From: Tom Lane <tgl@sss.pgh.pa.us>
To: Alvaro Herrera <alvherre@2ndquadrant.com>
cc: Thomas Munro <thomas.munro@enterprisedb.com>,
        Andrew Gierth <andrew@tao11.riddles.org.uk>,
        David Rowley <david.rowley@2ndquadrant.com>,
        PostgreSQL Hackers <pgsql-hackers@lists.postgresql.org>
Subject: Re: Using POPCNT and other advanced bit manipulation instructions
In-reply-to: <20190215145425.GA32071@alvherre.pgsql>
References: <20190215145425.GA32071@alvherre.pgsql>
Comments: In-reply-to Alvaro Herrera <alvherre@2ndquadrant.com>
	message dated "Fri, 15 Feb 2019 11:54:25 -0300"
MIME-Version: 1.0
Content-Type: text/plain; charset="us-ascii"
Content-ID: <3292.1550246240.1@sss.pgh.pa.us>
Date: Fri, 15 Feb 2019 10:57:20 -0500
Message-ID: <3293.1550246240@sss.pgh.pa.us>
Precedence: bulk

Alvaro Herrera <alvherre@2ndquadrant.com> writes:
> I noticed in Compiler Explorer that some (ancient?) Power cpus
> implement instruction "popcntb", and GCC support for those uses
> -mpopcntb switch enabling __builtin_popcount() to use it.  I added the
> switch to configure.in but I'm not sure how well that will work ...  I
> don't know if this is represented in buildfarm.

I experimented a bit with this on an old Apple laptop.  Apple's
compiler rejects -mpopcntb altogether.  FreeBSD's compiler
(gcc 4.2.1) recognizes the switch, but I could not get it to
emit the instruction, even when specifying -mcpu=power5,
which ought to enable it according to the gcc docs:

     ... The `-mpopcntb' option allows GCC to generate the
     popcount and double precision FP reciprocal estimate instruction
     implemented on the POWER5 processor and other processors that
     support the PowerPC V2.02 architecture.

A more recent gcc info file also mentions

     The `-mpopcntd' option
     allows GCC to generate the popcount instruction implemented on the
     POWER7 processor and other processors that support the PowerPC
     V2.06 architecture.

but the gcc version I have on this laptop doesn't know that switch.
In any case, I'm pretty sure Apple never shipped a CPU that could
run either instruction.

I suspect that probing for either option may not be worth the
configure cycles it'd consume :-( ... there are just way too
few of those specific POWER variants out there anymore, even
granting that you have a compiler that will play along.

Moreover, you can't turn on -mpopcntb without having some POWER
equivalent to the CPUID test.

However, if you want to leave the option for this open in
future, it really makes the file name pg_bitutils_sse42.c
quite inappropriate.  How about pg_bitutils_hwpopcnt.c
or something like that?

			regards, tom lane