xref: /openbsd/lib/libm/src/ld128/s_cbrtl.c (revision cca36db2)
1 /*-
2  * ====================================================
3  * Copyright (C) 1993 by Sun Microsystems, Inc. All rights reserved.
4  * Copyright (c) 2009-2011, Bruce D. Evans, Steven G. Kargl, David Schultz.
5  *
6  * Developed at SunPro, a Sun Microsystems, Inc. business.
7  * Permission to use, copy, modify, and distribute this
8  * software is freely granted, provided that this notice
9  * is preserved.
10  * ====================================================
11  *
12  * The argument reduction and testing for exceptional cases was
13  * written by Steven G. Kargl with input from Bruce D. Evans
14  * and David A. Schultz.
15  */
16 
17 #include <sys/cdefs.h>
18 #if 0
19 __FBSDID("$FreeBSD: src/lib/msun/src/s_cbrtl.c,v 1.1 2011/03/12 19:37:35 kargl Exp $");
20 #endif
21 
22 #include <float.h>
23 #include <ieeefp.h>
24 #include <math.h>
25 
26 #include "math_private.h"
27 
28 #define	BIAS	(LDBL_MAX_EXP - 1)
29 
30 static const unsigned
31     B1 = 709958130;	/* B1 = (127-127.0/3-0.03306235651)*2**23 */
32 
33 long double
34 cbrtl(long double x)
35 {
36 	long double v, r, s, t, w;
37 	double dr, dt, dx;
38 	float ft, fx;
39 	uint64_t hx, lx;
40 	uint16_t expsign;
41 	int k;
42 
43 	GET_LDOUBLE_MSW64(hx,x);
44 	k = (hx>>48)&0x7fff;
45 
46 	/*
47 	 * If x = +-Inf, then cbrt(x) = +-Inf.
48 	 * If x = NaN, then cbrt(x) = NaN.
49 	 */
50 	if (k == BIAS + LDBL_MAX_EXP)
51 		return (x + x);
52 
53 	if (k == 0) {
54 		/* If x = +-0, then cbrt(x) = +-0. */
55 		GET_LDOUBLE_WORDS64(hx,lx,x);
56 		if (((hx&0x7fffffffffffffffLL)|lx) == 0) {
57 			return (x);
58 		}
59 		/* Adjust subnormal numbers. */
60 		x *= 0x1.0p514;
61 		GET_LDOUBLE_MSW64(hx,x);
62 		k = (hx>>48)&0x7fff;
63 		k -= BIAS + 514;
64 	} else
65 		k -= BIAS;
66 	GET_LDOUBLE_MSW64(hx,x);
67 	hx = (hx&0x8000ffffffffffffLL)|((uint64_t)BIAS<<48);
68 	SET_LDOUBLE_MSW64(x,hx);
69 	v = 1;
70 
71 	switch (k % 3) {
72 	case 1:
73 	case -2:
74 		x = 2*x;
75 		k--;
76 		break;
77 	case 2:
78 	case -1:
79 		x = 4*x;
80 		k -= 2;
81 		break;
82 	}
83 	GET_LDOUBLE_MSW64(hx,x);
84 	expsign = ((hx>>48) & 0x8000) | (BIAS + k / 3);
85 	hx = (hx&0x8000ffffffffffffLL)|((uint64_t)expsign<<48);
86 	SET_LDOUBLE_MSW64(x,hx);
87 
88 	/*
89 	 * The following is the guts of s_cbrtf, with the handling of
90 	 * special values removed and extra care for accuracy not taken,
91 	 * but with most of the extra accuracy not discarded.
92 	 */
93 
94 	/* ~5-bit estimate: */
95 	fx = x;
96 	GET_FLOAT_WORD(hx, fx);
97 	SET_FLOAT_WORD(ft, ((hx & 0x7fffffff) / 3 + B1));
98 
99 	/* ~16-bit estimate: */
100 	dx = x;
101 	dt = ft;
102 	dr = dt * dt * dt;
103 	dt = dt * (dx + dx + dr) / (dx + dr + dr);
104 
105 	/* ~47-bit estimate: */
106 	dr = dt * dt * dt;
107 	dt = dt * (dx + dx + dr) / (dx + dr + dr);
108 
109 	/*
110 	 * Round dt away from zero to 47 bits.  Since we don't trust the 47,
111 	 * add 2 47-bit ulps instead of 1 to round up.  Rounding is slow and
112 	 * might be avoidable in this case, since on most machines dt will
113 	 * have been evaluated in 53-bit precision and the technical reasons
114 	 * for rounding up might not apply to either case in cbrtl() since
115 	 * dt is much more accurate than needed.
116 	 */
117 	t = dt + 0x2.0p-46 + 0x1.0p60L - 0x1.0p60;
118 
119 	/*
120 	 * Final step Newton iteration to 64 or 113 bits with
121 	 * error < 0.667 ulps
122 	 */
123 	s=t*t;				/* t*t is exact */
124 	r=x/s;				/* error <= 0.5 ulps; |r| < |t| */
125 	w=t+t;				/* t+t is exact */
126 	r=(r-t)/(w+r);			/* r-t is exact; w+r ~= 3*t */
127 	t=t+t*r;			/* error <= 0.5 + 0.5/3 + epsilon */
128 
129 	t *= v;
130 	return (t);
131 }
132