REFL037 should not warn here

[JohanLarsson]

typeof(BinaryExpression).Assembly.GetType("System.Linq.Expressions.AssignBinaryExpression", throwOnError: true)

Invisible strikes again?

[JohanLarsson]

typeof(AdornedElementPlaceholder).Assembly.GetType("MS.Internal.Controls.TemplatedAdorner", throwOnError: true);

[brian-reichle]

Maybe the only thing you can assert is that the 'type name' is well formed :)

[JohanLarsson]

Yes we need to disable it for now, it would be a useful analyzer so nice if we can find a way to make it work.
Maybe we can re-enable it when we upgrade to Roslyn 3.x and we can detect if internals are visible.

I've been thinking about a huge hack where we pack a small exe in the analyzer dll and use the exe to dump the internals using standard reflection.

[jnm2]

The analyzer needs to be able to quantify what it doesn't know. Right now it's based on whether it can prove the member exists, but it should be based on whether it can prove the member doesn't exist.

Scenarios it needs to detect to properly quantify the level of uncertainty, in order:

• Is there any chance that this is a reference assembly? ❌ Cannot prove nonexistence.
• Is there any chance that this is an assembly reference that does not end up in build output?
  High chance it will run against a different version with a different set of private members, e.g. the BCL. ❌ Cannot prove nonexistence.
• Is the GetX search limited to BindingFlags.Public? Use Roslyn's semantic model. ✔ Can prove nonexistence.
• Can you read compilation.Options.MetadataImportOptions via reflection? If so:
  • Is it set to All? Use Roslyn's semantic model. ✔ Can prove nonexistence.
• Can you obtain a MetadataReader or reflection-only Assembly object? ✔ Can prove nonexistence.
• Otherwise: ❌ Cannot prove nonexistence.

REFL037 should only be reported when you reach a green checkmark and prove that the member does not exist. If the path to the green checkmark is not implemented yet, REFL037 shouldn't light up for that scenario yet.

[jnm2]

Wait, it's not a [Flags] enum.

compilation.Options.MetadataImportOptions tells you whether you can stop at the second green checkmark at the top, but it also has three levels, not two. It sometimes reports internal members without reporting private members:

    /// <summary>
    /// Specifies what symbols to import from metadata.
    /// </summary>
    public enum MetadataImportOptions : byte
    {
        /// <summary>
        /// Only import public and protected symbols.
        /// </summary>
        Public = 0,

        /// <summary>
        /// Import public, protected and internal symbols.
        /// </summary>
        Internal = 1,

        /// <summary>
        /// Import all symbols.
        /// </summary>
        All = 2,
    }

https://github.com/dotnet/roslyn/blob/f792750b7c2a052fb07e349221b6a99eac6b475c/src/Compilers/Core/Portable/MetadataReader/MetadataImportOptions.cs#L7-L26

[JohanLarsson]

Compliation exposes it with a public property in more recent roslyn versions.

[jnm2]

I'd expect the analyzer to either skip that step then, or else only skip that step if you can't find and invoke the members via reflection.

[JohanLarsson]

Yes, I'll disable this analyzer for now until we have better handling. As it is now it is 100% annoying.

[jnm2]

I updated the logic that I'd use.

[JohanLarsson]

I had a horrible bug, closed this issue with the fix for that. We should resurrect the discussion as it is still valid.