Provide isless

[haampie]

For tuples there is

(1, 2) < (1, 3) #true

Would it make sense to have it for SVectors as well?

SVector(1, 2) < SVector(1, 3) # ERROR: MethodError: no method matching isless(::SVector{2,Int64}, ::SVector{2,Int64})

My use case is to sort a long Vector of small SVector elements.

[mschauer]

I prefer not, the lexical order is not a natural order for vectors.

[haampie]

Do you have an example in mind where it doesn't make sense?

[mschauer]

In ℝⁿ the expression x ≤ y almost exclusively refers to the partial order (I have a bias here though).

[andyferris]

We strictly follow Base in these matters - doing so aids in code compatibility between static and dynamically sized arrays in the long run.

Your best chance would be to open an issue on the main Julia webpage. While lexical ordering is not a natural operation on vector spaces, you might argue it makes just as much sense on a “list” as a “string” (in some languages they are the same (strings being just a list of characters), and Vector is our list type).

[haampie]

Actually we have lexicographical isless for vectors in Julia 0.7 now, so reopening this!

[andyferris]

Yep - thanks for checking that. As an implementation detail you might consider overloading cmp.

I note that the existing Base definition could be made more efficient. If we compare the length of the vectors before the loop (find the minimum), use indices instead of zip, and, use @inbounds, we probably get something closer to what we could squeeze out of fully unrolled code via a generated function anyway (either way involves branching).

[mschauer]

I was also wrong on how it should be: isless tests "whether x is less than y, according to a canonical total order" and the lexical order could well be considered the canonical total order on vectors.

[andyferris]

Yes, that’s right. isless doesn’t necessarily provide semantic meaning but rather some way of implementing sorted search and so-on. It could be a Hilbert curve or some-such and it would still be ok, but lexical order works for a variety of element types, is fast, and matches String (which in many ways has some similarity with Vector{Char}).

[c42f]

Well, this is no longer a bug (base implements a fallback), but the generated code for SVector is fairly horrible compared to Tuple.

Would be a good first issue for someone to implement this by just calling Tuple on the input SVectors and passing that to the comparison.