Comparing Performance: Loops vs. Iterators

To determine whether to use loops or iterators, you need to know which implementation is faster: the version of the search function with an explicit for loop or the version with iterators.

We ran a benchmark by loading the entire contents of The Adventures of Sherlock Holmes by Sir Arthur Conan Doyle into a String and looking for the word the in the contents. Here are the results of the benchmark on the version of search using the for loop and the version using iterators:

test bench_search_for  ... bench:  19,620,300 ns/iter (+/- 915,700)
test bench_search_iter ... bench:  19,234,900 ns/iter (+/- 657,200)

The iterator version was slightly faster! We won’t explain the benchmark code here, because the point is not to prove that the two versions are equivalent but to get a general sense of how these two implementations compare performance-wise.

For a more comprehensive benchmark, you should check using various texts of various sizes as the contents, different words and words of different lengths as the query, and all kinds of other variations. The point is this: iterators, although a high-level abstraction, get compiled down to roughly the same code as if you’d written the lower-level code yourself. Iterators are one of Rust’s zero-cost abstractions, by which we mean using the abstraction imposes no additional runtime overhead. This is analogous to how Bjarne Stroustrup, the original designer and implementor of C++, defines zero-overhead in “Foundations of C++” (2012):

In general, C++ implementations obey the zero-overhead principle: What you don’t use, you don’t pay for. And further: What you do use, you couldn’t hand code any better.

As another example, the following code is taken from an audio decoder. The decoding algorithm uses the linear prediction mathematical operation to estimate future values based on a linear function of the previous samples. This code uses an iterator chain to do some math on three variables in scope: a buffer slice of data, an array of 12 coefficients, and an amount by which to shift data in qlp_shift. We’ve declared the variables within this example but not given them any values; although this code doesn’t have much meaning outside of its context, it’s still a concise, real-world example of how Rust translates high-level ideas to low-level code.

let buffer: &mut [i32];
let coefficients: [i64; 12];
let qlp_shift: i16;

for i in 12..buffer.len() {
    let prediction = coefficients.iter()
                                 .zip(&buffer[i - 12..i])
                                 .map(|(&c, &s)| c * s as i64)
                                 .sum::<i64>() >> qlp_shift;
    let delta = buffer[i];
    buffer[i] = prediction as i32 + delta;
}

To calculate the value of prediction, this code iterates through each of the 12 values in coefficients and uses the zip method to pair the coefficient values with the previous 12 values in buffer. Then, for each pair, we multiply the values together, sum all the results, and shift the bits in the sum qlp_shift bits to the right.

Calculations in applications like audio decoders often prioritize performance most highly. Here, we’re creating an iterator, using two adapters, and then consuming the value. What assembly code would this Rust code compile to? Well, as of this writing, it compiles down to the same assembly you’d write by hand. There’s no loop at all corresponding to the iteration over the values in coefficients: Rust knows that there are 12 iterations, so it “unrolls” the loop. Unrolling is an optimization that removes the overhead of the loop controlling code and instead generates repetitive code for each iteration of the loop.

All of the coefficients get stored in registers, which means accessing the values is very fast. There are no bounds checks on the array access at runtime. All these optimizations that Rust is able to apply make the resulting code extremely efficient. Now that you know this, you can use iterators and closures without fear! They make code seem like it’s higher level but don’t impose a runtime performance penalty for doing so.

Summary

Closures and iterators are Rust features inspired by functional programming language ideas. They contribute to Rust’s capability to clearly express high-level ideas at low-level performance. The implementations of closures and iterators are such that runtime performance is not affected. This is part of Rust’s goal to strive to provide zero-cost abstractions.

Now that we’ve improved the expressiveness of our I/O project, let’s look at some more features of cargo that will help us share the project with the world.