On 25 March 2011 10:25, Toon Verwaest <toon.verwaest@gmail.com> wrote:
which paper? Is there something more than the blog? I read the old VW5 paper but eliot told me that this is old and not accurate with Cog anymore.
Mmh... he was referring me to that "old" paper indeed. http://www.esug.org/data/Articles/misc/oopsla99-contexts.pdf
@igor that was exactly my point. You avoid copying temps around which you might not need, and accesses to shared temps from your own frame keep on having direct access rather than 1 indirection. But then again, this might not pay off vs the speedup of the remote array in the long run depending on the implementation details of the stack/jit VM, I clearly have no idea. It would be cool to see some benchmarks that compare exactly having remote arrays vs linked context frames but both on a stack.
If i'm not mistaken, remote temp vectors are rare.. (Eliot could give you a nice doit to count all closures in system which using temps vector) So if its even slow, its not a big deal anyways.
Unless I'm still completely missing the point and there's a grand JIT-reason or other to not take the whole captured stackframe off... If this hasn't been tried out yet I'd be interested in collaborating on that. I really wonder what the speed diff is.
cheers, Toon
-- Best regards, Igor Stasenko AKA sig.