[embind] Fix misindexed template parameters which prevented policies from being applied to class functions or constructors

[canislupaster]

Hi, trying to allow_raw_pointers() in a class method bound with embind when a custom signature was given as a template argument to .class_function failed. It's easy to see why by comparing the definition of either RegisterClassMethod or RegisterClassConstructor with their usage:

template<typename ReturnType, typename ThisType, typename... Args>
struct RegisterClassMethod<ReturnType (ThisType, Args...)> {

    template <typename ClassType, typename Callable, typename... Policies>
    static void invoke(const char* methodName,
...

and its invocation in class_function:

using invoker = internal::RegisterClassConstructor<
            typename std::conditional<std::is_same<Signature, internal::DeduceArgumentsTag>::value,
                                      Callable,
                                      Signature>::type>;

invoker::template invoke<ClassType, Policies...>(callable);

which will skip the Callable template argument in some specializations (and Callable will become the return value policy). To fix this, I used the same approach used by the getters/setter PropertyTags (c.f. diff). Please let me know if anything is off.

[brendandahl]

Looks good. I think you'll need to run
test/runner other.test_minimal_runtime_code_size_hello_embind* --rebaseline
to fix up the code size tests.

[canislupaster]

I'm not sure if something with my environment is different from CircleCI's (MacOS / Apple Silicon, latest emsdk, no warnings about version mismatches for clang/binaryen/etc), but I get the old/incorrect code size locally. This seems like a terrible solution, but I can substitute the "correct" sizes from #24055 .

[sbc100]

I'm not sure if something with my environment is different from CircleCI's (MacOS / Apple Silicon, latest emsdk, no warnings about version mismatches for clang/binaryen/etc), but I get the old/incorrect code size locally. This seems like a terrible solution, but I can substitute the "correct" sizes from #24055 .

Did you do emsdk instal tot to get the tip-of-tree release?