Exploring Rust's unspecified drop order
Dec 12, 2016
5 minutes read

After my previous post, I thought it would be interesting to run some experiments to determine the unspecified drop order within different constructs of Rust. After you read this, I guarantee you will understand why there is so much discussion about changing the current drop order before stabilizing it :)

TLDR: the current drop order is really weird!

In this post we are going to look at:

  • Local variables
  • Tuples
  • Structs and enums
  • Slices
  • Closure captures

We will be reusing the PrintDrop struct, so here is the definition in case you forgot it:

struct PrintDrop(&'static str);

impl Drop for PrintDrop {
    fn drop(&mut self) {
        println!("Dropping {}", self.0)
    }
}

Now we are ready to go!

Local variables

Let’s start with the following piece of code, testing the drop order of local variables:

fn main() {
    let x = PrintDrop("x");
    let y = PrintDrop("y");
}

And the output is…

Dropping y
Dropping x

As you can see, local variables are dropped in the reverse order of their declaration. This should come as no surprise, since new objects can store references to previously declared ones. Therefore a different drop order would result in dangling pointers.

The drop order of function parameters is similar, so the first parameter in the list is the last one to be dropped. The code is omitted for the sake of brevity, but this behavior can be trivially reproduced by writing a function with two by-value parameters.

If you think about drop order from the perspective of data structures, the behavior of local variables resembles the way a stack works. You could say that they are pushed onto the stack and popped at the end of the scope.

Tuples

After seeing the stack-like behavior of local variables, one would expect to see something similar in other constructs. However, tuples have a little surprise for us…

fn main() {
    let tup = (PrintDrop("x"), PrintDrop("y"), PrintDrop("z"));
}

And the output is…

Dropping x
Dropping y
Dropping z

Wait! Are you telling me that the variables are dropped in the same order as they are declared? So it seems! To continue with the data structures story, we could say that tuples behave like a queue, in which elements are enqueued in their order of appearance and dequeued at the end of the scope.

But this is not all! There is a subtle surprise lurking around the corner… If there is a panic during construction of the tuple, the drop order is reversed! If you don’t believe me, just run the code below:

fn main() {
    let tup = (PrintDrop("x"), PrintDrop("y"), panic!());
}

As I told you, the output is:

Dropping y
Dropping x

In other words, a tuple shows a queue-like drop order, unless one of the expressions in the tuple constructor panics. In case of a panic during construction, the drop order will be stack-like!

EDIT: as pointed out by birkenfeld on Reddit, the stack-like drop order actually makes sense in case of a panic. There is at this stage no tuple! Therefore, the expressions are dropped according to the rules of local variables.

Structs and enums

Structs present the same weird behavior as tuples. To a certain extent this seems consistent, since a struct is arguably a tuple with named fields instead of indices. It seems logical that they share the same drop order. The same holds for enums.

For the sake of brevity, the code below only tests the drop order of a struct. Of course, the same behavior is expected from tuple structs, tuple enum variants and struct enum variants.

struct Foo {
    x: PrintDrop,
    y: PrintDrop,
    z: PrintDrop,
}

fn main() {
    let foo = Foo {
        x: PrintDrop("x"),
        y: PrintDrop("y"),
        z: PrintDrop("z"),
    };
}

And the output is…

Dropping x
Dropping y
Dropping z

Again, the order is reversed when a panic occurs during construction:

fn main() {
    let foo = Foo {
        x: PrintDrop("x"),
        y: PrintDrop("y"),
        z: panic!(),
    };
}

As can be observed in the output below:

Dropping y
Dropping x

Looking at things from the bright side, at least we can say that this behavior is consistent across all data types. Still, it feels completely arbitrary to use a queue-like drop order.

Slices

Slices show the same queue-like behavior under normal circumstances and a stack-like behavior in the presence of a panic during construction.

Panic-free construction:

fn main() {
    let xs = [PrintDrop("x"), PrintDrop("y"), PrintDrop("z")];
}

Results in queue-like drop order:

Dropping x
Dropping y
Dropping z

Panic during construction:

fn main() {
    let xs = [PrintDrop("x"), PrintDrop("y"), panic!()];
}

Results in stack-like drop order:

Dropping y
Dropping x

Interestingly, Vec<T> shows the same drop order. As we are used to, a panic in the vec![] macro will reverse the drop order. However, if you panic after constructing the Vec by manually calling push a couple of times, the drop order will be queue-like (from Rust’s perspective you are dropping a fully constructed Vec).

Closure captures

An intriguing construct to close this post is the case of closure captures. We know that, under the hood, closures are actually structs that implement the Fn, FnMut or FnOnce traits. This means that the drop order depends on the order in which captures are declared in the generated struct.

Let’s start with a simple code example. Note that the order in which the variables are declared is different than the order in which they are used in the closure.

fn main() {
    let z = PrintDrop("z");
    let x = PrintDrop("x");
    let y = PrintDrop("y");

    let closure = move || {
        x; y; z;
    };
}

And the output is…

Dropping x
Dropping y
Dropping z

Based on the output it seems that the drop order is the same as the order in which the variables appear in the closure. However, we should test what happens in the scenario below:

let closure = move || {
    {
        let z_ref = &z;
    }
    x; y; z;
};

Again, the output is:

Dropping x
Dropping y
Dropping z

As you can see, even though z appears first as a reference, it is still the last one to be dropped. Therefore we should reformulate our hypothesis and say that the order in which captures are dropped is the same as the order in which they are moved. This way we ignore any references that may appear before.

Of course, we could perform more experiments to see if there are any edge cases to be aware about, but in the end the best approach would be to look at the source code of the compiler. This will certainly be necessary when drop order is stabilized.


Back to posts