Re: [PATCH] Fix part of pr25505

[Josh Conner <jconner at apple dot com>]

Diego Novillo wrote:
> Diego Novillo wrote on 08/31/06 18:59:
> 
>>   FOR_EACH_SSA_DEF_OPERAND (def_p, stmt, op_iter, SSA_OP_ALL_DEFS)
>>     {
>>       tree def = DEF_FROM_PTR (def_p);
>>       def = SSA_NAME_VAR (def);
>>       if (TREE_TYPE (def) == TREE_TYPE (call)
>>           && MTAG_P (def)
>>           && is_call_clobbered (def))
>>         goto unsafe;
>>     }
>>
>> Note that I'm not 100% sure if this type check is actually valid.  But I
>> think it is.
>>
> Never mind.  Silly idea.  The individual tags will *not* have the type
> of the call.  Sigh.

Now that I understand how to generate these trees, I've started looking
at a solution.  Here's my general approach -- it's not perfect (it may
claim some things as clobbered that a finer investigation or pointer
aliasing would reveal aren't), but it's a whole lot better than what's
there now.  If this seems like a reasonable direction, I'll get a formal
patch for review and run tests.

static unsigned int
dest_safe_for_nrv_p (tree dest)
{
  switch (TREE_CODE (dest))
    {
      case VAR_DECL:
        {
          subvar_t subvar;
          if (is_call_clobbered (dest))
            return 0;
          for (subvar = get_subvars_for_var (dest);
               subvar;
               subvar = subvar->next)
            if (is_call_clobbered (subvar->var))
              return 0;
          return 1;
        }
      case ARRAY_REF:
      case COMPONENT_REF:
        return dest_safe_for_nrv_p (TREE_OPERAND (dest, 0));
      default:
        return 0;
    }
}

- Josh