[PATCH] LTO: Handle __real_SYM reference in IR

[Alan Modra]

On Fri, Apr 29, 2022 at 10:59:32AM -0700, H.J. Lu via Binutils wrote:
> --- a/bfd/elflink.c
> +++ b/bfd/elflink.c
> @@ -1101,8 +1101,19 @@ _bfd_elf_merge_symbol (bfd *abfd,
>    if (! bfd_is_und_section (sec))
>      h = elf_link_hash_lookup (elf_hash_table (info), name, true, false, false);
>    else
> -    h = ((struct elf_link_hash_entry *)
> -	 bfd_wrapped_link_hash_lookup (abfd, info, name, true, false, false));
> +    {
> +      h = ((struct elf_link_hash_entry *)
> +	   bfd_wrapped_link_hash_lookup (abfd, info, name, true, false,
> +					 false));
> +      /* Mark SYM is referenced in IR if __real_SYM is referenced in
> +	 IR.  */
> +      if (h != NULL
> +	  && info->wrap_hash != NULL
> +	  && (abfd->flags & BFD_PLUGIN) != 0
> +	  && strcmp (h->root.root.string,
> +		     name + sizeof "__real_" - 1) == 0)
> +	h->root.ir_ref_real = 1;
> +    }
>    if (h == NULL)
>      return false;
>    *sym_hash = h;

How about putting the above in linker.c:bfd_wrapped_link_hash_lookup
instead, to avoid the extra strcmp?  That seems like a better place to
me since there isn't really anything ELF specific about wrapped symbol
lookup.

Hmm, and maybe without the BFD_PLUGIN test and losing "ir" from the
flag name.

If that works for you, OK with those changes.

-- 
Alan Modra
Australia Development Lab, IBM