Mohannad Al-Awad (Dec 21 2022 at 04:55): Mohannad Al-Awad (Dec 21 2022 at 04:55):Hello :wave: Other than Rust, do we have choices to integrate with the Cranelift API, say from Haskell and C# via a shard library of some sort?
The CLI is there but I'm thinking of something better. Thank you!
Jamey Sharp (Dec 21 2022 at 17:50):There aren't currently any Cranelift bindings for other languages. Rust has good support for exposing C-compatible APIs, and Haskell has good support for using C APIs, so it's certainly possible. Tools like cbindgen (to generate a C header for a Rust API) and hsc2hs (to generate Haskell bindings from a C header) should help. (I assume there are similar options for C#, but I'm not familiar with that ecosystem.) But nobody has done any of that work for Cranelift.
After some thought, I don't see a super easy path to providing a Cranelift C API, because it has a lot of entry points. Though one piece that you might think would be a pain actually is the easiest part: the InstBuilder API is auto-generated, so auto-generating additional FFI-compatible entry points for that API would be relatively straightforward. I'd definitely consider merging a PR that did that. Designing a reasonable C API for everything else you might want is not as clear though.
I should note that Wasmtime does have a C API and bindings for several languages, so if you don't mind using WebAssembly as your IR, you can get a JIT/AOT compiler built on Cranelift that you can use from whatever language you want. That will be a more stable API over time than using Cranelift directly, too.
Chris Fallin (Dec 21 2022 at 18:08):I would agree that a builder-focused API would not be too hard (conceptually, though there's a lot of plumbing and implementation-level thought to do here). I suspect that an API that allows introspection and mutation of the IR might be a bit more difficult, but maybe also not as immediately useful (I suspect that most of the demand for this is from projects that want to use Cranelift as a backend; @Mohannad Al-Awad is that the case for you?). In that case, an API of "constructors" -- for instructions, blocks, and functions -- is maybe ~a month of work total? I'd also be happy to discuss further and help with design here, along with Jamey, if someone is interested.
Jamey Sharp (Dec 21 2022 at 18:29):@Chris Fallin, does your time estimate include cranelift_codegen::Context, cranelift-module, and either cranelift-jit or cranelift-object? Do you think there's anything especially challenging with any of those?
Chris Fallin (Dec 21 2022 at 19:36):yeah, I was thinking about what it would take to give a basic "compile this function" toplevel entry point, given a CLIFFunction* and CompileOptions* or somesuch; some of the fiddly plumbing would come about in communicating the result of that (relocations, etc). Wrapping either the jit or object crate to get module-level support is another bit of work, maybe not necessary for a very first step...
Mohannad Al-Awad (Dec 22 2022 at 15:28):Hi @Jamey Sharp and thank you for the suggestions. To be honest, I have considered WASM/WASI setup before for a different reason: portability in the sense that I can exchange backend compilers. However, I am not sure, due to my ignorance, if WASM/WASI is ready for general purpose programming while interfacing with other system/runtime APIs. Another consideration was also whether I will be able to pass the IR as an in-memory data structure to the backend compiler rather than passing it in some text/binary format that will then need to be parsed again, for performance. If you may comment on these points, I will be thankful. As for your point of multiple entry points, are you pointing out the large API surface that will need to be made C/FFI-compatible?
Mohannad Al-Awad (Dec 22 2022 at 15:31):Hi @Chris Fallin and thank you also for your extended help. Yes, you are correct about my demand of Cranelift for backend compilation. Now, I need to decide whether to go with WASM/WASI IR or Cranelift-native IR, considering the points that I have mentioned above.
Mohannad Al-Awad (Dec 22 2022 at 15:45):One more point for that decision is also possible optimizations with WASM/WASI due to potentially lost metadata from the previous steps that could benefit Cranelift having used its IR natively... not sure.
Last updated: Nov 22 2024 at 16:03 UTC