From the lecture you gave on prediction, I get that you predict exits.
In this case, there can be up to 3 exits, one for each call and the return. In fact, in this particular function there will always be at least 2 exits taken. At least one of the calls will be taken and the return will be taken. If I understood your talk correctly, you only have 1 prediction per EBB. Does that mean that your prediction would fail at least 50% of the time in this function? Or you have some mechanism to predict both exits?
I would think this is pretty common, we (developers) often make functions that call other functions and then return.
Also, on the lecture you gave some pretty interesting numbers at the beginning of the talk about how much time is wasted on cold runs in conventional hardware. You explain what the bottlenecks are and how you are addressing them in the Mill. But you don’t mention how effective your solutions are. Do you have any numbers as to how your solution performs vs conventional hardware regarding mispredictions?
