• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /* Copyright JS Foundation and other contributors, http://js.foundation
2  *
3  * Licensed under the Apache License, Version 2.0 (the "License");
4  * you may not use this file except in compliance with the License.
5  * You may obtain a copy of the License at
6  *
7  *     http://www.apache.org/licenses/LICENSE-2.0
8  *
9  * Unless required by applicable law or agreed to in writing, software
10  * distributed under the License is distributed on an "AS IS" BASIS
11  * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12  * See the License for the specific language governing permissions and
13  * limitations under the License.
14  *
15  * This file is based on work under the following copyright and permission
16  * notice:
17  *
18  *     Copyright (C) 1993 by Sun Microsystems, Inc. All rights reserved.
19  *
20  *     Developed at SunSoft, a Sun Microsystems, Inc. business.
21  *     Permission to use, copy, modify, and distribute this
22  *     software is freely granted, provided that this notice
23  *     is preserved.
24  *
25  *     @(#)e_sqrt.c 1.3 95/01/18
26  */
27 
28 #include "jerry-libm-internal.h"
29 
30 /* sqrt(x)
31  * Return correctly rounded sqrt.
32  *
33  *           ------------------------------------------
34  *           |  Use the hardware sqrt if you have one |
35  *           ------------------------------------------
36  *
37  * Method:
38  *   Bit by bit method using integer arithmetic. (Slow, but portable)
39  *   1. Normalization
40  *      Scale x to y in [1,4) with even powers of 2:
41  *      find an integer k such that  1 <= (y=x*2^(2k)) < 4, then
42  *              sqrt(x) = 2^k * sqrt(y)
43  *   2. Bit by bit computation
44  *      Let q  = sqrt(y) truncated to i bit after binary point (q = 1),
45  *           i                                                   0
46  *                                     i+1         2
47  *          s  = 2*q , and      y  =  2   * ( y - q  ).         (1)
48  *           i      i            i                 i
49  *
50  *      To compute q    from q , one checks whether
51  *                  i+1       i
52  *
53  *                            -(i+1) 2
54  *                      (q + 2      ) <= y.                     (2)
55  *                        i
56  *                                                            -(i+1)
57  *      If (2) is false, then q   = q ; otherwise q   = q  + 2      .
58  *                             i+1   i             i+1   i
59  *
60  *      With some algebric manipulation, it is not difficult to see
61  *      that (2) is equivalent to
62  *                             -(i+1)
63  *                      s  +  2       <= y                      (3)
64  *                       i                i
65  *
66  *      The advantage of (3) is that s  and y  can be computed by
67  *                                    i      i
68  *      the following recurrence formula:
69  *          if (3) is false
70  *
71  *          s     =  s  ,       y    = y   ;                    (4)
72  *           i+1      i          i+1    i
73  *
74  *          otherwise,
75  *                         -i                     -(i+1)
76  *          s     =  s  + 2  ,  y    = y  -  s  - 2             (5)
77  *           i+1      i          i+1    i     i
78  *
79  *      One may easily use induction to prove (4) and (5).
80  *      Note. Since the left hand side of (3) contain only i+2 bits,
81  *            it does not necessary to do a full (53-bit) comparison
82  *            in (3).
83  *   3. Final rounding
84  *      After generating the 53 bits result, we compute one more bit.
85  *      Together with the remainder, we can decide whether the
86  *      result is exact, bigger than 1/2ulp, or less than 1/2ulp
87  *      (it will never equal to 1/2ulp).
88  *      The rounding mode can be detected by checking whether
89  *      huge + tiny is equal to huge, and whether huge - tiny is
90  *      equal to huge for some floating point number "huge" and "tiny".
91  *
92  * Special cases:
93  *      sqrt(+-0) = +-0         ... exact
94  *      sqrt(inf) = inf
95  *      sqrt(-ve) = NaN         ... with invalid signal
96  *      sqrt(NaN) = NaN         ... with invalid signal for signaling NaN
97  *
98  * Other methods: see the appended file at the end of the program below.
99  */
100 
101 #define one  1.0
102 #define tiny 1.0e-300
103 
104 double
sqrt(double x)105 sqrt (double x)
106 {
107   int sign = (int) 0x80000000;
108   unsigned r, t1, s1, ix1, q1;
109   int ix0, s0, q, m, t, i;
110 
111   ix0 = __HI (x); /* high word of x */
112   ix1 = __LO (x); /* low word of x */
113 
114   /* take care of Inf and NaN */
115   if ((ix0 & 0x7ff00000) == 0x7ff00000)
116   {
117     return x * x + x; /* sqrt(NaN) = NaN, sqrt(+inf) = +inf, sqrt(-inf) = sNaN */
118   }
119   /* take care of zero */
120   if (ix0 <= 0)
121   {
122     if (((ix0 & (~sign)) | ix1) == 0) /* sqrt(+-0) = +-0 */
123     {
124       return x;
125     }
126     else if (ix0 < 0) /* sqrt(-ve) = sNaN */
127     {
128       return NAN;
129     }
130   }
131   /* normalize x */
132   m = (ix0 >> 20);
133   if (m == 0) /* subnormal x */
134   {
135     while (ix0 == 0)
136     {
137       m -= 21;
138       ix0 |= (ix1 >> 11);
139       ix1 <<= 21;
140     }
141     for (i = 0; (ix0 & 0x00100000) == 0; i++)
142     {
143       ix0 <<= 1;
144     }
145     m -= i - 1;
146     ix0 |= (ix1 >> (32 - i));
147     ix1 <<= i;
148   }
149   m -= 1023; /* unbias exponent */
150   ix0 = (ix0 & 0x000fffff) | 0x00100000;
151   if (m & 1) /* odd m, double x to make it even */
152   {
153     ix0 += ix0 + ((ix1 & sign) >> 31);
154     ix1 += ix1;
155   }
156   m >>= 1; /* m = [m / 2] */
157 
158   /* generate sqrt(x) bit by bit */
159   ix0 += ix0 + ((ix1 & sign) >> 31);
160   ix1 += ix1;
161   q = q1 = s0 = s1 = 0; /* [q,q1] = sqrt(x) */
162   r = 0x00200000; /* r = moving bit from right to left */
163 
164   while (r != 0)
165   {
166     t = s0 + r;
167     if (t <= ix0)
168     {
169       s0 = t + r;
170       ix0 -= t;
171       q += r;
172     }
173     ix0 += ix0 + ((ix1 & sign) >> 31);
174     ix1 += ix1;
175     r >>= 1;
176   }
177 
178   r = sign;
179   while (r != 0)
180   {
181     t1 = s1 + r;
182     t = s0;
183     if ((t < ix0) || ((t == ix0) && (t1 <= ix1)))
184     {
185       s1 = t1 + r;
186       if (((t1 & sign) == sign) && (s1 & sign) == 0)
187       {
188         s0 += 1;
189       }
190       ix0 -= t;
191       if (ix1 < t1)
192       {
193         ix0 -= 1;
194       }
195       ix1 -= t1;
196       q1 += r;
197     }
198     ix0 += ix0 + ((ix1 & sign) >> 31);
199     ix1 += ix1;
200     r >>= 1;
201   }
202 
203   double_accessor ret;
204 
205   /* use floating add to find out rounding direction */
206   if ((ix0 | ix1) != 0)
207   {
208     ret.dbl = one - tiny; /* trigger inexact flag */
209     if (ret.dbl >= one)
210     {
211       ret.dbl = one + tiny;
212       if (q1 == (unsigned) 0xffffffff)
213       {
214         q1 = 0;
215         q += 1;
216       }
217       else if (ret.dbl > one)
218       {
219         if (q1 == (unsigned) 0xfffffffe)
220         {
221           q += 1;
222         }
223         q1 += 2;
224       }
225       else
226       {
227         q1 += (q1 & 1);
228       }
229     }
230   }
231   ix0 = (q >> 1) + 0x3fe00000;
232   ix1 = q1 >> 1;
233   if ((q & 1) == 1)
234   {
235     ix1 |= sign;
236   }
237   ix0 += (m << 20);
238   ret.as_int.hi = ix0;
239   ret.as_int.lo = ix1;
240   return ret.dbl;
241 } /* sqrt */
242 
243 #undef one
244 #undef tiny
245 
246 /*
247 Other methods  (use floating-point arithmetic)
248 -------------
249 (This is a copy of a drafted paper by Prof W. Kahan
250 and K.C. Ng, written in May, 1986)
251 
252         Two algorithms are given here to implement sqrt(x)
253         (IEEE double precision arithmetic) in software.
254         Both supply sqrt(x) correctly rounded. The first algorithm (in
255         Section A) uses newton iterations and involves four divisions.
256         The second one uses reciproot iterations to avoid division, but
257         requires more multiplications. Both algorithms need the ability
258         to chop results of arithmetic operations instead of round them,
259         and the INEXACT flag to indicate when an arithmetic operation
260         is executed exactly with no roundoff error, all part of the
261         standard (IEEE 754-1985). The ability to perform shift, add,
262         subtract and logical AND operations upon 32-bit words is needed
263         too, though not part of the standard.
264 
265 A.  sqrt(x) by Newton Iteration
266 
267    (1)  Initial approximation
268 
269         Let x0 and x1 be the leading and the trailing 32-bit words of
270         a floating point number x (in IEEE double format) respectively
271 
272             1    11                  52                           ...widths
273            ------------------------------------------------------
274         x: |s|    e     |             f                         |
275            ------------------------------------------------------
276               msb    lsb  msb                                 lsb ...order
277 
278              ------------------------        ------------------------
279         x0:  |s|   e    |    f1     |    x1: |          f2           |
280              ------------------------        ------------------------
281 
282         By performing shifts and subtracts on x0 and x1 (both regarded
283         as integers), we obtain an 8-bit approximation of sqrt(x) as
284         follows.
285 
286                 k  := (x0>>1) + 0x1ff80000;
287                 y0 := k - T1[31&(k>>15)].       ... y ~ sqrt(x) to 8 bits
288         Here k is a 32-bit integer and T1[] is an integer array containing
289         correction terms. Now magically the floating value of y (y's
290         leading 32-bit word is y0, the value of its trailing word is 0)
291         approximates sqrt(x) to almost 8-bit.
292 
293         Value of T1:
294         static int T1[32]= {
295         0,      1024,   3062,   5746,   9193,   13348,  18162,  23592,
296         29598,  36145,  43202,  50740,  58733,  67158,  75992,  85215,
297         83599,  71378,  60428,  50647,  41945,  34246,  27478,  21581,
298         16499,  12183,  8588,   5674,   3403,   1742,   661,    130,};
299 
300     (2) Iterative refinement
301 
302         Apply Heron's rule three times to y, we have y approximates
303         sqrt(x) to within 1 ulp (Unit in the Last Place):
304 
305                 y := (y+x/y)/2          ... almost 17 sig. bits
306                 y := (y+x/y)/2          ... almost 35 sig. bits
307                 y := y-(y-x/y)/2        ... within 1 ulp
308 
309         Remark 1.
310             Another way to improve y to within 1 ulp is:
311 
312                 y := (y+x/y)            ... almost 17 sig. bits to 2*sqrt(x)
313                 y := y - 0x00100006     ... almost 18 sig. bits to sqrt(x)
314 
315                                 2
316                             (x-y )*y
317                 y := y + 2* ----------  ...within 1 ulp
318                                2
319                              3y  + x
320 
321         This formula has one division fewer than the one above; however,
322         it requires more multiplications and additions. Also x must be
323         scaled in advance to avoid spurious overflow in evaluating the
324         expression 3y*y+x. Hence it is not recommended uless division
325         is slow. If division is very slow, then one should use the
326         reciproot algorithm given in section B.
327 
328     (3) Final adjustment
329 
330         By twiddling y's last bit it is possible to force y to be
331         correctly rounded according to the prevailing rounding mode
332         as follows. Let r and i be copies of the rounding mode and
333         inexact flag before entering the square root program. Also we
334         use the expression y+-ulp for the next representable floating
335         numbers (up and down) of y. Note that y+-ulp = either fixed
336         point y+-1, or multiply y by nextafter(1,+-inf) in chopped
337         mode.
338 
339         I := FALSE;     ... reset INEXACT flag I
340         R := RZ;        ... set rounding mode to round-toward-zero
341                 z := x/y;       ... chopped quotient, possibly inexact
342                 If(not I) then {        ... if the quotient is exact
343                     if(z=y) {
344                         I := i;  ... restore inexact flag
345                         R := r;  ... restore rounded mode
346                         return sqrt(x):=y.
347                     } else {
348                         z := z - ulp;   ... special rounding
349                     }
350                 }
351                 i := TRUE;              ... sqrt(x) is inexact
352                 If (r=RN) then z=z+ulp  ... rounded-to-nearest
353                 If (r=RP) then {        ... round-toward-+inf
354                     y = y+ulp; z=z+ulp;
355                 }
356                 y := y+z;               ... chopped sum
357                 y0:=y0-0x00100000;      ... y := y/2 is correctly rounded.
358                 I := i;                 ... restore inexact flag
359                 R := r;                 ... restore rounded mode
360                 return sqrt(x):=y.
361 
362     (4) Special cases
363 
364         Square root of +inf, +-0, or NaN is itself;
365         Square root of a negative number is NaN with invalid signal.
366 
367 B.  sqrt(x) by Reciproot Iteration
368 
369    (1)  Initial approximation
370 
371         Let x0 and x1 be the leading and the trailing 32-bit words of
372         a floating point number x (in IEEE double format) respectively
373         (see section A). By performing shifs and subtracts on x0 and y0,
374         we obtain a 7.8-bit approximation of 1/sqrt(x) as follows.
375 
376             k := 0x5fe80000 - (x0>>1);
377             y0:= k - T2[63&(k>>14)].    ... y ~ 1/sqrt(x) to 7.8 bits
378 
379         Here k is a 32-bit integer and T2[] is an integer array
380         containing correction terms. Now magically the floating
381         value of y (y's leading 32-bit word is y0, the value of
382         its trailing word y1 is set to zero) approximates 1/sqrt(x)
383         to almost 7.8-bit.
384 
385         Value of T2:
386         static int T2[64]= {
387         0x1500, 0x2ef8, 0x4d67, 0x6b02, 0x87be, 0xa395, 0xbe7a, 0xd866,
388         0xf14a, 0x1091b,0x11fcd,0x13552,0x14999,0x15c98,0x16e34,0x17e5f,
389         0x18d03,0x19a01,0x1a545,0x1ae8a,0x1b5c4,0x1bb01,0x1bfde,0x1c28d,
390         0x1c2de,0x1c0db,0x1ba73,0x1b11c,0x1a4b5,0x1953d,0x18266,0x16be0,
391         0x1683e,0x179d8,0x18a4d,0x19992,0x1a789,0x1b445,0x1bf61,0x1c989,
392         0x1d16d,0x1d77b,0x1dddf,0x1e2ad,0x1e5bf,0x1e6e8,0x1e654,0x1e3cd,
393         0x1df2a,0x1d635,0x1cb16,0x1be2c,0x1ae4e,0x19bde,0x1868e,0x16e2e,
394         0x1527f,0x1334a,0x11051,0xe951, 0xbe01, 0x8e0d, 0x5924, 0x1edd,};
395 
396     (2) Iterative refinement
397 
398         Apply Reciproot iteration three times to y and multiply the
399         result by x to get an approximation z that matches sqrt(x)
400         to about 1 ulp. To be exact, we will have
401                 -1ulp < sqrt(x)-z<1.0625ulp.
402 
403         ... set rounding mode to Round-to-nearest
404            y := y*(1.5-0.5*x*y*y)       ... almost 15 sig. bits to 1/sqrt(x)
405            y := y*((1.5-2^-30)+0.5*x*y*y)... about 29 sig. bits to 1/sqrt(x)
406         ... special arrangement for better accuracy
407            z := x*y                     ... 29 bits to sqrt(x), with z*y<1
408            z := z + 0.5*z*(1-z*y)       ... about 1 ulp to sqrt(x)
409 
410         Remark 2. The constant 1.5-2^-30 is chosen to bias the error so that
411         (a) the term z*y in the final iteration is always less than 1;
412         (b) the error in the final result is biased upward so that
413                 -1 ulp < sqrt(x) - z < 1.0625 ulp
414             instead of |sqrt(x)-z|<1.03125ulp.
415 
416     (3) Final adjustment
417 
418         By twiddling y's last bit it is possible to force y to be
419         correctly rounded according to the prevailing rounding mode
420         as follows. Let r and i be copies of the rounding mode and
421         inexact flag before entering the square root program. Also we
422         use the expression y+-ulp for the next representable floating
423         numbers (up and down) of y. Note that y+-ulp = either fixed
424         point y+-1, or multiply y by nextafter(1,+-inf) in chopped
425         mode.
426 
427         R := RZ;                ... set rounding mode to round-toward-zero
428         switch(r) {
429             case RN:            ... round-to-nearest
430                if(x<= z*(z-ulp)...chopped) z = z - ulp; else
431                if(x<= z*(z+ulp)...chopped) z = z; else z = z+ulp;
432                break;
433             case RZ:case RM:    ... round-to-zero or round-to--inf
434                R:=RP;           ... reset rounding mod to round-to-+inf
435                if(x<z*z ... rounded up) z = z - ulp; else
436                if(x>=(z+ulp)*(z+ulp) ...rounded up) z = z+ulp;
437                break;
438             case RP:            ... round-to-+inf
439                if(x>(z+ulp)*(z+ulp)...chopped) z = z+2*ulp; else
440                if(x>z*z ...chopped) z = z+ulp;
441                break;
442         }
443 
444         Remark 3. The above comparisons can be done in fixed point. For
445         example, to compare x and w=z*z chopped, it suffices to compare
446         x1 and w1 (the trailing parts of x and w), regarding them as
447         two's complement integers.
448 
449         ...Is z an exact square root?
450         To determine whether z is an exact square root of x, let z1 be the
451         trailing part of z, and also let x0 and x1 be the leading and
452         trailing parts of x.
453 
454         If ((z1&0x03ffffff)!=0) ... not exact if trailing 26 bits of z!=0
455             I := 1;             ... Raise Inexact flag: z is not exact
456         else {
457             j := 1 - [(x0>>20)&1]       ... j = logb(x) mod 2
458             k := z1 >> 26;              ... get z's 25-th and 26-th
459                                             fraction bits
460             I := i or (k&j) or ((k&(j+j+1))!=(x1&3));
461         }
462         R:= r           ... restore rounded mode
463         return sqrt(x):=z.
464 
465         If multiplication is cheaper then the foregoing red tape, the
466         Inexact flag can be evaluated by
467 
468             I := i;
469             I := (z*z!=x) or I.
470 
471         Note that z*z can overwrite I; this value must be sensed if it is
472         True.
473 
474         Remark 4. If z*z = x exactly, then bit 25 to bit 0 of z1 must be
475         zero.
476 
477                     --------------------
478                 z1: |        f2        |
479                     --------------------
480                 bit 31             bit 0
481 
482         Further more, bit 27 and 26 of z1, bit 0 and 1 of x1, and the odd
483         or even of logb(x) have the following relations:
484 
485         -------------------------------------------------
486         bit 27,26 of z1         bit 1,0 of x1   logb(x)
487         -------------------------------------------------
488         00                      00              odd and even
489         01                      01              even
490         10                      10              odd
491         10                      00              even
492         11                      01              even
493         -------------------------------------------------
494 
495     (4) Special cases (see (4) of Section A).
496  */
497