Simon Etter

It seems to me that the "correct" way to distinguish these cases is just standard dispatch: ``` function rrule(::typeof(det), x::AbstractMatrix) F = lu(x; check = false) Ω = det(F) function...

I skimmed through the mpmath codes (and a few others), but at most I took ideas, no code.

The main reason for the generated functions is to avoid heap allocation. The structure of the algorithms is to first compute `k_i for i = 1:nsteps` and then use them...

Right, I agree I overused `@generated` for these two functions. Sorry about that, fixed it. I also agree that there probably isn't much point in avoiding heap allocations for `BigFloat`s,...

The only remaining issues with this PR seem to be about the names of the functions. Currently, we have - `ellipj` following Matlab and SciPy - `jpq` for `p,q in...

This is intentional and consistent with `Base.sqrt`: ``` julia> sqrt(-1) ERROR: DomainError: sqrt will only return a complex result if called with a complex argument. Try sqrt(complex(x)). Stacktrace: [1] sqrt(::Int64)...

Repeating Mason Protter's question: Is there a timeline for this to merge?

Once again, is there any timeline on getting this merged? If not, why?

Here's a list of remaining issues with this PR: - [ ] Decide on function names (see https://github.com/JuliaMath/SpecialFunctions.jl/pull/79#pullrequestreview-101299389). Fixable in 15 minutes once the decision is made. - [ ]...

Continuing the discussion from https://github.com/JuliaDiff/ChainRules.jl/pull/196#issuecomment-648746715: > Whether or not it plays nicely with Unitfuls out of the box is an implementation detail. Well it's part of the interface, so not...