wit-bindgen Draft: Symmetric ABI tracking and minor side functionality

My fork has aggregated several potentially separable parts over the past years.

I would like to come up with a plan to incrementally merge in interest order, so I would love to get feed-back about which features are most wanted/urgent to separate them in the most efficient way. Perhaps we want to merge Joel's work first to minimize friction on his work/merge.

[X] C++ guest support https://github.com/bytecodealliance/wit-bindgen/issues/826
- [ ] symmetric C++ API (I think the old one is horrible to understand, because move semantic for arguments is alien to C++)
- [ ] planned: C++ WASI 0.3 async (painful because concurrency TS (continuation) isn't standard and shared everything threads are too far off)
[ ] C++ host support (WAMR, stale)
[ ] (outdated?) Canonical ABI compiled to native (mesh, see https://wasmcon24.sched.com/event/1qvIG slide 40)
[ ] shared everything symmetric ABI https://github.com/WebAssembly/component-model/issues/386
- [X] 64bit compatible code generation (offset + alignment)
- [ ] wasm2c bridging support, see wasmcon slide 55
[X] Joel's work on async (WASI 0.3), identical to https://github.com/bytecodealliance/wit-bindgen/pull/1082
- [ ] async WASI 0.3 with symmetric ABI (PoC exists)
[ ] planned: Caller provided buffers https://github.com/WebAssembly/component-model/issues/369

Some of these features are unfinished but given enough interest I clearly want to make them upstreamable.

Nov 30 '24 09:11 cpetig

Thanks for this! Personally though I'd probably leave it mostly up to you as to how best to chunk up this work and land it. The 0.3 bits may take a bit to land so if you can avoid depending on that it might make sense, although if it is likely to heavily conflict with @dicej's work we might want to strategize in the near term.

Otherwise though I find it difficult to predict how to split up "unknown code" in the sense that I haven't reviewed this yet so having an opinion on how best to chunk it up would need a review/understanding on my behalf first, so this is also why I'll trust to you break it up since that makes the review load easier too :)

Dec 04 '24 21:12 alexcrichton

C++ guest support C++ language support #826
    symmetric C++ API (I think the old one is horrible to understand, because move semantic for arguments is alien to C++)
    planned: C++ WASI 0.3 async (painful because concurrency TS (continuation) isn't standard and shared everything threads are too far off)

Is "symmetric C++ API" related to the symmmetric ABI? If so, then see my comments below about the symmetric ABI. And for the async parts, see my comments below about async :-).

C++ host support (WAMR, stale)

As discussed in #395, host support no longer lives in the wit-bindgen repository, so this would make sense to split out, either into wamr itself or a separate repo somewhere. See this section for a discussion of what other hosts are doing.

(outdated?) Canonical ABI compiled to native (mesh, see https://wasmcon24.sched.com/event/1qvIG slide 40)
shared everything symmetric ABI Symmetrical ABI for component fusion and shared everything WebAssembly/component-model#386
    64bit compatible code generation (offset + alignment)
    wasm2c bridging support, see wasmcon slide 55

What would you think about factoring these changes out to live in their own repository? Native ABI interfaces sound like they have some similar needs to cool projects like wRPC, wasm-http-tools, and hopefully more such tools coming, which live in separate repositories, sharing factored-out building blocks like wit-parser.

Joel's work on async (WASI 0.3), identical to Add support for async/streams/futures to Rust generator #1082
    planned: async WASI 0.3 with symmetric ABI (PoC exists)

I think the main thing to do here rhg tnow is to just avoid creating more work for Joel, which I expect means waiting on any changes that would depend on these parts for now.

planned: Caller provided buffers Caller provided buffers question WebAssembly/component-model#369

This would be useful to a lot of people, so if you're looking for pieces to prioritize here, this is my suggestion :-).

Dec 07 '24 17:12 sunfishcode

C++ guest support C++ language support #826
    symmetric C++ API (I think the old one is horrible to understand, because move semantic for arguments is alien to C++)
    planned: C++ WASI 0.3 async (painful because concurrency TS (continuation) isn't standard and shared everything threads are too far off)
Is "symmetric C++ API" related to the symmmetric ABI? If so, then see my comments below about the symmetric ABI. And for the async parts, see my comments below about async :-).

Yes and no, https://github.com/cpetig/wit-bindgen/blob/main/crates/cpp/DESIGN.md will give you a good idea why the old API is really horrible, although the new API introduces some allocation for canonical as it is tailored towards zero allocations for the symmetric ABI.

C++ host support (WAMR, stale)
As discussed in #395, host support no longer lives in the wit-bindgen repository, so this would make sense to split out, either into wamr itself or a separate repo somewhere. See this section for a discussion of what other hosts are doing.

I agree, that other host parts are in a separate repo, but most of the code is shared between the two C++ codegens, and I really want to avoid creating more forks of my code. I will think of a viable way.

(outdated?) Canonical ABI compiled to native (mesh, see https://wasmcon24.sched.com/event/1qvIG slide 40)
shared everything symmetric ABI Symmetrical ABI for component fusion and shared everything WebAssembly/component-model#386
    64bit compatible code generation (offset + alignment)
    wasm2c bridging support, see wasmcon slide 55
What would you think about factoring these changes out to live in their own repository? Native ABI interfaces sound like they have some similar needs to cool projects like wRPC, wasm-http-tools, and hopefully more such tools coming, which live in separate repositories, sharing factored-out building blocks like wit-parser.

Right now it is just another command line switch for Rust and C++ to support the symmetric ABI, and as both bindgens are meant to be a drop-in replacement depending on which ABI you target, I think having the code in one place makes sense.

Joel's work on async (WASI 0.3), identical to Add support for async/streams/futures to Rust generator #1082
    planned: async WASI 0.3 with symmetric ABI (PoC exists)
I think the main thing to do here rhg tnow is to just avoid creating more work for Joel, which I expect means waiting on any changes that would depend on these parts for now.

Right now I experiment with streams for symmetric, async calls already work well (no codegen yet). I plan to separate this into parts after https://github.com/bytecodealliance/wit-bindgen/pull/1082 is merged (less work for me, less chance for conflicts).

planned: Caller provided buffers Caller provided buffers question WebAssembly/component-model#369
This would be useful to a lot of people, so if you're looking for pieces to prioritize here, this is my suggestion :-).

Yes, but this is the most advanced part, see also the discussion on zulip about zero-copy and shared memory which is identical to this change,

Dec 28 '24 11:12 cpetig

My fork has aggregated several potentially separable parts over the past years.

I would like to come up with a plan to incrementally merge in interest order, so I would love to get feed-back about which features are most wanted/urgent to separate them in the most efficient way. Perhaps we want to merge Joel's work first to minimize friction on his work/merge.

[ ] C++ guest support C++ language support #826

[ ] symmetric C++ API (I think the old one is horrible to understand, because move semantic for arguments is alien to C++)

[ ] planned: C++ WASI 0.3 async (painful because concurrency TS (continuation) isn't standard and shared everything threads are too far off)

[ ] C++ host support (WAMR, stale)

[ ] (outdated?) Canonical ABI compiled to native (mesh, see https://wasmcon24.sched.com/event/1qvIG slide 40)

[ ] shared everything symmetric ABI Symmetrical ABI for component fusion and shared everything WebAssembly/component-model#386

[ ] 64bit compatible code generation (offset + alignment)

[ ] wasm2c bridging support, see wasmcon slide 55

That sort of tool, a component compiler, is a blocker for component model support in my projects, which could support WIT-using core modules but cannot handle the intracies of the component model's format. I have, given existing tools, created many issues for using them to compile components, though their methods also turn out just as compilcated.

[ ] Joel's work on async (WASI 0.3), identical to Add support for async/streams/futures to Rust generator #1082

[ ] planned: async WASI 0.3 with symmetric ABI (PoC exists)

[ ] planned: Caller provided buffers Caller provided buffers question WebAssembly/component-model#369

Some of these features are unfinished but given enough interest I clearly want to make them upstreamable.

Feb 04 '25 01:02 gkgoat1

Update: I just got future<string> and stream<string> fully operational mixing native compiled C++ and Rust in a fully automated test case.

Also the alien move semantic was a misunderstanding between me and the C++ standard. @TartanLlama explained the details and (move-) passing by value will become the new standard.

This branch contains a symmetric ABI which still avoids allocation overhead using the symmetric API (passing arguments by reference in both directions).

Jun 22 '25 22:06 cpetig