On Fri, Jan 6, 2012 at 6:05 PM, Aldy Hernandez <al...@redhat.com> wrote:
> On 01/05/12 09:36, Richard Guenther wrote:
>>
>> On Thu, Jan 5, 2012 at 3:48 PM, Aldy Hernandez<al...@redhat.com> wrote:
>>>
>>> As you suggested here:
>>>
>>> http://gcc.gnu.org/ml/gcc-patches/2012-01/msg00232.html
>>>
>>> Bootregged on x86-64 Linux.
>>>
>>> OK?
>>
>>
>> No, this pessimizes things too much at gimplification time, please simply
>> inline the predicate instead.
>>
>> Richard.
>
>
> It wasn't clear from your other message whether you want me to replace the
> call with:
>
> !is_gimple_reg () && !TREE_ADDRESSABLE ()
>
> or outright inline the exact function call as we have now.
>
> I have done the later. Let me know if you want the former.
>
> Tested on x86-64 Linux.
>
> OK?
Ok with simply excluding SSA_NAMEs like
@@ -4411,16 +4411,27 @@ gimplify_modify_expr_rhs (tree *expr_p,
/* It's OK to use the target directly if it's being
initialized. */
use_target = true;
- else if (TREE_CODE (*to_p) != SSA_NAME
&& ....)
- /* Don't use the original target if it's already addressable;
- if its address escapes, and the called function uses the
- NRV optimization, a conforming program could see *to_p
- change before the called function returns; see c++/19317.
- When optimizing, the return_slot pass marks more functions
- as safe after we have escape info. */
- use_target = false;
- else
- use_target = true;
as SSA_NAMEs can never become addressable during gimplification.
Thanks,
Richard.
