Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::3:7 as permitted sender) client-ip=2620:137:e000::3:7;
Date:   Tue, 19 Sep 2023 14:04:48 -0400
From:   Charlie Jenkins <charlie@rivosinc.com>
To:     David Laight <David.Laight@aculab.com>
Cc:     Palmer Dabbelt <palmer@dabbelt.com>,
        Conor Dooley <conor@kernel.org>,
        Samuel Holland <samuel.holland@sifive.com>,
        "linux-riscv@lists.infradead.org" <linux-riscv@lists.infradead.org>,
        "linux-kernel@vger.kernel.org" <linux-kernel@vger.kernel.org>,
        "linux-arch@vger.kernel.org" <linux-arch@vger.kernel.org>,
        Paul Walmsley <paul.walmsley@sifive.com>,
        Albert Ou <aou@eecs.berkeley.edu>,
        Arnd Bergmann <arnd@arndb.de>
Subject: Re: [PATCH v6 3/4] riscv: Add checksum library
Message-ID: <ZQniwNEoYLo52HI7@ghost>
References: <20230915-optimize_checksum-v6-0-14a6cf61c618@rivosinc.com>
 <20230915-optimize_checksum-v6-3-14a6cf61c618@rivosinc.com>
 <0357e092c05043fba13eccad77ba799f@AcuMS.aculab.com>
 <ZQkOSf1b66lHzjaf@ghost>
 <0fe9694900c7492c96dce6b67710173f@AcuMS.aculab.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <0fe9694900c7492c96dce6b67710173f@AcuMS.aculab.com>
Precedence: bulk

On Tue, Sep 19, 2023 at 08:00:12AM +0000, David Laight wrote:
> ...
> > > So ending up with (something like):
> > > 	end = buff + length;
> > > 	...
> > > 	while (++ptr < end) {
> > > 		csum += data;
> > > 		carry += csum < data;
> > > 		data = ptr[-1];
> > > 	}
> > > (Although a do-while loop tends to generate better code
> > > and gcc will pretty much always make that transformation.)
> > >
> > > I think that is 4 instructions per word (load, add, cmp+set, add).
> > > In principle they could be completely pipelined and all
> > > execute (for different loop iterations) in the same clock.
> > > (But that is pretty unlikely to happen - even x86 isn't that good.)
> > > But taking two clocks is quite plausible.
> > > Plus 2 instructions per loop (inc, cmp+jmp).
> > > They might execute in parallel, but unrolling once
> > > may be required.
> > >
> > It looks like GCC actually ends up generating 7 total instructions:
> > ffffffff808d2acc:	97b6                	add	a5,a5,a3
> > ffffffff808d2ace:	00d7b533          	sltu	a0,a5,a3
> > ffffffff808d2ad2:	0721                	add	a4,a4,8
> > ffffffff808d2ad4:	86be                	mv	a3,a5
> > ffffffff808d2ad6:	962a                	add	a2,a2,a0
> > ffffffff808d2ad8:	ff873783          	ld	a5,-8(a4)
> > ffffffff808d2adc:	feb768e3          	bltu	a4,a1,ffffffff808d2acc <do_csum+0x34>
> > 
> > This mv instruction could be avoided if the registers were shuffled
> > around, but perhaps this way reduces some dependency chains.
> 
> gcc managed to do 'data += csum' so had add 'csum = data'.
> If you unroll once that might go away.
> It might then be 10 instructions for 16 bytes.
> Although you then need slightly larger alignment code.
> 
> 	David
> 
> -
> Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK
> Registration No: 1397386 (Wales)
> 
I messed with it a bit and couldn't get the mv to go away. I would expect
mv to be very cheap so it should be fine, and I would like to avoid adding
too much to the alignment code since it is already large, and I assume
that buff will be aligned more often than not.

Interestingly, the mv does not appear pre gcc 12, and does not appear on clang.

- Charlie