I said:
> none of this works if the index isn't unique.
On the other hand --- if the index *is* unique, and we are checking
equality on all columns (a fairly easily checked condition), then we
know we should retrieve at most one visible tuple. So, without making
any incorrect assumptions, we could terminate the indexscan after the
first successful match. Hmm ... you might be right that there's a
cheap win to be had there. I still think that we also need to do
something with propagating tuple deadness flags into the index, though.
regards, tom lane