glibc/sysdeps/powerpc/fpu/e_sqrtf.c
Joseph Myers 95c26233a1 Fix powerpc software sqrtf (bug 17967).
Similarly to sqrt in
<https://sourceware.org/ml/libc-alpha/2015-02/msg00353.html>, the
powerpc sqrtf implementation for when _ARCH_PPCSQ is not defined also
relies on a * b + c being contracted into a fused multiply-add.
Although this contraction is not explicitly disabled for e_sqrtf.c, it
still seems appropriate to make the file explicit about its
requirements by using __builtin_fmaf; this patch does so.
Furthermore, it turns out that doing so fixes the observed inaccuracy
and missing exceptions (that is, that without explicit __builtin_fmaf
usage, it was not being compiled as intended).

Tested for powerpc32 (hard float).

	[BZ #17967]
	* sysdeps/powerpc/fpu/e_sqrtf.c (__slow_ieee754_sqrtf): Use
	__builtin_fmaf instead of relying on contraction of a * b + c.
2015-02-13 16:20:36 +00:00

151 lines
5.0 KiB
C

/* Single-precision floating point square root.
Copyright (C) 1997-2015 Free Software Foundation, Inc.
This file is part of the GNU C Library.
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 <math.h>
#include <math_private.h>
#include <fenv_libc.h>
#include <inttypes.h>
#include <stdint.h>
#include <sysdep.h>
#include <ldsodefs.h>
#ifndef _ARCH_PPCSQ
static const float almost_half = 0.50000006; /* 0.5 + 2^-24 */
static const ieee_float_shape_type a_nan = {.word = 0x7fc00000 };
static const ieee_float_shape_type a_inf = {.word = 0x7f800000 };
static const float two48 = 281474976710656.0;
static const float twom24 = 5.9604644775390625e-8;
extern const float __t_sqrt[1024];
/* The method is based on a description in
Computation of elementary functions on the IBM RISC System/6000 processor,
P. W. Markstein, IBM J. Res. Develop, 34(1) 1990.
Basically, it consists of two interleaved Newton-Raphson approximations,
one to find the actual square root, and one to find its reciprocal
without the expense of a division operation. The tricky bit here
is the use of the POWER/PowerPC multiply-add operation to get the
required accuracy with high speed.
The argument reduction works by a combination of table lookup to
obtain the initial guesses, and some careful modification of the
generated guesses (which mostly runs on the integer unit, while the
Newton-Raphson is running on the FPU). */
float
__slow_ieee754_sqrtf (float x)
{
const float inf = a_inf.value;
if (x > 0)
{
if (x != inf)
{
/* Variables named starting with 's' exist in the
argument-reduced space, so that 2 > sx >= 0.5,
1.41... > sg >= 0.70.., 0.70.. >= sy > 0.35... .
Variables named ending with 'i' are integer versions of
floating-point values. */
float sx; /* The value of which we're trying to find the square
root. */
float sg, g; /* Guess of the square root of x. */
float sd, d; /* Difference between the square of the guess and x. */
float sy; /* Estimate of 1/2g (overestimated by 1ulp). */
float sy2; /* 2*sy */
float e; /* Difference between y*g and 1/2 (note that e==se). */
float shx; /* == sx * fsg */
float fsg; /* sg*fsg == g. */
fenv_t fe; /* Saved floating-point environment (stores rounding
mode and whether the inexact exception is
enabled). */
uint32_t xi, sxi, fsgi;
const float *t_sqrt;
GET_FLOAT_WORD (xi, x);
fe = fegetenv_register ();
relax_fenv_state ();
sxi = (xi & 0x3fffffff) | 0x3f000000;
SET_FLOAT_WORD (sx, sxi);
t_sqrt = __t_sqrt + (xi >> (23 - 8 - 1) & 0x3fe);
sg = t_sqrt[0];
sy = t_sqrt[1];
/* Here we have three Newton-Raphson iterations each of a
division and a square root and the remainder of the
argument reduction, all interleaved. */
sd = -__builtin_fmaf (sg, sg, -sx);
fsgi = (xi + 0x40000000) >> 1 & 0x7f800000;
sy2 = sy + sy;
sg = __builtin_fmaf (sy, sd, sg); /* 16-bit approximation to
sqrt(sx). */
e = -__builtin_fmaf (sy, sg, -almost_half);
SET_FLOAT_WORD (fsg, fsgi);
sd = -__builtin_fmaf (sg, sg, -sx);
sy = __builtin_fmaf (e, sy2, sy);
if ((xi & 0x7f800000) == 0)
goto denorm;
shx = sx * fsg;
sg = __builtin_fmaf (sy, sd, sg); /* 32-bit approximation to
sqrt(sx), but perhaps
rounded incorrectly. */
sy2 = sy + sy;
g = sg * fsg;
e = -__builtin_fmaf (sy, sg, -almost_half);
d = -__builtin_fmaf (g, sg, -shx);
sy = __builtin_fmaf (e, sy2, sy);
fesetenv_register (fe);
return __builtin_fmaf (sy, d, g);
denorm:
/* For denormalised numbers, we normalise, calculate the
square root, and return an adjusted result. */
fesetenv_register (fe);
return __slow_ieee754_sqrtf (x * two48) * twom24;
}
}
else if (x < 0)
{
/* For some reason, some PowerPC32 processors don't implement
FE_INVALID_SQRT. */
#ifdef FE_INVALID_SQRT
feraiseexcept (FE_INVALID_SQRT);
fenv_union_t u = { .fenv = fegetenv_register () };
if ((u.l & FE_INVALID) == 0)
#endif
feraiseexcept (FE_INVALID);
x = a_nan.value;
}
return f_washf (x);
}
#endif /* _ARCH_PPCSQ */
#undef __ieee754_sqrtf
float
__ieee754_sqrtf (float x)
{
double z;
#ifdef _ARCH_PPCSQ
asm ("fsqrts %0,%1\n" :"=f" (z):"f" (x));
#else
z = __slow_ieee754_sqrtf (x);
#endif
return z;
}
strong_alias (__ieee754_sqrtf, __sqrtf_finite)