that sounds like more to string interning for constant string, right? I thought vm can hold a vec of PyStringRef for that.

In this PR, I am trying short string embedding here. It will reuse fat pointer memory allocated for Box.

Ah, okay, that makes sense, I guess I misinterpreted what you meant by embedding

Looking at the source for embed-str -- is there a reason you did (1..=MAX_SIZE).contains(&s.len()) and not s.is_empty()?

Also, when you assume_init() for EmbeddedStr, I don't think that's actually all initialized -- you only copy_nonoverlapping() len bytes, after that I think it's uninitialized. You might want to store EmbeddedStr/EmbeddingStr as a MaybeUninit<[usize; 2]/[u8; SIZE]> and read the mode by reading only the exact byte that actually holds the metadata. Have you tried running the tests with cargo miri? That would probably help with ironing out UB

@Kroisse introduced https://docs.rs/string_cache/ for string interning

Looking at the source for embed-str -- is there a reason you did (1..=MAX_SIZE).contains(&s.len()) and not s.is_empty()?

I think that's compact expression for !s.is_empty() && s <= MAX_SIZE

About mode, i'd better to use exact byte from byte array. But does it actually need to be stored as MaybeUninit? we can access only to initailized bytes and it doesn't have any drop behavior.

I think this PR requires benchmark before merging

I think that's compact expression for !s.is_empty() && s <= MAX_SIZE

Oh, lol, I completely missed the const MAX_SIZE before that, I thought MAX_SIZE was like usize::MAX or something

But does it actually need to be stored as MaybeUninit? we can access only to initailized bytes and it doesn't have any drop behavior.

Yeah, unfortunately, it's still UB even if you're not accessing them, since you can't trust that the compiler won't optimize it to access some of the uninitialized memory -- there's a good article I read about it but I can't find it right now.

It seems that the embed_str crate does not check the invalid unicode points?