Understanding #[derive(Clone)] in Rust

13 minute read

This post assumes that you have an entry-level familiarity with Rust: you’ve fought with the borrow checker enough to start to internalize some of its model; you’ve defined structs, implemented traits on those structs, and derived implementations of common traits using macros; you’ve seen trait bounds and maybe used one or two. Most importantly, you know what the Clone trait is and why you would want to use it1.

A Rust riddle

Now that we’ve set expectations, let’s start with a Rust riddle. This code compiles:

use std::rc::Rc;

trait NotNecessarilyClone {}

#[derive(Clone)]
struct CloneByRc<NNC> where NNC: NotNecessarilyClone {
  nnc: Rc<NNC>,
}

This on its own is not particularly surprising. We have a trait NotNecessarilyClone. Things that implement this trait might or might not also implement Clone. We have a struct CloneByRc that stores an object that implements NotNecessarilyClone, but it stores it behind an Rc2 so that CloneByRc can implement Clone. This is kind of the point of an Rc: Rc<T> implements Clone for all T. It takes something that isn’t normally safe to clone and hides it behind a reference counter so that cloning just means incrementing the reference counter.

The surprising part is this: CloneByRc<NNC> doesn’t actually implement Clone, even though we’ve used #[derive(Clone)]! If we try to use CloneByRc<NNC> somewhere that requires it to implement Clone, we get a compilation error:

trait MustBeClone: Clone {}

// Compilation error!
impl<NNC> MustBeClone for CloneByRc<NNC> where NNC: NotNecessarilyClone {}

Specifically, we get the error

error[E0277]: the trait bound `NNC: Clone` is not satisfied
  --> src/lib.rs:12:11
   |
10 | trait MustBeClone: Clone {}
   |                    ----- required by this bound in `MustBeClone`
11 |
12 | impl<NNC> MustBeClone for CloneByRc<NNC> where NNC: NotNecessarilyClone {}
   |           ^^^^^^^^^^^ the trait `Clone` is not implemented for `NNC`
   |
   = note: required because of the requirements on the impl of `Clone` for `CloneByRc<NNC>`
help: consider further restricting this bound
   |
12 | impl<NNC> MustBeClone for CloneByRc<NNC> where NNC: NotNecessarilyClone + Clone {}
   |                                                                          ^^^^^^^

At first, I was confused about why this failed to compile. Why does Rust complain that the trait Clone is not implemented for NNC? I only want CloneByRc<NNC> to implement Clone. The whole point of using an Rc was so that NNC wouldn’t need to implement Clone.

Once I understood the compilation error, I was confused about why the original code compiled at all. If CloneByRc<NCC> doesn’t actually implement Clone, why didn’t the original code throw an error when I tried to #[derive(Clone)]?

Eventually, I figured out both the compilation error and the successful compilation of the original code, which is why I’m writing this blog post.

What #[derive(Clone)] promises

Let’s start with what #[derive(Clone)] claims it does:

This trait can be used with #[derive] if all fields are Clone. The derived implementation of Clone calls clone on each field.

Ok, that’s pretty straightforward. If every field in a struct implements Clone, then you can just call clone on each field and now you’ve cloned the whole struct.

So why doesn’t this work for my CloneByRc<NNC> struct? It only has one field, and that field implements Clone. The derived implementation should just be able to call ncc.clone() and be done.

To understand what’s going on, we need to look at what happens when generics (in this case, the <NNC> type parameter) enter the picture.

Implementing #[derive(Clone)]: a naive approach

Let’s consider how we’d implement #[derive(Clone)] for a struct. We want to write a deriver function that will take in a description of the underlying struct, then emit Rust code that defines the implementation of Clone for that struct3.

To keep things simple, let’s just look at the Rust code that we’re going to emit. “The derived implementation of Clone calls clone on each field” sounds pretty straightforward. We can imagine our implementation looking like this:

Struct
Implementation
struct Foo {
  a: u32,
  b: Vec<u32>,
  c: bool
}




impl Clone for Foo {
  fn clone(&self) -> Self {
    Foo {
      a: self.a.clone(),
      b: self.b.clone(),
      c: self.c.clone(),
    }
  }
}

Once our macro emits this Rust code, the compiler will come along and check whether the code we emitted actually compiles. In this case, the code just calls clone on each field of the input struct. If every field implements Clone, then this will typecheck and the compiler will be happy. If some field doesn’t implement Clone, the compiler will see us calling field.clone() on that field and throw an error.

This approach works fine as long as the input struct doesn’t have any type parameters. But what happens if it does?

Struct
Implementation
struct Foo<B, C> {
  a: u32,
  b: Rc<B>,
  c: C,
}




impl<B, C> Clone for Foo<B, C> {
  fn clone(&self) -> Self {
    Foo {
      a: self.a.clone(),
      b: self.b.clone(),
      c: self.c.clone(), // problem!
    }
  }
}

As before, the compiler will check our emitted code. self.a.clone() is fine, since u32 implements Clone. self.b.clone() is fine, since Rc<B> implements Clone no matter what B is. self.c.clone() is not fine: we have no information about C, so the compiler can’t guarantee that it implements Clone.

We could fix the problem by adding a type constraint on C:

struct Foo<B, C: Clone> {
  a: u32,
  b: Rc<B>,
  c: C,
}

Now self.c.clone() will compile. But this constraint means that we can only build a Foo<B, C> when C implements Clone. This is actually a reasonable solution4, but it’s not the only reasonable solution. It’s also not the one the team implementing #[derive(Clone)] chose.

How #[derive(Clone)] is actually implemented

A nifty thing about Rust macros expanding to real Rust code is that we can inspect the expanded Rust code directly. Let’s make a new crate and put the following code into src/lib.rs:

use std::rc::Rc;

#[derive(Clone)]
struct Foo<B, C> {
  a: u32,
  b: Rc<B>,
  c: C,
}

Now we can run cargo expand5 to see how #[derive(Clone)] actually works:

#[automatically_derived]
#[allow(unused_qualifications)]
impl<B: ::core::clone::Clone, C: ::core::clone::Clone> ::core::clone::Clone for Foo<B, C> {
  #[inline]
  fn clone(&self) -> Foo<B, C> {
    match *self {
      Foo {
        a: ref __self_0_0,
        b: ref __self_0_1,
        c: ref __self_0_2,
      } => Foo {
        a: ::core::clone::Clone::clone(&(*__self_0_0)),
        b: ::core::clone::Clone::clone(&(*__self_0_1)),
        c: ::core::clone::Clone::clone(&(*__self_0_2)),
      },
    }
  }
}

There’s a lot of name mangling and weird reference stuff going on, but we can ignore that. The important part is the trait implementation signature:

impl<B: ::core::clone::Clone, C: ::core::clone::Clone> ::core::clone::Clone for Foo<B, C>

This is how #[derive(Clone)] guarantees that it can call self.c.clone(): it adds a type constraint on C so that Foo<B, C> only implements Clone when C implements Clone6. This is another reasonable solution to the problem: whereas the struct Foo<B, C: Clone> implementation above says that you can only create a Foo<B, C> when C implements Clone, this implementation says that you can create a Foo<B, C> with any B and C you like, but that you can only clone your Foo<B, C> if C implements Clone.

But there’s something else going on here. This implementation also adds the B: Clone constraint, even though we don’t need that constraint for Foo<B, C> to implement Clone. Remember, the b field of Foo<B, C> is a Rc<B>. b.clone() will typecheck for any B.

This at least explains the compilation error I was getting in my CloneByRc<NNC> example in the riddle I started with: Rust complained that the trait bound NNC: Clone is not satisfied because the implementation of #[derive(Clone)] was adding an NNC: Clone type constraint. But why does #[derive(Clone)] add this unnecessary constraint?

The limitations of Rust macros

What we’d really want is an implementation that adds type constraints only when we need them. For this example, we’d like something like

impl<B, C: Clone> Clone for Foo<B, C>

Why doesn’t #[derive(Clone)] produce this implementation instead?

The problem is that this implementation requires the macro to inspect the fields of Foo<B, C>, figure out which ones are already Clone, and only add the type constraint for the other fields. The short answer is that Rust’s procedural macros just aren’t powerful enough to do this. They don’t have access to the Rust compiler, so our optimal #[derive(Clone)] macro would have to implement all of the necessary type checking by itself.

The long answer is that this kind of inspection is actually impossible in practice. The input to the macro is just the list of tokens that define Foo<B, C>. It only gets syntactic information (“the name of the type of this field is Rc<B>”) rather than semantic information (the definition of the Rc type), so the macro would need to do its type checking given only the names of the types being referenced. The definition of Foo<B, C> might include references to arbitrary other types, including types defined elsewhere in the crate, so there’s no way the macro can know how to check if the type implements Clone given only its name.

Do we have to inspect the fields?

So we can’t check if our fields are guaranteed to implement Clone. But there’s still got to be a way to avoid unnecessary type constraints, right? What if we try being a little more clever? For example, we could imagine an implementation that adds exactly the type constraints it needs:

impl<B, C> Clone for Foo<B, C> where Rc<B>: Clone, C: Clone

Here we’ve just taken the exact types of the fields that use our generic type parameters and constrained those types to implement Clone. This seems like exactly what we want: Rc<B>: Clone is true for all B, so C: Clone is the only nontrivial constraint.

This works for our current example, but it has some unfortunate downsides. First of all, our implementation now reveals that Foo<B, C> uses an Rc<B>. Since b was a private field of Foo, it’s a little weird to leak this information.

The bigger downside, though, is that we might start leaking private types. Let’s take a look at this example:

Struct
Implementation
struct PrivateFoo<B> {
  b: B,
}

pub struct Foo<B, C> {
  a: u32,
  b: PrivateFoo<B>,
  c: C,
}




impl<B, C> Clone for Foo<B, C>
where
  PrivateFoo<B>: Clone,
  C: Clone,
{
  fn clone(&self) -> Self {
    Foo {
      a: self.a.clone(),
      b: self.b.clone(),
      c: self.c.clone(),
    }
  }
}

Now our private type PrivateFoo<B> is part of the public signature of our Clone implementation. This will break compilation: the Rust compiler will throw error[E0446]: private type PrivateFoo<B> in public interface7. I think this error is a lot worse than the trait bound NNC: Clone is not satisfied error we get with the current implementation of #[derive(Clone)], since this error sounds scarier and doesn’t seem obviously related to implementing Clone. If I saw this error, I’d probably think something was wrong elsewhere in my program and waste a bunch of time trying to track it down8.

The workaround

Ok, so we’ve looked at a bunch of alternative ways to derive Clone and none of them seem like they would actually work. My struct is still broken. How do I fix it?

The answer, it turns out, is pretty straightforward: just implement Clone manually instead of deriving it. Unlike #[derive(Clone)], you only need to solve the problem for one struct and you have access to semantic information about all of the types you’re using. You can pick the exact type constraints you want.

In fact, I suspect that part of the reason #[derive(Clone)] hasn’t fixed this issue is that the workaround is so easy.

Conclusion

The answer to the first part of my riddle (“why do I get an error message complaining that the trait Clone is not implemented for NNC?”) is that the implementation of Clone produced by #[derive(Clone)] adds a type constraint NNC: Clone. The answer to the second part (“why did the declaration of CloneByRc<NNC> compile if CloneByRc<NNC> isn’t actually guaranteed to be Clone?”) also comes down to that type constraint: because of the type constraint, #[derive(Clone)] doesn’t actually guarantee that CloneByRc<NNC> will be Clone for all NNC. It only guarantees that CloneByRc<NNC> will be Clone whenever NNC is Clone.

Seeing that type constraint was the first step on our tour through many hypothetical implementations for #[derive(Clone)]. Because of the limitations on Rust’s derive macros, each of these implementations has a downside: the perfect ones are impossible and the possible ones are imperfect. For any implementation that could be written in Rust today, you’ll have to implement Clone manually for some types of structs. You might choose to optimize for the most common case or for the case where implementing manually is the most painful, but whichever one you pick you’re going to leave somebody wondering why they have to implement Clone manually.

Here are the two most plausible implementations:

Implementation No type constraints Type constraints on all type parameters
Example impl<B, C> Clone for Foo<B, C> { ... } impl<B: Clone, C: Clone> Clone for Foo<B, C> { ... }
Works when All fields implement Clone regardless of the input types All input types implement Clone
Breaks when At least one field might not implement Clone for some input type At least one field implements Clone regardless of the input types (for example, an Rc<T>), but at least one of that field’s input types doesn’t implement Clone

Given the situations in which these two implementations break down and the fact that the workaround is so straightforward, I’m not surprised that the authors of #[derive(Clone)] chose the “type constraints on all type parameters” approach.

  1. I’m using Clone as my example in this blog post, but this discussion applies to any derivable trait. If you’re not familiar with Clone, here’s a good explainer

  2. Short for ‘Reference Counted’. An Rc stores a pointer to an object that lives on the heap and a count of the number of live references to that object. When you create a new reference to the object, the Rc increments the counter. When you drop a reference to the object, the Rc decrements the counter. When the counter drops to 0 (when the last reference to the object is dropped) the object itself is dropped. This means that you can keep a bunch of references to the same underlying object instead of making multiple copies of the object (which might break the object’s functionality if it’s something like a shared counter or an HTTP client). 

  3. To be precise, #[derive(Clone)] is a procedural macro, which means that it takes a stream of tokens (the individual words of Rust code that make up the struct definition) and returns a stream of tokens (the words that make up the trait implementation). 

  4. Or at least, I think it’s reasonable. As we’ll see, there isn’t really a perfect solution here so any solution is going to end up missing some edge cases. Which edge cases you choose to leave out depends on your mental model of which edge cases are most common/most painful to work around. I doubt that anybody has done any meaningful data analysis to measure that in an objective sense, so it doesn’t seem like this solution is obviously worse than the one chosen. 

  5. If you’re having trouble getting cargo expand to run, I have a blog post just for you. 

  6. If you’re wondering why the expanded code uses the fully-qualified trait name ::core::clone::Clone instead of just Clone, it has to do with the way Rust’s procedural macros work. When the compiler invokes the macro, the macro returns a block of Rust code that the compiler just inserts directly in-place at the macro’s invocation site. The compiler doesn’t provide any namespacing for the emitted code, so the macro emits fully-qualified names to avoid name collisions. If the macro just used Clone and the file happened to import some other trait named Clone, the type checker would expect the emitted code to implement that other Clone instead of ::core::clone::Clone

  7. This is by no means the only thing that could go wrong with this type of implementation. There’s a pretty good discussion of other scary and exciting consequences on this Github issue, which was the source for most of this blog post. 

  8. While the error message would tell me that this error came from the macro implementation of Clone, before going down this rabbit hole I would have had no idea why an implementation of Clone would leak a private type. 

Thanks to Bianca Homberg for giving me feedback on this post

Tags:

Categories:

Updated: