Incorrect Swing GUI behaviour on Mac AArch64

[0xdaryl]

There is an issue running a Swing-based application on native macOS AArch64. Menus and dialog boxes repaint as white and do not refresh, and hangs are sometimes observed.

I have not seen this behave incorrectly when run with -Xint.

To reproduce:

1) Reproduce with a 17.0.6 Semeru build:

https://github.com/ibmruntimes/semeru17-binaries/releases/download/jdk-17.0.6%2B10_openj9-0.36.0/ibm-semeru-open-jdk_aarch64_mac_17.0.6_10_openj9-0.36.0.tar.gz

2) Download IBM Thread and Monitor Dump Analyzer

See link at the "Download the jar" step here -> https://www.ibm.com/support/pages/ibm-thread-and-monitor-dump-analyzer-java-tmda

A problem like this was reported on two Swing-based applications, but the TMDA application is one that is easy to get and reproduce with.

3) java -jar jca4614.jar

4) When the application starts, play around with the GUI a bit: open menus, click boxes, drag windows, etc. You could even load a javacore into the analyzer as well. In about 10-15 seconds the bad behaviour should be seen.

[0xdaryl]

Reproducible with -Xshareclasses:none -Xjit:optlevel=cold. Not seen with noopt.

[0xdaryl]

One of the problematic methods appears to be:

sun/java2d/opengl/OGLTextureToSurfaceBlit.Blit(Lsun/java2d/SurfaceData;Lsun/java2d/SurfaceData;Ljava/awt/Composite;Lsun/java2d/pipe/Region;IIIIII)V

I suspect this problem appears in other methods as well because excluding just this one method does not avoid the problem. However, running with only this method compiled will reproduce the problem.

[0xdaryl]

This appears to be a codegen bug passing one of the in-memory arguments on a method dispatch sequence to:

sun/java2d/opengl/OGLBlitLoops.IsoBlit(Lsun/java2d/SurfaceData;Lsun/java2d/SurfaceData;Ljava/awt/image/BufferedImage;Ljava/awt/image/BufferedImageOp;Ljava/awt/Composite;Lsun/java2d/pipe/Region;Ljava/awt/geom/AffineTransform;IIIIIDDDDZ)V

The argument is being clobbered during the dispatch sequence before being stored on the frame. Not sure why yet...

[knn-k]

I haven't looked into the generated code, but PR #16458 might have fixed this issue. The fix is not available in v0.36.

[0xdaryl]

Yes, the problem seemed familiar to me and I've been trying to get a build to run locally to verify that particular fix.

The problem is indeed because we're not copying one of the args into a temp.

[0xdaryl]

Not verified, but reviewing that PR again that is exactly the fix I would have implemented to fix this. But will confirm...

[knn-k]

I ran the application with the following binaries, and I observed the weird behavior with neither of them.

• v0.38 build from https://openj9-jenkins.osuosl.org/view/Build_Release/job/Build_JDK17_aarch64_mac_Release/47/
• Personal build using the head streams

[0xdaryl]

This has been fixed in 0.38+ builds. The user confirms they no longer see the problem running a 0.38 build on their application either. Closing.