Subj : Re: Memory visibility and MS Interlocked instructions
To   : comp.programming.threads
From : Peter Dimov
Date : Sat Sep 03 2005 01:59 pm

Alexander Terekhov wrote:
> David Hopwood wrote:

> > it, but most lock-free stuff doesn't need it, because it doesn't rely on
> > total store (or load) ordering.

It's been my impression that lock-free usually relies on a transitive
happens-before... but I may be wrong.

> Yep, most lock-free stuff doesn't rely on "atomic visibility" (remote
> write atomicity).

I find that very odd. Not the absence of atomic visibility per se,
but...

P1: X = 1; #storestore; Y = 1;
P2: if( Y == 1) { #loadstore; Z = 1; }
P3: if( Z == 1) { #loadload; assert( X == 1 ); }

If Z == 1 happens before X == 1, Z = 1 happens-before Z == 1, Y == 1
happens-before Z = 1, Y = 1 happens-before Y == 1 and X = 1
happens-before Y = 1, then X = 1 happens-before X == 1, right? If this
is not true, the happens-before model breaks, and it's impossible to
write programs against it. (I don't think that a non-broken
happens-before requires write atomicity - it's still weaker than that -
but I may be wrong about that as well.)

I'm obviously missing something.

.