• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /* origin: FreeBSD /usr/src/lib/msun/src/s_cbrtl.c */
2 /*-
3  * ====================================================
4  * Copyright (C) 1993 by Sun Microsystems, Inc. All rights reserved.
5  * Copyright (c) 2009-2011, Bruce D. Evans, Steven G. Kargl, David Schultz.
6  *
7  * Developed at SunPro, a Sun Microsystems, Inc. business.
8  * Permission to use, copy, modify, and distribute this
9  * software is freely granted, provided that this notice
10  * is preserved.
11  * ====================================================
12  *
13  * The argument reduction and testing for exceptional cases was
14  * written by Steven G. Kargl with input from Bruce D. Evans
15  * and David A. Schultz.
16  */
17 
18 #include "libm.h"
19 
20 #if LDBL_MANT_DIG == 53 && LDBL_MAX_EXP == 1024
cbrtl(long double x)21 long double cbrtl(long double x)
22 {
23 	return cbrt(x);
24 }
25 #elif (LDBL_MANT_DIG == 64 || LDBL_MANT_DIG == 113) && LDBL_MAX_EXP == 16384
26 static const unsigned B1 = 709958130; /* B1 = (127-127.0/3-0.03306235651)*2**23 */
27 
cbrtl(long double x)28 long double cbrtl(long double x)
29 {
30 	union ldshape u = {x}, v;
31 	union {float f; uint32_t i;} uft;
32 	long double r, s, t, w;
33 	double_t dr, dt, dx;
34 	float_t ft;
35 	int e = u.i.se & 0x7fff;
36 	int sign = u.i.se & 0x8000;
37 
38 	/*
39 	 * If x = +-Inf, then cbrt(x) = +-Inf.
40 	 * If x = NaN, then cbrt(x) = NaN.
41 	 */
42 	if (e == 0x7fff)
43 		return x + x;
44 	if (e == 0) {
45 		/* Adjust subnormal numbers. */
46 		u.f *= 0x1p120;
47 		e = u.i.se & 0x7fff;
48 		/* If x = +-0, then cbrt(x) = +-0. */
49 		if (e == 0)
50 			return x;
51 		e -= 120;
52 	}
53 	e -= 0x3fff;
54 	u.i.se = 0x3fff;
55 	x = u.f;
56 	switch (e % 3) {
57 	case 1:
58 	case -2:
59 		x *= 2;
60 		e--;
61 		break;
62 	case 2:
63 	case -1:
64 		x *= 4;
65 		e -= 2;
66 		break;
67 	}
68 	v.f = 1.0;
69 	v.i.se = sign | (0x3fff + e/3);
70 
71 	/*
72 	 * The following is the guts of s_cbrtf, with the handling of
73 	 * special values removed and extra care for accuracy not taken,
74 	 * but with most of the extra accuracy not discarded.
75 	 */
76 
77 	/* ~5-bit estimate: */
78 	uft.f = x;
79 	uft.i = (uft.i & 0x7fffffff)/3 + B1;
80 	ft = uft.f;
81 
82 	/* ~16-bit estimate: */
83 	dx = x;
84 	dt = ft;
85 	dr = dt * dt * dt;
86 	dt = dt * (dx + dx + dr) / (dx + dr + dr);
87 
88 	/* ~47-bit estimate: */
89 	dr = dt * dt * dt;
90 	dt = dt * (dx + dx + dr) / (dx + dr + dr);
91 
92 #if LDBL_MANT_DIG == 64
93 	/*
94 	 * dt is cbrtl(x) to ~47 bits (after x has been reduced to 1 <= x < 8).
95 	 * Round it away from zero to 32 bits (32 so that t*t is exact, and
96 	 * away from zero for technical reasons).
97 	 */
98 	t = dt + (0x1.0p32L + 0x1.0p-31L) - 0x1.0p32;
99 #elif LDBL_MANT_DIG == 113
100 	/*
101 	 * Round dt away from zero to 47 bits.  Since we don't trust the 47,
102 	 * add 2 47-bit ulps instead of 1 to round up.  Rounding is slow and
103 	 * might be avoidable in this case, since on most machines dt will
104 	 * have been evaluated in 53-bit precision and the technical reasons
105 	 * for rounding up might not apply to either case in cbrtl() since
106 	 * dt is much more accurate than needed.
107 	 */
108 	t = dt + 0x2.0p-46 + 0x1.0p60L - 0x1.0p60;
109 #endif
110 
111 	/*
112 	 * Final step Newton iteration to 64 or 113 bits with
113 	 * error < 0.667 ulps
114 	 */
115 	s = t*t;         /* t*t is exact */
116 	r = x/s;         /* error <= 0.5 ulps; |r| < |t| */
117 	w = t+t;         /* t+t is exact */
118 	r = (r-t)/(w+r); /* r-t is exact; w+r ~= 3*t */
119 	t = t+t*r;       /* error <= 0.5 + 0.5/3 + epsilon */
120 
121 	t *= v.f;
122 	return t;
123 }
124 #endif
125