Re: Confusing meaning of --no-undefined

[Alan Modra <amodra at bigpond dot net dot au>]

On Tue, Feb 11, 2003 at 05:30:45PM +0000, Nick Clifton wrote:
>   concerning the fact that:
> 
>      ld --share --no-undefined foo.o bar.so
> 
>   does not report an undefined symbol in bar.so.

I think there's general agreement that this is desirable default
behaviour.  ie. --no-undefined should only report undefined references
in the shared lib being built, _not_ in other libraries linked
against.

> *** bfd/elflink.h	4 Feb 2003 12:49:57 -0000	1.204
> --- bfd/elflink.h	11 Feb 2003 17:09:10 -0000
> *************** elf_link_output_extsym (h, data)
> *** 6096,6103 ****
>        referenced by regular files, because we will already have issued
>        warnings for them.  */
>     if (! finfo->info->relocateable
> !       && ! finfo->info->allow_shlib_undefined
> !       && ! finfo->info->shared
>         && h->root.type == bfd_link_hash_undefined
>         && (h->elf_link_hash_flags & ELF_LINK_HASH_REF_DYNAMIC) != 0
>         && (h->elf_link_hash_flags & ELF_LINK_HASH_REF_REGULAR) == 0
> --- 6097,6103 ----
>        referenced by regular files, because we will already have issued
>        warnings for them.  */
>     if (! finfo->info->relocateable
> !       && (! finfo->info->shared || ! finfo->info->allow_shlib_undefined)
>         && h->root.type == bfd_link_hash_undefined
>         && (h->elf_link_hash_flags & ELF_LINK_HASH_REF_DYNAMIC) != 0
>         && (h->elf_link_hash_flags & ELF_LINK_HASH_REF_REGULAR) == 0
> 
>   and sure enough the error message in Jan's test was now correctly
>   generated.  When I ran the linker tests however, I discovered that
>   this patch causes the ld-elfweak tests to fail:

Yeah, this code does look wrong, and I was the last to touch it.  :-(
I think your patch makes sense, but you should also update the comment
and default allow_shlib_undefined to 1.  The other thing is that I see
code like the following from elf32-i386.c:

	  else if (info->shared
		   && (!info->symbolic || info->allow_shlib_undefined)
		   && !info->no_undefined
		   && ELF_ST_VISIBILITY (h->other) == STV_DEFAULT)
	    ;
	  else
	    report an error

This seems like another bogus test of allow_shlib_undefined.  Or at
least it's an entirely different usage only affecting -Bsymbolic libs.
You could accomplish the same effect by not requesting --no-undefined.
ie. I think we want

	  else if (info->shared
		   && !info->no_undefined
		   && ELF_ST_VISIBILITY (h->other) == STV_DEFAULT)
	    ;
	  else
	    report an error

-- 
Alan Modra
IBM OzLabs - Linux Technology Centre