The code of `lib.closePropagation` was internally using a
recursion on the dependencies and returns all the derivation directly or
indirectly referenced by buildInputs.
`lib.closeProgation` is implemented in pure nix and uses an unique
function for list which is quadratic and does "true" equality, which
needs deep set comparison.
Instead, we use the `builtins.genericClosure` which is implemented as a
builtin and uses a more efficient sorting feature.
Note that `genericClosure` needs a `key` to discriminate the values, we
used the `outPath` which is unique and orderable.
On benchmarks, it performs up to 15x time faster on a benchmark related
to haskellPackages.ghcWithPackages.
Personally, I think that warnings such as
warning: The option `services.redis.enable' defined in `/home/ma27/Projects/nixpkgs/test.nix@node-vm' has been renamed to `services.redis.servers..enable'.
are fairly confusing because of the `..` and it's more correct to
actually quote that. With this change the warning now looks like this:
warning: The option `services.redis.enable' defined in `/home/ma27/Projects/nixpkgs/test.nix@node-vm' has been renamed to `services.redis.servers."".enable'.
While implementing that I realized that you'd have
a similar problem whenever you use attribute-names that aren't
identifiers, e.g.
services.nginx.virtualHosts."example.org".locations."/".invalid = 23;
now results in the following error:
error: The option `interactive.nodes.vm.services.nginx.virtualHosts."example.org".locations."/".invalid' does not exist. Definition values:
- In `/home/ma27/Projects/nixpkgs/test.nix@node-vm': 23
Of course there are some corner-cases where this won't work: when
generating the manual, you display submodules like this:
services.nginx.virtualHosts.<name>
Since `<name>` isn't a value, but an indicator for a submodule, it must
not be quoted. This also applies to the following identifiers:
* `*` for `listOf submodule`
* `<function body>` for `functionTo`
This might not be correct if you actually have a submodule with an
attribute name called `<name>`, but I think it's an improvement over the
current situation and for this you'd probably need to make even more
complex changes to the module system.
The motivation is to have a single identifier for that. Useful for the
next commit where I'll try to escape option-parts correctly (options can
be any kind of strings, but unless these are Nix identifiers, they must
be quoted).
Since `<function body>` (or `<name>`/`*`) are special identifiers in
error messages and the manual, we need a unique way to mark an option
part as function call because these are not to be quoted.
Move already implemented functionality to the upper level so
it could be used in a more generic way.
Signed-off-by: Ivan Nikolaenko <ivan.nikolaenko@unikie.com>
This brings two benefits:
1. The complete list of collisions is printed in the whenever any colliding
attribute is accessed.
2. The sets are intersected using a C++ primitive, which runs in O(n) time
(intersecting pre-sorted lists) with small constants rather than interpreted
Nix code.
Thanks to @toonn for prompting this improvement.
