Hi,

I just saw commit cbc5d70e... , enabling MFENCE and SFENCE on SSE2 systems,
rather than the earlier LOCK ADD [ESP], $0 for cpu_lfence and cpu_mfence.

When the Sun JVM folks were working, they found LOCK ADD was faster on Intel
systems, by a substantial amount; on AMD systems LOCK ADD was the same speed
as *FENCE, but "pipelined better", whatever that means:

http://blogs.sun.com/dave/resource/NHM-Pipeline-Blog-V2.txt

Perhaps this commit should be measured closely? I'd love to hear data either
confirming or disagreeing with Dave Dice and the Sun JVM team...

Thanks,
-- vs

Reply via email to