When dealing with a case where a 2TB table had 3 billion dead tuples I discovered that vacuum currently can't make use of more than 1GB of maintenance_work_mem - 179M tuples. This caused excessive amounts of index scanning even though there was plenty of memory available.
I didn't see any good reason for having this limit, so here is a patch that makes use of MemoryContextAllocHuge, and converts the array indexing to use size_t to lift a second limit at 12GB.
One potential problem with allowing larger arrays is that bsearch might no longer be the best way of determining if a ctid was marked dead. It might pay off to convert the dead tuples array to a hash table to avoid O(n log n) runtime when scanning indexes. I haven't done any profiling yet to see how big of a problem this is.
Second issue I noticed is that the dead_tuples array is always allocated max allowed size, unless the table can't possibly have that many tuples. It may make sense to allocate it based on estimated number of dead tuples and resize if needed.
