My ".NET GC written in C#" project is coded by hand, bit by bit, as an educational project. But I couldn't resist the temptation of having Fable "speedrun" the project and show me a glimpse of what it could look like if I had more time.
So I let Fable work autonomously on it, with a few constraints:
- Mark&Sweep GC, no compaction
- Supports 64-bit applications only
- Using only the GC API, can't fork the runtime (which means using the default write barrier, and not being able to use a read barrier)
- And of course, in C#
Within those constraints, the model was free to make all the technical choices it wanted.
Two days later, it got within range of the .NET GC, beating it in GCPerfSim but behind on more realistic workloads (TechEmpower Fortunes benchmark).
My ".NET GC written in C#" project is coded by hand, bit by bit, as an educational project. But I couldn't resist the temptation of having Fable "speedrun" the project and show me a glimpse of what it could look like if I had more time.
So I let Fable work autonomously on it, with a few constraints: - Mark&Sweep GC, no compaction - Supports 64-bit applications only - Using only the GC API, can't fork the runtime (which means using the default write barrier, and not being able to use a read barrier) - And of course, in C#
Within those constraints, the model was free to make all the technical choices it wanted.
Two days later, it got within range of the .NET GC, beating it in GCPerfSim but behind on more realistic workloads (TechEmpower Fortunes benchmark).
Source code and description are available on the github repository: https://github.com/kevingosse/ManagedDotnetGC/tree/claude/ex...