[WIP] Attempt to revamp the closure interpreter

[andreaTP]

Now that we have a mean to measure with JMH I rebased @enebo 's effort to latest main:
https://github.com/enebo/chicory/tree/closure_intpreter

I haven't noticed a massive performance improvement, thought, will check on the CI results again.

[electrum]

My initial thought is that this is more difficult to read, debug, etc. Given we have AOT, keeping the interpreter simpler seems better, even if it's slightly slower. It depends on the goals for the project.

[andreaTP]

I don't want to push for this, I agree that this is slightly more difficult to read/debug etc., the CI also doesn't show any game-changing performance improvement ( ref ).

It was worth giving it a try to make sure we have not been missing an opportunity.

[electrum]

If I'm reading the benchmark correctly, it's actually worse by 23% and 36%?

Doing these experiments is helpful. I think it shows that the JVM is good at optimizing and that simpler code is often easier for the JVM to understand.

From a code perspective, you could structure this similar to the existing switch statement by keeping the static methods:

instructions[OpCode.I32_GE_U.ordinal()] =
        (frame, stack, instance, callStack, instruction, operands) -> I32_GE_U(stack);

With Java 21, this could be simplified to

instructions[OpCode.I32_GE_U.ordinal()] =
        (_, stack, _, _, _, _) -> I32_GE_U(stack);

But I like the switch better (and seems easier for the JVM to understand).