Andreas Noack

Sounds reasonable. This could potentially also make the functions type stable but now that we have constant propagation this might not be an issue anymore (but would have to check...

This should probably go on https://discourse.julialang.org/ instead of the issue tracker but it should be possible to do something along the lines of what you describe. See this example https://github.com/JuliaParallel/Elemental.jl#truncated-svd...

Just repeating my reply from Slack here for the record. So `teig` isn't documented and exported because it's not complete although it's pretty close. That means that `F[2]` doesn't contain...

So https://github.com/JuliaLinearAlgebra/TSVD.jl/pull/26 fixes your minimal reproducer but I'd be interested in knowing if it also fixes your real use cases so please comment if the issue remains after the patch...

Thanks for the new example. I think a proper fix here will require a some real effort. One of the new vectors is proportional to the previous basis vector so...

> Yes you are right, after 0.7 I'm planning to clean these up. Awesome > Are you saying that the vector `A` might be freed in the following code? weird....

> But there's a guarantee that overriding `unsafe_convert(::Type{Ptr{T}}, ::MyMatrixType`) will not garbage collect when used in a ccall? Yes but that only takes care of the pointer within the `ccall`....

The default solver for Toeplitz matrices is an iterative solver since it's generally not possible to benefit from the Toeplitz structure in a direct solver (to my knowledge). The iterative...

Wouldn't the simplest solution be to provide MKL_jll from a centralized package installation on the cluster?

Yeah. It would make sense to reexport `LinearAlgebra`. It's unlikely that you'd only like to load this package without `LinearAlgebra`.