Add a Nix flake #111
Add a Nix flake #111
Conversation
sellout
force-pushed
the
updated-flake
branch
from
09ccc3c
to
240baf5
Compare
|
I recommend adding https://garnix.io/ to the repo, which will do Nix flake builds/checks without any additional configuration. It can’t yet replace GH Actions, since the flake doesn’t yet run tests, but it’ll make sure the flake stays in working order, and is a cache for anyone using the flake. |
There was a problem hiding this comment.
My Nix-expert coworker @kenranunderscore reviewed, and it looks good!
Many thanks!
Maybe should squash before merging?
flake.nix
Outdated
| "concat-hardware" | ||
| # TODO: Tests pass with `cabal new-test`, but not with `cabal test`, and | ||
| # I’m guessing the latter is what `callCabal2nix` uses. | ||
| "concat-plugin" |
There was a problem hiding this comment.
I would really like to remove this line – i.e., get the gold-tests working via Nix.
I’m ok with merging before that happens, but this prevents us from replacing GH Actions with Garnix.
There was a problem hiding this comment.
@sellout What exactly fails with the gold tests?
There was a problem hiding this comment.
One thing I can think of is that misc-examples and misc-trace aren't proper test suites.
Just turned them into executables in a0f2aff
There was a problem hiding this comment.
Ah, great – the same should be done for graphics-trace in concat-graphics.
Then we should also expose the executables in the flake 😄
There was a problem hiding this comment.
Your executable change fixes some of what I was running into, but there are still a few issues. Here’s what happens on GHC 9.0.2 (the default):
$ nix develop github:con-kitty/concat/updated-flake#ghc902_concat-plugin
⋮
❄️$ cabal test gold-tests
⋮
[3 of 4] Compiling BasicTests ( test/BasicTests.hs, dist/build/gold-tests/gold-tests-tmp/BasicTests.o )
[4 of 4] Compiling Main ( test/GoldTests.hs, dist/build/gold-tests/gold-tests-tmp/Main.o )
Linking dist/build/gold-tests/gold-tests ...
/nix/store/…-clang-wrapper-11.1.0/bin/ld: line 256: 51086 Segmentation fault: 11 /nix/store/…-cctools-binutils-darwin-973.0.1/bin/ld ${extraBefore+"${extraBefore[@]}"} ${params+"${params[@]}"} ${extraAfter+"${extraAfter[@]}"}
clang-11: error: linker command failed with exit code 139 (use -v to see invocation)
`cc' failed in phase `Linker'. (Exit code: 139)
❄️$ cabal v2-test gold-tests
⋮
5 out of 83 tests failed (0.14s)
Test suite gold-tests: FAIL
Test suite logged to:
/…/build/aarch64-osx/ghc-9.0.2/concat-plugin-0.3.0.0/t/gold-tests/test/concat-plugin-0.3.0.0-gold-tests.log
0 of 1 test suites (0 of 1 test cases) passed.
Error: cabal: Tests failed for test:gold-tests from concat-plugin-0.3.0.0.
❄️$
cabal test falls during linking, but cabal v2-test builds (then fails some test cases). Once the linking step succeeds, cabal test gold-tests will thereafter have the same result as cabal v2-test gold-tests.
When replicating this issue, I noticed a “bug”. Everything gets built with the correct GHC, but if you enter a dev shell (as above) the GHC that’s exposed in the shell is always the default (currently GHC 9.0.2). So,
$ nix develop github:con-kitty/concat/updated-flake#ghc8107_concat-plugin
⋮
❄️$ cabal test gold-tests
will build and run tests with GHC 9.0.2, not 8.10.7. I tried a few things locally to fix that, but didn’t get anything that works.
However nix build github:con-kitty/concat/updated-flake#ghc8107_concat-plugin builds just fine, no linking error, unlike 9.0.2 – and the tests pass on 8.10.7.
There was a problem hiding this comment.
Ok, I pushed a commit that changed the other things that should be executable and then enables all checks (which will still result in some builds failing).
However, I don’t think any of the failing builds are depended on by Categorifier 😁
There was a problem hiding this comment.
So, to summarize the current state:
concat-hardwarepanics during tests (on all GHCs)concat-pluginfails tests only on GHC 9.0.2- all of the other failures are probably fine, but are blocked by the above failures.
There was a problem hiding this comment.
I just enabled darwin builds, too, so we’ll see how that goes.
There was a problem hiding this comment.
The "5 out of 83 tests failed" are probably false negatives: You can see that the github ci on master also fails similarly, as the output is equivalent but similar. I'm guessing a slight skew in base versions, leading to different inlining results. If that's the case here (of which I'm pretty sure), we can just re-generate the gold output and be done with it.
sellout
force-pushed
the
updated-flake
branch
3 times, most recently
from
545e93c
to
b94caec
Compare
They really are – they don’t run tests, and are hard to use without human interaction.
Mostly because Haddock thinks commented-out guards are docs.
- add a Nix formatter; - add a default package & devShell; - restructure the overlays; - move the inputs to the end; - don’t repeat the GHC version list; - move lib into its own file (and add more functions); - add nixConfig; - provide a “Haskell overlay”; - disable GHC 9.4 and HEAD builds, as they fail; - enable all packages now that they work with newer GHC (except for concat-hardware); - enable `checkPhase` for all Cabal packages; - replace string passing with a package set; - extract "ghc902" to `defaultCompiler`; - add a Garnix config to enable darwin builds; and - expose executables from Cabal packages.
|
Ok, this is mergable now, IMO. The failures that are happening are known. concat-plugin on 9.0.2 sounds fixable after this, and then we can figure out concat-hardware. Everything else is good. |
|
Thanks - will fix 9.0.2 build subsequently. |
No description provided.