Casting performance issues

[PascalBoeschoten]

Hello, I'm using pybind in some performance sensitive code and profiling show pybind hammering two casting-related functions pretty hard, get_type_info and try_load_foreign_module_local:

pybind11/include/pybind11/detail/type_caster_base.h

Line 216 in 296615a

PYBIND11_NOINLINE detail::type_info *get_type_info(const std::type_index &tp,

pybind11/include/pybind11/detail/type_caster_base.h

Line 652 in 296615a

PYBIND11_NOINLINE bool try_load_foreign_module_local(handle src) {

While try_load_foreign_module_local takes up a significant amount of runtime it always ends up returning false at the first branch. Is this indicating some issue in my code? Would it be risky to remove/disable this function? (E.g. replacing the implementation with just a return false)

For get_type_info I think in my situation this function could really benefit from a small cache since the hot code tends to use a small number of types repeatedly. Is there a safety/correctness reason why this function doesn't cache already?