RFR: 8342601: AArch64: Micro-optimize bit shift in copy_memory

[Dean Long]

On Fri, 18 Oct 2024 23:59:28 GMT, John R Rose <jrose at openjdk.org> wrote:

> In other words, the cost model justifying this supposed improvement is probably about a half century out of date. Maybe an expert on AARCH64 can correct me on this point?

Apparently some moves are "0 latency" and skip the pipeline on some aarch64 hardware.

Yes, we usually expect Assembler APIs to do exactly what we ask, but when the API is in MacroAssembler and doesn't correspond to an actual aarch64 instruction or alias, I think we normally allow optimizations.  As LSR is an alias, I think we would expect it to generate the underlying `ubfm` encoding, so if we were going to optimize based on the shift value, we could introduce a new API with a name like shift_right().

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

PR Comment: https://git.openjdk.org/jdk/pull/21589#issuecomment-2423403430