Issue #5864: format(1234.5, '.4') gives misleading result
[python.git] / Python / pystrtod.c
blob2df4e3d0ad23a7c49746218ee2814ddae4c4f4c6
1 /* -*- Mode: C; c-file-style: "python" -*- */
3 #include <Python.h>
4 #include <locale.h>
6 /**
7 * PyOS_ascii_strtod:
8 * @nptr: the string to convert to a numeric value.
9 * @endptr: if non-%NULL, it returns the character after
10 * the last character used in the conversion.
12 * Converts a string to a #gdouble value.
13 * This function behaves like the standard strtod() function
14 * does in the C locale. It does this without actually
15 * changing the current locale, since that would not be
16 * thread-safe.
18 * This function is typically used when reading configuration
19 * files or other non-user input that should be locale independent.
20 * To handle input from the user you should normally use the
21 * locale-sensitive system strtod() function.
23 * If the correct value would cause overflow, plus or minus %HUGE_VAL
24 * is returned (according to the sign of the value), and %ERANGE is
25 * stored in %errno. If the correct value would cause underflow,
26 * zero is returned and %ERANGE is stored in %errno.
27 * If memory allocation fails, %ENOMEM is stored in %errno.
29 * This function resets %errno before calling strtod() so that
30 * you can reliably detect overflow and underflow.
32 * Return value: the #gdouble value.
33 **/
36 Use system strtod; since strtod is locale aware, we may
37 have to first fix the decimal separator.
39 Note that unlike _Py_dg_strtod, the system strtod may not always give
40 correctly rounded results.
43 double
44 PyOS_ascii_strtod(const char *nptr, char **endptr)
46 char *fail_pos;
47 double val = -1.0;
48 struct lconv *locale_data;
49 const char *decimal_point;
50 size_t decimal_point_len;
51 const char *p, *decimal_point_pos;
52 const char *end = NULL; /* Silence gcc */
53 const char *digits_pos = NULL;
54 int negate = 0;
56 assert(nptr != NULL);
58 fail_pos = NULL;
60 locale_data = localeconv();
61 decimal_point = locale_data->decimal_point;
62 decimal_point_len = strlen(decimal_point);
64 assert(decimal_point_len != 0);
66 decimal_point_pos = NULL;
68 /* Set errno to zero, so that we can distinguish zero results
69 and underflows */
70 errno = 0;
72 /* We process any leading whitespace and the optional sign manually,
73 then pass the remainder to the system strtod. This ensures that
74 the result of an underflow has the correct sign. (bug #1725) */
76 p = nptr;
77 /* Skip leading space */
78 while (Py_ISSPACE(*p))
79 p++;
81 /* Process leading sign, if present */
82 if (*p == '-') {
83 negate = 1;
84 p++;
86 else if (*p == '+') {
87 p++;
90 /* Parse infinities and nans */
91 if (*p == 'i' || *p == 'I') {
92 if (PyOS_strnicmp(p, "inf", 3) == 0) {
93 val = Py_HUGE_VAL;
94 if (PyOS_strnicmp(p+3, "inity", 5) == 0)
95 fail_pos = (char *)p+8;
96 else
97 fail_pos = (char *)p+3;
98 goto got_val;
100 else
101 goto invalid_string;
103 #ifdef Py_NAN
104 if (*p == 'n' || *p == 'N') {
105 if (PyOS_strnicmp(p, "nan", 3) == 0) {
106 val = Py_NAN;
107 fail_pos = (char *)p+3;
108 goto got_val;
110 else
111 goto invalid_string;
113 #endif
115 /* Some platform strtods accept hex floats; Python shouldn't (at the
116 moment), so we check explicitly for strings starting with '0x'. */
117 if (*p == '0' && (*(p+1) == 'x' || *(p+1) == 'X'))
118 goto invalid_string;
120 /* Check that what's left begins with a digit or decimal point */
121 if (!Py_ISDIGIT(*p) && *p != '.')
122 goto invalid_string;
124 digits_pos = p;
125 if (decimal_point[0] != '.' ||
126 decimal_point[1] != 0)
128 /* Look for a '.' in the input; if present, it'll need to be
129 swapped for the current locale's decimal point before we
130 call strtod. On the other hand, if we find the current
131 locale's decimal point then the input is invalid. */
132 while (Py_ISDIGIT(*p))
133 p++;
135 if (*p == '.')
137 decimal_point_pos = p++;
139 /* locate end of number */
140 while (Py_ISDIGIT(*p))
141 p++;
143 if (*p == 'e' || *p == 'E')
144 p++;
145 if (*p == '+' || *p == '-')
146 p++;
147 while (Py_ISDIGIT(*p))
148 p++;
149 end = p;
151 else if (strncmp(p, decimal_point, decimal_point_len) == 0)
152 /* Python bug #1417699 */
153 goto invalid_string;
154 /* For the other cases, we need not convert the decimal
155 point */
158 if (decimal_point_pos) {
159 char *copy, *c;
160 /* Create a copy of the input, with the '.' converted to the
161 locale-specific decimal point */
162 copy = (char *)PyMem_MALLOC(end - digits_pos +
163 1 + decimal_point_len);
164 if (copy == NULL) {
165 if (endptr)
166 *endptr = (char *)nptr;
167 errno = ENOMEM;
168 return val;
171 c = copy;
172 memcpy(c, digits_pos, decimal_point_pos - digits_pos);
173 c += decimal_point_pos - digits_pos;
174 memcpy(c, decimal_point, decimal_point_len);
175 c += decimal_point_len;
176 memcpy(c, decimal_point_pos + 1,
177 end - (decimal_point_pos + 1));
178 c += end - (decimal_point_pos + 1);
179 *c = 0;
181 val = strtod(copy, &fail_pos);
183 if (fail_pos)
185 if (fail_pos > decimal_point_pos)
186 fail_pos = (char *)digits_pos +
187 (fail_pos - copy) -
188 (decimal_point_len - 1);
189 else
190 fail_pos = (char *)digits_pos +
191 (fail_pos - copy);
194 PyMem_FREE(copy);
197 else {
198 val = strtod(digits_pos, &fail_pos);
201 if (fail_pos == digits_pos)
202 goto invalid_string;
204 got_val:
205 if (negate && fail_pos != nptr)
206 val = -val;
208 if (endptr)
209 *endptr = fail_pos;
211 return val;
213 invalid_string:
214 if (endptr)
215 *endptr = (char*)nptr;
216 errno = EINVAL;
217 return -1.0;
220 double
221 PyOS_ascii_atof(const char *nptr)
223 return PyOS_ascii_strtod(nptr, NULL);
227 /* Given a string that may have a decimal point in the current
228 locale, change it back to a dot. Since the string cannot get
229 longer, no need for a maximum buffer size parameter. */
230 Py_LOCAL_INLINE(void)
231 change_decimal_from_locale_to_dot(char* buffer)
233 struct lconv *locale_data = localeconv();
234 const char *decimal_point = locale_data->decimal_point;
236 if (decimal_point[0] != '.' || decimal_point[1] != 0) {
237 size_t decimal_point_len = strlen(decimal_point);
239 if (*buffer == '+' || *buffer == '-')
240 buffer++;
241 while (Py_ISDIGIT(*buffer))
242 buffer++;
243 if (strncmp(buffer, decimal_point, decimal_point_len) == 0) {
244 *buffer = '.';
245 buffer++;
246 if (decimal_point_len > 1) {
247 /* buffer needs to get smaller */
248 size_t rest_len = strlen(buffer +
249 (decimal_point_len - 1));
250 memmove(buffer,
251 buffer + (decimal_point_len - 1),
252 rest_len);
253 buffer[rest_len] = 0;
260 Py_LOCAL_INLINE(void)
261 ensure_sign(char* buffer, size_t buf_size)
263 size_t len;
265 if (buffer[0] == '-')
266 /* Already have a sign. */
267 return;
269 /* Include the trailing 0 byte. */
270 len = strlen(buffer)+1;
271 if (len >= buf_size+1)
272 /* No room for the sign, don't do anything. */
273 return;
275 memmove(buffer+1, buffer, len);
276 buffer[0] = '+';
279 /* From the C99 standard, section 7.19.6:
280 The exponent always contains at least two digits, and only as many more digits
281 as necessary to represent the exponent.
283 #define MIN_EXPONENT_DIGITS 2
285 /* Ensure that any exponent, if present, is at least MIN_EXPONENT_DIGITS
286 in length. */
287 Py_LOCAL_INLINE(void)
288 ensure_minimum_exponent_length(char* buffer, size_t buf_size)
290 char *p = strpbrk(buffer, "eE");
291 if (p && (*(p + 1) == '-' || *(p + 1) == '+')) {
292 char *start = p + 2;
293 int exponent_digit_cnt = 0;
294 int leading_zero_cnt = 0;
295 int in_leading_zeros = 1;
296 int significant_digit_cnt;
298 /* Skip over the exponent and the sign. */
299 p += 2;
301 /* Find the end of the exponent, keeping track of leading
302 zeros. */
303 while (*p && Py_ISDIGIT(*p)) {
304 if (in_leading_zeros && *p == '0')
305 ++leading_zero_cnt;
306 if (*p != '0')
307 in_leading_zeros = 0;
308 ++p;
309 ++exponent_digit_cnt;
312 significant_digit_cnt = exponent_digit_cnt - leading_zero_cnt;
313 if (exponent_digit_cnt == MIN_EXPONENT_DIGITS) {
314 /* If there are 2 exactly digits, we're done,
315 regardless of what they contain */
317 else if (exponent_digit_cnt > MIN_EXPONENT_DIGITS) {
318 int extra_zeros_cnt;
320 /* There are more than 2 digits in the exponent. See
321 if we can delete some of the leading zeros */
322 if (significant_digit_cnt < MIN_EXPONENT_DIGITS)
323 significant_digit_cnt = MIN_EXPONENT_DIGITS;
324 extra_zeros_cnt = exponent_digit_cnt -
325 significant_digit_cnt;
327 /* Delete extra_zeros_cnt worth of characters from the
328 front of the exponent */
329 assert(extra_zeros_cnt >= 0);
331 /* Add one to significant_digit_cnt to copy the
332 trailing 0 byte, thus setting the length */
333 memmove(start,
334 start + extra_zeros_cnt,
335 significant_digit_cnt + 1);
337 else {
338 /* If there are fewer than 2 digits, add zeros
339 until there are 2, if there's enough room */
340 int zeros = MIN_EXPONENT_DIGITS - exponent_digit_cnt;
341 if (start + zeros + exponent_digit_cnt + 1
342 < buffer + buf_size) {
343 memmove(start + zeros, start,
344 exponent_digit_cnt + 1);
345 memset(start, '0', zeros);
351 /* Remove trailing zeros after the decimal point from a numeric string; also
352 remove the decimal point if all digits following it are zero. The numeric
353 string must end in '\0', and should not have any leading or trailing
354 whitespace. Assumes that the decimal point is '.'. */
355 Py_LOCAL_INLINE(void)
356 remove_trailing_zeros(char *buffer)
358 char *old_fraction_end, *new_fraction_end, *end, *p;
360 p = buffer;
361 if (*p == '-' || *p == '+')
362 /* Skip leading sign, if present */
363 ++p;
364 while (Py_ISDIGIT(*p))
365 ++p;
367 /* if there's no decimal point there's nothing to do */
368 if (*p++ != '.')
369 return;
371 /* scan any digits after the point */
372 while (Py_ISDIGIT(*p))
373 ++p;
374 old_fraction_end = p;
376 /* scan up to ending '\0' */
377 while (*p != '\0')
378 p++;
379 /* +1 to make sure that we move the null byte as well */
380 end = p+1;
382 /* scan back from fraction_end, looking for removable zeros */
383 p = old_fraction_end;
384 while (*(p-1) == '0')
385 --p;
386 /* and remove point if we've got that far */
387 if (*(p-1) == '.')
388 --p;
389 new_fraction_end = p;
391 memmove(new_fraction_end, old_fraction_end, end-old_fraction_end);
394 /* Ensure that buffer has a decimal point in it. The decimal point will not
395 be in the current locale, it will always be '.'. Don't add a decimal point
396 if an exponent is present. Also, convert to exponential notation where
397 adding a '.0' would produce too many significant digits (see issue 5864).
399 Returns a pointer to the fixed buffer, or NULL on failure.
401 Py_LOCAL_INLINE(char *)
402 ensure_decimal_point(char* buffer, size_t buf_size, int precision)
404 int digit_count, insert_count = 0, convert_to_exp = 0;
405 char* chars_to_insert, *digits_start;
407 /* search for the first non-digit character */
408 char *p = buffer;
409 if (*p == '-' || *p == '+')
410 /* Skip leading sign, if present. I think this could only
411 ever be '-', but it can't hurt to check for both. */
412 ++p;
413 digits_start = p;
414 while (*p && Py_ISDIGIT(*p))
415 ++p;
416 digit_count = Py_SAFE_DOWNCAST(p - digits_start, Py_ssize_t, int);
418 if (*p == '.') {
419 if (Py_ISDIGIT(*(p+1))) {
420 /* Nothing to do, we already have a decimal
421 point and a digit after it */
423 else {
424 /* We have a decimal point, but no following
425 digit. Insert a zero after the decimal. */
426 /* can't ever get here via PyOS_double_to_string */
427 assert(precision == -1);
428 ++p;
429 chars_to_insert = "0";
430 insert_count = 1;
433 else if (!(*p == 'e' || *p == 'E')) {
434 /* Don't add ".0" if we have an exponent. */
435 if (digit_count == precision) {
436 /* issue 5864: don't add a trailing .0 in the case
437 where the '%g'-formatted result already has as many
438 significant digits as were requested. Switch to
439 exponential notation instead. */
440 convert_to_exp = 1;
441 /* no exponent, no point, and we shouldn't land here
442 for infs and nans, so we must be at the end of the
443 string. */
444 assert(*p == '\0');
446 else {
447 assert(precision == -1 || digit_count < precision);
448 chars_to_insert = ".0";
449 insert_count = 2;
452 if (insert_count) {
453 size_t buf_len = strlen(buffer);
454 if (buf_len + insert_count + 1 >= buf_size) {
455 /* If there is not enough room in the buffer
456 for the additional text, just skip it. It's
457 not worth generating an error over. */
459 else {
460 memmove(p + insert_count, p,
461 buffer + strlen(buffer) - p + 1);
462 memcpy(p, chars_to_insert, insert_count);
465 if (convert_to_exp) {
466 int written;
467 size_t buf_avail;
468 p = digits_start;
469 /* insert decimal point */
470 assert(digit_count >= 1);
471 memmove(p+2, p+1, digit_count); /* safe, but overwrites nul */
472 p[1] = '.';
473 p += digit_count+1;
474 assert(p <= buf_size+buffer);
475 buf_avail = buf_size+buffer-p;
476 if (buf_avail == 0)
477 return NULL;
478 /* Add exponent. It's okay to use lower case 'e': we only
479 arrive here as a result of using the empty format code or
480 repr/str builtins and those never want an upper case 'E' */
481 written = PyOS_snprintf(p, buf_avail, "e%+.02d", digit_count-1);
482 if (!(0 <= written &&
483 written < Py_SAFE_DOWNCAST(buf_avail, size_t, int)))
484 /* output truncated, or something else bad happened */
485 return NULL;
486 remove_trailing_zeros(buffer);
488 return buffer;
491 /* see FORMATBUFLEN in unicodeobject.c */
492 #define FLOAT_FORMATBUFLEN 120
495 * _PyOS_ascii_formatd:
496 * @buffer: A buffer to place the resulting string in
497 * @buf_size: The length of the buffer.
498 * @format: The printf()-style format to use for the
499 * code to use for converting.
500 * @d: The #gdouble to convert
502 * Converts a #gdouble to a string, using the '.' as
503 * decimal point. To format the number you pass in
504 * a printf()-style format string. Allowed conversion
505 * specifiers are 'e', 'E', 'f', 'F', 'g', 'G', and 'Z'.
507 * 'Z' is the same as 'g', except it always has a decimal and
508 * at least one digit after the decimal.
510 * Return value: The pointer to the buffer with the converted string.
511 * On failure returns NULL but does not set any Python exception.
513 /* DEPRECATED, will be deleted in 2.8 and 3.2 */
514 PyAPI_FUNC(char *)
515 PyOS_ascii_formatd(char *buffer,
516 size_t buf_size,
517 const char *format,
518 double d)
520 char format_char;
521 size_t format_len = strlen(format);
523 /* Issue 2264: code 'Z' requires copying the format. 'Z' is 'g', but
524 also with at least one character past the decimal. */
525 char tmp_format[FLOAT_FORMATBUFLEN];
527 if (PyErr_WarnEx(PyExc_DeprecationWarning,
528 "PyOS_ascii_formatd is deprecated, "
529 "use PyOS_double_to_string instead", 1) < 0)
530 return NULL;
532 /* The last character in the format string must be the format char */
533 format_char = format[format_len - 1];
535 if (format[0] != '%')
536 return NULL;
538 /* I'm not sure why this test is here. It's ensuring that the format
539 string after the first character doesn't have a single quote, a
540 lowercase l, or a percent. This is the reverse of the commented-out
541 test about 10 lines ago. */
542 if (strpbrk(format + 1, "'l%"))
543 return NULL;
545 /* Also curious about this function is that it accepts format strings
546 like "%xg", which are invalid for floats. In general, the
547 interface to this function is not very good, but changing it is
548 difficult because it's a public API. */
550 if (!(format_char == 'e' || format_char == 'E' ||
551 format_char == 'f' || format_char == 'F' ||
552 format_char == 'g' || format_char == 'G' ||
553 format_char == 'Z'))
554 return NULL;
556 /* Map 'Z' format_char to 'g', by copying the format string and
557 replacing the final char with a 'g' */
558 if (format_char == 'Z') {
559 if (format_len + 1 >= sizeof(tmp_format)) {
560 /* The format won't fit in our copy. Error out. In
561 practice, this will never happen and will be
562 detected by returning NULL */
563 return NULL;
565 strcpy(tmp_format, format);
566 tmp_format[format_len - 1] = 'g';
567 format = tmp_format;
571 /* Have PyOS_snprintf do the hard work */
572 PyOS_snprintf(buffer, buf_size, format, d);
574 /* Do various fixups on the return string */
576 /* Get the current locale, and find the decimal point string.
577 Convert that string back to a dot. */
578 change_decimal_from_locale_to_dot(buffer);
580 /* If an exponent exists, ensure that the exponent is at least
581 MIN_EXPONENT_DIGITS digits, providing the buffer is large enough
582 for the extra zeros. Also, if there are more than
583 MIN_EXPONENT_DIGITS, remove as many zeros as possible until we get
584 back to MIN_EXPONENT_DIGITS */
585 ensure_minimum_exponent_length(buffer, buf_size);
587 /* If format_char is 'Z', make sure we have at least one character
588 after the decimal point (and make sure we have a decimal point);
589 also switch to exponential notation in some edge cases where the
590 extra character would produce more significant digits that we
591 really want. */
592 if (format_char == 'Z')
593 buffer = ensure_decimal_point(buffer, buf_size, -1);
595 return buffer;
598 PyAPI_FUNC(void)
599 _PyOS_double_to_string(char *buf, size_t buf_len, double val,
600 char format_code, int precision,
601 int flags, int *ptype)
603 char format[32];
604 int t;
605 int upper = 0;
607 if (buf_len < 1) {
608 assert(0);
609 /* There's no way to signal this error. Just return. */
610 return;
612 buf[0] = 0;
614 /* Validate format_code, and map upper and lower case */
615 switch (format_code) {
616 case 'e': /* exponent */
617 case 'f': /* fixed */
618 case 'g': /* general */
619 break;
620 case 'E':
621 upper = 1;
622 format_code = 'e';
623 break;
624 case 'F':
625 upper = 1;
626 format_code = 'f';
627 break;
628 case 'G':
629 upper = 1;
630 format_code = 'g';
631 break;
632 case 'r': /* repr format */
633 /* Supplied precision is unused, must be 0. */
634 if (precision != 0)
635 return;
636 precision = 17;
637 format_code = 'g';
638 break;
639 case 's': /* str format */
640 /* Supplied precision is unused, must be 0. */
641 if (precision != 0)
642 return;
643 precision = 12;
644 format_code = 'g';
645 break;
646 default:
647 assert(0);
648 return;
651 /* Check for buf too small to fit "-inf". Other buffer too small
652 conditions are dealt with when converting or formatting finite
653 numbers. */
654 if (buf_len < 5) {
655 assert(0);
656 return;
659 /* Handle nan and inf. */
660 if (Py_IS_NAN(val)) {
661 strcpy(buf, "nan");
662 t = Py_DTST_NAN;
663 } else if (Py_IS_INFINITY(val)) {
664 if (copysign(1., val) == 1.)
665 strcpy(buf, "inf");
666 else
667 strcpy(buf, "-inf");
668 t = Py_DTST_INFINITE;
669 } else {
670 t = Py_DTST_FINITE;
672 /* Build the format string. */
673 PyOS_snprintf(format, sizeof(format), "%%%s.%i%c",
674 (flags & Py_DTSF_ALT ? "#" : ""), precision,
675 format_code);
677 /* Have PyOS_snprintf do the hard work. */
678 PyOS_snprintf(buf, buf_len, format, val);
680 /* Do various fixups on the return string */
682 /* Get the current locale, and find the decimal point string.
683 Convert that string back to a dot. */
684 change_decimal_from_locale_to_dot(buf);
686 /* If an exponent exists, ensure that the exponent is at least
687 MIN_EXPONENT_DIGITS digits, providing the buffer is large
688 enough for the extra zeros. Also, if there are more than
689 MIN_EXPONENT_DIGITS, remove as many zeros as possible until
690 we get back to MIN_EXPONENT_DIGITS */
691 ensure_minimum_exponent_length(buf, buf_len);
693 /* Possibly make sure we have at least one character after the
694 decimal point (and make sure we have a decimal point). */
695 if (flags & Py_DTSF_ADD_DOT_0)
696 buf = ensure_decimal_point(buf, buf_len, precision);
699 /* Add the sign if asked and the result isn't negative. */
700 if (flags & Py_DTSF_SIGN && buf[0] != '-')
701 ensure_sign(buf, buf_len);
703 if (upper) {
704 /* Convert to upper case. */
705 char *p;
706 for (p = buf; *p; p++)
707 *p = toupper(*p);
710 if (ptype)
711 *ptype = t;
715 PyAPI_FUNC(char *) PyOS_double_to_string(double val,
716 char format_code,
717 int precision,
718 int flags,
719 int *ptype)
721 char buf[128];
722 Py_ssize_t len;
723 char *result;
725 _PyOS_double_to_string(buf, sizeof(buf), val, format_code, precision,
726 flags, ptype);
727 len = strlen(buf);
728 if (len == 0) {
729 PyErr_BadInternalCall();
730 return NULL;
733 /* Add 1 for the trailing 0 byte. */
734 result = PyMem_Malloc(len + 1);
735 if (result == NULL) {
736 PyErr_NoMemory();
737 return NULL;
739 strcpy(result, buf);
741 return result;