1 /* Single-precision floating point square root.
2 Copyright (C) 1997-2018 Free Software Foundation, Inc.
3 This file is part of the GNU C Library.
5 The GNU C Library is free software; you can redistribute it and/or
6 modify it under the terms of the GNU Lesser General Public
7 License as published by the Free Software Foundation; either
8 version 2.1 of the License, or (at your option) any later version.
10 The GNU C Library is distributed in the hope that it will be useful,
11 but WITHOUT ANY WARRANTY; without even the implied warranty of
12 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 Lesser General Public License for more details.
15 You should have received a copy of the GNU Lesser General Public
16 License along with the GNU C Library; if not, see
17 <http://www.gnu.org/licenses/>. */
20 #include <math_private.h>
21 #include <fenv_private.h>
22 #include <fenv_libc.h>
29 static const float almost_half
= 0.50000006; /* 0.5 + 2^-24 */
30 static const ieee_float_shape_type a_nan
= {.word
= 0x7fc00000 };
31 static const ieee_float_shape_type a_inf
= {.word
= 0x7f800000 };
32 static const float two48
= 281474976710656.0;
33 static const float twom24
= 5.9604644775390625e-8;
34 extern const float __t_sqrt
[1024];
36 /* The method is based on a description in
37 Computation of elementary functions on the IBM RISC System/6000 processor,
38 P. W. Markstein, IBM J. Res. Develop, 34(1) 1990.
39 Basically, it consists of two interleaved Newton-Raphson approximations,
40 one to find the actual square root, and one to find its reciprocal
41 without the expense of a division operation. The tricky bit here
42 is the use of the POWER/PowerPC multiply-add operation to get the
43 required accuracy with high speed.
45 The argument reduction works by a combination of table lookup to
46 obtain the initial guesses, and some careful modification of the
47 generated guesses (which mostly runs on the integer unit, while the
48 Newton-Raphson is running on the FPU). */
51 __slow_ieee754_sqrtf (float x
)
53 const float inf
= a_inf
.value
;
59 /* Variables named starting with 's' exist in the
60 argument-reduced space, so that 2 > sx >= 0.5,
61 1.41... > sg >= 0.70.., 0.70.. >= sy > 0.35... .
62 Variables named ending with 'i' are integer versions of
63 floating-point values. */
64 float sx
; /* The value of which we're trying to find the square
66 float sg
, g
; /* Guess of the square root of x. */
67 float sd
, d
; /* Difference between the square of the guess and x. */
68 float sy
; /* Estimate of 1/2g (overestimated by 1ulp). */
70 float e
; /* Difference between y*g and 1/2 (note that e==se). */
71 float shx
; /* == sx * fsg */
72 float fsg
; /* sg*fsg == g. */
73 fenv_t fe
; /* Saved floating-point environment (stores rounding
74 mode and whether the inexact exception is
76 uint32_t xi
, sxi
, fsgi
;
79 GET_FLOAT_WORD (xi
, x
);
80 fe
= fegetenv_register ();
82 sxi
= (xi
& 0x3fffffff) | 0x3f000000;
83 SET_FLOAT_WORD (sx
, sxi
);
84 t_sqrt
= __t_sqrt
+ (xi
>> (23 - 8 - 1) & 0x3fe);
88 /* Here we have three Newton-Raphson iterations each of a
89 division and a square root and the remainder of the
90 argument reduction, all interleaved. */
91 sd
= -__builtin_fmaf (sg
, sg
, -sx
);
92 fsgi
= (xi
+ 0x40000000) >> 1 & 0x7f800000;
94 sg
= __builtin_fmaf (sy
, sd
, sg
); /* 16-bit approximation to
96 e
= -__builtin_fmaf (sy
, sg
, -almost_half
);
97 SET_FLOAT_WORD (fsg
, fsgi
);
98 sd
= -__builtin_fmaf (sg
, sg
, -sx
);
99 sy
= __builtin_fmaf (e
, sy2
, sy
);
100 if ((xi
& 0x7f800000) == 0)
103 sg
= __builtin_fmaf (sy
, sd
, sg
); /* 32-bit approximation to
104 sqrt(sx), but perhaps
105 rounded incorrectly. */
108 e
= -__builtin_fmaf (sy
, sg
, -almost_half
);
109 d
= -__builtin_fmaf (g
, sg
, -shx
);
110 sy
= __builtin_fmaf (e
, sy2
, sy
);
111 fesetenv_register (fe
);
112 return __builtin_fmaf (sy
, d
, g
);
114 /* For denormalised numbers, we normalise, calculate the
115 square root, and return an adjusted result. */
116 fesetenv_register (fe
);
117 return __slow_ieee754_sqrtf (x
* two48
) * twom24
;
122 /* For some reason, some PowerPC32 processors don't implement
124 #ifdef FE_INVALID_SQRT
125 feraiseexcept (FE_INVALID_SQRT
);
127 fenv_union_t u
= { .fenv
= fegetenv_register () };
128 if ((u
.l
& FE_INVALID
) == 0)
130 feraiseexcept (FE_INVALID
);
135 #endif /* _ARCH_PPCSQ */
137 #undef __ieee754_sqrtf
139 __ieee754_sqrtf (float x
)
144 asm ("fsqrts %0,%1\n" :"=f" (z
):"f" (x
));
146 z
= __slow_ieee754_sqrtf (x
);
151 strong_alias (__ieee754_sqrtf
, __sqrtf_finite
)