Re: [PATCH] [GIMPLE] Simplify (_Float16) ceil ((double) x) to .CEIL (x) when available.

[Hongtao Liu <crazylht@gmail.com>]

On Mon, Sep 27, 2021 at 8:53 PM Richard Biener
<richard.guenther@gmail.com> wrote:
>
> On Fri, Sep 24, 2021 at 1:26 PM liuhongt <hongtao.liu@intel.com> wrote:
> >
> > Hi:
> >   Related discussion in [1] and PR.
> >
> >   Bootstrapped and regtest on x86_64-linux-gnu{-m32,}.
> >   Ok for trunk?
> >
> > [1] https://gcc.gnu.org/pipermail/gcc-patches/2021-July/574330.html
> >
> > gcc/ChangeLog:
> >
> >         PR target/102464
> >         * config/i386/i386.c (ix86_optab_supported_p):
> >         Return true for HFmode.
> >         * match.pd: Simplify (_Float16) ceil ((double) x) to
> >         __builtin_ceilf16 (a) when a is _Float16 type and
> >         direct_internal_fn_supported_p.
> >
> > gcc/testsuite/ChangeLog:
> >
> >         * gcc.target/i386/pr102464.c: New test.
> > ---
> >  gcc/config/i386/i386.c                   | 20 +++++++-----
> >  gcc/match.pd                             | 28 +++++++++++++++++
> >  gcc/testsuite/gcc.target/i386/pr102464.c | 39 ++++++++++++++++++++++++
> >  3 files changed, 79 insertions(+), 8 deletions(-)
> >  create mode 100644 gcc/testsuite/gcc.target/i386/pr102464.c
> >
> > diff --git a/gcc/config/i386/i386.c b/gcc/config/i386/i386.c
> > index ba89e111d28..3767fe9806d 100644
> > --- a/gcc/config/i386/i386.c
> > +++ b/gcc/config/i386/i386.c
> > @@ -23582,20 +23582,24 @@ ix86_optab_supported_p (int op, machine_mode mode1, machine_mode,
> >        return opt_type == OPTIMIZE_FOR_SPEED;
> >
> >      case rint_optab:
> > -      if (SSE_FLOAT_MODE_P (mode1)
> > -         && TARGET_SSE_MATH
> > -         && !flag_trapping_math
> > -         && !TARGET_SSE4_1)
> > +      if (mode1 == HFmode)
> > +       return true;
> > +      else if (SSE_FLOAT_MODE_P (mode1)
> > +              && TARGET_SSE_MATH
> > +              && !flag_trapping_math
> > +              && !TARGET_SSE4_1)
> >         return opt_type == OPTIMIZE_FOR_SPEED;
> >        return true;
> >
> >      case floor_optab:
> >      case ceil_optab:
> >      case btrunc_optab:
> > -      if (SSE_FLOAT_MODE_P (mode1)
> > -         && TARGET_SSE_MATH
> > -         && !flag_trapping_math
> > -         && TARGET_SSE4_1)
> > +      if (mode1 == HFmode)
> > +       return true;
> > +      else if (SSE_FLOAT_MODE_P (mode1)
> > +              && TARGET_SSE_MATH
> > +              && !flag_trapping_math
> > +              && TARGET_SSE4_1)
> >         return true;
> >        return opt_type == OPTIMIZE_FOR_SPEED;
> >
> > diff --git a/gcc/match.pd b/gcc/match.pd
> > index a9791ceb74a..9ccec8b6ce3 100644
> > --- a/gcc/match.pd
> > +++ b/gcc/match.pd
> > @@ -6191,6 +6191,34 @@ DEFINE_INT_AND_FLOAT_ROUND_FN (RINT)
> >     (froms (convert float_value_p@0))
> >     (convert (tos @0)))))
> >
> > +#if GIMPLE
> > +(match float16_value_p
> > + @0
> > + (if (TYPE_MAIN_VARIANT (TREE_TYPE (@0)) == float16_type_node)))
> > +(for froms (BUILT_IN_TRUNCL BUILT_IN_TRUNC BUILT_IN_TRUNCF
> > +           BUILT_IN_FLOORL BUILT_IN_FLOOR BUILT_IN_FLOORF
> > +           BUILT_IN_CEILL BUILT_IN_CEIL BUILT_IN_CEILF
> > +           BUILT_IN_ROUNDEVENL BUILT_IN_ROUNDEVEN BUILT_IN_ROUNDEVENF
> > +           BUILT_IN_ROUNDL BUILT_IN_ROUND BUILT_IN_ROUNDF
> > +           BUILT_IN_NEARBYINTL BUILT_IN_NEARBYINT BUILT_IN_NEARBYINTF
> > +           BUILT_IN_RINTL BUILT_IN_RINT BUILT_IN_RINTF)
>
> we do have patterns that convert (truncl (convert floatval)) to
> (float)truncf (val),
> your's does (_Float16)trunc ((double) float16) -> truncF16 (float16), doesn't it
> make sense to have trunc ((double) float16) -> (double)trunfF16
> (float16) as well?
>
> Why do you conditionalize on GIMPLE here?
To avoid
error: ‘direct_internal_fn_supported_p’ was not declared in this scope

>
> That said, I wonder whether we can somehow address pattern explosion here,
> eliding the outer (convert ...) from the match would help a bit already.
>
> The related patterns use optimize && canonicalize_math_p as well btw., not
> sure whether either is appropriate here since there are no _Float16 math
> functions available.
Yes, that's why I didn't follow the existing pattern, i think we can
add optimize back to the condition, but not canonicalize_math_p ()
since there's no math function for _Float16.
Also w/o the outer (convert ..), it looks like a canonicalization to
transform ceil ((double) a) to (double) __builtin_ceilf16 (a) but not
an optimization.
>
> > +     tos (IFN_TRUNC IFN_TRUNC IFN_TRUNC
> > +         IFN_FLOOR IFN_FLOOR IFN_FLOOR
> > +         IFN_CEIL IFN_CEIL IFN_CEIL
> > +         IFN_ROUNDEVEN IFN_ROUNDEVEN IFN_ROUNDEVEN
> > +         IFN_ROUND IFN_ROUND IFN_ROUND
> > +         IFN_NEARBYINT IFN_NEARBYINT IFN_NEARBYINT
> > +         IFN_RINT IFN_RINT IFN_RINT)
> > + /* (_Float16) round ((doube) x) -> __built_in_roundf16 (x), etc.,
> > +    if x is a _Float16.  */
> > + (simplify
> > +   (convert (froms (convert float16_value_p@0)))
> > +     (if (types_match (type, TREE_TYPE (@0))
> > +         && direct_internal_fn_supported_p (as_internal_fn (tos),
> > +                                            type, OPTIMIZE_FOR_BOTH))
> > +       (tos @0))))
> > +#endif
> > +
> >  (for froms (XFLOORL XCEILL XROUNDL XRINTL)
> >       tos (XFLOOR XCEIL XROUND XRINT)
> >   /* llfloorl(extend(x)) -> llfloor(x), etc., if x is a double.  */
> > diff --git a/gcc/testsuite/gcc.target/i386/pr102464.c b/gcc/testsuite/gcc.target/i386/pr102464.c
> > new file mode 100644
> > index 00000000000..e3e060ee80b
> > --- /dev/null
> > +++ b/gcc/testsuite/gcc.target/i386/pr102464.c
> > @@ -0,0 +1,39 @@
> > +/* PR target/102464.  */
> > +/* { dg-do compile } */
> > +/* { dg-options "-O2 -mavx512fp16" } */
> > +
> > +#define FOO(FUNC,SUFFIX)                       \
> > +  _Float16                                     \
> > +  foo_##FUNC##_##SUFFIX (_Float16 a)           \
> > +  {                                            \
> > +    return __builtin_##FUNC##SUFFIX (a);       \
> > +  }
> > +
> > +FOO (roundeven, f16);
> > +FOO (roundeven, f);
> > +FOO (roundeven, );
> > +FOO (roundeven, l);
> > +FOO (trunc, f16);
> > +FOO (trunc, f);
> > +FOO (trunc, );
> > +FOO (trunc, l);
> > +FOO (ceil, f16);
> > +FOO (ceil, f);
> > +FOO (ceil, );
> > +FOO (ceil, l);
> > +FOO (floor, f16);
> > +FOO (floor, f);
> > +FOO (floor, );
> > +FOO (floor, l);
> > +FOO (nearbyint, f16);
> > +FOO (nearbyint, f);
> > +FOO (nearbyint, );
> > +FOO (nearbyint, l);
> > +FOO (rint, f16);
> > +FOO (rint, f);
> > +FOO (rint, );
> > +FOO (rint, l);
> > +
> > +/* { dg-final { scan-assembler-not "vcvtsh2s\[sd\]" } } */
> > +/* { dg-final { scan-assembler-not "extendhfxf" } } */
> > +/* { dg-final { scan-assembler-times "vrndscalesh\[^\n\r\]*xmm\[0-9\]" 24 } } */
> > --
> > 2.27.0
> >



-- 
BR,
Hongtao