There's not a ton of information here that would be helpful to optimization (yet), but it's really interesting to see!

This is pretty cool! Is it mandatory to add the flame-it attribute to each function you wish to track? Can we just follow all functions?

Off topic: why does the importlib loading take so long? Is it a whole lot of code?

re: importlib: The main problem is that it is a fair bit of code, and RustPython isn't all that optimized. See #1047.

re: automatically tagging all fns: see llogiq/flamer#45. I tried to do that initially, and there were a lot of issues with it. Even so, manually tagging all fns in frame.rs with #[flame] gave way too much irrelevant data, and the file size was huge as well (36MB compared to the 1MB with the current version here). e.g., do we really need to see how long each individual exception creation function takes?

I had to use a modified local clone of flame to get this flamegraph due to a bug (llogiq/flame#47), but something interesting that you can see is that in _install_external_importers, when it imports _frozen_importlib_external, a fairly large portion of the time spent in __import__ is before it even starts executing the module (Frame::run(<module>)). Maybe we can detect when it's importing _frozen_importlib_external from _frozen_importlib and use Rust to import it?

Here it is with a simple optimization to skip __import__ in that case: link. Shaved off 10ms, not bad!

CPython actually have large parts of the __import__ function implemented in C. We can do the same in Rust for optimization.