This is the mail archive of the libc-alpha@sourceware.org mailing list for the glibc project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: [PATCH][BZ #16824] Fix failing y1 due to too large ulps in downward/upward rounding mode.


On 11-04-2014 08:56, Stefan Liebler wrote:
> Hi,
>
> on s390 the y1 tests for long double fails due to too large ulps in downward / upward rounding mode (See Bug 16824).
> According to Joseph Myers patch comment
> (https://sourceware.org/ml/libc-alpha/2014-03/msg00620.html),
> round-to-nearest is used internally to reduce error accumulation.
> Tests on s390/s390x shows following ulps:
> y1: 2
> y1_downward: 4
> y1_towardzero: 2
> y1_upward: 5
>
> Is this okay? On other architectures too?
>
> Bye
> ---
> 2014-04-11  Stefan Liebler  <stli@linux.vnet.ibm.com>
>
>     [BZ #16824]
>     * sysdeps/ieee754/ldbl-128/e_j1l.c (__ieee754_y1l):
>     Set round-to-nearest internally to reduce error accumulation.
> ---
>
> diff --git a/sysdeps/ieee754/ldbl-128/e_j1l.c b/sysdeps/ieee754/ldbl-128/e_j1l.c
> index 70a1c86..db96c94 100644
> --- a/sysdeps/ieee754/ldbl-128/e_j1l.c
> +++ b/sysdeps/ieee754/ldbl-128/e_j1l.c
> @@ -855,11 +855,14 @@ __ieee754_y1l (long double x)
>      return -TWOOPI / x;
>    if (xx <= 2.0L)
>      {
> +      int save_round = fegetround ();
> +      libc_fesetround (FE_TONEAREST);
>        /* 0 <= x <= 2 */
>        z = xx * xx;
>        p = xx * neval (z, Y0_2N, NY0_2N) / deval (z, Y0_2D, NY0_2D);
>        p = -TWOOPI / xx + p;
>        p = TWOOPI * __ieee754_logl (x) * __ieee754_j1l (x) + p;
> +      libc_fesetround (save_round);
>        return p;
>      }

Although s390 does not implemented the optimization to set/restore rounding mode only
when needed (check the x86_64 and powerpc fenv_private.h), the way to set/reset 
rounding mode should using SET_RESTORE_ROUND.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]