FWIW, the non-parallel version yields a profile like: - 25.69% postgres jitted-24373-18.so [.] evalexpr_3_3 - 96.00% evalexpr_3_3

FWIW, the non-parallel version yields a profile like: - 25.69% postgres jitted-24373-18.so [.] evalexpr_3_3 - 96.00% evalexpr_3_3 ExecAgg + 2.89% __isnan - 12.15% postgres jitted-24373-16.so [.] evalexpr_3_0 - 99.38% evalexpr_3_0 ExecSeqScan ExecAgg + 0.56% ExecSeqScan - 11.74% postgres postgres [.] heap_getnext 99.83% heap_getnext ExecSeqScan ExecAgg + 9.31% postgres postgres [.] ExecSeqScan + 7.66% postgres libc-2.27.so [.] __memmove_avx_unaligned_erms + 5.61% postgres postgres [.] heapgetpage + 5.36% postgres postgres [.] ExecAgg + 5.35% postgres postgres [.] MemoryContextReset + 4.11% postgres postgres [.] HeapTupleSatisfiesMVCC + 3.53% postgres postgres [.] ExecStoreTuple + 2.55% postgres postgres [.] hash_search_with_hash_value + 1.63% postgres libc-2.27.so [.] __isnan + 1.16% postgres postgres [.] CheckForSerializableConflictOut so it's pretty clear that there's plenty additional work can be done to improve the generated code (it's pretty crappy right now), and that there's potential around improving things due to batching (mainly for increased cache efficiency).