RFC: Improved implementation of SubArray

[lindahua]

Based on the discussion in #3496

[lindahua]

It is just in the status of initial setup. I appreciate your feedback as this work proceeds.

[lindahua]

I agree with allowing dim > ndims(a).

I am going to first work on the main functionality. It is easy to add bound checking afterwards, when the main interface stabilize.

[lindahua]

There is a potential safety issue here:

Consider the code below:

a = rand(100)

s = sub(a, 1:10)
push!(a, 2.0)   

x = s[1]   # this can be dangerous, as the base address of a may have been changed during push!

Even when though s maintains a reference to a which prevents a from being released, s.ptr can become invalid after an element is pushed to a (this operation may or may not change the base address of the pointer).

[lindahua]

I suspect that we don't have to cache the pointer for performance when Jeff's optimization work is ready.

[ViralBShah]

@JeffBezanson Do you think it is ok to go ahead with this approach for now, to improve SubArray indexing?

@lindahua Why not use ArrayView instead of MultiDimView as a name?

This will also give a fast implementation for #3701 and help move towards the new design.

[timholy]

@ViralBShah, I think we still need some additional technology. See #1917 (comment)

[ViralBShah]

@StefanKarpinski What do we do here, given the new work on array views?

[ViralBShah]

I suggest we should close this and focus our attention on views.

[ViralBShah]

While this PR implements views, what I mean is to focus on getting our indexing operations to return views rather than copies, and work on making that really fast.

[StefanKarpinski]

I've been working on some fast linear indexing for multidimensional array views. Should have something to share soon.

[lindahua]

Close. My latest efforts are at #5513.