src/vasprintf.c - master - niceSOFT/mpfr

src/vasprintf.c
2 600 строк88 KB
Vincent Lefevre
[src/vasprintf.c] Fixed MPFR_LOG_FUNC usage.
27 фев 2026, 14:40
27 фев 2026, 14:402ac982b
/* mpfr_vasnprintf_aux -- helper function for the formatted output functions
   (printf functions family).

Copyright 2007-2026 Free Software Foundation, Inc.
Contributed by the Pascaline and Caramba projects, INRIA.

This file is part of the GNU MPFR Library.

The GNU MPFR Library is free software; you can redistribute it and/or modify
it under the terms of the GNU Lesser General Public License as published by
the Free Software Foundation; either version 3 of the License, or (at your
option) any later version.

The GNU MPFR Library is distributed in the hope that it will be useful, but
WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY
or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU Lesser General Public
License for more details.

You should have received a copy of the GNU Lesser General Public License
along with the GNU MPFR Library; see the file COPYING.LESSER.
If not, see <https://www.gnu.org/licenses/>. */

/* If the number of output characters is larger than INT_MAX, the
   ISO C99 / C11 standards are silent, but POSIX[*] requires the
   function to return a negative value and set errno to EOVERFLOW.
   [*] The Open Group Base Specifications Issue 7, 2018 edition
       IEEE Std 1003.1-2017 (Revision of IEEE Std 1003.1-2008)
   https://pubs.opengroup.org/onlinepubs/9699919799/functions/fprintf.html
   This follows a defect report submitted in 2007 to austin-review-l.
   Even in case of such a failure (just because of the limitation on int),
   we try to support %n, %ln, %jn when possible. That's why the sizes (or
   lengths) are expressed using mpfr_intmax_t in the code below. */

/* Notes about limitations on some platforms:

   Due to limitations from the C standard and GMP, if size_t < unsigned int
   (which is allowed by the C standard but unlikely to occur on any
   platform), the behavior is undefined for output that would reach
   SIZE_MAX = (size_t) -1 (if the result cannot be delivered, there should
   be an assertion failure, but this could not be tested).

   The stdarg(3) Linux man page says:
      On some systems, va_end contains a closing '}' matching a '{' in
      va_start, so that both macros must occur in the same function,
      and in a way that allows this.
   However, the only requirement from ISO C is that both macros must be
   invoked in the same function (MPFR uses va_copy instead of va_start,
   but the requirement is the same). Here, MPFR just follows ISO C.
*/

/* FIXME: For the native types, the gmp_asprintf and gmp_vasprintf functions
   are currently used, thus a full output is generated before being truncated
   in case of mpfr_snprintf(). This is potentially very inefficient and could
   unnecessarily exhaust the memory and/or crash.
   (Not tested yet.)

   A solution could be the use of gmp_snprintf / gmp_vsnprintf, but
   note that as a consequence, the case n > buffer size might not be
   supported everywhere. This would affect the snprintf_size() test
   in "tests/tsprintf.c". See discussion
     https://sourceware.org/pipermail/libc-alpha/2023-March/146343.html
   and about MPFR:
     https://sourceware.org/pipermail/libc-alpha/2023-March/146515.html

   This should be better, though the implementation of the standard
   formatted output functions may be inefficient, but this would be
   the best we can do. For instance, for the GNU C Library (glibc):
     https://sourceware.org/bugzilla/show_bug.cgi?id=17829
     https://sourceware.org/bugzilla/show_bug.cgi?id=21127
*/

/* Needed due to the tests on HAVE_STDARG and MPFR_USE_MINI_GMP */
#ifdef HAVE_CONFIG_H
# include "config.h"
#endif

/* The mpfr_printf-like functions are defined only if <stdarg.h> exists.
   Since they use mpf_t, they cannot be defined with mini-gmp. */
#if defined(HAVE_STDARG) && !defined(MPFR_USE_MINI_GMP)

#include <stdarg.h>

#ifndef HAVE_VA_COPY
# ifdef HAVE___VA_COPY
#  define va_copy(dst,src) __va_copy(dst, src)
# else
/* autoconf manual advocates this fallback.
   This is also the solution chosen by gmp */
#  define va_copy(dst,src) \
  do { memcpy(&(dst), &(src), sizeof(va_list)); } while (0)
# endif /* HAVE___VA_COPY */
#endif /* HAVE_VA_COPY */

#ifdef HAVE_WCHAR_H
#include <wchar.h>
#endif

#if defined (__cplusplus)
#include <cstddef>
#else
#include <stddef.h>             /* for ptrdiff_t */
#endif

#include <errno.h>

#define MPFR_NEED_LONGLONG_H
#define MPFR_NEED_INTMAX_H
#include "mpfr-impl.h"

/* Define a length modifier corresponding to mpfr_prec_t.
   We use literal string instead of literal character so as to permit future
   extension to long long int ("ll"). */
#if   _MPFR_PREC_FORMAT == 1
#define MPFR_PREC_FORMAT_TYPE "h"
#define MPFR_PREC_FORMAT_SIZE 1
#elif _MPFR_PREC_FORMAT == 2
#define MPFR_PREC_FORMAT_TYPE ""
#define MPFR_PREC_FORMAT_SIZE 0
#elif _MPFR_PREC_FORMAT == 3
#define MPFR_PREC_FORMAT_TYPE "l"
#define MPFR_PREC_FORMAT_SIZE 1
#else
#error "mpfr_prec_t size not supported"
#endif

/* Output for special values defined in the C99 standard */
#define MPFR_NAN_STRING_LC "nan"
#define MPFR_NAN_STRING_UC "NAN"
#define MPFR_NAN_STRING_LENGTH 3
#define MPFR_INF_STRING_LC "inf"
#define MPFR_INF_STRING_UC "INF"
#define MPFR_INF_STRING_LENGTH 3

#define DEFAULT_DECIMAL_PREC 6

/* The implicit \0 is useless, but we do not write num_to_text[16]
   otherwise g++ complains. */
static const char num_to_text[] = "0123456789abcdef";

/* some macro and functions for parsing format string */

/* Read an integer var of type mpfr_intmax_t. In case of overflow, set
   overflow to 1.
   The variable var must be 0 on input. If there are no digits, it is
   left to 0.
   This macro will be used to read the field width and the precision.
   The behavior will be similar to ISO C99. Note that unless "*" is
   used, the result will be non-negative (ISO C99 and C11 just specify
   "optional decimal integer" for the precision, but the behavior with
   a hardcoded negative integer is not explicitly defined, thus it is
   undefined, so that it is fine to reject such integers; the C2x draft
   now clarifies this: "an optional non-negative decimal integer").
   Note: Since mpfr_intmax_t = int is theoretically possible, all values
   of var are potentially valid values (via '*'). Hence the need of an
   overflow flag instead of a special value that would indicate overflow.
   Just saturating would not be OK either as the maximum value could be
   meaningful with %jn and/or in the case mpfr_intmax_t = int, for
   MPFR_PREC_ARG, i.e. one must be able to distinguish the maximum value
   from an overflow.
*/
#define READ_INT(ap, format, var)                                       \
  do {                                                                  \
    MPFR_ASSERTD ((var) == 0);                                          \
    if (*(format) == '*')                                               \
      {                                                                 \
        (var) = va_arg ((ap), int);                                     \
        ++(format);                                                     \
      }                                                                 \
    else                                                                \
      for ( ; *(format) >= '0' && *(format) <= '9' ; ++(format))        \
        if (!(overflow))                                                \
          {                                                             \
            if ((var) > MPFR_INTMAX_MAX / 10)                           \
              (overflow) = 1;                                           \
            else                                                        \
              {                                                         \
                int _i;                                                 \
                (var) *= 10;                                            \
                _i = *(format) - '0';                                   \
                MPFR_ASSERTN (_i >= 0 && _i <= 9);                      \
                if ((var) > MPFR_INTMAX_MAX - _i)                       \
                  (overflow) = 1;                                       \
                else                                                    \
                  (var) += _i;                                          \
              }                                                         \
          }                                                             \
  } while (0)

/* arg_t contains all the types described by the 'type' field of the
   format string */
enum arg_t
  {
    NONE,
    CHAR_ARG,
    SHORT_ARG,
    LONG_ARG,
    LONG_LONG_ARG,
    INTMAX_ARG,
    SIZE_ARG,
    PTRDIFF_ARG,
    LONG_DOUBLE_ARG,
    MPF_ARG,
    MPQ_ARG,
    MP_LIMB_ARG,
    MP_LIMB_ARRAY_ARG,
    MPZ_ARG,
    MPFR_PREC_ARG,
    MPFR_ARG,
    UNSUPPORTED
  };

/* Each conversion specification of the format string will be translated
   into a printf_spec structure by the parser.
   This structure is adapted from the GNU libc one (and augmented). */
struct printf_spec
{
  unsigned int alt:1;           /* # flag */
  unsigned int space:1;         /* Space flag */
  unsigned int left:1;          /* - flag */
  unsigned int showsign:1;      /* + flag */
  unsigned int group:1;         /* ' flag */

  mpfr_intmax_t width;          /* Width */
  mpfr_intmax_t prec;           /* Precision, or negative if omitted */
  int size;                     /* If 0 (i.e. snprintf with size = 0),
                                   no output. */

  enum arg_t arg_type;          /* Type of argument */
  mpfr_rnd_t rnd_mode;          /* Rounding mode */
  char spec;                    /* Conversion specifier */

  char pad;                     /* Padding character */
};

/* Note: LONG_ARG is unusual, but is accepted (ISO C99 says "as no effect
   on a following a, A, e, E, f, F, g, or G conversion specifier"). */
#define FLOATING_POINT_ARG_TYPE(at) \
  ((at) == MPFR_ARG || (at) == MPF_ARG \
   || (at) == LONG_ARG || (at) == LONG_DOUBLE_ARG)

#define INTEGER_LIKE_ARG_TYPE(at)                                       \
  ((at) == SHORT_ARG || (at) == LONG_ARG || (at) == LONG_LONG_ARG       \
   || (at) == INTMAX_ARG  || (at) == MPFR_PREC_ARG || (at) == MPZ_ARG   \
   || (at) == MPQ_ARG || (at) == MP_LIMB_ARG || (at) == MP_LIMB_ARRAY_ARG \
   || (at) == CHAR_ARG || (at) == SIZE_ARG || (at) == PTRDIFF_ARG)

static int
specinfo_is_valid (struct printf_spec spec)
{
  switch (spec.spec)
    {
    case 'n':
      return -1;

    case 'a':    case 'A':
    case 'e':    case 'E':
    case 'f':    /* 'F': see below */
    case 'g':    case 'G':
      return (spec.arg_type == NONE
              || FLOATING_POINT_ARG_TYPE (spec.arg_type));

    case 'F':  /* only MPFR_ARG is supported since GMP doesn't support it
                  due to its use as the mpf_t type specifier */
    case 'b':
      return spec.arg_type == MPFR_ARG;

    case 'd':    case 'i':
    case 'o':    case 'u':
    case 'x':    case 'X':
      return (spec.arg_type == NONE
              || INTEGER_LIKE_ARG_TYPE (spec.arg_type));

    case 'c':
    case 's':
      return (spec.arg_type == NONE || spec.arg_type == LONG_ARG);

    case 'p':
      return spec.arg_type == NONE;

    default:
      return 0;
    }
}

/* Note: additional flags should be added to the MPFR_PREC_ARG code
   for gmp_asprintf (when supported). */
MPFR_RETURNS_NONNULL static const char *
parse_flags (const char *format, struct printf_spec *specinfo)
{
  while (*format)
    {
      switch (*format)
        {
        case '0':
          specinfo->pad = '0';
          ++format;
          break;
        case '#':
          specinfo->alt = 1;
          ++format;
          break;
        case '+':
          specinfo->showsign = 1;
          ++format;
          break;
        case ' ':
          specinfo->space = 1;
          ++format;
          break;
        case '-':
          specinfo->left = 1;
          ++format;
          break;
        case '\'':
          /* Single UNIX Specification for thousand separator */
          specinfo->group = 1;
          ++format;
          break;
        default:
          return format;
        }
    }
  return format;
}

MPFR_RETURNS_NONNULL static const char *
parse_arg_type (const char *format, struct printf_spec *specinfo)
{
  switch (*format)
    {
    case '\0':
      break;
    case 'h':
      if (*++format == 'h')
        {
          ++format;
          specinfo->arg_type = CHAR_ARG;
        }
      else
        specinfo->arg_type = SHORT_ARG;
      break;
    case 'l':
      if (*++format == 'l')
        {
          ++format;
#if defined (HAVE_LONG_LONG)
          specinfo->arg_type = LONG_LONG_ARG;
#else
          specinfo->arg_type = UNSUPPORTED;
#endif
          break;
        }
      else
        {
          specinfo->arg_type = LONG_ARG;
          break;
        }
    case 'j':
      ++format;
#if defined(_MPFR_H_HAVE_INTMAX_T)
      specinfo->arg_type = INTMAX_ARG;
#else
      specinfo->arg_type = UNSUPPORTED;
#endif
      break;
    case 'z':
      ++format;
      specinfo->arg_type = SIZE_ARG;
      break;
    case 't':
      ++format;
      specinfo->arg_type = PTRDIFF_ARG;
      break;
    case 'L':
      ++format;
      specinfo->arg_type = LONG_DOUBLE_ARG;
      break;
    case 'F':
      ++format;
      specinfo->arg_type = MPF_ARG;
      break;
    case 'Q':
      ++format;
      specinfo->arg_type = MPQ_ARG;
      break;
    case 'M':
      ++format;
      /* The 'M' specifier was added in gmp 4.2.0 */
      specinfo->arg_type = MP_LIMB_ARG;
      break;
    case 'N':
      ++format;
      specinfo->arg_type = MP_LIMB_ARRAY_ARG;
      break;
    case 'Z':
      ++format;
      specinfo->arg_type = MPZ_ARG;
      break;

      /* mpfr specific specifiers */
    case 'P':
      ++format;
      specinfo->arg_type = MPFR_PREC_ARG;
      break;
    case 'R':
      ++format;
      specinfo->arg_type = MPFR_ARG;
    }
  return format;
}


/* some macros and functions filling the buffer */

/* CONSUME_VA_ARG removes from va_list ap the type expected by specinfo */

/* With a C++ compiler wchar_t and enumeration in va_list are converted to
   integer type : int, unsigned int, long or unsigned long (unfortunately,
   this is implementation dependent).
   We follow gmp which assumes in print/doprnt.c that wchar_t is converted
   to int (because wchar_t <= int).
   For wint_t, we assume that the case WINT_MAX < INT_MAX yields an
   integer promotion. */
#if defined(WINT_MAX) && WINT_MAX < INT_MAX
typedef int    mpfr_va_wint;  /* integer promotion */
#else
typedef wint_t mpfr_va_wint;
#endif
#define CASE_LONG_ARG(specinfo, ap)                                     \
  case LONG_ARG:                                                        \
  if ((specinfo).spec == 'd' || (specinfo).spec == 'i'                  \
      || (specinfo).spec == 'o' || (specinfo).spec == 'u'               \
      || (specinfo).spec == 'x' || (specinfo).spec == 'X')              \
    (void) va_arg ((ap), long);                                         \
  else if ((specinfo).spec == 'c')                                      \
    (void) va_arg ((ap), mpfr_va_wint);                                 \
  else if ((specinfo).spec == 's')                                      \
    (void) va_arg ((ap), int); /* we assume integer promotion */        \
  else if ((specinfo).spec == 'a' || (specinfo).spec == 'A'             \
           || (specinfo).spec == 'e' || (specinfo).spec == 'E'          \
           || (specinfo).spec == 'f' /* 'F' impossible */               \
           || (specinfo).spec == 'g' || (specinfo).spec == 'G')         \
    (void) va_arg ((ap), double);                                       \
  else                                                                  \
    MPFR_RET_NEVER_GO_HERE();                                           \
  break;

#if defined(_MPFR_H_HAVE_INTMAX_T)
#define CASE_INTMAX_ARG(specinfo, ap)           \
  case INTMAX_ARG:                              \
  (void) va_arg ((ap), intmax_t);               \
  break;
#else
#define CASE_INTMAX_ARG(specinfo, ap)
#endif

#ifdef HAVE_LONG_LONG
#define CASE_LONG_LONG_ARG(specinfo, ap)        \
  case LONG_LONG_ARG:                           \
  (void) va_arg ((ap), long long);              \
  break;
#else
#define CASE_LONG_LONG_ARG(specinfo, ap)
#endif

/* Note: (specinfo).width may be incorrect in case of overflow,
   but it is not used by CONSUME_VA_ARG. */
#define CONSUME_VA_ARG(specinfo, ap)            \
  do {                                          \
    switch ((specinfo).arg_type)                \
      {                                         \
      case CHAR_ARG:                            \
      case SHORT_ARG:                           \
        (void) va_arg ((ap), int);              \
        break;                                  \
      CASE_LONG_ARG (specinfo, ap)              \
      CASE_LONG_LONG_ARG (specinfo, ap)         \
      CASE_INTMAX_ARG (specinfo, ap)            \
      case SIZE_ARG:                            \
        (void) va_arg ((ap), size_t);           \
        break;                                  \
      case PTRDIFF_ARG:                         \
        (void) va_arg ((ap), ptrdiff_t);        \
        break;                                  \
      case LONG_DOUBLE_ARG:                     \
        (void) va_arg ((ap), long double);      \
        break;                                  \
      case MPF_ARG:                             \
        (void) va_arg ((ap), mpf_srcptr);       \
        break;                                  \
      case MPQ_ARG:                             \
        (void) va_arg ((ap), mpq_srcptr);       \
        break;                                  \
      case MP_LIMB_ARG:                         \
        (void) va_arg ((ap), mp_limb_t);        \
        break;                                  \
      case MP_LIMB_ARRAY_ARG:                   \
        (void) va_arg ((ap), mpfr_limb_ptr);    \
        (void) va_arg ((ap), mp_size_t);        \
        break;                                  \
      case MPZ_ARG:                             \
        (void) va_arg ((ap), mpz_srcptr);       \
        break;                                  \
      default:                                  \
        switch ((specinfo).spec)                \
          {                                     \
          case 'd':                             \
          case 'i':                             \
          case 'o':                             \
          case 'u':                             \
          case 'x':                             \
          case 'X':                             \
          case 'c':                             \
            (void) va_arg ((ap), int);          \
            break;                              \
          case 'a':                             \
          case 'A':                             \
          case 'e':                             \
          case 'E':                             \
          case 'f':                             \
          /* 'F' impossible */                  \
          case 'g':                             \
          case 'G':                             \
            (void) va_arg ((ap), double);       \
            break;                              \
          case 's':                             \
            (void) va_arg ((ap), char *);       \
            break;                              \
          case 'p':                             \
            (void) va_arg ((ap), void *);       \
          }                                     \
      }                                         \
  } while (0)

/* Process the format part that does not deal with mpfr types,
   from start to end (not included).
   Jump to external label 'error' if gmp_asprintf return -1.
   Note: start and end are pointers to the format string, so that
   size_t is the best type to express the difference.
   Warning! Since the output from gmp_vasprintf may contain non-terminating
   null characters (if %c is used with the value 0), the mpfr_free_str
   function must not be used to free the allocated memory, because the size
   may matter with some custom allocation functions. Anyway, mpfr_free_func
   is more efficient here, as the size does not need to be recomputed.
   FIXME: If buf.size = 0 or size != 0, gmp_vsnprintf should be called
   instead of gmp_vasprintf, outputting data directly to the buffer
   when applicable.
*/
#define FLUSH(flag, start, end, ap, buf_ptr)                            \
  do {                                                                  \
    const size_t n = (end) - (start);                                   \
    if ((flag))                                                         \
      /* previous specifiers are understood by gmp_printf */            \
      {                                                                 \
        MPFR_TMP_DECL (marker);                                         \
        char *fmt_copy, *s;                                             \
        int length;                                                     \
                                                                        \
        MPFR_TMP_MARK (marker);                                         \
        fmt_copy = (char *) MPFR_TMP_ALLOC (n + 1);                     \
        memcpy (fmt_copy, (start), n);                                  \
        fmt_copy[n] = '\0';                                             \
        length = gmp_vasprintf (&s, fmt_copy, (ap));                    \
        if (length < 0)                                                 \
          {                                                             \
            MPFR_TMP_FREE (marker);                                     \
            goto error;                                                 \
          }                                                             \
        buffer_cat ((buf_ptr), s, length);                              \
        mpfr_free_func (s, length + 1);                                 \
        (flag) = 0;                                                     \
        MPFR_TMP_FREE (marker);                                         \
      }                                                                 \
    else if ((start) != (end))                                          \
      /* no conversion specification, just simple characters */         \
      buffer_cat ((buf_ptr), (start), n);                               \
  } while (0)

/* Note: in case some form of %n is used in the format string,
   we may need the maximum signed integer type for len. */
struct string_buffer
{
  char *start;                  /* beginning of the buffer */
  char *curr;                   /* null terminating character */
  size_t size;                  /* buffer capacity */
  mpfr_intmax_t len;            /* string length or -1 if overflow */
};

static void
buffer_init (struct string_buffer *b, size_t s)
{
  if (s != 0)
    {
      b->start = (char *) mpfr_allocate_func (s);
      b->start[0] = '\0';
      b->curr = b->start;
    }
  b->size = s;
  b->len = 0;
}

/* Increase the len field of the buffer. Return non-zero iff overflow. */
static int
buffer_incr_len (struct string_buffer *b, mpfr_intmax_t len)
{
  if (b->len == -1)
    return 1;
  else
    {
      /* We need to take mpfr_uintmax_t as the type must be as large
         as both size_t (which is unsigned) and mpfr_intmax_t (which
         is used for the 'n' format specifier). */
      mpfr_uintmax_t newlen = (mpfr_uintmax_t) b->len + len;

      /* mpfr_uintmax_t is unsigned, thus the above is valid, but one
         has newlen < len in case of overflow. */

      if (MPFR_UNLIKELY (newlen < len || newlen > MPFR_INTMAX_MAX))
        {
          MPFR_LOG_MSG (("Overflow\n", 0));
          b->len = -1;
          return 1;
        }
      else
        {
          b->len = newlen;
          return 0;
        }
    }
}

/* Increase buffer size by a number of character being the least multiple of
   4096 greater than len+1. */
static void
buffer_widen (struct string_buffer *b, size_t len)
{
  const size_t pos = b->curr - b->start;
  const size_t n = 0x1000 + (len & ~((size_t) 0xfff));

  /* There are currently limitations here. We would need to switch to
     the null-size behavior once there is an overflow in the buffer. */

  MPFR_ASSERTN (n >= 0x1000 && n >= len);

  MPFR_ASSERTD (*b->curr == '\0');
  MPFR_ASSERTD (pos < b->size);

  MPFR_ASSERTN (b->size < ((size_t) -1) - n);

  b->start = (char *) mpfr_reallocate_func (b->start, b->size, b->size + n);
  b->size += n;
  b->curr = b->start + pos;

  MPFR_ASSERTD (pos < b->size);
  MPFR_ASSERTD (*b->curr == '\0');
}

/* Concatenate the first len characters of the array s to the buffer b,
   and expand it if needed. Return non-zero if overflow.
   Warning! The array s may contain null characters in addition to the
   terminating one, in case %c has been used with the value 0.
 */
static int
buffer_cat (struct string_buffer *b, const char *s, size_t len)
{
  /* If len == 0, which is possible when outputting an integer 0
     (either a native one or mpfr_prec_t) with precision field = 0,
     do nothing. This test is not necessary since the code below is
     valid for len == 0, but this is safer, just in case. */
  if (len == 0)
    return 0;

  if (buffer_incr_len (b, len))
    return 1;

  if (b->size != 0)
    {
      MPFR_ASSERTD (*b->curr == '\0');
      MPFR_ASSERTN (b->size < ((size_t) -1) - len);
      if (MPFR_UNLIKELY (b->curr + len >= b->start + b->size))
        buffer_widen (b, len);

      MPFR_ASSERTD (*b->curr == '\0');
      memcpy (b->curr, s, len);
      b->curr += len;
      MPFR_ASSERTD (b->curr < b->start + b->size);
      *b->curr = '\0';
    }

  return 0;
}

/* Add n characters c to the end of buffer b. Return non-zero if overflow.
   Note: this function is called only when there is output (b->size != 0),
   since it is needed only by sprnt_fp(). */
static int
buffer_pad (struct string_buffer *b, const char c, const mpfr_intmax_t n)
{
  MPFR_ASSERTD (b->size != 0);
  MPFR_ASSERTD (n > 0);

  if (buffer_incr_len (b, n))
    return 1;

  MPFR_ASSERTD (*b->curr == '\0');

  if (n > (size_t) -1 || b->size > ((size_t) -1) - n)
    {
      /* Reallocation will not be possible. Regard this as an overflow. */
      b->len = -1;
      return 1;
    }

  if (MPFR_UNLIKELY (b->curr + n >= b->start + b->size))
    buffer_widen (b, n);

  if (n == 1)
    *b->curr = c;
  else
    memset (b->curr, c, n);
  b->curr += n;
  *b->curr = '\0';

  MPFR_ASSERTD (b->curr < b->start + b->size);

  return 0;
}

/* Form a string by concatenating the first len characters of str to tz
   zero(s), insert into one character c each 3 characters starting from end
   to beginning and concatenate the result to the buffer b.
   Assume c is not null (\0). Return non-zero if overflow.
   Note: this function is called only when there is output (b->size != 0),
   since it is needed only by sprnt_fp(). */
static int
buffer_sandwich (struct string_buffer *b, char *str, size_t len,
                 const size_t tz, const char c)
{
  const size_t step = 3;
  size_t size, q, r, fullsize, i;
  char *oldcurr;

  MPFR_ASSERTD (b->size != 0);
  MPFR_ASSERTD (tz == 0 || tz == 1);

  if (len <= ULONG_MAX)
    MPFR_LOG_MSG (("len=%lu\n", (unsigned long) len));
  if (tz <= ULONG_MAX)
    MPFR_LOG_MSG (("tz=%lu\n", (unsigned long) tz));

  MPFR_ASSERTD (len <= strlen (str));
  MPFR_ASSERTD (c != '\0');

  /* check that len + tz does not overflow */
  if (len > (size_t) -1 - tz)
    return 1;

  size = len + tz;              /* number of digits */
  MPFR_ASSERTD (size > 0);

  q = (size - 1) / step;        /* number of separators C */
  r = ((size - 1) % step) + 1;  /* number of digits in the leftmost block */
  MPFR_ASSERTD (r >= 1 && r <= step);

  /* check that size + q does not overflow */
  if (size > (size_t) -1 - q)
    return 1;

  fullsize = size + q;          /* number of digits and separators */

  if (buffer_incr_len (b, fullsize))
    return 1;

  MPFR_ASSERTD (*b->curr == '\0');
  MPFR_ASSERTN (b->size < ((size_t) -1) - fullsize);
  if (MPFR_UNLIKELY (b->curr + fullsize >= b->start + b->size))
    buffer_widen (b, fullsize);

  MPFR_DBGRES (oldcurr = b->curr);

  /* first r significant digits (leftmost block) */
  if (r <= len)
    {
      memcpy (b->curr, str, r);
      str += r;
      len -= r;
    }
  else
    {
      MPFR_ASSERTD (r > len);
      MPFR_ASSERTD (len < step);    /* as a consequence */
      MPFR_ASSERTD (size <= step);  /* as a consequence */
      MPFR_ASSERTD (q == 0);        /* as a consequence */
      MPFR_ASSERTD (r == size);     /* as a consequence */
      MPFR_ASSERTD (tz == 1);       /* as a consequence */
      memcpy (b->curr, str, len);
      *(b->curr + len) = '0';  /* trailing zero */
      /* We do not need to set len to 0 since it will not be read again
         (q = 0, so that the loop below will have 0 iterations). */
    }
  b->curr += r;

  for (i = 0; i < q; ++i)
    {
      *b->curr++ = c;
      if (MPFR_LIKELY (len >= step))
        {
          memcpy (b->curr, str, step);
          len -= step;
          str += step;
        }
      else
        {
          /* last digits */
          MPFR_ASSERTD (i == q - 1 && step - len == 1);
          memcpy (b->curr, str, len);
          *(b->curr + len) = '0';  /* trailing zero */
        }
      b->curr += step;
    }

  MPFR_ASSERTD (b->curr - oldcurr == fullsize);

  *b->curr = '\0';

  MPFR_ASSERTD (b->curr < b->start + b->size);

  return 0;
}

/* Helper struct and functions for temporary strings management */
/* struct for easy string clearing */
struct string_list
{
  char *string;
  struct string_list *next; /* NULL in last node */
};

/* initialization */
static void
init_string_list (struct string_list *sl)
{
  sl->string = NULL;
  sl->next = NULL;
}

/* clear all strings in the list */
static void
clear_string_list (struct string_list *sl)
{
  struct string_list *n;

  while (sl)
    {
      if (sl->string)
        mpfr_free_str (sl->string);
      n = sl->next;
      mpfr_free_func (sl, sizeof(struct string_list));
      sl = n;
    }
}

/* add a string in the list */
static char *
register_string (struct string_list *sl, char *new_string)
{
  /* look for the last node */
  while (sl->next)
    sl = sl->next;

  sl->next = (struct string_list *)
    mpfr_allocate_func (sizeof (struct string_list));

  sl = sl->next;
  sl->next = NULL;
  return sl->string = new_string;
}

/* padding type: where are the padding characters */
enum pad_t
  {
    LEFT,          /* spaces in left hand side for right justification */
    LEADING_ZEROS, /* padding with '0' characters in integral part */
    RIGHT          /* spaces in right hand side for left justification */
  };

/* number_parts details how much characters are needed in each part of a float
   print.  */
struct number_parts
{
  enum pad_t pad_type;    /* Padding type */
  mpfr_intmax_t pad_size; /* Number of padding characters */

  char sign;              /* Sign character ('-', '+', ' ', or '\0') */

  char *prefix_ptr;       /* Pointer to prefix part */
  size_t prefix_size;     /* Number of characters in prefix_ptr */

  char thousands_sep;     /* Thousands separator (only with style 'f') */

  /* The integral part is given by the following 3 members
     (excluding the possible padding with zeros). */
  char *ip_ptr;           /* Pointer to integral part characters */
  size_t ip_size;         /* Number of digits of the integral part */
  int ip_trailing_digits; /* Number of additional digits in integral part
                             (if spec.size != 0, this can only be a zero) */

  char point;             /* Decimal point character */

  mpfr_intmax_t fp_leading_zeros;  /* Number of additional leading zeros in
                                      the fractional part */
  char *fp_ptr;           /* Pointer to fractional part characters */
  size_t fp_size;         /* Number of digits of the fractional part,
                             without leading and trailing zeros */
  mpfr_intmax_t fp_trailing_zeros;  /* Number of additional trailing zeros in
                                       the fractional part */

  char *exp_ptr;          /* Pointer to exponent part */
  size_t exp_size;        /* Number of characters in exp_ptr */

  struct string_list *sl; /* List of string buffers in use: we need such a
                             mechanism because fp_ptr may point into the same
                             string as ip_ptr */
};

/* For a real non zero number x, what is the base exponent f when rounding x
   with rounding mode r to r(x) = m*b^f, where m is a digit and 1 <= m < b ?
   Return non zero value if x is rounded up to b^f, return zero otherwise */
/* FIXME: It seems that the base-2 exponent is taken into account, which is
   what is expected. In this case, the description is incorrect. */
static int
next_base_power_p (mpfr_srcptr x, int base, mpfr_rnd_t rnd)
{
  mpfr_prec_t nbits;
  mp_limb_t pm;
  mp_limb_t xm;

  MPFR_ASSERTD (MPFR_IS_PURE_FP (x));
  MPFR_ASSERTD (base == 2 || base == 16);

  /* Warning: the decimal point is AFTER THE FIRST DIGIT in this output
     representation. */
  nbits = base == 2 ? 1 : 4;

  if (rnd == MPFR_RNDZ
      || (rnd == MPFR_RNDD && MPFR_IS_POS (x))
      || (rnd == MPFR_RNDU && MPFR_IS_NEG (x))
      || MPFR_PREC (x) <= nbits)
    /* no rounding when printing x with 1 digit */
    return 0;

  xm = MPFR_MANT (x) [MPFR_LIMB_SIZE (x) - 1];
  pm = MPFR_LIMB_MASK (GMP_NUMB_BITS - nbits);
  if ((xm & ~pm) ^ ~pm)
    /* do no round up if some of the nbits first bits are 0s. */
    return 0;

  if (rnd == MPFR_RNDN)
    /* mask for rounding bit */
    pm = (MPFR_LIMB_ONE << (GMP_NUMB_BITS - nbits - 1));

  /* round up if some remaining bits are 1 */
  /* warning: the return value must be an int */
  return xm & pm ? 1 : 0;
}

/* Record information from mpfr_get_str() so as to avoid multiple
   calls to this expensive function. */
struct decimal_info
{
  mpfr_exp_t exp;
  char *str;
};

/* For a real non zero number x, what is the exponent f so that
   10^f <= x < 10^(f+1). */
static mpfr_exp_t
floor_log10 (mpfr_srcptr x)
{
  mpfr_t y;
  mpfr_exp_t exp;

  /* make sure first that y can represent a mpfr_exp_t exactly
     and can compare with x */
  mpfr_prec_t prec = sizeof (mpfr_exp_t) * CHAR_BIT;
  mpfr_init2 (y, MAX (prec, MPFR_PREC (x)));

  exp = mpfr_ceil_mul (MPFR_GET_EXP (x), 10, 1) - 1;
  mpfr_set_exp_t (y, exp, MPFR_RNDU);
  /* The following call to mpfr_ui_pow should be fast: y is an integer
     (not too large), so that mpfr_pow_z will be used internally. */
  mpfr_ui_pow (y, 10, y, MPFR_RNDU);
  if (mpfr_cmpabs (x, y) < 0)
    exp--;

  mpfr_clear (y);
  return exp;
}

#define NDIGITS 8

/* If output is needed (spec.size != 0), this is mpfr_get_str.
   Otherwise...
   TODO: Explain the meaning of the returned string. Is it really used?
   Note that this wrapper is not used for %Rg, since trailing zeros may
   need to be removed.
*/
MPFR_RETURNS_NONNULL static char *
mpfr_get_str_wrapper (mpfr_exp_t *exp, int base, size_t n, mpfr_srcptr op,
                      const struct printf_spec spec)
{
  size_t ndigits;
  char *str, *s, nine;
  int neg;

  /* Possibles bases for the *printf functions. */
  MPFR_ASSERTD (base == 2 || base == 10 || base == 16);

  if (spec.size != 0)
    return mpfr_get_str (NULL, exp, base, n, op, spec.rnd_mode);

  /* Special case size = 0, i.e., xxx_snprintf with size = 0: we only want
     to compute the number of printed characters. Try to deduce it from
     a small number of significant digits, and increase the number of
     digits until this is possible. */
  nine = base == 2 ? '1' : base == 10 ? '9' : 'f';
  for (ndigits = NDIGITS; ; ndigits *= 2)
    {
      mpfr_rnd_t rnd = MPFR_RNDZ;
      /* When ndigits > n, we reduce it to the target size n, and then we use
         the wanted rounding mode, to avoid errors for example when n=1 and
         x = 9.5 with spec.rnd_mode = RNDU.
         WARNING! If n == 0, then it is important that ndigits is set to 0
         too, so that the correct number of characters is generated. Indeed,
         if n == 0, the number of characters is determined by this function;
         and it is important that it is correct, since the value returned by
         mpfr_*snprintf (NULL, 0, "%Re", ...) will be based on it. */
      if (ndigits >= n)
        {
          ndigits = n;
          rnd = spec.rnd_mode;
        }
      str = mpfr_get_str (NULL, exp, base, ndigits, op, rnd);
      if (ndigits == n)
        break;
      neg = str[0] == '-';
      s = str + neg;
      while (*s == nine)
        s++;
      /* Due to the above code, ndigits < n, and in this case, it is
         important that ndigits is at least 2 in order to have a non-empty
         fractional part (see the warning comments later). */
      MPFR_ASSERTD (ndigits < n && ndigits >= 2);
      if (s < str + neg + ndigits) /* we don't have ndigits 'nines' */
        break;
      mpfr_free_str (str);
      MPFR_ASSERTN (ndigits <= ((size_t) -1) / 2);
      /* to make sure that the product by 2 is representable. */
    }
  return str;
}

/* Determine the different parts of the string representation of the regular
   number P when spec.spec is 'a', 'A', or 'b'.

   Return -1 in case of overflow on the sizes.

   Note for 'a'/'A': If the precision field is non-zero, the output is the
   one with a binary exponent that is a multiple of 4 (thus this is similar
   to base 16, where base-16 exponent = binary exponent / 4). But if the
   precision field is 0, the exponent is no longer restricted to a multiple
   of 4; the precision is maximized, but the displayed digit may be 1; this
   is completely unintuitive.
   The obtained output for 4 values with precision fields 0 and 1:
               0         1
      30     0xfp+1   0x1.ep+4
      31     0x1p+5   0x1.fp+4
      32     0x8p+2   0x2.0p+4
      33     0x8p+2   0x2.1p+4
   First, the output for numbers that round up to the next power of 16
   with a precision field 0, like 31 here, has an unexpected form: here
   with 31, "0x1p+5" instead of "0x8p+2".
   Moreover, if one increases the output precision, the output form
   changes (even if no rounding is involved). For instance, for 32,
   "0x8p+2" changes to "0x2.0p+4" instead of "0x8.0p+2".
   FIXME: choose first digit = always 1. Discussion:
     https://sympa.inria.fr/sympa/arc/mpfr/2021-05/msg00002.html
*/
static int
regular_ab (struct number_parts *np, mpfr_srcptr p,
            const struct printf_spec spec)
{
  int uppercase;
  int base;
  char *str;
  mpfr_exp_t exp;

  uppercase = spec.spec == 'A';

  if (spec.spec == 'a' || spec.spec == 'A')
    /* prefix part */
    {
      np->prefix_size = 2;
      str = (char *) mpfr_allocate_func (1 + np->prefix_size);
      str[0] = '0';
      str[1] = uppercase ? 'X' : 'x';
      str[2] = '\0';
      np->prefix_ptr = register_string (np->sl, str);
    }

  /* integral part */
  np->ip_size = 1;
  base = spec.spec == 'b' ? 2 : 16;

  if (spec.prec != 0)
    {
      size_t nsd;

      /* Number of significant digits:
         - if no given precision, let mpfr_get_str determine it;
         - if a non-zero precision is specified, then one digit before decimal
         point plus spec.prec after it (which will give nsd > 1 below). */
      MPFR_ASSERTD (np->ip_size == 1);  /* thus the + 1 below */
      if (spec.prec < 0)
        nsd = 0;
      else
        {
          if (MPFR_UNLIKELY (spec.prec > (size_t) -2))  /* overflow */
            return -1;
          nsd = (size_t) spec.prec + 1;
          MPFR_ASSERTD (nsd > 1);
        }
      str = mpfr_get_str_wrapper (&exp, base, nsd, p, spec);
      register_string (np->sl, str);
      if (MPFR_IS_NEG (p))
        str++;  /* skip sign */

      if (base == 16)
        /* exp is the exponent for radix sixteen with decimal point BEFORE the
           first digit, we want the exponent for radix two and the decimal
           point AFTER the first digit. */
        {
          /* An integer overflow is normally not possible since MPFR_EXP_MIN
             is twice as large as MPFR_EMIN_MIN. */
          MPFR_ASSERTN (exp > (MPFR_EXP_MIN + 3) / 4);
          exp = (exp - 1) * 4;
        }
      else
        /* exp is the exponent for decimal point BEFORE the first digit, we
           want the exponent for decimal point AFTER the first digit. */
        {
          /* An integer overflow is normally not possible since MPFR_EXP_MIN
             is twice as large as MPFR_EMIN_MIN. */
          MPFR_ASSERTN (exp > MPFR_EXP_MIN);
          --exp;
        }
    }
  else
    {
      exp = MPFR_GET_EXP (p);
      str = (char *) mpfr_allocate_func (2);
      register_string (np->sl, str);
      if (next_base_power_p (p, base, spec.rnd_mode))
        {
          str[0] = '1';
        }
      else if (base == 2)
        {
          str[0] = '1';
          exp--;
        }
      else
        {
          int digit;
          mp_limb_t msl = (MPFR_MANT (p))[MPFR_LAST_LIMB (p)];
          int rnd_bit = GMP_NUMB_BITS - 5;

          /* pick up the 4 first bits */
          digit = msl >> (rnd_bit + 1);
          if (MPFR_IS_LIKE_RNDA (spec.rnd_mode, MPFR_IS_NEG (p)) ||
              (spec.rnd_mode == MPFR_RNDN &&
               (msl & (MPFR_LIMB_ONE << rnd_bit))))
            digit++;
          MPFR_ASSERTD (0 <= digit && digit <= 15);

          str[0] = num_to_text [digit];
          exp -= 4;
        }
      str[1] = '\0';
    }
  np->ip_ptr = str;

  if (uppercase && spec.size != 0)
    /* All digits in upper case (not needed if spec.size == 0) */
    {
      char *s1 = str;
      while (*s1)
        {
          switch (*s1)
            {
            case 'a':
              *s1 = 'A';
              break;
            case 'b':
              *s1 = 'B';
              break;
            case 'c':
              *s1 = 'C';
              break;
            case 'd':
              *s1 = 'D';
              break;
            case 'e':
              *s1 = 'E';
              break;
            case 'f':
              *s1 = 'F';
              break;
            }
          s1++;
        }
    }

  if (spec.spec == 'b' || spec.prec != 0)
    /* compute the number of digits in the fractional part */
    {
      char *ptr;
      size_t str_len;

      /* the sign has been skipped, skip also the first digit */
      ++str;
      str_len = strlen (str);
      ptr = str + str_len - 1; /* points to the end of str */

      if (spec.prec < 0)
        /* remove trailing zeros, if any */
        {
          while (*ptr == '0' && str_len != 0)
            {
              --ptr;
              --str_len;
            }
        }

      /* WARNING! Due to mpfr_get_str_wrapper, str_len may be smaller
         than the correct value when there is no output. However, the
         code is correct because:
         * If str_len == 0, i.e. the fractional part is empty, then
           if mpfr_get_str_wrapper was called, the number of digits
           has not been reduced; in summary, if str_len == 0, then
           it has the correct value. Otherwise, the first "if" below
           is entered and...
         * If spec.prec <= 0, then mpfr_get_str_wrapper could have been
           called only with spec.prec < 0 (see "if" above), and in this
           case, it was called with n == 0, so that str_len has the
           correct value.
         * If str_len >= spec.prec, then str_len == spec.prec, which is
           the correct value.
         * Otherwise the "if" condition below is true and one will have
           np->fp_size + np->fp_trailing_zeros == spec.prec, which is
           the correct value (both terms of the sum may be incorrect,
           but only the sum will matter). */
      if (str_len != 0)
        /* there are some digits in the fractional part */
        {
          np->fp_ptr = str;
          np->fp_size = str_len;
          /* Warning! str_len has type size_t, which is unsigned. */
          if (spec.prec > 0 && str_len < spec.prec)
            {
              np->fp_trailing_zeros = spec.prec - str_len;
              MPFR_ASSERTD (np->fp_trailing_zeros >= 0);
            }
        }
    }

  /* decimal point */
  if (np->fp_size != 0 || spec.alt)
    np->point = MPFR_DECIMAL_POINT;

  /* the exponent part contains the character 'p', or 'P' plus the sign
     character plus at least one digit and only as many more digits as
     necessary to represent the exponent.
     We assume that |exp| < 10^INT_MAX. */
  np->exp_size = 3;
  {
    mpfr_uexp_t x;

    x = SAFE_ABS (mpfr_uexp_t, exp);
    while (x > 9)
      {
        np->exp_size++;
        x /= 10;
      }
  }
  str = (char *) mpfr_allocate_func (1 + np->exp_size);
  np->exp_ptr = register_string (np->sl, str);
  {
    char exp_fmt[8];  /* contains at most 7 characters like in "p%+.1i",
                         or "P%+.2li" */

    exp_fmt[0] = uppercase ? 'P' : 'p';
    exp_fmt[1] = '\0';
    strcat (exp_fmt, "%+.1" MPFR_EXP_FSPEC "d");

    if (MPFR_UNLIKELY (sprintf (str, exp_fmt, (mpfr_eexp_t) exp) < 0))
      return -1;
  }

  return 0;
}

/* Determine the different parts of the string representation of the regular
   number P when spec.spec is 'e', 'E', 'g', or 'G'.
   For 'e' / 'E', dec_info is a null pointer and trailing zeros will be kept.
   For 'g' / 'G', dec_info contains the previously computed exponent and
   string. Note that if trailing zeros are not kept, we need the full string.

   Return -1 in case of overflow on the sizes. */
static int
regular_eg (struct number_parts *np, mpfr_srcptr p,
            const struct printf_spec spec, struct decimal_info *dec_info,
            int keep_trailing_zeros)
{
  char *str;
  mpfr_exp_t exp;

  const int uppercase = spec.spec == 'E' || spec.spec == 'G';

  /* integral part */
  np->ip_size = 1;
  if (dec_info == NULL)
    {
      size_t nsd;

      /* We will keep the trailing zeros. Thus we can use
         mpfr_get_str_wrapper below. */
      MPFR_ASSERTD (keep_trailing_zeros);

      /* Number of significant digits:
         - if no given precision, then let mpfr_get_str determine it,
         - if a precision is specified, then one digit before decimal point
         plus spec.prec after it.
         We use the fact here that mpfr_get_str allows us to ask for only one
         significant digit when the base is not a power of 2. */
      MPFR_ASSERTD (np->ip_size == 1);  /* thus the + 1 below */
      if (spec.prec < 0)
        nsd = 0;
      else
        {
          if (MPFR_UNLIKELY (spec.prec > (size_t) -2))  /* overflow */
            return -1;
          nsd = (size_t) spec.prec + 1;
        }
      str = mpfr_get_str_wrapper (&exp, 10, nsd, p, spec);
      register_string (np->sl, str);
    }
  else
    {
      exp = dec_info->exp;
      str = dec_info->str;
    }
  np->ip_ptr = MPFR_IS_NEG (p) ? ++str : str;  /* skip sign if any */

  if (spec.prec != 0)
    /* compute the number of digits in the fractional part */
    {
      char *ptr;
      size_t str_len;

      /* the sign has been skipped, skip also the first digit */
      ++str;
      str_len = strlen (str);
      ptr = str + str_len - 1; /* points to the end of str */

      if (!keep_trailing_zeros)
        /* remove trailing zeros, if any */
        {
          while (*ptr == '0' && str_len != 0)
            {
              --ptr;
              --str_len;
            }
        }

      /* WARNING! Due to mpfr_get_str_wrapper, str_len may be smaller
         than the correct value when there is no output. However, the
         code is correct because:
         * If keep_trailing_zeros is false, then mpfr_get_str_wrapper
           is not used. In the following items, we consider the case
           where keep_trailing_zeros is true, so that str_len has not
           been modified above.
         * If str_len == 0, i.e. the fractional part is empty, then
           if mpfr_get_str_wrapper was called, the number of digits
           has not been reduced; in summary, if str_len == 0, then
           it has the correct value. Otherwise, the first "if" below
           is entered and...
         * If spec.prec <= 0, then spec.prec < 0 (see "if" above), and
           in this case, mpfr_get_str_wrapper was called with n == 0,
           so that str_len has the correct value.
         * If str_len >= spec.prec, then str_len == spec.prec, which is
           the correct value.
         * Otherwise the "if" condition below is true and one will have
           np->fp_size + np->fp_trailing_zeros == spec.prec, which is
           the correct value (both terms of the sum may be incorrect,
           but only the sum will matter). */
      if (str_len != 0)
        /* there are some digits in the fractional part */
        {
          np->fp_ptr = str;
          np->fp_size = str_len;
          MPFR_ASSERTD (spec.prec < 0 || str_len <= spec.prec);
          /* Warning! str_len has type size_t, which is unsigned. */
          if (keep_trailing_zeros && spec.prec > 0 && str_len < spec.prec)
            {
              /* add missing trailing zeros or missing digits
                 (in case mpfr_get_str_wrapper has been used
                 and there is no output) */
              np->fp_trailing_zeros = spec.prec - str_len;
              MPFR_ASSERTD (np->fp_trailing_zeros >= 0);
            }
        }
    }

  /* decimal point */
  if (np->fp_size != 0 || spec.alt)
    np->point = MPFR_DECIMAL_POINT;

  /* exp is the exponent for decimal point BEFORE the first digit, we want
     the exponent for decimal point AFTER the first digit.
     Here, no possible overflow because exp < MPFR_EXP (p) / 3 */
  exp--;

  /* the exponent part contains the character 'e', or 'E' plus the sign
     character plus at least two digits and only as many more digits as
     necessary to represent the exponent.
     We assume that |exp| < 10^INT_MAX. */
  np->exp_size = 3;
  {
    mpfr_uexp_t x;

    x = SAFE_ABS (mpfr_uexp_t, exp);
    while (x > 9)
      {
        np->exp_size++;
        x /= 10;
      }
  }
  if (np->exp_size < 4)
    np->exp_size = 4;

  str = (char *) mpfr_allocate_func (1 + np->exp_size);
  np->exp_ptr = register_string (np->sl, str);

  {
    char exp_fmt[8];  /* e.g. "e%+.2i", or "E%+.2li" */

    exp_fmt[0] = uppercase ? 'E' : 'e';
    exp_fmt[1] = '\0';
    strcat (exp_fmt, "%+.2" MPFR_EXP_FSPEC "d");

    if (MPFR_UNLIKELY (sprintf (str, exp_fmt, (mpfr_eexp_t) exp) < 0))
      return -1;
  }

  return 0;
}

/* Determine the different parts of the string representation of the regular
   number P when spec.spec is 'f', 'F', 'g', or 'G'.
   dec_info contains the previously computed exponent and string or is
   a null pointer.

   Return -1 in case of overflow on the sizes. */
static int
regular_fg (struct number_parts *np, mpfr_srcptr p,
            const struct printf_spec spec, struct decimal_info *dec_info,
            int keep_trailing_zeros)
{
  mpfr_exp_t exp;
  char * str;

  /* WARNING: an empty precision field is forbidden (it means precision = 6
     and it should have been changed to 6 before the function call) */
  MPFR_ASSERTD (spec.prec >= 0);

  if (MPFR_GET_EXP (p) <= 0)
    /* 0 < |p| < 1 */
    {
      /* Most of the time, integral part is 0 */
      np->ip_size = 1;
      str = (char *) mpfr_allocate_func (1 + np->ip_size);
      str[0] = '0';
      str[1] = '\0';
      np->ip_ptr = register_string (np->sl, str);

      if (spec.prec == 0)
        /* only two possibilities: either 1 or 0. */
        {
          mpfr_t y;
          /* y = abs(p) */
          MPFR_ALIAS (y, p, 1, MPFR_EXP (p));

          if (spec.rnd_mode == MPFR_RNDA
              || (spec.rnd_mode == MPFR_RNDD && MPFR_IS_NEG (p))
              || (spec.rnd_mode == MPFR_RNDU && MPFR_IS_POS (p))
              || (spec.rnd_mode == MPFR_RNDN && mpfr_cmp_d (y, 0.5) > 0))
            /* rounded up to 1: one digit '1' in integral part.
               note that 0.5 is rounded to 0 with RNDN (round ties to even) */
            np->ip_ptr[0] = '1';
        }
      else
        {
          /* exp =  position of the most significant decimal digit. */
          exp = floor_log10 (p);
          MPFR_ASSERTD (exp < 0);

          if (exp < -spec.prec)
            /* only the last digit may be non zero */
            {
              int round_away;

              /* Due to mpfr_set_si below... */
              if (MPFR_UNLIKELY (spec.prec > LONG_MAX))  /* overflow */
                return -1;

              switch (spec.rnd_mode)
                {
                case MPFR_RNDA:
                case MPFR_RNDF:  /* round_away = 1 needed for %Rg */
                  round_away = 1;
                  break;
                case MPFR_RNDZ:
                  round_away = 0;
                  break;
                case MPFR_RNDD:
                  round_away = MPFR_IS_NEG (p);
                  break;
                case MPFR_RNDU:
                  round_away = MPFR_IS_POS (p);
                  break;
                default:
                  {
                    /* compare |p| to y = 0.5*10^(-spec.prec) */
                    mpfr_t y;
                    mpfr_exp_t e = MAX (MPFR_PREC (p), 56);
                    int cmp;

                    MPFR_ASSERTN (spec.rnd_mode == MPFR_RNDN);
                    mpfr_init2 (y, e + 8);

                    do
                      {
                        /* find a lower approximation of
                           0.5*10^(-spec.prec) different from |p| */
                        e += 8;
                        mpfr_set_prec (y, e);
                        mpfr_set_si (y, -spec.prec, MPFR_RNDN);
                        mpfr_exp10 (y, y, MPFR_RNDD);
                        mpfr_div_2ui (y, y, 1, MPFR_RNDN);
                        cmp = mpfr_cmpabs (y, p);
                      }
                    while (cmp == 0);

                    round_away = cmp < 0;
                    mpfr_clear (y);
                  }
                  break;
                }

              if (round_away)
                /* round away from zero: the last output digit is '1' */
                {
                  np->fp_leading_zeros = spec.prec - 1;

                  np->fp_size = 1;
                  str = (char *) mpfr_allocate_func (1 + np->fp_size);
                  str[0] = '1';
                  str[1] = '\0';
                  np->fp_ptr = register_string (np->sl, str);
                }
              else
                /* only zeros in the fractional part */
                {
                  MPFR_ASSERTD (spec.spec == 'f' || spec.spec == 'F');
                  np->fp_leading_zeros = spec.prec;
                }
            }
          else  /* exp >= -spec.prec */
            /* the most significant digits are the last
               spec.prec + exp + 1 digits in the fractional part */
            {
              char *ptr;
              size_t str_len;

              MPFR_ASSERTD (exp >= -spec.prec);
              if (dec_info == NULL)
                {
                  size_t nsd;

                  /* We will keep the trailing zeros. Thus we can use
                     mpfr_get_str_wrapper below. */
                  MPFR_ASSERTD (keep_trailing_zeros);

                  MPFR_ASSERTD (exp <= -1);
                  MPFR_ASSERTD (spec.prec + (exp + 1) >= 0);
                  if (MPFR_UNLIKELY (spec.prec + (exp + 1) > (size_t) -1))
                    return -1;
                  nsd = spec.prec + (exp + 1);
                  /* WARNING: nsd may equal 1, but here we use the
                     fact that mpfr_get_str can return one digit with
                     base ten (undocumented feature, see comments in
                     get_str.c) */

                  str = mpfr_get_str_wrapper (&exp, 10, nsd, p, spec);
                  register_string (np->sl, str);
                }
              else
                {
                  exp = dec_info->exp;
                  str = dec_info->str;
                }
              if (MPFR_IS_NEG (p))
                /* skip sign */
                ++str;
              if (exp == 1)
                /* round up to 1 */
                {
                  MPFR_ASSERTD (str[0] == '1');
                  np->ip_ptr[0] = '1';
                  if (keep_trailing_zeros)
                    np->fp_leading_zeros = spec.prec;
                }
              else
                {
                  np->fp_ptr = str;
                  np->fp_leading_zeros = -exp;
                  MPFR_ASSERTD (exp <= 0);

                  str_len = strlen (str); /* the sign has been skipped */
                  ptr = str + str_len - 1; /* points to the end of str */

                  if (!keep_trailing_zeros)
                    /* remove trailing zeros, if any */
                    {
                      while (*ptr == '0' && str_len != 0)
                        {
                          --ptr;
                          --str_len;
                        }
                    }

                  /* Before it is potentially reduced, str_len comes from the
                     size argument of mpfr_get_str or mpfr_get_str_wrapper,
                     and this argument is bounded by a mpfr_intmax_t value
                     in both cases. */
                  MPFR_ASSERTD (str_len <= MPFR_INTMAX_MAX);

                  MPFR_ASSERTD (str_len > 0);
                  np->fp_size = str_len;

                  /* WARNING! Due to mpfr_get_str_wrapper, str_len may be
                     smaller than the correct value when there is no output.
                     However, the code is correct because this function is
                     used only when keep_trailing_zeros is true, and in this
                     case, the "np->fp_size = str_len;" above and the code
                     below ensure that
                     np->fp_size + np->fp_trailing_zeros == spec.prec + exp,
                     which is the correct value. */
                  if (keep_trailing_zeros)
                    {
                      /* add missing trailing zeros */
                      /* Since spec.prec >= 0 and exp <= 0
                         and since spec.prec + exp >= 0 and
                         0 <= str_len <= MPFR_INTMAX_MAX, there is
                         no integer overflow in the operations below
                         (done in the mpfr_intmax_t type). */
                      np->fp_trailing_zeros =
                        (spec.prec + exp) - (mpfr_intmax_t) str_len;
                      MPFR_ASSERTD (np->fp_trailing_zeros >= 0);
                    }
                }
            }
        }

      if (spec.alt || np->fp_leading_zeros != 0 || np->fp_size != 0
          || np->fp_trailing_zeros != 0)
        np->point = MPFR_DECIMAL_POINT;
    }
  else
    /* 1 <= |p| */
    {
      size_t str_len;

      /* Determine the position of the most significant decimal digit. */
      exp = floor_log10 (p);
      MPFR_ASSERTD (exp >= 0);

      if (dec_info == NULL)
        {
          /* %f case */
          mpfr_uintmax_t n;

          n = (mpfr_uintmax_t) spec.prec + (exp + 1);
          if (MPFR_UNLIKELY (n > (size_t) -1))
            return -1;
          str = mpfr_get_str_wrapper (&exp, 10, n, p, spec);
          register_string (np->sl, str);
        }
      else
        {
          /* %g case */
          exp = dec_info->exp;
          str = dec_info->str;
        }
      np->ip_ptr = MPFR_IS_NEG (p) ? ++str : str; /* skip sign */
      str_len = strlen (str);

      /* integral part */
      if (exp > str_len)
        {
          /* When spec.size == 0, mpfr_get_str may be called in a reduced
             precision, so that some trailing digits may have been ignored.
             When spec.size != 0, this case is also possible in the case
             where p is rounded up to the next power of 10: a zero must be
             added since the exponent has been increased by 1. */
          np->ip_trailing_digits = exp - str_len;
          np->ip_size = str_len;
        }
      else
        np->ip_size = exp;

      if (spec.group)
        /* thousands separator in integral part */
        np->thousands_sep = MPFR_THOUSANDS_SEPARATOR;

      /* fractional part */
      str += np->ip_size;
      str_len -= np->ip_size;
      if (!keep_trailing_zeros)
        /* remove trailing zeros, if any */
        {
          char *ptr = str + str_len - 1; /* pointer to the last digit of
                                            str */
          while (*ptr == '0' && str_len != 0)
            {
              --ptr;
              --str_len;
            }
        }

      if (str_len > 0)
        /* some digits in the fractional part */
        {
          np->point = MPFR_DECIMAL_POINT;
          np->fp_ptr = str;
          np->fp_size = str_len;
        }

      /* Warning! str_len has type size_t, which is unsigned. */
      MPFR_ASSERTD (spec.prec >= 0);  /* let's recall this */
      if (keep_trailing_zeros && str_len < spec.prec)
        /* add missing trailing zeros */
        {
          np->point = MPFR_DECIMAL_POINT;
          np->fp_trailing_zeros = spec.prec - np->fp_size;
          MPFR_ASSERTD (np->fp_trailing_zeros >= 0);
        }

      if (spec.alt)
        /* add decimal point even if no digits follow it */
        np->point = MPFR_DECIMAL_POINT;
    }

  return 0;
}

/* partition_number determines the different parts of the string
   representation of the number p according to the given specification.
   partition_number initializes the given structure np, so all previous
   information in that variable is lost.
   Return the total number of characters to be written.
   Return -1 if an error occurred, in that case np's fields are in an
   undefined state but all string buffers have been freed. */
static mpfr_intmax_t
partition_number (struct number_parts *np, mpfr_srcptr p,
                  struct printf_spec spec)
{
  char *str;
  mpfr_uintmax_t total;  /* can hold the sum of two non-negative
                            signed integers + 1 */
  int uppercase;

  /* WARNING: left justification means right space padding */
  np->pad_type = spec.left ? RIGHT : spec.pad == '0' ? LEADING_ZEROS : LEFT;
  np->pad_size = 0;
  np->prefix_ptr = NULL;
  np->prefix_size = 0;
  np->thousands_sep = '\0';
  np->ip_ptr = NULL;
  np->ip_size = 0;
  np->ip_trailing_digits = 0;
  np->point = '\0';
  np->fp_leading_zeros = 0;
  np->fp_ptr = NULL;
  np->fp_size = 0;
  np->fp_trailing_zeros = 0;
  np->exp_ptr = NULL;
  np->exp_size = 0;
  np->sl = (struct string_list *)
    mpfr_allocate_func (sizeof (struct string_list));
  init_string_list (np->sl);

  uppercase = spec.spec == 'A' || spec.spec == 'E' || spec.spec == 'F'
    || spec.spec == 'G';

  /* The sign/space rule is the same for all cases. */
  np->sign =
    MPFR_IS_NEG (p) ? '-' :
    spec.showsign ? '+' :
    spec.space ? ' ' : '\0';

  if (MPFR_UNLIKELY (MPFR_IS_SINGULAR (p)))
    {
      if (MPFR_IS_NAN (p))
        {
          if (np->pad_type == LEADING_ZEROS)
            /* don't want "0000nan", change to right justification padding
               with left spaces instead */
            np->pad_type = LEFT;

          np->ip_size = MPFR_NAN_STRING_LENGTH;
          str = (char *) mpfr_allocate_func (1 + np->ip_size);
          strcpy (str, uppercase ? MPFR_NAN_STRING_UC : MPFR_NAN_STRING_LC);
          np->ip_ptr = register_string (np->sl, str);
        }
      else if (MPFR_IS_INF (p))
        {
          if (np->pad_type == LEADING_ZEROS)
            /* don't want "0000inf", change to right justification padding
               with left spaces instead */
            np->pad_type = LEFT;

          np->ip_size = MPFR_INF_STRING_LENGTH;
          str = (char *) mpfr_allocate_func (1 + np->ip_size);
          strcpy (str, uppercase ? MPFR_INF_STRING_UC : MPFR_INF_STRING_LC);
          np->ip_ptr = register_string (np->sl, str);
        }
      else
        {
          MPFR_ASSERTD (MPFR_IS_ZERO (p));
          /* note: for 'g' spec, zero is always displayed with 'f'-style with
             precision spec.prec - 1 and the trailing zeros are removed unless
             the flag '#' is used. */

          if (spec.spec == 'a' || spec.spec == 'A')
            /* prefix part */
            {
              np->prefix_size = 2;
              str = (char *) mpfr_allocate_func (1 + np->prefix_size);
              str[0] = '0';
              str[1] = uppercase ? 'X' : 'x';
              str[2] = '\0';
              np->prefix_ptr = register_string (np->sl, str);
            }

          /* integral part */
          np->ip_size = 1;
          str = (char *) mpfr_allocate_func (1 + np->ip_size);
          str[0] = '0';
          str[1] = '\0';
          np->ip_ptr = register_string (np->sl, str);

          if (spec.prec < 0)  /* empty precision field */
            {
              if (spec.spec == 'e' || spec.spec == 'E')
                spec.prec = mpfr_get_str_ndigits (10, MPFR_GET_PREC (p)) - 1;
              else if (spec.spec == 'f' || spec.spec == 'F' ||
                       spec.spec == 'g' || spec.spec == 'G')
                spec.prec = DEFAULT_DECIMAL_PREC;
            }

          if (spec.prec > 0
              && ((spec.spec != 'g' && spec.spec != 'G') || spec.alt))
            /* fractional part */
            {
              np->point = MPFR_DECIMAL_POINT;
              np->fp_trailing_zeros = (spec.spec == 'g' || spec.spec == 'G') ?
                spec.prec - 1 : spec.prec;
              MPFR_ASSERTD (np->fp_trailing_zeros >= 0);
            }
          else if (spec.alt)
            np->point = MPFR_DECIMAL_POINT;

          if (spec.spec == 'a' || spec.spec == 'A' || spec.spec == 'b'
              || spec.spec == 'e' || spec.spec == 'E')
            /* exponent part */
            {
              np->exp_size = (spec.spec == 'e' || spec.spec == 'E') ? 4 : 3;
              str = (char *) mpfr_allocate_func (1 + np->exp_size);
              if (spec.spec == 'e' || spec.spec == 'E')
                strcpy (str, uppercase ? "E+00" : "e+00");
              else
                strcpy (str, uppercase ? "P+0" : "p+0");
              np->exp_ptr = register_string (np->sl, str);
            }
        }
    }
  else if (MPFR_UNLIKELY (MPFR_IS_UBF (p)))
    {
      /* mpfr_get_str does not support UBF, so that UBF numbers are regarded
         as special cases here. This is not much a problem since UBF numbers
         are internal to MPFR and here, they only for logging. */
      if (np->pad_type == LEADING_ZEROS)
        /* change to right justification padding with left spaces */
        np->pad_type = LEFT;

      np->ip_size = 3;
      str = (char *) mpfr_allocate_func (1 + np->ip_size);
      strcpy (str, uppercase ? "UBF" : "ubf");
      np->ip_ptr = register_string (np->sl, str);
      /* TODO: output more information (e.g. the exponent) if need be. */
    }
  else
    {
      MPFR_ASSERTD (MPFR_IS_PURE_FP (p));
      if (spec.spec == 'a' || spec.spec == 'A' || spec.spec == 'b')
        {
          if (regular_ab (np, p, spec) == -1)
            goto error;
        }
      else if (spec.spec == 'f' || spec.spec == 'F')
        {
          if (spec.prec < 0)
            spec.prec = DEFAULT_DECIMAL_PREC;
          if (regular_fg (np, p, spec, NULL, 1) == -1)
            goto error;
        }
      else if (spec.spec == 'e' || spec.spec == 'E')
        {
          if (regular_eg (np, p, spec, NULL, 1) == -1)
            goto error;
        }
      else
        /* %g case */
        {
          /* Use the C99 rules:
             if T > X >= -4 then the conversion is with style 'f'/'F' and
             precision T-(X+1).
             otherwise, the conversion is with style 'e'/'E' and
             precision T-1.
             where T is the threshold computed below and X is the exponent
             that would be displayed with style 'e' and precision T-1. */
          mpfr_intmax_t threshold;
          mpfr_exp_t x, e, k;
          struct decimal_info dec_info;

          threshold = spec.prec < 0 ? DEFAULT_DECIMAL_PREC :
            spec.prec == 0 ? 1 : spec.prec;
          MPFR_ASSERTD (threshold >= 1);

          /* Here we cannot call mpfr_get_str_wrapper since we need the full
             significand in dec_info.str.
             Moreover, threshold may be huge while one can know that the
             number of digits that are not trailing zeros remains limited;
             such a limit occurs in practical cases, e.g. with numbers
             representable in the IEEE 754-2008 basic formats. Since the
             trailing zeros are not necessarily output, we do not want to
             waste time and memory by making mpfr_get_str generate them.
             So, let us try to find a smaller threshold for mpfr_get_str.
             |p| < 2^EXP(p) = 10^(EXP(p)*log10(2)). So, the integer part
             takes at most ceil(EXP(p)*log10(2)) digits (unless p rounds
             to the next power of 10, but in this case any threshold will
             be OK). So, for the integer part, we will take:
             max(0,floor((EXP(p)+2)/3)).
             Let k = PREC(p) - EXP(p), so that the last bit of p has
             weight 2^(-k). If k <= 0, then p is an integer, otherwise
             the fractional part in base 10 may have up to k digits
             (this bound is reached if the last bit is 1).
             Note: The bound could be improved, but this is not critical.
             Note also that even when there is no output, this may take much
             memory when both the exponent magnitude and the format precision
             are huge. The exponent of the most significant decimal digit
             can be obtained with mpfr_get_str_wrapper, and the exponent of
             the least significant decimal digit can easily be computed in
             at most linear space, but this allows one to deduce the number
             of decimal digits of the rounded value only if the requested
             format precision is at least as large as the number of decimal
             digits of the exact value. The code could be improved in this
             case, though. */
          e = MPFR_GET_EXP (p);
          k = MPFR_PREC (p) - e;
          e = e <= 0 ? k : (e + 2) / 3 + (k <= 0 ? 0 : k);
          MPFR_ASSERTD (e >= 1);

          if (e > threshold)
            e = threshold;

          /* error if e does not fit in size_t (for mpfr_get_str) */
          if (e > (size_t) -1)
            goto error;

          dec_info.str = mpfr_get_str (NULL, &dec_info.exp, 10,
                                       e, p, spec.rnd_mode);
          register_string (np->sl, dec_info.str);
          /* mpfr_get_str corresponds to a significand between 0.1 and 1,
             whereas here we want a significand between 1 and 10. */
          x = dec_info.exp - 1;

          if (threshold > x && x >= -4)
            {
              /* the conversion is with style 'f' */
              spec.prec = threshold - x - 1;

              if (regular_fg (np, p, spec, &dec_info, spec.alt) == -1)
                goto error;
            }
          else
            {
              spec.prec = threshold - 1;

              if (regular_eg (np, p, spec, &dec_info, spec.alt) == -1)
                goto error;
            }
        }
    }

  /* compute the number of characters to be written verifying it is not too
     much */

#define INCR_TOTAL(v)                                   \
  do {                                                  \
    MPFR_ASSERTD ((v) >= 0);                            \
    if (MPFR_UNLIKELY ((v) > MPFR_INTMAX_MAX))          \
      goto error;                                       \
    total += (v);                                       \
    if (MPFR_UNLIKELY (total > MPFR_INTMAX_MAX))        \
      goto error;                                       \
  } while (0)

  total = np->sign ? 1 : 0;
  INCR_TOTAL (np->prefix_size);
  INCR_TOTAL (np->ip_size);
  INCR_TOTAL (np->ip_trailing_digits);
  MPFR_ASSERTD (np->ip_size + np->ip_trailing_digits >= 1);
  if (np->thousands_sep)
    /* ' flag, style f and the thousands separator in current locale is not
       reduced to the null character */
    INCR_TOTAL ((np->ip_size + np->ip_trailing_digits - 1) / 3);
  if (np->point)
    ++total;
  INCR_TOTAL (np->fp_leading_zeros);
  INCR_TOTAL (np->fp_size);
  INCR_TOTAL (np->fp_trailing_zeros);
  INCR_TOTAL (np->exp_size);

  if (spec.width > total)
    /* pad with spaces or zeros depending on np->pad_type */
    {
      np->pad_size = spec.width - total;
      total = spec.width;
    }

  MPFR_ASSERTD (total > 0 && total <= MPFR_INTMAX_MAX);
  return total;

 error:
  clear_string_list (np->sl);
  np->prefix_ptr = NULL;
  np->ip_ptr = NULL;
  np->fp_ptr = NULL;
  np->exp_ptr = NULL;
  return -1;
}

/* sprnt_fp prints a mpfr_t according to spec.spec specification.

   Return the size of the string (not counting the terminating '\0').
   Return -1 if the built string is too long (i.e. has more than
   INT_MAX or MPFR_INTMAX_MAX characters).

   If spec.size is 0, we only want the number of characters (no output).
*/
static int
sprnt_fp (struct string_buffer *buf, mpfr_srcptr p,
          const struct printf_spec spec)
{
  mpfr_intmax_t length, start;
  struct number_parts np;

  length = partition_number (&np, p, spec);
  if (MPFR_UNLIKELY (length < 0))
    {
      buf->len = -1;
      return -1;
    }

  if (spec.size == 0)  /* no output */
    {
      /* This is equivalent to the code below this "if" (since there is
         no need to fill the buffer and length is already  known). */
      buffer_incr_len (buf, length);
      goto clear_and_exit;
    }

  MPFR_DBGRES (start = buf->len);

  /* right justification padding with left spaces */
  if (np.pad_type == LEFT && np.pad_size != 0)
    buffer_pad (buf, ' ', np.pad_size);

  /* sign character (may be '-', '+', ' ', or '\0') */
  if (np.sign)
    buffer_pad (buf, np.sign, 1);

  /* prefix part */
  if (np.prefix_ptr)
    buffer_cat (buf, np.prefix_ptr, np.prefix_size);

  /* right justification padding with leading zeros */
  if (np.pad_type == LEADING_ZEROS && np.pad_size != 0)
    buffer_pad (buf, '0', np.pad_size);

  /* integral part (may also be "nan" or "inf") */
  MPFR_ASSERTN (np.ip_ptr != NULL); /* never empty */
  if (MPFR_UNLIKELY (np.thousands_sep))
    {
      if (buffer_sandwich (buf, np.ip_ptr, np.ip_size, np.ip_trailing_digits,
                           np.thousands_sep))
        {
          buf->len = -1;
          goto clear_and_exit;
        }
    }
  else
    {
      buffer_cat (buf, np.ip_ptr, np.ip_size);

      /* possible trailing zero in integral part (spec.size != 0) */
      MPFR_ASSERTD (np.ip_trailing_digits <= 1);
      if (np.ip_trailing_digits != 0)
        buffer_pad (buf, '0', 1);
    }

  /* decimal point */
  if (np.point)
    buffer_pad (buf, np.point, 1);

  /* leading zeros in the fractional part */
  if (np.fp_leading_zeros != 0)
    buffer_pad (buf, '0', np.fp_leading_zeros);

  /* significant digits in the fractional part */
  if (np.fp_ptr)
    buffer_cat (buf, np.fp_ptr, np.fp_size);

  /* trailing zeros in the fractional part */
  if (np.fp_trailing_zeros != 0)
    buffer_pad (buf, '0', np.fp_trailing_zeros);

  /* exponent part */
  if (np.exp_ptr)
    buffer_cat (buf, np.exp_ptr, np.exp_size);

  /* left justification padding with right spaces */
  if (np.pad_type == RIGHT && np.pad_size != 0)
    buffer_pad (buf, ' ', np.pad_size);

  MPFR_ASSERTD (buf->len == -1 || buf->len - start == length);

 clear_and_exit:
  clear_string_list (np.sl);
  return buf->len == -1 ? -1 : length;
}

/* The following internal function implements both mpfr_vasprintf and
   mpfr_vsnprintf:
   (a) either ptr <> NULL, and then Buf and size are not used, and it
       implements mpfr_vasprintf (ptr, fmt, ap)
   (b) or ptr = NULL, and it implements mpfr_vsnprintf (Buf, size, fmt, ap)
   It returns the number of characters that would have been written had 'size'
   been sufficiently large, not counting the terminating null character, or -1
   if this number is too large for the return type 'int' (overflow), in which
   case, if ptr <> NULL, the memory is deallocated (otherwise, there is no way
   to deallocate it later, since the actual size, needed by mpfr_free_func, is
   unknown as there may be non-terminating null characters).
*/
int
mpfr_vasnprintf_aux (char **ptr, char *Buf, size_t size, const char *fmt,
                     va_list ap)
{
  struct string_buffer buf;
  int nbchar;

  /* information on the conversion specification filled by the parser */
  struct printf_spec spec;
  /* flag raised when previous part of fmt need to be processed by
     gmp_vsnprintf */
  int xgmp_fmt_flag;
  /* beginning and end of the previous unprocessed part of fmt */
  const char *start, *end;
  /* pointer to arguments for gmp_vasprintf */
  va_list ap2;
  MPFR_SAVE_EXPO_DECL (expo);

  MPFR_LOG_FUNC
    (("size=%zu", size),
     ("", 0));

  MPFR_SAVE_EXPO_MARK (expo);

  /* FIXME: For ptr = NULL, once buf.len >= size, switch to size = 0 for
     efficiency and avoid potential DoS? i.e. we no longer need to generate
     the strings (potentially huge), just compute the lengths. */

  spec.size = ptr != NULL || size != 0;  /* true iff do output */
  buffer_init (&buf, spec.size ? 4096 : 0);
  xgmp_fmt_flag = 0;
  va_copy (ap2, ap);
  start = fmt;
  while (*fmt != '\0')
    {
      int overflow = 0;

      /* Look for the next format specification */
      while (*fmt != '\0' && *fmt != '%')
        ++fmt;

      if (*fmt == '\0')
        break;

      if (*++fmt == '%')
        /* %%: go one step further otherwise the second '%' would be
           considered as a new conversion specification introducing
           character */
        {
          ++fmt;
          xgmp_fmt_flag = 1;
          continue;
        }

      end = fmt - 1;

      /* Initialize spec (of type struct printf_spec). */
      spec.alt = 0;
      spec.space = 0;
      spec.left = 0;
      spec.showsign = 0;
      spec.group = 0;
      spec.width = 0;
      spec.prec = 0;
      spec.arg_type = NONE;
      spec.rnd_mode = MPFR_RNDN;
      spec.spec = '\0';
      spec.pad = ' ';

      /* format string analysis */
      fmt = parse_flags (fmt, &spec);

      READ_INT (ap, fmt, spec.width);
      if (spec.width < 0)  /* integer read via '*', no overflow */
        {
          spec.left = 1;
          /* Since the type of the integer is int, spec.width >= INT_MIN,
             so that an overflow is possible here only if mpfr_intmax_t
             has the same size of int. The INT_MIN < - MPFR_INTMAX_MAX
             test allows the compiler to optimize when it is false. */
          if (MPFR_UNLIKELY (INT_MIN < - MPFR_INTMAX_MAX &&
                             spec.width < - MPFR_INTMAX_MAX))
            overflow = 1;
          else
            spec.width = - spec.width;
        }
      /* Note: We will make sure that spec.width is not used in case of
         overflow. */
      MPFR_ASSERTD (overflow || spec.width >= 0);

      if (*fmt == '.')
        {
          ++fmt;
          READ_INT (ap, fmt, spec.prec);
          /* A negative value is possible with ".*" and it will be regarded
             as a missing precision (ISO C). We need to make sure that such
             a value is representable in an int (see its use below). */
          if (spec.prec < 0)
            spec.prec = -1;
        }
      else
        spec.prec = -1;  /* missing precision */
      MPFR_ASSERTD (spec.prec >= -1);

      fmt = parse_arg_type (fmt, &spec);
      if (spec.arg_type == UNSUPPORTED)
        /* the current architecture doesn't support the type corresponding to
           the format specifier; according to the ISO C99 standard, the
           behavior is undefined. We choose to print the format specifier as a
           literal string, what may be printed after this string is
           undefined. */
        continue;
      else if (spec.arg_type == MPFR_ARG)
        {
          switch (*fmt)
            {
            case '\0':
              break;
            case '*':
              ++fmt;
              spec.rnd_mode = (mpfr_rnd_t) va_arg (ap, int);
              break;
            case 'D':
              ++fmt;
              spec.rnd_mode = MPFR_RNDD;
              break;
            case 'U':
              ++fmt;
              spec.rnd_mode = MPFR_RNDU;
              break;
            case 'Y':
              ++fmt;
              spec.rnd_mode = MPFR_RNDA;
              break;
            case 'Z':
              ++fmt;
              spec.rnd_mode = MPFR_RNDZ;
              break;
            case 'N':
              ++fmt;
              MPFR_FALLTHROUGH;
            default:
              spec.rnd_mode = MPFR_RNDN;
            }
        }

      spec.spec = *fmt;
      if (!specinfo_is_valid (spec))
        /* the format specifier is invalid; according to the ISO C99 standard,
           the behavior is undefined. We choose to print the invalid format
           specifier as a literal string, what may be printed after this
           string is undefined. */
        continue;

      if (*fmt != '\0')
        fmt++;

      /* Format processing */
      if (spec.spec == '\0')
        /* end of the format string */
        break;
      else if (spec.spec == 'n')
        /* put the number of characters written so far in the location pointed
           by the next va_list argument; the types of pointer accepted are the
           same as in GMP (except unsupported quad_t) plus pointer to a mpfr_t
           so as to be able to accept the same format strings. */
        {
          void *p;

          p = va_arg (ap, void *);
          FLUSH (xgmp_fmt_flag, start, end, ap2, &buf);
          va_end (ap2);
          start = fmt;

          switch (spec.arg_type)
            {
            case CHAR_ARG:
              *(char *) p = (char) buf.len;
              break;
            case SHORT_ARG:
              *(short *) p = (short) buf.len;
              break;
            case LONG_ARG:
              *(long *) p = (long) buf.len;
              break;
#ifdef HAVE_LONG_LONG
            case LONG_LONG_ARG:
              *(long long *) p = (long long) buf.len;
              break;
#endif
#ifdef _MPFR_H_HAVE_INTMAX_T
            case INTMAX_ARG:
              *(intmax_t *) p = (intmax_t) buf.len;
              break;
#endif
            case SIZE_ARG:
              *(size_t *) p = buf.len;
              break;
            case PTRDIFF_ARG:
              *(ptrdiff_t *) p = (ptrdiff_t) buf.len;
              break;
            case MPF_ARG:
              mpf_set_ui ((mpf_ptr) p, (unsigned long) buf.len);
              break;
            case MPQ_ARG:
              mpq_set_ui ((mpq_ptr) p, (unsigned long) buf.len, 1L);
              break;
            case MP_LIMB_ARG:
              *(mp_limb_t *) p = (mp_limb_t) buf.len;
              break;
            case MP_LIMB_ARRAY_ARG:
              {
                mp_limb_t *q = (mp_limb_t *) p;
                mp_size_t n;
                n = va_arg (ap, mp_size_t);
                if (n < 0)
                  n = -n;
                else if (n == 0)
                  break;

                /* we assume here that mp_limb_t is wider than int */
                *q = (mp_limb_t) buf.len;
                while (--n != 0)
                  {
                    q++;
                    *q = MPFR_LIMB_ZERO;
                  }
              }
              break;
            case MPZ_ARG:
              mpz_set_ui ((mpz_ptr) p, (unsigned long) buf.len);
              break;

            case MPFR_ARG:
              mpfr_set_ui ((mpfr_ptr) p, (unsigned long) buf.len,
                           spec.rnd_mode);
              break;

            default:
              *(int *) p = (int) buf.len;
            }
          va_copy (ap2, ap); /* after the switch, due to MP_LIMB_ARRAY_ARG
                                case */
        }
      else if (spec.arg_type == MPFR_PREC_ARG)
        /* output mpfr_prec_t variable */
        {
          char *s;
          char format[MPFR_PREC_FORMAT_SIZE + 12]; /* e.g. "%0#+ -'*.*ld\0" */
          size_t length;
          mpfr_prec_t prec;

          /* FIXME: With buf.size = 0 and a huge width or precision, this
             can uselessly take much memory. And even with buf.size != 0,
             this would take more memory than necessary and need a large
             buffer_cat. A solution: compute a bound on the maximum
             number of significant digits, and handle the additional
             characters separately. Moreover, if buf.size = 0 or size != 0,
             gmp_snprintf should be called instead of gmp_asprintf,
             outputting data directly to the buffer when applicable.
             See also: https://sourceware.org/bugzilla/show_bug.cgi?id=23432
             Add testcases. */

          prec = va_arg (ap, mpfr_prec_t);

          FLUSH (xgmp_fmt_flag, start, end, ap2, &buf);
          va_end (ap2);
          va_copy (ap2, ap);
          start = fmt;

          /* The restriction to INT_MAX is a limitation due to the fact
             that *.* is used below. If the width or precision field is
             larger than INT_MAX, then there is a real overflow on the
             return value due to the padding characters, thus the error
             is correct. The only minor drawback is that some variables
             corresponding to the 'n' conversion specifier with a type
             larger than int may not be set. This is not a bug, as there
             are no strong guarantees for such variables in case of error.
             FIXME: If size = 0 and max(spec.width,spec.prec) is large
             enough, there is no need to call gmp_asprintf since we are
             just interested in the length, which should be this maximum;
             in particular, this should avoid the overflow issue. */
          if (overflow || spec.width > INT_MAX || spec.prec > INT_MAX)
            {
              buf.len = -1;
              goto error;
            }

          /* Recalled from above. */
          MPFR_ASSERTD (spec.width >= 0);
          MPFR_ASSERTD (spec.prec >= -1);

          /* construct format string, like "%*.*hd" "%*.*d" or "%*.*ld" */
          sprintf (format, "%%%s%s%s%s%s%s*.*" MPFR_PREC_FORMAT_TYPE "%c",
                   spec.pad == '0' ? "0" : "",
                   spec.alt ? "#" : "",
                   spec.showsign ? "+" : "",
                   spec.space ? " " : "",
                   spec.left ? "-" : "",
                   spec.group ? "'" : "",
                   spec.spec);
          MPFR_LOG_MSG (("MPFR_PREC_ARG: format for gmp_asprintf: \"%s\"\n",
                         format));
          MPFR_LOG_MSG (("MPFR_PREC_ARG: width = %d, prec = %d, value = %"
                         MPFR_PREC_FORMAT_TYPE "d\n",
                         (int) spec.width, (int) spec.prec, prec));
          length = gmp_asprintf (&s, format,
                                 (int) spec.width, (int) spec.prec, prec);
          MPFR_ASSERTN (length >= 0);  /* guaranteed by GMP 6 */
          buffer_cat (&buf, s, length);
          mpfr_free_str (s);
        }
      else if (spec.arg_type == MPFR_ARG)
        /* output a mpfr_t variable */
        {
          mpfr_srcptr p;

          if (spec.spec != 'a' && spec.spec != 'A'
              && spec.spec != 'b'
              && spec.spec != 'e' && spec.spec != 'E'
              && spec.spec != 'f' && spec.spec != 'F'
              && spec.spec != 'g' && spec.spec != 'G')
            /* The format specifier is invalid; skip the invalid format
               specifier so as to print it as a literal string. What may
               be printed after this string is undefined. */
            continue;

          p = va_arg (ap, mpfr_srcptr);

          FLUSH (xgmp_fmt_flag, start, end, ap2, &buf);
          va_end (ap2);
          va_copy (ap2, ap);
          start = fmt;

          if (overflow)
            {
              buf.len = -1;
              goto error;
            }

          sprnt_fp (&buf, p, spec);
        }
      else
        /* gmp_printf specification, step forward in the va_list */
        {
          CONSUME_VA_ARG (spec, ap);
          xgmp_fmt_flag = 1;
        }
    }

  if (start != fmt)
    FLUSH (xgmp_fmt_flag, start, fmt, ap2, &buf);

  va_end (ap2);

  if (buf.len == -1 || buf.len > INT_MAX)  /* overflow */
    goto overflow;

  nbchar = buf.len;
  MPFR_ASSERTD (nbchar >= 0);

  /* Warning! Be careful that the buffer may contain null characters
     in addition to the terminating one, in case %c has been used with
     the value 0. */

  if (ptr != NULL)  /* implement mpfr_vasprintf */
    {
      *ptr = (char *) mpfr_reallocate_func (buf.start, buf.size, nbchar + 1);
    }
  else if (size != 0)  /* implement mpfr_vsnprintf */
    {
      /* The size is limited to int (see above). */
      int len = nbchar < size ? nbchar : size - 1;

      memcpy (Buf, buf.start, len);
      Buf[len] = '\0';
      mpfr_free_func (buf.start, buf.size);
    }

  MPFR_SAVE_EXPO_FREE (expo);
  MPFR_LOG_MSG (("nbchar=%d\n", nbchar));
  return nbchar; /* return the number of characters that would have
                    been written had 'size' been sufficiently large,
                    not counting the terminating null character */

 error:
  va_end (ap2);
  if (buf.len == -1)  /* overflow */
    {
    overflow:
      MPFR_LOG_MSG (("Overflow\n", 0));
      MPFR_SAVE_EXPO_UPDATE_FLAGS (expo, MPFR_FLAGS_ERANGE);
#ifdef EOVERFLOW
      MPFR_LOG_MSG (("Setting errno to EOVERFLOW\n", 0));
      errno = EOVERFLOW;
#endif
    }

  MPFR_SAVE_EXPO_FREE (expo);
  if (ptr != NULL)  /* implement mpfr_vasprintf */
    *ptr = NULL;
  if (ptr != NULL || size != 0)
    mpfr_free_func (buf.start, buf.size);

  return -1;
}

#else /* HAVE_STDARG */

/* Avoid an empty translation unit (see ISO C99, 6.9) */
typedef int foo;

#endif /* HAVE_STDARG */