Tons of `unexpected_cfgs` failures in portable-simd and backtrace doctests

[RalfJung]

Running std doctests in Miri results in tons of errors since today's nightly:

  0.000010   ---- library/core/src/../../portable-simd/crates/core_simd/src/vector.rs - core_simd::vector::Simd<T,N>::store_select (line 640) stdout ----
  0.000008   error: unexpected `cfg` condition value: `as_crate`
  0.000010    --> library/core/src/../../portable-simd/crates/core_simd/src/vector.rs:642:7
  0.000008     |
  0.000011   5 | #[cfg(feature = "as_crate")] use core_simd::simd;
  0.000010     |       ^^^^^^^^^^^^^^^^^^^^
  0.000011     |
  0.000013     = note: expected values for `feature` are: `debug_refcell`, `panic_immediate_abort`
  0.000007     = help: consider adding `as_crate` as a feature in `Cargo.toml`
  0.000011     = note: see <https://doc.rust-lang.org/nightly/cargo/reference/build-scripts.html#rustc-check-cfg> for more information about checking conditional configuration
  0.000007   note: the lint level is defined here
  0.000010    --> library/core/src/../../portable-simd/crates/core_simd/src/vector.rs:638:9
  0.000008     |
  0.000010   1 | #![deny(warnings)]
  0.000008     |         ^^^^^^^^
  0.000010     = note: `#[deny(unexpected_cfgs)]` implied by `#[deny(warnings)]`
  0.000010   
  0.000010   error: unexpected `cfg` condition value: `as_crate`
  0.000008    --> library/core/src/../../portable-simd/crates/core_simd/src/vector.rs:643:11
  0.000010     |
  0.000008   6 | #[cfg(not(feature = "as_crate"))] use core::simd;
  0.000010     |           ^^^^^^^^^^^^^^^^^^^^
  0.000008     |
  0.000010     = note: expected values for `feature` are: `debug_refcell`, `panic_immediate_abort`
  0.000008     = help: consider adding `as_crate` as a feature in `Cargo.toml`
  0.000011     = note: see <https://doc.rust-lang.org/nightly/cargo/reference/build-scripts.html#rustc-check-cfg> for more information about checking conditional configuration
  0.000009   
  0.000010   error: aborting due to 2 previous errors
  0.000008   
  0.000010   Couldn't compile the test.

Interestingly, the in-tree test suite is working, so there's likely some special bootstrap magic somewhere that out-of-tree tests do not benefit from?

cc @Urgau

[RalfJung]

These files are from portable-simd, which means they are sometimes their own crate and sometimes a module imported into core. There is indeed no as_crate feature in core, so no idea how these tests are passing in our CI.

Do we inject a -A unexpected_cfgs somewhere for doctests, or something like that?

[RalfJung]

There's also one in std, from the backtrace module, which similarly is also sometimes a crate and sometimes a module:

  0.000008   ---- library/std/src/../../backtrace/src/lib.rs - backtrace_rs (line 20) stdout ----
  0.000007   error: unexpected `cfg` condition value: `std`
  0.000007    --> library/std/src/../../backtrace/src/lib.rs:23:7
  0.000008     |
  0.000007   5 | #[cfg(feature = "std")] {
  0.000008     |       ^^^^^^^^^^^^^^^
  0.000007     |
  0.000008     = note: expected values for `feature` are: `addr2line`, `backtrace`, `compiler-builtins-c`, `compiler-builtins-mangled-names`, `compiler-builtins-mem`, `compiler-builtins-no-asm`, `compiler-builtins-weak-intrinsics`, `llvm-libunwind`, `miniz_oxide`, `object`, `panic-unwind`, `panic_immediate_abort`, `panic_unwind`, `profiler`, `profiler_builtins`, `std_detect_dlsym_getauxval`, `std_detect_env_override`, `std_detect_file_io`, `system-llvm-libunwind`
  0.000008     = help: consider adding `std` as a feature in `Cargo.toml`
  0.000007     = note: see <https://doc.rust-lang.org/nightly/cargo/reference/build-scripts.html#rustc-check-cfg> for more information about checking conditional configuration
  0.000007   note: the lint level is defined here
  0.000007    --> library/std/src/../../backtrace/src/lib.rs:19:9
  0.000007     |
  0.000007   1 | #![deny(warnings)]
  0.000007     |         ^^^^^^^^
  0.000007     = note: `#[deny(unexpected_cfgs)]` implied by `#[deny(warnings)]`
  0.000008   
  0.000007   error: aborting due to 1 previous error
  0.000007   
  0.000007   Couldn't compile the test.

[workingjubilee]

Huh.

[Kobzol]

It's quite possible that this is caused by the CI mess. One of the insta-merged PRs was a miri sync.

[RalfJung]

It's possible but seems unlikely to me. I guess we'll have to see.

[Urgau]

Interestingly, the in-tree test suite is working, so there's likely some special bootstrap magic somewhere that out-of-tree tests do not benefit from?

Yes, there is, https://github.com/rust-lang/rust/blob/69ffc0d3a3c619009bcb27b8f61d810e27b12612/src/bootstrap/src/core/builder.rs#L1545-L1555

I would suggest doing the same and adding as part of the build --check-cfg=cfg(feature,values(any())), possibly in RUSTFLAGS or somewhere else.

[RalfJung]

That seems to help, thanks. :)

[Urgau]

Another thing we could do now that check-cfg is stable, is adding this exception directly in the build.rs of core, alloc and std with the Cargo instruction cargo::rustc-check-cfg.

[RalfJung]

If you end up doing that, please ping me so I can revert https://github.com/rust-lang/miri-test-libstd/pull/66.

[Urgau]

@RalfJung I don't think allowing all values for the feature cfg is going to be enough. There are many other custom cfgs used in core/alloc/std, https://github.com/rust-lang/rust/blob/69ffc0d3a3c619009bcb27b8f61d810e27b12612/src/bootstrap/src/lib.rs#L87-L97

In fact I can see them in the latest CI run under the "Setup environment". If you want to completely silence them, allowing the lint (-Aunexpected_cfgs in RUSTFLAGS) is probably (in the short term) the best course of action.

[RalfJung]

This was enough though to make the tests pass again... :shrug:

[RalfJung]

Oh, they are shown as warning, not hard errors.

Sure, I can also allow the lint entirely.

[RalfJung]

Actually, those warnings come from the sysroot build, not the local test build. So this will need a patch in https://github.com/RalfJung/rustc-build-sysroot/ . Also, -Aunexpected_cfgs does not help for the doctests, for some reason.

(FWIW, debugging rustc invocations became a lot harder when they all grew to more than twice their size due to all these --check-cfg flags. If there's any way to hide these flags e.g. in argument files or so, that would be great...)