I wrote: > I attach your original 0003 here (it still applies, with some line > offsets). That's just so the cfbot doesn't get confused about what > it's supposed to test now.
Pushed that part now, too. BTW, the first test run I did on this (on x86_64) was actually several percent *slower* than HEAD. I couldn't reproduce that after restarting the postmaster; all later tests concurred that there was a speedup. So I suppose that was just some phase-of-the-moon effect, perhaps caused by an ASLR-dependent collision of bits of code in processor cache. Still, that illustrates the difficulty of getting useful, reproducible improvements when doing this kind of hacking. I tend to think that most of the time we're better off leaving this to the compiler. regards, tom lane