Talin wrote:
I'm thinking along similar lines, but my approach is to eliminate refcounting entirely.
That's a possibility, although refcounting does have some nice properties -- it's cache-friendly, and it's usually fairly easy to get it to work with other libraries that have their own scheme for managing memory and don't know about Python's one.
For example, take the case of a dictionary in which more than one thread is inserting values. .. I think we want to avoid the Java situation where every object has its own lock.
Having to lock dictionaries mightn't be so bad, as long as it can be done using special instructions. It's still a much larger-grained locking unit than an incref or decref. But I'm wondering whether the problem might get solved for us from the hardware end if we wait long enough. If we start seeing massively-multicore CPUs, I expect there will be a lot of pressure to come up with more efficient ways of doing fine- grained locking in order to make effective use of them. Maybe a special lump of high-speed multi-port memory used just for locks, with surrounding hardware designed for using it as such. -- Greg