Обсуждение: Use bsearch() instead of a manual binary search in syscache.c

cca5507 <cca5507@qq.com> wrote:

> I make a patch for the $subject, which make the code simpler, thoughts?

I proposed something like that earlier [1] but did not get too far. The short
discussion might be useful for you though.

[1] https://www.postgresql.org/message-id/36977.1720623613@antos

-- 
Antonin Houska
Web: https://www.cybertec-postgresql.com

On Sun, Nov 9, 2025 at 1:12 AM Antonin Houska <ah@cybertec.at> wrote:
> cca5507 <cca5507@qq.com> wrote:
> > I make a patch for the $subject, which make the code simpler, thoughts?
>
> I proposed something like that earlier [1] but did not get too far. The short
> discussion might be useful for you though.

One factor is that libc bsearch() implementations might not all be
header-only and inlineable.  I vaguely recall that being discussed in
some round of hacking on qsort() and qunique().

Thomas Munro <thomas.munro@gmail.com> writes:
> One factor is that libc bsearch() implementations might not all be
> header-only and inlineable.  I vaguely recall that being discussed in
> some round of hacking on qsort() and qunique().

I'm quite certain that years ago we determined that bsearch()
was slower than a manually written-out loop, probably because of
exactly the point that the comparisons would be inline.  Don't
know whether modern compilers have changed that conclusion.

There are places where we wouldn't care about such microscopic
performance details, but I think syscache.c is not one of them.

            regards, tom lane

On Sun, Nov 9, 2025 at 6:01 AM Tom Lane <tgl@sss.pgh.pa.us> wrote:
> I'm quite certain that years ago we determined that bsearch()
> was slower than a manually written-out loop, probably because of
> exactly the point that the comparisons would be inline.  Don't
> know whether modern compilers have changed that conclusion.

It looks like glibc's version is inlined, but others I checked aren't.

> There are places where we wouldn't care about such microscopic
> performance details, but I think syscache.c is not one of them.

So we'd probably need our own inline function to keep the playing
field level.  Some tweaked algorithms[1] are also said to speed up
small integer tables, Unicode tables etc.

[1] https://github.com/scandum/binary_search

So we'd probably need our own inline function to keep the playing

field level.  Some tweaked algorithms[1] are also said to speed up

small integer tables, Unicode tables etc.