What is a COW type22.04.21

Creating a Rust function that returns a &str or String

Written by Herman J. Radtke III on 29 May 2015

Russian Translation

We learned how to create a function that accepts String or &str as an argument. Now I want to show you how to create a function that returns either String or &str. I also want to discuss why we would want to do this. To start, let us write a function to remove all the spaces from a given string. Our function might look something like this:

fn remove_spaces(input: &str) -> String {
   let mut buf = String::with_capacity(input.len());

for c in input.chars() {
      if c != ' ' {
         buf.push(c);
      }
   }

buf
}

This function allocates memory for a string buffer, loops through each character of input and appends all non-space characters to the string buffer. Now I ask: what if my input did not contain spaces at all? The value input would be the same as buf. In that case, it would be more efficient to not create buf in the first place. Instead, we would like to just return the given input back to the caller. The type of input is a &str but our function returns a String though. We could change the type of input to a String:

fn remove_spaces(input: String) -> String { ... }

but this causes two problems. First, by making input of type String we are forcing the caller to move the ownership of input into our function. This prevents the caller from using that value in the future. We should only take ownership of input if we actually need it. Second, the input might already be of type &str and we are now forcing the caller to convert it into a String which defeats our attempts to not allocate new memory when creating buf.
Clone-on-write

What we really want is the ability to return our input string (&str) if there are no spaces and to return a new string (String) if there are spaces we need to remove. This is where the clone-on-write or Cow type can be used. The Cow type allows us to abstract away whether something is Owned or Borrowed. In our example, the &str is a reference to an existing string so that would be borrowed data. If there are spaces, then we need to allocate memory for a new String. That new String is owned by the buf variable. Normally, we would move the ownership of buf by returning it to the caller. When using Cow, we want to move the ownership of buf into the Cow type and return that.

use std::borrow::Cow;

fn remove_spaces<'a>(input: &'a str) -> Cow<'a, str> {
    if input.contains(' ') {
        let mut buf = String::with_capacity(input.len());

for c in input.chars() {
            if c != ' ' {
                buf.push(c);
            }
        }

return Cow::Owned(buf);
    }

return Cow::Borrowed(input);
}

Our function now checks to see if the given input contains a space and only then allocates memory for a new buffer. If the input does not contain a space, the input is simply returned. We are adding a bit of runtime complexity to optimize how we allocate memory. Notice that our Cow type has the same lifetime of the &str type. As we discussed previously, the compiler needs to track the &str reference to know when it can safely free (or Drop) the memory.

The beauty of Cow is that it implements the Deref trait so you can call immutable functions without knowing whether or not the result is a new string buffer or not. Example:

let s = remove_spaces("Herman Radtke");
println!("Length of string is {}", s.len());

If I do need to mutate s, then I can convert it into an owned variable using the into_owned() function. If the variant of Cow was already Owned then we are simply moving ownership. If the variant of Cow is Borrowed, then we are allocating memory. This allows us to lazily clone (allocate memory) only when we want to write (or mutate) the variable.

Example where a Cow::Borrowed is mutated:

let s = remove_spaces("Herman"); // s is a Cow::Borrowed variant
let len = s.len(); // immutable function call using Deref
let owned: String = s.into_owned(); // memory is allocated for a new string

Example where a Cow::Owned is mutated:

let s = remove_spaces("Herman Radtke"); // s is a Cow::Owned variant
let len = s.len(); // immutable function call using Deref
let owned: String = s.into_owned(); // no new memory allocated as we already had a String

The idea behind Cow is two-fold:

Delay the allocation of memory for as long as possible. In the best case, we never have to allocate any new memory.
    Allow the caller of our remove_spaces function to not care if memory was allocated or not. The usage of the Cow type is the same in either case.

Leveraging the Into Trait

We previously discussed using the Into trait to convert a &str into a String. We can also use the Into trait to convert the &str or String into the proper Cow variant. By calling .into() the compiler will perform the conversion automatically. Using .into() will not speed up or slow down the code. It is simply an option to avoid having to specify Cow::Owned or Cow::Borrowed explicitly.

fn remove_spaces<'a>(input: &'a str) -> Cow<'a, str> {
    if input.contains(' ') {
        let mut buf = String::with_capacity(input.len());
        let v: Vec<char> = input.chars().collect();

for c in v {
            if c != ' ' {
                buf.push(c);
            }
        }

return buf.into();
    }
    return input.into();
}

We can also clean this up a bit using just iterators:

fn remove_spaces<'a>(input: &'a str) -> Cow<'a, str> {
    if input.contains(' ') {
        input
        .chars()
        .filter(|&x| x != ' ')
        .collect::<std::string::String>()
        .into()
    } else {
        input.into()
    }
}

Real World Uses of Cow

My example of removing spaces may seem a bit contrived, but there are some great real-world applications of this strategy. Inside of Rust core there is a function that converts bytes to UTF-8 in a lossy manner and a function that will translate CRLF to LF. Both of these functions have a case where a &str can be returned in the optimal case and another case where a String has to be allocated. Other examples I can think of are properly encoding an xml/html string or properly escaping a SQL query. In many cases, the input is already properly encoded or escaped. In those cases, it is better to just return the input string back to the caller. When the input does need to be modified we are forced to allocate new memory, in the form of a String buffer, and return that to the caller.
Why use String::with_capacity() ?

While we are on the topic of efficient memory management, notice that I used String::with_capacity() instead of String::new() when creating the string buffer. You can use String::new() instead of String::with_capacity(), but it is more efficient to allocate memory for the buffer all at once instead of re-allocating memory as we push more chars onto the buffer. Let us walk through what Rust does when we use String::new() and then push characters onto the string.

A String is really a Vec of UTF-8 code points. When String::new() is called, Rust creates a vector with zero bytes of capacity. If we then push the character a onto the string buffer, like input.push('a') , Rust has to increase the capacity of the vector. In this case, it will allocate 2 bytes of memory. As we push more characters and exceed the capacity, Rust will double the size of the string by re-allocating memory. It will continue to double the size each time the capacity is exceeded. The sequence of memory allocation is 0, 2, 4, 8, 16, 32 ... 2^n where n is the number of times Rust detected that capacity was exceeded. Re-allocating memory is really slow (edit: kmc_v3 explained that it might not be as slow as I thought). Not only does Rust have to ask the kernel for new memory, it must also copy the contents of the vector from the old memory space to the new memory space. Check out the source code for Vec::push to see the resizing logic first-hand.

In general, we want to allocate new memory only when we need it and only allocate as much as we need. For small strings, like remove_spaces("Herman Radtke"), the overheard of re-allocating memory is not a big deal. What if I wanted to remove all of the spaces in each JavaScript file for my website? The overhead of re-allocating memory for a buffer is much higher. When pushing data onto a vector (String or otherwise) it can be a good idea to specify a capacity to start with. The best situation is when you already know the length and the capacity can be exactly set. The code comments for Vec give a similar warning.

https://hermanradtke.com/2015/05/29/creating-a-rust-function-that-returns-string-or-str.html

Anonymous

What is a COW type22.04.21

Information Epoch 1713527226