Well, that's fun:

AssertionError: 
Items are not equal:
 ACTUAL: 'masked_array(data=empty((0L, 2L)),\n             mask=False,\n       fill_value=1e+20,\n            dtype=float64)'
 DESIRED: 'masked_array(data=empty((0, 2)),\n             mask=False,\n       fill_value=1e+20,\n            dtype=float64)'

Only one of our test instances actually catches this. I suppose we could do repr(shape).replace('L','') # for python 2 to work around it

Also, we didn't consider subclasses. Currently:

>>> class MyArray(np.ndarray): pass
>>> np.empty((0, 20)).view(MyArray)
MyArray([], shape=(0, 20), dtype=float64)
>>> np.ma.array(_)
masked_MyArray(data=[],
               mask=False,
         fill_value=1e+20)

With this patch, I think we lose the subclass information

The old behavior was to print as np.array([], dtype=float64) no matte the shape, which could be misleading.

Not true, it prints as array([], shape=(2, 0, 1), dtype=float64).

We could also print it as ndarray(shape=(2, 0, 1), dtype=float64), which would actually be evaluable.

Hmm yeah. The reprs for subclasses also aren't evaluable currently.

The ouptut for the size-0 array in your example is not evaluable, it gives MyArray([], shape=(0, 20), dtype=float64). But it needs to be MyArray(shape=(0, 20), dtype=float64).

Non-empty subclassed ndarrays also don't work:

>>> MyArray(shape=(2), dtype=float64)
MyArray([  7.74860419e-304,   7.74860419e-304])
>>> MyArray([  7.74860419e-304,   7.74860419e-304])
TypeError: 'float' object cannot be interpreted as an index

Since there seem to be so many cases already that aren't evaluable, I am tempted to give up on this PR and just say "the repr often isn't evaluable", and leave the non-evaluable array([], shape=(xx))) repr.

I'm in favor of leaving this the way it is. If anyone else wants to give it a shot feel free, I'm going to close this one.