Michael Abbott

What are you picturing this returning? Zygote's `OneElement` is something very close to this, which could perhaps be moved upstream if it has other uses. Allows any `ndims`, any axes....

And worse if you repeat them: ``` julia> tmp[[2, 1, 1], :] 2-dimensional AxisArray{Float64,2,...} with axes: :x, [:b, :a, :a] :y, [:A, :B, :C] And data, a 3×3 Array{Float64,2}: 0.352127...

Yes, would be nice if this worked. My impression is that `*` and `mul!` are teetering right on the edge of what's possible with multiple dispatch without going crazy, and...

Can't this be simpler though? `mul!` simply has to check that `axes(A,2) == axes(B,1)` etc, and then strip everything off (I guess with some `unoffset` function which OffsetArrays can extend),...

Here's a sketch of what I was trying to say above: https://gist.github.com/mcabbott/03225572217660aa7694f46caaa6e15f There aren't so many calls to `similar` which need to be changed from `size` to `axes`. Anything with...

Thanks for having a look. I guess this is an attempt to change as little as possible. It wouldn't preclude doing a better `matmul3x3!` later... but wouldn't you would still...

I guess I was thinking of the pointer as being the maximally unwrapped form, just the memory (& the eltype). Before you pass that along, you'd need some logic to...

OK, good example. It's a bit ugly if `remove_offset_indexing(:: ZeroIndexedSVector)` adds an `OffsetArray` wrapper, just to feed the existing `generic_matmul!` something it understands. The other objection to this design is...

Yes to strides-of-the-parent, and JuliaLang/julia#30432 seems like a very neat mechanism to let this pass information about non-stridedness too. Not sure I'm volunteering to re-write `mul!` in terms of that...

This seems like a bug to me, since it has an explicit `atvalue`. Compare: ```julia julia> z = AxisArray(rand(3), sy=[:a, :b, :c]); julia> z[atvalue(:b)] # explicit 0.5205382494635875 julia> z[:b] #...