Copy clone semantics

[strega-nil]

Specifying that <T as Clone>::clone(&t) where T: Copy should be equivalent to ptr::read(&t).

[strega-nil]

Note that these comments were originally line comments, but that fact seems to have been lost by github.

[briansmith]

I originally thought this was a good idea, and even suggested something very much like this on the internals mailing list, but I've changed my mind. Every type that implements Copy and Clone should derive Clone and the derived implementation of Clone should be implemented in terms of Copy. Further, the optimizer should be able to notice that a hierarchy of copies is equivalent to a single flattened copy and optimize it to a memcpy. If the optimizer doesn't do that, then it is going to generate terrible code for lots of common cases that aren't calls to Clone::clone such as record updates that modify a single field.

[strega-nil]

@briansmith The issue isn't derived Clone implementations, it's that we allow you to not derive Clone. If your T: Copy, and you manually implement Clone, this would make it invalid to not make Clone::clone(&t) and ptr::read(&t) equivalent.

The optimizer is also not turned on in debug mode. Allowing us to turn Vec::clone into a memcpy, even in debug mode, would make for much faster code.

Edit: not invalid like undefined behavior or anything, just to be clear.

[strega-nil]

Note, we've already made breaking changes based on this assumption.

The change in #[derive(Copy, Clone)] to fn clone(&self) -> Self { *self } would break any code that relies on the fact that each of the inner Clone::clones were called.

[briansmith]

it's that we allow you to not derive Clone

IMO, it is OK to say that if you want Clone to be fast for your Copy type, you need to derive Clone.

The change in #[derive(Copy, Clone)] to fn clone(&self) -> Self { *self } would break any code that relies on the fact that each of the inner Clone::clones were called.

Yes, but IMO that was also a short-term mistake. Once the optimizer is doing the right thing, that shortcut won't be necessary and we can return to the simpler semantics.

In particular, the optimizer can already notice that a loop is equivalent to memcpy. It just needs to get smarter to recognize non-loop equivalents, including recursive invocations of cone/copy, are equivalent to memcpy are equivalent to memcpy too.

Consider let x = y { u32_field: new_value }. How would your proposed optimization improve this situation? It seems it wouldn't, but this kind of pattern is also important to optimize to memcpy + 32-bit memory write.

[strega-nil]

@briansmith This isn't an optimization. This is specification. It's saying that the standard library (and anyone else) may rely on the fact that a properly implemented Clone::clone for T: Copy == ptr::read. This allows for optimization through specialization, for example, the #[derive(Copy, Clone)] thing, and many other optimizations based around memcpy.

[briansmith]

Yes, but I'm saying that if the optimizer is working well, then you don't need to specify anything special here. It would be useful to see an example where this rule is needed that can't be done without the rule by simply improving the optimizer.

[Amanieu]

The RFC should specify what the consequences are for types that violate this rule. In particular this must not result in memory unsafety.

[strega-nil]

@briansmith The issue isn't where the optimizer is working. It's where it isn't. Optimization is not always on, like in debug mode. Also, building a language for a sufficiently smart compiler doesn't seem like the correct way of doing things. Often we will have to specifically specialize.

[strega-nil]

@Amanieu yes, that is the intention, I will write it down.

[briansmith]

Why not specify that even in debug mode, the copy coalescing optimizations are done? Then there is no semantic change to the language needed and debug build results are faster.

[strega-nil]

@briansmith The semantics would still change, and then you would have different semantics between debug and release.

[briansmith]

To be clear, there's nothing that would change semantically in my definition. The only thing that would change is that debug builds would be specified to have an optimization pass enabled, for at least Clone::clone that they currently don't have enabled, and that optimization pass would be made smarter. The language definition would stay exactly the same.

[strega-nil]

@briansmith Ah. You can't really do that. The necessary optimizations to realize that:

assert!(src.len() == dst.len());
let len = dst.len();
let src = &src[..len];
for i in 0..len {
    self[i].clone_from(&src[i]);
}

is equivalent to

assert!(src.len() == dst.len());
memcpy(src.as_ptr(), dst.as_mut_ptr(), dst.len());

is... most of them.

[petrochenkov]

Alternative wording: "When Copy type has user-defined implementation of Clone it's unspecified whether clone or memcpy is called when cloning a value of this type." or (shamelessly stolen from somewhere) "<...if Copy...> An implementation is allowed to omit the call to clone even if is has side effects. Such elision of calls to clone is called clone elision."

[strega-nil]

@petrochenkov it's not that it's unspecified whether clone or ptr::read is called when you actually call .clone(). It's just that libraries (including the standard library) are able to assume that Clone::clone == ptr::read for T: Copy; and, if that requirement is not met, libraries are allowed to exhibit unexpected (but defined) behavior. (Just like if an Ord type doesn't actually have an absolute ordering, or if an Eq type doesn't actually have transitive equality).

[petrochenkov]

it's not that it's unspecified whether clone or ptr::read is called when you actually call .clone()

Why not? If we go this route anyway, let's give the compiler/libraries all the freedom.
I'd even write "user-defined implementations of Clone for Copy types are ignored", but I'm not sure about generics:

fn f<T: Clone>(arg: T) -> T { arg.clone() } // It's not known if `T` is `Copy` or not and I'm not sure it's convenient to turn `clone`s into `memcpy`s during monomorphisation.

[petrochenkov]

It's just that libraries (including the standard library) are able to assume that Clone::clone == ptr::read for T: Copy

derive is not a library, so the language already can omit clones as well.

[petrochenkov]

Also, "clone is always ignored" is better than "clone is sometimes ignored" from pedagogical point of view.

[strega-nil]

hmm...

At first I thought this was a bad idea, but then I started thinking. Not "always ignored", but if your clone has different semantics from a ptr::read then it should be implementation defined whether it is called.

[arielb1]

I really don't like the "optimizer magic overriding of clone".

[strega-nil]

@arielb1 Yeah, I guess. I do think it would be better if it's just kept to libraries and language constructs like #[derive]

[glaebhoerl]

Is there any reason to write a manual Clone impl for a type that's Copy in the first place, instead of deriving it? If we want to depend on the two being equivalent at the library level (as proposed), we should lint against it. (If we were to want to depend on it at the language level (as not proposed), then we should just make it an error outright.)

[comex]

@glaebhoerl Unfortunately, derived impls on generic structs sometimes have the wrong trait bounds, so you have to implement them manually.

[oli-obk]

@comex: isn't that "just a bug"?

[strega-nil]

Also, it's possible to implement manually, so it's possible to implement manually wrong. This RFC isn't about all those people who #[derive(Copy, Clone)]. It's about anyone who manually implements Clone, for whatever reason (including the reasons @comex stated, or just reasons of style, or to rely on Copy for older compilers, or even really silly reasons).

[retep998]

@ubsan I manually implement Clone in winapi to force the implementation to rely on Copy and thereby improve compile times.

[strega-nil]

Some people do have to implement manually, like @retep998.

[arielb1]

I am not sure how much teeth this RFC has.

We already have both Clone::clone and Clone::clone_from, which may be implemented separately (and allocate differently, even). I don't think that we ever intended to specify which of them is called by standard library functions.

[strega-nil]

@arielb1 Yes, because it's specified that Clone::clone_from(x, y) shall be equivalent to *x = Clone::clone(y). We have to specify how cloning Copy types should work, or else we must do exactly what we say we are doing (i.e., cloning each individual value).

[arielb1]

Right. I think the "equivalent in functionally" language fits this case quite well.

[strega-nil]

I think that the RFC as written follows what you're saying?

[durka]

We did not yet make the #[derive(Copy, Clone)] change cited above, this potential behavior change being one of the reasons. If this language goes in then that obstacle would presumably be removed. However, I have an idea for how to do that optimization better anyway, after specialization stabilizes :)

[strega-nil]

@bluss ptr::read is specified in the document. It's not defined whether memcpy copies padding bytes.

[Kimundi]

@ubsan My point was not that there should be Clone specific compiler optimizations, but that there could be.

As in, if Clone is specified to not do more than Copy, then it would not matter if the compiler did this since there would be no observable behavior difference.

[arielb1]

@Kimundi

I see this as pretty similar to the situation with PartialEq and co. The compiler will not assume that your implementation of PartialEq is symmetric in any way (calls to a == b will not become calls to PartialEq::eq(&b, &a)). However, library functions might get confused.

[alexcrichton]

🔔 This RFC is now entering its week-long final comment period 🔔

[e-oz]

Hate breaking changes. Have you checked how many cargo crates will be affected? If yes - please share the numbers. From discussion it looks like breaking change just to make debug mode faster, but I think it can't be true.

[retep998]

@e-oz This won't cause anything to stop compiling. This merely makes an assumption that the standard library and compiler can optimize based on. If you can find any crates in the wild that would be affected negatively by this change, please do bring them up as there isn't really any way to find out in an automated fashion.

[ticki]

I think it is very unlikely that such a change will break anything at all.

[nikomatsakis]

(I also don't consider this a change, merely formally stating an existing assumption.)

Pre-emptive update: I don't really want to get into dickering over what is or is not a breaking change. What I meant here was that there is surely existing code that would exhibit bugs if clone and copy are not aligned (think generics), and that we've used the assumption that copy and clone align in their semantics elsewhere, but failed to document that assumption.

[e-oz]

Committed code contains words "this is a breaking change", that's why I asked. Sorry for being suspicious:)

[aturon]

I'm in favor of documenting these expectations for the semantics of Clone -- clearly, any other behavior for Copy data would be incredibly surprising and likely to lead to bugs.

[nikomatsakis]

Huzzah! The @rust-lang/lang team has decided to accept this RFC. However, since the libs team also claims jurisdiction, I'll let them merge it. :)

[alexcrichton]

The libs team discussed this RFC at triage today and the decision was to merge. It was brought up that the user-facing documentation probably wants to indicate *self instead of ptr::read as it's a bit more understandable, but that doesn't need to change the body of this RFC.

Thanks again for the RFC @ubsan!

Tracking issue: rust-lang/rust#33416

[bluss]

Vec::clone specialization for T: Copy is realized since rust-lang/rust/pull/38182 (Through extend_from_slice).