It looks like there's a bunch of "garbage collection" type activity where you may have some bytes which were once part of a string but aren't now used, and you're always paying for the overhead of this optimisation even if it's useless for your problem.

Suppose you work only with 500-4000 byte strings, maybe they're short reviews, and each ends with a rating in star emoji, ***** is the best * is the worst. [[HN ate my star emoji of course]]

So your reviews never fit in the "optimised" string slot, but also the prefix is just opening words from a review, which in some review styles will be the start of seemingly unrelated anecdotes. "My grandfather used to tell me" it'll get to the review eventually, and you'll see why they're connected, but the suffix is useful and that's not stored in a "German string" data structure.

Or maybe you have a high turnover of somewhat related medium size strings, so then that garbage collection step costs quite a lot of overhead.

Any code built closely around String's power-of-2 reallocation pattern may have to be reworked. I don't think there's any case when it cannot be used as a String replacement at all, except maybe when interfacing with an API that expects a &mut String as an output parameter.

My uninformed guess is that at least it will cost you the branching because you need to check if it is inlined or not, and you pay that for every string. Branch prediction is likely very good for this case though.