frange: flush denormals to zero for -funsafe-math-optimizations.

Message ID 20220917082403.1573721-1-aldyh@redhat.com
State New, archived
Headers
Series frange: flush denormals to zero for -funsafe-math-optimizations. |

Commit Message

Aldy Hernandez Sept. 17, 2022, 8:24 a.m. UTC
  Jakub has mentioned that for -funsafe-math-optimizations we may flush
denormals to zero, in which case we need to be careful to extend the
ranges to the appropriate zero.  This patch does exactly that.  For a
range of [x, -DENORMAL] we flush to [x, -0.0] and for [+DENORMAL, x]
we flush to [+0.0, x].

It is unclear whether we should do this for Alpha, since I believe
flushing to zero is the default, and the port requires -mieee for IEEE
sanity.  If so, perhaps we should add a target hook so backends are
free to request flushing to zero.

Thoughts?

gcc/ChangeLog:

	* value-range.cc (frange::flush_denormals_to_zero): New.
	(frange::set): Call flush_denormals_to_zero.
	* value-range.h (class frange): Add flush_denormals_to_zero.
---
 gcc/value-range.cc | 24 ++++++++++++++++++++++++
 gcc/value-range.h  |  1 +
 2 files changed, 25 insertions(+)
  

Comments

Richard Biener Sept. 19, 2022, 7:37 a.m. UTC | #1
On Sat, Sep 17, 2022 at 10:25 AM Aldy Hernandez via Gcc-patches
<gcc-patches@gcc.gnu.org> wrote:
>
> Jakub has mentioned that for -funsafe-math-optimizations we may flush
> denormals to zero, in which case we need to be careful to extend the
> ranges to the appropriate zero.  This patch does exactly that.  For a
> range of [x, -DENORMAL] we flush to [x, -0.0] and for [+DENORMAL, x]
> we flush to [+0.0, x].
>
> It is unclear whether we should do this for Alpha, since I believe
> flushing to zero is the default, and the port requires -mieee for IEEE
> sanity.  If so, perhaps we should add a target hook so backends are
> free to request flushing to zero.
>
> Thoughts?

I'm not sure what the intention of this is - it effectively results in
more conservative ranges for -funsafe-math-optimizations.  That is,
if -funsafe-math-optimizations says that denormals don't exist and
are all zero then doesn't that mean we can instead use the smalles
non-denormal number less than (or greater than) zero here?

That said, the flushing you do is also "valid" for
-fno-unsafe-math-optimizations
in case we don't want to deal with denormals in range boundaries.

It might also be a correctness requirement in case we don't know how
targets handle denormals (IIRC even OS defaults might matter here) so
we conservatively have to treat them as being flushed to zero.

So ... if we want to be on the "safe" side then please always do this.

At the same point you could do

 if (!HONOR_SIGNED_ZEROS ())
   if (real_iszero (&m_max))
     {
        if (real_iszero (&m_min))
           m.min.sign = 1;
        m_max.sign = 1;
     }
   else if (real_iszero (&m_min))
     m_min.sign = 0;

so that we canonicalize a zero bound so that the sign is known for a range.

Richard.

> gcc/ChangeLog:
>
>         * value-range.cc (frange::flush_denormals_to_zero): New.
>         (frange::set): Call flush_denormals_to_zero.
>         * value-range.h (class frange): Add flush_denormals_to_zero.
> ---
>  gcc/value-range.cc | 24 ++++++++++++++++++++++++
>  gcc/value-range.h  |  1 +
>  2 files changed, 25 insertions(+)
>
> diff --git a/gcc/value-range.cc b/gcc/value-range.cc
> index 67d5d7fa90f..f285734f0e0 100644
> --- a/gcc/value-range.cc
> +++ b/gcc/value-range.cc
> @@ -267,6 +267,26 @@ tree_compare (tree_code code, tree op1, tree op2)
>    return !integer_zerop (fold_build2 (code, integer_type_node, op1, op2));
>  }
>
> +// Flush denormal endpoints to the appropriate 0.0.
> +
> +void
> +frange::flush_denormals_to_zero ()
> +{
> +  if (undefined_p () || known_isnan ())
> +    return;
> +
> +  // Flush [x, -DENORMAL] to [x, -0.0].
> +  if (real_isdenormal (&m_max) && real_isneg (&m_max))
> +    {
> +      m_max = dconst0;
> +      if (HONOR_SIGNED_ZEROS (m_type))
> +       m_max.sign = 1;
> +    }
> +  // Flush [+DENORMAL, x] to [+0.0, x].
> +  if (real_isdenormal (&m_min) && !real_isneg (&m_min))
> +    m_min = dconst0;
> +}
> +
>  // Setter for franges.
>
>  void
> @@ -317,6 +337,10 @@ frange::set (tree min, tree max, value_range_kind kind)
>    gcc_checking_assert (tree_compare (LE_EXPR, min, max));
>
>    normalize_kind ();
> +
> +  if (flag_unsafe_math_optimizations)
> +    flush_denormals_to_zero ();
> +
>    if (flag_checking)
>      verify_range ();
>  }
> diff --git a/gcc/value-range.h b/gcc/value-range.h
> index 3a401f3e4e2..795b1f00fdc 100644
> --- a/gcc/value-range.h
> +++ b/gcc/value-range.h
> @@ -327,6 +327,7 @@ private:
>    bool union_nans (const frange &);
>    bool intersect_nans (const frange &);
>    bool combine_zeros (const frange &, bool union_p);
> +  void flush_denormals_to_zero ();
>
>    tree m_type;
>    REAL_VALUE_TYPE m_min;
> --
> 2.37.1
>
  
Jakub Jelinek Sept. 19, 2022, 8:25 a.m. UTC | #2
On Mon, Sep 19, 2022 at 09:37:22AM +0200, Richard Biener wrote:
> On Sat, Sep 17, 2022 at 10:25 AM Aldy Hernandez via Gcc-patches
> <gcc-patches@gcc.gnu.org> wrote:
> >
> > Jakub has mentioned that for -funsafe-math-optimizations we may flush
> > denormals to zero, in which case we need to be careful to extend the
> > ranges to the appropriate zero.  This patch does exactly that.  For a
> > range of [x, -DENORMAL] we flush to [x, -0.0] and for [+DENORMAL, x]
> > we flush to [+0.0, x].
> >
> > It is unclear whether we should do this for Alpha, since I believe
> > flushing to zero is the default, and the port requires -mieee for IEEE
> > sanity.  If so, perhaps we should add a target hook so backends are
> > free to request flushing to zero.
> >
> > Thoughts?
> 
> I'm not sure what the intention of this is - it effectively results in
> more conservative ranges for -funsafe-math-optimizations.  That is,
> if -funsafe-math-optimizations says that denormals don't exist and
> are all zero then doesn't that mean we can instead use the smalles
> non-denormal number less than (or greater than) zero here?

Yes, with -funsafe-math-optimizations we will sometimes have slightly
larger ranges, while when we know or assume we know that denormals
will be honored we can be more precise.
It is similar to the -fno-signed-zeros case, when we are more relaxed
and say we don't care about sign of zeros and optimizations can be
performed to leave the sign of zero in whatever state they like,
we can't have [+0.0, xx] or [xx, -0.0] ranges and need to extend them
to [-0.0, xx] or [xx, +0.0].  The honor denormals case is similar,
if there is the possibility of having denormals flushed to zero,
[+denormal, xx] needs to be extended to [+0.0, xx] and
[xx, -denormal] to [xx, -0.0].  Now, I think we just should add
a separate option whether denormals are honored or not, because
always extending the ranges with denormals on the boundaries might
prevent some useful optimizations.  If we add such an option, it
could default to not honoring them for -funsafe-math-optimizations,
and say on alpha it should default to not honoring them by default
always unless -mieee option is used, but all of this if the user
didn't use the option explicitly, in that case follow what the user
said.  That way if users know they set DTZ flag themselves or use
libraries that do so, they can tell the compiler about it.

> That said, the flushing you do is also "valid" for
> -fno-unsafe-math-optimizations
> in case we don't want to deal with denormals in range boundaries.

It is always valid to extend the range more, but could be harmful to
optimizations.  Say if we have then x > 0.0 test...
Of course, without -ffast-math we won't DCE various floating point
operations because they can raise exceptions, but at least we can DCE
the basic block after it that depends on x > 0.0.

	Jakub
  
Aldy Hernandez Sept. 19, 2022, 1:04 p.m. UTC | #3
On Mon, Sep 19, 2022 at 9:37 AM Richard Biener
<richard.guenther@gmail.com> wrote:
>
> On Sat, Sep 17, 2022 at 10:25 AM Aldy Hernandez via Gcc-patches
> <gcc-patches@gcc.gnu.org> wrote:
> >
> > Jakub has mentioned that for -funsafe-math-optimizations we may flush
> > denormals to zero, in which case we need to be careful to extend the
> > ranges to the appropriate zero.  This patch does exactly that.  For a
> > range of [x, -DENORMAL] we flush to [x, -0.0] and for [+DENORMAL, x]
> > we flush to [+0.0, x].
> >
> > It is unclear whether we should do this for Alpha, since I believe
> > flushing to zero is the default, and the port requires -mieee for IEEE
> > sanity.  If so, perhaps we should add a target hook so backends are
> > free to request flushing to zero.
> >
> > Thoughts?
>
> I'm not sure what the intention of this is - it effectively results in
> more conservative ranges for -funsafe-math-optimizations.  That is,
> if -funsafe-math-optimizations says that denormals don't exist and
> are all zero then doesn't that mean we can instead use the smalles
> non-denormal number less than (or greater than) zero here?
>
> That said, the flushing you do is also "valid" for
> -fno-unsafe-math-optimizations
> in case we don't want to deal with denormals in range boundaries.
>
> It might also be a correctness requirement in case we don't know how
> targets handle denormals (IIRC even OS defaults might matter here) so
> we conservatively have to treat them as being flushed to zero.

Actually, rth suggested we always flush to zero because we don't know
what the target would do.  Again, I'm happy to do whatever you agree
on.  I have no opinion.  My main goal here is correctness.

>
> So ... if we want to be on the "safe" side then please always do this.
>
> At the same point you could do
>
>  if (!HONOR_SIGNED_ZEROS ())
>    if (real_iszero (&m_max))
>      {
>         if (real_iszero (&m_min))
>            m.min.sign = 1;
>         m_max.sign = 1;
>      }

But wouldn't that set [-0.0, -0.0] when encountering [+0, +0] ??

>    else if (real_iszero (&m_min))
>      m_min.sign = 0;

Jakub actually suggested something completely different...just set +0
always for !HONOR_SIGNED_ZEROS.

Aldy

>
> so that we canonicalize a zero bound so that the sign is known for a range.
>
> Richard.
>
> > gcc/ChangeLog:
> >
> >         * value-range.cc (frange::flush_denormals_to_zero): New.
> >         (frange::set): Call flush_denormals_to_zero.
> >         * value-range.h (class frange): Add flush_denormals_to_zero.
> > ---
> >  gcc/value-range.cc | 24 ++++++++++++++++++++++++
> >  gcc/value-range.h  |  1 +
> >  2 files changed, 25 insertions(+)
> >
> > diff --git a/gcc/value-range.cc b/gcc/value-range.cc
> > index 67d5d7fa90f..f285734f0e0 100644
> > --- a/gcc/value-range.cc
> > +++ b/gcc/value-range.cc
> > @@ -267,6 +267,26 @@ tree_compare (tree_code code, tree op1, tree op2)
> >    return !integer_zerop (fold_build2 (code, integer_type_node, op1, op2));
> >  }
> >
> > +// Flush denormal endpoints to the appropriate 0.0.
> > +
> > +void
> > +frange::flush_denormals_to_zero ()
> > +{
> > +  if (undefined_p () || known_isnan ())
> > +    return;
> > +
> > +  // Flush [x, -DENORMAL] to [x, -0.0].
> > +  if (real_isdenormal (&m_max) && real_isneg (&m_max))
> > +    {
> > +      m_max = dconst0;
> > +      if (HONOR_SIGNED_ZEROS (m_type))
> > +       m_max.sign = 1;
> > +    }
> > +  // Flush [+DENORMAL, x] to [+0.0, x].
> > +  if (real_isdenormal (&m_min) && !real_isneg (&m_min))
> > +    m_min = dconst0;
> > +}
> > +
> >  // Setter for franges.
> >
> >  void
> > @@ -317,6 +337,10 @@ frange::set (tree min, tree max, value_range_kind kind)
> >    gcc_checking_assert (tree_compare (LE_EXPR, min, max));
> >
> >    normalize_kind ();
> > +
> > +  if (flag_unsafe_math_optimizations)
> > +    flush_denormals_to_zero ();
> > +
> >    if (flag_checking)
> >      verify_range ();
> >  }
> > diff --git a/gcc/value-range.h b/gcc/value-range.h
> > index 3a401f3e4e2..795b1f00fdc 100644
> > --- a/gcc/value-range.h
> > +++ b/gcc/value-range.h
> > @@ -327,6 +327,7 @@ private:
> >    bool union_nans (const frange &);
> >    bool intersect_nans (const frange &);
> >    bool combine_zeros (const frange &, bool union_p);
> > +  void flush_denormals_to_zero ();
> >
> >    tree m_type;
> >    REAL_VALUE_TYPE m_min;
> > --
> > 2.37.1
> >
>
  
Aldy Hernandez Sept. 19, 2022, 1:06 p.m. UTC | #4
On Mon, Sep 19, 2022 at 3:04 PM Aldy Hernandez <aldyh@redhat.com> wrote:
>
> On Mon, Sep 19, 2022 at 9:37 AM Richard Biener
> <richard.guenther@gmail.com> wrote:
> >
> > On Sat, Sep 17, 2022 at 10:25 AM Aldy Hernandez via Gcc-patches
> > <gcc-patches@gcc.gnu.org> wrote:
> > >
> > > Jakub has mentioned that for -funsafe-math-optimizations we may flush
> > > denormals to zero, in which case we need to be careful to extend the
> > > ranges to the appropriate zero.  This patch does exactly that.  For a
> > > range of [x, -DENORMAL] we flush to [x, -0.0] and for [+DENORMAL, x]
> > > we flush to [+0.0, x].
> > >
> > > It is unclear whether we should do this for Alpha, since I believe
> > > flushing to zero is the default, and the port requires -mieee for IEEE
> > > sanity.  If so, perhaps we should add a target hook so backends are
> > > free to request flushing to zero.
> > >
> > > Thoughts?
> >
> > I'm not sure what the intention of this is - it effectively results in
> > more conservative ranges for -funsafe-math-optimizations.  That is,
> > if -funsafe-math-optimizations says that denormals don't exist and
> > are all zero then doesn't that mean we can instead use the smalles
> > non-denormal number less than (or greater than) zero here?
> >
> > That said, the flushing you do is also "valid" for
> > -fno-unsafe-math-optimizations
> > in case we don't want to deal with denormals in range boundaries.
> >
> > It might also be a correctness requirement in case we don't know how
> > targets handle denormals (IIRC even OS defaults might matter here) so
> > we conservatively have to treat them as being flushed to zero.
>
> Actually, rth suggested we always flush to zero because we don't know
> what the target would do.  Again, I'm happy to do whatever you agree

More specifically, we don't know what the OS will do, so we either
should have a flag of some kind set to TRUE by default when unsafe
math optimizations, or always assume denormal flushing could happen.

Again, no opinion.  Up to y'all.

Aldy

> on.  I have no opinion.  My main goal here is correctness.
>
> >
> > So ... if we want to be on the "safe" side then please always do this.
> >
> > At the same point you could do
> >
> >  if (!HONOR_SIGNED_ZEROS ())
> >    if (real_iszero (&m_max))
> >      {
> >         if (real_iszero (&m_min))
> >            m.min.sign = 1;
> >         m_max.sign = 1;
> >      }
>
> But wouldn't that set [-0.0, -0.0] when encountering [+0, +0] ??
>
> >    else if (real_iszero (&m_min))
> >      m_min.sign = 0;
>
> Jakub actually suggested something completely different...just set +0
> always for !HONOR_SIGNED_ZEROS.
>
> Aldy
>
> >
> > so that we canonicalize a zero bound so that the sign is known for a range.
> >
> > Richard.
> >
> > > gcc/ChangeLog:
> > >
> > >         * value-range.cc (frange::flush_denormals_to_zero): New.
> > >         (frange::set): Call flush_denormals_to_zero.
> > >         * value-range.h (class frange): Add flush_denormals_to_zero.
> > > ---
> > >  gcc/value-range.cc | 24 ++++++++++++++++++++++++
> > >  gcc/value-range.h  |  1 +
> > >  2 files changed, 25 insertions(+)
> > >
> > > diff --git a/gcc/value-range.cc b/gcc/value-range.cc
> > > index 67d5d7fa90f..f285734f0e0 100644
> > > --- a/gcc/value-range.cc
> > > +++ b/gcc/value-range.cc
> > > @@ -267,6 +267,26 @@ tree_compare (tree_code code, tree op1, tree op2)
> > >    return !integer_zerop (fold_build2 (code, integer_type_node, op1, op2));
> > >  }
> > >
> > > +// Flush denormal endpoints to the appropriate 0.0.
> > > +
> > > +void
> > > +frange::flush_denormals_to_zero ()
> > > +{
> > > +  if (undefined_p () || known_isnan ())
> > > +    return;
> > > +
> > > +  // Flush [x, -DENORMAL] to [x, -0.0].
> > > +  if (real_isdenormal (&m_max) && real_isneg (&m_max))
> > > +    {
> > > +      m_max = dconst0;
> > > +      if (HONOR_SIGNED_ZEROS (m_type))
> > > +       m_max.sign = 1;
> > > +    }
> > > +  // Flush [+DENORMAL, x] to [+0.0, x].
> > > +  if (real_isdenormal (&m_min) && !real_isneg (&m_min))
> > > +    m_min = dconst0;
> > > +}
> > > +
> > >  // Setter for franges.
> > >
> > >  void
> > > @@ -317,6 +337,10 @@ frange::set (tree min, tree max, value_range_kind kind)
> > >    gcc_checking_assert (tree_compare (LE_EXPR, min, max));
> > >
> > >    normalize_kind ();
> > > +
> > > +  if (flag_unsafe_math_optimizations)
> > > +    flush_denormals_to_zero ();
> > > +
> > >    if (flag_checking)
> > >      verify_range ();
> > >  }
> > > diff --git a/gcc/value-range.h b/gcc/value-range.h
> > > index 3a401f3e4e2..795b1f00fdc 100644
> > > --- a/gcc/value-range.h
> > > +++ b/gcc/value-range.h
> > > @@ -327,6 +327,7 @@ private:
> > >    bool union_nans (const frange &);
> > >    bool intersect_nans (const frange &);
> > >    bool combine_zeros (const frange &, bool union_p);
> > > +  void flush_denormals_to_zero ();
> > >
> > >    tree m_type;
> > >    REAL_VALUE_TYPE m_min;
> > > --
> > > 2.37.1
> > >
> >
  
Richard Biener Sept. 19, 2022, 1:44 p.m. UTC | #5
On Mon, Sep 19, 2022 at 3:04 PM Aldy Hernandez <aldyh@redhat.com> wrote:
>
> On Mon, Sep 19, 2022 at 9:37 AM Richard Biener
> <richard.guenther@gmail.com> wrote:
> >
> > On Sat, Sep 17, 2022 at 10:25 AM Aldy Hernandez via Gcc-patches
> > <gcc-patches@gcc.gnu.org> wrote:
> > >
> > > Jakub has mentioned that for -funsafe-math-optimizations we may flush
> > > denormals to zero, in which case we need to be careful to extend the
> > > ranges to the appropriate zero.  This patch does exactly that.  For a
> > > range of [x, -DENORMAL] we flush to [x, -0.0] and for [+DENORMAL, x]
> > > we flush to [+0.0, x].
> > >
> > > It is unclear whether we should do this for Alpha, since I believe
> > > flushing to zero is the default, and the port requires -mieee for IEEE
> > > sanity.  If so, perhaps we should add a target hook so backends are
> > > free to request flushing to zero.
> > >
> > > Thoughts?
> >
> > I'm not sure what the intention of this is - it effectively results in
> > more conservative ranges for -funsafe-math-optimizations.  That is,
> > if -funsafe-math-optimizations says that denormals don't exist and
> > are all zero then doesn't that mean we can instead use the smalles
> > non-denormal number less than (or greater than) zero here?
> >
> > That said, the flushing you do is also "valid" for
> > -fno-unsafe-math-optimizations
> > in case we don't want to deal with denormals in range boundaries.
> >
> > It might also be a correctness requirement in case we don't know how
> > targets handle denormals (IIRC even OS defaults might matter here) so
> > we conservatively have to treat them as being flushed to zero.
>
> Actually, rth suggested we always flush to zero because we don't know
> what the target would do.  Again, I'm happy to do whatever you agree
> on.  I have no opinion.  My main goal here is correctness.

Yes, I think we should do this.

> >
> > So ... if we want to be on the "safe" side then please always do this.
> >
> > At the same point you could do
> >
> >  if (!HONOR_SIGNED_ZEROS ())
> >    if (real_iszero (&m_max))
> >      {
> >         if (real_iszero (&m_min))
> >            m.min.sign = 1;
> >         m_max.sign = 1;
> >      }
>
> But wouldn't that set [-0.0, -0.0] when encountering [+0, +0] ??

Yeah, that's my laziness not adding a special case for m_min == m_max.

>
> >    else if (real_iszero (&m_min))
> >      m_min.sign = 0;
>
> Jakub actually suggested something completely different...just set +0
> always for !HONOR_SIGNED_ZEROS.

Hmm, but the [-1, -0.] with known sign becomes [-1, +0.] with unknown sign?

Richard.

> Aldy
>
> >
> > so that we canonicalize a zero bound so that the sign is known for a range.
> >
> > Richard.
> >
> > > gcc/ChangeLog:
> > >
> > >         * value-range.cc (frange::flush_denormals_to_zero): New.
> > >         (frange::set): Call flush_denormals_to_zero.
> > >         * value-range.h (class frange): Add flush_denormals_to_zero.
> > > ---
> > >  gcc/value-range.cc | 24 ++++++++++++++++++++++++
> > >  gcc/value-range.h  |  1 +
> > >  2 files changed, 25 insertions(+)
> > >
> > > diff --git a/gcc/value-range.cc b/gcc/value-range.cc
> > > index 67d5d7fa90f..f285734f0e0 100644
> > > --- a/gcc/value-range.cc
> > > +++ b/gcc/value-range.cc
> > > @@ -267,6 +267,26 @@ tree_compare (tree_code code, tree op1, tree op2)
> > >    return !integer_zerop (fold_build2 (code, integer_type_node, op1, op2));
> > >  }
> > >
> > > +// Flush denormal endpoints to the appropriate 0.0.
> > > +
> > > +void
> > > +frange::flush_denormals_to_zero ()
> > > +{
> > > +  if (undefined_p () || known_isnan ())
> > > +    return;
> > > +
> > > +  // Flush [x, -DENORMAL] to [x, -0.0].
> > > +  if (real_isdenormal (&m_max) && real_isneg (&m_max))
> > > +    {
> > > +      m_max = dconst0;
> > > +      if (HONOR_SIGNED_ZEROS (m_type))
> > > +       m_max.sign = 1;
> > > +    }
> > > +  // Flush [+DENORMAL, x] to [+0.0, x].
> > > +  if (real_isdenormal (&m_min) && !real_isneg (&m_min))
> > > +    m_min = dconst0;
> > > +}
> > > +
> > >  // Setter for franges.
> > >
> > >  void
> > > @@ -317,6 +337,10 @@ frange::set (tree min, tree max, value_range_kind kind)
> > >    gcc_checking_assert (tree_compare (LE_EXPR, min, max));
> > >
> > >    normalize_kind ();
> > > +
> > > +  if (flag_unsafe_math_optimizations)
> > > +    flush_denormals_to_zero ();
> > > +
> > >    if (flag_checking)
> > >      verify_range ();
> > >  }
> > > diff --git a/gcc/value-range.h b/gcc/value-range.h
> > > index 3a401f3e4e2..795b1f00fdc 100644
> > > --- a/gcc/value-range.h
> > > +++ b/gcc/value-range.h
> > > @@ -327,6 +327,7 @@ private:
> > >    bool union_nans (const frange &);
> > >    bool intersect_nans (const frange &);
> > >    bool combine_zeros (const frange &, bool union_p);
> > > +  void flush_denormals_to_zero ();
> > >
> > >    tree m_type;
> > >    REAL_VALUE_TYPE m_min;
> > > --
> > > 2.37.1
> > >
> >
>
  
Aldy Hernandez Sept. 20, 2022, 5:22 a.m. UTC | #6
On Mon, Sep 19, 2022 at 3:45 PM Richard Biener
<richard.guenther@gmail.com> wrote:
>
> On Mon, Sep 19, 2022 at 3:04 PM Aldy Hernandez <aldyh@redhat.com> wrote:
> >
> > On Mon, Sep 19, 2022 at 9:37 AM Richard Biener
> > <richard.guenther@gmail.com> wrote:
> > >
> > > On Sat, Sep 17, 2022 at 10:25 AM Aldy Hernandez via Gcc-patches
> > > <gcc-patches@gcc.gnu.org> wrote:
> > > >
> > > > Jakub has mentioned that for -funsafe-math-optimizations we may flush
> > > > denormals to zero, in which case we need to be careful to extend the
> > > > ranges to the appropriate zero.  This patch does exactly that.  For a
> > > > range of [x, -DENORMAL] we flush to [x, -0.0] and for [+DENORMAL, x]
> > > > we flush to [+0.0, x].
> > > >
> > > > It is unclear whether we should do this for Alpha, since I believe
> > > > flushing to zero is the default, and the port requires -mieee for IEEE
> > > > sanity.  If so, perhaps we should add a target hook so backends are
> > > > free to request flushing to zero.
> > > >
> > > > Thoughts?
> > >
> > > I'm not sure what the intention of this is - it effectively results in
> > > more conservative ranges for -funsafe-math-optimizations.  That is,
> > > if -funsafe-math-optimizations says that denormals don't exist and
> > > are all zero then doesn't that mean we can instead use the smalles
> > > non-denormal number less than (or greater than) zero here?
> > >
> > > That said, the flushing you do is also "valid" for
> > > -fno-unsafe-math-optimizations
> > > in case we don't want to deal with denormals in range boundaries.
> > >
> > > It might also be a correctness requirement in case we don't know how
> > > targets handle denormals (IIRC even OS defaults might matter here) so
> > > we conservatively have to treat them as being flushed to zero.
> >
> > Actually, rth suggested we always flush to zero because we don't know
> > what the target would do.  Again, I'm happy to do whatever you agree
> > on.  I have no opinion.  My main goal here is correctness.
>
> Yes, I think we should do this.

Ok, I'm pushing the attached patch because it's conservatively correct
everywhere.  We can tweak it when y'all reach consensus.

>
> > >
> > > So ... if we want to be on the "safe" side then please always do this.
> > >
> > > At the same point you could do
> > >
> > >  if (!HONOR_SIGNED_ZEROS ())
> > >    if (real_iszero (&m_max))
> > >      {
> > >         if (real_iszero (&m_min))
> > >            m.min.sign = 1;
> > >         m_max.sign = 1;
> > >      }
> >
> > But wouldn't that set [-0.0, -0.0] when encountering [+0, +0] ??
>
> Yeah, that's my laziness not adding a special case for m_min == m_max.
>
> >
> > >    else if (real_iszero (&m_min))
> > >      m_min.sign = 0;
> >
> > Jakub actually suggested something completely different...just set +0
> > always for !HONOR_SIGNED_ZEROS.
>
> Hmm, but the [-1, -0.] with known sign becomes [-1, +0.] with unknown sign?

Huh.  I guess that's true.  Does that happen often enough in practice
to matter?  I suppose we're going into uncharted territory with folks
asking for no signs on zeros, but them appearing in the source code?
My gut feeling is that we should take whatever signs are in the source
code (similar to what we're doing for NANs), but don't introduce any
additional ones.

Again, no strong opinions.  Feel free to add whatever adjustment you
think is necessary.

Aldy
  
Jakub Jelinek Sept. 20, 2022, 8:26 a.m. UTC | #7
On Tue, Sep 20, 2022 at 07:22:03AM +0200, Aldy Hernandez wrote:
> > > Jakub actually suggested something completely different...just set +0
> > > always for !HONOR_SIGNED_ZEROS.
> >
> > Hmm, but the [-1, -0.] with known sign becomes [-1, +0.] with unknown sign?
> 
> Huh.  I guess that's true.  Does that happen often enough in practice

Sure, if you -fno-signed-zeros/-ffast-math and some variable can be zero,
copysign/signbit is undefined.  The option basically asserts you don't care
about it...

	Jakub
  

Patch

diff --git a/gcc/value-range.cc b/gcc/value-range.cc
index 67d5d7fa90f..f285734f0e0 100644
--- a/gcc/value-range.cc
+++ b/gcc/value-range.cc
@@ -267,6 +267,26 @@  tree_compare (tree_code code, tree op1, tree op2)
   return !integer_zerop (fold_build2 (code, integer_type_node, op1, op2));
 }
 
+// Flush denormal endpoints to the appropriate 0.0.
+
+void
+frange::flush_denormals_to_zero ()
+{
+  if (undefined_p () || known_isnan ())
+    return;
+
+  // Flush [x, -DENORMAL] to [x, -0.0].
+  if (real_isdenormal (&m_max) && real_isneg (&m_max))
+    {
+      m_max = dconst0;
+      if (HONOR_SIGNED_ZEROS (m_type))
+	m_max.sign = 1;
+    }
+  // Flush [+DENORMAL, x] to [+0.0, x].
+  if (real_isdenormal (&m_min) && !real_isneg (&m_min))
+    m_min = dconst0;
+}
+
 // Setter for franges.
 
 void
@@ -317,6 +337,10 @@  frange::set (tree min, tree max, value_range_kind kind)
   gcc_checking_assert (tree_compare (LE_EXPR, min, max));
 
   normalize_kind ();
+
+  if (flag_unsafe_math_optimizations)
+    flush_denormals_to_zero ();
+
   if (flag_checking)
     verify_range ();
 }
diff --git a/gcc/value-range.h b/gcc/value-range.h
index 3a401f3e4e2..795b1f00fdc 100644
--- a/gcc/value-range.h
+++ b/gcc/value-range.h
@@ -327,6 +327,7 @@  private:
   bool union_nans (const frange &);
   bool intersect_nans (const frange &);
   bool combine_zeros (const frange &, bool union_p);
+  void flush_denormals_to_zero ();
 
   tree m_type;
   REAL_VALUE_TYPE m_min;