Make Core.checked_dims public

[LilithHafner]

It turns out it's nontrivial to implement Core.checked_dims (e.g. #54255). I want it to be public from base so folks don't have to re-implement it in packages.

[KristofferC]

I want it to be public from base so folks don't have to re-implement it in packages.

Any examples of such implementations that folks had to reimplement?

[giordano]

Any examples of such implementations that folks had to reimplement?

https://github.com/JuliaArrays/FixedSizeArrays.jl/blob/e1c3b345e60492ef057af3578d0f7a8a6ad99541/src/FixedSizeArrays.jl#L56-L82

[vtjnash]

It also turned out to be occasionally useful just inside base, if you grep a bit (it is roughly equivalent to reduce(checked_mul, dims), but with special handling of zeros and typemax to correct some edge cases of those), so I think it is reasonable to export. I know FSA decided to implement it differently for themselves, though that will be necessary for packages that want to support older versions anyways.

[nsajko]

IMO it would be a bit weird to export Core.checked_dims, given that it's not generic and it's implementation is necessarily clumsy due to bootstrapping concerns.

FTR the function in FSA also supports just Int, but it would be easy to make it generic.

[LilithHafner]

Any examples of such implementations that folks had to reimplement?

https://github.com/LilithHafner/MallocArrays.jl/blob/0b6dbdc012e1058b2b64d0f94863eff4120def85/src/MallocArrays.jl#L10-L27

---

Yes, we'd need to make it a bit more generic and move it to Base. Possibly by defining Base.checked_dims(x...) = Core.checked_dims(Int.(x)...). Supporting Int only is okay/good b.c. that's what dims are represented/retrieved as, but auto-conversion is also reasonable.

[nsajko]

Supporting Int only is okay/good b.c. that's what dims are represented

This is not a rule, e.g.:

julia> map(BigInt, 1:7)
1:7

julia> typeof(size(ans))
Tuple{BigInt}