Having the closure as a generic parameter (and not type-erasing it internally) leads to a lot of code bloat compared to the corresponding stable APIs that take &dyn Debug as value to format. According to cargo +nightly llvm-lines a simple program formatting a struct with two fields generates 694 lines of LLVM IR while the same program using field() generates 42 lines. So at least with the current implementation, these functions have an annoying downside compared to e.g. .field("foo", &fmt::from_fn(|f| ...)

I think this can be addressed after stabilization, without changing the signatures. But I wanted to flag it so the reviewer can think about it as well. It's not entirely obvious to me for the helpers that deal in FnOnce. I'm aware of one workaround (put the closure into an Option, then pass a &dyn FnMut that unwraps this option to call the original impl FnOnce) but it's not quite zero cost in several dimensions.

Could this be solved by putting the main logic back into field(), and having field_with be an #[inline] wrapper around it?

I don't want to send in any PRs that might cause Git conflicts with the stabilization, but if the public API is good then I'd be happy to experiment with internal reorganization / optimization after both have landed.

field takes a &dyn Debug. How do you invoke a FnOnce or FnMut from the &self argument of Debug::fmt?

Cell<Option<F>> , then .replace(None).unwrap()? It doesn't need to survive more than one call.