scoder
BTW, I agree that class creations should give a good granularity for a split. I think that's worth trying out.
It seems that quite a bit of work went into this already and there are only a couple of problems left here. @0dminnimda, could you find the time to finish...
It might also be interesting to see the differences in symbol sizes that the C compiler generates from 0.29.x and 3.0 C sources. You can list those for a given...
> it seems each of them corresponds to the expansion of `__PYX_ERR`. That macro is marked as cold code, which is probably why you see a bulk of them stuffed...
> By the way, something much simpler might be to move each type initialization to its own function, so as to avoid having a single extremely large init function. See...
> I kind of wonder if any of this "large integer" type code is reliable covered by the test-suite at all. This should exercise it: https://github.com/cython/cython/blob/master/tests/run/int128.pyx
``` FAIL: signed_conversion (int128) Doctest: int128.signed_conversion ---------------------------------------------------------------------- File "/home/runner/work/cython/cython/TEST_TMP/0/run/cpp/int128/int128.cpython-313-x86_64-linux-gnu.so", line ?, in int128.signed_conversion Failed example: bigint(signed_conversion(-2**127)) Exception raised: Traceback (most recent call last): File "/opt/hostedtoolcache/Python/3.13.0-alpha.4/x64/lib/python3.13/doctest.py", line 1378, in __run exec(compile(example.source,...
I found several ref-leak issues in the shift+mask fallback conversion implementation while working on this that should also be fixed in 3.0.x. It's probably easier to backport the PR than...
I agree with @da-woods that having a concrete (toy, if necessary) example that exposes this issue would help. I would like to avoid covering up a bug that we can...