• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*	$NetBSD: strtod.c,v 1.45.2.1 2005/04/19 13:35:54 tron Exp $	*/
2 
3 /****************************************************************
4  *
5  * The author of this software is David M. Gay.
6  *
7  * Copyright (c) 1991 by AT&T.
8  *
9  * Permission to use, copy, modify, and distribute this software for any
10  * purpose without fee is hereby granted, provided that this entire notice
11  * is included in all copies of any software which is or includes a copy
12  * or modification of this software and in all copies of the supporting
13  * documentation for such software.
14  *
15  * THIS SOFTWARE IS BEING PROVIDED "AS IS", WITHOUT ANY EXPRESS OR IMPLIED
16  * WARRANTY.  IN PARTICULAR, NEITHER THE AUTHOR NOR AT&T MAKES ANY
17  * REPRESENTATION OR WARRANTY OF ANY KIND CONCERNING THE MERCHANTABILITY
18  * OF THIS SOFTWARE OR ITS FITNESS FOR ANY PARTICULAR PURPOSE.
19  *
20  ***************************************************************/
21 
22 /* Please send bug reports to
23 	David M. Gay
24 	AT&T Bell Laboratories, Room 2C-463
25 	600 Mountain Avenue
26 	Murray Hill, NJ 07974-2070
27 	U.S.A.
28 	dmg@research.att.com or research!dmg
29  */
30 
31 /* strtod for IEEE-, VAX-, and IBM-arithmetic machines.
32  *
33  * This strtod returns a nearest machine number to the input decimal
34  * string (or sets errno to ERANGE).  With IEEE arithmetic, ties are
35  * broken by the IEEE round-even rule.  Otherwise ties are broken by
36  * biased rounding (add half and chop).
37  *
38  * Inspired loosely by William D. Clinger's paper "How to Read Floating
39  * Point Numbers Accurately" [Proc. ACM SIGPLAN '90, pp. 92-101].
40  *
41  * Modifications:
42  *
43  *	1. We only require IEEE, IBM, or VAX double-precision
44  *		arithmetic (not IEEE double-extended).
45  *	2. We get by with floating-point arithmetic in a case that
46  *		Clinger missed -- when we're computing d * 10^n
47  *		for a small integer d and the integer n is not too
48  *		much larger than 22 (the maximum integer k for which
49  *		we can represent 10^k exactly), we may be able to
50  *		compute (d*10^k) * 10^(e-k) with just one roundoff.
51  *	3. Rather than a bit-at-a-time adjustment of the binary
52  *		result in the hard case, we use floating-point
53  *		arithmetic to determine the adjustment to within
54  *		one bit; only in really hard cases do we need to
55  *		compute a second residual.
56  *	4. Because of 3., we don't need a large table of powers of 10
57  *		for ten-to-e (just some small tables, e.g. of 10^k
58  *		for 0 <= k <= 22).
59  */
60 
61 /*
62  * #define IEEE_LITTLE_ENDIAN for IEEE-arithmetic machines where the least
63  *	significant byte has the lowest address.
64  * #define IEEE_BIG_ENDIAN for IEEE-arithmetic machines where the most
65  *	significant byte has the lowest address.
66  * #define Long int on machines with 32-bit ints and 64-bit longs.
67  * #define Sudden_Underflow for IEEE-format machines without gradual
68  *	underflow (i.e., that flush to zero on underflow).
69  * #define IBM for IBM mainframe-style floating-point arithmetic.
70  * #define VAX for VAX-style floating-point arithmetic.
71  * #define Unsigned_Shifts if >> does treats its left operand as unsigned.
72  * #define No_leftright to omit left-right logic in fast floating-point
73  *	computation of dtoa.
74  * #define Check_FLT_ROUNDS if FLT_ROUNDS can assume the values 2 or 3.
75  * #define RND_PRODQUOT to use rnd_prod and rnd_quot (assembly routines
76  *	that use extended-precision instructions to compute rounded
77  *	products and quotients) with IBM.
78  * #define ROUND_BIASED for IEEE-format with biased rounding.
79  * #define Inaccurate_Divide for IEEE-format with correctly rounded
80  *	products but inaccurate quotients, e.g., for Intel i860.
81  * #define Just_16 to store 16 bits per 32-bit Long when doing high-precision
82  *	integer arithmetic.  Whether this speeds things up or slows things
83  *	down depends on the machine and the number being converted.
84  * #define KR_headers for old-style C function headers.
85  * #define Bad_float_h if your system lacks a float.h or if it does not
86  *	define some or all of DBL_DIG, DBL_MAX_10_EXP, DBL_MAX_EXP,
87  *	FLT_RADIX, FLT_ROUNDS, and DBL_MAX.
88  * #define MALLOC your_malloc, where your_malloc(n) acts like malloc(n)
89  *	if memory is available and otherwise does something you deem
90  *	appropriate.  If MALLOC is undefined, malloc will be invoked
91  *	directly -- and assumed always to succeed.
92  */
93 
94 #ifdef ANDROID_CHANGES
95 #include <pthread.h>
96 #define mutex_lock(x) pthread_mutex_lock(x)
97 #define mutex_unlock(x) pthread_mutex_unlock(x)
98 #endif
99 
100 #include <sys/cdefs.h>
101 #if defined(LIBC_SCCS) && !defined(lint)
102 __RCSID("$NetBSD: strtod.c,v 1.45.2.1 2005/04/19 13:35:54 tron Exp $");
103 #endif /* LIBC_SCCS and not lint */
104 
105 #define Unsigned_Shifts
106 #if defined(__m68k__) || defined(__sparc__) || defined(__i386__) || \
107     defined(__mips__) || defined(__ns32k__) || defined(__alpha__) || \
108     defined(__powerpc__) || defined(__sh__) || defined(__x86_64__) || \
109     defined(__hppa__) || \
110     (defined(__arm__) && defined(__VFP_FP__))
111 #include <endian.h>
112 #if BYTE_ORDER == BIG_ENDIAN
113 #define IEEE_BIG_ENDIAN
114 #else
115 #define IEEE_LITTLE_ENDIAN
116 #endif
117 #endif
118 
119 #if defined(__arm__) && !defined(__VFP_FP__)
120 /*
121  * Although the CPU is little endian the FP has different
122  * byte and word endianness. The byte order is still little endian
123  * but the word order is big endian.
124  */
125 #define IEEE_BIG_ENDIAN
126 #endif
127 
128 #ifdef __vax__
129 #define VAX
130 #endif
131 
132 #if defined(__hppa__) || defined(__mips__) || defined(__sh__)
133 #define	NAN_WORD0	0x7ff40000
134 #else
135 #define	NAN_WORD0	0x7ff80000
136 #endif
137 #define	NAN_WORD1	0
138 
139 #define Long	int32_t
140 #define ULong	u_int32_t
141 
142 #ifdef DEBUG
143 #include "stdio.h"
144 #define Bug(x) {fprintf(stderr, "%s\n", x); exit(1);}
145 #endif
146 
147 #ifdef __cplusplus
148 #include "malloc.h"
149 #include "memory.h"
150 #else
151 #ifndef KR_headers
152 #include "stdlib.h"
153 #include "string.h"
154 #ifndef ANDROID_CHANGES
155 #include "locale.h"
156 #endif /* ANDROID_CHANGES */
157 #else
158 #include "malloc.h"
159 #include "memory.h"
160 #endif
161 #endif
162 #ifndef ANDROID_CHANGES
163 #include "extern.h"
164 #include "reentrant.h"
165 #endif /* ANDROID_CHANGES */
166 
167 #ifdef MALLOC
168 #ifdef KR_headers
169 extern char *MALLOC();
170 #else
171 extern void *MALLOC(size_t);
172 #endif
173 #else
174 #define MALLOC malloc
175 #endif
176 
177 #include "ctype.h"
178 #include "errno.h"
179 #include "float.h"
180 
181 #ifndef __MATH_H__
182 #include "math.h"
183 #endif
184 
185 #ifdef __cplusplus
186 extern "C" {
187 #endif
188 
189 #ifndef CONST
190 #ifdef KR_headers
191 #define CONST /* blank */
192 #else
193 #define CONST const
194 #endif
195 #endif
196 
197 #ifdef Unsigned_Shifts
198 #define Sign_Extend(a,b) if (b < 0) a |= 0xffff0000;
199 #else
200 #define Sign_Extend(a,b) /*no-op*/
201 #endif
202 
203 #if defined(IEEE_LITTLE_ENDIAN) + defined(IEEE_BIG_ENDIAN) + defined(VAX) + \
204     defined(IBM) != 1
205 Exactly one of IEEE_LITTLE_ENDIAN IEEE_BIG_ENDIAN, VAX, or
206 IBM should be defined.
207 #endif
208 
209 typedef union {
210 	double d;
211 	ULong ul[2];
212 } _double;
213 #define value(x) ((x).d)
214 #ifdef IEEE_LITTLE_ENDIAN
215 #define word0(x) ((x).ul[1])
216 #define word1(x) ((x).ul[0])
217 #else
218 #define word0(x) ((x).ul[0])
219 #define word1(x) ((x).ul[1])
220 #endif
221 
222 /* The following definition of Storeinc is appropriate for MIPS processors.
223  * An alternative that might be better on some machines is
224  * #define Storeinc(a,b,c) (*a++ = b << 16 | c & 0xffff)
225  */
226 #if defined(IEEE_LITTLE_ENDIAN) + defined(VAX) + defined(__arm__)
227 #define Storeinc(a,b,c) \
228     (((u_short *)(void *)a)[1] = \
229 	(u_short)b, ((u_short *)(void *)a)[0] = (u_short)c, a++)
230 #else
231 #define Storeinc(a,b,c) \
232     (((u_short *)(void *)a)[0] = \
233 	(u_short)b, ((u_short *)(void *)a)[1] = (u_short)c, a++)
234 #endif
235 
236 /* #define P DBL_MANT_DIG */
237 /* Ten_pmax = floor(P*log(2)/log(5)) */
238 /* Bletch = (highest power of 2 < DBL_MAX_10_EXP) / 16 */
239 /* Quick_max = floor((P-1)*log(FLT_RADIX)/log(10) - 1) */
240 /* Int_max = floor(P*log(FLT_RADIX)/log(10) - 1) */
241 
242 #if defined(IEEE_LITTLE_ENDIAN) + defined(IEEE_BIG_ENDIAN)
243 #define Exp_shift  20
244 #define Exp_shift1 20
245 #define Exp_msk1    0x100000
246 #define Exp_msk11   0x100000
247 #define Exp_mask  0x7ff00000
248 #define P 53
249 #define Bias 1023
250 #define IEEE_Arith
251 #define Emin (-1022)
252 #define Exp_1  0x3ff00000
253 #define Exp_11 0x3ff00000
254 #define Ebits 11
255 #define Frac_mask  0xfffff
256 #define Frac_mask1 0xfffff
257 #define Ten_pmax 22
258 #define Bletch 0x10
259 #define Bndry_mask  0xfffff
260 #define Bndry_mask1 0xfffff
261 #define LSB 1
262 #define Sign_bit 0x80000000
263 #define Log2P 1
264 #define Tiny0 0
265 #define Tiny1 1
266 #define Quick_max 14
267 #define Int_max 14
268 #define Infinite(x) (word0(x) == 0x7ff00000) /* sufficient test for here */
269 #else
270 #undef  Sudden_Underflow
271 #define Sudden_Underflow
272 #ifdef IBM
273 #define Exp_shift  24
274 #define Exp_shift1 24
275 #define Exp_msk1   0x1000000
276 #define Exp_msk11  0x1000000
277 #define Exp_mask  0x7f000000
278 #define P 14
279 #define Bias 65
280 #define Exp_1  0x41000000
281 #define Exp_11 0x41000000
282 #define Ebits 8	/* exponent has 7 bits, but 8 is the right value in b2d */
283 #define Frac_mask  0xffffff
284 #define Frac_mask1 0xffffff
285 #define Bletch 4
286 #define Ten_pmax 22
287 #define Bndry_mask  0xefffff
288 #define Bndry_mask1 0xffffff
289 #define LSB 1
290 #define Sign_bit 0x80000000
291 #define Log2P 4
292 #define Tiny0 0x100000
293 #define Tiny1 0
294 #define Quick_max 14
295 #define Int_max 15
296 #else /* VAX */
297 #define Exp_shift  23
298 #define Exp_shift1 7
299 #define Exp_msk1    0x80
300 #define Exp_msk11   0x800000
301 #define Exp_mask  0x7f80
302 #define P 56
303 #define Bias 129
304 #define Exp_1  0x40800000
305 #define Exp_11 0x4080
306 #define Ebits 8
307 #define Frac_mask  0x7fffff
308 #define Frac_mask1 0xffff007f
309 #define Ten_pmax 24
310 #define Bletch 2
311 #define Bndry_mask  0xffff007f
312 #define Bndry_mask1 0xffff007f
313 #define LSB 0x10000
314 #define Sign_bit 0x8000
315 #define Log2P 1
316 #define Tiny0 0x80
317 #define Tiny1 0
318 #define Quick_max 15
319 #define Int_max 15
320 #endif
321 #endif
322 
323 #ifndef IEEE_Arith
324 #define ROUND_BIASED
325 #endif
326 
327 #ifdef RND_PRODQUOT
328 #define rounded_product(a,b) a = rnd_prod(a, b)
329 #define rounded_quotient(a,b) a = rnd_quot(a, b)
330 #ifdef KR_headers
331 extern double rnd_prod(), rnd_quot();
332 #else
333 extern double rnd_prod(double, double), rnd_quot(double, double);
334 #endif
335 #else
336 #define rounded_product(a,b) a *= b
337 #define rounded_quotient(a,b) a /= b
338 #endif
339 
340 #define Big0 (Frac_mask1 | Exp_msk1*(DBL_MAX_EXP+Bias-1))
341 #define Big1 0xffffffff
342 
343 #ifndef Just_16
344 /* When Pack_32 is not defined, we store 16 bits per 32-bit Long.
345  * This makes some inner loops simpler and sometimes saves work
346  * during multiplications, but it often seems to make things slightly
347  * slower.  Hence the default is now to store 32 bits per Long.
348  */
349 #ifndef Pack_32
350 #define Pack_32
351 #endif
352 #endif
353 
354 #define Kmax 15
355 
356 #ifdef __cplusplus
357 extern "C" double strtod(const char *s00, char **se);
358 extern "C" char *__dtoa(double d, int mode, int ndigits,
359 			int *decpt, int *sign, char **rve);
360 #endif
361 
362  struct
363 Bigint {
364 	struct Bigint *next;
365 	int k, maxwds, sign, wds;
366 	ULong x[1];
367 };
368 
369  typedef struct Bigint Bigint;
370 
371  static Bigint *freelist[Kmax+1];
372 
373 #ifdef ANDROID_CHANGES
374  static pthread_mutex_t freelist_mutex = PTHREAD_MUTEX_INITIALIZER;
375 #else
376 #ifdef _REENTRANT
377  static mutex_t freelist_mutex = MUTEX_INITIALIZER;
378 #endif
379 #endif
380 
381 /* Special value used to indicate an invalid Bigint value,
382  * e.g. when a memory allocation fails. The idea is that we
383  * want to avoid introducing NULL checks everytime a bigint
384  * computation is performed. Also the NULL value can also be
385  * already used to indicate "value not initialized yet" and
386  * returning NULL might alter the execution code path in
387  * case of OOM.
388  */
389 #define  BIGINT_INVALID   ((Bigint *)&bigint_invalid_value)
390 
391 static const Bigint bigint_invalid_value;
392 
393 
394 /* Return BIGINT_INVALID on allocation failure.
395  *
396  * Most of the code here depends on the fact that this function
397  * never returns NULL.
398  */
399  static Bigint *
Balloc(k)400 Balloc
401 #ifdef KR_headers
402 	(k) int k;
403 #else
404 	(int k)
405 #endif
406 {
407 	int x;
408 	Bigint *rv;
409 
410 	mutex_lock(&freelist_mutex);
411 
412 	if ((rv = freelist[k]) != NULL) {
413 		freelist[k] = rv->next;
414 	}
415 	else {
416 		x = 1 << k;
417 		rv = (Bigint *)MALLOC(sizeof(Bigint) + (x-1)*sizeof(Long));
418 		if (rv == NULL) {
419 		        rv = BIGINT_INVALID;
420 			goto EXIT;
421 		}
422 		rv->k = k;
423 		rv->maxwds = x;
424 	}
425 	rv->sign = rv->wds = 0;
426 EXIT:
427 	mutex_unlock(&freelist_mutex);
428 
429 	return rv;
430 }
431 
432  static void
Bfree(v)433 Bfree
434 #ifdef KR_headers
435 	(v) Bigint *v;
436 #else
437 	(Bigint *v)
438 #endif
439 {
440 	if (v && v != BIGINT_INVALID) {
441 		mutex_lock(&freelist_mutex);
442 
443 		v->next = freelist[v->k];
444 		freelist[v->k] = v;
445 
446 		mutex_unlock(&freelist_mutex);
447 	}
448 }
449 
450 #define Bcopy_valid(x,y) memcpy(&(x)->sign, &(y)->sign, \
451     (y)->wds*sizeof(Long) + 2*sizeof(int))
452 
453 #define Bcopy(x,y)  Bcopy_ptr(&(x),(y))
454 
455  static void
Bcopy_ptr(Bigint ** px,Bigint * y)456 Bcopy_ptr(Bigint **px, Bigint *y)
457 {
458 	if (*px == BIGINT_INVALID)
459 		return; /* no space to store copy */
460 	if (y == BIGINT_INVALID) {
461 		Bfree(*px); /* invalid input */
462 		*px = BIGINT_INVALID;
463 	} else {
464 		Bcopy_valid(*px,y);
465 	}
466 }
467 
468  static Bigint *
multadd(b,m,a)469 multadd
470 #ifdef KR_headers
471 	(b, m, a) Bigint *b; int m, a;
472 #else
473 	(Bigint *b, int m, int a)	/* multiply by m and add a */
474 #endif
475 {
476 	int i, wds;
477 	ULong *x, y;
478 #ifdef Pack_32
479 	ULong xi, z;
480 #endif
481 	Bigint *b1;
482 
483 	if (b == BIGINT_INVALID)
484 		return b;
485 
486 	wds = b->wds;
487 	x = b->x;
488 	i = 0;
489 	do {
490 #ifdef Pack_32
491 		xi = *x;
492 		y = (xi & 0xffff) * m + a;
493 		z = (xi >> 16) * m + (y >> 16);
494 		a = (int)(z >> 16);
495 		*x++ = (z << 16) + (y & 0xffff);
496 #else
497 		y = *x * m + a;
498 		a = (int)(y >> 16);
499 		*x++ = y & 0xffff;
500 #endif
501 	}
502 	while(++i < wds);
503 	if (a) {
504 		if (wds >= b->maxwds) {
505 			b1 = Balloc(b->k+1);
506 			if (b1 == BIGINT_INVALID) {
507 				Bfree(b);
508 				return b1;
509 			}
510 			Bcopy_valid(b1, b);
511 			Bfree(b);
512 			b = b1;
513 			}
514 		b->x[wds++] = a;
515 		b->wds = wds;
516 	}
517 	return b;
518 }
519 
520  static Bigint *
s2b(s,nd0,nd,y9)521 s2b
522 #ifdef KR_headers
523 	(s, nd0, nd, y9) CONST char *s; int nd0, nd; ULong y9;
524 #else
525 	(CONST char *s, int nd0, int nd, ULong y9)
526 #endif
527 {
528 	Bigint *b;
529 	int i, k;
530 	Long x, y;
531 
532 	x = (nd + 8) / 9;
533 	for(k = 0, y = 1; x > y; y <<= 1, k++) ;
534 #ifdef Pack_32
535 	b = Balloc(k);
536 	if (b == BIGINT_INVALID)
537 		return b;
538 	b->x[0] = y9;
539 	b->wds = 1;
540 #else
541 	b = Balloc(k+1);
542 	if (b == BIGINT_INVALID)
543 		return b;
544 
545 	b->x[0] = y9 & 0xffff;
546 	b->wds = (b->x[1] = y9 >> 16) ? 2 : 1;
547 #endif
548 
549 	i = 9;
550 	if (9 < nd0) {
551 		s += 9;
552 		do b = multadd(b, 10, *s++ - '0');
553 			while(++i < nd0);
554 		s++;
555 	}
556 	else
557 		s += 10;
558 	for(; i < nd; i++)
559 		b = multadd(b, 10, *s++ - '0');
560 	return b;
561 }
562 
563  static int
hi0bits(x)564 hi0bits
565 #ifdef KR_headers
566 	(x) ULong x;
567 #else
568 	(ULong x)
569 #endif
570 {
571 	int k = 0;
572 
573 	if (!(x & 0xffff0000)) {
574 		k = 16;
575 		x <<= 16;
576 	}
577 	if (!(x & 0xff000000)) {
578 		k += 8;
579 		x <<= 8;
580 	}
581 	if (!(x & 0xf0000000)) {
582 		k += 4;
583 		x <<= 4;
584 	}
585 	if (!(x & 0xc0000000)) {
586 		k += 2;
587 		x <<= 2;
588 	}
589 	if (!(x & 0x80000000)) {
590 		k++;
591 		if (!(x & 0x40000000))
592 			return 32;
593 	}
594 	return k;
595 }
596 
597  static int
lo0bits(y)598 lo0bits
599 #ifdef KR_headers
600 	(y) ULong *y;
601 #else
602 	(ULong *y)
603 #endif
604 {
605 	int k;
606 	ULong x = *y;
607 
608 	if (x & 7) {
609 		if (x & 1)
610 			return 0;
611 		if (x & 2) {
612 			*y = x >> 1;
613 			return 1;
614 			}
615 		*y = x >> 2;
616 		return 2;
617 	}
618 	k = 0;
619 	if (!(x & 0xffff)) {
620 		k = 16;
621 		x >>= 16;
622 	}
623 	if (!(x & 0xff)) {
624 		k += 8;
625 		x >>= 8;
626 	}
627 	if (!(x & 0xf)) {
628 		k += 4;
629 		x >>= 4;
630 	}
631 	if (!(x & 0x3)) {
632 		k += 2;
633 		x >>= 2;
634 	}
635 	if (!(x & 1)) {
636 		k++;
637 		x >>= 1;
638 		if (!x & 1)
639 			return 32;
640 	}
641 	*y = x;
642 	return k;
643 }
644 
645  static Bigint *
i2b(i)646 i2b
647 #ifdef KR_headers
648 	(i) int i;
649 #else
650 	(int i)
651 #endif
652 {
653 	Bigint *b;
654 
655 	b = Balloc(1);
656 	if (b != BIGINT_INVALID) {
657 		b->x[0] = i;
658 		b->wds = 1;
659 		}
660 	return b;
661 }
662 
663  static Bigint *
mult(a,b)664 mult
665 #ifdef KR_headers
666 	(a, b) Bigint *a, *b;
667 #else
668 	(Bigint *a, Bigint *b)
669 #endif
670 {
671 	Bigint *c;
672 	int k, wa, wb, wc;
673 	ULong carry, y, z;
674 	ULong *x, *xa, *xae, *xb, *xbe, *xc, *xc0;
675 #ifdef Pack_32
676 	ULong z2;
677 #endif
678 
679 	if (a == BIGINT_INVALID || b == BIGINT_INVALID)
680 		return BIGINT_INVALID;
681 
682 	if (a->wds < b->wds) {
683 		c = a;
684 		a = b;
685 		b = c;
686 	}
687 	k = a->k;
688 	wa = a->wds;
689 	wb = b->wds;
690 	wc = wa + wb;
691 	if (wc > a->maxwds)
692 		k++;
693 	c = Balloc(k);
694 	if (c == BIGINT_INVALID)
695 		return c;
696 	for(x = c->x, xa = x + wc; x < xa; x++)
697 		*x = 0;
698 	xa = a->x;
699 	xae = xa + wa;
700 	xb = b->x;
701 	xbe = xb + wb;
702 	xc0 = c->x;
703 #ifdef Pack_32
704 	for(; xb < xbe; xb++, xc0++) {
705 		if ((y = *xb & 0xffff) != 0) {
706 			x = xa;
707 			xc = xc0;
708 			carry = 0;
709 			do {
710 				z = (*x & 0xffff) * y + (*xc & 0xffff) + carry;
711 				carry = z >> 16;
712 				z2 = (*x++ >> 16) * y + (*xc >> 16) + carry;
713 				carry = z2 >> 16;
714 				Storeinc(xc, z2, z);
715 			}
716 			while(x < xae);
717 			*xc = carry;
718 		}
719 		if ((y = *xb >> 16) != 0) {
720 			x = xa;
721 			xc = xc0;
722 			carry = 0;
723 			z2 = *xc;
724 			do {
725 				z = (*x & 0xffff) * y + (*xc >> 16) + carry;
726 				carry = z >> 16;
727 				Storeinc(xc, z, z2);
728 				z2 = (*x++ >> 16) * y + (*xc & 0xffff) + carry;
729 				carry = z2 >> 16;
730 			}
731 			while(x < xae);
732 			*xc = z2;
733 		}
734 	}
735 #else
736 	for(; xb < xbe; xc0++) {
737 		if (y = *xb++) {
738 			x = xa;
739 			xc = xc0;
740 			carry = 0;
741 			do {
742 				z = *x++ * y + *xc + carry;
743 				carry = z >> 16;
744 				*xc++ = z & 0xffff;
745 			}
746 			while(x < xae);
747 			*xc = carry;
748 		}
749 	}
750 #endif
751 	for(xc0 = c->x, xc = xc0 + wc; wc > 0 && !*--xc; --wc) ;
752 	c->wds = wc;
753 	return c;
754 }
755 
756  static Bigint *p5s;
757 
758  static Bigint *
pow5mult(b,k)759 pow5mult
760 #ifdef KR_headers
761 	(b, k) Bigint *b; int k;
762 #else
763 	(Bigint *b, int k)
764 #endif
765 {
766 	Bigint *b1, *p5, *p51;
767 	int i;
768 	static const int p05[3] = { 5, 25, 125 };
769 
770 	if (b == BIGINT_INVALID)
771 		return b;
772 
773 	if ((i = k & 3) != 0)
774 		b = multadd(b, p05[i-1], 0);
775 
776 	if (!(k = (unsigned int) k >> 2))
777 		return b;
778 	if (!(p5 = p5s)) {
779 		/* first time */
780 		p5 = i2b(625);
781 		if (p5 == BIGINT_INVALID) {
782 			Bfree(b);
783 			return p5;
784 		}
785 		p5s = p5;
786 		p5->next = 0;
787 	}
788 	for(;;) {
789 		if (k & 1) {
790 			b1 = mult(b, p5);
791 			Bfree(b);
792 			b = b1;
793 		}
794 		if (!(k = (unsigned int) k >> 1))
795 			break;
796 		if (!(p51 = p5->next)) {
797 			p51 = mult(p5,p5);
798 			if (p51 == BIGINT_INVALID) {
799 				Bfree(b);
800 				return p51;
801 			}
802 			p5->next = p51;
803 			p51->next = 0;
804 		}
805 		p5 = p51;
806 	}
807 	return b;
808 }
809 
810  static Bigint *
lshift(b,k)811 lshift
812 #ifdef KR_headers
813 	(b, k) Bigint *b; int k;
814 #else
815 	(Bigint *b, int k)
816 #endif
817 {
818 	int i, k1, n, n1;
819 	Bigint *b1;
820 	ULong *x, *x1, *xe, z;
821 
822 	if (b == BIGINT_INVALID)
823 		return b;
824 
825 #ifdef Pack_32
826 	n = (unsigned int)k >> 5;
827 #else
828 	n = (unsigned int)k >> 4;
829 #endif
830 	k1 = b->k;
831 	n1 = n + b->wds + 1;
832 	for(i = b->maxwds; n1 > i; i <<= 1)
833 		k1++;
834 	b1 = Balloc(k1);
835 	if (b1 == BIGINT_INVALID) {
836 		Bfree(b);
837 		return b1;
838 	}
839 	x1 = b1->x;
840 	for(i = 0; i < n; i++)
841 		*x1++ = 0;
842 	x = b->x;
843 	xe = x + b->wds;
844 #ifdef Pack_32
845 	if (k &= 0x1f) {
846 		k1 = 32 - k;
847 		z = 0;
848 		do {
849 			*x1++ = *x << k | z;
850 			z = *x++ >> k1;
851 		}
852 		while(x < xe);
853 		if ((*x1 = z) != 0)
854 			++n1;
855 	}
856 #else
857 	if (k &= 0xf) {
858 		k1 = 16 - k;
859 		z = 0;
860 		do {
861 			*x1++ = *x << k  & 0xffff | z;
862 			z = *x++ >> k1;
863 		}
864 		while(x < xe);
865 		if (*x1 = z)
866 			++n1;
867 	}
868 #endif
869 	else do
870 		*x1++ = *x++;
871 		while(x < xe);
872 	b1->wds = n1 - 1;
873 	Bfree(b);
874 	return b1;
875 }
876 
877  static int
cmp(a,b)878 cmp
879 #ifdef KR_headers
880 	(a, b) Bigint *a, *b;
881 #else
882 	(Bigint *a, Bigint *b)
883 #endif
884 {
885 	ULong *xa, *xa0, *xb, *xb0;
886 	int i, j;
887 
888 	if (a == BIGINT_INVALID || b == BIGINT_INVALID)
889 #ifdef DEBUG
890 		Bug("cmp called with a or b invalid");
891 #else
892 		return 0; /* equal - the best we can do right now */
893 #endif
894 
895 	i = a->wds;
896 	j = b->wds;
897 #ifdef DEBUG
898 	if (i > 1 && !a->x[i-1])
899 		Bug("cmp called with a->x[a->wds-1] == 0");
900 	if (j > 1 && !b->x[j-1])
901 		Bug("cmp called with b->x[b->wds-1] == 0");
902 #endif
903 	if (i -= j)
904 		return i;
905 	xa0 = a->x;
906 	xa = xa0 + j;
907 	xb0 = b->x;
908 	xb = xb0 + j;
909 	for(;;) {
910 		if (*--xa != *--xb)
911 			return *xa < *xb ? -1 : 1;
912 		if (xa <= xa0)
913 			break;
914 	}
915 	return 0;
916 }
917 
918  static Bigint *
diff(a,b)919 diff
920 #ifdef KR_headers
921 	(a, b) Bigint *a, *b;
922 #else
923 	(Bigint *a, Bigint *b)
924 #endif
925 {
926 	Bigint *c;
927 	int i, wa, wb;
928 	Long borrow, y;	/* We need signed shifts here. */
929 	ULong *xa, *xae, *xb, *xbe, *xc;
930 #ifdef Pack_32
931 	Long z;
932 #endif
933 
934 	if (a == BIGINT_INVALID || b == BIGINT_INVALID)
935 		return BIGINT_INVALID;
936 
937 	i = cmp(a,b);
938 	if (!i) {
939 		c = Balloc(0);
940 		if (c != BIGINT_INVALID) {
941 			c->wds = 1;
942 			c->x[0] = 0;
943 			}
944 		return c;
945 	}
946 	if (i < 0) {
947 		c = a;
948 		a = b;
949 		b = c;
950 		i = 1;
951 	}
952 	else
953 		i = 0;
954 	c = Balloc(a->k);
955 	if (c == BIGINT_INVALID)
956 		return c;
957 	c->sign = i;
958 	wa = a->wds;
959 	xa = a->x;
960 	xae = xa + wa;
961 	wb = b->wds;
962 	xb = b->x;
963 	xbe = xb + wb;
964 	xc = c->x;
965 	borrow = 0;
966 #ifdef Pack_32
967 	do {
968 		y = (*xa & 0xffff) - (*xb & 0xffff) + borrow;
969 		borrow = (ULong)y >> 16;
970 		Sign_Extend(borrow, y);
971 		z = (*xa++ >> 16) - (*xb++ >> 16) + borrow;
972 		borrow = (ULong)z >> 16;
973 		Sign_Extend(borrow, z);
974 		Storeinc(xc, z, y);
975 	}
976 	while(xb < xbe);
977 	while(xa < xae) {
978 		y = (*xa & 0xffff) + borrow;
979 		borrow = (ULong)y >> 16;
980 		Sign_Extend(borrow, y);
981 		z = (*xa++ >> 16) + borrow;
982 		borrow = (ULong)z >> 16;
983 		Sign_Extend(borrow, z);
984 		Storeinc(xc, z, y);
985 	}
986 #else
987 	do {
988 		y = *xa++ - *xb++ + borrow;
989 		borrow = y >> 16;
990 		Sign_Extend(borrow, y);
991 		*xc++ = y & 0xffff;
992 	}
993 	while(xb < xbe);
994 	while(xa < xae) {
995 		y = *xa++ + borrow;
996 		borrow = y >> 16;
997 		Sign_Extend(borrow, y);
998 		*xc++ = y & 0xffff;
999 	}
1000 #endif
1001 	while(!*--xc)
1002 		wa--;
1003 	c->wds = wa;
1004 	return c;
1005 }
1006 
1007  static double
ulp(_x)1008 ulp
1009 #ifdef KR_headers
1010 	(_x) double _x;
1011 #else
1012 	(double _x)
1013 #endif
1014 {
1015 	_double x;
1016 	Long L;
1017 	_double a;
1018 
1019 	value(x) = _x;
1020 	L = (word0(x) & Exp_mask) - (P-1)*Exp_msk1;
1021 #ifndef Sudden_Underflow
1022 	if (L > 0) {
1023 #endif
1024 #ifdef IBM
1025 		L |= Exp_msk1 >> 4;
1026 #endif
1027 		word0(a) = L;
1028 		word1(a) = 0;
1029 #ifndef Sudden_Underflow
1030 	}
1031 	else {
1032 		L = (ULong)-L >> Exp_shift;
1033 		if (L < Exp_shift) {
1034 			word0(a) = 0x80000 >> L;
1035 			word1(a) = 0;
1036 		}
1037 		else {
1038 			word0(a) = 0;
1039 			L -= Exp_shift;
1040 			word1(a) = L >= 31 ? 1 : 1 << (31 - L);
1041 		}
1042 	}
1043 #endif
1044 	return value(a);
1045 }
1046 
1047  static double
b2d(a,e)1048 b2d
1049 #ifdef KR_headers
1050 	(a, e) Bigint *a; int *e;
1051 #else
1052 	(Bigint *a, int *e)
1053 #endif
1054 {
1055 	ULong *xa, *xa0, w, y, z;
1056 	int k;
1057 	_double d;
1058 #ifdef VAX
1059 	ULong d0, d1;
1060 #else
1061 #define d0 word0(d)
1062 #define d1 word1(d)
1063 #endif
1064 
1065 	if (a == BIGINT_INVALID)
1066 		return NAN;
1067 
1068 	xa0 = a->x;
1069 	xa = xa0 + a->wds;
1070 	y = *--xa;
1071 #ifdef DEBUG
1072 	if (!y) Bug("zero y in b2d");
1073 #endif
1074 	k = hi0bits(y);
1075 	*e = 32 - k;
1076 #ifdef Pack_32
1077 	if (k < Ebits) {
1078 		d0 = Exp_1 | y >> (Ebits - k);
1079 		w = xa > xa0 ? *--xa : 0;
1080 		d1 = y << ((32-Ebits) + k) | w >> (Ebits - k);
1081 		goto ret_d;
1082 	}
1083 	z = xa > xa0 ? *--xa : 0;
1084 	if (k -= Ebits) {
1085 		d0 = Exp_1 | y << k | z >> (32 - k);
1086 		y = xa > xa0 ? *--xa : 0;
1087 		d1 = z << k | y >> (32 - k);
1088 	}
1089 	else {
1090 		d0 = Exp_1 | y;
1091 		d1 = z;
1092 	}
1093 #else
1094 	if (k < Ebits + 16) {
1095 		z = xa > xa0 ? *--xa : 0;
1096 		d0 = Exp_1 | y << k - Ebits | z >> Ebits + 16 - k;
1097 		w = xa > xa0 ? *--xa : 0;
1098 		y = xa > xa0 ? *--xa : 0;
1099 		d1 = z << k + 16 - Ebits | w << k - Ebits | y >> 16 + Ebits - k;
1100 		goto ret_d;
1101 	}
1102 	z = xa > xa0 ? *--xa : 0;
1103 	w = xa > xa0 ? *--xa : 0;
1104 	k -= Ebits + 16;
1105 	d0 = Exp_1 | y << k + 16 | z << k | w >> 16 - k;
1106 	y = xa > xa0 ? *--xa : 0;
1107 	d1 = w << k + 16 | y << k;
1108 #endif
1109  ret_d:
1110 #ifdef VAX
1111 	word0(d) = d0 >> 16 | d0 << 16;
1112 	word1(d) = d1 >> 16 | d1 << 16;
1113 #else
1114 #undef d0
1115 #undef d1
1116 #endif
1117 	return value(d);
1118 }
1119 
1120  static Bigint *
d2b(_d,e,bits)1121 d2b
1122 #ifdef KR_headers
1123 	(_d, e, bits) double d; int *e, *bits;
1124 #else
1125 	(double _d, int *e, int *bits)
1126 #endif
1127 {
1128 	Bigint *b;
1129 	int de, i, k;
1130 	ULong *x, y, z;
1131 	_double d;
1132 #ifdef VAX
1133 	ULong d0, d1;
1134 #endif
1135 
1136 	value(d) = _d;
1137 #ifdef VAX
1138 	d0 = word0(d) >> 16 | word0(d) << 16;
1139 	d1 = word1(d) >> 16 | word1(d) << 16;
1140 #else
1141 #define d0 word0(d)
1142 #define d1 word1(d)
1143 #endif
1144 
1145 #ifdef Pack_32
1146 	b = Balloc(1);
1147 #else
1148 	b = Balloc(2);
1149 #endif
1150 	if (b == BIGINT_INVALID)
1151 		return b;
1152 	x = b->x;
1153 
1154 	z = d0 & Frac_mask;
1155 	d0 &= 0x7fffffff;	/* clear sign bit, which we ignore */
1156 #ifdef Sudden_Underflow
1157 	de = (int)(d0 >> Exp_shift);
1158 #ifndef IBM
1159 	z |= Exp_msk11;
1160 #endif
1161 #else
1162 	if ((de = (int)(d0 >> Exp_shift)) != 0)
1163 		z |= Exp_msk1;
1164 #endif
1165 #ifdef Pack_32
1166 	if ((y = d1) != 0) {
1167 		if ((k = lo0bits(&y)) != 0) {
1168 			x[0] = y | z << (32 - k);
1169 			z >>= k;
1170 		}
1171 		else
1172 			x[0] = y;
1173 		i = b->wds = (x[1] = z) ? 2 : 1;
1174 	}
1175 	else {
1176 #ifdef DEBUG
1177 		if (!z)
1178 			Bug("Zero passed to d2b");
1179 #endif
1180 		k = lo0bits(&z);
1181 		x[0] = z;
1182 		i = b->wds = 1;
1183 		k += 32;
1184 	}
1185 #else
1186 	if (y = d1) {
1187 		if (k = lo0bits(&y))
1188 			if (k >= 16) {
1189 				x[0] = y | z << 32 - k & 0xffff;
1190 				x[1] = z >> k - 16 & 0xffff;
1191 				x[2] = z >> k;
1192 				i = 2;
1193 			}
1194 			else {
1195 				x[0] = y & 0xffff;
1196 				x[1] = y >> 16 | z << 16 - k & 0xffff;
1197 				x[2] = z >> k & 0xffff;
1198 				x[3] = z >> k+16;
1199 				i = 3;
1200 			}
1201 		else {
1202 			x[0] = y & 0xffff;
1203 			x[1] = y >> 16;
1204 			x[2] = z & 0xffff;
1205 			x[3] = z >> 16;
1206 			i = 3;
1207 		}
1208 	}
1209 	else {
1210 #ifdef DEBUG
1211 		if (!z)
1212 			Bug("Zero passed to d2b");
1213 #endif
1214 		k = lo0bits(&z);
1215 		if (k >= 16) {
1216 			x[0] = z;
1217 			i = 0;
1218 		}
1219 		else {
1220 			x[0] = z & 0xffff;
1221 			x[1] = z >> 16;
1222 			i = 1;
1223 		}
1224 		k += 32;
1225 	}
1226 	while(!x[i])
1227 		--i;
1228 	b->wds = i + 1;
1229 #endif
1230 #ifndef Sudden_Underflow
1231 	if (de) {
1232 #endif
1233 #ifdef IBM
1234 		*e = (de - Bias - (P-1) << 2) + k;
1235 		*bits = 4*P + 8 - k - hi0bits(word0(d) & Frac_mask);
1236 #else
1237 		*e = de - Bias - (P-1) + k;
1238 		*bits = P - k;
1239 #endif
1240 #ifndef Sudden_Underflow
1241 	}
1242 	else {
1243 		*e = de - Bias - (P-1) + 1 + k;
1244 #ifdef Pack_32
1245 		*bits = 32*i - hi0bits(x[i-1]);
1246 #else
1247 		*bits = (i+2)*16 - hi0bits(x[i]);
1248 #endif
1249 		}
1250 #endif
1251 	return b;
1252 }
1253 #undef d0
1254 #undef d1
1255 
1256  static double
ratio(a,b)1257 ratio
1258 #ifdef KR_headers
1259 	(a, b) Bigint *a, *b;
1260 #else
1261 	(Bigint *a, Bigint *b)
1262 #endif
1263 {
1264 	_double da, db;
1265 	int k, ka, kb;
1266 
1267 	if (a == BIGINT_INVALID || b == BIGINT_INVALID)
1268 		return NAN; /* for lack of better value ? */
1269 
1270 	value(da) = b2d(a, &ka);
1271 	value(db) = b2d(b, &kb);
1272 #ifdef Pack_32
1273 	k = ka - kb + 32*(a->wds - b->wds);
1274 #else
1275 	k = ka - kb + 16*(a->wds - b->wds);
1276 #endif
1277 #ifdef IBM
1278 	if (k > 0) {
1279 		word0(da) += (k >> 2)*Exp_msk1;
1280 		if (k &= 3)
1281 			da *= 1 << k;
1282 	}
1283 	else {
1284 		k = -k;
1285 		word0(db) += (k >> 2)*Exp_msk1;
1286 		if (k &= 3)
1287 			db *= 1 << k;
1288 	}
1289 #else
1290 	if (k > 0)
1291 		word0(da) += k*Exp_msk1;
1292 	else {
1293 		k = -k;
1294 		word0(db) += k*Exp_msk1;
1295 	}
1296 #endif
1297 	return value(da) / value(db);
1298 }
1299 
1300 static CONST double
1301 tens[] = {
1302 		1e0, 1e1, 1e2, 1e3, 1e4, 1e5, 1e6, 1e7, 1e8, 1e9,
1303 		1e10, 1e11, 1e12, 1e13, 1e14, 1e15, 1e16, 1e17, 1e18, 1e19,
1304 		1e20, 1e21, 1e22
1305 #ifdef VAX
1306 		, 1e23, 1e24
1307 #endif
1308 };
1309 
1310 #ifdef IEEE_Arith
1311 static CONST double bigtens[] = { 1e16, 1e32, 1e64, 1e128, 1e256 };
1312 static CONST double tinytens[] = { 1e-16, 1e-32, 1e-64, 1e-128, 1e-256 };
1313 #define n_bigtens 5
1314 #else
1315 #ifdef IBM
1316 static CONST double bigtens[] = { 1e16, 1e32, 1e64 };
1317 static CONST double tinytens[] = { 1e-16, 1e-32, 1e-64 };
1318 #define n_bigtens 3
1319 #else
1320 static CONST double bigtens[] = { 1e16, 1e32 };
1321 static CONST double tinytens[] = { 1e-16, 1e-32 };
1322 #define n_bigtens 2
1323 #endif
1324 #endif
1325 
1326  double
strtod(s00,se)1327 strtod
1328 #ifdef KR_headers
1329 	(s00, se) CONST char *s00; char **se;
1330 #else
1331 	(CONST char *s00, char **se)
1332 #endif
1333 {
1334 	int bb2, bb5, bbe, bd2, bd5, bbbits, bs2, c, dsign,
1335 		 e, e1, esign, i, j, k, nd, nd0, nf, nz, nz0, sign;
1336 	CONST char *s, *s0, *s1;
1337 	double aadj, aadj1, adj;
1338 	_double rv, rv0;
1339 	Long L;
1340 	ULong y, z;
1341 	Bigint *bb1, *bd0;
1342 	Bigint *bb = NULL, *bd = NULL, *bs = NULL, *delta = NULL;/* pacify gcc */
1343 
1344 #ifdef ANDROID_CHANGES
1345 	CONST char decimal_point = '.';
1346 #else /* ANDROID_CHANGES */
1347 #ifndef KR_headers
1348 	CONST char decimal_point = localeconv()->decimal_point[0];
1349 #else
1350 	CONST char decimal_point = '.';
1351 #endif
1352 
1353 #endif /* ANDROID_CHANGES */
1354 
1355 	sign = nz0 = nz = 0;
1356 	value(rv) = 0.;
1357 
1358 
1359 	for(s = s00; isspace((unsigned char) *s); s++)
1360 		;
1361 
1362 	if (*s == '-') {
1363 		sign = 1;
1364 		s++;
1365 	} else if (*s == '+') {
1366 		s++;
1367 	}
1368 
1369 	if (*s == '\0') {
1370 		s = s00;
1371 		goto ret;
1372 	}
1373 
1374 	/* "INF" or "INFINITY" */
1375 	if (tolower((unsigned char)*s) == 'i' && strncasecmp(s, "inf", 3) == 0) {
1376 		if (strncasecmp(s + 3, "inity", 5) == 0)
1377 			s += 8;
1378 		else
1379 			s += 3;
1380 
1381 		value(rv) = HUGE_VAL;
1382 		goto ret;
1383 	}
1384 
1385 #ifdef IEEE_Arith
1386 	/* "NAN" or "NAN(n-char-sequence-opt)" */
1387 	if (tolower((unsigned char)*s) == 'n' && strncasecmp(s, "nan", 3) == 0) {
1388 		/* Build a quiet NaN. */
1389 		word0(rv) = NAN_WORD0;
1390 		word1(rv) = NAN_WORD1;
1391 		s+= 3;
1392 
1393 		/* Don't interpret (n-char-sequence-opt), for now. */
1394 		if (*s == '(') {
1395 			s0 = s;
1396 			for (s++; *s != ')' && *s != '\0'; s++)
1397 				;
1398 			if (*s == ')')
1399 				s++;	/* Skip over closing paren ... */
1400 			else
1401 				s = s0;	/* ... otherwise go back. */
1402 		}
1403 
1404 		goto ret;
1405 	}
1406 #endif
1407 
1408 	if (*s == '0') {
1409 		nz0 = 1;
1410 		while(*++s == '0') ;
1411 		if (!*s)
1412 			goto ret;
1413 	}
1414 	s0 = s;
1415 	y = z = 0;
1416 	for(nd = nf = 0; (c = *s) >= '0' && c <= '9'; nd++, s++)
1417 		if (nd < 9)
1418 			y = 10*y + c - '0';
1419 		else if (nd < 16)
1420 			z = 10*z + c - '0';
1421 	nd0 = nd;
1422 	if (c == decimal_point) {
1423 		c = *++s;
1424 		if (!nd) {
1425 			for(; c == '0'; c = *++s)
1426 				nz++;
1427 			if (c > '0' && c <= '9') {
1428 				s0 = s;
1429 				nf += nz;
1430 				nz = 0;
1431 				goto have_dig;
1432 				}
1433 			goto dig_done;
1434 		}
1435 		for(; c >= '0' && c <= '9'; c = *++s) {
1436  have_dig:
1437 			nz++;
1438 			if (c -= '0') {
1439 				nf += nz;
1440 				for(i = 1; i < nz; i++)
1441 					if (nd++ < 9)
1442 						y *= 10;
1443 					else if (nd <= DBL_DIG + 1)
1444 						z *= 10;
1445 				if (nd++ < 9)
1446 					y = 10*y + c;
1447 				else if (nd <= DBL_DIG + 1)
1448 					z = 10*z + c;
1449 				nz = 0;
1450 			}
1451 		}
1452 	}
1453  dig_done:
1454 	e = 0;
1455 	if (c == 'e' || c == 'E') {
1456 		if (!nd && !nz && !nz0) {
1457 			s = s00;
1458 			goto ret;
1459 		}
1460 		s00 = s;
1461 		esign = 0;
1462 		switch(c = *++s) {
1463 			case '-':
1464 				esign = 1;
1465 				/* FALLTHROUGH */
1466 			case '+':
1467 				c = *++s;
1468 		}
1469 		if (c >= '0' && c <= '9') {
1470 			while(c == '0')
1471 				c = *++s;
1472 			if (c > '0' && c <= '9') {
1473 				L = c - '0';
1474 				s1 = s;
1475 				while((c = *++s) >= '0' && c <= '9')
1476 					L = 10*L + c - '0';
1477 				if (s - s1 > 8 || L > 19999)
1478 					/* Avoid confusion from exponents
1479 					 * so large that e might overflow.
1480 					 */
1481 					e = 19999; /* safe for 16 bit ints */
1482 				else
1483 					e = (int)L;
1484 				if (esign)
1485 					e = -e;
1486 			}
1487 			else
1488 				e = 0;
1489 		}
1490 		else
1491 			s = s00;
1492 	}
1493 	if (!nd) {
1494 		if (!nz && !nz0)
1495 			s = s00;
1496 		goto ret;
1497 	}
1498 	e1 = e -= nf;
1499 
1500 	/* Now we have nd0 digits, starting at s0, followed by a
1501 	 * decimal point, followed by nd-nd0 digits.  The number we're
1502 	 * after is the integer represented by those digits times
1503 	 * 10**e */
1504 
1505 	if (!nd0)
1506 		nd0 = nd;
1507 	k = nd < DBL_DIG + 1 ? nd : DBL_DIG + 1;
1508 	value(rv) = y;
1509 	if (k > 9)
1510 		value(rv) = tens[k - 9] * value(rv) + z;
1511 	bd0 = 0;
1512 	if (nd <= DBL_DIG
1513 #ifndef RND_PRODQUOT
1514 		&& FLT_ROUNDS == 1
1515 #endif
1516 		) {
1517 		if (!e)
1518 			goto ret;
1519 		if (e > 0) {
1520 			if (e <= Ten_pmax) {
1521 #ifdef VAX
1522 				goto vax_ovfl_check;
1523 #else
1524 				/* value(rv) = */ rounded_product(value(rv),
1525 				    tens[e]);
1526 				goto ret;
1527 #endif
1528 			}
1529 			i = DBL_DIG - nd;
1530 			if (e <= Ten_pmax + i) {
1531 				/* A fancier test would sometimes let us do
1532 				 * this for larger i values.
1533 				 */
1534 				e -= i;
1535 				value(rv) *= tens[i];
1536 #ifdef VAX
1537 				/* VAX exponent range is so narrow we must
1538 				 * worry about overflow here...
1539 				 */
1540  vax_ovfl_check:
1541 				word0(rv) -= P*Exp_msk1;
1542 				/* value(rv) = */ rounded_product(value(rv),
1543 				    tens[e]);
1544 				if ((word0(rv) & Exp_mask)
1545 				 > Exp_msk1*(DBL_MAX_EXP+Bias-1-P))
1546 					goto ovfl;
1547 				word0(rv) += P*Exp_msk1;
1548 #else
1549 				/* value(rv) = */ rounded_product(value(rv),
1550 				    tens[e]);
1551 #endif
1552 				goto ret;
1553 			}
1554 		}
1555 #ifndef Inaccurate_Divide
1556 		else if (e >= -Ten_pmax) {
1557 			/* value(rv) = */ rounded_quotient(value(rv),
1558 			    tens[-e]);
1559 			goto ret;
1560 		}
1561 #endif
1562 	}
1563 	e1 += nd - k;
1564 
1565 	/* Get starting approximation = rv * 10**e1 */
1566 
1567 	if (e1 > 0) {
1568 		if ((i = e1 & 15) != 0)
1569 			value(rv) *= tens[i];
1570 		if (e1 &= ~15) {
1571 			if (e1 > DBL_MAX_10_EXP) {
1572  ovfl:
1573 				errno = ERANGE;
1574 				value(rv) = HUGE_VAL;
1575 				if (bd0)
1576 					goto retfree;
1577 				goto ret;
1578 			}
1579 			if ((e1 = (unsigned int)e1 >> 4) != 0) {
1580 				for(j = 0; e1 > 1; j++,
1581 				    e1 = (unsigned int)e1 >> 1)
1582 					if (e1 & 1)
1583 						value(rv) *= bigtens[j];
1584 			/* The last multiplication could overflow. */
1585 				word0(rv) -= P*Exp_msk1;
1586 				value(rv) *= bigtens[j];
1587 				if ((z = word0(rv) & Exp_mask)
1588 				 > Exp_msk1*(DBL_MAX_EXP+Bias-P))
1589 					goto ovfl;
1590 				if (z > Exp_msk1*(DBL_MAX_EXP+Bias-1-P)) {
1591 					/* set to largest number */
1592 					/* (Can't trust DBL_MAX) */
1593 					word0(rv) = Big0;
1594 					word1(rv) = Big1;
1595 					}
1596 				else
1597 					word0(rv) += P*Exp_msk1;
1598 			}
1599 		}
1600 	}
1601 	else if (e1 < 0) {
1602 		e1 = -e1;
1603 		if ((i = e1 & 15) != 0)
1604 			value(rv) /= tens[i];
1605 		if (e1 &= ~15) {
1606 			e1 = (unsigned int)e1 >> 4;
1607 			if (e1 >= 1 << n_bigtens)
1608 				goto undfl;
1609 			for(j = 0; e1 > 1; j++,
1610 			    e1 = (unsigned int)e1 >> 1)
1611 				if (e1 & 1)
1612 					value(rv) *= tinytens[j];
1613 			/* The last multiplication could underflow. */
1614 			value(rv0) = value(rv);
1615 			value(rv) *= tinytens[j];
1616 			if (!value(rv)) {
1617 				value(rv) = 2.*value(rv0);
1618 				value(rv) *= tinytens[j];
1619 				if (!value(rv)) {
1620  undfl:
1621 					value(rv) = 0.;
1622 					errno = ERANGE;
1623 					if (bd0)
1624 						goto retfree;
1625 					goto ret;
1626 				}
1627 				word0(rv) = Tiny0;
1628 				word1(rv) = Tiny1;
1629 				/* The refinement below will clean
1630 				 * this approximation up.
1631 				 */
1632 			}
1633 		}
1634 	}
1635 
1636 	/* Now the hard part -- adjusting rv to the correct value.*/
1637 
1638 	/* Put digits into bd: true value = bd * 10^e */
1639 
1640 	bd0 = s2b(s0, nd0, nd, y);
1641 
1642 	for(;;) {
1643 		bd = Balloc(bd0->k);
1644 		Bcopy(bd, bd0);
1645 		bb = d2b(value(rv), &bbe, &bbbits);	/* rv = bb * 2^bbe */
1646 		bs = i2b(1);
1647 
1648 		if (e >= 0) {
1649 			bb2 = bb5 = 0;
1650 			bd2 = bd5 = e;
1651 		}
1652 		else {
1653 			bb2 = bb5 = -e;
1654 			bd2 = bd5 = 0;
1655 		}
1656 		if (bbe >= 0)
1657 			bb2 += bbe;
1658 		else
1659 			bd2 -= bbe;
1660 		bs2 = bb2;
1661 #ifdef Sudden_Underflow
1662 #ifdef IBM
1663 		j = 1 + 4*P - 3 - bbbits + ((bbe + bbbits - 1) & 3);
1664 #else
1665 		j = P + 1 - bbbits;
1666 #endif
1667 #else
1668 		i = bbe + bbbits - 1;	/* logb(rv) */
1669 		if (i < Emin)	/* denormal */
1670 			j = bbe + (P-Emin);
1671 		else
1672 			j = P + 1 - bbbits;
1673 #endif
1674 		bb2 += j;
1675 		bd2 += j;
1676 		i = bb2 < bd2 ? bb2 : bd2;
1677 		if (i > bs2)
1678 			i = bs2;
1679 		if (i > 0) {
1680 			bb2 -= i;
1681 			bd2 -= i;
1682 			bs2 -= i;
1683 		}
1684 		if (bb5 > 0) {
1685 			bs = pow5mult(bs, bb5);
1686 			bb1 = mult(bs, bb);
1687 			Bfree(bb);
1688 			bb = bb1;
1689 		}
1690 		if (bb2 > 0)
1691 			bb = lshift(bb, bb2);
1692 		if (bd5 > 0)
1693 			bd = pow5mult(bd, bd5);
1694 		if (bd2 > 0)
1695 			bd = lshift(bd, bd2);
1696 		if (bs2 > 0)
1697 			bs = lshift(bs, bs2);
1698 		delta = diff(bb, bd);
1699 		dsign = delta->sign;
1700 		delta->sign = 0;
1701 		i = cmp(delta, bs);
1702 		if (i < 0) {
1703 			/* Error is less than half an ulp -- check for
1704 			 * special case of mantissa a power of two.
1705 			 */
1706 			if (dsign || word1(rv) || word0(rv) & Bndry_mask)
1707 				break;
1708 			delta = lshift(delta,Log2P);
1709 			if (cmp(delta, bs) > 0)
1710 				goto drop_down;
1711 			break;
1712 		}
1713 		if (i == 0) {
1714 			/* exactly half-way between */
1715 			if (dsign) {
1716 				if ((word0(rv) & Bndry_mask1) == Bndry_mask1
1717 				 &&  word1(rv) == 0xffffffff) {
1718 					/*boundary case -- increment exponent*/
1719 					word0(rv) = (word0(rv) & Exp_mask)
1720 						+ Exp_msk1
1721 #ifdef IBM
1722 						| Exp_msk1 >> 4
1723 #endif
1724 						;
1725 					word1(rv) = 0;
1726 					break;
1727 				}
1728 			}
1729 			else if (!(word0(rv) & Bndry_mask) && !word1(rv)) {
1730  drop_down:
1731 				/* boundary case -- decrement exponent */
1732 #ifdef Sudden_Underflow
1733 				L = word0(rv) & Exp_mask;
1734 #ifdef IBM
1735 				if (L <  Exp_msk1)
1736 #else
1737 				if (L <= Exp_msk1)
1738 #endif
1739 					goto undfl;
1740 				L -= Exp_msk1;
1741 #else
1742 				L = (word0(rv) & Exp_mask) - Exp_msk1;
1743 #endif
1744 				word0(rv) = L | Bndry_mask1;
1745 				word1(rv) = 0xffffffff;
1746 #ifdef IBM
1747 				goto cont;
1748 #else
1749 				break;
1750 #endif
1751 			}
1752 #ifndef ROUND_BIASED
1753 			if (!(word1(rv) & LSB))
1754 				break;
1755 #endif
1756 			if (dsign)
1757 				value(rv) += ulp(value(rv));
1758 #ifndef ROUND_BIASED
1759 			else {
1760 				value(rv) -= ulp(value(rv));
1761 #ifndef Sudden_Underflow
1762 				if (!value(rv))
1763 					goto undfl;
1764 #endif
1765 			}
1766 #endif
1767 			break;
1768 		}
1769 		if ((aadj = ratio(delta, bs)) <= 2.) {
1770 			if (dsign)
1771 				aadj = aadj1 = 1.;
1772 			else if (word1(rv) || word0(rv) & Bndry_mask) {
1773 #ifndef Sudden_Underflow
1774 				if (word1(rv) == Tiny1 && !word0(rv))
1775 					goto undfl;
1776 #endif
1777 				aadj = 1.;
1778 				aadj1 = -1.;
1779 			}
1780 			else {
1781 				/* special case -- power of FLT_RADIX to be */
1782 				/* rounded down... */
1783 
1784 				if (aadj < 2./FLT_RADIX)
1785 					aadj = 1./FLT_RADIX;
1786 				else
1787 					aadj *= 0.5;
1788 				aadj1 = -aadj;
1789 				}
1790 		}
1791 		else {
1792 			aadj *= 0.5;
1793 			aadj1 = dsign ? aadj : -aadj;
1794 #ifdef Check_FLT_ROUNDS
1795 			switch(FLT_ROUNDS) {
1796 				case 2: /* towards +infinity */
1797 					aadj1 -= 0.5;
1798 					break;
1799 				case 0: /* towards 0 */
1800 				case 3: /* towards -infinity */
1801 					aadj1 += 0.5;
1802 			}
1803 #else
1804 			if (FLT_ROUNDS == 0)
1805 				aadj1 += 0.5;
1806 #endif
1807 		}
1808 		y = word0(rv) & Exp_mask;
1809 
1810 		/* Check for overflow */
1811 
1812 		if (y == Exp_msk1*(DBL_MAX_EXP+Bias-1)) {
1813 			value(rv0) = value(rv);
1814 			word0(rv) -= P*Exp_msk1;
1815 			adj = aadj1 * ulp(value(rv));
1816 			value(rv) += adj;
1817 			if ((word0(rv) & Exp_mask) >=
1818 					Exp_msk1*(DBL_MAX_EXP+Bias-P)) {
1819 				if (word0(rv0) == Big0 && word1(rv0) == Big1)
1820 					goto ovfl;
1821 				word0(rv) = Big0;
1822 				word1(rv) = Big1;
1823 				goto cont;
1824 			}
1825 			else
1826 				word0(rv) += P*Exp_msk1;
1827 		}
1828 		else {
1829 #ifdef Sudden_Underflow
1830 			if ((word0(rv) & Exp_mask) <= P*Exp_msk1) {
1831 				value(rv0) = value(rv);
1832 				word0(rv) += P*Exp_msk1;
1833 				adj = aadj1 * ulp(value(rv));
1834 				value(rv) += adj;
1835 #ifdef IBM
1836 				if ((word0(rv) & Exp_mask) <  P*Exp_msk1)
1837 #else
1838 				if ((word0(rv) & Exp_mask) <= P*Exp_msk1)
1839 #endif
1840 				{
1841 					if (word0(rv0) == Tiny0
1842 					 && word1(rv0) == Tiny1)
1843 						goto undfl;
1844 					word0(rv) = Tiny0;
1845 					word1(rv) = Tiny1;
1846 					goto cont;
1847 				}
1848 				else
1849 					word0(rv) -= P*Exp_msk1;
1850 				}
1851 			else {
1852 				adj = aadj1 * ulp(value(rv));
1853 				value(rv) += adj;
1854 			}
1855 #else
1856 			/* Compute adj so that the IEEE rounding rules will
1857 			 * correctly round rv + adj in some half-way cases.
1858 			 * If rv * ulp(rv) is denormalized (i.e.,
1859 			 * y <= (P-1)*Exp_msk1), we must adjust aadj to avoid
1860 			 * trouble from bits lost to denormalization;
1861 			 * example: 1.2e-307 .
1862 			 */
1863 			if (y <= (P-1)*Exp_msk1 && aadj >= 1.) {
1864 				aadj1 = (double)(int)(aadj + 0.5);
1865 				if (!dsign)
1866 					aadj1 = -aadj1;
1867 			}
1868 			adj = aadj1 * ulp(value(rv));
1869 			value(rv) += adj;
1870 #endif
1871 		}
1872 		z = word0(rv) & Exp_mask;
1873 		if (y == z) {
1874 			/* Can we stop now? */
1875 			L = aadj;
1876 			aadj -= L;
1877 			/* The tolerances below are conservative. */
1878 			if (dsign || word1(rv) || word0(rv) & Bndry_mask) {
1879 				if (aadj < .4999999 || aadj > .5000001)
1880 					break;
1881 			}
1882 			else if (aadj < .4999999/FLT_RADIX)
1883 				break;
1884 		}
1885  cont:
1886 		Bfree(bb);
1887 		Bfree(bd);
1888 		Bfree(bs);
1889 		Bfree(delta);
1890 	}
1891  retfree:
1892 	Bfree(bb);
1893 	Bfree(bd);
1894 	Bfree(bs);
1895 	Bfree(bd0);
1896 	Bfree(delta);
1897  ret:
1898 	if (se)
1899 		/* LINTED interface specification */
1900 		*se = (char *)s;
1901 	return sign ? -value(rv) : value(rv);
1902 }
1903 
1904  static int
quorem(b,S)1905 quorem
1906 #ifdef KR_headers
1907 	(b, S) Bigint *b, *S;
1908 #else
1909 	(Bigint *b, Bigint *S)
1910 #endif
1911 {
1912 	int n;
1913 	Long borrow, y;
1914 	ULong carry, q, ys;
1915 	ULong *bx, *bxe, *sx, *sxe;
1916 #ifdef Pack_32
1917 	Long z;
1918 	ULong si, zs;
1919 #endif
1920 
1921 	if (b == BIGINT_INVALID || S == BIGINT_INVALID)
1922 		return 0;
1923 
1924 	n = S->wds;
1925 #ifdef DEBUG
1926 	/*debug*/ if (b->wds > n)
1927 	/*debug*/	Bug("oversize b in quorem");
1928 #endif
1929 	if (b->wds < n)
1930 		return 0;
1931 	sx = S->x;
1932 	sxe = sx + --n;
1933 	bx = b->x;
1934 	bxe = bx + n;
1935 	q = *bxe / (*sxe + 1);	/* ensure q <= true quotient */
1936 #ifdef DEBUG
1937 	/*debug*/ if (q > 9)
1938 	/*debug*/	Bug("oversized quotient in quorem");
1939 #endif
1940 	if (q) {
1941 		borrow = 0;
1942 		carry = 0;
1943 		do {
1944 #ifdef Pack_32
1945 			si = *sx++;
1946 			ys = (si & 0xffff) * q + carry;
1947 			zs = (si >> 16) * q + (ys >> 16);
1948 			carry = zs >> 16;
1949 			y = (*bx & 0xffff) - (ys & 0xffff) + borrow;
1950 			borrow = (ULong)y >> 16;
1951 			Sign_Extend(borrow, y);
1952 			z = (*bx >> 16) - (zs & 0xffff) + borrow;
1953 			borrow = (ULong)z >> 16;
1954 			Sign_Extend(borrow, z);
1955 			Storeinc(bx, z, y);
1956 #else
1957 			ys = *sx++ * q + carry;
1958 			carry = ys >> 16;
1959 			y = *bx - (ys & 0xffff) + borrow;
1960 			borrow = y >> 16;
1961 			Sign_Extend(borrow, y);
1962 			*bx++ = y & 0xffff;
1963 #endif
1964 		}
1965 		while(sx <= sxe);
1966 		if (!*bxe) {
1967 			bx = b->x;
1968 			while(--bxe > bx && !*bxe)
1969 				--n;
1970 			b->wds = n;
1971 		}
1972 	}
1973 	if (cmp(b, S) >= 0) {
1974 		q++;
1975 		borrow = 0;
1976 		carry = 0;
1977 		bx = b->x;
1978 		sx = S->x;
1979 		do {
1980 #ifdef Pack_32
1981 			si = *sx++;
1982 			ys = (si & 0xffff) + carry;
1983 			zs = (si >> 16) + (ys >> 16);
1984 			carry = zs >> 16;
1985 			y = (*bx & 0xffff) - (ys & 0xffff) + borrow;
1986 			borrow = (ULong)y >> 16;
1987 			Sign_Extend(borrow, y);
1988 			z = (*bx >> 16) - (zs & 0xffff) + borrow;
1989 			borrow = (ULong)z >> 16;
1990 			Sign_Extend(borrow, z);
1991 			Storeinc(bx, z, y);
1992 #else
1993 			ys = *sx++ + carry;
1994 			carry = ys >> 16;
1995 			y = *bx - (ys & 0xffff) + borrow;
1996 			borrow = y >> 16;
1997 			Sign_Extend(borrow, y);
1998 			*bx++ = y & 0xffff;
1999 #endif
2000 		}
2001 		while(sx <= sxe);
2002 		bx = b->x;
2003 		bxe = bx + n;
2004 		if (!*bxe) {
2005 			while(--bxe > bx && !*bxe)
2006 				--n;
2007 			b->wds = n;
2008 		}
2009 	}
2010 	return q;
2011 }
2012 
2013 /* freedtoa(s) must be used to free values s returned by dtoa
2014  * when MULTIPLE_THREADS is #defined.  It should be used in all cases,
2015  * but for consistency with earlier versions of dtoa, it is optional
2016  * when MULTIPLE_THREADS is not defined.
2017  */
2018 
2019 void
2020 #ifdef KR_headers
freedtoa(s)2021 freedtoa(s) char *s;
2022 #else
2023 freedtoa(char *s)
2024 #endif
2025 {
2026 	free(s);
2027 }
2028 
2029 
2030 
2031 /* dtoa for IEEE arithmetic (dmg): convert double to ASCII string.
2032  *
2033  * Inspired by "How to Print Floating-Point Numbers Accurately" by
2034  * Guy L. Steele, Jr. and Jon L. White [Proc. ACM SIGPLAN '90, pp. 92-101].
2035  *
2036  * Modifications:
2037  *	1. Rather than iterating, we use a simple numeric overestimate
2038  *	   to determine k = floor(log10(d)).  We scale relevant
2039  *	   quantities using O(log2(k)) rather than O(k) multiplications.
2040  *	2. For some modes > 2 (corresponding to ecvt and fcvt), we don't
2041  *	   try to generate digits strictly left to right.  Instead, we
2042  *	   compute with fewer bits and propagate the carry if necessary
2043  *	   when rounding the final digit up.  This is often faster.
2044  *	3. Under the assumption that input will be rounded nearest,
2045  *	   mode 0 renders 1e23 as 1e23 rather than 9.999999999999999e22.
2046  *	   That is, we allow equality in stopping tests when the
2047  *	   round-nearest rule will give the same floating-point value
2048  *	   as would satisfaction of the stopping test with strict
2049  *	   inequality.
2050  *	4. We remove common factors of powers of 2 from relevant
2051  *	   quantities.
2052  *	5. When converting floating-point integers less than 1e16,
2053  *	   we use floating-point arithmetic rather than resorting
2054  *	   to multiple-precision integers.
2055  *	6. When asked to produce fewer than 15 digits, we first try
2056  *	   to get by with floating-point arithmetic; we resort to
2057  *	   multiple-precision integer arithmetic only if we cannot
2058  *	   guarantee that the floating-point calculation has given
2059  *	   the correctly rounded result.  For k requested digits and
2060  *	   "uniformly" distributed input, the probability is
2061  *	   something like 10^(k-15) that we must resort to the Long
2062  *	   calculation.
2063  */
2064 
2065  char *
__dtoa(_d,mode,ndigits,decpt,sign,rve)2066 __dtoa
2067 #ifdef KR_headers
2068 	(_d, mode, ndigits, decpt, sign, rve)
2069 	double _d; int mode, ndigits, *decpt, *sign; char **rve;
2070 #else
2071 	(double _d, int mode, int ndigits, int *decpt, int *sign, char **rve)
2072 #endif
2073 {
2074  /*	Arguments ndigits, decpt, sign are similar to those
2075 	of ecvt and fcvt; trailing zeros are suppressed from
2076 	the returned string.  If not null, *rve is set to point
2077 	to the end of the return value.  If d is +-Infinity or NaN,
2078 	then *decpt is set to 9999.
2079 
2080 	mode:
2081 		0 ==> shortest string that yields d when read in
2082 			and rounded to nearest.
2083 		1 ==> like 0, but with Steele & White stopping rule;
2084 			e.g. with IEEE P754 arithmetic , mode 0 gives
2085 			1e23 whereas mode 1 gives 9.999999999999999e22.
2086 		2 ==> max(1,ndigits) significant digits.  This gives a
2087 			return value similar to that of ecvt, except
2088 			that trailing zeros are suppressed.
2089 		3 ==> through ndigits past the decimal point.  This
2090 			gives a return value similar to that from fcvt,
2091 			except that trailing zeros are suppressed, and
2092 			ndigits can be negative.
2093 		4-9 should give the same return values as 2-3, i.e.,
2094 			4 <= mode <= 9 ==> same return as mode
2095 			2 + (mode & 1).  These modes are mainly for
2096 			debugging; often they run slower but sometimes
2097 			faster than modes 2-3.
2098 		4,5,8,9 ==> left-to-right digit generation.
2099 		6-9 ==> don't try fast floating-point estimate
2100 			(if applicable).
2101 
2102 		Values of mode other than 0-9 are treated as mode 0.
2103 
2104 		Sufficient space is allocated to the return value
2105 		to hold the suppressed trailing zeros.
2106 	*/
2107 
2108 	int bbits, b2, b5, be, dig, i, ieps, ilim0,
2109 		j, jj1, k, k0, k_check, leftright, m2, m5, s2, s5,
2110 		try_quick;
2111 	int ilim = 0, ilim1 = 0, spec_case = 0;	/* pacify gcc */
2112 	Long L;
2113 #ifndef Sudden_Underflow
2114 	int denorm;
2115 	ULong x;
2116 #endif
2117 	Bigint *b, *b1, *delta, *mhi, *S;
2118 	Bigint *mlo = NULL; /* pacify gcc */
2119 	double ds;
2120 	char *s, *s0;
2121 	Bigint *result = NULL;
2122 	int result_k = 0;
2123 	_double d, d2, eps;
2124 
2125 	value(d) = _d;
2126 
2127 	if (word0(d) & Sign_bit) {
2128 		/* set sign for everything, including 0's and NaNs */
2129 		*sign = 1;
2130 		word0(d) &= ~Sign_bit;	/* clear sign bit */
2131 	}
2132 	else
2133 		*sign = 0;
2134 
2135 #if defined(IEEE_Arith) + defined(VAX)
2136 #ifdef IEEE_Arith
2137 	if ((word0(d) & Exp_mask) == Exp_mask)
2138 #else
2139 	if (word0(d)  == 0x8000)
2140 #endif
2141 	{
2142 		/* Infinity or NaN */
2143 		*decpt = 9999;
2144 		s =
2145 #ifdef IEEE_Arith
2146 			!word1(d) && !(word0(d) & 0xfffff) ? "Infinity" :
2147 #endif
2148 				"NaN";
2149 		result = Balloc(strlen(s)+1);
2150 		if (result == BIGINT_INVALID)
2151 			return NULL;
2152 		s0 = (char *)(void *)result;
2153 		strcpy(s0, s);
2154 		if (rve)
2155 			*rve =
2156 #ifdef IEEE_Arith
2157 				s0[3] ? s0 + 8 :
2158 #endif
2159 				s0 + 3;
2160 		return s0;
2161 	}
2162 #endif
2163 #ifdef IBM
2164 	value(d) += 0; /* normalize */
2165 #endif
2166 	if (!value(d)) {
2167 		*decpt = 1;
2168 		result = Balloc(2);
2169 		if (result == BIGINT_INVALID)
2170 			return NULL;
2171 		s0 = (char *)(void *)result;
2172 		strcpy(s0, "0");
2173 		if (rve)
2174 			*rve = s0 + 1;
2175 		return s0;
2176 	}
2177 
2178 	b = d2b(value(d), &be, &bbits);
2179 #ifdef Sudden_Underflow
2180 	i = (int)(word0(d) >> Exp_shift1 & (Exp_mask>>Exp_shift1));
2181 #else
2182 	if ((i = (int)(word0(d) >> Exp_shift1 & (Exp_mask>>Exp_shift1))) != 0) {
2183 #endif
2184 		value(d2) = value(d);
2185 		word0(d2) &= Frac_mask1;
2186 		word0(d2) |= Exp_11;
2187 #ifdef IBM
2188 		if (j = 11 - hi0bits(word0(d2) & Frac_mask))
2189 			value(d2) /= 1 << j;
2190 #endif
2191 
2192 		/* log(x)	~=~ log(1.5) + (x-1.5)/1.5
2193 		 * log10(x)	 =  log(x) / log(10)
2194 		 *		~=~ log(1.5)/log(10) + (x-1.5)/(1.5*log(10))
2195 		 * log10(d) = (i-Bias)*log(2)/log(10) + log10(d2)
2196 		 *
2197 		 * This suggests computing an approximation k to log10(d) by
2198 		 *
2199 		 * k = (i - Bias)*0.301029995663981
2200 		 *	+ ( (d2-1.5)*0.289529654602168 + 0.176091259055681 );
2201 		 *
2202 		 * We want k to be too large rather than too small.
2203 		 * The error in the first-order Taylor series approximation
2204 		 * is in our favor, so we just round up the constant enough
2205 		 * to compensate for any error in the multiplication of
2206 		 * (i - Bias) by 0.301029995663981; since |i - Bias| <= 1077,
2207 		 * and 1077 * 0.30103 * 2^-52 ~=~ 7.2e-14,
2208 		 * adding 1e-13 to the constant term more than suffices.
2209 		 * Hence we adjust the constant term to 0.1760912590558.
2210 		 * (We could get a more accurate k by invoking log10,
2211 		 *  but this is probably not worthwhile.)
2212 		 */
2213 
2214 		i -= Bias;
2215 #ifdef IBM
2216 		i <<= 2;
2217 		i += j;
2218 #endif
2219 #ifndef Sudden_Underflow
2220 		denorm = 0;
2221 	}
2222 	else {
2223 		/* d is denormalized */
2224 
2225 		i = bbits + be + (Bias + (P-1) - 1);
2226 		x = i > 32  ? word0(d) << (64 - i) | word1(d) >> (i - 32)
2227 			    : word1(d) << (32 - i);
2228 		value(d2) = x;
2229 		word0(d2) -= 31*Exp_msk1; /* adjust exponent */
2230 		i -= (Bias + (P-1) - 1) + 1;
2231 		denorm = 1;
2232 	}
2233 #endif
2234 	ds = (value(d2)-1.5)*0.289529654602168 + 0.1760912590558 +
2235 	    i*0.301029995663981;
2236 	k = (int)ds;
2237 	if (ds < 0. && ds != k)
2238 		k--;	/* want k = floor(ds) */
2239 	k_check = 1;
2240 	if (k >= 0 && k <= Ten_pmax) {
2241 		if (value(d) < tens[k])
2242 			k--;
2243 		k_check = 0;
2244 	}
2245 	j = bbits - i - 1;
2246 	if (j >= 0) {
2247 		b2 = 0;
2248 		s2 = j;
2249 	}
2250 	else {
2251 		b2 = -j;
2252 		s2 = 0;
2253 	}
2254 	if (k >= 0) {
2255 		b5 = 0;
2256 		s5 = k;
2257 		s2 += k;
2258 	}
2259 	else {
2260 		b2 -= k;
2261 		b5 = -k;
2262 		s5 = 0;
2263 	}
2264 	if (mode < 0 || mode > 9)
2265 		mode = 0;
2266 	try_quick = 1;
2267 	if (mode > 5) {
2268 		mode -= 4;
2269 		try_quick = 0;
2270 	}
2271 	leftright = 1;
2272 	switch(mode) {
2273 		case 0:
2274 		case 1:
2275 			ilim = ilim1 = -1;
2276 			i = 18;
2277 			ndigits = 0;
2278 			break;
2279 		case 2:
2280 			leftright = 0;
2281 			/* FALLTHROUGH */
2282 		case 4:
2283 			if (ndigits <= 0)
2284 				ndigits = 1;
2285 			ilim = ilim1 = i = ndigits;
2286 			break;
2287 		case 3:
2288 			leftright = 0;
2289 			/* FALLTHROUGH */
2290 		case 5:
2291 			i = ndigits + k + 1;
2292 			ilim = i;
2293 			ilim1 = i - 1;
2294 			if (i <= 0)
2295 				i = 1;
2296 	}
2297 	j = sizeof(ULong);
2298         for(result_k = 0; (int)(sizeof(Bigint) - sizeof(ULong)) + j <= i;
2299 		j <<= 1) result_k++;
2300         // this is really a ugly hack, the code uses Balloc
2301         // instead of malloc, but casts the result into a char*
2302         // it seems the only reason to do that is due to the
2303         // complicated way the block size need to be computed
2304         // buuurk....
2305 	result = Balloc(result_k);
2306 	if (result == BIGINT_INVALID) {
2307 		Bfree(b);
2308 		return NULL;
2309 	}
2310 	s = s0 = (char *)(void *)result;
2311 
2312 	if (ilim >= 0 && ilim <= Quick_max && try_quick) {
2313 
2314 		/* Try to get by with floating-point arithmetic. */
2315 
2316 		i = 0;
2317 		value(d2) = value(d);
2318 		k0 = k;
2319 		ilim0 = ilim;
2320 		ieps = 2; /* conservative */
2321 		if (k > 0) {
2322 			ds = tens[k&0xf];
2323 			j = (unsigned int)k >> 4;
2324 			if (j & Bletch) {
2325 				/* prevent overflows */
2326 				j &= Bletch - 1;
2327 				value(d) /= bigtens[n_bigtens-1];
2328 				ieps++;
2329 				}
2330 			for(; j; j = (unsigned int)j >> 1, i++)
2331 				if (j & 1) {
2332 					ieps++;
2333 					ds *= bigtens[i];
2334 					}
2335 			value(d) /= ds;
2336 		}
2337 		else if ((jj1 = -k) != 0) {
2338 			value(d) *= tens[jj1 & 0xf];
2339 			for(j = (unsigned int)jj1 >> 4; j;
2340 			    j = (unsigned int)j >> 1, i++)
2341 				if (j & 1) {
2342 					ieps++;
2343 					value(d) *= bigtens[i];
2344 				}
2345 		}
2346 		if (k_check && value(d) < 1. && ilim > 0) {
2347 			if (ilim1 <= 0)
2348 				goto fast_failed;
2349 			ilim = ilim1;
2350 			k--;
2351 			value(d) *= 10.;
2352 			ieps++;
2353 		}
2354 		value(eps) = ieps*value(d) + 7.;
2355 		word0(eps) -= (P-1)*Exp_msk1;
2356 		if (ilim == 0) {
2357 			S = mhi = 0;
2358 			value(d) -= 5.;
2359 			if (value(d) > value(eps))
2360 				goto one_digit;
2361 			if (value(d) < -value(eps))
2362 				goto no_digits;
2363 			goto fast_failed;
2364 		}
2365 #ifndef No_leftright
2366 		if (leftright) {
2367 			/* Use Steele & White method of only
2368 			 * generating digits needed.
2369 			 */
2370 			value(eps) = 0.5/tens[ilim-1] - value(eps);
2371 			for(i = 0;;) {
2372 				L = value(d);
2373 				value(d) -= L;
2374 				*s++ = '0' + (int)L;
2375 				if (value(d) < value(eps))
2376 					goto ret1;
2377 				if (1. - value(d) < value(eps))
2378 					goto bump_up;
2379 				if (++i >= ilim)
2380 					break;
2381 				value(eps) *= 10.;
2382 				value(d) *= 10.;
2383 				}
2384 		}
2385 		else {
2386 #endif
2387 			/* Generate ilim digits, then fix them up. */
2388 			value(eps) *= tens[ilim-1];
2389 			for(i = 1;; i++, value(d) *= 10.) {
2390 				L = value(d);
2391 				value(d) -= L;
2392 				*s++ = '0' + (int)L;
2393 				if (i == ilim) {
2394 					if (value(d) > 0.5 + value(eps))
2395 						goto bump_up;
2396 					else if (value(d) < 0.5 - value(eps)) {
2397 						while(*--s == '0');
2398 						s++;
2399 						goto ret1;
2400 						}
2401 					break;
2402 				}
2403 			}
2404 #ifndef No_leftright
2405 		}
2406 #endif
2407  fast_failed:
2408 		s = s0;
2409 		value(d) = value(d2);
2410 		k = k0;
2411 		ilim = ilim0;
2412 	}
2413 
2414 	/* Do we have a "small" integer? */
2415 
2416 	if (be >= 0 && k <= Int_max) {
2417 		/* Yes. */
2418 		ds = tens[k];
2419 		if (ndigits < 0 && ilim <= 0) {
2420 			S = mhi = 0;
2421 			if (ilim < 0 || value(d) <= 5*ds)
2422 				goto no_digits;
2423 			goto one_digit;
2424 		}
2425 		for(i = 1;; i++) {
2426 			L = value(d) / ds;
2427 			value(d) -= L*ds;
2428 #ifdef Check_FLT_ROUNDS
2429 			/* If FLT_ROUNDS == 2, L will usually be high by 1 */
2430 			if (value(d) < 0) {
2431 				L--;
2432 				value(d) += ds;
2433 			}
2434 #endif
2435 			*s++ = '0' + (int)L;
2436 			if (i == ilim) {
2437 				value(d) += value(d);
2438 				if (value(d) > ds || (value(d) == ds && L & 1)) {
2439  bump_up:
2440 					while(*--s == '9')
2441 						if (s == s0) {
2442 							k++;
2443 							*s = '0';
2444 							break;
2445 						}
2446 					++*s++;
2447 				}
2448 				break;
2449 			}
2450 			if (!(value(d) *= 10.))
2451 				break;
2452 			}
2453 		goto ret1;
2454 	}
2455 
2456 	m2 = b2;
2457 	m5 = b5;
2458 	mhi = mlo = 0;
2459 	if (leftright) {
2460 		if (mode < 2) {
2461 			i =
2462 #ifndef Sudden_Underflow
2463 				denorm ? be + (Bias + (P-1) - 1 + 1) :
2464 #endif
2465 #ifdef IBM
2466 				1 + 4*P - 3 - bbits + ((bbits + be - 1) & 3);
2467 #else
2468 				1 + P - bbits;
2469 #endif
2470 		}
2471 		else {
2472 			j = ilim - 1;
2473 			if (m5 >= j)
2474 				m5 -= j;
2475 			else {
2476 				s5 += j -= m5;
2477 				b5 += j;
2478 				m5 = 0;
2479 			}
2480 			if ((i = ilim) < 0) {
2481 				m2 -= i;
2482 				i = 0;
2483 			}
2484 		}
2485 		b2 += i;
2486 		s2 += i;
2487 		mhi = i2b(1);
2488 	}
2489 	if (m2 > 0 && s2 > 0) {
2490 		i = m2 < s2 ? m2 : s2;
2491 		b2 -= i;
2492 		m2 -= i;
2493 		s2 -= i;
2494 	}
2495 	if (b5 > 0) {
2496 		if (leftright) {
2497 			if (m5 > 0) {
2498 				mhi = pow5mult(mhi, m5);
2499 				b1 = mult(mhi, b);
2500 				Bfree(b);
2501 				b = b1;
2502 			}
2503 			if ((j = b5 - m5) != 0)
2504 				b = pow5mult(b, j);
2505 			}
2506 		else
2507 			b = pow5mult(b, b5);
2508 	}
2509 	S = i2b(1);
2510 	if (s5 > 0)
2511 		S = pow5mult(S, s5);
2512 
2513 	/* Check for special case that d is a normalized power of 2. */
2514 
2515 	if (mode < 2) {
2516 		if (!word1(d) && !(word0(d) & Bndry_mask)
2517 #ifndef Sudden_Underflow
2518 		 && word0(d) & Exp_mask
2519 #endif
2520 				) {
2521 			/* The special case */
2522 			b2 += Log2P;
2523 			s2 += Log2P;
2524 			spec_case = 1;
2525 			}
2526 		else
2527 			spec_case = 0;
2528 	}
2529 
2530 	/* Arrange for convenient computation of quotients:
2531 	 * shift left if necessary so divisor has 4 leading 0 bits.
2532 	 *
2533 	 * Perhaps we should just compute leading 28 bits of S once
2534 	 * and for all and pass them and a shift to quorem, so it
2535 	 * can do shifts and ors to compute the numerator for q.
2536 	 */
2537 	if (S == BIGINT_INVALID) {
2538 		i = 0;
2539 	} else {
2540 #ifdef Pack_32
2541 		if ((i = ((s5 ? 32 - hi0bits(S->x[S->wds-1]) : 1) + s2) & 0x1f) != 0)
2542 			i = 32 - i;
2543 #else
2544 		if (i = ((s5 ? 32 - hi0bits(S->x[S->wds-1]) : 1) + s2) & 0xf)
2545 			i = 16 - i;
2546 #endif
2547 	}
2548 
2549 	if (i > 4) {
2550 		i -= 4;
2551 		b2 += i;
2552 		m2 += i;
2553 		s2 += i;
2554 	}
2555 	else if (i < 4) {
2556 		i += 28;
2557 		b2 += i;
2558 		m2 += i;
2559 		s2 += i;
2560 	}
2561 	if (b2 > 0)
2562 		b = lshift(b, b2);
2563 	if (s2 > 0)
2564 		S = lshift(S, s2);
2565 	if (k_check) {
2566 		if (cmp(b,S) < 0) {
2567 			k--;
2568 			b = multadd(b, 10, 0);	/* we botched the k estimate */
2569 			if (leftright)
2570 				mhi = multadd(mhi, 10, 0);
2571 			ilim = ilim1;
2572 			}
2573 	}
2574 	if (ilim <= 0 && mode > 2) {
2575 		if (ilim < 0 || cmp(b,S = multadd(S,5,0)) <= 0) {
2576 			/* no digits, fcvt style */
2577  no_digits:
2578 			k = -1 - ndigits;
2579 			goto ret;
2580 		}
2581  one_digit:
2582 		*s++ = '1';
2583 		k++;
2584 		goto ret;
2585 	}
2586 	if (leftright) {
2587 		if (m2 > 0)
2588 			mhi = lshift(mhi, m2);
2589 
2590 		/* Compute mlo -- check for special case
2591 		 * that d is a normalized power of 2.
2592 		 */
2593 
2594 		mlo = mhi;
2595 		if (spec_case) {
2596 			mhi = Balloc(mhi->k);
2597 			Bcopy(mhi, mlo);
2598 			mhi = lshift(mhi, Log2P);
2599 		}
2600 
2601 		for(i = 1;;i++) {
2602 			dig = quorem(b,S) + '0';
2603 			/* Do we yet have the shortest decimal string
2604 			 * that will round to d?
2605 			 */
2606 			j = cmp(b, mlo);
2607 			delta = diff(S, mhi);
2608 			jj1 = delta->sign ? 1 : cmp(b, delta);
2609 			Bfree(delta);
2610 #ifndef ROUND_BIASED
2611 			if (jj1 == 0 && !mode && !(word1(d) & 1)) {
2612 				if (dig == '9')
2613 					goto round_9_up;
2614 				if (j > 0)
2615 					dig++;
2616 				*s++ = dig;
2617 				goto ret;
2618 			}
2619 #endif
2620 			if (j < 0 || (j == 0 && !mode
2621 #ifndef ROUND_BIASED
2622 							&& !(word1(d) & 1)
2623 #endif
2624 					)) {
2625 				if (jj1 > 0) {
2626 					b = lshift(b, 1);
2627 					jj1 = cmp(b, S);
2628 					if ((jj1 > 0 || (jj1 == 0 && dig & 1))
2629 					&& dig++ == '9')
2630 						goto round_9_up;
2631 					}
2632 				*s++ = dig;
2633 				goto ret;
2634 			}
2635 			if (jj1 > 0) {
2636 				if (dig == '9') { /* possible if i == 1 */
2637  round_9_up:
2638 					*s++ = '9';
2639 					goto roundoff;
2640 					}
2641 				*s++ = dig + 1;
2642 				goto ret;
2643 			}
2644 			*s++ = dig;
2645 			if (i == ilim)
2646 				break;
2647 			b = multadd(b, 10, 0);
2648 			if (mlo == mhi)
2649 				mlo = mhi = multadd(mhi, 10, 0);
2650 			else {
2651 				mlo = multadd(mlo, 10, 0);
2652 				mhi = multadd(mhi, 10, 0);
2653 			}
2654 		}
2655 	}
2656 	else
2657 		for(i = 1;; i++) {
2658 			*s++ = dig = quorem(b,S) + '0';
2659 			if (i >= ilim)
2660 				break;
2661 			b = multadd(b, 10, 0);
2662 		}
2663 
2664 	/* Round off last digit */
2665 
2666 	b = lshift(b, 1);
2667 	j = cmp(b, S);
2668 	if (j > 0 || (j == 0 && dig & 1)) {
2669  roundoff:
2670 		while(*--s == '9')
2671 			if (s == s0) {
2672 				k++;
2673 				*s++ = '1';
2674 				goto ret;
2675 				}
2676 		++*s++;
2677 	}
2678 	else {
2679 		while(*--s == '0');
2680 		s++;
2681 	}
2682  ret:
2683 	Bfree(S);
2684 	if (mhi) {
2685 		if (mlo && mlo != mhi)
2686 			Bfree(mlo);
2687 		Bfree(mhi);
2688 	}
2689  ret1:
2690 	Bfree(b);
2691 	if (s == s0) {				/* don't return empty string */
2692 		*s++ = '0';
2693 		k = 0;
2694 	}
2695 	*s = 0;
2696 	*decpt = k + 1;
2697 	if (rve)
2698 		*rve = s;
2699 	return s0;
2700 }
2701 #ifdef __cplusplus
2702 }
2703 #endif
2704