We iterate so that you can recurse

It's been a while since I've read the Garbage Collection Handbook, but it cites several studies showing some workloads are better served by depth-first traversal for compacting GC and other workloads are better served by breadth-first traversal. In general, you want to place an object near the field references it frequently de-references.

The problem with depth-first traversal is that an object ends up near its first child and its first grand-child and its first great-grandchild. The problem with breadth-first traversal is that objects reached early wind up near their children (good), but there's no space to put the grandchildren near the children (bad). You might have a sampling profiler show you which field accesses are causing pipeline stalls, and when traversing the object graph, use a double-ended queue for the grey list. Push the frequently stalling references at the front of the queue and push all of the others to the rear of the queue. Though, you'd need to keep long enough history to avoid alternating between good layouts and poor layouts every-other GC compaction.

Anything with a read barrier might also be able to see which slots are hotter, and thus prioritise copying those first; there was one paper from the 90s showing profiling helped get a better layout.