[BZ,#16824] Fix failing y1 due to too large ulps in downward/upward rounding mode.
diff mbox

Message ID lig7bv$9gj$1@ger.gmane.org
State New
Headers show

Commit Message

Stefan Liebler April 14, 2014, 8:47 a.m. UTC
Thanks for the hint with SET_RESTORE_ROUND macro.
Here is the updated version.
Tested on s390/s390x.

Ok for commit?

Bye

---
2014-04-14  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.
---

On 04/11/2014 02:50 PM, Adhemerval Zanella wrote:
> 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.
>
>

Comments

Adhemerval Zanella April 14, 2014, 5:42 p.m. UTC | #1
On 14-04-2014 05:47, Stefan Liebler wrote:
> Thanks for the hint with SET_RESTORE_ROUND macro.
> Here is the updated version.
> Tested on s390/s390x.
>
> Ok for commit?
>
> Bye
>
LGTM.
Andreas Krebbel April 16, 2014, 11:06 a.m. UTC | #2
On 04/14/2014 10:47 AM, Stefan Liebler wrote:
>>> 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.

Applied. Thanks!

-Andreas-
Andreas Schwab April 17, 2014, 2:30 p.m. UTC | #3
Stefan Liebler <stli@linux.vnet.ibm.com> writes:

> diff --git a/sysdeps/ieee754/ldbl-128/e_j1l.c b/sysdeps/ieee754/ldbl-128/e_j1l.c
> index 70a1c86..1264c95 100644
> --- a/sysdeps/ieee754/ldbl-128/e_j1l.c
> +++ b/sysdeps/ieee754/ldbl-128/e_j1l.c
> @@ -856,6 +856,7 @@ __ieee754_y1l (long double x)
>    if (xx <= 2.0L)
>      {
>        /* 0 <= x <= 2 */
> +      SET_RESTORE_ROUNDL (FE_TONEAREST);

In file included from ../sysdeps/ieee754/flt-32/math_private.h:3:0,
                 from ../sysdeps/powerpc/fpu/math_private.h:26,
                 from ../sysdeps/ieee754/ldbl-128/e_j1l.c:99,
                 from ../sysdeps/ieee754/ldbl-128ibm/e_j1l.c:2:
../sysdeps/ieee754/ldbl-128/e_j1l.c: In function '__ieee754_y1l':
../sysdeps/generic/math_private.h:579:10: error: cleanup argument not a function
   struct rm_ctx ctx __attribute__((cleanup(CLEANUPFUNC ## _ctx)));       \
          ^
../sysdeps/generic/math_private.h:594:3: note: in expansion of macro 'SET_RESTORE_ROUND_GENERIC'
   SET_RESTORE_ROUND_GENERIC (RM, libc_feholdsetroundl, libc_feresetroundl)
   ^
../sysdeps/ieee754/ldbl-128/e_j1l.c:859:7: note: in expansion of macro 'SET_RESTORE_ROUNDL'
       SET_RESTORE_ROUNDL (FE_TONEAREST);
       ^
../sysdeps/generic/math_private.h:579:10: warning: implicit declaration of function 'libc_feholdsetroundl_ctx' [-Wimplicit-function-declaration]
   struct rm_ctx ctx __attribute__((cleanup(CLEANUPFUNC ## _ctx)));       \
          ^
../sysdeps/generic/math_private.h:594:3: note: in expansion of macro 'SET_RESTORE_ROUND_GENERIC'
   SET_RESTORE_ROUND_GENERIC (RM, libc_feholdsetroundl, libc_feresetroundl)
   ^
../sysdeps/ieee754/ldbl-128/e_j1l.c:859:7: note: in expansion of macro 'SET_RESTORE_ROUNDL'
       SET_RESTORE_ROUNDL (FE_TONEAREST);
       ^
/home/abuild/rpmbuild/BUILD/glibc-2.19.90/cc-base/sysd-rules:2775: recipe for target '/home/abuild/rpmbuild/BUILD/glibc-2.19.90/cc-base/math/e_j1l.o' failed

Andreas.

Patch
diff mbox

diff --git a/sysdeps/ieee754/ldbl-128/e_j1l.c b/sysdeps/ieee754/ldbl-128/e_j1l.c
index 70a1c86..1264c95 100644
--- a/sysdeps/ieee754/ldbl-128/e_j1l.c
+++ b/sysdeps/ieee754/ldbl-128/e_j1l.c
@@ -856,6 +856,7 @@  __ieee754_y1l (long double x)
   if (xx <= 2.0L)
     {
       /* 0 <= x <= 2 */
+      SET_RESTORE_ROUNDL (FE_TONEAREST);
       z = xx * xx;
       p = xx * neval (z, Y0_2N, NY0_2N) / deval (z, Y0_2D, NY0_2D);
       p = -TWOOPI / xx + p;