Yes, something like that would work.

When built against nanobind (+ #116), this provides a modest speedup for the microbenchmark on M1:

• pybind11: 85s
• nanobind in Switch to nanobind #107: 57s
• With nanobind: optimize build #116: 45s
• With nanobind: optimize build #116 and this PR: 42s

I'm OK with option b) in both cases. I'm not sure option a) would work unmodified. You would likely need to call the casting function manually somewhere in there in order to get it to work, and this may be incrementally faster than option b). But a lot depends on exactly how much faster, and whether that justifies the extra effort (and generated code).

At the same time, there's a small amount of speed-up here, and we're net deleting code: I'd call it a win regardless and would be happy to merge this, I think.

I'm OK with option b) in both cases. I'm not sure option a) would work unmodified. You would likely need to call the casting function manually somewhere in there in order to get it to work, and this may be incrementally faster than option b). But a lot depends on exactly how much faster, and whether that justifies the extra effort (and generated code).

Option a) works in the way it is written above (possibly also due to making use of implicitly_convertible?). Both options are very close performance-wise in my test.

Thanks!