Hi,
On 2026-01-26 16:47:31 +0900, Amit Langote wrote:
> I tried my patch over your committed SeqNext inlining patch and ran
> the following benchmark but didn't notice in material difference:
>
> CREATE TABLE t (a int);
> INSERT INTO t SELECT generate_series(1, 1000000);
> ANALYZE t;
Because the table isn't frozen, visibility checks will probably add enough
per-row overhead to make any per-row micro-optimization harder to see. On my
somewhat older workstation freezing is a 17% improvement.
> SET max_parallel_workers_per_gather = 0;
> SELECT * FROM t WHERE a = -1;
>
> Perhaps not too surprising given it's just eliminating a couple of
> instructions per row that the branch predictor probably handles well
> anyway? Still seems worth having for code hygiene if nothing else.
>
> Same result (no diff in perf) when I apply it over your patch to move
> the scandesc == NULL check.
FWIW, on my cascade lake workstation it's a, surprisingly large, 3.5%, after
freezing. Without freezing there maybe still is a difference, but it's very
close to the noise floor.
Greetings,
Andres Freund
