Re: Fix e_logl (128ibm) spurious underflow

[Adhemerval Zanella <azanella at linux dot vnet dot ibm dot com>]

Ping.

On 02/28/2013 02:02 AM, Adhemerval Zanella wrote:
> On 02/27/2013 08:50 PM, Joseph S. Myers wrote:
>> On Wed, 27 Feb 2013, Adhemerval Zanella wrote:
>>
>>> 2013-02-27  Adhemerval Zanella  <azanella@linux.vnet.ibm.com>
>>>
>>> 	* sysdeps/ieee754/ldbl-128ibm/e_logl.c (__ieee754_logl): Fix spurious
>>> 	underflow.
>>> 	* sysdeps/powerpc/fpu/libm-test-ulps: Update. 
>> What is the bug report in Bugzilla for this issue, which presumably is 
>> user-visible in a release so should have a bug report?  If there isn't 
>> one, please file one; in any case, include the [BZ #N] annotation in the 
>> ChangeLog entry (and a commit of a fix should include updating NEWS).  
>> (If the issue isn't user-visible in a release and so doesn't need a bug 
>> report, please explain why not.  But any bug being fixed that was 
>> user-visible in a release should have a corresponding Bugzilla entry.)
>>
>> In my view you should add a testcase for this issue for log to 
>> libm-test.inc, even if there are tests for other functions that show the 
>> problem as well.
>>
> I create BZ #15214 to track this issue. I didn't add the specific testcase yet
> because it is showing some higher ULPs that I'd like to take a look (this is 
> showing without my patch also).
>
> --
>
> 2013-02-28  Adhemerval Zanella  <azanella@linux.vnet.ibm.com>
>
> 	[BZ #15214]
> 	* NEWS: Mention BZ #15214.
> 	* sysdeps/ieee754/ldbl-128ibm/e_logl.c (__ieee754_logl): Fix spurious
> 	underflow.
> 	* sysdeps/powerpc/fpu/libm-test-ulps: Update. 
>
> diff --git a/NEWS b/NEWS
> index a4c5638..945aa9d 100644
> --- a/NEWS
> +++ b/NEWS
> @@ -11,7 +11,7 @@ Version 2.18
>
>    11561, 13951, 14142, 14200, 14317, 14327, 14496, 14920, 14964, 14981,
>    14982, 14985, 14994, 14996, 15003, 15006, 15020, 15023, 15036, 15054,
> -  15062, 15078, 15160.
> +  15062, 15078, 15160, 15214.
>
>  * 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/sysdeps/ieee754/ldbl-128ibm/e_logl.c b/sysdeps/ieee754/ldbl-128ibm/e_logl.c
> index 14f47eb..15b5edf 100644
> --- a/sysdeps/ieee754/ldbl-128ibm/e_logl.c
> +++ b/sysdeps/ieee754/ldbl-128ibm/e_logl.c
> @@ -182,6 +182,9 @@ static const long double
>    ln2a = 6.93145751953125e-1L,
>    ln2b = 1.4286068203094172321214581765680755001344E-6L;
>
> +static const long double
> +  ldbl_epsilon = 0x1p-106L;
> +
>  long double
>  __ieee754_logl(long double x)
>  {
> @@ -258,7 +261,12 @@ __ieee754_logl(long double x)
>      }
>    /* Series expansion of log(1+z).  */
>    w = z * z;
> -  y = ((((((((((((l15 * z
> +  /* Avoid spurious underflows.  */
> +  if (__glibc_unlikely(w <= ldbl_epsilon))
> +    y = 0.0L;
> +  else
> +    {
> +      y = ((((((((((((l15 * z
>  		  + l14) * z
>  		 + l13) * z
>  		+ l12) * z
> @@ -271,7 +279,8 @@ __ieee754_logl(long double x)
>  	 + l5) * z
>  	+ l4) * z
>         + l3) * z * w;
> -  y -= 0.5 * w;
> +      y -= 0.5 * w;
> +    }
>    y += e * ln2b;  /* Base 2 exponent offset times ln(2).  */
>    y += z;
>    y += logtbl[k-26]; /* log(t) - (t-1) */
> diff --git a/sysdeps/powerpc/fpu/libm-test-ulps b/sysdeps/powerpc/fpu/libm-test-ulps
> index 4221967..5072190 100644
> --- a/sysdeps/powerpc/fpu/libm-test-ulps
> +++ b/sysdeps/powerpc/fpu/libm-test-ulps
> @@ -243,9 +243,11 @@ idouble: 1
>  Test "Real part of: cacos (-0.5 + +0 i) == 2.094395102393195492308428922186335256131 - 0 i":
>  double: 1
>  idouble: 1
> +ldouble: 1
>  Test "Real part of: cacos (-0.5 - 0 i) == 2.094395102393195492308428922186335256131 + +0 i":
>  double: 1
>  idouble: 1
> +ldouble: 1
>  Test "Imaginary part of: cacos (-1.5 + +0 i) == pi - 0.9624236501192068949955178268487368462704 i":
>  double: 1
>  float: 1
> @@ -265,6 +267,10 @@ double: 1
>  float: 1
>  idouble: 1
>  ifloat: 1
> +Test "Imaginary part of: cacos (0x1.fp1023 + 0x1.fp1023 i) == 7.853981633974483096156608458198757210493e-1 - 7.107906849659093345062145442726115449315e2 i":
> +ldouble: 1
> +Test "Imaginary part of: cacos (0x1.fp127 + 0x1.fp127 i) == 7.853981633974483096156608458198757210493e-1 - 8.973081118419833726837456344608533993585e1 i":
> +ldouble: 1
>
>  # cacosh
>  Test "Real part of: cacosh (+0 + 0.5 i) == 0.4812118250596034474977589134243684231352 + pi/2 i":