Speculative multithreading (SpMT) aims to achieve speedup of a program
written for a single processor by allowing it to use multiple
processors. The motivation for the technique is that writing a program
that runs on multiple processors is difficult and needs as much
automation as possible. SpMT works by automatically dividing up a
program while it is running and executing pieces of the program code
in the wrong order on free processors. If the speculative computation
ends up being correct then execution can jump ahead, otherwise it is
simply discarded.

There have been many hardware simulation studies showing that SpMT is
useful but few software imple- mentations. Our goal is to make SpMT
work in software for the Java programming language on existing
multiprocessor machines. We first implemented a prototype in a
research Java virtual machine developed at McGill. Afterwards, IBM
became interested in the pro ject and so we made the core SpMT
functionality available in a software library. We are now working to
adapt their Java virtual machine to use our library.

We describe details of our implementation, initial performance
results, and three main directions for future optimization. The first
optimization is return value prediction. It attempts to predict the
values returned by a function, enabling speculative execution that
depends upon prior computation of these values. The second
optimization is nested speculation. This allows a speculative
computation to create an even more speculative computation. This is
fairly straightforward in hardware but difficult in software. The
third optimization is to apply heuristics that determine when a
speculative computation should be created, as opposed to blindly
attempting speculation everywhere.