Reference interferes with optimization

[tspiteri]

(This comes from a Reddit thread.)

In the code below, slow takes double the time fast takes. If print is changed to take a value rather than reference, the difference goes away.

fn print(s: &u64) {
    println!("{}", s);
}
fn fast() {
    let mut s: u64 = 0;
    for x in 0..10000000000 {
        if x % 16 < 4 {
            s += x;
        }
    }
    let s = s;
    print(&s);
}
fn slow() {
    let mut s: u64 = 0;
    for x in 0..10000000000 {
        if x % 16 < 4 {
            s += x;
        }
    }
    print(&s);
}
fn main() {
    if std::env::var("FAST").is_ok() {
        fast();
    } else {
        slow();
    }
}

Timings:

$ rustc -O main.rs
$ time FAST=1 ./main
12499999983750000000

real	0m4.334s
user	0m4.328s
sys	0m0.001s
$ time ./main
12499999983750000000

real	0m8.788s
user	0m8.776s
sys	0m0.002s

[smmalis37]

This sounds somewhat similar to https://medium.com/@robertgrosse/how-copying-an-int-made-my-code-11-times-faster-f76c66312e0f

[krdln]

Could it also be a similar thing that enables the assert_eq optimization?

[memoryruins]

The fast and slow functions now generate effectively the same codegen (with a minor difference, godbolt); the current nightly on godbolt is rustc 1.64.0-nightly (06754d885 2022-07-08).

Based on the issue's creation date, stable rustc at that time was 1.32. Because the codegen for this example improved from 1.56 onward, I presume this was fixed by the upgrade to LLVM 13.