BLD: 32-bit builds failing

[jreback]

https://travis-ci.org/MacPython/pandas-wheels/jobs/475157492

failing with

Exception ignored in: 'pandas._libs.tslibs.conversion.tz_convert_single'
Traceback (most recent call last):
  File "pandas/_libs/tslibs/conversion.pyx", line 663, in pandas._libs.tslibs.conversion._tz_convert_dst
ValueError: Buffer dtype mismatch, expected 'int64_t' but got 'int'
Traceback (most recent call last):
  File "pandas/_libs/tslibs/conversion.pyx", line 663, in pandas._libs.tslibs.conversion._tz_convert_dst
ValueError: Buffer dtype mismatch, expected 'int64_t' but got 'int'
Exception ignored in: 'pandas._libs.tslibs.conversion.tz_convert_single'

[jreback]

cc @mroeschke @qwhelan I think the perf enhancement changed this

[jorisvandenbossche]

I suppose you meant to tag @mroeschke instead of @mrocklin :)

It's about this PR: #24491

[jreback]

yep thanks

[qwhelan]

@jreback Certainly seems like my PR is to blame - I'll take a look

[qwhelan]

So what's happening is:

• np.searchsorted() always returns a result of type np.intp, which is 32-bit on 32, 64 on 64. It's described as "Integer used for indexing" here: https://docs.scipy.org/doc/numpy-1.13.0/user/basics.types.html
• int64_t[:] pos needs to be intp_t[:] pos as a result
• Seems we've always been unable to properly index into an array of size greater than int32 on 32-bit via np.searchsorted. I'm guessing there's no test coverage for something like this, but would be easy to construct a test case for.

Just opened #24621 to implement the above.