Buck2 build system for Haskell projects

Short summary: this is a cabal external command cabal buck2 that allows you to use Buck2 as the build system for your Cabal project.

Quick start

First download a buck2 binary, unpack it and put it on your PATH.

Then

cabal install cabal-buck2

Then in the root of your project or package:

git clone https://github.com/simonmar/haskell-buck2.git buck2 -b <version>
cabal buck2 --enable-tests

Where <version> is the version of cabal-buck, e.g. 0.2.0.0. Checking out the branch ensures that you get a compatible copy of the Buck2 support code. cabal buck2 checks that buck2/ is for its release (the CABAL_BUCK2_VERSION in buck2/version.bzl), and so does buck2 when it reads the files that cabal buck2 generated, and the error says which branch to switch to.

Then you can use buck2 as the build tool, e.g.

buck2 build //...

To build all the components, or

buck2 test //...

To run your tests.

Note that you need to re-run cabal buck2 --enable-tests if you modify the cabal.project or any of the .cabal files.

Not all Cabal features are supported - see limitations.

Why?

(skip this section if you know why you want buck2)

Why might you want to use buck2 as the build system compared with just using cabal? Well, first off let me be clear that you still need Cabal, because the Buck2 support doesn't know how to solve package dependencies or build them. So the workflow consists of first running cabal buck2 to solve and build the dependencies, but once you've done that you can switch to buck2 for building. The idea is that buck2 is a more pleasant experience because:

• It's faster than Cabal, particularly for rebuilds.

• It supports different build modes out of the box: the default is to build in dev mode (unoptimised with dynamic linking) but adding -m opt gives you optimisation and static linking. Note that Cabal doesn't have a purely dynamic build mode: it always uses -dynamic-too for libraries, which has a significant built-time performance cost.

• The Buck2 build is extensible. If you have anything that needs to be generated as part of your build, or any non-standard tooling, then hooking that up using Buck2 is far easier than Cabal. Furthermore Buck2 knows how to rebuild things correctly when either the build system or the code generator components change.

• It works a lot better than Cabal when you have non-Haskell code (e.g. C/C++ or Rust) in your project, because

  • Buck2 understands dependencies between C/C++ source files and header files (Cabal doesn't: issue #4306), so when you modify a C/C++ header the correct things are rebuilt.
  • Buck2 builds C/C++ files in parallel, while Cabal doesn't (issue #7127)

• You can use remote execution and caching. There is built-in support for using a build cache, see Sharing build results (a build cache) below.

Finally, if you have an existing codebase using Buck2 then this is the basis of something that could "buckify" Cabal packages to integrate into your build system. It needs a bit of work to be suitable for that use case, though: cabal buck2 builds all the external dependencies and installs them in the Cabal store, whereas to integrate with an existing build system you would want to satisfy those external dependencies from the build system itself.

How complete is it?

I've used it to build a few large projects, in particular the Cabal project itself which consists of about 16 packages and a few hundred source files. It can also build Glean, which has some complex build requirements including custom codegen, FFI & hsc2hs.

There are a few limitations, however.

What cabal buck2 does

You can run cabal buck2 in a project or a single Cabal package. It does the following:

1. Solves the build-depends constraints of your package(s) and builds all the dependencies, much like cabal build all --only-dependencies would.

2. Generates some files, notably:

   • BUCK and BUCK.cabal.bzl in each package, these are the Buck2 build targets

   • cabal-buck2/autogen in each package, this is where we put the files that Cabal autogenerates, such as cabal_macros.h and Paths_<pkg>.hs.

   • cabal-buck2/packages.bzl in the root of the project, which says where each of your packages is. Nothing in a package's BUCK.cabal.bzl says where it is or where the packages it depends on are, so moving a package, or building it as part of a bigger project, only changes this file. (All the packages have to be in the one Buck2 cell, the root.)

   • third-party/haskell: tells Buck2 about all the prebuilt package dependencies, either in the Cabal store or in GHC's package DB (unless you build the dependencies from source, see below). In here we also record the GHC version you're using, and the paths to any tool dependencies.

A component that can't be built with buck2 (e.g. build-type: Custom) is an error: cabal buck2 lists them and aborts, unless you give the --keep-going flag in which case it will translate everything it can and warn about the missing components.

Buck2 quick start

To build your code:

buck2 build //...

The //... is Buck2's syntax for "all targets recursively below the current directory". You can also build specific target(s), for example buck2 build cabal-install:cabal would build the cabal target in the cabal-install package. For more details see Target Pattern in the Buck2 docs.

Next you can run your tests:

buck2 test //...

Customising the build

cabal buck2 will generate all the BUCK files if they don't exist, but you can also write your own if you want (cabal buck2 won't overwrite them).

The BUCK file usually goes in the same directory as your source files. The targets that cabal buck2 produces go in the BUCK.cabal.bzl file, and are generated by a call to generated_targets() from the BUCK file. This call takes some arguments that you can use to override or transform the generated targets - take a look at the comments in the generated code to see how.

The overrides can also go in a cabal-buck2.bzl file in the root of the project, for all of its packages at once, which is useful when a package's directory isn't yours to add files to (a git submodule, say):

OVERRIDES = {
    # package name -> component name -> extra attributes
    "my-package": {
        "my-package": {"compiler_flags": ["-O2"]},
    },
}

Re-run cabal buck2 after adding or removing cabal-buck2.bzl.

You can also completely override the generated targets and write a BUCK file with your own rules, while still making use of the pre-built dependencies that cabal buck2 produces. For example, the BUCK file for a simple Haskell library might look something like

load("//buck2:haskell.bzl", "haskell_library")

haskell_library(
    name = "my-package",
    srcs = [
        "Some/Module.hs",
    ],
    packages = [
        "unordered-containers",
    ],
    visibility = ["PUBLIC"],
)

and the BUCK file for a test might look like

load("//buck2:haskell.bzl", "haskell_test")

haskell_test(
    name = "my-test",
    srcs = {
        "Main.hs" : "my-test.hs",
    },
    deps = [
        "//:my-package",
    ],
    packages = [
        "test-framework",
        "test-framework-hunit",
        "HUnit",
    ],
)

You can find docs on how to write BUCK files in the Buck2 docs, e.g. haskell_library.

Build modes

The Buck2 build system has two build modes:

• dev: the default, builds everything with -O0 and dynamic linking. This is intended to give you the quickest edit-compile-test turnaround.
• opt: enable -O and link statically. This takes longer but the code runs faster.

To build with opt, use -m opt, e.g.

buck2 build my-package:my-program -m opt

There are other build options that can be selected in a similar way, such as -m prof to enable profiling. See constraints/BUCK for details.

Sharing build results (a build cache)

By default buck2 keeps nothing between runs of its daemon, nor between checkouts: after buck2 kill, or in a new worktree, everything is built again. A cache of build results fixes that. Buck2 can use any server that implements the Bazel remote execution API's action cache and CAS, only to look results up and store them: nothing is run remotely. For example bazel-remote:

bazel-remote --dir ~/.cache/buck2 --max_size 20 --grpc_address 127.0.0.1:9092 --http_address 127.0.0.1:8080
cabal buck2 --cache=grpc://127.0.0.1:9092

--cache adds a block to .buckconfig (between # >>> cabal buck2: cache and # <<< cabal buck2: cache <<<; anything else in the file is left alone), which later runs keep. cabal buck2 --no-cache removes it. Run buck2 kill after changing it: buck2 reads the cache's address when its daemon starts, and a daemon that was already running keeps the old settings.

Things to know:

• The server has to be running. When it isn't, buck2 retries connecting for about 45 seconds on each build before carrying on without the cache. Re-run cabal buck2 with --no-cache to stop using the cache.
• The key of a cached result includes the command line, the environment and the contents of the inputs. It includes the exact packages from the Cabal store (their unit ids), a fingerprint of the GHC installation (its version, platform, source commit and the interface hashes of its boot packages) and a fingerprint of the C toolchain (the versions of the C compiler, ld, the C library and libstdc++). It does not include other files that are found on the system, such as headers and libraries that are not part of those. That is fine on one machine; sharing a cache between machines with different system software is not safe yet.
• Compiling and linking C/C++ code is cached, pkg-config queries are not.
• buck2 only downloads what is needed. A result that is found in the cache is not downloaded until something needs its files: a local action that has it as an input, buck2 run or buck2 test, or it is what you asked to build. So most intermediate results (the compiled modules of a library that was cached as a whole, say) are never fetched. buck2 build -M none //... goes further and does not download what you asked for either, which is a fast way to find out whether everything is already in the cache: the summary line shows how many actions were cache hits.

Building the dependencies with buck2

By default the dependencies of your packages are built by cabal, into its store, and Buck2 uses them from there. With --source-deps Buck2 builds them too:

cabal buck2 --source-deps

The source of each dependency is unpacked under dist-newstyle/src and the package gets a BUCK and BUCK.cabal.bzl like those of your own packages. Only the packages that come with GHC are used from its package database; the build then doesn't depend on the Cabal store at all. Together with a build cache that means a dependency is built once, and then found in the cache by every other checkout, which is the job the store does for cabal.

Things to know:

• If any dependency package isn't supported by cabal buck2, --source-deps will fail. The most common cause of this is a package that uses build-type: Custom (see Custom build type). In case of failure you can either fall back to not using --source-deps or use --keep-going to continue without the unsupported dependencies.
• Running cabal buck2 again without --source-deps goes back to the store, and removes the packages that were unpacked for the previous run.
• The tools that dependencies need to preprocess sources (alex and happy) are built by buck2 too, when the project needs them.
• The build plan can only have one version of each package, because a package in a .cabal file is referred to by its name. cabal buck2 stops and lists the packages that need more than one.

Performance

I ran some experiments building the Cabal project itself - 16 packages and 641 source files (one package, hackage-security, is not part of the project but has to be built locally nonetheless because it depends on Cabal-syntax which is part of the project).

Buck2 shines when it comes to rebuilds: the dependency graph is cached in memory, and it knows when build steps can be omitted because the inputs haven't changed.

Caveats

• Results tend to be +/- a few seconds from run to run
• I didn't dig into the results in any detail
• It's just one set of data points. Different projects and different choices of edits could give different results. However, I did perform a similar experiment with the persistent project, and got similar results.

Raw results and details

Clean build

• Optimised:

  • Default Cabal build: 280s
    • cabal build all --enable-tests --enable-benchmarks -j
  • Buck2 build (opt mode, including cabal buck2): 259s
    • cabal buck2 --enable-tests --enable-benchmarks && buck2 build //... -m opt
    • Not much difference here, as we expect.

• Unoptimised / dynamic:

  • Cabal build with -O0 -dynamic: 136s
    • cabal build all --enable-tests --enable-benchmarks -j --disable-optimisation --enable-executable-dynamic
  • Buck2 build (dev mode, including cabal buck2): 78s
    • cabal buck2 --enable-tests --enable-benchmarks && buck2 build //... -m dev
    • Cabal is using -dynamic-too for libraries, while Buck2 is building everything purely dynamic.

Edit + rebuild

Next I made a single edit (added an extension to Language.Haskell.Extension) and rebuilt everything:

• Optimised:

  • Cabal: 197s
  • Buck2: 179s

• Unoptimised / dynamic:

  • Cabal: 85s
  • Buck2: 55s

Limitations

It's an external command, not builtin to cabal-install

This has some implications:

• cabal passes only the arguments after buck2, so global flags given before it (cabal --store-dir=... buck2) don't reach the tool. Use the environment (CABAL_DIR) instead.

• Nothing checks that your cabal binary matches the version of the cabal-install library that cabal-buck2 was built against. Try to make sure they match, or confusion will undoubtedly ensue.

Builds currently use --make

The current Buck2 prelude uses ghc --make to build each component (library, executable). Ideally we should expose the full per-module dependencies to Buck2 so that it can exploit parallelism across packages for faster builds/rebuilds. It's entirely possible to do this, indeed the functionality already exists in Tweag's Haskell/Buck2 integration.

Custom build type

Build-type Custom isn't supported, and can't be in general, so cabal buck2 will fail if any package requires it. Note that this also applies to dependencies when --source-deps is being used. As with other unsupported features, you can use --keep-going to just skip all the affected packages.

If a package's Setup.hs isn't needed by your build, or what it does can be done some other way, --enable-custom-setup builds it as if it had build-type: Simple (or Configure, if it has a configure script). The modules that its Setup.hs would generate (its autogen-modules) must then be given by the project, in the overrides of the component (see Customising the build), for example

OVERRIDES = {
    "my-package": {
        "my-package": {"srcs": {"My.Generated.Module": "//gen:my-module"}},
    },
}

Hooks build type

There is partial support for build-type: Hooks: only configure hooks are supported so far; files generated by pre-build rules are not.

The SetupHooks executable is built (by cabal) during cabal buck2, including the package's setup-depends. It has to be built with the same Cabal version as cabal-buck2, so cabal-buck2 injects a solver constraint to ensure that.

Template Haskell and prof

A module that defines a splice must live in a different haskell_library() from any module that uses it, when profiling (-m prof). If not, the build will likely complain about a link error or a missing object file at compile-time.

The situation with Template Haskell and profiling is complex, as is the reason for this limitation.

• Without -fexternal-interpreter: GHC loads object code at compile-time into its own process. Since GHC is itself a dynamically-linked non-profiled executable, the objects it loads must be shared, non-profiled, objects. So we have to build all the dependencies of the current packages as shared libraries. This is fine, except for the current package: GHC expects to find the .dyn_o objects for the current package in the current -odir. But Buck2 doesn't work this way: it builds the two instances of the package separately. It's not clear if this is easily fixable.

• With -fexternal-interpreter, we could load the profiled non-shared objects. However, this method uses the RTS runtime linker, which is known to have some limitations and can't load some objects, particularly on certain architectures. This is the main reason that GHC switched to dynamic linking. So we don't go this route.

No support for Cabal's foreign-library

Nothing fundamental blocking this, it's just a TODO.

Preprocessors like hspec-discover

The hspec-discover preprocessor is designed to be invoked by GHC via the -pgmF flag to specify a custom preprocessor. The problem is that hspec-discover searches the filesystem to find other source files; these other source files amount to implicit inputs to the compilation, but when using Buck2 all inputs must be explicit (this is so that compilation steps can be executed remotely).

To build an hspec-discover test with Buck2, you have to run the preprocessor using a genrule() that takes all the source files as an input. For example, if your test is in test/Spec.hs:

filegroup(
    name = "srcs",
    srcs = glob(["**/*.hs"])
)

genrule(
    name = 'spec-gen',
    cmd = "$(location third-party-haskell//:hspec-discover-exe) $(location :srcs)/test/Spec.hs test/Spec.hs ${OUT}",
    out = "test/Spec.hs"
)

haskell_test(
    name = 'spec',
    srcs = {
        'Main.hs': ':spec-gen',
        ...
    },
    ...
)

Acknowledgments

Most of the code and modifications to the standard Buck2 prelude were developed with the help of Claude Code using Claude Sonnet 5/5.5.

The Haskell support already in the Buck2 prelude was developed by Meta and is in production use internally for building Glean. This project just fixes a few things and adds some functionality needed to support building Cabal projects.

Related projects

Tweag also worked on a Haskell integration for Buck2. This project has no code in common with theirs, except for the shared upstream prelude code. Tweag's integration is more sophisticated and was aimed at using Buck2's improved scalability to build large Haskell projects.