[PATCH v3 11/18] BFD: Correct archive format determination for non-default link emulations

[Maciej W. Rozycki]

On Wed, 10 Dec 2025, Alan Modra wrote:

> > > Looks plausible to me, but given the history could really do with Alan also
> > > taking a look. And of course ...
> > 
> >  FWIW I yet want to go back to the place of the segmentation fault and add 
> > an assertion there, as it's a format mismatch case, i.e. an a.out handler 
> > handed an ELF BFD, which is obviously not a user error.  Somehow it didn't 
> > struck me right away to do that.  It can/will be a change on its own of 
> > course.
> 
> I think we should keep the target_defaulted test.  The idea being that
> if a user supplies a target, then that should be only target allowed.

 So far so good.

> The problem in this case is that ld does set the target (via
> ldlang.c:3714 s->input_statement.target = current_target), and that
> goes back to commit 2fa0b342a5cd in 1991.  So a better solution might
> be to change the test to
> 
>   if ((abfd->target_defaulted || abfd->is_linker_input)
>       && bfd_has_map (abfd))

 I think the problem here is that LD also sets the input target in the 
absence of any explicitly setting supplied by the user, in which case I 
think no default target ought to be implied.  Then such hacks wouldn't be 
needed in BFD anymore.  But that would require a code audit and a rewrite 
large enough I'd rather not do on this occasion.  So I agree that sticking 
to the hacks is probably the most efficient approach to move forwards.

> That does happen to fix the segfault.  Even if it didn't, a new
> test that exposes a segfault in the aout backend due to code using
> aout target data before testing that the bfd was actually aout, is
> fine.

 Included in v4, as 05/15.

  Maciej