RFR: 8354572: Turn off AlwaysMergeDMB for Ampere CPU by default

[Liming Liu]

On Tue, 15 Apr 2025 12:03:45 GMT, Andrew Haley <aph at openjdk.org> wrote:

> > > Please post the numbers here.
> > 
> > 
> > Hi @theRealAph, I ran the benchmark from [JDK-8354572](https://bugs.openjdk.org/browse/JDK-8354572) with `-UseStoreStoreForCtor`, and the thoughput is ~51.1k ops/ms for `-AlwaysMergeDMB` on Ampere1, while it is ~23.8k ops/ms for `+AlwaysMergeDMB`.
> 
> OK, I think I understand now. However, have there been any other tests to ensure that we don't see regressions elsewhere?

Testcases from Renaissance mentioned in [JDK-8325449](https://bugs.openjdk.org/browse/JDK-8325449) have improvements, and there is no regression on SPECjbb2015 as well. Basically, the benchmark from [JDK-8354572](https://bugs.openjdk.org/browse/JDK-8354572) shows that "dmb ishst + ishld" is faster than "dmb ish" on Ampere CPU with some load and store instructions around. If a benchmark had regression due to this, it would benefit from a slower fence.

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

PR Comment: https://git.openjdk.org/jdk/pull/24652#issuecomment-2811710445