1 /*
2  * Copyright (C) 2015 The Android Open Source Project
3  *
4  * Licensed under the Apache License, Version 2.0 (the "License");
5  * you may not use this file except in compliance with the License.
6  * You may obtain a copy of the License at
7  *
8  *   http://www.apache.org/licenses/LICENSE-2.0
9  *
10  * Unless required by applicable law or agreed to in writing, software
11  * distributed under the License is distributed on an "AS IS" BASIS,
12  * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13  * See the License for the specific language governing permissions and
14  * limitations under the License.
15  */
16 
17 /*
18  * The above license covers additions and changes by AOSP authors.
19  * The original code is licensed as follows:
20  */
21 
22 //
23 // Copyright (c) 1999, Silicon Graphics, Inc. -- ALL RIGHTS RESERVED
24 //
25 // Permission is granted free of charge to copy, modify, use and distribute
26 // this software  provided you include the entirety of this notice in all
27 // copies made.
28 //
29 // THIS SOFTWARE IS PROVIDED ON AN AS IS BASIS, WITHOUT WARRANTY OF ANY
30 // KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, WITHOUT LIMITATION,
31 // WARRANTIES THAT THE SUBJECT SOFTWARE IS FREE OF DEFECTS, MERCHANTABLE, FIT
32 // FOR A PARTICULAR PURPOSE OR NON-INFRINGING.   SGI ASSUMES NO RISK AS TO THE
33 // QUALITY AND PERFORMANCE OF THE SOFTWARE.   SHOULD THE SOFTWARE PROVE
34 // DEFECTIVE IN ANY RESPECT, SGI ASSUMES NO COST OR LIABILITY FOR ANY
35 // SERVICING, REPAIR OR CORRECTION.  THIS DISCLAIMER OF WARRANTY CONSTITUTES
36 // AN ESSENTIAL PART OF THIS LICENSE. NO USE OF ANY SUBJECT SOFTWARE IS
37 // AUTHORIZED HEREUNDER EXCEPT UNDER THIS DISCLAIMER.
38 //
39 // UNDER NO CIRCUMSTANCES AND UNDER NO LEGAL THEORY, WHETHER TORT (INCLUDING,
40 // WITHOUT LIMITATION, NEGLIGENCE OR STRICT LIABILITY), CONTRACT, OR
41 // OTHERWISE, SHALL SGI BE LIABLE FOR ANY DIRECT, INDIRECT, SPECIAL,
42 // INCIDENTAL, OR CONSEQUENTIAL DAMAGES OF ANY CHARACTER WITH RESPECT TO THE
43 // SOFTWARE INCLUDING, WITHOUT LIMITATION, DAMAGES FOR LOSS OF GOODWILL, WORK
44 // STOPPAGE, LOSS OF DATA, COMPUTER FAILURE OR MALFUNCTION, OR ANY AND ALL
45 // OTHER COMMERCIAL DAMAGES OR LOSSES, EVEN IF SGI SHALL HAVE BEEN INFORMED OF
46 // THE POSSIBILITY OF SUCH DAMAGES.  THIS LIMITATION OF LIABILITY SHALL NOT
47 // APPLY TO LIABILITY RESULTING FROM SGI's NEGLIGENCE TO THE EXTENT APPLICABLE
48 // LAW PROHIBITS SUCH LIMITATION.  SOME JURISDICTIONS DO NOT ALLOW THE
49 // EXCLUSION OR LIMITATION OF INCIDENTAL OR CONSEQUENTIAL DAMAGES, SO THAT
50 // EXCLUSION AND LIMITATION MAY NOT APPLY TO YOU.
51 //
52 // These license terms shall be governed by and construed in accordance with
53 // the laws of the United States and the State of California as applied to
54 // agreements entered into and to be performed entirely within California
55 // between California residents.  Any litigation relating to these license
56 // terms shall be subject to the exclusive jurisdiction of the Federal Courts
57 // of the Northern District of California (or, absent subject matter
58 // jurisdiction in such courts, the courts of the State of California), with
59 // venue lying exclusively in Santa Clara County, California.
60 
61 // Copyright (c) 2001-2004, Hewlett-Packard Development Company, L.P.
62 //
63 // Permission is granted free of charge to copy, modify, use and distribute
64 // this software  provided you include the entirety of this notice in all
65 // copies made.
66 //
67 // THIS SOFTWARE IS PROVIDED ON AN AS IS BASIS, WITHOUT WARRANTY OF ANY
68 // KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, WITHOUT LIMITATION,
69 // WARRANTIES THAT THE SUBJECT SOFTWARE IS FREE OF DEFECTS, MERCHANTABLE, FIT
70 // FOR A PARTICULAR PURPOSE OR NON-INFRINGING.   HEWLETT-PACKARD ASSUMES
71 // NO RISK AS TO THE QUALITY AND PERFORMANCE OF THE SOFTWARE.
72 // SHOULD THE SOFTWARE PROVE DEFECTIVE IN ANY RESPECT,
73 // HEWLETT-PACKARD ASSUMES NO COST OR LIABILITY FOR ANY
74 // SERVICING, REPAIR OR CORRECTION.  THIS DISCLAIMER OF WARRANTY CONSTITUTES
75 // AN ESSENTIAL PART OF THIS LICENSE. NO USE OF ANY SUBJECT SOFTWARE IS
76 // AUTHORIZED HEREUNDER EXCEPT UNDER THIS DISCLAIMER.
77 //
78 // UNDER NO CIRCUMSTANCES AND UNDER NO LEGAL THEORY, WHETHER TORT (INCLUDING,
79 // WITHOUT LIMITATION, NEGLIGENCE OR STRICT LIABILITY), CONTRACT, OR
80 // OTHERWISE, SHALL HEWLETT-PACKARD BE LIABLE FOR ANY DIRECT, INDIRECT, SPECIAL,
81 // INCIDENTAL, OR CONSEQUENTIAL DAMAGES OF ANY CHARACTER WITH RESPECT TO THE
82 // SOFTWARE INCLUDING, WITHOUT LIMITATION, DAMAGES FOR LOSS OF GOODWILL, WORK
83 // STOPPAGE, LOSS OF DATA, COMPUTER FAILURE OR MALFUNCTION, OR ANY AND ALL
84 // OTHER COMMERCIAL DAMAGES OR LOSSES, EVEN IF HEWLETT-PACKARD SHALL
85 // HAVE BEEN INFORMED OF THE POSSIBILITY OF SUCH DAMAGES.
86 // THIS LIMITATION OF LIABILITY SHALL NOT APPLY TO LIABILITY RESULTING
87 // FROM HEWLETT-PACKARD's NEGLIGENCE TO THE EXTENT APPLICABLE
88 // LAW PROHIBITS SUCH LIMITATION.  SOME JURISDICTIONS DO NOT ALLOW THE
89 // EXCLUSION OR LIMITATION OF INCIDENTAL OR CONSEQUENTIAL DAMAGES, SO THAT
90 // EXCLUSION AND LIMITATION MAY NOT APPLY TO YOU.
91 //
92 
93 // Added valueOf(string, radix), fixed some documentation comments.
94 //              Hans_Boehm@hp.com 1/12/2001
95 // Fixed a serious typo in inv_CR():  For negative arguments it produced
96 //              the wrong sign.  This affected the sign of divisions.
97 // Added byteValue and fixed some comments.  Hans.Boehm@hp.com 12/17/2002
98 // Added toStringFloatRep.      Hans.Boehm@hp.com 4/1/2004
99 // Added get_appr() synchronization to allow access from multiple threads
100 // hboehm@google.com 4/25/2014
101 // Changed cos() prescaling to avoid logarithmic depth tree.
102 // hboehm@google.com 6/30/2014
103 // Added explicit asin() implementation.  Remove one.  Add ZERO and ONE and
104 // make them public.  hboehm@google.com 5/21/2015
105 // Added Gauss-Legendre PI implementation.  Removed two.
106 // hboehm@google.com 4/12/2016
107 // Fix shift operation in doubleValue. That produced incorrect values for
108 // large negative exponents.
109 // Don't negate argument and compute inverse for exp(). That causes severe
110 // performance problems for (-huge).exp()
111 // hboehm@google.com 8/21/2017
112 // Have comparison check for interruption. hboehm@google.com 10/31/2017
113 
114 package com.hp.creals;
115 
116 import java.math.BigInteger;
117 import java.util.ArrayList;
118 
119 /**
120 * Constructive real numbers, also known as recursive, or computable reals.
121 * Each recursive real number is represented as an object that provides an
122 * approximation function for the real number.
123 * The approximation function guarantees that the generated approximation
124 * is accurate to the specified precision.
125 * Arithmetic operations on constructive reals produce new such objects;
126 * they typically do not perform any real computation.
127 * In this sense, arithmetic computations are exact: They produce
128 * a description which describes the exact answer, and can be used to
129 * later approximate it to arbitrary precision.
130 * <P>
131 * When approximations are generated, <I>e.g.</i> for output, they are
132 * accurate to the requested precision; no cumulative rounding errors
133 * are visible.
134 * In order to achieve this precision, the approximation function will often
135 * need to approximate subexpressions to greater precision than was originally
136 * demanded.  Thus the approximation of a constructive real number
137 * generated through a complex sequence of operations may eventually require
138 * evaluation to very high precision.  This usually makes such computations
139 * prohibitively expensive for large numerical problems.
140 * But it is perfectly appropriate for use in a desk calculator,
141 * for small numerical problems, for the evaluation of expressions
142 * computated by a symbolic algebra system, for testing of accuracy claims
143 * for floating point code on small inputs, or the like.
144 * <P>
145 * We expect that the vast majority of uses will ignore the particular
146 * implementation, and the member functons <TT>approximate</tt>
147 * and <TT>get_appr</tt>.  Such applications will treat <TT>CR</tt> as
148 * a conventional numerical type, with an interface modelled on
149 * <TT>java.math.BigInteger</tt>.  No subclasses of <TT>CR</tt>
150 * will be explicitly mentioned by such a program.
151 * <P>
152 * All standard arithmetic operations, as well as a few algebraic
153 * and transcendal functions are provided.  Constructive reals are
154 * immutable; thus all of these operations return a new constructive real.
155 * <P>
156 * A few uses will require explicit construction of approximation functions.
157 * The requires the construction of a subclass of <TT>CR</tt> with
158 * an overridden <TT>approximate</tt> function.  Note that <TT>approximate</tt>
159 * should only be defined, but never called.  <TT>get_appr</tt>
160 * provides the same functionality, but adds the caching necessary to obtain
161 * reasonable performance.
162 * <P>
163 * Any operation may throw <TT>com.hp.creals.AbortedException</tt> if the thread
164 * in which it is executing is interrupted.  (<TT>InterruptedException</tt>
165 * cannot be used for this purpose, since CR inherits from <TT>Number</tt>.)
166 * <P>
167 * Any operation may also throw <TT>com.hp.creals.PrecisionOverflowException</tt>
168 * If the precision request generated during any subcalculation overflows
169 * a 28-bit integer.  (This should be extremely unlikely, except as an
170 * outcome of a division by zero, or other erroneous computation.)
171 *
172 */
173 public abstract class CR extends Number {
174     // CR is the basic representation of a number.
175     // Abstractly this is a function for computing an approximation
176     // plus the current best approximation.
177     // We could do without the latter, but that would
178     // be atrociously slow.
179 
180 /**
181  * Indicates a constructive real operation was interrupted.
182  * Most constructive real operations may throw such an exception.
183  * This is unchecked, since Number methods may not raise checked
184  * exceptions.
185 */
186 public static class AbortedException extends RuntimeException {
AbortedException()187     public AbortedException() { super(); }
AbortedException(String s)188     public AbortedException(String s) { super(s); }
189 }
190 
191 /**
192  * Indicates that the number of bits of precision requested by
193  * a computation on constructive reals required more than 28 bits,
194  * and was thus in danger of overflowing an int.
195  * This is likely to be a symptom of a diverging computation,
196  * <I>e.g.</i> division by zero.
197 */
198 public static class PrecisionOverflowException extends RuntimeException {
PrecisionOverflowException()199     public PrecisionOverflowException() { super(); }
PrecisionOverflowException(String s)200     public PrecisionOverflowException(String s) { super(s); }
201 }
202 
203     // First some frequently used constants, so we don't have to
204     // recompute these all over the place.
205       static final BigInteger big0 = BigInteger.ZERO;
206       static final BigInteger big1 = BigInteger.ONE;
207       static final BigInteger bigm1 = BigInteger.valueOf(-1);
208       static final BigInteger big2 = BigInteger.valueOf(2);
209       static final BigInteger bigm2 = BigInteger.valueOf(-2);
210       static final BigInteger big3 = BigInteger.valueOf(3);
211       static final BigInteger big6 = BigInteger.valueOf(6);
212       static final BigInteger big8 = BigInteger.valueOf(8);
213       static final BigInteger big10 = BigInteger.TEN;
214       static final BigInteger big750 = BigInteger.valueOf(750);
215       static final BigInteger bigm750 = BigInteger.valueOf(-750);
216 
217 /**
218 * Setting this to true requests that  all computations be aborted by
219 * throwing AbortedException.  Must be rest to false before any further
220 * computation.  Ideally Thread.interrupt() should be used instead, but
221 * that doesn't appear to be consistently supported by browser VMs.
222 */
223 public volatile static boolean please_stop = false;
224 
225 /**
226 * Must be defined in subclasses of <TT>CR</tt>.
227 * Most users can ignore the existence of this method, and will
228 * not ever need to define a <TT>CR</tt> subclass.
229 * Returns value / 2 ** precision rounded to an integer.
230 * The error in the result is strictly < 1.
231 * Informally, approximate(n) gives a scaled approximation
232 * accurate to 2**n.
233 * Implementations may safely assume that precision is
234 * at least a factor of 8 away from overflow.
235 * Called only with the lock on the <TT>CR</tt> object
236 * already held.
237 */
approximate(int precision)238       protected abstract BigInteger approximate(int precision);
239       transient int min_prec;
240         // The smallest precision value with which the above
241         // has been called.
242       transient BigInteger max_appr;
243         // The scaled approximation corresponding to min_prec.
244       transient boolean appr_valid = false;
245         // min_prec and max_val are valid.
246 
247     // Helper functions
bound_log2(int n)248       static int bound_log2(int n) {
249         int abs_n = Math.abs(n);
250         return (int)Math.ceil(Math.log((double)(abs_n + 1))/Math.log(2.0));
251       }
252       // Check that a precision is at least a factor of 8 away from
253       // overflowng the integer used to hold a precision spec.
254       // We generally perform this check early on, and then convince
255       // ourselves that none of the operations performed on precisions
256       // inside a function can generate an overflow.
check_prec(int n)257       static void check_prec(int n) {
258         int high = n >> 28;
259         // if n is not in danger of overflowing, then the 4 high order
260         // bits should be identical.  Thus high is either 0 or -1.
261         // The rest of this is to test for either of those in a way
262         // that should be as cheap as possible.
263         int high_shifted = n >> 29;
264         if (0 != (high ^ high_shifted)) {
265             throw new PrecisionOverflowException();
266         }
267       }
268 
269 /**
270 * The constructive real number corresponding to a
271 * <TT>BigInteger</tt>.
272 */
valueOf(BigInteger n)273       public static CR valueOf(BigInteger n) {
274         return new int_CR(n);
275       }
276 
277 /**
278 * The constructive real number corresponding to a
279 * Java <TT>int</tt>.
280 */
valueOf(int n)281       public static CR valueOf(int n) {
282         return valueOf(BigInteger.valueOf(n));
283       }
284 
285 /**
286 * The constructive real number corresponding to a
287 * Java <TT>long</tt>.
288 */
valueOf(long n)289       public static CR valueOf(long n) {
290         return valueOf(BigInteger.valueOf(n));
291       }
292 
293 /**
294 * The constructive real number corresponding to a
295 * Java <TT>double</tt>.
296 * The result is undefined if argument is infinite or NaN.
297 */
valueOf(double n)298       public static CR valueOf(double n) {
299         if (Double.isNaN(n)) throw new ArithmeticException("Nan argument");
300         if (Double.isInfinite(n)) {
301             throw new ArithmeticException("Infinite argument");
302         }
303         boolean negative = (n < 0.0);
304         long bits = Double.doubleToLongBits(Math.abs(n));
305         long mantissa = (bits & 0xfffffffffffffL);
306         int biased_exp = (int)(bits >> 52);
307         int exp = biased_exp - 1075;
308         if (biased_exp != 0) {
309             mantissa += (1L << 52);
310         } else {
311             mantissa <<= 1;
312         }
313         CR result = valueOf(mantissa).shiftLeft(exp);
314         if (negative) result = result.negate();
315         return result;
316       }
317 
318 /**
319 * The constructive real number corresponding to a
320 * Java <TT>float</tt>.
321 * The result is undefined if argument is infinite or NaN.
322 */
valueOf(float n)323       public static CR valueOf(float n) {
324         return valueOf((double) n);
325       }
326 
327       public static CR ZERO = valueOf(0);
328       public static CR ONE = valueOf(1);
329 
330     // Multiply k by 2**n.
shift(BigInteger k, int n)331       static BigInteger shift(BigInteger k, int n) {
332         if (n == 0) return k;
333         if (n < 0) return k.shiftRight(-n);
334         return k.shiftLeft(n);
335       }
336 
337     // Multiply by 2**n, rounding result
scale(BigInteger k, int n)338       static BigInteger scale(BigInteger k, int n) {
339         if (n >= 0) {
340             return k.shiftLeft(n);
341         } else {
342             BigInteger adj_k = shift(k, n+1).add(big1);
343             return adj_k.shiftRight(1);
344         }
345       }
346 
347     // Identical to approximate(), but maintain and update cache.
348 /**
349 * Returns value / 2 ** prec rounded to an integer.
350 * The error in the result is strictly < 1.
351 * Produces the same answer as <TT>approximate</tt>, but uses and
352 * maintains a cached approximation.
353 * Normally not overridden, and called only from <TT>approximate</tt>
354 * methods in subclasses.  Not needed if the provided operations
355 * on constructive reals suffice.
356 */
get_appr(int precision)357       public synchronized BigInteger get_appr(int precision) {
358         check_prec(precision);
359         if (appr_valid && precision >= min_prec) {
360             return scale(max_appr, min_prec - precision);
361         } else {
362             BigInteger result = approximate(precision);
363             min_prec = precision;
364             max_appr = result;
365             appr_valid = true;
366             return result;
367         }
368       }
369 
370     // Return the position of the msd.
371     // If x.msd() == n then
372     // 2**(n-1) < abs(x) < 2**(n+1)
373     // This initial version assumes that max_appr is valid
374     // and sufficiently removed from zero
375     // that the msd is determined.
known_msd()376       int known_msd() {
377         int first_digit;
378         int length;
379         if (max_appr.signum() >= 0) {
380             length = max_appr.bitLength();
381         } else {
382             length = max_appr.negate().bitLength();
383         }
384         first_digit = min_prec + length - 1;
385         return first_digit;
386       }
387 
388     // This version may return Integer.MIN_VALUE if the correct
389     // answer is < n.
msd(int n)390       int msd(int n) {
391         if (!appr_valid ||
392                 max_appr.compareTo(big1) <= 0
393                 && max_appr.compareTo(bigm1) >= 0) {
394             get_appr(n - 1);
395             if (max_appr.abs().compareTo(big1) <= 0) {
396                 // msd could still be arbitrarily far to the right.
397                 return Integer.MIN_VALUE;
398             }
399         }
400         return known_msd();
401       }
402 
403 
404     // Functionally equivalent, but iteratively evaluates to higher
405     // precision.
iter_msd(int n)406       int iter_msd(int n)
407       {
408         int prec = 0;
409 
410         for (;prec > n + 30; prec = (prec * 3)/2 - 16) {
411             int msd = msd(prec);
412             if (msd != Integer.MIN_VALUE) return msd;
413             check_prec(prec);
414             if (Thread.interrupted() || please_stop) {
415                 throw new AbortedException();
416             }
417         }
418         return msd(n);
419       }
420 
421     // This version returns a correct answer eventually, except
422     // that it loops forever (or throws an exception when the
423     // requested precision overflows) if this constructive real is zero.
msd()424       int msd() {
425           return iter_msd(Integer.MIN_VALUE);
426       }
427 
428     // A helper function for toString.
429     // Generate a String containing n zeroes.
zeroes(int n)430       private static String zeroes(int n) {
431         char[] a = new char[n];
432         for (int i = 0; i < n; ++i) {
433             a[i] = '0';
434         }
435         return new String(a);
436       }
437 
438     // Natural log of 2.  Needed for some prescaling below.
439     // ln(2) = 7ln(10/9) - 2ln(25/24) + 3ln(81/80)
simple_ln()440         CR simple_ln() {
441             return new prescaled_ln_CR(this.subtract(ONE));
442         }
443         static CR ten_ninths = valueOf(10).divide(valueOf(9));
444         static CR twentyfive_twentyfourths = valueOf(25).divide(valueOf(24));
445         static CR eightyone_eightyeths = valueOf(81).divide(valueOf(80));
446         static CR ln2_1 = valueOf(7).multiply(ten_ninths.simple_ln());
447         static CR ln2_2 =
448                 valueOf(2).multiply(twentyfive_twentyfourths.simple_ln());
449         static CR ln2_3 = valueOf(3).multiply(eightyone_eightyeths.simple_ln());
450         static CR ln2 = ln2_1.subtract(ln2_2).add(ln2_3);
451 
452     // Atan of integer reciprocal.  Used for atan_PI.  Could perhaps be made
453     // public.
atan_reciprocal(int n)454         static CR atan_reciprocal(int n) {
455             return new integral_atan_CR(n);
456         }
457     // Other constants used for PI computation.
458         static CR four = valueOf(4);
459 
460   // Public operations.
461 /**
462 * Return 0 if x = y to within the indicated tolerance,
463 * -1 if x < y, and +1 if x > y.  If x and y are indeed
464 * equal, it is guaranteed that 0 will be returned.  If
465 * they differ by less than the tolerance, anything
466 * may happen.  The tolerance allowed is
467 * the maximum of (abs(this)+abs(x))*(2**r) and 2**a
468 *       @param x        The other constructive real
469 *       @param r        Relative tolerance in bits
470 *       @param a        Absolute tolerance in bits
471 */
compareTo(CR x, int r, int a)472       public int compareTo(CR x, int r, int a) {
473         int this_msd = iter_msd(a);
474         int x_msd = x.iter_msd(this_msd > a? this_msd : a);
475         int max_msd = (x_msd > this_msd? x_msd : this_msd);
476         int rel = max_msd + r;
477             // This can't approach overflow, since r and a are
478             // effectively divided by 2, and msds are checked.
479         int abs_prec = (rel > a? rel : a);
480         return compareTo(x, abs_prec);
481       }
482 
483 /**
484 * Approximate comparison with only an absolute tolerance.
485 * Identical to the three argument version, but without a relative
486 * tolerance.
487 * Result is 0 if both constructive reals are equal, indeterminate
488 * if they differ by less than 2**a.
489 *
490 *       @param x        The other constructive real
491 *       @param a        Absolute tolerance in bits
492 */
compareTo(CR x, int a)493       public int compareTo(CR x, int a) {
494         int needed_prec = a - 1;
495         BigInteger this_appr = get_appr(needed_prec);
496         BigInteger x_appr = x.get_appr(needed_prec);
497         int comp1 = this_appr.compareTo(x_appr.add(big1));
498         if (comp1 > 0) return 1;
499         int comp2 = this_appr.compareTo(x_appr.subtract(big1));
500         if (comp2 < 0) return -1;
501         return 0;
502       }
503 
504 /**
505 * Return -1 if <TT>this &lt; x</tt>, or +1 if <TT>this &gt; x</tt>.
506 * Should be called only if <TT>this != x</tt>.
507 * If <TT>this == x</tt>, this will not terminate correctly; typically it
508 * will run until it exhausts memory.
509 * If the two constructive reals may be equal, the two or 3 argument
510 * version of compareTo should be used.
511 */
compareTo(CR x)512       public int compareTo(CR x) {
513         for (int a = -20; ; a *= 2) {
514             check_prec(a);
515             int result = compareTo(x, a);
516             if (0 != result) return result;
517             if (Thread.interrupted() || please_stop) {
518                 throw new AbortedException();
519             }
520         }
521       }
522 
523 /**
524 * Equivalent to <TT>compareTo(CR.valueOf(0), a)</tt>
525 */
signum(int a)526       public int signum(int a) {
527         if (appr_valid) {
528             int quick_try = max_appr.signum();
529             if (0 != quick_try) return quick_try;
530         }
531         int needed_prec = a - 1;
532         BigInteger this_appr = get_appr(needed_prec);
533         return this_appr.signum();
534       }
535 
536 /**
537 * Return -1 if negative, +1 if positive.
538 * Should be called only if <TT>this != 0</tt>.
539 * In the 0 case, this will not terminate correctly; typically it
540 * will run until it exhausts memory.
541 * If the two constructive reals may be equal, the one or two argument
542 * version of signum should be used.
543 */
signum()544       public int signum() {
545         for (int a = -20; ; a *= 2) {
546             check_prec(a);
547             int result = signum(a);
548             if (0 != result) return result;
549             if (Thread.interrupted() || please_stop) {
550                 throw new AbortedException();
551             }
552         }
553       }
554 
555 /**
556 * Return the constructive real number corresponding to the given
557 * textual representation and radix.
558 *
559 *       @param s        [-] digit* [. digit*]
560 *       @param radix
561 */
562 
valueOf(String s, int radix)563       public static CR valueOf(String s, int radix)
564              throws NumberFormatException {
565           int len = s.length();
566           int start_pos = 0, point_pos;
567           String fraction;
568           while (s.charAt(start_pos) == ' ') ++start_pos;
569           while (s.charAt(len - 1) == ' ') --len;
570           point_pos = s.indexOf('.', start_pos);
571           if (point_pos == -1) {
572               point_pos = len;
573               fraction = "0";
574           } else {
575               fraction = s.substring(point_pos + 1, len);
576           }
577           String whole = s.substring(start_pos, point_pos);
578           BigInteger scaled_result = new BigInteger(whole + fraction, radix);
579           BigInteger divisor = BigInteger.valueOf(radix).pow(fraction.length());
580           return CR.valueOf(scaled_result).divide(CR.valueOf(divisor));
581       }
582 
583 /**
584 * Return a textual representation accurate to <TT>n</tt> places
585 * to the right of the decimal point.  <TT>n</tt> must be nonnegative.
586 *
587 *       @param  n       Number of digits (>= 0) included to the right of decimal point
588 *       @param  radix   Base ( >= 2, <= 16) for the resulting representation.
589 */
toString(int n, int radix)590       public String toString(int n, int radix) {
591           CR scaled_CR;
592           if (16 == radix) {
593             scaled_CR = shiftLeft(4*n);
594           } else {
595             BigInteger scale_factor = BigInteger.valueOf(radix).pow(n);
596             scaled_CR = multiply(new int_CR(scale_factor));
597           }
598           BigInteger scaled_int = scaled_CR.get_appr(0);
599           String scaled_string = scaled_int.abs().toString(radix);
600           String result;
601           if (0 == n) {
602               result = scaled_string;
603           } else {
604               int len = scaled_string.length();
605               if (len <= n) {
606                 // Add sufficient leading zeroes
607                   String z = zeroes(n + 1 - len);
608                   scaled_string = z + scaled_string;
609                   len = n + 1;
610               }
611               String whole = scaled_string.substring(0, len - n);
612               String fraction = scaled_string.substring(len - n);
613               result = whole + "." + fraction;
614           }
615           if (scaled_int.signum() < 0) {
616               result = "-" + result;
617           }
618           return result;
619       }
620 
621 
622 /**
623 * Equivalent to <TT>toString(n,10)</tt>
624 *
625 *       @param  n       Number of digits included to the right of decimal point
626 */
toString(int n)627     public String toString(int n) {
628         return toString(n, 10);
629     }
630 
631 /**
632 * Equivalent to <TT>toString(10, 10)</tt>
633 */
toString()634     public String toString() {
635         return toString(10);
636     }
637 
638     static double doubleLog2 = Math.log(2.0);
639 /**
640 * Return a textual scientific notation representation accurate
641 * to <TT>n</tt> places to the right of the decimal point.
642 * <TT>n</tt> must be nonnegative.  A value smaller than
643 * <TT>radix</tt>**-<TT>m</tt> may be displayed as 0.
644 * The <TT>mantissa</tt> component of the result is either "0"
645 * or exactly <TT>n</tt> digits long.  The <TT>sign</tt>
646 * component is zero exactly when the mantissa is "0".
647 *
648 *       @param  n       Number of digits (&gt; 0) included to the right of decimal point.
649 *       @param  radix   Base ( &ge; 2, &le; 16) for the resulting representation.
650 *       @param  m       Precision used to distinguish number from zero.
651 *                       Expressed as a power of m.
652 */
toStringFloatRep(int n, int radix, int m)653     public StringFloatRep toStringFloatRep(int n, int radix, int m) {
654         if (n <= 0) throw new ArithmeticException("Bad precision argument");
655         double log2_radix = Math.log((double)radix)/doubleLog2;
656         BigInteger big_radix = BigInteger.valueOf(radix);
657         long long_msd_prec = (long)(log2_radix * (double)m);
658         if (long_msd_prec > (long)Integer.MAX_VALUE
659             || long_msd_prec < (long)Integer.MIN_VALUE)
660             throw new PrecisionOverflowException();
661         int msd_prec = (int)long_msd_prec;
662         check_prec(msd_prec);
663         int msd = iter_msd(msd_prec - 2);
664         if (msd == Integer.MIN_VALUE)
665             return new StringFloatRep(0, "0", radix, 0);
666         int exponent = (int)Math.ceil((double)msd / log2_radix);
667                 // Guess for the exponent.  Try to get it usually right.
668         int scale_exp = exponent - n;
669         CR scale;
670         if (scale_exp > 0) {
671             scale = CR.valueOf(big_radix.pow(scale_exp)).inverse();
672         } else {
673             scale = CR.valueOf(big_radix.pow(-scale_exp));
674         }
675         CR scaled_res = multiply(scale);
676         BigInteger scaled_int = scaled_res.get_appr(0);
677         int sign = scaled_int.signum();
678         String scaled_string = scaled_int.abs().toString(radix);
679         while (scaled_string.length() < n) {
680             // exponent was too large.  Adjust.
681             scaled_res = scaled_res.multiply(CR.valueOf(big_radix));
682             exponent -= 1;
683             scaled_int = scaled_res.get_appr(0);
684             sign = scaled_int.signum();
685             scaled_string = scaled_int.abs().toString(radix);
686         }
687         if (scaled_string.length() > n) {
688             // exponent was too small.  Adjust by truncating.
689             exponent += (scaled_string.length() - n);
690             scaled_string = scaled_string.substring(0, n);
691         }
692         return new StringFloatRep(sign, scaled_string, radix, exponent);
693     }
694 
695 /**
696 * Return a BigInteger which differs by less than one from the
697 * constructive real.
698 */
BigIntegerValue()699     public BigInteger BigIntegerValue() {
700         return get_appr(0);
701     }
702 
703 /**
704 * Return an int which differs by less than one from the
705 * constructive real.  Behavior on overflow is undefined.
706 */
intValue()707     public int intValue() {
708         return BigIntegerValue().intValue();
709     }
710 
711 /**
712 * Return an int which differs by less than one from the
713 * constructive real.  Behavior on overflow is undefined.
714 */
byteValue()715     public byte byteValue() {
716         return BigIntegerValue().byteValue();
717     }
718 
719 /**
720 * Return a long which differs by less than one from the
721 * constructive real.  Behavior on overflow is undefined.
722 */
longValue()723     public long longValue() {
724         return BigIntegerValue().longValue();
725     }
726 
727 /**
728 * Return a double which differs by less than one in the least
729 * represented bit from the constructive real.
730 * (We're in fact closer to round-to-nearest than that, but we can't and
731 * don't promise correct rounding.)
732 */
doubleValue()733     public double doubleValue() {
734         int my_msd = iter_msd(-1080 /* slightly > exp. range */);
735         if (Integer.MIN_VALUE == my_msd) return 0.0;
736         int needed_prec = my_msd - 60;
737         double scaled_int = get_appr(needed_prec).doubleValue();
738         boolean may_underflow = (needed_prec < -1000);
739         long scaled_int_rep = Double.doubleToLongBits(scaled_int);
740         long exp_adj = may_underflow? needed_prec + 96 : needed_prec;
741         long orig_exp = (scaled_int_rep >> 52) & 0x7ff;
742         if (((orig_exp + exp_adj) & ~0x7ff) != 0) {
743             // Original unbiased exponent is > 50. Exp_adj > -1050.
744             // Thus this can overflow the 11 bit exponent only if the result
745             // itself overflows.
746             if (scaled_int < 0.0) {
747                 return Double.NEGATIVE_INFINITY;
748             } else {
749                 return Double.POSITIVE_INFINITY;
750             }
751         }
752         scaled_int_rep += exp_adj << 52;
753         double result = Double.longBitsToDouble(scaled_int_rep);
754         if (may_underflow) {
755             double two48 = (double)(1L << 48);
756             return result/two48/two48;
757         } else {
758             return result;
759         }
760     }
761 
762 /**
763 * Return a float which differs by less than one in the least
764 * represented bit from the constructive real.
765 */
floatValue()766     public float floatValue() {
767         return (float)doubleValue();
768         // Note that double-rounding is not a problem here, since we
769         // cannot, and do not, guarantee correct rounding.
770     }
771 
772 /**
773 * Add two constructive reals.
774 */
add(CR x)775     public CR add(CR x) {
776         return new add_CR(this, x);
777     }
778 
779 /**
780 * Multiply a constructive real by 2**n.
781 * @param n      shift count, may be negative
782 */
shiftLeft(int n)783     public CR shiftLeft(int n) {
784         check_prec(n);
785         return new shifted_CR(this, n);
786     }
787 
788 /**
789 * Multiply a constructive real by 2**(-n).
790 * @param n      shift count, may be negative
791 */
shiftRight(int n)792     public CR shiftRight(int n) {
793         check_prec(n);
794         return new shifted_CR(this, -n);
795     }
796 
797 /**
798 * Produce a constructive real equivalent to the original, assuming
799 * the original was an integer.  Undefined results if the original
800 * was not an integer.  Prevents evaluation of digits to the right
801 * of the decimal point, and may thus improve performance.
802 */
assumeInt()803     public CR assumeInt() {
804         return new assumed_int_CR(this);
805     }
806 
807 /**
808 * The additive inverse of a constructive real
809 */
negate()810     public CR negate() {
811         return new neg_CR(this);
812     }
813 
814 /**
815 * The difference between two constructive reals
816 */
subtract(CR x)817     public CR subtract(CR x) {
818         return new add_CR(this, x.negate());
819     }
820 
821 /**
822 * The product of two constructive reals
823 */
multiply(CR x)824     public CR multiply(CR x) {
825         return new mult_CR(this, x);
826     }
827 
828 /**
829 * The multiplicative inverse of a constructive real.
830 * <TT>x.inverse()</tt> is equivalent to <TT>CR.valueOf(1).divide(x)</tt>.
831 */
inverse()832     public CR inverse() {
833         return new inv_CR(this);
834     }
835 
836 /**
837 * The quotient of two constructive reals.
838 */
divide(CR x)839     public CR divide(CR x) {
840         return new mult_CR(this, x.inverse());
841     }
842 
843 /**
844 * The real number <TT>x</tt> if <TT>this</tt> < 0, or <TT>y</tt> otherwise.
845 * Requires <TT>x</tt> = <TT>y</tt> if <TT>this</tt> = 0.
846 * Since comparisons may diverge, this is often
847 * a useful alternative to conditionals.
848 */
select(CR x, CR y)849     public CR select(CR x, CR y) {
850         return new select_CR(this, x, y);
851     }
852 
853 /**
854 * The maximum of two constructive reals.
855 */
max(CR x)856     public CR max(CR x) {
857         return subtract(x).select(x, this);
858     }
859 
860 /**
861 * The minimum of two constructive reals.
862 */
min(CR x)863     public CR min(CR x) {
864         return subtract(x).select(this, x);
865     }
866 
867 /**
868 * The absolute value of a constructive reals.
869 * Note that this cannot be written as a conditional.
870 */
abs()871     public CR abs() {
872         return select(negate(), this);
873     }
874 
875 /**
876 * The exponential function, that is e**<TT>this</tt>.
877 */
exp()878     public CR exp() {
879         final int low_prec = -10;
880         BigInteger rough_appr = get_appr(low_prec);
881         // Handle negative arguments directly; negating and computing inverse
882         // can be very expensive.
883         if (rough_appr.compareTo(big2) > 0 || rough_appr.compareTo(bigm2) < 0) {
884             CR square_root = shiftRight(1).exp();
885             return square_root.multiply(square_root);
886         } else {
887             return new prescaled_exp_CR(this);
888         }
889     }
890 
891 /**
892 * The ratio of a circle's circumference to its diameter.
893 */
894     public static CR PI = new gl_pi_CR();
895 
896     // Our old PI implementation. Keep this around for now to allow checking.
897     // This implementation may also be faster for BigInteger implementations
898     // that support only quadratic multiplication, but exhibit high performance
899     // for small computations.  (The standard Android 6 implementation supports
900     // subquadratic multiplication, but has high constant overhead.) Many other
901     // atan-based formulas are possible, but based on superficial
902     // experimentation, this is roughly as good as the more complex formulas.
903     public static CR atan_PI = four.multiply(four.multiply(atan_reciprocal(5))
904                                             .subtract(atan_reciprocal(239)));
905         // pi/4 = 4*atan(1/5) - atan(1/239)
906     static CR half_pi = PI.shiftRight(1);
907 
908 /**
909 * The trigonometric cosine function.
910 */
cos()911     public CR cos() {
912         BigInteger halfpi_multiples = divide(PI).get_appr(-1);
913         BigInteger abs_halfpi_multiples = halfpi_multiples.abs();
914         if (abs_halfpi_multiples.compareTo(big2) >= 0) {
915             // Subtract multiples of PI
916             BigInteger pi_multiples = scale(halfpi_multiples, -1);
917             CR adjustment = PI.multiply(CR.valueOf(pi_multiples));
918             if (pi_multiples.and(big1).signum() != 0) {
919                 return subtract(adjustment).cos().negate();
920             } else {
921                 return subtract(adjustment).cos();
922             }
923         } else if (get_appr(-1).abs().compareTo(big2) >= 0) {
924             // Scale further with double angle formula
925             CR cos_half = shiftRight(1).cos();
926             return cos_half.multiply(cos_half).shiftLeft(1).subtract(ONE);
927         } else {
928             return new prescaled_cos_CR(this);
929         }
930     }
931 
932 /**
933 * The trigonometric sine function.
934 */
sin()935     public CR sin() {
936         return half_pi.subtract(this).cos();
937     }
938 
939 /**
940 * The trignonometric arc (inverse) sine function.
941 */
asin()942     public CR asin() {
943         BigInteger rough_appr = get_appr(-10);
944         if (rough_appr.compareTo(big750) /* 1/sqrt(2) + a bit */ > 0){
945             CR new_arg = ONE.subtract(multiply(this)).sqrt();
946             return new_arg.acos();
947         } else if (rough_appr.compareTo(bigm750) < 0) {
948             return negate().asin().negate();
949         } else {
950             return new prescaled_asin_CR(this);
951         }
952     }
953 
954 /**
955 * The trignonometric arc (inverse) cosine function.
956 */
acos()957     public CR acos() {
958         return half_pi.subtract(asin());
959     }
960 
961     static final BigInteger low_ln_limit = big8; /* sixteenths, i.e. 1/2 */
962     static final BigInteger high_ln_limit =
963                         BigInteger.valueOf(16 + 8 /* 1.5 */);
964     static final BigInteger scaled_4 =
965                         BigInteger.valueOf(4*16);
966 
967 /**
968 * The natural (base e) logarithm.
969 */
ln()970     public CR ln() {
971         final int low_prec = -4;
972         BigInteger rough_appr = get_appr(low_prec); /* In sixteenths */
973         if (rough_appr.compareTo(big0) < 0) {
974             throw new ArithmeticException("ln(negative)");
975         }
976         if (rough_appr.compareTo(low_ln_limit) <= 0) {
977             return inverse().ln().negate();
978         }
979         if (rough_appr.compareTo(high_ln_limit) >= 0) {
980             if (rough_appr.compareTo(scaled_4) <= 0) {
981                 CR quarter = sqrt().sqrt().ln();
982                 return quarter.shiftLeft(2);
983             } else {
984                 int extra_bits = rough_appr.bitLength() - 3;
985                 CR scaled_result = shiftRight(extra_bits).ln();
986                 return scaled_result.add(CR.valueOf(extra_bits).multiply(ln2));
987             }
988         }
989         return simple_ln();
990     }
991 
992 /**
993 * The square root of a constructive real.
994 */
sqrt()995     public CR sqrt() {
996         return new sqrt_CR(this);
997     }
998 
999 }  // end of CR
1000 
1001 
1002 //
1003 // A specialization of CR for cases in which approximate() calls
1004 // to increase evaluation precision are somewhat expensive.
1005 // If we need to (re)evaluate, we speculatively evaluate to slightly
1006 // higher precision, miminimizing reevaluations.
1007 // Note that this requires any arguments to be evaluated to higher
1008 // precision than absolutely necessary.  It can thus potentially
1009 // result in lots of wasted effort, and should be used judiciously.
1010 // This assumes that the order of magnitude of the number is roughly one.
1011 //
1012 abstract class slow_CR extends CR {
1013     static int max_prec = -64;
1014     static int prec_incr = 32;
get_appr(int precision)1015     public synchronized BigInteger get_appr(int precision) {
1016         check_prec(precision);
1017         if (appr_valid && precision >= min_prec) {
1018             return scale(max_appr, min_prec - precision);
1019         } else {
1020             int eval_prec = (precision >= max_prec? max_prec :
1021                              (precision - prec_incr + 1) & ~(prec_incr - 1));
1022             BigInteger result = approximate(eval_prec);
1023             min_prec = eval_prec;
1024             max_appr = result;
1025             appr_valid = true;
1026             return scale(result, eval_prec - precision);
1027         }
1028     }
1029 }
1030 
1031 
1032 // Representation of an integer constant.  Private.
1033 class int_CR extends CR {
1034     BigInteger value;
int_CR(BigInteger n)1035     int_CR(BigInteger n) {
1036         value = n;
1037     }
approximate(int p)1038     protected BigInteger approximate(int p) {
1039         return scale(value, -p) ;
1040     }
1041 }
1042 
1043 // Representation of a number that may not have been completely
1044 // evaluated, but is assumed to be an integer.  Hence we never
1045 // evaluate beyond the decimal point.
1046 class assumed_int_CR extends CR {
1047     CR value;
assumed_int_CR(CR x)1048     assumed_int_CR(CR x) {
1049         value = x;
1050     }
approximate(int p)1051     protected BigInteger approximate(int p) {
1052         if (p >= 0) {
1053             return value.get_appr(p);
1054         } else {
1055             return scale(value.get_appr(0), -p) ;
1056         }
1057     }
1058 }
1059 
1060 // Representation of the sum of 2 constructive reals.  Private.
1061 class add_CR extends CR {
1062     CR op1;
1063     CR op2;
add_CR(CR x, CR y)1064     add_CR(CR x, CR y) {
1065         op1 = x;
1066         op2 = y;
1067     }
approximate(int p)1068     protected BigInteger approximate(int p) {
1069         // Args need to be evaluated so that each error is < 1/4 ulp.
1070         // Rounding error from the cale call is <= 1/2 ulp, so that
1071         // final error is < 1 ulp.
1072         return scale(op1.get_appr(p-2).add(op2.get_appr(p-2)), -2);
1073     }
1074 }
1075 
1076 // Representation of a CR multiplied by 2**n
1077 class shifted_CR extends CR {
1078     CR op;
1079     int count;
shifted_CR(CR x, int n)1080     shifted_CR(CR x, int n) {
1081         op = x;
1082         count = n;
1083     }
approximate(int p)1084     protected BigInteger approximate(int p) {
1085         return op.get_appr(p - count);
1086     }
1087 }
1088 
1089 // Representation of the negation of a constructive real.  Private.
1090 class neg_CR extends CR {
1091     CR op;
neg_CR(CR x)1092     neg_CR(CR x) {
1093         op = x;
1094     }
approximate(int p)1095     protected BigInteger approximate(int p) {
1096         return op.get_appr(p).negate();
1097     }
1098 }
1099 
1100 // Representation of:
1101 //      op1     if selector < 0
1102 //      op2     if selector >= 0
1103 // Assumes x = y if s = 0
1104 class select_CR extends CR {
1105     CR selector;
1106     int selector_sign;
1107     CR op1;
1108     CR op2;
select_CR(CR s, CR x, CR y)1109     select_CR(CR s, CR x, CR y) {
1110         selector = s;
1111         int selector_sign = selector.get_appr(-20).signum();
1112         op1 = x;
1113         op2 = y;
1114     }
approximate(int p)1115     protected BigInteger approximate(int p) {
1116         if (selector_sign < 0) return op1.get_appr(p);
1117         if (selector_sign > 0) return op2.get_appr(p);
1118         BigInteger op1_appr = op1.get_appr(p-1);
1119         BigInteger op2_appr = op2.get_appr(p-1);
1120         BigInteger diff = op1_appr.subtract(op2_appr).abs();
1121         if (diff.compareTo(big1) <= 0) {
1122             // close enough; use either
1123             return scale(op1_appr, -1);
1124         }
1125         // op1 and op2 are different; selector != 0;
1126         // safe to get sign of selector.
1127         if (selector.signum() < 0) {
1128             selector_sign = -1;
1129             return scale(op1_appr, -1);
1130         } else {
1131             selector_sign = 1;
1132             return scale(op2_appr, -1);
1133         }
1134     }
1135 }
1136 
1137 // Representation of the product of 2 constructive reals. Private.
1138 class mult_CR extends CR {
1139     CR op1;
1140     CR op2;
mult_CR(CR x, CR y)1141     mult_CR(CR x, CR y) {
1142         op1 = x;
1143         op2 = y;
1144     }
approximate(int p)1145     protected BigInteger approximate(int p) {
1146         int half_prec = (p >> 1) - 1;
1147         int msd_op1 = op1.msd(half_prec);
1148         int msd_op2;
1149 
1150         if (msd_op1 == Integer.MIN_VALUE) {
1151             msd_op2 = op2.msd(half_prec);
1152             if (msd_op2 == Integer.MIN_VALUE) {
1153                 // Product is small enough that zero will do as an
1154                 // approximation.
1155                 return big0;
1156             } else {
1157                 // Swap them, so the larger operand (in absolute value)
1158                 // is first.
1159                 CR tmp;
1160                 tmp = op1;
1161                 op1 = op2;
1162                 op2 = tmp;
1163                 msd_op1 = msd_op2;
1164             }
1165         }
1166         // msd_op1 is valid at this point.
1167         int prec2 = p - msd_op1 - 3;    // Precision needed for op2.
1168                 // The appr. error is multiplied by at most
1169                 // 2 ** (msd_op1 + 1)
1170                 // Thus each approximation contributes 1/4 ulp
1171                 // to the rounding error, and the final rounding adds
1172                 // another 1/2 ulp.
1173         BigInteger appr2 = op2.get_appr(prec2);
1174         if (appr2.signum() == 0) return big0;
1175         msd_op2 = op2.known_msd();
1176         int prec1 = p - msd_op2 - 3;    // Precision needed for op1.
1177         BigInteger appr1 = op1.get_appr(prec1);
1178         int scale_digits =  prec1 + prec2 - p;
1179         return scale(appr1.multiply(appr2), scale_digits);
1180     }
1181 }
1182 
1183 // Representation of the multiplicative inverse of a constructive
1184 // real.  Private.  Should use Newton iteration to refine estimates.
1185 class inv_CR extends CR {
1186     CR op;
inv_CR(CR x)1187     inv_CR(CR x) { op = x; }
approximate(int p)1188     protected BigInteger approximate(int p) {
1189         int msd = op.msd();
1190         int inv_msd = 1 - msd;
1191         int digits_needed = inv_msd - p + 3;
1192                                 // Number of SIGNIFICANT digits needed for
1193                                 // argument, excl. msd position, which may
1194                                 // be fictitious, since msd routine can be
1195                                 // off by 1.  Roughly 1 extra digit is
1196                                 // needed since the relative error is the
1197                                 // same in the argument and result, but
1198                                 // this isn't quite the same as the number
1199                                 // of significant digits.  Another digit
1200                                 // is needed to compensate for slop in the
1201                                 // calculation.
1202                                 // One further bit is required, since the
1203                                 // final rounding introduces a 0.5 ulp
1204                                 // error.
1205         int prec_needed = msd - digits_needed;
1206         int log_scale_factor = -p - prec_needed;
1207         if (log_scale_factor < 0) return big0;
1208         BigInteger dividend = big1.shiftLeft(log_scale_factor);
1209         BigInteger scaled_divisor = op.get_appr(prec_needed);
1210         BigInteger abs_scaled_divisor = scaled_divisor.abs();
1211         BigInteger adj_dividend = dividend.add(
1212                                         abs_scaled_divisor.shiftRight(1));
1213                 // Adjustment so that final result is rounded.
1214         BigInteger result = adj_dividend.divide(abs_scaled_divisor);
1215         if (scaled_divisor.signum() < 0) {
1216           return result.negate();
1217         } else {
1218           return result;
1219         }
1220     }
1221 }
1222 
1223 
1224 // Representation of the exponential of a constructive real.  Private.
1225 // Uses a Taylor series expansion.  Assumes |x| < 1/2.
1226 // Note: this is known to be a bad algorithm for
1227 // floating point.  Unfortunately, other alternatives
1228 // appear to require precomputed information.
1229 class prescaled_exp_CR extends CR {
1230     CR op;
prescaled_exp_CR(CR x)1231     prescaled_exp_CR(CR x) { op = x; }
approximate(int p)1232     protected BigInteger approximate(int p) {
1233         if (p >= 1) return big0;
1234         int iterations_needed = -p/2 + 2;  // conservative estimate > 0.
1235           //  Claim: each intermediate term is accurate
1236           //  to 2*2^calc_precision.
1237           //  Total rounding error in series computation is
1238           //  2*iterations_needed*2^calc_precision,
1239           //  exclusive of error in op.
1240         int calc_precision = p - bound_log2(2*iterations_needed)
1241                                - 4; // for error in op, truncation.
1242         int op_prec = p - 3;
1243         BigInteger op_appr = op.get_appr(op_prec);
1244           // Error in argument results in error of < 3/8 ulp.
1245           // Sum of term eval. rounding error is < 1/16 ulp.
1246           // Series truncation error < 1/16 ulp.
1247           // Final rounding error is <= 1/2 ulp.
1248           // Thus final error is < 1 ulp.
1249         BigInteger scaled_1 = big1.shiftLeft(-calc_precision);
1250         BigInteger current_term = scaled_1;
1251         BigInteger current_sum = scaled_1;
1252         int n = 0;
1253         BigInteger max_trunc_error =
1254                 big1.shiftLeft(p - 4 - calc_precision);
1255         while (current_term.abs().compareTo(max_trunc_error) >= 0) {
1256           if (Thread.interrupted() || please_stop) throw new AbortedException();
1257           n += 1;
1258           /* current_term = current_term * op / n */
1259           current_term = scale(current_term.multiply(op_appr), op_prec);
1260           current_term = current_term.divide(BigInteger.valueOf(n));
1261           current_sum = current_sum.add(current_term);
1262         }
1263         return scale(current_sum, calc_precision - p);
1264     }
1265 }
1266 
1267 // Representation of the cosine of a constructive real.  Private.
1268 // Uses a Taylor series expansion.  Assumes |x| < 1.
1269 class prescaled_cos_CR extends slow_CR {
1270     CR op;
prescaled_cos_CR(CR x)1271     prescaled_cos_CR(CR x) {
1272         op = x;
1273     }
approximate(int p)1274     protected BigInteger approximate(int p) {
1275         if (p >= 1) return big0;
1276         int iterations_needed = -p/2 + 4;  // conservative estimate > 0.
1277           //  Claim: each intermediate term is accurate
1278           //  to 2*2^calc_precision.
1279           //  Total rounding error in series computation is
1280           //  2*iterations_needed*2^calc_precision,
1281           //  exclusive of error in op.
1282         int calc_precision = p - bound_log2(2*iterations_needed)
1283                                - 4; // for error in op, truncation.
1284         int op_prec = p - 2;
1285         BigInteger op_appr = op.get_appr(op_prec);
1286           // Error in argument results in error of < 1/4 ulp.
1287           // Cumulative arithmetic rounding error is < 1/16 ulp.
1288           // Series truncation error < 1/16 ulp.
1289           // Final rounding error is <= 1/2 ulp.
1290           // Thus final error is < 1 ulp.
1291         BigInteger current_term;
1292         int n;
1293         BigInteger max_trunc_error =
1294                 big1.shiftLeft(p - 4 - calc_precision);
1295         n = 0;
1296         current_term = big1.shiftLeft(-calc_precision);
1297         BigInteger current_sum = current_term;
1298         while (current_term.abs().compareTo(max_trunc_error) >= 0) {
1299           if (Thread.interrupted() || please_stop) throw new AbortedException();
1300           n += 2;
1301           /* current_term = - current_term * op * op / n * (n - 1)   */
1302           current_term = scale(current_term.multiply(op_appr), op_prec);
1303           current_term = scale(current_term.multiply(op_appr), op_prec);
1304           BigInteger divisor = BigInteger.valueOf(-n)
1305                                   .multiply(BigInteger.valueOf(n-1));
1306           current_term = current_term.divide(divisor);
1307           current_sum = current_sum.add(current_term);
1308         }
1309         return scale(current_sum, calc_precision - p);
1310     }
1311 }
1312 
1313 // The constructive real atan(1/n), where n is a small integer
1314 // > base.
1315 // This gives a simple and moderately fast way to compute PI.
1316 class integral_atan_CR extends slow_CR {
1317     int op;
integral_atan_CR(int x)1318     integral_atan_CR(int x) { op = x; }
approximate(int p)1319     protected BigInteger approximate(int p) {
1320         if (p >= 1) return big0;
1321         int iterations_needed = -p/2 + 2;  // conservative estimate > 0.
1322           //  Claim: each intermediate term is accurate
1323           //  to 2*base^calc_precision.
1324           //  Total rounding error in series computation is
1325           //  2*iterations_needed*base^calc_precision,
1326           //  exclusive of error in op.
1327         int calc_precision = p - bound_log2(2*iterations_needed)
1328                                - 2; // for error in op, truncation.
1329           // Error in argument results in error of < 3/8 ulp.
1330           // Cumulative arithmetic rounding error is < 1/4 ulp.
1331           // Series truncation error < 1/4 ulp.
1332           // Final rounding error is <= 1/2 ulp.
1333           // Thus final error is < 1 ulp.
1334         BigInteger scaled_1 = big1.shiftLeft(-calc_precision);
1335         BigInteger big_op = BigInteger.valueOf(op);
1336         BigInteger big_op_squared = BigInteger.valueOf(op*op);
1337         BigInteger op_inverse = scaled_1.divide(big_op);
1338         BigInteger current_power = op_inverse;
1339         BigInteger current_term = op_inverse;
1340         BigInteger current_sum = op_inverse;
1341         int current_sign = 1;
1342         int n = 1;
1343         BigInteger max_trunc_error =
1344                 big1.shiftLeft(p - 2 - calc_precision);
1345         while (current_term.abs().compareTo(max_trunc_error) >= 0) {
1346           if (Thread.interrupted() || please_stop) throw new AbortedException();
1347           n += 2;
1348           current_power = current_power.divide(big_op_squared);
1349           current_sign = -current_sign;
1350           current_term =
1351             current_power.divide(BigInteger.valueOf(current_sign*n));
1352           current_sum = current_sum.add(current_term);
1353         }
1354         return scale(current_sum, calc_precision - p);
1355     }
1356 }
1357 
1358 // Representation for ln(1 + op)
1359 class prescaled_ln_CR extends slow_CR {
1360     CR op;
prescaled_ln_CR(CR x)1361     prescaled_ln_CR(CR x) { op = x; }
1362     // Compute an approximation of ln(1+x) to precision
1363     // prec. This assumes |x| < 1/2.
1364     // It uses a Taylor series expansion.
1365     // Unfortunately there appears to be no way to take
1366     // advantage of old information.
1367     // Note: this is known to be a bad algorithm for
1368     // floating point.  Unfortunately, other alternatives
1369     // appear to require precomputed tabular information.
approximate(int p)1370     protected BigInteger approximate(int p) {
1371         if (p >= 0) return big0;
1372         int iterations_needed = -p;  // conservative estimate > 0.
1373           //  Claim: each intermediate term is accurate
1374           //  to 2*2^calc_precision.  Total error is
1375           //  2*iterations_needed*2^calc_precision
1376           //  exclusive of error in op.
1377         int calc_precision = p - bound_log2(2*iterations_needed)
1378                                - 4; // for error in op, truncation.
1379         int op_prec = p - 3;
1380         BigInteger op_appr = op.get_appr(op_prec);
1381           // Error analysis as for exponential.
1382         BigInteger scaled_1 = big1.shiftLeft(-calc_precision);
1383         BigInteger x_nth = scale(op_appr, op_prec - calc_precision);
1384         BigInteger current_term = x_nth;  // x**n
1385         BigInteger current_sum = current_term;
1386         int n = 1;
1387         int current_sign = 1;   // (-1)^(n-1)
1388         BigInteger max_trunc_error =
1389                 big1.shiftLeft(p - 4 - calc_precision);
1390         while (current_term.abs().compareTo(max_trunc_error) >= 0) {
1391           if (Thread.interrupted() || please_stop) throw new AbortedException();
1392           n += 1;
1393           current_sign = -current_sign;
1394           x_nth = scale(x_nth.multiply(op_appr), op_prec);
1395           current_term = x_nth.divide(BigInteger.valueOf(n * current_sign));
1396                                 // x**n / (n * (-1)**(n-1))
1397           current_sum = current_sum.add(current_term);
1398         }
1399         return scale(current_sum, calc_precision - p);
1400     }
1401 }
1402 
1403 // Representation of the arcsine of a constructive real.  Private.
1404 // Uses a Taylor series expansion.  Assumes |x| < (1/2)^(1/3).
1405 class prescaled_asin_CR extends slow_CR {
1406     CR op;
prescaled_asin_CR(CR x)1407     prescaled_asin_CR(CR x) {
1408         op = x;
1409     }
approximate(int p)1410     protected BigInteger approximate(int p) {
1411         // The Taylor series is the sum of x^(2n+1) * (2n)!/(4^n n!^2 (2n+1))
1412         // Note that (2n)!/(4^n n!^2) is always less than one.
1413         // (The denominator is effectively 2n*2n*(2n-2)*(2n-2)*...*2*2
1414         // which is clearly > (2n)!)
1415         // Thus all terms are bounded by x^(2n+1).
1416         // Unfortunately, there's no easy way to prescale the argument
1417         // to less than 1/sqrt(2), and we can only approximate that.
1418         // Thus the worst case iteration count is fairly high.
1419         // But it doesn't make much difference.
1420         if (p >= 2) return big0;  // Never bigger than 4.
1421         int iterations_needed = -3 * p / 2 + 4;
1422                                 // conservative estimate > 0.
1423                                 // Follows from assumed bound on x and
1424                                 // the fact that only every other Taylor
1425                                 // Series term is present.
1426           //  Claim: each intermediate term is accurate
1427           //  to 2*2^calc_precision.
1428           //  Total rounding error in series computation is
1429           //  2*iterations_needed*2^calc_precision,
1430           //  exclusive of error in op.
1431         int calc_precision = p - bound_log2(2*iterations_needed)
1432                                - 4; // for error in op, truncation.
1433         int op_prec = p - 3;  // always <= -2
1434         BigInteger op_appr = op.get_appr(op_prec);
1435           // Error in argument results in error of < 1/4 ulp.
1436           // (Derivative is bounded by 2 in the specified range and we use
1437           // 3 extra digits.)
1438           // Ignoring the argument error, each term has an error of
1439           // < 3ulps relative to calc_precision, which is more precise than p.
1440           // Cumulative arithmetic rounding error is < 3/16 ulp (relative to p).
1441           // Series truncation error < 2/16 ulp.  (Each computed term
1442           // is at most 2/3 of last one, so some of remaining series <
1443           // 3/2 * current term.)
1444           // Final rounding error is <= 1/2 ulp.
1445           // Thus final error is < 1 ulp (relative to p).
1446         BigInteger max_last_term =
1447                 big1.shiftLeft(p - 4 - calc_precision);
1448         int exp = 1; // Current exponent, = 2n+1 in above expression
1449         BigInteger current_term = op_appr.shiftLeft(op_prec - calc_precision);
1450         BigInteger current_sum = current_term;
1451         BigInteger current_factor = current_term;
1452                                     // Current scaled Taylor series term
1453                                     // before division by the exponent.
1454                                     // Accurate to 3 ulp at calc_precision.
1455         while (current_term.abs().compareTo(max_last_term) >= 0) {
1456           if (Thread.interrupted() || please_stop) throw new AbortedException();
1457           exp += 2;
1458           // current_factor = current_factor * op * op * (exp-1) * (exp-2) /
1459           // (exp-1) * (exp-1), with the two exp-1 factors cancelling,
1460           // giving
1461           // current_factor = current_factor * op * op * (exp-2) / (exp-1)
1462           // Thus the error any in the previous term is multiplied by
1463           // op^2, adding an error of < (1/2)^(2/3) < 2/3 the original
1464           // error.
1465           current_factor = current_factor.multiply(BigInteger.valueOf(exp - 2));
1466           current_factor = scale(current_factor.multiply(op_appr), op_prec + 2);
1467                 // Carry 2 extra bits of precision forward; thus
1468                 // this effectively introduces 1/8 ulp error.
1469           current_factor = current_factor.multiply(op_appr);
1470           BigInteger divisor = BigInteger.valueOf(exp - 1);
1471           current_factor = current_factor.divide(divisor);
1472                 // Another 1/4 ulp error here.
1473           current_factor = scale(current_factor, op_prec - 2);
1474                 // Remove extra 2 bits.  1/2 ulp rounding error.
1475           // Current_factor has original 3 ulp rounding error, which we
1476           // reduced by 1, plus < 1 ulp new rounding error.
1477           current_term = current_factor.divide(BigInteger.valueOf(exp));
1478                 // Contributes 1 ulp error to sum plus at most 3 ulp
1479                 // from current_factor.
1480           current_sum = current_sum.add(current_term);
1481         }
1482         return scale(current_sum, calc_precision - p);
1483       }
1484   }
1485 
1486 
1487 class sqrt_CR extends CR {
1488     CR op;
sqrt_CR(CR x)1489     sqrt_CR(CR x) { op = x; }
1490     // Explicitly provide an initial approximation.
1491     // Useful for arithmetic geometric mean algorithms, where we've previously
1492     // computed a very similar square root.
sqrt_CR(CR x, int min_p, BigInteger max_a)1493     sqrt_CR(CR x, int min_p, BigInteger max_a) {
1494         op = x;
1495         min_prec = min_p;
1496         max_appr = max_a;
1497         appr_valid = true;
1498     }
1499     final int fp_prec = 50;     // Conservative estimate of number of
1500                                 // significant bits in double precision
1501                                 // computation.
1502     final int fp_op_prec = 60;
approximate(int p)1503     protected BigInteger approximate(int p) {
1504         int max_op_prec_needed = 2*p - 1;
1505         int msd = op.iter_msd(max_op_prec_needed);
1506         if (msd <= max_op_prec_needed) return big0;
1507         int result_msd = msd/2;                 // +- 1
1508         int result_digits = result_msd - p;     // +- 2
1509         if (result_digits > fp_prec) {
1510           // Compute less precise approximation and use a Newton iter.
1511             int appr_digits = result_digits/2 + 6;
1512                 // This should be conservative.  Is fewer enough?
1513             int appr_prec = result_msd - appr_digits;
1514             int prod_prec = 2*appr_prec;
1515             // First compute the argument to maximal precision, so we don't end up
1516             // reevaluating it incrementally.
1517             BigInteger op_appr = op.get_appr(prod_prec);
1518             BigInteger last_appr = get_appr(appr_prec);
1519             // Compute (last_appr * last_appr + op_appr) / last_appr / 2
1520             // while adjusting the scaling to make everything work
1521             BigInteger prod_prec_scaled_numerator =
1522                 last_appr.multiply(last_appr).add(op_appr);
1523             BigInteger scaled_numerator =
1524                 scale(prod_prec_scaled_numerator, appr_prec - p);
1525             BigInteger shifted_result = scaled_numerator.divide(last_appr);
1526             return shifted_result.add(big1).shiftRight(1);
1527         } else {
1528           // Use a double precision floating point approximation.
1529             // Make sure all precisions are even
1530             int op_prec = (msd - fp_op_prec) & ~1;
1531             int working_prec = op_prec - fp_op_prec;
1532             BigInteger scaled_bi_appr = op.get_appr(op_prec)
1533                                         .shiftLeft(fp_op_prec);
1534             double scaled_appr = scaled_bi_appr.doubleValue();
1535             if (scaled_appr < 0.0)
1536                 throw new ArithmeticException("sqrt(negative)");
1537             double scaled_fp_sqrt = Math.sqrt(scaled_appr);
1538             BigInteger scaled_sqrt = BigInteger.valueOf((long)scaled_fp_sqrt);
1539             int shift_count = working_prec/2 - p;
1540             return shift(scaled_sqrt, shift_count);
1541         }
1542     }
1543 }
1544 
1545 // The constant PI, computed using the Gauss-Legendre alternating
1546 // arithmetic-geometric mean algorithm:
1547 //      a[0] = 1
1548 //      b[0] = 1/sqrt(2)
1549 //      t[0] = 1/4
1550 //      p[0] = 1
1551 //
1552 //      a[n+1] = (a[n] + b[n])/2        (arithmetic mean, between 0.8 and 1)
1553 //      b[n+1] = sqrt(a[n] * b[n])      (geometric mean, between 0.7 and 1)
1554 //      t[n+1] = t[n] - (2^n)(a[n]-a[n+1])^2,  (always between 0.2 and 0.25)
1555 //
1556 //      pi is then approximated as (a[n+1]+b[n+1])^2 / 4*t[n+1].
1557 //
1558 class gl_pi_CR extends slow_CR {
1559     // In addition to the best approximation kept by the CR base class, we keep
1560     // the entire sequence b[n], to the extent we've needed it so far.  Each
1561     // reevaluation leads to slightly different sqrt arguments, but the
1562     // previous result can be used to avoid repeating low precision Newton
1563     // iterations for the sqrt approximation.
1564     ArrayList<Integer> b_prec = new ArrayList<Integer>();
1565     ArrayList<BigInteger> b_val = new ArrayList<BigInteger>();
gl_pi_CR()1566     gl_pi_CR() {
1567         b_prec.add(null);  // Zeroth entry unused.
1568         b_val.add(null);
1569     }
1570     private static BigInteger TOLERANCE = BigInteger.valueOf(4);
1571     // sqrt(1/2)
1572     private static CR SQRT_HALF = new sqrt_CR(ONE.shiftRight(1));
1573 
approximate(int p)1574     protected BigInteger approximate(int p) {
1575         // Rough approximations are easy.
1576         if (p >= 0) return scale(BigInteger.valueOf(3), -p);
1577         // We need roughly log2(p) iterations.  Each iteration should
1578         // contribute no more than 2 ulps to the error in the corresponding
1579         // term (a[n], b[n], or t[n]).  Thus 2log2(n) bits plus a few for the
1580         // final calulation and rounding suffice.
1581         final int extra_eval_prec =
1582                 (int)Math.ceil(Math.log(-p) / Math.log(2)) + 10;
1583         // All our terms are implicitly scaled by eval_prec.
1584         final int eval_prec = p - extra_eval_prec;
1585         BigInteger a = BigInteger.ONE.shiftLeft(-eval_prec);
1586         BigInteger b = SQRT_HALF.get_appr(eval_prec);
1587         BigInteger t = BigInteger.ONE.shiftLeft(-eval_prec - 2);
1588         int n = 0;
1589         while (a.subtract(b).subtract(TOLERANCE).signum() > 0) {
1590             // Current values correspond to n, next_ values to n + 1
1591             // b_prec.size() == b_val.size() >= n + 1
1592             final BigInteger next_a = a.add(b).shiftRight(1);
1593             final BigInteger a_diff = a.subtract(next_a);
1594             CR next_b_as_CR;
1595             final BigInteger b_prod = a.multiply(b).shiftRight(-eval_prec);
1596             // We the compute square root approximations using a nested
1597             // temporary CR computation, to avoid implementing BigInteger
1598             // square roots separately.
1599             final CR b_prod_as_CR = CR.valueOf(b_prod).shiftRight(-eval_prec);
1600             if (b_prec.size() == n + 1) {
1601                 // Need an n+1st slot.
1602                 b_prec.add(null);
1603                 b_val.add(null);
1604                 next_b_as_CR = b_prod_as_CR.sqrt();
1605             } else {
1606                 // Reuse previous approximation to reduce sqrt iterations,
1607                 // hopefully to one.
1608                 next_b_as_CR = new sqrt_CR(b_prod_as_CR, b_prec.get(n + 1),
1609                                            b_val.get(n + 1));
1610             }
1611             // b_prec.size() == b_val.size() >= n + 2
1612             final BigInteger next_b = next_b_as_CR.get_appr(eval_prec);
1613             b_prec.set(n + 1, Integer.valueOf(p));
1614             b_val.set(n + 1, scale(next_b, -extra_eval_prec));
1615             final BigInteger next_t =
1616                     t.subtract(a_diff.multiply(a_diff)
1617                      .shiftLeft(n + eval_prec));  // shift dist. usually neg.
1618             a = next_a;
1619             b = next_b;
1620             t = next_t;
1621             ++n;
1622         }
1623         final BigInteger sum = a.add(b);
1624         final BigInteger result = sum.multiply(sum).divide(t).shiftRight(2);
1625         return scale(result, -extra_eval_prec);
1626     }
1627 }
1628