xref: /freebsd/contrib/libpcap/optimize.c (revision daf1cffce2e07931f27c6c6998652e90df6ba87e)
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[] =
25     "@(#) $Header: /tcpdump/master/libpcap/optimize.c,v 1.61 1999/10/19 15:18:30 itojun Exp $ (LBL)";
26 #endif
27 
28 #include <sys/types.h>
29 #include <sys/time.h>
30 
31 #include <stdio.h>
32 #include <stdlib.h>
33 #include <memory.h>
34 
35 #include "pcap-int.h"
36 
37 #include "gencode.h"
38 
39 #include "gnuc.h"
40 #ifdef HAVE_OS_PROTO_H
41 #include "os-proto.h"
42 #endif
43 
44 #ifdef BDEBUG
45 extern int dflag;
46 #endif
47 
48 #define A_ATOM BPF_MEMWORDS
49 #define X_ATOM (BPF_MEMWORDS+1)
50 
51 #define NOP -1
52 
53 /*
54  * This define is used to represent *both* the accumulator and
55  * x register in use-def computations.
56  * Currently, the use-def code assumes only one definition per instruction.
57  */
58 #define AX_ATOM N_ATOMS
59 
60 /*
61  * A flag to indicate that further optimization is needed.
62  * Iterative passes are continued until a given pass yields no
63  * branch movement.
64  */
65 static int done;
66 
67 /*
68  * A block is marked if only if its mark equals the current mark.
69  * Rather than traverse the code array, marking each item, 'cur_mark' is
70  * incremented.  This automatically makes each element unmarked.
71  */
72 static int cur_mark;
73 #define isMarked(p) ((p)->mark == cur_mark)
74 #define unMarkAll() cur_mark += 1
75 #define Mark(p) ((p)->mark = cur_mark)
76 
77 static void opt_init(struct block *);
78 static void opt_cleanup(void);
79 
80 static void make_marks(struct block *);
81 static void mark_code(struct block *);
82 
83 static void intern_blocks(struct block *);
84 
85 static int eq_slist(struct slist *, struct slist *);
86 
87 static void find_levels_r(struct block *);
88 
89 static void find_levels(struct block *);
90 static void find_dom(struct block *);
91 static void propedom(struct edge *);
92 static void find_edom(struct block *);
93 static void find_closure(struct block *);
94 static int atomuse(struct stmt *);
95 static int atomdef(struct stmt *);
96 static void compute_local_ud(struct block *);
97 static void find_ud(struct block *);
98 static void init_val(void);
99 static int F(int, int, int);
100 static inline void vstore(struct stmt *, int *, int, int);
101 static void opt_blk(struct block *, int);
102 static int use_conflict(struct block *, struct block *);
103 static void opt_j(struct edge *);
104 static void or_pullup(struct block *);
105 static void and_pullup(struct block *);
106 static void opt_blks(struct block *, int);
107 static inline void link_inedge(struct edge *, struct block *);
108 static void find_inedges(struct block *);
109 static void opt_root(struct block **);
110 static void opt_loop(struct block *, int);
111 static void fold_op(struct stmt *, int, int);
112 static inline struct slist *this_op(struct slist *);
113 static void opt_not(struct block *);
114 static void opt_peep(struct block *);
115 static void opt_stmt(struct stmt *, int[], int);
116 static void deadstmt(struct stmt *, struct stmt *[]);
117 static void opt_deadstores(struct block *);
118 static void opt_blk(struct block *, int);
119 static int use_conflict(struct block *, struct block *);
120 static void opt_j(struct edge *);
121 static struct block *fold_edge(struct block *, struct edge *);
122 static inline int eq_blk(struct block *, struct block *);
123 static int slength(struct slist *);
124 static int count_blocks(struct block *);
125 static void number_blks_r(struct block *);
126 static int count_stmts(struct block *);
127 static int convert_code_r(struct block *);
128 #ifdef BDEBUG
129 static void opt_dump(struct block *);
130 #endif
131 
132 static int n_blocks;
133 struct block **blocks;
134 static int n_edges;
135 struct edge **edges;
136 
137 /*
138  * A bit vector set representation of the dominators.
139  * We round up the set size to the next power of two.
140  */
141 static int nodewords;
142 static int edgewords;
143 struct block **levels;
144 bpf_u_int32 *space;
145 #define BITS_PER_WORD (8*sizeof(bpf_u_int32))
146 /*
147  * True if a is in uset {p}
148  */
149 #define SET_MEMBER(p, a) \
150 ((p)[(unsigned)(a) / BITS_PER_WORD] & (1 << ((unsigned)(a) % BITS_PER_WORD)))
151 
152 /*
153  * Add 'a' to uset p.
154  */
155 #define SET_INSERT(p, a) \
156 (p)[(unsigned)(a) / BITS_PER_WORD] |= (1 << ((unsigned)(a) % BITS_PER_WORD))
157 
158 /*
159  * Delete 'a' from uset p.
160  */
161 #define SET_DELETE(p, a) \
162 (p)[(unsigned)(a) / BITS_PER_WORD] &= ~(1 << ((unsigned)(a) % BITS_PER_WORD))
163 
164 /*
165  * a := a intersect b
166  */
167 #define SET_INTERSECT(a, b, n)\
168 {\
169 	register bpf_u_int32 *_x = a, *_y = b;\
170 	register int _n = n;\
171 	while (--_n >= 0) *_x++ &= *_y++;\
172 }
173 
174 /*
175  * a := a - b
176  */
177 #define SET_SUBTRACT(a, b, n)\
178 {\
179 	register bpf_u_int32 *_x = a, *_y = b;\
180 	register int _n = n;\
181 	while (--_n >= 0) *_x++ &=~ *_y++;\
182 }
183 
184 /*
185  * a := a union b
186  */
187 #define SET_UNION(a, b, n)\
188 {\
189 	register bpf_u_int32 *_x = a, *_y = b;\
190 	register int _n = n;\
191 	while (--_n >= 0) *_x++ |= *_y++;\
192 }
193 
194 static uset all_dom_sets;
195 static uset all_closure_sets;
196 static uset all_edge_sets;
197 
198 #ifndef MAX
199 #define MAX(a,b) ((a)>(b)?(a):(b))
200 #endif
201 
202 static void
203 find_levels_r(b)
204 	struct block *b;
205 {
206 	int level;
207 
208 	if (isMarked(b))
209 		return;
210 
211 	Mark(b);
212 	b->link = 0;
213 
214 	if (JT(b)) {
215 		find_levels_r(JT(b));
216 		find_levels_r(JF(b));
217 		level = MAX(JT(b)->level, JF(b)->level) + 1;
218 	} else
219 		level = 0;
220 	b->level = level;
221 	b->link = levels[level];
222 	levels[level] = b;
223 }
224 
225 /*
226  * Level graph.  The levels go from 0 at the leaves to
227  * N_LEVELS at the root.  The levels[] array points to the
228  * first node of the level list, whose elements are linked
229  * with the 'link' field of the struct block.
230  */
231 static void
232 find_levels(root)
233 	struct block *root;
234 {
235 	memset((char *)levels, 0, n_blocks * sizeof(*levels));
236 	unMarkAll();
237 	find_levels_r(root);
238 }
239 
240 /*
241  * Find dominator relationships.
242  * Assumes graph has been leveled.
243  */
244 static void
245 find_dom(root)
246 	struct block *root;
247 {
248 	int i;
249 	struct block *b;
250 	bpf_u_int32 *x;
251 
252 	/*
253 	 * Initialize sets to contain all nodes.
254 	 */
255 	x = all_dom_sets;
256 	i = n_blocks * nodewords;
257 	while (--i >= 0)
258 		*x++ = ~0;
259 	/* Root starts off empty. */
260 	for (i = nodewords; --i >= 0;)
261 		root->dom[i] = 0;
262 
263 	/* root->level is the highest level no found. */
264 	for (i = root->level; i >= 0; --i) {
265 		for (b = levels[i]; b; b = b->link) {
266 			SET_INSERT(b->dom, b->id);
267 			if (JT(b) == 0)
268 				continue;
269 			SET_INTERSECT(JT(b)->dom, b->dom, nodewords);
270 			SET_INTERSECT(JF(b)->dom, b->dom, nodewords);
271 		}
272 	}
273 }
274 
275 static void
276 propedom(ep)
277 	struct edge *ep;
278 {
279 	SET_INSERT(ep->edom, ep->id);
280 	if (ep->succ) {
281 		SET_INTERSECT(ep->succ->et.edom, ep->edom, edgewords);
282 		SET_INTERSECT(ep->succ->ef.edom, ep->edom, edgewords);
283 	}
284 }
285 
286 /*
287  * Compute edge dominators.
288  * Assumes graph has been leveled and predecessors established.
289  */
290 static void
291 find_edom(root)
292 	struct block *root;
293 {
294 	int i;
295 	uset x;
296 	struct block *b;
297 
298 	x = all_edge_sets;
299 	for (i = n_edges * edgewords; --i >= 0; )
300 		x[i] = ~0;
301 
302 	/* root->level is the highest level no found. */
303 	memset(root->et.edom, 0, edgewords * sizeof(*(uset)0));
304 	memset(root->ef.edom, 0, edgewords * sizeof(*(uset)0));
305 	for (i = root->level; i >= 0; --i) {
306 		for (b = levels[i]; b != 0; b = b->link) {
307 			propedom(&b->et);
308 			propedom(&b->ef);
309 		}
310 	}
311 }
312 
313 /*
314  * Find the backwards transitive closure of the flow graph.  These sets
315  * are backwards in the sense that we find the set of nodes that reach
316  * a given node, not the set of nodes that can be reached by a node.
317  *
318  * Assumes graph has been leveled.
319  */
320 static void
321 find_closure(root)
322 	struct block *root;
323 {
324 	int i;
325 	struct block *b;
326 
327 	/*
328 	 * Initialize sets to contain no nodes.
329 	 */
330 	memset((char *)all_closure_sets, 0,
331 	      n_blocks * nodewords * sizeof(*all_closure_sets));
332 
333 	/* root->level is the highest level no found. */
334 	for (i = root->level; i >= 0; --i) {
335 		for (b = levels[i]; b; b = b->link) {
336 			SET_INSERT(b->closure, b->id);
337 			if (JT(b) == 0)
338 				continue;
339 			SET_UNION(JT(b)->closure, b->closure, nodewords);
340 			SET_UNION(JF(b)->closure, b->closure, nodewords);
341 		}
342 	}
343 }
344 
345 /*
346  * Return the register number that is used by s.  If A and X are both
347  * used, return AX_ATOM.  If no register is used, return -1.
348  *
349  * The implementation should probably change to an array access.
350  */
351 static int
352 atomuse(s)
353 	struct stmt *s;
354 {
355 	register int c = s->code;
356 
357 	if (c == NOP)
358 		return -1;
359 
360 	switch (BPF_CLASS(c)) {
361 
362 	case BPF_RET:
363 		return (BPF_RVAL(c) == BPF_A) ? A_ATOM :
364 			(BPF_RVAL(c) == BPF_X) ? X_ATOM : -1;
365 
366 	case BPF_LD:
367 	case BPF_LDX:
368 		return (BPF_MODE(c) == BPF_IND) ? X_ATOM :
369 			(BPF_MODE(c) == BPF_MEM) ? s->k : -1;
370 
371 	case BPF_ST:
372 		return A_ATOM;
373 
374 	case BPF_STX:
375 		return X_ATOM;
376 
377 	case BPF_JMP:
378 	case BPF_ALU:
379 		if (BPF_SRC(c) == BPF_X)
380 			return AX_ATOM;
381 		return A_ATOM;
382 
383 	case BPF_MISC:
384 		return BPF_MISCOP(c) == BPF_TXA ? X_ATOM : A_ATOM;
385 	}
386 	abort();
387 	/* NOTREACHED */
388 }
389 
390 /*
391  * Return the register number that is defined by 's'.  We assume that
392  * a single stmt cannot define more than one register.  If no register
393  * is defined, return -1.
394  *
395  * The implementation should probably change to an array access.
396  */
397 static int
398 atomdef(s)
399 	struct stmt *s;
400 {
401 	if (s->code == NOP)
402 		return -1;
403 
404 	switch (BPF_CLASS(s->code)) {
405 
406 	case BPF_LD:
407 	case BPF_ALU:
408 		return A_ATOM;
409 
410 	case BPF_LDX:
411 		return X_ATOM;
412 
413 	case BPF_ST:
414 	case BPF_STX:
415 		return s->k;
416 
417 	case BPF_MISC:
418 		return BPF_MISCOP(s->code) == BPF_TAX ? X_ATOM : A_ATOM;
419 	}
420 	return -1;
421 }
422 
423 static void
424 compute_local_ud(b)
425 	struct block *b;
426 {
427 	struct slist *s;
428 	atomset def = 0, use = 0, kill = 0;
429 	int atom;
430 
431 	for (s = b->stmts; s; s = s->next) {
432 		if (s->s.code == NOP)
433 			continue;
434 		atom = atomuse(&s->s);
435 		if (atom >= 0) {
436 			if (atom == AX_ATOM) {
437 				if (!ATOMELEM(def, X_ATOM))
438 					use |= ATOMMASK(X_ATOM);
439 				if (!ATOMELEM(def, A_ATOM))
440 					use |= ATOMMASK(A_ATOM);
441 			}
442 			else if (atom < N_ATOMS) {
443 				if (!ATOMELEM(def, atom))
444 					use |= ATOMMASK(atom);
445 			}
446 			else
447 				abort();
448 		}
449 		atom = atomdef(&s->s);
450 		if (atom >= 0) {
451 			if (!ATOMELEM(use, atom))
452 				kill |= ATOMMASK(atom);
453 			def |= ATOMMASK(atom);
454 		}
455 	}
456 	if (!ATOMELEM(def, A_ATOM) && BPF_CLASS(b->s.code) == BPF_JMP)
457 		use |= ATOMMASK(A_ATOM);
458 
459 	b->def = def;
460 	b->kill = kill;
461 	b->in_use = use;
462 }
463 
464 /*
465  * Assume graph is already leveled.
466  */
467 static void
468 find_ud(root)
469 	struct block *root;
470 {
471 	int i, maxlevel;
472 	struct block *p;
473 
474 	/*
475 	 * root->level is the highest level no found;
476 	 * count down from there.
477 	 */
478 	maxlevel = root->level;
479 	for (i = maxlevel; i >= 0; --i)
480 		for (p = levels[i]; p; p = p->link) {
481 			compute_local_ud(p);
482 			p->out_use = 0;
483 		}
484 
485 	for (i = 1; i <= maxlevel; ++i) {
486 		for (p = levels[i]; p; p = p->link) {
487 			p->out_use |= JT(p)->in_use | JF(p)->in_use;
488 			p->in_use |= p->out_use &~ p->kill;
489 		}
490 	}
491 }
492 
493 /*
494  * These data structures are used in a Cocke and Shwarz style
495  * value numbering scheme.  Since the flowgraph is acyclic,
496  * exit values can be propagated from a node's predecessors
497  * provided it is uniquely defined.
498  */
499 struct valnode {
500 	int code;
501 	int v0, v1;
502 	int val;
503 	struct valnode *next;
504 };
505 
506 #define MODULUS 213
507 static struct valnode *hashtbl[MODULUS];
508 static int curval;
509 static int maxval;
510 
511 /* Integer constants mapped with the load immediate opcode. */
512 #define K(i) F(BPF_LD|BPF_IMM|BPF_W, i, 0L)
513 
514 struct vmapinfo {
515 	int is_const;
516 	bpf_int32 const_val;
517 };
518 
519 struct vmapinfo *vmap;
520 struct valnode *vnode_base;
521 struct valnode *next_vnode;
522 
523 static void
524 init_val()
525 {
526 	curval = 0;
527 	next_vnode = vnode_base;
528 	memset((char *)vmap, 0, maxval * sizeof(*vmap));
529 	memset((char *)hashtbl, 0, sizeof hashtbl);
530 }
531 
532 /* Because we really don't have an IR, this stuff is a little messy. */
533 static int
534 F(code, v0, v1)
535 	int code;
536 	int v0, v1;
537 {
538 	u_int hash;
539 	int val;
540 	struct valnode *p;
541 
542 	hash = (u_int)code ^ (v0 << 4) ^ (v1 << 8);
543 	hash %= MODULUS;
544 
545 	for (p = hashtbl[hash]; p; p = p->next)
546 		if (p->code == code && p->v0 == v0 && p->v1 == v1)
547 			return p->val;
548 
549 	val = ++curval;
550 	if (BPF_MODE(code) == BPF_IMM &&
551 	    (BPF_CLASS(code) == BPF_LD || BPF_CLASS(code) == BPF_LDX)) {
552 		vmap[val].const_val = v0;
553 		vmap[val].is_const = 1;
554 	}
555 	p = next_vnode++;
556 	p->val = val;
557 	p->code = code;
558 	p->v0 = v0;
559 	p->v1 = v1;
560 	p->next = hashtbl[hash];
561 	hashtbl[hash] = p;
562 
563 	return val;
564 }
565 
566 static inline void
567 vstore(s, valp, newval, alter)
568 	struct stmt *s;
569 	int *valp;
570 	int newval;
571 	int alter;
572 {
573 	if (alter && *valp == newval)
574 		s->code = NOP;
575 	else
576 		*valp = newval;
577 }
578 
579 static void
580 fold_op(s, v0, v1)
581 	struct stmt *s;
582 	int v0, v1;
583 {
584 	bpf_int32 a, b;
585 
586 	a = vmap[v0].const_val;
587 	b = vmap[v1].const_val;
588 
589 	switch (BPF_OP(s->code)) {
590 	case BPF_ADD:
591 		a += b;
592 		break;
593 
594 	case BPF_SUB:
595 		a -= b;
596 		break;
597 
598 	case BPF_MUL:
599 		a *= b;
600 		break;
601 
602 	case BPF_DIV:
603 		if (b == 0)
604 			bpf_error("division by zero");
605 		a /= b;
606 		break;
607 
608 	case BPF_AND:
609 		a &= b;
610 		break;
611 
612 	case BPF_OR:
613 		a |= b;
614 		break;
615 
616 	case BPF_LSH:
617 		a <<= b;
618 		break;
619 
620 	case BPF_RSH:
621 		a >>= b;
622 		break;
623 
624 	case BPF_NEG:
625 		a = -a;
626 		break;
627 
628 	default:
629 		abort();
630 	}
631 	s->k = a;
632 	s->code = BPF_LD|BPF_IMM;
633 	done = 0;
634 }
635 
636 static inline struct slist *
637 this_op(s)
638 	struct slist *s;
639 {
640 	while (s != 0 && s->s.code == NOP)
641 		s = s->next;
642 	return s;
643 }
644 
645 static void
646 opt_not(b)
647 	struct block *b;
648 {
649 	struct block *tmp = JT(b);
650 
651 	JT(b) = JF(b);
652 	JF(b) = tmp;
653 }
654 
655 static void
656 opt_peep(b)
657 	struct block *b;
658 {
659 	struct slist *s;
660 	struct slist *next, *last;
661 	int val;
662 
663 	s = b->stmts;
664 	if (s == 0)
665 		return;
666 
667 	last = s;
668 	while (1) {
669 		s = this_op(s);
670 		if (s == 0)
671 			break;
672 		next = this_op(s->next);
673 		if (next == 0)
674 			break;
675 		last = next;
676 
677 		/*
678 		 * st  M[k]	-->	st  M[k]
679 		 * ldx M[k]		tax
680 		 */
681 		if (s->s.code == BPF_ST &&
682 		    next->s.code == (BPF_LDX|BPF_MEM) &&
683 		    s->s.k == next->s.k) {
684 			done = 0;
685 			next->s.code = BPF_MISC|BPF_TAX;
686 		}
687 		/*
688 		 * ld  #k	-->	ldx  #k
689 		 * tax			txa
690 		 */
691 		if (s->s.code == (BPF_LD|BPF_IMM) &&
692 		    next->s.code == (BPF_MISC|BPF_TAX)) {
693 			s->s.code = BPF_LDX|BPF_IMM;
694 			next->s.code = BPF_MISC|BPF_TXA;
695 			done = 0;
696 		}
697 		/*
698 		 * This is an ugly special case, but it happens
699 		 * when you say tcp[k] or udp[k] where k is a constant.
700 		 */
701 		if (s->s.code == (BPF_LD|BPF_IMM)) {
702 			struct slist *add, *tax, *ild;
703 
704 			/*
705 			 * Check that X isn't used on exit from this
706 			 * block (which the optimizer might cause).
707 			 * We know the code generator won't generate
708 			 * any local dependencies.
709 			 */
710 			if (ATOMELEM(b->out_use, X_ATOM))
711 				break;
712 
713 			if (next->s.code != (BPF_LDX|BPF_MSH|BPF_B))
714 				add = next;
715 			else
716 				add = this_op(next->next);
717 			if (add == 0 || add->s.code != (BPF_ALU|BPF_ADD|BPF_X))
718 				break;
719 
720 			tax = this_op(add->next);
721 			if (tax == 0 || tax->s.code != (BPF_MISC|BPF_TAX))
722 				break;
723 
724 			ild = this_op(tax->next);
725 			if (ild == 0 || BPF_CLASS(ild->s.code) != BPF_LD ||
726 			    BPF_MODE(ild->s.code) != BPF_IND)
727 				break;
728 			/*
729 			 * XXX We need to check that X is not
730 			 * subsequently used.  We know we can eliminate the
731 			 * accumulator modifications since it is defined
732 			 * by the last stmt of this sequence.
733 			 *
734 			 * We want to turn this sequence:
735 			 *
736 			 * (004) ldi     #0x2		{s}
737 			 * (005) ldxms   [14]		{next}  -- optional
738 			 * (006) addx			{add}
739 			 * (007) tax			{tax}
740 			 * (008) ild     [x+0]		{ild}
741 			 *
742 			 * into this sequence:
743 			 *
744 			 * (004) nop
745 			 * (005) ldxms   [14]
746 			 * (006) nop
747 			 * (007) nop
748 			 * (008) ild     [x+2]
749 			 *
750 			 */
751 			ild->s.k += s->s.k;
752 			s->s.code = NOP;
753 			add->s.code = NOP;
754 			tax->s.code = NOP;
755 			done = 0;
756 		}
757 		s = next;
758 	}
759 	/*
760 	 * If we have a subtract to do a comparison, and the X register
761 	 * is a known constant, we can merge this value into the
762 	 * comparison.
763 	 */
764 	if (last->s.code == (BPF_ALU|BPF_SUB|BPF_X) &&
765 	    !ATOMELEM(b->out_use, A_ATOM)) {
766 		val = b->val[X_ATOM];
767 		if (vmap[val].is_const) {
768 			int op;
769 
770 			b->s.k += vmap[val].const_val;
771 			op = BPF_OP(b->s.code);
772 			if (op == BPF_JGT || op == BPF_JGE) {
773 				struct block *t = JT(b);
774 				JT(b) = JF(b);
775 				JF(b) = t;
776 				b->s.k += 0x80000000;
777 			}
778 			last->s.code = NOP;
779 			done = 0;
780 		} else if (b->s.k == 0) {
781 			/*
782 			 * sub x  ->	nop
783 			 * j  #0	j  x
784 			 */
785 			last->s.code = NOP;
786 			b->s.code = BPF_CLASS(b->s.code) | BPF_OP(b->s.code) |
787 				BPF_X;
788 			done = 0;
789 		}
790 	}
791 	/*
792 	 * Likewise, a constant subtract can be simplified.
793 	 */
794 	else if (last->s.code == (BPF_ALU|BPF_SUB|BPF_K) &&
795 		 !ATOMELEM(b->out_use, A_ATOM)) {
796 		int op;
797 
798 		b->s.k += last->s.k;
799 		last->s.code = NOP;
800 		op = BPF_OP(b->s.code);
801 		if (op == BPF_JGT || op == BPF_JGE) {
802 			struct block *t = JT(b);
803 			JT(b) = JF(b);
804 			JF(b) = t;
805 			b->s.k += 0x80000000;
806 		}
807 		done = 0;
808 	}
809 	/*
810 	 * and #k	nop
811 	 * jeq #0  ->	jset #k
812 	 */
813 	if (last->s.code == (BPF_ALU|BPF_AND|BPF_K) &&
814 	    !ATOMELEM(b->out_use, A_ATOM) && b->s.k == 0) {
815 		b->s.k = last->s.k;
816 		b->s.code = BPF_JMP|BPF_K|BPF_JSET;
817 		last->s.code = NOP;
818 		done = 0;
819 		opt_not(b);
820 	}
821 	/*
822 	 * If the accumulator is a known constant, we can compute the
823 	 * comparison result.
824 	 */
825 	val = b->val[A_ATOM];
826 	if (vmap[val].is_const && BPF_SRC(b->s.code) == BPF_K) {
827 		bpf_int32 v = vmap[val].const_val;
828 		switch (BPF_OP(b->s.code)) {
829 
830 		case BPF_JEQ:
831 			v = v == b->s.k;
832 			break;
833 
834 		case BPF_JGT:
835 			v = (unsigned)v > b->s.k;
836 			break;
837 
838 		case BPF_JGE:
839 			v = (unsigned)v >= b->s.k;
840 			break;
841 
842 		case BPF_JSET:
843 			v &= b->s.k;
844 			break;
845 
846 		default:
847 			abort();
848 		}
849 		if (JF(b) != JT(b))
850 			done = 0;
851 		if (v)
852 			JF(b) = JT(b);
853 		else
854 			JT(b) = JF(b);
855 	}
856 }
857 
858 /*
859  * Compute the symbolic value of expression of 's', and update
860  * anything it defines in the value table 'val'.  If 'alter' is true,
861  * do various optimizations.  This code would be cleaner if symbolic
862  * evaluation and code transformations weren't folded together.
863  */
864 static void
865 opt_stmt(s, val, alter)
866 	struct stmt *s;
867 	int val[];
868 	int alter;
869 {
870 	int op;
871 	int v;
872 
873 	switch (s->code) {
874 
875 	case BPF_LD|BPF_ABS|BPF_W:
876 	case BPF_LD|BPF_ABS|BPF_H:
877 	case BPF_LD|BPF_ABS|BPF_B:
878 		v = F(s->code, s->k, 0L);
879 		vstore(s, &val[A_ATOM], v, alter);
880 		break;
881 
882 	case BPF_LD|BPF_IND|BPF_W:
883 	case BPF_LD|BPF_IND|BPF_H:
884 	case BPF_LD|BPF_IND|BPF_B:
885 		v = val[X_ATOM];
886 		if (alter && vmap[v].is_const) {
887 			s->code = BPF_LD|BPF_ABS|BPF_SIZE(s->code);
888 			s->k += vmap[v].const_val;
889 			v = F(s->code, s->k, 0L);
890 			done = 0;
891 		}
892 		else
893 			v = F(s->code, s->k, v);
894 		vstore(s, &val[A_ATOM], v, alter);
895 		break;
896 
897 	case BPF_LD|BPF_LEN:
898 		v = F(s->code, 0L, 0L);
899 		vstore(s, &val[A_ATOM], v, alter);
900 		break;
901 
902 	case BPF_LD|BPF_IMM:
903 		v = K(s->k);
904 		vstore(s, &val[A_ATOM], v, alter);
905 		break;
906 
907 	case BPF_LDX|BPF_IMM:
908 		v = K(s->k);
909 		vstore(s, &val[X_ATOM], v, alter);
910 		break;
911 
912 	case BPF_LDX|BPF_MSH|BPF_B:
913 		v = F(s->code, s->k, 0L);
914 		vstore(s, &val[X_ATOM], v, alter);
915 		break;
916 
917 	case BPF_ALU|BPF_NEG:
918 		if (alter && vmap[val[A_ATOM]].is_const) {
919 			s->code = BPF_LD|BPF_IMM;
920 			s->k = -vmap[val[A_ATOM]].const_val;
921 			val[A_ATOM] = K(s->k);
922 		}
923 		else
924 			val[A_ATOM] = F(s->code, val[A_ATOM], 0L);
925 		break;
926 
927 	case BPF_ALU|BPF_ADD|BPF_K:
928 	case BPF_ALU|BPF_SUB|BPF_K:
929 	case BPF_ALU|BPF_MUL|BPF_K:
930 	case BPF_ALU|BPF_DIV|BPF_K:
931 	case BPF_ALU|BPF_AND|BPF_K:
932 	case BPF_ALU|BPF_OR|BPF_K:
933 	case BPF_ALU|BPF_LSH|BPF_K:
934 	case BPF_ALU|BPF_RSH|BPF_K:
935 		op = BPF_OP(s->code);
936 		if (alter) {
937 			if (s->k == 0) {
938 				if (op == BPF_ADD || op == BPF_SUB ||
939 				    op == BPF_LSH || op == BPF_RSH ||
940 				    op == BPF_OR) {
941 					s->code = NOP;
942 					break;
943 				}
944 				if (op == BPF_MUL || op == BPF_AND) {
945 					s->code = BPF_LD|BPF_IMM;
946 					val[A_ATOM] = K(s->k);
947 					break;
948 				}
949 			}
950 			if (vmap[val[A_ATOM]].is_const) {
951 				fold_op(s, val[A_ATOM], K(s->k));
952 				val[A_ATOM] = K(s->k);
953 				break;
954 			}
955 		}
956 		val[A_ATOM] = F(s->code, val[A_ATOM], K(s->k));
957 		break;
958 
959 	case BPF_ALU|BPF_ADD|BPF_X:
960 	case BPF_ALU|BPF_SUB|BPF_X:
961 	case BPF_ALU|BPF_MUL|BPF_X:
962 	case BPF_ALU|BPF_DIV|BPF_X:
963 	case BPF_ALU|BPF_AND|BPF_X:
964 	case BPF_ALU|BPF_OR|BPF_X:
965 	case BPF_ALU|BPF_LSH|BPF_X:
966 	case BPF_ALU|BPF_RSH|BPF_X:
967 		op = BPF_OP(s->code);
968 		if (alter && vmap[val[X_ATOM]].is_const) {
969 			if (vmap[val[A_ATOM]].is_const) {
970 				fold_op(s, val[A_ATOM], val[X_ATOM]);
971 				val[A_ATOM] = K(s->k);
972 			}
973 			else {
974 				s->code = BPF_ALU|BPF_K|op;
975 				s->k = vmap[val[X_ATOM]].const_val;
976 				done = 0;
977 				val[A_ATOM] =
978 					F(s->code, val[A_ATOM], K(s->k));
979 			}
980 			break;
981 		}
982 		/*
983 		 * Check if we're doing something to an accumulator
984 		 * that is 0, and simplify.  This may not seem like
985 		 * much of a simplification but it could open up further
986 		 * optimizations.
987 		 * XXX We could also check for mul by 1, and -1, etc.
988 		 */
989 		if (alter && vmap[val[A_ATOM]].is_const
990 		    && vmap[val[A_ATOM]].const_val == 0) {
991 			if (op == BPF_ADD || op == BPF_OR ||
992 			    op == BPF_LSH || op == BPF_RSH || op == BPF_SUB) {
993 				s->code = BPF_MISC|BPF_TXA;
994 				vstore(s, &val[A_ATOM], val[X_ATOM], alter);
995 				break;
996 			}
997 			else if (op == BPF_MUL || op == BPF_DIV ||
998 				 op == BPF_AND) {
999 				s->code = BPF_LD|BPF_IMM;
1000 				s->k = 0;
1001 				vstore(s, &val[A_ATOM], K(s->k), alter);
1002 				break;
1003 			}
1004 			else if (op == BPF_NEG) {
1005 				s->code = NOP;
1006 				break;
1007 			}
1008 		}
1009 		val[A_ATOM] = F(s->code, val[A_ATOM], val[X_ATOM]);
1010 		break;
1011 
1012 	case BPF_MISC|BPF_TXA:
1013 		vstore(s, &val[A_ATOM], val[X_ATOM], alter);
1014 		break;
1015 
1016 	case BPF_LD|BPF_MEM:
1017 		v = val[s->k];
1018 		if (alter && vmap[v].is_const) {
1019 			s->code = BPF_LD|BPF_IMM;
1020 			s->k = vmap[v].const_val;
1021 			done = 0;
1022 		}
1023 		vstore(s, &val[A_ATOM], v, alter);
1024 		break;
1025 
1026 	case BPF_MISC|BPF_TAX:
1027 		vstore(s, &val[X_ATOM], val[A_ATOM], alter);
1028 		break;
1029 
1030 	case BPF_LDX|BPF_MEM:
1031 		v = val[s->k];
1032 		if (alter && vmap[v].is_const) {
1033 			s->code = BPF_LDX|BPF_IMM;
1034 			s->k = vmap[v].const_val;
1035 			done = 0;
1036 		}
1037 		vstore(s, &val[X_ATOM], v, alter);
1038 		break;
1039 
1040 	case BPF_ST:
1041 		vstore(s, &val[s->k], val[A_ATOM], alter);
1042 		break;
1043 
1044 	case BPF_STX:
1045 		vstore(s, &val[s->k], val[X_ATOM], alter);
1046 		break;
1047 	}
1048 }
1049 
1050 static void
1051 deadstmt(s, last)
1052 	register struct stmt *s;
1053 	register struct stmt *last[];
1054 {
1055 	register int atom;
1056 
1057 	atom = atomuse(s);
1058 	if (atom >= 0) {
1059 		if (atom == AX_ATOM) {
1060 			last[X_ATOM] = 0;
1061 			last[A_ATOM] = 0;
1062 		}
1063 		else
1064 			last[atom] = 0;
1065 	}
1066 	atom = atomdef(s);
1067 	if (atom >= 0) {
1068 		if (last[atom]) {
1069 			done = 0;
1070 			last[atom]->code = NOP;
1071 		}
1072 		last[atom] = s;
1073 	}
1074 }
1075 
1076 static void
1077 opt_deadstores(b)
1078 	register struct block *b;
1079 {
1080 	register struct slist *s;
1081 	register int atom;
1082 	struct stmt *last[N_ATOMS];
1083 
1084 	memset((char *)last, 0, sizeof last);
1085 
1086 	for (s = b->stmts; s != 0; s = s->next)
1087 		deadstmt(&s->s, last);
1088 	deadstmt(&b->s, last);
1089 
1090 	for (atom = 0; atom < N_ATOMS; ++atom)
1091 		if (last[atom] && !ATOMELEM(b->out_use, atom)) {
1092 			last[atom]->code = NOP;
1093 			done = 0;
1094 		}
1095 }
1096 
1097 static void
1098 opt_blk(b, do_stmts)
1099 	struct block *b;
1100 	int do_stmts;
1101 {
1102 	struct slist *s;
1103 	struct edge *p;
1104 	int i;
1105 	bpf_int32 aval;
1106 
1107 #if 0
1108 	for (s = b->stmts; s && s->next; s = s->next)
1109 		if (BPF_CLASS(s->s.code) == BPF_JMP) {
1110 			do_stmts = 0;
1111 			break;
1112 		}
1113 #endif
1114 
1115 	/*
1116 	 * Initialize the atom values.
1117 	 * If we have no predecessors, everything is undefined.
1118 	 * Otherwise, we inherent our values from our predecessors.
1119 	 * If any register has an ambiguous value (i.e. control paths are
1120 	 * merging) give it the undefined value of 0.
1121 	 */
1122 	p = b->in_edges;
1123 	if (p == 0)
1124 		memset((char *)b->val, 0, sizeof(b->val));
1125 	else {
1126 		memcpy((char *)b->val, (char *)p->pred->val, sizeof(b->val));
1127 		while ((p = p->next) != NULL) {
1128 			for (i = 0; i < N_ATOMS; ++i)
1129 				if (b->val[i] != p->pred->val[i])
1130 					b->val[i] = 0;
1131 		}
1132 	}
1133 	aval = b->val[A_ATOM];
1134 	for (s = b->stmts; s; s = s->next)
1135 		opt_stmt(&s->s, b->val, do_stmts);
1136 
1137 	/*
1138 	 * This is a special case: if we don't use anything from this
1139 	 * block, and we load the accumulator with value that is
1140 	 * already there, or if this block is a return,
1141 	 * eliminate all the statements.
1142 	 */
1143 	if (do_stmts &&
1144 	    ((b->out_use == 0 && aval != 0 &&b->val[A_ATOM] == aval) ||
1145 	     BPF_CLASS(b->s.code) == BPF_RET)) {
1146 		if (b->stmts != 0) {
1147 			b->stmts = 0;
1148 			done = 0;
1149 		}
1150 	} else {
1151 		opt_peep(b);
1152 		opt_deadstores(b);
1153 	}
1154 	/*
1155 	 * Set up values for branch optimizer.
1156 	 */
1157 	if (BPF_SRC(b->s.code) == BPF_K)
1158 		b->oval = K(b->s.k);
1159 	else
1160 		b->oval = b->val[X_ATOM];
1161 	b->et.code = b->s.code;
1162 	b->ef.code = -b->s.code;
1163 }
1164 
1165 /*
1166  * Return true if any register that is used on exit from 'succ', has
1167  * an exit value that is different from the corresponding exit value
1168  * from 'b'.
1169  */
1170 static int
1171 use_conflict(b, succ)
1172 	struct block *b, *succ;
1173 {
1174 	int atom;
1175 	atomset use = succ->out_use;
1176 
1177 	if (use == 0)
1178 		return 0;
1179 
1180 	for (atom = 0; atom < N_ATOMS; ++atom)
1181 		if (ATOMELEM(use, atom))
1182 			if (b->val[atom] != succ->val[atom])
1183 				return 1;
1184 	return 0;
1185 }
1186 
1187 static struct block *
1188 fold_edge(child, ep)
1189 	struct block *child;
1190 	struct edge *ep;
1191 {
1192 	int sense;
1193 	int aval0, aval1, oval0, oval1;
1194 	int code = ep->code;
1195 
1196 	if (code < 0) {
1197 		code = -code;
1198 		sense = 0;
1199 	} else
1200 		sense = 1;
1201 
1202 	if (child->s.code != code)
1203 		return 0;
1204 
1205 	aval0 = child->val[A_ATOM];
1206 	oval0 = child->oval;
1207 	aval1 = ep->pred->val[A_ATOM];
1208 	oval1 = ep->pred->oval;
1209 
1210 	if (aval0 != aval1)
1211 		return 0;
1212 
1213 	if (oval0 == oval1)
1214 		/*
1215 		 * The operands are identical, so the
1216 		 * result is true if a true branch was
1217 		 * taken to get here, otherwise false.
1218 		 */
1219 		return sense ? JT(child) : JF(child);
1220 
1221 	if (sense && code == (BPF_JMP|BPF_JEQ|BPF_K))
1222 		/*
1223 		 * At this point, we only know the comparison if we
1224 		 * came down the true branch, and it was an equality
1225 		 * comparison with a constant.  We rely on the fact that
1226 		 * distinct constants have distinct value numbers.
1227 		 */
1228 		return JF(child);
1229 
1230 	return 0;
1231 }
1232 
1233 static void
1234 opt_j(ep)
1235 	struct edge *ep;
1236 {
1237 	register int i, k;
1238 	register struct block *target;
1239 
1240 	if (JT(ep->succ) == 0)
1241 		return;
1242 
1243 	if (JT(ep->succ) == JF(ep->succ)) {
1244 		/*
1245 		 * Common branch targets can be eliminated, provided
1246 		 * there is no data dependency.
1247 		 */
1248 		if (!use_conflict(ep->pred, ep->succ->et.succ)) {
1249 			done = 0;
1250 			ep->succ = JT(ep->succ);
1251 		}
1252 	}
1253 	/*
1254 	 * For each edge dominator that matches the successor of this
1255 	 * edge, promote the edge successor to the its grandchild.
1256 	 *
1257 	 * XXX We violate the set abstraction here in favor a reasonably
1258 	 * efficient loop.
1259 	 */
1260  top:
1261 	for (i = 0; i < edgewords; ++i) {
1262 		register bpf_u_int32 x = ep->edom[i];
1263 
1264 		while (x != 0) {
1265 			k = ffs(x) - 1;
1266 			x &=~ (1 << k);
1267 			k += i * BITS_PER_WORD;
1268 
1269 			target = fold_edge(ep->succ, edges[k]);
1270 			/*
1271 			 * Check that there is no data dependency between
1272 			 * nodes that will be violated if we move the edge.
1273 			 */
1274 			if (target != 0 && !use_conflict(ep->pred, target)) {
1275 				done = 0;
1276 				ep->succ = target;
1277 				if (JT(target) != 0)
1278 					/*
1279 					 * Start over unless we hit a leaf.
1280 					 */
1281 					goto top;
1282 				return;
1283 			}
1284 		}
1285 	}
1286 }
1287 
1288 
1289 static void
1290 or_pullup(b)
1291 	struct block *b;
1292 {
1293 	int val, at_top;
1294 	struct block *pull;
1295 	struct block **diffp, **samep;
1296 	struct edge *ep;
1297 
1298 	ep = b->in_edges;
1299 	if (ep == 0)
1300 		return;
1301 
1302 	/*
1303 	 * Make sure each predecessor loads the same value.
1304 	 * XXX why?
1305 	 */
1306 	val = ep->pred->val[A_ATOM];
1307 	for (ep = ep->next; ep != 0; ep = ep->next)
1308 		if (val != ep->pred->val[A_ATOM])
1309 			return;
1310 
1311 	if (JT(b->in_edges->pred) == b)
1312 		diffp = &JT(b->in_edges->pred);
1313 	else
1314 		diffp = &JF(b->in_edges->pred);
1315 
1316 	at_top = 1;
1317 	while (1) {
1318 		if (*diffp == 0)
1319 			return;
1320 
1321 		if (JT(*diffp) != JT(b))
1322 			return;
1323 
1324 		if (!SET_MEMBER((*diffp)->dom, b->id))
1325 			return;
1326 
1327 		if ((*diffp)->val[A_ATOM] != val)
1328 			break;
1329 
1330 		diffp = &JF(*diffp);
1331 		at_top = 0;
1332 	}
1333 	samep = &JF(*diffp);
1334 	while (1) {
1335 		if (*samep == 0)
1336 			return;
1337 
1338 		if (JT(*samep) != JT(b))
1339 			return;
1340 
1341 		if (!SET_MEMBER((*samep)->dom, b->id))
1342 			return;
1343 
1344 		if ((*samep)->val[A_ATOM] == val)
1345 			break;
1346 
1347 		/* XXX Need to check that there are no data dependencies
1348 		   between dp0 and dp1.  Currently, the code generator
1349 		   will not produce such dependencies. */
1350 		samep = &JF(*samep);
1351 	}
1352 #ifdef notdef
1353 	/* XXX This doesn't cover everything. */
1354 	for (i = 0; i < N_ATOMS; ++i)
1355 		if ((*samep)->val[i] != pred->val[i])
1356 			return;
1357 #endif
1358 	/* Pull up the node. */
1359 	pull = *samep;
1360 	*samep = JF(pull);
1361 	JF(pull) = *diffp;
1362 
1363 	/*
1364 	 * At the top of the chain, each predecessor needs to point at the
1365 	 * pulled up node.  Inside the chain, there is only one predecessor
1366 	 * to worry about.
1367 	 */
1368 	if (at_top) {
1369 		for (ep = b->in_edges; ep != 0; ep = ep->next) {
1370 			if (JT(ep->pred) == b)
1371 				JT(ep->pred) = pull;
1372 			else
1373 				JF(ep->pred) = pull;
1374 		}
1375 	}
1376 	else
1377 		*diffp = pull;
1378 
1379 	done = 0;
1380 }
1381 
1382 static void
1383 and_pullup(b)
1384 	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 	 */
1398 	val = ep->pred->val[A_ATOM];
1399 	for (ep = ep->next; ep != 0; ep = ep->next)
1400 		if (val != ep->pred->val[A_ATOM])
1401 			return;
1402 
1403 	if (JT(b->in_edges->pred) == b)
1404 		diffp = &JT(b->in_edges->pred);
1405 	else
1406 		diffp = &JF(b->in_edges->pred);
1407 
1408 	at_top = 1;
1409 	while (1) {
1410 		if (*diffp == 0)
1411 			return;
1412 
1413 		if (JF(*diffp) != JF(b))
1414 			return;
1415 
1416 		if (!SET_MEMBER((*diffp)->dom, b->id))
1417 			return;
1418 
1419 		if ((*diffp)->val[A_ATOM] != val)
1420 			break;
1421 
1422 		diffp = &JT(*diffp);
1423 		at_top = 0;
1424 	}
1425 	samep = &JT(*diffp);
1426 	while (1) {
1427 		if (*samep == 0)
1428 			return;
1429 
1430 		if (JF(*samep) != JF(b))
1431 			return;
1432 
1433 		if (!SET_MEMBER((*samep)->dom, b->id))
1434 			return;
1435 
1436 		if ((*samep)->val[A_ATOM] == val)
1437 			break;
1438 
1439 		/* XXX Need to check that there are no data dependencies
1440 		   between diffp and samep.  Currently, the code generator
1441 		   will not produce such dependencies. */
1442 		samep = &JT(*samep);
1443 	}
1444 #ifdef notdef
1445 	/* XXX This doesn't cover everything. */
1446 	for (i = 0; i < N_ATOMS; ++i)
1447 		if ((*samep)->val[i] != pred->val[i])
1448 			return;
1449 #endif
1450 	/* Pull up the node. */
1451 	pull = *samep;
1452 	*samep = JT(pull);
1453 	JT(pull) = *diffp;
1454 
1455 	/*
1456 	 * At the top of the chain, each predecessor needs to point at the
1457 	 * pulled up node.  Inside the chain, there is only one predecessor
1458 	 * to worry about.
1459 	 */
1460 	if (at_top) {
1461 		for (ep = b->in_edges; ep != 0; ep = ep->next) {
1462 			if (JT(ep->pred) == b)
1463 				JT(ep->pred) = pull;
1464 			else
1465 				JF(ep->pred) = pull;
1466 		}
1467 	}
1468 	else
1469 		*diffp = pull;
1470 
1471 	done = 0;
1472 }
1473 
1474 static void
1475 opt_blks(root, do_stmts)
1476 	struct block *root;
1477 	int do_stmts;
1478 {
1479 	int i, maxlevel;
1480 	struct block *p;
1481 
1482 	init_val();
1483 	maxlevel = root->level;
1484 	for (i = maxlevel; i >= 0; --i)
1485 		for (p = levels[i]; p; p = p->link)
1486 			opt_blk(p, do_stmts);
1487 
1488 	if (do_stmts)
1489 		/*
1490 		 * No point trying to move branches; it can't possibly
1491 		 * make a difference at this point.
1492 		 */
1493 		return;
1494 
1495 	for (i = 1; i <= maxlevel; ++i) {
1496 		for (p = levels[i]; p; p = p->link) {
1497 			opt_j(&p->et);
1498 			opt_j(&p->ef);
1499 		}
1500 	}
1501 	for (i = 1; i <= maxlevel; ++i) {
1502 		for (p = levels[i]; p; p = p->link) {
1503 			or_pullup(p);
1504 			and_pullup(p);
1505 		}
1506 	}
1507 }
1508 
1509 static inline void
1510 link_inedge(parent, child)
1511 	struct edge *parent;
1512 	struct block *child;
1513 {
1514 	parent->next = child->in_edges;
1515 	child->in_edges = parent;
1516 }
1517 
1518 static void
1519 find_inedges(root)
1520 	struct block *root;
1521 {
1522 	int i;
1523 	struct block *b;
1524 
1525 	for (i = 0; i < n_blocks; ++i)
1526 		blocks[i]->in_edges = 0;
1527 
1528 	/*
1529 	 * Traverse the graph, adding each edge to the predecessor
1530 	 * list of its successors.  Skip the leaves (i.e. level 0).
1531 	 */
1532 	for (i = root->level; i > 0; --i) {
1533 		for (b = levels[i]; b != 0; b = b->link) {
1534 			link_inedge(&b->et, JT(b));
1535 			link_inedge(&b->ef, JF(b));
1536 		}
1537 	}
1538 }
1539 
1540 static void
1541 opt_root(b)
1542 	struct block **b;
1543 {
1544 	struct slist *tmp, *s;
1545 
1546 	s = (*b)->stmts;
1547 	(*b)->stmts = 0;
1548 	while (BPF_CLASS((*b)->s.code) == BPF_JMP && JT(*b) == JF(*b))
1549 		*b = JT(*b);
1550 
1551 	tmp = (*b)->stmts;
1552 	if (tmp != 0)
1553 		sappend(s, tmp);
1554 	(*b)->stmts = s;
1555 
1556 	/*
1557 	 * If the root node is a return, then there is no
1558 	 * point executing any statements (since the bpf machine
1559 	 * has no side effects).
1560 	 */
1561 	if (BPF_CLASS((*b)->s.code) == BPF_RET)
1562 		(*b)->stmts = 0;
1563 }
1564 
1565 static void
1566 opt_loop(root, do_stmts)
1567 	struct block *root;
1568 	int do_stmts;
1569 {
1570 
1571 #ifdef BDEBUG
1572 	if (dflag > 1)
1573 		opt_dump(root);
1574 #endif
1575 	do {
1576 		done = 1;
1577 		find_levels(root);
1578 		find_dom(root);
1579 		find_closure(root);
1580 		find_inedges(root);
1581 		find_ud(root);
1582 		find_edom(root);
1583 		opt_blks(root, do_stmts);
1584 #ifdef BDEBUG
1585 		if (dflag > 1)
1586 			opt_dump(root);
1587 #endif
1588 	} while (!done);
1589 }
1590 
1591 /*
1592  * Optimize the filter code in its dag representation.
1593  */
1594 void
1595 bpf_optimize(rootp)
1596 	struct block **rootp;
1597 {
1598 	struct block *root;
1599 
1600 	root = *rootp;
1601 
1602 	opt_init(root);
1603 	opt_loop(root, 0);
1604 	opt_loop(root, 1);
1605 	intern_blocks(root);
1606 	opt_root(rootp);
1607 	opt_cleanup();
1608 }
1609 
1610 static void
1611 make_marks(p)
1612 	struct block *p;
1613 {
1614 	if (!isMarked(p)) {
1615 		Mark(p);
1616 		if (BPF_CLASS(p->s.code) != BPF_RET) {
1617 			make_marks(JT(p));
1618 			make_marks(JF(p));
1619 		}
1620 	}
1621 }
1622 
1623 /*
1624  * Mark code array such that isMarked(i) is true
1625  * only for nodes that are alive.
1626  */
1627 static void
1628 mark_code(p)
1629 	struct block *p;
1630 {
1631 	cur_mark += 1;
1632 	make_marks(p);
1633 }
1634 
1635 /*
1636  * True iff the two stmt lists load the same value from the packet into
1637  * the accumulator.
1638  */
1639 static int
1640 eq_slist(x, y)
1641 	struct slist *x, *y;
1642 {
1643 	while (1) {
1644 		while (x && x->s.code == NOP)
1645 			x = x->next;
1646 		while (y && y->s.code == NOP)
1647 			y = y->next;
1648 		if (x == 0)
1649 			return y == 0;
1650 		if (y == 0)
1651 			return x == 0;
1652 		if (x->s.code != y->s.code || x->s.k != y->s.k)
1653 			return 0;
1654 		x = x->next;
1655 		y = y->next;
1656 	}
1657 }
1658 
1659 static inline int
1660 eq_blk(b0, b1)
1661 	struct block *b0, *b1;
1662 {
1663 	if (b0->s.code == b1->s.code &&
1664 	    b0->s.k == b1->s.k &&
1665 	    b0->et.succ == b1->et.succ &&
1666 	    b0->ef.succ == b1->ef.succ)
1667 		return eq_slist(b0->stmts, b1->stmts);
1668 	return 0;
1669 }
1670 
1671 static void
1672 intern_blocks(root)
1673 	struct block *root;
1674 {
1675 	struct block *p;
1676 	int i, j;
1677 	int done;
1678  top:
1679 	done = 1;
1680 	for (i = 0; i < n_blocks; ++i)
1681 		blocks[i]->link = 0;
1682 
1683 	mark_code(root);
1684 
1685 	for (i = n_blocks - 1; --i >= 0; ) {
1686 		if (!isMarked(blocks[i]))
1687 			continue;
1688 		for (j = i + 1; j < n_blocks; ++j) {
1689 			if (!isMarked(blocks[j]))
1690 				continue;
1691 			if (eq_blk(blocks[i], blocks[j])) {
1692 				blocks[i]->link = blocks[j]->link ?
1693 					blocks[j]->link : blocks[j];
1694 				break;
1695 			}
1696 		}
1697 	}
1698 	for (i = 0; i < n_blocks; ++i) {
1699 		p = blocks[i];
1700 		if (JT(p) == 0)
1701 			continue;
1702 		if (JT(p)->link) {
1703 			done = 0;
1704 			JT(p) = JT(p)->link;
1705 		}
1706 		if (JF(p)->link) {
1707 			done = 0;
1708 			JF(p) = JF(p)->link;
1709 		}
1710 	}
1711 	if (!done)
1712 		goto top;
1713 }
1714 
1715 static void
1716 opt_cleanup()
1717 {
1718 	free((void *)vnode_base);
1719 	free((void *)vmap);
1720 	free((void *)edges);
1721 	free((void *)space);
1722 	free((void *)levels);
1723 	free((void *)blocks);
1724 }
1725 
1726 /*
1727  * Return the number of stmts in 's'.
1728  */
1729 static int
1730 slength(s)
1731 	struct slist *s;
1732 {
1733 	int n = 0;
1734 
1735 	for (; s; s = s->next)
1736 		if (s->s.code != NOP)
1737 			++n;
1738 	return n;
1739 }
1740 
1741 /*
1742  * Return the number of nodes reachable by 'p'.
1743  * All nodes should be initially unmarked.
1744  */
1745 static int
1746 count_blocks(p)
1747 	struct block *p;
1748 {
1749 	if (p == 0 || isMarked(p))
1750 		return 0;
1751 	Mark(p);
1752 	return count_blocks(JT(p)) + count_blocks(JF(p)) + 1;
1753 }
1754 
1755 /*
1756  * Do a depth first search on the flow graph, numbering the
1757  * the basic blocks, and entering them into the 'blocks' array.`
1758  */
1759 static void
1760 number_blks_r(p)
1761 	struct block *p;
1762 {
1763 	int n;
1764 
1765 	if (p == 0 || isMarked(p))
1766 		return;
1767 
1768 	Mark(p);
1769 	n = n_blocks++;
1770 	p->id = n;
1771 	blocks[n] = p;
1772 
1773 	number_blks_r(JT(p));
1774 	number_blks_r(JF(p));
1775 }
1776 
1777 /*
1778  * Return the number of stmts in the flowgraph reachable by 'p'.
1779  * The nodes should be unmarked before calling.
1780  */
1781 static int
1782 count_stmts(p)
1783 	struct block *p;
1784 {
1785 	int n;
1786 
1787 	if (p == 0 || isMarked(p))
1788 		return 0;
1789 	Mark(p);
1790 	n = count_stmts(JT(p)) + count_stmts(JF(p));
1791 	return slength(p->stmts) + n + 1;
1792 }
1793 
1794 /*
1795  * Allocate memory.  All allocation is done before optimization
1796  * is begun.  A linear bound on the size of all data structures is computed
1797  * from the total number of blocks and/or statements.
1798  */
1799 static void
1800 opt_init(root)
1801 	struct block *root;
1802 {
1803 	bpf_u_int32 *p;
1804 	int i, n, max_stmts;
1805 
1806 	/*
1807 	 * First, count the blocks, so we can malloc an array to map
1808 	 * block number to block.  Then, put the blocks into the array.
1809 	 */
1810 	unMarkAll();
1811 	n = count_blocks(root);
1812 	blocks = (struct block **)malloc(n * sizeof(*blocks));
1813 	unMarkAll();
1814 	n_blocks = 0;
1815 	number_blks_r(root);
1816 
1817 	n_edges = 2 * n_blocks;
1818 	edges = (struct edge **)malloc(n_edges * sizeof(*edges));
1819 
1820 	/*
1821 	 * The number of levels is bounded by the number of nodes.
1822 	 */
1823 	levels = (struct block **)malloc(n_blocks * sizeof(*levels));
1824 
1825 	edgewords = n_edges / (8 * sizeof(bpf_u_int32)) + 1;
1826 	nodewords = n_blocks / (8 * sizeof(bpf_u_int32)) + 1;
1827 
1828 	/* XXX */
1829 	space = (bpf_u_int32 *)malloc(2 * n_blocks * nodewords * sizeof(*space)
1830 				 + n_edges * edgewords * sizeof(*space));
1831 	p = space;
1832 	all_dom_sets = p;
1833 	for (i = 0; i < n; ++i) {
1834 		blocks[i]->dom = p;
1835 		p += nodewords;
1836 	}
1837 	all_closure_sets = p;
1838 	for (i = 0; i < n; ++i) {
1839 		blocks[i]->closure = p;
1840 		p += nodewords;
1841 	}
1842 	all_edge_sets = p;
1843 	for (i = 0; i < n; ++i) {
1844 		register struct block *b = blocks[i];
1845 
1846 		b->et.edom = p;
1847 		p += edgewords;
1848 		b->ef.edom = p;
1849 		p += edgewords;
1850 		b->et.id = i;
1851 		edges[i] = &b->et;
1852 		b->ef.id = n_blocks + i;
1853 		edges[n_blocks + i] = &b->ef;
1854 		b->et.pred = b;
1855 		b->ef.pred = b;
1856 	}
1857 	max_stmts = 0;
1858 	for (i = 0; i < n; ++i)
1859 		max_stmts += slength(blocks[i]->stmts) + 1;
1860 	/*
1861 	 * We allocate at most 3 value numbers per statement,
1862 	 * so this is an upper bound on the number of valnodes
1863 	 * we'll need.
1864 	 */
1865 	maxval = 3 * max_stmts;
1866 	vmap = (struct vmapinfo *)malloc(maxval * sizeof(*vmap));
1867 	vnode_base = (struct valnode *)malloc(maxval * sizeof(*vmap));
1868 }
1869 
1870 /*
1871  * Some pointers used to convert the basic block form of the code,
1872  * into the array form that BPF requires.  'fstart' will point to
1873  * the malloc'd array while 'ftail' is used during the recursive traversal.
1874  */
1875 static struct bpf_insn *fstart;
1876 static struct bpf_insn *ftail;
1877 
1878 #ifdef BDEBUG
1879 int bids[1000];
1880 #endif
1881 
1882 /*
1883  * Returns true if successful.  Returns false if a branch has
1884  * an offset that is too large.  If so, we have marked that
1885  * branch so that on a subsequent iteration, it will be treated
1886  * properly.
1887  */
1888 static int
1889 convert_code_r(p)
1890 	struct block *p;
1891 {
1892 	struct bpf_insn *dst;
1893 	struct slist *src;
1894 	int slen;
1895 	u_int off;
1896 	int extrajmps;		/* number of extra jumps inserted */
1897 	struct slist **offset = NULL;
1898 
1899 	if (p == 0 || isMarked(p))
1900 		return (1);
1901 	Mark(p);
1902 
1903 	if (convert_code_r(JF(p)) == 0)
1904 		return (0);
1905 	if (convert_code_r(JT(p)) == 0)
1906 		return (0);
1907 
1908 	slen = slength(p->stmts);
1909 	dst = ftail -= (slen + 1 + p->longjt + p->longjf);
1910 		/* inflate length by any extra jumps */
1911 
1912 	p->offset = dst - fstart;
1913 
1914 	/* generate offset[] for convenience  */
1915 	if (slen) {
1916 		offset = (struct slist **)calloc(sizeof(struct slist *), slen);
1917 		if (!offset) {
1918 			bpf_error("not enough core");
1919 			/*NOTREACHED*/
1920 		}
1921 	}
1922 	src = p->stmts;
1923 	for (off = 0; off < slen && src; off++) {
1924 #if 0
1925 		printf("off=%d src=%x\n", off, src);
1926 #endif
1927 		offset[off] = src;
1928 		src = src->next;
1929 	}
1930 
1931 	off = 0;
1932 	for (src = p->stmts; src; src = src->next) {
1933 		if (src->s.code == NOP)
1934 			continue;
1935 		dst->code = (u_short)src->s.code;
1936 		dst->k = src->s.k;
1937 
1938 		/* fill block-local relative jump */
1939 		if (BPF_CLASS(src->s.code) != BPF_JMP || src->s.code == BPF_JMP|BPF_JA) {
1940 #if 0
1941 			if (src->s.jt || src->s.jf) {
1942 				bpf_error("illegal jmp destination");
1943 				/*NOTREACHED*/
1944 			}
1945 #endif
1946 			goto filled;
1947 		}
1948 		if (off == slen - 2)	/*???*/
1949 			goto filled;
1950 
1951 	    {
1952 		int i;
1953 		int jt, jf;
1954 		char *ljerr = "%s for block-local relative jump: off=%d";
1955 
1956 #if 0
1957 		printf("code=%x off=%d %x %x\n", src->s.code,
1958 			off, src->s.jt, src->s.jf);
1959 #endif
1960 
1961 		if (!src->s.jt || !src->s.jf) {
1962 			bpf_error(ljerr, "no jmp destination", off);
1963 			/*NOTREACHED*/
1964 		}
1965 
1966 		jt = jf = 0;
1967 		for (i = 0; i < slen; i++) {
1968 			if (offset[i] == src->s.jt) {
1969 				if (jt) {
1970 					bpf_error(ljerr, "multiple matches", off);
1971 					/*NOTREACHED*/
1972 				}
1973 
1974 				dst->jt = i - off - 1;
1975 				jt++;
1976 			}
1977 			if (offset[i] == src->s.jf) {
1978 				if (jf) {
1979 					bpf_error(ljerr, "multiple matches", off);
1980 					/*NOTREACHED*/
1981 				}
1982 				dst->jf = i - off - 1;
1983 				jf++;
1984 			}
1985 		}
1986 		if (!jt || !jf) {
1987 			bpf_error(ljerr, "no destination found", off);
1988 			/*NOTREACHED*/
1989 		}
1990 	    }
1991 filled:
1992 		++dst;
1993 		++off;
1994 	}
1995 	if (offset)
1996 		free(offset);
1997 
1998 #ifdef BDEBUG
1999 	bids[dst - fstart] = p->id + 1;
2000 #endif
2001 	dst->code = (u_short)p->s.code;
2002 	dst->k = p->s.k;
2003 	if (JT(p)) {
2004 		extrajmps = 0;
2005 		off = JT(p)->offset - (p->offset + slen) - 1;
2006 		if (off >= 256) {
2007 		    /* offset too large for branch, must add a jump */
2008 		    if (p->longjt == 0) {
2009 		    	/* mark this instruction and retry */
2010 			p->longjt++;
2011 			return(0);
2012 		    }
2013 		    /* branch if T to following jump */
2014 		    dst->jt = extrajmps;
2015 		    extrajmps++;
2016 		    dst[extrajmps].code = BPF_JMP|BPF_JA;
2017 		    dst[extrajmps].k = off - extrajmps;
2018 		}
2019 		else
2020 		    dst->jt = off;
2021 		off = JF(p)->offset - (p->offset + slen) - 1;
2022 		if (off >= 256) {
2023 		    /* offset too large for branch, must add a jump */
2024 		    if (p->longjf == 0) {
2025 		    	/* mark this instruction and retry */
2026 			p->longjf++;
2027 			return(0);
2028 		    }
2029 		    /* branch if F to following jump */
2030 		    /* if two jumps are inserted, F goes to second one */
2031 		    dst->jf = extrajmps;
2032 		    extrajmps++;
2033 		    dst[extrajmps].code = BPF_JMP|BPF_JA;
2034 		    dst[extrajmps].k = off - extrajmps;
2035 		}
2036 		else
2037 		    dst->jf = off;
2038 	}
2039 	return (1);
2040 }
2041 
2042 
2043 /*
2044  * Convert flowgraph intermediate representation to the
2045  * BPF array representation.  Set *lenp to the number of instructions.
2046  */
2047 struct bpf_insn *
2048 icode_to_fcode(root, lenp)
2049 	struct block *root;
2050 	int *lenp;
2051 {
2052 	int n;
2053 	struct bpf_insn *fp;
2054 
2055 	/*
2056 	 * Loop doing convert_codr_r() until no branches remain
2057 	 * with too-large offsets.
2058 	 */
2059 	while (1) {
2060 	    unMarkAll();
2061 	    n = *lenp = count_stmts(root);
2062 
2063 	    fp = (struct bpf_insn *)malloc(sizeof(*fp) * n);
2064 	    memset((char *)fp, 0, sizeof(*fp) * n);
2065 	    fstart = fp;
2066 	    ftail = fp + n;
2067 
2068 	    unMarkAll();
2069 	    if (convert_code_r(root))
2070 		break;
2071 	    free(fp);
2072 	}
2073 
2074 	return fp;
2075 }
2076 
2077 #ifdef BDEBUG
2078 static void
2079 opt_dump(root)
2080 	struct block *root;
2081 {
2082 	struct bpf_program f;
2083 
2084 	memset(bids, 0, sizeof bids);
2085 	f.bf_insns = icode_to_fcode(root, &f.bf_len);
2086 	bpf_dump(&f, 1);
2087 	putchar('\n');
2088 	free((char *)f.bf_insns);
2089 }
2090 #endif
2091