Reduce stack size of ShortVec in icu_locid, helping reduce Locale & LanguageIdentifier

[sffc]

Currently, we define ShortVec to be

enum ShortVec<T> {
    Empty,
    Single(T),
    Multi(Vec<T>),
}

This requires 32 bytes on x86_64: Vec<T> is 3*usize (24 bytes), and the discriminant requires one additional word.

Pending some dependencies, for T = NonZeroUsize and smaller, 16 bytes should be achievable as follows:

enum ShortBoxSlice<T> {
    ZeroOne(Option<T>),
    Multi(Box<[T]>),
}

(thanks @Manishearth and @mikebenfield)

Some notes:

1. This depends on the Rust upstream PR Use niche-filling optimization even when multiple variants have data. rust-lang/rust#94075 landing first
2. For this to work for Variant, which is the motivating use case, we need to fix Make Option<TinyAsciiStr> be the same size as TinyAsciiStr #2083 first
3. It hopefully won't be necessary to expand the Box into an explicit NonZeroPtr pointer and NonZeroUsize length, but that is an option we could pursue if needed

[robertbastian]

If we want to go further than the ShortVec regression we could put the Extensions struct on the heap and cut the size of Locale by 75% (216 to 56 bytes with the current ShortVec).

Once on the heap we can make Extensions ?Sized, which would unlock some optimizations that could offset the extra deref, such as storing the most common extension type (unicode?) inside the struct instead of its own allocation.

[robertbastian]

Edited, had a typo that messed up results before. I didn't sleep much.

I just tried this with today's nightly:

enum ShortVec<T> {
    Empty,
    Single(T),
    Multi(Vec<T>),
}

assert_eq!(size_of::<ShortVec<usize>>(), 32);
assert_eq!(size_of::<ShortVec<NonZeroUsize>>(), 32);

enum ShortVec2<T> {
    Empty,
    Single(T),
    Multi(Box<[T]>),
}

assert_eq!(size_of::<ShortVec2<usize>>(), 24);
assert_eq!(size_of::<ShortVec2<NonZeroUsize>>(), 24);

enum ShortVec3<T> {
    ZeroOne(Option<T>),
    Multi(Box<[T]>),
}

assert_eq!(size_of::<ShortVec3<usize>>(), 24);
assert_eq!(size_of::<ShortVec3<NonZeroUsize>>(), 16); // this is 24 on 1.64

[robertbastian]

Don't celebrate, this seems very brittle and won't work with TinyAsciiStr:

https://play.rust-lang.org/?version=nightly&mode=release&edition=2021&gist=89c6184a24634bf80c245ffbcbfc0928

[sffc]

^ @mikebenfield Can you take a look at that playground? Not sure why that niche is not being found.

[sffc]

Oh, wait, it's fine. It should be an array of NonZeroU8 not NonZeroUsize

https://play.rust-lang.org/?version=nightly&mode=release&edition=2021&gist=cd9d9465aa1b58927f7668ac1148e8e5