Re: [PING][PATCH/RFT] Fix PR middle-end/PR28690, modify swap_commutative_operands_p

["H. J. Lu" <hjl at lucon dot org>]

On Sat, Apr 28, 2007 at 04:20:25PM -0500, Peter Bergner wrote:
> On Sat, 2007-04-28 at 08:31 -0700, H. J. Lu wrote:
> > I tried your patch on gcc 4.3 revision 123850. But eon in SPEC
> > CPU 2K failed on Linux/ia32 at -O2. I have
> > 
> > 252.eon=default=default=default:
> > CXXPORTABILITY = -DHAS_ERRLIST
> > srcalt=fmax_errno
> > 
> > and got
> > 
> >   Running 252.eon ref base o2 default
> > *** Miscompare of pixels_out.kajiya, see
> > /export/spec/src/2000/spec/benchspec/CINT2000/252.eon/run/00000002/pixels_out.kajiya.mis
> > Invalid run; unable to continue.  If you wish to ignore errors please
> > use '-I' or ignore_errors
> > 
> > There is no problem with gcc 4.3 revision 123850 without your patch.
> > Does gcc 4.3 revision 123850 plus your patch work with SPEC CPU 2K
> > on Linux/ia32? What is your config on eon?
> 
> Well when we ran SPEC2000 (only used -O2), eon passed with our patch and
> failed without.  However, given what Paolo said in this post:
> 
>     http://gcc.gnu.org/ml/gcc-patches/2007-04/msg01153.html
> 
> Without -ffast-math, it's probably just luck that is passes with or
> without the patch.  So the short answer is, for eon, use -ffast-math.

That is very interesting. On Linux/i686, I found

1. 252.eon worked with gcc 4.3 revision 123850 at -O2.
2. 252.eon failed with gcc 4.3 revision 123850 at -O2 -ffast-math.
3. 252.eon failed with gcc 4.3 revision 123850 plus your patch at -O2.
3. 252.eon worked with gcc 4.3 revision 123850 plus your patch at
-O2 -ffast-math.

It is hard to measure SPEC CPU 2K impact of your change on Linux/i686.


H.J.