[BZ,#16823] Fix log1pl returning wrong infinity sign
Commit Message
Hi,
log1pl (-1) now returns with -inf instead of +inf in FE_DOWNWARD
rounding mode. The source code producing the -inf for the double/float
variants is the same, but on S390 the compiler generates a division by
constant +zero and ignores the (x-x).
Thus i also changed it to the constant zero in the double/float functions.
I´ve tested it on S390.
Please test it on other platforms because there are also reports for the
same failure on aarch64 (see bug comment 1)
and sparc (https://www.sourceware.org/ml/libc-alpha/2014-04/msg00410.html).
Give feedback in order to commit the patch.
Bye
---
2014-04-22 Stefan Liebler <stli@linux.vnet.ibm.com>
[BZ #16823]
* sysdeps/ieee754/ldbl-128/s_log1pl.c (__log1pl):
Always divide by positive zero when computing -Inf result.
* sysdeps/ieee754/dbl-64/s_log1p.c (__log1p): Likewise.
* sysdeps/ieee754/flt-32/s_log1pf.c (__log1pf): Likewise.
---
On 04/09/2014 03:24 PM, Stefan Liebler wrote:
> Hi,
>
> on S390 test-double fails for log1pl(-1) with infinity has wrong sign in
> rounding mode FE_DOWNWARD. See Bug 16823.
> In this rounding mode, (x-x) = -0.
> In all other rounding modes, (x-x) = +0.
> A division with divisor -0 leads to +inf, while +0 results in -inf as
> expected. This patch changes the divisor to a const +0.
> Tested on s390/s390x.
>
> Is this okay?
>
> Bye
>
> ---
> 2014-04-09 Stefan Liebler <stli@linux.vnet.ibm.com>
>
> [BZ #16823]
> * sysdeps/ieee754/ldbl-128/s_log1pl.c (__log1pl):
> Use const positive zero instead of (x-x).
> ---
Comments
From: Stefan Liebler <stli@linux.vnet.ibm.com>
Date: Tue, 22 Apr 2014 10:24:52 +0200
> log1pl (-1) now returns with -inf instead of +inf in FE_DOWNWARD
> rounding mode. The source code producing the -inf for the double/float
> variants is the same, but on S390 the compiler generates a division by
> constant +zero and ignores the (x-x).
> Thus i also changed it to the constant zero in the double/float
> functions.
> I´ve tested it on S390.
> Please test it on other platforms because there are also reports for
> the same failure on aarch64 (see bug comment 1)
> and sparc
> (https://www.sourceware.org/ml/libc-alpha/2014-04/msg00410.html).
>
> Give feedback in order to commit the patch.
I can confirm that this fixes the testsuite failures on sparc.
On 22 April 2014 09:24, Stefan Liebler <stli@linux.vnet.ibm.com> wrote:
> Hi,
>
> log1pl (-1) now returns with -inf instead of +inf in FE_DOWNWARD rounding
> mode. The source code producing the -inf for the double/float variants is
> the same, but on S390 the compiler generates a division by constant +zero
> and ignores the (x-x).
> Thus i also changed it to the constant zero in the double/float functions.
> I´ve tested it on S390.
> Please test it on other platforms because there are also reports for the
> same failure on aarch64 (see bug comment 1)
> and sparc (https://www.sourceware.org/ml/libc-alpha/2014-04/msg00410.html).
>
> Give feedback in order to commit the patch.
This resolves the failure on aarch64....
Cheers
/Marcus
On 04/22/2014 10:24 AM, Stefan Liebler wrote:
> 2014-04-22 Stefan Liebler <stli@linux.vnet.ibm.com>
>
> [BZ #16823]
> * sysdeps/ieee754/ldbl-128/s_log1pl.c (__log1pl):
> Always divide by positive zero when computing -Inf result.
> * sysdeps/ieee754/dbl-64/s_log1p.c (__log1p): Likewise.
> * sysdeps/ieee754/flt-32/s_log1pf.c (__log1pf): Likewise.
Applied. Thanks!
-Andreas-
@@ -110,7 +110,7 @@ __log1p (double x)
if (__glibc_unlikely (ax >= 0x3ff00000)) /* x <= -1.0 */
{
if (x == -1.0)
- return -two54 / (x - x); /* log1p(-1)=+inf */
+ return -two54 / zero; /* log1p(-1)=-inf */
else
return (x - x) / (x - x); /* log1p(x<-1)=NaN */
}
@@ -42,7 +42,7 @@ __log1pf(float x)
k = 1;
if (hx < 0x3ed413d7) { /* x < 0.41422 */
if(ax>=0x3f800000) { /* x <= -1.0 */
- if(x==(float)-1.0) return -two25/(x-x); /* log1p(-1)=+inf */
+ if(x==(float)-1.0) return -two25/zero; /* log1p(-1)=-inf */
else return (x-x)/(x-x); /* log1p(x<-1)=NaN */
}
if(ax<0x31000000) { /* |x| < 2**-29 */
@@ -150,7 +150,7 @@ __log1pl (long double xm1)
if (x <= 0.0L)
{
if (x == 0.0L)
- return (-1.0L / (x - x));
+ return (-1.0L / zero); /* log1p(-1) = -inf */
else
return (zero / (x - x));
}