> The more interesting question is whether requiring that additional barrier buys more performance in the typical case where you're not touching shared data. P

> The more interesting question is whether requiring that additional barrier buys more performance in the typical case where you're not touching shared data. Probably yes, i.e., there was some hardware reason for it. You can find one explanation [1] floating about that explains the effect was due to the cache banking design, with separate invalidation queues for each bank. Presumably that was better in some respect than the alternative designs available to implementors. That said, the difference is probably fairly small and if Alpha were around today they would almost certain want to abandon that particular reordering since they are the only one doing it and there would be a lot of pressure on them to make it fast (this comes up in all sorts of non-trivial scenarios like double checked locking, reads of final fields in Java, etc). That's generally the pattern, I think: if you are the odd man out with the weakest model, and you aren't the dominant player, you will have a lot of pressure to at least offer fast ways of lining up with the stronger models. So even if a weak model is a better design point in a vacuum, it might not be true when real-world implementation pressures are considered. --- [1] http://lse.sourceforge.net/locking/wmbdd.html http://lse.sourceforge.net/locking/wmbdd.html