[aardvark179]

I would say that AST interpreters are surprisingly fast on the JVM, and byte code interpreters are surprisingly slow. Maybe one day explicit tail calls combined with invoke dynamic will offer us a path to byte code interpreters that the JIT can optimise based on the predictable nature of byte code sequences, but we’re not there yet.

Edit: I should point out that byte code interpreters still have a place because they tend to represent code in a much smaller form. Again this is something that will change over time as things like project Lilliput shrink object headers, but it’s unlikely to go away entirely.

[larsga]

What really gave clear performance improvements was generating Java bytecode. In the first attempt I got a 25% speedup, and I'm sure I could have gotten more out of it if I'd spent more time on it.

[aardvark179]

Oh yeah, absolutely. The trade off with byte code is that it can feel like threading your language through the class file format’s very particular needle, and you have to pay a larger initial cost generating and loading that class, and you have ongoing costs in terms of the meta space classes consume.