make [] give an Any array

[JeffBezanson]

This change seems to be pretty popular from what I gather.

[jiahao]

👍

[ivarne]

🎆 🍰

[johnmyleswhite]

+100

[IainNZ]

https://www.youtube.com/watch?v=uvqJ1mTkEuY

[StefanKarpinski]

I haven't watched this but I'm assuming you rickrolled us.

On Sep 26, 2014, at 8:23 PM, Iain Dunning [email protected] wrote:

https://www.youtube.com/watch?v=uvqJ1mTkEuY

—
Reply to this email directly or view it on GitHub.

[JeffBezanson]

You know, Array(None,0) is generally pretty bad, but it does have one advantage: it lets you pass [] to any function accepting Array{<:T}, since None is a subtype of everything. It's handy not to need to specify an element type when you want to pass zero elements.

[ivarne]

@JeffBezanson most of those functions will not work though.

[StefanKarpinski]

Without covariant arrays, that's not all that useful.

[JeffBezanson]

Fair enough.

[nalimilan]

A plotting function with a keyword array argument containing zero or more values, e.g. plot(...; marks::Vector{Int}=[0]), could be called via plot(.., marks=Int[]), but not via f(..., marks=[])? Could be slightly annoying.

But making [] give an Any array has clear advantages in other places.

Or wouldn't it be possible to acknowledge the fact that all empty arrays are alike, so that their element type does not matter? Is there any situation where the element type of an empty array is important?

[JeffBezanson]

[] cannot be passed to f(x::Vector{Int}) today, which is why my point above is not very hard-hitting. Only f{T<:something}(x::Vector{T}) would work.

The element type of an empty array absolutely matters, because vectors can grow, and sum(Int[]) will give a different type of zero than sum(Float32[]), etc.

[nalimilan]

Ah, right. Sometimes you don't care about the type, but sometimes you do. One could imagine converting e.g. Any[] to Int[] when the only signature that matches expects the latter, but that sounds a bit hackish.

[eschnett]

Converting None[] to Int[] would make more sense from a point of view of types, since None is a sub-type of Int...

[StefanKarpinski]

Objects can't change type in Julia and in a dynamic language – especially one with as much polymorphism as Julia – reverse propagation of types is not really feasible.

[IainNZ]

Hard to tell how much this broke as so much still broken from #8420, but, didn't add too many bodies to the dead packages pile.