Re: Misleading error message from lt_dlopen()

[Jeff Squyres]

On Oct 24, 2008, at 3:54 PM, Ralf Wildenhues wrote:

> For fixing this, it would be really helpful to have a test case,
> anything that exposes the bug and is otherwise as simple/portable
> as possible.  "regression" screams for "testsuite exposure".
>
> If you could provide that, it would save us a lot of time.
> If you are even adventurous to learn Autotest, be our guest.  ;-)


I'm afraid that I wasn't adventurous enough to learn Autotest, but  
here's a tarball that shows the problem:

- it build a "main" program that basically calls lt_dlinit()/ 
ltdl_openext() on argv[1].
- it builds 2 DSOs: good-plugin and bad-plugin
- good-plugin will lt_dlopen() successfully
- bad-plugin will fail to lt_dlopen() because of a missing symbol
- simple build: ./configure && make -j 4
- "good" run: ./main good-plugin.la
  --> loads successfully
- "bad" run: ./main bad-plugin.la
  --> fails to load, and lt_dlerror() shows "file not found", not the  
"symbol missing" error

The tarball ended up being a bit big, so I put it here rather than  
attaching:

    http://www.open-mpi.org/~jsquyres/libtool/

--
Jeff Squyres
Cisco Systems