Re: [PATCH, v3] Potential solution to librt issue.

[Martin Sebor <sebor at roguewave dot com>]

Paolo Carlini wrote:
> Replying to myself... ;)
>
> > I'm still wondering if we could somehow
> > have a mechanism working *inside* the .so, like having
> > clock_gettime considered weak (Martin ;) inside the .so, and
> > then the linking of librt triggering at the end of the linking
> > process if the symbol is actually referenced somewhere in the
> > user code and remained undefined. Is that vague idea clear,
> > does it make sense (I'm asking the experts)?
>
> Would it work *not* including <time.h> in chrono.cc, instead declaring it "by hand" decorated with the "famous" weakref attribute. Then the .so would export a __clock_gettime symbol, which essentially would wrap everything is now in the body of definition of system_clock::now. In the header, the inline definition of system_clock::now just forwards to __clock_gettime.

IIUC, what you're suggesting is:

Declare in chrono.cc (part of libstdc++.so):

    static int __clock_gettime (clockid_t, struct timespec*)
        __attribute__((weakref ("clock_gettime")));

and change all references to clock_gettime() to something like

    if (__clock_gettime)
        __clock_gettime(clk_id, &tspec);
    else
        perror("must link with librt.so")

Then users of chrono would get a runtime error unless they
linked with librt.so.

Is this close to what you have in mind?

Martin

> The above, + Paolo's mechanism in the specs, - the unconditional linking of librt (added a few days ago).
>
> Paolo.