PostgreSQL Source Code  git master
trgm_regexp.c
Go to the documentation of this file.
1 /*-------------------------------------------------------------------------
2  *
3  * trgm_regexp.c
4  * Regular expression matching using trigrams.
5  *
6  * The general idea of trigram index support for a regular expression (regex)
7  * search is to transform the regex into a logical expression on trigrams.
8  * For example:
9  *
10  * (ab|cd)efg => ((abe & bef) | (cde & def)) & efg
11  *
12  * If a string matches the regex, then it must match the logical expression on
13  * trigrams. The opposite is not necessarily true, however: a string that
14  * matches the logical expression might not match the original regex. Such
15  * false positives are removed via recheck, by running the regular regex match
16  * operator on the retrieved heap tuple.
17  *
18  * Since the trigram expression involves both AND and OR operators, we can't
19  * expect the core index machinery to evaluate it completely. Instead, the
20  * result of regex analysis is a list of trigrams to be sought in the index,
21  * plus a simplified graph that is used by trigramsMatchGraph() to determine
22  * whether a particular indexed value matches the expression.
23  *
24  * Converting a regex to a trigram expression is based on analysis of an
25  * automaton corresponding to the regex. The algorithm consists of four
26  * stages:
27  *
28  * 1) Compile the regexp to NFA form. This is handled by the PostgreSQL
29  * regexp library, which provides accessors for its opaque regex_t struct
30  * to expose the NFA state graph and the "colors" (sets of equivalent
31  * characters) used as state transition labels.
32  *
33  * 2) Transform the original NFA into an expanded graph, where arcs
34  * are labeled with trigrams that must be present in order to move from
35  * one state to another via the arcs. The trigrams used in this stage
36  * consist of colors, not characters, as in the original NFA.
37  *
38  * 3) Expand the color trigrams into regular trigrams consisting of
39  * characters. If too many distinct trigrams are produced, trigrams are
40  * eliminated and the graph is simplified until it's simple enough.
41  *
42  * 4) Finally, the resulting graph is packed into a TrgmPackedGraph struct,
43  * and returned to the caller.
44  *
45  * 1) Compile the regexp to NFA form
46  * ---------------------------------
47  * The automaton returned by the regexp compiler is a graph where vertices
48  * are "states" and arcs are labeled with colors. Each color represents
49  * a set of characters, so that all characters assigned to the same color
50  * are interchangeable, so far as matching the regexp is concerned. There
51  * are two special states: "initial" and "final". A state can have multiple
52  * outgoing arcs labeled with the same color, which makes the automaton
53  * non-deterministic, because it can be in many states simultaneously.
54  *
55  * Note that this NFA is already lossy compared to the original regexp,
56  * since it ignores some regex features such as lookahead constraints and
57  * backref matching. This is OK for our purposes since it's still the case
58  * that only strings matching the NFA can possibly satisfy the regexp.
59  *
60  * 2) Transform the original NFA into an expanded graph
61  * ----------------------------------------------------
62  * In the 2nd stage, the automaton is transformed into a graph based on the
63  * original NFA. Each state in the expanded graph represents a state from
64  * the original NFA, plus a prefix identifying the last two characters
65  * (colors, to be precise) seen before entering the state. There can be
66  * multiple states in the expanded graph for each state in the original NFA,
67  * depending on what characters can precede it. A prefix position can be
68  * "unknown" if it's uncertain what the preceding character was, or "blank"
69  * if the character was a non-word character (we don't need to distinguish
70  * which non-word character it was, so just think of all of them as blanks).
71  *
72  * For convenience in description, call an expanded-state identifier
73  * (two prefix colors plus a state number from the original NFA) an
74  * "enter key".
75  *
76  * Each arc of the expanded graph is labeled with a trigram that must be
77  * present in the string to match. We can construct this from an out-arc of
78  * the underlying NFA state by combining the expanded state's prefix with the
79  * color label of the underlying out-arc, if neither prefix position is
80  * "unknown". But note that some of the colors in the trigram might be
81  * "blank". This is OK since we want to generate word-boundary trigrams as
82  * the regular trigram machinery would, if we know that some word characters
83  * must be adjacent to a word boundary in all strings matching the NFA.
84  *
85  * The expanded graph can also have fewer states than the original NFA,
86  * because we don't bother to make a separate state entry unless the state
87  * is reachable by a valid arc. When an enter key is reachable from a state
88  * of the expanded graph, but we do not know a complete trigram associated
89  * with that transition, we cannot make a valid arc; instead we insert the
90  * enter key into the enterKeys list of the source state. This effectively
91  * means that the two expanded states are not reliably distinguishable based
92  * on examining trigrams.
93  *
94  * So the expanded graph resembles the original NFA, but the arcs are
95  * labeled with trigrams instead of individual characters, and there may be
96  * more or fewer states. It is a lossy representation of the original NFA:
97  * any string that matches the original regexp must match the expanded graph,
98  * but the reverse is not true.
99  *
100  * We build the expanded graph through a breadth-first traversal of states
101  * reachable from the initial state. At each reachable state, we identify the
102  * states reachable from it without traversing a predictable trigram, and add
103  * those states' enter keys to the current state. Then we generate all
104  * out-arcs leading out of this collection of states that have predictable
105  * trigrams, adding their target states to the queue of states to examine.
106  *
107  * When building the graph, if the number of states or arcs exceed pre-defined
108  * limits, we give up and simply mark any states not yet processed as final
109  * states. Roughly speaking, that means that we make use of some portion from
110  * the beginning of the regexp. Also, any colors that have too many member
111  * characters are treated as "unknown", so that we can't derive trigrams
112  * from them.
113  *
114  * 3) Expand the color trigrams into regular trigrams
115  * --------------------------------------------------
116  * The trigrams in the expanded graph are "color trigrams", consisting
117  * of three consecutive colors that must be present in the string. But for
118  * search, we need regular trigrams consisting of characters. In the 3rd
119  * stage, the color trigrams are expanded into regular trigrams. Since each
120  * color can represent many characters, the total number of regular trigrams
121  * after expansion could be very large. Because searching the index for
122  * thousands of trigrams would be slow, and would likely produce so many
123  * false positives that we would have to traverse a large fraction of the
124  * index, the graph is simplified further in a lossy fashion by removing
125  * color trigrams. When a color trigram is removed, the states connected by
126  * any arcs labeled with that trigram are merged.
127  *
128  * Trigrams do not all have equivalent value for searching: some of them are
129  * more frequent and some of them are less frequent. Ideally, we would like
130  * to know the distribution of trigrams, but we don't. But because of padding
131  * we know for sure that the empty character is more frequent than others,
132  * so we can penalize trigrams according to presence of whitespace. The
133  * penalty assigned to each color trigram is the number of simple trigrams
134  * it would produce, times the penalties[] multiplier associated with its
135  * whitespace content. (The penalties[] constants were calculated by analysis
136  * of some real-life text.) We eliminate color trigrams starting with the
137  * highest-penalty one, until we get to a total penalty of no more than
138  * WISH_TRGM_PENALTY. However, we cannot remove a color trigram if that would
139  * lead to merging the initial and final states, so we may not be able to
140  * reach WISH_TRGM_PENALTY. It's still okay so long as we have no more than
141  * MAX_TRGM_COUNT simple trigrams in total, otherwise we fail.
142  *
143  * 4) Pack the graph into a compact representation
144  * -----------------------------------------------
145  * The 2nd and 3rd stages might have eliminated or merged many of the states
146  * and trigrams created earlier, so in this final stage, the graph is
147  * compacted and packed into a simpler struct that contains only the
148  * information needed to evaluate it.
149  *
150  * ALGORITHM EXAMPLE:
151  *
152  * Consider the example regex "ab[cd]". This regex is transformed into the
153  * following NFA (for simplicity we show colors as their single members):
154  *
155  * 4#
156  * c/
157  * a b /
158  * 1* --- 2 ---- 3
159  * \
160  * d\
161  * 5#
162  *
163  * We use * to mark initial state and # to mark final state. It's not depicted,
164  * but states 1, 4, 5 have self-referencing arcs for all possible characters,
165  * because this pattern can match to any part of a string.
166  *
167  * As the result of stage 2 we will have the following graph:
168  *
169  * abc abd
170  * 2# <---- 1* ----> 3#
171  *
172  * The process for generating this graph is:
173  * 1) Create state 1 with enter key (UNKNOWN, UNKNOWN, 1).
174  * 2) Add key (UNKNOWN, "a", 2) to state 1.
175  * 3) Add key ("a", "b", 3) to state 1.
176  * 4) Create new state 2 with enter key ("b", "c", 4). Add an arc
177  * from state 1 to state 2 with label trigram "abc".
178  * 5) Mark state 2 final because state 4 of source NFA is marked as final.
179  * 6) Create new state 3 with enter key ("b", "d", 5). Add an arc
180  * from state 1 to state 3 with label trigram "abd".
181  * 7) Mark state 3 final because state 5 of source NFA is marked as final.
182  *
183  *
184  * Portions Copyright (c) 1996-2021, PostgreSQL Global Development Group
185  * Portions Copyright (c) 1994, Regents of the University of California
186  *
187  * IDENTIFICATION
188  * contrib/pg_trgm/trgm_regexp.c
189  *
190  *-------------------------------------------------------------------------
191  */
192 #include "postgres.h"
193 
194 #include "regex/regexport.h"
195 #include "trgm.h"
196 #include "tsearch/ts_locale.h"
197 #include "utils/hsearch.h"
198 #include "utils/memutils.h"
199 
200 /*
201  * Uncomment (or use -DTRGM_REGEXP_DEBUG) to print debug info,
202  * for exploring and debugging the algorithm implementation.
203  * This produces three graph files in /tmp, in Graphviz .gv format.
204  * Some progress information is also printed to postmaster stderr.
205  */
206 /* #define TRGM_REGEXP_DEBUG */
207 
208 /*
209  * These parameters are used to limit the amount of work done.
210  * Otherwise regex processing could be too slow and memory-consuming.
211  *
212  * MAX_EXPANDED_STATES - How many states we allow in expanded graph
213  * MAX_EXPANDED_ARCS - How many arcs we allow in expanded graph
214  * MAX_TRGM_COUNT - How many simple trigrams we allow to be extracted
215  * WISH_TRGM_PENALTY - Maximum desired sum of color trigram penalties
216  * COLOR_COUNT_LIMIT - Maximum number of characters per color
217  */
218 #define MAX_EXPANDED_STATES 128
219 #define MAX_EXPANDED_ARCS 1024
220 #define MAX_TRGM_COUNT 256
221 #define WISH_TRGM_PENALTY 16
222 #define COLOR_COUNT_LIMIT 256
223 
224 /*
225  * Penalty multipliers for trigram counts depending on whitespace contents.
226  * Numbers based on analysis of real-life texts.
227  */
228 static const float4 penalties[8] = {
229  1.0f, /* "aaa" */
230  3.5f, /* "aa " */
231  0.0f, /* "a a" (impossible) */
232  0.0f, /* "a " (impossible) */
233  4.2f, /* " aa" */
234  2.1f, /* " a " */
235  25.0f, /* " a" */
236  0.0f /* " " (impossible) */
237 };
238 
239 /* Struct representing a single pg_wchar, converted back to multibyte form */
240 typedef struct
241 {
243 } trgm_mb_char;
244 
245 /*
246  * Attributes of NFA colors:
247  *
248  * expandable - we know the character expansion of this color
249  * containsNonWord - color contains non-word characters
250  * (which will not be extracted into trigrams)
251  * wordCharsCount - count of word characters in color
252  * wordChars - array of this color's word characters
253  * (which can be extracted into trigrams)
254  *
255  * When expandable is false, the other attributes don't matter; we just
256  * assume this color represents unknown character(s).
257  */
258 typedef struct
259 {
264 } TrgmColorInfo;
265 
266 /*
267  * A "prefix" is information about the colors of the last two characters read
268  * before reaching a specific NFA state. These colors can have special values
269  * COLOR_UNKNOWN and COLOR_BLANK. COLOR_UNKNOWN means that we have no
270  * information, for example because we read some character of an unexpandable
271  * color. COLOR_BLANK means that we read a non-word character.
272  *
273  * We call a prefix ambiguous if at least one of its colors is unknown. It's
274  * fully ambiguous if both are unknown, partially ambiguous if only the first
275  * is unknown. (The case of first color known, second unknown is not valid.)
276  *
277  * Wholly- or partly-blank prefixes are mostly handled the same as regular
278  * color prefixes. This allows us to generate appropriate partly-blank
279  * trigrams when the NFA requires word character(s) to appear adjacent to
280  * non-word character(s).
281  */
282 typedef int TrgmColor;
283 
284 /* We assume that colors returned by the regexp engine cannot be these: */
285 #define COLOR_UNKNOWN (-3)
286 #define COLOR_BLANK (-4)
287 
288 typedef struct
289 {
290  TrgmColor colors[2];
291 } TrgmPrefix;
292 
293 /*
294  * Color-trigram data type. Note that some elements of the trigram can be
295  * COLOR_BLANK, but we don't allow COLOR_UNKNOWN.
296  */
297 typedef struct
298 {
299  TrgmColor colors[3];
300 } ColorTrgm;
301 
302 /*
303  * Key identifying a state of our expanded graph: color prefix, and number
304  * of the corresponding state in the underlying regex NFA. The color prefix
305  * shows how we reached the regex state (to the extent that we know it).
306  */
307 typedef struct
308 {
310  int nstate;
311 } TrgmStateKey;
312 
313 /*
314  * One state of the expanded graph.
315  *
316  * stateKey - ID of this state
317  * arcs - outgoing arcs of this state (List of TrgmArc)
318  * enterKeys - enter keys reachable from this state without reading any
319  * predictable trigram (List of TrgmStateKey)
320  * flags - flag bits
321  * snumber - number of this state (initially assigned as -1, -2, etc,
322  * for debugging purposes only; then at the packaging stage,
323  * surviving states are renumbered with positive numbers)
324  * parent - parent state, if this state has been merged into another
325  * tentFlags - flags this state would acquire via planned merges
326  * tentParent - planned parent state, if considering a merge
327  */
328 #define TSTATE_INIT 0x01 /* flag indicating this state is initial */
329 #define TSTATE_FIN 0x02 /* flag indicating this state is final */
330 
331 typedef struct TrgmState
332 {
333  TrgmStateKey stateKey; /* hashtable key: must be first field */
336  int flags;
337  int snumber;
338  struct TrgmState *parent;
341 } TrgmState;
342 
343 /*
344  * One arc in the expanded graph.
345  */
346 typedef struct
347 {
348  ColorTrgm ctrgm; /* trigram needed to traverse arc */
349  TrgmState *target; /* next state */
350 } TrgmArc;
351 
352 /*
353  * Information about arc of specific color trigram (used in stage 3)
354  *
355  * Contains pointers to the source and target states.
356  */
357 typedef struct
358 {
361 } TrgmArcInfo;
362 
363 /*
364  * Information about color trigram (used in stage 3)
365  *
366  * ctrgm - trigram itself
367  * cnumber - number of this trigram (used in the packaging stage)
368  * count - number of simple trigrams created from this color trigram
369  * expanded - indicates this color trigram is expanded into simple trigrams
370  * arcs - list of all arcs labeled with this color trigram.
371  */
372 typedef struct
373 {
375  int cnumber;
376  int count;
378  bool expanded;
380 } ColorTrgmInfo;
381 
382 /*
383  * Data structure representing all the data we need during regex processing.
384  *
385  * regex - compiled regex
386  * colorInfo - extracted information about regex's colors
387  * ncolors - number of colors in colorInfo[]
388  * states - hashtable of TrgmStates (states of expanded graph)
389  * initState - pointer to initial state of expanded graph
390  * queue - queue of to-be-processed TrgmStates
391  * keysQueue - queue of to-be-processed TrgmStateKeys
392  * arcsCount - total number of arcs of expanded graph (for resource
393  * limiting)
394  * overflowed - we have exceeded resource limit for transformation
395  * colorTrgms - array of all color trigrams present in graph
396  * colorTrgmsCount - count of those color trigrams
397  * totalTrgmCount - total count of extracted simple trigrams
398  */
399 typedef struct
400 {
401  /* Source regexp, and color information extracted from it (stage 1) */
404  int ncolors;
405 
406  /* Expanded graph (stage 2) */
409  int nstates;
410 
411  /* Workspace for stage 2 */
416 
417  /* Information about distinct color trigrams in the graph (stage 3) */
421 } TrgmNFA;
422 
423 /*
424  * Final, compact representation of expanded graph.
425  */
426 typedef struct
427 {
428  int targetState; /* index of target state (zero-based) */
429  int colorTrgm; /* index of color trigram for transition */
430 } TrgmPackedArc;
431 
432 typedef struct
433 {
434  int arcsCount; /* number of out-arcs for this state */
435  TrgmPackedArc *arcs; /* array of arcsCount packed arcs */
437 
438 /* "typedef struct TrgmPackedGraph TrgmPackedGraph" appears in trgm.h */
440 {
441  /*
442  * colorTrigramsCount and colorTrigramGroups contain information about how
443  * trigrams are grouped into color trigrams. "colorTrigramsCount" is the
444  * count of color trigrams and "colorTrigramGroups" contains number of
445  * simple trigrams for each color trigram. The array of simple trigrams
446  * (stored separately from this struct) is ordered so that the simple
447  * trigrams for each color trigram are consecutive, and they're in order
448  * by color trigram number.
449  */
451  int *colorTrigramGroups; /* array of size colorTrigramsCount */
452 
453  /*
454  * The states of the simplified NFA. State number 0 is always initial
455  * state and state number 1 is always final state.
456  */
458  TrgmPackedState *states; /* array of size statesCount */
459 
460  /* Temporary work space for trigramsMatchGraph() */
461  bool *colorTrigramsActive; /* array of size colorTrigramsCount */
462  bool *statesActive; /* array of size statesCount */
463  int *statesQueue; /* array of size statesCount */
464 };
465 
466 /*
467  * Temporary structure for representing an arc during packaging.
468  */
469 typedef struct
470 {
475 
476 
477 /* prototypes for private functions */
478 static TRGM *createTrgmNFAInternal(regex_t *regex, TrgmPackedGraph **graph,
479  MemoryContext rcontext);
480 static void RE_compile(regex_t *regex, text *text_re,
481  int cflags, Oid collation);
482 static void getColorInfo(regex_t *regex, TrgmNFA *trgmNFA);
483 static bool convertPgWchar(pg_wchar c, trgm_mb_char *result);
484 static void transformGraph(TrgmNFA *trgmNFA);
485 static void processState(TrgmNFA *trgmNFA, TrgmState *state);
486 static void addKey(TrgmNFA *trgmNFA, TrgmState *state, TrgmStateKey *key);
487 static void addKeyToQueue(TrgmNFA *trgmNFA, TrgmStateKey *key);
488 static void addArcs(TrgmNFA *trgmNFA, TrgmState *state);
489 static void addArc(TrgmNFA *trgmNFA, TrgmState *state, TrgmStateKey *key,
490  TrgmColor co, TrgmStateKey *destKey);
491 static bool validArcLabel(TrgmStateKey *key, TrgmColor co);
492 static TrgmState *getState(TrgmNFA *trgmNFA, TrgmStateKey *key);
493 static bool prefixContains(TrgmPrefix *prefix1, TrgmPrefix *prefix2);
494 static bool selectColorTrigrams(TrgmNFA *trgmNFA);
495 static TRGM *expandColorTrigrams(TrgmNFA *trgmNFA, MemoryContext rcontext);
496 static void fillTrgm(trgm *ptrgm, trgm_mb_char s[3]);
497 static void mergeStates(TrgmState *state1, TrgmState *state2);
498 static int colorTrgmInfoCmp(const void *p1, const void *p2);
499 static int colorTrgmInfoPenaltyCmp(const void *p1, const void *p2);
500 static TrgmPackedGraph *packGraph(TrgmNFA *trgmNFA, MemoryContext rcontext);
501 static int packArcInfoCmp(const void *a1, const void *a2);
502 
503 #ifdef TRGM_REGEXP_DEBUG
504 static void printSourceNFA(regex_t *regex, TrgmColorInfo *colors, int ncolors);
505 static void printTrgmNFA(TrgmNFA *trgmNFA);
506 static void printTrgmColor(StringInfo buf, TrgmColor co);
507 static void printTrgmPackedGraph(TrgmPackedGraph *packedGraph, TRGM *trigrams);
508 #endif
509 
510 
511 /*
512  * Main entry point to process a regular expression.
513  *
514  * Returns an array of trigrams required by the regular expression, or NULL if
515  * the regular expression was too complex to analyze. In addition, a packed
516  * graph representation of the regex is returned into *graph. The results
517  * must be allocated in rcontext (which might or might not be the current
518  * context).
519  */
520 TRGM *
521 createTrgmNFA(text *text_re, Oid collation,
522  TrgmPackedGraph **graph, MemoryContext rcontext)
523 {
524  TRGM *trg;
525  regex_t regex;
526  MemoryContext tmpcontext;
527  MemoryContext oldcontext;
528 
529  /*
530  * This processing generates a great deal of cruft, which we'd like to
531  * clean up before returning (since this function may be called in a
532  * query-lifespan memory context). Make a temp context we can work in so
533  * that cleanup is easy.
534  */
536  "createTrgmNFA temporary context",
538  oldcontext = MemoryContextSwitchTo(tmpcontext);
539 
540  /*
541  * Stage 1: Compile the regexp into a NFA, using the regexp library.
542  */
543 #ifdef IGNORECASE
544  RE_compile(&regex, text_re, REG_ADVANCED | REG_ICASE, collation);
545 #else
546  RE_compile(&regex, text_re, REG_ADVANCED, collation);
547 #endif
548 
549  /*
550  * Since the regexp library allocates its internal data structures with
551  * malloc, we need to use a PG_TRY block to ensure that pg_regfree() gets
552  * done even if there's an error.
553  */
554  PG_TRY();
555  {
556  trg = createTrgmNFAInternal(&regex, graph, rcontext);
557  }
558  PG_FINALLY();
559  {
560  pg_regfree(&regex);
561  }
562  PG_END_TRY();
563 
564  /* Clean up all the cruft we created */
565  MemoryContextSwitchTo(oldcontext);
566  MemoryContextDelete(tmpcontext);
567 
568  return trg;
569 }
570 
571 /*
572  * Body of createTrgmNFA, exclusive of regex compilation/freeing.
573  */
574 static TRGM *
576  MemoryContext rcontext)
577 {
578  TRGM *trg;
579  TrgmNFA trgmNFA;
580 
581  trgmNFA.regex = regex;
582 
583  /* Collect color information from the regex */
584  getColorInfo(regex, &trgmNFA);
585 
586 #ifdef TRGM_REGEXP_DEBUG
587  printSourceNFA(regex, trgmNFA.colorInfo, trgmNFA.ncolors);
588 #endif
589 
590  /*
591  * Stage 2: Create an expanded graph from the source NFA.
592  */
593  transformGraph(&trgmNFA);
594 
595 #ifdef TRGM_REGEXP_DEBUG
596  printTrgmNFA(&trgmNFA);
597 #endif
598 
599  /*
600  * Fail if we were unable to make a nontrivial graph, ie it is possible to
601  * get from the initial state to the final state without reading any
602  * predictable trigram.
603  */
604  if (trgmNFA.initState->flags & TSTATE_FIN)
605  return NULL;
606 
607  /*
608  * Stage 3: Select color trigrams to expand. Fail if too many trigrams.
609  */
610  if (!selectColorTrigrams(&trgmNFA))
611  return NULL;
612 
613  /*
614  * Stage 4: Expand color trigrams and pack graph into final
615  * representation.
616  */
617  trg = expandColorTrigrams(&trgmNFA, rcontext);
618 
619  *graph = packGraph(&trgmNFA, rcontext);
620 
621 #ifdef TRGM_REGEXP_DEBUG
622  printTrgmPackedGraph(*graph, trg);
623 #endif
624 
625  return trg;
626 }
627 
628 /*
629  * Main entry point for evaluating a graph during index scanning.
630  *
631  * The check[] array is indexed by trigram number (in the array of simple
632  * trigrams returned by createTrgmNFA), and holds true for those trigrams
633  * that are present in the index entry being checked.
634  */
635 bool
637 {
638  int i,
639  j,
640  k,
641  queueIn,
642  queueOut;
643 
644  /*
645  * Reset temporary working areas.
646  */
647  memset(graph->colorTrigramsActive, 0,
648  sizeof(bool) * graph->colorTrigramsCount);
649  memset(graph->statesActive, 0, sizeof(bool) * graph->statesCount);
650 
651  /*
652  * Check which color trigrams were matched. A match for any simple
653  * trigram associated with a color trigram counts as a match of the color
654  * trigram.
655  */
656  j = 0;
657  for (i = 0; i < graph->colorTrigramsCount; i++)
658  {
659  int cnt = graph->colorTrigramGroups[i];
660 
661  for (k = j; k < j + cnt; k++)
662  {
663  if (check[k])
664  {
665  /*
666  * Found one matched trigram in the group. Can skip the rest
667  * of them and go to the next group.
668  */
669  graph->colorTrigramsActive[i] = true;
670  break;
671  }
672  }
673  j = j + cnt;
674  }
675 
676  /*
677  * Initialize the statesQueue to hold just the initial state. Note:
678  * statesQueue has room for statesCount entries, which is certainly enough
679  * since no state will be put in the queue more than once. The
680  * statesActive array marks which states have been queued.
681  */
682  graph->statesActive[0] = true;
683  graph->statesQueue[0] = 0;
684  queueIn = 0;
685  queueOut = 1;
686 
687  /* Process queued states as long as there are any. */
688  while (queueIn < queueOut)
689  {
690  int stateno = graph->statesQueue[queueIn++];
691  TrgmPackedState *state = &graph->states[stateno];
692  int cnt = state->arcsCount;
693 
694  /* Loop over state's out-arcs */
695  for (i = 0; i < cnt; i++)
696  {
697  TrgmPackedArc *arc = &state->arcs[i];
698 
699  /*
700  * If corresponding color trigram is present then activate the
701  * corresponding state. We're done if that's the final state,
702  * otherwise queue the state if it's not been queued already.
703  */
704  if (graph->colorTrigramsActive[arc->colorTrgm])
705  {
706  int nextstate = arc->targetState;
707 
708  if (nextstate == 1)
709  return true; /* success: final state is reachable */
710 
711  if (!graph->statesActive[nextstate])
712  {
713  graph->statesActive[nextstate] = true;
714  graph->statesQueue[queueOut++] = nextstate;
715  }
716  }
717  }
718  }
719 
720  /* Queue is empty, so match fails. */
721  return false;
722 }
723 
724 /*
725  * Compile regex string into struct at *regex.
726  * NB: pg_regfree must be applied to regex if this completes successfully.
727  */
728 static void
729 RE_compile(regex_t *regex, text *text_re, int cflags, Oid collation)
730 {
731  int text_re_len = VARSIZE_ANY_EXHDR(text_re);
732  char *text_re_val = VARDATA_ANY(text_re);
733  pg_wchar *pattern;
734  int pattern_len;
735  int regcomp_result;
736  char errMsg[100];
737 
738  /* Convert pattern string to wide characters */
739  pattern = (pg_wchar *) palloc((text_re_len + 1) * sizeof(pg_wchar));
740  pattern_len = pg_mb2wchar_with_len(text_re_val,
741  pattern,
742  text_re_len);
743 
744  /* Compile regex */
745  regcomp_result = pg_regcomp(regex,
746  pattern,
747  pattern_len,
748  cflags,
749  collation);
750 
751  pfree(pattern);
752 
753  if (regcomp_result != REG_OKAY)
754  {
755  /* re didn't compile (no need for pg_regfree, if so) */
756  pg_regerror(regcomp_result, regex, errMsg, sizeof(errMsg));
757  ereport(ERROR,
758  (errcode(ERRCODE_INVALID_REGULAR_EXPRESSION),
759  errmsg("invalid regular expression: %s", errMsg)));
760  }
761 }
762 
763 
764 /*---------------------
765  * Subroutines for pre-processing the color map (stage 1).
766  *---------------------
767  */
768 
769 /*
770  * Fill TrgmColorInfo structure for each color using regex export functions.
771  */
772 static void
773 getColorInfo(regex_t *regex, TrgmNFA *trgmNFA)
774 {
775  int colorsCount = pg_reg_getnumcolors(regex);
776  int i;
777 
778  trgmNFA->ncolors = colorsCount;
779  trgmNFA->colorInfo = (TrgmColorInfo *)
780  palloc0(colorsCount * sizeof(TrgmColorInfo));
781 
782  /*
783  * Loop over colors, filling TrgmColorInfo about each. Note we include
784  * WHITE (0) even though we know it'll be reported as non-expandable.
785  */
786  for (i = 0; i < colorsCount; i++)
787  {
788  TrgmColorInfo *colorInfo = &trgmNFA->colorInfo[i];
789  int charsCount = pg_reg_getnumcharacters(regex, i);
790  pg_wchar *chars;
791  int j;
792 
793  if (charsCount < 0 || charsCount > COLOR_COUNT_LIMIT)
794  {
795  /* Non expandable, or too large to work with */
796  colorInfo->expandable = false;
797  continue;
798  }
799 
800  colorInfo->expandable = true;
801  colorInfo->containsNonWord = false;
802  colorInfo->wordChars = (trgm_mb_char *)
803  palloc(sizeof(trgm_mb_char) * charsCount);
804  colorInfo->wordCharsCount = 0;
805 
806  /* Extract all the chars in this color */
807  chars = (pg_wchar *) palloc(sizeof(pg_wchar) * charsCount);
808  pg_reg_getcharacters(regex, i, chars, charsCount);
809 
810  /*
811  * Convert characters back to multibyte form, and save only those that
812  * are word characters. Set "containsNonWord" if any non-word
813  * character. (Note: it'd probably be nicer to keep the chars in
814  * pg_wchar format for now, but ISWORDCHR wants to see multibyte.)
815  */
816  for (j = 0; j < charsCount; j++)
817  {
818  trgm_mb_char c;
819 
820  if (!convertPgWchar(chars[j], &c))
821  continue; /* ok to ignore it altogether */
822  if (ISWORDCHR(c.bytes))
823  colorInfo->wordChars[colorInfo->wordCharsCount++] = c;
824  else
825  colorInfo->containsNonWord = true;
826  }
827 
828  pfree(chars);
829  }
830 }
831 
832 /*
833  * Convert pg_wchar to multibyte format.
834  * Returns false if the character should be ignored completely.
835  */
836 static bool
838 {
839  /* "s" has enough space for a multibyte character and a trailing NUL */
840  char s[MAX_MULTIBYTE_CHAR_LEN + 1];
841 
842  /*
843  * We can ignore the NUL character, since it can never appear in a PG text
844  * string. This avoids the need for various special cases when
845  * reconstructing trigrams.
846  */
847  if (c == 0)
848  return false;
849 
850  /* Do the conversion, making sure the result is NUL-terminated */
851  memset(s, 0, sizeof(s));
852  pg_wchar2mb_with_len(&c, s, 1);
853 
854  /*
855  * In IGNORECASE mode, we can ignore uppercase characters. We assume that
856  * the regex engine generated both uppercase and lowercase equivalents
857  * within each color, since we used the REG_ICASE option; so there's no
858  * need to process the uppercase version.
859  *
860  * XXX this code is dependent on the assumption that lowerstr() works the
861  * same as the regex engine's internal case folding machinery. Might be
862  * wiser to expose pg_wc_tolower and test whether c == pg_wc_tolower(c).
863  * On the other hand, the trigrams in the index were created using
864  * lowerstr(), so we're probably screwed if there's any incompatibility
865  * anyway.
866  */
867 #ifdef IGNORECASE
868  {
869  char *lowerCased = lowerstr(s);
870 
871  if (strcmp(lowerCased, s) != 0)
872  {
873  pfree(lowerCased);
874  return false;
875  }
876  pfree(lowerCased);
877  }
878 #endif
879 
880  /* Fill result with exactly MAX_MULTIBYTE_CHAR_LEN bytes */
881  memcpy(result->bytes, s, MAX_MULTIBYTE_CHAR_LEN);
882  return true;
883 }
884 
885 
886 /*---------------------
887  * Subroutines for expanding original NFA graph into a trigram graph (stage 2).
888  *---------------------
889  */
890 
891 /*
892  * Transform the graph, given a regex and extracted color information.
893  *
894  * We create and process a queue of expanded-graph states until all the states
895  * are processed.
896  *
897  * This algorithm may be stopped due to resource limitation. In this case we
898  * force every unprocessed branch to immediately finish with matching (this
899  * can give us false positives but no false negatives) by marking all
900  * unprocessed states as final.
901  */
902 static void
904 {
905  HASHCTL hashCtl;
906  TrgmStateKey initkey;
907  TrgmState *initstate;
908 
909  /* Initialize this stage's workspace in trgmNFA struct */
910  trgmNFA->queue = NIL;
911  trgmNFA->keysQueue = NIL;
912  trgmNFA->arcsCount = 0;
913  trgmNFA->overflowed = false;
914 
915  /* Create hashtable for states */
916  hashCtl.keysize = sizeof(TrgmStateKey);
917  hashCtl.entrysize = sizeof(TrgmState);
918  hashCtl.hcxt = CurrentMemoryContext;
919  trgmNFA->states = hash_create("Trigram NFA",
920  1024,
921  &hashCtl,
923  trgmNFA->nstates = 0;
924 
925  /* Create initial state: ambiguous prefix, NFA's initial state */
926  MemSet(&initkey, 0, sizeof(initkey));
927  initkey.prefix.colors[0] = COLOR_UNKNOWN;
928  initkey.prefix.colors[1] = COLOR_UNKNOWN;
929  initkey.nstate = pg_reg_getinitialstate(trgmNFA->regex);
930 
931  initstate = getState(trgmNFA, &initkey);
932  initstate->flags |= TSTATE_INIT;
933  trgmNFA->initState = initstate;
934 
935  /*
936  * Recursively build the expanded graph by processing queue of states
937  * (breadth-first search). getState already put initstate in the queue.
938  */
939  while (trgmNFA->queue != NIL)
940  {
941  TrgmState *state = (TrgmState *) linitial(trgmNFA->queue);
942 
943  trgmNFA->queue = list_delete_first(trgmNFA->queue);
944 
945  /*
946  * If we overflowed then just mark state as final. Otherwise do
947  * actual processing.
948  */
949  if (trgmNFA->overflowed)
950  state->flags |= TSTATE_FIN;
951  else
952  processState(trgmNFA, state);
953 
954  /* Did we overflow? */
955  if (trgmNFA->arcsCount > MAX_EXPANDED_ARCS ||
957  trgmNFA->overflowed = true;
958  }
959 }
960 
961 /*
962  * Process one state: add enter keys and then add outgoing arcs.
963  */
964 static void
966 {
967  /* keysQueue should be NIL already, but make sure */
968  trgmNFA->keysQueue = NIL;
969 
970  /*
971  * Add state's own key, and then process all keys added to keysQueue until
972  * queue is empty. But we can quit if the state gets marked final.
973  */
974  addKey(trgmNFA, state, &state->stateKey);
975  while (trgmNFA->keysQueue != NIL && !(state->flags & TSTATE_FIN))
976  {
978 
979  trgmNFA->keysQueue = list_delete_first(trgmNFA->keysQueue);
980  addKey(trgmNFA, state, key);
981  }
982 
983  /*
984  * Add outgoing arcs only if state isn't final (we have no interest in
985  * outgoing arcs if we already match)
986  */
987  if (!(state->flags & TSTATE_FIN))
988  addArcs(trgmNFA, state);
989 }
990 
991 /*
992  * Add the given enter key into the state's enterKeys list, and determine
993  * whether this should result in any further enter keys being added.
994  * If so, add those keys to keysQueue so that processState will handle them.
995  *
996  * If the enter key is for the NFA's final state, mark state as TSTATE_FIN.
997  * This situation means that we can reach the final state from this expanded
998  * state without reading any predictable trigram, so we must consider this
999  * state as an accepting one.
1000  *
1001  * The given key could be a duplicate of one already in enterKeys, or be
1002  * redundant with some enterKeys. So we check that before doing anything.
1003  *
1004  * Note that we don't generate any actual arcs here. addArcs will do that
1005  * later, after we have identified all the enter keys for this state.
1006  */
1007 static void
1009 {
1010  regex_arc_t *arcs;
1011  TrgmStateKey destKey;
1012  ListCell *cell;
1013  int i,
1014  arcsCount;
1015 
1016  /*
1017  * Ensure any pad bytes in destKey are zero, since it may get used as a
1018  * hashtable key by getState.
1019  */
1020  MemSet(&destKey, 0, sizeof(destKey));
1021 
1022  /*
1023  * Compare key to each existing enter key of the state to check for
1024  * redundancy. We can drop either old key(s) or the new key if we find
1025  * redundancy.
1026  */
1027  foreach(cell, state->enterKeys)
1028  {
1029  TrgmStateKey *existingKey = (TrgmStateKey *) lfirst(cell);
1030 
1031  if (existingKey->nstate == key->nstate)
1032  {
1033  if (prefixContains(&existingKey->prefix, &key->prefix))
1034  {
1035  /* This old key already covers the new key. Nothing to do */
1036  return;
1037  }
1038  if (prefixContains(&key->prefix, &existingKey->prefix))
1039  {
1040  /*
1041  * The new key covers this old key. Remove the old key, it's
1042  * no longer needed once we add this key to the list.
1043  */
1044  state->enterKeys = foreach_delete_current(state->enterKeys,
1045  cell);
1046  }
1047  }
1048  }
1049 
1050  /* No redundancy, so add this key to the state's list */
1051  state->enterKeys = lappend(state->enterKeys, key);
1052 
1053  /* If state is now known final, mark it and we're done */
1054  if (key->nstate == pg_reg_getfinalstate(trgmNFA->regex))
1055  {
1056  state->flags |= TSTATE_FIN;
1057  return;
1058  }
1059 
1060  /*
1061  * Loop through all outgoing arcs of the corresponding state in the
1062  * original NFA.
1063  */
1064  arcsCount = pg_reg_getnumoutarcs(trgmNFA->regex, key->nstate);
1065  arcs = (regex_arc_t *) palloc(sizeof(regex_arc_t) * arcsCount);
1066  pg_reg_getoutarcs(trgmNFA->regex, key->nstate, arcs, arcsCount);
1067 
1068  for (i = 0; i < arcsCount; i++)
1069  {
1070  regex_arc_t *arc = &arcs[i];
1071 
1072  if (pg_reg_colorisbegin(trgmNFA->regex, arc->co))
1073  {
1074  /*
1075  * Start of line/string (^). Trigram extraction treats start of
1076  * line same as start of word: double space prefix is added.
1077  * Hence, make an enter key showing we can reach the arc
1078  * destination with all-blank prefix.
1079  */
1080  destKey.prefix.colors[0] = COLOR_BLANK;
1081  destKey.prefix.colors[1] = COLOR_BLANK;
1082  destKey.nstate = arc->to;
1083 
1084  /* Add enter key to this state */
1085  addKeyToQueue(trgmNFA, &destKey);
1086  }
1087  else if (pg_reg_colorisend(trgmNFA->regex, arc->co))
1088  {
1089  /*
1090  * End of line/string ($). We must consider this arc as a
1091  * transition that doesn't read anything. The reason for adding
1092  * this enter key to the state is that if the arc leads to the
1093  * NFA's final state, we must mark this expanded state as final.
1094  */
1095  destKey.prefix.colors[0] = COLOR_UNKNOWN;
1096  destKey.prefix.colors[1] = COLOR_UNKNOWN;
1097  destKey.nstate = arc->to;
1098 
1099  /* Add enter key to this state */
1100  addKeyToQueue(trgmNFA, &destKey);
1101  }
1102  else if (arc->co >= 0)
1103  {
1104  /* Regular color (including WHITE) */
1105  TrgmColorInfo *colorInfo = &trgmNFA->colorInfo[arc->co];
1106 
1107  if (colorInfo->expandable)
1108  {
1109  if (colorInfo->containsNonWord &&
1110  !validArcLabel(key, COLOR_BLANK))
1111  {
1112  /*
1113  * We can reach the arc destination after reading a
1114  * non-word character, but the prefix is not something
1115  * that addArc will accept with COLOR_BLANK, so no trigram
1116  * arc can get made for this transition. We must make an
1117  * enter key to show that the arc destination is
1118  * reachable. Set it up with an all-blank prefix, since
1119  * that corresponds to what the trigram extraction code
1120  * will do at a word starting boundary.
1121  */
1122  destKey.prefix.colors[0] = COLOR_BLANK;
1123  destKey.prefix.colors[1] = COLOR_BLANK;
1124  destKey.nstate = arc->to;
1125  addKeyToQueue(trgmNFA, &destKey);
1126  }
1127 
1128  if (colorInfo->wordCharsCount > 0 &&
1129  !validArcLabel(key, arc->co))
1130  {
1131  /*
1132  * We can reach the arc destination after reading a word
1133  * character, but the prefix is not something that addArc
1134  * will accept, so no trigram arc can get made for this
1135  * transition. We must make an enter key to show that the
1136  * arc destination is reachable. The prefix for the enter
1137  * key should reflect the info we have for this arc.
1138  */
1139  destKey.prefix.colors[0] = key->prefix.colors[1];
1140  destKey.prefix.colors[1] = arc->co;
1141  destKey.nstate = arc->to;
1142  addKeyToQueue(trgmNFA, &destKey);
1143  }
1144  }
1145  else
1146  {
1147  /*
1148  * Unexpandable color. Add enter key with ambiguous prefix,
1149  * showing we can reach the destination from this state, but
1150  * the preceding colors will be uncertain. (We do not set the
1151  * first prefix color to key->prefix.colors[1], because a
1152  * prefix of known followed by unknown is invalid.)
1153  */
1154  destKey.prefix.colors[0] = COLOR_UNKNOWN;
1155  destKey.prefix.colors[1] = COLOR_UNKNOWN;
1156  destKey.nstate = arc->to;
1157  addKeyToQueue(trgmNFA, &destKey);
1158  }
1159  }
1160  else
1161  {
1162  /* RAINBOW: treat as unexpandable color */
1163  destKey.prefix.colors[0] = COLOR_UNKNOWN;
1164  destKey.prefix.colors[1] = COLOR_UNKNOWN;
1165  destKey.nstate = arc->to;
1166  addKeyToQueue(trgmNFA, &destKey);
1167  }
1168  }
1169 
1170  pfree(arcs);
1171 }
1172 
1173 /*
1174  * Add copy of given key to keysQueue for later processing.
1175  */
1176 static void
1178 {
1179  TrgmStateKey *keyCopy = (TrgmStateKey *) palloc(sizeof(TrgmStateKey));
1180 
1181  memcpy(keyCopy, key, sizeof(TrgmStateKey));
1182  trgmNFA->keysQueue = lappend(trgmNFA->keysQueue, keyCopy);
1183 }
1184 
1185 /*
1186  * Add outgoing arcs from given state, whose enter keys are all now known.
1187  */
1188 static void
1190 {
1191  TrgmStateKey destKey;
1192  ListCell *cell;
1193  regex_arc_t *arcs;
1194  int arcsCount,
1195  i;
1196 
1197  /*
1198  * Ensure any pad bytes in destKey are zero, since it may get used as a
1199  * hashtable key by getState.
1200  */
1201  MemSet(&destKey, 0, sizeof(destKey));
1202 
1203  /*
1204  * Iterate over enter keys associated with this expanded-graph state. This
1205  * includes both the state's own stateKey, and any enter keys we added to
1206  * it during addKey (which represent expanded-graph states that are not
1207  * distinguishable from this one by means of trigrams). For each such
1208  * enter key, examine all the out-arcs of the key's underlying NFA state,
1209  * and try to make a trigram arc leading to where the out-arc leads.
1210  * (addArc will deal with whether the arc is valid or not.)
1211  */
1212  foreach(cell, state->enterKeys)
1213  {
1214  TrgmStateKey *key = (TrgmStateKey *) lfirst(cell);
1215 
1216  arcsCount = pg_reg_getnumoutarcs(trgmNFA->regex, key->nstate);
1217  arcs = (regex_arc_t *) palloc(sizeof(regex_arc_t) * arcsCount);
1218  pg_reg_getoutarcs(trgmNFA->regex, key->nstate, arcs, arcsCount);
1219 
1220  for (i = 0; i < arcsCount; i++)
1221  {
1222  regex_arc_t *arc = &arcs[i];
1223  TrgmColorInfo *colorInfo;
1224 
1225  /*
1226  * Ignore non-expandable colors; addKey already handled the case.
1227  *
1228  * We need no special check for WHITE or begin/end pseudocolors
1229  * here. We don't need to do any processing for them, and they
1230  * will be marked non-expandable since the regex engine will have
1231  * reported them that way. We do have to watch out for RAINBOW,
1232  * which has a negative color number.
1233  */
1234  if (arc->co < 0)
1235  continue;
1236  Assert(arc->co < trgmNFA->ncolors);
1237 
1238  colorInfo = &trgmNFA->colorInfo[arc->co];
1239  if (!colorInfo->expandable)
1240  continue;
1241 
1242  if (colorInfo->containsNonWord)
1243  {
1244  /*
1245  * Color includes non-word character(s).
1246  *
1247  * Generate an arc, treating this transition as occurring on
1248  * BLANK. This allows word-ending trigrams to be manufactured
1249  * if possible.
1250  */
1251  destKey.prefix.colors[0] = key->prefix.colors[1];
1252  destKey.prefix.colors[1] = COLOR_BLANK;
1253  destKey.nstate = arc->to;
1254 
1255  addArc(trgmNFA, state, key, COLOR_BLANK, &destKey);
1256  }
1257 
1258  if (colorInfo->wordCharsCount > 0)
1259  {
1260  /*
1261  * Color includes word character(s).
1262  *
1263  * Generate an arc. Color is pushed into prefix of target
1264  * state.
1265  */
1266  destKey.prefix.colors[0] = key->prefix.colors[1];
1267  destKey.prefix.colors[1] = arc->co;
1268  destKey.nstate = arc->to;
1269 
1270  addArc(trgmNFA, state, key, arc->co, &destKey);
1271  }
1272  }
1273 
1274  pfree(arcs);
1275  }
1276 }
1277 
1278 /*
1279  * Generate an out-arc of the expanded graph, if it's valid and not redundant.
1280  *
1281  * state: expanded-graph state we want to add an out-arc to
1282  * key: provides prefix colors (key->nstate is not used)
1283  * co: transition color
1284  * destKey: identifier for destination state of expanded graph
1285  */
1286 static void
1288  TrgmColor co, TrgmStateKey *destKey)
1289 {
1290  TrgmArc *arc;
1291  ListCell *cell;
1292 
1293  /* Do nothing if this wouldn't be a valid arc label trigram */
1294  if (!validArcLabel(key, co))
1295  return;
1296 
1297  /*
1298  * Check if we are going to reach key which is covered by a key which is
1299  * already listed in this state. If so arc is useless: the NFA can bypass
1300  * it through a path that doesn't require any predictable trigram, so
1301  * whether the arc's trigram is present or not doesn't really matter.
1302  */
1303  foreach(cell, state->enterKeys)
1304  {
1305  TrgmStateKey *existingKey = (TrgmStateKey *) lfirst(cell);
1306 
1307  if (existingKey->nstate == destKey->nstate &&
1308  prefixContains(&existingKey->prefix, &destKey->prefix))
1309  return;
1310  }
1311 
1312  /* Checks were successful, add new arc */
1313  arc = (TrgmArc *) palloc(sizeof(TrgmArc));
1314  arc->target = getState(trgmNFA, destKey);
1315  arc->ctrgm.colors[0] = key->prefix.colors[0];
1316  arc->ctrgm.colors[1] = key->prefix.colors[1];
1317  arc->ctrgm.colors[2] = co;
1318 
1319  state->arcs = lappend(state->arcs, arc);
1320  trgmNFA->arcsCount++;
1321 }
1322 
1323 /*
1324  * Can we make a valid trigram arc label from the given prefix and arc color?
1325  *
1326  * This is split out so that tests in addKey and addArc will stay in sync.
1327  */
1328 static bool
1330 {
1331  /*
1332  * We have to know full trigram in order to add outgoing arc. So we can't
1333  * do it if prefix is ambiguous.
1334  */
1335  if (key->prefix.colors[0] == COLOR_UNKNOWN)
1336  return false;
1337 
1338  /* If key->prefix.colors[0] isn't unknown, its second color isn't either */
1339  Assert(key->prefix.colors[1] != COLOR_UNKNOWN);
1340  /* And we should not be called with an unknown arc color anytime */
1341  Assert(co != COLOR_UNKNOWN);
1342 
1343  /*
1344  * We don't bother with making arcs representing three non-word
1345  * characters, since that's useless for trigram extraction.
1346  */
1347  if (key->prefix.colors[0] == COLOR_BLANK &&
1348  key->prefix.colors[1] == COLOR_BLANK &&
1349  co == COLOR_BLANK)
1350  return false;
1351 
1352  /*
1353  * We also reject nonblank-blank-anything. The nonblank-blank-nonblank
1354  * case doesn't correspond to any trigram the trigram extraction code
1355  * would make. The nonblank-blank-blank case is also not possible with
1356  * RPADDING = 1. (Note that in many cases we'd fail to generate such a
1357  * trigram even if it were valid, for example processing "foo bar" will
1358  * not result in considering the trigram "o ". So if you want to support
1359  * RPADDING = 2, there's more to do than just twiddle this test.)
1360  */
1361  if (key->prefix.colors[0] != COLOR_BLANK &&
1362  key->prefix.colors[1] == COLOR_BLANK)
1363  return false;
1364 
1365  /*
1366  * Other combinations involving blank are valid, in particular we assume
1367  * blank-blank-nonblank is valid, which presumes that LPADDING is 2.
1368  *
1369  * Note: Using again the example "foo bar", we will not consider the
1370  * trigram " b", though this trigram would be found by the trigram
1371  * extraction code. Since we will find " ba", it doesn't seem worth
1372  * trying to hack the algorithm to generate the additional trigram.
1373  */
1374 
1375  /* arc label is valid */
1376  return true;
1377 }
1378 
1379 /*
1380  * Get state of expanded graph for given state key,
1381  * and queue the state for processing if it didn't already exist.
1382  */
1383 static TrgmState *
1385 {
1386  TrgmState *state;
1387  bool found;
1388 
1389  state = (TrgmState *) hash_search(trgmNFA->states, key, HASH_ENTER,
1390  &found);
1391  if (!found)
1392  {
1393  /* New state: initialize and queue it */
1394  state->arcs = NIL;
1395  state->enterKeys = NIL;
1396  state->flags = 0;
1397  /* states are initially given negative numbers */
1398  state->snumber = -(++trgmNFA->nstates);
1399  state->parent = NULL;
1400  state->tentFlags = 0;
1401  state->tentParent = NULL;
1402 
1403  trgmNFA->queue = lappend(trgmNFA->queue, state);
1404  }
1405  return state;
1406 }
1407 
1408 /*
1409  * Check if prefix1 "contains" prefix2.
1410  *
1411  * "contains" means that any exact prefix (with no ambiguity) that satisfies
1412  * prefix2 also satisfies prefix1.
1413  */
1414 static bool
1416 {
1417  if (prefix1->colors[1] == COLOR_UNKNOWN)
1418  {
1419  /* Fully ambiguous prefix contains everything */
1420  return true;
1421  }
1422  else if (prefix1->colors[0] == COLOR_UNKNOWN)
1423  {
1424  /*
1425  * Prefix with only first unknown color contains every prefix with
1426  * same second color.
1427  */
1428  if (prefix1->colors[1] == prefix2->colors[1])
1429  return true;
1430  else
1431  return false;
1432  }
1433  else
1434  {
1435  /* Exact prefix contains only the exact same prefix */
1436  if (prefix1->colors[0] == prefix2->colors[0] &&
1437  prefix1->colors[1] == prefix2->colors[1])
1438  return true;
1439  else
1440  return false;
1441  }
1442 }
1443 
1444 
1445 /*---------------------
1446  * Subroutines for expanding color trigrams into regular trigrams (stage 3).
1447  *---------------------
1448  */
1449 
1450 /*
1451  * Get vector of all color trigrams in graph and select which of them
1452  * to expand into simple trigrams.
1453  *
1454  * Returns true if OK, false if exhausted resource limits.
1455  */
1456 static bool
1458 {
1459  HASH_SEQ_STATUS scan_status;
1460  int arcsCount = trgmNFA->arcsCount,
1461  i;
1462  TrgmState *state;
1463  ColorTrgmInfo *colorTrgms;
1464  int64 totalTrgmCount;
1465  float4 totalTrgmPenalty;
1466  int cnumber;
1467 
1468  /* Collect color trigrams from all arcs */
1469  colorTrgms = (ColorTrgmInfo *) palloc0(sizeof(ColorTrgmInfo) * arcsCount);
1470  trgmNFA->colorTrgms = colorTrgms;
1471 
1472  i = 0;
1473  hash_seq_init(&scan_status, trgmNFA->states);
1474  while ((state = (TrgmState *) hash_seq_search(&scan_status)) != NULL)
1475  {
1476  ListCell *cell;
1477 
1478  foreach(cell, state->arcs)
1479  {
1480  TrgmArc *arc = (TrgmArc *) lfirst(cell);
1481  TrgmArcInfo *arcInfo = (TrgmArcInfo *) palloc(sizeof(TrgmArcInfo));
1482  ColorTrgmInfo *trgmInfo = &colorTrgms[i];
1483 
1484  arcInfo->source = state;
1485  arcInfo->target = arc->target;
1486  trgmInfo->ctrgm = arc->ctrgm;
1487  trgmInfo->cnumber = -1;
1488  /* count and penalty will be set below */
1489  trgmInfo->expanded = true;
1490  trgmInfo->arcs = list_make1(arcInfo);
1491  i++;
1492  }
1493  }
1494  Assert(i == arcsCount);
1495 
1496  /* Remove duplicates, merging their arcs lists */
1497  if (arcsCount >= 2)
1498  {
1499  ColorTrgmInfo *p1,
1500  *p2;
1501 
1502  /* Sort trigrams to ease duplicate detection */
1503  qsort(colorTrgms, arcsCount, sizeof(ColorTrgmInfo), colorTrgmInfoCmp);
1504 
1505  /* p1 is probe point, p2 is last known non-duplicate. */
1506  p2 = colorTrgms;
1507  for (p1 = colorTrgms + 1; p1 < colorTrgms + arcsCount; p1++)
1508  {
1509  if (colorTrgmInfoCmp(p1, p2) > 0)
1510  {
1511  p2++;
1512  *p2 = *p1;
1513  }
1514  else
1515  {
1516  p2->arcs = list_concat(p2->arcs, p1->arcs);
1517  }
1518  }
1519  trgmNFA->colorTrgmsCount = (p2 - colorTrgms) + 1;
1520  }
1521  else
1522  {
1523  trgmNFA->colorTrgmsCount = arcsCount;
1524  }
1525 
1526  /*
1527  * Count number of simple trigrams generated by each color trigram, and
1528  * also compute a penalty value, which is the number of simple trigrams
1529  * times a multiplier that depends on its whitespace content.
1530  *
1531  * Note: per-color-trigram counts cannot overflow an int so long as
1532  * COLOR_COUNT_LIMIT is not more than the cube root of INT_MAX, ie about
1533  * 1290. However, the grand total totalTrgmCount might conceivably
1534  * overflow an int, so we use int64 for that within this routine. Also,
1535  * penalties are calculated in float4 arithmetic to avoid any overflow
1536  * worries.
1537  */
1538  totalTrgmCount = 0;
1539  totalTrgmPenalty = 0.0f;
1540  for (i = 0; i < trgmNFA->colorTrgmsCount; i++)
1541  {
1542  ColorTrgmInfo *trgmInfo = &colorTrgms[i];
1543  int j,
1544  count = 1,
1545  typeIndex = 0;
1546 
1547  for (j = 0; j < 3; j++)
1548  {
1549  TrgmColor c = trgmInfo->ctrgm.colors[j];
1550 
1551  typeIndex *= 2;
1552  if (c == COLOR_BLANK)
1553  typeIndex++;
1554  else
1555  count *= trgmNFA->colorInfo[c].wordCharsCount;
1556  }
1557  trgmInfo->count = count;
1558  totalTrgmCount += count;
1559  trgmInfo->penalty = penalties[typeIndex] * (float4) count;
1560  totalTrgmPenalty += trgmInfo->penalty;
1561  }
1562 
1563  /* Sort color trigrams in descending order of their penalties */
1564  qsort(colorTrgms, trgmNFA->colorTrgmsCount, sizeof(ColorTrgmInfo),
1566 
1567  /*
1568  * Remove color trigrams from the graph so long as total penalty of color
1569  * trigrams exceeds WISH_TRGM_PENALTY. (If we fail to get down to
1570  * WISH_TRGM_PENALTY, it's OK so long as total count is no more than
1571  * MAX_TRGM_COUNT.) We prefer to remove color trigrams with higher
1572  * penalty, since those are the most promising for reducing the total
1573  * penalty. When removing a color trigram we have to merge states
1574  * connected by arcs labeled with that trigram. It's necessary to not
1575  * merge initial and final states, because our graph becomes useless if
1576  * that happens; so we cannot always remove the trigram we'd prefer to.
1577  */
1578  for (i = 0; i < trgmNFA->colorTrgmsCount; i++)
1579  {
1580  ColorTrgmInfo *trgmInfo = &colorTrgms[i];
1581  bool canRemove = true;
1582  ListCell *cell;
1583 
1584  /* Done if we've reached the target */
1585  if (totalTrgmPenalty <= WISH_TRGM_PENALTY)
1586  break;
1587 
1588 #ifdef TRGM_REGEXP_DEBUG
1589  fprintf(stderr, "considering ctrgm %d %d %d, penalty %f, %d arcs\n",
1590  trgmInfo->ctrgm.colors[0],
1591  trgmInfo->ctrgm.colors[1],
1592  trgmInfo->ctrgm.colors[2],
1593  trgmInfo->penalty,
1594  list_length(trgmInfo->arcs));
1595 #endif
1596 
1597  /*
1598  * Does any arc of this color trigram connect initial and final
1599  * states? If so we can't remove it.
1600  */
1601  foreach(cell, trgmInfo->arcs)
1602  {
1603  TrgmArcInfo *arcInfo = (TrgmArcInfo *) lfirst(cell);
1604  TrgmState *source = arcInfo->source,
1605  *target = arcInfo->target;
1606  int source_flags,
1607  target_flags;
1608 
1609 #ifdef TRGM_REGEXP_DEBUG
1610  fprintf(stderr, "examining arc to s%d (%x) from s%d (%x)\n",
1611  -target->snumber, target->flags,
1612  -source->snumber, source->flags);
1613 #endif
1614 
1615  /* examine parent states, if any merging has already happened */
1616  while (source->parent)
1617  source = source->parent;
1618  while (target->parent)
1619  target = target->parent;
1620 
1621 #ifdef TRGM_REGEXP_DEBUG
1622  fprintf(stderr, " ... after completed merges: to s%d (%x) from s%d (%x)\n",
1623  -target->snumber, target->flags,
1624  -source->snumber, source->flags);
1625 #endif
1626 
1627  /* we must also consider merges we are planning right now */
1628  source_flags = source->flags | source->tentFlags;
1629  while (source->tentParent)
1630  {
1631  source = source->tentParent;
1632  source_flags |= source->flags | source->tentFlags;
1633  }
1634  target_flags = target->flags | target->tentFlags;
1635  while (target->tentParent)
1636  {
1637  target = target->tentParent;
1638  target_flags |= target->flags | target->tentFlags;
1639  }
1640 
1641 #ifdef TRGM_REGEXP_DEBUG
1642  fprintf(stderr, " ... after tentative merges: to s%d (%x) from s%d (%x)\n",
1643  -target->snumber, target_flags,
1644  -source->snumber, source_flags);
1645 #endif
1646 
1647  /* would fully-merged state have both INIT and FIN set? */
1648  if (((source_flags | target_flags) & (TSTATE_INIT | TSTATE_FIN)) ==
1649  (TSTATE_INIT | TSTATE_FIN))
1650  {
1651  canRemove = false;
1652  break;
1653  }
1654 
1655  /* ok so far, so remember planned merge */
1656  if (source != target)
1657  {
1658 #ifdef TRGM_REGEXP_DEBUG
1659  fprintf(stderr, " ... tentatively merging s%d into s%d\n",
1660  -target->snumber, -source->snumber);
1661 #endif
1662  target->tentParent = source;
1663  source->tentFlags |= target_flags;
1664  }
1665  }
1666 
1667  /*
1668  * We must reset all the tentFlags/tentParent fields before
1669  * continuing. tentFlags could only have become set in states that
1670  * are the source or parent or tentative parent of one of the current
1671  * arcs; likewise tentParent could only have become set in states that
1672  * are the target or parent or tentative parent of one of the current
1673  * arcs. There might be some overlap between those sets, but if we
1674  * clear tentFlags in target states as well as source states, we
1675  * should be okay even if we visit a state as target before visiting
1676  * it as a source.
1677  */
1678  foreach(cell, trgmInfo->arcs)
1679  {
1680  TrgmArcInfo *arcInfo = (TrgmArcInfo *) lfirst(cell);
1681  TrgmState *source = arcInfo->source,
1682  *target = arcInfo->target;
1683  TrgmState *ttarget;
1684 
1685  /* no need to touch previously-merged states */
1686  while (source->parent)
1687  source = source->parent;
1688  while (target->parent)
1689  target = target->parent;
1690 
1691  while (source)
1692  {
1693  source->tentFlags = 0;
1694  source = source->tentParent;
1695  }
1696 
1697  while ((ttarget = target->tentParent) != NULL)
1698  {
1699  target->tentParent = NULL;
1700  target->tentFlags = 0; /* in case it was also a source */
1701  target = ttarget;
1702  }
1703  }
1704 
1705  /* Now, move on if we can't drop this trigram */
1706  if (!canRemove)
1707  {
1708 #ifdef TRGM_REGEXP_DEBUG
1709  fprintf(stderr, " ... not ok to merge\n");
1710 #endif
1711  continue;
1712  }
1713 
1714  /* OK, merge states linked by each arc labeled by the trigram */
1715  foreach(cell, trgmInfo->arcs)
1716  {
1717  TrgmArcInfo *arcInfo = (TrgmArcInfo *) lfirst(cell);
1718  TrgmState *source = arcInfo->source,
1719  *target = arcInfo->target;
1720 
1721  while (source->parent)
1722  source = source->parent;
1723  while (target->parent)
1724  target = target->parent;
1725  if (source != target)
1726  {
1727 #ifdef TRGM_REGEXP_DEBUG
1728  fprintf(stderr, "merging s%d into s%d\n",
1729  -target->snumber, -source->snumber);
1730 #endif
1731  mergeStates(source, target);
1732  /* Assert we didn't merge initial and final states */
1733  Assert((source->flags & (TSTATE_INIT | TSTATE_FIN)) !=
1734  (TSTATE_INIT | TSTATE_FIN));
1735  }
1736  }
1737 
1738  /* Mark trigram unexpanded, and update totals */
1739  trgmInfo->expanded = false;
1740  totalTrgmCount -= trgmInfo->count;
1741  totalTrgmPenalty -= trgmInfo->penalty;
1742  }
1743 
1744  /* Did we succeed in fitting into MAX_TRGM_COUNT? */
1745  if (totalTrgmCount > MAX_TRGM_COUNT)
1746  return false;
1747 
1748  trgmNFA->totalTrgmCount = (int) totalTrgmCount;
1749 
1750  /*
1751  * Sort color trigrams by colors (will be useful for bsearch in packGraph)
1752  * and enumerate the color trigrams that are expanded.
1753  */
1754  cnumber = 0;
1755  qsort(colorTrgms, trgmNFA->colorTrgmsCount, sizeof(ColorTrgmInfo),
1757  for (i = 0; i < trgmNFA->colorTrgmsCount; i++)
1758  {
1759  if (colorTrgms[i].expanded)
1760  {
1761  colorTrgms[i].cnumber = cnumber;
1762  cnumber++;
1763  }
1764  }
1765 
1766  return true;
1767 }
1768 
1769 /*
1770  * Expand selected color trigrams into regular trigrams.
1771  *
1772  * Returns the TRGM array to be passed to the index machinery.
1773  * The array must be allocated in rcontext.
1774  */
1775 static TRGM *
1777 {
1778  TRGM *trg;
1779  trgm *p;
1780  int i;
1781  TrgmColorInfo blankColor;
1782  trgm_mb_char blankChar;
1783 
1784  /* Set up "blank" color structure containing a single zero character */
1785  memset(blankChar.bytes, 0, sizeof(blankChar.bytes));
1786  blankColor.wordCharsCount = 1;
1787  blankColor.wordChars = &blankChar;
1788 
1789  /* Construct the trgm array */
1790  trg = (TRGM *)
1791  MemoryContextAllocZero(rcontext,
1792  TRGMHDRSIZE +
1793  trgmNFA->totalTrgmCount * sizeof(trgm));
1794  trg->flag = ARRKEY;
1795  SET_VARSIZE(trg, CALCGTSIZE(ARRKEY, trgmNFA->totalTrgmCount));
1796  p = GETARR(trg);
1797  for (i = 0; i < trgmNFA->colorTrgmsCount; i++)
1798  {
1799  ColorTrgmInfo *colorTrgm = &trgmNFA->colorTrgms[i];
1800  TrgmColorInfo *c[3];
1801  trgm_mb_char s[3];
1802  int j,
1803  i1,
1804  i2,
1805  i3;
1806 
1807  /* Ignore any unexpanded trigrams ... */
1808  if (!colorTrgm->expanded)
1809  continue;
1810 
1811  /* Get colors, substituting the dummy struct for COLOR_BLANK */
1812  for (j = 0; j < 3; j++)
1813  {
1814  if (colorTrgm->ctrgm.colors[j] != COLOR_BLANK)
1815  c[j] = &trgmNFA->colorInfo[colorTrgm->ctrgm.colors[j]];
1816  else
1817  c[j] = &blankColor;
1818  }
1819 
1820  /* Iterate over all possible combinations of colors' characters */
1821  for (i1 = 0; i1 < c[0]->wordCharsCount; i1++)
1822  {
1823  s[0] = c[0]->wordChars[i1];
1824  for (i2 = 0; i2 < c[1]->wordCharsCount; i2++)
1825  {
1826  s[1] = c[1]->wordChars[i2];
1827  for (i3 = 0; i3 < c[2]->wordCharsCount; i3++)
1828  {
1829  s[2] = c[2]->wordChars[i3];
1830  fillTrgm(p, s);
1831  p++;
1832  }
1833  }
1834  }
1835  }
1836 
1837  return trg;
1838 }
1839 
1840 /*
1841  * Convert trigram into trgm datatype.
1842  */
1843 static void
1845 {
1846  char str[3 * MAX_MULTIBYTE_CHAR_LEN],
1847  *p;
1848  int i,
1849  j;
1850 
1851  /* Write multibyte string into "str" (we don't need null termination) */
1852  p = str;
1853 
1854  for (i = 0; i < 3; i++)
1855  {
1856  if (s[i].bytes[0] != 0)
1857  {
1858  for (j = 0; j < MAX_MULTIBYTE_CHAR_LEN && s[i].bytes[j]; j++)
1859  *p++ = s[i].bytes[j];
1860  }
1861  else
1862  {
1863  /* Emit a space in place of COLOR_BLANK */
1864  *p++ = ' ';
1865  }
1866  }
1867 
1868  /* Convert "str" to a standard trigram (possibly hashing it) */
1869  compact_trigram(ptrgm, str, p - str);
1870 }
1871 
1872 /*
1873  * Merge two states of graph.
1874  */
1875 static void
1877 {
1878  Assert(state1 != state2);
1879  Assert(!state1->parent);
1880  Assert(!state2->parent);
1881 
1882  /* state1 absorbs state2's flags */
1883  state1->flags |= state2->flags;
1884 
1885  /* state2, and indirectly all its children, become children of state1 */
1886  state2->parent = state1;
1887 }
1888 
1889 /*
1890  * Compare function for sorting of color trigrams by their colors.
1891  */
1892 static int
1893 colorTrgmInfoCmp(const void *p1, const void *p2)
1894 {
1895  const ColorTrgmInfo *c1 = (const ColorTrgmInfo *) p1;
1896  const ColorTrgmInfo *c2 = (const ColorTrgmInfo *) p2;
1897 
1898  return memcmp(&c1->ctrgm, &c2->ctrgm, sizeof(ColorTrgm));
1899 }
1900 
1901 /*
1902  * Compare function for sorting color trigrams in descending order of
1903  * their penalty fields.
1904  */
1905 static int
1906 colorTrgmInfoPenaltyCmp(const void *p1, const void *p2)
1907 {
1908  float4 penalty1 = ((const ColorTrgmInfo *) p1)->penalty;
1909  float4 penalty2 = ((const ColorTrgmInfo *) p2)->penalty;
1910 
1911  if (penalty1 < penalty2)
1912  return 1;
1913  else if (penalty1 == penalty2)
1914  return 0;
1915  else
1916  return -1;
1917 }
1918 
1919 
1920 /*---------------------
1921  * Subroutines for packing the graph into final representation (stage 4).
1922  *---------------------
1923  */
1924 
1925 /*
1926  * Pack expanded graph into final representation.
1927  *
1928  * The result data must be allocated in rcontext.
1929  */
1930 static TrgmPackedGraph *
1931 packGraph(TrgmNFA *trgmNFA, MemoryContext rcontext)
1932 {
1933  int snumber = 2,
1934  arcIndex,
1935  arcsCount;
1936  HASH_SEQ_STATUS scan_status;
1937  TrgmState *state;
1939  *p1,
1940  *p2;
1941  TrgmPackedArc *packedArcs;
1942  TrgmPackedGraph *result;
1943  int i,
1944  j;
1945 
1946  /* Enumerate surviving states, giving init and fin reserved numbers */
1947  hash_seq_init(&scan_status, trgmNFA->states);
1948  while ((state = (TrgmState *) hash_seq_search(&scan_status)) != NULL)
1949  {
1950  while (state->parent)
1951  state = state->parent;
1952 
1953  if (state->snumber < 0)
1954  {
1955  if (state->flags & TSTATE_INIT)
1956  state->snumber = 0;
1957  else if (state->flags & TSTATE_FIN)
1958  state->snumber = 1;
1959  else
1960  {
1961  state->snumber = snumber;
1962  snumber++;
1963  }
1964  }
1965  }
1966 
1967  /* Collect array of all arcs */
1968  arcs = (TrgmPackArcInfo *)
1969  palloc(sizeof(TrgmPackArcInfo) * trgmNFA->arcsCount);
1970  arcIndex = 0;
1971  hash_seq_init(&scan_status, trgmNFA->states);
1972  while ((state = (TrgmState *) hash_seq_search(&scan_status)) != NULL)
1973  {
1974  TrgmState *source = state;
1975  ListCell *cell;
1976 
1977  while (source->parent)
1978  source = source->parent;
1979 
1980  foreach(cell, state->arcs)
1981  {
1982  TrgmArc *arc = (TrgmArc *) lfirst(cell);
1983  TrgmState *target = arc->target;
1984 
1985  while (target->parent)
1986  target = target->parent;
1987 
1988  if (source->snumber != target->snumber)
1989  {
1990  ColorTrgmInfo *ctrgm;
1991 
1992  ctrgm = (ColorTrgmInfo *) bsearch(&arc->ctrgm,
1993  trgmNFA->colorTrgms,
1994  trgmNFA->colorTrgmsCount,
1995  sizeof(ColorTrgmInfo),
1997  Assert(ctrgm != NULL);
1998  Assert(ctrgm->expanded);
1999 
2000  arcs[arcIndex].sourceState = source->snumber;
2001  arcs[arcIndex].targetState = target->snumber;
2002  arcs[arcIndex].colorTrgm = ctrgm->cnumber;
2003  arcIndex++;
2004  }
2005  }
2006  }
2007 
2008  /* Sort arcs to ease duplicate detection */
2009  qsort(arcs, arcIndex, sizeof(TrgmPackArcInfo), packArcInfoCmp);
2010 
2011  /* We could have duplicates because states were merged. Remove them. */
2012  /* p1 is probe point, p2 is last known non-duplicate. */
2013  p2 = arcs;
2014  for (p1 = arcs + 1; p1 < arcs + arcIndex; p1++)
2015  {
2016  if (packArcInfoCmp(p1, p2) > 0)
2017  {
2018  p2++;
2019  *p2 = *p1;
2020  }
2021  }
2022  arcsCount = (p2 - arcs) + 1;
2023 
2024  /* Create packed representation */
2025  result = (TrgmPackedGraph *)
2026  MemoryContextAlloc(rcontext, sizeof(TrgmPackedGraph));
2027 
2028  /* Pack color trigrams information */
2029  result->colorTrigramsCount = 0;
2030  for (i = 0; i < trgmNFA->colorTrgmsCount; i++)
2031  {
2032  if (trgmNFA->colorTrgms[i].expanded)
2033  result->colorTrigramsCount++;
2034  }
2035  result->colorTrigramGroups = (int *)
2036  MemoryContextAlloc(rcontext, sizeof(int) * result->colorTrigramsCount);
2037  j = 0;
2038  for (i = 0; i < trgmNFA->colorTrgmsCount; i++)
2039  {
2040  if (trgmNFA->colorTrgms[i].expanded)
2041  {
2042  result->colorTrigramGroups[j] = trgmNFA->colorTrgms[i].count;
2043  j++;
2044  }
2045  }
2046 
2047  /* Pack states and arcs information */
2048  result->statesCount = snumber;
2049  result->states = (TrgmPackedState *)
2050  MemoryContextAlloc(rcontext, snumber * sizeof(TrgmPackedState));
2051  packedArcs = (TrgmPackedArc *)
2052  MemoryContextAlloc(rcontext, arcsCount * sizeof(TrgmPackedArc));
2053  j = 0;
2054  for (i = 0; i < snumber; i++)
2055  {
2056  int cnt = 0;
2057 
2058  result->states[i].arcs = &packedArcs[j];
2059  while (j < arcsCount && arcs[j].sourceState == i)
2060  {
2061  packedArcs[j].targetState = arcs[j].targetState;
2062  packedArcs[j].colorTrgm = arcs[j].colorTrgm;
2063  cnt++;
2064  j++;
2065  }
2066  result->states[i].arcsCount = cnt;
2067  }
2068 
2069  /* Allocate working memory for trigramsMatchGraph() */
2070  result->colorTrigramsActive = (bool *)
2071  MemoryContextAlloc(rcontext, sizeof(bool) * result->colorTrigramsCount);
2072  result->statesActive = (bool *)
2073  MemoryContextAlloc(rcontext, sizeof(bool) * result->statesCount);
2074  result->statesQueue = (int *)
2075  MemoryContextAlloc(rcontext, sizeof(int) * result->statesCount);
2076 
2077  return result;
2078 }
2079 
2080 /*
2081  * Comparison function for sorting TrgmPackArcInfos.
2082  *
2083  * Compares arcs in following order: sourceState, colorTrgm, targetState.
2084  */
2085 static int
2086 packArcInfoCmp(const void *a1, const void *a2)
2087 {
2088  const TrgmPackArcInfo *p1 = (const TrgmPackArcInfo *) a1;
2089  const TrgmPackArcInfo *p2 = (const TrgmPackArcInfo *) a2;
2090 
2091  if (p1->sourceState < p2->sourceState)
2092  return -1;
2093  if (p1->sourceState > p2->sourceState)
2094  return 1;
2095  if (p1->colorTrgm < p2->colorTrgm)
2096  return -1;
2097  if (p1->colorTrgm > p2->colorTrgm)
2098  return 1;
2099  if (p1->targetState < p2->targetState)
2100  return -1;
2101  if (p1->targetState > p2->targetState)
2102  return 1;
2103  return 0;
2104 }
2105 
2106 
2107 /*---------------------
2108  * Debugging functions
2109  *
2110  * These are designed to emit GraphViz files.
2111  *---------------------
2112  */
2113 
2114 #ifdef TRGM_REGEXP_DEBUG
2115 
2116 /*
2117  * Print initial NFA, in regexp library's representation
2118  */
2119 static void
2120 printSourceNFA(regex_t *regex, TrgmColorInfo *colors, int ncolors)
2121 {
2123  int nstates = pg_reg_getnumstates(regex);
2124  int state;
2125  int i;
2126 
2127  initStringInfo(&buf);
2128 
2129  appendStringInfoString(&buf, "\ndigraph sourceNFA {\n");
2130 
2131  for (state = 0; state < nstates; state++)
2132  {
2133  regex_arc_t *arcs;
2134  int i,
2135  arcsCount;
2136 
2137  appendStringInfo(&buf, "s%d", state);
2138  if (pg_reg_getfinalstate(regex) == state)
2139  appendStringInfoString(&buf, " [shape = doublecircle]");
2140  appendStringInfoString(&buf, ";\n");
2141 
2142  arcsCount = pg_reg_getnumoutarcs(regex, state);
2143  arcs = (regex_arc_t *) palloc(sizeof(regex_arc_t) * arcsCount);
2144  pg_reg_getoutarcs(regex, state, arcs, arcsCount);
2145 
2146  for (i = 0; i < arcsCount; i++)
2147  {
2148  appendStringInfo(&buf, " s%d -> s%d [label = \"%d\"];\n",
2149  state, arcs[i].to, arcs[i].co);
2150  }
2151 
2152  pfree(arcs);
2153  }
2154 
2155  appendStringInfoString(&buf, " node [shape = point ]; initial;\n");
2156  appendStringInfo(&buf, " initial -> s%d;\n",
2157  pg_reg_getinitialstate(regex));
2158 
2159  /* Print colors */
2160  appendStringInfoString(&buf, " { rank = sink;\n");
2161  appendStringInfoString(&buf, " Colors [shape = none, margin=0, label=<\n");
2162 
2163  for (i = 0; i < ncolors; i++)
2164  {
2165  TrgmColorInfo *color = &colors[i];
2166  int j;
2167 
2168  appendStringInfo(&buf, "<br/>Color %d: ", i);
2169  if (color->expandable)
2170  {
2171  for (j = 0; j < color->wordCharsCount; j++)
2172  {
2173  char s[MAX_MULTIBYTE_CHAR_LEN + 1];
2174 
2175  memcpy(s, color->wordChars[j].bytes, MAX_MULTIBYTE_CHAR_LEN);
2176  s[MAX_MULTIBYTE_CHAR_LEN] = '\0';
2177  appendStringInfoString(&buf, s);
2178  }
2179  }
2180  else
2181  appendStringInfoString(&buf, "not expandable");
2182  appendStringInfoChar(&buf, '\n');
2183  }
2184 
2185  appendStringInfoString(&buf, " >];\n");
2186  appendStringInfoString(&buf, " }\n");
2187  appendStringInfoString(&buf, "}\n");
2188 
2189  {
2190  /* dot -Tpng -o /tmp/source.png < /tmp/source.gv */
2191  FILE *fp = fopen("/tmp/source.gv", "w");
2192 
2193  fprintf(fp, "%s", buf.data);
2194  fclose(fp);
2195  }
2196 
2197  pfree(buf.data);
2198 }
2199 
2200 /*
2201  * Print expanded graph.
2202  */
2203 static void
2204 printTrgmNFA(TrgmNFA *trgmNFA)
2205 {
2207  HASH_SEQ_STATUS scan_status;
2208  TrgmState *state;
2209  TrgmState *initstate = NULL;
2210 
2211  initStringInfo(&buf);
2212 
2213  appendStringInfoString(&buf, "\ndigraph transformedNFA {\n");
2214 
2215  hash_seq_init(&scan_status, trgmNFA->states);
2216  while ((state = (TrgmState *) hash_seq_search(&scan_status)) != NULL)
2217  {
2218  ListCell *cell;
2219 
2220  appendStringInfo(&buf, "s%d", -state->snumber);
2221  if (state->flags & TSTATE_FIN)
2222  appendStringInfoString(&buf, " [shape = doublecircle]");
2223  if (state->flags & TSTATE_INIT)
2224  initstate = state;
2225  appendStringInfo(&buf, " [label = \"%d\"]", state->stateKey.nstate);
2226  appendStringInfoString(&buf, ";\n");
2227 
2228  foreach(cell, state->arcs)
2229  {
2230  TrgmArc *arc = (TrgmArc *) lfirst(cell);
2231 
2232  appendStringInfo(&buf, " s%d -> s%d [label = \"",
2233  -state->snumber, -arc->target->snumber);
2234  printTrgmColor(&buf, arc->ctrgm.colors[0]);
2235  appendStringInfoChar(&buf, ' ');
2236  printTrgmColor(&buf, arc->ctrgm.colors[1]);
2237  appendStringInfoChar(&buf, ' ');
2238  printTrgmColor(&buf, arc->ctrgm.colors[2]);
2239  appendStringInfoString(&buf, "\"];\n");
2240  }
2241  }
2242 
2243  if (initstate)
2244  {
2245  appendStringInfoString(&buf, " node [shape = point ]; initial;\n");
2246  appendStringInfo(&buf, " initial -> s%d;\n", -initstate->snumber);
2247  }
2248 
2249  appendStringInfoString(&buf, "}\n");
2250 
2251  {
2252  /* dot -Tpng -o /tmp/transformed.png < /tmp/transformed.gv */
2253  FILE *fp = fopen("/tmp/transformed.gv", "w");
2254 
2255  fprintf(fp, "%s", buf.data);
2256  fclose(fp);
2257  }
2258 
2259  pfree(buf.data);
2260 }
2261 
2262 /*
2263  * Print a TrgmColor readably.
2264  */
2265 static void
2266 printTrgmColor(StringInfo buf, TrgmColor co)
2267 {
2268  if (co == COLOR_UNKNOWN)
2269  appendStringInfoChar(buf, 'u');
2270  else if (co == COLOR_BLANK)
2271  appendStringInfoChar(buf, 'b');
2272  else
2273  appendStringInfo(buf, "%d", (int) co);
2274 }
2275 
2276 /*
2277  * Print final packed representation of trigram-based expanded graph.
2278  */
2279 static void
2280 printTrgmPackedGraph(TrgmPackedGraph *packedGraph, TRGM *trigrams)
2281 {
2283  trgm *p;
2284  int i;
2285 
2286  initStringInfo(&buf);
2287 
2288  appendStringInfoString(&buf, "\ndigraph packedGraph {\n");
2289 
2290  for (i = 0; i < packedGraph->statesCount; i++)
2291  {
2292  TrgmPackedState *state = &packedGraph->states[i];
2293  int j;
2294 
2295  appendStringInfo(&buf, " s%d", i);
2296  if (i == 1)
2297  appendStringInfoString(&buf, " [shape = doublecircle]");
2298 
2299  appendStringInfo(&buf, " [label = <s%d>];\n", i);
2300 
2301  for (j = 0; j < state->arcsCount; j++)
2302  {
2303  TrgmPackedArc *arc = &state->arcs[j];
2304 
2305  appendStringInfo(&buf, " s%d -> s%d [label = \"trigram %d\"];\n",
2306  i, arc->targetState, arc->colorTrgm);
2307  }
2308  }
2309 
2310  appendStringInfoString(&buf, " node [shape = point ]; initial;\n");
2311  appendStringInfo(&buf, " initial -> s%d;\n", 0);
2312 
2313  /* Print trigrams */
2314  appendStringInfoString(&buf, " { rank = sink;\n");
2315  appendStringInfoString(&buf, " Trigrams [shape = none, margin=0, label=<\n");
2316 
2317  p = GETARR(trigrams);
2318  for (i = 0; i < packedGraph->colorTrigramsCount; i++)
2319  {
2320  int count = packedGraph->colorTrigramGroups[i];
2321  int j;
2322 
2323  appendStringInfo(&buf, "<br/>Trigram %d: ", i);
2324 
2325  for (j = 0; j < count; j++)
2326  {
2327  if (j > 0)
2328  appendStringInfoString(&buf, ", ");
2329 
2330  /*
2331  * XXX This representation is nice only for all-ASCII trigrams.
2332  */
2333  appendStringInfo(&buf, "\"%c%c%c\"", (*p)[0], (*p)[1], (*p)[2]);
2334  p++;
2335  }
2336  }
2337 
2338  appendStringInfoString(&buf, " >];\n");
2339  appendStringInfoString(&buf, " }\n");
2340  appendStringInfoString(&buf, "}\n");
2341 
2342  {
2343  /* dot -Tpng -o /tmp/packed.png < /tmp/packed.gv */
2344  FILE *fp = fopen("/tmp/packed.gv", "w");
2345 
2346  fprintf(fp, "%s", buf.data);
2347  fclose(fp);
2348  }
2349 
2350  pfree(buf.data);
2351 }
2352 
2353 #endif /* TRGM_REGEXP_DEBUG */
#define MAX_EXPANDED_ARCS
Definition: trgm_regexp.c:219
#define NIL
Definition: pg_list.h:65
static void mergeStates(TrgmState *state1, TrgmState *state2)
Definition: trgm_regexp.c:1876
bool * colorTrigramsActive
Definition: trgm_regexp.c:461
int pg_reg_getnumstates(const regex_t *regex)
Definition: regexport.c:36
void pg_reg_getoutarcs(const regex_t *regex, int st, regex_arc_t *arcs, int arcs_len)
Definition: regexport.c:155
struct TrgmState TrgmState
void MemoryContextDelete(MemoryContext context)
Definition: mcxt.c:218
#define AllocSetContextCreate
Definition: memutils.h:173
static void getColorInfo(regex_t *regex, TrgmNFA *trgmNFA)
Definition: trgm_regexp.c:773
ColorTrgm ctrgm
Definition: trgm_regexp.c:348
#define VARDATA_ANY(PTR)
Definition: postgres.h:361
static bool prefixContains(TrgmPrefix *prefix1, TrgmPrefix *prefix2)
Definition: trgm_regexp.c:1415
#define HASH_CONTEXT
Definition: hsearch.h:102
#define HASH_ELEM
Definition: hsearch.h:95
struct TrgmState * tentParent
Definition: trgm_regexp.c:340
MemoryContext hcxt
Definition: hsearch.h:86
static void addArc(TrgmNFA *trgmNFA, TrgmState *state, TrgmStateKey *key, TrgmColor co, TrgmStateKey *destKey)
Definition: trgm_regexp.c:1287
int pg_reg_getfinalstate(const regex_t *regex)
Definition: regexport.c:64
TrgmColorInfo * colorInfo
Definition: trgm_regexp.c:403
static TRGM * createTrgmNFAInternal(regex_t *regex, TrgmPackedGraph **graph, MemoryContext rcontext)
Definition: trgm_regexp.c:575
int pg_reg_getnumcharacters(const regex_t *regex, int co)
Definition: regexport.c:230
regex_t * regex
Definition: trgm_regexp.c:402
float4 penalty
Definition: trgm_regexp.c:377
Definition: regguts.h:290
def bytes(source, encoding='ascii', errors='strict')
short color
Definition: regguts.h:150
TrgmColor colors[3]
Definition: trgm_regexp.c:299
static MemoryContext MemoryContextSwitchTo(MemoryContext context)
Definition: palloc.h:109
HTAB * states
Definition: trgm_regexp.c:407
static void addKey(TrgmNFA *trgmNFA, TrgmState *state, TrgmStateKey *key)
Definition: trgm_regexp.c:1008
Size entrysize
Definition: hsearch.h:76
void pg_reg_getcharacters(const regex_t *regex, int co, pg_wchar *chars, int chars_len)
Definition: regexport.c:266
#define ISWORDCHR(c)
Definition: trgm.h:55
#define WISH_TRGM_PENALTY
Definition: trgm_regexp.c:221
#define ARRKEY
Definition: trgm.h:97
List * list_concat(List *list1, const List *list2)
Definition: list.c:530
int errcode(int sqlerrcode)
Definition: elog.c:698
#define MemSet(start, val, len)
Definition: c.h:1008
long hash_get_num_entries(HTAB *hashp)
Definition: dynahash.c:1382
int pg_reg_colorisbegin(const regex_t *regex, int co)
Definition: regexport.c:191
TrgmPackedState * states
Definition: trgm_regexp.c:458
TrgmState * initState
Definition: trgm_regexp.c:408
#define TSTATE_FIN
Definition: trgm_regexp.c:329
char * lowerstr(const char *str)
Definition: ts_locale.c:244
void * hash_search(HTAB *hashp, const void *keyPtr, HASHACTION action, bool *foundPtr)
Definition: dynahash.c:954
unsigned int Oid
Definition: postgres_ext.h:31
#define TSTATE_INIT
Definition: trgm_regexp.c:328
int totalTrgmCount
Definition: trgm_regexp.c:420
#define fprintf
Definition: port.h:220
bool overflowed
Definition: trgm_regexp.c:415
#define REG_ICASE
Definition: regex.h:108
int pg_regcomp(regex_t *re, const chr *string, size_t len, int flags, Oid collation)
Definition: regcomp.c:328
int pg_reg_getnumcolors(const regex_t *regex)
Definition: regexport.c:174
int nstates
Definition: trgm_regexp.c:409
#define CALCGTSIZE(flag, siglen)
Definition: hstore_gist.c:59
#define foreach_delete_current(lst, cell)
Definition: pg_list.h:369
bool containsNonWord
Definition: trgm_regexp.c:261
static bool convertPgWchar(pg_wchar c, trgm_mb_char *result)
Definition: trgm_regexp.c:837
#define list_make1(x1)
Definition: pg_list.h:206
TrgmPackedArc * arcs
Definition: trgm_regexp.c:435
static bool selectColorTrigrams(TrgmNFA *trgmNFA)
Definition: trgm_regexp.c:1457
static int packArcInfoCmp(const void *a1, const void *a2)
Definition: trgm_regexp.c:2086
#define GETARR(x)
Definition: trgm.h:107
static const FormData_pg_attribute a2
Definition: heap.c:166
Definition: dynahash.c:219
bool * statesActive
Definition: trgm_regexp.c:462
void pfree(void *pointer)
Definition: mcxt.c:1169
#define REG_OKAY
Definition: regex.h:139
void appendStringInfo(StringInfo str, const char *fmt,...)
Definition: stringinfo.c:91
#define linitial(l)
Definition: pg_list.h:174
TrgmColor colors[2]
Definition: trgm_regexp.c:290
#define ERROR
Definition: elog.h:46
Definition: trgm.h:67
int pg_reg_getinitialstate(const regex_t *regex)
Definition: regexport.c:50
#define ALLOCSET_DEFAULT_SIZES
Definition: memutils.h:195
void appendStringInfoString(StringInfo str, const char *s)
Definition: stringinfo.c:176
char * c
static char * buf
Definition: pg_test_fsync.c:68
List * enterKeys
Definition: trgm_regexp.c:335
TrgmState * target
Definition: trgm_regexp.c:360
ColorTrgmInfo * colorTrgms
Definition: trgm_regexp.c:418
HTAB * hash_create(const char *tabname, long nelem, const HASHCTL *info, int flags)
Definition: dynahash.c:349
size_t pg_regerror(int errcode, const regex_t *preg, char *errbuf, size_t errbuf_size)
Definition: regerror.c:60
bool trigramsMatchGraph(TrgmPackedGraph *graph, bool *check)
Definition: trgm_regexp.c:636
MemoryContext CurrentMemoryContext
Definition: mcxt.c:42
#define COLOR_COUNT_LIMIT
Definition: trgm_regexp.c:222
#define REG_ADVANCED
Definition: regex.h:105
TrgmState * target
Definition: trgm_regexp.c:349
List * arcs
Definition: trgm_regexp.c:334
#define TRGMHDRSIZE
Definition: trgm.h:74
unsigned int pg_wchar
Definition: mbprint.c:31
TrgmState * source
Definition: trgm_regexp.c:359
List * lappend(List *list, void *datum)
Definition: list.c:336
#define MAX_MULTIBYTE_CHAR_LEN
Definition: pg_wchar.h:30
int arcsCount
Definition: trgm_regexp.c:414
void appendStringInfoChar(StringInfo str, char ch)
Definition: stringinfo.c:188
void initStringInfo(StringInfo str)
Definition: stringinfo.c:59
#define COLOR_BLANK
Definition: trgm_regexp.c:286
float float4
Definition: c.h:564
List * queue
Definition: trgm_regexp.c:412
#define HASH_BLOBS
Definition: hsearch.h:97
#define PG_FINALLY()
Definition: elog.h:330
static TrgmPackedGraph * packGraph(TrgmNFA *trgmNFA, MemoryContext rcontext)
Definition: trgm_regexp.c:1931
void * palloc0(Size size)
Definition: mcxt.c:1093
TRGM * createTrgmNFA(text *text_re, Oid collation, TrgmPackedGraph **graph, MemoryContext rcontext)
Definition: trgm_regexp.c:521
static void addKeyToQueue(TrgmNFA *trgmNFA, TrgmStateKey *key)
Definition: trgm_regexp.c:1177
Size keysize
Definition: hsearch.h:75
uint8 flag
Definition: trgm.h:70
int pg_reg_colorisend(const regex_t *regex, int co)
Definition: regexport.c:208
void * MemoryContextAllocZero(MemoryContext context, Size size)
Definition: mcxt.c:906
static int colorTrgmInfoPenaltyCmp(const void *p1, const void *p2)
Definition: trgm_regexp.c:1906
int colorTrgmsCount
Definition: trgm_regexp.c:419
int pg_mb2wchar_with_len(const char *from, pg_wchar *to, int len)
Definition: mbutils.c:929
void compact_trigram(trgm *tptr, char *str, int bytelen)
Definition: trgm_op.c:198
#define ereport(elevel,...)
Definition: elog.h:157
int pg_reg_getnumoutarcs(const regex_t *regex, int st)
Definition: regexport.c:134
char trgm[3]
Definition: trgm.h:42
#define MAX_TRGM_COUNT
Definition: trgm_regexp.c:220
int tentFlags
Definition: trgm_regexp.c:339
#define Assert(condition)
Definition: c.h:804
#define lfirst(lc)
Definition: pg_list.h:169
Definition: regguts.h:317
char bytes[MAX_MULTIBYTE_CHAR_LEN]
Definition: trgm_regexp.c:242
TrgmStateKey stateKey
Definition: trgm_regexp.c:333
static TRGM * expandColorTrigrams(TrgmNFA *trgmNFA, MemoryContext rcontext)
Definition: trgm_regexp.c:1776
static void RE_compile(regex_t *regex, text *text_re, int cflags, Oid collation)
Definition: trgm_regexp.c:729
static rewind_source * source
Definition: pg_rewind.c:79
static bool validArcLabel(TrgmStateKey *key, TrgmColor co)
Definition: trgm_regexp.c:1329
static int list_length(const List *l)
Definition: pg_list.h:149
static TrgmState * getState(TrgmNFA *trgmNFA, TrgmStateKey *key)
Definition: trgm_regexp.c:1384
void * hash_seq_search(HASH_SEQ_STATUS *status)
Definition: dynahash.c:1436
static void addArcs(TrgmNFA *trgmNFA, TrgmState *state)
Definition: trgm_regexp.c:1189
void hash_seq_init(HASH_SEQ_STATUS *status, HTAB *hashp)
Definition: dynahash.c:1426
TrgmPrefix prefix
Definition: trgm_regexp.c:309
static const float4 penalties[8]
Definition: trgm_regexp.c:228
#define MAX_EXPANDED_STATES
Definition: trgm_regexp.c:218
static int colorTrgmInfoCmp(const void *p1, const void *p2)
Definition: trgm_regexp.c:1893
struct TrgmState * parent
Definition: trgm_regexp.c:338
#define COLOR_UNKNOWN
Definition: trgm_regexp.c:285
static void transformGraph(TrgmNFA *trgmNFA)
Definition: trgm_regexp.c:903
#define VARSIZE_ANY_EXHDR(PTR)
Definition: postgres.h:354
void * palloc(Size size)
Definition: mcxt.c:1062
int errmsg(const char *fmt,...)
Definition: elog.c:909
void * MemoryContextAlloc(MemoryContext context, Size size)
Definition: mcxt.c:863
int i
List * keysQueue
Definition: trgm_regexp.c:413
ColorTrgm ctrgm
Definition: trgm_regexp.c:374
Definition: c.h:621
#define SET_VARSIZE(PTR, len)
Definition: postgres.h:342
static void processState(TrgmNFA *trgmNFA, TrgmState *state)
Definition: trgm_regexp.c:965
int pg_wchar2mb_with_len(const pg_wchar *from, char *to, int len)
Definition: mbutils.c:951
static const FormData_pg_attribute a1
Definition: heap.c:152
#define qsort(a, b, c, d)
Definition: port.h:504
trgm_mb_char * wordChars
Definition: trgm_regexp.c:263
#define PG_TRY()
Definition: elog.h:313
void pg_regfree(regex_t *re)
Definition: regfree.c:49
Definition: pg_list.h:50
static char chars[TZ_MAX_CHARS]
Definition: zic.c:399
static void fillTrgm(trgm *ptrgm, trgm_mb_char s[3])
Definition: trgm_regexp.c:1844
Definition: regex.h:55
#define PG_END_TRY()
Definition: elog.h:338
List * list_delete_first(List *list)
Definition: list.c:875
int * colorTrigramGroups
Definition: trgm_regexp.c:451
int ncolors
Definition: trgm_regexp.c:404
int TrgmColor
Definition: trgm_regexp.c:282