[PATCH] BZ#13889: expl (709.75) wrongly overflows for ldbl-128ibm

[Andreas Jaeger]

On 03/19/2013 07:44 PM, Adhemerval Zanella wrote:
> Ping. I rebase against master (with updated NEWS) and add the float guard.
>
>
> On 06-03-2013 15:42, Adhemerval Zanella wrote:
>> On 03/06/2013 01:28 PM, Joseph S. Myers wrote:
>>> On Wed, 6 Mar 2013, Adhemerval Zanella wrote:
>>>
>>>> +  TEST_f_f (exp, 709.75L, 1.739836873264160557698252711673830393864768e+308L);
>>> Needs disabling for TEST_FLOAT; I'd expect you to get an OVERFLOW
>>> exception for float, causing test-float to fail, with the patch as-is (and
>>> if you *don't* get an OVERFLOW exception for expf (709.75), file a bug,
>>> because you should get one).
>>>
>> Yes you are right, I added TEST_FLOAT guards for the test now.
>>
> 2013-03-19  Adhemerval Zanella  <azanella@linux.vnet.ibm.com>
>
> 	[BZ #13889]
> 	* sysdeps/ieee754/ldbl-128ibm/e_expl.c (__ieee754_expl): Increase the high
> 	value to check if expl overflow.
> 	* sysdeps/ieee754/ldbl-128ibm/w_expl.c (__expl): Fix threshold constants
> 	to check for underflow and overflow.
> 	* math/libm-test.inc: Add exp test.
>
> diff --git a/NEWS b/NEWS
> index a02bf16..17159c5 100644
> --- a/NEWS
> +++ b/NEWS
> @@ -9,9 +9,9 @@ Version 2.18
>
>   * The following bugs are resolved with this release:
>
> -  11561, 12723, 13550, 13951, 14142, 14200, 14317, 14327, 14496, 14920,
> -  14964, 14981, 14982, 14985, 14994, 14996, 15003, 15006, 15020, 15023,
> -  15036, 15054, 15055, 15062, 15078, 15160, 15232, 15234, 15283.
> +  11561, 12723, 13889, 13550, 13951, 14142, 14200, 14317, 14327, 14496,
> +  14920, 14964, 14981, 14982, 14985, 14994, 14996, 15003, 15006, 15020,
> +  15023, 15036, 15054, 15055, 15062, 15078, 15160, 15232, 15234, 15283.
>
>   * Add support for calling C++11 thread_local object destructors on thread
>     and program exit.  This needs compiler support for offloading C++11
> diff --git a/math/libm-test.inc b/math/libm-test.inc
> index 914aab3..b617533 100644
> --- a/math/libm-test.inc
> +++ b/math/libm-test.inc
> @@ -4309,6 +4309,9 @@ exp_test (void)
>     TEST_f_f (exp, 0.75L, 2.11700001661267466854536981983709561L);
>     TEST_f_f (exp, 50.0L, 5184705528587072464087.45332293348538L);
>     TEST_f_f (exp, 88.72269439697265625L, 3.40233126623160774937554134772290447915e38L);
> +#ifndef TEST_FLOAT
> +  TEST_f_f (exp, 709.75L, 1.739836873264160557698252711673830393864768e+308L);
> +#endif
>   #if defined TEST_LDOUBLE && __LDBL_MAX_EXP__ > 1024
>     /* The result can only be represented in sane long double.  */
>     TEST_f_f (exp, 1000.0L, 0.197007111401704699388887935224332313e435L);
> diff --git a/sysdeps/ieee754/ldbl-128ibm/e_expl.c b/sysdeps/ieee754/ldbl-128ibm/e_expl.c
> index 8236390..9fd6198 100644
> --- a/sysdeps/ieee754/ldbl-128ibm/e_expl.c
> +++ b/sysdeps/ieee754/ldbl-128ibm/e_expl.c
> @@ -70,11 +70,11 @@
>   static const long double C[] = {
>   /* Smallest integer x for which e^x overflows.  */
>   #define himark C[0]
> - 709.08956571282405153382846025171462914L,
> + 709.78271289338399678773454114191496482L,
>
>   /* Largest integer x for which e^x underflows.  */
>   #define lomark C[1]
> --744.44007192138121808966388925909996033L,
> +-744.44007192138126231410729844608163411L,
>
>   /* 3x2^96 */
>   #define THREEp96 C[2]
> diff --git a/sysdeps/ieee754/ldbl-128ibm/w_expl.c b/sysdeps/ieee754/ldbl-128ibm/w_expl.c
> index a5e72b2..5531636 100644
> --- a/sysdeps/ieee754/ldbl-128ibm/w_expl.c
> +++ b/sysdeps/ieee754/ldbl-128ibm/w_expl.c
> @@ -1,6 +1,30 @@
> -/* Looks like we can use ieee854 w_expl.c as is for IBM extended format. */
> +#include <math.h>
> +#include <math_private.h>
>   #include <math_ldbl_opt.h>
> -#undef weak_alias
> -#define weak_alias(n,a)
> -#include <sysdeps/ieee754/ldbl-128/w_expl.c>
> +
> +static const long double
> +o_threshold = 709.78271289338399678773454114191496482L,
> +u_threshold = -744.44007192138126231410729844608163411L;
> +
> +long double __expl(long double x)	/* wrapper exp */

Could you fix the formatting issues, please? Add spaces before the 
opening braces in all cases.

> +{
> +#ifdef _IEEE_LIBM

I know you copy & pasted this code but I think that we can remove this 
check for _IEEE_LIBM completely. Care for a followup patch to cleanup 
all usages?

> +  return __ieee754_expl(x);
> +#else
> +  long double z;
> +  z = __ieee754_expl(x);
> +  if (_LIB_VERSION == _IEEE_)
> +    return z;
> +  if (__finitel(x))
> +    {
> +      if (x >= o_threshold)
> +	return __kernel_standard_l(x,x,206); /* exp overflow */
> +      else if (x <= u_threshold)
> +	return __kernel_standard_l(x,x,207); /* exp underflow */
> +    }
> +  return z;
> +#endif
> +}
> +hidden_def (__expl)
>   long_double_symbol (libm, __expl, expl);
> +

git push will not allow an empty line at the end of the file, better fix 
it directly ;)

Ok with the proposed changes,

Andreas
-- 
  Andreas Jaeger aj@{suse.com,opensuse.org} Twitter/Identica: jaegerandi
   SUSE LINUX Products GmbH, Maxfeldstr. 5, 90409 Nürnberg, Germany
    GF: Jeff Hawn,Jennifer Guild,Felix Imendörffer,HRB16746 (AG Nürnberg)
     GPG fingerprint = 93A3 365E CE47 B889 DF7F  FED1 389A 563C C272 A126