On Wed, Sep 9, 2020 at 7:33 AM Peter Geoghegan <p...@bowt.ie> wrote: > On Tue, Sep 8, 2020 at 1:23 AM Masahiko Sawada > <masahiko.saw...@2ndquadrant.com> wrote: > > > > I wonder if you would also see a speed-up with a bsearch() replacement > > > > that is inlineable, so it can inline the comparator (instead of > > > > calling it through a function pointer). I wonder if something more > > > > like (lblk << 32 | loff) - (rblk << 32 | roff) would go faster than > > > > the branchy comparator. > > > > > > Erm, off course that expression won't work... should be << 16, but > > > even then it would only work with a bsearch that uses int64_t > > > comparators, so I take that part back. > > > > Yeah, it seems to worth benchmarking the speed-up with an inlining. > > I'll do some performance tests with/without inlining on top of > > checking boundary values. > > It sounds like Thomas was talking about something like > itemptr_encode() + itemptr_decode(). In case you didn't know, we > actually do something like this for the TID tuplesort used for CREATE > INDEX CONCURRENTLY.
BTW I got around to trying this idea out for a specialised bsearch_itemptr() using a wide comparator, over here: https://www.postgresql.org/message-id/CA%2BhUKGKztHEWm676csTFjYzortziWmOcf8HDss2Zr0muZ2xfEg%40mail.gmail.com