The note in [atomics.types.operations] p23 says "if the underlying type has ..." but this is not the same meaning of "underlying type" used elsewhere ([basic.fundamental] and [dcl.enum]).

[atomics.types.generic] uses the phrase "the template argument for T", and the nested type value_type is also defined. Either would be better than "the underlying type", but I think the note would be improved by referring to "object representation of the type", possibly even making use of std::has_unique_object_representations.