Cell should be Copy

[pythonesque]

I'm not really sure why it isn't. It can only ever hold Copy data and Cell::new(some_cell.get()) has the same effect as a memcpy, so the effect of not making it Copy is mostly to just make it annoying to use types with Cells in them (since you can't #[derive(Copy)] on them, use them with the [Foo ; n] instantiation syntax in fixed-size arrays, put them in other Cells, etc.).

[huonw]

I think Cell and RefCell are explicitly non-Copy as a sort-of lint against accidentally not sharing them, since they're usually meant to be used for shared data. E.g. with Cell: Copy

let x = Cell::new(1);

Thread::scoped(move || {
    x.set(2);
}).join();

println!("{}", x.get());

won't do what one might be hoping.

[reem]

The lack of a [Foo; n] initializer when using Cell's is kind of a pain, but I agree with @huonw that it enables a lot of unintentional non-sharing.

[pythonesque]

@huonw I would rather not use Copy as a lint. Not being able to use Celllike this is just irritating ergonomically. This clearly isn't a memory safety concern and I don't think it enables unintentional sharing more than, say, [U8; 3]. It's not the absolute end of the world if it doesn't for my own data structures (since I can just make my own version of Cell) but it means I can't put many other data structures inside a Cell, which actually affects program functionality. This is why I wanted "can copy" to be a different trait from "pod", BTW.

[ranma42]

Does the same rationale from #20199 (comment) apply here?

[Gankra]

Yeah, I think we've settled on a "tricky things shouldn't be Copy" convention here. It's back-compat to add later, and right now the consequences are troubling.

[pythonesque]

I don't see what's "troubling" about the consequences here, can someone clarify? The only substantive thing I've heard is the concern that someone might accidentally not share memory, which can already happen with any other Copy type. It's neither a memory safety concern nor a security concern (like it is in the linked bug). And how often do you use a bare Cell (outside of another struct)? With Copy being opt-in I don't think this is a real concern outside of toy examples.

[huonw]

I don't personally have a huge opinion (other than the back-compat consideration), but Cell does have a fundamental difference to most other Copy types in that its essentially only purpose is in the context of shared memory.

It seems like this, and the considerations in #20199 (same thing applying to RNGs), would be entirely resolved by a extra trait, say trait MemcpyIsSafe { fn memcpy(&self) -> Self { /* do the memcpy */ } }, where a type still moves if it implements MemcpyIsSafe, but is known to be safe to use with Cell. (Copy could then inherit from MemcpyIsSafe and would only remove the move-by-default aspect.)

(The complexity may not be worth it.)

[pythonesque]

Sure, we could abbreviate MemcpyIsSafe as Pod. I don't think it would be that much more complex, since most of the time it would be the same thing, but since we haven't done it from the beginning it would be hard to add backwards-compatibly (how do you decide whether an existing potentially Copy type that doesn't #[derive(Copy)] should be Pod?).

[Gankra]

Whoops, sorry I keep ADD'ing away every time I start writing a comment here. You know, like, re-installing my entire OS. 🐫

Basically what @huonw said. In my mind Cell is basically a container or smaht-pointer type. It's not really "plain old data" (it contains an UnsafeCell!). But mostly I think we just want to be conservative here.

[pythonesque]

But it is safe, regardless of how it exists in your, my, or anyone's mind. It's only ever allowed to contain plain old data. The fact that it contains UnsafeCell is sort of immaterial, I feel. There's no point in being conservative for the sake of it--what's going to change in the future that makes it more attractive to make Cell Copy than it already is?

Your UnsafeCell point is good though, because I just checked and it's not #[derive(Copy)] either! This actually affects me, because it means I can't make a type like Cell that is Copy. Should I open a separate ticket for this? I'd rather not have to entirely abandon the standard library just to get a Copy Cell.

[Gankra]

CC @pcwalton (you're captain OIBIT, right?)

[Gankra]

I think this can be a general "what to do about iffy cases and Cell" meta-issue. Tracking it on multiple issues doesn't seem super helpful at this point.

[DiamondLovesYou]

This should be reopened. There are legitimate uses of Cell for caching in Copy structures. Despite what the Rust documentation leads us to believe, Cell should be thought as a tool to allow interior mutability, not whatever "shared mutable containers" means, since neither are box types (ie basically a pointer) so can't be shared by value anyway, outside of references.

I apologize, I know this is a pretty old issue, but the current behavior is surprising. At the very least UnsafeCell needs to be copy so one can at least bypass this issue using that structure.

[Mark-Simulacrum]

I think making UnsafeCell implement Copy would need an RFC at the least, though it seems surface-level potentially reasonable.