scoder
> For projects using type annotations for both type-checking and documentation, semantic naming would normally be accompanied by [type aliases](https://typing.python.org/en/latest/guides/modernizing.html#type-aliases) and/or [`NewType`s](https://typing.python.org/en/latest/spec/aliases.html#newtype). … with the obvious drawback of adding different...
> it will get worse because annotations will end up misleading Specifically in the case of Cython, we use the visitor pattern to traverse the syntax tree and make replacements...
For functions that already have docstrings, this is resolved in https://github.com/cython/cython/pull/6764
> We should look into these step by step to see if we can collect some low hanging fruit. I don't think anything more than that is on the plate...
> That's possibly worthwhile to do generally for limited API loops. We might be bending the intentions of the Limited C-API with this a little, but yeah, as long as...
After trying a couple of more benchmarks, it seems that there are some more regressions. Most benchmarks I tried are slightly to visibly slower with the master branch than with...
I added a benchmark runner to the CI jobs. Here's the latest run: ``` 2025-03-16 18:01:08,682 ### Benchmark 'bm_async_generators' (min/median/max): 2025-03-16 18:01:08,682 origin/3.0.x = 2.286 msec, 2.301 msec, 2.607 msec...
The raytrace benchmark is dominated by attribute lookups: ``` 7,93% python3 python3.12 [.] _PyObject_GenericGetAttrWithDict 4,26% python3 python3.12 [.] _PyType_Lookup ``` I'm running these with CPython 3.12/3.13 which received a lot...
According to `callgrind`, 10% of the runtime of the `bm_async_generators` benchmark are spent in 10% of the calls to `__Pyx_async_gen_anext()` that set up the asyncio hooks with `__Pyx_async_gen_init_hooks()` (which executes...
> `__Pyx__ReturnWithStopIteration()` tries to look up a vectorcall function for the `StopIteration` instantiation and fails https://github.com/cython/cython/pull/6738