xref: /freebsd/lib/msun/src/e_sqrt.c (revision c19fb1f963e3dc88a82b20d1b17f94a4cd321e74)
1 
2 /*
3  * ====================================================
4  * Copyright (C) 1993 by Sun Microsystems, Inc. All rights reserved.
5  *
6  * Developed at SunSoft, 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 
13 #include <sys/cdefs.h>
14 #include <float.h>
15 
16 #include "math.h"
17 #include "math_private.h"
18 
19 #ifdef USE_BUILTIN_SQRT
20 double
21 sqrt(double x)
22 {
23 	return (__builtin_sqrt(x));
24 }
25 #else
26 /* sqrt(x)
27  * Return correctly rounded sqrt.
28  *           ------------------------------------------
29  *	     |  Use the hardware sqrt if you have one |
30  *           ------------------------------------------
31  * Method:
32  *   Bit by bit method using integer arithmetic. (Slow, but portable)
33  *   1. Normalization
34  *	Scale x to y in [1,4) with even powers of 2:
35  *	find an integer k such that  1 <= (y=x*2^(2k)) < 4, then
36  *		sqrt(x) = 2^k * sqrt(y)
37  *   2. Bit by bit computation
38  *	Let q  = sqrt(y) truncated to i bit after binary point (q = 1),
39  *	     i							 0
40  *                                     i+1         2
41  *	    s  = 2*q , and	y  =  2   * ( y - q  ).		(1)
42  *	     i      i            i                 i
43  *
44  *	To compute q    from q , one checks whether
45  *		    i+1       i
46  *
47  *			      -(i+1) 2
48  *			(q + 2      ) <= y.			(2)
49  *     			  i
50  *							      -(i+1)
51  *	If (2) is false, then q   = q ; otherwise q   = q  + 2      .
52  *		 	       i+1   i             i+1   i
53  *
54  *	With some algebric manipulation, it is not difficult to see
55  *	that (2) is equivalent to
56  *                             -(i+1)
57  *			s  +  2       <= y			(3)
58  *			 i                i
59  *
60  *	The advantage of (3) is that s  and y  can be computed by
61  *				      i      i
62  *	the following recurrence formula:
63  *	    if (3) is false
64  *
65  *	    s     =  s  ,	y    = y   ;			(4)
66  *	     i+1      i		 i+1    i
67  *
68  *	    otherwise,
69  *                         -i                     -(i+1)
70  *	    s	  =  s  + 2  ,  y    = y  -  s  - 2  		(5)
71  *           i+1      i          i+1    i     i
72  *
73  *	One may easily use induction to prove (4) and (5).
74  *	Note. Since the left hand side of (3) contain only i+2 bits,
75  *	      it does not necessary to do a full (53-bit) comparison
76  *	      in (3).
77  *   3. Final rounding
78  *	After generating the 53 bits result, we compute one more bit.
79  *	Together with the remainder, we can decide whether the
80  *	result is exact, bigger than 1/2ulp, or less than 1/2ulp
81  *	(it will never equal to 1/2ulp).
82  *	The rounding mode can be detected by checking whether
83  *	huge + tiny is equal to huge, and whether huge - tiny is
84  *	equal to huge for some floating point number "huge" and "tiny".
85  *
86  * Special cases:
87  *	sqrt(+-0) = +-0 	... exact
88  *	sqrt(inf) = inf
89  *	sqrt(-ve) = NaN		... with invalid signal
90  *	sqrt(NaN) = NaN		... with invalid signal for signaling NaN
91  *
92  * Other methods : see the appended file at the end of the program below.
93  *---------------
94  */
95 
96 static	const double	one	= 1.0, tiny=1.0e-300;
97 
98 double
99 sqrt(double x)
100 {
101 	double z;
102 	int32_t sign = (int)0x80000000;
103 	int32_t ix0,s0,q,m,t,i;
104 	u_int32_t r,t1,s1,ix1,q1;
105 
106 	EXTRACT_WORDS(ix0,ix1,x);
107 
108     /* take care of Inf and NaN */
109 	if((ix0&0x7ff00000)==0x7ff00000) {
110 	    return x*x+x;		/* sqrt(NaN)=NaN, sqrt(+inf)=+inf
111 					   sqrt(-inf)=sNaN */
112 	}
113     /* take care of zero */
114 	if(ix0<=0) {
115 	    if(((ix0&(~sign))|ix1)==0) return x;/* sqrt(+-0) = +-0 */
116 	    else if(ix0<0)
117 		return (x-x)/(x-x);		/* sqrt(-ve) = sNaN */
118 	}
119     /* normalize x */
120 	m = (ix0>>20);
121 	if(m==0) {				/* subnormal x */
122 	    while(ix0==0) {
123 		m -= 21;
124 		ix0 |= (ix1>>11); ix1 <<= 21;
125 	    }
126 	    for(i=0;(ix0&0x00100000)==0;i++) ix0<<=1;
127 	    m -= i-1;
128 	    ix0 |= (ix1>>(32-i));
129 	    ix1 <<= i;
130 	}
131 	m -= 1023;	/* unbias exponent */
132 	ix0 = (ix0&0x000fffff)|0x00100000;
133 	if(m&1){	/* odd m, double x to make it even */
134 	    ix0 += ix0 + ((ix1&sign)>>31);
135 	    ix1 += ix1;
136 	}
137 	m >>= 1;	/* m = [m/2] */
138 
139     /* generate sqrt(x) bit by bit */
140 	ix0 += ix0 + ((ix1&sign)>>31);
141 	ix1 += ix1;
142 	q = q1 = s0 = s1 = 0;	/* [q,q1] = sqrt(x) */
143 	r = 0x00200000;		/* r = moving bit from right to left */
144 
145 	while(r!=0) {
146 	    t = s0+r;
147 	    if(t<=ix0) {
148 		s0   = t+r;
149 		ix0 -= t;
150 		q   += r;
151 	    }
152 	    ix0 += ix0 + ((ix1&sign)>>31);
153 	    ix1 += ix1;
154 	    r>>=1;
155 	}
156 
157 	r = sign;
158 	while(r!=0) {
159 	    t1 = s1+r;
160 	    t  = s0;
161 	    if((t<ix0)||((t==ix0)&&(t1<=ix1))) {
162 		s1  = t1+r;
163 		if(((t1&sign)==sign)&&(s1&sign)==0) s0 += 1;
164 		ix0 -= t;
165 		if (ix1 < t1) ix0 -= 1;
166 		ix1 -= t1;
167 		q1  += r;
168 	    }
169 	    ix0 += ix0 + ((ix1&sign)>>31);
170 	    ix1 += ix1;
171 	    r>>=1;
172 	}
173 
174     /* use floating add to find out rounding direction */
175 	if((ix0|ix1)!=0) {
176 	    z = one-tiny; /* trigger inexact flag */
177 	    if (z>=one) {
178 	        z = one+tiny;
179 	        if (q1==(u_int32_t)0xffffffff) { q1=0; q += 1;}
180 		else if (z>one) {
181 		    if (q1==(u_int32_t)0xfffffffe) q+=1;
182 		    q1+=2;
183 		} else
184 	            q1 += (q1&1);
185 	    }
186 	}
187 	ix0 = (q>>1)+0x3fe00000;
188 	ix1 =  q1>>1;
189 	if ((q&1)==1) ix1 |= sign;
190 	ix0 += (m <<20);
191 	INSERT_WORDS(z,ix0,ix1);
192 	return z;
193 }
194 #endif
195 
196 #if (LDBL_MANT_DIG == 53)
197 __weak_reference(sqrt, sqrtl);
198 #endif
199 
200 /*
201 Other methods  (use floating-point arithmetic)
202 -------------
203 (This is a copy of a drafted paper by Prof W. Kahan
204 and K.C. Ng, written in May, 1986)
205 
206 	Two algorithms are given here to implement sqrt(x)
207 	(IEEE double precision arithmetic) in software.
208 	Both supply sqrt(x) correctly rounded. The first algorithm (in
209 	Section A) uses newton iterations and involves four divisions.
210 	The second one uses reciproot iterations to avoid division, but
211 	requires more multiplications. Both algorithms need the ability
212 	to chop results of arithmetic operations instead of round them,
213 	and the INEXACT flag to indicate when an arithmetic operation
214 	is executed exactly with no roundoff error, all part of the
215 	standard (IEEE 754-1985). The ability to perform shift, add,
216 	subtract and logical AND operations upon 32-bit words is needed
217 	too, though not part of the standard.
218 
219 A.  sqrt(x) by Newton Iteration
220 
221    (1)	Initial approximation
222 
223 	Let x0 and x1 be the leading and the trailing 32-bit words of
224 	a floating point number x (in IEEE double format) respectively
225 
226 	    1    11		     52				  ...widths
227 	   ------------------------------------------------------
228 	x: |s|	  e     |	      f				|
229 	   ------------------------------------------------------
230 	      msb    lsb  msb				      lsb ...order
231 
232 
233 	     ------------------------  	     ------------------------
234 	x0:  |s|   e    |    f1     |	 x1: |          f2           |
235 	     ------------------------  	     ------------------------
236 
237 	By performing shifts and subtracts on x0 and x1 (both regarded
238 	as integers), we obtain an 8-bit approximation of sqrt(x) as
239 	follows.
240 
241 		k  := (x0>>1) + 0x1ff80000;
242 		y0 := k - T1[31&(k>>15)].	... y ~ sqrt(x) to 8 bits
243 	Here k is a 32-bit integer and T1[] is an integer array containing
244 	correction terms. Now magically the floating value of y (y's
245 	leading 32-bit word is y0, the value of its trailing word is 0)
246 	approximates sqrt(x) to almost 8-bit.
247 
248 	Value of T1:
249 	static int T1[32]= {
250 	0,	1024,	3062,	5746,	9193,	13348,	18162,	23592,
251 	29598,	36145,	43202,	50740,	58733,	67158,	75992,	85215,
252 	83599,	71378,	60428,	50647,	41945,	34246,	27478,	21581,
253 	16499,	12183,	8588,	5674,	3403,	1742,	661,	130,};
254 
255     (2)	Iterative refinement
256 
257 	Apply Heron's rule three times to y, we have y approximates
258 	sqrt(x) to within 1 ulp (Unit in the Last Place):
259 
260 		y := (y+x/y)/2		... almost 17 sig. bits
261 		y := (y+x/y)/2		... almost 35 sig. bits
262 		y := y-(y-x/y)/2	... within 1 ulp
263 
264 
265 	Remark 1.
266 	    Another way to improve y to within 1 ulp is:
267 
268 		y := (y+x/y)		... almost 17 sig. bits to 2*sqrt(x)
269 		y := y - 0x00100006	... almost 18 sig. bits to sqrt(x)
270 
271 				2
272 			    (x-y )*y
273 		y := y + 2* ----------	...within 1 ulp
274 			       2
275 			     3y  + x
276 
277 
278 	This formula has one division fewer than the one above; however,
279 	it requires more multiplications and additions. Also x must be
280 	scaled in advance to avoid spurious overflow in evaluating the
281 	expression 3y*y+x. Hence it is not recommended uless division
282 	is slow. If division is very slow, then one should use the
283 	reciproot algorithm given in section B.
284 
285     (3) Final adjustment
286 
287 	By twiddling y's last bit it is possible to force y to be
288 	correctly rounded according to the prevailing rounding mode
289 	as follows. Let r and i be copies of the rounding mode and
290 	inexact flag before entering the square root program. Also we
291 	use the expression y+-ulp for the next representable floating
292 	numbers (up and down) of y. Note that y+-ulp = either fixed
293 	point y+-1, or multiply y by nextafter(1,+-inf) in chopped
294 	mode.
295 
296 		I := FALSE;	... reset INEXACT flag I
297 		R := RZ;	... set rounding mode to round-toward-zero
298 		z := x/y;	... chopped quotient, possibly inexact
299 		If(not I) then {	... if the quotient is exact
300 		    if(z=y) {
301 		        I := i;	 ... restore inexact flag
302 		        R := r;  ... restore rounded mode
303 		        return sqrt(x):=y.
304 		    } else {
305 			z := z - ulp;	... special rounding
306 		    }
307 		}
308 		i := TRUE;		... sqrt(x) is inexact
309 		If (r=RN) then z=z+ulp	... rounded-to-nearest
310 		If (r=RP) then {	... round-toward-+inf
311 		    y = y+ulp; z=z+ulp;
312 		}
313 		y := y+z;		... chopped sum
314 		y0:=y0-0x00100000;	... y := y/2 is correctly rounded.
315 	        I := i;	 		... restore inexact flag
316 	        R := r;  		... restore rounded mode
317 	        return sqrt(x):=y.
318 
319     (4)	Special cases
320 
321 	Square root of +inf, +-0, or NaN is itself;
322 	Square root of a negative number is NaN with invalid signal.
323 
324 
325 B.  sqrt(x) by Reciproot Iteration
326 
327    (1)	Initial approximation
328 
329 	Let x0 and x1 be the leading and the trailing 32-bit words of
330 	a floating point number x (in IEEE double format) respectively
331 	(see section A). By performing shifs and subtracts on x0 and y0,
332 	we obtain a 7.8-bit approximation of 1/sqrt(x) as follows.
333 
334 	    k := 0x5fe80000 - (x0>>1);
335 	    y0:= k - T2[63&(k>>14)].	... y ~ 1/sqrt(x) to 7.8 bits
336 
337 	Here k is a 32-bit integer and T2[] is an integer array
338 	containing correction terms. Now magically the floating
339 	value of y (y's leading 32-bit word is y0, the value of
340 	its trailing word y1 is set to zero) approximates 1/sqrt(x)
341 	to almost 7.8-bit.
342 
343 	Value of T2:
344 	static int T2[64]= {
345 	0x1500,	0x2ef8,	0x4d67,	0x6b02,	0x87be,	0xa395,	0xbe7a,	0xd866,
346 	0xf14a,	0x1091b,0x11fcd,0x13552,0x14999,0x15c98,0x16e34,0x17e5f,
347 	0x18d03,0x19a01,0x1a545,0x1ae8a,0x1b5c4,0x1bb01,0x1bfde,0x1c28d,
348 	0x1c2de,0x1c0db,0x1ba73,0x1b11c,0x1a4b5,0x1953d,0x18266,0x16be0,
349 	0x1683e,0x179d8,0x18a4d,0x19992,0x1a789,0x1b445,0x1bf61,0x1c989,
350 	0x1d16d,0x1d77b,0x1dddf,0x1e2ad,0x1e5bf,0x1e6e8,0x1e654,0x1e3cd,
351 	0x1df2a,0x1d635,0x1cb16,0x1be2c,0x1ae4e,0x19bde,0x1868e,0x16e2e,
352 	0x1527f,0x1334a,0x11051,0xe951,	0xbe01,	0x8e0d,	0x5924,	0x1edd,};
353 
354     (2)	Iterative refinement
355 
356 	Apply Reciproot iteration three times to y and multiply the
357 	result by x to get an approximation z that matches sqrt(x)
358 	to about 1 ulp. To be exact, we will have
359 		-1ulp < sqrt(x)-z<1.0625ulp.
360 
361 	... set rounding mode to Round-to-nearest
362 	   y := y*(1.5-0.5*x*y*y)	... almost 15 sig. bits to 1/sqrt(x)
363 	   y := y*((1.5-2^-30)+0.5*x*y*y)... about 29 sig. bits to 1/sqrt(x)
364 	... special arrangement for better accuracy
365 	   z := x*y			... 29 bits to sqrt(x), with z*y<1
366 	   z := z + 0.5*z*(1-z*y)	... about 1 ulp to sqrt(x)
367 
368 	Remark 2. The constant 1.5-2^-30 is chosen to bias the error so that
369 	(a) the term z*y in the final iteration is always less than 1;
370 	(b) the error in the final result is biased upward so that
371 		-1 ulp < sqrt(x) - z < 1.0625 ulp
372 	    instead of |sqrt(x)-z|<1.03125ulp.
373 
374     (3)	Final adjustment
375 
376 	By twiddling y's last bit it is possible to force y to be
377 	correctly rounded according to the prevailing rounding mode
378 	as follows. Let r and i be copies of the rounding mode and
379 	inexact flag before entering the square root program. Also we
380 	use the expression y+-ulp for the next representable floating
381 	numbers (up and down) of y. Note that y+-ulp = either fixed
382 	point y+-1, or multiply y by nextafter(1,+-inf) in chopped
383 	mode.
384 
385 	R := RZ;		... set rounding mode to round-toward-zero
386 	switch(r) {
387 	    case RN:		... round-to-nearest
388 	       if(x<= z*(z-ulp)...chopped) z = z - ulp; else
389 	       if(x<= z*(z+ulp)...chopped) z = z; else z = z+ulp;
390 	       break;
391 	    case RZ:case RM:	... round-to-zero or round-to--inf
392 	       R:=RP;		... reset rounding mod to round-to-+inf
393 	       if(x<z*z ... rounded up) z = z - ulp; else
394 	       if(x>=(z+ulp)*(z+ulp) ...rounded up) z = z+ulp;
395 	       break;
396 	    case RP:		... round-to-+inf
397 	       if(x>(z+ulp)*(z+ulp)...chopped) z = z+2*ulp; else
398 	       if(x>z*z ...chopped) z = z+ulp;
399 	       break;
400 	}
401 
402 	Remark 3. The above comparisons can be done in fixed point. For
403 	example, to compare x and w=z*z chopped, it suffices to compare
404 	x1 and w1 (the trailing parts of x and w), regarding them as
405 	two's complement integers.
406 
407 	...Is z an exact square root?
408 	To determine whether z is an exact square root of x, let z1 be the
409 	trailing part of z, and also let x0 and x1 be the leading and
410 	trailing parts of x.
411 
412 	If ((z1&0x03ffffff)!=0)	... not exact if trailing 26 bits of z!=0
413 	    I := 1;		... Raise Inexact flag: z is not exact
414 	else {
415 	    j := 1 - [(x0>>20)&1]	... j = logb(x) mod 2
416 	    k := z1 >> 26;		... get z's 25-th and 26-th
417 					    fraction bits
418 	    I := i or (k&j) or ((k&(j+j+1))!=(x1&3));
419 	}
420 	R:= r		... restore rounded mode
421 	return sqrt(x):=z.
422 
423 	If multiplication is cheaper then the foregoing red tape, the
424 	Inexact flag can be evaluated by
425 
426 	    I := i;
427 	    I := (z*z!=x) or I.
428 
429 	Note that z*z can overwrite I; this value must be sensed if it is
430 	True.
431 
432 	Remark 4. If z*z = x exactly, then bit 25 to bit 0 of z1 must be
433 	zero.
434 
435 		    --------------------
436 		z1: |        f2        |
437 		    --------------------
438 		bit 31		   bit 0
439 
440 	Further more, bit 27 and 26 of z1, bit 0 and 1 of x1, and the odd
441 	or even of logb(x) have the following relations:
442 
443 	-------------------------------------------------
444 	bit 27,26 of z1		bit 1,0 of x1	logb(x)
445 	-------------------------------------------------
446 	00			00		odd and even
447 	01			01		even
448 	10			10		odd
449 	10			00		even
450 	11			01		even
451 	-------------------------------------------------
452 
453     (4)	Special cases (see (4) of Section A).
454 
455  */
456 
457