nalimilan
commented
7 years ago One thing that would be nice to resolve is if we could remove the need for explicitly typing the predicate function as ::Function without running into ambiguous signatures.
Right, but I'm not sure it's possible to distinguish, when two arguments are passed, whether they correspond to (pred, A)/(A, pred) or to (A, start), at least if we want to support arbitrary indices (which I think is a good idea in the long term).
Maybe a new function is needed to return a Union of all possible index types supported by a collection, which would be used to restrict the signature. Or conversely a Callable trait could be added to pred, which would never be used for indices.
I also don't think we necessarily have to constrain ourselves to having the predicate function first. The syntax it enables ( find(A) do elt …) reads to me like it'll only operate on the nonzero elements instead of what it's actually doing: transforming elt to a test to see if its index should be included in the output. In fact, I think the reverse argument ordering might read a bit more naturally: find(A, #= where =# x -> x == 2).
At least the current order is consistent with sum and mean. But I agree it feels a bit weird. I'd be inclined to choose whatever solution is simpler to implement, both to avoid ambiguities in signatures, and regarding the deprecation plan.
If you're satisfied with this first draft, I'd be inclined to merge it and allow others to propose PRs to specific sections instead of dragging this PR out.
Actually, I find it more practical to add comments to the PR itself. One it is merged, people need to file issues, which are not automatically grouped together, and which don't automatically get context like comments do. So I'd prefer continuing that way unless you prefer merging it soon.
StefanKarpinski
Very nicely done; this is a great summary. Thanks for doing this!
One thing that would be nice to resolve is if we could remove the need for explicitly typing the predicate function as
::Functionwithout running into ambiguous signatures. I also don't think we necessarily have to constrain ourselves to having the predicate function first. The syntax it enables (find(A) do elt …) reads to me like it'll only operate on the nonzero elements instead of what it's actually doing: transformingeltto a test to see if its index should be included in the output. In fact, I think the reverse argument ordering might read a bit more naturally:find(A, #= where =# x -> x == 2).If you're satisfied with this first draft, I'd be inclined to merge it and allow others to propose PRs to specific sections instead of dragging this PR out.