xref: /illumos-gate/usr/src/lib/libc/sparc/fp/_Q_sqrt.c (revision 7c478bd9)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License, Version 1.0 only
6  * (the "License").  You may not use this file except in compliance
7  * with the License.
8  *
9  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
10  * or http://www.opensolaris.org/os/licensing.
11  * See the License for the specific language governing permissions
12  * and limitations under the License.
13  *
14  * When distributing Covered Code, include this CDDL HEADER in each
15  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
16  * If applicable, add the following below this CDDL HEADER, with the
17  * fields enclosed by brackets "[]" replaced with your own identifying
18  * information: Portions Copyright [yyyy] [name of copyright owner]
19  *
20  * CDDL HEADER END
21  */
22 /*
23  * Copyright 2003 Sun Microsystems, Inc.  All rights reserved.
24  * Use is subject to license terms.
25  */
26 
27 #pragma ident	"%Z%%M%	%I%	%E% SMI"
28 
29 #include "quad.h"
30 
31 static const double C[] = {
32 	0.0,
33 	0.5,
34 	1.0,
35 	68719476736.0,
36 	536870912.0,
37 	48.0,
38 	16.0,
39 	1.52587890625000000000e-05,
40 	2.86102294921875000000e-06,
41 	5.96046447753906250000e-08,
42 	3.72529029846191406250e-09,
43 	1.70530256582424044609e-13,
44 	7.10542735760100185871e-15,
45 	8.67361737988403547206e-19,
46 	2.16840434497100886801e-19,
47 	1.27054942088145050860e-21,
48 	1.21169035041947413311e-27,
49 	9.62964972193617926528e-35,
50 	4.70197740328915003187e-38
51 };
52 
53 #define	zero		C[0]
54 #define	half		C[1]
55 #define	one		C[2]
56 #define	two36		C[3]
57 #define	two29		C[4]
58 #define	three2p4	C[5]
59 #define	two4		C[6]
60 #define	twom16		C[7]
61 #define	three2m20	C[8]
62 #define	twom24		C[9]
63 #define	twom28		C[10]
64 #define	three2m44	C[11]
65 #define	twom47		C[12]
66 #define	twom60		C[13]
67 #define	twom62		C[14]
68 #define	three2m71	C[15]
69 #define	three2m91	C[16]
70 #define	twom113		C[17]
71 #define	twom124		C[18]
72 
73 static const unsigned
74 	fsr_re = 0x00000000u,
75 	fsr_rn = 0xc0000000u;
76 
77 #ifdef __sparcv9
78 
79 /*
80  * _Qp_sqrt(pz, x) sets *pz = sqrt(*x).
81  */
82 void
83 _Qp_sqrt(union longdouble *pz, const union longdouble *x)
84 
85 #else
86 
87 /*
88  * _Q_sqrt(x) returns sqrt(*x).
89  */
90 union longdouble
91 _Q_sqrt(const union longdouble *x)
92 
93 #endif	/* __sparcv9 */
94 
95 {
96 	union longdouble	z;
97 	union xdouble		u;
98 	double			c, d, rr, r[2], tt[3], xx[4], zz[5];
99 	unsigned int		xm, fsr, lx, wx[3];
100 	unsigned int		msw, frac2, frac3, frac4, rm;
101 	int			ex, ez;
102 
103 	if (QUAD_ISZERO(*x)) {
104 		Z = *x;
105 		QUAD_RETURN(Z);
106 	}
107 
108 	xm = x->l.msw;
109 
110 	__quad_getfsrp(&fsr);
111 
112 	/* handle nan and inf cases */
113 	if ((xm & 0x7fffffff) >= 0x7fff0000) {
114 		if ((x->l.msw & 0xffff) | x->l.frac2 | x->l.frac3 |
115 		    x->l.frac4) {
116 			if (!(x->l.msw & 0x8000)) {
117 				/* snan, signal invalid */
118 				if (fsr & FSR_NVM) {
119 					__quad_fsqrtq(x, &Z);
120 				} else {
121 					Z = *x;
122 					Z.l.msw |= 0x8000;
123 					fsr = (fsr & ~FSR_CEXC) | FSR_NVA |
124 					    FSR_NVC;
125 					__quad_setfsrp(&fsr);
126 				}
127 				QUAD_RETURN(Z);
128 			}
129 			Z = *x;
130 			QUAD_RETURN(Z);
131 		}
132 		if (x->l.msw & 0x80000000) {
133 			/* sqrt(-inf), signal invalid */
134 			if (fsr & FSR_NVM) {
135 				__quad_fsqrtq(x, &Z);
136 			} else {
137 				Z.l.msw = 0x7fffffff;
138 				Z.l.frac2 = Z.l.frac3 = Z.l.frac4 = 0xffffffff;
139 				fsr = (fsr & ~FSR_CEXC) | FSR_NVA | FSR_NVC;
140 				__quad_setfsrp(&fsr);
141 			}
142 			QUAD_RETURN(Z);
143 		}
144 		/* sqrt(inf), return inf */
145 		Z = *x;
146 		QUAD_RETURN(Z);
147 	}
148 
149 	/* handle negative numbers */
150 	if (xm & 0x80000000) {
151 		if (fsr & FSR_NVM) {
152 			__quad_fsqrtq(x, &Z);
153 		} else {
154 			Z.l.msw = 0x7fffffff;
155 			Z.l.frac2 = Z.l.frac3 = Z.l.frac4 = 0xffffffff;
156 			fsr = (fsr & ~FSR_CEXC) | FSR_NVA | FSR_NVC;
157 			__quad_setfsrp(&fsr);
158 		}
159 		QUAD_RETURN(Z);
160 	}
161 
162 	/* now x is finite, positive */
163 	__quad_setfsrp((unsigned *)&fsr_re);
164 
165 	/* get the normalized significand and exponent */
166 	ex = (int)(xm >> 16);
167 	lx = xm & 0xffff;
168 	if (ex) {
169 		lx |= 0x10000;
170 		wx[0] = x->l.frac2;
171 		wx[1] = x->l.frac3;
172 		wx[2] = x->l.frac4;
173 	} else {
174 		if (lx | (x->l.frac2 & 0xfffe0000)) {
175 			wx[0] = x->l.frac2;
176 			wx[1] = x->l.frac3;
177 			wx[2] = x->l.frac4;
178 			ex = 1;
179 		} else if (x->l.frac2 | (x->l.frac3 & 0xfffe0000)) {
180 			lx = x->l.frac2;
181 			wx[0] = x->l.frac3;
182 			wx[1] = x->l.frac4;
183 			wx[2] = 0;
184 			ex = -31;
185 		} else if (x->l.frac3 | (x->l.frac4 & 0xfffe0000)) {
186 			lx = x->l.frac3;
187 			wx[0] = x->l.frac4;
188 			wx[1] = wx[2] = 0;
189 			ex = -63;
190 		} else {
191 			lx = x->l.frac4;
192 			wx[0] = wx[1] = wx[2] = 0;
193 			ex = -95;
194 		}
195 		while ((lx & 0x10000) == 0) {
196 			lx = (lx << 1) | (wx[0] >> 31);
197 			wx[0] = (wx[0] << 1) | (wx[1] >> 31);
198 			wx[1] = (wx[1] << 1) | (wx[2] >> 31);
199 			wx[2] <<= 1;
200 			ex--;
201 		}
202 	}
203 	ez = ex - 0x3fff;
204 	if (ez & 1) {
205 		/* make exponent even */
206 		lx = (lx << 1) | (wx[0] >> 31);
207 		wx[0] = (wx[0] << 1) | (wx[1] >> 31);
208 		wx[1] = (wx[1] << 1) | (wx[2] >> 31);
209 		wx[2] <<= 1;
210 		ez--;
211 	}
212 
213 	/* extract the significands into doubles */
214 	c = twom16;
215 	xx[0] = (double)((int)lx) * c;
216 
217 	c *= twom24;
218 	xx[0] += (double)((int)(wx[0] >> 8)) * c;
219 
220 	c *= twom24;
221 	xx[1] = (double)((int)(((wx[0] << 16) | (wx[1] >> 16)) &
222 	    0xffffff)) * c;
223 
224 	c *= twom24;
225 	xx[2] = (double)((int)(((wx[1] << 8) | (wx[2] >> 24)) &
226 	    0xffffff)) * c;
227 
228 	c *= twom24;
229 	xx[3] = (double)((int)(wx[2] & 0xffffff)) * c;
230 
231 	/* approximate the divisor for the Newton iteration */
232 	c = xx[0] + xx[1];
233 	c = __quad_dp_sqrt(&c);
234 	rr = half / c;
235 
236 	/* compute the first five "digits" of the square root */
237 	zz[0] = (c + two29) - two29;
238 	tt[0] = zz[0] + zz[0];
239 	r[0] = (xx[0] - zz[0] * zz[0]) + xx[1];
240 
241 	zz[1] = (rr * (r[0] + xx[2]) + three2p4) - three2p4;
242 	tt[1] = zz[1] + zz[1];
243 	r[0] -= tt[0] * zz[1];
244 	r[1] = xx[2] - zz[1] * zz[1];
245 	c = (r[1] + three2m20) - three2m20;
246 	r[0] += c;
247 	r[1] = (r[1] - c) + xx[3];
248 
249 	zz[2] = (rr * (r[0] + r[1]) + three2m20) - three2m20;
250 	tt[2] = zz[2] + zz[2];
251 	r[0] -= tt[0] * zz[2];
252 	r[1] -= tt[1] * zz[2];
253 	c = (r[1] + three2m44) - three2m44;
254 	r[0] += c;
255 	r[1] = (r[1] - c) - zz[2] * zz[2];
256 
257 	zz[3] = (rr * (r[0] + r[1]) + three2m44) - three2m44;
258 	r[0] = ((r[0] - tt[0] * zz[3]) + r[1]) - tt[1] * zz[3];
259 	r[1] = -tt[2] * zz[3];
260 	c = (r[1] + three2m91) - three2m91;
261 	r[0] += c;
262 	r[1] = (r[1] - c) - zz[3] * zz[3];
263 
264 	zz[4] = (rr * (r[0] + r[1]) + three2m71) - three2m71;
265 
266 	/* reduce to three doubles, making sure zz[1] is positive */
267 	zz[0] += zz[1] - twom47;
268 	zz[1] = twom47 + zz[2] + zz[3];
269 	zz[2] = zz[4];
270 
271 	/* if the third term might lie on a rounding boundary, perturb it */
272 	if (zz[2] == (twom62 + zz[2]) - twom62) {
273 		/* here we just need to get the sign of the remainder */
274 		c = (((((r[0] - tt[0] * zz[4]) - tt[1] * zz[4]) + r[1])
275 		    - tt[2] * zz[4]) - (zz[3] + zz[3]) * zz[4]) - zz[4] * zz[4];
276 		if (c < zero)
277 			zz[2] -= twom124;
278 		else if (c > zero)
279 			zz[2] += twom124;
280 	}
281 
282 	/*
283 	 * propagate carries/borrows, using round-to-negative-infinity mode
284 	 * to make all terms nonnegative (note that we can't encounter a
285 	 * borrow so large that the roundoff is unrepresentable because
286 	 * we took care to make zz[1] positive above)
287 	 */
288 	__quad_setfsrp(&fsr_rn);
289 	c = zz[1] + zz[2];
290 	zz[2] += (zz[1] - c);
291 	zz[1] = c;
292 	c = zz[0] + zz[1];
293 	zz[1] += (zz[0] - c);
294 	zz[0] = c;
295 
296 	/* adjust exponent and strip off integer bit */
297 	ez = (ez >> 1) + 0x3fff;
298 	zz[0] -= one;
299 
300 	/* the first 48 bits of fraction come from zz[0] */
301 	u.d = d = two36 + zz[0];
302 	msw = u.l.lo;
303 	zz[0] -= (d - two36);
304 
305 	u.d = d = two4 + zz[0];
306 	frac2 = u.l.lo;
307 	zz[0] -= (d - two4);
308 
309 	/* the next 32 come from zz[0] and zz[1] */
310 	u.d = d = twom28 + (zz[0] + zz[1]);
311 	frac3 = u.l.lo;
312 	zz[0] -= (d - twom28);
313 
314 	/* condense the remaining fraction; errors here won't matter */
315 	c = zz[0] + zz[1];
316 	zz[1] = ((zz[0] - c) + zz[1]) + zz[2];
317 	zz[0] = c;
318 
319 	/* get the last word of fraction */
320 	u.d = d = twom60 + (zz[0] + zz[1]);
321 	frac4 = u.l.lo;
322 	zz[0] -= (d - twom60);
323 
324 	/* keep track of what's left for rounding; note that the error */
325 	/* in computing c will be non-negative due to rounding mode */
326 	c = zz[0] + zz[1];
327 
328 	/* get the rounding mode */
329 	rm = fsr >> 30;
330 
331 	/* round and raise exceptions */
332 	fsr &= ~FSR_CEXC;
333 	if (c != zero) {
334 		fsr |= FSR_NXC;
335 
336 		/* decide whether to round the fraction up */
337 		if (rm == FSR_RP || (rm == FSR_RN && (c > twom113 ||
338 		    (c == twom113 && ((frac4 & 1) || (c - zz[0] != zz[1])))))) {
339 			/* round up and renormalize if necessary */
340 			if (++frac4 == 0)
341 				if (++frac3 == 0)
342 					if (++frac2 == 0)
343 						if (++msw == 0x10000) {
344 							msw = 0;
345 							ez++;
346 						}
347 		}
348 	}
349 
350 	/* stow the result */
351 	z.l.msw = (ez << 16) | msw;
352 	z.l.frac2 = frac2;
353 	z.l.frac3 = frac3;
354 	z.l.frac4 = frac4;
355 
356 	if ((fsr & FSR_CEXC) & (fsr >> 23)) {
357 		__quad_setfsrp(&fsr);
358 		__quad_fsqrtq(x, &Z);
359 	} else {
360 		Z = z;
361 		fsr |= (fsr & 0x1f) << 5;
362 		__quad_setfsrp(&fsr);
363 	}
364 	QUAD_RETURN(Z);
365 }
366