[dzheng]

Aha, that makes sense, thanks!

I agree that the performance benefits of such stack allocation (over heap allocation) aren't quite clear in practice.

I feel the bigger win of delimited cont./closure-based AD approaches is that they can model the control flow of reverse-mode AD without AD-specific code transformations. Delimited cont. is especially great at making things modular: each differentiable function performs primal computation, calls the callback with primal result, then performs adjoint computation.