Source Generation

Arch Forum — 2026-02-12

Basics

  • Generates code at compile time, with access to the same information as the compiler
  • We already use it:
    • Mapperly – object mapping
    • Regex source generators – docs

Why is this good?

  • Less boilerplate for common concerns (Firebase, mapping, regex)
  • Generated code is just normal C# – compiled like everything else
  • Slight performance advantage over runtime reflection
  • Opens the door to AOT compilation

Problems with Castle.Core

  • Handles things at runtime – hard to know what it does without debugging
  • Even debugging is difficult to follow
  • We are stuck on an old version; upgrading is risky because the code is hard to understand
  • The overall .NET ecosystem trend is away from runtime reflection

Understanding & Debugging

Not everyone needs to understand the generator code itself.

Everyone can inspect the generated output – it’s plain C# – making debugging much easier.

Specifics

  • Source generators are treated as Analyzers (even when they do more than analyze)
  • The generator must target netstandard2.0
    • All dependencies must also support netstandard2.0
  • Debugging the generator itself is not perfectly straightforward
  • Two main ways to emit code:
    • Basic strings written to .cs files (what we use – and the most popular method)
    • Typed SyntaxFactory
  • Generators are additive only – they can add new code but never modify existing user code

Our Use Cases

  1. Firebase
  2. SqlRepository
  3. RestClient
  4. Controllers

Firebase Generator

~400 lines of code

  1. Find all interfaces that need code generated
  2. For each interface:
    1. Create a file named after the interface without the I prefix, with a .g.cs extension
    2. Write class header (namespace, usings etc)
    3. Loop through methods & generate a method body for each
    4. Write class footer

SqlRepository Generator

  • PR is done
  • ~600 lines of code – ~200 lines less than the old Castle.Core implementation
  • 1–2 days of AI-assisted work to write
  • Haven’t inspected the generator code closely
  • Not fully tested yet, but looks promising

When not to use a generator

Source generators are powerful, but in many cases it’s even more clear to just write the code directly.

Demo

  1. Look at the generator
  2. Look at the generated code

End.