PERF: use fastpath=True in Index methods (delete/drop/insert/etc)?

[immerrr]

Been hit by this when optimizing index-oblivious Blocks. In my case,

is_deleted = np.zeros(len(index), dtype=np.bool_)
is_deleted[deleted_loc] = True
index = index[~is_deleted]

was still a lot faster then than index.delete(deleted_loc).

It appears, that delete doesn't add fastpath=True to ctor and that triggers type inference for string (object) indices. There seems to be plenty of methods that do it the same way and hence are slow too, can we do something about that?

[jreback]

sure
want to do a quick pr ?

vbenches a plus but not that necessary
maybe spot check ok

[jtratner]

so what's the downside here?

[immerrr]

@jtratner it's the same as one of my earlier PRs about __getitem__: we drop type inference that might've changed index type after deletion of non-matching elements. That is if we go with "trivial" solution.

If trivial is not necessary, then dtype inference may be preserved without the necessity to traverse each element of dtype=O array. If one made an int8 array of object classes (or int 16 if 255 classes is not enough) to accompany the main one, then instead of enumerating all entries each time it would suffice to do single np.bincount(x.ravel()).nonzero() to find out which classes are present. Memory overhead for that int8 array should be acceptable, but it may require some non-trivial effort to make sure they're in sync.