165 releases (stable)

1.0.122 May 7, 2024
1.0.120 Mar 23, 2024
1.0.111 Dec 17, 2023
1.0.110 Oct 28, 2023
0.0.0 Dec 28, 2019

#3 in FFI

Download history 254665/week @ 2024-01-31 242095/week @ 2024-02-07 265008/week @ 2024-02-14 308496/week @ 2024-02-21 284774/week @ 2024-02-28 256894/week @ 2024-03-06 235831/week @ 2024-03-13 248312/week @ 2024-03-20 237288/week @ 2024-03-27 249678/week @ 2024-04-03 272826/week @ 2024-04-10 266540/week @ 2024-04-17 279205/week @ 2024-04-24 245263/week @ 2024-05-01 270495/week @ 2024-05-08 218218/week @ 2024-05-15

1,060,876 downloads per month
Used in 321 crates (169 directly)



Rust 3K SLoC // 0.0% comments C++ 657 SLoC // 0.0% comments Bazel 200 SLoC // 0.0% comments JavaScript 94 SLoC Shell 11 SLoC Handlebars 7 SLoC

CXX — safe FFI between Rust and C++

github crates.io docs.rs build status

This library provides a safe mechanism for calling C++ code from Rust and Rust code from C++, not subject to the many ways that things can go wrong when using bindgen or cbindgen to generate unsafe C-style bindings.

This doesn't change the fact that 100% of C++ code is unsafe. When auditing a project, you would be on the hook for auditing all the unsafe Rust code and all the C++ code. The core safety claim under this new model is that auditing just the C++ side would be sufficient to catch all problems, i.e. the Rust side can be 100% safe.

cxx = "1.0"

cxx-build = "1.0"

Compiler support: requires rustc 1.63+ and c++11 or newer
Release notes


Please see https://cxx.rs for a tutorial, reference material, and example code.


The idea is that we define the signatures of both sides of our FFI boundary embedded together in one Rust module (the next section shows an example). From this, CXX receives a complete picture of the boundary to perform static analyses against the types and function signatures to uphold both Rust's and C++'s invariants and requirements.

If everything checks out statically, then CXX uses a pair of code generators to emit the relevant extern "C" signatures on both sides together with any necessary static assertions for later in the build process to verify correctness. On the Rust side this code generator is simply an attribute procedural macro. On the C++ side it can be a small Cargo build script if your build is managed by Cargo, or for other build systems like Bazel or Buck we provide a command line tool which generates the header and source file and should be easy to integrate.

The resulting FFI bridge operates at zero or negligible overhead, i.e. no copying, no serialization, no memory allocation, no runtime checks needed.

The FFI signatures are able to use native types from whichever side they please, such as Rust's String or C++'s std::string, Rust's Box or C++'s std::unique_ptr, Rust's Vec or C++'s std::vector, etc in any combination. CXX guarantees an ABI-compatible signature that both sides understand, based on builtin bindings for key standard library types to expose an idiomatic API on those types to the other language. For example when manipulating a C++ string from Rust, its len() method becomes a call of the size() member function defined by C++; when manipulating a Rust string from C++, its size() member function calls Rust's len().


In this example we are writing a Rust application that wishes to take advantage of an existing C++ client for a large-file blobstore service. The blobstore supports a put operation for a discontiguous buffer upload. For example we might be uploading snapshots of a circular buffer which would tend to consist of 2 chunks, or fragments of a file spread across memory for some other reason.

A runnable version of this example is provided under the demo directory of this repo. To try it out, run cargo run from that directory.

mod ffi {
    // Any shared structs, whose fields will be visible to both languages.
    struct BlobMetadata {
        size: usize,
        tags: Vec<String>,

    extern "Rust" {
        // Zero or more opaque types which both languages can pass around but
        // only Rust can see the fields.
        type MultiBuf;

        // Functions implemented in Rust.
        fn next_chunk(buf: &mut MultiBuf) -> &[u8];

    unsafe extern "C++" {
        // One or more headers with the matching C++ declarations. Our code
        // generators don't read it but it gets #include'd and used in static
        // assertions to ensure our picture of the FFI boundary is accurate.

        // Zero or more opaque types which both languages can pass around but
        // only C++ can see the fields.
        type BlobstoreClient;

        // Functions implemented in C++.
        fn new_blobstore_client() -> UniquePtr<BlobstoreClient>;
        fn put(&self, parts: &mut MultiBuf) -> u64;
        fn tag(&self, blobid: u64, tag: &str);
        fn metadata(&self, blobid: u64) -> BlobMetadata;

Now we simply provide Rust definitions of all the things in the extern "Rust" block and C++ definitions of all the things in the extern "C++" block, and get to call back and forth safely.

Here are links to the complete set of source files involved in the demo:

To look at the code generated in both languages for the example by the CXX code generators:

   # run Rust code generator and print to stdout
   # (requires https://github.com/dtolnay/cargo-expand)
$ cargo expand --manifest-path demo/Cargo.toml

   # run C++ code generator and print to stdout
$ cargo run --manifest-path gen/cmd/Cargo.toml -- demo/src/main.rs


As seen in the example, the language of the FFI boundary involves 3 kinds of items:

  • Shared structs — their fields are made visible to both languages. The definition written within cxx::bridge is the single source of truth.

  • Opaque types — their fields are secret from the other language. These cannot be passed across the FFI by value but only behind an indirection, such as a reference &, a Rust Box, or a UniquePtr. Can be a type alias for an arbitrarily complicated generic language-specific type depending on your use case.

  • Functions — implemented in either language, callable from the other language.

Within the extern "Rust" part of the CXX bridge we list the types and functions for which Rust is the source of truth. These all implicitly refer to the super module, the parent module of the CXX bridge. You can think of the two items listed in the example above as being like use super::MultiBuf and use super::next_chunk except re-exported to C++. The parent module will either contain the definitions directly for simple things, or contain the relevant use statements to bring them into scope from elsewhere.

Within the extern "C++" part, we list types and functions for which C++ is the source of truth, as well as the header(s) that declare those APIs. In the future it's possible that this section could be generated bindgen-style from the headers but for now we need the signatures written out; static assertions will verify that they are accurate.

Your function implementations themselves, whether in C++ or Rust, do not need to be defined as extern "C" ABI or no_mangle. CXX will put in the right shims where necessary to make it all work.

Comparison vs bindgen and cbindgen

Notice that with CXX there is repetition of all the function signatures: they are typed out once where the implementation is defined (in C++ or Rust) and again inside the cxx::bridge module, though compile-time assertions guarantee these are kept in sync. This is different from bindgen and cbindgen where function signatures are typed by a human once and the tool consumes them in one language and emits them in the other language.

This is because CXX fills a somewhat different role. It is a lower level tool than bindgen or cbindgen in a sense; you can think of it as being a replacement for the concept of extern "C" signatures as we know them, rather than a replacement for a bindgen. It would be reasonable to build a higher level bindgen-like tool on top of CXX which consumes a C++ header and/or Rust module (and/or IDL like Thrift) as source of truth and generates the cxx::bridge, eliminating the repetition while leveraging the static analysis safety guarantees of CXX.

But note in other ways CXX is higher level than the bindgens, with rich support for common standard library types. Frequently with bindgen when we are dealing with an idiomatic C++ API we would end up manually wrapping that API in C-style raw pointer functions, applying bindgen to get unsafe raw pointer Rust functions, and replicating the API again to expose those idiomatically in Rust. That's a much worse form of repetition because it is unsafe all the way through.

By using a CXX bridge as the shared understanding between the languages, rather than extern "C" C-style signatures as the shared understanding, common FFI use cases become expressible using 100% safe code.

It would also be reasonable to mix and match, using CXX bridge for the 95% of your FFI that is straightforward and doing the remaining few oddball signatures the old fashioned way with bindgen and cbindgen, if for some reason CXX's static restrictions get in the way. Please file an issue if you end up taking this approach so that we know what ways it would be worthwhile to make the tool more expressive.

Cargo-based setup

For builds that are orchestrated by Cargo, you will use a build script that runs CXX's C++ code generator and compiles the resulting C++ code along with any other C++ code for your crate.

The canonical build script is as follows. The indicated line returns a cc::Build instance (from the usual widely used cc crate) on which you can set up any additional source files and compiler flags as normal.

# Cargo.toml

cxx-build = "1.0"
// build.rs

fn main() {
    cxx_build::bridge("src/main.rs")  // returns a cc::Build


Non-Cargo setup

For use in non-Cargo builds like Bazel or Buck, CXX provides an alternate way of invoking the C++ code generator as a standalone command line tool. The tool is packaged as the cxxbridge-cmd crate on crates.io or can be built from the gen/cmd directory of this repo.

$ cargo install cxxbridge-cmd

$ cxxbridge src/main.rs --header > path/to/mybridge.h
$ cxxbridge src/main.rs > path/to/mybridge.cc


Be aware that the design of this library is intentionally restrictive and opinionated! It isn't a goal to be powerful enough to handle arbitrary signatures in either language. Instead this project is about carving out a reasonably expressive set of functionality about which we can make useful safety guarantees today and maybe extend over time. You may find that it takes some practice to use CXX bridge effectively as it won't work in all the ways that you are used to.

Some of the considerations that go into ensuring safety are:

  • By design, our paired code generators work together to control both sides of the FFI boundary. Ordinarily in Rust writing your own extern "C" blocks is unsafe because the Rust compiler has no way to know whether the signatures you've written actually match the signatures implemented in the other language. With CXX we achieve that visibility and know what's on the other side.

  • Our static analysis detects and prevents passing types by value that shouldn't be passed by value from C++ to Rust, for example because they may contain internal pointers that would be screwed up by Rust's move behavior.

  • To many people's surprise, it is possible to have a struct in Rust and a struct in C++ with exactly the same layout / fields / alignment / everything, and still not the same ABI when passed by value. This is a longstanding bindgen bug that leads to segfaults in absolutely correct-looking code (rust-lang/rust-bindgen#778). CXX knows about this and can insert the necessary zero-cost workaround transparently where needed, so go ahead and pass your structs by value without worries. This is made possible by owning both sides of the boundary rather than just one.

  • Template instantiations: for example in order to expose a UniquePtr<T> type in Rust backed by a real C++ unique_ptr, we have a way of using a Rust trait to connect the behavior back to the template instantiations performed by the other language.

Builtin types

In addition to all the primitive types (i32 <=> int32_t), the following common types may be used in the fields of shared structs and the arguments and returns of functions.

name in Rustname in C++restrictions
&[T]rust::Slice<const T>cannot hold opaque C++ type
&mut [T]rust::Slice<T>cannot hold opaque C++ type
CxxStringstd::stringcannot be passed by value
Box<T>rust::Box<T>cannot hold opaque C++ type
UniquePtr<T>std::unique_ptr<T>cannot hold opaque Rust type
SharedPtr<T>std::shared_ptr<T>cannot hold opaque Rust type
[T; N]std::array<T, N>cannot hold opaque C++ type
Vec<T>rust::Vec<T>cannot hold opaque C++ type
CxxVector<T>std::vector<T>cannot be passed by value, cannot hold opaque Rust type
*mut T, *const TT*, const T*fn with a raw pointer argument must be declared unsafe to call
fn(T, U) -> Vrust::Fn<V(T, U)>only passing from Rust to C++ is implemented so far
Result<T>throw/catchallowed as return type only

The C++ API of the rust namespace is defined by the include/cxx.h file in this repo. You will need to include this header in your C++ code when working with those types.

The following types are intended to be supported "soon" but are just not implemented yet. I don't expect any of these to be hard to make work but it's a matter of designing a nice API for each in its non-native language.

name in Rustname in C++
BTreeMap<K, V>tbd
HashMap<K, V>tbd
tbdstd::map<K, V>
tbdstd::unordered_map<K, V>

Remaining work

This is still early days for CXX; I am releasing it as a minimum viable product to collect feedback on the direction and invite collaborators. Please check the open issues.

Especially please report issues if you run into trouble building or linking any of this stuff. I'm sure there are ways to make the build aspects friendlier or more robust.

Finally, I know more about Rust library design than C++ library design so I would appreciate help making the C++ APIs in this project more idiomatic where anyone has suggestions.


Licensed under either of Apache License, Version 2.0 or MIT license at your option.
Unless you explicitly state otherwise, any contribution intentionally submitted for inclusion in this project by you, as defined in the Apache-2.0 license, shall be dual licensed as above, without any additional terms or conditions.


~22K SLoC