You might wonder why we need to micro-optimize this stuff and the answer is that this is something like 2% of all Herbie allocations. Working on the PR.

The require/typed will be pretty slow here, we should just fix that in TR directly.

How does one do that? In theory there's an unsound version of flbit-field which the compiler could produce that would be super fast.

You'd want to add it somewhere around here: https://github.com/racket/typed-racket/blob/c08e1a4fbbb6382d71fe3f4cc9362c6d588f71ef/typed-racket-lib/typed-racket/base-env/base-env-numeric.rkt#L2005

Done in racket/typed-racket#1484. How do we typically do this? Do we merge that and then remove the require/typed here and just expect it all to go in at once?

We should just merge the TR PR (possibly with a version bump) and then merge this one without the require/typed.

We could also add an optimization to transform flbit-field to unsafe-flbit-field when we know that the two fixnum arguments are in [0,63] (because they're usually going to be constants).

Yeah, I couldn't (in a cursory look) figure out how the optimization stuff worked so I didn't touch that, but yes, in practice they'll always be constants so one could detect the results and optimize. In theory you could also detect constants and use that for type checking (to assert that it's an Index if the bounds are tight enough), dunno if that gets you anything but maybe.