RFR: 8342826: Improve performance of oopDesc::klass() after JDK-8305895

[Roman Kennke]

On Fri, 15 Nov 2024 08:40:19 GMT, Stefan Karlsson <stefank at openjdk.org> wrote:

> > > Could you make an experiment where you just force-inline `oopDesc::klass()` with `ALWAYSINLINE` and compare that to the proposed patch? When I investigated one regression a non-inlined `klass()` was prominent in the JEP-450 code but not in the baseline.
> > 
> > 
> > Force-inlining doesn't help at all. dacapo:pmd scores at 737.51ms, which is about what I currently see in mainline.
> 
> OK. Funny enough. When I force inlined klass() and the backtrace push function I got back the pmd scores. Still, for future reference can you add the benchmark scores for all the listed benchmarks for 'mainline with ALWAYSINLINE klass()'?

It *does* seem to help with the Crypto benchmark, though. Also, I am not sure why, but when I tested JEP450, I could see a very reproducible regression with JEP450, but current benchmark shows quite an improvement, instead. Something else must have changed since then.

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

PR Comment: https://git.openjdk.org/jdk/pull/22020#issuecomment-2478666307