I wrote:
> Have two memtuple arrays, one for first sortkey null and one for first sortkey non-null:
Hacking on this has gotten only as far as the "compiles but segfaults" stage, but I wanted to note an idea that occurred to me:
Internal qsort doesn't need the srctape member, and removing both that and isnull1 would allow 16-byte "init-tuples" for qsort, which would save a bit of work_mem space, binary space for qsort specializations, and work done during swaps.
During heap sort, we already copy one entry into a stack variable to keep from clobbering it, so it's not a big step to read a member from the init array and form a regular sorttuple from it.