RFR: 8366441: AArch64: Support WFET in OnSpinWait [v3]

[Per Minborg]

On Tue, 10 Feb 2026 13:09:05 GMT, Andrew Haley <aph at openjdk.org> wrote:

> > Unless the default value is changed, this will have a relatively large impact on certain low-latency applications where one waits for a CAS-lock in a spin loop. Latencies in such applications are often measured in ns. So, this is something that should be communicated to the community in some way.
> 
> Thanks. Do you have an idea of what you'd like to see?
> 
> I think cache coherence protocols have about 40ns latency between cores in a socket, so would that be a sensible default?

I think something like that would be a good default. A handful of years ago, I wrote an article on the concept, and at the time,  we saw just above 250 ns end-to-end messaging between two Java processes. I think these days it is faster. See https://minborgsjavapot.blogspot.com/2022/05/which-jvm-version-is-fastest.html

-------------

PR Comment: https://git.openjdk.org/jdk/pull/27030#issuecomment-3877592665