1 /****************************************************************
3 * The author of this software is David M. Gay.
5 * Copyright (c) 1991, 2000, 2001 by Lucent Technologies.
7 * Permission to use, copy, modify, and distribute this software for any
8 * purpose without fee is hereby granted, provided that this entire notice
9 * is included in all copies of any software which is or includes a copy
10 * or modification of this software and in all copies of the supporting
11 * documentation for such software.
13 * THIS SOFTWARE IS BEING PROVIDED "AS IS", WITHOUT ANY EXPRESS OR IMPLIED
14 * WARRANTY. IN PARTICULAR, NEITHER THE AUTHOR NOR LUCENT MAKES ANY
15 * REPRESENTATION OR WARRANTY OF ANY KIND CONCERNING THE MERCHANTABILITY
16 * OF THIS SOFTWARE OR ITS FITNESS FOR ANY PARTICULAR PURPOSE.
18 ***************************************************************/
21 #define freedtoa __freedtoa
24 G_LOCK_DEFINE_STATIC(str_mutex0
);
25 G_LOCK_DEFINE_STATIC(str_mutex1
);
26 #define MULTIPLE_THREADS 1
27 #define ACQUIRE_DTOA_LOCK(n) G_LOCK (str_mutex##n)
28 #define FREE_DTOA_LOCK(n) G_UNLOCK (str_mutex##n)
30 /* Please send bug reports to David M. Gay (dmg at acm dot org,
31 * with " at " changed at "@" and " dot " changed to "."). */
33 /* On a machine with IEEE extended-precision registers, it is
34 * necessary to specify double-precision (53-bit) rounding precision
35 * before invoking strtod or dtoa. If the machine uses (the equivalent
36 * of) Intel 80x87 arithmetic, the call
37 * _control87(PC_53, MCW_PC);
38 * does this with many compilers. Whether this or another call is
39 * appropriate depends on the compiler; for this to work, it may be
40 * necessary to #include "float.h" or another system-dependent header
44 /* strtod for IEEE-, VAX-, and IBM-arithmetic machines.
46 * This strtod returns a nearest machine number to the input decimal
47 * string (or sets errno to ERANGE). With IEEE arithmetic, ties are
48 * broken by the IEEE round-even rule. Otherwise ties are broken by
49 * biased rounding (add half and chop).
51 * Inspired loosely by William D. Clinger's paper "How to Read Floating
52 * Point Numbers Accurately" [Proc. ACM SIGPLAN '90, pp. 92-101].
56 * 1. We only require IEEE, IBM, or VAX double-precision
57 * arithmetic (not IEEE double-extended).
58 * 2. We get by with floating-point arithmetic in a case that
59 * Clinger missed -- when we're computing d * 10^n
60 * for a small integer d and the integer n is not too
61 * much larger than 22 (the maximum integer k for which
62 * we can represent 10^k exactly), we may be able to
63 * compute (d*10^k) * 10^(e-k) with just one roundoff.
64 * 3. Rather than a bit-at-a-time adjustment of the binary
65 * result in the hard case, we use floating-point
66 * arithmetic to determine the adjustment to within
67 * one bit; only in really hard cases do we need to
68 * compute a second residual.
69 * 4. Because of 3., we don't need a large table of powers of 10
70 * for ten-to-e (just some small tables, e.g. of 10^k
75 * #define IEEE_8087 for IEEE-arithmetic machines where the least
76 * significant byte has the lowest address.
77 * #define IEEE_MC68k for IEEE-arithmetic machines where the most
78 * significant byte has the lowest address.
79 * #define Long int on machines with 32-bit ints and 64-bit longs.
80 * #define IBM for IBM mainframe-style floating-point arithmetic.
81 * #define VAX for VAX-style floating-point arithmetic (D_floating).
82 * #define No_leftright to omit left-right logic in fast floating-point
83 * computation of dtoa.
84 * #define Honor_FLT_ROUNDS if FLT_ROUNDS can assume the values 2 or 3
85 * and strtod and dtoa should round accordingly.
86 * #define Check_FLT_ROUNDS if FLT_ROUNDS can assume the values 2 or 3
87 * and Honor_FLT_ROUNDS is not #defined.
88 * #define RND_PRODQUOT to use rnd_prod and rnd_quot (assembly routines
89 * that use extended-precision instructions to compute rounded
90 * products and quotients) with IBM.
91 * #define ROUND_BIASED for IEEE-format with biased rounding.
92 * #define Inaccurate_Divide for IEEE-format with correctly rounded
93 * products but inaccurate quotients, e.g., for Intel i860.
94 * #define NO_LONG_LONG on machines that do not have a "long long"
95 * integer type (of >= 64 bits). On such machines, you can
96 * #define Just_16 to store 16 bits per 32-bit Long when doing
97 * high-precision integer arithmetic. Whether this speeds things
98 * up or slows things down depends on the machine and the number
99 * being converted. If long long is available and the name is
100 * something other than "long long", #define Llong to be the name,
101 * and if "unsigned Llong" does not work as an unsigned version of
102 * Llong, #define #ULLong to be the corresponding unsigned type.
103 * #define KR_headers for old-style C function headers.
104 * #define Bad_float_h if your system lacks a float.h or if it does not
105 * define some or all of DBL_DIG, DBL_MAX_10_EXP, DBL_MAX_EXP,
106 * FLT_RADIX, FLT_ROUNDS, and DBL_MAX.
107 * #define MALLOC your_malloc, where your_malloc(n) acts like malloc(n)
108 * if memory is available and otherwise does something you deem
109 * appropriate. If MALLOC is undefined, malloc will be invoked
110 * directly -- and assumed always to succeed.
111 * #define Omit_Private_Memory to omit logic (added Jan. 1998) for making
112 * memory allocations from a private pool of memory when possible.
113 * When used, the private pool is PRIVATE_MEM bytes long: 2304 bytes,
114 * unless #defined to be a different length. This default length
115 * suffices to get rid of MALLOC calls except for unusual cases,
116 * such as decimal-to-binary conversion of a very long string of
117 * digits. The longest string dtoa can return is about 751 bytes
118 * long. For conversions by strtod of strings of 800 digits and
119 * all dtoa conversions in single-threaded executions with 8-byte
120 * pointers, PRIVATE_MEM >= 7400 appears to suffice; with 4-byte
121 * pointers, PRIVATE_MEM >= 7112 appears adequate.
122 * #define INFNAN_CHECK on IEEE systems to cause strtod to check for
123 * Infinity and NaN (case insensitively). On some systems (e.g.,
124 * some HP systems), it may be necessary to #define NAN_WORD0
125 * appropriately -- to the most significant word of a quiet NaN.
126 * (On HP Series 700/800 machines, -DNAN_WORD0=0x7ff40000 works.)
127 * When INFNAN_CHECK is #defined and No_Hex_NaN is not #defined,
128 * strtod also accepts (case insensitively) strings of the form
129 * NaN(x), where x is a string of hexadecimal digits and spaces;
130 * if there is only one string of hexadecimal digits, it is taken
131 * for the 52 fraction bits of the resulting NaN; if there are two
132 * or more strings of hex digits, the first is for the high 20 bits,
133 * the second and subsequent for the low 32 bits, with intervening
134 * white space ignored; but if this results in none of the 52
135 * fraction bits being on (an IEEE Infinity symbol), then NAN_WORD0
136 * and NAN_WORD1 are used instead.
137 * #define MULTIPLE_THREADS if the system offers preemptively scheduled
138 * multiple threads. In this case, you must provide (or suitably
139 * #define) two locks, acquired by ACQUIRE_DTOA_LOCK(n) and freed
140 * by FREE_DTOA_LOCK(n) for n = 0 or 1. (The second lock, accessed
141 * in pow5mult, ensures lazy evaluation of only one copy of high
142 * powers of 5; omitting this lock would introduce a small
143 * probability of wasting memory, but would otherwise be harmless.)
144 * You must also invoke freedtoa(s) to free the value s returned by
145 * dtoa. You may do so whether or not MULTIPLE_THREADS is #defined.
146 * #define NO_IEEE_Scale to disable new (Feb. 1997) logic in strtod that
147 * avoids underflows on inputs whose result does not underflow.
148 * If you #define NO_IEEE_Scale on a machine that uses IEEE-format
149 * floating-point numbers and flushes underflows to zero rather
150 * than implementing gradual underflow, then you must also #define
152 * #define YES_ALIAS to permit aliasing certain double values with
153 * arrays of ULongs. This leads to slightly better code with
154 * some compilers and was always used prior to 19990916, but it
155 * is not strictly legal and can cause trouble with aggressively
156 * optimizing compilers (e.g., gcc 2.95.1 under -O2).
157 * #define USE_LOCALE to use the current locale's decimal_point value.
158 * #define SET_INEXACT if IEEE arithmetic is being used and extra
159 * computation should be done to set the inexact flag when the
160 * result is inexact and avoid setting inexact when the result
161 * is exact. In this case, dtoa.c must be compiled in
162 * an environment, perhaps provided by #include "dtoa.c" in a
163 * suitable wrapper, that defines two functions,
164 * int get_inexact(void);
165 * void clear_inexact(void);
166 * such that get_inexact() returns a nonzero value if the
167 * inexact bit is already set, and clear_inexact() sets the
168 * inexact bit to 0. When SET_INEXACT is #defined, strtod
169 * also does extra computations to set the underflow and overflow
170 * flags when appropriate (i.e., when the result is tiny and
171 * inexact or when it is a numeric value rounded to +-infinity).
172 * #define NO_ERRNO if strtod should not assign errno = ERANGE when
173 * the result overflows to +-Infinity or underflows to 0.
175 #if defined(i386) || defined(mips) && defined(MIPSEL) || defined (__arm__)
179 #elif defined(__x86_64__) || defined(__alpha__)
183 #elif defined(__ia64)
191 #elif defined(__hppa)
200 #define ULong guint32
204 #define Bug(x) {fprintf(stderr, "%s\n", x); exit(1);}
217 extern char *MALLOC();
219 extern void *MALLOC(size_t);
222 #define MALLOC malloc
225 #define Omit_Private_Memory
226 #ifndef Omit_Private_Memory
228 #define PRIVATE_MEM 2304
230 #define PRIVATE_mem ((PRIVATE_MEM+sizeof(double)-1)/sizeof(double))
231 static double private_mem
[PRIVATE_mem
], *pmem_next
= private_mem
;
235 #undef Avoid_Underflow
249 #define DBL_MAX_10_EXP 308
250 #define DBL_MAX_EXP 1024
252 #endif /*IEEE_Arith*/
256 #define DBL_MAX_10_EXP 75
257 #define DBL_MAX_EXP 63
259 #define DBL_MAX 7.2370055773322621e+75
264 #define DBL_MAX_10_EXP 38
265 #define DBL_MAX_EXP 127
267 #define DBL_MAX 1.7014118346046923e+38
271 #define LONG_MAX 2147483647
274 #else /* ifndef Bad_float_h */
276 #endif /* Bad_float_h */
288 #define CONST /* blank */
294 #if defined(IEEE_8087) + defined(IEEE_MC68k) + defined(VAX) + defined(IBM) != 1
295 Exactly one of IEEE_8087
, IEEE_MC68k
, VAX
, or IBM should be defined
.
298 typedef union { double d
; ULong L
[2]; } U
;
303 #define word0(x) ((ULong *)&x)[1]
304 #define word1(x) ((ULong *)&x)[0]
306 #define word0(x) ((ULong *)&x)[0]
307 #define word1(x) ((ULong *)&x)[1]
311 #define word0(x) ((U*)&x)->L[1]
312 #define word1(x) ((U*)&x)->L[0]
314 #define word0(x) ((U*)&x)->L[0]
315 #define word1(x) ((U*)&x)->L[1]
317 #define dval(x) ((U*)&x)->d
320 /* The following definition of Storeinc is appropriate for MIPS processors.
321 * An alternative that might be better on some machines is
322 * #define Storeinc(a,b,c) (*a++ = b << 16 | c & 0xffff)
324 #if defined(IEEE_8087) + defined(VAX)
325 #define Storeinc(a,b,c) do { (((unsigned short *)a)[1] = (unsigned short)b, \
326 ((unsigned short *)a)[0] = (unsigned short)c, a++) } while (0)
328 #define Storeinc(a,b,c) do { (((unsigned short *)a)[0] = (unsigned short)b, \
329 ((unsigned short *)a)[1] = (unsigned short)c, a++) } while (0)
332 /* #define P DBL_MANT_DIG */
333 /* Ten_pmax = floor(P*log(2)/log(5)) */
334 /* Bletch = (highest power of 2 < DBL_MAX_10_EXP) / 16 */
335 /* Quick_max = floor((P-1)*log(FLT_RADIX)/log(10) - 1) */
336 /* Int_max = floor(P*log(FLT_RADIX)/log(10) - 1) */
340 #define Exp_shift1 20
341 #define Exp_msk1 0x100000
342 #define Exp_msk11 0x100000
343 #define Exp_mask 0x7ff00000
347 #define Exp_1 0x3ff00000
348 #define Exp_11 0x3ff00000
350 #define Frac_mask 0xfffff
351 #define Frac_mask1 0xfffff
354 #define Bndry_mask 0xfffff
355 #define Bndry_mask1 0xfffff
357 #define Sign_bit 0x80000000
363 #ifndef NO_IEEE_Scale
364 #define Avoid_Underflow
365 #ifdef Flush_Denorm /* debugging option */
366 #undef Sudden_Underflow
372 #define Flt_Rounds FLT_ROUNDS
376 #endif /*Flt_Rounds*/
378 #ifdef Honor_FLT_ROUNDS
379 #define Rounding rounding
380 #undef Check_FLT_ROUNDS
381 #define Check_FLT_ROUNDS
383 #define Rounding Flt_Rounds
386 #else /* ifndef IEEE_Arith */
387 #undef Check_FLT_ROUNDS
388 #undef Honor_FLT_ROUNDS
390 #undef Sudden_Underflow
391 #define Sudden_Underflow
396 #define Exp_shift1 24
397 #define Exp_msk1 0x1000000
398 #define Exp_msk11 0x1000000
399 #define Exp_mask 0x7f000000
402 #define Exp_1 0x41000000
403 #define Exp_11 0x41000000
404 #define Ebits 8 /* exponent has 7 bits, but 8 is the right value in b2d */
405 #define Frac_mask 0xffffff
406 #define Frac_mask1 0xffffff
409 #define Bndry_mask 0xefffff
410 #define Bndry_mask1 0xffffff
412 #define Sign_bit 0x80000000
414 #define Tiny0 0x100000
423 #define Exp_msk1 0x80
424 #define Exp_msk11 0x800000
425 #define Exp_mask 0x7f80
428 #define Exp_1 0x40800000
429 #define Exp_11 0x4080
431 #define Frac_mask 0x7fffff
432 #define Frac_mask1 0xffff007f
435 #define Bndry_mask 0xffff007f
436 #define Bndry_mask1 0xffff007f
438 #define Sign_bit 0x8000
444 #endif /* IBM, VAX */
445 #endif /* IEEE_Arith */
452 #define rounded_product(a,b) a = rnd_prod(a, b)
453 #define rounded_quotient(a,b) a = rnd_quot(a, b)
455 extern double rnd_prod(), rnd_quot();
457 extern double rnd_prod(double, double), rnd_quot(double, double);
460 #define rounded_product(a,b) a *= b
461 #define rounded_quotient(a,b) a /= b
464 #define Big0 (Frac_mask1 | Exp_msk1*(DBL_MAX_EXP+Bias-1))
465 #define Big1 0xffffffff
472 #define FFFFFFFF ((((unsigned long)0xffff)<<16)|(unsigned long)0xffff)
474 #define FFFFFFFF 0xffffffffUL
481 /* When Pack_32 is not defined, we store 16 bits per 32-bit Long.
482 * This makes some inner loops simpler and sometimes saves work
483 * during multiplications, but it often seems to make things slightly
484 * slower. Hence the default is now to store 32 bits per Long.
487 #else /* long long available */
489 #define Llong long long
492 #define ULLong unsigned Llong
494 #endif /* NO_LONG_LONG */
496 #ifndef MULTIPLE_THREADS
497 #define ACQUIRE_DTOA_LOCK(n) /*nothing*/
498 #define FREE_DTOA_LOCK(n) /*nothing*/
504 extern "C" double strtod(const char *s00
, char **se
);
505 extern "C" char *dtoa(double d
, int mode
, int ndigits
,
506 int *decpt
, int *sign
, char **rve
);
512 int k
, maxwds
, sign
, wds
;
516 typedef struct Bigint Bigint
;
518 static Bigint
*freelist
[Kmax
+1];
530 #ifndef Omit_Private_Memory
534 ACQUIRE_DTOA_LOCK(0);
535 if ((rv
= freelist
[k
])) {
536 freelist
[k
] = rv
->next
;
540 #ifdef Omit_Private_Memory
541 rv
= (Bigint
*)MALLOC(sizeof(Bigint
) + (x
-1)*sizeof(ULong
));
543 len
= (sizeof(Bigint
) + (x
-1)*sizeof(ULong
) + sizeof(double) - 1)
545 if (pmem_next
- private_mem
+ len
<= PRIVATE_mem
) {
546 rv
= (Bigint
*)pmem_next
;
550 rv
= (Bigint
*)MALLOC(len
*sizeof(double));
556 rv
->sign
= rv
->wds
= 0;
569 ACQUIRE_DTOA_LOCK(0);
570 v
->next
= freelist
[v
->k
];
576 #define Bcopy(x,y) memcpy((char *)&x->sign, (char *)&y->sign, \
577 y->wds*sizeof(Long) + 2*sizeof(int))
582 (b
, m
, a
) Bigint
*b
; int m
, a
;
584 (Bigint
*b
, int m
, int a
) /* multiply by m and add a */
605 y
= *x
* (ULLong
)m
+ carry
;
611 y
= (xi
& 0xffff) * m
+ carry
;
612 z
= (xi
>> 16) * m
+ (y
>> 16);
614 *x
++ = (z
<< 16) + (y
& 0xffff);
624 if (wds
>= b
->maxwds
) {
639 (s
, nd0
, nd
, y9
) CONST
char *s
; int nd0
, nd
; ULong y9
;
641 (CONST
char *s
, int nd0
, int nd
, ULong y9
)
649 for(k
= 0, y
= 1; x
> y
; y
<<= 1, k
++) ;
656 b
->x
[0] = y9
& 0xffff;
657 b
->wds
= (b
->x
[1] = y9
>> 16) ? 2 : 1;
663 do b
= multadd(b
, 10, *s
++ - '0');
670 b
= multadd(b
, 10, *s
++ - '0');
677 (x
) register ULong x
;
684 if (!(x
& 0xffff0000)) {
688 if (!(x
& 0xff000000)) {
692 if (!(x
& 0xf0000000)) {
696 if (!(x
& 0xc0000000)) {
700 if (!(x
& 0x80000000)) {
702 if (!(x
& 0x40000000))
717 register ULong x
= *y
;
775 (a
, b
) Bigint
*a
, *b
;
777 (Bigint
*a
, Bigint
*b
)
782 ULong
*x
, *xa
, *xae
, *xb
, *xbe
, *xc
, *xc0
;
793 if (a
->wds
< b
->wds
) {
805 for(x
= c
->x
, xa
= x
+ wc
; x
< xa
; x
++)
813 for(; xb
< xbe
; xc0
++) {
819 z
= *x
++ * (ULLong
)y
+ *xc
+ carry
;
821 *xc
++ = z
& FFFFFFFF
;
829 for(; xb
< xbe
; xb
++, xc0
++) {
830 if (y
= *xb
& 0xffff) {
835 z
= (*x
& 0xffff) * y
+ (*xc
& 0xffff) + carry
;
837 z2
= (*x
++ >> 16) * y
+ (*xc
>> 16) + carry
;
850 z
= (*x
& 0xffff) * y
+ (*xc
>> 16) + carry
;
853 z2
= (*x
++ >> 16) * y
+ (*xc
& 0xffff) + carry
;
861 for(; xb
< xbe
; xc0
++) {
867 z
= *x
++ * y
+ *xc
+ carry
;
877 for(xc0
= c
->x
, xc
= xc0
+ wc
; wc
> 0 && !*--xc
; --wc
) ;
887 (b
, k
) Bigint
*b
; int k
;
892 Bigint
*b1
, *p5
, *p51
;
894 static int p05
[3] = { 5, 25, 125 };
897 b
= multadd(b
, p05
[i
-1], 0);
903 #ifdef MULTIPLE_THREADS
904 ACQUIRE_DTOA_LOCK(1);
923 if (!(p51
= p5
->next
)) {
924 #ifdef MULTIPLE_THREADS
925 ACQUIRE_DTOA_LOCK(1);
926 if (!(p51
= p5
->next
)) {
927 p51
= p5
->next
= mult(p5
,p5
);
932 p51
= p5
->next
= mult(p5
,p5
);
944 (b
, k
) Bigint
*b
; int k
;
951 ULong
*x
, *x1
, *xe
, z
;
960 for(i
= b
->maxwds
; n1
> i
; i
<<= 1)
964 for(i
= 0; i
< n
; i
++)
985 *x1
++ = *x
<< k
& 0xffff | z
;
1004 (a
, b
) Bigint
*a
, *b
;
1006 (Bigint
*a
, Bigint
*b
)
1009 ULong
*xa
, *xa0
, *xb
, *xb0
;
1015 if (i
> 1 && !a
->x
[i
-1])
1016 Bug("cmp called with a->x[a->wds-1] == 0");
1017 if (j
> 1 && !b
->x
[j
-1])
1018 Bug("cmp called with b->x[b->wds-1] == 0");
1028 return *xa
< *xb
? -1 : 1;
1038 (a
, b
) Bigint
*a
, *b
;
1040 (Bigint
*a
, Bigint
*b
)
1045 ULong
*xa
, *xae
, *xb
, *xbe
, *xc
;
1082 y
= (ULLong
)*xa
++ - *xb
++ - borrow
;
1083 borrow
= y
>> 32 & (ULong
)1;
1084 *xc
++ = y
& FFFFFFFF
;
1089 borrow
= y
>> 32 & (ULong
)1;
1090 *xc
++ = y
& FFFFFFFF
;
1095 y
= (*xa
& 0xffff) - (*xb
& 0xffff) - borrow
;
1096 borrow
= (y
& 0x10000) >> 16;
1097 z
= (*xa
++ >> 16) - (*xb
++ >> 16) - borrow
;
1098 borrow
= (z
& 0x10000) >> 16;
1103 y
= (*xa
& 0xffff) - borrow
;
1104 borrow
= (y
& 0x10000) >> 16;
1105 z
= (*xa
++ >> 16) - borrow
;
1106 borrow
= (z
& 0x10000) >> 16;
1111 y
= *xa
++ - *xb
++ - borrow
;
1112 borrow
= (y
& 0x10000) >> 16;
1118 borrow
= (y
& 0x10000) >> 16;
1140 L
= (word0(x
) & Exp_mask
) - (P
-1)*Exp_msk1
;
1141 #ifndef Avoid_Underflow
1142 #ifndef Sudden_Underflow
1151 #ifndef Avoid_Underflow
1152 #ifndef Sudden_Underflow
1155 L
= -L
>> Exp_shift
;
1156 if (L
< Exp_shift
) {
1157 word0(a
) = 0x80000 >> L
;
1163 word1(a
) = L
>= 31 ? 1 : 1 << 31 - L
;
1174 (a
, e
) Bigint
*a
; int *e
;
1179 ULong
*xa
, *xa0
, w
, y
, z
;
1193 if (!y
) Bug("zero y in b2d");
1199 d0
= Exp_1
| (y
>> (Ebits
- k
));
1200 w
= xa
> xa0
? *--xa
: 0;
1201 d1
= y
<< ((32-Ebits
) + k
) | (w
>> (Ebits
- k
));
1204 z
= xa
> xa0
? *--xa
: 0;
1206 d0
= Exp_1
| y
<< k
| (z
>> (32 - k
));
1207 y
= xa
> xa0
? *--xa
: 0;
1208 d1
= z
<< k
| (y
>> (32 - k
));
1215 if (k
< Ebits
+ 16) {
1216 z
= xa
> xa0
? *--xa
: 0;
1217 d0
= Exp_1
| y
<< k
- Ebits
| z
>> Ebits
+ 16 - k
;
1218 w
= xa
> xa0
? *--xa
: 0;
1219 y
= xa
> xa0
? *--xa
: 0;
1220 d1
= z
<< k
+ 16 - Ebits
| w
<< k
- Ebits
| y
>> 16 + Ebits
- k
;
1223 z
= xa
> xa0
? *--xa
: 0;
1224 w
= xa
> xa0
? *--xa
: 0;
1226 d0
= Exp_1
| y
<< k
+ 16 | z
<< k
| w
>> 16 - k
;
1227 y
= xa
> xa0
? *--xa
: 0;
1228 d1
= w
<< k
+ 16 | y
<< k
;
1232 word0(d
) = d0
>> 16 | d0
<< 16;
1233 word1(d
) = d1
>> 16 | d1
<< 16;
1244 (d
, e
, bits
) double d
; int *e
, *bits
;
1246 (double d
, int *e
, int *bits
)
1252 #ifndef Sudden_Underflow
1257 d0
= word0(d
) >> 16 | word0(d
) << 16;
1258 d1
= word1(d
) >> 16 | word1(d
) << 16;
1272 d0
&= 0x7fffffff; /* clear sign bit, which we ignore */
1273 #ifdef Sudden_Underflow
1274 de
= (int)(d0
>> Exp_shift
);
1279 if ((de
= (int)(d0
>> Exp_shift
)))
1284 if ((k
= lo0bits(&y
))) {
1285 x
[0] = y
| (z
<< (32 - k
));
1290 #ifndef Sudden_Underflow
1293 b
->wds
= (x
[1] = z
) ? 2 : 1;
1298 Bug("Zero passed to d2b");
1302 #ifndef Sudden_Underflow
1310 if (k
= lo0bits(&y
))
1312 x
[0] = y
| z
<< 32 - k
& 0xffff;
1313 x
[1] = z
>> k
- 16 & 0xffff;
1319 x
[1] = y
>> 16 | z
<< 16 - k
& 0xffff;
1320 x
[2] = z
>> k
& 0xffff;
1335 Bug("Zero passed to d2b");
1353 #ifndef Sudden_Underflow
1357 *e
= (de
- Bias
- (P
-1) << 2) + k
;
1358 *bits
= 4*P
+ 8 - k
- hi0bits(word0(d
) & Frac_mask
);
1360 *e
= de
- Bias
- (P
-1) + k
;
1363 #ifndef Sudden_Underflow
1366 *e
= de
- Bias
- (P
-1) + 1 + k
;
1368 *bits
= 32*i
- hi0bits(x
[i
-1]);
1370 *bits
= (i
+2)*16 - hi0bits(x
[i
]);
1382 (a
, b
) Bigint
*a
, *b
;
1384 (Bigint
*a
, Bigint
*b
)
1390 dval(da
) = b2d(a
, &ka
);
1391 dval(db
) = b2d(b
, &kb
);
1393 k
= ka
- kb
+ 32*(a
->wds
- b
->wds
);
1395 k
= ka
- kb
+ 16*(a
->wds
- b
->wds
);
1399 word0(da
) += (k
>> 2)*Exp_msk1
;
1405 word0(db
) += (k
>> 2)*Exp_msk1
;
1411 word0(da
) += k
*Exp_msk1
;
1414 word0(db
) += k
*Exp_msk1
;
1417 return dval(da
) / dval(db
);
1422 1e0
, 1e1
, 1e2
, 1e3
, 1e4
, 1e5
, 1e6
, 1e7
, 1e8
, 1e9
,
1423 1e10
, 1e11
, 1e12
, 1e13
, 1e14
, 1e15
, 1e16
, 1e17
, 1e18
, 1e19
,
1432 bigtens
[] = { 1e16
, 1e32
, 1e64
, 1e128
, 1e256
};
1433 static CONST
double tinytens
[] = { 1e-16, 1e-32, 1e-64, 1e-128,
1434 #ifdef Avoid_Underflow
1435 9007199254740992.*9007199254740992.e
-256
1436 /* = 2^106 * 1e-53 */
1441 /* The factor of 2^53 in tinytens[4] helps us avoid setting the underflow */
1442 /* flag unnecessarily. It leads to a song and dance at the end of strtod. */
1443 #define Scale_Bit 0x10
1447 bigtens
[] = { 1e16
, 1e32
, 1e64
};
1448 static CONST
double tinytens
[] = { 1e-16, 1e-32, 1e-64 };
1451 bigtens
[] = { 1e16
, 1e32
};
1452 static CONST
double tinytens
[] = { 1e-16, 1e-32 };
1464 #define NAN_WORD0 0x7ff80000
1474 (sp
, t
) char **sp
, *t
;
1476 (CONST
char **sp
, char *t
)
1480 CONST
char *s
= *sp
;
1483 if ((c
= *++s
) >= 'A' && c
<= 'Z')
1496 (rvp
, sp
) double *rvp
; CONST
char **sp
;
1498 (double *rvp
, CONST
char **sp
)
1503 int havedig
, udx0
, xshift
;
1506 havedig
= xshift
= 0;
1509 while(c
= *(CONST
unsigned char*)++s
) {
1510 if (c
>= '0' && c
<= '9')
1512 else if (c
>= 'a' && c
<= 'f')
1514 else if (c
>= 'A' && c
<= 'F')
1516 else if (c
<= ' ') {
1517 if (udx0
&& havedig
) {
1523 else if (/*(*/ c
== ')' && havedig
) {
1528 return; /* invalid form: don't change *sp */
1536 x
[0] = (x
[0] << 4) | (x
[1] >> 28);
1537 x
[1] = (x
[1] << 4) | c
;
1539 if ((x
[0] &= 0xfffff) || x
[1]) {
1540 word0(*rvp
) = Exp_mask
| x
[0];
1544 #endif /*No_Hex_NaN*/
1545 #endif /* INFNAN_CHECK */
1550 (s00
, se
) CONST
char *s00
; char **se
;
1552 (CONST
char *s00
, char **se
)
1555 #ifdef Avoid_Underflow
1558 int bb2
, bb5
, bbe
, bd2
, bd5
, bbbits
, bs2
, c
, dsign
,
1559 e
, e1
, esign
, i
, j
, k
, nd
, nd0
, nf
, nz
, nz0
, sign
;
1560 CONST
char *s
, *s0
, *s1
;
1561 double aadj
, aadj1
, adj
, rv
, rv0
;
1564 Bigint
*bb
, *bb1
, *bd
, *bd0
, *bs
, *delta
;
1566 int inexact
, oldinexact
;
1568 #ifdef Honor_FLT_ROUNDS
1575 sign
= nz0
= nz
= 0;
1577 for(s
= s00
;;s
++) switch(*s
) {
1600 while(*++s
== '0') ;
1606 for(nd
= nf
= 0; (c
= *s
) >= '0' && c
<= '9'; nd
++, s
++)
1613 s1
= localeconv()->decimal_point
;
1634 for(; c
== '0'; c
= *++s
)
1636 if (c
> '0' && c
<= '9') {
1644 for(; c
>= '0' && c
<= '9'; c
= *++s
) {
1649 for(i
= 1; i
< nz
; i
++)
1652 else if (nd
<= DBL_DIG
+ 1)
1656 else if (nd
<= DBL_DIG
+ 1)
1664 if (c
== 'e' || c
== 'E') {
1665 if (!nd
&& !nz
&& !nz0
) {
1676 if (c
>= '0' && c
<= '9') {
1679 if (c
> '0' && c
<= '9') {
1682 while((c
= *++s
) >= '0' && c
<= '9')
1684 if (s
- s1
> 8 || L
> 19999)
1685 /* Avoid confusion from exponents
1686 * so large that e might overflow.
1688 e
= 19999; /* safe for 16 bit ints */
1703 /* Check for Nan and Infinity */
1707 if (match(&s
,"nf")) {
1709 if (!match(&s
,"inity"))
1711 word0(rv
) = 0x7ff00000;
1718 if (match(&s
, "an")) {
1719 word0(rv
) = NAN_WORD0
;
1720 word1(rv
) = NAN_WORD1
;
1722 if (*s
== '(') /*)*/
1728 #endif /* INFNAN_CHECK */
1737 /* Now we have nd0 digits, starting at s0, followed by a
1738 * decimal point, followed by nd-nd0 digits. The number we're
1739 * after is the integer represented by those digits times
1744 k
= nd
< DBL_DIG
+ 1 ? nd
: DBL_DIG
+ 1;
1749 oldinexact
= get_inexact();
1751 dval(rv
) = tens
[k
- 9] * dval(rv
) + z
;
1755 #ifndef RND_PRODQUOT
1756 #ifndef Honor_FLT_ROUNDS
1764 if (e
<= Ten_pmax
) {
1766 goto vax_ovfl_check
;
1768 #ifdef Honor_FLT_ROUNDS
1769 /* round correctly FLT_ROUNDS = 2 or 3 */
1775 /* rv = */ rounded_product(dval(rv
), tens
[e
]);
1780 if (e
<= Ten_pmax
+ i
) {
1781 /* A fancier test would sometimes let us do
1782 * this for larger i values.
1784 #ifdef Honor_FLT_ROUNDS
1785 /* round correctly FLT_ROUNDS = 2 or 3 */
1792 dval(rv
) *= tens
[i
];
1794 /* VAX exponent range is so narrow we must
1795 * worry about overflow here...
1798 word0(rv
) -= P
*Exp_msk1
;
1799 /* rv = */ rounded_product(dval(rv
), tens
[e
]);
1800 if ((word0(rv
) & Exp_mask
)
1801 > Exp_msk1
*(DBL_MAX_EXP
+Bias
-1-P
))
1803 word0(rv
) += P
*Exp_msk1
;
1805 /* rv = */ rounded_product(dval(rv
), tens
[e
]);
1810 #ifndef Inaccurate_Divide
1811 else if (e
>= -Ten_pmax
) {
1812 #ifdef Honor_FLT_ROUNDS
1813 /* round correctly FLT_ROUNDS = 2 or 3 */
1819 /* rv = */ rounded_quotient(dval(rv
), tens
[-e
]);
1830 oldinexact
= get_inexact();
1832 #ifdef Avoid_Underflow
1835 #ifdef Honor_FLT_ROUNDS
1836 if ((rounding
= Flt_Rounds
) >= 2) {
1838 rounding
= rounding
== 2 ? 0 : 2;
1844 #endif /*IEEE_Arith*/
1846 /* Get starting approximation = rv * 10**e1 */
1850 dval(rv
) *= tens
[i
];
1852 if (e1
> DBL_MAX_10_EXP
) {
1857 /* Can't trust HUGE_VAL */
1859 #ifdef Honor_FLT_ROUNDS
1861 case 0: /* toward 0 */
1862 case 3: /* toward -infinity */
1867 word0(rv
) = Exp_mask
;
1870 #else /*Honor_FLT_ROUNDS*/
1871 word0(rv
) = Exp_mask
;
1873 #endif /*Honor_FLT_ROUNDS*/
1875 /* set overflow bit */
1877 dval(rv0
) *= dval(rv0
);
1879 #else /*IEEE_Arith*/
1882 #endif /*IEEE_Arith*/
1888 for(j
= 0; e1
> 1; j
++, e1
>>= 1)
1890 dval(rv
) *= bigtens
[j
];
1891 /* The last multiplication could overflow. */
1892 word0(rv
) -= P
*Exp_msk1
;
1893 dval(rv
) *= bigtens
[j
];
1894 if ((z
= word0(rv
) & Exp_mask
)
1895 > Exp_msk1
*(DBL_MAX_EXP
+Bias
-P
))
1897 if (z
> Exp_msk1
*(DBL_MAX_EXP
+Bias
-1-P
)) {
1898 /* set to largest number */
1899 /* (Can't trust DBL_MAX) */
1904 word0(rv
) += P
*Exp_msk1
;
1910 dval(rv
) /= tens
[i
];
1912 if (e1
>= 1 << n_bigtens
)
1914 #ifdef Avoid_Underflow
1917 for(j
= 0; e1
> 0; j
++, e1
>>= 1)
1919 dval(rv
) *= tinytens
[j
];
1920 if (scale
&& (j
= 2*P
+ 1 - ((word0(rv
) & Exp_mask
)
1921 >> Exp_shift
)) > 0) {
1922 /* scaled rv is denormal; zap j low bits */
1926 word0(rv
) = (P
+2)*Exp_msk1
;
1928 word0(rv
) &= 0xffffffff << (j
-32);
1931 word1(rv
) &= 0xffffffff << j
;
1934 for(j
= 0; e1
> 1; j
++, e1
>>= 1)
1936 dval(rv
) *= tinytens
[j
];
1937 /* The last multiplication could underflow. */
1938 dval(rv0
) = dval(rv
);
1939 dval(rv
) *= tinytens
[j
];
1941 dval(rv
) = 2.*dval(rv0
);
1942 dval(rv
) *= tinytens
[j
];
1954 #ifndef Avoid_Underflow
1957 /* The refinement below will clean
1958 * this approximation up.
1965 /* Now the hard part -- adjusting rv to the correct value.*/
1967 /* Put digits into bd: true value = bd * 10^e */
1969 bd0
= s2b(s0
, nd0
, nd
, y
);
1972 bd
= Balloc(bd0
->k
);
1974 bb
= d2b(dval(rv
), &bbe
, &bbbits
); /* rv = bb * 2^bbe */
1990 #ifdef Honor_FLT_ROUNDS
1994 #ifdef Avoid_Underflow
1996 i
= j
+ bbbits
- 1; /* logb(rv) */
1997 if (i
< Emin
) /* denormal */
2001 #else /*Avoid_Underflow*/
2002 #ifdef Sudden_Underflow
2004 j
= 1 + 4*P
- 3 - bbbits
+ ((bbe
+ bbbits
- 1) & 3);
2008 #else /*Sudden_Underflow*/
2010 i
= j
+ bbbits
- 1; /* logb(rv) */
2011 if (i
< Emin
) /* denormal */
2015 #endif /*Sudden_Underflow*/
2016 #endif /*Avoid_Underflow*/
2019 #ifdef Avoid_Underflow
2022 i
= bb2
< bd2
? bb2
: bd2
;
2031 bs
= pow5mult(bs
, bb5
);
2037 bb
= lshift(bb
, bb2
);
2039 bd
= pow5mult(bd
, bd5
);
2041 bd
= lshift(bd
, bd2
);
2043 bs
= lshift(bs
, bs2
);
2044 delta
= diff(bb
, bd
);
2045 dsign
= delta
->sign
;
2048 #ifdef Honor_FLT_ROUNDS
2049 if (rounding
!= 1) {
2051 /* Error is less than an ulp */
2052 if (!delta
->x
[0] && delta
->wds
<= 1) {
2068 && !(word0(rv
) & Frac_mask
)) {
2069 y
= word0(rv
) & Exp_mask
;
2070 #ifdef Avoid_Underflow
2071 if (!scale
|| y
> 2*P
*Exp_msk1
)
2076 delta
= lshift(delta
,Log2P
);
2077 if (cmp(delta
, bs
) <= 0)
2082 #ifdef Avoid_Underflow
2083 if (scale
&& (y
= word0(rv
) & Exp_mask
)
2085 word0(adj
) += (2*P
+1)*Exp_msk1
- y
;
2087 #ifdef Sudden_Underflow
2088 if ((word0(rv
) & Exp_mask
) <=
2090 word0(rv
) += P
*Exp_msk1
;
2091 dval(rv
) += adj
*ulp(dval(rv
));
2092 word0(rv
) -= P
*Exp_msk1
;
2095 #endif /*Sudden_Underflow*/
2096 #endif /*Avoid_Underflow*/
2097 dval(rv
) += adj
*ulp(dval(rv
));
2101 adj
= ratio(delta
, bs
);
2104 if (adj
<= 0x7ffffffe) {
2105 /* adj = rounding ? ceil(adj) : floor(adj); */
2108 if (!((rounding
>>1) ^ dsign
))
2113 #ifdef Avoid_Underflow
2114 if (scale
&& (y
= word0(rv
) & Exp_mask
) <= 2*P
*Exp_msk1
)
2115 word0(adj
) += (2*P
+1)*Exp_msk1
- y
;
2117 #ifdef Sudden_Underflow
2118 if ((word0(rv
) & Exp_mask
) <= P
*Exp_msk1
) {
2119 word0(rv
) += P
*Exp_msk1
;
2120 adj
*= ulp(dval(rv
));
2125 word0(rv
) -= P
*Exp_msk1
;
2128 #endif /*Sudden_Underflow*/
2129 #endif /*Avoid_Underflow*/
2130 adj
*= ulp(dval(rv
));
2137 #endif /*Honor_FLT_ROUNDS*/
2140 /* Error is less than half an ulp -- check for
2141 * special case of mantissa a power of two.
2143 if (dsign
|| word1(rv
) || word0(rv
) & Bndry_mask
2145 #ifdef Avoid_Underflow
2146 || (word0(rv
) & Exp_mask
) <= (2*P
+1)*Exp_msk1
2148 || (word0(rv
) & Exp_mask
) <= Exp_msk1
2153 if (!delta
->x
[0] && delta
->wds
<= 1)
2158 if (!delta
->x
[0] && delta
->wds
<= 1) {
2165 delta
= lshift(delta
,Log2P
);
2166 if (cmp(delta
, bs
) > 0)
2171 /* exactly half-way between */
2173 if ((word0(rv
) & Bndry_mask1
) == Bndry_mask1
2175 #ifdef Avoid_Underflow
2176 (scale
&& (y
= word0(rv
) & Exp_mask
) <= 2*P
*Exp_msk1
)
2177 ? (0xffffffff & (0xffffffff << (2*P
+1-(y
>>Exp_shift
)))) :
2180 /*boundary case -- increment exponent*/
2181 word0(rv
) = (word0(rv
) & Exp_mask
)
2188 #ifdef Avoid_Underflow
2194 else if (!(word0(rv
) & Bndry_mask
) && !word1(rv
)) {
2196 /* boundary case -- decrement exponent */
2197 #ifdef Sudden_Underflow /*{{*/
2198 L
= word0(rv
) & Exp_mask
;
2202 #ifdef Avoid_Underflow
2203 if (L
<= (scale
? (2*P
+1)*Exp_msk1
: Exp_msk1
))
2206 #endif /*Avoid_Underflow*/
2210 #else /*Sudden_Underflow}{*/
2211 #ifdef Avoid_Underflow
2213 L
= word0(rv
) & Exp_mask
;
2214 if (L
<= (2*P
+1)*Exp_msk1
) {
2215 if (L
> (P
+2)*Exp_msk1
)
2216 /* round even ==> */
2219 /* rv = smallest denormal */
2223 #endif /*Avoid_Underflow*/
2224 L
= (word0(rv
) & Exp_mask
) - Exp_msk1
;
2225 #endif /*Sudden_Underflow}}*/
2226 word0(rv
) = L
| Bndry_mask1
;
2227 word1(rv
) = 0xffffffff;
2234 #ifndef ROUND_BIASED
2235 if (!(word1(rv
) & LSB
))
2239 dval(rv
) += ulp(dval(rv
));
2240 #ifndef ROUND_BIASED
2242 dval(rv
) -= ulp(dval(rv
));
2243 #ifndef Sudden_Underflow
2248 #ifdef Avoid_Underflow
2254 if ((aadj
= ratio(delta
, bs
)) <= 2.) {
2257 else if (word1(rv
) || word0(rv
) & Bndry_mask
) {
2258 #ifndef Sudden_Underflow
2259 if (word1(rv
) == Tiny1
&& !word0(rv
))
2266 /* special case -- power of FLT_RADIX to be */
2267 /* rounded down... */
2269 if (aadj
< 2./FLT_RADIX
)
2270 aadj
= 1./FLT_RADIX
;
2278 aadj1
= dsign
? aadj
: -aadj
;
2279 #ifdef Check_FLT_ROUNDS
2281 case 2: /* towards +infinity */
2284 case 0: /* towards 0 */
2285 case 3: /* towards -infinity */
2289 if (Flt_Rounds
== 0)
2291 #endif /*Check_FLT_ROUNDS*/
2293 y
= word0(rv
) & Exp_mask
;
2295 /* Check for overflow */
2297 if (y
== Exp_msk1
*(DBL_MAX_EXP
+Bias
-1)) {
2298 dval(rv0
) = dval(rv
);
2299 word0(rv
) -= P
*Exp_msk1
;
2300 adj
= aadj1
* ulp(dval(rv
));
2302 if ((word0(rv
) & Exp_mask
) >=
2303 Exp_msk1
*(DBL_MAX_EXP
+Bias
-P
)) {
2304 if (word0(rv0
) == Big0
&& word1(rv0
) == Big1
)
2311 word0(rv
) += P
*Exp_msk1
;
2314 #ifdef Avoid_Underflow
2315 if (scale
&& y
<= 2*P
*Exp_msk1
) {
2316 if (aadj
<= 0x7fffffff) {
2317 if ((z
= aadj
) <= 0)
2320 aadj1
= dsign
? aadj
: -aadj
;
2322 word0(aadj1
) += (2*P
+1)*Exp_msk1
- y
;
2324 adj
= aadj1
* ulp(dval(rv
));
2327 #ifdef Sudden_Underflow
2328 if ((word0(rv
) & Exp_mask
) <= P
*Exp_msk1
) {
2329 dval(rv0
) = dval(rv
);
2330 word0(rv
) += P
*Exp_msk1
;
2331 adj
= aadj1
* ulp(dval(rv
));
2334 if ((word0(rv
) & Exp_mask
) < P
*Exp_msk1
)
2336 if ((word0(rv
) & Exp_mask
) <= P
*Exp_msk1
)
2339 if (word0(rv0
) == Tiny0
2340 && word1(rv0
) == Tiny1
)
2347 word0(rv
) -= P
*Exp_msk1
;
2350 adj
= aadj1
* ulp(dval(rv
));
2353 #else /*Sudden_Underflow*/
2354 /* Compute adj so that the IEEE rounding rules will
2355 * correctly round rv + adj in some half-way cases.
2356 * If rv * ulp(rv) is denormalized (i.e.,
2357 * y <= (P-1)*Exp_msk1), we must adjust aadj to avoid
2358 * trouble from bits lost to denormalization;
2359 * example: 1.2e-307 .
2361 if (y
<= (P
-1)*Exp_msk1
&& aadj
> 1.) {
2362 aadj1
= (double)(int)(aadj
+ 0.5);
2366 adj
= aadj1
* ulp(dval(rv
));
2368 #endif /*Sudden_Underflow*/
2369 #endif /*Avoid_Underflow*/
2371 z
= word0(rv
) & Exp_mask
;
2373 #ifdef Avoid_Underflow
2377 /* Can we stop now? */
2380 /* The tolerances below are conservative. */
2381 if (dsign
|| word1(rv
) || word0(rv
) & Bndry_mask
) {
2382 if (aadj
< .4999999 || aadj
> .5000001)
2385 else if (aadj
< .4999999/FLT_RADIX
)
2398 word0(rv0
) = Exp_1
+ (70 << Exp_shift
);
2403 else if (!oldinexact
)
2406 #ifdef Avoid_Underflow
2408 word0(rv0
) = Exp_1
- 2*P
*Exp_msk1
;
2410 dval(rv
) *= dval(rv0
);
2412 /* try to avoid the bug of testing an 8087 register value */
2413 if (word0(rv
) == 0 && word1(rv
) == 0)
2417 #endif /* Avoid_Underflow */
2419 if (inexact
&& !(word0(rv
) & Exp_mask
)) {
2420 /* set underflow bit */
2422 dval(rv0
) *= dval(rv0
);
2434 return sign
? -dval(rv
) : dval(rv
);
2440 (b
, S
) Bigint
*b
, *S
;
2442 (Bigint
*b
, Bigint
*S
)
2446 ULong
*bx
, *bxe
, q
, *sx
, *sxe
;
2448 ULLong borrow
, carry
, y
, ys
;
2450 ULong borrow
, carry
, y
, ys
;
2458 /*debug*/ if (b
->wds
> n
)
2459 /*debug*/ Bug("oversize b in quorem");
2467 q
= *bxe
/ (*sxe
+ 1); /* ensure q <= true quotient */
2469 /*debug*/ if (q
> 9)
2470 /*debug*/ Bug("oversized quotient in quorem");
2477 ys
= *sx
++ * (ULLong
)q
+ carry
;
2479 y
= *bx
- (ys
& FFFFFFFF
) - borrow
;
2480 borrow
= y
>> 32 & (ULong
)1;
2481 *bx
++ = y
& FFFFFFFF
;
2485 ys
= (si
& 0xffff) * q
+ carry
;
2486 zs
= (si
>> 16) * q
+ (ys
>> 16);
2488 y
= (*bx
& 0xffff) - (ys
& 0xffff) - borrow
;
2489 borrow
= (y
& 0x10000) >> 16;
2490 z
= (*bx
>> 16) - (zs
& 0xffff) - borrow
;
2491 borrow
= (z
& 0x10000) >> 16;
2494 ys
= *sx
++ * q
+ carry
;
2496 y
= *bx
- (ys
& 0xffff) - borrow
;
2497 borrow
= (y
& 0x10000) >> 16;
2505 while(--bxe
> bx
&& !*bxe
)
2510 if (cmp(b
, S
) >= 0) {
2520 y
= *bx
- (ys
& FFFFFFFF
) - borrow
;
2521 borrow
= y
>> 32 & (ULong
)1;
2522 *bx
++ = y
& FFFFFFFF
;
2526 ys
= (si
& 0xffff) + carry
;
2527 zs
= (si
>> 16) + (ys
>> 16);
2529 y
= (*bx
& 0xffff) - (ys
& 0xffff) - borrow
;
2530 borrow
= (y
& 0x10000) >> 16;
2531 z
= (*bx
>> 16) - (zs
& 0xffff) - borrow
;
2532 borrow
= (z
& 0x10000) >> 16;
2537 y
= *bx
- (ys
& 0xffff) - borrow
;
2538 borrow
= (y
& 0x10000) >> 16;
2547 while(--bxe
> bx
&& !*bxe
)
2555 #ifndef MULTIPLE_THREADS
2556 static char *dtoa_result
;
2570 sizeof(Bigint
) - sizeof(ULong
) - sizeof(int) + j
<= i
;
2573 r
= (int*)Balloc(k
);
2576 #ifndef MULTIPLE_THREADS
2584 nrv_alloc(s
, rve
, n
) char *s
, **rve
; int n
;
2586 nrv_alloc(char *s
, char **rve
, int n
)
2591 t
= rv
= rv_alloc(n
);
2592 while((*t
= *s
++)) t
++;
2598 /* freedtoa(s) must be used to free values s returned by dtoa
2599 * when MULTIPLE_THREADS is #defined. It should be used in all cases,
2600 * but for consistency with earlier versions of dtoa, it is optional
2601 * when MULTIPLE_THREADS is not defined.
2604 static void freedtoa (char *s
);
2608 freedtoa(s
) char *s
;
2613 Bigint
*b
= (Bigint
*)((int *)s
- 1);
2614 b
->maxwds
= 1 << (b
->k
= *(int*)b
);
2616 #ifndef MULTIPLE_THREADS
2617 if (s
== dtoa_result
)
2623 /* dtoa for IEEE arithmetic (dmg): convert double to ASCII string.
2625 * Inspired by "How to Print Floating-Point Numbers Accurately" by
2626 * Guy L. Steele, Jr. and Jon L. White [Proc. ACM SIGPLAN '90, pp. 112-126].
2629 * 1. Rather than iterating, we use a simple numeric overestimate
2630 * to determine k = floor(log10(d)). We scale relevant
2631 * quantities using O(log2(k)) rather than O(k) multiplications.
2632 * 2. For some modes > 2 (corresponding to ecvt and fcvt), we don't
2633 * try to generate digits strictly left to right. Instead, we
2634 * compute with fewer bits and propagate the carry if necessary
2635 * when rounding the final digit up. This is often faster.
2636 * 3. Under the assumption that input will be rounded nearest,
2637 * mode 0 renders 1e23 as 1e23 rather than 9.999999999999999e22.
2638 * That is, we allow equality in stopping tests when the
2639 * round-nearest rule will give the same floating-point value
2640 * as would satisfaction of the stopping test with strict
2642 * 4. We remove common factors of powers of 2 from relevant
2644 * 5. When converting floating-point integers less than 1e16,
2645 * we use floating-point arithmetic rather than resorting
2646 * to multiple-precision integers.
2647 * 6. When asked to produce fewer than 15 digits, we first try
2648 * to get by with floating-point arithmetic; we resort to
2649 * multiple-precision integer arithmetic only if we cannot
2650 * guarantee that the floating-point calculation has given
2651 * the correctly rounded result. For k requested digits and
2652 * "uniformly" distributed input, the probability is
2653 * something like 10^(k-15) that we must resort to the Long
2660 (d
, mode
, ndigits
, decpt
, sign
, rve
)
2661 double d
; int mode
, ndigits
, *decpt
, *sign
; char **rve
;
2663 (double d
, int mode
, int ndigits
, int *decpt
, int *sign
, char **rve
)
2666 /* Arguments ndigits, decpt, sign are similar to those
2667 of ecvt and fcvt; trailing zeros are suppressed from
2668 the returned string. If not null, *rve is set to point
2669 to the end of the return value. If d is +-Infinity or NaN,
2670 then *decpt is set to 9999.
2673 0 ==> shortest string that yields d when read in
2674 and rounded to nearest.
2675 1 ==> like 0, but with Steele & White stopping rule;
2676 e.g. with IEEE P754 arithmetic , mode 0 gives
2677 1e23 whereas mode 1 gives 9.999999999999999e22.
2678 2 ==> max(1,ndigits) significant digits. This gives a
2679 return value similar to that of ecvt, except
2680 that trailing zeros are suppressed.
2681 3 ==> through ndigits past the decimal point. This
2682 gives a return value similar to that from fcvt,
2683 except that trailing zeros are suppressed, and
2684 ndigits can be negative.
2685 4,5 ==> similar to 2 and 3, respectively, but (in
2686 round-nearest mode) with the tests of mode 0 to
2687 possibly return a shorter string that rounds to d.
2688 With IEEE arithmetic and compilation with
2689 -DHonor_FLT_ROUNDS, modes 4 and 5 behave the same
2690 as modes 2 and 3 when FLT_ROUNDS != 1.
2691 6-9 ==> Debugging modes similar to mode - 4: don't try
2692 fast floating-point estimate (if applicable).
2694 Values of mode other than 0-9 are treated as mode 0.
2696 Sufficient space is allocated to the return value
2697 to hold the suppressed trailing zeros.
2700 int bbits
, b2
, b5
, be
, dig
, i
, ieps
, ilim
, ilim0
, ilim1
,
2701 j
, j1
, k
, k0
, k_check
, leftright
, m2
, m5
, s2
, s5
,
2702 spec_case
, try_quick
;
2704 #ifndef Sudden_Underflow
2708 Bigint
*b
, *b1
, *delta
, *mlo
, *mhi
, *S
;
2711 #ifdef Honor_FLT_ROUNDS
2715 int inexact
, oldinexact
;
2718 #ifndef MULTIPLE_THREADS
2720 freedtoa(dtoa_result
);
2725 if (word0(d
) & Sign_bit
) {
2726 /* set sign for everything, including 0's and NaNs */
2728 word0(d
) &= ~Sign_bit
; /* clear sign bit */
2733 #if defined(IEEE_Arith) + defined(VAX)
2735 if ((word0(d
) & Exp_mask
) == Exp_mask
)
2737 if (word0(d
) == 0x8000)
2740 /* Infinity or NaN */
2743 if (!word1(d
) && !(word0(d
) & 0xfffff))
2744 return nrv_alloc("Infinity", rve
, 8);
2746 return nrv_alloc("NaN", rve
, 3);
2750 dval(d
) += 0; /* normalize */
2754 return nrv_alloc("0", rve
, 1);
2758 try_quick
= oldinexact
= get_inexact();
2761 #ifdef Honor_FLT_ROUNDS
2762 if ((rounding
= Flt_Rounds
) >= 2) {
2764 rounding
= rounding
== 2 ? 0 : 2;
2771 b
= d2b(dval(d
), &be
, &bbits
);
2772 #ifdef Sudden_Underflow
2773 i
= (int)(word0(d
) >> Exp_shift1
& (Exp_mask
>>Exp_shift1
));
2775 if (i
= (int)(word0(d
) >> Exp_shift1
& (Exp_mask
>>Exp_shift1
))) {
2778 word0(d2
) &= Frac_mask1
;
2779 word0(d2
) |= Exp_11
;
2781 if (j
= 11 - hi0bits(word0(d2
) & Frac_mask
))
2785 /* log(x) ~=~ log(1.5) + (x-1.5)/1.5
2786 * log10(x) = log(x) / log(10)
2787 * ~=~ log(1.5)/log(10) + (x-1.5)/(1.5*log(10))
2788 * log10(d) = (i-Bias)*log(2)/log(10) + log10(d2)
2790 * This suggests computing an approximation k to log10(d) by
2792 * k = (i - Bias)*0.301029995663981
2793 * + ( (d2-1.5)*0.289529654602168 + 0.176091259055681 );
2795 * We want k to be too large rather than too small.
2796 * The error in the first-order Taylor series approximation
2797 * is in our favor, so we just round up the constant enough
2798 * to compensate for any error in the multiplication of
2799 * (i - Bias) by 0.301029995663981; since |i - Bias| <= 1077,
2800 * and 1077 * 0.30103 * 2^-52 ~=~ 7.2e-14,
2801 * adding 1e-13 to the constant term more than suffices.
2802 * Hence we adjust the constant term to 0.1760912590558.
2803 * (We could get a more accurate k by invoking log10,
2804 * but this is probably not worthwhile.)
2812 #ifndef Sudden_Underflow
2816 /* d is denormalized */
2818 i
= bbits
+ be
+ (Bias
+ (P
-1) - 1);
2819 x
= i
> 32 ? word0(d
) << 64 - i
| word1(d
) >> i
- 32
2820 : word1(d
) << 32 - i
;
2822 word0(d2
) -= 31*Exp_msk1
; /* adjust exponent */
2823 i
-= (Bias
+ (P
-1) - 1) + 1;
2827 ds
= (dval(d2
)-1.5)*0.289529654602168 + 0.1760912590558 + i
*0.301029995663981;
2829 if (ds
< 0. && ds
!= k
)
2830 k
--; /* want k = floor(ds) */
2832 if (k
>= 0 && k
<= Ten_pmax
) {
2833 if (dval(d
) < tens
[k
])
2856 if (mode
< 0 || mode
> 9)
2860 #ifdef Check_FLT_ROUNDS
2861 try_quick
= Rounding
== 1;
2865 #endif /*SET_INEXACT*/
2885 ilim
= ilim1
= i
= ndigits
;
2891 i
= ndigits
+ k
+ 1;
2897 s
= s0
= rv_alloc(i
);
2899 #ifdef Honor_FLT_ROUNDS
2900 if (mode
> 1 && rounding
!= 1)
2904 if (ilim
>= 0 && ilim
<= Quick_max
&& try_quick
) {
2906 /* Try to get by with floating-point arithmetic. */
2912 ieps
= 2; /* conservative */
2917 /* prevent overflows */
2919 dval(d
) /= bigtens
[n_bigtens
-1];
2922 for(; j
; j
>>= 1, i
++)
2930 dval(d
) *= tens
[j1
& 0xf];
2931 for(j
= j1
>> 4; j
; j
>>= 1, i
++)
2934 dval(d
) *= bigtens
[i
];
2937 if (k_check
&& dval(d
) < 1. && ilim
> 0) {
2945 dval(eps
) = ieps
*dval(d
) + 7.;
2946 word0(eps
) -= (P
-1)*Exp_msk1
;
2950 if (dval(d
) > dval(eps
))
2952 if (dval(d
) < -dval(eps
))
2956 #ifndef No_leftright
2958 /* Use Steele & White method of only
2959 * generating digits needed.
2961 dval(eps
) = 0.5/tens
[ilim
-1] - dval(eps
);
2965 *s
++ = '0' + (int)L
;
2966 if (dval(d
) < dval(eps
))
2968 if (1. - dval(d
) < dval(eps
))
2978 /* Generate ilim digits, then fix them up. */
2979 dval(eps
) *= tens
[ilim
-1];
2980 for(i
= 1;; i
++, dval(d
) *= 10.) {
2981 L
= (Long
)(dval(d
));
2982 if (!(dval(d
) -= L
))
2984 *s
++ = '0' + (int)L
;
2986 if (dval(d
) > 0.5 + dval(eps
))
2988 else if (dval(d
) < 0.5 - dval(eps
)) {
2996 #ifndef No_leftright
3006 /* Do we have a "small" integer? */
3008 if (be
>= 0 && k
<= Int_max
) {
3011 if (ndigits
< 0 && ilim
<= 0) {
3013 if (ilim
< 0 || dval(d
) <= 5*ds
)
3017 for(i
= 1;; i
++, dval(d
) *= 10.) {
3018 L
= (Long
)(dval(d
) / ds
);
3020 #ifdef Check_FLT_ROUNDS
3021 /* If FLT_ROUNDS == 2, L will usually be high by 1 */
3027 *s
++ = '0' + (int)L
;
3035 #ifdef Honor_FLT_ROUNDS
3039 case 2: goto bump_up
;
3043 if (dval(d
) > ds
|| dval(d
) == ds
&& L
& 1) {
3064 #ifndef Sudden_Underflow
3065 denorm
? be
+ (Bias
+ (P
-1) - 1 + 1) :
3068 1 + 4*P
- 3 - bbits
+ ((bbits
+ be
- 1) & 3);
3076 if (m2
> 0 && s2
> 0) {
3077 i
= m2
< s2
? m2
: s2
;
3085 mhi
= pow5mult(mhi
, m5
);
3094 b
= pow5mult(b
, b5
);
3098 S
= pow5mult(S
, s5
);
3100 /* Check for special case that d is a normalized power of 2. */
3103 if ((mode
< 2 || leftright
)
3104 #ifdef Honor_FLT_ROUNDS
3108 if (!word1(d
) && !(word0(d
) & Bndry_mask
)
3109 #ifndef Sudden_Underflow
3110 && word0(d
) & (Exp_mask
& ~Exp_msk1
)
3113 /* The special case */
3120 /* Arrange for convenient computation of quotients:
3121 * shift left if necessary so divisor has 4 leading 0 bits.
3123 * Perhaps we should just compute leading 28 bits of S once
3124 * and for all and pass them and a shift to quorem, so it
3125 * can do shifts and ors to compute the numerator for q.
3128 if (i
= ((s5
? 32 - hi0bits(S
->x
[S
->wds
-1]) : 1) + s2
) & 0x1f)
3131 if (i
= ((s5
? 32 - hi0bits(S
->x
[S
->wds
-1]) : 1) + s2
) & 0xf)
3153 b
= multadd(b
, 10, 0); /* we botched the k estimate */
3155 mhi
= multadd(mhi
, 10, 0);
3159 if (ilim
<= 0 && (mode
== 3 || mode
== 5)) {
3160 if (ilim
< 0 || cmp(b
,S
= multadd(S
,5,0)) <= 0) {
3161 /* no digits, fcvt style */
3173 mhi
= lshift(mhi
, m2
);
3175 /* Compute mlo -- check for special case
3176 * that d is a normalized power of 2.
3181 mhi
= Balloc(mhi
->k
);
3183 mhi
= lshift(mhi
, Log2P
);
3187 dig
= quorem(b
,S
) + '0';
3188 /* Do we yet have the shortest decimal string
3189 * that will round to d?
3192 delta
= diff(S
, mhi
);
3193 j1
= delta
->sign
? 1 : cmp(b
, delta
);
3195 #ifndef ROUND_BIASED
3196 if (j1
== 0 && mode
!= 1 && !(word1(d
) & 1)
3197 #ifdef Honor_FLT_ROUNDS
3206 else if (!b
->x
[0] && b
->wds
<= 1)
3213 if (j
< 0 || j
== 0 && mode
!= 1
3214 #ifndef ROUND_BIASED
3218 if (!b
->x
[0] && b
->wds
<= 1) {
3224 #ifdef Honor_FLT_ROUNDS
3227 case 0: goto accept_dig
;
3228 case 2: goto keep_dig
;
3230 #endif /*Honor_FLT_ROUNDS*/
3234 if ((j1
> 0 || j1
== 0 && dig
& 1)
3243 #ifdef Honor_FLT_ROUNDS
3247 if (dig
== '9') { /* possible if i == 1 */
3255 #ifdef Honor_FLT_ROUNDS
3261 b
= multadd(b
, 10, 0);
3263 mlo
= mhi
= multadd(mhi
, 10, 0);
3265 mlo
= multadd(mlo
, 10, 0);
3266 mhi
= multadd(mhi
, 10, 0);
3272 *s
++ = dig
= quorem(b
,S
) + '0';
3273 if (!b
->x
[0] && b
->wds
<= 1) {
3281 b
= multadd(b
, 10, 0);
3284 /* Round off last digit */
3286 #ifdef Honor_FLT_ROUNDS
3288 case 0: goto trimzeros
;
3289 case 2: goto roundoff
;
3294 if (j
> 0 || j
== 0 && dig
& 1) {
3312 if (mlo
&& mlo
!= mhi
)
3320 word0(d
) = Exp_1
+ (70 << Exp_shift
);
3325 else if (!oldinexact
)