On 24/02/2011 22:59, H.J. Lu wrote:
> We should never mark entry symbol IR only. I checked in this patch as
> an obvious fix.

> --- a/ld/plugin.c
> +++ b/ld/plugin.c
> @@ -490,8 +490,10 @@ get_symbols (const void *handle, int nsyms, struct ld_plugin_symbol *syms)
>  	 even potentially-referenced, perhaps in a future final link if
>  	 this is a partial one, perhaps dynamically at load-time if the
>  	 symbol is externally visible.  */
> -      ironly = !is_visible_from_outside (&syms[n], owner_sec, blhe)
> -	&& !bfd_hash_lookup (non_ironly_hash, syms[n].name, FALSE, FALSE);
> +      ironly = (!is_visible_from_outside (&syms[n], owner_sec, blhe)
> +		&& !bfd_hash_lookup (non_ironly_hash, syms[n].name,
> +				     FALSE, FALSE)
> +		&& strcmp (syms[n].name, entry_symbol.name) != 0);

  This caused a bunch of regressions for me:

> +FAIL: plugin claimfile resolve symbol
> +FAIL: plugin claimfile replace file
> +FAIL: plugin ignore lib
> +FAIL: plugin claimfile replace lib

  It turns out that entry_symbol.name can be NULL, and strcmp doesn't have to
handle null pointers gracefully; it segfaulted on cygwin.  We need to guard
the test, like the attached.

ld/ChangeLog:

2011-02-25  Dave Korn  <....

	* plugin.c (get_symbols): Guard against NULL name of entry_symbol.

  OK?

    cheers,
      DaveK