Rust and C++ interoperability
(written August 2020)
Chrome engineers are experimenting with Rust. For the foreseeable future, C++ is the reigning monarch in our codebase, and any use of Rust will need to fit in with C++ — not the other way around. This seems to present some C++/Rust interoperability challenges which nobody else has faced.
We'd need to solve these before considering Rust as (nearly) a first-class citizen in our codebase. If we can’t solve these, Rust would at best be isolated to “leaf nodes” which don’t interact much with the rest of our codebase. And if that’s all we can use Rust for, that calls into question whether the costs of an extra language are justified in the first place.
As C++ is the ruler, we are primarily concerned with the ability for new Rust code to call into existing C++ code, rather than C++ to Rust calls.
We think it’s important for Rust to be able to call C++ functions in a way that meets the following criteria:
No need for the “unsafe” keyword unless something is known to be less safe than normal C++
For a Rustacean, this is controversial - all C++ is unsafe! But “unsafe” should be a really bad code smell. If “unsafe” is needed for all C++ calls, there will be thousands of them, and the “unsafe” keyword will lose its meaning. Where objects are simply passed backwards and forwards between Rust and C++, we must avoid the word ‘unsafe’. It should be restricted to patches of genuinely unsafe Rust code, and for C++ interoperability code where there’s shared ownership or other complexities. This particular property is satisfied by dtolnay’s marvellous cxx library already.
No overhead in the general case
LTO and cross-language inlining already solve this in principle. There are cases where overhead is necessary at the C++ boundary — especially, the UTF check required when strings are passed from C++ to Rust. This can be dealt with by handling such strings as &[u8] in Rust code, until string manipulation is really necessary, so we do not need any further innovations here. This box is checked.
No boilerplate or redeclarations. No C++ annotations. Ideally, no allowlist
If a C++ API exists, Rust should be able to call it. It’s that simple. The declaration in C++ should be sufficient. There should be no need for an allowlist, a redeclaration in Rust, or any Rust shim. Rare exceptions will exist (e.g. overloaded functions) and in some cases we’ll want to make an idiomatic Rust wrapper, but in general, that shouldn’t be necessary. This is not just aesthetic preference. Our codebase is complex and polluting it with extra annotations would be a small, but noticeable, tax on how everyone works.
Broad type support - with safety
cxx is the current state-of-the-art for safely exchanging data between C++ and Rust. Our “base” library exposes 1768 APIs which are used by other parts of Chrome. 1052 of those functions only take parameters which are types that can already be supported by cxx. 12 more are planned in the near term for cxx (e.g. more flexible slices). That’s ~60% of our APIs, which is good but not great. Another 12% can be supported if we are able to pass std::string and similar string types into existing C++ APIs. These can’t be represented in a Rust struct due to an internal pointer, but as cxx generates code on both the C++ and Rust side, it should be possible to own a UniquePtr<CxxString> on the Rust side, yet pass it into an existing C++ API which takes a std::string by value. (That sounds fairly straightforward, but it becomes much more complex when you’re talking about structs containing std::strings, such as url::Origin. Such a struct could only be owned as a UniquePtr<opaque type> from the Rust side, which would prevent field access. Solutions can be imagined but need more thought.) Another ~20% are functions which take pointer parameters - in our case, these are very often out parameters. We need to see how we can programmatically identify those which are ‘simple’ out parameters and allow Rust to populate them safely. The good news is that this leaves just 8% of our functions which can’t be supported by the cxx model of interop. Most of these are passing C++ structs (by value) which have raw pointers within them. This seems largely insoluble in Rust but they’re so rare that we can create case-by-case idiomatic wrappers. There are some caveats here: this analysis is based on symbols exported by the binary, rather than source code analysis. In some cases these APIs would be wrapped by inline functions, templates or macros, which this analysis ignores. It also ignores return values and direct field access. And of course, “base” isn’t the only set of APIs which our code would need to call
- it’s probably that higher-level functions would have more complex arguments on average so be less likely to fall in the ‘good’ bucket.
Ergonomics - with safety
From Rust code, we need to be able to instantiate C++ objects, safely pass around ownership (there are no significant problems with cxx’s UniquePtr here), call methods on them (both plain and virtual). For “plain old data” types in C++, containing simple, cxx-compatible fields, we need to be able to manipulate those fields. Most of this can be achieved with cxx already (though we need a way to call through to std::make_unique from Rust code for a type that’s opaque to Rust). We need this to be smooth enough that we do not need to wrap a typical C++ type in a Rust wrapper. So far, so good. But we also need: to act (at Rust build time) upon #defines set up by our C++ headers and build-time rules, figure out a plan for calling C++ overloaded functions and operators, call macros (e.g. LOG(ERROR) << “eek”), make templated functions and types available (possibly very hard, though bindgen does a remarkably good job here), and probably many other things we haven’t yet thought of. It may be that the best way to handle some of these cases is some inline C++ code within Rust (like the cpp crate but with the benefit of cxx’s safety). One specific challenge is reference-counted objects. We need the reference count to be shared between referees on the Rust and Chrome side. The bigger challenge here is how to deal with the prevalence of multiple mutable references on the C++ side, without the ability to do something like a RefCell::borrow_mut to ensure even runtime safety. It may be that we need to mark involvement with all such reference-counted objects as truly ‘unsafe’ from the Rust side. In general we think we can live without Rust types inheriting from C++ types, but there’s one exception: pure virtual observers. cxx provides the ability to pass function pointers from Rust to C++, so it’s quite possible for us to make wrapper types here. Ideally, though, this becomes ergonomic and seamless as well. More investigation is needed here.
What we don’t need
We believe we can live without: self-referential C++ types being passed by value into Rust (except for strings), Rust types inheriting from non-pure-virtual C++ types; variadic arguments, “safe” reference counting. (There will be cases where the absence of these features is annoying, but hopefully rare.) All this may be wrong: we still have much learning to do.
We think the hardest part of this is imagining a safe way to pass types between Rust and C++. That requires auto-generated shim code on both the Rust and C++ side. That’s already achieved by cxx with terrific emergent safety properties. And so that’s our basic model.
But, we don’t want to specify a cxx::bridge section for every API. We therefore need the cxx::bridge to be generated using a bindgen-like tool.
We don’t believe Rust language changes are needed. Some C++ types can’t be owned by value in Rust — for example std::string with its self-referential pointer — but we believe that good C++ interoperability can be achieved even if Rust can only own such objects by pointer. We may be wrong here as well!
For now, Chrome investment in Rust will remain a background investigation (mostly directed towards prototyping these tools and techniques). If we become convinced this sort of interoperability is possible, we’ll revisit widespread use of Rust in Chrome, and at that point we plan to work hard to achieve this with robust production-quality solutions.
(update September 2021) There's progress across the Rust community in solving many of these problems - see for example moveit, autocxx and mosaic.