Re: [boost] [atomic] Support for specialized instructions

Subject: Re: [boost] [atomic] Support for specialized instructions
From: Andrey Semashev (andrey.semashev_at_[hidden])
Date: 2013-11-21 06:48:34

On Thu, Nov 21, 2013 at 1:04 PM, Tim Blechmann <tim_at_[hidden]> wrote:
>
> hmmm, i see your points ... however in a way, i'd prefer if this
> functionality could be built upon boost.atomic, mainly to keep API
> compatibility with std::atomics.

I suppose, the extensions could be made just as functions. What I
don't like about this approach is the need to cast pointers to
atomic<> to pointers to the underlying integers. We have a few places
in Boost.Sync with such code and it really bothers me. Do you think we
could at least add a method to get the pointer or reference to the
internal storage inside atomic<>?

> with compiler-support for c++11 atomics, the first two instructions
> could probably be generated by a smart compiler ...

Although theoretically possible, I don't think that current compilers
are able to transform a CAS loop with an operation to an in-place
operation. Especially, if CAS is implemented in inline assembler. It's
a little easier with inc/dec but in practice I've never seen such
transformation done automatically.

> bts/btr may be
> useful, but i suppose they are rather specific to x86? i wonder, how
> would they map to arm?

AFAIK, ARM implements atomic ops with LL/SC instructions, so it should
be flexible enough to implement it. I'm not very familiar with the
architecture though.

> but, yes ... if you think it is reasonable to add them, please go ahead!

Thanks.