On Tue, 9 Jun 2020 at 19:24, Andrew Gierth <and...@tao11.riddles.org.uk> wrote: > > >>>>> "David" == David Rowley <dgrowle...@gmail.com> writes: > > David> As you can see, this squeezes about 5% extra out of a copy of a > David> 10 million row bigint table but costs us almost 3% on an > David> equivalent int table. > > And once again I have to issue the reminder: you can have gains or > losses of several percent on microbenchmarks of this kind just by > touching unrelated pieces of code that are never used in the test. In > order to demonstrate a consistent difference, you have to do each set of > tests multiple times, with random amounts of padding added to some > unrelated part of the code.
Thanks for the reminder. Instead of that, I tried with clang 10.0.0. I was previously using gcc 9.3. BIGINT test Master: latency average = 1842.182 ms Patched: latency average = 1715.418 ms INT test Master: latency average = 1650.583 ms Patched: latency average = 1617.783 ms There's nothing in the patch that makes the INT test faster, so I guess that's noise. The BIGINT test is about 7.3% faster in this case. David
