On 11/4/2023 11:44 AM, Chris M. Thomasson wrote:
> On 11/4/2023 8:53 AM, Scott Lurndal wrote:
>> "Chris M. Thomasson" <chris.m.thomasson.1@gmail.com> writes:
>>> On 11/3/2023 3:13 PM, Chris M. Thomasson wrote:
>>>> On 11/3/2023 5:36 AM, Bonita Montero wrote:
>>>>> Am 01.11.2023 um 21:02 schrieb Chris M. Thomasson:
>>>>>
>>>>>> If you can avoid using a CAS, aka CMPXCHG over on intel, then do it!
>>>>>> Although, wrt intel, there is a way to use CMPXCHG without looping.
>>>>>
>>>>> I don't believe that, give me some code-example _here_. The expected
>>>>> value of a compare exchange might always have been asynchronously
>>>>> changed, so you need looping every time.
>>>>>
>>>>>> Think, strong CAS vs weak CAS...
>>>>>
>>>>> Strong CAS vs. weak CAS only makes a difference of architectures with
>>>>> LL/SC. With these architectures a CAS may fail although the actual
>>>>> data corresponds with the expected value because the cacheline might
>>>>> have been evicted in the meantime. compare_exchange_strong checks for
>>>>> this case and loops internally, whereas compare_exchange_weak gives a
>>>>> false result after one try.
>>>>
>>>> Strong CAS cannot fail spuriously. Put on your thinking cap.
>>>>
>>>
>>> CMPXCHG cannot fail spuriously. However, LL/SC can.
>>
>> LL/SC is not compare and swap.
>>
>>
>
> LL/SC can be used to implement a weak CAS. Also, wrt the algorithm, it
> can be used to implement a strong CAS.
>
> CMPXCHG over on intel cannot be used to implement a weak CAS.

Well, there is a "way" wrt the algorithm that can mimic weak CAS even
with CMPXCHG. It involves checking the comparand before using CMPXCHG.
This would not be ideal, but it can be done. I would stay far away from
it. I do not really like optimistic sync, like LL/SC. Imvvho, CMPXCHG is
more, pessimistic, so to speak...