Re: Fix Fortran rounding issues, PR fortran/96983.

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

On Mon, Apr 19, 2021 at 9:40 PM Michael Meissner via Fortran
<fortran@gcc.gnu.org> wrote:
>
> Fix Fortran rounding issues, PR fortran/96983.
>
> I was looking at Fortran PR 96983, which fails on the PowerPC when trying to
> run the test PR96711.F90.  The compiler ICEs because the PowerPC does not have
> a floating point type with a type precision of 128.  The reason is that the
> PowerPC has 3 different 128 bit floating point types (__float128/_Float128,
> __ibm128, and long double).  Currently long double uses the IBM extended double
> type, but we would like to switch to using IEEE 128-bit long doubles in the
> future.
>
> In order to prevent the compiler from converting explicit __ibm128 types to
> long double when long double uses the IEEE 128-bit representation, we have set
> up the precision for __ibm128 to be 128, long double to be 127, and
> __float128/_Float128 to be 126.
>
> Originally, I was trying to see if for Fortran, I could change the precision of
> long double to be 128 (Fortran doesn't access __ibm128), but it quickly became
> hard to get the changes to work.
>
> I looked at the Fortran code in build_round_expr, and I came to the conclusion
> that there is no reason to promote the floating point type.  If you just do a
> normal round of the value using the current floating point format and then
> convert it to the integer type.  We don't have an appropriate built-in function
> that provides the equivalent of llround for 128-bit integer types.
>
> This patch fixes the compiler crash.
>
> However, while with this patch, the PowerPC compiler will not crash when
> building the test case, it will not run on the current default installation.
> The failure is because the test is explicitly expecting 128-bit floating point
> to handle 10384593717069655257060992658440192_16 (i.e. 2**113).
>
> By default, the PowerPC uses IBM extended double used for 128-bit floating
> point.  The IBM extended double format is a pair of doubles that provides more
> mantissa bits but does not grow the expoenent range.  The value in the test is
> fine for IEEE 128-bit floating point, but it is too large for the PowerPC
> extended double setup.
>
> I have built the following tests with this patch:
>
>    * I have built a bootstrap compiler on a little endian power9 Linux system
>      with the default long double format (IBM extended double).  The
>      pr96711.f90 test builds, but it does not run due to the range of the
>      real*16 exponent.  There were no other regressions in the C/C++/Fortran
>      tests.
>
>    * I have built a bootstrap compiler on a little endian power9 Linux system
>      with the default long double format set to IEEE 128-bit. I used the
>      Advance Toolchain 14.0-2 to provide the IEEE 128-bits.  The compiler was
>      configured to build power9 code by default, so the test generated native
>      power9 IEEE 128-bit instructions.  The pr96711.f90 test builds and runs
>      correctly in this setup.
>
>    * I have built a bootstrap compiler on a big endian power8 Linux system with
>      the default long double format (IBM extended double).  Like the first
>      case, the pr96711.f90 test does not crash the compiler, but the test fails
>      due to the range of the real*16 exponent.    There were no other
>      regressions in the C/C++/Fortran tests.
>
>    * I built a bootstrap compiler on my x86_64 laptop.  There were no
>      regressions in the tests.
>
>
> Can I check this change into the GCC trunk?

The patch looks fine technically and is definitely an improvement since the
intermediate conversion looks odd.  But it might be that the standard
requires such dance though the preceeding cases handled don't seem to
care.  I'm thinking of a FP format where round(1.6) == 3 because of lack
of precision but using an intermediate FP format with higher precision
would "correctly" compute 2.

Of course the current code doesn't handle this correctly for the
case if llroundf either.

Richard.

> I've not contributed to the Fortran front end before.  If the maintainers like
> the patch, can somebody point out if I need to do additional things to commit
> the patch?
>
> gcc/fortran/
> 2021-04-19  Michael Meissner  <meissner@linux.ibm.com>
>
>         PR gfortran/96983
>         * trans-intrinsic.c (build_round_expr): If int type is larger than
>         long long, do the round and convert to the integer type.  Do not
>         try to find a floating point type the exact size of the integer
>         type.
> ---
>  gcc/fortran/trans-intrinsic.c | 26 ++++++++------------------
>  1 file changed, 8 insertions(+), 18 deletions(-)
>
> diff --git a/gcc/fortran/trans-intrinsic.c b/gcc/fortran/trans-intrinsic.c
> index 5e53d1162fa..cceef8f34ac 100644
> --- a/gcc/fortran/trans-intrinsic.c
> +++ b/gcc/fortran/trans-intrinsic.c
> @@ -386,30 +386,20 @@ build_round_expr (tree arg, tree restype)
>    argprec = TYPE_PRECISION (argtype);
>    resprec = TYPE_PRECISION (restype);
>
> -  /* Depending on the type of the result, choose the int intrinsic
> -     (iround, available only as a builtin, therefore cannot use it for
> -     __float128), long int intrinsic (lround family) or long long
> -     intrinsic (llround).  We might also need to convert the result
> -     afterwards.  */
> +  /* Depending on the type of the result, choose the int intrinsic (iround,
> +     available only as a builtin, therefore cannot use it for __float128), long
> +     int intrinsic (lround family) or long long intrinsic (llround).  If we
> +     don't have an appropriate function that converts directly to the integer
> +     type (such as kind == 16), just use ROUND, and then convert the result to
> +     an integer.  We might also need to convert the result afterwards.  */
>    if (resprec <= INT_TYPE_SIZE && argprec <= LONG_DOUBLE_TYPE_SIZE)
>      fn = builtin_decl_for_precision (BUILT_IN_IROUND, argprec);
>    else if (resprec <= LONG_TYPE_SIZE)
>      fn = builtin_decl_for_precision (BUILT_IN_LROUND, argprec);
>    else if (resprec <= LONG_LONG_TYPE_SIZE)
>      fn = builtin_decl_for_precision (BUILT_IN_LLROUND, argprec);
> -  else if (resprec >= argprec && resprec == 128)
> -    {
> -      /* Search for a real kind suitable as temporary for conversion.  */
> -      int kind = -1;
> -      for (int i = 0; kind < 0 && gfc_real_kinds[i].kind != 0; i++)
> -       if (gfc_real_kinds[i].mode_precision >= resprec)
> -         kind = gfc_real_kinds[i].kind;
> -      if (kind < 0)
> -       gfc_internal_error ("Could not find real kind with at least %d bits",
> -                           resprec);
> -      arg = fold_convert (gfc_get_real_type (kind), arg);
> -      fn = gfc_builtin_decl_for_float_kind (BUILT_IN_ROUND, kind);
> -    }
> +  else if (resprec >= argprec)
> +    fn = builtin_decl_for_precision (BUILT_IN_ROUND, argprec);
>    else
>      gcc_unreachable ();
>
> --
> 2.22.0
>
>
> --
> Michael Meissner, IBM
> IBM, M/S 2506R, 550 King Street, Littleton, MA 01460-6245, USA
> email: meissner@linux.ibm.com, phone: +1 (978) 899-4797