Error with at-pattern and optional params

[lucabrunox]

$ echo 'let f = o@{x ? "x"}: o.x; in f {}'|nix-instantiate -
error: attribute `x' missing, at (string):1:20

It only happens with optional parameters.

[shlevy]

@-patterns bind the actual argument passed, not the optional arguments. This also applies to ...: If you pass an argument to a function that has a ... param, any extra attributes will still be bound by the @-pattern

[shlevy]

Think of foo@{ params... }: body as sugar for foo: ({ params...}: body) foo

[lucabrunox]

Didn't understand much the semantics, but it's very inconvenient. Now I have to write:

{ opt1 ? "def1", opt2 ? "def2" }: 
let opts = { inherit opt1 opt2; };
...

[edolstra]

You can just write x instead of o.x, right?

Anyway, closing since it's intended behaviour (and the alternative is less efficient to implement, because it would require constructing a new attrset).

[lucabrunox]

@edolstra no because I want to use o as a set for union ecc., as explained in the above comment, otherwise I wouldn't use the @-pattern no? That's the purpose of @-pattern, to reuse the argument set as set.
Also I don't think choosing performance over semantics is a sane approach, sorry.

[edolstra]

Given that plenty of people complain about the performance and memory consumption of the evaluator, it's a pretty relevant argument.

[lucabrunox]

If you want this semantics as you said being intended, I really dislike this choice because it makes @-pattern useless with optional arguments, and deserves a big warning in the docs. Currenly it's written as follows:

  args@{ x, y, z, ... }: z + y + x + args.a
Here args is bound to the **entire argument**

Which for me means that I can use o.x even if it's optional, unless I'm missing some other relevant line about argument sets.

I don't know the implementation details of nix and I don't know why another set is needed to be created, but if another set can even remotely (theoretically) be avoided, please keep this issue open.

[shlevy]

@lethalman Are you prepared to go through all existing uses of this and ensure they don't depend on the current semantics?

[lucabrunox]

@shlevy yes. I'd like to highlight this line: https://github.com/NixOS/nixpkgs/blob/master/pkgs/desktops/kde-4.12/kde-package/default.nix#L57

Apparently, who wrote that line had my same wrong thought, that meta, name and cmakeFlags are part of a.

[edolstra]

Well, the performance aspect is moot anyway, since the current behaviour is intended. The @ in Nix is just like the @ in Haskell: it's an alias for a pattern-matched argument. It's not intended to construct a new value.