PERF: is_bool_indexer

[jbrockmendel]

Besides avoiding an array allocation, thus avoids the np.asarray call that was one of the main offenders in #41632

[WillAyd]

Is there any concern with the benchmarks being somewhat misleading due to short circuiting? At least in the Python space similar functions don't seem to benefit much from this

In [6]: %timeit all(isinstance(x, bool) for x in obj)
6.7 µs ± 136 ns per loop (mean ± std. dev. of 7 runs, 100000 loops each)

In [7]: %timeit np.array(obj).dtype.kind == "b"
5.98 µs ± 121 ns per loop (mean ± std. dev. of 7 runs, 100000 loops each)

[jbrockmendel]

Is there any concern with the benchmarks being somewhat misleading due to short circuiting?

Only one of the 4 cases posted short-circuits. Am I misunderstanding the question?

At least in the Python space similar functions don't seem to benefit much from this

I was surprised by this too, found that doing it in cython was much more performant after talking with @seberg yesterday about how numpy does this inference.

[seberg]

Wow, the difference is big...

NumPy should maybe try to be a bit faster, but that will just make the gap a bit smaller (e.g. by assuming that usually all elements are the same type, I guess a true casting support np.asarray(..., dtype=bool, casting="equiv") would fill the gap, but no idea how likely that would be). Unlike NumPy, you already know what you want here (bools) and what you got (a 1-d list).

[WillAyd]

Ah my bad on short circuiting - misread it. Thanks for comments

[jreback]

thanks this looks good