RFR: 8370947: Mitigate Neoverse-N1 erratum 1542419 negative impact on GCs and JIT performance [v25]

[Evgeny Astigeevich]

On Tue, 10 Feb 2026 13:10:44 GMT, Andrew Haley <aph at openjdk.org> wrote:

>>> > > @theRealAph @shipilev @fisk Any other tests to run in addition to the suggested tests above?
>>> > 
>>> > 
>>> > It all goes back to the question I asked before. If we run this test on all Neoverse N1, not just the affected stepping, would there be any disadvantage?
>>> 
>>> Sorry, i mean if we run the same workaround code on all Neoverse N1. Why not? Would it be slower?
>> 
>> I don't have unaffected N1 to check. Out of curiosity, does unaffected N1 exist in any cloud?
>> 
>> I think it would not be slower.
>> 
>> Unaffected N1 will be running the following code, as V1 and V2, once per patched nmethod:
>> 
>> dsb ish
>> isb
>> 
>> 
>> [Graviton 3(Neoverse V1) results](https://github.com/openjdk/jdk/pull/28328#issuecomment-3585923078) show benefits of the deferred invalidation as well. So should unaffected N1. An explanation of the improvement is in the comment: https://github.com/openjdk/jdk/pull/28328#issuecomment-3585923078.
>
>> I don't have unaffected N1 to check. Out of curiosity, does unaffected N1 exist in any cloud?
>> 
>> I think it would not be slower.
> 
> Right, so let's turn this on for all N1. Then we can be sure that it gets tested.

@theRealAph @shipilev @fisk 
I ran jcstress with ZGC used on Graviton 2. 100% tests passed.

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

PR Comment: https://git.openjdk.org/jdk/pull/28328#issuecomment-3891149711