xref: /freebsd/lib/msun/src/e_sqrt.c (revision 2008043f386721d58158e37e0d7e50df8095942d)
1 
2 /* @(#)e_sqrt.c 1.3 95/01/18 */
3 /*
4  * ====================================================
5  * Copyright (C) 1993 by Sun Microsystems, Inc. All rights reserved.
6  *
7  * Developed at SunSoft, a Sun Microsystems, Inc. business.
8  * Permission to use, copy, modify, and distribute this
9  * software is freely granted, provided that this notice
10  * is preserved.
11  * ====================================================
12  */
13 
14 #include <sys/cdefs.h>
15 #include <float.h>
16 
17 #include "math.h"
18 #include "math_private.h"
19 
20 #ifdef USE_BUILTIN_SQRT
21 double
22 sqrt(double x)
23 {
24 	return (__builtin_sqrt(x));
25 }
26 #else
27 /* sqrt(x)
28  * Return correctly rounded sqrt.
29  *           ------------------------------------------
30  *	     |  Use the hardware sqrt if you have one |
31  *           ------------------------------------------
32  * Method:
33  *   Bit by bit method using integer arithmetic. (Slow, but portable)
34  *   1. Normalization
35  *	Scale x to y in [1,4) with even powers of 2:
36  *	find an integer k such that  1 <= (y=x*2^(2k)) < 4, then
37  *		sqrt(x) = 2^k * sqrt(y)
38  *   2. Bit by bit computation
39  *	Let q  = sqrt(y) truncated to i bit after binary point (q = 1),
40  *	     i							 0
41  *                                     i+1         2
42  *	    s  = 2*q , and	y  =  2   * ( y - q  ).		(1)
43  *	     i      i            i                 i
44  *
45  *	To compute q    from q , one checks whether
46  *		    i+1       i
47  *
48  *			      -(i+1) 2
49  *			(q + 2      ) <= y.			(2)
50  *     			  i
51  *							      -(i+1)
52  *	If (2) is false, then q   = q ; otherwise q   = q  + 2      .
53  *		 	       i+1   i             i+1   i
54  *
55  *	With some algebric manipulation, it is not difficult to see
56  *	that (2) is equivalent to
57  *                             -(i+1)
58  *			s  +  2       <= y			(3)
59  *			 i                i
60  *
61  *	The advantage of (3) is that s  and y  can be computed by
62  *				      i      i
63  *	the following recurrence formula:
64  *	    if (3) is false
65  *
66  *	    s     =  s  ,	y    = y   ;			(4)
67  *	     i+1      i		 i+1    i
68  *
69  *	    otherwise,
70  *                         -i                     -(i+1)
71  *	    s	  =  s  + 2  ,  y    = y  -  s  - 2  		(5)
72  *           i+1      i          i+1    i     i
73  *
74  *	One may easily use induction to prove (4) and (5).
75  *	Note. Since the left hand side of (3) contain only i+2 bits,
76  *	      it does not necessary to do a full (53-bit) comparison
77  *	      in (3).
78  *   3. Final rounding
79  *	After generating the 53 bits result, we compute one more bit.
80  *	Together with the remainder, we can decide whether the
81  *	result is exact, bigger than 1/2ulp, or less than 1/2ulp
82  *	(it will never equal to 1/2ulp).
83  *	The rounding mode can be detected by checking whether
84  *	huge + tiny is equal to huge, and whether huge - tiny is
85  *	equal to huge for some floating point number "huge" and "tiny".
86  *
87  * Special cases:
88  *	sqrt(+-0) = +-0 	... exact
89  *	sqrt(inf) = inf
90  *	sqrt(-ve) = NaN		... with invalid signal
91  *	sqrt(NaN) = NaN		... with invalid signal for signaling NaN
92  *
93  * Other methods : see the appended file at the end of the program below.
94  *---------------
95  */
96 
97 static	const double	one	= 1.0, tiny=1.0e-300;
98 
99 double
100 sqrt(double x)
101 {
102 	double z;
103 	int32_t sign = (int)0x80000000;
104 	int32_t ix0,s0,q,m,t,i;
105 	u_int32_t r,t1,s1,ix1,q1;
106 
107 	EXTRACT_WORDS(ix0,ix1,x);
108 
109     /* take care of Inf and NaN */
110 	if((ix0&0x7ff00000)==0x7ff00000) {
111 	    return x*x+x;		/* sqrt(NaN)=NaN, sqrt(+inf)=+inf
112 					   sqrt(-inf)=sNaN */
113 	}
114     /* take care of zero */
115 	if(ix0<=0) {
116 	    if(((ix0&(~sign))|ix1)==0) return x;/* sqrt(+-0) = +-0 */
117 	    else if(ix0<0)
118 		return (x-x)/(x-x);		/* sqrt(-ve) = sNaN */
119 	}
120     /* normalize x */
121 	m = (ix0>>20);
122 	if(m==0) {				/* subnormal x */
123 	    while(ix0==0) {
124 		m -= 21;
125 		ix0 |= (ix1>>11); ix1 <<= 21;
126 	    }
127 	    for(i=0;(ix0&0x00100000)==0;i++) ix0<<=1;
128 	    m -= i-1;
129 	    ix0 |= (ix1>>(32-i));
130 	    ix1 <<= i;
131 	}
132 	m -= 1023;	/* unbias exponent */
133 	ix0 = (ix0&0x000fffff)|0x00100000;
134 	if(m&1){	/* odd m, double x to make it even */
135 	    ix0 += ix0 + ((ix1&sign)>>31);
136 	    ix1 += ix1;
137 	}
138 	m >>= 1;	/* m = [m/2] */
139 
140     /* generate sqrt(x) bit by bit */
141 	ix0 += ix0 + ((ix1&sign)>>31);
142 	ix1 += ix1;
143 	q = q1 = s0 = s1 = 0;	/* [q,q1] = sqrt(x) */
144 	r = 0x00200000;		/* r = moving bit from right to left */
145 
146 	while(r!=0) {
147 	    t = s0+r;
148 	    if(t<=ix0) {
149 		s0   = t+r;
150 		ix0 -= t;
151 		q   += r;
152 	    }
153 	    ix0 += ix0 + ((ix1&sign)>>31);
154 	    ix1 += ix1;
155 	    r>>=1;
156 	}
157 
158 	r = sign;
159 	while(r!=0) {
160 	    t1 = s1+r;
161 	    t  = s0;
162 	    if((t<ix0)||((t==ix0)&&(t1<=ix1))) {
163 		s1  = t1+r;
164 		if(((t1&sign)==sign)&&(s1&sign)==0) s0 += 1;
165 		ix0 -= t;
166 		if (ix1 < t1) ix0 -= 1;
167 		ix1 -= t1;
168 		q1  += r;
169 	    }
170 	    ix0 += ix0 + ((ix1&sign)>>31);
171 	    ix1 += ix1;
172 	    r>>=1;
173 	}
174 
175     /* use floating add to find out rounding direction */
176 	if((ix0|ix1)!=0) {
177 	    z = one-tiny; /* trigger inexact flag */
178 	    if (z>=one) {
179 	        z = one+tiny;
180 	        if (q1==(u_int32_t)0xffffffff) { q1=0; q += 1;}
181 		else if (z>one) {
182 		    if (q1==(u_int32_t)0xfffffffe) q+=1;
183 		    q1+=2;
184 		} else
185 	            q1 += (q1&1);
186 	    }
187 	}
188 	ix0 = (q>>1)+0x3fe00000;
189 	ix1 =  q1>>1;
190 	if ((q&1)==1) ix1 |= sign;
191 	ix0 += (m <<20);
192 	INSERT_WORDS(z,ix0,ix1);
193 	return z;
194 }
195 #endif
196 
197 #if (LDBL_MANT_DIG == 53)
198 __weak_reference(sqrt, sqrtl);
199 #endif
200 
201 /*
202 Other methods  (use floating-point arithmetic)
203 -------------
204 (This is a copy of a drafted paper by Prof W. Kahan
205 and K.C. Ng, written in May, 1986)
206 
207 	Two algorithms are given here to implement sqrt(x)
208 	(IEEE double precision arithmetic) in software.
209 	Both supply sqrt(x) correctly rounded. The first algorithm (in
210 	Section A) uses newton iterations and involves four divisions.
211 	The second one uses reciproot iterations to avoid division, but
212 	requires more multiplications. Both algorithms need the ability
213 	to chop results of arithmetic operations instead of round them,
214 	and the INEXACT flag to indicate when an arithmetic operation
215 	is executed exactly with no roundoff error, all part of the
216 	standard (IEEE 754-1985). The ability to perform shift, add,
217 	subtract and logical AND operations upon 32-bit words is needed
218 	too, though not part of the standard.
219 
220 A.  sqrt(x) by Newton Iteration
221 
222    (1)	Initial approximation
223 
224 	Let x0 and x1 be the leading and the trailing 32-bit words of
225 	a floating point number x (in IEEE double format) respectively
226 
227 	    1    11		     52				  ...widths
228 	   ------------------------------------------------------
229 	x: |s|	  e     |	      f				|
230 	   ------------------------------------------------------
231 	      msb    lsb  msb				      lsb ...order
232 
233 
234 	     ------------------------  	     ------------------------
235 	x0:  |s|   e    |    f1     |	 x1: |          f2           |
236 	     ------------------------  	     ------------------------
237 
238 	By performing shifts and subtracts on x0 and x1 (both regarded
239 	as integers), we obtain an 8-bit approximation of sqrt(x) as
240 	follows.
241 
242 		k  := (x0>>1) + 0x1ff80000;
243 		y0 := k - T1[31&(k>>15)].	... y ~ sqrt(x) to 8 bits
244 	Here k is a 32-bit integer and T1[] is an integer array containing
245 	correction terms. Now magically the floating value of y (y's
246 	leading 32-bit word is y0, the value of its trailing word is 0)
247 	approximates sqrt(x) to almost 8-bit.
248 
249 	Value of T1:
250 	static int T1[32]= {
251 	0,	1024,	3062,	5746,	9193,	13348,	18162,	23592,
252 	29598,	36145,	43202,	50740,	58733,	67158,	75992,	85215,
253 	83599,	71378,	60428,	50647,	41945,	34246,	27478,	21581,
254 	16499,	12183,	8588,	5674,	3403,	1742,	661,	130,};
255 
256     (2)	Iterative refinement
257 
258 	Apply Heron's rule three times to y, we have y approximates
259 	sqrt(x) to within 1 ulp (Unit in the Last Place):
260 
261 		y := (y+x/y)/2		... almost 17 sig. bits
262 		y := (y+x/y)/2		... almost 35 sig. bits
263 		y := y-(y-x/y)/2	... within 1 ulp
264 
265 
266 	Remark 1.
267 	    Another way to improve y to within 1 ulp is:
268 
269 		y := (y+x/y)		... almost 17 sig. bits to 2*sqrt(x)
270 		y := y - 0x00100006	... almost 18 sig. bits to sqrt(x)
271 
272 				2
273 			    (x-y )*y
274 		y := y + 2* ----------	...within 1 ulp
275 			       2
276 			     3y  + x
277 
278 
279 	This formula has one division fewer than the one above; however,
280 	it requires more multiplications and additions. Also x must be
281 	scaled in advance to avoid spurious overflow in evaluating the
282 	expression 3y*y+x. Hence it is not recommended uless division
283 	is slow. If division is very slow, then one should use the
284 	reciproot algorithm given in section B.
285 
286     (3) Final adjustment
287 
288 	By twiddling y's last bit it is possible to force y to be
289 	correctly rounded according to the prevailing rounding mode
290 	as follows. Let r and i be copies of the rounding mode and
291 	inexact flag before entering the square root program. Also we
292 	use the expression y+-ulp for the next representable floating
293 	numbers (up and down) of y. Note that y+-ulp = either fixed
294 	point y+-1, or multiply y by nextafter(1,+-inf) in chopped
295 	mode.
296 
297 		I := FALSE;	... reset INEXACT flag I
298 		R := RZ;	... set rounding mode to round-toward-zero
299 		z := x/y;	... chopped quotient, possibly inexact
300 		If(not I) then {	... if the quotient is exact
301 		    if(z=y) {
302 		        I := i;	 ... restore inexact flag
303 		        R := r;  ... restore rounded mode
304 		        return sqrt(x):=y.
305 		    } else {
306 			z := z - ulp;	... special rounding
307 		    }
308 		}
309 		i := TRUE;		... sqrt(x) is inexact
310 		If (r=RN) then z=z+ulp	... rounded-to-nearest
311 		If (r=RP) then {	... round-toward-+inf
312 		    y = y+ulp; z=z+ulp;
313 		}
314 		y := y+z;		... chopped sum
315 		y0:=y0-0x00100000;	... y := y/2 is correctly rounded.
316 	        I := i;	 		... restore inexact flag
317 	        R := r;  		... restore rounded mode
318 	        return sqrt(x):=y.
319 
320     (4)	Special cases
321 
322 	Square root of +inf, +-0, or NaN is itself;
323 	Square root of a negative number is NaN with invalid signal.
324 
325 
326 B.  sqrt(x) by Reciproot Iteration
327 
328    (1)	Initial approximation
329 
330 	Let x0 and x1 be the leading and the trailing 32-bit words of
331 	a floating point number x (in IEEE double format) respectively
332 	(see section A). By performing shifs and subtracts on x0 and y0,
333 	we obtain a 7.8-bit approximation of 1/sqrt(x) as follows.
334 
335 	    k := 0x5fe80000 - (x0>>1);
336 	    y0:= k - T2[63&(k>>14)].	... y ~ 1/sqrt(x) to 7.8 bits
337 
338 	Here k is a 32-bit integer and T2[] is an integer array
339 	containing correction terms. Now magically the floating
340 	value of y (y's leading 32-bit word is y0, the value of
341 	its trailing word y1 is set to zero) approximates 1/sqrt(x)
342 	to almost 7.8-bit.
343 
344 	Value of T2:
345 	static int T2[64]= {
346 	0x1500,	0x2ef8,	0x4d67,	0x6b02,	0x87be,	0xa395,	0xbe7a,	0xd866,
347 	0xf14a,	0x1091b,0x11fcd,0x13552,0x14999,0x15c98,0x16e34,0x17e5f,
348 	0x18d03,0x19a01,0x1a545,0x1ae8a,0x1b5c4,0x1bb01,0x1bfde,0x1c28d,
349 	0x1c2de,0x1c0db,0x1ba73,0x1b11c,0x1a4b5,0x1953d,0x18266,0x16be0,
350 	0x1683e,0x179d8,0x18a4d,0x19992,0x1a789,0x1b445,0x1bf61,0x1c989,
351 	0x1d16d,0x1d77b,0x1dddf,0x1e2ad,0x1e5bf,0x1e6e8,0x1e654,0x1e3cd,
352 	0x1df2a,0x1d635,0x1cb16,0x1be2c,0x1ae4e,0x19bde,0x1868e,0x16e2e,
353 	0x1527f,0x1334a,0x11051,0xe951,	0xbe01,	0x8e0d,	0x5924,	0x1edd,};
354 
355     (2)	Iterative refinement
356 
357 	Apply Reciproot iteration three times to y and multiply the
358 	result by x to get an approximation z that matches sqrt(x)
359 	to about 1 ulp. To be exact, we will have
360 		-1ulp < sqrt(x)-z<1.0625ulp.
361 
362 	... set rounding mode to Round-to-nearest
363 	   y := y*(1.5-0.5*x*y*y)	... almost 15 sig. bits to 1/sqrt(x)
364 	   y := y*((1.5-2^-30)+0.5*x*y*y)... about 29 sig. bits to 1/sqrt(x)
365 	... special arrangement for better accuracy
366 	   z := x*y			... 29 bits to sqrt(x), with z*y<1
367 	   z := z + 0.5*z*(1-z*y)	... about 1 ulp to sqrt(x)
368 
369 	Remark 2. The constant 1.5-2^-30 is chosen to bias the error so that
370 	(a) the term z*y in the final iteration is always less than 1;
371 	(b) the error in the final result is biased upward so that
372 		-1 ulp < sqrt(x) - z < 1.0625 ulp
373 	    instead of |sqrt(x)-z|<1.03125ulp.
374 
375     (3)	Final adjustment
376 
377 	By twiddling y's last bit it is possible to force y to be
378 	correctly rounded according to the prevailing rounding mode
379 	as follows. Let r and i be copies of the rounding mode and
380 	inexact flag before entering the square root program. Also we
381 	use the expression y+-ulp for the next representable floating
382 	numbers (up and down) of y. Note that y+-ulp = either fixed
383 	point y+-1, or multiply y by nextafter(1,+-inf) in chopped
384 	mode.
385 
386 	R := RZ;		... set rounding mode to round-toward-zero
387 	switch(r) {
388 	    case RN:		... round-to-nearest
389 	       if(x<= z*(z-ulp)...chopped) z = z - ulp; else
390 	       if(x<= z*(z+ulp)...chopped) z = z; else z = z+ulp;
391 	       break;
392 	    case RZ:case RM:	... round-to-zero or round-to--inf
393 	       R:=RP;		... reset rounding mod to round-to-+inf
394 	       if(x<z*z ... rounded up) z = z - ulp; else
395 	       if(x>=(z+ulp)*(z+ulp) ...rounded up) z = z+ulp;
396 	       break;
397 	    case RP:		... round-to-+inf
398 	       if(x>(z+ulp)*(z+ulp)...chopped) z = z+2*ulp; else
399 	       if(x>z*z ...chopped) z = z+ulp;
400 	       break;
401 	}
402 
403 	Remark 3. The above comparisons can be done in fixed point. For
404 	example, to compare x and w=z*z chopped, it suffices to compare
405 	x1 and w1 (the trailing parts of x and w), regarding them as
406 	two's complement integers.
407 
408 	...Is z an exact square root?
409 	To determine whether z is an exact square root of x, let z1 be the
410 	trailing part of z, and also let x0 and x1 be the leading and
411 	trailing parts of x.
412 
413 	If ((z1&0x03ffffff)!=0)	... not exact if trailing 26 bits of z!=0
414 	    I := 1;		... Raise Inexact flag: z is not exact
415 	else {
416 	    j := 1 - [(x0>>20)&1]	... j = logb(x) mod 2
417 	    k := z1 >> 26;		... get z's 25-th and 26-th
418 					    fraction bits
419 	    I := i or (k&j) or ((k&(j+j+1))!=(x1&3));
420 	}
421 	R:= r		... restore rounded mode
422 	return sqrt(x):=z.
423 
424 	If multiplication is cheaper then the foregoing red tape, the
425 	Inexact flag can be evaluated by
426 
427 	    I := i;
428 	    I := (z*z!=x) or I.
429 
430 	Note that z*z can overwrite I; this value must be sensed if it is
431 	True.
432 
433 	Remark 4. If z*z = x exactly, then bit 25 to bit 0 of z1 must be
434 	zero.
435 
436 		    --------------------
437 		z1: |        f2        |
438 		    --------------------
439 		bit 31		   bit 0
440 
441 	Further more, bit 27 and 26 of z1, bit 0 and 1 of x1, and the odd
442 	or even of logb(x) have the following relations:
443 
444 	-------------------------------------------------
445 	bit 27,26 of z1		bit 1,0 of x1	logb(x)
446 	-------------------------------------------------
447 	00			00		odd and even
448 	01			01		even
449 	10			10		odd
450 	10			00		even
451 	11			01		even
452 	-------------------------------------------------
453 
454     (4)	Special cases (see (4) of Section A).
455 
456  */
457 
458