[Patch] C1X threading support

[Niall Douglas]

On 18 Dec 2011 at 2:06, Dag-Erling Smørgrav wrote:

> "Poul-Henning Kamp" <phk at phk.freebsd.dk> writes:
> > Ohhh, but I know:  Lets make a rival to the POSIX threads, we can do it
> > much better and slightly incompatible, big market there I'm sure.
> 
> That's not the point.  The point is that C until now did not have a
> concurrency model.  The threading API in itself is not important; I'm
> sure the committee knows perfectly well that nobody is going to use it.
> What's important is that the standard now defines how C behaves in a
> concurrent environment.

Actually we don't go far enough at all in specifying how C behaves in 
a concurrent environment. The problem is that in the expected 
lifetime of the standard (10 years) there are a number of known big 
changes coming to commodity hardware e.g. memory transactions and 
non-SMP cache coherency. No one on the committee wants to 
accidentally break future hardware features, so we have left them 
unspecified.

BTW we entirely expect the C1X threading API to supplant all others, 
including POSIX whose threading support will be mostly for backwards 
compatibility and may even become deprecated. A lot of resources and 
effort has been directed into getting the memory model right and 
future safe if not future proof. Indeed, the chances are that the 
next POSIX revision will copy C1X in some areas rather than the other 
way round.

The complaints about API bloating are obvious. Bear in mind that 
while POSIX based platforms are in a majority, there are significant 
large implementors of C1X for whom a lot of new support code will 
have to be written e.g. MS Windows, whereas POSIX platforms have a 
much easier time of things in supporting C1X.

Regarding implementing support for the C1X threading API, I'd suggest 
a different approach. Windows PE and Apple Mach-O both support symbol 
aliasing so you can declare an API to be identical to another API. 
Interestingly, ELF has no such feature. See 
http://blog.omega-prime.co.uk/?p=121.

I would suggest the best and easiest way to implement much of the C1X 
library is to add symbol aliasing to ELF. Then all ELF binary based 
systems could support C1X features without having to resort to inline 
headers or thin wrappers. Linux will be needing this too, and it's 
the easiest way out.

Niall

-- 
Technology & Consulting Services - ned Productions Limited.
http://www.nedproductions.biz/. VAT reg: IE 9708311Q. Company no: 
472909.