Date: Tue, 22 Apr 2003 21:43:38 +1000 (EST) From: Bruce Evans <bde@zeta.org.au> To: David Schultz <das@freebsd.org> Cc: current@freebsd.org Subject: Re: HEADS UP: UFS2 now the default creation type on 5.0-CURRENT Message-ID: <20030422210632.U16843@gamplex.bde.org> In-Reply-To: <20030421211440.GA5507@HAL9000.homeunix.com> References: <20030420192319.GB4963@HAL9000.homeunix.com> <Pine.NEB.3.96L.1030420174551.16891t-100000@fledge.watson.org> <20030421231756.H11214@gamplex.bde.org> <20030421211440.GA5507@HAL9000.homeunix.com>
next in thread | previous in thread | raw e-mail | index | archive | help
On Mon, 21 Apr 2003, David Schultz wrote:
> On Mon, Apr 21, 2003, Bruce Evans wrote:
> > On Mon, 21 Apr 2003, David Schultz wrote:
> > > Index: ufsread.c
> > > ===================================================================
> > > RCS file: /cvs/src/sys/boot/common/ufsread.c,v
> > > retrieving revision 1.11
> > > diff -u -r1.11 ufsread.c
> > > --- ufsread.c 25 Feb 2003 00:10:20 -0000 1.11
> > > +++ ufsread.c 21 Apr 2003 10:10:01 -0000
> > > ...
> > > @@ -47,11 +59,11 @@
> > > ...
> > > -#define FS_TO_VBA(fs, fsb, off) (fsbtodb(fs, fsb) + \
> > > - ((off) / VBLKSIZE) * DBPERVBLK)
> > > +#define FS_TO_VBA(fs, fsb, off) ma((off) / VBLKSIZE, DBPERVBLK, \
> > > + fsbtodb((fs), (fsb)))
> >
> > The division by VBLKSIZE should probably be a shift. ufsread.c has
> > VBLKSHIFT and uses it for all multiplications and divisions by VBLKSIZE
> > except this one. gcc can't optimize to just a shift since all the
> > types are signed and C99 specifies that division of negative integers
> > by positive ones has the usual hardware brokenness.
>
> As I recall, signed division gets optimized into a sign test, some
^ by a power of 2
> bit fiddling for negative numbers, and a division. The additional
shift
> cost is nominal if you only care about speed, but I'm sure using a
> shift directly would save a few more bytes.
I tried this, but it had no effect since FS_TO_VBA() is never actually used.
So there is a much better optimization for it :-). I think this makes ma()
unused too.
I thought that the savings for unsigned division were more for long
longs than for longs, but they are actually relatively smaller. On
i386's, signed division (when optimized to shifts) of %edx:%eax by
2^12 takes 19 bytes and right shift takes 7 bytes; the corresponding
numbers for %eax are 12 bytes and 3 bytes. Optimization for space
should not use shifts for the signed case.
Bruce
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?20030422210632.U16843>