This reverts commit c59c6b1c57.
This was a bit too ambitious, because no warnings were previously
triggered when `string` was nested e.g. `attrsOf string`, `nullOr
string`, etc.
Support for nested type deprecation warnings was introduced in
4b54aedee5, but had to be reverted in
a36e6760e9 because it caused infinite
recursion for some users, and I couldn't remember that it was reverted.
The type has given a warning on use since [4 years ago](03392cd336b128a1639c648baf0f6c1a1271e0d2), I think it's safe to move the deprecation to the next stage: An error instead of a warning.
Before this commit, getExe assumes that if `meta.mainProgram` is unset,
it has a main program that's named after the package name.
While this is probable, it leads to a bad error when the assumption does
not hold. If the user called `getExe` themselves, they might narrow down
the location of the assumption quite easily, but if that's not the case,
they'll have to go down the rabbit hole to figure out what went wrong.
For example, a NixOS module may use `lib.getExe` on a package-typed option
which is then used in the system configuration. This then typically leads
to a failure *after* deployment, which is bad, and it's quite likely that
the user will debug the package output contents before digging through the
NixOS module, which is bad.
Furthermore the `getExe` call is rather inconspicuous as it does not
contain something like "/bin/foo", which is bad.
Also modules can be hard to read for a newbie, which is bad.
All of this can be avoided by requiring `meta.mainProgram`.
Many packages already have the attribute, and I would expect 80% of
`getExe` usages to be covered by 20% of packages, because plenty of
packages aren't used with `getExe` anyway.
Finally we could make an effort to set `mainProgram` semi-automatically
using `nix-index`.
Deeply-curried functions are pretty error-prone in untyped languages
like Nix. This is a particularly bad case because
`top-level/splice.nix` *also* declares a makeScopeWithSplicing, but
it takes *two fewer arguments*.
Let's switch to attrset-passing form, to provide some minimal level
of sanity-checking.
mkOption does not require a `type` argument and does not set the
resulting attribute if it is not given. Consequently, we need to be
prepared to merge options that have no type information.
Add support for Nvidia's Bluefield 2 plattform as a compilation
target. There exists a version with and without crypto support,
while the crypto supported version is the most common one.
Support for the non-crypto version can be easily added in the future,
if needed.
For a datasheet of the hardware, see:
https://www.nvidia.com/content/dam/en-zz/Solutions/Data-Center/documents/datasheet-nvidia-bluefield-2-dpu.pdf
Signed-off-by: Markus Theil <theil.markus@gmail.com>
This is a non-trivial refactor that slightly changes the semantics
of the internal definition lists.
Whereas previously only individual list items would trigger the exception,
now the error is promoted to the whole list.
This is mostly ok, because we compute the value, it is wrong to ignore a definition.
However, we don't always compute the value. For instance `readOnly`
only needs to count definitions. That won't be possible anymore when
the error is raised for one of the items. As a consequence, an error
will be raised for the errant definition instead of the number of
definitions.
