RFR: JDK-8309410: [aarch64] Use pre-computed tables for logical immediates

Thomas Stuefe stuefe at openjdk.org
Mon Jun 5 09:38:05 UTC 2023


On Mon, 5 Jun 2023 08:56:14 GMT, Andrew Haley <aph at openjdk.org> wrote:

> Whether this is a win depends on a bunch of things that are hard to predict, such as the time take to page in the table from mass storage. If we're really going to revisit this, and I'm not at all sure we should, we should be thorough.

The table is ~64K, and we only use it very sparsely, since we only look up a small subset of possible encodings. So we'd only ever page in a small subset of pages. Computing the table at runtime involves paging in the whole space for both tables to populate them, apart from the bit fiddling. I assumed that would be more expensive. Measuring this would be an effort though.

> If we're really going to revisit this, and I'm not at all sure we should, we should be thorough.

My proposal was an incremental improvement and it deliberately does not touch the table format and the decoding. If you think it is not worth it, I'll close it - this was just a small improvement and there is no time for more.

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

PR Comment: https://git.openjdk.org/jdk/pull/14304#issuecomment-1576461199


More information about the hotspot-compiler-dev mailing list