+#if !defined(SQLITE_CORE) || defined(SQLITE_ENABLE_FTS3)
+
+#if defined(SQLITE_ENABLE_FTS3) && !defined(SQLITE_CORE)
+# define SQLITE_CORE 1
+#endif
+
+
+/************** Include fts3_hash.h in the middle of fts3.c ******************/
+/************** Begin file fts3_hash.h ***************************************/
+/*
+** 2001 September 22
+**
+** The author disclaims copyright to this source code. In place of
+** a legal notice, here is a blessing:
+**
+** May you do good and not evil.
+** May you find forgiveness for yourself and forgive others.
+** May you share freely, never taking more than you give.
+**
+*************************************************************************
+** This is the header file for the generic hash-table implemenation
+** used in SQLite. We've modified it slightly to serve as a standalone
+** hash table implementation for the full-text indexing module.
+**
+*/
+#ifndef _FTS3_HASH_H_
+#define _FTS3_HASH_H_
+
+/* Forward declarations of structures. */
+typedef struct fts3Hash fts3Hash;
+typedef struct fts3HashElem fts3HashElem;
+
+/* A complete hash table is an instance of the following structure.
+** The internals of this structure are intended to be opaque -- client
+** code should not attempt to access or modify the fields of this structure
+** directly. Change this structure only by using the routines below.
+** However, many of the "procedures" and "functions" for modifying and
+** accessing this structure are really macros, so we can't really make
+** this structure opaque.
+*/
+struct fts3Hash {
+ char keyClass; /* HASH_INT, _POINTER, _STRING, _BINARY */
+ char copyKey; /* True if copy of key made on insert */
+ int count; /* Number of entries in this table */
+ fts3HashElem *first; /* The first element of the array */
+ int htsize; /* Number of buckets in the hash table */
+ struct _fts3ht { /* the hash table */
+ int count; /* Number of entries with this hash */
+ fts3HashElem *chain; /* Pointer to first entry with this hash */
+ } *ht;
+};
+
+/* Each element in the hash table is an instance of the following
+** structure. All elements are stored on a single doubly-linked list.
+**
+** Again, this structure is intended to be opaque, but it can't really
+** be opaque because it is used by macros.
+*/
+struct fts3HashElem {
+ fts3HashElem *next, *prev; /* Next and previous elements in the table */
+ void *data; /* Data associated with this element */
+ void *pKey; int nKey; /* Key associated with this element */
+};
+
+/*
+** There are 2 different modes of operation for a hash table:
+**
+** FTS3_HASH_STRING pKey points to a string that is nKey bytes long
+** (including the null-terminator, if any). Case
+** is respected in comparisons.
+**
+** FTS3_HASH_BINARY pKey points to binary data nKey bytes long.
+** memcmp() is used to compare keys.
+**
+** A copy of the key is made if the copyKey parameter to fts3HashInit is 1.
+*/
+#define FTS3_HASH_STRING 1
+#define FTS3_HASH_BINARY 2
+
+/*
+** Access routines. To delete, insert a NULL pointer.
+*/
+SQLITE_PRIVATE void sqlite3Fts3HashInit(fts3Hash*, int keytype, int copyKey);
+SQLITE_PRIVATE void *sqlite3Fts3HashInsert(fts3Hash*, const void *pKey, int nKey, void *pData);
+SQLITE_PRIVATE void *sqlite3Fts3HashFind(const fts3Hash*, const void *pKey, int nKey);
+SQLITE_PRIVATE void sqlite3Fts3HashClear(fts3Hash*);
+
+/*
+** Shorthand for the functions above
+*/
+#define fts3HashInit sqlite3Fts3HashInit
+#define fts3HashInsert sqlite3Fts3HashInsert
+#define fts3HashFind sqlite3Fts3HashFind
+#define fts3HashClear sqlite3Fts3HashClear
+
+/*
+** Macros for looping over all elements of a hash table. The idiom is
+** like this:
+**
+** fts3Hash h;
+** fts3HashElem *p;
+** ...
+** for(p=fts3HashFirst(&h); p; p=fts3HashNext(p)){
+** SomeStructure *pData = fts3HashData(p);
+** // do something with pData
+** }
+*/
+#define fts3HashFirst(H) ((H)->first)
+#define fts3HashNext(E) ((E)->next)
+#define fts3HashData(E) ((E)->data)
+#define fts3HashKey(E) ((E)->pKey)
+#define fts3HashKeysize(E) ((E)->nKey)
+
+/*
+** Number of entries in a hash table
+*/
+#define fts3HashCount(H) ((H)->count)
+
+#endif /* _FTS3_HASH_H_ */
+
+/************** End of fts3_hash.h *******************************************/
+/************** Continuing where we left off in fts3.c ***********************/
+/************** Include fts3_tokenizer.h in the middle of fts3.c *************/
+/************** Begin file fts3_tokenizer.h **********************************/
+/*
+** 2006 July 10
+**
+** The author disclaims copyright to this source code.
+**
+*************************************************************************
+** Defines the interface to tokenizers used by fulltext-search. There
+** are three basic components:
+**
+** sqlite3_tokenizer_module is a singleton defining the tokenizer
+** interface functions. This is essentially the class structure for
+** tokenizers.
+**
+** sqlite3_tokenizer is used to define a particular tokenizer, perhaps
+** including customization information defined at creation time.
+**
+** sqlite3_tokenizer_cursor is generated by a tokenizer to generate
+** tokens from a particular input.
+*/
+#ifndef _FTS3_TOKENIZER_H_
+#define _FTS3_TOKENIZER_H_
+
+/* TODO(shess) Only used for SQLITE_OK and SQLITE_DONE at this time.
+** If tokenizers are to be allowed to call sqlite3_*() functions, then
+** we will need a way to register the API consistently.
+*/
+
+/*
+** Structures used by the tokenizer interface. When a new tokenizer
+** implementation is registered, the caller provides a pointer to
+** an sqlite3_tokenizer_module containing pointers to the callback
+** functions that make up an implementation.
+**
+** When an fts3 table is created, it passes any arguments passed to
+** the tokenizer clause of the CREATE VIRTUAL TABLE statement to the
+** sqlite3_tokenizer_module.xCreate() function of the requested tokenizer
+** implementation. The xCreate() function in turn returns an
+** sqlite3_tokenizer structure representing the specific tokenizer to
+** be used for the fts3 table (customized by the tokenizer clause arguments).
+**
+** To tokenize an input buffer, the sqlite3_tokenizer_module.xOpen()
+** method is called. It returns an sqlite3_tokenizer_cursor object
+** that may be used to tokenize a specific input buffer based on
+** the tokenization rules supplied by a specific sqlite3_tokenizer
+** object.
+*/
+typedef struct sqlite3_tokenizer_module sqlite3_tokenizer_module;
+typedef struct sqlite3_tokenizer sqlite3_tokenizer;
+typedef struct sqlite3_tokenizer_cursor sqlite3_tokenizer_cursor;
+
+struct sqlite3_tokenizer_module {
+
+ /*
+ ** Structure version. Should always be set to 0.
+ */
+ int iVersion;
+
+ /*
+ ** Create a new tokenizer. The values in the argv[] array are the
+ ** arguments passed to the "tokenizer" clause of the CREATE VIRTUAL
+ ** TABLE statement that created the fts3 table. For example, if
+ ** the following SQL is executed:
+ **
+ ** CREATE .. USING fts3( ... , tokenizer <tokenizer-name> arg1 arg2)
+ **
+ ** then argc is set to 2, and the argv[] array contains pointers
+ ** to the strings "arg1" and "arg2".
+ **
+ ** This method should return either SQLITE_OK (0), or an SQLite error
+ ** code. If SQLITE_OK is returned, then *ppTokenizer should be set
+ ** to point at the newly created tokenizer structure. The generic
+ ** sqlite3_tokenizer.pModule variable should not be initialised by
+ ** this callback. The caller will do so.
+ */
+ int (*xCreate)(
+ int argc, /* Size of argv array */
+ const char *const*argv, /* Tokenizer argument strings */
+ sqlite3_tokenizer **ppTokenizer /* OUT: Created tokenizer */
+ );
+
+ /*
+ ** Destroy an existing tokenizer. The fts3 module calls this method
+ ** exactly once for each successful call to xCreate().
+ */
+ int (*xDestroy)(sqlite3_tokenizer *pTokenizer);
+
+ /*
+ ** Create a tokenizer cursor to tokenize an input buffer. The caller
+ ** is responsible for ensuring that the input buffer remains valid
+ ** until the cursor is closed (using the xClose() method).
+ */
+ int (*xOpen)(
+ sqlite3_tokenizer *pTokenizer, /* Tokenizer object */
+ const char *pInput, int nBytes, /* Input buffer */
+ sqlite3_tokenizer_cursor **ppCursor /* OUT: Created tokenizer cursor */
+ );
+
+ /*
+ ** Destroy an existing tokenizer cursor. The fts3 module calls this
+ ** method exactly once for each successful call to xOpen().
+ */
+ int (*xClose)(sqlite3_tokenizer_cursor *pCursor);
+
+ /*
+ ** Retrieve the next token from the tokenizer cursor pCursor. This
+ ** method should either return SQLITE_OK and set the values of the
+ ** "OUT" variables identified below, or SQLITE_DONE to indicate that
+ ** the end of the buffer has been reached, or an SQLite error code.
+ **
+ ** *ppToken should be set to point at a buffer containing the
+ ** normalized version of the token (i.e. after any case-folding and/or
+ ** stemming has been performed). *pnBytes should be set to the length
+ ** of this buffer in bytes. The input text that generated the token is
+ ** identified by the byte offsets returned in *piStartOffset and
+ ** *piEndOffset.
+ **
+ ** The buffer *ppToken is set to point at is managed by the tokenizer
+ ** implementation. It is only required to be valid until the next call
+ ** to xNext() or xClose().
+ */
+ /* TODO(shess) current implementation requires pInput to be
+ ** nul-terminated. This should either be fixed, or pInput/nBytes
+ ** should be converted to zInput.
+ */
+ int (*xNext)(
+ sqlite3_tokenizer_cursor *pCursor, /* Tokenizer cursor */
+ const char **ppToken, int *pnBytes, /* OUT: Normalized text for token */
+ int *piStartOffset, /* OUT: Byte offset of token in input buffer */
+ int *piEndOffset, /* OUT: Byte offset of end of token in input buffer */
+ int *piPosition /* OUT: Number of tokens returned before this one */
+ );
+};
+
+struct sqlite3_tokenizer {
+ const sqlite3_tokenizer_module *pModule; /* The module for this tokenizer */
+ /* Tokenizer implementations will typically add additional fields */
+};
+
+struct sqlite3_tokenizer_cursor {
+ sqlite3_tokenizer *pTokenizer; /* Tokenizer for this cursor. */
+ /* Tokenizer implementations will typically add additional fields */
+};
+
+#endif /* _FTS3_TOKENIZER_H_ */
+
+/************** End of fts3_tokenizer.h **************************************/
+/************** Continuing where we left off in fts3.c ***********************/
+#ifndef SQLITE_CORE
+ SQLITE_EXTENSION_INIT1
+#endif
+
+
+/* TODO(shess) MAN, this thing needs some refactoring. At minimum, it
+** would be nice to order the file better, perhaps something along the
+** lines of:
+**
+** - utility functions
+** - table setup functions
+** - table update functions
+** - table query functions
+**
+** Put the query functions last because they're likely to reference
+** typedefs or functions from the table update section.
+*/
+
+#if 0
+# define FTSTRACE(A) printf A; fflush(stdout)
+#else
+# define FTSTRACE(A)
+#endif
+
+/*
+** Default span for NEAR operators.
+*/
+#define SQLITE_FTS3_DEFAULT_NEAR_PARAM 10
+
+/* It is not safe to call isspace(), tolower(), or isalnum() on
+** hi-bit-set characters. This is the same solution used in the
+** tokenizer.
+*/
+/* TODO(shess) The snippet-generation code should be using the
+** tokenizer-generated tokens rather than doing its own local
+** tokenization.
+*/
+/* TODO(shess) Is __isascii() a portable version of (c&0x80)==0? */
+static int safe_isspace(char c){
+ return (c&0x80)==0 ? isspace(c) : 0;
+}
+static int safe_tolower(char c){
+ return (c&0x80)==0 ? tolower(c) : c;
+}
+static int safe_isalnum(char c){
+ return (c&0x80)==0 ? isalnum(c) : 0;
+}
+
+typedef enum DocListType {
+ DL_DOCIDS, /* docids only */
+ DL_POSITIONS, /* docids + positions */
+ DL_POSITIONS_OFFSETS /* docids + positions + offsets */
+} DocListType;
+
+/*
+** By default, only positions and not offsets are stored in the doclists.
+** To change this so that offsets are stored too, compile with
+**
+** -DDL_DEFAULT=DL_POSITIONS_OFFSETS
+**
+** If DL_DEFAULT is set to DL_DOCIDS, your table can only be inserted
+** into (no deletes or updates).
+*/
+#ifndef DL_DEFAULT
+# define DL_DEFAULT DL_POSITIONS
+#endif
+
+enum {
+ POS_END = 0, /* end of this position list */
+ POS_COLUMN, /* followed by new column number */
+ POS_BASE
+};
+
+/* MERGE_COUNT controls how often we merge segments (see comment at
+** top of file).
+*/
+#define MERGE_COUNT 16
+
+/* utility functions */
+
+/* CLEAR() and SCRAMBLE() abstract memset() on a pointer to a single
+** record to prevent errors of the form:
+**
+** my_function(SomeType *b){
+** memset(b, '\0', sizeof(b)); // sizeof(b)!=sizeof(*b)
+** }
+*/
+/* TODO(shess) Obvious candidates for a header file. */
+#define CLEAR(b) memset(b, '\0', sizeof(*(b)))
+
+#ifndef NDEBUG
+# define SCRAMBLE(b) memset(b, 0x55, sizeof(*(b)))
+#else
+# define SCRAMBLE(b)
+#endif
+
+/* We may need up to VARINT_MAX bytes to store an encoded 64-bit integer. */
+#define VARINT_MAX 10
+
+/* Write a 64-bit variable-length integer to memory starting at p[0].
+ * The length of data written will be between 1 and VARINT_MAX bytes.
+ * The number of bytes written is returned. */
+static int fts3PutVarint(char *p, sqlite_int64 v){
+ unsigned char *q = (unsigned char *) p;
+ sqlite_uint64 vu = v;
+ do{
+ *q++ = (unsigned char) ((vu & 0x7f) | 0x80);
+ vu >>= 7;
+ }while( vu!=0 );
+ q[-1] &= 0x7f; /* turn off high bit in final byte */
+ assert( q - (unsigned char *)p <= VARINT_MAX );
+ return (int) (q - (unsigned char *)p);
+}
+
+/* Read a 64-bit variable-length integer from memory starting at p[0].
+ * Return the number of bytes read, or 0 on error.
+ * The value is stored in *v. */
+static int fts3GetVarint(const char *p, sqlite_int64 *v){
+ const unsigned char *q = (const unsigned char *) p;
+ sqlite_uint64 x = 0, y = 1;
+ while( (*q & 0x80) == 0x80 ){
+ x += y * (*q++ & 0x7f);
+ y <<= 7;
+ if( q - (unsigned char *)p >= VARINT_MAX ){ /* bad data */
+ assert( 0 );
+ return 0;
+ }
+ }
+ x += y * (*q++);
+ *v = (sqlite_int64) x;
+ return (int) (q - (unsigned char *)p);
+}
+
+static int fts3GetVarint32(const char *p, int *pi){
+ sqlite_int64 i;
+ int ret = fts3GetVarint(p, &i);
+ *pi = (int) i;
+ assert( *pi==i );
+ return ret;
+}
+
+/*******************************************************************/
+/* DataBuffer is used to collect data into a buffer in piecemeal
+** fashion. It implements the usual distinction between amount of
+** data currently stored (nData) and buffer capacity (nCapacity).
+**
+** dataBufferInit - create a buffer with given initial capacity.
+** dataBufferReset - forget buffer's data, retaining capacity.
+** dataBufferDestroy - free buffer's data.
+** dataBufferSwap - swap contents of two buffers.
+** dataBufferExpand - expand capacity without adding data.
+** dataBufferAppend - append data.
+** dataBufferAppend2 - append two pieces of data at once.
+** dataBufferReplace - replace buffer's data.
+*/
+typedef struct DataBuffer {
+ char *pData; /* Pointer to malloc'ed buffer. */
+ int nCapacity; /* Size of pData buffer. */
+ int nData; /* End of data loaded into pData. */
+} DataBuffer;
+
+static void dataBufferInit(DataBuffer *pBuffer, int nCapacity){
+ assert( nCapacity>=0 );
+ pBuffer->nData = 0;
+ pBuffer->nCapacity = nCapacity;
+ pBuffer->pData = nCapacity==0 ? NULL : sqlite3_malloc(nCapacity);
+}
+static void dataBufferReset(DataBuffer *pBuffer){
+ pBuffer->nData = 0;
+}
+static void dataBufferDestroy(DataBuffer *pBuffer){
+ if( pBuffer->pData!=NULL ) sqlite3_free(pBuffer->pData);
+ SCRAMBLE(pBuffer);
+}
+static void dataBufferSwap(DataBuffer *pBuffer1, DataBuffer *pBuffer2){
+ DataBuffer tmp = *pBuffer1;
+ *pBuffer1 = *pBuffer2;
+ *pBuffer2 = tmp;
+}
+static void dataBufferExpand(DataBuffer *pBuffer, int nAddCapacity){
+ assert( nAddCapacity>0 );
+ /* TODO(shess) Consider expanding more aggressively. Note that the
+ ** underlying malloc implementation may take care of such things for
+ ** us already.
+ */
+ if( pBuffer->nData+nAddCapacity>pBuffer->nCapacity ){
+ pBuffer->nCapacity = pBuffer->nData+nAddCapacity;
+ pBuffer->pData = sqlite3_realloc(pBuffer->pData, pBuffer->nCapacity);
+ }
+}
+static void dataBufferAppend(DataBuffer *pBuffer,
+ const char *pSource, int nSource){
+ assert( nSource>0 && pSource!=NULL );
+ dataBufferExpand(pBuffer, nSource);
+ memcpy(pBuffer->pData+pBuffer->nData, pSource, nSource);
+ pBuffer->nData += nSource;
+}
+static void dataBufferAppend2(DataBuffer *pBuffer,
+ const char *pSource1, int nSource1,
+ const char *pSource2, int nSource2){
+ assert( nSource1>0 && pSource1!=NULL );
+ assert( nSource2>0 && pSource2!=NULL );
+ dataBufferExpand(pBuffer, nSource1+nSource2);
+ memcpy(pBuffer->pData+pBuffer->nData, pSource1, nSource1);
+ memcpy(pBuffer->pData+pBuffer->nData+nSource1, pSource2, nSource2);
+ pBuffer->nData += nSource1+nSource2;
+}
+static void dataBufferReplace(DataBuffer *pBuffer,
+ const char *pSource, int nSource){
+ dataBufferReset(pBuffer);
+ dataBufferAppend(pBuffer, pSource, nSource);
+}
+
+/* StringBuffer is a null-terminated version of DataBuffer. */
+typedef struct StringBuffer {
+ DataBuffer b; /* Includes null terminator. */
+} StringBuffer;
+
+static void initStringBuffer(StringBuffer *sb){
+ dataBufferInit(&sb->b, 100);
+ dataBufferReplace(&sb->b, "", 1);
+}
+static int stringBufferLength(StringBuffer *sb){
+ return sb->b.nData-1;
+}
+static char *stringBufferData(StringBuffer *sb){
+ return sb->b.pData;
+}
+static void stringBufferDestroy(StringBuffer *sb){
+ dataBufferDestroy(&sb->b);
+}
+
+static void nappend(StringBuffer *sb, const char *zFrom, int nFrom){
+ assert( sb->b.nData>0 );
+ if( nFrom>0 ){
+ sb->b.nData--;
+ dataBufferAppend2(&sb->b, zFrom, nFrom, "", 1);
+ }
+}
+static void append(StringBuffer *sb, const char *zFrom){
+ nappend(sb, zFrom, strlen(zFrom));
+}
+
+/* Append a list of strings separated by commas. */
+static void appendList(StringBuffer *sb, int nString, char **azString){
+ int i;
+ for(i=0; i<nString; ++i){
+ if( i>0 ) append(sb, ", ");
+ append(sb, azString[i]);
+ }
+}
+
+static int endsInWhiteSpace(StringBuffer *p){
+ return stringBufferLength(p)>0 &&
+ safe_isspace(stringBufferData(p)[stringBufferLength(p)-1]);
+}
+
+/* If the StringBuffer ends in something other than white space, add a
+** single space character to the end.
+*/
+static void appendWhiteSpace(StringBuffer *p){
+ if( stringBufferLength(p)==0 ) return;
+ if( !endsInWhiteSpace(p) ) append(p, " ");
+}
+
+/* Remove white space from the end of the StringBuffer */
+static void trimWhiteSpace(StringBuffer *p){
+ while( endsInWhiteSpace(p) ){
+ p->b.pData[--p->b.nData-1] = '\0';
+ }
+}
+
+/*******************************************************************/
+/* DLReader is used to read document elements from a doclist. The
+** current docid is cached, so dlrDocid() is fast. DLReader does not
+** own the doclist buffer.
+**
+** dlrAtEnd - true if there's no more data to read.
+** dlrDocid - docid of current document.
+** dlrDocData - doclist data for current document (including docid).
+** dlrDocDataBytes - length of same.
+** dlrAllDataBytes - length of all remaining data.
+** dlrPosData - position data for current document.
+** dlrPosDataLen - length of pos data for current document (incl POS_END).
+** dlrStep - step to current document.
+** dlrInit - initial for doclist of given type against given data.
+** dlrDestroy - clean up.
+**
+** Expected usage is something like:
+**
+** DLReader reader;
+** dlrInit(&reader, pData, nData);
+** while( !dlrAtEnd(&reader) ){
+** // calls to dlrDocid() and kin.
+** dlrStep(&reader);
+** }
+** dlrDestroy(&reader);
+*/
+typedef struct DLReader {
+ DocListType iType;
+ const char *pData;
+ int nData;
+
+ sqlite_int64 iDocid;
+ int nElement;
+} DLReader;
+
+static int dlrAtEnd(DLReader *pReader){
+ assert( pReader->nData>=0 );
+ return pReader->nData==0;
+}
+static sqlite_int64 dlrDocid(DLReader *pReader){
+ assert( !dlrAtEnd(pReader) );
+ return pReader->iDocid;
+}
+static const char *dlrDocData(DLReader *pReader){
+ assert( !dlrAtEnd(pReader) );
+ return pReader->pData;
+}
+static int dlrDocDataBytes(DLReader *pReader){
+ assert( !dlrAtEnd(pReader) );
+ return pReader->nElement;
+}
+static int dlrAllDataBytes(DLReader *pReader){
+ assert( !dlrAtEnd(pReader) );
+ return pReader->nData;
+}
+/* TODO(shess) Consider adding a field to track iDocid varint length
+** to make these two functions faster. This might matter (a tiny bit)
+** for queries.
+*/
+static const char *dlrPosData(DLReader *pReader){
+ sqlite_int64 iDummy;
+ int n = fts3GetVarint(pReader->pData, &iDummy);
+ assert( !dlrAtEnd(pReader) );
+ return pReader->pData+n;
+}
+static int dlrPosDataLen(DLReader *pReader){
+ sqlite_int64 iDummy;
+ int n = fts3GetVarint(pReader->pData, &iDummy);
+ assert( !dlrAtEnd(pReader) );
+ return pReader->nElement-n;
+}
+static void dlrStep(DLReader *pReader){
+ assert( !dlrAtEnd(pReader) );
+
+ /* Skip past current doclist element. */
+ assert( pReader->nElement<=pReader->nData );
+ pReader->pData += pReader->nElement;
+ pReader->nData -= pReader->nElement;
+
+ /* If there is more data, read the next doclist element. */
+ if( pReader->nData!=0 ){
+ sqlite_int64 iDocidDelta;
+ int iDummy, n = fts3GetVarint(pReader->pData, &iDocidDelta);
+ pReader->iDocid += iDocidDelta;
+ if( pReader->iType>=DL_POSITIONS ){
+ assert( n<pReader->nData );
+ while( 1 ){
+ n += fts3GetVarint32(pReader->pData+n, &iDummy);
+ assert( n<=pReader->nData );
+ if( iDummy==POS_END ) break;
+ if( iDummy==POS_COLUMN ){
+ n += fts3GetVarint32(pReader->pData+n, &iDummy);
+ assert( n<pReader->nData );
+ }else if( pReader->iType==DL_POSITIONS_OFFSETS ){
+ n += fts3GetVarint32(pReader->pData+n, &iDummy);
+ n += fts3GetVarint32(pReader->pData+n, &iDummy);
+ assert( n<pReader->nData );
+ }
+ }
+ }
+ pReader->nElement = n;
+ assert( pReader->nElement<=pReader->nData );
+ }
+}
+static void dlrInit(DLReader *pReader, DocListType iType,
+ const char *pData, int nData){
+ assert( pData!=NULL && nData!=0 );
+ pReader->iType = iType;
+ pReader->pData = pData;
+ pReader->nData = nData;
+ pReader->nElement = 0;
+ pReader->iDocid = 0;
+
+ /* Load the first element's data. There must be a first element. */
+ dlrStep(pReader);
+}
+static void dlrDestroy(DLReader *pReader){
+ SCRAMBLE(pReader);
+}
+
+#ifndef NDEBUG
+/* Verify that the doclist can be validly decoded. Also returns the
+** last docid found because it is convenient in other assertions for
+** DLWriter.
+*/
+static void docListValidate(DocListType iType, const char *pData, int nData,
+ sqlite_int64 *pLastDocid){
+ sqlite_int64 iPrevDocid = 0;
+ assert( nData>0 );
+ assert( pData!=0 );
+ assert( pData+nData>pData );
+ while( nData!=0 ){
+ sqlite_int64 iDocidDelta;
+ int n = fts3GetVarint(pData, &iDocidDelta);
+ iPrevDocid += iDocidDelta;
+ if( iType>DL_DOCIDS ){
+ int iDummy;
+ while( 1 ){
+ n += fts3GetVarint32(pData+n, &iDummy);
+ if( iDummy==POS_END ) break;
+ if( iDummy==POS_COLUMN ){
+ n += fts3GetVarint32(pData+n, &iDummy);
+ }else if( iType>DL_POSITIONS ){
+ n += fts3GetVarint32(pData+n, &iDummy);
+ n += fts3GetVarint32(pData+n, &iDummy);
+ }
+ assert( n<=nData );
+ }
+ }
+ assert( n<=nData );
+ pData += n;
+ nData -= n;
+ }
+ if( pLastDocid ) *pLastDocid = iPrevDocid;
+}
+#define ASSERT_VALID_DOCLIST(i, p, n, o) docListValidate(i, p, n, o)
+#else
+#define ASSERT_VALID_DOCLIST(i, p, n, o) assert( 1 )
+#endif
+
+/*******************************************************************/
+/* DLWriter is used to write doclist data to a DataBuffer. DLWriter
+** always appends to the buffer and does not own it.
+**
+** dlwInit - initialize to write a given type doclistto a buffer.
+** dlwDestroy - clear the writer's memory. Does not free buffer.
+** dlwAppend - append raw doclist data to buffer.
+** dlwCopy - copy next doclist from reader to writer.
+** dlwAdd - construct doclist element and append to buffer.
+** Only apply dlwAdd() to DL_DOCIDS doclists (else use PLWriter).
+*/
+typedef struct DLWriter {
+ DocListType iType;
+ DataBuffer *b;
+ sqlite_int64 iPrevDocid;
+#ifndef NDEBUG
+ int has_iPrevDocid;
+#endif
+} DLWriter;
+
+static void dlwInit(DLWriter *pWriter, DocListType iType, DataBuffer *b){
+ pWriter->b = b;
+ pWriter->iType = iType;
+ pWriter->iPrevDocid = 0;
+#ifndef NDEBUG
+ pWriter->has_iPrevDocid = 0;
+#endif
+}
+static void dlwDestroy(DLWriter *pWriter){
+ SCRAMBLE(pWriter);
+}
+/* iFirstDocid is the first docid in the doclist in pData. It is
+** needed because pData may point within a larger doclist, in which
+** case the first item would be delta-encoded.
+**
+** iLastDocid is the final docid in the doclist in pData. It is
+** needed to create the new iPrevDocid for future delta-encoding. The
+** code could decode the passed doclist to recreate iLastDocid, but
+** the only current user (docListMerge) already has decoded this
+** information.
+*/
+/* TODO(shess) This has become just a helper for docListMerge.
+** Consider a refactor to make this cleaner.
+*/
+static void dlwAppend(DLWriter *pWriter,
+ const char *pData, int nData,
+ sqlite_int64 iFirstDocid, sqlite_int64 iLastDocid){
+ sqlite_int64 iDocid = 0;
+ char c[VARINT_MAX];
+ int nFirstOld, nFirstNew; /* Old and new varint len of first docid. */
+#ifndef NDEBUG
+ sqlite_int64 iLastDocidDelta;
+#endif
+
+ /* Recode the initial docid as delta from iPrevDocid. */
+ nFirstOld = fts3GetVarint(pData, &iDocid);
+ assert( nFirstOld<nData || (nFirstOld==nData && pWriter->iType==DL_DOCIDS) );
+ nFirstNew = fts3PutVarint(c, iFirstDocid-pWriter->iPrevDocid);
+
+ /* Verify that the incoming doclist is valid AND that it ends with
+ ** the expected docid. This is essential because we'll trust this
+ ** docid in future delta-encoding.
+ */
+ ASSERT_VALID_DOCLIST(pWriter->iType, pData, nData, &iLastDocidDelta);
+ assert( iLastDocid==iFirstDocid-iDocid+iLastDocidDelta );
+
+ /* Append recoded initial docid and everything else. Rest of docids
+ ** should have been delta-encoded from previous initial docid.
+ */
+ if( nFirstOld<nData ){
+ dataBufferAppend2(pWriter->b, c, nFirstNew,
+ pData+nFirstOld, nData-nFirstOld);
+ }else{
+ dataBufferAppend(pWriter->b, c, nFirstNew);
+ }
+ pWriter->iPrevDocid = iLastDocid;
+}
+static void dlwCopy(DLWriter *pWriter, DLReader *pReader){
+ dlwAppend(pWriter, dlrDocData(pReader), dlrDocDataBytes(pReader),
+ dlrDocid(pReader), dlrDocid(pReader));
+}
+static void dlwAdd(DLWriter *pWriter, sqlite_int64 iDocid){
+ char c[VARINT_MAX];
+ int n = fts3PutVarint(c, iDocid-pWriter->iPrevDocid);
+
+ /* Docids must ascend. */
+ assert( !pWriter->has_iPrevDocid || iDocid>pWriter->iPrevDocid );
+ assert( pWriter->iType==DL_DOCIDS );
+
+ dataBufferAppend(pWriter->b, c, n);
+ pWriter->iPrevDocid = iDocid;
+#ifndef NDEBUG
+ pWriter->has_iPrevDocid = 1;
+#endif
+}
+
+/*******************************************************************/
+/* PLReader is used to read data from a document's position list. As
+** the caller steps through the list, data is cached so that varints
+** only need to be decoded once.
+**
+** plrInit, plrDestroy - create/destroy a reader.
+** plrColumn, plrPosition, plrStartOffset, plrEndOffset - accessors
+** plrAtEnd - at end of stream, only call plrDestroy once true.
+** plrStep - step to the next element.
+*/
+typedef struct PLReader {
+ /* These refer to the next position's data. nData will reach 0 when
+ ** reading the last position, so plrStep() signals EOF by setting
+ ** pData to NULL.
+ */
+ const char *pData;
+ int nData;
+
+ DocListType iType;
+ int iColumn; /* the last column read */
+ int iPosition; /* the last position read */
+ int iStartOffset; /* the last start offset read */
+ int iEndOffset; /* the last end offset read */
+} PLReader;
+
+static int plrAtEnd(PLReader *pReader){
+ return pReader->pData==NULL;
+}
+static int plrColumn(PLReader *pReader){
+ assert( !plrAtEnd(pReader) );
+ return pReader->iColumn;
+}
+static int plrPosition(PLReader *pReader){
+ assert( !plrAtEnd(pReader) );
+ return pReader->iPosition;
+}
+static int plrStartOffset(PLReader *pReader){
+ assert( !plrAtEnd(pReader) );
+ return pReader->iStartOffset;
+}
+static int plrEndOffset(PLReader *pReader){
+ assert( !plrAtEnd(pReader) );
+ return pReader->iEndOffset;
+}
+static void plrStep(PLReader *pReader){
+ int i, n;
+
+ assert( !plrAtEnd(pReader) );
+
+ if( pReader->nData==0 ){
+ pReader->pData = NULL;
+ return;
+ }
+
+ n = fts3GetVarint32(pReader->pData, &i);
+ if( i==POS_COLUMN ){
+ n += fts3GetVarint32(pReader->pData+n, &pReader->iColumn);
+ pReader->iPosition = 0;
+ pReader->iStartOffset = 0;
+ n += fts3GetVarint32(pReader->pData+n, &i);
+ }
+ /* Should never see adjacent column changes. */
+ assert( i!=POS_COLUMN );
+
+ if( i==POS_END ){
+ pReader->nData = 0;
+ pReader->pData = NULL;
+ return;
+ }
+
+ pReader->iPosition += i-POS_BASE;
+ if( pReader->iType==DL_POSITIONS_OFFSETS ){
+ n += fts3GetVarint32(pReader->pData+n, &i);
+ pReader->iStartOffset += i;
+ n += fts3GetVarint32(pReader->pData+n, &i);
+ pReader->iEndOffset = pReader->iStartOffset+i;
+ }
+ assert( n<=pReader->nData );
+ pReader->pData += n;
+ pReader->nData -= n;
+}
+
+static void plrInit(PLReader *pReader, DLReader *pDLReader){
+ pReader->pData = dlrPosData(pDLReader);
+ pReader->nData = dlrPosDataLen(pDLReader);
+ pReader->iType = pDLReader->iType;
+ pReader->iColumn = 0;
+ pReader->iPosition = 0;
+ pReader->iStartOffset = 0;
+ pReader->iEndOffset = 0;
+ plrStep(pReader);
+}
+static void plrDestroy(PLReader *pReader){
+ SCRAMBLE(pReader);
+}
+
+/*******************************************************************/
+/* PLWriter is used in constructing a document's position list. As a
+** convenience, if iType is DL_DOCIDS, PLWriter becomes a no-op.
+** PLWriter writes to the associated DLWriter's buffer.
+**
+** plwInit - init for writing a document's poslist.
+** plwDestroy - clear a writer.
+** plwAdd - append position and offset information.
+** plwCopy - copy next position's data from reader to writer.
+** plwTerminate - add any necessary doclist terminator.
+**
+** Calling plwAdd() after plwTerminate() may result in a corrupt
+** doclist.
+*/
+/* TODO(shess) Until we've written the second item, we can cache the
+** first item's information. Then we'd have three states:
+**
+** - initialized with docid, no positions.
+** - docid and one position.
+** - docid and multiple positions.
+**
+** Only the last state needs to actually write to dlw->b, which would
+** be an improvement in the DLCollector case.
+*/
+typedef struct PLWriter {
+ DLWriter *dlw;
+
+ int iColumn; /* the last column written */
+ int iPos; /* the last position written */
+ int iOffset; /* the last start offset written */
+} PLWriter;
+
+/* TODO(shess) In the case where the parent is reading these values
+** from a PLReader, we could optimize to a copy if that PLReader has
+** the same type as pWriter.
+*/
+static void plwAdd(PLWriter *pWriter, int iColumn, int iPos,
+ int iStartOffset, int iEndOffset){
+ /* Worst-case space for POS_COLUMN, iColumn, iPosDelta,
+ ** iStartOffsetDelta, and iEndOffsetDelta.
+ */
+ char c[5*VARINT_MAX];
+ int n = 0;
+
+ /* Ban plwAdd() after plwTerminate(). */
+ assert( pWriter->iPos!=-1 );
+
+ if( pWriter->dlw->iType==DL_DOCIDS ) return;
+
+ if( iColumn!=pWriter->iColumn ){
+ n += fts3PutVarint(c+n, POS_COLUMN);
+ n += fts3PutVarint(c+n, iColumn);
+ pWriter->iColumn = iColumn;
+ pWriter->iPos = 0;
+ pWriter->iOffset = 0;
+ }
+ assert( iPos>=pWriter->iPos );
+ n += fts3PutVarint(c+n, POS_BASE+(iPos-pWriter->iPos));
+ pWriter->iPos = iPos;
+ if( pWriter->dlw->iType==DL_POSITIONS_OFFSETS ){
+ assert( iStartOffset>=pWriter->iOffset );
+ n += fts3PutVarint(c+n, iStartOffset-pWriter->iOffset);
+ pWriter->iOffset = iStartOffset;
+ assert( iEndOffset>=iStartOffset );
+ n += fts3PutVarint(c+n, iEndOffset-iStartOffset);
+ }
+ dataBufferAppend(pWriter->dlw->b, c, n);
+}
+static void plwCopy(PLWriter *pWriter, PLReader *pReader){
+ plwAdd(pWriter, plrColumn(pReader), plrPosition(pReader),
+ plrStartOffset(pReader), plrEndOffset(pReader));
+}
+static void plwInit(PLWriter *pWriter, DLWriter *dlw, sqlite_int64 iDocid){
+ char c[VARINT_MAX];
+ int n;
+
+ pWriter->dlw = dlw;
+
+ /* Docids must ascend. */
+ assert( !pWriter->dlw->has_iPrevDocid || iDocid>pWriter->dlw->iPrevDocid );
+ n = fts3PutVarint(c, iDocid-pWriter->dlw->iPrevDocid);
+ dataBufferAppend(pWriter->dlw->b, c, n);
+ pWriter->dlw->iPrevDocid = iDocid;
+#ifndef NDEBUG
+ pWriter->dlw->has_iPrevDocid = 1;
+#endif
+
+ pWriter->iColumn = 0;
+ pWriter->iPos = 0;
+ pWriter->iOffset = 0;
+}
+/* TODO(shess) Should plwDestroy() also terminate the doclist? But
+** then plwDestroy() would no longer be just a destructor, it would
+** also be doing work, which isn't consistent with the overall idiom.
+** Another option would be for plwAdd() to always append any necessary
+** terminator, so that the output is always correct. But that would
+** add incremental work to the common case with the only benefit being
+** API elegance. Punt for now.
+*/
+static void plwTerminate(PLWriter *pWriter){
+ if( pWriter->dlw->iType>DL_DOCIDS ){
+ char c[VARINT_MAX];
+ int n = fts3PutVarint(c, POS_END);
+ dataBufferAppend(pWriter->dlw->b, c, n);
+ }
+#ifndef NDEBUG
+ /* Mark as terminated for assert in plwAdd(). */
+ pWriter->iPos = -1;
+#endif
+}
+static void plwDestroy(PLWriter *pWriter){
+ SCRAMBLE(pWriter);
+}
+
+/*******************************************************************/
+/* DLCollector wraps PLWriter and DLWriter to provide a
+** dynamically-allocated doclist area to use during tokenization.
+**
+** dlcNew - malloc up and initialize a collector.
+** dlcDelete - destroy a collector and all contained items.
+** dlcAddPos - append position and offset information.
+** dlcAddDoclist - add the collected doclist to the given buffer.
+** dlcNext - terminate the current document and open another.
+*/
+typedef struct DLCollector {
+ DataBuffer b;
+ DLWriter dlw;
+ PLWriter plw;
+} DLCollector;
+
+/* TODO(shess) This could also be done by calling plwTerminate() and
+** dataBufferAppend(). I tried that, expecting nominal performance
+** differences, but it seemed to pretty reliably be worth 1% to code
+** it this way. I suspect it is the incremental malloc overhead (some
+** percentage of the plwTerminate() calls will cause a realloc), so
+** this might be worth revisiting if the DataBuffer implementation
+** changes.
+*/
+static void dlcAddDoclist(DLCollector *pCollector, DataBuffer *b){
+ if( pCollector->dlw.iType>DL_DOCIDS ){
+ char c[VARINT_MAX];
+ int n = fts3PutVarint(c, POS_END);
+ dataBufferAppend2(b, pCollector->b.pData, pCollector->b.nData, c, n);
+ }else{
+ dataBufferAppend(b, pCollector->b.pData, pCollector->b.nData);
+ }
+}
+static void dlcNext(DLCollector *pCollector, sqlite_int64 iDocid){
+ plwTerminate(&pCollector->plw);
+ plwDestroy(&pCollector->plw);
+ plwInit(&pCollector->plw, &pCollector->dlw, iDocid);
+}
+static void dlcAddPos(DLCollector *pCollector, int iColumn, int iPos,
+ int iStartOffset, int iEndOffset){
+ plwAdd(&pCollector->plw, iColumn, iPos, iStartOffset, iEndOffset);
+}
+
+static DLCollector *dlcNew(sqlite_int64 iDocid, DocListType iType){
+ DLCollector *pCollector = sqlite3_malloc(sizeof(DLCollector));
+ dataBufferInit(&pCollector->b, 0);
+ dlwInit(&pCollector->dlw, iType, &pCollector->b);
+ plwInit(&pCollector->plw, &pCollector->dlw, iDocid);
+ return pCollector;
+}
+static void dlcDelete(DLCollector *pCollector){
+ plwDestroy(&pCollector->plw);
+ dlwDestroy(&pCollector->dlw);
+ dataBufferDestroy(&pCollector->b);
+ SCRAMBLE(pCollector);
+ sqlite3_free(pCollector);
+}
+
+
+/* Copy the doclist data of iType in pData/nData into *out, trimming
+** unnecessary data as we go. Only columns matching iColumn are
+** copied, all columns copied if iColumn is -1. Elements with no
+** matching columns are dropped. The output is an iOutType doclist.
+*/
+/* NOTE(shess) This code is only valid after all doclists are merged.
+** If this is run before merges, then doclist items which represent
+** deletion will be trimmed, and will thus not effect a deletion
+** during the merge.
+*/
+static void docListTrim(DocListType iType, const char *pData, int nData,
+ int iColumn, DocListType iOutType, DataBuffer *out){
+ DLReader dlReader;
+ DLWriter dlWriter;
+
+ assert( iOutType<=iType );
+
+ dlrInit(&dlReader, iType, pData, nData);
+ dlwInit(&dlWriter, iOutType, out);
+
+ while( !dlrAtEnd(&dlReader) ){
+ PLReader plReader;
+ PLWriter plWriter;
+ int match = 0;
+
+ plrInit(&plReader, &dlReader);
+
+ while( !plrAtEnd(&plReader) ){
+ if( iColumn==-1 || plrColumn(&plReader)==iColumn ){
+ if( !match ){
+ plwInit(&plWriter, &dlWriter, dlrDocid(&dlReader));
+ match = 1;
+ }
+ plwAdd(&plWriter, plrColumn(&plReader), plrPosition(&plReader),
+ plrStartOffset(&plReader), plrEndOffset(&plReader));
+ }
+ plrStep(&plReader);
+ }
+ if( match ){
+ plwTerminate(&plWriter);
+ plwDestroy(&plWriter);
+ }
+
+ plrDestroy(&plReader);
+ dlrStep(&dlReader);
+ }
+ dlwDestroy(&dlWriter);
+ dlrDestroy(&dlReader);
+}
+
+/* Used by docListMerge() to keep doclists in the ascending order by
+** docid, then ascending order by age (so the newest comes first).
+*/
+typedef struct OrderedDLReader {
+ DLReader *pReader;
+
+ /* TODO(shess) If we assume that docListMerge pReaders is ordered by
+ ** age (which we do), then we could use pReader comparisons to break
+ ** ties.
+ */
+ int idx;
+} OrderedDLReader;
+
+/* Order eof to end, then by docid asc, idx desc. */
+static int orderedDLReaderCmp(OrderedDLReader *r1, OrderedDLReader *r2){
+ if( dlrAtEnd(r1->pReader) ){
+ if( dlrAtEnd(r2->pReader) ) return 0; /* Both atEnd(). */
+ return 1; /* Only r1 atEnd(). */
+ }
+ if( dlrAtEnd(r2->pReader) ) return -1; /* Only r2 atEnd(). */
+
+ if( dlrDocid(r1->pReader)<dlrDocid(r2->pReader) ) return -1;
+ if( dlrDocid(r1->pReader)>dlrDocid(r2->pReader) ) return 1;
+
+ /* Descending on idx. */
+ return r2->idx-r1->idx;
+}
+
+/* Bubble p[0] to appropriate place in p[1..n-1]. Assumes that
+** p[1..n-1] is already sorted.
+*/
+/* TODO(shess) Is this frequent enough to warrant a binary search?
+** Before implementing that, instrument the code to check. In most
+** current usage, I expect that p[0] will be less than p[1] a very
+** high proportion of the time.
+*/
+static void orderedDLReaderReorder(OrderedDLReader *p, int n){
+ while( n>1 && orderedDLReaderCmp(p, p+1)>0 ){
+ OrderedDLReader tmp = p[0];
+ p[0] = p[1];
+ p[1] = tmp;
+ n--;
+ p++;
+ }
+}
+
+/* Given an array of doclist readers, merge their doclist elements
+** into out in sorted order (by docid), dropping elements from older
+** readers when there is a duplicate docid. pReaders is assumed to be
+** ordered by age, oldest first.
+*/
+/* TODO(shess) nReaders must be <= MERGE_COUNT. This should probably
+** be fixed.
+*/
+static void docListMerge(DataBuffer *out,
+ DLReader *pReaders, int nReaders){
+ OrderedDLReader readers[MERGE_COUNT];
+ DLWriter writer;
+ int i, n;
+ const char *pStart = 0;
+ int nStart = 0;
+ sqlite_int64 iFirstDocid = 0, iLastDocid = 0;
+
+ assert( nReaders>0 );
+ if( nReaders==1 ){
+ dataBufferAppend(out, dlrDocData(pReaders), dlrAllDataBytes(pReaders));
+ return;
+ }
+
+ assert( nReaders<=MERGE_COUNT );
+ n = 0;
+ for(i=0; i<nReaders; i++){
+ assert( pReaders[i].iType==pReaders[0].iType );
+ readers[i].pReader = pReaders+i;
+ readers[i].idx = i;
+ n += dlrAllDataBytes(&pReaders[i]);
+ }
+ /* Conservatively size output to sum of inputs. Output should end
+ ** up strictly smaller than input.
+ */
+ dataBufferExpand(out, n);
+
+ /* Get the readers into sorted order. */
+ while( i-->0 ){
+ orderedDLReaderReorder(readers+i, nReaders-i);
+ }
+
+ dlwInit(&writer, pReaders[0].iType, out);
+ while( !dlrAtEnd(readers[0].pReader) ){
+ sqlite_int64 iDocid = dlrDocid(readers[0].pReader);
+
+ /* If this is a continuation of the current buffer to copy, extend
+ ** that buffer. memcpy() seems to be more efficient if it has a
+ ** lots of data to copy.
+ */
+ if( dlrDocData(readers[0].pReader)==pStart+nStart ){
+ nStart += dlrDocDataBytes(readers[0].pReader);
+ }else{
+ if( pStart!=0 ){
+ dlwAppend(&writer, pStart, nStart, iFirstDocid, iLastDocid);
+ }
+ pStart = dlrDocData(readers[0].pReader);
+ nStart = dlrDocDataBytes(readers[0].pReader);
+ iFirstDocid = iDocid;
+ }
+ iLastDocid = iDocid;
+ dlrStep(readers[0].pReader);
+
+ /* Drop all of the older elements with the same docid. */
+ for(i=1; i<nReaders &&
+ !dlrAtEnd(readers[i].pReader) &&
+ dlrDocid(readers[i].pReader)==iDocid; i++){
+ dlrStep(readers[i].pReader);
+ }
+
+ /* Get the readers back into order. */
+ while( i-->0 ){
+ orderedDLReaderReorder(readers+i, nReaders-i);
+ }
+ }
+
+ /* Copy over any remaining elements. */
+ if( nStart>0 ) dlwAppend(&writer, pStart, nStart, iFirstDocid, iLastDocid);
+ dlwDestroy(&writer);
+}
+
+/* Helper function for posListUnion(). Compares the current position
+** between left and right, returning as standard C idiom of <0 if
+** left<right, >0 if left>right, and 0 if left==right. "End" always
+** compares greater.
+*/
+static int posListCmp(PLReader *pLeft, PLReader *pRight){
+ assert( pLeft->iType==pRight->iType );
+ if( pLeft->iType==DL_DOCIDS ) return 0;
+
+ if( plrAtEnd(pLeft) ) return plrAtEnd(pRight) ? 0 : 1;
+ if( plrAtEnd(pRight) ) return -1;
+
+ if( plrColumn(pLeft)<plrColumn(pRight) ) return -1;
+ if( plrColumn(pLeft)>plrColumn(pRight) ) return 1;
+
+ if( plrPosition(pLeft)<plrPosition(pRight) ) return -1;
+ if( plrPosition(pLeft)>plrPosition(pRight) ) return 1;
+ if( pLeft->iType==DL_POSITIONS ) return 0;
+
+ if( plrStartOffset(pLeft)<plrStartOffset(pRight) ) return -1;
+ if( plrStartOffset(pLeft)>plrStartOffset(pRight) ) return 1;
+
+ if( plrEndOffset(pLeft)<plrEndOffset(pRight) ) return -1;
+ if( plrEndOffset(pLeft)>plrEndOffset(pRight) ) return 1;
+
+ return 0;
+}
+
+/* Write the union of position lists in pLeft and pRight to pOut.
+** "Union" in this case meaning "All unique position tuples". Should
+** work with any doclist type, though both inputs and the output
+** should be the same type.
+*/
+static void posListUnion(DLReader *pLeft, DLReader *pRight, DLWriter *pOut){
+ PLReader left, right;
+ PLWriter writer;
+
+ assert( dlrDocid(pLeft)==dlrDocid(pRight) );
+ assert( pLeft->iType==pRight->iType );
+ assert( pLeft->iType==pOut->iType );
+
+ plrInit(&left, pLeft);
+ plrInit(&right, pRight);
+ plwInit(&writer, pOut, dlrDocid(pLeft));
+
+ while( !plrAtEnd(&left) || !plrAtEnd(&right) ){
+ int c = posListCmp(&left, &right);
+ if( c<0 ){
+ plwCopy(&writer, &left);
+ plrStep(&left);
+ }else if( c>0 ){
+ plwCopy(&writer, &right);
+ plrStep(&right);
+ }else{
+ plwCopy(&writer, &left);
+ plrStep(&left);
+ plrStep(&right);
+ }
+ }
+
+ plwTerminate(&writer);
+ plwDestroy(&writer);
+ plrDestroy(&left);
+ plrDestroy(&right);
+}
+
+/* Write the union of doclists in pLeft and pRight to pOut. For
+** docids in common between the inputs, the union of the position
+** lists is written. Inputs and outputs are always type DL_DEFAULT.
+*/
+static void docListUnion(
+ const char *pLeft, int nLeft,
+ const char *pRight, int nRight,
+ DataBuffer *pOut /* Write the combined doclist here */
+){
+ DLReader left, right;
+ DLWriter writer;
+
+ if( nLeft==0 ){
+ if( nRight!=0) dataBufferAppend(pOut, pRight, nRight);
+ return;
+ }
+ if( nRight==0 ){
+ dataBufferAppend(pOut, pLeft, nLeft);
+ return;
+ }
+
+ dlrInit(&left, DL_DEFAULT, pLeft, nLeft);
+ dlrInit(&right, DL_DEFAULT, pRight, nRight);
+ dlwInit(&writer, DL_DEFAULT, pOut);
+
+ while( !dlrAtEnd(&left) || !dlrAtEnd(&right) ){
+ if( dlrAtEnd(&right) ){
+ dlwCopy(&writer, &left);
+ dlrStep(&left);
+ }else if( dlrAtEnd(&left) ){
+ dlwCopy(&writer, &right);
+ dlrStep(&right);
+ }else if( dlrDocid(&left)<dlrDocid(&right) ){
+ dlwCopy(&writer, &left);
+ dlrStep(&left);
+ }else if( dlrDocid(&left)>dlrDocid(&right) ){
+ dlwCopy(&writer, &right);
+ dlrStep(&right);
+ }else{
+ posListUnion(&left, &right, &writer);
+ dlrStep(&left);
+ dlrStep(&right);
+ }
+ }
+
+ dlrDestroy(&left);
+ dlrDestroy(&right);
+ dlwDestroy(&writer);
+}
+
+/*
+** This function is used as part of the implementation of phrase and
+** NEAR matching.
+**
+** pLeft and pRight are DLReaders positioned to the same docid in
+** lists of type DL_POSITION. This function writes an entry to the
+** DLWriter pOut for each position in pRight that is less than
+** (nNear+1) greater (but not equal to or smaller) than a position
+** in pLeft. For example, if nNear is 0, and the positions contained
+** by pLeft and pRight are:
+**
+** pLeft: 5 10 15 20
+** pRight: 6 9 17 21
+**
+** then the docid is added to pOut. If pOut is of type DL_POSITIONS,
+** then a positionids "6" and "21" are also added to pOut.
+**
+** If boolean argument isSaveLeft is true, then positionids are copied
+** from pLeft instead of pRight. In the example above, the positions "5"
+** and "20" would be added instead of "6" and "21".
+*/
+static void posListPhraseMerge(
+ DLReader *pLeft,
+ DLReader *pRight,
+ int nNear,
+ int isSaveLeft,
+ DLWriter *pOut
+){
+ PLReader left, right;
+ PLWriter writer;
+ int match = 0;
+
+ assert( dlrDocid(pLeft)==dlrDocid(pRight) );
+ assert( pOut->iType!=DL_POSITIONS_OFFSETS );
+
+ plrInit(&left, pLeft);
+ plrInit(&right, pRight);
+
+ while( !plrAtEnd(&left) && !plrAtEnd(&right) ){
+ if( plrColumn(&left)<plrColumn(&right) ){
+ plrStep(&left);
+ }else if( plrColumn(&left)>plrColumn(&right) ){
+ plrStep(&right);
+ }else if( plrPosition(&left)>=plrPosition(&right) ){
+ plrStep(&right);
+ }else{
+ if( (plrPosition(&right)-plrPosition(&left))<=(nNear+1) ){
+ if( !match ){
+ plwInit(&writer, pOut, dlrDocid(pLeft));
+ match = 1;
+ }
+ if( !isSaveLeft ){
+ plwAdd(&writer, plrColumn(&right), plrPosition(&right), 0, 0);
+ }else{
+ plwAdd(&writer, plrColumn(&left), plrPosition(&left), 0, 0);
+ }
+ plrStep(&right);
+ }else{
+ plrStep(&left);
+ }
+ }
+ }
+
+ if( match ){
+ plwTerminate(&writer);
+ plwDestroy(&writer);
+ }
+
+ plrDestroy(&left);
+ plrDestroy(&right);
+}
+
+/*
+** Compare the values pointed to by the PLReaders passed as arguments.
+** Return -1 if the value pointed to by pLeft is considered less than
+** the value pointed to by pRight, +1 if it is considered greater
+** than it, or 0 if it is equal. i.e.
+**
+** (*pLeft - *pRight)
+**
+** A PLReader that is in the EOF condition is considered greater than
+** any other. If neither argument is in EOF state, the return value of
+** plrColumn() is used. If the plrColumn() values are equal, the
+** comparison is on the basis of plrPosition().
+*/
+static int plrCompare(PLReader *pLeft, PLReader *pRight){
+ assert(!plrAtEnd(pLeft) || !plrAtEnd(pRight));
+
+ if( plrAtEnd(pRight) || plrAtEnd(pLeft) ){
+ return (plrAtEnd(pRight) ? -1 : 1);
+ }
+ if( plrColumn(pLeft)!=plrColumn(pRight) ){
+ return ((plrColumn(pLeft)<plrColumn(pRight)) ? -1 : 1);
+ }
+ if( plrPosition(pLeft)!=plrPosition(pRight) ){
+ return ((plrPosition(pLeft)<plrPosition(pRight)) ? -1 : 1);
+ }
+ return 0;
+}
+
+/* We have two doclists with positions: pLeft and pRight. Depending
+** on the value of the nNear parameter, perform either a phrase
+** intersection (if nNear==0) or a NEAR intersection (if nNear>0)
+** and write the results into pOut.
+**
+** A phrase intersection means that two documents only match
+** if pLeft.iPos+1==pRight.iPos.
+**
+** A NEAR intersection means that two documents only match if
+** (abs(pLeft.iPos-pRight.iPos)<nNear).
+**
+** If a NEAR intersection is requested, then the nPhrase argument should
+** be passed the number of tokens in the two operands to the NEAR operator
+** combined. For example:
+**
+** Query syntax nPhrase
+** ------------------------------------
+** "A B C" NEAR "D E" 5
+** A NEAR B 2
+**
+** iType controls the type of data written to pOut. If iType is
+** DL_POSITIONS, the positions are those from pRight.
+*/
+static void docListPhraseMerge(
+ const char *pLeft, int nLeft,
+ const char *pRight, int nRight,
+ int nNear, /* 0 for a phrase merge, non-zero for a NEAR merge */
+ int nPhrase, /* Number of tokens in left+right operands to NEAR */
+ DocListType iType, /* Type of doclist to write to pOut */
+ DataBuffer *pOut /* Write the combined doclist here */
+){
+ DLReader left, right;
+ DLWriter writer;
+
+ if( nLeft==0 || nRight==0 ) return;
+
+ assert( iType!=DL_POSITIONS_OFFSETS );
+
+ dlrInit(&left, DL_POSITIONS, pLeft, nLeft);
+ dlrInit(&right, DL_POSITIONS, pRight, nRight);
+ dlwInit(&writer, iType, pOut);
+
+ while( !dlrAtEnd(&left) && !dlrAtEnd(&right) ){
+ if( dlrDocid(&left)<dlrDocid(&right) ){
+ dlrStep(&left);
+ }else if( dlrDocid(&right)<dlrDocid(&left) ){
+ dlrStep(&right);
+ }else{
+ if( nNear==0 ){
+ posListPhraseMerge(&left, &right, 0, 0, &writer);
+ }else{
+ /* This case occurs when two terms (simple terms or phrases) are
+ * connected by a NEAR operator, span (nNear+1). i.e.
+ *
+ * '"terrible company" NEAR widget'
+ */
+ DataBuffer one = {0, 0, 0};
+ DataBuffer two = {0, 0, 0};
+
+ DLWriter dlwriter2;
+ DLReader dr1 = {0, 0, 0, 0, 0};
+ DLReader dr2 = {0, 0, 0, 0, 0};
+
+ dlwInit(&dlwriter2, iType, &one);
+ posListPhraseMerge(&right, &left, nNear-3+nPhrase, 1, &dlwriter2);
+ dlwInit(&dlwriter2, iType, &two);
+ posListPhraseMerge(&left, &right, nNear-1, 0, &dlwriter2);
+
+ if( one.nData) dlrInit(&dr1, iType, one.pData, one.nData);
+ if( two.nData) dlrInit(&dr2, iType, two.pData, two.nData);
+
+ if( !dlrAtEnd(&dr1) || !dlrAtEnd(&dr2) ){
+ PLReader pr1 = {0};
+ PLReader pr2 = {0};
+
+ PLWriter plwriter;
+ plwInit(&plwriter, &writer, dlrDocid(dlrAtEnd(&dr1)?&dr2:&dr1));
+
+ if( one.nData ) plrInit(&pr1, &dr1);
+ if( two.nData ) plrInit(&pr2, &dr2);
+ while( !plrAtEnd(&pr1) || !plrAtEnd(&pr2) ){
+ int iCompare = plrCompare(&pr1, &pr2);
+ switch( iCompare ){
+ case -1:
+ plwCopy(&plwriter, &pr1);
+ plrStep(&pr1);
+ break;
+ case 1:
+ plwCopy(&plwriter, &pr2);
+ plrStep(&pr2);
+ break;
+ case 0:
+ plwCopy(&plwriter, &pr1);
+ plrStep(&pr1);
+ plrStep(&pr2);
+ break;
+ }
+ }
+ plwTerminate(&plwriter);
+ }
+ dataBufferDestroy(&one);
+ dataBufferDestroy(&two);
+ }
+ dlrStep(&left);
+ dlrStep(&right);
+ }
+ }
+
+ dlrDestroy(&left);
+ dlrDestroy(&right);
+ dlwDestroy(&writer);
+}
+
+/* We have two DL_DOCIDS doclists: pLeft and pRight.
+** Write the intersection of these two doclists into pOut as a
+** DL_DOCIDS doclist.
+*/
+static void docListAndMerge(
+ const char *pLeft, int nLeft,
+ const char *pRight, int nRight,
+ DataBuffer *pOut /* Write the combined doclist here */
+){
+ DLReader left, right;
+ DLWriter writer;
+
+ if( nLeft==0 || nRight==0 ) return;
+
+ dlrInit(&left, DL_DOCIDS, pLeft, nLeft);
+ dlrInit(&right, DL_DOCIDS, pRight, nRight);
+ dlwInit(&writer, DL_DOCIDS, pOut);
+
+ while( !dlrAtEnd(&left) && !dlrAtEnd(&right) ){
+ if( dlrDocid(&left)<dlrDocid(&right) ){
+ dlrStep(&left);
+ }else if( dlrDocid(&right)<dlrDocid(&left) ){
+ dlrStep(&right);
+ }else{
+ dlwAdd(&writer, dlrDocid(&left));
+ dlrStep(&left);
+ dlrStep(&right);
+ }
+ }
+
+ dlrDestroy(&left);
+ dlrDestroy(&right);
+ dlwDestroy(&writer);
+}
+
+/* We have two DL_DOCIDS doclists: pLeft and pRight.
+** Write the union of these two doclists into pOut as a
+** DL_DOCIDS doclist.
+*/
+static void docListOrMerge(
+ const char *pLeft, int nLeft,
+ const char *pRight, int nRight,
+ DataBuffer *pOut /* Write the combined doclist here */
+){
+ DLReader left, right;
+ DLWriter writer;
+
+ if( nLeft==0 ){
+ if( nRight!=0 ) dataBufferAppend(pOut, pRight, nRight);
+ return;
+ }
+ if( nRight==0 ){
+ dataBufferAppend(pOut, pLeft, nLeft);
+ return;
+ }
+
+ dlrInit(&left, DL_DOCIDS, pLeft, nLeft);
+ dlrInit(&right, DL_DOCIDS, pRight, nRight);
+ dlwInit(&writer, DL_DOCIDS, pOut);
+
+ while( !dlrAtEnd(&left) || !dlrAtEnd(&right) ){
+ if( dlrAtEnd(&right) ){
+ dlwAdd(&writer, dlrDocid(&left));
+ dlrStep(&left);
+ }else if( dlrAtEnd(&left) ){
+ dlwAdd(&writer, dlrDocid(&right));
+ dlrStep(&right);
+ }else if( dlrDocid(&left)<dlrDocid(&right) ){
+ dlwAdd(&writer, dlrDocid(&left));
+ dlrStep(&left);
+ }else if( dlrDocid(&right)<dlrDocid(&left) ){
+ dlwAdd(&writer, dlrDocid(&right));
+ dlrStep(&right);
+ }else{
+ dlwAdd(&writer, dlrDocid(&left));
+ dlrStep(&left);
+ dlrStep(&right);
+ }
+ }
+
+ dlrDestroy(&left);
+ dlrDestroy(&right);
+ dlwDestroy(&writer);
+}
+
+/* We have two DL_DOCIDS doclists: pLeft and pRight.
+** Write into pOut as DL_DOCIDS doclist containing all documents that
+** occur in pLeft but not in pRight.
+*/
+static void docListExceptMerge(
+ const char *pLeft, int nLeft,
+ const char *pRight, int nRight,
+ DataBuffer *pOut /* Write the combined doclist here */
+){
+ DLReader left, right;
+ DLWriter writer;
+
+ if( nLeft==0 ) return;
+ if( nRight==0 ){
+ dataBufferAppend(pOut, pLeft, nLeft);
+ return;
+ }
+
+ dlrInit(&left, DL_DOCIDS, pLeft, nLeft);
+ dlrInit(&right, DL_DOCIDS, pRight, nRight);
+ dlwInit(&writer, DL_DOCIDS, pOut);
+
+ while( !dlrAtEnd(&left) ){
+ while( !dlrAtEnd(&right) && dlrDocid(&right)<dlrDocid(&left) ){
+ dlrStep(&right);
+ }
+ if( dlrAtEnd(&right) || dlrDocid(&left)<dlrDocid(&right) ){
+ dlwAdd(&writer, dlrDocid(&left));
+ }
+ dlrStep(&left);
+ }
+
+ dlrDestroy(&left);
+ dlrDestroy(&right);
+ dlwDestroy(&writer);
+}
+
+static char *string_dup_n(const char *s, int n){
+ char *str = sqlite3_malloc(n + 1);
+ memcpy(str, s, n);
+ str[n] = '\0';
+ return str;
+}
+
+/* Duplicate a string; the caller must free() the returned string.
+ * (We don't use strdup() since it is not part of the standard C library and
+ * may not be available everywhere.) */
+static char *string_dup(const char *s){
+ return string_dup_n(s, strlen(s));
+}
+
+/* Format a string, replacing each occurrence of the % character with
+ * zDb.zName. This may be more convenient than sqlite_mprintf()
+ * when one string is used repeatedly in a format string.
+ * The caller must free() the returned string. */
+static char *string_format(const char *zFormat,
+ const char *zDb, const char *zName){
+ const char *p;
+ size_t len = 0;
+ size_t nDb = strlen(zDb);
+ size_t nName = strlen(zName);
+ size_t nFullTableName = nDb+1+nName;
+ char *result;
+ char *r;
+
+ /* first compute length needed */
+ for(p = zFormat ; *p ; ++p){
+ len += (*p=='%' ? nFullTableName : 1);
+ }
+ len += 1; /* for null terminator */
+
+ r = result = sqlite3_malloc(len);
+ for(p = zFormat; *p; ++p){
+ if( *p=='%' ){
+ memcpy(r, zDb, nDb);
+ r += nDb;
+ *r++ = '.';
+ memcpy(r, zName, nName);
+ r += nName;
+ } else {
+ *r++ = *p;
+ }
+ }
+ *r++ = '\0';
+ assert( r == result + len );
+ return result;
+}
+
+static int sql_exec(sqlite3 *db, const char *zDb, const char *zName,
+ const char *zFormat){
+ char *zCommand = string_format(zFormat, zDb, zName);
+ int rc;
+ FTSTRACE(("FTS3 sql: %s\n", zCommand));
+ rc = sqlite3_exec(db, zCommand, NULL, 0, NULL);
+ sqlite3_free(zCommand);
+ return rc;
+}
+
+static int sql_prepare(sqlite3 *db, const char *zDb, const char *zName,
+ sqlite3_stmt **ppStmt, const char *zFormat){
+ char *zCommand = string_format(zFormat, zDb, zName);
+ int rc;
+ FTSTRACE(("FTS3 prepare: %s\n", zCommand));
+ rc = sqlite3_prepare_v2(db, zCommand, -1, ppStmt, NULL);
+ sqlite3_free(zCommand);
+ return rc;
+}
+
+/* end utility functions */
+
+/* Forward reference */
+typedef struct fulltext_vtab fulltext_vtab;
+
+/* A single term in a query is represented by an instances of
+** the following structure. Each word which may match against
+** document content is a term. Operators, like NEAR or OR, are
+** not terms. Query terms are organized as a flat list stored
+** in the Query.pTerms array.
+**
+** If the QueryTerm.nPhrase variable is non-zero, then the QueryTerm
+** is the first in a contiguous string of terms that are either part
+** of the same phrase, or connected by the NEAR operator.
+**
+** If the QueryTerm.nNear variable is non-zero, then the token is followed
+** by a NEAR operator with span set to (nNear-1). For example, the
+** following query:
+**
+** The QueryTerm.iPhrase variable stores the index of the token within
+** its phrase, indexed starting at 1, or 1 if the token is not part
+** of any phrase.
+**
+** For example, the data structure used to represent the following query:
+**
+** ... MATCH 'sqlite NEAR/5 google NEAR/2 "search engine"'
+**
+** is:
+**
+** {nPhrase=4, iPhrase=1, nNear=6, pTerm="sqlite"},
+** {nPhrase=0, iPhrase=1, nNear=3, pTerm="google"},
+** {nPhrase=0, iPhrase=1, nNear=0, pTerm="search"},
+** {nPhrase=0, iPhrase=2, nNear=0, pTerm="engine"},
+**
+** compiling the FTS3 syntax to Query structures is done by the parseQuery()
+** function.
+*/
+typedef struct QueryTerm {
+ short int nPhrase; /* How many following terms are part of the same phrase */
+ short int iPhrase; /* This is the i-th term of a phrase. */
+ short int iColumn; /* Column of the index that must match this term */
+ short int nNear; /* term followed by a NEAR operator with span=(nNear-1) */
+ signed char isOr; /* this term is preceded by "OR" */
+ signed char isNot; /* this term is preceded by "-" */
+ signed char isPrefix; /* this term is followed by "*" */
+ char *pTerm; /* text of the term. '\000' terminated. malloced */
+ int nTerm; /* Number of bytes in pTerm[] */
+} QueryTerm;
+
+
+/* A query string is parsed into a Query structure.
+ *
+ * We could, in theory, allow query strings to be complicated
+ * nested expressions with precedence determined by parentheses.
+ * But none of the major search engines do this. (Perhaps the
+ * feeling is that an parenthesized expression is two complex of
+ * an idea for the average user to grasp.) Taking our lead from
+ * the major search engines, we will allow queries to be a list
+ * of terms (with an implied AND operator) or phrases in double-quotes,
+ * with a single optional "-" before each non-phrase term to designate
+ * negation and an optional OR connector.
+ *
+ * OR binds more tightly than the implied AND, which is what the
+ * major search engines seem to do. So, for example:
+ *
+ * [one two OR three] ==> one AND (two OR three)
+ * [one OR two three] ==> (one OR two) AND three
+ *
+ * A "-" before a term matches all entries that lack that term.
+ * The "-" must occur immediately before the term with in intervening
+ * space. This is how the search engines do it.
+ *
+ * A NOT term cannot be the right-hand operand of an OR. If this
+ * occurs in the query string, the NOT is ignored:
+ *
+ * [one OR -two] ==> one OR two
+ *
+ */
+typedef struct Query {
+ fulltext_vtab *pFts; /* The full text index */
+ int nTerms; /* Number of terms in the query */
+ QueryTerm *pTerms; /* Array of terms. Space obtained from malloc() */
+ int nextIsOr; /* Set the isOr flag on the next inserted term */
+ int nextIsNear; /* Set the isOr flag on the next inserted term */
+ int nextColumn; /* Next word parsed must be in this column */
+ int dfltColumn; /* The default column */
+} Query;
+
+
+/*
+** An instance of the following structure keeps track of generated
+** matching-word offset information and snippets.
+*/
+typedef struct Snippet {
+ int nMatch; /* Total number of matches */
+ int nAlloc; /* Space allocated for aMatch[] */
+ struct snippetMatch { /* One entry for each matching term */
+ char snStatus; /* Status flag for use while constructing snippets */
+ short int iCol; /* The column that contains the match */
+ short int iTerm; /* The index in Query.pTerms[] of the matching term */
+ int iToken; /* The index of the matching document token */
+ short int nByte; /* Number of bytes in the term */
+ int iStart; /* The offset to the first character of the term */
+ } *aMatch; /* Points to space obtained from malloc */
+ char *zOffset; /* Text rendering of aMatch[] */
+ int nOffset; /* strlen(zOffset) */
+ char *zSnippet; /* Snippet text */
+ int nSnippet; /* strlen(zSnippet) */
+} Snippet;
+
+
+typedef enum QueryType {
+ QUERY_GENERIC, /* table scan */
+ QUERY_DOCID, /* lookup by docid */
+ QUERY_FULLTEXT /* QUERY_FULLTEXT + [i] is a full-text search for column i*/
+} QueryType;
+
+typedef enum fulltext_statement {
+ CONTENT_INSERT_STMT,
+ CONTENT_SELECT_STMT,
+ CONTENT_UPDATE_STMT,
+ CONTENT_DELETE_STMT,
+ CONTENT_EXISTS_STMT,
+
+ BLOCK_INSERT_STMT,
+ BLOCK_SELECT_STMT,
+ BLOCK_DELETE_STMT,
+ BLOCK_DELETE_ALL_STMT,
+
+ SEGDIR_MAX_INDEX_STMT,
+ SEGDIR_SET_STMT,
+ SEGDIR_SELECT_LEVEL_STMT,
+ SEGDIR_SPAN_STMT,
+ SEGDIR_DELETE_STMT,
+ SEGDIR_SELECT_SEGMENT_STMT,
+ SEGDIR_SELECT_ALL_STMT,
+ SEGDIR_DELETE_ALL_STMT,
+ SEGDIR_COUNT_STMT,
+
+ MAX_STMT /* Always at end! */
+} fulltext_statement;
+
+/* These must exactly match the enum above. */
+/* TODO(shess): Is there some risk that a statement will be used in two
+** cursors at once, e.g. if a query joins a virtual table to itself?
+** If so perhaps we should move some of these to the cursor object.
+*/
+static const char *const fulltext_zStatement[MAX_STMT] = {
+ /* CONTENT_INSERT */ NULL, /* generated in contentInsertStatement() */
+ /* CONTENT_SELECT */ NULL, /* generated in contentSelectStatement() */
+ /* CONTENT_UPDATE */ NULL, /* generated in contentUpdateStatement() */
+ /* CONTENT_DELETE */ "delete from %_content where docid = ?",
+ /* CONTENT_EXISTS */ "select docid from %_content limit 1",
+
+ /* BLOCK_INSERT */
+ "insert into %_segments (blockid, block) values (null, ?)",
+ /* BLOCK_SELECT */ "select block from %_segments where blockid = ?",
+ /* BLOCK_DELETE */ "delete from %_segments where blockid between ? and ?",
+ /* BLOCK_DELETE_ALL */ "delete from %_segments",
+
+ /* SEGDIR_MAX_INDEX */ "select max(idx) from %_segdir where level = ?",
+ /* SEGDIR_SET */ "insert into %_segdir values (?, ?, ?, ?, ?, ?)",
+ /* SEGDIR_SELECT_LEVEL */
+ "select start_block, leaves_end_block, root from %_segdir "
+ " where level = ? order by idx",
+ /* SEGDIR_SPAN */
+ "select min(start_block), max(end_block) from %_segdir "
+ " where level = ? and start_block <> 0",
+ /* SEGDIR_DELETE */ "delete from %_segdir where level = ?",
+
+ /* NOTE(shess): The first three results of the following two
+ ** statements must match.
+ */
+ /* SEGDIR_SELECT_SEGMENT */
+ "select start_block, leaves_end_block, root from %_segdir "
+ " where level = ? and idx = ?",
+ /* SEGDIR_SELECT_ALL */
+ "select start_block, leaves_end_block, root from %_segdir "
+ " order by level desc, idx asc",
+ /* SEGDIR_DELETE_ALL */ "delete from %_segdir",
+ /* SEGDIR_COUNT */ "select count(*), ifnull(max(level),0) from %_segdir",
+};
+
+/*
+** A connection to a fulltext index is an instance of the following
+** structure. The xCreate and xConnect methods create an instance
+** of this structure and xDestroy and xDisconnect free that instance.
+** All other methods receive a pointer to the structure as one of their
+** arguments.
+*/
+struct fulltext_vtab {
+ sqlite3_vtab base; /* Base class used by SQLite core */
+ sqlite3 *db; /* The database connection */
+ const char *zDb; /* logical database name */
+ const char *zName; /* virtual table name */
+ int nColumn; /* number of columns in virtual table */
+ char **azColumn; /* column names. malloced */
+ char **azContentColumn; /* column names in content table; malloced */
+ sqlite3_tokenizer *pTokenizer; /* tokenizer for inserts and queries */
+
+ /* Precompiled statements which we keep as long as the table is
+ ** open.
+ */
+ sqlite3_stmt *pFulltextStatements[MAX_STMT];
+
+ /* Precompiled statements used for segment merges. We run a
+ ** separate select across the leaf level of each tree being merged.
+ */
+ sqlite3_stmt *pLeafSelectStmts[MERGE_COUNT];
+ /* The statement used to prepare pLeafSelectStmts. */
+#define LEAF_SELECT \
+ "select block from %_segments where blockid between ? and ? order by blockid"
+
+ /* These buffer pending index updates during transactions.
+ ** nPendingData estimates the memory size of the pending data. It
+ ** doesn't include the hash-bucket overhead, nor any malloc
+ ** overhead. When nPendingData exceeds kPendingThreshold, the
+ ** buffer is flushed even before the transaction closes.
+ ** pendingTerms stores the data, and is only valid when nPendingData
+ ** is >=0 (nPendingData<0 means pendingTerms has not been
+ ** initialized). iPrevDocid is the last docid written, used to make
+ ** certain we're inserting in sorted order.
+ */
+ int nPendingData;
+#define kPendingThreshold (1*1024*1024)
+ sqlite_int64 iPrevDocid;
+ fts3Hash pendingTerms;
+};
+
+/*
+** When the core wants to do a query, it create a cursor using a
+** call to xOpen. This structure is an instance of a cursor. It
+** is destroyed by xClose.
+*/
+typedef struct fulltext_cursor {
+ sqlite3_vtab_cursor base; /* Base class used by SQLite core */
+ QueryType iCursorType; /* Copy of sqlite3_index_info.idxNum */
+ sqlite3_stmt *pStmt; /* Prepared statement in use by the cursor */
+ int eof; /* True if at End Of Results */
+ Query q; /* Parsed query string */
+ Snippet snippet; /* Cached snippet for the current row */
+ int iColumn; /* Column being searched */
+ DataBuffer result; /* Doclist results from fulltextQuery */
+ DLReader reader; /* Result reader if result not empty */
+} fulltext_cursor;
+
+static struct fulltext_vtab *cursor_vtab(fulltext_cursor *c){
+ return (fulltext_vtab *) c->base.pVtab;
+}
+
+static const sqlite3_module fts3Module; /* forward declaration */
+
+/* Return a dynamically generated statement of the form
+ * insert into %_content (docid, ...) values (?, ...)
+ */
+static const char *contentInsertStatement(fulltext_vtab *v){
+ StringBuffer sb;
+ int i;
+
+ initStringBuffer(&sb);
+ append(&sb, "insert into %_content (docid, ");
+ appendList(&sb, v->nColumn, v->azContentColumn);
+ append(&sb, ") values (?");
+ for(i=0; i<v->nColumn; ++i)
+ append(&sb, ", ?");
+ append(&sb, ")");
+ return stringBufferData(&sb);
+}
+
+/* Return a dynamically generated statement of the form
+ * select <content columns> from %_content where docid = ?
+ */
+static const char *contentSelectStatement(fulltext_vtab *v){
+ StringBuffer sb;
+ initStringBuffer(&sb);
+ append(&sb, "SELECT ");
+ appendList(&sb, v->nColumn, v->azContentColumn);
+ append(&sb, " FROM %_content WHERE docid = ?");
+ return stringBufferData(&sb);
+}
+
+/* Return a dynamically generated statement of the form
+ * update %_content set [col_0] = ?, [col_1] = ?, ...
+ * where docid = ?
+ */
+static const char *contentUpdateStatement(fulltext_vtab *v){
+ StringBuffer sb;
+ int i;
+
+ initStringBuffer(&sb);
+ append(&sb, "update %_content set ");
+ for(i=0; i<v->nColumn; ++i) {
+ if( i>0 ){
+ append(&sb, ", ");
+ }
+ append(&sb, v->azContentColumn[i]);
+ append(&sb, " = ?");
+ }
+ append(&sb, " where docid = ?");
+ return stringBufferData(&sb);
+}
+
+/* Puts a freshly-prepared statement determined by iStmt in *ppStmt.
+** If the indicated statement has never been prepared, it is prepared
+** and cached, otherwise the cached version is reset.
+*/
+static int sql_get_statement(fulltext_vtab *v, fulltext_statement iStmt,
+ sqlite3_stmt **ppStmt){
+ assert( iStmt<MAX_STMT );
+ if( v->pFulltextStatements[iStmt]==NULL ){
+ const char *zStmt;
+ int rc;
+ switch( iStmt ){
+ case CONTENT_INSERT_STMT:
+ zStmt = contentInsertStatement(v); break;
+ case CONTENT_SELECT_STMT:
+ zStmt = contentSelectStatement(v); break;
+ case CONTENT_UPDATE_STMT:
+ zStmt = contentUpdateStatement(v); break;
+ default:
+ zStmt = fulltext_zStatement[iStmt];
+ }
+ rc = sql_prepare(v->db, v->zDb, v->zName, &v->pFulltextStatements[iStmt],
+ zStmt);
+ if( zStmt != fulltext_zStatement[iStmt]) sqlite3_free((void *) zStmt);
+ if( rc!=SQLITE_OK ) return rc;
+ } else {
+ int rc = sqlite3_reset(v->pFulltextStatements[iStmt]);
+ if( rc!=SQLITE_OK ) return rc;
+ }
+
+ *ppStmt = v->pFulltextStatements[iStmt];
+ return SQLITE_OK;
+}
+
+/* Like sqlite3_step(), but convert SQLITE_DONE to SQLITE_OK and
+** SQLITE_ROW to SQLITE_ERROR. Useful for statements like UPDATE,
+** where we expect no results.
+*/
+static int sql_single_step(sqlite3_stmt *s){
+ int rc = sqlite3_step(s);
+ return (rc==SQLITE_DONE) ? SQLITE_OK : rc;
+}
+
+/* Like sql_get_statement(), but for special replicated LEAF_SELECT
+** statements. idx -1 is a special case for an uncached version of
+** the statement (used in the optimize implementation).
+*/
+/* TODO(shess) Write version for generic statements and then share
+** that between the cached-statement functions.
+*/
+static int sql_get_leaf_statement(fulltext_vtab *v, int idx,
+ sqlite3_stmt **ppStmt){
+ assert( idx>=-1 && idx<MERGE_COUNT );
+ if( idx==-1 ){
+ return sql_prepare(v->db, v->zDb, v->zName, ppStmt, LEAF_SELECT);
+ }else if( v->pLeafSelectStmts[idx]==NULL ){
+ int rc = sql_prepare(v->db, v->zDb, v->zName, &v->pLeafSelectStmts[idx],
+ LEAF_SELECT);
+ if( rc!=SQLITE_OK ) return rc;
+ }else{
+ int rc = sqlite3_reset(v->pLeafSelectStmts[idx]);
+ if( rc!=SQLITE_OK ) return rc;
+ }