On Mon, 9 Jun 2025 22:09:44 GMT, Kevin Rushforth <k...@openjdk.org> wrote:
> > > I'd looked at that but it was not obvious. And it definitely is not > > > obvious that it will be the same thread that initialises the > > > MaskMarlinAlphaConsumer.java class > > > > Yes, I see what you are saying. The `ALPHA_MASK_XXX` arrays are static > > fields of `MaskMarlinAlphaConsumer`, and the renderer context class has an > > instance field of `MaskMarlinAlphaConsumer`. So it does seem possible that > > `MaskMarlinAlphaConsumer` could be initialized by a thread other than the > > prism renderer thread. Hmmm. > > Fortunately, the presence of following field in RendererConext does not cause > the MaskMarlinAlphaConsumer class to be initialized: > > ``` > public MaskMarlinAlphaConsumer consumer = null; > ``` > > I instrumented the code, adding something that _did_ force initialization and > was able to provoke a WrongThreadException. Without my modification, > MaskMarlinAlphaConsumer doesn't get initialized until rendering. > > This seems _very_ fragile, though. At best, presuming we can prove that > initialization and access to MaskMarlinAlphaConsumer always happens on the > Prism rendering thread, this is an accident waiting to happen. > > @bourgesl Do you have any thoughts on this? Using the shared arena is > definitely safer, but there might be a performance penalty. These alpha masks are constant (static), so reused in marlin java2d where multiple threads & rendering contexts can work in parallel. With FFM, the global arena could be used, but the current patch is good enough as javafx uses only 1 thread to perform rendering. I hope using the shared arena does not hurt performance in this case (1 thread). ------------- PR Comment: https://git.openjdk.org/jfx/pull/1814#issuecomment-2959886137
