wasmtime::component::bindgen! macro generates invalid code

[pimeys]

Test Case

In our repository, we have versioned wit shared by our host and guest code. One of the bindgen calls started generating invalid Rust code when we updated wasmtime from version 30 to version 31:

https://github.com/grafbase/grafbase/blob/renovate/wasmtime/crates/wasi-component-loader/src/extension/api/since_0_8_0/wit.rs

Then again the next version of this wit, which is just reorganized and a few things are changed in the method calls, this generates correct code:

https://github.com/grafbase/grafbase/blob/renovate/wasmtime/crates/wasi-component-loader/src/extension/api/since_0_9_0/wit.rs

Steps to Reproduce

• Clone repo https://github.com/grafbase/grafbase
• Compile the crate crates/wasi-component-loader with cargo build, it should work.
• Change the branch to renovate/wasmtime which updates wasmtime from 30 to 31
• See the bindgen macro giving an error: use of undeclared lifetime name 'a

Expected Results

Updating wasmtime should not lead to invalid Rust code.

Actual Results

If you expand the macro with an error, we see code like this generated:

pub type DirectiveSite = super::super::super::__with_name0::DirectiveSite<'a>;
pub type FieldDefinitionDirective = super::super::super::__with_name0::FieldDefinitionDirective<'a>;

The 0.9.0 wit macro does not generate code like this, nor does wasmtime 30 for the exact same macro call.

Versions and Environment

Wasmtime: 31.0.0
Operating system: Aurora Linux
Architecture: x86_64

[pimeys]

This looks like the source of the issue:

#10311

Ping @pchickey

[pchickey]

Thanks for the bug report - will look into this soon

[pimeys]

Is the fix for this included in the latest release? https://github.com/bytecodealliance/wasmtime/releases/tag/v32.0.0

[fbjork]

@pchickey is this issue fixed?

[pchickey]

Hi - sorry, this fell off my radar, no its not fixed. I will try to get to it but I'm on vacation for the second half of May so I'm working to finish other projects before then.

[alexcrichton]

@pimeys or @fbjork would one of y'all be up for reducing this down to something smaller? If you can get a reproduction that's just wasmtime + some wits + some bindgen! calls I can probably take it from there

[pimeys]

Ok, so this took me a while (been quite busy too), but here's a reproduction:

https://github.com/pimeys/wasmtime-regression-poc

You can change wasmtime to version 30 and it compiles. This is not the only lifetime error I've witnessed... There's also cases when the macro renders something like

pub type MeowMeow:  &'a __with_name0::MeowMeow;

and I've also seen

pub type MeowMeow: __with_name0::MeowMeow<'a>;

all wrong. The correct way to expand this would be:

pub type MeowMeow<'a> = __with_name0::MeowMeow<'a>;

If you want more examples, you have to dig into our big codebase and update wasmtime there. Hopefully this repro is enough though.

[alexcrichton]

Could you expand on how pub struct MeowMeow<'a> is generated. I'm assuming it's generated by a separate bindgen! invocation as opposed to hand written, but if it's hand-written that may also explain the mismatch here. In any case I'm basically curious to dig into the provenance of this definition and from whence it came.

[pimeys]

Could you expand on how pub struct MeowMeow<'a> is generated. I'm assuming it's generated by a separate bindgen! invocation as opposed to hand written, but if it's hand-written that may also explain the mismatch here. In any case I'm basically curious to dig into the provenance of this definition and from whence it came.

It is hand-written in these certain types. We try to avoid extra allocations in certain very hot paths. Therefore we have types with a lifetime parameter, allowing us only allocate once when going from the host to the guest.