Deprecate frexp{T<:AbstractFloat}(A::Array{T}) in favor of broadcast ?

[musm]

Now that we have . for broadcasting, the method frexp{T<:AbstractFloat}(A::Array{T}) doesn't seem as useful as it once did?

Currently the method returns (F,E) where F is an array of 'significands' and E is an array of 'exponents'. The behavior with the broadcast would instead return an array of tuples.

@ScottPJones also mentioned on gitter that it is also over specified, it only works for Array, not any AbstractArray

[simonbyrne]

Yes, we probably should deprecate it, though as you point out, the fact that there is no way to recreate this functionality via . is a little annoying, which is why I suspect it hasn't been done already.

[JeffBezanson]

The vectorized version of this is pretty weird. Is there a use case for getting these two arrays of values? Unlike other vectorized functions, returning a tuple of two arrays doesn't compose naturally with most other functions.

[stevengj]

If you just want an array of F you could do e.g. first.(frexp.(A))

[stevengj]

(I can't find any found one registered packages using frexp, vectorized or not: Compat uses the scalar frexp function in its linspace code. So I doubt that removing the vectorized version will cause too much trouble. The ArbFloats, GSL, Unitful, and Measurement packages define frexp methods, but don't actually use them except in tests.)

[simonbyrne]

I think we just deprecate and say it is no longer available. We don't provide vectorised forms of other 2-output functions (e.g. modf, divmod)

[simonbyrne]

Sorry, I accidentally pushed to master. Fixed by 2e0b001.