about summary refs log tree commit diff
path: root/sysdeps/powerpc/fpu/e_sqrtf.c
blob: 3b2e243bb82002776ce35fce1dde5c727bf5a5b8 (plain) (blame)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
/* Single-precision floating point square root.
   Copyright (C) 1997-2013 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 <sysdep.h>
#include <ldsodefs.h>

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 = -(sg * sg - sx);
	  fsgi = (xi + 0x40000000) >> 1 & 0x7f800000;
	  sy2 = sy + sy;
	  sg = sy * sd + sg;	/* 16-bit approximation to sqrt(sx). */
	  e = -(sy * sg - almost_half);
	  SET_FLOAT_WORD (fsg, fsgi);
	  sd = -(sg * sg - sx);
	  sy = sy + e * sy2;
	  if ((xi & 0x7f800000) == 0)
	    goto denorm;
	  shx = sx * fsg;
	  sg = sg + sy * sd;	/* 32-bit approximation to sqrt(sx),
				   but perhaps rounded incorrectly.  */
	  sy2 = sy + sy;
	  g = sg * fsg;
	  e = -(sy * sg - almost_half);
	  d = -(g * sg - shx);
	  sy = sy + e * sy2;
	  fesetenv_register (fe);
	  return g + sy * d;
	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[1] & FE_INVALID) == 0)
#endif
	feraiseexcept (FE_INVALID);
      x = a_nan.value;
    }
  return f_washf (x);
}

#undef __ieee754_sqrtf
float
__ieee754_sqrtf (float x)
{
  double z;

  /* If the CPU is 64-bit we can use the optional FP instructions.  */
  if (__CPU_HAS_FSQRT)
    {
      /* Volatile is required to prevent the compiler from moving the
	 fsqrt instruction above the branch.  */
      __asm __volatile ("	fsqrts	%0,%1\n"
				:"=f" (z):"f" (x));
    }
  else
    z = __slow_ieee754_sqrtf (x);

  return z;
}
strong_alias (__ieee754_sqrtf, __sqrtf_finite)