[PATCH] Fix part of pr25505

[Daniel Berlin]

>
> if (VAR_DECL)
>   check decl and subdecls for call-clobber
> else
>   { existing logic here }
>
> How is that going to give strictly worse results?

No, I mean that the the SSA_NAME logic that lives in "existing logic
here" will give strictly worse results than it needs to, because by
doing

def = SSA_NAME_VAR (def)

if is_call_clobbered (def)
  goto bad;

instead of

def = SSA_NAME_VAR (def)
if memory_tag (def) && is_call_clobbered (def)
  goto bad;

it will ask about random VAR_DECL that happen to be the base of SSA
NAMES, which is nonsensical.

>
> - Josh
>
>
>