Hello,
On Tue, Mar 5, 2013 at 10:24 PM, Mark H Weaver <m...@netris.org> wrote:
> Hi Noah,
>
> Noah Lavine <noah.b.lav...@gmail.com> writes:
> > I've only read the most recent article you posted, but if I understand
> > correctly, there is a third option: (3) somehow find a way to generate
> > a portable memory barrier instruction. Is that currently possible?
>
> If we were to do that, we'd have to add memory barriers in two places:
> (1) after writing to the lazily-initialized variable, and (2) before
> reading from it. While memory barriers are somewhat more efficient than
> mutexes, they are still very expensive.
>
I'm not sure I understand the issue, but I think I was imagining something
like
if (variable == SCM_BOOL_F) {
acquire_mutex(var_mutex);
if (variable == SCM_BOOL_F) {
variable = initialize_variable();
memory_barrier();
}
release_mutex(var_mutex);
}
That's really just a normal locking scheme with an added memory barrier to
make sure that all threads see an update after the first thread updates the
variable. Would that work?
> As for portability, C11 is the first C standard to support memory
> barriers. For now, our best bet would probably be to use libatomic_ops,
> which is also used by libgc.
>
That means we already depend on libatomic_ops, which is good. However, I
see that the website for that library recommends using C11 instead. But I
really doubt that this issue is a big enough justification to use either
libatomic_ops or C11.
> > Probably option (2) is best if we can do it.
>
> Agreed. Unfortunately, in these cases option (2) would significantly
> increase the number of modules that need to be loaded at initialization
> time. That's why I reluctantly chose option (1).
>
That makes sense.
Thanks a lot for helping me understand these memory issues.
Noah
