Dave Dagastine has done some performance measurements that indicate
that doing prefetches on SSE-enabled x86 and amd platforms improves
specjbb by 1-1.5%, vis Dave's email:

Considering we know that c2 code generation has opportunities for improvement, I did a quick review its current tuning flags. I found that AllocatePrefetchStyle is set to 2 on ia64 and sparc, and set to 0 (disabled) on amd64 and i486. I did a few tests on linux_i486 and the following gives us another 1-1.5% on SPECjbb:
-XX:AllocatePrefetchStyle=2 -XX:AllocatePrefetchDistance=32

it also helps Solx86.