'ComWrappers.RegisterObjectForInterface' should skip a QI if 'inner' is provided

[Sergio0694]

Description

Currently, this method is always querying for inner, to handle back-compat with .NET 5, where there was no overload passing it:

runtime/src/coreclr/interop/interoplib.cpp

Lines 126 to 128 in 493650a

	HRESULT hr = external->QueryInterface(IID_IReferenceTracker, (void**)&trackerObject);
	if (SUCCEEDED(hr))
	checkForIdentity = trackerObject.p;

Now that the overload exists and callers can pass inner, we should be able to skip this QI if we do have a value.

cc. @jkoritzinsky

[dotnet-policy-service]

Tagging subscribers to this area: @dotnet/interop-contrib
See info in area-owners.md if you want to be subscribed.

[AaronRobinsonMSFT]

This isn't a performance issue. This is enabling a scenarios with an API that doesn't provide the needed details. I don't see any reason this should be changed at present.

[Sergio0694]

Jeremy suggested it should be possible while chatting on Discord. The idea is that when inner is already passed by callers, that QueryInterface call isn't actually necessary, so it should be skipped. That's why I used the "performance" label, as the change would just make that method a little bit faster for callers providing that argument explicitly.

[AaronRobinsonMSFT]

Yes, that is true. The reason I avoided that is because (A) you're never going to see that performance impact given all the other things going on in aggregation and (B) the predicates are already complicated and I was trying to avoid checking things multiple times as it makes the flow that much more complicated. If someone can demonstrate a performance benefit, then sure. Otherwise, I would leave it alone.

[Sergio0694]

It might be tricky to precisely isolate the performance improvement here (I mean, I suppose we could put together a microbenchmark perhaps), but in general this would just fall into the broader effort we've been doing in CsWinRT to try to minimize QueryInterface calls wherever possible. @manodasanW did lots of performance investigations on this, and it's also why the codegen was updated a few years ago to explicitly cache all object references for WinRT interfaces (and we're extending that even more in the future). This change feels like it'd fall into the same bucket: avoid as many QueryInterface calls as possible, if we can.

[AaronRobinsonMSFT]

Closing in deference to #113622