Re: [PATCH] buffer: a small optimization in grow_buffers

From: Mikulas Patocka <mpatocka@redhat.com>
To: Matthew Wilcox <willy@infradead.org>
Cc: Alexander Viro <viro@zeniv.linux.org.uk>,
	Jens Axboe <axboe@kernel.dk>,
	linux-fsdevel@vger.kernel.org, linux-block@vger.kernel.org
Subject: Re: [PATCH] buffer: a small optimization in grow_buffers
Date: Mon, 22 Mar 2021 12:36:30 -0400 (EDT)	[thread overview]
Message-ID: <alpine.LRH.2.02.2103221225330.1530@file01.intranet.prod.int.rdu2.redhat.com> (raw)
In-Reply-To: <20210322151124.GP1719932@casper.infradead.org>

On Mon, 22 Mar 2021, Matthew Wilcox wrote:

> On Mon, Mar 22, 2021 at 10:05:05AM -0400, Mikulas Patocka wrote:
> > This patch replaces a loop with a "tzcnt" instruction.
> 
> Are you sure that's an optimisation?  The loop would execute very few
> times under normal circumstances (a maximum of three times on x86).
> Some numbers would be nice.

According to Agner's instruction tables, tzcnt has latency 3 on Skylake 
and 2 on Zen. (386, 486, Pentium and K6 didn't have hardware for the bsf 
instruction, they executed it in microcode bit-by-bit, but newer 
processors execute it quickly).

The patch reduces code size by 16 bytes.

Mikulas

> > Signed-off-by: Mikulas Patocka <mpatocka@redhat.com>
> > 
> > Index: linux-2.6/fs/buffer.c
> > ===================================================================
> > --- linux-2.6.orig/fs/buffer.c
> > +++ linux-2.6/fs/buffer.c
> 
> Are ... are you still using CVS?!
> 
> > @@ -1020,11 +1020,7 @@ grow_buffers(struct block_device *bdev,
> >  	pgoff_t index;
> >  	int sizebits;
> >  
> > -	sizebits = -1;
> > -	do {
> > -		sizebits++;
> > -	} while ((size << sizebits) < PAGE_SIZE);
> > -
> > +	sizebits = PAGE_SHIFT - __ffs(size);
> >  	index = block >> sizebits;
> >  
> >  	/*
> > 
> 