rustdoc should skip unresolvable private members

[asomers]

When using the doc_cfg feature it's useful to run rustdoc on structures that aren't defined for the running platform. However, compilation will fail if those structures have private members that also aren't defined on the running platform. The only solution is to laboriously #[cfg] gate each private member, even if the struct itself is already gated. This issue severely restricts the usefulness of doc_cfg. Here's an example:

#[cfg(any(target_os = "freebsd", rustdoc))]
struct NewType {
    x: libc::something_freebsd_specific
}

rustdoc will fail unless you add another #[cfg(target_os = "freebsd")] attribute directly on x. It should be possible for rustdoc to simply ignore x, since private members aren't normally documented anyway.

For a real-world example, see https://github.com/nix-rust/nix/pull/1019/files#diff-78f451c970a63ad3a93cab0d2670b05dR12 . That PR only builds because of the extra #[cfg()] gate on the libc::ucontext variable, even though the entire module is already restricted to Linux.

doc_cfg tracking issue: #43781

[Centril]

cc @QuietMisdreavus

[QuietMisdreavus]

(@Centril please ping the whole rustdoc team for rustdoc issues @rust-lang/rustdoc)

It's possible to extend everybody_loops (or create a new pretty-print pass and run it after everybody_loops when rustdoc is running) to scrub out private struct/variant members. (cc @pnkfelix) One problem with this approach is that we'd have to only run that when --document-private-items was not passed - otherwise we still have to display that information somehow.

[steveklabnik]

Triage: no change I'm aware of

[jyn514]

It's possible to extend everybody_loops (or create a new pretty-print pass and run it after everybody_loops when rustdoc is running) to scrub out private struct/variant members. (cc @pnkfelix) One problem with this approach is that we'd have to only run that when --document-private-items was not passed - otherwise we still have to display that information somehow.

everybody_loops is no longer being run, so I don't think we should reuse that.

One possible way is to special case this in rustc_resolve. However there are already a lot of hacks in resolve and I'm not terribly comfortable adding that. Maybe instead rustdoc could remove private fields of structs in a new AST pass? It would have to keep track of how many were reviewed so it could add /* Some private fields hidden */ in the documentation.