It seems like we're at the point where something might be able to be done, but having the set up be

assembly -> RNN -> C

Would be asking for a bit much IMO. If you develop something like snowman that outputs many possible equivalent C ASTs and train something that tries to choose the one that was the original code.