Feature request: stdatomic.h and threads.h

Feature request: stdatomic.h and threads.h

[Daniel Cegiełka]

Hi,

http://www.openwall.com/lists/musl/2013/06/29/17

Is there a plan to add support for stdatomic.h and threads.h?

Daniel

Re: Feature request: stdatomic.h and threads.h

[Szabolcs Nagy]

* Daniel Cegie?ka <daniel.cegielka@gmail.com> [2013-11-21 11:31:35 +0100]:
> Is there a plan to add support for stdatomic.h and threads.h?

how would that work?
there is still no c11 atomic support in gcc

(there is an experimental c11 atomic branch with a gcc
provided stdatomic.h which uses some gcc extensions,
but there are still compiler issues eg there is no
solution for _Atomic float yet because x+=0.1 requires
fenv manipulation without libc support and that means
new asm for all archs, meanwhile wg14 and wg21 are still
busy filling the holes in the specs..)

Re: Feature request: stdatomic.h and threads.h

[Jens Gustedt]

[-- Attachment #1: Type: text/plain, Size: 2228 bytes --]

Hi,

Am Donnerstag, den 21.11.2013, 15:16 +0100 schrieb Szabolcs Nagy:
> * Daniel Cegie?ka <daniel.cegielka@gmail.com> [2013-11-21 11:31:35 +0100]:
> > Is there a plan to add support for stdatomic.h and threads.h?
> 
> how would that work?
> there is still no c11 atomic support in gcc
> (there is an experimental c11 atomic branch with a gcc
> provided stdatomic.h which uses some gcc extensions,
> but there are still compiler issues eg there is no
> solution for _Atomic float yet because x+=0.1 requires
> fenv manipulation without libc support and that means
> new asm for all archs,

As far as I understood this will be merged in the next gcc version

> meanwhile wg14 and wg21 are still busy filling the holes in the
> specs..)

I haven't seen much about atomics discussed there.

For all that is integer type atomics at least the support in gcc is
sufficient to implement everything that is function call based (so not
the operations like += etc).

In P99 I have a shallow wrapper on top of the __sync_ builtins that
implements all type generic macros in atomic.h. It works fine for
everyday's use and efficiently produces optimal assembler, AFAIKS.

Restriction besides the operations mentioned above are

 -  variable declaration. Out of the two forms

    _Atomic int toto;
    _Atomic(int) toto;

    I only support the later (as a macro) if the compiler doesn't
    implement the first.

 - the different modes of memory consistency are not supported and
   always resolve to the strongest possible consistency

So yes a good deal of the *library* support is possible directly, even
with oldish versions of gcc, and can be implemented conforming to the
specs.

Also, I discovered recently that icc now has full support for
atomic.h, so definitively this is something that will see increased
use in a relatively near future.

Jens


-- 
:: INRIA Nancy Grand Est :: http://www.loria.fr/~gustedt/   ::
:: AlGorille ::::::::::::::: office Nancy : +33 383593090   ::
:: ICube :::::::::::::: office Strasbourg : +33 368854536   ::
:: ::::::::::::::::::::::::::: gsm France : +33 651400183   ::
:: :::::::::::::::::::: gsm international : +49 15737185122 ::



[-- Attachment #2: This is a digitally signed message part --]
[-- Type: application/pgp-signature, Size: 198 bytes --]

Re: Feature request: stdatomic.h and threads.h

[Daniel Cegiełka]

Jens, I know your P99 :) I even have to mention it, but you were faster.

I was convinced that the work on this are much more advanced, but it
seems that we have to wait for gcc-4.9:

http://gcc.gnu.org/ml/gcc-patches/2013-11/msg00493.html


2013/11/21 Jens Gustedt <jens.gustedt@inria.fr>:

> Also, I discovered recently that icc now has full support for
> atomic.h, so definitively this is something that will see increased
> use in a relatively near future.
>
> Jens

btw. I changed glibc to musl libc in icc... using symlink or
permanently with hex editor.. +small patches :) It works very well. I
even built a system base in this way (but compiled code is a bit
larger compared to gcc).

thx,
Daniel

Re: Feature request: stdatomic.h and threads.h

[Rich Felker]

On Thu, Nov 21, 2013 at 11:31:35AM +0100, Daniel Cegiełka wrote:
> Hi,
> 
> http://www.openwall.com/lists/musl/2013/06/29/17
> 
> Is there a plan to add support for stdatomic.h and threads.h?

Yes, but not before 1.0. stdatomic.h is probably not hard if you
assume GCC __sync_* builtins (or it could be written in an
arch-specific way), but threads.h is much more complicated and
involves ABI decisions where I'm partly waiting to see what glibc
does.

Rich

Re: Feature request: stdatomic.h and threads.h

[Daniel Cegiełka]

2013/11/21 Rich Felker <dalias@aerifal.cx>:
> On Thu, Nov 21, 2013 at 11:31:35AM +0100, Daniel Cegiełka wrote:
>> Hi,
>>
>> http://www.openwall.com/lists/musl/2013/06/29/17
>>
>> Is there a plan to add support for stdatomic.h and threads.h?
>
> Yes, but not before 1.0. stdatomic.h is probably not hard if you
> assume GCC __sync_* builtins (or it could be written in an
> arch-specific way),

like for stdalign.h

http://git.musl-libc.org/cgit/musl/tree/include/stdalign.h

I use a few tricks to support atomic operations (with __sync_*), but
C(11) has a huge backlog vs c++.

> but threads.h is much more complicated and
> involves ABI decisions where I'm partly waiting to see what glibc
> does.
>
> Rich

thx,
Daniel

Re: Feature request: stdatomic.h and threads.h

[Daniel Cegiełka]

2013/11/21 Rich Felker <dalias@aerifal.cx>:
> On Thu, Nov 21, 2013 at 11:31:35AM +0100, Daniel Cegiełka wrote:
>> Hi,
>>
>> http://www.openwall.com/lists/musl/2013/06/29/17
>>
>> Is there a plan to add support for stdatomic.h and threads.h?
>
> Yes, but not before 1.0. stdatomic.h is probably not hard if you
> assume GCC __sync_* builtins (or it could be written in an
> arch-specific way)

this may be helpful: stdatomic.h from FreeBSD.

http://svnweb.freebsd.org/base/head/include/stdatomic.h?revision=234958&view=co&pathrev=234958

Support for gcc and cflang. This has been added to newlib:

https://sourceware.org/ml/newlib/2013/msg00436.html

Daniel