[API contract regression] SymbolDisplay.ToDisplayString(IntPtrTypeSymbol, FullnameFormat) returns nint instead of System.IntPtr with .NET 8.0 SDK

[ceztko]

Description:

In a simple compilation SymbolDisplay.ToDisplayString(IntPtrTypeSymbol, FullnameFormat), with IntPtrTypeSymbol the type symbol for System.IntPtr and FullnameFormat defined below, returns nint instead of System.IntPtr when compiled depending on .NET 8.0 SDK, despite no SymbolDisplayMiscellaneousOptions.UseSpecialTypes is being used. If the same source is compiled depending on .NET 6.0 SDK the returned value is System.IntPtr, as expected.

FullnameFormat is defined as:

static readonly SymbolDisplayFormat FullnameFormat = new SymbolDisplayFormat(
    globalNamespaceStyle: SymbolDisplayGlobalNamespaceStyle.OmittedAsContaining,
    typeQualificationStyle: SymbolDisplayTypeQualificationStyle.NameAndContainingTypesAndNamespaces,
    localOptions: SymbolDisplayLocalOptions.IncludeType,
    genericsOptions: SymbolDisplayGenericsOptions.IncludeTypeParameters
    | SymbolDisplayGenericsOptions.IncludeVariance,
    memberOptions:
        SymbolDisplayMemberOptions.IncludeContainingType |
        SymbolDisplayMemberOptions.IncludeExplicitInterface,
    miscellaneousOptions:
        SymbolDisplayMiscellaneousOptions.UseErrorTypeSymbolName |
        SymbolDisplayMiscellaneousOptions.ExpandNullable
        // NOTE: SymbolDisplayMiscellaneousOptions.UseSpecialTypes is not used!

Version Used:

Microsoft.CodeAnalysis.CSharp 4.12

Steps to Reproduce:

1. Decompress IntPtrTest.zip and launch the IntPtrTest.sln solution
2. Run the Net6SystemIntPtr project, it will print net6.0 dependency: IntPtr fullname is System.IntPtr;
3. Run the Net8nint project, it will print net8.0 dependency: IntPtr fullname is nint.

Expected Behavior:

I understand that using .NET 8.0 SDK nint is now an alias for System.IntPtr (or the other way round). Despite this SymbolDisplayMiscellaneousOptions.UseSpecialTypes is documented in this way:

/// <summary>
/// Uses keywords for predefined types. 
/// For example, "int" instead of "System.Int32" in C#
/// or "Integer" instead of "System.Integer" in Visual Basic.
/// </summary>
UseSpecialTypes = 1 << 0,

Similarly I expect nint to be a special type, and not having used this flag I expect SymbolDisplay.ToDisplayString(IntPtrTypeSymbol, FullnameFormat) to return System.IntPtr.

Actual Behavior:

Depending on the dependent SDK used for the compilation,SymbolDisplay.ToDisplayString returns a different value for the ubiquitous System.IntPtr type, which is burden when maintaining analyzers that will process code that can be dependent on a unpredictable SDK.

[CyrusNajmabadi]

which is burden when maintaining analyzers that will process code that can be dependent on a unpredictable SDK.

Could you clarify this. I would not expect analyzers to need the display string, or find themselves in a difficult to maintain state due to the display string of a particular symbol. Thanks!

[ceztko]

Could you clarify this. I would not expect analyzers to need the display string, or find themselves in a difficult to maintain state due to the display string of a particular symbol. Thanks!

Sure thing. I mentioned "analyzers", but I really meant anything using Roslyn. I'm maintaining a transpiler for a subset of C# to different languages/frameworks. I'm often in the need of checking against my custom list of "special types". While I understand it may be unreliable in some circumstances, I'm using the SymbolDisplayFormat I defined above to create a string identifier for the type, and I'm actively checking against it in multiple places. Knowing that the approach is not 100% correct (for example the same qualified type name may be defined in multiple assemblies), it works well at my conditions so I would focus on the API contract for SymbolDisplayMiscellaneousOptions.UseSpecialTypes which is enough clear. The situation now is that with the SymbolDisplayFormat above all older "special" type symbols like void, int, float, etc. are still correctly displayed respectively as System.Void, System.Int32, System.Single, etc., while IntPtr, nint or whatever maps to System.IntPtr now is displayed unconditionally nint (but only when targeting >= .NET 8.0) which is odd/slack. My guess here is that for the same reason using SymbolDisplay.ToDisplayString is not the most reliable way to create a identifier for types, a fix here could be done without creating too much concern (handling nuint/UIntPtr as well).

[ceztko]

I did some research and debugging session with the roslyn compiler myself. The branch conditional where Roslyn chooses to display nint instead of System.IntPtr is here. In the branch conditional,SymbolDisplayCompilerInternalOptions.UseNativeIntegerUnderlyingType would also allow to display exactly what I want (System.IntPtr instead of nint). SymbolDisplayCompilerInternalOptions is an internal enum and UseNativeIntegerUnderlyingType is a really little used enum value inside Roslyn code (I could find only 5 references). From a brief look it appears it's used just in unit tests. As I wild guess it could be just removed, and the tests could just ensure the already mentioned public SymbolDisplayMiscellaneousOptions.UseSpecialTypes is not used when they want to read the underlying struct integer type.