blob: 6b81c363ffb8cfd6bc0d782d80e33f034736dd4b [file] [log] [blame]
drh75897232000-05-29 14:26:00 +00001/*
drhb19a2bc2001-09-16 00:13:26 +00002** 2001 September 15
drh75897232000-05-29 14:26:00 +00003**
drhb19a2bc2001-09-16 00:13:26 +00004** The author disclaims copyright to this source code. In place of
5** a legal notice, here is a blessing:
drh75897232000-05-29 14:26:00 +00006**
drhb19a2bc2001-09-16 00:13:26 +00007** 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.
drh75897232000-05-29 14:26:00 +000010**
11*************************************************************************
12** This module contains C code that generates VDBE code used to process
drh909626d2008-05-30 14:58:37 +000013** the WHERE clause of SQL statements. This module is responsible for
drh51669862004-12-18 18:40:26 +000014** generating the code that loops through a table looking for applicable
15** rows. Indices are selected and used to speed the search when doing
16** so is applicable. Because this module is responsible for selecting
17** indices, you might also think of this module as the "query optimizer".
drh75897232000-05-29 14:26:00 +000018**
drh61495262009-04-22 15:32:59 +000019** $Id: where.c,v 1.386 2009/04/22 15:32:59 drh Exp $
drh75897232000-05-29 14:26:00 +000020*/
21#include "sqliteInt.h"
22
23/*
drh51147ba2005-07-23 22:59:55 +000024** Trace output macros
25*/
26#if defined(SQLITE_TEST) || defined(SQLITE_DEBUG)
mlcreech3a00f902008-03-04 17:45:01 +000027int sqlite3WhereTrace = 0;
drhe8f52c52008-07-12 14:52:20 +000028#endif
drh85799a42009-04-07 13:48:11 +000029#if defined(SQLITE_TEST) && defined(SQLITE_DEBUG)
mlcreech3a00f902008-03-04 17:45:01 +000030# define WHERETRACE(X) if(sqlite3WhereTrace) sqlite3DebugPrintf X
drh51147ba2005-07-23 22:59:55 +000031#else
drh4f0c5872007-03-26 22:05:01 +000032# define WHERETRACE(X)
drh51147ba2005-07-23 22:59:55 +000033#endif
34
drh0fcef5e2005-07-19 17:38:22 +000035/* Forward reference
36*/
37typedef struct WhereClause WhereClause;
drh111a6a72008-12-21 03:51:16 +000038typedef struct WhereMaskSet WhereMaskSet;
drh700a2262008-12-17 19:22:15 +000039typedef struct WhereOrInfo WhereOrInfo;
40typedef struct WhereAndInfo WhereAndInfo;
drh111a6a72008-12-21 03:51:16 +000041typedef struct WhereCost WhereCost;
drh0aa74ed2005-07-16 13:33:20 +000042
43/*
drh75897232000-05-29 14:26:00 +000044** The query generator uses an array of instances of this structure to
45** help it analyze the subexpressions of the WHERE clause. Each WHERE
drh61495262009-04-22 15:32:59 +000046** clause subexpression is separated from the others by AND operators,
47** usually, or sometimes subexpressions separated by OR.
drh51669862004-12-18 18:40:26 +000048**
drh0fcef5e2005-07-19 17:38:22 +000049** All WhereTerms are collected into a single WhereClause structure.
50** The following identity holds:
drh51669862004-12-18 18:40:26 +000051**
drh0fcef5e2005-07-19 17:38:22 +000052** WhereTerm.pWC->a[WhereTerm.idx] == WhereTerm
drh51669862004-12-18 18:40:26 +000053**
drh0fcef5e2005-07-19 17:38:22 +000054** When a term is of the form:
55**
56** X <op> <expr>
57**
58** where X is a column name and <op> is one of certain operators,
drh700a2262008-12-17 19:22:15 +000059** then WhereTerm.leftCursor and WhereTerm.u.leftColumn record the
60** cursor number and column number for X. WhereTerm.eOperator records
drh51147ba2005-07-23 22:59:55 +000061** the <op> using a bitmask encoding defined by WO_xxx below. The
62** use of a bitmask encoding for the operator allows us to search
63** quickly for terms that match any of several different operators.
drh0fcef5e2005-07-19 17:38:22 +000064**
drh700a2262008-12-17 19:22:15 +000065** A WhereTerm might also be two or more subterms connected by OR:
66**
67** (t1.X <op> <expr>) OR (t1.Y <op> <expr>) OR ....
68**
69** In this second case, wtFlag as the TERM_ORINFO set and eOperator==WO_OR
70** and the WhereTerm.u.pOrInfo field points to auxiliary information that
71** is collected about the
72**
73** If a term in the WHERE clause does not match either of the two previous
74** categories, then eOperator==0. The WhereTerm.pExpr field is still set
75** to the original subexpression content and wtFlags is set up appropriately
76** but no other fields in the WhereTerm object are meaningful.
77**
78** When eOperator!=0, prereqRight and prereqAll record sets of cursor numbers,
drh111a6a72008-12-21 03:51:16 +000079** but they do so indirectly. A single WhereMaskSet structure translates
drh51669862004-12-18 18:40:26 +000080** cursor number into bits and the translated bit is stored in the prereq
81** fields. The translation is used in order to maximize the number of
82** bits that will fit in a Bitmask. The VDBE cursor numbers might be
83** spread out over the non-negative integers. For example, the cursor
drh111a6a72008-12-21 03:51:16 +000084** numbers might be 3, 8, 9, 10, 20, 23, 41, and 45. The WhereMaskSet
drh51669862004-12-18 18:40:26 +000085** translates these sparse cursor numbers into consecutive integers
86** beginning with 0 in order to make the best possible use of the available
87** bits in the Bitmask. So, in the example above, the cursor numbers
88** would be mapped into integers 0 through 7.
drh6a1e0712008-12-05 15:24:15 +000089**
90** The number of terms in a join is limited by the number of bits
91** in prereqRight and prereqAll. The default is 64 bits, hence SQLite
92** is only able to process joins with 64 or fewer tables.
drh75897232000-05-29 14:26:00 +000093*/
drh0aa74ed2005-07-16 13:33:20 +000094typedef struct WhereTerm WhereTerm;
95struct WhereTerm {
drh165be382008-12-05 02:36:33 +000096 Expr *pExpr; /* Pointer to the subexpression that is this term */
drhec1724e2008-12-09 01:32:03 +000097 int iParent; /* Disable pWC->a[iParent] when this term disabled */
98 int leftCursor; /* Cursor number of X in "X <op> <expr>" */
drh700a2262008-12-17 19:22:15 +000099 union {
100 int leftColumn; /* Column number of X in "X <op> <expr>" */
101 WhereOrInfo *pOrInfo; /* Extra information if eOperator==WO_OR */
102 WhereAndInfo *pAndInfo; /* Extra information if eOperator==WO_AND */
103 } u;
drhb52076c2006-01-23 13:22:09 +0000104 u16 eOperator; /* A WO_xx value describing <op> */
drh165be382008-12-05 02:36:33 +0000105 u8 wtFlags; /* TERM_xxx bit flags. See below */
drh45b1ee42005-08-02 17:48:22 +0000106 u8 nChild; /* Number of children that must disable us */
drh0fcef5e2005-07-19 17:38:22 +0000107 WhereClause *pWC; /* The clause this term is part of */
drh165be382008-12-05 02:36:33 +0000108 Bitmask prereqRight; /* Bitmask of tables used by pExpr->pRight */
109 Bitmask prereqAll; /* Bitmask of tables referenced by pExpr */
drh75897232000-05-29 14:26:00 +0000110};
111
112/*
drh165be382008-12-05 02:36:33 +0000113** Allowed values of WhereTerm.wtFlags
drh0aa74ed2005-07-16 13:33:20 +0000114*/
drh633e6d52008-07-28 19:34:53 +0000115#define TERM_DYNAMIC 0x01 /* Need to call sqlite3ExprDelete(db, pExpr) */
drh6c30be82005-07-29 15:10:17 +0000116#define TERM_VIRTUAL 0x02 /* Added by the optimizer. Do not code */
117#define TERM_CODED 0x04 /* This term is already coded */
drh45b1ee42005-08-02 17:48:22 +0000118#define TERM_COPIED 0x08 /* Has a child */
drh700a2262008-12-17 19:22:15 +0000119#define TERM_ORINFO 0x10 /* Need to free the WhereTerm.u.pOrInfo object */
120#define TERM_ANDINFO 0x20 /* Need to free the WhereTerm.u.pAndInfo obj */
121#define TERM_OR_OK 0x40 /* Used during OR-clause processing */
drh0aa74ed2005-07-16 13:33:20 +0000122
123/*
124** An instance of the following structure holds all information about a
125** WHERE clause. Mostly this is a container for one or more WhereTerms.
126*/
drh0aa74ed2005-07-16 13:33:20 +0000127struct WhereClause {
drhfe05af82005-07-21 03:14:59 +0000128 Parse *pParse; /* The parser context */
drh111a6a72008-12-21 03:51:16 +0000129 WhereMaskSet *pMaskSet; /* Mapping of table cursor numbers to bitmasks */
drh29435252008-12-28 18:35:08 +0000130 u8 op; /* Split operator. TK_AND or TK_OR */
drh0aa74ed2005-07-16 13:33:20 +0000131 int nTerm; /* Number of terms */
132 int nSlot; /* Number of entries in a[] */
drh51147ba2005-07-23 22:59:55 +0000133 WhereTerm *a; /* Each a[] describes a term of the WHERE cluase */
drhec1724e2008-12-09 01:32:03 +0000134 WhereTerm aStatic[4]; /* Initial static space for a[] */
drhe23399f2005-07-22 00:31:39 +0000135};
136
137/*
drh700a2262008-12-17 19:22:15 +0000138** A WhereTerm with eOperator==WO_OR has its u.pOrInfo pointer set to
139** a dynamically allocated instance of the following structure.
140*/
141struct WhereOrInfo {
drh111a6a72008-12-21 03:51:16 +0000142 WhereClause wc; /* Decomposition into subterms */
drh1a58fe02008-12-20 02:06:13 +0000143 Bitmask indexable; /* Bitmask of all indexable tables in the clause */
drh700a2262008-12-17 19:22:15 +0000144};
145
146/*
147** A WhereTerm with eOperator==WO_AND has its u.pAndInfo pointer set to
148** a dynamically allocated instance of the following structure.
149*/
150struct WhereAndInfo {
drh29435252008-12-28 18:35:08 +0000151 WhereClause wc; /* The subexpression broken out */
drh700a2262008-12-17 19:22:15 +0000152};
153
154/*
drh6a3ea0e2003-05-02 14:32:12 +0000155** An instance of the following structure keeps track of a mapping
drh0aa74ed2005-07-16 13:33:20 +0000156** between VDBE cursor numbers and bits of the bitmasks in WhereTerm.
drh51669862004-12-18 18:40:26 +0000157**
158** The VDBE cursor numbers are small integers contained in
159** SrcList_item.iCursor and Expr.iTable fields. For any given WHERE
160** clause, the cursor numbers might not begin with 0 and they might
161** contain gaps in the numbering sequence. But we want to make maximum
162** use of the bits in our bitmasks. This structure provides a mapping
163** from the sparse cursor numbers into consecutive integers beginning
164** with 0.
165**
drh111a6a72008-12-21 03:51:16 +0000166** If WhereMaskSet.ix[A]==B it means that The A-th bit of a Bitmask
drh51669862004-12-18 18:40:26 +0000167** corresponds VDBE cursor number B. The A-th bit of a bitmask is 1<<A.
168**
169** For example, if the WHERE clause expression used these VDBE
drh111a6a72008-12-21 03:51:16 +0000170** cursors: 4, 5, 8, 29, 57, 73. Then the WhereMaskSet structure
drh51669862004-12-18 18:40:26 +0000171** would map those cursor numbers into bits 0 through 5.
172**
173** Note that the mapping is not necessarily ordered. In the example
174** above, the mapping might go like this: 4->3, 5->1, 8->2, 29->0,
175** 57->5, 73->4. Or one of 719 other combinations might be used. It
176** does not really matter. What is important is that sparse cursor
177** numbers all get mapped into bit numbers that begin with 0 and contain
178** no gaps.
drh6a3ea0e2003-05-02 14:32:12 +0000179*/
drh111a6a72008-12-21 03:51:16 +0000180struct WhereMaskSet {
drh1398ad32005-01-19 23:24:50 +0000181 int n; /* Number of assigned cursor values */
danielk197723432972008-11-17 16:42:00 +0000182 int ix[BMS]; /* Cursor assigned to each bit */
drh6a3ea0e2003-05-02 14:32:12 +0000183};
184
drh111a6a72008-12-21 03:51:16 +0000185/*
186** A WhereCost object records a lookup strategy and the estimated
187** cost of pursuing that strategy.
188*/
189struct WhereCost {
190 WherePlan plan; /* The lookup strategy */
191 double rCost; /* Overall cost of pursuing this search strategy */
192 double nRow; /* Estimated number of output rows */
193};
drh0aa74ed2005-07-16 13:33:20 +0000194
drh6a3ea0e2003-05-02 14:32:12 +0000195/*
drh51147ba2005-07-23 22:59:55 +0000196** Bitmasks for the operators that indices are able to exploit. An
197** OR-ed combination of these values can be used when searching for
198** terms in the where clause.
199*/
drh165be382008-12-05 02:36:33 +0000200#define WO_IN 0x001
201#define WO_EQ 0x002
drh51147ba2005-07-23 22:59:55 +0000202#define WO_LT (WO_EQ<<(TK_LT-TK_EQ))
203#define WO_LE (WO_EQ<<(TK_LE-TK_EQ))
204#define WO_GT (WO_EQ<<(TK_GT-TK_EQ))
205#define WO_GE (WO_EQ<<(TK_GE-TK_EQ))
drh165be382008-12-05 02:36:33 +0000206#define WO_MATCH 0x040
207#define WO_ISNULL 0x080
drh700a2262008-12-17 19:22:15 +0000208#define WO_OR 0x100 /* Two or more OR-connected terms */
209#define WO_AND 0x200 /* Two or more AND-connected terms */
drh51147ba2005-07-23 22:59:55 +0000210
drhec1724e2008-12-09 01:32:03 +0000211#define WO_ALL 0xfff /* Mask of all possible WO_* values */
drh1a58fe02008-12-20 02:06:13 +0000212#define WO_SINGLE 0x0ff /* Mask of all non-compound WO_* values */
drhec1724e2008-12-09 01:32:03 +0000213
drh51147ba2005-07-23 22:59:55 +0000214/*
drh700a2262008-12-17 19:22:15 +0000215** Value for wsFlags returned by bestIndex() and stored in
216** WhereLevel.wsFlags. These flags determine which search
217** strategies are appropriate.
drhf2d315d2007-01-25 16:56:06 +0000218**
drh165be382008-12-05 02:36:33 +0000219** The least significant 12 bits is reserved as a mask for WO_ values above.
drh700a2262008-12-17 19:22:15 +0000220** The WhereLevel.wsFlags field is usually set to WO_IN|WO_EQ|WO_ISNULL.
221** But if the table is the right table of a left join, WhereLevel.wsFlags
222** is set to WO_IN|WO_EQ. The WhereLevel.wsFlags field can then be used as
drhf2d315d2007-01-25 16:56:06 +0000223** the "op" parameter to findTerm when we are resolving equality constraints.
224** ISNULL constraints will then not be used on the right table of a left
225** join. Tickets #2177 and #2189.
drh51147ba2005-07-23 22:59:55 +0000226*/
drh165be382008-12-05 02:36:33 +0000227#define WHERE_ROWID_EQ 0x00001000 /* rowid=EXPR or rowid IN (...) */
228#define WHERE_ROWID_RANGE 0x00002000 /* rowid<EXPR and/or rowid>EXPR */
229#define WHERE_COLUMN_EQ 0x00010000 /* x=EXPR or x IN (...) */
230#define WHERE_COLUMN_RANGE 0x00020000 /* x<EXPR and/or x>EXPR */
231#define WHERE_COLUMN_IN 0x00040000 /* x IN (...) */
drh111a6a72008-12-21 03:51:16 +0000232#define WHERE_INDEXED 0x00070000 /* Anything that uses an index */
233#define WHERE_IN_ABLE 0x00071000 /* Able to support an IN operator */
drh165be382008-12-05 02:36:33 +0000234#define WHERE_TOP_LIMIT 0x00100000 /* x<EXPR or x<=EXPR constraint */
235#define WHERE_BTM_LIMIT 0x00200000 /* x>EXPR or x>=EXPR constraint */
236#define WHERE_IDX_ONLY 0x00800000 /* Use index only - omit table */
237#define WHERE_ORDERBY 0x01000000 /* Output will appear in correct order */
238#define WHERE_REVERSE 0x02000000 /* Scan in reverse order */
239#define WHERE_UNIQUE 0x04000000 /* Selects no more than one row */
240#define WHERE_VIRTUALTABLE 0x08000000 /* Use virtual-table processing */
241#define WHERE_MULTI_OR 0x10000000 /* OR using multiple indices */
drh51147ba2005-07-23 22:59:55 +0000242
243/*
drh0aa74ed2005-07-16 13:33:20 +0000244** Initialize a preallocated WhereClause structure.
drh75897232000-05-29 14:26:00 +0000245*/
drh7b4fc6a2007-02-06 13:26:32 +0000246static void whereClauseInit(
247 WhereClause *pWC, /* The WhereClause to be initialized */
248 Parse *pParse, /* The parsing context */
drh111a6a72008-12-21 03:51:16 +0000249 WhereMaskSet *pMaskSet /* Mapping from table cursor numbers to bitmasks */
drh7b4fc6a2007-02-06 13:26:32 +0000250){
drhfe05af82005-07-21 03:14:59 +0000251 pWC->pParse = pParse;
drh7b4fc6a2007-02-06 13:26:32 +0000252 pWC->pMaskSet = pMaskSet;
drh0aa74ed2005-07-16 13:33:20 +0000253 pWC->nTerm = 0;
drhcad651e2007-04-20 12:22:01 +0000254 pWC->nSlot = ArraySize(pWC->aStatic);
drh0aa74ed2005-07-16 13:33:20 +0000255 pWC->a = pWC->aStatic;
256}
257
drh700a2262008-12-17 19:22:15 +0000258/* Forward reference */
259static void whereClauseClear(WhereClause*);
260
261/*
262** Deallocate all memory associated with a WhereOrInfo object.
263*/
264static void whereOrInfoDelete(sqlite3 *db, WhereOrInfo *p){
drh5bd98ae2009-01-07 18:24:03 +0000265 whereClauseClear(&p->wc);
266 sqlite3DbFree(db, p);
drh700a2262008-12-17 19:22:15 +0000267}
268
269/*
270** Deallocate all memory associated with a WhereAndInfo object.
271*/
272static void whereAndInfoDelete(sqlite3 *db, WhereAndInfo *p){
drh5bd98ae2009-01-07 18:24:03 +0000273 whereClauseClear(&p->wc);
274 sqlite3DbFree(db, p);
drh700a2262008-12-17 19:22:15 +0000275}
276
drh0aa74ed2005-07-16 13:33:20 +0000277/*
278** Deallocate a WhereClause structure. The WhereClause structure
279** itself is not freed. This routine is the inverse of whereClauseInit().
280*/
281static void whereClauseClear(WhereClause *pWC){
282 int i;
283 WhereTerm *a;
drh633e6d52008-07-28 19:34:53 +0000284 sqlite3 *db = pWC->pParse->db;
drh0aa74ed2005-07-16 13:33:20 +0000285 for(i=pWC->nTerm-1, a=pWC->a; i>=0; i--, a++){
drh165be382008-12-05 02:36:33 +0000286 if( a->wtFlags & TERM_DYNAMIC ){
drh633e6d52008-07-28 19:34:53 +0000287 sqlite3ExprDelete(db, a->pExpr);
drh0aa74ed2005-07-16 13:33:20 +0000288 }
drh700a2262008-12-17 19:22:15 +0000289 if( a->wtFlags & TERM_ORINFO ){
290 whereOrInfoDelete(db, a->u.pOrInfo);
291 }else if( a->wtFlags & TERM_ANDINFO ){
292 whereAndInfoDelete(db, a->u.pAndInfo);
293 }
drh0aa74ed2005-07-16 13:33:20 +0000294 }
295 if( pWC->a!=pWC->aStatic ){
drh633e6d52008-07-28 19:34:53 +0000296 sqlite3DbFree(db, pWC->a);
drh0aa74ed2005-07-16 13:33:20 +0000297 }
298}
299
300/*
drh6a1e0712008-12-05 15:24:15 +0000301** Add a single new WhereTerm entry to the WhereClause object pWC.
302** The new WhereTerm object is constructed from Expr p and with wtFlags.
303** The index in pWC->a[] of the new WhereTerm is returned on success.
304** 0 is returned if the new WhereTerm could not be added due to a memory
305** allocation error. The memory allocation failure will be recorded in
306** the db->mallocFailed flag so that higher-level functions can detect it.
307**
308** This routine will increase the size of the pWC->a[] array as necessary.
drh9eb20282005-08-24 03:52:18 +0000309**
drh165be382008-12-05 02:36:33 +0000310** If the wtFlags argument includes TERM_DYNAMIC, then responsibility
drh6a1e0712008-12-05 15:24:15 +0000311** for freeing the expression p is assumed by the WhereClause object pWC.
312** This is true even if this routine fails to allocate a new WhereTerm.
drhb63a53d2007-03-31 01:34:44 +0000313**
drh9eb20282005-08-24 03:52:18 +0000314** WARNING: This routine might reallocate the space used to store
drh909626d2008-05-30 14:58:37 +0000315** WhereTerms. All pointers to WhereTerms should be invalidated after
drh9eb20282005-08-24 03:52:18 +0000316** calling this routine. Such pointers may be reinitialized by referencing
317** the pWC->a[] array.
drh0aa74ed2005-07-16 13:33:20 +0000318*/
drhec1724e2008-12-09 01:32:03 +0000319static int whereClauseInsert(WhereClause *pWC, Expr *p, u8 wtFlags){
drh0aa74ed2005-07-16 13:33:20 +0000320 WhereTerm *pTerm;
drh9eb20282005-08-24 03:52:18 +0000321 int idx;
drh0aa74ed2005-07-16 13:33:20 +0000322 if( pWC->nTerm>=pWC->nSlot ){
323 WhereTerm *pOld = pWC->a;
drh633e6d52008-07-28 19:34:53 +0000324 sqlite3 *db = pWC->pParse->db;
325 pWC->a = sqlite3DbMallocRaw(db, sizeof(pWC->a[0])*pWC->nSlot*2 );
drhb63a53d2007-03-31 01:34:44 +0000326 if( pWC->a==0 ){
drh165be382008-12-05 02:36:33 +0000327 if( wtFlags & TERM_DYNAMIC ){
drh633e6d52008-07-28 19:34:53 +0000328 sqlite3ExprDelete(db, p);
drhb63a53d2007-03-31 01:34:44 +0000329 }
drhf998b732007-11-26 13:36:00 +0000330 pWC->a = pOld;
drhb63a53d2007-03-31 01:34:44 +0000331 return 0;
332 }
drh0aa74ed2005-07-16 13:33:20 +0000333 memcpy(pWC->a, pOld, sizeof(pWC->a[0])*pWC->nTerm);
334 if( pOld!=pWC->aStatic ){
drh633e6d52008-07-28 19:34:53 +0000335 sqlite3DbFree(db, pOld);
drh0aa74ed2005-07-16 13:33:20 +0000336 }
drh6a1e0712008-12-05 15:24:15 +0000337 pWC->nSlot = sqlite3DbMallocSize(db, pWC->a)/sizeof(pWC->a[0]);
drh0aa74ed2005-07-16 13:33:20 +0000338 }
drh6a1e0712008-12-05 15:24:15 +0000339 pTerm = &pWC->a[idx = pWC->nTerm++];
drh0fcef5e2005-07-19 17:38:22 +0000340 pTerm->pExpr = p;
drh165be382008-12-05 02:36:33 +0000341 pTerm->wtFlags = wtFlags;
drh0fcef5e2005-07-19 17:38:22 +0000342 pTerm->pWC = pWC;
drh45b1ee42005-08-02 17:48:22 +0000343 pTerm->iParent = -1;
drh9eb20282005-08-24 03:52:18 +0000344 return idx;
drh0aa74ed2005-07-16 13:33:20 +0000345}
drh75897232000-05-29 14:26:00 +0000346
347/*
drh51669862004-12-18 18:40:26 +0000348** This routine identifies subexpressions in the WHERE clause where
drhb6fb62d2005-09-20 08:47:20 +0000349** each subexpression is separated by the AND operator or some other
drh6c30be82005-07-29 15:10:17 +0000350** operator specified in the op parameter. The WhereClause structure
351** is filled with pointers to subexpressions. For example:
drh75897232000-05-29 14:26:00 +0000352**
drh51669862004-12-18 18:40:26 +0000353** WHERE a=='hello' AND coalesce(b,11)<10 AND (c+12!=d OR c==22)
354** \________/ \_______________/ \________________/
355** slot[0] slot[1] slot[2]
356**
357** The original WHERE clause in pExpr is unaltered. All this routine
drh51147ba2005-07-23 22:59:55 +0000358** does is make slot[] entries point to substructure within pExpr.
drh51669862004-12-18 18:40:26 +0000359**
drh51147ba2005-07-23 22:59:55 +0000360** In the previous sentence and in the diagram, "slot[]" refers to
drh902b9ee2008-12-05 17:17:07 +0000361** the WhereClause.a[] array. The slot[] array grows as needed to contain
drh51147ba2005-07-23 22:59:55 +0000362** all terms of the WHERE clause.
drh75897232000-05-29 14:26:00 +0000363*/
drh6c30be82005-07-29 15:10:17 +0000364static void whereSplit(WhereClause *pWC, Expr *pExpr, int op){
drh29435252008-12-28 18:35:08 +0000365 pWC->op = (u8)op;
drh0aa74ed2005-07-16 13:33:20 +0000366 if( pExpr==0 ) return;
drh6c30be82005-07-29 15:10:17 +0000367 if( pExpr->op!=op ){
drh0aa74ed2005-07-16 13:33:20 +0000368 whereClauseInsert(pWC, pExpr, 0);
drh75897232000-05-29 14:26:00 +0000369 }else{
drh6c30be82005-07-29 15:10:17 +0000370 whereSplit(pWC, pExpr->pLeft, op);
371 whereSplit(pWC, pExpr->pRight, op);
drh75897232000-05-29 14:26:00 +0000372 }
drh75897232000-05-29 14:26:00 +0000373}
374
375/*
drh61495262009-04-22 15:32:59 +0000376** Initialize an expression mask set (a WhereMaskSet object)
drh6a3ea0e2003-05-02 14:32:12 +0000377*/
378#define initMaskSet(P) memset(P, 0, sizeof(*P))
379
380/*
drh1398ad32005-01-19 23:24:50 +0000381** Return the bitmask for the given cursor number. Return 0 if
382** iCursor is not in the set.
drh6a3ea0e2003-05-02 14:32:12 +0000383*/
drh111a6a72008-12-21 03:51:16 +0000384static Bitmask getMask(WhereMaskSet *pMaskSet, int iCursor){
drh6a3ea0e2003-05-02 14:32:12 +0000385 int i;
386 for(i=0; i<pMaskSet->n; i++){
drh51669862004-12-18 18:40:26 +0000387 if( pMaskSet->ix[i]==iCursor ){
388 return ((Bitmask)1)<<i;
389 }
drh6a3ea0e2003-05-02 14:32:12 +0000390 }
drh6a3ea0e2003-05-02 14:32:12 +0000391 return 0;
392}
393
394/*
drh1398ad32005-01-19 23:24:50 +0000395** Create a new mask for cursor iCursor.
drh0fcef5e2005-07-19 17:38:22 +0000396**
397** There is one cursor per table in the FROM clause. The number of
398** tables in the FROM clause is limited by a test early in the
drhb6fb62d2005-09-20 08:47:20 +0000399** sqlite3WhereBegin() routine. So we know that the pMaskSet->ix[]
drh0fcef5e2005-07-19 17:38:22 +0000400** array will never overflow.
drh1398ad32005-01-19 23:24:50 +0000401*/
drh111a6a72008-12-21 03:51:16 +0000402static void createMask(WhereMaskSet *pMaskSet, int iCursor){
drhcad651e2007-04-20 12:22:01 +0000403 assert( pMaskSet->n < ArraySize(pMaskSet->ix) );
drh0fcef5e2005-07-19 17:38:22 +0000404 pMaskSet->ix[pMaskSet->n++] = iCursor;
drh1398ad32005-01-19 23:24:50 +0000405}
406
407/*
drh75897232000-05-29 14:26:00 +0000408** This routine walks (recursively) an expression tree and generates
409** a bitmask indicating which tables are used in that expression
drh6a3ea0e2003-05-02 14:32:12 +0000410** tree.
drh75897232000-05-29 14:26:00 +0000411**
412** In order for this routine to work, the calling function must have
drh7d10d5a2008-08-20 16:35:10 +0000413** previously invoked sqlite3ResolveExprNames() on the expression. See
drh75897232000-05-29 14:26:00 +0000414** the header comment on that routine for additional information.
drh7d10d5a2008-08-20 16:35:10 +0000415** The sqlite3ResolveExprNames() routines looks for column names and
drh6a3ea0e2003-05-02 14:32:12 +0000416** sets their opcodes to TK_COLUMN and their Expr.iTable fields to
drh51147ba2005-07-23 22:59:55 +0000417** the VDBE cursor number of the table. This routine just has to
418** translate the cursor numbers into bitmask values and OR all
419** the bitmasks together.
drh75897232000-05-29 14:26:00 +0000420*/
drh111a6a72008-12-21 03:51:16 +0000421static Bitmask exprListTableUsage(WhereMaskSet*, ExprList*);
422static Bitmask exprSelectTableUsage(WhereMaskSet*, Select*);
423static Bitmask exprTableUsage(WhereMaskSet *pMaskSet, Expr *p){
drh51669862004-12-18 18:40:26 +0000424 Bitmask mask = 0;
drh75897232000-05-29 14:26:00 +0000425 if( p==0 ) return 0;
drh967e8b72000-06-21 13:59:10 +0000426 if( p->op==TK_COLUMN ){
drh8feb4b12004-07-19 02:12:14 +0000427 mask = getMask(pMaskSet, p->iTable);
drh8feb4b12004-07-19 02:12:14 +0000428 return mask;
drh75897232000-05-29 14:26:00 +0000429 }
danielk1977b3bce662005-01-29 08:32:43 +0000430 mask = exprTableUsage(pMaskSet, p->pRight);
431 mask |= exprTableUsage(pMaskSet, p->pLeft);
danielk19776ab3a2e2009-02-19 14:39:25 +0000432 if( ExprHasProperty(p, EP_xIsSelect) ){
433 mask |= exprSelectTableUsage(pMaskSet, p->x.pSelect);
434 }else{
435 mask |= exprListTableUsage(pMaskSet, p->x.pList);
436 }
danielk1977b3bce662005-01-29 08:32:43 +0000437 return mask;
438}
drh111a6a72008-12-21 03:51:16 +0000439static Bitmask exprListTableUsage(WhereMaskSet *pMaskSet, ExprList *pList){
danielk1977b3bce662005-01-29 08:32:43 +0000440 int i;
441 Bitmask mask = 0;
442 if( pList ){
443 for(i=0; i<pList->nExpr; i++){
444 mask |= exprTableUsage(pMaskSet, pList->a[i].pExpr);
drhdd579122002-04-02 01:58:57 +0000445 }
446 }
drh75897232000-05-29 14:26:00 +0000447 return mask;
448}
drh111a6a72008-12-21 03:51:16 +0000449static Bitmask exprSelectTableUsage(WhereMaskSet *pMaskSet, Select *pS){
drha430ae82007-09-12 15:41:01 +0000450 Bitmask mask = 0;
451 while( pS ){
452 mask |= exprListTableUsage(pMaskSet, pS->pEList);
drhf5b11382005-09-17 13:07:13 +0000453 mask |= exprListTableUsage(pMaskSet, pS->pGroupBy);
454 mask |= exprListTableUsage(pMaskSet, pS->pOrderBy);
455 mask |= exprTableUsage(pMaskSet, pS->pWhere);
456 mask |= exprTableUsage(pMaskSet, pS->pHaving);
drha430ae82007-09-12 15:41:01 +0000457 pS = pS->pPrior;
drhf5b11382005-09-17 13:07:13 +0000458 }
459 return mask;
460}
drh75897232000-05-29 14:26:00 +0000461
462/*
drh487ab3c2001-11-08 00:45:21 +0000463** Return TRUE if the given operator is one of the operators that is
drh51669862004-12-18 18:40:26 +0000464** allowed for an indexable WHERE clause term. The allowed operators are
drhc27a1ce2002-06-14 20:58:45 +0000465** "=", "<", ">", "<=", ">=", and "IN".
drh487ab3c2001-11-08 00:45:21 +0000466*/
467static int allowedOp(int op){
drhfe05af82005-07-21 03:14:59 +0000468 assert( TK_GT>TK_EQ && TK_GT<TK_GE );
469 assert( TK_LT>TK_EQ && TK_LT<TK_GE );
470 assert( TK_LE>TK_EQ && TK_LE<TK_GE );
471 assert( TK_GE==TK_EQ+4 );
drh50b39962006-10-28 00:28:09 +0000472 return op==TK_IN || (op>=TK_EQ && op<=TK_GE) || op==TK_ISNULL;
drh487ab3c2001-11-08 00:45:21 +0000473}
474
475/*
drh902b9ee2008-12-05 17:17:07 +0000476** Swap two objects of type TYPE.
drh193bd772004-07-20 18:23:14 +0000477*/
478#define SWAP(TYPE,A,B) {TYPE t=A; A=B; B=t;}
479
480/*
drh909626d2008-05-30 14:58:37 +0000481** Commute a comparison operator. Expressions of the form "X op Y"
drh0fcef5e2005-07-19 17:38:22 +0000482** are converted into "Y op X".
danielk1977eb5453d2007-07-30 14:40:48 +0000483**
484** If a collation sequence is associated with either the left or right
485** side of the comparison, it remains associated with the same side after
486** the commutation. So "Y collate NOCASE op X" becomes
487** "X collate NOCASE op Y". This is because any collation sequence on
488** the left hand side of a comparison overrides any collation sequence
489** attached to the right. For the same reason the EP_ExpCollate flag
490** is not commuted.
drh193bd772004-07-20 18:23:14 +0000491*/
drh7d10d5a2008-08-20 16:35:10 +0000492static void exprCommute(Parse *pParse, Expr *pExpr){
danielk1977eb5453d2007-07-30 14:40:48 +0000493 u16 expRight = (pExpr->pRight->flags & EP_ExpCollate);
494 u16 expLeft = (pExpr->pLeft->flags & EP_ExpCollate);
drhfe05af82005-07-21 03:14:59 +0000495 assert( allowedOp(pExpr->op) && pExpr->op!=TK_IN );
drh7d10d5a2008-08-20 16:35:10 +0000496 pExpr->pRight->pColl = sqlite3ExprCollSeq(pParse, pExpr->pRight);
497 pExpr->pLeft->pColl = sqlite3ExprCollSeq(pParse, pExpr->pLeft);
drh0fcef5e2005-07-19 17:38:22 +0000498 SWAP(CollSeq*,pExpr->pRight->pColl,pExpr->pLeft->pColl);
danielk1977eb5453d2007-07-30 14:40:48 +0000499 pExpr->pRight->flags = (pExpr->pRight->flags & ~EP_ExpCollate) | expLeft;
500 pExpr->pLeft->flags = (pExpr->pLeft->flags & ~EP_ExpCollate) | expRight;
drh0fcef5e2005-07-19 17:38:22 +0000501 SWAP(Expr*,pExpr->pRight,pExpr->pLeft);
502 if( pExpr->op>=TK_GT ){
503 assert( TK_LT==TK_GT+2 );
504 assert( TK_GE==TK_LE+2 );
505 assert( TK_GT>TK_EQ );
506 assert( TK_GT<TK_LE );
507 assert( pExpr->op>=TK_GT && pExpr->op<=TK_GE );
508 pExpr->op = ((pExpr->op-TK_GT)^2)+TK_GT;
drh193bd772004-07-20 18:23:14 +0000509 }
drh193bd772004-07-20 18:23:14 +0000510}
511
512/*
drhfe05af82005-07-21 03:14:59 +0000513** Translate from TK_xx operator to WO_xx bitmask.
514*/
drhec1724e2008-12-09 01:32:03 +0000515static u16 operatorMask(int op){
516 u16 c;
drhfe05af82005-07-21 03:14:59 +0000517 assert( allowedOp(op) );
518 if( op==TK_IN ){
drh51147ba2005-07-23 22:59:55 +0000519 c = WO_IN;
drh50b39962006-10-28 00:28:09 +0000520 }else if( op==TK_ISNULL ){
521 c = WO_ISNULL;
drhfe05af82005-07-21 03:14:59 +0000522 }else{
drhec1724e2008-12-09 01:32:03 +0000523 assert( (WO_EQ<<(op-TK_EQ)) < 0x7fff );
524 c = (u16)(WO_EQ<<(op-TK_EQ));
drhfe05af82005-07-21 03:14:59 +0000525 }
drh50b39962006-10-28 00:28:09 +0000526 assert( op!=TK_ISNULL || c==WO_ISNULL );
drh51147ba2005-07-23 22:59:55 +0000527 assert( op!=TK_IN || c==WO_IN );
528 assert( op!=TK_EQ || c==WO_EQ );
529 assert( op!=TK_LT || c==WO_LT );
530 assert( op!=TK_LE || c==WO_LE );
531 assert( op!=TK_GT || c==WO_GT );
532 assert( op!=TK_GE || c==WO_GE );
533 return c;
drhfe05af82005-07-21 03:14:59 +0000534}
535
536/*
537** Search for a term in the WHERE clause that is of the form "X <op> <expr>"
538** where X is a reference to the iColumn of table iCur and <op> is one of
539** the WO_xx operator codes specified by the op parameter.
540** Return a pointer to the term. Return 0 if not found.
541*/
542static WhereTerm *findTerm(
543 WhereClause *pWC, /* The WHERE clause to be searched */
544 int iCur, /* Cursor number of LHS */
545 int iColumn, /* Column number of LHS */
546 Bitmask notReady, /* RHS must not overlap with this mask */
drhec1724e2008-12-09 01:32:03 +0000547 u32 op, /* Mask of WO_xx values describing operator */
drhfe05af82005-07-21 03:14:59 +0000548 Index *pIdx /* Must be compatible with this index, if not NULL */
549){
550 WhereTerm *pTerm;
551 int k;
drh22c24032008-07-09 13:28:53 +0000552 assert( iCur>=0 );
drhec1724e2008-12-09 01:32:03 +0000553 op &= WO_ALL;
drhfe05af82005-07-21 03:14:59 +0000554 for(pTerm=pWC->a, k=pWC->nTerm; k; k--, pTerm++){
555 if( pTerm->leftCursor==iCur
556 && (pTerm->prereqRight & notReady)==0
drh700a2262008-12-17 19:22:15 +0000557 && pTerm->u.leftColumn==iColumn
drhb52076c2006-01-23 13:22:09 +0000558 && (pTerm->eOperator & op)!=0
drhfe05af82005-07-21 03:14:59 +0000559 ){
drh22c24032008-07-09 13:28:53 +0000560 if( pIdx && pTerm->eOperator!=WO_ISNULL ){
drhfe05af82005-07-21 03:14:59 +0000561 Expr *pX = pTerm->pExpr;
562 CollSeq *pColl;
563 char idxaff;
danielk1977f0113002006-01-24 12:09:17 +0000564 int j;
drhfe05af82005-07-21 03:14:59 +0000565 Parse *pParse = pWC->pParse;
566
567 idxaff = pIdx->pTable->aCol[iColumn].affinity;
568 if( !sqlite3IndexAffinityOk(pX, idxaff) ) continue;
danielk1977bcbb04e2007-05-29 12:11:29 +0000569
570 /* Figure out the collation sequence required from an index for
571 ** it to be useful for optimising expression pX. Store this
572 ** value in variable pColl.
573 */
574 assert(pX->pLeft);
575 pColl = sqlite3BinaryCompareCollSeq(pParse, pX->pLeft, pX->pRight);
danielk197793574162008-12-30 15:26:29 +0000576 assert(pColl || pParse->nErr);
danielk1977bcbb04e2007-05-29 12:11:29 +0000577
drh22c24032008-07-09 13:28:53 +0000578 for(j=0; pIdx->aiColumn[j]!=iColumn; j++){
drh34004ce2008-07-11 16:15:17 +0000579 if( NEVER(j>=pIdx->nColumn) ) return 0;
drh22c24032008-07-09 13:28:53 +0000580 }
danielk197793574162008-12-30 15:26:29 +0000581 if( pColl && sqlite3StrICmp(pColl->zName, pIdx->azColl[j]) ) continue;
drhfe05af82005-07-21 03:14:59 +0000582 }
583 return pTerm;
584 }
585 }
586 return 0;
587}
588
drh6c30be82005-07-29 15:10:17 +0000589/* Forward reference */
drh7b4fc6a2007-02-06 13:26:32 +0000590static void exprAnalyze(SrcList*, WhereClause*, int);
drh6c30be82005-07-29 15:10:17 +0000591
592/*
593** Call exprAnalyze on all terms in a WHERE clause.
594**
595**
596*/
597static void exprAnalyzeAll(
598 SrcList *pTabList, /* the FROM clause */
drh6c30be82005-07-29 15:10:17 +0000599 WhereClause *pWC /* the WHERE clause to be analyzed */
600){
drh6c30be82005-07-29 15:10:17 +0000601 int i;
drh9eb20282005-08-24 03:52:18 +0000602 for(i=pWC->nTerm-1; i>=0; i--){
drh7b4fc6a2007-02-06 13:26:32 +0000603 exprAnalyze(pTabList, pWC, i);
drh6c30be82005-07-29 15:10:17 +0000604 }
605}
606
drhd2687b72005-08-12 22:56:09 +0000607#ifndef SQLITE_OMIT_LIKE_OPTIMIZATION
608/*
609** Check to see if the given expression is a LIKE or GLOB operator that
610** can be optimized using inequality constraints. Return TRUE if it is
611** so and false if not.
612**
613** In order for the operator to be optimizible, the RHS must be a string
614** literal that does not begin with a wildcard.
615*/
616static int isLikeOrGlob(
drh7d10d5a2008-08-20 16:35:10 +0000617 Parse *pParse, /* Parsing and code generating context */
drhd2687b72005-08-12 22:56:09 +0000618 Expr *pExpr, /* Test this expression */
619 int *pnPattern, /* Number of non-wildcard prefix characters */
drh9f504ea2008-02-23 21:55:39 +0000620 int *pisComplete, /* True if the only wildcard is % in the last character */
621 int *pnoCase /* True if uppercase is equivalent to lowercase */
drhd2687b72005-08-12 22:56:09 +0000622){
drh5bd98ae2009-01-07 18:24:03 +0000623 const char *z; /* String on RHS of LIKE operator */
624 Expr *pRight, *pLeft; /* Right and left size of LIKE operator */
625 ExprList *pList; /* List of operands to the LIKE operator */
626 int c; /* One character in z[] */
627 int cnt; /* Number of non-wildcard prefix characters */
628 char wc[3]; /* Wildcard characters */
629 CollSeq *pColl; /* Collating sequence for LHS */
630 sqlite3 *db = pParse->db; /* Database connection */
drhd64fe2f2005-08-28 17:00:23 +0000631
drh9f504ea2008-02-23 21:55:39 +0000632 if( !sqlite3IsLikeFunction(db, pExpr, pnoCase, wc) ){
drhd2687b72005-08-12 22:56:09 +0000633 return 0;
634 }
drh9f504ea2008-02-23 21:55:39 +0000635#ifdef SQLITE_EBCDIC
636 if( *pnoCase ) return 0;
637#endif
danielk19776ab3a2e2009-02-19 14:39:25 +0000638 pList = pExpr->x.pList;
drh55ef4d92005-08-14 01:20:37 +0000639 pRight = pList->a[0].pExpr;
drh5bd98ae2009-01-07 18:24:03 +0000640 if( pRight->op!=TK_STRING ){
drhd2687b72005-08-12 22:56:09 +0000641 return 0;
642 }
drh55ef4d92005-08-14 01:20:37 +0000643 pLeft = pList->a[1].pExpr;
drhd2687b72005-08-12 22:56:09 +0000644 if( pLeft->op!=TK_COLUMN ){
645 return 0;
646 }
drh7d10d5a2008-08-20 16:35:10 +0000647 pColl = sqlite3ExprCollSeq(pParse, pLeft);
drh01495b92008-01-23 12:52:40 +0000648 assert( pColl!=0 || pLeft->iColumn==-1 );
drhd64fe2f2005-08-28 17:00:23 +0000649 if( pColl==0 ){
drh01495b92008-01-23 12:52:40 +0000650 /* No collation is defined for the ROWID. Use the default. */
drhd64fe2f2005-08-28 17:00:23 +0000651 pColl = db->pDfltColl;
652 }
drh9f504ea2008-02-23 21:55:39 +0000653 if( (pColl->type!=SQLITE_COLL_BINARY || *pnoCase) &&
654 (pColl->type!=SQLITE_COLL_NOCASE || !*pnoCase) ){
drhd64fe2f2005-08-28 17:00:23 +0000655 return 0;
656 }
drh7c01f1d2009-03-25 16:51:43 +0000657 sqlite3DequoteExpr(pRight);
danielk197700fd9572005-12-07 06:27:43 +0000658 z = (char *)pRight->token.z;
drhf998b732007-11-26 13:36:00 +0000659 cnt = 0;
660 if( z ){
661 while( (c=z[cnt])!=0 && c!=wc[0] && c!=wc[1] && c!=wc[2] ){ cnt++; }
662 }
drh5bd98ae2009-01-07 18:24:03 +0000663 if( cnt==0 || 255==(u8)z[cnt-1] ){
drhd2687b72005-08-12 22:56:09 +0000664 return 0;
665 }
drh55ef4d92005-08-14 01:20:37 +0000666 *pisComplete = z[cnt]==wc[0] && z[cnt+1]==0;
drhd2687b72005-08-12 22:56:09 +0000667 *pnPattern = cnt;
668 return 1;
669}
670#endif /* SQLITE_OMIT_LIKE_OPTIMIZATION */
671
drhedb193b2006-06-27 13:20:21 +0000672
673#ifndef SQLITE_OMIT_VIRTUALTABLE
drhfe05af82005-07-21 03:14:59 +0000674/*
drh7f375902006-06-13 17:38:59 +0000675** Check to see if the given expression is of the form
676**
677** column MATCH expr
678**
679** If it is then return TRUE. If not, return FALSE.
680*/
681static int isMatchOfColumn(
682 Expr *pExpr /* Test this expression */
683){
684 ExprList *pList;
685
686 if( pExpr->op!=TK_FUNCTION ){
687 return 0;
688 }
drhedb193b2006-06-27 13:20:21 +0000689 if( pExpr->token.n!=5 ||
690 sqlite3StrNICmp((const char*)pExpr->token.z,"match",5)!=0 ){
drh7f375902006-06-13 17:38:59 +0000691 return 0;
692 }
danielk19776ab3a2e2009-02-19 14:39:25 +0000693 pList = pExpr->x.pList;
drh7f375902006-06-13 17:38:59 +0000694 if( pList->nExpr!=2 ){
695 return 0;
696 }
697 if( pList->a[1].pExpr->op != TK_COLUMN ){
698 return 0;
699 }
700 return 1;
701}
drhedb193b2006-06-27 13:20:21 +0000702#endif /* SQLITE_OMIT_VIRTUALTABLE */
drh7f375902006-06-13 17:38:59 +0000703
704/*
drh54a167d2005-11-26 14:08:07 +0000705** If the pBase expression originated in the ON or USING clause of
706** a join, then transfer the appropriate markings over to derived.
707*/
708static void transferJoinMarkings(Expr *pDerived, Expr *pBase){
709 pDerived->flags |= pBase->flags & EP_FromJoin;
710 pDerived->iRightJoinTable = pBase->iRightJoinTable;
711}
712
drh3e355802007-02-23 23:13:33 +0000713#if !defined(SQLITE_OMIT_OR_OPTIMIZATION) && !defined(SQLITE_OMIT_SUBQUERY)
714/*
drh1a58fe02008-12-20 02:06:13 +0000715** Analyze a term that consists of two or more OR-connected
716** subterms. So in:
drh3e355802007-02-23 23:13:33 +0000717**
drh1a58fe02008-12-20 02:06:13 +0000718** ... WHERE (a=5) AND (b=7 OR c=9 OR d=13) AND (d=13)
719** ^^^^^^^^^^^^^^^^^^^^
drh3e355802007-02-23 23:13:33 +0000720**
drh1a58fe02008-12-20 02:06:13 +0000721** This routine analyzes terms such as the middle term in the above example.
722** A WhereOrTerm object is computed and attached to the term under
723** analysis, regardless of the outcome of the analysis. Hence:
drh3e355802007-02-23 23:13:33 +0000724**
drh1a58fe02008-12-20 02:06:13 +0000725** WhereTerm.wtFlags |= TERM_ORINFO
726** WhereTerm.u.pOrInfo = a dynamically allocated WhereOrTerm object
drh3e355802007-02-23 23:13:33 +0000727**
drh1a58fe02008-12-20 02:06:13 +0000728** The term being analyzed must have two or more of OR-connected subterms.
danielk1977fdc40192008-12-29 18:33:32 +0000729** A single subterm might be a set of AND-connected sub-subterms.
drh1a58fe02008-12-20 02:06:13 +0000730** Examples of terms under analysis:
drh3e355802007-02-23 23:13:33 +0000731**
drh1a58fe02008-12-20 02:06:13 +0000732** (A) t1.x=t2.y OR t1.x=t2.z OR t1.y=15 OR t1.z=t3.a+5
733** (B) x=expr1 OR expr2=x OR x=expr3
734** (C) t1.x=t2.y OR (t1.x=t2.z AND t1.y=15)
735** (D) x=expr1 OR (y>11 AND y<22 AND z LIKE '*hello*')
736** (E) (p.a=1 AND q.b=2 AND r.c=3) OR (p.x=4 AND q.y=5 AND r.z=6)
drh3e355802007-02-23 23:13:33 +0000737**
drh1a58fe02008-12-20 02:06:13 +0000738** CASE 1:
739**
740** If all subterms are of the form T.C=expr for some single column of C
741** a single table T (as shown in example B above) then create a new virtual
742** term that is an equivalent IN expression. In other words, if the term
743** being analyzed is:
744**
745** x = expr1 OR expr2 = x OR x = expr3
746**
747** then create a new virtual term like this:
748**
749** x IN (expr1,expr2,expr3)
750**
751** CASE 2:
752**
753** If all subterms are indexable by a single table T, then set
754**
755** WhereTerm.eOperator = WO_OR
756** WhereTerm.u.pOrInfo->indexable |= the cursor number for table T
757**
758** A subterm is "indexable" if it is of the form
759** "T.C <op> <expr>" where C is any column of table T and
760** <op> is one of "=", "<", "<=", ">", ">=", "IS NULL", or "IN".
761** A subterm is also indexable if it is an AND of two or more
762** subsubterms at least one of which is indexable. Indexable AND
763** subterms have their eOperator set to WO_AND and they have
764** u.pAndInfo set to a dynamically allocated WhereAndTerm object.
765**
766** From another point of view, "indexable" means that the subterm could
767** potentially be used with an index if an appropriate index exists.
768** This analysis does not consider whether or not the index exists; that
769** is something the bestIndex() routine will determine. This analysis
770** only looks at whether subterms appropriate for indexing exist.
771**
772** All examples A through E above all satisfy case 2. But if a term
773** also statisfies case 1 (such as B) we know that the optimizer will
774** always prefer case 1, so in that case we pretend that case 2 is not
775** satisfied.
776**
777** It might be the case that multiple tables are indexable. For example,
778** (E) above is indexable on tables P, Q, and R.
779**
780** Terms that satisfy case 2 are candidates for lookup by using
781** separate indices to find rowids for each subterm and composing
782** the union of all rowids using a RowSet object. This is similar
783** to "bitmap indices" in other database engines.
784**
785** OTHERWISE:
786**
787** If neither case 1 nor case 2 apply, then leave the eOperator set to
788** zero. This term is not useful for search.
drh3e355802007-02-23 23:13:33 +0000789*/
drh1a58fe02008-12-20 02:06:13 +0000790static void exprAnalyzeOrTerm(
791 SrcList *pSrc, /* the FROM clause */
792 WhereClause *pWC, /* the complete WHERE clause */
793 int idxTerm /* Index of the OR-term to be analyzed */
794){
795 Parse *pParse = pWC->pParse; /* Parser context */
796 sqlite3 *db = pParse->db; /* Database connection */
797 WhereTerm *pTerm = &pWC->a[idxTerm]; /* The term to be analyzed */
798 Expr *pExpr = pTerm->pExpr; /* The expression of the term */
drh111a6a72008-12-21 03:51:16 +0000799 WhereMaskSet *pMaskSet = pWC->pMaskSet; /* Table use masks */
drh1a58fe02008-12-20 02:06:13 +0000800 int i; /* Loop counters */
801 WhereClause *pOrWc; /* Breakup of pTerm into subterms */
802 WhereTerm *pOrTerm; /* A Sub-term within the pOrWc */
803 WhereOrInfo *pOrInfo; /* Additional information associated with pTerm */
804 Bitmask chngToIN; /* Tables that might satisfy case 1 */
805 Bitmask indexable; /* Tables that are indexable, satisfying case 2 */
drh3e355802007-02-23 23:13:33 +0000806
drh1a58fe02008-12-20 02:06:13 +0000807 /*
808 ** Break the OR clause into its separate subterms. The subterms are
809 ** stored in a WhereClause structure containing within the WhereOrInfo
810 ** object that is attached to the original OR clause term.
811 */
812 assert( (pTerm->wtFlags & (TERM_DYNAMIC|TERM_ORINFO|TERM_ANDINFO))==0 );
813 assert( pExpr->op==TK_OR );
drh954701a2008-12-29 23:45:07 +0000814 pTerm->u.pOrInfo = pOrInfo = sqlite3DbMallocZero(db, sizeof(*pOrInfo));
drh1a58fe02008-12-20 02:06:13 +0000815 if( pOrInfo==0 ) return;
816 pTerm->wtFlags |= TERM_ORINFO;
817 pOrWc = &pOrInfo->wc;
818 whereClauseInit(pOrWc, pWC->pParse, pMaskSet);
819 whereSplit(pOrWc, pExpr, TK_OR);
820 exprAnalyzeAll(pSrc, pOrWc);
821 if( db->mallocFailed ) return;
822 assert( pOrWc->nTerm>=2 );
823
824 /*
825 ** Compute the set of tables that might satisfy cases 1 or 2.
826 */
827 indexable = chngToIN = ~(Bitmask)0;
828 for(i=pOrWc->nTerm-1, pOrTerm=pOrWc->a; i>=0 && indexable; i--, pOrTerm++){
829 if( (pOrTerm->eOperator & WO_SINGLE)==0 ){
drh29435252008-12-28 18:35:08 +0000830 WhereAndInfo *pAndInfo;
831 assert( pOrTerm->eOperator==0 );
832 assert( (pOrTerm->wtFlags & (TERM_ANDINFO|TERM_ORINFO))==0 );
drh1a58fe02008-12-20 02:06:13 +0000833 chngToIN = 0;
drh29435252008-12-28 18:35:08 +0000834 pAndInfo = sqlite3DbMallocRaw(db, sizeof(*pAndInfo));
835 if( pAndInfo ){
836 WhereClause *pAndWC;
837 WhereTerm *pAndTerm;
838 int j;
839 Bitmask b = 0;
840 pOrTerm->u.pAndInfo = pAndInfo;
841 pOrTerm->wtFlags |= TERM_ANDINFO;
842 pOrTerm->eOperator = WO_AND;
843 pAndWC = &pAndInfo->wc;
844 whereClauseInit(pAndWC, pWC->pParse, pMaskSet);
845 whereSplit(pAndWC, pOrTerm->pExpr, TK_AND);
846 exprAnalyzeAll(pSrc, pAndWC);
drh7c2fbde2009-01-07 20:58:57 +0000847 testcase( db->mallocFailed );
drh96c7a7d2009-01-10 15:34:12 +0000848 if( !db->mallocFailed ){
849 for(j=0, pAndTerm=pAndWC->a; j<pAndWC->nTerm; j++, pAndTerm++){
850 assert( pAndTerm->pExpr );
851 if( allowedOp(pAndTerm->pExpr->op) ){
852 b |= getMask(pMaskSet, pAndTerm->leftCursor);
853 }
drh29435252008-12-28 18:35:08 +0000854 }
855 }
856 indexable &= b;
857 }
drh1a58fe02008-12-20 02:06:13 +0000858 }else if( pOrTerm->wtFlags & TERM_COPIED ){
859 /* Skip this term for now. We revisit it when we process the
860 ** corresponding TERM_VIRTUAL term */
861 }else{
862 Bitmask b;
863 b = getMask(pMaskSet, pOrTerm->leftCursor);
864 if( pOrTerm->wtFlags & TERM_VIRTUAL ){
865 WhereTerm *pOther = &pOrWc->a[pOrTerm->iParent];
866 b |= getMask(pMaskSet, pOther->leftCursor);
867 }
868 indexable &= b;
869 if( pOrTerm->eOperator!=WO_EQ ){
870 chngToIN = 0;
871 }else{
872 chngToIN &= b;
873 }
874 }
drh3e355802007-02-23 23:13:33 +0000875 }
drh1a58fe02008-12-20 02:06:13 +0000876
877 /*
878 ** Record the set of tables that satisfy case 2. The set might be
drh111a6a72008-12-21 03:51:16 +0000879 ** empty.
drh1a58fe02008-12-20 02:06:13 +0000880 */
881 pOrInfo->indexable = indexable;
drh111a6a72008-12-21 03:51:16 +0000882 pTerm->eOperator = indexable==0 ? 0 : WO_OR;
drh1a58fe02008-12-20 02:06:13 +0000883
884 /*
885 ** chngToIN holds a set of tables that *might* satisfy case 1. But
886 ** we have to do some additional checking to see if case 1 really
887 ** is satisfied.
888 */
889 if( chngToIN ){
890 int okToChngToIN = 0; /* True if the conversion to IN is valid */
891 int iColumn = -1; /* Column index on lhs of IN operator */
shane63207ab2009-02-04 01:49:30 +0000892 int iCursor = -1; /* Table cursor common to all terms */
drh1a58fe02008-12-20 02:06:13 +0000893 int j = 0; /* Loop counter */
894
895 /* Search for a table and column that appears on one side or the
896 ** other of the == operator in every subterm. That table and column
897 ** will be recorded in iCursor and iColumn. There might not be any
898 ** such table and column. Set okToChngToIN if an appropriate table
899 ** and column is found but leave okToChngToIN false if not found.
900 */
901 for(j=0; j<2 && !okToChngToIN; j++){
902 pOrTerm = pOrWc->a;
903 for(i=pOrWc->nTerm-1; i>=0; i--, pOrTerm++){
904 assert( pOrTerm->eOperator==WO_EQ );
905 pOrTerm->wtFlags &= ~TERM_OR_OK;
906 if( pOrTerm->leftCursor==iColumn ) continue;
907 if( (chngToIN & getMask(pMaskSet, pOrTerm->leftCursor))==0 ) continue;
908 iColumn = pOrTerm->u.leftColumn;
909 iCursor = pOrTerm->leftCursor;
910 break;
911 }
912 if( i<0 ){
913 assert( j==1 );
914 assert( (chngToIN&(chngToIN-1))==0 );
915 assert( chngToIN==getMask(pMaskSet, iColumn) );
916 break;
917 }
918 okToChngToIN = 1;
919 for(; i>=0 && okToChngToIN; i--, pOrTerm++){
920 assert( pOrTerm->eOperator==WO_EQ );
921 if( pOrTerm->leftCursor!=iCursor ){
922 pOrTerm->wtFlags &= ~TERM_OR_OK;
923 }else if( pOrTerm->u.leftColumn!=iColumn ){
924 okToChngToIN = 0;
925 }else{
926 int affLeft, affRight;
927 /* If the right-hand side is also a column, then the affinities
928 ** of both right and left sides must be such that no type
929 ** conversions are required on the right. (Ticket #2249)
930 */
931 affRight = sqlite3ExprAffinity(pOrTerm->pExpr->pRight);
932 affLeft = sqlite3ExprAffinity(pOrTerm->pExpr->pLeft);
933 if( affRight!=0 && affRight!=affLeft ){
934 okToChngToIN = 0;
935 }else{
936 pOrTerm->wtFlags |= TERM_OR_OK;
937 }
938 }
939 }
940 }
941
942 /* At this point, okToChngToIN is true if original pTerm satisfies
943 ** case 1. In that case, construct a new virtual term that is
944 ** pTerm converted into an IN operator.
945 */
946 if( okToChngToIN ){
947 Expr *pDup; /* A transient duplicate expression */
948 ExprList *pList = 0; /* The RHS of the IN operator */
949 Expr *pLeft = 0; /* The LHS of the IN operator */
950 Expr *pNew; /* The complete IN operator */
951
952 for(i=pOrWc->nTerm-1, pOrTerm=pOrWc->a; i>=0; i--, pOrTerm++){
953 if( (pOrTerm->wtFlags & TERM_OR_OK)==0 ) continue;
954 assert( pOrTerm->eOperator==WO_EQ );
955 assert( pOrTerm->leftCursor==iCursor );
956 assert( pOrTerm->u.leftColumn==iColumn );
danielk19776ab3a2e2009-02-19 14:39:25 +0000957 pDup = sqlite3ExprDup(db, pOrTerm->pExpr->pRight, 0);
drh1a58fe02008-12-20 02:06:13 +0000958 pList = sqlite3ExprListAppend(pWC->pParse, pList, pDup, 0);
959 pLeft = pOrTerm->pExpr->pLeft;
960 }
961 assert( pLeft!=0 );
danielk19776ab3a2e2009-02-19 14:39:25 +0000962 pDup = sqlite3ExprDup(db, pLeft, 0);
drh1a58fe02008-12-20 02:06:13 +0000963 pNew = sqlite3Expr(db, TK_IN, pDup, 0, 0);
964 if( pNew ){
965 int idxNew;
966 transferJoinMarkings(pNew, pExpr);
danielk19776ab3a2e2009-02-19 14:39:25 +0000967 assert( !ExprHasProperty(pNew, EP_xIsSelect) );
968 pNew->x.pList = pList;
drh1a58fe02008-12-20 02:06:13 +0000969 idxNew = whereClauseInsert(pWC, pNew, TERM_VIRTUAL|TERM_DYNAMIC);
970 testcase( idxNew==0 );
971 exprAnalyze(pSrc, pWC, idxNew);
972 pTerm = &pWC->a[idxTerm];
973 pWC->a[idxNew].iParent = idxTerm;
974 pTerm->nChild = 1;
975 }else{
976 sqlite3ExprListDelete(db, pList);
977 }
978 pTerm->eOperator = 0; /* case 1 trumps case 2 */
979 }
drh3e355802007-02-23 23:13:33 +0000980 }
drh3e355802007-02-23 23:13:33 +0000981}
982#endif /* !SQLITE_OMIT_OR_OPTIMIZATION && !SQLITE_OMIT_SUBQUERY */
drh54a167d2005-11-26 14:08:07 +0000983
drh1a58fe02008-12-20 02:06:13 +0000984
drh54a167d2005-11-26 14:08:07 +0000985/*
drh0aa74ed2005-07-16 13:33:20 +0000986** The input to this routine is an WhereTerm structure with only the
drh51147ba2005-07-23 22:59:55 +0000987** "pExpr" field filled in. The job of this routine is to analyze the
drh0aa74ed2005-07-16 13:33:20 +0000988** subexpression and populate all the other fields of the WhereTerm
drh75897232000-05-29 14:26:00 +0000989** structure.
drh51147ba2005-07-23 22:59:55 +0000990**
991** If the expression is of the form "<expr> <op> X" it gets commuted
drh1a58fe02008-12-20 02:06:13 +0000992** to the standard form of "X <op> <expr>".
993**
994** If the expression is of the form "X <op> Y" where both X and Y are
995** columns, then the original expression is unchanged and a new virtual
996** term of the form "Y <op> X" is added to the WHERE clause and
997** analyzed separately. The original term is marked with TERM_COPIED
998** and the new term is marked with TERM_DYNAMIC (because it's pExpr
999** needs to be freed with the WhereClause) and TERM_VIRTUAL (because it
1000** is a commuted copy of a prior term.) The original term has nChild=1
1001** and the copy has idxParent set to the index of the original term.
drh75897232000-05-29 14:26:00 +00001002*/
drh0fcef5e2005-07-19 17:38:22 +00001003static void exprAnalyze(
1004 SrcList *pSrc, /* the FROM clause */
drh9eb20282005-08-24 03:52:18 +00001005 WhereClause *pWC, /* the WHERE clause */
1006 int idxTerm /* Index of the term to be analyzed */
drh0fcef5e2005-07-19 17:38:22 +00001007){
drh1a58fe02008-12-20 02:06:13 +00001008 WhereTerm *pTerm; /* The term to be analyzed */
drh111a6a72008-12-21 03:51:16 +00001009 WhereMaskSet *pMaskSet; /* Set of table index masks */
drh1a58fe02008-12-20 02:06:13 +00001010 Expr *pExpr; /* The expression to be analyzed */
1011 Bitmask prereqLeft; /* Prerequesites of the pExpr->pLeft */
1012 Bitmask prereqAll; /* Prerequesites of pExpr */
drhdafc0ce2008-04-17 19:14:02 +00001013 Bitmask extraRight = 0;
drhd2687b72005-08-12 22:56:09 +00001014 int nPattern;
1015 int isComplete;
drh9f504ea2008-02-23 21:55:39 +00001016 int noCase;
drh1a58fe02008-12-20 02:06:13 +00001017 int op; /* Top-level operator. pExpr->op */
1018 Parse *pParse = pWC->pParse; /* Parsing context */
1019 sqlite3 *db = pParse->db; /* Database connection */
drh0fcef5e2005-07-19 17:38:22 +00001020
drhf998b732007-11-26 13:36:00 +00001021 if( db->mallocFailed ){
1022 return;
1023 }
1024 pTerm = &pWC->a[idxTerm];
1025 pMaskSet = pWC->pMaskSet;
1026 pExpr = pTerm->pExpr;
drh0fcef5e2005-07-19 17:38:22 +00001027 prereqLeft = exprTableUsage(pMaskSet, pExpr->pLeft);
drh50b39962006-10-28 00:28:09 +00001028 op = pExpr->op;
1029 if( op==TK_IN ){
drhf5b11382005-09-17 13:07:13 +00001030 assert( pExpr->pRight==0 );
danielk19776ab3a2e2009-02-19 14:39:25 +00001031 if( ExprHasProperty(pExpr, EP_xIsSelect) ){
1032 pTerm->prereqRight = exprSelectTableUsage(pMaskSet, pExpr->x.pSelect);
1033 }else{
1034 pTerm->prereqRight = exprListTableUsage(pMaskSet, pExpr->x.pList);
1035 }
drh50b39962006-10-28 00:28:09 +00001036 }else if( op==TK_ISNULL ){
1037 pTerm->prereqRight = 0;
drhf5b11382005-09-17 13:07:13 +00001038 }else{
1039 pTerm->prereqRight = exprTableUsage(pMaskSet, pExpr->pRight);
1040 }
drh22d6a532005-09-19 21:05:48 +00001041 prereqAll = exprTableUsage(pMaskSet, pExpr);
1042 if( ExprHasProperty(pExpr, EP_FromJoin) ){
drh42165be2008-03-26 14:56:34 +00001043 Bitmask x = getMask(pMaskSet, pExpr->iRightJoinTable);
1044 prereqAll |= x;
drhdafc0ce2008-04-17 19:14:02 +00001045 extraRight = x-1; /* ON clause terms may not be used with an index
1046 ** on left table of a LEFT JOIN. Ticket #3015 */
drh22d6a532005-09-19 21:05:48 +00001047 }
1048 pTerm->prereqAll = prereqAll;
drh0fcef5e2005-07-19 17:38:22 +00001049 pTerm->leftCursor = -1;
drh45b1ee42005-08-02 17:48:22 +00001050 pTerm->iParent = -1;
drhb52076c2006-01-23 13:22:09 +00001051 pTerm->eOperator = 0;
drh50b39962006-10-28 00:28:09 +00001052 if( allowedOp(op) && (pTerm->prereqRight & prereqLeft)==0 ){
drh0fcef5e2005-07-19 17:38:22 +00001053 Expr *pLeft = pExpr->pLeft;
1054 Expr *pRight = pExpr->pRight;
1055 if( pLeft->op==TK_COLUMN ){
1056 pTerm->leftCursor = pLeft->iTable;
drh700a2262008-12-17 19:22:15 +00001057 pTerm->u.leftColumn = pLeft->iColumn;
drh50b39962006-10-28 00:28:09 +00001058 pTerm->eOperator = operatorMask(op);
drh75897232000-05-29 14:26:00 +00001059 }
drh0fcef5e2005-07-19 17:38:22 +00001060 if( pRight && pRight->op==TK_COLUMN ){
1061 WhereTerm *pNew;
1062 Expr *pDup;
1063 if( pTerm->leftCursor>=0 ){
drh9eb20282005-08-24 03:52:18 +00001064 int idxNew;
danielk19776ab3a2e2009-02-19 14:39:25 +00001065 pDup = sqlite3ExprDup(db, pExpr, 0);
drh17435752007-08-16 04:30:38 +00001066 if( db->mallocFailed ){
drh633e6d52008-07-28 19:34:53 +00001067 sqlite3ExprDelete(db, pDup);
drh28f45912006-10-18 23:26:38 +00001068 return;
1069 }
drh9eb20282005-08-24 03:52:18 +00001070 idxNew = whereClauseInsert(pWC, pDup, TERM_VIRTUAL|TERM_DYNAMIC);
1071 if( idxNew==0 ) return;
1072 pNew = &pWC->a[idxNew];
1073 pNew->iParent = idxTerm;
1074 pTerm = &pWC->a[idxTerm];
drh45b1ee42005-08-02 17:48:22 +00001075 pTerm->nChild = 1;
drh165be382008-12-05 02:36:33 +00001076 pTerm->wtFlags |= TERM_COPIED;
drh0fcef5e2005-07-19 17:38:22 +00001077 }else{
1078 pDup = pExpr;
1079 pNew = pTerm;
1080 }
drh7d10d5a2008-08-20 16:35:10 +00001081 exprCommute(pParse, pDup);
drh0fcef5e2005-07-19 17:38:22 +00001082 pLeft = pDup->pLeft;
1083 pNew->leftCursor = pLeft->iTable;
drh700a2262008-12-17 19:22:15 +00001084 pNew->u.leftColumn = pLeft->iColumn;
drh0fcef5e2005-07-19 17:38:22 +00001085 pNew->prereqRight = prereqLeft;
1086 pNew->prereqAll = prereqAll;
drhb52076c2006-01-23 13:22:09 +00001087 pNew->eOperator = operatorMask(pDup->op);
drh75897232000-05-29 14:26:00 +00001088 }
1089 }
drhed378002005-07-28 23:12:08 +00001090
drhd2687b72005-08-12 22:56:09 +00001091#ifndef SQLITE_OMIT_BETWEEN_OPTIMIZATION
drhed378002005-07-28 23:12:08 +00001092 /* If a term is the BETWEEN operator, create two new virtual terms
drh1a58fe02008-12-20 02:06:13 +00001093 ** that define the range that the BETWEEN implements. For example:
1094 **
1095 ** a BETWEEN b AND c
1096 **
1097 ** is converted into:
1098 **
1099 ** (a BETWEEN b AND c) AND (a>=b) AND (a<=c)
1100 **
1101 ** The two new terms are added onto the end of the WhereClause object.
1102 ** The new terms are "dynamic" and are children of the original BETWEEN
1103 ** term. That means that if the BETWEEN term is coded, the children are
1104 ** skipped. Or, if the children are satisfied by an index, the original
1105 ** BETWEEN term is skipped.
drhed378002005-07-28 23:12:08 +00001106 */
drh29435252008-12-28 18:35:08 +00001107 else if( pExpr->op==TK_BETWEEN && pWC->op==TK_AND ){
danielk19776ab3a2e2009-02-19 14:39:25 +00001108 ExprList *pList = pExpr->x.pList;
drhed378002005-07-28 23:12:08 +00001109 int i;
1110 static const u8 ops[] = {TK_GE, TK_LE};
1111 assert( pList!=0 );
1112 assert( pList->nExpr==2 );
1113 for(i=0; i<2; i++){
1114 Expr *pNewExpr;
drh9eb20282005-08-24 03:52:18 +00001115 int idxNew;
danielk19776ab3a2e2009-02-19 14:39:25 +00001116 pNewExpr = sqlite3Expr(db, ops[i], sqlite3ExprDup(db, pExpr->pLeft, 0),
1117 sqlite3ExprDup(db, pList->a[i].pExpr, 0), 0);
drh9eb20282005-08-24 03:52:18 +00001118 idxNew = whereClauseInsert(pWC, pNewExpr, TERM_VIRTUAL|TERM_DYNAMIC);
drh6a1e0712008-12-05 15:24:15 +00001119 testcase( idxNew==0 );
drh7b4fc6a2007-02-06 13:26:32 +00001120 exprAnalyze(pSrc, pWC, idxNew);
drh9eb20282005-08-24 03:52:18 +00001121 pTerm = &pWC->a[idxTerm];
1122 pWC->a[idxNew].iParent = idxTerm;
drhed378002005-07-28 23:12:08 +00001123 }
drh45b1ee42005-08-02 17:48:22 +00001124 pTerm->nChild = 2;
drhed378002005-07-28 23:12:08 +00001125 }
drhd2687b72005-08-12 22:56:09 +00001126#endif /* SQLITE_OMIT_BETWEEN_OPTIMIZATION */
drhed378002005-07-28 23:12:08 +00001127
danielk19771576cd92006-01-14 08:02:28 +00001128#if !defined(SQLITE_OMIT_OR_OPTIMIZATION) && !defined(SQLITE_OMIT_SUBQUERY)
drh1a58fe02008-12-20 02:06:13 +00001129 /* Analyze a term that is composed of two or more subterms connected by
1130 ** an OR operator.
drh6c30be82005-07-29 15:10:17 +00001131 */
1132 else if( pExpr->op==TK_OR ){
drh29435252008-12-28 18:35:08 +00001133 assert( pWC->op==TK_AND );
drh1a58fe02008-12-20 02:06:13 +00001134 exprAnalyzeOrTerm(pSrc, pWC, idxTerm);
drh6c30be82005-07-29 15:10:17 +00001135 }
drhd2687b72005-08-12 22:56:09 +00001136#endif /* SQLITE_OMIT_OR_OPTIMIZATION */
1137
1138#ifndef SQLITE_OMIT_LIKE_OPTIMIZATION
1139 /* Add constraints to reduce the search space on a LIKE or GLOB
1140 ** operator.
drh9f504ea2008-02-23 21:55:39 +00001141 **
1142 ** A like pattern of the form "x LIKE 'abc%'" is changed into constraints
1143 **
1144 ** x>='abc' AND x<'abd' AND x LIKE 'abc%'
1145 **
1146 ** The last character of the prefix "abc" is incremented to form the
shane7bc71e52008-05-28 18:01:44 +00001147 ** termination condition "abd".
drhd2687b72005-08-12 22:56:09 +00001148 */
drh29435252008-12-28 18:35:08 +00001149 if( isLikeOrGlob(pParse, pExpr, &nPattern, &isComplete, &noCase)
1150 && pWC->op==TK_AND ){
drhd2687b72005-08-12 22:56:09 +00001151 Expr *pLeft, *pRight;
1152 Expr *pStr1, *pStr2;
1153 Expr *pNewExpr1, *pNewExpr2;
drh9eb20282005-08-24 03:52:18 +00001154 int idxNew1, idxNew2;
1155
danielk19776ab3a2e2009-02-19 14:39:25 +00001156 pLeft = pExpr->x.pList->a[1].pExpr;
1157 pRight = pExpr->x.pList->a[0].pExpr;
drh17435752007-08-16 04:30:38 +00001158 pStr1 = sqlite3PExpr(pParse, TK_STRING, 0, 0, 0);
drhd2687b72005-08-12 22:56:09 +00001159 if( pStr1 ){
drh17435752007-08-16 04:30:38 +00001160 sqlite3TokenCopy(db, &pStr1->token, &pRight->token);
drhd2687b72005-08-12 22:56:09 +00001161 pStr1->token.n = nPattern;
drh9c86df52007-06-11 12:56:15 +00001162 pStr1->flags = EP_Dequoted;
drhd2687b72005-08-12 22:56:09 +00001163 }
danielk19776ab3a2e2009-02-19 14:39:25 +00001164 pStr2 = sqlite3ExprDup(db, pStr1, 0);
drhf998b732007-11-26 13:36:00 +00001165 if( !db->mallocFailed ){
drh9f504ea2008-02-23 21:55:39 +00001166 u8 c, *pC;
danielk19776ab3a2e2009-02-19 14:39:25 +00001167 /* assert( pStr2->token.dyn ); */
drh9f504ea2008-02-23 21:55:39 +00001168 pC = (u8*)&pStr2->token.z[nPattern-1];
1169 c = *pC;
drh02a50b72008-05-26 18:33:40 +00001170 if( noCase ){
1171 if( c=='@' ) isComplete = 0;
1172 c = sqlite3UpperToLower[c];
1173 }
drh9f504ea2008-02-23 21:55:39 +00001174 *pC = c + 1;
drhd2687b72005-08-12 22:56:09 +00001175 }
danielk19776ab3a2e2009-02-19 14:39:25 +00001176 pNewExpr1 = sqlite3PExpr(pParse, TK_GE, sqlite3ExprDup(db,pLeft,0),pStr1,0);
drh9eb20282005-08-24 03:52:18 +00001177 idxNew1 = whereClauseInsert(pWC, pNewExpr1, TERM_VIRTUAL|TERM_DYNAMIC);
drh6a1e0712008-12-05 15:24:15 +00001178 testcase( idxNew1==0 );
drh7b4fc6a2007-02-06 13:26:32 +00001179 exprAnalyze(pSrc, pWC, idxNew1);
danielk19776ab3a2e2009-02-19 14:39:25 +00001180 pNewExpr2 = sqlite3PExpr(pParse, TK_LT, sqlite3ExprDup(db,pLeft,0),pStr2,0);
drh9eb20282005-08-24 03:52:18 +00001181 idxNew2 = whereClauseInsert(pWC, pNewExpr2, TERM_VIRTUAL|TERM_DYNAMIC);
drh6a1e0712008-12-05 15:24:15 +00001182 testcase( idxNew2==0 );
drh7b4fc6a2007-02-06 13:26:32 +00001183 exprAnalyze(pSrc, pWC, idxNew2);
drh9eb20282005-08-24 03:52:18 +00001184 pTerm = &pWC->a[idxTerm];
drhd2687b72005-08-12 22:56:09 +00001185 if( isComplete ){
drh9eb20282005-08-24 03:52:18 +00001186 pWC->a[idxNew1].iParent = idxTerm;
1187 pWC->a[idxNew2].iParent = idxTerm;
drhd2687b72005-08-12 22:56:09 +00001188 pTerm->nChild = 2;
1189 }
1190 }
1191#endif /* SQLITE_OMIT_LIKE_OPTIMIZATION */
drh7f375902006-06-13 17:38:59 +00001192
1193#ifndef SQLITE_OMIT_VIRTUALTABLE
1194 /* Add a WO_MATCH auxiliary term to the constraint set if the
1195 ** current expression is of the form: column MATCH expr.
1196 ** This information is used by the xBestIndex methods of
1197 ** virtual tables. The native query optimizer does not attempt
1198 ** to do anything with MATCH functions.
1199 */
1200 if( isMatchOfColumn(pExpr) ){
1201 int idxNew;
1202 Expr *pRight, *pLeft;
1203 WhereTerm *pNewTerm;
1204 Bitmask prereqColumn, prereqExpr;
1205
danielk19776ab3a2e2009-02-19 14:39:25 +00001206 pRight = pExpr->x.pList->a[0].pExpr;
1207 pLeft = pExpr->x.pList->a[1].pExpr;
drh7f375902006-06-13 17:38:59 +00001208 prereqExpr = exprTableUsage(pMaskSet, pRight);
1209 prereqColumn = exprTableUsage(pMaskSet, pLeft);
1210 if( (prereqExpr & prereqColumn)==0 ){
drh1a90e092006-06-14 22:07:10 +00001211 Expr *pNewExpr;
danielk19776ab3a2e2009-02-19 14:39:25 +00001212 pNewExpr = sqlite3Expr(db, TK_MATCH, 0, sqlite3ExprDup(db, pRight, 0), 0);
drh1a90e092006-06-14 22:07:10 +00001213 idxNew = whereClauseInsert(pWC, pNewExpr, TERM_VIRTUAL|TERM_DYNAMIC);
drh6a1e0712008-12-05 15:24:15 +00001214 testcase( idxNew==0 );
drh7f375902006-06-13 17:38:59 +00001215 pNewTerm = &pWC->a[idxNew];
1216 pNewTerm->prereqRight = prereqExpr;
1217 pNewTerm->leftCursor = pLeft->iTable;
drh700a2262008-12-17 19:22:15 +00001218 pNewTerm->u.leftColumn = pLeft->iColumn;
drh7f375902006-06-13 17:38:59 +00001219 pNewTerm->eOperator = WO_MATCH;
1220 pNewTerm->iParent = idxTerm;
drhd2ca60d2006-06-27 02:36:58 +00001221 pTerm = &pWC->a[idxTerm];
drh7f375902006-06-13 17:38:59 +00001222 pTerm->nChild = 1;
drh165be382008-12-05 02:36:33 +00001223 pTerm->wtFlags |= TERM_COPIED;
drh7f375902006-06-13 17:38:59 +00001224 pNewTerm->prereqAll = pTerm->prereqAll;
1225 }
1226 }
1227#endif /* SQLITE_OMIT_VIRTUALTABLE */
drhdafc0ce2008-04-17 19:14:02 +00001228
1229 /* Prevent ON clause terms of a LEFT JOIN from being used to drive
1230 ** an index for tables to the left of the join.
1231 */
1232 pTerm->prereqRight |= extraRight;
drh75897232000-05-29 14:26:00 +00001233}
1234
drh7b4fc6a2007-02-06 13:26:32 +00001235/*
1236** Return TRUE if any of the expressions in pList->a[iFirst...] contain
1237** a reference to any table other than the iBase table.
1238*/
1239static int referencesOtherTables(
1240 ExprList *pList, /* Search expressions in ths list */
drh111a6a72008-12-21 03:51:16 +00001241 WhereMaskSet *pMaskSet, /* Mapping from tables to bitmaps */
drh7b4fc6a2007-02-06 13:26:32 +00001242 int iFirst, /* Be searching with the iFirst-th expression */
1243 int iBase /* Ignore references to this table */
1244){
1245 Bitmask allowed = ~getMask(pMaskSet, iBase);
1246 while( iFirst<pList->nExpr ){
1247 if( (exprTableUsage(pMaskSet, pList->a[iFirst++].pExpr)&allowed)!=0 ){
1248 return 1;
1249 }
1250 }
1251 return 0;
1252}
1253
drh0fcef5e2005-07-19 17:38:22 +00001254
drh75897232000-05-29 14:26:00 +00001255/*
drh51669862004-12-18 18:40:26 +00001256** This routine decides if pIdx can be used to satisfy the ORDER BY
1257** clause. If it can, it returns 1. If pIdx cannot satisfy the
1258** ORDER BY clause, this routine returns 0.
1259**
1260** pOrderBy is an ORDER BY clause from a SELECT statement. pTab is the
1261** left-most table in the FROM clause of that same SELECT statement and
1262** the table has a cursor number of "base". pIdx is an index on pTab.
1263**
1264** nEqCol is the number of columns of pIdx that are used as equality
1265** constraints. Any of these columns may be missing from the ORDER BY
1266** clause and the match can still be a success.
1267**
drh51669862004-12-18 18:40:26 +00001268** All terms of the ORDER BY that match against the index must be either
1269** ASC or DESC. (Terms of the ORDER BY clause past the end of a UNIQUE
1270** index do not need to satisfy this constraint.) The *pbRev value is
1271** set to 1 if the ORDER BY clause is all DESC and it is set to 0 if
1272** the ORDER BY clause is all ASC.
1273*/
1274static int isSortingIndex(
1275 Parse *pParse, /* Parsing context */
drh111a6a72008-12-21 03:51:16 +00001276 WhereMaskSet *pMaskSet, /* Mapping from table cursor numbers to bitmaps */
drh51669862004-12-18 18:40:26 +00001277 Index *pIdx, /* The index we are testing */
drh74161702006-02-24 02:53:49 +00001278 int base, /* Cursor number for the table to be sorted */
drh51669862004-12-18 18:40:26 +00001279 ExprList *pOrderBy, /* The ORDER BY clause */
1280 int nEqCol, /* Number of index columns with == constraints */
1281 int *pbRev /* Set to 1 if ORDER BY is DESC */
1282){
drhb46b5772005-08-29 16:40:52 +00001283 int i, j; /* Loop counters */
drh85eeb692005-12-21 03:16:42 +00001284 int sortOrder = 0; /* XOR of index and ORDER BY sort direction */
drhb46b5772005-08-29 16:40:52 +00001285 int nTerm; /* Number of ORDER BY terms */
1286 struct ExprList_item *pTerm; /* A term of the ORDER BY clause */
drh51669862004-12-18 18:40:26 +00001287 sqlite3 *db = pParse->db;
1288
1289 assert( pOrderBy!=0 );
1290 nTerm = pOrderBy->nExpr;
1291 assert( nTerm>0 );
1292
1293 /* Match terms of the ORDER BY clause against columns of
1294 ** the index.
drhcc192542006-12-20 03:24:19 +00001295 **
1296 ** Note that indices have pIdx->nColumn regular columns plus
1297 ** one additional column containing the rowid. The rowid column
1298 ** of the index is also allowed to match against the ORDER BY
1299 ** clause.
drh51669862004-12-18 18:40:26 +00001300 */
drhcc192542006-12-20 03:24:19 +00001301 for(i=j=0, pTerm=pOrderBy->a; j<nTerm && i<=pIdx->nColumn; i++){
drh51669862004-12-18 18:40:26 +00001302 Expr *pExpr; /* The expression of the ORDER BY pTerm */
1303 CollSeq *pColl; /* The collating sequence of pExpr */
drh85eeb692005-12-21 03:16:42 +00001304 int termSortOrder; /* Sort order for this term */
drhcc192542006-12-20 03:24:19 +00001305 int iColumn; /* The i-th column of the index. -1 for rowid */
1306 int iSortOrder; /* 1 for DESC, 0 for ASC on the i-th index term */
1307 const char *zColl; /* Name of the collating sequence for i-th index term */
drh51669862004-12-18 18:40:26 +00001308
1309 pExpr = pTerm->pExpr;
1310 if( pExpr->op!=TK_COLUMN || pExpr->iTable!=base ){
1311 /* Can not use an index sort on anything that is not a column in the
1312 ** left-most table of the FROM clause */
drh7b4fc6a2007-02-06 13:26:32 +00001313 break;
drh51669862004-12-18 18:40:26 +00001314 }
1315 pColl = sqlite3ExprCollSeq(pParse, pExpr);
drhcc192542006-12-20 03:24:19 +00001316 if( !pColl ){
1317 pColl = db->pDfltColl;
1318 }
1319 if( i<pIdx->nColumn ){
1320 iColumn = pIdx->aiColumn[i];
1321 if( iColumn==pIdx->pTable->iPKey ){
1322 iColumn = -1;
1323 }
1324 iSortOrder = pIdx->aSortOrder[i];
1325 zColl = pIdx->azColl[i];
1326 }else{
1327 iColumn = -1;
1328 iSortOrder = 0;
1329 zColl = pColl->zName;
1330 }
1331 if( pExpr->iColumn!=iColumn || sqlite3StrICmp(pColl->zName, zColl) ){
drh9012bcb2004-12-19 00:11:35 +00001332 /* Term j of the ORDER BY clause does not match column i of the index */
1333 if( i<nEqCol ){
drh51669862004-12-18 18:40:26 +00001334 /* If an index column that is constrained by == fails to match an
1335 ** ORDER BY term, that is OK. Just ignore that column of the index
1336 */
1337 continue;
drhff354e92008-06-25 02:47:57 +00001338 }else if( i==pIdx->nColumn ){
1339 /* Index column i is the rowid. All other terms match. */
1340 break;
drh51669862004-12-18 18:40:26 +00001341 }else{
1342 /* If an index column fails to match and is not constrained by ==
1343 ** then the index cannot satisfy the ORDER BY constraint.
1344 */
1345 return 0;
1346 }
1347 }
danielk1977b3bf5562006-01-10 17:58:23 +00001348 assert( pIdx->aSortOrder!=0 );
drh85eeb692005-12-21 03:16:42 +00001349 assert( pTerm->sortOrder==0 || pTerm->sortOrder==1 );
drhcc192542006-12-20 03:24:19 +00001350 assert( iSortOrder==0 || iSortOrder==1 );
1351 termSortOrder = iSortOrder ^ pTerm->sortOrder;
drh51669862004-12-18 18:40:26 +00001352 if( i>nEqCol ){
drh85eeb692005-12-21 03:16:42 +00001353 if( termSortOrder!=sortOrder ){
drh51669862004-12-18 18:40:26 +00001354 /* Indices can only be used if all ORDER BY terms past the
1355 ** equality constraints are all either DESC or ASC. */
1356 return 0;
1357 }
1358 }else{
drh85eeb692005-12-21 03:16:42 +00001359 sortOrder = termSortOrder;
drh51669862004-12-18 18:40:26 +00001360 }
1361 j++;
1362 pTerm++;
drh7b4fc6a2007-02-06 13:26:32 +00001363 if( iColumn<0 && !referencesOtherTables(pOrderBy, pMaskSet, j, base) ){
drhcc192542006-12-20 03:24:19 +00001364 /* If the indexed column is the primary key and everything matches
drh7b4fc6a2007-02-06 13:26:32 +00001365 ** so far and none of the ORDER BY terms to the right reference other
1366 ** tables in the join, then we are assured that the index can be used
1367 ** to sort because the primary key is unique and so none of the other
1368 ** columns will make any difference
drhcc192542006-12-20 03:24:19 +00001369 */
1370 j = nTerm;
1371 }
drh51669862004-12-18 18:40:26 +00001372 }
1373
drhcc192542006-12-20 03:24:19 +00001374 *pbRev = sortOrder!=0;
drh8718f522005-08-13 16:13:04 +00001375 if( j>=nTerm ){
drhcc192542006-12-20 03:24:19 +00001376 /* All terms of the ORDER BY clause are covered by this index so
1377 ** this index can be used for sorting. */
1378 return 1;
1379 }
drh7b4fc6a2007-02-06 13:26:32 +00001380 if( pIdx->onError!=OE_None && i==pIdx->nColumn
1381 && !referencesOtherTables(pOrderBy, pMaskSet, j, base) ){
drhcc192542006-12-20 03:24:19 +00001382 /* All terms of this index match some prefix of the ORDER BY clause
drh7b4fc6a2007-02-06 13:26:32 +00001383 ** and the index is UNIQUE and no terms on the tail of the ORDER BY
1384 ** clause reference other tables in a join. If this is all true then
1385 ** the order by clause is superfluous. */
drh51669862004-12-18 18:40:26 +00001386 return 1;
1387 }
1388 return 0;
1389}
1390
1391/*
drhb6c29892004-11-22 19:12:19 +00001392** Check table to see if the ORDER BY clause in pOrderBy can be satisfied
1393** by sorting in order of ROWID. Return true if so and set *pbRev to be
1394** true for reverse ROWID and false for forward ROWID order.
1395*/
1396static int sortableByRowid(
1397 int base, /* Cursor number for table to be sorted */
1398 ExprList *pOrderBy, /* The ORDER BY clause */
drh111a6a72008-12-21 03:51:16 +00001399 WhereMaskSet *pMaskSet, /* Mapping from table cursors to bitmaps */
drhb6c29892004-11-22 19:12:19 +00001400 int *pbRev /* Set to 1 if ORDER BY is DESC */
1401){
1402 Expr *p;
1403
1404 assert( pOrderBy!=0 );
1405 assert( pOrderBy->nExpr>0 );
1406 p = pOrderBy->a[0].pExpr;
drh7b4fc6a2007-02-06 13:26:32 +00001407 if( p->op==TK_COLUMN && p->iTable==base && p->iColumn==-1
1408 && !referencesOtherTables(pOrderBy, pMaskSet, 1, base) ){
drhb6c29892004-11-22 19:12:19 +00001409 *pbRev = pOrderBy->a[0].sortOrder;
1410 return 1;
1411 }
1412 return 0;
1413}
1414
drhfe05af82005-07-21 03:14:59 +00001415/*
drhb6fb62d2005-09-20 08:47:20 +00001416** Prepare a crude estimate of the logarithm of the input value.
drh28c4cf42005-07-27 20:41:43 +00001417** The results need not be exact. This is only used for estimating
drh909626d2008-05-30 14:58:37 +00001418** the total cost of performing operations with O(logN) or O(NlogN)
drh28c4cf42005-07-27 20:41:43 +00001419** complexity. Because N is just a guess, it is no great tragedy if
1420** logN is a little off.
drh28c4cf42005-07-27 20:41:43 +00001421*/
1422static double estLog(double N){
drhb37df7b2005-10-13 02:09:49 +00001423 double logN = 1;
1424 double x = 10;
drh28c4cf42005-07-27 20:41:43 +00001425 while( N>x ){
drhb37df7b2005-10-13 02:09:49 +00001426 logN += 1;
drh28c4cf42005-07-27 20:41:43 +00001427 x *= 10;
1428 }
1429 return logN;
1430}
1431
drh6d209d82006-06-27 01:54:26 +00001432/*
1433** Two routines for printing the content of an sqlite3_index_info
1434** structure. Used for testing and debugging only. If neither
1435** SQLITE_TEST or SQLITE_DEBUG are defined, then these routines
1436** are no-ops.
1437*/
drh77a2a5e2007-04-06 01:04:39 +00001438#if !defined(SQLITE_OMIT_VIRTUALTABLE) && defined(SQLITE_DEBUG)
drh6d209d82006-06-27 01:54:26 +00001439static void TRACE_IDX_INPUTS(sqlite3_index_info *p){
1440 int i;
mlcreech3a00f902008-03-04 17:45:01 +00001441 if( !sqlite3WhereTrace ) return;
drh6d209d82006-06-27 01:54:26 +00001442 for(i=0; i<p->nConstraint; i++){
1443 sqlite3DebugPrintf(" constraint[%d]: col=%d termid=%d op=%d usabled=%d\n",
1444 i,
1445 p->aConstraint[i].iColumn,
1446 p->aConstraint[i].iTermOffset,
1447 p->aConstraint[i].op,
1448 p->aConstraint[i].usable);
1449 }
1450 for(i=0; i<p->nOrderBy; i++){
1451 sqlite3DebugPrintf(" orderby[%d]: col=%d desc=%d\n",
1452 i,
1453 p->aOrderBy[i].iColumn,
1454 p->aOrderBy[i].desc);
1455 }
1456}
1457static void TRACE_IDX_OUTPUTS(sqlite3_index_info *p){
1458 int i;
mlcreech3a00f902008-03-04 17:45:01 +00001459 if( !sqlite3WhereTrace ) return;
drh6d209d82006-06-27 01:54:26 +00001460 for(i=0; i<p->nConstraint; i++){
1461 sqlite3DebugPrintf(" usage[%d]: argvIdx=%d omit=%d\n",
1462 i,
1463 p->aConstraintUsage[i].argvIndex,
1464 p->aConstraintUsage[i].omit);
1465 }
1466 sqlite3DebugPrintf(" idxNum=%d\n", p->idxNum);
1467 sqlite3DebugPrintf(" idxStr=%s\n", p->idxStr);
1468 sqlite3DebugPrintf(" orderByConsumed=%d\n", p->orderByConsumed);
1469 sqlite3DebugPrintf(" estimatedCost=%g\n", p->estimatedCost);
1470}
1471#else
1472#define TRACE_IDX_INPUTS(A)
1473#define TRACE_IDX_OUTPUTS(A)
1474#endif
1475
danielk19771d461462009-04-21 09:02:45 +00001476/*
1477** Required because bestIndex() is called by bestOrClauseIndex()
1478*/
1479static void bestIndex(
1480 Parse*, WhereClause*, struct SrcList_item*, Bitmask, ExprList*, WhereCost*);
1481
1482/*
1483** This routine attempts to find an scanning strategy that can be used
1484** to optimize an 'OR' expression that is part of a WHERE clause.
1485**
1486** The table associated with FROM clause term pSrc may be either a
1487** regular B-Tree table or a virtual table.
1488*/
1489static void bestOrClauseIndex(
1490 Parse *pParse, /* The parsing context */
1491 WhereClause *pWC, /* The WHERE clause */
1492 struct SrcList_item *pSrc, /* The FROM clause term to search */
1493 Bitmask notReady, /* Mask of cursors that are not available */
1494 ExprList *pOrderBy, /* The ORDER BY clause */
1495 WhereCost *pCost /* Lowest cost query plan */
1496){
1497#ifndef SQLITE_OMIT_OR_OPTIMIZATION
1498 const int iCur = pSrc->iCursor; /* The cursor of the table to be accessed */
1499 const Bitmask maskSrc = getMask(pWC->pMaskSet, iCur); /* Bitmask for pSrc */
1500 WhereTerm * const pWCEnd = &pWC->a[pWC->nTerm]; /* End of pWC->a[] */
1501 WhereTerm *pTerm; /* A single term of the WHERE clause */
1502
1503 /* Search the WHERE clause terms for a usable WO_OR term. */
1504 for(pTerm=pWC->a; pTerm<pWCEnd; pTerm++){
1505 if( pTerm->eOperator==WO_OR
1506 && ((pTerm->prereqAll & ~maskSrc) & notReady)==0
1507 && (pTerm->u.pOrInfo->indexable & maskSrc)!=0
1508 ){
1509 WhereClause * const pOrWC = &pTerm->u.pOrInfo->wc;
1510 WhereTerm * const pOrWCEnd = &pOrWC->a[pOrWC->nTerm];
1511 WhereTerm *pOrTerm;
1512 int flags = WHERE_MULTI_OR;
1513 double rTotal = 0;
1514 double nRow = 0;
1515
1516 for(pOrTerm=pOrWC->a; pOrTerm<pOrWCEnd; pOrTerm++){
1517 WhereCost sTermCost;
1518 WHERETRACE(("... Multi-index OR testing for term %d of %d....\n",
1519 (pOrTerm - pOrWC->a), (pTerm - pWC->a)
1520 ));
1521 if( pOrTerm->eOperator==WO_AND ){
1522 WhereClause *pAndWC = &pOrTerm->u.pAndInfo->wc;
1523 bestIndex(pParse, pAndWC, pSrc, notReady, 0, &sTermCost);
1524 }else if( pOrTerm->leftCursor==iCur ){
1525 WhereClause tempWC;
1526 tempWC.pParse = pWC->pParse;
1527 tempWC.pMaskSet = pWC->pMaskSet;
1528 tempWC.op = TK_AND;
1529 tempWC.a = pOrTerm;
1530 tempWC.nTerm = 1;
1531 bestIndex(pParse, &tempWC, pSrc, notReady, 0, &sTermCost);
1532 }else{
1533 continue;
1534 }
1535 rTotal += sTermCost.rCost;
1536 nRow += sTermCost.nRow;
1537 if( rTotal>=pCost->rCost ) break;
1538 }
1539
1540 /* If there is an ORDER BY clause, increase the scan cost to account
1541 ** for the cost of the sort. */
1542 if( pOrderBy!=0 ){
1543 rTotal += nRow*estLog(nRow);
1544 WHERETRACE(("... sorting increases OR cost to %.9g\n", rTotal));
1545 }
1546
1547 /* If the cost of scanning using this OR term for optimization is
1548 ** less than the current cost stored in pCost, replace the contents
1549 ** of pCost. */
1550 WHERETRACE(("... multi-index OR cost=%.9g nrow=%.9g\n", rTotal, nRow));
1551 if( rTotal<pCost->rCost ){
1552 pCost->rCost = rTotal;
1553 pCost->nRow = nRow;
1554 pCost->plan.wsFlags = flags;
1555 pCost->plan.u.pTerm = pTerm;
1556 }
1557 }
1558 }
1559#endif /* SQLITE_OMIT_OR_OPTIMIZATION */
1560}
1561
drh9eff6162006-06-12 21:59:13 +00001562#ifndef SQLITE_OMIT_VIRTUALTABLE
1563/*
danielk19771d461462009-04-21 09:02:45 +00001564** Allocate and populate an sqlite3_index_info structure. It is the
1565** responsibility of the caller to eventually release the structure
1566** by passing the pointer returned by this function to sqlite3_free().
1567*/
1568static sqlite3_index_info *allocateIndexInfo(
1569 Parse *pParse,
1570 WhereClause *pWC,
1571 struct SrcList_item *pSrc,
1572 ExprList *pOrderBy
1573){
1574 int i, j;
1575 int nTerm;
1576 struct sqlite3_index_constraint *pIdxCons;
1577 struct sqlite3_index_orderby *pIdxOrderBy;
1578 struct sqlite3_index_constraint_usage *pUsage;
1579 WhereTerm *pTerm;
1580 int nOrderBy;
1581 sqlite3_index_info *pIdxInfo;
1582
1583 WHERETRACE(("Recomputing index info for %s...\n", pSrc->pTab->zName));
1584
1585 /* Count the number of possible WHERE clause constraints referring
1586 ** to this virtual table */
1587 for(i=nTerm=0, pTerm=pWC->a; i<pWC->nTerm; i++, pTerm++){
1588 if( pTerm->leftCursor != pSrc->iCursor ) continue;
1589 assert( (pTerm->eOperator&(pTerm->eOperator-1))==0 );
1590 testcase( pTerm->eOperator==WO_IN );
1591 testcase( pTerm->eOperator==WO_ISNULL );
1592 if( pTerm->eOperator & (WO_IN|WO_ISNULL) ) continue;
1593 nTerm++;
1594 }
1595
1596 /* If the ORDER BY clause contains only columns in the current
1597 ** virtual table then allocate space for the aOrderBy part of
1598 ** the sqlite3_index_info structure.
1599 */
1600 nOrderBy = 0;
1601 if( pOrderBy ){
1602 for(i=0; i<pOrderBy->nExpr; i++){
1603 Expr *pExpr = pOrderBy->a[i].pExpr;
1604 if( pExpr->op!=TK_COLUMN || pExpr->iTable!=pSrc->iCursor ) break;
1605 }
1606 if( i==pOrderBy->nExpr ){
1607 nOrderBy = pOrderBy->nExpr;
1608 }
1609 }
1610
1611 /* Allocate the sqlite3_index_info structure
1612 */
1613 pIdxInfo = sqlite3DbMallocZero(pParse->db, sizeof(*pIdxInfo)
1614 + (sizeof(*pIdxCons) + sizeof(*pUsage))*nTerm
1615 + sizeof(*pIdxOrderBy)*nOrderBy );
1616 if( pIdxInfo==0 ){
1617 sqlite3ErrorMsg(pParse, "out of memory");
1618 /* (double)0 In case of SQLITE_OMIT_FLOATING_POINT... */
1619 return 0;
1620 }
1621
1622 /* Initialize the structure. The sqlite3_index_info structure contains
1623 ** many fields that are declared "const" to prevent xBestIndex from
1624 ** changing them. We have to do some funky casting in order to
1625 ** initialize those fields.
1626 */
1627 pIdxCons = (struct sqlite3_index_constraint*)&pIdxInfo[1];
1628 pIdxOrderBy = (struct sqlite3_index_orderby*)&pIdxCons[nTerm];
1629 pUsage = (struct sqlite3_index_constraint_usage*)&pIdxOrderBy[nOrderBy];
1630 *(int*)&pIdxInfo->nConstraint = nTerm;
1631 *(int*)&pIdxInfo->nOrderBy = nOrderBy;
1632 *(struct sqlite3_index_constraint**)&pIdxInfo->aConstraint = pIdxCons;
1633 *(struct sqlite3_index_orderby**)&pIdxInfo->aOrderBy = pIdxOrderBy;
1634 *(struct sqlite3_index_constraint_usage**)&pIdxInfo->aConstraintUsage =
1635 pUsage;
1636
1637 for(i=j=0, pTerm=pWC->a; i<pWC->nTerm; i++, pTerm++){
1638 if( pTerm->leftCursor != pSrc->iCursor ) continue;
1639 assert( (pTerm->eOperator&(pTerm->eOperator-1))==0 );
1640 testcase( pTerm->eOperator==WO_IN );
1641 testcase( pTerm->eOperator==WO_ISNULL );
1642 if( pTerm->eOperator & (WO_IN|WO_ISNULL) ) continue;
1643 pIdxCons[j].iColumn = pTerm->u.leftColumn;
1644 pIdxCons[j].iTermOffset = i;
1645 pIdxCons[j].op = (u8)pTerm->eOperator;
1646 /* The direct assignment in the previous line is possible only because
1647 ** the WO_ and SQLITE_INDEX_CONSTRAINT_ codes are identical. The
1648 ** following asserts verify this fact. */
1649 assert( WO_EQ==SQLITE_INDEX_CONSTRAINT_EQ );
1650 assert( WO_LT==SQLITE_INDEX_CONSTRAINT_LT );
1651 assert( WO_LE==SQLITE_INDEX_CONSTRAINT_LE );
1652 assert( WO_GT==SQLITE_INDEX_CONSTRAINT_GT );
1653 assert( WO_GE==SQLITE_INDEX_CONSTRAINT_GE );
1654 assert( WO_MATCH==SQLITE_INDEX_CONSTRAINT_MATCH );
1655 assert( pTerm->eOperator & (WO_EQ|WO_LT|WO_LE|WO_GT|WO_GE|WO_MATCH) );
1656 j++;
1657 }
1658 for(i=0; i<nOrderBy; i++){
1659 Expr *pExpr = pOrderBy->a[i].pExpr;
1660 pIdxOrderBy[i].iColumn = pExpr->iColumn;
1661 pIdxOrderBy[i].desc = pOrderBy->a[i].sortOrder;
1662 }
1663
1664 return pIdxInfo;
1665}
1666
1667/*
1668** The table object reference passed as the second argument to this function
1669** must represent a virtual table. This function invokes the xBestIndex()
1670** method of the virtual table with the sqlite3_index_info pointer passed
1671** as the argument.
1672**
1673** If an error occurs, pParse is populated with an error message and a
1674** non-zero value is returned. Otherwise, 0 is returned and the output
1675** part of the sqlite3_index_info structure is left populated.
1676**
1677** Whether or not an error is returned, it is the responsibility of the
1678** caller to eventually free p->idxStr if p->needToFreeIdxStr indicates
1679** that this is required.
1680*/
1681static int vtabBestIndex(Parse *pParse, Table *pTab, sqlite3_index_info *p){
1682 sqlite3_vtab *pVtab = pTab->pVtab;
1683 int i;
1684 int rc;
1685
1686 (void)sqlite3SafetyOff(pParse->db);
1687 WHERETRACE(("xBestIndex for %s\n", pTab->zName));
1688 TRACE_IDX_INPUTS(p);
1689 rc = pVtab->pModule->xBestIndex(pVtab, p);
1690 TRACE_IDX_OUTPUTS(p);
1691 (void)sqlite3SafetyOn(pParse->db);
1692
1693 if( rc!=SQLITE_OK ){
1694 if( rc==SQLITE_NOMEM ){
1695 pParse->db->mallocFailed = 1;
1696 }else if( !pVtab->zErrMsg ){
1697 sqlite3ErrorMsg(pParse, "%s", sqlite3ErrStr(rc));
1698 }else{
1699 sqlite3ErrorMsg(pParse, "%s", pVtab->zErrMsg);
1700 }
1701 }
1702 sqlite3DbFree(pParse->db, pVtab->zErrMsg);
1703 pVtab->zErrMsg = 0;
1704
1705 for(i=0; i<p->nConstraint; i++){
1706 if( !p->aConstraint[i].usable && p->aConstraintUsage[i].argvIndex>0 ){
1707 sqlite3ErrorMsg(pParse,
1708 "table %s: xBestIndex returned an invalid plan", pTab->zName);
1709 }
1710 }
1711
1712 return pParse->nErr;
1713}
1714
1715
1716/*
drh7f375902006-06-13 17:38:59 +00001717** Compute the best index for a virtual table.
1718**
1719** The best index is computed by the xBestIndex method of the virtual
1720** table module. This routine is really just a wrapper that sets up
1721** the sqlite3_index_info structure that is used to communicate with
1722** xBestIndex.
1723**
1724** In a join, this routine might be called multiple times for the
1725** same virtual table. The sqlite3_index_info structure is created
1726** and initialized on the first invocation and reused on all subsequent
1727** invocations. The sqlite3_index_info structure is also used when
1728** code is generated to access the virtual table. The whereInfoDelete()
1729** routine takes care of freeing the sqlite3_index_info structure after
1730** everybody has finished with it.
drh9eff6162006-06-12 21:59:13 +00001731*/
danielk19771d461462009-04-21 09:02:45 +00001732static void bestVirtualIndex(
1733 Parse *pParse, /* The parsing context */
1734 WhereClause *pWC, /* The WHERE clause */
1735 struct SrcList_item *pSrc, /* The FROM clause term to search */
1736 Bitmask notReady, /* Mask of cursors that are not available */
1737 ExprList *pOrderBy, /* The order by clause */
1738 WhereCost *pCost, /* Lowest cost query plan */
1739 sqlite3_index_info **ppIdxInfo /* Index information passed to xBestIndex */
drh9eff6162006-06-12 21:59:13 +00001740){
1741 Table *pTab = pSrc->pTab;
1742 sqlite3_index_info *pIdxInfo;
1743 struct sqlite3_index_constraint *pIdxCons;
drh9eff6162006-06-12 21:59:13 +00001744 struct sqlite3_index_constraint_usage *pUsage;
1745 WhereTerm *pTerm;
1746 int i, j;
1747 int nOrderBy;
1748
1749 /* If the sqlite3_index_info structure has not been previously
danielk19771d461462009-04-21 09:02:45 +00001750 ** allocated and initialized, then allocate and initialize it now.
drh9eff6162006-06-12 21:59:13 +00001751 */
1752 pIdxInfo = *ppIdxInfo;
1753 if( pIdxInfo==0 ){
danielk19771d461462009-04-21 09:02:45 +00001754 *ppIdxInfo = pIdxInfo = allocateIndexInfo(pParse, pWC, pSrc, pOrderBy);
drh9eff6162006-06-12 21:59:13 +00001755 }
danielk1977732dc552009-04-21 17:23:04 +00001756 if( pIdxInfo==0 ){
1757 return;
1758 }
drh9eff6162006-06-12 21:59:13 +00001759
drh7f375902006-06-13 17:38:59 +00001760 /* At this point, the sqlite3_index_info structure that pIdxInfo points
1761 ** to will have been initialized, either during the current invocation or
1762 ** during some prior invocation. Now we just have to customize the
1763 ** details of pIdxInfo for the current invocation and pass it to
1764 ** xBestIndex.
1765 */
1766
danielk1977935ed5e2007-03-30 09:13:13 +00001767 /* The module name must be defined. Also, by this point there must
1768 ** be a pointer to an sqlite3_vtab structure. Otherwise
1769 ** sqlite3ViewGetColumnNames() would have picked up the error.
1770 */
drh9eff6162006-06-12 21:59:13 +00001771 assert( pTab->azModuleArg && pTab->azModuleArg[0] );
danielk19771d461462009-04-21 09:02:45 +00001772 assert( pTab->pVtab );
drh9eff6162006-06-12 21:59:13 +00001773
1774 /* Set the aConstraint[].usable fields and initialize all
drh7f375902006-06-13 17:38:59 +00001775 ** output variables to zero.
1776 **
1777 ** aConstraint[].usable is true for constraints where the right-hand
1778 ** side contains only references to tables to the left of the current
1779 ** table. In other words, if the constraint is of the form:
1780 **
1781 ** column = expr
1782 **
1783 ** and we are evaluating a join, then the constraint on column is
1784 ** only valid if all tables referenced in expr occur to the left
1785 ** of the table containing column.
1786 **
1787 ** The aConstraints[] array contains entries for all constraints
1788 ** on the current table. That way we only have to compute it once
1789 ** even though we might try to pick the best index multiple times.
1790 ** For each attempt at picking an index, the order of tables in the
1791 ** join might be different so we have to recompute the usable flag
1792 ** each time.
drh9eff6162006-06-12 21:59:13 +00001793 */
1794 pIdxCons = *(struct sqlite3_index_constraint**)&pIdxInfo->aConstraint;
1795 pUsage = pIdxInfo->aConstraintUsage;
1796 for(i=0; i<pIdxInfo->nConstraint; i++, pIdxCons++){
1797 j = pIdxCons->iTermOffset;
1798 pTerm = &pWC->a[j];
drhec1724e2008-12-09 01:32:03 +00001799 pIdxCons->usable = (pTerm->prereqRight & notReady)==0 ?1:0;
drh9eff6162006-06-12 21:59:13 +00001800 }
1801 memset(pUsage, 0, sizeof(pUsage[0])*pIdxInfo->nConstraint);
drh4be8b512006-06-13 23:51:34 +00001802 if( pIdxInfo->needToFreeIdxStr ){
1803 sqlite3_free(pIdxInfo->idxStr);
1804 }
1805 pIdxInfo->idxStr = 0;
1806 pIdxInfo->idxNum = 0;
1807 pIdxInfo->needToFreeIdxStr = 0;
drh9eff6162006-06-12 21:59:13 +00001808 pIdxInfo->orderByConsumed = 0;
shanefbd60f82009-02-04 03:59:25 +00001809 /* ((double)2) In case of SQLITE_OMIT_FLOATING_POINT... */
1810 pIdxInfo->estimatedCost = SQLITE_BIG_DBL / ((double)2);
drh9eff6162006-06-12 21:59:13 +00001811 nOrderBy = pIdxInfo->nOrderBy;
danielk19771d461462009-04-21 09:02:45 +00001812 if( !pOrderBy ){
1813 pIdxInfo->nOrderBy = 0;
drh9eff6162006-06-12 21:59:13 +00001814 }
danielk197774cdba42006-06-19 12:02:58 +00001815
danielk19771d461462009-04-21 09:02:45 +00001816 if( vtabBestIndex(pParse, pTab, pIdxInfo) ){
1817 return;
danielk197739359dc2008-03-17 09:36:44 +00001818 }
1819
danielk19771d461462009-04-21 09:02:45 +00001820 /* The cost is not allowed to be larger than SQLITE_BIG_DBL (the
1821 ** inital value of lowestCost in this loop. If it is, then the
1822 ** (cost<lowestCost) test below will never be true.
1823 **
1824 ** Use "(double)2" instead of "2.0" in case OMIT_FLOATING_POINT
1825 ** is defined.
1826 */
1827 if( (SQLITE_BIG_DBL/((double)2))<pIdxInfo->estimatedCost ){
1828 pCost->rCost = (SQLITE_BIG_DBL/((double)2));
1829 }else{
1830 pCost->rCost = pIdxInfo->estimatedCost;
1831 }
1832 pCost->plan.wsFlags = WHERE_VIRTUALTABLE;
1833 pCost->plan.u.pVtabIdx = pIdxInfo;
1834 if( pIdxInfo && pIdxInfo->orderByConsumed ){
1835 pCost->plan.wsFlags |= WHERE_ORDERBY;
1836 }
1837 pCost->plan.nEq = 0;
1838 pIdxInfo->nOrderBy = nOrderBy;
1839
1840 /* Try to find a more efficient access pattern by using multiple indexes
1841 ** to optimize an OR expression within the WHERE clause.
1842 */
1843 bestOrClauseIndex(pParse, pWC, pSrc, notReady, pOrderBy, pCost);
drh9eff6162006-06-12 21:59:13 +00001844}
1845#endif /* SQLITE_OMIT_VIRTUALTABLE */
1846
drh28c4cf42005-07-27 20:41:43 +00001847/*
drh111a6a72008-12-21 03:51:16 +00001848** Find the query plan for accessing a particular table. Write the
1849** best query plan and its cost into the WhereCost object supplied as the
1850** last parameter.
drh51147ba2005-07-23 22:59:55 +00001851**
drh111a6a72008-12-21 03:51:16 +00001852** The lowest cost plan wins. The cost is an estimate of the amount of
1853** CPU and disk I/O need to process the request using the selected plan.
drh51147ba2005-07-23 22:59:55 +00001854** Factors that influence cost include:
1855**
1856** * The estimated number of rows that will be retrieved. (The
1857** fewer the better.)
1858**
1859** * Whether or not sorting must occur.
1860**
1861** * Whether or not there must be separate lookups in the
1862** index and in the main table.
1863**
danielk1977e2d7b242009-02-23 17:33:49 +00001864** If there was an INDEXED BY clause (pSrc->pIndex) attached to the table in
1865** the SQL statement, then this function only considers plans using the
drh296a4832009-03-22 20:36:18 +00001866** named index. If no such plan is found, then the returned cost is
1867** SQLITE_BIG_DBL. If a plan is found that uses the named index,
danielk197785574e32008-10-06 05:32:18 +00001868** then the cost is calculated in the usual way.
1869**
danielk1977e2d7b242009-02-23 17:33:49 +00001870** If a NOT INDEXED clause (pSrc->notIndexed!=0) was attached to the table
1871** in the SELECT statement, then no indexes are considered. However, the
1872** selected plan may still take advantage of the tables built-in rowid
danielk197785574e32008-10-06 05:32:18 +00001873** index.
drhfe05af82005-07-21 03:14:59 +00001874*/
danielk19771d461462009-04-21 09:02:45 +00001875static void bestBtreeIndex(
drhfe05af82005-07-21 03:14:59 +00001876 Parse *pParse, /* The parsing context */
1877 WhereClause *pWC, /* The WHERE clause */
1878 struct SrcList_item *pSrc, /* The FROM clause term to search */
1879 Bitmask notReady, /* Mask of cursors that are not available */
drh111a6a72008-12-21 03:51:16 +00001880 ExprList *pOrderBy, /* The ORDER BY clause */
1881 WhereCost *pCost /* Lowest cost query plan */
drhfe05af82005-07-21 03:14:59 +00001882){
drh111a6a72008-12-21 03:51:16 +00001883 WhereTerm *pTerm; /* A single term of the WHERE clause */
drh51147ba2005-07-23 22:59:55 +00001884 int iCur = pSrc->iCursor; /* The cursor of the table to be accessed */
1885 Index *pProbe; /* An index we are evaluating */
1886 int rev; /* True to scan in reverse order */
drh165be382008-12-05 02:36:33 +00001887 int wsFlags; /* Flags associated with pProbe */
drh51147ba2005-07-23 22:59:55 +00001888 int nEq; /* Number of == or IN constraints */
drhc49de5d2007-01-19 01:06:01 +00001889 int eqTermMask; /* Mask of valid equality operators */
drh51147ba2005-07-23 22:59:55 +00001890 double cost; /* Cost of using pProbe */
drh111a6a72008-12-21 03:51:16 +00001891 double nRow; /* Estimated number of rows in result set */
drhdd5f5a62008-12-23 13:35:23 +00001892 int i; /* Loop counter */
drhfe05af82005-07-21 03:14:59 +00001893
drh165be382008-12-05 02:36:33 +00001894 WHERETRACE(("bestIndex: tbl=%s notReady=%llx\n", pSrc->pTab->zName,notReady));
drh4dd238a2006-03-28 23:55:57 +00001895 pProbe = pSrc->pTab->pIndex;
danielk197785574e32008-10-06 05:32:18 +00001896 if( pSrc->notIndexed ){
1897 pProbe = 0;
1898 }
drh4dd238a2006-03-28 23:55:57 +00001899
1900 /* If the table has no indices and there are no terms in the where
1901 ** clause that refer to the ROWID, then we will never be able to do
1902 ** anything other than a full table scan on this table. We might as
1903 ** well put it first in the join order. That way, perhaps it can be
1904 ** referenced by other tables in the join.
1905 */
drh111a6a72008-12-21 03:51:16 +00001906 memset(pCost, 0, sizeof(*pCost));
drh4dd238a2006-03-28 23:55:57 +00001907 if( pProbe==0 &&
1908 findTerm(pWC, iCur, -1, 0, WO_EQ|WO_IN|WO_LT|WO_LE|WO_GT|WO_GE,0)==0 &&
drh7b4fc6a2007-02-06 13:26:32 +00001909 (pOrderBy==0 || !sortableByRowid(iCur, pOrderBy, pWC->pMaskSet, &rev)) ){
drh69a442e2009-04-06 12:26:57 +00001910 if( pParse->db->flags & SQLITE_ReverseOrder ){
1911 /* For application testing, randomly reverse the output order for
1912 ** SELECT statements that omit the ORDER BY clause. This will help
1913 ** to find cases where
1914 */
1915 pCost->plan.wsFlags |= WHERE_REVERSE;
1916 }
drh111a6a72008-12-21 03:51:16 +00001917 return;
drh4dd238a2006-03-28 23:55:57 +00001918 }
drh111a6a72008-12-21 03:51:16 +00001919 pCost->rCost = SQLITE_BIG_DBL;
drh51147ba2005-07-23 22:59:55 +00001920
danielk197785574e32008-10-06 05:32:18 +00001921 /* Check for a rowid=EXPR or rowid IN (...) constraints. If there was
1922 ** an INDEXED BY clause attached to this table, skip this step.
drhfe05af82005-07-21 03:14:59 +00001923 */
danielk197785574e32008-10-06 05:32:18 +00001924 if( !pSrc->pIndex ){
1925 pTerm = findTerm(pWC, iCur, -1, notReady, WO_EQ|WO_IN, 0);
1926 if( pTerm ){
1927 Expr *pExpr;
drh111a6a72008-12-21 03:51:16 +00001928 pCost->plan.wsFlags = WHERE_ROWID_EQ;
danielk197785574e32008-10-06 05:32:18 +00001929 if( pTerm->eOperator & WO_EQ ){
1930 /* Rowid== is always the best pick. Look no further. Because only
1931 ** a single row is generated, output is always in sorted order */
drh111a6a72008-12-21 03:51:16 +00001932 pCost->plan.wsFlags = WHERE_ROWID_EQ | WHERE_UNIQUE;
1933 pCost->plan.nEq = 1;
danielk197785574e32008-10-06 05:32:18 +00001934 WHERETRACE(("... best is rowid\n"));
drh111a6a72008-12-21 03:51:16 +00001935 pCost->rCost = 0;
1936 pCost->nRow = 1;
1937 return;
danielk19776ab3a2e2009-02-19 14:39:25 +00001938 }else if( !ExprHasProperty((pExpr = pTerm->pExpr), EP_xIsSelect)
1939 && pExpr->x.pList
1940 ){
danielk197785574e32008-10-06 05:32:18 +00001941 /* Rowid IN (LIST): cost is NlogN where N is the number of list
1942 ** elements. */
danielk19776ab3a2e2009-02-19 14:39:25 +00001943 pCost->rCost = pCost->nRow = pExpr->x.pList->nExpr;
drh111a6a72008-12-21 03:51:16 +00001944 pCost->rCost *= estLog(pCost->rCost);
danielk197785574e32008-10-06 05:32:18 +00001945 }else{
1946 /* Rowid IN (SELECT): cost is NlogN where N is the number of rows
1947 ** in the result of the inner select. We have no way to estimate
1948 ** that value so make a wild guess. */
drh111a6a72008-12-21 03:51:16 +00001949 pCost->nRow = 100;
1950 pCost->rCost = 200;
drh28c4cf42005-07-27 20:41:43 +00001951 }
drh111a6a72008-12-21 03:51:16 +00001952 WHERETRACE(("... rowid IN cost: %.9g\n", pCost->rCost));
drh51147ba2005-07-23 22:59:55 +00001953 }
danielk197785574e32008-10-06 05:32:18 +00001954
1955 /* Estimate the cost of a table scan. If we do not know how many
1956 ** entries are in the table, use 1 million as a guess.
1957 */
1958 cost = pProbe ? pProbe->aiRowEst[0] : 1000000;
1959 WHERETRACE(("... table scan base cost: %.9g\n", cost));
drh165be382008-12-05 02:36:33 +00001960 wsFlags = WHERE_ROWID_RANGE;
danielk197785574e32008-10-06 05:32:18 +00001961
1962 /* Check for constraints on a range of rowids in a table scan.
1963 */
1964 pTerm = findTerm(pWC, iCur, -1, notReady, WO_LT|WO_LE|WO_GT|WO_GE, 0);
1965 if( pTerm ){
1966 if( findTerm(pWC, iCur, -1, notReady, WO_LT|WO_LE, 0) ){
drh165be382008-12-05 02:36:33 +00001967 wsFlags |= WHERE_TOP_LIMIT;
drh700a2262008-12-17 19:22:15 +00001968 cost /= 3; /* Guess that rowid<EXPR eliminates two-thirds of rows */
danielk197785574e32008-10-06 05:32:18 +00001969 }
1970 if( findTerm(pWC, iCur, -1, notReady, WO_GT|WO_GE, 0) ){
drh165be382008-12-05 02:36:33 +00001971 wsFlags |= WHERE_BTM_LIMIT;
danielk197785574e32008-10-06 05:32:18 +00001972 cost /= 3; /* Guess that rowid>EXPR eliminates two-thirds of rows */
1973 }
1974 WHERETRACE(("... rowid range reduces cost to %.9g\n", cost));
1975 }else{
drh165be382008-12-05 02:36:33 +00001976 wsFlags = 0;
danielk197785574e32008-10-06 05:32:18 +00001977 }
drh111a6a72008-12-21 03:51:16 +00001978 nRow = cost;
danielk197785574e32008-10-06 05:32:18 +00001979
1980 /* If the table scan does not satisfy the ORDER BY clause, increase
1981 ** the cost by NlogN to cover the expense of sorting. */
1982 if( pOrderBy ){
1983 if( sortableByRowid(iCur, pOrderBy, pWC->pMaskSet, &rev) ){
drh165be382008-12-05 02:36:33 +00001984 wsFlags |= WHERE_ORDERBY|WHERE_ROWID_RANGE;
danielk197785574e32008-10-06 05:32:18 +00001985 if( rev ){
drh165be382008-12-05 02:36:33 +00001986 wsFlags |= WHERE_REVERSE;
danielk197785574e32008-10-06 05:32:18 +00001987 }
1988 }else{
1989 cost += cost*estLog(cost);
1990 WHERETRACE(("... sorting increases cost to %.9g\n", cost));
1991 }
drh699b3d42009-02-23 16:52:07 +00001992 }else if( pParse->db->flags & SQLITE_ReverseOrder ){
1993 /* For application testing, randomly reverse the output order for
1994 ** SELECT statements that omit the ORDER BY clause. This will help
1995 ** to find cases where
1996 */
1997 wsFlags |= WHERE_REVERSE;
danielk197785574e32008-10-06 05:32:18 +00001998 }
drh699b3d42009-02-23 16:52:07 +00001999
2000 /* Remember this case if it is the best so far */
drh111a6a72008-12-21 03:51:16 +00002001 if( cost<pCost->rCost ){
2002 pCost->rCost = cost;
2003 pCost->nRow = nRow;
2004 pCost->plan.wsFlags = wsFlags;
danielk197785574e32008-10-06 05:32:18 +00002005 }
drhfe05af82005-07-21 03:14:59 +00002006 }
2007
danielk19771d461462009-04-21 09:02:45 +00002008 bestOrClauseIndex(pParse, pWC, pSrc, notReady, pOrderBy, pCost);
drhdd5f5a62008-12-23 13:35:23 +00002009
drhc49de5d2007-01-19 01:06:01 +00002010 /* If the pSrc table is the right table of a LEFT JOIN then we may not
2011 ** use an index to satisfy IS NULL constraints on that table. This is
2012 ** because columns might end up being NULL if the table does not match -
2013 ** a circumstance which the index cannot help us discover. Ticket #2177.
2014 */
2015 if( (pSrc->jointype & JT_LEFT)!=0 ){
2016 eqTermMask = WO_EQ|WO_IN;
2017 }else{
2018 eqTermMask = WO_EQ|WO_IN|WO_ISNULL;
2019 }
2020
drhfe05af82005-07-21 03:14:59 +00002021 /* Look at each index.
2022 */
danielk197785574e32008-10-06 05:32:18 +00002023 if( pSrc->pIndex ){
2024 pProbe = pSrc->pIndex;
2025 }
2026 for(; pProbe; pProbe=(pSrc->pIndex ? 0 : pProbe->pNext)){
drh75572e92009-03-29 00:13:03 +00002027 double inMultiplier = 1; /* Number of equality look-ups needed */
2028 int inMultIsEst = 0; /* True if inMultiplier is an estimate */
drh51147ba2005-07-23 22:59:55 +00002029
drh4f0c5872007-03-26 22:05:01 +00002030 WHERETRACE(("... index %s:\n", pProbe->zName));
drhfe05af82005-07-21 03:14:59 +00002031
2032 /* Count the number of columns in the index that are satisfied
drh75572e92009-03-29 00:13:03 +00002033 ** by x=EXPR constraints or x IN (...) constraints. For a term
2034 ** of the form x=EXPR we only have to do a single binary search.
2035 ** But for x IN (...) we have to do a number of binary searched
2036 ** equal to the number of entries on the RHS of the IN operator.
2037 ** The inMultipler variable with try to estimate the number of
2038 ** binary searches needed.
drhfe05af82005-07-21 03:14:59 +00002039 */
drh165be382008-12-05 02:36:33 +00002040 wsFlags = 0;
drhfe05af82005-07-21 03:14:59 +00002041 for(i=0; i<pProbe->nColumn; i++){
2042 int j = pProbe->aiColumn[i];
drhc49de5d2007-01-19 01:06:01 +00002043 pTerm = findTerm(pWC, iCur, j, notReady, eqTermMask, pProbe);
drhfe05af82005-07-21 03:14:59 +00002044 if( pTerm==0 ) break;
drh165be382008-12-05 02:36:33 +00002045 wsFlags |= WHERE_COLUMN_EQ;
drhb52076c2006-01-23 13:22:09 +00002046 if( pTerm->eOperator & WO_IN ){
drha6110402005-07-28 20:51:19 +00002047 Expr *pExpr = pTerm->pExpr;
drh165be382008-12-05 02:36:33 +00002048 wsFlags |= WHERE_COLUMN_IN;
danielk19776ab3a2e2009-02-19 14:39:25 +00002049 if( ExprHasProperty(pExpr, EP_xIsSelect) ){
drhffe0f892006-05-11 13:26:25 +00002050 inMultiplier *= 25;
drh75572e92009-03-29 00:13:03 +00002051 inMultIsEst = 1;
danielk19776ab3a2e2009-02-19 14:39:25 +00002052 }else if( pExpr->x.pList ){
2053 inMultiplier *= pExpr->x.pList->nExpr + 1;
drhfe05af82005-07-21 03:14:59 +00002054 }
2055 }
2056 }
drh111a6a72008-12-21 03:51:16 +00002057 nRow = pProbe->aiRowEst[i] * inMultiplier;
drh75572e92009-03-29 00:13:03 +00002058 /* If inMultiplier is an estimate and that estimate results in an
2059 ** nRow it that is more than half number of rows in the table,
2060 ** then reduce inMultipler */
2061 if( inMultIsEst && nRow*2 > pProbe->aiRowEst[0] ){
2062 nRow = pProbe->aiRowEst[0]/2;
2063 inMultiplier = nRow/pProbe->aiRowEst[i];
2064 }
2065 cost = nRow + inMultiplier*estLog(pProbe->aiRowEst[0]);
drh51147ba2005-07-23 22:59:55 +00002066 nEq = i;
drh165be382008-12-05 02:36:33 +00002067 if( pProbe->onError!=OE_None && (wsFlags & WHERE_COLUMN_IN)==0
drh943af3c2005-07-29 19:43:58 +00002068 && nEq==pProbe->nColumn ){
drh165be382008-12-05 02:36:33 +00002069 wsFlags |= WHERE_UNIQUE;
drh943af3c2005-07-29 19:43:58 +00002070 }
drh75572e92009-03-29 00:13:03 +00002071 WHERETRACE(("...... nEq=%d inMult=%.9g nRow=%.9g cost=%.9g\n",
2072 nEq, inMultiplier, nRow, cost));
drhfe05af82005-07-21 03:14:59 +00002073
drh75572e92009-03-29 00:13:03 +00002074 /* Look for range constraints. Assume that each range constraint
2075 ** makes the search space 1/3rd smaller.
drhfe05af82005-07-21 03:14:59 +00002076 */
drh51147ba2005-07-23 22:59:55 +00002077 if( nEq<pProbe->nColumn ){
2078 int j = pProbe->aiColumn[nEq];
2079 pTerm = findTerm(pWC, iCur, j, notReady, WO_LT|WO_LE|WO_GT|WO_GE, pProbe);
2080 if( pTerm ){
drh165be382008-12-05 02:36:33 +00002081 wsFlags |= WHERE_COLUMN_RANGE;
drh51147ba2005-07-23 22:59:55 +00002082 if( findTerm(pWC, iCur, j, notReady, WO_LT|WO_LE, pProbe) ){
drh165be382008-12-05 02:36:33 +00002083 wsFlags |= WHERE_TOP_LIMIT;
drhb37df7b2005-10-13 02:09:49 +00002084 cost /= 3;
drh111a6a72008-12-21 03:51:16 +00002085 nRow /= 3;
drh51147ba2005-07-23 22:59:55 +00002086 }
2087 if( findTerm(pWC, iCur, j, notReady, WO_GT|WO_GE, pProbe) ){
drh165be382008-12-05 02:36:33 +00002088 wsFlags |= WHERE_BTM_LIMIT;
drhb37df7b2005-10-13 02:09:49 +00002089 cost /= 3;
drh111a6a72008-12-21 03:51:16 +00002090 nRow /= 3;
drh51147ba2005-07-23 22:59:55 +00002091 }
drh75572e92009-03-29 00:13:03 +00002092 WHERETRACE(("...... range reduces nRow to %.9g and cost to %.9g\n",
2093 nRow, cost));
drh51147ba2005-07-23 22:59:55 +00002094 }
2095 }
2096
drh28c4cf42005-07-27 20:41:43 +00002097 /* Add the additional cost of sorting if that is a factor.
drh51147ba2005-07-23 22:59:55 +00002098 */
drh28c4cf42005-07-27 20:41:43 +00002099 if( pOrderBy ){
drh165be382008-12-05 02:36:33 +00002100 if( (wsFlags & WHERE_COLUMN_IN)==0 &&
drh7b4fc6a2007-02-06 13:26:32 +00002101 isSortingIndex(pParse,pWC->pMaskSet,pProbe,iCur,pOrderBy,nEq,&rev) ){
drh165be382008-12-05 02:36:33 +00002102 if( wsFlags==0 ){
2103 wsFlags = WHERE_COLUMN_RANGE;
drh28c4cf42005-07-27 20:41:43 +00002104 }
drh165be382008-12-05 02:36:33 +00002105 wsFlags |= WHERE_ORDERBY;
drh28c4cf42005-07-27 20:41:43 +00002106 if( rev ){
drh165be382008-12-05 02:36:33 +00002107 wsFlags |= WHERE_REVERSE;
drh28c4cf42005-07-27 20:41:43 +00002108 }
2109 }else{
2110 cost += cost*estLog(cost);
drh4f0c5872007-03-26 22:05:01 +00002111 WHERETRACE(("...... orderby increases cost to %.9g\n", cost));
drh51147ba2005-07-23 22:59:55 +00002112 }
drh699b3d42009-02-23 16:52:07 +00002113 }else if( pParse->db->flags & SQLITE_ReverseOrder ){
2114 /* For application testing, randomly reverse the output order for
2115 ** SELECT statements that omit the ORDER BY clause. This will help
2116 ** to find cases where
2117 */
2118 wsFlags |= WHERE_REVERSE;
drhfe05af82005-07-21 03:14:59 +00002119 }
2120
2121 /* Check to see if we can get away with using just the index without
drh51147ba2005-07-23 22:59:55 +00002122 ** ever reading the table. If that is the case, then halve the
2123 ** cost of this index.
drhfe05af82005-07-21 03:14:59 +00002124 */
drh165be382008-12-05 02:36:33 +00002125 if( wsFlags && pSrc->colUsed < (((Bitmask)1)<<(BMS-1)) ){
drhfe05af82005-07-21 03:14:59 +00002126 Bitmask m = pSrc->colUsed;
2127 int j;
2128 for(j=0; j<pProbe->nColumn; j++){
2129 int x = pProbe->aiColumn[j];
2130 if( x<BMS-1 ){
2131 m &= ~(((Bitmask)1)<<x);
2132 }
2133 }
2134 if( m==0 ){
drh165be382008-12-05 02:36:33 +00002135 wsFlags |= WHERE_IDX_ONLY;
drhb37df7b2005-10-13 02:09:49 +00002136 cost /= 2;
drh4f0c5872007-03-26 22:05:01 +00002137 WHERETRACE(("...... idx-only reduces cost to %.9g\n", cost));
drhfe05af82005-07-21 03:14:59 +00002138 }
2139 }
2140
drh51147ba2005-07-23 22:59:55 +00002141 /* If this index has achieved the lowest cost so far, then use it.
drhfe05af82005-07-21 03:14:59 +00002142 */
drh111a6a72008-12-21 03:51:16 +00002143 if( wsFlags!=0 && cost < pCost->rCost ){
2144 pCost->rCost = cost;
2145 pCost->nRow = nRow;
2146 pCost->plan.wsFlags = wsFlags;
2147 pCost->plan.nEq = nEq;
2148 assert( pCost->plan.wsFlags & WHERE_INDEXED );
2149 pCost->plan.u.pIdx = pProbe;
drhfe05af82005-07-21 03:14:59 +00002150 }
2151 }
2152
drhfe05af82005-07-21 03:14:59 +00002153 /* Report the best result
2154 */
drh111a6a72008-12-21 03:51:16 +00002155 pCost->plan.wsFlags |= eqTermMask;
2156 WHERETRACE(("best index is %s, cost=%.9g, nrow=%.9g, wsFlags=%x, nEq=%d\n",
2157 (pCost->plan.wsFlags & WHERE_INDEXED)!=0 ?
2158 pCost->plan.u.pIdx->zName : "(none)", pCost->nRow,
2159 pCost->rCost, pCost->plan.wsFlags, pCost->plan.nEq));
drhfe05af82005-07-21 03:14:59 +00002160}
2161
danielk19771d461462009-04-21 09:02:45 +00002162/*
2163** Find the query plan for accessing table pSrc->pTab. Write the
2164** best query plan and its cost into the WhereCost object supplied
2165** as the last parameter. This function may calculate the cost of
2166** both real and virtual table scans.
2167*/
2168static void bestIndex(
2169 Parse *pParse, /* The parsing context */
2170 WhereClause *pWC, /* The WHERE clause */
2171 struct SrcList_item *pSrc, /* The FROM clause term to search */
2172 Bitmask notReady, /* Mask of cursors that are not available */
2173 ExprList *pOrderBy, /* The ORDER BY clause */
2174 WhereCost *pCost /* Lowest cost query plan */
2175){
2176 if( IsVirtual(pSrc->pTab) ){
2177 sqlite3_index_info *p = 0;
2178 bestVirtualIndex(pParse, pWC, pSrc, notReady, pOrderBy, pCost, &p);
2179 if( p->needToFreeIdxStr ){
2180 sqlite3_free(p->idxStr);
2181 }
2182 sqlite3DbFree(pParse->db, p);
2183 }else{
2184 bestBtreeIndex(pParse, pWC, pSrc, notReady, pOrderBy, pCost);
2185 }
2186}
drhb6c29892004-11-22 19:12:19 +00002187
2188/*
drh2ffb1182004-07-19 19:14:01 +00002189** Disable a term in the WHERE clause. Except, do not disable the term
2190** if it controls a LEFT OUTER JOIN and it did not originate in the ON
2191** or USING clause of that join.
2192**
2193** Consider the term t2.z='ok' in the following queries:
2194**
2195** (1) SELECT * FROM t1 LEFT JOIN t2 ON t1.a=t2.x WHERE t2.z='ok'
2196** (2) SELECT * FROM t1 LEFT JOIN t2 ON t1.a=t2.x AND t2.z='ok'
2197** (3) SELECT * FROM t1, t2 WHERE t1.a=t2.x AND t2.z='ok'
2198**
drh23bf66d2004-12-14 03:34:34 +00002199** The t2.z='ok' is disabled in the in (2) because it originates
drh2ffb1182004-07-19 19:14:01 +00002200** in the ON clause. The term is disabled in (3) because it is not part
2201** of a LEFT OUTER JOIN. In (1), the term is not disabled.
2202**
2203** Disabling a term causes that term to not be tested in the inner loop
drhb6fb62d2005-09-20 08:47:20 +00002204** of the join. Disabling is an optimization. When terms are satisfied
2205** by indices, we disable them to prevent redundant tests in the inner
2206** loop. We would get the correct results if nothing were ever disabled,
2207** but joins might run a little slower. The trick is to disable as much
2208** as we can without disabling too much. If we disabled in (1), we'd get
2209** the wrong answer. See ticket #813.
drh2ffb1182004-07-19 19:14:01 +00002210*/
drh0fcef5e2005-07-19 17:38:22 +00002211static void disableTerm(WhereLevel *pLevel, WhereTerm *pTerm){
2212 if( pTerm
drh165be382008-12-05 02:36:33 +00002213 && ALWAYS((pTerm->wtFlags & TERM_CODED)==0)
drh0fcef5e2005-07-19 17:38:22 +00002214 && (pLevel->iLeftJoin==0 || ExprHasProperty(pTerm->pExpr, EP_FromJoin))
2215 ){
drh165be382008-12-05 02:36:33 +00002216 pTerm->wtFlags |= TERM_CODED;
drh45b1ee42005-08-02 17:48:22 +00002217 if( pTerm->iParent>=0 ){
2218 WhereTerm *pOther = &pTerm->pWC->a[pTerm->iParent];
2219 if( (--pOther->nChild)==0 ){
drhed378002005-07-28 23:12:08 +00002220 disableTerm(pLevel, pOther);
2221 }
drh0fcef5e2005-07-19 17:38:22 +00002222 }
drh2ffb1182004-07-19 19:14:01 +00002223 }
2224}
2225
2226/*
danielk1977b790c6c2008-04-18 10:25:24 +00002227** Apply the affinities associated with the first n columns of index
2228** pIdx to the values in the n registers starting at base.
drh94a11212004-09-25 13:12:14 +00002229*/
danielk1977b790c6c2008-04-18 10:25:24 +00002230static void codeApplyAffinity(Parse *pParse, int base, int n, Index *pIdx){
2231 if( n>0 ){
2232 Vdbe *v = pParse->pVdbe;
2233 assert( v!=0 );
2234 sqlite3VdbeAddOp2(v, OP_Affinity, base, n);
2235 sqlite3IndexAffinityStr(v, pIdx);
2236 sqlite3ExprCacheAffinityChange(pParse, base, n);
2237 }
drh94a11212004-09-25 13:12:14 +00002238}
2239
drhe8b97272005-07-19 22:22:12 +00002240
2241/*
drh51147ba2005-07-23 22:59:55 +00002242** Generate code for a single equality term of the WHERE clause. An equality
2243** term can be either X=expr or X IN (...). pTerm is the term to be
2244** coded.
2245**
drh1db639c2008-01-17 02:36:28 +00002246** The current value for the constraint is left in register iReg.
drh51147ba2005-07-23 22:59:55 +00002247**
2248** For a constraint of the form X=expr, the expression is evaluated and its
2249** result is left on the stack. For constraints of the form X IN (...)
2250** this routine sets up a loop that will iterate over all values of X.
drh94a11212004-09-25 13:12:14 +00002251*/
drh678ccce2008-03-31 18:19:54 +00002252static int codeEqualityTerm(
drh94a11212004-09-25 13:12:14 +00002253 Parse *pParse, /* The parsing context */
drhe23399f2005-07-22 00:31:39 +00002254 WhereTerm *pTerm, /* The term of the WHERE clause to be coded */
drh1db639c2008-01-17 02:36:28 +00002255 WhereLevel *pLevel, /* When level of the FROM clause we are working on */
drh678ccce2008-03-31 18:19:54 +00002256 int iTarget /* Attempt to leave results in this register */
drh94a11212004-09-25 13:12:14 +00002257){
drh0fcef5e2005-07-19 17:38:22 +00002258 Expr *pX = pTerm->pExpr;
drh50b39962006-10-28 00:28:09 +00002259 Vdbe *v = pParse->pVdbe;
drh678ccce2008-03-31 18:19:54 +00002260 int iReg; /* Register holding results */
drh1db639c2008-01-17 02:36:28 +00002261
danielk19772d605492008-10-01 08:43:03 +00002262 assert( iTarget>0 );
drh50b39962006-10-28 00:28:09 +00002263 if( pX->op==TK_EQ ){
drh678ccce2008-03-31 18:19:54 +00002264 iReg = sqlite3ExprCodeTarget(pParse, pX->pRight, iTarget);
drh50b39962006-10-28 00:28:09 +00002265 }else if( pX->op==TK_ISNULL ){
drh678ccce2008-03-31 18:19:54 +00002266 iReg = iTarget;
drh1db639c2008-01-17 02:36:28 +00002267 sqlite3VdbeAddOp2(v, OP_Null, 0, iReg);
danielk1977b3bce662005-01-29 08:32:43 +00002268#ifndef SQLITE_OMIT_SUBQUERY
drh94a11212004-09-25 13:12:14 +00002269 }else{
danielk19779a96b662007-11-29 17:05:18 +00002270 int eType;
danielk1977b3bce662005-01-29 08:32:43 +00002271 int iTab;
drh72e8fa42007-03-28 14:30:06 +00002272 struct InLoop *pIn;
danielk1977b3bce662005-01-29 08:32:43 +00002273
drh50b39962006-10-28 00:28:09 +00002274 assert( pX->op==TK_IN );
drh678ccce2008-03-31 18:19:54 +00002275 iReg = iTarget;
danielk19770cdc0222008-06-26 18:04:03 +00002276 eType = sqlite3FindInIndex(pParse, pX, 0);
danielk1977b3bce662005-01-29 08:32:43 +00002277 iTab = pX->iTable;
drh66a51672008-01-03 00:01:23 +00002278 sqlite3VdbeAddOp2(v, OP_Rewind, iTab, 0);
drhd4e70eb2008-01-02 00:34:36 +00002279 VdbeComment((v, "%.*s", pX->span.n, pX->span.z));
drh111a6a72008-12-21 03:51:16 +00002280 assert( pLevel->plan.wsFlags & WHERE_IN_ABLE );
2281 if( pLevel->u.in.nIn==0 ){
drhb3190c12008-12-08 21:37:14 +00002282 pLevel->addrNxt = sqlite3VdbeMakeLabel(v);
drh72e8fa42007-03-28 14:30:06 +00002283 }
drh111a6a72008-12-21 03:51:16 +00002284 pLevel->u.in.nIn++;
2285 pLevel->u.in.aInLoop =
2286 sqlite3DbReallocOrFree(pParse->db, pLevel->u.in.aInLoop,
2287 sizeof(pLevel->u.in.aInLoop[0])*pLevel->u.in.nIn);
2288 pIn = pLevel->u.in.aInLoop;
drh72e8fa42007-03-28 14:30:06 +00002289 if( pIn ){
drh111a6a72008-12-21 03:51:16 +00002290 pIn += pLevel->u.in.nIn - 1;
drh72e8fa42007-03-28 14:30:06 +00002291 pIn->iCur = iTab;
drh1db639c2008-01-17 02:36:28 +00002292 if( eType==IN_INDEX_ROWID ){
drhb3190c12008-12-08 21:37:14 +00002293 pIn->addrInTop = sqlite3VdbeAddOp2(v, OP_Rowid, iTab, iReg);
drh1db639c2008-01-17 02:36:28 +00002294 }else{
drhb3190c12008-12-08 21:37:14 +00002295 pIn->addrInTop = sqlite3VdbeAddOp3(v, OP_Column, iTab, 0, iReg);
drh1db639c2008-01-17 02:36:28 +00002296 }
2297 sqlite3VdbeAddOp1(v, OP_IsNull, iReg);
drha6110402005-07-28 20:51:19 +00002298 }else{
drh111a6a72008-12-21 03:51:16 +00002299 pLevel->u.in.nIn = 0;
drhe23399f2005-07-22 00:31:39 +00002300 }
danielk1977b3bce662005-01-29 08:32:43 +00002301#endif
drh94a11212004-09-25 13:12:14 +00002302 }
drh0fcef5e2005-07-19 17:38:22 +00002303 disableTerm(pLevel, pTerm);
drh678ccce2008-03-31 18:19:54 +00002304 return iReg;
drh94a11212004-09-25 13:12:14 +00002305}
2306
drh51147ba2005-07-23 22:59:55 +00002307/*
2308** Generate code that will evaluate all == and IN constraints for an
2309** index. The values for all constraints are left on the stack.
2310**
2311** For example, consider table t1(a,b,c,d,e,f) with index i1(a,b,c).
2312** Suppose the WHERE clause is this: a==5 AND b IN (1,2,3) AND c>5 AND c<10
2313** The index has as many as three equality constraints, but in this
2314** example, the third "c" value is an inequality. So only two
2315** constraints are coded. This routine will generate code to evaluate
drh6df2acd2008-12-28 16:55:25 +00002316** a==5 and b IN (1,2,3). The current values for a and b will be stored
2317** in consecutive registers and the index of the first register is returned.
drh51147ba2005-07-23 22:59:55 +00002318**
2319** In the example above nEq==2. But this subroutine works for any value
2320** of nEq including 0. If nEq==0, this routine is nearly a no-op.
2321** The only thing it does is allocate the pLevel->iMem memory cell.
2322**
drh700a2262008-12-17 19:22:15 +00002323** This routine always allocates at least one memory cell and returns
2324** the index of that memory cell. The code that
2325** calls this routine will use that memory cell to store the termination
drh51147ba2005-07-23 22:59:55 +00002326** key value of the loop. If one or more IN operators appear, then
2327** this routine allocates an additional nEq memory cells for internal
2328** use.
2329*/
drh1db639c2008-01-17 02:36:28 +00002330static int codeAllEqualityTerms(
drh51147ba2005-07-23 22:59:55 +00002331 Parse *pParse, /* Parsing context */
2332 WhereLevel *pLevel, /* Which nested loop of the FROM we are coding */
2333 WhereClause *pWC, /* The WHERE clause */
drh1db639c2008-01-17 02:36:28 +00002334 Bitmask notReady, /* Which parts of FROM have not yet been coded */
2335 int nExtraReg /* Number of extra registers to allocate */
drh51147ba2005-07-23 22:59:55 +00002336){
drh111a6a72008-12-21 03:51:16 +00002337 int nEq = pLevel->plan.nEq; /* The number of == or IN constraints to code */
2338 Vdbe *v = pParse->pVdbe; /* The vm under construction */
2339 Index *pIdx; /* The index being used for this loop */
drh51147ba2005-07-23 22:59:55 +00002340 int iCur = pLevel->iTabCur; /* The cursor of the table */
2341 WhereTerm *pTerm; /* A single constraint term */
2342 int j; /* Loop counter */
drh1db639c2008-01-17 02:36:28 +00002343 int regBase; /* Base register */
drh6df2acd2008-12-28 16:55:25 +00002344 int nReg; /* Number of registers to allocate */
drh51147ba2005-07-23 22:59:55 +00002345
drh111a6a72008-12-21 03:51:16 +00002346 /* This module is only called on query plans that use an index. */
2347 assert( pLevel->plan.wsFlags & WHERE_INDEXED );
2348 pIdx = pLevel->plan.u.pIdx;
2349
drh51147ba2005-07-23 22:59:55 +00002350 /* Figure out how many memory cells we will need then allocate them.
drh51147ba2005-07-23 22:59:55 +00002351 */
drh700a2262008-12-17 19:22:15 +00002352 regBase = pParse->nMem + 1;
drh6df2acd2008-12-28 16:55:25 +00002353 nReg = pLevel->plan.nEq + nExtraReg;
2354 pParse->nMem += nReg;
drh51147ba2005-07-23 22:59:55 +00002355
2356 /* Evaluate the equality constraints
2357 */
drhc49de5d2007-01-19 01:06:01 +00002358 assert( pIdx->nColumn>=nEq );
2359 for(j=0; j<nEq; j++){
drh678ccce2008-03-31 18:19:54 +00002360 int r1;
drh51147ba2005-07-23 22:59:55 +00002361 int k = pIdx->aiColumn[j];
drh111a6a72008-12-21 03:51:16 +00002362 pTerm = findTerm(pWC, iCur, k, notReady, pLevel->plan.wsFlags, pIdx);
drh34004ce2008-07-11 16:15:17 +00002363 if( NEVER(pTerm==0) ) break;
drh165be382008-12-05 02:36:33 +00002364 assert( (pTerm->wtFlags & TERM_CODED)==0 );
drh678ccce2008-03-31 18:19:54 +00002365 r1 = codeEqualityTerm(pParse, pTerm, pLevel, regBase+j);
2366 if( r1!=regBase+j ){
drh6df2acd2008-12-28 16:55:25 +00002367 if( nReg==1 ){
2368 sqlite3ReleaseTempReg(pParse, regBase);
2369 regBase = r1;
2370 }else{
2371 sqlite3VdbeAddOp2(v, OP_SCopy, r1, regBase+j);
2372 }
drh678ccce2008-03-31 18:19:54 +00002373 }
drh981642f2008-04-19 14:40:43 +00002374 testcase( pTerm->eOperator & WO_ISNULL );
2375 testcase( pTerm->eOperator & WO_IN );
drh72e8fa42007-03-28 14:30:06 +00002376 if( (pTerm->eOperator & (WO_ISNULL|WO_IN))==0 ){
drhb3190c12008-12-08 21:37:14 +00002377 sqlite3VdbeAddOp2(v, OP_IsNull, regBase+j, pLevel->addrBrk);
drh51147ba2005-07-23 22:59:55 +00002378 }
2379 }
drh1db639c2008-01-17 02:36:28 +00002380 return regBase;
drh51147ba2005-07-23 22:59:55 +00002381}
2382
drh111a6a72008-12-21 03:51:16 +00002383/*
2384** Generate code for the start of the iLevel-th loop in the WHERE clause
2385** implementation described by pWInfo.
2386*/
2387static Bitmask codeOneLoopStart(
2388 WhereInfo *pWInfo, /* Complete information about the WHERE clause */
2389 int iLevel, /* Which level of pWInfo->a[] should be coded */
2390 u8 wctrlFlags, /* One of the WHERE_* flags defined in sqliteInt.h */
2391 Bitmask notReady /* Which tables are currently available */
2392){
2393 int j, k; /* Loop counters */
2394 int iCur; /* The VDBE cursor for the table */
2395 int addrNxt; /* Where to jump to continue with the next IN case */
2396 int omitTable; /* True if we use the index only */
2397 int bRev; /* True if we need to scan in reverse order */
2398 WhereLevel *pLevel; /* The where level to be coded */
2399 WhereClause *pWC; /* Decomposition of the entire WHERE clause */
2400 WhereTerm *pTerm; /* A WHERE clause term */
2401 Parse *pParse; /* Parsing context */
2402 Vdbe *v; /* The prepared stmt under constructions */
2403 struct SrcList_item *pTabItem; /* FROM clause term being coded */
drh23d04d52008-12-23 23:56:22 +00002404 int addrBrk; /* Jump here to break out of the loop */
2405 int addrCont; /* Jump here to continue with next cycle */
2406 int regRowSet; /* Write rowids to this RowSet if non-negative */
danielk19771d461462009-04-21 09:02:45 +00002407
2408 /* Sometimes, this function is required to generate code to do
drh61495262009-04-22 15:32:59 +00002409 ** something with the rowid of each row scanned. Specifically,
2410 ** If pWInfo->regRowSet is non-zero, then the rowid must be inserted
2411 ** into the RowSet object stored in register pWInfo->regRowSet.
danielk19771d461462009-04-21 09:02:45 +00002412 **
danielk19771d461462009-04-21 09:02:45 +00002413 ** Extracting a rowid value from a VDBE cursor is not always a cheap
2414 ** operation, especially if the rowid is being extracted from an index
2415 ** cursor. If the rowid value is available as a by-product of the code
2416 ** generated to create the top of the scan loop, then it can be reused
drh61495262009-04-22 15:32:59 +00002417 ** without extracting it from a cursor. The following two variables are
2418 ** used to communicate the availability of the rowid value to the C-code
2419 ** at the end of this function that generates the rowid-handling VDBE code.
danielk19771d461462009-04-21 09:02:45 +00002420 */
drh61495262009-04-22 15:32:59 +00002421 int iRowidReg = 0; /* Rowid is stored in this register, if not zero */
2422 int iReleaseReg = 0; /* Temp register to free before returning */
drh111a6a72008-12-21 03:51:16 +00002423
2424 pParse = pWInfo->pParse;
2425 v = pParse->pVdbe;
2426 pWC = pWInfo->pWC;
2427 pLevel = &pWInfo->a[iLevel];
2428 pTabItem = &pWInfo->pTabList->a[pLevel->iFrom];
2429 iCur = pTabItem->iCursor;
2430 bRev = (pLevel->plan.wsFlags & WHERE_REVERSE)!=0;
danielk19771d461462009-04-21 09:02:45 +00002431 omitTable = (pLevel->plan.wsFlags & WHERE_IDX_ONLY)!=0
drh1b26c7c2009-04-22 02:15:47 +00002432 && (wctrlFlags & WHERE_FILL_ROWTEST)==0;
drh23d04d52008-12-23 23:56:22 +00002433 regRowSet = pWInfo->regRowSet;
drh111a6a72008-12-21 03:51:16 +00002434
2435 /* Create labels for the "break" and "continue" instructions
2436 ** for the current loop. Jump to addrBrk to break out of a loop.
2437 ** Jump to cont to go immediately to the next iteration of the
2438 ** loop.
2439 **
2440 ** When there is an IN operator, we also have a "addrNxt" label that
2441 ** means to continue with the next IN value combination. When
2442 ** there are no IN operators in the constraints, the "addrNxt" label
2443 ** is the same as "addrBrk".
2444 */
2445 addrBrk = pLevel->addrBrk = pLevel->addrNxt = sqlite3VdbeMakeLabel(v);
2446 addrCont = pLevel->addrCont = sqlite3VdbeMakeLabel(v);
2447
2448 /* If this is the right table of a LEFT OUTER JOIN, allocate and
2449 ** initialize a memory cell that records if this table matches any
2450 ** row of the left table of the join.
2451 */
2452 if( pLevel->iFrom>0 && (pTabItem[0].jointype & JT_LEFT)!=0 ){
2453 pLevel->iLeftJoin = ++pParse->nMem;
2454 sqlite3VdbeAddOp2(v, OP_Integer, 0, pLevel->iLeftJoin);
2455 VdbeComment((v, "init LEFT JOIN no-match flag"));
2456 }
2457
2458#ifndef SQLITE_OMIT_VIRTUALTABLE
2459 if( (pLevel->plan.wsFlags & WHERE_VIRTUALTABLE)!=0 ){
2460 /* Case 0: The table is a virtual-table. Use the VFilter and VNext
2461 ** to access the data.
2462 */
2463 int iReg; /* P3 Value for OP_VFilter */
2464 sqlite3_index_info *pVtabIdx = pLevel->plan.u.pVtabIdx;
2465 int nConstraint = pVtabIdx->nConstraint;
2466 struct sqlite3_index_constraint_usage *aUsage =
2467 pVtabIdx->aConstraintUsage;
2468 const struct sqlite3_index_constraint *aConstraint =
2469 pVtabIdx->aConstraint;
2470
2471 iReg = sqlite3GetTempRange(pParse, nConstraint+2);
2472 pParse->disableColCache++;
2473 for(j=1; j<=nConstraint; j++){
2474 for(k=0; k<nConstraint; k++){
2475 if( aUsage[k].argvIndex==j ){
2476 int iTerm = aConstraint[k].iTermOffset;
2477 assert( pParse->disableColCache );
2478 sqlite3ExprCode(pParse, pWC->a[iTerm].pExpr->pRight, iReg+j+1);
2479 break;
2480 }
2481 }
2482 if( k==nConstraint ) break;
2483 }
2484 assert( pParse->disableColCache );
2485 pParse->disableColCache--;
2486 sqlite3VdbeAddOp2(v, OP_Integer, pVtabIdx->idxNum, iReg);
2487 sqlite3VdbeAddOp2(v, OP_Integer, j-1, iReg+1);
2488 sqlite3VdbeAddOp4(v, OP_VFilter, iCur, addrBrk, iReg, pVtabIdx->idxStr,
2489 pVtabIdx->needToFreeIdxStr ? P4_MPRINTF : P4_STATIC);
drh111a6a72008-12-21 03:51:16 +00002490 pVtabIdx->needToFreeIdxStr = 0;
2491 for(j=0; j<nConstraint; j++){
2492 if( aUsage[j].omit ){
2493 int iTerm = aConstraint[j].iTermOffset;
2494 disableTerm(pLevel, &pWC->a[iTerm]);
2495 }
2496 }
2497 pLevel->op = OP_VNext;
2498 pLevel->p1 = iCur;
2499 pLevel->p2 = sqlite3VdbeCurrentAddr(v);
drh23d04d52008-12-23 23:56:22 +00002500 sqlite3ReleaseTempRange(pParse, iReg, nConstraint+2);
drh111a6a72008-12-21 03:51:16 +00002501 }else
2502#endif /* SQLITE_OMIT_VIRTUALTABLE */
2503
2504 if( pLevel->plan.wsFlags & WHERE_ROWID_EQ ){
2505 /* Case 1: We can directly reference a single row using an
2506 ** equality comparison against the ROWID field. Or
2507 ** we reference multiple rows using a "rowid IN (...)"
2508 ** construct.
2509 */
danielk19771d461462009-04-21 09:02:45 +00002510 iReleaseReg = sqlite3GetTempReg(pParse);
drh111a6a72008-12-21 03:51:16 +00002511 pTerm = findTerm(pWC, iCur, -1, notReady, WO_EQ|WO_IN, 0);
2512 assert( pTerm!=0 );
2513 assert( pTerm->pExpr!=0 );
2514 assert( pTerm->leftCursor==iCur );
2515 assert( omitTable==0 );
danielk19771d461462009-04-21 09:02:45 +00002516 iRowidReg = codeEqualityTerm(pParse, pTerm, pLevel, iReleaseReg);
drh111a6a72008-12-21 03:51:16 +00002517 addrNxt = pLevel->addrNxt;
danielk19771d461462009-04-21 09:02:45 +00002518 sqlite3VdbeAddOp2(v, OP_MustBeInt, iRowidReg, addrNxt);
2519 sqlite3VdbeAddOp3(v, OP_NotExists, iCur, addrNxt, iRowidReg);
drh111a6a72008-12-21 03:51:16 +00002520 VdbeComment((v, "pk"));
2521 pLevel->op = OP_Noop;
2522 }else if( pLevel->plan.wsFlags & WHERE_ROWID_RANGE ){
2523 /* Case 2: We have an inequality comparison against the ROWID field.
2524 */
2525 int testOp = OP_Noop;
2526 int start;
2527 int memEndValue = 0;
2528 WhereTerm *pStart, *pEnd;
2529
2530 assert( omitTable==0 );
2531 pStart = findTerm(pWC, iCur, -1, notReady, WO_GT|WO_GE, 0);
2532 pEnd = findTerm(pWC, iCur, -1, notReady, WO_LT|WO_LE, 0);
2533 if( bRev ){
2534 pTerm = pStart;
2535 pStart = pEnd;
2536 pEnd = pTerm;
2537 }
2538 if( pStart ){
2539 Expr *pX; /* The expression that defines the start bound */
2540 int r1, rTemp; /* Registers for holding the start boundary */
2541
2542 /* The following constant maps TK_xx codes into corresponding
2543 ** seek opcodes. It depends on a particular ordering of TK_xx
2544 */
2545 const u8 aMoveOp[] = {
2546 /* TK_GT */ OP_SeekGt,
2547 /* TK_LE */ OP_SeekLe,
2548 /* TK_LT */ OP_SeekLt,
2549 /* TK_GE */ OP_SeekGe
2550 };
2551 assert( TK_LE==TK_GT+1 ); /* Make sure the ordering.. */
2552 assert( TK_LT==TK_GT+2 ); /* ... of the TK_xx values... */
2553 assert( TK_GE==TK_GT+3 ); /* ... is correcct. */
2554
2555 pX = pStart->pExpr;
2556 assert( pX!=0 );
2557 assert( pStart->leftCursor==iCur );
2558 r1 = sqlite3ExprCodeTemp(pParse, pX->pRight, &rTemp);
2559 sqlite3VdbeAddOp3(v, aMoveOp[pX->op-TK_GT], iCur, addrBrk, r1);
2560 VdbeComment((v, "pk"));
2561 sqlite3ExprCacheAffinityChange(pParse, r1, 1);
2562 sqlite3ReleaseTempReg(pParse, rTemp);
2563 disableTerm(pLevel, pStart);
2564 }else{
2565 sqlite3VdbeAddOp2(v, bRev ? OP_Last : OP_Rewind, iCur, addrBrk);
2566 }
2567 if( pEnd ){
2568 Expr *pX;
2569 pX = pEnd->pExpr;
2570 assert( pX!=0 );
2571 assert( pEnd->leftCursor==iCur );
2572 memEndValue = ++pParse->nMem;
2573 sqlite3ExprCode(pParse, pX->pRight, memEndValue);
2574 if( pX->op==TK_LT || pX->op==TK_GT ){
2575 testOp = bRev ? OP_Le : OP_Ge;
2576 }else{
2577 testOp = bRev ? OP_Lt : OP_Gt;
2578 }
2579 disableTerm(pLevel, pEnd);
2580 }
2581 start = sqlite3VdbeCurrentAddr(v);
2582 pLevel->op = bRev ? OP_Prev : OP_Next;
2583 pLevel->p1 = iCur;
2584 pLevel->p2 = start;
drhca8c4662008-12-28 20:47:02 +00002585 pLevel->p5 = (pStart==0 && pEnd==0) ?1:0;
danielk19771d461462009-04-21 09:02:45 +00002586 if( testOp!=OP_Noop ){
2587 iRowidReg = iReleaseReg = sqlite3GetTempReg(pParse);
2588 sqlite3VdbeAddOp2(v, OP_Rowid, iCur, iRowidReg);
2589 sqlite3VdbeAddOp3(v, testOp, memEndValue, addrBrk, iRowidReg);
2590 sqlite3VdbeChangeP5(v, SQLITE_AFF_NUMERIC | SQLITE_JUMPIFNULL);
drh111a6a72008-12-21 03:51:16 +00002591 }
2592 }else if( pLevel->plan.wsFlags & (WHERE_COLUMN_RANGE|WHERE_COLUMN_EQ) ){
2593 /* Case 3: A scan using an index.
2594 **
2595 ** The WHERE clause may contain zero or more equality
2596 ** terms ("==" or "IN" operators) that refer to the N
2597 ** left-most columns of the index. It may also contain
2598 ** inequality constraints (>, <, >= or <=) on the indexed
2599 ** column that immediately follows the N equalities. Only
2600 ** the right-most column can be an inequality - the rest must
2601 ** use the "==" and "IN" operators. For example, if the
2602 ** index is on (x,y,z), then the following clauses are all
2603 ** optimized:
2604 **
2605 ** x=5
2606 ** x=5 AND y=10
2607 ** x=5 AND y<10
2608 ** x=5 AND y>5 AND y<10
2609 ** x=5 AND y=5 AND z<=10
2610 **
2611 ** The z<10 term of the following cannot be used, only
2612 ** the x=5 term:
2613 **
2614 ** x=5 AND z<10
2615 **
2616 ** N may be zero if there are inequality constraints.
2617 ** If there are no inequality constraints, then N is at
2618 ** least one.
2619 **
2620 ** This case is also used when there are no WHERE clause
2621 ** constraints but an index is selected anyway, in order
2622 ** to force the output order to conform to an ORDER BY.
2623 */
2624 int aStartOp[] = {
2625 0,
2626 0,
2627 OP_Rewind, /* 2: (!start_constraints && startEq && !bRev) */
2628 OP_Last, /* 3: (!start_constraints && startEq && bRev) */
2629 OP_SeekGt, /* 4: (start_constraints && !startEq && !bRev) */
2630 OP_SeekLt, /* 5: (start_constraints && !startEq && bRev) */
2631 OP_SeekGe, /* 6: (start_constraints && startEq && !bRev) */
2632 OP_SeekLe /* 7: (start_constraints && startEq && bRev) */
2633 };
2634 int aEndOp[] = {
2635 OP_Noop, /* 0: (!end_constraints) */
2636 OP_IdxGE, /* 1: (end_constraints && !bRev) */
2637 OP_IdxLT /* 2: (end_constraints && bRev) */
2638 };
2639 int nEq = pLevel->plan.nEq;
2640 int isMinQuery = 0; /* If this is an optimized SELECT min(x).. */
2641 int regBase; /* Base register holding constraint values */
2642 int r1; /* Temp register */
2643 WhereTerm *pRangeStart = 0; /* Inequality constraint at range start */
2644 WhereTerm *pRangeEnd = 0; /* Inequality constraint at range end */
2645 int startEq; /* True if range start uses ==, >= or <= */
2646 int endEq; /* True if range end uses ==, >= or <= */
2647 int start_constraints; /* Start of range is constrained */
2648 int nConstraint; /* Number of constraint terms */
2649 Index *pIdx; /* The index we will be using */
2650 int iIdxCur; /* The VDBE cursor for the index */
drh6df2acd2008-12-28 16:55:25 +00002651 int nExtraReg = 0; /* Number of extra registers needed */
2652 int op; /* Instruction opcode */
drh111a6a72008-12-21 03:51:16 +00002653
2654 pIdx = pLevel->plan.u.pIdx;
2655 iIdxCur = pLevel->iIdxCur;
2656 k = pIdx->aiColumn[nEq]; /* Column for inequality constraints */
2657
drh111a6a72008-12-21 03:51:16 +00002658 /* If this loop satisfies a sort order (pOrderBy) request that
2659 ** was passed to this function to implement a "SELECT min(x) ..."
2660 ** query, then the caller will only allow the loop to run for
2661 ** a single iteration. This means that the first row returned
2662 ** should not have a NULL value stored in 'x'. If column 'x' is
2663 ** the first one after the nEq equality constraints in the index,
2664 ** this requires some special handling.
2665 */
2666 if( (wctrlFlags&WHERE_ORDERBY_MIN)!=0
2667 && (pLevel->plan.wsFlags&WHERE_ORDERBY)
2668 && (pIdx->nColumn>nEq)
2669 ){
2670 /* assert( pOrderBy->nExpr==1 ); */
2671 /* assert( pOrderBy->a[0].pExpr->iColumn==pIdx->aiColumn[nEq] ); */
2672 isMinQuery = 1;
drh6df2acd2008-12-28 16:55:25 +00002673 nExtraReg = 1;
drh111a6a72008-12-21 03:51:16 +00002674 }
2675
2676 /* Find any inequality constraint terms for the start and end
2677 ** of the range.
2678 */
2679 if( pLevel->plan.wsFlags & WHERE_TOP_LIMIT ){
2680 pRangeEnd = findTerm(pWC, iCur, k, notReady, (WO_LT|WO_LE), pIdx);
drh6df2acd2008-12-28 16:55:25 +00002681 nExtraReg = 1;
drh111a6a72008-12-21 03:51:16 +00002682 }
2683 if( pLevel->plan.wsFlags & WHERE_BTM_LIMIT ){
2684 pRangeStart = findTerm(pWC, iCur, k, notReady, (WO_GT|WO_GE), pIdx);
drh6df2acd2008-12-28 16:55:25 +00002685 nExtraReg = 1;
drh111a6a72008-12-21 03:51:16 +00002686 }
2687
drh6df2acd2008-12-28 16:55:25 +00002688 /* Generate code to evaluate all constraint terms using == or IN
2689 ** and store the values of those terms in an array of registers
2690 ** starting at regBase.
2691 */
2692 regBase = codeAllEqualityTerms(pParse, pLevel, pWC, notReady, nExtraReg);
2693 addrNxt = pLevel->addrNxt;
2694
2695
drh111a6a72008-12-21 03:51:16 +00002696 /* If we are doing a reverse order scan on an ascending index, or
2697 ** a forward order scan on a descending index, interchange the
2698 ** start and end terms (pRangeStart and pRangeEnd).
2699 */
2700 if( bRev==(pIdx->aSortOrder[nEq]==SQLITE_SO_ASC) ){
2701 SWAP(WhereTerm *, pRangeEnd, pRangeStart);
2702 }
2703
2704 testcase( pRangeStart && pRangeStart->eOperator & WO_LE );
2705 testcase( pRangeStart && pRangeStart->eOperator & WO_GE );
2706 testcase( pRangeEnd && pRangeEnd->eOperator & WO_LE );
2707 testcase( pRangeEnd && pRangeEnd->eOperator & WO_GE );
2708 startEq = !pRangeStart || pRangeStart->eOperator & (WO_LE|WO_GE);
2709 endEq = !pRangeEnd || pRangeEnd->eOperator & (WO_LE|WO_GE);
2710 start_constraints = pRangeStart || nEq>0;
2711
2712 /* Seek the index cursor to the start of the range. */
2713 nConstraint = nEq;
2714 if( pRangeStart ){
2715 int dcc = pParse->disableColCache;
2716 if( pRangeEnd ){
2717 pParse->disableColCache++;
2718 }
2719 sqlite3ExprCode(pParse, pRangeStart->pExpr->pRight, regBase+nEq);
2720 pParse->disableColCache = dcc;
2721 sqlite3VdbeAddOp2(v, OP_IsNull, regBase+nEq, addrNxt);
2722 nConstraint++;
2723 }else if( isMinQuery ){
2724 sqlite3VdbeAddOp2(v, OP_Null, 0, regBase+nEq);
2725 nConstraint++;
2726 startEq = 0;
2727 start_constraints = 1;
2728 }
2729 codeApplyAffinity(pParse, regBase, nConstraint, pIdx);
2730 op = aStartOp[(start_constraints<<2) + (startEq<<1) + bRev];
2731 assert( op!=0 );
2732 testcase( op==OP_Rewind );
2733 testcase( op==OP_Last );
2734 testcase( op==OP_SeekGt );
2735 testcase( op==OP_SeekGe );
2736 testcase( op==OP_SeekLe );
2737 testcase( op==OP_SeekLt );
2738 sqlite3VdbeAddOp4(v, op, iIdxCur, addrNxt, regBase,
2739 SQLITE_INT_TO_PTR(nConstraint), P4_INT32);
2740
2741 /* Load the value for the inequality constraint at the end of the
2742 ** range (if any).
2743 */
2744 nConstraint = nEq;
2745 if( pRangeEnd ){
2746 sqlite3ExprCode(pParse, pRangeEnd->pExpr->pRight, regBase+nEq);
2747 sqlite3VdbeAddOp2(v, OP_IsNull, regBase+nEq, addrNxt);
2748 codeApplyAffinity(pParse, regBase, nEq+1, pIdx);
2749 nConstraint++;
2750 }
2751
2752 /* Top of the loop body */
2753 pLevel->p2 = sqlite3VdbeCurrentAddr(v);
2754
2755 /* Check if the index cursor is past the end of the range. */
2756 op = aEndOp[(pRangeEnd || nEq) * (1 + bRev)];
2757 testcase( op==OP_Noop );
2758 testcase( op==OP_IdxGE );
2759 testcase( op==OP_IdxLT );
drh6df2acd2008-12-28 16:55:25 +00002760 if( op!=OP_Noop ){
2761 sqlite3VdbeAddOp4(v, op, iIdxCur, addrNxt, regBase,
2762 SQLITE_INT_TO_PTR(nConstraint), P4_INT32);
2763 sqlite3VdbeChangeP5(v, endEq!=bRev ?1:0);
2764 }
drh111a6a72008-12-21 03:51:16 +00002765
2766 /* If there are inequality constraints, check that the value
2767 ** of the table column that the inequality contrains is not NULL.
2768 ** If it is, jump to the next iteration of the loop.
2769 */
2770 r1 = sqlite3GetTempReg(pParse);
2771 testcase( pLevel->plan.wsFlags & WHERE_BTM_LIMIT );
2772 testcase( pLevel->plan.wsFlags & WHERE_TOP_LIMIT );
2773 if( pLevel->plan.wsFlags & (WHERE_BTM_LIMIT|WHERE_TOP_LIMIT) ){
2774 sqlite3VdbeAddOp3(v, OP_Column, iIdxCur, nEq, r1);
2775 sqlite3VdbeAddOp2(v, OP_IsNull, r1, addrCont);
2776 }
danielk19771d461462009-04-21 09:02:45 +00002777 sqlite3ReleaseTempReg(pParse, r1);
drh111a6a72008-12-21 03:51:16 +00002778
2779 /* Seek the table cursor, if required */
drh23d04d52008-12-23 23:56:22 +00002780 disableTerm(pLevel, pRangeStart);
2781 disableTerm(pLevel, pRangeEnd);
danielk19771d461462009-04-21 09:02:45 +00002782 if( !omitTable ){
2783 iRowidReg = iReleaseReg = sqlite3GetTempReg(pParse);
2784 sqlite3VdbeAddOp2(v, OP_IdxRowid, iIdxCur, iRowidReg);
2785 sqlite3VdbeAddOp2(v, OP_Seek, iCur, iRowidReg); /* Deferred seek */
drh111a6a72008-12-21 03:51:16 +00002786 }
drh111a6a72008-12-21 03:51:16 +00002787
2788 /* Record the instruction used to terminate the loop. Disable
2789 ** WHERE clause terms made redundant by the index range scan.
2790 */
2791 pLevel->op = bRev ? OP_Prev : OP_Next;
2792 pLevel->p1 = iIdxCur;
drhdd5f5a62008-12-23 13:35:23 +00002793 }else
2794
drh23d04d52008-12-23 23:56:22 +00002795#ifndef SQLITE_OMIT_OR_OPTIMIZATION
drhdd5f5a62008-12-23 13:35:23 +00002796 if( pLevel->plan.wsFlags & WHERE_MULTI_OR ){
drh111a6a72008-12-21 03:51:16 +00002797 /* Case 4: Two or more separately indexed terms connected by OR
2798 **
2799 ** Example:
2800 **
2801 ** CREATE TABLE t1(a,b,c,d);
2802 ** CREATE INDEX i1 ON t1(a);
2803 ** CREATE INDEX i2 ON t1(b);
2804 ** CREATE INDEX i3 ON t1(c);
2805 **
2806 ** SELECT * FROM t1 WHERE a=5 OR b=7 OR (c=11 AND d=13)
2807 **
2808 ** In the example, there are three indexed terms connected by OR.
danielk19771d461462009-04-21 09:02:45 +00002809 ** The top of the loop looks like this:
drh111a6a72008-12-21 03:51:16 +00002810 **
drh1b26c7c2009-04-22 02:15:47 +00002811 ** Null 1 # Zero the rowset in reg 1
drh111a6a72008-12-21 03:51:16 +00002812 **
danielk19771d461462009-04-21 09:02:45 +00002813 ** Then, for each indexed term, the following. The arguments to
drh1b26c7c2009-04-22 02:15:47 +00002814 ** RowSetTest are such that the rowid of the current row is inserted
2815 ** into the RowSet. If it is already present, control skips the
danielk19771d461462009-04-21 09:02:45 +00002816 ** Gosub opcode and jumps straight to the code generated by WhereEnd().
drh111a6a72008-12-21 03:51:16 +00002817 **
danielk19771d461462009-04-21 09:02:45 +00002818 ** sqlite3WhereBegin(<term>)
drh1b26c7c2009-04-22 02:15:47 +00002819 ** RowSetTest # Insert rowid into rowset
danielk19771d461462009-04-21 09:02:45 +00002820 ** Gosub 2 A
2821 ** sqlite3WhereEnd()
2822 **
2823 ** Following the above, code to terminate the loop. Label A, the target
2824 ** of the Gosub above, jumps to the instruction right after the Goto.
2825 **
drh1b26c7c2009-04-22 02:15:47 +00002826 ** Null 1 # Zero the rowset in reg 1
danielk19771d461462009-04-21 09:02:45 +00002827 ** Goto B # The loop is finished.
2828 **
2829 ** A: <loop body> # Return data, whatever.
2830 **
2831 ** Return 2 # Jump back to the Gosub
2832 **
2833 ** B: <after the loop>
2834 **
drh111a6a72008-12-21 03:51:16 +00002835 */
drh1b26c7c2009-04-22 02:15:47 +00002836 const int f = WHERE_OMIT_OPEN | WHERE_OMIT_CLOSE | WHERE_FILL_ROWTEST;
danielk19771d461462009-04-21 09:02:45 +00002837
drh111a6a72008-12-21 03:51:16 +00002838 WhereClause *pOrWc; /* The OR-clause broken out into subterms */
danielk19771d461462009-04-21 09:02:45 +00002839 WhereTerm *pFinal; /* Final subterm within the OR-clause. */
drhdd5f5a62008-12-23 13:35:23 +00002840 SrcList oneTab; /* Shortened table list */
danielk19771d461462009-04-21 09:02:45 +00002841
2842 int regReturn = ++pParse->nMem; /* Register used with OP_Gosub */
drh1b26c7c2009-04-22 02:15:47 +00002843 int regRowset = ++pParse->nMem; /* Register for RowSet object */
danielk19771d461462009-04-21 09:02:45 +00002844 int iLoopBody = sqlite3VdbeMakeLabel(v); /* Start of loop body */
2845 int iRetInit; /* Address of regReturn init */
2846 int ii;
drh111a6a72008-12-21 03:51:16 +00002847
2848 pTerm = pLevel->plan.u.pTerm;
2849 assert( pTerm!=0 );
2850 assert( pTerm->eOperator==WO_OR );
2851 assert( (pTerm->wtFlags & TERM_ORINFO)!=0 );
2852 pOrWc = &pTerm->u.pOrInfo->wc;
danielk19771d461462009-04-21 09:02:45 +00002853 pFinal = &pOrWc->a[pOrWc->nTerm-1];
drh23d04d52008-12-23 23:56:22 +00002854
danielk19771d461462009-04-21 09:02:45 +00002855 /* Set up a SrcList containing just the table being scanned by this loop. */
drhdd5f5a62008-12-23 13:35:23 +00002856 oneTab.nSrc = 1;
2857 oneTab.nAlloc = 1;
2858 oneTab.a[0] = *pTabItem;
danielk19771d461462009-04-21 09:02:45 +00002859
drh1b26c7c2009-04-22 02:15:47 +00002860 /* Initialize the rowset register to contain NULL. An SQL NULL is
2861 ** equivalent to an empty rowset.
danielk19771d461462009-04-21 09:02:45 +00002862 **
2863 ** Also initialize regReturn to contain the address of the instruction
2864 ** immediately following the OP_Return at the bottom of the loop. This
2865 ** is required in a few obscure LEFT JOIN cases where control jumps
2866 ** over the top of the loop into the body of it. In this case the
2867 ** correct response for the end-of-loop code (the OP_Return) is to
2868 ** fall through to the next instruction, just as an OP_Next does if
2869 ** called on an uninitialized cursor.
2870 */
drh1b26c7c2009-04-22 02:15:47 +00002871 sqlite3VdbeAddOp2(v, OP_Null, 0, regRowset);
danielk19771d461462009-04-21 09:02:45 +00002872 iRetInit = sqlite3VdbeAddOp2(v, OP_Integer, 0, regReturn);
2873
danielk19771d461462009-04-21 09:02:45 +00002874 for(ii=0; ii<pOrWc->nTerm; ii++){
2875 WhereTerm *pOrTerm = &pOrWc->a[ii];
2876 if( pOrTerm->leftCursor==iCur || pOrTerm->eOperator==WO_AND ){
2877 WhereInfo *pSubWInfo; /* Info for single OR-term scan */
2878
2879 /* Loop through table entries that match term pOrTerm. */
2880 pSubWInfo = sqlite3WhereBegin(
drh1b26c7c2009-04-22 02:15:47 +00002881 pParse, &oneTab, pOrTerm->pExpr, 0, f, regRowset);
danielk19771d461462009-04-21 09:02:45 +00002882 if( pSubWInfo ){
2883 int iSet = ((ii==pOrWc->nTerm-1)?-1:ii);
drh1b26c7c2009-04-22 02:15:47 +00002884 /* The call to sqlite3WhereBegin has coded an OP_RowSetTest
2885 ** at instruction iRowSet. Set P2 (the jump target) of this
danielk19771d461462009-04-21 09:02:45 +00002886 ** instruction to jump past the OP_Gosub coded below. This way,
2887 ** if the rowid is already in the hash-table, the body of the
2888 ** loop is not executed.
2889 */
drh1b26c7c2009-04-22 02:15:47 +00002890 int iRowSet = pSubWInfo->iRowidHandler;
drh61495262009-04-22 15:32:59 +00002891 assert( iRowSet>0 || pWInfo->pParse->db->mallocFailed );
drh1b26c7c2009-04-22 02:15:47 +00002892 sqlite3VdbeChangeP2(v, iRowSet, sqlite3VdbeCurrentAddr(v) + 1);
2893 sqlite3VdbeChangeP4(v, iRowSet, (char *)iSet, P4_INT32);
danielk19771d461462009-04-21 09:02:45 +00002894 sqlite3VdbeAddOp2(v, OP_Gosub, regReturn, iLoopBody);
2895
2896 /* Finish the loop through table entries that match term pOrTerm. */
2897 sqlite3WhereEnd(pSubWInfo);
2898 }
drhdd5f5a62008-12-23 13:35:23 +00002899 }
2900 }
danielk19771d461462009-04-21 09:02:45 +00002901 sqlite3VdbeChangeP1(v, iRetInit, sqlite3VdbeCurrentAddr(v));
drh1b26c7c2009-04-22 02:15:47 +00002902 sqlite3VdbeAddOp2(v, OP_Null, 0, regRowset);
danielk19771d461462009-04-21 09:02:45 +00002903 sqlite3VdbeAddOp2(v, OP_Goto, 0, pLevel->addrBrk);
2904 sqlite3VdbeResolveLabel(v, iLoopBody);
2905
2906 pLevel->op = OP_Return;
2907 pLevel->p1 = regReturn;
drh23d04d52008-12-23 23:56:22 +00002908 disableTerm(pLevel, pTerm);
drhdd5f5a62008-12-23 13:35:23 +00002909 }else
drh23d04d52008-12-23 23:56:22 +00002910#endif /* SQLITE_OMIT_OR_OPTIMIZATION */
drhdd5f5a62008-12-23 13:35:23 +00002911
2912 {
drh111a6a72008-12-21 03:51:16 +00002913 /* Case 5: There is no usable index. We must do a complete
2914 ** scan of the entire table.
2915 */
drh699b3d42009-02-23 16:52:07 +00002916 static const u8 aStep[] = { OP_Next, OP_Prev };
2917 static const u8 aStart[] = { OP_Rewind, OP_Last };
2918 assert( bRev==0 || bRev==1 );
drh111a6a72008-12-21 03:51:16 +00002919 assert( omitTable==0 );
drh699b3d42009-02-23 16:52:07 +00002920 pLevel->op = aStep[bRev];
drh111a6a72008-12-21 03:51:16 +00002921 pLevel->p1 = iCur;
drh699b3d42009-02-23 16:52:07 +00002922 pLevel->p2 = 1 + sqlite3VdbeAddOp2(v, aStart[bRev], iCur, addrBrk);
drh111a6a72008-12-21 03:51:16 +00002923 pLevel->p5 = SQLITE_STMTSTATUS_FULLSCAN_STEP;
2924 }
2925 notReady &= ~getMask(pWC->pMaskSet, iCur);
2926
2927 /* Insert code to test every subexpression that can be completely
2928 ** computed using the current set of tables.
2929 */
2930 k = 0;
2931 for(pTerm=pWC->a, j=pWC->nTerm; j>0; j--, pTerm++){
2932 Expr *pE;
2933 testcase( pTerm->wtFlags & TERM_VIRTUAL );
2934 testcase( pTerm->wtFlags & TERM_CODED );
2935 if( pTerm->wtFlags & (TERM_VIRTUAL|TERM_CODED) ) continue;
2936 if( (pTerm->prereqAll & notReady)!=0 ) continue;
2937 pE = pTerm->pExpr;
2938 assert( pE!=0 );
2939 if( pLevel->iLeftJoin && !ExprHasProperty(pE, EP_FromJoin) ){
2940 continue;
2941 }
2942 pParse->disableColCache += k;
2943 sqlite3ExprIfFalse(pParse, pE, addrCont, SQLITE_JUMPIFNULL);
2944 pParse->disableColCache -= k;
2945 k = 1;
2946 pTerm->wtFlags |= TERM_CODED;
2947 }
2948
2949 /* For a LEFT OUTER JOIN, generate code that will record the fact that
2950 ** at least one row of the right table has matched the left table.
2951 */
2952 if( pLevel->iLeftJoin ){
2953 pLevel->addrFirst = sqlite3VdbeCurrentAddr(v);
2954 sqlite3VdbeAddOp2(v, OP_Integer, 1, pLevel->iLeftJoin);
2955 VdbeComment((v, "record LEFT JOIN hit"));
2956 sqlite3ExprClearColumnCache(pParse, pLevel->iTabCur);
2957 sqlite3ExprClearColumnCache(pParse, pLevel->iIdxCur);
2958 for(pTerm=pWC->a, j=0; j<pWC->nTerm; j++, pTerm++){
2959 testcase( pTerm->wtFlags & TERM_VIRTUAL );
2960 testcase( pTerm->wtFlags & TERM_CODED );
2961 if( pTerm->wtFlags & (TERM_VIRTUAL|TERM_CODED) ) continue;
2962 if( (pTerm->prereqAll & notReady)!=0 ) continue;
2963 assert( pTerm->pExpr );
2964 sqlite3ExprIfFalse(pParse, pTerm->pExpr, addrCont, SQLITE_JUMPIFNULL);
2965 pTerm->wtFlags |= TERM_CODED;
2966 }
2967 }
drh23d04d52008-12-23 23:56:22 +00002968
danielk19771d461462009-04-21 09:02:45 +00002969 /* Do the special rowid handling now. */
2970 if( regRowSet ){
2971 assert( regRowSet>0 );
2972 if( iRowidReg==0 ){
2973 /* The rowid was not available as a side-effect of the code
2974 ** genenerated above. So extract it from the cursor now.
2975 */
2976 assert( iReleaseReg==0 );
2977 iReleaseReg = iRowidReg = sqlite3GetTempReg(pParse);
drh23d04d52008-12-23 23:56:22 +00002978#ifndef SQLITE_OMIT_VIRTUALTABLE
danielk19771d461462009-04-21 09:02:45 +00002979 if( (pLevel->plan.wsFlags & WHERE_VIRTUALTABLE)!=0 ){
2980 sqlite3VdbeAddOp2(v, OP_VRowid, iCur, iRowidReg);
2981 }else
drh23d04d52008-12-23 23:56:22 +00002982#endif
danielk19771d461462009-04-21 09:02:45 +00002983 {
2984 sqlite3VdbeAddOp2(v, OP_Rowid, iCur, iRowidReg);
2985 }
drh23d04d52008-12-23 23:56:22 +00002986 }
danielk19771d461462009-04-21 09:02:45 +00002987
danielk19771d461462009-04-21 09:02:45 +00002988 if( pWInfo->wctrlFlags&WHERE_FILL_ROWSET ){
2989 sqlite3VdbeAddOp2(v, OP_RowSetAdd, regRowSet, iRowidReg);
drh61495262009-04-22 15:32:59 +00002990 VVA_ONLY( pWInfo->iRowidHandler = 0; )
danielk19771d461462009-04-21 09:02:45 +00002991 }else{
drh1b26c7c2009-04-22 02:15:47 +00002992 assert( pWInfo->wctrlFlags&WHERE_FILL_ROWTEST );
drh61495262009-04-22 15:32:59 +00002993 pWInfo->iRowidHandler =
2994 sqlite3VdbeAddOp3(v, OP_RowSetTest, regRowSet, 0, iRowidReg);
danielk19771d461462009-04-21 09:02:45 +00002995 }
drh23d04d52008-12-23 23:56:22 +00002996 }
danielk19771d461462009-04-21 09:02:45 +00002997 sqlite3ReleaseTempReg(pParse, iReleaseReg);
drh23d04d52008-12-23 23:56:22 +00002998
drh111a6a72008-12-21 03:51:16 +00002999 return notReady;
3000}
3001
drh549c8b62005-09-19 13:15:23 +00003002#if defined(SQLITE_TEST)
drh84bfda42005-07-15 13:05:21 +00003003/*
3004** The following variable holds a text description of query plan generated
3005** by the most recent call to sqlite3WhereBegin(). Each call to WhereBegin
3006** overwrites the previous. This information is used for testing and
3007** analysis only.
3008*/
3009char sqlite3_query_plan[BMS*2*40]; /* Text of the join */
3010static int nQPlan = 0; /* Next free slow in _query_plan[] */
3011
3012#endif /* SQLITE_TEST */
3013
3014
drh9eff6162006-06-12 21:59:13 +00003015/*
3016** Free a WhereInfo structure
3017*/
drh10fe8402008-10-11 16:47:35 +00003018static void whereInfoFree(sqlite3 *db, WhereInfo *pWInfo){
drh9eff6162006-06-12 21:59:13 +00003019 if( pWInfo ){
3020 int i;
3021 for(i=0; i<pWInfo->nLevel; i++){
drh4be8b512006-06-13 23:51:34 +00003022 sqlite3_index_info *pInfo = pWInfo->a[i].pIdxInfo;
3023 if( pInfo ){
danielk19771d461462009-04-21 09:02:45 +00003024 /* assert( pInfo->needToFreeIdxStr==0 || db->mallocFailed ); */
danielk197780442942008-12-24 11:25:39 +00003025 if( pInfo->needToFreeIdxStr ){
3026 sqlite3_free(pInfo->idxStr);
danielk1977be229652009-03-20 14:18:51 +00003027 }
drh633e6d52008-07-28 19:34:53 +00003028 sqlite3DbFree(db, pInfo);
danielk1977be8a7832006-06-13 15:00:54 +00003029 }
drh9eff6162006-06-12 21:59:13 +00003030 }
drh111a6a72008-12-21 03:51:16 +00003031 whereClauseClear(pWInfo->pWC);
drh633e6d52008-07-28 19:34:53 +00003032 sqlite3DbFree(db, pWInfo);
drh9eff6162006-06-12 21:59:13 +00003033 }
3034}
3035
drh94a11212004-09-25 13:12:14 +00003036
3037/*
drhe3184742002-06-19 14:27:05 +00003038** Generate the beginning of the loop used for WHERE clause processing.
drhacf3b982005-01-03 01:27:18 +00003039** The return value is a pointer to an opaque structure that contains
drh75897232000-05-29 14:26:00 +00003040** information needed to terminate the loop. Later, the calling routine
danielk19774adee202004-05-08 08:23:19 +00003041** should invoke sqlite3WhereEnd() with the return value of this function
drh75897232000-05-29 14:26:00 +00003042** in order to complete the WHERE clause processing.
3043**
3044** If an error occurs, this routine returns NULL.
drhc27a1ce2002-06-14 20:58:45 +00003045**
3046** The basic idea is to do a nested loop, one loop for each table in
3047** the FROM clause of a select. (INSERT and UPDATE statements are the
3048** same as a SELECT with only a single table in the FROM clause.) For
3049** example, if the SQL is this:
3050**
3051** SELECT * FROM t1, t2, t3 WHERE ...;
3052**
3053** Then the code generated is conceptually like the following:
3054**
3055** foreach row1 in t1 do \ Code generated
danielk19774adee202004-05-08 08:23:19 +00003056** foreach row2 in t2 do |-- by sqlite3WhereBegin()
drhc27a1ce2002-06-14 20:58:45 +00003057** foreach row3 in t3 do /
3058** ...
3059** end \ Code generated
danielk19774adee202004-05-08 08:23:19 +00003060** end |-- by sqlite3WhereEnd()
drhc27a1ce2002-06-14 20:58:45 +00003061** end /
3062**
drh29dda4a2005-07-21 18:23:20 +00003063** Note that the loops might not be nested in the order in which they
3064** appear in the FROM clause if a different order is better able to make
drh51147ba2005-07-23 22:59:55 +00003065** use of indices. Note also that when the IN operator appears in
3066** the WHERE clause, it might result in additional nested loops for
3067** scanning through all values on the right-hand side of the IN.
drh29dda4a2005-07-21 18:23:20 +00003068**
drhc27a1ce2002-06-14 20:58:45 +00003069** There are Btree cursors associated with each table. t1 uses cursor
drh6a3ea0e2003-05-02 14:32:12 +00003070** number pTabList->a[0].iCursor. t2 uses the cursor pTabList->a[1].iCursor.
3071** And so forth. This routine generates code to open those VDBE cursors
danielk19774adee202004-05-08 08:23:19 +00003072** and sqlite3WhereEnd() generates the code to close them.
drhc27a1ce2002-06-14 20:58:45 +00003073**
drhe6f85e72004-12-25 01:03:13 +00003074** The code that sqlite3WhereBegin() generates leaves the cursors named
3075** in pTabList pointing at their appropriate entries. The [...] code
drhf0863fe2005-06-12 21:35:51 +00003076** can use OP_Column and OP_Rowid opcodes on these cursors to extract
drhe6f85e72004-12-25 01:03:13 +00003077** data from the various tables of the loop.
3078**
drhc27a1ce2002-06-14 20:58:45 +00003079** If the WHERE clause is empty, the foreach loops must each scan their
3080** entire tables. Thus a three-way join is an O(N^3) operation. But if
3081** the tables have indices and there are terms in the WHERE clause that
3082** refer to those indices, a complete table scan can be avoided and the
3083** code will run much faster. Most of the work of this routine is checking
3084** to see if there are indices that can be used to speed up the loop.
3085**
3086** Terms of the WHERE clause are also used to limit which rows actually
3087** make it to the "..." in the middle of the loop. After each "foreach",
3088** terms of the WHERE clause that use only terms in that loop and outer
3089** loops are evaluated and if false a jump is made around all subsequent
3090** inner loops (or around the "..." if the test occurs within the inner-
3091** most loop)
3092**
3093** OUTER JOINS
3094**
3095** An outer join of tables t1 and t2 is conceptally coded as follows:
3096**
3097** foreach row1 in t1 do
3098** flag = 0
3099** foreach row2 in t2 do
3100** start:
3101** ...
3102** flag = 1
3103** end
drhe3184742002-06-19 14:27:05 +00003104** if flag==0 then
3105** move the row2 cursor to a null row
3106** goto start
3107** fi
drhc27a1ce2002-06-14 20:58:45 +00003108** end
3109**
drhe3184742002-06-19 14:27:05 +00003110** ORDER BY CLAUSE PROCESSING
3111**
3112** *ppOrderBy is a pointer to the ORDER BY clause of a SELECT statement,
3113** if there is one. If there is no ORDER BY clause or if this routine
3114** is called from an UPDATE or DELETE statement, then ppOrderBy is NULL.
3115**
3116** If an index can be used so that the natural output order of the table
3117** scan is correct for the ORDER BY clause, then that index is used and
3118** *ppOrderBy is set to NULL. This is an optimization that prevents an
3119** unnecessary sort of the result set if an index appropriate for the
3120** ORDER BY clause already exists.
3121**
3122** If the where clause loops cannot be arranged to provide the correct
3123** output order, then the *ppOrderBy is unchanged.
drh75897232000-05-29 14:26:00 +00003124*/
danielk19774adee202004-05-08 08:23:19 +00003125WhereInfo *sqlite3WhereBegin(
danielk1977ed326d72004-11-16 15:50:19 +00003126 Parse *pParse, /* The parser context */
3127 SrcList *pTabList, /* A list of all tables to be scanned */
3128 Expr *pWhere, /* The WHERE clause */
danielk1977a9d1ccb2008-01-05 17:39:29 +00003129 ExprList **ppOrderBy, /* An ORDER BY clause, or NULL */
drh23d04d52008-12-23 23:56:22 +00003130 u8 wctrlFlags, /* One of the WHERE_* flags defined in sqliteInt.h */
3131 int regRowSet /* Register hold RowSet if WHERE_FILL_ROWSET is set */
drh75897232000-05-29 14:26:00 +00003132){
3133 int i; /* Loop counter */
danielk1977be229652009-03-20 14:18:51 +00003134 int nByteWInfo; /* Num. bytes allocated for WhereInfo struct */
drh75897232000-05-29 14:26:00 +00003135 WhereInfo *pWInfo; /* Will become the return value of this function */
3136 Vdbe *v = pParse->pVdbe; /* The virtual database engine */
drhfe05af82005-07-21 03:14:59 +00003137 Bitmask notReady; /* Cursors that are not yet positioned */
drh111a6a72008-12-21 03:51:16 +00003138 WhereMaskSet *pMaskSet; /* The expression mask set */
drh111a6a72008-12-21 03:51:16 +00003139 WhereClause *pWC; /* Decomposition of the WHERE clause */
drh9012bcb2004-12-19 00:11:35 +00003140 struct SrcList_item *pTabItem; /* A single entry from pTabList */
3141 WhereLevel *pLevel; /* A single level in the pWInfo list */
drh29dda4a2005-07-21 18:23:20 +00003142 int iFrom; /* First unused FROM clause element */
drh111a6a72008-12-21 03:51:16 +00003143 int andFlags; /* AND-ed combination of all pWC->a[].wtFlags */
drh17435752007-08-16 04:30:38 +00003144 sqlite3 *db; /* Database connection */
danielk1977a9d1ccb2008-01-05 17:39:29 +00003145 ExprList *pOrderBy = 0;
drh75897232000-05-29 14:26:00 +00003146
drh29dda4a2005-07-21 18:23:20 +00003147 /* The number of tables in the FROM clause is limited by the number of
drh1398ad32005-01-19 23:24:50 +00003148 ** bits in a Bitmask
3149 */
drh29dda4a2005-07-21 18:23:20 +00003150 if( pTabList->nSrc>BMS ){
3151 sqlite3ErrorMsg(pParse, "at most %d tables in a join", BMS);
drh1398ad32005-01-19 23:24:50 +00003152 return 0;
3153 }
3154
danielk1977a9d1ccb2008-01-05 17:39:29 +00003155 if( ppOrderBy ){
3156 pOrderBy = *ppOrderBy;
3157 }
3158
drh75897232000-05-29 14:26:00 +00003159 /* Allocate and initialize the WhereInfo structure that will become the
danielk1977be229652009-03-20 14:18:51 +00003160 ** return value. A single allocation is used to store the WhereInfo
3161 ** struct, the contents of WhereInfo.a[], the WhereClause structure
3162 ** and the WhereMaskSet structure. Since WhereClause contains an 8-byte
3163 ** field (type Bitmask) it must be aligned on an 8-byte boundary on
3164 ** some architectures. Hence the ROUND8() below.
drh75897232000-05-29 14:26:00 +00003165 */
drh17435752007-08-16 04:30:38 +00003166 db = pParse->db;
danielk1977be229652009-03-20 14:18:51 +00003167 nByteWInfo = ROUND8(sizeof(WhereInfo)+(pTabList->nSrc-1)*sizeof(WhereLevel));
3168 pWInfo = sqlite3DbMallocZero(db,
3169 nByteWInfo +
3170 sizeof(WhereClause) +
3171 sizeof(WhereMaskSet)
3172 );
drh17435752007-08-16 04:30:38 +00003173 if( db->mallocFailed ){
danielk197785574e32008-10-06 05:32:18 +00003174 goto whereBeginError;
drh75897232000-05-29 14:26:00 +00003175 }
danielk197770b6d572006-06-19 04:49:34 +00003176 pWInfo->nLevel = pTabList->nSrc;
drh75897232000-05-29 14:26:00 +00003177 pWInfo->pParse = pParse;
3178 pWInfo->pTabList = pTabList;
danielk19774adee202004-05-08 08:23:19 +00003179 pWInfo->iBreak = sqlite3VdbeMakeLabel(v);
danielk19771d461462009-04-21 09:02:45 +00003180 pWInfo->regRowSet = regRowSet;
danielk1977be229652009-03-20 14:18:51 +00003181 pWInfo->pWC = pWC = (WhereClause *)&((u8 *)pWInfo)[nByteWInfo];
drh6df2acd2008-12-28 16:55:25 +00003182 pWInfo->wctrlFlags = wctrlFlags;
drh111a6a72008-12-21 03:51:16 +00003183 pMaskSet = (WhereMaskSet*)&pWC[1];
drh1b26c7c2009-04-22 02:15:47 +00003184 assert( regRowSet==0 || (wctrlFlags&(WHERE_FILL_ROWSET|WHERE_FILL_ROWTEST)) );
drh08192d52002-04-30 19:20:28 +00003185
drh111a6a72008-12-21 03:51:16 +00003186 /* Split the WHERE clause into separate subexpressions where each
3187 ** subexpression is separated by an AND operator.
3188 */
3189 initMaskSet(pMaskSet);
3190 whereClauseInit(pWC, pParse, pMaskSet);
3191 sqlite3ExprCodeConstants(pParse, pWhere);
3192 whereSplit(pWC, pWhere, TK_AND);
3193
drh08192d52002-04-30 19:20:28 +00003194 /* Special case: a WHERE clause that is constant. Evaluate the
3195 ** expression and either jump over all of the code or fall thru.
3196 */
drh0a168372007-06-08 00:20:47 +00003197 if( pWhere && (pTabList->nSrc==0 || sqlite3ExprIsConstantNotJoin(pWhere)) ){
drh35573352008-01-08 23:54:25 +00003198 sqlite3ExprIfFalse(pParse, pWhere, pWInfo->iBreak, SQLITE_JUMPIFNULL);
drhdf199a22002-06-14 22:38:41 +00003199 pWhere = 0;
drh08192d52002-04-30 19:20:28 +00003200 }
drh75897232000-05-29 14:26:00 +00003201
drh42165be2008-03-26 14:56:34 +00003202 /* Assign a bit from the bitmask to every term in the FROM clause.
3203 **
3204 ** When assigning bitmask values to FROM clause cursors, it must be
3205 ** the case that if X is the bitmask for the N-th FROM clause term then
3206 ** the bitmask for all FROM clause terms to the left of the N-th term
3207 ** is (X-1). An expression from the ON clause of a LEFT JOIN can use
3208 ** its Expr.iRightJoinTable value to find the bitmask of the right table
3209 ** of the join. Subtracting one from the right table bitmask gives a
3210 ** bitmask for all tables to the left of the join. Knowing the bitmask
3211 ** for all tables to the left of a left join is important. Ticket #3015.
3212 */
3213 for(i=0; i<pTabList->nSrc; i++){
drh111a6a72008-12-21 03:51:16 +00003214 createMask(pMaskSet, pTabList->a[i].iCursor);
drh42165be2008-03-26 14:56:34 +00003215 }
3216#ifndef NDEBUG
3217 {
3218 Bitmask toTheLeft = 0;
3219 for(i=0; i<pTabList->nSrc; i++){
drh111a6a72008-12-21 03:51:16 +00003220 Bitmask m = getMask(pMaskSet, pTabList->a[i].iCursor);
drh42165be2008-03-26 14:56:34 +00003221 assert( (m-1)==toTheLeft );
3222 toTheLeft |= m;
3223 }
3224 }
3225#endif
3226
drh29dda4a2005-07-21 18:23:20 +00003227 /* Analyze all of the subexpressions. Note that exprAnalyze() might
3228 ** add new virtual terms onto the end of the WHERE clause. We do not
3229 ** want to analyze these virtual terms, so start analyzing at the end
drhb6fb62d2005-09-20 08:47:20 +00003230 ** and work forward so that the added virtual terms are never processed.
drh75897232000-05-29 14:26:00 +00003231 */
drh111a6a72008-12-21 03:51:16 +00003232 exprAnalyzeAll(pTabList, pWC);
drh17435752007-08-16 04:30:38 +00003233 if( db->mallocFailed ){
danielk197785574e32008-10-06 05:32:18 +00003234 goto whereBeginError;
drh0bbaa1b2005-08-19 19:14:12 +00003235 }
drh75897232000-05-29 14:26:00 +00003236
drh29dda4a2005-07-21 18:23:20 +00003237 /* Chose the best index to use for each table in the FROM clause.
3238 **
drh51147ba2005-07-23 22:59:55 +00003239 ** This loop fills in the following fields:
3240 **
3241 ** pWInfo->a[].pIdx The index to use for this level of the loop.
drh165be382008-12-05 02:36:33 +00003242 ** pWInfo->a[].wsFlags WHERE_xxx flags associated with pIdx
drh51147ba2005-07-23 22:59:55 +00003243 ** pWInfo->a[].nEq The number of == and IN constraints
danielk197785574e32008-10-06 05:32:18 +00003244 ** pWInfo->a[].iFrom Which term of the FROM clause is being coded
drh51147ba2005-07-23 22:59:55 +00003245 ** pWInfo->a[].iTabCur The VDBE cursor for the database table
3246 ** pWInfo->a[].iIdxCur The VDBE cursor for the index
drh111a6a72008-12-21 03:51:16 +00003247 ** pWInfo->a[].pTerm When wsFlags==WO_OR, the OR-clause term
drh51147ba2005-07-23 22:59:55 +00003248 **
3249 ** This loop also figures out the nesting order of tables in the FROM
3250 ** clause.
drh75897232000-05-29 14:26:00 +00003251 */
drhfe05af82005-07-21 03:14:59 +00003252 notReady = ~(Bitmask)0;
drh9012bcb2004-12-19 00:11:35 +00003253 pTabItem = pTabList->a;
3254 pLevel = pWInfo->a;
drh943af3c2005-07-29 19:43:58 +00003255 andFlags = ~0;
drh4f0c5872007-03-26 22:05:01 +00003256 WHERETRACE(("*** Optimizer Start ***\n"));
drh29dda4a2005-07-21 18:23:20 +00003257 for(i=iFrom=0, pLevel=pWInfo->a; i<pTabList->nSrc; i++, pLevel++){
drh111a6a72008-12-21 03:51:16 +00003258 WhereCost bestPlan; /* Most efficient plan seen so far */
drh29dda4a2005-07-21 18:23:20 +00003259 Index *pIdx; /* Index for FROM table at pTabItem */
drh29dda4a2005-07-21 18:23:20 +00003260 int j; /* For looping over FROM tables */
drh02afc862006-01-20 18:10:57 +00003261 int bestJ = 0; /* The value of j */
drh29dda4a2005-07-21 18:23:20 +00003262 Bitmask m; /* Bitmask value for j or bestJ */
drh570b9352006-02-01 02:45:02 +00003263 int once = 0; /* True when first table is seen */
drh29dda4a2005-07-21 18:23:20 +00003264
drh111a6a72008-12-21 03:51:16 +00003265 memset(&bestPlan, 0, sizeof(bestPlan));
3266 bestPlan.rCost = SQLITE_BIG_DBL;
drh29dda4a2005-07-21 18:23:20 +00003267 for(j=iFrom, pTabItem=&pTabList->a[j]; j<pTabList->nSrc; j++, pTabItem++){
danielk19771d461462009-04-21 09:02:45 +00003268 int doNotReorder; /* True if this table should not be reordered */
3269 WhereCost sCost; /* Cost information from best[Virtual]Index() */
3270 ExprList *pOrderBy; /* ORDER BY clause for index to optimize */
drhdf26fd52006-06-06 11:45:54 +00003271
drh61dfc312006-12-16 16:25:15 +00003272 doNotReorder = (pTabItem->jointype & (JT_LEFT|JT_CROSS))!=0;
drhdf26fd52006-06-06 11:45:54 +00003273 if( once && doNotReorder ) break;
drh111a6a72008-12-21 03:51:16 +00003274 m = getMask(pMaskSet, pTabItem->iCursor);
drh29dda4a2005-07-21 18:23:20 +00003275 if( (m & notReady)==0 ){
3276 if( j==iFrom ) iFrom++;
3277 continue;
3278 }
danielk19771d461462009-04-21 09:02:45 +00003279 pOrderBy = ((i==0 && ppOrderBy )?*ppOrderBy:0);
3280
drh9eff6162006-06-12 21:59:13 +00003281 assert( pTabItem->pTab );
3282#ifndef SQLITE_OMIT_VIRTUALTABLE
drh4cbdda92006-06-14 19:00:20 +00003283 if( IsVirtual(pTabItem->pTab) ){
danielk19771d461462009-04-21 09:02:45 +00003284 sqlite3_index_info **pp = &pWInfo->a[j].pIdxInfo;
3285 bestVirtualIndex(pParse, pWC, pTabItem, notReady, pOrderBy, &sCost, pp);
drh9eff6162006-06-12 21:59:13 +00003286 }else
3287#endif
3288 {
danielk19771d461462009-04-21 09:02:45 +00003289 bestBtreeIndex(pParse, pWC, pTabItem, notReady, pOrderBy, &sCost);
drh9eff6162006-06-12 21:59:13 +00003290 }
danielk1977992347f2008-12-30 09:45:45 +00003291 if( once==0 || sCost.rCost<bestPlan.rCost ){
drh570b9352006-02-01 02:45:02 +00003292 once = 1;
drh111a6a72008-12-21 03:51:16 +00003293 bestPlan = sCost;
drh29dda4a2005-07-21 18:23:20 +00003294 bestJ = j;
3295 }
drhdf26fd52006-06-06 11:45:54 +00003296 if( doNotReorder ) break;
drh29dda4a2005-07-21 18:23:20 +00003297 }
danielk1977992347f2008-12-30 09:45:45 +00003298 assert( once );
3299 assert( notReady & getMask(pMaskSet, pTabList->a[bestJ].iCursor) );
drhcb041342008-06-12 00:07:29 +00003300 WHERETRACE(("*** Optimizer selects table %d for loop %d\n", bestJ,
drh3dec2232005-09-10 15:28:09 +00003301 pLevel-pWInfo->a));
drh111a6a72008-12-21 03:51:16 +00003302 if( (bestPlan.plan.wsFlags & WHERE_ORDERBY)!=0 ){
drhfe05af82005-07-21 03:14:59 +00003303 *ppOrderBy = 0;
drhc4a3c772001-04-04 11:48:57 +00003304 }
drh111a6a72008-12-21 03:51:16 +00003305 andFlags &= bestPlan.plan.wsFlags;
3306 pLevel->plan = bestPlan.plan;
3307 if( bestPlan.plan.wsFlags & WHERE_INDEXED ){
drh9012bcb2004-12-19 00:11:35 +00003308 pLevel->iIdxCur = pParse->nTab++;
drhfe05af82005-07-21 03:14:59 +00003309 }else{
3310 pLevel->iIdxCur = -1;
drh6b563442001-11-07 16:48:26 +00003311 }
drh111a6a72008-12-21 03:51:16 +00003312 notReady &= ~getMask(pMaskSet, pTabList->a[bestJ].iCursor);
shaned87897d2009-01-30 05:40:27 +00003313 pLevel->iFrom = (u8)bestJ;
danielk197785574e32008-10-06 05:32:18 +00003314
3315 /* Check that if the table scanned by this loop iteration had an
3316 ** INDEXED BY clause attached to it, that the named index is being
3317 ** used for the scan. If not, then query compilation has failed.
3318 ** Return an error.
3319 */
3320 pIdx = pTabList->a[bestJ].pIndex;
drh171256c2009-01-08 03:11:19 +00003321 if( pIdx ){
3322 if( (bestPlan.plan.wsFlags & WHERE_INDEXED)==0 ){
3323 sqlite3ErrorMsg(pParse, "cannot use index: %s", pIdx->zName);
3324 goto whereBeginError;
3325 }else{
3326 /* If an INDEXED BY clause is used, the bestIndex() function is
3327 ** guaranteed to find the index specified in the INDEXED BY clause
3328 ** if it find an index at all. */
3329 assert( bestPlan.plan.u.pIdx==pIdx );
3330 }
danielk197785574e32008-10-06 05:32:18 +00003331 }
drh75897232000-05-29 14:26:00 +00003332 }
drh4f0c5872007-03-26 22:05:01 +00003333 WHERETRACE(("*** Optimizer Finished ***\n"));
danielk19771d461462009-04-21 09:02:45 +00003334 if( pParse->nErr || db->mallocFailed ){
danielk197780442942008-12-24 11:25:39 +00003335 goto whereBeginError;
3336 }
drh75897232000-05-29 14:26:00 +00003337
drh943af3c2005-07-29 19:43:58 +00003338 /* If the total query only selects a single row, then the ORDER BY
3339 ** clause is irrelevant.
3340 */
3341 if( (andFlags & WHERE_UNIQUE)!=0 && ppOrderBy ){
3342 *ppOrderBy = 0;
3343 }
3344
drh08c88eb2008-04-10 13:33:18 +00003345 /* If the caller is an UPDATE or DELETE statement that is requesting
3346 ** to use a one-pass algorithm, determine if this is appropriate.
3347 ** The one-pass algorithm only works if the WHERE clause constraints
3348 ** the statement to update a single row.
3349 */
drh165be382008-12-05 02:36:33 +00003350 assert( (wctrlFlags & WHERE_ONEPASS_DESIRED)==0 || pWInfo->nLevel==1 );
3351 if( (wctrlFlags & WHERE_ONEPASS_DESIRED)!=0 && (andFlags & WHERE_UNIQUE)!=0 ){
drh08c88eb2008-04-10 13:33:18 +00003352 pWInfo->okOnePass = 1;
drh111a6a72008-12-21 03:51:16 +00003353 pWInfo->a[0].plan.wsFlags &= ~WHERE_IDX_ONLY;
drh08c88eb2008-04-10 13:33:18 +00003354 }
3355
drh9012bcb2004-12-19 00:11:35 +00003356 /* Open all tables in the pTabList and any indices selected for
3357 ** searching those tables.
3358 */
3359 sqlite3CodeVerifySchema(pParse, -1); /* Insert the cookie verifier Goto */
drh29dda4a2005-07-21 18:23:20 +00003360 for(i=0, pLevel=pWInfo->a; i<pTabList->nSrc; i++, pLevel++){
danielk1977da184232006-01-05 11:34:32 +00003361 Table *pTab; /* Table to open */
danielk1977da184232006-01-05 11:34:32 +00003362 int iDb; /* Index of database containing table/index */
drh9012bcb2004-12-19 00:11:35 +00003363
drhecc92422005-09-10 16:46:12 +00003364#ifndef SQLITE_OMIT_EXPLAIN
3365 if( pParse->explain==2 ){
3366 char *zMsg;
3367 struct SrcList_item *pItem = &pTabList->a[pLevel->iFrom];
danielk19771e536952007-08-16 10:09:01 +00003368 zMsg = sqlite3MPrintf(db, "TABLE %s", pItem->zName);
drhecc92422005-09-10 16:46:12 +00003369 if( pItem->zAlias ){
drh633e6d52008-07-28 19:34:53 +00003370 zMsg = sqlite3MAppendf(db, zMsg, "%s AS %s", zMsg, pItem->zAlias);
drhecc92422005-09-10 16:46:12 +00003371 }
drh111a6a72008-12-21 03:51:16 +00003372 if( (pLevel->plan.wsFlags & WHERE_INDEXED)!=0 ){
3373 zMsg = sqlite3MAppendf(db, zMsg, "%s WITH INDEX %s",
3374 zMsg, pLevel->plan.u.pIdx->zName);
drh46129af2008-12-30 16:18:47 +00003375 }else if( pLevel->plan.wsFlags & WHERE_MULTI_OR ){
3376 zMsg = sqlite3MAppendf(db, zMsg, "%s VIA MULTI-INDEX UNION", zMsg);
drh111a6a72008-12-21 03:51:16 +00003377 }else if( pLevel->plan.wsFlags & (WHERE_ROWID_EQ|WHERE_ROWID_RANGE) ){
drh633e6d52008-07-28 19:34:53 +00003378 zMsg = sqlite3MAppendf(db, zMsg, "%s USING PRIMARY KEY", zMsg);
drhecc92422005-09-10 16:46:12 +00003379 }
drh9eff6162006-06-12 21:59:13 +00003380#ifndef SQLITE_OMIT_VIRTUALTABLE
drh111a6a72008-12-21 03:51:16 +00003381 else if( (pLevel->plan.wsFlags & WHERE_VIRTUALTABLE)!=0 ){
3382 sqlite3_index_info *pVtabIdx = pLevel->plan.u.pVtabIdx;
drh633e6d52008-07-28 19:34:53 +00003383 zMsg = sqlite3MAppendf(db, zMsg, "%s VIRTUAL TABLE INDEX %d:%s", zMsg,
drh111a6a72008-12-21 03:51:16 +00003384 pVtabIdx->idxNum, pVtabIdx->idxStr);
drh9eff6162006-06-12 21:59:13 +00003385 }
3386#endif
drh111a6a72008-12-21 03:51:16 +00003387 if( pLevel->plan.wsFlags & WHERE_ORDERBY ){
drh633e6d52008-07-28 19:34:53 +00003388 zMsg = sqlite3MAppendf(db, zMsg, "%s ORDER BY", zMsg);
drhe2b39092006-04-21 09:38:36 +00003389 }
drh66a51672008-01-03 00:01:23 +00003390 sqlite3VdbeAddOp4(v, OP_Explain, i, pLevel->iFrom, 0, zMsg, P4_DYNAMIC);
drhecc92422005-09-10 16:46:12 +00003391 }
3392#endif /* SQLITE_OMIT_EXPLAIN */
drh29dda4a2005-07-21 18:23:20 +00003393 pTabItem = &pTabList->a[pLevel->iFrom];
drh9012bcb2004-12-19 00:11:35 +00003394 pTab = pTabItem->pTab;
danielk1977da184232006-01-05 11:34:32 +00003395 iDb = sqlite3SchemaToIndex(pParse->db, pTab->pSchema);
drh7d10d5a2008-08-20 16:35:10 +00003396 if( (pTab->tabFlags & TF_Ephemeral)!=0 || pTab->pSelect ) continue;
drh9eff6162006-06-12 21:59:13 +00003397#ifndef SQLITE_OMIT_VIRTUALTABLE
drh111a6a72008-12-21 03:51:16 +00003398 if( (pLevel->plan.wsFlags & WHERE_VIRTUALTABLE)!=0 ){
danielk197793626f42006-06-20 13:07:27 +00003399 int iCur = pTabItem->iCursor;
drh66a51672008-01-03 00:01:23 +00003400 sqlite3VdbeAddOp4(v, OP_VOpen, iCur, 0, 0,
3401 (const char*)pTab->pVtab, P4_VTAB);
drh9eff6162006-06-12 21:59:13 +00003402 }else
3403#endif
drh6df2acd2008-12-28 16:55:25 +00003404 if( (pLevel->plan.wsFlags & WHERE_IDX_ONLY)==0
3405 && (wctrlFlags & WHERE_OMIT_OPEN)==0 ){
drh08c88eb2008-04-10 13:33:18 +00003406 int op = pWInfo->okOnePass ? OP_OpenWrite : OP_OpenRead;
3407 sqlite3OpenTable(pParse, pTabItem->iCursor, iDb, pTab, op);
danielk197723432972008-11-17 16:42:00 +00003408 if( !pWInfo->okOnePass && pTab->nCol<BMS ){
danielk19779792eef2006-01-13 15:58:43 +00003409 Bitmask b = pTabItem->colUsed;
3410 int n = 0;
drh74161702006-02-24 02:53:49 +00003411 for(; b; b=b>>1, n++){}
shanec0688ea2009-03-05 03:48:06 +00003412 sqlite3VdbeChangeP4(v, sqlite3VdbeCurrentAddr(v)-1, SQLITE_INT_TO_PTR(n), P4_INT32);
danielk19779792eef2006-01-13 15:58:43 +00003413 assert( n<=pTab->nCol );
3414 }
danielk1977c00da102006-01-07 13:21:04 +00003415 }else{
3416 sqlite3TableLock(pParse, iDb, pTab->tnum, 0, pTab->zName);
drh9012bcb2004-12-19 00:11:35 +00003417 }
3418 pLevel->iTabCur = pTabItem->iCursor;
drh111a6a72008-12-21 03:51:16 +00003419 if( (pLevel->plan.wsFlags & WHERE_INDEXED)!=0 ){
3420 Index *pIx = pLevel->plan.u.pIdx;
danielk1977b3bf5562006-01-10 17:58:23 +00003421 KeyInfo *pKey = sqlite3IndexKeyinfo(pParse, pIx);
drh111a6a72008-12-21 03:51:16 +00003422 int iIdxCur = pLevel->iIdxCur;
danielk1977da184232006-01-05 11:34:32 +00003423 assert( pIx->pSchema==pTab->pSchema );
drh111a6a72008-12-21 03:51:16 +00003424 assert( iIdxCur>=0 );
danielk1977207872a2008-01-03 07:54:23 +00003425 sqlite3VdbeAddOp4(v, OP_OpenRead, iIdxCur, pIx->tnum, iDb,
drh66a51672008-01-03 00:01:23 +00003426 (char*)pKey, P4_KEYINFO_HANDOFF);
danielk1977207872a2008-01-03 07:54:23 +00003427 VdbeComment((v, "%s", pIx->zName));
drh9012bcb2004-12-19 00:11:35 +00003428 }
danielk1977da184232006-01-05 11:34:32 +00003429 sqlite3CodeVerifySchema(pParse, iDb);
drh9012bcb2004-12-19 00:11:35 +00003430 }
3431 pWInfo->iTop = sqlite3VdbeCurrentAddr(v);
3432
drh29dda4a2005-07-21 18:23:20 +00003433 /* Generate the code to do the search. Each iteration of the for
3434 ** loop below generates code for a single nested loop of the VM
3435 ** program.
drh75897232000-05-29 14:26:00 +00003436 */
drhfe05af82005-07-21 03:14:59 +00003437 notReady = ~(Bitmask)0;
drh111a6a72008-12-21 03:51:16 +00003438 for(i=0; i<pTabList->nSrc; i++){
3439 notReady = codeOneLoopStart(pWInfo, i, wctrlFlags, notReady);
drh813f31e2009-01-06 00:08:02 +00003440 pWInfo->iContinue = pWInfo->a[i].addrCont;
drh75897232000-05-29 14:26:00 +00003441 }
drh7ec764a2005-07-21 03:48:20 +00003442
3443#ifdef SQLITE_TEST /* For testing and debugging use only */
3444 /* Record in the query plan information about the current table
3445 ** and the index used to access it (if any). If the table itself
3446 ** is not used, its name is just '{}'. If no index is used
3447 ** the index is listed as "{}". If the primary key is used the
3448 ** index name is '*'.
3449 */
3450 for(i=0; i<pTabList->nSrc; i++){
3451 char *z;
3452 int n;
drh7ec764a2005-07-21 03:48:20 +00003453 pLevel = &pWInfo->a[i];
drh29dda4a2005-07-21 18:23:20 +00003454 pTabItem = &pTabList->a[pLevel->iFrom];
drh7ec764a2005-07-21 03:48:20 +00003455 z = pTabItem->zAlias;
3456 if( z==0 ) z = pTabItem->pTab->zName;
drhea678832008-12-10 19:26:22 +00003457 n = sqlite3Strlen30(z);
drh7ec764a2005-07-21 03:48:20 +00003458 if( n+nQPlan < sizeof(sqlite3_query_plan)-10 ){
drh111a6a72008-12-21 03:51:16 +00003459 if( pLevel->plan.wsFlags & WHERE_IDX_ONLY ){
drh5bb3eb92007-05-04 13:15:55 +00003460 memcpy(&sqlite3_query_plan[nQPlan], "{}", 2);
drh7ec764a2005-07-21 03:48:20 +00003461 nQPlan += 2;
3462 }else{
drh5bb3eb92007-05-04 13:15:55 +00003463 memcpy(&sqlite3_query_plan[nQPlan], z, n);
drh7ec764a2005-07-21 03:48:20 +00003464 nQPlan += n;
3465 }
3466 sqlite3_query_plan[nQPlan++] = ' ';
3467 }
drh111a6a72008-12-21 03:51:16 +00003468 testcase( pLevel->plan.wsFlags & WHERE_ROWID_EQ );
3469 testcase( pLevel->plan.wsFlags & WHERE_ROWID_RANGE );
3470 if( pLevel->plan.wsFlags & (WHERE_ROWID_EQ|WHERE_ROWID_RANGE) ){
drh5bb3eb92007-05-04 13:15:55 +00003471 memcpy(&sqlite3_query_plan[nQPlan], "* ", 2);
drh7ec764a2005-07-21 03:48:20 +00003472 nQPlan += 2;
drh111a6a72008-12-21 03:51:16 +00003473 }else if( (pLevel->plan.wsFlags & WHERE_INDEXED)!=0 ){
3474 n = sqlite3Strlen30(pLevel->plan.u.pIdx->zName);
drh7ec764a2005-07-21 03:48:20 +00003475 if( n+nQPlan < sizeof(sqlite3_query_plan)-2 ){
drh111a6a72008-12-21 03:51:16 +00003476 memcpy(&sqlite3_query_plan[nQPlan], pLevel->plan.u.pIdx->zName, n);
drh7ec764a2005-07-21 03:48:20 +00003477 nQPlan += n;
3478 sqlite3_query_plan[nQPlan++] = ' ';
3479 }
drh111a6a72008-12-21 03:51:16 +00003480 }else{
3481 memcpy(&sqlite3_query_plan[nQPlan], "{} ", 3);
3482 nQPlan += 3;
drh7ec764a2005-07-21 03:48:20 +00003483 }
3484 }
3485 while( nQPlan>0 && sqlite3_query_plan[nQPlan-1]==' ' ){
3486 sqlite3_query_plan[--nQPlan] = 0;
3487 }
3488 sqlite3_query_plan[nQPlan] = 0;
3489 nQPlan = 0;
3490#endif /* SQLITE_TEST // Testing and debugging use only */
3491
drh29dda4a2005-07-21 18:23:20 +00003492 /* Record the continuation address in the WhereInfo structure. Then
3493 ** clean up and return.
3494 */
drh75897232000-05-29 14:26:00 +00003495 return pWInfo;
drhe23399f2005-07-22 00:31:39 +00003496
3497 /* Jump here if malloc fails */
danielk197785574e32008-10-06 05:32:18 +00003498whereBeginError:
drh10fe8402008-10-11 16:47:35 +00003499 whereInfoFree(db, pWInfo);
drhe23399f2005-07-22 00:31:39 +00003500 return 0;
drh75897232000-05-29 14:26:00 +00003501}
3502
3503/*
drhc27a1ce2002-06-14 20:58:45 +00003504** Generate the end of the WHERE loop. See comments on
danielk19774adee202004-05-08 08:23:19 +00003505** sqlite3WhereBegin() for additional information.
drh75897232000-05-29 14:26:00 +00003506*/
danielk19774adee202004-05-08 08:23:19 +00003507void sqlite3WhereEnd(WhereInfo *pWInfo){
drh633e6d52008-07-28 19:34:53 +00003508 Parse *pParse = pWInfo->pParse;
3509 Vdbe *v = pParse->pVdbe;
drh19a775c2000-06-05 18:54:46 +00003510 int i;
drh6b563442001-11-07 16:48:26 +00003511 WhereLevel *pLevel;
drhad3cab52002-05-24 02:04:32 +00003512 SrcList *pTabList = pWInfo->pTabList;
drh633e6d52008-07-28 19:34:53 +00003513 sqlite3 *db = pParse->db;
drh19a775c2000-06-05 18:54:46 +00003514
drh9012bcb2004-12-19 00:11:35 +00003515 /* Generate loop termination code.
3516 */
drh633e6d52008-07-28 19:34:53 +00003517 sqlite3ExprClearColumnCache(pParse, -1);
drhad3cab52002-05-24 02:04:32 +00003518 for(i=pTabList->nSrc-1; i>=0; i--){
drh6b563442001-11-07 16:48:26 +00003519 pLevel = &pWInfo->a[i];
drhb3190c12008-12-08 21:37:14 +00003520 sqlite3VdbeResolveLabel(v, pLevel->addrCont);
drh6b563442001-11-07 16:48:26 +00003521 if( pLevel->op!=OP_Noop ){
drh66a51672008-01-03 00:01:23 +00003522 sqlite3VdbeAddOp2(v, pLevel->op, pLevel->p1, pLevel->p2);
drhd1d38482008-10-07 23:46:38 +00003523 sqlite3VdbeChangeP5(v, pLevel->p5);
drh19a775c2000-06-05 18:54:46 +00003524 }
drh111a6a72008-12-21 03:51:16 +00003525 if( pLevel->plan.wsFlags & WHERE_IN_ABLE && pLevel->u.in.nIn>0 ){
drh72e8fa42007-03-28 14:30:06 +00003526 struct InLoop *pIn;
drhe23399f2005-07-22 00:31:39 +00003527 int j;
drhb3190c12008-12-08 21:37:14 +00003528 sqlite3VdbeResolveLabel(v, pLevel->addrNxt);
drh111a6a72008-12-21 03:51:16 +00003529 for(j=pLevel->u.in.nIn, pIn=&pLevel->u.in.aInLoop[j-1]; j>0; j--, pIn--){
drhb3190c12008-12-08 21:37:14 +00003530 sqlite3VdbeJumpHere(v, pIn->addrInTop+1);
3531 sqlite3VdbeAddOp2(v, OP_Next, pIn->iCur, pIn->addrInTop);
3532 sqlite3VdbeJumpHere(v, pIn->addrInTop-1);
drhe23399f2005-07-22 00:31:39 +00003533 }
drh111a6a72008-12-21 03:51:16 +00003534 sqlite3DbFree(db, pLevel->u.in.aInLoop);
drhd99f7062002-06-08 23:25:08 +00003535 }
drhb3190c12008-12-08 21:37:14 +00003536 sqlite3VdbeResolveLabel(v, pLevel->addrBrk);
drhad2d8302002-05-24 20:31:36 +00003537 if( pLevel->iLeftJoin ){
3538 int addr;
drh3c84ddf2008-01-09 02:15:38 +00003539 addr = sqlite3VdbeAddOp1(v, OP_IfPos, pLevel->iLeftJoin);
3540 sqlite3VdbeAddOp1(v, OP_NullRow, pTabList->a[i].iCursor);
drh9012bcb2004-12-19 00:11:35 +00003541 if( pLevel->iIdxCur>=0 ){
drh3c84ddf2008-01-09 02:15:38 +00003542 sqlite3VdbeAddOp1(v, OP_NullRow, pLevel->iIdxCur);
drh7f09b3e2002-08-13 13:15:49 +00003543 }
drhb3190c12008-12-08 21:37:14 +00003544 sqlite3VdbeAddOp2(v, OP_Goto, 0, pLevel->addrFirst);
drhd654be82005-09-20 17:42:23 +00003545 sqlite3VdbeJumpHere(v, addr);
drhad2d8302002-05-24 20:31:36 +00003546 }
drh19a775c2000-06-05 18:54:46 +00003547 }
drh9012bcb2004-12-19 00:11:35 +00003548
3549 /* The "break" point is here, just past the end of the outer loop.
3550 ** Set it.
3551 */
danielk19774adee202004-05-08 08:23:19 +00003552 sqlite3VdbeResolveLabel(v, pWInfo->iBreak);
drh9012bcb2004-12-19 00:11:35 +00003553
drh29dda4a2005-07-21 18:23:20 +00003554 /* Close all of the cursors that were opened by sqlite3WhereBegin.
drh9012bcb2004-12-19 00:11:35 +00003555 */
drh29dda4a2005-07-21 18:23:20 +00003556 for(i=0, pLevel=pWInfo->a; i<pTabList->nSrc; i++, pLevel++){
3557 struct SrcList_item *pTabItem = &pTabList->a[pLevel->iFrom];
drh9012bcb2004-12-19 00:11:35 +00003558 Table *pTab = pTabItem->pTab;
drh5cf590c2003-04-24 01:45:04 +00003559 assert( pTab!=0 );
drh7d10d5a2008-08-20 16:35:10 +00003560 if( (pTab->tabFlags & TF_Ephemeral)!=0 || pTab->pSelect ) continue;
drh6df2acd2008-12-28 16:55:25 +00003561 if( (pWInfo->wctrlFlags & WHERE_OMIT_CLOSE)==0 ){
3562 if( !pWInfo->okOnePass && (pLevel->plan.wsFlags & WHERE_IDX_ONLY)==0 ){
3563 sqlite3VdbeAddOp1(v, OP_Close, pTabItem->iCursor);
3564 }
3565 if( (pLevel->plan.wsFlags & WHERE_INDEXED)!=0 ){
3566 sqlite3VdbeAddOp1(v, OP_Close, pLevel->iIdxCur);
3567 }
drh9012bcb2004-12-19 00:11:35 +00003568 }
3569
danielk197721de2e72007-11-29 17:43:27 +00003570 /* If this scan uses an index, make code substitutions to read data
3571 ** from the index in preference to the table. Sometimes, this means
3572 ** the table need never be read from. This is a performance boost,
3573 ** as the vdbe level waits until the table is read before actually
3574 ** seeking the table cursor to the record corresponding to the current
3575 ** position in the index.
drh9012bcb2004-12-19 00:11:35 +00003576 **
3577 ** Calls to the code generator in between sqlite3WhereBegin and
3578 ** sqlite3WhereEnd will have created code that references the table
3579 ** directly. This loop scans all that code looking for opcodes
3580 ** that reference the table and converts them into opcodes that
3581 ** reference the index.
3582 */
drh111a6a72008-12-21 03:51:16 +00003583 if( (pLevel->plan.wsFlags & WHERE_INDEXED)!=0 ){
danielk1977f0113002006-01-24 12:09:17 +00003584 int k, j, last;
drh9012bcb2004-12-19 00:11:35 +00003585 VdbeOp *pOp;
drh111a6a72008-12-21 03:51:16 +00003586 Index *pIdx = pLevel->plan.u.pIdx;
3587 int useIndexOnly = pLevel->plan.wsFlags & WHERE_IDX_ONLY;
drh9012bcb2004-12-19 00:11:35 +00003588
3589 assert( pIdx!=0 );
3590 pOp = sqlite3VdbeGetOp(v, pWInfo->iTop);
3591 last = sqlite3VdbeCurrentAddr(v);
danielk1977f0113002006-01-24 12:09:17 +00003592 for(k=pWInfo->iTop; k<last; k++, pOp++){
drh9012bcb2004-12-19 00:11:35 +00003593 if( pOp->p1!=pLevel->iTabCur ) continue;
3594 if( pOp->opcode==OP_Column ){
drh9012bcb2004-12-19 00:11:35 +00003595 for(j=0; j<pIdx->nColumn; j++){
3596 if( pOp->p2==pIdx->aiColumn[j] ){
3597 pOp->p2 = j;
danielk197721de2e72007-11-29 17:43:27 +00003598 pOp->p1 = pLevel->iIdxCur;
drh9012bcb2004-12-19 00:11:35 +00003599 break;
3600 }
3601 }
danielk197721de2e72007-11-29 17:43:27 +00003602 assert(!useIndexOnly || j<pIdx->nColumn);
drhf0863fe2005-06-12 21:35:51 +00003603 }else if( pOp->opcode==OP_Rowid ){
drh9012bcb2004-12-19 00:11:35 +00003604 pOp->p1 = pLevel->iIdxCur;
drhf0863fe2005-06-12 21:35:51 +00003605 pOp->opcode = OP_IdxRowid;
danielk197721de2e72007-11-29 17:43:27 +00003606 }else if( pOp->opcode==OP_NullRow && useIndexOnly ){
danielk19776c18b6e2005-01-30 09:17:58 +00003607 pOp->opcode = OP_Noop;
drh9012bcb2004-12-19 00:11:35 +00003608 }
3609 }
drh6b563442001-11-07 16:48:26 +00003610 }
drh19a775c2000-06-05 18:54:46 +00003611 }
drh9012bcb2004-12-19 00:11:35 +00003612
3613 /* Final cleanup
3614 */
drh10fe8402008-10-11 16:47:35 +00003615 whereInfoFree(db, pWInfo);
drh75897232000-05-29 14:26:00 +00003616 return;
3617}