Re: [PATCH] range-op-float: Fix up -ffinite-math-only range extension and don't extend into infinities [PR109008]

[Aldy Hernandez <aldyh@redhat.com>]

On Mon, Mar 20, 2023 at 5:14 PM Jakub Jelinek <jakub@redhat.com> wrote:
>
> On Mon, Mar 13, 2023 at 09:41:47AM +0100, Aldy Hernandez wrote:
> > On 3/13/23 09:06, Jakub Jelinek wrote:
> > > On Mon, Mar 13, 2023 at 08:59:15AM +0100, Aldy Hernandez wrote:
> > > > > Yes, sure - I just noticed that we're forced to use high-level API for
> > > > > something that's quite low-level and should be internal (a range
> > > > > "breaking" internal consistency checks).
> > > >
> > > > Yeah, let's fix the API.  No sense hacking around things if what we need is
> > > > to tweak the design.
> > > >
> > > > I don't like hacking around things.  It always comes back to bite me ;-).
> > >
> > > Sure.  The current state is that I think the actual bugs are fixed except
> > > for the !MODE_HAS_INFINITIES case which people rarely use, so fixing up the
> > > API can wait even to next release.
> > >
> > > For !MODE_HAS_INFINITIES, I wonder if the best fix wouldn't be to change
> > > set and a few other spots, so that if the boundaries are
> > > real_min_representable/real_max_representable, we widen them to -inf and inf
> > > and change frange_val_min/max to also be dconstninf/dconstinf for
> > > !MODE_HAS_INFINITIES, because the min/max for that case (probably) really work as
> > > infinities.  Whenever we actually round that value to mode, it will become
> > > real_min_representable/real_max_representable again.
> > > But that can also wait a week.
> >
> > That sounds very reasonable.  It would remove special casing and would make
> > the code easier to read.  For that matter, that was what I had in the
> > original implementation.
>
> I think we don't want to remove the special casing for -ffinite-math-only
> on types which do support infinities.
> Thinking further on it, perhaps for !MODE_HAS_INFINITIES a better fix would
> be to do something like the patch below.
> Consider say having a range of VAX float type:
> #define M0 -FLT_MAX
> #define M1 nextafterf (F0, FLT_MAX)
> #define M2 nextafterf (M1, FLT_MAX)
> [M2, M2] - [M0, M1]
> Or perhaps if one or both of the operands are in such a case a min and max,
> perform real_arithmetic recurse on the argument replaced with
> dconstninf/dconstinf and then depending on inf pick the mininum or maximum
> of the two results (and carefully think about what to do if both operands
> are min/max).

LGTM.
Aldy

>
> 2023-03-20  Jakub Jelinek  <jakub@redhat.com>
>
>         * range-op-float.cc (frange_arithmetic): For !MODE_HAS_INFINITIES
>         types, pretend operands with minimum or maximum values are actually
>         infinities.
>
> --- gcc/range-op-float.cc.jj    2023-03-10 12:40:19.673108938 +0100
> +++ gcc/range-op-float.cc       2023-03-20 16:58:36.604981486 +0100
> @@ -313,8 +313,26 @@ frange_arithmetic (enum tree_code code,
>    REAL_VALUE_TYPE value;
>    enum machine_mode mode = TYPE_MODE (type);
>    bool mode_composite = MODE_COMPOSITE_P (mode);
> +  const REAL_VALUE_TYPE *pop1 = &op1;
> +  const REAL_VALUE_TYPE *pop2 = &op2;
>
> -  bool inexact = real_arithmetic (&value, code, &op1, &op2);
> +  if (!MODE_HAS_INFINITIES (mode))
> +    {
> +      // If mode doesn't have infinities, the minimum and maximum
> +      // values are saturating.  Pretend for real_arithmetic such
> +      // values are actual infinities.  real_convert will then
> +      // canonicalize the result not to be an infinity.
> +      if (frange_val_is_min (op1, type))
> +       pop1 = &dconstninf;
> +      else if (frange_val_is_max (op1, type))
> +       pop1 = &dconstinf;
> +      if (frange_val_is_min (op2, type))
> +       pop2 = &dconstninf;
> +      else if (frange_val_is_max (op2, type))
> +       pop2 = &dconstinf;
> +    }
> +
> +  bool inexact = real_arithmetic (&value, code, pop1, pop2);
>    real_convert (&result, mode, &value);
>
>    // Be extra careful if there may be discrepancies between the
>
>
>         Jakub
>