[jdk17u-dev] RFR: 8268231: Aarch64: Use Ldp in intrinsics for String.compareTo

[Andrew Dinn]

On Thu, 13 Jan 2022 14:48:40 GMT, Dmitry Chuyko <dchuyko at openjdk.org> wrote:

> Original patch applies cleanly.
> 
> Benchmark results of test/micro/org/openjdk/bench/java/lang/StringCompareToDifferentLength.java correspond to original issue, 10-25% 17u improvements on Graviton 2 for long enough strings (72+ chars).
> 
> Testing: tier1, tier2 (aarch64).

I am not a maintainer so it is not in the end my call as to whether this goes in. However, I am against backporting this change.

I recall the original discussion where the conclusion was that this patch only managed to scrape is way into jdk18 because

1. it caused no visible harm
2. it improved a rare use case on some micro-architectures

That's maybe ok as a reason to commit to a dev branch but to me it's a poor reason to backport.

Standing against whatever limited gain this might offer are not just the risk of destabilising existing deployments that is present with any unnecessary change but also an increase in code size for the intrinsic. That latter outcome may not show in the micro-benchmark results but it may well increase icache pressure, which in itself might undermine the validity of point 1 above.

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

PR: https://git.openjdk.java.net/jdk17u-dev/pull/91