devel/readline: Fix build for dependent ports.

[des]

Summary

The previous commit made a bad situation worse by replacing declarations
with incorrect prototypes, thus breaking any software that calls
rl_message() and doesn't know the secret sauce to get the correct
prototype (e.g. devel/rlwrap). Fix this by dropping K&R support and
instead offering an annotated prototype for compilers that understand
the GCC format attribute and an unannotated one for those that don't.

Fixes: ee994524fffd
PR: 281633

[des]

Just noticed the second half of the patch...

[arrowd]

devel/readline/files/patch-clang18

8

The rlstdc.h header contains this:

/* Moved from config.h.in because readline.h:rl_message depends on these
   defines. */
#if defined (__STDC__) && defined (HAVE_STDARG_H)
#  define PREFER_STDARG
#  define USE_VARARGS
#else
#  if defined (HAVE_VARARGS_H)
#    define PREFER_VARARGS
#    define USE_VARARGS
#  endif
#endif

So we should end with USE_VARARGS and PREFER_STDARG defined and this replacement shouldn't be needed.

[des]

Only if the translation unit that includes <readline/readline.h> also includes <stdarg.h>, which is required to define rl_message() since it uses va_start() etc but should not be required to merely call it.

[des]

Sorry, I misread HAVE_STDARG_H as STDARG_H_INCLUDED which would indicate that <stdarg.h> has been included. This is even worse: it requires the caller to use something like autoconf and to explicitly check for the existence of <stdarg.h> which, again, is not needed to simply use libreadline.

[arrowd]

Oh, right, this is indeed silly from the readline side.

[sunpoet]

LGTM but I'd like to ask dim@ to take a look. Thanks.

[dim]

For Function, VFunction etc this goes back to the old situation, where for C the prototypes do not say anything about their arguments, which would not really solve the issue with samba (as in https://bugs.freebsd.org/bugzilla/show_bug.cgi?id=281776). The problem there was that it didn't use rl_completion_func_t as it ought to have done.

I don't really like the #if defined(__GNUC__) || defined(__clang__) trick, but it is probably the easiest way to work around the detection of varargs support. In fact you could maybe make it unconditional, as I expect the number of people building FreeBSD with a compiler without varargs support to be zero. :)

[des]

It's not trying to detect varargs support, it's trying to detect support for __attribute__((__format__)).