Pass 2.3

[Hans Boehm]

I guess you alreaday submitted.  And that's fine.  Thanks.

I'm still not sure there are actually important architectures with the
problem you mention below.  It is something to look at, though.

Hans

On Fri, 10 Sep 2004, Andrei Alexandrescu wrote:

> Oh, thanks. I also added one more paragraph in the "Extra writes"
> section, reintroducing the word tearing theme. Let me know if that's cool:
>
> Another instance of this problem is with hardware that doesn't support
> memory writes below a certain size (often, \verb|sizeof(int)|). In
> that case, if \verb|S| is ``packed'' such that \verb|sizeof(S) ==
> sizeof(int)|, the compiler must generate code for \verb|s.a = 0| to
> read the entire \verb|s| object in a register, mask a portion of it,
> and write back the entire object to memory. Effectively, simply
> writing the the field \verb|S::a| became a read-modify-write operation
> as far as the neighboring fields \verb|S::b| and \verb|S::c| are
> concerned---operation that has racing problems if the operation is not
> atomic, and efficiency problems if it is not.
>
>
>
> Andrei
>
> Boehm, Hans wrote:
> > Looks good.  But I think the comparison in "Fun" is backwards.  Delete the return and reindent.
> >
> > Minor point for future discussion:
> > If you implement atomics with private locks, you also need to disable signals.  (Atomic_ops can do this as a last resort.)  It may be cheaper to just do the emulation in the kernel.  I dimly recall that this was done for AIX on Power a long time ago.
> >
> > Hans
> >
> >
> >>-----Original Message-----
> >>From: Andrei Alexandrescu [mailto:andrei at metalanguage.com]
> >>Sent: Friday, September 10, 2004 4:33 PM
> >>To: Kevlin Henney
> >>Cc: Doug Lea; Boehm, Hans; 'Ben Hutchings'; Doug Lea; pugh at cs.umd.edu
> >>Subject: Re: Pass 2.3
> >>
> >>
> >>Ok, I've made a quick pass, put some pizzazz in, etc. If anyone could
> >>give it a read and comment on it, that would be great. Thanks!
> >>Andrei
> >>P.S. If you want to compile the doc, you'll need the listings package.
> >>
>