[PATCH] match.pd: Simplify 1 / X for integer X [PR95424]

[Jakub Jelinek]

On Thu, Jan 06, 2022 at 05:52:03PM +0800, Zhao Wei Liew wrote:
> On Wed, 5 Jan 2022 at 17:55, Richard Biener <richard.guenther@gmail.com>
> wrote:
> 
> > On Wed, Jan 5, 2022 at 10:42 AM Jakub Jelinek <jakub@redhat.com> wrote:
> > >
> > > On Wed, Jan 05, 2022 at 10:38:53AM +0100, Richard Biener via Gcc-patches
> > wrote:
> > > > On Wed, Jan 5, 2022 at 10:18 AM Zhao Wei Liew <zhaoweiliew@gmail.com>
> > wrote:
> > > > >
> > > > > > X >= -1 && X <= 1 is (hopefully?) going to be simplified
> > > > > > as  (unsigned)X + 1 <= 2, it might be good to simplify it this way
> > > > > > here as well?
> > > > >
> > > > > Yup, GCC does simplify it that way in the end, so I didn't really
> > bother to simplify it here. That said, I'm open to simplifying it here as
> > well, but I'm not sure how to do the unsigned cast.
> > > >
> > > > You'd do sth like
> > > >   (with
> > > >     {  tree utype = unsigned_type_for (type); }
> > > >     (cond (le (plus (convert:utype @0) { build_one_cst (utype); }) {
> > > > build_int_cst (utype, 2); }) ...)
> > > >
> > > > extra tricky will be 1 bit integer types, I guess it might be easiest
> > > > to exclude them
> > > > and special case them separately - X / Y is always X for them I think,
> > >
> > > Note, we already have:
> > >  /* X / bool_range_Y is X.  */
> > >  (simplify
> > >   (div @0 SSA_NAME@1)
> > >   (if (INTEGRAL_TYPE_P (type) && ssa_name_has_boolean_range (@1))
> > >    @0))
> > > for those.
> >
> > Ah, it might not handle the signed : 1 case though since -1 is not in the
> > bool range.  We could generalize the above though.
> >
> > Richard.
> >
> > >
> > >         Jakub
> > >
> >
> 
> Perhaps it is possible to exclude 1-bit cases and rely on other existing
> simplifications to deal with them?
> In particular, GCC seems to already optimally simplify the signed and
> unsigned 1-bit cases [1].

Sure, I thought that was what Richi was suggesting, so ad some
 && TYPE_PRECISION (...) > 1
somewhere (after making sure it is INTEGRAL_TYPE_P), or element_precision
if it is ANY_INTEGRAL_TYPE_P and could be vector type.

The discussion was about whether we optimize even those : 1 cases
sufficiently but that can be done in different simplifications...

	Jakub