xref: /third_party/jerryscript/jerry-libm/sqrt.c (revision 425bb815)
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
104double
105sqrt (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/*
247Other methods  (use floating-point arithmetic)
248-------------
249(This is a copy of a drafted paper by Prof W. Kahan
250and 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
265A.  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
367B.  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