[PATCH] Fix PR25620, pow() expansion missed-optimization, 2nd try

[Richard Guenther]

On Mon, 13 Nov 2006, Richard Guenther wrote:

> On Sun, 12 Nov 2006, Roger Sayle wrote:
> 
> > 
> > Hi Richard,
> > 
> > On Sun, 12 Nov 2006, Richard Guenther wrote:
> > > Bootstrapped and regtested on x86_64-unknown-linux-gnu.  SPEC 2k
> > > doesn't show a difference, Polyhedron has an improvement for air
> > > (down from 21s to 16.5s runtime).
> > 
> > Cool!  I hadn't appreciated this had such a significant impact on
> > polyhedron.
> > 
> > 
> > This is much better, thanks.  However, there are still a number of
> > (new) issues that need to be resolved.
> > 
> > The first is a potential bug...
> > 
> > > ! 	      expand_simple_binop (mode, MULT, op, op2, op, 0, OPTAB_DIRECT);
> > 
> > The expand_binop, and expand_simple_binop, APIs treat their target
> > argument as a hint.  Hence the return result of expand_simple_binop
> > isn't guaranteed to be in target_rtx.  Hence you need to use
> > 
> > 	op = expand_simple_binop (....)
> 
> Whoops.  Looks like I have to fix some of the i386 rounding expanders
> as well here...

One other thing here is when to use OPTAB_DIRECT vs. other methods.  I
see that expand_powi for creating the reciprocal uses OPTAB_LIB_WIDEN -
does this mean that for example on soft-float targets OPTAB_DIRECT will
"fail" and so we need OPTAB_LIB_WIDEN?

Thanks,
Richard.

--
Richard Guenther <rguenther@suse.de>
Novell / SUSE Labs