Supporting @lens _[end] etc?

[tkf]

Are you planning to support those cases?

l = @lens _[end]
@test get(l, [1, 2, 3]) == 3

l = @lens _[end-1]
@test get(l, [1, 2, 3]) == 2

l = @lens _[end-1][end]  # some complex case
@test get(l, [[1, 2, 3], []]) == 3

l = @lens _[:end]  # just make sure that symbol index is still ok
@test get(l, Dict(:end => 1)) == 1

[tkf]

Also with colons, like @lens _[end-2:end].

[tkf]

I actually looked at the code and tried to fix it. But then I realized that you need a little interpreter for the indexing operations involving the arithmetic operations and : (and what else?). You also need to be careful about what to interpret at the macro eval time; e.g., @lens _[f()-end] should eval f(), replace the result x with the original indexing AST, and store x-end. This is not a super hairy problem but requires some non-trivial code. Is it the only way to do it?

[tkf]

I had a quick try: #16. It turned out the indexing lenses lose type stability unless we make the indexing interpreter type stable as well. It means to infer the type of the expression without evaluating it when @lens is invoked. I guess it can be done by, e.g., injecting some dummy integer to end, eval the index expression, and store the type of the result as the type parameter of IndexLens. But I'm still not sure if it makes the indexing interpreter type stable.

[tkf]

Maybe more reasonable approach is to do the eval in @set and only allow concrete indices in IndexLens.

[jw3126]

I would definitely like to support end (I think in future there will also be begin). However I don't really now how x[1:end-1] works internally etc. Does it get lowered to something like x[1:lastindex(x) - 1]?

[jw3126]

Okay I think I have an idea how to implement this treating the 1:end mechanism as a blackbox.

struct OOPLens{S,G,M} <: Lens
    set::S
    get::G
    modify::M
end

and @lens _[fancy_index_expression] expands to

let 
    set(obj, val) = obj[fancy_index_expression] = val; obj
    ...
   OOPLens(set,get,modify)
end

One detail is, that obj[fancy_index_expression] = val involves setindex! and we might like setindex. We could however patch this by wrapping the object

struct SetindexWrapper{O}
    obj::O
end
setindex!(w::SetindexWrapper, args...) = SetindexWrapper(setindex(w.obj, args...)

and replace e.g

    set(obj, val) = obj[fancy_index_expression] = val; obj

by

    set(obj, val) = w = SetindexWrapper(obj); w[fancy_index_expression] = val; w.obj

[tkf]

Ooooh, this is much cleaner solution.