Cholesky Update

[Deduction42]

There's a couple of functions in LinearAlgebra.jl around updating Cholesky decompositions:

• lowrankupdate!
• lowrandowndate!

You can find them starting here:
https://github.com/JuliaLang/LinearAlgebra.jl/blob/27b0f1321777fd2f935b8b9f1cefc61e12114b29/src/cholesky.jl#L882

This looks like a lot of mutating code, but would it be possible to use Accessors.jl to emulate this mutating behaviour? The annoying thing about lowrankupdate! is that it destroys the input vector, so the non-mutating versions lowrankupdate have to copy everything (hence allocations). Being able to get this to work for StaticArrays allows me to bypass these allocations and wierd side-effects. How feasible do you think this is?

[mateuszbaran]

The easiest way of handling that in StaticArrays.jl would be doing the mutation in non-escaping MArrays. Then the allocation is optimized away. I think implementing that wouldn't be too much work.

[Deduction42]

Nice! But wow, I thought I was keeping up with Julia features but I totally missed escape analysis! I guess I learned something new today.

[mateuszbaran]

Julia has had escape analysis for mutable structs inside a single function for a few years IIRC. The big thing people are talking about is inter-procedural escape analysis but that's still not ready.