v0.3.74 fails to build on Windows

[KapJI]

Here are the errors: backtrace-errors.txt

We are building project with buck2 but third party dependencies are resolved and vendored by cargo.

Is windows-sys dependency wrong?

v0.3.73 and below work fine. Other platforms seems to be ok as well.

[ChrisDenton]

We don't use the windows-sys crate, we use windows-bindgen to pre-generate our own windows_sys.rs module. We do have a dependency on windows-targets though.

The errors look like they come from our code that tries to validate that the APIs are correct (aisde: this code is probably redundant now we're not hand-writing anything). This is failing for you for some reason but not for our CI (or in the rust-lang/rust repo). My guess would be you're using the windows raw dylib environment variable and that's causing a difference somewhere.

[KapJI]

I can bisect it to specific commit tomorrow.

[ChrisDenton]

Following up on my hunch it seems that the windows-targets crate using extern "C" for all ABIs on 64-bit systems when windows_raw_dylib is used https://github.com/microsoft/windows-rs/blob/a8e14c75581bb89fc244ad480f98567be972aea2/crates/libs/targets/src/lib.rs#L23C9-L23C10 which would account for the ABI mismatch assertions failing. I seem to recall there was a reason for that, but I can't quite remember. I guess it's because all the x86 ABIs just resolve to extern "C" on x64 and arm64.

[KapJI]

We add windows_raw_dylib to windows-targets indeed. But it was there before that as well and backtrace build was passing. What has changed with the latest release? For builds we have regular windows-latest Github runner.

[KapJI]

Bisected to #653 which added dependency of windows-targets.

[dtolnay]

Non-Buck repro, as of the current master branch of this repo (4f3acf7):

rustup target add x86_64-pc-windows-msvc
RUSTFLAGS=--cfg=windows_raw_dylib cargo check --target x86_64-pc-windows-msvc

[ChrisDenton]

I seem to recall there was a reason for that,

Ah jogged my memory: this was because of microsoft/windows-rs#2458. Varargs weren't being correctly handled by rustc so this was a workaround. It's been fixed in rustc now but the MSRV of windows-related crates means it's unlikely to revert the workaround any time soon. So best thing to do would be to update our validation or remove them entirely if nobody can think of a reason why they're still needed.