Re: [PATCH] Use 128-bit math to accelerate numeric division, when 8 < divisor digits <= 16

On Sun, 22 Jan 2023 at 15:41, Joel Jacobson <joel(at)compiler(dot)org> wrote:
>
> On Sun, Jan 22, 2023, at 11:06, Dean Rasheed wrote:
> > Seems like a reasonable idea, with some pretty decent gains.
> >
> > Note, however, that for a divisor having fewer than 5 or 6 digits,
> > it's now significantly slower because it's forced to go through
> > div_var_int64() instead of div_var_int() for all small divisors. So
> > the var2ndigits <= 2 case needs to come first.
>
> Can you give a measurable example of when the patch
> the way it's written is significantly slower for a divisor having
> fewer than 5 or 6 digits, on some platform?
>

I just modified the previous test you posted:

\timing on
SELECT count(numeric_div_volatile(1e131071,123456)) FROM generate_series(1,1e4);

Time: 2048.060 ms (00:02.048) -- HEAD
Time: 2422.720 ms (00:02.423) -- With patch

> I did write the code like you suggest first, but changed it,
> since I realised the extra "else if" needed could be eliminated,
> and thought div_var_int64() wouldn't be slower than div_var_int() since
> I thought 64-bit instructions in general are as fast as 32-bit instructions,
> on 64-bit platforms.
>

Apparently it can make a difference. Probably something to do with
having less data to move around. I remember noticing that when I wrote
div_var_int(), which is why I split it into 2 branches in that way.

Regards,
Dean