use NTuple as the representation of a CartesianIndex

[JeffBezanson]

cc @timholy @mbauman @Jutho
What should I run to make sure there are no performance regressions here?

[timholy]

👍. Use the new perf/array tests.

[Jutho]

+1

[JeffBezanson]

Great, ran the tests, examined results with DataFrames, looks good.

[mbauman]

Awesome.

[mbauman]

Bonus 🍰 for speeding up scalar SubArray indexing of small arrays by 6-8x.

[timholy]

I don't really understand how that's possible, but I'll take it! (An NTuple shouldn't be better than a custom immutable.)

Overall I have the sense that the loss of regular perf testing means that we've kind of been bouncing around.

[IainNZ]

AFAIK a new machine has been ordered that we should be able to use for performance testing, so I'm hoping we can grind out (or at least identify) perf regressions before 0.4.0

[ScottPJones]

When will you get this new machine? (would that also be the machine for running PackageEvaluator?)
I'd be happy to do nightly runs on my MacPro, if that would help, until you get it (it's not in use at night anyway)

[jakebolewski]

Realistically I think end of August would be the soonest we can hope for.