Miri doesn't respect `target-feature` in doc tests

[Kixunil]

When code under miri reaches a function annotated with #[target_feature(enable = "...")] in doc tests miri throws an error saying "calling a function that requires unavailable target features" despite this being in the if block that checks the features using is_x86_feature_enabled! and despite having -C target-feature=... on the command line. The same issue is not present in non-doc tests and the same code works fine.

[RalfJung]

Are you locally using the exact same Miri version as CI?

[Kixunil]

I guess e-mail didn't show you my edit, I found it's actually doc tests that have this problem (on both machines).

[RalfJung]

How are you setting -Ctarget-feature? If you are using RUSTFLAGS, then you need to also set RUSTDOCFLAGS for doctests. MIRIFLAGS should apply everywhere... hopefully...

Please show the full command you are using. It's hard to guess that and details matter. :)

[Kixunil]

you need to also set RUSTDOCFLAGS

Oh, that helped, thanks!

[RalfJung]

FWIW that's a cargo thing, RUSTFLAGS is simply not set on doctests.

[RalfJung]

Reading rust-lang/rust#67533 I actually don't understand how setting RUSTDOCFLAGS helps... if rustdoc doesn't pass the flags to the rustc it invokes, then it shouldn't help with -Ctarget-features either?

[RalfJung]

I guess rustdoc must be forwarding -C flags to the rustc it invokes, but not other flags like --cap-lints.

[Kixunil]

Not sure which flags rustdoc passes but FTR I set both RUSTFLAGS and RUSTDOCFLAGS in CI to configure target-features and it works fine.