https://gcc.gnu.org/bugzilla/show_bug.cgi?id=99846
Patrick Palka <ppalka at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
CC| |ppalka at gcc dot gnu.org
--- Comment #7 from Patrick Palka <ppalka at gcc dot gnu.org> ---
Before CWG 2369 (i.e. in GCC 10), for the original testcase, overload
resolution for 'left < right' would reject the operator<=> candidate due to
substitution failure into its return type before ever checking its loopy
constraints, and because we also define the (non-constrained) relational
operators for std::variant alongside operator<=>, overload resolution would
just end up selecting the explicitly defined operator<.
After CWG 2369 (i.e. in GCC 11), we check constraints sooner during overload
resolution, so when considering the operator<=> candidate we end up looping
during constraint checking, which we consider to be a fatal error.
Distilled testcase that exhibits the issue:
template <typename L, typename R> concept comparable_with = requires(L l, R r)
{ l < r; };
template <typename b> concept comparable = comparable_with<b, b>;
template <typename> class variant;
template <typename T> void operator<(variant<T>, variant<T>);
template <typename T> requires(comparable<T>)
typename T::fail operator<=>(variant<T>, variant<T>);
template <typename> class variant {};
struct g;
struct g : variant<g> {};
int main()
{
g left, right;
left < right;
}
