RFR: 8353237: [AArch64] Incorrect result of VectorizedHashCode intrinsic on Cortex-A53

[Aleksei Voitylov]

On Wed, 9 Apr 2025 15:46:41 GMT, Andrew Haley <aph at openjdk.org> wrote:

>> Maybe, but I'm not sure this is the right thing to enforce in the build system across the board either. For example, a cloud vendor with a big Arm fleet without A53s may find such a setting undesirable.
>
>> Maybe, but I'm not sure this is the right thing to enforce in the build system across the board either. For example, a cloud vendor with a big Arm fleet without A53s may find such a setting undesirable.
> 
> I don't think there is any possibility that a big cloud vendor would notice. From what I see of GCC-generated code, added NOPs are very rare, because GCC tends to schedule loads long before uses. It's likely that there would be no difference to any GCC-generated code.

@theRealAph I was under the impression that building with, say, `-mcpu=cortex-a75 -mfix-cortex-a53-835769` would fail, but surprisingly (to me), it does not.

After randomly checking Alma Linux 9, Red Hat 8, Debian 12, Ubuntu 24, Fedora 41 and our Alpaquita, and finding out that for all these distros it’s on by default: 


$ gcc -Q --help=target | grep -i fix-cortex-a53-835769 
-mfix-cortex-a53-835769 [enabled]


I find no compelling reason to enforce this erratum in OpenJDK native build system. Those who disable it probably know what they are doing.

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

PR Comment: https://git.openjdk.org/jdk/pull/24489#issuecomment-2805347492