glibc/math/s_csqrtl.c
Joseph Myers 718d34a309 Fix csqrt spurious underflows (bug 18371).
The csqrt implementations in glibc can cause spurious underflows in
some cases as a side-effect of the scaling for large arguments (when
underflow is correct for the square root of the argument that was
scaled down to avoid overflow, but not for the original argument).
This patch arranges to avoid the underflowing intermediate computation
(eliminating a multiplication in 0.5 in the problem cases where a
subsequent scaling by 2 would follow).

Tested for x86_64 and x86 and ulps updated accordingly (only needed
for x86).

	[BZ #18371]
	* math/s_csqrt.c (__csqrt): Avoid multiplication by 0.5 where
	intermediate but not final result might underflow.
	* math/s_csqrtf.c (__csqrtf): Likewise.
	* math/s_csqrtl.c (__csqrtl): Likewise.
	* math/auto-libm-test-in: Add more tests of csqrt.
	* math/auto-libm-test-out: Regenerated.
	* sysdeps/i386/fpu/libm-test-ulps: Update.
2015-06-23 16:01:54 +00:00

159 lines
4.2 KiB
C

/* Complex square root of long double value.
Copyright (C) 1997-2015 Free Software Foundation, Inc.
This file is part of the GNU C Library.
Based on an algorithm by Stephen L. Moshier <moshier@world.std.com>.
Contributed by Ulrich Drepper <drepper@cygnus.com>, 1997.
The GNU C Library is free software; you can redistribute it and/or
modify it under the terms of the GNU Lesser General Public
License as published by the Free Software Foundation; either
version 2.1 of the License, or (at your option) any later version.
The GNU C Library is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
Lesser General Public License for more details.
You should have received a copy of the GNU Lesser General Public
License along with the GNU C Library; if not, see
<http://www.gnu.org/licenses/>. */
#include <complex.h>
#include <math.h>
#include <math_private.h>
#include <float.h>
__complex__ long double
__csqrtl (__complex__ long double x)
{
__complex__ long double res;
int rcls = fpclassify (__real__ x);
int icls = fpclassify (__imag__ x);
if (__glibc_unlikely (rcls <= FP_INFINITE || icls <= FP_INFINITE))
{
if (icls == FP_INFINITE)
{
__real__ res = HUGE_VALL;
__imag__ res = __imag__ x;
}
else if (rcls == FP_INFINITE)
{
if (__real__ x < 0.0)
{
__real__ res = icls == FP_NAN ? __nanl ("") : 0;
__imag__ res = __copysignl (HUGE_VALL, __imag__ x);
}
else
{
__real__ res = __real__ x;
__imag__ res = (icls == FP_NAN
? __nanl ("") : __copysignl (0.0, __imag__ x));
}
}
else
{
__real__ res = __nanl ("");
__imag__ res = __nanl ("");
}
}
else
{
if (__glibc_unlikely (icls == FP_ZERO))
{
if (__real__ x < 0.0)
{
__real__ res = 0.0;
__imag__ res = __copysignl (__ieee754_sqrtl (-__real__ x),
__imag__ x);
}
else
{
__real__ res = fabsl (__ieee754_sqrtl (__real__ x));
__imag__ res = __copysignl (0.0, __imag__ x);
}
}
else if (__glibc_unlikely (rcls == FP_ZERO))
{
long double r;
if (fabsl (__imag__ x) >= 2.0L * LDBL_MIN)
r = __ieee754_sqrtl (0.5L * fabsl (__imag__ x));
else
r = 0.5L * __ieee754_sqrtl (2.0L * fabsl (__imag__ x));
__real__ res = r;
__imag__ res = __copysignl (r, __imag__ x);
}
else
{
long double d, r, s;
int scale = 0;
if (fabsl (__real__ x) > LDBL_MAX / 4.0L)
{
scale = 1;
__real__ x = __scalbnl (__real__ x, -2 * scale);
__imag__ x = __scalbnl (__imag__ x, -2 * scale);
}
else if (fabsl (__imag__ x) > LDBL_MAX / 4.0L)
{
scale = 1;
if (fabsl (__real__ x) >= 4.0L * LDBL_MIN)
__real__ x = __scalbnl (__real__ x, -2 * scale);
else
__real__ x = 0.0L;
__imag__ x = __scalbnl (__imag__ x, -2 * scale);
}
else if (fabsl (__real__ x) < LDBL_MIN
&& fabsl (__imag__ x) < LDBL_MIN)
{
scale = -(LDBL_MANT_DIG / 2);
__real__ x = __scalbnl (__real__ x, -2 * scale);
__imag__ x = __scalbnl (__imag__ x, -2 * scale);
}
d = __ieee754_hypotl (__real__ x, __imag__ x);
/* Use the identity 2 Re res Im res = Im x
to avoid cancellation error in d +/- Re x. */
if (__real__ x > 0)
{
r = __ieee754_sqrtl (0.5L * (d + __real__ x));
if (scale == 1 && fabsl (__imag__ x) < 1.0L)
{
/* Avoid possible intermediate underflow. */
s = __imag__ x / r;
r = __scalbnl (r, scale);
scale = 0;
}
else
s = 0.5L * (__imag__ x / r);
}
else
{
s = __ieee754_sqrtl (0.5L * (d - __real__ x));
if (scale == 1 && fabsl (__imag__ x) < 1.0L)
{
/* Avoid possible intermediate underflow. */
r = fabsl (__imag__ x / s);
s = __scalbnl (s, scale);
scale = 0;
}
else
r = fabsl (0.5L * (__imag__ x / s));
}
if (scale)
{
r = __scalbnl (r, scale);
s = __scalbnl (s, scale);
}
__real__ res = r;
__imag__ res = __copysignl (s, __imag__ x);
}
}
return res;
}
weak_alias (__csqrtl, csqrtl)