Coverage Report

Created: 2025-07-04 06:49

/src/cpython/Python/dtoa.c
Line
Count
Source (jump to first uncovered line)
1
/****************************************************************
2
 *
3
 * The author of this software is David M. Gay.
4
 *
5
 * Copyright (c) 1991, 2000, 2001 by Lucent Technologies.
6
 *
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.
12
 *
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.
17
 *
18
 ***************************************************************/
19
20
/****************************************************************
21
 * This is dtoa.c by David M. Gay, downloaded from
22
 * http://www.netlib.org/fp/dtoa.c on April 15, 2009 and modified for
23
 * inclusion into the Python core by Mark E. T. Dickinson and Eric V. Smith.
24
 *
25
 * Please remember to check http://www.netlib.org/fp regularly (and especially
26
 * before any Python release) for bugfixes and updates.
27
 *
28
 * The major modifications from Gay's original code are as follows:
29
 *
30
 *  0. The original code has been specialized to Python's needs by removing
31
 *     many of the #ifdef'd sections.  In particular, code to support VAX and
32
 *     IBM floating-point formats, hex NaNs, hex floats, locale-aware
33
 *     treatment of the decimal point, and setting of the inexact flag have
34
 *     been removed.
35
 *
36
 *  1. We use PyMem_Malloc and PyMem_Free in place of malloc and free.
37
 *
38
 *  2. The public functions strtod, dtoa and freedtoa all now have
39
 *     a _Py_dg_ prefix.
40
 *
41
 *  3. Instead of assuming that PyMem_Malloc always succeeds, we thread
42
 *     PyMem_Malloc failures through the code.  The functions
43
 *
44
 *       Balloc, multadd, s2b, i2b, mult, pow5mult, lshift, diff, d2b
45
 *
46
 *     of return type *Bigint all return NULL to indicate a malloc failure.
47
 *     Similarly, rv_alloc and nrv_alloc (return type char *) return NULL on
48
 *     failure.  bigcomp now has return type int (it used to be void) and
49
 *     returns -1 on failure and 0 otherwise.  _Py_dg_dtoa returns NULL
50
 *     on failure.  _Py_dg_strtod indicates failure due to malloc failure
51
 *     by returning -1.0, setting errno=ENOMEM and *se to s00.
52
 *
53
 *  4. The static variable dtoa_result has been removed.  Callers of
54
 *     _Py_dg_dtoa are expected to call _Py_dg_freedtoa to free
55
 *     the memory allocated by _Py_dg_dtoa.
56
 *
57
 *  5. The code has been reformatted to better fit with Python's
58
 *     C style guide (PEP 7).
59
 *
60
 *  6. A bug in the memory allocation has been fixed: to avoid FREEing memory
61
 *     that hasn't been MALLOC'ed, private_mem should only be used when k <=
62
 *     Kmax.
63
 *
64
 *  7. _Py_dg_strtod has been modified so that it doesn't accept strings with
65
 *     leading whitespace.
66
 *
67
 *  8. A corner case where _Py_dg_dtoa didn't strip trailing zeros has been
68
 *     fixed. (bugs.python.org/issue40780)
69
 *
70
 ***************************************************************/
71
72
/* Please send bug reports for the original dtoa.c code to David M. Gay (dmg
73
 * at acm dot org, with " at " changed at "@" and " dot " changed to ".").
74
 * Please report bugs for this modified version using the Python issue tracker
75
 * as detailed at (https://devguide.python.org/triage/issue-tracker/). */
76
77
/* On a machine with IEEE extended-precision registers, it is
78
 * necessary to specify double-precision (53-bit) rounding precision
79
 * before invoking strtod or dtoa.  If the machine uses (the equivalent
80
 * of) Intel 80x87 arithmetic, the call
81
 *      _control87(PC_53, MCW_PC);
82
 * does this with many compilers.  Whether this or another call is
83
 * appropriate depends on the compiler; for this to work, it may be
84
 * necessary to #include "float.h" or another system-dependent header
85
 * file.
86
 */
87
88
/* strtod for IEEE-, VAX-, and IBM-arithmetic machines.
89
 *
90
 * This strtod returns a nearest machine number to the input decimal
91
 * string (or sets errno to ERANGE).  With IEEE arithmetic, ties are
92
 * broken by the IEEE round-even rule.  Otherwise ties are broken by
93
 * biased rounding (add half and chop).
94
 *
95
 * Inspired loosely by William D. Clinger's paper "How to Read Floating
96
 * Point Numbers Accurately" [Proc. ACM SIGPLAN '90, pp. 92-101].
97
 *
98
 * Modifications:
99
 *
100
 *      1. We only require IEEE, IBM, or VAX double-precision
101
 *              arithmetic (not IEEE double-extended).
102
 *      2. We get by with floating-point arithmetic in a case that
103
 *              Clinger missed -- when we're computing d * 10^n
104
 *              for a small integer d and the integer n is not too
105
 *              much larger than 22 (the maximum integer k for which
106
 *              we can represent 10^k exactly), we may be able to
107
 *              compute (d*10^k) * 10^(e-k) with just one roundoff.
108
 *      3. Rather than a bit-at-a-time adjustment of the binary
109
 *              result in the hard case, we use floating-point
110
 *              arithmetic to determine the adjustment to within
111
 *              one bit; only in really hard cases do we need to
112
 *              compute a second residual.
113
 *      4. Because of 3., we don't need a large table of powers of 10
114
 *              for ten-to-e (just some small tables, e.g. of 10^k
115
 *              for 0 <= k <= 22).
116
 */
117
118
/* Linking of Python's #defines to Gay's #defines starts here. */
119
120
#include "Python.h"
121
#include "pycore_dtoa.h"          // _PY_SHORT_FLOAT_REPR
122
#include "pycore_interp_structs.h"// struct Bigint
123
#include "pycore_pystate.h"       // _PyInterpreterState_GET()
124
#include <stdlib.h>               // exit()
125
126
127
/* if _PY_SHORT_FLOAT_REPR == 0, then don't even try to compile
128
   the following code */
129
#if _PY_SHORT_FLOAT_REPR == 1
130
131
#include "float.h"
132
133
34
#define MALLOC PyMem_Malloc
134
0
#define FREE PyMem_Free
135
136
/* This code should also work for ARM mixed-endian format on little-endian
137
   machines, where doubles have byte order 45670123 (in increasing address
138
   order, 0 being the least significant byte). */
139
#ifdef DOUBLE_IS_LITTLE_ENDIAN_IEEE754
140
#  define IEEE_8087
141
#endif
142
#if defined(DOUBLE_IS_BIG_ENDIAN_IEEE754) ||  \
143
  defined(DOUBLE_IS_ARM_MIXED_ENDIAN_IEEE754)
144
#  define IEEE_MC68k
145
#endif
146
#if defined(IEEE_8087) + defined(IEEE_MC68k) != 1
147
#error "Exactly one of IEEE_8087 or IEEE_MC68k should be defined."
148
#endif
149
150
/* The code below assumes that the endianness of integers matches the
151
   endianness of the two 32-bit words of a double.  Check this. */
152
#if defined(WORDS_BIGENDIAN) && (defined(DOUBLE_IS_LITTLE_ENDIAN_IEEE754) || \
153
                                 defined(DOUBLE_IS_ARM_MIXED_ENDIAN_IEEE754))
154
#error "doubles and ints have incompatible endianness"
155
#endif
156
157
#if !defined(WORDS_BIGENDIAN) && defined(DOUBLE_IS_BIG_ENDIAN_IEEE754)
158
#error "doubles and ints have incompatible endianness"
159
#endif
160
161
162
typedef uint32_t ULong;
163
typedef int32_t Long;
164
typedef uint64_t ULLong;
165
166
#undef DEBUG
167
#ifdef Py_DEBUG
168
#define DEBUG
169
#endif
170
171
/* End Python #define linking */
172
173
#ifdef DEBUG
174
#define Bug(x) {fprintf(stderr, "%s\n", x); exit(1);}
175
#endif
176
177
typedef union { double d; ULong L[2]; } U;
178
179
#ifdef IEEE_8087
180
1.59M
#define word0(x) (x)->L[1]
181
1.04M
#define word1(x) (x)->L[0]
182
#else
183
#define word0(x) (x)->L[0]
184
#define word1(x) (x)->L[1]
185
#endif
186
4.05M
#define dval(x) (x)->d
187
188
#ifndef STRTOD_DIGLIM
189
75.3k
#define STRTOD_DIGLIM 40
190
#endif
191
192
/* maximum permitted exponent value for strtod; exponents larger than
193
   MAX_ABS_EXP in absolute value get truncated to +-MAX_ABS_EXP.  MAX_ABS_EXP
194
   should fit into an int. */
195
#ifndef MAX_ABS_EXP
196
568k
#define MAX_ABS_EXP 1100000000U
197
#endif
198
/* Bound on length of pieces of input strings in _Py_dg_strtod; specifically,
199
   this is used to bound the total number of digits ignoring leading zeros and
200
   the number of digits that follow the decimal point.  Ideally, MAX_DIGITS
201
   should satisfy MAX_DIGITS + 400 < MAX_ABS_EXP; that ensures that the
202
   exponent clipping in _Py_dg_strtod can't affect the value of the output. */
203
#ifndef MAX_DIGITS
204
1.90M
#define MAX_DIGITS 1000000000U
205
#endif
206
207
/* Guard against trying to use the above values on unusual platforms with ints
208
 * of width less than 32 bits. */
209
#if MAX_ABS_EXP > INT_MAX
210
#error "MAX_ABS_EXP should fit in an int"
211
#endif
212
#if MAX_DIGITS > INT_MAX
213
#error "MAX_DIGITS should fit in an int"
214
#endif
215
216
/* The following definition of Storeinc is appropriate for MIPS processors.
217
 * An alternative that might be better on some machines is
218
 * #define Storeinc(a,b,c) (*a++ = b << 16 | c & 0xffff)
219
 */
220
#if defined(IEEE_8087)
221
#define Storeinc(a,b,c) (((unsigned short *)a)[1] = (unsigned short)b,  \
222
                         ((unsigned short *)a)[0] = (unsigned short)c, a++)
223
#else
224
#define Storeinc(a,b,c) (((unsigned short *)a)[0] = (unsigned short)b,  \
225
                         ((unsigned short *)a)[1] = (unsigned short)c, a++)
226
#endif
227
228
/* #define P DBL_MANT_DIG */
229
/* Ten_pmax = floor(P*log(2)/log(5)) */
230
/* Bletch = (highest power of 2 < DBL_MAX_10_EXP) / 16 */
231
/* Quick_max = floor((P-1)*log(FLT_RADIX)/log(10) - 1) */
232
/* Int_max = floor(P*log(FLT_RADIX)/log(10) - 1) */
233
234
172k
#define Exp_shift  20
235
99.7k
#define Exp_shift1 20
236
477k
#define Exp_msk1    0x100000
237
#define Exp_msk11   0x100000
238
928k
#define Exp_mask  0x7ff00000
239
421k
#define P 53
240
#define Nbits 53
241
223k
#define Bias 1023
242
#define Emax 1023
243
#define Emin (-1022)
244
289k
#define Etiny (-1074)  /* smallest denormal is 2**Etiny */
245
98.4k
#define Exp_1  0x3ff00000
246
45.7k
#define Exp_11 0x3ff00000
247
200k
#define Ebits 11
248
152k
#define Frac_mask  0xfffff
249
48.4k
#define Frac_mask1 0xfffff
250
1.12M
#define Ten_pmax 22
251
0
#define Bletch 0x10
252
63.0k
#define Bndry_mask  0xfffff
253
6.66k
#define Bndry_mask1 0xfffff
254
69.3k
#define Sign_bit 0x80000000
255
5.77k
#define Log2P 1
256
#define Tiny0 0
257
25.7k
#define Tiny1 1
258
49.8k
#define Quick_max 14
259
31.9k
#define Int_max 14
260
261
#ifndef Flt_Rounds
262
#ifdef FLT_ROUNDS
263
605k
#define Flt_Rounds FLT_ROUNDS
264
#else
265
#define Flt_Rounds 1
266
#endif
267
#endif /*Flt_Rounds*/
268
269
#define Rounding Flt_Rounds
270
271
4.26k
#define Big0 (Frac_mask1 | Exp_msk1*(DBL_MAX_EXP+Bias-1))
272
2.67k
#define Big1 0xffffffff
273
274
/* Bits of the representation of positive infinity. */
275
276
#define POSINF_WORD0 0x7ff00000
277
#define POSINF_WORD1 0
278
279
/* struct BCinfo is used to pass information from _Py_dg_strtod to bigcomp */
280
281
typedef struct BCinfo BCinfo;
282
struct
283
BCinfo {
284
    int e0, nd, nd0, scale;
285
};
286
287
16.2M
#define FFFFFFFF 0xffffffffUL
288
289
/* struct Bigint is used to represent arbitrary-precision integers.  These
290
   integers are stored in sign-magnitude format, with the magnitude stored as
291
   an array of base 2**32 digits.  Bigints are always normalized: if x is a
292
   Bigint then x->wds >= 1, and either x->wds == 1 or x[wds-1] is nonzero.
293
294
   The Bigint fields are as follows:
295
296
     - next is a header used by Balloc and Bfree to keep track of lists
297
         of freed Bigints;  it's also used for the linked list of
298
         powers of 5 of the form 5**2**i used by pow5mult.
299
     - k indicates which pool this Bigint was allocated from
300
     - maxwds is the maximum number of words space was allocated for
301
       (usually maxwds == 2**k)
302
     - sign is 1 for negative Bigints, 0 for positive.  The sign is unused
303
       (ignored on inputs, set to 0 on outputs) in almost all operations
304
       involving Bigints: a notable exception is the diff function, which
305
       ignores signs on inputs but sets the sign of the output correctly.
306
     - wds is the actual number of significant words
307
     - x contains the vector of words (digits) for this Bigint, from least
308
       significant (x[0]) to most significant (x[wds-1]).
309
*/
310
311
// struct Bigint is defined in pycore_dtoa.h.
312
typedef struct Bigint Bigint;
313
314
#if !defined(Py_GIL_DISABLED) && !defined(Py_USING_MEMORY_DEBUGGER)
315
316
/* Memory management: memory is allocated from, and returned to, Kmax+1 pools
317
   of memory, where pool k (0 <= k <= Kmax) is for Bigints b with b->maxwds ==
318
   1 << k.  These pools are maintained as linked lists, with freelist[k]
319
   pointing to the head of the list for pool k.
320
321
   On allocation, if there's no free slot in the appropriate pool, MALLOC is
322
   called to get more memory.  This memory is not returned to the system until
323
   Python quits.  There's also a private memory pool that's allocated from
324
   in preference to using MALLOC.
325
326
   For Bigints with more than (1 << Kmax) digits (which implies at least 1233
327
   decimal digits), memory is directly allocated using MALLOC, and freed using
328
   FREE.
329
330
   XXX: it would be easy to bypass this memory-management system and
331
   translate each call to Balloc into a call to PyMem_Malloc, and each
332
   Bfree to PyMem_Free.  Investigate whether this has any significant
333
   performance on impact. */
334
335
6.17M
#define freelist interp->dtoa.freelist
336
207
#define private_mem interp->dtoa.preallocated
337
553
#define pmem_next interp->dtoa.preallocated_next
338
339
/* Allocate space for a Bigint with up to 1<<k digits */
340
341
static Bigint *
342
Balloc(int k)
343
1.54M
{
344
1.54M
    int x;
345
1.54M
    Bigint *rv;
346
1.54M
    unsigned int len;
347
1.54M
    PyInterpreterState *interp = _PyInterpreterState_GET();
348
349
1.54M
    if (k <= Bigint_Kmax && (rv = freelist[k]))
350
1.54M
        freelist[k] = rv->next;
351
207
    else {
352
207
        x = 1 << k;
353
207
        len = (sizeof(Bigint) + (x-1)*sizeof(ULong) + sizeof(double) - 1)
354
207
            /sizeof(double);
355
207
        if (k <= Bigint_Kmax &&
356
207
            pmem_next - private_mem + len <= (Py_ssize_t)Bigint_PREALLOC_SIZE
357
207
        ) {
358
173
            rv = (Bigint*)pmem_next;
359
173
            pmem_next += len;
360
173
        }
361
34
        else {
362
34
            rv = (Bigint*)MALLOC(len*sizeof(double));
363
34
            if (rv == NULL)
364
0
                return NULL;
365
34
        }
366
207
        rv->k = k;
367
207
        rv->maxwds = x;
368
207
    }
369
1.54M
    rv->sign = rv->wds = 0;
370
1.54M
    return rv;
371
1.54M
}
372
373
/* Free a Bigint allocated with Balloc */
374
375
static void
376
Bfree(Bigint *v)
377
4.34M
{
378
4.34M
    if (v) {
379
1.54M
        if (v->k > Bigint_Kmax)
380
0
            FREE((void*)v);
381
1.54M
        else {
382
1.54M
            PyInterpreterState *interp = _PyInterpreterState_GET();
383
1.54M
            v->next = freelist[v->k];
384
1.54M
            freelist[v->k] = v;
385
1.54M
        }
386
1.54M
    }
387
4.34M
}
388
389
#undef pmem_next
390
#undef private_mem
391
#undef freelist
392
393
#else
394
395
/* Alternative versions of Balloc and Bfree that use PyMem_Malloc and
396
   PyMem_Free directly in place of the custom memory allocation scheme above.
397
   These are provided for the benefit of memory debugging tools like
398
   Valgrind. */
399
400
/* Allocate space for a Bigint with up to 1<<k digits */
401
402
static Bigint *
403
Balloc(int k)
404
{
405
    int x;
406
    Bigint *rv;
407
    unsigned int len;
408
409
    x = 1 << k;
410
    len = (sizeof(Bigint) + (x-1)*sizeof(ULong) + sizeof(double) - 1)
411
        /sizeof(double);
412
413
    rv = (Bigint*)MALLOC(len*sizeof(double));
414
    if (rv == NULL)
415
        return NULL;
416
417
    rv->k = k;
418
    rv->maxwds = x;
419
    rv->sign = rv->wds = 0;
420
    return rv;
421
}
422
423
/* Free a Bigint allocated with Balloc */
424
425
static void
426
Bfree(Bigint *v)
427
{
428
    if (v) {
429
        FREE((void*)v);
430
    }
431
}
432
433
#endif /* !defined(Py_GIL_DISABLED) && !defined(Py_USING_MEMORY_DEBUGGER) */
434
435
98.2k
#define Bcopy(x,y) memcpy((char *)&x->sign, (char *)&y->sign,   \
436
98.2k
                          y->wds*sizeof(Long) + 2*sizeof(int))
437
438
/* Multiply a Bigint b by m and add a.  Either modifies b in place and returns
439
   a pointer to the modified b, or Bfrees b and returns a pointer to a copy.
440
   On failure, return NULL.  In this case, b will have been already freed. */
441
442
static Bigint *
443
multadd(Bigint *b, int m, int a)       /* multiply by m and add a */
444
736k
{
445
736k
    int i, wds;
446
736k
    ULong *x;
447
736k
    ULLong carry, y;
448
736k
    Bigint *b1;
449
450
736k
    wds = b->wds;
451
736k
    x = b->x;
452
736k
    i = 0;
453
736k
    carry = a;
454
2.55M
    do {
455
2.55M
        y = *x * (ULLong)m + carry;
456
2.55M
        carry = y >> 32;
457
2.55M
        *x++ = (ULong)(y & FFFFFFFF);
458
2.55M
    }
459
2.55M
    while(++i < wds);
460
736k
    if (carry) {
461
47.4k
        if (wds >= b->maxwds) {
462
1.77k
            b1 = Balloc(b->k+1);
463
1.77k
            if (b1 == NULL){
464
0
                Bfree(b);
465
0
                return NULL;
466
0
            }
467
1.77k
            Bcopy(b1, b);
468
1.77k
            Bfree(b);
469
1.77k
            b = b1;
470
1.77k
        }
471
47.4k
        b->x[wds++] = (ULong)carry;
472
47.4k
        b->wds = wds;
473
47.4k
    }
474
736k
    return b;
475
736k
}
476
477
/* convert a string s containing nd decimal digits (possibly containing a
478
   decimal separator at position nd0, which is ignored) to a Bigint.  This
479
   function carries on where the parsing code in _Py_dg_strtod leaves off: on
480
   entry, y9 contains the result of converting the first 9 digits.  Returns
481
   NULL on failure. */
482
483
static Bigint *
484
s2b(const char *s, int nd0, int nd, ULong y9)
485
75.3k
{
486
75.3k
    Bigint *b;
487
75.3k
    int i, k;
488
75.3k
    Long x, y;
489
490
75.3k
    x = (nd + 8) / 9;
491
111k
    for(k = 0, y = 1; x > y; y <<= 1, k++) ;
492
75.3k
    b = Balloc(k);
493
75.3k
    if (b == NULL)
494
0
        return NULL;
495
75.3k
    b->x[0] = y9;
496
75.3k
    b->wds = 1;
497
498
75.3k
    if (nd <= 9)
499
47.0k
      return b;
500
501
28.3k
    s += 9;
502
211k
    for (i = 9; i < nd0; i++) {
503
183k
        b = multadd(b, 10, *s++ - '0');
504
183k
        if (b == NULL)
505
0
            return NULL;
506
183k
    }
507
28.3k
    s++;
508
106k
    for(; i < nd; i++) {
509
77.9k
        b = multadd(b, 10, *s++ - '0');
510
77.9k
        if (b == NULL)
511
0
            return NULL;
512
77.9k
    }
513
28.3k
    return b;
514
28.3k
}
515
516
/* count leading 0 bits in the 32-bit integer x. */
517
518
static int
519
hi0bits(ULong x)
520
127k
{
521
127k
    int k = 0;
522
523
127k
    if (!(x & 0xffff0000)) {
524
77.1k
        k = 16;
525
77.1k
        x <<= 16;
526
77.1k
    }
527
127k
    if (!(x & 0xff000000)) {
528
80.5k
        k += 8;
529
80.5k
        x <<= 8;
530
80.5k
    }
531
127k
    if (!(x & 0xf0000000)) {
532
74.2k
        k += 4;
533
74.2k
        x <<= 4;
534
74.2k
    }
535
127k
    if (!(x & 0xc0000000)) {
536
72.4k
        k += 2;
537
72.4k
        x <<= 2;
538
72.4k
    }
539
127k
    if (!(x & 0x80000000)) {
540
75.5k
        k++;
541
75.5k
        if (!(x & 0x40000000))
542
0
            return 32;
543
75.5k
    }
544
127k
    return k;
545
127k
}
546
547
/* count trailing 0 bits in the 32-bit integer y, and shift y right by that
548
   number of bits. */
549
550
static int
551
lo0bits(ULong *y)
552
49.8k
{
553
49.8k
    int k;
554
49.8k
    ULong x = *y;
555
556
49.8k
    if (x & 7) {
557
29.9k
        if (x & 1)
558
15.3k
            return 0;
559
14.5k
        if (x & 2) {
560
8.89k
            *y = x >> 1;
561
8.89k
            return 1;
562
8.89k
        }
563
5.67k
        *y = x >> 2;
564
5.67k
        return 2;
565
14.5k
    }
566
19.9k
    k = 0;
567
19.9k
    if (!(x & 0xffff)) {
568
8.72k
        k = 16;
569
8.72k
        x >>= 16;
570
8.72k
    }
571
19.9k
    if (!(x & 0xff)) {
572
4.88k
        k += 8;
573
4.88k
        x >>= 8;
574
4.88k
    }
575
19.9k
    if (!(x & 0xf)) {
576
11.2k
        k += 4;
577
11.2k
        x >>= 4;
578
11.2k
    }
579
19.9k
    if (!(x & 0x3)) {
580
10.9k
        k += 2;
581
10.9k
        x >>= 2;
582
10.9k
    }
583
19.9k
    if (!(x & 1)) {
584
12.5k
        k++;
585
12.5k
        x >>= 1;
586
12.5k
        if (!x)
587
0
            return 32;
588
12.5k
    }
589
19.9k
    *y = x;
590
19.9k
    return k;
591
19.9k
}
592
593
/* convert a small nonnegative integer to a Bigint */
594
595
static Bigint *
596
i2b(int i)
597
175k
{
598
175k
    Bigint *b;
599
600
175k
    b = Balloc(1);
601
175k
    if (b == NULL)
602
0
        return NULL;
603
175k
    b->x[0] = i;
604
175k
    b->wds = 1;
605
175k
    return b;
606
175k
}
607
608
/* multiply two Bigints.  Returns a new Bigint, or NULL on failure.  Ignores
609
   the signs of a and b. */
610
611
static Bigint *
612
mult(Bigint *a, Bigint *b)
613
410k
{
614
410k
    Bigint *c;
615
410k
    int k, wa, wb, wc;
616
410k
    ULong *x, *xa, *xae, *xb, *xbe, *xc, *xc0;
617
410k
    ULong y;
618
410k
    ULLong carry, z;
619
620
410k
    if ((!a->x[0] && a->wds == 1) || (!b->x[0] && b->wds == 1)) {
621
4.45k
        c = Balloc(0);
622
4.45k
        if (c == NULL)
623
0
            return NULL;
624
4.45k
        c->wds = 1;
625
4.45k
        c->x[0] = 0;
626
4.45k
        return c;
627
4.45k
    }
628
629
406k
    if (a->wds < b->wds) {
630
201k
        c = a;
631
201k
        a = b;
632
201k
        b = c;
633
201k
    }
634
406k
    k = a->k;
635
406k
    wa = a->wds;
636
406k
    wb = b->wds;
637
406k
    wc = wa + wb;
638
406k
    if (wc > a->maxwds)
639
179k
        k++;
640
406k
    c = Balloc(k);
641
406k
    if (c == NULL)
642
0
        return NULL;
643
3.79M
    for(x = c->x, xa = x + wc; x < xa; x++)
644
3.38M
        *x = 0;
645
406k
    xa = a->x;
646
406k
    xae = xa + wa;
647
406k
    xb = b->x;
648
406k
    xbe = xb + wb;
649
406k
    xc0 = c->x;
650
1.29M
    for(; xb < xbe; xc0++) {
651
883k
        if ((y = *xb++)) {
652
879k
            x = xa;
653
879k
            xc = xc0;
654
879k
            carry = 0;
655
8.72M
            do {
656
8.72M
                z = *x++ * (ULLong)y + *xc + carry;
657
8.72M
                carry = z >> 32;
658
8.72M
                *xc++ = (ULong)(z & FFFFFFFF);
659
8.72M
            }
660
8.72M
            while(x < xae);
661
879k
            *xc = (ULong)carry;
662
879k
        }
663
883k
    }
664
696k
    for(xc0 = c->x, xc = xc0 + wc; wc > 0 && !*--xc; --wc) ;
665
406k
    c->wds = wc;
666
406k
    return c;
667
406k
}
668
669
#ifndef Py_USING_MEMORY_DEBUGGER
670
671
/* multiply the Bigint b by 5**k.  Returns a pointer to the result, or NULL on
672
   failure; if the returned pointer is distinct from b then the original
673
   Bigint b will have been Bfree'd.   Ignores the sign of b. */
674
675
static Bigint *
676
pow5mult(Bigint *b, int k)
677
129k
{
678
129k
    Bigint *b1, *p5, **p5s;
679
129k
    int i;
680
129k
    static const int p05[3] = { 5, 25, 125 };
681
682
    // For double-to-string conversion, the maximum value of k is limited by
683
    // DBL_MAX_10_EXP (308), the maximum decimal base-10 exponent for binary64.
684
    // For string-to-double conversion, the extreme case is constrained by our
685
    // hardcoded exponent limit before we underflow of -512, adjusted by
686
    // STRTOD_DIGLIM-DBL_DIG-1, giving a maximum of k=535.
687
129k
    assert(0 <= k && k < 1024);
688
689
129k
    if ((i = k & 3)) {
690
90.7k
        b = multadd(b, p05[i-1], 0);
691
90.7k
        if (b == NULL)
692
0
            return NULL;
693
90.7k
    }
694
695
129k
    if (!(k >>= 2))
696
10.1k
        return b;
697
119k
    PyInterpreterState *interp = _PyInterpreterState_GET();
698
119k
    p5s = interp->dtoa.p5s;
699
595k
    for(;;) {
700
595k
        assert(p5s != interp->dtoa.p5s + Bigint_Pow5size);
701
595k
        p5 = *p5s;
702
595k
        p5s++;
703
595k
        if (k & 1) {
704
349k
            b1 = mult(b, p5);
705
349k
            Bfree(b);
706
349k
            b = b1;
707
349k
            if (b == NULL)
708
0
                return NULL;
709
349k
        }
710
595k
        if (!(k >>= 1))
711
119k
            break;
712
595k
    }
713
119k
    return b;
714
119k
}
715
716
#else
717
718
/* Version of pow5mult that doesn't cache powers of 5. Provided for
719
   the benefit of memory debugging tools like Valgrind. */
720
721
static Bigint *
722
pow5mult(Bigint *b, int k)
723
{
724
    Bigint *b1, *p5, *p51;
725
    int i;
726
    static const int p05[3] = { 5, 25, 125 };
727
728
    if ((i = k & 3)) {
729
        b = multadd(b, p05[i-1], 0);
730
        if (b == NULL)
731
            return NULL;
732
    }
733
734
    if (!(k >>= 2))
735
        return b;
736
    p5 = i2b(625);
737
    if (p5 == NULL) {
738
        Bfree(b);
739
        return NULL;
740
    }
741
742
    for(;;) {
743
        if (k & 1) {
744
            b1 = mult(b, p5);
745
            Bfree(b);
746
            b = b1;
747
            if (b == NULL) {
748
                Bfree(p5);
749
                return NULL;
750
            }
751
        }
752
        if (!(k >>= 1))
753
            break;
754
        p51 = mult(p5, p5);
755
        Bfree(p5);
756
        p5 = p51;
757
        if (p5 == NULL) {
758
            Bfree(b);
759
            return NULL;
760
        }
761
    }
762
    Bfree(p5);
763
    return b;
764
}
765
766
#endif /* Py_USING_MEMORY_DEBUGGER */
767
768
/* shift a Bigint b left by k bits.  Return a pointer to the shifted result,
769
   or NULL on failure.  If the returned pointer is distinct from b then the
770
   original b will have been Bfree'd.   Ignores the sign of b. */
771
772
static Bigint *
773
lshift(Bigint *b, int k)
774
322k
{
775
322k
    int i, k1, n, n1;
776
322k
    Bigint *b1;
777
322k
    ULong *x, *x1, *xe, z;
778
779
322k
    if (!k || (!b->x[0] && b->wds == 1))
780
4.98k
        return b;
781
782
317k
    n = k >> 5;
783
317k
    k1 = b->k;
784
317k
    n1 = n + b->wds + 1;
785
793k
    for(i = b->maxwds; n1 > i; i <<= 1)
786
475k
        k1++;
787
317k
    b1 = Balloc(k1);
788
317k
    if (b1 == NULL) {
789
0
        Bfree(b);
790
0
        return NULL;
791
0
    }
792
317k
    x1 = b1->x;
793
1.89M
    for(i = 0; i < n; i++)
794
1.57M
        *x1++ = 0;
795
317k
    x = b->x;
796
317k
    xe = x + b->wds;
797
317k
    if (k &= 0x1f) {
798
314k
        k1 = 32 - k;
799
314k
        z = 0;
800
1.52M
        do {
801
1.52M
            *x1++ = *x << k | z;
802
1.52M
            z = *x++ >> k1;
803
1.52M
        }
804
1.52M
        while(x < xe);
805
314k
        if ((*x1 = z))
806
49.4k
            ++n1;
807
314k
    }
808
3.13k
    else do
809
6.25k
             *x1++ = *x++;
810
6.25k
        while(x < xe);
811
317k
    b1->wds = n1 - 1;
812
317k
    Bfree(b);
813
317k
    return b1;
814
317k
}
815
816
/* Do a three-way compare of a and b, returning -1 if a < b, 0 if a == b and
817
   1 if a > b.  Ignores signs of a and b. */
818
819
static int
820
cmp(Bigint *a, Bigint *b)
821
996k
{
822
996k
    ULong *xa, *xa0, *xb, *xb0;
823
996k
    int i, j;
824
825
996k
    i = a->wds;
826
996k
    j = b->wds;
827
#ifdef DEBUG
828
    if (i > 1 && !a->x[i-1])
829
        Bug("cmp called with a->x[a->wds-1] == 0");
830
    if (j > 1 && !b->x[j-1])
831
        Bug("cmp called with b->x[b->wds-1] == 0");
832
#endif
833
996k
    if (i -= j)
834
227k
        return i;
835
768k
    xa0 = a->x;
836
768k
    xa = xa0 + j;
837
768k
    xb0 = b->x;
838
768k
    xb = xb0 + j;
839
926k
    for(;;) {
840
926k
        if (*--xa != *--xb)
841
751k
            return *xa < *xb ? -1 : 1;
842
175k
        if (xa <= xa0)
843
16.8k
            break;
844
175k
    }
845
16.8k
    return 0;
846
768k
}
847
848
/* Take the difference of Bigints a and b, returning a new Bigint.  Returns
849
   NULL on failure.  The signs of a and b are ignored, but the sign of the
850
   result is set appropriately. */
851
852
static Bigint *
853
diff(Bigint *a, Bigint *b)
854
257k
{
855
257k
    Bigint *c;
856
257k
    int i, wa, wb;
857
257k
    ULong *xa, *xae, *xb, *xbe, *xc;
858
257k
    ULLong borrow, y;
859
860
257k
    i = cmp(a,b);
861
257k
    if (!i) {
862
2.56k
        c = Balloc(0);
863
2.56k
        if (c == NULL)
864
0
            return NULL;
865
2.56k
        c->wds = 1;
866
2.56k
        c->x[0] = 0;
867
2.56k
        return c;
868
2.56k
    }
869
255k
    if (i < 0) {
870
44.5k
        c = a;
871
44.5k
        a = b;
872
44.5k
        b = c;
873
44.5k
        i = 1;
874
44.5k
    }
875
210k
    else
876
210k
        i = 0;
877
255k
    c = Balloc(a->k);
878
255k
    if (c == NULL)
879
0
        return NULL;
880
255k
    c->sign = i;
881
255k
    wa = a->wds;
882
255k
    xa = a->x;
883
255k
    xae = xa + wa;
884
255k
    wb = b->wds;
885
255k
    xb = b->x;
886
255k
    xbe = xb + wb;
887
255k
    xc = c->x;
888
255k
    borrow = 0;
889
1.86M
    do {
890
1.86M
        y = (ULLong)*xa++ - *xb++ - borrow;
891
1.86M
        borrow = y >> 32 & (ULong)1;
892
1.86M
        *xc++ = (ULong)(y & FFFFFFFF);
893
1.86M
    }
894
1.86M
    while(xb < xbe);
895
494k
    while(xa < xae) {
896
239k
        y = *xa++ - borrow;
897
239k
        borrow = y >> 32 & (ULong)1;
898
239k
        *xc++ = (ULong)(y & FFFFFFFF);
899
239k
    }
900
375k
    while(!*--xc)
901
120k
        wa--;
902
255k
    c->wds = wa;
903
255k
    return c;
904
255k
}
905
906
/* Given a positive normal double x, return the difference between x and the
907
   next double up.  Doesn't give correct results for subnormals. */
908
909
static double
910
ulp(U *x)
911
42.8k
{
912
42.8k
    Long L;
913
42.8k
    U u;
914
915
42.8k
    L = (word0(x) & Exp_mask) - (P-1)*Exp_msk1;
916
42.8k
    word0(&u) = L;
917
42.8k
    word1(&u) = 0;
918
42.8k
    return dval(&u);
919
42.8k
}
920
921
/* Convert a Bigint to a double plus an exponent */
922
923
static double
924
b2d(Bigint *a, int *e)
925
80.4k
{
926
80.4k
    ULong *xa, *xa0, w, y, z;
927
80.4k
    int k;
928
80.4k
    U d;
929
930
80.4k
    xa0 = a->x;
931
80.4k
    xa = xa0 + a->wds;
932
80.4k
    y = *--xa;
933
#ifdef DEBUG
934
    if (!y) Bug("zero y in b2d");
935
#endif
936
80.4k
    k = hi0bits(y);
937
80.4k
    *e = 32 - k;
938
80.4k
    if (k < Ebits) {
939
19.8k
        word0(&d) = Exp_1 | y >> (Ebits - k);
940
19.8k
        w = xa > xa0 ? *--xa : 0;
941
19.8k
        word1(&d) = y << ((32-Ebits) + k) | w >> (Ebits - k);
942
19.8k
        goto ret_d;
943
19.8k
    }
944
60.5k
    z = xa > xa0 ? *--xa : 0;
945
60.5k
    if (k -= Ebits) {
946
56.1k
        word0(&d) = Exp_1 | y << k | z >> (32 - k);
947
56.1k
        y = xa > xa0 ? *--xa : 0;
948
56.1k
        word1(&d) = z << k | y >> (32 - k);
949
56.1k
    }
950
4.43k
    else {
951
4.43k
        word0(&d) = Exp_1 | y;
952
4.43k
        word1(&d) = z;
953
4.43k
    }
954
80.4k
  ret_d:
955
80.4k
    return dval(&d);
956
60.5k
}
957
958
/* Convert a scaled double to a Bigint plus an exponent.  Similar to d2b,
959
   except that it accepts the scale parameter used in _Py_dg_strtod (which
960
   should be either 0 or 2*P), and the normalization for the return value is
961
   different (see below).  On input, d should be finite and nonnegative, and d
962
   / 2**scale should be exactly representable as an IEEE 754 double.
963
964
   Returns a Bigint b and an integer e such that
965
966
     dval(d) / 2**scale = b * 2**e.
967
968
   Unlike d2b, b is not necessarily odd: b and e are normalized so
969
   that either 2**(P-1) <= b < 2**P and e >= Etiny, or b < 2**P
970
   and e == Etiny.  This applies equally to an input of 0.0: in that
971
   case the return values are b = 0 and e = Etiny.
972
973
   The above normalization ensures that for all possible inputs d,
974
   2**e gives ulp(d/2**scale).
975
976
   Returns NULL on failure.
977
*/
978
979
static Bigint *
980
sd2b(U *d, int scale, int *e)
981
102k
{
982
102k
    Bigint *b;
983
984
102k
    b = Balloc(1);
985
102k
    if (b == NULL)
986
0
        return NULL;
987
988
    /* First construct b and e assuming that scale == 0. */
989
102k
    b->wds = 2;
990
102k
    b->x[0] = word1(d);
991
102k
    b->x[1] = word0(d) & Frac_mask;
992
102k
    *e = Etiny - 1 + (int)((word0(d) & Exp_mask) >> Exp_shift);
993
102k
    if (*e < Etiny)
994
4.98k
        *e = Etiny;
995
97.4k
    else
996
97.4k
        b->x[1] |= Exp_msk1;
997
998
    /* Now adjust for scale, provided that b != 0. */
999
102k
    if (scale && (b->x[0] || b->x[1])) {
1000
28.8k
        *e -= scale;
1001
28.8k
        if (*e < Etiny) {
1002
25.4k
            scale = Etiny - *e;
1003
25.4k
            *e = Etiny;
1004
            /* We can't shift more than P-1 bits without shifting out a 1. */
1005
25.4k
            assert(0 < scale && scale <= P - 1);
1006
25.4k
            if (scale >= 32) {
1007
                /* The bits shifted out should all be zero. */
1008
12.1k
                assert(b->x[0] == 0);
1009
12.1k
                b->x[0] = b->x[1];
1010
12.1k
                b->x[1] = 0;
1011
12.1k
                scale -= 32;
1012
12.1k
            }
1013
25.4k
            if (scale) {
1014
                /* The bits shifted out should all be zero. */
1015
23.9k
                assert(b->x[0] << (32 - scale) == 0);
1016
23.9k
                b->x[0] = (b->x[0] >> scale) | (b->x[1] << (32 - scale));
1017
23.9k
                b->x[1] >>= scale;
1018
23.9k
            }
1019
25.4k
        }
1020
28.8k
    }
1021
    /* Ensure b is normalized. */
1022
102k
    if (!b->x[1])
1023
20.9k
        b->wds = 1;
1024
1025
102k
    return b;
1026
102k
}
1027
1028
/* Convert a double to a Bigint plus an exponent.  Return NULL on failure.
1029
1030
   Given a finite nonzero double d, return an odd Bigint b and exponent *e
1031
   such that fabs(d) = b * 2**e.  On return, *bbits gives the number of
1032
   significant bits of b; that is, 2**(*bbits-1) <= b < 2**(*bbits).
1033
1034
   If d is zero, then b == 0, *e == -1010, *bbits = 0.
1035
 */
1036
1037
static Bigint *
1038
d2b(U *d, int *e, int *bits)
1039
49.8k
{
1040
49.8k
    Bigint *b;
1041
49.8k
    int de, k;
1042
49.8k
    ULong *x, y, z;
1043
49.8k
    int i;
1044
1045
49.8k
    b = Balloc(1);
1046
49.8k
    if (b == NULL)
1047
0
        return NULL;
1048
49.8k
    x = b->x;
1049
1050
49.8k
    z = word0(d) & Frac_mask;
1051
49.8k
    word0(d) &= 0x7fffffff;   /* clear sign bit, which we ignore */
1052
49.8k
    if ((de = (int)(word0(d) >> Exp_shift)))
1053
45.7k
        z |= Exp_msk1;
1054
49.8k
    if ((y = word1(d))) {
1055
37.6k
        if ((k = lo0bits(&y))) {
1056
23.2k
            x[0] = y | z << (32 - k);
1057
23.2k
            z >>= k;
1058
23.2k
        }
1059
14.3k
        else
1060
14.3k
            x[0] = y;
1061
37.6k
        i =
1062
37.6k
            b->wds = (x[1] = z) ? 2 : 1;
1063
37.6k
    }
1064
12.2k
    else {
1065
12.2k
        k = lo0bits(&z);
1066
12.2k
        x[0] = z;
1067
12.2k
        i =
1068
12.2k
            b->wds = 1;
1069
12.2k
        k += 32;
1070
12.2k
    }
1071
49.8k
    if (de) {
1072
45.7k
        *e = de - Bias - (P-1) + k;
1073
45.7k
        *bits = P - k;
1074
45.7k
    }
1075
4.12k
    else {
1076
4.12k
        *e = de - Bias - (P-1) + 1 + k;
1077
4.12k
        *bits = 32*i - hi0bits(x[i-1]);
1078
4.12k
    }
1079
49.8k
    return b;
1080
49.8k
}
1081
1082
/* Compute the ratio of two Bigints, as a double.  The result may have an
1083
   error of up to 2.5 ulps. */
1084
1085
static double
1086
ratio(Bigint *a, Bigint *b)
1087
40.2k
{
1088
40.2k
    U da, db;
1089
40.2k
    int k, ka, kb;
1090
1091
40.2k
    dval(&da) = b2d(a, &ka);
1092
40.2k
    dval(&db) = b2d(b, &kb);
1093
40.2k
    k = ka - kb + 32*(a->wds - b->wds);
1094
40.2k
    if (k > 0)
1095
23.1k
        word0(&da) += k*Exp_msk1;
1096
17.0k
    else {
1097
17.0k
        k = -k;
1098
17.0k
        word0(&db) += k*Exp_msk1;
1099
17.0k
    }
1100
40.2k
    return dval(&da) / dval(&db);
1101
40.2k
}
1102
1103
static const double
1104
tens[] = {
1105
    1e0, 1e1, 1e2, 1e3, 1e4, 1e5, 1e6, 1e7, 1e8, 1e9,
1106
    1e10, 1e11, 1e12, 1e13, 1e14, 1e15, 1e16, 1e17, 1e18, 1e19,
1107
    1e20, 1e21, 1e22
1108
};
1109
1110
static const double
1111
bigtens[] = { 1e16, 1e32, 1e64, 1e128, 1e256 };
1112
static const double tinytens[] = { 1e-16, 1e-32, 1e-64, 1e-128,
1113
                                   9007199254740992.*9007199254740992.e-256
1114
                                   /* = 2^106 * 1e-256 */
1115
};
1116
/* The factor of 2^53 in tinytens[4] helps us avoid setting the underflow */
1117
/* flag unnecessarily.  It leads to a song and dance at the end of strtod. */
1118
22.4k
#define Scale_Bit 0x10
1119
23.7k
#define n_bigtens 5
1120
1121
#define ULbits 32
1122
#define kshift 5
1123
43.4k
#define kmask 31
1124
1125
1126
static int
1127
dshift(Bigint *b, int p2)
1128
43.4k
{
1129
43.4k
    int rv = hi0bits(b->x[b->wds-1]) - 4;
1130
43.4k
    if (p2 > 0)
1131
24.4k
        rv -= p2;
1132
43.4k
    return rv & kmask;
1133
43.4k
}
1134
1135
/* special case of Bigint division.  The quotient is always in the range 0 <=
1136
   quotient < 10, and on entry the divisor S is normalized so that its top 4
1137
   bits (28--31) are zero and bit 27 is set. */
1138
1139
static int
1140
quorem(Bigint *b, Bigint *S)
1141
286k
{
1142
286k
    int n;
1143
286k
    ULong *bx, *bxe, q, *sx, *sxe;
1144
286k
    ULLong borrow, carry, y, ys;
1145
1146
286k
    n = S->wds;
1147
#ifdef DEBUG
1148
    /*debug*/ if (b->wds > n)
1149
        /*debug*/       Bug("oversize b in quorem");
1150
#endif
1151
286k
    if (b->wds < n)
1152
6.23k
        return 0;
1153
280k
    sx = S->x;
1154
280k
    sxe = sx + --n;
1155
280k
    bx = b->x;
1156
280k
    bxe = bx + n;
1157
280k
    q = *bxe / (*sxe + 1);      /* ensure q <= true quotient */
1158
#ifdef DEBUG
1159
    /*debug*/ if (q > 9)
1160
        /*debug*/       Bug("oversized quotient in quorem");
1161
#endif
1162
280k
    if (q) {
1163
232k
        borrow = 0;
1164
232k
        carry = 0;
1165
1.34M
        do {
1166
1.34M
            ys = *sx++ * (ULLong)q + carry;
1167
1.34M
            carry = ys >> 32;
1168
1.34M
            y = *bx - (ys & FFFFFFFF) - borrow;
1169
1.34M
            borrow = y >> 32 & (ULong)1;
1170
1.34M
            *bx++ = (ULong)(y & FFFFFFFF);
1171
1.34M
        }
1172
1.34M
        while(sx <= sxe);
1173
232k
        if (!*bxe) {
1174
790
            bx = b->x;
1175
790
            while(--bxe > bx && !*bxe)
1176
0
                --n;
1177
790
            b->wds = n;
1178
790
        }
1179
232k
    }
1180
280k
    if (cmp(b, S) >= 0) {
1181
18.8k
        q++;
1182
18.8k
        borrow = 0;
1183
18.8k
        carry = 0;
1184
18.8k
        bx = b->x;
1185
18.8k
        sx = S->x;
1186
115k
        do {
1187
115k
            ys = *sx++ + carry;
1188
115k
            carry = ys >> 32;
1189
115k
            y = *bx - (ys & FFFFFFFF) - borrow;
1190
115k
            borrow = y >> 32 & (ULong)1;
1191
115k
            *bx++ = (ULong)(y & FFFFFFFF);
1192
115k
        }
1193
115k
        while(sx <= sxe);
1194
18.8k
        bx = b->x;
1195
18.8k
        bxe = bx + n;
1196
18.8k
        if (!*bxe) {
1197
19.5k
            while(--bxe > bx && !*bxe)
1198
1.46k
                --n;
1199
18.1k
            b->wds = n;
1200
18.1k
        }
1201
18.8k
    }
1202
280k
    return q;
1203
286k
}
1204
1205
/* sulp(x) is a version of ulp(x) that takes bc.scale into account.
1206
1207
   Assuming that x is finite and nonnegative (positive zero is fine
1208
   here) and x / 2^bc.scale is exactly representable as a double,
1209
   sulp(x) is equivalent to 2^bc.scale * ulp(x / 2^bc.scale). */
1210
1211
static double
1212
sulp(U *x, BCinfo *bc)
1213
2.98k
{
1214
2.98k
    U u;
1215
1216
2.98k
    if (bc->scale && 2*P + 1 > (int)((word0(x) & Exp_mask) >> Exp_shift)) {
1217
        /* rv/2^bc->scale is subnormal */
1218
399
        word0(&u) = (P+2)*Exp_msk1;
1219
399
        word1(&u) = 0;
1220
399
        return u.d;
1221
399
    }
1222
2.58k
    else {
1223
2.58k
        assert(word0(x) || word1(x)); /* x != 0.0 */
1224
2.58k
        return ulp(x);
1225
2.58k
    }
1226
2.98k
}
1227
1228
/* The bigcomp function handles some hard cases for strtod, for inputs
1229
   with more than STRTOD_DIGLIM digits.  It's called once an initial
1230
   estimate for the double corresponding to the input string has
1231
   already been obtained by the code in _Py_dg_strtod.
1232
1233
   The bigcomp function is only called after _Py_dg_strtod has found a
1234
   double value rv such that either rv or rv + 1ulp represents the
1235
   correctly rounded value corresponding to the original string.  It
1236
   determines which of these two values is the correct one by
1237
   computing the decimal digits of rv + 0.5ulp and comparing them with
1238
   the corresponding digits of s0.
1239
1240
   In the following, write dv for the absolute value of the number represented
1241
   by the input string.
1242
1243
   Inputs:
1244
1245
     s0 points to the first significant digit of the input string.
1246
1247
     rv is a (possibly scaled) estimate for the closest double value to the
1248
        value represented by the original input to _Py_dg_strtod.  If
1249
        bc->scale is nonzero, then rv/2^(bc->scale) is the approximation to
1250
        the input value.
1251
1252
     bc is a struct containing information gathered during the parsing and
1253
        estimation steps of _Py_dg_strtod.  Description of fields follows:
1254
1255
        bc->e0 gives the exponent of the input value, such that dv = (integer
1256
           given by the bd->nd digits of s0) * 10**e0
1257
1258
        bc->nd gives the total number of significant digits of s0.  It will
1259
           be at least 1.
1260
1261
        bc->nd0 gives the number of significant digits of s0 before the
1262
           decimal separator.  If there's no decimal separator, bc->nd0 ==
1263
           bc->nd.
1264
1265
        bc->scale is the value used to scale rv to avoid doing arithmetic with
1266
           subnormal values.  It's either 0 or 2*P (=106).
1267
1268
   Outputs:
1269
1270
     On successful exit, rv/2^(bc->scale) is the closest double to dv.
1271
1272
     Returns 0 on success, -1 on failure (e.g., due to a failed malloc call). */
1273
1274
static int
1275
bigcomp(U *rv, const char *s0, BCinfo *bc)
1276
7.03k
{
1277
7.03k
    Bigint *b, *d;
1278
7.03k
    int b2, d2, dd, i, nd, nd0, odd, p2, p5;
1279
1280
7.03k
    nd = bc->nd;
1281
7.03k
    nd0 = bc->nd0;
1282
7.03k
    p5 = nd + bc->e0;
1283
7.03k
    b = sd2b(rv, bc->scale, &p2);
1284
7.03k
    if (b == NULL)
1285
0
        return -1;
1286
1287
    /* record whether the lsb of rv/2^(bc->scale) is odd:  in the exact halfway
1288
       case, this is used for round to even. */
1289
7.03k
    odd = b->x[0] & 1;
1290
1291
    /* left shift b by 1 bit and or a 1 into the least significant bit;
1292
       this gives us b * 2**p2 = rv/2^(bc->scale) + 0.5 ulp. */
1293
7.03k
    b = lshift(b, 1);
1294
7.03k
    if (b == NULL)
1295
0
        return -1;
1296
7.03k
    b->x[0] |= 1;
1297
7.03k
    p2--;
1298
1299
7.03k
    p2 -= p5;
1300
7.03k
    d = i2b(1);
1301
7.03k
    if (d == NULL) {
1302
0
        Bfree(b);
1303
0
        return -1;
1304
0
    }
1305
    /* Arrange for convenient computation of quotients:
1306
     * shift left if necessary so divisor has 4 leading 0 bits.
1307
     */
1308
7.03k
    if (p5 > 0) {
1309
5.02k
        d = pow5mult(d, p5);
1310
5.02k
        if (d == NULL) {
1311
0
            Bfree(b);
1312
0
            return -1;
1313
0
        }
1314
5.02k
    }
1315
2.01k
    else if (p5 < 0) {
1316
1.43k
        b = pow5mult(b, -p5);
1317
1.43k
        if (b == NULL) {
1318
0
            Bfree(d);
1319
0
            return -1;
1320
0
        }
1321
1.43k
    }
1322
7.03k
    if (p2 > 0) {
1323
3.75k
        b2 = p2;
1324
3.75k
        d2 = 0;
1325
3.75k
    }
1326
3.27k
    else {
1327
3.27k
        b2 = 0;
1328
3.27k
        d2 = -p2;
1329
3.27k
    }
1330
7.03k
    i = dshift(d, d2);
1331
7.03k
    if ((b2 += i) > 0) {
1332
6.88k
        b = lshift(b, b2);
1333
6.88k
        if (b == NULL) {
1334
0
            Bfree(d);
1335
0
            return -1;
1336
0
        }
1337
6.88k
    }
1338
7.03k
    if ((d2 += i) > 0) {
1339
6.41k
        d = lshift(d, d2);
1340
6.41k
        if (d == NULL) {
1341
0
            Bfree(b);
1342
0
            return -1;
1343
0
        }
1344
6.41k
    }
1345
1346
    /* Compare s0 with b/d: set dd to -1, 0, or 1 according as s0 < b/d, s0 ==
1347
     * b/d, or s0 > b/d.  Here the digits of s0 are thought of as representing
1348
     * a number in the range [0.1, 1). */
1349
7.03k
    if (cmp(b, d) >= 0)
1350
        /* b/d >= 1 */
1351
900
        dd = -1;
1352
6.13k
    else {
1353
6.13k
        i = 0;
1354
123k
        for(;;) {
1355
123k
            b = multadd(b, 10, 0);
1356
123k
            if (b == NULL) {
1357
0
                Bfree(d);
1358
0
                return -1;
1359
0
            }
1360
123k
            dd = s0[i < nd0 ? i : i+1] - '0' - quorem(b, d);
1361
123k
            i++;
1362
1363
123k
            if (dd)
1364
4.82k
                break;
1365
119k
            if (!b->x[0] && b->wds == 1) {
1366
                /* b/d == 0 */
1367
839
                dd = i < nd;
1368
839
                break;
1369
839
            }
1370
118k
            if (!(i < nd)) {
1371
                /* b/d != 0, but digits of s0 exhausted */
1372
474
                dd = -1;
1373
474
                break;
1374
474
            }
1375
118k
        }
1376
6.13k
    }
1377
7.03k
    Bfree(b);
1378
7.03k
    Bfree(d);
1379
7.03k
    if (dd > 0 || (dd == 0 && odd))
1380
1.55k
        dval(rv) += sulp(rv, bc);
1381
7.03k
    return 0;
1382
7.03k
}
1383
1384
1385
double
1386
_Py_dg_strtod(const char *s00, char **se)
1387
636k
{
1388
636k
    int bb2, bb5, bbe, bd2, bd5, bs2, c, dsign, e, e1, error;
1389
636k
    int esign, i, j, k, lz, nd, nd0, odd, sign;
1390
636k
    const char *s, *s0, *s1;
1391
636k
    double aadj, aadj1;
1392
636k
    U aadj2, adj, rv, rv0;
1393
636k
    ULong y, z, abs_exp;
1394
636k
    Long L;
1395
636k
    BCinfo bc;
1396
636k
    Bigint *bb = NULL, *bd = NULL, *bd0 = NULL, *bs = NULL, *delta = NULL;
1397
636k
    size_t ndigits, fraclen;
1398
636k
    double result;
1399
1400
636k
    dval(&rv) = 0.;
1401
1402
    /* Start parsing. */
1403
636k
    c = *(s = s00);
1404
1405
    /* Parse optional sign, if present. */
1406
636k
    sign = 0;
1407
636k
    switch (c) {
1408
496k
    case '-':
1409
496k
        sign = 1;
1410
496k
        _Py_FALLTHROUGH;
1411
496k
    case '+':
1412
496k
        c = *++s;
1413
636k
    }
1414
1415
    /* Skip leading zeros: lz is true iff there were leading zeros. */
1416
636k
    s1 = s;
1417
655k
    while (c == '0')
1418
19.1k
        c = *++s;
1419
636k
    lz = s != s1;
1420
1421
    /* Point s0 at the first nonzero digit (if any).  fraclen will be the
1422
       number of digits between the decimal point and the end of the
1423
       digit string.  ndigits will be the total number of digits ignoring
1424
       leading zeros. */
1425
636k
    s0 = s1 = s;
1426
4.31M
    while ('0' <= c && c <= '9')
1427
3.67M
        c = *++s;
1428
636k
    ndigits = s - s1;
1429
636k
    fraclen = 0;
1430
1431
    /* Parse decimal point and following digits. */
1432
636k
    if (c == '.') {
1433
69.1k
        c = *++s;
1434
69.1k
        if (!ndigits) {
1435
20.1k
            s1 = s;
1436
677k
            while (c == '0')
1437
657k
                c = *++s;
1438
20.1k
            lz = lz || s != s1;
1439
20.1k
            fraclen += (s - s1);
1440
20.1k
            s0 = s;
1441
20.1k
        }
1442
69.1k
        s1 = s;
1443
11.1M
        while ('0' <= c && c <= '9')
1444
11.1M
            c = *++s;
1445
69.1k
        ndigits += s - s1;
1446
69.1k
        fraclen += s - s1;
1447
69.1k
    }
1448
1449
    /* Now lz is true if and only if there were leading zero digits, and
1450
       ndigits gives the total number of digits ignoring leading zeros.  A
1451
       valid input must have at least one digit. */
1452
636k
    if (!ndigits && !lz) {
1453
23
        if (se)
1454
23
            *se = (char *)s00;
1455
23
        goto parse_error;
1456
23
    }
1457
1458
    /* Range check ndigits and fraclen to make sure that they, and values
1459
       computed with them, can safely fit in an int. */
1460
636k
    if (ndigits > MAX_DIGITS || fraclen > MAX_DIGITS) {
1461
0
        if (se)
1462
0
            *se = (char *)s00;
1463
0
        goto parse_error;
1464
0
    }
1465
636k
    nd = (int)ndigits;
1466
636k
    nd0 = (int)ndigits - (int)fraclen;
1467
1468
    /* Parse exponent. */
1469
636k
    e = 0;
1470
636k
    if (c == 'e' || c == 'E') {
1471
568k
        s00 = s;
1472
568k
        c = *++s;
1473
1474
        /* Exponent sign. */
1475
568k
        esign = 0;
1476
568k
        switch (c) {
1477
27.2k
        case '-':
1478
27.2k
            esign = 1;
1479
27.2k
            _Py_FALLTHROUGH;
1480
44.3k
        case '+':
1481
44.3k
            c = *++s;
1482
568k
        }
1483
1484
        /* Skip zeros.  lz is true iff there are leading zeros. */
1485
568k
        s1 = s;
1486
573k
        while (c == '0')
1487
4.97k
            c = *++s;
1488
568k
        lz = s != s1;
1489
1490
        /* Get absolute value of the exponent. */
1491
568k
        s1 = s;
1492
568k
        abs_exp = 0;
1493
2.37M
        while ('0' <= c && c <= '9') {
1494
1.80M
            abs_exp = 10*abs_exp + (c - '0');
1495
1.80M
            c = *++s;
1496
1.80M
        }
1497
1498
        /* abs_exp will be correct modulo 2**32.  But 10**9 < 2**32, so if
1499
           there are at most 9 significant exponent digits then overflow is
1500
           impossible. */
1501
568k
        if (s - s1 > 9 || abs_exp > MAX_ABS_EXP)
1502
4.92k
            e = (int)MAX_ABS_EXP;
1503
563k
        else
1504
563k
            e = (int)abs_exp;
1505
568k
        if (esign)
1506
27.2k
            e = -e;
1507
1508
        /* A valid exponent must have at least one digit. */
1509
568k
        if (s == s1 && !lz)
1510
0
            s = s00;
1511
568k
    }
1512
1513
    /* Adjust exponent to take into account position of the point. */
1514
636k
    e -= nd - nd0;
1515
636k
    if (nd0 <= 0)
1516
24.0k
        nd0 = nd;
1517
1518
    /* Finished parsing.  Set se to indicate how far we parsed */
1519
636k
    if (se)
1520
636k
        *se = (char *)s;
1521
1522
    /* If all digits were zero, exit with return value +-0.0.  Otherwise,
1523
       strip trailing zeros: scan back until we hit a nonzero digit. */
1524
636k
    if (!nd)
1525
10.1k
        goto ret;
1526
2.61M
    for (i = nd; i > 0; ) {
1527
2.61M
        --i;
1528
2.61M
        if (s0[i < nd0 ? i : i+1] != '0') {
1529
626k
            ++i;
1530
626k
            break;
1531
626k
        }
1532
2.61M
    }
1533
626k
    e += nd - i;
1534
626k
    nd = i;
1535
626k
    if (nd0 > nd)
1536
11.0k
        nd0 = nd;
1537
1538
    /* Summary of parsing results.  After parsing, and dealing with zero
1539
     * inputs, we have values s0, nd0, nd, e, sign, where:
1540
     *
1541
     *  - s0 points to the first significant digit of the input string
1542
     *
1543
     *  - nd is the total number of significant digits (here, and
1544
     *    below, 'significant digits' means the set of digits of the
1545
     *    significand of the input that remain after ignoring leading
1546
     *    and trailing zeros).
1547
     *
1548
     *  - nd0 indicates the position of the decimal point, if present; it
1549
     *    satisfies 1 <= nd0 <= nd.  The nd significant digits are in
1550
     *    s0[0:nd0] and s0[nd0+1:nd+1] using the usual Python half-open slice
1551
     *    notation.  (If nd0 < nd, then s0[nd0] contains a '.'  character; if
1552
     *    nd0 == nd, then s0[nd0] could be any non-digit character.)
1553
     *
1554
     *  - e is the adjusted exponent: the absolute value of the number
1555
     *    represented by the original input string is n * 10**e, where
1556
     *    n is the integer represented by the concatenation of
1557
     *    s0[0:nd0] and s0[nd0+1:nd+1]
1558
     *
1559
     *  - sign gives the sign of the input:  1 for negative, 0 for positive
1560
     *
1561
     *  - the first and last significant digits are nonzero
1562
     */
1563
1564
    /* put first DBL_DIG+1 digits into integer y and z.
1565
     *
1566
     *  - y contains the value represented by the first min(9, nd)
1567
     *    significant digits
1568
     *
1569
     *  - if nd > 9, z contains the value represented by significant digits
1570
     *    with indices in [9, min(16, nd)).  So y * 10**(min(16, nd) - 9) + z
1571
     *    gives the value represented by the first min(16, nd) sig. digits.
1572
     */
1573
1574
626k
    bc.e0 = e1 = e;
1575
626k
    y = z = 0;
1576
1.90M
    for (i = 0; i < nd; i++) {
1577
1.30M
        if (i < 9)
1578
1.04M
            y = 10*y + s0[i < nd0 ? i : i+1] - '0';
1579
257k
        else if (i < DBL_DIG+1)
1580
234k
            z = 10*z + s0[i < nd0 ? i : i+1] - '0';
1581
23.3k
        else
1582
23.3k
            break;
1583
1.30M
    }
1584
1585
626k
    k = nd < DBL_DIG + 1 ? nd : DBL_DIG + 1;
1586
626k
    dval(&rv) = y;
1587
626k
    if (k > 9) {
1588
40.1k
        dval(&rv) = tens[k - 9] * dval(&rv) + z;
1589
40.1k
    }
1590
626k
    if (nd <= DBL_DIG
1591
626k
        && Flt_Rounds == 1
1592
626k
        ) {
1593
595k
        if (!e)
1594
12.6k
            goto ret;
1595
582k
        if (e > 0) {
1596
537k
            if (e <= Ten_pmax) {
1597
27.5k
                dval(&rv) *= tens[e];
1598
27.5k
                goto ret;
1599
27.5k
            }
1600
509k
            i = DBL_DIG - nd;
1601
509k
            if (e <= Ten_pmax + i) {
1602
                /* A fancier test would sometimes let us do
1603
                 * this for larger i values.
1604
                 */
1605
5.28k
                e -= i;
1606
5.28k
                dval(&rv) *= tens[i];
1607
5.28k
                dval(&rv) *= tens[e];
1608
5.28k
                goto ret;
1609
5.28k
            }
1610
509k
        }
1611
45.0k
        else if (e >= -Ten_pmax) {
1612
26.1k
            dval(&rv) /= tens[-e];
1613
26.1k
            goto ret;
1614
26.1k
        }
1615
582k
    }
1616
554k
    e1 += nd - k;
1617
1618
554k
    bc.scale = 0;
1619
1620
    /* Get starting approximation = rv * 10**e1 */
1621
1622
554k
    if (e1 > 0) {
1623
520k
        if ((i = e1 & 15))
1624
508k
            dval(&rv) *= tens[i];
1625
520k
        if (e1 &= ~15) {
1626
512k
            if (e1 > DBL_MAX_10_EXP)
1627
476k
                goto ovfl;
1628
35.5k
            e1 >>= 4;
1629
104k
            for(j = 0; e1 > 1; j++, e1 >>= 1)
1630
69.4k
                if (e1 & 1)
1631
32.7k
                    dval(&rv) *= bigtens[j];
1632
            /* The last multiplication could overflow. */
1633
35.5k
            word0(&rv) -= P*Exp_msk1;
1634
35.5k
            dval(&rv) *= bigtens[j];
1635
35.5k
            if ((z = word0(&rv) & Exp_mask)
1636
35.5k
                > Exp_msk1*(DBL_MAX_EXP+Bias-P))
1637
937
                goto ovfl;
1638
34.5k
            if (z > Exp_msk1*(DBL_MAX_EXP+Bias-1-P)) {
1639
                /* set to largest number */
1640
                /* (Can't trust DBL_MAX) */
1641
577
                word0(&rv) = Big0;
1642
577
                word1(&rv) = Big1;
1643
577
            }
1644
34.0k
            else
1645
34.0k
                word0(&rv) += P*Exp_msk1;
1646
34.5k
        }
1647
520k
    }
1648
34.4k
    else if (e1 < 0) {
1649
        /* The input decimal value lies in [10**e1, 10**(e1+16)).
1650
1651
           If e1 <= -512, underflow immediately.
1652
           If e1 <= -256, set bc.scale to 2*P.
1653
1654
           So for input value < 1e-256, bc.scale is always set;
1655
           for input value >= 1e-240, bc.scale is never set.
1656
           For input values in [1e-256, 1e-240), bc.scale may or may
1657
           not be set. */
1658
1659
31.3k
        e1 = -e1;
1660
31.3k
        if ((i = e1 & 15))
1661
27.5k
            dval(&rv) /= tens[i];
1662
31.3k
        if (e1 >>= 4) {
1663
23.7k
            if (e1 >= 1 << n_bigtens)
1664
1.30k
                goto undfl;
1665
22.4k
            if (e1 & Scale_Bit)
1666
18.0k
                bc.scale = 2*P;
1667
119k
            for(j = 0; e1 > 0; j++, e1 >>= 1)
1668
97.2k
                if (e1 & 1)
1669
56.7k
                    dval(&rv) *= tinytens[j];
1670
22.4k
            if (bc.scale && (j = 2*P + 1 - ((word0(&rv) & Exp_mask)
1671
18.0k
                                            >> Exp_shift)) > 0) {
1672
                /* scaled rv is denormal; clear j low bits */
1673
16.5k
                if (j >= 32) {
1674
9.43k
                    word1(&rv) = 0;
1675
9.43k
                    if (j >= 53)
1676
5.24k
                        word0(&rv) = (P+2)*Exp_msk1;
1677
4.19k
                    else
1678
4.19k
                        word0(&rv) &= 0xffffffff << (j-32);
1679
9.43k
                }
1680
7.16k
                else
1681
7.16k
                    word1(&rv) &= 0xffffffff << j;
1682
16.5k
            }
1683
22.4k
            if (!dval(&rv))
1684
0
                goto undfl;
1685
22.4k
        }
1686
31.3k
    }
1687
1688
    /* Now the hard part -- adjusting rv to the correct value.*/
1689
1690
    /* Put digits into bd: true value = bd * 10^e */
1691
1692
75.3k
    bc.nd = nd;
1693
75.3k
    bc.nd0 = nd0;       /* Only needed if nd > STRTOD_DIGLIM, but done here */
1694
                        /* to silence an erroneous warning about bc.nd0 */
1695
                        /* possibly not being initialized. */
1696
75.3k
    if (nd > STRTOD_DIGLIM) {
1697
        /* ASSERT(STRTOD_DIGLIM >= 18); 18 == one more than the */
1698
        /* minimum number of decimal digits to distinguish double values */
1699
        /* in IEEE arithmetic. */
1700
1701
        /* Truncate input to 18 significant digits, then discard any trailing
1702
           zeros on the result by updating nd, nd0, e and y suitably. (There's
1703
           no need to update z; it's not reused beyond this point.) */
1704
60.6k
        for (i = 18; i > 0; ) {
1705
            /* scan back until we hit a nonzero digit.  significant digit 'i'
1706
            is s0[i] if i < nd0, s0[i+1] if i >= nd0. */
1707
60.6k
            --i;
1708
60.6k
            if (s0[i < nd0 ? i : i+1] != '0') {
1709
9.97k
                ++i;
1710
9.97k
                break;
1711
9.97k
            }
1712
60.6k
        }
1713
9.97k
        e += nd - i;
1714
9.97k
        nd = i;
1715
9.97k
        if (nd0 > nd)
1716
7.06k
            nd0 = nd;
1717
9.97k
        if (nd < 9) { /* must recompute y */
1718
3.84k
            y = 0;
1719
20.5k
            for(i = 0; i < nd0; ++i)
1720
16.7k
                y = 10*y + s0[i] - '0';
1721
10.2k
            for(; i < nd; ++i)
1722
6.43k
                y = 10*y + s0[i+1] - '0';
1723
3.84k
        }
1724
9.97k
    }
1725
75.3k
    bd0 = s2b(s0, nd0, nd, y);
1726
75.3k
    if (bd0 == NULL)
1727
0
        goto failed_malloc;
1728
1729
    /* Notation for the comments below.  Write:
1730
1731
         - dv for the absolute value of the number represented by the original
1732
           decimal input string.
1733
1734
         - if we've truncated dv, write tdv for the truncated value.
1735
           Otherwise, set tdv == dv.
1736
1737
         - srv for the quantity rv/2^bc.scale; so srv is the current binary
1738
           approximation to tdv (and dv).  It should be exactly representable
1739
           in an IEEE 754 double.
1740
    */
1741
1742
95.3k
    for(;;) {
1743
1744
        /* This is the main correction loop for _Py_dg_strtod.
1745
1746
           We've got a decimal value tdv, and a floating-point approximation
1747
           srv=rv/2^bc.scale to tdv.  The aim is to determine whether srv is
1748
           close enough (i.e., within 0.5 ulps) to tdv, and to compute a new
1749
           approximation if not.
1750
1751
           To determine whether srv is close enough to tdv, compute integers
1752
           bd, bb and bs proportional to tdv, srv and 0.5 ulp(srv)
1753
           respectively, and then use integer arithmetic to determine whether
1754
           |tdv - srv| is less than, equal to, or greater than 0.5 ulp(srv).
1755
        */
1756
1757
95.3k
        bd = Balloc(bd0->k);
1758
95.3k
        if (bd == NULL) {
1759
0
            goto failed_malloc;
1760
0
        }
1761
95.3k
        Bcopy(bd, bd0);
1762
95.3k
        bb = sd2b(&rv, bc.scale, &bbe);   /* srv = bb * 2^bbe */
1763
95.3k
        if (bb == NULL) {
1764
0
            goto failed_malloc;
1765
0
        }
1766
        /* Record whether lsb of bb is odd, in case we need this
1767
           for the round-to-even step later. */
1768
95.3k
        odd = bb->x[0] & 1;
1769
1770
        /* tdv = bd * 10**e;  srv = bb * 2**bbe */
1771
95.3k
        bs = i2b(1);
1772
95.3k
        if (bs == NULL) {
1773
0
            goto failed_malloc;
1774
0
        }
1775
1776
95.3k
        if (e >= 0) {
1777
46.9k
            bb2 = bb5 = 0;
1778
46.9k
            bd2 = bd5 = e;
1779
46.9k
        }
1780
48.4k
        else {
1781
48.4k
            bb2 = bb5 = -e;
1782
48.4k
            bd2 = bd5 = 0;
1783
48.4k
        }
1784
95.3k
        if (bbe >= 0)
1785
48.8k
            bb2 += bbe;
1786
46.5k
        else
1787
46.5k
            bd2 -= bbe;
1788
95.3k
        bs2 = bb2;
1789
95.3k
        bb2++;
1790
95.3k
        bd2++;
1791
1792
        /* At this stage bd5 - bb5 == e == bd2 - bb2 + bbe, bb2 - bs2 == 1,
1793
           and bs == 1, so:
1794
1795
              tdv == bd * 10**e = bd * 2**(bbe - bb2 + bd2) * 5**(bd5 - bb5)
1796
              srv == bb * 2**bbe = bb * 2**(bbe - bb2 + bb2)
1797
              0.5 ulp(srv) == 2**(bbe-1) = bs * 2**(bbe - bb2 + bs2)
1798
1799
           It follows that:
1800
1801
              M * tdv = bd * 2**bd2 * 5**bd5
1802
              M * srv = bb * 2**bb2 * 5**bb5
1803
              M * 0.5 ulp(srv) = bs * 2**bs2 * 5**bb5
1804
1805
           for some constant M.  (Actually, M == 2**(bb2 - bbe) * 5**bb5, but
1806
           this fact is not needed below.)
1807
        */
1808
1809
        /* Remove factor of 2**i, where i = min(bb2, bd2, bs2). */
1810
95.3k
        i = bb2 < bd2 ? bb2 : bd2;
1811
95.3k
        if (i > bs2)
1812
46.2k
            i = bs2;
1813
95.3k
        if (i > 0) {
1814
94.6k
            bb2 -= i;
1815
94.6k
            bd2 -= i;
1816
94.6k
            bs2 -= i;
1817
94.6k
        }
1818
1819
        /* Scale bb, bd, bs by the appropriate powers of 2 and 5. */
1820
95.3k
        if (bb5 > 0) {
1821
48.4k
            bs = pow5mult(bs, bb5);
1822
48.4k
            if (bs == NULL) {
1823
0
                goto failed_malloc;
1824
0
            }
1825
48.4k
            Bigint *bb1 = mult(bs, bb);
1826
48.4k
            Bfree(bb);
1827
48.4k
            bb = bb1;
1828
48.4k
            if (bb == NULL) {
1829
0
                goto failed_malloc;
1830
0
            }
1831
48.4k
        }
1832
95.3k
        if (bb2 > 0) {
1833
95.3k
            bb = lshift(bb, bb2);
1834
95.3k
            if (bb == NULL) {
1835
0
                goto failed_malloc;
1836
0
            }
1837
95.3k
        }
1838
95.3k
        if (bd5 > 0) {
1839
41.0k
            bd = pow5mult(bd, bd5);
1840
41.0k
            if (bd == NULL) {
1841
0
                goto failed_malloc;
1842
0
            }
1843
41.0k
        }
1844
95.3k
        if (bd2 > 0) {
1845
46.2k
            bd = lshift(bd, bd2);
1846
46.2k
            if (bd == NULL) {
1847
0
                goto failed_malloc;
1848
0
            }
1849
46.2k
        }
1850
95.3k
        if (bs2 > 0) {
1851
46.0k
            bs = lshift(bs, bs2);
1852
46.0k
            if (bs == NULL) {
1853
0
                goto failed_malloc;
1854
0
            }
1855
46.0k
        }
1856
1857
        /* Now bd, bb and bs are scaled versions of tdv, srv and 0.5 ulp(srv),
1858
           respectively.  Compute the difference |tdv - srv|, and compare
1859
           with 0.5 ulp(srv). */
1860
1861
95.3k
        delta = diff(bb, bd);
1862
95.3k
        if (delta == NULL) {
1863
0
            goto failed_malloc;
1864
0
        }
1865
95.3k
        dsign = delta->sign;
1866
95.3k
        delta->sign = 0;
1867
95.3k
        i = cmp(delta, bs);
1868
95.3k
        if (bc.nd > nd && i <= 0) {
1869
9.97k
            if (dsign)
1870
6.26k
                break;  /* Must use bigcomp(). */
1871
1872
            /* Here rv overestimates the truncated decimal value by at most
1873
               0.5 ulp(rv).  Hence rv either overestimates the true decimal
1874
               value by <= 0.5 ulp(rv), or underestimates it by some small
1875
               amount (< 0.1 ulp(rv)); either way, rv is within 0.5 ulps of
1876
               the true decimal value, so it's possible to exit.
1877
1878
               Exception: if scaled rv is a normal exact power of 2, but not
1879
               DBL_MIN, then rv - 0.5 ulp(rv) takes us all the way down to the
1880
               next double, so the correctly rounded result is either rv - 0.5
1881
               ulp(rv) or rv; in this case, use bigcomp to distinguish. */
1882
1883
3.70k
            if (!word1(&rv) && !(word0(&rv) & Bndry_mask)) {
1884
                /* rv can't be 0, since it's an overestimate for some
1885
                   nonzero value.  So rv is a normal power of 2. */
1886
1.16k
                j = (int)(word0(&rv) & Exp_mask) >> Exp_shift;
1887
                /* rv / 2^bc.scale = 2^(j - 1023 - bc.scale); use bigcomp if
1888
                   rv / 2^bc.scale >= 2^-1021. */
1889
1.16k
                if (j - bc.scale >= 2) {
1890
768
                    dval(&rv) -= 0.5 * sulp(&rv, &bc);
1891
768
                    break; /* Use bigcomp. */
1892
768
                }
1893
1.16k
            }
1894
1895
2.93k
            {
1896
2.93k
                bc.nd = nd;
1897
2.93k
                i = -1; /* Discarded digits make delta smaller. */
1898
2.93k
            }
1899
2.93k
        }
1900
1901
88.3k
        if (i < 0) {
1902
            /* Error is less than half an ulp -- check for
1903
             * special case of mantissa a power of two.
1904
             */
1905
44.5k
            if (dsign || word1(&rv) || word0(&rv) & Bndry_mask
1906
44.5k
                || (word0(&rv) & Exp_mask) <= (2*P+1)*Exp_msk1
1907
44.5k
                ) {
1908
41.4k
                break;
1909
41.4k
            }
1910
3.08k
            if (!delta->x[0] && delta->wds <= 1) {
1911
                /* exact result */
1912
417
                break;
1913
417
            }
1914
2.66k
            delta = lshift(delta,Log2P);
1915
2.66k
            if (delta == NULL) {
1916
0
                goto failed_malloc;
1917
0
            }
1918
2.66k
            if (cmp(delta, bs) > 0)
1919
882
                goto drop_down;
1920
1.78k
            break;
1921
2.66k
        }
1922
43.7k
        if (i == 0) {
1923
            /* exactly half-way between */
1924
3.57k
            if (dsign) {
1925
1.92k
                if ((word0(&rv) & Bndry_mask1) == Bndry_mask1
1926
1.92k
                    &&  word1(&rv) == (
1927
1.05k
                        (bc.scale &&
1928
1.05k
                         (y = word0(&rv) & Exp_mask) <= 2*P*Exp_msk1) ?
1929
0
                        (0xffffffff & (0xffffffff << (2*P+1-(y>>Exp_shift)))) :
1930
1.05k
                        0xffffffff)) {
1931
                    /*boundary case -- increment exponent*/
1932
643
                    word0(&rv) = (word0(&rv) & Exp_mask)
1933
643
                        + Exp_msk1
1934
643
                        ;
1935
643
                    word1(&rv) = 0;
1936
                    /* dsign = 0; */
1937
643
                    break;
1938
643
                }
1939
1.92k
            }
1940
1.64k
            else if (!(word0(&rv) & Bndry_mask) && !word1(&rv)) {
1941
882
              drop_down:
1942
                /* boundary case -- decrement exponent */
1943
882
                if (bc.scale) {
1944
0
                    L = word0(&rv) & Exp_mask;
1945
0
                    if (L <= (2*P+1)*Exp_msk1) {
1946
0
                        if (L > (P+2)*Exp_msk1)
1947
                            /* round even ==> */
1948
                            /* accept rv */
1949
0
                            break;
1950
                        /* rv = smallest denormal */
1951
0
                        if (bc.nd > nd)
1952
0
                            break;
1953
0
                        goto undfl;
1954
0
                    }
1955
0
                }
1956
882
                L = (word0(&rv) & Exp_mask) - Exp_msk1;
1957
882
                word0(&rv) = L | Bndry_mask1;
1958
882
                word1(&rv) = 0xffffffff;
1959
882
                break;
1960
882
            }
1961
2.93k
            if (!odd)
1962
2.26k
                break;
1963
664
            if (dsign)
1964
307
                dval(&rv) += sulp(&rv, &bc);
1965
357
            else {
1966
357
                dval(&rv) -= sulp(&rv, &bc);
1967
357
                if (!dval(&rv)) {
1968
0
                    if (bc.nd >nd)
1969
0
                        break;
1970
0
                    goto undfl;
1971
0
                }
1972
357
            }
1973
            /* dsign = 1 - dsign; */
1974
664
            break;
1975
664
        }
1976
40.2k
        if ((aadj = ratio(delta, bs)) <= 2.) {
1977
29.4k
            if (dsign)
1978
12.1k
                aadj = aadj1 = 1.;
1979
17.3k
            else if (word1(&rv) || word0(&rv) & Bndry_mask) {
1980
12.8k
                if (word1(&rv) == Tiny1 && !word0(&rv)) {
1981
0
                    if (bc.nd >nd)
1982
0
                        break;
1983
0
                    goto undfl;
1984
0
                }
1985
12.8k
                aadj = 1.;
1986
12.8k
                aadj1 = -1.;
1987
12.8k
            }
1988
4.45k
            else {
1989
                /* special case -- power of FLT_RADIX to be */
1990
                /* rounded down... */
1991
1992
4.45k
                if (aadj < 2./FLT_RADIX)
1993
0
                    aadj = 1./FLT_RADIX;
1994
4.45k
                else
1995
4.45k
                    aadj *= 0.5;
1996
4.45k
                aadj1 = -aadj;
1997
4.45k
            }
1998
29.4k
        }
1999
10.8k
        else {
2000
10.8k
            aadj *= 0.5;
2001
10.8k
            aadj1 = dsign ? aadj : -aadj;
2002
10.8k
            if (Flt_Rounds == 0)
2003
0
                aadj1 += 0.5;
2004
10.8k
        }
2005
40.2k
        y = word0(&rv) & Exp_mask;
2006
2007
        /* Check for overflow */
2008
2009
40.2k
        if (y == Exp_msk1*(DBL_MAX_EXP+Bias-1)) {
2010
2.23k
            dval(&rv0) = dval(&rv);
2011
2.23k
            word0(&rv) -= P*Exp_msk1;
2012
2.23k
            adj.d = aadj1 * ulp(&rv);
2013
2.23k
            dval(&rv) += adj.d;
2014
2.23k
            if ((word0(&rv) & Exp_mask) >=
2015
2.23k
                Exp_msk1*(DBL_MAX_EXP+Bias-P)) {
2016
1.58k
                if (word0(&rv0) == Big0 && word1(&rv0) == Big1) {
2017
1.08k
                    goto ovfl;
2018
1.08k
                }
2019
506
                word0(&rv) = Big0;
2020
506
                word1(&rv) = Big1;
2021
506
                goto cont;
2022
1.58k
            }
2023
642
            else
2024
642
                word0(&rv) += P*Exp_msk1;
2025
2.23k
        }
2026
37.9k
        else {
2027
37.9k
            if (bc.scale && y <= 2*P*Exp_msk1) {
2028
13.3k
                if (aadj <= 0x7fffffff) {
2029
13.3k
                    if ((z = (ULong)aadj) <= 0)
2030
679
                        z = 1;
2031
13.3k
                    aadj = z;
2032
13.3k
                    aadj1 = dsign ? aadj : -aadj;
2033
13.3k
                }
2034
13.3k
                dval(&aadj2) = aadj1;
2035
13.3k
                word0(&aadj2) += (2*P+1)*Exp_msk1 - y;
2036
13.3k
                aadj1 = dval(&aadj2);
2037
13.3k
            }
2038
37.9k
            adj.d = aadj1 * ulp(&rv);
2039
37.9k
            dval(&rv) += adj.d;
2040
37.9k
        }
2041
38.6k
        z = word0(&rv) & Exp_mask;
2042
38.6k
        if (bc.nd == nd) {
2043
34.4k
            if (!bc.scale)
2044
21.1k
                if (y == z) {
2045
                    /* Can we stop now? */
2046
19.7k
                    L = (Long)aadj;
2047
19.7k
                    aadj -= L;
2048
                    /* The tolerances below are conservative. */
2049
19.7k
                    if (dsign || word1(&rv) || word0(&rv) & Bndry_mask) {
2050
19.7k
                        if (aadj < .4999999 || aadj > .5000001)
2051
19.1k
                            break;
2052
19.7k
                    }
2053
3
                    else if (aadj < .4999999/FLT_RADIX)
2054
3
                        break;
2055
19.7k
                }
2056
34.4k
        }
2057
20.0k
      cont:
2058
20.0k
        Bfree(bb); bb = NULL;
2059
20.0k
        Bfree(bd); bd = NULL;
2060
20.0k
        Bfree(bs); bs = NULL;
2061
20.0k
        Bfree(delta); delta = NULL;
2062
20.0k
    }
2063
74.3k
    if (bc.nd > nd) {
2064
7.03k
        error = bigcomp(&rv, s0, &bc);
2065
7.03k
        if (error)
2066
0
            goto failed_malloc;
2067
7.03k
    }
2068
2069
74.3k
    if (bc.scale) {
2070
18.0k
        word0(&rv0) = Exp_1 - 2*P*Exp_msk1;
2071
18.0k
        word1(&rv0) = 0;
2072
18.0k
        dval(&rv) *= dval(&rv0);
2073
18.0k
    }
2074
2075
156k
  ret:
2076
156k
    result = sign ? -dval(&rv) : dval(&rv);
2077
156k
    goto done;
2078
2079
23
  parse_error:
2080
23
    result = 0.0;
2081
23
    goto done;
2082
2083
0
  failed_malloc:
2084
0
    errno = ENOMEM;
2085
0
    result = -1.0;
2086
0
    goto done;
2087
2088
1.30k
  undfl:
2089
1.30k
    result = sign ? -0.0 : 0.0;
2090
1.30k
    goto done;
2091
2092
478k
  ovfl:
2093
478k
    errno = ERANGE;
2094
    /* Can't trust HUGE_VAL */
2095
478k
    word0(&rv) = Exp_mask;
2096
478k
    word1(&rv) = 0;
2097
478k
    result = sign ? -dval(&rv) : dval(&rv);
2098
478k
    goto done;
2099
2100
636k
  done:
2101
636k
    Bfree(bb);
2102
636k
    Bfree(bd);
2103
636k
    Bfree(bs);
2104
636k
    Bfree(bd0);
2105
636k
    Bfree(delta);
2106
636k
    return result;
2107
2108
74.3k
}
2109
2110
static char *
2111
rv_alloc(int i)
2112
55.1k
{
2113
55.1k
    int j, k, *r;
2114
2115
55.1k
    j = sizeof(ULong);
2116
55.1k
    for(k = 0;
2117
55.1k
        sizeof(Bigint) - sizeof(ULong) - sizeof(int) + j <= (unsigned)i;
2118
55.1k
        j <<= 1)
2119
0
        k++;
2120
55.1k
    r = (int*)Balloc(k);
2121
55.1k
    if (r == NULL)
2122
0
        return NULL;
2123
55.1k
    *r = k;
2124
55.1k
    return (char *)(r+1);
2125
55.1k
}
2126
2127
static char *
2128
nrv_alloc(const char *s, char **rve, int n)
2129
5.32k
{
2130
5.32k
    char *rv, *t;
2131
2132
5.32k
    rv = rv_alloc(n);
2133
5.32k
    if (rv == NULL)
2134
0
        return NULL;
2135
5.32k
    t = rv;
2136
15.4k
    while((*t = *s++)) t++;
2137
5.32k
    if (rve)
2138
5.32k
        *rve = t;
2139
5.32k
    return rv;
2140
5.32k
}
2141
2142
/* freedtoa(s) must be used to free values s returned by dtoa
2143
 * when MULTIPLE_THREADS is #defined.  It should be used in all cases,
2144
 * but for consistency with earlier versions of dtoa, it is optional
2145
 * when MULTIPLE_THREADS is not defined.
2146
 */
2147
2148
void
2149
_Py_dg_freedtoa(char *s)
2150
55.1k
{
2151
55.1k
    Bigint *b = (Bigint *)((int *)s - 1);
2152
55.1k
    b->maxwds = 1 << (b->k = *(int*)b);
2153
55.1k
    Bfree(b);
2154
55.1k
}
2155
2156
/* dtoa for IEEE arithmetic (dmg): convert double to ASCII string.
2157
 *
2158
 * Inspired by "How to Print Floating-Point Numbers Accurately" by
2159
 * Guy L. Steele, Jr. and Jon L. White [Proc. ACM SIGPLAN '90, pp. 112-126].
2160
 *
2161
 * Modifications:
2162
 *      1. Rather than iterating, we use a simple numeric overestimate
2163
 *         to determine k = floor(log10(d)).  We scale relevant
2164
 *         quantities using O(log2(k)) rather than O(k) multiplications.
2165
 *      2. For some modes > 2 (corresponding to ecvt and fcvt), we don't
2166
 *         try to generate digits strictly left to right.  Instead, we
2167
 *         compute with fewer bits and propagate the carry if necessary
2168
 *         when rounding the final digit up.  This is often faster.
2169
 *      3. Under the assumption that input will be rounded nearest,
2170
 *         mode 0 renders 1e23 as 1e23 rather than 9.999999999999999e22.
2171
 *         That is, we allow equality in stopping tests when the
2172
 *         round-nearest rule will give the same floating-point value
2173
 *         as would satisfaction of the stopping test with strict
2174
 *         inequality.
2175
 *      4. We remove common factors of powers of 2 from relevant
2176
 *         quantities.
2177
 *      5. When converting floating-point integers less than 1e16,
2178
 *         we use floating-point arithmetic rather than resorting
2179
 *         to multiple-precision integers.
2180
 *      6. When asked to produce fewer than 15 digits, we first try
2181
 *         to get by with floating-point arithmetic; we resort to
2182
 *         multiple-precision integer arithmetic only if we cannot
2183
 *         guarantee that the floating-point calculation has given
2184
 *         the correctly rounded result.  For k requested digits and
2185
 *         "uniformly" distributed input, the probability is
2186
 *         something like 10^(k-15) that we must resort to the Long
2187
 *         calculation.
2188
 */
2189
2190
/* Additional notes (METD): (1) returns NULL on failure.  (2) to avoid memory
2191
   leakage, a successful call to _Py_dg_dtoa should always be matched by a
2192
   call to _Py_dg_freedtoa. */
2193
2194
char *
2195
_Py_dg_dtoa(double dd, int mode, int ndigits,
2196
            int *decpt, int *sign, char **rve)
2197
55.1k
{
2198
    /*  Arguments ndigits, decpt, sign are similar to those
2199
        of ecvt and fcvt; trailing zeros are suppressed from
2200
        the returned string.  If not null, *rve is set to point
2201
        to the end of the return value.  If d is +-Infinity or NaN,
2202
        then *decpt is set to 9999.
2203
2204
        mode:
2205
        0 ==> shortest string that yields d when read in
2206
        and rounded to nearest.
2207
        1 ==> like 0, but with Steele & White stopping rule;
2208
        e.g. with IEEE P754 arithmetic , mode 0 gives
2209
        1e23 whereas mode 1 gives 9.999999999999999e22.
2210
        2 ==> max(1,ndigits) significant digits.  This gives a
2211
        return value similar to that of ecvt, except
2212
        that trailing zeros are suppressed.
2213
        3 ==> through ndigits past the decimal point.  This
2214
        gives a return value similar to that from fcvt,
2215
        except that trailing zeros are suppressed, and
2216
        ndigits can be negative.
2217
        4,5 ==> similar to 2 and 3, respectively, but (in
2218
        round-nearest mode) with the tests of mode 0 to
2219
        possibly return a shorter string that rounds to d.
2220
        With IEEE arithmetic and compilation with
2221
        -DHonor_FLT_ROUNDS, modes 4 and 5 behave the same
2222
        as modes 2 and 3 when FLT_ROUNDS != 1.
2223
        6-9 ==> Debugging modes similar to mode - 4:  don't try
2224
        fast floating-point estimate (if applicable).
2225
2226
        Values of mode other than 0-9 are treated as mode 0.
2227
2228
        Sufficient space is allocated to the return value
2229
        to hold the suppressed trailing zeros.
2230
    */
2231
2232
55.1k
    int bbits, b2, b5, be, dig, i, ieps, ilim, ilim0, ilim1,
2233
55.1k
        j, j1, k, k0, k_check, leftright, m2, m5, s2, s5,
2234
55.1k
        spec_case, try_quick;
2235
55.1k
    Long L;
2236
55.1k
    int denorm;
2237
55.1k
    ULong x;
2238
55.1k
    Bigint *b, *b1, *delta, *mlo, *mhi, *S;
2239
55.1k
    U d2, eps, u;
2240
55.1k
    double ds;
2241
55.1k
    char *s, *s0;
2242
2243
    /* set pointers to NULL, to silence gcc compiler warnings and make
2244
       cleanup easier on error */
2245
55.1k
    mlo = mhi = S = 0;
2246
55.1k
    s0 = 0;
2247
2248
55.1k
    u.d = dd;
2249
55.1k
    if (word0(&u) & Sign_bit) {
2250
        /* set sign for everything, including 0's and NaNs */
2251
14.1k
        *sign = 1;
2252
14.1k
        word0(&u) &= ~Sign_bit; /* clear sign bit */
2253
14.1k
    }
2254
41.0k
    else
2255
41.0k
        *sign = 0;
2256
2257
    /* quick return for Infinities, NaNs and zeros */
2258
55.1k
    if ((word0(&u) & Exp_mask) == Exp_mask)
2259
692
    {
2260
        /* Infinity or NaN */
2261
692
        *decpt = 9999;
2262
692
        if (!word1(&u) && !(word0(&u) & 0xfffff))
2263
692
            return nrv_alloc("Infinity", rve, 8);
2264
0
        return nrv_alloc("NaN", rve, 3);
2265
692
    }
2266
54.4k
    if (!dval(&u)) {
2267
4.63k
        *decpt = 1;
2268
4.63k
        return nrv_alloc("0", rve, 1);
2269
4.63k
    }
2270
2271
    /* compute k = floor(log10(d)).  The computation may leave k
2272
       one too large, but should never leave k too small. */
2273
49.8k
    b = d2b(&u, &be, &bbits);
2274
49.8k
    if (b == NULL)
2275
0
        goto failed_malloc;
2276
49.8k
    if ((i = (int)(word0(&u) >> Exp_shift1 & (Exp_mask>>Exp_shift1)))) {
2277
45.7k
        dval(&d2) = dval(&u);
2278
45.7k
        word0(&d2) &= Frac_mask1;
2279
45.7k
        word0(&d2) |= Exp_11;
2280
2281
        /* log(x)       ~=~ log(1.5) + (x-1.5)/1.5
2282
         * log10(x)      =  log(x) / log(10)
2283
         *              ~=~ log(1.5)/log(10) + (x-1.5)/(1.5*log(10))
2284
         * log10(d) = (i-Bias)*log(2)/log(10) + log10(d2)
2285
         *
2286
         * This suggests computing an approximation k to log10(d) by
2287
         *
2288
         * k = (i - Bias)*0.301029995663981
2289
         *      + ( (d2-1.5)*0.289529654602168 + 0.176091259055681 );
2290
         *
2291
         * We want k to be too large rather than too small.
2292
         * The error in the first-order Taylor series approximation
2293
         * is in our favor, so we just round up the constant enough
2294
         * to compensate for any error in the multiplication of
2295
         * (i - Bias) by 0.301029995663981; since |i - Bias| <= 1077,
2296
         * and 1077 * 0.30103 * 2^-52 ~=~ 7.2e-14,
2297
         * adding 1e-13 to the constant term more than suffices.
2298
         * Hence we adjust the constant term to 0.1760912590558.
2299
         * (We could get a more accurate k by invoking log10,
2300
         *  but this is probably not worthwhile.)
2301
         */
2302
2303
45.7k
        i -= Bias;
2304
45.7k
        denorm = 0;
2305
45.7k
    }
2306
4.12k
    else {
2307
        /* d is denormalized */
2308
2309
4.12k
        i = bbits + be + (Bias + (P-1) - 1);
2310
4.12k
        x = i > 32  ? word0(&u) << (64 - i) | word1(&u) >> (i - 32)
2311
4.12k
            : word1(&u) << (32 - i);
2312
4.12k
        dval(&d2) = x;
2313
4.12k
        word0(&d2) -= 31*Exp_msk1; /* adjust exponent */
2314
4.12k
        i -= (Bias + (P-1) - 1) + 1;
2315
4.12k
        denorm = 1;
2316
4.12k
    }
2317
49.8k
    ds = (dval(&d2)-1.5)*0.289529654602168 + 0.1760912590558 +
2318
49.8k
        i*0.301029995663981;
2319
49.8k
    k = (int)ds;
2320
49.8k
    if (ds < 0. && ds != k)
2321
12.1k
        k--;    /* want k = floor(ds) */
2322
49.8k
    k_check = 1;
2323
49.8k
    if (k >= 0 && k <= Ten_pmax) {
2324
23.0k
        if (dval(&u) < tens[k])
2325
2.18k
            k--;
2326
23.0k
        k_check = 0;
2327
23.0k
    }
2328
49.8k
    j = bbits - i - 1;
2329
49.8k
    if (j >= 0) {
2330
19.7k
        b2 = 0;
2331
19.7k
        s2 = j;
2332
19.7k
    }
2333
30.1k
    else {
2334
30.1k
        b2 = -j;
2335
30.1k
        s2 = 0;
2336
30.1k
    }
2337
49.8k
    if (k >= 0) {
2338
37.1k
        b5 = 0;
2339
37.1k
        s5 = k;
2340
37.1k
        s2 += k;
2341
37.1k
    }
2342
12.6k
    else {
2343
12.6k
        b2 -= k;
2344
12.6k
        b5 = -k;
2345
12.6k
        s5 = 0;
2346
12.6k
    }
2347
49.8k
    if (mode < 0 || mode > 9)
2348
0
        mode = 0;
2349
2350
49.8k
    try_quick = 1;
2351
2352
49.8k
    if (mode > 5) {
2353
0
        mode -= 4;
2354
0
        try_quick = 0;
2355
0
    }
2356
49.8k
    leftright = 1;
2357
49.8k
    ilim = ilim1 = -1;  /* Values for cases 0 and 1; done here to */
2358
    /* silence erroneous "gcc -Wall" warning. */
2359
49.8k
    switch(mode) {
2360
49.8k
    case 0:
2361
49.8k
    case 1:
2362
49.8k
        i = 18;
2363
49.8k
        ndigits = 0;
2364
49.8k
        break;
2365
0
    case 2:
2366
0
        leftright = 0;
2367
0
        _Py_FALLTHROUGH;
2368
0
    case 4:
2369
0
        if (ndigits <= 0)
2370
0
            ndigits = 1;
2371
0
        ilim = ilim1 = i = ndigits;
2372
0
        break;
2373
0
    case 3:
2374
0
        leftright = 0;
2375
0
        _Py_FALLTHROUGH;
2376
0
    case 5:
2377
0
        i = ndigits + k + 1;
2378
0
        ilim = i;
2379
0
        ilim1 = i - 1;
2380
0
        if (i <= 0)
2381
0
            i = 1;
2382
49.8k
    }
2383
49.8k
    s0 = rv_alloc(i);
2384
49.8k
    if (s0 == NULL)
2385
0
        goto failed_malloc;
2386
49.8k
    s = s0;
2387
2388
2389
49.8k
    if (ilim >= 0 && ilim <= Quick_max && try_quick) {
2390
2391
        /* Try to get by with floating-point arithmetic. */
2392
2393
0
        i = 0;
2394
0
        dval(&d2) = dval(&u);
2395
0
        k0 = k;
2396
0
        ilim0 = ilim;
2397
0
        ieps = 2; /* conservative */
2398
0
        if (k > 0) {
2399
0
            ds = tens[k&0xf];
2400
0
            j = k >> 4;
2401
0
            if (j & Bletch) {
2402
                /* prevent overflows */
2403
0
                j &= Bletch - 1;
2404
0
                dval(&u) /= bigtens[n_bigtens-1];
2405
0
                ieps++;
2406
0
            }
2407
0
            for(; j; j >>= 1, i++)
2408
0
                if (j & 1) {
2409
0
                    ieps++;
2410
0
                    ds *= bigtens[i];
2411
0
                }
2412
0
            dval(&u) /= ds;
2413
0
        }
2414
0
        else if ((j1 = -k)) {
2415
0
            dval(&u) *= tens[j1 & 0xf];
2416
0
            for(j = j1 >> 4; j; j >>= 1, i++)
2417
0
                if (j & 1) {
2418
0
                    ieps++;
2419
0
                    dval(&u) *= bigtens[i];
2420
0
                }
2421
0
        }
2422
0
        if (k_check && dval(&u) < 1. && ilim > 0) {
2423
0
            if (ilim1 <= 0)
2424
0
                goto fast_failed;
2425
0
            ilim = ilim1;
2426
0
            k--;
2427
0
            dval(&u) *= 10.;
2428
0
            ieps++;
2429
0
        }
2430
0
        dval(&eps) = ieps*dval(&u) + 7.;
2431
0
        word0(&eps) -= (P-1)*Exp_msk1;
2432
0
        if (ilim == 0) {
2433
0
            S = mhi = 0;
2434
0
            dval(&u) -= 5.;
2435
0
            if (dval(&u) > dval(&eps))
2436
0
                goto one_digit;
2437
0
            if (dval(&u) < -dval(&eps))
2438
0
                goto no_digits;
2439
0
            goto fast_failed;
2440
0
        }
2441
0
        if (leftright) {
2442
            /* Use Steele & White method of only
2443
             * generating digits needed.
2444
             */
2445
0
            dval(&eps) = 0.5/tens[ilim-1] - dval(&eps);
2446
0
            for(i = 0;;) {
2447
0
                L = (Long)dval(&u);
2448
0
                dval(&u) -= L;
2449
0
                *s++ = '0' + (int)L;
2450
0
                if (dval(&u) < dval(&eps))
2451
0
                    goto ret1;
2452
0
                if (1. - dval(&u) < dval(&eps))
2453
0
                    goto bump_up;
2454
0
                if (++i >= ilim)
2455
0
                    break;
2456
0
                dval(&eps) *= 10.;
2457
0
                dval(&u) *= 10.;
2458
0
            }
2459
0
        }
2460
0
        else {
2461
            /* Generate ilim digits, then fix them up. */
2462
0
            dval(&eps) *= tens[ilim-1];
2463
0
            for(i = 1;; i++, dval(&u) *= 10.) {
2464
0
                L = (Long)(dval(&u));
2465
0
                if (!(dval(&u) -= L))
2466
0
                    ilim = i;
2467
0
                *s++ = '0' + (int)L;
2468
0
                if (i == ilim) {
2469
0
                    if (dval(&u) > 0.5 + dval(&eps))
2470
0
                        goto bump_up;
2471
0
                    else if (dval(&u) < 0.5 - dval(&eps)) {
2472
0
                        while(*--s == '0');
2473
0
                        s++;
2474
0
                        goto ret1;
2475
0
                    }
2476
0
                    break;
2477
0
                }
2478
0
            }
2479
0
        }
2480
0
      fast_failed:
2481
0
        s = s0;
2482
0
        dval(&u) = dval(&d2);
2483
0
        k = k0;
2484
0
        ilim = ilim0;
2485
0
    }
2486
2487
    /* Do we have a "small" integer? */
2488
2489
49.8k
    if (be >= 0 && k <= Int_max) {
2490
        /* Yes. */
2491
13.4k
        ds = tens[k];
2492
13.4k
        if (ndigits < 0 && ilim <= 0) {
2493
0
            S = mhi = 0;
2494
0
            if (ilim < 0 || dval(&u) <= 5*ds)
2495
0
                goto no_digits;
2496
0
            goto one_digit;
2497
0
        }
2498
22.1k
        for(i = 1;; i++, dval(&u) *= 10.) {
2499
22.1k
            L = (Long)(dval(&u) / ds);
2500
22.1k
            dval(&u) -= L*ds;
2501
22.1k
            *s++ = '0' + (int)L;
2502
22.1k
            if (!dval(&u)) {
2503
13.4k
                break;
2504
13.4k
            }
2505
8.69k
            if (i == ilim) {
2506
0
                dval(&u) += dval(&u);
2507
0
                if (dval(&u) > ds || (dval(&u) == ds && L & 1)) {
2508
0
                  bump_up:
2509
0
                    while(*--s == '9')
2510
0
                        if (s == s0) {
2511
0
                            k++;
2512
0
                            *s = '0';
2513
0
                            break;
2514
0
                        }
2515
0
                    ++*s++;
2516
0
                }
2517
0
                else {
2518
                    /* Strip trailing zeros. This branch was missing from the
2519
                       original dtoa.c, leading to surplus trailing zeros in
2520
                       some cases. See bugs.python.org/issue40780. */
2521
0
                    while (s > s0 && s[-1] == '0') {
2522
0
                        --s;
2523
0
                    }
2524
0
                }
2525
0
                break;
2526
0
            }
2527
8.69k
        }
2528
13.4k
        goto ret1;
2529
13.4k
    }
2530
2531
36.3k
    m2 = b2;
2532
36.3k
    m5 = b5;
2533
36.3k
    if (leftright) {
2534
36.3k
        i =
2535
36.3k
            denorm ? be + (Bias + (P-1) - 1 + 1) :
2536
36.3k
            1 + P - bbits;
2537
36.3k
        b2 += i;
2538
36.3k
        s2 += i;
2539
36.3k
        mhi = i2b(1);
2540
36.3k
        if (mhi == NULL)
2541
0
            goto failed_malloc;
2542
36.3k
    }
2543
36.3k
    if (m2 > 0 && s2 > 0) {
2544
31.1k
        i = m2 < s2 ? m2 : s2;
2545
31.1k
        b2 -= i;
2546
31.1k
        m2 -= i;
2547
31.1k
        s2 -= i;
2548
31.1k
    }
2549
36.3k
    if (b5 > 0) {
2550
12.6k
        if (leftright) {
2551
12.6k
            if (m5 > 0) {
2552
12.6k
                mhi = pow5mult(mhi, m5);
2553
12.6k
                if (mhi == NULL)
2554
0
                    goto failed_malloc;
2555
12.6k
                b1 = mult(mhi, b);
2556
12.6k
                Bfree(b);
2557
12.6k
                b = b1;
2558
12.6k
                if (b == NULL)
2559
0
                    goto failed_malloc;
2560
12.6k
            }
2561
12.6k
            if ((j = b5 - m5)) {
2562
0
                b = pow5mult(b, j);
2563
0
                if (b == NULL)
2564
0
                    goto failed_malloc;
2565
0
            }
2566
12.6k
        }
2567
0
        else {
2568
0
            b = pow5mult(b, b5);
2569
0
            if (b == NULL)
2570
0
                goto failed_malloc;
2571
0
        }
2572
12.6k
    }
2573
36.3k
    S = i2b(1);
2574
36.3k
    if (S == NULL)
2575
0
        goto failed_malloc;
2576
36.3k
    if (s5 > 0) {
2577
20.8k
        S = pow5mult(S, s5);
2578
20.8k
        if (S == NULL)
2579
0
            goto failed_malloc;
2580
20.8k
    }
2581
2582
    /* Check for special case that d is a normalized power of 2. */
2583
2584
36.3k
    spec_case = 0;
2585
36.3k
    if ((mode < 2 || leftright)
2586
36.3k
        ) {
2587
36.3k
        if (!word1(&u) && !(word0(&u) & Bndry_mask)
2588
36.3k
            && word0(&u) & (Exp_mask & ~Exp_msk1)
2589
36.3k
            ) {
2590
            /* The special case */
2591
1.03k
            b2 += Log2P;
2592
1.03k
            s2 += Log2P;
2593
1.03k
            spec_case = 1;
2594
1.03k
        }
2595
36.3k
    }
2596
2597
    /* Arrange for convenient computation of quotients:
2598
     * shift left if necessary so divisor has 4 leading 0 bits.
2599
     *
2600
     * Perhaps we should just compute leading 28 bits of S once
2601
     * and for all and pass them and a shift to quorem, so it
2602
     * can do shifts and ors to compute the numerator for q.
2603
     */
2604
36.3k
#define iInc 28
2605
36.3k
    i = dshift(S, s2);
2606
36.3k
    b2 += i;
2607
36.3k
    m2 += i;
2608
36.3k
    s2 += i;
2609
36.3k
    if (b2 > 0) {
2610
36.3k
        b = lshift(b, b2);
2611
36.3k
        if (b == NULL)
2612
0
            goto failed_malloc;
2613
36.3k
    }
2614
36.3k
    if (s2 > 0) {
2615
35.7k
        S = lshift(S, s2);
2616
35.7k
        if (S == NULL)
2617
0
            goto failed_malloc;
2618
35.7k
    }
2619
36.3k
    if (k_check) {
2620
26.8k
        if (cmp(b,S) < 0) {
2621
2.91k
            k--;
2622
2.91k
            b = multadd(b, 10, 0);      /* we botched the k estimate */
2623
2.91k
            if (b == NULL)
2624
0
                goto failed_malloc;
2625
2.91k
            if (leftright) {
2626
2.91k
                mhi = multadd(mhi, 10, 0);
2627
2.91k
                if (mhi == NULL)
2628
0
                    goto failed_malloc;
2629
2.91k
            }
2630
2.91k
            ilim = ilim1;
2631
2.91k
        }
2632
26.8k
    }
2633
36.3k
    if (ilim <= 0 && (mode == 3 || mode == 5)) {
2634
0
        if (ilim < 0) {
2635
            /* no digits, fcvt style */
2636
0
          no_digits:
2637
0
            k = -1 - ndigits;
2638
0
            goto ret;
2639
0
        }
2640
0
        else {
2641
0
            S = multadd(S, 5, 0);
2642
0
            if (S == NULL)
2643
0
                goto failed_malloc;
2644
0
            if (cmp(b, S) <= 0)
2645
0
                goto no_digits;
2646
0
        }
2647
0
      one_digit:
2648
0
        *s++ = '1';
2649
0
        k++;
2650
0
        goto ret;
2651
0
    }
2652
36.3k
    if (leftright) {
2653
36.3k
        if (m2 > 0) {
2654
35.4k
            mhi = lshift(mhi, m2);
2655
35.4k
            if (mhi == NULL)
2656
0
                goto failed_malloc;
2657
35.4k
        }
2658
2659
        /* Compute mlo -- check for special case
2660
         * that d is a normalized power of 2.
2661
         */
2662
2663
36.3k
        mlo = mhi;
2664
36.3k
        if (spec_case) {
2665
1.03k
            mhi = Balloc(mhi->k);
2666
1.03k
            if (mhi == NULL)
2667
0
                goto failed_malloc;
2668
1.03k
            Bcopy(mhi, mlo);
2669
1.03k
            mhi = lshift(mhi, Log2P);
2670
1.03k
            if (mhi == NULL)
2671
0
                goto failed_malloc;
2672
1.03k
        }
2673
2674
162k
        for(i = 1;;i++) {
2675
162k
            dig = quorem(b,S) + '0';
2676
            /* Do we yet have the shortest decimal string
2677
             * that will round to d?
2678
             */
2679
162k
            j = cmp(b, mlo);
2680
162k
            delta = diff(S, mhi);
2681
162k
            if (delta == NULL)
2682
0
                goto failed_malloc;
2683
162k
            j1 = delta->sign ? 1 : cmp(b, delta);
2684
162k
            Bfree(delta);
2685
162k
            if (j1 == 0 && mode != 1 && !(word1(&u) & 1)
2686
162k
                ) {
2687
2.03k
                if (dig == '9')
2688
533
                    goto round_9_up;
2689
1.50k
                if (j > 0)
2690
782
                    dig++;
2691
1.50k
                *s++ = dig;
2692
1.50k
                goto ret;
2693
2.03k
            }
2694
160k
            if (j < 0 || (j == 0 && mode != 1
2695
140k
                          && !(word1(&u) & 1)
2696
140k
                    )) {
2697
20.7k
                if (!b->x[0] && b->wds <= 1) {
2698
2.86k
                    goto accept_dig;
2699
2.86k
                }
2700
17.8k
                if (j1 > 0) {
2701
3.42k
                    b = lshift(b, 1);
2702
3.42k
                    if (b == NULL)
2703
0
                        goto failed_malloc;
2704
3.42k
                    j1 = cmp(b, S);
2705
3.42k
                    if ((j1 > 0 || (j1 == 0 && dig & 1))
2706
3.42k
                        && dig++ == '9')
2707
274
                        goto round_9_up;
2708
3.42k
                }
2709
20.4k
              accept_dig:
2710
20.4k
                *s++ = dig;
2711
20.4k
                goto ret;
2712
17.8k
            }
2713
139k
            if (j1 > 0) {
2714
13.6k
                if (dig == '9') { /* possible if i == 1 */
2715
2.72k
                  round_9_up:
2716
2.72k
                    *s++ = '9';
2717
2.72k
                    goto roundoff;
2718
1.91k
                }
2719
11.7k
                *s++ = dig + 1;
2720
11.7k
                goto ret;
2721
13.6k
            }
2722
126k
            *s++ = dig;
2723
126k
            if (i == ilim)
2724
0
                break;
2725
126k
            b = multadd(b, 10, 0);
2726
126k
            if (b == NULL)
2727
0
                goto failed_malloc;
2728
126k
            if (mlo == mhi) {
2729
123k
                mlo = mhi = multadd(mhi, 10, 0);
2730
123k
                if (mlo == NULL)
2731
0
                    goto failed_malloc;
2732
123k
            }
2733
2.34k
            else {
2734
2.34k
                mlo = multadd(mlo, 10, 0);
2735
2.34k
                if (mlo == NULL)
2736
0
                    goto failed_malloc;
2737
2.34k
                mhi = multadd(mhi, 10, 0);
2738
2.34k
                if (mhi == NULL)
2739
0
                    goto failed_malloc;
2740
2.34k
            }
2741
126k
        }
2742
36.3k
    }
2743
0
    else
2744
0
        for(i = 1;; i++) {
2745
0
            *s++ = dig = quorem(b,S) + '0';
2746
0
            if (!b->x[0] && b->wds <= 1) {
2747
0
                goto ret;
2748
0
            }
2749
0
            if (i >= ilim)
2750
0
                break;
2751
0
            b = multadd(b, 10, 0);
2752
0
            if (b == NULL)
2753
0
                goto failed_malloc;
2754
0
        }
2755
2756
    /* Round off last digit */
2757
2758
0
    b = lshift(b, 1);
2759
0
    if (b == NULL)
2760
0
        goto failed_malloc;
2761
0
    j = cmp(b, S);
2762
0
    if (j > 0 || (j == 0 && dig & 1)) {
2763
2.72k
      roundoff:
2764
2.72k
        while(*--s == '9')
2765
2.72k
            if (s == s0) {
2766
2.72k
                k++;
2767
2.72k
                *s++ = '1';
2768
2.72k
                goto ret;
2769
2.72k
            }
2770
0
        ++*s++;
2771
0
    }
2772
0
    else {
2773
0
        while(*--s == '0');
2774
0
        s++;
2775
0
    }
2776
36.3k
  ret:
2777
36.3k
    Bfree(S);
2778
36.3k
    if (mhi) {
2779
36.3k
        if (mlo && mlo != mhi)
2780
1.03k
            Bfree(mlo);
2781
36.3k
        Bfree(mhi);
2782
36.3k
    }
2783
49.8k
  ret1:
2784
49.8k
    Bfree(b);
2785
49.8k
    *s = 0;
2786
49.8k
    *decpt = k + 1;
2787
49.8k
    if (rve)
2788
49.8k
        *rve = s;
2789
49.8k
    return s0;
2790
0
  failed_malloc:
2791
0
    if (S)
2792
0
        Bfree(S);
2793
0
    if (mlo && mlo != mhi)
2794
0
        Bfree(mlo);
2795
0
    if (mhi)
2796
0
        Bfree(mhi);
2797
0
    if (b)
2798
0
        Bfree(b);
2799
0
    if (s0)
2800
0
        _Py_dg_freedtoa(s0);
2801
0
    return NULL;
2802
36.3k
}
2803
2804
#endif  // _PY_SHORT_FLOAT_REPR == 1
2805
2806
PyStatus
2807
_PyDtoa_Init(PyInterpreterState *interp)
2808
16
{
2809
16
#if _PY_SHORT_FLOAT_REPR == 1 && !defined(Py_USING_MEMORY_DEBUGGER)
2810
16
    Bigint **p5s = interp->dtoa.p5s;
2811
2812
    // 5**4 = 625
2813
16
    Bigint *p5 = i2b(625);
2814
16
    if (p5 == NULL) {
2815
0
        return PyStatus_NoMemory();
2816
0
    }
2817
16
    p5s[0] = p5;
2818
2819
    // compute 5**8, 5**16, 5**32, ..., 5**512
2820
128
    for (Py_ssize_t i = 1; i < Bigint_Pow5size; i++) {
2821
112
        p5 = mult(p5, p5);
2822
112
        if (p5 == NULL) {
2823
0
            return PyStatus_NoMemory();
2824
0
        }
2825
112
        p5s[i] = p5;
2826
112
    }
2827
2828
16
#endif
2829
16
    return PyStatus_Ok();
2830
16
}
2831
2832
void
2833
_PyDtoa_Fini(PyInterpreterState *interp)
2834
0
{
2835
0
#if _PY_SHORT_FLOAT_REPR == 1 && !defined(Py_USING_MEMORY_DEBUGGER)
2836
0
    Bigint **p5s = interp->dtoa.p5s;
2837
0
    for (Py_ssize_t i = 0; i < Bigint_Pow5size; i++) {
2838
0
        Bigint *p5 = p5s[i];
2839
0
        p5s[i] = NULL;
2840
0
        Bfree(p5);
2841
0
    }
2842
0
#endif
2843
0
}