RFR: 8352075: Perf regression accessing fields [v23]

[Radim Vansa]

On Fri, 6 Jun 2025 16:13:11 GMT, Ioi Lam <iklam at openjdk.org> wrote:

>  While the benefit can be seen in a synthetic benchmark, do we have any data that shows a benefit in real world applications?

@iklam Regrettably I cannot disclose the reproducer that came from a customer, but it is not synthetic - it caused problems when migrating to 21.

I admit that I was not looking on the unaligned behaviour of platforms besides x86_64 and aarch64, where this was supposed to be OK (with perf penalty at worst). I've replaced this with `__builtin_memcpy` as I haven't found a different platform-agnostic way to convince GCC handle an unaligned read/write. I don't see any difference in performance on the reproducer, so it's probably good enough.

Added the debugging print method as well.

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

PR Comment: https://git.openjdk.org/jdk/pull/24847#issuecomment-2955355271