ppc_8xx-gcc 2.95.3 Monta Vista does not do ANY loop unrolling

[Joakim Tjernlund]

> On Tue, Nov 12, 2002 at 05:09:11PM +0100, Joakim Tjernlund wrote:
> > > On Tue, Nov 12, 2002 at 10:40:41AM +0100, Joakim Tjernlund wrote:
> > >
> > > > I optimized the crc32() in JFFS2(fs/jffs2/crc.h) by manually unrolling
> > > > the crc32 loop. This gave me a speed increase of 22% in mounting JFFS2 FS
> > > >
> > > > Later Alan Cox pointed out that my changes makes x86 run slower and it turns
> > > > out that on x86 and a fairly new gcc will automatically unroll loops 'where appropriate'
> > > >
> > > > Removed my hand coded unrolling and added -funroll-loops to the JFFS2 Makefile,
> > > > I got similar results as my hand coded unrolling (a little better).
> > > >
> > > > I therefore conclude that ppc_8xx-gcc 2.95.3 from Monta Vista does not do ANY unrolling
> > > > unless you specify -funroll-loops. Doing this for the whole kernel is NOT a good idea,
> > > > it will run slower due to big increase of size.
> > >
> > > I'm sort-of supprised that gcc-2.95.x (or gcc-*, for that matter) will
> > > unroll some loops with only -O2 since the info page on gcc-3.2 and
> > > gcc-2.95 both say that -funroll-loops isn't turned on my any of the -O
> > > levels.
> > >
> > > So I suspect someone decided that small loops can safely be unrolled on
> > > i386 at some optimization level, but that same decision (with possibly
> > > good reason) was not made for PPC32.  So it's a gcc feature, not a
> > > MVista-specific issue.
> >
> > Newer gcc(>=3.0) may do the same for PPC32. We only know that newer gcc's(Alan Cox knows more)
> > will do it for x86 and 2.95.3 for ppc_8xx won't, so there is a big ? in the middle.
>
> Did Alan say what version of gcc Alan was talking about?

No, I did not ask at the time :-(

>
> > Now to the trick question(s):
> > Where might it be suitable to add -funroll-loops or, better yet, can it be done
> > with a pragma or attribute attached to the function in question? It's pretty
> > hard to unroll inline functions otherwise (and only the inline function).
>
> Well, to lib/Makefile:
> ifeq ($(CONFIG_PPC32),y)
> CFLAGS_crc32.o += -funroll-loops
> endif
>
> Should work.  And it's not unheard of.

Yes, that much I already figured, but are there OTHER places in
the kernel that also might benefit from unrolling. I don't know the
kernel as well as you do and was hoping for a lead or two.

>
> >  Any progress on the i2c-algo-8xx.c and/or 8xx_io/enet.c patches I sent earlier?
>
> As I said privatly, Dan Malek is handling the enet patch, and I'm
> looking for time to do the i2c one.  Right now I'm working on making the
> kernel easier to tweak (in some ways) for 2.5.

I know Dan is handling the enet stuff, but since you both work
for MV(don't you?) I figured you might know, being an insider and all :-)

Maybe your tweak stuff could make use of forced unrolling?

 Jocke


** Sent via the linuxppc-embedded mail list. See http://lists.linuxppc.org/