Re: [boost] [atomic] Generalized read-modify-write operation

Subject: Re: [boost] [atomic] Generalized read-modify-write operation
From: Andrey Semashev (andrey.semashev_at_[hidden])
Date: 2015-09-01 08:24:48

On Tue, Sep 1, 2015 at 2:14 PM, Giovanni Piero Deretta
<gpderetta_at_[hidden]> wrote:
> On 1 Sep 2015 10:35 am, "Andrey Semashev" <andrey.semashev_at_[hidden]> wrote:
> [...]
>> I know there's probably no way to ensure that the function meets these
>> requirements, and it is possible to get an infinite loop in runtime,
>> so maybe always using a CAS loop would be a safer approach, even on an
>> architecture with LL/SC instructions.
>
> The implementation could fall back on the CAS loop after the first failed
> LL/SC. You might want to do it even in the case if contention to reduce the
> window the cachelines must be held. It is not optimal in case that the
> reservation always fail, but is always correct and the performance can be
> corrected with profiling and some sort of CAS hint.

Thanks for the suggestion. Yes, that's probably the best way to
implement it. I can provide two functions in the API - one that always
implements a CAS loop right away and the other that might try
injecting modify function between LL and SC and do the CAS loop if
that fails. On a CAS-only architecture the two functions would be
equivalent.