sternenseemann 571a07d774 doc/haskell: don't use lib.recursiveUpdate in overlays

`lib.recursiveUpdate` indiscriminately recurses into all attribute sets,
also into derivations. This means that it is possible that evaluating a
derivation in the final haskell package set can cause something in
`prev.haskell` to be forced by `recursiveUpdate`, potentially causing an
evaluation error that should not happen.

It can be fixed using a well-crafted predicate for
`lib.recursiveUpdateUntil`, but most robust is just explicitly writing
out the desired merging manually.

2024-01-31 23:49:32 +01:00

54 KiB

Raw Blame History

Haskell

The Haskell infrastructure in Nixpkgs has two main purposes: The primary purpose is to provide a Haskell compiler and build tools as well as infrastructure for packaging Haskell-based packages.

The secondary purpose is to provide support for Haskell development environments including prebuilt Haskell libraries. However, in this area sacrifices have been made due to self-imposed restrictions in Nixpkgs, to lessen the maintenance effort and to improve performance. (More details in the subsection Limitations.)

Available packages

The compiler and most build tools are exposed at the top level:

ghc is the default version of GHC
Language specific tools: cabal-install, stack, hpack, …

Many “normal” user facing packages written in Haskell, like niv or cachix, are also exposed at the top level, and there is nothing Haskell specific to installing and using them.

All of these packages are originally defined in the haskellPackages package set and are re-exposed with a reduced dependency closure for convenience. (see justStaticExecutables or separateBinOutput below)

The haskellPackages set includes at least one version of every package from Hackage as well as some manually injected packages. This amounts to a lot of packages, so it is hidden from nix-env -qa by default for performance reasons. You can still list all packages in the set like this:

$ nix-env -f '<nixpkgs>' -qaP -A haskellPackages
haskellPackages.a50                                                         a50-0.5
haskellPackages.AAI                                                         AAI-0.2.0.1
haskellPackages.aasam                                                       aasam-0.2.0.0
haskellPackages.abacate                                                     abacate-0.0.0.0
haskellPackages.abc-puzzle                                                  abc-puzzle-0.2.1
…

Also, the haskellPackages set is included on search.nixos.org.

The attribute names in haskellPackages always correspond with their name on Hackage. Since Hackage allows names that are not valid Nix without escaping, you need to take care when handling attribute names like 3dmodels.

For packages that are part of Stackage (a curated set of known to be compatible packages), we use the version prescribed by a Stackage snapshot (usually the current LTS one) as the default version. For all other packages we use the latest version from Hackage (the repository of basically all open source Haskell packages). See [below](#haskell-available- versions) for a few more details on this.

Roughly half of the 16K packages contained in haskellPackages don’t actually build and are marked as broken semi-automatically. Most of those packages are deprecated or unmaintained, but sometimes packages that should build, do not build. Very often fixing them is not a lot of work.

haskellPackages is built with our default compiler, but we also provide other releases of GHC and package sets built with them. You can list all available compilers like this:

$ nix-env -f '<nixpkgs>' -qaP -A haskell.compiler
haskell.compiler.ghc810                  ghc-8.10.7
haskell.compiler.ghc90                   ghc-9.0.2
haskell.compiler.ghc925                  ghc-9.2.5
haskell.compiler.ghc926                  ghc-9.2.6
haskell.compiler.ghc927                  ghc-9.2.7
haskell.compiler.ghc92                   ghc-9.2.8
haskell.compiler.ghc945                  ghc-9.4.5
haskell.compiler.ghc946                  ghc-9.4.6
haskell.compiler.ghc947                  ghc-9.4.7
haskell.compiler.ghc94                   ghc-9.4.8
haskell.compiler.ghc963                  ghc-9.6.3
haskell.compiler.ghc96                   ghc-9.6.4
haskell.compiler.ghc98                   ghc-9.8.1
haskell.compiler.ghcHEAD                 ghc-9.9.20231121
haskell.compiler.ghc8107Binary           ghc-binary-8.10.7
haskell.compiler.ghc865Binary            ghc-binary-8.6.5
haskell.compiler.ghc924Binary            ghc-binary-9.2.4
haskell.compiler.integer-simple.ghc8107  ghc-integer-simple-8.10.7
haskell.compiler.integer-simple.ghc810   ghc-integer-simple-8.10.7
haskell.compiler.native-bignum.ghc90     ghc-native-bignum-9.0.2
haskell.compiler.native-bignum.ghc902    ghc-native-bignum-9.0.2
haskell.compiler.native-bignum.ghc925    ghc-native-bignum-9.2.5
haskell.compiler.native-bignum.ghc926    ghc-native-bignum-9.2.6
haskell.compiler.native-bignum.ghc927    ghc-native-bignum-9.2.7
haskell.compiler.native-bignum.ghc92     ghc-native-bignum-9.2.8
haskell.compiler.native-bignum.ghc928    ghc-native-bignum-9.2.8
haskell.compiler.native-bignum.ghc945    ghc-native-bignum-9.4.5
haskell.compiler.native-bignum.ghc946    ghc-native-bignum-9.4.6
haskell.compiler.native-bignum.ghc947    ghc-native-bignum-9.4.7
haskell.compiler.native-bignum.ghc94     ghc-native-bignum-9.4.8
haskell.compiler.native-bignum.ghc948    ghc-native-bignum-9.4.8
haskell.compiler.native-bignum.ghc963    ghc-native-bignum-9.6.3
haskell.compiler.native-bignum.ghc96     ghc-native-bignum-9.6.4
haskell.compiler.native-bignum.ghc964    ghc-native-bignum-9.6.4
haskell.compiler.native-bignum.ghc98     ghc-native-bignum-9.8.1
haskell.compiler.native-bignum.ghc981    ghc-native-bignum-9.8.1
haskell.compiler.native-bignum.ghcHEAD   ghc-native-bignum-9.9.20231121
haskell.compiler.ghcjs                   ghcjs-8.10.7

Each of those compiler versions has a corresponding attribute set built using it. However, the non-standard package sets are not tested regularly and, as a result, contain fewer working packages. The corresponding package set for GHC 9.4.5 is haskell.packages.ghc945. In fact haskellPackages is just an alias for haskell.packages.ghc927:

$ nix-env -f '<nixpkgs>' -qaP -A haskell.packages.ghc927
haskell.packages.ghc927.a50                                                         a50-0.5
haskell.packages.ghc927.AAI                                                         AAI-0.2.0.1
haskell.packages.ghc927.aasam                                                       aasam-0.2.0.0
haskell.packages.ghc927.abacate                                                     abacate-0.0.0.0
haskell.packages.ghc927.abc-puzzle                                                  abc-puzzle-0.2.1
…

Every package set also re-exposes the GHC used to build its packages as haskell.packages.*.ghc.

Available package versions

We aim for a “blessed” package set which only contains one version of each package, like Stackage, which is a curated set of known to be compatible packages. We use the version information from Stackage snapshots and extend it with more packages. Normally in Nixpkgs the number of building Haskell packages is roughly two to three times the size of Stackage. For choosing the version to use for a certain package we use the following rules:

By default, for haskellPackages.foo is the newest version of the package foo found on Hackage, which is the central registry of all open source Haskell packages. Nixpkgs contains a reference to a pinned Hackage snapshot, thus we use the state of Hackage as of the last time we updated this pin.
If the Stackage snapshot that we use (usually the newest LTS snapshot) contains a package, we use instead the version in the Stackage snapshot as default version for that package.
For some packages, which are not on Stackage, we have if necessary manual overrides to set the default version to a version older than the newest on Hackage.
For all packages, for which the newest Hackage version is not the default version, there will also be a haskellPackages.foo_x_y_z package with the newest version. The x_y_z part encodes the version with dots replaced by underscores. When the newest version changes by a new release to Hackage the old package will disappear under that name and be replaced by a newer one under the name with the new version. The package name including the version will also disappear when the default version e.g. from Stackage catches up with the newest version from Hackage. E.g. if haskellPackages.foo gets updated from 1.0.0 to 1.1.0 the package haskellPackages.foo_1_1_0 becomes obsolete and gets dropped.
For some packages, we also manually add other haskellPackages.foo_x_y_z versions, if they are required for a certain build.

Relying on haskellPackages.foo_x_y_z attributes in derivations outside nixpkgs is discouraged because they may change or disappear with every package set update.

All haskell.packages.* package sets use the same package descriptions and the same sets of versions by default. There are however GHC version specific override .nix files to loosen this a bit.

Dependency resolution

Normally when you build Haskell packages with cabal-install, cabal-install does dependency resolution. It will look at all Haskell package versions known on Hackage and tries to pick for every (transitive) dependency of your build exactly one version. Those versions need to satisfy all the version constraints given in the .cabal file of your package and all its dependencies.

The Haskell builder in nixpkgs does no such thing. It will take as input packages with names off the desired dependencies and just check whether they fulfill the version bounds and fail if they don’t (by default, see jailbreak to circumvent this).

The haskellPackages.callPackage function does the package resolution. It will, e.g., use haskellPackages.aesonwhich has the default version as described above for a package input of name aeson. (More general: <packages>.callPackage f will call f with named inputs provided from the package set <packages>.) While this is the default behavior, it is possible to override the dependencies for a specific package, see override and overrideScope.

Limitations

Our main objective with haskellPackages is to package Haskell software in nixpkgs. This entails some limitations, partially due to self-imposed restrictions of nixpkgs, partially in the name of maintainability:

Only the packages built with the default compiler see extensive testing of the whole package set. For other GHC versions only a few essential packages are tested and cached.
As described above we only build one version of most packages.

The experience using an older or newer packaged compiler or using different versions may be worse, because builds will not be cached on cache.nixos.org or may fail.

Thus, to get the best experience, make sure that your project can be compiled using the default compiler of nixpkgs and recent versions of its dependencies.

A result of this setup is, that getting a valid build plan for a given package can sometimes be quite painful, and in fact this is where most of the maintenance work for haskellPackages is required. Besides that, it is not possible to get the dependencies of a legacy project from nixpkgs or to use a specific stack solver for compiling a project.

Even though we couldn’t use them directly in nixpkgs, it would be desirable to have tooling to generate working Nix package sets from build plans generated by cabal-install or a specific Stackage snapshot via import-from-derivation. Sadly we currently don’t have tooling for this. For this you might be interested in the alternative haskell.nix framework, which, be warned, is completely incompatible with packages from haskellPackages.

`haskellPackages.mkDerivation`

Every haskell package set has its own haskell-aware mkDerivation which is used to build its packages. Generally you won't have to interact with this builder since cabal2nix can generate packages using it for an arbitrary cabal package definition. Still it is useful to know the parameters it takes when you need to override a generated Nix expression.

haskellPackages.mkDerivation is a wrapper around stdenv.mkDerivation which re-defines the default phases to be haskell aware and handles dependency specification, test suites, benchmarks etc. by compiling and invoking the package's Setup.hs. It does not use or invoke the cabal-install binary, but uses the underlying Cabal library instead.

General arguments

pname: Package name, assumed to be the same as on Hackage (if applicable)
version: Packaged version, assumed to be the same as on Hackage (if applicable)
src: Source of the package. If omitted, fetch package corresponding to pname and version from Hackage.
sha256: Hash to use for the default case of src.
revision: Revision number of the updated cabal file to fetch from Hackage. If null (which is the default value), the one included in src is used.
editedCabalFile: sha256 hash of the cabal file identified by revision or null.
configureFlags: Extra flags passed when executing the configure command of Setup.hs.
buildFlags: Extra flags passed when executing the build command of Setup.hs.
haddockFlags: Extra flags passed to Setup.hs haddock when building the documentation.
doCheck: Whether to execute the package's test suite if it has one. Defaults to true unless cross-compiling.
doBenchmark: Whether to execute the package's benchmark if it has one. Defaults to false.
doHoogle: Whether to generate an index file for hoogle as part of haddockPhase by passing the --hoogle option. Defaults to true.
doHaddockQuickjump: Whether to generate an index for interactive navigation of the HTML documentation. Defaults to true if supported.
doInstallIntermediates: Whether to install intermediate build products (files written to dist/build by GHC during the build process). With enableSeparateIntermediatesOutput, these files are instead installed to a separate intermediates output. The output can then be passed into a future build of the same package with the previousIntermediates argument to support incremental builds. See “Incremental builds” for more information. Defaults to false.
enableLibraryProfiling: Whether to enable profiling for libraries contained in the package. Enabled by default if supported.
enableExecutableProfiling: Whether to enable profiling for executables contained in the package. Disabled by default.
profilingDetail: Profiling detail level to set. Defaults to exported-functions.
enableSharedExecutables: Whether to link executables dynamically. By default, executables are linked statically.
enableSharedLibraries: Whether to build shared Haskell libraries. This is enabled by default unless we are using pkgsStatic or shared libraries have been disabled in GHC.
enableStaticLibraries: Whether to build static libraries. Enabled by default if supported.
enableDeadCodeElimination: Whether to enable linker based dead code elimination in GHC. Enabled by default if supported.
enableHsc2hsViaAsm: Whether to pass --via-asm to hsc2hs. Enabled by default only on Windows.
hyperlinkSource: Whether to render the source as well as part of the haddock documentation by passing the --hyperlinked-source flag. Defaults to true.
isExecutable: Whether the package contains an executable.
isLibrary: Whether the package contains a library.
jailbreak: Whether to execute jailbreak-cabal before configurePhase to lift any version constraints in the cabal file. Note that this can't lift version bounds if they are conditional, i.e. if a dependency is hidden behind a flag.
enableParallelBuilding: Whether to use the -j flag to make GHC/Cabal start multiple jobs in parallel.
maxBuildCores: Upper limit of jobs to use in parallel for compilation regardless of $NIX_BUILD_CORES. Defaults to 16 as Haskell compilation with GHC currently sees a performance regression if too many parallel jobs are used.
doCoverage: Whether to generate and install files needed for HPC. Defaults to false.
doHaddock: Whether to build (HTML) documentation using haddock. Defaults to true if supported.
testTarget: Name of the test suite to build and run. If unset, all test suites will be executed.
preCompileBuildDriver: Shell code to run before compiling Setup.hs.
postCompileBuildDriver: Shell code to run after compiling Setup.hs.
preHaddock: Shell code to run before building documentation using haddock.
postHaddock: Shell code to run after building documentation using haddock.
coreSetup: Whether to only allow core libraries to be used while building Setup.hs. Defaults to false.
useCpphs: Whether to enable the cpphs preprocessor. Defaults to false.
enableSeparateBinOutput: Whether to install executables to a separate bin output. Defaults to false.
enableSeparateDataOutput: Whether to install data files shipped with the package to a separate data output. Defaults to false.
enableSeparateDocOutput: Whether to install documentation to a separate doc output. Is automatically enabled if doHaddock is true.
enableSeparateIntermediatesOutput: When doInstallIntermediates is true, whether to install intermediate build products to a separate intermediates output. See “Incremental builds” for more information. Defaults to false.
allowInconsistentDependencies: If enabled, allow multiple versions of the same Haskell package in the dependency tree at configure time. Often in such a situation compilation would later fail because of type mismatches. Defaults to false.
enableLibraryForGhci: Build and install a special object file for GHCi. This improves performance when loading the library in the REPL, but requires extra build time and disk space. Defaults to false.
previousIntermediates: If non-null, intermediate build artifacts are copied from this input to dist/build before performing compiling. See “Incremental builds” for more information. Defaults to null.
buildTarget: Name of the executable or library to build and install. If unset, all available targets are built and installed.

Specifying dependencies

Since haskellPackages.mkDerivation is intended to be generated from cabal files, it reflects cabal's way of specifying dependencies. For one, dependencies are grouped by what part of the package they belong to. This helps to reduce the dependency closure of a derivation, for example benchmark dependencies are not included if doBenchmark == false.

setup*Depends: dependencies necessary to compile Setup.hs
library*Depends: dependencies of a library contained in the package
executable*Depends: dependencies of an executable contained in the package
test*Depends: dependencies of a test suite contained in the package
benchmark*Depends: dependencies of a benchmark contained in the package

The other categorization relates to the way the package depends on the dependency:

*ToolDepends: Tools we need to run as part of the build process. They are added to the derivation's nativeBuildInputs.
*HaskellDepends: Haskell libraries the package depends on. They are added to propagatedBuildInputs.
*SystemDepends: Non-Haskell libraries the package depends on. They are added to buildInputs
*PkgconfigDepends: *SystemDepends which are discovered using pkg-config. They are added to buildInputs and it is additionally ensured that pkg-config is available at build time.
*FrameworkDepends: Apple SDK Framework which the package depends on when compiling it on Darwin.

Using these two distinctions, you should be able to categorize most of the dependency specifications that are available: benchmarkFrameworkDepends, benchmarkHaskellDepends, benchmarkPkgconfigDepends, benchmarkSystemDepends, benchmarkToolDepends, executableFrameworkDepends, executableHaskellDepends, executablePkgconfigDepends, executableSystemDepends, executableToolDepends, libraryFrameworkDepends, libraryHaskellDepends, libraryPkgconfigDepends, librarySystemDepends, libraryToolDepends, setupHaskellDepends, testFrameworkDepends, testHaskellDepends, testPkgconfigDepends, testSystemDepends and testToolDepends.

That only leaves the following extra ways for specifying dependencies:

buildDepends: Allows specifying Haskell dependencies which are added to propagatedBuildInputs unconditionally.
buildTools: Like *ToolDepends, but are added to nativeBuildInputs unconditionally.
extraLibraries: Like *SystemDepends, but are added to buildInputs unconditionally.
pkg-configDepends: Like *PkgconfigDepends, but are added to buildInputs unconditionally.
testDepends: Deprecated, use either testHaskellDepends or testSystemDepends.
benchmarkDepends: Deprecated, use either benchmarkHaskellDepends or benchmarkSystemDepends.

The dependency specification methods in this list which are unconditional are especially useful when writing overrides when you want to make sure that they are definitely included. However, it is recommended to use the more accurate ones listed above when possible.

Meta attributes

haskellPackages.mkDerivation accepts the following attributes as direct arguments which are transparently set in meta of the resulting derivation. See the Meta-attributes section for their documentation.

These attributes are populated with a default value if omitted:
- homepage: defaults to the Hackage page for pname.
- platforms: defaults to lib.platforms.all (since GHC can cross-compile)
These attributes are only set if given:
- description
- license
- changelog
- maintainers
- broken
- hydraPlatforms

Incremental builds

haskellPackages.mkDerivation supports incremental builds for GHC 9.4 and newer with the doInstallIntermediates, enableSeparateIntermediatesOutput, and previousIntermediates arguments.

The basic idea is to first perform a full build of the package in question, save its intermediate build products for later, and then copy those build products into the build directory of an incremental build performed later. Then, GHC will use those build artifacts to avoid recompiling unchanged modules.

For more detail on how to store and use incremental build products, see Gabriella Gonzalez’ blog post “Nixpkgs support for incremental Haskell builds”. motivation behind this feature.

An incremental build for the turtle package can be performed like so:

let
  pkgs = import <nixpkgs> {};
  inherit (pkgs) haskell;
  inherit (haskell.lib.compose) overrideCabal;

  # Incremental builds work with GHC >=9.4.
  turtle = haskell.packages.ghc944.turtle;

  # This will do a full build of `turtle`, while writing the intermediate build products
  # (compiled modules, etc.) to the `intermediates` output.
  turtle-full-build-with-incremental-output = overrideCabal (drv: {
    doInstallIntermediates = true;
    enableSeparateIntermediatesOutput = true;
  }) turtle;

  # This will do an incremental build of `turtle` by copying the previously
  # compiled modules and intermediate build products into the source tree
  # before running the build.
  #
  # GHC will then naturally pick up and reuse these products, making this build
  # complete much more quickly than the previous one.
  turtle-incremental-build = overrideCabal (drv: {
    previousIntermediates = turtle-full-build-with-incremental-output.intermediates;
  }) turtle;
in
  turtle-incremental-build

Development environments

In addition to building and installing Haskell software, nixpkgs can also provide development environments for Haskell projects. This has the obvious advantage that you benefit from cache.nixos.org and no longer need to compile all project dependencies yourself. While it is often very useful, this is not the primary use case of our package set. Have a look at the section available package versions to learn which versions of packages we provide and the section limitations, to judge whether a haskellPackages based development environment for your project is feasible.

By default, every derivation built using haskellPackages.mkDerivation exposes an environment suitable for building it interactively as the env attribute. For example, if you have a local checkout of random, you can enter a development environment for it like this (if the dependencies in the development and packaged version match):

$ cd ~/src/random
$ nix-shell -A haskellPackages.random.env '<nixpkgs>'
[nix-shell:~/src/random]$ ghc-pkg list
/nix/store/a8hhl54xlzfizrhcf03c1l3f6l9l8qwv-ghc-9.2.4-with-packages/lib/ghc-9.2.4/package.conf.d
    Cabal-3.6.3.0
    array-0.5.4.0
    base-4.16.3.0
    binary-0.8.9.0
    …
    ghc-9.2.4
    …

As you can see, the environment contains a GHC which is set up so it finds all dependencies of random. Note that this environment does not mirror the environment used to build the package, but is intended as a convenient tool for development and simple debugging. env relies on the ghcWithPackages wrapper which automatically injects a pre-populated package-db into every GHC invocation. In contrast, using nix-shell -A haskellPackages.random will not result in an environment in which the dependencies are in GHCs package database. Instead, the Haskell builder will pass in all dependencies explicitly via configure flags.

env mirrors the normal derivation environment in one aspect: It does not include familiar development tools like cabal-install, since we rely on plain Setup.hs to build all packages. However, cabal-install will work as expected if in PATH (e.g. when installed globally and using a nix-shell without --pure). A declarative and pure way of adding arbitrary development tools is provided via shellFor.

When using cabal-install for dependency resolution you need to be a bit careful to achieve build purity. cabal-install will find and use all dependencies installed from the packages env via Nix, but it will also consult Hackage to potentially download and compile dependencies if it can’t find a valid build plan locally. To prevent this you can either never run cabal update, remove the cabal database from your ~/.cabal folder or run cabal with --offline. Note though, that for some usecases cabal2nix needs the local Hackage db.

Often you won't work on a package that is already part of haskellPackages or Hackage, so we first need to write a Nix expression to obtain the development environment from. Luckily, we can generate one very easily from an already existing cabal file using cabal2nix:

$ ls
my-project.cabal src …
$ cabal2nix ./. > my-project.nix

The generated Nix expression evaluates to a function ready to be callPackage-ed. For now, we can add a minimal default.nix which does just that:

# Retrieve nixpkgs impurely from NIX_PATH for now, you can pin it instead, of course.
{ pkgs ? import <nixpkgs> {} }:

# use the nixpkgs default haskell package set
pkgs.haskellPackages.callPackage ./my-project.nix { }

Using nix-build default.nix we can now build our project, but we can also enter a shell with all the package's dependencies available using nix-shell -A env default.nix. If you have cabal-install installed globally, it'll work inside the shell as expected.

shellFor

Having to install tools globally is obviously not great, especially if you want to provide a batteries-included shell.nix with your project. Luckily there's a proper tool for making development environments out of packages' build environments: shellFor, a function exposed by every haskell package set. It takes the following arguments and returns a derivation which is suitable as a development environment inside nix-shell:

packages: This argument is used to select the packages for which to build the development environment. This should be a function which takes a haskell package set and returns a list of packages. shellFor will pass the used package set to this function and include all dependencies of the returned package in the build environment. This means you can reuse Nix expressions of packages included in nixpkgs, but also use local Nix expressions like this: hpkgs: [ (hpkgs.callPackage ./my-project.nix { }) ].
nativeBuildInputs: Expects a list of derivations to add as build tools to the build environment. This is the place to add packages like cabal-install, doctest or hlint. Defaults to [].
buildInputs: Expects a list of derivations to add as library dependencies, like openssl. This is rarely necessary as the haskell package expressions usually track system dependencies as well. Defaults to []. (see also derivation dependencies)
withHoogle: If this is true, hoogle will be added to nativeBuildInputs. Additionally, its database will be populated with all included dependencies, so you'll be able search through the documentation of your dependencies. Defaults to false.
genericBuilderArgsModifier: This argument accepts a function allowing you to modify the arguments passed to mkDerivation in order to create the development environment. For example, args: { doCheck = false; } would cause the environment to not include any test dependencies. Defaults to lib.id.
doBenchmark: This is a shortcut for enabling doBenchmark via genericBuilderArgsModifier. Setting it to true will cause the development environment to include all benchmark dependencies which would be excluded by default. Defaults to false.

One neat property of shellFor is that it allows you to work on multiple packages using the same environment in conjunction with cabal.project files. Say our example above depends on distribution-nixpkgs and we have a project file set up for both, we can add the following shell.nix expression:

{ pkgs ? import <nixpkgs> {} }:

pkgs.haskellPackages.shellFor {
  packages = hpkgs: [
    # reuse the nixpkgs for this package
    hpkgs.distribution-nixpkgs
    # call our generated Nix expression manually
    (hpkgs.callPackage ./my-project/my-project.nix { })
  ];

  # development tools we use
  nativeBuildInputs = [
    pkgs.cabal-install
    pkgs.haskellPackages.doctest
    pkgs.cabal2nix
  ];

  # Extra arguments are added to mkDerivation's arguments as-is.
  # Since it adds all passed arguments to the shell environment,
  # we can use this to set the environment variable the `Paths_`
  # module of distribution-nixpkgs uses to search for bundled
  # files.
  # See also: https://cabal.readthedocs.io/en/latest/cabal-package.html#accessing-data-files-from-package-code
  distribution_nixpkgs_datadir = toString ./distribution-nixpkgs;
}

haskell-language-server

To use HLS in short: Install pkgs.haskell-language-server e.g. in nativeBuildInputs in shellFor and use the haskell-language-server-wrapper command to run it. See the HLS user guide on how to configure your text editor to use HLS and how to test your setup.

HLS needs to be compiled with the GHC version of the project you use it on.

pkgs.haskell-language-server provides haskell-language-server-wrapper, haskell-language-server and haskell-language-server-x.x.x binaries, where x.x.x is the GHC version for which it is compiled. By default, it only includes binaries for the current GHC version, to reduce closure size. The closure size is large, because HLS needs to be dynamically linked to work reliably. You can override the list of supported GHC versions with e.g.

pkgs.haskell-language-server.override { supportedGhcVersions = [ "90" "94" ]; }

Where all strings version are allowed such that haskell.packages.ghc${version} is an existing package set.

When you run haskell-language-server-wrapper it will detect the GHC version used by the project you are working on (by asking e.g. cabal or stack) and pick the appropriate versioned binary from your path.

Be careful when installing HLS globally and using a pinned nixpkgs for a Haskell project in a nix-shell. If the nixpkgs versions deviate to much (e.g., use different glibc versions) the haskell-language-server-?.?.? executable will try to detect these situations and refuse to start. It is recommended to obtain HLS via nix-shell from the nixpkgs version pinned in there instead.

The top level pkgs.haskell-language-server attribute is just a convenience wrapper to make it possible to install HLS for multiple GHC versions at the same time. If you know, that you only use one GHC version, e.g., in a project specific nix-shell you can use pkgs.haskellPackages.haskell-language-server or pkgs.haskell.packages.*.haskell-language-server from the package set you use.

If you use nix-shell for your development environments remember to start your editor in that environment. You may want to use something like direnv and/or an editor plugin to achieve this.

Overriding Haskell packages

Overriding a single package

Like many language specific subsystems in nixpkgs, the Haskell infrastructure also has its own quirks when it comes to overriding. Overriding of the inputs to a package at least follows the standard procedure. For example, imagine you need to build nix-tree with a more recent version of brick than the default one provided by haskellPackages:

haskellPackages.nix-tree.override {
  brick = haskellPackages.brick_0_67;
}

The custom interface comes into play when you want to override the arguments passed to haskellPackages.mkDerivation. For this, the function overrideCabal from haskell.lib.compose is used. E.g., if you want to install a man page that is distributed with the package, you can do something like this:

haskell.lib.compose.overrideCabal (drv: {
  postInstall = ''
    ${drv.postInstall or ""}
    install -Dm644 man/pnbackup.1 -t $out/share/man/man1
  '';
}) haskellPackages.pnbackup

overrideCabal takes two arguments:

A function which receives all arguments passed to haskellPackages.mkDerivation before and returns a set of arguments to replace (or add) with a new value.
The Haskell derivation to override.

The arguments are ordered so that you can easily create helper functions by making use of currying:

let
  installManPage = haskell.lib.compose.overrideCabal (drv: {
    postInstall = ''
      ${drv.postInstall or ""}
      install -Dm644 man/${drv.pname}.1 -t "$out/share/man/man1"
    '';
  });
in

installManPage haskellPackages.pnbackup

In fact, haskell.lib.compose already provides lots of useful helpers for common tasks, detailed in the next section. They are also structured in such a way that they can be combined using lib.pipe:

lib.pipe my-haskell-package [
  # lift version bounds on dependencies
  haskell.lib.compose.doJailbreak
  # disable building the haddock documentation
  haskell.lib.compose.dontHaddock
  # pass extra package flag to Cabal's configure step
  (haskell.lib.compose.enableCabalFlag "myflag")
]

`haskell.lib.compose`

The base interface for all overriding is the following function:

overrideCabal f drv: Takes the arguments passed to obtain drv to f and uses the resulting attribute set to update the argument set. Then a recomputed version of drv using the new argument set is returned.

All other helper functions are implemented in terms of overrideCabal and make common overrides shorter and more complicate ones trivial. The simple overrides which only change a single argument are only described very briefly in the following overview. Refer to the documentation of haskellPackages.mkDerivation for a more detailed description of the effects of the respective arguments.

Packaging Helpers

overrideSrc { src, version } drv: Replace the source used for building drv with the path or derivation given as src. The version attribute is optional. Prefer this function over overriding src via overrideCabal, since it also automatically takes care of removing any Hackage revisions.

justStaticExecutables drv: Only build and install the executables produced by drv, removing everything that may refer to other Haskell packages' store paths (like libraries and documentation). This dramatically reduces the closure size of the resulting derivation. Note that the executables are only statically linked against their Haskell dependencies, but will still link dynamically against libc, GMP and other system library dependencies. If dependencies use their Cabal-generated Paths_* module, this may not work as well if GHC's dead code elimination is unable to remove the references to the dependency's store path that module contains.
enableSeparateBinOutput drv: Install executables produced by drv to a separate bin output. This has a similar effect as justStaticExecutables, but preserves the libraries and documentation in the out output alongside the bin output with a much smaller closure size.
markBroken drv: Sets the broken flag to true for drv.
markUnbroken drv, unmarkBroken drv: Set the broken flag to false for drv.
doDistribute drv: Updates hydraPlatforms so that Hydra will build drv. This is sometimes necessary when working with versioned packages in haskellPackages which are not built by default.
dontDistribute drv: Sets hydraPlatforms to [], causing Hydra to skip this package altogether. Useful if it fails to evaluate cleanly and is causing noise in the evaluation errors tab on Hydra.

Development Helpers

sdistTarball drv: Create a source distribution tarball like those found on Hackage instead of building the package drv.
documentationTarball drv: Create a documentation tarball suitable for uploading to Hackage instead of building the package drv.
buildFromSdist drv: Uses sdistTarball drv as the source to compile drv. This helps to catch packaging bugs when building from a local directory, e.g. when required files are missing from extra-source-files.
failOnAllWarnings drv: Enables all warnings GHC supports and makes it fail the build if any of them are emitted.

enableDWARFDebugging drv: Compiles the package with additional debug symbols enabled, useful for debugging with e.g. gdb.
doStrip drv: Sets doStrip to true for drv.
dontStrip drv: Sets doStrip to false for drv.

Trivial Helpers

doJailbreak drv: Sets the jailbreak argument to true for drv.
dontJailbreak drv: Sets the jailbreak argument to false for drv.
doHaddock drv: Sets doHaddock to true for drv.
dontHaddock drv: Sets doHaddock to false for drv. Useful if the build of a package is failing because of e.g. a syntax error in the Haddock documentation.
doHyperlinkSource drv: Sets hyperlinkSource to true for drv.
dontHyperlinkSource drv: Sets hyperlinkSource to false for drv.
doCheck drv: Sets doCheck to true for drv.
dontCheck drv: Sets doCheck to false for drv. Useful if a package has a broken, flaky or otherwise problematic test suite breaking the build.

appendConfigureFlags list drv: Adds the strings in list to the configureFlags argument for drv.
enableCabalFlag flag drv: Makes sure that the Cabal flag flag is enabled in Cabal's configure step.
disableCabalFlag flag drv: Makes sure that the Cabal flag flag is disabled in Cabal's configure step.
appendBuildFlags list drv: Adds the strings in list to the buildFlags argument for drv.

appendPatches list drv: Adds the list of derivations or paths to the patches argument for drv.

addBuildTools list drv: Adds the list of derivations to the buildTools argument for drv.
addExtraLibraries list drv: Adds the list of derivations to the extraLibraries argument for drv.
addBuildDepends list drv: Adds the list of derivations to the buildDepends argument for drv.
addTestToolDepends list drv: Adds the list of derivations to the testToolDepends argument for drv.
addPkgconfigDepends list drv: Adds the list of derivations to the pkg-configDepends argument for drv.
addSetupDepends list drv: Adds the list of derivations to the setupHaskellDepends argument for drv.
doBenchmark drv: Set doBenchmark to true for drv. Useful if your development environment is missing the dependencies necessary for compiling the benchmark component.
dontBenchmark drv: Set doBenchmark to false for drv.
setBuildTargets drv list: Sets the buildTarget argument for drv so that the targets specified in list are built.
doCoverage drv: Sets the doCoverage argument to true for drv.
dontCoverage drv: Sets the doCoverage argument to false for drv.
enableExecutableProfiling drv: Sets the enableExecutableProfiling argument to true for drv.
disableExecutableProfiling drv: Sets the enableExecutableProfiling argument to false for drv.
enableLibraryProfiling drv: Sets the enableLibraryProfiling argument to true for drv.
disableLibraryProfiling drv: Sets the enableLibraryProfiling argument to false for drv.

Library functions in the Haskell package sets

Some library functions depend on packages from the Haskell package sets. Thus they are exposed from those instead of from haskell.lib.compose which can only access what is passed directly to it. When using the functions below, make sure that you are obtaining them from the same package set (haskellPackages, haskell.packages.ghc944 etc.) as the packages you are working with or – even better – from the self/final fix point of your overlay to haskellPackages.

Note: Some functions like shellFor that are not intended for overriding per se, are omitted in this section.

cabalSdist { src, name ? ... }: Generates the Cabal sdist tarball for src, suitable for uploading to Hackage. Contrary to haskell.lib.compose.sdistTarball, it uses cabal-install over Setup.hs, so it is usually faster: No build dependencies need to be downloaded, and we can skip compiling Setup.hs.
buildFromCabalSdist drv: Build drv, but run its src attribute through cabalSdist first. Useful for catching files necessary for compilation that are missing from the sdist.
generateOptparseApplicativeCompletions list drv: Generate and install shell completion files for the installed executables whose names are given via list. The executables need to be using optparse-applicative for this to work. Note that this feature is automatically disabled when cross-compiling, since it requires executing the binaries in question.

F.A.Q.

Why is topic X not covered in this section? Why is section Y missing?

We have been working on moving the nixpkgs Haskell documentation back into the nixpkgs manual. Since this process has not been completed yet, you may find some topics missing here covered in the old haskell4nix docs.

If you feel any important topic is not documented at all, feel free to comment on the issue linked above.

How to enable or disable profiling builds globally?

By default, Nixpkgs builds a profiling version of each Haskell library. The exception to this rule are some platforms where it is disabled due to concerns over output size. You may want to…

…enable profiling globally so that you can build a project you are working on with profiling ability giving you insight in the time spent across your code and code you depend on using GHC's profiling feature.
…disable profiling (globally) to reduce the time spent building the profiling versions of libraries which a significant amount of build time is spent on (although they are not as expensive as the “normal” build of a Haskell library).

::: {.note} The method described below affects the build of all libraries in the respective Haskell package set as well as GHC. If your choices differ from Nixpkgs' default for your (host) platform, you will lose the ability to substitute from the official binary cache.

If you are concerned about build times and thus want to disable profiling, it probably makes sense to use haskell.lib.compose.disableLibraryProfiling (see ) on the packages you are building locally while continuing to substitute their dependencies and GHC. :::

Since we need to change the profiling settings for the desired Haskell package set and GHC (as the core libraries like base, filepath etc. are bundled with GHC), it is recommended to use overlays for Nixpkgs to change them. Since the interrelated parts, i.e. the package set and GHC, are connected via the Nixpkgs fixpoint, we need to modify them both in a way that preserves their connection (or else we'd have to wire it up again manually). This is achieved by changing GHC and the package set in separate overlays to prevent the package set from pulling in GHC from prev.

The result is two overlays like the ones shown below. Adjustable parts are annotated with comments, as are any optional or alternative ways to achieve the desired profiling settings without causing too many rebuilds.

let
  # Name of the compiler and package set you want to change. If you are using
  # the default package set `haskellPackages`, you need to look up what version
  # of GHC it currently uses (note that this is subject to change).
  ghcName = "ghc92";
  # Desired new setting
  enableProfiling = true;
in

[
  # The first overlay modifies the GHC derivation so that it does or does not
  # build profiling versions of the core libraries bundled with it. It is
  # recommended to only use such an overlay if you are enabling profiling on a
  # platform that doesn't by default, because compiling GHC from scratch is
  # quite expensive.
  (final: prev:
  let
    inherit (final) lib;
  in

  {
    haskell = prev.haskell // {
      compiler = prev.haskell.compiler // {
        ${ghcName} = prev.haskell.compiler.${ghcName}.override {
          # Unfortunately, the GHC setting is named differently for historical reasons
          enableProfiledLibs = enableProfiling;
        };
      };
    };
  })

  (final: prev:
  let
    inherit (final) lib;
    haskellLib = final.haskell.lib.compose;
  in

  {
    haskell = prev.haskell // {
      packages = prev.haskell.packages // {
        ${ghcName} = prev.haskell.packages.${ghcName}.override {
          overrides = hfinal: hprev: {
            mkDerivation = args: hprev.mkDerivation (args // {
              # Since we are forcing our ideas upon mkDerivation, this change will
              # affect every package in the package set.
              enableLibraryProfiling = enableProfiling;

              # To actually use profiling on an executable, executable profiling
              # needs to be enabled for the executable you want to profile. You
              # can either do this globally or…
              enableExecutableProfiling = enableProfiling;
            });

            # …only for the package that contains an executable you want to profile.
            # That saves on unnecessary rebuilds for packages that you only depend
            # on for their library, but also contain executables (e.g. pandoc).
            my-executable = haskellLib.enableExecutableProfiling hprev.my-executable;

            # If you are disabling profiling to save on build time, but want to
            # retain the ability to substitute from the binary cache. Drop the
            # override for mkDerivation above and instead have an override like
            # this for the specific packages you are building locally and want
            # to make cheaper to build.
            my-library = haskellLib.disableLibraryProfiling hprev.my-library;
          };
        };
      };
    };
  })
]

54 KiB Raw Blame History Unescape Escape