xref: /dragonfly/contrib/libpcap/optimize.c (revision a563ca70)
1 /*
2  * Copyright (c) 1988, 1989, 1990, 1991, 1993, 1994, 1995, 1996
3  *	The Regents of the University of California.  All rights reserved.
4  *
5  * Redistribution and use in source and binary forms, with or without
6  * modification, are permitted provided that: (1) source code distributions
7  * retain the above copyright notice and this paragraph in its entirety, (2)
8  * distributions including binary code include the above copyright notice and
9  * this paragraph in its entirety in the documentation or other materials
10  * provided with the distribution, and (3) all advertising materials mentioning
11  * features or use of this software display the following acknowledgement:
12  * ``This product includes software developed by the University of California,
13  * Lawrence Berkeley Laboratory and its contributors.'' Neither the name of
14  * the University nor the names of its contributors may be used to endorse
15  * or promote products derived from this software without specific prior
16  * written permission.
17  * THIS SOFTWARE IS PROVIDED ``AS IS'' AND WITHOUT ANY EXPRESS OR IMPLIED
18  * WARRANTIES, INCLUDING, WITHOUT LIMITATION, THE IMPLIED WARRANTIES OF
19  * MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE.
20  *
21  *  Optimization module for tcpdump intermediate representation.
22  */
23 #ifndef lint
24 static const char rcsid[] _U_ =
25     "@(#) $Header: /tcpdump/master/libpcap/optimize.c,v 1.90.2.1 2008/01/02 04:22:16 guy Exp $ (LBL)";
26 #endif
27 
28 #ifdef HAVE_CONFIG_H
29 #include "config.h"
30 #endif
31 
32 #include <stdio.h>
33 #include <stdlib.h>
34 #include <memory.h>
35 #include <string.h>
36 
37 #include <errno.h>
38 
39 #include "pcap-int.h"
40 
41 #include "gencode.h"
42 
43 #ifdef HAVE_OS_PROTO_H
44 #include "os-proto.h"
45 #endif
46 
47 #ifdef BDEBUG
48 extern int dflag;
49 #endif
50 
51 #if defined(MSDOS) && !defined(__DJGPP__)
52 extern int _w32_ffs (int mask);
53 #define ffs _w32_ffs
54 #endif
55 
56 #if defined(WIN32) && defined (_MSC_VER)
57 int ffs(int mask);
58 #endif
59 
60 /*
61  * Represents a deleted instruction.
62  */
63 #define NOP -1
64 
65 /*
66  * Register numbers for use-def values.
67  * 0 through BPF_MEMWORDS-1 represent the corresponding scratch memory
68  * location.  A_ATOM is the accumulator and X_ATOM is the index
69  * register.
70  */
71 #define A_ATOM BPF_MEMWORDS
72 #define X_ATOM (BPF_MEMWORDS+1)
73 
74 /*
75  * This define is used to represent *both* the accumulator and
76  * x register in use-def computations.
77  * Currently, the use-def code assumes only one definition per instruction.
78  */
79 #define AX_ATOM N_ATOMS
80 
81 /*
82  * A flag to indicate that further optimization is needed.
83  * Iterative passes are continued until a given pass yields no
84  * branch movement.
85  */
86 static int done;
87 
88 /*
89  * A block is marked if only if its mark equals the current mark.
90  * Rather than traverse the code array, marking each item, 'cur_mark' is
91  * incremented.  This automatically makes each element unmarked.
92  */
93 static int cur_mark;
94 #define isMarked(p) ((p)->mark == cur_mark)
95 #define unMarkAll() cur_mark += 1
96 #define Mark(p) ((p)->mark = cur_mark)
97 
98 static void opt_init(struct block *);
99 static void opt_cleanup(void);
100 
101 static void make_marks(struct block *);
102 static void mark_code(struct block *);
103 
104 static void intern_blocks(struct block *);
105 
106 static int eq_slist(struct slist *, struct slist *);
107 
108 static void find_levels_r(struct block *);
109 
110 static void find_levels(struct block *);
111 static void find_dom(struct block *);
112 static void propedom(struct edge *);
113 static void find_edom(struct block *);
114 static void find_closure(struct block *);
115 static int atomuse(struct stmt *);
116 static int atomdef(struct stmt *);
117 static void compute_local_ud(struct block *);
118 static void find_ud(struct block *);
119 static void init_val(void);
120 static int F(int, int, int);
121 static inline void vstore(struct stmt *, int *, int, int);
122 static void opt_blk(struct block *, int);
123 static int use_conflict(struct block *, struct block *);
124 static void opt_j(struct edge *);
125 static void or_pullup(struct block *);
126 static void and_pullup(struct block *);
127 static void opt_blks(struct block *, int);
128 static inline void link_inedge(struct edge *, struct block *);
129 static void find_inedges(struct block *);
130 static void opt_root(struct block **);
131 static void opt_loop(struct block *, int);
132 static void fold_op(struct stmt *, int, int);
133 static inline struct slist *this_op(struct slist *);
134 static void opt_not(struct block *);
135 static void opt_peep(struct block *);
136 static void opt_stmt(struct stmt *, int[], int);
137 static void deadstmt(struct stmt *, struct stmt *[]);
138 static void opt_deadstores(struct block *);
139 static struct block *fold_edge(struct block *, struct edge *);
140 static inline int eq_blk(struct block *, struct block *);
141 static int slength(struct slist *);
142 static int count_blocks(struct block *);
143 static void number_blks_r(struct block *);
144 static int count_stmts(struct block *);
145 static int convert_code_r(struct block *);
146 #ifdef BDEBUG
147 static void opt_dump(struct block *);
148 #endif
149 
150 static int n_blocks;
151 struct block **blocks;
152 static int n_edges;
153 struct edge **edges;
154 
155 /*
156  * A bit vector set representation of the dominators.
157  * We round up the set size to the next power of two.
158  */
159 static int nodewords;
160 static int edgewords;
161 struct block **levels;
162 bpf_u_int32 *space;
163 #define BITS_PER_WORD (8*sizeof(bpf_u_int32))
164 /*
165  * True if a is in uset {p}
166  */
167 #define SET_MEMBER(p, a) \
168 ((p)[(unsigned)(a) / BITS_PER_WORD] & (1 << ((unsigned)(a) % BITS_PER_WORD)))
169 
170 /*
171  * Add 'a' to uset p.
172  */
173 #define SET_INSERT(p, a) \
174 (p)[(unsigned)(a) / BITS_PER_WORD] |= (1 << ((unsigned)(a) % BITS_PER_WORD))
175 
176 /*
177  * Delete 'a' from uset p.
178  */
179 #define SET_DELETE(p, a) \
180 (p)[(unsigned)(a) / BITS_PER_WORD] &= ~(1 << ((unsigned)(a) % BITS_PER_WORD))
181 
182 /*
183  * a := a intersect b
184  */
185 #define SET_INTERSECT(a, b, n)\
186 {\
187 	register bpf_u_int32 *_x = a, *_y = b;\
188 	register int _n = n;\
189 	while (--_n >= 0) *_x++ &= *_y++;\
190 }
191 
192 /*
193  * a := a - b
194  */
195 #define SET_SUBTRACT(a, b, n)\
196 {\
197 	register bpf_u_int32 *_x = a, *_y = b;\
198 	register int _n = n;\
199 	while (--_n >= 0) *_x++ &=~ *_y++;\
200 }
201 
202 /*
203  * a := a union b
204  */
205 #define SET_UNION(a, b, n)\
206 {\
207 	register bpf_u_int32 *_x = a, *_y = b;\
208 	register int _n = n;\
209 	while (--_n >= 0) *_x++ |= *_y++;\
210 }
211 
212 static uset all_dom_sets;
213 static uset all_closure_sets;
214 static uset all_edge_sets;
215 
216 #ifndef MAX
217 #define MAX(a,b) ((a)>(b)?(a):(b))
218 #endif
219 
220 static void
221 find_levels_r(b)
222 	struct block *b;
223 {
224 	int level;
225 
226 	if (isMarked(b))
227 		return;
228 
229 	Mark(b);
230 	b->link = 0;
231 
232 	if (JT(b)) {
233 		find_levels_r(JT(b));
234 		find_levels_r(JF(b));
235 		level = MAX(JT(b)->level, JF(b)->level) + 1;
236 	} else
237 		level = 0;
238 	b->level = level;
239 	b->link = levels[level];
240 	levels[level] = b;
241 }
242 
243 /*
244  * Level graph.  The levels go from 0 at the leaves to
245  * N_LEVELS at the root.  The levels[] array points to the
246  * first node of the level list, whose elements are linked
247  * with the 'link' field of the struct block.
248  */
249 static void
250 find_levels(root)
251 	struct block *root;
252 {
253 	memset((char *)levels, 0, n_blocks * sizeof(*levels));
254 	unMarkAll();
255 	find_levels_r(root);
256 }
257 
258 /*
259  * Find dominator relationships.
260  * Assumes graph has been leveled.
261  */
262 static void
263 find_dom(root)
264 	struct block *root;
265 {
266 	int i;
267 	struct block *b;
268 	bpf_u_int32 *x;
269 
270 	/*
271 	 * Initialize sets to contain all nodes.
272 	 */
273 	x = all_dom_sets;
274 	i = n_blocks * nodewords;
275 	while (--i >= 0)
276 		*x++ = ~0;
277 	/* Root starts off empty. */
278 	for (i = nodewords; --i >= 0;)
279 		root->dom[i] = 0;
280 
281 	/* root->level is the highest level no found. */
282 	for (i = root->level; i >= 0; --i) {
283 		for (b = levels[i]; b; b = b->link) {
284 			SET_INSERT(b->dom, b->id);
285 			if (JT(b) == 0)
286 				continue;
287 			SET_INTERSECT(JT(b)->dom, b->dom, nodewords);
288 			SET_INTERSECT(JF(b)->dom, b->dom, nodewords);
289 		}
290 	}
291 }
292 
293 static void
294 propedom(ep)
295 	struct edge *ep;
296 {
297 	SET_INSERT(ep->edom, ep->id);
298 	if (ep->succ) {
299 		SET_INTERSECT(ep->succ->et.edom, ep->edom, edgewords);
300 		SET_INTERSECT(ep->succ->ef.edom, ep->edom, edgewords);
301 	}
302 }
303 
304 /*
305  * Compute edge dominators.
306  * Assumes graph has been leveled and predecessors established.
307  */
308 static void
309 find_edom(root)
310 	struct block *root;
311 {
312 	int i;
313 	uset x;
314 	struct block *b;
315 
316 	x = all_edge_sets;
317 	for (i = n_edges * edgewords; --i >= 0; )
318 		x[i] = ~0;
319 
320 	/* root->level is the highest level no found. */
321 	memset(root->et.edom, 0, edgewords * sizeof(*(uset)0));
322 	memset(root->ef.edom, 0, edgewords * sizeof(*(uset)0));
323 	for (i = root->level; i >= 0; --i) {
324 		for (b = levels[i]; b != 0; b = b->link) {
325 			propedom(&b->et);
326 			propedom(&b->ef);
327 		}
328 	}
329 }
330 
331 /*
332  * Find the backwards transitive closure of the flow graph.  These sets
333  * are backwards in the sense that we find the set of nodes that reach
334  * a given node, not the set of nodes that can be reached by a node.
335  *
336  * Assumes graph has been leveled.
337  */
338 static void
339 find_closure(root)
340 	struct block *root;
341 {
342 	int i;
343 	struct block *b;
344 
345 	/*
346 	 * Initialize sets to contain no nodes.
347 	 */
348 	memset((char *)all_closure_sets, 0,
349 	      n_blocks * nodewords * sizeof(*all_closure_sets));
350 
351 	/* root->level is the highest level no found. */
352 	for (i = root->level; i >= 0; --i) {
353 		for (b = levels[i]; b; b = b->link) {
354 			SET_INSERT(b->closure, b->id);
355 			if (JT(b) == 0)
356 				continue;
357 			SET_UNION(JT(b)->closure, b->closure, nodewords);
358 			SET_UNION(JF(b)->closure, b->closure, nodewords);
359 		}
360 	}
361 }
362 
363 /*
364  * Return the register number that is used by s.  If A and X are both
365  * used, return AX_ATOM.  If no register is used, return -1.
366  *
367  * The implementation should probably change to an array access.
368  */
369 static int
370 atomuse(s)
371 	struct stmt *s;
372 {
373 	register int c = s->code;
374 
375 	if (c == NOP)
376 		return -1;
377 
378 	switch (BPF_CLASS(c)) {
379 
380 	case BPF_RET:
381 		return (BPF_RVAL(c) == BPF_A) ? A_ATOM :
382 			(BPF_RVAL(c) == BPF_X) ? X_ATOM : -1;
383 
384 	case BPF_LD:
385 	case BPF_LDX:
386 		return (BPF_MODE(c) == BPF_IND) ? X_ATOM :
387 			(BPF_MODE(c) == BPF_MEM) ? s->k : -1;
388 
389 	case BPF_ST:
390 		return A_ATOM;
391 
392 	case BPF_STX:
393 		return X_ATOM;
394 
395 	case BPF_JMP:
396 	case BPF_ALU:
397 		if (BPF_SRC(c) == BPF_X)
398 			return AX_ATOM;
399 		return A_ATOM;
400 
401 	case BPF_MISC:
402 		return BPF_MISCOP(c) == BPF_TXA ? X_ATOM : A_ATOM;
403 	}
404 	abort();
405 	/* NOTREACHED */
406 }
407 
408 /*
409  * Return the register number that is defined by 's'.  We assume that
410  * a single stmt cannot define more than one register.  If no register
411  * is defined, return -1.
412  *
413  * The implementation should probably change to an array access.
414  */
415 static int
416 atomdef(s)
417 	struct stmt *s;
418 {
419 	if (s->code == NOP)
420 		return -1;
421 
422 	switch (BPF_CLASS(s->code)) {
423 
424 	case BPF_LD:
425 	case BPF_ALU:
426 		return A_ATOM;
427 
428 	case BPF_LDX:
429 		return X_ATOM;
430 
431 	case BPF_ST:
432 	case BPF_STX:
433 		return s->k;
434 
435 	case BPF_MISC:
436 		return BPF_MISCOP(s->code) == BPF_TAX ? X_ATOM : A_ATOM;
437 	}
438 	return -1;
439 }
440 
441 /*
442  * Compute the sets of registers used, defined, and killed by 'b'.
443  *
444  * "Used" means that a statement in 'b' uses the register before any
445  * statement in 'b' defines it, i.e. it uses the value left in
446  * that register by a predecessor block of this block.
447  * "Defined" means that a statement in 'b' defines it.
448  * "Killed" means that a statement in 'b' defines it before any
449  * statement in 'b' uses it, i.e. it kills the value left in that
450  * register by a predecessor block of this block.
451  */
452 static void
453 compute_local_ud(b)
454 	struct block *b;
455 {
456 	struct slist *s;
457 	atomset def = 0, use = 0, kill = 0;
458 	int atom;
459 
460 	for (s = b->stmts; s; s = s->next) {
461 		if (s->s.code == NOP)
462 			continue;
463 		atom = atomuse(&s->s);
464 		if (atom >= 0) {
465 			if (atom == AX_ATOM) {
466 				if (!ATOMELEM(def, X_ATOM))
467 					use |= ATOMMASK(X_ATOM);
468 				if (!ATOMELEM(def, A_ATOM))
469 					use |= ATOMMASK(A_ATOM);
470 			}
471 			else if (atom < N_ATOMS) {
472 				if (!ATOMELEM(def, atom))
473 					use |= ATOMMASK(atom);
474 			}
475 			else
476 				abort();
477 		}
478 		atom = atomdef(&s->s);
479 		if (atom >= 0) {
480 			if (!ATOMELEM(use, atom))
481 				kill |= ATOMMASK(atom);
482 			def |= ATOMMASK(atom);
483 		}
484 	}
485 	if (BPF_CLASS(b->s.code) == BPF_JMP) {
486 		/*
487 		 * XXX - what about RET?
488 		 */
489 		atom = atomuse(&b->s);
490 		if (atom >= 0) {
491 			if (atom == AX_ATOM) {
492 				if (!ATOMELEM(def, X_ATOM))
493 					use |= ATOMMASK(X_ATOM);
494 				if (!ATOMELEM(def, A_ATOM))
495 					use |= ATOMMASK(A_ATOM);
496 			}
497 			else if (atom < N_ATOMS) {
498 				if (!ATOMELEM(def, atom))
499 					use |= ATOMMASK(atom);
500 			}
501 			else
502 				abort();
503 		}
504 	}
505 
506 	b->def = def;
507 	b->kill = kill;
508 	b->in_use = use;
509 }
510 
511 /*
512  * Assume graph is already leveled.
513  */
514 static void
515 find_ud(root)
516 	struct block *root;
517 {
518 	int i, maxlevel;
519 	struct block *p;
520 
521 	/*
522 	 * root->level is the highest level no found;
523 	 * count down from there.
524 	 */
525 	maxlevel = root->level;
526 	for (i = maxlevel; i >= 0; --i)
527 		for (p = levels[i]; p; p = p->link) {
528 			compute_local_ud(p);
529 			p->out_use = 0;
530 		}
531 
532 	for (i = 1; i <= maxlevel; ++i) {
533 		for (p = levels[i]; p; p = p->link) {
534 			p->out_use |= JT(p)->in_use | JF(p)->in_use;
535 			p->in_use |= p->out_use &~ p->kill;
536 		}
537 	}
538 }
539 
540 /*
541  * These data structures are used in a Cocke and Shwarz style
542  * value numbering scheme.  Since the flowgraph is acyclic,
543  * exit values can be propagated from a node's predecessors
544  * provided it is uniquely defined.
545  */
546 struct valnode {
547 	int code;
548 	int v0, v1;
549 	int val;
550 	struct valnode *next;
551 };
552 
553 #define MODULUS 213
554 static struct valnode *hashtbl[MODULUS];
555 static int curval;
556 static int maxval;
557 
558 /* Integer constants mapped with the load immediate opcode. */
559 #define K(i) F(BPF_LD|BPF_IMM|BPF_W, i, 0L)
560 
561 struct vmapinfo {
562 	int is_const;
563 	bpf_int32 const_val;
564 };
565 
566 struct vmapinfo *vmap;
567 struct valnode *vnode_base;
568 struct valnode *next_vnode;
569 
570 static void
571 init_val()
572 {
573 	curval = 0;
574 	next_vnode = vnode_base;
575 	memset((char *)vmap, 0, maxval * sizeof(*vmap));
576 	memset((char *)hashtbl, 0, sizeof hashtbl);
577 }
578 
579 /* Because we really don't have an IR, this stuff is a little messy. */
580 static int
581 F(code, v0, v1)
582 	int code;
583 	int v0, v1;
584 {
585 	u_int hash;
586 	int val;
587 	struct valnode *p;
588 
589 	hash = (u_int)code ^ (v0 << 4) ^ (v1 << 8);
590 	hash %= MODULUS;
591 
592 	for (p = hashtbl[hash]; p; p = p->next)
593 		if (p->code == code && p->v0 == v0 && p->v1 == v1)
594 			return p->val;
595 
596 	val = ++curval;
597 	if (BPF_MODE(code) == BPF_IMM &&
598 	    (BPF_CLASS(code) == BPF_LD || BPF_CLASS(code) == BPF_LDX)) {
599 		vmap[val].const_val = v0;
600 		vmap[val].is_const = 1;
601 	}
602 	p = next_vnode++;
603 	p->val = val;
604 	p->code = code;
605 	p->v0 = v0;
606 	p->v1 = v1;
607 	p->next = hashtbl[hash];
608 	hashtbl[hash] = p;
609 
610 	return val;
611 }
612 
613 static inline void
614 vstore(s, valp, newval, alter)
615 	struct stmt *s;
616 	int *valp;
617 	int newval;
618 	int alter;
619 {
620 	if (alter && *valp == newval)
621 		s->code = NOP;
622 	else
623 		*valp = newval;
624 }
625 
626 static void
627 fold_op(s, v0, v1)
628 	struct stmt *s;
629 	int v0, v1;
630 {
631 	bpf_u_int32 a, b;
632 
633 	a = vmap[v0].const_val;
634 	b = vmap[v1].const_val;
635 
636 	switch (BPF_OP(s->code)) {
637 	case BPF_ADD:
638 		a += b;
639 		break;
640 
641 	case BPF_SUB:
642 		a -= b;
643 		break;
644 
645 	case BPF_MUL:
646 		a *= b;
647 		break;
648 
649 	case BPF_DIV:
650 		if (b == 0)
651 			bpf_error("division by zero");
652 		a /= b;
653 		break;
654 
655 	case BPF_AND:
656 		a &= b;
657 		break;
658 
659 	case BPF_OR:
660 		a |= b;
661 		break;
662 
663 	case BPF_LSH:
664 		a <<= b;
665 		break;
666 
667 	case BPF_RSH:
668 		a >>= b;
669 		break;
670 
671 	case BPF_NEG:
672 		a = -a;
673 		break;
674 
675 	default:
676 		abort();
677 	}
678 	s->k = a;
679 	s->code = BPF_LD|BPF_IMM;
680 	done = 0;
681 }
682 
683 static inline struct slist *
684 this_op(s)
685 	struct slist *s;
686 {
687 	while (s != 0 && s->s.code == NOP)
688 		s = s->next;
689 	return s;
690 }
691 
692 static void
693 opt_not(b)
694 	struct block *b;
695 {
696 	struct block *tmp = JT(b);
697 
698 	JT(b) = JF(b);
699 	JF(b) = tmp;
700 }
701 
702 static void
703 opt_peep(b)
704 	struct block *b;
705 {
706 	struct slist *s;
707 	struct slist *next, *last;
708 	int val;
709 
710 	s = b->stmts;
711 	if (s == 0)
712 		return;
713 
714 	last = s;
715 	for (/*empty*/; /*empty*/; s = next) {
716 		/*
717 		 * Skip over nops.
718 		 */
719 		s = this_op(s);
720 		if (s == 0)
721 			break;	/* nothing left in the block */
722 
723 		/*
724 		 * Find the next real instruction after that one
725 		 * (skipping nops).
726 		 */
727 		next = this_op(s->next);
728 		if (next == 0)
729 			break;	/* no next instruction */
730 		last = next;
731 
732 		/*
733 		 * st  M[k]	-->	st  M[k]
734 		 * ldx M[k]		tax
735 		 */
736 		if (s->s.code == BPF_ST &&
737 		    next->s.code == (BPF_LDX|BPF_MEM) &&
738 		    s->s.k == next->s.k) {
739 			done = 0;
740 			next->s.code = BPF_MISC|BPF_TAX;
741 		}
742 		/*
743 		 * ld  #k	-->	ldx  #k
744 		 * tax			txa
745 		 */
746 		if (s->s.code == (BPF_LD|BPF_IMM) &&
747 		    next->s.code == (BPF_MISC|BPF_TAX)) {
748 			s->s.code = BPF_LDX|BPF_IMM;
749 			next->s.code = BPF_MISC|BPF_TXA;
750 			done = 0;
751 		}
752 		/*
753 		 * This is an ugly special case, but it happens
754 		 * when you say tcp[k] or udp[k] where k is a constant.
755 		 */
756 		if (s->s.code == (BPF_LD|BPF_IMM)) {
757 			struct slist *add, *tax, *ild;
758 
759 			/*
760 			 * Check that X isn't used on exit from this
761 			 * block (which the optimizer might cause).
762 			 * We know the code generator won't generate
763 			 * any local dependencies.
764 			 */
765 			if (ATOMELEM(b->out_use, X_ATOM))
766 				continue;
767 
768 			/*
769 			 * Check that the instruction following the ldi
770 			 * is an addx, or it's an ldxms with an addx
771 			 * following it (with 0 or more nops between the
772 			 * ldxms and addx).
773 			 */
774 			if (next->s.code != (BPF_LDX|BPF_MSH|BPF_B))
775 				add = next;
776 			else
777 				add = this_op(next->next);
778 			if (add == 0 || add->s.code != (BPF_ALU|BPF_ADD|BPF_X))
779 				continue;
780 
781 			/*
782 			 * Check that a tax follows that (with 0 or more
783 			 * nops between them).
784 			 */
785 			tax = this_op(add->next);
786 			if (tax == 0 || tax->s.code != (BPF_MISC|BPF_TAX))
787 				continue;
788 
789 			/*
790 			 * Check that an ild follows that (with 0 or more
791 			 * nops between them).
792 			 */
793 			ild = this_op(tax->next);
794 			if (ild == 0 || BPF_CLASS(ild->s.code) != BPF_LD ||
795 			    BPF_MODE(ild->s.code) != BPF_IND)
796 				continue;
797 			/*
798 			 * We want to turn this sequence:
799 			 *
800 			 * (004) ldi     #0x2		{s}
801 			 * (005) ldxms   [14]		{next}  -- optional
802 			 * (006) addx			{add}
803 			 * (007) tax			{tax}
804 			 * (008) ild     [x+0]		{ild}
805 			 *
806 			 * into this sequence:
807 			 *
808 			 * (004) nop
809 			 * (005) ldxms   [14]
810 			 * (006) nop
811 			 * (007) nop
812 			 * (008) ild     [x+2]
813 			 *
814 			 * XXX We need to check that X is not
815 			 * subsequently used, because we want to change
816 			 * what'll be in it after this sequence.
817 			 *
818 			 * We know we can eliminate the accumulator
819 			 * modifications earlier in the sequence since
820 			 * it is defined by the last stmt of this sequence
821 			 * (i.e., the last statement of the sequence loads
822 			 * a value into the accumulator, so we can eliminate
823 			 * earlier operations on the accumulator).
824 			 */
825 			ild->s.k += s->s.k;
826 			s->s.code = NOP;
827 			add->s.code = NOP;
828 			tax->s.code = NOP;
829 			done = 0;
830 		}
831 	}
832 	/*
833 	 * If the comparison at the end of a block is an equality
834 	 * comparison against a constant, and nobody uses the value
835 	 * we leave in the A register at the end of a block, and
836 	 * the operation preceding the comparison is an arithmetic
837 	 * operation, we can sometime optimize it away.
838 	 */
839 	if (b->s.code == (BPF_JMP|BPF_JEQ|BPF_K) &&
840 	    !ATOMELEM(b->out_use, A_ATOM)) {
841 	    	/*
842 	    	 * We can optimize away certain subtractions of the
843 	    	 * X register.
844 	    	 */
845 		if (last->s.code == (BPF_ALU|BPF_SUB|BPF_X)) {
846 			val = b->val[X_ATOM];
847 			if (vmap[val].is_const) {
848 				/*
849 				 * If we have a subtract to do a comparison,
850 				 * and the X register is a known constant,
851 				 * we can merge this value into the
852 				 * comparison:
853 				 *
854 				 * sub x  ->	nop
855 				 * jeq #y	jeq #(x+y)
856 				 */
857 				b->s.k += vmap[val].const_val;
858 				last->s.code = NOP;
859 				done = 0;
860 			} else if (b->s.k == 0) {
861 				/*
862 				 * If the X register isn't a constant,
863 				 * and the comparison in the test is
864 				 * against 0, we can compare with the
865 				 * X register, instead:
866 				 *
867 				 * sub x  ->	nop
868 				 * jeq #0	jeq x
869 				 */
870 				last->s.code = NOP;
871 				b->s.code = BPF_JMP|BPF_JEQ|BPF_X;
872 				done = 0;
873 			}
874 		}
875 		/*
876 		 * Likewise, a constant subtract can be simplified:
877 		 *
878 		 * sub #x ->	nop
879 		 * jeq #y ->	jeq #(x+y)
880 		 */
881 		else if (last->s.code == (BPF_ALU|BPF_SUB|BPF_K)) {
882 			last->s.code = NOP;
883 			b->s.k += last->s.k;
884 			done = 0;
885 		}
886 		/*
887 		 * And, similarly, a constant AND can be simplified
888 		 * if we're testing against 0, i.e.:
889 		 *
890 		 * and #k	nop
891 		 * jeq #0  ->	jset #k
892 		 */
893 		else if (last->s.code == (BPF_ALU|BPF_AND|BPF_K) &&
894 		    b->s.k == 0) {
895 			b->s.k = last->s.k;
896 			b->s.code = BPF_JMP|BPF_K|BPF_JSET;
897 			last->s.code = NOP;
898 			done = 0;
899 			opt_not(b);
900 		}
901 	}
902 	/*
903 	 * jset #0        ->   never
904 	 * jset #ffffffff ->   always
905 	 */
906 	if (b->s.code == (BPF_JMP|BPF_K|BPF_JSET)) {
907 		if (b->s.k == 0)
908 			JT(b) = JF(b);
909 		if (b->s.k == 0xffffffff)
910 			JF(b) = JT(b);
911 	}
912 	/*
913 	 * If we're comparing against the index register, and the index
914 	 * register is a known constant, we can just compare against that
915 	 * constant.
916 	 */
917 	val = b->val[X_ATOM];
918 	if (vmap[val].is_const && BPF_SRC(b->s.code) == BPF_X) {
919 		bpf_int32 v = vmap[val].const_val;
920 		b->s.code &= ~BPF_X;
921 		b->s.k = v;
922 	}
923 	/*
924 	 * If the accumulator is a known constant, we can compute the
925 	 * comparison result.
926 	 */
927 	val = b->val[A_ATOM];
928 	if (vmap[val].is_const && BPF_SRC(b->s.code) == BPF_K) {
929 		bpf_int32 v = vmap[val].const_val;
930 		switch (BPF_OP(b->s.code)) {
931 
932 		case BPF_JEQ:
933 			v = v == b->s.k;
934 			break;
935 
936 		case BPF_JGT:
937 			v = (unsigned)v > b->s.k;
938 			break;
939 
940 		case BPF_JGE:
941 			v = (unsigned)v >= b->s.k;
942 			break;
943 
944 		case BPF_JSET:
945 			v &= b->s.k;
946 			break;
947 
948 		default:
949 			abort();
950 		}
951 		if (JF(b) != JT(b))
952 			done = 0;
953 		if (v)
954 			JF(b) = JT(b);
955 		else
956 			JT(b) = JF(b);
957 	}
958 }
959 
960 /*
961  * Compute the symbolic value of expression of 's', and update
962  * anything it defines in the value table 'val'.  If 'alter' is true,
963  * do various optimizations.  This code would be cleaner if symbolic
964  * evaluation and code transformations weren't folded together.
965  */
966 static void
967 opt_stmt(s, val, alter)
968 	struct stmt *s;
969 	int val[];
970 	int alter;
971 {
972 	int op;
973 	int v;
974 
975 	switch (s->code) {
976 
977 	case BPF_LD|BPF_ABS|BPF_W:
978 	case BPF_LD|BPF_ABS|BPF_H:
979 	case BPF_LD|BPF_ABS|BPF_B:
980 		v = F(s->code, s->k, 0L);
981 		vstore(s, &val[A_ATOM], v, alter);
982 		break;
983 
984 	case BPF_LD|BPF_IND|BPF_W:
985 	case BPF_LD|BPF_IND|BPF_H:
986 	case BPF_LD|BPF_IND|BPF_B:
987 		v = val[X_ATOM];
988 		if (alter && vmap[v].is_const) {
989 			s->code = BPF_LD|BPF_ABS|BPF_SIZE(s->code);
990 			s->k += vmap[v].const_val;
991 			v = F(s->code, s->k, 0L);
992 			done = 0;
993 		}
994 		else
995 			v = F(s->code, s->k, v);
996 		vstore(s, &val[A_ATOM], v, alter);
997 		break;
998 
999 	case BPF_LD|BPF_LEN:
1000 		v = F(s->code, 0L, 0L);
1001 		vstore(s, &val[A_ATOM], v, alter);
1002 		break;
1003 
1004 	case BPF_LD|BPF_IMM:
1005 		v = K(s->k);
1006 		vstore(s, &val[A_ATOM], v, alter);
1007 		break;
1008 
1009 	case BPF_LDX|BPF_IMM:
1010 		v = K(s->k);
1011 		vstore(s, &val[X_ATOM], v, alter);
1012 		break;
1013 
1014 	case BPF_LDX|BPF_MSH|BPF_B:
1015 		v = F(s->code, s->k, 0L);
1016 		vstore(s, &val[X_ATOM], v, alter);
1017 		break;
1018 
1019 	case BPF_ALU|BPF_NEG:
1020 		if (alter && vmap[val[A_ATOM]].is_const) {
1021 			s->code = BPF_LD|BPF_IMM;
1022 			s->k = -vmap[val[A_ATOM]].const_val;
1023 			val[A_ATOM] = K(s->k);
1024 		}
1025 		else
1026 			val[A_ATOM] = F(s->code, val[A_ATOM], 0L);
1027 		break;
1028 
1029 	case BPF_ALU|BPF_ADD|BPF_K:
1030 	case BPF_ALU|BPF_SUB|BPF_K:
1031 	case BPF_ALU|BPF_MUL|BPF_K:
1032 	case BPF_ALU|BPF_DIV|BPF_K:
1033 	case BPF_ALU|BPF_AND|BPF_K:
1034 	case BPF_ALU|BPF_OR|BPF_K:
1035 	case BPF_ALU|BPF_LSH|BPF_K:
1036 	case BPF_ALU|BPF_RSH|BPF_K:
1037 		op = BPF_OP(s->code);
1038 		if (alter) {
1039 			if (s->k == 0) {
1040 				/* don't optimize away "sub #0"
1041 				 * as it may be needed later to
1042 				 * fixup the generated math code */
1043 				if (op == BPF_ADD ||
1044 				    op == BPF_LSH || op == BPF_RSH ||
1045 				    op == BPF_OR) {
1046 					s->code = NOP;
1047 					break;
1048 				}
1049 				if (op == BPF_MUL || op == BPF_AND) {
1050 					s->code = BPF_LD|BPF_IMM;
1051 					val[A_ATOM] = K(s->k);
1052 					break;
1053 				}
1054 			}
1055 			if (vmap[val[A_ATOM]].is_const) {
1056 				fold_op(s, val[A_ATOM], K(s->k));
1057 				val[A_ATOM] = K(s->k);
1058 				break;
1059 			}
1060 		}
1061 		val[A_ATOM] = F(s->code, val[A_ATOM], K(s->k));
1062 		break;
1063 
1064 	case BPF_ALU|BPF_ADD|BPF_X:
1065 	case BPF_ALU|BPF_SUB|BPF_X:
1066 	case BPF_ALU|BPF_MUL|BPF_X:
1067 	case BPF_ALU|BPF_DIV|BPF_X:
1068 	case BPF_ALU|BPF_AND|BPF_X:
1069 	case BPF_ALU|BPF_OR|BPF_X:
1070 	case BPF_ALU|BPF_LSH|BPF_X:
1071 	case BPF_ALU|BPF_RSH|BPF_X:
1072 		op = BPF_OP(s->code);
1073 		if (alter && vmap[val[X_ATOM]].is_const) {
1074 			if (vmap[val[A_ATOM]].is_const) {
1075 				fold_op(s, val[A_ATOM], val[X_ATOM]);
1076 				val[A_ATOM] = K(s->k);
1077 			}
1078 			else {
1079 				s->code = BPF_ALU|BPF_K|op;
1080 				s->k = vmap[val[X_ATOM]].const_val;
1081 				done = 0;
1082 				val[A_ATOM] =
1083 					F(s->code, val[A_ATOM], K(s->k));
1084 			}
1085 			break;
1086 		}
1087 		/*
1088 		 * Check if we're doing something to an accumulator
1089 		 * that is 0, and simplify.  This may not seem like
1090 		 * much of a simplification but it could open up further
1091 		 * optimizations.
1092 		 * XXX We could also check for mul by 1, etc.
1093 		 */
1094 		if (alter && vmap[val[A_ATOM]].is_const
1095 		    && vmap[val[A_ATOM]].const_val == 0) {
1096 			if (op == BPF_ADD || op == BPF_OR) {
1097 				s->code = BPF_MISC|BPF_TXA;
1098 				vstore(s, &val[A_ATOM], val[X_ATOM], alter);
1099 				break;
1100 			}
1101 			else if (op == BPF_MUL || op == BPF_DIV ||
1102 				 op == BPF_AND || op == BPF_LSH || op == BPF_RSH) {
1103 				s->code = BPF_LD|BPF_IMM;
1104 				s->k = 0;
1105 				vstore(s, &val[A_ATOM], K(s->k), alter);
1106 				break;
1107 			}
1108 			else if (op == BPF_NEG) {
1109 				s->code = NOP;
1110 				break;
1111 			}
1112 		}
1113 		val[A_ATOM] = F(s->code, val[A_ATOM], val[X_ATOM]);
1114 		break;
1115 
1116 	case BPF_MISC|BPF_TXA:
1117 		vstore(s, &val[A_ATOM], val[X_ATOM], alter);
1118 		break;
1119 
1120 	case BPF_LD|BPF_MEM:
1121 		v = val[s->k];
1122 		if (alter && vmap[v].is_const) {
1123 			s->code = BPF_LD|BPF_IMM;
1124 			s->k = vmap[v].const_val;
1125 			done = 0;
1126 		}
1127 		vstore(s, &val[A_ATOM], v, alter);
1128 		break;
1129 
1130 	case BPF_MISC|BPF_TAX:
1131 		vstore(s, &val[X_ATOM], val[A_ATOM], alter);
1132 		break;
1133 
1134 	case BPF_LDX|BPF_MEM:
1135 		v = val[s->k];
1136 		if (alter && vmap[v].is_const) {
1137 			s->code = BPF_LDX|BPF_IMM;
1138 			s->k = vmap[v].const_val;
1139 			done = 0;
1140 		}
1141 		vstore(s, &val[X_ATOM], v, alter);
1142 		break;
1143 
1144 	case BPF_ST:
1145 		vstore(s, &val[s->k], val[A_ATOM], alter);
1146 		break;
1147 
1148 	case BPF_STX:
1149 		vstore(s, &val[s->k], val[X_ATOM], alter);
1150 		break;
1151 	}
1152 }
1153 
1154 static void
1155 deadstmt(s, last)
1156 	register struct stmt *s;
1157 	register struct stmt *last[];
1158 {
1159 	register int atom;
1160 
1161 	atom = atomuse(s);
1162 	if (atom >= 0) {
1163 		if (atom == AX_ATOM) {
1164 			last[X_ATOM] = 0;
1165 			last[A_ATOM] = 0;
1166 		}
1167 		else
1168 			last[atom] = 0;
1169 	}
1170 	atom = atomdef(s);
1171 	if (atom >= 0) {
1172 		if (last[atom]) {
1173 			done = 0;
1174 			last[atom]->code = NOP;
1175 		}
1176 		last[atom] = s;
1177 	}
1178 }
1179 
1180 static void
1181 opt_deadstores(b)
1182 	register struct block *b;
1183 {
1184 	register struct slist *s;
1185 	register int atom;
1186 	struct stmt *last[N_ATOMS];
1187 
1188 	memset((char *)last, 0, sizeof last);
1189 
1190 	for (s = b->stmts; s != 0; s = s->next)
1191 		deadstmt(&s->s, last);
1192 	deadstmt(&b->s, last);
1193 
1194 	for (atom = 0; atom < N_ATOMS; ++atom)
1195 		if (last[atom] && !ATOMELEM(b->out_use, atom)) {
1196 			last[atom]->code = NOP;
1197 			done = 0;
1198 		}
1199 }
1200 
1201 static void
1202 opt_blk(b, do_stmts)
1203 	struct block *b;
1204 	int do_stmts;
1205 {
1206 	struct slist *s;
1207 	struct edge *p;
1208 	int i;
1209 	bpf_int32 aval, xval;
1210 
1211 #if 0
1212 	for (s = b->stmts; s && s->next; s = s->next)
1213 		if (BPF_CLASS(s->s.code) == BPF_JMP) {
1214 			do_stmts = 0;
1215 			break;
1216 		}
1217 #endif
1218 
1219 	/*
1220 	 * Initialize the atom values.
1221 	 */
1222 	p = b->in_edges;
1223 	if (p == 0) {
1224 		/*
1225 		 * We have no predecessors, so everything is undefined
1226 		 * upon entry to this block.
1227 		 */
1228 		memset((char *)b->val, 0, sizeof(b->val));
1229 	} else {
1230 		/*
1231 		 * Inherit values from our predecessors.
1232 		 *
1233 		 * First, get the values from the predecessor along the
1234 		 * first edge leading to this node.
1235 		 */
1236 		memcpy((char *)b->val, (char *)p->pred->val, sizeof(b->val));
1237 		/*
1238 		 * Now look at all the other nodes leading to this node.
1239 		 * If, for the predecessor along that edge, a register
1240 		 * has a different value from the one we have (i.e.,
1241 		 * control paths are merging, and the merging paths
1242 		 * assign different values to that register), give the
1243 		 * register the undefined value of 0.
1244 		 */
1245 		while ((p = p->next) != NULL) {
1246 			for (i = 0; i < N_ATOMS; ++i)
1247 				if (b->val[i] != p->pred->val[i])
1248 					b->val[i] = 0;
1249 		}
1250 	}
1251 	aval = b->val[A_ATOM];
1252 	xval = b->val[X_ATOM];
1253 	for (s = b->stmts; s; s = s->next)
1254 		opt_stmt(&s->s, b->val, do_stmts);
1255 
1256 	/*
1257 	 * This is a special case: if we don't use anything from this
1258 	 * block, and we load the accumulator or index register with a
1259 	 * value that is already there, or if this block is a return,
1260 	 * eliminate all the statements.
1261 	 *
1262 	 * XXX - what if it does a store?
1263 	 *
1264 	 * XXX - why does it matter whether we use anything from this
1265 	 * block?  If the accumulator or index register doesn't change
1266 	 * its value, isn't that OK even if we use that value?
1267 	 *
1268 	 * XXX - if we load the accumulator with a different value,
1269 	 * and the block ends with a conditional branch, we obviously
1270 	 * can't eliminate it, as the branch depends on that value.
1271 	 * For the index register, the conditional branch only depends
1272 	 * on the index register value if the test is against the index
1273 	 * register value rather than a constant; if nothing uses the
1274 	 * value we put into the index register, and we're not testing
1275 	 * against the index register's value, and there aren't any
1276 	 * other problems that would keep us from eliminating this
1277 	 * block, can we eliminate it?
1278 	 */
1279 	if (do_stmts &&
1280 	    ((b->out_use == 0 && aval != 0 && b->val[A_ATOM] == aval &&
1281 	      xval != 0 && b->val[X_ATOM] == xval) ||
1282 	     BPF_CLASS(b->s.code) == BPF_RET)) {
1283 		if (b->stmts != 0) {
1284 			b->stmts = 0;
1285 			done = 0;
1286 		}
1287 	} else {
1288 		opt_peep(b);
1289 		opt_deadstores(b);
1290 	}
1291 	/*
1292 	 * Set up values for branch optimizer.
1293 	 */
1294 	if (BPF_SRC(b->s.code) == BPF_K)
1295 		b->oval = K(b->s.k);
1296 	else
1297 		b->oval = b->val[X_ATOM];
1298 	b->et.code = b->s.code;
1299 	b->ef.code = -b->s.code;
1300 }
1301 
1302 /*
1303  * Return true if any register that is used on exit from 'succ', has
1304  * an exit value that is different from the corresponding exit value
1305  * from 'b'.
1306  */
1307 static int
1308 use_conflict(b, succ)
1309 	struct block *b, *succ;
1310 {
1311 	int atom;
1312 	atomset use = succ->out_use;
1313 
1314 	if (use == 0)
1315 		return 0;
1316 
1317 	for (atom = 0; atom < N_ATOMS; ++atom)
1318 		if (ATOMELEM(use, atom))
1319 			if (b->val[atom] != succ->val[atom])
1320 				return 1;
1321 	return 0;
1322 }
1323 
1324 static struct block *
1325 fold_edge(child, ep)
1326 	struct block *child;
1327 	struct edge *ep;
1328 {
1329 	int sense;
1330 	int aval0, aval1, oval0, oval1;
1331 	int code = ep->code;
1332 
1333 	if (code < 0) {
1334 		code = -code;
1335 		sense = 0;
1336 	} else
1337 		sense = 1;
1338 
1339 	if (child->s.code != code)
1340 		return 0;
1341 
1342 	aval0 = child->val[A_ATOM];
1343 	oval0 = child->oval;
1344 	aval1 = ep->pred->val[A_ATOM];
1345 	oval1 = ep->pred->oval;
1346 
1347 	if (aval0 != aval1)
1348 		return 0;
1349 
1350 	if (oval0 == oval1)
1351 		/*
1352 		 * The operands of the branch instructions are
1353 		 * identical, so the result is true if a true
1354 		 * branch was taken to get here, otherwise false.
1355 		 */
1356 		return sense ? JT(child) : JF(child);
1357 
1358 	if (sense && code == (BPF_JMP|BPF_JEQ|BPF_K))
1359 		/*
1360 		 * At this point, we only know the comparison if we
1361 		 * came down the true branch, and it was an equality
1362 		 * comparison with a constant.
1363 		 *
1364 		 * I.e., if we came down the true branch, and the branch
1365 		 * was an equality comparison with a constant, we know the
1366 		 * accumulator contains that constant.  If we came down
1367 		 * the false branch, or the comparison wasn't with a
1368 		 * constant, we don't know what was in the accumulator.
1369 		 *
1370 		 * We rely on the fact that distinct constants have distinct
1371 		 * value numbers.
1372 		 */
1373 		return JF(child);
1374 
1375 	return 0;
1376 }
1377 
1378 static void
1379 opt_j(ep)
1380 	struct edge *ep;
1381 {
1382 	register int i, k;
1383 	register struct block *target;
1384 
1385 	if (JT(ep->succ) == 0)
1386 		return;
1387 
1388 	if (JT(ep->succ) == JF(ep->succ)) {
1389 		/*
1390 		 * Common branch targets can be eliminated, provided
1391 		 * there is no data dependency.
1392 		 */
1393 		if (!use_conflict(ep->pred, ep->succ->et.succ)) {
1394 			done = 0;
1395 			ep->succ = JT(ep->succ);
1396 		}
1397 	}
1398 	/*
1399 	 * For each edge dominator that matches the successor of this
1400 	 * edge, promote the edge successor to the its grandchild.
1401 	 *
1402 	 * XXX We violate the set abstraction here in favor a reasonably
1403 	 * efficient loop.
1404 	 */
1405  top:
1406 	for (i = 0; i < edgewords; ++i) {
1407 		register bpf_u_int32 x = ep->edom[i];
1408 
1409 		while (x != 0) {
1410 			k = ffs(x) - 1;
1411 			x &=~ (1 << k);
1412 			k += i * BITS_PER_WORD;
1413 
1414 			target = fold_edge(ep->succ, edges[k]);
1415 			/*
1416 			 * Check that there is no data dependency between
1417 			 * nodes that will be violated if we move the edge.
1418 			 */
1419 			if (target != 0 && !use_conflict(ep->pred, target)) {
1420 				done = 0;
1421 				ep->succ = target;
1422 				if (JT(target) != 0)
1423 					/*
1424 					 * Start over unless we hit a leaf.
1425 					 */
1426 					goto top;
1427 				return;
1428 			}
1429 		}
1430 	}
1431 }
1432 
1433 
1434 static void
1435 or_pullup(b)
1436 	struct block *b;
1437 {
1438 	int val, at_top;
1439 	struct block *pull;
1440 	struct block **diffp, **samep;
1441 	struct edge *ep;
1442 
1443 	ep = b->in_edges;
1444 	if (ep == 0)
1445 		return;
1446 
1447 	/*
1448 	 * Make sure each predecessor loads the same value.
1449 	 * XXX why?
1450 	 */
1451 	val = ep->pred->val[A_ATOM];
1452 	for (ep = ep->next; ep != 0; ep = ep->next)
1453 		if (val != ep->pred->val[A_ATOM])
1454 			return;
1455 
1456 	if (JT(b->in_edges->pred) == b)
1457 		diffp = &JT(b->in_edges->pred);
1458 	else
1459 		diffp = &JF(b->in_edges->pred);
1460 
1461 	at_top = 1;
1462 	while (1) {
1463 		if (*diffp == 0)
1464 			return;
1465 
1466 		if (JT(*diffp) != JT(b))
1467 			return;
1468 
1469 		if (!SET_MEMBER((*diffp)->dom, b->id))
1470 			return;
1471 
1472 		if ((*diffp)->val[A_ATOM] != val)
1473 			break;
1474 
1475 		diffp = &JF(*diffp);
1476 		at_top = 0;
1477 	}
1478 	samep = &JF(*diffp);
1479 	while (1) {
1480 		if (*samep == 0)
1481 			return;
1482 
1483 		if (JT(*samep) != JT(b))
1484 			return;
1485 
1486 		if (!SET_MEMBER((*samep)->dom, b->id))
1487 			return;
1488 
1489 		if ((*samep)->val[A_ATOM] == val)
1490 			break;
1491 
1492 		/* XXX Need to check that there are no data dependencies
1493 		   between dp0 and dp1.  Currently, the code generator
1494 		   will not produce such dependencies. */
1495 		samep = &JF(*samep);
1496 	}
1497 #ifdef notdef
1498 	/* XXX This doesn't cover everything. */
1499 	for (i = 0; i < N_ATOMS; ++i)
1500 		if ((*samep)->val[i] != pred->val[i])
1501 			return;
1502 #endif
1503 	/* Pull up the node. */
1504 	pull = *samep;
1505 	*samep = JF(pull);
1506 	JF(pull) = *diffp;
1507 
1508 	/*
1509 	 * At the top of the chain, each predecessor needs to point at the
1510 	 * pulled up node.  Inside the chain, there is only one predecessor
1511 	 * to worry about.
1512 	 */
1513 	if (at_top) {
1514 		for (ep = b->in_edges; ep != 0; ep = ep->next) {
1515 			if (JT(ep->pred) == b)
1516 				JT(ep->pred) = pull;
1517 			else
1518 				JF(ep->pred) = pull;
1519 		}
1520 	}
1521 	else
1522 		*diffp = pull;
1523 
1524 	done = 0;
1525 }
1526 
1527 static void
1528 and_pullup(b)
1529 	struct block *b;
1530 {
1531 	int val, at_top;
1532 	struct block *pull;
1533 	struct block **diffp, **samep;
1534 	struct edge *ep;
1535 
1536 	ep = b->in_edges;
1537 	if (ep == 0)
1538 		return;
1539 
1540 	/*
1541 	 * Make sure each predecessor loads the same value.
1542 	 */
1543 	val = ep->pred->val[A_ATOM];
1544 	for (ep = ep->next; ep != 0; ep = ep->next)
1545 		if (val != ep->pred->val[A_ATOM])
1546 			return;
1547 
1548 	if (JT(b->in_edges->pred) == b)
1549 		diffp = &JT(b->in_edges->pred);
1550 	else
1551 		diffp = &JF(b->in_edges->pred);
1552 
1553 	at_top = 1;
1554 	while (1) {
1555 		if (*diffp == 0)
1556 			return;
1557 
1558 		if (JF(*diffp) != JF(b))
1559 			return;
1560 
1561 		if (!SET_MEMBER((*diffp)->dom, b->id))
1562 			return;
1563 
1564 		if ((*diffp)->val[A_ATOM] != val)
1565 			break;
1566 
1567 		diffp = &JT(*diffp);
1568 		at_top = 0;
1569 	}
1570 	samep = &JT(*diffp);
1571 	while (1) {
1572 		if (*samep == 0)
1573 			return;
1574 
1575 		if (JF(*samep) != JF(b))
1576 			return;
1577 
1578 		if (!SET_MEMBER((*samep)->dom, b->id))
1579 			return;
1580 
1581 		if ((*samep)->val[A_ATOM] == val)
1582 			break;
1583 
1584 		/* XXX Need to check that there are no data dependencies
1585 		   between diffp and samep.  Currently, the code generator
1586 		   will not produce such dependencies. */
1587 		samep = &JT(*samep);
1588 	}
1589 #ifdef notdef
1590 	/* XXX This doesn't cover everything. */
1591 	for (i = 0; i < N_ATOMS; ++i)
1592 		if ((*samep)->val[i] != pred->val[i])
1593 			return;
1594 #endif
1595 	/* Pull up the node. */
1596 	pull = *samep;
1597 	*samep = JT(pull);
1598 	JT(pull) = *diffp;
1599 
1600 	/*
1601 	 * At the top of the chain, each predecessor needs to point at the
1602 	 * pulled up node.  Inside the chain, there is only one predecessor
1603 	 * to worry about.
1604 	 */
1605 	if (at_top) {
1606 		for (ep = b->in_edges; ep != 0; ep = ep->next) {
1607 			if (JT(ep->pred) == b)
1608 				JT(ep->pred) = pull;
1609 			else
1610 				JF(ep->pred) = pull;
1611 		}
1612 	}
1613 	else
1614 		*diffp = pull;
1615 
1616 	done = 0;
1617 }
1618 
1619 static void
1620 opt_blks(root, do_stmts)
1621 	struct block *root;
1622 	int do_stmts;
1623 {
1624 	int i, maxlevel;
1625 	struct block *p;
1626 
1627 	init_val();
1628 	maxlevel = root->level;
1629 
1630 	find_inedges(root);
1631 	for (i = maxlevel; i >= 0; --i)
1632 		for (p = levels[i]; p; p = p->link)
1633 			opt_blk(p, do_stmts);
1634 
1635 	if (do_stmts)
1636 		/*
1637 		 * No point trying to move branches; it can't possibly
1638 		 * make a difference at this point.
1639 		 */
1640 		return;
1641 
1642 	for (i = 1; i <= maxlevel; ++i) {
1643 		for (p = levels[i]; p; p = p->link) {
1644 			opt_j(&p->et);
1645 			opt_j(&p->ef);
1646 		}
1647 	}
1648 
1649 	find_inedges(root);
1650 	for (i = 1; i <= maxlevel; ++i) {
1651 		for (p = levels[i]; p; p = p->link) {
1652 			or_pullup(p);
1653 			and_pullup(p);
1654 		}
1655 	}
1656 }
1657 
1658 static inline void
1659 link_inedge(parent, child)
1660 	struct edge *parent;
1661 	struct block *child;
1662 {
1663 	parent->next = child->in_edges;
1664 	child->in_edges = parent;
1665 }
1666 
1667 static void
1668 find_inedges(root)
1669 	struct block *root;
1670 {
1671 	int i;
1672 	struct block *b;
1673 
1674 	for (i = 0; i < n_blocks; ++i)
1675 		blocks[i]->in_edges = 0;
1676 
1677 	/*
1678 	 * Traverse the graph, adding each edge to the predecessor
1679 	 * list of its successors.  Skip the leaves (i.e. level 0).
1680 	 */
1681 	for (i = root->level; i > 0; --i) {
1682 		for (b = levels[i]; b != 0; b = b->link) {
1683 			link_inedge(&b->et, JT(b));
1684 			link_inedge(&b->ef, JF(b));
1685 		}
1686 	}
1687 }
1688 
1689 static void
1690 opt_root(b)
1691 	struct block **b;
1692 {
1693 	struct slist *tmp, *s;
1694 
1695 	s = (*b)->stmts;
1696 	(*b)->stmts = 0;
1697 	while (BPF_CLASS((*b)->s.code) == BPF_JMP && JT(*b) == JF(*b))
1698 		*b = JT(*b);
1699 
1700 	tmp = (*b)->stmts;
1701 	if (tmp != 0)
1702 		sappend(s, tmp);
1703 	(*b)->stmts = s;
1704 
1705 	/*
1706 	 * If the root node is a return, then there is no
1707 	 * point executing any statements (since the bpf machine
1708 	 * has no side effects).
1709 	 */
1710 	if (BPF_CLASS((*b)->s.code) == BPF_RET)
1711 		(*b)->stmts = 0;
1712 }
1713 
1714 static void
1715 opt_loop(root, do_stmts)
1716 	struct block *root;
1717 	int do_stmts;
1718 {
1719 
1720 #ifdef BDEBUG
1721 	if (dflag > 1) {
1722 		printf("opt_loop(root, %d) begin\n", do_stmts);
1723 		opt_dump(root);
1724 	}
1725 #endif
1726 	do {
1727 		done = 1;
1728 		find_levels(root);
1729 		find_dom(root);
1730 		find_closure(root);
1731 		find_ud(root);
1732 		find_edom(root);
1733 		opt_blks(root, do_stmts);
1734 #ifdef BDEBUG
1735 		if (dflag > 1) {
1736 			printf("opt_loop(root, %d) bottom, done=%d\n", do_stmts, done);
1737 			opt_dump(root);
1738 		}
1739 #endif
1740 	} while (!done);
1741 }
1742 
1743 /*
1744  * Optimize the filter code in its dag representation.
1745  */
1746 void
1747 bpf_optimize(rootp)
1748 	struct block **rootp;
1749 {
1750 	struct block *root;
1751 
1752 	root = *rootp;
1753 
1754 	opt_init(root);
1755 	opt_loop(root, 0);
1756 	opt_loop(root, 1);
1757 	intern_blocks(root);
1758 #ifdef BDEBUG
1759 	if (dflag > 1) {
1760 		printf("after intern_blocks()\n");
1761 		opt_dump(root);
1762 	}
1763 #endif
1764 	opt_root(rootp);
1765 #ifdef BDEBUG
1766 	if (dflag > 1) {
1767 		printf("after opt_root()\n");
1768 		opt_dump(root);
1769 	}
1770 #endif
1771 	opt_cleanup();
1772 }
1773 
1774 static void
1775 make_marks(p)
1776 	struct block *p;
1777 {
1778 	if (!isMarked(p)) {
1779 		Mark(p);
1780 		if (BPF_CLASS(p->s.code) != BPF_RET) {
1781 			make_marks(JT(p));
1782 			make_marks(JF(p));
1783 		}
1784 	}
1785 }
1786 
1787 /*
1788  * Mark code array such that isMarked(i) is true
1789  * only for nodes that are alive.
1790  */
1791 static void
1792 mark_code(p)
1793 	struct block *p;
1794 {
1795 	cur_mark += 1;
1796 	make_marks(p);
1797 }
1798 
1799 /*
1800  * True iff the two stmt lists load the same value from the packet into
1801  * the accumulator.
1802  */
1803 static int
1804 eq_slist(x, y)
1805 	struct slist *x, *y;
1806 {
1807 	while (1) {
1808 		while (x && x->s.code == NOP)
1809 			x = x->next;
1810 		while (y && y->s.code == NOP)
1811 			y = y->next;
1812 		if (x == 0)
1813 			return y == 0;
1814 		if (y == 0)
1815 			return x == 0;
1816 		if (x->s.code != y->s.code || x->s.k != y->s.k)
1817 			return 0;
1818 		x = x->next;
1819 		y = y->next;
1820 	}
1821 }
1822 
1823 static inline int
1824 eq_blk(b0, b1)
1825 	struct block *b0, *b1;
1826 {
1827 	if (b0->s.code == b1->s.code &&
1828 	    b0->s.k == b1->s.k &&
1829 	    b0->et.succ == b1->et.succ &&
1830 	    b0->ef.succ == b1->ef.succ)
1831 		return eq_slist(b0->stmts, b1->stmts);
1832 	return 0;
1833 }
1834 
1835 static void
1836 intern_blocks(root)
1837 	struct block *root;
1838 {
1839 	struct block *p;
1840 	int i, j;
1841 	int done1; /* don't shadow global */
1842  top:
1843 	done1 = 1;
1844 	for (i = 0; i < n_blocks; ++i)
1845 		blocks[i]->link = 0;
1846 
1847 	mark_code(root);
1848 
1849 	for (i = n_blocks - 1; --i >= 0; ) {
1850 		if (!isMarked(blocks[i]))
1851 			continue;
1852 		for (j = i + 1; j < n_blocks; ++j) {
1853 			if (!isMarked(blocks[j]))
1854 				continue;
1855 			if (eq_blk(blocks[i], blocks[j])) {
1856 				blocks[i]->link = blocks[j]->link ?
1857 					blocks[j]->link : blocks[j];
1858 				break;
1859 			}
1860 		}
1861 	}
1862 	for (i = 0; i < n_blocks; ++i) {
1863 		p = blocks[i];
1864 		if (JT(p) == 0)
1865 			continue;
1866 		if (JT(p)->link) {
1867 			done1 = 0;
1868 			JT(p) = JT(p)->link;
1869 		}
1870 		if (JF(p)->link) {
1871 			done1 = 0;
1872 			JF(p) = JF(p)->link;
1873 		}
1874 	}
1875 	if (!done1)
1876 		goto top;
1877 }
1878 
1879 static void
1880 opt_cleanup()
1881 {
1882 	free((void *)vnode_base);
1883 	free((void *)vmap);
1884 	free((void *)edges);
1885 	free((void *)space);
1886 	free((void *)levels);
1887 	free((void *)blocks);
1888 }
1889 
1890 /*
1891  * Return the number of stmts in 's'.
1892  */
1893 static int
1894 slength(s)
1895 	struct slist *s;
1896 {
1897 	int n = 0;
1898 
1899 	for (; s; s = s->next)
1900 		if (s->s.code != NOP)
1901 			++n;
1902 	return n;
1903 }
1904 
1905 /*
1906  * Return the number of nodes reachable by 'p'.
1907  * All nodes should be initially unmarked.
1908  */
1909 static int
1910 count_blocks(p)
1911 	struct block *p;
1912 {
1913 	if (p == 0 || isMarked(p))
1914 		return 0;
1915 	Mark(p);
1916 	return count_blocks(JT(p)) + count_blocks(JF(p)) + 1;
1917 }
1918 
1919 /*
1920  * Do a depth first search on the flow graph, numbering the
1921  * the basic blocks, and entering them into the 'blocks' array.`
1922  */
1923 static void
1924 number_blks_r(p)
1925 	struct block *p;
1926 {
1927 	int n;
1928 
1929 	if (p == 0 || isMarked(p))
1930 		return;
1931 
1932 	Mark(p);
1933 	n = n_blocks++;
1934 	p->id = n;
1935 	blocks[n] = p;
1936 
1937 	number_blks_r(JT(p));
1938 	number_blks_r(JF(p));
1939 }
1940 
1941 /*
1942  * Return the number of stmts in the flowgraph reachable by 'p'.
1943  * The nodes should be unmarked before calling.
1944  *
1945  * Note that "stmts" means "instructions", and that this includes
1946  *
1947  *	side-effect statements in 'p' (slength(p->stmts));
1948  *
1949  *	statements in the true branch from 'p' (count_stmts(JT(p)));
1950  *
1951  *	statements in the false branch from 'p' (count_stmts(JF(p)));
1952  *
1953  *	the conditional jump itself (1);
1954  *
1955  *	an extra long jump if the true branch requires it (p->longjt);
1956  *
1957  *	an extra long jump if the false branch requires it (p->longjf).
1958  */
1959 static int
1960 count_stmts(p)
1961 	struct block *p;
1962 {
1963 	int n;
1964 
1965 	if (p == 0 || isMarked(p))
1966 		return 0;
1967 	Mark(p);
1968 	n = count_stmts(JT(p)) + count_stmts(JF(p));
1969 	return slength(p->stmts) + n + 1 + p->longjt + p->longjf;
1970 }
1971 
1972 /*
1973  * Allocate memory.  All allocation is done before optimization
1974  * is begun.  A linear bound on the size of all data structures is computed
1975  * from the total number of blocks and/or statements.
1976  */
1977 static void
1978 opt_init(root)
1979 	struct block *root;
1980 {
1981 	bpf_u_int32 *p;
1982 	int i, n, max_stmts;
1983 
1984 	/*
1985 	 * First, count the blocks, so we can malloc an array to map
1986 	 * block number to block.  Then, put the blocks into the array.
1987 	 */
1988 	unMarkAll();
1989 	n = count_blocks(root);
1990 	blocks = (struct block **)calloc(n, sizeof(*blocks));
1991 	if (blocks == NULL)
1992 		bpf_error("malloc");
1993 	unMarkAll();
1994 	n_blocks = 0;
1995 	number_blks_r(root);
1996 
1997 	n_edges = 2 * n_blocks;
1998 	edges = (struct edge **)calloc(n_edges, sizeof(*edges));
1999 	if (edges == NULL)
2000 		bpf_error("malloc");
2001 
2002 	/*
2003 	 * The number of levels is bounded by the number of nodes.
2004 	 */
2005 	levels = (struct block **)calloc(n_blocks, sizeof(*levels));
2006 	if (levels == NULL)
2007 		bpf_error("malloc");
2008 
2009 	edgewords = n_edges / (8 * sizeof(bpf_u_int32)) + 1;
2010 	nodewords = n_blocks / (8 * sizeof(bpf_u_int32)) + 1;
2011 
2012 	/* XXX */
2013 	space = (bpf_u_int32 *)malloc(2 * n_blocks * nodewords * sizeof(*space)
2014 				 + n_edges * edgewords * sizeof(*space));
2015 	if (space == NULL)
2016 		bpf_error("malloc");
2017 	p = space;
2018 	all_dom_sets = p;
2019 	for (i = 0; i < n; ++i) {
2020 		blocks[i]->dom = p;
2021 		p += nodewords;
2022 	}
2023 	all_closure_sets = p;
2024 	for (i = 0; i < n; ++i) {
2025 		blocks[i]->closure = p;
2026 		p += nodewords;
2027 	}
2028 	all_edge_sets = p;
2029 	for (i = 0; i < n; ++i) {
2030 		register struct block *b = blocks[i];
2031 
2032 		b->et.edom = p;
2033 		p += edgewords;
2034 		b->ef.edom = p;
2035 		p += edgewords;
2036 		b->et.id = i;
2037 		edges[i] = &b->et;
2038 		b->ef.id = n_blocks + i;
2039 		edges[n_blocks + i] = &b->ef;
2040 		b->et.pred = b;
2041 		b->ef.pred = b;
2042 	}
2043 	max_stmts = 0;
2044 	for (i = 0; i < n; ++i)
2045 		max_stmts += slength(blocks[i]->stmts) + 1;
2046 	/*
2047 	 * We allocate at most 3 value numbers per statement,
2048 	 * so this is an upper bound on the number of valnodes
2049 	 * we'll need.
2050 	 */
2051 	maxval = 3 * max_stmts;
2052 	vmap = (struct vmapinfo *)calloc(maxval, sizeof(*vmap));
2053 	vnode_base = (struct valnode *)calloc(maxval, sizeof(*vnode_base));
2054 	if (vmap == NULL || vnode_base == NULL)
2055 		bpf_error("malloc");
2056 }
2057 
2058 /*
2059  * Some pointers used to convert the basic block form of the code,
2060  * into the array form that BPF requires.  'fstart' will point to
2061  * the malloc'd array while 'ftail' is used during the recursive traversal.
2062  */
2063 static struct bpf_insn *fstart;
2064 static struct bpf_insn *ftail;
2065 
2066 #ifdef BDEBUG
2067 int bids[1000];
2068 #endif
2069 
2070 /*
2071  * Returns true if successful.  Returns false if a branch has
2072  * an offset that is too large.  If so, we have marked that
2073  * branch so that on a subsequent iteration, it will be treated
2074  * properly.
2075  */
2076 static int
2077 convert_code_r(p)
2078 	struct block *p;
2079 {
2080 	struct bpf_insn *dst;
2081 	struct slist *src;
2082 	int slen;
2083 	u_int off;
2084 	int extrajmps;		/* number of extra jumps inserted */
2085 	struct slist **offset = NULL;
2086 
2087 	if (p == 0 || isMarked(p))
2088 		return (1);
2089 	Mark(p);
2090 
2091 	if (convert_code_r(JF(p)) == 0)
2092 		return (0);
2093 	if (convert_code_r(JT(p)) == 0)
2094 		return (0);
2095 
2096 	slen = slength(p->stmts);
2097 	dst = ftail -= (slen + 1 + p->longjt + p->longjf);
2098 		/* inflate length by any extra jumps */
2099 
2100 	p->offset = dst - fstart;
2101 
2102 	/* generate offset[] for convenience  */
2103 	if (slen) {
2104 		offset = (struct slist **)calloc(slen, sizeof(struct slist *));
2105 		if (!offset) {
2106 			bpf_error("not enough core");
2107 			/*NOTREACHED*/
2108 		}
2109 	}
2110 	src = p->stmts;
2111 	for (off = 0; off < slen && src; off++) {
2112 #if 0
2113 		printf("off=%d src=%x\n", off, src);
2114 #endif
2115 		offset[off] = src;
2116 		src = src->next;
2117 	}
2118 
2119 	off = 0;
2120 	for (src = p->stmts; src; src = src->next) {
2121 		if (src->s.code == NOP)
2122 			continue;
2123 		dst->code = (u_short)src->s.code;
2124 		dst->k = src->s.k;
2125 
2126 		/* fill block-local relative jump */
2127 		if (BPF_CLASS(src->s.code) != BPF_JMP || src->s.code == (BPF_JMP|BPF_JA)) {
2128 #if 0
2129 			if (src->s.jt || src->s.jf) {
2130 				bpf_error("illegal jmp destination");
2131 				/*NOTREACHED*/
2132 			}
2133 #endif
2134 			goto filled;
2135 		}
2136 		if (off == slen - 2)	/*???*/
2137 			goto filled;
2138 
2139 	    {
2140 		int i;
2141 		int jt, jf;
2142 		const char *ljerr = "%s for block-local relative jump: off=%d";
2143 
2144 #if 0
2145 		printf("code=%x off=%d %x %x\n", src->s.code,
2146 			off, src->s.jt, src->s.jf);
2147 #endif
2148 
2149 		if (!src->s.jt || !src->s.jf) {
2150 			bpf_error(ljerr, "no jmp destination", off);
2151 			/*NOTREACHED*/
2152 		}
2153 
2154 		jt = jf = 0;
2155 		for (i = 0; i < slen; i++) {
2156 			if (offset[i] == src->s.jt) {
2157 				if (jt) {
2158 					bpf_error(ljerr, "multiple matches", off);
2159 					/*NOTREACHED*/
2160 				}
2161 
2162 				dst->jt = i - off - 1;
2163 				jt++;
2164 			}
2165 			if (offset[i] == src->s.jf) {
2166 				if (jf) {
2167 					bpf_error(ljerr, "multiple matches", off);
2168 					/*NOTREACHED*/
2169 				}
2170 				dst->jf = i - off - 1;
2171 				jf++;
2172 			}
2173 		}
2174 		if (!jt || !jf) {
2175 			bpf_error(ljerr, "no destination found", off);
2176 			/*NOTREACHED*/
2177 		}
2178 	    }
2179 filled:
2180 		++dst;
2181 		++off;
2182 	}
2183 	if (offset)
2184 		free(offset);
2185 
2186 #ifdef BDEBUG
2187 	bids[dst - fstart] = p->id + 1;
2188 #endif
2189 	dst->code = (u_short)p->s.code;
2190 	dst->k = p->s.k;
2191 	if (JT(p)) {
2192 		extrajmps = 0;
2193 		off = JT(p)->offset - (p->offset + slen) - 1;
2194 		if (off >= 256) {
2195 		    /* offset too large for branch, must add a jump */
2196 		    if (p->longjt == 0) {
2197 		    	/* mark this instruction and retry */
2198 			p->longjt++;
2199 			return(0);
2200 		    }
2201 		    /* branch if T to following jump */
2202 		    dst->jt = extrajmps;
2203 		    extrajmps++;
2204 		    dst[extrajmps].code = BPF_JMP|BPF_JA;
2205 		    dst[extrajmps].k = off - extrajmps;
2206 		}
2207 		else
2208 		    dst->jt = off;
2209 		off = JF(p)->offset - (p->offset + slen) - 1;
2210 		if (off >= 256) {
2211 		    /* offset too large for branch, must add a jump */
2212 		    if (p->longjf == 0) {
2213 		    	/* mark this instruction and retry */
2214 			p->longjf++;
2215 			return(0);
2216 		    }
2217 		    /* branch if F to following jump */
2218 		    /* if two jumps are inserted, F goes to second one */
2219 		    dst->jf = extrajmps;
2220 		    extrajmps++;
2221 		    dst[extrajmps].code = BPF_JMP|BPF_JA;
2222 		    dst[extrajmps].k = off - extrajmps;
2223 		}
2224 		else
2225 		    dst->jf = off;
2226 	}
2227 	return (1);
2228 }
2229 
2230 
2231 /*
2232  * Convert flowgraph intermediate representation to the
2233  * BPF array representation.  Set *lenp to the number of instructions.
2234  *
2235  * This routine does *NOT* leak the memory pointed to by fp.  It *must
2236  * not* do free(fp) before returning fp; doing so would make no sense,
2237  * as the BPF array pointed to by the return value of icode_to_fcode()
2238  * must be valid - it's being returned for use in a bpf_program structure.
2239  *
2240  * If it appears that icode_to_fcode() is leaking, the problem is that
2241  * the program using pcap_compile() is failing to free the memory in
2242  * the BPF program when it's done - the leak is in the program, not in
2243  * the routine that happens to be allocating the memory.  (By analogy, if
2244  * a program calls fopen() without ever calling fclose() on the FILE *,
2245  * it will leak the FILE structure; the leak is not in fopen(), it's in
2246  * the program.)  Change the program to use pcap_freecode() when it's
2247  * done with the filter program.  See the pcap man page.
2248  */
2249 struct bpf_insn *
2250 icode_to_fcode(root, lenp)
2251 	struct block *root;
2252 	int *lenp;
2253 {
2254 	int n;
2255 	struct bpf_insn *fp;
2256 
2257 	/*
2258 	 * Loop doing convert_code_r() until no branches remain
2259 	 * with too-large offsets.
2260 	 */
2261 	while (1) {
2262 	    unMarkAll();
2263 	    n = *lenp = count_stmts(root);
2264 
2265 	    fp = (struct bpf_insn *)malloc(sizeof(*fp) * n);
2266 	    if (fp == NULL)
2267 		    bpf_error("malloc");
2268 	    memset((char *)fp, 0, sizeof(*fp) * n);
2269 	    fstart = fp;
2270 	    ftail = fp + n;
2271 
2272 	    unMarkAll();
2273 	    if (convert_code_r(root))
2274 		break;
2275 	    free(fp);
2276 	}
2277 
2278 	return fp;
2279 }
2280 
2281 /*
2282  * Make a copy of a BPF program and put it in the "fcode" member of
2283  * a "pcap_t".
2284  *
2285  * If we fail to allocate memory for the copy, fill in the "errbuf"
2286  * member of the "pcap_t" with an error message, and return -1;
2287  * otherwise, return 0.
2288  */
2289 int
2290 install_bpf_program(pcap_t *p, struct bpf_program *fp)
2291 {
2292 	size_t prog_size;
2293 
2294 	/*
2295 	 * Validate the program.
2296 	 */
2297 	if (!bpf_validate(fp->bf_insns, fp->bf_len)) {
2298 		snprintf(p->errbuf, sizeof(p->errbuf),
2299 			"BPF program is not valid");
2300 		return (-1);
2301 	}
2302 
2303 	/*
2304 	 * Free up any already installed program.
2305 	 */
2306 	pcap_freecode(&p->fcode);
2307 
2308 	prog_size = sizeof(*fp->bf_insns) * fp->bf_len;
2309 	p->fcode.bf_len = fp->bf_len;
2310 	p->fcode.bf_insns = (struct bpf_insn *)malloc(prog_size);
2311 	if (p->fcode.bf_insns == NULL) {
2312 		snprintf(p->errbuf, sizeof(p->errbuf),
2313 			 "malloc: %s", pcap_strerror(errno));
2314 		return (-1);
2315 	}
2316 	memcpy(p->fcode.bf_insns, fp->bf_insns, prog_size);
2317 	return (0);
2318 }
2319 
2320 #ifdef BDEBUG
2321 static void
2322 opt_dump(root)
2323 	struct block *root;
2324 {
2325 	struct bpf_program f;
2326 
2327 	memset(bids, 0, sizeof bids);
2328 	f.bf_insns = icode_to_fcode(root, &f.bf_len);
2329 	bpf_dump(&f, 1);
2330 	putchar('\n');
2331 	free((char *)f.bf_insns);
2332 }
2333 #endif
2334