Alexis Montoison

HI @lijas! `Krylov.jl` allocates a workspace based on the type of `b`. It's the best way that we find to support complex numbers, multiple precision and GPUs. The issue here...

If you are able to compute a complete factorization of `K`, you can also use the preconditioner `P^{-1} = [K^{-1} 0; 0 diag(B'*K^{-1}*B)^{-1}]` and use `TriMR` / `GPMR` instead of...

The issue with `similar(b)` is that you want dense vectors (`Vector`) even if your right-hand side `b` is sparse (`SparseVector`). The other issue with `BlockVector` is that `similar(b, n)` returns...

The user should provide `S` (`Vector{ComplexF32}`, `CuVector{Float64}`, ...) `T` and `FC` are determined from `S`. Is it not easier to just build a new solver with the constructor `XyzSolver(n, m,...

Can I close this PR ?

> This needs to be debugged in a Docker container... Yes :( I rebased the PR to see if the `mul!` revamp changed something.

`ARM` architectures have [Tier 2 / Tier 3 support](https://julialang.org/downloads/). The issue is probably not related to Krylov.jl.

We should not merge this PR because we have limited credit with DroneCI but we can keep it open and just rebase it at each major release to verify that...

Version 1.0 - [x] LNLQ (#128) - [ ] USYMLQR (#164) - [x] Regularization in CRAIG and CRAIGMR (#171) Version ≥ 1.0 - [ ] Support for complex linear systems...

Do you use the Julia binary built for ARM processor ? [macOS ARM (M-series Processor)](https://julialang.org/downloads/) We should open an issue to the Julia repository, this Julia binary has only `Tier...