blob: 15a17b7a7f57bcaa07b0dec844f0da5c2739d469 [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**
danielk1977f51d1bd2009-07-31 06:14:51 +000019** $Id: where.c,v 1.411 2009/07/31 06:14:52 danielk1977 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 */
danielk1977e672c8e2009-05-22 15:43:26 +0000130 Bitmask vmask; /* Bitmask identifying virtual table cursors */
drh29435252008-12-28 18:35:08 +0000131 u8 op; /* Split operator. TK_AND or TK_OR */
drh0aa74ed2005-07-16 13:33:20 +0000132 int nTerm; /* Number of terms */
133 int nSlot; /* Number of entries in a[] */
drh51147ba2005-07-23 22:59:55 +0000134 WhereTerm *a; /* Each a[] describes a term of the WHERE cluase */
drh50d654d2009-06-03 01:24:54 +0000135#if defined(SQLITE_SMALL_STACK)
136 WhereTerm aStatic[1]; /* Initial static space for a[] */
137#else
138 WhereTerm aStatic[8]; /* Initial static space for a[] */
139#endif
drhe23399f2005-07-22 00:31:39 +0000140};
141
142/*
drh700a2262008-12-17 19:22:15 +0000143** A WhereTerm with eOperator==WO_OR has its u.pOrInfo pointer set to
144** a dynamically allocated instance of the following structure.
145*/
146struct WhereOrInfo {
drh111a6a72008-12-21 03:51:16 +0000147 WhereClause wc; /* Decomposition into subterms */
drh1a58fe02008-12-20 02:06:13 +0000148 Bitmask indexable; /* Bitmask of all indexable tables in the clause */
drh700a2262008-12-17 19:22:15 +0000149};
150
151/*
152** A WhereTerm with eOperator==WO_AND has its u.pAndInfo pointer set to
153** a dynamically allocated instance of the following structure.
154*/
155struct WhereAndInfo {
drh29435252008-12-28 18:35:08 +0000156 WhereClause wc; /* The subexpression broken out */
drh700a2262008-12-17 19:22:15 +0000157};
158
159/*
drh6a3ea0e2003-05-02 14:32:12 +0000160** An instance of the following structure keeps track of a mapping
drh0aa74ed2005-07-16 13:33:20 +0000161** between VDBE cursor numbers and bits of the bitmasks in WhereTerm.
drh51669862004-12-18 18:40:26 +0000162**
163** The VDBE cursor numbers are small integers contained in
164** SrcList_item.iCursor and Expr.iTable fields. For any given WHERE
165** clause, the cursor numbers might not begin with 0 and they might
166** contain gaps in the numbering sequence. But we want to make maximum
167** use of the bits in our bitmasks. This structure provides a mapping
168** from the sparse cursor numbers into consecutive integers beginning
169** with 0.
170**
drh111a6a72008-12-21 03:51:16 +0000171** If WhereMaskSet.ix[A]==B it means that The A-th bit of a Bitmask
drh51669862004-12-18 18:40:26 +0000172** corresponds VDBE cursor number B. The A-th bit of a bitmask is 1<<A.
173**
174** For example, if the WHERE clause expression used these VDBE
drh111a6a72008-12-21 03:51:16 +0000175** cursors: 4, 5, 8, 29, 57, 73. Then the WhereMaskSet structure
drh51669862004-12-18 18:40:26 +0000176** would map those cursor numbers into bits 0 through 5.
177**
178** Note that the mapping is not necessarily ordered. In the example
179** above, the mapping might go like this: 4->3, 5->1, 8->2, 29->0,
180** 57->5, 73->4. Or one of 719 other combinations might be used. It
181** does not really matter. What is important is that sparse cursor
182** numbers all get mapped into bit numbers that begin with 0 and contain
183** no gaps.
drh6a3ea0e2003-05-02 14:32:12 +0000184*/
drh111a6a72008-12-21 03:51:16 +0000185struct WhereMaskSet {
drh1398ad32005-01-19 23:24:50 +0000186 int n; /* Number of assigned cursor values */
danielk197723432972008-11-17 16:42:00 +0000187 int ix[BMS]; /* Cursor assigned to each bit */
drh6a3ea0e2003-05-02 14:32:12 +0000188};
189
drh111a6a72008-12-21 03:51:16 +0000190/*
191** A WhereCost object records a lookup strategy and the estimated
192** cost of pursuing that strategy.
193*/
194struct WhereCost {
195 WherePlan plan; /* The lookup strategy */
196 double rCost; /* Overall cost of pursuing this search strategy */
197 double nRow; /* Estimated number of output rows */
dan5236ac12009-08-13 07:09:33 +0000198 Bitmask used; /* Bitmask of cursors used by this plan */
drh111a6a72008-12-21 03:51:16 +0000199};
drh0aa74ed2005-07-16 13:33:20 +0000200
drh6a3ea0e2003-05-02 14:32:12 +0000201/*
drh51147ba2005-07-23 22:59:55 +0000202** Bitmasks for the operators that indices are able to exploit. An
203** OR-ed combination of these values can be used when searching for
204** terms in the where clause.
205*/
drh165be382008-12-05 02:36:33 +0000206#define WO_IN 0x001
207#define WO_EQ 0x002
drh51147ba2005-07-23 22:59:55 +0000208#define WO_LT (WO_EQ<<(TK_LT-TK_EQ))
209#define WO_LE (WO_EQ<<(TK_LE-TK_EQ))
210#define WO_GT (WO_EQ<<(TK_GT-TK_EQ))
211#define WO_GE (WO_EQ<<(TK_GE-TK_EQ))
drh165be382008-12-05 02:36:33 +0000212#define WO_MATCH 0x040
213#define WO_ISNULL 0x080
drh700a2262008-12-17 19:22:15 +0000214#define WO_OR 0x100 /* Two or more OR-connected terms */
215#define WO_AND 0x200 /* Two or more AND-connected terms */
drh51147ba2005-07-23 22:59:55 +0000216
drhec1724e2008-12-09 01:32:03 +0000217#define WO_ALL 0xfff /* Mask of all possible WO_* values */
drh1a58fe02008-12-20 02:06:13 +0000218#define WO_SINGLE 0x0ff /* Mask of all non-compound WO_* values */
drhec1724e2008-12-09 01:32:03 +0000219
drh51147ba2005-07-23 22:59:55 +0000220/*
drh700a2262008-12-17 19:22:15 +0000221** Value for wsFlags returned by bestIndex() and stored in
222** WhereLevel.wsFlags. These flags determine which search
223** strategies are appropriate.
drhf2d315d2007-01-25 16:56:06 +0000224**
drh165be382008-12-05 02:36:33 +0000225** The least significant 12 bits is reserved as a mask for WO_ values above.
drh700a2262008-12-17 19:22:15 +0000226** The WhereLevel.wsFlags field is usually set to WO_IN|WO_EQ|WO_ISNULL.
227** But if the table is the right table of a left join, WhereLevel.wsFlags
228** is set to WO_IN|WO_EQ. The WhereLevel.wsFlags field can then be used as
drhf2d315d2007-01-25 16:56:06 +0000229** the "op" parameter to findTerm when we are resolving equality constraints.
230** ISNULL constraints will then not be used on the right table of a left
231** join. Tickets #2177 and #2189.
drh51147ba2005-07-23 22:59:55 +0000232*/
drh165be382008-12-05 02:36:33 +0000233#define WHERE_ROWID_EQ 0x00001000 /* rowid=EXPR or rowid IN (...) */
234#define WHERE_ROWID_RANGE 0x00002000 /* rowid<EXPR and/or rowid>EXPR */
drh46619d62009-04-24 14:51:42 +0000235#define WHERE_COLUMN_EQ 0x00010000 /* x=EXPR or x IN (...) or x IS NULL */
drh165be382008-12-05 02:36:33 +0000236#define WHERE_COLUMN_RANGE 0x00020000 /* x<EXPR and/or x>EXPR */
237#define WHERE_COLUMN_IN 0x00040000 /* x IN (...) */
drh46619d62009-04-24 14:51:42 +0000238#define WHERE_COLUMN_NULL 0x00080000 /* x IS NULL */
239#define WHERE_INDEXED 0x000f0000 /* Anything that uses an index */
240#define WHERE_IN_ABLE 0x000f1000 /* Able to support an IN operator */
drh165be382008-12-05 02:36:33 +0000241#define WHERE_TOP_LIMIT 0x00100000 /* x<EXPR or x<=EXPR constraint */
242#define WHERE_BTM_LIMIT 0x00200000 /* x>EXPR or x>=EXPR constraint */
243#define WHERE_IDX_ONLY 0x00800000 /* Use index only - omit table */
244#define WHERE_ORDERBY 0x01000000 /* Output will appear in correct order */
245#define WHERE_REVERSE 0x02000000 /* Scan in reverse order */
246#define WHERE_UNIQUE 0x04000000 /* Selects no more than one row */
247#define WHERE_VIRTUALTABLE 0x08000000 /* Use virtual-table processing */
248#define WHERE_MULTI_OR 0x10000000 /* OR using multiple indices */
drh51147ba2005-07-23 22:59:55 +0000249
250/*
drh0aa74ed2005-07-16 13:33:20 +0000251** Initialize a preallocated WhereClause structure.
drh75897232000-05-29 14:26:00 +0000252*/
drh7b4fc6a2007-02-06 13:26:32 +0000253static void whereClauseInit(
254 WhereClause *pWC, /* The WhereClause to be initialized */
255 Parse *pParse, /* The parsing context */
drh111a6a72008-12-21 03:51:16 +0000256 WhereMaskSet *pMaskSet /* Mapping from table cursor numbers to bitmasks */
drh7b4fc6a2007-02-06 13:26:32 +0000257){
drhfe05af82005-07-21 03:14:59 +0000258 pWC->pParse = pParse;
drh7b4fc6a2007-02-06 13:26:32 +0000259 pWC->pMaskSet = pMaskSet;
drh0aa74ed2005-07-16 13:33:20 +0000260 pWC->nTerm = 0;
drhcad651e2007-04-20 12:22:01 +0000261 pWC->nSlot = ArraySize(pWC->aStatic);
drh0aa74ed2005-07-16 13:33:20 +0000262 pWC->a = pWC->aStatic;
danielk1977e672c8e2009-05-22 15:43:26 +0000263 pWC->vmask = 0;
drh0aa74ed2005-07-16 13:33:20 +0000264}
265
drh700a2262008-12-17 19:22:15 +0000266/* Forward reference */
267static void whereClauseClear(WhereClause*);
268
269/*
270** Deallocate all memory associated with a WhereOrInfo object.
271*/
272static void whereOrInfoDelete(sqlite3 *db, WhereOrInfo *p){
drh5bd98ae2009-01-07 18:24:03 +0000273 whereClauseClear(&p->wc);
274 sqlite3DbFree(db, p);
drh700a2262008-12-17 19:22:15 +0000275}
276
277/*
278** Deallocate all memory associated with a WhereAndInfo object.
279*/
280static void whereAndInfoDelete(sqlite3 *db, WhereAndInfo *p){
drh5bd98ae2009-01-07 18:24:03 +0000281 whereClauseClear(&p->wc);
282 sqlite3DbFree(db, p);
drh700a2262008-12-17 19:22:15 +0000283}
284
drh0aa74ed2005-07-16 13:33:20 +0000285/*
286** Deallocate a WhereClause structure. The WhereClause structure
287** itself is not freed. This routine is the inverse of whereClauseInit().
288*/
289static void whereClauseClear(WhereClause *pWC){
290 int i;
291 WhereTerm *a;
drh633e6d52008-07-28 19:34:53 +0000292 sqlite3 *db = pWC->pParse->db;
drh0aa74ed2005-07-16 13:33:20 +0000293 for(i=pWC->nTerm-1, a=pWC->a; i>=0; i--, a++){
drh165be382008-12-05 02:36:33 +0000294 if( a->wtFlags & TERM_DYNAMIC ){
drh633e6d52008-07-28 19:34:53 +0000295 sqlite3ExprDelete(db, a->pExpr);
drh0aa74ed2005-07-16 13:33:20 +0000296 }
drh700a2262008-12-17 19:22:15 +0000297 if( a->wtFlags & TERM_ORINFO ){
298 whereOrInfoDelete(db, a->u.pOrInfo);
299 }else if( a->wtFlags & TERM_ANDINFO ){
300 whereAndInfoDelete(db, a->u.pAndInfo);
301 }
drh0aa74ed2005-07-16 13:33:20 +0000302 }
303 if( pWC->a!=pWC->aStatic ){
drh633e6d52008-07-28 19:34:53 +0000304 sqlite3DbFree(db, pWC->a);
drh0aa74ed2005-07-16 13:33:20 +0000305 }
306}
307
308/*
drh6a1e0712008-12-05 15:24:15 +0000309** Add a single new WhereTerm entry to the WhereClause object pWC.
310** The new WhereTerm object is constructed from Expr p and with wtFlags.
311** The index in pWC->a[] of the new WhereTerm is returned on success.
312** 0 is returned if the new WhereTerm could not be added due to a memory
313** allocation error. The memory allocation failure will be recorded in
314** the db->mallocFailed flag so that higher-level functions can detect it.
315**
316** This routine will increase the size of the pWC->a[] array as necessary.
drh9eb20282005-08-24 03:52:18 +0000317**
drh165be382008-12-05 02:36:33 +0000318** If the wtFlags argument includes TERM_DYNAMIC, then responsibility
drh6a1e0712008-12-05 15:24:15 +0000319** for freeing the expression p is assumed by the WhereClause object pWC.
320** This is true even if this routine fails to allocate a new WhereTerm.
drhb63a53d2007-03-31 01:34:44 +0000321**
drh9eb20282005-08-24 03:52:18 +0000322** WARNING: This routine might reallocate the space used to store
drh909626d2008-05-30 14:58:37 +0000323** WhereTerms. All pointers to WhereTerms should be invalidated after
drh9eb20282005-08-24 03:52:18 +0000324** calling this routine. Such pointers may be reinitialized by referencing
325** the pWC->a[] array.
drh0aa74ed2005-07-16 13:33:20 +0000326*/
drhec1724e2008-12-09 01:32:03 +0000327static int whereClauseInsert(WhereClause *pWC, Expr *p, u8 wtFlags){
drh0aa74ed2005-07-16 13:33:20 +0000328 WhereTerm *pTerm;
drh9eb20282005-08-24 03:52:18 +0000329 int idx;
drh0aa74ed2005-07-16 13:33:20 +0000330 if( pWC->nTerm>=pWC->nSlot ){
331 WhereTerm *pOld = pWC->a;
drh633e6d52008-07-28 19:34:53 +0000332 sqlite3 *db = pWC->pParse->db;
333 pWC->a = sqlite3DbMallocRaw(db, sizeof(pWC->a[0])*pWC->nSlot*2 );
drhb63a53d2007-03-31 01:34:44 +0000334 if( pWC->a==0 ){
drh165be382008-12-05 02:36:33 +0000335 if( wtFlags & TERM_DYNAMIC ){
drh633e6d52008-07-28 19:34:53 +0000336 sqlite3ExprDelete(db, p);
drhb63a53d2007-03-31 01:34:44 +0000337 }
drhf998b732007-11-26 13:36:00 +0000338 pWC->a = pOld;
drhb63a53d2007-03-31 01:34:44 +0000339 return 0;
340 }
drh0aa74ed2005-07-16 13:33:20 +0000341 memcpy(pWC->a, pOld, sizeof(pWC->a[0])*pWC->nTerm);
342 if( pOld!=pWC->aStatic ){
drh633e6d52008-07-28 19:34:53 +0000343 sqlite3DbFree(db, pOld);
drh0aa74ed2005-07-16 13:33:20 +0000344 }
drh6a1e0712008-12-05 15:24:15 +0000345 pWC->nSlot = sqlite3DbMallocSize(db, pWC->a)/sizeof(pWC->a[0]);
drh0aa74ed2005-07-16 13:33:20 +0000346 }
drh6a1e0712008-12-05 15:24:15 +0000347 pTerm = &pWC->a[idx = pWC->nTerm++];
drh0fcef5e2005-07-19 17:38:22 +0000348 pTerm->pExpr = p;
drh165be382008-12-05 02:36:33 +0000349 pTerm->wtFlags = wtFlags;
drh0fcef5e2005-07-19 17:38:22 +0000350 pTerm->pWC = pWC;
drh45b1ee42005-08-02 17:48:22 +0000351 pTerm->iParent = -1;
drh9eb20282005-08-24 03:52:18 +0000352 return idx;
drh0aa74ed2005-07-16 13:33:20 +0000353}
drh75897232000-05-29 14:26:00 +0000354
355/*
drh51669862004-12-18 18:40:26 +0000356** This routine identifies subexpressions in the WHERE clause where
drhb6fb62d2005-09-20 08:47:20 +0000357** each subexpression is separated by the AND operator or some other
drh6c30be82005-07-29 15:10:17 +0000358** operator specified in the op parameter. The WhereClause structure
359** is filled with pointers to subexpressions. For example:
drh75897232000-05-29 14:26:00 +0000360**
drh51669862004-12-18 18:40:26 +0000361** WHERE a=='hello' AND coalesce(b,11)<10 AND (c+12!=d OR c==22)
362** \________/ \_______________/ \________________/
363** slot[0] slot[1] slot[2]
364**
365** The original WHERE clause in pExpr is unaltered. All this routine
drh51147ba2005-07-23 22:59:55 +0000366** does is make slot[] entries point to substructure within pExpr.
drh51669862004-12-18 18:40:26 +0000367**
drh51147ba2005-07-23 22:59:55 +0000368** In the previous sentence and in the diagram, "slot[]" refers to
drh902b9ee2008-12-05 17:17:07 +0000369** the WhereClause.a[] array. The slot[] array grows as needed to contain
drh51147ba2005-07-23 22:59:55 +0000370** all terms of the WHERE clause.
drh75897232000-05-29 14:26:00 +0000371*/
drh6c30be82005-07-29 15:10:17 +0000372static void whereSplit(WhereClause *pWC, Expr *pExpr, int op){
drh29435252008-12-28 18:35:08 +0000373 pWC->op = (u8)op;
drh0aa74ed2005-07-16 13:33:20 +0000374 if( pExpr==0 ) return;
drh6c30be82005-07-29 15:10:17 +0000375 if( pExpr->op!=op ){
drh0aa74ed2005-07-16 13:33:20 +0000376 whereClauseInsert(pWC, pExpr, 0);
drh75897232000-05-29 14:26:00 +0000377 }else{
drh6c30be82005-07-29 15:10:17 +0000378 whereSplit(pWC, pExpr->pLeft, op);
379 whereSplit(pWC, pExpr->pRight, op);
drh75897232000-05-29 14:26:00 +0000380 }
drh75897232000-05-29 14:26:00 +0000381}
382
383/*
drh61495262009-04-22 15:32:59 +0000384** Initialize an expression mask set (a WhereMaskSet object)
drh6a3ea0e2003-05-02 14:32:12 +0000385*/
386#define initMaskSet(P) memset(P, 0, sizeof(*P))
387
388/*
drh1398ad32005-01-19 23:24:50 +0000389** Return the bitmask for the given cursor number. Return 0 if
390** iCursor is not in the set.
drh6a3ea0e2003-05-02 14:32:12 +0000391*/
drh111a6a72008-12-21 03:51:16 +0000392static Bitmask getMask(WhereMaskSet *pMaskSet, int iCursor){
drh6a3ea0e2003-05-02 14:32:12 +0000393 int i;
drh3500ed62009-05-05 15:46:43 +0000394 assert( pMaskSet->n<=sizeof(Bitmask)*8 );
drh6a3ea0e2003-05-02 14:32:12 +0000395 for(i=0; i<pMaskSet->n; i++){
drh51669862004-12-18 18:40:26 +0000396 if( pMaskSet->ix[i]==iCursor ){
397 return ((Bitmask)1)<<i;
398 }
drh6a3ea0e2003-05-02 14:32:12 +0000399 }
drh6a3ea0e2003-05-02 14:32:12 +0000400 return 0;
401}
402
403/*
drh1398ad32005-01-19 23:24:50 +0000404** Create a new mask for cursor iCursor.
drh0fcef5e2005-07-19 17:38:22 +0000405**
406** There is one cursor per table in the FROM clause. The number of
407** tables in the FROM clause is limited by a test early in the
drhb6fb62d2005-09-20 08:47:20 +0000408** sqlite3WhereBegin() routine. So we know that the pMaskSet->ix[]
drh0fcef5e2005-07-19 17:38:22 +0000409** array will never overflow.
drh1398ad32005-01-19 23:24:50 +0000410*/
drh111a6a72008-12-21 03:51:16 +0000411static void createMask(WhereMaskSet *pMaskSet, int iCursor){
drhcad651e2007-04-20 12:22:01 +0000412 assert( pMaskSet->n < ArraySize(pMaskSet->ix) );
drh0fcef5e2005-07-19 17:38:22 +0000413 pMaskSet->ix[pMaskSet->n++] = iCursor;
drh1398ad32005-01-19 23:24:50 +0000414}
415
416/*
drh75897232000-05-29 14:26:00 +0000417** This routine walks (recursively) an expression tree and generates
418** a bitmask indicating which tables are used in that expression
drh6a3ea0e2003-05-02 14:32:12 +0000419** tree.
drh75897232000-05-29 14:26:00 +0000420**
421** In order for this routine to work, the calling function must have
drh7d10d5a2008-08-20 16:35:10 +0000422** previously invoked sqlite3ResolveExprNames() on the expression. See
drh75897232000-05-29 14:26:00 +0000423** the header comment on that routine for additional information.
drh7d10d5a2008-08-20 16:35:10 +0000424** The sqlite3ResolveExprNames() routines looks for column names and
drh6a3ea0e2003-05-02 14:32:12 +0000425** sets their opcodes to TK_COLUMN and their Expr.iTable fields to
drh51147ba2005-07-23 22:59:55 +0000426** the VDBE cursor number of the table. This routine just has to
427** translate the cursor numbers into bitmask values and OR all
428** the bitmasks together.
drh75897232000-05-29 14:26:00 +0000429*/
drh111a6a72008-12-21 03:51:16 +0000430static Bitmask exprListTableUsage(WhereMaskSet*, ExprList*);
431static Bitmask exprSelectTableUsage(WhereMaskSet*, Select*);
432static Bitmask exprTableUsage(WhereMaskSet *pMaskSet, Expr *p){
drh51669862004-12-18 18:40:26 +0000433 Bitmask mask = 0;
drh75897232000-05-29 14:26:00 +0000434 if( p==0 ) return 0;
drh967e8b72000-06-21 13:59:10 +0000435 if( p->op==TK_COLUMN ){
drh8feb4b12004-07-19 02:12:14 +0000436 mask = getMask(pMaskSet, p->iTable);
drh8feb4b12004-07-19 02:12:14 +0000437 return mask;
drh75897232000-05-29 14:26:00 +0000438 }
danielk1977b3bce662005-01-29 08:32:43 +0000439 mask = exprTableUsage(pMaskSet, p->pRight);
440 mask |= exprTableUsage(pMaskSet, p->pLeft);
danielk19776ab3a2e2009-02-19 14:39:25 +0000441 if( ExprHasProperty(p, EP_xIsSelect) ){
442 mask |= exprSelectTableUsage(pMaskSet, p->x.pSelect);
443 }else{
444 mask |= exprListTableUsage(pMaskSet, p->x.pList);
445 }
danielk1977b3bce662005-01-29 08:32:43 +0000446 return mask;
447}
drh111a6a72008-12-21 03:51:16 +0000448static Bitmask exprListTableUsage(WhereMaskSet *pMaskSet, ExprList *pList){
danielk1977b3bce662005-01-29 08:32:43 +0000449 int i;
450 Bitmask mask = 0;
451 if( pList ){
452 for(i=0; i<pList->nExpr; i++){
453 mask |= exprTableUsage(pMaskSet, pList->a[i].pExpr);
drhdd579122002-04-02 01:58:57 +0000454 }
455 }
drh75897232000-05-29 14:26:00 +0000456 return mask;
457}
drh111a6a72008-12-21 03:51:16 +0000458static Bitmask exprSelectTableUsage(WhereMaskSet *pMaskSet, Select *pS){
drha430ae82007-09-12 15:41:01 +0000459 Bitmask mask = 0;
460 while( pS ){
461 mask |= exprListTableUsage(pMaskSet, pS->pEList);
drhf5b11382005-09-17 13:07:13 +0000462 mask |= exprListTableUsage(pMaskSet, pS->pGroupBy);
463 mask |= exprListTableUsage(pMaskSet, pS->pOrderBy);
464 mask |= exprTableUsage(pMaskSet, pS->pWhere);
465 mask |= exprTableUsage(pMaskSet, pS->pHaving);
drha430ae82007-09-12 15:41:01 +0000466 pS = pS->pPrior;
drhf5b11382005-09-17 13:07:13 +0000467 }
468 return mask;
469}
drh75897232000-05-29 14:26:00 +0000470
471/*
drh487ab3c2001-11-08 00:45:21 +0000472** Return TRUE if the given operator is one of the operators that is
drh51669862004-12-18 18:40:26 +0000473** allowed for an indexable WHERE clause term. The allowed operators are
drhc27a1ce2002-06-14 20:58:45 +0000474** "=", "<", ">", "<=", ">=", and "IN".
drh487ab3c2001-11-08 00:45:21 +0000475*/
476static int allowedOp(int op){
drhfe05af82005-07-21 03:14:59 +0000477 assert( TK_GT>TK_EQ && TK_GT<TK_GE );
478 assert( TK_LT>TK_EQ && TK_LT<TK_GE );
479 assert( TK_LE>TK_EQ && TK_LE<TK_GE );
480 assert( TK_GE==TK_EQ+4 );
drh50b39962006-10-28 00:28:09 +0000481 return op==TK_IN || (op>=TK_EQ && op<=TK_GE) || op==TK_ISNULL;
drh487ab3c2001-11-08 00:45:21 +0000482}
483
484/*
drh902b9ee2008-12-05 17:17:07 +0000485** Swap two objects of type TYPE.
drh193bd772004-07-20 18:23:14 +0000486*/
487#define SWAP(TYPE,A,B) {TYPE t=A; A=B; B=t;}
488
489/*
drh909626d2008-05-30 14:58:37 +0000490** Commute a comparison operator. Expressions of the form "X op Y"
drh0fcef5e2005-07-19 17:38:22 +0000491** are converted into "Y op X".
danielk1977eb5453d2007-07-30 14:40:48 +0000492**
493** If a collation sequence is associated with either the left or right
494** side of the comparison, it remains associated with the same side after
495** the commutation. So "Y collate NOCASE op X" becomes
496** "X collate NOCASE op Y". This is because any collation sequence on
497** the left hand side of a comparison overrides any collation sequence
498** attached to the right. For the same reason the EP_ExpCollate flag
499** is not commuted.
drh193bd772004-07-20 18:23:14 +0000500*/
drh7d10d5a2008-08-20 16:35:10 +0000501static void exprCommute(Parse *pParse, Expr *pExpr){
danielk1977eb5453d2007-07-30 14:40:48 +0000502 u16 expRight = (pExpr->pRight->flags & EP_ExpCollate);
503 u16 expLeft = (pExpr->pLeft->flags & EP_ExpCollate);
drhfe05af82005-07-21 03:14:59 +0000504 assert( allowedOp(pExpr->op) && pExpr->op!=TK_IN );
drh7d10d5a2008-08-20 16:35:10 +0000505 pExpr->pRight->pColl = sqlite3ExprCollSeq(pParse, pExpr->pRight);
506 pExpr->pLeft->pColl = sqlite3ExprCollSeq(pParse, pExpr->pLeft);
drh0fcef5e2005-07-19 17:38:22 +0000507 SWAP(CollSeq*,pExpr->pRight->pColl,pExpr->pLeft->pColl);
danielk1977eb5453d2007-07-30 14:40:48 +0000508 pExpr->pRight->flags = (pExpr->pRight->flags & ~EP_ExpCollate) | expLeft;
509 pExpr->pLeft->flags = (pExpr->pLeft->flags & ~EP_ExpCollate) | expRight;
drh0fcef5e2005-07-19 17:38:22 +0000510 SWAP(Expr*,pExpr->pRight,pExpr->pLeft);
511 if( pExpr->op>=TK_GT ){
512 assert( TK_LT==TK_GT+2 );
513 assert( TK_GE==TK_LE+2 );
514 assert( TK_GT>TK_EQ );
515 assert( TK_GT<TK_LE );
516 assert( pExpr->op>=TK_GT && pExpr->op<=TK_GE );
517 pExpr->op = ((pExpr->op-TK_GT)^2)+TK_GT;
drh193bd772004-07-20 18:23:14 +0000518 }
drh193bd772004-07-20 18:23:14 +0000519}
520
521/*
drhfe05af82005-07-21 03:14:59 +0000522** Translate from TK_xx operator to WO_xx bitmask.
523*/
drhec1724e2008-12-09 01:32:03 +0000524static u16 operatorMask(int op){
525 u16 c;
drhfe05af82005-07-21 03:14:59 +0000526 assert( allowedOp(op) );
527 if( op==TK_IN ){
drh51147ba2005-07-23 22:59:55 +0000528 c = WO_IN;
drh50b39962006-10-28 00:28:09 +0000529 }else if( op==TK_ISNULL ){
530 c = WO_ISNULL;
drhfe05af82005-07-21 03:14:59 +0000531 }else{
drhec1724e2008-12-09 01:32:03 +0000532 assert( (WO_EQ<<(op-TK_EQ)) < 0x7fff );
533 c = (u16)(WO_EQ<<(op-TK_EQ));
drhfe05af82005-07-21 03:14:59 +0000534 }
drh50b39962006-10-28 00:28:09 +0000535 assert( op!=TK_ISNULL || c==WO_ISNULL );
drh51147ba2005-07-23 22:59:55 +0000536 assert( op!=TK_IN || c==WO_IN );
537 assert( op!=TK_EQ || c==WO_EQ );
538 assert( op!=TK_LT || c==WO_LT );
539 assert( op!=TK_LE || c==WO_LE );
540 assert( op!=TK_GT || c==WO_GT );
541 assert( op!=TK_GE || c==WO_GE );
542 return c;
drhfe05af82005-07-21 03:14:59 +0000543}
544
545/*
546** Search for a term in the WHERE clause that is of the form "X <op> <expr>"
547** where X is a reference to the iColumn of table iCur and <op> is one of
548** the WO_xx operator codes specified by the op parameter.
549** Return a pointer to the term. Return 0 if not found.
550*/
551static WhereTerm *findTerm(
552 WhereClause *pWC, /* The WHERE clause to be searched */
553 int iCur, /* Cursor number of LHS */
554 int iColumn, /* Column number of LHS */
555 Bitmask notReady, /* RHS must not overlap with this mask */
drhec1724e2008-12-09 01:32:03 +0000556 u32 op, /* Mask of WO_xx values describing operator */
drhfe05af82005-07-21 03:14:59 +0000557 Index *pIdx /* Must be compatible with this index, if not NULL */
558){
559 WhereTerm *pTerm;
560 int k;
drh22c24032008-07-09 13:28:53 +0000561 assert( iCur>=0 );
drhec1724e2008-12-09 01:32:03 +0000562 op &= WO_ALL;
drhfe05af82005-07-21 03:14:59 +0000563 for(pTerm=pWC->a, k=pWC->nTerm; k; k--, pTerm++){
564 if( pTerm->leftCursor==iCur
565 && (pTerm->prereqRight & notReady)==0
drh700a2262008-12-17 19:22:15 +0000566 && pTerm->u.leftColumn==iColumn
drhb52076c2006-01-23 13:22:09 +0000567 && (pTerm->eOperator & op)!=0
drhfe05af82005-07-21 03:14:59 +0000568 ){
drh22c24032008-07-09 13:28:53 +0000569 if( pIdx && pTerm->eOperator!=WO_ISNULL ){
drhfe05af82005-07-21 03:14:59 +0000570 Expr *pX = pTerm->pExpr;
571 CollSeq *pColl;
572 char idxaff;
danielk1977f0113002006-01-24 12:09:17 +0000573 int j;
drhfe05af82005-07-21 03:14:59 +0000574 Parse *pParse = pWC->pParse;
575
576 idxaff = pIdx->pTable->aCol[iColumn].affinity;
577 if( !sqlite3IndexAffinityOk(pX, idxaff) ) continue;
danielk1977bcbb04e2007-05-29 12:11:29 +0000578
579 /* Figure out the collation sequence required from an index for
580 ** it to be useful for optimising expression pX. Store this
581 ** value in variable pColl.
582 */
583 assert(pX->pLeft);
584 pColl = sqlite3BinaryCompareCollSeq(pParse, pX->pLeft, pX->pRight);
danielk197793574162008-12-30 15:26:29 +0000585 assert(pColl || pParse->nErr);
danielk1977bcbb04e2007-05-29 12:11:29 +0000586
drh22c24032008-07-09 13:28:53 +0000587 for(j=0; pIdx->aiColumn[j]!=iColumn; j++){
drh34004ce2008-07-11 16:15:17 +0000588 if( NEVER(j>=pIdx->nColumn) ) return 0;
drh22c24032008-07-09 13:28:53 +0000589 }
danielk197793574162008-12-30 15:26:29 +0000590 if( pColl && sqlite3StrICmp(pColl->zName, pIdx->azColl[j]) ) continue;
drhfe05af82005-07-21 03:14:59 +0000591 }
592 return pTerm;
593 }
594 }
595 return 0;
596}
597
drh6c30be82005-07-29 15:10:17 +0000598/* Forward reference */
drh7b4fc6a2007-02-06 13:26:32 +0000599static void exprAnalyze(SrcList*, WhereClause*, int);
drh6c30be82005-07-29 15:10:17 +0000600
601/*
602** Call exprAnalyze on all terms in a WHERE clause.
603**
604**
605*/
606static void exprAnalyzeAll(
607 SrcList *pTabList, /* the FROM clause */
drh6c30be82005-07-29 15:10:17 +0000608 WhereClause *pWC /* the WHERE clause to be analyzed */
609){
drh6c30be82005-07-29 15:10:17 +0000610 int i;
drh9eb20282005-08-24 03:52:18 +0000611 for(i=pWC->nTerm-1; i>=0; i--){
drh7b4fc6a2007-02-06 13:26:32 +0000612 exprAnalyze(pTabList, pWC, i);
drh6c30be82005-07-29 15:10:17 +0000613 }
614}
615
drhd2687b72005-08-12 22:56:09 +0000616#ifndef SQLITE_OMIT_LIKE_OPTIMIZATION
617/*
618** Check to see if the given expression is a LIKE or GLOB operator that
619** can be optimized using inequality constraints. Return TRUE if it is
620** so and false if not.
621**
622** In order for the operator to be optimizible, the RHS must be a string
623** literal that does not begin with a wildcard.
624*/
625static int isLikeOrGlob(
drh7d10d5a2008-08-20 16:35:10 +0000626 Parse *pParse, /* Parsing and code generating context */
drhd2687b72005-08-12 22:56:09 +0000627 Expr *pExpr, /* Test this expression */
dan937d0de2009-10-15 18:35:38 +0000628 Expr **ppPrefix, /* Pointer to TK_STRING expression with pattern prefix */
drh9f504ea2008-02-23 21:55:39 +0000629 int *pisComplete, /* True if the only wildcard is % in the last character */
630 int *pnoCase /* True if uppercase is equivalent to lowercase */
drhd2687b72005-08-12 22:56:09 +0000631){
dan937d0de2009-10-15 18:35:38 +0000632 const char *z = 0; /* String on RHS of LIKE operator */
drh5bd98ae2009-01-07 18:24:03 +0000633 Expr *pRight, *pLeft; /* Right and left size of LIKE operator */
634 ExprList *pList; /* List of operands to the LIKE operator */
635 int c; /* One character in z[] */
636 int cnt; /* Number of non-wildcard prefix characters */
637 char wc[3]; /* Wildcard characters */
638 CollSeq *pColl; /* Collating sequence for LHS */
639 sqlite3 *db = pParse->db; /* Database connection */
dan937d0de2009-10-15 18:35:38 +0000640 sqlite3_value *pVal = 0;
641 int op; /* Opcode of pRight */
drhd64fe2f2005-08-28 17:00:23 +0000642
drh9f504ea2008-02-23 21:55:39 +0000643 if( !sqlite3IsLikeFunction(db, pExpr, pnoCase, wc) ){
drhd2687b72005-08-12 22:56:09 +0000644 return 0;
645 }
drh9f504ea2008-02-23 21:55:39 +0000646#ifdef SQLITE_EBCDIC
647 if( *pnoCase ) return 0;
648#endif
danielk19776ab3a2e2009-02-19 14:39:25 +0000649 pList = pExpr->x.pList;
drh55ef4d92005-08-14 01:20:37 +0000650 pLeft = pList->a[1].pExpr;
drhd2687b72005-08-12 22:56:09 +0000651 if( pLeft->op!=TK_COLUMN ){
652 return 0;
653 }
drh7d10d5a2008-08-20 16:35:10 +0000654 pColl = sqlite3ExprCollSeq(pParse, pLeft);
drh01495b92008-01-23 12:52:40 +0000655 assert( pColl!=0 || pLeft->iColumn==-1 );
drhc4ac22e2009-06-07 23:45:10 +0000656 if( pColl==0 ) return 0;
drh9f504ea2008-02-23 21:55:39 +0000657 if( (pColl->type!=SQLITE_COLL_BINARY || *pnoCase) &&
658 (pColl->type!=SQLITE_COLL_NOCASE || !*pnoCase) ){
drhd64fe2f2005-08-28 17:00:23 +0000659 return 0;
660 }
drhc4ac22e2009-06-07 23:45:10 +0000661 if( sqlite3ExprAffinity(pLeft)!=SQLITE_AFF_TEXT ) return 0;
dan937d0de2009-10-15 18:35:38 +0000662
663 pRight = pList->a[0].pExpr;
664 op = pRight->op;
665 if( op==TK_REGISTER ){
666 op = pRight->op2;
667 }
668 if( op==TK_VARIABLE ){
669 Vdbe *pReprepare = pParse->pReprepare;
670 pVal = sqlite3VdbeGetValue(pReprepare, pRight->iColumn, SQLITE_AFF_NONE);
671 if( pVal && sqlite3_value_type(pVal)==SQLITE_TEXT ){
672 z = (char *)sqlite3_value_text(pVal);
673 }
dan1d2ce4f2009-10-19 18:11:09 +0000674 sqlite3VdbeSetVarmask(pParse->pVdbe, pRight->iColumn);
dan937d0de2009-10-15 18:35:38 +0000675 assert( pRight->op==TK_VARIABLE || pRight->op==TK_REGISTER );
676 }else if( op==TK_STRING ){
677 z = pRight->u.zToken;
678 }
679 if( z ){
shane85095702009-06-15 16:27:08 +0000680 cnt = 0;
drhb7916a72009-05-27 10:31:29 +0000681 while( (c=z[cnt])!=0 && c!=wc[0] && c!=wc[1] && c!=wc[2] ){
drh24fb6272009-05-01 21:13:36 +0000682 cnt++;
683 }
shane85095702009-06-15 16:27:08 +0000684 if( cnt!=0 && c!=0 && 255!=(u8)z[cnt-1] ){
dan937d0de2009-10-15 18:35:38 +0000685 Expr *pPrefix;
shane85095702009-06-15 16:27:08 +0000686 *pisComplete = z[cnt]==wc[0] && z[cnt+1]==0;
dan937d0de2009-10-15 18:35:38 +0000687 pPrefix = sqlite3Expr(db, TK_STRING, z);
688 if( pPrefix ) pPrefix->u.zToken[cnt] = 0;
689 *ppPrefix = pPrefix;
690 if( op==TK_VARIABLE ){
691 Vdbe *v = pParse->pVdbe;
dan1d2ce4f2009-10-19 18:11:09 +0000692 sqlite3VdbeSetVarmask(v, pRight->iColumn);
dan937d0de2009-10-15 18:35:38 +0000693 if( *pisComplete && pRight->u.zToken[1] ){
694 /* If the rhs of the LIKE expression is a variable, and the current
695 ** value of the variable means there is no need to invoke the LIKE
696 ** function, then no OP_Variable will be added to the program.
697 ** This causes problems for the sqlite3_bind_parameter_name()
drhbec451f2009-10-17 13:13:02 +0000698 ** API. To workaround them, add a dummy OP_Variable here.
699 */
700 int r1 = sqlite3GetTempReg(pParse);
701 sqlite3ExprCodeTarget(pParse, pRight, r1);
dan937d0de2009-10-15 18:35:38 +0000702 sqlite3VdbeChangeP3(v, sqlite3VdbeCurrentAddr(v)-1, 0);
drhbec451f2009-10-17 13:13:02 +0000703 sqlite3ReleaseTempReg(pParse, r1);
dan937d0de2009-10-15 18:35:38 +0000704 }
705 }
706 }else{
707 z = 0;
shane85095702009-06-15 16:27:08 +0000708 }
drhf998b732007-11-26 13:36:00 +0000709 }
dan937d0de2009-10-15 18:35:38 +0000710
711 sqlite3ValueFree(pVal);
712 return (z!=0);
drhd2687b72005-08-12 22:56:09 +0000713}
714#endif /* SQLITE_OMIT_LIKE_OPTIMIZATION */
715
drhedb193b2006-06-27 13:20:21 +0000716
717#ifndef SQLITE_OMIT_VIRTUALTABLE
drhfe05af82005-07-21 03:14:59 +0000718/*
drh7f375902006-06-13 17:38:59 +0000719** Check to see if the given expression is of the form
720**
721** column MATCH expr
722**
723** If it is then return TRUE. If not, return FALSE.
724*/
725static int isMatchOfColumn(
726 Expr *pExpr /* Test this expression */
727){
728 ExprList *pList;
729
730 if( pExpr->op!=TK_FUNCTION ){
731 return 0;
732 }
drh33e619f2009-05-28 01:00:55 +0000733 if( sqlite3StrICmp(pExpr->u.zToken,"match")!=0 ){
drh7f375902006-06-13 17:38:59 +0000734 return 0;
735 }
danielk19776ab3a2e2009-02-19 14:39:25 +0000736 pList = pExpr->x.pList;
drh7f375902006-06-13 17:38:59 +0000737 if( pList->nExpr!=2 ){
738 return 0;
739 }
740 if( pList->a[1].pExpr->op != TK_COLUMN ){
741 return 0;
742 }
743 return 1;
744}
drhedb193b2006-06-27 13:20:21 +0000745#endif /* SQLITE_OMIT_VIRTUALTABLE */
drh7f375902006-06-13 17:38:59 +0000746
747/*
drh54a167d2005-11-26 14:08:07 +0000748** If the pBase expression originated in the ON or USING clause of
749** a join, then transfer the appropriate markings over to derived.
750*/
751static void transferJoinMarkings(Expr *pDerived, Expr *pBase){
752 pDerived->flags |= pBase->flags & EP_FromJoin;
753 pDerived->iRightJoinTable = pBase->iRightJoinTable;
754}
755
drh3e355802007-02-23 23:13:33 +0000756#if !defined(SQLITE_OMIT_OR_OPTIMIZATION) && !defined(SQLITE_OMIT_SUBQUERY)
757/*
drh1a58fe02008-12-20 02:06:13 +0000758** Analyze a term that consists of two or more OR-connected
759** subterms. So in:
drh3e355802007-02-23 23:13:33 +0000760**
drh1a58fe02008-12-20 02:06:13 +0000761** ... WHERE (a=5) AND (b=7 OR c=9 OR d=13) AND (d=13)
762** ^^^^^^^^^^^^^^^^^^^^
drh3e355802007-02-23 23:13:33 +0000763**
drh1a58fe02008-12-20 02:06:13 +0000764** This routine analyzes terms such as the middle term in the above example.
765** A WhereOrTerm object is computed and attached to the term under
766** analysis, regardless of the outcome of the analysis. Hence:
drh3e355802007-02-23 23:13:33 +0000767**
drh1a58fe02008-12-20 02:06:13 +0000768** WhereTerm.wtFlags |= TERM_ORINFO
769** WhereTerm.u.pOrInfo = a dynamically allocated WhereOrTerm object
drh3e355802007-02-23 23:13:33 +0000770**
drh1a58fe02008-12-20 02:06:13 +0000771** The term being analyzed must have two or more of OR-connected subterms.
danielk1977fdc40192008-12-29 18:33:32 +0000772** A single subterm might be a set of AND-connected sub-subterms.
drh1a58fe02008-12-20 02:06:13 +0000773** Examples of terms under analysis:
drh3e355802007-02-23 23:13:33 +0000774**
drh1a58fe02008-12-20 02:06:13 +0000775** (A) t1.x=t2.y OR t1.x=t2.z OR t1.y=15 OR t1.z=t3.a+5
776** (B) x=expr1 OR expr2=x OR x=expr3
777** (C) t1.x=t2.y OR (t1.x=t2.z AND t1.y=15)
778** (D) x=expr1 OR (y>11 AND y<22 AND z LIKE '*hello*')
779** (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 +0000780**
drh1a58fe02008-12-20 02:06:13 +0000781** CASE 1:
782**
783** If all subterms are of the form T.C=expr for some single column of C
784** a single table T (as shown in example B above) then create a new virtual
785** term that is an equivalent IN expression. In other words, if the term
786** being analyzed is:
787**
788** x = expr1 OR expr2 = x OR x = expr3
789**
790** then create a new virtual term like this:
791**
792** x IN (expr1,expr2,expr3)
793**
794** CASE 2:
795**
796** If all subterms are indexable by a single table T, then set
797**
798** WhereTerm.eOperator = WO_OR
799** WhereTerm.u.pOrInfo->indexable |= the cursor number for table T
800**
801** A subterm is "indexable" if it is of the form
802** "T.C <op> <expr>" where C is any column of table T and
803** <op> is one of "=", "<", "<=", ">", ">=", "IS NULL", or "IN".
804** A subterm is also indexable if it is an AND of two or more
805** subsubterms at least one of which is indexable. Indexable AND
806** subterms have their eOperator set to WO_AND and they have
807** u.pAndInfo set to a dynamically allocated WhereAndTerm object.
808**
809** From another point of view, "indexable" means that the subterm could
810** potentially be used with an index if an appropriate index exists.
811** This analysis does not consider whether or not the index exists; that
812** is something the bestIndex() routine will determine. This analysis
813** only looks at whether subterms appropriate for indexing exist.
814**
815** All examples A through E above all satisfy case 2. But if a term
816** also statisfies case 1 (such as B) we know that the optimizer will
817** always prefer case 1, so in that case we pretend that case 2 is not
818** satisfied.
819**
820** It might be the case that multiple tables are indexable. For example,
821** (E) above is indexable on tables P, Q, and R.
822**
823** Terms that satisfy case 2 are candidates for lookup by using
824** separate indices to find rowids for each subterm and composing
825** the union of all rowids using a RowSet object. This is similar
826** to "bitmap indices" in other database engines.
827**
828** OTHERWISE:
829**
830** If neither case 1 nor case 2 apply, then leave the eOperator set to
831** zero. This term is not useful for search.
drh3e355802007-02-23 23:13:33 +0000832*/
drh1a58fe02008-12-20 02:06:13 +0000833static void exprAnalyzeOrTerm(
834 SrcList *pSrc, /* the FROM clause */
835 WhereClause *pWC, /* the complete WHERE clause */
836 int idxTerm /* Index of the OR-term to be analyzed */
837){
838 Parse *pParse = pWC->pParse; /* Parser context */
839 sqlite3 *db = pParse->db; /* Database connection */
840 WhereTerm *pTerm = &pWC->a[idxTerm]; /* The term to be analyzed */
841 Expr *pExpr = pTerm->pExpr; /* The expression of the term */
drh111a6a72008-12-21 03:51:16 +0000842 WhereMaskSet *pMaskSet = pWC->pMaskSet; /* Table use masks */
drh1a58fe02008-12-20 02:06:13 +0000843 int i; /* Loop counters */
844 WhereClause *pOrWc; /* Breakup of pTerm into subterms */
845 WhereTerm *pOrTerm; /* A Sub-term within the pOrWc */
846 WhereOrInfo *pOrInfo; /* Additional information associated with pTerm */
847 Bitmask chngToIN; /* Tables that might satisfy case 1 */
848 Bitmask indexable; /* Tables that are indexable, satisfying case 2 */
drh3e355802007-02-23 23:13:33 +0000849
drh1a58fe02008-12-20 02:06:13 +0000850 /*
851 ** Break the OR clause into its separate subterms. The subterms are
852 ** stored in a WhereClause structure containing within the WhereOrInfo
853 ** object that is attached to the original OR clause term.
854 */
855 assert( (pTerm->wtFlags & (TERM_DYNAMIC|TERM_ORINFO|TERM_ANDINFO))==0 );
856 assert( pExpr->op==TK_OR );
drh954701a2008-12-29 23:45:07 +0000857 pTerm->u.pOrInfo = pOrInfo = sqlite3DbMallocZero(db, sizeof(*pOrInfo));
drh1a58fe02008-12-20 02:06:13 +0000858 if( pOrInfo==0 ) return;
859 pTerm->wtFlags |= TERM_ORINFO;
860 pOrWc = &pOrInfo->wc;
861 whereClauseInit(pOrWc, pWC->pParse, pMaskSet);
862 whereSplit(pOrWc, pExpr, TK_OR);
863 exprAnalyzeAll(pSrc, pOrWc);
864 if( db->mallocFailed ) return;
865 assert( pOrWc->nTerm>=2 );
866
867 /*
868 ** Compute the set of tables that might satisfy cases 1 or 2.
869 */
danielk1977e672c8e2009-05-22 15:43:26 +0000870 indexable = ~(Bitmask)0;
871 chngToIN = ~(pWC->vmask);
drh1a58fe02008-12-20 02:06:13 +0000872 for(i=pOrWc->nTerm-1, pOrTerm=pOrWc->a; i>=0 && indexable; i--, pOrTerm++){
873 if( (pOrTerm->eOperator & WO_SINGLE)==0 ){
drh29435252008-12-28 18:35:08 +0000874 WhereAndInfo *pAndInfo;
875 assert( pOrTerm->eOperator==0 );
876 assert( (pOrTerm->wtFlags & (TERM_ANDINFO|TERM_ORINFO))==0 );
drh1a58fe02008-12-20 02:06:13 +0000877 chngToIN = 0;
drh29435252008-12-28 18:35:08 +0000878 pAndInfo = sqlite3DbMallocRaw(db, sizeof(*pAndInfo));
879 if( pAndInfo ){
880 WhereClause *pAndWC;
881 WhereTerm *pAndTerm;
882 int j;
883 Bitmask b = 0;
884 pOrTerm->u.pAndInfo = pAndInfo;
885 pOrTerm->wtFlags |= TERM_ANDINFO;
886 pOrTerm->eOperator = WO_AND;
887 pAndWC = &pAndInfo->wc;
888 whereClauseInit(pAndWC, pWC->pParse, pMaskSet);
889 whereSplit(pAndWC, pOrTerm->pExpr, TK_AND);
890 exprAnalyzeAll(pSrc, pAndWC);
drh7c2fbde2009-01-07 20:58:57 +0000891 testcase( db->mallocFailed );
drh96c7a7d2009-01-10 15:34:12 +0000892 if( !db->mallocFailed ){
893 for(j=0, pAndTerm=pAndWC->a; j<pAndWC->nTerm; j++, pAndTerm++){
894 assert( pAndTerm->pExpr );
895 if( allowedOp(pAndTerm->pExpr->op) ){
896 b |= getMask(pMaskSet, pAndTerm->leftCursor);
897 }
drh29435252008-12-28 18:35:08 +0000898 }
899 }
900 indexable &= b;
901 }
drh1a58fe02008-12-20 02:06:13 +0000902 }else if( pOrTerm->wtFlags & TERM_COPIED ){
903 /* Skip this term for now. We revisit it when we process the
904 ** corresponding TERM_VIRTUAL term */
905 }else{
906 Bitmask b;
907 b = getMask(pMaskSet, pOrTerm->leftCursor);
908 if( pOrTerm->wtFlags & TERM_VIRTUAL ){
909 WhereTerm *pOther = &pOrWc->a[pOrTerm->iParent];
910 b |= getMask(pMaskSet, pOther->leftCursor);
911 }
912 indexable &= b;
913 if( pOrTerm->eOperator!=WO_EQ ){
914 chngToIN = 0;
915 }else{
916 chngToIN &= b;
917 }
918 }
drh3e355802007-02-23 23:13:33 +0000919 }
drh1a58fe02008-12-20 02:06:13 +0000920
921 /*
922 ** Record the set of tables that satisfy case 2. The set might be
drh111a6a72008-12-21 03:51:16 +0000923 ** empty.
drh1a58fe02008-12-20 02:06:13 +0000924 */
925 pOrInfo->indexable = indexable;
drh111a6a72008-12-21 03:51:16 +0000926 pTerm->eOperator = indexable==0 ? 0 : WO_OR;
drh1a58fe02008-12-20 02:06:13 +0000927
928 /*
929 ** chngToIN holds a set of tables that *might* satisfy case 1. But
930 ** we have to do some additional checking to see if case 1 really
931 ** is satisfied.
drh4e8be3b2009-06-08 17:11:08 +0000932 **
933 ** chngToIN will hold either 0, 1, or 2 bits. The 0-bit case means
934 ** that there is no possibility of transforming the OR clause into an
935 ** IN operator because one or more terms in the OR clause contain
936 ** something other than == on a column in the single table. The 1-bit
937 ** case means that every term of the OR clause is of the form
938 ** "table.column=expr" for some single table. The one bit that is set
939 ** will correspond to the common table. We still need to check to make
940 ** sure the same column is used on all terms. The 2-bit case is when
941 ** the all terms are of the form "table1.column=table2.column". It
942 ** might be possible to form an IN operator with either table1.column
943 ** or table2.column as the LHS if either is common to every term of
944 ** the OR clause.
945 **
946 ** Note that terms of the form "table.column1=table.column2" (the
947 ** same table on both sizes of the ==) cannot be optimized.
drh1a58fe02008-12-20 02:06:13 +0000948 */
949 if( chngToIN ){
950 int okToChngToIN = 0; /* True if the conversion to IN is valid */
951 int iColumn = -1; /* Column index on lhs of IN operator */
shane63207ab2009-02-04 01:49:30 +0000952 int iCursor = -1; /* Table cursor common to all terms */
drh1a58fe02008-12-20 02:06:13 +0000953 int j = 0; /* Loop counter */
954
955 /* Search for a table and column that appears on one side or the
956 ** other of the == operator in every subterm. That table and column
957 ** will be recorded in iCursor and iColumn. There might not be any
958 ** such table and column. Set okToChngToIN if an appropriate table
959 ** and column is found but leave okToChngToIN false if not found.
960 */
961 for(j=0; j<2 && !okToChngToIN; j++){
962 pOrTerm = pOrWc->a;
963 for(i=pOrWc->nTerm-1; i>=0; i--, pOrTerm++){
964 assert( pOrTerm->eOperator==WO_EQ );
965 pOrTerm->wtFlags &= ~TERM_OR_OK;
drh4e8be3b2009-06-08 17:11:08 +0000966 if( pOrTerm->leftCursor==iCursor ){
967 /* This is the 2-bit case and we are on the second iteration and
968 ** current term is from the first iteration. So skip this term. */
969 assert( j==1 );
970 continue;
971 }
972 if( (chngToIN & getMask(pMaskSet, pOrTerm->leftCursor))==0 ){
973 /* This term must be of the form t1.a==t2.b where t2 is in the
974 ** chngToIN set but t1 is not. This term will be either preceeded
975 ** or follwed by an inverted copy (t2.b==t1.a). Skip this term
976 ** and use its inversion. */
977 testcase( pOrTerm->wtFlags & TERM_COPIED );
978 testcase( pOrTerm->wtFlags & TERM_VIRTUAL );
979 assert( pOrTerm->wtFlags & (TERM_COPIED|TERM_VIRTUAL) );
980 continue;
981 }
drh1a58fe02008-12-20 02:06:13 +0000982 iColumn = pOrTerm->u.leftColumn;
983 iCursor = pOrTerm->leftCursor;
984 break;
985 }
986 if( i<0 ){
drh4e8be3b2009-06-08 17:11:08 +0000987 /* No candidate table+column was found. This can only occur
988 ** on the second iteration */
drh1a58fe02008-12-20 02:06:13 +0000989 assert( j==1 );
990 assert( (chngToIN&(chngToIN-1))==0 );
drh4e8be3b2009-06-08 17:11:08 +0000991 assert( chngToIN==getMask(pMaskSet, iCursor) );
drh1a58fe02008-12-20 02:06:13 +0000992 break;
993 }
drh4e8be3b2009-06-08 17:11:08 +0000994 testcase( j==1 );
995
996 /* We have found a candidate table and column. Check to see if that
997 ** table and column is common to every term in the OR clause */
drh1a58fe02008-12-20 02:06:13 +0000998 okToChngToIN = 1;
999 for(; i>=0 && okToChngToIN; i--, pOrTerm++){
1000 assert( pOrTerm->eOperator==WO_EQ );
1001 if( pOrTerm->leftCursor!=iCursor ){
1002 pOrTerm->wtFlags &= ~TERM_OR_OK;
1003 }else if( pOrTerm->u.leftColumn!=iColumn ){
1004 okToChngToIN = 0;
1005 }else{
1006 int affLeft, affRight;
1007 /* If the right-hand side is also a column, then the affinities
1008 ** of both right and left sides must be such that no type
1009 ** conversions are required on the right. (Ticket #2249)
1010 */
1011 affRight = sqlite3ExprAffinity(pOrTerm->pExpr->pRight);
1012 affLeft = sqlite3ExprAffinity(pOrTerm->pExpr->pLeft);
1013 if( affRight!=0 && affRight!=affLeft ){
1014 okToChngToIN = 0;
1015 }else{
1016 pOrTerm->wtFlags |= TERM_OR_OK;
1017 }
1018 }
1019 }
1020 }
1021
1022 /* At this point, okToChngToIN is true if original pTerm satisfies
1023 ** case 1. In that case, construct a new virtual term that is
1024 ** pTerm converted into an IN operator.
1025 */
1026 if( okToChngToIN ){
1027 Expr *pDup; /* A transient duplicate expression */
1028 ExprList *pList = 0; /* The RHS of the IN operator */
1029 Expr *pLeft = 0; /* The LHS of the IN operator */
1030 Expr *pNew; /* The complete IN operator */
1031
1032 for(i=pOrWc->nTerm-1, pOrTerm=pOrWc->a; i>=0; i--, pOrTerm++){
1033 if( (pOrTerm->wtFlags & TERM_OR_OK)==0 ) continue;
1034 assert( pOrTerm->eOperator==WO_EQ );
1035 assert( pOrTerm->leftCursor==iCursor );
1036 assert( pOrTerm->u.leftColumn==iColumn );
danielk19776ab3a2e2009-02-19 14:39:25 +00001037 pDup = sqlite3ExprDup(db, pOrTerm->pExpr->pRight, 0);
drhb7916a72009-05-27 10:31:29 +00001038 pList = sqlite3ExprListAppend(pWC->pParse, pList, pDup);
drh1a58fe02008-12-20 02:06:13 +00001039 pLeft = pOrTerm->pExpr->pLeft;
1040 }
1041 assert( pLeft!=0 );
danielk19776ab3a2e2009-02-19 14:39:25 +00001042 pDup = sqlite3ExprDup(db, pLeft, 0);
drhb7916a72009-05-27 10:31:29 +00001043 pNew = sqlite3PExpr(pParse, TK_IN, pDup, 0, 0);
drh1a58fe02008-12-20 02:06:13 +00001044 if( pNew ){
1045 int idxNew;
1046 transferJoinMarkings(pNew, pExpr);
danielk19776ab3a2e2009-02-19 14:39:25 +00001047 assert( !ExprHasProperty(pNew, EP_xIsSelect) );
1048 pNew->x.pList = pList;
drh1a58fe02008-12-20 02:06:13 +00001049 idxNew = whereClauseInsert(pWC, pNew, TERM_VIRTUAL|TERM_DYNAMIC);
1050 testcase( idxNew==0 );
1051 exprAnalyze(pSrc, pWC, idxNew);
1052 pTerm = &pWC->a[idxTerm];
1053 pWC->a[idxNew].iParent = idxTerm;
1054 pTerm->nChild = 1;
1055 }else{
1056 sqlite3ExprListDelete(db, pList);
1057 }
1058 pTerm->eOperator = 0; /* case 1 trumps case 2 */
1059 }
drh3e355802007-02-23 23:13:33 +00001060 }
drh3e355802007-02-23 23:13:33 +00001061}
1062#endif /* !SQLITE_OMIT_OR_OPTIMIZATION && !SQLITE_OMIT_SUBQUERY */
drh54a167d2005-11-26 14:08:07 +00001063
drh1a58fe02008-12-20 02:06:13 +00001064
drh54a167d2005-11-26 14:08:07 +00001065/*
drh0aa74ed2005-07-16 13:33:20 +00001066** The input to this routine is an WhereTerm structure with only the
drh51147ba2005-07-23 22:59:55 +00001067** "pExpr" field filled in. The job of this routine is to analyze the
drh0aa74ed2005-07-16 13:33:20 +00001068** subexpression and populate all the other fields of the WhereTerm
drh75897232000-05-29 14:26:00 +00001069** structure.
drh51147ba2005-07-23 22:59:55 +00001070**
1071** If the expression is of the form "<expr> <op> X" it gets commuted
drh1a58fe02008-12-20 02:06:13 +00001072** to the standard form of "X <op> <expr>".
1073**
1074** If the expression is of the form "X <op> Y" where both X and Y are
1075** columns, then the original expression is unchanged and a new virtual
1076** term of the form "Y <op> X" is added to the WHERE clause and
1077** analyzed separately. The original term is marked with TERM_COPIED
1078** and the new term is marked with TERM_DYNAMIC (because it's pExpr
1079** needs to be freed with the WhereClause) and TERM_VIRTUAL (because it
1080** is a commuted copy of a prior term.) The original term has nChild=1
1081** and the copy has idxParent set to the index of the original term.
drh75897232000-05-29 14:26:00 +00001082*/
drh0fcef5e2005-07-19 17:38:22 +00001083static void exprAnalyze(
1084 SrcList *pSrc, /* the FROM clause */
drh9eb20282005-08-24 03:52:18 +00001085 WhereClause *pWC, /* the WHERE clause */
1086 int idxTerm /* Index of the term to be analyzed */
drh0fcef5e2005-07-19 17:38:22 +00001087){
drh1a58fe02008-12-20 02:06:13 +00001088 WhereTerm *pTerm; /* The term to be analyzed */
drh111a6a72008-12-21 03:51:16 +00001089 WhereMaskSet *pMaskSet; /* Set of table index masks */
drh1a58fe02008-12-20 02:06:13 +00001090 Expr *pExpr; /* The expression to be analyzed */
1091 Bitmask prereqLeft; /* Prerequesites of the pExpr->pLeft */
1092 Bitmask prereqAll; /* Prerequesites of pExpr */
drhdafc0ce2008-04-17 19:14:02 +00001093 Bitmask extraRight = 0;
drhd2687b72005-08-12 22:56:09 +00001094 int isComplete;
drh9f504ea2008-02-23 21:55:39 +00001095 int noCase;
drh1a58fe02008-12-20 02:06:13 +00001096 int op; /* Top-level operator. pExpr->op */
1097 Parse *pParse = pWC->pParse; /* Parsing context */
1098 sqlite3 *db = pParse->db; /* Database connection */
dan937d0de2009-10-15 18:35:38 +00001099 Expr *pStr1;
drh0fcef5e2005-07-19 17:38:22 +00001100
drhf998b732007-11-26 13:36:00 +00001101 if( db->mallocFailed ){
1102 return;
1103 }
1104 pTerm = &pWC->a[idxTerm];
1105 pMaskSet = pWC->pMaskSet;
1106 pExpr = pTerm->pExpr;
drh0fcef5e2005-07-19 17:38:22 +00001107 prereqLeft = exprTableUsage(pMaskSet, pExpr->pLeft);
drh50b39962006-10-28 00:28:09 +00001108 op = pExpr->op;
1109 if( op==TK_IN ){
drhf5b11382005-09-17 13:07:13 +00001110 assert( pExpr->pRight==0 );
danielk19776ab3a2e2009-02-19 14:39:25 +00001111 if( ExprHasProperty(pExpr, EP_xIsSelect) ){
1112 pTerm->prereqRight = exprSelectTableUsage(pMaskSet, pExpr->x.pSelect);
1113 }else{
1114 pTerm->prereqRight = exprListTableUsage(pMaskSet, pExpr->x.pList);
1115 }
drh50b39962006-10-28 00:28:09 +00001116 }else if( op==TK_ISNULL ){
1117 pTerm->prereqRight = 0;
drhf5b11382005-09-17 13:07:13 +00001118 }else{
1119 pTerm->prereqRight = exprTableUsage(pMaskSet, pExpr->pRight);
1120 }
drh22d6a532005-09-19 21:05:48 +00001121 prereqAll = exprTableUsage(pMaskSet, pExpr);
1122 if( ExprHasProperty(pExpr, EP_FromJoin) ){
drh42165be2008-03-26 14:56:34 +00001123 Bitmask x = getMask(pMaskSet, pExpr->iRightJoinTable);
1124 prereqAll |= x;
drhdafc0ce2008-04-17 19:14:02 +00001125 extraRight = x-1; /* ON clause terms may not be used with an index
1126 ** on left table of a LEFT JOIN. Ticket #3015 */
drh22d6a532005-09-19 21:05:48 +00001127 }
1128 pTerm->prereqAll = prereqAll;
drh0fcef5e2005-07-19 17:38:22 +00001129 pTerm->leftCursor = -1;
drh45b1ee42005-08-02 17:48:22 +00001130 pTerm->iParent = -1;
drhb52076c2006-01-23 13:22:09 +00001131 pTerm->eOperator = 0;
drh50b39962006-10-28 00:28:09 +00001132 if( allowedOp(op) && (pTerm->prereqRight & prereqLeft)==0 ){
drh0fcef5e2005-07-19 17:38:22 +00001133 Expr *pLeft = pExpr->pLeft;
1134 Expr *pRight = pExpr->pRight;
1135 if( pLeft->op==TK_COLUMN ){
1136 pTerm->leftCursor = pLeft->iTable;
drh700a2262008-12-17 19:22:15 +00001137 pTerm->u.leftColumn = pLeft->iColumn;
drh50b39962006-10-28 00:28:09 +00001138 pTerm->eOperator = operatorMask(op);
drh75897232000-05-29 14:26:00 +00001139 }
drh0fcef5e2005-07-19 17:38:22 +00001140 if( pRight && pRight->op==TK_COLUMN ){
1141 WhereTerm *pNew;
1142 Expr *pDup;
1143 if( pTerm->leftCursor>=0 ){
drh9eb20282005-08-24 03:52:18 +00001144 int idxNew;
danielk19776ab3a2e2009-02-19 14:39:25 +00001145 pDup = sqlite3ExprDup(db, pExpr, 0);
drh17435752007-08-16 04:30:38 +00001146 if( db->mallocFailed ){
drh633e6d52008-07-28 19:34:53 +00001147 sqlite3ExprDelete(db, pDup);
drh28f45912006-10-18 23:26:38 +00001148 return;
1149 }
drh9eb20282005-08-24 03:52:18 +00001150 idxNew = whereClauseInsert(pWC, pDup, TERM_VIRTUAL|TERM_DYNAMIC);
1151 if( idxNew==0 ) return;
1152 pNew = &pWC->a[idxNew];
1153 pNew->iParent = idxTerm;
1154 pTerm = &pWC->a[idxTerm];
drh45b1ee42005-08-02 17:48:22 +00001155 pTerm->nChild = 1;
drh165be382008-12-05 02:36:33 +00001156 pTerm->wtFlags |= TERM_COPIED;
drh0fcef5e2005-07-19 17:38:22 +00001157 }else{
1158 pDup = pExpr;
1159 pNew = pTerm;
1160 }
drh7d10d5a2008-08-20 16:35:10 +00001161 exprCommute(pParse, pDup);
drh0fcef5e2005-07-19 17:38:22 +00001162 pLeft = pDup->pLeft;
1163 pNew->leftCursor = pLeft->iTable;
drh700a2262008-12-17 19:22:15 +00001164 pNew->u.leftColumn = pLeft->iColumn;
drh0fcef5e2005-07-19 17:38:22 +00001165 pNew->prereqRight = prereqLeft;
1166 pNew->prereqAll = prereqAll;
drhb52076c2006-01-23 13:22:09 +00001167 pNew->eOperator = operatorMask(pDup->op);
drh75897232000-05-29 14:26:00 +00001168 }
1169 }
drhed378002005-07-28 23:12:08 +00001170
drhd2687b72005-08-12 22:56:09 +00001171#ifndef SQLITE_OMIT_BETWEEN_OPTIMIZATION
drhed378002005-07-28 23:12:08 +00001172 /* If a term is the BETWEEN operator, create two new virtual terms
drh1a58fe02008-12-20 02:06:13 +00001173 ** that define the range that the BETWEEN implements. For example:
1174 **
1175 ** a BETWEEN b AND c
1176 **
1177 ** is converted into:
1178 **
1179 ** (a BETWEEN b AND c) AND (a>=b) AND (a<=c)
1180 **
1181 ** The two new terms are added onto the end of the WhereClause object.
1182 ** The new terms are "dynamic" and are children of the original BETWEEN
1183 ** term. That means that if the BETWEEN term is coded, the children are
1184 ** skipped. Or, if the children are satisfied by an index, the original
1185 ** BETWEEN term is skipped.
drhed378002005-07-28 23:12:08 +00001186 */
drh29435252008-12-28 18:35:08 +00001187 else if( pExpr->op==TK_BETWEEN && pWC->op==TK_AND ){
danielk19776ab3a2e2009-02-19 14:39:25 +00001188 ExprList *pList = pExpr->x.pList;
drhed378002005-07-28 23:12:08 +00001189 int i;
1190 static const u8 ops[] = {TK_GE, TK_LE};
1191 assert( pList!=0 );
1192 assert( pList->nExpr==2 );
1193 for(i=0; i<2; i++){
1194 Expr *pNewExpr;
drh9eb20282005-08-24 03:52:18 +00001195 int idxNew;
drhb7916a72009-05-27 10:31:29 +00001196 pNewExpr = sqlite3PExpr(pParse, ops[i],
1197 sqlite3ExprDup(db, pExpr->pLeft, 0),
danielk19776ab3a2e2009-02-19 14:39:25 +00001198 sqlite3ExprDup(db, pList->a[i].pExpr, 0), 0);
drh9eb20282005-08-24 03:52:18 +00001199 idxNew = whereClauseInsert(pWC, pNewExpr, TERM_VIRTUAL|TERM_DYNAMIC);
drh6a1e0712008-12-05 15:24:15 +00001200 testcase( idxNew==0 );
drh7b4fc6a2007-02-06 13:26:32 +00001201 exprAnalyze(pSrc, pWC, idxNew);
drh9eb20282005-08-24 03:52:18 +00001202 pTerm = &pWC->a[idxTerm];
1203 pWC->a[idxNew].iParent = idxTerm;
drhed378002005-07-28 23:12:08 +00001204 }
drh45b1ee42005-08-02 17:48:22 +00001205 pTerm->nChild = 2;
drhed378002005-07-28 23:12:08 +00001206 }
drhd2687b72005-08-12 22:56:09 +00001207#endif /* SQLITE_OMIT_BETWEEN_OPTIMIZATION */
drhed378002005-07-28 23:12:08 +00001208
danielk19771576cd92006-01-14 08:02:28 +00001209#if !defined(SQLITE_OMIT_OR_OPTIMIZATION) && !defined(SQLITE_OMIT_SUBQUERY)
drh1a58fe02008-12-20 02:06:13 +00001210 /* Analyze a term that is composed of two or more subterms connected by
1211 ** an OR operator.
drh6c30be82005-07-29 15:10:17 +00001212 */
1213 else if( pExpr->op==TK_OR ){
drh29435252008-12-28 18:35:08 +00001214 assert( pWC->op==TK_AND );
drh1a58fe02008-12-20 02:06:13 +00001215 exprAnalyzeOrTerm(pSrc, pWC, idxTerm);
danielk1977f51d1bd2009-07-31 06:14:51 +00001216 pTerm = &pWC->a[idxTerm];
drh6c30be82005-07-29 15:10:17 +00001217 }
drhd2687b72005-08-12 22:56:09 +00001218#endif /* SQLITE_OMIT_OR_OPTIMIZATION */
1219
1220#ifndef SQLITE_OMIT_LIKE_OPTIMIZATION
1221 /* Add constraints to reduce the search space on a LIKE or GLOB
1222 ** operator.
drh9f504ea2008-02-23 21:55:39 +00001223 **
1224 ** A like pattern of the form "x LIKE 'abc%'" is changed into constraints
1225 **
1226 ** x>='abc' AND x<'abd' AND x LIKE 'abc%'
1227 **
1228 ** The last character of the prefix "abc" is incremented to form the
shane7bc71e52008-05-28 18:01:44 +00001229 ** termination condition "abd".
drhd2687b72005-08-12 22:56:09 +00001230 */
dan937d0de2009-10-15 18:35:38 +00001231 if( pWC->op==TK_AND
1232 && isLikeOrGlob(pParse, pExpr, &pStr1, &isComplete, &noCase)
1233 ){
1234 Expr *pLeft;
1235 Expr *pStr2;
drhd2687b72005-08-12 22:56:09 +00001236 Expr *pNewExpr1, *pNewExpr2;
drh9eb20282005-08-24 03:52:18 +00001237 int idxNew1, idxNew2;
1238
danielk19776ab3a2e2009-02-19 14:39:25 +00001239 pLeft = pExpr->x.pList->a[1].pExpr;
danielk19776ab3a2e2009-02-19 14:39:25 +00001240 pStr2 = sqlite3ExprDup(db, pStr1, 0);
drhf998b732007-11-26 13:36:00 +00001241 if( !db->mallocFailed ){
drh254993e2009-06-08 19:44:36 +00001242 u8 c, *pC; /* Last character before the first wildcard */
dan937d0de2009-10-15 18:35:38 +00001243 pC = (u8*)&pStr2->u.zToken[sqlite3Strlen30(pStr2->u.zToken)-1];
drh9f504ea2008-02-23 21:55:39 +00001244 c = *pC;
drh02a50b72008-05-26 18:33:40 +00001245 if( noCase ){
drh254993e2009-06-08 19:44:36 +00001246 /* The point is to increment the last character before the first
1247 ** wildcard. But if we increment '@', that will push it into the
1248 ** alphabetic range where case conversions will mess up the
1249 ** inequality. To avoid this, make sure to also run the full
1250 ** LIKE on all candidate expressions by clearing the isComplete flag
1251 */
1252 if( c=='A'-1 ) isComplete = 0;
1253
drh02a50b72008-05-26 18:33:40 +00001254 c = sqlite3UpperToLower[c];
1255 }
drh9f504ea2008-02-23 21:55:39 +00001256 *pC = c + 1;
drhd2687b72005-08-12 22:56:09 +00001257 }
danielk19776ab3a2e2009-02-19 14:39:25 +00001258 pNewExpr1 = sqlite3PExpr(pParse, TK_GE, sqlite3ExprDup(db,pLeft,0),pStr1,0);
drh9eb20282005-08-24 03:52:18 +00001259 idxNew1 = whereClauseInsert(pWC, pNewExpr1, TERM_VIRTUAL|TERM_DYNAMIC);
drh6a1e0712008-12-05 15:24:15 +00001260 testcase( idxNew1==0 );
drh7b4fc6a2007-02-06 13:26:32 +00001261 exprAnalyze(pSrc, pWC, idxNew1);
danielk19776ab3a2e2009-02-19 14:39:25 +00001262 pNewExpr2 = sqlite3PExpr(pParse, TK_LT, sqlite3ExprDup(db,pLeft,0),pStr2,0);
drh9eb20282005-08-24 03:52:18 +00001263 idxNew2 = whereClauseInsert(pWC, pNewExpr2, TERM_VIRTUAL|TERM_DYNAMIC);
drh6a1e0712008-12-05 15:24:15 +00001264 testcase( idxNew2==0 );
drh7b4fc6a2007-02-06 13:26:32 +00001265 exprAnalyze(pSrc, pWC, idxNew2);
drh9eb20282005-08-24 03:52:18 +00001266 pTerm = &pWC->a[idxTerm];
drhd2687b72005-08-12 22:56:09 +00001267 if( isComplete ){
drh9eb20282005-08-24 03:52:18 +00001268 pWC->a[idxNew1].iParent = idxTerm;
1269 pWC->a[idxNew2].iParent = idxTerm;
drhd2687b72005-08-12 22:56:09 +00001270 pTerm->nChild = 2;
1271 }
1272 }
1273#endif /* SQLITE_OMIT_LIKE_OPTIMIZATION */
drh7f375902006-06-13 17:38:59 +00001274
1275#ifndef SQLITE_OMIT_VIRTUALTABLE
1276 /* Add a WO_MATCH auxiliary term to the constraint set if the
1277 ** current expression is of the form: column MATCH expr.
1278 ** This information is used by the xBestIndex methods of
1279 ** virtual tables. The native query optimizer does not attempt
1280 ** to do anything with MATCH functions.
1281 */
1282 if( isMatchOfColumn(pExpr) ){
1283 int idxNew;
1284 Expr *pRight, *pLeft;
1285 WhereTerm *pNewTerm;
1286 Bitmask prereqColumn, prereqExpr;
1287
danielk19776ab3a2e2009-02-19 14:39:25 +00001288 pRight = pExpr->x.pList->a[0].pExpr;
1289 pLeft = pExpr->x.pList->a[1].pExpr;
drh7f375902006-06-13 17:38:59 +00001290 prereqExpr = exprTableUsage(pMaskSet, pRight);
1291 prereqColumn = exprTableUsage(pMaskSet, pLeft);
1292 if( (prereqExpr & prereqColumn)==0 ){
drh1a90e092006-06-14 22:07:10 +00001293 Expr *pNewExpr;
drhb7916a72009-05-27 10:31:29 +00001294 pNewExpr = sqlite3PExpr(pParse, TK_MATCH,
1295 0, sqlite3ExprDup(db, pRight, 0), 0);
drh1a90e092006-06-14 22:07:10 +00001296 idxNew = whereClauseInsert(pWC, pNewExpr, TERM_VIRTUAL|TERM_DYNAMIC);
drh6a1e0712008-12-05 15:24:15 +00001297 testcase( idxNew==0 );
drh7f375902006-06-13 17:38:59 +00001298 pNewTerm = &pWC->a[idxNew];
1299 pNewTerm->prereqRight = prereqExpr;
1300 pNewTerm->leftCursor = pLeft->iTable;
drh700a2262008-12-17 19:22:15 +00001301 pNewTerm->u.leftColumn = pLeft->iColumn;
drh7f375902006-06-13 17:38:59 +00001302 pNewTerm->eOperator = WO_MATCH;
1303 pNewTerm->iParent = idxTerm;
drhd2ca60d2006-06-27 02:36:58 +00001304 pTerm = &pWC->a[idxTerm];
drh7f375902006-06-13 17:38:59 +00001305 pTerm->nChild = 1;
drh165be382008-12-05 02:36:33 +00001306 pTerm->wtFlags |= TERM_COPIED;
drh7f375902006-06-13 17:38:59 +00001307 pNewTerm->prereqAll = pTerm->prereqAll;
1308 }
1309 }
1310#endif /* SQLITE_OMIT_VIRTUALTABLE */
drhdafc0ce2008-04-17 19:14:02 +00001311
1312 /* Prevent ON clause terms of a LEFT JOIN from being used to drive
1313 ** an index for tables to the left of the join.
1314 */
1315 pTerm->prereqRight |= extraRight;
drh75897232000-05-29 14:26:00 +00001316}
1317
drh7b4fc6a2007-02-06 13:26:32 +00001318/*
1319** Return TRUE if any of the expressions in pList->a[iFirst...] contain
1320** a reference to any table other than the iBase table.
1321*/
1322static int referencesOtherTables(
1323 ExprList *pList, /* Search expressions in ths list */
drh111a6a72008-12-21 03:51:16 +00001324 WhereMaskSet *pMaskSet, /* Mapping from tables to bitmaps */
drh7b4fc6a2007-02-06 13:26:32 +00001325 int iFirst, /* Be searching with the iFirst-th expression */
1326 int iBase /* Ignore references to this table */
1327){
1328 Bitmask allowed = ~getMask(pMaskSet, iBase);
1329 while( iFirst<pList->nExpr ){
1330 if( (exprTableUsage(pMaskSet, pList->a[iFirst++].pExpr)&allowed)!=0 ){
1331 return 1;
1332 }
1333 }
1334 return 0;
1335}
1336
drh0fcef5e2005-07-19 17:38:22 +00001337
drh75897232000-05-29 14:26:00 +00001338/*
drh51669862004-12-18 18:40:26 +00001339** This routine decides if pIdx can be used to satisfy the ORDER BY
1340** clause. If it can, it returns 1. If pIdx cannot satisfy the
1341** ORDER BY clause, this routine returns 0.
1342**
1343** pOrderBy is an ORDER BY clause from a SELECT statement. pTab is the
1344** left-most table in the FROM clause of that same SELECT statement and
1345** the table has a cursor number of "base". pIdx is an index on pTab.
1346**
1347** nEqCol is the number of columns of pIdx that are used as equality
1348** constraints. Any of these columns may be missing from the ORDER BY
1349** clause and the match can still be a success.
1350**
drh51669862004-12-18 18:40:26 +00001351** All terms of the ORDER BY that match against the index must be either
1352** ASC or DESC. (Terms of the ORDER BY clause past the end of a UNIQUE
1353** index do not need to satisfy this constraint.) The *pbRev value is
1354** set to 1 if the ORDER BY clause is all DESC and it is set to 0 if
1355** the ORDER BY clause is all ASC.
1356*/
1357static int isSortingIndex(
1358 Parse *pParse, /* Parsing context */
drh111a6a72008-12-21 03:51:16 +00001359 WhereMaskSet *pMaskSet, /* Mapping from table cursor numbers to bitmaps */
drh51669862004-12-18 18:40:26 +00001360 Index *pIdx, /* The index we are testing */
drh74161702006-02-24 02:53:49 +00001361 int base, /* Cursor number for the table to be sorted */
drh51669862004-12-18 18:40:26 +00001362 ExprList *pOrderBy, /* The ORDER BY clause */
1363 int nEqCol, /* Number of index columns with == constraints */
1364 int *pbRev /* Set to 1 if ORDER BY is DESC */
1365){
drhb46b5772005-08-29 16:40:52 +00001366 int i, j; /* Loop counters */
drh85eeb692005-12-21 03:16:42 +00001367 int sortOrder = 0; /* XOR of index and ORDER BY sort direction */
drhb46b5772005-08-29 16:40:52 +00001368 int nTerm; /* Number of ORDER BY terms */
1369 struct ExprList_item *pTerm; /* A term of the ORDER BY clause */
drh51669862004-12-18 18:40:26 +00001370 sqlite3 *db = pParse->db;
1371
1372 assert( pOrderBy!=0 );
1373 nTerm = pOrderBy->nExpr;
1374 assert( nTerm>0 );
1375
dan5236ac12009-08-13 07:09:33 +00001376 /* Argument pIdx must either point to a 'real' named index structure,
1377 ** or an index structure allocated on the stack by bestBtreeIndex() to
1378 ** represent the rowid index that is part of every table. */
1379 assert( pIdx->zName || (pIdx->nColumn==1 && pIdx->aiColumn[0]==-1) );
1380
drh51669862004-12-18 18:40:26 +00001381 /* Match terms of the ORDER BY clause against columns of
1382 ** the index.
drhcc192542006-12-20 03:24:19 +00001383 **
1384 ** Note that indices have pIdx->nColumn regular columns plus
1385 ** one additional column containing the rowid. The rowid column
1386 ** of the index is also allowed to match against the ORDER BY
1387 ** clause.
drh51669862004-12-18 18:40:26 +00001388 */
drhcc192542006-12-20 03:24:19 +00001389 for(i=j=0, pTerm=pOrderBy->a; j<nTerm && i<=pIdx->nColumn; i++){
drh51669862004-12-18 18:40:26 +00001390 Expr *pExpr; /* The expression of the ORDER BY pTerm */
1391 CollSeq *pColl; /* The collating sequence of pExpr */
drh85eeb692005-12-21 03:16:42 +00001392 int termSortOrder; /* Sort order for this term */
drhcc192542006-12-20 03:24:19 +00001393 int iColumn; /* The i-th column of the index. -1 for rowid */
1394 int iSortOrder; /* 1 for DESC, 0 for ASC on the i-th index term */
1395 const char *zColl; /* Name of the collating sequence for i-th index term */
drh51669862004-12-18 18:40:26 +00001396
1397 pExpr = pTerm->pExpr;
1398 if( pExpr->op!=TK_COLUMN || pExpr->iTable!=base ){
1399 /* Can not use an index sort on anything that is not a column in the
1400 ** left-most table of the FROM clause */
drh7b4fc6a2007-02-06 13:26:32 +00001401 break;
drh51669862004-12-18 18:40:26 +00001402 }
1403 pColl = sqlite3ExprCollSeq(pParse, pExpr);
drhcc192542006-12-20 03:24:19 +00001404 if( !pColl ){
1405 pColl = db->pDfltColl;
1406 }
dan5236ac12009-08-13 07:09:33 +00001407 if( pIdx->zName && i<pIdx->nColumn ){
drhcc192542006-12-20 03:24:19 +00001408 iColumn = pIdx->aiColumn[i];
1409 if( iColumn==pIdx->pTable->iPKey ){
1410 iColumn = -1;
1411 }
1412 iSortOrder = pIdx->aSortOrder[i];
1413 zColl = pIdx->azColl[i];
1414 }else{
1415 iColumn = -1;
1416 iSortOrder = 0;
1417 zColl = pColl->zName;
1418 }
1419 if( pExpr->iColumn!=iColumn || sqlite3StrICmp(pColl->zName, zColl) ){
drh9012bcb2004-12-19 00:11:35 +00001420 /* Term j of the ORDER BY clause does not match column i of the index */
1421 if( i<nEqCol ){
drh51669862004-12-18 18:40:26 +00001422 /* If an index column that is constrained by == fails to match an
1423 ** ORDER BY term, that is OK. Just ignore that column of the index
1424 */
1425 continue;
drhff354e92008-06-25 02:47:57 +00001426 }else if( i==pIdx->nColumn ){
1427 /* Index column i is the rowid. All other terms match. */
1428 break;
drh51669862004-12-18 18:40:26 +00001429 }else{
1430 /* If an index column fails to match and is not constrained by ==
1431 ** then the index cannot satisfy the ORDER BY constraint.
1432 */
1433 return 0;
1434 }
1435 }
dan5236ac12009-08-13 07:09:33 +00001436 assert( pIdx->aSortOrder!=0 || iColumn==-1 );
drh85eeb692005-12-21 03:16:42 +00001437 assert( pTerm->sortOrder==0 || pTerm->sortOrder==1 );
drhcc192542006-12-20 03:24:19 +00001438 assert( iSortOrder==0 || iSortOrder==1 );
1439 termSortOrder = iSortOrder ^ pTerm->sortOrder;
drh51669862004-12-18 18:40:26 +00001440 if( i>nEqCol ){
drh85eeb692005-12-21 03:16:42 +00001441 if( termSortOrder!=sortOrder ){
drh51669862004-12-18 18:40:26 +00001442 /* Indices can only be used if all ORDER BY terms past the
1443 ** equality constraints are all either DESC or ASC. */
1444 return 0;
1445 }
1446 }else{
drh85eeb692005-12-21 03:16:42 +00001447 sortOrder = termSortOrder;
drh51669862004-12-18 18:40:26 +00001448 }
1449 j++;
1450 pTerm++;
drh7b4fc6a2007-02-06 13:26:32 +00001451 if( iColumn<0 && !referencesOtherTables(pOrderBy, pMaskSet, j, base) ){
drhcc192542006-12-20 03:24:19 +00001452 /* If the indexed column is the primary key and everything matches
drh7b4fc6a2007-02-06 13:26:32 +00001453 ** so far and none of the ORDER BY terms to the right reference other
1454 ** tables in the join, then we are assured that the index can be used
1455 ** to sort because the primary key is unique and so none of the other
1456 ** columns will make any difference
drhcc192542006-12-20 03:24:19 +00001457 */
1458 j = nTerm;
1459 }
drh51669862004-12-18 18:40:26 +00001460 }
1461
drhcc192542006-12-20 03:24:19 +00001462 *pbRev = sortOrder!=0;
drh8718f522005-08-13 16:13:04 +00001463 if( j>=nTerm ){
drhcc192542006-12-20 03:24:19 +00001464 /* All terms of the ORDER BY clause are covered by this index so
1465 ** this index can be used for sorting. */
1466 return 1;
1467 }
drh7b4fc6a2007-02-06 13:26:32 +00001468 if( pIdx->onError!=OE_None && i==pIdx->nColumn
1469 && !referencesOtherTables(pOrderBy, pMaskSet, j, base) ){
drhcc192542006-12-20 03:24:19 +00001470 /* All terms of this index match some prefix of the ORDER BY clause
drh7b4fc6a2007-02-06 13:26:32 +00001471 ** and the index is UNIQUE and no terms on the tail of the ORDER BY
1472 ** clause reference other tables in a join. If this is all true then
1473 ** the order by clause is superfluous. */
drh51669862004-12-18 18:40:26 +00001474 return 1;
1475 }
1476 return 0;
1477}
1478
1479/*
drhb6fb62d2005-09-20 08:47:20 +00001480** Prepare a crude estimate of the logarithm of the input value.
drh28c4cf42005-07-27 20:41:43 +00001481** The results need not be exact. This is only used for estimating
drh909626d2008-05-30 14:58:37 +00001482** the total cost of performing operations with O(logN) or O(NlogN)
drh28c4cf42005-07-27 20:41:43 +00001483** complexity. Because N is just a guess, it is no great tragedy if
1484** logN is a little off.
drh28c4cf42005-07-27 20:41:43 +00001485*/
1486static double estLog(double N){
drhb37df7b2005-10-13 02:09:49 +00001487 double logN = 1;
1488 double x = 10;
drh28c4cf42005-07-27 20:41:43 +00001489 while( N>x ){
drhb37df7b2005-10-13 02:09:49 +00001490 logN += 1;
drh28c4cf42005-07-27 20:41:43 +00001491 x *= 10;
1492 }
1493 return logN;
1494}
1495
drh6d209d82006-06-27 01:54:26 +00001496/*
1497** Two routines for printing the content of an sqlite3_index_info
1498** structure. Used for testing and debugging only. If neither
1499** SQLITE_TEST or SQLITE_DEBUG are defined, then these routines
1500** are no-ops.
1501*/
drh77a2a5e2007-04-06 01:04:39 +00001502#if !defined(SQLITE_OMIT_VIRTUALTABLE) && defined(SQLITE_DEBUG)
drh6d209d82006-06-27 01:54:26 +00001503static void TRACE_IDX_INPUTS(sqlite3_index_info *p){
1504 int i;
mlcreech3a00f902008-03-04 17:45:01 +00001505 if( !sqlite3WhereTrace ) return;
drh6d209d82006-06-27 01:54:26 +00001506 for(i=0; i<p->nConstraint; i++){
1507 sqlite3DebugPrintf(" constraint[%d]: col=%d termid=%d op=%d usabled=%d\n",
1508 i,
1509 p->aConstraint[i].iColumn,
1510 p->aConstraint[i].iTermOffset,
1511 p->aConstraint[i].op,
1512 p->aConstraint[i].usable);
1513 }
1514 for(i=0; i<p->nOrderBy; i++){
1515 sqlite3DebugPrintf(" orderby[%d]: col=%d desc=%d\n",
1516 i,
1517 p->aOrderBy[i].iColumn,
1518 p->aOrderBy[i].desc);
1519 }
1520}
1521static void TRACE_IDX_OUTPUTS(sqlite3_index_info *p){
1522 int i;
mlcreech3a00f902008-03-04 17:45:01 +00001523 if( !sqlite3WhereTrace ) return;
drh6d209d82006-06-27 01:54:26 +00001524 for(i=0; i<p->nConstraint; i++){
1525 sqlite3DebugPrintf(" usage[%d]: argvIdx=%d omit=%d\n",
1526 i,
1527 p->aConstraintUsage[i].argvIndex,
1528 p->aConstraintUsage[i].omit);
1529 }
1530 sqlite3DebugPrintf(" idxNum=%d\n", p->idxNum);
1531 sqlite3DebugPrintf(" idxStr=%s\n", p->idxStr);
1532 sqlite3DebugPrintf(" orderByConsumed=%d\n", p->orderByConsumed);
1533 sqlite3DebugPrintf(" estimatedCost=%g\n", p->estimatedCost);
1534}
1535#else
1536#define TRACE_IDX_INPUTS(A)
1537#define TRACE_IDX_OUTPUTS(A)
1538#endif
1539
danielk19771d461462009-04-21 09:02:45 +00001540/*
1541** Required because bestIndex() is called by bestOrClauseIndex()
1542*/
1543static void bestIndex(
1544 Parse*, WhereClause*, struct SrcList_item*, Bitmask, ExprList*, WhereCost*);
1545
1546/*
1547** This routine attempts to find an scanning strategy that can be used
1548** to optimize an 'OR' expression that is part of a WHERE clause.
1549**
1550** The table associated with FROM clause term pSrc may be either a
1551** regular B-Tree table or a virtual table.
1552*/
1553static void bestOrClauseIndex(
1554 Parse *pParse, /* The parsing context */
1555 WhereClause *pWC, /* The WHERE clause */
1556 struct SrcList_item *pSrc, /* The FROM clause term to search */
1557 Bitmask notReady, /* Mask of cursors that are not available */
1558 ExprList *pOrderBy, /* The ORDER BY clause */
1559 WhereCost *pCost /* Lowest cost query plan */
1560){
1561#ifndef SQLITE_OMIT_OR_OPTIMIZATION
1562 const int iCur = pSrc->iCursor; /* The cursor of the table to be accessed */
1563 const Bitmask maskSrc = getMask(pWC->pMaskSet, iCur); /* Bitmask for pSrc */
1564 WhereTerm * const pWCEnd = &pWC->a[pWC->nTerm]; /* End of pWC->a[] */
1565 WhereTerm *pTerm; /* A single term of the WHERE clause */
1566
1567 /* Search the WHERE clause terms for a usable WO_OR term. */
1568 for(pTerm=pWC->a; pTerm<pWCEnd; pTerm++){
1569 if( pTerm->eOperator==WO_OR
1570 && ((pTerm->prereqAll & ~maskSrc) & notReady)==0
1571 && (pTerm->u.pOrInfo->indexable & maskSrc)!=0
1572 ){
1573 WhereClause * const pOrWC = &pTerm->u.pOrInfo->wc;
1574 WhereTerm * const pOrWCEnd = &pOrWC->a[pOrWC->nTerm];
1575 WhereTerm *pOrTerm;
1576 int flags = WHERE_MULTI_OR;
1577 double rTotal = 0;
1578 double nRow = 0;
dan5236ac12009-08-13 07:09:33 +00001579 Bitmask used = 0;
danielk19771d461462009-04-21 09:02:45 +00001580
1581 for(pOrTerm=pOrWC->a; pOrTerm<pOrWCEnd; pOrTerm++){
1582 WhereCost sTermCost;
1583 WHERETRACE(("... Multi-index OR testing for term %d of %d....\n",
1584 (pOrTerm - pOrWC->a), (pTerm - pWC->a)
1585 ));
1586 if( pOrTerm->eOperator==WO_AND ){
1587 WhereClause *pAndWC = &pOrTerm->u.pAndInfo->wc;
1588 bestIndex(pParse, pAndWC, pSrc, notReady, 0, &sTermCost);
1589 }else if( pOrTerm->leftCursor==iCur ){
1590 WhereClause tempWC;
1591 tempWC.pParse = pWC->pParse;
1592 tempWC.pMaskSet = pWC->pMaskSet;
1593 tempWC.op = TK_AND;
1594 tempWC.a = pOrTerm;
1595 tempWC.nTerm = 1;
1596 bestIndex(pParse, &tempWC, pSrc, notReady, 0, &sTermCost);
1597 }else{
1598 continue;
1599 }
1600 rTotal += sTermCost.rCost;
1601 nRow += sTermCost.nRow;
dan5236ac12009-08-13 07:09:33 +00001602 used |= sTermCost.used;
danielk19771d461462009-04-21 09:02:45 +00001603 if( rTotal>=pCost->rCost ) break;
1604 }
1605
1606 /* If there is an ORDER BY clause, increase the scan cost to account
1607 ** for the cost of the sort. */
1608 if( pOrderBy!=0 ){
1609 rTotal += nRow*estLog(nRow);
1610 WHERETRACE(("... sorting increases OR cost to %.9g\n", rTotal));
1611 }
1612
1613 /* If the cost of scanning using this OR term for optimization is
1614 ** less than the current cost stored in pCost, replace the contents
1615 ** of pCost. */
1616 WHERETRACE(("... multi-index OR cost=%.9g nrow=%.9g\n", rTotal, nRow));
1617 if( rTotal<pCost->rCost ){
1618 pCost->rCost = rTotal;
1619 pCost->nRow = nRow;
dan5236ac12009-08-13 07:09:33 +00001620 pCost->used = used;
danielk19771d461462009-04-21 09:02:45 +00001621 pCost->plan.wsFlags = flags;
1622 pCost->plan.u.pTerm = pTerm;
1623 }
1624 }
1625 }
1626#endif /* SQLITE_OMIT_OR_OPTIMIZATION */
1627}
1628
drh9eff6162006-06-12 21:59:13 +00001629#ifndef SQLITE_OMIT_VIRTUALTABLE
1630/*
danielk19771d461462009-04-21 09:02:45 +00001631** Allocate and populate an sqlite3_index_info structure. It is the
1632** responsibility of the caller to eventually release the structure
1633** by passing the pointer returned by this function to sqlite3_free().
1634*/
1635static sqlite3_index_info *allocateIndexInfo(
1636 Parse *pParse,
1637 WhereClause *pWC,
1638 struct SrcList_item *pSrc,
1639 ExprList *pOrderBy
1640){
1641 int i, j;
1642 int nTerm;
1643 struct sqlite3_index_constraint *pIdxCons;
1644 struct sqlite3_index_orderby *pIdxOrderBy;
1645 struct sqlite3_index_constraint_usage *pUsage;
1646 WhereTerm *pTerm;
1647 int nOrderBy;
1648 sqlite3_index_info *pIdxInfo;
1649
1650 WHERETRACE(("Recomputing index info for %s...\n", pSrc->pTab->zName));
1651
1652 /* Count the number of possible WHERE clause constraints referring
1653 ** to this virtual table */
1654 for(i=nTerm=0, pTerm=pWC->a; i<pWC->nTerm; i++, pTerm++){
1655 if( pTerm->leftCursor != pSrc->iCursor ) continue;
1656 assert( (pTerm->eOperator&(pTerm->eOperator-1))==0 );
1657 testcase( pTerm->eOperator==WO_IN );
1658 testcase( pTerm->eOperator==WO_ISNULL );
1659 if( pTerm->eOperator & (WO_IN|WO_ISNULL) ) continue;
1660 nTerm++;
1661 }
1662
1663 /* If the ORDER BY clause contains only columns in the current
1664 ** virtual table then allocate space for the aOrderBy part of
1665 ** the sqlite3_index_info structure.
1666 */
1667 nOrderBy = 0;
1668 if( pOrderBy ){
1669 for(i=0; i<pOrderBy->nExpr; i++){
1670 Expr *pExpr = pOrderBy->a[i].pExpr;
1671 if( pExpr->op!=TK_COLUMN || pExpr->iTable!=pSrc->iCursor ) break;
1672 }
1673 if( i==pOrderBy->nExpr ){
1674 nOrderBy = pOrderBy->nExpr;
1675 }
1676 }
1677
1678 /* Allocate the sqlite3_index_info structure
1679 */
1680 pIdxInfo = sqlite3DbMallocZero(pParse->db, sizeof(*pIdxInfo)
1681 + (sizeof(*pIdxCons) + sizeof(*pUsage))*nTerm
1682 + sizeof(*pIdxOrderBy)*nOrderBy );
1683 if( pIdxInfo==0 ){
1684 sqlite3ErrorMsg(pParse, "out of memory");
1685 /* (double)0 In case of SQLITE_OMIT_FLOATING_POINT... */
1686 return 0;
1687 }
1688
1689 /* Initialize the structure. The sqlite3_index_info structure contains
1690 ** many fields that are declared "const" to prevent xBestIndex from
1691 ** changing them. We have to do some funky casting in order to
1692 ** initialize those fields.
1693 */
1694 pIdxCons = (struct sqlite3_index_constraint*)&pIdxInfo[1];
1695 pIdxOrderBy = (struct sqlite3_index_orderby*)&pIdxCons[nTerm];
1696 pUsage = (struct sqlite3_index_constraint_usage*)&pIdxOrderBy[nOrderBy];
1697 *(int*)&pIdxInfo->nConstraint = nTerm;
1698 *(int*)&pIdxInfo->nOrderBy = nOrderBy;
1699 *(struct sqlite3_index_constraint**)&pIdxInfo->aConstraint = pIdxCons;
1700 *(struct sqlite3_index_orderby**)&pIdxInfo->aOrderBy = pIdxOrderBy;
1701 *(struct sqlite3_index_constraint_usage**)&pIdxInfo->aConstraintUsage =
1702 pUsage;
1703
1704 for(i=j=0, pTerm=pWC->a; i<pWC->nTerm; i++, pTerm++){
1705 if( pTerm->leftCursor != pSrc->iCursor ) continue;
1706 assert( (pTerm->eOperator&(pTerm->eOperator-1))==0 );
1707 testcase( pTerm->eOperator==WO_IN );
1708 testcase( pTerm->eOperator==WO_ISNULL );
1709 if( pTerm->eOperator & (WO_IN|WO_ISNULL) ) continue;
1710 pIdxCons[j].iColumn = pTerm->u.leftColumn;
1711 pIdxCons[j].iTermOffset = i;
1712 pIdxCons[j].op = (u8)pTerm->eOperator;
1713 /* The direct assignment in the previous line is possible only because
1714 ** the WO_ and SQLITE_INDEX_CONSTRAINT_ codes are identical. The
1715 ** following asserts verify this fact. */
1716 assert( WO_EQ==SQLITE_INDEX_CONSTRAINT_EQ );
1717 assert( WO_LT==SQLITE_INDEX_CONSTRAINT_LT );
1718 assert( WO_LE==SQLITE_INDEX_CONSTRAINT_LE );
1719 assert( WO_GT==SQLITE_INDEX_CONSTRAINT_GT );
1720 assert( WO_GE==SQLITE_INDEX_CONSTRAINT_GE );
1721 assert( WO_MATCH==SQLITE_INDEX_CONSTRAINT_MATCH );
1722 assert( pTerm->eOperator & (WO_EQ|WO_LT|WO_LE|WO_GT|WO_GE|WO_MATCH) );
1723 j++;
1724 }
1725 for(i=0; i<nOrderBy; i++){
1726 Expr *pExpr = pOrderBy->a[i].pExpr;
1727 pIdxOrderBy[i].iColumn = pExpr->iColumn;
1728 pIdxOrderBy[i].desc = pOrderBy->a[i].sortOrder;
1729 }
1730
1731 return pIdxInfo;
1732}
1733
1734/*
1735** The table object reference passed as the second argument to this function
1736** must represent a virtual table. This function invokes the xBestIndex()
1737** method of the virtual table with the sqlite3_index_info pointer passed
1738** as the argument.
1739**
1740** If an error occurs, pParse is populated with an error message and a
1741** non-zero value is returned. Otherwise, 0 is returned and the output
1742** part of the sqlite3_index_info structure is left populated.
1743**
1744** Whether or not an error is returned, it is the responsibility of the
1745** caller to eventually free p->idxStr if p->needToFreeIdxStr indicates
1746** that this is required.
1747*/
1748static int vtabBestIndex(Parse *pParse, Table *pTab, sqlite3_index_info *p){
danielk1977595a5232009-07-24 17:58:53 +00001749 sqlite3_vtab *pVtab = sqlite3GetVTable(pParse->db, pTab)->pVtab;
danielk19771d461462009-04-21 09:02:45 +00001750 int i;
1751 int rc;
1752
1753 (void)sqlite3SafetyOff(pParse->db);
1754 WHERETRACE(("xBestIndex for %s\n", pTab->zName));
1755 TRACE_IDX_INPUTS(p);
1756 rc = pVtab->pModule->xBestIndex(pVtab, p);
1757 TRACE_IDX_OUTPUTS(p);
1758 (void)sqlite3SafetyOn(pParse->db);
1759
1760 if( rc!=SQLITE_OK ){
1761 if( rc==SQLITE_NOMEM ){
1762 pParse->db->mallocFailed = 1;
1763 }else if( !pVtab->zErrMsg ){
1764 sqlite3ErrorMsg(pParse, "%s", sqlite3ErrStr(rc));
1765 }else{
1766 sqlite3ErrorMsg(pParse, "%s", pVtab->zErrMsg);
1767 }
1768 }
1769 sqlite3DbFree(pParse->db, pVtab->zErrMsg);
1770 pVtab->zErrMsg = 0;
1771
1772 for(i=0; i<p->nConstraint; i++){
1773 if( !p->aConstraint[i].usable && p->aConstraintUsage[i].argvIndex>0 ){
1774 sqlite3ErrorMsg(pParse,
1775 "table %s: xBestIndex returned an invalid plan", pTab->zName);
1776 }
1777 }
1778
1779 return pParse->nErr;
1780}
1781
1782
1783/*
drh7f375902006-06-13 17:38:59 +00001784** Compute the best index for a virtual table.
1785**
1786** The best index is computed by the xBestIndex method of the virtual
1787** table module. This routine is really just a wrapper that sets up
1788** the sqlite3_index_info structure that is used to communicate with
1789** xBestIndex.
1790**
1791** In a join, this routine might be called multiple times for the
1792** same virtual table. The sqlite3_index_info structure is created
1793** and initialized on the first invocation and reused on all subsequent
1794** invocations. The sqlite3_index_info structure is also used when
1795** code is generated to access the virtual table. The whereInfoDelete()
1796** routine takes care of freeing the sqlite3_index_info structure after
1797** everybody has finished with it.
drh9eff6162006-06-12 21:59:13 +00001798*/
danielk19771d461462009-04-21 09:02:45 +00001799static void bestVirtualIndex(
1800 Parse *pParse, /* The parsing context */
1801 WhereClause *pWC, /* The WHERE clause */
1802 struct SrcList_item *pSrc, /* The FROM clause term to search */
1803 Bitmask notReady, /* Mask of cursors that are not available */
1804 ExprList *pOrderBy, /* The order by clause */
1805 WhereCost *pCost, /* Lowest cost query plan */
1806 sqlite3_index_info **ppIdxInfo /* Index information passed to xBestIndex */
drh9eff6162006-06-12 21:59:13 +00001807){
1808 Table *pTab = pSrc->pTab;
1809 sqlite3_index_info *pIdxInfo;
1810 struct sqlite3_index_constraint *pIdxCons;
drh9eff6162006-06-12 21:59:13 +00001811 struct sqlite3_index_constraint_usage *pUsage;
1812 WhereTerm *pTerm;
1813 int i, j;
1814 int nOrderBy;
1815
danielk19776eacd282009-04-29 11:50:53 +00001816 /* Make sure wsFlags is initialized to some sane value. Otherwise, if the
1817 ** malloc in allocateIndexInfo() fails and this function returns leaving
1818 ** wsFlags in an uninitialized state, the caller may behave unpredictably.
1819 */
drh6a863cd2009-05-06 18:42:21 +00001820 memset(pCost, 0, sizeof(*pCost));
danielk19776eacd282009-04-29 11:50:53 +00001821 pCost->plan.wsFlags = WHERE_VIRTUALTABLE;
1822
drh9eff6162006-06-12 21:59:13 +00001823 /* If the sqlite3_index_info structure has not been previously
danielk19771d461462009-04-21 09:02:45 +00001824 ** allocated and initialized, then allocate and initialize it now.
drh9eff6162006-06-12 21:59:13 +00001825 */
1826 pIdxInfo = *ppIdxInfo;
1827 if( pIdxInfo==0 ){
danielk19771d461462009-04-21 09:02:45 +00001828 *ppIdxInfo = pIdxInfo = allocateIndexInfo(pParse, pWC, pSrc, pOrderBy);
drh9eff6162006-06-12 21:59:13 +00001829 }
danielk1977732dc552009-04-21 17:23:04 +00001830 if( pIdxInfo==0 ){
1831 return;
1832 }
drh9eff6162006-06-12 21:59:13 +00001833
drh7f375902006-06-13 17:38:59 +00001834 /* At this point, the sqlite3_index_info structure that pIdxInfo points
1835 ** to will have been initialized, either during the current invocation or
1836 ** during some prior invocation. Now we just have to customize the
1837 ** details of pIdxInfo for the current invocation and pass it to
1838 ** xBestIndex.
1839 */
1840
danielk1977935ed5e2007-03-30 09:13:13 +00001841 /* The module name must be defined. Also, by this point there must
1842 ** be a pointer to an sqlite3_vtab structure. Otherwise
1843 ** sqlite3ViewGetColumnNames() would have picked up the error.
1844 */
drh9eff6162006-06-12 21:59:13 +00001845 assert( pTab->azModuleArg && pTab->azModuleArg[0] );
danielk1977595a5232009-07-24 17:58:53 +00001846 assert( sqlite3GetVTable(pParse->db, pTab) );
drh9eff6162006-06-12 21:59:13 +00001847
1848 /* Set the aConstraint[].usable fields and initialize all
drh7f375902006-06-13 17:38:59 +00001849 ** output variables to zero.
1850 **
1851 ** aConstraint[].usable is true for constraints where the right-hand
1852 ** side contains only references to tables to the left of the current
1853 ** table. In other words, if the constraint is of the form:
1854 **
1855 ** column = expr
1856 **
1857 ** and we are evaluating a join, then the constraint on column is
1858 ** only valid if all tables referenced in expr occur to the left
1859 ** of the table containing column.
1860 **
1861 ** The aConstraints[] array contains entries for all constraints
1862 ** on the current table. That way we only have to compute it once
1863 ** even though we might try to pick the best index multiple times.
1864 ** For each attempt at picking an index, the order of tables in the
1865 ** join might be different so we have to recompute the usable flag
1866 ** each time.
drh9eff6162006-06-12 21:59:13 +00001867 */
1868 pIdxCons = *(struct sqlite3_index_constraint**)&pIdxInfo->aConstraint;
1869 pUsage = pIdxInfo->aConstraintUsage;
1870 for(i=0; i<pIdxInfo->nConstraint; i++, pIdxCons++){
1871 j = pIdxCons->iTermOffset;
1872 pTerm = &pWC->a[j];
dan5236ac12009-08-13 07:09:33 +00001873 pIdxCons->usable = (pTerm->prereqRight&notReady) ? 0 : 1;
drh9eff6162006-06-12 21:59:13 +00001874 }
1875 memset(pUsage, 0, sizeof(pUsage[0])*pIdxInfo->nConstraint);
drh4be8b512006-06-13 23:51:34 +00001876 if( pIdxInfo->needToFreeIdxStr ){
1877 sqlite3_free(pIdxInfo->idxStr);
1878 }
1879 pIdxInfo->idxStr = 0;
1880 pIdxInfo->idxNum = 0;
1881 pIdxInfo->needToFreeIdxStr = 0;
drh9eff6162006-06-12 21:59:13 +00001882 pIdxInfo->orderByConsumed = 0;
shanefbd60f82009-02-04 03:59:25 +00001883 /* ((double)2) In case of SQLITE_OMIT_FLOATING_POINT... */
1884 pIdxInfo->estimatedCost = SQLITE_BIG_DBL / ((double)2);
drh9eff6162006-06-12 21:59:13 +00001885 nOrderBy = pIdxInfo->nOrderBy;
danielk19771d461462009-04-21 09:02:45 +00001886 if( !pOrderBy ){
1887 pIdxInfo->nOrderBy = 0;
drh9eff6162006-06-12 21:59:13 +00001888 }
danielk197774cdba42006-06-19 12:02:58 +00001889
danielk19771d461462009-04-21 09:02:45 +00001890 if( vtabBestIndex(pParse, pTab, pIdxInfo) ){
1891 return;
danielk197739359dc2008-03-17 09:36:44 +00001892 }
1893
dan5236ac12009-08-13 07:09:33 +00001894 pIdxCons = *(struct sqlite3_index_constraint**)&pIdxInfo->aConstraint;
1895 for(i=0; i<pIdxInfo->nConstraint; i++){
1896 if( pUsage[i].argvIndex>0 ){
1897 pCost->used |= pWC->a[pIdxCons[i].iTermOffset].prereqRight;
1898 }
1899 }
1900
danielk19771d461462009-04-21 09:02:45 +00001901 /* The cost is not allowed to be larger than SQLITE_BIG_DBL (the
1902 ** inital value of lowestCost in this loop. If it is, then the
1903 ** (cost<lowestCost) test below will never be true.
1904 **
1905 ** Use "(double)2" instead of "2.0" in case OMIT_FLOATING_POINT
1906 ** is defined.
1907 */
1908 if( (SQLITE_BIG_DBL/((double)2))<pIdxInfo->estimatedCost ){
1909 pCost->rCost = (SQLITE_BIG_DBL/((double)2));
1910 }else{
1911 pCost->rCost = pIdxInfo->estimatedCost;
1912 }
danielk19771d461462009-04-21 09:02:45 +00001913 pCost->plan.u.pVtabIdx = pIdxInfo;
drh5901b572009-06-10 19:33:28 +00001914 if( pIdxInfo->orderByConsumed ){
danielk19771d461462009-04-21 09:02:45 +00001915 pCost->plan.wsFlags |= WHERE_ORDERBY;
1916 }
1917 pCost->plan.nEq = 0;
1918 pIdxInfo->nOrderBy = nOrderBy;
1919
1920 /* Try to find a more efficient access pattern by using multiple indexes
1921 ** to optimize an OR expression within the WHERE clause.
1922 */
1923 bestOrClauseIndex(pParse, pWC, pSrc, notReady, pOrderBy, pCost);
drh9eff6162006-06-12 21:59:13 +00001924}
1925#endif /* SQLITE_OMIT_VIRTUALTABLE */
1926
drh28c4cf42005-07-27 20:41:43 +00001927/*
dan02fa4692009-08-17 17:06:58 +00001928** Argument pIdx is a pointer to an index structure that has an array of
1929** SQLITE_INDEX_SAMPLES evenly spaced samples of the first indexed column
1930** stored in Index.aSample. The domain of values stored in said column
1931** may be thought of as divided into (SQLITE_INDEX_SAMPLES+1) regions.
1932** Region 0 contains all values smaller than the first sample value. Region
1933** 1 contains values larger than or equal to the value of the first sample,
1934** but smaller than the value of the second. And so on.
1935**
1936** If successful, this function determines which of the regions value
drh98cdf622009-08-20 18:14:42 +00001937** pVal lies in, sets *piRegion to the region index (a value between 0
1938** and SQLITE_INDEX_SAMPLES+1, inclusive) and returns SQLITE_OK.
dan02fa4692009-08-17 17:06:58 +00001939** Or, if an OOM occurs while converting text values between encodings,
drh98cdf622009-08-20 18:14:42 +00001940** SQLITE_NOMEM is returned and *piRegion is undefined.
dan02fa4692009-08-17 17:06:58 +00001941*/
dan69188d92009-08-19 08:18:32 +00001942#ifdef SQLITE_ENABLE_STAT2
dan02fa4692009-08-17 17:06:58 +00001943static int whereRangeRegion(
1944 Parse *pParse, /* Database connection */
1945 Index *pIdx, /* Index to consider domain of */
1946 sqlite3_value *pVal, /* Value to consider */
1947 int *piRegion /* OUT: Region of domain in which value lies */
1948){
drhdaf4a9f2009-08-20 20:05:55 +00001949 if( ALWAYS(pVal) ){
dan02fa4692009-08-17 17:06:58 +00001950 IndexSample *aSample = pIdx->aSample;
1951 int i = 0;
1952 int eType = sqlite3_value_type(pVal);
1953
1954 if( eType==SQLITE_INTEGER || eType==SQLITE_FLOAT ){
1955 double r = sqlite3_value_double(pVal);
1956 for(i=0; i<SQLITE_INDEX_SAMPLES; i++){
1957 if( aSample[i].eType==SQLITE_NULL ) continue;
1958 if( aSample[i].eType>=SQLITE_TEXT || aSample[i].u.r>r ) break;
1959 }
drhcdaca552009-08-20 13:45:07 +00001960 }else{
dan02fa4692009-08-17 17:06:58 +00001961 sqlite3 *db = pParse->db;
1962 CollSeq *pColl;
1963 const u8 *z;
1964 int n;
drhcdaca552009-08-20 13:45:07 +00001965
1966 /* pVal comes from sqlite3ValueFromExpr() so the type cannot be NULL */
1967 assert( eType==SQLITE_TEXT || eType==SQLITE_BLOB );
1968
dan02fa4692009-08-17 17:06:58 +00001969 if( eType==SQLITE_BLOB ){
1970 z = (const u8 *)sqlite3_value_blob(pVal);
1971 pColl = db->pDfltColl;
dane275dc32009-08-18 16:24:58 +00001972 assert( pColl->enc==SQLITE_UTF8 );
dan02fa4692009-08-17 17:06:58 +00001973 }else{
drh9aeda792009-08-20 02:34:15 +00001974 pColl = sqlite3GetCollSeq(db, SQLITE_UTF8, 0, *pIdx->azColl);
1975 if( pColl==0 ){
1976 sqlite3ErrorMsg(pParse, "no such collation sequence: %s",
1977 *pIdx->azColl);
dane275dc32009-08-18 16:24:58 +00001978 return SQLITE_ERROR;
1979 }
dan02fa4692009-08-17 17:06:58 +00001980 z = (const u8 *)sqlite3ValueText(pVal, pColl->enc);
dane275dc32009-08-18 16:24:58 +00001981 if( !z ){
1982 return SQLITE_NOMEM;
1983 }
dan02fa4692009-08-17 17:06:58 +00001984 assert( z && pColl && pColl->xCmp );
1985 }
1986 n = sqlite3ValueBytes(pVal, pColl->enc);
1987
1988 for(i=0; i<SQLITE_INDEX_SAMPLES; i++){
dane275dc32009-08-18 16:24:58 +00001989 int r;
dan02fa4692009-08-17 17:06:58 +00001990 int eSampletype = aSample[i].eType;
1991 if( eSampletype==SQLITE_NULL || eSampletype<eType ) continue;
1992 if( (eSampletype!=eType) ) break;
dane83c4f32009-09-21 16:34:24 +00001993#ifndef SQLITE_OMIT_UTF16
1994 if( pColl->enc!=SQLITE_UTF8 ){
dane275dc32009-08-18 16:24:58 +00001995 int nSample;
1996 char *zSample = sqlite3Utf8to16(
dan02fa4692009-08-17 17:06:58 +00001997 db, pColl->enc, aSample[i].u.z, aSample[i].nByte, &nSample
1998 );
dane275dc32009-08-18 16:24:58 +00001999 if( !zSample ){
2000 assert( db->mallocFailed );
2001 return SQLITE_NOMEM;
2002 }
2003 r = pColl->xCmp(pColl->pUser, nSample, zSample, n, z);
2004 sqlite3DbFree(db, zSample);
dane83c4f32009-09-21 16:34:24 +00002005 }else
2006#endif
2007 {
2008 r = pColl->xCmp(pColl->pUser, aSample[i].nByte, aSample[i].u.z, n, z);
dan02fa4692009-08-17 17:06:58 +00002009 }
dane275dc32009-08-18 16:24:58 +00002010 if( r>0 ) break;
dan02fa4692009-08-17 17:06:58 +00002011 }
2012 }
2013
drha8f57612009-08-25 16:28:14 +00002014 assert( i>=0 && i<=SQLITE_INDEX_SAMPLES );
dan02fa4692009-08-17 17:06:58 +00002015 *piRegion = i;
2016 }
2017 return SQLITE_OK;
2018}
dan69188d92009-08-19 08:18:32 +00002019#endif /* #ifdef SQLITE_ENABLE_STAT2 */
dan02fa4692009-08-17 17:06:58 +00002020
2021/*
dan937d0de2009-10-15 18:35:38 +00002022** If expression pExpr represents a literal value, set *pp to point to
2023** an sqlite3_value structure containing the same value, with affinity
2024** aff applied to it, before returning. It is the responsibility of the
2025** caller to eventually release this structure by passing it to
2026** sqlite3ValueFree().
2027**
2028** If the current parse is a recompile (sqlite3Reprepare()) and pExpr
2029** is an SQL variable that currently has a non-NULL value bound to it,
2030** create an sqlite3_value structure containing this value, again with
2031** affinity aff applied to it, instead.
2032**
2033** If neither of the above apply, set *pp to NULL.
2034**
2035** If an error occurs, return an error code. Otherwise, SQLITE_OK.
2036*/
danf7b0b0a2009-10-19 15:52:32 +00002037#ifdef SQLITE_ENABLE_STAT2
dan937d0de2009-10-15 18:35:38 +00002038static int valueFromExpr(
2039 Parse *pParse,
2040 Expr *pExpr,
2041 u8 aff,
2042 sqlite3_value **pp
2043){
drhb4138de2009-10-19 22:41:06 +00002044 /* The evalConstExpr() function will have already converted any TK_VARIABLE
2045 ** expression involved in an comparison into a TK_REGISTER. */
2046 assert( pExpr->op!=TK_VARIABLE );
2047 if( pExpr->op==TK_REGISTER && pExpr->op2==TK_VARIABLE ){
dan937d0de2009-10-15 18:35:38 +00002048 int iVar = pExpr->iColumn;
dan1d2ce4f2009-10-19 18:11:09 +00002049 sqlite3VdbeSetVarmask(pParse->pVdbe, iVar);
dan937d0de2009-10-15 18:35:38 +00002050 *pp = sqlite3VdbeGetValue(pParse->pReprepare, iVar, aff);
2051 return SQLITE_OK;
2052 }
2053 return sqlite3ValueFromExpr(pParse->db, pExpr, SQLITE_UTF8, aff, pp);
2054}
danf7b0b0a2009-10-19 15:52:32 +00002055#endif
dan937d0de2009-10-15 18:35:38 +00002056
2057/*
dan02fa4692009-08-17 17:06:58 +00002058** This function is used to estimate the number of rows that will be visited
2059** by scanning an index for a range of values. The range may have an upper
2060** bound, a lower bound, or both. The WHERE clause terms that set the upper
2061** and lower bounds are represented by pLower and pUpper respectively. For
2062** example, assuming that index p is on t1(a):
2063**
2064** ... FROM t1 WHERE a > ? AND a < ? ...
2065** |_____| |_____|
2066** | |
2067** pLower pUpper
2068**
drh98cdf622009-08-20 18:14:42 +00002069** If either of the upper or lower bound is not present, then NULL is passed in
drhcdaca552009-08-20 13:45:07 +00002070** place of the corresponding WhereTerm.
dan02fa4692009-08-17 17:06:58 +00002071**
2072** The nEq parameter is passed the index of the index column subject to the
2073** range constraint. Or, equivalently, the number of equality constraints
2074** optimized by the proposed index scan. For example, assuming index p is
2075** on t1(a, b), and the SQL query is:
2076**
2077** ... FROM t1 WHERE a = ? AND b > ? AND b < ? ...
2078**
2079** then nEq should be passed the value 1 (as the range restricted column,
2080** b, is the second left-most column of the index). Or, if the query is:
2081**
2082** ... FROM t1 WHERE a > ? AND a < ? ...
2083**
2084** then nEq should be passed 0.
2085**
drh98cdf622009-08-20 18:14:42 +00002086** The returned value is an integer between 1 and 100, inclusive. A return
dan02fa4692009-08-17 17:06:58 +00002087** value of 1 indicates that the proposed range scan is expected to visit
drh98cdf622009-08-20 18:14:42 +00002088** approximately 1/100th (1%) of the rows selected by the nEq equality
2089** constraints (if any). A return value of 100 indicates that it is expected
2090** that the range scan will visit every row (100%) selected by the equality
dan02fa4692009-08-17 17:06:58 +00002091** constraints.
drh98cdf622009-08-20 18:14:42 +00002092**
2093** In the absence of sqlite_stat2 ANALYZE data, each range inequality
2094** reduces the search space by 2/3rds. Hence a single constraint (x>?)
2095** results in a return of 33 and a range constraint (x>? AND x<?) results
2096** in a return of 11.
dan02fa4692009-08-17 17:06:58 +00002097*/
2098static int whereRangeScanEst(
drhcdaca552009-08-20 13:45:07 +00002099 Parse *pParse, /* Parsing & code generating context */
2100 Index *p, /* The index containing the range-compared column; "x" */
2101 int nEq, /* index into p->aCol[] of the range-compared column */
2102 WhereTerm *pLower, /* Lower bound on the range. ex: "x>123" Might be NULL */
2103 WhereTerm *pUpper, /* Upper bound on the range. ex: "x<455" Might be NULL */
2104 int *piEst /* OUT: Return value */
dan02fa4692009-08-17 17:06:58 +00002105){
dan69188d92009-08-19 08:18:32 +00002106 int rc = SQLITE_OK;
2107
2108#ifdef SQLITE_ENABLE_STAT2
dan02fa4692009-08-17 17:06:58 +00002109
2110 if( nEq==0 && p->aSample ){
dan937d0de2009-10-15 18:35:38 +00002111 sqlite3_value *pLowerVal = 0;
2112 sqlite3_value *pUpperVal = 0;
dan02fa4692009-08-17 17:06:58 +00002113 int iEst;
drh011cfca2009-08-25 15:56:51 +00002114 int iLower = 0;
2115 int iUpper = SQLITE_INDEX_SAMPLES;
dan937d0de2009-10-15 18:35:38 +00002116 u8 aff = p->pTable->aCol[p->aiColumn[0]].affinity;
drh98cdf622009-08-20 18:14:42 +00002117
dan02fa4692009-08-17 17:06:58 +00002118 if( pLower ){
2119 Expr *pExpr = pLower->pExpr->pRight;
dan937d0de2009-10-15 18:35:38 +00002120 rc = valueFromExpr(pParse, pExpr, aff, &pLowerVal);
dan02fa4692009-08-17 17:06:58 +00002121 }
drh98cdf622009-08-20 18:14:42 +00002122 if( rc==SQLITE_OK && pUpper ){
dan02fa4692009-08-17 17:06:58 +00002123 Expr *pExpr = pUpper->pExpr->pRight;
dan937d0de2009-10-15 18:35:38 +00002124 rc = valueFromExpr(pParse, pExpr, aff, &pUpperVal);
drh98cdf622009-08-20 18:14:42 +00002125 }
2126
2127 if( rc!=SQLITE_OK || (pLowerVal==0 && pUpperVal==0) ){
2128 sqlite3ValueFree(pLowerVal);
2129 sqlite3ValueFree(pUpperVal);
2130 goto range_est_fallback;
2131 }else if( pLowerVal==0 ){
2132 rc = whereRangeRegion(pParse, p, pUpperVal, &iUpper);
drh011cfca2009-08-25 15:56:51 +00002133 if( pLower ) iLower = iUpper/2;
drh98cdf622009-08-20 18:14:42 +00002134 }else if( pUpperVal==0 ){
2135 rc = whereRangeRegion(pParse, p, pLowerVal, &iLower);
drh011cfca2009-08-25 15:56:51 +00002136 if( pUpper ) iUpper = (iLower + SQLITE_INDEX_SAMPLES + 1)/2;
drh98cdf622009-08-20 18:14:42 +00002137 }else{
2138 rc = whereRangeRegion(pParse, p, pUpperVal, &iUpper);
2139 if( rc==SQLITE_OK ){
2140 rc = whereRangeRegion(pParse, p, pLowerVal, &iLower);
dan02fa4692009-08-17 17:06:58 +00002141 }
2142 }
2143
dan02fa4692009-08-17 17:06:58 +00002144 iEst = iUpper - iLower;
drha8f57612009-08-25 16:28:14 +00002145 testcase( iEst==SQLITE_INDEX_SAMPLES );
2146 assert( iEst<=SQLITE_INDEX_SAMPLES );
2147 if( iEst<1 ){
drh98cdf622009-08-20 18:14:42 +00002148 iEst = 1;
2149 }
dan02fa4692009-08-17 17:06:58 +00002150
2151 sqlite3ValueFree(pLowerVal);
2152 sqlite3ValueFree(pUpperVal);
drh98cdf622009-08-20 18:14:42 +00002153 *piEst = (iEst * 100)/SQLITE_INDEX_SAMPLES;
dan02fa4692009-08-17 17:06:58 +00002154 return rc;
2155 }
drh98cdf622009-08-20 18:14:42 +00002156range_est_fallback:
drh3f022182009-09-09 16:10:50 +00002157#else
2158 UNUSED_PARAMETER(pParse);
2159 UNUSED_PARAMETER(p);
2160 UNUSED_PARAMETER(nEq);
dan69188d92009-08-19 08:18:32 +00002161#endif
dan02fa4692009-08-17 17:06:58 +00002162 assert( pLower || pUpper );
drh98cdf622009-08-20 18:14:42 +00002163 if( pLower && pUpper ){
2164 *piEst = 11;
2165 }else{
2166 *piEst = 33;
2167 }
dan02fa4692009-08-17 17:06:58 +00002168 return rc;
2169}
2170
2171
2172/*
drh111a6a72008-12-21 03:51:16 +00002173** Find the query plan for accessing a particular table. Write the
2174** best query plan and its cost into the WhereCost object supplied as the
2175** last parameter.
drh51147ba2005-07-23 22:59:55 +00002176**
drh111a6a72008-12-21 03:51:16 +00002177** The lowest cost plan wins. The cost is an estimate of the amount of
2178** CPU and disk I/O need to process the request using the selected plan.
drh51147ba2005-07-23 22:59:55 +00002179** Factors that influence cost include:
2180**
2181** * The estimated number of rows that will be retrieved. (The
2182** fewer the better.)
2183**
2184** * Whether or not sorting must occur.
2185**
2186** * Whether or not there must be separate lookups in the
2187** index and in the main table.
2188**
danielk1977e2d7b242009-02-23 17:33:49 +00002189** If there was an INDEXED BY clause (pSrc->pIndex) attached to the table in
2190** the SQL statement, then this function only considers plans using the
drh296a4832009-03-22 20:36:18 +00002191** named index. If no such plan is found, then the returned cost is
2192** SQLITE_BIG_DBL. If a plan is found that uses the named index,
danielk197785574e32008-10-06 05:32:18 +00002193** then the cost is calculated in the usual way.
2194**
danielk1977e2d7b242009-02-23 17:33:49 +00002195** If a NOT INDEXED clause (pSrc->notIndexed!=0) was attached to the table
2196** in the SELECT statement, then no indexes are considered. However, the
2197** selected plan may still take advantage of the tables built-in rowid
danielk197785574e32008-10-06 05:32:18 +00002198** index.
drhfe05af82005-07-21 03:14:59 +00002199*/
danielk19771d461462009-04-21 09:02:45 +00002200static void bestBtreeIndex(
drhfe05af82005-07-21 03:14:59 +00002201 Parse *pParse, /* The parsing context */
2202 WhereClause *pWC, /* The WHERE clause */
2203 struct SrcList_item *pSrc, /* The FROM clause term to search */
2204 Bitmask notReady, /* Mask of cursors that are not available */
drh111a6a72008-12-21 03:51:16 +00002205 ExprList *pOrderBy, /* The ORDER BY clause */
2206 WhereCost *pCost /* Lowest cost query plan */
drhfe05af82005-07-21 03:14:59 +00002207){
drh51147ba2005-07-23 22:59:55 +00002208 int iCur = pSrc->iCursor; /* The cursor of the table to be accessed */
2209 Index *pProbe; /* An index we are evaluating */
dan5236ac12009-08-13 07:09:33 +00002210 Index *pIdx; /* Copy of pProbe, or zero for IPK index */
2211 int eqTermMask; /* Current mask of valid equality operators */
2212 int idxEqTermMask; /* Index mask of valid equality operators */
drhcdaca552009-08-20 13:45:07 +00002213 Index sPk; /* A fake index object for the primary key */
2214 unsigned int aiRowEstPk[2]; /* The aiRowEst[] value for the sPk index */
2215 int aiColumnPk = -1; /* The aColumn[] value for the sPk index */
2216 int wsFlagMask; /* Allowed flags in pCost->plan.wsFlag */
drhfe05af82005-07-21 03:14:59 +00002217
drhcdaca552009-08-20 13:45:07 +00002218 /* Initialize the cost to a worst-case value */
drh111a6a72008-12-21 03:51:16 +00002219 memset(pCost, 0, sizeof(*pCost));
drh111a6a72008-12-21 03:51:16 +00002220 pCost->rCost = SQLITE_BIG_DBL;
drh51147ba2005-07-23 22:59:55 +00002221
drhc49de5d2007-01-19 01:06:01 +00002222 /* If the pSrc table is the right table of a LEFT JOIN then we may not
2223 ** use an index to satisfy IS NULL constraints on that table. This is
2224 ** because columns might end up being NULL if the table does not match -
2225 ** a circumstance which the index cannot help us discover. Ticket #2177.
2226 */
dan5236ac12009-08-13 07:09:33 +00002227 if( pSrc->jointype & JT_LEFT ){
2228 idxEqTermMask = WO_EQ|WO_IN;
drhc49de5d2007-01-19 01:06:01 +00002229 }else{
dan5236ac12009-08-13 07:09:33 +00002230 idxEqTermMask = WO_EQ|WO_IN|WO_ISNULL;
drhc49de5d2007-01-19 01:06:01 +00002231 }
2232
danielk197785574e32008-10-06 05:32:18 +00002233 if( pSrc->pIndex ){
drhcdaca552009-08-20 13:45:07 +00002234 /* An INDEXED BY clause specifies a particular index to use */
dan5236ac12009-08-13 07:09:33 +00002235 pIdx = pProbe = pSrc->pIndex;
2236 wsFlagMask = ~(WHERE_ROWID_EQ|WHERE_ROWID_RANGE);
2237 eqTermMask = idxEqTermMask;
2238 }else{
drhcdaca552009-08-20 13:45:07 +00002239 /* There is no INDEXED BY clause. Create a fake Index object to
2240 ** represent the primary key */
2241 Index *pFirst; /* Any other index on the table */
2242 memset(&sPk, 0, sizeof(Index));
2243 sPk.nColumn = 1;
2244 sPk.aiColumn = &aiColumnPk;
2245 sPk.aiRowEst = aiRowEstPk;
2246 aiRowEstPk[1] = 1;
2247 sPk.onError = OE_Replace;
2248 sPk.pTable = pSrc->pTab;
2249 pFirst = pSrc->pTab->pIndex;
dan5236ac12009-08-13 07:09:33 +00002250 if( pSrc->notIndexed==0 ){
drhcdaca552009-08-20 13:45:07 +00002251 sPk.pNext = pFirst;
dan5236ac12009-08-13 07:09:33 +00002252 }
drhcdaca552009-08-20 13:45:07 +00002253 /* The aiRowEstPk[0] is an estimate of the total number of rows in the
2254 ** table. Get this information from the ANALYZE information if it is
2255 ** available. If not available, assume the table 1 million rows in size.
2256 */
2257 if( pFirst ){
2258 assert( pFirst->aiRowEst!=0 ); /* Allocated together with pFirst */
2259 aiRowEstPk[0] = pFirst->aiRowEst[0];
2260 }else{
2261 aiRowEstPk[0] = 1000000;
dan5236ac12009-08-13 07:09:33 +00002262 }
drhcdaca552009-08-20 13:45:07 +00002263 pProbe = &sPk;
dan5236ac12009-08-13 07:09:33 +00002264 wsFlagMask = ~(
2265 WHERE_COLUMN_IN|WHERE_COLUMN_EQ|WHERE_COLUMN_NULL|WHERE_COLUMN_RANGE
2266 );
2267 eqTermMask = WO_EQ|WO_IN;
2268 pIdx = 0;
danielk197785574e32008-10-06 05:32:18 +00002269 }
drh51147ba2005-07-23 22:59:55 +00002270
drhcdaca552009-08-20 13:45:07 +00002271 /* Loop over all indices looking for the best one to use
2272 */
dan5236ac12009-08-13 07:09:33 +00002273 for(; pProbe; pIdx=pProbe=pProbe->pNext){
2274 const unsigned int * const aiRowEst = pProbe->aiRowEst;
2275 double cost; /* Cost of using pProbe */
2276 double nRow; /* Estimated number of rows in result set */
2277 int rev; /* True to scan in reverse order */
2278 int wsFlags = 0;
2279 Bitmask used = 0;
2280
2281 /* The following variables are populated based on the properties of
2282 ** scan being evaluated. They are then used to determine the expected
2283 ** cost and number of rows returned.
2284 **
2285 ** nEq:
2286 ** Number of equality terms that can be implemented using the index.
2287 **
2288 ** nInMul:
2289 ** The "in-multiplier". This is an estimate of how many seek operations
2290 ** SQLite must perform on the index in question. For example, if the
2291 ** WHERE clause is:
2292 **
2293 ** WHERE a IN (1, 2, 3) AND b IN (4, 5, 6)
2294 **
2295 ** SQLite must perform 9 lookups on an index on (a, b), so nInMul is
2296 ** set to 9. Given the same schema and either of the following WHERE
2297 ** clauses:
2298 **
2299 ** WHERE a = 1
2300 ** WHERE a >= 2
2301 **
2302 ** nInMul is set to 1.
2303 **
2304 ** If there exists a WHERE term of the form "x IN (SELECT ...)", then
2305 ** the sub-select is assumed to return 25 rows for the purposes of
2306 ** determining nInMul.
2307 **
2308 ** bInEst:
2309 ** Set to true if there was at least one "x IN (SELECT ...)" term used
2310 ** in determining the value of nInMul.
2311 **
drhcdaca552009-08-20 13:45:07 +00002312 ** nBound:
drh98cdf622009-08-20 18:14:42 +00002313 ** An estimate on the amount of the table that must be searched. A
2314 ** value of 100 means the entire table is searched. Range constraints
2315 ** might reduce this to a value less than 100 to indicate that only
2316 ** a fraction of the table needs searching. In the absence of
2317 ** sqlite_stat2 ANALYZE data, a single inequality reduces the search
2318 ** space to 1/3rd its original size. So an x>? constraint reduces
2319 ** nBound to 33. Two constraints (x>? AND x<?) reduce nBound to 11.
dan5236ac12009-08-13 07:09:33 +00002320 **
2321 ** bSort:
2322 ** Boolean. True if there is an ORDER BY clause that will require an
2323 ** external sort (i.e. scanning the index being evaluated will not
2324 ** correctly order records).
2325 **
2326 ** bLookup:
2327 ** Boolean. True if for each index entry visited a lookup on the
2328 ** corresponding table b-tree is required. This is always false
2329 ** for the rowid index. For other indexes, it is true unless all the
2330 ** columns of the table used by the SELECT statement are present in
2331 ** the index (such an index is sometimes described as a covering index).
2332 ** For example, given the index on (a, b), the second of the following
2333 ** two queries requires table b-tree lookups, but the first does not.
2334 **
2335 ** SELECT a, b FROM tbl WHERE a = 1;
2336 ** SELECT a, b, c FROM tbl WHERE a = 1;
drhfe05af82005-07-21 03:14:59 +00002337 */
dan5236ac12009-08-13 07:09:33 +00002338 int nEq;
2339 int bInEst = 0;
2340 int nInMul = 1;
drh98cdf622009-08-20 18:14:42 +00002341 int nBound = 100;
dan5236ac12009-08-13 07:09:33 +00002342 int bSort = 0;
2343 int bLookup = 0;
2344
2345 /* Determine the values of nEq and nInMul */
2346 for(nEq=0; nEq<pProbe->nColumn; nEq++){
2347 WhereTerm *pTerm; /* A single term of the WHERE clause */
2348 int j = pProbe->aiColumn[nEq];
2349 pTerm = findTerm(pWC, iCur, j, notReady, eqTermMask, pIdx);
drhfe05af82005-07-21 03:14:59 +00002350 if( pTerm==0 ) break;
dan5236ac12009-08-13 07:09:33 +00002351 wsFlags |= (WHERE_COLUMN_EQ|WHERE_ROWID_EQ);
drhb52076c2006-01-23 13:22:09 +00002352 if( pTerm->eOperator & WO_IN ){
drha6110402005-07-28 20:51:19 +00002353 Expr *pExpr = pTerm->pExpr;
drh165be382008-12-05 02:36:33 +00002354 wsFlags |= WHERE_COLUMN_IN;
danielk19776ab3a2e2009-02-19 14:39:25 +00002355 if( ExprHasProperty(pExpr, EP_xIsSelect) ){
dan5236ac12009-08-13 07:09:33 +00002356 nInMul *= 25;
2357 bInEst = 1;
danielk19776ab3a2e2009-02-19 14:39:25 +00002358 }else if( pExpr->x.pList ){
dan5236ac12009-08-13 07:09:33 +00002359 nInMul *= pExpr->x.pList->nExpr + 1;
drhfe05af82005-07-21 03:14:59 +00002360 }
drh46619d62009-04-24 14:51:42 +00002361 }else if( pTerm->eOperator & WO_ISNULL ){
2362 wsFlags |= WHERE_COLUMN_NULL;
drhfe05af82005-07-21 03:14:59 +00002363 }
dan5236ac12009-08-13 07:09:33 +00002364 used |= pTerm->prereqRight;
drhfe05af82005-07-21 03:14:59 +00002365 }
dan5236ac12009-08-13 07:09:33 +00002366
2367 /* Determine the value of nBound. */
2368 if( nEq<pProbe->nColumn ){
2369 int j = pProbe->aiColumn[nEq];
2370 if( findTerm(pWC, iCur, j, notReady, WO_LT|WO_LE|WO_GT|WO_GE, pIdx) ){
2371 WhereTerm *pTop = findTerm(pWC, iCur, j, notReady, WO_LT|WO_LE, pIdx);
2372 WhereTerm *pBtm = findTerm(pWC, iCur, j, notReady, WO_GT|WO_GE, pIdx);
dane275dc32009-08-18 16:24:58 +00002373 whereRangeScanEst(pParse, pProbe, nEq, pBtm, pTop, &nBound);
dan5236ac12009-08-13 07:09:33 +00002374 if( pTop ){
2375 wsFlags |= WHERE_TOP_LIMIT;
dan5236ac12009-08-13 07:09:33 +00002376 used |= pTop->prereqRight;
2377 }
2378 if( pBtm ){
2379 wsFlags |= WHERE_BTM_LIMIT;
dan5236ac12009-08-13 07:09:33 +00002380 used |= pBtm->prereqRight;
2381 }
2382 wsFlags |= (WHERE_COLUMN_RANGE|WHERE_ROWID_RANGE);
2383 }
2384 }else if( pProbe->onError!=OE_None ){
drh46619d62009-04-24 14:51:42 +00002385 testcase( wsFlags & WHERE_COLUMN_IN );
2386 testcase( wsFlags & WHERE_COLUMN_NULL );
2387 if( (wsFlags & (WHERE_COLUMN_IN|WHERE_COLUMN_NULL))==0 ){
2388 wsFlags |= WHERE_UNIQUE;
2389 }
drh943af3c2005-07-29 19:43:58 +00002390 }
drhfe05af82005-07-21 03:14:59 +00002391
dan5236ac12009-08-13 07:09:33 +00002392 /* If there is an ORDER BY clause and the index being considered will
2393 ** naturally scan rows in the required order, set the appropriate flags
2394 ** in wsFlags. Otherwise, if there is an ORDER BY clause but the index
2395 ** will scan rows in a different order, set the bSort variable. */
drh28c4cf42005-07-27 20:41:43 +00002396 if( pOrderBy ){
drh46619d62009-04-24 14:51:42 +00002397 if( (wsFlags & (WHERE_COLUMN_IN|WHERE_COLUMN_NULL))==0
dan5236ac12009-08-13 07:09:33 +00002398 && isSortingIndex(pParse,pWC->pMaskSet,pProbe,iCur,pOrderBy,nEq,&rev)
drh46619d62009-04-24 14:51:42 +00002399 ){
dan5236ac12009-08-13 07:09:33 +00002400 wsFlags |= WHERE_ROWID_RANGE|WHERE_COLUMN_RANGE|WHERE_ORDERBY;
2401 wsFlags |= (rev ? WHERE_REVERSE : 0);
drh28c4cf42005-07-27 20:41:43 +00002402 }else{
dan5236ac12009-08-13 07:09:33 +00002403 bSort = 1;
drh51147ba2005-07-23 22:59:55 +00002404 }
drhfe05af82005-07-21 03:14:59 +00002405 }
2406
dan5236ac12009-08-13 07:09:33 +00002407 /* If currently calculating the cost of using an index (not the IPK
2408 ** index), determine if all required column data may be obtained without
2409 ** seeking to entries in the main table (i.e. if the index is a covering
2410 ** index for this query). If it is, set the WHERE_IDX_ONLY flag in
2411 ** wsFlags. Otherwise, set the bLookup variable to true. */
2412 if( pIdx && wsFlags ){
drhfe05af82005-07-21 03:14:59 +00002413 Bitmask m = pSrc->colUsed;
2414 int j;
dan5236ac12009-08-13 07:09:33 +00002415 for(j=0; j<pIdx->nColumn; j++){
2416 int x = pIdx->aiColumn[j];
drhfe05af82005-07-21 03:14:59 +00002417 if( x<BMS-1 ){
2418 m &= ~(((Bitmask)1)<<x);
2419 }
2420 }
2421 if( m==0 ){
drh165be382008-12-05 02:36:33 +00002422 wsFlags |= WHERE_IDX_ONLY;
dan5236ac12009-08-13 07:09:33 +00002423 }else{
2424 bLookup = 1;
drhfe05af82005-07-21 03:14:59 +00002425 }
2426 }
2427
drhcdaca552009-08-20 13:45:07 +00002428 /**** Begin adding up the cost of using this index (Needs improvements)
2429 **
2430 ** Estimate the number of rows of output. For an IN operator,
2431 ** do not let the estimate exceed half the rows in the table.
2432 */
dan5236ac12009-08-13 07:09:33 +00002433 nRow = (double)(aiRowEst[nEq] * nInMul);
2434 if( bInEst && nRow*2>aiRowEst[0] ){
2435 nRow = aiRowEst[0]/2;
shanecea72b22009-09-07 04:38:36 +00002436 nInMul = (int)(nRow / aiRowEst[nEq]);
dan5236ac12009-08-13 07:09:33 +00002437 }
drhcdaca552009-08-20 13:45:07 +00002438
2439 /* Assume constant cost to access a row and logarithmic cost to
2440 ** do a binary search. Hence, the initial cost is the number of output
2441 ** rows plus log2(table-size) times the number of binary searches.
2442 */
dan5236ac12009-08-13 07:09:33 +00002443 cost = nRow + nInMul*estLog(aiRowEst[0]);
drhcdaca552009-08-20 13:45:07 +00002444
2445 /* Adjust the number of rows and the cost downward to reflect rows
2446 ** that are excluded by range constraints.
2447 */
drh98cdf622009-08-20 18:14:42 +00002448 nRow = (nRow * (double)nBound) / (double)100;
2449 cost = (cost * (double)nBound) / (double)100;
drhcdaca552009-08-20 13:45:07 +00002450
2451 /* Add in the estimated cost of sorting the result
2452 */
dan5236ac12009-08-13 07:09:33 +00002453 if( bSort ){
2454 cost += cost*estLog(cost);
2455 }
drhcdaca552009-08-20 13:45:07 +00002456
2457 /* If all information can be taken directly from the index, we avoid
2458 ** doing table lookups. This reduces the cost by half. (Not really -
2459 ** this needs to be fixed.)
2460 */
dan5236ac12009-08-13 07:09:33 +00002461 if( pIdx && bLookup==0 ){
drhcdaca552009-08-20 13:45:07 +00002462 cost /= (double)2;
dan5236ac12009-08-13 07:09:33 +00002463 }
drhcdaca552009-08-20 13:45:07 +00002464 /**** Cost of using this index has now been computed ****/
dan5236ac12009-08-13 07:09:33 +00002465
2466 WHERETRACE((
2467 "tbl=%s idx=%s nEq=%d nInMul=%d nBound=%d bSort=%d bLookup=%d"
2468 " wsFlags=%d (nRow=%.2f cost=%.2f)\n",
2469 pSrc->pTab->zName, (pIdx ? pIdx->zName : "ipk"),
2470 nEq, nInMul, nBound, bSort, bLookup, wsFlags, nRow, cost
2471 ));
2472
drhcdaca552009-08-20 13:45:07 +00002473 /* If this index is the best we have seen so far, then record this
2474 ** index and its cost in the pCost structure.
2475 */
dan5236ac12009-08-13 07:09:33 +00002476 if( (!pIdx || wsFlags) && cost<pCost->rCost ){
drh111a6a72008-12-21 03:51:16 +00002477 pCost->rCost = cost;
2478 pCost->nRow = nRow;
dan5236ac12009-08-13 07:09:33 +00002479 pCost->used = used;
2480 pCost->plan.wsFlags = (wsFlags&wsFlagMask);
drh111a6a72008-12-21 03:51:16 +00002481 pCost->plan.nEq = nEq;
dan5236ac12009-08-13 07:09:33 +00002482 pCost->plan.u.pIdx = pIdx;
drhfe05af82005-07-21 03:14:59 +00002483 }
dan5236ac12009-08-13 07:09:33 +00002484
drhcdaca552009-08-20 13:45:07 +00002485 /* If there was an INDEXED BY clause, then only that one index is
2486 ** considered. */
dan5236ac12009-08-13 07:09:33 +00002487 if( pSrc->pIndex ) break;
drhcdaca552009-08-20 13:45:07 +00002488
2489 /* Reset masks for the next index in the loop */
dan5236ac12009-08-13 07:09:33 +00002490 wsFlagMask = ~(WHERE_ROWID_EQ|WHERE_ROWID_RANGE);
2491 eqTermMask = idxEqTermMask;
drhfe05af82005-07-21 03:14:59 +00002492 }
2493
dan5236ac12009-08-13 07:09:33 +00002494 /* If there is no ORDER BY clause and the SQLITE_ReverseOrder flag
2495 ** is set, then reverse the order that the index will be scanned
2496 ** in. This is used for application testing, to help find cases
2497 ** where application behaviour depends on the (undefined) order that
2498 ** SQLite outputs rows in in the absence of an ORDER BY clause. */
2499 if( !pOrderBy && pParse->db->flags & SQLITE_ReverseOrder ){
2500 pCost->plan.wsFlags |= WHERE_REVERSE;
2501 }
2502
2503 assert( pOrderBy || (pCost->plan.wsFlags&WHERE_ORDERBY)==0 );
2504 assert( pCost->plan.u.pIdx==0 || (pCost->plan.wsFlags&WHERE_ROWID_EQ)==0 );
2505 assert( pSrc->pIndex==0
2506 || pCost->plan.u.pIdx==0
2507 || pCost->plan.u.pIdx==pSrc->pIndex
2508 );
2509
2510 WHERETRACE(("best index is: %s\n",
2511 (pCost->plan.u.pIdx ? pCost->plan.u.pIdx->zName : "ipk")
2512 ));
2513
2514 bestOrClauseIndex(pParse, pWC, pSrc, notReady, pOrderBy, pCost);
drh111a6a72008-12-21 03:51:16 +00002515 pCost->plan.wsFlags |= eqTermMask;
drhfe05af82005-07-21 03:14:59 +00002516}
2517
danielk19771d461462009-04-21 09:02:45 +00002518/*
2519** Find the query plan for accessing table pSrc->pTab. Write the
2520** best query plan and its cost into the WhereCost object supplied
2521** as the last parameter. This function may calculate the cost of
2522** both real and virtual table scans.
2523*/
2524static void bestIndex(
2525 Parse *pParse, /* The parsing context */
2526 WhereClause *pWC, /* The WHERE clause */
2527 struct SrcList_item *pSrc, /* The FROM clause term to search */
2528 Bitmask notReady, /* Mask of cursors that are not available */
2529 ExprList *pOrderBy, /* The ORDER BY clause */
2530 WhereCost *pCost /* Lowest cost query plan */
2531){
shanee26fa4c2009-06-16 14:15:22 +00002532#ifndef SQLITE_OMIT_VIRTUALTABLE
danielk19771d461462009-04-21 09:02:45 +00002533 if( IsVirtual(pSrc->pTab) ){
2534 sqlite3_index_info *p = 0;
2535 bestVirtualIndex(pParse, pWC, pSrc, notReady, pOrderBy, pCost, &p);
2536 if( p->needToFreeIdxStr ){
2537 sqlite3_free(p->idxStr);
2538 }
2539 sqlite3DbFree(pParse->db, p);
shanee26fa4c2009-06-16 14:15:22 +00002540 }else
2541#endif
2542 {
danielk19771d461462009-04-21 09:02:45 +00002543 bestBtreeIndex(pParse, pWC, pSrc, notReady, pOrderBy, pCost);
2544 }
2545}
drhb6c29892004-11-22 19:12:19 +00002546
2547/*
drh2ffb1182004-07-19 19:14:01 +00002548** Disable a term in the WHERE clause. Except, do not disable the term
2549** if it controls a LEFT OUTER JOIN and it did not originate in the ON
2550** or USING clause of that join.
2551**
2552** Consider the term t2.z='ok' in the following queries:
2553**
2554** (1) SELECT * FROM t1 LEFT JOIN t2 ON t1.a=t2.x WHERE t2.z='ok'
2555** (2) SELECT * FROM t1 LEFT JOIN t2 ON t1.a=t2.x AND t2.z='ok'
2556** (3) SELECT * FROM t1, t2 WHERE t1.a=t2.x AND t2.z='ok'
2557**
drh23bf66d2004-12-14 03:34:34 +00002558** The t2.z='ok' is disabled in the in (2) because it originates
drh2ffb1182004-07-19 19:14:01 +00002559** in the ON clause. The term is disabled in (3) because it is not part
2560** of a LEFT OUTER JOIN. In (1), the term is not disabled.
2561**
2562** Disabling a term causes that term to not be tested in the inner loop
drhb6fb62d2005-09-20 08:47:20 +00002563** of the join. Disabling is an optimization. When terms are satisfied
2564** by indices, we disable them to prevent redundant tests in the inner
2565** loop. We would get the correct results if nothing were ever disabled,
2566** but joins might run a little slower. The trick is to disable as much
2567** as we can without disabling too much. If we disabled in (1), we'd get
2568** the wrong answer. See ticket #813.
drh2ffb1182004-07-19 19:14:01 +00002569*/
drh0fcef5e2005-07-19 17:38:22 +00002570static void disableTerm(WhereLevel *pLevel, WhereTerm *pTerm){
2571 if( pTerm
drh165be382008-12-05 02:36:33 +00002572 && ALWAYS((pTerm->wtFlags & TERM_CODED)==0)
drh0fcef5e2005-07-19 17:38:22 +00002573 && (pLevel->iLeftJoin==0 || ExprHasProperty(pTerm->pExpr, EP_FromJoin))
2574 ){
drh165be382008-12-05 02:36:33 +00002575 pTerm->wtFlags |= TERM_CODED;
drh45b1ee42005-08-02 17:48:22 +00002576 if( pTerm->iParent>=0 ){
2577 WhereTerm *pOther = &pTerm->pWC->a[pTerm->iParent];
2578 if( (--pOther->nChild)==0 ){
drhed378002005-07-28 23:12:08 +00002579 disableTerm(pLevel, pOther);
2580 }
drh0fcef5e2005-07-19 17:38:22 +00002581 }
drh2ffb1182004-07-19 19:14:01 +00002582 }
2583}
2584
2585/*
dan69f8bb92009-08-13 19:21:16 +00002586** Code an OP_Affinity opcode to apply the column affinity string zAff
2587** to the n registers starting at base.
2588**
2589** Buffer zAff was allocated using sqlite3DbMalloc(). It is the
2590** responsibility of this function to arrange for it to be eventually
2591** freed using sqlite3DbFree().
drh94a11212004-09-25 13:12:14 +00002592*/
dan69f8bb92009-08-13 19:21:16 +00002593static void codeApplyAffinity(Parse *pParse, int base, int n, char *zAff){
2594 Vdbe *v = pParse->pVdbe;
2595 assert( v!=0 );
2596 sqlite3VdbeAddOp2(v, OP_Affinity, base, n);
2597 sqlite3VdbeChangeP4(v, -1, zAff, P4_DYNAMIC);
2598 sqlite3ExprCacheAffinityChange(pParse, base, n);
drh94a11212004-09-25 13:12:14 +00002599}
2600
drhe8b97272005-07-19 22:22:12 +00002601
2602/*
drh51147ba2005-07-23 22:59:55 +00002603** Generate code for a single equality term of the WHERE clause. An equality
2604** term can be either X=expr or X IN (...). pTerm is the term to be
2605** coded.
2606**
drh1db639c2008-01-17 02:36:28 +00002607** The current value for the constraint is left in register iReg.
drh51147ba2005-07-23 22:59:55 +00002608**
2609** For a constraint of the form X=expr, the expression is evaluated and its
2610** result is left on the stack. For constraints of the form X IN (...)
2611** this routine sets up a loop that will iterate over all values of X.
drh94a11212004-09-25 13:12:14 +00002612*/
drh678ccce2008-03-31 18:19:54 +00002613static int codeEqualityTerm(
drh94a11212004-09-25 13:12:14 +00002614 Parse *pParse, /* The parsing context */
drhe23399f2005-07-22 00:31:39 +00002615 WhereTerm *pTerm, /* The term of the WHERE clause to be coded */
drh1db639c2008-01-17 02:36:28 +00002616 WhereLevel *pLevel, /* When level of the FROM clause we are working on */
drh678ccce2008-03-31 18:19:54 +00002617 int iTarget /* Attempt to leave results in this register */
drh94a11212004-09-25 13:12:14 +00002618){
drh0fcef5e2005-07-19 17:38:22 +00002619 Expr *pX = pTerm->pExpr;
drh50b39962006-10-28 00:28:09 +00002620 Vdbe *v = pParse->pVdbe;
drh678ccce2008-03-31 18:19:54 +00002621 int iReg; /* Register holding results */
drh1db639c2008-01-17 02:36:28 +00002622
danielk19772d605492008-10-01 08:43:03 +00002623 assert( iTarget>0 );
drh50b39962006-10-28 00:28:09 +00002624 if( pX->op==TK_EQ ){
drh678ccce2008-03-31 18:19:54 +00002625 iReg = sqlite3ExprCodeTarget(pParse, pX->pRight, iTarget);
drh50b39962006-10-28 00:28:09 +00002626 }else if( pX->op==TK_ISNULL ){
drh678ccce2008-03-31 18:19:54 +00002627 iReg = iTarget;
drh1db639c2008-01-17 02:36:28 +00002628 sqlite3VdbeAddOp2(v, OP_Null, 0, iReg);
danielk1977b3bce662005-01-29 08:32:43 +00002629#ifndef SQLITE_OMIT_SUBQUERY
drh94a11212004-09-25 13:12:14 +00002630 }else{
danielk19779a96b662007-11-29 17:05:18 +00002631 int eType;
danielk1977b3bce662005-01-29 08:32:43 +00002632 int iTab;
drh72e8fa42007-03-28 14:30:06 +00002633 struct InLoop *pIn;
danielk1977b3bce662005-01-29 08:32:43 +00002634
drh50b39962006-10-28 00:28:09 +00002635 assert( pX->op==TK_IN );
drh678ccce2008-03-31 18:19:54 +00002636 iReg = iTarget;
danielk19770cdc0222008-06-26 18:04:03 +00002637 eType = sqlite3FindInIndex(pParse, pX, 0);
danielk1977b3bce662005-01-29 08:32:43 +00002638 iTab = pX->iTable;
drh66a51672008-01-03 00:01:23 +00002639 sqlite3VdbeAddOp2(v, OP_Rewind, iTab, 0);
drh111a6a72008-12-21 03:51:16 +00002640 assert( pLevel->plan.wsFlags & WHERE_IN_ABLE );
2641 if( pLevel->u.in.nIn==0 ){
drhb3190c12008-12-08 21:37:14 +00002642 pLevel->addrNxt = sqlite3VdbeMakeLabel(v);
drh72e8fa42007-03-28 14:30:06 +00002643 }
drh111a6a72008-12-21 03:51:16 +00002644 pLevel->u.in.nIn++;
2645 pLevel->u.in.aInLoop =
2646 sqlite3DbReallocOrFree(pParse->db, pLevel->u.in.aInLoop,
2647 sizeof(pLevel->u.in.aInLoop[0])*pLevel->u.in.nIn);
2648 pIn = pLevel->u.in.aInLoop;
drh72e8fa42007-03-28 14:30:06 +00002649 if( pIn ){
drh111a6a72008-12-21 03:51:16 +00002650 pIn += pLevel->u.in.nIn - 1;
drh72e8fa42007-03-28 14:30:06 +00002651 pIn->iCur = iTab;
drh1db639c2008-01-17 02:36:28 +00002652 if( eType==IN_INDEX_ROWID ){
drhb3190c12008-12-08 21:37:14 +00002653 pIn->addrInTop = sqlite3VdbeAddOp2(v, OP_Rowid, iTab, iReg);
drh1db639c2008-01-17 02:36:28 +00002654 }else{
drhb3190c12008-12-08 21:37:14 +00002655 pIn->addrInTop = sqlite3VdbeAddOp3(v, OP_Column, iTab, 0, iReg);
drh1db639c2008-01-17 02:36:28 +00002656 }
2657 sqlite3VdbeAddOp1(v, OP_IsNull, iReg);
drha6110402005-07-28 20:51:19 +00002658 }else{
drh111a6a72008-12-21 03:51:16 +00002659 pLevel->u.in.nIn = 0;
drhe23399f2005-07-22 00:31:39 +00002660 }
danielk1977b3bce662005-01-29 08:32:43 +00002661#endif
drh94a11212004-09-25 13:12:14 +00002662 }
drh0fcef5e2005-07-19 17:38:22 +00002663 disableTerm(pLevel, pTerm);
drh678ccce2008-03-31 18:19:54 +00002664 return iReg;
drh94a11212004-09-25 13:12:14 +00002665}
2666
drh51147ba2005-07-23 22:59:55 +00002667/*
2668** Generate code that will evaluate all == and IN constraints for an
2669** index. The values for all constraints are left on the stack.
2670**
2671** For example, consider table t1(a,b,c,d,e,f) with index i1(a,b,c).
2672** Suppose the WHERE clause is this: a==5 AND b IN (1,2,3) AND c>5 AND c<10
2673** The index has as many as three equality constraints, but in this
2674** example, the third "c" value is an inequality. So only two
2675** constraints are coded. This routine will generate code to evaluate
drh6df2acd2008-12-28 16:55:25 +00002676** a==5 and b IN (1,2,3). The current values for a and b will be stored
2677** in consecutive registers and the index of the first register is returned.
drh51147ba2005-07-23 22:59:55 +00002678**
2679** In the example above nEq==2. But this subroutine works for any value
2680** of nEq including 0. If nEq==0, this routine is nearly a no-op.
2681** The only thing it does is allocate the pLevel->iMem memory cell.
2682**
drh700a2262008-12-17 19:22:15 +00002683** This routine always allocates at least one memory cell and returns
2684** the index of that memory cell. The code that
2685** calls this routine will use that memory cell to store the termination
drh51147ba2005-07-23 22:59:55 +00002686** key value of the loop. If one or more IN operators appear, then
2687** this routine allocates an additional nEq memory cells for internal
2688** use.
dan69f8bb92009-08-13 19:21:16 +00002689**
2690** Before returning, *pzAff is set to point to a buffer containing a
2691** copy of the column affinity string of the index allocated using
2692** sqlite3DbMalloc(). Except, entries in the copy of the string associated
2693** with equality constraints that use NONE affinity are set to
2694** SQLITE_AFF_NONE. This is to deal with SQL such as the following:
2695**
2696** CREATE TABLE t1(a TEXT PRIMARY KEY, b);
2697** SELECT ... FROM t1 AS t2, t1 WHERE t1.a = t2.b;
2698**
2699** In the example above, the index on t1(a) has TEXT affinity. But since
2700** the right hand side of the equality constraint (t2.b) has NONE affinity,
2701** no conversion should be attempted before using a t2.b value as part of
2702** a key to search the index. Hence the first byte in the returned affinity
2703** string in this example would be set to SQLITE_AFF_NONE.
drh51147ba2005-07-23 22:59:55 +00002704*/
drh1db639c2008-01-17 02:36:28 +00002705static int codeAllEqualityTerms(
drh51147ba2005-07-23 22:59:55 +00002706 Parse *pParse, /* Parsing context */
2707 WhereLevel *pLevel, /* Which nested loop of the FROM we are coding */
2708 WhereClause *pWC, /* The WHERE clause */
drh1db639c2008-01-17 02:36:28 +00002709 Bitmask notReady, /* Which parts of FROM have not yet been coded */
dan69f8bb92009-08-13 19:21:16 +00002710 int nExtraReg, /* Number of extra registers to allocate */
2711 char **pzAff /* OUT: Set to point to affinity string */
drh51147ba2005-07-23 22:59:55 +00002712){
drh111a6a72008-12-21 03:51:16 +00002713 int nEq = pLevel->plan.nEq; /* The number of == or IN constraints to code */
2714 Vdbe *v = pParse->pVdbe; /* The vm under construction */
2715 Index *pIdx; /* The index being used for this loop */
drh51147ba2005-07-23 22:59:55 +00002716 int iCur = pLevel->iTabCur; /* The cursor of the table */
2717 WhereTerm *pTerm; /* A single constraint term */
2718 int j; /* Loop counter */
drh1db639c2008-01-17 02:36:28 +00002719 int regBase; /* Base register */
drh6df2acd2008-12-28 16:55:25 +00002720 int nReg; /* Number of registers to allocate */
dan69f8bb92009-08-13 19:21:16 +00002721 char *zAff; /* Affinity string to return */
drh51147ba2005-07-23 22:59:55 +00002722
drh111a6a72008-12-21 03:51:16 +00002723 /* This module is only called on query plans that use an index. */
2724 assert( pLevel->plan.wsFlags & WHERE_INDEXED );
2725 pIdx = pLevel->plan.u.pIdx;
2726
drh51147ba2005-07-23 22:59:55 +00002727 /* Figure out how many memory cells we will need then allocate them.
drh51147ba2005-07-23 22:59:55 +00002728 */
drh700a2262008-12-17 19:22:15 +00002729 regBase = pParse->nMem + 1;
drh6df2acd2008-12-28 16:55:25 +00002730 nReg = pLevel->plan.nEq + nExtraReg;
2731 pParse->nMem += nReg;
drh51147ba2005-07-23 22:59:55 +00002732
dan69f8bb92009-08-13 19:21:16 +00002733 zAff = sqlite3DbStrDup(pParse->db, sqlite3IndexAffinityStr(v, pIdx));
2734 if( !zAff ){
2735 pParse->db->mallocFailed = 1;
2736 }
2737
drh51147ba2005-07-23 22:59:55 +00002738 /* Evaluate the equality constraints
2739 */
drhc49de5d2007-01-19 01:06:01 +00002740 assert( pIdx->nColumn>=nEq );
2741 for(j=0; j<nEq; j++){
drh678ccce2008-03-31 18:19:54 +00002742 int r1;
drh51147ba2005-07-23 22:59:55 +00002743 int k = pIdx->aiColumn[j];
drh111a6a72008-12-21 03:51:16 +00002744 pTerm = findTerm(pWC, iCur, k, notReady, pLevel->plan.wsFlags, pIdx);
drh34004ce2008-07-11 16:15:17 +00002745 if( NEVER(pTerm==0) ) break;
drh165be382008-12-05 02:36:33 +00002746 assert( (pTerm->wtFlags & TERM_CODED)==0 );
drh678ccce2008-03-31 18:19:54 +00002747 r1 = codeEqualityTerm(pParse, pTerm, pLevel, regBase+j);
2748 if( r1!=regBase+j ){
drh6df2acd2008-12-28 16:55:25 +00002749 if( nReg==1 ){
2750 sqlite3ReleaseTempReg(pParse, regBase);
2751 regBase = r1;
2752 }else{
2753 sqlite3VdbeAddOp2(v, OP_SCopy, r1, regBase+j);
2754 }
drh678ccce2008-03-31 18:19:54 +00002755 }
drh981642f2008-04-19 14:40:43 +00002756 testcase( pTerm->eOperator & WO_ISNULL );
2757 testcase( pTerm->eOperator & WO_IN );
drh72e8fa42007-03-28 14:30:06 +00002758 if( (pTerm->eOperator & (WO_ISNULL|WO_IN))==0 ){
drhb3190c12008-12-08 21:37:14 +00002759 sqlite3VdbeAddOp2(v, OP_IsNull, regBase+j, pLevel->addrBrk);
dan69f8bb92009-08-13 19:21:16 +00002760 if( zAff
2761 && sqlite3CompareAffinity(pTerm->pExpr->pRight, zAff[j])==SQLITE_AFF_NONE
2762 ){
2763 zAff[j] = SQLITE_AFF_NONE;
2764 }
drh51147ba2005-07-23 22:59:55 +00002765 }
2766 }
dan69f8bb92009-08-13 19:21:16 +00002767 *pzAff = zAff;
drh1db639c2008-01-17 02:36:28 +00002768 return regBase;
drh51147ba2005-07-23 22:59:55 +00002769}
2770
drh111a6a72008-12-21 03:51:16 +00002771/*
2772** Generate code for the start of the iLevel-th loop in the WHERE clause
2773** implementation described by pWInfo.
2774*/
2775static Bitmask codeOneLoopStart(
2776 WhereInfo *pWInfo, /* Complete information about the WHERE clause */
2777 int iLevel, /* Which level of pWInfo->a[] should be coded */
drh336a5302009-04-24 15:46:21 +00002778 u16 wctrlFlags, /* One of the WHERE_* flags defined in sqliteInt.h */
drh111a6a72008-12-21 03:51:16 +00002779 Bitmask notReady /* Which tables are currently available */
2780){
2781 int j, k; /* Loop counters */
2782 int iCur; /* The VDBE cursor for the table */
2783 int addrNxt; /* Where to jump to continue with the next IN case */
2784 int omitTable; /* True if we use the index only */
2785 int bRev; /* True if we need to scan in reverse order */
2786 WhereLevel *pLevel; /* The where level to be coded */
2787 WhereClause *pWC; /* Decomposition of the entire WHERE clause */
2788 WhereTerm *pTerm; /* A WHERE clause term */
2789 Parse *pParse; /* Parsing context */
2790 Vdbe *v; /* The prepared stmt under constructions */
2791 struct SrcList_item *pTabItem; /* FROM clause term being coded */
drh23d04d52008-12-23 23:56:22 +00002792 int addrBrk; /* Jump here to break out of the loop */
2793 int addrCont; /* Jump here to continue with next cycle */
drh61495262009-04-22 15:32:59 +00002794 int iRowidReg = 0; /* Rowid is stored in this register, if not zero */
2795 int iReleaseReg = 0; /* Temp register to free before returning */
drh111a6a72008-12-21 03:51:16 +00002796
2797 pParse = pWInfo->pParse;
2798 v = pParse->pVdbe;
2799 pWC = pWInfo->pWC;
2800 pLevel = &pWInfo->a[iLevel];
2801 pTabItem = &pWInfo->pTabList->a[pLevel->iFrom];
2802 iCur = pTabItem->iCursor;
2803 bRev = (pLevel->plan.wsFlags & WHERE_REVERSE)!=0;
danielk19771d461462009-04-21 09:02:45 +00002804 omitTable = (pLevel->plan.wsFlags & WHERE_IDX_ONLY)!=0
drh336a5302009-04-24 15:46:21 +00002805 && (wctrlFlags & WHERE_FORCE_TABLE)==0;
drh111a6a72008-12-21 03:51:16 +00002806
2807 /* Create labels for the "break" and "continue" instructions
2808 ** for the current loop. Jump to addrBrk to break out of a loop.
2809 ** Jump to cont to go immediately to the next iteration of the
2810 ** loop.
2811 **
2812 ** When there is an IN operator, we also have a "addrNxt" label that
2813 ** means to continue with the next IN value combination. When
2814 ** there are no IN operators in the constraints, the "addrNxt" label
2815 ** is the same as "addrBrk".
2816 */
2817 addrBrk = pLevel->addrBrk = pLevel->addrNxt = sqlite3VdbeMakeLabel(v);
2818 addrCont = pLevel->addrCont = sqlite3VdbeMakeLabel(v);
2819
2820 /* If this is the right table of a LEFT OUTER JOIN, allocate and
2821 ** initialize a memory cell that records if this table matches any
2822 ** row of the left table of the join.
2823 */
2824 if( pLevel->iFrom>0 && (pTabItem[0].jointype & JT_LEFT)!=0 ){
2825 pLevel->iLeftJoin = ++pParse->nMem;
2826 sqlite3VdbeAddOp2(v, OP_Integer, 0, pLevel->iLeftJoin);
2827 VdbeComment((v, "init LEFT JOIN no-match flag"));
2828 }
2829
2830#ifndef SQLITE_OMIT_VIRTUALTABLE
2831 if( (pLevel->plan.wsFlags & WHERE_VIRTUALTABLE)!=0 ){
2832 /* Case 0: The table is a virtual-table. Use the VFilter and VNext
2833 ** to access the data.
2834 */
2835 int iReg; /* P3 Value for OP_VFilter */
2836 sqlite3_index_info *pVtabIdx = pLevel->plan.u.pVtabIdx;
2837 int nConstraint = pVtabIdx->nConstraint;
2838 struct sqlite3_index_constraint_usage *aUsage =
2839 pVtabIdx->aConstraintUsage;
2840 const struct sqlite3_index_constraint *aConstraint =
2841 pVtabIdx->aConstraint;
2842
2843 iReg = sqlite3GetTempRange(pParse, nConstraint+2);
drh111a6a72008-12-21 03:51:16 +00002844 for(j=1; j<=nConstraint; j++){
2845 for(k=0; k<nConstraint; k++){
2846 if( aUsage[k].argvIndex==j ){
2847 int iTerm = aConstraint[k].iTermOffset;
drh111a6a72008-12-21 03:51:16 +00002848 sqlite3ExprCode(pParse, pWC->a[iTerm].pExpr->pRight, iReg+j+1);
2849 break;
2850 }
2851 }
2852 if( k==nConstraint ) break;
2853 }
drh111a6a72008-12-21 03:51:16 +00002854 sqlite3VdbeAddOp2(v, OP_Integer, pVtabIdx->idxNum, iReg);
2855 sqlite3VdbeAddOp2(v, OP_Integer, j-1, iReg+1);
2856 sqlite3VdbeAddOp4(v, OP_VFilter, iCur, addrBrk, iReg, pVtabIdx->idxStr,
2857 pVtabIdx->needToFreeIdxStr ? P4_MPRINTF : P4_STATIC);
drh111a6a72008-12-21 03:51:16 +00002858 pVtabIdx->needToFreeIdxStr = 0;
2859 for(j=0; j<nConstraint; j++){
2860 if( aUsage[j].omit ){
2861 int iTerm = aConstraint[j].iTermOffset;
2862 disableTerm(pLevel, &pWC->a[iTerm]);
2863 }
2864 }
2865 pLevel->op = OP_VNext;
2866 pLevel->p1 = iCur;
2867 pLevel->p2 = sqlite3VdbeCurrentAddr(v);
drh23d04d52008-12-23 23:56:22 +00002868 sqlite3ReleaseTempRange(pParse, iReg, nConstraint+2);
drh111a6a72008-12-21 03:51:16 +00002869 }else
2870#endif /* SQLITE_OMIT_VIRTUALTABLE */
2871
2872 if( pLevel->plan.wsFlags & WHERE_ROWID_EQ ){
2873 /* Case 1: We can directly reference a single row using an
2874 ** equality comparison against the ROWID field. Or
2875 ** we reference multiple rows using a "rowid IN (...)"
2876 ** construct.
2877 */
danielk19771d461462009-04-21 09:02:45 +00002878 iReleaseReg = sqlite3GetTempReg(pParse);
drh111a6a72008-12-21 03:51:16 +00002879 pTerm = findTerm(pWC, iCur, -1, notReady, WO_EQ|WO_IN, 0);
2880 assert( pTerm!=0 );
2881 assert( pTerm->pExpr!=0 );
2882 assert( pTerm->leftCursor==iCur );
2883 assert( omitTable==0 );
danielk19771d461462009-04-21 09:02:45 +00002884 iRowidReg = codeEqualityTerm(pParse, pTerm, pLevel, iReleaseReg);
drh111a6a72008-12-21 03:51:16 +00002885 addrNxt = pLevel->addrNxt;
danielk19771d461462009-04-21 09:02:45 +00002886 sqlite3VdbeAddOp2(v, OP_MustBeInt, iRowidReg, addrNxt);
2887 sqlite3VdbeAddOp3(v, OP_NotExists, iCur, addrNxt, iRowidReg);
drhceea3322009-04-23 13:22:42 +00002888 sqlite3ExprCacheStore(pParse, iCur, -1, iRowidReg);
drh111a6a72008-12-21 03:51:16 +00002889 VdbeComment((v, "pk"));
2890 pLevel->op = OP_Noop;
2891 }else if( pLevel->plan.wsFlags & WHERE_ROWID_RANGE ){
2892 /* Case 2: We have an inequality comparison against the ROWID field.
2893 */
2894 int testOp = OP_Noop;
2895 int start;
2896 int memEndValue = 0;
2897 WhereTerm *pStart, *pEnd;
2898
2899 assert( omitTable==0 );
2900 pStart = findTerm(pWC, iCur, -1, notReady, WO_GT|WO_GE, 0);
2901 pEnd = findTerm(pWC, iCur, -1, notReady, WO_LT|WO_LE, 0);
2902 if( bRev ){
2903 pTerm = pStart;
2904 pStart = pEnd;
2905 pEnd = pTerm;
2906 }
2907 if( pStart ){
2908 Expr *pX; /* The expression that defines the start bound */
2909 int r1, rTemp; /* Registers for holding the start boundary */
2910
2911 /* The following constant maps TK_xx codes into corresponding
2912 ** seek opcodes. It depends on a particular ordering of TK_xx
2913 */
2914 const u8 aMoveOp[] = {
2915 /* TK_GT */ OP_SeekGt,
2916 /* TK_LE */ OP_SeekLe,
2917 /* TK_LT */ OP_SeekLt,
2918 /* TK_GE */ OP_SeekGe
2919 };
2920 assert( TK_LE==TK_GT+1 ); /* Make sure the ordering.. */
2921 assert( TK_LT==TK_GT+2 ); /* ... of the TK_xx values... */
2922 assert( TK_GE==TK_GT+3 ); /* ... is correcct. */
2923
2924 pX = pStart->pExpr;
2925 assert( pX!=0 );
2926 assert( pStart->leftCursor==iCur );
2927 r1 = sqlite3ExprCodeTemp(pParse, pX->pRight, &rTemp);
2928 sqlite3VdbeAddOp3(v, aMoveOp[pX->op-TK_GT], iCur, addrBrk, r1);
2929 VdbeComment((v, "pk"));
2930 sqlite3ExprCacheAffinityChange(pParse, r1, 1);
2931 sqlite3ReleaseTempReg(pParse, rTemp);
2932 disableTerm(pLevel, pStart);
2933 }else{
2934 sqlite3VdbeAddOp2(v, bRev ? OP_Last : OP_Rewind, iCur, addrBrk);
2935 }
2936 if( pEnd ){
2937 Expr *pX;
2938 pX = pEnd->pExpr;
2939 assert( pX!=0 );
2940 assert( pEnd->leftCursor==iCur );
2941 memEndValue = ++pParse->nMem;
2942 sqlite3ExprCode(pParse, pX->pRight, memEndValue);
2943 if( pX->op==TK_LT || pX->op==TK_GT ){
2944 testOp = bRev ? OP_Le : OP_Ge;
2945 }else{
2946 testOp = bRev ? OP_Lt : OP_Gt;
2947 }
2948 disableTerm(pLevel, pEnd);
2949 }
2950 start = sqlite3VdbeCurrentAddr(v);
2951 pLevel->op = bRev ? OP_Prev : OP_Next;
2952 pLevel->p1 = iCur;
2953 pLevel->p2 = start;
drhca8c4662008-12-28 20:47:02 +00002954 pLevel->p5 = (pStart==0 && pEnd==0) ?1:0;
danielk19771d461462009-04-21 09:02:45 +00002955 if( testOp!=OP_Noop ){
2956 iRowidReg = iReleaseReg = sqlite3GetTempReg(pParse);
2957 sqlite3VdbeAddOp2(v, OP_Rowid, iCur, iRowidReg);
drhceea3322009-04-23 13:22:42 +00002958 sqlite3ExprCacheStore(pParse, iCur, -1, iRowidReg);
danielk19771d461462009-04-21 09:02:45 +00002959 sqlite3VdbeAddOp3(v, testOp, memEndValue, addrBrk, iRowidReg);
2960 sqlite3VdbeChangeP5(v, SQLITE_AFF_NUMERIC | SQLITE_JUMPIFNULL);
drh111a6a72008-12-21 03:51:16 +00002961 }
2962 }else if( pLevel->plan.wsFlags & (WHERE_COLUMN_RANGE|WHERE_COLUMN_EQ) ){
2963 /* Case 3: A scan using an index.
2964 **
2965 ** The WHERE clause may contain zero or more equality
2966 ** terms ("==" or "IN" operators) that refer to the N
2967 ** left-most columns of the index. It may also contain
2968 ** inequality constraints (>, <, >= or <=) on the indexed
2969 ** column that immediately follows the N equalities. Only
2970 ** the right-most column can be an inequality - the rest must
2971 ** use the "==" and "IN" operators. For example, if the
2972 ** index is on (x,y,z), then the following clauses are all
2973 ** optimized:
2974 **
2975 ** x=5
2976 ** x=5 AND y=10
2977 ** x=5 AND y<10
2978 ** x=5 AND y>5 AND y<10
2979 ** x=5 AND y=5 AND z<=10
2980 **
2981 ** The z<10 term of the following cannot be used, only
2982 ** the x=5 term:
2983 **
2984 ** x=5 AND z<10
2985 **
2986 ** N may be zero if there are inequality constraints.
2987 ** If there are no inequality constraints, then N is at
2988 ** least one.
2989 **
2990 ** This case is also used when there are no WHERE clause
2991 ** constraints but an index is selected anyway, in order
2992 ** to force the output order to conform to an ORDER BY.
2993 */
2994 int aStartOp[] = {
2995 0,
2996 0,
2997 OP_Rewind, /* 2: (!start_constraints && startEq && !bRev) */
2998 OP_Last, /* 3: (!start_constraints && startEq && bRev) */
2999 OP_SeekGt, /* 4: (start_constraints && !startEq && !bRev) */
3000 OP_SeekLt, /* 5: (start_constraints && !startEq && bRev) */
3001 OP_SeekGe, /* 6: (start_constraints && startEq && !bRev) */
3002 OP_SeekLe /* 7: (start_constraints && startEq && bRev) */
3003 };
3004 int aEndOp[] = {
3005 OP_Noop, /* 0: (!end_constraints) */
3006 OP_IdxGE, /* 1: (end_constraints && !bRev) */
3007 OP_IdxLT /* 2: (end_constraints && bRev) */
3008 };
3009 int nEq = pLevel->plan.nEq;
3010 int isMinQuery = 0; /* If this is an optimized SELECT min(x).. */
3011 int regBase; /* Base register holding constraint values */
3012 int r1; /* Temp register */
3013 WhereTerm *pRangeStart = 0; /* Inequality constraint at range start */
3014 WhereTerm *pRangeEnd = 0; /* Inequality constraint at range end */
3015 int startEq; /* True if range start uses ==, >= or <= */
3016 int endEq; /* True if range end uses ==, >= or <= */
3017 int start_constraints; /* Start of range is constrained */
3018 int nConstraint; /* Number of constraint terms */
3019 Index *pIdx; /* The index we will be using */
3020 int iIdxCur; /* The VDBE cursor for the index */
drh6df2acd2008-12-28 16:55:25 +00003021 int nExtraReg = 0; /* Number of extra registers needed */
3022 int op; /* Instruction opcode */
dan69f8bb92009-08-13 19:21:16 +00003023 char *zAff;
drh111a6a72008-12-21 03:51:16 +00003024
3025 pIdx = pLevel->plan.u.pIdx;
3026 iIdxCur = pLevel->iIdxCur;
3027 k = pIdx->aiColumn[nEq]; /* Column for inequality constraints */
3028
drh111a6a72008-12-21 03:51:16 +00003029 /* If this loop satisfies a sort order (pOrderBy) request that
3030 ** was passed to this function to implement a "SELECT min(x) ..."
3031 ** query, then the caller will only allow the loop to run for
3032 ** a single iteration. This means that the first row returned
3033 ** should not have a NULL value stored in 'x'. If column 'x' is
3034 ** the first one after the nEq equality constraints in the index,
3035 ** this requires some special handling.
3036 */
3037 if( (wctrlFlags&WHERE_ORDERBY_MIN)!=0
3038 && (pLevel->plan.wsFlags&WHERE_ORDERBY)
3039 && (pIdx->nColumn>nEq)
3040 ){
3041 /* assert( pOrderBy->nExpr==1 ); */
3042 /* assert( pOrderBy->a[0].pExpr->iColumn==pIdx->aiColumn[nEq] ); */
3043 isMinQuery = 1;
drh6df2acd2008-12-28 16:55:25 +00003044 nExtraReg = 1;
drh111a6a72008-12-21 03:51:16 +00003045 }
3046
3047 /* Find any inequality constraint terms for the start and end
3048 ** of the range.
3049 */
3050 if( pLevel->plan.wsFlags & WHERE_TOP_LIMIT ){
3051 pRangeEnd = findTerm(pWC, iCur, k, notReady, (WO_LT|WO_LE), pIdx);
drh6df2acd2008-12-28 16:55:25 +00003052 nExtraReg = 1;
drh111a6a72008-12-21 03:51:16 +00003053 }
3054 if( pLevel->plan.wsFlags & WHERE_BTM_LIMIT ){
3055 pRangeStart = findTerm(pWC, iCur, k, notReady, (WO_GT|WO_GE), pIdx);
drh6df2acd2008-12-28 16:55:25 +00003056 nExtraReg = 1;
drh111a6a72008-12-21 03:51:16 +00003057 }
3058
drh6df2acd2008-12-28 16:55:25 +00003059 /* Generate code to evaluate all constraint terms using == or IN
3060 ** and store the values of those terms in an array of registers
3061 ** starting at regBase.
3062 */
dan69f8bb92009-08-13 19:21:16 +00003063 regBase = codeAllEqualityTerms(
3064 pParse, pLevel, pWC, notReady, nExtraReg, &zAff
3065 );
drh6df2acd2008-12-28 16:55:25 +00003066 addrNxt = pLevel->addrNxt;
3067
drh111a6a72008-12-21 03:51:16 +00003068 /* If we are doing a reverse order scan on an ascending index, or
3069 ** a forward order scan on a descending index, interchange the
3070 ** start and end terms (pRangeStart and pRangeEnd).
3071 */
3072 if( bRev==(pIdx->aSortOrder[nEq]==SQLITE_SO_ASC) ){
3073 SWAP(WhereTerm *, pRangeEnd, pRangeStart);
3074 }
3075
3076 testcase( pRangeStart && pRangeStart->eOperator & WO_LE );
3077 testcase( pRangeStart && pRangeStart->eOperator & WO_GE );
3078 testcase( pRangeEnd && pRangeEnd->eOperator & WO_LE );
3079 testcase( pRangeEnd && pRangeEnd->eOperator & WO_GE );
3080 startEq = !pRangeStart || pRangeStart->eOperator & (WO_LE|WO_GE);
3081 endEq = !pRangeEnd || pRangeEnd->eOperator & (WO_LE|WO_GE);
3082 start_constraints = pRangeStart || nEq>0;
3083
3084 /* Seek the index cursor to the start of the range. */
3085 nConstraint = nEq;
3086 if( pRangeStart ){
dan69f8bb92009-08-13 19:21:16 +00003087 Expr *pRight = pRangeStart->pExpr->pRight;
3088 sqlite3ExprCode(pParse, pRight, regBase+nEq);
drh111a6a72008-12-21 03:51:16 +00003089 sqlite3VdbeAddOp2(v, OP_IsNull, regBase+nEq, addrNxt);
dan69f8bb92009-08-13 19:21:16 +00003090 if( zAff
3091 && sqlite3CompareAffinity(pRight, zAff[nConstraint])==SQLITE_AFF_NONE
3092 ){
3093 /* Since the comparison is to be performed with no conversions applied
3094 ** to the operands, set the affinity to apply to pRight to
3095 ** SQLITE_AFF_NONE. */
3096 zAff[nConstraint] = SQLITE_AFF_NONE;
3097 }
drh111a6a72008-12-21 03:51:16 +00003098 nConstraint++;
3099 }else if( isMinQuery ){
3100 sqlite3VdbeAddOp2(v, OP_Null, 0, regBase+nEq);
3101 nConstraint++;
3102 startEq = 0;
3103 start_constraints = 1;
3104 }
dan69f8bb92009-08-13 19:21:16 +00003105 codeApplyAffinity(pParse, regBase, nConstraint, zAff);
drh111a6a72008-12-21 03:51:16 +00003106 op = aStartOp[(start_constraints<<2) + (startEq<<1) + bRev];
3107 assert( op!=0 );
3108 testcase( op==OP_Rewind );
3109 testcase( op==OP_Last );
3110 testcase( op==OP_SeekGt );
3111 testcase( op==OP_SeekGe );
3112 testcase( op==OP_SeekLe );
3113 testcase( op==OP_SeekLt );
3114 sqlite3VdbeAddOp4(v, op, iIdxCur, addrNxt, regBase,
3115 SQLITE_INT_TO_PTR(nConstraint), P4_INT32);
3116
3117 /* Load the value for the inequality constraint at the end of the
3118 ** range (if any).
3119 */
3120 nConstraint = nEq;
3121 if( pRangeEnd ){
dan69f8bb92009-08-13 19:21:16 +00003122 Expr *pRight = pRangeEnd->pExpr->pRight;
drhceea3322009-04-23 13:22:42 +00003123 sqlite3ExprCacheRemove(pParse, regBase+nEq);
dan69f8bb92009-08-13 19:21:16 +00003124 sqlite3ExprCode(pParse, pRight, regBase+nEq);
drh111a6a72008-12-21 03:51:16 +00003125 sqlite3VdbeAddOp2(v, OP_IsNull, regBase+nEq, addrNxt);
dan69f8bb92009-08-13 19:21:16 +00003126 zAff = sqlite3DbStrDup(pParse->db, zAff);
3127 if( zAff
3128 && sqlite3CompareAffinity(pRight, zAff[nConstraint])==SQLITE_AFF_NONE
3129 ){
3130 /* Since the comparison is to be performed with no conversions applied
3131 ** to the operands, set the affinity to apply to pRight to
3132 ** SQLITE_AFF_NONE. */
3133 zAff[nConstraint] = SQLITE_AFF_NONE;
3134 }
3135 codeApplyAffinity(pParse, regBase, nEq+1, zAff);
drh111a6a72008-12-21 03:51:16 +00003136 nConstraint++;
3137 }
3138
3139 /* Top of the loop body */
3140 pLevel->p2 = sqlite3VdbeCurrentAddr(v);
3141
3142 /* Check if the index cursor is past the end of the range. */
3143 op = aEndOp[(pRangeEnd || nEq) * (1 + bRev)];
3144 testcase( op==OP_Noop );
3145 testcase( op==OP_IdxGE );
3146 testcase( op==OP_IdxLT );
drh6df2acd2008-12-28 16:55:25 +00003147 if( op!=OP_Noop ){
3148 sqlite3VdbeAddOp4(v, op, iIdxCur, addrNxt, regBase,
3149 SQLITE_INT_TO_PTR(nConstraint), P4_INT32);
3150 sqlite3VdbeChangeP5(v, endEq!=bRev ?1:0);
3151 }
drh111a6a72008-12-21 03:51:16 +00003152
3153 /* If there are inequality constraints, check that the value
3154 ** of the table column that the inequality contrains is not NULL.
3155 ** If it is, jump to the next iteration of the loop.
3156 */
3157 r1 = sqlite3GetTempReg(pParse);
3158 testcase( pLevel->plan.wsFlags & WHERE_BTM_LIMIT );
3159 testcase( pLevel->plan.wsFlags & WHERE_TOP_LIMIT );
3160 if( pLevel->plan.wsFlags & (WHERE_BTM_LIMIT|WHERE_TOP_LIMIT) ){
3161 sqlite3VdbeAddOp3(v, OP_Column, iIdxCur, nEq, r1);
3162 sqlite3VdbeAddOp2(v, OP_IsNull, r1, addrCont);
3163 }
danielk19771d461462009-04-21 09:02:45 +00003164 sqlite3ReleaseTempReg(pParse, r1);
drh111a6a72008-12-21 03:51:16 +00003165
3166 /* Seek the table cursor, if required */
drh23d04d52008-12-23 23:56:22 +00003167 disableTerm(pLevel, pRangeStart);
3168 disableTerm(pLevel, pRangeEnd);
danielk19771d461462009-04-21 09:02:45 +00003169 if( !omitTable ){
3170 iRowidReg = iReleaseReg = sqlite3GetTempReg(pParse);
3171 sqlite3VdbeAddOp2(v, OP_IdxRowid, iIdxCur, iRowidReg);
drhceea3322009-04-23 13:22:42 +00003172 sqlite3ExprCacheStore(pParse, iCur, -1, iRowidReg);
danielk19771d461462009-04-21 09:02:45 +00003173 sqlite3VdbeAddOp2(v, OP_Seek, iCur, iRowidReg); /* Deferred seek */
drh111a6a72008-12-21 03:51:16 +00003174 }
drh111a6a72008-12-21 03:51:16 +00003175
3176 /* Record the instruction used to terminate the loop. Disable
3177 ** WHERE clause terms made redundant by the index range scan.
3178 */
3179 pLevel->op = bRev ? OP_Prev : OP_Next;
3180 pLevel->p1 = iIdxCur;
drhdd5f5a62008-12-23 13:35:23 +00003181 }else
3182
drh23d04d52008-12-23 23:56:22 +00003183#ifndef SQLITE_OMIT_OR_OPTIMIZATION
drhdd5f5a62008-12-23 13:35:23 +00003184 if( pLevel->plan.wsFlags & WHERE_MULTI_OR ){
drh111a6a72008-12-21 03:51:16 +00003185 /* Case 4: Two or more separately indexed terms connected by OR
3186 **
3187 ** Example:
3188 **
3189 ** CREATE TABLE t1(a,b,c,d);
3190 ** CREATE INDEX i1 ON t1(a);
3191 ** CREATE INDEX i2 ON t1(b);
3192 ** CREATE INDEX i3 ON t1(c);
3193 **
3194 ** SELECT * FROM t1 WHERE a=5 OR b=7 OR (c=11 AND d=13)
3195 **
3196 ** In the example, there are three indexed terms connected by OR.
danielk19771d461462009-04-21 09:02:45 +00003197 ** The top of the loop looks like this:
drh111a6a72008-12-21 03:51:16 +00003198 **
drh1b26c7c2009-04-22 02:15:47 +00003199 ** Null 1 # Zero the rowset in reg 1
drh111a6a72008-12-21 03:51:16 +00003200 **
danielk19771d461462009-04-21 09:02:45 +00003201 ** Then, for each indexed term, the following. The arguments to
drh1b26c7c2009-04-22 02:15:47 +00003202 ** RowSetTest are such that the rowid of the current row is inserted
3203 ** into the RowSet. If it is already present, control skips the
danielk19771d461462009-04-21 09:02:45 +00003204 ** Gosub opcode and jumps straight to the code generated by WhereEnd().
drh111a6a72008-12-21 03:51:16 +00003205 **
danielk19771d461462009-04-21 09:02:45 +00003206 ** sqlite3WhereBegin(<term>)
drh1b26c7c2009-04-22 02:15:47 +00003207 ** RowSetTest # Insert rowid into rowset
danielk19771d461462009-04-21 09:02:45 +00003208 ** Gosub 2 A
3209 ** sqlite3WhereEnd()
3210 **
3211 ** Following the above, code to terminate the loop. Label A, the target
3212 ** of the Gosub above, jumps to the instruction right after the Goto.
3213 **
drh1b26c7c2009-04-22 02:15:47 +00003214 ** Null 1 # Zero the rowset in reg 1
danielk19771d461462009-04-21 09:02:45 +00003215 ** Goto B # The loop is finished.
3216 **
3217 ** A: <loop body> # Return data, whatever.
3218 **
3219 ** Return 2 # Jump back to the Gosub
3220 **
3221 ** B: <after the loop>
3222 **
drh111a6a72008-12-21 03:51:16 +00003223 */
drh111a6a72008-12-21 03:51:16 +00003224 WhereClause *pOrWc; /* The OR-clause broken out into subterms */
danielk19771d461462009-04-21 09:02:45 +00003225 WhereTerm *pFinal; /* Final subterm within the OR-clause. */
drhdd5f5a62008-12-23 13:35:23 +00003226 SrcList oneTab; /* Shortened table list */
danielk19771d461462009-04-21 09:02:45 +00003227
3228 int regReturn = ++pParse->nMem; /* Register used with OP_Gosub */
shane85095702009-06-15 16:27:08 +00003229 int regRowset = 0; /* Register for RowSet object */
3230 int regRowid = 0; /* Register holding rowid */
danielk19771d461462009-04-21 09:02:45 +00003231 int iLoopBody = sqlite3VdbeMakeLabel(v); /* Start of loop body */
3232 int iRetInit; /* Address of regReturn init */
3233 int ii;
drh111a6a72008-12-21 03:51:16 +00003234
3235 pTerm = pLevel->plan.u.pTerm;
3236 assert( pTerm!=0 );
3237 assert( pTerm->eOperator==WO_OR );
3238 assert( (pTerm->wtFlags & TERM_ORINFO)!=0 );
3239 pOrWc = &pTerm->u.pOrInfo->wc;
danielk19771d461462009-04-21 09:02:45 +00003240 pFinal = &pOrWc->a[pOrWc->nTerm-1];
drh23d04d52008-12-23 23:56:22 +00003241
danielk19771d461462009-04-21 09:02:45 +00003242 /* Set up a SrcList containing just the table being scanned by this loop. */
drhdd5f5a62008-12-23 13:35:23 +00003243 oneTab.nSrc = 1;
3244 oneTab.nAlloc = 1;
3245 oneTab.a[0] = *pTabItem;
danielk19771d461462009-04-21 09:02:45 +00003246
drh1b26c7c2009-04-22 02:15:47 +00003247 /* Initialize the rowset register to contain NULL. An SQL NULL is
3248 ** equivalent to an empty rowset.
danielk19771d461462009-04-21 09:02:45 +00003249 **
3250 ** Also initialize regReturn to contain the address of the instruction
3251 ** immediately following the OP_Return at the bottom of the loop. This
3252 ** is required in a few obscure LEFT JOIN cases where control jumps
3253 ** over the top of the loop into the body of it. In this case the
3254 ** correct response for the end-of-loop code (the OP_Return) is to
3255 ** fall through to the next instruction, just as an OP_Next does if
3256 ** called on an uninitialized cursor.
3257 */
drh336a5302009-04-24 15:46:21 +00003258 if( (wctrlFlags & WHERE_DUPLICATES_OK)==0 ){
3259 regRowset = ++pParse->nMem;
3260 regRowid = ++pParse->nMem;
3261 sqlite3VdbeAddOp2(v, OP_Null, 0, regRowset);
3262 }
danielk19771d461462009-04-21 09:02:45 +00003263 iRetInit = sqlite3VdbeAddOp2(v, OP_Integer, 0, regReturn);
3264
danielk19771d461462009-04-21 09:02:45 +00003265 for(ii=0; ii<pOrWc->nTerm; ii++){
3266 WhereTerm *pOrTerm = &pOrWc->a[ii];
3267 if( pOrTerm->leftCursor==iCur || pOrTerm->eOperator==WO_AND ){
3268 WhereInfo *pSubWInfo; /* Info for single OR-term scan */
danielk19771d461462009-04-21 09:02:45 +00003269 /* Loop through table entries that match term pOrTerm. */
drh336a5302009-04-24 15:46:21 +00003270 pSubWInfo = sqlite3WhereBegin(pParse, &oneTab, pOrTerm->pExpr, 0,
3271 WHERE_OMIT_OPEN | WHERE_OMIT_CLOSE | WHERE_FORCE_TABLE);
danielk19771d461462009-04-21 09:02:45 +00003272 if( pSubWInfo ){
drh336a5302009-04-24 15:46:21 +00003273 if( (wctrlFlags & WHERE_DUPLICATES_OK)==0 ){
3274 int iSet = ((ii==pOrWc->nTerm-1)?-1:ii);
3275 int r;
3276 r = sqlite3ExprCodeGetColumn(pParse, pTabItem->pTab, -1, iCur,
3277 regRowid, 0);
shane85095702009-06-15 16:27:08 +00003278 sqlite3VdbeAddOp4(v, OP_RowSetTest, regRowset,
shane60a4b532009-05-06 18:57:09 +00003279 sqlite3VdbeCurrentAddr(v)+2,
3280 r, SQLITE_INT_TO_PTR(iSet), P4_INT32);
drh336a5302009-04-24 15:46:21 +00003281 }
danielk19771d461462009-04-21 09:02:45 +00003282 sqlite3VdbeAddOp2(v, OP_Gosub, regReturn, iLoopBody);
3283
3284 /* Finish the loop through table entries that match term pOrTerm. */
3285 sqlite3WhereEnd(pSubWInfo);
3286 }
drhdd5f5a62008-12-23 13:35:23 +00003287 }
3288 }
danielk19771d461462009-04-21 09:02:45 +00003289 sqlite3VdbeChangeP1(v, iRetInit, sqlite3VdbeCurrentAddr(v));
drh336a5302009-04-24 15:46:21 +00003290 /* sqlite3VdbeAddOp2(v, OP_Null, 0, regRowset); */
danielk19771d461462009-04-21 09:02:45 +00003291 sqlite3VdbeAddOp2(v, OP_Goto, 0, pLevel->addrBrk);
3292 sqlite3VdbeResolveLabel(v, iLoopBody);
3293
3294 pLevel->op = OP_Return;
3295 pLevel->p1 = regReturn;
drh23d04d52008-12-23 23:56:22 +00003296 disableTerm(pLevel, pTerm);
drhdd5f5a62008-12-23 13:35:23 +00003297 }else
drh23d04d52008-12-23 23:56:22 +00003298#endif /* SQLITE_OMIT_OR_OPTIMIZATION */
drhdd5f5a62008-12-23 13:35:23 +00003299
3300 {
drh111a6a72008-12-21 03:51:16 +00003301 /* Case 5: There is no usable index. We must do a complete
3302 ** scan of the entire table.
3303 */
drh699b3d42009-02-23 16:52:07 +00003304 static const u8 aStep[] = { OP_Next, OP_Prev };
3305 static const u8 aStart[] = { OP_Rewind, OP_Last };
3306 assert( bRev==0 || bRev==1 );
drh111a6a72008-12-21 03:51:16 +00003307 assert( omitTable==0 );
drh699b3d42009-02-23 16:52:07 +00003308 pLevel->op = aStep[bRev];
drh111a6a72008-12-21 03:51:16 +00003309 pLevel->p1 = iCur;
drh699b3d42009-02-23 16:52:07 +00003310 pLevel->p2 = 1 + sqlite3VdbeAddOp2(v, aStart[bRev], iCur, addrBrk);
drh111a6a72008-12-21 03:51:16 +00003311 pLevel->p5 = SQLITE_STMTSTATUS_FULLSCAN_STEP;
3312 }
3313 notReady &= ~getMask(pWC->pMaskSet, iCur);
3314
3315 /* Insert code to test every subexpression that can be completely
3316 ** computed using the current set of tables.
3317 */
3318 k = 0;
3319 for(pTerm=pWC->a, j=pWC->nTerm; j>0; j--, pTerm++){
3320 Expr *pE;
3321 testcase( pTerm->wtFlags & TERM_VIRTUAL );
3322 testcase( pTerm->wtFlags & TERM_CODED );
3323 if( pTerm->wtFlags & (TERM_VIRTUAL|TERM_CODED) ) continue;
3324 if( (pTerm->prereqAll & notReady)!=0 ) continue;
3325 pE = pTerm->pExpr;
3326 assert( pE!=0 );
3327 if( pLevel->iLeftJoin && !ExprHasProperty(pE, EP_FromJoin) ){
3328 continue;
3329 }
drh111a6a72008-12-21 03:51:16 +00003330 sqlite3ExprIfFalse(pParse, pE, addrCont, SQLITE_JUMPIFNULL);
drh111a6a72008-12-21 03:51:16 +00003331 k = 1;
3332 pTerm->wtFlags |= TERM_CODED;
3333 }
3334
3335 /* For a LEFT OUTER JOIN, generate code that will record the fact that
3336 ** at least one row of the right table has matched the left table.
3337 */
3338 if( pLevel->iLeftJoin ){
3339 pLevel->addrFirst = sqlite3VdbeCurrentAddr(v);
3340 sqlite3VdbeAddOp2(v, OP_Integer, 1, pLevel->iLeftJoin);
3341 VdbeComment((v, "record LEFT JOIN hit"));
drhceea3322009-04-23 13:22:42 +00003342 sqlite3ExprCacheClear(pParse);
drh111a6a72008-12-21 03:51:16 +00003343 for(pTerm=pWC->a, j=0; j<pWC->nTerm; j++, pTerm++){
3344 testcase( pTerm->wtFlags & TERM_VIRTUAL );
3345 testcase( pTerm->wtFlags & TERM_CODED );
3346 if( pTerm->wtFlags & (TERM_VIRTUAL|TERM_CODED) ) continue;
3347 if( (pTerm->prereqAll & notReady)!=0 ) continue;
3348 assert( pTerm->pExpr );
3349 sqlite3ExprIfFalse(pParse, pTerm->pExpr, addrCont, SQLITE_JUMPIFNULL);
3350 pTerm->wtFlags |= TERM_CODED;
3351 }
3352 }
danielk19771d461462009-04-21 09:02:45 +00003353 sqlite3ReleaseTempReg(pParse, iReleaseReg);
drh23d04d52008-12-23 23:56:22 +00003354
drh111a6a72008-12-21 03:51:16 +00003355 return notReady;
3356}
3357
drh549c8b62005-09-19 13:15:23 +00003358#if defined(SQLITE_TEST)
drh84bfda42005-07-15 13:05:21 +00003359/*
3360** The following variable holds a text description of query plan generated
3361** by the most recent call to sqlite3WhereBegin(). Each call to WhereBegin
3362** overwrites the previous. This information is used for testing and
3363** analysis only.
3364*/
3365char sqlite3_query_plan[BMS*2*40]; /* Text of the join */
3366static int nQPlan = 0; /* Next free slow in _query_plan[] */
3367
3368#endif /* SQLITE_TEST */
3369
3370
drh9eff6162006-06-12 21:59:13 +00003371/*
3372** Free a WhereInfo structure
3373*/
drh10fe8402008-10-11 16:47:35 +00003374static void whereInfoFree(sqlite3 *db, WhereInfo *pWInfo){
drh9eff6162006-06-12 21:59:13 +00003375 if( pWInfo ){
3376 int i;
3377 for(i=0; i<pWInfo->nLevel; i++){
drh4be8b512006-06-13 23:51:34 +00003378 sqlite3_index_info *pInfo = pWInfo->a[i].pIdxInfo;
3379 if( pInfo ){
danielk19771d461462009-04-21 09:02:45 +00003380 /* assert( pInfo->needToFreeIdxStr==0 || db->mallocFailed ); */
danielk197780442942008-12-24 11:25:39 +00003381 if( pInfo->needToFreeIdxStr ){
3382 sqlite3_free(pInfo->idxStr);
danielk1977be229652009-03-20 14:18:51 +00003383 }
drh633e6d52008-07-28 19:34:53 +00003384 sqlite3DbFree(db, pInfo);
danielk1977be8a7832006-06-13 15:00:54 +00003385 }
drh9eff6162006-06-12 21:59:13 +00003386 }
drh111a6a72008-12-21 03:51:16 +00003387 whereClauseClear(pWInfo->pWC);
drh633e6d52008-07-28 19:34:53 +00003388 sqlite3DbFree(db, pWInfo);
drh9eff6162006-06-12 21:59:13 +00003389 }
3390}
3391
drh94a11212004-09-25 13:12:14 +00003392
3393/*
drhe3184742002-06-19 14:27:05 +00003394** Generate the beginning of the loop used for WHERE clause processing.
drhacf3b982005-01-03 01:27:18 +00003395** The return value is a pointer to an opaque structure that contains
drh75897232000-05-29 14:26:00 +00003396** information needed to terminate the loop. Later, the calling routine
danielk19774adee202004-05-08 08:23:19 +00003397** should invoke sqlite3WhereEnd() with the return value of this function
drh75897232000-05-29 14:26:00 +00003398** in order to complete the WHERE clause processing.
3399**
3400** If an error occurs, this routine returns NULL.
drhc27a1ce2002-06-14 20:58:45 +00003401**
3402** The basic idea is to do a nested loop, one loop for each table in
3403** the FROM clause of a select. (INSERT and UPDATE statements are the
3404** same as a SELECT with only a single table in the FROM clause.) For
3405** example, if the SQL is this:
3406**
3407** SELECT * FROM t1, t2, t3 WHERE ...;
3408**
3409** Then the code generated is conceptually like the following:
3410**
3411** foreach row1 in t1 do \ Code generated
danielk19774adee202004-05-08 08:23:19 +00003412** foreach row2 in t2 do |-- by sqlite3WhereBegin()
drhc27a1ce2002-06-14 20:58:45 +00003413** foreach row3 in t3 do /
3414** ...
3415** end \ Code generated
danielk19774adee202004-05-08 08:23:19 +00003416** end |-- by sqlite3WhereEnd()
drhc27a1ce2002-06-14 20:58:45 +00003417** end /
3418**
drh29dda4a2005-07-21 18:23:20 +00003419** Note that the loops might not be nested in the order in which they
3420** appear in the FROM clause if a different order is better able to make
drh51147ba2005-07-23 22:59:55 +00003421** use of indices. Note also that when the IN operator appears in
3422** the WHERE clause, it might result in additional nested loops for
3423** scanning through all values on the right-hand side of the IN.
drh29dda4a2005-07-21 18:23:20 +00003424**
drhc27a1ce2002-06-14 20:58:45 +00003425** There are Btree cursors associated with each table. t1 uses cursor
drh6a3ea0e2003-05-02 14:32:12 +00003426** number pTabList->a[0].iCursor. t2 uses the cursor pTabList->a[1].iCursor.
3427** And so forth. This routine generates code to open those VDBE cursors
danielk19774adee202004-05-08 08:23:19 +00003428** and sqlite3WhereEnd() generates the code to close them.
drhc27a1ce2002-06-14 20:58:45 +00003429**
drhe6f85e72004-12-25 01:03:13 +00003430** The code that sqlite3WhereBegin() generates leaves the cursors named
3431** in pTabList pointing at their appropriate entries. The [...] code
drhf0863fe2005-06-12 21:35:51 +00003432** can use OP_Column and OP_Rowid opcodes on these cursors to extract
drhe6f85e72004-12-25 01:03:13 +00003433** data from the various tables of the loop.
3434**
drhc27a1ce2002-06-14 20:58:45 +00003435** If the WHERE clause is empty, the foreach loops must each scan their
3436** entire tables. Thus a three-way join is an O(N^3) operation. But if
3437** the tables have indices and there are terms in the WHERE clause that
3438** refer to those indices, a complete table scan can be avoided and the
3439** code will run much faster. Most of the work of this routine is checking
3440** to see if there are indices that can be used to speed up the loop.
3441**
3442** Terms of the WHERE clause are also used to limit which rows actually
3443** make it to the "..." in the middle of the loop. After each "foreach",
3444** terms of the WHERE clause that use only terms in that loop and outer
3445** loops are evaluated and if false a jump is made around all subsequent
3446** inner loops (or around the "..." if the test occurs within the inner-
3447** most loop)
3448**
3449** OUTER JOINS
3450**
3451** An outer join of tables t1 and t2 is conceptally coded as follows:
3452**
3453** foreach row1 in t1 do
3454** flag = 0
3455** foreach row2 in t2 do
3456** start:
3457** ...
3458** flag = 1
3459** end
drhe3184742002-06-19 14:27:05 +00003460** if flag==0 then
3461** move the row2 cursor to a null row
3462** goto start
3463** fi
drhc27a1ce2002-06-14 20:58:45 +00003464** end
3465**
drhe3184742002-06-19 14:27:05 +00003466** ORDER BY CLAUSE PROCESSING
3467**
3468** *ppOrderBy is a pointer to the ORDER BY clause of a SELECT statement,
3469** if there is one. If there is no ORDER BY clause or if this routine
3470** is called from an UPDATE or DELETE statement, then ppOrderBy is NULL.
3471**
3472** If an index can be used so that the natural output order of the table
3473** scan is correct for the ORDER BY clause, then that index is used and
3474** *ppOrderBy is set to NULL. This is an optimization that prevents an
3475** unnecessary sort of the result set if an index appropriate for the
3476** ORDER BY clause already exists.
3477**
3478** If the where clause loops cannot be arranged to provide the correct
3479** output order, then the *ppOrderBy is unchanged.
drh75897232000-05-29 14:26:00 +00003480*/
danielk19774adee202004-05-08 08:23:19 +00003481WhereInfo *sqlite3WhereBegin(
danielk1977ed326d72004-11-16 15:50:19 +00003482 Parse *pParse, /* The parser context */
3483 SrcList *pTabList, /* A list of all tables to be scanned */
3484 Expr *pWhere, /* The WHERE clause */
danielk1977a9d1ccb2008-01-05 17:39:29 +00003485 ExprList **ppOrderBy, /* An ORDER BY clause, or NULL */
drh336a5302009-04-24 15:46:21 +00003486 u16 wctrlFlags /* One of the WHERE_* flags defined in sqliteInt.h */
drh75897232000-05-29 14:26:00 +00003487){
3488 int i; /* Loop counter */
danielk1977be229652009-03-20 14:18:51 +00003489 int nByteWInfo; /* Num. bytes allocated for WhereInfo struct */
drh75897232000-05-29 14:26:00 +00003490 WhereInfo *pWInfo; /* Will become the return value of this function */
3491 Vdbe *v = pParse->pVdbe; /* The virtual database engine */
drhfe05af82005-07-21 03:14:59 +00003492 Bitmask notReady; /* Cursors that are not yet positioned */
drh111a6a72008-12-21 03:51:16 +00003493 WhereMaskSet *pMaskSet; /* The expression mask set */
drh111a6a72008-12-21 03:51:16 +00003494 WhereClause *pWC; /* Decomposition of the WHERE clause */
drh9012bcb2004-12-19 00:11:35 +00003495 struct SrcList_item *pTabItem; /* A single entry from pTabList */
3496 WhereLevel *pLevel; /* A single level in the pWInfo list */
drh29dda4a2005-07-21 18:23:20 +00003497 int iFrom; /* First unused FROM clause element */
drh111a6a72008-12-21 03:51:16 +00003498 int andFlags; /* AND-ed combination of all pWC->a[].wtFlags */
drh17435752007-08-16 04:30:38 +00003499 sqlite3 *db; /* Database connection */
drh75897232000-05-29 14:26:00 +00003500
drh29dda4a2005-07-21 18:23:20 +00003501 /* The number of tables in the FROM clause is limited by the number of
drh1398ad32005-01-19 23:24:50 +00003502 ** bits in a Bitmask
3503 */
drh29dda4a2005-07-21 18:23:20 +00003504 if( pTabList->nSrc>BMS ){
3505 sqlite3ErrorMsg(pParse, "at most %d tables in a join", BMS);
drh1398ad32005-01-19 23:24:50 +00003506 return 0;
3507 }
3508
drh75897232000-05-29 14:26:00 +00003509 /* Allocate and initialize the WhereInfo structure that will become the
danielk1977be229652009-03-20 14:18:51 +00003510 ** return value. A single allocation is used to store the WhereInfo
3511 ** struct, the contents of WhereInfo.a[], the WhereClause structure
3512 ** and the WhereMaskSet structure. Since WhereClause contains an 8-byte
3513 ** field (type Bitmask) it must be aligned on an 8-byte boundary on
3514 ** some architectures. Hence the ROUND8() below.
drh75897232000-05-29 14:26:00 +00003515 */
drh17435752007-08-16 04:30:38 +00003516 db = pParse->db;
danielk1977be229652009-03-20 14:18:51 +00003517 nByteWInfo = ROUND8(sizeof(WhereInfo)+(pTabList->nSrc-1)*sizeof(WhereLevel));
3518 pWInfo = sqlite3DbMallocZero(db,
3519 nByteWInfo +
3520 sizeof(WhereClause) +
3521 sizeof(WhereMaskSet)
3522 );
drh17435752007-08-16 04:30:38 +00003523 if( db->mallocFailed ){
danielk197785574e32008-10-06 05:32:18 +00003524 goto whereBeginError;
drh75897232000-05-29 14:26:00 +00003525 }
danielk197770b6d572006-06-19 04:49:34 +00003526 pWInfo->nLevel = pTabList->nSrc;
drh75897232000-05-29 14:26:00 +00003527 pWInfo->pParse = pParse;
3528 pWInfo->pTabList = pTabList;
danielk19774adee202004-05-08 08:23:19 +00003529 pWInfo->iBreak = sqlite3VdbeMakeLabel(v);
danielk1977be229652009-03-20 14:18:51 +00003530 pWInfo->pWC = pWC = (WhereClause *)&((u8 *)pWInfo)[nByteWInfo];
drh6df2acd2008-12-28 16:55:25 +00003531 pWInfo->wctrlFlags = wctrlFlags;
drh111a6a72008-12-21 03:51:16 +00003532 pMaskSet = (WhereMaskSet*)&pWC[1];
drh08192d52002-04-30 19:20:28 +00003533
drh111a6a72008-12-21 03:51:16 +00003534 /* Split the WHERE clause into separate subexpressions where each
3535 ** subexpression is separated by an AND operator.
3536 */
3537 initMaskSet(pMaskSet);
3538 whereClauseInit(pWC, pParse, pMaskSet);
3539 sqlite3ExprCodeConstants(pParse, pWhere);
3540 whereSplit(pWC, pWhere, TK_AND);
3541
drh08192d52002-04-30 19:20:28 +00003542 /* Special case: a WHERE clause that is constant. Evaluate the
3543 ** expression and either jump over all of the code or fall thru.
3544 */
drh0a168372007-06-08 00:20:47 +00003545 if( pWhere && (pTabList->nSrc==0 || sqlite3ExprIsConstantNotJoin(pWhere)) ){
drh35573352008-01-08 23:54:25 +00003546 sqlite3ExprIfFalse(pParse, pWhere, pWInfo->iBreak, SQLITE_JUMPIFNULL);
drhdf199a22002-06-14 22:38:41 +00003547 pWhere = 0;
drh08192d52002-04-30 19:20:28 +00003548 }
drh75897232000-05-29 14:26:00 +00003549
drh42165be2008-03-26 14:56:34 +00003550 /* Assign a bit from the bitmask to every term in the FROM clause.
3551 **
3552 ** When assigning bitmask values to FROM clause cursors, it must be
3553 ** the case that if X is the bitmask for the N-th FROM clause term then
3554 ** the bitmask for all FROM clause terms to the left of the N-th term
3555 ** is (X-1). An expression from the ON clause of a LEFT JOIN can use
3556 ** its Expr.iRightJoinTable value to find the bitmask of the right table
3557 ** of the join. Subtracting one from the right table bitmask gives a
3558 ** bitmask for all tables to the left of the join. Knowing the bitmask
3559 ** for all tables to the left of a left join is important. Ticket #3015.
danielk1977e672c8e2009-05-22 15:43:26 +00003560 **
3561 ** Configure the WhereClause.vmask variable so that bits that correspond
3562 ** to virtual table cursors are set. This is used to selectively disable
3563 ** the OR-to-IN transformation in exprAnalyzeOrTerm(). It is not helpful
3564 ** with virtual tables.
drh42165be2008-03-26 14:56:34 +00003565 */
danielk1977e672c8e2009-05-22 15:43:26 +00003566 assert( pWC->vmask==0 && pMaskSet->n==0 );
drh42165be2008-03-26 14:56:34 +00003567 for(i=0; i<pTabList->nSrc; i++){
drh111a6a72008-12-21 03:51:16 +00003568 createMask(pMaskSet, pTabList->a[i].iCursor);
shanee26fa4c2009-06-16 14:15:22 +00003569#ifndef SQLITE_OMIT_VIRTUALTABLE
drh2c1a0c52009-06-11 17:04:28 +00003570 if( ALWAYS(pTabList->a[i].pTab) && IsVirtual(pTabList->a[i].pTab) ){
danielk1977e672c8e2009-05-22 15:43:26 +00003571 pWC->vmask |= ((Bitmask)1 << i);
3572 }
shanee26fa4c2009-06-16 14:15:22 +00003573#endif
drh42165be2008-03-26 14:56:34 +00003574 }
3575#ifndef NDEBUG
3576 {
3577 Bitmask toTheLeft = 0;
3578 for(i=0; i<pTabList->nSrc; i++){
drh111a6a72008-12-21 03:51:16 +00003579 Bitmask m = getMask(pMaskSet, pTabList->a[i].iCursor);
drh42165be2008-03-26 14:56:34 +00003580 assert( (m-1)==toTheLeft );
3581 toTheLeft |= m;
3582 }
3583 }
3584#endif
3585
drh29dda4a2005-07-21 18:23:20 +00003586 /* Analyze all of the subexpressions. Note that exprAnalyze() might
3587 ** add new virtual terms onto the end of the WHERE clause. We do not
3588 ** want to analyze these virtual terms, so start analyzing at the end
drhb6fb62d2005-09-20 08:47:20 +00003589 ** and work forward so that the added virtual terms are never processed.
drh75897232000-05-29 14:26:00 +00003590 */
drh111a6a72008-12-21 03:51:16 +00003591 exprAnalyzeAll(pTabList, pWC);
drh17435752007-08-16 04:30:38 +00003592 if( db->mallocFailed ){
danielk197785574e32008-10-06 05:32:18 +00003593 goto whereBeginError;
drh0bbaa1b2005-08-19 19:14:12 +00003594 }
drh75897232000-05-29 14:26:00 +00003595
drh29dda4a2005-07-21 18:23:20 +00003596 /* Chose the best index to use for each table in the FROM clause.
3597 **
drh51147ba2005-07-23 22:59:55 +00003598 ** This loop fills in the following fields:
3599 **
3600 ** pWInfo->a[].pIdx The index to use for this level of the loop.
drh165be382008-12-05 02:36:33 +00003601 ** pWInfo->a[].wsFlags WHERE_xxx flags associated with pIdx
drh51147ba2005-07-23 22:59:55 +00003602 ** pWInfo->a[].nEq The number of == and IN constraints
danielk197785574e32008-10-06 05:32:18 +00003603 ** pWInfo->a[].iFrom Which term of the FROM clause is being coded
drh51147ba2005-07-23 22:59:55 +00003604 ** pWInfo->a[].iTabCur The VDBE cursor for the database table
3605 ** pWInfo->a[].iIdxCur The VDBE cursor for the index
drh111a6a72008-12-21 03:51:16 +00003606 ** pWInfo->a[].pTerm When wsFlags==WO_OR, the OR-clause term
drh51147ba2005-07-23 22:59:55 +00003607 **
3608 ** This loop also figures out the nesting order of tables in the FROM
3609 ** clause.
drh75897232000-05-29 14:26:00 +00003610 */
drhfe05af82005-07-21 03:14:59 +00003611 notReady = ~(Bitmask)0;
drh9012bcb2004-12-19 00:11:35 +00003612 pTabItem = pTabList->a;
3613 pLevel = pWInfo->a;
drh943af3c2005-07-29 19:43:58 +00003614 andFlags = ~0;
drh4f0c5872007-03-26 22:05:01 +00003615 WHERETRACE(("*** Optimizer Start ***\n"));
drh29dda4a2005-07-21 18:23:20 +00003616 for(i=iFrom=0, pLevel=pWInfo->a; i<pTabList->nSrc; i++, pLevel++){
drh111a6a72008-12-21 03:51:16 +00003617 WhereCost bestPlan; /* Most efficient plan seen so far */
drh29dda4a2005-07-21 18:23:20 +00003618 Index *pIdx; /* Index for FROM table at pTabItem */
drh29dda4a2005-07-21 18:23:20 +00003619 int j; /* For looping over FROM tables */
dan5236ac12009-08-13 07:09:33 +00003620 int bestJ = -1; /* The value of j */
drh29dda4a2005-07-21 18:23:20 +00003621 Bitmask m; /* Bitmask value for j or bestJ */
dan5236ac12009-08-13 07:09:33 +00003622 int isOptimal; /* Iterator for optimal/non-optimal search */
drh29dda4a2005-07-21 18:23:20 +00003623
drh111a6a72008-12-21 03:51:16 +00003624 memset(&bestPlan, 0, sizeof(bestPlan));
3625 bestPlan.rCost = SQLITE_BIG_DBL;
drhdf26fd52006-06-06 11:45:54 +00003626
dan5236ac12009-08-13 07:09:33 +00003627 /* Loop through the remaining entries in the FROM clause to find the
3628 ** next nested loop. The FROM clause entries may be iterated through
3629 ** either once or twice.
3630 **
3631 ** The first iteration, which is always performed, searches for the
3632 ** FROM clause entry that permits the lowest-cost, "optimal" scan. In
3633 ** this context an optimal scan is one that uses the same strategy
3634 ** for the given FROM clause entry as would be selected if the entry
drhd0015162009-08-21 13:22:25 +00003635 ** were used as the innermost nested loop. In other words, a table
3636 ** is chosen such that the cost of running that table cannot be reduced
3637 ** by waiting for other tables to run first.
dan5236ac12009-08-13 07:09:33 +00003638 **
3639 ** The second iteration is only performed if no optimal scan strategies
3640 ** were found by the first. This iteration is used to search for the
3641 ** lowest cost scan overall.
3642 **
3643 ** Previous versions of SQLite performed only the second iteration -
3644 ** the next outermost loop was always that with the lowest overall
3645 ** cost. However, this meant that SQLite could select the wrong plan
3646 ** for scripts such as the following:
3647 **
3648 ** CREATE TABLE t1(a, b);
3649 ** CREATE TABLE t2(c, d);
3650 ** SELECT * FROM t2, t1 WHERE t2.rowid = t1.a;
3651 **
3652 ** The best strategy is to iterate through table t1 first. However it
3653 ** is not possible to determine this with a simple greedy algorithm.
3654 ** However, since the cost of a linear scan through table t2 is the same
3655 ** as the cost of a linear scan through table t1, a simple greedy
3656 ** algorithm may choose to use t2 for the outer loop, which is a much
3657 ** costlier approach.
3658 */
3659 for(isOptimal=1; isOptimal>=0 && bestJ<0; isOptimal--){
3660 Bitmask mask = (isOptimal ? 0 : notReady);
3661 assert( (pTabList->nSrc-iFrom)>1 || isOptimal );
3662 for(j=iFrom, pTabItem=&pTabList->a[j]; j<pTabList->nSrc; j++, pTabItem++){
3663 int doNotReorder; /* True if this table should not be reordered */
3664 WhereCost sCost; /* Cost information from best[Virtual]Index() */
3665 ExprList *pOrderBy; /* ORDER BY clause for index to optimize */
3666
3667 doNotReorder = (pTabItem->jointype & (JT_LEFT|JT_CROSS))!=0;
3668 if( j!=iFrom && doNotReorder ) break;
3669 m = getMask(pMaskSet, pTabItem->iCursor);
3670 if( (m & notReady)==0 ){
3671 if( j==iFrom ) iFrom++;
3672 continue;
3673 }
3674 pOrderBy = ((i==0 && ppOrderBy )?*ppOrderBy:0);
3675
3676 assert( pTabItem->pTab );
drh9eff6162006-06-12 21:59:13 +00003677#ifndef SQLITE_OMIT_VIRTUALTABLE
dan5236ac12009-08-13 07:09:33 +00003678 if( IsVirtual(pTabItem->pTab) ){
3679 sqlite3_index_info **pp = &pWInfo->a[j].pIdxInfo;
3680 bestVirtualIndex(pParse, pWC, pTabItem, mask, pOrderBy, &sCost, pp);
3681 }else
drh9eff6162006-06-12 21:59:13 +00003682#endif
dan5236ac12009-08-13 07:09:33 +00003683 {
3684 bestBtreeIndex(pParse, pWC, pTabItem, mask, pOrderBy, &sCost);
3685 }
3686 assert( isOptimal || (sCost.used&notReady)==0 );
3687
3688 if( (sCost.used&notReady)==0
3689 && (j==iFrom || sCost.rCost<bestPlan.rCost)
3690 ){
3691 bestPlan = sCost;
3692 bestJ = j;
3693 }
3694 if( doNotReorder ) break;
drh9eff6162006-06-12 21:59:13 +00003695 }
drh29dda4a2005-07-21 18:23:20 +00003696 }
dan5236ac12009-08-13 07:09:33 +00003697 assert( bestJ>=0 );
danielk1977992347f2008-12-30 09:45:45 +00003698 assert( notReady & getMask(pMaskSet, pTabList->a[bestJ].iCursor) );
drhcb041342008-06-12 00:07:29 +00003699 WHERETRACE(("*** Optimizer selects table %d for loop %d\n", bestJ,
drh3dec2232005-09-10 15:28:09 +00003700 pLevel-pWInfo->a));
drh111a6a72008-12-21 03:51:16 +00003701 if( (bestPlan.plan.wsFlags & WHERE_ORDERBY)!=0 ){
drhfe05af82005-07-21 03:14:59 +00003702 *ppOrderBy = 0;
drhc4a3c772001-04-04 11:48:57 +00003703 }
drh111a6a72008-12-21 03:51:16 +00003704 andFlags &= bestPlan.plan.wsFlags;
3705 pLevel->plan = bestPlan.plan;
3706 if( bestPlan.plan.wsFlags & WHERE_INDEXED ){
drh9012bcb2004-12-19 00:11:35 +00003707 pLevel->iIdxCur = pParse->nTab++;
drhfe05af82005-07-21 03:14:59 +00003708 }else{
3709 pLevel->iIdxCur = -1;
drh6b563442001-11-07 16:48:26 +00003710 }
drh111a6a72008-12-21 03:51:16 +00003711 notReady &= ~getMask(pMaskSet, pTabList->a[bestJ].iCursor);
shaned87897d2009-01-30 05:40:27 +00003712 pLevel->iFrom = (u8)bestJ;
danielk197785574e32008-10-06 05:32:18 +00003713
3714 /* Check that if the table scanned by this loop iteration had an
3715 ** INDEXED BY clause attached to it, that the named index is being
3716 ** used for the scan. If not, then query compilation has failed.
3717 ** Return an error.
3718 */
3719 pIdx = pTabList->a[bestJ].pIndex;
drh171256c2009-01-08 03:11:19 +00003720 if( pIdx ){
3721 if( (bestPlan.plan.wsFlags & WHERE_INDEXED)==0 ){
3722 sqlite3ErrorMsg(pParse, "cannot use index: %s", pIdx->zName);
3723 goto whereBeginError;
3724 }else{
3725 /* If an INDEXED BY clause is used, the bestIndex() function is
3726 ** guaranteed to find the index specified in the INDEXED BY clause
3727 ** if it find an index at all. */
3728 assert( bestPlan.plan.u.pIdx==pIdx );
3729 }
danielk197785574e32008-10-06 05:32:18 +00003730 }
drh75897232000-05-29 14:26:00 +00003731 }
drh4f0c5872007-03-26 22:05:01 +00003732 WHERETRACE(("*** Optimizer Finished ***\n"));
danielk19771d461462009-04-21 09:02:45 +00003733 if( pParse->nErr || db->mallocFailed ){
danielk197780442942008-12-24 11:25:39 +00003734 goto whereBeginError;
3735 }
drh75897232000-05-29 14:26:00 +00003736
drh943af3c2005-07-29 19:43:58 +00003737 /* If the total query only selects a single row, then the ORDER BY
3738 ** clause is irrelevant.
3739 */
3740 if( (andFlags & WHERE_UNIQUE)!=0 && ppOrderBy ){
3741 *ppOrderBy = 0;
3742 }
3743
drh08c88eb2008-04-10 13:33:18 +00003744 /* If the caller is an UPDATE or DELETE statement that is requesting
3745 ** to use a one-pass algorithm, determine if this is appropriate.
3746 ** The one-pass algorithm only works if the WHERE clause constraints
3747 ** the statement to update a single row.
3748 */
drh165be382008-12-05 02:36:33 +00003749 assert( (wctrlFlags & WHERE_ONEPASS_DESIRED)==0 || pWInfo->nLevel==1 );
3750 if( (wctrlFlags & WHERE_ONEPASS_DESIRED)!=0 && (andFlags & WHERE_UNIQUE)!=0 ){
drh08c88eb2008-04-10 13:33:18 +00003751 pWInfo->okOnePass = 1;
drh111a6a72008-12-21 03:51:16 +00003752 pWInfo->a[0].plan.wsFlags &= ~WHERE_IDX_ONLY;
drh08c88eb2008-04-10 13:33:18 +00003753 }
3754
drh9012bcb2004-12-19 00:11:35 +00003755 /* Open all tables in the pTabList and any indices selected for
3756 ** searching those tables.
3757 */
3758 sqlite3CodeVerifySchema(pParse, -1); /* Insert the cookie verifier Goto */
drh29dda4a2005-07-21 18:23:20 +00003759 for(i=0, pLevel=pWInfo->a; i<pTabList->nSrc; i++, pLevel++){
danielk1977da184232006-01-05 11:34:32 +00003760 Table *pTab; /* Table to open */
danielk1977da184232006-01-05 11:34:32 +00003761 int iDb; /* Index of database containing table/index */
drh9012bcb2004-12-19 00:11:35 +00003762
drhecc92422005-09-10 16:46:12 +00003763#ifndef SQLITE_OMIT_EXPLAIN
3764 if( pParse->explain==2 ){
3765 char *zMsg;
3766 struct SrcList_item *pItem = &pTabList->a[pLevel->iFrom];
danielk19771e536952007-08-16 10:09:01 +00003767 zMsg = sqlite3MPrintf(db, "TABLE %s", pItem->zName);
drhecc92422005-09-10 16:46:12 +00003768 if( pItem->zAlias ){
drh633e6d52008-07-28 19:34:53 +00003769 zMsg = sqlite3MAppendf(db, zMsg, "%s AS %s", zMsg, pItem->zAlias);
drhecc92422005-09-10 16:46:12 +00003770 }
drh111a6a72008-12-21 03:51:16 +00003771 if( (pLevel->plan.wsFlags & WHERE_INDEXED)!=0 ){
3772 zMsg = sqlite3MAppendf(db, zMsg, "%s WITH INDEX %s",
3773 zMsg, pLevel->plan.u.pIdx->zName);
drh46129af2008-12-30 16:18:47 +00003774 }else if( pLevel->plan.wsFlags & WHERE_MULTI_OR ){
3775 zMsg = sqlite3MAppendf(db, zMsg, "%s VIA MULTI-INDEX UNION", zMsg);
drh111a6a72008-12-21 03:51:16 +00003776 }else if( pLevel->plan.wsFlags & (WHERE_ROWID_EQ|WHERE_ROWID_RANGE) ){
drh633e6d52008-07-28 19:34:53 +00003777 zMsg = sqlite3MAppendf(db, zMsg, "%s USING PRIMARY KEY", zMsg);
drhecc92422005-09-10 16:46:12 +00003778 }
drh9eff6162006-06-12 21:59:13 +00003779#ifndef SQLITE_OMIT_VIRTUALTABLE
drh111a6a72008-12-21 03:51:16 +00003780 else if( (pLevel->plan.wsFlags & WHERE_VIRTUALTABLE)!=0 ){
3781 sqlite3_index_info *pVtabIdx = pLevel->plan.u.pVtabIdx;
drh633e6d52008-07-28 19:34:53 +00003782 zMsg = sqlite3MAppendf(db, zMsg, "%s VIRTUAL TABLE INDEX %d:%s", zMsg,
drh111a6a72008-12-21 03:51:16 +00003783 pVtabIdx->idxNum, pVtabIdx->idxStr);
drh9eff6162006-06-12 21:59:13 +00003784 }
3785#endif
drh111a6a72008-12-21 03:51:16 +00003786 if( pLevel->plan.wsFlags & WHERE_ORDERBY ){
drh633e6d52008-07-28 19:34:53 +00003787 zMsg = sqlite3MAppendf(db, zMsg, "%s ORDER BY", zMsg);
drhe2b39092006-04-21 09:38:36 +00003788 }
drh66a51672008-01-03 00:01:23 +00003789 sqlite3VdbeAddOp4(v, OP_Explain, i, pLevel->iFrom, 0, zMsg, P4_DYNAMIC);
drhecc92422005-09-10 16:46:12 +00003790 }
3791#endif /* SQLITE_OMIT_EXPLAIN */
drh29dda4a2005-07-21 18:23:20 +00003792 pTabItem = &pTabList->a[pLevel->iFrom];
drh9012bcb2004-12-19 00:11:35 +00003793 pTab = pTabItem->pTab;
danielk1977595a5232009-07-24 17:58:53 +00003794 iDb = sqlite3SchemaToIndex(db, pTab->pSchema);
drh7d10d5a2008-08-20 16:35:10 +00003795 if( (pTab->tabFlags & TF_Ephemeral)!=0 || pTab->pSelect ) continue;
drh9eff6162006-06-12 21:59:13 +00003796#ifndef SQLITE_OMIT_VIRTUALTABLE
drh111a6a72008-12-21 03:51:16 +00003797 if( (pLevel->plan.wsFlags & WHERE_VIRTUALTABLE)!=0 ){
danielk1977595a5232009-07-24 17:58:53 +00003798 const char *pVTab = (const char *)sqlite3GetVTable(db, pTab);
danielk197793626f42006-06-20 13:07:27 +00003799 int iCur = pTabItem->iCursor;
danielk1977595a5232009-07-24 17:58:53 +00003800 sqlite3VdbeAddOp4(v, OP_VOpen, iCur, 0, 0, pVTab, P4_VTAB);
drh9eff6162006-06-12 21:59:13 +00003801 }else
3802#endif
drh6df2acd2008-12-28 16:55:25 +00003803 if( (pLevel->plan.wsFlags & WHERE_IDX_ONLY)==0
3804 && (wctrlFlags & WHERE_OMIT_OPEN)==0 ){
drh08c88eb2008-04-10 13:33:18 +00003805 int op = pWInfo->okOnePass ? OP_OpenWrite : OP_OpenRead;
3806 sqlite3OpenTable(pParse, pTabItem->iCursor, iDb, pTab, op);
danielk197723432972008-11-17 16:42:00 +00003807 if( !pWInfo->okOnePass && pTab->nCol<BMS ){
danielk19779792eef2006-01-13 15:58:43 +00003808 Bitmask b = pTabItem->colUsed;
3809 int n = 0;
drh74161702006-02-24 02:53:49 +00003810 for(; b; b=b>>1, n++){}
shanec0688ea2009-03-05 03:48:06 +00003811 sqlite3VdbeChangeP4(v, sqlite3VdbeCurrentAddr(v)-1, SQLITE_INT_TO_PTR(n), P4_INT32);
danielk19779792eef2006-01-13 15:58:43 +00003812 assert( n<=pTab->nCol );
3813 }
danielk1977c00da102006-01-07 13:21:04 +00003814 }else{
3815 sqlite3TableLock(pParse, iDb, pTab->tnum, 0, pTab->zName);
drh9012bcb2004-12-19 00:11:35 +00003816 }
3817 pLevel->iTabCur = pTabItem->iCursor;
drh111a6a72008-12-21 03:51:16 +00003818 if( (pLevel->plan.wsFlags & WHERE_INDEXED)!=0 ){
3819 Index *pIx = pLevel->plan.u.pIdx;
danielk1977b3bf5562006-01-10 17:58:23 +00003820 KeyInfo *pKey = sqlite3IndexKeyinfo(pParse, pIx);
drh111a6a72008-12-21 03:51:16 +00003821 int iIdxCur = pLevel->iIdxCur;
danielk1977da184232006-01-05 11:34:32 +00003822 assert( pIx->pSchema==pTab->pSchema );
drh111a6a72008-12-21 03:51:16 +00003823 assert( iIdxCur>=0 );
danielk1977207872a2008-01-03 07:54:23 +00003824 sqlite3VdbeAddOp4(v, OP_OpenRead, iIdxCur, pIx->tnum, iDb,
drh66a51672008-01-03 00:01:23 +00003825 (char*)pKey, P4_KEYINFO_HANDOFF);
danielk1977207872a2008-01-03 07:54:23 +00003826 VdbeComment((v, "%s", pIx->zName));
drh9012bcb2004-12-19 00:11:35 +00003827 }
danielk1977da184232006-01-05 11:34:32 +00003828 sqlite3CodeVerifySchema(pParse, iDb);
drh9012bcb2004-12-19 00:11:35 +00003829 }
3830 pWInfo->iTop = sqlite3VdbeCurrentAddr(v);
3831
drh29dda4a2005-07-21 18:23:20 +00003832 /* Generate the code to do the search. Each iteration of the for
3833 ** loop below generates code for a single nested loop of the VM
3834 ** program.
drh75897232000-05-29 14:26:00 +00003835 */
drhfe05af82005-07-21 03:14:59 +00003836 notReady = ~(Bitmask)0;
drh111a6a72008-12-21 03:51:16 +00003837 for(i=0; i<pTabList->nSrc; i++){
3838 notReady = codeOneLoopStart(pWInfo, i, wctrlFlags, notReady);
drh813f31e2009-01-06 00:08:02 +00003839 pWInfo->iContinue = pWInfo->a[i].addrCont;
drh75897232000-05-29 14:26:00 +00003840 }
drh7ec764a2005-07-21 03:48:20 +00003841
3842#ifdef SQLITE_TEST /* For testing and debugging use only */
3843 /* Record in the query plan information about the current table
3844 ** and the index used to access it (if any). If the table itself
3845 ** is not used, its name is just '{}'. If no index is used
3846 ** the index is listed as "{}". If the primary key is used the
3847 ** index name is '*'.
3848 */
3849 for(i=0; i<pTabList->nSrc; i++){
3850 char *z;
3851 int n;
drh7ec764a2005-07-21 03:48:20 +00003852 pLevel = &pWInfo->a[i];
drh29dda4a2005-07-21 18:23:20 +00003853 pTabItem = &pTabList->a[pLevel->iFrom];
drh7ec764a2005-07-21 03:48:20 +00003854 z = pTabItem->zAlias;
3855 if( z==0 ) z = pTabItem->pTab->zName;
drhea678832008-12-10 19:26:22 +00003856 n = sqlite3Strlen30(z);
drh7ec764a2005-07-21 03:48:20 +00003857 if( n+nQPlan < sizeof(sqlite3_query_plan)-10 ){
drh111a6a72008-12-21 03:51:16 +00003858 if( pLevel->plan.wsFlags & WHERE_IDX_ONLY ){
drh5bb3eb92007-05-04 13:15:55 +00003859 memcpy(&sqlite3_query_plan[nQPlan], "{}", 2);
drh7ec764a2005-07-21 03:48:20 +00003860 nQPlan += 2;
3861 }else{
drh5bb3eb92007-05-04 13:15:55 +00003862 memcpy(&sqlite3_query_plan[nQPlan], z, n);
drh7ec764a2005-07-21 03:48:20 +00003863 nQPlan += n;
3864 }
3865 sqlite3_query_plan[nQPlan++] = ' ';
3866 }
drh111a6a72008-12-21 03:51:16 +00003867 testcase( pLevel->plan.wsFlags & WHERE_ROWID_EQ );
3868 testcase( pLevel->plan.wsFlags & WHERE_ROWID_RANGE );
3869 if( pLevel->plan.wsFlags & (WHERE_ROWID_EQ|WHERE_ROWID_RANGE) ){
drh5bb3eb92007-05-04 13:15:55 +00003870 memcpy(&sqlite3_query_plan[nQPlan], "* ", 2);
drh7ec764a2005-07-21 03:48:20 +00003871 nQPlan += 2;
drh111a6a72008-12-21 03:51:16 +00003872 }else if( (pLevel->plan.wsFlags & WHERE_INDEXED)!=0 ){
3873 n = sqlite3Strlen30(pLevel->plan.u.pIdx->zName);
drh7ec764a2005-07-21 03:48:20 +00003874 if( n+nQPlan < sizeof(sqlite3_query_plan)-2 ){
drh111a6a72008-12-21 03:51:16 +00003875 memcpy(&sqlite3_query_plan[nQPlan], pLevel->plan.u.pIdx->zName, n);
drh7ec764a2005-07-21 03:48:20 +00003876 nQPlan += n;
3877 sqlite3_query_plan[nQPlan++] = ' ';
3878 }
drh111a6a72008-12-21 03:51:16 +00003879 }else{
3880 memcpy(&sqlite3_query_plan[nQPlan], "{} ", 3);
3881 nQPlan += 3;
drh7ec764a2005-07-21 03:48:20 +00003882 }
3883 }
3884 while( nQPlan>0 && sqlite3_query_plan[nQPlan-1]==' ' ){
3885 sqlite3_query_plan[--nQPlan] = 0;
3886 }
3887 sqlite3_query_plan[nQPlan] = 0;
3888 nQPlan = 0;
3889#endif /* SQLITE_TEST // Testing and debugging use only */
3890
drh29dda4a2005-07-21 18:23:20 +00003891 /* Record the continuation address in the WhereInfo structure. Then
3892 ** clean up and return.
3893 */
drh75897232000-05-29 14:26:00 +00003894 return pWInfo;
drhe23399f2005-07-22 00:31:39 +00003895
3896 /* Jump here if malloc fails */
danielk197785574e32008-10-06 05:32:18 +00003897whereBeginError:
drh10fe8402008-10-11 16:47:35 +00003898 whereInfoFree(db, pWInfo);
drhe23399f2005-07-22 00:31:39 +00003899 return 0;
drh75897232000-05-29 14:26:00 +00003900}
3901
3902/*
drhc27a1ce2002-06-14 20:58:45 +00003903** Generate the end of the WHERE loop. See comments on
danielk19774adee202004-05-08 08:23:19 +00003904** sqlite3WhereBegin() for additional information.
drh75897232000-05-29 14:26:00 +00003905*/
danielk19774adee202004-05-08 08:23:19 +00003906void sqlite3WhereEnd(WhereInfo *pWInfo){
drh633e6d52008-07-28 19:34:53 +00003907 Parse *pParse = pWInfo->pParse;
3908 Vdbe *v = pParse->pVdbe;
drh19a775c2000-06-05 18:54:46 +00003909 int i;
drh6b563442001-11-07 16:48:26 +00003910 WhereLevel *pLevel;
drhad3cab52002-05-24 02:04:32 +00003911 SrcList *pTabList = pWInfo->pTabList;
drh633e6d52008-07-28 19:34:53 +00003912 sqlite3 *db = pParse->db;
drh19a775c2000-06-05 18:54:46 +00003913
drh9012bcb2004-12-19 00:11:35 +00003914 /* Generate loop termination code.
3915 */
drhceea3322009-04-23 13:22:42 +00003916 sqlite3ExprCacheClear(pParse);
drhad3cab52002-05-24 02:04:32 +00003917 for(i=pTabList->nSrc-1; i>=0; i--){
drh6b563442001-11-07 16:48:26 +00003918 pLevel = &pWInfo->a[i];
drhb3190c12008-12-08 21:37:14 +00003919 sqlite3VdbeResolveLabel(v, pLevel->addrCont);
drh6b563442001-11-07 16:48:26 +00003920 if( pLevel->op!=OP_Noop ){
drh66a51672008-01-03 00:01:23 +00003921 sqlite3VdbeAddOp2(v, pLevel->op, pLevel->p1, pLevel->p2);
drhd1d38482008-10-07 23:46:38 +00003922 sqlite3VdbeChangeP5(v, pLevel->p5);
drh19a775c2000-06-05 18:54:46 +00003923 }
drh111a6a72008-12-21 03:51:16 +00003924 if( pLevel->plan.wsFlags & WHERE_IN_ABLE && pLevel->u.in.nIn>0 ){
drh72e8fa42007-03-28 14:30:06 +00003925 struct InLoop *pIn;
drhe23399f2005-07-22 00:31:39 +00003926 int j;
drhb3190c12008-12-08 21:37:14 +00003927 sqlite3VdbeResolveLabel(v, pLevel->addrNxt);
drh111a6a72008-12-21 03:51:16 +00003928 for(j=pLevel->u.in.nIn, pIn=&pLevel->u.in.aInLoop[j-1]; j>0; j--, pIn--){
drhb3190c12008-12-08 21:37:14 +00003929 sqlite3VdbeJumpHere(v, pIn->addrInTop+1);
3930 sqlite3VdbeAddOp2(v, OP_Next, pIn->iCur, pIn->addrInTop);
3931 sqlite3VdbeJumpHere(v, pIn->addrInTop-1);
drhe23399f2005-07-22 00:31:39 +00003932 }
drh111a6a72008-12-21 03:51:16 +00003933 sqlite3DbFree(db, pLevel->u.in.aInLoop);
drhd99f7062002-06-08 23:25:08 +00003934 }
drhb3190c12008-12-08 21:37:14 +00003935 sqlite3VdbeResolveLabel(v, pLevel->addrBrk);
drhad2d8302002-05-24 20:31:36 +00003936 if( pLevel->iLeftJoin ){
3937 int addr;
drh3c84ddf2008-01-09 02:15:38 +00003938 addr = sqlite3VdbeAddOp1(v, OP_IfPos, pLevel->iLeftJoin);
3939 sqlite3VdbeAddOp1(v, OP_NullRow, pTabList->a[i].iCursor);
drh9012bcb2004-12-19 00:11:35 +00003940 if( pLevel->iIdxCur>=0 ){
drh3c84ddf2008-01-09 02:15:38 +00003941 sqlite3VdbeAddOp1(v, OP_NullRow, pLevel->iIdxCur);
drh7f09b3e2002-08-13 13:15:49 +00003942 }
drh336a5302009-04-24 15:46:21 +00003943 if( pLevel->op==OP_Return ){
3944 sqlite3VdbeAddOp2(v, OP_Gosub, pLevel->p1, pLevel->addrFirst);
3945 }else{
3946 sqlite3VdbeAddOp2(v, OP_Goto, 0, pLevel->addrFirst);
3947 }
drhd654be82005-09-20 17:42:23 +00003948 sqlite3VdbeJumpHere(v, addr);
drhad2d8302002-05-24 20:31:36 +00003949 }
drh19a775c2000-06-05 18:54:46 +00003950 }
drh9012bcb2004-12-19 00:11:35 +00003951
3952 /* The "break" point is here, just past the end of the outer loop.
3953 ** Set it.
3954 */
danielk19774adee202004-05-08 08:23:19 +00003955 sqlite3VdbeResolveLabel(v, pWInfo->iBreak);
drh9012bcb2004-12-19 00:11:35 +00003956
drh29dda4a2005-07-21 18:23:20 +00003957 /* Close all of the cursors that were opened by sqlite3WhereBegin.
drh9012bcb2004-12-19 00:11:35 +00003958 */
drh29dda4a2005-07-21 18:23:20 +00003959 for(i=0, pLevel=pWInfo->a; i<pTabList->nSrc; i++, pLevel++){
3960 struct SrcList_item *pTabItem = &pTabList->a[pLevel->iFrom];
drh9012bcb2004-12-19 00:11:35 +00003961 Table *pTab = pTabItem->pTab;
drh5cf590c2003-04-24 01:45:04 +00003962 assert( pTab!=0 );
drh7d10d5a2008-08-20 16:35:10 +00003963 if( (pTab->tabFlags & TF_Ephemeral)!=0 || pTab->pSelect ) continue;
drh6df2acd2008-12-28 16:55:25 +00003964 if( (pWInfo->wctrlFlags & WHERE_OMIT_CLOSE)==0 ){
3965 if( !pWInfo->okOnePass && (pLevel->plan.wsFlags & WHERE_IDX_ONLY)==0 ){
3966 sqlite3VdbeAddOp1(v, OP_Close, pTabItem->iCursor);
3967 }
3968 if( (pLevel->plan.wsFlags & WHERE_INDEXED)!=0 ){
3969 sqlite3VdbeAddOp1(v, OP_Close, pLevel->iIdxCur);
3970 }
drh9012bcb2004-12-19 00:11:35 +00003971 }
3972
danielk197721de2e72007-11-29 17:43:27 +00003973 /* If this scan uses an index, make code substitutions to read data
3974 ** from the index in preference to the table. Sometimes, this means
3975 ** the table need never be read from. This is a performance boost,
3976 ** as the vdbe level waits until the table is read before actually
3977 ** seeking the table cursor to the record corresponding to the current
3978 ** position in the index.
drh9012bcb2004-12-19 00:11:35 +00003979 **
3980 ** Calls to the code generator in between sqlite3WhereBegin and
3981 ** sqlite3WhereEnd will have created code that references the table
3982 ** directly. This loop scans all that code looking for opcodes
3983 ** that reference the table and converts them into opcodes that
3984 ** reference the index.
3985 */
drh125feff2009-06-06 15:17:27 +00003986 if( (pLevel->plan.wsFlags & WHERE_INDEXED)!=0 && !db->mallocFailed){
danielk1977f0113002006-01-24 12:09:17 +00003987 int k, j, last;
drh9012bcb2004-12-19 00:11:35 +00003988 VdbeOp *pOp;
drh111a6a72008-12-21 03:51:16 +00003989 Index *pIdx = pLevel->plan.u.pIdx;
3990 int useIndexOnly = pLevel->plan.wsFlags & WHERE_IDX_ONLY;
drh9012bcb2004-12-19 00:11:35 +00003991
3992 assert( pIdx!=0 );
3993 pOp = sqlite3VdbeGetOp(v, pWInfo->iTop);
3994 last = sqlite3VdbeCurrentAddr(v);
danielk1977f0113002006-01-24 12:09:17 +00003995 for(k=pWInfo->iTop; k<last; k++, pOp++){
drh9012bcb2004-12-19 00:11:35 +00003996 if( pOp->p1!=pLevel->iTabCur ) continue;
3997 if( pOp->opcode==OP_Column ){
drh9012bcb2004-12-19 00:11:35 +00003998 for(j=0; j<pIdx->nColumn; j++){
3999 if( pOp->p2==pIdx->aiColumn[j] ){
4000 pOp->p2 = j;
danielk197721de2e72007-11-29 17:43:27 +00004001 pOp->p1 = pLevel->iIdxCur;
drh9012bcb2004-12-19 00:11:35 +00004002 break;
4003 }
4004 }
danielk197721de2e72007-11-29 17:43:27 +00004005 assert(!useIndexOnly || j<pIdx->nColumn);
drhf0863fe2005-06-12 21:35:51 +00004006 }else if( pOp->opcode==OP_Rowid ){
drh9012bcb2004-12-19 00:11:35 +00004007 pOp->p1 = pLevel->iIdxCur;
drhf0863fe2005-06-12 21:35:51 +00004008 pOp->opcode = OP_IdxRowid;
danielk197721de2e72007-11-29 17:43:27 +00004009 }else if( pOp->opcode==OP_NullRow && useIndexOnly ){
danielk19776c18b6e2005-01-30 09:17:58 +00004010 pOp->opcode = OP_Noop;
drh9012bcb2004-12-19 00:11:35 +00004011 }
4012 }
drh6b563442001-11-07 16:48:26 +00004013 }
drh19a775c2000-06-05 18:54:46 +00004014 }
drh9012bcb2004-12-19 00:11:35 +00004015
4016 /* Final cleanup
4017 */
drh10fe8402008-10-11 16:47:35 +00004018 whereInfoFree(db, pWInfo);
drh75897232000-05-29 14:26:00 +00004019 return;
4020}