Optimize factory caching for apps and pinned DLLs

[kennykerr]

C++/WinRT provides WINRT_NO_MODULE_LOCK that may be used by apps or pinned DLLs (that don't support unloading) to avoid the overhead of tracking of a module reference count. This update further optimizes the factory cache in such cases since unloading has an impact on caching performance. With this change, C++/WinRT is on par with the Rust language. Previously, Rust was just slightly faster.

Data from: https://github.com/kennykerr/lang-perf

cc @axelandrejs

[JamesMcNellis]

LGTM

[jonwis]

A DLL that compiles with this should be super careful to pin itself, yes? Is there a "pin" type that should be turned on to grab and leak a DLL reference, or is that up to the DLL to do?

[kennykerr]

C++/WinRT will do this automatically. If you define WINRT_NO_MODULE_LOCK then C++/WinRT's implementation of DllCanUnloadNow will always return false and you are effectively pinned. Such DLLs can also not export DllCanUnloadNow as added insurance.