Box::new() performance regression between 1.18 and 1.19 beta

[rom1v]

It seems there is a performance regression related to Box::new() between 1.18 and 1.19 beta.

use std::time::Instant;

fn main() {
    let timer = Instant::now();
    for _ in 0..100000 {
        Box::new([0; 1000]);
    }
    println!("{:?}", timer.elapsed());
}

In Release/Stable:

rustc 1.18.0 (03fc9d622 2017-06-06)

Duration { secs: 0, nanos: 107624 }

In Release/Beta, it's ~200× slower:

rustc 1.19.0-beta.1 (a87984118 2017-06-06)

Duration { secs: 0, nanos: 22355940 }

https://is.gd/AaDx3A

[crumblingstatue]

I tested on nightly, it's about 140 times slower than 1.18 stable on my system.

[jonas-schievink]

You're not measuring Box::new, you're measuring the compiler's ability to remove unnecessary allocations. For some reason 1.19+ isn't able to eliminate the allocations, while 1.18 is.

[wesleywiser]

I think this is related #24194 (comment)

[jonas-schievink]

Fix is #42410

[joshlf]

You're not measuring Box::new, you're measuring the compiler's ability to remove unnecessary allocations. For some reason 1.19+ isn't able to eliminate the allocations, while 1.18 is.

You could use test::black_box to prevent this optimization to make sure that that's really the issue.

[nagisa]

Closing as a duplicate of #24194