Arraypocalypse slicing compatibility

[c42f]

Not a lot has happened over at JuliaLang/LinearAlgebra.jl#255 yet, but the breakage from APL-style slicing (JuliaLang/julia#13612) is already apparent. I hope there's something that can be done in Compat to ease the slicing transition. Here's one possibility:

1. For 0.4, @compat A[:,1] rewrites the :ref AST node to include a placeholder type which can be dispatched on - something like A[SliceAPL, :, 1].
2. We then implement special versions of getindex and setindex! taking a SliceAPL as the first index, which inspect the types of the incoming slice expression and squeeze out any singleton dimensions as required.

[c42f]

Ugh, the above approach doesn't work cleanly due to a raft of ambiguities when defining the appropriate getindex. Might be better to wrap the array itself in a new array wrapper type, ie A[:,1] rewrites to ArrayAPL(A)[:,1]. Then define methods on the ArrayAPL type. This would probably generalize to some of the other anticipated breakage.

[martinholters]

Are you (or is anyone else) working on this? The occasions where I think this would come handy seem to become more often, lately...

[c42f]

No, I'm afraid I haven't had time to look at it. I've been stuck on 0.4 lately.

[andreasnoack]

Is it the fact that A[1,:] returns a Vector that is an issue here? If so then remember that A[1:1,:] and vec(A[1,:]) return the same types on 0.4 and 0.5.

[martinholters]

Right, and vec(A[1,:]) is even shorter than @compat A[1,:]. But is there a similarly nice construct for more dimensions?

Oh, and beyond the scope of this issue maybe, but what about sparse matrices?

[c42f]

@andreasnoack good point. It's not an ideal solution though - it litters package code with constructs which are not obviously related to backward compatibility.

[andreasnoack]

But is there a similarly nice construct for more dimensions?

I cannot come up with anything nice right now. I guess that squeeze(1, A[1:1,:,:]) would work but that is not nice.

@c42f I'm fine with Compat support. I just wanted to point out that a solution existed.

[timholy]

The best approach would be to "backport" the getindex changes to Compat, using a different name and having @compat call the new name. Might not be a small job, however.

[c42f]

@timholy I guess this would mean the @compat macro would need to understand the lowering from :ref AST nodes to getindex/setindex? Does that present problems?

[timholy]

I think it would be manageable, but probably the challenges will be clear only by trying it. I would think about adding a global variable is_lhs ("is left-hand-side") and have @compat set it/check it when parsing Expr(:(=), args...).

[martinholters]

I've tried @c42f's idea of rewriting A[...] to ArrayAPL(A)[...] and then providing getindex for that - looks like it might be working. I'll open a PR for discussion soonish.

[andreasnoack]

Is this still relevant?

[martinholters]

Theoretically, yes. We still support Julia 0.4 which differs from Julia 0.5 and later in this respect and have no @compat for it. Practically, it looks like no-one is missing it, given how dormant this thread is.

I'd say we keep this open until we bump minimum Julia version to 0.5, close it then (along with JuliaLang/julia#179).

[c42f]

In practice I haven't found this to be a major pain point in packages I maintain. It's a bit of an annoyance on occasion, but working around it manually isn't too hard.