On Sat, 20 Sep 2025 01:23:25 GMT, Ioi Lam <[email protected]> wrote:

> > @iklam Reading up the description of the bug tracker, it seems we are 
> > trying to find out more opportunities to AOT initialize classes than the 
> > current state. Is that correct? I may be missing something here, why not 
> > "force initialize" classes annotated with `AOTSafeClassInitializer` in 
> > `AOTMetaspace::try_link_class`, instead of introducing new annotation? 
> > Wouldn't that serve the same purpose? And I think we also need to track the 
> > classes that are actually initialized in the training phase, and apply 
> > "force initialization" only on that subset. Otherwise we can end up in 
> > situation where a classes is loaded and not initialized in training phase, 
> > but still gets AOT initialized because it is annotated with `AOTInitialize`.
> 
> I think this is a better alternative. I will fix the implementation to do 
> what you described (by tracking the set of initialized classes in 
> finalImageRecipes..cpp).
> 
> To avoid confusion, I will post a separate PR.

I have created an alternative PR. https://github.com/openjdk/jdk/pull/27402.
@adinn @ashu-mehra @liach please take a look.

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

PR Comment: https://git.openjdk.org/jdk/pull/27024#issuecomment-3315384179

Reply via email to