Prefetch is way that the compiler can tell the processor "In a little while, I'm going to touch this particular memory address, so why don't you start loading it into the cache for me if it's not already there." This is a way of reducing memory latency and can give a 10-20% boost in performance for some applications.
The compiler looks at memory reference patterns, for example, stepping through an array, and automatically inserts prefetch instructions ahead of when the data will be used, improving the chance that the data will be in the cache when needed.
This is a big help on Pentium III, but not so much on Pentium 4 where the processor itself tries to predict memory use patterns and does its own prefetching. We found, for example, that applying a Pentium III prefetch model to Pentium 4 actually made performance worse! CVF 6.6 uses a more appropriate memory system model for Pentium 4, resulting in fewer prefetch instructions issued, and better performance.
We were able to add this to CVF 6.5 because our optimizer already knew how to do prefetching for Alpha, so it was just a matter of tuning the memory model.
Steve
