1 
2 /* @(#)e_sqrt.c 5.1 93/09/24 */
3 /*
4  * ====================================================
5  * Copyright (C) 1993 by Sun Microsystems, Inc. All rights reserved.
6  *
7  * Developed at SunPro, 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 /* __ieee754_sqrt(x)
15  * Return correctly rounded sqrt.
16  *           ------------------------------------------
17  *	     |  Use the hardware sqrt if you have one |
18  *           ------------------------------------------
19  * Method:
20  *   Bit by bit method using integer arithmetic. (Slow, but portable)
21  *   1. Normalization
22  *	Scale x to y in [1,4) with even powers of 2:
23  *	find an integer k such that  1 <= (y=x*2^(2k)) < 4, then
24  *		sqrt(x) = 2^k * sqrt(y)
25  *   2. Bit by bit computation
26  *	Let q  = sqrt(y) truncated to i bit after binary point (q = 1),
27  *	     i							 0
28  *                                     i+1         2
29  *	    s  = 2*q , and	y  =  2   * ( y - q  ).		(1)
30  *	     i      i            i                 i
31  *
32  *	To compute q    from q , one checks whether
33  *		    i+1       i
34  *
35  *			      -(i+1) 2
36  *			(q + 2      ) <= y.			(2)
37  *     			  i
38  *							      -(i+1)
39  *	If (2) is false, then q   = q ; otherwise q   = q  + 2      .
40  *		 	       i+1   i             i+1   i
41  *
42  *	With some algebric manipulation, it is not difficult to see
43  *	that (2) is equivalent to
44  *                             -(i+1)
45  *			s  +  2       <= y			(3)
46  *			 i                i
47  *
48  *	The advantage of (3) is that s  and y  can be computed by
49  *				      i      i
50  *	the following recurrence formula:
51  *	    if (3) is false
52  *
53  *	    s     =  s  ,	y    = y   ;			(4)
54  *	     i+1      i		 i+1    i
55  *
56  *	    otherwise,
57  *                         -i                     -(i+1)
58  *	    s	  =  s  + 2  ,  y    = y  -  s  - 2  		(5)
59  *           i+1      i          i+1    i     i
60  *
61  *	One may easily use induction to prove (4) and (5).
62  *	Note. Since the left hand side of (3) contain only i+2 bits,
63  *	      it does not necessary to do a full (53-bit) comparison
64  *	      in (3).
65  *   3. Final rounding
66  *	After generating the 53 bits result, we compute one more bit.
67  *	Together with the remainder, we can decide whether the
68  *	result is exact, bigger than 1/2ulp, or less than 1/2ulp
69  *	(it will never equal to 1/2ulp).
70  *	The rounding mode can be detected by checking whether
71  *	huge + tiny is equal to huge, and whether huge - tiny is
72  *	equal to huge for some floating point number "huge" and "tiny".
73  *
74  * Special cases:
75  *	sqrt(+-0) = +-0 	... exact
76  *	sqrt(inf) = inf
77  *	sqrt(-ve) = NaN		... with invalid signal
78  *	sqrt(NaN) = NaN		... with invalid signal for signaling NaN
79  *
80  * Other methods : see the appended file at the end of the program below.
81  *---------------
82  */
83 
84 #include "fdlibm.h"
85 
86 #ifndef _DOUBLE_IS_32BITS
87 
88 #ifdef __STDC__
89 static	const double	one	= 1.0, tiny=1.0e-300;
90 #else
91 static	double	one	= 1.0, tiny=1.0e-300;
92 #endif
93 
94 #ifdef __STDC__
__ieee754_sqrt(double x)95 	double __ieee754_sqrt(double x)
96 #else
97 	double __ieee754_sqrt(x)
98 	double x;
99 #endif
100 {
101 	double z;
102 	__int32_t sign = 0x80000000;
103 	__uint32_t r,t1,s1,ix1,q1;
104 	__int32_t ix0,s0,q,m,t,i;
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==(__uint32_t)0xffffffff) { q1=0; q += 1;}
180 		else if (z>one) {
181 		    if (q1==(__uint32_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 
195 #endif /* defined(_DOUBLE_IS_32BITS) */
196 
197 /*
198 Other methods  (use floating-point arithmetic)
199 -------------
200 (This is a copy of a drafted paper by Prof W. Kahan
201 and K.C. Ng, written in May, 1986)
202 
203 	Two algorithms are given here to implement sqrt(x)
204 	(IEEE double precision arithmetic) in software.
205 	Both supply sqrt(x) correctly rounded. The first algorithm (in
206 	Section A) uses newton iterations and involves four divisions.
207 	The second one uses reciproot iterations to avoid division, but
208 	requires more multiplications. Both algorithms need the ability
209 	to chop results of arithmetic operations instead of round them,
210 	and the INEXACT flag to indicate when an arithmetic operation
211 	is executed exactly with no roundoff error, all part of the
212 	standard (IEEE 754-1985). The ability to perform shift, add,
213 	subtract and logical AND operations upon 32-bit words is needed
214 	too, though not part of the standard.
215 
216 A.  sqrt(x) by Newton Iteration
217 
218    (1)	Initial approximation
219 
220 	Let x0 and x1 be the leading and the trailing 32-bit words of
221 	a floating point number x (in IEEE double format) respectively
222 
223 	    1    11		     52				  ...widths
224 	   ------------------------------------------------------
225 	x: |s|	  e     |	      f				|
226 	   ------------------------------------------------------
227 	      msb    lsb  msb				      lsb ...order
228 
229 
230 	     ------------------------  	     ------------------------
231 	x0:  |s|   e    |    f1     |	 x1: |          f2           |
232 	     ------------------------  	     ------------------------
233 
234 	By performing shifts and subtracts on x0 and x1 (both regarded
235 	as integers), we obtain an 8-bit approximation of sqrt(x) as
236 	follows.
237 
238 		k  := (x0>>1) + 0x1ff80000;
239 		y0 := k - T1[31&(k>>15)].	... y ~ sqrt(x) to 8 bits
240 	Here k is a 32-bit integer and T1[] is an integer array containing
241 	correction terms. Now magically the floating value of y (y's
242 	leading 32-bit word is y0, the value of its trailing word is 0)
243 	approximates sqrt(x) to almost 8-bit.
244 
245 	Value of T1:
246 	static int T1[32]= {
247 	0,	1024,	3062,	5746,	9193,	13348,	18162,	23592,
248 	29598,	36145,	43202,	50740,	58733,	67158,	75992,	85215,
249 	83599,	71378,	60428,	50647,	41945,	34246,	27478,	21581,
250 	16499,	12183,	8588,	5674,	3403,	1742,	661,	130,};
251 
252     (2)	Iterative refinement
253 
254 	Apply Heron's rule three times to y, we have y approximates
255 	sqrt(x) to within 1 ulp (Unit in the Last Place):
256 
257 		y := (y+x/y)/2		... almost 17 sig. bits
258 		y := (y+x/y)/2		... almost 35 sig. bits
259 		y := y-(y-x/y)/2	... within 1 ulp
260 
261 
262 	Remark 1.
263 	    Another way to improve y to within 1 ulp is:
264 
265 		y := (y+x/y)		... almost 17 sig. bits to 2*sqrt(x)
266 		y := y - 0x00100006	... almost 18 sig. bits to sqrt(x)
267 
268 				2
269 			    (x-y )*y
270 		y := y + 2* ----------	...within 1 ulp
271 			       2
272 			     3y  + x
273 
274 
275 	This formula has one division fewer than the one above; however,
276 	it requires more multiplications and additions. Also x must be
277 	scaled in advance to avoid spurious overflow in evaluating the
278 	expression 3y*y+x. Hence it is not recommended uless division
279 	is slow. If division is very slow, then one should use the
280 	reciproot algorithm given in section B.
281 
282     (3) Final adjustment
283 
284 	By twiddling y's last bit it is possible to force y to be
285 	correctly rounded according to the prevailing rounding mode
286 	as follows. Let r and i be copies of the rounding mode and
287 	inexact flag before entering the square root program. Also we
288 	use the expression y+-ulp for the next representable floating
289 	numbers (up and down) of y. Note that y+-ulp = either fixed
290 	point y+-1, or multiply y by nextafter(1,+-inf) in chopped
291 	mode.
292 
293 		I := FALSE;	... reset INEXACT flag I
294 		R := RZ;	... set rounding mode to round-toward-zero
295 		z := x/y;	... chopped quotient, possibly inexact
296 		If(not I) then {	... if the quotient is exact
297 		    if(z=y) {
298 		        I := i;	 ... restore inexact flag
299 		        R := r;  ... restore rounded mode
300 		        return sqrt(x):=y.
301 		    } else {
302 			z := z - ulp;	... special rounding
303 		    }
304 		}
305 		i := TRUE;		... sqrt(x) is inexact
306 		If (r=RN) then z=z+ulp	... rounded-to-nearest
307 		If (r=RP) then {	... round-toward-+inf
308 		    y = y+ulp; z=z+ulp;
309 		}
310 		y := y+z;		... chopped sum
311 		y0:=y0-0x00100000;	... y := y/2 is correctly rounded.
312 	        I := i;	 		... restore inexact flag
313 	        R := r;  		... restore rounded mode
314 	        return sqrt(x):=y.
315 
316     (4)	Special cases
317 
318 	Square root of +inf, +-0, or NaN is itself;
319 	Square root of a negative number is NaN with invalid signal.
320 
321 
322 B.  sqrt(x) by Reciproot Iteration
323 
324    (1)	Initial approximation
325 
326 	Let x0 and x1 be the leading and the trailing 32-bit words of
327 	a floating point number x (in IEEE double format) respectively
328 	(see section A). By performing shifs and subtracts on x0 and y0,
329 	we obtain a 7.8-bit approximation of 1/sqrt(x) as follows.
330 
331 	    k := 0x5fe80000 - (x0>>1);
332 	    y0:= k - T2[63&(k>>14)].	... y ~ 1/sqrt(x) to 7.8 bits
333 
334 	Here k is a 32-bit integer and T2[] is an integer array
335 	containing correction terms. Now magically the floating
336 	value of y (y's leading 32-bit word is y0, the value of
337 	its trailing word y1 is set to zero) approximates 1/sqrt(x)
338 	to almost 7.8-bit.
339 
340 	Value of T2:
341 	static int T2[64]= {
342 	0x1500,	0x2ef8,	0x4d67,	0x6b02,	0x87be,	0xa395,	0xbe7a,	0xd866,
343 	0xf14a,	0x1091b,0x11fcd,0x13552,0x14999,0x15c98,0x16e34,0x17e5f,
344 	0x18d03,0x19a01,0x1a545,0x1ae8a,0x1b5c4,0x1bb01,0x1bfde,0x1c28d,
345 	0x1c2de,0x1c0db,0x1ba73,0x1b11c,0x1a4b5,0x1953d,0x18266,0x16be0,
346 	0x1683e,0x179d8,0x18a4d,0x19992,0x1a789,0x1b445,0x1bf61,0x1c989,
347 	0x1d16d,0x1d77b,0x1dddf,0x1e2ad,0x1e5bf,0x1e6e8,0x1e654,0x1e3cd,
348 	0x1df2a,0x1d635,0x1cb16,0x1be2c,0x1ae4e,0x19bde,0x1868e,0x16e2e,
349 	0x1527f,0x1334a,0x11051,0xe951,	0xbe01,	0x8e0d,	0x5924,	0x1edd,};
350 
351     (2)	Iterative refinement
352 
353 	Apply Reciproot iteration three times to y and multiply the
354 	result by x to get an approximation z that matches sqrt(x)
355 	to about 1 ulp. To be exact, we will have
356 		-1ulp < sqrt(x)-z<1.0625ulp.
357 
358 	... set rounding mode to Round-to-nearest
359 	   y := y*(1.5-0.5*x*y*y)	... almost 15 sig. bits to 1/sqrt(x)
360 	   y := y*((1.5-2^-30)+0.5*x*y*y)... about 29 sig. bits to 1/sqrt(x)
361 	... special arrangement for better accuracy
362 	   z := x*y			... 29 bits to sqrt(x), with z*y<1
363 	   z := z + 0.5*z*(1-z*y)	... about 1 ulp to sqrt(x)
364 
365 	Remark 2. The constant 1.5-2^-30 is chosen to bias the error so that
366 	(a) the term z*y in the final iteration is always less than 1;
367 	(b) the error in the final result is biased upward so that
368 		-1 ulp < sqrt(x) - z < 1.0625 ulp
369 	    instead of |sqrt(x)-z|<1.03125ulp.
370 
371     (3)	Final adjustment
372 
373 	By twiddling y's last bit it is possible to force y to be
374 	correctly rounded according to the prevailing rounding mode
375 	as follows. Let r and i be copies of the rounding mode and
376 	inexact flag before entering the square root program. Also we
377 	use the expression y+-ulp for the next representable floating
378 	numbers (up and down) of y. Note that y+-ulp = either fixed
379 	point y+-1, or multiply y by nextafter(1,+-inf) in chopped
380 	mode.
381 
382 	R := RZ;		... set rounding mode to round-toward-zero
383 	switch(r) {
384 	    case RN:		... round-to-nearest
385 	       if(x<= z*(z-ulp)...chopped) z = z - ulp; else
386 	       if(x<= z*(z+ulp)...chopped) z = z; else z = z+ulp;
387 	       break;
388 	    case RZ:case RM:	... round-to-zero or round-to--inf
389 	       R:=RP;		... reset rounding mod to round-to-+inf
390 	       if(x<z*z ... rounded up) z = z - ulp; else
391 	       if(x>=(z+ulp)*(z+ulp) ...rounded up) z = z+ulp;
392 	       break;
393 	    case RP:		... round-to-+inf
394 	       if(x>(z+ulp)*(z+ulp)...chopped) z = z+2*ulp; else
395 	       if(x>z*z ...chopped) z = z+ulp;
396 	       break;
397 	}
398 
399 	Remark 3. The above comparisons can be done in fixed point. For
400 	example, to compare x and w=z*z chopped, it suffices to compare
401 	x1 and w1 (the trailing parts of x and w), regarding them as
402 	two's complement integers.
403 
404 	...Is z an exact square root?
405 	To determine whether z is an exact square root of x, let z1 be the
406 	trailing part of z, and also let x0 and x1 be the leading and
407 	trailing parts of x.
408 
409 	If ((z1&0x03ffffff)!=0)	... not exact if trailing 26 bits of z!=0
410 	    I := 1;		... Raise Inexact flag: z is not exact
411 	else {
412 	    j := 1 - [(x0>>20)&1]	... j = logb(x) mod 2
413 	    k := z1 >> 26;		... get z's 25-th and 26-th
414 					    fraction bits
415 	    I := i or (k&j) or ((k&(j+j+1))!=(x1&3));
416 	}
417 	R:= r		... restore rounded mode
418 	return sqrt(x):=z.
419 
420 	If multiplication is cheaper then the foregoing red tape, the
421 	Inexact flag can be evaluated by
422 
423 	    I := i;
424 	    I := (z*z!=x) or I.
425 
426 	Note that z*z can overwrite I; this value must be sensed if it is
427 	True.
428 
429 	Remark 4. If z*z = x exactly, then bit 25 to bit 0 of z1 must be
430 	zero.
431 
432 		    --------------------
433 		z1: |        f2        |
434 		    --------------------
435 		bit 31		   bit 0
436 
437 	Further more, bit 27 and 26 of z1, bit 0 and 1 of x1, and the odd
438 	or even of logb(x) have the following relations:
439 
440 	-------------------------------------------------
441 	bit 27,26 of z1		bit 1,0 of x1	logb(x)
442 	-------------------------------------------------
443 	00			00		odd and even
444 	01			01		even
445 	10			10		odd
446 	10			00		even
447 	11			01		even
448 	-------------------------------------------------
449 
450     (4)	Special cases (see (4) of Section A).
451 
452  */
453