RFR: JDK-8318119: Invalid narrow Klass base on aarch64 post 8312018

[Thomas Stuefe]

On Fri, 3 Nov 2023 14:39:35 GMT, Andrew Haley <aph at openjdk.org> wrote:

> > Why not do the optimized (for AArch64) stuff before trying the immediate + shift?
> 
> For AArch64, the optimum is zero base, no shift, then immediate base, no shift, then shift, and last of all immediate + shift.

Wait, immediate + shift=0 before no immediate=0 + shift? Why? Looking at https://dougallj.github.io/applecpu/firestorm-int.html , both movk and lsl need the same number of cycles. Why would movk be better than lsl?

The reason I ask is that this would mean switching up the order of things I do in Metaspace::reserve_address_space_for_compressed_classes. That function attempts to do the best for all platforms, and all other platforms prefer zero immediate + shift. It would mean even more ifdefs than we have before.

(the best way would be to move this function into CPU specific files. We would have some duplication, but at least every CPU can then decide how to layout memory, and the comments would not have to be so darn vague. But last time I attempted that I ran into objections from reviewers).

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

PR Review Comment: https://git.openjdk.org/jdk/pull/16215#discussion_r1383499653