Currently our copying collection occurs simultaneously with the marking phase. We'd like to be able to reuse CopiedBlocks as soon as they become fully evacuated, but this is not currently possible because we don't know the liveness statistics of each old CopiedBlock until marking/copying has already finished. Instead, we have to allocate additional memory from the OS to use as our working set of CopiedBlocks while copying. We then return the fully evacuated old CopiedBlocks back to the block allocator, thus giving our copying phase an O(live bytes) overhead. To fix this, we should instead split the copying phase apart from the marking phase. This way we have full liveness data for each CopiedBlock during the copying phase so that we can reuse them the instant they become fully evacuated. With the additional liveness data that each CopiedBlock accumulates, we can add some additional heuristics to the collector. For example, we can calculate our global Heap fragmentation and only choose to do a copying phase if that fragmentation exceeds some limit. As another example, we can skip copying blocks that are already above a particular fragmentation limit, which allows older objects to coalesce into blocks that are rarely copied.