1 /*
2 * Copyright (c) 1988, 1989, 1990, 1991, 1993, 1994, 1995, 1996
3 * The Regents of the University of California. All rights reserved.
4 *
5 * Redistribution and use in source and binary forms, with or without
6 * modification, are permitted provided that: (1) source code distributions
7 * retain the above copyright notice and this paragraph in its entirety, (2)
8 * distributions including binary code include the above copyright notice and
9 * this paragraph in its entirety in the documentation or other materials
10 * provided with the distribution, and (3) all advertising materials mentioning
11 * features or use of this software display the following acknowledgement:
12 * ``This product includes software developed by the University of California,
13 * Lawrence Berkeley Laboratory and its contributors.'' Neither the name of
14 * the University nor the names of its contributors may be used to endorse
15 * or promote products derived from this software without specific prior
16 * written permission.
17 * THIS SOFTWARE IS PROVIDED ``AS IS'' AND WITHOUT ANY EXPRESS OR IMPLIED
18 * WARRANTIES, INCLUDING, WITHOUT LIMITATION, THE IMPLIED WARRANTIES OF
19 * MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE.
20 *
21 * Optimization module for tcpdump intermediate representation.
22 */
23 #ifndef lint
24 static const char rcsid[] _U_ =
25 "@(#) $Header: /tcpdump/master/libpcap/optimize.c,v 1.85.2.3 2007/09/12 21:29:45 guy Exp $ (LBL)";
26 #endif
27
28 #ifdef HAVE_CONFIG_H
29 #include "config.h"
30 #endif
31
32 #include <stdio.h>
33 #include <stdlib.h>
34 #include <memory.h>
35 #include <string.h>
36
37 #include <errno.h>
38
39 #include "pcap-int.h"
40
41 #include "gencode.h"
42
43 #ifdef HAVE_OS_PROTO_H
44 #include "os-proto.h"
45 #endif
46
47 #ifdef BDEBUG
48 extern int dflag;
49 #endif
50
51 #if defined(MSDOS) && !defined(__DJGPP__)
52 extern int _w32_ffs (int mask);
53 #define ffs _w32_ffs
54 #endif
55
56 /*
57 * Represents a deleted instruction.
58 */
59 #define NOP -1
60
61 /*
62 * Register numbers for use-def values.
63 * 0 through BPF_MEMWORDS-1 represent the corresponding scratch memory
64 * location. A_ATOM is the accumulator and X_ATOM is the index
65 * register.
66 */
67 #define A_ATOM BPF_MEMWORDS
68 #define X_ATOM (BPF_MEMWORDS+1)
69
70 /*
71 * This define is used to represent *both* the accumulator and
72 * x register in use-def computations.
73 * Currently, the use-def code assumes only one definition per instruction.
74 */
75 #define AX_ATOM N_ATOMS
76
77 /*
78 * A flag to indicate that further optimization is needed.
79 * Iterative passes are continued until a given pass yields no
80 * branch movement.
81 */
82 static int done;
83
84 /*
85 * A block is marked if only if its mark equals the current mark.
86 * Rather than traverse the code array, marking each item, 'cur_mark' is
87 * incremented. This automatically makes each element unmarked.
88 */
89 static int cur_mark;
90 #define isMarked(p) ((p)->mark == cur_mark)
91 #define unMarkAll() cur_mark += 1
92 #define Mark(p) ((p)->mark = cur_mark)
93
94 static void opt_init(struct block *);
95 static void opt_cleanup(void);
96
97 static void make_marks(struct block *);
98 static void mark_code(struct block *);
99
100 static void intern_blocks(struct block *);
101
102 static int eq_slist(struct slist *, struct slist *);
103
104 static void find_levels_r(struct block *);
105
106 static void find_levels(struct block *);
107 static void find_dom(struct block *);
108 static void propedom(struct edge *);
109 static void find_edom(struct block *);
110 static void find_closure(struct block *);
111 static int atomuse(struct stmt *);
112 static int atomdef(struct stmt *);
113 static void compute_local_ud(struct block *);
114 static void find_ud(struct block *);
115 static void init_val(void);
116 static int F(int, int, int);
117 static inline void vstore(struct stmt *, int *, int, int);
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 void or_pullup(struct block *);
122 static void and_pullup(struct block *);
123 static void opt_blks(struct block *, int);
124 static inline void link_inedge(struct edge *, struct block *);
125 static void find_inedges(struct block *);
126 static void opt_root(struct block **);
127 static void opt_loop(struct block *, int);
128 static void fold_op(struct stmt *, int, int);
129 static inline struct slist *this_op(struct slist *);
130 static void opt_not(struct block *);
131 static void opt_peep(struct block *);
132 static void opt_stmt(struct stmt *, int[], int);
133 static void deadstmt(struct stmt *, struct stmt *[]);
134 static void opt_deadstores(struct block *);
135 static struct block *fold_edge(struct block *, struct edge *);
136 static inline int eq_blk(struct block *, struct block *);
137 static int slength(struct slist *);
138 static int count_blocks(struct block *);
139 static void number_blks_r(struct block *);
140 static int count_stmts(struct block *);
141 static int convert_code_r(struct block *);
142 #ifdef BDEBUG
143 static void opt_dump(struct block *);
144 #endif
145
146 static int n_blocks;
147 struct block **blocks;
148 static int n_edges;
149 struct edge **edges;
150
151 /*
152 * A bit vector set representation of the dominators.
153 * We round up the set size to the next power of two.
154 */
155 static int nodewords;
156 static int edgewords;
157 struct block **levels;
158 bpf_u_int32 *space;
159 #define BITS_PER_WORD (8*sizeof(bpf_u_int32))
160 /*
161 * True if a is in uset {p}
162 */
163 #define SET_MEMBER(p, a) \
164 ((p)[(unsigned)(a) / BITS_PER_WORD] & (1 << ((unsigned)(a) % BITS_PER_WORD)))
165
166 /*
167 * Add 'a' to uset p.
168 */
169 #define SET_INSERT(p, a) \
170 (p)[(unsigned)(a) / BITS_PER_WORD] |= (1 << ((unsigned)(a) % BITS_PER_WORD))
171
172 /*
173 * Delete 'a' from uset p.
174 */
175 #define SET_DELETE(p, a) \
176 (p)[(unsigned)(a) / BITS_PER_WORD] &= ~(1 << ((unsigned)(a) % BITS_PER_WORD))
177
178 /*
179 * a := a intersect b
180 */
181 #define SET_INTERSECT(a, b, n)\
182 {\
183 register bpf_u_int32 *_x = a, *_y = b;\
184 register int _n = n;\
185 while (--_n >= 0) *_x++ &= *_y++;\
186 }
187
188 /*
189 * a := a - b
190 */
191 #define SET_SUBTRACT(a, b, n)\
192 {\
193 register bpf_u_int32 *_x = a, *_y = b;\
194 register int _n = n;\
195 while (--_n >= 0) *_x++ &=~ *_y++;\
196 }
197
198 /*
199 * a := a union b
200 */
201 #define SET_UNION(a, b, n)\
202 {\
203 register bpf_u_int32 *_x = a, *_y = b;\
204 register int _n = n;\
205 while (--_n >= 0) *_x++ |= *_y++;\
206 }
207
208 static uset all_dom_sets;
209 static uset all_closure_sets;
210 static uset all_edge_sets;
211
212 #ifndef MAX
213 #define MAX(a,b) ((a)>(b)?(a):(b))
214 #endif
215
216 static void
find_levels_r(b)217 find_levels_r(b)
218 struct block *b;
219 {
220 int level;
221
222 if (isMarked(b))
223 return;
224
225 Mark(b);
226 b->link = 0;
227
228 if (JT(b)) {
229 find_levels_r(JT(b));
230 find_levels_r(JF(b));
231 level = MAX(JT(b)->level, JF(b)->level) + 1;
232 } else
233 level = 0;
234 b->level = level;
235 b->link = levels[level];
236 levels[level] = b;
237 }
238
239 /*
240 * Level graph. The levels go from 0 at the leaves to
241 * N_LEVELS at the root. The levels[] array points to the
242 * first node of the level list, whose elements are linked
243 * with the 'link' field of the struct block.
244 */
245 static void
find_levels(root)246 find_levels(root)
247 struct block *root;
248 {
249 memset((char *)levels, 0, n_blocks * sizeof(*levels));
250 unMarkAll();
251 find_levels_r(root);
252 }
253
254 /*
255 * Find dominator relationships.
256 * Assumes graph has been leveled.
257 */
258 static void
find_dom(root)259 find_dom(root)
260 struct block *root;
261 {
262 int i;
263 struct block *b;
264 bpf_u_int32 *x;
265
266 /*
267 * Initialize sets to contain all nodes.
268 */
269 x = all_dom_sets;
270 i = n_blocks * nodewords;
271 while (--i >= 0)
272 *x++ = ~0;
273 /* Root starts off empty. */
274 for (i = nodewords; --i >= 0;)
275 root->dom[i] = 0;
276
277 /* root->level is the highest level no found. */
278 for (i = root->level; i >= 0; --i) {
279 for (b = levels[i]; b; b = b->link) {
280 SET_INSERT(b->dom, b->id);
281 if (JT(b) == 0)
282 continue;
283 SET_INTERSECT(JT(b)->dom, b->dom, nodewords);
284 SET_INTERSECT(JF(b)->dom, b->dom, nodewords);
285 }
286 }
287 }
288
289 static void
propedom(ep)290 propedom(ep)
291 struct edge *ep;
292 {
293 SET_INSERT(ep->edom, ep->id);
294 if (ep->succ) {
295 SET_INTERSECT(ep->succ->et.edom, ep->edom, edgewords);
296 SET_INTERSECT(ep->succ->ef.edom, ep->edom, edgewords);
297 }
298 }
299
300 /*
301 * Compute edge dominators.
302 * Assumes graph has been leveled and predecessors established.
303 */
304 static void
find_edom(root)305 find_edom(root)
306 struct block *root;
307 {
308 int i;
309 uset x;
310 struct block *b;
311
312 x = all_edge_sets;
313 for (i = n_edges * edgewords; --i >= 0; )
314 x[i] = ~0;
315
316 /* root->level is the highest level no found. */
317 memset(root->et.edom, 0, edgewords * sizeof(*(uset)0));
318 memset(root->ef.edom, 0, edgewords * sizeof(*(uset)0));
319 for (i = root->level; i >= 0; --i) {
320 for (b = levels[i]; b != 0; b = b->link) {
321 propedom(&b->et);
322 propedom(&b->ef);
323 }
324 }
325 }
326
327 /*
328 * Find the backwards transitive closure of the flow graph. These sets
329 * are backwards in the sense that we find the set of nodes that reach
330 * a given node, not the set of nodes that can be reached by a node.
331 *
332 * Assumes graph has been leveled.
333 */
334 static void
find_closure(root)335 find_closure(root)
336 struct block *root;
337 {
338 int i;
339 struct block *b;
340
341 /*
342 * Initialize sets to contain no nodes.
343 */
344 memset((char *)all_closure_sets, 0,
345 n_blocks * nodewords * sizeof(*all_closure_sets));
346
347 /* root->level is the highest level no found. */
348 for (i = root->level; i >= 0; --i) {
349 for (b = levels[i]; b; b = b->link) {
350 SET_INSERT(b->closure, b->id);
351 if (JT(b) == 0)
352 continue;
353 SET_UNION(JT(b)->closure, b->closure, nodewords);
354 SET_UNION(JF(b)->closure, b->closure, nodewords);
355 }
356 }
357 }
358
359 /*
360 * Return the register number that is used by s. If A and X are both
361 * used, return AX_ATOM. If no register is used, return -1.
362 *
363 * The implementation should probably change to an array access.
364 */
365 static int
atomuse(s)366 atomuse(s)
367 struct stmt *s;
368 {
369 register int c = s->code;
370
371 if (c == NOP)
372 return -1;
373
374 switch (BPF_CLASS(c)) {
375
376 case BPF_RET:
377 return (BPF_RVAL(c) == BPF_A) ? A_ATOM :
378 (BPF_RVAL(c) == BPF_X) ? X_ATOM : -1;
379
380 case BPF_LD:
381 case BPF_LDX:
382 return (BPF_MODE(c) == BPF_IND) ? X_ATOM :
383 (BPF_MODE(c) == BPF_MEM) ? s->k : -1;
384
385 case BPF_ST:
386 return A_ATOM;
387
388 case BPF_STX:
389 return X_ATOM;
390
391 case BPF_JMP:
392 case BPF_ALU:
393 if (BPF_SRC(c) == BPF_X)
394 return AX_ATOM;
395 return A_ATOM;
396
397 case BPF_MISC:
398 return BPF_MISCOP(c) == BPF_TXA ? X_ATOM : A_ATOM;
399 }
400 abort();
401 /* NOTREACHED */
402 }
403
404 /*
405 * Return the register number that is defined by 's'. We assume that
406 * a single stmt cannot define more than one register. If no register
407 * is defined, return -1.
408 *
409 * The implementation should probably change to an array access.
410 */
411 static int
atomdef(s)412 atomdef(s)
413 struct stmt *s;
414 {
415 if (s->code == NOP)
416 return -1;
417
418 switch (BPF_CLASS(s->code)) {
419
420 case BPF_LD:
421 case BPF_ALU:
422 return A_ATOM;
423
424 case BPF_LDX:
425 return X_ATOM;
426
427 case BPF_ST:
428 case BPF_STX:
429 return s->k;
430
431 case BPF_MISC:
432 return BPF_MISCOP(s->code) == BPF_TAX ? X_ATOM : A_ATOM;
433 }
434 return -1;
435 }
436
437 /*
438 * Compute the sets of registers used, defined, and killed by 'b'.
439 *
440 * "Used" means that a statement in 'b' uses the register before any
441 * statement in 'b' defines it, i.e. it uses the value left in
442 * that register by a predecessor block of this block.
443 * "Defined" means that a statement in 'b' defines it.
444 * "Killed" means that a statement in 'b' defines it before any
445 * statement in 'b' uses it, i.e. it kills the value left in that
446 * register by a predecessor block of this block.
447 */
448 static void
compute_local_ud(b)449 compute_local_ud(b)
450 struct block *b;
451 {
452 struct slist *s;
453 atomset def = 0, use = 0, kill = 0;
454 int atom;
455
456 for (s = b->stmts; s; s = s->next) {
457 if (s->s.code == NOP)
458 continue;
459 atom = atomuse(&s->s);
460 if (atom >= 0) {
461 if (atom == AX_ATOM) {
462 if (!ATOMELEM(def, X_ATOM))
463 use |= ATOMMASK(X_ATOM);
464 if (!ATOMELEM(def, A_ATOM))
465 use |= ATOMMASK(A_ATOM);
466 }
467 else if (atom < N_ATOMS) {
468 if (!ATOMELEM(def, atom))
469 use |= ATOMMASK(atom);
470 }
471 else
472 abort();
473 }
474 atom = atomdef(&s->s);
475 if (atom >= 0) {
476 if (!ATOMELEM(use, atom))
477 kill |= ATOMMASK(atom);
478 def |= ATOMMASK(atom);
479 }
480 }
481 if (BPF_CLASS(b->s.code) == BPF_JMP) {
482 /*
483 * XXX - what about RET?
484 */
485 atom = atomuse(&b->s);
486 if (atom >= 0) {
487 if (atom == AX_ATOM) {
488 if (!ATOMELEM(def, X_ATOM))
489 use |= ATOMMASK(X_ATOM);
490 if (!ATOMELEM(def, A_ATOM))
491 use |= ATOMMASK(A_ATOM);
492 }
493 else if (atom < N_ATOMS) {
494 if (!ATOMELEM(def, atom))
495 use |= ATOMMASK(atom);
496 }
497 else
498 abort();
499 }
500 }
501
502 b->def = def;
503 b->kill = kill;
504 b->in_use = use;
505 }
506
507 /*
508 * Assume graph is already leveled.
509 */
510 static void
find_ud(root)511 find_ud(root)
512 struct block *root;
513 {
514 int i, maxlevel;
515 struct block *p;
516
517 /*
518 * root->level is the highest level no found;
519 * count down from there.
520 */
521 maxlevel = root->level;
522 for (i = maxlevel; i >= 0; --i)
523 for (p = levels[i]; p; p = p->link) {
524 compute_local_ud(p);
525 p->out_use = 0;
526 }
527
528 for (i = 1; i <= maxlevel; ++i) {
529 for (p = levels[i]; p; p = p->link) {
530 p->out_use |= JT(p)->in_use | JF(p)->in_use;
531 p->in_use |= p->out_use &~ p->kill;
532 }
533 }
534 }
535
536 /*
537 * These data structures are used in a Cocke and Shwarz style
538 * value numbering scheme. Since the flowgraph is acyclic,
539 * exit values can be propagated from a node's predecessors
540 * provided it is uniquely defined.
541 */
542 struct valnode {
543 int code;
544 int v0, v1;
545 int val;
546 struct valnode *next;
547 };
548
549 #define MODULUS 213
550 static struct valnode *hashtbl[MODULUS];
551 static int curval;
552 static int maxval;
553
554 /* Integer constants mapped with the load immediate opcode. */
555 #define K(i) F(BPF_LD|BPF_IMM|BPF_W, i, 0L)
556
557 struct vmapinfo {
558 int is_const;
559 bpf_int32 const_val;
560 };
561
562 struct vmapinfo *vmap;
563 struct valnode *vnode_base;
564 struct valnode *next_vnode;
565
566 static void
init_val()567 init_val()
568 {
569 curval = 0;
570 next_vnode = vnode_base;
571 memset((char *)vmap, 0, maxval * sizeof(*vmap));
572 memset((char *)hashtbl, 0, sizeof hashtbl);
573 }
574
575 /* Because we really don't have an IR, this stuff is a little messy. */
576 static int
F(code,v0,v1)577 F(code, v0, v1)
578 int code;
579 int v0, v1;
580 {
581 u_int hash;
582 int val;
583 struct valnode *p;
584
585 hash = (u_int)code ^ (v0 << 4) ^ (v1 << 8);
586 hash %= MODULUS;
587
588 for (p = hashtbl[hash]; p; p = p->next)
589 if (p->code == code && p->v0 == v0 && p->v1 == v1)
590 return p->val;
591
592 val = ++curval;
593 if (BPF_MODE(code) == BPF_IMM &&
594 (BPF_CLASS(code) == BPF_LD || BPF_CLASS(code) == BPF_LDX)) {
595 vmap[val].const_val = v0;
596 vmap[val].is_const = 1;
597 }
598 p = next_vnode++;
599 p->val = val;
600 p->code = code;
601 p->v0 = v0;
602 p->v1 = v1;
603 p->next = hashtbl[hash];
604 hashtbl[hash] = p;
605
606 return val;
607 }
608
609 static inline void
vstore(s,valp,newval,alter)610 vstore(s, valp, newval, alter)
611 struct stmt *s;
612 int *valp;
613 int newval;
614 int alter;
615 {
616 if (alter && *valp == newval)
617 s->code = NOP;
618 else
619 *valp = newval;
620 }
621
622 static void
fold_op(s,v0,v1)623 fold_op(s, v0, v1)
624 struct stmt *s;
625 int v0, v1;
626 {
627 bpf_u_int32 a, b;
628
629 a = vmap[v0].const_val;
630 b = vmap[v1].const_val;
631
632 switch (BPF_OP(s->code)) {
633 case BPF_ADD:
634 a += b;
635 break;
636
637 case BPF_SUB:
638 a -= b;
639 break;
640
641 case BPF_MUL:
642 a *= b;
643 break;
644
645 case BPF_DIV:
646 if (b == 0)
647 bpf_error("division by zero");
648 a /= b;
649 break;
650
651 case BPF_AND:
652 a &= b;
653 break;
654
655 case BPF_OR:
656 a |= b;
657 break;
658
659 case BPF_LSH:
660 a <<= b;
661 break;
662
663 case BPF_RSH:
664 a >>= b;
665 break;
666
667 case BPF_NEG:
668 a = -a;
669 break;
670
671 default:
672 abort();
673 }
674 s->k = a;
675 s->code = BPF_LD|BPF_IMM;
676 done = 0;
677 }
678
679 static inline struct slist *
this_op(s)680 this_op(s)
681 struct slist *s;
682 {
683 while (s != 0 && s->s.code == NOP)
684 s = s->next;
685 return s;
686 }
687
688 static void
opt_not(b)689 opt_not(b)
690 struct block *b;
691 {
692 struct block *tmp = JT(b);
693
694 JT(b) = JF(b);
695 JF(b) = tmp;
696 }
697
698 static void
opt_peep(b)699 opt_peep(b)
700 struct block *b;
701 {
702 struct slist *s;
703 struct slist *next, *last;
704 int val;
705
706 s = b->stmts;
707 if (s == 0)
708 return;
709
710 last = s;
711 for (/*empty*/; /*empty*/; s = next) {
712 /*
713 * Skip over nops.
714 */
715 s = this_op(s);
716 if (s == 0)
717 break; /* nothing left in the block */
718
719 /*
720 * Find the next real instruction after that one
721 * (skipping nops).
722 */
723 next = this_op(s->next);
724 if (next == 0)
725 break; /* no next instruction */
726 last = next;
727
728 /*
729 * st M[k] --> st M[k]
730 * ldx M[k] tax
731 */
732 if (s->s.code == BPF_ST &&
733 next->s.code == (BPF_LDX|BPF_MEM) &&
734 s->s.k == next->s.k) {
735 done = 0;
736 next->s.code = BPF_MISC|BPF_TAX;
737 }
738 /*
739 * ld #k --> ldx #k
740 * tax txa
741 */
742 if (s->s.code == (BPF_LD|BPF_IMM) &&
743 next->s.code == (BPF_MISC|BPF_TAX)) {
744 s->s.code = BPF_LDX|BPF_IMM;
745 next->s.code = BPF_MISC|BPF_TXA;
746 done = 0;
747 }
748 /*
749 * This is an ugly special case, but it happens
750 * when you say tcp[k] or udp[k] where k is a constant.
751 */
752 if (s->s.code == (BPF_LD|BPF_IMM)) {
753 struct slist *add, *tax, *ild;
754
755 /*
756 * Check that X isn't used on exit from this
757 * block (which the optimizer might cause).
758 * We know the code generator won't generate
759 * any local dependencies.
760 */
761 if (ATOMELEM(b->out_use, X_ATOM))
762 continue;
763
764 /*
765 * Check that the instruction following the ldi
766 * is an addx, or it's an ldxms with an addx
767 * following it (with 0 or more nops between the
768 * ldxms and addx).
769 */
770 if (next->s.code != (BPF_LDX|BPF_MSH|BPF_B))
771 add = next;
772 else
773 add = this_op(next->next);
774 if (add == 0 || add->s.code != (BPF_ALU|BPF_ADD|BPF_X))
775 continue;
776
777 /*
778 * Check that a tax follows that (with 0 or more
779 * nops between them).
780 */
781 tax = this_op(add->next);
782 if (tax == 0 || tax->s.code != (BPF_MISC|BPF_TAX))
783 continue;
784
785 /*
786 * Check that an ild follows that (with 0 or more
787 * nops between them).
788 */
789 ild = this_op(tax->next);
790 if (ild == 0 || BPF_CLASS(ild->s.code) != BPF_LD ||
791 BPF_MODE(ild->s.code) != BPF_IND)
792 continue;
793 /*
794 * We want to turn this sequence:
795 *
796 * (004) ldi #0x2 {s}
797 * (005) ldxms [14] {next} -- optional
798 * (006) addx {add}
799 * (007) tax {tax}
800 * (008) ild [x+0] {ild}
801 *
802 * into this sequence:
803 *
804 * (004) nop
805 * (005) ldxms [14]
806 * (006) nop
807 * (007) nop
808 * (008) ild [x+2]
809 *
810 * XXX We need to check that X is not
811 * subsequently used, because we want to change
812 * what'll be in it after this sequence.
813 *
814 * We know we can eliminate the accumulator
815 * modifications earlier in the sequence since
816 * it is defined by the last stmt of this sequence
817 * (i.e., the last statement of the sequence loads
818 * a value into the accumulator, so we can eliminate
819 * earlier operations on the accumulator).
820 */
821 ild->s.k += s->s.k;
822 s->s.code = NOP;
823 add->s.code = NOP;
824 tax->s.code = NOP;
825 done = 0;
826 }
827 }
828 /*
829 * If the comparison at the end of a block is an equality
830 * comparison against a constant, and nobody uses the value
831 * we leave in the A register at the end of a block, and
832 * the operation preceding the comparison is an arithmetic
833 * operation, we can sometime optimize it away.
834 */
835 if (b->s.code == (BPF_JMP|BPF_JEQ|BPF_K) &&
836 !ATOMELEM(b->out_use, A_ATOM)) {
837 /*
838 * We can optimize away certain subtractions of the
839 * X register.
840 */
841 if (last->s.code == (BPF_ALU|BPF_SUB|BPF_X)) {
842 val = b->val[X_ATOM];
843 if (vmap[val].is_const) {
844 /*
845 * If we have a subtract to do a comparison,
846 * and the X register is a known constant,
847 * we can merge this value into the
848 * comparison:
849 *
850 * sub x -> nop
851 * jeq #y jeq #(x+y)
852 */
853 b->s.k += vmap[val].const_val;
854 last->s.code = NOP;
855 done = 0;
856 } else if (b->s.k == 0) {
857 /*
858 * If the X register isn't a constant,
859 * and the comparison in the test is
860 * against 0, we can compare with the
861 * X register, instead:
862 *
863 * sub x -> nop
864 * jeq #0 jeq x
865 */
866 last->s.code = NOP;
867 b->s.code = BPF_JMP|BPF_JEQ|BPF_X;
868 done = 0;
869 }
870 }
871 /*
872 * Likewise, a constant subtract can be simplified:
873 *
874 * sub #x -> nop
875 * jeq #y -> jeq #(x+y)
876 */
877 else if (last->s.code == (BPF_ALU|BPF_SUB|BPF_K)) {
878 last->s.code = NOP;
879 b->s.k += last->s.k;
880 done = 0;
881 }
882 /*
883 * And, similarly, a constant AND can be simplified
884 * if we're testing against 0, i.e.:
885 *
886 * and #k nop
887 * jeq #0 -> jset #k
888 */
889 else if (last->s.code == (BPF_ALU|BPF_AND|BPF_K) &&
890 b->s.k == 0) {
891 b->s.k = last->s.k;
892 b->s.code = BPF_JMP|BPF_K|BPF_JSET;
893 last->s.code = NOP;
894 done = 0;
895 opt_not(b);
896 }
897 }
898 /*
899 * jset #0 -> never
900 * jset #ffffffff -> always
901 */
902 if (b->s.code == (BPF_JMP|BPF_K|BPF_JSET)) {
903 if (b->s.k == 0)
904 JT(b) = JF(b);
905 if (b->s.k == 0xffffffff)
906 JF(b) = JT(b);
907 }
908 /*
909 * If the accumulator is a known constant, we can compute the
910 * comparison result.
911 */
912 val = b->val[A_ATOM];
913 if (vmap[val].is_const && BPF_SRC(b->s.code) == BPF_K) {
914 bpf_int32 v = vmap[val].const_val;
915 switch (BPF_OP(b->s.code)) {
916
917 case BPF_JEQ:
918 v = v == b->s.k;
919 break;
920
921 case BPF_JGT:
922 v = (unsigned)v > b->s.k;
923 break;
924
925 case BPF_JGE:
926 v = (unsigned)v >= b->s.k;
927 break;
928
929 case BPF_JSET:
930 v &= b->s.k;
931 break;
932
933 default:
934 abort();
935 }
936 if (JF(b) != JT(b))
937 done = 0;
938 if (v)
939 JF(b) = JT(b);
940 else
941 JT(b) = JF(b);
942 }
943 }
944
945 /*
946 * Compute the symbolic value of expression of 's', and update
947 * anything it defines in the value table 'val'. If 'alter' is true,
948 * do various optimizations. This code would be cleaner if symbolic
949 * evaluation and code transformations weren't folded together.
950 */
951 static void
opt_stmt(s,val,alter)952 opt_stmt(s, val, alter)
953 struct stmt *s;
954 int val[];
955 int alter;
956 {
957 int op;
958 int v;
959
960 switch (s->code) {
961
962 case BPF_LD|BPF_ABS|BPF_W:
963 case BPF_LD|BPF_ABS|BPF_H:
964 case BPF_LD|BPF_ABS|BPF_B:
965 v = F(s->code, s->k, 0L);
966 vstore(s, &val[A_ATOM], v, alter);
967 break;
968
969 case BPF_LD|BPF_IND|BPF_W:
970 case BPF_LD|BPF_IND|BPF_H:
971 case BPF_LD|BPF_IND|BPF_B:
972 v = val[X_ATOM];
973 if (alter && vmap[v].is_const) {
974 s->code = BPF_LD|BPF_ABS|BPF_SIZE(s->code);
975 s->k += vmap[v].const_val;
976 v = F(s->code, s->k, 0L);
977 done = 0;
978 }
979 else
980 v = F(s->code, s->k, v);
981 vstore(s, &val[A_ATOM], v, alter);
982 break;
983
984 case BPF_LD|BPF_LEN:
985 v = F(s->code, 0L, 0L);
986 vstore(s, &val[A_ATOM], v, alter);
987 break;
988
989 case BPF_LD|BPF_IMM:
990 v = K(s->k);
991 vstore(s, &val[A_ATOM], v, alter);
992 break;
993
994 case BPF_LDX|BPF_IMM:
995 v = K(s->k);
996 vstore(s, &val[X_ATOM], v, alter);
997 break;
998
999 case BPF_LDX|BPF_MSH|BPF_B:
1000 v = F(s->code, s->k, 0L);
1001 vstore(s, &val[X_ATOM], v, alter);
1002 break;
1003
1004 case BPF_ALU|BPF_NEG:
1005 if (alter && vmap[val[A_ATOM]].is_const) {
1006 s->code = BPF_LD|BPF_IMM;
1007 s->k = -vmap[val[A_ATOM]].const_val;
1008 val[A_ATOM] = K(s->k);
1009 }
1010 else
1011 val[A_ATOM] = F(s->code, val[A_ATOM], 0L);
1012 break;
1013
1014 case BPF_ALU|BPF_ADD|BPF_K:
1015 case BPF_ALU|BPF_SUB|BPF_K:
1016 case BPF_ALU|BPF_MUL|BPF_K:
1017 case BPF_ALU|BPF_DIV|BPF_K:
1018 case BPF_ALU|BPF_AND|BPF_K:
1019 case BPF_ALU|BPF_OR|BPF_K:
1020 case BPF_ALU|BPF_LSH|BPF_K:
1021 case BPF_ALU|BPF_RSH|BPF_K:
1022 op = BPF_OP(s->code);
1023 if (alter) {
1024 if (s->k == 0) {
1025 /* don't optimize away "sub #0"
1026 * as it may be needed later to
1027 * fixup the generated math code */
1028 if (op == BPF_ADD ||
1029 op == BPF_LSH || op == BPF_RSH ||
1030 op == BPF_OR) {
1031 s->code = NOP;
1032 break;
1033 }
1034 if (op == BPF_MUL || op == BPF_AND) {
1035 s->code = BPF_LD|BPF_IMM;
1036 val[A_ATOM] = K(s->k);
1037 break;
1038 }
1039 }
1040 if (vmap[val[A_ATOM]].is_const) {
1041 fold_op(s, val[A_ATOM], K(s->k));
1042 val[A_ATOM] = K(s->k);
1043 break;
1044 }
1045 }
1046 val[A_ATOM] = F(s->code, val[A_ATOM], K(s->k));
1047 break;
1048
1049 case BPF_ALU|BPF_ADD|BPF_X:
1050 case BPF_ALU|BPF_SUB|BPF_X:
1051 case BPF_ALU|BPF_MUL|BPF_X:
1052 case BPF_ALU|BPF_DIV|BPF_X:
1053 case BPF_ALU|BPF_AND|BPF_X:
1054 case BPF_ALU|BPF_OR|BPF_X:
1055 case BPF_ALU|BPF_LSH|BPF_X:
1056 case BPF_ALU|BPF_RSH|BPF_X:
1057 op = BPF_OP(s->code);
1058 if (alter && vmap[val[X_ATOM]].is_const) {
1059 if (vmap[val[A_ATOM]].is_const) {
1060 fold_op(s, val[A_ATOM], val[X_ATOM]);
1061 val[A_ATOM] = K(s->k);
1062 }
1063 else {
1064 s->code = BPF_ALU|BPF_K|op;
1065 s->k = vmap[val[X_ATOM]].const_val;
1066 done = 0;
1067 val[A_ATOM] =
1068 F(s->code, val[A_ATOM], K(s->k));
1069 }
1070 break;
1071 }
1072 /*
1073 * Check if we're doing something to an accumulator
1074 * that is 0, and simplify. This may not seem like
1075 * much of a simplification but it could open up further
1076 * optimizations.
1077 * XXX We could also check for mul by 1, etc.
1078 */
1079 if (alter && vmap[val[A_ATOM]].is_const
1080 && vmap[val[A_ATOM]].const_val == 0) {
1081 if (op == BPF_ADD || op == BPF_OR) {
1082 s->code = BPF_MISC|BPF_TXA;
1083 vstore(s, &val[A_ATOM], val[X_ATOM], alter);
1084 break;
1085 }
1086 else if (op == BPF_MUL || op == BPF_DIV ||
1087 op == BPF_AND || op == BPF_LSH || op == BPF_RSH) {
1088 s->code = BPF_LD|BPF_IMM;
1089 s->k = 0;
1090 vstore(s, &val[A_ATOM], K(s->k), alter);
1091 break;
1092 }
1093 else if (op == BPF_NEG) {
1094 s->code = NOP;
1095 break;
1096 }
1097 }
1098 val[A_ATOM] = F(s->code, val[A_ATOM], val[X_ATOM]);
1099 break;
1100
1101 case BPF_MISC|BPF_TXA:
1102 vstore(s, &val[A_ATOM], val[X_ATOM], alter);
1103 break;
1104
1105 case BPF_LD|BPF_MEM:
1106 v = val[s->k];
1107 if (alter && vmap[v].is_const) {
1108 s->code = BPF_LD|BPF_IMM;
1109 s->k = vmap[v].const_val;
1110 done = 0;
1111 }
1112 vstore(s, &val[A_ATOM], v, alter);
1113 break;
1114
1115 case BPF_MISC|BPF_TAX:
1116 vstore(s, &val[X_ATOM], val[A_ATOM], alter);
1117 break;
1118
1119 case BPF_LDX|BPF_MEM:
1120 v = val[s->k];
1121 if (alter && vmap[v].is_const) {
1122 s->code = BPF_LDX|BPF_IMM;
1123 s->k = vmap[v].const_val;
1124 done = 0;
1125 }
1126 vstore(s, &val[X_ATOM], v, alter);
1127 break;
1128
1129 case BPF_ST:
1130 vstore(s, &val[s->k], val[A_ATOM], alter);
1131 break;
1132
1133 case BPF_STX:
1134 vstore(s, &val[s->k], val[X_ATOM], alter);
1135 break;
1136 }
1137 }
1138
1139 static void
deadstmt(s,last)1140 deadstmt(s, last)
1141 register struct stmt *s;
1142 register struct stmt *last[];
1143 {
1144 register int atom;
1145
1146 atom = atomuse(s);
1147 if (atom >= 0) {
1148 if (atom == AX_ATOM) {
1149 last[X_ATOM] = 0;
1150 last[A_ATOM] = 0;
1151 }
1152 else
1153 last[atom] = 0;
1154 }
1155 atom = atomdef(s);
1156 if (atom >= 0) {
1157 if (last[atom]) {
1158 done = 0;
1159 last[atom]->code = NOP;
1160 }
1161 last[atom] = s;
1162 }
1163 }
1164
1165 static void
opt_deadstores(b)1166 opt_deadstores(b)
1167 register struct block *b;
1168 {
1169 register struct slist *s;
1170 register int atom;
1171 struct stmt *last[N_ATOMS];
1172
1173 memset((char *)last, 0, sizeof last);
1174
1175 for (s = b->stmts; s != 0; s = s->next)
1176 deadstmt(&s->s, last);
1177 deadstmt(&b->s, last);
1178
1179 for (atom = 0; atom < N_ATOMS; ++atom)
1180 if (last[atom] && !ATOMELEM(b->out_use, atom)) {
1181 last[atom]->code = NOP;
1182 done = 0;
1183 }
1184 }
1185
1186 static void
opt_blk(b,do_stmts)1187 opt_blk(b, do_stmts)
1188 struct block *b;
1189 int do_stmts;
1190 {
1191 struct slist *s;
1192 struct edge *p;
1193 int i;
1194 bpf_int32 aval, xval;
1195
1196 #if 0
1197 for (s = b->stmts; s && s->next; s = s->next)
1198 if (BPF_CLASS(s->s.code) == BPF_JMP) {
1199 do_stmts = 0;
1200 break;
1201 }
1202 #endif
1203
1204 /*
1205 * Initialize the atom values.
1206 */
1207 p = b->in_edges;
1208 if (p == 0) {
1209 /*
1210 * We have no predecessors, so everything is undefined
1211 * upon entry to this block.
1212 */
1213 memset((char *)b->val, 0, sizeof(b->val));
1214 } else {
1215 /*
1216 * Inherit values from our predecessors.
1217 *
1218 * First, get the values from the predecessor along the
1219 * first edge leading to this node.
1220 */
1221 memcpy((char *)b->val, (char *)p->pred->val, sizeof(b->val));
1222 /*
1223 * Now look at all the other nodes leading to this node.
1224 * If, for the predecessor along that edge, a register
1225 * has a different value from the one we have (i.e.,
1226 * control paths are merging, and the merging paths
1227 * assign different values to that register), give the
1228 * register the undefined value of 0.
1229 */
1230 while ((p = p->next) != NULL) {
1231 for (i = 0; i < N_ATOMS; ++i)
1232 if (b->val[i] != p->pred->val[i])
1233 b->val[i] = 0;
1234 }
1235 }
1236 aval = b->val[A_ATOM];
1237 xval = b->val[X_ATOM];
1238 for (s = b->stmts; s; s = s->next)
1239 opt_stmt(&s->s, b->val, do_stmts);
1240
1241 /*
1242 * This is a special case: if we don't use anything from this
1243 * block, and we load the accumulator or index register with a
1244 * value that is already there, or if this block is a return,
1245 * eliminate all the statements.
1246 *
1247 * XXX - what if it does a store?
1248 *
1249 * XXX - why does it matter whether we use anything from this
1250 * block? If the accumulator or index register doesn't change
1251 * its value, isn't that OK even if we use that value?
1252 *
1253 * XXX - if we load the accumulator with a different value,
1254 * and the block ends with a conditional branch, we obviously
1255 * can't eliminate it, as the branch depends on that value.
1256 * For the index register, the conditional branch only depends
1257 * on the index register value if the test is against the index
1258 * register value rather than a constant; if nothing uses the
1259 * value we put into the index register, and we're not testing
1260 * against the index register's value, and there aren't any
1261 * other problems that would keep us from eliminating this
1262 * block, can we eliminate it?
1263 */
1264 if (do_stmts &&
1265 ((b->out_use == 0 && aval != 0 && b->val[A_ATOM] == aval &&
1266 xval != 0 && b->val[X_ATOM] == xval) ||
1267 BPF_CLASS(b->s.code) == BPF_RET)) {
1268 if (b->stmts != 0) {
1269 b->stmts = 0;
1270 done = 0;
1271 }
1272 } else {
1273 opt_peep(b);
1274 opt_deadstores(b);
1275 }
1276 /*
1277 * Set up values for branch optimizer.
1278 */
1279 if (BPF_SRC(b->s.code) == BPF_K)
1280 b->oval = K(b->s.k);
1281 else
1282 b->oval = b->val[X_ATOM];
1283 b->et.code = b->s.code;
1284 b->ef.code = -b->s.code;
1285 }
1286
1287 /*
1288 * Return true if any register that is used on exit from 'succ', has
1289 * an exit value that is different from the corresponding exit value
1290 * from 'b'.
1291 */
1292 static int
use_conflict(b,succ)1293 use_conflict(b, succ)
1294 struct block *b, *succ;
1295 {
1296 int atom;
1297 atomset use = succ->out_use;
1298
1299 if (use == 0)
1300 return 0;
1301
1302 for (atom = 0; atom < N_ATOMS; ++atom)
1303 if (ATOMELEM(use, atom))
1304 if (b->val[atom] != succ->val[atom])
1305 return 1;
1306 return 0;
1307 }
1308
1309 static struct block *
fold_edge(child,ep)1310 fold_edge(child, ep)
1311 struct block *child;
1312 struct edge *ep;
1313 {
1314 int sense;
1315 int aval0, aval1, oval0, oval1;
1316 int code = ep->code;
1317
1318 if (code < 0) {
1319 code = -code;
1320 sense = 0;
1321 } else
1322 sense = 1;
1323
1324 if (child->s.code != code)
1325 return 0;
1326
1327 aval0 = child->val[A_ATOM];
1328 oval0 = child->oval;
1329 aval1 = ep->pred->val[A_ATOM];
1330 oval1 = ep->pred->oval;
1331
1332 if (aval0 != aval1)
1333 return 0;
1334
1335 if (oval0 == oval1)
1336 /*
1337 * The operands of the branch instructions are
1338 * identical, so the result is true if a true
1339 * branch was taken to get here, otherwise false.
1340 */
1341 return sense ? JT(child) : JF(child);
1342
1343 if (sense && code == (BPF_JMP|BPF_JEQ|BPF_K))
1344 /*
1345 * At this point, we only know the comparison if we
1346 * came down the true branch, and it was an equality
1347 * comparison with a constant.
1348 *
1349 * I.e., if we came down the true branch, and the branch
1350 * was an equality comparison with a constant, we know the
1351 * accumulator contains that constant. If we came down
1352 * the false branch, or the comparison wasn't with a
1353 * constant, we don't know what was in the accumulator.
1354 *
1355 * We rely on the fact that distinct constants have distinct
1356 * value numbers.
1357 */
1358 return JF(child);
1359
1360 return 0;
1361 }
1362
1363 #include "ffs.h"
1364 static void
opt_j(ep)1365 opt_j(ep)
1366 struct edge *ep;
1367 {
1368 register int i, k;
1369 register struct block *target;
1370
1371 if (JT(ep->succ) == 0)
1372 return;
1373
1374 if (JT(ep->succ) == JF(ep->succ)) {
1375 /*
1376 * Common branch targets can be eliminated, provided
1377 * there is no data dependency.
1378 */
1379 if (!use_conflict(ep->pred, ep->succ->et.succ)) {
1380 done = 0;
1381 ep->succ = JT(ep->succ);
1382 }
1383 }
1384 /*
1385 * For each edge dominator that matches the successor of this
1386 * edge, promote the edge successor to the its grandchild.
1387 *
1388 * XXX We violate the set abstraction here in favor a reasonably
1389 * efficient loop.
1390 */
1391 top:
1392 for (i = 0; i < edgewords; ++i) {
1393 register bpf_u_int32 x = ep->edom[i];
1394
1395 while (x != 0) {
1396 k = ffs(x) - 1;
1397 x &=~ (1 << k);
1398 k += i * BITS_PER_WORD;
1399
1400 target = fold_edge(ep->succ, edges[k]);
1401 /*
1402 * Check that there is no data dependency between
1403 * nodes that will be violated if we move the edge.
1404 */
1405 if (target != 0 && !use_conflict(ep->pred, target)) {
1406 done = 0;
1407 ep->succ = target;
1408 if (JT(target) != 0)
1409 /*
1410 * Start over unless we hit a leaf.
1411 */
1412 goto top;
1413 return;
1414 }
1415 }
1416 }
1417 }
1418
1419
1420 static void
or_pullup(b)1421 or_pullup(b)
1422 struct block *b;
1423 {
1424 int val, at_top;
1425 struct block *pull;
1426 struct block **diffp, **samep;
1427 struct edge *ep;
1428
1429 ep = b->in_edges;
1430 if (ep == 0)
1431 return;
1432
1433 /*
1434 * Make sure each predecessor loads the same value.
1435 * XXX why?
1436 */
1437 val = ep->pred->val[A_ATOM];
1438 for (ep = ep->next; ep != 0; ep = ep->next)
1439 if (val != ep->pred->val[A_ATOM])
1440 return;
1441
1442 if (JT(b->in_edges->pred) == b)
1443 diffp = &JT(b->in_edges->pred);
1444 else
1445 diffp = &JF(b->in_edges->pred);
1446
1447 at_top = 1;
1448 while (1) {
1449 if (*diffp == 0)
1450 return;
1451
1452 if (JT(*diffp) != JT(b))
1453 return;
1454
1455 if (!SET_MEMBER((*diffp)->dom, b->id))
1456 return;
1457
1458 if ((*diffp)->val[A_ATOM] != val)
1459 break;
1460
1461 diffp = &JF(*diffp);
1462 at_top = 0;
1463 }
1464 samep = &JF(*diffp);
1465 while (1) {
1466 if (*samep == 0)
1467 return;
1468
1469 if (JT(*samep) != JT(b))
1470 return;
1471
1472 if (!SET_MEMBER((*samep)->dom, b->id))
1473 return;
1474
1475 if ((*samep)->val[A_ATOM] == val)
1476 break;
1477
1478 /* XXX Need to check that there are no data dependencies
1479 between dp0 and dp1. Currently, the code generator
1480 will not produce such dependencies. */
1481 samep = &JF(*samep);
1482 }
1483 #ifdef notdef
1484 /* XXX This doesn't cover everything. */
1485 for (i = 0; i < N_ATOMS; ++i)
1486 if ((*samep)->val[i] != pred->val[i])
1487 return;
1488 #endif
1489 /* Pull up the node. */
1490 pull = *samep;
1491 *samep = JF(pull);
1492 JF(pull) = *diffp;
1493
1494 /*
1495 * At the top of the chain, each predecessor needs to point at the
1496 * pulled up node. Inside the chain, there is only one predecessor
1497 * to worry about.
1498 */
1499 if (at_top) {
1500 for (ep = b->in_edges; ep != 0; ep = ep->next) {
1501 if (JT(ep->pred) == b)
1502 JT(ep->pred) = pull;
1503 else
1504 JF(ep->pred) = pull;
1505 }
1506 }
1507 else
1508 *diffp = pull;
1509
1510 done = 0;
1511 }
1512
1513 static void
and_pullup(b)1514 and_pullup(b)
1515 struct block *b;
1516 {
1517 int val, at_top;
1518 struct block *pull;
1519 struct block **diffp, **samep;
1520 struct edge *ep;
1521
1522 ep = b->in_edges;
1523 if (ep == 0)
1524 return;
1525
1526 /*
1527 * Make sure each predecessor loads the same value.
1528 */
1529 val = ep->pred->val[A_ATOM];
1530 for (ep = ep->next; ep != 0; ep = ep->next)
1531 if (val != ep->pred->val[A_ATOM])
1532 return;
1533
1534 if (JT(b->in_edges->pred) == b)
1535 diffp = &JT(b->in_edges->pred);
1536 else
1537 diffp = &JF(b->in_edges->pred);
1538
1539 at_top = 1;
1540 while (1) {
1541 if (*diffp == 0)
1542 return;
1543
1544 if (JF(*diffp) != JF(b))
1545 return;
1546
1547 if (!SET_MEMBER((*diffp)->dom, b->id))
1548 return;
1549
1550 if ((*diffp)->val[A_ATOM] != val)
1551 break;
1552
1553 diffp = &JT(*diffp);
1554 at_top = 0;
1555 }
1556 samep = &JT(*diffp);
1557 while (1) {
1558 if (*samep == 0)
1559 return;
1560
1561 if (JF(*samep) != JF(b))
1562 return;
1563
1564 if (!SET_MEMBER((*samep)->dom, b->id))
1565 return;
1566
1567 if ((*samep)->val[A_ATOM] == val)
1568 break;
1569
1570 /* XXX Need to check that there are no data dependencies
1571 between diffp and samep. Currently, the code generator
1572 will not produce such dependencies. */
1573 samep = &JT(*samep);
1574 }
1575 #ifdef notdef
1576 /* XXX This doesn't cover everything. */
1577 for (i = 0; i < N_ATOMS; ++i)
1578 if ((*samep)->val[i] != pred->val[i])
1579 return;
1580 #endif
1581 /* Pull up the node. */
1582 pull = *samep;
1583 *samep = JT(pull);
1584 JT(pull) = *diffp;
1585
1586 /*
1587 * At the top of the chain, each predecessor needs to point at the
1588 * pulled up node. Inside the chain, there is only one predecessor
1589 * to worry about.
1590 */
1591 if (at_top) {
1592 for (ep = b->in_edges; ep != 0; ep = ep->next) {
1593 if (JT(ep->pred) == b)
1594 JT(ep->pred) = pull;
1595 else
1596 JF(ep->pred) = pull;
1597 }
1598 }
1599 else
1600 *diffp = pull;
1601
1602 done = 0;
1603 }
1604
1605 static void
opt_blks(root,do_stmts)1606 opt_blks(root, do_stmts)
1607 struct block *root;
1608 int do_stmts;
1609 {
1610 int i, maxlevel;
1611 struct block *p;
1612
1613 init_val();
1614 maxlevel = root->level;
1615
1616 find_inedges(root);
1617 for (i = maxlevel; i >= 0; --i)
1618 for (p = levels[i]; p; p = p->link)
1619 opt_blk(p, do_stmts);
1620
1621 if (do_stmts)
1622 /*
1623 * No point trying to move branches; it can't possibly
1624 * make a difference at this point.
1625 */
1626 return;
1627
1628 for (i = 1; i <= maxlevel; ++i) {
1629 for (p = levels[i]; p; p = p->link) {
1630 opt_j(&p->et);
1631 opt_j(&p->ef);
1632 }
1633 }
1634
1635 find_inedges(root);
1636 for (i = 1; i <= maxlevel; ++i) {
1637 for (p = levels[i]; p; p = p->link) {
1638 or_pullup(p);
1639 and_pullup(p);
1640 }
1641 }
1642 }
1643
1644 static inline void
link_inedge(parent,child)1645 link_inedge(parent, child)
1646 struct edge *parent;
1647 struct block *child;
1648 {
1649 parent->next = child->in_edges;
1650 child->in_edges = parent;
1651 }
1652
1653 static void
find_inedges(root)1654 find_inedges(root)
1655 struct block *root;
1656 {
1657 int i;
1658 struct block *b;
1659
1660 for (i = 0; i < n_blocks; ++i)
1661 blocks[i]->in_edges = 0;
1662
1663 /*
1664 * Traverse the graph, adding each edge to the predecessor
1665 * list of its successors. Skip the leaves (i.e. level 0).
1666 */
1667 for (i = root->level; i > 0; --i) {
1668 for (b = levels[i]; b != 0; b = b->link) {
1669 link_inedge(&b->et, JT(b));
1670 link_inedge(&b->ef, JF(b));
1671 }
1672 }
1673 }
1674
1675 static void
opt_root(b)1676 opt_root(b)
1677 struct block **b;
1678 {
1679 struct slist *tmp, *s;
1680
1681 s = (*b)->stmts;
1682 (*b)->stmts = 0;
1683 while (BPF_CLASS((*b)->s.code) == BPF_JMP && JT(*b) == JF(*b))
1684 *b = JT(*b);
1685
1686 tmp = (*b)->stmts;
1687 if (tmp != 0)
1688 sappend(s, tmp);
1689 (*b)->stmts = s;
1690
1691 /*
1692 * If the root node is a return, then there is no
1693 * point executing any statements (since the bpf machine
1694 * has no side effects).
1695 */
1696 if (BPF_CLASS((*b)->s.code) == BPF_RET)
1697 (*b)->stmts = 0;
1698 }
1699
1700 static void
opt_loop(root,do_stmts)1701 opt_loop(root, do_stmts)
1702 struct block *root;
1703 int do_stmts;
1704 {
1705
1706 #ifdef BDEBUG
1707 if (dflag > 1) {
1708 printf("opt_loop(root, %d) begin\n", do_stmts);
1709 opt_dump(root);
1710 }
1711 #endif
1712 do {
1713 done = 1;
1714 find_levels(root);
1715 find_dom(root);
1716 find_closure(root);
1717 find_ud(root);
1718 find_edom(root);
1719 opt_blks(root, do_stmts);
1720 #ifdef BDEBUG
1721 if (dflag > 1) {
1722 printf("opt_loop(root, %d) bottom, done=%d\n", do_stmts, done);
1723 opt_dump(root);
1724 }
1725 #endif
1726 } while (!done);
1727 }
1728
1729 /*
1730 * Optimize the filter code in its dag representation.
1731 */
1732 void
bpf_optimize(rootp)1733 bpf_optimize(rootp)
1734 struct block **rootp;
1735 {
1736 struct block *root;
1737
1738 root = *rootp;
1739
1740 opt_init(root);
1741 opt_loop(root, 0);
1742 opt_loop(root, 1);
1743 intern_blocks(root);
1744 #ifdef BDEBUG
1745 if (dflag > 1) {
1746 printf("after intern_blocks()\n");
1747 opt_dump(root);
1748 }
1749 #endif
1750 opt_root(rootp);
1751 #ifdef BDEBUG
1752 if (dflag > 1) {
1753 printf("after opt_root()\n");
1754 opt_dump(root);
1755 }
1756 #endif
1757 opt_cleanup();
1758 }
1759
1760 static void
make_marks(p)1761 make_marks(p)
1762 struct block *p;
1763 {
1764 if (!isMarked(p)) {
1765 Mark(p);
1766 if (BPF_CLASS(p->s.code) != BPF_RET) {
1767 make_marks(JT(p));
1768 make_marks(JF(p));
1769 }
1770 }
1771 }
1772
1773 /*
1774 * Mark code array such that isMarked(i) is true
1775 * only for nodes that are alive.
1776 */
1777 static void
mark_code(p)1778 mark_code(p)
1779 struct block *p;
1780 {
1781 cur_mark += 1;
1782 make_marks(p);
1783 }
1784
1785 /*
1786 * True iff the two stmt lists load the same value from the packet into
1787 * the accumulator.
1788 */
1789 static int
eq_slist(x,y)1790 eq_slist(x, y)
1791 struct slist *x, *y;
1792 {
1793 while (1) {
1794 while (x && x->s.code == NOP)
1795 x = x->next;
1796 while (y && y->s.code == NOP)
1797 y = y->next;
1798 if (x == 0)
1799 return y == 0;
1800 if (y == 0)
1801 return x == 0;
1802 if (x->s.code != y->s.code || x->s.k != y->s.k)
1803 return 0;
1804 x = x->next;
1805 y = y->next;
1806 }
1807 }
1808
1809 static inline int
eq_blk(b0,b1)1810 eq_blk(b0, b1)
1811 struct block *b0, *b1;
1812 {
1813 if (b0->s.code == b1->s.code &&
1814 b0->s.k == b1->s.k &&
1815 b0->et.succ == b1->et.succ &&
1816 b0->ef.succ == b1->ef.succ)
1817 return eq_slist(b0->stmts, b1->stmts);
1818 return 0;
1819 }
1820
1821 static void
intern_blocks(root)1822 intern_blocks(root)
1823 struct block *root;
1824 {
1825 struct block *p;
1826 int i, j;
1827 int done1; /* don't shadow global */
1828 top:
1829 done1 = 1;
1830 for (i = 0; i < n_blocks; ++i)
1831 blocks[i]->link = 0;
1832
1833 mark_code(root);
1834
1835 for (i = n_blocks - 1; --i >= 0; ) {
1836 if (!isMarked(blocks[i]))
1837 continue;
1838 for (j = i + 1; j < n_blocks; ++j) {
1839 if (!isMarked(blocks[j]))
1840 continue;
1841 if (eq_blk(blocks[i], blocks[j])) {
1842 blocks[i]->link = blocks[j]->link ?
1843 blocks[j]->link : blocks[j];
1844 break;
1845 }
1846 }
1847 }
1848 for (i = 0; i < n_blocks; ++i) {
1849 p = blocks[i];
1850 if (JT(p) == 0)
1851 continue;
1852 if (JT(p)->link) {
1853 done1 = 0;
1854 JT(p) = JT(p)->link;
1855 }
1856 if (JF(p)->link) {
1857 done1 = 0;
1858 JF(p) = JF(p)->link;
1859 }
1860 }
1861 if (!done1)
1862 goto top;
1863 }
1864
1865 static void
opt_cleanup()1866 opt_cleanup()
1867 {
1868 free((void *)vnode_base);
1869 free((void *)vmap);
1870 free((void *)edges);
1871 free((void *)space);
1872 free((void *)levels);
1873 free((void *)blocks);
1874 }
1875
1876 /*
1877 * Return the number of stmts in 's'.
1878 */
1879 static int
slength(s)1880 slength(s)
1881 struct slist *s;
1882 {
1883 int n = 0;
1884
1885 for (; s; s = s->next)
1886 if (s->s.code != NOP)
1887 ++n;
1888 return n;
1889 }
1890
1891 /*
1892 * Return the number of nodes reachable by 'p'.
1893 * All nodes should be initially unmarked.
1894 */
1895 static int
count_blocks(p)1896 count_blocks(p)
1897 struct block *p;
1898 {
1899 if (p == 0 || isMarked(p))
1900 return 0;
1901 Mark(p);
1902 return count_blocks(JT(p)) + count_blocks(JF(p)) + 1;
1903 }
1904
1905 /*
1906 * Do a depth first search on the flow graph, numbering the
1907 * the basic blocks, and entering them into the 'blocks' array.`
1908 */
1909 static void
number_blks_r(p)1910 number_blks_r(p)
1911 struct block *p;
1912 {
1913 int n;
1914
1915 if (p == 0 || isMarked(p))
1916 return;
1917
1918 Mark(p);
1919 n = n_blocks++;
1920 p->id = n;
1921 blocks[n] = p;
1922
1923 number_blks_r(JT(p));
1924 number_blks_r(JF(p));
1925 }
1926
1927 /*
1928 * Return the number of stmts in the flowgraph reachable by 'p'.
1929 * The nodes should be unmarked before calling.
1930 *
1931 * Note that "stmts" means "instructions", and that this includes
1932 *
1933 * side-effect statements in 'p' (slength(p->stmts));
1934 *
1935 * statements in the true branch from 'p' (count_stmts(JT(p)));
1936 *
1937 * statements in the false branch from 'p' (count_stmts(JF(p)));
1938 *
1939 * the conditional jump itself (1);
1940 *
1941 * an extra long jump if the true branch requires it (p->longjt);
1942 *
1943 * an extra long jump if the false branch requires it (p->longjf).
1944 */
1945 static int
count_stmts(p)1946 count_stmts(p)
1947 struct block *p;
1948 {
1949 int n;
1950
1951 if (p == 0 || isMarked(p))
1952 return 0;
1953 Mark(p);
1954 n = count_stmts(JT(p)) + count_stmts(JF(p));
1955 return slength(p->stmts) + n + 1 + p->longjt + p->longjf;
1956 }
1957
1958 /*
1959 * Allocate memory. All allocation is done before optimization
1960 * is begun. A linear bound on the size of all data structures is computed
1961 * from the total number of blocks and/or statements.
1962 */
1963 static void
opt_init(root)1964 opt_init(root)
1965 struct block *root;
1966 {
1967 bpf_u_int32 *p;
1968 int i, n, max_stmts;
1969
1970 /*
1971 * First, count the blocks, so we can malloc an array to map
1972 * block number to block. Then, put the blocks into the array.
1973 */
1974 unMarkAll();
1975 n = count_blocks(root);
1976 blocks = (struct block **)calloc(n, sizeof(*blocks));
1977 if (blocks == NULL)
1978 bpf_error("malloc");
1979 unMarkAll();
1980 n_blocks = 0;
1981 number_blks_r(root);
1982
1983 n_edges = 2 * n_blocks;
1984 edges = (struct edge **)calloc(n_edges, sizeof(*edges));
1985 if (edges == NULL)
1986 bpf_error("malloc");
1987
1988 /*
1989 * The number of levels is bounded by the number of nodes.
1990 */
1991 levels = (struct block **)calloc(n_blocks, sizeof(*levels));
1992 if (levels == NULL)
1993 bpf_error("malloc");
1994
1995 edgewords = n_edges / (8 * sizeof(bpf_u_int32)) + 1;
1996 nodewords = n_blocks / (8 * sizeof(bpf_u_int32)) + 1;
1997
1998 /* XXX */
1999 space = (bpf_u_int32 *)malloc(2 * n_blocks * nodewords * sizeof(*space)
2000 + n_edges * edgewords * sizeof(*space));
2001 if (space == NULL)
2002 bpf_error("malloc");
2003 p = space;
2004 all_dom_sets = p;
2005 for (i = 0; i < n; ++i) {
2006 blocks[i]->dom = p;
2007 p += nodewords;
2008 }
2009 all_closure_sets = p;
2010 for (i = 0; i < n; ++i) {
2011 blocks[i]->closure = p;
2012 p += nodewords;
2013 }
2014 all_edge_sets = p;
2015 for (i = 0; i < n; ++i) {
2016 register struct block *b = blocks[i];
2017
2018 b->et.edom = p;
2019 p += edgewords;
2020 b->ef.edom = p;
2021 p += edgewords;
2022 b->et.id = i;
2023 edges[i] = &b->et;
2024 b->ef.id = n_blocks + i;
2025 edges[n_blocks + i] = &b->ef;
2026 b->et.pred = b;
2027 b->ef.pred = b;
2028 }
2029 max_stmts = 0;
2030 for (i = 0; i < n; ++i)
2031 max_stmts += slength(blocks[i]->stmts) + 1;
2032 /*
2033 * We allocate at most 3 value numbers per statement,
2034 * so this is an upper bound on the number of valnodes
2035 * we'll need.
2036 */
2037 maxval = 3 * max_stmts;
2038 vmap = (struct vmapinfo *)calloc(maxval, sizeof(*vmap));
2039 vnode_base = (struct valnode *)calloc(maxval, sizeof(*vnode_base));
2040 if (vmap == NULL || vnode_base == NULL)
2041 bpf_error("malloc");
2042 }
2043
2044 /*
2045 * Some pointers used to convert the basic block form of the code,
2046 * into the array form that BPF requires. 'fstart' will point to
2047 * the malloc'd array while 'ftail' is used during the recursive traversal.
2048 */
2049 static struct bpf_insn *fstart;
2050 static struct bpf_insn *ftail;
2051
2052 #ifdef BDEBUG
2053 int bids[1000];
2054 #endif
2055
2056 /*
2057 * Returns true if successful. Returns false if a branch has
2058 * an offset that is too large. If so, we have marked that
2059 * branch so that on a subsequent iteration, it will be treated
2060 * properly.
2061 */
2062 static int
convert_code_r(p)2063 convert_code_r(p)
2064 struct block *p;
2065 {
2066 struct bpf_insn *dst;
2067 struct slist *src;
2068 int slen;
2069 u_int off;
2070 int extrajmps; /* number of extra jumps inserted */
2071 struct slist **offset = NULL;
2072
2073 if (p == 0 || isMarked(p))
2074 return (1);
2075 Mark(p);
2076
2077 if (convert_code_r(JF(p)) == 0)
2078 return (0);
2079 if (convert_code_r(JT(p)) == 0)
2080 return (0);
2081
2082 slen = slength(p->stmts);
2083 dst = ftail -= (slen + 1 + p->longjt + p->longjf);
2084 /* inflate length by any extra jumps */
2085
2086 p->offset = dst - fstart;
2087
2088 /* generate offset[] for convenience */
2089 if (slen) {
2090 offset = (struct slist **)calloc(slen, sizeof(struct slist *));
2091 if (!offset) {
2092 bpf_error("not enough core");
2093 /*NOTREACHED*/
2094 }
2095 }
2096 src = p->stmts;
2097 for (off = 0; off < slen && src; off++) {
2098 #if 0
2099 printf("off=%d src=%x\n", off, src);
2100 #endif
2101 offset[off] = src;
2102 src = src->next;
2103 }
2104
2105 off = 0;
2106 for (src = p->stmts; src; src = src->next) {
2107 if (src->s.code == NOP)
2108 continue;
2109 dst->code = (u_short)src->s.code;
2110 dst->k = src->s.k;
2111
2112 /* fill block-local relative jump */
2113 if (BPF_CLASS(src->s.code) != BPF_JMP || src->s.code == (BPF_JMP|BPF_JA)) {
2114 #if 0
2115 if (src->s.jt || src->s.jf) {
2116 bpf_error("illegal jmp destination");
2117 /*NOTREACHED*/
2118 }
2119 #endif
2120 goto filled;
2121 }
2122 if (off == slen - 2) /*???*/
2123 goto filled;
2124
2125 {
2126 int i;
2127 int jt, jf;
2128 const char *ljerr = "%s for block-local relative jump: off=%d";
2129
2130 #if 0
2131 printf("code=%x off=%d %x %x\n", src->s.code,
2132 off, src->s.jt, src->s.jf);
2133 #endif
2134
2135 if (!src->s.jt || !src->s.jf) {
2136 bpf_error(ljerr, "no jmp destination", off);
2137 /*NOTREACHED*/
2138 }
2139
2140 jt = jf = 0;
2141 for (i = 0; i < slen; i++) {
2142 if (offset[i] == src->s.jt) {
2143 if (jt) {
2144 bpf_error(ljerr, "multiple matches", off);
2145 /*NOTREACHED*/
2146 }
2147
2148 dst->jt = i - off - 1;
2149 jt++;
2150 }
2151 if (offset[i] == src->s.jf) {
2152 if (jf) {
2153 bpf_error(ljerr, "multiple matches", off);
2154 /*NOTREACHED*/
2155 }
2156 dst->jf = i - off - 1;
2157 jf++;
2158 }
2159 }
2160 if (!jt || !jf) {
2161 bpf_error(ljerr, "no destination found", off);
2162 /*NOTREACHED*/
2163 }
2164 }
2165 filled:
2166 ++dst;
2167 ++off;
2168 }
2169 if (offset)
2170 free(offset);
2171
2172 #ifdef BDEBUG
2173 bids[dst - fstart] = p->id + 1;
2174 #endif
2175 dst->code = (u_short)p->s.code;
2176 dst->k = p->s.k;
2177 if (JT(p)) {
2178 extrajmps = 0;
2179 off = JT(p)->offset - (p->offset + slen) - 1;
2180 if (off >= 256) {
2181 /* offset too large for branch, must add a jump */
2182 if (p->longjt == 0) {
2183 /* mark this instruction and retry */
2184 p->longjt++;
2185 return(0);
2186 }
2187 /* branch if T to following jump */
2188 dst->jt = extrajmps;
2189 extrajmps++;
2190 dst[extrajmps].code = BPF_JMP|BPF_JA;
2191 dst[extrajmps].k = off - extrajmps;
2192 }
2193 else
2194 dst->jt = off;
2195 off = JF(p)->offset - (p->offset + slen) - 1;
2196 if (off >= 256) {
2197 /* offset too large for branch, must add a jump */
2198 if (p->longjf == 0) {
2199 /* mark this instruction and retry */
2200 p->longjf++;
2201 return(0);
2202 }
2203 /* branch if F to following jump */
2204 /* if two jumps are inserted, F goes to second one */
2205 dst->jf = extrajmps;
2206 extrajmps++;
2207 dst[extrajmps].code = BPF_JMP|BPF_JA;
2208 dst[extrajmps].k = off - extrajmps;
2209 }
2210 else
2211 dst->jf = off;
2212 }
2213 return (1);
2214 }
2215
2216
2217 /*
2218 * Convert flowgraph intermediate representation to the
2219 * BPF array representation. Set *lenp to the number of instructions.
2220 *
2221 * This routine does *NOT* leak the memory pointed to by fp. It *must
2222 * not* do free(fp) before returning fp; doing so would make no sense,
2223 * as the BPF array pointed to by the return value of icode_to_fcode()
2224 * must be valid - it's being returned for use in a bpf_program structure.
2225 *
2226 * If it appears that icode_to_fcode() is leaking, the problem is that
2227 * the program using pcap_compile() is failing to free the memory in
2228 * the BPF program when it's done - the leak is in the program, not in
2229 * the routine that happens to be allocating the memory. (By analogy, if
2230 * a program calls fopen() without ever calling fclose() on the FILE *,
2231 * it will leak the FILE structure; the leak is not in fopen(), it's in
2232 * the program.) Change the program to use pcap_freecode() when it's
2233 * done with the filter program. See the pcap man page.
2234 */
2235 struct bpf_insn *
icode_to_fcode(root,lenp)2236 icode_to_fcode(root, lenp)
2237 struct block *root;
2238 int *lenp;
2239 {
2240 int n;
2241 struct bpf_insn *fp;
2242
2243 /*
2244 * Loop doing convert_code_r() until no branches remain
2245 * with too-large offsets.
2246 */
2247 while (1) {
2248 unMarkAll();
2249 n = *lenp = count_stmts(root);
2250
2251 fp = (struct bpf_insn *)malloc(sizeof(*fp) * n);
2252 if (fp == NULL)
2253 bpf_error("malloc");
2254 memset((char *)fp, 0, sizeof(*fp) * n);
2255 fstart = fp;
2256 ftail = fp + n;
2257
2258 unMarkAll();
2259 if (convert_code_r(root))
2260 break;
2261 free(fp);
2262 }
2263
2264 return fp;
2265 }
2266
2267 /*
2268 * Make a copy of a BPF program and put it in the "fcode" member of
2269 * a "pcap_t".
2270 *
2271 * If we fail to allocate memory for the copy, fill in the "errbuf"
2272 * member of the "pcap_t" with an error message, and return -1;
2273 * otherwise, return 0.
2274 */
2275 int
install_bpf_program(pcap_t * p,struct bpf_program * fp)2276 install_bpf_program(pcap_t *p, struct bpf_program *fp)
2277 {
2278 size_t prog_size;
2279
2280 /*
2281 * Free up any already installed program.
2282 */
2283 pcap_freecode(&p->fcode);
2284
2285 prog_size = sizeof(*fp->bf_insns) * fp->bf_len;
2286 p->fcode.bf_len = fp->bf_len;
2287 p->fcode.bf_insns = (struct bpf_insn *)malloc(prog_size);
2288 if (p->fcode.bf_insns == NULL) {
2289 snprintf(p->errbuf, sizeof(p->errbuf),
2290 "malloc: %s", pcap_strerror(errno));
2291 return (-1);
2292 }
2293 memcpy(p->fcode.bf_insns, fp->bf_insns, prog_size);
2294 return (0);
2295 }
2296
2297 #ifdef BDEBUG
2298 static void
opt_dump(root)2299 opt_dump(root)
2300 struct block *root;
2301 {
2302 struct bpf_program f;
2303
2304 memset(bids, 0, sizeof bids);
2305 f.bf_insns = icode_to_fcode(root, &f.bf_len);
2306 bpf_dump(&f, 1);
2307 putchar('\n');
2308 free((char *)f.bf_insns);
2309 }
2310 #endif
2311