On Tue, 15 Sep 2026 07:34:19 GMT, Fei Yang <[email protected]> wrote:

>> Yes, we can create a new JBS for this optimization later.
>
>> However, a more intuitive way of thinking is that a feature has two distinct 
>> dimensions: supported (whether CPU/OS support has been detected) and enabled 
>> (whether the JVM chooses to use it). A supported feature may remain 
>> disabled, but attempting to enable an unsupported feature should trigger a 
>> warning and leave it disabled, rather than override capability detection.
> 
> The situation is quite complex here on this CPU platform. Currently, we 
> mainly rely on the hwprobe linux syscall for hardware feature/extension 
> detection and option enablement. As I remembered, there are several 
> considerations. First, the number of extensions is non-trivial. It's not that 
> feasible for end users to manually select and enable all of them.  But they 
> are still able to disable ones they don't like. The diagnostic options are 
> still there. Second, it's not easy to say which extensions are not supported. 
> The ecosystem is still maturing. And you have different kernel versions with 
> / without hwprobe syscall. The hwprobe syscall is evolving too adding more 
> extensions as we go.  So it could happen that a feature is there, but is not 
> detected on old kernels. Still the user could enable such ones through these 
> diagnostic options on the command line.

Thanks, that makes sense. I missed the "unknown" state where hardware support 
exists but the kernel cannot report it.

It still feels slightly unsafe to treat "not reported" as "assume supported", 
but I understand why the diagnostic override can be useful here.

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

PR Review Comment: https://git.openjdk.org/jdk/pull/32309#discussion_r4021373984

Reply via email to