perf: avoid inflating UnmaskedArrays in broadcasting when you can

[jpivarski]

@martindurant

[martindurant]

Thanks, @jpivarski .

Have you considered implementing something like asv for ak, to make sure that new changes don't accidentally hurt performance?

[jpivarski]

Airspeed Velocity? Maybe, although we'd have to define some large samples and a suite of computations to test. We have a huge number of tests, since they were added as features were added and bugs were fixed, but adding a suite of performance tests would be starting from scratch because all of the functionality tests are way too small to be meaningful for performance.

This particular PR, however, won't make anything slower. There's a special case that will get faster. The only cost here is the added complexity for maintenance. I thought this one was worth the cost because it cleanly splits out as a special case and it wouldn't be a very rare special case: conversions from Arrow are likely to introduce UnmaskedArrays (because Arrow nodes are option-type by default, but often don't actually have missing values).

[martindurant]

You mean when mask is None? Makes sense.

About asv, agree that it would be quite an undertaking, maybe a good project for a student, for a select few workflows.

[pfackeldey]

Looks good to me.

If anything you may want to add a dedicated test with a comment to this commit. That may make future maintenance a bit easier.

[jpivarski]

I couldn't think of a test because it doesn't affect functionality, just performance. And the array I was using to test performance is too big to save in scikit-hep-testdata.

Thanks for checking! I think I'll merge this now.