Re: solaris 10: address@hidden

[David Lee]

On Wed, 29 Jun 2005, Ralf Wildenhues wrote:

> * David Lee wrote on Tue, Jun 28, 2005 at 07:10:59PM CEST:
>
> > (Incidentally: this may or may not be relevant to the above:  in my own
> > case, I know that our GNU ld and libtool are somewhat out-of-date.  So
> > I've attempted to get libtool 1.5.18 installed.  "make check" is fine on
> > S8, but has a failure in "mdemo2-make.test" on S10.  If someone can guide
> > me, I'd be happy to try to pursue this futher.)
>
> I bet this is related, it tests dlopening.
>
> Execute
>  make check VERBOSE=x TESTS='mdemo-conf.test mdemo-make.test mdemo2-conf.test 
> mdemo2-make.test mdemo2-exec.test'
> and post the output (This is a subset of the check above).

(See earlier email for that output.  Also, Ralf and I had a very useful 
off-list email chat.  I think it is now time to come back on-list.)

Possible progress.

Over the weekend, someone on the Sun forum:
   http://forum.sun.com/thread.jspa?threadID=24107&tstart=0

dug a little deeper with their applications, indicating that a generated 
setting "build_libtool_need_lc=yes" may be problematical on S10.  Setting 
it to "no" seemed to improve things.  So I have tried something similar on 
my application (heartbeat, a.k.a. linux-ha) which also seems to look 
better.

I should probably now try this on libtool-1.5.18 and its similar failure 
on S10 in "mdemo2-make.test", but don't know the best way to do this, so 
am looking for some guidance.

The basic idea of the fix (workaround, fudge) is to try to strip out all 
traces of "-lc" and "-ldl" when building shared objects (".so").

Now I don't the insides of libtool at all (not in the slightest!).  So 
what follows is supposition.

I suspect that the problematic code is in "mdemo2/acinclude.m4" around 
line 6000:
   AC_MSG_CHECKING([whether -lc should be explicitly linked in])

which is generating a "yes", which I then think becomes the value of 
variable "build_libtool_need_lc" in the generated "libtool".

I think there is some overloading happening, with this result being used 
in two different contexts which may require different values.  This code 
(mdemo2/acinclude.m4:6000) is using the result of linking a final, 
runnable progam (which probably _does_ need "-lc" and/or "-ldl") to set a 
"yes" (probably needed for final programs).  But that result also seems to 
be used for linking shared-objects (".so"), and it seems that (for reasons 
yet unknown) this probably should _not_ be done for S10 (fine on earlier 
releases).

Does that seem plausible?

Could someone who knows libtool possibly suggest some possible experiments 
(e.g. edits, pseudo-patches) that I, a complete novice to libtool 
internals, might try?  Do we need an additional variable alongside that 
"build_libtool_need_lc", to indicate how to build shared-objects?

I'm using GNU/ld; but I'd be happy,as part of the experiments, to try to 
redirect our "gcc" to use the Solaris/ld.  (How to redirect it?)




--

:  David Lee                                I.T. Service          :
:  Senior Systems Programmer                Computer Centre       :
:                                           Durham University     :
:  http://www.dur.ac.uk/t.d.lee/            South Road            :
:                                           Durham DH1 3LE        :
:  Phone: +44 191 334 2752                  U.K.                  :