custom section that documents producing toolchain?

[lukewagner]

In a few contexts now, it has seemed useful to have a standardized custom section that declares the toolchain that produced the .wasm module. For example, browsers could use such a section to measure the rate of usage of various toolchains and languages via telemetry to help understand usage trends on the web. WDYT?

[xtuc]

I often had to ask where that wasm file is comming from, but that could ideally be encoded in the binary.

[sbc100]

LLVM produces a .comment section with this information but we currently don't include that in the wasm object format. I tried adding in the past but it relies on the SHF_STRINGS + SHF_MERGE support in the linker. I can take another look.

One question is weather to include this as custom section called ".comment" or to include it as a data segment. From a clarity point of view its probably best as custom section, but I think the linker implementation might be a little harder in this case.

[dschuff]

The general idea makes sense to me. As @sbc100 said this is common in ELF. I think it would be much simpler (to make use of) as a custom section rather than putting it in the data section. I agree that would probably need some extra linker support; I think something like SHF_MERGE should be enough though. I'm i'm interpreting the sun doc correctly, adding SHF_STRINGS is partly an optimization allowing it to interned in the string table (which we don't have), and partly to allow the linker to locate divisions between multiple items. But for this use case (and probably some others) we could have a simple merge that would treat the whole section as a unit.

[lukewagner]

Great to see the interest! Next suggestion: rather than having a single string that will inevitably go the way of browser UA-strings, how about having a few fields that are intentionally coarse-grained and contain a fixed (but growable) set of values that we maintain in this repo? This would assist in implementing privacy-preserving telemetry that gave a good high-level picture of the toolchain usage. E.g.:

• language (such as: wat, C++, Rust, C#, AssemblyScript, ...)
• processed-by (a comma-delimited set containing all tools that were used in the pipeline to produce the final wasm, such as: llvm, binaryen, wabt, webpack, wasm-bindgen, ...)
• sdk (the top-level tool the developer installed to produce the wasm: Emscripten, Blazor, Unity, Rust-wasm-pack, ...)

Edit: additionally, each of the above field values could be followed by a parenthesized version string, allowing versions that could be easily skipped by tools trying to build coarse-grained summaries.

[est31]

rather than having a single string that will inevitably go the way of browser UA-strings

Just wanting to point out that this is precisely what happened in the non-wasm analogon of such a string value, the producer value. Please see this commit that made the Rust compiler claim to be clang because of a check by gdb, as well as the corresponding issue here.

[xtuc]

I would also suggest adding a date and ensure that the processed-by/sdk includes a version?

[ashleygwilliams]

i'm def in favor of this with @xtuc's caveat of adding a version!

[MaxGraey]

Also will be great add signature of check sum like do in this project: https://github.com/frehberg/wasm-sign

[lukewagner]

@xtuc @ashleygwilliams Good point. Perhaps then we would standardize how the version was embedded (say, between parens right after the tool/sdk name) so that it could be mechanically skipped over.

@MaxGraey Signature/check-sum mostly seem orthogonal to the use cases raised here and probably belongs in a separate custom section which could be proposed in a separate issue.

[binji]

We discussed this in the Oct 2 CG meeting. There was general agreement that this is useful. There was a question about how "official" this custom section should be. We agreed that for now it should be documented in the tool-conventions repo, but not in the core spec.

[lukewagner]

Yep! Unless anyone else is itching to, I'll make a PR to add a BinaryEncoding.md-esque description of the binary format.

[MaxGraey]

Also will be great have metadata field which recognize which post-MVP features used during compilation.

[lukewagner]

That's an interesting idea, but I wonder if the use case is slightly different, deserving a different custom section. In particular, I can see the utility of a wasm module being able to declare which post-MVP features are allowed to be used in the wasm module, as a way to prevent intermediate optimization steps from mistakenly using something they shouldn't. Such a section would be stripped at the end of the compilation, though, since in theory the wasm engine can directly observe which post-MVP features are actually used in the module body for the purpose of telemetry etc.