//
// Another limitation is that it assumes all code will be executed. A store
// through a null pointer in a basic block which is never reached is harmless,
-// but this pass will warn about it anyway.
+// but this pass will warn about it anyway. This is the main reason why most
+// of these checks live here instead of in the Verifier pass.
//
// Optimization passes may make conditions that this pass checks for more or
// less obvious. If an optimization pass appears to be introducing a warning,
#include "llvm/Analysis/Loads.h"
#include "llvm/Analysis/ValueTracking.h"
#include "llvm/Assembly/Writer.h"
-#include "llvm/Target/TargetData.h"
+#include "llvm/DataLayout.h"
+#include "llvm/Target/TargetLibraryInfo.h"
#include "llvm/Pass.h"
#include "llvm/PassManager.h"
#include "llvm/IntrinsicInst.h"
void visitFunction(Function &F);
void visitCallSite(CallSite CS);
- void visitMemoryReference(Instruction &I, Value *Ptr, unsigned Align,
- const Type *Ty, unsigned Flags);
+ void visitMemoryReference(Instruction &I, Value *Ptr,
+ uint64_t Size, unsigned Align,
+ Type *Ty, unsigned Flags);
void visitCallInst(CallInst &I);
void visitInvokeInst(InvokeInst &I);
void visitUnreachableInst(UnreachableInst &I);
Value *findValue(Value *V, bool OffsetOk) const;
+ Value *findValueImpl(Value *V, bool OffsetOk,
+ SmallPtrSet<Value *, 4> &Visited) const;
public:
Module *Mod;
AliasAnalysis *AA;
DominatorTree *DT;
- TargetData *TD;
+ DataLayout *TD;
+ TargetLibraryInfo *TLI;
std::string Messages;
raw_string_ostream MessagesStr;
static char ID; // Pass identification, replacement for typeid
- Lint() : FunctionPass(&ID), MessagesStr(Messages) {}
+ Lint() : FunctionPass(ID), MessagesStr(Messages) {
+ initializeLintPass(*PassRegistry::getPassRegistry());
+ }
virtual bool runOnFunction(Function &F);
virtual void getAnalysisUsage(AnalysisUsage &AU) const {
AU.setPreservesAll();
AU.addRequired<AliasAnalysis>();
+ AU.addRequired<TargetLibraryInfo>();
AU.addRequired<DominatorTree>();
}
virtual void print(raw_ostream &O, const Module *M) const {}
}
}
- void WriteType(const Type *T) {
- if (!T) return;
- MessagesStr << ' ';
- WriteTypeSymbolic(MessagesStr, T, Mod);
- }
-
// CheckFailed - A check failed, so print out the condition and the message
// that failed. This provides a nice place to put a breakpoint if you want
// to see why something is not correct.
WriteValue(V3);
WriteValue(V4);
}
-
- void CheckFailed(const Twine &Message, const Value *V1,
- const Type *T2, const Value *V3 = 0) {
- MessagesStr << Message.str() << "\n";
- WriteValue(V1);
- WriteType(T2);
- WriteValue(V3);
- }
-
- void CheckFailed(const Twine &Message, const Type *T1,
- const Type *T2 = 0, const Type *T3 = 0) {
- MessagesStr << Message.str() << "\n";
- WriteType(T1);
- WriteType(T2);
- WriteType(T3);
- }
};
}
char Lint::ID = 0;
-static RegisterPass<Lint>
-X("lint", "Statically lint-checks LLVM IR", false, true);
+INITIALIZE_PASS_BEGIN(Lint, "lint", "Statically lint-checks LLVM IR",
+ false, true)
+INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfo)
+INITIALIZE_PASS_DEPENDENCY(DominatorTree)
+INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
+INITIALIZE_PASS_END(Lint, "lint", "Statically lint-checks LLVM IR",
+ false, true)
// Assert - We know that cond should be true, if not print an error message.
#define Assert(C, M) \
Mod = F.getParent();
AA = &getAnalysis<AliasAnalysis>();
DT = &getAnalysis<DominatorTree>();
- TD = getAnalysisIfAvailable<TargetData>();
+ TD = getAnalysisIfAvailable<DataLayout>();
+ TLI = &getAnalysis<TargetLibraryInfo>();
visit(F);
dbgs() << MessagesStr.str();
Messages.clear();
// fairly common mistake to neglect to name a function.
Assert1(F.hasName() || F.hasLocalLinkage(),
"Unusual: Unnamed function with non-local linkage", &F);
+
+ // TODO: Check for irreducible control flow.
}
void Lint::visitCallSite(CallSite CS) {
Instruction &I = *CS.getInstruction();
Value *Callee = CS.getCalledValue();
- visitMemoryReference(I, Callee, 0, 0, MemRef::Callee);
+ visitMemoryReference(I, Callee, AliasAnalysis::UnknownSize,
+ 0, 0, MemRef::Callee);
if (Function *F = dyn_cast<Function>(findValue(Callee, /*OffsetOk=*/false))) {
Assert1(CS.getCallingConv() == F->getCallingConv(),
"Undefined behavior: Caller and callee calling convention differ",
&I);
- const FunctionType *FT = F->getFunctionType();
+ FunctionType *FT = F->getFunctionType();
unsigned NumActualArgs = unsigned(CS.arg_end()-CS.arg_begin());
Assert1(FT->isVarArg() ?
FT->getNumParams() == NumActualArgs,
"Undefined behavior: Call argument count mismatches callee "
"argument count", &I);
-
- // TODO: Check argument types (in case the callee was casted)
- // TODO: Check ABI-significant attributes.
-
- // TODO: Check noalias attribute.
-
- // TODO: Check sret attribute.
+ Assert1(FT->getReturnType() == I.getType(),
+ "Undefined behavior: Call return type mismatches "
+ "callee return type", &I);
+
+ // Check argument types (in case the callee was casted) and attributes.
+ // TODO: Verify that caller and callee attributes are compatible.
+ Function::arg_iterator PI = F->arg_begin(), PE = F->arg_end();
+ CallSite::arg_iterator AI = CS.arg_begin(), AE = CS.arg_end();
+ for (; AI != AE; ++AI) {
+ Value *Actual = *AI;
+ if (PI != PE) {
+ Argument *Formal = PI++;
+ Assert1(Formal->getType() == Actual->getType(),
+ "Undefined behavior: Call argument type mismatches "
+ "callee parameter type", &I);
+
+ // Check that noalias arguments don't alias other arguments. This is
+ // not fully precise because we don't know the sizes of the dereferenced
+ // memory regions.
+ if (Formal->hasNoAliasAttr() && Actual->getType()->isPointerTy())
+ for (CallSite::arg_iterator BI = CS.arg_begin(); BI != AE; ++BI)
+ if (AI != BI && (*BI)->getType()->isPointerTy()) {
+ AliasAnalysis::AliasResult Result = AA->alias(*AI, *BI);
+ Assert1(Result != AliasAnalysis::MustAlias &&
+ Result != AliasAnalysis::PartialAlias,
+ "Unusual: noalias argument aliases another argument", &I);
+ }
+
+ // Check that an sret argument points to valid memory.
+ if (Formal->hasStructRetAttr() && Actual->getType()->isPointerTy()) {
+ Type *Ty =
+ cast<PointerType>(Formal->getType())->getElementType();
+ visitMemoryReference(I, Actual, AA->getTypeStoreSize(Ty),
+ TD ? TD->getABITypeAlignment(Ty) : 0,
+ Ty, MemRef::Read | MemRef::Write);
+ }
+ }
+ }
}
if (CS.isCall() && cast<CallInst>(CS.getInstruction())->isTailCall())
case Intrinsic::memcpy: {
MemCpyInst *MCI = cast<MemCpyInst>(&I);
- visitMemoryReference(I, MCI->getSource(), MCI->getAlignment(), 0,
+ // TODO: If the size is known, use it.
+ visitMemoryReference(I, MCI->getDest(), AliasAnalysis::UnknownSize,
+ MCI->getAlignment(), 0,
MemRef::Write);
- visitMemoryReference(I, MCI->getDest(), MCI->getAlignment(), 0,
+ visitMemoryReference(I, MCI->getSource(), AliasAnalysis::UnknownSize,
+ MCI->getAlignment(), 0,
MemRef::Read);
// Check that the memcpy arguments don't overlap. The AliasAnalysis API
// isn't expressive enough for what we really want to do. Known partial
// overlap is not distinguished from the case where nothing is known.
- unsigned Size = 0;
+ uint64_t Size = 0;
if (const ConstantInt *Len =
dyn_cast<ConstantInt>(findValue(MCI->getLength(),
/*OffsetOk=*/false)))
}
case Intrinsic::memmove: {
MemMoveInst *MMI = cast<MemMoveInst>(&I);
- visitMemoryReference(I, MMI->getSource(), MMI->getAlignment(), 0,
+ // TODO: If the size is known, use it.
+ visitMemoryReference(I, MMI->getDest(), AliasAnalysis::UnknownSize,
+ MMI->getAlignment(), 0,
MemRef::Write);
- visitMemoryReference(I, MMI->getDest(), MMI->getAlignment(), 0,
+ visitMemoryReference(I, MMI->getSource(), AliasAnalysis::UnknownSize,
+ MMI->getAlignment(), 0,
MemRef::Read);
break;
}
case Intrinsic::memset: {
MemSetInst *MSI = cast<MemSetInst>(&I);
- visitMemoryReference(I, MSI->getDest(), MSI->getAlignment(), 0,
+ // TODO: If the size is known, use it.
+ visitMemoryReference(I, MSI->getDest(), AliasAnalysis::UnknownSize,
+ MSI->getAlignment(), 0,
MemRef::Write);
break;
}
"Undefined behavior: va_start called in a non-varargs function",
&I);
- visitMemoryReference(I, CS.getArgument(0), 0, 0,
- MemRef::Read | MemRef::Write);
+ visitMemoryReference(I, CS.getArgument(0), AliasAnalysis::UnknownSize,
+ 0, 0, MemRef::Read | MemRef::Write);
break;
case Intrinsic::vacopy:
- visitMemoryReference(I, CS.getArgument(0), 0, 0, MemRef::Write);
- visitMemoryReference(I, CS.getArgument(1), 0, 0, MemRef::Read);
+ visitMemoryReference(I, CS.getArgument(0), AliasAnalysis::UnknownSize,
+ 0, 0, MemRef::Write);
+ visitMemoryReference(I, CS.getArgument(1), AliasAnalysis::UnknownSize,
+ 0, 0, MemRef::Read);
break;
case Intrinsic::vaend:
- visitMemoryReference(I, CS.getArgument(0), 0, 0,
- MemRef::Read | MemRef::Write);
+ visitMemoryReference(I, CS.getArgument(0), AliasAnalysis::UnknownSize,
+ 0, 0, MemRef::Read | MemRef::Write);
break;
case Intrinsic::stackrestore:
// Stackrestore doesn't read or write memory, but it sets the
// stack pointer, which the compiler may read from or write to
// at any time, so check it for both readability and writeability.
- visitMemoryReference(I, CS.getArgument(0), 0, 0,
- MemRef::Read | MemRef::Write);
+ visitMemoryReference(I, CS.getArgument(0), AliasAnalysis::UnknownSize,
+ 0, 0, MemRef::Read | MemRef::Write);
break;
}
}
}
}
-// TODO: Add a length argument and check that the reference is in bounds
+// TODO: Check that the reference is in bounds.
+// TODO: Check readnone/readonly function attributes.
void Lint::visitMemoryReference(Instruction &I,
- Value *Ptr, unsigned Align, const Type *Ty,
- unsigned Flags) {
+ Value *Ptr, uint64_t Size, unsigned Align,
+ Type *Ty, unsigned Flags) {
+ // If no memory is being referenced, it doesn't matter if the pointer
+ // is valid.
+ if (Size == 0)
+ return;
+
Value *UnderlyingObject = findValue(Ptr, /*OffsetOk=*/true);
Assert1(!isa<ConstantPointerNull>(UnderlyingObject),
"Undefined behavior: Null pointer dereference", &I);
Assert1(!isa<UndefValue>(UnderlyingObject),
"Undefined behavior: Undef pointer dereference", &I);
+ Assert1(!isa<ConstantInt>(UnderlyingObject) ||
+ !cast<ConstantInt>(UnderlyingObject)->isAllOnesValue(),
+ "Unusual: All-ones pointer dereference", &I);
+ Assert1(!isa<ConstantInt>(UnderlyingObject) ||
+ !cast<ConstantInt>(UnderlyingObject)->isOne(),
+ "Unusual: Address one pointer dereference", &I);
if (Flags & MemRef::Write) {
if (const GlobalVariable *GV = dyn_cast<GlobalVariable>(UnderlyingObject))
"Undefined behavior: Branch to non-blockaddress", &I);
}
+ // Check for buffer overflows and misalignment.
if (TD) {
- if (Align == 0 && Ty) Align = TD->getABITypeAlignment(Ty);
-
- if (Align != 0) {
- unsigned BitWidth = TD->getTypeSizeInBits(Ptr->getType());
- APInt Mask = APInt::getAllOnesValue(BitWidth),
- KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
- ComputeMaskedBits(Ptr, Mask, KnownZero, KnownOne, TD);
- Assert1(!(KnownOne & APInt::getLowBitsSet(BitWidth, Log2_32(Align))),
+ // Only handles memory references that read/write something simple like an
+ // alloca instruction or a global variable.
+ int64_t Offset = 0;
+ if (Value *Base = GetPointerBaseWithConstantOffset(Ptr, Offset, *TD)) {
+ // OK, so the access is to a constant offset from Ptr. Check that Ptr is
+ // something we can handle and if so extract the size of this base object
+ // along with its alignment.
+ uint64_t BaseSize = AliasAnalysis::UnknownSize;
+ unsigned BaseAlign = 0;
+
+ if (AllocaInst *AI = dyn_cast<AllocaInst>(Base)) {
+ Type *ATy = AI->getAllocatedType();
+ if (!AI->isArrayAllocation() && ATy->isSized())
+ BaseSize = TD->getTypeAllocSize(ATy);
+ BaseAlign = AI->getAlignment();
+ if (BaseAlign == 0 && ATy->isSized())
+ BaseAlign = TD->getABITypeAlignment(ATy);
+ } else if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Base)) {
+ // If the global may be defined differently in another compilation unit
+ // then don't warn about funky memory accesses.
+ if (GV->hasDefinitiveInitializer()) {
+ Type *GTy = GV->getType()->getElementType();
+ if (GTy->isSized())
+ BaseSize = TD->getTypeAllocSize(GTy);
+ BaseAlign = GV->getAlignment();
+ if (BaseAlign == 0 && GTy->isSized())
+ BaseAlign = TD->getABITypeAlignment(GTy);
+ }
+ }
+
+ // Accesses from before the start or after the end of the object are not
+ // defined.
+ Assert1(Size == AliasAnalysis::UnknownSize ||
+ BaseSize == AliasAnalysis::UnknownSize ||
+ (Offset >= 0 && Offset + Size <= BaseSize),
+ "Undefined behavior: Buffer overflow", &I);
+
+ // Accesses that say that the memory is more aligned than it is are not
+ // defined.
+ if (Align == 0 && Ty && Ty->isSized())
+ Align = TD->getABITypeAlignment(Ty);
+ Assert1(!BaseAlign || Align <= MinAlign(BaseAlign, Offset),
"Undefined behavior: Memory reference address is misaligned", &I);
}
}
}
void Lint::visitLoadInst(LoadInst &I) {
- visitMemoryReference(I, I.getPointerOperand(), I.getAlignment(), I.getType(),
- MemRef::Read);
+ visitMemoryReference(I, I.getPointerOperand(),
+ AA->getTypeStoreSize(I.getType()), I.getAlignment(),
+ I.getType(), MemRef::Read);
}
void Lint::visitStoreInst(StoreInst &I) {
- visitMemoryReference(I, I.getPointerOperand(), I.getAlignment(),
- I.getOperand(0)->getType(), MemRef::Write);
+ visitMemoryReference(I, I.getPointerOperand(),
+ AA->getTypeStoreSize(I.getOperand(0)->getType()),
+ I.getAlignment(),
+ I.getOperand(0)->getType(), MemRef::Write);
}
void Lint::visitXor(BinaryOperator &I) {
"Undefined result: Shift count out of range", &I);
}
-static bool isZero(Value *V, TargetData *TD) {
+static bool isZero(Value *V, DataLayout *TD) {
// Assume undef could be zero.
if (isa<UndefValue>(V)) return true;
unsigned BitWidth = cast<IntegerType>(V->getType())->getBitWidth();
- APInt Mask = APInt::getAllOnesValue(BitWidth),
- KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
- ComputeMaskedBits(V, Mask, KnownZero, KnownOne, TD);
+ APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
+ ComputeMaskedBits(V, KnownZero, KnownOne, TD);
return KnownZero.isAllOnesValue();
}
// This isn't undefined behavior, it's just an obvious pessimization.
Assert1(&I.getParent()->getParent()->getEntryBlock() == I.getParent(),
"Pessimization: Static alloca outside of entry block", &I);
+
+ // TODO: Check for an unusual size (MSB set?)
}
void Lint::visitVAArgInst(VAArgInst &I) {
- visitMemoryReference(I, I.getOperand(0), 0, 0,
+ visitMemoryReference(I, I.getOperand(0), AliasAnalysis::UnknownSize, 0, 0,
MemRef::Read | MemRef::Write);
}
void Lint::visitIndirectBrInst(IndirectBrInst &I) {
- visitMemoryReference(I, I.getAddress(), 0, 0, MemRef::Branchee);
+ visitMemoryReference(I, I.getAddress(), AliasAnalysis::UnknownSize, 0, 0,
+ MemRef::Branchee);
+
+ Assert1(I.getNumDestinations() != 0,
+ "Undefined behavior: indirectbr with no destinations", &I);
}
void Lint::visitExtractElementInst(ExtractElementInst &I) {
/// will simplify most of these kinds of things away. But it's a goal of
/// this Lint pass to be useful even on non-optimized IR.
Value *Lint::findValue(Value *V, bool OffsetOk) const {
+ SmallPtrSet<Value *, 4> Visited;
+ return findValueImpl(V, OffsetOk, Visited);
+}
+
+/// findValueImpl - Implementation helper for findValue.
+Value *Lint::findValueImpl(Value *V, bool OffsetOk,
+ SmallPtrSet<Value *, 4> &Visited) const {
+ // Detect self-referential values.
+ if (!Visited.insert(V))
+ return UndefValue::get(V->getType());
+
// TODO: Look through sext or zext cast, when the result is known to
// be interpreted as signed or unsigned, respectively.
+ // TODO: Look through eliminable cast pairs.
// TODO: Look through calls with unique return values.
// TODO: Look through vector insert/extract/shuffle.
- V = OffsetOk ? V->getUnderlyingObject() : V->stripPointerCasts();
+ V = OffsetOk ? GetUnderlyingObject(V, TD) : V->stripPointerCasts();
if (LoadInst *L = dyn_cast<LoadInst>(V)) {
BasicBlock::iterator BBI = L;
BasicBlock *BB = L->getParent();
+ SmallPtrSet<BasicBlock *, 4> VisitedBlocks;
for (;;) {
+ if (!VisitedBlocks.insert(BB)) break;
if (Value *U = FindAvailableLoadedValue(L->getPointerOperand(),
BB, BBI, 6, AA))
- return findValue(U, OffsetOk);
- BB = L->getParent()->getUniquePredecessor();
+ return findValueImpl(U, OffsetOk, Visited);
+ if (BBI != BB->begin()) break;
+ BB = BB->getUniquePredecessor();
if (!BB) break;
BBI = BB->end();
}
+ } else if (PHINode *PN = dyn_cast<PHINode>(V)) {
+ if (Value *W = PN->hasConstantValue())
+ if (W != V)
+ return findValueImpl(W, OffsetOk, Visited);
} else if (CastInst *CI = dyn_cast<CastInst>(V)) {
if (CI->isNoopCast(TD ? TD->getIntPtrType(V->getContext()) :
Type::getInt64Ty(V->getContext())))
- return findValue(CI->getOperand(0), OffsetOk);
- } else if (PHINode *PN = dyn_cast<PHINode>(V)) {
- if (Value *W = PN->hasConstantValue(DT))
- return findValue(W, OffsetOk);
+ return findValueImpl(CI->getOperand(0), OffsetOk, Visited);
} else if (ExtractValueInst *Ex = dyn_cast<ExtractValueInst>(V)) {
if (Value *W = FindInsertedValue(Ex->getAggregateOperand(),
- Ex->idx_begin(),
- Ex->idx_end()))
+ Ex->getIndices()))
if (W != V)
- return findValue(W, OffsetOk);
+ return findValueImpl(W, OffsetOk, Visited);
+ } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) {
+ // Same as above, but for ConstantExpr instead of Instruction.
+ if (Instruction::isCast(CE->getOpcode())) {
+ if (CastInst::isNoopCast(Instruction::CastOps(CE->getOpcode()),
+ CE->getOperand(0)->getType(),
+ CE->getType(),
+ TD ? TD->getIntPtrType(V->getContext()) :
+ Type::getInt64Ty(V->getContext())))
+ return findValueImpl(CE->getOperand(0), OffsetOk, Visited);
+ } else if (CE->getOpcode() == Instruction::ExtractValue) {
+ ArrayRef<unsigned> Indices = CE->getIndices();
+ if (Value *W = FindInsertedValue(CE->getOperand(0), Indices))
+ if (W != V)
+ return findValueImpl(W, OffsetOk, Visited);
+ }
}
// As a last resort, try SimplifyInstruction or constant folding.
if (Instruction *Inst = dyn_cast<Instruction>(V)) {
- if (Value *W = SimplifyInstruction(Inst, TD))
- if (W != Inst)
- return findValue(W, OffsetOk);
+ if (Value *W = SimplifyInstruction(Inst, TD, TLI, DT))
+ return findValueImpl(W, OffsetOk, Visited);
} else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) {
- if (Value *W = ConstantFoldConstantExpression(CE, TD))
+ if (Value *W = ConstantFoldConstantExpression(CE, TD, TLI))
if (W != V)
- return findValue(W, OffsetOk);
+ return findValueImpl(W, OffsetOk, Visited);
}
return V;
projects / oota-llvm.git / blobdiff