[GR-58381] Synchronize accesses to class init state

[graalvmbot]

This PR adds memory ordering to accesses to HotSpot's InstanceKlass::_init_state field in the Graal JIT, adapting a change introduced in the upstream PR openjdk/jdk#21110.

[zakkak]

Hello @c-refice, could you please provide a description for this PR?

I am wondering if this patch would help with the (theoretical) issue described in quarkusio/quarkus#40737 (comment):

if we add concurrency in class initialization there is no way for GraalVM to detect whether all threads have completed or not. That said if a class initializer spawns a thread and returns, the spawned thread can potentially run in parallel with the analysis (e.g. the reachable callbacks) and image building in general which can be problematic...

[alina-yur]

@zakkak thank you, provided.

[c-refice]

I am wondering if this patch would help with the (theoretical) issue described in quarkusio/quarkus#40737 (comment):

From what I understood the issue is with a native image build, whereas this change is specific to HotSpot compilations in Graal JIT.

[zakkak]

Thank you @alina-yur and @c-refice for the clarifications