src/sort.c - coreutils-8.23

Global variables defined

Data types defined

Functions defined

Macros defined

Source code

/* sort - sort lines of text (with all kinds of options).
   Copyright (C) 1988-2014 Free Software Foundation, Inc.

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

   This program 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 General Public License for more details.

   You should have received a copy of the GNU General Public License
   along with this program.  If not, see <http://www.gnu.org/licenses/>.

   Written December 1988 by Mike Haertel.
   The author may be reached (Email) at the address mike@gnu.ai.mit.edu,
   or (US mail) as Mike Haertel c/o Free Software Foundation.

   Ørn E. Hansen added NLS support in 1997.  */

#include <config.h>

#include <getopt.h>
#include <pthread.h>
#include <sys/resource.h>
#include <sys/types.h>
#include <sys/wait.h>
#include <signal.h>
#include <assert.h>
#include "system.h"
#include "argmatch.h"
#include "error.h"
#include "fadvise.h"
#include "filevercmp.h"
#include "hard-locale.h"
#include "hash.h"
#include "heap.h"
#include "ignore-value.h"
#include "md5.h"
#include "mbswidth.h"
#include "nproc.h"
#include "physmem.h"
#include "posixver.h"
#include "quote.h"
#include "quotearg.h"
#include "randread.h"
#include "readtokens0.h"
#include "stdio--.h"
#include "stdlib--.h"
#include "strnumcmp.h"
#include "xmemcoll.h"
#include "xnanosleep.h"
#include "xstrtol.h"

#ifndef RLIMIT_DATA
‌struct rlimit { size_t rlim_cur; };
‌# define getrlimit(Resource, Rlp) (-1)
#endif

/* The official name of this program (e.g., no 'g' prefix).  */
‌#define PROGRAM_NAME "sort"

‌#define AUTHORS \
  proper_name ("Mike Haertel"), \
  proper_name ("Paul Eggert")

#if HAVE_LANGINFO_CODESET
# include <langinfo.h>
#endif

/* Use SA_NOCLDSTOP as a proxy for whether the sigaction machinery is
   present.  */
#ifndef SA_NOCLDSTOP
‌# define SA_NOCLDSTOP 0
/* No sigprocmask.  Always 'return' zero. */
‌# define sigprocmask(How, Set, Oset) (0)
‌# define sigset_t int
# if ! HAVE_SIGINTERRUPT
‌#  define siginterrupt(sig, flag) /* empty */
# endif
#endif

#if !defined OPEN_MAX && defined NR_OPEN
‌# define OPEN_MAX NR_OPEN
#endif
#if !defined OPEN_MAX
‌# define OPEN_MAX 20
#endif

‌#define UCHAR_LIM (UCHAR_MAX + 1)

#if HAVE_C99_STRTOLD
‌# define long_double long double
#else
‌# define long_double double
‌# undef strtold
‌# define strtold strtod
#endif

#ifndef DEFAULT_TMPDIR
‌# define DEFAULT_TMPDIR "/tmp"
#endif

/* Maximum number of lines to merge every time a NODE is taken from
   the merge queue.  Node is at LEVEL in the binary merge tree,
   and is responsible for merging TOTAL lines. */
‌#define MAX_MERGE(total, level) (((total) >> (2 * ((level) + 1))) + 1)

/* Heuristic value for the number of lines for which it is worth creating
   a subthread, during an internal merge sort.  I.e., it is a small number
   of "average" lines for which sorting via two threads is faster than
   sorting via one on an "average" system.  On a dual-core 2.0 GHz i686
   system with 3GB of RAM and 2MB of L2 cache, a file containing 128K
   lines of gensort -a output is sorted slightly faster with --parallel=2
   than with --parallel=1.  By contrast, using --parallel=1 is about 10%
   faster than using --parallel=2 with a 64K-line input.  */
enum { SUBTHREAD_LINES_HEURISTIC = 128 * 1024 };
verify (4 <= SUBTHREAD_LINES_HEURISTIC);

/* The number of threads after which there are
   diminishing performance gains.  */
enum { DEFAULT_MAX_THREADS = 8 };

/* Exit statuses.  */
enum
  {
    /* POSIX says to exit with status 1 if invoked with -c and the
       input is not properly sorted.  */
    SORT_OUT_OF_ORDER = 1,

    /* POSIX says any other irregular exit must exit with a status
       code greater than 1.  */
    SORT_FAILURE = 2
  };

enum
  {
    /* The number of times we should try to fork a compression process
       (we retry if the fork call fails).  We don't _need_ to compress
       temp files, this is just to reduce disk access, so this number
       can be small.  Each retry doubles in duration.  */
    MAX_FORK_TRIES_COMPRESS = 4,

    /* The number of times we should try to fork a decompression process.
       If we can't fork a decompression process, we can't sort, so this
       number should be big.  Each retry doubles in duration.  */
    MAX_FORK_TRIES_DECOMPRESS = 9
  };

enum
  {
    /* Level of the end-of-merge node, one level above the root. */
    MERGE_END = 0,

    /* Level of the root node in merge tree. */
    MERGE_ROOT = 1
  };

/* The representation of the decimal point in the current locale.  */
‌static int decimal_point;

/* Thousands separator; if -1, then there isn't one.  */
‌static int thousands_sep;

/* Nonzero if the corresponding locales are hard.  */
‌static bool hard_LC_COLLATE;
#if HAVE_NL_LANGINFO
‌static bool hard_LC_TIME;
#endif

‌#define NONZERO(x) ((x) != 0)

/* The kind of blanks for '-b' to skip in various options. */
‌enum blanktype { bl_start, bl_end, bl_both };

/* The character marking end of line. Default to \n. */
‌static char eolchar = '\n';

/* Lines are held in core as counted strings. */
‌struct line
{
  char *text;            /* Text of the line. */
  size_t length;        /* Length including final newline. */
  char *keybeg;            /* Start of first key. */
  char *keylim;            /* Limit of first key. */
};

/* Input buffers. */
‌struct buffer
{
  char *buf;            /* Dynamically allocated buffer,
                                   partitioned into 3 regions:
                                   - input data;
                                   - unused area;
                                   - an array of lines, in reverse order.  */
  size_t used;            /* Number of bytes used for input data.  */
  size_t nlines;        /* Number of lines in the line array.  */
  size_t alloc;            /* Number of bytes allocated. */
  size_t left;            /* Number of bytes left from previous reads. */
  size_t line_bytes;        /* Number of bytes to reserve for each line. */
  bool eof;            /* An EOF has been read.  */
};

/* Sort key.  */
‌struct keyfield
{
  size_t sword;            /* Zero-origin 'word' to start at. */
  size_t schar;            /* Additional characters to skip. */
  size_t eword;            /* Zero-origin last 'word' of key. */
  size_t echar;            /* Additional characters in field. */
  bool const *ignore;        /* Boolean array of characters to ignore. */
  char const *translate;    /* Translation applied to characters. */
  bool skipsblanks;        /* Skip leading blanks when finding start.  */
  bool skipeblanks;        /* Skip leading blanks when finding end.  */
  bool numeric;            /* Flag for numeric comparison.  Handle
                                   strings of digits with optional decimal
                                   point, but no exponential notation. */
  bool random;            /* Sort by random hash of key.  */
  bool general_numeric;        /* Flag for general, numeric comparison.
                                   Handle numbers in exponential notation. */
  bool human_numeric;        /* Flag for sorting by human readable
                                   units with either SI or IEC prefixes. */
  bool month;            /* Flag for comparison by month name. */
  bool reverse;            /* Reverse the sense of comparison. */
  bool version;            /* sort by version number */
  bool obsolete_used;        /* obsolescent key option format is used. */
  struct keyfield *next;    /* Next keyfield to try. */
};

‌struct month
{
  char const *name;
  int val;
};

/* Binary merge tree node. */
‌struct merge_node
{
  struct line *lo;              /* Lines to merge from LO child node. */
  struct line *hi;              /* Lines to merge from HI child ndoe. */
  struct line *end_lo;          /* End of available lines from LO. */
  struct line *end_hi;          /* End of available lines from HI. */
  struct line **dest;           /* Pointer to destination of merge. */
  size_t nlo;                   /* Total Lines remaining from LO. */
  size_t nhi;                   /* Total lines remaining from HI. */
  struct merge_node *parent;    /* Parent node. */
  struct merge_node *lo_child;  /* LO child node. */
  struct merge_node *hi_child;  /* HI child node. */
  unsigned int level;           /* Level in merge tree. */
  bool queued;                  /* Node is already in heap. */
  pthread_mutex_t lock;         /* Lock for node operations. */
};

/* Priority queue of merge nodes. */
‌struct merge_node_queue
{
  struct heap *priority_queue;  /* Priority queue of merge tree nodes. */
  pthread_mutex_t mutex;        /* Lock for queue operations. */
  pthread_cond_t cond;          /* Conditional wait for empty queue to populate
                                   when popping. */
};

/* Used to implement --unique (-u).  */
‌static struct line saved_line;

/* FIXME: None of these tables work with multibyte character sets.
   Also, there are many other bugs when handling multibyte characters.
   One way to fix this is to rewrite 'sort' to use wide characters
   internally, but doing this with good performance is a bit
   tricky.  */

/* Table of blanks.  */
‌static bool blanks[UCHAR_LIM];

/* Table of non-printing characters. */
‌static bool nonprinting[UCHAR_LIM];

/* Table of non-dictionary characters (not letters, digits, or blanks). */
‌static bool nondictionary[UCHAR_LIM];

/* Translation table folding lower case to upper.  */
‌static char fold_toupper[UCHAR_LIM];

‌#define MONTHS_PER_YEAR 12

/* Table mapping month names to integers.
   Alphabetic order allows binary search. */
‌static struct month monthtab[] =
{
  {"APR", 4},
  {"AUG", 8},
  {"DEC", 12},
  {"FEB", 2},
  {"JAN", 1},
  {"JUL", 7},
  {"JUN", 6},
  {"MAR", 3},
  {"MAY", 5},
  {"NOV", 11},
  {"OCT", 10},
  {"SEP", 9}
};

/* During the merge phase, the number of files to merge at once. */
‌#define NMERGE_DEFAULT 16

/* Minimum size for a merge or check buffer.  */
‌#define MIN_MERGE_BUFFER_SIZE (2 + sizeof (struct line))

/* Minimum sort size; the code might not work with smaller sizes.  */
‌#define MIN_SORT_SIZE (nmerge * MIN_MERGE_BUFFER_SIZE)

/* The number of bytes needed for a merge or check buffer, which can
   function relatively efficiently even if it holds only one line.  If
   a longer line is seen, this value is increased.  */
‌static size_t merge_buffer_size = MAX (MIN_MERGE_BUFFER_SIZE, 256 * 1024);

/* The approximate maximum number of bytes of main memory to use, as
   specified by the user.  Zero if the user has not specified a size.  */
‌static size_t sort_size;

/* The initial allocation factor for non-regular files.
   This is used, e.g., when reading from a pipe.
   Don't make it too big, since it is multiplied by ~130 to
   obtain the size of the actual buffer sort will allocate.
   Also, there may be 8 threads all doing this at the same time.  */
‌#define INPUT_FILE_SIZE_GUESS (128 * 1024)

/* Array of directory names in which any temporary files are to be created. */
‌static char const **temp_dirs;

/* Number of temporary directory names used.  */
‌static size_t temp_dir_count;

/* Number of allocated slots in temp_dirs.  */
‌static size_t temp_dir_alloc;

/* Flag to reverse the order of all comparisons. */
‌static bool reverse;

/* Flag for stable sort.  This turns off the last ditch bytewise
   comparison of lines, and instead leaves lines in the same order
   they were read if all keys compare equal.  */
‌static bool stable;

/* If TAB has this value, blanks separate fields.  */
enum { TAB_DEFAULT = CHAR_MAX + 1 };

/* Tab character separating fields.  If TAB_DEFAULT, then fields are
   separated by the empty string between a non-blank character and a blank
   character. */
‌static int tab = TAB_DEFAULT;

/* Flag to remove consecutive duplicate lines from the output.
   Only the last of a sequence of equal lines will be output. */
‌static bool unique;

/* Nonzero if any of the input files are the standard input. */
‌static bool have_read_stdin;

/* List of key field comparisons to be tried.  */
‌static struct keyfield *keylist;

/* Program used to (de)compress temp files.  Must accept -d.  */
‌static char const *compress_program;

/* Annotate the output with extra info to aid the user.  */
‌static bool debug;

/* Maximum number of files to merge in one go.  If more than this
   number are present, temp files will be used. */
‌static unsigned int nmerge = NMERGE_DEFAULT;

/* Output an error to stderr using async-signal-safe routines, and _exit().
   This can be used safely from signal handlers,
   and between fork() and exec() of multithreaded processes.  */

static void async_safe_die (int, const char *) ATTRIBUTE_NORETURN;
static void
‌async_safe_die (int errnum, const char *errstr)
{
  ignore_value (write (STDERR_FILENO, errstr, strlen (errstr)));

  /* Even if defined HAVE_STRERROR_R, we can't use it,
     as it may return a translated string etc. and even if not
     may malloc() which is unsafe.  We might improve this
     by testing for sys_errlist and using that if available.
     For now just report the error number.  */
  if (errnum)
    {
      char errbuf[INT_BUFSIZE_BOUND (errnum)];
      char *p = inttostr (errnum, errbuf);
      ignore_value (write (STDERR_FILENO, ": errno ", 8));
      ignore_value (write (STDERR_FILENO, p, strlen (p)));
    }

  ignore_value (write (STDERR_FILENO, "\n", 1));

  _exit (SORT_FAILURE);
}

/* Report MESSAGE for FILE, then clean up and exit.
   If FILE is null, it represents standard output.  */

static void die (char const *, char const *) ATTRIBUTE_NORETURN;
static void
‌die (char const *message, char const *file)
{
  error (0, errno, "%s: %s", message, file ? file : _("standard output"));
  exit (SORT_FAILURE);
}

void
‌usage (int status)
{
  if (status != EXIT_SUCCESS)
    emit_try_help ();
  else
    {
      printf (_("\
Usage: %s [OPTION]... [FILE]...\n\
  or:  %s [OPTION]... --files0-from=F\n\
"),
              program_name, program_name);
      fputs (_("\
Write sorted concatenation of all FILE(s) to standard output.\n\
"), stdout);

      emit_mandatory_arg_note ();

      fputs (_("\
Ordering options:\n\
\n\
"), stdout);
      fputs (_("\
  -b, --ignore-leading-blanks  ignore leading blanks\n\
  -d, --dictionary-order      consider only blanks and alphanumeric characters\
\n\
  -f, --ignore-case           fold lower case to upper case characters\n\
"), stdout);
      fputs (_("\
  -g, --general-numeric-sort  compare according to general numerical value\n\
  -i, --ignore-nonprinting    consider only printable characters\n\
  -M, --month-sort            compare (unknown) < 'JAN' < ... < 'DEC'\n\
"), stdout);
      fputs (_("\
  -h, --human-numeric-sort    compare human readable numbers (e.g., 2K 1G)\n\
"), stdout);
      fputs (_("\
  -n, --numeric-sort          compare according to string numerical value\n\
  -R, --random-sort           sort by random hash of keys\n\
      --random-source=FILE    get random bytes from FILE\n\
  -r, --reverse               reverse the result of comparisons\n\
"), stdout);
      fputs (_("\
      --sort=WORD             sort according to WORD:\n\
                                general-numeric -g, human-numeric -h, month -M,\
\n\
                                numeric -n, random -R, version -V\n\
  -V, --version-sort          natural sort of (version) numbers within text\n\
\n\
"), stdout);
      fputs (_("\
Other options:\n\
\n\
"), stdout);
      fputs (_("\
      --batch-size=NMERGE   merge at most NMERGE inputs at once;\n\
                            for more use temp files\n\
"), stdout);
      fputs (_("\
  -c, --check, --check=diagnose-first  check for sorted input; do not sort\n\
  -C, --check=quiet, --check=silent  like -c, but do not report first bad line\
\n\
      --compress-program=PROG  compress temporaries with PROG;\n\
                              decompress them with PROG -d\n\
"), stdout);
      fputs (_("\
      --debug               annotate the part of the line used to sort,\n\
                              and warn about questionable usage to stderr\n\
      --files0-from=F       read input from the files specified by\n\
                            NUL-terminated names in file F;\n\
                            If F is - then read names from standard input\n\
"), stdout);
      fputs (_("\
  -k, --key=KEYDEF          sort via a key; KEYDEF gives location and type\n\
  -m, --merge               merge already sorted files; do not sort\n\
"), stdout);
      fputs (_("\
  -o, --output=FILE         write result to FILE instead of standard output\n\
  -s, --stable              stabilize sort by disabling last-resort comparison\
\n\
  -S, --buffer-size=SIZE    use SIZE for main memory buffer\n\
"), stdout);
      printf (_("\
  -t, --field-separator=SEP  use SEP instead of non-blank to blank transition\n\
  -T, --temporary-directory=DIR  use DIR for temporaries, not $TMPDIR or %s;\n\
                              multiple options specify multiple directories\n\
      --parallel=N          change the number of sorts run concurrently to N\n\
  -u, --unique              with -c, check for strict ordering;\n\
                              without -c, output only the first of an equal run\
\n\
"), DEFAULT_TMPDIR);
      fputs (_("\
  -z, --zero-terminated     line delimiter is NUL, not newline\n\
"), stdout);
      fputs (HELP_OPTION_DESCRIPTION, stdout);
      fputs (VERSION_OPTION_DESCRIPTION, stdout);
      fputs (_("\
\n\
KEYDEF is F[.C][OPTS][,F[.C][OPTS]] for start and stop position, where F is a\n\
field number and C a character position in the field; both are origin 1, and\n\
the stop position defaults to the line's end.  If neither -t nor -b is in\n\
effect, characters in a field are counted from the beginning of the preceding\n\
whitespace.  OPTS is one or more single-letter ordering options [bdfgiMhnRrV],\
\n\
which override global ordering options for that key.  If no key is given, use\n\
the entire line as the key.\n\
\n\
SIZE may be followed by the following multiplicative suffixes:\n\
"), stdout);
      fputs (_("\
% 1% of memory, b 1, K 1024 (default), and so on for M, G, T, P, E, Z, Y.\n\
\n\
With no FILE, or when FILE is -, read standard input.\n\
\n\
*** WARNING ***\n\
The locale specified by the environment affects sort order.\n\
Set LC_ALL=C to get the traditional sort order that uses\n\
native byte values.\n\
"), stdout );
      emit_ancillary_info ();
    }

  exit (status);
}

/* For long options that have no equivalent short option, use a
   non-character as a pseudo short option, starting with CHAR_MAX + 1.  */
enum
{
  CHECK_OPTION = CHAR_MAX + 1,
  COMPRESS_PROGRAM_OPTION,
  DEBUG_PROGRAM_OPTION,
  FILES0_FROM_OPTION,
  NMERGE_OPTION,
  RANDOM_SOURCE_OPTION,
  SORT_OPTION,
  PARALLEL_OPTION
};

‌static char const short_options[] = "-bcCdfghik:mMno:rRsS:t:T:uVy:z";

‌static struct option const long_options[] =
{
  {"ignore-leading-blanks", no_argument, NULL, 'b'},
  {"check", optional_argument, NULL, CHECK_OPTION},
  {"compress-program", required_argument, NULL, COMPRESS_PROGRAM_OPTION},
  {"debug", no_argument, NULL, DEBUG_PROGRAM_OPTION},
  {"dictionary-order", no_argument, NULL, 'd'},
  {"ignore-case", no_argument, NULL, 'f'},
  {"files0-from", required_argument, NULL, FILES0_FROM_OPTION},
  {"general-numeric-sort", no_argument, NULL, 'g'},
  {"ignore-nonprinting", no_argument, NULL, 'i'},
  {"key", required_argument, NULL, 'k'},
  {"merge", no_argument, NULL, 'm'},
  {"month-sort", no_argument, NULL, 'M'},
  {"numeric-sort", no_argument, NULL, 'n'},
  {"human-numeric-sort", no_argument, NULL, 'h'},
  {"version-sort", no_argument, NULL, 'V'},
  {"random-sort", no_argument, NULL, 'R'},
  {"random-source", required_argument, NULL, RANDOM_SOURCE_OPTION},
  {"sort", required_argument, NULL, SORT_OPTION},
  {"output", required_argument, NULL, 'o'},
  {"reverse", no_argument, NULL, 'r'},
  {"stable", no_argument, NULL, 's'},
  {"batch-size", required_argument, NULL, NMERGE_OPTION},
  {"buffer-size", required_argument, NULL, 'S'},
  {"field-separator", required_argument, NULL, 't'},
  {"temporary-directory", required_argument, NULL, 'T'},
  {"unique", no_argument, NULL, 'u'},
  {"zero-terminated", no_argument, NULL, 'z'},
  {"parallel", required_argument, NULL, PARALLEL_OPTION},
  {GETOPT_HELP_OPTION_DECL},
  {GETOPT_VERSION_OPTION_DECL},
  {NULL, 0, NULL, 0},
};

‌#define CHECK_TABLE \
  _ct_("quiet",          'C') \
  _ct_("silent",         'C') \
  _ct_("diagnose-first", 'c')

‌static char const *const check_args[] =
{
‌#define _ct_(_s, _c) _s,
  CHECK_TABLE NULL
‌#undef  _ct_
};
‌static char const check_types[] =
{
‌#define _ct_(_s, _c) _c,
  CHECK_TABLE
‌#undef  _ct_
};

‌#define SORT_TABLE \
  _st_("general-numeric", 'g') \
  _st_("human-numeric",   'h') \
  _st_("month",           'M') \
  _st_("numeric",         'n') \
  _st_("random",          'R') \
  _st_("version",         'V')

‌static char const *const sort_args[] =
{
‌#define _st_(_s, _c) _s,
  SORT_TABLE NULL
‌#undef  _st_
};
‌static char const sort_types[] =
{
‌#define _st_(_s, _c) _c,
  SORT_TABLE
‌#undef  _st_
};

/* The set of signals that are caught.  */
‌static sigset_t caught_signals;

/* Critical section status.  */
‌struct cs_status
{
  bool valid;
  sigset_t sigs;
};

/* Enter a critical section.  */
static struct cs_status
‌cs_enter (void)
{
  struct cs_status status;
  status.valid = (sigprocmask (SIG_BLOCK, &caught_signals, &status.sigs) == 0);
  return status;
}

/* Leave a critical section.  */
static void
‌cs_leave (struct cs_status status)
{
  if (status.valid)
    {
      /* Ignore failure when restoring the signal mask. */
      sigprocmask (SIG_SETMASK, &status.sigs, NULL);
    }
}

/* Possible states for a temp file.  If compressed, the file's status
   is unreaped or reaped, depending on whether 'sort' has waited for
   the subprocess to finish.  */
enum { UNCOMPRESSED, UNREAPED, REAPED };

/* The list of temporary files. */
‌struct tempnode
{
  struct tempnode *volatile next;
  pid_t pid;     /* The subprocess PID; undefined if state == UNCOMPRESSED.  */
  char state;
  char name[1];  /* Actual size is 1 + file name length.  */
};
‌static struct tempnode *volatile temphead;
‌static struct tempnode *volatile *temptail = &temphead;

/* A file to be sorted.  */
‌struct sortfile
{
  /* The file's name.  */
  char const *name;

  /* Nonnull if this is a temporary file, in which case NAME == TEMP->name.  */
  struct tempnode *temp;
};

/* Map PIDs of unreaped subprocesses to their struct tempnode objects.  */
‌static Hash_table *proctab;

enum { INIT_PROCTAB_SIZE = 47 };

static size_t
‌proctab_hasher (void const *entry, size_t tabsize)
{
  struct tempnode const *node = entry;
  return node->pid % tabsize;
}

static bool
‌proctab_comparator (void const *e1, void const *e2)
{
  struct tempnode const *n1 = e1;
  struct tempnode const *n2 = e2;
  return n1->pid == n2->pid;
}

/* The number of unreaped child processes.  */
‌static pid_t nprocs;

static bool delete_proc (pid_t);

/* If PID is positive, wait for the child process with that PID to
   exit, and assume that PID has already been removed from the process
   table.  If PID is 0 or -1, clean up some child that has exited (by
   waiting for it, and removing it from the proc table) and return the
   child's process ID.  However, if PID is 0 and no children have
   exited, return 0 without waiting.  */

static pid_t
‌reap (pid_t pid)
{
  int status;
  pid_t cpid = waitpid ((pid ? pid : -1), &status, (pid ? 0 : WNOHANG));

  if (cpid < 0)
    error (SORT_FAILURE, errno, _("waiting for %s [-d]"),
           compress_program);
  else if (0 < cpid && (0 < pid || delete_proc (cpid)))
    {
      if (! WIFEXITED (status) || WEXITSTATUS (status))
        error (SORT_FAILURE, 0, _("%s [-d] terminated abnormally"),
               compress_program);
      --nprocs;
    }

  return cpid;
}

/* TEMP represents a new process; add it to the process table.  Create
   the process table the first time it's called.  */

static void
‌register_proc (struct tempnode *temp)
{
  if (! proctab)
    {
      proctab = hash_initialize (INIT_PROCTAB_SIZE, NULL,
                                 proctab_hasher,
                                 proctab_comparator,
                                 NULL);
      if (! proctab)
        xalloc_die ();
    }

  temp->state = UNREAPED;

  if (! hash_insert (proctab, temp))
    xalloc_die ();
}

/* If PID is in the process table, remove it and return true.
   Otherwise, return false.  */

static bool
‌delete_proc (pid_t pid)
{
  struct tempnode test;

  test.pid = pid;
  struct tempnode *node = hash_delete (proctab, &test);
  if (! node)
    return false;
  node->state = REAPED;
  return true;
}

/* Remove PID from the process table, and wait for it to exit if it
   hasn't already.  */

static void
‌wait_proc (pid_t pid)
{
  if (delete_proc (pid))
    reap (pid);
}

/* Reap any exited children.  Do not block; reap only those that have
   already exited.  */

static void
‌reap_exited (void)
{
  while (0 < nprocs && reap (0))
    continue;
}

/* Reap at least one exited child, waiting if necessary.  */

static void
‌reap_some (void)
{
  reap (-1);
  reap_exited ();
}

/* Reap all children, waiting if necessary.  */

static void
‌reap_all (void)
{
  while (0 < nprocs)
    reap (-1);
}

/* Clean up any remaining temporary files.  */

static void
‌cleanup (void)
{
  struct tempnode const *node;

  for (node = temphead; node; node = node->next)
    unlink (node->name);
  temphead = NULL;
}

/* Cleanup actions to take when exiting.  */

static void
‌exit_cleanup (void)
{
  if (temphead)
    {
      /* Clean up any remaining temporary files in a critical section so
         that a signal handler does not try to clean them too.  */
      struct cs_status cs = cs_enter ();
      cleanup ();
      cs_leave (cs);
    }

  close_stdout ();
}

/* Create a new temporary file, returning its newly allocated tempnode.
   Store into *PFD the file descriptor open for writing.
   If the creation fails, return NULL and store -1 into *PFD if the
   failure is due to file descriptor exhaustion and
   SURVIVE_FD_EXHAUSTION; otherwise, die.  */

static struct tempnode *
‌create_temp_file (int *pfd, bool survive_fd_exhaustion)
{
  static char const slashbase[] = "/sortXXXXXX";
  static size_t temp_dir_index;
  int fd;
  int saved_errno;
  char const *temp_dir = temp_dirs[temp_dir_index];
  size_t len = strlen (temp_dir);
  struct tempnode *node =
    xmalloc (offsetof (struct tempnode, name) + len + sizeof slashbase);
  char *file = node->name;
  struct cs_status cs;

  memcpy (file, temp_dir, len);
  memcpy (file + len, slashbase, sizeof slashbase);
  node->next = NULL;
  if (++temp_dir_index == temp_dir_count)
    temp_dir_index = 0;

  /* Create the temporary file in a critical section, to avoid races.  */
  cs = cs_enter ();
  fd = mkstemp (file);
  if (0 <= fd)
    {
      *temptail = node;
      temptail = &node->next;
    }
  saved_errno = errno;
  cs_leave (cs);
  errno = saved_errno;

  if (fd < 0)
    {
      if (! (survive_fd_exhaustion && errno == EMFILE))
        error (SORT_FAILURE, errno, _("cannot create temporary file in %s"),
               quote (temp_dir));
      free (node);
      node = NULL;
    }

  *pfd = fd;
  return node;
}

/* Return a stream for FILE, opened with mode HOW.  A null FILE means
   standard output; HOW should be "w".  When opening for input, "-"
   means standard input.  To avoid confusion, do not return file
   descriptors STDIN_FILENO, STDOUT_FILENO, or STDERR_FILENO when
   opening an ordinary FILE.  Return NULL if unsuccessful.

   fadvise() is used to specify an access pattern for input files.
   There are a few hints we could possibly provide,
   and after careful testing it was decided that
   specifying POSIX_FADV_SEQUENTIAL was not detrimental
   to any cases.  On Linux 2.6.31, this option doubles
   the size of read ahead performed and thus was seen to
   benefit these cases:
     Merging
     Sorting with a smaller internal buffer
     Reading from faster flash devices

   In _addition_ one could also specify other hints...

   POSIX_FADV_WILLNEED was tested, but Linux 2.6.31
   at least uses that to _synchronously_ prepopulate the cache
   with the specified range.  While sort does need to
   read all of its input before outputting, a synchronous
   read of the whole file up front precludes any processing
   that sort could do in parallel with the system doing
   read ahead of the data. This was seen to have negative effects
   in a couple of cases:
     Merging
     Sorting with a smaller internal buffer
   Note this option was seen to shorten the runtime for sort
   on a multicore system with lots of RAM and other processes
   competing for CPU.  It could be argued that more explicit
   scheduling hints with 'nice' et. al. are more appropriate
   for this situation.

   POSIX_FADV_NOREUSE is a possibility as it could lower
   the priority of input data in the cache as sort will
   only need to process it once.  However its functionality
   has changed over Linux kernel versions and as of 2.6.31
   it does nothing and thus we can't depend on what it might
   do in future.

   POSIX_FADV_DONTNEED is not appropriate for user specified
   input files, but for temp files we do want to drop the
   cache immediately after processing.  This is done implicitly
   however when the files are unlinked.  */

static FILE *
‌stream_open (char const *file, char const *how)
{
  FILE *fp;

  if (*how == 'r')
    {
      if (STREQ (file, "-"))
        {
          have_read_stdin = true;
          fp = stdin;
        }
      else
        fp = fopen (file, how);
      fadvise (fp, FADVISE_SEQUENTIAL);
    }
  else if (*how == 'w')
    {
      if (file && ftruncate (STDOUT_FILENO, 0) != 0)
        error (SORT_FAILURE, errno, _("%s: error truncating"),
               quote (file));
      fp = stdout;
    }
  else
    assert (!"unexpected mode passed to stream_open");

  return fp;
}

/* Same as stream_open, except always return a non-null value; die on
   failure.  */

static FILE *
‌xfopen (char const *file, char const *how)
{
  FILE *fp = stream_open (file, how);
  if (!fp)
    die (_("open failed"), file);
  return fp;
}

/* Close FP, whose name is FILE, and report any errors.  */

static void
‌xfclose (FILE *fp, char const *file)
{
  switch (fileno (fp))
    {
    case STDIN_FILENO:
      /* Allow reading stdin from tty more than once.  */
      if (feof (fp))
        clearerr (fp);
      break;

    case STDOUT_FILENO:
      /* Don't close stdout just yet.  close_stdout does that.  */
      if (fflush (fp) != 0)
        die (_("fflush failed"), file);
      break;

    default:
      if (fclose (fp) != 0)
        die (_("close failed"), file);
      break;
    }
}

static void
‌move_fd_or_die (int oldfd, int newfd)
{
  if (oldfd != newfd)
    {
      /* This should never fail for our usage.  */
      dup2 (oldfd, newfd);
      close (oldfd);
    }
}

/* Fork a child process for piping to and do common cleanup.  The
   TRIES parameter tells us how many times to try to fork before
   giving up.  Return the PID of the child, or -1 (setting errno)
   on failure. */

static pid_t
‌pipe_fork (int pipefds[2], size_t tries)
{
#if HAVE_WORKING_FORK
  struct tempnode *saved_temphead;
  int saved_errno;
  double wait_retry = 0.25;
  pid_t pid IF_LINT ( = -1);
  struct cs_status cs;

  if (pipe (pipefds) < 0)
    return -1;

  /* At least NMERGE + 1 subprocesses are needed.  More could be created, but
     uncontrolled subprocess generation can hurt performance significantly.
     Allow at most NMERGE + 2 subprocesses, on the theory that there
     may be some useful parallelism by letting compression for the
     previous merge finish (1 subprocess) in parallel with the current
     merge (NMERGE + 1 subprocesses).  */

  if (nmerge + 1 < nprocs)
    reap_some ();

  while (tries--)
    {
      /* This is so the child process won't delete our temp files
         if it receives a signal before exec-ing.  */
      cs = cs_enter ();
      saved_temphead = temphead;
      temphead = NULL;

      pid = fork ();
      saved_errno = errno;
      if (pid)
        temphead = saved_temphead;

      cs_leave (cs);
      errno = saved_errno;

      if (0 <= pid || errno != EAGAIN)
        break;
      else
        {
          xnanosleep (wait_retry);
          wait_retry *= 2;
          reap_exited ();
        }
    }

  if (pid < 0)
    {
      saved_errno = errno;
      close (pipefds[0]);
      close (pipefds[1]);
      errno = saved_errno;
    }
  else if (pid == 0)
    {
      close (STDIN_FILENO);
      close (STDOUT_FILENO);
    }
  else
    ++nprocs;

  return pid;

#else  /* ! HAVE_WORKING_FORK */
  return -1;
#endif
}

/* Create a temporary file and, if asked for, start a compressor
   to that file.  Set *PFP to the file handle and return
   the address of the new temp node.  If the creation
   fails, return NULL if the failure is due to file descriptor
   exhaustion and SURVIVE_FD_EXHAUSTION; otherwise, die.  */

static struct tempnode *
‌maybe_create_temp (FILE **pfp, bool survive_fd_exhaustion)
{
  int tempfd;
  struct tempnode *node = create_temp_file (&tempfd, survive_fd_exhaustion);
  if (! node)
    return NULL;

  node->state = UNCOMPRESSED;

  if (compress_program)
    {
      int pipefds[2];

      node->pid = pipe_fork (pipefds, MAX_FORK_TRIES_COMPRESS);
      if (0 < node->pid)
        {
          close (tempfd);
          close (pipefds[0]);
          tempfd = pipefds[1];

          register_proc (node);
        }
      else if (node->pid == 0)
        {
          /* Being the child of a multithreaded program before exec(),
             we're restricted to calling async-signal-safe routines here.  */
          close (pipefds[1]);
          move_fd_or_die (tempfd, STDOUT_FILENO);
          move_fd_or_die (pipefds[0], STDIN_FILENO);

          execlp (compress_program, compress_program, (char *) NULL);

          async_safe_die (errno, "couldn't execute compress program");
        }
    }

  *pfp = fdopen (tempfd, "w");
  if (! *pfp)
    die (_("couldn't create temporary file"), node->name);

  return node;
}

/* Create a temporary file and, if asked for, start a compressor
   to that file.  Set *PFP to the file handle and return the address
   of the new temp node.  Die on failure.  */

static struct tempnode *
‌create_temp (FILE **pfp)
{
  return maybe_create_temp (pfp, false);
}

/* Open a compressed temp file and start a decompression process through
   which to filter the input.  Return NULL (setting errno to
   EMFILE) if we ran out of file descriptors, and die on any other
   kind of failure.  */

static FILE *
‌open_temp (struct tempnode *temp)
{
  int tempfd, pipefds[2];
  FILE *fp = NULL;

  if (temp->state == UNREAPED)
    wait_proc (temp->pid);

  tempfd = open (temp->name, O_RDONLY);
  if (tempfd < 0)
    return NULL;

  pid_t child = pipe_fork (pipefds, MAX_FORK_TRIES_DECOMPRESS);

  switch (child)
    {
    case -1:
      if (errno != EMFILE)
        error (SORT_FAILURE, errno, _("couldn't create process for %s -d"),
               compress_program);
      close (tempfd);
      errno = EMFILE;
      break;

    case 0:
      /* Being the child of a multithreaded program before exec(),
         we're restricted to calling async-signal-safe routines here.  */
      close (pipefds[0]);
      move_fd_or_die (tempfd, STDIN_FILENO);
      move_fd_or_die (pipefds[1], STDOUT_FILENO);

      execlp (compress_program, compress_program, "-d", (char *) NULL);

      async_safe_die (errno, "couldn't execute compress program (with -d)");

    default:
      temp->pid = child;
      register_proc (temp);
      close (tempfd);
      close (pipefds[1]);

      fp = fdopen (pipefds[0], "r");
      if (! fp)
        {
          int saved_errno = errno;
          close (pipefds[0]);
          errno = saved_errno;
        }
      break;
    }

  return fp;
}

/* Append DIR to the array of temporary directory names.  */
static void
‌add_temp_dir (char const *dir)
{
  if (temp_dir_count == temp_dir_alloc)
    temp_dirs = X2NREALLOC (temp_dirs, &temp_dir_alloc);

  temp_dirs[temp_dir_count++] = dir;
}

/* Remove NAME from the list of temporary files.  */

static void
‌zaptemp (char const *name)
{
  struct tempnode *volatile *pnode;
  struct tempnode *node;
  struct tempnode *next;
  int unlink_status;
  int unlink_errno = 0;
  struct cs_status cs;

  for (pnode = &temphead; (node = *pnode)->name != name; pnode = &node->next)
    continue;

  if (node->state == UNREAPED)
    wait_proc (node->pid);

  /* Unlink the temporary file in a critical section to avoid races.  */
  next = node->next;
  cs = cs_enter ();
  unlink_status = unlink (name);
  unlink_errno = errno;
  *pnode = next;
  cs_leave (cs);

  if (unlink_status != 0)
    error (0, unlink_errno, _("warning: cannot remove: %s"), name);
  if (! next)
    temptail = pnode;
  free (node);
}

#if HAVE_NL_LANGINFO

static int
‌struct_month_cmp (void const *m1, void const *m2)
{
  struct month const *month1 = m1;
  struct month const *month2 = m2;
  return strcmp (month1->name, month2->name);
}

#endif

/* Initialize the character class tables. */

static void
‌inittables (void)
{
  size_t i;

  for (i = 0; i < UCHAR_LIM; ++i)
    {
      blanks[i] = !! isblank (i);
      nonprinting[i] = ! isprint (i);
      nondictionary[i] = ! isalnum (i) && ! isblank (i);
      fold_toupper[i] = toupper (i);
    }

#if HAVE_NL_LANGINFO
  /* If we're not in the "C" locale, read different names for months.  */
  if (hard_LC_TIME)
    {
      for (i = 0; i < MONTHS_PER_YEAR; i++)
        {
          char const *s;
          size_t s_len;
          size_t j, k;
          char *name;

          s = nl_langinfo (ABMON_1 + i);
          s_len = strlen (s);
          monthtab[i].name = name = xmalloc (s_len + 1);
          monthtab[i].val = i + 1;

          for (j = k = 0; j < s_len; j++)
            if (! isblank (to_uchar (s[j])))
              name[k++] = fold_toupper[to_uchar (s[j])];
          name[k] = '\0';
        }
      qsort (monthtab, MONTHS_PER_YEAR, sizeof *monthtab, struct_month_cmp);
    }
#endif
}

/* Specify how many inputs may be merged at once.
   This may be set on the command-line with the
   --batch-size option. */
static void
‌specify_nmerge (int oi, char c, char const *s)
{
  uintmax_t n;
  struct rlimit rlimit;
  enum strtol_error e = xstrtoumax (s, NULL, 10, &n, NULL);

  /* Try to find out how many file descriptors we'll be able
     to open.  We need at least nmerge + 3 (STDIN_FILENO,
     STDOUT_FILENO and STDERR_FILENO). */
  unsigned int max_nmerge = ((getrlimit (RLIMIT_NOFILE, &rlimit) == 0
                              ? rlimit.rlim_cur
                              : OPEN_MAX)
                             - 3);

  if (e == LONGINT_OK)
    {
      nmerge = n;
      if (nmerge != n)
        e = LONGINT_OVERFLOW;
      else
        {
          if (nmerge < 2)
            {
              error (0, 0, _("invalid --%s argument %s"),
                     long_options[oi].name, quote (s));
              error (SORT_FAILURE, 0,
                     _("minimum --%s argument is %s"),
                     long_options[oi].name, quote ("2"));
            }
          else if (max_nmerge < nmerge)
            {
              e = LONGINT_OVERFLOW;
            }
          else
            return;
        }
    }

  if (e == LONGINT_OVERFLOW)
    {
      char max_nmerge_buf[INT_BUFSIZE_BOUND (max_nmerge)];
      error (0, 0, _("--%s argument %s too large"),
             long_options[oi].name, quote (s));
      error (SORT_FAILURE, 0,
             _("maximum --%s argument with current rlimit is %s"),
             long_options[oi].name,
             uinttostr (max_nmerge, max_nmerge_buf));
    }
  else
    xstrtol_fatal (e, oi, c, long_options, s);
}

/* Specify the amount of main memory to use when sorting.  */
static void
‌specify_sort_size (int oi, char c, char const *s)
{
  uintmax_t n;
  char *suffix;
  enum strtol_error e = xstrtoumax (s, &suffix, 10, &n, "EgGkKmMPtTYZ");

  /* The default unit is KiB.  */
  if (e == LONGINT_OK && ISDIGIT (suffix[-1]))
    {
      if (n <= UINTMAX_MAX / 1024)
        n *= 1024;
      else
        e = LONGINT_OVERFLOW;
    }

  /* A 'b' suffix means bytes; a '%' suffix means percent of memory.  */
  if (e == LONGINT_INVALID_SUFFIX_CHAR && ISDIGIT (suffix[-1]) && ! suffix[1])
    switch (suffix[0])
      {
      case 'b':
        e = LONGINT_OK;
        break;

      case '%':
        {
          double mem = physmem_total () * n / 100;

          /* Use "<", not "<=", to avoid problems with rounding.  */
          if (mem < UINTMAX_MAX)
            {
              n = mem;
              e = LONGINT_OK;
            }
          else
            e = LONGINT_OVERFLOW;
        }
        break;
      }

  if (e == LONGINT_OK)
    {
      /* If multiple sort sizes are specified, take the maximum, so
         that option order does not matter.  */
      if (n < sort_size)
        return;

      sort_size = n;
      if (sort_size == n)
        {
          sort_size = MAX (sort_size, MIN_SORT_SIZE);
          return;
        }

      e = LONGINT_OVERFLOW;
    }

  xstrtol_fatal (e, oi, c, long_options, s);
}

/* Specify the number of threads to spawn during internal sort.  */
static size_t
‌specify_nthreads (int oi, char c, char const *s)
{
  unsigned long int nthreads;
  enum strtol_error e = xstrtoul (s, NULL, 10, &nthreads, "");
  if (e == LONGINT_OVERFLOW)
    return SIZE_MAX;
  if (e != LONGINT_OK)
    xstrtol_fatal (e, oi, c, long_options, s);
  if (SIZE_MAX < nthreads)
    nthreads = SIZE_MAX;
  if (nthreads == 0)
    error (SORT_FAILURE, 0, _("number in parallel must be nonzero"));
  return nthreads;
}

/* Return the default sort size.  */
static size_t
‌default_sort_size (void)
{
  /* Let SIZE be MEM, but no more than the maximum object size,
     total memory, or system resource limits.  Don't bother to check
     for values like RLIM_INFINITY since in practice they are not much
     less than SIZE_MAX.  */
  size_t size = SIZE_MAX;
  struct rlimit rlimit;
  if (getrlimit (RLIMIT_DATA, &rlimit) == 0 && rlimit.rlim_cur < size)
    size = rlimit.rlim_cur;
#ifdef RLIMIT_AS
  if (getrlimit (RLIMIT_AS, &rlimit) == 0 && rlimit.rlim_cur < size)
    size = rlimit.rlim_cur;
#endif

  /* Leave a large safety margin for the above limits, as failure can
     occur when they are exceeded.  */
  size /= 2;

#ifdef RLIMIT_RSS
  /* Leave a 1/16 margin for RSS to leave room for code, stack, etc.
     Exceeding RSS is not fatal, but can be quite slow.  */
  if (getrlimit (RLIMIT_RSS, &rlimit) == 0 && rlimit.rlim_cur / 16 * 15 < size)
    size = rlimit.rlim_cur / 16 * 15;
#endif

  /* Let MEM be available memory or 1/8 of total memory, whichever
     is greater.  */
  double avail = physmem_available ();
  double total = physmem_total ();
  double mem = MAX (avail, total / 8);

  /* Leave a 1/4 margin for physical memory.  */
  if (total * 0.75 < size)
    size = total * 0.75;

  /* Return the minimum of MEM and SIZE, but no less than
     MIN_SORT_SIZE.  Avoid the MIN macro here, as it is not quite
     right when only one argument is floating point.  */
  if (mem < size)
    size = mem;
  return MAX (size, MIN_SORT_SIZE);
}

/* Return the sort buffer size to use with the input files identified
   by FPS and FILES, which are alternate names of the same files.
   NFILES gives the number of input files; NFPS may be less.  Assume
   that each input line requires LINE_BYTES extra bytes' worth of line
   information.  Do not exceed the size bound specified by the user
   (or a default size bound, if the user does not specify one).  */

static size_t
‌sort_buffer_size (FILE *const *fps, size_t nfps,
                  char *const *files, size_t nfiles,
                  size_t line_bytes)
{
  /* A bound on the input size.  If zero, the bound hasn't been
     determined yet.  */
  static size_t size_bound;

  /* In the worst case, each input byte is a newline.  */
  size_t worst_case_per_input_byte = line_bytes + 1;

  /* Keep enough room for one extra input line and an extra byte.
     This extra room might be needed when preparing to read EOF.  */
  size_t size = worst_case_per_input_byte + 1;

  size_t i;

  for (i = 0; i < nfiles; i++)
    {
      struct stat st;
      off_t file_size;
      size_t worst_case;

      if ((i < nfps ? fstat (fileno (fps[i]), &st)
           : STREQ (files[i], "-") ? fstat (STDIN_FILENO, &st)
           : stat (files[i], &st))
          != 0)
        die (_("stat failed"), files[i]);

      if (S_ISREG (st.st_mode))
        file_size = st.st_size;
      else
        {
          /* The file has unknown size.  If the user specified a sort
             buffer size, use that; otherwise, guess the size.  */
          if (sort_size)
            return sort_size;
          file_size = INPUT_FILE_SIZE_GUESS;
        }

      if (! size_bound)
        {
          size_bound = sort_size;
          if (! size_bound)
            size_bound = default_sort_size ();
        }

      /* Add the amount of memory needed to represent the worst case
         where the input consists entirely of newlines followed by a
         single non-newline.  Check for overflow.  */
      worst_case = file_size * worst_case_per_input_byte + 1;
      if (file_size != worst_case / worst_case_per_input_byte
          || size_bound - size <= worst_case)
        return size_bound;
      size += worst_case;
    }

  return size;
}

/* Initialize BUF.  Reserve LINE_BYTES bytes for each line; LINE_BYTES
   must be at least sizeof (struct line).  Allocate ALLOC bytes
   initially.  */

static void
‌initbuf (struct buffer *buf, size_t line_bytes, size_t alloc)
{
  /* Ensure that the line array is properly aligned.  If the desired
     size cannot be allocated, repeatedly halve it until allocation
     succeeds.  The smaller allocation may hurt overall performance,
     but that's better than failing.  */
  while (true)
    {
      alloc += sizeof (struct line) - alloc % sizeof (struct line);
      buf->buf = malloc (alloc);
      if (buf->buf)
        break;
      alloc /= 2;
      if (alloc <= line_bytes + 1)
        xalloc_die ();
    }

  buf->line_bytes = line_bytes;
  buf->alloc = alloc;
  buf->used = buf->left = buf->nlines = 0;
  buf->eof = false;
}

/* Return one past the limit of the line array.  */

static inline struct line *
‌buffer_linelim (struct buffer const *buf)
{
  void *linelim = buf->buf + buf->alloc;
  return linelim;
}

/* Return a pointer to the first character of the field specified
   by KEY in LINE. */

static char *
‌begfield (struct line const *line, struct keyfield const *key)
{
  char *ptr = line->text, *lim = ptr + line->length - 1;
  size_t sword = key->sword;
  size_t schar = key->schar;

  /* The leading field separator itself is included in a field when -t
     is absent.  */

  if (tab != TAB_DEFAULT)
    while (ptr < lim && sword--)
      {
        while (ptr < lim && *ptr != tab)
          ++ptr;
        if (ptr < lim)
          ++ptr;
      }
  else
    while (ptr < lim && sword--)
      {
        while (ptr < lim && blanks[to_uchar (*ptr)])
          ++ptr;
        while (ptr < lim && !blanks[to_uchar (*ptr)])
          ++ptr;
      }

  /* If we're ignoring leading blanks when computing the Start
     of the field, skip past them here.  */
  if (key->skipsblanks)
    while (ptr < lim && blanks[to_uchar (*ptr)])
      ++ptr;

  /* Advance PTR by SCHAR (if possible), but no further than LIM.  */
  ptr = MIN (lim, ptr + schar);

  return ptr;
}

/* Return the limit of (a pointer to the first character after) the field
   in LINE specified by KEY. */

static char *
‌limfield (struct line const *line, struct keyfield const *key)
{
  char *ptr = line->text, *lim = ptr + line->length - 1;
  size_t eword = key->eword, echar = key->echar;

  if (echar == 0)
    eword++; /* Skip all of end field.  */

  /* Move PTR past EWORD fields or to one past the last byte on LINE,
     whichever comes first.  If there are more than EWORD fields, leave
     PTR pointing at the beginning of the field having zero-based index,
     EWORD.  If a delimiter character was specified (via -t), then that
     'beginning' is the first character following the delimiting TAB.
     Otherwise, leave PTR pointing at the first 'blank' character after
     the preceding field.  */
  if (tab != TAB_DEFAULT)
    while (ptr < lim && eword--)
      {
        while (ptr < lim && *ptr != tab)
          ++ptr;
        if (ptr < lim && (eword || echar))
          ++ptr;
      }
  else
    while (ptr < lim && eword--)
      {
        while (ptr < lim && blanks[to_uchar (*ptr)])
          ++ptr;
        while (ptr < lim && !blanks[to_uchar (*ptr)])
          ++ptr;
      }

#ifdef POSIX_UNSPECIFIED
  /* The following block of code makes GNU sort incompatible with
     standard Unix sort, so it's ifdef'd out for now.
     The POSIX spec isn't clear on how to interpret this.
     FIXME: request clarification.

     From: kwzh@gnu.ai.mit.edu (Karl Heuer)
     Date: Thu, 30 May 96 12:20:41 -0400
     [Translated to POSIX 1003.1-2001 terminology by Paul Eggert.]

     [...]I believe I've found another bug in 'sort'.

     $ cat /tmp/sort.in
     a b c 2 d
     pq rs 1 t
     $ textutils-1.15/src/sort -k1.7,1.7 </tmp/sort.in
     a b c 2 d
     pq rs 1 t
     $ /bin/sort -k1.7,1.7 </tmp/sort.in
     pq rs 1 t
     a b c 2 d

     Unix sort produced the answer I expected: sort on the single character
     in column 7.  GNU sort produced different results, because it disagrees
     on the interpretation of the key-end spec "M.N".  Unix sort reads this
     as "skip M-1 fields, then N-1 characters"; but GNU sort wants it to mean
     "skip M-1 fields, then either N-1 characters or the rest of the current
     field, whichever comes first".  This extra clause applies only to
     key-ends, not key-starts.
     */

  /* Make LIM point to the end of (one byte past) the current field.  */
  if (tab != TAB_DEFAULT)
    {
      char *newlim;
      newlim = memchr (ptr, tab, lim - ptr);
      if (newlim)
        lim = newlim;
    }
  else
    {
      char *newlim;
      newlim = ptr;
      while (newlim < lim && blanks[to_uchar (*newlim)])
        ++newlim;
      while (newlim < lim && !blanks[to_uchar (*newlim)])
        ++newlim;
      lim = newlim;
    }
#endif

  if (echar != 0) /* We need to skip over a portion of the end field.  */
    {
      /* If we're ignoring leading blanks when computing the End
         of the field, skip past them here.  */
      if (key->skipeblanks)
        while (ptr < lim && blanks[to_uchar (*ptr)])
          ++ptr;

      /* Advance PTR by ECHAR (if possible), but no further than LIM.  */
      ptr = MIN (lim, ptr + echar);
    }

  return ptr;
}

/* Fill BUF reading from FP, moving buf->left bytes from the end
   of buf->buf to the beginning first.  If EOF is reached and the
   file wasn't terminated by a newline, supply one.  Set up BUF's line
   table too.  FILE is the name of the file corresponding to FP.
   Return true if some input was read.  */

static bool
‌fillbuf (struct buffer *buf, FILE *fp, char const *file)
{
  struct keyfield const *key = keylist;
  char eol = eolchar;
  size_t line_bytes = buf->line_bytes;
  size_t mergesize = merge_buffer_size - MIN_MERGE_BUFFER_SIZE;

  if (buf->eof)
    return false;

  if (buf->used != buf->left)
    {
      memmove (buf->buf, buf->buf + buf->used - buf->left, buf->left);
      buf->used = buf->left;
      buf->nlines = 0;
    }

  while (true)
    {
      char *ptr = buf->buf + buf->used;
      struct line *linelim = buffer_linelim (buf);
      struct line *line = linelim - buf->nlines;
      size_t avail = (char *) linelim - buf->nlines * line_bytes - ptr;
      char *line_start = buf->nlines ? line->text + line->length : buf->buf;

      while (line_bytes + 1 < avail)
        {
          /* Read as many bytes as possible, but do not read so many
             bytes that there might not be enough room for the
             corresponding line array.  The worst case is when the
             rest of the input file consists entirely of newlines,
             except that the last byte is not a newline.  */
          size_t readsize = (avail - 1) / (line_bytes + 1);
          size_t bytes_read = fread (ptr, 1, readsize, fp);
          char *ptrlim = ptr + bytes_read;
          char *p;
          avail -= bytes_read;

          if (bytes_read != readsize)
            {
              if (ferror (fp))
                die (_("read failed"), file);
              if (feof (fp))
                {
                  buf->eof = true;
                  if (buf->buf == ptrlim)
                    return false;
                  if (line_start != ptrlim && ptrlim[-1] != eol)
                    *ptrlim++ = eol;
                }
            }

          /* Find and record each line in the just-read input.  */
          while ((p = memchr (ptr, eol, ptrlim - ptr)))
            {
              /* Delimit the line with NUL. This eliminates the need to
                 temporarily replace the last byte with NUL when calling
                 xmemcoll(), which increases performance.  */
              *p = '\0';
              ptr = p + 1;
              line--;
              line->text = line_start;
              line->length = ptr - line_start;
              mergesize = MAX (mergesize, line->length);
              avail -= line_bytes;

              if (key)
                {
                  /* Precompute the position of the first key for
                     efficiency.  */
                  line->keylim = (key->eword == SIZE_MAX
                                  ? p
                                  : limfield (line, key));

                  if (key->sword != SIZE_MAX)
                    line->keybeg = begfield (line, key);
                  else
                    {
                      if (key->skipsblanks)
                        while (blanks[to_uchar (*line_start)])
                          line_start++;
                      line->keybeg = line_start;
                    }
                }

              line_start = ptr;
            }

          ptr = ptrlim;
          if (buf->eof)
            break;
        }

      buf->used = ptr - buf->buf;
      buf->nlines = buffer_linelim (buf) - line;
      if (buf->nlines != 0)
        {
          buf->left = ptr - line_start;
          merge_buffer_size = mergesize + MIN_MERGE_BUFFER_SIZE;
          return true;
        }

      {
        /* The current input line is too long to fit in the buffer.
           Increase the buffer size and try again, keeping it properly
           aligned.  */
        size_t line_alloc = buf->alloc / sizeof (struct line);
        buf->buf = x2nrealloc (buf->buf, &line_alloc, sizeof (struct line));
        buf->alloc = line_alloc * sizeof (struct line);
      }
    }
}

/* Table that maps characters to order-of-magnitude values.  */
‌static char const unit_order[UCHAR_LIM] =
  {
#if ! ('K' == 75 && 'M' == 77 && 'G' == 71 && 'T' == 84 && 'P' == 80 \
     && 'E' == 69 && 'Z' == 90 && 'Y' == 89 && 'k' == 107)
    /* This initializer syntax works on all C99 hosts.  For now, use
       it only on non-ASCII hosts, to ease the pain of porting to
       pre-C99 ASCII hosts.  */
    ['K']=1, ['M']=2, ['G']=3, ['T']=4, ['P']=5, ['E']=6, ['Z']=7, ['Y']=8,
    ['k']=1,
#else
    /* Generate the following table with this command:
       perl -e 'my %a=(k=>1, K=>1, M=>2, G=>3, T=>4, P=>5, E=>6, Z=>7, Y=>8);
       foreach my $i (0..255) {my $c=chr($i); $a{$c} ||= 0;print "$a{$c}, "}'\
       |fmt  */
    0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
    0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
    0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 6, 0, 3,
    0, 0, 0, 1, 0, 2, 0, 0, 5, 0, 0, 0, 4, 0, 0, 0, 0, 8, 7, 0, 0, 0, 0, 0,
    0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
    0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
    0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
    0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
    0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
    0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
    0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0
#endif
  };

/* Return an integer that represents the order of magnitude of the
   unit following the number.  The number may contain thousands
   separators and a decimal point, but it may not contain leading blanks.
   Negative numbers get negative orders; zero numbers have a zero order.  */

static int _GL_ATTRIBUTE_PURE
‌find_unit_order (char const *number)
{
  bool minus_sign = (*number == '-');
  char const *p = number + minus_sign;
  int nonzero = 0;
  unsigned char ch;

  /* Scan to end of number.
     Decimals or separators not followed by digits stop the scan.
     Numbers ending in decimals or separators are thus considered
     to be lacking in units.
     FIXME: add support for multibyte thousands_sep and decimal_point.  */

  do
    {
      while (ISDIGIT (ch = *p++))
        nonzero |= ch - '0';
    }
  while (ch == thousands_sep);

  if (ch == decimal_point)
    while (ISDIGIT (ch = *p++))
      nonzero |= ch - '0';

  if (nonzero)
    {
      int order = unit_order[ch];
      return (minus_sign ? -order : order);
    }
  else
    return 0;
}

/* Compare numbers A and B ending in units with SI or IEC prefixes
       <none/unknown> < K/k < M < G < T < P < E < Z < Y  */

static int
‌human_numcompare (char const *a, char const *b)
{
  while (blanks[to_uchar (*a)])
    a++;
  while (blanks[to_uchar (*b)])
    b++;

  int diff = find_unit_order (a) - find_unit_order (b);
  return (diff ? diff : strnumcmp (a, b, decimal_point, thousands_sep));
}

/* Compare strings A and B as numbers without explicitly converting them to
   machine numbers.  Comparatively slow for short strings, but asymptotically
   hideously fast. */

static int
‌numcompare (char const *a, char const *b)
{
  while (blanks[to_uchar (*a)])
    a++;
  while (blanks[to_uchar (*b)])
    b++;

  return strnumcmp (a, b, decimal_point, thousands_sep);
}

/* Work around a problem whereby the long double value returned by glibc's
   strtold ("NaN", ...) contains uninitialized bits: clear all bytes of
   A and B before calling strtold.  FIXME: remove this function once
   gnulib guarantees that strtold's result is always well defined.  */
static int
‌nan_compare (char const *sa, char const *sb)
{
  long_double a;
  memset (&a, 0, sizeof a);
  a = strtold (sa, NULL);

  long_double b;
  memset (&b, 0, sizeof b);
  b = strtold (sb, NULL);

  return memcmp (&a, &b, sizeof a);
}

static int
‌general_numcompare (char const *sa, char const *sb)
{
  /* FIXME: maybe add option to try expensive FP conversion
     only if A and B can't be compared more cheaply/accurately.  */

  char *ea;
  char *eb;
  long_double a = strtold (sa, &ea);
  long_double b = strtold (sb, &eb);

  /* Put conversion errors at the start of the collating sequence.  */
  if (sa == ea)
    return sb == eb ? 0 : -1;
  if (sb == eb)
    return 1;

  /* Sort numbers in the usual way, where -0 == +0.  Put NaNs after
     conversion errors but before numbers; sort them by internal
     bit-pattern, for lack of a more portable alternative.  */
  return (a < b ? -1
          : a > b ? 1
          : a == b ? 0
          : b == b ? -1
          : a == a ? 1
          : nan_compare (sa, sb));
}

/* Return an integer in 1..12 of the month name MONTH.
   Return 0 if the name in S is not recognized.  */

static int
‌getmonth (char const *month, char **ea)
{
  size_t lo = 0;
  size_t hi = MONTHS_PER_YEAR;

  while (blanks[to_uchar (*month)])
    month++;

  do
    {
      size_t ix = (lo + hi) / 2;
      char const *m = month;
      char const *n = monthtab[ix].name;

      for (;; m++, n++)
        {
          if (!*n)
            {
              if (ea)
                *ea = (char *) m;
              return monthtab[ix].val;
            }
          if (to_uchar (fold_toupper[to_uchar (*m)]) < to_uchar (*n))
            {
              hi = ix;
              break;
            }
          else if (to_uchar (fold_toupper[to_uchar (*m)]) > to_uchar (*n))
            {
              lo = ix + 1;
              break;
            }
        }
    }
  while (lo < hi);

  return 0;
}

/* A randomly chosen MD5 state, used for random comparison.  */
‌static struct md5_ctx random_md5_state;

/* Initialize the randomly chosen MD5 state.  */

static void
‌random_md5_state_init (char const *random_source)
{
  unsigned char buf[MD5_DIGEST_SIZE];
  struct randread_source *r = randread_new (random_source, sizeof buf);
  if (! r)
    die (_("open failed"), random_source);
  randread (r, buf, sizeof buf);
  if (randread_free (r) != 0)
    die (_("close failed"), random_source);
  md5_init_ctx (&random_md5_state);
  md5_process_bytes (buf, sizeof buf, &random_md5_state);
}

/* This is like strxfrm, except it reports any error and exits.  */

static size_t
‌xstrxfrm (char *restrict dest, char const *restrict src, size_t destsize)
{
  errno = 0;
  size_t translated_size = strxfrm (dest, src, destsize);

  if (errno)
    {
      error (0, errno, _("string transformation failed"));
      error (0, 0, _("set LC_ALL='C' to work around the problem"));
      error (SORT_FAILURE, 0,
             _("the untransformed string was %s"),
             quotearg_n_style (0, locale_quoting_style, src));
    }

  return translated_size;
}

/* Compare the keys TEXTA (of length LENA) and TEXTB (of length LENB)
   using one or more random hash functions.  TEXTA[LENA] and
   TEXTB[LENB] must be zero.  */

static int
‌compare_random (char *restrict texta, size_t lena,
                char *restrict textb, size_t lenb)
{
  /* XFRM_DIFF records the equivalent of memcmp on the transformed
     data.  This is used to break ties if there is a checksum
     collision, and this is good enough given the astronomically low
     probability of a collision.  */
  int xfrm_diff = 0;

  char stackbuf[4000];
  char *buf = stackbuf;
  size_t bufsize = sizeof stackbuf;
  void *allocated = NULL;
  uint32_t dig[2][MD5_DIGEST_SIZE / sizeof (uint32_t)];
  struct md5_ctx s[2];
  s[0] = s[1] = random_md5_state;

  if (hard_LC_COLLATE)
    {
      char const *lima = texta + lena;
      char const *limb = textb + lenb;

      while (true)
        {
          /* Transform the text into the basis of comparison, so that byte
             strings that would otherwise considered to be equal are
             considered equal here even if their bytes differ.

             Each time through this loop, transform one
             null-terminated string's worth from TEXTA or from TEXTB
             or both.  That way, there's no need to store the
             transformation of the whole line, if it contains many
             null-terminated strings.  */

          /* Store the transformed data into a big-enough buffer.  */

          /* A 3X size guess avoids the overhead of calling strxfrm
             twice on typical implementations.  Don't worry about
             size_t overflow, as the guess need not be correct.  */
          size_t guess_bufsize = 3 * (lena + lenb) + 2;
          if (bufsize < guess_bufsize)
            {
              bufsize = MAX (guess_bufsize, bufsize * 3 / 2);
              free (allocated);
              buf = allocated = malloc (bufsize);
              if (! buf)
                {
                  buf = stackbuf;
                  bufsize = sizeof stackbuf;
                }
            }

          size_t sizea =
            (texta < lima ? xstrxfrm (buf, texta, bufsize) + 1 : 0);
          bool a_fits = sizea <= bufsize;
          size_t sizeb =
            (textb < limb
             ? (xstrxfrm ((a_fits ? buf + sizea : NULL), textb,
                          (a_fits ? bufsize - sizea : 0))
                + 1)
             : 0);

          if (! (a_fits && sizea + sizeb <= bufsize))
            {
              bufsize = sizea + sizeb;
              if (bufsize < SIZE_MAX / 3)
                bufsize = bufsize * 3 / 2;
              free (allocated);
              buf = allocated = xmalloc (bufsize);
              if (texta < lima)
                strxfrm (buf, texta, sizea);
              if (textb < limb)
                strxfrm (buf + sizea, textb, sizeb);
            }

          /* Advance past NULs to the next part of each input string,
             exiting the loop if both strings are exhausted.  When
             exiting the loop, prepare to finish off the tiebreaker
             comparison properly.  */
          if (texta < lima)
            texta += strlen (texta) + 1;
          if (textb < limb)
            textb += strlen (textb) + 1;
          if (! (texta < lima || textb < limb))
            {
              lena = sizea; texta = buf;
              lenb = sizeb; textb = buf + sizea;
              break;
            }

          /* Accumulate the transformed data in the corresponding
             checksums.  */
          md5_process_bytes (buf, sizea, &s[0]);
          md5_process_bytes (buf + sizea, sizeb, &s[1]);

          /* Update the tiebreaker comparison of the transformed data.  */
          if (! xfrm_diff)
            {
              xfrm_diff = memcmp (buf, buf + sizea, MIN (sizea, sizeb));
              if (! xfrm_diff)
                xfrm_diff = (sizea > sizeb) - (sizea < sizeb);
            }
        }
    }

  /* Compute and compare the checksums.  */
  md5_process_bytes (texta, lena, &s[0]); md5_finish_ctx (&s[0], dig[0]);
  md5_process_bytes (textb, lenb, &s[1]); md5_finish_ctx (&s[1], dig[1]);
  int diff = memcmp (dig[0], dig[1], sizeof dig[0]);

  /* Fall back on the tiebreaker if the checksums collide.  */
  if (! diff)
    {
      if (! xfrm_diff)
        {
          xfrm_diff = memcmp (texta, textb, MIN (lena, lenb));
          if (! xfrm_diff)
            xfrm_diff = (lena > lenb) - (lena < lenb);
        }

      diff = xfrm_diff;
    }

  free (allocated);

  return diff;
}

/* Return the printable width of the block of memory starting at
   TEXT and ending just before LIM, counting each tab as one byte.
   FIXME: Should we generally be counting non printable chars?  */

static size_t
‌debug_width (char const *text, char const *lim)
{
  size_t width = mbsnwidth (text, lim - text, 0);
  while (text < lim)
    width += (*text++ == '\t');
  return width;
}

/* For debug mode, "underline" a key at the
   specified offset and screen width.  */

static void
‌mark_key (size_t offset, size_t width)
{
  while (offset--)
    putchar (' ');

  if (!width)
    printf (_("^ no match for key\n"));
  else
    {
      do
        putchar ('_');
      while (--width);

      putchar ('\n');
    }
}

/* Return true if KEY is a numeric key.  */

static inline bool
‌key_numeric (struct keyfield const *key)
{
  return key->numeric || key->general_numeric || key->human_numeric;
}

/* For LINE, output a debugging line that underlines KEY in LINE.
   If KEY is null, underline the whole line.  */

static void
‌debug_key (struct line const *line, struct keyfield const *key)
{
  char *text = line->text;
  char *beg = text;
  char *lim = text + line->length - 1;

  if (key)
    {
      if (key->sword != SIZE_MAX)
        beg = begfield (line, key);
      if (key->eword != SIZE_MAX)
        lim = limfield (line, key);

      if (key->skipsblanks || key->month || key_numeric (key))
        {
          char saved = *lim;
          *lim = '\0';

          while (blanks[to_uchar (*beg)])
            beg++;

          char *tighter_lim = beg;

          if (lim < beg)
            tighter_lim = lim;
          else if (key->month)
            getmonth (beg, &tighter_lim);
          else if (key->general_numeric)
            ignore_value (strtold (beg, &tighter_lim));
          else if (key->numeric || key->human_numeric)
            {
              char *p = beg + (beg < lim && *beg == '-');
              bool found_digit = false;
              unsigned char ch;

              do
                {
                  while (ISDIGIT (ch = *p++))
                    found_digit = true;
                }
              while (ch == thousands_sep);

              if (ch == decimal_point)
                while (ISDIGIT (ch = *p++))
                  found_digit = true;

              if (found_digit)
                tighter_lim = p - ! (key->human_numeric && unit_order[ch]);
            }
          else
            tighter_lim = lim;

          *lim = saved;
          lim = tighter_lim;
        }
    }

  size_t offset = debug_width (text, beg);
  size_t width = debug_width (beg, lim);
  mark_key (offset, width);
}

/* Debug LINE by underlining its keys.  */

static void
‌debug_line (struct line const *line)
{
  struct keyfield const *key = keylist;

  do
    debug_key (line, key);
  while (key && ((key = key->next) || ! (unique || stable)));
}

/* Return whether sorting options specified for key.  */

static bool
‌default_key_compare (struct keyfield const *key)
{
  return ! (key->ignore
            || key->translate
            || key->skipsblanks
            || key->skipeblanks
            || key_numeric (key)
            || key->month
            || key->version
            || key->random
            /* || key->reverse */
           );
}

/* Convert a key to the short options used to specify it.  */

static void
‌key_to_opts (struct keyfield const *key, char *opts)
{
  if (key->skipsblanks || key->skipeblanks)
    *opts++ = 'b';/* either disables global -b  */
  if (key->ignore == nondictionary)
    *opts++ = 'd';
  if (key->translate)
    *opts++ = 'f';
  if (key->general_numeric)
    *opts++ = 'g';
  if (key->human_numeric)
    *opts++ = 'h';
  if (key->ignore == nonprinting)
    *opts++ = 'i';
  if (key->month)
    *opts++ = 'M';
  if (key->numeric)
    *opts++ = 'n';
  if (key->random)
    *opts++ = 'R';
  if (key->reverse)
    *opts++ = 'r';
  if (key->version)
    *opts++ = 'V';
  *opts = '\0';
}

/* Output data independent key warnings to stderr.  */

static void
‌key_warnings (struct keyfield const *gkey, bool gkey_only)
{
  struct keyfield const *key;
  struct keyfield ugkey = *gkey;
  unsigned long keynum = 1;

  for (key = keylist; key; key = key->next, keynum++)
    {
      if (key->obsolete_used)
        {
          size_t sword = key->sword;
          size_t eword = key->eword;
          char tmp[INT_BUFSIZE_BOUND (uintmax_t)];
          /* obsolescent syntax +A.x -B.y is equivalent to:
               -k A+1.x+1,B.y   (when y = 0)
               -k A+1.x+1,B+1.y (when y > 0)  */
          char obuf[INT_BUFSIZE_BOUND (sword) * 2 + 4]; /* +# -#  */
          char nbuf[INT_BUFSIZE_BOUND (sword) * 2 + 5]; /* -k #,#  */
          char *po = obuf;
          char *pn = nbuf;

          if (sword == SIZE_MAX)
            sword++;

          po = stpcpy (stpcpy (po, "+"), umaxtostr (sword, tmp));
          pn = stpcpy (stpcpy (pn, "-k "), umaxtostr (sword + 1, tmp));
          if (key->eword != SIZE_MAX)
            {
              stpcpy (stpcpy (po, " -"), umaxtostr (eword + 1, tmp));
              stpcpy (stpcpy (pn, ","),
                      umaxtostr (eword + 1
                                 + (key->echar == SIZE_MAX), tmp));
            }
          error (0, 0, _("obsolescent key %s used; consider %s instead"),
                 quote_n (0, obuf), quote_n (1, nbuf));
        }

      /* Warn about field specs that will never match.  */
      if (key->sword != SIZE_MAX && key->eword < key->sword)
        error (0, 0, _("key %lu has zero width and will be ignored"), keynum);

      /* Warn about significant leading blanks.  */
      bool implicit_skip = key_numeric (key) || key->month;
      bool maybe_space_aligned = !hard_LC_COLLATE && default_key_compare (key)
                                 && !(key->schar || key->echar);
      bool line_offset = key->eword == 0 && key->echar != 0; /* -k1.x,1.y  */
      if (!gkey_only && tab == TAB_DEFAULT && !line_offset
          && ((!key->skipsblanks && !(implicit_skip || maybe_space_aligned))
              || (!key->skipsblanks && key->schar)
              || (!key->skipeblanks && key->echar)))
        error (0, 0, _("leading blanks are significant in key %lu; "
                       "consider also specifying 'b'"), keynum);

      /* Warn about numeric comparisons spanning fields,
         as field delimiters could be interpreted as part
         of the number (maybe only in other locales).  */
      if (!gkey_only && key_numeric (key))
        {
          size_t sword = key->sword + 1;
          size_t eword = key->eword + 1;
          if (!sword)
            sword++;
          if (!eword || sword < eword)
            error (0, 0, _("key %lu is numeric and spans multiple fields"),
                   keynum);
        }

      /* Flag global options not copied or specified in any key.  */
      if (ugkey.ignore && (ugkey.ignore == key->ignore))
        ugkey.ignore = NULL;
      if (ugkey.translate && (ugkey.translate == key->translate))
        ugkey.translate = NULL;
      ugkey.skipsblanks &= !key->skipsblanks;
      ugkey.skipeblanks &= !key->skipeblanks;
      ugkey.month &= !key->month;
      ugkey.numeric &= !key->numeric;
      ugkey.general_numeric &= !key->general_numeric;
      ugkey.human_numeric &= !key->human_numeric;
      ugkey.random &= !key->random;
      ugkey.version &= !key->version;
      ugkey.reverse &= !key->reverse;
    }

  /* Warn about ignored global options flagged above.
     Note if gkey is the only one in the list, all flags are cleared.  */
  if (!default_key_compare (&ugkey)
      || (ugkey.reverse && (stable || unique) && keylist))
    {
      bool ugkey_reverse = ugkey.reverse;
      if (!(stable || unique))
        ugkey.reverse = false;
      /* The following is too big, but guaranteed to be "big enough".  */
      char opts[sizeof short_options];
      key_to_opts (&ugkey, opts);
      error (0, 0,
             ngettext ("option '-%s' is ignored",
                       "options '-%s' are ignored",
                       select_plural (strlen (opts))), opts);
      ugkey.reverse = ugkey_reverse;
    }
  if (ugkey.reverse && !(stable || unique) && keylist)
    error (0, 0, _("option '-r' only applies to last-resort comparison"));
}

/* Compare two lines A and B trying every key in sequence until there
   are no more keys or a difference is found. */

static int
‌keycompare (struct line const *a, struct line const *b)
{
  struct keyfield *key = keylist;

  /* For the first iteration only, the key positions have been
     precomputed for us. */
  char *texta = a->keybeg;
  char *textb = b->keybeg;
  char *lima = a->keylim;
  char *limb = b->keylim;

  int diff;

  while (true)
    {
      char const *translate = key->translate;
      bool const *ignore = key->ignore;

      /* Treat field ends before field starts as empty fields.  */
      lima = MAX (texta, lima);
      limb = MAX (textb, limb);

      /* Find the lengths. */
      size_t lena = lima - texta;
      size_t lenb = limb - textb;

      if (hard_LC_COLLATE || key_numeric (key)
          || key->month || key->random || key->version)
        {
          char *ta;
          char *tb;
          size_t tlena;
          size_t tlenb;

          char enda IF_LINT (= 0);
          char endb IF_LINT (= 0);
          void *allocated IF_LINT (= NULL);
          char stackbuf[4000];

          if (ignore || translate)
            {
              /* Compute with copies of the keys, which are the result of
                 translating or ignoring characters, and which need their
                 own storage.  */

              size_t i;

              /* Allocate space for copies.  */
              size_t size = lena + 1 + lenb + 1;
              if (size <= sizeof stackbuf)
                ta = stackbuf, allocated = NULL;
              else
                ta = allocated = xmalloc (size);
              tb = ta + lena + 1;

              /* Put into each copy a version of the key in which the
                 requested characters are ignored or translated.  */
              for (tlena = i = 0; i < lena; i++)
                if (! (ignore && ignore[to_uchar (texta[i])]))
                  ta[tlena++] = (translate
                                 ? translate[to_uchar (texta[i])]
                                 : texta[i]);
              ta[tlena] = '\0';

              for (tlenb = i = 0; i < lenb; i++)
                if (! (ignore && ignore[to_uchar (textb[i])]))
                  tb[tlenb++] = (translate
                                 ? translate[to_uchar (textb[i])]
                                 : textb[i]);
              tb[tlenb] = '\0';
            }
          else
            {
              /* Use the keys in-place, temporarily null-terminated.  */
              ta = texta; tlena = lena; enda = ta[tlena]; ta[tlena] = '\0';
              tb = textb; tlenb = lenb; endb = tb[tlenb]; tb[tlenb] = '\0';
            }

          if (key->numeric)
            diff = numcompare (ta, tb);
          else if (key->general_numeric)
            diff = general_numcompare (ta, tb);
          else if (key->human_numeric)
            diff = human_numcompare (ta, tb);
          else if (key->month)
            diff = getmonth (ta, NULL) - getmonth (tb, NULL);
          else if (key->random)
            diff = compare_random (ta, tlena, tb, tlenb);
          else if (key->version)
            diff = filevercmp (ta, tb);
          else
            {
              /* Locale-dependent string sorting.  This is slower than
                 C-locale sorting, which is implemented below.  */
              if (tlena == 0)
                diff = - NONZERO (tlenb);
              else if (tlenb == 0)
                diff = 1;
              else
                diff = xmemcoll0 (ta, tlena + 1, tb, tlenb + 1);
            }

          if (ignore || translate)
            free (allocated);
          else
            {
              ta[tlena] = enda;
              tb[tlenb] = endb;
            }
        }
      else if (ignore)
        {
‌#define CMP_WITH_IGNORE(A, B)                        \
  do                                    \
    {                                    \
          while (true)                            \
            {                                \
              while (texta < lima && ignore[to_uchar (*texta)])        \
                ++texta;                        \
              while (textb < limb && ignore[to_uchar (*textb)])        \
                ++textb;                        \
              if (! (texta < lima && textb < limb))            \
                break;                            \
              diff = to_uchar (A) - to_uchar (B);            \
              if (diff)                            \
                goto not_equal;                        \
              ++texta;                            \
              ++textb;                            \
            }                                \
                                                                        \
          diff = (texta < lima) - (textb < limb);            \
    }                                    \
  while (0)

          if (translate)
            CMP_WITH_IGNORE (translate[to_uchar (*texta)],
                             translate[to_uchar (*textb)]);
          else
            CMP_WITH_IGNORE (*texta, *textb);
        }
      else if (lena == 0)
        diff = - NONZERO (lenb);
      else if (lenb == 0)
        goto greater;
      else
        {
          if (translate)
            {
              while (texta < lima && textb < limb)
                {
                  diff = (to_uchar (translate[to_uchar (*texta++)])
                          - to_uchar (translate[to_uchar (*textb++)]));
                  if (diff)
                    goto not_equal;
                }
            }
          else
            {
              diff = memcmp (texta, textb, MIN (lena, lenb));
              if (diff)
                goto not_equal;
            }
          diff = lena < lenb ? -1 : lena != lenb;
        }

      if (diff)
        goto not_equal;

      key = key->next;
      if (! key)
        break;

      /* Find the beginning and limit of the next field.  */
      if (key->eword != SIZE_MAX)
        lima = limfield (a, key), limb = limfield (b, key);
      else
        lima = a->text + a->length - 1, limb = b->text + b->length - 1;

      if (key->sword != SIZE_MAX)
        texta = begfield (a, key), textb = begfield (b, key);
      else
        {
          texta = a->text, textb = b->text;
          if (key->skipsblanks)
            {
              while (texta < lima && blanks[to_uchar (*texta)])
                ++texta;
              while (textb < limb && blanks[to_uchar (*textb)])
                ++textb;
            }
        }
    }

  return 0;

greater:
  diff = 1;
not_equal:
  return key->reverse ? -diff : diff;
}

/* Compare two lines A and B, returning negative, zero, or positive
   depending on whether A compares less than, equal to, or greater than B. */

static int
‌compare (struct line const *a, struct line const *b)
{
  int diff;
  size_t alen, blen;

  /* First try to compare on the specified keys (if any).
     The only two cases with no key at all are unadorned sort,
     and unadorned sort -r. */
  if (keylist)
    {
      diff = keycompare (a, b);
      if (diff || unique || stable)
        return diff;
    }

  /* If the keys all compare equal (or no keys were specified)
     fall through to the default comparison.  */
  alen = a->length - 1, blen = b->length - 1;

  if (alen == 0)
    diff = - NONZERO (blen);
  else if (blen == 0)
    diff = 1;
  else if (hard_LC_COLLATE)
    {
      /* Note xmemcoll0 is a performance enhancement as
         it will not unconditionally write '\0' after the
         passed in buffers, which was seen to give around
         a 3% increase in performance for short lines.  */
      diff = xmemcoll0 (a->text, alen + 1, b->text, blen + 1);
    }
  else if (! (diff = memcmp (a->text, b->text, MIN (alen, blen))))
    diff = alen < blen ? -1 : alen != blen;

  return reverse ? -diff : diff;
}

/* Write LINE to output stream FP; the output file's name is
   OUTPUT_FILE if OUTPUT_FILE is nonnull, and is the standard output
   otherwise.  If debugging is enabled and FP is standard output,
   append some debugging information.  */

static void
‌write_line (struct line const *line, FILE *fp, char const *output_file)
{
  char *buf = line->text;
  size_t n_bytes = line->length;
  char *ebuf = buf + n_bytes;

  if (!output_file && debug)
    {
      /* Convert TAB to '>' and EOL to \n, and then output debugging info.  */
      char const *c = buf;

      while (c < ebuf)
        {
          char wc = *c++;
          if (wc == '\t')
            wc = '>';
          else if (c == ebuf)
            wc = '\n';
          if (fputc (wc, fp) == EOF)
            die (_("write failed"), output_file);
        }

      debug_line (line);
    }
  else
    {
      ebuf[-1] = eolchar;
      if (fwrite (buf, 1, n_bytes, fp) != n_bytes)
        die (_("write failed"), output_file);
      ebuf[-1] = '\0';
    }
}

/* Check that the lines read from FILE_NAME come in order.  Return
   true if they are in order.  If CHECKONLY == 'c', also print a
   diagnostic (FILE_NAME, line number, contents of line) to stderr if
   they are not in order.  */

static bool
‌check (char const *file_name, char checkonly)
{
  FILE *fp = xfopen (file_name, "r");
  struct buffer buf;        /* Input buffer. */
  struct line temp;        /* Copy of previous line. */
  size_t alloc = 0;
  uintmax_t line_number = 0;
  struct keyfield const *key = keylist;
  bool nonunique = ! unique;
  bool ordered = true;

  initbuf (&buf, sizeof (struct line),
           MAX (merge_buffer_size, sort_size));
  temp.text = NULL;

  while (fillbuf (&buf, fp, file_name))
    {
      struct line const *line = buffer_linelim (&buf);
      struct line const *linebase = line - buf.nlines;

      /* Make sure the line saved from the old buffer contents is
         less than or equal to the first line of the new buffer. */
      if (alloc && nonunique <= compare (&temp, line - 1))
        {
        found_disorder:
          {
            if (checkonly == 'c')
              {
                struct line const *disorder_line = line - 1;
                uintmax_t disorder_line_number =
                  buffer_linelim (&buf) - disorder_line + line_number;
                char hr_buf[INT_BUFSIZE_BOUND (disorder_line_number)];
                fprintf (stderr, _("%s: %s:%s: disorder: "),
                         program_name, file_name,
                         umaxtostr (disorder_line_number, hr_buf));
                write_line (disorder_line, stderr, _("standard error"));
              }

            ordered = false;
            break;
          }
        }

      /* Compare each line in the buffer with its successor.  */
      while (linebase < --line)
        if (nonunique <= compare (line, line - 1))
          goto found_disorder;

      line_number += buf.nlines;

      /* Save the last line of the buffer.  */
      if (alloc < line->length)
        {
          do
            {
              alloc *= 2;
              if (! alloc)
                {
                  alloc = line->length;
                  break;
                }
            }
          while (alloc < line->length);

          free (temp.text);
          temp.text = xmalloc (alloc);
        }
      memcpy (temp.text, line->text, line->length);
      temp.length = line->length;
      if (key)
        {
          temp.keybeg = temp.text + (line->keybeg - line->text);
          temp.keylim = temp.text + (line->keylim - line->text);
        }
    }

  xfclose (fp, file_name);
  free (buf.buf);
  free (temp.text);
  return ordered;
}

/* Open FILES (there are NFILES of them) and store the resulting array
   of stream pointers into (*PFPS).  Allocate the array.  Return the
   number of successfully opened files, setting errno if this value is
   less than NFILES.  */

static size_t
‌open_input_files (struct sortfile *files, size_t nfiles, FILE ***pfps)
{
  FILE **fps = *pfps = xnmalloc (nfiles, sizeof *fps);
  int i;

  /* Open as many input files as we can.  */
  for (i = 0; i < nfiles; i++)
    {
      fps[i] = (files[i].temp && files[i].temp->state != UNCOMPRESSED
                ? open_temp (files[i].temp)
                : stream_open (files[i].name, "r"));
      if (!fps[i])
        break;
    }

  return i;
}

/* Merge lines from FILES onto OFP.  NTEMPS is the number of temporary
   files (all of which are at the start of the FILES array), and
   NFILES is the number of files; 0 <= NTEMPS <= NFILES <= NMERGE.
   FPS is the vector of open stream corresponding to the files.
   Close input and output streams before returning.
   OUTPUT_FILE gives the name of the output file.  If it is NULL,
   the output file is standard output.  */

static void
‌mergefps (struct sortfile *files, size_t ntemps, size_t nfiles,
          FILE *ofp, char const *output_file, FILE **fps)
{
  struct buffer *buffer = xnmalloc (nfiles, sizeof *buffer);
                                /* Input buffers for each file. */
  struct line saved;        /* Saved line storage for unique check. */
  struct line const *savedline = NULL;
                                /* &saved if there is a saved line. */
  size_t savealloc = 0;        /* Size allocated for the saved line. */
  struct line const **cur = xnmalloc (nfiles, sizeof *cur);
                                /* Current line in each line table. */
  struct line const **base = xnmalloc (nfiles, sizeof *base);
                                /* Base of each line table.  */
  size_t *ord = xnmalloc (nfiles, sizeof *ord);
                                /* Table representing a permutation of fps,
                                   such that cur[ord[0]] is the smallest line
                                   and will be next output. */
  size_t i;
  size_t j;
  size_t t;
  struct keyfield const *key = keylist;
  saved.text = NULL;

  /* Read initial lines from each input file. */
  for (i = 0; i < nfiles; )
    {
      initbuf (&buffer[i], sizeof (struct line),
               MAX (merge_buffer_size, sort_size / nfiles));
      if (fillbuf (&buffer[i], fps[i], files[i].name))
        {
          struct line const *linelim = buffer_linelim (&buffer[i]);
          cur[i] = linelim - 1;
          base[i] = linelim - buffer[i].nlines;
          i++;
        }
      else
        {
          /* fps[i] is empty; eliminate it from future consideration.  */
          xfclose (fps[i], files[i].name);
          if (i < ntemps)
            {
              ntemps--;
              zaptemp (files[i].name);
            }
          free (buffer[i].buf);
          --nfiles;
          for (j = i; j < nfiles; ++j)
            {
              files[j] = files[j + 1];
              fps[j] = fps[j + 1];
            }
        }
    }

  /* Set up the ord table according to comparisons among input lines.
     Since this only reorders two items if one is strictly greater than
     the other, it is stable. */
  for (i = 0; i < nfiles; ++i)
    ord[i] = i;
  for (i = 1; i < nfiles; ++i)
    if (0 < compare (cur[ord[i - 1]], cur[ord[i]]))
      t = ord[i - 1], ord[i - 1] = ord[i], ord[i] = t, i = 0;

  /* Repeatedly output the smallest line until no input remains. */
  while (nfiles)
    {
      struct line const *smallest = cur[ord[0]];

      /* If uniquified output is turned on, output only the first of
         an identical series of lines. */
      if (unique)
        {
          if (savedline && compare (savedline, smallest))
            {
              savedline = NULL;
              write_line (&saved, ofp, output_file);
            }
          if (!savedline)
            {
              savedline = &saved;
              if (savealloc < smallest->length)
                {
                  do
                    if (! savealloc)
                      {
                        savealloc = smallest->length;
                        break;
                      }
                  while ((savealloc *= 2) < smallest->length);

                  free (saved.text);
                  saved.text = xmalloc (savealloc);
                }
              saved.length = smallest->length;
              memcpy (saved.text, smallest->text, saved.length);
              if (key)
                {
                  saved.keybeg =
                    saved.text + (smallest->keybeg - smallest->text);
                  saved.keylim =
                    saved.text + (smallest->keylim - smallest->text);
                }
            }
        }
      else
        write_line (smallest, ofp, output_file);

      /* Check if we need to read more lines into core. */
      if (base[ord[0]] < smallest)
        cur[ord[0]] = smallest - 1;
      else
        {
          if (fillbuf (&buffer[ord[0]], fps[ord[0]], files[ord[0]].name))
            {
              struct line const *linelim = buffer_linelim (&buffer[ord[0]]);
              cur[ord[0]] = linelim - 1;
              base[ord[0]] = linelim - buffer[ord[0]].nlines;
            }
          else
            {
              /* We reached EOF on fps[ord[0]].  */
              for (i = 1; i < nfiles; ++i)
                if (ord[i] > ord[0])
                  --ord[i];
              --nfiles;
              xfclose (fps[ord[0]], files[ord[0]].name);
              if (ord[0] < ntemps)
                {
                  ntemps--;
                  zaptemp (files[ord[0]].name);
                }
              free (buffer[ord[0]].buf);
              for (i = ord[0]; i < nfiles; ++i)
                {
                  fps[i] = fps[i + 1];
                  files[i] = files[i + 1];
                  buffer[i] = buffer[i + 1];
                  cur[i] = cur[i + 1];
                  base[i] = base[i + 1];
                }
              for (i = 0; i < nfiles; ++i)
                ord[i] = ord[i + 1];
              continue;
            }
        }

      /* The new line just read in may be larger than other lines
         already in main memory; push it back in the queue until we
         encounter a line larger than it.  Optimize for the common
         case where the new line is smallest.  */
      {
        size_t lo = 1;
        size_t hi = nfiles;
        size_t probe = lo;
        size_t ord0 = ord[0];
        size_t count_of_smaller_lines;

        while (lo < hi)
          {
            int cmp = compare (cur[ord0], cur[ord[probe]]);
            if (cmp < 0 || (cmp == 0 && ord0 < ord[probe]))
              hi = probe;
            else
              lo = probe + 1;
            probe = (lo + hi) / 2;
          }

        count_of_smaller_lines = lo - 1;
        for (j = 0; j < count_of_smaller_lines; j++)
          ord[j] = ord[j + 1];
        ord[count_of_smaller_lines] = ord0;
      }
    }

  if (unique && savedline)
    {
      write_line (&saved, ofp, output_file);
      free (saved.text);
    }

  xfclose (ofp, output_file);
  free (fps);
  free (buffer);
  free (ord);
  free (base);
  free (cur);
}

/* Merge lines from FILES onto OFP.  NTEMPS is the number of temporary
   files (all of which are at the start of the FILES array), and
   NFILES is the number of files; 0 <= NTEMPS <= NFILES <= NMERGE.
   Close input and output files before returning.
   OUTPUT_FILE gives the name of the output file.

   Return the number of files successfully merged.  This number can be
   less than NFILES if we ran low on file descriptors, but in this
   case it is never less than 2.  */

static size_t
‌mergefiles (struct sortfile *files, size_t ntemps, size_t nfiles,
            FILE *ofp, char const *output_file)
{
  FILE **fps;
  size_t nopened = open_input_files (files, nfiles, &fps);
  if (nopened < nfiles && nopened < 2)
    die (_("open failed"), files[nopened].name);
  mergefps (files, ntemps, nopened, ofp, output_file, fps);
  return nopened;
}

/* Merge into T (of size NLINES) the two sorted arrays of lines
   LO (with NLINES / 2 members), and
   T - (NLINES / 2) (with NLINES - NLINES / 2 members).
   T and LO point just past their respective arrays, and the arrays
   are in reverse order.  NLINES must be at least 2.  */

static void
‌mergelines (struct line *restrict t, size_t nlines,
            struct line const *restrict lo)
{
  size_t nlo = nlines / 2;
  size_t nhi = nlines - nlo;
  struct line *hi = t - nlo;

  while (true)
    if (compare (lo - 1, hi - 1) <= 0)
      {
        *--t = *--lo;
        if (! --nlo)
          {
            /* HI must equal T now, and there is no need to copy from
               HI to T. */
            return;
          }
      }
    else
      {
        *--t = *--hi;
        if (! --nhi)
          {
            do
              *--t = *--lo;
            while (--nlo);

            return;
          }
      }
}

/* Sort the array LINES with NLINES members, using TEMP for temporary space.
   Do this all within one thread.  NLINES must be at least 2.
   If TO_TEMP, put the sorted output into TEMP, and TEMP is as large as LINES.
   Otherwise the sort is in-place and TEMP is half-sized.
   The input and output arrays are in reverse order, and LINES and
   TEMP point just past the end of their respective arrays.

   Use a recursive divide-and-conquer algorithm, in the style
   suggested by Knuth volume 3 (2nd edition), exercise 5.2.4-23.  Use
   the optimization suggested by exercise 5.2.4-10; this requires room
   for only 1.5*N lines, rather than the usual 2*N lines.  Knuth
   writes that this memory optimization was originally published by
   D. A. Bell, Comp J. 1 (1958), 75.  */

static void
‌sequential_sort (struct line *restrict lines, size_t nlines,
                 struct line *restrict temp, bool to_temp)
{
  if (nlines == 2)
    {
      /* Declare 'swap' as int, not bool, to work around a bug
         <http://lists.gnu.org/archive/html/bug-coreutils/2005-10/msg00086.html>
         in the IBM xlc 6.0.0.0 compiler in 64-bit mode.  */
      int swap = (0 < compare (&lines[-1], &lines[-2]));
      if (to_temp)
        {
          temp[-1] = lines[-1 - swap];
          temp[-2] = lines[-2 + swap];
        }
      else if (swap)
        {
          temp[-1] = lines[-1];
          lines[-1] = lines[-2];
          lines[-2] = temp[-1];
        }
    }
  else
    {
      size_t nlo = nlines / 2;
      size_t nhi = nlines - nlo;
      struct line *lo = lines;
      struct line *hi = lines - nlo;

      sequential_sort (hi, nhi, temp - (to_temp ? nlo : 0), to_temp);
      if (1 < nlo)
        sequential_sort (lo, nlo, temp, !to_temp);
      else if (!to_temp)
        temp[-1] = lo[-1];

      struct line *dest;
      struct line const *sorted_lo;
      if (to_temp)
        {
          dest = temp;
          sorted_lo = lines;
        }
      else
        {
          dest = lines;
          sorted_lo = temp;
        }
      mergelines (dest, nlines, sorted_lo);
    }
}

static struct merge_node *init_node (struct merge_node *restrict,
                                     struct merge_node *restrict,
                                     struct line *, size_t, size_t, bool);


/* Create and return a merge tree for NTHREADS threads, sorting NLINES
   lines, with destination DEST.  */
static struct merge_node *
‌merge_tree_init (size_t nthreads, size_t nlines, struct line *dest)
{
  struct merge_node *merge_tree = xmalloc (2 * sizeof *merge_tree * nthreads);

  struct merge_node *root = merge_tree;
  root->lo = root->hi = root->end_lo = root->end_hi = NULL;
  root->dest = NULL;
  root->nlo = root->nhi = nlines;
  root->parent = NULL;
  root->level = MERGE_END;
  root->queued = false;
  pthread_mutex_init (&root->lock, NULL);

  init_node (root, root + 1, dest, nthreads, nlines, false);
  return merge_tree;
}

/* Destroy the merge tree. */
static void
‌merge_tree_destroy (size_t nthreads, struct merge_node *merge_tree)
{
  size_t n_nodes = nthreads * 2;
  struct merge_node *node = merge_tree;

  while (n_nodes--)
    {
      pthread_mutex_destroy (&node->lock);
      node++;
    }

  free (merge_tree);
}

/* Initialize a merge tree node and its descendants.  The node's
   parent is PARENT.  The node and its descendants are taken from the
   array of nodes NODE_POOL.  Their destination starts at DEST; they
   will consume NTHREADS threads.  The total number of sort lines is
   TOTAL_LINES.  IS_LO_CHILD is true if the node is the low child of
   its parent.  */

static struct merge_node *
‌init_node (struct merge_node *restrict parent,
           struct merge_node *restrict node_pool,
           struct line *dest, size_t nthreads,
           size_t total_lines, bool is_lo_child)
{
  size_t nlines = (is_lo_child ? parent->nlo : parent->nhi);
  size_t nlo = nlines / 2;
  size_t nhi = nlines - nlo;
  struct line *lo = dest - total_lines;
  struct line *hi = lo - nlo;
  struct line **parent_end = (is_lo_child ? &parent->end_lo : &parent->end_hi);

  struct merge_node *node = node_pool++;
  node->lo = node->end_lo = lo;
  node->hi = node->end_hi = hi;
  node->dest = parent_end;
  node->nlo = nlo;
  node->nhi = nhi;
  node->parent = parent;
  node->level = parent->level + 1;
  node->queued = false;
  pthread_mutex_init (&node->lock, NULL);

  if (nthreads > 1)
    {
      size_t lo_threads = nthreads / 2;
      size_t hi_threads = nthreads - lo_threads;
      node->lo_child = node_pool;
      node_pool = init_node (node, node_pool, lo, lo_threads,
                             total_lines, true);
      node->hi_child = node_pool;
      node_pool = init_node (node, node_pool, hi, hi_threads,
                             total_lines, false);
    }
  else
    {
      node->lo_child = NULL;
      node->hi_child = NULL;
    }
  return node_pool;
}


/* Compare two merge nodes A and B for priority.  */

static int
‌compare_nodes (void const *a, void const *b)
{
  struct merge_node const *nodea = a;
  struct merge_node const *nodeb = b;
  if (nodea->level == nodeb->level)
      return (nodea->nlo + nodea->nhi) < (nodeb->nlo + nodeb->nhi);
  return nodea->level < nodeb->level;
}

/* Lock a merge tree NODE.  */

static inline void
‌lock_node (struct merge_node *node)
{
  pthread_mutex_lock (&node->lock);
}

/* Unlock a merge tree NODE. */

static inline void
‌unlock_node (struct merge_node *node)
{
  pthread_mutex_unlock (&node->lock);
}

/* Destroy merge QUEUE. */

static void
‌queue_destroy (struct merge_node_queue *queue)
{
  heap_free (queue->priority_queue);
  pthread_cond_destroy (&queue->cond);
  pthread_mutex_destroy (&queue->mutex);
}

/* Initialize merge QUEUE, allocating space suitable for a maximum of
   NTHREADS threads.  */

static void
‌queue_init (struct merge_node_queue *queue, size_t nthreads)
{
  /* Though it's highly unlikely all nodes are in the heap at the same
     time, the heap should accommodate all of them.  Counting a NULL
     dummy head for the heap, reserve 2 * NTHREADS nodes.  */
  queue->priority_queue = heap_alloc (compare_nodes, 2 * nthreads);
  pthread_mutex_init (&queue->mutex, NULL);
  pthread_cond_init (&queue->cond, NULL);
}

/* Insert NODE into QUEUE.  The caller either holds a lock on NODE, or
   does not need to lock NODE.  */

static void
‌queue_insert (struct merge_node_queue *queue, struct merge_node *node)
{
  pthread_mutex_lock (&queue->mutex);
  heap_insert (queue->priority_queue, node);
  node->queued = true;
  pthread_cond_signal (&queue->cond);
  pthread_mutex_unlock (&queue->mutex);
}

/* Pop the top node off the priority QUEUE, lock the node, return it.  */

static struct merge_node *
‌queue_pop (struct merge_node_queue *queue)
{
  struct merge_node *node;
  pthread_mutex_lock (&queue->mutex);
  while (! (node = heap_remove_top (queue->priority_queue)))
    pthread_cond_wait (&queue->cond, &queue->mutex);
  pthread_mutex_unlock (&queue->mutex);
  lock_node (node);
  node->queued = false;
  return node;
}

/* Output LINE to TFP, unless -u is specified and the line compares
   equal to the previous line.  TEMP_OUTPUT is the name of TFP, or
   is null if TFP is standard output.

   This function does not save the line for comparison later, so it is
   appropriate only for internal sort.  */

static void
‌write_unique (struct line const *line, FILE *tfp, char const *temp_output)
{
  if (unique)
    {
      if (saved_line.text && ! compare (line, &saved_line))
        return;
      saved_line = *line;
    }

  write_line (line, tfp, temp_output);
}

/* Merge the lines currently available to a NODE in the binary
   merge tree.  Merge a number of lines appropriate for this merge
   level, assuming TOTAL_LINES is the total number of lines.

   If merging at the top level, send output to TFP.  TEMP_OUTPUT is
   the name of TFP, or is null if TFP is standard output.  */

static void
‌mergelines_node (struct merge_node *restrict node, size_t total_lines,
                 FILE *tfp, char const *temp_output)
{
  struct line *lo_orig = node->lo;
  struct line *hi_orig = node->hi;
  size_t to_merge = MAX_MERGE (total_lines, node->level);
  size_t merged_lo;
  size_t merged_hi;

  if (node->level > MERGE_ROOT)
    {
      /* Merge to destination buffer. */
      struct line *dest = *node->dest;
      while (node->lo != node->end_lo && node->hi != node->end_hi && to_merge--)
        if (compare (node->lo - 1, node->hi - 1) <= 0)
          *--dest = *--node->lo;
        else
          *--dest = *--node->hi;

      merged_lo = lo_orig - node->lo;
      merged_hi = hi_orig - node->hi;

      if (node->nhi == merged_hi)
        while (node->lo != node->end_lo && to_merge--)
          *--dest = *--node->lo;
      else if (node->nlo == merged_lo)
        while (node->hi != node->end_hi && to_merge--)
          *--dest = *--node->hi;
      *node->dest = dest;
    }
  else
    {
      /* Merge directly to output. */
      while (node->lo != node->end_lo && node->hi != node->end_hi && to_merge--)
        {
          if (compare (node->lo - 1, node->hi - 1) <= 0)
            write_unique (--node->lo, tfp, temp_output);
          else
            write_unique (--node->hi, tfp, temp_output);
        }

      merged_lo = lo_orig - node->lo;
      merged_hi = hi_orig - node->hi;

      if (node->nhi == merged_hi)
        {
          while (node->lo != node->end_lo && to_merge--)
            write_unique (--node->lo, tfp, temp_output);
        }
      else if (node->nlo == merged_lo)
        {
          while (node->hi != node->end_hi && to_merge--)
            write_unique (--node->hi, tfp, temp_output);
        }
    }

  /* Update NODE. */
  merged_lo = lo_orig - node->lo;
  merged_hi = hi_orig - node->hi;
  node->nlo -= merged_lo;
  node->nhi -= merged_hi;
}

/* Into QUEUE, insert NODE if it is not already queued, and if one of
   NODE's children has available lines and the other either has
   available lines or has exhausted its lines.  */

static void
‌queue_check_insert (struct merge_node_queue *queue, struct merge_node *node)
{
  if (! node->queued)
    {
      bool lo_avail = (node->lo - node->end_lo) != 0;
      bool hi_avail = (node->hi - node->end_hi) != 0;
      if (lo_avail ? hi_avail || ! node->nhi : hi_avail && ! node->nlo)
        queue_insert (queue, node);
    }
}

/* Into QUEUE, insert NODE's parent if the parent can now be worked on.  */

static void
‌queue_check_insert_parent (struct merge_node_queue *queue,
                           struct merge_node *node)
{
  if (node->level > MERGE_ROOT)
    {
      lock_node (node->parent);
      queue_check_insert (queue, node->parent);
      unlock_node (node->parent);
    }
  else if (node->nlo + node->nhi == 0)
    {
      /* If the MERGE_ROOT NODE has finished merging, insert the
         MERGE_END node.  */
      queue_insert (queue, node->parent);
    }
}

/* Repeatedly pop QUEUE for a node with lines to merge, and merge at least
   some of those lines, until the MERGE_END node is popped.
   TOTAL_LINES is the total number of lines.  If merging at the top
   level, send output to TFP.  TEMP_OUTPUT is the name of TFP, or is
   null if TFP is standard output.  */

static void
‌merge_loop (struct merge_node_queue *queue,
            size_t total_lines, FILE *tfp, char const *temp_output)
{
  while (1)
    {
      struct merge_node *node = queue_pop (queue);

      if (node->level == MERGE_END)
        {
          unlock_node (node);
          /* Reinsert so other threads can pop it. */
          queue_insert (queue, node);
          break;
        }
      mergelines_node (node, total_lines, tfp, temp_output);
      queue_check_insert (queue, node);
      queue_check_insert_parent (queue, node);

      unlock_node (node);
    }
}


static void sortlines (struct line *restrict, size_t, size_t,
                       struct merge_node *, struct merge_node_queue *,
                       FILE *, char const *);

/* Thread arguments for sortlines_thread. */

‌struct thread_args
{
  /* Source, i.e., the array of lines to sort.  This points just past
     the end of the array.  */
  struct line *lines;

  /* Number of threads to use.  If 0 or 1, sort single-threaded.  */
  size_t nthreads;

  /* Number of lines in LINES and DEST.  */
  size_t const total_lines;

  /* Merge node. Lines from this node and this node's sibling will merged
     to this node's parent. */
  struct merge_node *const node;

  /* The priority queue controlling available work for the entire
     internal sort.  */
  struct merge_node_queue *const queue;

  /* If at the top level, the file to output to, and the file's name.
     If the file is standard output, the file's name is null.  */
  FILE *tfp;
  char const *output_temp;
};

/* Like sortlines, except with a signature acceptable to pthread_create.  */

static void *
‌sortlines_thread (void *data)
{
  struct thread_args const *args = data;
  sortlines (args->lines, args->nthreads, args->total_lines,
             args->node, args->queue, args->tfp,
             args->output_temp);
  return NULL;
}

/* Sort lines, possibly in parallel.  The arguments are as in struct
   thread_args above.

   The algorithm has three phases: node creation, sequential sort,
   and binary merge.

   During node creation, sortlines recursively visits each node in the
   binary merge tree and creates a NODE structure corresponding to all the
   future line merging NODE is responsible for. For each call to
   sortlines, half the available threads are assigned to each recursive
   call, until a leaf node having only 1 available thread is reached.

   Each leaf node then performs two sequential sorts, one on each half of
   the lines it is responsible for. It records in its NODE structure that
   there are two sorted sublists available to merge from, and inserts its
   NODE into the priority queue.

   The binary merge phase then begins. Each thread drops into a loop
   where the thread retrieves a NODE from the priority queue, merges lines
   available to that NODE, and potentially insert NODE or its parent back
   into the queue if there are sufficient available lines for them to
   merge. This continues until all lines at all nodes of the merge tree
   have been merged. */

static void
‌sortlines (struct line *restrict lines, size_t nthreads,
           size_t total_lines, struct merge_node *node,
           struct merge_node_queue *queue, FILE *tfp, char const *temp_output)
{
  size_t nlines = node->nlo + node->nhi;

  /* Calculate thread arguments. */
  size_t lo_threads = nthreads / 2;
  size_t hi_threads = nthreads - lo_threads;
  pthread_t thread;
  struct thread_args args = {lines, lo_threads, total_lines,
                             node->lo_child, queue, tfp, temp_output};

  if (nthreads > 1 && SUBTHREAD_LINES_HEURISTIC <= nlines
      && pthread_create (&thread, NULL, sortlines_thread, &args) == 0)
    {
      sortlines (lines - node->nlo, hi_threads, total_lines,
                 node->hi_child, queue, tfp, temp_output);
      pthread_join (thread, NULL);
    }
  else
    {
      /* Nthreads = 1, this is a leaf NODE, or pthread_create failed.
         Sort with 1 thread. */
      size_t nlo = node->nlo;
      size_t nhi = node->nhi;
      struct line *temp = lines - total_lines;
      if (1 < nhi)
        sequential_sort (lines - nlo, nhi, temp - nlo / 2, false);
      if (1 < nlo)
        sequential_sort (lines, nlo, temp, false);

      /* Update merge NODE. No need to lock yet. */
      node->lo = lines;
      node->hi = lines - nlo;
      node->end_lo = lines - nlo;
      node->end_hi = lines - nlo - nhi;

      queue_insert (queue, node);
      merge_loop (queue, total_lines, tfp, temp_output);
    }
}

/* Scan through FILES[NTEMPS .. NFILES-1] looking for files that are
   the same as OUTFILE.  If found, replace each with the same
   temporary copy that can be merged into OUTFILE without destroying
   OUTFILE before it is completely read.  This temporary copy does not
   count as a merge temp, so don't worry about incrementing NTEMPS in
   the caller; final cleanup will remove it, not zaptemp.

   This test ensures that an otherwise-erroneous use like
   "sort -m -o FILE ... FILE ..." copies FILE before writing to it.
   It's not clear that POSIX requires this nicety.
   Detect common error cases, but don't try to catch obscure cases like
   "cat ... FILE ... | sort -m -o FILE"
   where traditional "sort" doesn't copy the input and where
   people should know that they're getting into trouble anyway.
   Catching these obscure cases would slow down performance in
   common cases.  */

static void
‌avoid_trashing_input (struct sortfile *files, size_t ntemps,
                      size_t nfiles, char const *outfile)
{
  size_t i;
  bool got_outstat = false;
  struct stat outstat;
  struct tempnode *tempcopy = NULL;

  for (i = ntemps; i < nfiles; i++)
    {
      bool is_stdin = STREQ (files[i].name, "-");
      bool same;
      struct stat instat;

      if (outfile && STREQ (outfile, files[i].name) && !is_stdin)
        same = true;
      else
        {
          if (! got_outstat)
            {
              if (fstat (STDOUT_FILENO, &outstat) != 0)
                break;
              got_outstat = true;
            }

          same = (((is_stdin
                    ? fstat (STDIN_FILENO, &instat)
                    : stat (files[i].name, &instat))
                   == 0)
                  && SAME_INODE (instat, outstat));
        }

      if (same)
        {
          if (! tempcopy)
            {
              FILE *tftp;
              tempcopy = create_temp (&tftp);
              mergefiles (&files[i], 0, 1, tftp, tempcopy->name);
            }

          files[i].name = tempcopy->name;
          files[i].temp = tempcopy;
        }
    }
}

/* Scan the input files to ensure all are accessible.
   Otherwise exit with a diagnostic.

   Note this will catch common issues with permissions etc.
   but will fail to notice issues where you can open() but not read(),
   like when a directory is specified on some systems.
   Catching these obscure cases could slow down performance in
   common cases.  */

static void
‌check_inputs (char *const *files, size_t nfiles)
{
  size_t i;
  for (i = 0; i < nfiles; i++)
    {
      if (STREQ (files[i], "-"))
        continue;

      if (euidaccess (files[i], R_OK) != 0)
        die (_("cannot read"), files[i]);
    }
}

/* Ensure a specified output file can be created or written to,
   and point stdout to it.  Do not truncate the file.
   Exit with a diagnostic on failure.  */

static void
‌check_output (char const *outfile)
{
  if (outfile)
    {
      int outfd = open (outfile, O_WRONLY | O_CREAT | O_BINARY, MODE_RW_UGO);
      if (outfd < 0)
        die (_("open failed"), outfile);
      move_fd_or_die (outfd, STDOUT_FILENO);
    }
}

/* Merge the input FILES.  NTEMPS is the number of files at the
   start of FILES that are temporary; it is zero at the top level.
   NFILES is the total number of files.  Put the output in
   OUTPUT_FILE; a null OUTPUT_FILE stands for standard output.  */

static void
‌merge (struct sortfile *files, size_t ntemps, size_t nfiles,
       char const *output_file)
{
  while (nmerge < nfiles)
    {
      /* Number of input files processed so far.  */
      size_t in;

      /* Number of output files generated so far.  */
      size_t out;

      /* nfiles % NMERGE; this counts input files that are left over
         after all full-sized merges have been done.  */
      size_t remainder;

      /* Number of easily-available slots at the next loop iteration.  */
      size_t cheap_slots;

      /* Do as many NMERGE-size merges as possible. In the case that
         nmerge is bogus, increment by the maximum number of file
         descriptors allowed.  */
      for (out = in = 0; nmerge <= nfiles - in; out++)
        {
          FILE *tfp;
          struct tempnode *temp = create_temp (&tfp);
          size_t num_merged = mergefiles (&files[in], MIN (ntemps, nmerge),
                                          nmerge, tfp, temp->name);
          ntemps -= MIN (ntemps, num_merged);
          files[out].name = temp->name;
          files[out].temp = temp;
          in += num_merged;
        }

      remainder = nfiles - in;
      cheap_slots = nmerge - out % nmerge;

      if (cheap_slots < remainder)
        {
          /* So many files remain that they can't all be put into the last
             NMERGE-sized output window.  Do one more merge.  Merge as few
             files as possible, to avoid needless I/O.  */
          size_t nshortmerge = remainder - cheap_slots + 1;
          FILE *tfp;
          struct tempnode *temp = create_temp (&tfp);
          size_t num_merged = mergefiles (&files[in], MIN (ntemps, nshortmerge),
                                          nshortmerge, tfp, temp->name);
          ntemps -= MIN (ntemps, num_merged);
          files[out].name = temp->name;
          files[out++].temp = temp;
          in += num_merged;
        }

      /* Put the remaining input files into the last NMERGE-sized output
         window, so they will be merged in the next pass.  */
      memmove (&files[out], &files[in], (nfiles - in) * sizeof *files);
      ntemps += out;
      nfiles -= in - out;
    }

  avoid_trashing_input (files, ntemps, nfiles, output_file);

  /* We aren't guaranteed that this final mergefiles will work, therefore we
     try to merge into the output, and then merge as much as we can into a
     temp file if we can't. Repeat.  */

  while (true)
    {
      /* Merge directly into the output file if possible.  */
      FILE **fps;
      size_t nopened = open_input_files (files, nfiles, &fps);

      if (nopened == nfiles)
        {
          FILE *ofp = stream_open (output_file, "w");
          if (ofp)
            {
              mergefps (files, ntemps, nfiles, ofp, output_file, fps);
              break;
            }
          if (errno != EMFILE || nopened <= 2)
            die (_("open failed"), output_file);
        }
      else if (nopened <= 2)
        die (_("open failed"), files[nopened].name);

      /* We ran out of file descriptors.  Close one of the input
         files, to gain a file descriptor.  Then create a temporary
         file with our spare file descriptor.  Retry if that failed
         (e.g., some other process could open a file between the time
         we closed and tried to create).  */
      FILE *tfp;
      struct tempnode *temp;
      do
        {
          nopened--;
          xfclose (fps[nopened], files[nopened].name);
          temp = maybe_create_temp (&tfp, ! (nopened <= 2));
        }
      while (!temp);

      /* Merge into the newly allocated temporary.  */
      mergefps (&files[0], MIN (ntemps, nopened), nopened, tfp, temp->name,
                fps);
      ntemps -= MIN (ntemps, nopened);
      files[0].name = temp->name;
      files[0].temp = temp;

      memmove (&files[1], &files[nopened], (nfiles - nopened) * sizeof *files);
      ntemps++;
      nfiles -= nopened - 1;
    }
}

/* Sort NFILES FILES onto OUTPUT_FILE.  Use at most NTHREADS threads.  */

static void
‌sort (char *const *files, size_t nfiles, char const *output_file,
      size_t nthreads)
{
  struct buffer buf;
  IF_LINT (buf.buf = NULL);
  size_t ntemps = 0;
  bool output_file_created = false;

  buf.alloc = 0;

  while (nfiles)
    {
      char const *temp_output;
      char const *file = *files;
      FILE *fp = xfopen (file, "r");
      FILE *tfp;

      size_t bytes_per_line;
      if (nthreads > 1)
        {
          /* Get log P. */
          size_t tmp = 1;
          size_t mult = 1;
          while (tmp < nthreads)
            {
              tmp *= 2;
              mult++;
            }
          bytes_per_line = (mult * sizeof (struct line));
        }
      else
        bytes_per_line = sizeof (struct line) * 3 / 2;

      if (! buf.alloc)
        initbuf (&buf, bytes_per_line,
                 sort_buffer_size (&fp, 1, files, nfiles, bytes_per_line));
      buf.eof = false;
      files++;
      nfiles--;

      while (fillbuf (&buf, fp, file))
        {
          struct line *line;

          if (buf.eof && nfiles
              && (bytes_per_line + 1
                  < (buf.alloc - buf.used - bytes_per_line * buf.nlines)))
            {
              /* End of file, but there is more input and buffer room.
                 Concatenate the next input file; this is faster in
                 the usual case.  */
              buf.left = buf.used;
              break;
            }

          saved_line.text = NULL;
          line = buffer_linelim (&buf);
          if (buf.eof && !nfiles && !ntemps && !buf.left)
            {
              xfclose (fp, file);
              tfp = xfopen (output_file, "w");
              temp_output = output_file;
              output_file_created = true;
            }
          else
            {
              ++ntemps;
              temp_output = create_temp (&tfp)->name;
            }
          if (1 < buf.nlines)
            {
              struct merge_node_queue queue;
              queue_init (&queue, nthreads);
              struct merge_node *merge_tree =
                merge_tree_init (nthreads, buf.nlines, line);

              sortlines (line, nthreads, buf.nlines, merge_tree + 1,
                         &queue, tfp, temp_output);

#ifdef lint
              merge_tree_destroy (nthreads, merge_tree);
              queue_destroy (&queue);
#endif
            }
          else
            write_unique (line - 1, tfp, temp_output);

          xfclose (tfp, temp_output);

          if (output_file_created)
            goto finish;
        }
      xfclose (fp, file);
    }

finish:
  free (buf.buf);

  if (! output_file_created)
    {
      size_t i;
      struct tempnode *node = temphead;
      struct sortfile *tempfiles = xnmalloc (ntemps, sizeof *tempfiles);
      for (i = 0; node; i++)
        {
          tempfiles[i].name = node->name;
          tempfiles[i].temp = node;
          node = node->next;
        }
      merge (tempfiles, ntemps, ntemps, output_file);
      free (tempfiles);
    }

  reap_all ();
}

/* Insert a malloc'd copy of key KEY_ARG at the end of the key list.  */

static void
‌insertkey (struct keyfield *key_arg)
{
  struct keyfield **p;
  struct keyfield *key = xmemdup (key_arg, sizeof *key);

  for (p = &keylist; *p; p = &(*p)->next)
    continue;
  *p = key;
  key->next = NULL;
}

/* Report a bad field specification SPEC, with extra info MSGID.  */

static void badfieldspec (char const *, char const *)
     ATTRIBUTE_NORETURN;
static void
‌badfieldspec (char const *spec, char const *msgid)
{
  error (SORT_FAILURE, 0, _("%s: invalid field specification %s"),
         _(msgid), quote (spec));
  abort ();
}

/* Report incompatible options.  */

static void incompatible_options (char const *) ATTRIBUTE_NORETURN;
static void
‌incompatible_options (char const *opts)
{
  error (SORT_FAILURE, 0, _("options '-%s' are incompatible"), opts);
  abort ();
}

/* Check compatibility of ordering options.  */

static void
‌check_ordering_compatibility (void)
{
  struct keyfield *key;

  for (key = keylist; key; key = key->next)
    if (1 < (key->numeric + key->general_numeric + key->human_numeric
             + key->month + (key->version | key->random | !!key->ignore)))
      {
        /* The following is too big, but guaranteed to be "big enough".  */
        char opts[sizeof short_options];
        /* Clear flags we're not interested in.  */
        key->skipsblanks = key->skipeblanks = key->reverse = false;
        key_to_opts (key, opts);
        incompatible_options (opts);
      }
}

/* Parse the leading integer in STRING and store the resulting value
   (which must fit into size_t) into *VAL.  Return the address of the
   suffix after the integer.  If the value is too large, silently
   substitute SIZE_MAX.  If MSGID is NULL, return NULL after
   failure; otherwise, report MSGID and exit on failure.  */

static char const *
‌parse_field_count (char const *string, size_t *val, char const *msgid)
{
  char *suffix;
  uintmax_t n;

  switch (xstrtoumax (string, &suffix, 10, &n, ""))
    {
    case LONGINT_OK:
    case LONGINT_INVALID_SUFFIX_CHAR:
      *val = n;
      if (*val == n)
        break;
      /* Fall through.  */
    case LONGINT_OVERFLOW:
    case LONGINT_OVERFLOW | LONGINT_INVALID_SUFFIX_CHAR:
      *val = SIZE_MAX;
      break;

    case LONGINT_INVALID:
      if (msgid)
        error (SORT_FAILURE, 0, _("%s: invalid count at start of %s"),
               _(msgid), quote (string));
      return NULL;
    }

  return suffix;
}

/* Handle interrupts and hangups. */

static void
‌sighandler (int sig)
{
  if (! SA_NOCLDSTOP)
    signal (sig, SIG_IGN);

  cleanup ();

  signal (sig, SIG_DFL);
  raise (sig);
}

/* Set the ordering options for KEY specified in S.
   Return the address of the first character in S that
   is not a valid ordering option.
   BLANKTYPE is the kind of blanks that 'b' should skip. */

static char *
‌set_ordering (char const *s, struct keyfield *key, enum blanktype blanktype)
{
  while (*s)
    {
      switch (*s)
        {
        case 'b':
          if (blanktype == bl_start || blanktype == bl_both)
            key->skipsblanks = true;
          if (blanktype == bl_end || blanktype == bl_both)
            key->skipeblanks = true;
          break;
        case 'd':
          key->ignore = nondictionary;
          break;
        case 'f':
          key->translate = fold_toupper;
          break;
        case 'g':
          key->general_numeric = true;
          break;
        case 'h':
          key->human_numeric = true;
          break;
        case 'i':
          /* Option order should not matter, so don't let -i override
             -d.  -d implies -i, but -i does not imply -d.  */
          if (! key->ignore)
            key->ignore = nonprinting;
          break;
        case 'M':
          key->month = true;
          break;
        case 'n':
          key->numeric = true;
          break;
        case 'R':
          key->random = true;
          break;
        case 'r':
          key->reverse = true;
          break;
        case 'V':
          key->version = true;
          break;
        default:
          return (char *) s;
        }
      ++s;
    }
  return (char *) s;
}

/* Initialize KEY.  */

static struct keyfield *
‌key_init (struct keyfield *key)
{
  memset (key, 0, sizeof *key);
  key->eword = SIZE_MAX;
  return key;
}

int
‌main (int argc, char **argv)
{
  struct keyfield *key;
  struct keyfield key_buf;
  struct keyfield gkey;
  bool gkey_only = false;
  char const *s;
  int c = 0;
  char checkonly = 0;
  bool mergeonly = false;
  char *random_source = NULL;
  bool need_random = false;
  size_t nthreads = 0;
  size_t nfiles = 0;
  bool posixly_correct = (getenv ("POSIXLY_CORRECT") != NULL);
  bool obsolete_usage = (posix2_version () < 200112);
  char **files;
  char *files_from = NULL;
  struct Tokens tok;
  char const *outfile = NULL;

  initialize_main (&argc, &argv);
  set_program_name (argv[0]);
  setlocale (LC_ALL, "");
  bindtextdomain (PACKAGE, LOCALEDIR);
  textdomain (PACKAGE);

  initialize_exit_failure (SORT_FAILURE);

  hard_LC_COLLATE = hard_locale (LC_COLLATE);
#if HAVE_NL_LANGINFO
  hard_LC_TIME = hard_locale (LC_TIME);
#endif

  /* Get locale's representation of the decimal point.  */
  {
    struct lconv const *locale = localeconv ();

    /* If the locale doesn't define a decimal point, or if the decimal
       point is multibyte, use the C locale's decimal point.  FIXME:
       add support for multibyte decimal points.  */
    decimal_point = to_uchar (locale->decimal_point[0]);
    if (! decimal_point || locale->decimal_point[1])
      decimal_point = '.';

    /* FIXME: add support for multibyte thousands separators.  */
    thousands_sep = to_uchar (*locale->thousands_sep);
    if (! thousands_sep || locale->thousands_sep[1])
      thousands_sep = -1;
  }

  have_read_stdin = false;
  inittables ();

  {
    size_t i;
    static int const sig[] =
      {
        /* The usual suspects.  */
        SIGALRM, SIGHUP, SIGINT, SIGPIPE, SIGQUIT, SIGTERM,
#ifdef SIGPOLL
        SIGPOLL,
#endif
#ifdef SIGPROF
        SIGPROF,
#endif
#ifdef SIGVTALRM
        SIGVTALRM,
#endif
#ifdef SIGXCPU
        SIGXCPU,
#endif
#ifdef SIGXFSZ
        SIGXFSZ,
#endif
      };
    enum { nsigs = ARRAY_CARDINALITY (sig) };

#if SA_NOCLDSTOP
    struct sigaction act;

    sigemptyset (&caught_signals);
    for (i = 0; i < nsigs; i++)
      {
        sigaction (sig[i], NULL, &act);
        if (act.sa_handler != SIG_IGN)
          sigaddset (&caught_signals, sig[i]);
      }

    act.sa_handler = sighandler;
    act.sa_mask = caught_signals;
    act.sa_flags = 0;

    for (i = 0; i < nsigs; i++)
      if (sigismember (&caught_signals, sig[i]))
        sigaction (sig[i], &act, NULL);
#else
    for (i = 0; i < nsigs; i++)
      if (signal (sig[i], SIG_IGN) != SIG_IGN)
        {
          signal (sig[i], sighandler);
          siginterrupt (sig[i], 1);
        }
#endif
  }
  signal (SIGCHLD, SIG_DFL); /* Don't inherit CHLD handling from parent.  */

  /* The signal mask is known, so it is safe to invoke exit_cleanup.  */
  atexit (exit_cleanup);

  key_init (&gkey);
  gkey.sword = SIZE_MAX;

  files = xnmalloc (argc, sizeof *files);

  while (true)
    {
      /* Parse an operand as a file after "--" was seen; or if
         pedantic and a file was seen, unless the POSIX version
         predates 1003.1-2001 and -c was not seen and the operand is
         "-o FILE" or "-oFILE".  */
      int oi = -1;

      if (c == -1
          || (posixly_correct && nfiles != 0
              && ! (obsolete_usage
                    && ! checkonly
                    && optind != argc
                    && argv[optind][0] == '-' && argv[optind][1] == 'o'
                    && (argv[optind][2] || optind + 1 != argc)))
          || ((c = getopt_long (argc, argv, short_options,
                                long_options, &oi))
              == -1))
        {
          if (argc <= optind)
            break;
          files[nfiles++] = argv[optind++];
        }
      else switch (c)
        {
        case 1:
          key = NULL;
          if (optarg[0] == '+')
            {
              bool minus_pos_usage = (optind != argc && argv[optind][0] == '-'
                                      && ISDIGIT (argv[optind][1]));
              obsolete_usage |= minus_pos_usage && !posixly_correct;
              if (obsolete_usage)
                {
                  /* Treat +POS1 [-POS2] as a key if possible; but silently
                     treat an operand as a file if it is not a valid +POS1.  */
                  key = key_init (&key_buf);
                  s = parse_field_count (optarg + 1, &key->sword, NULL);
                  if (s && *s == '.')
                    s = parse_field_count (s + 1, &key->schar, NULL);
                  if (! (key->sword || key->schar))
                    key->sword = SIZE_MAX;
                  if (! s || *set_ordering (s, key, bl_start))
                    key = NULL;
                  else
                    {
                      if (minus_pos_usage)
                        {
                          char const *optarg1 = argv[optind++];
                          s = parse_field_count (optarg1 + 1, &key->eword,
                                             N_("invalid number after '-'"));
                          /* When called with a non-NULL message ID,
                             parse_field_count cannot return NULL.  Tell static
                             analysis tools that dereferencing S is safe.  */
                          assert (s);
                          if (*s == '.')
                            s = parse_field_count (s + 1, &key->echar,
                                               N_("invalid number after '.'"));
                          if (!key->echar && key->eword)
                            {
                              /* obsolescent syntax +A.x -B.y is equivalent to:
                                   -k A+1.x+1,B.y   (when y = 0)
                                   -k A+1.x+1,B+1.y (when y > 0)
                                 So eword is decremented as in the -k case
                                 only when the end field (B) is specified and
                                 echar (y) is 0.  */
                              key->eword--;
                            }
                          if (*set_ordering (s, key, bl_end))
                            badfieldspec (optarg1,
                                      N_("stray character in field spec"));
                        }
                      key->obsolete_used = true;
                      insertkey (key);
                    }
                }
            }
          if (! key)
            files[nfiles++] = optarg;
          break;

        case SORT_OPTION:
          c = XARGMATCH ("--sort", optarg, sort_args, sort_types);
          /* Fall through. */
        case 'b':
        case 'd':
        case 'f':
        case 'g':
        case 'h':
        case 'i':
        case 'M':
        case 'n':
        case 'r':
        case 'R':
        case 'V':
          {
            char str[2];
            str[0] = c;
            str[1] = '\0';
            set_ordering (str, &gkey, bl_both);
          }
          break;

        case CHECK_OPTION:
          c = (optarg
               ? XARGMATCH ("--check", optarg, check_args, check_types)
               : 'c');
          /* Fall through.  */
        case 'c':
        case 'C':
          if (checkonly && checkonly != c)
            incompatible_options ("cC");
          checkonly = c;
          break;

        case COMPRESS_PROGRAM_OPTION:
          if (compress_program && !STREQ (compress_program, optarg))
            error (SORT_FAILURE, 0, _("multiple compress programs specified"));
          compress_program = optarg;
          break;

        case DEBUG_PROGRAM_OPTION:
          debug = true;
          break;

        case FILES0_FROM_OPTION:
          files_from = optarg;
          break;

        case 'k':
          key = key_init (&key_buf);

          /* Get POS1. */
          s = parse_field_count (optarg, &key->sword,
                                 N_("invalid number at field start"));
          if (! key->sword--)
            {
              /* Provoke with 'sort -k0' */
              badfieldspec (optarg, N_("field number is zero"));
            }
          if (*s == '.')
            {
              s = parse_field_count (s + 1, &key->schar,
                                     N_("invalid number after '.'"));
              if (! key->schar--)
                {
                  /* Provoke with 'sort -k1.0' */
                  badfieldspec (optarg, N_("character offset is zero"));
                }
            }
          if (! (key->sword || key->schar))
            key->sword = SIZE_MAX;
          s = set_ordering (s, key, bl_start);
          if (*s != ',')
            {
              key->eword = SIZE_MAX;
              key->echar = 0;
            }
          else
            {
              /* Get POS2. */
              s = parse_field_count (s + 1, &key->eword,
                                     N_("invalid number after ','"));
              if (! key->eword--)
                {
                  /* Provoke with 'sort -k1,0' */
                  badfieldspec (optarg, N_("field number is zero"));
                }
              if (*s == '.')
                {
                  s = parse_field_count (s + 1, &key->echar,
                                         N_("invalid number after '.'"));
                }
              s = set_ordering (s, key, bl_end);
            }
          if (*s)
            badfieldspec (optarg, N_("stray character in field spec"));
          insertkey (key);
          break;

        case 'm':
          mergeonly = true;
          break;

        case NMERGE_OPTION:
          specify_nmerge (oi, c, optarg);
          break;

        case 'o':
          if (outfile && !STREQ (outfile, optarg))
            error (SORT_FAILURE, 0, _("multiple output files specified"));
          outfile = optarg;
          break;

        case RANDOM_SOURCE_OPTION:
          if (random_source && !STREQ (random_source, optarg))
            error (SORT_FAILURE, 0, _("multiple random sources specified"));
          random_source = optarg;
          break;

        case 's':
          stable = true;
          break;

        case 'S':
          specify_sort_size (oi, c, optarg);
          break;

        case 't':
          {
            char newtab = optarg[0];
            if (! newtab)
              error (SORT_FAILURE, 0, _("empty tab"));
            if (optarg[1])
              {
                if (STREQ (optarg, "\\0"))
                  newtab = '\0';
                else
                  {
                    /* Provoke with 'sort -txx'.  Complain about
                       "multi-character tab" instead of "multibyte tab", so
                       that the diagnostic's wording does not need to be
                       changed once multibyte characters are supported.  */
                    error (SORT_FAILURE, 0, _("multi-character tab %s"),
                           quote (optarg));
                  }
              }
            if (tab != TAB_DEFAULT && tab != newtab)
              error (SORT_FAILURE, 0, _("incompatible tabs"));
            tab = newtab;
          }
          break;

        case 'T':
          add_temp_dir (optarg);
          break;

        case PARALLEL_OPTION:
          nthreads = specify_nthreads (oi, c, optarg);
          break;

        case 'u':
          unique = true;
          break;

        case 'y':
          /* Accept and ignore e.g. -y0 for compatibility with Solaris 2.x
             through Solaris 7.  It is also accepted by many non-Solaris
             "sort" implementations, e.g., AIX 5.2, HP-UX 11i v2, IRIX 6.5.
             -y is marked as obsolete starting with Solaris 8 (1999), but is
             still accepted as of Solaris 10 prerelease (2004).

             Solaris 2.5.1 "sort -y 100" reads the input file "100", but
             emulate Solaris 8 and 9 "sort -y 100" which ignores the "100",
             and which in general ignores the argument after "-y" if it
             consists entirely of digits (it can even be empty).  */
          if (optarg == argv[optind - 1])
            {
              char const *p;
              for (p = optarg; ISDIGIT (*p); p++)
                continue;
              optind -= (*p != '\0');
            }
          break;

        case 'z':
          eolchar = 0;
          break;

        case_GETOPT_HELP_CHAR;

        case_GETOPT_VERSION_CHAR (PROGRAM_NAME, AUTHORS);

        default:
          usage (SORT_FAILURE);
        }
    }

  if (files_from)
    {
      FILE *stream;

      /* When using --files0-from=F, you may not specify any files
         on the command-line.  */
      if (nfiles)
        {
          error (0, 0, _("extra operand %s"), quote (files[0]));
          fprintf (stderr, "%s\n",
                   _("file operands cannot be combined with --files0-from"));
          usage (SORT_FAILURE);
        }

      if (STREQ (files_from, "-"))
        stream = stdin;
      else
        {
          stream = fopen (files_from, "r");
          if (stream == NULL)
            error (SORT_FAILURE, errno, _("cannot open %s for reading"),
                   quote (files_from));
        }

      readtokens0_init (&tok);

      if (! readtokens0 (stream, &tok) || fclose (stream) != 0)
        error (SORT_FAILURE, 0, _("cannot read file names from %s"),
               quote (files_from));

      if (tok.n_tok)
        {
          size_t i;
          free (files);
          files = tok.tok;
          nfiles = tok.n_tok;
          for (i = 0; i < nfiles; i++)
            {
              if (STREQ (files[i], "-"))
                error (SORT_FAILURE, 0, _("when reading file names from stdin, "
                                          "no file name of %s allowed"),
                       quote (files[i]));
              else if (files[i][0] == '\0')
                {
                  /* Using the standard 'filename:line-number:' prefix here is
                     not totally appropriate, since NUL is the separator,
                     not NL, but it might be better than nothing.  */
                  unsigned long int file_number = i + 1;
                  error (SORT_FAILURE, 0,
                         _("%s:%lu: invalid zero-length file name"),
                         quotearg_colon (files_from), file_number);
                }
            }
        }
      else
        error (SORT_FAILURE, 0, _("no input from %s"),
               quote (files_from));
    }

  /* Inheritance of global options to individual keys. */
  for (key = keylist; key; key = key->next)
    {
      if (default_key_compare (key) && !key->reverse)
        {
          key->ignore = gkey.ignore;
          key->translate = gkey.translate;
          key->skipsblanks = gkey.skipsblanks;
          key->skipeblanks = gkey.skipeblanks;
          key->month = gkey.month;
          key->numeric = gkey.numeric;
          key->general_numeric = gkey.general_numeric;
          key->human_numeric = gkey.human_numeric;
          key->version = gkey.version;
          key->random = gkey.random;
          key->reverse = gkey.reverse;
        }

      need_random |= key->random;
    }

  if (!keylist && !default_key_compare (&gkey))
    {
      gkey_only = true;
      insertkey (&gkey);
      need_random |= gkey.random;
    }

  check_ordering_compatibility ();

  if (debug)
    {
      if (checkonly || outfile)
        {
          static char opts[] = "X --debug";
          opts[0] = (checkonly ? checkonly : 'o');
          incompatible_options (opts);
        }

      /* Always output the locale in debug mode, since this
         is such a common source of confusion.  */
      if (hard_LC_COLLATE)
        error (0, 0, _("using %s sorting rules"),
               quote (setlocale (LC_COLLATE, NULL)));
      else
        error (0, 0, _("using simple byte comparison"));
      key_warnings (&gkey, gkey_only);
    }

  reverse = gkey.reverse;

  if (need_random)
    random_md5_state_init (random_source);

  if (temp_dir_count == 0)
    {
      char const *tmp_dir = getenv ("TMPDIR");
      add_temp_dir (tmp_dir ? tmp_dir : DEFAULT_TMPDIR);
    }

  if (nfiles == 0)
    {
      static char *minus = (char *) "-";
      nfiles = 1;
      free (files);
      files = &minus;
    }

  /* Need to re-check that we meet the minimum requirement for memory
     usage with the final value for NMERGE. */
  if (0 < sort_size)
    sort_size = MAX (sort_size, MIN_SORT_SIZE);

  if (checkonly)
    {
      if (nfiles > 1)
        error (SORT_FAILURE, 0, _("extra operand %s not allowed with -%c"),
               quote (files[1]), checkonly);

      if (outfile)
        {
          static char opts[] = {0, 'o', 0};
          opts[0] = checkonly;
          incompatible_options (opts);
        }

      /* POSIX requires that sort return 1 IFF invoked with -c or -C and the
         input is not properly sorted.  */
      exit (check (files[0], checkonly) ? EXIT_SUCCESS : SORT_OUT_OF_ORDER);
    }

  /* Check all inputs are accessible, or exit immediately.  */
  check_inputs (files, nfiles);

  /* Check output is writable, or exit immediately.  */
  check_output (outfile);

  if (mergeonly)
    {
      struct sortfile *sortfiles = xcalloc (nfiles, sizeof *sortfiles);
      size_t i;

      for (i = 0; i < nfiles; ++i)
        sortfiles[i].name = files[i];

      merge (sortfiles, 0, nfiles, outfile);
      IF_LINT (free (sortfiles));
    }
  else
    {
      if (!nthreads)
        {
          unsigned long int np = num_processors (NPROC_CURRENT_OVERRIDABLE);
          nthreads = MIN (np, DEFAULT_MAX_THREADS);
        }

      /* Avoid integer overflow later.  */
      size_t nthreads_max = SIZE_MAX / (2 * sizeof (struct merge_node));
      nthreads = MIN (nthreads, nthreads_max);

      sort (files, nfiles, outfile, nthreads);
    }

  if (have_read_stdin && fclose (stdin) == EOF)
    die (_("close failed"), "-");

  exit (EXIT_SUCCESS);
}