1 // Rust-101, Part 06: Copy, Lifetimes
2 // ==================================
4 // We continue to work on our `BigInt`, so we start by importing what we already established.
7 // With `BigInt` being about numbers, we should be able to write a version of `vec_min`
8 // that computes the minimum of a list of `BigInt`. First, we have to write `min` for `BigInt`.
10 fn min_try1(self, other: Self) -> Self {
11 // Just to be sure, we first check that both operands actually satisfy our invariant. `debug_assert!` is a
12 // macro that checks that its argument (must be of type `bool`) is `true`, and panics otherwise. It gets
13 // removed in release builds, which you do with `cargo build --release`.
14 debug_assert!(self.test_invariant() && other.test_invariant());
15 // Now our assumption of having no trailing zeros comes in handy:
16 // If the lengths of the two numbers differ, we already know which is larger.
17 if self.data.len() < other.data.len() {
19 } else if self.data.len() > other.data.len() {
22 // **Exercise 06.1**: Fill in this code.
28 // Now we can write `vec_min`. In order to make it type-check, we have make a deep copy of e.
29 fn vec_min(v: &Vec<BigInt>) -> Option<BigInt> {
30 let mut min: Option<BigInt> = None;
33 min = Some(match min {
35 Some(n) => e.min_try1(n)
40 // Now, what's happening here? Why do we have to clone `e`, and why did we not
41 // have to do that in our previous version?
43 // The answer is already hidden in the type of `vec_min`: `v` is just borrowed, but
44 // the Option<BigInt> that it returns is *owned*. We can't just return one of the elements of `v`,
45 // as that would mean that it is no longer in the vector! In our code, this comes up when we update
46 // the intermediate variable `min`, which also has type `Option<BigInt>`. If you replace get rid of the
47 // `e.clone()`, Rust will complain "Cannot move out of borrowed content". That's because
48 // `e` is a `&BigInt`. Assigning `min = Some(*e)` works just like a function call: Ownership of the
49 // underlying data is transferred from where `e` borrows from to `min`. But that's not allowed, since
50 // we just borrowed `e`, so we cannot empty it! We can, however, call `clone()` on it. Then we own
51 // the copy that was created, and hence we can store it in `min`.<br/>
52 // Of course, making such a full copy is expensive, so we'd like to avoid it. We'll some to that in the next part.
55 // But before we go there, I should answer the second question I brought up above: Why did our old `vec_min` work?
56 // We stored the minimal `i32` locally without cloning, and Rust did not complain. That's because there isn't
57 // really much of an "ownership" when it comes to types like `i32` or `bool`: If you move the value from one
58 // place to another, then both instances are "complete". We also say the value has been *duplicated*. This is in
59 // stark contrast to types like `Vec<i32>`, where moving the value results in both the old and the new vector to
60 // point to the same underlying buffer. We don't have two vectors, there's no proper duplication.
62 // Rust calls types that can be easily duplicated `Copy` types. `Copy` is another trait, and it is implemented for
63 // types like `i32` and `bool`. Remember how we defined the trait `Minimum` by writing `trait Minimum : Copy { ...`?
64 // This tells Rust that every type that implements `Minimum` must also implement `Copy`, and that's why the compiler
65 // accepted our generic `vec_min` in part 02. `Copy` is the first *marker trait* that we encounter: It does not provide
66 // any methods, but makes a promise about the behavior of the type - in this case, being duplicable.
68 // If you try to implement `Copy` for `BigInt`, you will notice that Rust
69 // does not let you do that. A type can only be `Copy` if all its elements
70 // are `Copy`, and that's not the case for `BigInt`. However, we can make
71 // `SomethingOrNothing<T>` copy if `T` is `Copy`.
72 use part02::{SomethingOrNothing,Something,Nothing};
73 impl<T: Copy> Copy for SomethingOrNothing<T> {}
74 // Again, Rust can generate implementations of `Copy` automatically. If
75 // you add `#[derive(Copy,Clone)]` right before the definition of `SomethingOrNothing`,
76 // both `Copy` and `Clone` will automatically be implemented.
78 // ## An operational perspective
79 // Instead of looking at what happens "at the surface" (i.e., visible in Rust), one can also explain
80 // ownership passing and how `Copy` and `Clone` fit in by looking at what happens on the machine.<br/>
81 // When Rust code is executed, passing a value (like `i32` or `Vec<i32>`) to a function will always
82 // result in a shallow copy being performed: Rust just copies the bytes representing that value, and
83 // considers itself done. That's just like the default copy constructor in C++. Rust, however, will
84 // consider this a destructive operation: After copying the bytes elsewhere, the original value must
85 // no longer be used. After all, the two could now share a pointer! If, however, you mark a type `Copy`,
86 // then Rust will *not* consider a move destructive, and just like in C++, the old and new value
87 // can happily coexist. Now, Rust does not allow you to overload the copy constructor. This means that
88 // passing a value around will always be a fast operation, no allocation or any other kind of heap access
89 // will happen. In the situations where you would write a copy constructor in C++ (and hence
90 // incur a hidden cost on every copy of this type), you'd have the type *not* implement `Copy`, but only
91 // `Clone`. This makes the cost explicit.
94 // To fix the performance problems of `vec_min`, we need to avoid using `clone()`. We'd like
95 // the return value to not be owned (remember that this was the source of our need for cloning), but *borrowed*.
97 // The function `head` demonstrates how that could work: It borrows the first element of a vector if it is non-empty.
98 // The type of the function says that it will either return nothing, or it will return a borrowed `T`.
99 // We can then borrow the first element of `v` and use it to construct the return value.
100 fn head<T>(v: &Vec<T>) -> Option<&T> {
107 // Technically, we are returning a pointer to the first element. But doesn't that mean that callers have to be
108 // careful? Imagine `head` would be a C++ function, and we would write the following code.
110 int foo(std::vector<int> v) {
111 int *first = head(v);
116 // This is very much like our very first motivating example for ownership, at the beginning of part 04:
117 // `push_back` could reallocate the buffer, making `first` an invalid pointer. Again, we have aliasing (of `first`
118 // and `v`) and mutation. But this time, the bug is hidden behind the call to `head`. How does Rust solve this? If we translate
119 // the code above to Rust, it doesn't compile, so clearly we are good - but how and why?
120 // (Notice that have to explicitly assert using `unwrap` that `first` is not `None`, whereas the C++ code
121 // above would silently dereference a `NULL`-pointer. But that's another point.)
122 fn rust_foo(mut v: Vec<i32>) -> i32 {
123 let first: Option<&i32> = head(&v);
128 // To give the answer to this question, we have to talk about the *lifetime* of a borrow. The point is, saying that
129 // you borrowed your friend a `Vec<i32>`, or a book, is not good enough, unless you also agree on *how long*
130 // your friend can borrow it. After all, you need to know when you can rely on owning your data (or book) again.
132 // Every borrow in Rust has an associated lifetime, written `&'a T` for a borrow of type `T` with lifetime `'a`. The full
133 // type of `head` reads as follows: `fn<'a, T>(&'a Vec<T>) -> Option<&'a T>`. Here, `'a` is a *lifetime variable*, which
134 // represents how long the vector has been borrowed. The function type expresses that argument and return value have *the same lifetime*.
136 // When analyzing the code of `rust_foo`, Rust has to assign a lifetime to `first`. It will choose the scope
137 // where `first` is valid, which is the entire rest of the function. Because `head` ties the lifetime of its
138 // argument and return value together, this means that `&v` also has to borrow `v` for the entire duration of
139 // the function. So when we try to borrow `v` mutable for `push`, Rust complains that the two borrows (the one
140 // for `head`, and the one for `push`) overlap. Lucky us! Rust caught our mistake and made sure we don't crash the program.
142 // So, to sum this up: Lifetimes enable Rust to reason about *how long* a pointer has been borrowed. We can thus
143 // safely write functions like `head`, that return pointers into data they got as argument, and make sure they
144 // are used correctly, *while looking only at the function type*. At no point in our analysis of `rust_foo` did
145 // we have to look *into* `head`. That's, of course, crucial if we want to separate library code from application code.
146 // Most of the time, we don't have to explicitly add lifetimes to function types. This is thanks to *lifetimes elision*,
147 // where Rust will automatically insert lifetimes we did not specify, following some [simple, well-documented rules](http://doc.rust-lang.org/stable/book/lifetimes.html#lifetime-elision).
149 // [index](main.html) | [previous](part05.html) | [next](part07.html)