rustdoc incorrectly evaluates and fails due to panic in zerocopy

[Mark-Simulacrum]

This code:

extern crate zerocopy;

pub use std::cell::RefMut as LockGuard;

with Cargo.toml:

[package]
name = "test-rustdoc-bug"
version = "0.1.0"
edition = "2021"

[dependencies]
zerocopy = "0.7.32"

produces the following error:

$ RUSTDOCFLAGS="-Zunstable-options --document-hidden-items" cargo +nightly doc
 Documenting test-rustdoc-bug v0.1.0 (/Users/mark/test-rustdoc-bug)
error[E0080]: evaluation of `<std::cell::RefMut<'_, [u8]> as zerocopy::ByteSlice>::INTO_REF_INTO_MUT_ARE_SOUND` failed
    --> /Users/mark/.cargo/registry/src/index.crates.io-6f17d22bba15001f/zerocopy-0.7.32/src/lib.rs:5238:9
     |
5238 |         panic!("Ref::into_ref and Ref::into_mut are unsound when used with core::cell::RefMut; see https://github.com/google/zerocopy/issues/716")
     |         ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ the evaluated program panicked at 'Ref::into_ref and Ref::into_mut are unsound when used with core::cell::RefMut; see https://github.com/google/zerocopy/issues/716', /Users/mark/.cargo/registry/src/index.crates.io-6f17d22bba15001f/zerocopy-0.7.32/src/lib.rs:5238:9
     |
     = note: this error originates in the macro `$crate::panic::panic_2015` which comes from the expansion of the macro `panic` (in Nightly builds, run with -Z macro-backtrace for more info)

For more information about this error, try `rustc --explain E0080`.
error: could not document `test-rustdoc-bug`

This seems to happen on all(?) of current stable (1.77), beta (1.78), and nightly (1.79, 3c85e56 2024-03-18) but I'm suspicious that there's some kind of new breakage here in 1.78, because this broke rust-lang/rust's CI on similar code (in rustc_data_structures): #122754 (comment) while performing a bump of the bootstrap compiler from 1.77 to 1.78. This minimal example appears to reproduce across all versions. Tagging this as a regression for now despite lacking a minimal reproducer demonstrating that.

I think this is likely a bug of some kind in rustdoc regardless, because presumably it's not intentional that we should fail to document (benign-looking) code like this.

[Mark-Simulacrum]

Mark-Simulacrum@eb7c254 works locally to resolve this bug, probably by just avoiding the relevant code path in rustdoc.

[GuillaumeGomez]

This is strange. Tagging the compiler team as well.

[apiraino]

I've tried bisecting this error, but results are perhaps a bit iffy. Does any of these commit ring a bell?

Regression in nightly-2023-07-19
converted 2023-07-19 to 903e279

  commit[0] 2023-07-17: Auto merge of #89132 - Cyborus04:rc_allocator_support, r=Amanieu
  commit[1] 2023-07-18: Auto merge of #113061 - Amanieu:x86_64-ohos, r=compiler-errors
  commit[2] 2023-07-18: Auto merge of #113574 - GuillaumeGomez:rustdoc-json-strip-hidden-impl, r=aDotInTheVoid,notriddle
  commit[3] 2023-07-18: Auto merge of #113801 - compiler-errors:iter-instantiated, r=oli-obk
  commit[4] 2023-07-18: Auto merge of #113659 - ericmarkmartin:smir-refs-and-ptrs, r=spastorino
  commit[5] 2023-07-18: Auto merge of #113677 - bryangarza:unevaluated-const-ice_issue-110892, r=davidtwco
  commit[6] 2023-07-18: Auto merge of #113706 - Alexendoo:compiletest-backslash-re, r=oli-obk
  commit[7] 2023-07-18: Auto merge of #112374 - chenx97:better-mips64r6, r=jackh726
  commit[8] 2023-07-18: Auto merge of #113837 - matthiaskrgr:rollup-v4xud4s, r=matthiaskrgr
  commit[9] 2023-07-18: Auto merge of #113636 - compiler-errors:opaque-recursive-check-bad, r=oli-obk
  commit[10] 2023-07-18: Auto merge of #113841 - weihanglo:update-cargo, r=weihanglo

[Mark-Simulacrum]

This is hitting the 1.79 promotion too. Probably the fix in Mark-Simulacrum@eb7c254 wasn't quite enough -- that did land on master, so it's not entirely missing.

[GuillaumeGomez]

With everything going on in rustc_codegen_gcc, didn't have time to take a look yet...

[GuillaumeGomez]

Took a look at it and it seems normal. Can be reproduced with this foo crate as dependency (no need to have a current crate, just using foo as a dependency is enough):

[package]
name = "foo"
version = "0.1.0"
edition = "2021"

src/lib.rs:

pub const YEP: bool = if !cfg!(doc) {
    panic!("bad");
} else {
    false
};

The problem is that the dependencies are built with rustc before being built with rustdoc (to get crate information). So unless zerocopy is the final crate, this condition will always at some point be true. Problem seems to be solved in newest zerocopy versions so I think we can close this issue.