Figure out how to use push_unchecked in more places

[Lucretiel]

There's an issue in which using push_unchecked in loops, even with a fixed number of iterations, causes inferior assembly to be generated. For example:

staticvec/src/lib.rs

Lines 116 to 133 in 4b2534f

	#[inline]
	pub fn filled_with<F>(mut initializer: F) -> Self
	where F: FnMut() -> T {
	let mut res = Self::new();
	//You might think it would make more sense to use `push_unchecked` here.
	//Originally, I did also! However, as of today (November 19, 2019), doing so
	//both in this function and several others throughout the crate inhibits the ability
	//of `rustc` to fully unroll and autovectorize various constant-bounds loops. If this changes
	//in the future, feel free to open a PR switching out the manual code for `get_unchecked`, if
	//you happen to notice it before I do.
	for i in 0..N {
	unsafe {
	res.data.get_unchecked_mut(i).write(initializer());
	res.length += 1;
	}
	}
	res
	}

The workaround is to manually push by writing directly to self.data and incrementing self.length. However, we'd prefer to minimize the amount of unsafe code we have to write, and we'd rather use "safer unsafes" like push_unchecked that provide a friendlier contract than self.data.get_unchecked_mut().write.

[slightlyoutofphase]

I honestly think it has something to do with the fact that MaybeUninit::write in fact has a return value, which is a mutable reference to whatever you just wrote to it.

The docs state this is for your convenience, although IMO it seems like a somewhat strange API. So it may even be that using regular pointer writes in the troublesome spots turns out to work more efficiently (although IIRC that left me with Miri errors at some point early on, because I guess in some cases it wasn't considered the "same" as using the built-in MaybeUninit::write, although generally it should be.)

I'll experiment with some things when I have a bit more time later on to see if I can arrive at a solution that's both fast and sound.

[slightlyoutofphase]

I'd say push_unchecked is now basically being used where it makes sense to use it (except for filled_with still, but that seems to be a particularly special case.)