Re: PATCH: undefined symbols in shared libraries

["H. J. Lu" <hjl at lucon dot org>]

On Fri, Mar 12, 2004 at 01:29:30PM +1030, Alan Modra wrote:
> On Thu, Mar 11, 2004 at 01:18:08PM -0800, H. J. Lu wrote:
> > -	       ((info->shared && info->unresolved_syms_in_shared_libs == RM_GENERATE_ERROR)				\
> > -		|| (!info->shared && info->unresolved_syms_in_objects == RM_GENERATE_ERROR)				\
> > +	       (info->unresolved_syms_in_objects == RM_GENERATE_ERROR							\
> >  		|| ELF_ST_VISIBILITY (h->other))									\
> 
> Ick, this macro needs trimming down to 80 columns.

Can we turn it into a function?

> 
> I don't think this is correct, nor was the test correct before.  It
> seems wrong to me that we're testing info->shared here (and also
> !info->executable a little earlier), as that just confuses the meaning

That is correct since when building DSO, undefined reference is allowed
by default.

> of info->unresolved_syms_in_shared_libs and
> info->unresolved_syms_in_objects.  This is all completely untested,
> but I think we should be using
> 
>       else if (h->root.type == bfd_link_hash_undefweak)
> 	;
>       else if (((h->elf_link_hash_flags & ELF_LINK_HASH_REF_REGULAR) == 0
> 		|| info->unresolved_syms_in_objects == RM_IGNORE)
> 	       && ((h->elf_link_hash_flags & ELF_LINK_HASH_REF_DYNAMIC) == 0
> 		   || (info->unresolved_syms_in_shared_libs == RM_IGNORE))
> 	       && ELF_ST_VISIBILITY (h->other) == STV_DEFAULT)

H is referenced in the input object. Why do we care about DSO here?

	else if (!info->executable
		&& info->unresolved_syms_in_objects == RM_IGNORE
		&& ELF_ST_VISIBILITY (h->other) == STV_DEFAULT)
		;

looks OK to me.

> 	;
>       else
> 	{
> 	  bfd_boolean err;
> 	  err = (((h->elf_link_hash_flags & ELF_LINK_HASH_REF_REGULAR) != 0
> 		  && info->unresolved_syms_in_objects == RM_GENERATE_ERROR)
> 		 || ((h->elf_link_hash_flags & ELF_LINK_HASH_REF_DYNAMIC) != 0
> 		     && (info->unresolved_syms_in_shared_libs
> 			 == RM_GENERATE_ERROR))
> 		 || ELF_ST_VISIBILITY (h->other) != STV_DEFAULT);
> 
> 	  if (! info->callbacks->undefined_symbol (info, h->root.root.string,
> 						   input_bfd, input_section,
> 						   rel->r_offset, err))
> 	    return FALSE;
> 	  warned = TRUE;
> 	}

I believe the symbol is either defined or referenced when this macro
is used. I think my patch is correct. 

> 
> You might then need to tweak lexsup.c to set unresolved_syms_in_*

I think the current setting is OK.

> differently.  Note the comments in bfdlink.h disagree with the way ld
> currently sets these flags.

I think it meant to say executables and shared libraries instead of
static binaries and dynamic binaries. You can't have unresolved
references in dynamic binaries.


H.J.