/hermetic_cc_toolchain

Bazel C/C++ toolchain for cross-compiling C/C++ programs

Primary LanguageStarlarkMIT LicenseMIT

ci

Hermetic CC toolchain

This is a C/C++ toolchain that can (cross-)compile C/C++ programs on top of zig cc. It contains clang-17, musl, glibc 2.17-2.38, all in a ~40MB package. Read here about zig-cc; the rest of the README will present how to use this toolchain from Bazel.

Configuring toolchains in Bazel is complex and fraught with peril. We, the team behind hermetic_cc_toolchain, are still confused on how this all works, and often wonder why it works at all. That aside, we made our best effort to make hermetic_cc_toolchain usable for your C/C++/CGo projects, with as many guardrails can be installed.

While copy-pasting the code in your project, attempt to read and understand the text surrounding the code snippets. This will save you hours of head scratching.

Project Origin

This repository is cloned from and is based on Adam Bouhenguel's bazel-zig-cc, and was later developed at sr.ht/~motiejus/bazel-zig-cc. After a while this repository was moved to the Uber GitHub repository and renamed to hermetic_cc_toolchain.

Our special thanks to Adam for coming up with the idea - and creating the original version – of bazel-zig-cc and publishing it. His idea and work helped make the concept of using Zig with Bazel a reality; now we all can benefit from it.

Usage

Add this to your WORKSPACE:

load("@bazel_tools//tools/build_defs/repo:http.bzl", "http_archive")

HERMETIC_CC_TOOLCHAIN_VERSION = "v3.1.0"

http_archive(
    name = "hermetic_cc_toolchain",
    sha256 = "df091afc25d73b0948ed371d3d61beef29447f690508e02bc24e7001ccc12d38",
    urls = [
        "https://mirror.bazel.build/github.com/uber/hermetic_cc_toolchain/releases/download/{0}/hermetic_cc_toolchain-{0}.tar.gz".format(HERMETIC_CC_TOOLCHAIN_VERSION),
        "https://github.com/uber/hermetic_cc_toolchain/releases/download/{0}/hermetic_cc_toolchain-{0}.tar.gz".format(HERMETIC_CC_TOOLCHAIN_VERSION),
    ],
)

load("@hermetic_cc_toolchain//toolchain:defs.bzl", zig_toolchains = "toolchains")

# Plain zig_toolchains() will pick reasonable defaults. See
# toolchain/defs.bzl:toolchains on how to change the Zig SDK version and
# download URL.
zig_toolchains()

And this to .bazelrc on a Unix-y systems:

common --enable_platform_specific_config
build:linux --sandbox_add_mount_pair=/tmp
build:macos --sandbox_add_mount_pair=/var/tmp
build:windows --sandbox_add_mount_pair=C:\Temp

The directories can be narrowed down to /tmp/zig-cache (Linux), /var/tmp/zig-cache (MacOS) and C:\Temp\zig-cache respectively if it can be ensured they will be created before the invocation of bazel build. See #83 for more context. If a different place is preferred for zig cache, set:

build --repo_env=HERMETIC_CC_TOOLCHAIN_CACHE_PREFIX=/path/to/cache
build --sandbox_add_mount_pair=/path/to/cache

The snippets above will download the zig toolchain and make the bazel toolchains available for registration and usage. If nothing else is done, this will work for some minimal use cases. The .bazelrc snippet instructs Bazel to use the registered "new kinds of toolchains". The next steps depend on how one wants to use hermetic_cc_toolchain. The descriptions below is a gentle introduction to C++ toolchains from "user's perspective" too.

See examples for some other recommended .bazelrc flags, as well as how to use hermetic_cc_toolchain with bzlmod.

Use case: manually build a single target with a specific zig cc toolchain

This option is least disruptive to the workflow compared to no hermetic C++ toolchain, and works best when trying out or getting started with hermetic_cc_toolchain for a subset of targets.

To request Bazel to use a specific toolchain (compatible with the specified platform) for build/tests/whatever on linux-amd64-musl, do:

bazel build \
    --platforms @zig_sdk//platform:linux_arm64 \
    --extra_toolchains @zig_sdk//toolchain:linux_arm64_musl \
    //test/go:go

There are a few things going on here, let's try to dissect them.

Option --platforms @zig_sdk//platform:linux_arm64

Specifies that the our target platform is linux_arm64, which resolves into:

$ bazel query --output=build @zig_sdk//platform:linux_arm64
platform(
  name = "linux_arm64",
  generator_name = "linux_arm64",
  generator_function = "declare_platforms",
  generator_location = "platform/BUILD:7:18",
  constraint_values = ["@platforms//os:linux", "@platforms//cpu:aarch64"],
)

constraint_values instructs Bazel to be looking for a toolchain that is compatible with (in Bazelspeak, target_compatible_with) all of the ["@platforms//os:linux", "@platforms//cpu:aarch64"].

Option --toolchains=@zig_sdk//toolchain:linux_arm64_musl

Inspect first (@platforms//cpu:aarch64 is an alias to @platforms//cpu:arm64):

$ bazel query --output=build @zig_sdk//toolchain:linux_arm64_musl
toolchain(
  name = "linux_arm64_musl",
  generator_name = "linux_arm64_musl",
  generator_function = "declare_toolchains",
  generator_location = "toolchain/BUILD:7:19",
  toolchain_type = "@bazel_tools//tools/cpp:toolchain_type",
  target_compatible_with = ["@platforms//os:linux", "@platforms//cpu:aarch64", "@zig_sdk//libc:unconstrained"],
  toolchain = "@zig_sdk//:aarch64-linux-musl_cc",
)

For a platform to pick up the right toolchain, the platform's constraint_values must be a subset1 of the toolchain's target_compatible_with. Since the platform is a subset (therefore, toolchain's @zig_sdk//libc:unconstrained does not matter), this toolchain is selected for this platform. As a result, --platforms @zig_sdk//platform:linux_amd64 causes Bazel to select a toolchain @zig_sdk//platform:linux_arm64_musl (because it satisfies all constraints), which will compile and link the C/C++ code with musl.

@zig_sdk//libc:unconstrained will become important later.

Same as above, less typing (with --config)

Specifying the platform and toolchain for every target may become burdensome, so they can be put used via --config. For example, append this to .bazelrc:

build:linux_arm64 --platforms @zig_sdk//platform:linux_arm64
build:linux_arm64 --extra_toolchains @zig_sdk//toolchain:linux_arm64_musl

And then building to linux-arm64-musl boils down to:

bazel build --config=linux_arm64_musl //test/go:go

Use case: always compile with zig cc

Instead of adding the toolchains to .bazelrc, they can be added unconditionally. Append this to WORKSPACE after zig_toolchains(...):

register_toolchains(
    "@zig_sdk//toolchain:linux_amd64_gnu.2.28",
    "@zig_sdk//toolchain:linux_arm64_gnu.2.28",
    "@zig_sdk//toolchain:darwin_amd64",
    "@zig_sdk//toolchain:darwin_arm64",
    "@zig_sdk//toolchain:windows_amd64",
    "@zig_sdk//toolchain:windows_arm64",
    "@zig_sdk//toolchain:wasip1_wasm",
)

Append this to .bazelrc:

build --action_env BAZEL_DO_NOT_DETECT_CPP_TOOLCHAIN=1

From Bazel's perspective, this is almost equivalent to always specifying --extra_toolchains on every bazel <...> command-line invocation. It also means there is no way to disable the toolchain with the command line. This is useful if you find hermetic_cc_toolchain useful enough to compile for all of your targets and tools.

With BAZEL_DO_NOT_DETECT_CPP_TOOLCHAIN=1 Bazel stops detecting the default host toolchain. Configuring toolchains is complicated enough, and the auto-detection (read: fallback to non-hermetic toolchain) is a footgun best avoided. This option is not documented in bazel, so may break. If you intend to use the hermetic toolchain exclusively, it won't hurt.

Use case: zig-cc for targets for multiple libc variants

When some targets need to be build with different libcs (either different versions of glibc or musl), use a linux toolchain from @zig_sdk//libc_aware/toolchains:<...>. The toolchain will only be selected when building for a specific libc. For example, in WORKSPACE:

register_toolchains(
    "@zig_sdk//libc_aware/toolchain:linux_amd64_gnu.2.19",
    "@zig_sdk//libc_aware/toolchain:linux_arm64_gnu.2.28",
    "@zig_sdk//libc_aware/toolchain:x86_64-linux-musl",
)

What does @zig_sdk//libc_aware/toolchain:linux_amd64_gnu.2.19 mean?

$ bazel query --output=build @zig_sdk//libc_aware/toolchain:linux_amd64_gnu.2.19 |& grep target
  target_compatible_with = ["@platforms//os:linux", "@platforms//cpu:x86_64", "@zig_sdk//libc:gnu.2.19"],

To see how this relates to the platform:

$ bazel query --output=build @zig_sdk//libc_aware/platform:linux_amd64_gnu.2.19 |& grep constraint
  constraint_values = ["@platforms//os:linux", "@platforms//cpu:x86_64", "@zig_sdk//libc:gnu.2.19"],

In this case, the platform's constraint_values and toolchain's target_compatible_with are identical, causing Bazel to select the right toolchain for the requested platform. With these toolchains registered, one can build a project for a specific libc-aware platform; it will select the appropriate toolchain:

$ bazel run --platforms @zig_sdk//libc_aware/platform:linux_amd64_gnu.2.19 //test/c:which_libc
glibc_2.19
$ bazel run --platforms @zig_sdk//libc_aware/platform:linux_amd64_gnu.2.28 //test/c:which_libc
glibc_2.28
$ bazel run --platforms @zig_sdk//libc_aware/platform:linux_amd64_musl //test/c:which_libc
non_glibc
$ bazel run --run_under=file --platforms @zig_sdk//libc_aware/platform:linux_arm64_gnu.2.28 //test/c:which_libc
which_libc: ELF 64-bit LSB executable, ARM aarch64, version 1 (SYSV), dynamically linked, interpreter /lib/ld-linux-aarch64.so.1, for GNU/Linux 2.0.0, stripped

To the list of libc aware toolchains and platforms:

$ bazel query @zig_sdk//libc_aware/toolchain/...
$ bazel query @zig_sdk//libc_aware/platform/...

Libc-aware toolchains are especially useful when relying on transitions, as transitioning extra_platforms will cause the host tools to be rebuilt with the specific libc version, which takes time; also the build host may not be able to run them if, say, target glibc version is newer than on the host. Some tests in this repository (under test/) are using transitions; you may check out how it's done.

The @zig_sdk//libc:variant constraint is necessary to select a matching toolchain. Remember: the toolchain's target_compatible_with must be equivalent or a superset of the platform's constraint_values. This is why both libc-aware platforms and libc-aware toolchains reside in their own namespace; if we try to mix non-libc-aware to libc-aware, confusion ensues.

To use the libc constraints in the project's platform definitions, add a @zig_sdk//libc:variant constraint to them. See the list of available values:

$ bazel query "attr(constraint_setting, @zig_sdk//libc:variant, @zig_sdk//...)"

@zig_sdk//libc:unconstrained is a special value that indicates that no value for the constraint is specified. The non libc aware linux toolchains are only compatible with this value to prevent accidental silent fallthrough to them. This is a guardrail.

Note: Naming

Both Go and Bazel naming schemes are accepted. For convenience with Go, the following Go-style toolchain aliases are created:

Bazel (zig) name Go name
x86_64 amd64
aarch64 arm64
wasm32 wasm
macos darwin
wasi wasip1

For example, the toolchain linux_amd64_gnu.2.28 is aliased to x86_64-linux-gnu.2.28. To find out which toolchains can be registered or used, run:

$ bazel query @zig_sdk//toolchain/...

Incompatibilities with clang and gcc

zig cc is almost a drop-in replacement for clang/gcc. This section lists some of the discovered differences and ways to live with them.

UBSAN and "SIGILL: Illegal Instruction"

zig cc differs from "mainstream" compilers by enabling UBSAN by default. Which means your program may compile successfully and crash with:

SIGILL: illegal instruction

This flag encourages program authors to fix the undefined behavior. There are many ways to find the undefined behavior.

Known Issues In hermetic_cc_toolchain

These are the things you may stumble into when using hermetic_cc_toolchain. We are unlikely to implement them any time soon, but patches implementing those will be accepted.

Zig cache location

Currently zig cache is stored in /var/tmp/zig-cache, so bazel clean --expunge will not clear the zig cache. Zig's cache should be stored somewhere in the project's path. It is not clear how to do it.

See #83 for more context.

OSX: sysroot

For non-trivial programs (and for all darwin/arm64 cgo programs) MacOS SDK may be necessary. Read Jakub's comment about it. Support for OSX sysroot is currently not implemented, but patches implementing it will be accepted, as long as the OSX sysroot must come through an http_archive.

In essence, OSX target support is not well tested with hermetic_cc_toolchain. Also see #10.

Bazel 6 or earlier

Add to .bazelrc:

build --incompatible_enable_cc_toolchain_resolution

Host Environments

This repository is used on the following (host) platforms:

  • linux_amd64, a.k.a. x86_64.
  • linux_arm64, a.k.a. AArch64.
  • darwin_amd64, the 64-bit post-PowerPC models.
  • darwin_arm64, the M1.
  • windows_amd64, a.k.a. x64.

The tests are running (CId) on linux-amd64.

Transient docker environment

A standalone Docker environment to play with hermetic_cc_toolchain:

$ docker run -e CC=/usr/bin/false -ti --rm -v "$PWD:/x" -w /x debian:bookworm-slim
# apt update && apt install --no-install-recommends -y shellcheck ca-certificates python3 git
# git config --global --add safe.directory /x
# tools/bazel test //...
# ./ci/lint
# ./ci/release
# ./ci/zig-wrapper

Communication

We maintain two channels for comms:

  • Github issues and pull requests.
  • Slack: #zig in bazelbuild.slack.com.

Previous Commuications

Previous communications were done in a mailing list; the past archive can be accessed like this:

git checkout v2.0.0-rc2 mailing-list-archive.mbox
mutt -R -f mailing-list-archive.mbox

Maintainers

Guidelines for maintainers2:

  • Communicate intent precisely.
  • Edge cases matter.
  • Favor reading code over writing code.
  • Only one obvious way to do things.
  • Runtime crashes are better than bugs.
  • Compile errors are better than runtime crashes.
  • Incremental improvements.
  • Avoid local maximums.
  • Reduce the amount one must remember.
  • Focus on code rather than style.
  • Resource allocation may fail; resource deallocation must succeed.
  • Memory is a resource.
  • Together we serve the users.

On a more practical note:

  • Maintainers can merge others' pull requests following their best judgement. They may or may not ask for feedback from other maintainers. Follow the Zen of Zig.
  • Currently releases are coordinated with Uber employees, because they can test the version-to-be-released their big repository. If you use hermetic_cc_toolchain in production and, more importantly, have a heterogeneous environment (different languages, RBE, different platforms), we encourage you to make yourself known. That way we can work together to validate it before cutting the release.

Footnotes

  1. a mathematical subset: both can be equal.

  2. Credit: zig zen