For posterity, switching name_of<> and get_name<> to use std::wstring_view was basically a net wash. No test code compilation errors that would signal a breaking change, overall binary increased by about 512 bytes.

One area of interest is in GetRuntimeClassName. The inliner is super aggressive about splatting winrt::hstring(anything) inline (and ~hstring) rather than hoisting them out. In PGO-optimized binaries this ctor is normally de-inlined since it's replicated countless times throughout the code. Making a custom winrt::hstring get_runtime_class_name(std::wstring_view) would likely not really help, as PGO should automatically produce that that when used.

It looks like I can get another 8kb by regularizing the tail of query_interface. Running total after that is ~140kb.

It also looks like the cluster around make_agile_delegate could be connected to a simplification on how delegates are generated. Delegate ctors have a lot of overloads - lambda, function, raw-pointer member, com_ptr member, weak_ref member, shared_ptr member, weak_ptr member. Under the covers they all route through the lambda form, doing various forms of "lock" and "call method." I wonder if there's a simplification of those to a two-form version. Unless we were willing to try some stronger type-erasure through a std::function-alike, I'm not sure this can be reduced further. It's probably another 3.2kb on this test binary, so under 0.5%.

/azp run

Overall looks like goodness. Just a few comments.

One area of interest is in GetRuntimeClassName. The inliner is super aggressive about splatting winrt::hstring(anything) inline (and ~hstring) rather than hoisting them out. In PGO-optimized binaries this ctor is normally de-inlined since it's replicated countless times throughout the code. Making a custom winrt::hstring get_runtime_class_name(std::wstring_view) would likely not really help, as PGO should automatically produce that that when used.

For windows-rs I managed to write a "constexpr" version of hstring so that it can be declared as a compile-time constant. This is mainly used for literals, but it also turns these GetRuntimeClassName into a cheap pointer copy. There are some limitations, so I'm not sure its super practical for cppwinrt to adopt though.

Making the constructor of param::hstring constexpr seems pretty trivial, and it would enable constexpr fast-pass strings.

Making the constructor of param::hstring constexpr seems pretty trivial, and it would enable constexpr fast-pass strings.

I was referring to return values, which has a lot more optimizing value.