Getting inspired by faster-cpython project!

[AmirHmZz]

I've been following faster-cpython repository since python 3.10 was released. They are trying to make CPython faster and have already achieved some exciting results. CPython3.11 is about 25% faster than CPython3.10. Can MicroPython benefits from these ideas to increase its performance?

[jimmo]

Almost certainly!

MicroPython does have some additional constraints that make some of the ideas more difficult. For example the Specialising Adaptive Interpreter (https://peps.python.org/pep-0659/) is difficult because in many situations MicroPython bytecode is immutable (because it's executing from ROM).

Some history: MicroPython did have a bytecode-mutating optimisation for optimising map lookups, that made instructions like LOAD_ADDR and LOAD_GLOBAL much faster. But this was basically never able to be enabled except on Unix. I believe CPython subsequently implemented something similar (this predates faster-cpython though). We subsequently removed and implemented with a separate cache for almost the same overall benefit, but supported and enabled on all targets. Mutating the bytecode is faster, but just wasn't an option.

Many faster-cpython ideas are also around improving memory usage (which is faster on its own, but also reduces GC pressure). MicroPython obviously already takes this to pretty extreme levels (to the opposite end of the spectrum where it deliberately does things in slower ways to use even less RAM).

There's definitely some stuff worth exploring around exceptions and stack frames, I know faster-cpython is investigating this too. Damien did a bunch of work on this previously see the MICROPY_STACKLESS and MICROPY_ENABLE_PYSTACK options in the code.

[dpgeorge]

I also keep an eye on the faster-cpython progress. There are some nice ideas there. It looks like they will switch to a register-based VM (as opposed to stack-based) and that's probably never going to happen in MicroPython because it'll make the bytecode a lot bigger. But there are other things we can look into.