Re: tests on AIX 5

[Ralf Wildenhues]

* Bob Friesenhahn wrote on Tue, Nov 16, 2004 at 12:58:49AM CET:
> On Mon, 15 Nov 2004, Bob Friesenhahn wrote:
> >On Mon, 15 Nov 2004, Ralf Wildenhues wrote:
> >>
> >>Issue 2:
> >>mdemo*-exec.test all fail, because they can lt_dlopen(NULL), but not
> >>lt_dlsym() a symbol from the main program (the libraries are opened and
> >>used fine).
> >>
> >>A cursory search/test showed that it is possible to dlsym() symbols from
> >>the main program by doing one of
> >
> >Is dlsym() of a symbol from the main program (vs a symbol in shared 
> >libraries the program depends on) portable behavior?  Should we expect 
> >libtool/libltdl to support it?
> 
> To follow up on my own posting, it seems like the libltdl 
> documentation should include a summary of the capabilities it 
> claims/expects to portably support, and what it is expected to be used 
> for. Libltdl is not really a portable dlsym() wrapper since dlsym() 
> supports many non-portable behaviors.  In order to support its 
> portability objective, libltdl should support the lowest (acceptable) 
> common denominator functionality, and this functionality should be 
> documented.

This is *not* a sane requirement IMVHO.  This *is* a matter which can be
argued about.  Might do it now, while many new things are envisioned
and argued anyway:

I think we should very much encourage using the portable behavior.
But: If we never allow any features of the not-available-on-100% of all
obsolete and non-obsolete systems, then Libtool is doomed to fail.
Fail in the sense that it will actually become worthless over time.
Many software packages have no chance to be portable to many of the
older systems anyway, yet they can profit from a portable libtool (as
well as maybe a portable libltdl) which enables seamless use on ten to
twenty platforms, which may make up the vast majority of in-use systems.
Obviously, both this point of view and the distinction which system 
and which feature are worth including is to some extent a subjective
matter.

You might argue now that I try to change Libtool's objectives and goals.
I don't think I am really.  If Libtool were to do otherwise, we would
not have shell functions.  We would not have inter-library dependencies.
I don't want to hinder portability.  I just don't want to set in stone
that there cannot be progress.

Also, your argument does not apply to the `lt_dlsym from main' case.
We can always support this by falling back on dlpreloading (and, on AIX,
it's supported by the system anyway).

BTW, I'll demand someone showing me that current libtool works on a
system without shared libraries at all, if the lowest common denominator
is the continuing argument.  I'm ready to bet this hasn't been used long
enough that it broke silently along the way.

Regards,
Ralf