Re: [PATCH 07/12] x86: use cmpxchg_flag() where applicable

[Jeremy Fitzhardinge]

On 08/24/2011 03:01 PM, H. Peter Anvin wrote:
> On 08/24/2011 02:56 PM, Linus Torvalds wrote:
>> Ok, I see nothing horrible in this series.
>>
>> The one reaction I have is that the cmpxchg_flag() thing returns an
>> 8-bit value, but then a lot of the users end up having to extend it to
>> a full "int" purely for calling convention reasons (eg I think
>> 'down_write_trylock()' will have 'sete + movzl' - not a new problem,
>> but since the whole point was to remove extraneous instructions and we
>> no longer have the silly 'testl', it now annoys me more).
>>
>> So it seems a bit sad. But I guess it doesn't really matter.
>>
> I think it is a net lose.  The most common case is probably going to be
> to use it immediately, in which case we have:
>
> cmpxchg -> sete -> compare -> conditional
>
> versus
>
> cmpxchg -> compare -> conditional
>
> For doubleword cmpxchg it's another matter entirely, because doubleword
> comparisons are significantly more expensive that sete + test.
>
> So unless there is actual data showing this is better, I would like to
> see this dropped for now.

Well, we could keep the API (since it is convenient), but just implement
it with a compare.

    J
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/