Re: [Qemu-devel] [PATCH] atomic: using memory_order_relaxed for refcnt inc/dec ops

[liu ping fan]

On Sat, Jul 13, 2013 at 12:24 AM, Richard Henderson <address@hidden> wrote:
> On 07/11/2013 11:32 PM, Liu Ping Fan wrote:
>> Refcnt's atomic inc/dec ops are frequent and its idiom need no seq_cst
>> order. So to get better performance, it worth to adopt _relaxed
>> other than _seq_cst memory model on them.
>
> You'd need to update the documentation then.  As it stands, what you've 
> written
> looks like a bug.
>
Ok, will update atomic.txt

>> +#ifndef _GLIBCXX_ATOMIC_BUILTINS
>
> This will never be defined.  It's private to the libstdc++ implementation.  
> See
> how we've defined things using __atomic elsewhere in the file, looking at one
> of the __ATOMIC defines.
>
Oh, I misunderstood the description of _GLIBCXX_ATOMIC_BUILTINS.  Got
it, thanks.

> And in either case, it's better form to use positive tests than negative ones.
> I.e. #ifdef rather than #ifndef
>
Will fix.
>>  #define atomic_fetch_inc(ptr)  __sync_fetch_and_add(ptr, 1)
>>  #define atomic_fetch_dec(ptr)  __sync_fetch_and_add(ptr, -1)
>
> I'd prefer atomic_fetch_inc_relaxed, as that's more self-documenting.
>
But if _relaxed not supported, it will fall back on _seq_cst, then
atomic_fetch_inc_relaxed will be wrong named?

Thanks and regards,
Pingfan

> But I'll re-iterate the necessity of documentation in this area.
>
>
> r~