How do you implement Software Transactional Memory?

Question

In terms of actual low level atomic instructions and memory fences (I assume they\'re used), how do you implement STM?

The part that\'s mysterious to me is that given s

Answer 1

I suggest you watch this presentation: http://www.infoq.com/presentations/Value-Identity-State-Rich-Hickey

In the second half it explains how to update values without leaving them in an undefined state. For example - if you have a tree that you want to update in STM-style you don't change the previous version at all. Let's say that tree is a pointer to the root of the tree. The only thing you create is the nodes that changed (but they can refer to nodes in the original snapshot of the tree.

Then you do a compare-and-swap on the tree pointer. If it succeeded, then everyone will now see your new tree and the old one can be garbage-collected. If it hasn't, then you repeat the process and the tree you just constructed is garbage collected.

The big idea is that you don't need to detect if anyone else changed the tree until you actually swap the new and old values, so there are no "conflicts" or "clobbered values" from the typical multithreaded programming.