LogTM and SLE

I came up with this idea today - which appearantly James had actually thought of first, to implement a LogTM and SLE hybrid. Basically the processor would run with SLE as long as there are sufficient resoucres. When it runs out of resources, it only logs the memory values of the data items it evicts.

This scheme, as I first thought, would save a lot on aborts. Since all the abort would have to do is to rollback onthe they (hopefully) few log entries it creates (which should be significantly less than the ones with LogTM), and clear the cache. Moreover, it should reduce the overhead of logging.

The problem is though, the way LogTM does it, the values it saves in the log will most probably have been in the cache already, while the scheme I proposed here will definitely have to go to memory to fetch the old value to log. Thus it would end up being slower than LogTM.

Moreover, aborts are very infrequent, so there is no real need for optimising them.

The problem with LogTM is essentially that it cannot handle context switches. Plus James thinks that even though the idea is cool, the actual implementation is not elegant. I am not familair enough with LogTM yet to make any judgements personally...