FIX DataFrame diff with timedelta

[hayd]

fixes #4533

need to fix test in numpy 1.6...

[jreback]

hmm seems a bit special casey

[hayd]

true, perhaps better to change the dtype of arr_out / view, but couldn't get that to work...

[jreback]

I think you need to set the dtype of out_arr in the if block (e.g. if its float/int/datetime64/timedelta64) as it needs to be i8 (then you take a view before you return it, so need to remember to do that)

e.g. return out_arr.view('timedelta64[ns]')

[jreback]

@hayd ?

[jreback]

@hayd ?

[jreback]

@hayd ?

[hayd]

@jreback Did you think this was too special casey? (will change the view dtype and see if this passes).

[jreback]

@hayd it might be, but not sure how can avoid this, timedelta is ITSELF special casey (integer and datelike at the same time)

[hayd]

@jreback ok cool, in general it's good to be uncomfortable/point out/whine about special casey-ness. :)

[jreback]

actually I would do this slightly differently. I would define the dtype of out_arr (as the same as the input type), except for datetime64[ns], then I think you can use this as is.

[jreback]

closing this as picking it up in #8184

[hayd]

thanks @jreback, sorry for dropping the ball on a few PRs.