Recently I added support for Intel 64bit JNI compilation. To simplify this I implemented helper methods in JNIEnvironment. I think the same approach should be adopted for PPC as there is no obvious performance degradation:
This approach should also hopefully simplify changes to JNI code just requiring changes to Java code in the future. For example, the native threading code changes the entry and exit to/from JNI, this has been achieved currently by hand rewriting the assembler for Intel. Whilst the approach costs a method call it also exposes the potential for optimization within the helper method performed by the opt compiler. It is also more maintainable and easier to verify the behavior of.
I don't believe the performance drop is significant as the biggest cost is for native methods with reference arguments. All native methods have 1 reference argument but the other arguments are small in number and frequently not reference types.