value*> &GEP2Ops,
63 // Register this pass...
64 RegisterOpt<BasicAliasAnalysis>
65 X("basicaa", "Basic Alias Analysis (default AA impl)");
67 // Declare that we implement the AliasAnalysis interface
68 RegisterAnalysisGroup<AliasAnalysis, BasicAliasAnalysis, true> Y;
69 } // End of anonymous namespace
71 void BasicAliasAnalysis::initializePass() {
72 InitializeAliasAnalysis(this);
75 // hasUniqueAddress - Return true if the specified value points to something
76 // with a unique, discernable, address.
77 static inline bool hasUniqueAddress(const Value *V) {
78 return isa<GlobalValue>(V) || isa<AllocationInst>(V);
81 // getUnderlyingObject - This traverses the use chain to figure out what object
82 // the specified value points to. If the value points to, or is derived from, a
83 // unique object or an argument, return it.
84 static const Value *getUnderlyingObject(const Value *V) {
85 if (!isa<PointerType>(V->getType())) return 0;
87 // If we are at some type of object... return it.
88 if (hasUniqueAddress(V) || isa<Argument>(V)) return V;
90 // Traverse through different addressing mechanisms...
91 if (const Instruction *I = dyn_cast<Instruction>(V)) {
92 if (isa<CastInst>(I) || isa<GetElementPtrInst>(I))
93 return getUnderlyingObject(I->getOperand(0));
94 } else if (const ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) {
95 if (CE->getOpcode() == Instruction::Cast ||
96 CE->getOpcode() == Instruction::GetElementPtr)
97 return getUnderlyingObject(CE->getOperand(0));
98 } else if (const ConstantPointerRef *CPR = dyn_cast<ConstantPointerRef>(V)) {
99 return CPR->getValue();
104 static const User *isGEP(const Value *V) {
105 if (isa<GetElementPtrInst>(V) ||
106 (isa<ConstantExpr>(V) &&
107 cast<ConstantExpr>(V)->getOpcode() == Instruction::GetElementPtr))
108 return cast<User>(V);
112 static const Value *GetGEPOperands(const Value *V, std::vector<Value*> &GEPOps){
113 assert(GEPOps.empty() && "Expect empty list to populate!");
114 GEPOps.insert(GEPOps.end(), cast<User>(V)->op_begin()+1,
115 cast<User>(V)->op_end());
117 // Accumulate all of the chained indexes into the operand array
118 V = cast<User>(V)->getOperand(0);
120 while (const User *G = isGEP(V)) {
121 if (!isa<Constant>(GEPOps[0]) ||
122 !cast<Constant>(GEPOps[0])->isNullValue())
123 break; // Don't handle folding arbitrary pointer offsets yet...
124 GEPOps.erase(GEPOps.begin()); // Drop the zero index
125 GEPOps.insert(GEPOps.begin(), G->op_begin()+1, G->op_end());
126 V = G->getOperand(0);
131 /// pointsToConstantMemory - Chase pointers until we find a (constant
133 bool BasicAliasAnalysis::pointsToConstantMemory(const Value *P) {
134 if (const Value *V = getUnderlyingObject(P))
135 if (const GlobalVariable *GV = dyn_cast<GlobalVariable>(V))
136 return GV->isConstant();
140 // alias - Provide a bunch of ad-hoc rules to disambiguate in common cases, such
141 // as array references. Note that this function is heavily tail recursive.
142 // Hopefully we have a smart C++ compiler. :)
144 AliasAnalysis::AliasResult
145 BasicAliasAnalysis::alias(const Value *V1, unsigned V1Size,
146 const Value *V2, unsigned V2Size) {
147 // Strip off any constant expression casts if they exist
148 if (const ConstantExpr *CE = dyn_cast<ConstantExpr>(V1))
149 if (CE->getOpcode() == Instruction::Cast)
150 V1 = CE->getOperand(0);
151 if (const ConstantExpr *CE = dyn_cast<ConstantExpr>(V2))
152 if (CE->getOpcode() == Instruction::Cast)
153 V2 = CE->getOperand(0);
155 // Strip off constant pointer refs if they exist
156 if (const ConstantPointerRef *CPR = dyn_cast<ConstantPointerRef>(V1))
157 V1 = CPR->getValue();
158 if (const ConstantPointerRef *CPR = dyn_cast<ConstantPointerRef>(V2))
159 V2 = CPR->getValue();
161 // Are we checking for alias of the same value?
162 if (V1 == V2) return MustAlias;
164 if ((!isa<PointerType>(V1->getType()) || !isa<PointerType>(V2->getType())) &&
165 V1->getType() != Type::LongTy && V2->getType() != Type::LongTy)
166 return NoAlias; // Scalars cannot alias each other
168 // Strip off cast instructions...
169 if (const Instruction *I = dyn_cast<CastInst>(V1))
170 return alias(I->getOperand(0), V1Size, V2, V2Size);
171 if (const Instruction *I = dyn_cast<CastInst>(V2))
172 return alias(V1, V1Size, I->getOperand(0), V2Size);
174 // Figure out what objects these things are pointing to if we can...
175 const Value *O1 = getUnderlyingObject(V1);
176 const Value *O2 = getUnderlyingObject(V2);
178 // Pointing at a discernible object?
180 if (isa<Argument>(O1)) {
181 // Incoming argument cannot alias locally allocated object!
182 if (isa<AllocationInst>(O2)) return NoAlias;
183 // Otherwise, nothing is known...
184 } else if (isa<Argument>(O2)) {
185 // Incoming argument cannot alias locally allocated object!
186 if (isa<AllocationInst>(O1)) return NoAlias;
187 // Otherwise, nothing is known...
189 // If they are two different objects, we know that we have no alias...
190 if (O1 != O2) return NoAlias;
193 // If they are the same object, they we can look at the indexes. If they
194 // index off of the object is the same for both pointers, they must alias.
195 // If they are provably different, they must not alias. Otherwise, we can't
197 } else if (O1 && !isa<Argument>(O1) && isa<ConstantPointerNull>(V2)) {
198 return NoAlias; // Unique values don't alias null
199 } else if (O2 && !isa<Argument>(O2) && isa<ConstantPointerNull>(V1)) {
200 return NoAlias; // Unique values don't alias null
203 // If we have two gep instructions with must-alias'ing base pointers, figure
204 // out if the indexes to the GEP tell us anything about the derived pointer.
205 // Note that we also handle chains of getelementptr instructions as well as
206 // constant expression getelementptrs here.
208 if (isGEP(V1) && isGEP(V2)) {
209 // Drill down into the first non-gep value, to test for must-aliasing of
210 // the base pointers.
211 const Value *BasePtr1 = V1, *BasePtr2 = V2;
213 BasePtr1 = cast<User>(BasePtr1)->getOperand(0);
214 } while (isGEP(BasePtr1) &&
215 cast<User>(BasePtr1)->getOperand(1) ==
216 Constant::getNullValue(cast<User>(BasePtr1)->getOperand(1)->getType()));
218 BasePtr2 = cast<User>(BasePtr2)->getOperand(0);
219 } while (isGEP(BasePtr2) &&
220 cast<User>(BasePtr2)->getOperand(1) ==
221 Constant::getNullValue(cast<User>(BasePtr2)->getOperand(1)->getType()));
223 // Do the base pointers alias?
224 AliasResult BaseAlias = alias(BasePtr1, V1Size, BasePtr2, V2Size);
225 if (BaseAlias == NoAlias) return NoAlias;
226 if (BaseAlias == MustAlias) {
227 // If the base pointers alias each other exactly, check to see if we can
228 // figure out anything about the resultant pointers, to try to prove
231 // Collect all of the chained GEP operands together into one simple place
232 std::vector<Value*> GEP1Ops, GEP2Ops;
233 BasePtr1 = GetGEPOperands(V1, GEP1Ops);
234 BasePtr2 = GetGEPOperands(V2, GEP2Ops);
237 CheckGEPInstructions(BasePtr1->getType(), GEP1Ops, V1Size,
238 BasePtr2->getType(), GEP2Ops, V2Size);
239 if (GAlias != MayAlias)
244 // Check to see if these two pointers are related by a getelementptr
245 // instruction. If one pointer is a GEP with a non-zero index of the other
246 // pointer, we know they cannot alias.
250 std::swap(V1Size, V2Size);
253 if (V1Size != ~0U && V2Size != ~0U)
254 if (const User *GEP = isGEP(V1)) {
255 std::vector<Value*> GEPOperands;
256 const Value *BasePtr = GetGEPOperands(V1, GEPOperands);
258 AliasResult R = alias(BasePtr, V1Size, V2, V2Size);
259 if (R == MustAlias) {
260 // If there is at least one non-zero constant index, we know they cannot
262 bool ConstantFound = false;
263 bool AllZerosFound = true;
264 for (unsigned i = 0, e = GEPOperands.size(); i != e; ++i)
265 if (const Constant *C = dyn_cast<Constant>(GEPOperands[i])) {
266 if (!C->isNullValue()) {
267 ConstantFound = true;
268 AllZerosFound = false;
272 AllZerosFound = false;
275 // If we have getelementptr <ptr>, 0, 0, 0, 0, ... and V2 must aliases
276 // the ptr, the end result is a must alias also.
281 if (V2Size <= 1 && V1Size <= 1) // Just pointer check?
284 // Otherwise we have to check to see that the distance is more than
285 // the size of the argument... build an index vector that is equal to
286 // the arguments provided, except substitute 0's for any variable
287 // indexes we find...
288 for (unsigned i = 0; i != GEPOperands.size(); ++i)
289 if (!isa<Constant>(GEPOperands[i]) ||
290 isa<ConstantExpr>(GEPOperands[i]))
291 GEPOperands[i] =Constant::getNullValue(GEPOperands[i]->getType());
292 int64_t Offset = getTargetData().getIndexedOffset(BasePtr->getType(),
294 if (Offset >= (int64_t)V2Size || Offset <= -(int64_t)V1Size)
303 /// CheckGEPInstructions - Check two GEP instructions with known must-aliasing
304 /// base pointers. This checks to see if the index expressions preclude the
305 /// pointers from aliasing...
306 AliasAnalysis::AliasResult BasicAliasAnalysis::
307 CheckGEPInstructions(const Type* BasePtr1Ty, std::vector<Value*> &GEP1Ops,
309 const Type *BasePtr2Ty, std::vector<Value*> &GEP2Ops,
311 // We currently can't handle the case when the base pointers have different
312 // primitive types. Since this is uncommon anyway, we are happy being
313 // extremely conservative.
314 if (BasePtr1Ty != BasePtr2Ty)
317 const Type *GEPPointerTy = BasePtr1Ty;
319 // Find the (possibly empty) initial sequence of equal values... which are not
320 // necessarily constants.
321 unsigned NumGEP1Operands = GEP1Ops.size(), NumGEP2Operands = GEP2Ops.size();
322 unsigned MinOperands = std::min(NumGEP1Operands, NumGEP2Operands);
323 unsigned MaxOperands = std::max(NumGEP1Operands, NumGEP2Operands);
324 unsigned UnequalOper = 0;
325 while (UnequalOper != MinOperands &&
326 GEP1Ops[UnequalOper] == GEP2Ops[UnequalOper]) {
327 // Advance through the type as we go...
329 if (const CompositeType *CT = dyn_cast<CompositeType>(BasePtr1Ty))
330 BasePtr1Ty = CT->getTypeAtIndex(GEP1Ops[UnequalOper-1]);
332 // If all operands equal each other, then the derived pointers must
333 // alias each other...
335 assert(UnequalOper == NumGEP1Operands && UnequalOper == NumGEP2Operands &&
336 "Ran out of type nesting, but not out of operands?");
341 // If we have seen all constant operands, and run out of indexes on one of the
342 // getelementptrs, check to see if the tail of the leftover one is all zeros.
343 // If so, return mustalias.
344 if (UnequalOper == MinOperands) {
345 if (GEP1Ops.size() < GEP2Ops.size()) std::swap(GEP1Ops, GEP2Ops);
347 bool AllAreZeros = true;
348 for (unsigned i = UnequalOper; i != MaxOperands; ++i)
349 if (!isa<Constant>(GEP1Ops[i]) ||
350 !cast<Constant>(GEP1Ops[i])->isNullValue()) {
354 if (AllAreZeros) return MustAlias;
358 // So now we know that the indexes derived from the base pointers,
359 // which are known to alias, are different. We can still determine a
360 // no-alias result if there are differing constant pairs in the index
361 // chain. For example:
362 // A[i][0] != A[j][1] iff (&A[0][1]-&A[0][0] >= std::max(G1S, G2S))
364 unsigned SizeMax = std::max(G1S, G2S);
365 if (SizeMax == ~0U) return MayAlias; // Avoid frivolous work...
367 // Scan for the first operand that is constant and unequal in the
368 // two getelemenptrs...
369 unsigned FirstConstantOper = UnequalOper;
370 for (; FirstConstantOper != MinOperands; ++FirstConstantOper) {
371 const Value *G1Oper = GEP1Ops[FirstConstantOper];
372 const Value *G2Oper = GEP2Ops[FirstConstantOper];
374 if (G1Oper != G2Oper) // Found non-equal constant indexes...
375 if (Constant *G1OC = dyn_cast<Constant>(const_cast<Value*>(G1Oper)))
376 if (Constant *G2OC = dyn_cast<Constant>(const_cast<Value*>(G2Oper))) {
377 // Make sure they are comparable (ie, not constant expressions)...
378 // and make sure the GEP with the smaller leading constant is GEP1.
379 Constant *Compare = ConstantExpr::get(Instruction::SetGT, G1OC, G2OC);
380 if (ConstantBool *CV = dyn_cast<ConstantBool>(Compare)) {
381 if (CV->getValue()) // If they are comparable and G2 > G1
382 std::swap(GEP1Ops, GEP2Ops); // Make GEP1 < GEP2
386 BasePtr1Ty = cast<CompositeType>(BasePtr1Ty)->getTypeAtIndex(G1Oper);
389 // No shared constant operands, and we ran out of common operands. At this
390 // point, the GEP instructions have run through all of their operands, and we
391 // haven't found evidence that there are any deltas between the GEP's.
392 // However, one GEP may have more operands than the other. If this is the
393 // case, there may still be hope. This this now.
394 if (FirstConstantOper == MinOperands) {
395 // Make GEP1Ops be the longer one if there is a longer one.
396 if (GEP1Ops.size() < GEP2Ops.size())
397 std::swap(GEP1Ops, GEP2Ops);
399 // Is there anything to check?
400 if (GEP1Ops.size() > MinOperands) {
401 for (unsigned i = FirstConstantOper; i != MaxOperands; ++i)
402 if (isa<Constant>(GEP1Ops[i]) && !isa<ConstantExpr>(GEP1Ops[i]) &&
403 !cast<Constant>(GEP1Ops[i])->isNullValue()) {
404 // Yup, there's a constant in the tail. Set all variables to
405 // constants in the GEP instruction to make it suiteable for
406 // TargetData::getIndexedOffset.
407 for (i = 0; i != MaxOperands; ++i)
408 if (!isa<Constant>(GEP1Ops[i]) || isa<ConstantExpr>(GEP1Ops[i]))
409 GEP1Ops[i] = Constant::getNullValue(GEP1Ops[i]->getType());
410 // Okay, now get the offset. This is the relative offset for the full
412 const TargetData &TD = getTargetData();
413 int64_t Offset1 = TD.getIndexedOffset(GEPPointerTy, GEP1Ops);
415 // Now crop off any constants from the end...
416 GEP1Ops.resize(MinOperands);
417 int64_t Offset2 = TD.getIndexedOffset(GEPPointerTy, GEP1Ops);
419 // If the tail provided a bit enough offset, return noalias!
420 if ((uint64_t)(Offset2-Offset1) >= SizeMax)
425 // Couldn't find anything useful.
429 // If there are non-equal constants arguments, then we can figure
430 // out a minimum known delta between the two index expressions... at
431 // this point we know that the first constant index of GEP1 is less
432 // than the first constant index of GEP2.
434 // Advance BasePtr[12]Ty over this first differing constant operand.
435 BasePtr2Ty = cast<CompositeType>(BasePtr1Ty)->getTypeAtIndex(GEP2Ops[FirstConstantOper]);
436 BasePtr1Ty = cast<CompositeType>(BasePtr1Ty)->getTypeAtIndex(GEP1Ops[FirstConstantOper]);
438 // We are going to be using TargetData::getIndexedOffset to determine the
439 // offset that each of the GEP's is reaching. To do this, we have to convert
440 // all variable references to constant references. To do this, we convert the
441 // initial equal sequence of variables into constant zeros to start with.
442 for (unsigned i = 0; i != FirstConstantOper; ++i) {
443 if (!isa<Constant>(GEP1Ops[i]) || isa<ConstantExpr>(GEP1Ops[i]) ||
444 !isa<Constant>(GEP2Ops[i]) || isa<ConstantExpr>(GEP2Ops[i])) {
445 GEP1Ops[i] = Constant::getNullValue(GEP1Ops[i]->getType());
446 GEP2Ops[i] = Constant::getNullValue(GEP2Ops[i]->getType());
450 // We know that GEP1Ops[FirstConstantOper] & GEP2Ops[FirstConstantOper] are ok
452 // Loop over the rest of the operands...
453 for (unsigned i = FirstConstantOper+1; i != MaxOperands; ++i) {
454 const Value *Op1 = i < GEP1Ops.size() ? GEP1Ops[i] : 0;
455 const Value *Op2 = i < GEP2Ops.size() ? GEP2Ops[i] : 0;
456 // If they are equal, use a zero index...
457 if (Op1 == Op2 && BasePtr1Ty == BasePtr2Ty) {
458 if (!isa<Constant>(Op1) || isa<ConstantExpr>(Op1))
459 GEP1Ops[i] = GEP2Ops[i] = Constant::getNullValue(Op1->getType());
460 // Otherwise, just keep the constants we have.
463 if (const ConstantInt *Op1C = dyn_cast<ConstantInt>(Op1)) {
464 // If this is an array index, make sure the array element is in range.
465 if (const ArrayType *AT = dyn_cast<ArrayType>(BasePtr1Ty))
466 if (Op1C->getRawValue() >= AT->getNumElements())
467 return MayAlias; // Be conservative with out-of-range accesses
470 // GEP1 is known to produce a value less than GEP2. To be
471 // conservatively correct, we must assume the largest possible
472 // constant is used in this position. This cannot be the initial
473 // index to the GEP instructions (because we know we have at least one
474 // element before this one with the different constant arguments), so
475 // we know that the current index must be into either a struct or
476 // array. Because we know it's not constant, this cannot be a
477 // structure index. Because of this, we can calculate the maximum
480 if (const ArrayType *AT = dyn_cast<ArrayType>(BasePtr1Ty))
481 GEP1Ops[i] = ConstantSInt::get(Type::LongTy,AT->getNumElements()-1);
486 if (const ConstantInt *Op2C = dyn_cast<ConstantInt>(Op2)) {
487 // If this is an array index, make sure the array element is in range.
488 if (const ArrayType *AT = dyn_cast<ArrayType>(BasePtr1Ty))
489 if (Op2C->getRawValue() >= AT->getNumElements())
490 return MayAlias; // Be conservative with out-of-range accesses
491 } else { // Conservatively assume the minimum value for this index
492 GEP2Ops[i] = Constant::getNullValue(Op2->getType());
497 if (BasePtr1Ty && Op1) {
498 if (const CompositeType *CT = dyn_cast<CompositeType>(BasePtr1Ty))
499 BasePtr1Ty = CT->getTypeAtIndex(GEP1Ops[i]);
504 if (BasePtr2Ty && Op2) {
505 if (const CompositeType *CT = dyn_cast<CompositeType>(BasePtr2Ty))
506 BasePtr2Ty = CT->getTypeAtIndex(GEP2Ops[i]);
512 int64_t Offset1 = getTargetData().getIndexedOffset(GEPPointerTy, GEP1Ops);
513 int64_t Offset2 = getTargetData().getIndexedOffset(GEPPointerTy, GEP2Ops);
514 assert(Offset1 < Offset2 &&"There is at least one different constant here!");
516 if ((uint64_t)(Offset2-Offset1) >= SizeMax) {
517 //std::cerr << "Determined that these two GEP's don't alias ["
518 // << SizeMax << " bytes]: \n" << *GEP1 << *GEP2;