Add writeStringReferencesToFile, a builder which extracts a string's
references to derivations and paths and writes them to a text file,
removing the input string itself from the dependency graph. This is
useful when you want to make a derivation depend on the string's
references, but not its content (to avoid unnecessary rebuilds, for
example).
This fixes#126344, specifically with the goal of enabling overriding the
checkPhase argument. See `design notes` at the end for details.
This allows among other things, enabling bash extension for the `checkPhase`.
Previously using such bash extensions was prohibited by the `writeShellScript`
code because there was no way to enable the extension in the checker.
As an example:
```nix
(writeShellScript "foo" ''
shopt -s extglob
echo @(foo|bar)
'').overrideAttrs (old: {
checkPhase = ''
# use subshell to preserve outer environment
(
export BASHOPTS
shopt -s extglob
${old.checkPhase}
)
'';
})
```
This commit also adds tests for this feature to `pkgs/tests/default.nix`,
under `trivial-overriding`. The test code is located at
`pkgs/build-support/trivial-builders/test-overriding.nix`.
Design notes:
-------------
Per discussion with @sternenseemann, the original approach of just wrapping
`writeTextFile` in `makeOverridable` had the issue that combined with `callPackage`
in the following form, would shadow the `.override` attribute of the `writeTextFile`:
```nix
with import <nixpkgs>;
callPackage ({writeShellScript}: writeShellScript "foo" "echo foo")
```
A better approach can be seen in this commit, where `checkPhase` is moved
from an argument of `writeTextFile`, which is substituted into `buildCommand`,
into an `mkDerivation` argument, which is substituted from the environment
and `eval`-ed. (see the source)
This way we can simple use `.overideAttrs` as usual, and this also makes
`checkPhase` a bit more conformant to `mkDerivation` naming, with respect to
phases generally being overridable attrs.
Co-authored-by: sterni <sternenseemann@systemli.org>
Co-authored-by: Naïm Favier <n@monade.li>
> There is an issue in the test added by #123111.
> [it] introduces a dependency on the contents of nixpkgs,
> making every change evaluate with a different hash.
