https://gcc.gnu.org/bugzilla/show_bug.cgi?id=102975
Patrick Palka <ppalka at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
CC| |ppalka at gcc dot gnu.org
--- Comment #2 from Patrick Palka <ppalka at gcc dot gnu.org> ---
(In reply to Andrew Pinski from comment #1)
> Hmm, this is interesting:
> template<class> concept Never = false;
> template<class T> concept C = Never<typename T::type>;
> void f() {
> struct X {
> // using type = int;
> };
> static_assert(not C<X>);
> }
>
> is able to compile. I don't know enough about C++ concepts to say if this is
> valid or not but it looks like the type is really unused in the above case
> ...
> This is different from "constexpr bool" which requires the type to be
> defined ...
Yeah, when evaluating a concept-id such as C<X> (which is done by substituting
{X} into the normal form of C), the normalization process
(https://eel.is/c++draft/temp.constr.normal) ignores arguments to unused
template parameters.
So since 'Never' doesn't use its template parameter, during normalization of C
we discard the template argument 'typename T::type' passed to Never. The
normal form of C ends up being 'false (with empty parameter mapping)' which is
trivially satisfied for all T. So it seems to me the warning is correct since
evaluating C<X> doesn't actually use X::type, and the alias is never otherwise
used.
(If this discarding unused template parameters of a concept is undesirable, one
can define the concept in question in a way that trivially uses its template
parameter, e.g.:
template<class T> concept Never = requires { typename T; } && false;)
