Re: [PATCH 1/2] [Gimple] Simplify (trunc)fmax/fmin((extend)a, (extend)b) to MAX/MIN(a,b)

[Richard Biener <richard.guenther@gmail.com>]

On Mon, Nov 8, 2021 at 2:30 AM Hongtao Liu <crazylht@gmail.com> wrote:
>
> On Fri, Nov 5, 2021 at 5:52 PM Richard Biener
> <richard.guenther@gmail.com> wrote:
> >
> > On Fri, Nov 5, 2021 at 6:38 AM liuhongt <hongtao.liu@intel.com> wrote:
> > >
> > > a and b are same type as trunc type and has less precision than
> > > extend type, the transformation is guarded by flag_finite_math_only.
> > >
> > > Bootstrapped and regtested under x86_64-pc-linux-gnu{-m32,}
> > > Ok for trunk?
> > >
> > > gcc/ChangeLog:
> > >
> > >         PR target/102464
> > >         * match.pd: Simplify (trunc)fmax/fmin((extend)a, (extend)b) to
> > >         MAX/MIN(a,b)
> > >
> > > gcc/testsuite/ChangeLog:
> > >
> > >         * gcc.target/i386/pr102464-maxmin.c: New test.
> > > ---
> > >  gcc/match.pd                                  | 14 ++++++
> > >  .../gcc.target/i386/pr102464-maxmin.c         | 44 +++++++++++++++++++
> > >  2 files changed, 58 insertions(+)
> > >  create mode 100644 gcc/testsuite/gcc.target/i386/pr102464-maxmin.c
> > >
> > > diff --git a/gcc/match.pd b/gcc/match.pd
> > > index f63079023d0..857ce7f712a 100644
> > > --- a/gcc/match.pd
> > > +++ b/gcc/match.pd
> > > @@ -6182,6 +6182,20 @@ DEFINE_INT_AND_FLOAT_ROUND_FN (RINT)
> > >         && direct_internal_fn_supported_p (IFN_COPYSIGN,
> > >                                           type, OPTIMIZE_FOR_BOTH))
> > >      (IFN_COPYSIGN @0 @1))))
> > > +
> > > +(for maxmin (max min)
> > > + (simplify
> > > +  (convert (maxmin (convert@2 @0) (convert @1)))
> > > +   (if (flag_finite_math_only
> >
> > I suppose you are concerned about infinities, not about NaNs.
> > Please use !HONOR_INFINITIES (@2) then (in general testing
> > flag_* is frowned upon).  You may want to do the FLOAT_TYPE_P
> > tests first.
> I'm concerned about NANs since MAX/MIN_EXPR are different from IEEE
> minimum and maximum operations at NAN operations.

But you are already only handling non-IEEE MAX/MIN_EXPR where the
behavior with a NaN argument is unspecified?

> So i think i'd use MODE_HAS_NANS(@2)?
> >
> > > +       && optimize
> > > +       && FLOAT_TYPE_P (type)
> > > +       && FLOAT_TYPE_P (TREE_TYPE (@2))
> > > +       && types_match (type, TREE_TYPE (@0))
> > > +       && types_match (type, TREE_TYPE (@1))
> > > +       && TYPE_PRECISION (type) < TYPE_PRECISION (TREE_TYPE (@2))
> > > +       && optab_handler (maxmin == MAX_EXPR ? smax_optab : smin_optab,
> > > +                       TYPE_MODE (type)) != CODE_FOR_nothing)

And just noticing this now - since we're only changing the type a MAX/MIN_EXPR
operate on, we don't really need to do the optab check.  At RTL expansion
we'd eventually try a wider mode.

> > > +    (maxmin @0 @1))))
> > >  #endif
> > >
> > >  (for froms (XFLOORL XCEILL XROUNDL XRINTL)
> > > diff --git a/gcc/testsuite/gcc.target/i386/pr102464-maxmin.c b/gcc/testsuite/gcc.target/i386/pr102464-maxmin.c
> > > new file mode 100644
> > > index 00000000000..37867235a6c
> > > --- /dev/null
> > > +++ b/gcc/testsuite/gcc.target/i386/pr102464-maxmin.c
> > > @@ -0,0 +1,44 @@
> > > +/* PR target/102464.  */
> > > +/* { dg-do compile } */
> > > +/* { dg-options "-O2 -mavx512fp16 -mavx512vl -ffast-math -ftree-vectorize -mtune=generic -mfpmath=sse" } */
> > > +/* { dg-final { scan-assembler-times "vmaxph" 3 } }  */
> > > +/* { dg-final { scan-assembler-times "vminph" 3 } }  */
> > > +/* { dg-final { scan-assembler-times "vmaxsh" 3 } }  */
> > > +/* { dg-final { scan-assembler-times "vminsh" 3 } }  */
> > > +/* { dg-final { scan-assembler-times "vmaxps" 2 } }  */
> > > +/* { dg-final { scan-assembler-times "vminps" 2 } }  */
> > > +/* { dg-final { scan-assembler-times "vmaxss" 2 } }  */
> > > +/* { dg-final { scan-assembler-times "vminss" 2 } }  */
> > > +/* { dg-final { scan-assembler-times "vmaxpd" 1 } }  */
> > > +/* { dg-final { scan-assembler-times "vminpd" 1 } }  */
> > > +/* { dg-final { scan-assembler-times "vmaxsd" 1 } }  */
> > > +/* { dg-final { scan-assembler-times "vminsd" 1 } }  */
> > > +
> > > +#include<math.h>
> > > +#define FOO(CODE,TYPE,SUFFIX)                                          \
> > > +  void                                                                 \
> > > +  foo_vect_##CODE##TYPE##SUFFIX (TYPE* __restrict a, TYPE* b, TYPE* c) \
> > > +  {                                                                    \
> > > +    for (int i = 0; i != 8; i++)                                       \
> > > +      a[i] = CODE##SUFFIX (b[i], c[i]);                                        \
> > > +  }                                                                    \
> > > +  TYPE                                                                 \
> > > +  foo_##CODE##TYPE##SUFFIX (TYPE b, TYPE c)                            \
> > > +  {                                                                    \
> > > +    return CODE##l (b, c);                                             \
> > > +  }
> > > +
> > > +FOO (fmax, _Float16, f);
> > > +FOO (fmax, _Float16,);
> > > +FOO (fmax, _Float16, l);
> > > +FOO (fmin, _Float16, f);
> > > +FOO (fmin, _Float16,);
> > > +FOO (fmin, _Float16, l);
> > > +
> > > +FOO (fmax, float,);
> > > +FOO (fmax, float, l);
> > > +FOO (fmin, float,);
> > > +FOO (fmin, float, l);
> > > +
> > > +FOO (fmax, double, l);
> > > +FOO (fmin, double, l);
> > > --
> > > 2.18.1
> > >
>
>
>
> --
> BR,
> Hongtao