Re: Some introductory docs about C level threading

[Ken Raeburn]

On Feb 6, 2005, at 19:48, Kevin Ryde wrote:
> Ken Raeburn <address@hidden> writes:
> > The problem is, if there are any case where it would be needed, then 
> > we
> > need to use it everywhere there could be a potential problem.
> > Depending on just how it works out, that could mean essentially
> > declaring some of the scheme object types (like cons cells) as
> > volatile, which would be poor for performance.
>
> I'd think it shouldn't hurt too much.  A store written in the code
> probably means a store should be done :-).  Optimizations for local
> variables etc should be unchanged, just memory ops become exactly as
> written.

And a read becomes a read, with no option for combining multiple reads 
from the location.  I think that's where the performance might not be 
so good.

> > And volatile declarations may fix the possibility of storing 
> > incomplete
> > values, but I don't think it would deal with the cross-thread memory
> > access ordering issue.
>
> I think we have to assume all threads see memory ops in the order
> they're done (anything else is too crazy), and with that assumption
> and some care libguile ought to be ok.

Then you will never have a correct port to the Alpha, at least.  The 
Alpha architecture specification allows for rearrangement of accesses 
except when a memory barrier instruction is used.  (Memory operations 
issued before the MB must precede those issued after it.)  I'd be very 
surprised if other modern architectures didn't also use various weak 
memory models, for performance in multiprocessor configurations.