Paul Eggert wrote:
> If you're worried about other systems doing (d) you can run the attached
> check-undef script to see whether it's possible there.
Thanks a lot. I did that on the include files of various systems, and
found this:
1) On Cygwin:
cygwin-2.9.0/usr/include/w32api/ws2tcpip.h:#define gai_strerror
__MINGW_NAME_AW(gai_strerror)
cygwin-3.5.3/usr/include/w32api/ws2tcpip.h:#define gai_strerror
__MINGW_NAME_AW(gai_strerror)
cygwin-2.9.0/usr/include/w32api/_mingw_unicode.h:# define __MINGW_NAME_AW(func)
func##W
cygwin-2.9.0/usr/include/w32api/_mingw_unicode.h:# define __MINGW_NAME_AW(func)
func##A
These macros expand to a single token; therefore this case is harmless.
2) These seem relevant:
msvc14/usr/include/WDBGEXTS.H:#define dprintf
(ExtensionApis.lpOutputRoutine)
minix-3.3.0/usr/include/x86/mutex.h:#define mtx_lock
u.s.mtxs_lock
netbsd-10.0/usr/include/x86/mutex.h:#define mtx_lock
u.s.mtxs_lock
(for MSVC/clang and NetBSD).
3) Some others:
glibc-2.3.6/usr/include/tclPlatDecls.h:#define strncasecmp \
glibc-2.3.6/usr/include/tclPlatDecls.h:#define strcasecmp \
hpux-11.31/usr/include/sys/strenv.h:#define time streams_time()
irix-6.5/usr/include/shlib.h:#define malloc (* _libc_malloc)
irix-6.5/usr/include/shlib.h:#define free (* _libc_free)
irix-6.5/usr/include/shlib.h:#define realloc (* _libc_realloc)
We can ignore these, as they are in very specialized include files.
> Simpler would
> be checking that the argument of _GL_WARN_ON_USE is a single identifier
> rather than some complicated expression, as this should prevent any
> plausible problems arising from (d).
>
> The attached proposed patches do that. The first patch does the real
> work
I tested these invocations:
_GL_WARN_ID_LIKE (foo)
_GL_WARN_ID_LIKE (__)
_GL_WARN_ID_LIKE (77)
_GL_WARN_ID_LIKE (u.v.w)
_GL_WARN_ID_LIKE (u;v;w)
_GL_WARN_ID_LIKE (({ int a = c; a; }))
The last one produces a diagnostic, but unfortunately
_GL_WARN_ID_LIKE (u.v.w)
_GL_WARN_ID_LIKE (u;v;w)
don't give a warning or error.
> the second one is a near-copy of the patch I emailed earlier.
In summary, I think
- The first patch is a brilliant idea, but since it does not catch some
significant cases, I would not commit it.
- The second patch is fine, except for 'dprintf' and 'mtx_lock'. For these
two, it's better to keep the '#undef'.
Bruno
