Yes, I wrote the GC from scratch. I am not sure if it can be integrated well with llvm, I never thought about it. The concurrent GC algorithm (plus a single threaded incremental collector) in Dao was inspired by the algorithm proposed in http://researcher.watson.ibm.com/researcher/files/us-bacon/B....
IIRC reference counted GCs are easier to integrate with LLVM as there is no need to generate a global root set (which is one of the things that llvm makes hard).
Right, this property of reference counted GCs can make certain things simpler.
I just realized that you asked the question probably because of the LLVM-based JIT compiler for Dao. This JIT compiler actually does not deal with GC at all, it just compiles consecutive segments of VM instructions into functions that will be called by the VM. Only a subset of the total VM instructions are compilable this way. It was intended for programs with a significant amount of numeric computation.
