133965Sjdp/* origin: FreeBSD /usr/src/lib/msun/src/s_cbrtl.c */
277298Sobrien/*-
389857Sobrien * ====================================================
4218822Sdim * Copyright (C) 1993 by Sun Microsystems, Inc. All rights reserved.
5218822Sdim * Copyright (c) 2009-2011, Bruce D. Evans, Steven G. Kargl, David Schultz.
668765Sobrien *
7218822Sdim * Developed at SunPro, a Sun Microsystems, Inc. business.
868765Sobrien * Permission to use, copy, modify, and distribute this
933965Sjdp * software is freely granted, provided that this notice
1033965Sjdp * is preserved.
1133965Sjdp * ====================================================
1233965Sjdp *
1333965Sjdp * The argument reduction and testing for exceptional cases was
1433965Sjdp * written by Steven G. Kargl with input from Bruce D. Evans
1533965Sjdp * and David A. Schultz.
1633965Sjdp */
1733965Sjdp
1833965Sjdp#include "libm.h"
1933965Sjdp
2033965Sjdp#if LDBL_MANT_DIG == 53 && LDBL_MAX_EXP == 1024
2133965Sjdplong double cbrtl(long double x)
2233965Sjdp{
2333965Sjdp	return cbrt(x);
2433965Sjdp}
25218822Sdim#elif (LDBL_MANT_DIG == 64 || LDBL_MANT_DIG == 113) && LDBL_MAX_EXP == 16384
26218822Sdimstatic const unsigned B1 = 709958130; /* B1 = (127-127.0/3-0.03306235651)*2**23 */
27218822Sdim
2833965Sjdplong double cbrtl(long double x)
2933965Sjdp{
3033965Sjdp	union ldshape u = {x}, v;
3133965Sjdp	union {float f; uint32_t i;} uft;
3233965Sjdp	long double r, s, t, w;
33218822Sdim	double_t dr, dt, dx;
3489857Sobrien	float_t ft;
3589857Sobrien	int e = u.i.se & 0x7fff;
3660484Sobrien	int sign = u.i.se & 0x8000;
3733965Sjdp
3833965Sjdp	/*
3933965Sjdp	 * If x = +-Inf, then cbrt(x) = +-Inf.
4033965Sjdp	 * If x = NaN, then cbrt(x) = NaN.
4133965Sjdp	 */
4233965Sjdp	if (e == 0x7fff)
4333965Sjdp		return x + x;
4433965Sjdp	if (e == 0) {
4533965Sjdp		/* Adjust subnormal numbers. */
4633965Sjdp		u.f *= 0x1p120;
4733965Sjdp		e = u.i.se & 0x7fff;
4833965Sjdp		/* If x = +-0, then cbrt(x) = +-0. */
4933965Sjdp		if (e == 0)
5033965Sjdp			return x;
5133965Sjdp		e -= 120;
5233965Sjdp	}
5333965Sjdp	e -= 0x3fff;
5433965Sjdp	u.i.se = 0x3fff;
5533965Sjdp	x = u.f;
5633965Sjdp	switch (e % 3) {
5768765Sobrien	case 1:
5833965Sjdp	case -2:
5968765Sobrien		x *= 2;
6068765Sobrien		e--;
6168765Sobrien		break;
6268765Sobrien	case 2:
6368765Sobrien	case -1:
6468765Sobrien		x *= 4;
6568765Sobrien		e -= 2;
6677298Sobrien		break;
6777298Sobrien	}
6877298Sobrien	v.f = 1.0;
6968765Sobrien	v.i.se = sign | (0x3fff + e/3);
7077298Sobrien
7177298Sobrien	/*
7277298Sobrien	 * The following is the guts of s_cbrtf, with the handling of
7377298Sobrien	 * special values removed and extra care for accuracy not taken,
7477298Sobrien	 * but with most of the extra accuracy not discarded.
75218822Sdim	 */
7677298Sobrien
7777298Sobrien	/* ~5-bit estimate: */
7877298Sobrien	uft.f = x;
7977298Sobrien	uft.i = (uft.i & 0x7fffffff)/3 + B1;
8077298Sobrien	ft = uft.f;
8168765Sobrien
8268765Sobrien	/* ~16-bit estimate: */
8368765Sobrien	dx = x;
8468765Sobrien	dt = ft;
8568765Sobrien	dr = dt * dt * dt;
8677298Sobrien	dt = dt * (dx + dx + dr) / (dx + dr + dr);
8777298Sobrien
88218822Sdim	/* ~47-bit estimate: */
8977298Sobrien	dr = dt * dt * dt;
90218822Sdim	dt = dt * (dx + dx + dr) / (dx + dr + dr);
9168765Sobrien
92218822Sdim#if LDBL_MANT_DIG == 64
9368765Sobrien	/*
9468765Sobrien	 * dt is cbrtl(x) to ~47 bits (after x has been reduced to 1 <= x < 8).
9568765Sobrien	 * Round it away from zero to 32 bits (32 so that t*t is exact, and
9668765Sobrien	 * away from zero for technical reasons).
9768765Sobrien	 */
9877298Sobrien	t = dt + (0x1.0p32L + 0x1.0p-31L) - 0x1.0p32;
9968765Sobrien#elif LDBL_MANT_DIG == 113
10068765Sobrien	/*
10168765Sobrien	 * Round dt away from zero to 47 bits.  Since we don't trust the 47,
10268765Sobrien	 * add 2 47-bit ulps instead of 1 to round up.  Rounding is slow and
10368765Sobrien	 * might be avoidable in this case, since on most machines dt will
104218822Sdim	 * have been evaluated in 53-bit precision and the technical reasons
10568765Sobrien	 * for rounding up might not apply to either case in cbrtl() since
10668765Sobrien	 * dt is much more accurate than needed.
10768765Sobrien	 */
10868765Sobrien	t = dt + 0x2.0p-46 + 0x1.0p60L - 0x1.0p60;
10968765Sobrien#endif
11068765Sobrien
11168765Sobrien	/*
11268765Sobrien	 * Final step Newton iteration to 64 or 113 bits with
11368765Sobrien	 * error < 0.667 ulps
11468765Sobrien	 */
11568765Sobrien	s = t*t;         /* t*t is exact */
11668765Sobrien	r = x/s;         /* error <= 0.5 ulps; |r| < |t| */
11733965Sjdp	w = t+t;         /* t+t is exact */
11833965Sjdp	r = (r-t)/(w+r); /* r-t is exact; w+r ~= 3*t */
11933965Sjdp	t = t+t*r;       /* error <= 0.5 + 0.5/3 + epsilon */
12033965Sjdp
12133965Sjdp	t *= v.f;
12233965Sjdp	return t;
123218822Sdim}
124218822Sdim#endif
125218822Sdim