Is input arrays for the JacobianCache constructor supposed to alias sometimes?

[pkofod]

In this constructor https://github.com/JuliaDiffEq/DiffEqDiffTools.jl/blob/4c51240b7576327f09710ff6ace5db044ca4cfdc/src/jacobians.jl#L58 fx1 is aliased, but generally it seems that input is only used to construct new arrays. So if you input and x or fx, it'll generally be used in the context of similar(x), or fill(..., size(fx1)). I'm just wondering if this is intentional (you might use fx1, the forward evaluated vector somewhere) or not.

[dextorious]

I think the idea was that we might want to alias it somewhere but never actually got to needing to do that. It's been a while, so I don't have a firm recollection of the exact reasoning I had at the time, but it wasn't random at the time.

[ChrisRackauckas]

I think the idea was that we might want to alias it somewhere but never actually got to needing to do that

DiffEq uses it.

[pkofod]

I was thinking about changing this in NLSolversBase, so the constructors actually use the exact arrays you use to initialize the objective types. You could then turn it off with a copy_input kwarg

[ChrisRackauckas]

Closing because it is indeed intentional. You can build the cache using pre-existing arrays, and the cache is sized so that the largest cache methods (central differencing) don't need to allocate during the runs.