src dependencies/cleanup issue aggregator

[jbrockmendel]

• _libs/window both includes "skiplist.pyx" and cimports from skiplist. I don't know if it is necessary, but it is definitely confusing. (closed by simplify skiplist inclusion/cimport to be more cythonize-friendly #18420)

• _libs/reshape.pyx is an 18 line stub that just includes reshape_helper.pxi. Could we just change reshape_helper.pxi.in to reshape.pyx.in? updateIt is correct as is.

• src/reduce.pyx and src/inference.pyx are a minor PITA, should be merged directly into lib or made into their own modules.

• src/numpy.pxd should be dropped, just defer to np.get_include() (closed by Remove src/numpy.pxd #19418)

• Only a small fraction of src/datetime/ is used outside of src/, a lot of it is only used by src/ujson/python/objToJSON.c. On the one hand it would be nice to separate out the pieces that are only needed by objToJSON. On the other hand, most of src/datetime is copy/pasted from numpy, and diverging from the original is liable to cause headaches.

• src/helper.h defines PANDAS_INLINE. Cython has a CYTHON_INLINE that AFAICT would be a drop-in replacement. I have no idea how one would do that.

• src/skiplist.h has a duplicate definition of PANDAS_INLINE

• pandas_datetimestruct should either be aliased to or replaced by npy_datetimestruct (closed by use npy_datetimestruct instead of pandas_datetimestruct #21886)

• from libc.math cimport sqrt is not nogil, whereas cdef extern from src/headers/math.h... is. That appears to have a nontrivial perf impact. See simplify skiplist inclusion/cimport to be more cythonize-friendly #18420. Go through and make sure we're using the performant implementation.

Benchmarks (technically belongs elsewhere, but not worth opening a new issue)

• asv is particularly volatile on my machine, so this may be just me. But I'm consistently seeing these as outliers: timeseries.AsOfDataFrame.time_asof, timeseries.AsOfDataFrame.time_asof_nan, timeseries.DatetimeIndex.time_dti_tz_factorize, timeseries.DatetimeIndex.time_dti_factorize

[jbrockmendel]

@WillAyd Any thoughts on the suggestion to move timedelta_struct and associated conversion funcs out of the src/ directory and into cython? I'll volunteer to actually do the implementation, but want to get your blessing first.

Background: src/datetime/ is a big dependency, much of which is only used for src/ujson. If we can get a few things out of it, there are several tslibs modules that can drop that large dependency. Plus having things in cython makes it easier to test+profile+measure coverage.

[WillAyd]

I'm not an expert in this area. If you think it's a good idea then go for it. Let me know if I can help

[jbrockmendel]

Closing as all actionable items have been addressed.