I think that's an indication that the branch has a non-negligible branch misprediction rate. A mispredicts means the branch will often sit at the head of ROB as recovery happens, so will be heavily represented in the samples. That's also consistent with cmov providing a good speedup: that usually means the underlying branches were hard to predict (otherwise cmov usually ends up slower).

In this post I don't find any particular clusterting around branches, but they are all predicted essentially 100% of the time, so that's not surprising.