Matt Bauman

Not naive at all. I could even get behind doing that for _all_ `SubArray`s on the assumption that other `.parent`s may specialize on nonscalar indexing.

Am I correct this is specifically about division? Classic example: ```julia julia> spzeros(5,5) ./ ones(5) 5×5 SparseArrays.SparseMatrixCSC{Float64,Int64} with 25 stored entries: [1, 1] = 0.0 [2, 1] = 0.0 [3,...

Right. The general case is all n-ary functions `f` with `f(zeros(n)...) != 0` and with `f(args...) == 0` when some nonzero (array) arguments exist in `args`. Division is the most...

I'm not sure about what to call this function, but could we possibly do a pigeon-hole sort of optimization within the `broadcast!` definition itself when we detect this case?

I **_really_** appreciate what you’re trying to do, but I’m still finding this fairly complicated and confusing… and enough text that my eyes start glazing over. When I look at...

The crux of the difference, I think, is that when `Base.unalias` needs to make a copy of a view, it'll do so in an efficient manner (while still preserving type...

Here's a simpler reproducer. The difference is whether `b` is 1- or 2-strided: ``` julia> using LinearAlgebra julia> A = [-0.23190906957695134 1.1722448627021573 -0.26988492037793593 0.13457571047366457 -1.0745769949749375; 0.940390210248594 -1.696813019281842 0.5780036022342506 0.16554255091922548 -1.3819748518823003;...

Yeah, I'm sure this is happening somewhere inside the OpenBLAS assembly kernels, which will of course vary by arch. I still see the issue regardless of numbers of threads — ...

I have a few questions for which I don't have solid answers: * Does it make sense for the other methods of `empty` (the non-array ones) to allow types, too?...

The one trouble is that OffsetArrays adds reshape methods that use ranges (axes) to specify the resulting shape. That use-case becomes ambiguous here.