Mateusz Baran

> This would have the advantage that the two fields `cost` and `grad` that are currently within the problem would stay, > which is not possible in your sketch, since...

Yes, right, a unified cost/grad/etc. field would be a significant rework. So maybe a better (or at least easier to implement) idea would be to introduce additional fields in specific...

No, it's not urgent, actually the opposite. I doubt it will ever make anything more than 2x faster which is neat but rarely makes a difference :wink: . I think...

> Oh, then feel free to open more issues – there are several parts that I wrote to learn Julia around 2015/2016, so sure that might not all be optimal...

It may be beneficial to explain what could go wrong in different cases of type piracy instead of putting everything in one bag. ``` module PkgA struct C end bar(x::C)...

Most likely LLVM generates different SIMD instructions. Is this actually a problem in practice?

It might be fixable but I would guess that it's not worth the required effort. I really doubt this is actually serious. Lots of people use matrix multiplication from StaticArrays.jl...

Yeah, Julia's Base uses BLAS which does its own SIMD. Default `isapprox` is not ideal for comparing two numbers that are supposed to approximate 0 -- for that I usually...

That's a Julia broadcasting thing. This doesn't allocate: ```julia julia> function run2(N, x) y = zero(x) for i in 1:N y += map(z -> round(Int, z), x) end end julia>...

Yes, `LinearAlgebra` is correct. StaticArrays often works quite poorly with ill-conditioned matrices like yours. You'd have to replace the current solver in StaticArrays (there is a custom one for exactly...