[cosmos0072]

Compiling an expression to a tree of closures, and a list of statements to a slice of closures, is exactly how I optimized [gomacro](https://github.com/cosmos72/gomacro) my Go interpreter written in go.

There are more tricks available there, as for example unrolling the loop that calls the list of closures, and having a `nop` closure that is executed when there's nothing to run but execution is not yet at the end of the the unrolled loop.

[ALLTaken]

Impressive! I would love to learn howto implement that in Julia. Could you help me understand how you did that?

I'd love to see, if it's possible to create a libc-free, dependency-free executable without Nim (https://nim-lang.org/).

[stevekemp]

I put together this example after reading the article:

https://github.com/skx/simple-vm

Simpler than the full gomacro codebase, but perhaps helpful.

[cosmos0072]

For optimal speed, you should move as much code as possible outside the closures.

In particular, you should do the `switch op` at https://github.com/skx/simple-vm/blob/b3917aef0bd6c4178eed0c... outside the closure, and create a different, specialised closure for each case. Otherwise the "fast interpreter" may be almost as slow as a vanilla AST walker.

[stevekemp]

That's fair, thanks for taking the time to look and giving the feedback.

[cosmos0072]

The core idea is simple: do a type analysis on each expression you want to "compile" to a closure, and instantiate the correct closure for each type combination.

Here is a pseudocode example, adapted from gomacro sources:

https://gist.github.com/cosmos72/f971c172e71d08030f92a1fc5fa...

This works best for "compiling" statically typed languages, and while much faster than an AST interpreter, the "tree of closures" above is still ~10 times slower that natively compiled code. And it's usually also slower than JIT-compiled code