X-Git-Url: https://git.ralfj.de/rust-101.git/blobdiff_plain/a9c7d7471bf6f06a2a4710daa306c26fc5324557..3530068c5bbcbd9185757369dd23b9edb3bbc9ad:/src/part11.rs diff --git a/src/part11.rs b/src/part11.rs index 4e5c5fd..a5cafa1 100644 --- a/src/part11.rs +++ b/src/part11.rs @@ -1,150 +1,120 @@ -// Rust-101, Part 11: Trait Objects, Box, Rc, Lifetime bounds -// ========================================================== +// Rust-101, Part 11: Trait Objects, Box, Lifetime bounds +// ====================================================== //@ We will play around with closures a bit more. Let us implement some kind of generic "callback" -//@ mechanism, providing two functions: Registering a new callback, and calling all registered callbacks. There will be two -//@ versions, so to avoid clashes of names, we put them into modules. -mod callbacks { - //@ First of all, we need to find a way to store the callbacks. Clearly, there will be a `Vec` involved, so that we can - //@ always grow the number of registered callbacks. A callback will be a closure, i.e., something implementing - //@ `FnMut(i32)` (we want to call this multiple times, so clearly `FnOnce` would be no good). So our first attempt may be the following. - // For now, we just decide that the callbacks have an argument of type `i32`. - struct CallbacksV1 { - callbacks: Vec, - } - //@ However, this will not work. Remember how the "type" of a closure is specific to the environment of captured variables. Different closures - //@ all implementing `FnMut(i32)` may have different types. However, a `Vec` is a *uniformly typed* vector. - - //@ We will thus need a way to store things of *different* types in the same vector. We know all these types implement `FnMut(i32)`. For this scenario, - //@ Rust provides *trait objects*: The truth is, `FnMut(i32)` is not just a trait. It is also a type, that can be given to anything implementing - //@ this trait. So, we may write the following. - /* struct CallbacksV2 { - callbacks: Vec, - } */ - //@ But, Rust complains about this definition. It says something about "Sized". What's the trouble? See, for many things we want to do, it is crucial that - //@ Rust knows the precise, fixed size of the type - that is, how large this type will be when represented in memory. For example, for a `Vec`, the - //@ elements are stored one right after the other. How should that be possible, without a fixed size? The trouble is, `FnMut(i32)` could be of any size. - //@ We don't know how large that "type that implemenets `FnMut(i32)`" is. Rust calls this an *unsized* type. Whenever we introduce a type variable, Rust - //@ will implicitly add a bound to that variable, demanding that it is sized. That's why we did not have to worry about this so far.
- //@ You can opt-out of this implicit bound by saying `T: ?Sized`. Then `T` may or may not be sized. - - //@ So, what can we do, if we can't store the callbacks in a vector? We can put them in a box. Semantically, `Box` is a lot like `T`: You fully own - //@ the data stored there. On the machine, however, `Box` is a *pointer* to `T`. It is a lot like `std::unique_ptr` in C++. In our current example, - //@ the important bit is that since it's a pointer, `T` can be unsized, but `Box` itself will always be sized. So we can put it in a `Vec`. - pub struct Callbacks { - callbacks: Vec>, - } - - impl Callbacks { - // Now we can provide some functions. The constructor should be straight-forward. - pub fn new() -> Self { - Callbacks { callbacks: Vec::new() } /*@*/ - } - - // Registration simply stores the callback. - pub fn register(&mut self, callback: Box) { - self.callbacks.push(callback); /*@*/ - } - - // And here we call all the stored callbacks. - pub fn call(&mut self, val: i32) { - // Since they are of type `FnMut`, we need to mutably iterate. Notice that boxes dereference implicitly. - for callback in self.callbacks.iter_mut() { - callback(val); /*@*/ - } - } - } - - // Now we are ready for the demo. - pub fn demo(c: &mut Callbacks) { - c.register(Box::new(|val| println!("Callback 1: {}", val))); - c.call(0); - - //@ We can even register callbacks that modify their environment. Rust will again attempt to borrow `count`. However, - //@ that doesn't work out this time: Since we want to put this thing in a `Box`, it could live longer than the function - //@ we are in. Then the borrow of `count` would become invalid. We have to explicitly tell Rust to `move` ownership of the - //@ variable into the closure. Its environment will then contain a `usize` rather than a `&mut uszie`, and have - //@ no effect on this local variable anymore. - let mut count: usize = 0; - c.register(Box::new(move |val| { count = count+1; println!("Callback 2, {}. time: {}", count, val); } )); - c.call(1); c.call(2); - } +//@ mechanism, providing two functions: Registering a new callback, and calling all registered callbacks. + +//@ First of all, we need to find a way to store the callbacks. Clearly, there will be a `Vec` involved, so that we can +//@ always grow the number of registered callbacks. A callback will be a closure, i.e., something implementing +//@ `FnMut(i32)` (we want to call this multiple times, so clearly `FnOnce` would be no good). So our first attempt may be the following. +// For now, we just decide that the callbacks have an argument of type `i32`. +struct CallbacksV1 { + callbacks: Vec, } - -// Remember to edit `main.rs` to run the demo. -pub fn main() { - let mut c = callbacks::Callbacks::new(); - callbacks::demo(&mut c); +//@ However, this will not work. Remember how the "type" of a closure is specific to the environment of captured variables. Different closures +//@ all implementing `FnMut(i32)` may have different types. However, a `Vec` is a *uniformly typed* vector. + +//@ We will thus need a way to store things of *different* types in the same vector. We know all these types implement `FnMut(i32)`. For this scenario, +//@ Rust provides *trait objects*: The truth is, `FnMut(i32)` is not just a trait. It is also a type, that can be given to anything implementing +//@ this trait. So, we may write the following. +/* struct CallbacksV2 { + callbacks: Vec, +} */ +//@ But, Rust complains about this definition. It says something about "Sized". What's the trouble? See, for many things we want to do, it is crucial that +//@ Rust knows the precise, fixed size of the type - that is, how large this type will be when represented in memory. For example, for a `Vec`, the +//@ elements are stored one right after the other. How should that be possible, without a fixed size? The point is, `FnMut(i32)` could be of any size. +//@ We don't know how large that "type that implemenets `FnMut(i32)`" is. Rust calls this an *unsized* type. Whenever we introduce a type variable, Rust +//@ will implicitly add a bound to that variable, demanding that it is sized. That's why we did not have to worry about this so far.
+//@ You can opt-out of this implicit bound by saying `T: ?Sized`. Then `T` may or may not be sized. + +//@ So, what can we do, if we can't store the callbacks in a vector? We can put them in a box. Semantically, `Box` is a lot like `T`: You fully own +//@ the data stored there. On the machine, however, `Box` is a *pointer* to a heap-allocated `T`. It is a lot like `std::unique_ptr` in C++. In our current example, +//@ the important bit is that since it's a pointer, `T` can be unsized, but `Box` itself will always be sized. So we can put it in a `Vec`. +pub struct Callbacks { + callbacks: Vec>, } -mod callbacks_clone { - //@ So, this worked great, didn't it! There's one point though that I'd like to emphasize: One cannot `clone` a closure. - //@ Hence it becomes impossible to implement `Clone` for our `Callbacks` type. What could we do about this? - - //@ You already learned about `Box` above. `Box` is an example of a *smart pointer*: It's like a pointer (in the C - //@ sense), but with some additional smarts to it. For `Box`, that's the part about ownership. Once you drop the box, the - //@ content it points to will be deleted.
- //@ Another example of a smart pointer is `Rc`. This is short for *reference-counter*, so you can already guess how - //@ this pointer is smart: It has a reference count. You can `clone` an `Rc` as often as you want, that doesn't affect the - //@ data it contains at all. It only creates more references to the same data. Once all the references are gone, the data is deleted. - //@ - //@ Wait a moment, you may say here. Multiple references to the same data? That's aliasing! Indeed, we have to be careful. - //@ Once data is stored in an `Rc`, it is read-only: By dereferencing the smart `Rc`, you can only get a shared borrow of the data. - use std::rc::Rc; - - //@ Because of this read-only restriction, we cannot use `FnMut` here: We'd be unable to call the function with a mutable borrow - //@ of it's environment! So we have to go with `Fn`. We wrap that in an `Rc`, and then Rust happily derives `Clone` for us. - #[derive(Clone)] - pub struct Callbacks { - callbacks: Vec>, +impl Callbacks { + // Now we can provide some functions. The constructor should be straight-forward. + pub fn new() -> Self { + Callbacks { callbacks: Vec::new() } /*@*/ } - impl Callbacks { - pub fn new() -> Self { - Callbacks { callbacks: Vec::new() } /*@*/ - } + // Registration simply stores the callback. + pub fn register(&mut self, callback: Box) { + self.callbacks.push(callback); + } - // For the `register` function, we don't actually have to use trait objects in the argument. - //@ We can make this function generic, such that it will be instantiated with some concrete closure type `F` - //@ and do the creation of the `Rc` and the conversion to `Fn(i32)` itself. - - //@ For this to work, we need to demand that the type `F` does not contain any short-lived borrows. After all, we will store it - //@ in our list of callbacks indefinitely. `'static` is a lifetime, the lifetime of the entire program. We can use lifetimes - //@ as bounds on types, to demand that anything in (an element of) the type lives at least as long as this lifetime. That bound was implicit in the `Box` - //@ above, and it is the reason we could not have the borrowed `count` in the closure in `demo`. - pub fn register(&mut self, callback: F) { - self.callbacks.push(Rc::new(callback)); /*@*/ - } + // We can also write a generic version of `register`, such that it will be instantiated with some concrete closure type `F` + // and do the creation of the `Box` and the conversion from `F` to `FnMut(i32)` itself. + + //@ For this to work, we need to demand that the type `F` does not contain any short-lived references. After all, we will store it + //@ in our list of callbacks indefinitely. If the closure contained a pointer to our caller's stackframe, that pointer + //@ could be invalid by the time the closure is called. We can mitigate this by bounding `F` by a *lifetime*: `F: 'a` says + //@ that all data of type `F` will *outlive* (i.e., will be valid for at least as long as) lifetime `'a`. + //@ Here, we use the special lifetime `'static`, which is the lifetime of the entire program. + //@ The same bound has been implicitly added in the version of `register` above, and in the definition of + //@ `Callbacks`. + pub fn register_generic(&mut self, callback: F) { + self.callbacks.push(Box::new(callback)); /*@*/ + } - pub fn call(&mut self, val: i32) { - // We only need a shared iterator here. `Rc` also implicitly dereferences, so we can simply call the callback. - for callback in self.callbacks.iter() { - callback(val); /*@*/ - } + // And here we call all the stored callbacks. + pub fn call(&mut self, val: i32) { + // Since they are of type `FnMut`, we need to mutably iterate. + for callback in self.callbacks.iter_mut() { + //@ Here, `callback` has type `&mut Box`. We can make use of the fact that `Box` is a *smart pointer*: In + //@ particular, we can use it as if it were a normal reference, and use `*` to get to its contents. Then we obtain a + //@ mutable reference to these contents, because we call a `FnMut`. + (&mut *callback)(val); /*@*/ + //@ Just like it is the case with normal references, this typically happens implicitly with smart pointers, so we can also directly call the function. + //@ Try removing the `&mut *`. + //@ + //@ The difference to a reference is that `Box` implies full ownership: Once you drop the box (i.e., when the entire `Callbacks` instance is + //@ dropped), the content it points to on the heap will be deleted. } } +} - // The demo works just as above. Our counting callback doesn't work anymore though, because we are using `Fn` now. - fn demo(c: &mut Callbacks) { - c.register(|val| println!("Callback 1: {}", val)); - c.call(0); c.call(1); +// Now we are ready for the demo. Remember to edit `main.rs` to run it. +pub fn main() { + let mut c = Callbacks::new(); + c.register(Box::new(|val| println!("Callback 1: {}", val))); + c.call(0); + + { + //@ We can even register callbacks that modify their environment. Per default, Rust will attempt to capture a reference to `count`, to borrow it. However, + //@ that doesn't work out this time. Remember the `'static` bound above? Borrowing `count` in the environment would + //@ violate that bound, as the reference is only valid for this block. If the callbacks are triggered later, we'd be in trouble. + //@ We have to explicitly tell Rust to `move` ownership of the variable into the closure. Its environment will then contain a + //@ `usize` rather than a `&mut usize`, and the closure has no effect on this local variable anymore. + let mut count: usize = 0; + c.register_generic(move |val| { + count = count+1; + println!("Callback 2: {} ({}. time)", val, count); + } ); } + c.call(1); c.call(2); } -// **Exercise 11.1**: We made the arbitrary choice of using `i32` for the arguments. Generalize the data-structures above -// to work with an arbitrary type `T` that's passed to the callbacks. Since you need to call multiple callbacks with the -// same `t: T`, you will either have to restrict `T` to `Copy` types, or pass a borrow. - //@ ## Run-time behavior //@ When you run the program above, how does Rust know what to do with the callbacks? Since an unsized type lacks some information, -//@ a *pointer* to such a type (be it a `Box`, an `Rc` or a borrow) will need to complete this information. We say that pointers to +//@ a *pointer* to such a type (be it a `Box` or a reference) will need to complete this information. We say that pointers to //@ trait objects are *fat*. They store not only the address of the object, but (in the case of trait objects) also a *vtable*: A //@ table of function pointers, determining the code that's run when a trait method is called. There are some restrictions for traits to be usable -//@ as trait objects. This is called *object safety* and described in [the documentation](http://doc.rust-lang.org/stable/book/trait-objects.html) and [the reference](http://doc.rust-lang.org/reference.html#trait-objects). +//@ as trait objects. This is called *object safety* and described in [the documentation](https://doc.rust-lang.org/stable/book/trait-objects.html) and [the reference](https://doc.rust-lang.org/reference.html#trait-objects). +//@ In case of the `FnMut` trait, there's only a single action to be performed: Calling the closure. You can thus think of a pointer to `FnMut` as +//@ a pointer to the code, and a pointer to the environment. This is how Rust recovers the typical encoding of closures as a special case of a more +//@ general concept. //@ -//@ Whenever you write a generic function, you have a choice: You can make it polymorphic, like our `vec_min`. Or you -//@ can use trait objects, like the first `register` above. The latter will result in only a single compiled version (rather +//@ Whenever you write a generic function, you have a choice: You can make it generic, like `register_generic`. Or you +//@ can use trait objects, like `register`. The latter will result in only a single compiled version (rather //@ than one version per type it is instantiated with). This makes for smaller code, but you pay the overhead of the virtual function calls. -//@ Isn't it beautiful how traits can handle both of these cases (and much more, as we saw, like closures and operator overloading) nicely? +//@ (Of course, in the case of `register` above, there's no function called on the trait object.) +//@ Isn't it beautiful how traits can nicely handle this tradeoff (and much more, as we saw, like closures and operator overloading)? + +// **Exercise 11.1**: We made the arbitrary choice of using `i32` for the arguments. Generalize the data structures above +// to work with an arbitrary type `T` that's passed to the callbacks. Since you need to call multiple callbacks with the +// same `t: T`, you will either have to restrict `T` to `Copy` types, or pass a reference. -//@ [index](main.html) | [previous](part10.html) | [next](main.html) +//@ [index](main.html) | [previous](part10.html) | [raw source](https://www.ralfj.de/git/rust-101.git/blob_plain/HEAD:/workspace/src/part11.rs) | [next](part12.html)