This seems like the string version of --ffast-math. Optimize performance at the cost of everything else[1]. Except with --fast-math you explicitly turn it on.
I assume the underlying storage isn't reference counted, because if it was it seems it would be trivial to detect there's only a single reference left, and trigger a reallocation.
But having only one reference is rare, often you extract several pieces of data from a huge string.
Unless the underlying storage is reference counted with some small granularity. That might be a way to go, when a substring is taken, split the original and keep track of the pieces separately.
I assume the underlying storage isn't reference counted, because if it was it seems it would be trivial to detect there's only a single reference left, and trigger a reallocation.
[1]: https://gcc.gnu.org/wiki/FloatingPointMath