Last time I mentioned that with the integration of OpenJDK 8u45 MethodHandle performance went from awful to unusable. That was pretty literal as the JSR 292 test cases that I regularly run went from taking about 8 minutes to more than 30 minutes (when my patience ran out).
Using sophisticated profiling techniques (pressing Ctrl-Break a few times) I determined that a big part of the problem was MethodHandle.asType(). So I wrote a microbenchmark:
The numbers are times in milliseconds. Clearly not a good trend. I did not investigate deeply what changed in OpenJDK, but after looking at the 8u45 code it was clear that too many intermediate MethodHandles were being created. So I rewrote asType to create a single LambdaForm to do all the work at once. This improved the performance a bit, but the disturbing increase in time for the second iteration was still there. Once again I decided not to investigate the root cause of this, but simply to assume that it was because of anonymous type creation (the CLR has no anonymous types and creating a type is relatively expensive).
Avoiding anonymous type creation turned out to be easy (well, the high level design was easy, the actual implementation took a lot more time). I just had to replace the LambdaForm compiler. There is a single method that represents the exact point where I can come in and change the implementation:
static MemberName generateCustomizedCode(LambdaForm form, MethodType invokerType) { ... }
In OpenJDK this method compiles the LambdaForm into a static method in an anonymous class and returns a MemberName that points to the static method. All I had to do was replace this method with my own implementation that directly generates a .NET DynamicMethod. As I said before, the idea was simple, actually getting the implementation correct took a couple of weeks (part time).
With both these optimizations in place, MethodHandle performance is back to awful (actually, it is less afwul than it was before):
The running time of the JSR 292 test cases went down to less than 7 minutes. So I was satisfied. There are many more opportunities to improve the MethodHandle performance on IKVM, but so far no IKVM user has complained about it, so it is not a priority. Note that Java 8 lambdas are not implemented using MethodHandles on IKVM.
Changes:
Binaries available here: ikvmbin-8.1.5653.zip
Powered by: newtelligence dasBlog 2.3.12105.0
© Copyright 2021, Jeroen Frijters
E-mail