Document MaybeUninit bit validity

[joshlf]

Partially addresses rust-lang/unsafe-code-guidelines#555 by clarifying that it is sound to write any byte values (initialized or uninitialized) to any MaybeUninit<T> regardless of T.

r? @RalfJung

[RalfJung]

Cc @rust-lang/opsem

[RalfJung]

I don't think it makes sense to say that a type "contains initialized bytes" at some offset. That's a property of a representation.

The typical term for representation bytes that are lost here is "padding". I don't think we have rigorously defined padding anywhere yet, but the term is sufficiently widely-used (and generally with a consistent meaning) that we may just be able to use it here?

[joshlf]

IIUC, you're making two points:

• We should speak about a type's representation containing bytes, not about the type itself containing bytes
• In a representation, we should speak about padding bytes rather than uninitialized bytes

Is that right?

One thing that's probably worth distinguishing here is between values and layouts. In my mental model, an uninit byte is one of the possible values that a byte can have (e.g., it's the 257th value that can legally appear in a MaybeUninit<u8>). By contrast, padding is a property of a layout - namely, it's a sequence of bytes in a type's layout that happen to have the validity [MaybeUninit<u8>; PADDING_LEN].

Based on this, maybe it's best to say:

If byte offsets exists at which T's representation does not permit uninitialized bytes but U's representation does (e.g. due to padding), then the bytes in T at these offsets may not be preserved in u, and so transmute(u) may produce a T with uninitialized bytes at these offsets. This is an active area of discussion, and this code may become sound in the future.

[RalfJung]

Is that right?

No. I think both of the following concepts make sense:

• The representation of a particular value at a particular type contains uninitialized bytes.
• A type contains padding bytes. (These are bytes which are always ignored by the representation relation.)

But it makes less sense to talk about padding of a representation, or to talk about uninitialized bytes in a type.

So for this PR, the two key points (and they are separate points) are:

• If U has padding, those bytes may be reset to "uninitialized" as part of the round-trip. If those same bytes are not padding in T, this can therefore mean some of the information of the original T value is lost.
• If T does not permit uninitialized bytes on those positions, the round-trip is UB.

The second point is just a logical consequence of the first, it does not add any new information. Not sure if it is worth mentioning.

[joshlf]

• The representation of a particular value at a particular type contains uninitialized bytes.
• A type contains padding bytes. (These are bytes which are always ignored by the representation relation.)

Does this imply that a type contains padding bytes, not a type's representation?

I'm thinking through the implications of what you said, and I think I understand something new that I didn't before, and I want to run it by you: In my existing mental model, a padding byte is a location in a type's layout such that every byte value at that location (including uninit) is valid (enums complicate this model, but I don't think that complication is relevant for this discussion - we can just stick to thinking about structs). The problem with this mental model is that, interpreted naively, it implies that different byte values in a padding byte could correspond to different logical values of the type. So e.g. in the type #[repr(C)] struct T(u8, u16), [0, 0, 0, 0] and [0, 1, 0, 0] would correspond to different values of the type since we're treating the padding byte itself as part of the representation relation. Of course, that is not something we want.

IIUC, by contrast your model is that the representation relation simply doesn't include padding bytes at all. So it'd be more accurate to describe the representation of T as consisting of three bytes - at offsets 0, 2, and 3. Every representation of T has a "hole" at offset 1 which is not part of the representation. This ensures that there's a 1:1 mapping between logical values and representations. Is that right?

[RalfJung]

Does this imply that a type contains padding bytes, not a type's representation?

That's how I think about it. We can't tell which byte is a padding byte by looking at one representation -- it's a property of the type.

In my existing mental model, a padding byte is a location in a type's layout such that every byte value at that location (including uninit) is valid

That would make the only byte of MaybeUninit<u8> a padding byte, so I don't think this is the right definition.
That's why I said above: a padding byte is a byte that is ignored by the representation relation. Slightly more formally: if r is some representation valid for type T, and r' is equal to r everywhere except for padding bytes, then r and r' represent the same value.

So it'd be more accurate to describe the representation of T as consisting of three bytes

The representation has 4 bytes. But only 3 of them actually affect the represented value (which is a tuple of two [mathematical] integers).

---

We seem to be using the term "representation" slightly differently. For me, that's list a List<Byte> of appropriate length. You may be using that term to refer to what I call "representation relation"?

[joshlf]

We seem to be using the term "representation" slightly differently. For me, that's list a List<Byte> of appropriate length. You may be using that term to refer to what I call "representation relation"?

That's helpful, thank you!

To avoid rabbit holing too much on the definitions (although it's interesting and useful – just maybe a bit of a distraction here), maybe you could propose language you'd prefer to see in place of what I've written here?

[RalfJung]

@rustbot ready

[RalfJung]

@rustbot author

[rustbot]

Reminder, once the PR becomes ready for a review, use @rustbot ready.

[RalfJung]

I raised the question on Zulip whether it is wise to make a guarantee here that isn't, strictly speaking, documented in the LLVM LangRef. Nikita says he thinks that that's fine -- we may have to adjust how exactly we compile MaybeUninit in the future, but LLVM currently intends do support this case in a somewhat roundabout and incomplete way that seems to work well enough in practice, and LLVM can't more aggressively exploit the fuzziness along the edges of that approximation until a proper alternative exists.

[karolzwolak]

Thanks for your contribution @joshlf from wg-triage.
Could you address the comments above?

[joshlf]

Thanks for your contribution @joshlf from wg-triage. Could you address the comments above?

I likely won't have time to move this forward until mid-September or October, but I'll follow up at that point.

[joshlf]

Thanks for your contribution @joshlf from wg-triage. Could you address the comments above?

I likely won't have time to move this forward until mid-September or October, but I'll follow up at that point.

Nvm, found some time 🙂

I've responded to various comment threads.

[karolzwolak]

Awesome, you should also rebase your changes onto master, and do @rustbot ready when you're ready for review.

[joshlf]

@rustbot ready

[traviscross]

Checking in on checkboxes...

cc @rust-lang/opsem

[rust-rfcbot]

🔔 This is now entering its final comment period, as per the review above. 🔔

[rust-rfcbot]

The final comment period, with a disposition to merge, as per the review above, is now complete.

As the automated representative of the governance process, I would like to thank the author for their work and everyone else who contributed.

This will be merged soon.

[RalfJung]

Thanks @joshlf for the patience. :)
@bors r+ rollup

[bors]

📌 Commit 11627f0 has been approved by RalfJung

It is now in the queue for this repository.