Safe non-traditional array indexing

[timholy]

With reference to #16260 (comment):

• transitionally (julia-0.5 only) make size and length throw an error for arrays with unconventional indexing
• introduce @arraysafe to rewrite calls to size and length to something that doesn't throw an error
• merge allocate_for into similar

[StefanKarpinski]

@timholy, where are we at on this?

[timholy]

As far as I'm concerned, all important known issues are closed by one merged PR (#17137) and two open PRs: #17228 and #17355. It would be nice to have a better deprecation for first(::Colon), see discussion starting at #16260 (comment). But once we decide on something it's a couple of lines to implement it. EDIT: it occurs to me that one good strategy would be to deprecate it but not provide a replacement in Base. The warning could link to a section of the devdocs that provides guidance. I would guess that almost all uses would be covered by indexoffset as described in #16260 (comment).

However, I've gotten hints of some discontent, so most importantly see #17338 (comment) and #17228 (comment). FYI the DataArrays breakage is already fixed at JuliaStats/DataArrays.jl#205, with tkelman/DataArrays.jl#2 cued to improve the fix.

[timholy]

Oh, if folks are giving this some attention but need some time, my "merge by Monday" can obviously go on hold. I put a deadline on it because it seems like a bad idea to let them sit until moments before julia-0.5RC and THEN decide to either merge or revert (either of which could be disruptive). We should decide soon, but it doesn't have to be today.

[timholy]

Since @wildart seems to be coming to completion, I'm inclined to start hitting "merge" on my PRs. Speak now or...

[StefanKarpinski]

I think @JeffBezanson may have a few reservations, so maybe give him a chance to comment.

[timholy]

OK, that's the feedback I was waiting for. Silence is very hard to interpret.

[timholy]

#17228 seems ready for review. It also includes the revised deprecation for first(::Colon).

I guess there is one potential last dangling issue: what to do, if anything, about #16378? The fundamental problem is that (i,a) in enumerate(A) is ambiguous: does i count entries of A, or is it an index for A? The name suggests it's a counter, and that's occasionally useful, but I'd wager that in practice it's more commonly used as an index for A. We can merge #16378, but I've never been thrilled with the name eachindexvalue, nor with having something so drattedly similar to enumerate. EDIT: or we could declare that enumerate, despite the implication of counting, actually returns an index.

[eschnett]

For dictionaries, enumerate adds a one-based index. That is not really comparable to arrays since dictionaries iterate over (key,value) pairs by default, while arrays don't, but it indicates that enumerate uses one-based enumeration even if a key is present.

What about indexedvalues? Similar to enumerate, the prefix each is not really necessary, and adding an index could be called "indexing".