[jdk11u-dev] RFR: 8280007: Enable Neoverse N1 optimizations for Arm Neoverse V1 & N2

[Liang Mao]

On Wed, 3 Jan 2024 14:10:14 GMT, Andrew Haley <aph at openjdk.org> wrote:

> > > > > Have a look at "Fix Approvals" https://wiki.openjdk.org/display/JDKUpdates/JDK11u and see what you think. I'm not sure if this one qualifies.
> > > > 
> > > > 
> > > > I believe this backport is worth considering. JDK11 is still the most used JDK version and Arm Neoverse servers grow rapidly. And this patch is quite safe.
> > > 
> > > 
> > > Understood, but that only addresses the "is of low risk" part of the requirement, but not the "substantial benefit to OpenJDK 11u users" part. What is the substantial benefit?
> > 
> > 
> > UseSIMDForMemoryOps takes great advantage of ARM Neoverse on array copy ops which is substantial to applications on JDK11 running on ARM Neoverse servers.
> 
> OK, so let's have a look at the benchmark results. I guess there was a benchmark submitted with the original patch which introduced vectorized memcpy, so it should just be a matter of re-running that on Neoverse V1 & N2. We could try it on V2 while we're at it, so we don't have to go through all this again soon.

We have a bunch of array tests. UseSIMDForMemoryOps provides 10-60% improvements on several typical non-short array copy micro benchmarks on JDK11 with Neoverse N2.

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

PR Comment: https://git.openjdk.org/jdk11u-dev/pull/2425#issuecomment-1876475140