[prev in list] [next in list] [prev in thread] [next in thread] 

List:       postgresql-hackers
Subject:    Re: speed up unicode normalization quick check
From:       John Naylor <john.naylor () 2ndquadrant ! com>
Date:       2020-05-30 6:52:24
Message-ID: CACPNZCsWu2XSLuqFfdye8GrL+89QjfL7+i1Bp02H-2Z5k49v6g () mail ! gmail ! com
[Download RAW message or body]

On Sat, May 30, 2020 at 12:13 AM Mark Dilger
<mark.dilger@enterprisedb.com> wrote:
> 
> 
> I forgot in my first round of code review to mention, "thanks for the patch".  I \
> generally like what you are doing here, and am trying to review it so it gets \
> committed.

And I forgot to say thanks for taking a look!

> The reason I gave this feedback is that I saved the *kwlist_d.h files generated \
> before applying the patch, and compared them with the same files generated after \
> applying the patch, and noticed a very slight degradation.  Most of the files \
> changed without any expansion, but the largest of them, src/common/kwlist_d.h, \
> changed from 
> static const int16 h[901]
> 
> to
> 
> static const int16 h[902]

Interesting, I hadn't noticed. With 450 keywords, we need at least 901
elements in the table. Since 901 is divisible by the new hash
multiplier 17, this gets triggered:

# However, it would be very bad if $nverts were exactly equal to either
# $hash_mult1 or $hash_mult2: effectively, that hash function would be
# sensitive to only the last byte of each key.  Cases where $nverts is a
# multiple of either multiplier likewise lose information.  (But $nverts
# can't actually divide them, if they've been intelligently chosen as
# primes.)  We can avoid such problems by adjusting the table size.
while ($nverts % $hash_mult1 == 0
    || $nverts % $hash_mult2 == 0)
{
    $nverts++;
}

This is harmless, and will go away next time we add a keyword.

--
John Naylor                https://www.2ndQuadrant.com/
PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services


[prev in list] [next in list] [prev in thread] [next in thread]