add a lint against references to packed fields

[RalfJung]

Creating a reference to an insufficiently aligned packed field is UB and should be disallowed, both inside and outside of unsafe blocks. However, currently there is no stable alternative (#64490) so all we do right now is have a future incompatibility warning when doing this outside unsafe (#46043).

This adds an allow-by-default lint. @retep998 suggested this can help early adopters avoid issues. It also means we can then do a crater run where this is deny-by-default as suggested by @joshtriplett.

I guess the main thing to bikeshed is the lint name. I am not particularly happy with "packed_references" as it sounds like the packed field has reference type. I chose this because it is similar to "safe_packed_borrows". What about "reference_to_packed" or "unaligned_reference" or so?

[rust-highfive]

r? @varkor

(rust_highfive has picked a reviewer for you, use r? to override)

[retep998]

Another question is whether this should lint against any reference to a field of a struct that has any level of packing, or if it should only lint when a reference is created to a field where the field's alignment is strictly greater than the struct's alignment? Although this question also applies to the safe_packed_borrows future incompatibility warning.

[RalfJung]

It already does the smart thing here, as the aligned accesses in the test show. This is also what safe_packed_borrows does.

[RalfJung]

Well, so far the "smartness" is that it never complains about references to 1-aligned types. There might be more we can do, but that seems rather orthogonal to the linting part.

[oli-obk]

I don't even know what this means... I think this is a leftover from the pre-miri MIR based const evaluator. I think we could just merge the entire mir_const query into mir_validated

[RalfJung]

Hm, maybe we can, I'll leave that to someone else.^^

If you change this, please remember to update the rustc-dev-guide, which is where I learned that this is not just for consts. :)

[RalfJung]

Looks like an alternative to this would be to add the lint here:

rust/src/librustc_mir_build/build/mod.rs

Line 184 in 31add7e

lints::check(tcx, &body, def_id);

[oli-obk]

Our lint naming rules (https://rust-lang.github.io/rfcs/0344-conventions-galore.html#lints) say that we should have the "bad thing" in the name and use a plural, so unaligned_references?

[RalfJung]

Well this does not detect all unaligned references... references_to_packed_field is a mouthful though.

[RalfJung]

(I rebased so I wouldn't have to rebuild llvm to work on this again.)

[RalfJung]

@oli-obk I ended up renaming the lint as you suggested. False negatives are okay for lints after all (and this lint will hopefully eventually become a hard error anyway -- to fix a soundness issue -- at which point it stops having a name).

[oli-obk]

@bors r+

[bors]

📌 Commit d959a8f has been approved by oli-obk