Feature: unsafe push_unchecked

[trueb2]

I have a µcontroller that suffers with the push's including a branch on the len check (no branch predictor). I would like to reserve sufficient space, then push unchecked to avoid the overhead. I could similarly accomplish my goals with a spare_capacity_mut interface similar to Vec, but that provides slightly different internal data structure control rather than pushing initialized values through a more normal interface.

An Extend interface could benefit from this as well because it uses push internally. The Extend impl change would not resolve my purpose because it would require the overhead of managing state within the Iterator instance, so a pub push_unchecked would be preferred for me.

Would you be open to a PR with an unsafe push_unchecked impl?

[emilio]

I think it can be implemented using something like:

unsafe fn push_unchecked<T>(v: &mut ThinVec<T>, value: T) {
    debug_assert!(v.capacity() > v.len());
    let end = v.as_mut_ptr().add(v.len());
    std::ptr::write(end, value);
    v.set_len(v.len() + 1);
}

right?

[SOF3]

A push_unchecked() alone does not seem to serve any explicit purpose, but just best effort helping optimizer elide multiple v.len += 1 increments, which causes significant cache misses (edit: len still fits in L1 cache according to iai, but the number of accesses is 40 times greater than std::vec when copying [u16; 1027]). Would it be simpler to just write the entire extended slice and set_len once in extend() to avoid this?