xref: /freebsd/usr.bin/localedef/collate.c (revision b9f654b163bce26de79705e77b872427c9f2afa1)
1 /*-
2  * Copyright 2018 Nexenta Systems, Inc.
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 static int32_t subst_count[COLL_WEIGHTS_MAX];
257 static int32_t chain_count;
258 static int32_t large_count;
259 
260 static collpri_t	*prilist = NULL;
261 static int		numpri = 0;
262 static int		maxpri = 0;
263 
264 static void start_order(int);
265 
266 static int32_t
267 new_pri(void)
268 {
269 	int i;
270 
271 	if (numpri >= maxpri) {
272 		maxpri = maxpri ? maxpri * 2 : 1024;
273 		prilist = realloc(prilist, sizeof (collpri_t) * maxpri);
274 		if (prilist == NULL) {
275 			fprintf(stderr,"out of memory");
276 			return (-1);
277 		}
278 		for (i = numpri; i < maxpri; i++) {
279 			prilist[i].res = UNKNOWN;
280 			prilist[i].pri = 0;
281 			prilist[i].pass = 0;
282 		}
283 	}
284 	return (numpri++);
285 }
286 
287 static collpri_t *
288 get_pri(int32_t ref)
289 {
290 	if ((ref < 0) || (ref > numpri)) {
291 		INTERR;
292 		return (NULL);
293 	}
294 	return (&prilist[ref]);
295 }
296 
297 static void
298 set_pri(int32_t ref, int32_t v, res_t res)
299 {
300 	collpri_t	*pri;
301 
302 	pri = get_pri(ref);
303 
304 	if ((res == REFER) && ((v < 0) || (v >= numpri))) {
305 		INTERR;
306 	}
307 
308 	/* Resolve self references */
309 	if ((res == REFER) && (ref == v)) {
310 		v = nextpri;
311 		res = RESOLVED;
312 	}
313 
314 	if (pri->res != UNKNOWN) {
315 		warn("repeated item in order list (first on %d)",
316 		    pri->lineno);
317 		return;
318 	}
319 	pri->lineno = lineno;
320 	pri->pri = v;
321 	pri->res = res;
322 }
323 
324 static int32_t
325 resolve_pri(int32_t ref)
326 {
327 	collpri_t	*pri;
328 	static int32_t	pass = 0;
329 
330 	pri = get_pri(ref);
331 	pass++;
332 	while (pri->res == REFER) {
333 		if (pri->pass == pass) {
334 			/* report a line with the circular symbol */
335 			lineno = pri->lineno;
336 			fprintf(stderr,"circular reference in order list");
337 			return (-1);
338 		}
339 		if ((pri->pri < 0) || (pri->pri >= numpri)) {
340 			INTERR;
341 			return (-1);
342 		}
343 		pri->pass = pass;
344 		pri = &prilist[pri->pri];
345 	}
346 
347 	if (pri->res == UNKNOWN) {
348 		return (-1);
349 	}
350 	if (pri->res != RESOLVED)
351 		INTERR;
352 
353 	return (pri->pri);
354 }
355 
356 static int
357 weight_compare(const void *n1, const void *n2)
358 {
359 	int32_t	k1 = ((const weight_t *)n1)->pri;
360 	int32_t	k2 = ((const weight_t *)n2)->pri;
361 
362 	return (k1 < k2 ? -1 : k1 > k2 ? 1 : 0);
363 }
364 
365 RB_GENERATE_STATIC(weights, weight, entry, weight_compare);
366 
367 static int
368 collsym_compare(const void *n1, const void *n2)
369 {
370 	const collsym_t *c1 = n1;
371 	const collsym_t *c2 = n2;
372 	int rv;
373 
374 	rv = strcmp(c1->name, c2->name);
375 	return ((rv < 0) ? -1 : (rv > 0) ? 1 : 0);
376 }
377 
378 RB_GENERATE_STATIC(collsyms, collsym, entry, collsym_compare);
379 
380 static int
381 collundef_compare(const void *n1, const void *n2)
382 {
383 	const collundef_t *c1 = n1;
384 	const collundef_t *c2 = n2;
385 	int rv;
386 
387 	rv = strcmp(c1->name, c2->name);
388 	return ((rv < 0) ? -1 : (rv > 0) ? 1 : 0);
389 }
390 
391 RB_GENERATE_STATIC(collundefs, collundef, entry, collundef_compare);
392 
393 static int
394 element_compare_symbol(const void *n1, const void *n2)
395 {
396 	const collelem_t *c1 = n1;
397 	const collelem_t *c2 = n2;
398 	int rv;
399 
400 	rv = strcmp(c1->symbol, c2->symbol);
401 	return ((rv < 0) ? -1 : (rv > 0) ? 1 : 0);
402 }
403 
404 RB_GENERATE_STATIC(elem_by_symbol, collelem, rb_bysymbol, element_compare_symbol);
405 
406 static int
407 element_compare_expand(const void *n1, const void *n2)
408 {
409 	const collelem_t *c1 = n1;
410 	const collelem_t *c2 = n2;
411 	int rv;
412 
413 	rv = wcscmp(c1->expand, c2->expand);
414 	return ((rv < 0) ? -1 : (rv > 0) ? 1 : 0);
415 }
416 
417 RB_GENERATE_STATIC(elem_by_expand, collelem, rb_byexpand, element_compare_expand);
418 
419 static int
420 collchar_compare(const void *n1, const void *n2)
421 {
422 	wchar_t	k1 = ((const collchar_t *)n1)->wc;
423 	wchar_t	k2 = ((const collchar_t *)n2)->wc;
424 
425 	return (k1 < k2 ? -1 : k1 > k2 ? 1 : 0);
426 }
427 
428 RB_GENERATE_STATIC(collchars, collchar, entry, collchar_compare);
429 
430 static int
431 subst_compare(const void *n1, const void *n2)
432 {
433 	int32_t	k1 = ((const subst_t *)n1)->key;
434 	int32_t	k2 = ((const subst_t *)n2)->key;
435 
436 	return (k1 < k2 ? -1 : k1 > k2 ? 1 : 0);
437 }
438 
439 RB_GENERATE_STATIC(substs, subst, entry, subst_compare);
440 
441 static int
442 subst_compare_ref(const void *n1, const void *n2)
443 {
444 	const wchar_t *c1 = ((const subst_t *)n1)->ref;
445 	const wchar_t *c2 = ((const subst_t *)n2)->ref;
446 	int rv;
447 
448 	rv = wcscmp(c1, c2);
449 	return ((rv < 0) ? -1 : (rv > 0) ? 1 : 0);
450 }
451 
452 RB_GENERATE_STATIC(substs_ref, subst, entry_ref, subst_compare_ref);
453 
454 void
455 init_collate(void)
456 {
457 	int i;
458 
459 	RB_INIT(&collsyms);
460 
461 	RB_INIT(&collundefs);
462 
463 	RB_INIT(&elem_by_symbol);
464 
465 	RB_INIT(&elem_by_expand);
466 
467 	RB_INIT(&collchars);
468 
469 	for (i = 0; i < COLL_WEIGHTS_MAX; i++) {
470 		RB_INIT(&substs[i]);
471 		RB_INIT(&substs_ref[i]);
472 		RB_INIT(&weights[i]);
473 		nweight[i] = 1;
474 	}
475 
476 	(void) memset(&collinfo, 0, sizeof (collinfo));
477 
478 	/* allocate some initial priorities */
479 	pri_ignore = new_pri();
480 
481 	set_pri(pri_ignore, 0, RESOLVED);
482 
483 	for (i = 0; i < COLL_WEIGHTS_MAX; i++) {
484 		pri_undefined[i] = new_pri();
485 
486 		/* we will override this later */
487 		set_pri(pri_undefined[i], COLLATE_MAX_PRIORITY, UNKNOWN);
488 	}
489 }
490 
491 void
492 define_collsym(char *name)
493 {
494 	collsym_t	*sym;
495 
496 	if ((sym = calloc(1, sizeof(*sym))) == NULL) {
497 		fprintf(stderr,"out of memory");
498 		return;
499 	}
500 	sym->name = name;
501 	sym->ref = new_pri();
502 
503 	if (RB_FIND(collsyms, &collsyms, sym) != NULL) {
504 		/*
505 		 * This should never happen because we are only called
506 		 * for undefined symbols.
507 		 */
508 		free(sym);
509 		INTERR;
510 		return;
511 	}
512 	RB_INSERT(collsyms, &collsyms, sym);
513 }
514 
515 collsym_t *
516 lookup_collsym(char *name)
517 {
518 	collsym_t	srch;
519 
520 	srch.name = name;
521 	return (RB_FIND(collsyms, &collsyms, &srch));
522 }
523 
524 collelem_t *
525 lookup_collelem(char *symbol)
526 {
527 	collelem_t	srch;
528 
529 	srch.symbol = symbol;
530 	return (RB_FIND(elem_by_symbol, &elem_by_symbol, &srch));
531 }
532 
533 static collundef_t *
534 get_collundef(char *name)
535 {
536 	collundef_t	srch;
537 	collundef_t	*ud;
538 	int		i;
539 
540 	srch.name = name;
541 	if ((ud = RB_FIND(collundefs, &collundefs, &srch)) == NULL) {
542 		if (((ud = calloc(1, sizeof(*ud))) == NULL) ||
543 		    ((ud->name = strdup(name)) == NULL)) {
544 			fprintf(stderr,"out of memory");
545 			free(ud);
546 			return (NULL);
547 		}
548 		for (i = 0; i < NUM_WT; i++) {
549 			ud->ref[i] = new_pri();
550 		}
551 		RB_INSERT(collundefs, &collundefs, ud);
552 	}
553 	add_charmap_undefined(name);
554 	return (ud);
555 }
556 
557 static collchar_t *
558 get_collchar(wchar_t wc, int create)
559 {
560 	collchar_t	srch;
561 	collchar_t	*cc;
562 	int		i;
563 
564 	srch.wc = wc;
565 	cc = RB_FIND(collchars, &collchars, &srch);
566 	if ((cc == NULL) && create) {
567 		if ((cc = calloc(1, sizeof(*cc))) == NULL) {
568 			fprintf(stderr, "out of memory");
569 			return (NULL);
570 		}
571 		for (i = 0; i < NUM_WT; i++) {
572 			cc->ref[i] = new_pri();
573 		}
574 		cc->wc = wc;
575 		RB_INSERT(collchars, &collchars, cc);
576 	}
577 	return (cc);
578 }
579 
580 void
581 end_order_collsym(collsym_t *sym)
582 {
583 	start_order(T_COLLSYM);
584 	/* update the weight */
585 
586 	set_pri(sym->ref, nextpri, RESOLVED);
587 	nextpri++;
588 }
589 
590 void
591 end_order(void)
592 {
593 	int		i;
594 	int32_t		pri;
595 	int32_t		ref;
596 	collpri_t	*p;
597 
598 	/* advance the priority/weight */
599 	pri = nextpri;
600 
601 	switch (currorder) {
602 	case T_CHAR:
603 		for (i = 0; i < NUM_WT; i++) {
604 			if (((ref = order_weights[i]) < 0) ||
605 			    ((p = get_pri(ref)) == NULL) ||
606 			    (p->pri == -1)) {
607 				/* unspecified weight is a self reference */
608 				set_pri(currchar->ref[i], pri, RESOLVED);
609 			} else {
610 				set_pri(currchar->ref[i], ref, REFER);
611 			}
612 			order_weights[i] = -1;
613 		}
614 
615 		/* leave a cookie trail in case next symbol is ellipsis */
616 		ellipsis_start = currchar->wc + 1;
617 		currchar = NULL;
618 		break;
619 
620 	case T_ELLIPSIS:
621 		/* save off the weights were we can find them */
622 		for (i = 0; i < NUM_WT; i++) {
623 			ellipsis_weights[i] = order_weights[i];
624 			order_weights[i] = -1;
625 		}
626 		break;
627 
628 	case T_COLLELEM:
629 		if (currelem == NULL) {
630 			INTERR;
631 		} else {
632 			for (i = 0; i < NUM_WT; i++) {
633 
634 				if (((ref = order_weights[i]) < 0) ||
635 				    ((p = get_pri(ref)) == NULL) ||
636 				    (p->pri == -1)) {
637 					set_pri(currelem->ref[i], pri,
638 					    RESOLVED);
639 				} else {
640 					set_pri(currelem->ref[i], ref, REFER);
641 				}
642 				order_weights[i] = -1;
643 			}
644 		}
645 		break;
646 
647 	case T_UNDEFINED:
648 		for (i = 0; i < NUM_WT; i++) {
649 			if (((ref = order_weights[i]) < 0) ||
650 			    ((p = get_pri(ref)) == NULL) ||
651 			    (p->pri == -1)) {
652 				set_pri(pri_undefined[i], -1, RESOLVED);
653 			} else {
654 				set_pri(pri_undefined[i], ref, REFER);
655 			}
656 			order_weights[i] = -1;
657 		}
658 		break;
659 
660 	case T_SYMBOL:
661 		for (i = 0; i < NUM_WT; i++) {
662 			if (((ref = order_weights[i]) < 0) ||
663 			    ((p = get_pri(ref)) == NULL) ||
664 			    (p->pri == -1)) {
665 				set_pri(currundef->ref[i], pri, RESOLVED);
666 			} else {
667 				set_pri(currundef->ref[i], ref, REFER);
668 			}
669 			order_weights[i] = -1;
670 		}
671 		break;
672 
673 	default:
674 		INTERR;
675 	}
676 
677 	nextpri++;
678 }
679 
680 static void
681 start_order(int type)
682 {
683 	int	i;
684 
685 	lastorder = currorder;
686 	currorder = type;
687 
688 	/* this is used to protect ELLIPSIS processing */
689 	if ((lastorder == T_ELLIPSIS) && (type != T_CHAR)) {
690 		fprintf(stderr, "character value expected");
691 	}
692 
693 	for (i = 0; i < COLL_WEIGHTS_MAX; i++) {
694 		order_weights[i] = -1;
695 	}
696 	curr_weight = 0;
697 }
698 
699 void
700 start_order_undefined(void)
701 {
702 	start_order(T_UNDEFINED);
703 }
704 
705 void
706 start_order_symbol(char *name)
707 {
708 	currundef = get_collundef(name);
709 	start_order(T_SYMBOL);
710 }
711 
712 void
713 start_order_char(wchar_t wc)
714 {
715 	collchar_t	*cc;
716 	int32_t		ref;
717 
718 	start_order(T_CHAR);
719 
720 	/*
721 	 * If we last saw an ellipsis, then we need to close the range.
722 	 * Handle that here.  Note that we have to be careful because the
723 	 * items *inside* the range are treated exclusiveley to the items
724 	 * outside of the range.  The ends of the range can have quite
725 	 * different weights than the range members.
726 	 */
727 	if (lastorder == T_ELLIPSIS) {
728 		int		i;
729 
730 		if (wc < ellipsis_start) {
731 			fprintf(stderr, "malformed range!");
732 			return;
733 		}
734 		while (ellipsis_start < wc) {
735 			/*
736 			 * pick all of the saved weights for the
737 			 * ellipsis.  note that -1 encodes for the
738 			 * ellipsis itself, which means to take the
739 			 * current relative priority.
740 			 */
741 			if ((cc = get_collchar(ellipsis_start, 1)) == NULL) {
742 				INTERR;
743 				return;
744 			}
745 			for (i = 0; i < NUM_WT; i++) {
746 				collpri_t *p;
747 				if (((ref = ellipsis_weights[i]) == -1) ||
748 				    ((p = get_pri(ref)) == NULL) ||
749 				    (p->pri == -1)) {
750 					set_pri(cc->ref[i], nextpri, RESOLVED);
751 				} else {
752 					set_pri(cc->ref[i], ref, REFER);
753 				}
754 				ellipsis_weights[i] = 0;
755 			}
756 			ellipsis_start++;
757 			nextpri++;
758 		}
759 	}
760 
761 	currchar = get_collchar(wc, 1);
762 }
763 
764 void
765 start_order_collelem(collelem_t *e)
766 {
767 	start_order(T_COLLELEM);
768 	currelem = e;
769 }
770 
771 void
772 start_order_ellipsis(void)
773 {
774 	int	i;
775 
776 	start_order(T_ELLIPSIS);
777 
778 	if (lastorder != T_CHAR) {
779 		fprintf(stderr, "illegal starting point for range");
780 		return;
781 	}
782 
783 	for (i = 0; i < NUM_WT; i++) {
784 		ellipsis_weights[i] = order_weights[i];
785 	}
786 }
787 
788 void
789 define_collelem(char *name, wchar_t *wcs)
790 {
791 	collelem_t	*e;
792 	int		i;
793 
794 	if (wcslen(wcs) >= COLLATE_STR_LEN) {
795 		fprintf(stderr,"expanded collation element too long");
796 		return;
797 	}
798 
799 	if ((e = calloc(1, sizeof(*e))) == NULL) {
800 		fprintf(stderr, "out of memory");
801 		return;
802 	}
803 	e->expand = wcs;
804 	e->symbol = name;
805 
806 	/*
807 	 * This is executed before the order statement, so we don't
808 	 * know how many priorities we *really* need.  We allocate one
809 	 * for each possible weight.  Not a big deal, as collating-elements
810 	 * prove to be quite rare.
811 	 */
812 	for (i = 0; i < COLL_WEIGHTS_MAX; i++) {
813 		e->ref[i] = new_pri();
814 	}
815 
816 	/* A character sequence can only reduce to one element. */
817 	if ((RB_FIND(elem_by_symbol, &elem_by_symbol, e) != NULL) ||
818 	    (RB_FIND(elem_by_expand, &elem_by_expand, e) != NULL)) {
819 		fprintf(stderr, "duplicate collating element definition");
820 		free(e);
821 		return;
822 	}
823 	RB_INSERT(elem_by_symbol, &elem_by_symbol, e);
824 	RB_INSERT(elem_by_expand, &elem_by_expand, e);
825 }
826 
827 void
828 add_order_bit(int kw)
829 {
830 	uint8_t bit = DIRECTIVE_UNDEF;
831 
832 	switch (kw) {
833 	case T_FORWARD:
834 		bit = DIRECTIVE_FORWARD;
835 		break;
836 	case T_BACKWARD:
837 		bit = DIRECTIVE_BACKWARD;
838 		break;
839 	case T_POSITION:
840 		bit = DIRECTIVE_POSITION;
841 		break;
842 	default:
843 		INTERR;
844 		break;
845 	}
846 	collinfo.directive[collinfo.directive_count] |= bit;
847 }
848 
849 void
850 add_order_directive(void)
851 {
852 	if (collinfo.directive_count >= COLL_WEIGHTS_MAX) {
853 		fprintf(stderr,"too many directives (max %d)", COLL_WEIGHTS_MAX);
854 	}
855 	collinfo.directive_count++;
856 }
857 
858 static void
859 add_order_pri(int32_t ref)
860 {
861 	if (curr_weight >= NUM_WT) {
862 		fprintf(stderr,"too many weights (max %d)", NUM_WT);
863 		return;
864 	}
865 	order_weights[curr_weight] = ref;
866 	curr_weight++;
867 }
868 
869 void
870 add_order_collsym(collsym_t *s)
871 {
872 	add_order_pri(s->ref);
873 }
874 
875 void
876 add_order_char(wchar_t wc)
877 {
878 	collchar_t *cc;
879 
880 	if ((cc = get_collchar(wc, 1)) == NULL) {
881 		INTERR;
882 		return;
883 	}
884 
885 	add_order_pri(cc->ref[curr_weight]);
886 }
887 
888 void
889 add_order_collelem(collelem_t *e)
890 {
891 	add_order_pri(e->ref[curr_weight]);
892 }
893 
894 void
895 add_order_ignore(void)
896 {
897 	add_order_pri(pri_ignore);
898 }
899 
900 void
901 add_order_symbol(char *sym)
902 {
903 	collundef_t *c;
904 	if ((c = get_collundef(sym)) == NULL) {
905 		INTERR;
906 		return;
907 	}
908 	add_order_pri(c->ref[curr_weight]);
909 }
910 
911 void
912 add_order_ellipsis(void)
913 {
914 	/* special NULL value indicates self reference */
915 	add_order_pri(0);
916 }
917 
918 void
919 add_order_subst(void)
920 {
921 	subst_t srch;
922 	subst_t	*s;
923 	int i;
924 
925 	(void) memset(&srch, 0, sizeof (srch));
926 	for (i = 0; i < curr_subst; i++) {
927 		srch.ref[i] = subst_weights[i];
928 		subst_weights[i] = 0;
929 	}
930 	s = RB_FIND(substs_ref, &substs_ref[curr_weight], &srch);
931 
932 	if (s == NULL) {
933 		if ((s = calloc(1, sizeof(*s))) == NULL) {
934 			fprintf(stderr,"out of memory");
935 			return;
936 		}
937 		s->key = new_pri();
938 
939 		/*
940 		 * We use a self reference for our key, but we set a
941 		 * high bit to indicate that this is a substitution
942 		 * reference.  This will expedite table lookups later,
943 		 * and prevent table lookups for situations that don't
944 		 * require it.  (In short, its a big win, because we
945 		 * can skip a lot of binary searching.)
946 		 */
947 		set_pri(s->key,
948 		    (nextsubst[curr_weight] | COLLATE_SUBST_PRIORITY),
949 		    RESOLVED);
950 		nextsubst[curr_weight] += 1;
951 
952 		for (i = 0; i < curr_subst; i++) {
953 			s->ref[i] = srch.ref[i];
954 		}
955 
956 		RB_INSERT(substs_ref, &substs_ref[curr_weight], s);
957 
958 		if (RB_FIND(substs, &substs[curr_weight], s) != NULL) {
959 			INTERR;
960 			return;
961 		}
962 		RB_INSERT(substs, &substs[curr_weight], s);
963 	}
964 	curr_subst = 0;
965 
966 
967 	/*
968 	 * We are using the current (unique) priority as a search key
969 	 * in the substitution table.
970 	 */
971 	add_order_pri(s->key);
972 }
973 
974 static void
975 add_subst_pri(int32_t ref)
976 {
977 	if (curr_subst >= COLLATE_STR_LEN) {
978 		fprintf(stderr,"substitution string is too long");
979 		return;
980 	}
981 	subst_weights[curr_subst] = ref;
982 	curr_subst++;
983 }
984 
985 void
986 add_subst_char(wchar_t wc)
987 {
988 	collchar_t *cc;
989 
990 
991 	if (((cc = get_collchar(wc, 1)) == NULL) ||
992 	    (cc->wc != wc)) {
993 		INTERR;
994 		return;
995 	}
996 	/* we take the weight for the character at that position */
997 	add_subst_pri(cc->ref[curr_weight]);
998 }
999 
1000 void
1001 add_subst_collelem(collelem_t *e)
1002 {
1003 	add_subst_pri(e->ref[curr_weight]);
1004 }
1005 
1006 void
1007 add_subst_collsym(collsym_t *s)
1008 {
1009 	add_subst_pri(s->ref);
1010 }
1011 
1012 void
1013 add_subst_symbol(char *ptr)
1014 {
1015 	collundef_t *cu;
1016 
1017 	if ((cu = get_collundef(ptr)) != NULL) {
1018 		add_subst_pri(cu->ref[curr_weight]);
1019 	}
1020 }
1021 
1022 void
1023 add_weight(int32_t ref, int pass)
1024 {
1025 	weight_t srch;
1026 	weight_t *w;
1027 
1028 	srch.pri = resolve_pri(ref);
1029 
1030 	/* No translation of ignores */
1031 	if (srch.pri == 0)
1032 		return;
1033 
1034 	/* Substitution priorities are not weights */
1035 	if (srch.pri & COLLATE_SUBST_PRIORITY)
1036 		return;
1037 
1038 	if (RB_FIND(weights, &weights[pass], &srch) != NULL)
1039 		return;
1040 
1041 	if ((w = calloc(1, sizeof(*w))) == NULL) {
1042 		fprintf(stderr, "out of memory");
1043 		return;
1044 	}
1045 	w->pri = srch.pri;
1046 	RB_INSERT(weights, &weights[pass], w);
1047 }
1048 
1049 void
1050 add_weights(int32_t *refs)
1051 {
1052 	int i;
1053 	for (i = 0; i < NUM_WT; i++) {
1054 		add_weight(refs[i], i);
1055 	}
1056 }
1057 
1058 int32_t
1059 get_weight(int32_t ref, int pass)
1060 {
1061 	weight_t	srch;
1062 	weight_t	*w;
1063 	int32_t		pri;
1064 
1065 	pri = resolve_pri(ref);
1066 	if (pri & COLLATE_SUBST_PRIORITY) {
1067 		return (pri);
1068 	}
1069 	if (pri <= 0) {
1070 		return (pri);
1071 	}
1072 	srch.pri = pri;
1073 	if ((w = RB_FIND(weights, &weights[pass], &srch)) == NULL) {
1074 		INTERR;
1075 		return (-1);
1076 	}
1077 	return (w->opt);
1078 }
1079 
1080 wchar_t *
1081 wsncpy(wchar_t *s1, const wchar_t *s2, size_t n)
1082 {
1083 	wchar_t *os1 = s1;
1084 
1085 	n++;
1086 	while (--n > 0 && (*s1++ = htote(*s2++)) != 0)
1087 		continue;
1088 	if (n > 0)
1089 		while (--n > 0)
1090 			*s1++ = 0;
1091 	return (os1);
1092 }
1093 
1094 #define RB_COUNT(x, name, head, cnt) do { \
1095 	(cnt) = 0; \
1096 	RB_FOREACH(x, name, (head)) { \
1097 		(cnt)++; \
1098 	} \
1099 } while (0)
1100 
1101 #define RB_NUMNODES(type, name, head, cnt) do { \
1102 	type *t; \
1103 	cnt = 0; \
1104 	RB_FOREACH(t, name, head) { \
1105 		cnt++; \
1106 	} \
1107 } while (0)
1108 
1109 void
1110 dump_collate(void)
1111 {
1112 	FILE			*f;
1113 	int			i, j, n;
1114 	size_t			sz;
1115 	int32_t			pri;
1116 	collelem_t		*ce;
1117 	collchar_t		*cc;
1118 	subst_t			*sb;
1119 	char			vers[COLLATE_STR_LEN];
1120 	collate_char_t		chars[UCHAR_MAX + 1];
1121 	collate_large_t		*large;
1122 	collate_subst_t		*subst[COLL_WEIGHTS_MAX];
1123 	collate_chain_t		*chain;
1124 
1125 	/*
1126 	 * We have to run through a preliminary pass to identify all the
1127 	 * weights that we use for each sorting level.
1128 	 */
1129 	for (i = 0; i < NUM_WT; i++) {
1130 		add_weight(pri_ignore, i);
1131 	}
1132 	for (i = 0; i < NUM_WT; i++) {
1133 		RB_FOREACH(sb, substs, &substs[i]) {
1134 			for (j = 0; sb->ref[j]; j++) {
1135 				add_weight(sb->ref[j], i);
1136 			}
1137 		}
1138 	}
1139 	RB_FOREACH(ce, elem_by_expand, &elem_by_expand) {
1140 		add_weights(ce->ref);
1141 	}
1142 	RB_FOREACH(cc, collchars, &collchars) {
1143 		add_weights(cc->ref);
1144 	}
1145 
1146 	/*
1147 	 * Now we walk the entire set of weights, removing the gaps
1148 	 * in the weights.  This gives us optimum usage.  The walk
1149 	 * occurs in priority.
1150 	 */
1151 	for (i = 0; i < NUM_WT; i++) {
1152 		weight_t *w;
1153 		RB_FOREACH(w, weights, &weights[i]) {
1154 			w->opt = nweight[i];
1155 			nweight[i] += 1;
1156 		}
1157 	}
1158 
1159 	(void) memset(&chars, 0, sizeof (chars));
1160 	(void) memset(vers, 0, COLLATE_STR_LEN);
1161 	(void) strlcpy(vers, COLLATE_VERSION, sizeof (vers));
1162 
1163 	/*
1164 	 * We need to make sure we arrange for the UNDEFINED field
1165 	 * to show up.  Also, set the total weight counts.
1166 	 */
1167 	for (i = 0; i < NUM_WT; i++) {
1168 		if (resolve_pri(pri_undefined[i]) == -1) {
1169 			set_pri(pri_undefined[i], -1, RESOLVED);
1170 			/* they collate at the end of everything else */
1171 			collinfo.undef_pri[i] = htote(COLLATE_MAX_PRIORITY);
1172 		}
1173 		collinfo.pri_count[i] = htote(nweight[i]);
1174 	}
1175 
1176 	collinfo.pri_count[NUM_WT] = htote(max_wide());
1177 	collinfo.undef_pri[NUM_WT] = htote(COLLATE_MAX_PRIORITY);
1178 	collinfo.directive[NUM_WT] = DIRECTIVE_UNDEFINED;
1179 
1180 	/*
1181 	 * Ordinary character priorities
1182 	 */
1183 	for (i = 0; i <= UCHAR_MAX; i++) {
1184 		if ((cc = get_collchar(i, 0)) != NULL) {
1185 			for (j = 0; j < NUM_WT; j++) {
1186 				chars[i].pri[j] =
1187 				    htote(get_weight(cc->ref[j], j));
1188 			}
1189 		} else {
1190 			for (j = 0; j < NUM_WT; j++) {
1191 				chars[i].pri[j] =
1192 				    htote(get_weight(pri_undefined[j], j));
1193 			}
1194 			/*
1195 			 * Per POSIX, for undefined characters, we
1196 			 * also have to add a last item, which is the
1197 			 * character code.
1198 			 */
1199 			chars[i].pri[NUM_WT] = htote(i);
1200 		}
1201 	}
1202 
1203 	/*
1204 	 * Substitution tables
1205 	 */
1206 	for (i = 0; i < NUM_WT; i++) {
1207 		collate_subst_t *st = NULL;
1208 		subst_t *temp;
1209 		RB_COUNT(temp, substs, &substs[i], n);
1210 		subst_count[i] = n;
1211 		if ((st = calloc(n, sizeof(collate_subst_t))) == NULL) {
1212 			fprintf(stderr, "out of memory");
1213 			return;
1214 		}
1215 		n = 0;
1216 		RB_FOREACH(sb, substs, &substs[i]) {
1217 			if ((st[n].key = resolve_pri(sb->key)) < 0) {
1218 				/* by definition these resolve! */
1219 				INTERR;
1220 			}
1221 			if (st[n].key != (n | COLLATE_SUBST_PRIORITY)) {
1222 				INTERR;
1223 			}
1224 			st[n].key = htote(st[n].key);
1225 			for (j = 0; sb->ref[j]; j++) {
1226 				st[n].pri[j] = htote(get_weight(sb->ref[j],
1227 				    i));
1228 			}
1229 			n++;
1230 		}
1231 		if (n != subst_count[i])
1232 			INTERR;
1233 		subst[i] = st;
1234 	}
1235 
1236 
1237 	/*
1238 	 * Chains, i.e. collating elements
1239 	 */
1240 	RB_NUMNODES(collelem_t, elem_by_expand, &elem_by_expand, chain_count);
1241 	chain = calloc(chain_count, sizeof(collate_chain_t));
1242 	if (chain == NULL) {
1243 		fprintf(stderr, "out of memory");
1244 		return;
1245 	}
1246 	n = 0;
1247 	RB_FOREACH(ce, elem_by_expand, &elem_by_expand) {
1248 		(void) wsncpy(chain[n].str, ce->expand, COLLATE_STR_LEN);
1249 		for (i = 0; i < NUM_WT; i++) {
1250 			chain[n].pri[i] = htote(get_weight(ce->ref[i], i));
1251 		}
1252 		n++;
1253 	}
1254 	if (n != chain_count)
1255 		INTERR;
1256 
1257 	/*
1258 	 * Large (> UCHAR_MAX) character priorities
1259 	 */
1260 	RB_NUMNODES(collchar_t, collchars, &collchars, n);
1261 	large = calloc(n, sizeof(collate_large_t));
1262 	if (large == NULL) {
1263 		fprintf(stderr, "out of memory");
1264 		return;
1265 	}
1266 
1267 	i = 0;
1268 	RB_FOREACH(cc, collchars, &collchars) {
1269 		int	undef = 0;
1270 		/* we already gathered those */
1271 		if (cc->wc <= UCHAR_MAX)
1272 			continue;
1273 		for (j = 0; j < NUM_WT; j++) {
1274 			if ((pri = get_weight(cc->ref[j], j)) < 0) {
1275 				undef = 1;
1276 			}
1277 			if (undef && (pri >= 0)) {
1278 				/* if undefined, then all priorities are */
1279 				INTERR;
1280 			} else {
1281 				large[i].pri.pri[j] = htote(pri);
1282 			}
1283 		}
1284 		if (!undef) {
1285 			large[i].val = htote(cc->wc);
1286 			large_count = i++;
1287 		}
1288 	}
1289 
1290 	if ((f = open_category()) == NULL) {
1291 		return;
1292 	}
1293 
1294 	/* Time to write the entire data set out */
1295 
1296 	for (i = 0; i < NUM_WT; i++)
1297 		collinfo.subst_count[i] = htote(subst_count[i]);
1298 	collinfo.chain_count = htote(chain_count);
1299 	collinfo.large_count = htote(large_count);
1300 
1301 	if ((wr_category(vers, COLLATE_STR_LEN, f) < 0) ||
1302 	    (wr_category(&collinfo, sizeof (collinfo), f) < 0) ||
1303 	    (wr_category(&chars, sizeof (chars), f) < 0)) {
1304 		return;
1305 	}
1306 
1307 	for (i = 0; i < NUM_WT; i++) {
1308 		sz = sizeof (collate_subst_t) * subst_count[i];
1309 		if (wr_category(subst[i], sz, f) < 0) {
1310 			return;
1311 		}
1312 	}
1313 	sz = sizeof (collate_chain_t) * chain_count;
1314 	if (wr_category(chain, sz, f) < 0) {
1315 		return;
1316 	}
1317 	sz = sizeof (collate_large_t) * large_count;
1318 	if (wr_category(large, sz, f) < 0) {
1319 		return;
1320 	}
1321 
1322 	close_category(f);
1323 }
1324