Why does Vec::retain run slower after updating to Rust 1.38.0?

Question

After updating Rust to version 1.38.0 from 1.36.0, I noticed that my program runs slower — by about 50%.

Using perf, I discovered that half of the program time is spent in alloc::vec::Vec<T>::retain in the new version. In the older version, this function does not even show up. Why would retain take so much longer in 1.38.0?

The call to retain is done like this:

some_vec.retain(|&x| x < DEADLINE);

deadline is a constant u32 and some_vec is a Vec<u32>.

I ran the program without the retain calls in both versions. In this case, 1.38.0 was still slower on average, but only by ~10% instead of the > 50% seen before.

To recap what happened in the tests:

Version 1.36.0

• with retain: ~18sec
• without retain: ~11sec

Version 1.38.0

• with retain: ~28sec
• without retain: ~12sec

For a reproducible example, you can try:

use std::time::Instant;

fn main() {
    let start = Instant::now();
    let mut my_vec: Vec<u32>;
    for _ in 0..100_000 {
        my_vec = (0..10_000).collect();
        my_vec.retain(|&x| x < 9000);
        my_vec.retain(|&x| x < 8000);
        my_vec.retain(|&x| x < 7000);
        my_vec.retain(|&x| x < 6000);
        my_vec.retain(|&x| x < 5000);
        my_vec.retain(|&x| (x < 5) & (x > 2));
    }
    let duration = start.elapsed();
    println!("Program took: {:?}", duration);
}

With cargo +1.36.0 run --release and then cargo +1.38.0 run --release.

For this small example, I got:

$ cargo +1.36.0 run --release
Program took: 4.624297719s

$ cargo +1.38.0 run --release
Program took: 8.293383522s

Answer 1

In general, rust.godbolt.org is useful for checking quality of the generated code (but don't forget to add optimization flags!)

In your case code generated for retain has clearly changed for worse: https://rust.godbolt.org/z/ZhVCDg

So you should report that to Rust as a performance regression.