On Fri, 7 Nov 2025, Qing Zhao wrote:
> + gcc_assert (TYPE_NAME (outmost_struct_type) != NULL);
> + /* If the type of the containing structure is an anonymous struct/union,
> + get the first outer named structure/union type. */
> + while (TYPE_NAME (type) == NULL_TREE)
I'm not sure that TYPE_NAME being NULL_TREE is always the right check.
You can have a struct or union with no tag that is its own top-level type,
not an example of the anonymous structs and unions feature; that only
applies when the *field* with the struct or union type has no name (in
addition to the struct or union itself having no name - and some
extensions allow a named struct or union to be used with an unnamed field
as well) - and when the field is declared directly with a structure or
union specifier (not typedef or typeof for such - again, some extensions
allow that).
So, for example,
struct { int a; char b[] __attribute__ ((counted_by (a))); } *x;
is a legitimate use of an unnamed structure type, where the counted_by
checks should take place at the level of that type rather than expecting
it to appear in a larger type. Similarly, if you name the field using a
type without a tag,
struct s { struct { int a; char b[] __attribute__ ((counted_by (a))); } x; };
then again, the checks should take place in the context of the inner,
unnamed structure type. It's only when the *field* is unnamed that the
members of the unnamed struct or union are considered members of the
containing struct or union and so should all be considered together for
counted_by purposes.
This means the patch should include testcases (both valid, and invalid
ones being diagnosed) with unnamed structure or union types used in
contexts that *don't* insert their members into the namespace of some
outer structure or union, and verify that the compiler behaves as expected
on such cases (and I suspect the code based on TYPE_NAME will need
changing to make it behave correctly).
--
Joseph S. Myers
[email protected]
