[trans-mem] handle invisible references for decls (PR/47554)

[Richard Henderson]

On 02/04/2011 12:08 PM, Aldy Hernandez wrote:
> +      if (DECL_BY_REFERENCE (x))
> +	{
> +	  /* Stupid GCC magic.  In assign_params() we'll see the
> +	     DECL_BY_REFERENCE and put the whole thing in a register,
> +	     so we can't instrument this by calculating &<retval>.  We
> +	     will see the resulting INDIRECT_REF elsewhere and
> +	     instrument that.  See... this is why GCC hackers will
> +	     always have a job, cause nobody understands this and
> +	     takes too long to rewrite it.  */
> +	  return false;

Ha ha.

A better comment might be

  /* ??? This value is a pointer, but aggregate_value_p has been
     jigged to return true which confuses needs_to_live_in_memory.
     This ought to be cleaned up generically.  */

Ok with that change.


I sincerely believe that the test in aggregate_value_p is a mistake.
We should not be lying about the fact that this value is a pointer.
But looking at how DECL_BY_REFERENCE is used elsewhere, it'll be a
hard one to unravel.  And that certainly should not be done on this
branch.


r~