Increase foldability of various templates

[jonwis]

SizeBench revealed heavily non-foldability of certain C++/WinRT templates contributing to increased Windows binary sizes. Some small tweaks dropped the size of test.exe (release, x64) by ~132kb. (5180416 -> 5030400, or 2.8% ... which isn't a super lot, but it's nonzero.

Fixes #1334 and related

Changes

Enable nuspec iteration - Bump the default nuspec version to 3.0.0.0 so build_nuget.cmd and iterating with a local nuget source is easier.

Add solution files - Niceness for people who live in Visual Studio.

Fold delegate bases - Forced the layout of the delegate types to be regular by moving the reference count into a base before the non-ABI types. This folds nearly all the AddRef/QueryInterface instances into one. Releases are still unique because they call up into dtors.

Fold "report" in invoke - The invoke helper is a try/catch{report} but the body of "report" is identical on each invoke specialization. Move it to a unique non-inlineable method.

Fold event::add - Make the delegate first, then all the "expand the collection" instances are identical.

Further Work

SizeBench reports the following additional foldability wastage, quoted from the same test.exe:

• make_agile_delegate<> - 3.2kb; the agile thunk is unique basically per delegate ::Invoke vtable and offset adjusters.
• root_implements::<T1,T2,T3>::query_interface - 4.1kb; again unique per vtable instance, also related to inlining choices
• winrt::event::add - 3.2kb; all cost related to inlined use of make_agile_delegate
• winrt::impl::runtime_class_name<T1,T2>::get() - 1.4kb; should probably switch to std::wstring_view but this may be a breaking change; currently induces an hstring construction.

Testing

Tested by build_and_test.cmd - all built, all passed

[jonwis]

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.

[jonwis]

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%.

[jonwis]

/azp run

[azure-pipelines[bot]]

No pipelines are associated with this pull request.

[kennykerr]

Overall looks like goodness. Just a few comments.

[kennykerr]

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.

[sylveon]

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

[kennykerr]

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.