Currently G1 has special predictors for copy costs during the entire marking (including remembered set rebuild) that are different than otherwise.

However, G1 only does additional work in the closures during the initial mark pause, so there does not seem any reason to have separate predictors the whole concurrent operation.

Limit the special predictor to the initial mark pause; potentially investigate whether there is actually a noticeable difference between initial mark and the "regular" pauses, and eventually simply drop this distinction.