about summary refs log tree commit diff
path: root/sysdeps/ieee754/ldbl-96
diff options
context:
space:
mode:
authorJoseph Myers <joseph@codesourcery.com>2015-02-27 17:48:37 +0000
committerJoseph Myers <joseph@codesourcery.com>2015-02-27 17:48:37 +0000
commit2ca725c594e0c186d928dc0823be7d8b5976112c (patch)
tree8f477bda16d5f94bb925301798e653d3064ee75d /sysdeps/ieee754/ldbl-96
parentaf96be34825586536ebcfbf5c675e795ddd3c8fa (diff)
downloadglibc-2ca725c594e0c186d928dc0823be7d8b5976112c.tar.gz
glibc-2ca725c594e0c186d928dc0823be7d8b5976112c.tar.xz
glibc-2ca725c594e0c186d928dc0823be7d8b5976112c.zip
Fix ldbl-96, ldbl-128ibm atanhl inaccuracy (bug 18046, bug 18047).
The threshold in ldbl-96 atanhl for when to return the argument,
0x1p-28, is a bit too big, and that in ldbl-128ibm atanhl is much too
big (the relevant condition being x^3/3 being < 0.5ulp of x),
resulting in errors a bit above the limits of those considered
acceptable in glibc in the ldbl-96 case, and in large errors in the
ldbl-128ibm case.  This patch changes those implementations to use
more appropriate thresholds and adds tests around the thresholds for
various formats.

Tested for x86_64, x86 and powerpc.  x86_64 and x86 ulps updated
accordingly.

	[BZ #18046]
	[BZ #18047]
	* sysdeps/ieee754/ldbl-128ibm/e_atanhl.c (__ieee754_atanhl): Use
	0x1p-56L as threshold for just returning the argument.
	* sysdeps/ieee754/ldbl-96/e_atanhl.c (__ieee754_atanhl): Use
	0x1p-32L as threshold for just returning the argument.
	* math/auto-libm-test-in: Add more tests of atanh.
	* math/auto-libm-test-out: Regenerated.
	* sysdeps/i386/fpu/libm-test-ulps: Update.
	* sysdeps/x86_64/fpu/libm-test-ulp: Likewise.
Diffstat (limited to 'sysdeps/ieee754/ldbl-96')
-rw-r--r--sysdeps/ieee754/ldbl-96/e_atanhl.c4
1 files changed, 2 insertions, 2 deletions
diff --git a/sysdeps/ieee754/ldbl-96/e_atanhl.c b/sysdeps/ieee754/ldbl-96/e_atanhl.c
index 4c29243695..305d50e309 100644
--- a/sysdeps/ieee754/ldbl-96/e_atanhl.c
+++ b/sysdeps/ieee754/ldbl-96/e_atanhl.c
@@ -52,9 +52,9 @@ __ieee754_atanhl(long double x)
 	    return (x-x)/(x-x);
 	if(ix==0x3fff)
 	    return x/zero;
-	if(ix<0x3fe3) {
+	if(ix<0x3fdf) {
 	    math_force_eval(huge+x);
-	    return x;	/* x<2**-28 */
+	    return x;	/* x<2**-32 */
 	}
 	SET_LDOUBLE_EXP(x,ix);
 	if(ix<0x3ffe) {		/* x < 0.5 */