Stub generator emits assembly-qualified .NET names in Java class descriptors #735

I hit another problem with converted dlls. I was able to fix it locally with ChatGPT and wanted to share the fix here.

Problem

When IKVM generates stubs for converted assemblies, class-valued annotation metadata can contain a .NET assembly-qualified type name instead of a valid Java descriptor.

For example, metadata intended to contain:

Ljava/lang/Thread;

may instead contain:

Ljava/lang/Thread, IKVM.Java, Version=..., Culture=..., PublicKeyToken=...;

The comma and assembly metadata make the generated descriptor invalid.

Cause

StubGenerator encodes class-valued annotation elements directly from the stored type string. Some of these strings originate from Type.AssemblyQualifiedName, so they include .NET assembly metadata.

Required changes

In src/IKVM.Runtime/StubGen/StubGenerator.cs, add a helper that detects assembly-qualified descriptors, resolves the underlying IKVM Java class, and returns its canonical Java descriptor:

static string DecodeTypeName(string typeName)
{
#if !FIRST_PASS && !EXPORTER
    int index = typeName.IndexOf(',');
    if (index > 0)
    {
        // Resolve an assembly-qualified .NET type name back to
        // its canonical Java class descriptor.
        try
        {
            typeName =
                "L" +
                java.lang.Class
                    .forName(typeName.Substring(1, typeName.Length - 2).Replace('/', '.'))
                    .getName()
                    .Replace('.', '/') +
                ";";
        }
        catch
        {
            // Preserve the original value if it cannot be resolved.
        }
    }
#endif

    return typeName;
}

For FIRST_PASS and EXPORTER builds, the helper can remain a no-op if required by the existing conditional compilation structure:

#if FIRST_PASS || EXPORTER
static string DecodeTypeName(string typeName) => typeName;
#endif

Pass type descriptors through DecodeTypeName before adding them to the constant pool.

For object-array annotation data, update the class, enum, and nested-annotation cases:

encoder.Enum(
    builder.Constants.GetOrAddUtf8(DecodeTypeName((string)v[1])),
    builder.Constants.GetOrAddUtf8((string)v[2]));

encoder.Class(
    builder.Constants.GetOrAddUtf8(DecodeTypeName((string)v[1])));

e.Annotation(
    builder.Constants.GetOrAddUtf8(DecodeTypeName((string)v[1])),
    e2 =>
    {
        // Existing element encoding
    });

Also update the CustomAttributeTypedArgument TAG_CLASS path:

encoder.Class(
    builder.Constants.GetOrAddUtf8(
        DecodeTypeName((string)data[1].Value)));

Expected result

Assembly-qualified .NET names are normalized before being written to Java annotation metadata. Generated output contains valid descriptors such as Ljava/lang/Thread;, without the trailing comma, assembly name, version, culture, or public-key token.