Date: Fri, 17 Jan 2025 11:15:29 -0600
From: Nathan Bossart <nathandbossart@gmail.com>
To: Tom Lane <tgl@sss.pgh.pa.us>
Cc: David Rowley <dgrowleyml@gmail.com>,
	John Naylor <johncnaylorls@gmail.com>,
	"Chiranmoy.Bhattacharya@fujitsu.com" <Chiranmoy.Bhattacharya@fujitsu.com>,
	"Devanga.Susmitha@fujitsu.com" <Devanga.Susmitha@fujitsu.com>,
	"pgsql-hackers@postgresql.org" <pgsql-hackers@postgresql.org>,
	"Ragesh.Hajela@fujitsu.com" <Ragesh.Hajela@fujitsu.com>
Subject: Re: [PATCH] Hex-coding optimizations using SVE on ARM.
Message-ID: <Z4qQMS8EtYzN-eno@nathan>
References: <Z4E-5oQGux-2R4iL@nathan>
 <Z4GHNfhRKuA0r_Wn@nathan>
 <CANWCAZZvXuJMgqMN4u068Yqa19CEjS31tQKZp_qFFFbgYfaXqQ@mail.gmail.com>
 <622082.1736834344@sss.pgh.pa.us>
 <Z4aXj_fJYCQ6GBbC@nathan>
 <CANWCAZYNtrphdWSEakFbGrJZtjkFvvo8BB6nh5AGRiTfUu47Og@mail.gmail.com>
 <1194263.1736925288@sss.pgh.pa.us>
 <CANWCAZaq-hHGzEE9u=+ed3czdn6WUQdmGpmUpBiT7p+wNsOS+A@mail.gmail.com>
 <CAApHDvpJr7r0n6TY_0psumA1uLy7+nmdq==L3jjDT-RYPMoHRw@mail.gmail.com>
 <1519827.1736990839@sss.pgh.pa.us>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <1519827.1736990839@sss.pgh.pa.us>
Archived-At: <https://www.postgresql.org/message-id/Z4qQMS8EtYzN-eno%40nathan>
Precedence: bulk

With commit e24d770 in place, I took a closer look at hex_decode(), and I
concluded that doing anything better without intrinsics would likely
require either a huge lookup table or something with complexity rivalling
the instrinsics approach (while also not rivalling its performance).  So, I
took a closer look at the instrinsics patches and had the following
thoughts:

* The approach looks generally reasonable to me, but IMHO the code needs
  much more commentary to explain how it works.

* The functions that test the length before potentially calling a function
  pointer should probably be inlined (see pg_popcount() in pg_bitutils.h).
  I wouldn't be surprised if some compilers are inlining this stuff
  already, but it's probably worth being explicit about it.

* Finally, I think we should ensure we've established a really strong case
  for this optimization.  IME these intrinsics patches require a ton of
  time and energy, and the code is often extremely complex.  I would be
  interested to see how your bytea test compares with the improvements
  added in commit e24d770 and with sending the data in binary.

-- 
nathan