add note on eltype and getindex

[kalmarek]

following discussion in #38422
@nalimilan

[simeonschaub]

I don't think that is correct. The latter just has to be a subtype of the former.

[kalmarek]

I learned about these assumptions when broadcasting on a struct of mine got broken in julia-1.6. Could you point me to any reference for typeof(first(itr)) <: eltype(typeof(itr))?

[vtjnash]

The later is not required, or expected, to hold true. The eltype is for iteration, not indexing.

[vtjnash]

What he means is that eltype may be any supertype, for example, it may be Any. What is expected to hold true is:

all(isa(eltype(itr)), itr)

[nalimilan]

Yes subtyping is enough. But why should eltype apply to iteration but not indexing @vtjnash? I don't really care, but that sounds natural.

[kalmarek]

Else, there are numerous counter-examples where the requirement doesn't hold, even for array (such as indexing with a BitArray or Vector)

ok,I thought that the docstring could say

`typeof(itr[idx]) <: eltype(typeof(itr)) ` holds for each `idx` in `eachindex(itr)`.

but indexable types do not necessarily implement eachindex...

[mbauman]

This is largely already documented in https://docs.julialang.org/en/v1/manual/interfaces/ — but again, what's required is dependent upon the abstract super type.

[kalmarek]

#39185 tries to document a case that was broken by 1.6 and does follow everything(?) what is in written in https://docs.julialang.org/en/v1/manual/interfaces/.
I'll wait for the conclusion of discussions in #39185 on what to write here exactly.

EDIT: the type in real life is already fixed (by introducing a type iterator struct), I just wanted to document the requirement. If the docstring of eltype is not the place to document it, I'd be glad if you could suggest a better option.

[mbauman]

Sorry, I didn't mean to imply that this shouldn't be documented here — just that the requirements depend upon the interface and it's hard to document exactly what this means without reference to this.

[kalmarek]

@mbauman no offence was taken ;)
Are there any discussions here (github) or elsewhere on Number interface? It seems that this issue should belong there.

[vtjnash]

I don't think this adds additional clarity over what is already stated there. Perhaps we could mention AbstractArray instead of just AbstractDict, and especially reference the interface documentation?

Closing as I don't think we should merge this version, but feel free to re-open or to push a new PR with updates.

[kalmarek]

@vtjnash I'm happy to work on this, but there doesn't seem to be a consensus on what is actually required for eltype.

with #39295 there will be a need to implement eltype equal to the type inferred by getindex to keep broadcasting the same as manual loop(?). I'm happy to return to this pull then.