1 /* tblcmp - table compression routines */
4 * Copyright (c) 1990 The Regents of the University of California.
7 * This code is derived from software contributed to Berkeley by
10 * The United States Government has rights in this work pursuant
11 * to contract no. DE-AC03-76SF00098 between the United States
12 * Department of Energy and the University of California.
14 * Redistribution and use in source and binary forms are permitted provided
15 * that: (1) source distributions retain this entire copyright notice and
16 * comment, and (2) distributions including binaries display the following
17 * acknowledgement: ``This product includes software developed by the
18 * University of California, Berkeley and its contributors'' in the
19 * documentation or other materials provided with the distribution and in
20 * all advertising materials mentioning features or use of this software.
21 * Neither the name of the University nor the names of its contributors may
22 * be used to endorse or promote products derived from this software without
23 * specific prior written permission.
24 * THIS SOFTWARE IS PROVIDED ``AS IS'' AND WITHOUT ANY EXPRESS OR IMPLIED
25 * WARRANTIES, INCLUDING, WITHOUT LIMITATION, THE IMPLIED WARRANTIES OF
26 * MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE.
29 /* $Header: /home/daffy/u0/vern/flex/RCS/tblcmp.c,v 2.11 94/11/05 17:08:28 vern Exp $ */
30 /* $FreeBSD: src/usr.bin/lex/tblcmp.c,v 1.5 1999/10/27 07:56:47 obrien Exp $ */
31 /* $DragonFly: src/usr.bin/lex/tblcmp.c,v 1.4 2005/08/04 17:31:22 drhodus Exp $ */
36 /* declarations for functions that have forward references */
38 void mkentry
PROTO((int*, int, int, int, int));
39 void mkprot
PROTO((int[], int, int));
40 void mktemplate
PROTO((int[], int, int));
41 void mv2front
PROTO((int));
42 int tbldiff
PROTO((int[], int, int[]));
45 /* bldtbl - build table entries for dfa state
48 * int state[numecs], statenum, totaltrans, comstate, comfreq;
49 * bldtbl( state, statenum, totaltrans, comstate, comfreq );
51 * State is the statenum'th dfa state. It is indexed by equivalence class and
52 * gives the number of the state to enter for a given equivalence class.
53 * totaltrans is the total number of transitions out of the state. Comstate
54 * is that state which is the destination of the most transitions out of State.
55 * Comfreq is how many transitions there are out of State to Comstate.
57 * A note on terminology:
58 * "protos" are transition tables which have a high probability of
59 * either being redundant (a state processed later will have an identical
60 * transition table) or nearly redundant (a state processed later will have
61 * many of the same out-transitions). A "most recently used" queue of
62 * protos is kept around with the hope that most states will find a proto
63 * which is similar enough to be usable, and therefore compacting the
65 * "templates" are a special type of proto. If a transition table is
66 * homogeneous or nearly homogeneous (all transitions go to the same
67 * destination) then the odds are good that future states will also go
68 * to the same destination state on basically the same character set.
69 * These homogeneous states are so common when dealing with large rule
70 * sets that they merit special attention. If the transition table were
71 * simply made into a proto, then (typically) each subsequent, similar
72 * state will differ from the proto for two out-transitions. One of these
73 * out-transitions will be that character on which the proto does not go
74 * to the common destination, and one will be that character on which the
75 * state does not go to the common destination. Templates, on the other
76 * hand, go to the common state on EVERY transition character, and therefore
77 * cost only one difference.
80 void bldtbl(int *state
, int statenum
, int totaltrans
, int comstate
, int comfreq
)
82 int extptr
, extrct
[2][CSIZE
+ 1];
83 int mindiff
, minprot
, i
, d
;
85 /* If extptr is 0 then the first array of extrct holds the result
86 * of the "best difference" to date, which is those transitions
87 * which occur in "state" but not in the proto which, to date,
88 * has the fewest differences between itself and "state". If
89 * extptr is 1 then the second array of extrct hold the best
90 * difference. The two arrays are toggled between so that the
91 * best difference to date can be kept around and also a difference
92 * just created by checking against a candidate "best" proto.
97 /* If the state has too few out-transitions, don't bother trying to
101 if ( (totaltrans
* 100) < (numecs
* PROTO_SIZE_PERCENTAGE
) )
102 mkentry( state
, numecs
, statenum
, JAMSTATE
, totaltrans
);
106 /* "checkcom" is true if we should only check "state" against
107 * protos which have the same "comstate" value.
110 comfreq
* 100 > totaltrans
* CHECK_COM_PERCENTAGE
;
113 mindiff
= totaltrans
;
117 /* Find first proto which has the same "comstate". */
118 for ( i
= firstprot
; i
!= NIL
; i
= protnext
[i
] )
119 if ( protcomst
[i
] == comstate
)
122 mindiff
= tbldiff( state
, minprot
,
130 /* Since we've decided that the most common destination
131 * out of "state" does not occur with a high enough
132 * frequency, we set the "comstate" to zero, assuring
133 * that if this state is entered into the proto list,
134 * it will not be considered a template.
138 if ( firstprot
!= NIL
)
141 mindiff
= tbldiff( state
, minprot
,
146 /* We now have the first interesting proto in "minprot". If
147 * it matches within the tolerances set for the first proto,
148 * we don't want to bother scanning the rest of the proto list
149 * to see if we have any other reasonable matches.
152 if ( mindiff
* 100 > totaltrans
* FIRST_MATCH_DIFF_PERCENTAGE
)
154 /* Not a good enough match. Scan the rest of the
157 for ( i
= minprot
; i
!= NIL
; i
= protnext
[i
] )
159 d
= tbldiff( state
, i
, extrct
[1 - extptr
] );
169 /* Check if the proto we've decided on as our best bet is close
170 * enough to the state we want to match to be usable.
173 if ( mindiff
* 100 > totaltrans
* ACCEPTABLE_DIFF_PERCENTAGE
)
175 /* No good. If the state is homogeneous enough,
176 * we make a template out of it. Otherwise, we
180 if ( comfreq
* 100 >=
181 totaltrans
* TEMPLATE_SAME_PERCENTAGE
)
182 mktemplate( state
, statenum
, comstate
);
186 mkprot( state
, statenum
, comstate
);
187 mkentry( state
, numecs
, statenum
,
188 JAMSTATE
, totaltrans
);
193 { /* use the proto */
194 mkentry( extrct
[extptr
], numecs
, statenum
,
195 prottbl
[minprot
], mindiff
);
197 /* If this state was sufficiently different from the
198 * proto we built it from, make it, too, a proto.
201 if ( mindiff
* 100 >=
202 totaltrans
* NEW_PROTO_DIFF_PERCENTAGE
)
203 mkprot( state
, statenum
, comstate
);
205 /* Since mkprot added a new proto to the proto queue,
206 * it's possible that "minprot" is no longer on the
207 * proto queue (if it happened to have been the last
208 * entry, it would have been bumped off). If it's
209 * not there, then the new proto took its physical
210 * place (though logically the new proto is at the
211 * beginning of the queue), so in that case the
212 * following call will do nothing.
221 /* cmptmps - compress template table entries
223 * Template tables are compressed by using the 'template equivalence
224 * classes', which are collections of transition character equivalence
225 * classes which always appear together in templates - really meta-equivalence
231 int tmpstorage
[CSIZE
+ 1];
232 int *tmp
= tmpstorage
, i
, j
;
233 int totaltrans
, trans
;
235 peakpairs
= numtemps
* numecs
+ tblend
;
239 /* Create equivalence classes based on data gathered on
240 * template transitions.
242 nummecs
= cre8ecs( tecfwd
, tecbck
, numecs
);
248 while ( lastdfa
+ numtemps
+ 1 >= current_max_dfas
)
251 /* Loop through each template. */
253 for ( i
= 1; i
<= numtemps
; ++i
)
255 /* Number of non-jam transitions out of this template. */
258 for ( j
= 1; j
<= numecs
; ++j
)
260 trans
= tnxt
[numecs
* i
+ j
];
264 /* The absolute value of tecbck is the
265 * meta-equivalence class of a given
266 * equivalence class, as set up by cre8ecs().
270 tmp
[tecbck
[j
]] = trans
;
286 /* It is assumed (in a rather subtle way) in the skeleton
287 * that if we're using meta-equivalence classes, the def[]
288 * entry for all templates is the jam template, i.e.,
289 * templates never default to other non-jam table entries
290 * (e.g., another template)
293 /* Leave room for the jam-state after the last real state. */
294 mkentry( tmp
, nummecs
, lastdfa
+ i
+ 1, JAMSTATE
, totaltrans
);
300 /* expand_nxt_chk - expand the next check arrays */
302 void expand_nxt_chk(void)
304 int old_max
= current_max_xpairs
;
306 current_max_xpairs
+= MAX_XPAIRS_INCREMENT
;
310 nxt
= reallocate_integer_array( nxt
, current_max_xpairs
);
311 chk
= reallocate_integer_array( chk
, current_max_xpairs
);
313 zero_out( (char *) (chk
+ old_max
),
314 (size_t) (MAX_XPAIRS_INCREMENT
* sizeof( int )) );
318 /* find_table_space - finds a space in the table for a state to be placed
321 * int *state, numtrans, block_start;
322 * int find_table_space();
324 * block_start = find_table_space( state, numtrans );
326 * State is the state to be added to the full speed transition table.
327 * Numtrans is the number of out-transitions for the state.
329 * find_table_space() returns the position of the start of the first block (in
330 * chk) able to accommodate the state
332 * In determining if a state will or will not fit, find_table_space() must take
333 * into account the fact that an end-of-buffer state will be added at [0],
334 * and an action number will be added in [-1].
337 int find_table_space(int *state
, int numtrans
)
339 /* Firstfree is the position of the first possible occurrence of two
340 * consecutive unused records in the chk and nxt arrays.
343 int *state_ptr
, *chk_ptr
;
344 int *ptr_to_last_entry_in_state
;
346 /* If there are too many out-transitions, put the state at the end of
349 if ( numtrans
> MAX_XTIONS_FULL_INTERIOR_FIT
)
351 /* If table is empty, return the first available spot in
352 * chk/nxt, which should be 1.
357 /* Start searching for table space near the end of
364 /* Start searching for table space from the beginning
365 * (skipping only the elements which will definitely not
366 * hold the new state).
370 while ( 1 ) /* loops until a space is found */
372 while ( i
+ numecs
>= current_max_xpairs
)
375 /* Loops until space for end-of-buffer and action number
380 /* Check for action number space. */
381 if ( chk
[i
- 1] == 0 )
383 /* Check for end-of-buffer space. */
388 /* Since i != 0, there is no use
389 * checking to see if (++i) - 1 == 0,
390 * because that's the same as i == 0,
391 * so we skip a space.
399 while ( i
+ numecs
>= current_max_xpairs
)
403 /* If we started search from the beginning, store the new
404 * firstfree for the next call of find_table_space().
406 if ( numtrans
<= MAX_XTIONS_FULL_INTERIOR_FIT
)
409 /* Check to see if all elements in chk (and therefore nxt)
410 * that are needed for the new state have not yet been taken.
413 state_ptr
= &state
[1];
414 ptr_to_last_entry_in_state
= &chk
[i
+ numecs
+ 1];
416 for ( chk_ptr
= &chk
[i
+ 1];
417 chk_ptr
!= ptr_to_last_entry_in_state
; ++chk_ptr
)
418 if ( *(state_ptr
++) != 0 && *chk_ptr
!= 0 )
421 if ( chk_ptr
== ptr_to_last_entry_in_state
)
430 /* inittbl - initialize transition tables
432 * Initializes "firstfree" to be one beyond the end of the table. Initializes
433 * all "chk" entries to be zero.
439 zero_out( (char *) chk
, (size_t) (current_max_xpairs
* sizeof( int )) );
442 firstfree
= tblend
+ 1;
447 /* Set up doubly-linked meta-equivalence classes; these
448 * are sets of equivalence classes which all have identical
449 * transitions out of TEMPLATES.
454 for ( i
= 2; i
<= numecs
; ++i
)
460 tecfwd
[numecs
] = NIL
;
465 /* mkdeftbl - make the default, "jam" table entries */
471 jamstate
= lastdfa
+ 1;
473 ++tblend
; /* room for transition on end-of-buffer character */
475 while ( tblend
+ numecs
>= current_max_xpairs
)
478 /* Add in default end-of-buffer transition. */
479 nxt
[tblend
] = end_of_buffer_state
;
480 chk
[tblend
] = jamstate
;
482 for ( i
= 1; i
<= numecs
; ++i
)
485 chk
[tblend
+ i
] = jamstate
;
490 base
[jamstate
] = jambase
;
498 /* mkentry - create base/def and nxt/chk entries for transition array
501 * int state[numchars + 1], numchars, statenum, deflink, totaltrans;
502 * mkentry( state, numchars, statenum, deflink, totaltrans );
504 * "state" is a transition array "numchars" characters in size, "statenum"
505 * is the offset to be used into the base/def tables, and "deflink" is the
506 * entry to put in the "def" table entry. If "deflink" is equal to
507 * "JAMSTATE", then no attempt will be made to fit zero entries of "state"
508 * (i.e., jam entries) into the table. It is assumed that by linking to
509 * "JAMSTATE" they will be taken care of. In any case, entries in "state"
510 * marking transitions to "SAME_TRANS" are treated as though they will be
511 * taken care of by whereever "deflink" points. "totaltrans" is the total
512 * number of transitions out of the state. If it is below a certain threshold,
513 * the tables are searched for an interior spot that will accommodate the
517 void mkentry(int *state
, int numchars
, int statenum
, int deflink
,
520 int minec
, maxec
, i
, baseaddr
;
521 int tblbase
, tbllast
;
523 if ( totaltrans
== 0 )
524 { /* there are no out-transitions */
525 if ( deflink
== JAMSTATE
)
526 base
[statenum
] = JAMSTATE
;
530 def
[statenum
] = deflink
;
534 for ( minec
= 1; minec
<= numchars
; ++minec
)
536 if ( state
[minec
] != SAME_TRANS
)
537 if ( state
[minec
] != 0 || deflink
!= JAMSTATE
)
541 if ( totaltrans
== 1 )
543 /* There's only one out-transition. Save it for later to fill
544 * in holes in the tables.
546 stack1( statenum
, minec
, state
[minec
], deflink
);
550 for ( maxec
= numchars
; maxec
> 0; --maxec
)
552 if ( state
[maxec
] != SAME_TRANS
)
553 if ( state
[maxec
] != 0 || deflink
!= JAMSTATE
)
557 /* Whether we try to fit the state table in the middle of the table
558 * entries we have already generated, or if we just take the state
559 * table at the end of the nxt/chk tables, we must make sure that we
560 * have a valid base address (i.e., non-negative). Note that
561 * negative base addresses dangerous at run-time (because indexing
562 * the nxt array with one and a low-valued character will access
563 * memory before the start of the array.
566 /* Find the first transition of state that we need to worry about. */
567 if ( totaltrans
* 100 <= numchars
* INTERIOR_FIT_PERCENTAGE
)
569 /* Attempt to squeeze it into the middle of the tables. */
570 baseaddr
= firstfree
;
572 while ( baseaddr
< minec
)
574 /* Using baseaddr would result in a negative base
575 * address below; find the next free slot.
577 for ( ++baseaddr
; chk
[baseaddr
] != 0; ++baseaddr
)
581 while ( baseaddr
+ maxec
- minec
+ 1 >= current_max_xpairs
)
584 for ( i
= minec
; i
<= maxec
; ++i
)
585 if ( state
[i
] != SAME_TRANS
&&
586 (state
[i
] != 0 || deflink
!= JAMSTATE
) &&
587 chk
[baseaddr
+ i
- minec
] != 0 )
588 { /* baseaddr unsuitable - find another */
590 baseaddr
< current_max_xpairs
&&
591 chk
[baseaddr
] != 0; ++baseaddr
)
594 while ( baseaddr
+ maxec
- minec
+ 1 >=
598 /* Reset the loop counter so we'll start all
599 * over again next time it's incremented.
608 /* Ensure that the base address we eventually generate is
611 baseaddr
= MAX( tblend
+ 1, minec
);
614 tblbase
= baseaddr
- minec
;
615 tbllast
= tblbase
+ maxec
;
617 while ( tbllast
+ 1 >= current_max_xpairs
)
620 base
[statenum
] = tblbase
;
621 def
[statenum
] = deflink
;
623 for ( i
= minec
; i
<= maxec
; ++i
)
624 if ( state
[i
] != SAME_TRANS
)
625 if ( state
[i
] != 0 || deflink
!= JAMSTATE
)
627 nxt
[tblbase
+ i
] = state
[i
];
628 chk
[tblbase
+ i
] = statenum
;
631 if ( baseaddr
== firstfree
)
632 /* Find next free slot in tables. */
633 for ( ++firstfree
; chk
[firstfree
] != 0; ++firstfree
)
636 tblend
= MAX( tblend
, tbllast
);
640 /* mk1tbl - create table entries for a state (or state fragment) which
641 * has only one out-transition
644 void mk1tbl(int state
, int sym
, int onenxt
, int onedef
)
646 if ( firstfree
< sym
)
649 while ( chk
[firstfree
] != 0 )
650 if ( ++firstfree
>= current_max_xpairs
)
653 base
[state
] = firstfree
- sym
;
655 chk
[firstfree
] = state
;
656 nxt
[firstfree
] = onenxt
;
658 if ( firstfree
> tblend
)
660 tblend
= firstfree
++;
662 if ( firstfree
>= current_max_xpairs
)
668 /* mkprot - create new proto entry */
670 void mkprot(int *state
, int statenum
, int comstate
)
672 int i
, slot
, tblbase
;
674 if ( ++numprots
>= MSP
|| numecs
* numprots
>= PROT_SAVE_SIZE
)
676 /* Gotta make room for the new proto by dropping last entry in
680 lastprot
= protprev
[lastprot
];
681 protnext
[lastprot
] = NIL
;
687 protnext
[slot
] = firstprot
;
689 if ( firstprot
!= NIL
)
690 protprev
[firstprot
] = slot
;
693 prottbl
[slot
] = statenum
;
694 protcomst
[slot
] = comstate
;
696 /* Copy state into save area so it can be compared with rapidly. */
697 tblbase
= numecs
* (slot
- 1);
699 for ( i
= 1; i
<= numecs
; ++i
)
700 protsave
[tblbase
+ i
] = state
[i
];
704 /* mktemplate - create a template entry based on a state, and connect the state
708 void mktemplate(int *state
, int statenum
, int comstate
)
710 int i
, numdiff
, tmpbase
, tmp
[CSIZE
+ 1];
711 Char transset
[CSIZE
+ 1];
718 /* Calculate where we will temporarily store the transition table
719 * of the template in the tnxt[] array. The final transition table
720 * gets created by cmptmps().
723 tmpbase
= numtemps
* numecs
;
725 if ( tmpbase
+ numecs
>= current_max_template_xpairs
)
727 current_max_template_xpairs
+= MAX_TEMPLATE_XPAIRS_INCREMENT
;
731 tnxt
= reallocate_integer_array( tnxt
,
732 current_max_template_xpairs
);
735 for ( i
= 1; i
<= numecs
; ++i
)
737 tnxt
[tmpbase
+ i
] = 0;
740 transset
[tsptr
++] = i
;
741 tnxt
[tmpbase
+ i
] = comstate
;
745 mkeccl( transset
, tsptr
, tecfwd
, tecbck
, numecs
, 0 );
747 mkprot( tnxt
+ tmpbase
, -numtemps
, comstate
);
749 /* We rely on the fact that mkprot adds things to the beginning
750 * of the proto queue.
753 numdiff
= tbldiff( state
, firstprot
, tmp
);
754 mkentry( tmp
, numecs
, statenum
, -numtemps
, numdiff
);
758 /* mv2front - move proto queue element to front of queue */
760 void mv2front(int qelm
)
762 if ( firstprot
!= qelm
)
764 if ( qelm
== lastprot
)
765 lastprot
= protprev
[lastprot
];
767 protnext
[protprev
[qelm
]] = protnext
[qelm
];
769 if ( protnext
[qelm
] != NIL
)
770 protprev
[protnext
[qelm
]] = protprev
[qelm
];
772 protprev
[qelm
] = NIL
;
773 protnext
[qelm
] = firstprot
;
774 protprev
[firstprot
] = qelm
;
780 /* place_state - place a state into full speed transition table
782 * State is the statenum'th state. It is indexed by equivalence class and
783 * gives the number of the state to enter for a given equivalence class.
784 * Transnum is the number of out-transitions for the state.
787 void place_state(int *state
, int statenum
, int transnum
)
791 int position
= find_table_space( state
, transnum
);
793 /* "base" is the table of start positions. */
794 base
[statenum
] = position
;
796 /* Put in action number marker; this non-zero number makes sure that
797 * find_table_space() knows that this position in chk/nxt is taken
798 * and should not be used for another accepting number in another
801 chk
[position
- 1] = 1;
803 /* Put in end-of-buffer marker; this is for the same purposes as
808 /* Place the state into chk and nxt. */
809 state_ptr
= &state
[1];
811 for ( i
= 1; i
<= numecs
; ++i
, ++state_ptr
)
812 if ( *state_ptr
!= 0 )
814 chk
[position
+ i
] = i
;
815 nxt
[position
+ i
] = *state_ptr
;
818 if ( position
+ numecs
> tblend
)
819 tblend
= position
+ numecs
;
823 /* stack1 - save states with only one out-transition to be processed later
825 * If there's room for another state on the "one-transition" stack, the
826 * state is pushed onto it, to be processed later by mk1tbl. If there's
827 * no room, we process the sucker right now.
830 void stack1(int statenum
, int sym
, int nextstate
, int deflink
)
832 if ( onesp
>= ONE_STACK_SIZE
- 1 )
833 mk1tbl( statenum
, sym
, nextstate
, deflink
);
838 onestate
[onesp
] = statenum
;
840 onenext
[onesp
] = nextstate
;
841 onedef
[onesp
] = deflink
;
846 /* tbldiff - compute differences between two state tables
848 * "state" is the state array which is to be extracted from the pr'th
849 * proto. "pr" is both the number of the proto we are extracting from
850 * and an index into the save area where we can find the proto's complete
851 * state table. Each entry in "state" which differs from the corresponding
852 * entry of "pr" will appear in "ext".
854 * Entries which are the same in both "state" and "pr" will be marked
855 * as transitions to "SAME_TRANS" in "ext". The total number of differences
856 * between "state" and "pr" is returned as function value. Note that this
857 * number is "numecs" minus the number of "SAME_TRANS" entries in "ext".
860 int tbldiff(int *state
, int pr
, int *ext
)
862 int i
, *sp
= state
, *ep
= ext
, *protp
;
865 protp
= &protsave
[numecs
* (pr
- 1)];
867 for ( i
= numecs
; i
> 0; --i
)
869 if ( *++protp
== *++sp
)