[master] RFR: JDK-8325104: Lilliput: Shrink Classpointers [v3]

[John R Rose]

On Mon, 25 Mar 2024 19:38:11 GMT, John R Rose <jrose at openjdk.org> wrote:

> … interactions with data cache dynamics …

I see acknowledgement of this issue in a comment in this PR that says "smallest shift possible that still covers the whole range”.  This is a good solution as long as that smallest shift is 64 (or less).  A shift of 128 would dilute the D$ by 50%, but would reach 2^(22+7-10) distinct Klasses, half a million, which is great.  I guess VMs which configure themselves for a quarter million classes or fewer just use the cache line as the allocation modulus, and all is well.  So my solution (mod 1088) is really  useful only for huge workloads.  I retract my suggestion that performance experiments are needed at this time, given the solution you have in place.

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

PR Comment: https://git.openjdk.org/lilliput/pull/128#issuecomment-2018815640