we have a lot of should-be-unnecessary ensure_platform_int calls. these should be extremely cheap, so haven't been bothering to remove.

so haven't been bothering to remove.

And in the past I bothered making a fast-path take_1d which explicitly removed it. It did show up in reindex benchmarks IIRC (#40246, #39692)

Digging up the benchmarks. Doing one specific for take_1d:

from pandas.core.array_algos.take import take_1d
arr = np.arange(10000)
idx = np.arange(100, dtype=np.intp)

In [5]: %timeit take_1d(arr, idx, allow_fill=False)
498 ns ± 0.77 ns per loop (mean ± std. dev. of 7 runs, 1000000 loops each)  # <-- master
465 ns ± 0.962 ns per loop (mean ± std. dev. of 7 runs, 1000000 loops each)  # <-- PR

This is of course about nanoseconds, so in absolute terms only a tiny difference. But because this function is used repeatedly in things like unstack and reindex, it does make a noticeable difference. Eg on the unstack case as used in #40246, it gives ~ 3-5% difference (which is still small, but IMO worth it since this is actually removing a single line, and the constraint is explicitly documented / in the type information)

sure no objection

no objection, can you rebase