Block descriptors currently contain both `gen` (a `generation *`) and `gen_no` (a `uint16_t`, which is `gen->no` cached). I wonder whether it would be worthwhile to drop the former and always use `generations[bd->gen_no]`. Afterall, this would save an entire word per descriptor which may improve our caches' coverage of the bdescr region appreciably. Presumably the `generations` array would be hot enough to stay in cache in such a scheme, so I would guess that the cost would be small-ish. Simon, have you measured this in the past?