[Nightly] failing linsolve test

[jpthiele]

The nightly workflow fails on the 'linsolve' test.
It issues a bunch of warnings about LLVM calls with integer pointers being deprecated.
I suppose this leads to the wrong comparison result in the end.

However, all the warnings are issues for upstream packages like LoopVectorization or LinearSolve.

@j-fu what is the best way forward, should I open some issues to make the upstream package maintainers aware?

[j-fu]

The failing test is about allocation behavior. The integer pointer stuff are just warnings
which might go away with newer versions of nightly.

So these IMHO are essentially two issues.

I think it is worth to open an issue (probably in RecursiveFactorizations.jl), but not without an MWE and a bit of furher investigation. And we would need to look for already open issues/PRs relevant to this.

For context (note to self: document the test): it is about having a non-allocating inplace LU factorization for generic number types (including ForwardDiff.Dual). With mixture flows, I want to solve the Stefan-Maxwell ansatz to obtain fluxes from their implicit description in each flux function call.

RecursiveFactorizations.jl provides this. But it depends on LoopVectorization.jl whose fate is a bit brittle.

I have my own (albeit non-pivoting) implementation based on the Doolittle LU algorithm as a fallback. See vfvm_functions.jl which may be good enough.

Also, may be now there are more alternatives available besides RecursiveFactorizations.