Add mem_uninitialized lint to FCW lint on uses

[ghost]

This PR adds a lint that will detect uses of std::mem::uninitialized (using the mem_uninitialized diagnostic item) that are not definitely of "basic" types like int/float/raw pointer, and emit a FCW lint for it.

For generic cases, we can't say it's definitely UB, just that it might be.

Background: #98862 (comment)

[rust-highfive]

r? @wesleywiser

(rust-highfive has picked a reviewer for you, use r? to override)

[rustbot]

Hey! It looks like you've submitted a new PR for the library teams!

If this PR contains changes to any rust-lang/rust public library APIs then please comment with @rustbot label +T-libs-api -T-libs to tag it appropriately. If this PR contains changes to any unstable APIs please edit the PR description to add a link to the relevant API Change Proposal or create one if you haven't already. If you're unsure where your change falls no worries, just leave it as is and the reviewer will take a look and make a decision to forward on if necessary.

Examples of T-libs-api changes:

• Stabilizing library features
• Introducing insta-stable changes such as new implementations of existing stable traits on existing stable types
• Introducing new or changing existing unstable library APIs (excluding permanently unstable features / features without a tracking issue)
• Changing public documentation in ways that create new stability guarantees
• Changing observable runtime behavior of library APIs

[ghost]

Did some tests compiling some representative rust programs, and it's worth noting that this would cause likely a fair chunk of nontrivial programs to warn (For example, itoa 0.4.8 warns, and pre 1.0 itoa makes up a good 1/3 of the downloads at the moment).

Also the message might need adjusting, it currently says warning: the following packages contain code that will be rejected by a future version of Rust: bytes v0.4.12, itoa v0.4.8, which... isn't quite true? Well, it might panic in a future version of rust, which I guess is rejection?

I still feel a warning is really the best choice to push crates to update their dependencies / remove their use of mem::uninit, but it's also probably worth keeping in mind just how many people will see it.

[ghost]

and now that i think of it this is probably more @rustbot label -T-compiler T-lang

i think. maybe not API? Not sure!

[wesleywiser]

Nominating for Lang Team discussion: should we lint on (most) uses of std::mem::uninitalized?

[ghost]

This PR, as implemented, would be all uses, even definitely or possibly non-UB ones, since at the call-site, we might not know the type, and nearly all uses of mem::uninit are UB anyways. The best we could likely do and still be useful is to skip known non-UB calls, of which I doubt there really are any.

And yeah, I don't know if this is T-Lang or T-Libs or both, but #98862 (comment) was nominated for T-libs and consensus seemed to be that a FCW was a good idea.

[Mark-Simulacrum]

Did some tests compiling some representative rust programs, and it's worth noting that this would cause likely a fair chunk of nontrivial programs to warn.

Can we get a better understanding of the actual path most users will have to fixing this? Is it cargo update, or do they need to actually go and edit their Cargo.toml (moving to an entirely different crate or across semver versions)? And, if the majority of users do need to actually make a semver-incompatible change here, is that something we can help with (e.g., by pushing some key crates to release point releases of older versions)?

---

I am pretty wary of false-positive warnings for uses of uninitialized that aren't currently UB, particularly if we're telling a very large amount of people to take some non-trivial action (e.g., more than cargo update). I think that sets a bad precedent and may discourage people from trusting such notices in the future.

For example, looking at itoa (since you mentioned it), I believe we're warning as a result of this call. Given we've merged #99182, this is just slow but not actually language UB anymore. It's debatable whether it's library UB -- but it seems like, in general, not something we need a FCW to warn users to move off of in their dependencies: doing so would cause more noise than benefit, I think.

It seems likely that we can detect at least this kind of case (concrete types, no generics), so I would expect it to be possible to avoid linting here. Maybe the invalid_value lint could be reused here (not familiar with its exact semantics). If that buys us avoiding warnings for ~10-20% of Rust users, that's still very significant, and seems not unlikely.

[ghost]

And, if the majority of users do need to actually make a semver-incompatible change here, is that something we can help with (e.g., by pushing some key crates to release point releases of older versions)?

In the case of itoa specifically, this was brought up upstream in dtolnay/itoa#36 with the response

People getting a FCW for using an ancient version of the library sounds like a good outcome to me.

But yeah, the fix was made on a breaking change release of the library, (0.4.8 -> 1.0.0), so cargo update is not enough for direct users. Not too many libraries (on crates.io) actually use itoa directly, the majority of the downloads of pre-1.0 are from:

• csv (bumped in git to 1.0 but not released, will be a semver compatible change that nearly all of its dependencies will get)
• num-format which looks to be abandoned (no git activity for 2 years, 0.4.1 on git but not released)
• redis: bumped in git but unreleased, a large number of crates not using the latest version

Of the 1.7 million monthly downloads of itoa 0.4.8, that accounts for around 1.6 million of them.

itoa isn't the only crate that uses mem::uninitialized to make [u8; N]'s but it's the major one. http is the other one that I can think of at the moment, and that has a semver compatible fix that a lot of crates using it will get (https://lib.rs/crates/http/rev). There's a large number of 0.1.x dependents, but they have a relatively small download count.

As for more UB cases, a fair few cases of UB were found in #99389 (even where we needed to allow uninit &[u8]). It was mostly smallvec and crossbeam, the triage comment is #99389 (comment).

---

Yeah, we could use a modified form of the invalid_value lint here. Invalid_value lints about known to be UB calls, but it also lints for mem::zeroed and MaybeUninit::uninit().assume_init() (and even transmute(0), for zeroed). So for the specific case of "get people off mem::uninitialized", it wouldn't be quite appropriate. And it won't lint on any UB calls to

unsafefnuninit<T>() -> T{ mem::uninitialized()}

We could do a very targeted fix of "if it's known to be a [u8; N], ignore it" which is strange but will capture all the cases of "this is a buffer the caller didn't want to initialize", which is typically what mem::uninitialized is used for, but not ignore much else.

[ghost]

After looking at more failures (by using the https://github.com/mTvare6/hello-world.rs repo as a repo that has a lot of dependencies, so the check runs on a lot of somewhat representative code), ignoring just [u8; N] might not work, since a lot of crates will make ints or structs with all ints, with mem::uninitialized().

I'll try and write a more targeted check, as a "invalid_values but only caring about mem::uninitialized and falling back to linting, instead of ignoring". For reference, the current set of hit crates on that repo is bytes v0.4.12 dtoa v0.4.8 generic-array v0.12.4 generic-array v0.13.3 glsl-layout v0.3.2 hyper v0.13.10 itoa v0.4.8 json v0.11.15 mio v0.6.23 nalgebra v0.19.0 nix v0.14.1 parking_lot v0.4.8 parking_lot v0.9.0 parking_lot_core v0.2.14 winit v0.19.5. Most of them being for ints, as far as I could see.

[ghost]

Ignoring just known to be ints/floats/raw pointers (so a generic T will still lint) gets us generic-array v0.12.4, generic-array v0.13.3, glsl-layout v0.3.2, hyper v0.13.10, nalgebra v0.19.0, nix v0.14.1, winit v0.19.5 list of warning crates (I removed the ones that are actually for procedural-masquarade, that being rental and fluent-bundle (which uses rental)).

[RalfJung]

This is blocked on #98919.

[RalfJung]

Why is this so clever?

I think the idea is to just warn on all uses of this function. Basically this is a very strong form of deprecation. We might even do this via a new attribute rather than a specific lint -- #[extremely_deprecated] or so, which makes the existing deprecation lint into an FCW lint.

[RalfJung]

Oh, looks like this was specifically requested by @Mark-Simulacrum. However my understanding from previous discussions was that we didn't want to differentiate here -- we want to slowly work towards removing that function entirely. If we had a time machine we'd erase it from history; lacking that we will remove it as much as we possibly can.

[ghost]

Yeah, I'm not sure if we agreed on exactly when it should lint. #100342 (comment) seemed to say it's better off ignoring cases like [u8; 64] or similar.

"Basically just a lint looking for use of the mem_uninitialized diagnostic item" is what this was originally implemented as, and I can revert it back to that if we agree that's fine.

But if we wanted to reduce noise, then we could ignore things that are not UB (assuming the 0x01 filling or a hypothetical freeze).

Personally I think we've had MaybeUninit for long enough and that a full-on FCW for any mention of mem::uninitialized is warranted, but it would be noisy for some time while the ecosystem upgrades.

[RalfJung]

Looking at it again, the guidance we got from the lang team is ambiguous, so I asked for clarification in that thread.

[nikomatsakis]

@rustbot label -I-lang-nominated -T-lang +T-libs-api +I-libs-api-nominated

Given that this issue is spawned from #98862, which is tagged as libs-api, we are relabeling this to libs-api.

Our sense is that the decision of what is UB remains T-lang purview, but the decision of the stdlib surface (e.g., whether to deprecate, remove things, etc) is T-libs-api.

[ghost]

Okay, I cleaned the code up and squashed commits.

As for what we should do about things like bool: It's very simple to make them go from FCWing to not FCWing, so I'm not concerned about effort there. I'd rather us FCW for it, but I'm not going to fight too hard for doing that. (Things like &T definitely should FCW).

I suspect not FCWing for bool won't buy you very much, I'd need to check through common crates to see what they warn on, but it's presumably either references, or generic T. Neither of which I'm happy to make not warn, we all seem in agreement for references at least. (And for generic T, it usually is used as a data structure that references or worse are put in).

I'm happy to make the reference tracking issue once we're happy with the situations that this will FCW for. I'm assuming it's my job to make that issue, in any case.

As before, I'm unsure of the lint wording here, but I'll ask around to see if other people have better ideas for how to phrase it.

But now, I think this is ready to @rustbot review

[ghost]

Errors now look like:

(hyper, only the first one since there's 2 instances of the same exact error here)

The package `hyper v0.13.10` currently triggers the following future incompatibility lints:
> warning: the type `[httparse::Header;100]` does not permit being left uninitialized
> --> /home/jess/.cargo/registry/src/github.com-1ecc6299db9ec823/hyper-0.13.10/src/proto/h1/role.rs:120:77
> |
> 120 | let mut headers:[httparse::Header<'_>;MAX_HEADERS] = unsafe{ mem::uninitialized()};
> | ^^^^^^^^^^^^^^^^^^^^
> | |
> | this code causes undefined behavior when executed
> | help: use `MaybeUninit<T>` instead, and only call `assume_init` after initialization is done
> |
> = note: `#[allow(mem_uninitialized)]` on by default
> = note:for more information, see FIXME: fill this in
> note: references must be non-null(in this struct field)
> --> /home/jess/.cargo/registry/src/github.com-1ecc6299db9ec823/httparse-1.7.1/src/lib.rs:660:5
> |
> 660 | pub name:&'a str,
> | ^^^^^^^^^^^^^^^^^
> 

Old http and itoa do not lint, as all they do is make uninit [u8; N].

Old crossbeam looks like

The package `crossbeam v0.2.12` currently triggers the following future incompatibility lints:
> warning: the type `T` is generic, and might not permit being left uninitialized
> --> /home/jess/.cargo/registry/src/github.com-1ecc6299db9ec823/crossbeam-0.2.12/src/sync/ms_queue.rs:66:45
> |
> 66 | payload:Payload::Data(unsafe{ mem::uninitialized()}),
> | ^^^^^^^^^^^^^^^^^^^^
> | |
> | this code causes undefined behavior when executed
> | help: use `MaybeUninit<T>` instead, and only call `assume_init` after initialization is done
> |
> = note: `#[allow(mem_uninitialized)]` on by default
> = note:for more information, see FIXME: fill this in
> = note:type might not be allowed to be left uninitialized
> > warning: the type `[std::cell::UnsafeCell<(T, std::sync::atomic::AtomicBool)>;32]` is generic, and might not permit being left uninitialized
> --> /home/jess/.cargo/registry/src/github.com-1ecc6299db9ec823/crossbeam-0.2.12/src/sync/seg_queue.rs:40:28
> |
> 40 | data:unsafe{ mem::uninitialized()},
> | ^^^^^^^^^^^^^^^^^^^^
> | |
> | this code causes undefined behavior when executed
> | help: use `MaybeUninit<T>` instead, and only call `assume_init` after initialization is done
> |
> = note: `#[allow(mem_uninitialized)]` on by default
> = note:for more information, see FIXME: fill this in
> note:type might not be allowed to be left uninitialized(in this structfield)
> --> /home/jess/src/rust/library/core/src/cell.rs:1869:5
> |
> 1869 | value:T,
> | ^^^^^^^^
> 

[RalfJung]

So when I have something like what hyper does in my own crate, I now get two warnings -- one from invalid_values, and one from this one?

[ghost]

Yeah. It's unfortunate.

I could not emit an invalid_value warning if the mem_uninitialized warning would fire (I guess, just query the mem_uninitialized's ty_find_init_error?). Means that if you specifically silence the mem_uninitialized warning, but don't silence invalid_value, you'll get less warnings for things that both care about (But the FCW warning still applies, you'll be told to cargo report future-incompatibilities regardless of lint level). That might be an acceptable edge case.

[ghost]

I could not emit an invalid_value warning if the mem_uninitialized warning would fire

This is now implemented.

Output from doing

fnmain(){let _x:char = unsafe{ std::mem::uninitialized()};}

looks like

# cargo +source check
Checking abcdef v0.1.0(/tmp/scratchaBCBdAv6h)
warning:use of deprecated function `std::mem::uninitialized`:use `mem::MaybeUninit` instead
--> src/main.rs:2:39
|
2 | let _x: char = unsafe{ std::mem::uninitialized()};
| ^^^^^^^^^^^^^
|
= note: `#[warn(deprecated)]` on by default
warning: the type `char` does not permit being left uninitialized
--> src/main.rs:2:29
|
2 | let _x: char = unsafe{ std::mem::uninitialized()};
| ^^^^^^^^^^^^^^^^^^^^^^^^^
| |
| this code causes undefined behavior when executed
| help:use `MaybeUninit<T>` instead, and only call `assume_init` after initialization is done
|
= note: `#[warn(mem_uninitialized)]` on by default
= note:for more information, see FIXME: fill this in
= note: characters must be a valid Unicode codepoint
warning: `abcdef` (bin "abcdef") generated 2 warnings
Finished dev [unoptimized + debuginfo]target(s) in 0.07s
warning: the following packages contain code that will be rejected by a future version of Rust: abcdef v0.1.0(/tmp/scratchaBCBdAv6h)
note: to see what the problems were,use the option `--future-incompat-report`, or run `cargo report future-incompatibilities --id 8`

Might be a good idea to merge the checks for invalid_value and mem_uninitialized and pass in another argument indicating which form of the checks we're doing, since right now it is duplicated code.

[bors]

☔ The latest upstream changes (presumably #98919) made this pull request unmergeable. Please resolve the merge conflicts.

[wesleywiser]

The compiler changes look good to me but since this is T-libs-api, I'm going to hand it off to them at this time to make the final review of what should/should not trigger the lint.

r? rust-lang/libs

[thomcc]

I'm in favor of a FCW warning here, but will punt to someone on libs-api.

r? @joshtriplett

[ghost]

Created #101570 as a tracking issue for the FCW, and put that in the code.

[bors]

☔ The latest upstream changes (presumably #102377) made this pull request unmergeable. Please resolve the merge conflicts.

[Noratrieb]

any updates here? @joshtriplett @5225225

[Noratrieb]

I don't think it's a good idea to add an FCW for mem::uninitialized now. The danger of the function has been diffused, it just returns 0x01 now. Programs using it are completely fine (though they are a little bit slower than they originally anticipated :D).
That said, we should FCW on invalid uses of MaybeUninit::uninit().assume_init() because that one is actual UB and should be fixed in code. #100423 tries to make it panic, but imo we should FCW first before making it panic.

[ghost]

Makes sense. I'll adjust the code to instead lint for that.

[ghost]

Cleaning up old PRs that I won't get to, don't have the time. Happy to help anyone pick this up again, though.