[systemd-devel] compile with clang broken

[Djalal Harouni]

On Fri, Aug 15, 2014 at 10:55:57AM +0200, David Herrmann wrote:
> Hi
> 
> On Thu, Aug 14, 2014 at 8:07 PM, Lennart Poettering
> <lennart at poettering.net> wrote:
> > On Fri, 18.07.14 16:02, Thomas H.P. Andersen (phomes at gmail.com) wrote:
> >
> >> 1716f6dcf54d4c181c2e2558e3d5414f54c8d9ca (resolved: add LLMNR support
> >> for looking up names) broke the build on clang.
> >>
> >> src/resolve/resolved-manager.c:553:43: error: non-const static data
> >> member must be initialized out of line
> >> uint8_t buffer[CMSG_SPACE(MAX(sizeof(struct in_pktinfo), sizeof(struct
> >> in6_pktinfo)))
> >>
> >> Moving the MAX(...) to a separate line fixes that problem but another
> >> error then happens:
> >>
> >> src/resolve/resolved-manager.c:554:25: error: fields must have a
> >> constant size: 'variable length array in structure' extension will
> >> never be supported
> >> uint8_t buffer[CMSG_SPACE(size)
> >>
> >> We have encountered the same problem before and Lennart was able to
> >> write the code in a different way. Would this be possible here too?
> >
> > My sugegstion here would be to maybe rewrite the MAX() macro to use
> > __builtin_constant_p() so that it becomes constant if the params are
> > constant, and only uses code block when it isn't. Or so...
> >
> > http://lists.freedesktop.org/archives/systemd-devel/2014-August/021912.html
> 
> Hm, I don't know whether that works. See the description here:
>     https://gcc.gnu.org/onlinedocs/gcc-4.3.3/gcc/Other-Builtins.html
> 
> What you propose is something like my attached patch, I guess? Along
> the lines of:
>     (__builtin_constant_p(A) && __builtin_constant_p(B)) ?
>         ((A) > (B) ? (A) : (B)) :
>         ({ ....OLD_MAX.... })
> 
> Thing is, the ELSE case is not considered a compile-time constant by
> LLVM. Therefore, the whole expression is not considered a compile-time
> constant. I don't know whether conditions with __builtin_constant_p()
> are evaluated at the parser-step. The GCC example replaces the ELSE
> case with -1, effectively making both compile-time constants.
Sorry I didn't follow the thread, but:

Actually and IIRC it is more complicated, __builtin_constant_p() can
be computed during the SSA (optimization) passes on the GIMPLE form of
the code... so it depends on the code and parameters passed to
__builitin_constant_p(), not only preprocessor constants.

https://gcc.gnu.org/onlinedocs/gccint/Tree-SSA.html


-- 
Djalal Harouni
http://opendz.org