Export Base.RefValue

[jakobnissen]

RefValue is really useful - not just to create mutable const objects, but also to create mutable fields in large immutable structs. However, that requires importing Base.RefValue - which is not exported.

So what's the deal with RefValue? Is the type simply an internal implementation of Ref, not to be relied on in user code? If not, can we export it from Base? If so, can we have some exported type that can be used for a similar purpose?

[vtjnash]

It is a very inefficient (slow) way to create a mutable field in a large mutable struct (the efficient way is to make the whole struct mutable), so we try to dissuade people from using it directly. I believe there's old issues discussing this, but can't quickly find them.

[MasonProtter]

I had thought that these efficiency problems were basically solved in 1.5? Is that not the case?

[cscherrer]

I don't have the code any more, but I did some benchmarks about a year ago (using Julia 1.5 IIRC) and saw that immutable structs can be much faster than mutable. If you have an immutable struct with a single Ref field, wouldn't this only introduce over head to that particular field, so the others would still be faster?

If this is correct, then what's preferable would seem to depend strongly on the access patterns, even if accessing a single Ref is slower than accessing a field in a mutable struct. But maybe I'm missing something?

[vtjnash]

Using Ref / RefValue will nearly always be worse than choosing mutable struct when applicable. The general rule of thumb is that you want to put mutable at the outermost container that you can. This implies that most types will be struct, but that avoiding writing mutable when it is actually applicable would also cost significant performance. And that you should almost never use Ref (or subtypes), as that gets this exactly backwards.

[cscherrer]

Thanks @vtjnash . I don't understand how this can be, and it entirely breaks my mental model. But your understanding of this is a lot deeper than mine, so I'll just trust you and keep an eye out for ways to resolve the cognitive dissonance :)

[jakobnissen]

I'm closing this issue, as the original use case (including it as fields of a struct) is not as useful as I thought.

[nsajko]

A Ref-killer package, meant to supplant Ref for non-ccall-related use, implementing generic and efficient zero-dimensional arrays is being registered:

• https://github.com/JuliaArrays/ZeroDimensionalArrays.jl

Discourse announcement post:

• https://discourse.julialang.org/t/ann-zerodimensionalarrays-jl-zero-dimensional-arrays-references-boxes/128002

Once the package is registered I'll try to point to it from the Ref doc string.