Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 23.128.96.18 as permitted sender) client-ip=23.128.96.18;
IronPort-SDR: TEs/9Fnz7hcwNvlBPeJ4OVtvGPzLhNLQFtRw305M0D7cs5XyEVbTF+dRAgUgrnlluBdl7hcVhx
 KmFfpZ3Vg6hw==
IronPort-SDR: inL+gV1iNgpp4Pg9Jrp7kPw0b7fxvNXaWpgLg9468RJgLqjzIFFXB084vJwePYH/m1c+jM1v90
 Dxza4s6gBnMA==
Date:   Tue, 21 Apr 2020 00:27:23 +0300
From:   Andy Shevchenko <andriy.shevchenko@linux.intel.com>
To:     Alexey Dobriyan <adobriyan@gmail.com>
Cc:     akpm@linux-foundation.org, linux-kernel@vger.kernel.org,
        linux-fsdevel@vger.kernel.org, pmladek@suse.com,
        rostedt@goodmis.org, sergey.senozhatsky@gmail.com,
        linux@rasmusvillemoes.dk
Subject: Re: [PATCH 03/15] print_integer: new and improved way of printing
 integers
Message-ID: <20200420212723.GE185537@smile.fi.intel.com>
References: <20200420205743.19964-1-adobriyan@gmail.com>
 <20200420205743.19964-3-adobriyan@gmail.com>
 <20200420211911.GC185537@smile.fi.intel.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <20200420211911.GC185537@smile.fi.intel.com>
Organization: Intel Finland Oy - BIC 0357606-4 - Westendinkatu 7, 02160 Espoo
Sender: linux-kernel-owner@vger.kernel.org
Precedence: bulk

On Tue, Apr 21, 2020 at 12:19:11AM +0300, Andy Shevchenko wrote:
> On Mon, Apr 20, 2020 at 11:57:31PM +0300, Alexey Dobriyan wrote:
> > Time honored way to print integers via vsnprintf() or equivalent has
> > unavoidable slowdown of parsing format string. This can't be fixed in C,
> > without introducing external preprocessor.
> > 
> > seq_put_decimal_ull() partially saves the day, but there are a lot of
> > branches inside and overcopying still.
> > 
> > _print_integer_*() family of functions is meant to make printing
> > integers as fast as possible by deleting format string parsing and doing
> > as little work as possible.
> > 
> > It is based on the following observations:
> > 
> > 1) memcpy is done in forward direction
> > 	it can be done backwards but nobody does that,
> > 
> > 2) digits can be extracted in a very simple loop which costs only
> > 	1 multiplication and shift (division by constant is not division)
> > 
> > All the above asks for the following signature, semantics and pattern of
> > printing out beloved /proc files:
> > 
> > 	/* seq_printf(seq, "%u %llu\n", A, b); */
> > 
> > 	char buf[10 + 1 + 20 + 1];
> > 	char *p = buf + sizeof(buf);
> > 
> > 	*--p = '\n';
> > 	p = _print_integer_u64(p, B);
> > 	*--p = ' ';
> > 	p = _print_integer_u32(p, A);
> > 
> > 	seq_write(seq, p, buf + sizeof(buf) - p);
> > 
> > 1) stack buffer capable of holding the biggest string is allocated.
> > 
> > 2) "p" is pointer to start of the string. Initially it points past
> > 	the end of the buffer WHICH IS NOT NUL-TERMINATED!
> > 
> > 3) _print_integer_*() actually prints an integer from right to left
> > 	and returns new start of the string.
> > 
> > 			     <--------|
> > 				123
> > 				^
> > 				|
> > 				+-- p
> > 
> > 4) 1 character is printed with
> > 
> > 	*--p = 'x';
> > 
> > 	It generates very efficient code as multiple writes can be
> > 	merged.
> > 
> > 5) fixed string is printed with
> > 
> > 	p = memcpy(p - 3, "foo", 3);
> > 
> > 	Complers know what memcpy() does and write-combine it.
> > 	4/8-byte writes become 1 instruction and are very efficient.
> > 
> > 6) Once everything is printed, the result is written to seq_file buffer.
> > 	It does only one overflow check and 1 copy.
> > 
> > This generates very efficient code (and small!).
> > 
> > In regular seq_printf() calls, first argument and format string are
> > constantly reloaded. Format string will most likely with [rip+...] which
> > is quite verbose.
> > 
> > seq_put_decimal_ull() will do branches (and even more branches
> > with "width" argument)
> > 
> 
> > 	TODO
> > 	benchmark with mainline because nouveau is broken for me -(
> > 	vsnprintf() changes make the code slower
> 
> Exactly main point of this exercise. I don't believe that algos in vsprintf.c
> are too dumb to use division per digit (yes, division by constant which is not
> power of two is a heavy operation).
> 

And second point here, why not to use existing algos from vsprintf.c?

-- 
With Best Regards,
Andy Shevchenko