Re: [PATCH] [ranger] x == -0.0 does not mean we can replace x with -0.0

[Aldy Hernandez <aldyh@redhat.com>]

On Fri, Aug 26, 2022 at 6:40 PM Jakub Jelinek <jakub@redhat.com> wrote:
>
> On Fri, Aug 26, 2022 at 05:46:06PM +0200, Aldy Hernandez wrote:
> > On the true side of x == -0.0, we can't just blindly value propagate
> > the -0.0 into every use of x because x could be +0.0 (or vice versa).
> >
> > With this change, we only allow the transformation if
> > !HONOR_SIGNED_ZEROS or if the range is known not to contain 0.
> >
> > Will commit after tests complete.
> >
> > gcc/ChangeLog:
> >
> >       * range-op-float.cc (foperator_equal::op1_range): Do not blindly
> >       copy op2 range when honoring signed zeros.
> > ---
> >  gcc/range-op-float.cc | 17 +++++++++++++++--
> >  1 file changed, 15 insertions(+), 2 deletions(-)
> >
> > diff --git a/gcc/range-op-float.cc b/gcc/range-op-float.cc
> > index ad2fae578d2..ff9fe312acf 100644
> > --- a/gcc/range-op-float.cc
> > +++ b/gcc/range-op-float.cc
> > @@ -252,8 +252,21 @@ foperator_equal::op1_range (frange &r, tree type,
> >    switch (get_bool_state (r, lhs, type))
> >      {
> >      case BRS_TRUE:
> > -      // If it's true, the result is the same as OP2.
> > -      r = op2;
> > +      if (HONOR_SIGNED_ZEROS (type)
> > +       && op2.contains_p (build_zero_cst (type)))
>
> What exactly does op2.contains_p for zero?
> Does it use real_compare/real_equal under the hood, so it is
> equivalent to op2 == 0.0 or op2 == -0.0, where both will be
> true whether op2 is -0.0 or 0.0?  Or is it more strict and
> checks whether it is actually a positive zero?

frange::contains_p() uses real_compare(), so both -0.0 and 0.0 will
come out as true:

  return (real_compare (GE_EXPR, TREE_REAL_CST_PTR (cst), &m_min)
      && real_compare (LE_EXPR, TREE_REAL_CST_PTR (cst), &m_max));

I thought about this some more, and you're right, dropping to VARYING
is a big hammer.

It seems to me we can do this optimization regardless, but then treat
positive and negative zero the same throughout the frange class.
Particularly, in frange::singleton_p().  We should never return TRUE
for any version of 0.0.  This will keep VRP from propagating an
incorrect 0.0, since all VRP does is propagate when a range is
provably a singleton.  Also, frange::zero_p() shall return true for
any version of 0.0.

I fleshed out all the relational operators (with endpoints) with this
approach, and everything worked out...including go, ada, and fortran,
which had given me headaches.  As a bonus, we can get rid of the
INF/NINF property bits I was keeping around, since now the range will
have actual endpoints.  I will repost the full frange endpoints patch
(and CC you) in the appropriate thread.

Aldy

> In any case, for HONOR_SIGNED_ZEROS, VARYING is unnecessary, all you
> can do is extend the r range to contain both -0.0 and +0.0 if it contains
> at least one of them.
>
> > +     {
> > +       // With signed zeros, x == -0.0 does not mean we can replace
> > +       // x with -0.0, because x may be either +0.0 or -0.0.
> > +       r.set_varying (type);
> > +     }
> > +      else
> > +     {
> > +       // If it's true, the result is the same as OP2.
> > +       //
> > +       // If the range does not actually contain zeros, this should
> > +       // always be OK.
> > +       r = op2;
> > +     }
>
> !HONOR_SIGNED_ZEROS doesn't imply that negative zeros won't appear,
> but says that user doesn't care if he gets a positive or negative zero
> (unless !MODE_HAS_SIGNED_ZEROS - in that case -0.0 doesn't exist
> and one doesn't need to bother with it).
>
> Now, if all the code setting franges makes sure that for
> MODE_HAS_SIGNED_ZEROS && !HONOR_SIGNED_ZEROS if +0.0 or -0.0 are inside
> of a range, then both -0.0 and +0.0 are in the range, then yes,
> you can use r = op2;
>
> >        // The TRUE side of op1 == op2 implies op1 is !NAN.
> >        r.set_nan (fp_prop::NO);
> >        break;
>
>         Jakub
>