Re: pgsql: Optimize btree insertions for common case of increasing values

Suggested next steps to deal with this:

* A minor point: I don't think you should call
RelationSetTargetBlock() when the page P_ISROOT(), which, as I
mentioned, is a condition that can coexist with P_ISLEAF() with very
small B-Trees. There can be no point in doing so. No need to add
P_ISROOT() to the main "Is cached page stale?" test within
_bt_doinsert(), though.

* An assertion would make me feel a lot better about depending on not
having a page split from a significant distance.

Your optimization should continue to not be used when it would result
in a page split, but only because that would be slow. The comments
should say so, IMV.

Also, _bt_insertonpg() should have an assertion
against a page split actually occurring when the optimization was
used, just in case. When !BufferIsValid(cbuf), we know that we're
being called from _bt_doinsert() (see existing assertion at the top of
_bt_insertonpg() as an example of this), so at the point where it's
clear a page split is needed, we should assert that there is no target
block that we must have been passed as the target page.

* The indentation of the main _bt_doinsert() test does not follow
project guidelines. Please tweak that, too.