[foreign-memaccess+abi] RFR: Add benchmarks to MemorySegmentVsBits

[Uwe Schindler]

On Tue, 3 Jan 2023 21:33:21 GMT, Maurizio Cimadamore <mcimadamore at openjdk.org> wrote:

> All checks are hoisted out of the hot loop - so in principle the segment code should vectorize as well?

Hah, this may be the reason for some code being slower in Lucene under some circumstances. We have many parts which easily vectorize but some code is still faster when done on standard on-heap arrays, but are consistenly slower with segments and ByteBuffer. This is one reason why we sometimes copy stuff from MMapDirectory to heap, although it would be better to do the calculations off-heap, too (especially vector calculations).

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

PR: https://git.openjdk.org/panama-foreign/pull/762