Re: A patch for elf32.em (Re: GNU/Linux vs. libtool --no-undefined)

[Alexandre Oliva <aoliva at redhat dot com>]

On Feb  7, 2001, Ian Lance Taylor <ian@zembu.com> wrote:

> The original error appears to be warning about an undefined symbol
> reference from a shared library.  Is that so?  If it is, then I think
> the linker should be fixed to not warn about undefined references from
> shared libraries, even when --no-undefined is used.  We care whether
> each reference is satisfied somewhere.  I don't think we care whether
> each shared library included in the link has all symbols satisfied.
> Do we?

Assume I have a component library A, that is not self-contained, but
rather, it contains a few symbols that are supposed to be defined by
its users.  Now assume I want to create a library B, with another
component that builds upon A, defining all symbols that are missing in
A, plus a number of other symbols that implement B's components.

I want B to be a complete dlopenable module, i.e., one that doesn't
depend on any external symbols that are not provided by libraries it's
linked with.  So, I link B with --no-undefined --export-dynamic.

During the development phase, I keep A as a PIC static library, just
so that builds are faster (not having to compile PIC and non-PIC
versions of its object files).  Various times, --no-undefined detects
that symbols are missing when linking B, because of various changes in
A and B.

At some point, someone suggests I'd save on building time if I created
A as a shared library, because then the link step of B wouldn't have
to copy pretty much everything from A into B; the dynamic linker would
take care of that.  And that's indeed the case.

When my program dlopens B (with RTLD_LAZY, of course :-), it works
fine, and I'm a happy camper.  Until, some day, after refactoring some
components in A, I exercise some seldom-exercised feature, and the
program crashes.

Some investigation leads to the conclusion that B had an undefined
symbol that it expected to get from A, but that I had forgotten to
define in A.

Why didn't --no-undefined work?  Should the fact that A is shared or
static make any difference?

-- 
Alexandre Oliva   Enjoy Guarana', see http://www.ic.unicamp.br/~oliva/
Red Hat GCC Developer                  aoliva@{cygnus.com, redhat.com}
CS PhD student at IC-Unicamp        oliva@{lsd.ic.unicamp.br, gnu.org}
Free Software Evangelist    *Please* write to mailing lists, not to me