Optimize for small strings?

[pchampin]

Inspired by what is done in EcoString, MownStr could sacrifice one more bit in the len metadata to "internalize" small (≤15B) strings.

More precisely:

• xlen and addr would have to be stored in that order in memory;
• if the first bit of xlen is 0, addr points to borrowed data (like a &str), and xlen is the length;
• otherwise,
  • if the 2nd bit of xlen is 0, addr points to owned data (like a box), and xlen (without its first bit) is the length;
  • otherwise, the first byte of xlen (without its first two bits) is the length, and the remaining 7 bytes of xlen + the 8 bytes of addr are interpreted as a buffer of u8.

NB: this makes the assumption that xlen and addr are 64bits long... the code would actually need to be adapted for other architectures.

Maybe that's overly complicated for the potential benefit...

[pchampin]

thinking more about this, I'm reluctant to make that change. It requires to know precisely the layout of the internal struct (is the len stored before or after the pointer physically?), the endian-ness of usize (is the most significant bit really at the "beginning" of the usize?)...

And the recent experience with debugging UB (#5, #8) makes me wary of making the code more complex.