Once the compiler has computed the async graph, it could in theory fuse subgraphs of nodes (especially linear sequences) back together and re-express the execut

Once the compiler has computed the async graph, it could in theory fuse subgraphs of nodes (especially linear sequences) back together and re-express the execution flow in terms of spawning and joining threads. This would be simple and could even be a performance win for some graphs (i.e. embarrassingly parallel ones), and would be a complete trainwreck for many, probably most. I certainly wouldn’t want to be stuck reading the error messages produced by a compiler when it failed to find a satisfactory solution to a complex execution graph. Profile guided optimization might help I suppose, but still.