On Thu, Jun 17, 2010 at 6:32 PM, Alastair Bridgewater
> Hello all,
> [tl;dr: We need memory barrier operations for proper multi-thread
> operation, preferably following the linux-kernel model. Now is the
> time to raise objections or give advice.]
> A better solution is to define an explicit model for multi-thread
> memory synchronization, built around the nastiest architecture we
> purport to support, even if we never go to the trouble of updating it
> to support multi-threaded operation: the alpha. This would be about as
> future-proof as we can get, and an explicit model allows for the
> possibility of introducing automated correctness verification of
> lock-free synchronization. Or even just a well-formed basis for manual
> As it happens, the alpha is also the nastiest architecture supported
> by the linux kernel, essentially defining the linux kernel memory
> barrier model, making it the obvious model to adopt as it applies
> everywhere sbcl runs, is reasonably well explained, and is in use in
> software that sees far more users than sbcl does.
I checked out the linux memory barriers a number of times and I think
they are sane. Naming scheme is okay too.
"Reasonably well explained" does not do justice to the documentation.
It's very good.