xref: /freebsd/lib/msun/src/e_sqrt.c (revision 681ce946)
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 __FBSDID("$FreeBSD$");
16 
17 #include <float.h>
18 
19 #include "math.h"
20 #include "math_private.h"
21 
22 #ifdef USE_BUILTIN_SQRT
23 double
24 __ieee754_sqrt(double x)
25 {
26 	return (__builtin_sqrt(x));
27 }
28 #else
29 /* __ieee754_sqrt(x)
30  * Return correctly rounded sqrt.
31  *           ------------------------------------------
32  *	     |  Use the hardware sqrt if you have one |
33  *           ------------------------------------------
34  * Method:
35  *   Bit by bit method using integer arithmetic. (Slow, but portable)
36  *   1. Normalization
37  *	Scale x to y in [1,4) with even powers of 2:
38  *	find an integer k such that  1 <= (y=x*2^(2k)) < 4, then
39  *		sqrt(x) = 2^k * sqrt(y)
40  *   2. Bit by bit computation
41  *	Let q  = sqrt(y) truncated to i bit after binary point (q = 1),
42  *	     i							 0
43  *                                     i+1         2
44  *	    s  = 2*q , and	y  =  2   * ( y - q  ).		(1)
45  *	     i      i            i                 i
46  *
47  *	To compute q    from q , one checks whether
48  *		    i+1       i
49  *
50  *			      -(i+1) 2
51  *			(q + 2      ) <= y.			(2)
52  *     			  i
53  *							      -(i+1)
54  *	If (2) is false, then q   = q ; otherwise q   = q  + 2      .
55  *		 	       i+1   i             i+1   i
56  *
57  *	With some algebric manipulation, it is not difficult to see
58  *	that (2) is equivalent to
59  *                             -(i+1)
60  *			s  +  2       <= y			(3)
61  *			 i                i
62  *
63  *	The advantage of (3) is that s  and y  can be computed by
64  *				      i      i
65  *	the following recurrence formula:
66  *	    if (3) is false
67  *
68  *	    s     =  s  ,	y    = y   ;			(4)
69  *	     i+1      i		 i+1    i
70  *
71  *	    otherwise,
72  *                         -i                     -(i+1)
73  *	    s	  =  s  + 2  ,  y    = y  -  s  - 2  		(5)
74  *           i+1      i          i+1    i     i
75  *
76  *	One may easily use induction to prove (4) and (5).
77  *	Note. Since the left hand side of (3) contain only i+2 bits,
78  *	      it does not necessary to do a full (53-bit) comparison
79  *	      in (3).
80  *   3. Final rounding
81  *	After generating the 53 bits result, we compute one more bit.
82  *	Together with the remainder, we can decide whether the
83  *	result is exact, bigger than 1/2ulp, or less than 1/2ulp
84  *	(it will never equal to 1/2ulp).
85  *	The rounding mode can be detected by checking whether
86  *	huge + tiny is equal to huge, and whether huge - tiny is
87  *	equal to huge for some floating point number "huge" and "tiny".
88  *
89  * Special cases:
90  *	sqrt(+-0) = +-0 	... exact
91  *	sqrt(inf) = inf
92  *	sqrt(-ve) = NaN		... with invalid signal
93  *	sqrt(NaN) = NaN		... with invalid signal for signaling NaN
94  *
95  * Other methods : see the appended file at the end of the program below.
96  *---------------
97  */
98 
99 static	const double	one	= 1.0, tiny=1.0e-300;
100 
101 double
102 __ieee754_sqrt(double x)
103 {
104 	double z;
105 	int32_t sign = (int)0x80000000;
106 	int32_t ix0,s0,q,m,t,i;
107 	u_int32_t r,t1,s1,ix1,q1;
108 
109 	EXTRACT_WORDS(ix0,ix1,x);
110 
111     /* take care of Inf and NaN */
112 	if((ix0&0x7ff00000)==0x7ff00000) {
113 	    return x*x+x;		/* sqrt(NaN)=NaN, sqrt(+inf)=+inf
114 					   sqrt(-inf)=sNaN */
115 	}
116     /* take care of zero */
117 	if(ix0<=0) {
118 	    if(((ix0&(~sign))|ix1)==0) return x;/* sqrt(+-0) = +-0 */
119 	    else if(ix0<0)
120 		return (x-x)/(x-x);		/* sqrt(-ve) = sNaN */
121 	}
122     /* normalize x */
123 	m = (ix0>>20);
124 	if(m==0) {				/* subnormal x */
125 	    while(ix0==0) {
126 		m -= 21;
127 		ix0 |= (ix1>>11); ix1 <<= 21;
128 	    }
129 	    for(i=0;(ix0&0x00100000)==0;i++) ix0<<=1;
130 	    m -= i-1;
131 	    ix0 |= (ix1>>(32-i));
132 	    ix1 <<= i;
133 	}
134 	m -= 1023;	/* unbias exponent */
135 	ix0 = (ix0&0x000fffff)|0x00100000;
136 	if(m&1){	/* odd m, double x to make it even */
137 	    ix0 += ix0 + ((ix1&sign)>>31);
138 	    ix1 += ix1;
139 	}
140 	m >>= 1;	/* m = [m/2] */
141 
142     /* generate sqrt(x) bit by bit */
143 	ix0 += ix0 + ((ix1&sign)>>31);
144 	ix1 += ix1;
145 	q = q1 = s0 = s1 = 0;	/* [q,q1] = sqrt(x) */
146 	r = 0x00200000;		/* r = moving bit from right to left */
147 
148 	while(r!=0) {
149 	    t = s0+r;
150 	    if(t<=ix0) {
151 		s0   = t+r;
152 		ix0 -= t;
153 		q   += r;
154 	    }
155 	    ix0 += ix0 + ((ix1&sign)>>31);
156 	    ix1 += ix1;
157 	    r>>=1;
158 	}
159 
160 	r = sign;
161 	while(r!=0) {
162 	    t1 = s1+r;
163 	    t  = s0;
164 	    if((t<ix0)||((t==ix0)&&(t1<=ix1))) {
165 		s1  = t1+r;
166 		if(((t1&sign)==sign)&&(s1&sign)==0) s0 += 1;
167 		ix0 -= t;
168 		if (ix1 < t1) ix0 -= 1;
169 		ix1 -= t1;
170 		q1  += r;
171 	    }
172 	    ix0 += ix0 + ((ix1&sign)>>31);
173 	    ix1 += ix1;
174 	    r>>=1;
175 	}
176 
177     /* use floating add to find out rounding direction */
178 	if((ix0|ix1)!=0) {
179 	    z = one-tiny; /* trigger inexact flag */
180 	    if (z>=one) {
181 	        z = one+tiny;
182 	        if (q1==(u_int32_t)0xffffffff) { q1=0; q += 1;}
183 		else if (z>one) {
184 		    if (q1==(u_int32_t)0xfffffffe) q+=1;
185 		    q1+=2;
186 		} else
187 	            q1 += (q1&1);
188 	    }
189 	}
190 	ix0 = (q>>1)+0x3fe00000;
191 	ix1 =  q1>>1;
192 	if ((q&1)==1) ix1 |= sign;
193 	ix0 += (m <<20);
194 	INSERT_WORDS(z,ix0,ix1);
195 	return z;
196 }
197 #endif
198 
199 #if (LDBL_MANT_DIG == 53)
200 __weak_reference(sqrt, sqrtl);
201 #endif
202 
203 /*
204 Other methods  (use floating-point arithmetic)
205 -------------
206 (This is a copy of a drafted paper by Prof W. Kahan
207 and K.C. Ng, written in May, 1986)
208 
209 	Two algorithms are given here to implement sqrt(x)
210 	(IEEE double precision arithmetic) in software.
211 	Both supply sqrt(x) correctly rounded. The first algorithm (in
212 	Section A) uses newton iterations and involves four divisions.
213 	The second one uses reciproot iterations to avoid division, but
214 	requires more multiplications. Both algorithms need the ability
215 	to chop results of arithmetic operations instead of round them,
216 	and the INEXACT flag to indicate when an arithmetic operation
217 	is executed exactly with no roundoff error, all part of the
218 	standard (IEEE 754-1985). The ability to perform shift, add,
219 	subtract and logical AND operations upon 32-bit words is needed
220 	too, though not part of the standard.
221 
222 A.  sqrt(x) by Newton Iteration
223 
224    (1)	Initial approximation
225 
226 	Let x0 and x1 be the leading and the trailing 32-bit words of
227 	a floating point number x (in IEEE double format) respectively
228 
229 	    1    11		     52				  ...widths
230 	   ------------------------------------------------------
231 	x: |s|	  e     |	      f				|
232 	   ------------------------------------------------------
233 	      msb    lsb  msb				      lsb ...order
234 
235 
236 	     ------------------------  	     ------------------------
237 	x0:  |s|   e    |    f1     |	 x1: |          f2           |
238 	     ------------------------  	     ------------------------
239 
240 	By performing shifts and subtracts on x0 and x1 (both regarded
241 	as integers), we obtain an 8-bit approximation of sqrt(x) as
242 	follows.
243 
244 		k  := (x0>>1) + 0x1ff80000;
245 		y0 := k - T1[31&(k>>15)].	... y ~ sqrt(x) to 8 bits
246 	Here k is a 32-bit integer and T1[] is an integer array containing
247 	correction terms. Now magically the floating value of y (y's
248 	leading 32-bit word is y0, the value of its trailing word is 0)
249 	approximates sqrt(x) to almost 8-bit.
250 
251 	Value of T1:
252 	static int T1[32]= {
253 	0,	1024,	3062,	5746,	9193,	13348,	18162,	23592,
254 	29598,	36145,	43202,	50740,	58733,	67158,	75992,	85215,
255 	83599,	71378,	60428,	50647,	41945,	34246,	27478,	21581,
256 	16499,	12183,	8588,	5674,	3403,	1742,	661,	130,};
257 
258     (2)	Iterative refinement
259 
260 	Apply Heron's rule three times to y, we have y approximates
261 	sqrt(x) to within 1 ulp (Unit in the Last Place):
262 
263 		y := (y+x/y)/2		... almost 17 sig. bits
264 		y := (y+x/y)/2		... almost 35 sig. bits
265 		y := y-(y-x/y)/2	... within 1 ulp
266 
267 
268 	Remark 1.
269 	    Another way to improve y to within 1 ulp is:
270 
271 		y := (y+x/y)		... almost 17 sig. bits to 2*sqrt(x)
272 		y := y - 0x00100006	... almost 18 sig. bits to sqrt(x)
273 
274 				2
275 			    (x-y )*y
276 		y := y + 2* ----------	...within 1 ulp
277 			       2
278 			     3y  + x
279 
280 
281 	This formula has one division fewer than the one above; however,
282 	it requires more multiplications and additions. Also x must be
283 	scaled in advance to avoid spurious overflow in evaluating the
284 	expression 3y*y+x. Hence it is not recommended uless division
285 	is slow. If division is very slow, then one should use the
286 	reciproot algorithm given in section B.
287 
288     (3) Final adjustment
289 
290 	By twiddling y's last bit it is possible to force y to be
291 	correctly rounded according to the prevailing rounding mode
292 	as follows. Let r and i be copies of the rounding mode and
293 	inexact flag before entering the square root program. Also we
294 	use the expression y+-ulp for the next representable floating
295 	numbers (up and down) of y. Note that y+-ulp = either fixed
296 	point y+-1, or multiply y by nextafter(1,+-inf) in chopped
297 	mode.
298 
299 		I := FALSE;	... reset INEXACT flag I
300 		R := RZ;	... set rounding mode to round-toward-zero
301 		z := x/y;	... chopped quotient, possibly inexact
302 		If(not I) then {	... if the quotient is exact
303 		    if(z=y) {
304 		        I := i;	 ... restore inexact flag
305 		        R := r;  ... restore rounded mode
306 		        return sqrt(x):=y.
307 		    } else {
308 			z := z - ulp;	... special rounding
309 		    }
310 		}
311 		i := TRUE;		... sqrt(x) is inexact
312 		If (r=RN) then z=z+ulp	... rounded-to-nearest
313 		If (r=RP) then {	... round-toward-+inf
314 		    y = y+ulp; z=z+ulp;
315 		}
316 		y := y+z;		... chopped sum
317 		y0:=y0-0x00100000;	... y := y/2 is correctly rounded.
318 	        I := i;	 		... restore inexact flag
319 	        R := r;  		... restore rounded mode
320 	        return sqrt(x):=y.
321 
322     (4)	Special cases
323 
324 	Square root of +inf, +-0, or NaN is itself;
325 	Square root of a negative number is NaN with invalid signal.
326 
327 
328 B.  sqrt(x) by Reciproot Iteration
329 
330    (1)	Initial approximation
331 
332 	Let x0 and x1 be the leading and the trailing 32-bit words of
333 	a floating point number x (in IEEE double format) respectively
334 	(see section A). By performing shifs and subtracts on x0 and y0,
335 	we obtain a 7.8-bit approximation of 1/sqrt(x) as follows.
336 
337 	    k := 0x5fe80000 - (x0>>1);
338 	    y0:= k - T2[63&(k>>14)].	... y ~ 1/sqrt(x) to 7.8 bits
339 
340 	Here k is a 32-bit integer and T2[] is an integer array
341 	containing correction terms. Now magically the floating
342 	value of y (y's leading 32-bit word is y0, the value of
343 	its trailing word y1 is set to zero) approximates 1/sqrt(x)
344 	to almost 7.8-bit.
345 
346 	Value of T2:
347 	static int T2[64]= {
348 	0x1500,	0x2ef8,	0x4d67,	0x6b02,	0x87be,	0xa395,	0xbe7a,	0xd866,
349 	0xf14a,	0x1091b,0x11fcd,0x13552,0x14999,0x15c98,0x16e34,0x17e5f,
350 	0x18d03,0x19a01,0x1a545,0x1ae8a,0x1b5c4,0x1bb01,0x1bfde,0x1c28d,
351 	0x1c2de,0x1c0db,0x1ba73,0x1b11c,0x1a4b5,0x1953d,0x18266,0x16be0,
352 	0x1683e,0x179d8,0x18a4d,0x19992,0x1a789,0x1b445,0x1bf61,0x1c989,
353 	0x1d16d,0x1d77b,0x1dddf,0x1e2ad,0x1e5bf,0x1e6e8,0x1e654,0x1e3cd,
354 	0x1df2a,0x1d635,0x1cb16,0x1be2c,0x1ae4e,0x19bde,0x1868e,0x16e2e,
355 	0x1527f,0x1334a,0x11051,0xe951,	0xbe01,	0x8e0d,	0x5924,	0x1edd,};
356 
357     (2)	Iterative refinement
358 
359 	Apply Reciproot iteration three times to y and multiply the
360 	result by x to get an approximation z that matches sqrt(x)
361 	to about 1 ulp. To be exact, we will have
362 		-1ulp < sqrt(x)-z<1.0625ulp.
363 
364 	... set rounding mode to Round-to-nearest
365 	   y := y*(1.5-0.5*x*y*y)	... almost 15 sig. bits to 1/sqrt(x)
366 	   y := y*((1.5-2^-30)+0.5*x*y*y)... about 29 sig. bits to 1/sqrt(x)
367 	... special arrangement for better accuracy
368 	   z := x*y			... 29 bits to sqrt(x), with z*y<1
369 	   z := z + 0.5*z*(1-z*y)	... about 1 ulp to sqrt(x)
370 
371 	Remark 2. The constant 1.5-2^-30 is chosen to bias the error so that
372 	(a) the term z*y in the final iteration is always less than 1;
373 	(b) the error in the final result is biased upward so that
374 		-1 ulp < sqrt(x) - z < 1.0625 ulp
375 	    instead of |sqrt(x)-z|<1.03125ulp.
376 
377     (3)	Final adjustment
378 
379 	By twiddling y's last bit it is possible to force y to be
380 	correctly rounded according to the prevailing rounding mode
381 	as follows. Let r and i be copies of the rounding mode and
382 	inexact flag before entering the square root program. Also we
383 	use the expression y+-ulp for the next representable floating
384 	numbers (up and down) of y. Note that y+-ulp = either fixed
385 	point y+-1, or multiply y by nextafter(1,+-inf) in chopped
386 	mode.
387 
388 	R := RZ;		... set rounding mode to round-toward-zero
389 	switch(r) {
390 	    case RN:		... round-to-nearest
391 	       if(x<= z*(z-ulp)...chopped) z = z - ulp; else
392 	       if(x<= z*(z+ulp)...chopped) z = z; else z = z+ulp;
393 	       break;
394 	    case RZ:case RM:	... round-to-zero or round-to--inf
395 	       R:=RP;		... reset rounding mod to round-to-+inf
396 	       if(x<z*z ... rounded up) z = z - ulp; else
397 	       if(x>=(z+ulp)*(z+ulp) ...rounded up) z = z+ulp;
398 	       break;
399 	    case RP:		... round-to-+inf
400 	       if(x>(z+ulp)*(z+ulp)...chopped) z = z+2*ulp; else
401 	       if(x>z*z ...chopped) z = z+ulp;
402 	       break;
403 	}
404 
405 	Remark 3. The above comparisons can be done in fixed point. For
406 	example, to compare x and w=z*z chopped, it suffices to compare
407 	x1 and w1 (the trailing parts of x and w), regarding them as
408 	two's complement integers.
409 
410 	...Is z an exact square root?
411 	To determine whether z is an exact square root of x, let z1 be the
412 	trailing part of z, and also let x0 and x1 be the leading and
413 	trailing parts of x.
414 
415 	If ((z1&0x03ffffff)!=0)	... not exact if trailing 26 bits of z!=0
416 	    I := 1;		... Raise Inexact flag: z is not exact
417 	else {
418 	    j := 1 - [(x0>>20)&1]	... j = logb(x) mod 2
419 	    k := z1 >> 26;		... get z's 25-th and 26-th
420 					    fraction bits
421 	    I := i or (k&j) or ((k&(j+j+1))!=(x1&3));
422 	}
423 	R:= r		... restore rounded mode
424 	return sqrt(x):=z.
425 
426 	If multiplication is cheaper then the foregoing red tape, the
427 	Inexact flag can be evaluated by
428 
429 	    I := i;
430 	    I := (z*z!=x) or I.
431 
432 	Note that z*z can overwrite I; this value must be sensed if it is
433 	True.
434 
435 	Remark 4. If z*z = x exactly, then bit 25 to bit 0 of z1 must be
436 	zero.
437 
438 		    --------------------
439 		z1: |        f2        |
440 		    --------------------
441 		bit 31		   bit 0
442 
443 	Further more, bit 27 and 26 of z1, bit 0 and 1 of x1, and the odd
444 	or even of logb(x) have the following relations:
445 
446 	-------------------------------------------------
447 	bit 27,26 of z1		bit 1,0 of x1	logb(x)
448 	-------------------------------------------------
449 	00			00		odd and even
450 	01			01		even
451 	10			10		odd
452 	10			00		even
453 	11			01		even
454 	-------------------------------------------------
455 
456     (4)	Special cases (see (4) of Section A).
457 
458  */
459 
460