xref: /relibc/openlibm/src/s_cbrtl.c (revision 232ba9db6e4b36dd94ff2ca1022f9ba09970a2ec)
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 "cdefs-compat.h"
18 //__FBSDID("$FreeBSD: src/lib/msun/src/s_cbrtl.c,v 1.1 2011/03/12 19:37:35 kargl Exp $");
19 
20 #include <float.h>
21 #include <openlibm_math.h>
22 // VBS
23 //#include <ieeefp.h>
24 
25 #include "fpmath.h"
26 #include "math_private.h"
27 #if defined(__i386__)
28 #include "i387/bsd_ieeefp.h"
29 #endif
30 
31 #define	BIAS	(LDBL_MAX_EXP - 1)
32 
33 static const unsigned
34     B1 = 709958130;	/* B1 = (127-127.0/3-0.03306235651)*2**23 */
35 
36 OLM_DLLEXPORT long double
37 cbrtl(long double x)
38 {
39 	union IEEEl2bits u, v;
40 	long double r, s, t, w;
41 	double dr, dt, dx;
42 	float ft, fx;
43 	u_int32_t hx;
44 	u_int16_t expsign;
45 	int k;
46 
47 	u.e = x;
48 	expsign = u.xbits.expsign;
49 	k = expsign & 0x7fff;
50 
51 	/*
52 	 * If x = +-Inf, then cbrt(x) = +-Inf.
53 	 * If x = NaN, then cbrt(x) = NaN.
54 	 */
55 	if (k == BIAS + LDBL_MAX_EXP)
56 		return (x + x);
57 
58 #ifdef __i386__
59 	fp_prec_t oprec;
60 
61 	oprec = fpgetprec();
62 	if (oprec != FP_PE)
63 		fpsetprec(FP_PE);
64 #endif
65 
66 	if (k == 0) {
67 		/* If x = +-0, then cbrt(x) = +-0. */
68 		if ((u.bits.manh | u.bits.manl) == 0) {
69 #ifdef __i386__
70 			if (oprec != FP_PE)
71 				fpsetprec(oprec);
72 #endif
73 			return (x);
74 	    	}
75 		/* Adjust subnormal numbers. */
76 		u.e *= 0x1.0p514;
77 		k = u.bits.exp;
78 		k -= BIAS + 514;
79  	} else
80 		k -= BIAS;
81 	u.xbits.expsign = BIAS;
82 	v.e = 1;
83 
84 	x = u.e;
85 	switch (k % 3) {
86 	case 1:
87 	case -2:
88 		x = 2*x;
89 		k--;
90 		break;
91 	case 2:
92 	case -1:
93 		x = 4*x;
94 		k -= 2;
95 		break;
96 	}
97 	v.xbits.expsign = (expsign & 0x8000) | (BIAS + k / 3);
98 
99 	/*
100 	 * The following is the guts of s_cbrtf, with the handling of
101 	 * special values removed and extra care for accuracy not taken,
102 	 * but with most of the extra accuracy not discarded.
103 	 */
104 
105 	/* ~5-bit estimate: */
106 	fx = x;
107 	GET_FLOAT_WORD(hx, fx);
108 	SET_FLOAT_WORD(ft, ((hx & 0x7fffffff) / 3 + B1));
109 
110 	/* ~16-bit estimate: */
111 	dx = x;
112 	dt = ft;
113 	dr = dt * dt * dt;
114 	dt = dt * (dx + dx + dr) / (dx + dr + dr);
115 
116 	/* ~47-bit estimate: */
117 	dr = dt * dt * dt;
118 	dt = dt * (dx + dx + dr) / (dx + dr + dr);
119 
120 #if LDBL_MANT_DIG == 64
121 	/*
122 	 * dt is cbrtl(x) to ~47 bits (after x has been reduced to 1 <= x < 8).
123 	 * Round it away from zero to 32 bits (32 so that t*t is exact, and
124 	 * away from zero for technical reasons).
125 	 */
126 	volatile double vd2 = 0x1.0p32;
127 	volatile double vd1 = 0x1.0p-31;
128 	#define vd ((long double)vd2 + vd1)
129 
130 	t = dt + vd - 0x1.0p32;
131 #elif LDBL_MANT_DIG == 113
132 	/*
133 	 * Round dt away from zero to 47 bits.  Since we don't trust the 47,
134 	 * add 2 47-bit ulps instead of 1 to round up.  Rounding is slow and
135 	 * might be avoidable in this case, since on most machines dt will
136 	 * have been evaluated in 53-bit precision and the technical reasons
137 	 * for rounding up might not apply to either case in cbrtl() since
138 	 * dt is much more accurate than needed.
139 	 */
140 	t = dt + 0x2.0p-46 + 0x1.0p60L - 0x1.0p60;
141 #else
142 #error "Unsupported long double format"
143 #endif
144 
145 	/*
146      	 * Final step Newton iteration to 64 or 113 bits with
147 	 * error < 0.667 ulps
148 	 */
149 	s=t*t;				/* t*t is exact */
150 	r=x/s;				/* error <= 0.5 ulps; |r| < |t| */
151 	w=t+t;				/* t+t is exact */
152 	r=(r-t)/(w+r);			/* r-t is exact; w+r ~= 3*t */
153 	t=t+t*r;			/* error <= 0.5 + 0.5/3 + epsilon */
154 
155 	t *= v.e;
156 #ifdef __i386__
157 	if (oprec != FP_PE)
158 		fpsetprec(oprec);
159 #endif
160 	return (t);
161 }
162