Horton

Re: Improving and extending int128.h to more of numeric.c

Dean Rasheed <dean.a.rasheed@gmail.com>

View thread View on postgr.es
From: Dean Rasheed <dean.a.rasheed@gmail.com>
To: John Naylor <johncnaylorls@gmail.com>
Cc: Andres Freund <andres@anarazel.de>, PostgreSQL Hackers <pgsql-hackers@postgresql.org>
Date: 2025-07-17T07:29:53Z
Lists: pgsql-hackers 
On Wed, 16 Jul 2025 at 19:23, Dean Rasheed <dean.a.rasheed@gmail.com> wrote:
>
> On 64-bit Linux with gcc 14.2 and native int128 support disabled I got
> the following results:
>
> Query 1:
>   HEAD:  1404.096 ms
>   Patch:  992.818 ms
>
> Query 2:
>   HEAD:  1498.949 ms
>   Patch:  935.654 ms
>

BTW, my other motivation for doing this was to simplify the numeric
code. Even if this had zero performance benefit, as long as it didn't
make things any slower, I would argue that it's worth doing.

The other 2 places in numeric.c that have conditional 128-bit integer
code would require more complex hand-written code to replace, such as
128-bit-by-128-bit division. That's obviously doable, but perhaps not
worth the effort as long as it's only those 2 numeric functions that
need it. OTOH, if there's a wider demand for 128-bit integers, that
might change.

Regards,
Dean

Commits

  1. Guard against division by zero in test_int128 module.

  2. Extend int128.h to support more numeric code.

  3. Simplify non-native 64x64-bit multiplication in int128.h.

  4. Optimise non-native 128-bit addition in int128.h.

  5. Refactor int128.h, bringing the native and non-native code together.

  6. Convert src/tools/testint128.c into a test module.