Probably needs testing against the scipy and pandas testsuites

Is there a concrete benefit besides it being a little cleaner?

Primarily that it will smoke out bugs downstream where shape=() was intended to mean "create a scalar", but has been implemented to mean "ignore this argument".

Having this warning locally was what discovered the bugs #15881, #15882, and #15885.

The most unsure I was was whether descriptors should keep accepting None to mean no-subarray.

As in, np.dtype((int, None)) and np.dtype([('a', int, None)])? As far as I can tell, these are already an error, albeit a very cryptic one.

A test for array creation functions might be good (they seem the most visible users), or test it directly with the new testing functions.

I'd forgotten I now have tests for the converters. A test for PyArray_IntpConverter seems very sensible.

Thanks, was just going to say that the tests seem still needed. The only usage in scipy is fine. If anyone needs it, I guess we could consider making the optional version public, but dunno...

@eric-wieser sorry for the followup, those commands (dtype([('a', '<i8')])) work for me on master? Or is this platform dependend somehow?

Huh, I think I was testing those on 1.17.0. Did we add support for None on purpose?

I don't recall adding it on purpose, was probably an accident... Anyway, it doesn't matter too much? Just deprecate it again now. I don't mind if None works, but it also doesn't seem very useful for me (and () is identical, which actually makes it confusing rather than helpful).

The most unsure I was was whether descriptors should keep accepting None to mean no-subarray.

I still can't work out the case you're describing here. Can you give me the input/output pair that you're referring to

With this patch, the suggestions I mention above give ValueError: mismatch in size of old and new data-descriptor due to it attempting to construct a union type. I'm pretty sure master behaves the same way.

Eric:

In [1]: np.dtype((np.int64, None))                                                                                      
Out[1]: dtype('int64')

works for me, in master as well as 1.18

Got it, thanks. Works for me too in this branch with no warning. My issue was using int instead of int64.

Ah OK, I guess this is already special cased there, sorry for the noise.

I think I'm fine continuing to allow None in subdtype specs, to answer the original question.

Thanks for the test Eric, lets put this in.