Paul A. Clayton (paaronclayton.delete@this.gmail.com) on August 14, 2012 8:33 am wrote:
> "Essentially, the load-acquire ensures gets the most recent data at an address,
> while store-release guarantees that nobody else will read an old value of an
> address." (on page 5)
>
> might be better as:
>
> "Essentially, the load-acquire
> ensures retrieval of the most recent data at an address, while store-release
> guarantees that nobody else will read an old value at an address."

Actually this seems misleading.

Acquire and release are types of fences for *all* memory operations (not just at a particular address).

Also have to be careful because what is "old" or "most recent" is somewhat relative. A load operation will always return the most recent (from the POV of this CPU) data. But it may not be the most recent system-wide if another CPU has a store which not yet invalidated our cache line.

And (perhaps depending on ISA semantics) store release will not necessarily be visible to other CPUs at any particular time. Just that when it is visible, previous stores will also be visible.

In my opinion they really are best thought of as relative ordering operations and not absolute temporal ones. You should reason about them not in terms of ideas like "newest" and "oldest", but in terms of their relationship to other memory operations (A before B, C after D, etc).