Chris Hegarty

The overhead of `MS::isLoaded` is certainly not a good tradeoff here, as can be seen from the profile that you posted @original-brownbear. Given the new code, `consecutivePrefetchHitCount` doesn't seem effective...

/cc @uschindler

Might be that we allow the default sizes/values to be set, so e.g. 128 bit on Mac, etc. without disabling the Panama provider.

Those settings are really only useful when running the test which compares that the Panama and default implementations emit the same results. If we could somehow just limit setting them...

I really wanna spawn a new JVM with those options for a small set of tests, rather than have them infect all other running tests. For example, in the JDK...

I wonder if we can just avoid some of this complexity. For example, locally I can run all the tests with the Panama Vector implementation, by doing this: ``` $...

> By the way the empty string is allowed because of this: Yes, exactly. I'm exploiting this, which is a fine way to say "default" :-) > My quickest idea...

@uschindler your idea above is clearly preferable, but another alternative is to just add to the set of values in `vectorsize`, that will result in enabling Panama Vector randomly. Not...

I opened the above PR #13351, which implements your suggestion @uschindler. We can consider that independently, of the possibility of adding specific test tasks for different bit sizes of TestVectorUtilSupport...

Assuming , JDK 21+ and Lucene 9.12.x: I have not looked in detail, but `SegmentInfos.readLatestCommit` uses `IOContext.READONCE`, which means that we should use a _confined_ `Arena`, and that arena should...