@dbwiddis could you please have a look at this? It was triggered by your observations that there were problems on AIX. This might or might not improve the situation for you, but at least after this, on the gcc build farm both ant test and ant test-platform run with success.

What I'm not sure about is this:

It basicly maps everything Semeru might map to xyz.so to xyz.a. While this brings Semeru in line with the traditional AIX JDK it might not be what people expect. The alternative would be to only apply that mapping to libjnidispatch, which we control. What do you think?

I went from being able to reliably reproduce it to not reproducing it at all, so never really got to the root cause of the issues I was having. If you believe this would create a "mapped name" from the jna...tmp file ending in .so and change it to .a, then this change seems sane.

As described above it will not do that. However I doubt, that that is the problem. I read the issue you referenced and my reading is, that Semeru has a problem because it decided to drop using ".a" as the default shared library extension and switched to ".so". This in turn let different tests fails and they tried to fix this by introducing new API.