xref: /illumos-gate/usr/src/lib/libm/common/Q/sqrtl.c (revision 2aeafac3612e19716bf8164f89c3c9196342979c)
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 (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 
22 /*
23  * Copyright 2011 Nexenta Systems, Inc.  All rights reserved.
24  */
25 /*
26  * Copyright 2006 Sun Microsystems, Inc.  All rights reserved.
27  * Use is subject to license terms.
28  */
29 
30 #pragma weak __sqrtl = sqrtl
31 
32 #include "libm.h"
33 #include "longdouble.h"
34 
35 extern int __swapTE(int);
36 extern int __swapEX(int);
37 extern enum fp_direction_type __swapRD(enum fp_direction_type);
38 
39 /*
40  * in struct longdouble, msw consists of
41  *	unsigned short	sgn:1;
42  *	unsigned short	exp:15;
43  *	unsigned short	frac1:16;
44  */
45 
46 #ifdef __LITTLE_ENDIAN
47 
48 /* array indices used to access words within a double */
49 #define	HIWORD	1
50 #define	LOWORD	0
51 
52 /* structure used to access words within a quad */
53 union longdouble {
54 	struct {
55 		unsigned int	frac4;
56 		unsigned int	frac3;
57 		unsigned int	frac2;
58 		unsigned int	msw;
59 	} l;
60 	long double	d;
61 };
62 
63 /* default NaN returned for sqrt(neg) */
64 static const union longdouble
65 	qnan = { 0xffffffff, 0xffffffff, 0xffffffff, 0x7fffffff };
66 
67 /* signalling NaN used to raise invalid */
68 static const union {
69 	unsigned u[2];
70 	double d;
71 } snan = { 0, 0x7ff00001 };
72 
73 #else
74 
75 /* array indices used to access words within a double */
76 #define	HIWORD	0
77 #define	LOWORD	1
78 
79 /* structure used to access words within a quad */
80 union longdouble {
81 	struct {
82 		unsigned int	msw;
83 		unsigned int	frac2;
84 		unsigned int	frac3;
85 		unsigned int	frac4;
86 	} l;
87 	long double	d;
88 };
89 
90 /* default NaN returned for sqrt(neg) */
91 static const union longdouble
92 	qnan = { 0x7fffffff, 0xffffffff, 0xffffffff, 0xffffffff };
93 
94 /* signalling NaN used to raise invalid */
95 static const union {
96 	unsigned u[2];
97 	double d;
98 } snan = { 0x7ff00001, 0 };
99 
100 #endif /* __LITTLE_ENDIAN */
101 
102 
103 static const double
104 	zero = 0.0,
105 	half = 0.5,
106 	one = 1.0,
107 	huge = 1.0e300,
108 	tiny = 1.0e-300,
109 	two36 = 6.87194767360000000000e+10,
110 	two30 = 1.07374182400000000000e+09,
111 	two6 = 6.40000000000000000000e+01,
112 	two4 = 1.60000000000000000000e+01,
113 	twom18 = 3.81469726562500000000e-06,
114 	twom28 = 3.72529029846191406250e-09,
115 	twom42 = 2.27373675443232059479e-13,
116 	twom60 = 8.67361737988403547206e-19,
117 	twom62 = 2.16840434497100886801e-19,
118 	twom66 = 1.35525271560688054251e-20,
119 	twom90 = 8.07793566946316088742e-28,
120 	twom113 = 9.62964972193617926528e-35,
121 	twom124 = 4.70197740328915003187e-38;
122 
123 
124 /*
125 *	Extract the exponent and normalized significand (represented as
126 *	an array of five doubles) from a finite, nonzero quad.
127 */
128 static int
129 __q_unpack(const union longdouble *x, double *s)
130 {
131 	union {
132 		double			d;
133 		unsigned int	l[2];
134 	} u;
135 	double			b;
136 	unsigned int	lx, w[3];
137 	int				ex;
138 
139 	/* get the normalized significand and exponent */
140 	ex = (int) ((x->l.msw & 0x7fffffff) >> 16);
141 	lx = x->l.msw & 0xffff;
142 	if (ex)
143 	{
144 		lx |= 0x10000;
145 		w[0] = x->l.frac2;
146 		w[1] = x->l.frac3;
147 		w[2] = x->l.frac4;
148 	}
149 	else
150 	{
151 		if (lx | (x->l.frac2 & 0xfffe0000))
152 		{
153 			w[0] = x->l.frac2;
154 			w[1] = x->l.frac3;
155 			w[2] = x->l.frac4;
156 			ex = 1;
157 		}
158 		else if (x->l.frac2 | (x->l.frac3 & 0xfffe0000))
159 		{
160 			lx = x->l.frac2;
161 			w[0] = x->l.frac3;
162 			w[1] = x->l.frac4;
163 			w[2] = 0;
164 			ex = -31;
165 		}
166 		else if (x->l.frac3 | (x->l.frac4 & 0xfffe0000))
167 		{
168 			lx = x->l.frac3;
169 			w[0] = x->l.frac4;
170 			w[1] = w[2] = 0;
171 			ex = -63;
172 		}
173 		else
174 		{
175 			lx = x->l.frac4;
176 			w[0] = w[1] = w[2] = 0;
177 			ex = -95;
178 		}
179 		while ((lx & 0x10000) == 0)
180 		{
181 			lx = (lx << 1) | (w[0] >> 31);
182 			w[0] = (w[0] << 1) | (w[1] >> 31);
183 			w[1] = (w[1] << 1) | (w[2] >> 31);
184 			w[2] <<= 1;
185 			ex--;
186 		}
187 	}
188 
189 	/* extract the significand into five doubles */
190 	u.l[HIWORD] = 0x42300000;
191 	u.l[LOWORD] = 0;
192 	b = u.d;
193 	u.l[LOWORD] = lx;
194 	s[0] = u.d - b;
195 
196 	u.l[HIWORD] = 0x40300000;
197 	u.l[LOWORD] = 0;
198 	b = u.d;
199 	u.l[LOWORD] = w[0] & 0xffffff00;
200 	s[1] = u.d - b;
201 
202 	u.l[HIWORD] = 0x3e300000;
203 	u.l[LOWORD] = 0;
204 	b = u.d;
205 	u.l[HIWORD] |= w[0] & 0xff;
206 	u.l[LOWORD] = w[1] & 0xffff0000;
207 	s[2] = u.d - b;
208 
209 	u.l[HIWORD] = 0x3c300000;
210 	u.l[LOWORD] = 0;
211 	b = u.d;
212 	u.l[HIWORD] |= w[1] & 0xffff;
213 	u.l[LOWORD] = w[2] & 0xff000000;
214 	s[3] = u.d - b;
215 
216 	u.l[HIWORD] = 0x3c300000;
217 	u.l[LOWORD] = 0;
218 	b = u.d;
219 	u.l[LOWORD] = w[2] & 0xffffff;
220 	s[4] = u.d - b;
221 
222 	return ex - 0x3fff;
223 }
224 
225 
226 /*
227 *	Pack an exponent and array of three doubles representing a finite,
228 *	nonzero number into a quad.  Assume the sign is already there and
229 *	the rounding mode has been fudged accordingly.
230 */
231 static void
232 __q_pack(const double *z, int exp, enum fp_direction_type rm,
233 	union longdouble *x, int *inexact)
234 {
235 	union {
236 		double			d;
237 		unsigned int	l[2];
238 	} u;
239 	double			s[3], t, t2;
240 	unsigned int	msw, frac2, frac3, frac4;
241 
242 	/* bias exponent and strip off integer bit */
243 	exp += 0x3fff;
244 	s[0] = z[0] - one;
245 	s[1] = z[1];
246 	s[2] = z[2];
247 
248 	/*
249 	 * chop the significand to obtain the fraction;
250 	 * use round-to-minus-infinity to ensure chopping
251 	 */
252 	(void) __swapRD(fp_negative);
253 
254 	/* extract the first eighty bits of fraction */
255 	t = s[1] + s[2];
256 	u.d = two36 + (s[0] + t);
257 	msw = u.l[LOWORD];
258 	s[0] -= (u.d - two36);
259 
260 	u.d = two4 + (s[0] + t);
261 	frac2 = u.l[LOWORD];
262 	s[0] -= (u.d - two4);
263 
264 	u.d = twom28 + (s[0] + t);
265 	frac3 = u.l[LOWORD];
266 	s[0] -= (u.d - twom28);
267 
268 	/* condense the remaining fraction; errors here won't matter */
269 	t = s[0] + s[1];
270 	s[1] = ((s[0] - t) + s[1]) + s[2];
271 	s[0] = t;
272 
273 	/* get the last word of fraction */
274 	u.d = twom60 + (s[0] + s[1]);
275 	frac4 = u.l[LOWORD];
276 	s[0] -= (u.d - twom60);
277 
278 	/*
279 	 * keep track of what's left for rounding; note that
280 	 * t2 will be non-negative due to rounding mode
281 	 */
282 	t = s[0] + s[1];
283 	t2 = (s[0] - t) + s[1];
284 
285 	if (t != zero)
286 	{
287 		*inexact = 1;
288 
289 		/* decide whether to round the fraction up */
290 		if (rm == fp_positive || (rm == fp_nearest && (t > twom113 ||
291 			(t == twom113 && (t2 != zero || frac4 & 1)))))
292 		{
293 			/* round up and renormalize if necessary */
294 			if (++frac4 == 0)
295 				if (++frac3 == 0)
296 					if (++frac2 == 0)
297 						if (++msw == 0x10000)
298 						{
299 							msw = 0;
300 							exp++;
301 						}
302 		}
303 	}
304 
305 	/* assemble the result */
306 	x->l.msw |= msw | (exp << 16);
307 	x->l.frac2 = frac2;
308 	x->l.frac3 = frac3;
309 	x->l.frac4 = frac4;
310 }
311 
312 
313 /*
314 *	Compute the square root of x and place the TP result in s.
315 */
316 static void
317 __q_tp_sqrt(const double *x, double *s)
318 {
319 	double	c, rr, r[3], tt[3], t[5];
320 
321 	/* approximate the divisor for the Newton iteration */
322 	c = sqrt((x[0] + x[1]) + x[2]);
323 	rr = half / c;
324 
325 	/* compute the first five "digits" of the square root */
326 	t[0] = (c + two30) - two30;
327 	tt[0] = t[0] + t[0];
328 	r[0] = ((x[0] - t[0] * t[0]) + x[1]) + x[2];
329 
330 	t[1] = (rr * (r[0] + x[3]) + two6) - two6;
331 	tt[1] = t[1] + t[1];
332 	r[0] -= tt[0] * t[1];
333 	r[1] = x[3] - t[1] * t[1];
334 	c = (r[1] + twom18) - twom18;
335 	r[0] += c;
336 	r[1] = (r[1] - c) + x[4];
337 
338 	t[2] = (rr * (r[0] + r[1]) + twom18) - twom18;
339 	tt[2] = t[2] + t[2];
340 	r[0] -= tt[0] * t[2];
341 	r[1] -= tt[1] * t[2];
342 	c = (r[1] + twom42) - twom42;
343 	r[0] += c;
344 	r[1] = (r[1] - c) - t[2] * t[2];
345 
346 	t[3] = (rr * (r[0] + r[1]) + twom42) - twom42;
347 	r[0] = ((r[0] - tt[0] * t[3]) + r[1]) - tt[1] * t[3];
348 	r[1] = -tt[2] * t[3];
349 	c = (r[1] + twom90) - twom90;
350 	r[0] += c;
351 	r[1] = (r[1] - c) - t[3] * t[3];
352 
353 	t[4] = (rr * (r[0] + r[1]) + twom66) - twom66;
354 
355 	/* here we just need to get the sign of the remainder */
356 	c = (((((r[0] - tt[0] * t[4]) - tt[1] * t[4]) + r[1])
357 		- tt[2] * t[4]) - (t[3] + t[3]) * t[4]) - t[4] * t[4];
358 
359 	/* reduce to three doubles */
360 	t[0] += t[1];
361 	t[1] = t[2] + t[3];
362 	t[2] = t[4];
363 
364 	/* if the third term might lie on a rounding boundary, perturb it */
365 	if (c != zero && t[2] == (twom62 + t[2]) - twom62)
366 	{
367 		if (c < zero)
368 			t[2] -= twom124;
369 		else
370 			t[2] += twom124;
371 	}
372 
373 	/* condense the square root */
374 	c = t[1] + t[2];
375 	t[2] += (t[1] - c);
376 	t[1] = c;
377 	c = t[0] + t[1];
378 	s[1] = t[1] + (t[0] - c);
379 	s[0] = c;
380 	if (s[1] == zero)
381 	{
382 		c = s[0] + t[2];
383 		s[1] = t[2] + (s[0] - c);
384 		s[0] = c;
385 		s[2] = zero;
386 	}
387 	else
388 	{
389 		c = s[1] + t[2];
390 		s[2] = t[2] + (s[1] - c);
391 		s[1] = c;
392 	}
393 }
394 
395 
396 long double
397 sqrtl(long double ldx)
398 {
399 	union	longdouble		x;
400 	volatile double			t;
401 	double					xx[5], zz[3];
402 	enum fp_direction_type	rm;
403 	int				ex, inexact, exc, traps;
404 
405 	/* clear cexc */
406 	t = zero;
407 	t -= zero;
408 
409 	/* check for zero operand */
410 	x.d = ldx;
411 	if (!((x.l.msw & 0x7fffffff) | x.l.frac2 | x.l.frac3 | x.l.frac4))
412 		return ldx;
413 
414 	/* handle nan and inf cases */
415 	if ((x.l.msw & 0x7fffffff) >= 0x7fff0000)
416 	{
417 		if ((x.l.msw & 0xffff) | x.l.frac2 | x.l.frac3 | x.l.frac4)
418 		{
419 			if (!(x.l.msw & 0x8000))
420 			{
421 				/* snan, signal invalid */
422 				t += snan.d;
423 			}
424 			x.l.msw |= 0x8000;
425 			return x.d;
426 		}
427 		if (x.l.msw & 0x80000000)
428 		{
429 			/* sqrt(-inf), signal invalid */
430 			t = -one;
431 			t = sqrt(t);
432 			return qnan.d;
433 		}
434 		/* sqrt(inf), return inf */
435 		return x.d;
436 	}
437 
438 	/* handle negative numbers */
439 	if (x.l.msw & 0x80000000)
440 	{
441 		t = -one;
442 		t = sqrt(t);
443 		return qnan.d;
444 	}
445 
446 	/* now x is finite, positive */
447 
448 	traps = __swapTE(0);
449 	exc = __swapEX(0);
450 	rm = __swapRD(fp_nearest);
451 
452 	ex = __q_unpack(&x, xx);
453 	if (ex & 1)
454 	{
455 		/* make exponent even */
456 		xx[0] += xx[0];
457 		xx[1] += xx[1];
458 		xx[2] += xx[2];
459 		xx[3] += xx[3];
460 		xx[4] += xx[4];
461 		ex--;
462 	}
463 	__q_tp_sqrt(xx, zz);
464 
465 	/* put everything together */
466 	x.l.msw = 0;
467 	inexact = 0;
468 	__q_pack(zz, ex >> 1, rm, &x, &inexact);
469 
470 	(void) __swapRD(rm);
471 	(void) __swapEX(exc);
472 	(void) __swapTE(traps);
473 	if (inexact)
474 	{
475 		t = huge;
476 		t += tiny;
477 	}
478 	return x.d;
479 }
480