Performance of BinaryHeap::into_sorted_vec vs. Vec::sort_unstable

[ghost]

I think the implementation of BinaryHeap::into_sorted_vec should just use Vec::sort_unstable instead of sorting by repeated Heap operations. For very small N (N < 1000), the current implementation is maybe 5% faster some times, but for bigger N Vec::sort_unstable will be more than 10 to 30 times faster.

Also, I assume sorting by Heap operations requires 2*N*Log2(N) key comparisons in average and worst case when using the default Heapsort like in the standard library, while Vec::sort_unstable probably requires less than 1.5*N*Log2(N) key comparisons in average case and also a lower number of moves in average case.

I don't have good statistics, maybe others could add some.

Here is an example, to show the performance difference:

/*
[dependencies]
fastrand = "2.0.0"
*/

use std::{collections::BinaryHeap, time::Instant};

fn main() {
    let N: usize = 10_000_000;
    let mut a_elements_in_random_order: Vec<usize> = Vec::from_iter(0..N);
    fastrand::shuffle(&mut a_elements_in_random_order);
    let mut heap: BinaryHeap<usize> = BinaryHeap::with_capacity(N);
    for i in 0..N {
        heap.push(a_elements_in_random_order[i]);
    }
    // both Vecs are now identical
    let mut a_sorted_by_sort_unstable = heap.clone().into_vec();
    let now = Instant::now();
    a_sorted_by_sort_unstable.sort_unstable();
    let runtime_sort_unstable = now.elapsed();
    let now = Instant::now();
    let a_sorted_by_heap = heap.into_sorted_vec();
    let runtime_sorted_by_heap = now.elapsed();
    assert!(a_sorted_by_sort_unstable == a_sorted_by_heap);
    println!("runtime_sort_unstable={:?}", runtime_sort_unstable);
    println!("runtime_sorted_by_heap={:?}", runtime_sorted_by_heap);
}

N=10_000_000
runtime_sort_unstable=403.8169ms
runtime_sorted_by_heap=4.6621243s

N=100_000_000
runtime_sort_unstable=3.6869678s
runtime_sorted_by_heap=95.9239879s

[scottmcm]

cc #76234 (comment), which had a previous conversation about how into_sorted_vec is kinda weird as an API.

[finnbear]

For my WebAssembly client use-case, BinaryHeap::from(vec).into_sorted_vec() generates ~10X smaller code than vec.sort_unstable() while having good asymptotic complexity (unlike the similarly small custom insertion-sort).

Note: slice::heapsort from feature(sort_internals) is an option but that's an internal rustc feature, requires nightly, and has extraneous bounds checks.