Object mutating assignments are checked for old -> new direction and handled in one of two ways, let's call them A and B.
Under method A, when a old -> new assignment takes place, the new object is added to a "check" array. At the same time, its generation is changed from 0 to -1, so that this is wastefully not done twice. Objects in the check array are processed during the mark phase; they are marked as reachable (and that will promote them to the mature generation).
Under method B, the new object is not involved. An old object has been mutated and we record it in a "mutated" array. (So that this isn't done twice for the same object, we also change its generation to -1; that gets fixed back during GC.) The "mutated" array is subject to marking in the mark phase (even though mature objects are not), in order that we chase the references from that object to any new object. These (necessarily considered reachable) objects are also processed in the sweep phase to return them to gen 1.
When might you use method B? Say there is a bulk assignment operation, like hundreds of elements of a vector object are set to values, some of which may be new objects. In that situation, it's more efficient to put the aggregate object into the mutated array and not deal with any information about the right hand side objects at all.
So this is a choice made by the user? Or the choice is made automatically using some heuristic? Regardless: I don't really see the point of avoiding moving gc if you have to have barriers. Moving gc will be faster; the only reason I can think of for avoiding it is to avoid mutator overhead, but you are paying barriers anyway.
Under method A, when a old -> new assignment takes place, the new object is added to a "check" array. At the same time, its generation is changed from 0 to -1, so that this is wastefully not done twice. Objects in the check array are processed during the mark phase; they are marked as reachable (and that will promote them to the mature generation).
Under method B, the new object is not involved. An old object has been mutated and we record it in a "mutated" array. (So that this isn't done twice for the same object, we also change its generation to -1; that gets fixed back during GC.) The "mutated" array is subject to marking in the mark phase (even though mature objects are not), in order that we chase the references from that object to any new object. These (necessarily considered reachable) objects are also processed in the sweep phase to return them to gen 1.
When might you use method B? Say there is a bulk assignment operation, like hundreds of elements of a vector object are set to values, some of which may be new objects. In that situation, it's more efficient to put the aggregate object into the mutated array and not deal with any information about the right hand side objects at all.