what's the point of rpython?

[Carl Banks]

On Jan 19, 8:00 pm, "Brendan Miller" <catph... at catphive.net> wrote:
> Maybe I'm missing something here but a lock free algorithm for
> reference counting seems pretty trivial. As long as you can atomically
> increment and decrement an integer without locking you are pretty much
> done.

You're missing that most of the platforms that Python supports can't
actually do this.  Keep in mind that it has to be atomic across all
cores.

(For Python to use such a technique, it would have to be doable on
every platform Python supports.  They aren't going to get rid of the
GIL for some platforms and not others.)


> For a reference implementation of lock free reference counting on all
> common platforms check out boosts implementation of shared_ptr (a
> reference counting smart pointer designed around multithreaded use
> cases).

I just looked at the boost documentation, which claims that multiple
asynchronous writes to the same shared_ptr results in undefined
behavior.  That will not suffice for Python reference counting.



Carl Banks