Merge Chromium + Blink git repositories
[chromium-blink-merge.git] / third_party / sqlite / src / ext / fts3 / fts3.c
blobc43f3696ff4d50621501369b70e7563c954087fe
1 /*
2 ** 2006 Oct 10
3 **
4 ** The author disclaims copyright to this source code. In place of
5 ** a legal notice, here is a blessing:
6 **
7 ** May you do good and not evil.
8 ** May you find forgiveness for yourself and forgive others.
9 ** May you share freely, never taking more than you give.
11 ******************************************************************************
13 ** This is an SQLite module implementing full-text search.
17 ** The code in this file is only compiled if:
19 ** * The FTS3 module is being built as an extension
20 ** (in which case SQLITE_CORE is not defined), or
22 ** * The FTS3 module is being built into the core of
23 ** SQLite (in which case SQLITE_ENABLE_FTS3 is defined).
26 /* The full-text index is stored in a series of b+tree (-like)
27 ** structures called segments which map terms to doclists. The
28 ** structures are like b+trees in layout, but are constructed from the
29 ** bottom up in optimal fashion and are not updatable. Since trees
30 ** are built from the bottom up, things will be described from the
31 ** bottom up.
34 **** Varints ****
35 ** The basic unit of encoding is a variable-length integer called a
36 ** varint. We encode variable-length integers in little-endian order
37 ** using seven bits * per byte as follows:
39 ** KEY:
40 ** A = 0xxxxxxx 7 bits of data and one flag bit
41 ** B = 1xxxxxxx 7 bits of data and one flag bit
43 ** 7 bits - A
44 ** 14 bits - BA
45 ** 21 bits - BBA
46 ** and so on.
48 ** This is similar in concept to how sqlite encodes "varints" but
49 ** the encoding is not the same. SQLite varints are big-endian
50 ** are are limited to 9 bytes in length whereas FTS3 varints are
51 ** little-endian and can be up to 10 bytes in length (in theory).
53 ** Example encodings:
55 ** 1: 0x01
56 ** 127: 0x7f
57 ** 128: 0x81 0x00
60 **** Document lists ****
61 ** A doclist (document list) holds a docid-sorted list of hits for a
62 ** given term. Doclists hold docids and associated token positions.
63 ** A docid is the unique integer identifier for a single document.
64 ** A position is the index of a word within the document. The first
65 ** word of the document has a position of 0.
67 ** FTS3 used to optionally store character offsets using a compile-time
68 ** option. But that functionality is no longer supported.
70 ** A doclist is stored like this:
72 ** array {
73 ** varint docid; (delta from previous doclist)
74 ** array { (position list for column 0)
75 ** varint position; (2 more than the delta from previous position)
76 ** }
77 ** array {
78 ** varint POS_COLUMN; (marks start of position list for new column)
79 ** varint column; (index of new column)
80 ** array {
81 ** varint position; (2 more than the delta from previous position)
82 ** }
83 ** }
84 ** varint POS_END; (marks end of positions for this document.
85 ** }
87 ** Here, array { X } means zero or more occurrences of X, adjacent in
88 ** memory. A "position" is an index of a token in the token stream
89 ** generated by the tokenizer. Note that POS_END and POS_COLUMN occur
90 ** in the same logical place as the position element, and act as sentinals
91 ** ending a position list array. POS_END is 0. POS_COLUMN is 1.
92 ** The positions numbers are not stored literally but rather as two more
93 ** than the difference from the prior position, or the just the position plus
94 ** 2 for the first position. Example:
96 ** label: A B C D E F G H I J K
97 ** value: 123 5 9 1 1 14 35 0 234 72 0
99 ** The 123 value is the first docid. For column zero in this document
100 ** there are two matches at positions 3 and 10 (5-2 and 9-2+3). The 1
101 ** at D signals the start of a new column; the 1 at E indicates that the
102 ** new column is column number 1. There are two positions at 12 and 45
103 ** (14-2 and 35-2+12). The 0 at H indicate the end-of-document. The
104 ** 234 at I is the delta to next docid (357). It has one position 70
105 ** (72-2) and then terminates with the 0 at K.
107 ** A "position-list" is the list of positions for multiple columns for
108 ** a single docid. A "column-list" is the set of positions for a single
109 ** column. Hence, a position-list consists of one or more column-lists,
110 ** a document record consists of a docid followed by a position-list and
111 ** a doclist consists of one or more document records.
113 ** A bare doclist omits the position information, becoming an
114 ** array of varint-encoded docids.
116 **** Segment leaf nodes ****
117 ** Segment leaf nodes store terms and doclists, ordered by term. Leaf
118 ** nodes are written using LeafWriter, and read using LeafReader (to
119 ** iterate through a single leaf node's data) and LeavesReader (to
120 ** iterate through a segment's entire leaf layer). Leaf nodes have
121 ** the format:
123 ** varint iHeight; (height from leaf level, always 0)
124 ** varint nTerm; (length of first term)
125 ** char pTerm[nTerm]; (content of first term)
126 ** varint nDoclist; (length of term's associated doclist)
127 ** char pDoclist[nDoclist]; (content of doclist)
128 ** array {
129 ** (further terms are delta-encoded)
130 ** varint nPrefix; (length of prefix shared with previous term)
131 ** varint nSuffix; (length of unshared suffix)
132 ** char pTermSuffix[nSuffix];(unshared suffix of next term)
133 ** varint nDoclist; (length of term's associated doclist)
134 ** char pDoclist[nDoclist]; (content of doclist)
135 ** }
137 ** Here, array { X } means zero or more occurrences of X, adjacent in
138 ** memory.
140 ** Leaf nodes are broken into blocks which are stored contiguously in
141 ** the %_segments table in sorted order. This means that when the end
142 ** of a node is reached, the next term is in the node with the next
143 ** greater node id.
145 ** New data is spilled to a new leaf node when the current node
146 ** exceeds LEAF_MAX bytes (default 2048). New data which itself is
147 ** larger than STANDALONE_MIN (default 1024) is placed in a standalone
148 ** node (a leaf node with a single term and doclist). The goal of
149 ** these settings is to pack together groups of small doclists while
150 ** making it efficient to directly access large doclists. The
151 ** assumption is that large doclists represent terms which are more
152 ** likely to be query targets.
154 ** TODO(shess) It may be useful for blocking decisions to be more
155 ** dynamic. For instance, it may make more sense to have a 2.5k leaf
156 ** node rather than splitting into 2k and .5k nodes. My intuition is
157 ** that this might extend through 2x or 4x the pagesize.
160 **** Segment interior nodes ****
161 ** Segment interior nodes store blockids for subtree nodes and terms
162 ** to describe what data is stored by the each subtree. Interior
163 ** nodes are written using InteriorWriter, and read using
164 ** InteriorReader. InteriorWriters are created as needed when
165 ** SegmentWriter creates new leaf nodes, or when an interior node
166 ** itself grows too big and must be split. The format of interior
167 ** nodes:
169 ** varint iHeight; (height from leaf level, always >0)
170 ** varint iBlockid; (block id of node's leftmost subtree)
171 ** optional {
172 ** varint nTerm; (length of first term)
173 ** char pTerm[nTerm]; (content of first term)
174 ** array {
175 ** (further terms are delta-encoded)
176 ** varint nPrefix; (length of shared prefix with previous term)
177 ** varint nSuffix; (length of unshared suffix)
178 ** char pTermSuffix[nSuffix]; (unshared suffix of next term)
179 ** }
180 ** }
182 ** Here, optional { X } means an optional element, while array { X }
183 ** means zero or more occurrences of X, adjacent in memory.
185 ** An interior node encodes n terms separating n+1 subtrees. The
186 ** subtree blocks are contiguous, so only the first subtree's blockid
187 ** is encoded. The subtree at iBlockid will contain all terms less
188 ** than the first term encoded (or all terms if no term is encoded).
189 ** Otherwise, for terms greater than or equal to pTerm[i] but less
190 ** than pTerm[i+1], the subtree for that term will be rooted at
191 ** iBlockid+i. Interior nodes only store enough term data to
192 ** distinguish adjacent children (if the rightmost term of the left
193 ** child is "something", and the leftmost term of the right child is
194 ** "wicked", only "w" is stored).
196 ** New data is spilled to a new interior node at the same height when
197 ** the current node exceeds INTERIOR_MAX bytes (default 2048).
198 ** INTERIOR_MIN_TERMS (default 7) keeps large terms from monopolizing
199 ** interior nodes and making the tree too skinny. The interior nodes
200 ** at a given height are naturally tracked by interior nodes at
201 ** height+1, and so on.
204 **** Segment directory ****
205 ** The segment directory in table %_segdir stores meta-information for
206 ** merging and deleting segments, and also the root node of the
207 ** segment's tree.
209 ** The root node is the top node of the segment's tree after encoding
210 ** the entire segment, restricted to ROOT_MAX bytes (default 1024).
211 ** This could be either a leaf node or an interior node. If the top
212 ** node requires more than ROOT_MAX bytes, it is flushed to %_segments
213 ** and a new root interior node is generated (which should always fit
214 ** within ROOT_MAX because it only needs space for 2 varints, the
215 ** height and the blockid of the previous root).
217 ** The meta-information in the segment directory is:
218 ** level - segment level (see below)
219 ** idx - index within level
220 ** - (level,idx uniquely identify a segment)
221 ** start_block - first leaf node
222 ** leaves_end_block - last leaf node
223 ** end_block - last block (including interior nodes)
224 ** root - contents of root node
226 ** If the root node is a leaf node, then start_block,
227 ** leaves_end_block, and end_block are all 0.
230 **** Segment merging ****
231 ** To amortize update costs, segments are grouped into levels and
232 ** merged in batches. Each increase in level represents exponentially
233 ** more documents.
235 ** New documents (actually, document updates) are tokenized and
236 ** written individually (using LeafWriter) to a level 0 segment, with
237 ** incrementing idx. When idx reaches MERGE_COUNT (default 16), all
238 ** level 0 segments are merged into a single level 1 segment. Level 1
239 ** is populated like level 0, and eventually MERGE_COUNT level 1
240 ** segments are merged to a single level 2 segment (representing
241 ** MERGE_COUNT^2 updates), and so on.
243 ** A segment merge traverses all segments at a given level in
244 ** parallel, performing a straightforward sorted merge. Since segment
245 ** leaf nodes are written in to the %_segments table in order, this
246 ** merge traverses the underlying sqlite disk structures efficiently.
247 ** After the merge, all segment blocks from the merged level are
248 ** deleted.
250 ** MERGE_COUNT controls how often we merge segments. 16 seems to be
251 ** somewhat of a sweet spot for insertion performance. 32 and 64 show
252 ** very similar performance numbers to 16 on insertion, though they're
253 ** a tiny bit slower (perhaps due to more overhead in merge-time
254 ** sorting). 8 is about 20% slower than 16, 4 about 50% slower than
255 ** 16, 2 about 66% slower than 16.
257 ** At query time, high MERGE_COUNT increases the number of segments
258 ** which need to be scanned and merged. For instance, with 100k docs
259 ** inserted:
261 ** MERGE_COUNT segments
262 ** 16 25
263 ** 8 12
264 ** 4 10
265 ** 2 6
267 ** This appears to have only a moderate impact on queries for very
268 ** frequent terms (which are somewhat dominated by segment merge
269 ** costs), and infrequent and non-existent terms still seem to be fast
270 ** even with many segments.
272 ** TODO(shess) That said, it would be nice to have a better query-side
273 ** argument for MERGE_COUNT of 16. Also, it is possible/likely that
274 ** optimizations to things like doclist merging will swing the sweet
275 ** spot around.
279 **** Handling of deletions and updates ****
280 ** Since we're using a segmented structure, with no docid-oriented
281 ** index into the term index, we clearly cannot simply update the term
282 ** index when a document is deleted or updated. For deletions, we
283 ** write an empty doclist (varint(docid) varint(POS_END)), for updates
284 ** we simply write the new doclist. Segment merges overwrite older
285 ** data for a particular docid with newer data, so deletes or updates
286 ** will eventually overtake the earlier data and knock it out. The
287 ** query logic likewise merges doclists so that newer data knocks out
288 ** older data.
290 #define CHROMIUM_FTS3_CHANGES 1
292 #include "fts3Int.h"
293 #if !defined(SQLITE_CORE) || defined(SQLITE_ENABLE_FTS3)
295 #if defined(SQLITE_ENABLE_FTS3) && !defined(SQLITE_CORE)
296 # define SQLITE_CORE 1
297 #endif
299 #include <assert.h>
300 #include <stdlib.h>
301 #include <stddef.h>
302 #include <stdio.h>
303 #include <string.h>
304 #include <stdarg.h>
306 #include "fts3.h"
307 #ifndef SQLITE_CORE
308 # include "sqlite3ext.h"
309 SQLITE_EXTENSION_INIT1
310 #endif
312 static int fts3EvalNext(Fts3Cursor *pCsr);
313 static int fts3EvalStart(Fts3Cursor *pCsr);
314 static int fts3TermSegReaderCursor(
315 Fts3Cursor *, const char *, int, int, Fts3MultiSegReader **);
318 ** Write a 64-bit variable-length integer to memory starting at p[0].
319 ** The length of data written will be between 1 and FTS3_VARINT_MAX bytes.
320 ** The number of bytes written is returned.
322 int sqlite3Fts3PutVarint(char *p, sqlite_int64 v){
323 unsigned char *q = (unsigned char *) p;
324 sqlite_uint64 vu = v;
326 *q++ = (unsigned char) ((vu & 0x7f) | 0x80);
327 vu >>= 7;
328 }while( vu!=0 );
329 q[-1] &= 0x7f; /* turn off high bit in final byte */
330 assert( q - (unsigned char *)p <= FTS3_VARINT_MAX );
331 return (int) (q - (unsigned char *)p);
334 #define GETVARINT_STEP(v, ptr, shift, mask1, mask2, var, ret) \
335 v = (v & mask1) | ( (*ptr++) << shift ); \
336 if( (v & mask2)==0 ){ var = v; return ret; }
337 #define GETVARINT_INIT(v, ptr, shift, mask1, mask2, var, ret) \
338 v = (*ptr++); \
339 if( (v & mask2)==0 ){ var = v; return ret; }
342 ** Read a 64-bit variable-length integer from memory starting at p[0].
343 ** Return the number of bytes read, or 0 on error.
344 ** The value is stored in *v.
346 int sqlite3Fts3GetVarint(const char *p, sqlite_int64 *v){
347 const char *pStart = p;
348 u32 a;
349 u64 b;
350 int shift;
352 GETVARINT_INIT(a, p, 0, 0x00, 0x80, *v, 1);
353 GETVARINT_STEP(a, p, 7, 0x7F, 0x4000, *v, 2);
354 GETVARINT_STEP(a, p, 14, 0x3FFF, 0x200000, *v, 3);
355 GETVARINT_STEP(a, p, 21, 0x1FFFFF, 0x10000000, *v, 4);
356 b = (a & 0x0FFFFFFF );
358 for(shift=28; shift<=63; shift+=7){
359 u64 c = *p++;
360 b += (c&0x7F) << shift;
361 if( (c & 0x80)==0 ) break;
363 *v = b;
364 return (int)(p - pStart);
368 ** Similar to sqlite3Fts3GetVarint(), except that the output is truncated to a
369 ** 32-bit integer before it is returned.
371 int sqlite3Fts3GetVarint32(const char *p, int *pi){
372 u32 a;
374 #ifndef fts3GetVarint32
375 GETVARINT_INIT(a, p, 0, 0x00, 0x80, *pi, 1);
376 #else
377 a = (*p++);
378 assert( a & 0x80 );
379 #endif
381 GETVARINT_STEP(a, p, 7, 0x7F, 0x4000, *pi, 2);
382 GETVARINT_STEP(a, p, 14, 0x3FFF, 0x200000, *pi, 3);
383 GETVARINT_STEP(a, p, 21, 0x1FFFFF, 0x10000000, *pi, 4);
384 a = (a & 0x0FFFFFFF );
385 *pi = (int)(a | ((u32)(*p & 0x0F) << 28));
386 return 5;
390 ** Return the number of bytes required to encode v as a varint
392 int sqlite3Fts3VarintLen(sqlite3_uint64 v){
393 int i = 0;
395 i++;
396 v >>= 7;
397 }while( v!=0 );
398 return i;
402 ** Convert an SQL-style quoted string into a normal string by removing
403 ** the quote characters. The conversion is done in-place. If the
404 ** input does not begin with a quote character, then this routine
405 ** is a no-op.
407 ** Examples:
409 ** "abc" becomes abc
410 ** 'xyz' becomes xyz
411 ** [pqr] becomes pqr
412 ** `mno` becomes mno
415 void sqlite3Fts3Dequote(char *z){
416 char quote; /* Quote character (if any ) */
418 quote = z[0];
419 if( quote=='[' || quote=='\'' || quote=='"' || quote=='`' ){
420 int iIn = 1; /* Index of next byte to read from input */
421 int iOut = 0; /* Index of next byte to write to output */
423 /* If the first byte was a '[', then the close-quote character is a ']' */
424 if( quote=='[' ) quote = ']';
426 while( ALWAYS(z[iIn]) ){
427 if( z[iIn]==quote ){
428 if( z[iIn+1]!=quote ) break;
429 z[iOut++] = quote;
430 iIn += 2;
431 }else{
432 z[iOut++] = z[iIn++];
435 z[iOut] = '\0';
440 ** Read a single varint from the doclist at *pp and advance *pp to point
441 ** to the first byte past the end of the varint. Add the value of the varint
442 ** to *pVal.
444 static void fts3GetDeltaVarint(char **pp, sqlite3_int64 *pVal){
445 sqlite3_int64 iVal;
446 *pp += sqlite3Fts3GetVarint(*pp, &iVal);
447 *pVal += iVal;
451 ** When this function is called, *pp points to the first byte following a
452 ** varint that is part of a doclist (or position-list, or any other list
453 ** of varints). This function moves *pp to point to the start of that varint,
454 ** and sets *pVal by the varint value.
456 ** Argument pStart points to the first byte of the doclist that the
457 ** varint is part of.
459 static void fts3GetReverseVarint(
460 char **pp,
461 char *pStart,
462 sqlite3_int64 *pVal
464 sqlite3_int64 iVal;
465 char *p;
467 /* Pointer p now points at the first byte past the varint we are
468 ** interested in. So, unless the doclist is corrupt, the 0x80 bit is
469 ** clear on character p[-1]. */
470 for(p = (*pp)-2; p>=pStart && *p&0x80; p--);
471 p++;
472 *pp = p;
474 sqlite3Fts3GetVarint(p, &iVal);
475 *pVal = iVal;
479 ** The xDisconnect() virtual table method.
481 static int fts3DisconnectMethod(sqlite3_vtab *pVtab){
482 Fts3Table *p = (Fts3Table *)pVtab;
483 int i;
485 assert( p->nPendingData==0 );
486 assert( p->pSegments==0 );
488 /* Free any prepared statements held */
489 for(i=0; i<SizeofArray(p->aStmt); i++){
490 sqlite3_finalize(p->aStmt[i]);
492 sqlite3_free(p->zSegmentsTbl);
493 sqlite3_free(p->zReadExprlist);
494 sqlite3_free(p->zWriteExprlist);
495 sqlite3_free(p->zContentTbl);
496 sqlite3_free(p->zLanguageid);
498 /* Invoke the tokenizer destructor to free the tokenizer. */
499 p->pTokenizer->pModule->xDestroy(p->pTokenizer);
501 sqlite3_free(p);
502 return SQLITE_OK;
506 ** Construct one or more SQL statements from the format string given
507 ** and then evaluate those statements. The success code is written
508 ** into *pRc.
510 ** If *pRc is initially non-zero then this routine is a no-op.
512 static void fts3DbExec(
513 int *pRc, /* Success code */
514 sqlite3 *db, /* Database in which to run SQL */
515 const char *zFormat, /* Format string for SQL */
516 ... /* Arguments to the format string */
518 va_list ap;
519 char *zSql;
520 if( *pRc ) return;
521 va_start(ap, zFormat);
522 zSql = sqlite3_vmprintf(zFormat, ap);
523 va_end(ap);
524 if( zSql==0 ){
525 *pRc = SQLITE_NOMEM;
526 }else{
527 *pRc = sqlite3_exec(db, zSql, 0, 0, 0);
528 sqlite3_free(zSql);
533 ** The xDestroy() virtual table method.
535 static int fts3DestroyMethod(sqlite3_vtab *pVtab){
536 Fts3Table *p = (Fts3Table *)pVtab;
537 int rc = SQLITE_OK; /* Return code */
538 const char *zDb = p->zDb; /* Name of database (e.g. "main", "temp") */
539 sqlite3 *db = p->db; /* Database handle */
541 /* Drop the shadow tables */
542 if( p->zContentTbl==0 ){
543 fts3DbExec(&rc, db, "DROP TABLE IF EXISTS %Q.'%q_content'", zDb, p->zName);
545 fts3DbExec(&rc, db, "DROP TABLE IF EXISTS %Q.'%q_segments'", zDb,p->zName);
546 fts3DbExec(&rc, db, "DROP TABLE IF EXISTS %Q.'%q_segdir'", zDb, p->zName);
547 fts3DbExec(&rc, db, "DROP TABLE IF EXISTS %Q.'%q_docsize'", zDb, p->zName);
548 fts3DbExec(&rc, db, "DROP TABLE IF EXISTS %Q.'%q_stat'", zDb, p->zName);
550 /* If everything has worked, invoke fts3DisconnectMethod() to free the
551 ** memory associated with the Fts3Table structure and return SQLITE_OK.
552 ** Otherwise, return an SQLite error code.
554 return (rc==SQLITE_OK ? fts3DisconnectMethod(pVtab) : rc);
559 ** Invoke sqlite3_declare_vtab() to declare the schema for the FTS3 table
560 ** passed as the first argument. This is done as part of the xConnect()
561 ** and xCreate() methods.
563 ** If *pRc is non-zero when this function is called, it is a no-op.
564 ** Otherwise, if an error occurs, an SQLite error code is stored in *pRc
565 ** before returning.
567 static void fts3DeclareVtab(int *pRc, Fts3Table *p){
568 if( *pRc==SQLITE_OK ){
569 int i; /* Iterator variable */
570 int rc; /* Return code */
571 char *zSql; /* SQL statement passed to declare_vtab() */
572 char *zCols; /* List of user defined columns */
573 const char *zLanguageid;
575 zLanguageid = (p->zLanguageid ? p->zLanguageid : "__langid");
576 sqlite3_vtab_config(p->db, SQLITE_VTAB_CONSTRAINT_SUPPORT, 1);
578 /* Create a list of user columns for the virtual table */
579 zCols = sqlite3_mprintf("%Q, ", p->azColumn[0]);
580 for(i=1; zCols && i<p->nColumn; i++){
581 zCols = sqlite3_mprintf("%z%Q, ", zCols, p->azColumn[i]);
584 /* Create the whole "CREATE TABLE" statement to pass to SQLite */
585 zSql = sqlite3_mprintf(
586 "CREATE TABLE x(%s %Q HIDDEN, docid HIDDEN, %Q HIDDEN)",
587 zCols, p->zName, zLanguageid
589 if( !zCols || !zSql ){
590 rc = SQLITE_NOMEM;
591 }else{
592 rc = sqlite3_declare_vtab(p->db, zSql);
595 sqlite3_free(zSql);
596 sqlite3_free(zCols);
597 *pRc = rc;
602 ** Create the %_stat table if it does not already exist.
604 void sqlite3Fts3CreateStatTable(int *pRc, Fts3Table *p){
605 fts3DbExec(pRc, p->db,
606 "CREATE TABLE IF NOT EXISTS %Q.'%q_stat'"
607 "(id INTEGER PRIMARY KEY, value BLOB);",
608 p->zDb, p->zName
610 if( (*pRc)==SQLITE_OK ) p->bHasStat = 1;
614 ** Create the backing store tables (%_content, %_segments and %_segdir)
615 ** required by the FTS3 table passed as the only argument. This is done
616 ** as part of the vtab xCreate() method.
618 ** If the p->bHasDocsize boolean is true (indicating that this is an
619 ** FTS4 table, not an FTS3 table) then also create the %_docsize and
620 ** %_stat tables required by FTS4.
622 static int fts3CreateTables(Fts3Table *p){
623 int rc = SQLITE_OK; /* Return code */
624 int i; /* Iterator variable */
625 sqlite3 *db = p->db; /* The database connection */
627 if( p->zContentTbl==0 ){
628 const char *zLanguageid = p->zLanguageid;
629 char *zContentCols; /* Columns of %_content table */
631 /* Create a list of user columns for the content table */
632 zContentCols = sqlite3_mprintf("docid INTEGER PRIMARY KEY");
633 for(i=0; zContentCols && i<p->nColumn; i++){
634 char *z = p->azColumn[i];
635 zContentCols = sqlite3_mprintf("%z, 'c%d%q'", zContentCols, i, z);
637 if( zLanguageid && zContentCols ){
638 zContentCols = sqlite3_mprintf("%z, langid", zContentCols, zLanguageid);
640 if( zContentCols==0 ) rc = SQLITE_NOMEM;
642 /* Create the content table */
643 fts3DbExec(&rc, db,
644 "CREATE TABLE %Q.'%q_content'(%s)",
645 p->zDb, p->zName, zContentCols
647 sqlite3_free(zContentCols);
650 /* Create other tables */
651 fts3DbExec(&rc, db,
652 "CREATE TABLE %Q.'%q_segments'(blockid INTEGER PRIMARY KEY, block BLOB);",
653 p->zDb, p->zName
655 fts3DbExec(&rc, db,
656 "CREATE TABLE %Q.'%q_segdir'("
657 "level INTEGER,"
658 "idx INTEGER,"
659 "start_block INTEGER,"
660 "leaves_end_block INTEGER,"
661 "end_block INTEGER,"
662 "root BLOB,"
663 "PRIMARY KEY(level, idx)"
664 ");",
665 p->zDb, p->zName
667 if( p->bHasDocsize ){
668 fts3DbExec(&rc, db,
669 "CREATE TABLE %Q.'%q_docsize'(docid INTEGER PRIMARY KEY, size BLOB);",
670 p->zDb, p->zName
673 assert( p->bHasStat==p->bFts4 );
674 if( p->bHasStat ){
675 sqlite3Fts3CreateStatTable(&rc, p);
677 return rc;
681 ** Store the current database page-size in bytes in p->nPgsz.
683 ** If *pRc is non-zero when this function is called, it is a no-op.
684 ** Otherwise, if an error occurs, an SQLite error code is stored in *pRc
685 ** before returning.
687 static void fts3DatabasePageSize(int *pRc, Fts3Table *p){
688 if( *pRc==SQLITE_OK ){
689 int rc; /* Return code */
690 char *zSql; /* SQL text "PRAGMA %Q.page_size" */
691 sqlite3_stmt *pStmt; /* Compiled "PRAGMA %Q.page_size" statement */
693 zSql = sqlite3_mprintf("PRAGMA %Q.page_size", p->zDb);
694 if( !zSql ){
695 rc = SQLITE_NOMEM;
696 }else{
697 rc = sqlite3_prepare(p->db, zSql, -1, &pStmt, 0);
698 if( rc==SQLITE_OK ){
699 sqlite3_step(pStmt);
700 p->nPgsz = sqlite3_column_int(pStmt, 0);
701 rc = sqlite3_finalize(pStmt);
702 }else if( rc==SQLITE_AUTH ){
703 p->nPgsz = 1024;
704 rc = SQLITE_OK;
707 assert( p->nPgsz>0 || rc!=SQLITE_OK );
708 sqlite3_free(zSql);
709 *pRc = rc;
714 ** "Special" FTS4 arguments are column specifications of the following form:
716 ** <key> = <value>
718 ** There may not be whitespace surrounding the "=" character. The <value>
719 ** term may be quoted, but the <key> may not.
721 static int fts3IsSpecialColumn(
722 const char *z,
723 int *pnKey,
724 char **pzValue
726 char *zValue;
727 const char *zCsr = z;
729 while( *zCsr!='=' ){
730 if( *zCsr=='\0' ) return 0;
731 zCsr++;
734 *pnKey = (int)(zCsr-z);
735 zValue = sqlite3_mprintf("%s", &zCsr[1]);
736 if( zValue ){
737 sqlite3Fts3Dequote(zValue);
739 *pzValue = zValue;
740 return 1;
744 ** Append the output of a printf() style formatting to an existing string.
746 static void fts3Appendf(
747 int *pRc, /* IN/OUT: Error code */
748 char **pz, /* IN/OUT: Pointer to string buffer */
749 const char *zFormat, /* Printf format string to append */
750 ... /* Arguments for printf format string */
752 if( *pRc==SQLITE_OK ){
753 va_list ap;
754 char *z;
755 va_start(ap, zFormat);
756 z = sqlite3_vmprintf(zFormat, ap);
757 va_end(ap);
758 if( z && *pz ){
759 char *z2 = sqlite3_mprintf("%s%s", *pz, z);
760 sqlite3_free(z);
761 z = z2;
763 if( z==0 ) *pRc = SQLITE_NOMEM;
764 sqlite3_free(*pz);
765 *pz = z;
770 ** Return a copy of input string zInput enclosed in double-quotes (") and
771 ** with all double quote characters escaped. For example:
773 ** fts3QuoteId("un \"zip\"") -> "un \"\"zip\"\""
775 ** The pointer returned points to memory obtained from sqlite3_malloc(). It
776 ** is the callers responsibility to call sqlite3_free() to release this
777 ** memory.
779 static char *fts3QuoteId(char const *zInput){
780 int nRet;
781 char *zRet;
782 nRet = 2 + (int)strlen(zInput)*2 + 1;
783 zRet = sqlite3_malloc(nRet);
784 if( zRet ){
785 int i;
786 char *z = zRet;
787 *(z++) = '"';
788 for(i=0; zInput[i]; i++){
789 if( zInput[i]=='"' ) *(z++) = '"';
790 *(z++) = zInput[i];
792 *(z++) = '"';
793 *(z++) = '\0';
795 return zRet;
799 ** Return a list of comma separated SQL expressions and a FROM clause that
800 ** could be used in a SELECT statement such as the following:
802 ** SELECT <list of expressions> FROM %_content AS x ...
804 ** to return the docid, followed by each column of text data in order
805 ** from left to write. If parameter zFunc is not NULL, then instead of
806 ** being returned directly each column of text data is passed to an SQL
807 ** function named zFunc first. For example, if zFunc is "unzip" and the
808 ** table has the three user-defined columns "a", "b", and "c", the following
809 ** string is returned:
811 ** "docid, unzip(x.'a'), unzip(x.'b'), unzip(x.'c') FROM %_content AS x"
813 ** The pointer returned points to a buffer allocated by sqlite3_malloc(). It
814 ** is the responsibility of the caller to eventually free it.
816 ** If *pRc is not SQLITE_OK when this function is called, it is a no-op (and
817 ** a NULL pointer is returned). Otherwise, if an OOM error is encountered
818 ** by this function, NULL is returned and *pRc is set to SQLITE_NOMEM. If
819 ** no error occurs, *pRc is left unmodified.
821 static char *fts3ReadExprList(Fts3Table *p, const char *zFunc, int *pRc){
822 char *zRet = 0;
823 char *zFree = 0;
824 char *zFunction;
825 int i;
827 if( p->zContentTbl==0 ){
828 if( !zFunc ){
829 zFunction = "";
830 }else{
831 zFree = zFunction = fts3QuoteId(zFunc);
833 fts3Appendf(pRc, &zRet, "docid");
834 for(i=0; i<p->nColumn; i++){
835 fts3Appendf(pRc, &zRet, ",%s(x.'c%d%q')", zFunction, i, p->azColumn[i]);
837 if( p->zLanguageid ){
838 fts3Appendf(pRc, &zRet, ", x.%Q", "langid");
840 sqlite3_free(zFree);
841 }else{
842 fts3Appendf(pRc, &zRet, "rowid");
843 for(i=0; i<p->nColumn; i++){
844 fts3Appendf(pRc, &zRet, ", x.'%q'", p->azColumn[i]);
846 if( p->zLanguageid ){
847 fts3Appendf(pRc, &zRet, ", x.%Q", p->zLanguageid);
850 fts3Appendf(pRc, &zRet, " FROM '%q'.'%q%s' AS x",
851 p->zDb,
852 (p->zContentTbl ? p->zContentTbl : p->zName),
853 (p->zContentTbl ? "" : "_content")
855 return zRet;
859 ** Return a list of N comma separated question marks, where N is the number
860 ** of columns in the %_content table (one for the docid plus one for each
861 ** user-defined text column).
863 ** If argument zFunc is not NULL, then all but the first question mark
864 ** is preceded by zFunc and an open bracket, and followed by a closed
865 ** bracket. For example, if zFunc is "zip" and the FTS3 table has three
866 ** user-defined text columns, the following string is returned:
868 ** "?, zip(?), zip(?), zip(?)"
870 ** The pointer returned points to a buffer allocated by sqlite3_malloc(). It
871 ** is the responsibility of the caller to eventually free it.
873 ** If *pRc is not SQLITE_OK when this function is called, it is a no-op (and
874 ** a NULL pointer is returned). Otherwise, if an OOM error is encountered
875 ** by this function, NULL is returned and *pRc is set to SQLITE_NOMEM. If
876 ** no error occurs, *pRc is left unmodified.
878 static char *fts3WriteExprList(Fts3Table *p, const char *zFunc, int *pRc){
879 char *zRet = 0;
880 char *zFree = 0;
881 char *zFunction;
882 int i;
884 if( !zFunc ){
885 zFunction = "";
886 }else{
887 zFree = zFunction = fts3QuoteId(zFunc);
889 fts3Appendf(pRc, &zRet, "?");
890 for(i=0; i<p->nColumn; i++){
891 fts3Appendf(pRc, &zRet, ",%s(?)", zFunction);
893 if( p->zLanguageid ){
894 fts3Appendf(pRc, &zRet, ", ?");
896 sqlite3_free(zFree);
897 return zRet;
901 ** This function interprets the string at (*pp) as a non-negative integer
902 ** value. It reads the integer and sets *pnOut to the value read, then
903 ** sets *pp to point to the byte immediately following the last byte of
904 ** the integer value.
906 ** Only decimal digits ('0'..'9') may be part of an integer value.
908 ** If *pp does not being with a decimal digit SQLITE_ERROR is returned and
909 ** the output value undefined. Otherwise SQLITE_OK is returned.
911 ** This function is used when parsing the "prefix=" FTS4 parameter.
913 static int fts3GobbleInt(const char **pp, int *pnOut){
914 const char *p; /* Iterator pointer */
915 int nInt = 0; /* Output value */
917 for(p=*pp; p[0]>='0' && p[0]<='9'; p++){
918 nInt = nInt * 10 + (p[0] - '0');
920 if( p==*pp ) return SQLITE_ERROR;
921 *pnOut = nInt;
922 *pp = p;
923 return SQLITE_OK;
927 ** This function is called to allocate an array of Fts3Index structures
928 ** representing the indexes maintained by the current FTS table. FTS tables
929 ** always maintain the main "terms" index, but may also maintain one or
930 ** more "prefix" indexes, depending on the value of the "prefix=" parameter
931 ** (if any) specified as part of the CREATE VIRTUAL TABLE statement.
933 ** Argument zParam is passed the value of the "prefix=" option if one was
934 ** specified, or NULL otherwise.
936 ** If no error occurs, SQLITE_OK is returned and *apIndex set to point to
937 ** the allocated array. *pnIndex is set to the number of elements in the
938 ** array. If an error does occur, an SQLite error code is returned.
940 ** Regardless of whether or not an error is returned, it is the responsibility
941 ** of the caller to call sqlite3_free() on the output array to free it.
943 static int fts3PrefixParameter(
944 const char *zParam, /* ABC in prefix=ABC parameter to parse */
945 int *pnIndex, /* OUT: size of *apIndex[] array */
946 struct Fts3Index **apIndex /* OUT: Array of indexes for this table */
948 struct Fts3Index *aIndex; /* Allocated array */
949 int nIndex = 1; /* Number of entries in array */
951 if( zParam && zParam[0] ){
952 const char *p;
953 nIndex++;
954 for(p=zParam; *p; p++){
955 if( *p==',' ) nIndex++;
959 aIndex = sqlite3_malloc(sizeof(struct Fts3Index) * nIndex);
960 *apIndex = aIndex;
961 *pnIndex = nIndex;
962 if( !aIndex ){
963 return SQLITE_NOMEM;
966 memset(aIndex, 0, sizeof(struct Fts3Index) * nIndex);
967 if( zParam ){
968 const char *p = zParam;
969 int i;
970 for(i=1; i<nIndex; i++){
971 int nPrefix;
972 if( fts3GobbleInt(&p, &nPrefix) ) return SQLITE_ERROR;
973 aIndex[i].nPrefix = nPrefix;
974 p++;
978 return SQLITE_OK;
982 ** This function is called when initializing an FTS4 table that uses the
983 ** content=xxx option. It determines the number of and names of the columns
984 ** of the new FTS4 table.
986 ** The third argument passed to this function is the value passed to the
987 ** config=xxx option (i.e. "xxx"). This function queries the database for
988 ** a table of that name. If found, the output variables are populated
989 ** as follows:
991 ** *pnCol: Set to the number of columns table xxx has,
993 ** *pnStr: Set to the total amount of space required to store a copy
994 ** of each columns name, including the nul-terminator.
996 ** *pazCol: Set to point to an array of *pnCol strings. Each string is
997 ** the name of the corresponding column in table xxx. The array
998 ** and its contents are allocated using a single allocation. It
999 ** is the responsibility of the caller to free this allocation
1000 ** by eventually passing the *pazCol value to sqlite3_free().
1002 ** If the table cannot be found, an error code is returned and the output
1003 ** variables are undefined. Or, if an OOM is encountered, SQLITE_NOMEM is
1004 ** returned (and the output variables are undefined).
1006 static int fts3ContentColumns(
1007 sqlite3 *db, /* Database handle */
1008 const char *zDb, /* Name of db (i.e. "main", "temp" etc.) */
1009 const char *zTbl, /* Name of content table */
1010 const char ***pazCol, /* OUT: Malloc'd array of column names */
1011 int *pnCol, /* OUT: Size of array *pazCol */
1012 int *pnStr /* OUT: Bytes of string content */
1014 int rc = SQLITE_OK; /* Return code */
1015 char *zSql; /* "SELECT *" statement on zTbl */
1016 sqlite3_stmt *pStmt = 0; /* Compiled version of zSql */
1018 zSql = sqlite3_mprintf("SELECT * FROM %Q.%Q", zDb, zTbl);
1019 if( !zSql ){
1020 rc = SQLITE_NOMEM;
1021 }else{
1022 rc = sqlite3_prepare(db, zSql, -1, &pStmt, 0);
1024 sqlite3_free(zSql);
1026 if( rc==SQLITE_OK ){
1027 const char **azCol; /* Output array */
1028 int nStr = 0; /* Size of all column names (incl. 0x00) */
1029 int nCol; /* Number of table columns */
1030 int i; /* Used to iterate through columns */
1032 /* Loop through the returned columns. Set nStr to the number of bytes of
1033 ** space required to store a copy of each column name, including the
1034 ** nul-terminator byte. */
1035 nCol = sqlite3_column_count(pStmt);
1036 for(i=0; i<nCol; i++){
1037 const char *zCol = sqlite3_column_name(pStmt, i);
1038 nStr += (int)strlen(zCol) + 1;
1041 /* Allocate and populate the array to return. */
1042 azCol = (const char **)sqlite3_malloc(sizeof(char *) * nCol + nStr);
1043 if( azCol==0 ){
1044 rc = SQLITE_NOMEM;
1045 }else{
1046 char *p = (char *)&azCol[nCol];
1047 for(i=0; i<nCol; i++){
1048 const char *zCol = sqlite3_column_name(pStmt, i);
1049 int n = (int)strlen(zCol)+1;
1050 memcpy(p, zCol, n);
1051 azCol[i] = p;
1052 p += n;
1055 sqlite3_finalize(pStmt);
1057 /* Set the output variables. */
1058 *pnCol = nCol;
1059 *pnStr = nStr;
1060 *pazCol = azCol;
1063 return rc;
1067 ** This function is the implementation of both the xConnect and xCreate
1068 ** methods of the FTS3 virtual table.
1070 ** The argv[] array contains the following:
1072 ** argv[0] -> module name ("fts3" or "fts4")
1073 ** argv[1] -> database name
1074 ** argv[2] -> table name
1075 ** argv[...] -> "column name" and other module argument fields.
1077 static int fts3InitVtab(
1078 int isCreate, /* True for xCreate, false for xConnect */
1079 sqlite3 *db, /* The SQLite database connection */
1080 void *pAux, /* Hash table containing tokenizers */
1081 int argc, /* Number of elements in argv array */
1082 const char * const *argv, /* xCreate/xConnect argument array */
1083 sqlite3_vtab **ppVTab, /* Write the resulting vtab structure here */
1084 char **pzErr /* Write any error message here */
1086 Fts3Hash *pHash = (Fts3Hash *)pAux;
1087 Fts3Table *p = 0; /* Pointer to allocated vtab */
1088 int rc = SQLITE_OK; /* Return code */
1089 int i; /* Iterator variable */
1090 int nByte; /* Size of allocation used for *p */
1091 int iCol; /* Column index */
1092 int nString = 0; /* Bytes required to hold all column names */
1093 int nCol = 0; /* Number of columns in the FTS table */
1094 char *zCsr; /* Space for holding column names */
1095 int nDb; /* Bytes required to hold database name */
1096 int nName; /* Bytes required to hold table name */
1097 int isFts4 = (argv[0][3]=='4'); /* True for FTS4, false for FTS3 */
1098 const char **aCol; /* Array of column names */
1099 sqlite3_tokenizer *pTokenizer = 0; /* Tokenizer for this table */
1101 int nIndex; /* Size of aIndex[] array */
1102 struct Fts3Index *aIndex = 0; /* Array of indexes for this table */
1104 /* The results of parsing supported FTS4 key=value options: */
1105 int bNoDocsize = 0; /* True to omit %_docsize table */
1106 int bDescIdx = 0; /* True to store descending indexes */
1107 char *zPrefix = 0; /* Prefix parameter value (or NULL) */
1108 char *zCompress = 0; /* compress=? parameter (or NULL) */
1109 char *zUncompress = 0; /* uncompress=? parameter (or NULL) */
1110 char *zContent = 0; /* content=? parameter (or NULL) */
1111 char *zLanguageid = 0; /* languageid=? parameter (or NULL) */
1112 char **azNotindexed = 0; /* The set of notindexed= columns */
1113 int nNotindexed = 0; /* Size of azNotindexed[] array */
1115 assert( strlen(argv[0])==4 );
1116 assert( (sqlite3_strnicmp(argv[0], "fts4", 4)==0 && isFts4)
1117 || (sqlite3_strnicmp(argv[0], "fts3", 4)==0 && !isFts4)
1120 nDb = (int)strlen(argv[1]) + 1;
1121 nName = (int)strlen(argv[2]) + 1;
1123 nByte = sizeof(const char *) * (argc-2);
1124 aCol = (const char **)sqlite3_malloc(nByte);
1125 if( aCol ){
1126 memset((void*)aCol, 0, nByte);
1127 azNotindexed = (char **)sqlite3_malloc(nByte);
1129 if( azNotindexed ){
1130 memset(azNotindexed, 0, nByte);
1132 if( !aCol || !azNotindexed ){
1133 rc = SQLITE_NOMEM;
1134 goto fts3_init_out;
1137 /* Loop through all of the arguments passed by the user to the FTS3/4
1138 ** module (i.e. all the column names and special arguments). This loop
1139 ** does the following:
1141 ** + Figures out the number of columns the FTSX table will have, and
1142 ** the number of bytes of space that must be allocated to store copies
1143 ** of the column names.
1145 ** + If there is a tokenizer specification included in the arguments,
1146 ** initializes the tokenizer pTokenizer.
1148 for(i=3; rc==SQLITE_OK && i<argc; i++){
1149 char const *z = argv[i];
1150 int nKey;
1151 char *zVal;
1153 /* Check if this is a tokenizer specification */
1154 if( !pTokenizer
1155 && strlen(z)>8
1156 && 0==sqlite3_strnicmp(z, "tokenize", 8)
1157 && 0==sqlite3Fts3IsIdChar(z[8])
1159 rc = sqlite3Fts3InitTokenizer(pHash, &z[9], &pTokenizer, pzErr);
1162 /* Check if it is an FTS4 special argument. */
1163 else if( isFts4 && fts3IsSpecialColumn(z, &nKey, &zVal) ){
1164 struct Fts4Option {
1165 const char *zOpt;
1166 int nOpt;
1167 } aFts4Opt[] = {
1168 { "matchinfo", 9 }, /* 0 -> MATCHINFO */
1169 { "prefix", 6 }, /* 1 -> PREFIX */
1170 { "compress", 8 }, /* 2 -> COMPRESS */
1171 { "uncompress", 10 }, /* 3 -> UNCOMPRESS */
1172 { "order", 5 }, /* 4 -> ORDER */
1173 { "content", 7 }, /* 5 -> CONTENT */
1174 { "languageid", 10 }, /* 6 -> LANGUAGEID */
1175 { "notindexed", 10 } /* 7 -> NOTINDEXED */
1178 int iOpt;
1179 if( !zVal ){
1180 rc = SQLITE_NOMEM;
1181 }else{
1182 for(iOpt=0; iOpt<SizeofArray(aFts4Opt); iOpt++){
1183 struct Fts4Option *pOp = &aFts4Opt[iOpt];
1184 if( nKey==pOp->nOpt && !sqlite3_strnicmp(z, pOp->zOpt, pOp->nOpt) ){
1185 break;
1188 if( iOpt==SizeofArray(aFts4Opt) ){
1189 *pzErr = sqlite3_mprintf("unrecognized parameter: %s", z);
1190 rc = SQLITE_ERROR;
1191 }else{
1192 switch( iOpt ){
1193 case 0: /* MATCHINFO */
1194 if( strlen(zVal)!=4 || sqlite3_strnicmp(zVal, "fts3", 4) ){
1195 *pzErr = sqlite3_mprintf("unrecognized matchinfo: %s", zVal);
1196 rc = SQLITE_ERROR;
1198 bNoDocsize = 1;
1199 break;
1201 case 1: /* PREFIX */
1202 sqlite3_free(zPrefix);
1203 zPrefix = zVal;
1204 zVal = 0;
1205 break;
1207 case 2: /* COMPRESS */
1208 sqlite3_free(zCompress);
1209 zCompress = zVal;
1210 zVal = 0;
1211 break;
1213 case 3: /* UNCOMPRESS */
1214 sqlite3_free(zUncompress);
1215 zUncompress = zVal;
1216 zVal = 0;
1217 break;
1219 case 4: /* ORDER */
1220 if( (strlen(zVal)!=3 || sqlite3_strnicmp(zVal, "asc", 3))
1221 && (strlen(zVal)!=4 || sqlite3_strnicmp(zVal, "desc", 4))
1223 *pzErr = sqlite3_mprintf("unrecognized order: %s", zVal);
1224 rc = SQLITE_ERROR;
1226 bDescIdx = (zVal[0]=='d' || zVal[0]=='D');
1227 break;
1229 case 5: /* CONTENT */
1230 sqlite3_free(zContent);
1231 zContent = zVal;
1232 zVal = 0;
1233 break;
1235 case 6: /* LANGUAGEID */
1236 assert( iOpt==6 );
1237 sqlite3_free(zLanguageid);
1238 zLanguageid = zVal;
1239 zVal = 0;
1240 break;
1242 case 7: /* NOTINDEXED */
1243 azNotindexed[nNotindexed++] = zVal;
1244 zVal = 0;
1245 break;
1248 sqlite3_free(zVal);
1252 /* Otherwise, the argument is a column name. */
1253 else {
1254 nString += (int)(strlen(z) + 1);
1255 aCol[nCol++] = z;
1259 /* If a content=xxx option was specified, the following:
1261 ** 1. Ignore any compress= and uncompress= options.
1263 ** 2. If no column names were specified as part of the CREATE VIRTUAL
1264 ** TABLE statement, use all columns from the content table.
1266 if( rc==SQLITE_OK && zContent ){
1267 sqlite3_free(zCompress);
1268 sqlite3_free(zUncompress);
1269 zCompress = 0;
1270 zUncompress = 0;
1271 if( nCol==0 ){
1272 sqlite3_free((void*)aCol);
1273 aCol = 0;
1274 rc = fts3ContentColumns(db, argv[1], zContent, &aCol, &nCol, &nString);
1276 /* If a languageid= option was specified, remove the language id
1277 ** column from the aCol[] array. */
1278 if( rc==SQLITE_OK && zLanguageid ){
1279 int j;
1280 for(j=0; j<nCol; j++){
1281 if( sqlite3_stricmp(zLanguageid, aCol[j])==0 ){
1282 int k;
1283 for(k=j; k<nCol; k++) aCol[k] = aCol[k+1];
1284 nCol--;
1285 break;
1291 if( rc!=SQLITE_OK ) goto fts3_init_out;
1293 if( nCol==0 ){
1294 assert( nString==0 );
1295 aCol[0] = "content";
1296 nString = 8;
1297 nCol = 1;
1300 if( pTokenizer==0 ){
1301 rc = sqlite3Fts3InitTokenizer(pHash, "simple", &pTokenizer, pzErr);
1302 if( rc!=SQLITE_OK ) goto fts3_init_out;
1304 assert( pTokenizer );
1306 rc = fts3PrefixParameter(zPrefix, &nIndex, &aIndex);
1307 if( rc==SQLITE_ERROR ){
1308 assert( zPrefix );
1309 *pzErr = sqlite3_mprintf("error parsing prefix parameter: %s", zPrefix);
1311 if( rc!=SQLITE_OK ) goto fts3_init_out;
1313 /* Allocate and populate the Fts3Table structure. */
1314 nByte = sizeof(Fts3Table) + /* Fts3Table */
1315 nCol * sizeof(char *) + /* azColumn */
1316 nIndex * sizeof(struct Fts3Index) + /* aIndex */
1317 nCol * sizeof(u8) + /* abNotindexed */
1318 nName + /* zName */
1319 nDb + /* zDb */
1320 nString; /* Space for azColumn strings */
1321 p = (Fts3Table*)sqlite3_malloc(nByte);
1322 if( p==0 ){
1323 rc = SQLITE_NOMEM;
1324 goto fts3_init_out;
1326 memset(p, 0, nByte);
1327 p->db = db;
1328 p->nColumn = nCol;
1329 p->nPendingData = 0;
1330 p->azColumn = (char **)&p[1];
1331 p->pTokenizer = pTokenizer;
1332 p->nMaxPendingData = FTS3_MAX_PENDING_DATA;
1333 p->bHasDocsize = (isFts4 && bNoDocsize==0);
1334 p->bHasStat = isFts4;
1335 p->bFts4 = isFts4;
1336 p->bDescIdx = bDescIdx;
1337 p->nAutoincrmerge = 0xff; /* 0xff means setting unknown */
1338 p->zContentTbl = zContent;
1339 p->zLanguageid = zLanguageid;
1340 zContent = 0;
1341 zLanguageid = 0;
1342 TESTONLY( p->inTransaction = -1 );
1343 TESTONLY( p->mxSavepoint = -1 );
1345 p->aIndex = (struct Fts3Index *)&p->azColumn[nCol];
1346 memcpy(p->aIndex, aIndex, sizeof(struct Fts3Index) * nIndex);
1347 p->nIndex = nIndex;
1348 for(i=0; i<nIndex; i++){
1349 fts3HashInit(&p->aIndex[i].hPending, FTS3_HASH_STRING, 1);
1351 p->abNotindexed = (u8 *)&p->aIndex[nIndex];
1353 /* Fill in the zName and zDb fields of the vtab structure. */
1354 zCsr = (char *)&p->abNotindexed[nCol];
1355 p->zName = zCsr;
1356 memcpy(zCsr, argv[2], nName);
1357 zCsr += nName;
1358 p->zDb = zCsr;
1359 memcpy(zCsr, argv[1], nDb);
1360 zCsr += nDb;
1362 /* Fill in the azColumn array */
1363 for(iCol=0; iCol<nCol; iCol++){
1364 char *z;
1365 int n = 0;
1366 z = (char *)sqlite3Fts3NextToken(aCol[iCol], &n);
1367 memcpy(zCsr, z, n);
1368 zCsr[n] = '\0';
1369 sqlite3Fts3Dequote(zCsr);
1370 p->azColumn[iCol] = zCsr;
1371 zCsr += n+1;
1372 assert( zCsr <= &((char *)p)[nByte] );
1375 /* Fill in the abNotindexed array */
1376 for(iCol=0; iCol<nCol; iCol++){
1377 int n = (int)strlen(p->azColumn[iCol]);
1378 for(i=0; i<nNotindexed; i++){
1379 char *zNot = azNotindexed[i];
1380 if( zNot && n==(int)strlen(zNot)
1381 && 0==sqlite3_strnicmp(p->azColumn[iCol], zNot, n)
1383 p->abNotindexed[iCol] = 1;
1384 sqlite3_free(zNot);
1385 azNotindexed[i] = 0;
1389 for(i=0; i<nNotindexed; i++){
1390 if( azNotindexed[i] ){
1391 *pzErr = sqlite3_mprintf("no such column: %s", azNotindexed[i]);
1392 rc = SQLITE_ERROR;
1396 if( rc==SQLITE_OK && (zCompress==0)!=(zUncompress==0) ){
1397 char const *zMiss = (zCompress==0 ? "compress" : "uncompress");
1398 rc = SQLITE_ERROR;
1399 *pzErr = sqlite3_mprintf("missing %s parameter in fts4 constructor", zMiss);
1401 p->zReadExprlist = fts3ReadExprList(p, zUncompress, &rc);
1402 p->zWriteExprlist = fts3WriteExprList(p, zCompress, &rc);
1403 if( rc!=SQLITE_OK ) goto fts3_init_out;
1405 /* If this is an xCreate call, create the underlying tables in the
1406 ** database. TODO: For xConnect(), it could verify that said tables exist.
1408 if( isCreate ){
1409 rc = fts3CreateTables(p);
1412 /* Check to see if a legacy fts3 table has been "upgraded" by the
1413 ** addition of a %_stat table so that it can use incremental merge.
1415 if( !isFts4 && !isCreate ){
1416 p->bHasStat = 2;
1419 /* Figure out the page-size for the database. This is required in order to
1420 ** estimate the cost of loading large doclists from the database. */
1421 fts3DatabasePageSize(&rc, p);
1422 p->nNodeSize = p->nPgsz-35;
1424 /* Declare the table schema to SQLite. */
1425 fts3DeclareVtab(&rc, p);
1427 fts3_init_out:
1428 sqlite3_free(zPrefix);
1429 sqlite3_free(aIndex);
1430 sqlite3_free(zCompress);
1431 sqlite3_free(zUncompress);
1432 sqlite3_free(zContent);
1433 sqlite3_free(zLanguageid);
1434 for(i=0; i<nNotindexed; i++) sqlite3_free(azNotindexed[i]);
1435 sqlite3_free((void *)aCol);
1436 sqlite3_free((void *)azNotindexed);
1437 if( rc!=SQLITE_OK ){
1438 if( p ){
1439 fts3DisconnectMethod((sqlite3_vtab *)p);
1440 }else if( pTokenizer ){
1441 pTokenizer->pModule->xDestroy(pTokenizer);
1443 }else{
1444 assert( p->pSegments==0 );
1445 *ppVTab = &p->base;
1447 return rc;
1451 ** The xConnect() and xCreate() methods for the virtual table. All the
1452 ** work is done in function fts3InitVtab().
1454 static int fts3ConnectMethod(
1455 sqlite3 *db, /* Database connection */
1456 void *pAux, /* Pointer to tokenizer hash table */
1457 int argc, /* Number of elements in argv array */
1458 const char * const *argv, /* xCreate/xConnect argument array */
1459 sqlite3_vtab **ppVtab, /* OUT: New sqlite3_vtab object */
1460 char **pzErr /* OUT: sqlite3_malloc'd error message */
1462 return fts3InitVtab(0, db, pAux, argc, argv, ppVtab, pzErr);
1464 static int fts3CreateMethod(
1465 sqlite3 *db, /* Database connection */
1466 void *pAux, /* Pointer to tokenizer hash table */
1467 int argc, /* Number of elements in argv array */
1468 const char * const *argv, /* xCreate/xConnect argument array */
1469 sqlite3_vtab **ppVtab, /* OUT: New sqlite3_vtab object */
1470 char **pzErr /* OUT: sqlite3_malloc'd error message */
1472 return fts3InitVtab(1, db, pAux, argc, argv, ppVtab, pzErr);
1476 ** Set the pIdxInfo->estimatedRows variable to nRow. Unless this
1477 ** extension is currently being used by a version of SQLite too old to
1478 ** support estimatedRows. In that case this function is a no-op.
1480 static void fts3SetEstimatedRows(sqlite3_index_info *pIdxInfo, i64 nRow){
1481 #if SQLITE_VERSION_NUMBER>=3008002
1482 if( sqlite3_libversion_number()>=3008002 ){
1483 pIdxInfo->estimatedRows = nRow;
1485 #endif
1489 ** Implementation of the xBestIndex method for FTS3 tables. There
1490 ** are three possible strategies, in order of preference:
1492 ** 1. Direct lookup by rowid or docid.
1493 ** 2. Full-text search using a MATCH operator on a non-docid column.
1494 ** 3. Linear scan of %_content table.
1496 static int fts3BestIndexMethod(sqlite3_vtab *pVTab, sqlite3_index_info *pInfo){
1497 Fts3Table *p = (Fts3Table *)pVTab;
1498 int i; /* Iterator variable */
1499 int iCons = -1; /* Index of constraint to use */
1501 int iLangidCons = -1; /* Index of langid=x constraint, if present */
1502 int iDocidGe = -1; /* Index of docid>=x constraint, if present */
1503 int iDocidLe = -1; /* Index of docid<=x constraint, if present */
1504 int iIdx;
1506 /* By default use a full table scan. This is an expensive option,
1507 ** so search through the constraints to see if a more efficient
1508 ** strategy is possible.
1510 pInfo->idxNum = FTS3_FULLSCAN_SEARCH;
1511 pInfo->estimatedCost = 5000000;
1512 for(i=0; i<pInfo->nConstraint; i++){
1513 int bDocid; /* True if this constraint is on docid */
1514 struct sqlite3_index_constraint *pCons = &pInfo->aConstraint[i];
1515 if( pCons->usable==0 ){
1516 if( pCons->op==SQLITE_INDEX_CONSTRAINT_MATCH ){
1517 /* There exists an unusable MATCH constraint. This means that if
1518 ** the planner does elect to use the results of this call as part
1519 ** of the overall query plan the user will see an "unable to use
1520 ** function MATCH in the requested context" error. To discourage
1521 ** this, return a very high cost here. */
1522 pInfo->idxNum = FTS3_FULLSCAN_SEARCH;
1523 pInfo->estimatedCost = 1e50;
1524 fts3SetEstimatedRows(pInfo, ((sqlite3_int64)1) << 50);
1525 return SQLITE_OK;
1527 continue;
1530 bDocid = (pCons->iColumn<0 || pCons->iColumn==p->nColumn+1);
1532 /* A direct lookup on the rowid or docid column. Assign a cost of 1.0. */
1533 if( iCons<0 && pCons->op==SQLITE_INDEX_CONSTRAINT_EQ && bDocid ){
1534 pInfo->idxNum = FTS3_DOCID_SEARCH;
1535 pInfo->estimatedCost = 1.0;
1536 iCons = i;
1539 /* A MATCH constraint. Use a full-text search.
1541 ** If there is more than one MATCH constraint available, use the first
1542 ** one encountered. If there is both a MATCH constraint and a direct
1543 ** rowid/docid lookup, prefer the MATCH strategy. This is done even
1544 ** though the rowid/docid lookup is faster than a MATCH query, selecting
1545 ** it would lead to an "unable to use function MATCH in the requested
1546 ** context" error.
1548 if( pCons->op==SQLITE_INDEX_CONSTRAINT_MATCH
1549 && pCons->iColumn>=0 && pCons->iColumn<=p->nColumn
1551 pInfo->idxNum = FTS3_FULLTEXT_SEARCH + pCons->iColumn;
1552 pInfo->estimatedCost = 2.0;
1553 iCons = i;
1556 /* Equality constraint on the langid column */
1557 if( pCons->op==SQLITE_INDEX_CONSTRAINT_EQ
1558 && pCons->iColumn==p->nColumn + 2
1560 iLangidCons = i;
1563 if( bDocid ){
1564 switch( pCons->op ){
1565 case SQLITE_INDEX_CONSTRAINT_GE:
1566 case SQLITE_INDEX_CONSTRAINT_GT:
1567 iDocidGe = i;
1568 break;
1570 case SQLITE_INDEX_CONSTRAINT_LE:
1571 case SQLITE_INDEX_CONSTRAINT_LT:
1572 iDocidLe = i;
1573 break;
1578 iIdx = 1;
1579 if( iCons>=0 ){
1580 pInfo->aConstraintUsage[iCons].argvIndex = iIdx++;
1581 pInfo->aConstraintUsage[iCons].omit = 1;
1583 if( iLangidCons>=0 ){
1584 pInfo->idxNum |= FTS3_HAVE_LANGID;
1585 pInfo->aConstraintUsage[iLangidCons].argvIndex = iIdx++;
1587 if( iDocidGe>=0 ){
1588 pInfo->idxNum |= FTS3_HAVE_DOCID_GE;
1589 pInfo->aConstraintUsage[iDocidGe].argvIndex = iIdx++;
1591 if( iDocidLe>=0 ){
1592 pInfo->idxNum |= FTS3_HAVE_DOCID_LE;
1593 pInfo->aConstraintUsage[iDocidLe].argvIndex = iIdx++;
1596 /* Regardless of the strategy selected, FTS can deliver rows in rowid (or
1597 ** docid) order. Both ascending and descending are possible.
1599 if( pInfo->nOrderBy==1 ){
1600 struct sqlite3_index_orderby *pOrder = &pInfo->aOrderBy[0];
1601 if( pOrder->iColumn<0 || pOrder->iColumn==p->nColumn+1 ){
1602 if( pOrder->desc ){
1603 pInfo->idxStr = "DESC";
1604 }else{
1605 pInfo->idxStr = "ASC";
1607 pInfo->orderByConsumed = 1;
1611 assert( p->pSegments==0 );
1612 return SQLITE_OK;
1616 ** Implementation of xOpen method.
1618 static int fts3OpenMethod(sqlite3_vtab *pVTab, sqlite3_vtab_cursor **ppCsr){
1619 sqlite3_vtab_cursor *pCsr; /* Allocated cursor */
1621 UNUSED_PARAMETER(pVTab);
1623 /* Allocate a buffer large enough for an Fts3Cursor structure. If the
1624 ** allocation succeeds, zero it and return SQLITE_OK. Otherwise,
1625 ** if the allocation fails, return SQLITE_NOMEM.
1627 *ppCsr = pCsr = (sqlite3_vtab_cursor *)sqlite3_malloc(sizeof(Fts3Cursor));
1628 if( !pCsr ){
1629 return SQLITE_NOMEM;
1631 memset(pCsr, 0, sizeof(Fts3Cursor));
1632 return SQLITE_OK;
1636 ** Close the cursor. For additional information see the documentation
1637 ** on the xClose method of the virtual table interface.
1639 static int fts3CloseMethod(sqlite3_vtab_cursor *pCursor){
1640 Fts3Cursor *pCsr = (Fts3Cursor *)pCursor;
1641 assert( ((Fts3Table *)pCsr->base.pVtab)->pSegments==0 );
1642 sqlite3_finalize(pCsr->pStmt);
1643 sqlite3Fts3ExprFree(pCsr->pExpr);
1644 sqlite3Fts3FreeDeferredTokens(pCsr);
1645 sqlite3_free(pCsr->aDoclist);
1646 sqlite3_free(pCsr->aMatchinfo);
1647 assert( ((Fts3Table *)pCsr->base.pVtab)->pSegments==0 );
1648 sqlite3_free(pCsr);
1649 return SQLITE_OK;
1653 ** If pCsr->pStmt has not been prepared (i.e. if pCsr->pStmt==0), then
1654 ** compose and prepare an SQL statement of the form:
1656 ** "SELECT <columns> FROM %_content WHERE rowid = ?"
1658 ** (or the equivalent for a content=xxx table) and set pCsr->pStmt to
1659 ** it. If an error occurs, return an SQLite error code.
1661 ** Otherwise, set *ppStmt to point to pCsr->pStmt and return SQLITE_OK.
1663 static int fts3CursorSeekStmt(Fts3Cursor *pCsr, sqlite3_stmt **ppStmt){
1664 int rc = SQLITE_OK;
1665 if( pCsr->pStmt==0 ){
1666 Fts3Table *p = (Fts3Table *)pCsr->base.pVtab;
1667 char *zSql;
1668 zSql = sqlite3_mprintf("SELECT %s WHERE rowid = ?", p->zReadExprlist);
1669 if( !zSql ) return SQLITE_NOMEM;
1670 rc = sqlite3_prepare_v2(p->db, zSql, -1, &pCsr->pStmt, 0);
1671 sqlite3_free(zSql);
1673 *ppStmt = pCsr->pStmt;
1674 return rc;
1678 ** Position the pCsr->pStmt statement so that it is on the row
1679 ** of the %_content table that contains the last match. Return
1680 ** SQLITE_OK on success.
1682 static int fts3CursorSeek(sqlite3_context *pContext, Fts3Cursor *pCsr){
1683 int rc = SQLITE_OK;
1684 if( pCsr->isRequireSeek ){
1685 sqlite3_stmt *pStmt = 0;
1687 rc = fts3CursorSeekStmt(pCsr, &pStmt);
1688 if( rc==SQLITE_OK ){
1689 sqlite3_bind_int64(pCsr->pStmt, 1, pCsr->iPrevId);
1690 pCsr->isRequireSeek = 0;
1691 if( SQLITE_ROW==sqlite3_step(pCsr->pStmt) ){
1692 return SQLITE_OK;
1693 }else{
1694 rc = sqlite3_reset(pCsr->pStmt);
1695 if( rc==SQLITE_OK && ((Fts3Table *)pCsr->base.pVtab)->zContentTbl==0 ){
1696 /* If no row was found and no error has occurred, then the %_content
1697 ** table is missing a row that is present in the full-text index.
1698 ** The data structures are corrupt. */
1699 rc = FTS_CORRUPT_VTAB;
1700 pCsr->isEof = 1;
1706 if( rc!=SQLITE_OK && pContext ){
1707 sqlite3_result_error_code(pContext, rc);
1709 return rc;
1713 ** This function is used to process a single interior node when searching
1714 ** a b-tree for a term or term prefix. The node data is passed to this
1715 ** function via the zNode/nNode parameters. The term to search for is
1716 ** passed in zTerm/nTerm.
1718 ** If piFirst is not NULL, then this function sets *piFirst to the blockid
1719 ** of the child node that heads the sub-tree that may contain the term.
1721 ** If piLast is not NULL, then *piLast is set to the right-most child node
1722 ** that heads a sub-tree that may contain a term for which zTerm/nTerm is
1723 ** a prefix.
1725 ** If an OOM error occurs, SQLITE_NOMEM is returned. Otherwise, SQLITE_OK.
1727 static int fts3ScanInteriorNode(
1728 const char *zTerm, /* Term to select leaves for */
1729 int nTerm, /* Size of term zTerm in bytes */
1730 const char *zNode, /* Buffer containing segment interior node */
1731 int nNode, /* Size of buffer at zNode */
1732 sqlite3_int64 *piFirst, /* OUT: Selected child node */
1733 sqlite3_int64 *piLast /* OUT: Selected child node */
1735 int rc = SQLITE_OK; /* Return code */
1736 const char *zCsr = zNode; /* Cursor to iterate through node */
1737 const char *zEnd = &zCsr[nNode];/* End of interior node buffer */
1738 char *zBuffer = 0; /* Buffer to load terms into */
1739 int nAlloc = 0; /* Size of allocated buffer */
1740 int isFirstTerm = 1; /* True when processing first term on page */
1741 sqlite3_int64 iChild; /* Block id of child node to descend to */
1743 /* Skip over the 'height' varint that occurs at the start of every
1744 ** interior node. Then load the blockid of the left-child of the b-tree
1745 ** node into variable iChild.
1747 ** Even if the data structure on disk is corrupted, this (reading two
1748 ** varints from the buffer) does not risk an overread. If zNode is a
1749 ** root node, then the buffer comes from a SELECT statement. SQLite does
1750 ** not make this guarantee explicitly, but in practice there are always
1751 ** either more than 20 bytes of allocated space following the nNode bytes of
1752 ** contents, or two zero bytes. Or, if the node is read from the %_segments
1753 ** table, then there are always 20 bytes of zeroed padding following the
1754 ** nNode bytes of content (see sqlite3Fts3ReadBlock() for details).
1756 zCsr += sqlite3Fts3GetVarint(zCsr, &iChild);
1757 zCsr += sqlite3Fts3GetVarint(zCsr, &iChild);
1758 if( zCsr>zEnd ){
1759 return FTS_CORRUPT_VTAB;
1762 while( zCsr<zEnd && (piFirst || piLast) ){
1763 int cmp; /* memcmp() result */
1764 int nSuffix; /* Size of term suffix */
1765 int nPrefix = 0; /* Size of term prefix */
1766 int nBuffer; /* Total term size */
1768 /* Load the next term on the node into zBuffer. Use realloc() to expand
1769 ** the size of zBuffer if required. */
1770 if( !isFirstTerm ){
1771 zCsr += fts3GetVarint32(zCsr, &nPrefix);
1773 isFirstTerm = 0;
1774 zCsr += fts3GetVarint32(zCsr, &nSuffix);
1776 /* NOTE(shess): Previous code checked for negative nPrefix and
1777 ** nSuffix and suffix overrunning zEnd. Additionally corrupt if
1778 ** the prefix is longer than the previous term, or if the suffix
1779 ** causes overflow.
1781 if( nPrefix<0 || nSuffix<0 /* || nPrefix>nBuffer */
1782 || &zCsr[nSuffix]<zCsr || &zCsr[nSuffix]>zEnd ){
1783 rc = SQLITE_CORRUPT;
1784 goto finish_scan;
1786 if( nPrefix+nSuffix>nAlloc ){
1787 char *zNew;
1788 nAlloc = (nPrefix+nSuffix) * 2;
1789 zNew = (char *)sqlite3_realloc(zBuffer, nAlloc);
1790 if( !zNew ){
1791 rc = SQLITE_NOMEM;
1792 goto finish_scan;
1794 zBuffer = zNew;
1796 assert( zBuffer );
1797 memcpy(&zBuffer[nPrefix], zCsr, nSuffix);
1798 nBuffer = nPrefix + nSuffix;
1799 zCsr += nSuffix;
1801 /* Compare the term we are searching for with the term just loaded from
1802 ** the interior node. If the specified term is greater than or equal
1803 ** to the term from the interior node, then all terms on the sub-tree
1804 ** headed by node iChild are smaller than zTerm. No need to search
1805 ** iChild.
1807 ** If the interior node term is larger than the specified term, then
1808 ** the tree headed by iChild may contain the specified term.
1810 cmp = memcmp(zTerm, zBuffer, (nBuffer>nTerm ? nTerm : nBuffer));
1811 if( piFirst && (cmp<0 || (cmp==0 && nBuffer>nTerm)) ){
1812 *piFirst = iChild;
1813 piFirst = 0;
1816 if( piLast && cmp<0 ){
1817 *piLast = iChild;
1818 piLast = 0;
1821 iChild++;
1824 if( piFirst ) *piFirst = iChild;
1825 if( piLast ) *piLast = iChild;
1827 finish_scan:
1828 sqlite3_free(zBuffer);
1829 return rc;
1834 ** The buffer pointed to by argument zNode (size nNode bytes) contains an
1835 ** interior node of a b-tree segment. The zTerm buffer (size nTerm bytes)
1836 ** contains a term. This function searches the sub-tree headed by the zNode
1837 ** node for the range of leaf nodes that may contain the specified term
1838 ** or terms for which the specified term is a prefix.
1840 ** If piLeaf is not NULL, then *piLeaf is set to the blockid of the
1841 ** left-most leaf node in the tree that may contain the specified term.
1842 ** If piLeaf2 is not NULL, then *piLeaf2 is set to the blockid of the
1843 ** right-most leaf node that may contain a term for which the specified
1844 ** term is a prefix.
1846 ** It is possible that the range of returned leaf nodes does not contain
1847 ** the specified term or any terms for which it is a prefix. However, if the
1848 ** segment does contain any such terms, they are stored within the identified
1849 ** range. Because this function only inspects interior segment nodes (and
1850 ** never loads leaf nodes into memory), it is not possible to be sure.
1852 ** If an error occurs, an error code other than SQLITE_OK is returned.
1854 static int fts3SelectLeaf(
1855 Fts3Table *p, /* Virtual table handle */
1856 const char *zTerm, /* Term to select leaves for */
1857 int nTerm, /* Size of term zTerm in bytes */
1858 const char *zNode, /* Buffer containing segment interior node */
1859 int nNode, /* Size of buffer at zNode */
1860 sqlite3_int64 *piLeaf, /* Selected leaf node */
1861 sqlite3_int64 *piLeaf2 /* Selected leaf node */
1863 int rc; /* Return code */
1864 int iHeight; /* Height of this node in tree */
1866 assert( piLeaf || piLeaf2 );
1868 fts3GetVarint32(zNode, &iHeight);
1869 rc = fts3ScanInteriorNode(zTerm, nTerm, zNode, nNode, piLeaf, piLeaf2);
1870 assert( !piLeaf2 || !piLeaf || rc!=SQLITE_OK || (*piLeaf<=*piLeaf2) );
1872 if( rc==SQLITE_OK && iHeight>1 ){
1873 char *zBlob = 0; /* Blob read from %_segments table */
1874 int nBlob; /* Size of zBlob in bytes */
1876 if( piLeaf && piLeaf2 && (*piLeaf!=*piLeaf2) ){
1877 rc = sqlite3Fts3ReadBlock(p, *piLeaf, &zBlob, &nBlob, 0);
1878 if( rc==SQLITE_OK ){
1879 rc = fts3SelectLeaf(p, zTerm, nTerm, zBlob, nBlob, piLeaf, 0);
1881 sqlite3_free(zBlob);
1882 piLeaf = 0;
1883 zBlob = 0;
1886 if( rc==SQLITE_OK ){
1887 rc = sqlite3Fts3ReadBlock(p, piLeaf?*piLeaf:*piLeaf2, &zBlob, &nBlob, 0);
1889 if( rc==SQLITE_OK ){
1890 rc = fts3SelectLeaf(p, zTerm, nTerm, zBlob, nBlob, piLeaf, piLeaf2);
1892 sqlite3_free(zBlob);
1895 return rc;
1899 ** This function is used to create delta-encoded serialized lists of FTS3
1900 ** varints. Each call to this function appends a single varint to a list.
1902 static void fts3PutDeltaVarint(
1903 char **pp, /* IN/OUT: Output pointer */
1904 sqlite3_int64 *piPrev, /* IN/OUT: Previous value written to list */
1905 sqlite3_int64 iVal /* Write this value to the list */
1907 assert( iVal-*piPrev > 0 || (*piPrev==0 && iVal==0) );
1908 *pp += sqlite3Fts3PutVarint(*pp, iVal-*piPrev);
1909 *piPrev = iVal;
1913 ** When this function is called, *ppPoslist is assumed to point to the
1914 ** start of a position-list. After it returns, *ppPoslist points to the
1915 ** first byte after the position-list.
1917 ** A position list is list of positions (delta encoded) and columns for
1918 ** a single document record of a doclist. So, in other words, this
1919 ** routine advances *ppPoslist so that it points to the next docid in
1920 ** the doclist, or to the first byte past the end of the doclist.
1922 ** If pp is not NULL, then the contents of the position list are copied
1923 ** to *pp. *pp is set to point to the first byte past the last byte copied
1924 ** before this function returns.
1926 static void fts3PoslistCopy(char **pp, char **ppPoslist){
1927 char *pEnd = *ppPoslist;
1928 char c = 0;
1930 /* The end of a position list is marked by a zero encoded as an FTS3
1931 ** varint. A single POS_END (0) byte. Except, if the 0 byte is preceded by
1932 ** a byte with the 0x80 bit set, then it is not a varint 0, but the tail
1933 ** of some other, multi-byte, value.
1935 ** The following while-loop moves pEnd to point to the first byte that is not
1936 ** immediately preceded by a byte with the 0x80 bit set. Then increments
1937 ** pEnd once more so that it points to the byte immediately following the
1938 ** last byte in the position-list.
1940 while( *pEnd | c ){
1941 c = *pEnd++ & 0x80;
1942 testcase( c!=0 && (*pEnd)==0 );
1944 pEnd++; /* Advance past the POS_END terminator byte */
1946 if( pp ){
1947 int n = (int)(pEnd - *ppPoslist);
1948 char *p = *pp;
1949 memcpy(p, *ppPoslist, n);
1950 p += n;
1951 *pp = p;
1953 *ppPoslist = pEnd;
1957 ** When this function is called, *ppPoslist is assumed to point to the
1958 ** start of a column-list. After it returns, *ppPoslist points to the
1959 ** to the terminator (POS_COLUMN or POS_END) byte of the column-list.
1961 ** A column-list is list of delta-encoded positions for a single column
1962 ** within a single document within a doclist.
1964 ** The column-list is terminated either by a POS_COLUMN varint (1) or
1965 ** a POS_END varint (0). This routine leaves *ppPoslist pointing to
1966 ** the POS_COLUMN or POS_END that terminates the column-list.
1968 ** If pp is not NULL, then the contents of the column-list are copied
1969 ** to *pp. *pp is set to point to the first byte past the last byte copied
1970 ** before this function returns. The POS_COLUMN or POS_END terminator
1971 ** is not copied into *pp.
1973 static void fts3ColumnlistCopy(char **pp, char **ppPoslist){
1974 char *pEnd = *ppPoslist;
1975 char c = 0;
1977 /* A column-list is terminated by either a 0x01 or 0x00 byte that is
1978 ** not part of a multi-byte varint.
1980 while( 0xFE & (*pEnd | c) ){
1981 c = *pEnd++ & 0x80;
1982 testcase( c!=0 && ((*pEnd)&0xfe)==0 );
1984 if( pp ){
1985 int n = (int)(pEnd - *ppPoslist);
1986 char *p = *pp;
1987 memcpy(p, *ppPoslist, n);
1988 p += n;
1989 *pp = p;
1991 *ppPoslist = pEnd;
1995 ** Value used to signify the end of an position-list. This is safe because
1996 ** it is not possible to have a document with 2^31 terms.
1998 #define POSITION_LIST_END 0x7fffffff
2001 ** This function is used to help parse position-lists. When this function is
2002 ** called, *pp may point to the start of the next varint in the position-list
2003 ** being parsed, or it may point to 1 byte past the end of the position-list
2004 ** (in which case **pp will be a terminator bytes POS_END (0) or
2005 ** (1)).
2007 ** If *pp points past the end of the current position-list, set *pi to
2008 ** POSITION_LIST_END and return. Otherwise, read the next varint from *pp,
2009 ** increment the current value of *pi by the value read, and set *pp to
2010 ** point to the next value before returning.
2012 ** Before calling this routine *pi must be initialized to the value of
2013 ** the previous position, or zero if we are reading the first position
2014 ** in the position-list. Because positions are delta-encoded, the value
2015 ** of the previous position is needed in order to compute the value of
2016 ** the next position.
2018 static void fts3ReadNextPos(
2019 char **pp, /* IN/OUT: Pointer into position-list buffer */
2020 sqlite3_int64 *pi /* IN/OUT: Value read from position-list */
2022 if( (**pp)&0xFE ){
2023 fts3GetDeltaVarint(pp, pi);
2024 *pi -= 2;
2025 }else{
2026 *pi = POSITION_LIST_END;
2031 ** If parameter iCol is not 0, write an POS_COLUMN (1) byte followed by
2032 ** the value of iCol encoded as a varint to *pp. This will start a new
2033 ** column list.
2035 ** Set *pp to point to the byte just after the last byte written before
2036 ** returning (do not modify it if iCol==0). Return the total number of bytes
2037 ** written (0 if iCol==0).
2039 static int fts3PutColNumber(char **pp, int iCol){
2040 int n = 0; /* Number of bytes written */
2041 if( iCol ){
2042 char *p = *pp; /* Output pointer */
2043 n = 1 + sqlite3Fts3PutVarint(&p[1], iCol);
2044 *p = 0x01;
2045 *pp = &p[n];
2047 return n;
2051 ** Compute the union of two position lists. The output written
2052 ** into *pp contains all positions of both *pp1 and *pp2 in sorted
2053 ** order and with any duplicates removed. All pointers are
2054 ** updated appropriately. The caller is responsible for insuring
2055 ** that there is enough space in *pp to hold the complete output.
2057 static void fts3PoslistMerge(
2058 char **pp, /* Output buffer */
2059 char **pp1, /* Left input list */
2060 char **pp2 /* Right input list */
2062 char *p = *pp;
2063 char *p1 = *pp1;
2064 char *p2 = *pp2;
2066 while( *p1 || *p2 ){
2067 int iCol1; /* The current column index in pp1 */
2068 int iCol2; /* The current column index in pp2 */
2070 if( *p1==POS_COLUMN ) fts3GetVarint32(&p1[1], &iCol1);
2071 else if( *p1==POS_END ) iCol1 = POSITION_LIST_END;
2072 else iCol1 = 0;
2074 if( *p2==POS_COLUMN ) fts3GetVarint32(&p2[1], &iCol2);
2075 else if( *p2==POS_END ) iCol2 = POSITION_LIST_END;
2076 else iCol2 = 0;
2078 if( iCol1==iCol2 ){
2079 sqlite3_int64 i1 = 0; /* Last position from pp1 */
2080 sqlite3_int64 i2 = 0; /* Last position from pp2 */
2081 sqlite3_int64 iPrev = 0;
2082 int n = fts3PutColNumber(&p, iCol1);
2083 p1 += n;
2084 p2 += n;
2086 /* At this point, both p1 and p2 point to the start of column-lists
2087 ** for the same column (the column with index iCol1 and iCol2).
2088 ** A column-list is a list of non-negative delta-encoded varints, each
2089 ** incremented by 2 before being stored. Each list is terminated by a
2090 ** POS_END (0) or POS_COLUMN (1). The following block merges the two lists
2091 ** and writes the results to buffer p. p is left pointing to the byte
2092 ** after the list written. No terminator (POS_END or POS_COLUMN) is
2093 ** written to the output.
2095 fts3GetDeltaVarint(&p1, &i1);
2096 fts3GetDeltaVarint(&p2, &i2);
2097 do {
2098 fts3PutDeltaVarint(&p, &iPrev, (i1<i2) ? i1 : i2);
2099 iPrev -= 2;
2100 if( i1==i2 ){
2101 fts3ReadNextPos(&p1, &i1);
2102 fts3ReadNextPos(&p2, &i2);
2103 }else if( i1<i2 ){
2104 fts3ReadNextPos(&p1, &i1);
2105 }else{
2106 fts3ReadNextPos(&p2, &i2);
2108 }while( i1!=POSITION_LIST_END || i2!=POSITION_LIST_END );
2109 }else if( iCol1<iCol2 ){
2110 p1 += fts3PutColNumber(&p, iCol1);
2111 fts3ColumnlistCopy(&p, &p1);
2112 }else{
2113 p2 += fts3PutColNumber(&p, iCol2);
2114 fts3ColumnlistCopy(&p, &p2);
2118 *p++ = POS_END;
2119 *pp = p;
2120 *pp1 = p1 + 1;
2121 *pp2 = p2 + 1;
2125 ** This function is used to merge two position lists into one. When it is
2126 ** called, *pp1 and *pp2 must both point to position lists. A position-list is
2127 ** the part of a doclist that follows each document id. For example, if a row
2128 ** contains:
2130 ** 'a b c'|'x y z'|'a b b a'
2132 ** Then the position list for this row for token 'b' would consist of:
2134 ** 0x02 0x01 0x02 0x03 0x03 0x00
2136 ** When this function returns, both *pp1 and *pp2 are left pointing to the
2137 ** byte following the 0x00 terminator of their respective position lists.
2139 ** If isSaveLeft is 0, an entry is added to the output position list for
2140 ** each position in *pp2 for which there exists one or more positions in
2141 ** *pp1 so that (pos(*pp2)>pos(*pp1) && pos(*pp2)-pos(*pp1)<=nToken). i.e.
2142 ** when the *pp1 token appears before the *pp2 token, but not more than nToken
2143 ** slots before it.
2145 ** e.g. nToken==1 searches for adjacent positions.
2147 static int fts3PoslistPhraseMerge(
2148 char **pp, /* IN/OUT: Preallocated output buffer */
2149 int nToken, /* Maximum difference in token positions */
2150 int isSaveLeft, /* Save the left position */
2151 int isExact, /* If *pp1 is exactly nTokens before *pp2 */
2152 char **pp1, /* IN/OUT: Left input list */
2153 char **pp2 /* IN/OUT: Right input list */
2155 char *p = *pp;
2156 char *p1 = *pp1;
2157 char *p2 = *pp2;
2158 int iCol1 = 0;
2159 int iCol2 = 0;
2161 /* Never set both isSaveLeft and isExact for the same invocation. */
2162 assert( isSaveLeft==0 || isExact==0 );
2164 assert( p!=0 && *p1!=0 && *p2!=0 );
2165 if( *p1==POS_COLUMN ){
2166 p1++;
2167 p1 += fts3GetVarint32(p1, &iCol1);
2169 if( *p2==POS_COLUMN ){
2170 p2++;
2171 p2 += fts3GetVarint32(p2, &iCol2);
2174 while( 1 ){
2175 if( iCol1==iCol2 ){
2176 char *pSave = p;
2177 sqlite3_int64 iPrev = 0;
2178 sqlite3_int64 iPos1 = 0;
2179 sqlite3_int64 iPos2 = 0;
2181 if( iCol1 ){
2182 *p++ = POS_COLUMN;
2183 p += sqlite3Fts3PutVarint(p, iCol1);
2186 assert( *p1!=POS_END && *p1!=POS_COLUMN );
2187 assert( *p2!=POS_END && *p2!=POS_COLUMN );
2188 fts3GetDeltaVarint(&p1, &iPos1); iPos1 -= 2;
2189 fts3GetDeltaVarint(&p2, &iPos2); iPos2 -= 2;
2191 while( 1 ){
2192 if( iPos2==iPos1+nToken
2193 || (isExact==0 && iPos2>iPos1 && iPos2<=iPos1+nToken)
2195 sqlite3_int64 iSave;
2196 iSave = isSaveLeft ? iPos1 : iPos2;
2197 fts3PutDeltaVarint(&p, &iPrev, iSave+2); iPrev -= 2;
2198 pSave = 0;
2199 assert( p );
2201 if( (!isSaveLeft && iPos2<=(iPos1+nToken)) || iPos2<=iPos1 ){
2202 if( (*p2&0xFE)==0 ) break;
2203 fts3GetDeltaVarint(&p2, &iPos2); iPos2 -= 2;
2204 }else{
2205 if( (*p1&0xFE)==0 ) break;
2206 fts3GetDeltaVarint(&p1, &iPos1); iPos1 -= 2;
2210 if( pSave ){
2211 assert( pp && p );
2212 p = pSave;
2215 fts3ColumnlistCopy(0, &p1);
2216 fts3ColumnlistCopy(0, &p2);
2217 assert( (*p1&0xFE)==0 && (*p2&0xFE)==0 );
2218 if( 0==*p1 || 0==*p2 ) break;
2220 p1++;
2221 p1 += fts3GetVarint32(p1, &iCol1);
2222 p2++;
2223 p2 += fts3GetVarint32(p2, &iCol2);
2226 /* Advance pointer p1 or p2 (whichever corresponds to the smaller of
2227 ** iCol1 and iCol2) so that it points to either the 0x00 that marks the
2228 ** end of the position list, or the 0x01 that precedes the next
2229 ** column-number in the position list.
2231 else if( iCol1<iCol2 ){
2232 fts3ColumnlistCopy(0, &p1);
2233 if( 0==*p1 ) break;
2234 p1++;
2235 p1 += fts3GetVarint32(p1, &iCol1);
2236 }else{
2237 fts3ColumnlistCopy(0, &p2);
2238 if( 0==*p2 ) break;
2239 p2++;
2240 p2 += fts3GetVarint32(p2, &iCol2);
2244 fts3PoslistCopy(0, &p2);
2245 fts3PoslistCopy(0, &p1);
2246 *pp1 = p1;
2247 *pp2 = p2;
2248 if( *pp==p ){
2249 return 0;
2251 *p++ = 0x00;
2252 *pp = p;
2253 return 1;
2257 ** Merge two position-lists as required by the NEAR operator. The argument
2258 ** position lists correspond to the left and right phrases of an expression
2259 ** like:
2261 ** "phrase 1" NEAR "phrase number 2"
2263 ** Position list *pp1 corresponds to the left-hand side of the NEAR
2264 ** expression and *pp2 to the right. As usual, the indexes in the position
2265 ** lists are the offsets of the last token in each phrase (tokens "1" and "2"
2266 ** in the example above).
2268 ** The output position list - written to *pp - is a copy of *pp2 with those
2269 ** entries that are not sufficiently NEAR entries in *pp1 removed.
2271 static int fts3PoslistNearMerge(
2272 char **pp, /* Output buffer */
2273 char *aTmp, /* Temporary buffer space */
2274 int nRight, /* Maximum difference in token positions */
2275 int nLeft, /* Maximum difference in token positions */
2276 char **pp1, /* IN/OUT: Left input list */
2277 char **pp2 /* IN/OUT: Right input list */
2279 char *p1 = *pp1;
2280 char *p2 = *pp2;
2282 char *pTmp1 = aTmp;
2283 char *pTmp2;
2284 char *aTmp2;
2285 int res = 1;
2287 fts3PoslistPhraseMerge(&pTmp1, nRight, 0, 0, pp1, pp2);
2288 aTmp2 = pTmp2 = pTmp1;
2289 *pp1 = p1;
2290 *pp2 = p2;
2291 fts3PoslistPhraseMerge(&pTmp2, nLeft, 1, 0, pp2, pp1);
2292 if( pTmp1!=aTmp && pTmp2!=aTmp2 ){
2293 fts3PoslistMerge(pp, &aTmp, &aTmp2);
2294 }else if( pTmp1!=aTmp ){
2295 fts3PoslistCopy(pp, &aTmp);
2296 }else if( pTmp2!=aTmp2 ){
2297 fts3PoslistCopy(pp, &aTmp2);
2298 }else{
2299 res = 0;
2302 return res;
2306 ** An instance of this function is used to merge together the (potentially
2307 ** large number of) doclists for each term that matches a prefix query.
2308 ** See function fts3TermSelectMerge() for details.
2310 typedef struct TermSelect TermSelect;
2311 struct TermSelect {
2312 char *aaOutput[16]; /* Malloc'd output buffers */
2313 int anOutput[16]; /* Size each output buffer in bytes */
2317 ** This function is used to read a single varint from a buffer. Parameter
2318 ** pEnd points 1 byte past the end of the buffer. When this function is
2319 ** called, if *pp points to pEnd or greater, then the end of the buffer
2320 ** has been reached. In this case *pp is set to 0 and the function returns.
2322 ** If *pp does not point to or past pEnd, then a single varint is read
2323 ** from *pp. *pp is then set to point 1 byte past the end of the read varint.
2325 ** If bDescIdx is false, the value read is added to *pVal before returning.
2326 ** If it is true, the value read is subtracted from *pVal before this
2327 ** function returns.
2329 static void fts3GetDeltaVarint3(
2330 char **pp, /* IN/OUT: Point to read varint from */
2331 char *pEnd, /* End of buffer */
2332 int bDescIdx, /* True if docids are descending */
2333 sqlite3_int64 *pVal /* IN/OUT: Integer value */
2335 if( *pp>=pEnd ){
2336 *pp = 0;
2337 }else{
2338 sqlite3_int64 iVal;
2339 *pp += sqlite3Fts3GetVarint(*pp, &iVal);
2340 if( bDescIdx ){
2341 *pVal -= iVal;
2342 }else{
2343 *pVal += iVal;
2349 ** This function is used to write a single varint to a buffer. The varint
2350 ** is written to *pp. Before returning, *pp is set to point 1 byte past the
2351 ** end of the value written.
2353 ** If *pbFirst is zero when this function is called, the value written to
2354 ** the buffer is that of parameter iVal.
2356 ** If *pbFirst is non-zero when this function is called, then the value
2357 ** written is either (iVal-*piPrev) (if bDescIdx is zero) or (*piPrev-iVal)
2358 ** (if bDescIdx is non-zero).
2360 ** Before returning, this function always sets *pbFirst to 1 and *piPrev
2361 ** to the value of parameter iVal.
2363 static void fts3PutDeltaVarint3(
2364 char **pp, /* IN/OUT: Output pointer */
2365 int bDescIdx, /* True for descending docids */
2366 sqlite3_int64 *piPrev, /* IN/OUT: Previous value written to list */
2367 int *pbFirst, /* IN/OUT: True after first int written */
2368 sqlite3_int64 iVal /* Write this value to the list */
2370 sqlite3_int64 iWrite;
2371 if( bDescIdx==0 || *pbFirst==0 ){
2372 iWrite = iVal - *piPrev;
2373 }else{
2374 iWrite = *piPrev - iVal;
2376 assert( *pbFirst || *piPrev==0 );
2377 assert( *pbFirst==0 || iWrite>0 );
2378 *pp += sqlite3Fts3PutVarint(*pp, iWrite);
2379 *piPrev = iVal;
2380 *pbFirst = 1;
2385 ** This macro is used by various functions that merge doclists. The two
2386 ** arguments are 64-bit docid values. If the value of the stack variable
2387 ** bDescDoclist is 0 when this macro is invoked, then it returns (i1-i2).
2388 ** Otherwise, (i2-i1).
2390 ** Using this makes it easier to write code that can merge doclists that are
2391 ** sorted in either ascending or descending order.
2393 #define DOCID_CMP(i1, i2) ((bDescDoclist?-1:1) * (i1-i2))
2396 ** This function does an "OR" merge of two doclists (output contains all
2397 ** positions contained in either argument doclist). If the docids in the
2398 ** input doclists are sorted in ascending order, parameter bDescDoclist
2399 ** should be false. If they are sorted in ascending order, it should be
2400 ** passed a non-zero value.
2402 ** If no error occurs, *paOut is set to point at an sqlite3_malloc'd buffer
2403 ** containing the output doclist and SQLITE_OK is returned. In this case
2404 ** *pnOut is set to the number of bytes in the output doclist.
2406 ** If an error occurs, an SQLite error code is returned. The output values
2407 ** are undefined in this case.
2409 static int fts3DoclistOrMerge(
2410 int bDescDoclist, /* True if arguments are desc */
2411 char *a1, int n1, /* First doclist */
2412 char *a2, int n2, /* Second doclist */
2413 char **paOut, int *pnOut /* OUT: Malloc'd doclist */
2415 sqlite3_int64 i1 = 0;
2416 sqlite3_int64 i2 = 0;
2417 sqlite3_int64 iPrev = 0;
2418 char *pEnd1 = &a1[n1];
2419 char *pEnd2 = &a2[n2];
2420 char *p1 = a1;
2421 char *p2 = a2;
2422 char *p;
2423 char *aOut;
2424 int bFirstOut = 0;
2426 *paOut = 0;
2427 *pnOut = 0;
2429 /* Allocate space for the output. Both the input and output doclists
2430 ** are delta encoded. If they are in ascending order (bDescDoclist==0),
2431 ** then the first docid in each list is simply encoded as a varint. For
2432 ** each subsequent docid, the varint stored is the difference between the
2433 ** current and previous docid (a positive number - since the list is in
2434 ** ascending order).
2436 ** The first docid written to the output is therefore encoded using the
2437 ** same number of bytes as it is in whichever of the input lists it is
2438 ** read from. And each subsequent docid read from the same input list
2439 ** consumes either the same or less bytes as it did in the input (since
2440 ** the difference between it and the previous value in the output must
2441 ** be a positive value less than or equal to the delta value read from
2442 ** the input list). The same argument applies to all but the first docid
2443 ** read from the 'other' list. And to the contents of all position lists
2444 ** that will be copied and merged from the input to the output.
2446 ** However, if the first docid copied to the output is a negative number,
2447 ** then the encoding of the first docid from the 'other' input list may
2448 ** be larger in the output than it was in the input (since the delta value
2449 ** may be a larger positive integer than the actual docid).
2451 ** The space required to store the output is therefore the sum of the
2452 ** sizes of the two inputs, plus enough space for exactly one of the input
2453 ** docids to grow.
2455 ** A symetric argument may be made if the doclists are in descending
2456 ** order.
2458 aOut = sqlite3_malloc(n1+n2+FTS3_VARINT_MAX-1);
2459 if( !aOut ) return SQLITE_NOMEM;
2461 p = aOut;
2462 fts3GetDeltaVarint3(&p1, pEnd1, 0, &i1);
2463 fts3GetDeltaVarint3(&p2, pEnd2, 0, &i2);
2464 while( p1 || p2 ){
2465 sqlite3_int64 iDiff = DOCID_CMP(i1, i2);
2467 if( p2 && p1 && iDiff==0 ){
2468 fts3PutDeltaVarint3(&p, bDescDoclist, &iPrev, &bFirstOut, i1);
2469 fts3PoslistMerge(&p, &p1, &p2);
2470 fts3GetDeltaVarint3(&p1, pEnd1, bDescDoclist, &i1);
2471 fts3GetDeltaVarint3(&p2, pEnd2, bDescDoclist, &i2);
2472 }else if( !p2 || (p1 && iDiff<0) ){
2473 fts3PutDeltaVarint3(&p, bDescDoclist, &iPrev, &bFirstOut, i1);
2474 fts3PoslistCopy(&p, &p1);
2475 fts3GetDeltaVarint3(&p1, pEnd1, bDescDoclist, &i1);
2476 }else{
2477 fts3PutDeltaVarint3(&p, bDescDoclist, &iPrev, &bFirstOut, i2);
2478 fts3PoslistCopy(&p, &p2);
2479 fts3GetDeltaVarint3(&p2, pEnd2, bDescDoclist, &i2);
2483 *paOut = aOut;
2484 *pnOut = (int)(p-aOut);
2485 assert( *pnOut<=n1+n2+FTS3_VARINT_MAX-1 );
2486 return SQLITE_OK;
2490 ** This function does a "phrase" merge of two doclists. In a phrase merge,
2491 ** the output contains a copy of each position from the right-hand input
2492 ** doclist for which there is a position in the left-hand input doclist
2493 ** exactly nDist tokens before it.
2495 ** If the docids in the input doclists are sorted in ascending order,
2496 ** parameter bDescDoclist should be false. If they are sorted in ascending
2497 ** order, it should be passed a non-zero value.
2499 ** The right-hand input doclist is overwritten by this function.
2501 static void fts3DoclistPhraseMerge(
2502 int bDescDoclist, /* True if arguments are desc */
2503 int nDist, /* Distance from left to right (1=adjacent) */
2504 char *aLeft, int nLeft, /* Left doclist */
2505 char *aRight, int *pnRight /* IN/OUT: Right/output doclist */
2507 sqlite3_int64 i1 = 0;
2508 sqlite3_int64 i2 = 0;
2509 sqlite3_int64 iPrev = 0;
2510 char *pEnd1 = &aLeft[nLeft];
2511 char *pEnd2 = &aRight[*pnRight];
2512 char *p1 = aLeft;
2513 char *p2 = aRight;
2514 char *p;
2515 int bFirstOut = 0;
2516 char *aOut = aRight;
2518 assert( nDist>0 );
2520 p = aOut;
2521 fts3GetDeltaVarint3(&p1, pEnd1, 0, &i1);
2522 fts3GetDeltaVarint3(&p2, pEnd2, 0, &i2);
2524 while( p1 && p2 ){
2525 sqlite3_int64 iDiff = DOCID_CMP(i1, i2);
2526 if( iDiff==0 ){
2527 char *pSave = p;
2528 sqlite3_int64 iPrevSave = iPrev;
2529 int bFirstOutSave = bFirstOut;
2531 fts3PutDeltaVarint3(&p, bDescDoclist, &iPrev, &bFirstOut, i1);
2532 if( 0==fts3PoslistPhraseMerge(&p, nDist, 0, 1, &p1, &p2) ){
2533 p = pSave;
2534 iPrev = iPrevSave;
2535 bFirstOut = bFirstOutSave;
2537 fts3GetDeltaVarint3(&p1, pEnd1, bDescDoclist, &i1);
2538 fts3GetDeltaVarint3(&p2, pEnd2, bDescDoclist, &i2);
2539 }else if( iDiff<0 ){
2540 fts3PoslistCopy(0, &p1);
2541 fts3GetDeltaVarint3(&p1, pEnd1, bDescDoclist, &i1);
2542 }else{
2543 fts3PoslistCopy(0, &p2);
2544 fts3GetDeltaVarint3(&p2, pEnd2, bDescDoclist, &i2);
2548 *pnRight = (int)(p - aOut);
2552 ** Argument pList points to a position list nList bytes in size. This
2553 ** function checks to see if the position list contains any entries for
2554 ** a token in position 0 (of any column). If so, it writes argument iDelta
2555 ** to the output buffer pOut, followed by a position list consisting only
2556 ** of the entries from pList at position 0, and terminated by an 0x00 byte.
2557 ** The value returned is the number of bytes written to pOut (if any).
2559 int sqlite3Fts3FirstFilter(
2560 sqlite3_int64 iDelta, /* Varint that may be written to pOut */
2561 char *pList, /* Position list (no 0x00 term) */
2562 int nList, /* Size of pList in bytes */
2563 char *pOut /* Write output here */
2565 int nOut = 0;
2566 int bWritten = 0; /* True once iDelta has been written */
2567 char *p = pList;
2568 char *pEnd = &pList[nList];
2570 if( *p!=0x01 ){
2571 if( *p==0x02 ){
2572 nOut += sqlite3Fts3PutVarint(&pOut[nOut], iDelta);
2573 pOut[nOut++] = 0x02;
2574 bWritten = 1;
2576 fts3ColumnlistCopy(0, &p);
2579 while( p<pEnd && *p==0x01 ){
2580 sqlite3_int64 iCol;
2581 p++;
2582 p += sqlite3Fts3GetVarint(p, &iCol);
2583 if( *p==0x02 ){
2584 if( bWritten==0 ){
2585 nOut += sqlite3Fts3PutVarint(&pOut[nOut], iDelta);
2586 bWritten = 1;
2588 pOut[nOut++] = 0x01;
2589 nOut += sqlite3Fts3PutVarint(&pOut[nOut], iCol);
2590 pOut[nOut++] = 0x02;
2592 fts3ColumnlistCopy(0, &p);
2594 if( bWritten ){
2595 pOut[nOut++] = 0x00;
2598 return nOut;
2603 ** Merge all doclists in the TermSelect.aaOutput[] array into a single
2604 ** doclist stored in TermSelect.aaOutput[0]. If successful, delete all
2605 ** other doclists (except the aaOutput[0] one) and return SQLITE_OK.
2607 ** If an OOM error occurs, return SQLITE_NOMEM. In this case it is
2608 ** the responsibility of the caller to free any doclists left in the
2609 ** TermSelect.aaOutput[] array.
2611 static int fts3TermSelectFinishMerge(Fts3Table *p, TermSelect *pTS){
2612 char *aOut = 0;
2613 int nOut = 0;
2614 int i;
2616 /* Loop through the doclists in the aaOutput[] array. Merge them all
2617 ** into a single doclist.
2619 for(i=0; i<SizeofArray(pTS->aaOutput); i++){
2620 if( pTS->aaOutput[i] ){
2621 if( !aOut ){
2622 aOut = pTS->aaOutput[i];
2623 nOut = pTS->anOutput[i];
2624 pTS->aaOutput[i] = 0;
2625 }else{
2626 int nNew;
2627 char *aNew;
2629 int rc = fts3DoclistOrMerge(p->bDescIdx,
2630 pTS->aaOutput[i], pTS->anOutput[i], aOut, nOut, &aNew, &nNew
2632 if( rc!=SQLITE_OK ){
2633 sqlite3_free(aOut);
2634 return rc;
2637 sqlite3_free(pTS->aaOutput[i]);
2638 sqlite3_free(aOut);
2639 pTS->aaOutput[i] = 0;
2640 aOut = aNew;
2641 nOut = nNew;
2646 pTS->aaOutput[0] = aOut;
2647 pTS->anOutput[0] = nOut;
2648 return SQLITE_OK;
2652 ** Merge the doclist aDoclist/nDoclist into the TermSelect object passed
2653 ** as the first argument. The merge is an "OR" merge (see function
2654 ** fts3DoclistOrMerge() for details).
2656 ** This function is called with the doclist for each term that matches
2657 ** a queried prefix. It merges all these doclists into one, the doclist
2658 ** for the specified prefix. Since there can be a very large number of
2659 ** doclists to merge, the merging is done pair-wise using the TermSelect
2660 ** object.
2662 ** This function returns SQLITE_OK if the merge is successful, or an
2663 ** SQLite error code (SQLITE_NOMEM) if an error occurs.
2665 static int fts3TermSelectMerge(
2666 Fts3Table *p, /* FTS table handle */
2667 TermSelect *pTS, /* TermSelect object to merge into */
2668 char *aDoclist, /* Pointer to doclist */
2669 int nDoclist /* Size of aDoclist in bytes */
2671 if( pTS->aaOutput[0]==0 ){
2672 /* If this is the first term selected, copy the doclist to the output
2673 ** buffer using memcpy(). */
2674 pTS->aaOutput[0] = sqlite3_malloc(nDoclist);
2675 pTS->anOutput[0] = nDoclist;
2676 if( pTS->aaOutput[0] ){
2677 memcpy(pTS->aaOutput[0], aDoclist, nDoclist);
2678 }else{
2679 return SQLITE_NOMEM;
2681 }else{
2682 char *aMerge = aDoclist;
2683 int nMerge = nDoclist;
2684 int iOut;
2686 for(iOut=0; iOut<SizeofArray(pTS->aaOutput); iOut++){
2687 if( pTS->aaOutput[iOut]==0 ){
2688 assert( iOut>0 );
2689 pTS->aaOutput[iOut] = aMerge;
2690 pTS->anOutput[iOut] = nMerge;
2691 break;
2692 }else{
2693 char *aNew;
2694 int nNew;
2696 int rc = fts3DoclistOrMerge(p->bDescIdx, aMerge, nMerge,
2697 pTS->aaOutput[iOut], pTS->anOutput[iOut], &aNew, &nNew
2699 if( rc!=SQLITE_OK ){
2700 if( aMerge!=aDoclist ) sqlite3_free(aMerge);
2701 return rc;
2704 if( aMerge!=aDoclist ) sqlite3_free(aMerge);
2705 sqlite3_free(pTS->aaOutput[iOut]);
2706 pTS->aaOutput[iOut] = 0;
2708 aMerge = aNew;
2709 nMerge = nNew;
2710 if( (iOut+1)==SizeofArray(pTS->aaOutput) ){
2711 pTS->aaOutput[iOut] = aMerge;
2712 pTS->anOutput[iOut] = nMerge;
2717 return SQLITE_OK;
2721 ** Append SegReader object pNew to the end of the pCsr->apSegment[] array.
2723 static int fts3SegReaderCursorAppend(
2724 Fts3MultiSegReader *pCsr,
2725 Fts3SegReader *pNew
2727 if( (pCsr->nSegment%16)==0 ){
2728 Fts3SegReader **apNew;
2729 int nByte = (pCsr->nSegment + 16)*sizeof(Fts3SegReader*);
2730 apNew = (Fts3SegReader **)sqlite3_realloc(pCsr->apSegment, nByte);
2731 if( !apNew ){
2732 sqlite3Fts3SegReaderFree(pNew);
2733 return SQLITE_NOMEM;
2735 pCsr->apSegment = apNew;
2737 pCsr->apSegment[pCsr->nSegment++] = pNew;
2738 return SQLITE_OK;
2742 ** Add seg-reader objects to the Fts3MultiSegReader object passed as the
2743 ** 8th argument.
2745 ** This function returns SQLITE_OK if successful, or an SQLite error code
2746 ** otherwise.
2748 static int fts3SegReaderCursor(
2749 Fts3Table *p, /* FTS3 table handle */
2750 int iLangid, /* Language id */
2751 int iIndex, /* Index to search (from 0 to p->nIndex-1) */
2752 int iLevel, /* Level of segments to scan */
2753 const char *zTerm, /* Term to query for */
2754 int nTerm, /* Size of zTerm in bytes */
2755 int isPrefix, /* True for a prefix search */
2756 int isScan, /* True to scan from zTerm to EOF */
2757 Fts3MultiSegReader *pCsr /* Cursor object to populate */
2759 int rc = SQLITE_OK; /* Error code */
2760 sqlite3_stmt *pStmt = 0; /* Statement to iterate through segments */
2761 int rc2; /* Result of sqlite3_reset() */
2763 /* If iLevel is less than 0 and this is not a scan, include a seg-reader
2764 ** for the pending-terms. If this is a scan, then this call must be being
2765 ** made by an fts4aux module, not an FTS table. In this case calling
2766 ** Fts3SegReaderPending might segfault, as the data structures used by
2767 ** fts4aux are not completely populated. So it's easiest to filter these
2768 ** calls out here. */
2769 if( iLevel<0 && p->aIndex ){
2770 Fts3SegReader *pSeg = 0;
2771 rc = sqlite3Fts3SegReaderPending(p, iIndex, zTerm, nTerm, isPrefix, &pSeg);
2772 if( rc==SQLITE_OK && pSeg ){
2773 rc = fts3SegReaderCursorAppend(pCsr, pSeg);
2777 if( iLevel!=FTS3_SEGCURSOR_PENDING ){
2778 if( rc==SQLITE_OK ){
2779 rc = sqlite3Fts3AllSegdirs(p, iLangid, iIndex, iLevel, &pStmt);
2782 while( rc==SQLITE_OK && SQLITE_ROW==(rc = sqlite3_step(pStmt)) ){
2783 Fts3SegReader *pSeg = 0;
2785 /* Read the values returned by the SELECT into local variables. */
2786 sqlite3_int64 iStartBlock = sqlite3_column_int64(pStmt, 1);
2787 sqlite3_int64 iLeavesEndBlock = sqlite3_column_int64(pStmt, 2);
2788 sqlite3_int64 iEndBlock = sqlite3_column_int64(pStmt, 3);
2789 int nRoot = sqlite3_column_bytes(pStmt, 4);
2790 char const *zRoot = sqlite3_column_blob(pStmt, 4);
2792 /* If zTerm is not NULL, and this segment is not stored entirely on its
2793 ** root node, the range of leaves scanned can be reduced. Do this. */
2794 if( iStartBlock && zTerm ){
2795 sqlite3_int64 *pi = (isPrefix ? &iLeavesEndBlock : 0);
2796 rc = fts3SelectLeaf(p, zTerm, nTerm, zRoot, nRoot, &iStartBlock, pi);
2797 if( rc!=SQLITE_OK ) goto finished;
2798 if( isPrefix==0 && isScan==0 ) iLeavesEndBlock = iStartBlock;
2801 rc = sqlite3Fts3SegReaderNew(pCsr->nSegment+1,
2802 (isPrefix==0 && isScan==0),
2803 iStartBlock, iLeavesEndBlock,
2804 iEndBlock, zRoot, nRoot, &pSeg
2806 if( rc!=SQLITE_OK ) goto finished;
2807 rc = fts3SegReaderCursorAppend(pCsr, pSeg);
2811 finished:
2812 rc2 = sqlite3_reset(pStmt);
2813 if( rc==SQLITE_DONE ) rc = rc2;
2815 return rc;
2819 ** Set up a cursor object for iterating through a full-text index or a
2820 ** single level therein.
2822 int sqlite3Fts3SegReaderCursor(
2823 Fts3Table *p, /* FTS3 table handle */
2824 int iLangid, /* Language-id to search */
2825 int iIndex, /* Index to search (from 0 to p->nIndex-1) */
2826 int iLevel, /* Level of segments to scan */
2827 const char *zTerm, /* Term to query for */
2828 int nTerm, /* Size of zTerm in bytes */
2829 int isPrefix, /* True for a prefix search */
2830 int isScan, /* True to scan from zTerm to EOF */
2831 Fts3MultiSegReader *pCsr /* Cursor object to populate */
2833 assert( iIndex>=0 && iIndex<p->nIndex );
2834 assert( iLevel==FTS3_SEGCURSOR_ALL
2835 || iLevel==FTS3_SEGCURSOR_PENDING
2836 || iLevel>=0
2838 assert( iLevel<FTS3_SEGDIR_MAXLEVEL );
2839 assert( FTS3_SEGCURSOR_ALL<0 && FTS3_SEGCURSOR_PENDING<0 );
2840 assert( isPrefix==0 || isScan==0 );
2842 memset(pCsr, 0, sizeof(Fts3MultiSegReader));
2843 return fts3SegReaderCursor(
2844 p, iLangid, iIndex, iLevel, zTerm, nTerm, isPrefix, isScan, pCsr
2849 ** In addition to its current configuration, have the Fts3MultiSegReader
2850 ** passed as the 4th argument also scan the doclist for term zTerm/nTerm.
2852 ** SQLITE_OK is returned if no error occurs, otherwise an SQLite error code.
2854 static int fts3SegReaderCursorAddZero(
2855 Fts3Table *p, /* FTS virtual table handle */
2856 int iLangid,
2857 const char *zTerm, /* Term to scan doclist of */
2858 int nTerm, /* Number of bytes in zTerm */
2859 Fts3MultiSegReader *pCsr /* Fts3MultiSegReader to modify */
2861 return fts3SegReaderCursor(p,
2862 iLangid, 0, FTS3_SEGCURSOR_ALL, zTerm, nTerm, 0, 0,pCsr
2867 ** Open an Fts3MultiSegReader to scan the doclist for term zTerm/nTerm. Or,
2868 ** if isPrefix is true, to scan the doclist for all terms for which
2869 ** zTerm/nTerm is a prefix. If successful, return SQLITE_OK and write
2870 ** a pointer to the new Fts3MultiSegReader to *ppSegcsr. Otherwise, return
2871 ** an SQLite error code.
2873 ** It is the responsibility of the caller to free this object by eventually
2874 ** passing it to fts3SegReaderCursorFree()
2876 ** SQLITE_OK is returned if no error occurs, otherwise an SQLite error code.
2877 ** Output parameter *ppSegcsr is set to 0 if an error occurs.
2879 static int fts3TermSegReaderCursor(
2880 Fts3Cursor *pCsr, /* Virtual table cursor handle */
2881 const char *zTerm, /* Term to query for */
2882 int nTerm, /* Size of zTerm in bytes */
2883 int isPrefix, /* True for a prefix search */
2884 Fts3MultiSegReader **ppSegcsr /* OUT: Allocated seg-reader cursor */
2886 Fts3MultiSegReader *pSegcsr; /* Object to allocate and return */
2887 int rc = SQLITE_NOMEM; /* Return code */
2889 pSegcsr = sqlite3_malloc(sizeof(Fts3MultiSegReader));
2890 if( pSegcsr ){
2891 int i;
2892 int bFound = 0; /* True once an index has been found */
2893 Fts3Table *p = (Fts3Table *)pCsr->base.pVtab;
2895 if( isPrefix ){
2896 for(i=1; bFound==0 && i<p->nIndex; i++){
2897 if( p->aIndex[i].nPrefix==nTerm ){
2898 bFound = 1;
2899 rc = sqlite3Fts3SegReaderCursor(p, pCsr->iLangid,
2900 i, FTS3_SEGCURSOR_ALL, zTerm, nTerm, 0, 0, pSegcsr
2902 pSegcsr->bLookup = 1;
2906 for(i=1; bFound==0 && i<p->nIndex; i++){
2907 if( p->aIndex[i].nPrefix==nTerm+1 ){
2908 bFound = 1;
2909 rc = sqlite3Fts3SegReaderCursor(p, pCsr->iLangid,
2910 i, FTS3_SEGCURSOR_ALL, zTerm, nTerm, 1, 0, pSegcsr
2912 if( rc==SQLITE_OK ){
2913 rc = fts3SegReaderCursorAddZero(
2914 p, pCsr->iLangid, zTerm, nTerm, pSegcsr
2921 if( bFound==0 ){
2922 rc = sqlite3Fts3SegReaderCursor(p, pCsr->iLangid,
2923 0, FTS3_SEGCURSOR_ALL, zTerm, nTerm, isPrefix, 0, pSegcsr
2925 pSegcsr->bLookup = !isPrefix;
2929 *ppSegcsr = pSegcsr;
2930 return rc;
2934 ** Free an Fts3MultiSegReader allocated by fts3TermSegReaderCursor().
2936 static void fts3SegReaderCursorFree(Fts3MultiSegReader *pSegcsr){
2937 sqlite3Fts3SegReaderFinish(pSegcsr);
2938 sqlite3_free(pSegcsr);
2942 ** This function retrieves the doclist for the specified term (or term
2943 ** prefix) from the database.
2945 static int fts3TermSelect(
2946 Fts3Table *p, /* Virtual table handle */
2947 Fts3PhraseToken *pTok, /* Token to query for */
2948 int iColumn, /* Column to query (or -ve for all columns) */
2949 int *pnOut, /* OUT: Size of buffer at *ppOut */
2950 char **ppOut /* OUT: Malloced result buffer */
2952 int rc; /* Return code */
2953 Fts3MultiSegReader *pSegcsr; /* Seg-reader cursor for this term */
2954 TermSelect tsc; /* Object for pair-wise doclist merging */
2955 Fts3SegFilter filter; /* Segment term filter configuration */
2957 pSegcsr = pTok->pSegcsr;
2958 memset(&tsc, 0, sizeof(TermSelect));
2960 filter.flags = FTS3_SEGMENT_IGNORE_EMPTY | FTS3_SEGMENT_REQUIRE_POS
2961 | (pTok->isPrefix ? FTS3_SEGMENT_PREFIX : 0)
2962 | (pTok->bFirst ? FTS3_SEGMENT_FIRST : 0)
2963 | (iColumn<p->nColumn ? FTS3_SEGMENT_COLUMN_FILTER : 0);
2964 filter.iCol = iColumn;
2965 filter.zTerm = pTok->z;
2966 filter.nTerm = pTok->n;
2968 rc = sqlite3Fts3SegReaderStart(p, pSegcsr, &filter);
2969 while( SQLITE_OK==rc
2970 && SQLITE_ROW==(rc = sqlite3Fts3SegReaderStep(p, pSegcsr))
2972 rc = fts3TermSelectMerge(p, &tsc, pSegcsr->aDoclist, pSegcsr->nDoclist);
2975 if( rc==SQLITE_OK ){
2976 rc = fts3TermSelectFinishMerge(p, &tsc);
2978 if( rc==SQLITE_OK ){
2979 *ppOut = tsc.aaOutput[0];
2980 *pnOut = tsc.anOutput[0];
2981 }else{
2982 int i;
2983 for(i=0; i<SizeofArray(tsc.aaOutput); i++){
2984 sqlite3_free(tsc.aaOutput[i]);
2988 fts3SegReaderCursorFree(pSegcsr);
2989 pTok->pSegcsr = 0;
2990 return rc;
2994 ** This function counts the total number of docids in the doclist stored
2995 ** in buffer aList[], size nList bytes.
2997 ** If the isPoslist argument is true, then it is assumed that the doclist
2998 ** contains a position-list following each docid. Otherwise, it is assumed
2999 ** that the doclist is simply a list of docids stored as delta encoded
3000 ** varints.
3002 static int fts3DoclistCountDocids(char *aList, int nList){
3003 int nDoc = 0; /* Return value */
3004 if( aList ){
3005 char *aEnd = &aList[nList]; /* Pointer to one byte after EOF */
3006 char *p = aList; /* Cursor */
3007 while( p<aEnd ){
3008 nDoc++;
3009 while( (*p++)&0x80 ); /* Skip docid varint */
3010 fts3PoslistCopy(0, &p); /* Skip over position list */
3014 return nDoc;
3018 ** Advance the cursor to the next row in the %_content table that
3019 ** matches the search criteria. For a MATCH search, this will be
3020 ** the next row that matches. For a full-table scan, this will be
3021 ** simply the next row in the %_content table. For a docid lookup,
3022 ** this routine simply sets the EOF flag.
3024 ** Return SQLITE_OK if nothing goes wrong. SQLITE_OK is returned
3025 ** even if we reach end-of-file. The fts3EofMethod() will be called
3026 ** subsequently to determine whether or not an EOF was hit.
3028 static int fts3NextMethod(sqlite3_vtab_cursor *pCursor){
3029 int rc;
3030 Fts3Cursor *pCsr = (Fts3Cursor *)pCursor;
3031 if( pCsr->eSearch==FTS3_DOCID_SEARCH || pCsr->eSearch==FTS3_FULLSCAN_SEARCH ){
3032 if( SQLITE_ROW!=sqlite3_step(pCsr->pStmt) ){
3033 pCsr->isEof = 1;
3034 rc = sqlite3_reset(pCsr->pStmt);
3035 }else{
3036 pCsr->iPrevId = sqlite3_column_int64(pCsr->pStmt, 0);
3037 rc = SQLITE_OK;
3039 }else{
3040 rc = fts3EvalNext((Fts3Cursor *)pCursor);
3042 assert( ((Fts3Table *)pCsr->base.pVtab)->pSegments==0 );
3043 return rc;
3047 ** The following are copied from sqliteInt.h.
3049 ** Constants for the largest and smallest possible 64-bit signed integers.
3050 ** These macros are designed to work correctly on both 32-bit and 64-bit
3051 ** compilers.
3053 #ifndef SQLITE_AMALGAMATION
3054 # define LARGEST_INT64 (0xffffffff|(((sqlite3_int64)0x7fffffff)<<32))
3055 # define SMALLEST_INT64 (((sqlite3_int64)-1) - LARGEST_INT64)
3056 #endif
3059 ** If the numeric type of argument pVal is "integer", then return it
3060 ** converted to a 64-bit signed integer. Otherwise, return a copy of
3061 ** the second parameter, iDefault.
3063 static sqlite3_int64 fts3DocidRange(sqlite3_value *pVal, i64 iDefault){
3064 if( pVal ){
3065 int eType = sqlite3_value_numeric_type(pVal);
3066 if( eType==SQLITE_INTEGER ){
3067 return sqlite3_value_int64(pVal);
3070 return iDefault;
3074 ** This is the xFilter interface for the virtual table. See
3075 ** the virtual table xFilter method documentation for additional
3076 ** information.
3078 ** If idxNum==FTS3_FULLSCAN_SEARCH then do a full table scan against
3079 ** the %_content table.
3081 ** If idxNum==FTS3_DOCID_SEARCH then do a docid lookup for a single entry
3082 ** in the %_content table.
3084 ** If idxNum>=FTS3_FULLTEXT_SEARCH then use the full text index. The
3085 ** column on the left-hand side of the MATCH operator is column
3086 ** number idxNum-FTS3_FULLTEXT_SEARCH, 0 indexed. argv[0] is the right-hand
3087 ** side of the MATCH operator.
3089 static int fts3FilterMethod(
3090 sqlite3_vtab_cursor *pCursor, /* The cursor used for this query */
3091 int idxNum, /* Strategy index */
3092 const char *idxStr, /* Unused */
3093 int nVal, /* Number of elements in apVal */
3094 sqlite3_value **apVal /* Arguments for the indexing scheme */
3096 int rc;
3097 char *zSql; /* SQL statement used to access %_content */
3098 int eSearch;
3099 Fts3Table *p = (Fts3Table *)pCursor->pVtab;
3100 Fts3Cursor *pCsr = (Fts3Cursor *)pCursor;
3102 sqlite3_value *pCons = 0; /* The MATCH or rowid constraint, if any */
3103 sqlite3_value *pLangid = 0; /* The "langid = ?" constraint, if any */
3104 sqlite3_value *pDocidGe = 0; /* The "docid >= ?" constraint, if any */
3105 sqlite3_value *pDocidLe = 0; /* The "docid <= ?" constraint, if any */
3106 int iIdx;
3108 UNUSED_PARAMETER(idxStr);
3109 UNUSED_PARAMETER(nVal);
3111 eSearch = (idxNum & 0x0000FFFF);
3112 assert( eSearch>=0 && eSearch<=(FTS3_FULLTEXT_SEARCH+p->nColumn) );
3113 assert( p->pSegments==0 );
3115 /* Collect arguments into local variables */
3116 iIdx = 0;
3117 if( eSearch!=FTS3_FULLSCAN_SEARCH ) pCons = apVal[iIdx++];
3118 if( idxNum & FTS3_HAVE_LANGID ) pLangid = apVal[iIdx++];
3119 if( idxNum & FTS3_HAVE_DOCID_GE ) pDocidGe = apVal[iIdx++];
3120 if( idxNum & FTS3_HAVE_DOCID_LE ) pDocidLe = apVal[iIdx++];
3121 assert( iIdx==nVal );
3123 /* In case the cursor has been used before, clear it now. */
3124 sqlite3_finalize(pCsr->pStmt);
3125 sqlite3_free(pCsr->aDoclist);
3126 sqlite3_free(pCsr->aMatchinfo);
3127 sqlite3Fts3ExprFree(pCsr->pExpr);
3128 memset(&pCursor[1], 0, sizeof(Fts3Cursor)-sizeof(sqlite3_vtab_cursor));
3130 /* Set the lower and upper bounds on docids to return */
3131 pCsr->iMinDocid = fts3DocidRange(pDocidGe, SMALLEST_INT64);
3132 pCsr->iMaxDocid = fts3DocidRange(pDocidLe, LARGEST_INT64);
3134 if( idxStr ){
3135 pCsr->bDesc = (idxStr[0]=='D');
3136 }else{
3137 pCsr->bDesc = p->bDescIdx;
3139 pCsr->eSearch = (i16)eSearch;
3141 if( eSearch!=FTS3_DOCID_SEARCH && eSearch!=FTS3_FULLSCAN_SEARCH ){
3142 int iCol = eSearch-FTS3_FULLTEXT_SEARCH;
3143 const char *zQuery = (const char *)sqlite3_value_text(pCons);
3145 if( zQuery==0 && sqlite3_value_type(pCons)!=SQLITE_NULL ){
3146 return SQLITE_NOMEM;
3149 pCsr->iLangid = 0;
3150 if( pLangid ) pCsr->iLangid = sqlite3_value_int(pLangid);
3152 assert( p->base.zErrMsg==0 );
3153 rc = sqlite3Fts3ExprParse(p->pTokenizer, pCsr->iLangid,
3154 p->azColumn, p->bFts4, p->nColumn, iCol, zQuery, -1, &pCsr->pExpr,
3155 &p->base.zErrMsg
3157 if( rc!=SQLITE_OK ){
3158 return rc;
3161 rc = fts3EvalStart(pCsr);
3162 sqlite3Fts3SegmentsClose(p);
3163 if( rc!=SQLITE_OK ) return rc;
3164 pCsr->pNextId = pCsr->aDoclist;
3165 pCsr->iPrevId = 0;
3168 /* Compile a SELECT statement for this cursor. For a full-table-scan, the
3169 ** statement loops through all rows of the %_content table. For a
3170 ** full-text query or docid lookup, the statement retrieves a single
3171 ** row by docid.
3173 if( eSearch==FTS3_FULLSCAN_SEARCH ){
3174 zSql = sqlite3_mprintf(
3175 "SELECT %s ORDER BY rowid %s",
3176 p->zReadExprlist, (pCsr->bDesc ? "DESC" : "ASC")
3178 if( zSql ){
3179 rc = sqlite3_prepare_v2(p->db, zSql, -1, &pCsr->pStmt, 0);
3180 sqlite3_free(zSql);
3181 }else{
3182 rc = SQLITE_NOMEM;
3184 }else if( eSearch==FTS3_DOCID_SEARCH ){
3185 rc = fts3CursorSeekStmt(pCsr, &pCsr->pStmt);
3186 if( rc==SQLITE_OK ){
3187 rc = sqlite3_bind_value(pCsr->pStmt, 1, pCons);
3190 if( rc!=SQLITE_OK ) return rc;
3192 return fts3NextMethod(pCursor);
3196 ** This is the xEof method of the virtual table. SQLite calls this
3197 ** routine to find out if it has reached the end of a result set.
3199 static int fts3EofMethod(sqlite3_vtab_cursor *pCursor){
3200 return ((Fts3Cursor *)pCursor)->isEof;
3204 ** This is the xRowid method. The SQLite core calls this routine to
3205 ** retrieve the rowid for the current row of the result set. fts3
3206 ** exposes %_content.docid as the rowid for the virtual table. The
3207 ** rowid should be written to *pRowid.
3209 static int fts3RowidMethod(sqlite3_vtab_cursor *pCursor, sqlite_int64 *pRowid){
3210 Fts3Cursor *pCsr = (Fts3Cursor *) pCursor;
3211 *pRowid = pCsr->iPrevId;
3212 return SQLITE_OK;
3216 ** This is the xColumn method, called by SQLite to request a value from
3217 ** the row that the supplied cursor currently points to.
3219 ** If:
3221 ** (iCol < p->nColumn) -> The value of the iCol'th user column.
3222 ** (iCol == p->nColumn) -> Magic column with the same name as the table.
3223 ** (iCol == p->nColumn+1) -> Docid column
3224 ** (iCol == p->nColumn+2) -> Langid column
3226 static int fts3ColumnMethod(
3227 sqlite3_vtab_cursor *pCursor, /* Cursor to retrieve value from */
3228 sqlite3_context *pCtx, /* Context for sqlite3_result_xxx() calls */
3229 int iCol /* Index of column to read value from */
3231 int rc = SQLITE_OK; /* Return Code */
3232 Fts3Cursor *pCsr = (Fts3Cursor *) pCursor;
3233 Fts3Table *p = (Fts3Table *)pCursor->pVtab;
3235 /* The column value supplied by SQLite must be in range. */
3236 assert( iCol>=0 && iCol<=p->nColumn+2 );
3238 if( iCol==p->nColumn+1 ){
3239 /* This call is a request for the "docid" column. Since "docid" is an
3240 ** alias for "rowid", use the xRowid() method to obtain the value.
3242 sqlite3_result_int64(pCtx, pCsr->iPrevId);
3243 }else if( iCol==p->nColumn ){
3244 /* The extra column whose name is the same as the table.
3245 ** Return a blob which is a pointer to the cursor. */
3246 sqlite3_result_blob(pCtx, &pCsr, sizeof(pCsr), SQLITE_TRANSIENT);
3247 }else if( iCol==p->nColumn+2 && pCsr->pExpr ){
3248 sqlite3_result_int64(pCtx, pCsr->iLangid);
3249 }else{
3250 /* The requested column is either a user column (one that contains
3251 ** indexed data), or the language-id column. */
3252 rc = fts3CursorSeek(0, pCsr);
3254 if( rc==SQLITE_OK ){
3255 if( iCol==p->nColumn+2 ){
3256 int iLangid = 0;
3257 if( p->zLanguageid ){
3258 iLangid = sqlite3_column_int(pCsr->pStmt, p->nColumn+1);
3260 sqlite3_result_int(pCtx, iLangid);
3261 }else if( sqlite3_data_count(pCsr->pStmt)>(iCol+1) ){
3262 sqlite3_result_value(pCtx, sqlite3_column_value(pCsr->pStmt, iCol+1));
3267 assert( ((Fts3Table *)pCsr->base.pVtab)->pSegments==0 );
3268 return rc;
3272 ** This function is the implementation of the xUpdate callback used by
3273 ** FTS3 virtual tables. It is invoked by SQLite each time a row is to be
3274 ** inserted, updated or deleted.
3276 static int fts3UpdateMethod(
3277 sqlite3_vtab *pVtab, /* Virtual table handle */
3278 int nArg, /* Size of argument array */
3279 sqlite3_value **apVal, /* Array of arguments */
3280 sqlite_int64 *pRowid /* OUT: The affected (or effected) rowid */
3282 return sqlite3Fts3UpdateMethod(pVtab, nArg, apVal, pRowid);
3286 ** Implementation of xSync() method. Flush the contents of the pending-terms
3287 ** hash-table to the database.
3289 static int fts3SyncMethod(sqlite3_vtab *pVtab){
3291 /* Following an incremental-merge operation, assuming that the input
3292 ** segments are not completely consumed (the usual case), they are updated
3293 ** in place to remove the entries that have already been merged. This
3294 ** involves updating the leaf block that contains the smallest unmerged
3295 ** entry and each block (if any) between the leaf and the root node. So
3296 ** if the height of the input segment b-trees is N, and input segments
3297 ** are merged eight at a time, updating the input segments at the end
3298 ** of an incremental-merge requires writing (8*(1+N)) blocks. N is usually
3299 ** small - often between 0 and 2. So the overhead of the incremental
3300 ** merge is somewhere between 8 and 24 blocks. To avoid this overhead
3301 ** dwarfing the actual productive work accomplished, the incremental merge
3302 ** is only attempted if it will write at least 64 leaf blocks. Hence
3303 ** nMinMerge.
3305 ** Of course, updating the input segments also involves deleting a bunch
3306 ** of blocks from the segments table. But this is not considered overhead
3307 ** as it would also be required by a crisis-merge that used the same input
3308 ** segments.
3310 const u32 nMinMerge = 64; /* Minimum amount of incr-merge work to do */
3312 Fts3Table *p = (Fts3Table*)pVtab;
3313 int rc = sqlite3Fts3PendingTermsFlush(p);
3315 if( rc==SQLITE_OK
3316 && p->nLeafAdd>(nMinMerge/16)
3317 && p->nAutoincrmerge && p->nAutoincrmerge!=0xff
3319 int mxLevel = 0; /* Maximum relative level value in db */
3320 int A; /* Incr-merge parameter A */
3322 rc = sqlite3Fts3MaxLevel(p, &mxLevel);
3323 assert( rc==SQLITE_OK || mxLevel==0 );
3324 A = p->nLeafAdd * mxLevel;
3325 A += (A/2);
3326 if( A>(int)nMinMerge ) rc = sqlite3Fts3Incrmerge(p, A, p->nAutoincrmerge);
3328 sqlite3Fts3SegmentsClose(p);
3329 return rc;
3333 ** If it is currently unknown whether or not the FTS table has an %_stat
3334 ** table (if p->bHasStat==2), attempt to determine this (set p->bHasStat
3335 ** to 0 or 1). Return SQLITE_OK if successful, or an SQLite error code
3336 ** if an error occurs.
3338 static int fts3SetHasStat(Fts3Table *p){
3339 int rc = SQLITE_OK;
3340 if( p->bHasStat==2 ){
3341 const char *zFmt ="SELECT 1 FROM %Q.sqlite_master WHERE tbl_name='%q_stat'";
3342 char *zSql = sqlite3_mprintf(zFmt, p->zDb, p->zName);
3343 if( zSql ){
3344 sqlite3_stmt *pStmt = 0;
3345 rc = sqlite3_prepare_v2(p->db, zSql, -1, &pStmt, 0);
3346 if( rc==SQLITE_OK ){
3347 int bHasStat = (sqlite3_step(pStmt)==SQLITE_ROW);
3348 rc = sqlite3_finalize(pStmt);
3349 if( rc==SQLITE_OK ) p->bHasStat = bHasStat;
3351 sqlite3_free(zSql);
3352 }else{
3353 rc = SQLITE_NOMEM;
3356 return rc;
3360 ** Implementation of xBegin() method.
3362 static int fts3BeginMethod(sqlite3_vtab *pVtab){
3363 Fts3Table *p = (Fts3Table*)pVtab;
3364 UNUSED_PARAMETER(pVtab);
3365 assert( p->pSegments==0 );
3366 assert( p->nPendingData==0 );
3367 assert( p->inTransaction!=1 );
3368 TESTONLY( p->inTransaction = 1 );
3369 TESTONLY( p->mxSavepoint = -1; );
3370 p->nLeafAdd = 0;
3371 return fts3SetHasStat(p);
3375 ** Implementation of xCommit() method. This is a no-op. The contents of
3376 ** the pending-terms hash-table have already been flushed into the database
3377 ** by fts3SyncMethod().
3379 static int fts3CommitMethod(sqlite3_vtab *pVtab){
3380 TESTONLY( Fts3Table *p = (Fts3Table*)pVtab );
3381 UNUSED_PARAMETER(pVtab);
3382 assert( p->nPendingData==0 );
3383 assert( p->inTransaction!=0 );
3384 assert( p->pSegments==0 );
3385 TESTONLY( p->inTransaction = 0 );
3386 TESTONLY( p->mxSavepoint = -1; );
3387 return SQLITE_OK;
3391 ** Implementation of xRollback(). Discard the contents of the pending-terms
3392 ** hash-table. Any changes made to the database are reverted by SQLite.
3394 static int fts3RollbackMethod(sqlite3_vtab *pVtab){
3395 Fts3Table *p = (Fts3Table*)pVtab;
3396 sqlite3Fts3PendingTermsClear(p);
3397 assert( p->inTransaction!=0 );
3398 TESTONLY( p->inTransaction = 0 );
3399 TESTONLY( p->mxSavepoint = -1; );
3400 return SQLITE_OK;
3404 ** When called, *ppPoslist must point to the byte immediately following the
3405 ** end of a position-list. i.e. ( (*ppPoslist)[-1]==POS_END ). This function
3406 ** moves *ppPoslist so that it instead points to the first byte of the
3407 ** same position list.
3409 static void fts3ReversePoslist(char *pStart, char **ppPoslist){
3410 char *p = &(*ppPoslist)[-2];
3411 char c = 0;
3413 while( p>pStart && (c=*p--)==0 );
3414 while( p>pStart && (*p & 0x80) | c ){
3415 c = *p--;
3417 if( p>pStart ){ p = &p[2]; }
3418 while( *p++&0x80 );
3419 *ppPoslist = p;
3423 ** Helper function used by the implementation of the overloaded snippet(),
3424 ** offsets() and optimize() SQL functions.
3426 ** If the value passed as the third argument is a blob of size
3427 ** sizeof(Fts3Cursor*), then the blob contents are copied to the
3428 ** output variable *ppCsr and SQLITE_OK is returned. Otherwise, an error
3429 ** message is written to context pContext and SQLITE_ERROR returned. The
3430 ** string passed via zFunc is used as part of the error message.
3432 static int fts3FunctionArg(
3433 sqlite3_context *pContext, /* SQL function call context */
3434 const char *zFunc, /* Function name */
3435 sqlite3_value *pVal, /* argv[0] passed to function */
3436 Fts3Cursor **ppCsr /* OUT: Store cursor handle here */
3438 Fts3Cursor *pRet;
3439 if( sqlite3_value_type(pVal)!=SQLITE_BLOB
3440 || sqlite3_value_bytes(pVal)!=sizeof(Fts3Cursor *)
3442 char *zErr = sqlite3_mprintf("illegal first argument to %s", zFunc);
3443 sqlite3_result_error(pContext, zErr, -1);
3444 sqlite3_free(zErr);
3445 return SQLITE_ERROR;
3447 memcpy(&pRet, sqlite3_value_blob(pVal), sizeof(Fts3Cursor *));
3448 *ppCsr = pRet;
3449 return SQLITE_OK;
3453 ** Implementation of the snippet() function for FTS3
3455 static void fts3SnippetFunc(
3456 sqlite3_context *pContext, /* SQLite function call context */
3457 int nVal, /* Size of apVal[] array */
3458 sqlite3_value **apVal /* Array of arguments */
3460 Fts3Cursor *pCsr; /* Cursor handle passed through apVal[0] */
3461 const char *zStart = "<b>";
3462 const char *zEnd = "</b>";
3463 const char *zEllipsis = "<b>...</b>";
3464 int iCol = -1;
3465 int nToken = 15; /* Default number of tokens in snippet */
3467 /* There must be at least one argument passed to this function (otherwise
3468 ** the non-overloaded version would have been called instead of this one).
3470 assert( nVal>=1 );
3472 if( nVal>6 ){
3473 sqlite3_result_error(pContext,
3474 "wrong number of arguments to function snippet()", -1);
3475 return;
3477 if( fts3FunctionArg(pContext, "snippet", apVal[0], &pCsr) ) return;
3479 switch( nVal ){
3480 case 6: nToken = sqlite3_value_int(apVal[5]);
3481 case 5: iCol = sqlite3_value_int(apVal[4]);
3482 case 4: zEllipsis = (const char*)sqlite3_value_text(apVal[3]);
3483 case 3: zEnd = (const char*)sqlite3_value_text(apVal[2]);
3484 case 2: zStart = (const char*)sqlite3_value_text(apVal[1]);
3486 if( !zEllipsis || !zEnd || !zStart ){
3487 sqlite3_result_error_nomem(pContext);
3488 }else if( SQLITE_OK==fts3CursorSeek(pContext, pCsr) ){
3489 sqlite3Fts3Snippet(pContext, pCsr, zStart, zEnd, zEllipsis, iCol, nToken);
3494 ** Implementation of the offsets() function for FTS3
3496 static void fts3OffsetsFunc(
3497 sqlite3_context *pContext, /* SQLite function call context */
3498 int nVal, /* Size of argument array */
3499 sqlite3_value **apVal /* Array of arguments */
3501 Fts3Cursor *pCsr; /* Cursor handle passed through apVal[0] */
3503 UNUSED_PARAMETER(nVal);
3505 assert( nVal==1 );
3506 if( fts3FunctionArg(pContext, "offsets", apVal[0], &pCsr) ) return;
3507 assert( pCsr );
3508 if( SQLITE_OK==fts3CursorSeek(pContext, pCsr) ){
3509 sqlite3Fts3Offsets(pContext, pCsr);
3514 ** Implementation of the special optimize() function for FTS3. This
3515 ** function merges all segments in the database to a single segment.
3516 ** Example usage is:
3518 ** SELECT optimize(t) FROM t LIMIT 1;
3520 ** where 't' is the name of an FTS3 table.
3522 static void fts3OptimizeFunc(
3523 sqlite3_context *pContext, /* SQLite function call context */
3524 int nVal, /* Size of argument array */
3525 sqlite3_value **apVal /* Array of arguments */
3527 int rc; /* Return code */
3528 Fts3Table *p; /* Virtual table handle */
3529 Fts3Cursor *pCursor; /* Cursor handle passed through apVal[0] */
3531 UNUSED_PARAMETER(nVal);
3533 assert( nVal==1 );
3534 if( fts3FunctionArg(pContext, "optimize", apVal[0], &pCursor) ) return;
3535 p = (Fts3Table *)pCursor->base.pVtab;
3536 assert( p );
3538 rc = sqlite3Fts3Optimize(p);
3540 switch( rc ){
3541 case SQLITE_OK:
3542 sqlite3_result_text(pContext, "Index optimized", -1, SQLITE_STATIC);
3543 break;
3544 case SQLITE_DONE:
3545 sqlite3_result_text(pContext, "Index already optimal", -1, SQLITE_STATIC);
3546 break;
3547 default:
3548 sqlite3_result_error_code(pContext, rc);
3549 break;
3554 ** Implementation of the matchinfo() function for FTS3
3556 static void fts3MatchinfoFunc(
3557 sqlite3_context *pContext, /* SQLite function call context */
3558 int nVal, /* Size of argument array */
3559 sqlite3_value **apVal /* Array of arguments */
3561 Fts3Cursor *pCsr; /* Cursor handle passed through apVal[0] */
3562 assert( nVal==1 || nVal==2 );
3563 if( SQLITE_OK==fts3FunctionArg(pContext, "matchinfo", apVal[0], &pCsr) ){
3564 const char *zArg = 0;
3565 if( nVal>1 ){
3566 zArg = (const char *)sqlite3_value_text(apVal[1]);
3568 sqlite3Fts3Matchinfo(pContext, pCsr, zArg);
3573 ** This routine implements the xFindFunction method for the FTS3
3574 ** virtual table.
3576 static int fts3FindFunctionMethod(
3577 sqlite3_vtab *pVtab, /* Virtual table handle */
3578 int nArg, /* Number of SQL function arguments */
3579 const char *zName, /* Name of SQL function */
3580 void (**pxFunc)(sqlite3_context*,int,sqlite3_value**), /* OUT: Result */
3581 void **ppArg /* Unused */
3583 struct Overloaded {
3584 const char *zName;
3585 void (*xFunc)(sqlite3_context*,int,sqlite3_value**);
3586 } aOverload[] = {
3587 { "snippet", fts3SnippetFunc },
3588 { "offsets", fts3OffsetsFunc },
3589 { "optimize", fts3OptimizeFunc },
3590 { "matchinfo", fts3MatchinfoFunc },
3592 int i; /* Iterator variable */
3594 UNUSED_PARAMETER(pVtab);
3595 UNUSED_PARAMETER(nArg);
3596 UNUSED_PARAMETER(ppArg);
3598 for(i=0; i<SizeofArray(aOverload); i++){
3599 if( strcmp(zName, aOverload[i].zName)==0 ){
3600 *pxFunc = aOverload[i].xFunc;
3601 return 1;
3605 /* No function of the specified name was found. Return 0. */
3606 return 0;
3610 ** Implementation of FTS3 xRename method. Rename an fts3 table.
3612 static int fts3RenameMethod(
3613 sqlite3_vtab *pVtab, /* Virtual table handle */
3614 const char *zName /* New name of table */
3616 Fts3Table *p = (Fts3Table *)pVtab;
3617 sqlite3 *db = p->db; /* Database connection */
3618 int rc; /* Return Code */
3620 /* At this point it must be known if the %_stat table exists or not.
3621 ** So bHasStat may not be 2. */
3622 rc = fts3SetHasStat(p);
3624 /* As it happens, the pending terms table is always empty here. This is
3625 ** because an "ALTER TABLE RENAME TABLE" statement inside a transaction
3626 ** always opens a savepoint transaction. And the xSavepoint() method
3627 ** flushes the pending terms table. But leave the (no-op) call to
3628 ** PendingTermsFlush() in in case that changes.
3630 assert( p->nPendingData==0 );
3631 if( rc==SQLITE_OK ){
3632 rc = sqlite3Fts3PendingTermsFlush(p);
3635 if( p->zContentTbl==0 ){
3636 fts3DbExec(&rc, db,
3637 "ALTER TABLE %Q.'%q_content' RENAME TO '%q_content';",
3638 p->zDb, p->zName, zName
3642 if( p->bHasDocsize ){
3643 fts3DbExec(&rc, db,
3644 "ALTER TABLE %Q.'%q_docsize' RENAME TO '%q_docsize';",
3645 p->zDb, p->zName, zName
3648 if( p->bHasStat ){
3649 fts3DbExec(&rc, db,
3650 "ALTER TABLE %Q.'%q_stat' RENAME TO '%q_stat';",
3651 p->zDb, p->zName, zName
3654 fts3DbExec(&rc, db,
3655 "ALTER TABLE %Q.'%q_segments' RENAME TO '%q_segments';",
3656 p->zDb, p->zName, zName
3658 fts3DbExec(&rc, db,
3659 "ALTER TABLE %Q.'%q_segdir' RENAME TO '%q_segdir';",
3660 p->zDb, p->zName, zName
3662 return rc;
3666 ** The xSavepoint() method.
3668 ** Flush the contents of the pending-terms table to disk.
3670 static int fts3SavepointMethod(sqlite3_vtab *pVtab, int iSavepoint){
3671 int rc = SQLITE_OK;
3672 UNUSED_PARAMETER(iSavepoint);
3673 assert( ((Fts3Table *)pVtab)->inTransaction );
3674 assert( ((Fts3Table *)pVtab)->mxSavepoint < iSavepoint );
3675 TESTONLY( ((Fts3Table *)pVtab)->mxSavepoint = iSavepoint );
3676 if( ((Fts3Table *)pVtab)->bIgnoreSavepoint==0 ){
3677 rc = fts3SyncMethod(pVtab);
3679 return rc;
3683 ** The xRelease() method.
3685 ** This is a no-op.
3687 static int fts3ReleaseMethod(sqlite3_vtab *pVtab, int iSavepoint){
3688 TESTONLY( Fts3Table *p = (Fts3Table*)pVtab );
3689 UNUSED_PARAMETER(iSavepoint);
3690 UNUSED_PARAMETER(pVtab);
3691 assert( p->inTransaction );
3692 assert( p->mxSavepoint >= iSavepoint );
3693 TESTONLY( p->mxSavepoint = iSavepoint-1 );
3694 return SQLITE_OK;
3698 ** The xRollbackTo() method.
3700 ** Discard the contents of the pending terms table.
3702 static int fts3RollbackToMethod(sqlite3_vtab *pVtab, int iSavepoint){
3703 Fts3Table *p = (Fts3Table*)pVtab;
3704 UNUSED_PARAMETER(iSavepoint);
3705 assert( p->inTransaction );
3706 assert( p->mxSavepoint >= iSavepoint );
3707 TESTONLY( p->mxSavepoint = iSavepoint );
3708 sqlite3Fts3PendingTermsClear(p);
3709 return SQLITE_OK;
3712 static const sqlite3_module fts3Module = {
3713 /* iVersion */ 2,
3714 /* xCreate */ fts3CreateMethod,
3715 /* xConnect */ fts3ConnectMethod,
3716 /* xBestIndex */ fts3BestIndexMethod,
3717 /* xDisconnect */ fts3DisconnectMethod,
3718 /* xDestroy */ fts3DestroyMethod,
3719 /* xOpen */ fts3OpenMethod,
3720 /* xClose */ fts3CloseMethod,
3721 /* xFilter */ fts3FilterMethod,
3722 /* xNext */ fts3NextMethod,
3723 /* xEof */ fts3EofMethod,
3724 /* xColumn */ fts3ColumnMethod,
3725 /* xRowid */ fts3RowidMethod,
3726 /* xUpdate */ fts3UpdateMethod,
3727 /* xBegin */ fts3BeginMethod,
3728 /* xSync */ fts3SyncMethod,
3729 /* xCommit */ fts3CommitMethod,
3730 /* xRollback */ fts3RollbackMethod,
3731 /* xFindFunction */ fts3FindFunctionMethod,
3732 /* xRename */ fts3RenameMethod,
3733 /* xSavepoint */ fts3SavepointMethod,
3734 /* xRelease */ fts3ReleaseMethod,
3735 /* xRollbackTo */ fts3RollbackToMethod,
3739 ** This function is registered as the module destructor (called when an
3740 ** FTS3 enabled database connection is closed). It frees the memory
3741 ** allocated for the tokenizer hash table.
3743 static void hashDestroy(void *p){
3744 Fts3Hash *pHash = (Fts3Hash *)p;
3745 sqlite3Fts3HashClear(pHash);
3746 sqlite3_free(pHash);
3750 ** The fts3 built-in tokenizers - "simple", "porter" and "icu"- are
3751 ** implemented in files fts3_tokenizer1.c, fts3_porter.c and fts3_icu.c
3752 ** respectively. The following three forward declarations are for functions
3753 ** declared in these files used to retrieve the respective implementations.
3755 ** Calling sqlite3Fts3SimpleTokenizerModule() sets the value pointed
3756 ** to by the argument to point to the "simple" tokenizer implementation.
3757 ** And so on.
3759 void sqlite3Fts3SimpleTokenizerModule(sqlite3_tokenizer_module const**ppModule);
3760 void sqlite3Fts3PorterTokenizerModule(sqlite3_tokenizer_module const**ppModule);
3761 #ifndef SQLITE_DISABLE_FTS3_UNICODE
3762 void sqlite3Fts3UnicodeTokenizer(sqlite3_tokenizer_module const**ppModule);
3763 #endif
3764 #ifdef SQLITE_ENABLE_ICU
3765 void sqlite3Fts3IcuTokenizerModule(sqlite3_tokenizer_module const**ppModule);
3766 #endif
3769 ** Initialize the fts3 extension. If this extension is built as part
3770 ** of the sqlite library, then this function is called directly by
3771 ** SQLite. If fts3 is built as a dynamically loadable extension, this
3772 ** function is called by the sqlite3_extension_init() entry point.
3774 int sqlite3Fts3Init(sqlite3 *db){
3775 int rc = SQLITE_OK;
3776 Fts3Hash *pHash = 0;
3777 const sqlite3_tokenizer_module *pSimple = 0;
3778 const sqlite3_tokenizer_module *pPorter = 0;
3779 #ifndef SQLITE_DISABLE_FTS3_UNICODE
3780 const sqlite3_tokenizer_module *pUnicode = 0;
3781 #endif
3783 #ifdef SQLITE_ENABLE_ICU
3784 const sqlite3_tokenizer_module *pIcu = 0;
3785 sqlite3Fts3IcuTokenizerModule(&pIcu);
3786 #endif
3788 #ifndef SQLITE_DISABLE_FTS3_UNICODE
3789 sqlite3Fts3UnicodeTokenizer(&pUnicode);
3790 #endif
3792 #ifdef SQLITE_TEST
3793 rc = sqlite3Fts3InitTerm(db);
3794 if( rc!=SQLITE_OK ) return rc;
3795 #endif
3797 rc = sqlite3Fts3InitAux(db);
3798 if( rc!=SQLITE_OK ) return rc;
3800 sqlite3Fts3SimpleTokenizerModule(&pSimple);
3801 sqlite3Fts3PorterTokenizerModule(&pPorter);
3803 /* Allocate and initialize the hash-table used to store tokenizers. */
3804 pHash = sqlite3_malloc(sizeof(Fts3Hash));
3805 if( !pHash ){
3806 rc = SQLITE_NOMEM;
3807 }else{
3808 sqlite3Fts3HashInit(pHash, FTS3_HASH_STRING, 1);
3811 /* Load the built-in tokenizers into the hash table */
3812 if( rc==SQLITE_OK ){
3813 if( sqlite3Fts3HashInsert(pHash, "simple", 7, (void *)pSimple)
3814 || sqlite3Fts3HashInsert(pHash, "porter", 7, (void *)pPorter)
3816 #ifndef SQLITE_DISABLE_FTS3_UNICODE
3817 || sqlite3Fts3HashInsert(pHash, "unicode61", 10, (void *)pUnicode)
3818 #endif
3819 #ifdef SQLITE_ENABLE_ICU
3820 || (pIcu && sqlite3Fts3HashInsert(pHash, "icu", 4, (void *)pIcu))
3821 #endif
3823 rc = SQLITE_NOMEM;
3827 #ifdef SQLITE_TEST
3828 if( rc==SQLITE_OK ){
3829 rc = sqlite3Fts3ExprInitTestInterface(db);
3831 #endif
3833 /* Create the virtual table wrapper around the hash-table and overload
3834 ** the two scalar functions. If this is successful, register the
3835 ** module with sqlite.
3837 if( SQLITE_OK==rc
3838 #if CHROMIUM_FTS3_CHANGES && !SQLITE_TEST
3839 /* fts3_tokenizer() disabled for security reasons. */
3840 #else
3841 && SQLITE_OK==(rc = sqlite3Fts3InitHashTable(db, pHash, "fts3_tokenizer"))
3842 #endif
3843 && SQLITE_OK==(rc = sqlite3_overload_function(db, "snippet", -1))
3844 && SQLITE_OK==(rc = sqlite3_overload_function(db, "offsets", 1))
3845 && SQLITE_OK==(rc = sqlite3_overload_function(db, "matchinfo", 1))
3846 && SQLITE_OK==(rc = sqlite3_overload_function(db, "matchinfo", 2))
3847 && SQLITE_OK==(rc = sqlite3_overload_function(db, "optimize", 1))
3849 rc = sqlite3_create_module_v2(
3850 db, "fts3", &fts3Module, (void *)pHash, hashDestroy
3852 #if CHROMIUM_FTS3_CHANGES && !SQLITE_TEST
3853 /* Disable fts4 and tokenizer vtab pending review. */
3854 #else
3855 if( rc==SQLITE_OK ){
3856 rc = sqlite3_create_module_v2(
3857 db, "fts4", &fts3Module, (void *)pHash, 0
3860 if( rc==SQLITE_OK ){
3861 rc = sqlite3Fts3InitTok(db, (void *)pHash);
3863 #endif
3864 return rc;
3868 /* An error has occurred. Delete the hash table and return the error code. */
3869 assert( rc!=SQLITE_OK );
3870 if( pHash ){
3871 sqlite3Fts3HashClear(pHash);
3872 sqlite3_free(pHash);
3874 return rc;
3878 ** Allocate an Fts3MultiSegReader for each token in the expression headed
3879 ** by pExpr.
3881 ** An Fts3SegReader object is a cursor that can seek or scan a range of
3882 ** entries within a single segment b-tree. An Fts3MultiSegReader uses multiple
3883 ** Fts3SegReader objects internally to provide an interface to seek or scan
3884 ** within the union of all segments of a b-tree. Hence the name.
3886 ** If the allocated Fts3MultiSegReader just seeks to a single entry in a
3887 ** segment b-tree (if the term is not a prefix or it is a prefix for which
3888 ** there exists prefix b-tree of the right length) then it may be traversed
3889 ** and merged incrementally. Otherwise, it has to be merged into an in-memory
3890 ** doclist and then traversed.
3892 static void fts3EvalAllocateReaders(
3893 Fts3Cursor *pCsr, /* FTS cursor handle */
3894 Fts3Expr *pExpr, /* Allocate readers for this expression */
3895 int *pnToken, /* OUT: Total number of tokens in phrase. */
3896 int *pnOr, /* OUT: Total number of OR nodes in expr. */
3897 int *pRc /* IN/OUT: Error code */
3899 if( pExpr && SQLITE_OK==*pRc ){
3900 if( pExpr->eType==FTSQUERY_PHRASE ){
3901 int i;
3902 int nToken = pExpr->pPhrase->nToken;
3903 *pnToken += nToken;
3904 for(i=0; i<nToken; i++){
3905 Fts3PhraseToken *pToken = &pExpr->pPhrase->aToken[i];
3906 int rc = fts3TermSegReaderCursor(pCsr,
3907 pToken->z, pToken->n, pToken->isPrefix, &pToken->pSegcsr
3909 if( rc!=SQLITE_OK ){
3910 *pRc = rc;
3911 return;
3914 assert( pExpr->pPhrase->iDoclistToken==0 );
3915 pExpr->pPhrase->iDoclistToken = -1;
3916 }else{
3917 *pnOr += (pExpr->eType==FTSQUERY_OR);
3918 fts3EvalAllocateReaders(pCsr, pExpr->pLeft, pnToken, pnOr, pRc);
3919 fts3EvalAllocateReaders(pCsr, pExpr->pRight, pnToken, pnOr, pRc);
3925 ** Arguments pList/nList contain the doclist for token iToken of phrase p.
3926 ** It is merged into the main doclist stored in p->doclist.aAll/nAll.
3928 ** This function assumes that pList points to a buffer allocated using
3929 ** sqlite3_malloc(). This function takes responsibility for eventually
3930 ** freeing the buffer.
3932 static void fts3EvalPhraseMergeToken(
3933 Fts3Table *pTab, /* FTS Table pointer */
3934 Fts3Phrase *p, /* Phrase to merge pList/nList into */
3935 int iToken, /* Token pList/nList corresponds to */
3936 char *pList, /* Pointer to doclist */
3937 int nList /* Number of bytes in pList */
3939 assert( iToken!=p->iDoclistToken );
3941 if( pList==0 ){
3942 sqlite3_free(p->doclist.aAll);
3943 p->doclist.aAll = 0;
3944 p->doclist.nAll = 0;
3947 else if( p->iDoclistToken<0 ){
3948 p->doclist.aAll = pList;
3949 p->doclist.nAll = nList;
3952 else if( p->doclist.aAll==0 ){
3953 sqlite3_free(pList);
3956 else {
3957 char *pLeft;
3958 char *pRight;
3959 int nLeft;
3960 int nRight;
3961 int nDiff;
3963 if( p->iDoclistToken<iToken ){
3964 pLeft = p->doclist.aAll;
3965 nLeft = p->doclist.nAll;
3966 pRight = pList;
3967 nRight = nList;
3968 nDiff = iToken - p->iDoclistToken;
3969 }else{
3970 pRight = p->doclist.aAll;
3971 nRight = p->doclist.nAll;
3972 pLeft = pList;
3973 nLeft = nList;
3974 nDiff = p->iDoclistToken - iToken;
3977 fts3DoclistPhraseMerge(pTab->bDescIdx, nDiff, pLeft, nLeft, pRight,&nRight);
3978 sqlite3_free(pLeft);
3979 p->doclist.aAll = pRight;
3980 p->doclist.nAll = nRight;
3983 if( iToken>p->iDoclistToken ) p->iDoclistToken = iToken;
3987 ** Load the doclist for phrase p into p->doclist.aAll/nAll. The loaded doclist
3988 ** does not take deferred tokens into account.
3990 ** SQLITE_OK is returned if no error occurs, otherwise an SQLite error code.
3992 static int fts3EvalPhraseLoad(
3993 Fts3Cursor *pCsr, /* FTS Cursor handle */
3994 Fts3Phrase *p /* Phrase object */
3996 Fts3Table *pTab = (Fts3Table *)pCsr->base.pVtab;
3997 int iToken;
3998 int rc = SQLITE_OK;
4000 for(iToken=0; rc==SQLITE_OK && iToken<p->nToken; iToken++){
4001 Fts3PhraseToken *pToken = &p->aToken[iToken];
4002 assert( pToken->pDeferred==0 || pToken->pSegcsr==0 );
4004 if( pToken->pSegcsr ){
4005 int nThis = 0;
4006 char *pThis = 0;
4007 rc = fts3TermSelect(pTab, pToken, p->iColumn, &nThis, &pThis);
4008 if( rc==SQLITE_OK ){
4009 fts3EvalPhraseMergeToken(pTab, p, iToken, pThis, nThis);
4012 assert( pToken->pSegcsr==0 );
4015 return rc;
4019 ** This function is called on each phrase after the position lists for
4020 ** any deferred tokens have been loaded into memory. It updates the phrases
4021 ** current position list to include only those positions that are really
4022 ** instances of the phrase (after considering deferred tokens). If this
4023 ** means that the phrase does not appear in the current row, doclist.pList
4024 ** and doclist.nList are both zeroed.
4026 ** SQLITE_OK is returned if no error occurs, otherwise an SQLite error code.
4028 static int fts3EvalDeferredPhrase(Fts3Cursor *pCsr, Fts3Phrase *pPhrase){
4029 int iToken; /* Used to iterate through phrase tokens */
4030 char *aPoslist = 0; /* Position list for deferred tokens */
4031 int nPoslist = 0; /* Number of bytes in aPoslist */
4032 int iPrev = -1; /* Token number of previous deferred token */
4034 assert( pPhrase->doclist.bFreeList==0 );
4036 for(iToken=0; iToken<pPhrase->nToken; iToken++){
4037 Fts3PhraseToken *pToken = &pPhrase->aToken[iToken];
4038 Fts3DeferredToken *pDeferred = pToken->pDeferred;
4040 if( pDeferred ){
4041 char *pList = 0;
4042 int nList = 0;
4043 int rc = sqlite3Fts3DeferredTokenList(pDeferred, &pList, &nList);
4044 if( rc!=SQLITE_OK ) return rc;
4046 if( pList==0 ){
4047 sqlite3_free(aPoslist);
4048 pPhrase->doclist.pList = 0;
4049 pPhrase->doclist.nList = 0;
4050 return SQLITE_OK;
4052 }else if( aPoslist==0 ){
4053 aPoslist = pList;
4054 nPoslist = nList;
4056 }else{
4057 char *aOut = pList;
4058 char *p1 = aPoslist;
4059 char *p2 = aOut;
4061 assert( iPrev>=0 );
4062 fts3PoslistPhraseMerge(&aOut, iToken-iPrev, 0, 1, &p1, &p2);
4063 sqlite3_free(aPoslist);
4064 aPoslist = pList;
4065 nPoslist = (int)(aOut - aPoslist);
4066 if( nPoslist==0 ){
4067 sqlite3_free(aPoslist);
4068 pPhrase->doclist.pList = 0;
4069 pPhrase->doclist.nList = 0;
4070 return SQLITE_OK;
4073 iPrev = iToken;
4077 if( iPrev>=0 ){
4078 int nMaxUndeferred = pPhrase->iDoclistToken;
4079 if( nMaxUndeferred<0 ){
4080 pPhrase->doclist.pList = aPoslist;
4081 pPhrase->doclist.nList = nPoslist;
4082 pPhrase->doclist.iDocid = pCsr->iPrevId;
4083 pPhrase->doclist.bFreeList = 1;
4084 }else{
4085 int nDistance;
4086 char *p1;
4087 char *p2;
4088 char *aOut;
4090 if( nMaxUndeferred>iPrev ){
4091 p1 = aPoslist;
4092 p2 = pPhrase->doclist.pList;
4093 nDistance = nMaxUndeferred - iPrev;
4094 }else{
4095 p1 = pPhrase->doclist.pList;
4096 p2 = aPoslist;
4097 nDistance = iPrev - nMaxUndeferred;
4100 aOut = (char *)sqlite3_malloc(nPoslist+8);
4101 if( !aOut ){
4102 sqlite3_free(aPoslist);
4103 return SQLITE_NOMEM;
4106 pPhrase->doclist.pList = aOut;
4107 if( fts3PoslistPhraseMerge(&aOut, nDistance, 0, 1, &p1, &p2) ){
4108 pPhrase->doclist.bFreeList = 1;
4109 pPhrase->doclist.nList = (int)(aOut - pPhrase->doclist.pList);
4110 }else{
4111 sqlite3_free(aOut);
4112 pPhrase->doclist.pList = 0;
4113 pPhrase->doclist.nList = 0;
4115 sqlite3_free(aPoslist);
4119 return SQLITE_OK;
4123 ** Maximum number of tokens a phrase may have to be considered for the
4124 ** incremental doclists strategy.
4126 #define MAX_INCR_PHRASE_TOKENS 4
4129 ** This function is called for each Fts3Phrase in a full-text query
4130 ** expression to initialize the mechanism for returning rows. Once this
4131 ** function has been called successfully on an Fts3Phrase, it may be
4132 ** used with fts3EvalPhraseNext() to iterate through the matching docids.
4134 ** If parameter bOptOk is true, then the phrase may (or may not) use the
4135 ** incremental loading strategy. Otherwise, the entire doclist is loaded into
4136 ** memory within this call.
4138 ** SQLITE_OK is returned if no error occurs, otherwise an SQLite error code.
4140 static int fts3EvalPhraseStart(Fts3Cursor *pCsr, int bOptOk, Fts3Phrase *p){
4141 Fts3Table *pTab = (Fts3Table *)pCsr->base.pVtab;
4142 int rc = SQLITE_OK; /* Error code */
4143 int i;
4145 /* Determine if doclists may be loaded from disk incrementally. This is
4146 ** possible if the bOptOk argument is true, the FTS doclists will be
4147 ** scanned in forward order, and the phrase consists of
4148 ** MAX_INCR_PHRASE_TOKENS or fewer tokens, none of which are are "^first"
4149 ** tokens or prefix tokens that cannot use a prefix-index. */
4150 int bHaveIncr = 0;
4151 int bIncrOk = (bOptOk
4152 && pCsr->bDesc==pTab->bDescIdx
4153 && p->nToken<=MAX_INCR_PHRASE_TOKENS && p->nToken>0
4154 && p->nToken<=MAX_INCR_PHRASE_TOKENS && p->nToken>0
4155 #ifdef SQLITE_TEST
4156 && pTab->bNoIncrDoclist==0
4157 #endif
4159 for(i=0; bIncrOk==1 && i<p->nToken; i++){
4160 Fts3PhraseToken *pToken = &p->aToken[i];
4161 if( pToken->bFirst || (pToken->pSegcsr!=0 && !pToken->pSegcsr->bLookup) ){
4162 bIncrOk = 0;
4164 if( pToken->pSegcsr ) bHaveIncr = 1;
4167 if( bIncrOk && bHaveIncr ){
4168 /* Use the incremental approach. */
4169 int iCol = (p->iColumn >= pTab->nColumn ? -1 : p->iColumn);
4170 for(i=0; rc==SQLITE_OK && i<p->nToken; i++){
4171 Fts3PhraseToken *pToken = &p->aToken[i];
4172 Fts3MultiSegReader *pSegcsr = pToken->pSegcsr;
4173 if( pSegcsr ){
4174 rc = sqlite3Fts3MsrIncrStart(pTab, pSegcsr, iCol, pToken->z, pToken->n);
4177 p->bIncr = 1;
4178 }else{
4179 /* Load the full doclist for the phrase into memory. */
4180 rc = fts3EvalPhraseLoad(pCsr, p);
4181 p->bIncr = 0;
4184 assert( rc!=SQLITE_OK || p->nToken<1 || p->aToken[0].pSegcsr==0 || p->bIncr );
4185 return rc;
4189 ** This function is used to iterate backwards (from the end to start)
4190 ** through doclists. It is used by this module to iterate through phrase
4191 ** doclists in reverse and by the fts3_write.c module to iterate through
4192 ** pending-terms lists when writing to databases with "order=desc".
4194 ** The doclist may be sorted in ascending (parameter bDescIdx==0) or
4195 ** descending (parameter bDescIdx==1) order of docid. Regardless, this
4196 ** function iterates from the end of the doclist to the beginning.
4198 void sqlite3Fts3DoclistPrev(
4199 int bDescIdx, /* True if the doclist is desc */
4200 char *aDoclist, /* Pointer to entire doclist */
4201 int nDoclist, /* Length of aDoclist in bytes */
4202 char **ppIter, /* IN/OUT: Iterator pointer */
4203 sqlite3_int64 *piDocid, /* IN/OUT: Docid pointer */
4204 int *pnList, /* OUT: List length pointer */
4205 u8 *pbEof /* OUT: End-of-file flag */
4207 char *p = *ppIter;
4209 assert( nDoclist>0 );
4210 assert( *pbEof==0 );
4211 assert( p || *piDocid==0 );
4212 assert( !p || (p>aDoclist && p<&aDoclist[nDoclist]) );
4214 if( p==0 ){
4215 sqlite3_int64 iDocid = 0;
4216 char *pNext = 0;
4217 char *pDocid = aDoclist;
4218 char *pEnd = &aDoclist[nDoclist];
4219 int iMul = 1;
4221 while( pDocid<pEnd ){
4222 sqlite3_int64 iDelta;
4223 pDocid += sqlite3Fts3GetVarint(pDocid, &iDelta);
4224 iDocid += (iMul * iDelta);
4225 pNext = pDocid;
4226 fts3PoslistCopy(0, &pDocid);
4227 while( pDocid<pEnd && *pDocid==0 ) pDocid++;
4228 iMul = (bDescIdx ? -1 : 1);
4231 *pnList = (int)(pEnd - pNext);
4232 *ppIter = pNext;
4233 *piDocid = iDocid;
4234 }else{
4235 int iMul = (bDescIdx ? -1 : 1);
4236 sqlite3_int64 iDelta;
4237 fts3GetReverseVarint(&p, aDoclist, &iDelta);
4238 *piDocid -= (iMul * iDelta);
4240 if( p==aDoclist ){
4241 *pbEof = 1;
4242 }else{
4243 char *pSave = p;
4244 fts3ReversePoslist(aDoclist, &p);
4245 *pnList = (int)(pSave - p);
4247 *ppIter = p;
4252 ** Iterate forwards through a doclist.
4254 void sqlite3Fts3DoclistNext(
4255 int bDescIdx, /* True if the doclist is desc */
4256 char *aDoclist, /* Pointer to entire doclist */
4257 int nDoclist, /* Length of aDoclist in bytes */
4258 char **ppIter, /* IN/OUT: Iterator pointer */
4259 sqlite3_int64 *piDocid, /* IN/OUT: Docid pointer */
4260 u8 *pbEof /* OUT: End-of-file flag */
4262 char *p = *ppIter;
4264 assert( nDoclist>0 );
4265 assert( *pbEof==0 );
4266 assert( p || *piDocid==0 );
4267 assert( !p || (p>=aDoclist && p<=&aDoclist[nDoclist]) );
4269 if( p==0 ){
4270 p = aDoclist;
4271 p += sqlite3Fts3GetVarint(p, piDocid);
4272 }else{
4273 fts3PoslistCopy(0, &p);
4274 if( p>=&aDoclist[nDoclist] ){
4275 *pbEof = 1;
4276 }else{
4277 sqlite3_int64 iVar;
4278 p += sqlite3Fts3GetVarint(p, &iVar);
4279 *piDocid += ((bDescIdx ? -1 : 1) * iVar);
4283 *ppIter = p;
4287 ** Advance the iterator pDL to the next entry in pDL->aAll/nAll. Set *pbEof
4288 ** to true if EOF is reached.
4290 static void fts3EvalDlPhraseNext(
4291 Fts3Table *pTab,
4292 Fts3Doclist *pDL,
4293 u8 *pbEof
4295 char *pIter; /* Used to iterate through aAll */
4296 char *pEnd = &pDL->aAll[pDL->nAll]; /* 1 byte past end of aAll */
4298 if( pDL->pNextDocid ){
4299 pIter = pDL->pNextDocid;
4300 }else{
4301 pIter = pDL->aAll;
4304 if( pIter>=pEnd ){
4305 /* We have already reached the end of this doclist. EOF. */
4306 *pbEof = 1;
4307 }else{
4308 sqlite3_int64 iDelta;
4309 pIter += sqlite3Fts3GetVarint(pIter, &iDelta);
4310 if( pTab->bDescIdx==0 || pDL->pNextDocid==0 ){
4311 pDL->iDocid += iDelta;
4312 }else{
4313 pDL->iDocid -= iDelta;
4315 pDL->pList = pIter;
4316 fts3PoslistCopy(0, &pIter);
4317 pDL->nList = (int)(pIter - pDL->pList);
4319 /* pIter now points just past the 0x00 that terminates the position-
4320 ** list for document pDL->iDocid. However, if this position-list was
4321 ** edited in place by fts3EvalNearTrim(), then pIter may not actually
4322 ** point to the start of the next docid value. The following line deals
4323 ** with this case by advancing pIter past the zero-padding added by
4324 ** fts3EvalNearTrim(). */
4325 while( pIter<pEnd && *pIter==0 ) pIter++;
4327 pDL->pNextDocid = pIter;
4328 assert( pIter>=&pDL->aAll[pDL->nAll] || *pIter );
4329 *pbEof = 0;
4334 ** Helper type used by fts3EvalIncrPhraseNext() and incrPhraseTokenNext().
4336 typedef struct TokenDoclist TokenDoclist;
4337 struct TokenDoclist {
4338 int bIgnore;
4339 sqlite3_int64 iDocid;
4340 char *pList;
4341 int nList;
4345 ** Token pToken is an incrementally loaded token that is part of a
4346 ** multi-token phrase. Advance it to the next matching document in the
4347 ** database and populate output variable *p with the details of the new
4348 ** entry. Or, if the iterator has reached EOF, set *pbEof to true.
4350 ** If an error occurs, return an SQLite error code. Otherwise, return
4351 ** SQLITE_OK.
4353 static int incrPhraseTokenNext(
4354 Fts3Table *pTab, /* Virtual table handle */
4355 Fts3Phrase *pPhrase, /* Phrase to advance token of */
4356 int iToken, /* Specific token to advance */
4357 TokenDoclist *p, /* OUT: Docid and doclist for new entry */
4358 u8 *pbEof /* OUT: True if iterator is at EOF */
4360 int rc = SQLITE_OK;
4362 if( pPhrase->iDoclistToken==iToken ){
4363 assert( p->bIgnore==0 );
4364 assert( pPhrase->aToken[iToken].pSegcsr==0 );
4365 fts3EvalDlPhraseNext(pTab, &pPhrase->doclist, pbEof);
4366 p->pList = pPhrase->doclist.pList;
4367 p->nList = pPhrase->doclist.nList;
4368 p->iDocid = pPhrase->doclist.iDocid;
4369 }else{
4370 Fts3PhraseToken *pToken = &pPhrase->aToken[iToken];
4371 assert( pToken->pDeferred==0 );
4372 assert( pToken->pSegcsr || pPhrase->iDoclistToken>=0 );
4373 if( pToken->pSegcsr ){
4374 assert( p->bIgnore==0 );
4375 rc = sqlite3Fts3MsrIncrNext(
4376 pTab, pToken->pSegcsr, &p->iDocid, &p->pList, &p->nList
4378 if( p->pList==0 ) *pbEof = 1;
4379 }else{
4380 p->bIgnore = 1;
4384 return rc;
4389 ** The phrase iterator passed as the second argument:
4391 ** * features at least one token that uses an incremental doclist, and
4393 ** * does not contain any deferred tokens.
4395 ** Advance it to the next matching documnent in the database and populate
4396 ** the Fts3Doclist.pList and nList fields.
4398 ** If there is no "next" entry and no error occurs, then *pbEof is set to
4399 ** 1 before returning. Otherwise, if no error occurs and the iterator is
4400 ** successfully advanced, *pbEof is set to 0.
4402 ** If an error occurs, return an SQLite error code. Otherwise, return
4403 ** SQLITE_OK.
4405 static int fts3EvalIncrPhraseNext(
4406 Fts3Cursor *pCsr, /* FTS Cursor handle */
4407 Fts3Phrase *p, /* Phrase object to advance to next docid */
4408 u8 *pbEof /* OUT: Set to 1 if EOF */
4410 int rc = SQLITE_OK;
4411 Fts3Doclist *pDL = &p->doclist;
4412 Fts3Table *pTab = (Fts3Table *)pCsr->base.pVtab;
4413 u8 bEof = 0;
4415 /* This is only called if it is guaranteed that the phrase has at least
4416 ** one incremental token. In which case the bIncr flag is set. */
4417 assert( p->bIncr==1 );
4419 if( p->nToken==1 && p->bIncr ){
4420 rc = sqlite3Fts3MsrIncrNext(pTab, p->aToken[0].pSegcsr,
4421 &pDL->iDocid, &pDL->pList, &pDL->nList
4423 if( pDL->pList==0 ) bEof = 1;
4424 }else{
4425 int bDescDoclist = pCsr->bDesc;
4426 struct TokenDoclist a[MAX_INCR_PHRASE_TOKENS];
4428 memset(a, 0, sizeof(a));
4429 assert( p->nToken<=MAX_INCR_PHRASE_TOKENS );
4430 assert( p->iDoclistToken<MAX_INCR_PHRASE_TOKENS );
4432 while( bEof==0 ){
4433 int bMaxSet = 0;
4434 sqlite3_int64 iMax = 0; /* Largest docid for all iterators */
4435 int i; /* Used to iterate through tokens */
4437 /* Advance the iterator for each token in the phrase once. */
4438 for(i=0; rc==SQLITE_OK && i<p->nToken && bEof==0; i++){
4439 rc = incrPhraseTokenNext(pTab, p, i, &a[i], &bEof);
4440 if( a[i].bIgnore==0 && (bMaxSet==0 || DOCID_CMP(iMax, a[i].iDocid)<0) ){
4441 iMax = a[i].iDocid;
4442 bMaxSet = 1;
4445 assert( rc!=SQLITE_OK || (p->nToken>=1 && a[p->nToken-1].bIgnore==0) );
4446 assert( rc!=SQLITE_OK || bMaxSet );
4448 /* Keep advancing iterators until they all point to the same document */
4449 for(i=0; i<p->nToken; i++){
4450 while( rc==SQLITE_OK && bEof==0
4451 && a[i].bIgnore==0 && DOCID_CMP(a[i].iDocid, iMax)<0
4453 rc = incrPhraseTokenNext(pTab, p, i, &a[i], &bEof);
4454 if( DOCID_CMP(a[i].iDocid, iMax)>0 ){
4455 iMax = a[i].iDocid;
4456 i = 0;
4461 /* Check if the current entries really are a phrase match */
4462 if( bEof==0 ){
4463 int nList = 0;
4464 int nByte = a[p->nToken-1].nList;
4465 char *aDoclist = sqlite3_malloc(nByte+1);
4466 if( !aDoclist ) return SQLITE_NOMEM;
4467 memcpy(aDoclist, a[p->nToken-1].pList, nByte+1);
4469 for(i=0; i<(p->nToken-1); i++){
4470 if( a[i].bIgnore==0 ){
4471 char *pL = a[i].pList;
4472 char *pR = aDoclist;
4473 char *pOut = aDoclist;
4474 int nDist = p->nToken-1-i;
4475 int res = fts3PoslistPhraseMerge(&pOut, nDist, 0, 1, &pL, &pR);
4476 if( res==0 ) break;
4477 nList = (int)(pOut - aDoclist);
4480 if( i==(p->nToken-1) ){
4481 pDL->iDocid = iMax;
4482 pDL->pList = aDoclist;
4483 pDL->nList = nList;
4484 pDL->bFreeList = 1;
4485 break;
4487 sqlite3_free(aDoclist);
4492 *pbEof = bEof;
4493 return rc;
4497 ** Attempt to move the phrase iterator to point to the next matching docid.
4498 ** If an error occurs, return an SQLite error code. Otherwise, return
4499 ** SQLITE_OK.
4501 ** If there is no "next" entry and no error occurs, then *pbEof is set to
4502 ** 1 before returning. Otherwise, if no error occurs and the iterator is
4503 ** successfully advanced, *pbEof is set to 0.
4505 static int fts3EvalPhraseNext(
4506 Fts3Cursor *pCsr, /* FTS Cursor handle */
4507 Fts3Phrase *p, /* Phrase object to advance to next docid */
4508 u8 *pbEof /* OUT: Set to 1 if EOF */
4510 int rc = SQLITE_OK;
4511 Fts3Doclist *pDL = &p->doclist;
4512 Fts3Table *pTab = (Fts3Table *)pCsr->base.pVtab;
4514 if( p->bIncr ){
4515 rc = fts3EvalIncrPhraseNext(pCsr, p, pbEof);
4516 }else if( pCsr->bDesc!=pTab->bDescIdx && pDL->nAll ){
4517 sqlite3Fts3DoclistPrev(pTab->bDescIdx, pDL->aAll, pDL->nAll,
4518 &pDL->pNextDocid, &pDL->iDocid, &pDL->nList, pbEof
4520 pDL->pList = pDL->pNextDocid;
4521 }else{
4522 fts3EvalDlPhraseNext(pTab, pDL, pbEof);
4525 return rc;
4530 ** If *pRc is not SQLITE_OK when this function is called, it is a no-op.
4531 ** Otherwise, fts3EvalPhraseStart() is called on all phrases within the
4532 ** expression. Also the Fts3Expr.bDeferred variable is set to true for any
4533 ** expressions for which all descendent tokens are deferred.
4535 ** If parameter bOptOk is zero, then it is guaranteed that the
4536 ** Fts3Phrase.doclist.aAll/nAll variables contain the entire doclist for
4537 ** each phrase in the expression (subject to deferred token processing).
4538 ** Or, if bOptOk is non-zero, then one or more tokens within the expression
4539 ** may be loaded incrementally, meaning doclist.aAll/nAll is not available.
4541 ** If an error occurs within this function, *pRc is set to an SQLite error
4542 ** code before returning.
4544 static void fts3EvalStartReaders(
4545 Fts3Cursor *pCsr, /* FTS Cursor handle */
4546 Fts3Expr *pExpr, /* Expression to initialize phrases in */
4547 int *pRc /* IN/OUT: Error code */
4549 if( pExpr && SQLITE_OK==*pRc ){
4550 if( pExpr->eType==FTSQUERY_PHRASE ){
4551 int i;
4552 int nToken = pExpr->pPhrase->nToken;
4553 for(i=0; i<nToken; i++){
4554 if( pExpr->pPhrase->aToken[i].pDeferred==0 ) break;
4556 pExpr->bDeferred = (i==nToken);
4557 *pRc = fts3EvalPhraseStart(pCsr, 1, pExpr->pPhrase);
4558 }else{
4559 fts3EvalStartReaders(pCsr, pExpr->pLeft, pRc);
4560 fts3EvalStartReaders(pCsr, pExpr->pRight, pRc);
4561 pExpr->bDeferred = (pExpr->pLeft->bDeferred && pExpr->pRight->bDeferred);
4567 ** An array of the following structures is assembled as part of the process
4568 ** of selecting tokens to defer before the query starts executing (as part
4569 ** of the xFilter() method). There is one element in the array for each
4570 ** token in the FTS expression.
4572 ** Tokens are divided into AND/NEAR clusters. All tokens in a cluster belong
4573 ** to phrases that are connected only by AND and NEAR operators (not OR or
4574 ** NOT). When determining tokens to defer, each AND/NEAR cluster is considered
4575 ** separately. The root of a tokens AND/NEAR cluster is stored in
4576 ** Fts3TokenAndCost.pRoot.
4578 typedef struct Fts3TokenAndCost Fts3TokenAndCost;
4579 struct Fts3TokenAndCost {
4580 Fts3Phrase *pPhrase; /* The phrase the token belongs to */
4581 int iToken; /* Position of token in phrase */
4582 Fts3PhraseToken *pToken; /* The token itself */
4583 Fts3Expr *pRoot; /* Root of NEAR/AND cluster */
4584 int nOvfl; /* Number of overflow pages to load doclist */
4585 int iCol; /* The column the token must match */
4589 ** This function is used to populate an allocated Fts3TokenAndCost array.
4591 ** If *pRc is not SQLITE_OK when this function is called, it is a no-op.
4592 ** Otherwise, if an error occurs during execution, *pRc is set to an
4593 ** SQLite error code.
4595 static void fts3EvalTokenCosts(
4596 Fts3Cursor *pCsr, /* FTS Cursor handle */
4597 Fts3Expr *pRoot, /* Root of current AND/NEAR cluster */
4598 Fts3Expr *pExpr, /* Expression to consider */
4599 Fts3TokenAndCost **ppTC, /* Write new entries to *(*ppTC)++ */
4600 Fts3Expr ***ppOr, /* Write new OR root to *(*ppOr)++ */
4601 int *pRc /* IN/OUT: Error code */
4603 if( *pRc==SQLITE_OK ){
4604 if( pExpr->eType==FTSQUERY_PHRASE ){
4605 Fts3Phrase *pPhrase = pExpr->pPhrase;
4606 int i;
4607 for(i=0; *pRc==SQLITE_OK && i<pPhrase->nToken; i++){
4608 Fts3TokenAndCost *pTC = (*ppTC)++;
4609 pTC->pPhrase = pPhrase;
4610 pTC->iToken = i;
4611 pTC->pRoot = pRoot;
4612 pTC->pToken = &pPhrase->aToken[i];
4613 pTC->iCol = pPhrase->iColumn;
4614 *pRc = sqlite3Fts3MsrOvfl(pCsr, pTC->pToken->pSegcsr, &pTC->nOvfl);
4616 }else if( pExpr->eType!=FTSQUERY_NOT ){
4617 assert( pExpr->eType==FTSQUERY_OR
4618 || pExpr->eType==FTSQUERY_AND
4619 || pExpr->eType==FTSQUERY_NEAR
4621 assert( pExpr->pLeft && pExpr->pRight );
4622 if( pExpr->eType==FTSQUERY_OR ){
4623 pRoot = pExpr->pLeft;
4624 **ppOr = pRoot;
4625 (*ppOr)++;
4627 fts3EvalTokenCosts(pCsr, pRoot, pExpr->pLeft, ppTC, ppOr, pRc);
4628 if( pExpr->eType==FTSQUERY_OR ){
4629 pRoot = pExpr->pRight;
4630 **ppOr = pRoot;
4631 (*ppOr)++;
4633 fts3EvalTokenCosts(pCsr, pRoot, pExpr->pRight, ppTC, ppOr, pRc);
4639 ** Determine the average document (row) size in pages. If successful,
4640 ** write this value to *pnPage and return SQLITE_OK. Otherwise, return
4641 ** an SQLite error code.
4643 ** The average document size in pages is calculated by first calculating
4644 ** determining the average size in bytes, B. If B is less than the amount
4645 ** of data that will fit on a single leaf page of an intkey table in
4646 ** this database, then the average docsize is 1. Otherwise, it is 1 plus
4647 ** the number of overflow pages consumed by a record B bytes in size.
4649 static int fts3EvalAverageDocsize(Fts3Cursor *pCsr, int *pnPage){
4650 if( pCsr->nRowAvg==0 ){
4651 /* The average document size, which is required to calculate the cost
4652 ** of each doclist, has not yet been determined. Read the required
4653 ** data from the %_stat table to calculate it.
4655 ** Entry 0 of the %_stat table is a blob containing (nCol+1) FTS3
4656 ** varints, where nCol is the number of columns in the FTS3 table.
4657 ** The first varint is the number of documents currently stored in
4658 ** the table. The following nCol varints contain the total amount of
4659 ** data stored in all rows of each column of the table, from left
4660 ** to right.
4662 int rc;
4663 Fts3Table *p = (Fts3Table*)pCsr->base.pVtab;
4664 sqlite3_stmt *pStmt;
4665 sqlite3_int64 nDoc = 0;
4666 sqlite3_int64 nByte = 0;
4667 const char *pEnd;
4668 const char *a;
4670 rc = sqlite3Fts3SelectDoctotal(p, &pStmt);
4671 if( rc!=SQLITE_OK ) return rc;
4672 a = sqlite3_column_blob(pStmt, 0);
4673 assert( a );
4675 pEnd = &a[sqlite3_column_bytes(pStmt, 0)];
4676 a += sqlite3Fts3GetVarint(a, &nDoc);
4677 while( a<pEnd ){
4678 a += sqlite3Fts3GetVarint(a, &nByte);
4680 if( nDoc==0 || nByte==0 ){
4681 sqlite3_reset(pStmt);
4682 return FTS_CORRUPT_VTAB;
4685 pCsr->nDoc = nDoc;
4686 pCsr->nRowAvg = (int)(((nByte / nDoc) + p->nPgsz) / p->nPgsz);
4687 assert( pCsr->nRowAvg>0 );
4688 rc = sqlite3_reset(pStmt);
4689 if( rc!=SQLITE_OK ) return rc;
4692 *pnPage = pCsr->nRowAvg;
4693 return SQLITE_OK;
4697 ** This function is called to select the tokens (if any) that will be
4698 ** deferred. The array aTC[] has already been populated when this is
4699 ** called.
4701 ** This function is called once for each AND/NEAR cluster in the
4702 ** expression. Each invocation determines which tokens to defer within
4703 ** the cluster with root node pRoot. See comments above the definition
4704 ** of struct Fts3TokenAndCost for more details.
4706 ** If no error occurs, SQLITE_OK is returned and sqlite3Fts3DeferToken()
4707 ** called on each token to defer. Otherwise, an SQLite error code is
4708 ** returned.
4710 static int fts3EvalSelectDeferred(
4711 Fts3Cursor *pCsr, /* FTS Cursor handle */
4712 Fts3Expr *pRoot, /* Consider tokens with this root node */
4713 Fts3TokenAndCost *aTC, /* Array of expression tokens and costs */
4714 int nTC /* Number of entries in aTC[] */
4716 Fts3Table *pTab = (Fts3Table *)pCsr->base.pVtab;
4717 int nDocSize = 0; /* Number of pages per doc loaded */
4718 int rc = SQLITE_OK; /* Return code */
4719 int ii; /* Iterator variable for various purposes */
4720 int nOvfl = 0; /* Total overflow pages used by doclists */
4721 int nToken = 0; /* Total number of tokens in cluster */
4723 int nMinEst = 0; /* The minimum count for any phrase so far. */
4724 int nLoad4 = 1; /* (Phrases that will be loaded)^4. */
4726 /* Tokens are never deferred for FTS tables created using the content=xxx
4727 ** option. The reason being that it is not guaranteed that the content
4728 ** table actually contains the same data as the index. To prevent this from
4729 ** causing any problems, the deferred token optimization is completely
4730 ** disabled for content=xxx tables. */
4731 if( pTab->zContentTbl ){
4732 return SQLITE_OK;
4735 /* Count the tokens in this AND/NEAR cluster. If none of the doclists
4736 ** associated with the tokens spill onto overflow pages, or if there is
4737 ** only 1 token, exit early. No tokens to defer in this case. */
4738 for(ii=0; ii<nTC; ii++){
4739 if( aTC[ii].pRoot==pRoot ){
4740 nOvfl += aTC[ii].nOvfl;
4741 nToken++;
4744 if( nOvfl==0 || nToken<2 ) return SQLITE_OK;
4746 /* Obtain the average docsize (in pages). */
4747 rc = fts3EvalAverageDocsize(pCsr, &nDocSize);
4748 assert( rc!=SQLITE_OK || nDocSize>0 );
4751 /* Iterate through all tokens in this AND/NEAR cluster, in ascending order
4752 ** of the number of overflow pages that will be loaded by the pager layer
4753 ** to retrieve the entire doclist for the token from the full-text index.
4754 ** Load the doclists for tokens that are either:
4756 ** a. The cheapest token in the entire query (i.e. the one visited by the
4757 ** first iteration of this loop), or
4759 ** b. Part of a multi-token phrase.
4761 ** After each token doclist is loaded, merge it with the others from the
4762 ** same phrase and count the number of documents that the merged doclist
4763 ** contains. Set variable "nMinEst" to the smallest number of documents in
4764 ** any phrase doclist for which 1 or more token doclists have been loaded.
4765 ** Let nOther be the number of other phrases for which it is certain that
4766 ** one or more tokens will not be deferred.
4768 ** Then, for each token, defer it if loading the doclist would result in
4769 ** loading N or more overflow pages into memory, where N is computed as:
4771 ** (nMinEst + 4^nOther - 1) / (4^nOther)
4773 for(ii=0; ii<nToken && rc==SQLITE_OK; ii++){
4774 int iTC; /* Used to iterate through aTC[] array. */
4775 Fts3TokenAndCost *pTC = 0; /* Set to cheapest remaining token. */
4777 /* Set pTC to point to the cheapest remaining token. */
4778 for(iTC=0; iTC<nTC; iTC++){
4779 if( aTC[iTC].pToken && aTC[iTC].pRoot==pRoot
4780 && (!pTC || aTC[iTC].nOvfl<pTC->nOvfl)
4782 pTC = &aTC[iTC];
4785 assert( pTC );
4787 if( ii && pTC->nOvfl>=((nMinEst+(nLoad4/4)-1)/(nLoad4/4))*nDocSize ){
4788 /* The number of overflow pages to load for this (and therefore all
4789 ** subsequent) tokens is greater than the estimated number of pages
4790 ** that will be loaded if all subsequent tokens are deferred.
4792 Fts3PhraseToken *pToken = pTC->pToken;
4793 rc = sqlite3Fts3DeferToken(pCsr, pToken, pTC->iCol);
4794 fts3SegReaderCursorFree(pToken->pSegcsr);
4795 pToken->pSegcsr = 0;
4796 }else{
4797 /* Set nLoad4 to the value of (4^nOther) for the next iteration of the
4798 ** for-loop. Except, limit the value to 2^24 to prevent it from
4799 ** overflowing the 32-bit integer it is stored in. */
4800 if( ii<12 ) nLoad4 = nLoad4*4;
4802 if( ii==0 || (pTC->pPhrase->nToken>1 && ii!=nToken-1) ){
4803 /* Either this is the cheapest token in the entire query, or it is
4804 ** part of a multi-token phrase. Either way, the entire doclist will
4805 ** (eventually) be loaded into memory. It may as well be now. */
4806 Fts3PhraseToken *pToken = pTC->pToken;
4807 int nList = 0;
4808 char *pList = 0;
4809 rc = fts3TermSelect(pTab, pToken, pTC->iCol, &nList, &pList);
4810 assert( rc==SQLITE_OK || pList==0 );
4811 if( rc==SQLITE_OK ){
4812 int nCount;
4813 fts3EvalPhraseMergeToken(pTab, pTC->pPhrase, pTC->iToken,pList,nList);
4814 nCount = fts3DoclistCountDocids(
4815 pTC->pPhrase->doclist.aAll, pTC->pPhrase->doclist.nAll
4817 if( ii==0 || nCount<nMinEst ) nMinEst = nCount;
4821 pTC->pToken = 0;
4824 return rc;
4828 ** This function is called from within the xFilter method. It initializes
4829 ** the full-text query currently stored in pCsr->pExpr. To iterate through
4830 ** the results of a query, the caller does:
4832 ** fts3EvalStart(pCsr);
4833 ** while( 1 ){
4834 ** fts3EvalNext(pCsr);
4835 ** if( pCsr->bEof ) break;
4836 ** ... return row pCsr->iPrevId to the caller ...
4837 ** }
4839 static int fts3EvalStart(Fts3Cursor *pCsr){
4840 Fts3Table *pTab = (Fts3Table *)pCsr->base.pVtab;
4841 int rc = SQLITE_OK;
4842 int nToken = 0;
4843 int nOr = 0;
4845 /* Allocate a MultiSegReader for each token in the expression. */
4846 fts3EvalAllocateReaders(pCsr, pCsr->pExpr, &nToken, &nOr, &rc);
4848 /* Determine which, if any, tokens in the expression should be deferred. */
4849 #ifndef SQLITE_DISABLE_FTS4_DEFERRED
4850 if( rc==SQLITE_OK && nToken>1 && pTab->bFts4 ){
4851 Fts3TokenAndCost *aTC;
4852 Fts3Expr **apOr;
4853 aTC = (Fts3TokenAndCost *)sqlite3_malloc(
4854 sizeof(Fts3TokenAndCost) * nToken
4855 + sizeof(Fts3Expr *) * nOr * 2
4857 apOr = (Fts3Expr **)&aTC[nToken];
4859 if( !aTC ){
4860 rc = SQLITE_NOMEM;
4861 }else{
4862 int ii;
4863 Fts3TokenAndCost *pTC = aTC;
4864 Fts3Expr **ppOr = apOr;
4866 fts3EvalTokenCosts(pCsr, 0, pCsr->pExpr, &pTC, &ppOr, &rc);
4867 nToken = (int)(pTC-aTC);
4868 nOr = (int)(ppOr-apOr);
4870 if( rc==SQLITE_OK ){
4871 rc = fts3EvalSelectDeferred(pCsr, 0, aTC, nToken);
4872 for(ii=0; rc==SQLITE_OK && ii<nOr; ii++){
4873 rc = fts3EvalSelectDeferred(pCsr, apOr[ii], aTC, nToken);
4877 sqlite3_free(aTC);
4880 #endif
4882 fts3EvalStartReaders(pCsr, pCsr->pExpr, &rc);
4883 return rc;
4887 ** Invalidate the current position list for phrase pPhrase.
4889 static void fts3EvalInvalidatePoslist(Fts3Phrase *pPhrase){
4890 if( pPhrase->doclist.bFreeList ){
4891 sqlite3_free(pPhrase->doclist.pList);
4893 pPhrase->doclist.pList = 0;
4894 pPhrase->doclist.nList = 0;
4895 pPhrase->doclist.bFreeList = 0;
4899 ** This function is called to edit the position list associated with
4900 ** the phrase object passed as the fifth argument according to a NEAR
4901 ** condition. For example:
4903 ** abc NEAR/5 "def ghi"
4905 ** Parameter nNear is passed the NEAR distance of the expression (5 in
4906 ** the example above). When this function is called, *paPoslist points to
4907 ** the position list, and *pnToken is the number of phrase tokens in, the
4908 ** phrase on the other side of the NEAR operator to pPhrase. For example,
4909 ** if pPhrase refers to the "def ghi" phrase, then *paPoslist points to
4910 ** the position list associated with phrase "abc".
4912 ** All positions in the pPhrase position list that are not sufficiently
4913 ** close to a position in the *paPoslist position list are removed. If this
4914 ** leaves 0 positions, zero is returned. Otherwise, non-zero.
4916 ** Before returning, *paPoslist is set to point to the position lsit
4917 ** associated with pPhrase. And *pnToken is set to the number of tokens in
4918 ** pPhrase.
4920 static int fts3EvalNearTrim(
4921 int nNear, /* NEAR distance. As in "NEAR/nNear". */
4922 char *aTmp, /* Temporary space to use */
4923 char **paPoslist, /* IN/OUT: Position list */
4924 int *pnToken, /* IN/OUT: Tokens in phrase of *paPoslist */
4925 Fts3Phrase *pPhrase /* The phrase object to trim the doclist of */
4927 int nParam1 = nNear + pPhrase->nToken;
4928 int nParam2 = nNear + *pnToken;
4929 int nNew;
4930 char *p2;
4931 char *pOut;
4932 int res;
4934 assert( pPhrase->doclist.pList );
4936 p2 = pOut = pPhrase->doclist.pList;
4937 res = fts3PoslistNearMerge(
4938 &pOut, aTmp, nParam1, nParam2, paPoslist, &p2
4940 if( res ){
4941 nNew = (int)(pOut - pPhrase->doclist.pList) - 1;
4942 assert( pPhrase->doclist.pList[nNew]=='\0' );
4943 assert( nNew<=pPhrase->doclist.nList && nNew>0 );
4944 memset(&pPhrase->doclist.pList[nNew], 0, pPhrase->doclist.nList - nNew);
4945 pPhrase->doclist.nList = nNew;
4946 *paPoslist = pPhrase->doclist.pList;
4947 *pnToken = pPhrase->nToken;
4950 return res;
4954 ** This function is a no-op if *pRc is other than SQLITE_OK when it is called.
4955 ** Otherwise, it advances the expression passed as the second argument to
4956 ** point to the next matching row in the database. Expressions iterate through
4957 ** matching rows in docid order. Ascending order if Fts3Cursor.bDesc is zero,
4958 ** or descending if it is non-zero.
4960 ** If an error occurs, *pRc is set to an SQLite error code. Otherwise, if
4961 ** successful, the following variables in pExpr are set:
4963 ** Fts3Expr.bEof (non-zero if EOF - there is no next row)
4964 ** Fts3Expr.iDocid (valid if bEof==0. The docid of the next row)
4966 ** If the expression is of type FTSQUERY_PHRASE, and the expression is not
4967 ** at EOF, then the following variables are populated with the position list
4968 ** for the phrase for the visited row:
4970 ** FTs3Expr.pPhrase->doclist.nList (length of pList in bytes)
4971 ** FTs3Expr.pPhrase->doclist.pList (pointer to position list)
4973 ** It says above that this function advances the expression to the next
4974 ** matching row. This is usually true, but there are the following exceptions:
4976 ** 1. Deferred tokens are not taken into account. If a phrase consists
4977 ** entirely of deferred tokens, it is assumed to match every row in
4978 ** the db. In this case the position-list is not populated at all.
4980 ** Or, if a phrase contains one or more deferred tokens and one or
4981 ** more non-deferred tokens, then the expression is advanced to the
4982 ** next possible match, considering only non-deferred tokens. In other
4983 ** words, if the phrase is "A B C", and "B" is deferred, the expression
4984 ** is advanced to the next row that contains an instance of "A * C",
4985 ** where "*" may match any single token. The position list in this case
4986 ** is populated as for "A * C" before returning.
4988 ** 2. NEAR is treated as AND. If the expression is "x NEAR y", it is
4989 ** advanced to point to the next row that matches "x AND y".
4991 ** See fts3EvalTestDeferredAndNear() for details on testing if a row is
4992 ** really a match, taking into account deferred tokens and NEAR operators.
4994 static void fts3EvalNextRow(
4995 Fts3Cursor *pCsr, /* FTS Cursor handle */
4996 Fts3Expr *pExpr, /* Expr. to advance to next matching row */
4997 int *pRc /* IN/OUT: Error code */
4999 if( *pRc==SQLITE_OK ){
5000 int bDescDoclist = pCsr->bDesc; /* Used by DOCID_CMP() macro */
5001 assert( pExpr->bEof==0 );
5002 pExpr->bStart = 1;
5004 switch( pExpr->eType ){
5005 case FTSQUERY_NEAR:
5006 case FTSQUERY_AND: {
5007 Fts3Expr *pLeft = pExpr->pLeft;
5008 Fts3Expr *pRight = pExpr->pRight;
5009 assert( !pLeft->bDeferred || !pRight->bDeferred );
5011 if( pLeft->bDeferred ){
5012 /* LHS is entirely deferred. So we assume it matches every row.
5013 ** Advance the RHS iterator to find the next row visited. */
5014 fts3EvalNextRow(pCsr, pRight, pRc);
5015 pExpr->iDocid = pRight->iDocid;
5016 pExpr->bEof = pRight->bEof;
5017 }else if( pRight->bDeferred ){
5018 /* RHS is entirely deferred. So we assume it matches every row.
5019 ** Advance the LHS iterator to find the next row visited. */
5020 fts3EvalNextRow(pCsr, pLeft, pRc);
5021 pExpr->iDocid = pLeft->iDocid;
5022 pExpr->bEof = pLeft->bEof;
5023 }else{
5024 /* Neither the RHS or LHS are deferred. */
5025 fts3EvalNextRow(pCsr, pLeft, pRc);
5026 fts3EvalNextRow(pCsr, pRight, pRc);
5027 while( !pLeft->bEof && !pRight->bEof && *pRc==SQLITE_OK ){
5028 sqlite3_int64 iDiff = DOCID_CMP(pLeft->iDocid, pRight->iDocid);
5029 if( iDiff==0 ) break;
5030 if( iDiff<0 ){
5031 fts3EvalNextRow(pCsr, pLeft, pRc);
5032 }else{
5033 fts3EvalNextRow(pCsr, pRight, pRc);
5036 pExpr->iDocid = pLeft->iDocid;
5037 pExpr->bEof = (pLeft->bEof || pRight->bEof);
5039 break;
5042 case FTSQUERY_OR: {
5043 Fts3Expr *pLeft = pExpr->pLeft;
5044 Fts3Expr *pRight = pExpr->pRight;
5045 sqlite3_int64 iCmp = DOCID_CMP(pLeft->iDocid, pRight->iDocid);
5047 assert( pLeft->bStart || pLeft->iDocid==pRight->iDocid );
5048 assert( pRight->bStart || pLeft->iDocid==pRight->iDocid );
5050 if( pRight->bEof || (pLeft->bEof==0 && iCmp<0) ){
5051 fts3EvalNextRow(pCsr, pLeft, pRc);
5052 }else if( pLeft->bEof || (pRight->bEof==0 && iCmp>0) ){
5053 fts3EvalNextRow(pCsr, pRight, pRc);
5054 }else{
5055 fts3EvalNextRow(pCsr, pLeft, pRc);
5056 fts3EvalNextRow(pCsr, pRight, pRc);
5059 pExpr->bEof = (pLeft->bEof && pRight->bEof);
5060 iCmp = DOCID_CMP(pLeft->iDocid, pRight->iDocid);
5061 if( pRight->bEof || (pLeft->bEof==0 && iCmp<0) ){
5062 pExpr->iDocid = pLeft->iDocid;
5063 }else{
5064 pExpr->iDocid = pRight->iDocid;
5067 break;
5070 case FTSQUERY_NOT: {
5071 Fts3Expr *pLeft = pExpr->pLeft;
5072 Fts3Expr *pRight = pExpr->pRight;
5074 if( pRight->bStart==0 ){
5075 fts3EvalNextRow(pCsr, pRight, pRc);
5076 assert( *pRc!=SQLITE_OK || pRight->bStart );
5079 fts3EvalNextRow(pCsr, pLeft, pRc);
5080 if( pLeft->bEof==0 ){
5081 while( !*pRc
5082 && !pRight->bEof
5083 && DOCID_CMP(pLeft->iDocid, pRight->iDocid)>0
5085 fts3EvalNextRow(pCsr, pRight, pRc);
5088 pExpr->iDocid = pLeft->iDocid;
5089 pExpr->bEof = pLeft->bEof;
5090 break;
5093 default: {
5094 Fts3Phrase *pPhrase = pExpr->pPhrase;
5095 fts3EvalInvalidatePoslist(pPhrase);
5096 *pRc = fts3EvalPhraseNext(pCsr, pPhrase, &pExpr->bEof);
5097 pExpr->iDocid = pPhrase->doclist.iDocid;
5098 break;
5105 ** If *pRc is not SQLITE_OK, or if pExpr is not the root node of a NEAR
5106 ** cluster, then this function returns 1 immediately.
5108 ** Otherwise, it checks if the current row really does match the NEAR
5109 ** expression, using the data currently stored in the position lists
5110 ** (Fts3Expr->pPhrase.doclist.pList/nList) for each phrase in the expression.
5112 ** If the current row is a match, the position list associated with each
5113 ** phrase in the NEAR expression is edited in place to contain only those
5114 ** phrase instances sufficiently close to their peers to satisfy all NEAR
5115 ** constraints. In this case it returns 1. If the NEAR expression does not
5116 ** match the current row, 0 is returned. The position lists may or may not
5117 ** be edited if 0 is returned.
5119 static int fts3EvalNearTest(Fts3Expr *pExpr, int *pRc){
5120 int res = 1;
5122 /* The following block runs if pExpr is the root of a NEAR query.
5123 ** For example, the query:
5125 ** "w" NEAR "x" NEAR "y" NEAR "z"
5127 ** which is represented in tree form as:
5129 ** |
5130 ** +--NEAR--+ <-- root of NEAR query
5131 ** | |
5132 ** +--NEAR--+ "z"
5133 ** | |
5134 ** +--NEAR--+ "y"
5135 ** | |
5136 ** "w" "x"
5138 ** The right-hand child of a NEAR node is always a phrase. The
5139 ** left-hand child may be either a phrase or a NEAR node. There are
5140 ** no exceptions to this - it's the way the parser in fts3_expr.c works.
5142 if( *pRc==SQLITE_OK
5143 && pExpr->eType==FTSQUERY_NEAR
5144 && pExpr->bEof==0
5145 && (pExpr->pParent==0 || pExpr->pParent->eType!=FTSQUERY_NEAR)
5147 Fts3Expr *p;
5148 int nTmp = 0; /* Bytes of temp space */
5149 char *aTmp; /* Temp space for PoslistNearMerge() */
5151 /* Allocate temporary working space. */
5152 for(p=pExpr; p->pLeft; p=p->pLeft){
5153 nTmp += p->pRight->pPhrase->doclist.nList;
5155 nTmp += p->pPhrase->doclist.nList;
5156 if( nTmp==0 ){
5157 res = 0;
5158 }else{
5159 aTmp = sqlite3_malloc(nTmp*2);
5160 if( !aTmp ){
5161 *pRc = SQLITE_NOMEM;
5162 res = 0;
5163 }else{
5164 char *aPoslist = p->pPhrase->doclist.pList;
5165 int nToken = p->pPhrase->nToken;
5167 for(p=p->pParent;res && p && p->eType==FTSQUERY_NEAR; p=p->pParent){
5168 Fts3Phrase *pPhrase = p->pRight->pPhrase;
5169 int nNear = p->nNear;
5170 res = fts3EvalNearTrim(nNear, aTmp, &aPoslist, &nToken, pPhrase);
5173 aPoslist = pExpr->pRight->pPhrase->doclist.pList;
5174 nToken = pExpr->pRight->pPhrase->nToken;
5175 for(p=pExpr->pLeft; p && res; p=p->pLeft){
5176 int nNear;
5177 Fts3Phrase *pPhrase;
5178 assert( p->pParent && p->pParent->pLeft==p );
5179 nNear = p->pParent->nNear;
5180 pPhrase = (
5181 p->eType==FTSQUERY_NEAR ? p->pRight->pPhrase : p->pPhrase
5183 res = fts3EvalNearTrim(nNear, aTmp, &aPoslist, &nToken, pPhrase);
5187 sqlite3_free(aTmp);
5191 return res;
5195 ** This function is a helper function for fts3EvalTestDeferredAndNear().
5196 ** Assuming no error occurs or has occurred, It returns non-zero if the
5197 ** expression passed as the second argument matches the row that pCsr
5198 ** currently points to, or zero if it does not.
5200 ** If *pRc is not SQLITE_OK when this function is called, it is a no-op.
5201 ** If an error occurs during execution of this function, *pRc is set to
5202 ** the appropriate SQLite error code. In this case the returned value is
5203 ** undefined.
5205 static int fts3EvalTestExpr(
5206 Fts3Cursor *pCsr, /* FTS cursor handle */
5207 Fts3Expr *pExpr, /* Expr to test. May or may not be root. */
5208 int *pRc /* IN/OUT: Error code */
5210 int bHit = 1; /* Return value */
5211 if( *pRc==SQLITE_OK ){
5212 switch( pExpr->eType ){
5213 case FTSQUERY_NEAR:
5214 case FTSQUERY_AND:
5215 bHit = (
5216 fts3EvalTestExpr(pCsr, pExpr->pLeft, pRc)
5217 && fts3EvalTestExpr(pCsr, pExpr->pRight, pRc)
5218 && fts3EvalNearTest(pExpr, pRc)
5221 /* If the NEAR expression does not match any rows, zero the doclist for
5222 ** all phrases involved in the NEAR. This is because the snippet(),
5223 ** offsets() and matchinfo() functions are not supposed to recognize
5224 ** any instances of phrases that are part of unmatched NEAR queries.
5225 ** For example if this expression:
5227 ** ... MATCH 'a OR (b NEAR c)'
5229 ** is matched against a row containing:
5231 ** 'a b d e'
5233 ** then any snippet() should ony highlight the "a" term, not the "b"
5234 ** (as "b" is part of a non-matching NEAR clause).
5236 if( bHit==0
5237 && pExpr->eType==FTSQUERY_NEAR
5238 && (pExpr->pParent==0 || pExpr->pParent->eType!=FTSQUERY_NEAR)
5240 Fts3Expr *p;
5241 for(p=pExpr; p->pPhrase==0; p=p->pLeft){
5242 if( p->pRight->iDocid==pCsr->iPrevId ){
5243 fts3EvalInvalidatePoslist(p->pRight->pPhrase);
5246 if( p->iDocid==pCsr->iPrevId ){
5247 fts3EvalInvalidatePoslist(p->pPhrase);
5251 break;
5253 case FTSQUERY_OR: {
5254 int bHit1 = fts3EvalTestExpr(pCsr, pExpr->pLeft, pRc);
5255 int bHit2 = fts3EvalTestExpr(pCsr, pExpr->pRight, pRc);
5256 bHit = bHit1 || bHit2;
5257 break;
5260 case FTSQUERY_NOT:
5261 bHit = (
5262 fts3EvalTestExpr(pCsr, pExpr->pLeft, pRc)
5263 && !fts3EvalTestExpr(pCsr, pExpr->pRight, pRc)
5265 break;
5267 default: {
5268 #ifndef SQLITE_DISABLE_FTS4_DEFERRED
5269 if( pCsr->pDeferred
5270 && (pExpr->iDocid==pCsr->iPrevId || pExpr->bDeferred)
5272 Fts3Phrase *pPhrase = pExpr->pPhrase;
5273 assert( pExpr->bDeferred || pPhrase->doclist.bFreeList==0 );
5274 if( pExpr->bDeferred ){
5275 fts3EvalInvalidatePoslist(pPhrase);
5277 *pRc = fts3EvalDeferredPhrase(pCsr, pPhrase);
5278 bHit = (pPhrase->doclist.pList!=0);
5279 pExpr->iDocid = pCsr->iPrevId;
5280 }else
5281 #endif
5283 bHit = (pExpr->bEof==0 && pExpr->iDocid==pCsr->iPrevId);
5285 break;
5289 return bHit;
5293 ** This function is called as the second part of each xNext operation when
5294 ** iterating through the results of a full-text query. At this point the
5295 ** cursor points to a row that matches the query expression, with the
5296 ** following caveats:
5298 ** * Up until this point, "NEAR" operators in the expression have been
5299 ** treated as "AND".
5301 ** * Deferred tokens have not yet been considered.
5303 ** If *pRc is not SQLITE_OK when this function is called, it immediately
5304 ** returns 0. Otherwise, it tests whether or not after considering NEAR
5305 ** operators and deferred tokens the current row is still a match for the
5306 ** expression. It returns 1 if both of the following are true:
5308 ** 1. *pRc is SQLITE_OK when this function returns, and
5310 ** 2. After scanning the current FTS table row for the deferred tokens,
5311 ** it is determined that the row does *not* match the query.
5313 ** Or, if no error occurs and it seems the current row does match the FTS
5314 ** query, return 0.
5316 static int fts3EvalTestDeferredAndNear(Fts3Cursor *pCsr, int *pRc){
5317 int rc = *pRc;
5318 int bMiss = 0;
5319 if( rc==SQLITE_OK ){
5321 /* If there are one or more deferred tokens, load the current row into
5322 ** memory and scan it to determine the position list for each deferred
5323 ** token. Then, see if this row is really a match, considering deferred
5324 ** tokens and NEAR operators (neither of which were taken into account
5325 ** earlier, by fts3EvalNextRow()).
5327 if( pCsr->pDeferred ){
5328 rc = fts3CursorSeek(0, pCsr);
5329 if( rc==SQLITE_OK ){
5330 rc = sqlite3Fts3CacheDeferredDoclists(pCsr);
5333 bMiss = (0==fts3EvalTestExpr(pCsr, pCsr->pExpr, &rc));
5335 /* Free the position-lists accumulated for each deferred token above. */
5336 sqlite3Fts3FreeDeferredDoclists(pCsr);
5337 *pRc = rc;
5339 return (rc==SQLITE_OK && bMiss);
5343 ** Advance to the next document that matches the FTS expression in
5344 ** Fts3Cursor.pExpr.
5346 static int fts3EvalNext(Fts3Cursor *pCsr){
5347 int rc = SQLITE_OK; /* Return Code */
5348 Fts3Expr *pExpr = pCsr->pExpr;
5349 assert( pCsr->isEof==0 );
5350 if( pExpr==0 ){
5351 pCsr->isEof = 1;
5352 }else{
5353 do {
5354 if( pCsr->isRequireSeek==0 ){
5355 sqlite3_reset(pCsr->pStmt);
5357 assert( sqlite3_data_count(pCsr->pStmt)==0 );
5358 fts3EvalNextRow(pCsr, pExpr, &rc);
5359 pCsr->isEof = pExpr->bEof;
5360 pCsr->isRequireSeek = 1;
5361 pCsr->isMatchinfoNeeded = 1;
5362 pCsr->iPrevId = pExpr->iDocid;
5363 }while( pCsr->isEof==0 && fts3EvalTestDeferredAndNear(pCsr, &rc) );
5366 /* Check if the cursor is past the end of the docid range specified
5367 ** by Fts3Cursor.iMinDocid/iMaxDocid. If so, set the EOF flag. */
5368 if( rc==SQLITE_OK && (
5369 (pCsr->bDesc==0 && pCsr->iPrevId>pCsr->iMaxDocid)
5370 || (pCsr->bDesc!=0 && pCsr->iPrevId<pCsr->iMinDocid)
5372 pCsr->isEof = 1;
5375 return rc;
5379 ** Restart interation for expression pExpr so that the next call to
5380 ** fts3EvalNext() visits the first row. Do not allow incremental
5381 ** loading or merging of phrase doclists for this iteration.
5383 ** If *pRc is other than SQLITE_OK when this function is called, it is
5384 ** a no-op. If an error occurs within this function, *pRc is set to an
5385 ** SQLite error code before returning.
5387 static void fts3EvalRestart(
5388 Fts3Cursor *pCsr,
5389 Fts3Expr *pExpr,
5390 int *pRc
5392 if( pExpr && *pRc==SQLITE_OK ){
5393 Fts3Phrase *pPhrase = pExpr->pPhrase;
5395 if( pPhrase ){
5396 fts3EvalInvalidatePoslist(pPhrase);
5397 if( pPhrase->bIncr ){
5398 int i;
5399 for(i=0; i<pPhrase->nToken; i++){
5400 Fts3PhraseToken *pToken = &pPhrase->aToken[i];
5401 assert( pToken->pDeferred==0 );
5402 if( pToken->pSegcsr ){
5403 sqlite3Fts3MsrIncrRestart(pToken->pSegcsr);
5406 *pRc = fts3EvalPhraseStart(pCsr, 0, pPhrase);
5408 pPhrase->doclist.pNextDocid = 0;
5409 pPhrase->doclist.iDocid = 0;
5412 pExpr->iDocid = 0;
5413 pExpr->bEof = 0;
5414 pExpr->bStart = 0;
5416 fts3EvalRestart(pCsr, pExpr->pLeft, pRc);
5417 fts3EvalRestart(pCsr, pExpr->pRight, pRc);
5422 ** After allocating the Fts3Expr.aMI[] array for each phrase in the
5423 ** expression rooted at pExpr, the cursor iterates through all rows matched
5424 ** by pExpr, calling this function for each row. This function increments
5425 ** the values in Fts3Expr.aMI[] according to the position-list currently
5426 ** found in Fts3Expr.pPhrase->doclist.pList for each of the phrase
5427 ** expression nodes.
5429 static void fts3EvalUpdateCounts(Fts3Expr *pExpr){
5430 if( pExpr ){
5431 Fts3Phrase *pPhrase = pExpr->pPhrase;
5432 if( pPhrase && pPhrase->doclist.pList ){
5433 int iCol = 0;
5434 char *p = pPhrase->doclist.pList;
5436 assert( *p );
5437 while( 1 ){
5438 u8 c = 0;
5439 int iCnt = 0;
5440 while( 0xFE & (*p | c) ){
5441 if( (c&0x80)==0 ) iCnt++;
5442 c = *p++ & 0x80;
5445 /* aMI[iCol*3 + 1] = Number of occurrences
5446 ** aMI[iCol*3 + 2] = Number of rows containing at least one instance
5448 pExpr->aMI[iCol*3 + 1] += iCnt;
5449 pExpr->aMI[iCol*3 + 2] += (iCnt>0);
5450 if( *p==0x00 ) break;
5451 p++;
5452 p += fts3GetVarint32(p, &iCol);
5456 fts3EvalUpdateCounts(pExpr->pLeft);
5457 fts3EvalUpdateCounts(pExpr->pRight);
5462 ** Expression pExpr must be of type FTSQUERY_PHRASE.
5464 ** If it is not already allocated and populated, this function allocates and
5465 ** populates the Fts3Expr.aMI[] array for expression pExpr. If pExpr is part
5466 ** of a NEAR expression, then it also allocates and populates the same array
5467 ** for all other phrases that are part of the NEAR expression.
5469 ** SQLITE_OK is returned if the aMI[] array is successfully allocated and
5470 ** populated. Otherwise, if an error occurs, an SQLite error code is returned.
5472 static int fts3EvalGatherStats(
5473 Fts3Cursor *pCsr, /* Cursor object */
5474 Fts3Expr *pExpr /* FTSQUERY_PHRASE expression */
5476 int rc = SQLITE_OK; /* Return code */
5478 assert( pExpr->eType==FTSQUERY_PHRASE );
5479 if( pExpr->aMI==0 ){
5480 Fts3Table *pTab = (Fts3Table *)pCsr->base.pVtab;
5481 Fts3Expr *pRoot; /* Root of NEAR expression */
5482 Fts3Expr *p; /* Iterator used for several purposes */
5484 sqlite3_int64 iPrevId = pCsr->iPrevId;
5485 sqlite3_int64 iDocid;
5486 u8 bEof;
5488 /* Find the root of the NEAR expression */
5489 pRoot = pExpr;
5490 while( pRoot->pParent && pRoot->pParent->eType==FTSQUERY_NEAR ){
5491 pRoot = pRoot->pParent;
5493 iDocid = pRoot->iDocid;
5494 bEof = pRoot->bEof;
5495 assert( pRoot->bStart );
5497 /* Allocate space for the aMSI[] array of each FTSQUERY_PHRASE node */
5498 for(p=pRoot; p; p=p->pLeft){
5499 Fts3Expr *pE = (p->eType==FTSQUERY_PHRASE?p:p->pRight);
5500 assert( pE->aMI==0 );
5501 pE->aMI = (u32 *)sqlite3_malloc(pTab->nColumn * 3 * sizeof(u32));
5502 if( !pE->aMI ) return SQLITE_NOMEM;
5503 memset(pE->aMI, 0, pTab->nColumn * 3 * sizeof(u32));
5506 fts3EvalRestart(pCsr, pRoot, &rc);
5508 while( pCsr->isEof==0 && rc==SQLITE_OK ){
5510 do {
5511 /* Ensure the %_content statement is reset. */
5512 if( pCsr->isRequireSeek==0 ) sqlite3_reset(pCsr->pStmt);
5513 assert( sqlite3_data_count(pCsr->pStmt)==0 );
5515 /* Advance to the next document */
5516 fts3EvalNextRow(pCsr, pRoot, &rc);
5517 pCsr->isEof = pRoot->bEof;
5518 pCsr->isRequireSeek = 1;
5519 pCsr->isMatchinfoNeeded = 1;
5520 pCsr->iPrevId = pRoot->iDocid;
5521 }while( pCsr->isEof==0
5522 && pRoot->eType==FTSQUERY_NEAR
5523 && fts3EvalTestDeferredAndNear(pCsr, &rc)
5526 if( rc==SQLITE_OK && pCsr->isEof==0 ){
5527 fts3EvalUpdateCounts(pRoot);
5531 pCsr->isEof = 0;
5532 pCsr->iPrevId = iPrevId;
5534 if( bEof ){
5535 pRoot->bEof = bEof;
5536 }else{
5537 /* Caution: pRoot may iterate through docids in ascending or descending
5538 ** order. For this reason, even though it seems more defensive, the
5539 ** do loop can not be written:
5541 ** do {...} while( pRoot->iDocid<iDocid && rc==SQLITE_OK );
5543 fts3EvalRestart(pCsr, pRoot, &rc);
5544 do {
5545 fts3EvalNextRow(pCsr, pRoot, &rc);
5546 assert( pRoot->bEof==0 );
5547 }while( pRoot->iDocid!=iDocid && rc==SQLITE_OK );
5548 fts3EvalTestDeferredAndNear(pCsr, &rc);
5551 return rc;
5555 ** This function is used by the matchinfo() module to query a phrase
5556 ** expression node for the following information:
5558 ** 1. The total number of occurrences of the phrase in each column of
5559 ** the FTS table (considering all rows), and
5561 ** 2. For each column, the number of rows in the table for which the
5562 ** column contains at least one instance of the phrase.
5564 ** If no error occurs, SQLITE_OK is returned and the values for each column
5565 ** written into the array aiOut as follows:
5567 ** aiOut[iCol*3 + 1] = Number of occurrences
5568 ** aiOut[iCol*3 + 2] = Number of rows containing at least one instance
5570 ** Caveats:
5572 ** * If a phrase consists entirely of deferred tokens, then all output
5573 ** values are set to the number of documents in the table. In other
5574 ** words we assume that very common tokens occur exactly once in each
5575 ** column of each row of the table.
5577 ** * If a phrase contains some deferred tokens (and some non-deferred
5578 ** tokens), count the potential occurrence identified by considering
5579 ** the non-deferred tokens instead of actual phrase occurrences.
5581 ** * If the phrase is part of a NEAR expression, then only phrase instances
5582 ** that meet the NEAR constraint are included in the counts.
5584 int sqlite3Fts3EvalPhraseStats(
5585 Fts3Cursor *pCsr, /* FTS cursor handle */
5586 Fts3Expr *pExpr, /* Phrase expression */
5587 u32 *aiOut /* Array to write results into (see above) */
5589 Fts3Table *pTab = (Fts3Table *)pCsr->base.pVtab;
5590 int rc = SQLITE_OK;
5591 int iCol;
5593 if( pExpr->bDeferred && pExpr->pParent->eType!=FTSQUERY_NEAR ){
5594 assert( pCsr->nDoc>0 );
5595 for(iCol=0; iCol<pTab->nColumn; iCol++){
5596 aiOut[iCol*3 + 1] = (u32)pCsr->nDoc;
5597 aiOut[iCol*3 + 2] = (u32)pCsr->nDoc;
5599 }else{
5600 rc = fts3EvalGatherStats(pCsr, pExpr);
5601 if( rc==SQLITE_OK ){
5602 assert( pExpr->aMI );
5603 for(iCol=0; iCol<pTab->nColumn; iCol++){
5604 aiOut[iCol*3 + 1] = pExpr->aMI[iCol*3 + 1];
5605 aiOut[iCol*3 + 2] = pExpr->aMI[iCol*3 + 2];
5610 return rc;
5614 ** The expression pExpr passed as the second argument to this function
5615 ** must be of type FTSQUERY_PHRASE.
5617 ** The returned value is either NULL or a pointer to a buffer containing
5618 ** a position-list indicating the occurrences of the phrase in column iCol
5619 ** of the current row.
5621 ** More specifically, the returned buffer contains 1 varint for each
5622 ** occurrence of the phrase in the column, stored using the normal (delta+2)
5623 ** compression and is terminated by either an 0x01 or 0x00 byte. For example,
5624 ** if the requested column contains "a b X c d X X" and the position-list
5625 ** for 'X' is requested, the buffer returned may contain:
5627 ** 0x04 0x05 0x03 0x01 or 0x04 0x05 0x03 0x00
5629 ** This function works regardless of whether or not the phrase is deferred,
5630 ** incremental, or neither.
5632 int sqlite3Fts3EvalPhrasePoslist(
5633 Fts3Cursor *pCsr, /* FTS3 cursor object */
5634 Fts3Expr *pExpr, /* Phrase to return doclist for */
5635 int iCol, /* Column to return position list for */
5636 char **ppOut /* OUT: Pointer to position list */
5638 Fts3Phrase *pPhrase = pExpr->pPhrase;
5639 Fts3Table *pTab = (Fts3Table *)pCsr->base.pVtab;
5640 char *pIter;
5641 int iThis;
5642 sqlite3_int64 iDocid;
5644 /* If this phrase is applies specifically to some column other than
5645 ** column iCol, return a NULL pointer. */
5646 *ppOut = 0;
5647 assert( iCol>=0 && iCol<pTab->nColumn );
5648 if( (pPhrase->iColumn<pTab->nColumn && pPhrase->iColumn!=iCol) ){
5649 return SQLITE_OK;
5652 iDocid = pExpr->iDocid;
5653 pIter = pPhrase->doclist.pList;
5654 if( iDocid!=pCsr->iPrevId || pExpr->bEof ){
5655 int bDescDoclist = pTab->bDescIdx; /* For DOCID_CMP macro */
5656 int iMul; /* +1 if csr dir matches index dir, else -1 */
5657 int bOr = 0;
5658 u8 bEof = 0;
5659 u8 bTreeEof = 0;
5660 Fts3Expr *p; /* Used to iterate from pExpr to root */
5661 Fts3Expr *pNear; /* Most senior NEAR ancestor (or pExpr) */
5663 /* Check if this phrase descends from an OR expression node. If not,
5664 ** return NULL. Otherwise, the entry that corresponds to docid
5665 ** pCsr->iPrevId may lie earlier in the doclist buffer. Or, if the
5666 ** tree that the node is part of has been marked as EOF, but the node
5667 ** itself is not EOF, then it may point to an earlier entry. */
5668 pNear = pExpr;
5669 for(p=pExpr->pParent; p; p=p->pParent){
5670 if( p->eType==FTSQUERY_OR ) bOr = 1;
5671 if( p->eType==FTSQUERY_NEAR ) pNear = p;
5672 if( p->bEof ) bTreeEof = 1;
5674 if( bOr==0 ) return SQLITE_OK;
5676 /* This is the descendent of an OR node. In this case we cannot use
5677 ** an incremental phrase. Load the entire doclist for the phrase
5678 ** into memory in this case. */
5679 if( pPhrase->bIncr ){
5680 int rc = SQLITE_OK;
5681 int bEofSave = pExpr->bEof;
5682 fts3EvalRestart(pCsr, pExpr, &rc);
5683 while( rc==SQLITE_OK && !pExpr->bEof ){
5684 fts3EvalNextRow(pCsr, pExpr, &rc);
5685 if( bEofSave==0 && pExpr->iDocid==iDocid ) break;
5687 pIter = pPhrase->doclist.pList;
5688 assert( rc!=SQLITE_OK || pPhrase->bIncr==0 );
5689 if( rc!=SQLITE_OK ) return rc;
5692 iMul = ((pCsr->bDesc==bDescDoclist) ? 1 : -1);
5693 while( bTreeEof==1
5694 && pNear->bEof==0
5695 && (DOCID_CMP(pNear->iDocid, pCsr->iPrevId) * iMul)<0
5697 int rc = SQLITE_OK;
5698 fts3EvalNextRow(pCsr, pExpr, &rc);
5699 if( rc!=SQLITE_OK ) return rc;
5700 iDocid = pExpr->iDocid;
5701 pIter = pPhrase->doclist.pList;
5704 bEof = (pPhrase->doclist.nAll==0);
5705 assert( bDescDoclist==0 || bDescDoclist==1 );
5706 assert( pCsr->bDesc==0 || pCsr->bDesc==1 );
5708 if( bEof==0 ){
5709 if( pCsr->bDesc==bDescDoclist ){
5710 int dummy;
5711 if( pNear->bEof ){
5712 /* This expression is already at EOF. So position it to point to the
5713 ** last entry in the doclist at pPhrase->doclist.aAll[]. Variable
5714 ** iDocid is already set for this entry, so all that is required is
5715 ** to set pIter to point to the first byte of the last position-list
5716 ** in the doclist.
5718 ** It would also be correct to set pIter and iDocid to zero. In
5719 ** this case, the first call to sqltie3Fts4DoclistPrev() below
5720 ** would also move the iterator to point to the last entry in the
5721 ** doclist. However, this is expensive, as to do so it has to
5722 ** iterate through the entire doclist from start to finish (since
5723 ** it does not know the docid for the last entry). */
5724 pIter = &pPhrase->doclist.aAll[pPhrase->doclist.nAll-1];
5725 fts3ReversePoslist(pPhrase->doclist.aAll, &pIter);
5727 while( (pIter==0 || DOCID_CMP(iDocid, pCsr->iPrevId)>0 ) && bEof==0 ){
5728 sqlite3Fts3DoclistPrev(
5729 bDescDoclist, pPhrase->doclist.aAll, pPhrase->doclist.nAll,
5730 &pIter, &iDocid, &dummy, &bEof
5733 }else{
5734 if( pNear->bEof ){
5735 pIter = 0;
5736 iDocid = 0;
5738 while( (pIter==0 || DOCID_CMP(iDocid, pCsr->iPrevId)<0 ) && bEof==0 ){
5739 sqlite3Fts3DoclistNext(
5740 bDescDoclist, pPhrase->doclist.aAll, pPhrase->doclist.nAll,
5741 &pIter, &iDocid, &bEof
5747 if( bEof || iDocid!=pCsr->iPrevId ) pIter = 0;
5749 if( pIter==0 ) return SQLITE_OK;
5751 if( *pIter==0x01 ){
5752 pIter++;
5753 pIter += fts3GetVarint32(pIter, &iThis);
5754 }else{
5755 iThis = 0;
5757 while( iThis<iCol ){
5758 fts3ColumnlistCopy(0, &pIter);
5759 if( *pIter==0x00 ) return 0;
5760 pIter++;
5761 pIter += fts3GetVarint32(pIter, &iThis);
5764 *ppOut = ((iCol==iThis)?pIter:0);
5765 return SQLITE_OK;
5769 ** Free all components of the Fts3Phrase structure that were allocated by
5770 ** the eval module. Specifically, this means to free:
5772 ** * the contents of pPhrase->doclist, and
5773 ** * any Fts3MultiSegReader objects held by phrase tokens.
5775 void sqlite3Fts3EvalPhraseCleanup(Fts3Phrase *pPhrase){
5776 if( pPhrase ){
5777 int i;
5778 sqlite3_free(pPhrase->doclist.aAll);
5779 fts3EvalInvalidatePoslist(pPhrase);
5780 memset(&pPhrase->doclist, 0, sizeof(Fts3Doclist));
5781 for(i=0; i<pPhrase->nToken; i++){
5782 fts3SegReaderCursorFree(pPhrase->aToken[i].pSegcsr);
5783 pPhrase->aToken[i].pSegcsr = 0;
5790 ** Return SQLITE_CORRUPT_VTAB.
5792 #ifdef SQLITE_DEBUG
5793 int sqlite3Fts3Corrupt(){
5794 return SQLITE_CORRUPT_VTAB;
5796 #endif
5798 #if !SQLITE_CORE
5800 ** Initialize API pointer table, if required.
5802 #ifdef _WIN32
5803 __declspec(dllexport)
5804 #endif
5805 int sqlite3_fts3_init(
5806 sqlite3 *db,
5807 char **pzErrMsg,
5808 const sqlite3_api_routines *pApi
5810 SQLITE_EXTENSION_INIT2(pApi)
5811 return sqlite3Fts3Init(db);
5813 #endif
5815 #endif