While you’re at it, allow for a few more ulp (units of least place) error for floating point ops.

Kidding aside, my understanding is this sort of thing cannot be microcode patched.

But I would be pleased to be proven wrong.

It can be patched, see example in "constant-time hardware division" section in https://misc0110.net/files/cpu_woot23.pdf (code: https://github.com/pietroborrello/CustomProcessingUnit/blob/...). You probably won't observe any improvement unless you know an algo massively better than reciprocals+newton.