RFR: 8341013: Optimize x86/aarch64 MD5 intrinsics by reducing data dependency [v2]
Oli Gillespie
ogillespie at openjdk.org
Fri Sep 27 15:46:35 UTC 2024
On Fri, 27 Sep 2024 15:02:54 GMT, Andrew Haley <aph at openjdk.org> wrote:
>> I see, Thanks!
>
> Unless you really want zero extension it's better to use `mov` than `movw` (or `orrw`).
> On many AArch64 implementations `mov` doesn't even issue. Instead, it is handled by the renamer during the decode stage. However, because it has to clear the upper 32 bits, `movw` does issue.
Thanks! I don't measure a throughput improvement (didn't check any perf counters like instructions retired) when changing the `movw` to `mov` for `t` across F, G, H and I on Neoverse N1 or Mac M1. I'm also not sure how to tell if it's safe, my knowledge is shallow here.
-------------
PR Review Comment: https://git.openjdk.org/jdk/pull/21203#discussion_r1778824238
More information about the hotspot-dev
mailing list