Attached are updated diffs for the ABI spec. Besides a couple of minor fixes, this includes an updated treatment of parameter references in signature expressions. There was a separate thread discussing nested function declarators, but at the recent Rapperswil meeting of WG21, an additional issue came up. Consider: template<class T> auto f(T const x)->decltype(x); // #1 template<class T> auto f(T x)->decltype(x); // #2 Ordinarily, top-level cv-qualifiers are ignored when forming the function type. E.g., "int(*)(T)" and "int(*)(T const)" are the same type. However, in expression contexts we cannot ignore the cv-qualifiers because the expression must mean the same there as it would if it appeared in the function definition (where cv-qualifiers matter). So #1 and #2 above must be distinct templates. Rather than encoding the cv-qualifier on the parameter type of the template signature (which would presumably be a serious break of backward compatibility), we're proposing to encode it in the reference in expression contexts (except for known non-class types in rvalue contexts, where the cv-qualifier is meaningless). E.g., rather than encoding "decltype(x)" in #1 as "Dtfp_E" (as we'd do for #2) we use "DtfpK_E". As usual, feedback is welcome. Thanks, Daveed
Attachment:
SFINAE_diffs-3
Description: Binary data