GenAPI omits internally defined types and attributes that are relevant to the langugage / Compiler

[sharwell]

Describe the Bug

Reference assembly source code does not include nullability attributes which are required parts of the signatures. For example, these methods failed to include the [NotNullWhen(true)] attribute on the parameters:

https://github.com/dotnet/source-build-reference-packages/blob/4163a212b06740e791c3dfe7ea1258565229af07/src/referencePackages/src/microsoft.codeanalysis.workspaces.common/3.7.0/lib/netstandard2.0/Microsoft.CodeAnalysis.Workspaces.cs#L231-L235

[sharwell]

dotnet/roslyn-analyzers#6705 disables CS8603 in source build scenarios due to this omission.

[MichaelSimons]

FYI @ViktorHofer, @ericstj

[MichaelSimons]

I did regenerate this package to validate it was up to date with respect to the latest generator changes. The code (*.cs) is up-to-date. There are a few csproj/nuspec updates that are irrelevant to this issue.

I couldn't figure out the best area label to add to this issue. If you have write-permissions please help me learn by adding exactly one area label.

[MichaelSimons]

Not sure where in the stack this issue is from. I did check to make sure it is not one of the attributes excluded from SBRP.

[ericstj]

Looks like this is for a netstandard2.0 build of Microsoft.CodeAnalysis.Workspaces. netstandard2.0 did not expose NotNullWhenAttribute and Microsoft.CodeAnalysis.Workspaces itself is defining this attribute as internal. I don't think we want to special case these internal attributes. GenAPI is behaving as expected in excluding them as they are not public API.

I wonder if we exposed an additional way to opt-in certain internal API if that might help in this case.

[sharwell]

I don't think we want to special case these internal attributes.

From the perspective of the compiler, the declared accessibility of these attributes is irrelevant. GenAPI is producing a reference assembly which does not expose the same public surface area of the original assembly.

It may be beneficial to have an opt-in for internal APIs, but that opt-in should not apply to attributes which affect compiler behavior for public APIs even when they are declared internal.

[ericstj]

The same thing is being hit with CollectionBuilderAttribute which was added to ImmutableCollections. This is another case of a user-authored internal attribute that matters to the compiler.

We should add support for inclusion of specific internal types (perhaps with a default set).