That probably closer to a best case scenario. A property predicted branch is still an extra instruction and a couple of entries in the branch target buffer and a global(?) history (I'm not really up to date on the newer perceptron predictors yet.
These are going to be limited resources so you'd rather not pollute them and make other branches predict wrong because they are aliases (shared) by many different instruction address.
Also, the fetcher can only get instructions in blocks and if there are multiple branches it will have to predict based on the first branch.
There is a lot of machinery in branch and branch the prediction because keeping the pipeline full of instructions is critically important for performance.
And while throuhput may not suffer much, the chances of increasing mispredictions in the hot path increase adversely harming latency.
These really need to be compiled out for release because you can't assume others can afford the cycles.
These are going to be limited resources so you'd rather not pollute them and make other branches predict wrong because they are aliases (shared) by many different instruction address.
Also, the fetcher can only get instructions in blocks and if there are multiple branches it will have to predict based on the first branch.
There is a lot of machinery in branch and branch the prediction because keeping the pipeline full of instructions is critically important for performance.
And while throuhput may not suffer much, the chances of increasing mispredictions in the hot path increase adversely harming latency.
These really need to be compiled out for release because you can't assume others can afford the cycles.