On Fri, 25 Mar 2022 at 05:58, Jason Merrill <jason@redhat.com> wrote:
> >
> >>> +      if (current_function_decl
> >>> +          && (DECL_CONSTRUCTOR_P (current_function_decl)
> >>> +              || DECL_DESTRUCTOR_P (current_function_decl))
> >>> +          && TREE_CODE (expr) == NOP_EXPR
> >>> +          && is_this_parameter (TREE_OPERAND (expr, 0)))
> >>
> >> I think the resolves_to_fixed_type_p function would be useful here;
> >> you're testing for a subset of the cases it handles.
> >
> > Does the new patch look like how you intended it to? The new patch
> > passes all regression tests and newly added tests. However, I don't
> > fully understand the code for resolves_to_fixed_type_p() and
> > fixed_type_or_null(), except that I see that they contain logic to
> > return -1 when within a ctor/dtor.
>
> > +      if (TREE_CODE (expr) == NOP_EXPR
> > +          && resolves_to_fixed_type_p (TREE_OPERAND (expr, 0)) == -1)
>
> Why not just pass expr itself to resolves_to_fixed_type_p?

You're right. I didn't realise that fixed_type_or_null handles the
NOP_EXPR case as part of CASE_CONVERT.

> We might as well also warn if it returns >0, e.g.
>
> struct A { } a;
> struct B: A { };
> auto p = static_cast<B*>(&a); // undefined
>
> You should also handle reference casts.

I've made some changes in v3 to handle reference casts and the case
where resolves_to_fixed_type_p > 0. I've also slightly modified the
patch title to reflect the new changes.

Thanks as always!

v2: https://gcc.gnu.org/pipermail/gcc-patches/2022-March/591894.html
Changes since v2:
- Handle reference casts and add new tests.
- Warn if resolves_to_fixed_type_p > 0 to handle more general cases.

v1: https://gcc.gnu.org/pipermail/gcc-patches/2022-March/591683.html
- Add new warning flag.
- Use resolves_to_fixed_type_p.