Add niche to ColdString to make Option<ColdString> the same size as ColdString

[luksan]

This changes the type of the internal "encdoded" from *const u8 to NonNull in order to create a niche so that Option has the same size as ColdString. To handle storing of WIDTH NUL bytes the NULL pointer is mapped to a static [0u8; WIDTH] array.

[tomtomwombat]

Can this handle "\0\0\0\0\0\0\0\0"?

[tomtomwombat]

I see, it represents "\0\0\0\0\0\0\0\0" as Self::PTR_TAG. Unfortunately, is_inline returns false for "\0\0\0\0\0\0\0\0". Is there a particular reason you picked Self::PTR_TAG to represent 8-null? Why not a new representation that's invalid for every other string, like usize::MAX? usize::MAX's first byte matches the inline tag 11111xxx, but is invalid UTF-8, so it can be treated as 8 null. There might be a better choice....

[luksan]

Since rust doesn't support custom niches in types NonNull is the only reasonable internal type. That means that we can't store an 8 byte all NUL str inline, so it has to be that value that is special cased.

[tomtomwombat]

I'm with you there. But, is Self::TAG_PTR the only reasonable 8-nul representation?

I'm suggesting

fn new_eight_nul() -> Self {
    // SAFETY: PTR_TAG is non-zero
    unsafe { Self::from_inline_buf(usize::MAX.to_ne_bytes()) }
}

fn inline_len(&self) -> usize {
    let addr = self.addr();
    match addr & Self::INLINE_TAG {
        Self::INLINE_TAG if addr != usize::MAX => (addr & Self::LEN_MASK).rotate_right(Self::ROT),
         _ => WIDTH,
    }
}

so new_eight_nul().is_inline() == true and we don't handle the 8 nul case in heap related functions: heap_ptr, decode_heap etc.

Would that work?

[luksan]

And having an invalid utf8 string as representation for 8NUL means that we can't return an &str to it, so some kind of tag and indirection is needed. I don't see any disadvantage of considering as a pointer, and then doing null ptr checks, as compared to having a completely different codepath for the niches.