This is awtoau's fork of upstream
immunant/c2rust, maintained to support
linux-rs — a port of Linux kernel lib/*.c
files to Rust. The goal for this fork is for c2rust transpile to
become linux-rs's primary translation source — not just a reference
tool — by fixing it until its output reliably conforms to linux-rs's
rule-cited translation discipline (every deviation from literal C
semantics must be either Rust-forced or explicitly licensed by a cited
rule) for real kernel lib/*.c source.
Getting there is staged. Right now (stage 1) the priority is
reliability: get c2rust transpile to run over the full corpus without
crashing or hanging, fixing genuine gaps in its C-construct coverage —
see issue #1 for the
running log. Once transpilation itself is reliable, later stages target
output quality directly, potentially including changes to
c2rust transpile itself and/or a linux-rs-side normalization pass, so
its output moves from "unsafe, unidiomatic, but functionally faithful"
(c2rust's own stated design goal, see below) toward what linux-rs's
rules require.
Regardless of how good the output gets, landing requires two things,
both of them, for every translation whether hand-written or
c2rust-derived: (1) linux-rs's full verification pipeline passing —
build, a differential oracle against the original C, a QEMU boot, and
the kernel's KUnit test suites — and (2) the translation actually
conforming to linux-rs's rules (rulesdb/rules/*.toml): every
deviation from literal C semantics traceable to either something Rust's
type system forces or a specific cited rule. Pipeline-green alone is
not sufficient and is not treated as sufficient — this project has
already seen, empirically, drafts that compiled clean, booted, and
passed KUnit while being semantically wrong in ways the tests didn't
happen to exercise (a silently swapped allocator was one concrete
case). A c2rust-derived translation that passes the pipeline but, say,
reimplements a function the target arch actually overrides, or invents
a deviation with no rule behind it, is not landable just because the
tests didn't catch it — the same rule-conformance review a hand
translation gets applies here too.
Stage-1 reliability patches are written to be general fixes (not narrow
kernel-only special-cases) wherever the underlying gap is a real one in
c2rust's C-construct coverage — things like GCC extended-asm label
operands (asm goto), architecture-conditional builtin vector types,
and _Pragma-based loop-unroll attributes, all real gaps against
kernel source that upstream's existing test corpus doesn't happen to
exercise, and are upstreamable in principle. This fork does not
currently open PRs against immunant/c2rust — fixes stay here.
Once a construct transpiles reliably (stage 1), later work can target
emitting linux-rs's required idiom directly instead of a literal
transliteration — e.g. WARN_ON(cond) as kernel::warn_on!(cond)
instead of the expanded let __ret = !!(cond); if unlikely(...) {...}
scaffolding, or the kernel's generic fls/__fls/__ffs/fls64
bit-scan functions as leading_zeros()/trailing_zeros() arithmetic
instead of a transliterated byte-scan loop. Unlike stage-1 fixes,
these are linux-rs-specific idiom choices, not general C-to-Rust
correctness fixes — plain c2rust transpile with no extra flags
must stay byte-for-byte identical to stock upstream behavior for
anyone else using this fork as an ordinary transpiler.
Each rewrite is registered in c2rust-transpile::kernel_idioms
(KernelIdiomRule) and only fires when explicitly requested via
--enable-rule=<rule>[,<rule>...] (or --enable-rule=all for every
rule this build knows about) on the c2rust transpile CLI. linux-rs's
own scripts/run_c2rust_baseline.py passes --enable-rule=all
unconditionally, so its baseline/regression runs always exercise every
landed rule without needing an update each time a new one is added —
adding rule N+1 means adding it to the registry and (if warranted)
rulesdb/rules/*.toml on the linux-rs side; the Python harness doesn't
change.
The pattern for adding a new one, established across three rules this
far (warn-on, fls-family, swap-mem-swap):
- Rank remaining rule-conformance violations by count
(
scripts/check_c2rust_rule_conformance.pyon linux-rs, or querypatterns.db'sc2rust_rule_conformancetable directly —rule_idis stored without therulesdb/rules/NNNN--style numeric prefix, e.g."fls-family"not"0006-fls-family") — fix the biggest gap first. - Find the real C definition of whatever's being mistranslated and
work out the exact Rust replacement BEFORE writing any translator
code — for anything involving arithmetic (not a pure 1:1 syntactic
substitution like a macro call), verify numerically against real
compiled C behavior across edge cases, not just one happy-path
value. A translation that compiles and looks plausible but is
silently wrong by a constant offset is strictly worse than the
literal transliteration it would replace (see
rulesdb/rules/'s own[validation] negativewarnings — several rules document exactly this failure mode from a prior attempt). - The Rust replacement must call into linux-riscv's real, vendored
Rust-for-Linux
kernelcrate (linux-riscv/rust/kernel/*.rs) — never reimplement kernel behavior independently inside the translated file.WARN_ONtargetskernel::warn_on!(rust/ kernel/bug.rs), itself backed by arust/helpers/bug.cC shim (rust_helper_WARN_ON) that bindgen exposes asbindings::WARN_ON— a thin bridge to the real C macro, not a Rust reimplementation of its behavior. If the target rule needs a kernel-crate API that doesn't exist yet inrust/kernel/, porting/wrapping that API (verified against the real C it wraps, with arust/helpers/*.cshim if the C side is a macro/inline function bindgen can't see directly) is a REQUIRED prerequisite step, done and verified before any c2rust rewrite targets it — not a shortcut to skip by having the rewrite emit ad hoc equivalent logic instead of the real API call. (Rust standard-library targets —leading_zeros(),core::mem::swap— don't need this: they're not kernel-crate APIs, just the correct idiomatic Rust for the C construct.) - Detect via the most precise signal available: macro-expansion origin (Clang reports which macro a given AST subtree expanded from) where the construct is a macro, exact function name + parameter/ return C-type-kind match where it's a real header-inline function — never bare AST-shape pattern-matching alone if a more precise origin signal exists, since shape-only matching risks misfiring on unrelated hand-written code that happens to look similar.
- Add the new arm to
KernelIdiomRuleand gate the rewrite behind it; confirm the DEFAULT (no--enable-ruleflags) path is provably unchanged before anything else. - Verify: full-corpus baseline before/after
(
dev.py c2rust-baselineon linux-rs) must show identical outcome counts (clean/dropped_decls/crash) — a rule rewrite changes what gets emitted for already-successful declarations, never whether a file succeeds or fails to transpile at all;dev.py c2rust-regress <before-rev> <after-rev>for a hard per-declaration regression gate; re-run the conformance checker and confirm the violation count for that rule drops to (ideally) zero, with any nonzero remainder explained (an excluded edge case, a name collision with an unrelated user function, etc.) rather than silently accepted.
C2Rust helps you migrate C99-compliant code to Rust.
The transpiler, c2rust transpile,
produces unsafe Rust code that closely mirrors the input C code.
The primary goal of the transpiler is to preserve functionality;
test suites should continue to pass after translation.
The output of c2rust transpile is unsafe and unidiomatic; it is merely the first step in a longer migration process.
Generating safe and idiomatic Rust code from C ultimately requires additional work.
This work can be done by a human, a large language model, a deterministic tool, or some combination thereof.
flowchart LR
A[C sources] --> B
subgraph "C2Rust toolchain"
B["`c2rust transpile`"] --> C["`c2rust refactor`"]
C --> D["`c2rust postprocess`"]
end
D --> E[Human and/or agentic workflow]
E --> F[Safe, idiomatic Rust]
click B "https://github.com/immunant/c2rust/tree/master/c2rust-transpile" "c2rust transpile"
click C "https://github.com/immunant/c2rust/tree/master/c2rust-refactor" "c2rust refactor"
click D "https://github.com/immunant/c2rust/tree/master/c2rust-postprocess" "c2rust postprocess"
For instance, we provide a deterministic refactoring tool to automate cleanup across the files produced by c2rust transpile.
We also provide an LLM-powered postprocessing tool for additional types of cleanup that are hard to do deterministically.
Even though the postprocessor validates the output of LLMs, it can introduce errors; we recommend using it in combination with a robust test suite.
You can also cross-check the translated code against the original (tutorial).
You can try c2rust transpile directly in the Compiler Explorer.
This uses the current master branch, updated every night.
To learn more about using and developing C2Rust, check out the manual. The manual is still a work-in-progress, so if you can't find something please let us know. c2rust.com/manual/ also has not been updated since ~2019, so refer to the in-tree ./manual/ for more up-to-date instructions.
C2Rust requires LLVM 15 or later with its corresponding clang compiler and libraries.
Python (through uv), CMake 3.5 or later and openssl (1.0) are also required.
These prerequisites may be installed with the following commands, depending on your platform:
Python:
curl -LsSf https://astral.sh/uv/install.sh | sh
uv venv
uv pip install -r scripts/requirements.txt-
Ubuntu 22.04, Debian 12, and later:
apt install build-essential llvm clang libclang-dev cmake libssl-dev pkg-config git
Depending on the LLVM distribution, the llvm-dev package may also be required.
For example, the official LLVM packages from apt.llvm.org require llvm-dev to be installed.
-
Arch Linux:
pacman -S base-devel llvm clang cmake openssl
-
NixOS / nix:
nix-shell
-
macOS: Xcode command-line tools and recent LLVM (we recommend the Homebrew version) are required.
xcode-select --install brew install llvm cmake openssl
The C2Rust transpiler now builds using a stable Rust compiler. If you are developing other features, you may need to install the correct nightly compiler version.
cargo install --locked c2rustYou can also set the LLVM version explicitly if you have multiple installed, like this, for example:
LLVM_CONFIG_PATH=llvm-config-15 cargo install --locked c2rustIf you're using LLVM from Homebrew (either on Apple Silicon, Intel Macs, or Linuxbrew), you can run:
LLVM_CONFIG_PATH="$(brew --prefix)/opt/llvm/bin/llvm-config" cargo install --locked c2rustor for a specific LLVM version,
LLVM_CONFIG_PATH="$(brew --prefix)/opt/llvm@22/bin/llvm-config" cargo install --locked c2rustOn Gentoo, you need to point the build system to
the location of libclang.so and llvm-config as follows:
LLVM_CONFIG_PATH=/path/to/llvm-config LIBCLANG_PATH=/path/to/libclang.so cargo install --locked c2rustIf you have trouble with building and installing, or want to build from the latest master, the developer docs provide more details on the build system.
If you'd like to check our recently developed features or you urgently require a bugfixed version of c2rust,
you can install it directly from Git:
cargo install --locked --git https://github.com/immunant/c2rust.git c2rustPlease note that the master branch is under constant development and you may experience issues or crashes.
You should also set LLVM_CONFIG_PATH accordingly if required as described above.
c2rust and c2rust-transpile are installed by default and can be built on stable rustc.
The other tools, such as c2rust-refactor, use rustc internal APIs, however,
and are thus pinned to a specific rustc nightly version: nightly-2022-11-03.
These are also not published to crates.io.
To install these, these can be installed with cargo with the pinned nightly. For example,
cargo +nightly-2022-11-03 install --locked --git https://github.com/immunant/c2rust.git c2rust-refactorHowever, we recommend installing them from a full checkout, as this will resolve the pinned nightly automatically:
git clone https://github.com/immunant/c2rust.git
cd c2rust
cargo build --releaseThese tools, like c2rust-refactor, can then also be invoked through c2rust
as c2rust refactor, assuming they are installed in the same directory.
To translate C files specified in compile_commands.json (see below),
run the c2rust tool with the transpile subcommand:
c2rust transpile compile_commands.jsonc2rust also supports a trivial transpile of source files, e.g.:
c2rust transpile project/*.c project/*.hFor non-trivial projects, the translator requires the exact compiler commands used to build the C code.
This information is provided via a compilation database
file named compile_commands.json (note that it must be named exactly compile_commands.json;
otherwise libclangTooling can have (silent) trouble resolving it correctly).
(Read more about compilation databases here).
Many build systems can automatically generate this file;
we show a few examples below.
Once you have a compile_commands.json file describing the C build,
translate the C code to Rust with the following command:
c2rust transpile path/to/compile_commands.jsonTo generate a Cargo.toml template for a Rust library, add the --emit-build-files option:
c2rust transpile --emit-build-files path/to/compile_commands.jsonTo generate a Cargo.toml template for a Rust binary, do this:
c2rust transpile --binary myprog path/to/compile_commands.jsonWhere --binary myprog tells the transpiler to use
the main function from myprog.rs as the entry point for a binary.
This can be repeated multiple times for multiple binaries.
The translated Rust files will not depend directly on each other like normal Rust modules. They will export and import functions through the C API. These modules can be compiled together into a single static Rust library or binary.
You can run with --reorganize-definitions (which invokes c2rust-refactor),
which should deduplicate definitions and directly import them
with uses instead of through the C API.
The refactorer can also be run on its own to run other refactoring passes:
c2rust refactor --cargo $transformThere are several known limitations in this translator. The translator will emit a warning and attempt to skip function definitions that cannot be translated.
The compile_commands.json file can be automatically created
using either cmake, meson, bear, intercept-build, or compiledb.
It may be a good idea to remove optimizations (-OX) from the compilation database,
as there are optimization builtins which we do not support translating.
When creating the initial build directory with cmake,
specify -DCMAKE_EXPORT_COMPILE_COMMANDS=1.
This only works on projects configured to be built by cmake.
This works on Linux and MacOS.
cmake -DCMAKE_EXPORT_COMPILE_COMMANDS=1 ...When creating the initial build directory with meson,
it will automatically generate a compile_commands.json
file inside of <build_dir>.
meson setup <build_dir>bear is recommended for projects whose build systems
don't generate compile_commands.json automatically
(make, for example, unlike cmake or meson). It can also be useful
for cmake and meson to generate a subset of the full compile_commands.json,
as it records all compilations that a subcommand does.
It can be installed with
apt install bearor
brew install bearUsage:
bear -- <build command><build command> can be make, make/cmake for a single target, or a single cc compilation:
bear -- make
bear -- cmake --build . --target $target
bear -- cc -c program.cNote that since it detects compilations,
if compilations are cached (by make for example),
you'll need a clean build first (e.g. make clean).
intercept-build (part of the scan-build)
is very similar, but not always as up-to-date and comprehensive as bear.
intercept-build is bundled with clang under tools/scan-build-py,
but a standalone version can be easily installed via pip with:
uv tool install scan-buildThe compiledb package can also be used for make projects if the other tools don't work.
Unlike the others, it doesn't require a clean build/make clean.
Install via pip with:
uv tool install compiledbUsage:
# After running
./autogen.sh && ./configure # etc.
# Run
compiledb makeTo report issues with translation or refactoring, please use our Issue Tracker.
To reach the development team, join our discord channel or email us at c2rust@immunant.com.
I translated code on platform X, but it didn't work correctly on platform Y.
We run the C preprocessor before translation to Rust.
This specializes the code to the target platform (usually the host platform).
We do, however, support cross-architecture transpilation with a different sysroot
(cross-OS transpilation is more difficult because
it can be difficult to get a sysroot for the target OS).
For example, on an aarch64-linux-gnu host, to cross-transpile to x86_64-linux-gnu,
you can run
sudo apt install gcc-x86-64-linux-gnu # install cross-compiler, which comes with a sysroot
c2rust transpile ${existing_args[@]} -- --target=x86_64-linux-gnu --sysroot=/usr/x86_64-linux-gnuThese extra args are passed to the libclangTooling that c2rust-transpile uses.
You sometimes also need to pass extra headers, as occasionally headers are installed globally
in the default sysroot and won't be found in the cross-compiling sysroot.
What platforms can C2Rust be run on?
The translator and refactoring tool support both macOS and Linux.
This is a list of all significant uses of c2rust transpile that we know of:
| Rust | C | By | Safety | Description |
|---|---|---|---|---|
rav1d |
dav1d |
@memorysafety, @immunant | fully safe | AV1 decoder |
rexpat |
libexpat |
@immunant | safety unfinished | streaming XML parser |
unsafe-libyaml |
libyaml |
@dtolnay | minor cleanup, fully unsafe | YAML parser and writer used by serde_yaml |
libyaml-safer |
libyaml |
@simonask | fully safe | safe fork of unsafe-libyaml |
libbzip2-rs |
bzip2 |
@trifectatechfoundation | fully safe | file compression |
tsuki |
lua |
@ultimaweapon | fully safe | Lua interpreter |
spiro.rlib |
spiro |
@ctrlcctrlv | fully safe | spline interpolation |
sapp-kms |
sokol |
@not-fl3 | cleaned up, still unsafe | application rendering library |
lhdcv5 |
(unknown) | (unknown) | fully safe | Bluetooth audio codec |
linux-rs |
Linux kernel lib/*.c |
@awtoau | unsafe FFI-shim style; landing bar is build + differential oracle + boot + KUnit, not c2rust output trusted blind |
riscv64 kernel lib/ port; this fork's patches target making c2rust a reliable primary source for it (see About this fork) |
If any other project successfully uses c2rust, feel free to add your ported project here.
This material is available under the BSD-3 style license as found in the LICENSE file.
The C2Rust translator is inspired by Jamey Sharp's Corrode translator. We rely on Emscripten's Relooper algorithm to translate arbitrary C control flows. Many individuals have contributed bug fixes and improvements to C2Rust; thank you so much!
This material is based upon work supported by the United States Air Force and DARPA under Contracts No. FA8750-15-C-0124, HR0011-22-C-0020, and HR00112590133. Any opinions, findings and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the United States Air Force or DARPA.
Distribution Statement A, "Approved for Public Release, Distribution Unlimited."