glibc/locale/programs/ld-collate.c

/* Copyright (C) 1995-1999, 2000 Free Software Foundation, Inc.
   This file is part of the GNU C Library.
   Contributed by Ulrich Drepper <drepper@gnu.org>, 1995.

   The GNU C Library is free software; you can redistribute it and/or
   modify it under the terms of the GNU Library General Public License as
   published by the Free Software Foundation; either version 2 of the
   License, or (at your option) any later version.

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

   You should have received a copy of the GNU Library General Public
   License along with the GNU C Library; see the file COPYING.LIB.  If not,
   write to the Free Software Foundation, Inc., 59 Temple Place - Suite 330,
   Boston, MA 02111-1307, USA.  */

#ifdef HAVE_CONFIG_H
# include <config.h>
#endif

#include <errno.h>
#include <error.h>
#include <stdlib.h>
#include <wchar.h>
#include <sys/param.h>

#include "charmap.h"
#include "localeinfo.h"
#include "linereader.h"
#include "locfile.h"
#include "localedef.h"
#include "elem-hash.h"

/* Uncomment the following line in the production version.  */
/* #define NDEBUG 1 */
#include <assert.h>

#define obstack_chunk_alloc malloc
#define obstack_chunk_free free

/* Forward declaration.  */
struct element_t;

/* Data type for list of strings.  */
struct section_list
{
  struct section_list *next;
  /* Name of the section.  */
  const char *name;
  /* First element of this section.  */
  struct element_t *first;
  /* Last element of this section.  */
  struct element_t *last;
  /* These are the rules for this section.  */
  enum coll_sort_rule *rules;
  /* Index of the rule set in the appropriate section of the output file.  */
  int ruleidx;
};

struct element_t;

struct element_list_t
{
  /* Number of elements.  */
  int cnt;

  struct element_t **w;
};

/* Data type for collating element.  */
struct element_t
{
  const char *name;

  const char *mbs;
  size_t nmbs;
  const uint32_t *wcs;
  size_t nwcs;
  int *mborder;
  int wcorder;

  /* The following is a bit mask which bits are set if this element is
     used in the appropriate level.  Interesting for the singlebyte
     weight computation.

     XXX The type here restricts the number of levels to 32.  It could
     we changed if necessary but I doubt this is necessary.  */
  unsigned int used_in_level;

  struct element_list_t *weights;

  /* Nonzero if this is a real character definition.  */
  int is_character;

  /* Where does the definition come from.  */
  const char *file;
  size_t line;

  /* Which section does this belong to.  */
  struct section_list *section;

  /* Predecessor and successor in the order list.  */
  struct element_t *last;
  struct element_t *next;

  /* Next element in multibyte output list.  */
  struct element_t *mbnext;
  struct element_t *mblast;

  /* Next element in wide character output list.  */
  struct element_t *wcnext;
  struct element_t *wclast;
};

/* Special element value.  */
#define ELEMENT_ELLIPSIS2	((struct element_t *) 1)
#define ELEMENT_ELLIPSIS3	((struct element_t *) 2)
#define ELEMENT_ELLIPSIS4	((struct element_t *) 3)

/* Data type for collating symbol.  */
struct symbol_t
{
  /* Point to place in the order list.  */
  struct element_t *order;

  /* Where does the definition come from.  */
  const char *file;
  size_t line;
};


/* The real definition of the struct for the LC_COLLATE locale.  */
struct locale_collate_t
{
  int col_weight_max;
  int cur_weight_max;

  /* List of known scripts.  */
  struct section_list *sections;
  /* Current section using definition.  */
  struct section_list *current_section;
  /* There always can be an unnamed section.  */
  struct section_list unnamed_section;
  /* To make handling of errors easier we have another section.  */
  struct section_list error_section;

  /* Start of the order list.  */
  struct element_t *start;

  /* The undefined element.  */
  struct element_t undefined;

  /* This is the cursor for `reorder_after' insertions.  */
  struct element_t *cursor;

  /* This value is used when handling ellipsis.  */
  struct element_t ellipsis_weight;

  /* Known collating elements.  */
  hash_table elem_table;

  /* Known collating symbols.  */
  hash_table sym_table;

  /* Known collation sequences.  */
  hash_table seq_table;

  struct obstack mempool;

  /* The LC_COLLATE category is a bit special as it is sometimes possible
     that the definitions from more than one input file contains information.
     Therefore we keep all relevant input in a list.  */
  struct locale_collate_t *next;

  /* Arrays with heads of the list for each of the leading bytes in
     the multibyte sequences.  */
  struct element_t *mbheads[256];

  /* Table size of wide character hash table.  */
  size_t plane_size;
  size_t plane_cnt;

  /* Arrays with heads of the list for each of the leading bytes in
     the multibyte sequences.  */
  struct element_t **wcheads;
};


/* We have a few global variables which are used for reading all
   LC_COLLATE category descriptions in all files.  */
static uint32_t nrules;


/* These are definitions used by some of the functions for handling
   UTF-8 encoding below.  */
static const uint32_t encoding_mask[] =
{
  ~0x7ff, ~0xffff, ~0x1fffff, ~0x3ffffff
};

static const unsigned char encoding_byte[] =
{
  0xc0, 0xe0, 0xf0, 0xf8, 0xfc
};


/* We need UTF-8 encoding of numbers.  */
static inline int
utf8_encode (char *buf, int val)
{
  int retval;

  if (val < 0x80)
    {
      *buf++ = (char) val;
      retval = 1;
    }
  else
    {
      int step;

      for (step = 2; step < 6; ++step)
	if ((val & encoding_mask[step - 2]) == 0)
	  break;
      retval = step;

      *buf = encoding_byte[step - 2];
      --step;
      do
	{
	  buf[step] = 0x80 | (val & 0x3f);
	  val >>= 6;
	}
      while (--step > 0);
      *buf |= val;
    }

  return retval;
}


static struct section_list *
make_seclist_elem (struct locale_collate_t *collate, const char *string,
		   struct section_list *next)
{
  struct section_list *newp;

  newp = (struct section_list *) obstack_alloc (&collate->mempool,
						sizeof (*newp));
  newp->next = next;
  newp->name = string;
  newp->first = NULL;

  return newp;
}


static struct element_t *
new_element (struct locale_collate_t *collate, const char *mbs, size_t mbslen,
	     const uint32_t *wcs, const char *name, size_t namelen,
	     int is_character)
{
  struct element_t *newp;

  newp = (struct element_t *) obstack_alloc (&collate->mempool,
					     sizeof (*newp));
  newp->name = name == NULL ? NULL : obstack_copy0 (&collate->mempool,
						    name, namelen);
  if (mbs != NULL)
    {
      newp->mbs = obstack_copy0 (&collate->mempool, mbs, mbslen);
      newp->nmbs = mbslen;
    }
  else
    {
      newp->mbs = NULL;
      newp->nmbs = 0;
    }
  if (wcs != NULL)
    {
      size_t nwcs = wcslen ((wchar_t *) wcs);
      uint32_t zero = 0;
      obstack_grow (&collate->mempool, wcs, nwcs * sizeof (uint32_t));
      obstack_grow (&collate->mempool, &zero, sizeof (uint32_t));
      newp->wcs = (uint32_t *) obstack_finish (&collate->mempool);
      newp->nwcs = nwcs;
    }
  else
    {
      newp->wcs = NULL;
      newp->nwcs = 0;
    }
  newp->mborder = NULL;
  newp->wcorder = 0;
  newp->used_in_level = 0;
  newp->is_character = is_character;

  /* Will be allocated later.  */
  newp->weights = NULL;

  newp->file = NULL;
  newp->line = 0;

  newp->section = collate->current_section;

  newp->last = NULL;
  newp->next = NULL;

  newp->mbnext = NULL;
  newp->mblast = NULL;

  return newp;
}


static struct symbol_t *
new_symbol (struct locale_collate_t *collate)
{
  struct symbol_t *newp;

  newp = (struct symbol_t *) obstack_alloc (&collate->mempool, sizeof (*newp));

  newp->order = NULL;

  newp->file = NULL;
  newp->line = 0;

  return newp;
}


/* Test whether this name is already defined somewhere.  */
static int
check_duplicate (struct linereader *ldfile, struct locale_collate_t *collate,
		 struct charmap_t *charmap, struct repertoire_t *repertoire,
		 const char *symbol, size_t symbol_len)
{
  void *ignore = NULL;

  if (find_entry (&charmap->char_table, symbol, symbol_len, &ignore) == 0)
    {
      lr_error (ldfile, _("`%.*s' already defined in charmap"),
		(int) symbol_len, symbol);
      return 1;
    }

  if (find_entry (&repertoire->char_table, symbol, symbol_len, &ignore) == 0)
    {
      lr_error (ldfile, _("`%.*s' already defined in repertoire"),
		(int) symbol_len, symbol);
      return 1;
    }

  if (find_entry (&collate->sym_table, symbol, symbol_len, &ignore) == 0)
    {
      lr_error (ldfile, _("`%.*s' already defined as collating symbol"),
		(int) symbol_len, symbol);
      return 1;
    }

  if (find_entry (&collate->elem_table, symbol, symbol_len, &ignore) == 0)
    {
      lr_error (ldfile, _("`%.*s' already defined as collating element"),
		(int) symbol_len, symbol);
      return 1;
    }

  return 0;
}


/* Read the direction specification.  */
static void
read_directions (struct linereader *ldfile, struct token *arg,
		 struct charmap_t *charmap, struct repertoire_t *repertoire,
		 struct locale_collate_t *collate)
{
  int cnt = 0;
  int max = nrules ?: 10;
  enum coll_sort_rule *rules = calloc (max, sizeof (*rules));
  int warned = 0;

  while (1)
    {
      int valid = 0;

      if (arg->tok == tok_forward)
	{
	  if (rules[cnt] & sort_backward)
	    {
	      if (! warned)
		{
		  lr_error (ldfile, _("\
%s: `forward' and `backward' are mutually excluding each other"),
			    "LC_COLLATE");
		  warned = 1;
		}
	    }
	  else if (rules[cnt] & sort_forward)
	    {
	      if (! warned)
		{
		  lr_error (ldfile, _("\
%s: `%s' mentioned more than once in definition of weight %d"),
			    "LC_COLLATE", "forward", cnt + 1);
		}
	    }
	  else
	    rules[cnt] |= sort_forward;

	  valid = 1;
	}
      else if (arg->tok == tok_backward)
	{
	  if (rules[cnt] & sort_forward)
	    {
	      if (! warned)
		{
		  lr_error (ldfile, _("\
%s: `forward' and `backward' are mutually excluding each other"),
			    "LC_COLLATE");
		  warned = 1;
		}
	    }
	  else if (rules[cnt] & sort_backward)
	    {
	      if (! warned)
		{
		  lr_error (ldfile, _("\
%s: `%s' mentioned more than once in definition of weight %d"),
			    "LC_COLLATE", "backward", cnt + 1);
		}
	    }
	  else
	    rules[cnt] |= sort_backward;

	  valid = 1;
	}
      else if (arg->tok == tok_position)
	{
	  if (rules[cnt] & sort_position)
	    {
	      if (! warned)
		{
		  lr_error (ldfile, _("\
%s: `%s' mentioned more than once in definition of weight %d"),
			    "LC_COLLATE", "position", cnt + 1);
		}
	    }
	  else
	    rules[cnt] |= sort_position;

	  valid = 1;
	}

      if (valid)
	arg = lr_token (ldfile, charmap, repertoire);

      if (arg->tok == tok_eof || arg->tok == tok_eol || arg->tok == tok_comma
	  || arg->tok == tok_semicolon)
	{
	  if (! valid && ! warned)
	    {
	      lr_error (ldfile, _("%s: syntax error"), "LC_COLLATE");
	      warned = 1;
	    }

	  /* See whether we have to increment the counter.  */
	  if (arg->tok != tok_comma && rules[cnt] != 0)
	    {
	      /* Add the default `forward' if we have seen only `position'.  */
	      if (rules[cnt] == sort_position)
		rules[cnt] = sort_position | sort_forward;

	      ++cnt;
	    }

	  if (arg->tok == tok_eof || arg->tok == tok_eol)
	    /* End of line or file, so we exit the loop.  */
	    break;

	  if (nrules == 0)
	    {
	      /* See whether we have enough room in the array.  */
	      if (cnt == max)
		{
		  max += 10;
		  rules = (enum coll_sort_rule *) xrealloc (rules,
							    max
							    * sizeof (*rules));
		  memset (&rules[cnt], '\0', (max - cnt) * sizeof (*rules));
		}
	    }
	  else
	    {
	      if (cnt == nrules)
		{
		  /* There must not be any more rule.  */
		  if (! warned)
		    {
		      lr_error (ldfile, _("\
%s: too many rules; first entry only had %d"),
				"LC_COLLATE", nrules);
		      warned = 1;
		    }

		  lr_ignore_rest (ldfile, 0);
		  break;
		}
	    }
	}
      else
	{
	  if (! warned)
	    {
	      lr_error (ldfile, _("%s: syntax error"), "LC_COLLATE");
	      warned = 1;
	    }
	}

      arg = lr_token (ldfile, charmap, repertoire);
    }

  if (nrules == 0)
    {
      /* Now we know how many rules we have.  */
      nrules = cnt;
      rules = (enum coll_sort_rule *) xrealloc (rules,
						nrules * sizeof (*rules));
    }
  else
    {
      if (cnt < nrules)
	{
	  /* Not enough rules in this specification.  */
	  if (! warned)
	    lr_error (ldfile, _("%s: not enough sorting rules"), "LC_COLLATE");

	  do
	    rules[cnt] = sort_forward;
	  while (++cnt < nrules);
	}
    }

  collate->current_section->rules = rules;
}


static struct element_t *
find_element (struct linereader *ldfile, struct locale_collate_t *collate,
	      const char *str, size_t len, uint32_t *wcstr)
{
  struct element_t *result = NULL;

  /* Search for the entries among the collation sequences already define.  */
  if (find_entry (&collate->seq_table, str, len, (void **) &result) != 0)
    {
      /* Nope, not define yet.  So we see whether it is a
         collation symbol.  */
      void *ptr;

      if (find_entry (&collate->sym_table, str, len, &ptr) == 0)
	{
	  /* It's a collation symbol.  */
	  struct symbol_t *sym = (struct symbol_t *) ptr;
	  result = sym->order;

	  if (result == NULL)
	    result = sym->order = new_element (collate, NULL, 0, NULL,
					       NULL, 0, 0);
	}
      else if (find_entry (&collate->elem_table, str, len,
			   (void **) &result) != 0)
	{
	  /* It's also no collation element.  So it is a character
	     element defined later.  */
	  result = new_element (collate, NULL, 0, NULL, str, len, 1);
	  if (result != NULL)
	    /* Insert it into the sequence table.  */
	    insert_entry (&collate->seq_table, str, len, result);
	}
    }

  return result;
}


static void
unlink_element (struct locale_collate_t *collate)
{
  if (collate->cursor == collate->start)
    {
      assert (collate->cursor->next == NULL);
      assert (collate->cursor->last == NULL);
      collate->cursor = NULL;
    }
  else
    {
      if (collate->cursor->next != NULL)
	collate->cursor->next->last = collate->cursor->last;
      if (collate->cursor->last != NULL)
	collate->cursor->last->next = collate->cursor->next;
      collate->cursor = collate->cursor->last;
    }
}


static void
insert_weights (struct linereader *ldfile, struct element_t *elem,
		struct charmap_t *charmap, struct repertoire_t *repertoire,
		struct locale_collate_t *collate, enum token_t ellipsis)
{
  int weight_cnt;
  struct token *arg;

  /* Initialize all the fields.  */
  elem->file = ldfile->fname;
  elem->line = ldfile->lineno;
  elem->last = collate->cursor;
  elem->next = collate->cursor ? collate->cursor->next : NULL;
  elem->section = collate->current_section;
  if (collate->cursor != NULL)
    collate->cursor->next = elem;
  if (collate->start == NULL)
    {
      assert (collate->cursor == NULL);
      collate->start = elem;
    }
  elem->weights = (struct element_list_t *)
    obstack_alloc (&collate->mempool, nrules * sizeof (struct element_list_t));
  memset (elem->weights, '\0', nrules * sizeof (struct element_list_t));

  if (collate->current_section->first == NULL)
    collate->current_section->first = elem;
  if (collate->current_section->last == collate->cursor)
    collate->current_section->last = elem;

  collate->cursor = elem;

  weight_cnt = 0;

  arg = lr_token (ldfile, charmap, repertoire);
  do
    {
      if (arg->tok == tok_eof || arg->tok == tok_eol)
	break;

      if (arg->tok == tok_ignore)
	{
	  /* The weight for this level has to be ignored.  We use the
	     null pointer to indicate this.  */
	  elem->weights[weight_cnt].w = (struct element_t **)
	    obstack_alloc (&collate->mempool, sizeof (struct element_t *));
	  elem->weights[weight_cnt].w[0] = NULL;
	  elem->weights[weight_cnt].cnt = 1;
	}
      else if (arg->tok == tok_bsymbol)
	{
	  struct element_t *val = find_element (ldfile, collate,
						arg->val.str.startmb,
						arg->val.str.lenmb,
						arg->val.str.startwc);

	  if (val == NULL)
	    break;

	  elem->weights[weight_cnt].w = (struct element_t **)
	    obstack_alloc (&collate->mempool, sizeof (struct element_t *));
	  elem->weights[weight_cnt].w[0] = val;
	  elem->weights[weight_cnt].cnt = 1;
	}
      else if (arg->tok == tok_string)
	{
	  /* Split the string up in the individual characters and put
	     the element definitions in the list.  */
	  const char *cp = arg->val.str.startmb;
	  int cnt = 0;
	  struct element_t *charelem;
	  struct element_t **weights = NULL;
	  int max = 0;

	  if (*cp == '\0')
	    {
	      lr_error (ldfile, _("%s: empty weight string not allowed"),
			"LC_COLLATE");
	      lr_ignore_rest (ldfile, 0);
	      break;
	    }

	  do
	    {
	      if (*cp == '<')
		{
		  /* Ahh, it's a bsymbol.  That's what we want.  */
		  const char *startp = ++cp;

		  while (*cp != '>')
		    {
		      if (*cp == ldfile->escape_char)
			++cp;
		      if (*cp == '\0')
			/* It's a syntax error.  */
			goto syntax;

		      ++cp;
		    }

		    charelem = find_element (ldfile, collate, startp,
					     cp - startp, NULL);
		    ++cp;
		}
	      else
		{
		  /* People really shouldn't use characters directly in
		     the string.  Especially since it's not really clear
		     what this means.  We interpret all characters in the
		     string as if that would be bsymbols.  Otherwise we
		     would have to match back to bsymbols somehow and this
		     is normally not what people normally expect.  */
		  charelem = find_element (ldfile, collate, cp++, 1, NULL);
		}

	      if (charelem == NULL)
		{
		  /* We ignore the rest of the line.  */
		  lr_ignore_rest (ldfile, 0);
		  break;
		}

	      /* Add the pointer.  */
	      if (cnt >= max)
		{
		  struct element_t **newp;
		  max += 10;
		  newp = (struct element_t **)
		    alloca (max * sizeof (struct element_t *));
		  memcpy (newp, weights, cnt * sizeof (struct element_t *));
		  weights = newp;
		}
	      weights[cnt++] = charelem;
	    }
	  while (*cp != '\0');

	  /* Now store the information.  */
	  elem->weights[weight_cnt].w = (struct element_t **)
	    obstack_alloc (&collate->mempool,
			   cnt * sizeof (struct element_t *));
	  memcpy (elem->weights[weight_cnt].w, weights,
		  cnt * sizeof (struct element_t *));
	  elem->weights[weight_cnt].cnt = cnt;

	  /* We don't need the string anymore.  */
	  free (arg->val.str.startmb);
	}
      else if (ellipsis != tok_none
	       && (arg->tok == tok_ellipsis2
		   || arg->tok == tok_ellipsis3
		   || arg->tok == tok_ellipsis4))
	{
	  /* It must be the same ellipsis as used in the initial column.  */
	  if (arg->tok != ellipsis)
	    lr_error (ldfile, _("\
%s: weights must use the same ellipsis symbol as the name"),
		      "LC_COLLATE");

	  /* The weight for this level has to be ignored.  We use the
	     null pointer to indicate this.  */
	  elem->weights[weight_cnt].w = (struct element_t **)
	    obstack_alloc (&collate->mempool, sizeof (struct element_t *));
	  elem->weights[weight_cnt].w[0] = ELEMENT_ELLIPSIS2;
	  elem->weights[weight_cnt].cnt = 1;
	}
      else
	{
	syntax:
	  /* It's a syntax error.  */
	  lr_error (ldfile, _("%s: syntax error"), "LC_COLLATE");
	  lr_ignore_rest (ldfile, 0);
	  break;
	}

      arg = lr_token (ldfile, charmap, repertoire);
      /* This better should be the end of the line or a semicolon.  */
      if (arg->tok == tok_semicolon)
	/* OK, ignore this and read the next token.  */
	arg = lr_token (ldfile, charmap, repertoire);
      else if (arg->tok != tok_eof && arg->tok != tok_eol)
	{
	  /* It's a syntax error.  */
	  lr_error (ldfile, _("%s: syntax error"), "LC_COLLATE");
	  lr_ignore_rest (ldfile, 0);
	  break;
	}
    }
  while (++weight_cnt < nrules);

  if (weight_cnt < nrules)
    {
      /* This means the rest of the line uses the current element as
	 the weight.  */
      do
	{
	  elem->weights[weight_cnt].w = (struct element_t **)
	    obstack_alloc (&collate->mempool, sizeof (struct element_t *));
	  if (ellipsis == tok_none)
	    elem->weights[weight_cnt].w[0] = elem;
	  else
	    elem->weights[weight_cnt].w[0] = ELEMENT_ELLIPSIS2;
	  elem->weights[weight_cnt].cnt = 1;
	}
      while (++weight_cnt < nrules);
    }
  else
    {
      if (arg->tok == tok_ignore || arg->tok == tok_bsymbol)
	{
	  /* Too many rule values.  */
	  lr_error (ldfile, _("%s: too many values"), "LC_COLLATE");
	  lr_ignore_rest (ldfile, 0);
	}
      else
	lr_ignore_rest (ldfile, arg->tok != tok_eol && arg->tok != tok_eof);
    }
}


static int
insert_value (struct linereader *ldfile, struct token *arg,
	      struct charmap_t *charmap, struct repertoire_t *repertoire,
	      struct locale_collate_t *collate)
{
  /* First find out what kind of symbol this is.  */
  struct charseq *seq;
  uint32_t wc;
  struct element_t *elem = NULL;

  /* Try to find the character in the charmap.  */
  seq = charmap_find_value (charmap, arg->val.str.startmb, arg->val.str.lenmb);

  /* Determine the wide character.  */
  if (seq == NULL || seq->ucs4 == UNINITIALIZED_CHAR_VALUE)
    {
      wc = repertoire_find_value (repertoire, arg->val.str.startmb,
				  arg->val.str.lenmb);
      if (seq != NULL)
	seq->ucs4 = wc;
    }
  else
    wc = seq->ucs4;

  if (wc == ILLEGAL_CHAR_VALUE && seq == NULL)
    {
      /* It's no character, so look through the collation elements and
	 symbol list.  */
      void *result;

      if (find_entry (&collate->sym_table, arg->val.str.startmb,
		      arg->val.str.lenmb, &result) == 0)
	{
	  /* It's a collation symbol.  */
	  struct symbol_t *sym = (struct symbol_t *) result;
	  elem = sym->order;

	  if (elem == NULL)
	    elem = sym->order = new_element (collate, NULL, 0, NULL, NULL, 0,
					     0);
	}
      else if (find_entry (&collate->elem_table, arg->val.str.startmb,
			   arg->val.str.lenmb, (void **) &elem) != 0)
	{
	  /* It's also no collation element.  Therefore ignore it.  */
	  lr_ignore_rest (ldfile, 0);
	  return 1;
	}
    }
  else
    {
      /* Otherwise the symbols stands for a character.  */
      if (find_entry (&collate->seq_table, arg->val.str.startmb,
		      arg->val.str.lenmb, (void **) &elem) != 0)
	{
	  uint32_t wcs[2] = { wc, 0 };

	  /* We have to allocate an entry.  */
	  elem = new_element (collate, seq != NULL ? seq->bytes : NULL,
			      seq != NULL ? seq->nbytes : 0,
			      wc == ILLEGAL_CHAR_VALUE ? NULL : wcs,
			      arg->val.str.startmb, arg->val.str.lenmb, 1);

	  /* And add it to the table.  */
	  if (insert_entry (&collate->seq_table, arg->val.str.startmb,
			    arg->val.str.lenmb, elem) != 0)
	    /* This cannot happen.  */
	    assert (! "Internal error");
	}
      else
	{
	  /* Maybe the character was used before the definition.  In this case
	     we have to insert the byte sequences now.  */
	  if (elem->mbs == NULL && seq != NULL)
	    {
	      elem->mbs = obstack_copy0 (&collate->mempool,
					 seq->bytes, seq->nbytes);
	      elem->nmbs = seq->nbytes;
	    }

	  if (elem->wcs == NULL && wc != ILLEGAL_CHAR_VALUE)
	    {
	      uint32_t wcs[2] = { wc, 0 };

	      elem->wcs = obstack_copy (&collate->mempool, wcs, sizeof (wcs));
	      elem->nwcs = 1;
	    }
	}
    }

  /* Test whether this element is not already in the list.  */
  if (elem->next != NULL || (collate->cursor != NULL
			     && elem->next == collate->cursor))
    {
      lr_error (ldfile, _("order for `%.*s' already defined at %s:%zu"),
		(int) arg->val.str.lenmb, arg->val.str.startmb,
		elem->file, elem->line);
      lr_ignore_rest (ldfile, 0);
      return 1;
    }

  insert_weights (ldfile, elem, charmap, repertoire, collate, tok_none);

  return 0;
}


static void
handle_ellipsis (struct linereader *ldfile, struct token *arg,
		 enum token_t ellipsis, struct charmap_t *charmap,
		 struct repertoire_t *repertoire,
		 struct locale_collate_t *collate)
{
  struct element_t *startp;
  struct element_t *endp;

  /* Unlink the entry added for the ellipsis.  */
  unlink_element (collate);
  startp = collate->cursor;

  /* Process and add the end-entry.  */
  if (arg != NULL
      && insert_value (ldfile, arg, charmap, repertoire, collate))
    /* Something went wrong with inserting the to-value.  This means
       we cannot process the ellipsis.  */
    return;

  /* Reset the cursor.  */
  collate->cursor = startp;

  /* Now we have to handle many different situations:
     - we have to distinguish between the three different ellipsis forms
     - the is the ellipsis at the beginning, in the middle, or at the end.
  */
  endp = collate->cursor->next;
  assert (arg == NULL || endp != NULL);

  /* Both, the start and the end symbol, must stand for characters.  */
  if ((startp != NULL && (startp->name == NULL || ! startp->is_character))
      || (endp != NULL && (endp->name == NULL|| ! endp->is_character)))
    {
      lr_error (ldfile, _("\
%s: the start end the end symbol of a range must stand for characters"),
		"LC_COLLATE");
      return;
    }

  if (ellipsis == tok_ellipsis3)
    {
      /* One requirement we make here: the length of the byte
	 sequences for the first and end character must be the same.
	 This is mainly to prevent unwanted effects and this is often
	 not what is wanted.  */
      size_t len = (startp->mbs != NULL ? startp->nmbs
		    : (endp->mbs != NULL ? endp->nmbs : 0));
      char mbcnt[len + 1];
      char mbend[len + 1];

      /* Well, this should be caught somewhere else already.  Just to
	 make sure.  */
      assert (startp == NULL || startp->wcs == NULL || startp->wcs[1] == 0);
      assert (endp == NULL || endp->wcs == NULL || endp->wcs[1] == 0);

      if (startp != NULL && endp != NULL
	  && startp->mbs != NULL && endp->mbs != NULL
	  && startp->nmbs != endp->nmbs)
	{
	  lr_error (ldfile, _("\
%s: byte sequences of first and last character must have the same length"),
		    "LC_COLLATE");
	  return;
	}

      /* Determine whether we have to generate multibyte sequences.  */
      if ((startp == NULL || startp->mbs != NULL)
	  && (endp == NULL || endp->mbs != NULL))
	{
	  int cnt;
	  int ret;

	  /* Prepare the beginning byte sequence.  This is either from the
	     beginning byte sequence or it is all nulls if it was an
	     initial ellipsis.  */
	  if (startp == NULL || startp->mbs == NULL)
	    memset (mbcnt, '\0', len);
	  else
	    {
	      memcpy (mbcnt, startp->mbs, len);

	      /* And increment it so that the value is the first one we will
		 try to insert.  */
	      for (cnt = len - 1; cnt >= 0; --cnt)
		if (++mbcnt[cnt] != '\0')
		  break;
	    }
	  mbcnt[len] = '\0';

	  /* And the end sequence.  */
	  if (endp == NULL || endp->mbs == NULL)
	    memset (mbend, '\0', len);
	  else
	    memcpy (mbend, endp->mbs, len);
	  mbend[len] = '\0';

	  /* Test whether we have a correct range.  */
	  ret = memcmp (mbcnt, mbend, len);
	  if (ret >= 0)
	    {
	      if (ret > 0)
		lr_error (ldfile, _("%s: byte sequence of first character of \
sequence is not lower than that of the last character"), "LC_COLLATE");
	      return;
	    }

	  /* Generate the byte sequences data.  */
	  while (1)
	    {
	      struct charseq *seq;

	      /* Quite a bit of work ahead.  We have to find the character
		 definition for the byte sequence and then determine the
		 wide character belonging to it.  */
	      seq = charmap_find_symbol (charmap, mbcnt, len);
	      if (seq != NULL)
		{
		  struct element_t *elem;
		  size_t namelen;

		  /* I don't this this can ever happen.  */
		  assert (seq->name != NULL);
		  namelen = strlen (seq->name);

		  if (seq->ucs4 == UNINITIALIZED_CHAR_VALUE)
		    seq->ucs4 = repertoire_find_value (repertoire, seq->name,
						       namelen);

		  /* Now we are ready to insert the new value in the
		     sequence.  Find out whether the element is
		     already known.  */
		  if (find_entry (&collate->seq_table, seq->name, namelen,
				  (void **) &elem) != 0)
		    {
		      uint32_t wcs[2] = { seq->ucs4, 0 };

		      /* We have to allocate an entry.  */
		      elem = new_element (collate, mbcnt, len,
					  seq->ucs4 == ILLEGAL_CHAR_VALUE
					  ? NULL : wcs, seq->name,
					  namelen, 1);

		      /* And add it to the table.  */
		      if (insert_entry (&collate->seq_table, seq->name,
					namelen, elem) != 0)
			/* This cannot happen.  */
			assert (! "Internal error");
		    }

		  /* Test whether this element is not already in the list.  */
		  if (elem->next != NULL || (collate->cursor != NULL
					     && elem->next == collate->cursor))
		    {
		      lr_error (ldfile, _("\
order for `%.*s' already defined at %s:%zu"),
				(int) namelen, seq->name,
				elem->file, elem->line);
		      goto increment;
		    }

		  /* Enqueue the new element.  */
		  elem->last = collate->cursor;
		  if (collate->cursor == NULL)
		    elem->next = NULL;
		  else
		    {
		      elem->next = collate->cursor->next;
		      elem->last->next = elem;
		      if (elem->next != NULL)
			elem->next->last = elem;
		    }
		  if (collate->start == NULL)
		    {
		      assert (collate->cursor == NULL);
		      collate->start = elem;
		    }
		  collate->cursor = elem;

		 /* Add the weight value.  We take them from the
		    `ellipsis_weights' member of `collate'.  */
		  elem->weights = (struct element_list_t *)
		    obstack_alloc (&collate->mempool,
				   nrules * sizeof (struct element_list_t));
		  for (cnt = 0; cnt < nrules; ++cnt)
		    if (collate->ellipsis_weight.weights[cnt].cnt == 1
			&& (collate->ellipsis_weight.weights[cnt].w[0]
			    == ELEMENT_ELLIPSIS2))
		      {
			elem->weights[cnt].w = (struct element_t **)
			  obstack_alloc (&collate->mempool,
					 sizeof (struct element_t *));
			elem->weights[cnt].w[0] = elem;
			elem->weights[cnt].cnt = 1;
		      }
		    else
		      {
			/* Simply use the weight from `ellipsis_weight'.  */
			elem->weights[cnt].w =
			  collate->ellipsis_weight.weights[cnt].w;
			elem->weights[cnt].cnt =
			  collate->ellipsis_weight.weights[cnt].cnt;
		      }
		}

	      /* Increment for the next round.  */
	    increment:
	      for (cnt = len - 1; cnt >= 0; --cnt)
		if (++mbcnt[cnt] != '\0')
		  break;

	      /* Find out whether this was all.  */
	      if (cnt < 0 || memcmp (mbcnt, mbend, len) >= 0)
		/* Yep, that's all.  */
		break;
	    }
	}
    }
  else
    {
      /* For symbolic range we naturally must have a beginning and an
	 end specified by the user.  */
      if (startp == NULL)
	lr_error (ldfile, _("\
%s: symbolic range ellipsis must not directly follow `order_start'"),
		  "LC_COLLATE");
      else if (endp == NULL)
	lr_error (ldfile, _("\
%s: symbolic range ellipsis must not be direct followed by `order_end'"),
		  "LC_COLLATE");
      else
	{
	  /* Determine the range.  To do so we have to determine the
             common prefix of the both names and then the numeric
             values of both ends.  */
	  size_t lenfrom = strlen (startp->name);
	  size_t lento = strlen (endp->name);
	  char buf[lento + 1];
	  int preflen = 0;
	  long int from;
	  long int to;
	  char *cp;
	  int base = ellipsis == tok_ellipsis2 ? 16 : 10;

	  if (lenfrom != lento)
	    {
	    invalid_range:
	      lr_error (ldfile, _("\
`%s' and `%.*s' are no valid names for symbolic range"),
			startp->name, (int) lento, endp->name);
	      return;
	    }

	  while (startp->name[preflen] == endp->name[preflen])
	    if (startp->name[preflen] == '\0')
	      /* Nothing to be done.  The start and end point are identical
		 and while inserting the end point we have already given
		 the user an error message.  */
	      return;
	    else
	      ++preflen;

	  errno = 0;
	  from = strtol (startp->name + preflen, &cp, base);
	  if ((from == UINT_MAX && errno == ERANGE) || *cp != '\0')
	    goto invalid_range;

	  errno = 0;
	  to = strtol (endp->name + preflen, &cp, base);
	  if ((to == UINT_MAX && errno == ERANGE) || *cp != '\0')
	    goto invalid_range;

	  /* Copy the prefix.  */
	  memcpy (buf, startp->name, preflen);

	  /* Loop over all values.  */
	  for (++from; from < to; ++from)
	    {
	      struct element_t *elem = NULL;
	      struct charseq *seq;
	      uint32_t wc;
	      int cnt;

	      /* Generate the the name.  */
	      sprintf (buf + preflen, base == 10 ? "%d" : "%x", from);

	      /* Look whether this name is already defined.  */
	      if (find_entry (&collate->seq_table, arg->val.str.startmb,
			      arg->val.str.lenmb, (void **) &elem) == 0)
		{
		  if (elem->next != NULL || (collate->cursor != NULL
					     && elem->next == collate->cursor))
		    {
		      lr_error (ldfile, _("\
%s: order for `%.*s' already defined at %s:%zu"),
				"LC_COLLATE", (int) lenfrom, buf,
				elem->file, elem->line);
		      continue;
		    }

		  if (elem->name == NULL)
		    {
		      lr_error (ldfile, _("%s: `%s' must be a charater"),
				"LC_COLLATE", buf);
		      continue;
		    }
		}

	      if (elem == NULL || (elem->mbs == NULL && elem->wcs == NULL))
		{
		  /* Search for a character of this name.  */
		  seq = charmap_find_value (charmap, buf, lenfrom);
		  if (seq == NULL || seq->ucs4 == UNINITIALIZED_CHAR_VALUE)
		    {
		      wc = repertoire_find_value (repertoire, buf, lenfrom);

		      if (seq != NULL)
			seq->ucs4 = wc;
		    }
		  else
		    wc = seq->ucs4;

		  if (wc == ILLEGAL_CHAR_VALUE && seq == NULL)
		    /* We don't know anything about a character with this
		       name.  XXX Should we warn?  */
		    continue;

		  if (elem == NULL)
		    {
		      uint32_t wcs[2] = { wc, 0 };

		      /* We have to allocate an entry.  */
		      elem = new_element (collate,
					  seq != NULL ? seq->bytes : NULL,
					  seq != NULL ? seq->nbytes : 0,
					  wc == ILLEGAL_CHAR_VALUE
					  ? NULL : wcs, buf, lenfrom, 1);
		    }
		  else
		    {
		      /* Update the element.  */
		      if (seq != NULL)
			{
			  elem->mbs = obstack_copy0 (&collate->mempool,
						     seq->bytes, seq->nbytes);
			  elem->nmbs = seq->nbytes;
			}

		      if (wc != ILLEGAL_CHAR_VALUE)
			{
			  uint32_t zero = 0;

			  obstack_grow (&collate->mempool,
					&wc, sizeof (uint32_t));
			  obstack_grow (&collate->mempool,
					&zero, sizeof (uint32_t));
			  elem->wcs = obstack_finish (&collate->mempool);
			  elem->nwcs = 1;
			}
		    }

		  elem->file = ldfile->fname;
		  elem->line = ldfile->lineno;
		  elem->section = collate->current_section;
		}

	      /* Enqueue the new element.  */
	      elem->last = collate->cursor;
	      elem->next = collate->cursor->next;
	      elem->last->next = elem;
	      if (elem->next != NULL)
		elem->next->last = elem;
	      collate->cursor = elem;

	      /* Now add the weights.  They come from the `ellipsis_weights'
		 member of `collate'.  */
	      elem->weights = (struct element_list_t *)
		obstack_alloc (&collate->mempool,
			       nrules * sizeof (struct element_list_t));
	      for (cnt = 0; cnt < nrules; ++cnt)
		if (collate->ellipsis_weight.weights[cnt].cnt == 1
		    && (collate->ellipsis_weight.weights[cnt].w[0]
			== ELEMENT_ELLIPSIS2))
		  {
		    elem->weights[cnt].w = (struct element_t **)
		      obstack_alloc (&collate->mempool,
				     sizeof (struct element_t *));
		    elem->weights[cnt].w[0] = elem;
		    elem->weights[cnt].cnt = 1;
		  }
		else
		  {
		    /* Simly use the weight from `ellipsis_weight'.  */
		    elem->weights[cnt].w =
		      collate->ellipsis_weight.weights[cnt].w;
		    elem->weights[cnt].cnt =
		      collate->ellipsis_weight.weights[cnt].cnt;
		  }
	    }
	}
    }
}


static void
collate_startup (struct linereader *ldfile, struct localedef_t *locale,
		 struct localedef_t *copy_locale, int ignore_content)
{
  if (!ignore_content)
    {
      struct locale_collate_t *collate;

      if (copy_locale == NULL)
	{
	  collate = locale->categories[LC_COLLATE].collate =
	    (struct locale_collate_t *)
	    xcalloc (1, sizeof (struct locale_collate_t));

	  /* Init the various data structures.  */
	  init_hash (&collate->elem_table, 100);
	  init_hash (&collate->sym_table, 100);
	  init_hash (&collate->seq_table, 500);
	  obstack_init (&collate->mempool);

	  collate->col_weight_max = -1;
	}
      else
	collate = locale->categories[LC_COLLATE].collate =
	  copy_locale->categories[LC_COLLATE].collate;
    }

  ldfile->translate_strings = 0;
  ldfile->return_widestr = 0;
}


void
collate_finish (struct localedef_t *locale, struct charmap_t *charmap)
{
  /* Now is the time when we can assign the individual collation
     values for all the symbols.  We have possibly different values
     for the wide- and the multibyte-character symbols.  This is done
     since it might make a difference in the encoding if there is in
     some cases no multibyte-character but there are wide-characters.
     (The other way around it is not important since theencoded
     collation value in the wide-character case is 32 bits wide and
     therefore requires no encoding).

     The lowest collation value assigned is 2.  Zero is reserved for
     the NUL byte terminating the strings in the `strxfrm'/`wcsxfrm'
     functions and 1 is used to separate the individual passes for the
     different rules.

     We also have to construct is list with all the bytes/words which
     can come first in a sequence, followed by all the elements which
     also start with this byte/word.  The order is reverse which has
     among others the important effect that longer strings are located
     first in the list.  This is required for the output data since
     the algorithm used in `strcoll' etc depends on this.

     The multibyte case is easy.  We simply sort into an array with
     256 elements.  */
  struct locale_collate_t *collate = locale->categories[LC_COLLATE].collate;
  int mbact[nrules];
  int wcact;
  struct element_t *runp;
  int i;
  int need_undefined = 0;
  struct section_list *sect;
  int ruleidx;
  int nr_wide_elems = 0;
  size_t min_total;
  size_t act_size;

  if (collate == NULL)
    {
      /* No data, no check.  */
      if (! be_quiet)
	error (0, 0, _("No definition for %s category found"), "LC_COLLATE");
      return;
    }

  /* If this assertion is hit change the type in `element_t'.  */
  assert (nrules <= sizeof (runp->used_in_level) * 8);

  /* Make sure that the `position' rule is used either in all sections
     or in none.  */
  for (i = 0; i < nrules; ++i)
    for (sect = collate->sections; sect != NULL; sect = sect->next)
      if ((sect->rules[i] & sort_position)
	  != (collate->sections->rules[i] & sort_position))
	{
	  error (0, 0, _("\
%s: `position' must be used for a specific level in all sections or none"),
		 "LC_COLLATE");
	  break;
	}

  /* Find out which elements are used at which level.  At the same
     time we find out whether we have any undefined symbols.  */
  runp = collate->start;
  while (runp != NULL)
    {
      if (runp->mbs != NULL)
	{
	  for (i = 0; i < nrules; ++i)
	    {
	      int j;

	      for (j = 0; j < runp->weights[i].cnt; ++j)
		/* A NULL pointer as the weight means IGNORE.  */
		if (runp->weights[i].w[j] != NULL)
		  {
		    if (runp->weights[i].w[j]->weights == NULL)
		      {
			error_at_line (0, 0, runp->file, runp->line,
				       _("symbol `%s' not defined"),
				       runp->weights[i].w[j]->name);

			need_undefined = 1;
			runp->weights[i].w[j] = &collate->undefined;
		      }
		    else
		      /* Set the bit for the level.  */
		      runp->weights[i].w[j]->used_in_level |= 1 << i;
		  }
	    }
	}

      /* Up to the next entry.  */
      runp = runp->next;
    }

  /* Walk through the list of defined sequences and assign weights.  Also
     create the data structure which will allow generating the single byte
     character based tables.

     Since at each time only the weights for each of the rules are
     only compared to other weights for this rule it is possible to
     assign more compact weight values than simply counting all
     weights in sequence.  We can assign weights from 3, one for each
     rule individually and only for those elements, which are actually
     used for this rule.

     Why is this important?  It is not for the wide char table.  But
     it is for the singlebyte output since here larger numbers have to
     be encoded to make it possible to emit the value as a byte
     string.  */
  for (i = 0; i < nrules; ++i)
    mbact[i] = 2;
  wcact = 2;
  runp = collate->start;
  while (runp != NULL)
    {
      /* Determine the order.  */
      if (runp->used_in_level != 0)
	{
	  runp->mborder = (int *) obstack_alloc (&collate->mempool,
						 nrules * sizeof (int));

	  for (i = 0; i < nrules; ++i)
	    if ((runp->used_in_level & (1 << i)) != 0)
	      runp->mborder[i] = mbact[i]++;
	    else
	      runp->mborder[i] = 0;
	}

      if (runp->mbs != NULL)
	{
	  struct element_t **eptr;
	  struct element_t *lastp = NULL;

	  /* Find the point where to insert in the list.  */
	  eptr = &collate->mbheads[((unsigned char *) runp->mbs)[0]];
	  while (*eptr != NULL)
	    {
	      if ((*eptr)->nmbs < runp->nmbs)
		break;

	      if ((*eptr)->nmbs == runp->nmbs)
		{
		  int c = memcmp ((*eptr)->mbs, runp->mbs, runp->nmbs);

		  if (c == 0)
		    {
		      /* This should not happen.  It means that we have
			 to symbols with the same byte sequence.  It is
			 of course an error.  */
		      error_at_line (0, 0, (*eptr)->file, (*eptr)->line,
				     _("symbol `%s' has the same encoding as"),
				     (*eptr)->name);
		      error_at_line (0, 0, runp->file, runp->line,
				     _("symbol `%s'"), runp->name);
		      goto dont_insert;
		    }
		  else if (c < 0)
		    /* Insert it here.  */
		    break;
		}

	      /* To the next entry.  */
	      lastp = *eptr;
	      eptr = &(*eptr)->mbnext;
	    }

	  /* Set the pointers.  */
	  runp->mbnext = *eptr;
	  runp->mblast = lastp;
	  if (*eptr != NULL)
	    (*eptr)->mblast = runp;
	  *eptr = runp;
	dont_insert:
	}

      if (runp->wcs != NULL)
	{
	  runp->wcorder = wcact++;

	  /* We take the opportunity to count the elements which have
	     wide characters.  */
	  ++nr_wide_elems;
	}

      /* Up to the next entry.  */
      runp = runp->next;
    }

  /* Find out whether any of the `mbheads' entries is unset.  In this
     case we use the UNDEFINED entry.  */
  for (i = 1; i < 256; ++i)
    if (collate->mbheads[i] == NULL)
      {
	need_undefined = 1;
	collate->mbheads[i] = &collate->undefined;
      }

  /* Now to the wide character case.  Here we have to find first a good
     mapping function to get the wide range of wide character values
     (0x00000000 to 0x7fffffff) to a managable table.  This might take
     some time so we issue a warning.

     We use a very trivial hashing function to store the sparse
     table.  CH % TABSIZE is used as an index.  To solve multiple hits
     we have N planes.  This guarantees a fixed search time for a
     character [N / 2].  In the following code we determine the minimum
     value for TABSIZE * N, where TABSIZE >= 256.

     Some people complained that this algorithm takes too long.  Well,
     go on, improve it.  But changing the step size is *not* an
     option.  Some people changed this to use only sizes of prime
     numbers.  Think again, do some math.  We are looking for the
     optimal solution, not something which works in general.  Unless
     somebody can provide a dynamic programming solution I think this
     implementation is as good as it can get.  */
  if (nr_wide_elems > 512 && !be_quiet)
    fputs (_("\
Computing table size for collation table might take a while..."),
	   stderr);

  min_total = UINT_MAX;
  act_size = 256;

  /* While we want to have a small total size we are willing to use a
     little bit larger table if this reduces the number of layers.
     Therefore we add a little penalty to the number of planes.
     Maybe this constant has to be adjusted a bit.  */
#define PENALTY 128
  do
    {
      size_t cnt[act_size];
      struct element_t *elem[act_size];
      size_t act_planes = 1;

      memset (cnt, '\0', sizeof cnt);
      memset (elem, '\0', sizeof elem);

      runp = collate->start;
      while (runp != NULL)
	{
	  if (runp->wcs != NULL)
	    {
	      size_t nr = runp->wcs[0] % act_size;
	      struct element_t *elemp = elem[nr];

	      while (elemp != NULL)
		{
		  if (elemp->wcs[0] == runp->wcs[0])
		    break;
		  elemp = elemp->wcnext;
		}

	      if (elemp == NULL && ++cnt[nr] > act_planes)
		{
		  act_planes = cnt[nr];

		  runp->wcnext = elem[nr];
		  elem[nr] = runp;

		  if ((act_size + PENALTY) * act_planes >= min_total)
		    break;
		}
	    }

	  /* Up to the next entry.  */
	  runp = runp->next;
	}

      if ((act_size + PENALTY) * act_planes < min_total)
	{
	  min_total = (act_size + PENALTY) * act_planes;
	  collate->plane_size = act_size;
	  collate->plane_cnt = act_planes;
	}

      ++act_size;
    }
  while (act_size < min_total);

  if (nr_wide_elems > 512 && !be_quiet)
    fputs (_(" done\n"), stderr);

  /* Now that we know how large the table has to be we are able to
     allocate the array and start adding the characters to the lists
     in the same way we did it for the multibyte characters.  */
  collate->wcheads = (struct element_t **)
    obstack_alloc (&collate->mempool, (collate->plane_size
				       * collate->plane_cnt
				       * sizeof (struct element_t *)));
  memset (collate->wcheads, '\0', (collate->plane_size
				   * collate->plane_cnt
				   * sizeof (struct element_t *)));

  /* Start adding.  */
  runp = collate->start;
  while (runp != NULL)
    {
      if (runp->wcs != NULL)
	{
	  struct element_t **eptr;
	  struct element_t *lastp = NULL;
	  size_t idx;

	  /* Find a free index.  */
	  idx = runp->wcs[0] % collate->plane_size;
	  while (collate->wcheads[idx] != NULL)
	    {
	      /* Stop if this is an entry with the same starting character.  */
	      if (collate->wcheads[idx]->wcs[0] == runp->wcs[0])
		break;

	      idx += collate->plane_size;
	    }

	  /* Find the point where to insert in the list.  */
	  eptr = &collate->wcheads[idx];
	  while (*eptr != NULL)
	    {
	      if ((*eptr)->nwcs < runp->nwcs)
		break;

	      if ((*eptr)->nwcs == runp->nwcs)
		{
		  int c = wmemcmp ((wchar_t *) (*eptr)->wcs,
				   (wchar_t *) runp->wcs, runp->nwcs);

		  if (c == 0)
		    {
		      /* This should not happen.  It means that we have
			 to symbols with the same byte sequence.  It is
			 of course an error.  */
		      error_at_line (0, 0, (*eptr)->file, (*eptr)->line,
				     _("symbol `%s' has the same encoding as"),
				     (*eptr)->name);
		      error_at_line (0, 0, runp->file, runp->line,
				     _("symbol `%s'"), runp->name);
		      goto dont_insertwc;
		    }
		  else if (c < 0)
		    /* Insert it here.  */
		    break;
		}

	      /* To the next entry.  */
	      lastp = *eptr;
	      eptr = &(*eptr)->wcnext;
	    }

	  /* Set the pointers.  */
	  runp->wcnext = *eptr;
	  runp->wclast = lastp;
	  if (*eptr != NULL)
	    (*eptr)->wclast = runp;
	  *eptr = runp;
	dont_insertwc:
	}

      /* Up to the next entry.  */
      runp = runp->next;
    }

  /* Now determine whether the UNDEFINED entry is needed and if yes,
     whether it was defined.  */
  collate->undefined.used_in_level = need_undefined ? ~0ul : 0;
  if (collate->undefined.file == NULL)
    {
      if (need_undefined)
	{
	  error (0, 0, _("no definition of `UNDEFINED'"));

	  /* Add UNDEFINED at the end.  */
	  collate->undefined.mborder =
	    (int *) obstack_alloc (&collate->mempool, nrules * sizeof (int));

	  for (i = 0; i < nrules; ++i)
	    collate->undefined.mborder[i] = mbact[i]++;
	}

      /* In any case we will need the definition for the wide character
	 case.  But we will not complain that it is missing since the
	 specification strangely enough does not seem to account for
	 this.  */
      collate->undefined.wcorder = wcact++;
    }

  /* Finally, try to unify the rules for the sections.  Whenever the rules
     for a section are the same as those for another section give the
     ruleset the same index.  Since there are never many section we can
     use an O(n^2) algorithm here.  */
  sect = collate->sections;
  assert (sect != NULL);
  ruleidx = 0;
  do
    {
      struct section_list *osect = collate->sections;

      while (osect != sect)
	if (memcmp (osect->rules, sect->rules, nrules) == 0)
	  break;
	else
	  osect = osect->next;

      if (osect == sect)
	sect->ruleidx = ruleidx++;
      else
	sect->ruleidx = osect->ruleidx;

      /* Next section.  */
      sect = sect->next;
    }
  while (sect != NULL);
  /* We are currently not prepared for more than 256 rulesets.  But this
     should never really be a problem.  */
  assert (ruleidx <= 256);
}


static int32_t
output_weight (struct obstack *pool, struct locale_collate_t *collate,
	       struct element_t *elem)
{
  size_t cnt;
  int32_t retval;

  /* Optimize the use of UNDEFINED.  */
  if (elem == &collate->undefined)
    /* The weights are already inserted.  */
    return 0;

  /* This byte can start exactly one collation element and this is
     a single byte.  We can directly give the index to the weights.  */
  retval = obstack_object_size (pool);

  /* Construct the weight.  */
  for (cnt = 0; cnt < nrules; ++cnt)
    {
      char buf[elem->weights[cnt].cnt * 7];
      int len = 0;
      int i;

      for (i = 0; i < elem->weights[cnt].cnt; ++i)
	/* Encode the weight value.  We do nothing for IGNORE entries.  */
	if (elem->weights[cnt].w[i] != NULL)
	  len += utf8_encode (&buf[len],
			      elem->weights[cnt].w[i]->mborder[cnt]);

      /* And add the buffer content.  */
      obstack_1grow (pool, len);
      obstack_grow (pool, buf, len);
    }

  return retval | ((elem->section->ruleidx & 0x7f) << 24);
}


static int32_t
output_weightwc (struct obstack *pool, struct locale_collate_t *collate,
		 struct element_t *elem)
{
  size_t cnt;
  int32_t retval;

  /* Optimize the use of UNDEFINED.  */
  if (elem == &collate->undefined)
    /* The weights are already inserted.  */
    return 0;

  /* This byte can start exactly one collation element and this is
     a single byte.  We can directly give the index to the weights.  */
  retval = obstack_object_size (pool) / sizeof (int32_t);

  /* Construct the weight.  */
  for (cnt = 0; cnt < nrules; ++cnt)
    {
      int32_t buf[elem->weights[cnt].cnt];
      int i;
      int32_t j;

      for (i = 0, j = 0; i < elem->weights[cnt].cnt; ++i)
	if (elem->weights[cnt].w[i] != NULL)
	  buf[j++] = elem->weights[cnt].w[i]->wcorder;

      /* And add the buffer content.  */
      if (sizeof (int) == sizeof (int32_t))
	obstack_int_grow (pool, j);
      else
	obstack_grow (pool, &j, sizeof (int32_t));

      obstack_grow (pool, buf, j * sizeof (int32_t));
    }

  return retval | ((elem->section->ruleidx & 0x7f) << 24);
}


void
collate_output (struct localedef_t *locale, struct charmap_t *charmap,
		const char *output_path)
{
  struct locale_collate_t *collate = locale->categories[LC_COLLATE].collate;
  const size_t nelems = _NL_ITEM_INDEX (_NL_NUM_LC_COLLATE);
  struct iovec iov[2 + nelems];
  struct locale_file data;
  uint32_t idx[nelems];
  size_t cnt;
  size_t ch;
  int32_t tablemb[256];
  struct obstack weightpool;
  struct obstack extrapool;
  struct obstack indirectpool;
  struct section_list *sect;
  uint32_t *names;
  uint32_t *tablewc;
  size_t table_size;
  uint32_t elem_size;
  uint32_t *elem_table;
  int i;
  struct element_t *runp;

  data.magic = LIMAGIC (LC_COLLATE);
  data.n = nelems;
  iov[0].iov_base = (void *) &data;
  iov[0].iov_len = sizeof (data);

  iov[1].iov_base = (void *) idx;
  iov[1].iov_len = sizeof (idx);

  idx[0] = iov[0].iov_len + iov[1].iov_len;
  cnt = 0;

  assert (cnt == _NL_ITEM_INDEX (_NL_COLLATE_NRULES));
  iov[2 + cnt].iov_base = &nrules;
  iov[2 + cnt].iov_len = sizeof (uint32_t);
  idx[1 + cnt] = idx[cnt] + iov[2 + cnt].iov_len;
  ++cnt;

  /* If we have no LC_COLLATE data emit only the number of rules as zero.  */
  if (collate == NULL)
    {
      int32_t dummy = 0;

      while (cnt < _NL_ITEM_INDEX (_NL_NUM_LC_COLLATE))
	{
	  /* The words have to be handled specially.  */
	  if (cnt == _NL_ITEM_INDEX (_NL_COLLATE_HASH_SIZE)
	      || cnt == _NL_ITEM_INDEX (_NL_COLLATE_HASH_LAYERS)
	      || cnt == _NL_ITEM_INDEX (_NL_COLLATE_SYMB_HASH_SIZEMB))
	    {
	      iov[2 + cnt].iov_base = &dummy;
	      iov[2 + cnt].iov_len = sizeof (int32_t);
	    }
	  else
	    {
	      iov[2 + cnt].iov_base = (char *) "";
	      iov[2 + cnt].iov_len = 0;
	    }

	  if (cnt + 1 < _NL_ITEM_INDEX (_NL_NUM_LC_COLLATE))
	    idx[1 + cnt] = idx[cnt] + iov[2 + cnt].iov_len;
	  ++cnt;
	}

      assert (cnt == _NL_ITEM_INDEX (_NL_NUM_LC_COLLATE));

      write_locale_data (output_path, "LC_COLLATE", 2 + cnt, iov);

      return;
    }

  obstack_init (&weightpool);
  obstack_init (&extrapool);
  obstack_init (&indirectpool);

  /* Since we are using the sign of an integer to mark indirection the
     offsets in the arrays we are indirectly referring to must not be
     zero since -0 == 0.  Therefore we add a bit of dummy content.  */
  if (sizeof (int) == sizeof (int32_t))
    {
      obstack_int_grow (&extrapool, 0);
      obstack_int_grow (&indirectpool, 0);
    }
  else
    {
      int32_t zero = 0;
      obstack_grow (&extrapool, &zero, sizeof (zero));
      obstack_grow (&indirectpool, &zero, sizeof (zero));
    }

  /* Prepare the ruleset table.  */
  for (sect = collate->sections, i = 0; sect != NULL; sect = sect->next)
    if (sect->ruleidx == i)
      {
	int j;

	obstack_make_room (&weightpool, nrules);

	for (j = 0; j < nrules; ++j)
	  obstack_1grow_fast (&weightpool, sect->rules[j]);
	++i;
      }
  /* And align the output.  */
  i = (nrules * i) % __alignof__ (int32_t);
  if (i > 0)
    do
      obstack_1grow (&weightpool, '\0');
    while (++i < __alignof__ (int32_t));

  assert (cnt == _NL_ITEM_INDEX (_NL_COLLATE_RULESETS));
  iov[2 + cnt].iov_len = obstack_object_size (&weightpool);
  iov[2 + cnt].iov_base = obstack_finish (&weightpool);
  idx[1 + cnt] = idx[cnt] + iov[2 + cnt].iov_len;
  ++cnt;

  /* Generate the 8-bit table.  Walk through the lists of sequences
     starting with the same byte and add them one after the other to
     the table.  In case we have more than one sequence starting with
     the same byte we have to use extra indirection.

     First add a record for the NUL byte.  This entry will never be used
     so it does not matter.  */
  tablemb[0] = 0;

  /* Now insert the `UNDEFINED' value if it is used.  Since this value
     will probably be used more than once it is good to store the
     weights only once.  */
  if (collate->undefined.used_in_level != 0)
    output_weight (&weightpool, collate, &collate->undefined);

  for (ch = 1; ch < 256; ++ch)
    if (collate->mbheads[ch]->mbnext == NULL
	&& collate->mbheads[ch]->nmbs <= 1)
      {
	tablemb[ch] = output_weight (&weightpool, collate,
				     collate->mbheads[ch]);
      }
    else
      {
	/* The entries in the list are sorted by length and then
           alphabetically.  This is the order in which we will add the
           elements to the collation table.  This allows to simply
           walk the table in sequence and stop at the first matching
           entry.  Since the longer sequences are coming first in the
           list they have the possibility to match first, just as it
           has to be.  In the worst case we are walking to the end of
           the list where we put, if no singlebyte sequence is defined
           in the locale definition, the weights for UNDEFINED.

	   To reduce the length of the search list we compress them a bit.
	   This happens by collecting sequences of consecutive byte
	   sequences in one entry (having and begin and end byte sequence)
	   and add only one index into the weight table.  We can find the
	   consecutive entries since they are also consecutive in the list.  */
	struct element_t *runp = collate->mbheads[ch];
	struct element_t *lastp;

	assert ((obstack_object_size (&extrapool)
		 & (__alignof__ (int32_t) - 1)) == 0);

	tablemb[ch] = -obstack_object_size (&extrapool);

	do
	  {
	    /* Store the current index in the weight table.  We know that
	       the current position in the `extrapool' is aligned on a
	       32-bit address.  */
	    int32_t weightidx;
	    int added;

	    /* Find out wether this is a single entry or we have more than
	       one consecutive entry.  */
	    if (runp->mbnext != NULL
		&& runp->nmbs == runp->mbnext->nmbs
		&& memcmp (runp->mbs, runp->mbnext->mbs, runp->nmbs - 1) == 0
		&& (runp->mbs[runp->nmbs - 1]
		    == runp->mbnext->mbs[runp->nmbs - 1] + 1))
	      {
		int i;
		struct element_t *series_startp = runp;
		struct element_t *curp;

		/* Compute how much space we will need.  */
		added = ((sizeof (int32_t) + 1 + 2 * (runp->nmbs - 1)
			  + __alignof__ (int32_t) - 1)
			 & ~(__alignof__ (int32_t) - 1));
		assert ((obstack_object_size (&extrapool)
			 & (__alignof__ (int32_t) - 1)) == 0);
		obstack_make_room (&extrapool, added);

		/* More than one consecutive entry.  We mark this by having
		   a negative index into the indirect table.  */
		if (sizeof (int32_t) == sizeof (int))
		  obstack_int_grow_fast (&extrapool,
					 -(obstack_object_size (&indirectpool)
					   / sizeof (int32_t)));
		else
		  {
		    int32_t i = -(obstack_object_size (&indirectpool)
				  / sizeof (int32_t));
		    obstack_grow (&extrapool, &i, sizeof (int32_t));
		  }

		/* Now search first the end of the series.  */
		do
		  runp = runp->mbnext;
		while (runp->mbnext != NULL
		       && runp->nmbs == runp->mbnext->nmbs
		       && memcmp (runp->mbs, runp->mbnext->mbs,
				  runp->nmbs - 1) == 0
		       && (runp->mbs[runp->nmbs - 1]
			   == runp->mbnext->mbs[runp->nmbs - 1] + 1));

		/* Now walk backward from here to the beginning.  */
		curp = runp;

		assert (runp->nmbs <= 256);
		obstack_1grow_fast (&extrapool, curp->nmbs - 1);
		for (i = 1; i < curp->nmbs; ++i)
		  obstack_1grow_fast (&extrapool, curp->mbs[i]);

		/* Now find the end of the consecutive sequence and
                   add all the indeces in the indirect pool.  */
		do
		  {
		    weightidx = output_weight (&weightpool, collate, curp);
		    if (sizeof (int32_t) == sizeof (int))
		      obstack_int_grow (&indirectpool, weightidx);
		    else
		      obstack_grow (&indirectpool, &weightidx,
				    sizeof (int32_t));

		    curp = curp->mblast;
		  }
		while (curp != series_startp);

		/* Add the final weight.  */
		weightidx = output_weight (&weightpool, collate, curp);
		if (sizeof (int32_t) == sizeof (int))
		  obstack_int_grow (&indirectpool, weightidx);
		else
		  obstack_grow (&indirectpool, &weightidx, sizeof (int32_t));

		/* And add the end byte sequence.  Without length this
                   time.  */
		for (i = 1; i < curp->nmbs; ++i)
		  obstack_1grow_fast (&extrapool, curp->mbs[i]);
	      }
	    else
	      {
		/* A single entry.  Simply add the index and the length and
		   string (except for the first character which is already
		   tested for).  */
		int i;

		/* Output the weight info.  */
		weightidx = output_weight (&weightpool, collate, runp);

		added = ((sizeof (int32_t) + 1 + runp->nmbs - 1
			  + __alignof__ (int32_t) - 1)
			 & ~(__alignof__ (int32_t) - 1));
		assert ((obstack_object_size (&extrapool)
			 & (__alignof__ (int32_t) - 1)) == 0);
		obstack_make_room (&extrapool, added);

		if (sizeof (int32_t) == sizeof (int))
		  obstack_int_grow_fast (&extrapool, weightidx);
		else
		  obstack_grow (&extrapool, &weightidx, sizeof (int32_t));
		assert (runp->nmbs <= 256);
		obstack_1grow_fast (&extrapool, runp->nmbs - 1);

		for (i = 1; i < runp->nmbs; ++i)
		  obstack_1grow_fast (&extrapool, runp->mbs[i]);
	      }

	    /* Add alignment bytes if necessary.  */
	    while ((obstack_object_size (&extrapool)
		    & (__alignof__ (int32_t) - 1)) != 0)
	      obstack_1grow_fast (&extrapool, '\0');

	    /* Next entry.  */
	    lastp = runp;
	    runp = runp->mbnext;
	  }
	while (runp != NULL);

	assert ((obstack_object_size (&extrapool)
		 & (__alignof__ (int32_t) - 1)) == 0);

	/* If the final entry in the list is not a single character we
           add an UNDEFINED entry here.  */
	if (lastp->nmbs != 1)
	  {
	    int added = ((sizeof (int32_t) + 1 + 1 + __alignof__ (int32_t) - 1)
			 & ~(__alignof__ (int32_t) - 1));
	    obstack_make_room (&extrapool, added);

	    if (sizeof (int32_t) == sizeof (int))
	      obstack_int_grow_fast (&extrapool, 0);
	    else
	      {
		int32_t zero = 0;
		obstack_grow (&extrapool, &zero, sizeof (int32_t));
	      }
	    /* XXX What rule? We just pick the first.  */
	    obstack_1grow_fast (&extrapool, 0);
	    /* Length is zero.  */
	    obstack_1grow_fast (&extrapool, 0);

	    /* Add alignment bytes if necessary.  */
	    while ((obstack_object_size (&extrapool)
		    & (__alignof__ (int32_t) - 1)) != 0)
	      obstack_1grow_fast (&extrapool, '\0');
	  }
      }

  /* Add padding to the tables if necessary.  */
  while ((obstack_object_size (&weightpool) & (__alignof__ (int32_t) - 1))
	 != 0)
    obstack_1grow (&weightpool, 0);

  /* Now add the four tables.  */
  assert (cnt == _NL_ITEM_INDEX (_NL_COLLATE_TABLEMB));
  iov[2 + cnt].iov_base = tablemb;
  iov[2 + cnt].iov_len = sizeof (tablemb);
  idx[1 + cnt] = idx[cnt] + iov[2 + cnt].iov_len;
  assert ((iov[2 + cnt].iov_len & (__alignof__ (int32_t) - 1)) == 0);
  ++cnt;

  assert (cnt == _NL_ITEM_INDEX (_NL_COLLATE_WEIGHTMB));
  iov[2 + cnt].iov_len = obstack_object_size (&weightpool);
  iov[2 + cnt].iov_base = obstack_finish (&weightpool);
  idx[1 + cnt] = idx[cnt] + iov[2 + cnt].iov_len;
  ++cnt;

  assert (cnt == _NL_ITEM_INDEX (_NL_COLLATE_EXTRAMB));
  iov[2 + cnt].iov_len = obstack_object_size (&extrapool);
  iov[2 + cnt].iov_base = obstack_finish (&extrapool);
  idx[1 + cnt] = idx[cnt] + iov[2 + cnt].iov_len;
  ++cnt;

  assert (cnt == _NL_ITEM_INDEX (_NL_COLLATE_INDIRECTMB));
  iov[2 + cnt].iov_len = obstack_object_size (&indirectpool);
  iov[2 + cnt].iov_base = obstack_finish (&indirectpool);
  idx[1 + cnt] = idx[cnt] + iov[2 + cnt].iov_len;
  assert ((iov[2 + cnt].iov_len & (__alignof__ (int32_t) - 1)) == 0);
  ++cnt;


  /* Now the same for the wide character table.  We need to store some
     more information here.  */
  assert (cnt == _NL_ITEM_INDEX (_NL_COLLATE_HASH_SIZE));
  iov[2 + cnt].iov_base = &collate->plane_size;
  iov[2 + cnt].iov_len = sizeof (collate->plane_size);
  idx[1 + cnt] = idx[cnt] + iov[2 + cnt].iov_len;
  ++cnt;

  assert (cnt == _NL_ITEM_INDEX (_NL_COLLATE_HASH_LAYERS));
  iov[2 + cnt].iov_base = &collate->plane_cnt;
  iov[2 + cnt].iov_len = sizeof (collate->plane_cnt);
  idx[1 + cnt] = idx[cnt] + iov[2 + cnt].iov_len;
  ++cnt;

  /* Construct a table with the names.  The size of the table is the same
     as the table with the pointers.  */
  table_size = collate->plane_size * collate->plane_cnt;
  names = (uint32_t *) alloca (table_size * sizeof (uint32_t));
  for (ch = 0; ch < table_size; ++ch)
    if (collate->wcheads[ch] == NULL)
      names[ch] = 0;
    else
      names[ch] = collate->wcheads[ch]->wcs[0];

  assert (cnt == _NL_ITEM_INDEX (_NL_COLLATE_NAMES));
  iov[2 + cnt].iov_base = names;
  iov[2 + cnt].iov_len = table_size * sizeof (uint32_t);
  idx[1 + cnt] = idx[cnt] + iov[2 + cnt].iov_len;
  ++cnt;

  /* Since we are using the sign of an integer to mark indirection the
     offsets in the arrays we are indirectly referring to must not be
     zero since -0 == 0.  Therefore we add a bit of dummy content.  */
  if (sizeof (int) == sizeof (int32_t))
    {
      obstack_int_grow (&extrapool, 0);
      obstack_int_grow (&indirectpool, 0);
    }
  else
    {
      int32_t zero = 0;
      obstack_grow (&extrapool, &zero, sizeof (zero));
      obstack_grow (&indirectpool, &zero, sizeof (zero));
    }

  /* Now insert the `UNDEFINED' value if it is used.  Since this value
     will probably be used more than once it is good to store the
     weights only once.  */
  if (output_weightwc (&weightpool, collate, &collate->undefined) != 0)
    abort ();

  /* Generate the table.  Walk through the lists of sequences starting
     with the same wide character and add them one after the other to
     the table.  In case we have more than one sequence starting with
     the same byte we have to use extra indirection.  */
  tablewc = (uint32_t *) alloca (table_size * sizeof (uint32_t));
  for (ch = 0; ch < table_size; ++ch)
    if (collate->wcheads[ch] == NULL)
      {
	/* Set the entry to zero.  */
	tablewc[ch] = 0;
      }
    else if (collate->wcheads[ch]->wcnext == NULL
	     && collate->wcheads[ch]->nwcs == 1)
      {
	tablewc[ch] = output_weightwc (&weightpool, collate,
				       collate->wcheads[ch]);
      }
    else
      {
	/* As for the singlebyte table, we recognize sequences and
	   compress them.  */
	struct element_t *runp = collate->wcheads[ch];
	struct element_t *lastp;

	tablewc[ch] = -(obstack_object_size (&extrapool) / sizeof (uint32_t));

	do
	  {
	    /* Store the current index in the weight table.  We know that
	       the current position in the `extrapool' is aligned on a
	       32-bit address.  */
	    int32_t weightidx;
	    int added;

	    /* Find out wether this is a single entry or we have more than
	       one consecutive entry.  */
	    if (runp->wcnext != NULL
		&& runp->nwcs == runp->wcnext->nwcs
		&& wmemcmp ((wchar_t *) runp->wcs,
			    (wchar_t *)runp->wcnext->wcs, runp->nwcs - 1) == 0
		&& (runp->wcs[runp->nwcs - 1]
		    == runp->wcnext->wcs[runp->nwcs - 1] + 1))
	      {
		int i;
		struct element_t *series_startp = runp;
		struct element_t *curp;

		/* Now add first the initial byte sequence.  */
		added = (1 + 1 + 2 * (runp->nwcs - 1)) * sizeof (int32_t);
		if (sizeof (int32_t) == sizeof (int))
		  obstack_make_room (&extrapool, added);

		/* More than one consecutive entry.  We mark this by having
		   a negative index into the indirect table.  */
		if (sizeof (int32_t) == sizeof (int))
		  {
		    obstack_int_grow_fast (&extrapool,
					   obstack_object_size (&indirectpool)
					   / sizeof (int32_t));
		    obstack_int_grow_fast (&extrapool, runp->nwcs - 1);
		  }
		else
		  {
		    int32_t i = (obstack_object_size (&indirectpool)
				 / sizeof (int32_t));
		    obstack_grow (&extrapool, &i, sizeof (int32_t));
		    i = runp->nwcs - 1;
		    obstack_grow (&extrapool, &i, sizeof (int32_t));
		  }

		do
		  runp = runp->wcnext;
		while (runp->wcnext != NULL
		       && runp->nwcs == runp->wcnext->nwcs
		       && wmemcmp ((wchar_t *) runp->wcs,
				   (wchar_t *)runp->wcnext->wcs,
				   runp->nwcs - 1) == 0
		       && (runp->wcs[runp->nwcs - 1]
			   == runp->wcnext->wcs[runp->nwcs - 1] + 1));

		/* Now walk backward from here to the beginning.  */
		curp = runp;

		for (i = 1; i < runp->nwcs; ++i)
		  if (sizeof (int32_t) == sizeof (int))
		    obstack_int_grow_fast (&extrapool, curp->wcs[i]);
		  else
		    obstack_grow (&extrapool, &curp->wcs[i], sizeof (int32_t));

		/* Now find the end of the consecutive sequence and
                   add all the indeces in the indirect pool.  */
		do
		  {
		    weightidx = output_weightwc (&weightpool, collate, curp);
		    if (sizeof (int32_t) == sizeof (int))
		      obstack_int_grow (&extrapool, weightidx);
		    else
		      obstack_grow (&extrapool, &weightidx, sizeof (int32_t));

		    curp = curp->wclast;
		  }
		while (curp != series_startp);

		/* Add the final weight.  */
		weightidx = output_weightwc (&weightpool, collate, curp);
		if (sizeof (int32_t) == sizeof (int))
		  obstack_int_grow (&indirectpool, weightidx);
		else
		  obstack_grow (&indirectpool, &weightidx, sizeof (int32_t));

		/* And add the end byte sequence.  Without length this
                   time.  */
		for (i = 1; i < runp->nwcs; ++i)
		  if (sizeof (int32_t) == sizeof (int))
		    obstack_int_grow (&extrapool, runp->wcs[i]);
		  else
		    obstack_grow (&extrapool, &runp->wcs[i], sizeof (int32_t));
	      }
	    else
	      {
		/* A single entry.  Simply add the index and the length and
		   string (except for the first character which is already
		   tested for).  */
		int i;

		/* Output the weight info.  */
		weightidx = output_weightwc (&weightpool, collate, runp);

		added = (1 + 1 + runp->nwcs - 1) * sizeof (int32_t);
		if (sizeof (int) == sizeof (int32_t))
		  obstack_make_room (&extrapool, added);

		if (sizeof (int32_t) == sizeof (int))
		  {
		    obstack_int_grow_fast (&extrapool, weightidx);
		    obstack_int_grow_fast (&extrapool, runp->nwcs - 1);
		  }
		else
		  {
		    int32_t l = runp->nwcs - 1;
		    obstack_grow (&extrapool, &weightidx, sizeof (int32_t));
		    obstack_grow (&extrapool, &l, sizeof (int32_t));
		  }
		for (i = 1; i < runp->nwcs; ++i)
		  if (sizeof (int32_t) == sizeof (int))
		    obstack_int_grow_fast (&extrapool, runp->wcs[i]);
		  else
		    obstack_grow (&extrapool, &runp->wcs[i], sizeof (int32_t));
	      }

	    /* Next entry.  */
	    lastp = runp;
	    runp = runp->wcnext;
	  }
	while (runp != NULL);
      }

  /* Now add the four tables.  */
  assert (cnt == _NL_ITEM_INDEX (_NL_COLLATE_TABLEWC));
  iov[2 + cnt].iov_base = tablewc;
  iov[2 + cnt].iov_len = table_size * sizeof (uint32_t);
  idx[1 + cnt] = idx[cnt] + iov[2 + cnt].iov_len;
  assert (iov[2 + cnt].iov_len % sizeof (int32_t) == 0);
  ++cnt;

  assert (cnt == _NL_ITEM_INDEX (_NL_COLLATE_WEIGHTWC));
  iov[2 + cnt].iov_len = obstack_object_size (&weightpool);
  iov[2 + cnt].iov_base = obstack_finish (&weightpool);
  idx[1 + cnt] = idx[cnt] + iov[2 + cnt].iov_len;
  assert (iov[2 + cnt].iov_len % sizeof (int32_t) == 0);
  ++cnt;

  assert (cnt == _NL_ITEM_INDEX (_NL_COLLATE_EXTRAWC));
  iov[2 + cnt].iov_len = obstack_object_size (&extrapool);
  iov[2 + cnt].iov_base = obstack_finish (&extrapool);
  idx[1 + cnt] = idx[cnt] + iov[2 + cnt].iov_len;
  assert (iov[2 + cnt].iov_len % sizeof (int32_t) == 0);
  assert (iov[2 + cnt].iov_len % sizeof (int32_t) == 0);
  ++cnt;

  assert (cnt == _NL_ITEM_INDEX (_NL_COLLATE_INDIRECTWC));
  iov[2 + cnt].iov_len = obstack_object_size (&indirectpool);
  iov[2 + cnt].iov_base = obstack_finish (&indirectpool);
  idx[1 + cnt] = idx[cnt] + iov[2 + cnt].iov_len;
  assert (iov[2 + cnt].iov_len % sizeof (int32_t) == 0);
  ++cnt;


  /* Finally write the table with collation element names out.  It is
     a hash table with a simple function which gets the name of the
     character as the input.  One character might have many names.  The
     value associated with the name is an index into the weight table
     where we are then interested in the first-level weight value.

     To determine how large the table should be we are counting the
     elements have to put in.  Since we are using internal chaining
     using a secondary hash function we have to make the table a bit
     larger to avoid extremely long search times.  We can achieve
     good results with a 40% larger table than there are entries.  */
  elem_size = 0;
  runp = collate->start;
  while (runp != NULL)
    {
      if (runp->mbs != NULL && runp->weights != NULL)
	/* Yep, the element really counts.  */
	++elem_size;

      runp = runp->next;
    }
  /* Add 40% and find the next prime number.  */
  elem_size = MIN (next_prime (elem_size * 1.4), 257);

  /* Allocate the table.  Each entry consists of two words: the hash
     value and an index in a secondary table which provides the index
     into the weight table and the string itself (so that a match can
     be determined).  */
  elem_table = (uint32_t *) obstack_alloc (&extrapool,
					   elem_size * 2 * sizeof (uint32_t));
  memset (elem_table, '\0', elem_size * 2 * sizeof (uint32_t));

  /* Now add the elements.  */
  runp = collate->start;
  while (runp != NULL)
    {
      if (runp->mbs != NULL && runp->weights != NULL)
	{
	  /* Compute the hash value of the name.  */
	  uint32_t namelen = strlen (runp->name);
	  uint32_t hash = elem_hash (runp->name, namelen);
	  size_t idx = hash % elem_size;

	  if (elem_table[idx * 2] != 0)
	    {
	      /* The spot is already take.  Try iterating using the value
		 from the secondary hashing function.  */
	      size_t iter = hash % (elem_size - 2);

	      do
		{
		  idx += iter;
		  if (idx >= elem_size)
		    idx -= elem_size;
		}
	      while (elem_table[idx * 2] != 0);

	      /* This is the spot where we will insert the value.  */
	      elem_table[idx * 2] = hash;
	      elem_table[idx * 2 + 1] = obstack_object_size (&extrapool);

	      /* The the string itself including length.  */
	      obstack_1grow (&extrapool, namelen);
	      obstack_grow (&extrapool, runp->name, namelen);

	      /* And the multibyte representation.  */
	      obstack_1grow (&extrapool, runp->nmbs);
	      obstack_grow (&extrapool, runp->mbs, runp->nmbs);

	      /* And align again to 32 bits.  */
	      if ((1 + namelen + 1 + runp->nmbs) % sizeof (int32_t) != 0)
		obstack_grow (&extrapool, "\0\0",
			      (sizeof (int32_t)
			       - ((1 + namelen + 1 + runp->nmbs)
				  % sizeof (int32_t))));
	    }
	}

      runp = runp->next;
    }

  /* Prepare to write out this data.  */
  assert (cnt == _NL_ITEM_INDEX (_NL_COLLATE_SYMB_HASH_SIZEMB));
  iov[2 + cnt].iov_base = &elem_size;
  iov[2 + cnt].iov_len = sizeof (int32_t);
  idx[1 + cnt] = idx[cnt] + iov[2 + cnt].iov_len;
  ++cnt;

  assert (cnt == _NL_ITEM_INDEX (_NL_COLLATE_SYMB_TABLEMB));
  iov[2 + cnt].iov_base = elem_table;
  iov[2 + cnt].iov_len = elem_size * 2 * sizeof (int32_t);
  idx[1 + cnt] = idx[cnt] + iov[2 + cnt].iov_len;
  ++cnt;

  assert (cnt == _NL_ITEM_INDEX (_NL_COLLATE_SYMB_EXTRAMB));
  iov[2 + cnt].iov_len = obstack_object_size (&extrapool);
  iov[2 + cnt].iov_base = obstack_finish (&extrapool);
  ++cnt;


  assert (cnt == _NL_ITEM_INDEX (_NL_NUM_LC_COLLATE));

  write_locale_data (output_path, "LC_COLLATE", 2 + cnt, iov);

  obstack_free (&weightpool, NULL);
  obstack_free (&extrapool, NULL);
  obstack_free (&indirectpool, NULL);
}


void
collate_read (struct linereader *ldfile, struct localedef_t *result,
	      struct charmap_t *charmap, const char *repertoire_name,
	      int ignore_content)
{
  struct repertoire_t *repertoire = NULL;
  struct locale_collate_t *collate;
  struct token *now;
  struct token *arg = NULL;
  enum token_t nowtok;
  int state = 0;
  enum token_t was_ellipsis = tok_none;
  struct localedef_t *copy_locale = NULL;

  /* Get the repertoire we have to use.  */
  if (repertoire_name != NULL)
    repertoire = repertoire_read (repertoire_name);

  /* The rest of the line containing `LC_COLLATE' must be free.  */
  lr_ignore_rest (ldfile, 1);

  do
    {
      now = lr_token (ldfile, charmap, NULL);
      nowtok = now->tok;
    }
  while (nowtok == tok_eol);

  if (nowtok == tok_copy)
    {
      state = 2;
      now = lr_token (ldfile, charmap, NULL);
      if (now->tok != tok_string)
	{
	  SYNTAX_ERROR (_("%s: syntax error"), "LC_COLLATE");

	skip_category:
	  do
	    now = lr_token (ldfile, charmap, NULL);
	  while (now->tok != tok_eof && now->tok != tok_end);

	  if (now->tok != tok_eof
	      || (now = lr_token (ldfile, charmap, NULL), now->tok == tok_eof))
	    lr_error (ldfile, _("%s: premature end of file"), "LC_COLLATE");
	  else if (now->tok != tok_lc_collate)
	    {
	      lr_error (ldfile, _("\
%1$s: definition does not end with `END %1$s'"), "LC_COLLATE");
	      lr_ignore_rest (ldfile, 0);
	    }
	  else
	    lr_ignore_rest (ldfile, 1);

	  return;
	}

      /* Get the locale definition.  */
      copy_locale = load_locale (LC_COLLATE, now->val.str.startmb,
				 repertoire_name, charmap);
      if ((copy_locale->avail & COLLATE_LOCALE) == 0)
	{
	  /* Not yet loaded.  So do it now.  */
	  if (locfile_read (copy_locale, charmap) != 0)
	    goto skip_category;
	}

      lr_ignore_rest (ldfile, 1);

      now = lr_token (ldfile, charmap, NULL);
      nowtok = now->tok;
    }

  /* Prepare the data structures.  */
  collate_startup (ldfile, result, copy_locale, ignore_content);
  collate = result->categories[LC_COLLATE].collate;

  while (1)
    {
      /* Of course we don't proceed beyond the end of file.  */
      if (nowtok == tok_eof)
	break;

      /* Ingore empty lines.  */
      if (nowtok == tok_eol)
	{
	  now = lr_token (ldfile, charmap, NULL);
	  nowtok = now->tok;
	  continue;
	}

      switch (nowtok)
	{
	case tok_coll_weight_max:
	  /* Ignore the rest of the line if we don't need the input of
	     this line.  */
	  if (ignore_content)
	    {
	      lr_ignore_rest (ldfile, 0);
	      break;
	    }

	  if (state != 0)
	    goto err_label;

	  arg = lr_token (ldfile, charmap, NULL);
	  if (arg->tok != tok_number)
	    goto err_label;
	  if (collate->col_weight_max != -1)
	    lr_error (ldfile, _("%s: duplicate definition of `%s'"),
		      "LC_COLLATE", "col_weight_max");
	  else
	    collate->col_weight_max = arg->val.num;
	  lr_ignore_rest (ldfile, 1);
	  break;

	case tok_section_symbol:
	  /* Ignore the rest of the line if we don't need the input of
	     this line.  */
	  if (ignore_content)
	    {
	      lr_ignore_rest (ldfile, 0);
	      break;
	    }

	  if (state != 0)
	    goto err_label;

	  arg = lr_token (ldfile, charmap, repertoire);
	  if (arg->tok != tok_bsymbol)
	    goto err_label;
	  else if (!ignore_content)
	    {
	      /* Check whether this section is already known.  */
	      struct section_list *known = collate->sections;
	      while (known != NULL)
		if (strcmp (known->name, arg->val.str.startmb) == 0)
		  break;

	      if (known != NULL)
		{
		  lr_error (ldfile,
			    _("%s: duplicate declaration of section `%s'"),
			    "LC_COLLATE", arg->val.str.startmb);
		  free (arg->val.str.startmb);
		}
	      else
		collate->sections = make_seclist_elem (collate,
						       arg->val.str.startmb,
						       collate->sections);

	      lr_ignore_rest (ldfile, known == NULL);
	    }
	  else
	    {
	      free (arg->val.str.startmb);
	      lr_ignore_rest (ldfile, 0);
	    }
	  break;

	case tok_collating_element:
	  /* Ignore the rest of the line if we don't need the input of
	     this line.  */
	  if (ignore_content)
	    {
	      lr_ignore_rest (ldfile, 0);
	      break;
	    }

	  if (state != 0)
	    goto err_label;

	  arg = lr_token (ldfile, charmap, repertoire);
	  if (arg->tok != tok_bsymbol)
	    goto err_label;
	  else
	    {
	      const char *symbol = arg->val.str.startmb;
	      size_t symbol_len = arg->val.str.lenmb;

	      /* Next the `from' keyword.  */
	      arg = lr_token (ldfile, charmap, repertoire);
	      if (arg->tok != tok_from)
		{
		  free ((char *) symbol);
		  goto err_label;
		}

	      ldfile->return_widestr = 1;
	      ldfile->translate_strings = 1;

	      /* Finally the string with the replacement.  */
	      arg = lr_token (ldfile, charmap, repertoire);

	      ldfile->return_widestr = 0;
	      ldfile->translate_strings = 0;

	      if (arg->tok != tok_string)
		goto err_label;

	      if (!ignore_content && symbol != NULL)
		{
		  /* The name is already defined.  */
		  if (check_duplicate (ldfile, collate, charmap,
				       repertoire, symbol, symbol_len))
		    goto col_elem_free;

		  if (insert_entry (&collate->elem_table,
				    symbol, symbol_len,
				    new_element (collate,
						 arg->val.str.startmb,
						 arg->val.str.lenmb - 1,
						 arg->val.str.startwc,
						 symbol, symbol_len, 0)) < 0)
		    lr_error (ldfile, _("\
error while adding collating element"));
		}
	      else
		{
		col_elem_free:
		  if (symbol != NULL)
		    free ((char *) symbol);
		  if (arg->val.str.startmb != NULL)
		    free (arg->val.str.startmb);
		  if (arg->val.str.startwc != NULL)
		    free (arg->val.str.startwc);
		}
	      lr_ignore_rest (ldfile, 1);
	    }
	  break;

	case tok_collating_symbol:
	  /* Ignore the rest of the line if we don't need the input of
	     this line.  */
	  if (ignore_content)
	    {
	      lr_ignore_rest (ldfile, 0);
	      break;
	    }

	  if (state != 0)
	    goto err_label;

	  arg = lr_token (ldfile, charmap, repertoire);
	  if (arg->tok != tok_bsymbol)
	    goto err_label;
	  else
	    {
	      const char *symbol = arg->val.str.startmb;
	      size_t symbol_len = arg->val.str.lenmb;

	      if (!ignore_content)
		{
		  if (symbol == NULL)
		    lr_error (ldfile, _("\
%s: unknown character in collating symbol name"),
			      "LC_COLLATE");
		  else
		    {
		      /* The name is already defined.  */
		      if (check_duplicate (ldfile, collate, charmap,
					   repertoire, symbol, symbol_len))
			goto col_sym_free;

		      if (insert_entry (&collate->sym_table,
					symbol, symbol_len,
					new_symbol (collate)) < 0)
			lr_error (ldfile, _("\
error while adding collating symbol"));
		    }
		}
	      else
		{
		col_sym_free:
		  if (symbol != NULL)
		    free ((char *) symbol);
		}
	      lr_ignore_rest (ldfile, 1);
	    }
	  break;

	case tok_symbol_equivalence:
	  /* Ignore the rest of the line if we don't need the input of
	     this line.  */
	  if (ignore_content)
	    {
	      lr_ignore_rest (ldfile, 0);
	      break;
	    }

	  if (state != 0)
	    goto err_label;

	  arg = lr_token (ldfile, charmap, repertoire);
	  if (arg->tok != tok_bsymbol)
	    goto err_label;
	  else
	    {
	      const char *newname = arg->val.str.startmb;
	      size_t newname_len = arg->val.str.lenmb;
	      const char *symname;
	      size_t symname_len;
	      struct symbol_t *symval;

	      arg = lr_token (ldfile, charmap, repertoire);
	      if (arg->tok != tok_bsymbol)
		{
		  if (newname != NULL)
		    free ((char *) newname);
		  goto err_label;
		}

	      symname = arg->val.str.startmb;
	      symname_len = arg->val.str.lenmb;

	      if (newname == NULL)
		{
		  lr_error (ldfile, _("\
%s: unknown character in equivalent definition name"),
			    "LC_COLLATE");

		sym_equiv_free:
		  if (newname != NULL)
		    free ((char *) newname);
		  if (symname != NULL)
		    free ((char *) symname);
		  break;
		}
	      if (symname == NULL)
		{
		  lr_error (ldfile, _("\
%s: unknown character in equivalent definition value"),
			    "LC_COLLATE");
		  goto sym_equiv_free;
		}
	      /* The name is already defined.  */
	      if (check_duplicate (ldfile, collate, charmap,
				   repertoire, symname, symname_len))
		goto col_sym_free;

	      /* See whether the symbol name is already defined.  */
	      if (find_entry (&collate->sym_table, symname, symname_len,
			      (void **) &symval) != 0)
		{
		  lr_error (ldfile, _("\
%s: unknown symbol `%s' in equivalent definition"),
			    "LC_COLLATE", symname);
		  goto col_sym_free;
		}

	      if (insert_entry (&collate->sym_table,
				newname, newname_len, symval) < 0)
		{
		  lr_error (ldfile, _("\
error while adding equivalent collating symbol"));
		  goto sym_equiv_free;
		}

	      free ((char *) symname);
	    }
	  lr_ignore_rest (ldfile, 1);
	  break;

	case tok_order_start:
	  /* Ignore the rest of the line if we don't need the input of
	     this line.  */
	  if (ignore_content)
	    {
	      lr_ignore_rest (ldfile, 0);
	      break;
	    }

	  if (state != 0 && state != 1)
	    goto err_label;
	  state = 1;

	  /* The 14652 draft does not specify whether all `order_start' lines
	     must contain the same number of sort-rules, but 14651 does.  So
	     we require this here as well.  */
	  arg = lr_token (ldfile, charmap, repertoire);
	  if (arg->tok == tok_bsymbol)
	    {
	      /* This better should be a section name.  */
	      struct section_list *sp = collate->sections;
	      while (sp != NULL
		     && strcmp (sp->name, arg->val.str.startmb) != 0)
		sp = sp->next;

	      if (sp == NULL)
		{
		  lr_error (ldfile, _("\
%s: unknown section name `%s'"),
			    "LC_COLLATE", arg->val.str.startmb);
		  /* We use the error section.  */
		  collate->current_section = &collate->error_section;

		  if (collate->error_section.first == NULL)
		    {
		      collate->error_section.next = collate->sections;
		      collate->sections = &collate->error_section;
		    }
		}
	      else
		{
		  /* Remember this section.  */
		  collate->current_section = sp;

		  /* One should not be allowed to open the same
                     section twice.  */
		  if (sp->first != NULL)
		    lr_error (ldfile, _("\
%s: multiple order definitions for section `%s'"),
			      "LC_COLLATE", sp->name);
		  else
		    {
		      sp->next = collate->sections;
		      collate->sections = sp;
		    }

		  /* Next should come the end of the line or a semicolon.  */
		  arg = lr_token (ldfile, charmap, repertoire);
		  if (arg->tok == tok_eol)
		    {
		      uint32_t cnt;

		      /* This means we have exactly one rule: `forward'.  */
		      if (nrules > 1)
			lr_error (ldfile, _("\
%s: invalid number of sorting rules"),
				  "LC_COLLATE");
		      else
			nrules = 1;
		      sp->rules = obstack_alloc (&collate->mempool,
						 (sizeof (enum coll_sort_rule)
						  * nrules));
		      for (cnt = 0; cnt < nrules; ++cnt)
			sp->rules[cnt] = sort_forward;

		      /* Next line.  */
		      break;
		    }

		  /* Get the next token.  */
		  arg = lr_token (ldfile, charmap, repertoire);
		}
	    }
	  else
	    {
	      /* There is no section symbol.  Therefore we use the unnamed
		 section.  */
	      collate->current_section = &collate->unnamed_section;

	      if (collate->unnamed_section.first != NULL)
		lr_error (ldfile, _("\
%s: multiple order definitions for unnamed section"),
			  "LC_COLLATE");
	      else
		{
		  collate->unnamed_section.next = collate->sections;
		  collate->sections = &collate->unnamed_section;
		}
	    }

	  /* Now read the direction names.  */
	  read_directions (ldfile, arg, charmap, repertoire, collate);

	  /* From now be need the strings untranslated.  */
	  ldfile->translate_strings = 0;
	  break;

	case tok_order_end:
	  /* Ignore the rest of the line if we don't need the input of
	     this line.  */
	  if (ignore_content)
	    {
	      lr_ignore_rest (ldfile, 0);
	      break;
	    }

	  if (state != 1)
	    goto err_label;

	  /* Handle ellipsis at end of list.  */
	  if (was_ellipsis != tok_none)
	    {
	      handle_ellipsis (ldfile, NULL, was_ellipsis, charmap, repertoire,
			       collate);
	      was_ellipsis = tok_none;
	    }

	  state = 2;
	  lr_ignore_rest (ldfile, 1);
	  break;

	case tok_reorder_after:
	  /* Ignore the rest of the line if we don't need the input of
	     this line.  */
	  if (ignore_content)
	    {
	      lr_ignore_rest (ldfile, 0);
	      break;
	    }

	  if (state == 1)
	    {
	      lr_error (ldfile, _("%s: missing `order_end' keyword"),
			"LC_COLLATE");
	      state = 2;

	      /* Handle ellipsis at end of list.  */
	      if (was_ellipsis != tok_none)
		{
		  handle_ellipsis (ldfile, arg, was_ellipsis, charmap,
				   repertoire, collate);
		  was_ellipsis = tok_none;
		}
	    }
	  else if (state != 2 && state != 3)
	    goto err_label;
	  state = 3;

	  arg = lr_token (ldfile, charmap, repertoire);
	  if (arg->tok == tok_bsymbol)
	    {
	      /* Find this symbol in the sequence table.  */
	      struct element_t *insp;
	      int no_error = 1;

	      if (find_entry (&collate->seq_table, arg->val.str.startmb,
			      arg->val.str.lenmb, (void **) &insp) == 0)
		/* Yes, the symbol exists.  Simply point the cursor
		   to it.  */
		  collate->cursor = insp;
	      else
		{
		  /* This is bad.  The symbol after which we have to
                     insert does not exist.  */
		  lr_error (ldfile, _("\
%s: cannot reorder after %.*s: symbol not known"),
			    "LC_COLLATE", (int) arg->val.str.lenmb,
			    arg->val.str.startmb);
		  collate->cursor = NULL;
		  no_error = 0;
		}

	      lr_ignore_rest (ldfile, no_error);
	    }
	  else
	    /* This must not happen.  */
	    goto err_label;
	  break;

	case tok_reorder_end:
	  /* Ignore the rest of the line if we don't need the input of
	     this line.  */
	  if (ignore_content)
	    break;

	  if (state != 3)
	    goto err_label;
	  state = 4;
	  lr_ignore_rest (ldfile, 1);
	  break;

	case tok_reorder_sections_after:
	  /* Ignore the rest of the line if we don't need the input of
	     this line.  */
	  if (ignore_content)
	    {
	      lr_ignore_rest (ldfile, 0);
	      break;
	    }

	  if (state == 1)
	    {
	      lr_error (ldfile, _("%s: missing `order_end' keyword"),
			"LC_COLLATE");
	      state = 2;

	      /* Handle ellipsis at end of list.  */
	      if (was_ellipsis != tok_none)
		{
		  handle_ellipsis (ldfile, NULL, was_ellipsis, charmap,
				   repertoire, collate);
		  was_ellipsis = tok_none;
		}
	    }
	  else if (state == 3)
	    {
	      error (0, 0, _("%s: missing `reorder-end' keyword"),
		     "LC_COLLATE");
	      state = 4;
	    }
	  else if (state != 2 && state != 4)
	    goto err_label;
	  state = 5;

	  /* Get the name of the sections we are adding after.  */
	  arg = lr_token (ldfile, charmap, repertoire);
	  if (arg->tok == tok_bsymbol)
	    {
	      /* Now find a section with this name.  */
	      struct section_list *runp = collate->sections;

	      while (runp != NULL)
		{
		  if (runp->name != NULL
		      && strlen (runp->name) == arg->val.str.lenmb
		      && memcmp (runp->name, arg->val.str.startmb,
				 arg->val.str.lenmb) == 0)
		    break;

		  runp = runp->next;
		}

	      if (runp != NULL)
		collate->current_section = runp;
	      else
		{
		  /* This is bad.  The section after which we have to
                     reorder does not exist.  Therefore we cannot
                     process the whole rest of this reorder
                     specification.  */
		  lr_error (ldfile, _("%s: section `%.*s' not known"),
			    "LC_COLLATE", (int) arg->val.str.lenmb,
			    arg->val.str.startmb);

		  do
		    {
		      lr_ignore_rest (ldfile, 0);

		      now = lr_token (ldfile, charmap, NULL);
		    }
		  while (now->tok == tok_reorder_sections_after
			 || now->tok == tok_reorder_sections_end
			 || now->tok == tok_end);

		  /* Process the token we just saw.  */
		  nowtok = now->tok;
		  continue;
		}
	    }
	  else
	    /* This must not happen.  */
	    goto err_label;
	  break;

	case tok_reorder_sections_end:
	  /* Ignore the rest of the line if we don't need the input of
	     this line.  */
	  if (ignore_content)
	    break;

	  if (state != 5)
	    goto err_label;
	  state = 6;
	  lr_ignore_rest (ldfile, 1);
	  break;

	case tok_bsymbol:
	  /* Ignore the rest of the line if we don't need the input of
	     this line.  */
	  if (ignore_content)
	    {
	      lr_ignore_rest (ldfile, 0);
	      break;
	    }

	  if (state != 1 && state != 3)
	    goto err_label;

	  if (state == 3)
	    {
	      /* It is possible that we already have this collation sequence.
		 In this case we move the entry.  */
	      struct element_t *seqp;

	      /* If the symbol after which we have to insert was not found
		 ignore all entries.  */
	      if (collate->cursor == NULL)
		{
		  lr_ignore_rest (ldfile, 0);
		  break;
		}

	      if (find_entry (&collate->seq_table, arg->val.str.startmb,
			      arg->val.str.lenmb, (void **) &seqp) == 0)
		{
		  /* Remove the entry from the old position.  */
		  if (seqp->last == NULL)
		    collate->start = seqp->next;
		  else
		    seqp->last->next = seqp->next;
		  if (seqp->next != NULL)
		    seqp->next->last = seqp->last;

		  /* We also have to check whether this entry is the
                     first or last of a section.  */
		  if (seqp->section->first == seqp)
		    {
		      if (seqp->section->first == seqp->section->last)
			/* This setion has no content anymore.  */
			seqp->section->first = seqp->section->last = NULL;
		      else
			seqp->section->first = seqp->next;
		    }
		  else if (seqp->section->last == seqp)
		    seqp->section->last = seqp->last;

		  /* Now insert it in the new place.  */
		  seqp->next = collate->cursor->next;
		  seqp->last = collate->cursor;
		  collate->cursor->next = seqp;
		  if (seqp->next != NULL)
		    seqp->next->last = seqp;

		  seqp->section = collate->cursor->section;
		  if (seqp->section->last == collate->cursor)
		    seqp->section->last = seqp;

		  break;
		}

	      /* Otherwise we just add a new entry.  */
	    }
	  else if (state == 5)
	    {
	      /* We are reordering sections.  Find the named section.  */
	      struct section_list *runp = collate->sections;
	      struct section_list *prevp = NULL;

	      while (runp != NULL)
		{
		  if (runp->name != NULL
		      && strlen (runp->name) == arg->val.str.lenmb
		      && memcmp (runp->name, arg->val.str.startmb,
				 arg->val.str.lenmb) == 0)
		    break;

		  prevp = runp;
		  runp = runp->next;
		}

	      if (runp == NULL)
		{
		  lr_error (ldfile, _("%s: section `%.*s' not known"),
			    "LC_COLLATE", (int) arg->val.str.lenmb,
			    arg->val.str.startmb);
		  lr_ignore_rest (ldfile, 0);
		}
	      else
		{
		  if (runp != collate->current_section)
		    {
		      /* Remove the named section from the old place and
			 insert it in the new one.  */
		      prevp->next = runp->next;

		      runp->next = collate->current_section->next;
		      collate->current_section->next = runp;
		      collate->current_section = runp;
		    }

		  /* Process the rest of the line which might change
                     the collation rules.  */
		  arg = lr_token (ldfile, charmap, repertoire);
		  if (arg->tok != tok_eof && arg->tok != tok_eol)
		    read_directions (ldfile, arg, charmap, repertoire,
				     collate);
		}
	      break;
	    }
	  else if (was_ellipsis != tok_none)
	    {
	      /* Using the information in the `ellipsis_weight'
                 element and this and the last value we have to handle
                 the ellipsis now.  */
	      assert (state == 1);

	      handle_ellipsis (ldfile, arg, was_ellipsis, charmap, repertoire,
			       collate);

	      /* Remember that we processed the ellipsis.  */
	      was_ellipsis = tok_none;

	      /* And don't add the value a second time.  */
	      break;
	    }

	  /* Now insert in the new place.  */
	  insert_value (ldfile, arg, charmap, repertoire, collate);
	  break;

	case tok_undefined:
	  /* Ignore the rest of the line if we don't need the input of
	     this line.  */
	  if (ignore_content)
	    {
	      lr_ignore_rest (ldfile, 0);
	      break;
	    }

	  if (state != 1)
	    goto err_label;

	  if (was_ellipsis != tok_none)
	    {
	      lr_error (ldfile,
			_("%s: cannot have `%s' as end of ellipsis range"),
			"LC_COLLATE", "UNDEFINED");

	      unlink_element (collate);
	      was_ellipsis = tok_none;
	    }

	  /* See whether UNDEFINED already appeared somewhere.  */
	  if (collate->undefined.next != NULL
	      || (collate->cursor != NULL
		  && collate->undefined.next == collate->cursor))
	    {
	      lr_error (ldfile,
			_("%s: order for `%.*s' already defined at %s:%zu"),
			"LC_COLLATE", 9, "UNDEFINED", collate->undefined.file,
			collate->undefined.line);
	      lr_ignore_rest (ldfile, 0);
	    }
	  else
	    /* Parse the weights.  */
	     insert_weights (ldfile, &collate->undefined, charmap,
			     repertoire, collate, tok_none);
	  break;

	case tok_ellipsis2:
	case tok_ellipsis3:
	case tok_ellipsis4:
	  /* This is the symbolic (decimal or hexadecimal) or absolute
             ellipsis.  */
	  if (was_ellipsis != tok_none)
	    goto err_label;

	  if (state != 1 && state != 3)
	    goto err_label;

	  was_ellipsis = nowtok;

	  insert_weights (ldfile, &collate->ellipsis_weight, charmap,
			  repertoire, collate, nowtok);
	  break;

	case tok_end:
	  /* Next we assume `LC_COLLATE'.  */
	  if (!ignore_content)
	    {
	      if (state == 0)
		/* We must either see a copy statement or have
		   ordering values.  */
		lr_error (ldfile,
			  _("%s: empty category description not allowed"),
			  "LC_COLLATE");
	      else if (state == 1)
		{
		  lr_error (ldfile, _("%s: missing `order_end' keyword"),
			    "LC_COLLATE");

		  /* Handle ellipsis at end of list.  */
		  if (was_ellipsis != tok_none)
		    {
		      handle_ellipsis (ldfile, NULL, was_ellipsis, charmap,
				       repertoire, collate);
		      was_ellipsis = tok_none;
		    }
		}
	      else if (state == 3)
		error (0, 0, _("%s: missing `reorder-end' keyword"),
		       "LC_COLLATE");
	      else if (state == 5)
		error (0, 0, _("%s: missing `reorder-sections-end' keyword"),
		       "LC_COLLATE");
	    }
	  arg = lr_token (ldfile, charmap, NULL);
	  if (arg->tok == tok_eof)
	    break;
	  if (arg->tok == tok_eol)
	    lr_error (ldfile, _("%s: incomplete `END' line"), "LC_COLLATE");
	  else if (arg->tok != tok_lc_collate)
	    lr_error (ldfile, _("\
%1$s: definition does not end with `END %1$s'"), "LC_COLLATE");
	  lr_ignore_rest (ldfile, arg->tok == tok_lc_collate);
	  return;

	default:
	err_label:
	  SYNTAX_ERROR (_("%s: syntax error"), "LC_COLLATE");
	}

      /* Prepare for the next round.  */
      now = lr_token (ldfile, charmap, NULL);
      nowtok = now->tok;
    }

  /* When we come here we reached the end of the file.  */
  lr_error (ldfile, _("%s: premature end of file"), "LC_COLLATE");
}