//===----------------------------------------------------------------------===//
#include "llvm/Analysis/AliasAnalysis.h"
+#include "llvm/Analysis/CaptureTracking.h"
#include "llvm/Analysis/Passes.h"
#include "llvm/Constants.h"
#include "llvm/DerivedTypes.h"
#include "llvm/GlobalVariable.h"
#include "llvm/Instructions.h"
#include "llvm/IntrinsicInst.h"
+#include "llvm/LLVMContext.h"
+#include "llvm/Operator.h"
#include "llvm/Pass.h"
#include "llvm/Target/TargetData.h"
#include "llvm/ADT/SmallVector.h"
#include "llvm/ADT/STLExtras.h"
#include "llvm/Support/Compiler.h"
+#include "llvm/Support/ErrorHandling.h"
#include "llvm/Support/GetElementPtrTypeIterator.h"
-#include "llvm/Support/ManagedStatic.h"
#include <algorithm>
using namespace llvm;
// Useful predicates
//===----------------------------------------------------------------------===//
-// Determine if an AllocationInst instruction escapes from the function it is
-// contained in. If it does not escape, there is no way for another function to
-// mod/ref it. We do this by looking at its uses and determining if the uses
-// can escape (recursively).
-static bool AddressMightEscape(const Value *V) {
- for (Value::use_const_iterator UI = V->use_begin(), E = V->use_end();
- UI != E; ++UI) {
- const Instruction *I = cast<Instruction>(*UI);
- switch (I->getOpcode()) {
- case Instruction::Load:
- break; //next use.
- case Instruction::Store:
- if (I->getOperand(0) == V)
- return true; // Escapes if the pointer is stored.
- break; // next use.
- case Instruction::GetElementPtr:
- if (AddressMightEscape(I))
- return true;
- break; // next use.
- case Instruction::BitCast:
- if (AddressMightEscape(I))
- return true;
- break; // next use
- case Instruction::Ret:
- // If returned, the address will escape to calling functions, but no
- // callees could modify it.
- break; // next use
- case Instruction::Call:
- // If the call is to a few known safe intrinsics, we know that it does
- // not escape.
- // TODO: Eventually just check the 'nocapture' attribute.
- if (!isa<MemIntrinsic>(I))
- return true;
- break; // next use
- default:
- return true;
- }
- }
- return false;
-}
-
-static const User *isGEP(const Value *V) {
- if (isa<GetElementPtrInst>(V) ||
- (isa<ConstantExpr>(V) &&
- cast<ConstantExpr>(V)->getOpcode() == Instruction::GetElementPtr))
- return cast<User>(V);
- return 0;
+static const GEPOperator *isGEP(const Value *V) {
+ return dyn_cast<GEPOperator>(V);
}
static const Value *GetGEPOperands(const Value *V,
- SmallVector<Value*, 16> &GEPOps){
+ SmallVector<Value*, 16> &GEPOps) {
assert(GEPOps.empty() && "Expect empty list to populate!");
GEPOps.insert(GEPOps.end(), cast<User>(V)->op_begin()+1,
cast<User>(V)->op_end());
return V;
}
-/// isNoAliasCall - Return true if this pointer is returned by a noalias
-/// function.
-static bool isNoAliasCall(const Value *V) {
- if (isa<CallInst>(V) || isa<InvokeInst>(V))
- return CallSite(const_cast<Instruction*>(cast<Instruction>(V)))
- .paramHasAttr(0, Attribute::NoAlias);
- return false;
-}
-
-/// isIdentifiedObject - Return true if this pointer refers to a distinct and
-/// identifiable object. This returns true for:
-/// Global Variables and Functions
-/// Allocas and Mallocs
-/// ByVal and NoAlias Arguments
-/// NoAlias returns
-///
-static bool isIdentifiedObject(const Value *V) {
- if (isa<GlobalValue>(V) || isa<AllocationInst>(V) || isNoAliasCall(V))
- return true;
- if (const Argument *A = dyn_cast<Argument>(V))
- return A->hasNoAliasAttr() || A->hasByValAttr();
- return false;
-}
-
/// isKnownNonNull - Return true if we know that the specified value is never
/// null.
static bool isKnownNonNull(const Value *V) {
static bool isNonEscapingLocalObject(const Value *V) {
// If this is a local allocation, check to see if it escapes.
if (isa<AllocationInst>(V) || isNoAliasCall(V))
- return !AddressMightEscape(V);
-
+ return !PointerMayBeCaptured(V, false);
+
// If this is an argument that corresponds to a byval or noalias argument,
- // it can't escape either.
+ // then it has not escaped before entering the function. Check if it escapes
+ // inside the function.
if (const Argument *A = dyn_cast<Argument>(V))
- if (A->hasByValAttr() || A->hasNoAliasAttr())
- return !AddressMightEscape(V);
+ if (A->hasByValAttr() || A->hasNoAliasAttr()) {
+ // Don't bother analyzing arguments already known not to escape.
+ if (A->hasNoCaptureAttr())
+ return true;
+ return !PointerMayBeCaptured(V, false);
+ }
return false;
}
}
if (AccessTy->isSized())
- return TD.getABITypeSize(AccessTy) < Size;
+ return TD.getTypeAllocSize(AccessTy) < Size;
return false;
}
return MayAlias;
}
- virtual ModRefBehavior getModRefBehavior(Function *F, CallSite CS,
- std::vector<PointerAccessInfo> *Info) {
- return UnknownModRefBehavior;
- }
-
virtual void getArgumentAccesses(Function *F, CallSite CS,
std::vector<PointerAccessInfo> &Info) {
- assert(0 && "This method may not be called on this function!");
+ llvm_unreachable("This method may not be called on this function!");
}
virtual void getMustAliases(Value *P, std::vector<Value*> &RetVals) { }
const Value *V2, unsigned V2Size);
ModRefResult getModRefInfo(CallSite CS, Value *P, unsigned Size);
- ModRefResult getModRefInfo(CallSite CS1, CallSite CS2) {
- return NoAA::getModRefInfo(CS1,CS2);
- }
+ ModRefResult getModRefInfo(CallSite CS1, CallSite CS2);
/// hasNoModRefInfoForCalls - We can provide mod/ref information against
/// non-escaping allocations.
return false;
}
+
// getModRefInfo - Check to see if the specified callsite can clobber the
// specified memory object. Since we only look at local properties of this
// function, we really can't say much about this query. We do, however, use
// If the pointer is to a locally allocated object that does not escape,
// then the call can not mod/ref the pointer unless the call takes the
// argument without capturing it.
- if (isNonEscapingLocalObject(Object)) {
+ if (isNonEscapingLocalObject(Object) && CS.getInstruction() != Object) {
bool passedAsArg = false;
// TODO: Eventually only check 'nocapture' arguments.
for (CallSite::arg_iterator CI = CS.arg_begin(), CE = CS.arg_end();
}
+AliasAnalysis::ModRefResult
+BasicAliasAnalysis::getModRefInfo(CallSite CS1, CallSite CS2) {
+ // If CS1 or CS2 are readnone, they don't interact.
+ ModRefBehavior CS1B = AliasAnalysis::getModRefBehavior(CS1);
+ if (CS1B == DoesNotAccessMemory) return NoModRef;
+
+ ModRefBehavior CS2B = AliasAnalysis::getModRefBehavior(CS2);
+ if (CS2B == DoesNotAccessMemory) return NoModRef;
+
+ // If they both only read from memory, just return ref.
+ if (CS1B == OnlyReadsMemory && CS2B == OnlyReadsMemory)
+ return Ref;
+
+ // Otherwise, fall back to NoAA (mod+ref).
+ return NoAA::getModRefInfo(CS1, CS2);
+}
+
+
// alias - Provide a bunch of ad-hoc rules to disambiguate in common cases, such
-// as array references. Note that this function is heavily tail recursive.
-// Hopefully we have a smart C++ compiler. :)
+// as array references.
//
AliasAnalysis::AliasResult
BasicAliasAnalysis::alias(const Value *V1, unsigned V1Size,
const Value *V2, unsigned V2Size) {
+ LLVMContext &Context = V1->getType()->getContext();
+
// Strip off any constant expression casts if they exist
if (const ConstantExpr *CE = dyn_cast<ConstantExpr>(V1))
if (CE->isCast() && isa<PointerType>(CE->getOperand(0)->getType()))
if (!isa<PointerType>(V1->getType()) || !isa<PointerType>(V2->getType()))
return NoAlias; // Scalars cannot alias each other
- // Strip off cast instructions...
+ // Strip off cast instructions. Since V1 and V2 are pointers, they must be
+ // pointer<->pointer bitcasts.
if (const BitCastInst *I = dyn_cast<BitCastInst>(V1))
return alias(I->getOperand(0), V1Size, V2, V2Size);
if (const BitCastInst *I = dyn_cast<BitCastInst>(V2))
return alias(V1, V1Size, I->getOperand(0), V2Size);
- // Figure out what objects these things are pointing to if we can...
+ // Figure out what objects these things are pointing to if we can.
const Value *O1 = V1->getUnderlyingObject();
const Value *O2 = V2->getUnderlyingObject();
// non-escaping local object, then we know the object couldn't escape to a
// point where the call could return it.
if ((isa<CallInst>(O1) || isa<InvokeInst>(O1)) &&
- isNonEscapingLocalObject(O2))
+ isNonEscapingLocalObject(O2) && O1 != O2)
return NoAlias;
if ((isa<CallInst>(O2) || isa<InvokeInst>(O2)) &&
- isNonEscapingLocalObject(O1))
+ isNonEscapingLocalObject(O1) && O1 != O2)
return NoAlias;
// If we have two gep instructions with must-alias'ing base pointers, figure
// constant expression getelementptrs here.
//
if (isGEP(V1) && isGEP(V2)) {
+ const User *GEP1 = cast<User>(V1);
+ const User *GEP2 = cast<User>(V2);
+
+ // If V1 and V2 are identical GEPs, just recurse down on both of them.
+ // This allows us to analyze things like:
+ // P = gep A, 0, i, 1
+ // Q = gep B, 0, i, 1
+ // by just analyzing A and B. This is even safe for variable indices.
+ if (GEP1->getType() == GEP2->getType() &&
+ GEP1->getNumOperands() == GEP2->getNumOperands() &&
+ GEP1->getOperand(0)->getType() == GEP2->getOperand(0)->getType() &&
+ // All operands are the same, ignoring the base.
+ std::equal(GEP1->op_begin()+1, GEP1->op_end(), GEP2->op_begin()+1))
+ return alias(GEP1->getOperand(0), V1Size, GEP2->getOperand(0), V2Size);
+
+
// Drill down into the first non-gep value, to test for must-aliasing of
// the base pointers.
- const User *G = cast<User>(V1);
- while (isGEP(G->getOperand(0)) &&
- G->getOperand(1) ==
- Constant::getNullValue(G->getOperand(1)->getType()))
- G = cast<User>(G->getOperand(0));
- const Value *BasePtr1 = G->getOperand(0);
-
- G = cast<User>(V2);
- while (isGEP(G->getOperand(0)) &&
- G->getOperand(1) ==
- Constant::getNullValue(G->getOperand(1)->getType()))
- G = cast<User>(G->getOperand(0));
- const Value *BasePtr2 = G->getOperand(0);
+ while (isGEP(GEP1->getOperand(0)) &&
+ GEP1->getOperand(1) ==
+ Context.getNullValue(GEP1->getOperand(1)->getType()))
+ GEP1 = cast<User>(GEP1->getOperand(0));
+ const Value *BasePtr1 = GEP1->getOperand(0);
+
+ while (isGEP(GEP2->getOperand(0)) &&
+ GEP2->getOperand(1) ==
+ Context.getNullValue(GEP2->getOperand(1)->getType()))
+ GEP2 = cast<User>(GEP2->getOperand(0));
+ const Value *BasePtr2 = GEP2->getOperand(0);
// Do the base pointers alias?
AliasResult BaseAlias = alias(BasePtr1, ~0U, BasePtr2, ~0U);
for (unsigned i = 0; i != GEPOperands.size(); ++i)
if (!isa<ConstantInt>(GEPOperands[i]))
GEPOperands[i] =
- Constant::getNullValue(GEPOperands[i]->getType());
+ Context.getNullValue(GEPOperands[i]->getType());
int64_t Offset =
getTargetData().getIndexedOffset(BasePtr->getType(),
&GEPOperands[0],
return MayAlias;
}
-// This function is used to determin if the indices of two GEP instructions are
+// This function is used to determine if the indices of two GEP instructions are
// equal. V1 and V2 are the indices.
-static bool IndexOperandsEqual(Value *V1, Value *V2) {
+static bool IndexOperandsEqual(Value *V1, Value *V2, LLVMContext &Context) {
if (V1->getType() == V2->getType())
return V1 == V2;
if (Constant *C1 = dyn_cast<Constant>(V1))
if (Constant *C2 = dyn_cast<Constant>(V2)) {
// Sign extend the constants to long types, if necessary
if (C1->getType() != Type::Int64Ty)
- C1 = ConstantExpr::getSExt(C1, Type::Int64Ty);
+ C1 = Context.getConstantExprSExt(C1, Type::Int64Ty);
if (C2->getType() != Type::Int64Ty)
- C2 = ConstantExpr::getSExt(C2, Type::Int64Ty);
+ C2 = Context.getConstantExprSExt(C2, Type::Int64Ty);
return C1 == C2;
}
return false;
const PointerType *GEPPointerTy = cast<PointerType>(BasePtr1Ty);
+ LLVMContext &Context = GEPPointerTy->getContext();
+
// Find the (possibly empty) initial sequence of equal values... which are not
// necessarily constants.
unsigned NumGEP1Operands = NumGEP1Ops, NumGEP2Operands = NumGEP2Ops;
unsigned MaxOperands = std::max(NumGEP1Operands, NumGEP2Operands);
unsigned UnequalOper = 0;
while (UnequalOper != MinOperands &&
- IndexOperandsEqual(GEP1Ops[UnequalOper], GEP2Ops[UnequalOper])) {
+ IndexOperandsEqual(GEP1Ops[UnequalOper], GEP2Ops[UnequalOper],
+ Context)) {
// Advance through the type as we go...
++UnequalOper;
if (const CompositeType *CT = dyn_cast<CompositeType>(BasePtr1Ty))
if (G1OC->getType() != G2OC->getType()) {
// Sign extend both operands to long.
if (G1OC->getType() != Type::Int64Ty)
- G1OC = ConstantExpr::getSExt(G1OC, Type::Int64Ty);
+ G1OC = Context.getConstantExprSExt(G1OC, Type::Int64Ty);
if (G2OC->getType() != Type::Int64Ty)
- G2OC = ConstantExpr::getSExt(G2OC, Type::Int64Ty);
+ G2OC = Context.getConstantExprSExt(G2OC, Type::Int64Ty);
GEP1Ops[FirstConstantOper] = G1OC;
GEP2Ops[FirstConstantOper] = G2OC;
}
if (G1OC != G2OC) {
// Handle the "be careful" case above: if this is an array/vector
// subscript, scan for a subsequent variable array index.
- if (isa<SequentialType>(BasePtr1Ty)) {
- const Type *NextTy =
- cast<SequentialType>(BasePtr1Ty)->getElementType();
+ if (const SequentialType *STy =
+ dyn_cast<SequentialType>(BasePtr1Ty)) {
+ const Type *NextTy = STy;
bool isBadCase = false;
- for (unsigned Idx = FirstConstantOper+1;
+ for (unsigned Idx = FirstConstantOper;
Idx != MinOperands && isa<SequentialType>(NextTy); ++Idx) {
const Value *V1 = GEP1Ops[Idx], *V2 = GEP2Ops[Idx];
if (!isa<Constant>(V1) || !isa<Constant>(V2)) {
isBadCase = true;
break;
}
+ // If the array is indexed beyond the bounds of the static type
+ // at this level, it will also fall into the "be careful" case.
+ // It would theoretically be possible to analyze these cases,
+ // but for now just be conservatively correct.
+ if (const ArrayType *ATy = dyn_cast<ArrayType>(STy))
+ if (cast<ConstantInt>(G1OC)->getZExtValue() >=
+ ATy->getNumElements() ||
+ cast<ConstantInt>(G2OC)->getZExtValue() >=
+ ATy->getNumElements()) {
+ isBadCase = true;
+ break;
+ }
+ if (const VectorType *VTy = dyn_cast<VectorType>(STy))
+ if (cast<ConstantInt>(G1OC)->getZExtValue() >=
+ VTy->getNumElements() ||
+ cast<ConstantInt>(G2OC)->getZExtValue() >=
+ VTy->getNumElements()) {
+ isBadCase = true;
+ break;
+ }
+ STy = cast<SequentialType>(NextTy);
NextTy = cast<SequentialType>(NextTy)->getElementType();
}
// TargetData::getIndexedOffset.
for (i = 0; i != MaxOperands; ++i)
if (!isa<ConstantInt>(GEP1Ops[i]))
- GEP1Ops[i] = Constant::getNullValue(GEP1Ops[i]->getType());
+ GEP1Ops[i] = Context.getNullValue(GEP1Ops[i]->getType());
// Okay, now get the offset. This is the relative offset for the full
// instruction.
const TargetData &TD = getTargetData();
const Type *ZeroIdxTy = GEPPointerTy;
for (unsigned i = 0; i != FirstConstantOper; ++i) {
if (!isa<StructType>(ZeroIdxTy))
- GEP1Ops[i] = GEP2Ops[i] = Constant::getNullValue(Type::Int32Ty);
+ GEP1Ops[i] = GEP2Ops[i] = Context.getNullValue(Type::Int32Ty);
if (const CompositeType *CT = dyn_cast<CompositeType>(ZeroIdxTy))
ZeroIdxTy = CT->getTypeAtIndex(GEP1Ops[i]);
// If they are equal, use a zero index...
if (Op1 == Op2 && BasePtr1Ty == BasePtr2Ty) {
if (!isa<ConstantInt>(Op1))
- GEP1Ops[i] = GEP2Ops[i] = Constant::getNullValue(Op1->getType());
+ GEP1Ops[i] = GEP2Ops[i] = Context.getNullValue(Op1->getType());
// Otherwise, just keep the constants we have.
} else {
if (Op1) {
// value possible.
//
if (const ArrayType *AT = dyn_cast<ArrayType>(BasePtr1Ty))
- GEP1Ops[i] = ConstantInt::get(Type::Int64Ty,AT->getNumElements()-1);
+ GEP1Ops[i] =
+ Context.getConstantInt(Type::Int64Ty,AT->getNumElements()-1);
else if (const VectorType *VT = dyn_cast<VectorType>(BasePtr1Ty))
- GEP1Ops[i] = ConstantInt::get(Type::Int64Ty,VT->getNumElements()-1);
+ GEP1Ops[i] =
+ Context.getConstantInt(Type::Int64Ty,VT->getNumElements()-1);
}
}
return MayAlias; // Be conservative with out-of-range accesses
}
} else { // Conservatively assume the minimum value for this index
- GEP2Ops[i] = Constant::getNullValue(Op2->getType());
+ GEP2Ops[i] = Context.getNullValue(Op2->getType());
}
}
}