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 
24 #ifdef HAVE_CONFIG_H
25 #include "config.h"
26 #endif
27 
28 #ifdef WIN32
29 #include <pcap-stdinc.h>
30 #else /* WIN32 */
31 #if HAVE_INTTYPES_H
32 #include <inttypes.h>
33 #elif HAVE_STDINT_H
34 #include <stdint.h>
35 #endif
36 #ifdef HAVE_SYS_BITYPES_H
37 #include <sys/bitypes.h>
38 #endif
39 #include <sys/types.h>
40 #endif /* WIN32 */
41 
42 #include <stdio.h>
43 #include <stdlib.h>
44 #include <memory.h>
45 #include <string.h>
46 
47 #include <errno.h>
48 
49 #include "pcap-int.h"
50 
51 #include "gencode.h"
52 
53 #ifdef HAVE_OS_PROTO_H
54 #include "os-proto.h"
55 #endif
56 
57 #ifdef BDEBUG
58 extern int dflag;
59 #endif
60 
61 #if defined(MSDOS) && !defined(__DJGPP__)
62 extern int _w32_ffs (int mask);
63 #define ffs _w32_ffs
64 #endif
65 
66 #if defined(WIN32) && defined (_MSC_VER)
67 int ffs(int mask);
68 #endif
69 
70 /*
71  * Represents a deleted instruction.
72  */
73 #define NOP -1
74 
75 /*
76  * Register numbers for use-def values.
77  * 0 through BPF_MEMWORDS-1 represent the corresponding scratch memory
78  * location.  A_ATOM is the accumulator and X_ATOM is the index
79  * register.
80  */
81 #define A_ATOM BPF_MEMWORDS
82 #define X_ATOM (BPF_MEMWORDS+1)
83 
84 /*
85  * This define is used to represent *both* the accumulator and
86  * x register in use-def computations.
87  * Currently, the use-def code assumes only one definition per instruction.
88  */
89 #define AX_ATOM N_ATOMS
90 
91 /*
92  * A flag to indicate that further optimization is needed.
93  * Iterative passes are continued until a given pass yields no
94  * branch movement.
95  */
96 static int done;
97 
98 /*
99  * A block is marked if only if its mark equals the current mark.
100  * Rather than traverse the code array, marking each item, 'cur_mark' is
101  * incremented.  This automatically makes each element unmarked.
102  */
103 static int cur_mark;
104 #define isMarked(p) ((p)->mark == cur_mark)
105 #define unMarkAll() cur_mark += 1
106 #define Mark(p) ((p)->mark = cur_mark)
107 
108 static void opt_init(struct block *);
109 static void opt_cleanup(void);
110 
111 static void intern_blocks(struct block *);
112 
113 static void find_inedges(struct block *);
114 #ifdef BDEBUG
115 static void opt_dump(struct block *);
116 #endif
117 
118 static int n_blocks;
119 struct block **blocks;
120 static int n_edges;
121 struct edge **edges;
122 
123 /*
124  * A bit vector set representation of the dominators.
125  * We round up the set size to the next power of two.
126  */
127 static int nodewords;
128 static int edgewords;
129 struct block **levels;
130 bpf_u_int32 *space;
131 #define BITS_PER_WORD (8*sizeof(bpf_u_int32))
132 /*
133  * True if a is in uset {p}
134  */
135 #define SET_MEMBER(p, a) \
136 ((p)[(unsigned)(a) / BITS_PER_WORD] & (1 << ((unsigned)(a) % BITS_PER_WORD)))
137 
138 /*
139  * Add 'a' to uset p.
140  */
141 #define SET_INSERT(p, a) \
142 (p)[(unsigned)(a) / BITS_PER_WORD] |= (1 << ((unsigned)(a) % BITS_PER_WORD))
143 
144 /*
145  * Delete 'a' from uset p.
146  */
147 #define SET_DELETE(p, a) \
148 (p)[(unsigned)(a) / BITS_PER_WORD] &= ~(1 << ((unsigned)(a) % BITS_PER_WORD))
149 
150 /*
151  * a := a intersect b
152  */
153 #define SET_INTERSECT(a, b, n)\
154 {\
155 	register bpf_u_int32 *_x = a, *_y = b;\
156 	register int _n = n;\
157 	while (--_n >= 0) *_x++ &= *_y++;\
158 }
159 
160 /*
161  * a := a - b
162  */
163 #define SET_SUBTRACT(a, b, n)\
164 {\
165 	register bpf_u_int32 *_x = a, *_y = b;\
166 	register int _n = n;\
167 	while (--_n >= 0) *_x++ &=~ *_y++;\
168 }
169 
170 /*
171  * a := a union b
172  */
173 #define SET_UNION(a, b, n)\
174 {\
175 	register bpf_u_int32 *_x = a, *_y = b;\
176 	register int _n = n;\
177 	while (--_n >= 0) *_x++ |= *_y++;\
178 }
179 
180 static uset all_dom_sets;
181 static uset all_closure_sets;
182 static uset all_edge_sets;
183 
184 #ifndef MAX
185 #define MAX(a,b) ((a)>(b)?(a):(b))
186 #endif
187 
188 static void
find_levels_r(struct block * b)189 find_levels_r(struct block *b)
190 {
191 	int level;
192 
193 	if (isMarked(b))
194 		return;
195 
196 	Mark(b);
197 	b->link = 0;
198 
199 	if (JT(b)) {
200 		find_levels_r(JT(b));
201 		find_levels_r(JF(b));
202 		level = MAX(JT(b)->level, JF(b)->level) + 1;
203 	} else
204 		level = 0;
205 	b->level = level;
206 	b->link = levels[level];
207 	levels[level] = b;
208 }
209 
210 /*
211  * Level graph.  The levels go from 0 at the leaves to
212  * N_LEVELS at the root.  The levels[] array points to the
213  * first node of the level list, whose elements are linked
214  * with the 'link' field of the struct block.
215  */
216 static void
find_levels(struct block * root)217 find_levels(struct block *root)
218 {
219 	memset((char *)levels, 0, n_blocks * sizeof(*levels));
220 	unMarkAll();
221 	find_levels_r(root);
222 }
223 
224 /*
225  * Find dominator relationships.
226  * Assumes graph has been leveled.
227  */
228 static void
find_dom(struct block * root)229 find_dom(struct block *root)
230 {
231 	int i;
232 	struct block *b;
233 	bpf_u_int32 *x;
234 
235 	/*
236 	 * Initialize sets to contain all nodes.
237 	 */
238 	x = all_dom_sets;
239 	i = n_blocks * nodewords;
240 	while (--i >= 0)
241 		*x++ = ~0;
242 	/* Root starts off empty. */
243 	for (i = nodewords; --i >= 0;)
244 		root->dom[i] = 0;
245 
246 	/* root->level is the highest level no found. */
247 	for (i = root->level; i >= 0; --i) {
248 		for (b = levels[i]; b; b = b->link) {
249 			SET_INSERT(b->dom, b->id);
250 			if (JT(b) == 0)
251 				continue;
252 			SET_INTERSECT(JT(b)->dom, b->dom, nodewords);
253 			SET_INTERSECT(JF(b)->dom, b->dom, nodewords);
254 		}
255 	}
256 }
257 
258 static void
propedom(struct edge * ep)259 propedom(struct edge *ep)
260 {
261 	SET_INSERT(ep->edom, ep->id);
262 	if (ep->succ) {
263 		SET_INTERSECT(ep->succ->et.edom, ep->edom, edgewords);
264 		SET_INTERSECT(ep->succ->ef.edom, ep->edom, edgewords);
265 	}
266 }
267 
268 /*
269  * Compute edge dominators.
270  * Assumes graph has been leveled and predecessors established.
271  */
272 static void
find_edom(struct block * root)273 find_edom(struct block *root)
274 {
275 	int i;
276 	uset x;
277 	struct block *b;
278 
279 	x = all_edge_sets;
280 	for (i = n_edges * edgewords; --i >= 0; )
281 		x[i] = ~0;
282 
283 	/* root->level is the highest level no found. */
284 	memset(root->et.edom, 0, edgewords * sizeof(*(uset)0));
285 	memset(root->ef.edom, 0, edgewords * sizeof(*(uset)0));
286 	for (i = root->level; i >= 0; --i) {
287 		for (b = levels[i]; b != 0; b = b->link) {
288 			propedom(&b->et);
289 			propedom(&b->ef);
290 		}
291 	}
292 }
293 
294 /*
295  * Find the backwards transitive closure of the flow graph.  These sets
296  * are backwards in the sense that we find the set of nodes that reach
297  * a given node, not the set of nodes that can be reached by a node.
298  *
299  * Assumes graph has been leveled.
300  */
301 static void
find_closure(struct block * root)302 find_closure(struct block *root)
303 {
304 	int i;
305 	struct block *b;
306 
307 	/*
308 	 * Initialize sets to contain no nodes.
309 	 */
310 	memset((char *)all_closure_sets, 0,
311 	      n_blocks * nodewords * sizeof(*all_closure_sets));
312 
313 	/* root->level is the highest level no found. */
314 	for (i = root->level; i >= 0; --i) {
315 		for (b = levels[i]; b; b = b->link) {
316 			SET_INSERT(b->closure, b->id);
317 			if (JT(b) == 0)
318 				continue;
319 			SET_UNION(JT(b)->closure, b->closure, nodewords);
320 			SET_UNION(JF(b)->closure, b->closure, nodewords);
321 		}
322 	}
323 }
324 
325 /*
326  * Return the register number that is used by s.  If A and X are both
327  * used, return AX_ATOM.  If no register is used, return -1.
328  *
329  * The implementation should probably change to an array access.
330  */
331 static int
atomuse(struct stmt * s)332 atomuse(struct stmt *s)
333 {
334 	register int c = s->code;
335 
336 	if (c == NOP)
337 		return -1;
338 
339 	switch (BPF_CLASS(c)) {
340 
341 	case BPF_RET:
342 		return (BPF_RVAL(c) == BPF_A) ? A_ATOM :
343 			(BPF_RVAL(c) == BPF_X) ? X_ATOM : -1;
344 
345 	case BPF_LD:
346 	case BPF_LDX:
347 		return (BPF_MODE(c) == BPF_IND) ? X_ATOM :
348 			(BPF_MODE(c) == BPF_MEM) ? s->k : -1;
349 
350 	case BPF_ST:
351 		return A_ATOM;
352 
353 	case BPF_STX:
354 		return X_ATOM;
355 
356 	case BPF_JMP:
357 	case BPF_ALU:
358 		if (BPF_SRC(c) == BPF_X)
359 			return AX_ATOM;
360 		return A_ATOM;
361 
362 	case BPF_MISC:
363 		return BPF_MISCOP(c) == BPF_TXA ? X_ATOM : A_ATOM;
364 	}
365 	abort();
366 	/* NOTREACHED */
367 }
368 
369 /*
370  * Return the register number that is defined by 's'.  We assume that
371  * a single stmt cannot define more than one register.  If no register
372  * is defined, return -1.
373  *
374  * The implementation should probably change to an array access.
375  */
376 static int
atomdef(struct stmt * s)377 atomdef(struct stmt *s)
378 {
379 	if (s->code == NOP)
380 		return -1;
381 
382 	switch (BPF_CLASS(s->code)) {
383 
384 	case BPF_LD:
385 	case BPF_ALU:
386 		return A_ATOM;
387 
388 	case BPF_LDX:
389 		return X_ATOM;
390 
391 	case BPF_ST:
392 	case BPF_STX:
393 		return s->k;
394 
395 	case BPF_MISC:
396 		return BPF_MISCOP(s->code) == BPF_TAX ? X_ATOM : A_ATOM;
397 	}
398 	return -1;
399 }
400 
401 /*
402  * Compute the sets of registers used, defined, and killed by 'b'.
403  *
404  * "Used" means that a statement in 'b' uses the register before any
405  * statement in 'b' defines it, i.e. it uses the value left in
406  * that register by a predecessor block of this block.
407  * "Defined" means that a statement in 'b' defines it.
408  * "Killed" means that a statement in 'b' defines it before any
409  * statement in 'b' uses it, i.e. it kills the value left in that
410  * register by a predecessor block of this block.
411  */
412 static void
compute_local_ud(struct block * b)413 compute_local_ud(struct block *b)
414 {
415 	struct slist *s;
416 	atomset def = 0, use = 0, kill = 0;
417 	int atom;
418 
419 	for (s = b->stmts; s; s = s->next) {
420 		if (s->s.code == NOP)
421 			continue;
422 		atom = atomuse(&s->s);
423 		if (atom >= 0) {
424 			if (atom == AX_ATOM) {
425 				if (!ATOMELEM(def, X_ATOM))
426 					use |= ATOMMASK(X_ATOM);
427 				if (!ATOMELEM(def, A_ATOM))
428 					use |= ATOMMASK(A_ATOM);
429 			}
430 			else if (atom < N_ATOMS) {
431 				if (!ATOMELEM(def, atom))
432 					use |= ATOMMASK(atom);
433 			}
434 			else
435 				abort();
436 		}
437 		atom = atomdef(&s->s);
438 		if (atom >= 0) {
439 			if (!ATOMELEM(use, atom))
440 				kill |= ATOMMASK(atom);
441 			def |= ATOMMASK(atom);
442 		}
443 	}
444 	if (BPF_CLASS(b->s.code) == BPF_JMP) {
445 		/*
446 		 * XXX - what about RET?
447 		 */
448 		atom = atomuse(&b->s);
449 		if (atom >= 0) {
450 			if (atom == AX_ATOM) {
451 				if (!ATOMELEM(def, X_ATOM))
452 					use |= ATOMMASK(X_ATOM);
453 				if (!ATOMELEM(def, A_ATOM))
454 					use |= ATOMMASK(A_ATOM);
455 			}
456 			else if (atom < N_ATOMS) {
457 				if (!ATOMELEM(def, atom))
458 					use |= ATOMMASK(atom);
459 			}
460 			else
461 				abort();
462 		}
463 	}
464 
465 	b->def = def;
466 	b->kill = kill;
467 	b->in_use = use;
468 }
469 
470 /*
471  * Assume graph is already leveled.
472  */
473 static void
find_ud(struct block * root)474 find_ud(struct block *root)
475 {
476 	int i, maxlevel;
477 	struct block *p;
478 
479 	/*
480 	 * root->level is the highest level no found;
481 	 * count down from there.
482 	 */
483 	maxlevel = root->level;
484 	for (i = maxlevel; i >= 0; --i)
485 		for (p = levels[i]; p; p = p->link) {
486 			compute_local_ud(p);
487 			p->out_use = 0;
488 		}
489 
490 	for (i = 1; i <= maxlevel; ++i) {
491 		for (p = levels[i]; p; p = p->link) {
492 			p->out_use |= JT(p)->in_use | JF(p)->in_use;
493 			p->in_use |= p->out_use &~ p->kill;
494 		}
495 	}
496 }
497 
498 /*
499  * These data structures are used in a Cocke and Shwarz style
500  * value numbering scheme.  Since the flowgraph is acyclic,
501  * exit values can be propagated from a node's predecessors
502  * provided it is uniquely defined.
503  */
504 struct valnode {
505 	int code;
506 	int v0, v1;
507 	int val;
508 	struct valnode *next;
509 };
510 
511 #define MODULUS 213
512 static struct valnode *hashtbl[MODULUS];
513 static int curval;
514 static int maxval;
515 
516 /* Integer constants mapped with the load immediate opcode. */
517 #define K(i) F(BPF_LD|BPF_IMM|BPF_W, i, 0L)
518 
519 struct vmapinfo {
520 	int is_const;
521 	bpf_int32 const_val;
522 };
523 
524 struct vmapinfo *vmap;
525 struct valnode *vnode_base;
526 struct valnode *next_vnode;
527 
528 static void
init_val(void)529 init_val(void)
530 {
531 	curval = 0;
532 	next_vnode = vnode_base;
533 	memset((char *)vmap, 0, maxval * sizeof(*vmap));
534 	memset((char *)hashtbl, 0, sizeof hashtbl);
535 }
536 
537 /* Because we really don't have an IR, this stuff is a little messy. */
538 static int
F(int code,int v0,int v1)539 F(int code, int v0, int v1)
540 {
541 	u_int hash;
542 	int val;
543 	struct valnode *p;
544 
545 	hash = (u_int)code ^ (v0 << 4) ^ (v1 << 8);
546 	hash %= MODULUS;
547 
548 	for (p = hashtbl[hash]; p; p = p->next)
549 		if (p->code == code && p->v0 == v0 && p->v1 == v1)
550 			return p->val;
551 
552 	val = ++curval;
553 	if (BPF_MODE(code) == BPF_IMM &&
554 	    (BPF_CLASS(code) == BPF_LD || BPF_CLASS(code) == BPF_LDX)) {
555 		vmap[val].const_val = v0;
556 		vmap[val].is_const = 1;
557 	}
558 	p = next_vnode++;
559 	p->val = val;
560 	p->code = code;
561 	p->v0 = v0;
562 	p->v1 = v1;
563 	p->next = hashtbl[hash];
564 	hashtbl[hash] = p;
565 
566 	return val;
567 }
568 
569 static inline void
vstore(struct stmt * s,int * valp,int newval,int alter)570 vstore(struct stmt *s, int *valp, int newval, int alter)
571 {
572 	if (alter && *valp == newval)
573 		s->code = NOP;
574 	else
575 		*valp = newval;
576 }
577 
578 /*
579  * Do constant-folding on binary operators.
580  * (Unary operators are handled elsewhere.)
581  */
582 static void
fold_op(struct stmt * s,int v0,int v1)583 fold_op(struct stmt *s, int v0, int v1)
584 {
585 	bpf_u_int32 a, b;
586 
587 	a = vmap[v0].const_val;
588 	b = vmap[v1].const_val;
589 
590 	switch (BPF_OP(s->code)) {
591 	case BPF_ADD:
592 		a += b;
593 		break;
594 
595 	case BPF_SUB:
596 		a -= b;
597 		break;
598 
599 	case BPF_MUL:
600 		a *= b;
601 		break;
602 
603 	case BPF_DIV:
604 		if (b == 0)
605 			bpf_error("division by zero");
606 		a /= b;
607 		break;
608 
609 	case BPF_MOD:
610 		if (b == 0)
611 			bpf_error("modulus by zero");
612 		a %= b;
613 		break;
614 
615 	case BPF_AND:
616 		a &= b;
617 		break;
618 
619 	case BPF_OR:
620 		a |= b;
621 		break;
622 
623 	case BPF_XOR:
624 		a ^= b;
625 		break;
626 
627 	case BPF_LSH:
628 		a <<= b;
629 		break;
630 
631 	case BPF_RSH:
632 		a >>= b;
633 		break;
634 
635 	default:
636 		abort();
637 	}
638 	s->k = a;
639 	s->code = BPF_LD|BPF_IMM;
640 	done = 0;
641 }
642 
643 static inline struct slist *
this_op(struct slist * s)644 this_op(struct slist *s)
645 {
646 	while (s != 0 && s->s.code == NOP)
647 		s = s->next;
648 	return s;
649 }
650 
651 static void
opt_not(struct block * b)652 opt_not(struct block *b)
653 {
654 	struct block *tmp = JT(b);
655 
656 	JT(b) = JF(b);
657 	JF(b) = tmp;
658 }
659 
660 static void
opt_peep(struct block * b)661 opt_peep(struct block *b)
662 {
663 	struct slist *s;
664 	struct slist *next, *last;
665 	int val;
666 
667 	s = b->stmts;
668 	if (s == 0)
669 		return;
670 
671 	last = s;
672 	for (/*empty*/; /*empty*/; s = next) {
673 		/*
674 		 * Skip over nops.
675 		 */
676 		s = this_op(s);
677 		if (s == 0)
678 			break;	/* nothing left in the block */
679 
680 		/*
681 		 * Find the next real instruction after that one
682 		 * (skipping nops).
683 		 */
684 		next = this_op(s->next);
685 		if (next == 0)
686 			break;	/* no next instruction */
687 		last = next;
688 
689 		/*
690 		 * st  M[k]	-->	st  M[k]
691 		 * ldx M[k]		tax
692 		 */
693 		if (s->s.code == BPF_ST &&
694 		    next->s.code == (BPF_LDX|BPF_MEM) &&
695 		    s->s.k == next->s.k) {
696 			done = 0;
697 			next->s.code = BPF_MISC|BPF_TAX;
698 		}
699 		/*
700 		 * ld  #k	-->	ldx  #k
701 		 * tax			txa
702 		 */
703 		if (s->s.code == (BPF_LD|BPF_IMM) &&
704 		    next->s.code == (BPF_MISC|BPF_TAX)) {
705 			s->s.code = BPF_LDX|BPF_IMM;
706 			next->s.code = BPF_MISC|BPF_TXA;
707 			done = 0;
708 		}
709 		/*
710 		 * This is an ugly special case, but it happens
711 		 * when you say tcp[k] or udp[k] where k is a constant.
712 		 */
713 		if (s->s.code == (BPF_LD|BPF_IMM)) {
714 			struct slist *add, *tax, *ild;
715 
716 			/*
717 			 * Check that X isn't used on exit from this
718 			 * block (which the optimizer might cause).
719 			 * We know the code generator won't generate
720 			 * any local dependencies.
721 			 */
722 			if (ATOMELEM(b->out_use, X_ATOM))
723 				continue;
724 
725 			/*
726 			 * Check that the instruction following the ldi
727 			 * is an addx, or it's an ldxms with an addx
728 			 * following it (with 0 or more nops between the
729 			 * ldxms and addx).
730 			 */
731 			if (next->s.code != (BPF_LDX|BPF_MSH|BPF_B))
732 				add = next;
733 			else
734 				add = this_op(next->next);
735 			if (add == 0 || add->s.code != (BPF_ALU|BPF_ADD|BPF_X))
736 				continue;
737 
738 			/*
739 			 * Check that a tax follows that (with 0 or more
740 			 * nops between them).
741 			 */
742 			tax = this_op(add->next);
743 			if (tax == 0 || tax->s.code != (BPF_MISC|BPF_TAX))
744 				continue;
745 
746 			/*
747 			 * Check that an ild follows that (with 0 or more
748 			 * nops between them).
749 			 */
750 			ild = this_op(tax->next);
751 			if (ild == 0 || BPF_CLASS(ild->s.code) != BPF_LD ||
752 			    BPF_MODE(ild->s.code) != BPF_IND)
753 				continue;
754 			/*
755 			 * We want to turn this sequence:
756 			 *
757 			 * (004) ldi     #0x2		{s}
758 			 * (005) ldxms   [14]		{next}  -- optional
759 			 * (006) addx			{add}
760 			 * (007) tax			{tax}
761 			 * (008) ild     [x+0]		{ild}
762 			 *
763 			 * into this sequence:
764 			 *
765 			 * (004) nop
766 			 * (005) ldxms   [14]
767 			 * (006) nop
768 			 * (007) nop
769 			 * (008) ild     [x+2]
770 			 *
771 			 * XXX We need to check that X is not
772 			 * subsequently used, because we want to change
773 			 * what'll be in it after this sequence.
774 			 *
775 			 * We know we can eliminate the accumulator
776 			 * modifications earlier in the sequence since
777 			 * it is defined by the last stmt of this sequence
778 			 * (i.e., the last statement of the sequence loads
779 			 * a value into the accumulator, so we can eliminate
780 			 * earlier operations on the accumulator).
781 			 */
782 			ild->s.k += s->s.k;
783 			s->s.code = NOP;
784 			add->s.code = NOP;
785 			tax->s.code = NOP;
786 			done = 0;
787 		}
788 	}
789 	/*
790 	 * If the comparison at the end of a block is an equality
791 	 * comparison against a constant, and nobody uses the value
792 	 * we leave in the A register at the end of a block, and
793 	 * the operation preceding the comparison is an arithmetic
794 	 * operation, we can sometime optimize it away.
795 	 */
796 	if (b->s.code == (BPF_JMP|BPF_JEQ|BPF_K) &&
797 	    !ATOMELEM(b->out_use, A_ATOM)) {
798 	    	/*
799 	    	 * We can optimize away certain subtractions of the
800 	    	 * X register.
801 	    	 */
802 		if (last->s.code == (BPF_ALU|BPF_SUB|BPF_X)) {
803 			val = b->val[X_ATOM];
804 			if (vmap[val].is_const) {
805 				/*
806 				 * If we have a subtract to do a comparison,
807 				 * and the X register is a known constant,
808 				 * we can merge this value into the
809 				 * comparison:
810 				 *
811 				 * sub x  ->	nop
812 				 * jeq #y	jeq #(x+y)
813 				 */
814 				b->s.k += vmap[val].const_val;
815 				last->s.code = NOP;
816 				done = 0;
817 			} else if (b->s.k == 0) {
818 				/*
819 				 * If the X register isn't a constant,
820 				 * and the comparison in the test is
821 				 * against 0, we can compare with the
822 				 * X register, instead:
823 				 *
824 				 * sub x  ->	nop
825 				 * jeq #0	jeq x
826 				 */
827 				last->s.code = NOP;
828 				b->s.code = BPF_JMP|BPF_JEQ|BPF_X;
829 				done = 0;
830 			}
831 		}
832 		/*
833 		 * Likewise, a constant subtract can be simplified:
834 		 *
835 		 * sub #x ->	nop
836 		 * jeq #y ->	jeq #(x+y)
837 		 */
838 		else if (last->s.code == (BPF_ALU|BPF_SUB|BPF_K)) {
839 			last->s.code = NOP;
840 			b->s.k += last->s.k;
841 			done = 0;
842 		}
843 		/*
844 		 * And, similarly, a constant AND can be simplified
845 		 * if we're testing against 0, i.e.:
846 		 *
847 		 * and #k	nop
848 		 * jeq #0  ->	jset #k
849 		 */
850 		else if (last->s.code == (BPF_ALU|BPF_AND|BPF_K) &&
851 		    b->s.k == 0) {
852 			b->s.k = last->s.k;
853 			b->s.code = BPF_JMP|BPF_K|BPF_JSET;
854 			last->s.code = NOP;
855 			done = 0;
856 			opt_not(b);
857 		}
858 	}
859 	/*
860 	 * jset #0        ->   never
861 	 * jset #ffffffff ->   always
862 	 */
863 	if (b->s.code == (BPF_JMP|BPF_K|BPF_JSET)) {
864 		if (b->s.k == 0)
865 			JT(b) = JF(b);
866 		if (b->s.k == 0xffffffff)
867 			JF(b) = JT(b);
868 	}
869 	/*
870 	 * If we're comparing against the index register, and the index
871 	 * register is a known constant, we can just compare against that
872 	 * constant.
873 	 */
874 	val = b->val[X_ATOM];
875 	if (vmap[val].is_const && BPF_SRC(b->s.code) == BPF_X) {
876 		bpf_int32 v = vmap[val].const_val;
877 		b->s.code &= ~BPF_X;
878 		b->s.k = v;
879 	}
880 	/*
881 	 * If the accumulator is a known constant, we can compute the
882 	 * comparison result.
883 	 */
884 	val = b->val[A_ATOM];
885 	if (vmap[val].is_const && BPF_SRC(b->s.code) == BPF_K) {
886 		bpf_int32 v = vmap[val].const_val;
887 		switch (BPF_OP(b->s.code)) {
888 
889 		case BPF_JEQ:
890 			v = v == b->s.k;
891 			break;
892 
893 		case BPF_JGT:
894 			v = (unsigned)v > b->s.k;
895 			break;
896 
897 		case BPF_JGE:
898 			v = (unsigned)v >= b->s.k;
899 			break;
900 
901 		case BPF_JSET:
902 			v &= b->s.k;
903 			break;
904 
905 		default:
906 			abort();
907 		}
908 		if (JF(b) != JT(b))
909 			done = 0;
910 		if (v)
911 			JF(b) = JT(b);
912 		else
913 			JT(b) = JF(b);
914 	}
915 }
916 
917 /*
918  * Compute the symbolic value of expression of 's', and update
919  * anything it defines in the value table 'val'.  If 'alter' is true,
920  * do various optimizations.  This code would be cleaner if symbolic
921  * evaluation and code transformations weren't folded together.
922  */
923 static void
opt_stmt(struct stmt * s,int val[],int alter)924 opt_stmt(struct stmt *s, int val[], int alter)
925 {
926 	int op;
927 	int v;
928 
929 	switch (s->code) {
930 
931 	case BPF_LD|BPF_ABS|BPF_W:
932 	case BPF_LD|BPF_ABS|BPF_H:
933 	case BPF_LD|BPF_ABS|BPF_B:
934 		v = F(s->code, s->k, 0L);
935 		vstore(s, &val[A_ATOM], v, alter);
936 		break;
937 
938 	case BPF_LD|BPF_IND|BPF_W:
939 	case BPF_LD|BPF_IND|BPF_H:
940 	case BPF_LD|BPF_IND|BPF_B:
941 		v = val[X_ATOM];
942 		if (alter && vmap[v].is_const) {
943 			s->code = BPF_LD|BPF_ABS|BPF_SIZE(s->code);
944 			s->k += vmap[v].const_val;
945 			v = F(s->code, s->k, 0L);
946 			done = 0;
947 		}
948 		else
949 			v = F(s->code, s->k, v);
950 		vstore(s, &val[A_ATOM], v, alter);
951 		break;
952 
953 	case BPF_LD|BPF_LEN:
954 		v = F(s->code, 0L, 0L);
955 		vstore(s, &val[A_ATOM], v, alter);
956 		break;
957 
958 	case BPF_LD|BPF_IMM:
959 		v = K(s->k);
960 		vstore(s, &val[A_ATOM], v, alter);
961 		break;
962 
963 	case BPF_LDX|BPF_IMM:
964 		v = K(s->k);
965 		vstore(s, &val[X_ATOM], v, alter);
966 		break;
967 
968 	case BPF_LDX|BPF_MSH|BPF_B:
969 		v = F(s->code, s->k, 0L);
970 		vstore(s, &val[X_ATOM], v, alter);
971 		break;
972 
973 	case BPF_ALU|BPF_NEG:
974 		if (alter && vmap[val[A_ATOM]].is_const) {
975 			s->code = BPF_LD|BPF_IMM;
976 			s->k = -vmap[val[A_ATOM]].const_val;
977 			val[A_ATOM] = K(s->k);
978 		}
979 		else
980 			val[A_ATOM] = F(s->code, val[A_ATOM], 0L);
981 		break;
982 
983 	case BPF_ALU|BPF_ADD|BPF_K:
984 	case BPF_ALU|BPF_SUB|BPF_K:
985 	case BPF_ALU|BPF_MUL|BPF_K:
986 	case BPF_ALU|BPF_DIV|BPF_K:
987 	case BPF_ALU|BPF_MOD|BPF_K:
988 	case BPF_ALU|BPF_AND|BPF_K:
989 	case BPF_ALU|BPF_OR|BPF_K:
990 	case BPF_ALU|BPF_XOR|BPF_K:
991 	case BPF_ALU|BPF_LSH|BPF_K:
992 	case BPF_ALU|BPF_RSH|BPF_K:
993 		op = BPF_OP(s->code);
994 		if (alter) {
995 			if (s->k == 0) {
996 				/* don't optimize away "sub #0"
997 				 * as it may be needed later to
998 				 * fixup the generated math code */
999 				if (op == BPF_ADD ||
1000 				    op == BPF_LSH || op == BPF_RSH ||
1001 				    op == BPF_OR || op == BPF_XOR) {
1002 					s->code = NOP;
1003 					break;
1004 				}
1005 				if (op == BPF_MUL || op == BPF_AND) {
1006 					s->code = BPF_LD|BPF_IMM;
1007 					val[A_ATOM] = K(s->k);
1008 					break;
1009 				}
1010 			}
1011 			if (vmap[val[A_ATOM]].is_const) {
1012 				fold_op(s, val[A_ATOM], K(s->k));
1013 				val[A_ATOM] = K(s->k);
1014 				break;
1015 			}
1016 		}
1017 		val[A_ATOM] = F(s->code, val[A_ATOM], K(s->k));
1018 		break;
1019 
1020 	case BPF_ALU|BPF_ADD|BPF_X:
1021 	case BPF_ALU|BPF_SUB|BPF_X:
1022 	case BPF_ALU|BPF_MUL|BPF_X:
1023 	case BPF_ALU|BPF_DIV|BPF_X:
1024 	case BPF_ALU|BPF_MOD|BPF_X:
1025 	case BPF_ALU|BPF_AND|BPF_X:
1026 	case BPF_ALU|BPF_OR|BPF_X:
1027 	case BPF_ALU|BPF_XOR|BPF_X:
1028 	case BPF_ALU|BPF_LSH|BPF_X:
1029 	case BPF_ALU|BPF_RSH|BPF_X:
1030 		op = BPF_OP(s->code);
1031 		if (alter && vmap[val[X_ATOM]].is_const) {
1032 			if (vmap[val[A_ATOM]].is_const) {
1033 				fold_op(s, val[A_ATOM], val[X_ATOM]);
1034 				val[A_ATOM] = K(s->k);
1035 			}
1036 			else {
1037 				s->code = BPF_ALU|BPF_K|op;
1038 				s->k = vmap[val[X_ATOM]].const_val;
1039 				done = 0;
1040 				val[A_ATOM] =
1041 					F(s->code, val[A_ATOM], K(s->k));
1042 			}
1043 			break;
1044 		}
1045 		/*
1046 		 * Check if we're doing something to an accumulator
1047 		 * that is 0, and simplify.  This may not seem like
1048 		 * much of a simplification but it could open up further
1049 		 * optimizations.
1050 		 * XXX We could also check for mul by 1, etc.
1051 		 */
1052 		if (alter && vmap[val[A_ATOM]].is_const
1053 		    && vmap[val[A_ATOM]].const_val == 0) {
1054 			if (op == BPF_ADD || op == BPF_OR || op == BPF_XOR) {
1055 				s->code = BPF_MISC|BPF_TXA;
1056 				vstore(s, &val[A_ATOM], val[X_ATOM], alter);
1057 				break;
1058 			}
1059 			else if (op == BPF_MUL || op == BPF_DIV || op == BPF_MOD ||
1060 				 op == BPF_AND || op == BPF_LSH || op == BPF_RSH) {
1061 				s->code = BPF_LD|BPF_IMM;
1062 				s->k = 0;
1063 				vstore(s, &val[A_ATOM], K(s->k), alter);
1064 				break;
1065 			}
1066 			else if (op == BPF_NEG) {
1067 				s->code = NOP;
1068 				break;
1069 			}
1070 		}
1071 		val[A_ATOM] = F(s->code, val[A_ATOM], val[X_ATOM]);
1072 		break;
1073 
1074 	case BPF_MISC|BPF_TXA:
1075 		vstore(s, &val[A_ATOM], val[X_ATOM], alter);
1076 		break;
1077 
1078 	case BPF_LD|BPF_MEM:
1079 		v = val[s->k];
1080 		if (alter && vmap[v].is_const) {
1081 			s->code = BPF_LD|BPF_IMM;
1082 			s->k = vmap[v].const_val;
1083 			done = 0;
1084 		}
1085 		vstore(s, &val[A_ATOM], v, alter);
1086 		break;
1087 
1088 	case BPF_MISC|BPF_TAX:
1089 		vstore(s, &val[X_ATOM], val[A_ATOM], alter);
1090 		break;
1091 
1092 	case BPF_LDX|BPF_MEM:
1093 		v = val[s->k];
1094 		if (alter && vmap[v].is_const) {
1095 			s->code = BPF_LDX|BPF_IMM;
1096 			s->k = vmap[v].const_val;
1097 			done = 0;
1098 		}
1099 		vstore(s, &val[X_ATOM], v, alter);
1100 		break;
1101 
1102 	case BPF_ST:
1103 		vstore(s, &val[s->k], val[A_ATOM], alter);
1104 		break;
1105 
1106 	case BPF_STX:
1107 		vstore(s, &val[s->k], val[X_ATOM], alter);
1108 		break;
1109 	}
1110 }
1111 
1112 static void
deadstmt(register struct stmt * s,register struct stmt * last[])1113 deadstmt(register struct stmt *s, register struct stmt *last[])
1114 {
1115 	register int atom;
1116 
1117 	atom = atomuse(s);
1118 	if (atom >= 0) {
1119 		if (atom == AX_ATOM) {
1120 			last[X_ATOM] = 0;
1121 			last[A_ATOM] = 0;
1122 		}
1123 		else
1124 			last[atom] = 0;
1125 	}
1126 	atom = atomdef(s);
1127 	if (atom >= 0) {
1128 		if (last[atom]) {
1129 			done = 0;
1130 			last[atom]->code = NOP;
1131 		}
1132 		last[atom] = s;
1133 	}
1134 }
1135 
1136 static void
opt_deadstores(register struct block * b)1137 opt_deadstores(register struct block *b)
1138 {
1139 	register struct slist *s;
1140 	register int atom;
1141 	struct stmt *last[N_ATOMS];
1142 
1143 	memset((char *)last, 0, sizeof last);
1144 
1145 	for (s = b->stmts; s != 0; s = s->next)
1146 		deadstmt(&s->s, last);
1147 	deadstmt(&b->s, last);
1148 
1149 	for (atom = 0; atom < N_ATOMS; ++atom)
1150 		if (last[atom] && !ATOMELEM(b->out_use, atom)) {
1151 			last[atom]->code = NOP;
1152 			done = 0;
1153 		}
1154 }
1155 
1156 static void
opt_blk(struct block * b,int do_stmts)1157 opt_blk(struct block *b, int do_stmts)
1158 {
1159 	struct slist *s;
1160 	struct edge *p;
1161 	int i;
1162 	bpf_int32 aval, xval;
1163 
1164 #if 0
1165 	for (s = b->stmts; s && s->next; s = s->next)
1166 		if (BPF_CLASS(s->s.code) == BPF_JMP) {
1167 			do_stmts = 0;
1168 			break;
1169 		}
1170 #endif
1171 
1172 	/*
1173 	 * Initialize the atom values.
1174 	 */
1175 	p = b->in_edges;
1176 	if (p == 0) {
1177 		/*
1178 		 * We have no predecessors, so everything is undefined
1179 		 * upon entry to this block.
1180 		 */
1181 		memset((char *)b->val, 0, sizeof(b->val));
1182 	} else {
1183 		/*
1184 		 * Inherit values from our predecessors.
1185 		 *
1186 		 * First, get the values from the predecessor along the
1187 		 * first edge leading to this node.
1188 		 */
1189 		memcpy((char *)b->val, (char *)p->pred->val, sizeof(b->val));
1190 		/*
1191 		 * Now look at all the other nodes leading to this node.
1192 		 * If, for the predecessor along that edge, a register
1193 		 * has a different value from the one we have (i.e.,
1194 		 * control paths are merging, and the merging paths
1195 		 * assign different values to that register), give the
1196 		 * register the undefined value of 0.
1197 		 */
1198 		while ((p = p->next) != NULL) {
1199 			for (i = 0; i < N_ATOMS; ++i)
1200 				if (b->val[i] != p->pred->val[i])
1201 					b->val[i] = 0;
1202 		}
1203 	}
1204 	aval = b->val[A_ATOM];
1205 	xval = b->val[X_ATOM];
1206 	for (s = b->stmts; s; s = s->next)
1207 		opt_stmt(&s->s, b->val, do_stmts);
1208 
1209 	/*
1210 	 * This is a special case: if we don't use anything from this
1211 	 * block, and we load the accumulator or index register with a
1212 	 * value that is already there, or if this block is a return,
1213 	 * eliminate all the statements.
1214 	 *
1215 	 * XXX - what if it does a store?
1216 	 *
1217 	 * XXX - why does it matter whether we use anything from this
1218 	 * block?  If the accumulator or index register doesn't change
1219 	 * its value, isn't that OK even if we use that value?
1220 	 *
1221 	 * XXX - if we load the accumulator with a different value,
1222 	 * and the block ends with a conditional branch, we obviously
1223 	 * can't eliminate it, as the branch depends on that value.
1224 	 * For the index register, the conditional branch only depends
1225 	 * on the index register value if the test is against the index
1226 	 * register value rather than a constant; if nothing uses the
1227 	 * value we put into the index register, and we're not testing
1228 	 * against the index register's value, and there aren't any
1229 	 * other problems that would keep us from eliminating this
1230 	 * block, can we eliminate it?
1231 	 */
1232 	if (do_stmts &&
1233 	    ((b->out_use == 0 && aval != 0 && b->val[A_ATOM] == aval &&
1234 	      xval != 0 && b->val[X_ATOM] == xval) ||
1235 	     BPF_CLASS(b->s.code) == BPF_RET)) {
1236 		if (b->stmts != 0) {
1237 			b->stmts = 0;
1238 			done = 0;
1239 		}
1240 	} else {
1241 		opt_peep(b);
1242 		opt_deadstores(b);
1243 	}
1244 	/*
1245 	 * Set up values for branch optimizer.
1246 	 */
1247 	if (BPF_SRC(b->s.code) == BPF_K)
1248 		b->oval = K(b->s.k);
1249 	else
1250 		b->oval = b->val[X_ATOM];
1251 	b->et.code = b->s.code;
1252 	b->ef.code = -b->s.code;
1253 }
1254 
1255 /*
1256  * Return true if any register that is used on exit from 'succ', has
1257  * an exit value that is different from the corresponding exit value
1258  * from 'b'.
1259  */
1260 static int
use_conflict(struct block * b,struct block * succ)1261 use_conflict(struct block *b, struct block *succ)
1262 {
1263 	int atom;
1264 	atomset use = succ->out_use;
1265 
1266 	if (use == 0)
1267 		return 0;
1268 
1269 	for (atom = 0; atom < N_ATOMS; ++atom)
1270 		if (ATOMELEM(use, atom))
1271 			if (b->val[atom] != succ->val[atom])
1272 				return 1;
1273 	return 0;
1274 }
1275 
1276 static struct block *
fold_edge(struct block * child,struct edge * ep)1277 fold_edge(struct block *child, struct edge *ep)
1278 {
1279 	int sense;
1280 	int aval0, aval1, oval0, oval1;
1281 	int code = ep->code;
1282 
1283 	if (code < 0) {
1284 		code = -code;
1285 		sense = 0;
1286 	} else
1287 		sense = 1;
1288 
1289 	if (child->s.code != code)
1290 		return 0;
1291 
1292 	aval0 = child->val[A_ATOM];
1293 	oval0 = child->oval;
1294 	aval1 = ep->pred->val[A_ATOM];
1295 	oval1 = ep->pred->oval;
1296 
1297 	if (aval0 != aval1)
1298 		return 0;
1299 
1300 	if (oval0 == oval1)
1301 		/*
1302 		 * The operands of the branch instructions are
1303 		 * identical, so the result is true if a true
1304 		 * branch was taken to get here, otherwise false.
1305 		 */
1306 		return sense ? JT(child) : JF(child);
1307 
1308 	if (sense && code == (BPF_JMP|BPF_JEQ|BPF_K))
1309 		/*
1310 		 * At this point, we only know the comparison if we
1311 		 * came down the true branch, and it was an equality
1312 		 * comparison with a constant.
1313 		 *
1314 		 * I.e., if we came down the true branch, and the branch
1315 		 * was an equality comparison with a constant, we know the
1316 		 * accumulator contains that constant.  If we came down
1317 		 * the false branch, or the comparison wasn't with a
1318 		 * constant, we don't know what was in the accumulator.
1319 		 *
1320 		 * We rely on the fact that distinct constants have distinct
1321 		 * value numbers.
1322 		 */
1323 		return JF(child);
1324 
1325 	return 0;
1326 }
1327 
1328 static void
opt_j(struct edge * ep)1329 opt_j(struct edge *ep)
1330 {
1331 	register int i, k;
1332 	register struct block *target;
1333 
1334 	if (JT(ep->succ) == 0)
1335 		return;
1336 
1337 	if (JT(ep->succ) == JF(ep->succ)) {
1338 		/*
1339 		 * Common branch targets can be eliminated, provided
1340 		 * there is no data dependency.
1341 		 */
1342 		if (!use_conflict(ep->pred, ep->succ->et.succ)) {
1343 			done = 0;
1344 			ep->succ = JT(ep->succ);
1345 		}
1346 	}
1347 	/*
1348 	 * For each edge dominator that matches the successor of this
1349 	 * edge, promote the edge successor to the its grandchild.
1350 	 *
1351 	 * XXX We violate the set abstraction here in favor a reasonably
1352 	 * efficient loop.
1353 	 */
1354  top:
1355 	for (i = 0; i < edgewords; ++i) {
1356 		register bpf_u_int32 x = ep->edom[i];
1357 
1358 		while (x != 0) {
1359 			k = ffs(x) - 1;
1360 			x &=~ (1 << k);
1361 			k += i * BITS_PER_WORD;
1362 
1363 			target = fold_edge(ep->succ, edges[k]);
1364 			/*
1365 			 * Check that there is no data dependency between
1366 			 * nodes that will be violated if we move the edge.
1367 			 */
1368 			if (target != 0 && !use_conflict(ep->pred, target)) {
1369 				done = 0;
1370 				ep->succ = target;
1371 				if (JT(target) != 0)
1372 					/*
1373 					 * Start over unless we hit a leaf.
1374 					 */
1375 					goto top;
1376 				return;
1377 			}
1378 		}
1379 	}
1380 }
1381 
1382 
1383 static void
or_pullup(struct block * b)1384 or_pullup(struct block *b)
1385 {
1386 	int val, at_top;
1387 	struct block *pull;
1388 	struct block **diffp, **samep;
1389 	struct edge *ep;
1390 
1391 	ep = b->in_edges;
1392 	if (ep == 0)
1393 		return;
1394 
1395 	/*
1396 	 * Make sure each predecessor loads the same value.
1397 	 * XXX why?
1398 	 */
1399 	val = ep->pred->val[A_ATOM];
1400 	for (ep = ep->next; ep != 0; ep = ep->next)
1401 		if (val != ep->pred->val[A_ATOM])
1402 			return;
1403 
1404 	if (JT(b->in_edges->pred) == b)
1405 		diffp = &JT(b->in_edges->pred);
1406 	else
1407 		diffp = &JF(b->in_edges->pred);
1408 
1409 	at_top = 1;
1410 	while (1) {
1411 		if (*diffp == 0)
1412 			return;
1413 
1414 		if (JT(*diffp) != JT(b))
1415 			return;
1416 
1417 		if (!SET_MEMBER((*diffp)->dom, b->id))
1418 			return;
1419 
1420 		if ((*diffp)->val[A_ATOM] != val)
1421 			break;
1422 
1423 		diffp = &JF(*diffp);
1424 		at_top = 0;
1425 	}
1426 	samep = &JF(*diffp);
1427 	while (1) {
1428 		if (*samep == 0)
1429 			return;
1430 
1431 		if (JT(*samep) != JT(b))
1432 			return;
1433 
1434 		if (!SET_MEMBER((*samep)->dom, b->id))
1435 			return;
1436 
1437 		if ((*samep)->val[A_ATOM] == val)
1438 			break;
1439 
1440 		/* XXX Need to check that there are no data dependencies
1441 		   between dp0 and dp1.  Currently, the code generator
1442 		   will not produce such dependencies. */
1443 		samep = &JF(*samep);
1444 	}
1445 #ifdef notdef
1446 	/* XXX This doesn't cover everything. */
1447 	for (i = 0; i < N_ATOMS; ++i)
1448 		if ((*samep)->val[i] != pred->val[i])
1449 			return;
1450 #endif
1451 	/* Pull up the node. */
1452 	pull = *samep;
1453 	*samep = JF(pull);
1454 	JF(pull) = *diffp;
1455 
1456 	/*
1457 	 * At the top of the chain, each predecessor needs to point at the
1458 	 * pulled up node.  Inside the chain, there is only one predecessor
1459 	 * to worry about.
1460 	 */
1461 	if (at_top) {
1462 		for (ep = b->in_edges; ep != 0; ep = ep->next) {
1463 			if (JT(ep->pred) == b)
1464 				JT(ep->pred) = pull;
1465 			else
1466 				JF(ep->pred) = pull;
1467 		}
1468 	}
1469 	else
1470 		*diffp = pull;
1471 
1472 	done = 0;
1473 }
1474 
1475 static void
and_pullup(struct block * b)1476 and_pullup(struct block *b)
1477 {
1478 	int val, at_top;
1479 	struct block *pull;
1480 	struct block **diffp, **samep;
1481 	struct edge *ep;
1482 
1483 	ep = b->in_edges;
1484 	if (ep == 0)
1485 		return;
1486 
1487 	/*
1488 	 * Make sure each predecessor loads the same value.
1489 	 */
1490 	val = ep->pred->val[A_ATOM];
1491 	for (ep = ep->next; ep != 0; ep = ep->next)
1492 		if (val != ep->pred->val[A_ATOM])
1493 			return;
1494 
1495 	if (JT(b->in_edges->pred) == b)
1496 		diffp = &JT(b->in_edges->pred);
1497 	else
1498 		diffp = &JF(b->in_edges->pred);
1499 
1500 	at_top = 1;
1501 	while (1) {
1502 		if (*diffp == 0)
1503 			return;
1504 
1505 		if (JF(*diffp) != JF(b))
1506 			return;
1507 
1508 		if (!SET_MEMBER((*diffp)->dom, b->id))
1509 			return;
1510 
1511 		if ((*diffp)->val[A_ATOM] != val)
1512 			break;
1513 
1514 		diffp = &JT(*diffp);
1515 		at_top = 0;
1516 	}
1517 	samep = &JT(*diffp);
1518 	while (1) {
1519 		if (*samep == 0)
1520 			return;
1521 
1522 		if (JF(*samep) != JF(b))
1523 			return;
1524 
1525 		if (!SET_MEMBER((*samep)->dom, b->id))
1526 			return;
1527 
1528 		if ((*samep)->val[A_ATOM] == val)
1529 			break;
1530 
1531 		/* XXX Need to check that there are no data dependencies
1532 		   between diffp and samep.  Currently, the code generator
1533 		   will not produce such dependencies. */
1534 		samep = &JT(*samep);
1535 	}
1536 #ifdef notdef
1537 	/* XXX This doesn't cover everything. */
1538 	for (i = 0; i < N_ATOMS; ++i)
1539 		if ((*samep)->val[i] != pred->val[i])
1540 			return;
1541 #endif
1542 	/* Pull up the node. */
1543 	pull = *samep;
1544 	*samep = JT(pull);
1545 	JT(pull) = *diffp;
1546 
1547 	/*
1548 	 * At the top of the chain, each predecessor needs to point at the
1549 	 * pulled up node.  Inside the chain, there is only one predecessor
1550 	 * to worry about.
1551 	 */
1552 	if (at_top) {
1553 		for (ep = b->in_edges; ep != 0; ep = ep->next) {
1554 			if (JT(ep->pred) == b)
1555 				JT(ep->pred) = pull;
1556 			else
1557 				JF(ep->pred) = pull;
1558 		}
1559 	}
1560 	else
1561 		*diffp = pull;
1562 
1563 	done = 0;
1564 }
1565 
1566 static void
opt_blks(struct block * root,int do_stmts)1567 opt_blks(struct block *root, int do_stmts)
1568 {
1569 	int i, maxlevel;
1570 	struct block *p;
1571 
1572 	init_val();
1573 	maxlevel = root->level;
1574 
1575 	find_inedges(root);
1576 	for (i = maxlevel; i >= 0; --i)
1577 		for (p = levels[i]; p; p = p->link)
1578 			opt_blk(p, do_stmts);
1579 
1580 	if (do_stmts)
1581 		/*
1582 		 * No point trying to move branches; it can't possibly
1583 		 * make a difference at this point.
1584 		 */
1585 		return;
1586 
1587 	for (i = 1; i <= maxlevel; ++i) {
1588 		for (p = levels[i]; p; p = p->link) {
1589 			opt_j(&p->et);
1590 			opt_j(&p->ef);
1591 		}
1592 	}
1593 
1594 	find_inedges(root);
1595 	for (i = 1; i <= maxlevel; ++i) {
1596 		for (p = levels[i]; p; p = p->link) {
1597 			or_pullup(p);
1598 			and_pullup(p);
1599 		}
1600 	}
1601 }
1602 
1603 static inline void
link_inedge(struct edge * parent,struct block * child)1604 link_inedge(struct edge *parent, struct block *child)
1605 {
1606 	parent->next = child->in_edges;
1607 	child->in_edges = parent;
1608 }
1609 
1610 static void
find_inedges(struct block * root)1611 find_inedges(struct block *root)
1612 {
1613 	int i;
1614 	struct block *b;
1615 
1616 	for (i = 0; i < n_blocks; ++i)
1617 		blocks[i]->in_edges = 0;
1618 
1619 	/*
1620 	 * Traverse the graph, adding each edge to the predecessor
1621 	 * list of its successors.  Skip the leaves (i.e. level 0).
1622 	 */
1623 	for (i = root->level; i > 0; --i) {
1624 		for (b = levels[i]; b != 0; b = b->link) {
1625 			link_inedge(&b->et, JT(b));
1626 			link_inedge(&b->ef, JF(b));
1627 		}
1628 	}
1629 }
1630 
1631 static void
opt_root(struct block ** b)1632 opt_root(struct block **b)
1633 {
1634 	struct slist *tmp, *s;
1635 
1636 	s = (*b)->stmts;
1637 	(*b)->stmts = 0;
1638 	while (BPF_CLASS((*b)->s.code) == BPF_JMP && JT(*b) == JF(*b))
1639 		*b = JT(*b);
1640 
1641 	tmp = (*b)->stmts;
1642 	if (tmp != 0)
1643 		sappend(s, tmp);
1644 	(*b)->stmts = s;
1645 
1646 	/*
1647 	 * If the root node is a return, then there is no
1648 	 * point executing any statements (since the bpf machine
1649 	 * has no side effects).
1650 	 */
1651 	if (BPF_CLASS((*b)->s.code) == BPF_RET)
1652 		(*b)->stmts = 0;
1653 }
1654 
1655 static void
opt_loop(struct block * root,int do_stmts)1656 opt_loop(struct block *root, int do_stmts)
1657 {
1658 
1659 #ifdef BDEBUG
1660 	if (dflag > 1) {
1661 		printf("opt_loop(root, %d) begin\n", do_stmts);
1662 		opt_dump(root);
1663 	}
1664 #endif
1665 	do {
1666 		done = 1;
1667 		find_levels(root);
1668 		find_dom(root);
1669 		find_closure(root);
1670 		find_ud(root);
1671 		find_edom(root);
1672 		opt_blks(root, do_stmts);
1673 #ifdef BDEBUG
1674 		if (dflag > 1) {
1675 			printf("opt_loop(root, %d) bottom, done=%d\n", do_stmts, done);
1676 			opt_dump(root);
1677 		}
1678 #endif
1679 	} while (!done);
1680 }
1681 
1682 /*
1683  * Optimize the filter code in its dag representation.
1684  */
1685 void
bpf_optimize(struct block ** rootp)1686 bpf_optimize(struct block **rootp)
1687 {
1688 	struct block *root;
1689 
1690 	root = *rootp;
1691 
1692 	opt_init(root);
1693 	opt_loop(root, 0);
1694 	opt_loop(root, 1);
1695 	intern_blocks(root);
1696 #ifdef BDEBUG
1697 	if (dflag > 1) {
1698 		printf("after intern_blocks()\n");
1699 		opt_dump(root);
1700 	}
1701 #endif
1702 	opt_root(rootp);
1703 #ifdef BDEBUG
1704 	if (dflag > 1) {
1705 		printf("after opt_root()\n");
1706 		opt_dump(root);
1707 	}
1708 #endif
1709 	opt_cleanup();
1710 }
1711 
1712 static void
make_marks(struct block * p)1713 make_marks(struct block *p)
1714 {
1715 	if (!isMarked(p)) {
1716 		Mark(p);
1717 		if (BPF_CLASS(p->s.code) != BPF_RET) {
1718 			make_marks(JT(p));
1719 			make_marks(JF(p));
1720 		}
1721 	}
1722 }
1723 
1724 /*
1725  * Mark code array such that isMarked(i) is true
1726  * only for nodes that are alive.
1727  */
1728 static void
mark_code(struct block * p)1729 mark_code(struct block *p)
1730 {
1731 	cur_mark += 1;
1732 	make_marks(p);
1733 }
1734 
1735 /*
1736  * True iff the two stmt lists load the same value from the packet into
1737  * the accumulator.
1738  */
1739 static int
eq_slist(struct slist * x,struct slist * y)1740 eq_slist(struct slist *x, struct slist *y)
1741 {
1742 	while (1) {
1743 		while (x && x->s.code == NOP)
1744 			x = x->next;
1745 		while (y && y->s.code == NOP)
1746 			y = y->next;
1747 		if (x == 0)
1748 			return y == 0;
1749 		if (y == 0)
1750 			return x == 0;
1751 		if (x->s.code != y->s.code || x->s.k != y->s.k)
1752 			return 0;
1753 		x = x->next;
1754 		y = y->next;
1755 	}
1756 }
1757 
1758 static inline int
eq_blk(struct block * b0,struct block * b1)1759 eq_blk(struct block *b0, struct block *b1)
1760 {
1761 	if (b0->s.code == b1->s.code &&
1762 	    b0->s.k == b1->s.k &&
1763 	    b0->et.succ == b1->et.succ &&
1764 	    b0->ef.succ == b1->ef.succ)
1765 		return eq_slist(b0->stmts, b1->stmts);
1766 	return 0;
1767 }
1768 
1769 static void
intern_blocks(struct block * root)1770 intern_blocks(struct block *root)
1771 {
1772 	struct block *p;
1773 	int i, j;
1774 	int done1; /* don't shadow global */
1775  top:
1776 	done1 = 1;
1777 	for (i = 0; i < n_blocks; ++i)
1778 		blocks[i]->link = 0;
1779 
1780 	mark_code(root);
1781 
1782 	for (i = n_blocks - 1; --i >= 0; ) {
1783 		if (!isMarked(blocks[i]))
1784 			continue;
1785 		for (j = i + 1; j < n_blocks; ++j) {
1786 			if (!isMarked(blocks[j]))
1787 				continue;
1788 			if (eq_blk(blocks[i], blocks[j])) {
1789 				blocks[i]->link = blocks[j]->link ?
1790 					blocks[j]->link : blocks[j];
1791 				break;
1792 			}
1793 		}
1794 	}
1795 	for (i = 0; i < n_blocks; ++i) {
1796 		p = blocks[i];
1797 		if (JT(p) == 0)
1798 			continue;
1799 		if (JT(p)->link) {
1800 			done1 = 0;
1801 			JT(p) = JT(p)->link;
1802 		}
1803 		if (JF(p)->link) {
1804 			done1 = 0;
1805 			JF(p) = JF(p)->link;
1806 		}
1807 	}
1808 	if (!done1)
1809 		goto top;
1810 }
1811 
1812 static void
opt_cleanup(void)1813 opt_cleanup(void)
1814 {
1815 	free((void *)vnode_base);
1816 	free((void *)vmap);
1817 	free((void *)edges);
1818 	free((void *)space);
1819 	free((void *)levels);
1820 	free((void *)blocks);
1821 }
1822 
1823 /*
1824  * Return the number of stmts in 's'.
1825  */
1826 static u_int
slength(struct slist * s)1827 slength(struct slist *s)
1828 {
1829 	u_int n = 0;
1830 
1831 	for (; s; s = s->next)
1832 		if (s->s.code != NOP)
1833 			++n;
1834 	return n;
1835 }
1836 
1837 /*
1838  * Return the number of nodes reachable by 'p'.
1839  * All nodes should be initially unmarked.
1840  */
1841 static int
count_blocks(struct block * p)1842 count_blocks(struct block *p)
1843 {
1844 	if (p == 0 || isMarked(p))
1845 		return 0;
1846 	Mark(p);
1847 	return count_blocks(JT(p)) + count_blocks(JF(p)) + 1;
1848 }
1849 
1850 /*
1851  * Do a depth first search on the flow graph, numbering the
1852  * the basic blocks, and entering them into the 'blocks' array.`
1853  */
1854 static void
number_blks_r(struct block * p)1855 number_blks_r(struct block *p)
1856 {
1857 	int n;
1858 
1859 	if (p == 0 || isMarked(p))
1860 		return;
1861 
1862 	Mark(p);
1863 	n = n_blocks++;
1864 	p->id = n;
1865 	blocks[n] = p;
1866 
1867 	number_blks_r(JT(p));
1868 	number_blks_r(JF(p));
1869 }
1870 
1871 /*
1872  * Return the number of stmts in the flowgraph reachable by 'p'.
1873  * The nodes should be unmarked before calling.
1874  *
1875  * Note that "stmts" means "instructions", and that this includes
1876  *
1877  *	side-effect statements in 'p' (slength(p->stmts));
1878  *
1879  *	statements in the true branch from 'p' (count_stmts(JT(p)));
1880  *
1881  *	statements in the false branch from 'p' (count_stmts(JF(p)));
1882  *
1883  *	the conditional jump itself (1);
1884  *
1885  *	an extra long jump if the true branch requires it (p->longjt);
1886  *
1887  *	an extra long jump if the false branch requires it (p->longjf).
1888  */
1889 static u_int
count_stmts(struct block * p)1890 count_stmts(struct block *p)
1891 {
1892 	u_int n;
1893 
1894 	if (p == 0 || isMarked(p))
1895 		return 0;
1896 	Mark(p);
1897 	n = count_stmts(JT(p)) + count_stmts(JF(p));
1898 	return slength(p->stmts) + n + 1 + p->longjt + p->longjf;
1899 }
1900 
1901 /*
1902  * Allocate memory.  All allocation is done before optimization
1903  * is begun.  A linear bound on the size of all data structures is computed
1904  * from the total number of blocks and/or statements.
1905  */
1906 static void
opt_init(struct block * root)1907 opt_init(struct block *root)
1908 {
1909 	bpf_u_int32 *p;
1910 	int i, n, max_stmts;
1911 
1912 	/*
1913 	 * First, count the blocks, so we can malloc an array to map
1914 	 * block number to block.  Then, put the blocks into the array.
1915 	 */
1916 	unMarkAll();
1917 	n = count_blocks(root);
1918 	blocks = (struct block **)calloc(n, sizeof(*blocks));
1919 	if (blocks == NULL)
1920 		bpf_error("malloc");
1921 	unMarkAll();
1922 	n_blocks = 0;
1923 	number_blks_r(root);
1924 
1925 	n_edges = 2 * n_blocks;
1926 	edges = (struct edge **)calloc(n_edges, sizeof(*edges));
1927 	if (edges == NULL)
1928 		bpf_error("malloc");
1929 
1930 	/*
1931 	 * The number of levels is bounded by the number of nodes.
1932 	 */
1933 	levels = (struct block **)calloc(n_blocks, sizeof(*levels));
1934 	if (levels == NULL)
1935 		bpf_error("malloc");
1936 
1937 	edgewords = n_edges / (8 * sizeof(bpf_u_int32)) + 1;
1938 	nodewords = n_blocks / (8 * sizeof(bpf_u_int32)) + 1;
1939 
1940 	/* XXX */
1941 	space = (bpf_u_int32 *)malloc(2 * n_blocks * nodewords * sizeof(*space)
1942 				 + n_edges * edgewords * sizeof(*space));
1943 	if (space == NULL)
1944 		bpf_error("malloc");
1945 	p = space;
1946 	all_dom_sets = p;
1947 	for (i = 0; i < n; ++i) {
1948 		blocks[i]->dom = p;
1949 		p += nodewords;
1950 	}
1951 	all_closure_sets = p;
1952 	for (i = 0; i < n; ++i) {
1953 		blocks[i]->closure = p;
1954 		p += nodewords;
1955 	}
1956 	all_edge_sets = p;
1957 	for (i = 0; i < n; ++i) {
1958 		register struct block *b = blocks[i];
1959 
1960 		b->et.edom = p;
1961 		p += edgewords;
1962 		b->ef.edom = p;
1963 		p += edgewords;
1964 		b->et.id = i;
1965 		edges[i] = &b->et;
1966 		b->ef.id = n_blocks + i;
1967 		edges[n_blocks + i] = &b->ef;
1968 		b->et.pred = b;
1969 		b->ef.pred = b;
1970 	}
1971 	max_stmts = 0;
1972 	for (i = 0; i < n; ++i)
1973 		max_stmts += slength(blocks[i]->stmts) + 1;
1974 	/*
1975 	 * We allocate at most 3 value numbers per statement,
1976 	 * so this is an upper bound on the number of valnodes
1977 	 * we'll need.
1978 	 */
1979 	maxval = 3 * max_stmts;
1980 	vmap = (struct vmapinfo *)calloc(maxval, sizeof(*vmap));
1981 	vnode_base = (struct valnode *)calloc(maxval, sizeof(*vnode_base));
1982 	if (vmap == NULL || vnode_base == NULL)
1983 		bpf_error("malloc");
1984 }
1985 
1986 /*
1987  * Some pointers used to convert the basic block form of the code,
1988  * into the array form that BPF requires.  'fstart' will point to
1989  * the malloc'd array while 'ftail' is used during the recursive traversal.
1990  */
1991 static struct bpf_insn *fstart;
1992 static struct bpf_insn *ftail;
1993 
1994 #ifdef BDEBUG
1995 int bids[1000];
1996 #endif
1997 
1998 /*
1999  * Returns true if successful.  Returns false if a branch has
2000  * an offset that is too large.  If so, we have marked that
2001  * branch so that on a subsequent iteration, it will be treated
2002  * properly.
2003  */
2004 static int
convert_code_r(struct block * p)2005 convert_code_r(struct block *p)
2006 {
2007 	struct bpf_insn *dst;
2008 	struct slist *src;
2009 	int slen;
2010 	u_int off;
2011 	int extrajmps;		/* number of extra jumps inserted */
2012 	struct slist **offset = NULL;
2013 
2014 	if (p == 0 || isMarked(p))
2015 		return (1);
2016 	Mark(p);
2017 
2018 	if (convert_code_r(JF(p)) == 0)
2019 		return (0);
2020 	if (convert_code_r(JT(p)) == 0)
2021 		return (0);
2022 
2023 	slen = slength(p->stmts);
2024 	dst = ftail -= (slen + 1 + p->longjt + p->longjf);
2025 		/* inflate length by any extra jumps */
2026 
2027 	p->offset = dst - fstart;
2028 
2029 	/* generate offset[] for convenience  */
2030 	if (slen) {
2031 		offset = (struct slist **)calloc(slen, sizeof(struct slist *));
2032 		if (!offset) {
2033 			bpf_error("not enough core");
2034 			/*NOTREACHED*/
2035 		}
2036 	}
2037 	src = p->stmts;
2038 	for (off = 0; off < slen && src; off++) {
2039 #if 0
2040 		printf("off=%d src=%x\n", off, src);
2041 #endif
2042 		offset[off] = src;
2043 		src = src->next;
2044 	}
2045 
2046 	off = 0;
2047 	for (src = p->stmts; src; src = src->next) {
2048 		if (src->s.code == NOP)
2049 			continue;
2050 		dst->code = (u_short)src->s.code;
2051 		dst->k = src->s.k;
2052 
2053 		/* fill block-local relative jump */
2054 		if (BPF_CLASS(src->s.code) != BPF_JMP || src->s.code == (BPF_JMP|BPF_JA)) {
2055 #if 0
2056 			if (src->s.jt || src->s.jf) {
2057 				bpf_error("illegal jmp destination");
2058 				/*NOTREACHED*/
2059 			}
2060 #endif
2061 			goto filled;
2062 		}
2063 		if (off == slen - 2)	/*???*/
2064 			goto filled;
2065 
2066 	    {
2067 		int i;
2068 		int jt, jf;
2069 		const char *ljerr = "%s for block-local relative jump: off=%d";
2070 
2071 #if 0
2072 		printf("code=%x off=%d %x %x\n", src->s.code,
2073 			off, src->s.jt, src->s.jf);
2074 #endif
2075 
2076 		if (!src->s.jt || !src->s.jf) {
2077 			bpf_error(ljerr, "no jmp destination", off);
2078 			/*NOTREACHED*/
2079 		}
2080 
2081 		jt = jf = 0;
2082 		for (i = 0; i < slen; i++) {
2083 			if (offset[i] == src->s.jt) {
2084 				if (jt) {
2085 					bpf_error(ljerr, "multiple matches", off);
2086 					/*NOTREACHED*/
2087 				}
2088 
2089 				dst->jt = i - off - 1;
2090 				jt++;
2091 			}
2092 			if (offset[i] == src->s.jf) {
2093 				if (jf) {
2094 					bpf_error(ljerr, "multiple matches", off);
2095 					/*NOTREACHED*/
2096 				}
2097 				dst->jf = i - off - 1;
2098 				jf++;
2099 			}
2100 		}
2101 		if (!jt || !jf) {
2102 			bpf_error(ljerr, "no destination found", off);
2103 			/*NOTREACHED*/
2104 		}
2105 	    }
2106 filled:
2107 		++dst;
2108 		++off;
2109 	}
2110 	if (offset)
2111 		free(offset);
2112 
2113 #ifdef BDEBUG
2114 	bids[dst - fstart] = p->id + 1;
2115 #endif
2116 	dst->code = (u_short)p->s.code;
2117 	dst->k = p->s.k;
2118 	if (JT(p)) {
2119 		extrajmps = 0;
2120 		off = JT(p)->offset - (p->offset + slen) - 1;
2121 		if (off >= 256) {
2122 		    /* offset too large for branch, must add a jump */
2123 		    if (p->longjt == 0) {
2124 		    	/* mark this instruction and retry */
2125 			p->longjt++;
2126 			return(0);
2127 		    }
2128 		    /* branch if T to following jump */
2129 		    dst->jt = extrajmps;
2130 		    extrajmps++;
2131 		    dst[extrajmps].code = BPF_JMP|BPF_JA;
2132 		    dst[extrajmps].k = off - extrajmps;
2133 		}
2134 		else
2135 		    dst->jt = off;
2136 		off = JF(p)->offset - (p->offset + slen) - 1;
2137 		if (off >= 256) {
2138 		    /* offset too large for branch, must add a jump */
2139 		    if (p->longjf == 0) {
2140 		    	/* mark this instruction and retry */
2141 			p->longjf++;
2142 			return(0);
2143 		    }
2144 		    /* branch if F to following jump */
2145 		    /* if two jumps are inserted, F goes to second one */
2146 		    dst->jf = extrajmps;
2147 		    extrajmps++;
2148 		    dst[extrajmps].code = BPF_JMP|BPF_JA;
2149 		    dst[extrajmps].k = off - extrajmps;
2150 		}
2151 		else
2152 		    dst->jf = off;
2153 	}
2154 	return (1);
2155 }
2156 
2157 
2158 /*
2159  * Convert flowgraph intermediate representation to the
2160  * BPF array representation.  Set *lenp to the number of instructions.
2161  *
2162  * This routine does *NOT* leak the memory pointed to by fp.  It *must
2163  * not* do free(fp) before returning fp; doing so would make no sense,
2164  * as the BPF array pointed to by the return value of icode_to_fcode()
2165  * must be valid - it's being returned for use in a bpf_program structure.
2166  *
2167  * If it appears that icode_to_fcode() is leaking, the problem is that
2168  * the program using pcap_compile() is failing to free the memory in
2169  * the BPF program when it's done - the leak is in the program, not in
2170  * the routine that happens to be allocating the memory.  (By analogy, if
2171  * a program calls fopen() without ever calling fclose() on the FILE *,
2172  * it will leak the FILE structure; the leak is not in fopen(), it's in
2173  * the program.)  Change the program to use pcap_freecode() when it's
2174  * done with the filter program.  See the pcap man page.
2175  */
2176 struct bpf_insn *
icode_to_fcode(struct block * root,u_int * lenp)2177 icode_to_fcode(struct block *root, u_int *lenp)
2178 {
2179 	u_int n;
2180 	struct bpf_insn *fp;
2181 
2182 	/*
2183 	 * Loop doing convert_code_r() until no branches remain
2184 	 * with too-large offsets.
2185 	 */
2186 	while (1) {
2187 	    unMarkAll();
2188 	    n = *lenp = count_stmts(root);
2189 
2190 	    fp = (struct bpf_insn *)malloc(sizeof(*fp) * n);
2191 	    if (fp == NULL)
2192 		    bpf_error("malloc");
2193 	    memset((char *)fp, 0, sizeof(*fp) * n);
2194 	    fstart = fp;
2195 	    ftail = fp + n;
2196 
2197 	    unMarkAll();
2198 	    if (convert_code_r(root))
2199 		break;
2200 	    free(fp);
2201 	}
2202 
2203 	return fp;
2204 }
2205 
2206 /*
2207  * Make a copy of a BPF program and put it in the "fcode" member of
2208  * a "pcap_t".
2209  *
2210  * If we fail to allocate memory for the copy, fill in the "errbuf"
2211  * member of the "pcap_t" with an error message, and return -1;
2212  * otherwise, return 0.
2213  */
2214 int
install_bpf_program(pcap_t * p,struct bpf_program * fp)2215 install_bpf_program(pcap_t *p, struct bpf_program *fp)
2216 {
2217 	size_t prog_size;
2218 
2219 	/*
2220 	 * Validate the program.
2221 	 */
2222 	if (!bpf_validate(fp->bf_insns, fp->bf_len)) {
2223 		snprintf(p->errbuf, sizeof(p->errbuf),
2224 			"BPF program is not valid");
2225 		return (-1);
2226 	}
2227 
2228 	/*
2229 	 * Free up any already installed program.
2230 	 */
2231 	pcap_freecode(&p->fcode);
2232 
2233 	prog_size = sizeof(*fp->bf_insns) * fp->bf_len;
2234 	p->fcode.bf_len = fp->bf_len;
2235 	p->fcode.bf_insns = (struct bpf_insn *)malloc(prog_size);
2236 	if (p->fcode.bf_insns == NULL) {
2237 		snprintf(p->errbuf, sizeof(p->errbuf),
2238 			 "malloc: %s", pcap_strerror(errno));
2239 		return (-1);
2240 	}
2241 	memcpy(p->fcode.bf_insns, fp->bf_insns, prog_size);
2242 	return (0);
2243 }
2244 
2245 #ifdef BDEBUG
2246 static void
dot_dump_node(struct block * block,struct bpf_program * prog,FILE * out)2247 dot_dump_node(struct block *block, struct bpf_program *prog, FILE *out)
2248 {
2249 	int icount, noffset;
2250 	int i;
2251 
2252 	if (block == NULL || isMarked(block))
2253 		return;
2254 	Mark(block);
2255 
2256 	icount = slength(block->stmts) + 1 + block->longjt + block->longjf;
2257 	noffset = min(block->offset + icount, (int)prog->bf_len);
2258 
2259 	fprintf(out, "\tblock%d [shape=ellipse, id=\"block-%d\" label=\"BLOCK%d\\n", block->id, block->id, block->id);
2260 	for (i = block->offset; i < noffset; i++) {
2261 		fprintf(out, "\\n%s", bpf_image(prog->bf_insns + i, i));
2262 	}
2263 	fprintf(out, "\" tooltip=\"");
2264 	for (i = 0; i < BPF_MEMWORDS; i++)
2265 		if (block->val[i] != 0)
2266 			fprintf(out, "val[%d]=%d ", i, block->val[i]);
2267 	fprintf(out, "val[A]=%d ", block->val[A_ATOM]);
2268 	fprintf(out, "val[X]=%d", block->val[X_ATOM]);
2269 	fprintf(out, "\"");
2270 	if (JT(block) == NULL)
2271 		fprintf(out, ", peripheries=2");
2272 	fprintf(out, "];\n");
2273 
2274 	dot_dump_node(JT(block), prog, out);
2275 	dot_dump_node(JF(block), prog, out);
2276 }
2277 static void
dot_dump_edge(struct block * block,FILE * out)2278 dot_dump_edge(struct block *block, FILE *out)
2279 {
2280 	if (block == NULL || isMarked(block))
2281 		return;
2282 	Mark(block);
2283 
2284 	if (JT(block)) {
2285 		fprintf(out, "\t\"block%d\":se -> \"block%d\":n [label=\"T\"]; \n",
2286 				block->id, JT(block)->id);
2287 		fprintf(out, "\t\"block%d\":sw -> \"block%d\":n [label=\"F\"]; \n",
2288 			   block->id, JF(block)->id);
2289 	}
2290 	dot_dump_edge(JT(block), out);
2291 	dot_dump_edge(JF(block), out);
2292 }
2293 /* Output the block CFG using graphviz/DOT language
2294  * In the CFG, block's code, value index for each registers at EXIT,
2295  * and the jump relationship is show.
2296  *
2297  * example DOT for BPF `ip src host 1.1.1.1' is:
2298     digraph BPF {
2299     	block0 [shape=ellipse, id="block-0" label="BLOCK0\n\n(000) ldh      [12]\n(001) jeq      #0x800           jt 2	jf 5" tooltip="val[A]=0 val[X]=0"];
2300     	block1 [shape=ellipse, id="block-1" label="BLOCK1\n\n(002) ld       [26]\n(003) jeq      #0x1010101       jt 4	jf 5" tooltip="val[A]=0 val[X]=0"];
2301     	block2 [shape=ellipse, id="block-2" label="BLOCK2\n\n(004) ret      #68" tooltip="val[A]=0 val[X]=0", peripheries=2];
2302     	block3 [shape=ellipse, id="block-3" label="BLOCK3\n\n(005) ret      #0" tooltip="val[A]=0 val[X]=0", peripheries=2];
2303     	"block0":se -> "block1":n [label="T"];
2304     	"block0":sw -> "block3":n [label="F"];
2305     	"block1":se -> "block2":n [label="T"];
2306     	"block1":sw -> "block3":n [label="F"];
2307     }
2308  *
2309  *  After install graphviz on http://www.graphviz.org/, save it as bpf.dot
2310  *  and run `dot -Tpng -O bpf.dot' to draw the graph.
2311  */
2312 static void
dot_dump(struct block * root)2313 dot_dump(struct block *root)
2314 {
2315 	struct bpf_program f;
2316 	FILE *out = stdout;
2317 
2318 	memset(bids, 0, sizeof bids);
2319 	f.bf_insns = icode_to_fcode(root, &f.bf_len);
2320 
2321 	fprintf(out, "digraph BPF {\n");
2322 	unMarkAll();
2323 	dot_dump_node(root, &f, out);
2324 	unMarkAll();
2325 	dot_dump_edge(root, out);
2326 	fprintf(out, "}\n");
2327 
2328 	free((char *)f.bf_insns);
2329 }
2330 
2331 static void
plain_dump(struct block * root)2332 plain_dump(struct block *root)
2333 {
2334 	struct bpf_program f;
2335 
2336 	memset(bids, 0, sizeof bids);
2337 	f.bf_insns = icode_to_fcode(root, &f.bf_len);
2338 	bpf_dump(&f, 1);
2339 	putchar('\n');
2340 	free((char *)f.bf_insns);
2341 }
2342 static void
opt_dump(struct block * root)2343 opt_dump(struct block *root)
2344 {
2345 	/* if optimizer debugging is enabled, output DOT graph
2346 	 * `dflag=4' is equivalent to -dddd to follow -d/-dd/-ddd
2347      * convention in tcpdump command line
2348 	 */
2349 	if (dflag > 3)
2350 		dot_dump(root);
2351 	else
2352 		plain_dump(root);
2353 }
2354 
2355 #endif
2356