The test is really nice with the injection points and the dynamically sized data.
> Ah, but this codepath is taken when HEAP_KEYS_UPDATED is *not* set. I > got that backwards. So I agree the ItemPointerEquals(&tuple->t_self, > ctid) check is redundant. Ok, I did not think about deletes. So the boolean updated here could mean both update and delete, that makes sense to me. > I chatted about that with Andres Freund just the other day, and he said > that compilers are not good at optimizing passing them by value. I'll > take his word on that. I believe that too, but in heap_lock_updated_tuple_rec() the first thing we do is ItemPointerCopy(tid, &tupid); , which makes it probably just as inefficient. Jasper
