Some utilities in std.array: remove_duplicates and filter_map

[Danl2620]

Is your feature request related to a problem? Please describe.

I need to take values that fit in to nice high-level schema and re-organize them for ad-hoc "back ends" like Ansible inventories. Some more array processing functions would help in this transformation.

Describe the solution you'd like

• A function std.array.remove_duplicates. This could be written to take an equality function and likely have the type forall a (a -> a -> Bool) -> Array a-> Array a. It could also use built-in equality testing for a simpler definition like forall a Array a -> Array a.
• A function std.array.filter_map. I'm not sure if the type of function is express-able -- i.e. it would need a type like forall a b. (a -> b) -> Array a -> Array b but b could be a Bool (with the only acceptable value being false).

Describe alternatives you've considered

I have workarounds that bloat the code a bit.

Additional context

Hopefully these are fairly understood functions -- they definitely exist in other functional languages. The type system of Nickel is a bit different than what I'm used to so perhaps the types would need adjustment.

[jneem]

I think the signature for std.array.filter_map should be something like forall a b. (a -> [| 'Some b, 'None |]) -> Array a -> Array b. (Or Just / Nothing)

[thufschmitt]

Quick passing thoughts on these (mostly from a performance pov):

• std.array.remove_duplicates is a dangerous thing because it's quadratic in complexity. Maybe a version that sorts would be better (I know that Nixpkgs' performance suffered quite a bit from an abusive usage of an equivalent function until it got replaced by an $n*log(n)$ sorting equivalent).
• std.array.filter_map is kinda expressible already through std.array.flat_map. Not exactly the same semantics, but at least it means that it can probably be defined in client code at very low cost:

  let
    filter_map : forall a b. (a -> [| 'Some b, 'None |]) -> Array a -> Array b = fun f =>
      std.array.flat_map (fun x => f x |> match { 'Some y => [y], 'None => [] })
  in
  filter_map (fun x => if x % 2 == 0 then 'Some (x / 2) else 'None) [ 2, 2, 3, 5 ,3, 4, 9 ,0 ]
  

[yannham]

std.array.remove_duplicates is a dangerous thing because it's quadratic in complexity. Maybe a version that sorts would be better (I know that Nixpkgs' performance suffered quite a bit from an abusive usage of an equivalent function until it got replaced by an sorting equivalent).

By the way, does Nix do anything fancy for sorting? I have a vague idea of sorting having a really bad performance in early Nickel. You'd like to have it as a fast primitive (using, say, Rust's sorting algorithm), but preserving laziness plus the fact that you take a user-defined order makes it quite hard to do so.

std.array.filter_map is kinda expressible already through std.array.flat_map. Not exactly the same semantics, but at least it means that it can probably be defined in client code at very low cost:

While I agree, IMHO it's still a common and useful addition that could be reused across Nickel projects. In the end many std.array functions are just a few lines of glue code on top of core combinators like maps and folds.

[thufschmitt]

does Nix do anything fancy for sorting? I have a vague idea of sorting having a really bad performance in early Nickel. You'd like to have it as a fast primitive (using, say, Rust's sorting algorithm), but preserving laziness plus the fact that you take a user-defined order makes it quite hard to do so.

It mostly falls-back to the STL sort algorithm. The main fancy thing is an optimisation to bypass the language layer if the ordering is the built-in one: https://github.com/tweag/nix/blob/84aa8e9f196e50549aa5cdef76f61f442ce40b82/src/libexpr/primops.cc#L3305-L3310

(note also that the built-in comparison operator of Nix is more permissive than the Nickel one as it accepts anything except records and functions)

[yannham]

It mostly falls-back to the STL sort algorithm. The main fancy thing is an optimisation to bypass the language layer if the ordering is the built-in one: https://github.com/tweag/nix/blob/84aa8e9f196e50549aa5cdef76f61f442ce40b82/src/libexpr/primops.cc#L3305-L3310

I see. Using the standard sort algorithm is easier in Nix because the interpreter just use mutually recursive functions to force values. In Nickel, as we have a stack machine (in practice a loop and a heap-allocated stack), its more annoying. But I suppose we could make it work by nesting call to vm.eval or something.