xref: /freebsd/usr.bin/localedef/collate.c (revision 3416500aef140042c64bc149cb1ec6620483bc44)
1 /*-
2  * Copyright 2010 Nexenta Systems, Inc.  All rights reserved.
3  * Copyright 2015 John Marino <draco@marino.st>
4  *
5  * This source code is derived from the illumos localedef command, and
6  * provided under BSD-style license terms by Nexenta Systems, Inc.
7  *
8  * Redistribution and use in source and binary forms, with or without
9  * modification, are permitted provided that the following conditions
10  * are met:
11  *
12  * 1. Redistributions of source code must retain the above copyright
13  *    notice, this list of conditions and the following disclaimer.
14  * 2. Redistributions in binary form must reproduce the above copyright
15  *    notice, this list of conditions and the following disclaimer in the
16  *    documentation and/or other materials provided with the distribution.
17  *
18  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
19  * AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
20  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
21  * ARE DISCLAIMED.  IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
22  * LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
23  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
24  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
25  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
26  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
27  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
28  * POSSIBILITY OF SUCH DAMAGE.
29  */
30 
31 /*
32  * LC_COLLATE database generation routines for localedef.
33  */
34 #include <sys/cdefs.h>
35 __FBSDID("$FreeBSD$");
36 
37 #include <sys/types.h>
38 #include <sys/tree.h>
39 
40 #include <stdio.h>
41 #include <stddef.h>
42 #include <stdlib.h>
43 #include <errno.h>
44 #include <string.h>
45 #include <unistd.h>
46 #include <wchar.h>
47 #include <limits.h>
48 #include "localedef.h"
49 #include "parser.h"
50 #include "collate.h"
51 
52 /*
53  * Design notes.
54  *
55  * It will be extremely helpful to the reader if they have access to
56  * the localedef and locale file format specifications available.
57  * Latest versions of these are available from www.opengroup.org.
58  *
59  * The design for the collation code is a bit complex.  The goal is a
60  * single collation database as described in collate.h (in
61  * libc/port/locale).  However, there are some other tidbits:
62  *
63  * a) The substitution entries are now a directly indexable array.  A
64  * priority elsewhere in the table is taken as an index into the
65  * substitution table if it has a high bit (COLLATE_SUBST_PRIORITY)
66  * set.  (The bit is cleared and the result is the index into the
67  * table.
68  *
69  * b) We eliminate duplicate entries into the substitution table.
70  * This saves a lot of space.
71  *
72  * c) The priorities for each level are "compressed", so that each
73  * sorting level has consecutively numbered priorities starting at 1.
74  * (O is reserved for the ignore priority.)  This means sort levels
75  * which only have a few distinct priorities can represent the
76  * priority level in fewer bits, which makes the strxfrm output
77  * smaller.
78  *
79  * d) We record the total number of priorities so that strxfrm can
80  * figure out how many bytes to expand a numeric priority into.
81  *
82  * e) For the UNDEFINED pass (the last pass), we record the maximum
83  * number of bits needed to uniquely prioritize these entries, so that
84  * the last pass can also use smaller strxfrm output when possible.
85  *
86  * f) Priorities with the sign bit set are verboten.  This works out
87  * because no active character set needs that bit to carry significant
88  * information once the character is in wide form.
89  *
90  * To process the entire data to make the database, we actually run
91  * multiple passes over the data.
92  *
93  * The first pass, which is done at parse time, identifies elements,
94  * substitutions, and such, and records them in priority order.  As
95  * some priorities can refer to other priorities, using forward
96  * references, we use a table of references indicating whether the
97  * priority's value has been resolved, or whether it is still a
98  * reference.
99  *
100  * The second pass walks over all the items in priority order, noting
101  * that they are used directly, and not just an indirect reference.
102  * This is done by creating a "weight" structure for the item.  The
103  * weights are stashed in an RB tree sorted by relative "priority".
104  *
105  * The third pass walks over all the weight structures, in priority
106  * order, and assigns a new monotonically increasing (per sort level)
107  * weight value to them.  These are the values that will actually be
108  * written to the file.
109  *
110  * The fourth pass just writes the data out.
111  */
112 
113 /*
114  * In order to resolve the priorities, we create a table of priorities.
115  * Entries in the table can be in one of three states.
116  *
117  * UNKNOWN is for newly allocated entries, and indicates that nothing
118  * is known about the priority.  (For example, when new entries are created
119  * for collating-symbols, this is the value assigned for them until the
120  * collating symbol's order has been determined.
121  *
122  * RESOLVED is used for an entry where the priority indicates the final
123  * numeric weight.
124  *
125  * REFER is used for entries that reference other entries.  Typically
126  * this is used for forward references.  A collating-symbol can never
127  * have this value.
128  *
129  * The "pass" field is used during final resolution to aid in detection
130  * of referencing loops.  (For example <A> depends on <B>, but <B> has its
131  * priority dependent on <A>.)
132  */
133 typedef enum {
134 	UNKNOWN,	/* priority is totally unknown */
135 	RESOLVED,	/* priority value fully resolved */
136 	REFER		/* priority is a reference (index) */
137 } res_t;
138 
139 typedef struct weight {
140 	int32_t		pri;
141 	int		opt;
142 	RB_ENTRY(weight) entry;
143 } weight_t;
144 
145 typedef struct priority {
146 	res_t		res;
147 	int32_t		pri;
148 	int		pass;
149 	int		lineno;
150 } collpri_t;
151 
152 #define	NUM_WT	collinfo.directive_count
153 
154 /*
155  * These are the abstract collating symbols, which are just a symbolic
156  * way to reference a priority.
157  */
158 struct collsym {
159 	char		*name;
160 	int32_t		ref;
161 	RB_ENTRY(collsym) entry;
162 };
163 
164 /*
165  * These are also abstract collating symbols, but we allow them to have
166  * different priorities at different levels.
167  */
168 typedef struct collundef {
169 	char		*name;
170 	int32_t		ref[COLL_WEIGHTS_MAX];
171 	RB_ENTRY(collundef) entry;
172 } collundef_t;
173 
174 /*
175  * These are called "chains" in libc.  This records the fact that two
176  * more characters should be treated as a single collating entity when
177  * they appear together.  For example, in Spanish <C><h> gets collated
178  * as a character between <C> and <D>.
179  */
180 struct collelem {
181 	char		*symbol;
182 	wchar_t		*expand;
183 	int32_t		ref[COLL_WEIGHTS_MAX];
184 	RB_ENTRY(collelem) rb_bysymbol;
185 	RB_ENTRY(collelem) rb_byexpand;
186 };
187 
188 /*
189  * Individual characters have a sequence of weights as well.
190  */
191 typedef struct collchar {
192 	wchar_t		wc;
193 	int32_t		ref[COLL_WEIGHTS_MAX];
194 	RB_ENTRY(collchar) entry;
195 } collchar_t;
196 
197 /*
198  * Substitution entries.  The key is itself a priority.  Note that
199  * when we create one of these, we *automatically* wind up with a
200  * fully resolved priority for the key, because creation of
201  * substitutions creates a resolved priority at the same time.
202  */
203 typedef struct subst{
204 	int32_t		key;
205 	int32_t		ref[COLLATE_STR_LEN];
206 	RB_ENTRY(subst)	entry;
207 	RB_ENTRY(subst)	entry_ref;
208 } subst_t;
209 
210 static RB_HEAD(collsyms, collsym) collsyms;
211 static RB_HEAD(collundefs, collundef) collundefs;
212 static RB_HEAD(elem_by_symbol, collelem) elem_by_symbol;
213 static RB_HEAD(elem_by_expand, collelem) elem_by_expand;
214 static RB_HEAD(collchars, collchar) collchars;
215 static RB_HEAD(substs, subst) substs[COLL_WEIGHTS_MAX];
216 static RB_HEAD(substs_ref, subst) substs_ref[COLL_WEIGHTS_MAX];
217 static RB_HEAD(weights, weight) weights[COLL_WEIGHTS_MAX];
218 static int32_t		nweight[COLL_WEIGHTS_MAX];
219 
220 /*
221  * This is state tracking for the ellipsis token.  Note that we start
222  * the initial values so that the ellipsis logic will think we got a
223  * magic starting value of NUL.  It starts at minus one because the
224  * starting point is exclusive -- i.e. the starting point is not
225  * itself handled by the ellipsis code.
226  */
227 static int currorder = EOF;
228 static int lastorder = EOF;
229 static collelem_t *currelem;
230 static collchar_t *currchar;
231 static collundef_t *currundef;
232 static wchar_t ellipsis_start = 0;
233 static int32_t ellipsis_weights[COLL_WEIGHTS_MAX];
234 
235 /*
236  * We keep a running tally of weights.
237  */
238 static int nextpri = 1;
239 static int nextsubst[COLL_WEIGHTS_MAX] = { 0 };
240 
241 /*
242  * This array collects up the weights for each level.
243  */
244 static int32_t order_weights[COLL_WEIGHTS_MAX];
245 static int curr_weight = 0;
246 static int32_t subst_weights[COLLATE_STR_LEN];
247 static int curr_subst = 0;
248 
249 /*
250  * Some initial priority values.
251  */
252 static int32_t pri_undefined[COLL_WEIGHTS_MAX];
253 static int32_t pri_ignore;
254 
255 static collate_info_t collinfo;
256 
257 static collpri_t	*prilist = NULL;
258 static int		numpri = 0;
259 static int		maxpri = 0;
260 
261 static void start_order(int);
262 
263 static int32_t
264 new_pri(void)
265 {
266 	int i;
267 
268 	if (numpri >= maxpri) {
269 		maxpri = maxpri ? maxpri * 2 : 1024;
270 		prilist = realloc(prilist, sizeof (collpri_t) * maxpri);
271 		if (prilist == NULL) {
272 			fprintf(stderr,"out of memory");
273 			return (-1);
274 		}
275 		for (i = numpri; i < maxpri; i++) {
276 			prilist[i].res = UNKNOWN;
277 			prilist[i].pri = 0;
278 			prilist[i].pass = 0;
279 		}
280 	}
281 	return (numpri++);
282 }
283 
284 static collpri_t *
285 get_pri(int32_t ref)
286 {
287 	if ((ref < 0) || (ref > numpri)) {
288 		INTERR;
289 		return (NULL);
290 	}
291 	return (&prilist[ref]);
292 }
293 
294 static void
295 set_pri(int32_t ref, int32_t v, res_t res)
296 {
297 	collpri_t	*pri;
298 
299 	pri = get_pri(ref);
300 
301 	if ((res == REFER) && ((v < 0) || (v >= numpri))) {
302 		INTERR;
303 	}
304 
305 	/* Resolve self references */
306 	if ((res == REFER) && (ref == v)) {
307 		v = nextpri;
308 		res = RESOLVED;
309 	}
310 
311 	if (pri->res != UNKNOWN) {
312 		warn("repeated item in order list (first on %d)",
313 		    pri->lineno);
314 		return;
315 	}
316 	pri->lineno = lineno;
317 	pri->pri = v;
318 	pri->res = res;
319 }
320 
321 static int32_t
322 resolve_pri(int32_t ref)
323 {
324 	collpri_t	*pri;
325 	static int32_t	pass = 0;
326 
327 	pri = get_pri(ref);
328 	pass++;
329 	while (pri->res == REFER) {
330 		if (pri->pass == pass) {
331 			/* report a line with the circular symbol */
332 			lineno = pri->lineno;
333 			fprintf(stderr,"circular reference in order list");
334 			return (-1);
335 		}
336 		if ((pri->pri < 0) || (pri->pri >= numpri)) {
337 			INTERR;
338 			return (-1);
339 		}
340 		pri->pass = pass;
341 		pri = &prilist[pri->pri];
342 	}
343 
344 	if (pri->res == UNKNOWN) {
345 		return (-1);
346 	}
347 	if (pri->res != RESOLVED)
348 		INTERR;
349 
350 	return (pri->pri);
351 }
352 
353 static int
354 weight_compare(const void *n1, const void *n2)
355 {
356 	int32_t	k1 = ((const weight_t *)n1)->pri;
357 	int32_t	k2 = ((const weight_t *)n2)->pri;
358 
359 	return (k1 < k2 ? -1 : k1 > k2 ? 1 : 0);
360 }
361 
362 RB_GENERATE_STATIC(weights, weight, entry, weight_compare);
363 
364 static int
365 collsym_compare(const void *n1, const void *n2)
366 {
367 	const collsym_t *c1 = n1;
368 	const collsym_t *c2 = n2;
369 	int rv;
370 
371 	rv = strcmp(c1->name, c2->name);
372 	return ((rv < 0) ? -1 : (rv > 0) ? 1 : 0);
373 }
374 
375 RB_GENERATE_STATIC(collsyms, collsym, entry, collsym_compare);
376 
377 static int
378 collundef_compare(const void *n1, const void *n2)
379 {
380 	const collundef_t *c1 = n1;
381 	const collundef_t *c2 = n2;
382 	int rv;
383 
384 	rv = strcmp(c1->name, c2->name);
385 	return ((rv < 0) ? -1 : (rv > 0) ? 1 : 0);
386 }
387 
388 RB_GENERATE_STATIC(collundefs, collundef, entry, collundef_compare);
389 
390 static int
391 element_compare_symbol(const void *n1, const void *n2)
392 {
393 	const collelem_t *c1 = n1;
394 	const collelem_t *c2 = n2;
395 	int rv;
396 
397 	rv = strcmp(c1->symbol, c2->symbol);
398 	return ((rv < 0) ? -1 : (rv > 0) ? 1 : 0);
399 }
400 
401 RB_GENERATE_STATIC(elem_by_symbol, collelem, rb_bysymbol, element_compare_symbol);
402 
403 static int
404 element_compare_expand(const void *n1, const void *n2)
405 {
406 	const collelem_t *c1 = n1;
407 	const collelem_t *c2 = n2;
408 	int rv;
409 
410 	rv = wcscmp(c1->expand, c2->expand);
411 	return ((rv < 0) ? -1 : (rv > 0) ? 1 : 0);
412 }
413 
414 RB_GENERATE_STATIC(elem_by_expand, collelem, rb_byexpand, element_compare_expand);
415 
416 static int
417 collchar_compare(const void *n1, const void *n2)
418 {
419 	wchar_t	k1 = ((const collchar_t *)n1)->wc;
420 	wchar_t	k2 = ((const collchar_t *)n2)->wc;
421 
422 	return (k1 < k2 ? -1 : k1 > k2 ? 1 : 0);
423 }
424 
425 RB_GENERATE_STATIC(collchars, collchar, entry, collchar_compare);
426 
427 static int
428 subst_compare(const void *n1, const void *n2)
429 {
430 	int32_t	k1 = ((const subst_t *)n1)->key;
431 	int32_t	k2 = ((const subst_t *)n2)->key;
432 
433 	return (k1 < k2 ? -1 : k1 > k2 ? 1 : 0);
434 }
435 
436 RB_GENERATE_STATIC(substs, subst, entry, subst_compare);
437 
438 static int
439 subst_compare_ref(const void *n1, const void *n2)
440 {
441 	const wchar_t *c1 = ((const subst_t *)n1)->ref;
442 	const wchar_t *c2 = ((const subst_t *)n2)->ref;
443 	int rv;
444 
445 	rv = wcscmp(c1, c2);
446 	return ((rv < 0) ? -1 : (rv > 0) ? 1 : 0);
447 }
448 
449 RB_GENERATE_STATIC(substs_ref, subst, entry_ref, subst_compare_ref);
450 
451 void
452 init_collate(void)
453 {
454 	int i;
455 
456 	RB_INIT(&collsyms);
457 
458 	RB_INIT(&collundefs);
459 
460 	RB_INIT(&elem_by_symbol);
461 
462 	RB_INIT(&elem_by_expand);
463 
464 	RB_INIT(&collchars);
465 
466 	for (i = 0; i < COLL_WEIGHTS_MAX; i++) {
467 		RB_INIT(&substs[i]);
468 		RB_INIT(&substs_ref[i]);
469 		RB_INIT(&weights[i]);
470 		nweight[i] = 1;
471 	}
472 
473 	(void) memset(&collinfo, 0, sizeof (collinfo));
474 
475 	/* allocate some initial priorities */
476 	pri_ignore = new_pri();
477 
478 	set_pri(pri_ignore, 0, RESOLVED);
479 
480 	for (i = 0; i < COLL_WEIGHTS_MAX; i++) {
481 		pri_undefined[i] = new_pri();
482 
483 		/* we will override this later */
484 		set_pri(pri_undefined[i], COLLATE_MAX_PRIORITY, UNKNOWN);
485 	}
486 }
487 
488 void
489 define_collsym(char *name)
490 {
491 	collsym_t	*sym;
492 
493 	if ((sym = calloc(1, sizeof(*sym))) == NULL) {
494 		fprintf(stderr,"out of memory");
495 		return;
496 	}
497 	sym->name = name;
498 	sym->ref = new_pri();
499 
500 	if (RB_FIND(collsyms, &collsyms, sym) != NULL) {
501 		/*
502 		 * This should never happen because we are only called
503 		 * for undefined symbols.
504 		 */
505 		free(sym);
506 		INTERR;
507 		return;
508 	}
509 	RB_INSERT(collsyms, &collsyms, sym);
510 }
511 
512 collsym_t *
513 lookup_collsym(char *name)
514 {
515 	collsym_t	srch;
516 
517 	srch.name = name;
518 	return (RB_FIND(collsyms, &collsyms, &srch));
519 }
520 
521 collelem_t *
522 lookup_collelem(char *symbol)
523 {
524 	collelem_t	srch;
525 
526 	srch.symbol = symbol;
527 	return (RB_FIND(elem_by_symbol, &elem_by_symbol, &srch));
528 }
529 
530 static collundef_t *
531 get_collundef(char *name)
532 {
533 	collundef_t	srch;
534 	collundef_t	*ud;
535 	int		i;
536 
537 	srch.name = name;
538 	if ((ud = RB_FIND(collundefs, &collundefs, &srch)) == NULL) {
539 		if (((ud = calloc(1, sizeof(*ud))) == NULL) ||
540 		    ((ud->name = strdup(name)) == NULL)) {
541 			fprintf(stderr,"out of memory");
542 			free(ud);
543 			return (NULL);
544 		}
545 		for (i = 0; i < NUM_WT; i++) {
546 			ud->ref[i] = new_pri();
547 		}
548 		RB_INSERT(collundefs, &collundefs, ud);
549 	}
550 	add_charmap_undefined(name);
551 	return (ud);
552 }
553 
554 static collchar_t *
555 get_collchar(wchar_t wc, int create)
556 {
557 	collchar_t	srch;
558 	collchar_t	*cc;
559 	int		i;
560 
561 	srch.wc = wc;
562 	cc = RB_FIND(collchars, &collchars, &srch);
563 	if ((cc == NULL) && create) {
564 		if ((cc = calloc(1, sizeof(*cc))) == NULL) {
565 			fprintf(stderr, "out of memory");
566 			return (NULL);
567 		}
568 		for (i = 0; i < NUM_WT; i++) {
569 			cc->ref[i] = new_pri();
570 		}
571 		cc->wc = wc;
572 		RB_INSERT(collchars, &collchars, cc);
573 	}
574 	return (cc);
575 }
576 
577 void
578 end_order_collsym(collsym_t *sym)
579 {
580 	start_order(T_COLLSYM);
581 	/* update the weight */
582 
583 	set_pri(sym->ref, nextpri, RESOLVED);
584 	nextpri++;
585 }
586 
587 void
588 end_order(void)
589 {
590 	int		i;
591 	int32_t		pri;
592 	int32_t		ref;
593 	collpri_t	*p;
594 
595 	/* advance the priority/weight */
596 	pri = nextpri;
597 
598 	switch (currorder) {
599 	case T_CHAR:
600 		for (i = 0; i < NUM_WT; i++) {
601 			if (((ref = order_weights[i]) < 0) ||
602 			    ((p = get_pri(ref)) == NULL) ||
603 			    (p->pri == -1)) {
604 				/* unspecified weight is a self reference */
605 				set_pri(currchar->ref[i], pri, RESOLVED);
606 			} else {
607 				set_pri(currchar->ref[i], ref, REFER);
608 			}
609 			order_weights[i] = -1;
610 		}
611 
612 		/* leave a cookie trail in case next symbol is ellipsis */
613 		ellipsis_start = currchar->wc + 1;
614 		currchar = NULL;
615 		break;
616 
617 	case T_ELLIPSIS:
618 		/* save off the weights were we can find them */
619 		for (i = 0; i < NUM_WT; i++) {
620 			ellipsis_weights[i] = order_weights[i];
621 			order_weights[i] = -1;
622 		}
623 		break;
624 
625 	case T_COLLELEM:
626 		if (currelem == NULL) {
627 			INTERR;
628 		} else {
629 			for (i = 0; i < NUM_WT; i++) {
630 
631 				if (((ref = order_weights[i]) < 0) ||
632 				    ((p = get_pri(ref)) == NULL) ||
633 				    (p->pri == -1)) {
634 					set_pri(currelem->ref[i], pri,
635 					    RESOLVED);
636 				} else {
637 					set_pri(currelem->ref[i], ref, REFER);
638 				}
639 				order_weights[i] = -1;
640 			}
641 		}
642 		break;
643 
644 	case T_UNDEFINED:
645 		for (i = 0; i < NUM_WT; i++) {
646 			if (((ref = order_weights[i]) < 0) ||
647 			    ((p = get_pri(ref)) == NULL) ||
648 			    (p->pri == -1)) {
649 				set_pri(pri_undefined[i], -1, RESOLVED);
650 			} else {
651 				set_pri(pri_undefined[i], ref, REFER);
652 			}
653 			order_weights[i] = -1;
654 		}
655 		break;
656 
657 	case T_SYMBOL:
658 		for (i = 0; i < NUM_WT; i++) {
659 			if (((ref = order_weights[i]) < 0) ||
660 			    ((p = get_pri(ref)) == NULL) ||
661 			    (p->pri == -1)) {
662 				set_pri(currundef->ref[i], pri, RESOLVED);
663 			} else {
664 				set_pri(currundef->ref[i], ref, REFER);
665 			}
666 			order_weights[i] = -1;
667 		}
668 		break;
669 
670 	default:
671 		INTERR;
672 	}
673 
674 	nextpri++;
675 }
676 
677 static void
678 start_order(int type)
679 {
680 	int	i;
681 
682 	lastorder = currorder;
683 	currorder = type;
684 
685 	/* this is used to protect ELLIPSIS processing */
686 	if ((lastorder == T_ELLIPSIS) && (type != T_CHAR)) {
687 		fprintf(stderr, "character value expected");
688 	}
689 
690 	for (i = 0; i < COLL_WEIGHTS_MAX; i++) {
691 		order_weights[i] = -1;
692 	}
693 	curr_weight = 0;
694 }
695 
696 void
697 start_order_undefined(void)
698 {
699 	start_order(T_UNDEFINED);
700 }
701 
702 void
703 start_order_symbol(char *name)
704 {
705 	currundef = get_collundef(name);
706 	start_order(T_SYMBOL);
707 }
708 
709 void
710 start_order_char(wchar_t wc)
711 {
712 	collchar_t	*cc;
713 	int32_t		ref;
714 
715 	start_order(T_CHAR);
716 
717 	/*
718 	 * If we last saw an ellipsis, then we need to close the range.
719 	 * Handle that here.  Note that we have to be careful because the
720 	 * items *inside* the range are treated exclusiveley to the items
721 	 * outside of the range.  The ends of the range can have quite
722 	 * different weights than the range members.
723 	 */
724 	if (lastorder == T_ELLIPSIS) {
725 		int		i;
726 
727 		if (wc < ellipsis_start) {
728 			fprintf(stderr, "malformed range!");
729 			return;
730 		}
731 		while (ellipsis_start < wc) {
732 			/*
733 			 * pick all of the saved weights for the
734 			 * ellipsis.  note that -1 encodes for the
735 			 * ellipsis itself, which means to take the
736 			 * current relative priority.
737 			 */
738 			if ((cc = get_collchar(ellipsis_start, 1)) == NULL) {
739 				INTERR;
740 				return;
741 			}
742 			for (i = 0; i < NUM_WT; i++) {
743 				collpri_t *p;
744 				if (((ref = ellipsis_weights[i]) == -1) ||
745 				    ((p = get_pri(ref)) == NULL) ||
746 				    (p->pri == -1)) {
747 					set_pri(cc->ref[i], nextpri, RESOLVED);
748 				} else {
749 					set_pri(cc->ref[i], ref, REFER);
750 				}
751 				ellipsis_weights[i] = 0;
752 			}
753 			ellipsis_start++;
754 			nextpri++;
755 		}
756 	}
757 
758 	currchar = get_collchar(wc, 1);
759 }
760 
761 void
762 start_order_collelem(collelem_t *e)
763 {
764 	start_order(T_COLLELEM);
765 	currelem = e;
766 }
767 
768 void
769 start_order_ellipsis(void)
770 {
771 	int	i;
772 
773 	start_order(T_ELLIPSIS);
774 
775 	if (lastorder != T_CHAR) {
776 		fprintf(stderr, "illegal starting point for range");
777 		return;
778 	}
779 
780 	for (i = 0; i < NUM_WT; i++) {
781 		ellipsis_weights[i] = order_weights[i];
782 	}
783 }
784 
785 void
786 define_collelem(char *name, wchar_t *wcs)
787 {
788 	collelem_t	*e;
789 	int		i;
790 
791 	if (wcslen(wcs) >= COLLATE_STR_LEN) {
792 		fprintf(stderr,"expanded collation element too long");
793 		return;
794 	}
795 
796 	if ((e = calloc(1, sizeof(*e))) == NULL) {
797 		fprintf(stderr, "out of memory");
798 		return;
799 	}
800 	e->expand = wcs;
801 	e->symbol = name;
802 
803 	/*
804 	 * This is executed before the order statement, so we don't
805 	 * know how many priorities we *really* need.  We allocate one
806 	 * for each possible weight.  Not a big deal, as collating-elements
807 	 * prove to be quite rare.
808 	 */
809 	for (i = 0; i < COLL_WEIGHTS_MAX; i++) {
810 		e->ref[i] = new_pri();
811 	}
812 
813 	/* A character sequence can only reduce to one element. */
814 	if ((RB_FIND(elem_by_symbol, &elem_by_symbol, e) != NULL) ||
815 	    (RB_FIND(elem_by_expand, &elem_by_expand, e) != NULL)) {
816 		fprintf(stderr, "duplicate collating element definition");
817 		free(e);
818 		return;
819 	}
820 	RB_INSERT(elem_by_symbol, &elem_by_symbol, e);
821 	RB_INSERT(elem_by_expand, &elem_by_expand, e);
822 }
823 
824 void
825 add_order_bit(int kw)
826 {
827 	uint8_t bit = DIRECTIVE_UNDEF;
828 
829 	switch (kw) {
830 	case T_FORWARD:
831 		bit = DIRECTIVE_FORWARD;
832 		break;
833 	case T_BACKWARD:
834 		bit = DIRECTIVE_BACKWARD;
835 		break;
836 	case T_POSITION:
837 		bit = DIRECTIVE_POSITION;
838 		break;
839 	default:
840 		INTERR;
841 		break;
842 	}
843 	collinfo.directive[collinfo.directive_count] |= bit;
844 }
845 
846 void
847 add_order_directive(void)
848 {
849 	if (collinfo.directive_count >= COLL_WEIGHTS_MAX) {
850 		fprintf(stderr,"too many directives (max %d)", COLL_WEIGHTS_MAX);
851 	}
852 	collinfo.directive_count++;
853 }
854 
855 static void
856 add_order_pri(int32_t ref)
857 {
858 	if (curr_weight >= NUM_WT) {
859 		fprintf(stderr,"too many weights (max %d)", NUM_WT);
860 		return;
861 	}
862 	order_weights[curr_weight] = ref;
863 	curr_weight++;
864 }
865 
866 void
867 add_order_collsym(collsym_t *s)
868 {
869 	add_order_pri(s->ref);
870 }
871 
872 void
873 add_order_char(wchar_t wc)
874 {
875 	collchar_t *cc;
876 
877 	if ((cc = get_collchar(wc, 1)) == NULL) {
878 		INTERR;
879 		return;
880 	}
881 
882 	add_order_pri(cc->ref[curr_weight]);
883 }
884 
885 void
886 add_order_collelem(collelem_t *e)
887 {
888 	add_order_pri(e->ref[curr_weight]);
889 }
890 
891 void
892 add_order_ignore(void)
893 {
894 	add_order_pri(pri_ignore);
895 }
896 
897 void
898 add_order_symbol(char *sym)
899 {
900 	collundef_t *c;
901 	if ((c = get_collundef(sym)) == NULL) {
902 		INTERR;
903 		return;
904 	}
905 	add_order_pri(c->ref[curr_weight]);
906 }
907 
908 void
909 add_order_ellipsis(void)
910 {
911 	/* special NULL value indicates self reference */
912 	add_order_pri(0);
913 }
914 
915 void
916 add_order_subst(void)
917 {
918 	subst_t srch;
919 	subst_t	*s;
920 	int i;
921 
922 	(void) memset(&srch, 0, sizeof (srch));
923 	for (i = 0; i < curr_subst; i++) {
924 		srch.ref[i] = subst_weights[i];
925 		subst_weights[i] = 0;
926 	}
927 	s = RB_FIND(substs_ref, &substs_ref[curr_weight], &srch);
928 
929 	if (s == NULL) {
930 		if ((s = calloc(1, sizeof(*s))) == NULL) {
931 			fprintf(stderr,"out of memory");
932 			return;
933 		}
934 		s->key = new_pri();
935 
936 		/*
937 		 * We use a self reference for our key, but we set a
938 		 * high bit to indicate that this is a substitution
939 		 * reference.  This will expedite table lookups later,
940 		 * and prevent table lookups for situations that don't
941 		 * require it.  (In short, its a big win, because we
942 		 * can skip a lot of binary searching.)
943 		 */
944 		set_pri(s->key,
945 		    (nextsubst[curr_weight] | COLLATE_SUBST_PRIORITY),
946 		    RESOLVED);
947 		nextsubst[curr_weight] += 1;
948 
949 		for (i = 0; i < curr_subst; i++) {
950 			s->ref[i] = srch.ref[i];
951 		}
952 
953 		RB_INSERT(substs_ref, &substs_ref[curr_weight], s);
954 
955 		if (RB_FIND(substs, &substs[curr_weight], s) != NULL) {
956 			INTERR;
957 			return;
958 		}
959 		RB_INSERT(substs, &substs[curr_weight], s);
960 	}
961 	curr_subst = 0;
962 
963 
964 	/*
965 	 * We are using the current (unique) priority as a search key
966 	 * in the substitution table.
967 	 */
968 	add_order_pri(s->key);
969 }
970 
971 static void
972 add_subst_pri(int32_t ref)
973 {
974 	if (curr_subst >= COLLATE_STR_LEN) {
975 		fprintf(stderr,"substitution string is too long");
976 		return;
977 	}
978 	subst_weights[curr_subst] = ref;
979 	curr_subst++;
980 }
981 
982 void
983 add_subst_char(wchar_t wc)
984 {
985 	collchar_t *cc;
986 
987 
988 	if (((cc = get_collchar(wc, 1)) == NULL) ||
989 	    (cc->wc != wc)) {
990 		INTERR;
991 		return;
992 	}
993 	/* we take the weight for the character at that position */
994 	add_subst_pri(cc->ref[curr_weight]);
995 }
996 
997 void
998 add_subst_collelem(collelem_t *e)
999 {
1000 	add_subst_pri(e->ref[curr_weight]);
1001 }
1002 
1003 void
1004 add_subst_collsym(collsym_t *s)
1005 {
1006 	add_subst_pri(s->ref);
1007 }
1008 
1009 void
1010 add_subst_symbol(char *ptr)
1011 {
1012 	collundef_t *cu;
1013 
1014 	if ((cu = get_collundef(ptr)) != NULL) {
1015 		add_subst_pri(cu->ref[curr_weight]);
1016 	}
1017 }
1018 
1019 void
1020 add_weight(int32_t ref, int pass)
1021 {
1022 	weight_t srch;
1023 	weight_t *w;
1024 
1025 	srch.pri = resolve_pri(ref);
1026 
1027 	/* No translation of ignores */
1028 	if (srch.pri == 0)
1029 		return;
1030 
1031 	/* Substitution priorities are not weights */
1032 	if (srch.pri & COLLATE_SUBST_PRIORITY)
1033 		return;
1034 
1035 	if (RB_FIND(weights, &weights[pass], &srch) != NULL)
1036 		return;
1037 
1038 	if ((w = calloc(1, sizeof(*w))) == NULL) {
1039 		fprintf(stderr, "out of memory");
1040 		return;
1041 	}
1042 	w->pri = srch.pri;
1043 	RB_INSERT(weights, &weights[pass], w);
1044 }
1045 
1046 void
1047 add_weights(int32_t *refs)
1048 {
1049 	int i;
1050 	for (i = 0; i < NUM_WT; i++) {
1051 		add_weight(refs[i], i);
1052 	}
1053 }
1054 
1055 int32_t
1056 get_weight(int32_t ref, int pass)
1057 {
1058 	weight_t	srch;
1059 	weight_t	*w;
1060 	int32_t		pri;
1061 
1062 	pri = resolve_pri(ref);
1063 	if (pri & COLLATE_SUBST_PRIORITY) {
1064 		return (pri);
1065 	}
1066 	if (pri <= 0) {
1067 		return (pri);
1068 	}
1069 	srch.pri = pri;
1070 	if ((w = RB_FIND(weights, &weights[pass], &srch)) == NULL) {
1071 		INTERR;
1072 		return (-1);
1073 	}
1074 	return (w->opt);
1075 }
1076 
1077 wchar_t *
1078 wsncpy(wchar_t *s1, const wchar_t *s2, size_t n)
1079 {
1080 	wchar_t *os1 = s1;
1081 
1082 	n++;
1083 	while (--n > 0 && (*s1++ = *s2++) != 0)
1084 		continue;
1085 	if (n > 0)
1086 		while (--n > 0)
1087 			*s1++ = 0;
1088 	return (os1);
1089 }
1090 
1091 #define RB_COUNT(x, name, head, cnt) do { \
1092 	(cnt) = 0; \
1093 	RB_FOREACH(x, name, (head)) { \
1094 		(cnt)++; \
1095 	} \
1096 } while (0)
1097 
1098 #define RB_NUMNODES(type, name, head, cnt) do { \
1099 	type *t; \
1100 	cnt = 0; \
1101 	RB_FOREACH(t, name, head) { \
1102 		cnt++; \
1103 	} \
1104 } while (0)
1105 
1106 void
1107 dump_collate(void)
1108 {
1109 	FILE			*f;
1110 	int			i, j, n;
1111 	size_t			sz;
1112 	int32_t			pri;
1113 	collelem_t		*ce;
1114 	collchar_t		*cc;
1115 	subst_t			*sb;
1116 	char			vers[COLLATE_STR_LEN];
1117 	collate_char_t		chars[UCHAR_MAX + 1];
1118 	collate_large_t		*large;
1119 	collate_subst_t		*subst[COLL_WEIGHTS_MAX];
1120 	collate_chain_t		*chain;
1121 
1122 	/*
1123 	 * We have to run through a preliminary pass to identify all the
1124 	 * weights that we use for each sorting level.
1125 	 */
1126 	for (i = 0; i < NUM_WT; i++) {
1127 		add_weight(pri_ignore, i);
1128 	}
1129 	for (i = 0; i < NUM_WT; i++) {
1130 		RB_FOREACH(sb, substs, &substs[i]) {
1131 			for (j = 0; sb->ref[j]; j++) {
1132 				add_weight(sb->ref[j], i);
1133 			}
1134 		}
1135 	}
1136 	RB_FOREACH(ce, elem_by_expand, &elem_by_expand) {
1137 		add_weights(ce->ref);
1138 	}
1139 	RB_FOREACH(cc, collchars, &collchars) {
1140 		add_weights(cc->ref);
1141 	}
1142 
1143 	/*
1144 	 * Now we walk the entire set of weights, removing the gaps
1145 	 * in the weights.  This gives us optimum usage.  The walk
1146 	 * occurs in priority.
1147 	 */
1148 	for (i = 0; i < NUM_WT; i++) {
1149 		weight_t *w;
1150 		RB_FOREACH(w, weights, &weights[i]) {
1151 			w->opt = nweight[i];
1152 			nweight[i] += 1;
1153 		}
1154 	}
1155 
1156 	(void) memset(&chars, 0, sizeof (chars));
1157 	(void) memset(vers, 0, COLLATE_STR_LEN);
1158 	(void) strlcpy(vers, COLLATE_VERSION, sizeof (vers));
1159 
1160 	/*
1161 	 * We need to make sure we arrange for the UNDEFINED field
1162 	 * to show up.  Also, set the total weight counts.
1163 	 */
1164 	for (i = 0; i < NUM_WT; i++) {
1165 		if (resolve_pri(pri_undefined[i]) == -1) {
1166 			set_pri(pri_undefined[i], -1, RESOLVED);
1167 			/* they collate at the end of everything else */
1168 			collinfo.undef_pri[i] = COLLATE_MAX_PRIORITY;
1169 		}
1170 		collinfo.pri_count[i] = nweight[i];
1171 	}
1172 
1173 	collinfo.pri_count[NUM_WT] = max_wide();
1174 	collinfo.undef_pri[NUM_WT] = COLLATE_MAX_PRIORITY;
1175 	collinfo.directive[NUM_WT] = DIRECTIVE_UNDEFINED;
1176 
1177 	/*
1178 	 * Ordinary character priorities
1179 	 */
1180 	for (i = 0; i <= UCHAR_MAX; i++) {
1181 		if ((cc = get_collchar(i, 0)) != NULL) {
1182 			for (j = 0; j < NUM_WT; j++) {
1183 				chars[i].pri[j] = get_weight(cc->ref[j], j);
1184 			}
1185 		} else {
1186 			for (j = 0; j < NUM_WT; j++) {
1187 				chars[i].pri[j] =
1188 				    get_weight(pri_undefined[j], j);
1189 			}
1190 			/*
1191 			 * Per POSIX, for undefined characters, we
1192 			 * also have to add a last item, which is the
1193 			 * character code.
1194 			 */
1195 			chars[i].pri[NUM_WT] = i;
1196 		}
1197 	}
1198 
1199 	/*
1200 	 * Substitution tables
1201 	 */
1202 	for (i = 0; i < NUM_WT; i++) {
1203 		collate_subst_t *st = NULL;
1204 		subst_t *temp;
1205 		RB_COUNT(temp, substs, &substs[i], n);
1206 		collinfo.subst_count[i] = n;
1207 		if ((st = calloc(n, sizeof(collate_subst_t))) == NULL) {
1208 			fprintf(stderr, "out of memory");
1209 			return;
1210 		}
1211 		n = 0;
1212 		RB_FOREACH(sb, substs, &substs[i]) {
1213 			if ((st[n].key = resolve_pri(sb->key)) < 0) {
1214 				/* by definition these resolve! */
1215 				INTERR;
1216 			}
1217 			if (st[n].key != (n | COLLATE_SUBST_PRIORITY)) {
1218 				INTERR;
1219 			}
1220 			for (j = 0; sb->ref[j]; j++) {
1221 				st[n].pri[j] = get_weight(sb->ref[j], i);
1222 			}
1223 			n++;
1224 		}
1225 		if (n != collinfo.subst_count[i])
1226 			INTERR;
1227 		subst[i] = st;
1228 	}
1229 
1230 
1231 	/*
1232 	 * Chains, i.e. collating elements
1233 	 */
1234 	RB_NUMNODES(collelem_t, elem_by_expand, &elem_by_expand,
1235 	    collinfo.chain_count);
1236 	chain = calloc(collinfo.chain_count, sizeof(collate_chain_t));
1237 	if (chain == NULL) {
1238 		fprintf(stderr, "out of memory");
1239 		return;
1240 	}
1241 	n = 0;
1242 	RB_FOREACH(ce, elem_by_expand, &elem_by_expand) {
1243 		(void) wsncpy(chain[n].str, ce->expand, COLLATE_STR_LEN);
1244 		for (i = 0; i < NUM_WT; i++) {
1245 			chain[n].pri[i] = get_weight(ce->ref[i], i);
1246 		}
1247 		n++;
1248 	}
1249 	if (n != collinfo.chain_count)
1250 		INTERR;
1251 
1252 	/*
1253 	 * Large (> UCHAR_MAX) character priorities
1254 	 */
1255 	RB_NUMNODES(collchar_t, collchars, &collchars, n);
1256 	large = calloc(n, sizeof(collate_large_t));
1257 	if (large == NULL) {
1258 		fprintf(stderr, "out of memory");
1259 		return;
1260 	}
1261 
1262 	i = 0;
1263 	RB_FOREACH(cc, collchars, &collchars) {
1264 		int	undef = 0;
1265 		/* we already gathered those */
1266 		if (cc->wc <= UCHAR_MAX)
1267 			continue;
1268 		for (j = 0; j < NUM_WT; j++) {
1269 			if ((pri = get_weight(cc->ref[j], j)) < 0) {
1270 				undef = 1;
1271 			}
1272 			if (undef && (pri >= 0)) {
1273 				/* if undefined, then all priorities are */
1274 				INTERR;
1275 			} else {
1276 				large[i].pri.pri[j] = pri;
1277 			}
1278 		}
1279 		if (!undef) {
1280 			large[i].val = cc->wc;
1281 			collinfo.large_count = i++;
1282 		}
1283 	}
1284 
1285 	if ((f = open_category()) == NULL) {
1286 		return;
1287 	}
1288 
1289 	/* Time to write the entire data set out */
1290 
1291 	if ((wr_category(vers, COLLATE_STR_LEN, f) < 0) ||
1292 	    (wr_category(&collinfo, sizeof (collinfo), f) < 0) ||
1293 	    (wr_category(&chars, sizeof (chars), f) < 0)) {
1294 		return;
1295 	}
1296 
1297 	for (i = 0; i < NUM_WT; i++) {
1298 		sz =  sizeof (collate_subst_t) * collinfo.subst_count[i];
1299 		if (wr_category(subst[i], sz, f) < 0) {
1300 			return;
1301 		}
1302 	}
1303 	sz = sizeof (collate_chain_t) * collinfo.chain_count;
1304 	if (wr_category(chain, sz, f) < 0) {
1305 		return;
1306 	}
1307 	sz = sizeof (collate_large_t) * collinfo.large_count;
1308 	if (wr_category(large, sz, f) < 0) {
1309 		return;
1310 	}
1311 
1312 	close_category(f);
1313 }
1314