[clang] Crash in clang codegen after CXXDependentScopeMemberExpr change #97483
Labels
clang:frontend
Language frontend issues, e.g. anything involving "Sema"
crash
Prefer [crash-on-valid] or [crash-on-invalid]
Comments
github-actions
bot
added
the
clang
|
@sdkrystian would you mind taking a look at this? It seems to be a regression from #92318 |
EugeneZelenko
added
clang:codegen
|
@llvm/issue-subscribers-clang-codegen Author: Mike Rice (mikerice1969)
Attached test crashes in clang codegen after fd87d76. The test has been creduced from some 600,000+ lines of preprocessed Windows blender code. The crash happens in the same place in the full code (but the expression is used).
|
|
The problem here is that we transform the first operand of binary |
|
Reduced to: struct A {
int x;
};
void operator&(A);
template<typename T>
struct B {
bool f() {
return T::x & 1; // should diagnose this as an invalid id-expression naming a non-static data member per [expr.prim.id.general] p4.
}
};
template struct B<A>; |
|
@mikerice1969 Have a fix ready in #97596 |
Thanks. Taking a look now. |
|
The full preprocessed test also seems to compile okay with this fix. @sdkrystian thanks for the quick fix! |
EugeneZelenko
added
clang:frontend
|
@llvm/issue-subscribers-clang-frontend Author: Mike Rice (mikerice1969)
Attached test crashes in clang codegen after fd87d76. The test has been creduced from some 600,000+ lines of preprocessed Windows blender code. The crash happens in the same place in the full code (but the expression is used).
|
kbluck
pushed a commit
to kbluck/llvm-project
that referenced
this issue
Jul 6, 2024
…nary operator& (llvm#97596) Currently, `TreeTransform::TransformCXXOperatorCallExpr` calls `TreeTransform::TransformAddressOfOperand` to transform the first operand of a `CXXOperatorCallExpr` when its `OverloadOperatorKind` is `OO_Amp` -- regardless of arity. This results in the first operand of binary `operator&` being incorrectly transformed as if it was the operand of the address of operator in cases such as the following: ``` struct A { int x; }; void operator&(A, A); template<typename T> struct B { int f() { return T::x & 1; // invalid reference to 'A::x' is not diagnosed because 'T::x' is incorrectly transformed as if it was the operand of unary operator& } }; template struct B<A>; ``` Prior to llvm#92318 we would build a `CXXDependentScopeMemberExpr` for `T::x` (as with most dependent qualified names that were not member qualified names). Since `TreeTransform::TransformAddressOfOperand` only differs from `TransformExpr` for `DependentScopeDeclRefExpr` and `UnresolvedLookupExpr` operands, `T::x` was transformed "correctly". Now that we build a `DependentScopeDeclRefExpr` for `T::x`, it is incorrectly transformed as if it was the operand of the address of operator and we fail to diagnose the invalid reference to a non-static data member. This patch fixes the issue by only calling `TreeTransform::TransformAddressOfOperand` for `CXXOperatorCallExpr`s with a single operand. This fixes llvm#97483.
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Attached test crashes in clang codegen after fd87d76. The test has been creduced from some 600,000+ lines of preprocessed Windows blender code. The crash happens in the same place in the full code (but the expression is used).
testcase.zip
The text was updated successfully, but these errors were encountered: