public inbox for gcc-patches@gcc.gnu.org
 help / color / mirror / Atom feed
From: Richard Biener <richard.guenther@gmail.com>
To: Aldy Hernandez <aldyh@redhat.com>
Cc: GCC patches <gcc-patches@gcc.gnu.org>,
	Jakub Jelinek <jakub@redhat.com>,
	 Andrew MacLeod <amacleod@redhat.com>
Subject: Re: [PATCH] frange: drop endpoints to min/max representable numbers for -ffinite-math-only.
Date: Fri, 23 Sep 2022 09:51:13 +0200	[thread overview]
Message-ID: <CAFiYyc0i=4vQ=mP=NWQtjiLyybbimN0zuuOC=jjFLKhGk6kMZw@mail.gmail.com> (raw)
In-Reply-To: <CAGm3qMUYD6u0g4wS6Y_JDXAY2aR=iECfEkTTobT1AOn8CiByRQ@mail.gmail.com>

On Fri, Sep 23, 2022 at 9:21 AM Aldy Hernandez <aldyh@redhat.com> wrote:
>
> Ughhh, my bad.  I had reworked this as soon as Jakub said we couldn't
> cache the min/max in TYPE_MIN/MAX_VALUE, but forgot to send it.  And
> yes...the incessant wrapping was very annoying.  It's all fixed.
>
> Let me know what you think.

Much better.

Thanks,
Richard.

> Aldy
>
> On Fri, Sep 23, 2022 at 9:04 AM Richard Biener
> <richard.guenther@gmail.com> wrote:
> >
> > On Thu, Sep 22, 2022 at 6:49 PM Aldy Hernandez <aldyh@redhat.com> wrote:
> > >
> > > Similarly to how we drop NANs to UNDEFINED when -ffinite-math-only, I
> > > think we can drop the numbers outside of the min/max representable
> > > numbers to the representable number.
> > >
> > > This means the endpoings to VR_VARYING for -ffinite-math-only can now
> > > be the min/max representable, instead of -INF and +INF.
> > >
> > > Saturating in the setter means that the upcoming implementation for
> > > binary operators no longer have to worry about doing the right
> > > thing for -ffinite-math-only.  If the range goes outside the limits,
> > > it'll get chopped down.
> > >
> > > How does this look?
> > >
> > > Tested on x86-64 Linux.
> > >
> > > gcc/ChangeLog:
> > >
> > >         * range-op-float.cc (build_le): Use vrp_val_*.
> > >         (build_lt): Same.
> > >         (build_ge): Same.
> > >         (build_gt): Same.
> > >         * value-range.cc (frange::set): Chop ranges outside of the
> > >         representable numbers for -ffinite-math-only.
> > >         (frange::normalize_kind): Use vrp_val*.
> > >         (frange::verify_range): Same.
> > >         (frange::set_nonnegative): Same.
> > >         (range_tests_floats): Remove tests that depend on -INF and +INF.
> > >         * value-range.h (real_max_representable): Add prototype.
> > >         (real_min_representable): Same.
> > >         (vrp_val_max): Set max representable number for
> > >         -ffinite-math-only.
> > >         (vrp_val_min): Same but for min.
> > >         (frange::set_varying): Use vrp_val*.
> > > ---
> > >  gcc/range-op-float.cc | 12 +++++++----
> > >  gcc/value-range.cc    | 46 ++++++++++++++++++++-----------------------
> > >  gcc/value-range.h     | 30 ++++++++++++++++++++++------
> > >  3 files changed, 53 insertions(+), 35 deletions(-)
> > >
> > > diff --git a/gcc/range-op-float.cc b/gcc/range-op-float.cc
> > > index 2bd3dc9253f..15ba19c2deb 100644
> > > --- a/gcc/range-op-float.cc
> > > +++ b/gcc/range-op-float.cc
> > > @@ -232,7 +232,8 @@ build_le (frange &r, tree type, const frange &val)
> > >  {
> > >    gcc_checking_assert (!val.known_isnan ());
> > >
> > > -  r.set (type, dconstninf, val.upper_bound ());
> > > +  REAL_VALUE_TYPE ninf = *TREE_REAL_CST_PTR (vrp_val_min (type));
> > > +  r.set (type, ninf, val.upper_bound ());
> > >
> > >    // Add both zeros if there's the possibility of zero equality.
> > >    frange_add_zeros (r, type);
> > > @@ -257,7 +258,8 @@ build_lt (frange &r, tree type, const frange &val)
> > >        return false;
> > >      }
> > >    // We only support closed intervals.
> > > -  r.set (type, dconstninf, val.upper_bound ());
> > > +  REAL_VALUE_TYPE ninf = *TREE_REAL_CST_PTR (vrp_val_min (type));
> > > +  r.set (type, ninf, val.upper_bound ());
> > >    return true;
> > >  }
> > >
> > > @@ -268,7 +270,8 @@ build_ge (frange &r, tree type, const frange &val)
> > >  {
> > >    gcc_checking_assert (!val.known_isnan ());
> > >
> > > -  r.set (type, val.lower_bound (), dconstinf);
> > > +  REAL_VALUE_TYPE inf = *TREE_REAL_CST_PTR (vrp_val_max (type));
> > > +  r.set (type, val.lower_bound (), inf);
> > >
> > >    // Add both zeros if there's the possibility of zero equality.
> > >    frange_add_zeros (r, type);
> > > @@ -294,7 +297,8 @@ build_gt (frange &r, tree type, const frange &val)
> > >      }
> > >
> > >    // We only support closed intervals.
> > > -  r.set (type, val.lower_bound (), dconstinf);
> > > +  REAL_VALUE_TYPE inf = *TREE_REAL_CST_PTR (vrp_val_max (type));
> > > +  r.set (type, val.lower_bound (), inf);
> > >    return true;
> > >  }
> > >
> > > diff --git a/gcc/value-range.cc b/gcc/value-range.cc
> > > index 7e8028eced2..e57d60e1bac 100644
> > > --- a/gcc/value-range.cc
> > > +++ b/gcc/value-range.cc
> > > @@ -338,6 +338,18 @@ frange::set (tree min, tree max, value_range_kind kind)
> > >        m_neg_nan = false;
> > >      }
> > >
> > > +  // For -ffinite-math-only we can drop ranges outside the
> > > +  // representable numbers to min/max for the type.
> > > +  if (flag_finite_math_only)
> > > +    {
> > > +      REAL_VALUE_TYPE min_repr = *TREE_REAL_CST_PTR (vrp_val_min (m_type));
> > > +      REAL_VALUE_TYPE max_repr = *TREE_REAL_CST_PTR (vrp_val_max (m_type));
> > > +      if (real_less (&m_min, &min_repr))
> > > +       m_min = min_repr;
> > > +      if (real_less (&max_repr, &m_max))
> > > +       m_max = max_repr;
> >
> > I think you want to re-formulate that in terms of real_isinf() and
> > change those to the max representable values.
> >
> > > +    }
> > > +
> > >    // Check for swapped ranges.
> > >    gcc_checking_assert (tree_compare (LE_EXPR, min, max));
> > >
> > > @@ -371,8 +383,8 @@ bool
> > >  frange::normalize_kind ()
> > >  {
> > >    if (m_kind == VR_RANGE
> > > -      && real_isinf (&m_min, 1)
> > > -      && real_isinf (&m_max, 0))
> > > +      && vrp_val_is_min (build_real (m_type, m_min))
> > > +      && vrp_val_is_max (build_real (m_type, m_max)))
> >
> > I think this suggests that keeping +-Inf might be the better choice.
> > At least you want to change vrp_val_is_min/max to take a REAL_VALUE_TYPE
> > to avoid building four tree nodes for just this compare?
> >
> > >      {
> > >        if (m_pos_nan && m_neg_nan)
> > >         {
> > > @@ -385,8 +397,8 @@ frange::normalize_kind ()
> > >        if (!m_pos_nan || !m_neg_nan)
> > >         {
> > >           m_kind = VR_RANGE;
> > > -         m_min = dconstninf;
> > > -         m_max = dconstinf;
> > > +         m_min = *TREE_REAL_CST_PTR (vrp_val_min (m_type));
> > > +         m_max = *TREE_REAL_CST_PTR (vrp_val_max (m_type));
> > >           return true;
> > >         }
> > >      }
> > > @@ -706,8 +718,8 @@ frange::verify_range ()
> > >      case VR_VARYING:
> > >        gcc_checking_assert (m_type);
> > >        gcc_checking_assert (m_pos_nan && m_neg_nan);
> > > -      gcc_checking_assert (real_isinf (&m_min, 1));
> > > -      gcc_checking_assert (real_isinf (&m_max, 0));
> > > +      gcc_checking_assert (vrp_val_is_min (build_real (m_type, m_min)));
> > > +      gcc_checking_assert (vrp_val_is_max (build_real (m_type, m_max)));
> > >        return;
> > >      case VR_RANGE:
> > >        gcc_checking_assert (m_type);
> > > @@ -732,7 +744,8 @@ frange::verify_range ()
> > >    // If all the properties are clear, we better not span the entire
> > >    // domain, because that would make us varying.
> > >    if (m_pos_nan && m_neg_nan)
> > > -    gcc_checking_assert (!real_isinf (&m_min, 1) || !real_isinf (&m_max, 0));
> > > +    gcc_checking_assert (!vrp_val_is_min (build_real (m_type, m_min))
> > > +                        || !vrp_val_is_max (build_real (m_type, m_max)));
> > >  }
> > >
> > >  // We can't do much with nonzeros yet.
> > > @@ -779,7 +792,7 @@ frange::zero_p () const
> > >  void
> > >  frange::set_nonnegative (tree type)
> > >  {
> > > -  set (type, dconst0, dconstinf);
> > > +  set (type, dconst0, *TREE_REAL_CST_PTR (vrp_val_max (type)));
> > >
> > >    // Set +NAN as the only possibility.
> > >    if (HONOR_NANS (type))
> > > @@ -3886,23 +3899,6 @@ range_tests_floats ()
> > >    r0.clear_nan ();
> > >    ASSERT_FALSE (r0.varying_p ());
> > >
> > > -  // The endpoints of a VARYING are +-INF.
> > > -  r0.set_varying (float_type_node);
> > > -  ASSERT_TRUE (real_identical (&r0.lower_bound (), &dconstninf));
> > > -  ASSERT_TRUE (real_identical (&r0.upper_bound (), &dconstinf));
> > > -
> > > -  // The maximum representable range for a type is still a subset of VARYING.
> > > -  REAL_VALUE_TYPE q, r;
> > > -  real_min_representable (&q, float_type_node);
> > > -  real_max_representable (&r, float_type_node);
> > > -  r0 = frange (float_type_node, q, r);
> > > -  // r0 is not a varying, because it does not include -INF/+INF.
> > > -  ASSERT_FALSE (r0.varying_p ());
> > > -  // The upper bound of r0 must be less than +INF.
> > > -  ASSERT_TRUE (real_less (&r0.upper_bound (), &dconstinf));
> > > -  // The lower bound of r0 must be greater than -INF.
> > > -  ASSERT_TRUE (real_less (&dconstninf, &r0.lower_bound ()));
> > > -
> > >    // For most architectures, where float and double are different
> > >    // sizes, having the same endpoints does not necessarily mean the
> > >    // ranges are equal.
> > > diff --git a/gcc/value-range.h b/gcc/value-range.h
> > > index 3668b331187..74879b76b40 100644
> > > --- a/gcc/value-range.h
> > > +++ b/gcc/value-range.h
> > > @@ -593,6 +593,8 @@ extern void dump_value_range (FILE *, const vrange *);
> > >  extern bool vrp_val_is_min (const_tree);
> > >  extern bool vrp_val_is_max (const_tree);
> > >  extern bool vrp_operand_equal_p (const_tree, const_tree);
> > > +inline void real_max_representable (REAL_VALUE_TYPE *r, const_tree type);
> > > +inline void real_min_representable (REAL_VALUE_TYPE *r, const_tree type);
> > >
> > >  inline value_range_kind
> > >  vrange::kind () const
> > > @@ -1009,7 +1011,15 @@ vrp_val_max (const_tree type)
> > >        return wide_int_to_tree (const_cast<tree> (type), max);
> > >      }
> > >    if (frange::supports_p (type))
> > > -    return build_real (const_cast <tree> (type), dconstinf);
> > > +    {
> > > +      if (flag_finite_math_only)
> > > +       {
> > > +         REAL_VALUE_TYPE r;
> > > +         real_max_representable (&r, type);
> > > +         return build_real (const_cast <tree> (type), r);
> > > +       }
> > > +      return build_real (const_cast <tree> (type), dconstinf);
> >
> > Overall there's too many places affected IMHO.  The above also shows
> > we're building trees a lot - note that REAL_CST are _not_ shared, so
> > each time you call build_real a new REAL_CST tree node is allocated
> > (and later GCed if no longer used).
> >
> > I thought vrp_val_min/max are legacy by now ...
> >
> > That said, using trees for the range representation might not be the
> > very best thing to do, it's unnecessarily wrapping things.
> >
> > > +    }
> > >    return NULL_TREE;
> > >  }
> > >
> > > @@ -1023,7 +1033,15 @@ vrp_val_min (const_tree type)
> > >    if (POINTER_TYPE_P (type))
> > >      return build_zero_cst (const_cast<tree> (type));
> > >    if (frange::supports_p (type))
> > > -    return build_real (const_cast <tree> (type), dconstninf);
> > > +    {
> > > +      if (flag_finite_math_only)
> > > +       {
> > > +         REAL_VALUE_TYPE r;
> > > +         real_min_representable (&r, type);
> > > +         return build_real (const_cast <tree> (type), r);
> > > +       }
> > > +      return build_real (const_cast <tree> (type), dconstninf);
> > > +    }
> > >    return NULL_TREE;
> > >  }
> > >
> > > @@ -1073,8 +1091,8 @@ frange::set_varying (tree type)
> > >  {
> > >    m_kind = VR_VARYING;
> > >    m_type = type;
> > > -  m_min = dconstninf;
> > > -  m_max = dconstinf;
> > > +  m_min = *TREE_REAL_CST_PTR (vrp_val_min (type));
> > > +  m_max = *TREE_REAL_CST_PTR (vrp_val_max (type));
> > >    m_pos_nan = true;
> > >    m_neg_nan = true;
> > >  }
> > > @@ -1133,7 +1151,7 @@ frange::clear_nan ()
> > >  // Set R to maximum representable value for TYPE.
> > >
> > >  inline void
> > > -real_max_representable (REAL_VALUE_TYPE *r, tree type)
> > > +real_max_representable (REAL_VALUE_TYPE *r, const_tree type)
> > >  {
> > >    char buf[128];
> > >    get_max_float (REAL_MODE_FORMAT (TYPE_MODE (type)),
> > > @@ -1145,7 +1163,7 @@ real_max_representable (REAL_VALUE_TYPE *r, tree type)
> > >  // Set R to minimum representable value for TYPE.
> > >
> > >  inline void
> > > -real_min_representable (REAL_VALUE_TYPE *r, tree type)
> > > +real_min_representable (REAL_VALUE_TYPE *r, const_tree type)
> > >  {
> > >    real_max_representable (r, type);
> > >    *r = real_value_negate (r);
> > > --
> > > 2.37.1
> > >
> >

  reply	other threads:[~2022-09-23  7:51 UTC|newest]

Thread overview: 9+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2022-09-22 16:49 [PATCH] frange: dump hex values when dumping FP numbers Aldy Hernandez
2022-09-22 16:49 ` [PATCH] frange: drop endpoints to min/max representable numbers for -ffinite-math-only Aldy Hernandez
2022-09-23  7:03   ` Richard Biener
2022-09-23  7:21     ` Aldy Hernandez
2022-09-23  7:51       ` Richard Biener [this message]
2022-09-22 19:23 ` [PATCH] frange: dump hex values when dumping FP numbers Toon Moene
2022-09-22 19:25 ` Jeff Law
2022-09-22 21:04 ` Jakub Jelinek
2022-09-23  6:56   ` Aldy Hernandez

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to='CAFiYyc0i=4vQ=mP=NWQtjiLyybbimN0zuuOC=jjFLKhGk6kMZw@mail.gmail.com' \
    --to=richard.guenther@gmail.com \
    --cc=aldyh@redhat.com \
    --cc=amacleod@redhat.com \
    --cc=gcc-patches@gcc.gnu.org \
    --cc=jakub@redhat.com \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).