Re: [PATCH] New if-combining pass, addresses PRs 15353 and 31657

["Andrew Pinski" <pinskia at gmail dot com>]

On 4/26/07, Richard Guenther <rguenther@suse.de> wrote:
> On Wed, 25 Apr 2007, Andrew Pinski wrote:
>
> > On 4/25/07, Richard Guenther <rguenther@suse.de> wrote:
> >
> > > +       /* returning 1 causes phiopt to trigger in */
> > If this is true, then this pass should run before phiopt so you can
> > catch more stuff.
>
> Well, I thought it is not profitable to do the transformation if phiopt
> can remove the same amount of control flow.  Basically replacing
>
>   x = (a >> b)
>   if (x & 1)
>     y = (a >> c)
>     if (y & 1)
>       return 1;
>     else
>       return 0;
>
> with
>
>   tmp = (1 << b) | (1 << c)
>   if (a & tmp == tmp)
>     return 1;
>   else
>     return 0;
>
> instead of
>
>   x = (a >> b)
>   if (x & 1)
>     y = (a >> c)
>     return y & 1;

Well it depends on the target :).  Some (like spu-elf) non hinted
branches (and mispredicted hinted branches) are really really bad so
the first case of getting rid of an extra branch is a good thing.

Thanks,
Andrew Pinski