add pragma cipher_default_use_hmac to toggle global HMAC setting
[sqlcipher.git] / src / where.c
blob05414da58b4cf02dcca176d42a25ab192eca4426
1 /*
2 ** 2001 September 15
3 **
4 ** The author disclaims copyright to this source code. In place of
5 ** a legal notice, here is a blessing:
6 **
7 ** May you do good and not evil.
8 ** May you find forgiveness for yourself and forgive others.
9 ** May you share freely, never taking more than you give.
11 *************************************************************************
12 ** This module contains C code that generates VDBE code used to process
13 ** the WHERE clause of SQL statements. This module is responsible for
14 ** 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".
19 #include "sqliteInt.h"
23 ** Trace output macros
25 #if defined(SQLITE_TEST) || defined(SQLITE_DEBUG)
26 int sqlite3WhereTrace = 0;
27 #endif
28 #if defined(SQLITE_TEST) && defined(SQLITE_DEBUG)
29 # define WHERETRACE(X) if(sqlite3WhereTrace) sqlite3DebugPrintf X
30 #else
31 # define WHERETRACE(X)
32 #endif
34 /* Forward reference
36 typedef struct WhereClause WhereClause;
37 typedef struct WhereMaskSet WhereMaskSet;
38 typedef struct WhereOrInfo WhereOrInfo;
39 typedef struct WhereAndInfo WhereAndInfo;
40 typedef struct WhereCost WhereCost;
43 ** The query generator uses an array of instances of this structure to
44 ** help it analyze the subexpressions of the WHERE clause. Each WHERE
45 ** clause subexpression is separated from the others by AND operators,
46 ** usually, or sometimes subexpressions separated by OR.
48 ** All WhereTerms are collected into a single WhereClause structure.
49 ** The following identity holds:
51 ** WhereTerm.pWC->a[WhereTerm.idx] == WhereTerm
53 ** When a term is of the form:
55 ** X <op> <expr>
57 ** where X is a column name and <op> is one of certain operators,
58 ** then WhereTerm.leftCursor and WhereTerm.u.leftColumn record the
59 ** cursor number and column number for X. WhereTerm.eOperator records
60 ** the <op> using a bitmask encoding defined by WO_xxx below. The
61 ** use of a bitmask encoding for the operator allows us to search
62 ** quickly for terms that match any of several different operators.
64 ** A WhereTerm might also be two or more subterms connected by OR:
66 ** (t1.X <op> <expr>) OR (t1.Y <op> <expr>) OR ....
68 ** In this second case, wtFlag as the TERM_ORINFO set and eOperator==WO_OR
69 ** and the WhereTerm.u.pOrInfo field points to auxiliary information that
70 ** is collected about the
72 ** If a term in the WHERE clause does not match either of the two previous
73 ** categories, then eOperator==0. The WhereTerm.pExpr field is still set
74 ** to the original subexpression content and wtFlags is set up appropriately
75 ** but no other fields in the WhereTerm object are meaningful.
77 ** When eOperator!=0, prereqRight and prereqAll record sets of cursor numbers,
78 ** but they do so indirectly. A single WhereMaskSet structure translates
79 ** cursor number into bits and the translated bit is stored in the prereq
80 ** fields. The translation is used in order to maximize the number of
81 ** bits that will fit in a Bitmask. The VDBE cursor numbers might be
82 ** spread out over the non-negative integers. For example, the cursor
83 ** numbers might be 3, 8, 9, 10, 20, 23, 41, and 45. The WhereMaskSet
84 ** translates these sparse cursor numbers into consecutive integers
85 ** beginning with 0 in order to make the best possible use of the available
86 ** bits in the Bitmask. So, in the example above, the cursor numbers
87 ** would be mapped into integers 0 through 7.
89 ** The number of terms in a join is limited by the number of bits
90 ** in prereqRight and prereqAll. The default is 64 bits, hence SQLite
91 ** is only able to process joins with 64 or fewer tables.
93 typedef struct WhereTerm WhereTerm;
94 struct WhereTerm {
95 Expr *pExpr; /* Pointer to the subexpression that is this term */
96 int iParent; /* Disable pWC->a[iParent] when this term disabled */
97 int leftCursor; /* Cursor number of X in "X <op> <expr>" */
98 union {
99 int leftColumn; /* Column number of X in "X <op> <expr>" */
100 WhereOrInfo *pOrInfo; /* Extra information if eOperator==WO_OR */
101 WhereAndInfo *pAndInfo; /* Extra information if eOperator==WO_AND */
102 } u;
103 u16 eOperator; /* A WO_xx value describing <op> */
104 u8 wtFlags; /* TERM_xxx bit flags. See below */
105 u8 nChild; /* Number of children that must disable us */
106 WhereClause *pWC; /* The clause this term is part of */
107 Bitmask prereqRight; /* Bitmask of tables used by pExpr->pRight */
108 Bitmask prereqAll; /* Bitmask of tables referenced by pExpr */
112 ** Allowed values of WhereTerm.wtFlags
114 #define TERM_DYNAMIC 0x01 /* Need to call sqlite3ExprDelete(db, pExpr) */
115 #define TERM_VIRTUAL 0x02 /* Added by the optimizer. Do not code */
116 #define TERM_CODED 0x04 /* This term is already coded */
117 #define TERM_COPIED 0x08 /* Has a child */
118 #define TERM_ORINFO 0x10 /* Need to free the WhereTerm.u.pOrInfo object */
119 #define TERM_ANDINFO 0x20 /* Need to free the WhereTerm.u.pAndInfo obj */
120 #define TERM_OR_OK 0x40 /* Used during OR-clause processing */
121 #ifdef SQLITE_ENABLE_STAT3
122 # define TERM_VNULL 0x80 /* Manufactured x>NULL or x<=NULL term */
123 #else
124 # define TERM_VNULL 0x00 /* Disabled if not using stat3 */
125 #endif
128 ** An instance of the following structure holds all information about a
129 ** WHERE clause. Mostly this is a container for one or more WhereTerms.
131 ** Explanation of pOuter: For a WHERE clause of the form
133 ** a AND ((b AND c) OR (d AND e)) AND f
135 ** There are separate WhereClause objects for the whole clause and for
136 ** the subclauses "(b AND c)" and "(d AND e)". The pOuter field of the
137 ** subclauses points to the WhereClause object for the whole clause.
139 struct WhereClause {
140 Parse *pParse; /* The parser context */
141 WhereMaskSet *pMaskSet; /* Mapping of table cursor numbers to bitmasks */
142 Bitmask vmask; /* Bitmask identifying virtual table cursors */
143 WhereClause *pOuter; /* Outer conjunction */
144 u8 op; /* Split operator. TK_AND or TK_OR */
145 u16 wctrlFlags; /* Might include WHERE_AND_ONLY */
146 int nTerm; /* Number of terms */
147 int nSlot; /* Number of entries in a[] */
148 WhereTerm *a; /* Each a[] describes a term of the WHERE cluase */
149 #if defined(SQLITE_SMALL_STACK)
150 WhereTerm aStatic[1]; /* Initial static space for a[] */
151 #else
152 WhereTerm aStatic[8]; /* Initial static space for a[] */
153 #endif
157 ** A WhereTerm with eOperator==WO_OR has its u.pOrInfo pointer set to
158 ** a dynamically allocated instance of the following structure.
160 struct WhereOrInfo {
161 WhereClause wc; /* Decomposition into subterms */
162 Bitmask indexable; /* Bitmask of all indexable tables in the clause */
166 ** A WhereTerm with eOperator==WO_AND has its u.pAndInfo pointer set to
167 ** a dynamically allocated instance of the following structure.
169 struct WhereAndInfo {
170 WhereClause wc; /* The subexpression broken out */
174 ** An instance of the following structure keeps track of a mapping
175 ** between VDBE cursor numbers and bits of the bitmasks in WhereTerm.
177 ** The VDBE cursor numbers are small integers contained in
178 ** SrcList_item.iCursor and Expr.iTable fields. For any given WHERE
179 ** clause, the cursor numbers might not begin with 0 and they might
180 ** contain gaps in the numbering sequence. But we want to make maximum
181 ** use of the bits in our bitmasks. This structure provides a mapping
182 ** from the sparse cursor numbers into consecutive integers beginning
183 ** with 0.
185 ** If WhereMaskSet.ix[A]==B it means that The A-th bit of a Bitmask
186 ** corresponds VDBE cursor number B. The A-th bit of a bitmask is 1<<A.
188 ** For example, if the WHERE clause expression used these VDBE
189 ** cursors: 4, 5, 8, 29, 57, 73. Then the WhereMaskSet structure
190 ** would map those cursor numbers into bits 0 through 5.
192 ** Note that the mapping is not necessarily ordered. In the example
193 ** above, the mapping might go like this: 4->3, 5->1, 8->2, 29->0,
194 ** 57->5, 73->4. Or one of 719 other combinations might be used. It
195 ** does not really matter. What is important is that sparse cursor
196 ** numbers all get mapped into bit numbers that begin with 0 and contain
197 ** no gaps.
199 struct WhereMaskSet {
200 int n; /* Number of assigned cursor values */
201 int ix[BMS]; /* Cursor assigned to each bit */
205 ** A WhereCost object records a lookup strategy and the estimated
206 ** cost of pursuing that strategy.
208 struct WhereCost {
209 WherePlan plan; /* The lookup strategy */
210 double rCost; /* Overall cost of pursuing this search strategy */
211 Bitmask used; /* Bitmask of cursors used by this plan */
215 ** Bitmasks for the operators that indices are able to exploit. An
216 ** OR-ed combination of these values can be used when searching for
217 ** terms in the where clause.
219 #define WO_IN 0x001
220 #define WO_EQ 0x002
221 #define WO_LT (WO_EQ<<(TK_LT-TK_EQ))
222 #define WO_LE (WO_EQ<<(TK_LE-TK_EQ))
223 #define WO_GT (WO_EQ<<(TK_GT-TK_EQ))
224 #define WO_GE (WO_EQ<<(TK_GE-TK_EQ))
225 #define WO_MATCH 0x040
226 #define WO_ISNULL 0x080
227 #define WO_OR 0x100 /* Two or more OR-connected terms */
228 #define WO_AND 0x200 /* Two or more AND-connected terms */
229 #define WO_NOOP 0x800 /* This term does not restrict search space */
231 #define WO_ALL 0xfff /* Mask of all possible WO_* values */
232 #define WO_SINGLE 0x0ff /* Mask of all non-compound WO_* values */
235 ** Value for wsFlags returned by bestIndex() and stored in
236 ** WhereLevel.wsFlags. These flags determine which search
237 ** strategies are appropriate.
239 ** The least significant 12 bits is reserved as a mask for WO_ values above.
240 ** The WhereLevel.wsFlags field is usually set to WO_IN|WO_EQ|WO_ISNULL.
241 ** But if the table is the right table of a left join, WhereLevel.wsFlags
242 ** is set to WO_IN|WO_EQ. The WhereLevel.wsFlags field can then be used as
243 ** the "op" parameter to findTerm when we are resolving equality constraints.
244 ** ISNULL constraints will then not be used on the right table of a left
245 ** join. Tickets #2177 and #2189.
247 #define WHERE_ROWID_EQ 0x00001000 /* rowid=EXPR or rowid IN (...) */
248 #define WHERE_ROWID_RANGE 0x00002000 /* rowid<EXPR and/or rowid>EXPR */
249 #define WHERE_COLUMN_EQ 0x00010000 /* x=EXPR or x IN (...) or x IS NULL */
250 #define WHERE_COLUMN_RANGE 0x00020000 /* x<EXPR and/or x>EXPR */
251 #define WHERE_COLUMN_IN 0x00040000 /* x IN (...) */
252 #define WHERE_COLUMN_NULL 0x00080000 /* x IS NULL */
253 #define WHERE_INDEXED 0x000f0000 /* Anything that uses an index */
254 #define WHERE_NOT_FULLSCAN 0x100f3000 /* Does not do a full table scan */
255 #define WHERE_IN_ABLE 0x000f1000 /* Able to support an IN operator */
256 #define WHERE_TOP_LIMIT 0x00100000 /* x<EXPR or x<=EXPR constraint */
257 #define WHERE_BTM_LIMIT 0x00200000 /* x>EXPR or x>=EXPR constraint */
258 #define WHERE_BOTH_LIMIT 0x00300000 /* Both x>EXPR and x<EXPR */
259 #define WHERE_IDX_ONLY 0x00800000 /* Use index only - omit table */
260 #define WHERE_ORDERBY 0x01000000 /* Output will appear in correct order */
261 #define WHERE_REVERSE 0x02000000 /* Scan in reverse order */
262 #define WHERE_UNIQUE 0x04000000 /* Selects no more than one row */
263 #define WHERE_VIRTUALTABLE 0x08000000 /* Use virtual-table processing */
264 #define WHERE_MULTI_OR 0x10000000 /* OR using multiple indices */
265 #define WHERE_TEMP_INDEX 0x20000000 /* Uses an ephemeral index */
266 #define WHERE_DISTINCT 0x40000000 /* Correct order for DISTINCT */
269 ** Initialize a preallocated WhereClause structure.
271 static void whereClauseInit(
272 WhereClause *pWC, /* The WhereClause to be initialized */
273 Parse *pParse, /* The parsing context */
274 WhereMaskSet *pMaskSet, /* Mapping from table cursor numbers to bitmasks */
275 u16 wctrlFlags /* Might include WHERE_AND_ONLY */
277 pWC->pParse = pParse;
278 pWC->pMaskSet = pMaskSet;
279 pWC->pOuter = 0;
280 pWC->nTerm = 0;
281 pWC->nSlot = ArraySize(pWC->aStatic);
282 pWC->a = pWC->aStatic;
283 pWC->vmask = 0;
284 pWC->wctrlFlags = wctrlFlags;
287 /* Forward reference */
288 static void whereClauseClear(WhereClause*);
291 ** Deallocate all memory associated with a WhereOrInfo object.
293 static void whereOrInfoDelete(sqlite3 *db, WhereOrInfo *p){
294 whereClauseClear(&p->wc);
295 sqlite3DbFree(db, p);
299 ** Deallocate all memory associated with a WhereAndInfo object.
301 static void whereAndInfoDelete(sqlite3 *db, WhereAndInfo *p){
302 whereClauseClear(&p->wc);
303 sqlite3DbFree(db, p);
307 ** Deallocate a WhereClause structure. The WhereClause structure
308 ** itself is not freed. This routine is the inverse of whereClauseInit().
310 static void whereClauseClear(WhereClause *pWC){
311 int i;
312 WhereTerm *a;
313 sqlite3 *db = pWC->pParse->db;
314 for(i=pWC->nTerm-1, a=pWC->a; i>=0; i--, a++){
315 if( a->wtFlags & TERM_DYNAMIC ){
316 sqlite3ExprDelete(db, a->pExpr);
318 if( a->wtFlags & TERM_ORINFO ){
319 whereOrInfoDelete(db, a->u.pOrInfo);
320 }else if( a->wtFlags & TERM_ANDINFO ){
321 whereAndInfoDelete(db, a->u.pAndInfo);
324 if( pWC->a!=pWC->aStatic ){
325 sqlite3DbFree(db, pWC->a);
330 ** Add a single new WhereTerm entry to the WhereClause object pWC.
331 ** The new WhereTerm object is constructed from Expr p and with wtFlags.
332 ** The index in pWC->a[] of the new WhereTerm is returned on success.
333 ** 0 is returned if the new WhereTerm could not be added due to a memory
334 ** allocation error. The memory allocation failure will be recorded in
335 ** the db->mallocFailed flag so that higher-level functions can detect it.
337 ** This routine will increase the size of the pWC->a[] array as necessary.
339 ** If the wtFlags argument includes TERM_DYNAMIC, then responsibility
340 ** for freeing the expression p is assumed by the WhereClause object pWC.
341 ** This is true even if this routine fails to allocate a new WhereTerm.
343 ** WARNING: This routine might reallocate the space used to store
344 ** WhereTerms. All pointers to WhereTerms should be invalidated after
345 ** calling this routine. Such pointers may be reinitialized by referencing
346 ** the pWC->a[] array.
348 static int whereClauseInsert(WhereClause *pWC, Expr *p, u8 wtFlags){
349 WhereTerm *pTerm;
350 int idx;
351 testcase( wtFlags & TERM_VIRTUAL ); /* EV: R-00211-15100 */
352 if( pWC->nTerm>=pWC->nSlot ){
353 WhereTerm *pOld = pWC->a;
354 sqlite3 *db = pWC->pParse->db;
355 pWC->a = sqlite3DbMallocRaw(db, sizeof(pWC->a[0])*pWC->nSlot*2 );
356 if( pWC->a==0 ){
357 if( wtFlags & TERM_DYNAMIC ){
358 sqlite3ExprDelete(db, p);
360 pWC->a = pOld;
361 return 0;
363 memcpy(pWC->a, pOld, sizeof(pWC->a[0])*pWC->nTerm);
364 if( pOld!=pWC->aStatic ){
365 sqlite3DbFree(db, pOld);
367 pWC->nSlot = sqlite3DbMallocSize(db, pWC->a)/sizeof(pWC->a[0]);
369 pTerm = &pWC->a[idx = pWC->nTerm++];
370 pTerm->pExpr = p;
371 pTerm->wtFlags = wtFlags;
372 pTerm->pWC = pWC;
373 pTerm->iParent = -1;
374 return idx;
378 ** This routine identifies subexpressions in the WHERE clause where
379 ** each subexpression is separated by the AND operator or some other
380 ** operator specified in the op parameter. The WhereClause structure
381 ** is filled with pointers to subexpressions. For example:
383 ** WHERE a=='hello' AND coalesce(b,11)<10 AND (c+12!=d OR c==22)
384 ** \________/ \_______________/ \________________/
385 ** slot[0] slot[1] slot[2]
387 ** The original WHERE clause in pExpr is unaltered. All this routine
388 ** does is make slot[] entries point to substructure within pExpr.
390 ** In the previous sentence and in the diagram, "slot[]" refers to
391 ** the WhereClause.a[] array. The slot[] array grows as needed to contain
392 ** all terms of the WHERE clause.
394 static void whereSplit(WhereClause *pWC, Expr *pExpr, int op){
395 pWC->op = (u8)op;
396 if( pExpr==0 ) return;
397 if( pExpr->op!=op ){
398 whereClauseInsert(pWC, pExpr, 0);
399 }else{
400 whereSplit(pWC, pExpr->pLeft, op);
401 whereSplit(pWC, pExpr->pRight, op);
406 ** Initialize an expression mask set (a WhereMaskSet object)
408 #define initMaskSet(P) memset(P, 0, sizeof(*P))
411 ** Return the bitmask for the given cursor number. Return 0 if
412 ** iCursor is not in the set.
414 static Bitmask getMask(WhereMaskSet *pMaskSet, int iCursor){
415 int i;
416 assert( pMaskSet->n<=(int)sizeof(Bitmask)*8 );
417 for(i=0; i<pMaskSet->n; i++){
418 if( pMaskSet->ix[i]==iCursor ){
419 return ((Bitmask)1)<<i;
422 return 0;
426 ** Create a new mask for cursor iCursor.
428 ** There is one cursor per table in the FROM clause. The number of
429 ** tables in the FROM clause is limited by a test early in the
430 ** sqlite3WhereBegin() routine. So we know that the pMaskSet->ix[]
431 ** array will never overflow.
433 static void createMask(WhereMaskSet *pMaskSet, int iCursor){
434 assert( pMaskSet->n < ArraySize(pMaskSet->ix) );
435 pMaskSet->ix[pMaskSet->n++] = iCursor;
439 ** This routine walks (recursively) an expression tree and generates
440 ** a bitmask indicating which tables are used in that expression
441 ** tree.
443 ** In order for this routine to work, the calling function must have
444 ** previously invoked sqlite3ResolveExprNames() on the expression. See
445 ** the header comment on that routine for additional information.
446 ** The sqlite3ResolveExprNames() routines looks for column names and
447 ** sets their opcodes to TK_COLUMN and their Expr.iTable fields to
448 ** the VDBE cursor number of the table. This routine just has to
449 ** translate the cursor numbers into bitmask values and OR all
450 ** the bitmasks together.
452 static Bitmask exprListTableUsage(WhereMaskSet*, ExprList*);
453 static Bitmask exprSelectTableUsage(WhereMaskSet*, Select*);
454 static Bitmask exprTableUsage(WhereMaskSet *pMaskSet, Expr *p){
455 Bitmask mask = 0;
456 if( p==0 ) return 0;
457 if( p->op==TK_COLUMN ){
458 mask = getMask(pMaskSet, p->iTable);
459 return mask;
461 mask = exprTableUsage(pMaskSet, p->pRight);
462 mask |= exprTableUsage(pMaskSet, p->pLeft);
463 if( ExprHasProperty(p, EP_xIsSelect) ){
464 mask |= exprSelectTableUsage(pMaskSet, p->x.pSelect);
465 }else{
466 mask |= exprListTableUsage(pMaskSet, p->x.pList);
468 return mask;
470 static Bitmask exprListTableUsage(WhereMaskSet *pMaskSet, ExprList *pList){
471 int i;
472 Bitmask mask = 0;
473 if( pList ){
474 for(i=0; i<pList->nExpr; i++){
475 mask |= exprTableUsage(pMaskSet, pList->a[i].pExpr);
478 return mask;
480 static Bitmask exprSelectTableUsage(WhereMaskSet *pMaskSet, Select *pS){
481 Bitmask mask = 0;
482 while( pS ){
483 SrcList *pSrc = pS->pSrc;
484 mask |= exprListTableUsage(pMaskSet, pS->pEList);
485 mask |= exprListTableUsage(pMaskSet, pS->pGroupBy);
486 mask |= exprListTableUsage(pMaskSet, pS->pOrderBy);
487 mask |= exprTableUsage(pMaskSet, pS->pWhere);
488 mask |= exprTableUsage(pMaskSet, pS->pHaving);
489 if( ALWAYS(pSrc!=0) ){
490 int i;
491 for(i=0; i<pSrc->nSrc; i++){
492 mask |= exprSelectTableUsage(pMaskSet, pSrc->a[i].pSelect);
493 mask |= exprTableUsage(pMaskSet, pSrc->a[i].pOn);
496 pS = pS->pPrior;
498 return mask;
502 ** Return TRUE if the given operator is one of the operators that is
503 ** allowed for an indexable WHERE clause term. The allowed operators are
504 ** "=", "<", ">", "<=", ">=", and "IN".
506 ** IMPLEMENTATION-OF: R-59926-26393 To be usable by an index a term must be
507 ** of one of the following forms: column = expression column > expression
508 ** column >= expression column < expression column <= expression
509 ** expression = column expression > column expression >= column
510 ** expression < column expression <= column column IN
511 ** (expression-list) column IN (subquery) column IS NULL
513 static int allowedOp(int op){
514 assert( TK_GT>TK_EQ && TK_GT<TK_GE );
515 assert( TK_LT>TK_EQ && TK_LT<TK_GE );
516 assert( TK_LE>TK_EQ && TK_LE<TK_GE );
517 assert( TK_GE==TK_EQ+4 );
518 return op==TK_IN || (op>=TK_EQ && op<=TK_GE) || op==TK_ISNULL;
522 ** Swap two objects of type TYPE.
524 #define SWAP(TYPE,A,B) {TYPE t=A; A=B; B=t;}
527 ** Commute a comparison operator. Expressions of the form "X op Y"
528 ** are converted into "Y op X".
530 ** If a collation sequence is associated with either the left or right
531 ** side of the comparison, it remains associated with the same side after
532 ** the commutation. So "Y collate NOCASE op X" becomes
533 ** "X collate NOCASE op Y". This is because any collation sequence on
534 ** the left hand side of a comparison overrides any collation sequence
535 ** attached to the right. For the same reason the EP_ExpCollate flag
536 ** is not commuted.
538 static void exprCommute(Parse *pParse, Expr *pExpr){
539 u16 expRight = (pExpr->pRight->flags & EP_ExpCollate);
540 u16 expLeft = (pExpr->pLeft->flags & EP_ExpCollate);
541 assert( allowedOp(pExpr->op) && pExpr->op!=TK_IN );
542 pExpr->pRight->pColl = sqlite3ExprCollSeq(pParse, pExpr->pRight);
543 pExpr->pLeft->pColl = sqlite3ExprCollSeq(pParse, pExpr->pLeft);
544 SWAP(CollSeq*,pExpr->pRight->pColl,pExpr->pLeft->pColl);
545 pExpr->pRight->flags = (pExpr->pRight->flags & ~EP_ExpCollate) | expLeft;
546 pExpr->pLeft->flags = (pExpr->pLeft->flags & ~EP_ExpCollate) | expRight;
547 SWAP(Expr*,pExpr->pRight,pExpr->pLeft);
548 if( pExpr->op>=TK_GT ){
549 assert( TK_LT==TK_GT+2 );
550 assert( TK_GE==TK_LE+2 );
551 assert( TK_GT>TK_EQ );
552 assert( TK_GT<TK_LE );
553 assert( pExpr->op>=TK_GT && pExpr->op<=TK_GE );
554 pExpr->op = ((pExpr->op-TK_GT)^2)+TK_GT;
559 ** Translate from TK_xx operator to WO_xx bitmask.
561 static u16 operatorMask(int op){
562 u16 c;
563 assert( allowedOp(op) );
564 if( op==TK_IN ){
565 c = WO_IN;
566 }else if( op==TK_ISNULL ){
567 c = WO_ISNULL;
568 }else{
569 assert( (WO_EQ<<(op-TK_EQ)) < 0x7fff );
570 c = (u16)(WO_EQ<<(op-TK_EQ));
572 assert( op!=TK_ISNULL || c==WO_ISNULL );
573 assert( op!=TK_IN || c==WO_IN );
574 assert( op!=TK_EQ || c==WO_EQ );
575 assert( op!=TK_LT || c==WO_LT );
576 assert( op!=TK_LE || c==WO_LE );
577 assert( op!=TK_GT || c==WO_GT );
578 assert( op!=TK_GE || c==WO_GE );
579 return c;
583 ** Search for a term in the WHERE clause that is of the form "X <op> <expr>"
584 ** where X is a reference to the iColumn of table iCur and <op> is one of
585 ** the WO_xx operator codes specified by the op parameter.
586 ** Return a pointer to the term. Return 0 if not found.
588 static WhereTerm *findTerm(
589 WhereClause *pWC, /* The WHERE clause to be searched */
590 int iCur, /* Cursor number of LHS */
591 int iColumn, /* Column number of LHS */
592 Bitmask notReady, /* RHS must not overlap with this mask */
593 u32 op, /* Mask of WO_xx values describing operator */
594 Index *pIdx /* Must be compatible with this index, if not NULL */
596 WhereTerm *pTerm;
597 int k;
598 assert( iCur>=0 );
599 op &= WO_ALL;
600 for(; pWC; pWC=pWC->pOuter){
601 for(pTerm=pWC->a, k=pWC->nTerm; k; k--, pTerm++){
602 if( pTerm->leftCursor==iCur
603 && (pTerm->prereqRight & notReady)==0
604 && pTerm->u.leftColumn==iColumn
605 && (pTerm->eOperator & op)!=0
607 if( pIdx && pTerm->eOperator!=WO_ISNULL ){
608 Expr *pX = pTerm->pExpr;
609 CollSeq *pColl;
610 char idxaff;
611 int j;
612 Parse *pParse = pWC->pParse;
614 idxaff = pIdx->pTable->aCol[iColumn].affinity;
615 if( !sqlite3IndexAffinityOk(pX, idxaff) ) continue;
617 /* Figure out the collation sequence required from an index for
618 ** it to be useful for optimising expression pX. Store this
619 ** value in variable pColl.
621 assert(pX->pLeft);
622 pColl = sqlite3BinaryCompareCollSeq(pParse, pX->pLeft, pX->pRight);
623 assert(pColl || pParse->nErr);
625 for(j=0; pIdx->aiColumn[j]!=iColumn; j++){
626 if( NEVER(j>=pIdx->nColumn) ) return 0;
628 if( pColl && sqlite3StrICmp(pColl->zName, pIdx->azColl[j]) ) continue;
630 return pTerm;
634 return 0;
637 /* Forward reference */
638 static void exprAnalyze(SrcList*, WhereClause*, int);
641 ** Call exprAnalyze on all terms in a WHERE clause.
645 static void exprAnalyzeAll(
646 SrcList *pTabList, /* the FROM clause */
647 WhereClause *pWC /* the WHERE clause to be analyzed */
649 int i;
650 for(i=pWC->nTerm-1; i>=0; i--){
651 exprAnalyze(pTabList, pWC, i);
655 #ifndef SQLITE_OMIT_LIKE_OPTIMIZATION
657 ** Check to see if the given expression is a LIKE or GLOB operator that
658 ** can be optimized using inequality constraints. Return TRUE if it is
659 ** so and false if not.
661 ** In order for the operator to be optimizible, the RHS must be a string
662 ** literal that does not begin with a wildcard.
664 static int isLikeOrGlob(
665 Parse *pParse, /* Parsing and code generating context */
666 Expr *pExpr, /* Test this expression */
667 Expr **ppPrefix, /* Pointer to TK_STRING expression with pattern prefix */
668 int *pisComplete, /* True if the only wildcard is % in the last character */
669 int *pnoCase /* True if uppercase is equivalent to lowercase */
671 const char *z = 0; /* String on RHS of LIKE operator */
672 Expr *pRight, *pLeft; /* Right and left size of LIKE operator */
673 ExprList *pList; /* List of operands to the LIKE operator */
674 int c; /* One character in z[] */
675 int cnt; /* Number of non-wildcard prefix characters */
676 char wc[3]; /* Wildcard characters */
677 sqlite3 *db = pParse->db; /* Database connection */
678 sqlite3_value *pVal = 0;
679 int op; /* Opcode of pRight */
681 if( !sqlite3IsLikeFunction(db, pExpr, pnoCase, wc) ){
682 return 0;
684 #ifdef SQLITE_EBCDIC
685 if( *pnoCase ) return 0;
686 #endif
687 pList = pExpr->x.pList;
688 pLeft = pList->a[1].pExpr;
689 if( pLeft->op!=TK_COLUMN || sqlite3ExprAffinity(pLeft)!=SQLITE_AFF_TEXT ){
690 /* IMP: R-02065-49465 The left-hand side of the LIKE or GLOB operator must
691 ** be the name of an indexed column with TEXT affinity. */
692 return 0;
694 assert( pLeft->iColumn!=(-1) ); /* Because IPK never has AFF_TEXT */
696 pRight = pList->a[0].pExpr;
697 op = pRight->op;
698 if( op==TK_REGISTER ){
699 op = pRight->op2;
701 if( op==TK_VARIABLE ){
702 Vdbe *pReprepare = pParse->pReprepare;
703 int iCol = pRight->iColumn;
704 pVal = sqlite3VdbeGetValue(pReprepare, iCol, SQLITE_AFF_NONE);
705 if( pVal && sqlite3_value_type(pVal)==SQLITE_TEXT ){
706 z = (char *)sqlite3_value_text(pVal);
708 sqlite3VdbeSetVarmask(pParse->pVdbe, iCol);
709 assert( pRight->op==TK_VARIABLE || pRight->op==TK_REGISTER );
710 }else if( op==TK_STRING ){
711 z = pRight->u.zToken;
713 if( z ){
714 cnt = 0;
715 while( (c=z[cnt])!=0 && c!=wc[0] && c!=wc[1] && c!=wc[2] ){
716 cnt++;
718 if( cnt!=0 && 255!=(u8)z[cnt-1] ){
719 Expr *pPrefix;
720 *pisComplete = c==wc[0] && z[cnt+1]==0;
721 pPrefix = sqlite3Expr(db, TK_STRING, z);
722 if( pPrefix ) pPrefix->u.zToken[cnt] = 0;
723 *ppPrefix = pPrefix;
724 if( op==TK_VARIABLE ){
725 Vdbe *v = pParse->pVdbe;
726 sqlite3VdbeSetVarmask(v, pRight->iColumn);
727 if( *pisComplete && pRight->u.zToken[1] ){
728 /* If the rhs of the LIKE expression is a variable, and the current
729 ** value of the variable means there is no need to invoke the LIKE
730 ** function, then no OP_Variable will be added to the program.
731 ** This causes problems for the sqlite3_bind_parameter_name()
732 ** API. To workaround them, add a dummy OP_Variable here.
734 int r1 = sqlite3GetTempReg(pParse);
735 sqlite3ExprCodeTarget(pParse, pRight, r1);
736 sqlite3VdbeChangeP3(v, sqlite3VdbeCurrentAddr(v)-1, 0);
737 sqlite3ReleaseTempReg(pParse, r1);
740 }else{
741 z = 0;
745 sqlite3ValueFree(pVal);
746 return (z!=0);
748 #endif /* SQLITE_OMIT_LIKE_OPTIMIZATION */
751 #ifndef SQLITE_OMIT_VIRTUALTABLE
753 ** Check to see if the given expression is of the form
755 ** column MATCH expr
757 ** If it is then return TRUE. If not, return FALSE.
759 static int isMatchOfColumn(
760 Expr *pExpr /* Test this expression */
762 ExprList *pList;
764 if( pExpr->op!=TK_FUNCTION ){
765 return 0;
767 if( sqlite3StrICmp(pExpr->u.zToken,"match")!=0 ){
768 return 0;
770 pList = pExpr->x.pList;
771 if( pList->nExpr!=2 ){
772 return 0;
774 if( pList->a[1].pExpr->op != TK_COLUMN ){
775 return 0;
777 return 1;
779 #endif /* SQLITE_OMIT_VIRTUALTABLE */
782 ** If the pBase expression originated in the ON or USING clause of
783 ** a join, then transfer the appropriate markings over to derived.
785 static void transferJoinMarkings(Expr *pDerived, Expr *pBase){
786 pDerived->flags |= pBase->flags & EP_FromJoin;
787 pDerived->iRightJoinTable = pBase->iRightJoinTable;
790 #if !defined(SQLITE_OMIT_OR_OPTIMIZATION) && !defined(SQLITE_OMIT_SUBQUERY)
792 ** Analyze a term that consists of two or more OR-connected
793 ** subterms. So in:
795 ** ... WHERE (a=5) AND (b=7 OR c=9 OR d=13) AND (d=13)
796 ** ^^^^^^^^^^^^^^^^^^^^
798 ** This routine analyzes terms such as the middle term in the above example.
799 ** A WhereOrTerm object is computed and attached to the term under
800 ** analysis, regardless of the outcome of the analysis. Hence:
802 ** WhereTerm.wtFlags |= TERM_ORINFO
803 ** WhereTerm.u.pOrInfo = a dynamically allocated WhereOrTerm object
805 ** The term being analyzed must have two or more of OR-connected subterms.
806 ** A single subterm might be a set of AND-connected sub-subterms.
807 ** Examples of terms under analysis:
809 ** (A) t1.x=t2.y OR t1.x=t2.z OR t1.y=15 OR t1.z=t3.a+5
810 ** (B) x=expr1 OR expr2=x OR x=expr3
811 ** (C) t1.x=t2.y OR (t1.x=t2.z AND t1.y=15)
812 ** (D) x=expr1 OR (y>11 AND y<22 AND z LIKE '*hello*')
813 ** (E) (p.a=1 AND q.b=2 AND r.c=3) OR (p.x=4 AND q.y=5 AND r.z=6)
815 ** CASE 1:
817 ** If all subterms are of the form T.C=expr for some single column of C
818 ** a single table T (as shown in example B above) then create a new virtual
819 ** term that is an equivalent IN expression. In other words, if the term
820 ** being analyzed is:
822 ** x = expr1 OR expr2 = x OR x = expr3
824 ** then create a new virtual term like this:
826 ** x IN (expr1,expr2,expr3)
828 ** CASE 2:
830 ** If all subterms are indexable by a single table T, then set
832 ** WhereTerm.eOperator = WO_OR
833 ** WhereTerm.u.pOrInfo->indexable |= the cursor number for table T
835 ** A subterm is "indexable" if it is of the form
836 ** "T.C <op> <expr>" where C is any column of table T and
837 ** <op> is one of "=", "<", "<=", ">", ">=", "IS NULL", or "IN".
838 ** A subterm is also indexable if it is an AND of two or more
839 ** subsubterms at least one of which is indexable. Indexable AND
840 ** subterms have their eOperator set to WO_AND and they have
841 ** u.pAndInfo set to a dynamically allocated WhereAndTerm object.
843 ** From another point of view, "indexable" means that the subterm could
844 ** potentially be used with an index if an appropriate index exists.
845 ** This analysis does not consider whether or not the index exists; that
846 ** is something the bestIndex() routine will determine. This analysis
847 ** only looks at whether subterms appropriate for indexing exist.
849 ** All examples A through E above all satisfy case 2. But if a term
850 ** also statisfies case 1 (such as B) we know that the optimizer will
851 ** always prefer case 1, so in that case we pretend that case 2 is not
852 ** satisfied.
854 ** It might be the case that multiple tables are indexable. For example,
855 ** (E) above is indexable on tables P, Q, and R.
857 ** Terms that satisfy case 2 are candidates for lookup by using
858 ** separate indices to find rowids for each subterm and composing
859 ** the union of all rowids using a RowSet object. This is similar
860 ** to "bitmap indices" in other database engines.
862 ** OTHERWISE:
864 ** If neither case 1 nor case 2 apply, then leave the eOperator set to
865 ** zero. This term is not useful for search.
867 static void exprAnalyzeOrTerm(
868 SrcList *pSrc, /* the FROM clause */
869 WhereClause *pWC, /* the complete WHERE clause */
870 int idxTerm /* Index of the OR-term to be analyzed */
872 Parse *pParse = pWC->pParse; /* Parser context */
873 sqlite3 *db = pParse->db; /* Database connection */
874 WhereTerm *pTerm = &pWC->a[idxTerm]; /* The term to be analyzed */
875 Expr *pExpr = pTerm->pExpr; /* The expression of the term */
876 WhereMaskSet *pMaskSet = pWC->pMaskSet; /* Table use masks */
877 int i; /* Loop counters */
878 WhereClause *pOrWc; /* Breakup of pTerm into subterms */
879 WhereTerm *pOrTerm; /* A Sub-term within the pOrWc */
880 WhereOrInfo *pOrInfo; /* Additional information associated with pTerm */
881 Bitmask chngToIN; /* Tables that might satisfy case 1 */
882 Bitmask indexable; /* Tables that are indexable, satisfying case 2 */
885 ** Break the OR clause into its separate subterms. The subterms are
886 ** stored in a WhereClause structure containing within the WhereOrInfo
887 ** object that is attached to the original OR clause term.
889 assert( (pTerm->wtFlags & (TERM_DYNAMIC|TERM_ORINFO|TERM_ANDINFO))==0 );
890 assert( pExpr->op==TK_OR );
891 pTerm->u.pOrInfo = pOrInfo = sqlite3DbMallocZero(db, sizeof(*pOrInfo));
892 if( pOrInfo==0 ) return;
893 pTerm->wtFlags |= TERM_ORINFO;
894 pOrWc = &pOrInfo->wc;
895 whereClauseInit(pOrWc, pWC->pParse, pMaskSet, pWC->wctrlFlags);
896 whereSplit(pOrWc, pExpr, TK_OR);
897 exprAnalyzeAll(pSrc, pOrWc);
898 if( db->mallocFailed ) return;
899 assert( pOrWc->nTerm>=2 );
902 ** Compute the set of tables that might satisfy cases 1 or 2.
904 indexable = ~(Bitmask)0;
905 chngToIN = ~(pWC->vmask);
906 for(i=pOrWc->nTerm-1, pOrTerm=pOrWc->a; i>=0 && indexable; i--, pOrTerm++){
907 if( (pOrTerm->eOperator & WO_SINGLE)==0 ){
908 WhereAndInfo *pAndInfo;
909 assert( pOrTerm->eOperator==0 );
910 assert( (pOrTerm->wtFlags & (TERM_ANDINFO|TERM_ORINFO))==0 );
911 chngToIN = 0;
912 pAndInfo = sqlite3DbMallocRaw(db, sizeof(*pAndInfo));
913 if( pAndInfo ){
914 WhereClause *pAndWC;
915 WhereTerm *pAndTerm;
916 int j;
917 Bitmask b = 0;
918 pOrTerm->u.pAndInfo = pAndInfo;
919 pOrTerm->wtFlags |= TERM_ANDINFO;
920 pOrTerm->eOperator = WO_AND;
921 pAndWC = &pAndInfo->wc;
922 whereClauseInit(pAndWC, pWC->pParse, pMaskSet, pWC->wctrlFlags);
923 whereSplit(pAndWC, pOrTerm->pExpr, TK_AND);
924 exprAnalyzeAll(pSrc, pAndWC);
925 pAndWC->pOuter = pWC;
926 testcase( db->mallocFailed );
927 if( !db->mallocFailed ){
928 for(j=0, pAndTerm=pAndWC->a; j<pAndWC->nTerm; j++, pAndTerm++){
929 assert( pAndTerm->pExpr );
930 if( allowedOp(pAndTerm->pExpr->op) ){
931 b |= getMask(pMaskSet, pAndTerm->leftCursor);
935 indexable &= b;
937 }else if( pOrTerm->wtFlags & TERM_COPIED ){
938 /* Skip this term for now. We revisit it when we process the
939 ** corresponding TERM_VIRTUAL term */
940 }else{
941 Bitmask b;
942 b = getMask(pMaskSet, pOrTerm->leftCursor);
943 if( pOrTerm->wtFlags & TERM_VIRTUAL ){
944 WhereTerm *pOther = &pOrWc->a[pOrTerm->iParent];
945 b |= getMask(pMaskSet, pOther->leftCursor);
947 indexable &= b;
948 if( pOrTerm->eOperator!=WO_EQ ){
949 chngToIN = 0;
950 }else{
951 chngToIN &= b;
957 ** Record the set of tables that satisfy case 2. The set might be
958 ** empty.
960 pOrInfo->indexable = indexable;
961 pTerm->eOperator = indexable==0 ? 0 : WO_OR;
964 ** chngToIN holds a set of tables that *might* satisfy case 1. But
965 ** we have to do some additional checking to see if case 1 really
966 ** is satisfied.
968 ** chngToIN will hold either 0, 1, or 2 bits. The 0-bit case means
969 ** that there is no possibility of transforming the OR clause into an
970 ** IN operator because one or more terms in the OR clause contain
971 ** something other than == on a column in the single table. The 1-bit
972 ** case means that every term of the OR clause is of the form
973 ** "table.column=expr" for some single table. The one bit that is set
974 ** will correspond to the common table. We still need to check to make
975 ** sure the same column is used on all terms. The 2-bit case is when
976 ** the all terms are of the form "table1.column=table2.column". It
977 ** might be possible to form an IN operator with either table1.column
978 ** or table2.column as the LHS if either is common to every term of
979 ** the OR clause.
981 ** Note that terms of the form "table.column1=table.column2" (the
982 ** same table on both sizes of the ==) cannot be optimized.
984 if( chngToIN ){
985 int okToChngToIN = 0; /* True if the conversion to IN is valid */
986 int iColumn = -1; /* Column index on lhs of IN operator */
987 int iCursor = -1; /* Table cursor common to all terms */
988 int j = 0; /* Loop counter */
990 /* Search for a table and column that appears on one side or the
991 ** other of the == operator in every subterm. That table and column
992 ** will be recorded in iCursor and iColumn. There might not be any
993 ** such table and column. Set okToChngToIN if an appropriate table
994 ** and column is found but leave okToChngToIN false if not found.
996 for(j=0; j<2 && !okToChngToIN; j++){
997 pOrTerm = pOrWc->a;
998 for(i=pOrWc->nTerm-1; i>=0; i--, pOrTerm++){
999 assert( pOrTerm->eOperator==WO_EQ );
1000 pOrTerm->wtFlags &= ~TERM_OR_OK;
1001 if( pOrTerm->leftCursor==iCursor ){
1002 /* This is the 2-bit case and we are on the second iteration and
1003 ** current term is from the first iteration. So skip this term. */
1004 assert( j==1 );
1005 continue;
1007 if( (chngToIN & getMask(pMaskSet, pOrTerm->leftCursor))==0 ){
1008 /* This term must be of the form t1.a==t2.b where t2 is in the
1009 ** chngToIN set but t1 is not. This term will be either preceeded
1010 ** or follwed by an inverted copy (t2.b==t1.a). Skip this term
1011 ** and use its inversion. */
1012 testcase( pOrTerm->wtFlags & TERM_COPIED );
1013 testcase( pOrTerm->wtFlags & TERM_VIRTUAL );
1014 assert( pOrTerm->wtFlags & (TERM_COPIED|TERM_VIRTUAL) );
1015 continue;
1017 iColumn = pOrTerm->u.leftColumn;
1018 iCursor = pOrTerm->leftCursor;
1019 break;
1021 if( i<0 ){
1022 /* No candidate table+column was found. This can only occur
1023 ** on the second iteration */
1024 assert( j==1 );
1025 assert( (chngToIN&(chngToIN-1))==0 );
1026 assert( chngToIN==getMask(pMaskSet, iCursor) );
1027 break;
1029 testcase( j==1 );
1031 /* We have found a candidate table and column. Check to see if that
1032 ** table and column is common to every term in the OR clause */
1033 okToChngToIN = 1;
1034 for(; i>=0 && okToChngToIN; i--, pOrTerm++){
1035 assert( pOrTerm->eOperator==WO_EQ );
1036 if( pOrTerm->leftCursor!=iCursor ){
1037 pOrTerm->wtFlags &= ~TERM_OR_OK;
1038 }else if( pOrTerm->u.leftColumn!=iColumn ){
1039 okToChngToIN = 0;
1040 }else{
1041 int affLeft, affRight;
1042 /* If the right-hand side is also a column, then the affinities
1043 ** of both right and left sides must be such that no type
1044 ** conversions are required on the right. (Ticket #2249)
1046 affRight = sqlite3ExprAffinity(pOrTerm->pExpr->pRight);
1047 affLeft = sqlite3ExprAffinity(pOrTerm->pExpr->pLeft);
1048 if( affRight!=0 && affRight!=affLeft ){
1049 okToChngToIN = 0;
1050 }else{
1051 pOrTerm->wtFlags |= TERM_OR_OK;
1057 /* At this point, okToChngToIN is true if original pTerm satisfies
1058 ** case 1. In that case, construct a new virtual term that is
1059 ** pTerm converted into an IN operator.
1061 ** EV: R-00211-15100
1063 if( okToChngToIN ){
1064 Expr *pDup; /* A transient duplicate expression */
1065 ExprList *pList = 0; /* The RHS of the IN operator */
1066 Expr *pLeft = 0; /* The LHS of the IN operator */
1067 Expr *pNew; /* The complete IN operator */
1069 for(i=pOrWc->nTerm-1, pOrTerm=pOrWc->a; i>=0; i--, pOrTerm++){
1070 if( (pOrTerm->wtFlags & TERM_OR_OK)==0 ) continue;
1071 assert( pOrTerm->eOperator==WO_EQ );
1072 assert( pOrTerm->leftCursor==iCursor );
1073 assert( pOrTerm->u.leftColumn==iColumn );
1074 pDup = sqlite3ExprDup(db, pOrTerm->pExpr->pRight, 0);
1075 pList = sqlite3ExprListAppend(pWC->pParse, pList, pDup);
1076 pLeft = pOrTerm->pExpr->pLeft;
1078 assert( pLeft!=0 );
1079 pDup = sqlite3ExprDup(db, pLeft, 0);
1080 pNew = sqlite3PExpr(pParse, TK_IN, pDup, 0, 0);
1081 if( pNew ){
1082 int idxNew;
1083 transferJoinMarkings(pNew, pExpr);
1084 assert( !ExprHasProperty(pNew, EP_xIsSelect) );
1085 pNew->x.pList = pList;
1086 idxNew = whereClauseInsert(pWC, pNew, TERM_VIRTUAL|TERM_DYNAMIC);
1087 testcase( idxNew==0 );
1088 exprAnalyze(pSrc, pWC, idxNew);
1089 pTerm = &pWC->a[idxTerm];
1090 pWC->a[idxNew].iParent = idxTerm;
1091 pTerm->nChild = 1;
1092 }else{
1093 sqlite3ExprListDelete(db, pList);
1095 pTerm->eOperator = WO_NOOP; /* case 1 trumps case 2 */
1099 #endif /* !SQLITE_OMIT_OR_OPTIMIZATION && !SQLITE_OMIT_SUBQUERY */
1103 ** The input to this routine is an WhereTerm structure with only the
1104 ** "pExpr" field filled in. The job of this routine is to analyze the
1105 ** subexpression and populate all the other fields of the WhereTerm
1106 ** structure.
1108 ** If the expression is of the form "<expr> <op> X" it gets commuted
1109 ** to the standard form of "X <op> <expr>".
1111 ** If the expression is of the form "X <op> Y" where both X and Y are
1112 ** columns, then the original expression is unchanged and a new virtual
1113 ** term of the form "Y <op> X" is added to the WHERE clause and
1114 ** analyzed separately. The original term is marked with TERM_COPIED
1115 ** and the new term is marked with TERM_DYNAMIC (because it's pExpr
1116 ** needs to be freed with the WhereClause) and TERM_VIRTUAL (because it
1117 ** is a commuted copy of a prior term.) The original term has nChild=1
1118 ** and the copy has idxParent set to the index of the original term.
1120 static void exprAnalyze(
1121 SrcList *pSrc, /* the FROM clause */
1122 WhereClause *pWC, /* the WHERE clause */
1123 int idxTerm /* Index of the term to be analyzed */
1125 WhereTerm *pTerm; /* The term to be analyzed */
1126 WhereMaskSet *pMaskSet; /* Set of table index masks */
1127 Expr *pExpr; /* The expression to be analyzed */
1128 Bitmask prereqLeft; /* Prerequesites of the pExpr->pLeft */
1129 Bitmask prereqAll; /* Prerequesites of pExpr */
1130 Bitmask extraRight = 0; /* Extra dependencies on LEFT JOIN */
1131 Expr *pStr1 = 0; /* RHS of LIKE/GLOB operator */
1132 int isComplete = 0; /* RHS of LIKE/GLOB ends with wildcard */
1133 int noCase = 0; /* LIKE/GLOB distinguishes case */
1134 int op; /* Top-level operator. pExpr->op */
1135 Parse *pParse = pWC->pParse; /* Parsing context */
1136 sqlite3 *db = pParse->db; /* Database connection */
1138 if( db->mallocFailed ){
1139 return;
1141 pTerm = &pWC->a[idxTerm];
1142 pMaskSet = pWC->pMaskSet;
1143 pExpr = pTerm->pExpr;
1144 prereqLeft = exprTableUsage(pMaskSet, pExpr->pLeft);
1145 op = pExpr->op;
1146 if( op==TK_IN ){
1147 assert( pExpr->pRight==0 );
1148 if( ExprHasProperty(pExpr, EP_xIsSelect) ){
1149 pTerm->prereqRight = exprSelectTableUsage(pMaskSet, pExpr->x.pSelect);
1150 }else{
1151 pTerm->prereqRight = exprListTableUsage(pMaskSet, pExpr->x.pList);
1153 }else if( op==TK_ISNULL ){
1154 pTerm->prereqRight = 0;
1155 }else{
1156 pTerm->prereqRight = exprTableUsage(pMaskSet, pExpr->pRight);
1158 prereqAll = exprTableUsage(pMaskSet, pExpr);
1159 if( ExprHasProperty(pExpr, EP_FromJoin) ){
1160 Bitmask x = getMask(pMaskSet, pExpr->iRightJoinTable);
1161 prereqAll |= x;
1162 extraRight = x-1; /* ON clause terms may not be used with an index
1163 ** on left table of a LEFT JOIN. Ticket #3015 */
1165 pTerm->prereqAll = prereqAll;
1166 pTerm->leftCursor = -1;
1167 pTerm->iParent = -1;
1168 pTerm->eOperator = 0;
1169 if( allowedOp(op) && (pTerm->prereqRight & prereqLeft)==0 ){
1170 Expr *pLeft = pExpr->pLeft;
1171 Expr *pRight = pExpr->pRight;
1172 if( pLeft->op==TK_COLUMN ){
1173 pTerm->leftCursor = pLeft->iTable;
1174 pTerm->u.leftColumn = pLeft->iColumn;
1175 pTerm->eOperator = operatorMask(op);
1177 if( pRight && pRight->op==TK_COLUMN ){
1178 WhereTerm *pNew;
1179 Expr *pDup;
1180 if( pTerm->leftCursor>=0 ){
1181 int idxNew;
1182 pDup = sqlite3ExprDup(db, pExpr, 0);
1183 if( db->mallocFailed ){
1184 sqlite3ExprDelete(db, pDup);
1185 return;
1187 idxNew = whereClauseInsert(pWC, pDup, TERM_VIRTUAL|TERM_DYNAMIC);
1188 if( idxNew==0 ) return;
1189 pNew = &pWC->a[idxNew];
1190 pNew->iParent = idxTerm;
1191 pTerm = &pWC->a[idxTerm];
1192 pTerm->nChild = 1;
1193 pTerm->wtFlags |= TERM_COPIED;
1194 }else{
1195 pDup = pExpr;
1196 pNew = pTerm;
1198 exprCommute(pParse, pDup);
1199 pLeft = pDup->pLeft;
1200 pNew->leftCursor = pLeft->iTable;
1201 pNew->u.leftColumn = pLeft->iColumn;
1202 testcase( (prereqLeft | extraRight) != prereqLeft );
1203 pNew->prereqRight = prereqLeft | extraRight;
1204 pNew->prereqAll = prereqAll;
1205 pNew->eOperator = operatorMask(pDup->op);
1209 #ifndef SQLITE_OMIT_BETWEEN_OPTIMIZATION
1210 /* If a term is the BETWEEN operator, create two new virtual terms
1211 ** that define the range that the BETWEEN implements. For example:
1213 ** a BETWEEN b AND c
1215 ** is converted into:
1217 ** (a BETWEEN b AND c) AND (a>=b) AND (a<=c)
1219 ** The two new terms are added onto the end of the WhereClause object.
1220 ** The new terms are "dynamic" and are children of the original BETWEEN
1221 ** term. That means that if the BETWEEN term is coded, the children are
1222 ** skipped. Or, if the children are satisfied by an index, the original
1223 ** BETWEEN term is skipped.
1225 else if( pExpr->op==TK_BETWEEN && pWC->op==TK_AND ){
1226 ExprList *pList = pExpr->x.pList;
1227 int i;
1228 static const u8 ops[] = {TK_GE, TK_LE};
1229 assert( pList!=0 );
1230 assert( pList->nExpr==2 );
1231 for(i=0; i<2; i++){
1232 Expr *pNewExpr;
1233 int idxNew;
1234 pNewExpr = sqlite3PExpr(pParse, ops[i],
1235 sqlite3ExprDup(db, pExpr->pLeft, 0),
1236 sqlite3ExprDup(db, pList->a[i].pExpr, 0), 0);
1237 idxNew = whereClauseInsert(pWC, pNewExpr, TERM_VIRTUAL|TERM_DYNAMIC);
1238 testcase( idxNew==0 );
1239 exprAnalyze(pSrc, pWC, idxNew);
1240 pTerm = &pWC->a[idxTerm];
1241 pWC->a[idxNew].iParent = idxTerm;
1243 pTerm->nChild = 2;
1245 #endif /* SQLITE_OMIT_BETWEEN_OPTIMIZATION */
1247 #if !defined(SQLITE_OMIT_OR_OPTIMIZATION) && !defined(SQLITE_OMIT_SUBQUERY)
1248 /* Analyze a term that is composed of two or more subterms connected by
1249 ** an OR operator.
1251 else if( pExpr->op==TK_OR ){
1252 assert( pWC->op==TK_AND );
1253 exprAnalyzeOrTerm(pSrc, pWC, idxTerm);
1254 pTerm = &pWC->a[idxTerm];
1256 #endif /* SQLITE_OMIT_OR_OPTIMIZATION */
1258 #ifndef SQLITE_OMIT_LIKE_OPTIMIZATION
1259 /* Add constraints to reduce the search space on a LIKE or GLOB
1260 ** operator.
1262 ** A like pattern of the form "x LIKE 'abc%'" is changed into constraints
1264 ** x>='abc' AND x<'abd' AND x LIKE 'abc%'
1266 ** The last character of the prefix "abc" is incremented to form the
1267 ** termination condition "abd".
1269 if( pWC->op==TK_AND
1270 && isLikeOrGlob(pParse, pExpr, &pStr1, &isComplete, &noCase)
1272 Expr *pLeft; /* LHS of LIKE/GLOB operator */
1273 Expr *pStr2; /* Copy of pStr1 - RHS of LIKE/GLOB operator */
1274 Expr *pNewExpr1;
1275 Expr *pNewExpr2;
1276 int idxNew1;
1277 int idxNew2;
1278 CollSeq *pColl; /* Collating sequence to use */
1280 pLeft = pExpr->x.pList->a[1].pExpr;
1281 pStr2 = sqlite3ExprDup(db, pStr1, 0);
1282 if( !db->mallocFailed ){
1283 u8 c, *pC; /* Last character before the first wildcard */
1284 pC = (u8*)&pStr2->u.zToken[sqlite3Strlen30(pStr2->u.zToken)-1];
1285 c = *pC;
1286 if( noCase ){
1287 /* The point is to increment the last character before the first
1288 ** wildcard. But if we increment '@', that will push it into the
1289 ** alphabetic range where case conversions will mess up the
1290 ** inequality. To avoid this, make sure to also run the full
1291 ** LIKE on all candidate expressions by clearing the isComplete flag
1293 if( c=='A'-1 ) isComplete = 0; /* EV: R-64339-08207 */
1296 c = sqlite3UpperToLower[c];
1298 *pC = c + 1;
1300 pColl = sqlite3FindCollSeq(db, SQLITE_UTF8, noCase ? "NOCASE" : "BINARY",0);
1301 pNewExpr1 = sqlite3PExpr(pParse, TK_GE,
1302 sqlite3ExprSetColl(sqlite3ExprDup(db,pLeft,0), pColl),
1303 pStr1, 0);
1304 idxNew1 = whereClauseInsert(pWC, pNewExpr1, TERM_VIRTUAL|TERM_DYNAMIC);
1305 testcase( idxNew1==0 );
1306 exprAnalyze(pSrc, pWC, idxNew1);
1307 pNewExpr2 = sqlite3PExpr(pParse, TK_LT,
1308 sqlite3ExprSetColl(sqlite3ExprDup(db,pLeft,0), pColl),
1309 pStr2, 0);
1310 idxNew2 = whereClauseInsert(pWC, pNewExpr2, TERM_VIRTUAL|TERM_DYNAMIC);
1311 testcase( idxNew2==0 );
1312 exprAnalyze(pSrc, pWC, idxNew2);
1313 pTerm = &pWC->a[idxTerm];
1314 if( isComplete ){
1315 pWC->a[idxNew1].iParent = idxTerm;
1316 pWC->a[idxNew2].iParent = idxTerm;
1317 pTerm->nChild = 2;
1320 #endif /* SQLITE_OMIT_LIKE_OPTIMIZATION */
1322 #ifndef SQLITE_OMIT_VIRTUALTABLE
1323 /* Add a WO_MATCH auxiliary term to the constraint set if the
1324 ** current expression is of the form: column MATCH expr.
1325 ** This information is used by the xBestIndex methods of
1326 ** virtual tables. The native query optimizer does not attempt
1327 ** to do anything with MATCH functions.
1329 if( isMatchOfColumn(pExpr) ){
1330 int idxNew;
1331 Expr *pRight, *pLeft;
1332 WhereTerm *pNewTerm;
1333 Bitmask prereqColumn, prereqExpr;
1335 pRight = pExpr->x.pList->a[0].pExpr;
1336 pLeft = pExpr->x.pList->a[1].pExpr;
1337 prereqExpr = exprTableUsage(pMaskSet, pRight);
1338 prereqColumn = exprTableUsage(pMaskSet, pLeft);
1339 if( (prereqExpr & prereqColumn)==0 ){
1340 Expr *pNewExpr;
1341 pNewExpr = sqlite3PExpr(pParse, TK_MATCH,
1342 0, sqlite3ExprDup(db, pRight, 0), 0);
1343 idxNew = whereClauseInsert(pWC, pNewExpr, TERM_VIRTUAL|TERM_DYNAMIC);
1344 testcase( idxNew==0 );
1345 pNewTerm = &pWC->a[idxNew];
1346 pNewTerm->prereqRight = prereqExpr;
1347 pNewTerm->leftCursor = pLeft->iTable;
1348 pNewTerm->u.leftColumn = pLeft->iColumn;
1349 pNewTerm->eOperator = WO_MATCH;
1350 pNewTerm->iParent = idxTerm;
1351 pTerm = &pWC->a[idxTerm];
1352 pTerm->nChild = 1;
1353 pTerm->wtFlags |= TERM_COPIED;
1354 pNewTerm->prereqAll = pTerm->prereqAll;
1357 #endif /* SQLITE_OMIT_VIRTUALTABLE */
1359 #ifdef SQLITE_ENABLE_STAT3
1360 /* When sqlite_stat3 histogram data is available an operator of the
1361 ** form "x IS NOT NULL" can sometimes be evaluated more efficiently
1362 ** as "x>NULL" if x is not an INTEGER PRIMARY KEY. So construct a
1363 ** virtual term of that form.
1365 ** Note that the virtual term must be tagged with TERM_VNULL. This
1366 ** TERM_VNULL tag will suppress the not-null check at the beginning
1367 ** of the loop. Without the TERM_VNULL flag, the not-null check at
1368 ** the start of the loop will prevent any results from being returned.
1370 if( pExpr->op==TK_NOTNULL
1371 && pExpr->pLeft->op==TK_COLUMN
1372 && pExpr->pLeft->iColumn>=0
1374 Expr *pNewExpr;
1375 Expr *pLeft = pExpr->pLeft;
1376 int idxNew;
1377 WhereTerm *pNewTerm;
1379 pNewExpr = sqlite3PExpr(pParse, TK_GT,
1380 sqlite3ExprDup(db, pLeft, 0),
1381 sqlite3PExpr(pParse, TK_NULL, 0, 0, 0), 0);
1383 idxNew = whereClauseInsert(pWC, pNewExpr,
1384 TERM_VIRTUAL|TERM_DYNAMIC|TERM_VNULL);
1385 if( idxNew ){
1386 pNewTerm = &pWC->a[idxNew];
1387 pNewTerm->prereqRight = 0;
1388 pNewTerm->leftCursor = pLeft->iTable;
1389 pNewTerm->u.leftColumn = pLeft->iColumn;
1390 pNewTerm->eOperator = WO_GT;
1391 pNewTerm->iParent = idxTerm;
1392 pTerm = &pWC->a[idxTerm];
1393 pTerm->nChild = 1;
1394 pTerm->wtFlags |= TERM_COPIED;
1395 pNewTerm->prereqAll = pTerm->prereqAll;
1398 #endif /* SQLITE_ENABLE_STAT */
1400 /* Prevent ON clause terms of a LEFT JOIN from being used to drive
1401 ** an index for tables to the left of the join.
1403 pTerm->prereqRight |= extraRight;
1407 ** Return TRUE if any of the expressions in pList->a[iFirst...] contain
1408 ** a reference to any table other than the iBase table.
1410 static int referencesOtherTables(
1411 ExprList *pList, /* Search expressions in ths list */
1412 WhereMaskSet *pMaskSet, /* Mapping from tables to bitmaps */
1413 int iFirst, /* Be searching with the iFirst-th expression */
1414 int iBase /* Ignore references to this table */
1416 Bitmask allowed = ~getMask(pMaskSet, iBase);
1417 while( iFirst<pList->nExpr ){
1418 if( (exprTableUsage(pMaskSet, pList->a[iFirst++].pExpr)&allowed)!=0 ){
1419 return 1;
1422 return 0;
1426 ** This function searches the expression list passed as the second argument
1427 ** for an expression of type TK_COLUMN that refers to the same column and
1428 ** uses the same collation sequence as the iCol'th column of index pIdx.
1429 ** Argument iBase is the cursor number used for the table that pIdx refers
1430 ** to.
1432 ** If such an expression is found, its index in pList->a[] is returned. If
1433 ** no expression is found, -1 is returned.
1435 static int findIndexCol(
1436 Parse *pParse, /* Parse context */
1437 ExprList *pList, /* Expression list to search */
1438 int iBase, /* Cursor for table associated with pIdx */
1439 Index *pIdx, /* Index to match column of */
1440 int iCol /* Column of index to match */
1442 int i;
1443 const char *zColl = pIdx->azColl[iCol];
1445 for(i=0; i<pList->nExpr; i++){
1446 Expr *p = pList->a[i].pExpr;
1447 if( p->op==TK_COLUMN
1448 && p->iColumn==pIdx->aiColumn[iCol]
1449 && p->iTable==iBase
1451 CollSeq *pColl = sqlite3ExprCollSeq(pParse, p);
1452 if( ALWAYS(pColl) && 0==sqlite3StrICmp(pColl->zName, zColl) ){
1453 return i;
1458 return -1;
1462 ** This routine determines if pIdx can be used to assist in processing a
1463 ** DISTINCT qualifier. In other words, it tests whether or not using this
1464 ** index for the outer loop guarantees that rows with equal values for
1465 ** all expressions in the pDistinct list are delivered grouped together.
1467 ** For example, the query
1469 ** SELECT DISTINCT a, b, c FROM tbl WHERE a = ?
1471 ** can benefit from any index on columns "b" and "c".
1473 static int isDistinctIndex(
1474 Parse *pParse, /* Parsing context */
1475 WhereClause *pWC, /* The WHERE clause */
1476 Index *pIdx, /* The index being considered */
1477 int base, /* Cursor number for the table pIdx is on */
1478 ExprList *pDistinct, /* The DISTINCT expressions */
1479 int nEqCol /* Number of index columns with == */
1481 Bitmask mask = 0; /* Mask of unaccounted for pDistinct exprs */
1482 int i; /* Iterator variable */
1484 if( pIdx->zName==0 || pDistinct==0 || pDistinct->nExpr>=BMS ) return 0;
1485 testcase( pDistinct->nExpr==BMS-1 );
1487 /* Loop through all the expressions in the distinct list. If any of them
1488 ** are not simple column references, return early. Otherwise, test if the
1489 ** WHERE clause contains a "col=X" clause. If it does, the expression
1490 ** can be ignored. If it does not, and the column does not belong to the
1491 ** same table as index pIdx, return early. Finally, if there is no
1492 ** matching "col=X" expression and the column is on the same table as pIdx,
1493 ** set the corresponding bit in variable mask.
1495 for(i=0; i<pDistinct->nExpr; i++){
1496 WhereTerm *pTerm;
1497 Expr *p = pDistinct->a[i].pExpr;
1498 if( p->op!=TK_COLUMN ) return 0;
1499 pTerm = findTerm(pWC, p->iTable, p->iColumn, ~(Bitmask)0, WO_EQ, 0);
1500 if( pTerm ){
1501 Expr *pX = pTerm->pExpr;
1502 CollSeq *p1 = sqlite3BinaryCompareCollSeq(pParse, pX->pLeft, pX->pRight);
1503 CollSeq *p2 = sqlite3ExprCollSeq(pParse, p);
1504 if( p1==p2 ) continue;
1506 if( p->iTable!=base ) return 0;
1507 mask |= (((Bitmask)1) << i);
1510 for(i=nEqCol; mask && i<pIdx->nColumn; i++){
1511 int iExpr = findIndexCol(pParse, pDistinct, base, pIdx, i);
1512 if( iExpr<0 ) break;
1513 mask &= ~(((Bitmask)1) << iExpr);
1516 return (mask==0);
1521 ** Return true if the DISTINCT expression-list passed as the third argument
1522 ** is redundant. A DISTINCT list is redundant if the database contains a
1523 ** UNIQUE index that guarantees that the result of the query will be distinct
1524 ** anyway.
1526 static int isDistinctRedundant(
1527 Parse *pParse,
1528 SrcList *pTabList,
1529 WhereClause *pWC,
1530 ExprList *pDistinct
1532 Table *pTab;
1533 Index *pIdx;
1534 int i;
1535 int iBase;
1537 /* If there is more than one table or sub-select in the FROM clause of
1538 ** this query, then it will not be possible to show that the DISTINCT
1539 ** clause is redundant. */
1540 if( pTabList->nSrc!=1 ) return 0;
1541 iBase = pTabList->a[0].iCursor;
1542 pTab = pTabList->a[0].pTab;
1544 /* If any of the expressions is an IPK column on table iBase, then return
1545 ** true. Note: The (p->iTable==iBase) part of this test may be false if the
1546 ** current SELECT is a correlated sub-query.
1548 for(i=0; i<pDistinct->nExpr; i++){
1549 Expr *p = pDistinct->a[i].pExpr;
1550 if( p->op==TK_COLUMN && p->iTable==iBase && p->iColumn<0 ) return 1;
1553 /* Loop through all indices on the table, checking each to see if it makes
1554 ** the DISTINCT qualifier redundant. It does so if:
1556 ** 1. The index is itself UNIQUE, and
1558 ** 2. All of the columns in the index are either part of the pDistinct
1559 ** list, or else the WHERE clause contains a term of the form "col=X",
1560 ** where X is a constant value. The collation sequences of the
1561 ** comparison and select-list expressions must match those of the index.
1563 for(pIdx=pTab->pIndex; pIdx; pIdx=pIdx->pNext){
1564 if( pIdx->onError==OE_None ) continue;
1565 for(i=0; i<pIdx->nColumn; i++){
1566 int iCol = pIdx->aiColumn[i];
1567 if( 0==findTerm(pWC, iBase, iCol, ~(Bitmask)0, WO_EQ, pIdx)
1568 && 0>findIndexCol(pParse, pDistinct, iBase, pIdx, i)
1570 break;
1573 if( i==pIdx->nColumn ){
1574 /* This index implies that the DISTINCT qualifier is redundant. */
1575 return 1;
1579 return 0;
1583 ** This routine decides if pIdx can be used to satisfy the ORDER BY
1584 ** clause. If it can, it returns 1. If pIdx cannot satisfy the
1585 ** ORDER BY clause, this routine returns 0.
1587 ** pOrderBy is an ORDER BY clause from a SELECT statement. pTab is the
1588 ** left-most table in the FROM clause of that same SELECT statement and
1589 ** the table has a cursor number of "base". pIdx is an index on pTab.
1591 ** nEqCol is the number of columns of pIdx that are used as equality
1592 ** constraints. Any of these columns may be missing from the ORDER BY
1593 ** clause and the match can still be a success.
1595 ** All terms of the ORDER BY that match against the index must be either
1596 ** ASC or DESC. (Terms of the ORDER BY clause past the end of a UNIQUE
1597 ** index do not need to satisfy this constraint.) The *pbRev value is
1598 ** set to 1 if the ORDER BY clause is all DESC and it is set to 0 if
1599 ** the ORDER BY clause is all ASC.
1601 static int isSortingIndex(
1602 Parse *pParse, /* Parsing context */
1603 WhereMaskSet *pMaskSet, /* Mapping from table cursor numbers to bitmaps */
1604 Index *pIdx, /* The index we are testing */
1605 int base, /* Cursor number for the table to be sorted */
1606 ExprList *pOrderBy, /* The ORDER BY clause */
1607 int nEqCol, /* Number of index columns with == constraints */
1608 int wsFlags, /* Index usages flags */
1609 int *pbRev /* Set to 1 if ORDER BY is DESC */
1611 int i, j; /* Loop counters */
1612 int sortOrder = 0; /* XOR of index and ORDER BY sort direction */
1613 int nTerm; /* Number of ORDER BY terms */
1614 struct ExprList_item *pTerm; /* A term of the ORDER BY clause */
1615 sqlite3 *db = pParse->db;
1617 if( !pOrderBy ) return 0;
1618 if( wsFlags & WHERE_COLUMN_IN ) return 0;
1619 if( pIdx->bUnordered ) return 0;
1621 nTerm = pOrderBy->nExpr;
1622 assert( nTerm>0 );
1624 /* Argument pIdx must either point to a 'real' named index structure,
1625 ** or an index structure allocated on the stack by bestBtreeIndex() to
1626 ** represent the rowid index that is part of every table. */
1627 assert( pIdx->zName || (pIdx->nColumn==1 && pIdx->aiColumn[0]==-1) );
1629 /* Match terms of the ORDER BY clause against columns of
1630 ** the index.
1632 ** Note that indices have pIdx->nColumn regular columns plus
1633 ** one additional column containing the rowid. The rowid column
1634 ** of the index is also allowed to match against the ORDER BY
1635 ** clause.
1637 for(i=j=0, pTerm=pOrderBy->a; j<nTerm && i<=pIdx->nColumn; i++){
1638 Expr *pExpr; /* The expression of the ORDER BY pTerm */
1639 CollSeq *pColl; /* The collating sequence of pExpr */
1640 int termSortOrder; /* Sort order for this term */
1641 int iColumn; /* The i-th column of the index. -1 for rowid */
1642 int iSortOrder; /* 1 for DESC, 0 for ASC on the i-th index term */
1643 const char *zColl; /* Name of the collating sequence for i-th index term */
1645 pExpr = pTerm->pExpr;
1646 if( pExpr->op!=TK_COLUMN || pExpr->iTable!=base ){
1647 /* Can not use an index sort on anything that is not a column in the
1648 ** left-most table of the FROM clause */
1649 break;
1651 pColl = sqlite3ExprCollSeq(pParse, pExpr);
1652 if( !pColl ){
1653 pColl = db->pDfltColl;
1655 if( pIdx->zName && i<pIdx->nColumn ){
1656 iColumn = pIdx->aiColumn[i];
1657 if( iColumn==pIdx->pTable->iPKey ){
1658 iColumn = -1;
1660 iSortOrder = pIdx->aSortOrder[i];
1661 zColl = pIdx->azColl[i];
1662 }else{
1663 iColumn = -1;
1664 iSortOrder = 0;
1665 zColl = pColl->zName;
1667 if( pExpr->iColumn!=iColumn || sqlite3StrICmp(pColl->zName, zColl) ){
1668 /* Term j of the ORDER BY clause does not match column i of the index */
1669 if( i<nEqCol ){
1670 /* If an index column that is constrained by == fails to match an
1671 ** ORDER BY term, that is OK. Just ignore that column of the index
1673 continue;
1674 }else if( i==pIdx->nColumn ){
1675 /* Index column i is the rowid. All other terms match. */
1676 break;
1677 }else{
1678 /* If an index column fails to match and is not constrained by ==
1679 ** then the index cannot satisfy the ORDER BY constraint.
1681 return 0;
1684 assert( pIdx->aSortOrder!=0 || iColumn==-1 );
1685 assert( pTerm->sortOrder==0 || pTerm->sortOrder==1 );
1686 assert( iSortOrder==0 || iSortOrder==1 );
1687 termSortOrder = iSortOrder ^ pTerm->sortOrder;
1688 if( i>nEqCol ){
1689 if( termSortOrder!=sortOrder ){
1690 /* Indices can only be used if all ORDER BY terms past the
1691 ** equality constraints are all either DESC or ASC. */
1692 return 0;
1694 }else{
1695 sortOrder = termSortOrder;
1697 j++;
1698 pTerm++;
1699 if( iColumn<0 && !referencesOtherTables(pOrderBy, pMaskSet, j, base) ){
1700 /* If the indexed column is the primary key and everything matches
1701 ** so far and none of the ORDER BY terms to the right reference other
1702 ** tables in the join, then we are assured that the index can be used
1703 ** to sort because the primary key is unique and so none of the other
1704 ** columns will make any difference
1706 j = nTerm;
1710 *pbRev = sortOrder!=0;
1711 if( j>=nTerm ){
1712 /* All terms of the ORDER BY clause are covered by this index so
1713 ** this index can be used for sorting. */
1714 return 1;
1716 if( pIdx->onError!=OE_None && i==pIdx->nColumn
1717 && (wsFlags & WHERE_COLUMN_NULL)==0
1718 && !referencesOtherTables(pOrderBy, pMaskSet, j, base) ){
1719 /* All terms of this index match some prefix of the ORDER BY clause
1720 ** and the index is UNIQUE and no terms on the tail of the ORDER BY
1721 ** clause reference other tables in a join. If this is all true then
1722 ** the order by clause is superfluous. Not that if the matching
1723 ** condition is IS NULL then the result is not necessarily unique
1724 ** even on a UNIQUE index, so disallow those cases. */
1725 return 1;
1727 return 0;
1731 ** Prepare a crude estimate of the logarithm of the input value.
1732 ** The results need not be exact. This is only used for estimating
1733 ** the total cost of performing operations with O(logN) or O(NlogN)
1734 ** complexity. Because N is just a guess, it is no great tragedy if
1735 ** logN is a little off.
1737 static double estLog(double N){
1738 double logN = 1;
1739 double x = 10;
1740 while( N>x ){
1741 logN += 1;
1742 x *= 10;
1744 return logN;
1748 ** Two routines for printing the content of an sqlite3_index_info
1749 ** structure. Used for testing and debugging only. If neither
1750 ** SQLITE_TEST or SQLITE_DEBUG are defined, then these routines
1751 ** are no-ops.
1753 #if !defined(SQLITE_OMIT_VIRTUALTABLE) && defined(SQLITE_DEBUG)
1754 static void TRACE_IDX_INPUTS(sqlite3_index_info *p){
1755 int i;
1756 if( !sqlite3WhereTrace ) return;
1757 for(i=0; i<p->nConstraint; i++){
1758 sqlite3DebugPrintf(" constraint[%d]: col=%d termid=%d op=%d usabled=%d\n",
1760 p->aConstraint[i].iColumn,
1761 p->aConstraint[i].iTermOffset,
1762 p->aConstraint[i].op,
1763 p->aConstraint[i].usable);
1765 for(i=0; i<p->nOrderBy; i++){
1766 sqlite3DebugPrintf(" orderby[%d]: col=%d desc=%d\n",
1768 p->aOrderBy[i].iColumn,
1769 p->aOrderBy[i].desc);
1772 static void TRACE_IDX_OUTPUTS(sqlite3_index_info *p){
1773 int i;
1774 if( !sqlite3WhereTrace ) return;
1775 for(i=0; i<p->nConstraint; i++){
1776 sqlite3DebugPrintf(" usage[%d]: argvIdx=%d omit=%d\n",
1778 p->aConstraintUsage[i].argvIndex,
1779 p->aConstraintUsage[i].omit);
1781 sqlite3DebugPrintf(" idxNum=%d\n", p->idxNum);
1782 sqlite3DebugPrintf(" idxStr=%s\n", p->idxStr);
1783 sqlite3DebugPrintf(" orderByConsumed=%d\n", p->orderByConsumed);
1784 sqlite3DebugPrintf(" estimatedCost=%g\n", p->estimatedCost);
1786 #else
1787 #define TRACE_IDX_INPUTS(A)
1788 #define TRACE_IDX_OUTPUTS(A)
1789 #endif
1792 ** Required because bestIndex() is called by bestOrClauseIndex()
1794 static void bestIndex(
1795 Parse*, WhereClause*, struct SrcList_item*,
1796 Bitmask, Bitmask, ExprList*, WhereCost*);
1799 ** This routine attempts to find an scanning strategy that can be used
1800 ** to optimize an 'OR' expression that is part of a WHERE clause.
1802 ** The table associated with FROM clause term pSrc may be either a
1803 ** regular B-Tree table or a virtual table.
1805 static void bestOrClauseIndex(
1806 Parse *pParse, /* The parsing context */
1807 WhereClause *pWC, /* The WHERE clause */
1808 struct SrcList_item *pSrc, /* The FROM clause term to search */
1809 Bitmask notReady, /* Mask of cursors not available for indexing */
1810 Bitmask notValid, /* Cursors not available for any purpose */
1811 ExprList *pOrderBy, /* The ORDER BY clause */
1812 WhereCost *pCost /* Lowest cost query plan */
1814 #ifndef SQLITE_OMIT_OR_OPTIMIZATION
1815 const int iCur = pSrc->iCursor; /* The cursor of the table to be accessed */
1816 const Bitmask maskSrc = getMask(pWC->pMaskSet, iCur); /* Bitmask for pSrc */
1817 WhereTerm * const pWCEnd = &pWC->a[pWC->nTerm]; /* End of pWC->a[] */
1818 WhereTerm *pTerm; /* A single term of the WHERE clause */
1820 /* The OR-clause optimization is disallowed if the INDEXED BY or
1821 ** NOT INDEXED clauses are used or if the WHERE_AND_ONLY bit is set. */
1822 if( pSrc->notIndexed || pSrc->pIndex!=0 ){
1823 return;
1825 if( pWC->wctrlFlags & WHERE_AND_ONLY ){
1826 return;
1829 /* Search the WHERE clause terms for a usable WO_OR term. */
1830 for(pTerm=pWC->a; pTerm<pWCEnd; pTerm++){
1831 if( pTerm->eOperator==WO_OR
1832 && ((pTerm->prereqAll & ~maskSrc) & notReady)==0
1833 && (pTerm->u.pOrInfo->indexable & maskSrc)!=0
1835 WhereClause * const pOrWC = &pTerm->u.pOrInfo->wc;
1836 WhereTerm * const pOrWCEnd = &pOrWC->a[pOrWC->nTerm];
1837 WhereTerm *pOrTerm;
1838 int flags = WHERE_MULTI_OR;
1839 double rTotal = 0;
1840 double nRow = 0;
1841 Bitmask used = 0;
1843 for(pOrTerm=pOrWC->a; pOrTerm<pOrWCEnd; pOrTerm++){
1844 WhereCost sTermCost;
1845 WHERETRACE(("... Multi-index OR testing for term %d of %d....\n",
1846 (pOrTerm - pOrWC->a), (pTerm - pWC->a)
1848 if( pOrTerm->eOperator==WO_AND ){
1849 WhereClause *pAndWC = &pOrTerm->u.pAndInfo->wc;
1850 bestIndex(pParse, pAndWC, pSrc, notReady, notValid, 0, &sTermCost);
1851 }else if( pOrTerm->leftCursor==iCur ){
1852 WhereClause tempWC;
1853 tempWC.pParse = pWC->pParse;
1854 tempWC.pMaskSet = pWC->pMaskSet;
1855 tempWC.pOuter = pWC;
1856 tempWC.op = TK_AND;
1857 tempWC.a = pOrTerm;
1858 tempWC.wctrlFlags = 0;
1859 tempWC.nTerm = 1;
1860 bestIndex(pParse, &tempWC, pSrc, notReady, notValid, 0, &sTermCost);
1861 }else{
1862 continue;
1864 rTotal += sTermCost.rCost;
1865 nRow += sTermCost.plan.nRow;
1866 used |= sTermCost.used;
1867 if( rTotal>=pCost->rCost ) break;
1870 /* If there is an ORDER BY clause, increase the scan cost to account
1871 ** for the cost of the sort. */
1872 if( pOrderBy!=0 ){
1873 WHERETRACE(("... sorting increases OR cost %.9g to %.9g\n",
1874 rTotal, rTotal+nRow*estLog(nRow)));
1875 rTotal += nRow*estLog(nRow);
1878 /* If the cost of scanning using this OR term for optimization is
1879 ** less than the current cost stored in pCost, replace the contents
1880 ** of pCost. */
1881 WHERETRACE(("... multi-index OR cost=%.9g nrow=%.9g\n", rTotal, nRow));
1882 if( rTotal<pCost->rCost ){
1883 pCost->rCost = rTotal;
1884 pCost->used = used;
1885 pCost->plan.nRow = nRow;
1886 pCost->plan.wsFlags = flags;
1887 pCost->plan.u.pTerm = pTerm;
1891 #endif /* SQLITE_OMIT_OR_OPTIMIZATION */
1894 #ifndef SQLITE_OMIT_AUTOMATIC_INDEX
1896 ** Return TRUE if the WHERE clause term pTerm is of a form where it
1897 ** could be used with an index to access pSrc, assuming an appropriate
1898 ** index existed.
1900 static int termCanDriveIndex(
1901 WhereTerm *pTerm, /* WHERE clause term to check */
1902 struct SrcList_item *pSrc, /* Table we are trying to access */
1903 Bitmask notReady /* Tables in outer loops of the join */
1905 char aff;
1906 if( pTerm->leftCursor!=pSrc->iCursor ) return 0;
1907 if( pTerm->eOperator!=WO_EQ ) return 0;
1908 if( (pTerm->prereqRight & notReady)!=0 ) return 0;
1909 aff = pSrc->pTab->aCol[pTerm->u.leftColumn].affinity;
1910 if( !sqlite3IndexAffinityOk(pTerm->pExpr, aff) ) return 0;
1911 return 1;
1913 #endif
1915 #ifndef SQLITE_OMIT_AUTOMATIC_INDEX
1917 ** If the query plan for pSrc specified in pCost is a full table scan
1918 ** and indexing is allows (if there is no NOT INDEXED clause) and it
1919 ** possible to construct a transient index that would perform better
1920 ** than a full table scan even when the cost of constructing the index
1921 ** is taken into account, then alter the query plan to use the
1922 ** transient index.
1924 static void bestAutomaticIndex(
1925 Parse *pParse, /* The parsing context */
1926 WhereClause *pWC, /* The WHERE clause */
1927 struct SrcList_item *pSrc, /* The FROM clause term to search */
1928 Bitmask notReady, /* Mask of cursors that are not available */
1929 WhereCost *pCost /* Lowest cost query plan */
1931 double nTableRow; /* Rows in the input table */
1932 double logN; /* log(nTableRow) */
1933 double costTempIdx; /* per-query cost of the transient index */
1934 WhereTerm *pTerm; /* A single term of the WHERE clause */
1935 WhereTerm *pWCEnd; /* End of pWC->a[] */
1936 Table *pTable; /* Table tht might be indexed */
1938 if( pParse->nQueryLoop<=(double)1 ){
1939 /* There is no point in building an automatic index for a single scan */
1940 return;
1942 if( (pParse->db->flags & SQLITE_AutoIndex)==0 ){
1943 /* Automatic indices are disabled at run-time */
1944 return;
1946 if( (pCost->plan.wsFlags & WHERE_NOT_FULLSCAN)!=0 ){
1947 /* We already have some kind of index in use for this query. */
1948 return;
1950 if( pSrc->notIndexed ){
1951 /* The NOT INDEXED clause appears in the SQL. */
1952 return;
1954 if( pSrc->isCorrelated ){
1955 /* The source is a correlated sub-query. No point in indexing it. */
1956 return;
1959 assert( pParse->nQueryLoop >= (double)1 );
1960 pTable = pSrc->pTab;
1961 nTableRow = pTable->nRowEst;
1962 logN = estLog(nTableRow);
1963 costTempIdx = 2*logN*(nTableRow/pParse->nQueryLoop + 1);
1964 if( costTempIdx>=pCost->rCost ){
1965 /* The cost of creating the transient table would be greater than
1966 ** doing the full table scan */
1967 return;
1970 /* Search for any equality comparison term */
1971 pWCEnd = &pWC->a[pWC->nTerm];
1972 for(pTerm=pWC->a; pTerm<pWCEnd; pTerm++){
1973 if( termCanDriveIndex(pTerm, pSrc, notReady) ){
1974 WHERETRACE(("auto-index reduces cost from %.1f to %.1f\n",
1975 pCost->rCost, costTempIdx));
1976 pCost->rCost = costTempIdx;
1977 pCost->plan.nRow = logN + 1;
1978 pCost->plan.wsFlags = WHERE_TEMP_INDEX;
1979 pCost->used = pTerm->prereqRight;
1980 break;
1984 #else
1985 # define bestAutomaticIndex(A,B,C,D,E) /* no-op */
1986 #endif /* SQLITE_OMIT_AUTOMATIC_INDEX */
1989 #ifndef SQLITE_OMIT_AUTOMATIC_INDEX
1991 ** Generate code to construct the Index object for an automatic index
1992 ** and to set up the WhereLevel object pLevel so that the code generator
1993 ** makes use of the automatic index.
1995 static void constructAutomaticIndex(
1996 Parse *pParse, /* The parsing context */
1997 WhereClause *pWC, /* The WHERE clause */
1998 struct SrcList_item *pSrc, /* The FROM clause term to get the next index */
1999 Bitmask notReady, /* Mask of cursors that are not available */
2000 WhereLevel *pLevel /* Write new index here */
2002 int nColumn; /* Number of columns in the constructed index */
2003 WhereTerm *pTerm; /* A single term of the WHERE clause */
2004 WhereTerm *pWCEnd; /* End of pWC->a[] */
2005 int nByte; /* Byte of memory needed for pIdx */
2006 Index *pIdx; /* Object describing the transient index */
2007 Vdbe *v; /* Prepared statement under construction */
2008 int regIsInit; /* Register set by initialization */
2009 int addrInit; /* Address of the initialization bypass jump */
2010 Table *pTable; /* The table being indexed */
2011 KeyInfo *pKeyinfo; /* Key information for the index */
2012 int addrTop; /* Top of the index fill loop */
2013 int regRecord; /* Register holding an index record */
2014 int n; /* Column counter */
2015 int i; /* Loop counter */
2016 int mxBitCol; /* Maximum column in pSrc->colUsed */
2017 CollSeq *pColl; /* Collating sequence to on a column */
2018 Bitmask idxCols; /* Bitmap of columns used for indexing */
2019 Bitmask extraCols; /* Bitmap of additional columns */
2021 /* Generate code to skip over the creation and initialization of the
2022 ** transient index on 2nd and subsequent iterations of the loop. */
2023 v = pParse->pVdbe;
2024 assert( v!=0 );
2025 regIsInit = ++pParse->nMem;
2026 addrInit = sqlite3VdbeAddOp1(v, OP_Once, regIsInit);
2028 /* Count the number of columns that will be added to the index
2029 ** and used to match WHERE clause constraints */
2030 nColumn = 0;
2031 pTable = pSrc->pTab;
2032 pWCEnd = &pWC->a[pWC->nTerm];
2033 idxCols = 0;
2034 for(pTerm=pWC->a; pTerm<pWCEnd; pTerm++){
2035 if( termCanDriveIndex(pTerm, pSrc, notReady) ){
2036 int iCol = pTerm->u.leftColumn;
2037 Bitmask cMask = iCol>=BMS ? ((Bitmask)1)<<(BMS-1) : ((Bitmask)1)<<iCol;
2038 testcase( iCol==BMS );
2039 testcase( iCol==BMS-1 );
2040 if( (idxCols & cMask)==0 ){
2041 nColumn++;
2042 idxCols |= cMask;
2046 assert( nColumn>0 );
2047 pLevel->plan.nEq = nColumn;
2049 /* Count the number of additional columns needed to create a
2050 ** covering index. A "covering index" is an index that contains all
2051 ** columns that are needed by the query. With a covering index, the
2052 ** original table never needs to be accessed. Automatic indices must
2053 ** be a covering index because the index will not be updated if the
2054 ** original table changes and the index and table cannot both be used
2055 ** if they go out of sync.
2057 extraCols = pSrc->colUsed & (~idxCols | (((Bitmask)1)<<(BMS-1)));
2058 mxBitCol = (pTable->nCol >= BMS-1) ? BMS-1 : pTable->nCol;
2059 testcase( pTable->nCol==BMS-1 );
2060 testcase( pTable->nCol==BMS-2 );
2061 for(i=0; i<mxBitCol; i++){
2062 if( extraCols & (((Bitmask)1)<<i) ) nColumn++;
2064 if( pSrc->colUsed & (((Bitmask)1)<<(BMS-1)) ){
2065 nColumn += pTable->nCol - BMS + 1;
2067 pLevel->plan.wsFlags |= WHERE_COLUMN_EQ | WHERE_IDX_ONLY | WO_EQ;
2069 /* Construct the Index object to describe this index */
2070 nByte = sizeof(Index);
2071 nByte += nColumn*sizeof(int); /* Index.aiColumn */
2072 nByte += nColumn*sizeof(char*); /* Index.azColl */
2073 nByte += nColumn; /* Index.aSortOrder */
2074 pIdx = sqlite3DbMallocZero(pParse->db, nByte);
2075 if( pIdx==0 ) return;
2076 pLevel->plan.u.pIdx = pIdx;
2077 pIdx->azColl = (char**)&pIdx[1];
2078 pIdx->aiColumn = (int*)&pIdx->azColl[nColumn];
2079 pIdx->aSortOrder = (u8*)&pIdx->aiColumn[nColumn];
2080 pIdx->zName = "auto-index";
2081 pIdx->nColumn = nColumn;
2082 pIdx->pTable = pTable;
2083 n = 0;
2084 idxCols = 0;
2085 for(pTerm=pWC->a; pTerm<pWCEnd; pTerm++){
2086 if( termCanDriveIndex(pTerm, pSrc, notReady) ){
2087 int iCol = pTerm->u.leftColumn;
2088 Bitmask cMask = iCol>=BMS ? ((Bitmask)1)<<(BMS-1) : ((Bitmask)1)<<iCol;
2089 if( (idxCols & cMask)==0 ){
2090 Expr *pX = pTerm->pExpr;
2091 idxCols |= cMask;
2092 pIdx->aiColumn[n] = pTerm->u.leftColumn;
2093 pColl = sqlite3BinaryCompareCollSeq(pParse, pX->pLeft, pX->pRight);
2094 pIdx->azColl[n] = ALWAYS(pColl) ? pColl->zName : "BINARY";
2095 n++;
2099 assert( (u32)n==pLevel->plan.nEq );
2101 /* Add additional columns needed to make the automatic index into
2102 ** a covering index */
2103 for(i=0; i<mxBitCol; i++){
2104 if( extraCols & (((Bitmask)1)<<i) ){
2105 pIdx->aiColumn[n] = i;
2106 pIdx->azColl[n] = "BINARY";
2107 n++;
2110 if( pSrc->colUsed & (((Bitmask)1)<<(BMS-1)) ){
2111 for(i=BMS-1; i<pTable->nCol; i++){
2112 pIdx->aiColumn[n] = i;
2113 pIdx->azColl[n] = "BINARY";
2114 n++;
2117 assert( n==nColumn );
2119 /* Create the automatic index */
2120 pKeyinfo = sqlite3IndexKeyinfo(pParse, pIdx);
2121 assert( pLevel->iIdxCur>=0 );
2122 sqlite3VdbeAddOp4(v, OP_OpenAutoindex, pLevel->iIdxCur, nColumn+1, 0,
2123 (char*)pKeyinfo, P4_KEYINFO_HANDOFF);
2124 VdbeComment((v, "for %s", pTable->zName));
2126 /* Fill the automatic index with content */
2127 addrTop = sqlite3VdbeAddOp1(v, OP_Rewind, pLevel->iTabCur);
2128 regRecord = sqlite3GetTempReg(pParse);
2129 sqlite3GenerateIndexKey(pParse, pIdx, pLevel->iTabCur, regRecord, 1);
2130 sqlite3VdbeAddOp2(v, OP_IdxInsert, pLevel->iIdxCur, regRecord);
2131 sqlite3VdbeChangeP5(v, OPFLAG_USESEEKRESULT);
2132 sqlite3VdbeAddOp2(v, OP_Next, pLevel->iTabCur, addrTop+1);
2133 sqlite3VdbeChangeP5(v, SQLITE_STMTSTATUS_AUTOINDEX);
2134 sqlite3VdbeJumpHere(v, addrTop);
2135 sqlite3ReleaseTempReg(pParse, regRecord);
2137 /* Jump here when skipping the initialization */
2138 sqlite3VdbeJumpHere(v, addrInit);
2140 #endif /* SQLITE_OMIT_AUTOMATIC_INDEX */
2142 #ifndef SQLITE_OMIT_VIRTUALTABLE
2144 ** Allocate and populate an sqlite3_index_info structure. It is the
2145 ** responsibility of the caller to eventually release the structure
2146 ** by passing the pointer returned by this function to sqlite3_free().
2148 static sqlite3_index_info *allocateIndexInfo(
2149 Parse *pParse,
2150 WhereClause *pWC,
2151 struct SrcList_item *pSrc,
2152 ExprList *pOrderBy
2154 int i, j;
2155 int nTerm;
2156 struct sqlite3_index_constraint *pIdxCons;
2157 struct sqlite3_index_orderby *pIdxOrderBy;
2158 struct sqlite3_index_constraint_usage *pUsage;
2159 WhereTerm *pTerm;
2160 int nOrderBy;
2161 sqlite3_index_info *pIdxInfo;
2163 WHERETRACE(("Recomputing index info for %s...\n", pSrc->pTab->zName));
2165 /* Count the number of possible WHERE clause constraints referring
2166 ** to this virtual table */
2167 for(i=nTerm=0, pTerm=pWC->a; i<pWC->nTerm; i++, pTerm++){
2168 if( pTerm->leftCursor != pSrc->iCursor ) continue;
2169 assert( (pTerm->eOperator&(pTerm->eOperator-1))==0 );
2170 testcase( pTerm->eOperator==WO_IN );
2171 testcase( pTerm->eOperator==WO_ISNULL );
2172 if( pTerm->eOperator & (WO_IN|WO_ISNULL) ) continue;
2173 if( pTerm->wtFlags & TERM_VNULL ) continue;
2174 nTerm++;
2177 /* If the ORDER BY clause contains only columns in the current
2178 ** virtual table then allocate space for the aOrderBy part of
2179 ** the sqlite3_index_info structure.
2181 nOrderBy = 0;
2182 if( pOrderBy ){
2183 for(i=0; i<pOrderBy->nExpr; i++){
2184 Expr *pExpr = pOrderBy->a[i].pExpr;
2185 if( pExpr->op!=TK_COLUMN || pExpr->iTable!=pSrc->iCursor ) break;
2187 if( i==pOrderBy->nExpr ){
2188 nOrderBy = pOrderBy->nExpr;
2192 /* Allocate the sqlite3_index_info structure
2194 pIdxInfo = sqlite3DbMallocZero(pParse->db, sizeof(*pIdxInfo)
2195 + (sizeof(*pIdxCons) + sizeof(*pUsage))*nTerm
2196 + sizeof(*pIdxOrderBy)*nOrderBy );
2197 if( pIdxInfo==0 ){
2198 sqlite3ErrorMsg(pParse, "out of memory");
2199 /* (double)0 In case of SQLITE_OMIT_FLOATING_POINT... */
2200 return 0;
2203 /* Initialize the structure. The sqlite3_index_info structure contains
2204 ** many fields that are declared "const" to prevent xBestIndex from
2205 ** changing them. We have to do some funky casting in order to
2206 ** initialize those fields.
2208 pIdxCons = (struct sqlite3_index_constraint*)&pIdxInfo[1];
2209 pIdxOrderBy = (struct sqlite3_index_orderby*)&pIdxCons[nTerm];
2210 pUsage = (struct sqlite3_index_constraint_usage*)&pIdxOrderBy[nOrderBy];
2211 *(int*)&pIdxInfo->nConstraint = nTerm;
2212 *(int*)&pIdxInfo->nOrderBy = nOrderBy;
2213 *(struct sqlite3_index_constraint**)&pIdxInfo->aConstraint = pIdxCons;
2214 *(struct sqlite3_index_orderby**)&pIdxInfo->aOrderBy = pIdxOrderBy;
2215 *(struct sqlite3_index_constraint_usage**)&pIdxInfo->aConstraintUsage =
2216 pUsage;
2218 for(i=j=0, pTerm=pWC->a; i<pWC->nTerm; i++, pTerm++){
2219 if( pTerm->leftCursor != pSrc->iCursor ) continue;
2220 assert( (pTerm->eOperator&(pTerm->eOperator-1))==0 );
2221 testcase( pTerm->eOperator==WO_IN );
2222 testcase( pTerm->eOperator==WO_ISNULL );
2223 if( pTerm->eOperator & (WO_IN|WO_ISNULL) ) continue;
2224 if( pTerm->wtFlags & TERM_VNULL ) continue;
2225 pIdxCons[j].iColumn = pTerm->u.leftColumn;
2226 pIdxCons[j].iTermOffset = i;
2227 pIdxCons[j].op = (u8)pTerm->eOperator;
2228 /* The direct assignment in the previous line is possible only because
2229 ** the WO_ and SQLITE_INDEX_CONSTRAINT_ codes are identical. The
2230 ** following asserts verify this fact. */
2231 assert( WO_EQ==SQLITE_INDEX_CONSTRAINT_EQ );
2232 assert( WO_LT==SQLITE_INDEX_CONSTRAINT_LT );
2233 assert( WO_LE==SQLITE_INDEX_CONSTRAINT_LE );
2234 assert( WO_GT==SQLITE_INDEX_CONSTRAINT_GT );
2235 assert( WO_GE==SQLITE_INDEX_CONSTRAINT_GE );
2236 assert( WO_MATCH==SQLITE_INDEX_CONSTRAINT_MATCH );
2237 assert( pTerm->eOperator & (WO_EQ|WO_LT|WO_LE|WO_GT|WO_GE|WO_MATCH) );
2238 j++;
2240 for(i=0; i<nOrderBy; i++){
2241 Expr *pExpr = pOrderBy->a[i].pExpr;
2242 pIdxOrderBy[i].iColumn = pExpr->iColumn;
2243 pIdxOrderBy[i].desc = pOrderBy->a[i].sortOrder;
2246 return pIdxInfo;
2250 ** The table object reference passed as the second argument to this function
2251 ** must represent a virtual table. This function invokes the xBestIndex()
2252 ** method of the virtual table with the sqlite3_index_info pointer passed
2253 ** as the argument.
2255 ** If an error occurs, pParse is populated with an error message and a
2256 ** non-zero value is returned. Otherwise, 0 is returned and the output
2257 ** part of the sqlite3_index_info structure is left populated.
2259 ** Whether or not an error is returned, it is the responsibility of the
2260 ** caller to eventually free p->idxStr if p->needToFreeIdxStr indicates
2261 ** that this is required.
2263 static int vtabBestIndex(Parse *pParse, Table *pTab, sqlite3_index_info *p){
2264 sqlite3_vtab *pVtab = sqlite3GetVTable(pParse->db, pTab)->pVtab;
2265 int i;
2266 int rc;
2268 WHERETRACE(("xBestIndex for %s\n", pTab->zName));
2269 TRACE_IDX_INPUTS(p);
2270 rc = pVtab->pModule->xBestIndex(pVtab, p);
2271 TRACE_IDX_OUTPUTS(p);
2273 if( rc!=SQLITE_OK ){
2274 if( rc==SQLITE_NOMEM ){
2275 pParse->db->mallocFailed = 1;
2276 }else if( !pVtab->zErrMsg ){
2277 sqlite3ErrorMsg(pParse, "%s", sqlite3ErrStr(rc));
2278 }else{
2279 sqlite3ErrorMsg(pParse, "%s", pVtab->zErrMsg);
2282 sqlite3_free(pVtab->zErrMsg);
2283 pVtab->zErrMsg = 0;
2285 for(i=0; i<p->nConstraint; i++){
2286 if( !p->aConstraint[i].usable && p->aConstraintUsage[i].argvIndex>0 ){
2287 sqlite3ErrorMsg(pParse,
2288 "table %s: xBestIndex returned an invalid plan", pTab->zName);
2292 return pParse->nErr;
2297 ** Compute the best index for a virtual table.
2299 ** The best index is computed by the xBestIndex method of the virtual
2300 ** table module. This routine is really just a wrapper that sets up
2301 ** the sqlite3_index_info structure that is used to communicate with
2302 ** xBestIndex.
2304 ** In a join, this routine might be called multiple times for the
2305 ** same virtual table. The sqlite3_index_info structure is created
2306 ** and initialized on the first invocation and reused on all subsequent
2307 ** invocations. The sqlite3_index_info structure is also used when
2308 ** code is generated to access the virtual table. The whereInfoDelete()
2309 ** routine takes care of freeing the sqlite3_index_info structure after
2310 ** everybody has finished with it.
2312 static void bestVirtualIndex(
2313 Parse *pParse, /* The parsing context */
2314 WhereClause *pWC, /* The WHERE clause */
2315 struct SrcList_item *pSrc, /* The FROM clause term to search */
2316 Bitmask notReady, /* Mask of cursors not available for index */
2317 Bitmask notValid, /* Cursors not valid for any purpose */
2318 ExprList *pOrderBy, /* The order by clause */
2319 WhereCost *pCost, /* Lowest cost query plan */
2320 sqlite3_index_info **ppIdxInfo /* Index information passed to xBestIndex */
2322 Table *pTab = pSrc->pTab;
2323 sqlite3_index_info *pIdxInfo;
2324 struct sqlite3_index_constraint *pIdxCons;
2325 struct sqlite3_index_constraint_usage *pUsage;
2326 WhereTerm *pTerm;
2327 int i, j;
2328 int nOrderBy;
2329 double rCost;
2331 /* Make sure wsFlags is initialized to some sane value. Otherwise, if the
2332 ** malloc in allocateIndexInfo() fails and this function returns leaving
2333 ** wsFlags in an uninitialized state, the caller may behave unpredictably.
2335 memset(pCost, 0, sizeof(*pCost));
2336 pCost->plan.wsFlags = WHERE_VIRTUALTABLE;
2338 /* If the sqlite3_index_info structure has not been previously
2339 ** allocated and initialized, then allocate and initialize it now.
2341 pIdxInfo = *ppIdxInfo;
2342 if( pIdxInfo==0 ){
2343 *ppIdxInfo = pIdxInfo = allocateIndexInfo(pParse, pWC, pSrc, pOrderBy);
2345 if( pIdxInfo==0 ){
2346 return;
2349 /* At this point, the sqlite3_index_info structure that pIdxInfo points
2350 ** to will have been initialized, either during the current invocation or
2351 ** during some prior invocation. Now we just have to customize the
2352 ** details of pIdxInfo for the current invocation and pass it to
2353 ** xBestIndex.
2356 /* The module name must be defined. Also, by this point there must
2357 ** be a pointer to an sqlite3_vtab structure. Otherwise
2358 ** sqlite3ViewGetColumnNames() would have picked up the error.
2360 assert( pTab->azModuleArg && pTab->azModuleArg[0] );
2361 assert( sqlite3GetVTable(pParse->db, pTab) );
2363 /* Set the aConstraint[].usable fields and initialize all
2364 ** output variables to zero.
2366 ** aConstraint[].usable is true for constraints where the right-hand
2367 ** side contains only references to tables to the left of the current
2368 ** table. In other words, if the constraint is of the form:
2370 ** column = expr
2372 ** and we are evaluating a join, then the constraint on column is
2373 ** only valid if all tables referenced in expr occur to the left
2374 ** of the table containing column.
2376 ** The aConstraints[] array contains entries for all constraints
2377 ** on the current table. That way we only have to compute it once
2378 ** even though we might try to pick the best index multiple times.
2379 ** For each attempt at picking an index, the order of tables in the
2380 ** join might be different so we have to recompute the usable flag
2381 ** each time.
2383 pIdxCons = *(struct sqlite3_index_constraint**)&pIdxInfo->aConstraint;
2384 pUsage = pIdxInfo->aConstraintUsage;
2385 for(i=0; i<pIdxInfo->nConstraint; i++, pIdxCons++){
2386 j = pIdxCons->iTermOffset;
2387 pTerm = &pWC->a[j];
2388 pIdxCons->usable = (pTerm->prereqRight&notReady) ? 0 : 1;
2390 memset(pUsage, 0, sizeof(pUsage[0])*pIdxInfo->nConstraint);
2391 if( pIdxInfo->needToFreeIdxStr ){
2392 sqlite3_free(pIdxInfo->idxStr);
2394 pIdxInfo->idxStr = 0;
2395 pIdxInfo->idxNum = 0;
2396 pIdxInfo->needToFreeIdxStr = 0;
2397 pIdxInfo->orderByConsumed = 0;
2398 /* ((double)2) In case of SQLITE_OMIT_FLOATING_POINT... */
2399 pIdxInfo->estimatedCost = SQLITE_BIG_DBL / ((double)2);
2400 nOrderBy = pIdxInfo->nOrderBy;
2401 if( !pOrderBy ){
2402 pIdxInfo->nOrderBy = 0;
2405 if( vtabBestIndex(pParse, pTab, pIdxInfo) ){
2406 return;
2409 pIdxCons = *(struct sqlite3_index_constraint**)&pIdxInfo->aConstraint;
2410 for(i=0; i<pIdxInfo->nConstraint; i++){
2411 if( pUsage[i].argvIndex>0 ){
2412 pCost->used |= pWC->a[pIdxCons[i].iTermOffset].prereqRight;
2416 /* If there is an ORDER BY clause, and the selected virtual table index
2417 ** does not satisfy it, increase the cost of the scan accordingly. This
2418 ** matches the processing for non-virtual tables in bestBtreeIndex().
2420 rCost = pIdxInfo->estimatedCost;
2421 if( pOrderBy && pIdxInfo->orderByConsumed==0 ){
2422 rCost += estLog(rCost)*rCost;
2425 /* The cost is not allowed to be larger than SQLITE_BIG_DBL (the
2426 ** inital value of lowestCost in this loop. If it is, then the
2427 ** (cost<lowestCost) test below will never be true.
2429 ** Use "(double)2" instead of "2.0" in case OMIT_FLOATING_POINT
2430 ** is defined.
2432 if( (SQLITE_BIG_DBL/((double)2))<rCost ){
2433 pCost->rCost = (SQLITE_BIG_DBL/((double)2));
2434 }else{
2435 pCost->rCost = rCost;
2437 pCost->plan.u.pVtabIdx = pIdxInfo;
2438 if( pIdxInfo->orderByConsumed ){
2439 pCost->plan.wsFlags |= WHERE_ORDERBY;
2441 pCost->plan.nEq = 0;
2442 pIdxInfo->nOrderBy = nOrderBy;
2444 /* Try to find a more efficient access pattern by using multiple indexes
2445 ** to optimize an OR expression within the WHERE clause.
2447 bestOrClauseIndex(pParse, pWC, pSrc, notReady, notValid, pOrderBy, pCost);
2449 #endif /* SQLITE_OMIT_VIRTUALTABLE */
2451 #ifdef SQLITE_ENABLE_STAT3
2453 ** Estimate the location of a particular key among all keys in an
2454 ** index. Store the results in aStat as follows:
2456 ** aStat[0] Est. number of rows less than pVal
2457 ** aStat[1] Est. number of rows equal to pVal
2459 ** Return SQLITE_OK on success.
2461 static int whereKeyStats(
2462 Parse *pParse, /* Database connection */
2463 Index *pIdx, /* Index to consider domain of */
2464 sqlite3_value *pVal, /* Value to consider */
2465 int roundUp, /* Round up if true. Round down if false */
2466 tRowcnt *aStat /* OUT: stats written here */
2468 tRowcnt n;
2469 IndexSample *aSample;
2470 int i, eType;
2471 int isEq = 0;
2472 i64 v;
2473 double r, rS;
2475 assert( roundUp==0 || roundUp==1 );
2476 assert( pIdx->nSample>0 );
2477 if( pVal==0 ) return SQLITE_ERROR;
2478 n = pIdx->aiRowEst[0];
2479 aSample = pIdx->aSample;
2480 eType = sqlite3_value_type(pVal);
2482 if( eType==SQLITE_INTEGER ){
2483 v = sqlite3_value_int64(pVal);
2484 r = (i64)v;
2485 for(i=0; i<pIdx->nSample; i++){
2486 if( aSample[i].eType==SQLITE_NULL ) continue;
2487 if( aSample[i].eType>=SQLITE_TEXT ) break;
2488 if( aSample[i].eType==SQLITE_INTEGER ){
2489 if( aSample[i].u.i>=v ){
2490 isEq = aSample[i].u.i==v;
2491 break;
2493 }else{
2494 assert( aSample[i].eType==SQLITE_FLOAT );
2495 if( aSample[i].u.r>=r ){
2496 isEq = aSample[i].u.r==r;
2497 break;
2501 }else if( eType==SQLITE_FLOAT ){
2502 r = sqlite3_value_double(pVal);
2503 for(i=0; i<pIdx->nSample; i++){
2504 if( aSample[i].eType==SQLITE_NULL ) continue;
2505 if( aSample[i].eType>=SQLITE_TEXT ) break;
2506 if( aSample[i].eType==SQLITE_FLOAT ){
2507 rS = aSample[i].u.r;
2508 }else{
2509 rS = aSample[i].u.i;
2511 if( rS>=r ){
2512 isEq = rS==r;
2513 break;
2516 }else if( eType==SQLITE_NULL ){
2517 i = 0;
2518 if( aSample[0].eType==SQLITE_NULL ) isEq = 1;
2519 }else{
2520 assert( eType==SQLITE_TEXT || eType==SQLITE_BLOB );
2521 for(i=0; i<pIdx->nSample; i++){
2522 if( aSample[i].eType==SQLITE_TEXT || aSample[i].eType==SQLITE_BLOB ){
2523 break;
2526 if( i<pIdx->nSample ){
2527 sqlite3 *db = pParse->db;
2528 CollSeq *pColl;
2529 const u8 *z;
2530 if( eType==SQLITE_BLOB ){
2531 z = (const u8 *)sqlite3_value_blob(pVal);
2532 pColl = db->pDfltColl;
2533 assert( pColl->enc==SQLITE_UTF8 );
2534 }else{
2535 pColl = sqlite3GetCollSeq(db, SQLITE_UTF8, 0, *pIdx->azColl);
2536 if( pColl==0 ){
2537 sqlite3ErrorMsg(pParse, "no such collation sequence: %s",
2538 *pIdx->azColl);
2539 return SQLITE_ERROR;
2541 z = (const u8 *)sqlite3ValueText(pVal, pColl->enc);
2542 if( !z ){
2543 return SQLITE_NOMEM;
2545 assert( z && pColl && pColl->xCmp );
2547 n = sqlite3ValueBytes(pVal, pColl->enc);
2549 for(; i<pIdx->nSample; i++){
2550 int c;
2551 int eSampletype = aSample[i].eType;
2552 if( eSampletype<eType ) continue;
2553 if( eSampletype!=eType ) break;
2554 #ifndef SQLITE_OMIT_UTF16
2555 if( pColl->enc!=SQLITE_UTF8 ){
2556 int nSample;
2557 char *zSample = sqlite3Utf8to16(
2558 db, pColl->enc, aSample[i].u.z, aSample[i].nByte, &nSample
2560 if( !zSample ){
2561 assert( db->mallocFailed );
2562 return SQLITE_NOMEM;
2564 c = pColl->xCmp(pColl->pUser, nSample, zSample, n, z);
2565 sqlite3DbFree(db, zSample);
2566 }else
2567 #endif
2569 c = pColl->xCmp(pColl->pUser, aSample[i].nByte, aSample[i].u.z, n, z);
2571 if( c>=0 ){
2572 if( c==0 ) isEq = 1;
2573 break;
2579 /* At this point, aSample[i] is the first sample that is greater than
2580 ** or equal to pVal. Or if i==pIdx->nSample, then all samples are less
2581 ** than pVal. If aSample[i]==pVal, then isEq==1.
2583 if( isEq ){
2584 assert( i<pIdx->nSample );
2585 aStat[0] = aSample[i].nLt;
2586 aStat[1] = aSample[i].nEq;
2587 }else{
2588 tRowcnt iLower, iUpper, iGap;
2589 if( i==0 ){
2590 iLower = 0;
2591 iUpper = aSample[0].nLt;
2592 }else{
2593 iUpper = i>=pIdx->nSample ? n : aSample[i].nLt;
2594 iLower = aSample[i-1].nEq + aSample[i-1].nLt;
2596 aStat[1] = pIdx->avgEq;
2597 if( iLower>=iUpper ){
2598 iGap = 0;
2599 }else{
2600 iGap = iUpper - iLower;
2602 if( roundUp ){
2603 iGap = (iGap*2)/3;
2604 }else{
2605 iGap = iGap/3;
2607 aStat[0] = iLower + iGap;
2609 return SQLITE_OK;
2611 #endif /* SQLITE_ENABLE_STAT3 */
2614 ** If expression pExpr represents a literal value, set *pp to point to
2615 ** an sqlite3_value structure containing the same value, with affinity
2616 ** aff applied to it, before returning. It is the responsibility of the
2617 ** caller to eventually release this structure by passing it to
2618 ** sqlite3ValueFree().
2620 ** If the current parse is a recompile (sqlite3Reprepare()) and pExpr
2621 ** is an SQL variable that currently has a non-NULL value bound to it,
2622 ** create an sqlite3_value structure containing this value, again with
2623 ** affinity aff applied to it, instead.
2625 ** If neither of the above apply, set *pp to NULL.
2627 ** If an error occurs, return an error code. Otherwise, SQLITE_OK.
2629 #ifdef SQLITE_ENABLE_STAT3
2630 static int valueFromExpr(
2631 Parse *pParse,
2632 Expr *pExpr,
2633 u8 aff,
2634 sqlite3_value **pp
2636 if( pExpr->op==TK_VARIABLE
2637 || (pExpr->op==TK_REGISTER && pExpr->op2==TK_VARIABLE)
2639 int iVar = pExpr->iColumn;
2640 sqlite3VdbeSetVarmask(pParse->pVdbe, iVar);
2641 *pp = sqlite3VdbeGetValue(pParse->pReprepare, iVar, aff);
2642 return SQLITE_OK;
2644 return sqlite3ValueFromExpr(pParse->db, pExpr, SQLITE_UTF8, aff, pp);
2646 #endif
2649 ** This function is used to estimate the number of rows that will be visited
2650 ** by scanning an index for a range of values. The range may have an upper
2651 ** bound, a lower bound, or both. The WHERE clause terms that set the upper
2652 ** and lower bounds are represented by pLower and pUpper respectively. For
2653 ** example, assuming that index p is on t1(a):
2655 ** ... FROM t1 WHERE a > ? AND a < ? ...
2656 ** |_____| |_____|
2657 ** | |
2658 ** pLower pUpper
2660 ** If either of the upper or lower bound is not present, then NULL is passed in
2661 ** place of the corresponding WhereTerm.
2663 ** The nEq parameter is passed the index of the index column subject to the
2664 ** range constraint. Or, equivalently, the number of equality constraints
2665 ** optimized by the proposed index scan. For example, assuming index p is
2666 ** on t1(a, b), and the SQL query is:
2668 ** ... FROM t1 WHERE a = ? AND b > ? AND b < ? ...
2670 ** then nEq should be passed the value 1 (as the range restricted column,
2671 ** b, is the second left-most column of the index). Or, if the query is:
2673 ** ... FROM t1 WHERE a > ? AND a < ? ...
2675 ** then nEq should be passed 0.
2677 ** The returned value is an integer divisor to reduce the estimated
2678 ** search space. A return value of 1 means that range constraints are
2679 ** no help at all. A return value of 2 means range constraints are
2680 ** expected to reduce the search space by half. And so forth...
2682 ** In the absence of sqlite_stat3 ANALYZE data, each range inequality
2683 ** reduces the search space by a factor of 4. Hence a single constraint (x>?)
2684 ** results in a return of 4 and a range constraint (x>? AND x<?) results
2685 ** in a return of 16.
2687 static int whereRangeScanEst(
2688 Parse *pParse, /* Parsing & code generating context */
2689 Index *p, /* The index containing the range-compared column; "x" */
2690 int nEq, /* index into p->aCol[] of the range-compared column */
2691 WhereTerm *pLower, /* Lower bound on the range. ex: "x>123" Might be NULL */
2692 WhereTerm *pUpper, /* Upper bound on the range. ex: "x<455" Might be NULL */
2693 double *pRangeDiv /* OUT: Reduce search space by this divisor */
2695 int rc = SQLITE_OK;
2697 #ifdef SQLITE_ENABLE_STAT3
2699 if( nEq==0 && p->nSample ){
2700 sqlite3_value *pRangeVal;
2701 tRowcnt iLower = 0;
2702 tRowcnt iUpper = p->aiRowEst[0];
2703 tRowcnt a[2];
2704 u8 aff = p->pTable->aCol[p->aiColumn[0]].affinity;
2706 if( pLower ){
2707 Expr *pExpr = pLower->pExpr->pRight;
2708 rc = valueFromExpr(pParse, pExpr, aff, &pRangeVal);
2709 assert( pLower->eOperator==WO_GT || pLower->eOperator==WO_GE );
2710 if( rc==SQLITE_OK
2711 && whereKeyStats(pParse, p, pRangeVal, 0, a)==SQLITE_OK
2713 iLower = a[0];
2714 if( pLower->eOperator==WO_GT ) iLower += a[1];
2716 sqlite3ValueFree(pRangeVal);
2718 if( rc==SQLITE_OK && pUpper ){
2719 Expr *pExpr = pUpper->pExpr->pRight;
2720 rc = valueFromExpr(pParse, pExpr, aff, &pRangeVal);
2721 assert( pUpper->eOperator==WO_LT || pUpper->eOperator==WO_LE );
2722 if( rc==SQLITE_OK
2723 && whereKeyStats(pParse, p, pRangeVal, 1, a)==SQLITE_OK
2725 iUpper = a[0];
2726 if( pUpper->eOperator==WO_LE ) iUpper += a[1];
2728 sqlite3ValueFree(pRangeVal);
2730 if( rc==SQLITE_OK ){
2731 if( iUpper<=iLower ){
2732 *pRangeDiv = (double)p->aiRowEst[0];
2733 }else{
2734 *pRangeDiv = (double)p->aiRowEst[0]/(double)(iUpper - iLower);
2736 WHERETRACE(("range scan regions: %u..%u div=%g\n",
2737 (u32)iLower, (u32)iUpper, *pRangeDiv));
2738 return SQLITE_OK;
2741 #else
2742 UNUSED_PARAMETER(pParse);
2743 UNUSED_PARAMETER(p);
2744 UNUSED_PARAMETER(nEq);
2745 #endif
2746 assert( pLower || pUpper );
2747 *pRangeDiv = (double)1;
2748 if( pLower && (pLower->wtFlags & TERM_VNULL)==0 ) *pRangeDiv *= (double)4;
2749 if( pUpper ) *pRangeDiv *= (double)4;
2750 return rc;
2753 #ifdef SQLITE_ENABLE_STAT3
2755 ** Estimate the number of rows that will be returned based on
2756 ** an equality constraint x=VALUE and where that VALUE occurs in
2757 ** the histogram data. This only works when x is the left-most
2758 ** column of an index and sqlite_stat3 histogram data is available
2759 ** for that index. When pExpr==NULL that means the constraint is
2760 ** "x IS NULL" instead of "x=VALUE".
2762 ** Write the estimated row count into *pnRow and return SQLITE_OK.
2763 ** If unable to make an estimate, leave *pnRow unchanged and return
2764 ** non-zero.
2766 ** This routine can fail if it is unable to load a collating sequence
2767 ** required for string comparison, or if unable to allocate memory
2768 ** for a UTF conversion required for comparison. The error is stored
2769 ** in the pParse structure.
2771 static int whereEqualScanEst(
2772 Parse *pParse, /* Parsing & code generating context */
2773 Index *p, /* The index whose left-most column is pTerm */
2774 Expr *pExpr, /* Expression for VALUE in the x=VALUE constraint */
2775 double *pnRow /* Write the revised row estimate here */
2777 sqlite3_value *pRhs = 0; /* VALUE on right-hand side of pTerm */
2778 u8 aff; /* Column affinity */
2779 int rc; /* Subfunction return code */
2780 tRowcnt a[2]; /* Statistics */
2782 assert( p->aSample!=0 );
2783 assert( p->nSample>0 );
2784 aff = p->pTable->aCol[p->aiColumn[0]].affinity;
2785 if( pExpr ){
2786 rc = valueFromExpr(pParse, pExpr, aff, &pRhs);
2787 if( rc ) goto whereEqualScanEst_cancel;
2788 }else{
2789 pRhs = sqlite3ValueNew(pParse->db);
2791 if( pRhs==0 ) return SQLITE_NOTFOUND;
2792 rc = whereKeyStats(pParse, p, pRhs, 0, a);
2793 if( rc==SQLITE_OK ){
2794 WHERETRACE(("equality scan regions: %d\n", (int)a[1]));
2795 *pnRow = a[1];
2797 whereEqualScanEst_cancel:
2798 sqlite3ValueFree(pRhs);
2799 return rc;
2801 #endif /* defined(SQLITE_ENABLE_STAT3) */
2803 #ifdef SQLITE_ENABLE_STAT3
2805 ** Estimate the number of rows that will be returned based on
2806 ** an IN constraint where the right-hand side of the IN operator
2807 ** is a list of values. Example:
2809 ** WHERE x IN (1,2,3,4)
2811 ** Write the estimated row count into *pnRow and return SQLITE_OK.
2812 ** If unable to make an estimate, leave *pnRow unchanged and return
2813 ** non-zero.
2815 ** This routine can fail if it is unable to load a collating sequence
2816 ** required for string comparison, or if unable to allocate memory
2817 ** for a UTF conversion required for comparison. The error is stored
2818 ** in the pParse structure.
2820 static int whereInScanEst(
2821 Parse *pParse, /* Parsing & code generating context */
2822 Index *p, /* The index whose left-most column is pTerm */
2823 ExprList *pList, /* The value list on the RHS of "x IN (v1,v2,v3,...)" */
2824 double *pnRow /* Write the revised row estimate here */
2826 int rc = SQLITE_OK; /* Subfunction return code */
2827 double nEst; /* Number of rows for a single term */
2828 double nRowEst = (double)0; /* New estimate of the number of rows */
2829 int i; /* Loop counter */
2831 assert( p->aSample!=0 );
2832 for(i=0; rc==SQLITE_OK && i<pList->nExpr; i++){
2833 nEst = p->aiRowEst[0];
2834 rc = whereEqualScanEst(pParse, p, pList->a[i].pExpr, &nEst);
2835 nRowEst += nEst;
2837 if( rc==SQLITE_OK ){
2838 if( nRowEst > p->aiRowEst[0] ) nRowEst = p->aiRowEst[0];
2839 *pnRow = nRowEst;
2840 WHERETRACE(("IN row estimate: est=%g\n", nRowEst));
2842 return rc;
2844 #endif /* defined(SQLITE_ENABLE_STAT3) */
2848 ** Find the best query plan for accessing a particular table. Write the
2849 ** best query plan and its cost into the WhereCost object supplied as the
2850 ** last parameter.
2852 ** The lowest cost plan wins. The cost is an estimate of the amount of
2853 ** CPU and disk I/O needed to process the requested result.
2854 ** Factors that influence cost include:
2856 ** * The estimated number of rows that will be retrieved. (The
2857 ** fewer the better.)
2859 ** * Whether or not sorting must occur.
2861 ** * Whether or not there must be separate lookups in the
2862 ** index and in the main table.
2864 ** If there was an INDEXED BY clause (pSrc->pIndex) attached to the table in
2865 ** the SQL statement, then this function only considers plans using the
2866 ** named index. If no such plan is found, then the returned cost is
2867 ** SQLITE_BIG_DBL. If a plan is found that uses the named index,
2868 ** then the cost is calculated in the usual way.
2870 ** If a NOT INDEXED clause (pSrc->notIndexed!=0) was attached to the table
2871 ** in the SELECT statement, then no indexes are considered. However, the
2872 ** selected plan may still take advantage of the built-in rowid primary key
2873 ** index.
2875 static void bestBtreeIndex(
2876 Parse *pParse, /* The parsing context */
2877 WhereClause *pWC, /* The WHERE clause */
2878 struct SrcList_item *pSrc, /* The FROM clause term to search */
2879 Bitmask notReady, /* Mask of cursors not available for indexing */
2880 Bitmask notValid, /* Cursors not available for any purpose */
2881 ExprList *pOrderBy, /* The ORDER BY clause */
2882 ExprList *pDistinct, /* The select-list if query is DISTINCT */
2883 WhereCost *pCost /* Lowest cost query plan */
2885 int iCur = pSrc->iCursor; /* The cursor of the table to be accessed */
2886 Index *pProbe; /* An index we are evaluating */
2887 Index *pIdx; /* Copy of pProbe, or zero for IPK index */
2888 int eqTermMask; /* Current mask of valid equality operators */
2889 int idxEqTermMask; /* Index mask of valid equality operators */
2890 Index sPk; /* A fake index object for the primary key */
2891 tRowcnt aiRowEstPk[2]; /* The aiRowEst[] value for the sPk index */
2892 int aiColumnPk = -1; /* The aColumn[] value for the sPk index */
2893 int wsFlagMask; /* Allowed flags in pCost->plan.wsFlag */
2895 /* Initialize the cost to a worst-case value */
2896 memset(pCost, 0, sizeof(*pCost));
2897 pCost->rCost = SQLITE_BIG_DBL;
2899 /* If the pSrc table is the right table of a LEFT JOIN then we may not
2900 ** use an index to satisfy IS NULL constraints on that table. This is
2901 ** because columns might end up being NULL if the table does not match -
2902 ** a circumstance which the index cannot help us discover. Ticket #2177.
2904 if( pSrc->jointype & JT_LEFT ){
2905 idxEqTermMask = WO_EQ|WO_IN;
2906 }else{
2907 idxEqTermMask = WO_EQ|WO_IN|WO_ISNULL;
2910 if( pSrc->pIndex ){
2911 /* An INDEXED BY clause specifies a particular index to use */
2912 pIdx = pProbe = pSrc->pIndex;
2913 wsFlagMask = ~(WHERE_ROWID_EQ|WHERE_ROWID_RANGE);
2914 eqTermMask = idxEqTermMask;
2915 }else{
2916 /* There is no INDEXED BY clause. Create a fake Index object in local
2917 ** variable sPk to represent the rowid primary key index. Make this
2918 ** fake index the first in a chain of Index objects with all of the real
2919 ** indices to follow */
2920 Index *pFirst; /* First of real indices on the table */
2921 memset(&sPk, 0, sizeof(Index));
2922 sPk.nColumn = 1;
2923 sPk.aiColumn = &aiColumnPk;
2924 sPk.aiRowEst = aiRowEstPk;
2925 sPk.onError = OE_Replace;
2926 sPk.pTable = pSrc->pTab;
2927 aiRowEstPk[0] = pSrc->pTab->nRowEst;
2928 aiRowEstPk[1] = 1;
2929 pFirst = pSrc->pTab->pIndex;
2930 if( pSrc->notIndexed==0 ){
2931 /* The real indices of the table are only considered if the
2932 ** NOT INDEXED qualifier is omitted from the FROM clause */
2933 sPk.pNext = pFirst;
2935 pProbe = &sPk;
2936 wsFlagMask = ~(
2937 WHERE_COLUMN_IN|WHERE_COLUMN_EQ|WHERE_COLUMN_NULL|WHERE_COLUMN_RANGE
2939 eqTermMask = WO_EQ|WO_IN;
2940 pIdx = 0;
2943 /* Loop over all indices looking for the best one to use
2945 for(; pProbe; pIdx=pProbe=pProbe->pNext){
2946 const tRowcnt * const aiRowEst = pProbe->aiRowEst;
2947 double cost; /* Cost of using pProbe */
2948 double nRow; /* Estimated number of rows in result set */
2949 double log10N = (double)1; /* base-10 logarithm of nRow (inexact) */
2950 int rev; /* True to scan in reverse order */
2951 int wsFlags = 0;
2952 Bitmask used = 0;
2954 /* The following variables are populated based on the properties of
2955 ** index being evaluated. They are then used to determine the expected
2956 ** cost and number of rows returned.
2958 ** nEq:
2959 ** Number of equality terms that can be implemented using the index.
2960 ** In other words, the number of initial fields in the index that
2961 ** are used in == or IN or NOT NULL constraints of the WHERE clause.
2963 ** nInMul:
2964 ** The "in-multiplier". This is an estimate of how many seek operations
2965 ** SQLite must perform on the index in question. For example, if the
2966 ** WHERE clause is:
2968 ** WHERE a IN (1, 2, 3) AND b IN (4, 5, 6)
2970 ** SQLite must perform 9 lookups on an index on (a, b), so nInMul is
2971 ** set to 9. Given the same schema and either of the following WHERE
2972 ** clauses:
2974 ** WHERE a = 1
2975 ** WHERE a >= 2
2977 ** nInMul is set to 1.
2979 ** If there exists a WHERE term of the form "x IN (SELECT ...)", then
2980 ** the sub-select is assumed to return 25 rows for the purposes of
2981 ** determining nInMul.
2983 ** bInEst:
2984 ** Set to true if there was at least one "x IN (SELECT ...)" term used
2985 ** in determining the value of nInMul. Note that the RHS of the
2986 ** IN operator must be a SELECT, not a value list, for this variable
2987 ** to be true.
2989 ** rangeDiv:
2990 ** An estimate of a divisor by which to reduce the search space due
2991 ** to inequality constraints. In the absence of sqlite_stat3 ANALYZE
2992 ** data, a single inequality reduces the search space to 1/4rd its
2993 ** original size (rangeDiv==4). Two inequalities reduce the search
2994 ** space to 1/16th of its original size (rangeDiv==16).
2996 ** bSort:
2997 ** Boolean. True if there is an ORDER BY clause that will require an
2998 ** external sort (i.e. scanning the index being evaluated will not
2999 ** correctly order records).
3001 ** bLookup:
3002 ** Boolean. True if a table lookup is required for each index entry
3003 ** visited. In other words, true if this is not a covering index.
3004 ** This is always false for the rowid primary key index of a table.
3005 ** For other indexes, it is true unless all the columns of the table
3006 ** used by the SELECT statement are present in the index (such an
3007 ** index is sometimes described as a covering index).
3008 ** For example, given the index on (a, b), the second of the following
3009 ** two queries requires table b-tree lookups in order to find the value
3010 ** of column c, but the first does not because columns a and b are
3011 ** both available in the index.
3013 ** SELECT a, b FROM tbl WHERE a = 1;
3014 ** SELECT a, b, c FROM tbl WHERE a = 1;
3016 int nEq; /* Number of == or IN terms matching index */
3017 int bInEst = 0; /* True if "x IN (SELECT...)" seen */
3018 int nInMul = 1; /* Number of distinct equalities to lookup */
3019 double rangeDiv = (double)1; /* Estimated reduction in search space */
3020 int nBound = 0; /* Number of range constraints seen */
3021 int bSort = !!pOrderBy; /* True if external sort required */
3022 int bDist = !!pDistinct; /* True if index cannot help with DISTINCT */
3023 int bLookup = 0; /* True if not a covering index */
3024 WhereTerm *pTerm; /* A single term of the WHERE clause */
3025 #ifdef SQLITE_ENABLE_STAT3
3026 WhereTerm *pFirstTerm = 0; /* First term matching the index */
3027 #endif
3029 /* Determine the values of nEq and nInMul */
3030 for(nEq=0; nEq<pProbe->nColumn; nEq++){
3031 int j = pProbe->aiColumn[nEq];
3032 pTerm = findTerm(pWC, iCur, j, notReady, eqTermMask, pIdx);
3033 if( pTerm==0 ) break;
3034 wsFlags |= (WHERE_COLUMN_EQ|WHERE_ROWID_EQ);
3035 testcase( pTerm->pWC!=pWC );
3036 if( pTerm->eOperator & WO_IN ){
3037 Expr *pExpr = pTerm->pExpr;
3038 wsFlags |= WHERE_COLUMN_IN;
3039 if( ExprHasProperty(pExpr, EP_xIsSelect) ){
3040 /* "x IN (SELECT ...)": Assume the SELECT returns 25 rows */
3041 nInMul *= 25;
3042 bInEst = 1;
3043 }else if( ALWAYS(pExpr->x.pList && pExpr->x.pList->nExpr) ){
3044 /* "x IN (value, value, ...)" */
3045 nInMul *= pExpr->x.pList->nExpr;
3047 }else if( pTerm->eOperator & WO_ISNULL ){
3048 wsFlags |= WHERE_COLUMN_NULL;
3050 #ifdef SQLITE_ENABLE_STAT3
3051 if( nEq==0 && pProbe->aSample ) pFirstTerm = pTerm;
3052 #endif
3053 used |= pTerm->prereqRight;
3056 /* Determine the value of rangeDiv */
3057 if( nEq<pProbe->nColumn && pProbe->bUnordered==0 ){
3058 int j = pProbe->aiColumn[nEq];
3059 if( findTerm(pWC, iCur, j, notReady, WO_LT|WO_LE|WO_GT|WO_GE, pIdx) ){
3060 WhereTerm *pTop = findTerm(pWC, iCur, j, notReady, WO_LT|WO_LE, pIdx);
3061 WhereTerm *pBtm = findTerm(pWC, iCur, j, notReady, WO_GT|WO_GE, pIdx);
3062 whereRangeScanEst(pParse, pProbe, nEq, pBtm, pTop, &rangeDiv);
3063 if( pTop ){
3064 nBound = 1;
3065 wsFlags |= WHERE_TOP_LIMIT;
3066 used |= pTop->prereqRight;
3067 testcase( pTop->pWC!=pWC );
3069 if( pBtm ){
3070 nBound++;
3071 wsFlags |= WHERE_BTM_LIMIT;
3072 used |= pBtm->prereqRight;
3073 testcase( pBtm->pWC!=pWC );
3075 wsFlags |= (WHERE_COLUMN_RANGE|WHERE_ROWID_RANGE);
3077 }else if( pProbe->onError!=OE_None ){
3078 testcase( wsFlags & WHERE_COLUMN_IN );
3079 testcase( wsFlags & WHERE_COLUMN_NULL );
3080 if( (wsFlags & (WHERE_COLUMN_IN|WHERE_COLUMN_NULL))==0 ){
3081 wsFlags |= WHERE_UNIQUE;
3085 /* If there is an ORDER BY clause and the index being considered will
3086 ** naturally scan rows in the required order, set the appropriate flags
3087 ** in wsFlags. Otherwise, if there is an ORDER BY clause but the index
3088 ** will scan rows in a different order, set the bSort variable. */
3089 if( isSortingIndex(
3090 pParse, pWC->pMaskSet, pProbe, iCur, pOrderBy, nEq, wsFlags, &rev)
3092 bSort = 0;
3093 wsFlags |= WHERE_ROWID_RANGE|WHERE_COLUMN_RANGE|WHERE_ORDERBY;
3094 wsFlags |= (rev ? WHERE_REVERSE : 0);
3097 /* If there is a DISTINCT qualifier and this index will scan rows in
3098 ** order of the DISTINCT expressions, clear bDist and set the appropriate
3099 ** flags in wsFlags. */
3100 if( isDistinctIndex(pParse, pWC, pProbe, iCur, pDistinct, nEq) ){
3101 bDist = 0;
3102 wsFlags |= WHERE_ROWID_RANGE|WHERE_COLUMN_RANGE|WHERE_DISTINCT;
3105 /* If currently calculating the cost of using an index (not the IPK
3106 ** index), determine if all required column data may be obtained without
3107 ** using the main table (i.e. if the index is a covering
3108 ** index for this query). If it is, set the WHERE_IDX_ONLY flag in
3109 ** wsFlags. Otherwise, set the bLookup variable to true. */
3110 if( pIdx && wsFlags ){
3111 Bitmask m = pSrc->colUsed;
3112 int j;
3113 for(j=0; j<pIdx->nColumn; j++){
3114 int x = pIdx->aiColumn[j];
3115 if( x<BMS-1 ){
3116 m &= ~(((Bitmask)1)<<x);
3119 if( m==0 ){
3120 wsFlags |= WHERE_IDX_ONLY;
3121 }else{
3122 bLookup = 1;
3127 ** Estimate the number of rows of output. For an "x IN (SELECT...)"
3128 ** constraint, do not let the estimate exceed half the rows in the table.
3130 nRow = (double)(aiRowEst[nEq] * nInMul);
3131 if( bInEst && nRow*2>aiRowEst[0] ){
3132 nRow = aiRowEst[0]/2;
3133 nInMul = (int)(nRow / aiRowEst[nEq]);
3136 #ifdef SQLITE_ENABLE_STAT3
3137 /* If the constraint is of the form x=VALUE or x IN (E1,E2,...)
3138 ** and we do not think that values of x are unique and if histogram
3139 ** data is available for column x, then it might be possible
3140 ** to get a better estimate on the number of rows based on
3141 ** VALUE and how common that value is according to the histogram.
3143 if( nRow>(double)1 && nEq==1 && pFirstTerm!=0 && aiRowEst[1]>1 ){
3144 assert( (pFirstTerm->eOperator & (WO_EQ|WO_ISNULL|WO_IN))!=0 );
3145 if( pFirstTerm->eOperator & (WO_EQ|WO_ISNULL) ){
3146 testcase( pFirstTerm->eOperator==WO_EQ );
3147 testcase( pFirstTerm->eOperator==WO_ISNULL );
3148 whereEqualScanEst(pParse, pProbe, pFirstTerm->pExpr->pRight, &nRow);
3149 }else if( bInEst==0 ){
3150 assert( pFirstTerm->eOperator==WO_IN );
3151 whereInScanEst(pParse, pProbe, pFirstTerm->pExpr->x.pList, &nRow);
3154 #endif /* SQLITE_ENABLE_STAT3 */
3156 /* Adjust the number of output rows and downward to reflect rows
3157 ** that are excluded by range constraints.
3159 nRow = nRow/rangeDiv;
3160 if( nRow<1 ) nRow = 1;
3162 /* Experiments run on real SQLite databases show that the time needed
3163 ** to do a binary search to locate a row in a table or index is roughly
3164 ** log10(N) times the time to move from one row to the next row within
3165 ** a table or index. The actual times can vary, with the size of
3166 ** records being an important factor. Both moves and searches are
3167 ** slower with larger records, presumably because fewer records fit
3168 ** on one page and hence more pages have to be fetched.
3170 ** The ANALYZE command and the sqlite_stat1 and sqlite_stat3 tables do
3171 ** not give us data on the relative sizes of table and index records.
3172 ** So this computation assumes table records are about twice as big
3173 ** as index records
3175 if( (wsFlags & WHERE_NOT_FULLSCAN)==0 ){
3176 /* The cost of a full table scan is a number of move operations equal
3177 ** to the number of rows in the table.
3179 ** We add an additional 4x penalty to full table scans. This causes
3180 ** the cost function to err on the side of choosing an index over
3181 ** choosing a full scan. This 4x full-scan penalty is an arguable
3182 ** decision and one which we expect to revisit in the future. But
3183 ** it seems to be working well enough at the moment.
3185 cost = aiRowEst[0]*4;
3186 }else{
3187 log10N = estLog(aiRowEst[0]);
3188 cost = nRow;
3189 if( pIdx ){
3190 if( bLookup ){
3191 /* For an index lookup followed by a table lookup:
3192 ** nInMul index searches to find the start of each index range
3193 ** + nRow steps through the index
3194 ** + nRow table searches to lookup the table entry using the rowid
3196 cost += (nInMul + nRow)*log10N;
3197 }else{
3198 /* For a covering index:
3199 ** nInMul index searches to find the initial entry
3200 ** + nRow steps through the index
3202 cost += nInMul*log10N;
3204 }else{
3205 /* For a rowid primary key lookup:
3206 ** nInMult table searches to find the initial entry for each range
3207 ** + nRow steps through the table
3209 cost += nInMul*log10N;
3213 /* Add in the estimated cost of sorting the result. Actual experimental
3214 ** measurements of sorting performance in SQLite show that sorting time
3215 ** adds C*N*log10(N) to the cost, where N is the number of rows to be
3216 ** sorted and C is a factor between 1.95 and 4.3. We will split the
3217 ** difference and select C of 3.0.
3219 if( bSort ){
3220 cost += nRow*estLog(nRow)*3;
3222 if( bDist ){
3223 cost += nRow*estLog(nRow)*3;
3226 /**** Cost of using this index has now been computed ****/
3228 /* If there are additional constraints on this table that cannot
3229 ** be used with the current index, but which might lower the number
3230 ** of output rows, adjust the nRow value accordingly. This only
3231 ** matters if the current index is the least costly, so do not bother
3232 ** with this step if we already know this index will not be chosen.
3233 ** Also, never reduce the output row count below 2 using this step.
3235 ** It is critical that the notValid mask be used here instead of
3236 ** the notReady mask. When computing an "optimal" index, the notReady
3237 ** mask will only have one bit set - the bit for the current table.
3238 ** The notValid mask, on the other hand, always has all bits set for
3239 ** tables that are not in outer loops. If notReady is used here instead
3240 ** of notValid, then a optimal index that depends on inner joins loops
3241 ** might be selected even when there exists an optimal index that has
3242 ** no such dependency.
3244 if( nRow>2 && cost<=pCost->rCost ){
3245 int k; /* Loop counter */
3246 int nSkipEq = nEq; /* Number of == constraints to skip */
3247 int nSkipRange = nBound; /* Number of < constraints to skip */
3248 Bitmask thisTab; /* Bitmap for pSrc */
3250 thisTab = getMask(pWC->pMaskSet, iCur);
3251 for(pTerm=pWC->a, k=pWC->nTerm; nRow>2 && k; k--, pTerm++){
3252 if( pTerm->wtFlags & TERM_VIRTUAL ) continue;
3253 if( (pTerm->prereqAll & notValid)!=thisTab ) continue;
3254 if( pTerm->eOperator & (WO_EQ|WO_IN|WO_ISNULL) ){
3255 if( nSkipEq ){
3256 /* Ignore the first nEq equality matches since the index
3257 ** has already accounted for these */
3258 nSkipEq--;
3259 }else{
3260 /* Assume each additional equality match reduces the result
3261 ** set size by a factor of 10 */
3262 nRow /= 10;
3264 }else if( pTerm->eOperator & (WO_LT|WO_LE|WO_GT|WO_GE) ){
3265 if( nSkipRange ){
3266 /* Ignore the first nSkipRange range constraints since the index
3267 ** has already accounted for these */
3268 nSkipRange--;
3269 }else{
3270 /* Assume each additional range constraint reduces the result
3271 ** set size by a factor of 3. Indexed range constraints reduce
3272 ** the search space by a larger factor: 4. We make indexed range
3273 ** more selective intentionally because of the subjective
3274 ** observation that indexed range constraints really are more
3275 ** selective in practice, on average. */
3276 nRow /= 3;
3278 }else if( pTerm->eOperator!=WO_NOOP ){
3279 /* Any other expression lowers the output row count by half */
3280 nRow /= 2;
3283 if( nRow<2 ) nRow = 2;
3287 WHERETRACE((
3288 "%s(%s): nEq=%d nInMul=%d rangeDiv=%d bSort=%d bLookup=%d wsFlags=0x%x\n"
3289 " notReady=0x%llx log10N=%.1f nRow=%.1f cost=%.1f used=0x%llx\n",
3290 pSrc->pTab->zName, (pIdx ? pIdx->zName : "ipk"),
3291 nEq, nInMul, (int)rangeDiv, bSort, bLookup, wsFlags,
3292 notReady, log10N, nRow, cost, used
3295 /* If this index is the best we have seen so far, then record this
3296 ** index and its cost in the pCost structure.
3298 if( (!pIdx || wsFlags)
3299 && (cost<pCost->rCost || (cost<=pCost->rCost && nRow<pCost->plan.nRow))
3301 pCost->rCost = cost;
3302 pCost->used = used;
3303 pCost->plan.nRow = nRow;
3304 pCost->plan.wsFlags = (wsFlags&wsFlagMask);
3305 pCost->plan.nEq = nEq;
3306 pCost->plan.u.pIdx = pIdx;
3309 /* If there was an INDEXED BY clause, then only that one index is
3310 ** considered. */
3311 if( pSrc->pIndex ) break;
3313 /* Reset masks for the next index in the loop */
3314 wsFlagMask = ~(WHERE_ROWID_EQ|WHERE_ROWID_RANGE);
3315 eqTermMask = idxEqTermMask;
3318 /* If there is no ORDER BY clause and the SQLITE_ReverseOrder flag
3319 ** is set, then reverse the order that the index will be scanned
3320 ** in. This is used for application testing, to help find cases
3321 ** where application behaviour depends on the (undefined) order that
3322 ** SQLite outputs rows in in the absence of an ORDER BY clause. */
3323 if( !pOrderBy && pParse->db->flags & SQLITE_ReverseOrder ){
3324 pCost->plan.wsFlags |= WHERE_REVERSE;
3327 assert( pOrderBy || (pCost->plan.wsFlags&WHERE_ORDERBY)==0 );
3328 assert( pCost->plan.u.pIdx==0 || (pCost->plan.wsFlags&WHERE_ROWID_EQ)==0 );
3329 assert( pSrc->pIndex==0
3330 || pCost->plan.u.pIdx==0
3331 || pCost->plan.u.pIdx==pSrc->pIndex
3334 WHERETRACE(("best index is: %s\n",
3335 ((pCost->plan.wsFlags & WHERE_NOT_FULLSCAN)==0 ? "none" :
3336 pCost->plan.u.pIdx ? pCost->plan.u.pIdx->zName : "ipk")
3339 bestOrClauseIndex(pParse, pWC, pSrc, notReady, notValid, pOrderBy, pCost);
3340 bestAutomaticIndex(pParse, pWC, pSrc, notReady, pCost);
3341 pCost->plan.wsFlags |= eqTermMask;
3345 ** Find the query plan for accessing table pSrc->pTab. Write the
3346 ** best query plan and its cost into the WhereCost object supplied
3347 ** as the last parameter. This function may calculate the cost of
3348 ** both real and virtual table scans.
3350 static void bestIndex(
3351 Parse *pParse, /* The parsing context */
3352 WhereClause *pWC, /* The WHERE clause */
3353 struct SrcList_item *pSrc, /* The FROM clause term to search */
3354 Bitmask notReady, /* Mask of cursors not available for indexing */
3355 Bitmask notValid, /* Cursors not available for any purpose */
3356 ExprList *pOrderBy, /* The ORDER BY clause */
3357 WhereCost *pCost /* Lowest cost query plan */
3359 #ifndef SQLITE_OMIT_VIRTUALTABLE
3360 if( IsVirtual(pSrc->pTab) ){
3361 sqlite3_index_info *p = 0;
3362 bestVirtualIndex(pParse, pWC, pSrc, notReady, notValid, pOrderBy, pCost,&p);
3363 if( p->needToFreeIdxStr ){
3364 sqlite3_free(p->idxStr);
3366 sqlite3DbFree(pParse->db, p);
3367 }else
3368 #endif
3370 bestBtreeIndex(pParse, pWC, pSrc, notReady, notValid, pOrderBy, 0, pCost);
3375 ** Disable a term in the WHERE clause. Except, do not disable the term
3376 ** if it controls a LEFT OUTER JOIN and it did not originate in the ON
3377 ** or USING clause of that join.
3379 ** Consider the term t2.z='ok' in the following queries:
3381 ** (1) SELECT * FROM t1 LEFT JOIN t2 ON t1.a=t2.x WHERE t2.z='ok'
3382 ** (2) SELECT * FROM t1 LEFT JOIN t2 ON t1.a=t2.x AND t2.z='ok'
3383 ** (3) SELECT * FROM t1, t2 WHERE t1.a=t2.x AND t2.z='ok'
3385 ** The t2.z='ok' is disabled in the in (2) because it originates
3386 ** in the ON clause. The term is disabled in (3) because it is not part
3387 ** of a LEFT OUTER JOIN. In (1), the term is not disabled.
3389 ** IMPLEMENTATION-OF: R-24597-58655 No tests are done for terms that are
3390 ** completely satisfied by indices.
3392 ** Disabling a term causes that term to not be tested in the inner loop
3393 ** of the join. Disabling is an optimization. When terms are satisfied
3394 ** by indices, we disable them to prevent redundant tests in the inner
3395 ** loop. We would get the correct results if nothing were ever disabled,
3396 ** but joins might run a little slower. The trick is to disable as much
3397 ** as we can without disabling too much. If we disabled in (1), we'd get
3398 ** the wrong answer. See ticket #813.
3400 static void disableTerm(WhereLevel *pLevel, WhereTerm *pTerm){
3401 if( pTerm
3402 && (pTerm->wtFlags & TERM_CODED)==0
3403 && (pLevel->iLeftJoin==0 || ExprHasProperty(pTerm->pExpr, EP_FromJoin))
3405 pTerm->wtFlags |= TERM_CODED;
3406 if( pTerm->iParent>=0 ){
3407 WhereTerm *pOther = &pTerm->pWC->a[pTerm->iParent];
3408 if( (--pOther->nChild)==0 ){
3409 disableTerm(pLevel, pOther);
3416 ** Code an OP_Affinity opcode to apply the column affinity string zAff
3417 ** to the n registers starting at base.
3419 ** As an optimization, SQLITE_AFF_NONE entries (which are no-ops) at the
3420 ** beginning and end of zAff are ignored. If all entries in zAff are
3421 ** SQLITE_AFF_NONE, then no code gets generated.
3423 ** This routine makes its own copy of zAff so that the caller is free
3424 ** to modify zAff after this routine returns.
3426 static void codeApplyAffinity(Parse *pParse, int base, int n, char *zAff){
3427 Vdbe *v = pParse->pVdbe;
3428 if( zAff==0 ){
3429 assert( pParse->db->mallocFailed );
3430 return;
3432 assert( v!=0 );
3434 /* Adjust base and n to skip over SQLITE_AFF_NONE entries at the beginning
3435 ** and end of the affinity string.
3437 while( n>0 && zAff[0]==SQLITE_AFF_NONE ){
3438 n--;
3439 base++;
3440 zAff++;
3442 while( n>1 && zAff[n-1]==SQLITE_AFF_NONE ){
3443 n--;
3446 /* Code the OP_Affinity opcode if there is anything left to do. */
3447 if( n>0 ){
3448 sqlite3VdbeAddOp2(v, OP_Affinity, base, n);
3449 sqlite3VdbeChangeP4(v, -1, zAff, n);
3450 sqlite3ExprCacheAffinityChange(pParse, base, n);
3456 ** Generate code for a single equality term of the WHERE clause. An equality
3457 ** term can be either X=expr or X IN (...). pTerm is the term to be
3458 ** coded.
3460 ** The current value for the constraint is left in register iReg.
3462 ** For a constraint of the form X=expr, the expression is evaluated and its
3463 ** result is left on the stack. For constraints of the form X IN (...)
3464 ** this routine sets up a loop that will iterate over all values of X.
3466 static int codeEqualityTerm(
3467 Parse *pParse, /* The parsing context */
3468 WhereTerm *pTerm, /* The term of the WHERE clause to be coded */
3469 WhereLevel *pLevel, /* When level of the FROM clause we are working on */
3470 int iTarget /* Attempt to leave results in this register */
3472 Expr *pX = pTerm->pExpr;
3473 Vdbe *v = pParse->pVdbe;
3474 int iReg; /* Register holding results */
3476 assert( iTarget>0 );
3477 if( pX->op==TK_EQ ){
3478 iReg = sqlite3ExprCodeTarget(pParse, pX->pRight, iTarget);
3479 }else if( pX->op==TK_ISNULL ){
3480 iReg = iTarget;
3481 sqlite3VdbeAddOp2(v, OP_Null, 0, iReg);
3482 #ifndef SQLITE_OMIT_SUBQUERY
3483 }else{
3484 int eType;
3485 int iTab;
3486 struct InLoop *pIn;
3488 assert( pX->op==TK_IN );
3489 iReg = iTarget;
3490 eType = sqlite3FindInIndex(pParse, pX, 0);
3491 iTab = pX->iTable;
3492 sqlite3VdbeAddOp2(v, OP_Rewind, iTab, 0);
3493 assert( pLevel->plan.wsFlags & WHERE_IN_ABLE );
3494 if( pLevel->u.in.nIn==0 ){
3495 pLevel->addrNxt = sqlite3VdbeMakeLabel(v);
3497 pLevel->u.in.nIn++;
3498 pLevel->u.in.aInLoop =
3499 sqlite3DbReallocOrFree(pParse->db, pLevel->u.in.aInLoop,
3500 sizeof(pLevel->u.in.aInLoop[0])*pLevel->u.in.nIn);
3501 pIn = pLevel->u.in.aInLoop;
3502 if( pIn ){
3503 pIn += pLevel->u.in.nIn - 1;
3504 pIn->iCur = iTab;
3505 if( eType==IN_INDEX_ROWID ){
3506 pIn->addrInTop = sqlite3VdbeAddOp2(v, OP_Rowid, iTab, iReg);
3507 }else{
3508 pIn->addrInTop = sqlite3VdbeAddOp3(v, OP_Column, iTab, 0, iReg);
3510 sqlite3VdbeAddOp1(v, OP_IsNull, iReg);
3511 }else{
3512 pLevel->u.in.nIn = 0;
3514 #endif
3516 disableTerm(pLevel, pTerm);
3517 return iReg;
3521 ** Generate code that will evaluate all == and IN constraints for an
3522 ** index.
3524 ** For example, consider table t1(a,b,c,d,e,f) with index i1(a,b,c).
3525 ** Suppose the WHERE clause is this: a==5 AND b IN (1,2,3) AND c>5 AND c<10
3526 ** The index has as many as three equality constraints, but in this
3527 ** example, the third "c" value is an inequality. So only two
3528 ** constraints are coded. This routine will generate code to evaluate
3529 ** a==5 and b IN (1,2,3). The current values for a and b will be stored
3530 ** in consecutive registers and the index of the first register is returned.
3532 ** In the example above nEq==2. But this subroutine works for any value
3533 ** of nEq including 0. If nEq==0, this routine is nearly a no-op.
3534 ** The only thing it does is allocate the pLevel->iMem memory cell and
3535 ** compute the affinity string.
3537 ** This routine always allocates at least one memory cell and returns
3538 ** the index of that memory cell. The code that
3539 ** calls this routine will use that memory cell to store the termination
3540 ** key value of the loop. If one or more IN operators appear, then
3541 ** this routine allocates an additional nEq memory cells for internal
3542 ** use.
3544 ** Before returning, *pzAff is set to point to a buffer containing a
3545 ** copy of the column affinity string of the index allocated using
3546 ** sqlite3DbMalloc(). Except, entries in the copy of the string associated
3547 ** with equality constraints that use NONE affinity are set to
3548 ** SQLITE_AFF_NONE. This is to deal with SQL such as the following:
3550 ** CREATE TABLE t1(a TEXT PRIMARY KEY, b);
3551 ** SELECT ... FROM t1 AS t2, t1 WHERE t1.a = t2.b;
3553 ** In the example above, the index on t1(a) has TEXT affinity. But since
3554 ** the right hand side of the equality constraint (t2.b) has NONE affinity,
3555 ** no conversion should be attempted before using a t2.b value as part of
3556 ** a key to search the index. Hence the first byte in the returned affinity
3557 ** string in this example would be set to SQLITE_AFF_NONE.
3559 static int codeAllEqualityTerms(
3560 Parse *pParse, /* Parsing context */
3561 WhereLevel *pLevel, /* Which nested loop of the FROM we are coding */
3562 WhereClause *pWC, /* The WHERE clause */
3563 Bitmask notReady, /* Which parts of FROM have not yet been coded */
3564 int nExtraReg, /* Number of extra registers to allocate */
3565 char **pzAff /* OUT: Set to point to affinity string */
3567 int nEq = pLevel->plan.nEq; /* The number of == or IN constraints to code */
3568 Vdbe *v = pParse->pVdbe; /* The vm under construction */
3569 Index *pIdx; /* The index being used for this loop */
3570 int iCur = pLevel->iTabCur; /* The cursor of the table */
3571 WhereTerm *pTerm; /* A single constraint term */
3572 int j; /* Loop counter */
3573 int regBase; /* Base register */
3574 int nReg; /* Number of registers to allocate */
3575 char *zAff; /* Affinity string to return */
3577 /* This module is only called on query plans that use an index. */
3578 assert( pLevel->plan.wsFlags & WHERE_INDEXED );
3579 pIdx = pLevel->plan.u.pIdx;
3581 /* Figure out how many memory cells we will need then allocate them.
3583 regBase = pParse->nMem + 1;
3584 nReg = pLevel->plan.nEq + nExtraReg;
3585 pParse->nMem += nReg;
3587 zAff = sqlite3DbStrDup(pParse->db, sqlite3IndexAffinityStr(v, pIdx));
3588 if( !zAff ){
3589 pParse->db->mallocFailed = 1;
3592 /* Evaluate the equality constraints
3594 assert( pIdx->nColumn>=nEq );
3595 for(j=0; j<nEq; j++){
3596 int r1;
3597 int k = pIdx->aiColumn[j];
3598 pTerm = findTerm(pWC, iCur, k, notReady, pLevel->plan.wsFlags, pIdx);
3599 if( NEVER(pTerm==0) ) break;
3600 /* The following true for indices with redundant columns.
3601 ** Ex: CREATE INDEX i1 ON t1(a,b,a); SELECT * FROM t1 WHERE a=0 AND b=0; */
3602 testcase( (pTerm->wtFlags & TERM_CODED)!=0 );
3603 testcase( pTerm->wtFlags & TERM_VIRTUAL ); /* EV: R-30575-11662 */
3604 r1 = codeEqualityTerm(pParse, pTerm, pLevel, regBase+j);
3605 if( r1!=regBase+j ){
3606 if( nReg==1 ){
3607 sqlite3ReleaseTempReg(pParse, regBase);
3608 regBase = r1;
3609 }else{
3610 sqlite3VdbeAddOp2(v, OP_SCopy, r1, regBase+j);
3613 testcase( pTerm->eOperator & WO_ISNULL );
3614 testcase( pTerm->eOperator & WO_IN );
3615 if( (pTerm->eOperator & (WO_ISNULL|WO_IN))==0 ){
3616 Expr *pRight = pTerm->pExpr->pRight;
3617 sqlite3ExprCodeIsNullJump(v, pRight, regBase+j, pLevel->addrBrk);
3618 if( zAff ){
3619 if( sqlite3CompareAffinity(pRight, zAff[j])==SQLITE_AFF_NONE ){
3620 zAff[j] = SQLITE_AFF_NONE;
3622 if( sqlite3ExprNeedsNoAffinityChange(pRight, zAff[j]) ){
3623 zAff[j] = SQLITE_AFF_NONE;
3628 *pzAff = zAff;
3629 return regBase;
3632 #ifndef SQLITE_OMIT_EXPLAIN
3634 ** This routine is a helper for explainIndexRange() below
3636 ** pStr holds the text of an expression that we are building up one term
3637 ** at a time. This routine adds a new term to the end of the expression.
3638 ** Terms are separated by AND so add the "AND" text for second and subsequent
3639 ** terms only.
3641 static void explainAppendTerm(
3642 StrAccum *pStr, /* The text expression being built */
3643 int iTerm, /* Index of this term. First is zero */
3644 const char *zColumn, /* Name of the column */
3645 const char *zOp /* Name of the operator */
3647 if( iTerm ) sqlite3StrAccumAppend(pStr, " AND ", 5);
3648 sqlite3StrAccumAppend(pStr, zColumn, -1);
3649 sqlite3StrAccumAppend(pStr, zOp, 1);
3650 sqlite3StrAccumAppend(pStr, "?", 1);
3654 ** Argument pLevel describes a strategy for scanning table pTab. This
3655 ** function returns a pointer to a string buffer containing a description
3656 ** of the subset of table rows scanned by the strategy in the form of an
3657 ** SQL expression. Or, if all rows are scanned, NULL is returned.
3659 ** For example, if the query:
3661 ** SELECT * FROM t1 WHERE a=1 AND b>2;
3663 ** is run and there is an index on (a, b), then this function returns a
3664 ** string similar to:
3666 ** "a=? AND b>?"
3668 ** The returned pointer points to memory obtained from sqlite3DbMalloc().
3669 ** It is the responsibility of the caller to free the buffer when it is
3670 ** no longer required.
3672 static char *explainIndexRange(sqlite3 *db, WhereLevel *pLevel, Table *pTab){
3673 WherePlan *pPlan = &pLevel->plan;
3674 Index *pIndex = pPlan->u.pIdx;
3675 int nEq = pPlan->nEq;
3676 int i, j;
3677 Column *aCol = pTab->aCol;
3678 int *aiColumn = pIndex->aiColumn;
3679 StrAccum txt;
3681 if( nEq==0 && (pPlan->wsFlags & (WHERE_BTM_LIMIT|WHERE_TOP_LIMIT))==0 ){
3682 return 0;
3684 sqlite3StrAccumInit(&txt, 0, 0, SQLITE_MAX_LENGTH);
3685 txt.db = db;
3686 sqlite3StrAccumAppend(&txt, " (", 2);
3687 for(i=0; i<nEq; i++){
3688 explainAppendTerm(&txt, i, aCol[aiColumn[i]].zName, "=");
3691 j = i;
3692 if( pPlan->wsFlags&WHERE_BTM_LIMIT ){
3693 explainAppendTerm(&txt, i++, aCol[aiColumn[j]].zName, ">");
3695 if( pPlan->wsFlags&WHERE_TOP_LIMIT ){
3696 explainAppendTerm(&txt, i, aCol[aiColumn[j]].zName, "<");
3698 sqlite3StrAccumAppend(&txt, ")", 1);
3699 return sqlite3StrAccumFinish(&txt);
3703 ** This function is a no-op unless currently processing an EXPLAIN QUERY PLAN
3704 ** command. If the query being compiled is an EXPLAIN QUERY PLAN, a single
3705 ** record is added to the output to describe the table scan strategy in
3706 ** pLevel.
3708 static void explainOneScan(
3709 Parse *pParse, /* Parse context */
3710 SrcList *pTabList, /* Table list this loop refers to */
3711 WhereLevel *pLevel, /* Scan to write OP_Explain opcode for */
3712 int iLevel, /* Value for "level" column of output */
3713 int iFrom, /* Value for "from" column of output */
3714 u16 wctrlFlags /* Flags passed to sqlite3WhereBegin() */
3716 if( pParse->explain==2 ){
3717 u32 flags = pLevel->plan.wsFlags;
3718 struct SrcList_item *pItem = &pTabList->a[pLevel->iFrom];
3719 Vdbe *v = pParse->pVdbe; /* VM being constructed */
3720 sqlite3 *db = pParse->db; /* Database handle */
3721 char *zMsg; /* Text to add to EQP output */
3722 sqlite3_int64 nRow; /* Expected number of rows visited by scan */
3723 int iId = pParse->iSelectId; /* Select id (left-most output column) */
3724 int isSearch; /* True for a SEARCH. False for SCAN. */
3726 if( (flags&WHERE_MULTI_OR) || (wctrlFlags&WHERE_ONETABLE_ONLY) ) return;
3728 isSearch = (pLevel->plan.nEq>0)
3729 || (flags&(WHERE_BTM_LIMIT|WHERE_TOP_LIMIT))!=0
3730 || (wctrlFlags&(WHERE_ORDERBY_MIN|WHERE_ORDERBY_MAX));
3732 zMsg = sqlite3MPrintf(db, "%s", isSearch?"SEARCH":"SCAN");
3733 if( pItem->pSelect ){
3734 zMsg = sqlite3MAppendf(db, zMsg, "%s SUBQUERY %d", zMsg,pItem->iSelectId);
3735 }else{
3736 zMsg = sqlite3MAppendf(db, zMsg, "%s TABLE %s", zMsg, pItem->zName);
3739 if( pItem->zAlias ){
3740 zMsg = sqlite3MAppendf(db, zMsg, "%s AS %s", zMsg, pItem->zAlias);
3742 if( (flags & WHERE_INDEXED)!=0 ){
3743 char *zWhere = explainIndexRange(db, pLevel, pItem->pTab);
3744 zMsg = sqlite3MAppendf(db, zMsg, "%s USING %s%sINDEX%s%s%s", zMsg,
3745 ((flags & WHERE_TEMP_INDEX)?"AUTOMATIC ":""),
3746 ((flags & WHERE_IDX_ONLY)?"COVERING ":""),
3747 ((flags & WHERE_TEMP_INDEX)?"":" "),
3748 ((flags & WHERE_TEMP_INDEX)?"": pLevel->plan.u.pIdx->zName),
3749 zWhere
3751 sqlite3DbFree(db, zWhere);
3752 }else if( flags & (WHERE_ROWID_EQ|WHERE_ROWID_RANGE) ){
3753 zMsg = sqlite3MAppendf(db, zMsg, "%s USING INTEGER PRIMARY KEY", zMsg);
3755 if( flags&WHERE_ROWID_EQ ){
3756 zMsg = sqlite3MAppendf(db, zMsg, "%s (rowid=?)", zMsg);
3757 }else if( (flags&WHERE_BOTH_LIMIT)==WHERE_BOTH_LIMIT ){
3758 zMsg = sqlite3MAppendf(db, zMsg, "%s (rowid>? AND rowid<?)", zMsg);
3759 }else if( flags&WHERE_BTM_LIMIT ){
3760 zMsg = sqlite3MAppendf(db, zMsg, "%s (rowid>?)", zMsg);
3761 }else if( flags&WHERE_TOP_LIMIT ){
3762 zMsg = sqlite3MAppendf(db, zMsg, "%s (rowid<?)", zMsg);
3765 #ifndef SQLITE_OMIT_VIRTUALTABLE
3766 else if( (flags & WHERE_VIRTUALTABLE)!=0 ){
3767 sqlite3_index_info *pVtabIdx = pLevel->plan.u.pVtabIdx;
3768 zMsg = sqlite3MAppendf(db, zMsg, "%s VIRTUAL TABLE INDEX %d:%s", zMsg,
3769 pVtabIdx->idxNum, pVtabIdx->idxStr);
3771 #endif
3772 if( wctrlFlags&(WHERE_ORDERBY_MIN|WHERE_ORDERBY_MAX) ){
3773 testcase( wctrlFlags & WHERE_ORDERBY_MIN );
3774 nRow = 1;
3775 }else{
3776 nRow = (sqlite3_int64)pLevel->plan.nRow;
3778 zMsg = sqlite3MAppendf(db, zMsg, "%s (~%lld rows)", zMsg, nRow);
3779 sqlite3VdbeAddOp4(v, OP_Explain, iId, iLevel, iFrom, zMsg, P4_DYNAMIC);
3782 #else
3783 # define explainOneScan(u,v,w,x,y,z)
3784 #endif /* SQLITE_OMIT_EXPLAIN */
3788 ** Generate code for the start of the iLevel-th loop in the WHERE clause
3789 ** implementation described by pWInfo.
3791 static Bitmask codeOneLoopStart(
3792 WhereInfo *pWInfo, /* Complete information about the WHERE clause */
3793 int iLevel, /* Which level of pWInfo->a[] should be coded */
3794 u16 wctrlFlags, /* One of the WHERE_* flags defined in sqliteInt.h */
3795 Bitmask notReady, /* Which tables are currently available */
3796 Expr *pWhere /* Complete WHERE clause */
3798 int j, k; /* Loop counters */
3799 int iCur; /* The VDBE cursor for the table */
3800 int addrNxt; /* Where to jump to continue with the next IN case */
3801 int omitTable; /* True if we use the index only */
3802 int bRev; /* True if we need to scan in reverse order */
3803 WhereLevel *pLevel; /* The where level to be coded */
3804 WhereClause *pWC; /* Decomposition of the entire WHERE clause */
3805 WhereTerm *pTerm; /* A WHERE clause term */
3806 Parse *pParse; /* Parsing context */
3807 Vdbe *v; /* The prepared stmt under constructions */
3808 struct SrcList_item *pTabItem; /* FROM clause term being coded */
3809 int addrBrk; /* Jump here to break out of the loop */
3810 int addrCont; /* Jump here to continue with next cycle */
3811 int iRowidReg = 0; /* Rowid is stored in this register, if not zero */
3812 int iReleaseReg = 0; /* Temp register to free before returning */
3814 pParse = pWInfo->pParse;
3815 v = pParse->pVdbe;
3816 pWC = pWInfo->pWC;
3817 pLevel = &pWInfo->a[iLevel];
3818 pTabItem = &pWInfo->pTabList->a[pLevel->iFrom];
3819 iCur = pTabItem->iCursor;
3820 bRev = (pLevel->plan.wsFlags & WHERE_REVERSE)!=0;
3821 omitTable = (pLevel->plan.wsFlags & WHERE_IDX_ONLY)!=0
3822 && (wctrlFlags & WHERE_FORCE_TABLE)==0;
3824 /* Create labels for the "break" and "continue" instructions
3825 ** for the current loop. Jump to addrBrk to break out of a loop.
3826 ** Jump to cont to go immediately to the next iteration of the
3827 ** loop.
3829 ** When there is an IN operator, we also have a "addrNxt" label that
3830 ** means to continue with the next IN value combination. When
3831 ** there are no IN operators in the constraints, the "addrNxt" label
3832 ** is the same as "addrBrk".
3834 addrBrk = pLevel->addrBrk = pLevel->addrNxt = sqlite3VdbeMakeLabel(v);
3835 addrCont = pLevel->addrCont = sqlite3VdbeMakeLabel(v);
3837 /* If this is the right table of a LEFT OUTER JOIN, allocate and
3838 ** initialize a memory cell that records if this table matches any
3839 ** row of the left table of the join.
3841 if( pLevel->iFrom>0 && (pTabItem[0].jointype & JT_LEFT)!=0 ){
3842 pLevel->iLeftJoin = ++pParse->nMem;
3843 sqlite3VdbeAddOp2(v, OP_Integer, 0, pLevel->iLeftJoin);
3844 VdbeComment((v, "init LEFT JOIN no-match flag"));
3847 #ifndef SQLITE_OMIT_VIRTUALTABLE
3848 if( (pLevel->plan.wsFlags & WHERE_VIRTUALTABLE)!=0 ){
3849 /* Case 0: The table is a virtual-table. Use the VFilter and VNext
3850 ** to access the data.
3852 int iReg; /* P3 Value for OP_VFilter */
3853 sqlite3_index_info *pVtabIdx = pLevel->plan.u.pVtabIdx;
3854 int nConstraint = pVtabIdx->nConstraint;
3855 struct sqlite3_index_constraint_usage *aUsage =
3856 pVtabIdx->aConstraintUsage;
3857 const struct sqlite3_index_constraint *aConstraint =
3858 pVtabIdx->aConstraint;
3860 sqlite3ExprCachePush(pParse);
3861 iReg = sqlite3GetTempRange(pParse, nConstraint+2);
3862 for(j=1; j<=nConstraint; j++){
3863 for(k=0; k<nConstraint; k++){
3864 if( aUsage[k].argvIndex==j ){
3865 int iTerm = aConstraint[k].iTermOffset;
3866 sqlite3ExprCode(pParse, pWC->a[iTerm].pExpr->pRight, iReg+j+1);
3867 break;
3870 if( k==nConstraint ) break;
3872 sqlite3VdbeAddOp2(v, OP_Integer, pVtabIdx->idxNum, iReg);
3873 sqlite3VdbeAddOp2(v, OP_Integer, j-1, iReg+1);
3874 sqlite3VdbeAddOp4(v, OP_VFilter, iCur, addrBrk, iReg, pVtabIdx->idxStr,
3875 pVtabIdx->needToFreeIdxStr ? P4_MPRINTF : P4_STATIC);
3876 pVtabIdx->needToFreeIdxStr = 0;
3877 for(j=0; j<nConstraint; j++){
3878 if( aUsage[j].omit ){
3879 int iTerm = aConstraint[j].iTermOffset;
3880 disableTerm(pLevel, &pWC->a[iTerm]);
3883 pLevel->op = OP_VNext;
3884 pLevel->p1 = iCur;
3885 pLevel->p2 = sqlite3VdbeCurrentAddr(v);
3886 sqlite3ReleaseTempRange(pParse, iReg, nConstraint+2);
3887 sqlite3ExprCachePop(pParse, 1);
3888 }else
3889 #endif /* SQLITE_OMIT_VIRTUALTABLE */
3891 if( pLevel->plan.wsFlags & WHERE_ROWID_EQ ){
3892 /* Case 1: We can directly reference a single row using an
3893 ** equality comparison against the ROWID field. Or
3894 ** we reference multiple rows using a "rowid IN (...)"
3895 ** construct.
3897 iReleaseReg = sqlite3GetTempReg(pParse);
3898 pTerm = findTerm(pWC, iCur, -1, notReady, WO_EQ|WO_IN, 0);
3899 assert( pTerm!=0 );
3900 assert( pTerm->pExpr!=0 );
3901 assert( pTerm->leftCursor==iCur );
3902 assert( omitTable==0 );
3903 testcase( pTerm->wtFlags & TERM_VIRTUAL ); /* EV: R-30575-11662 */
3904 iRowidReg = codeEqualityTerm(pParse, pTerm, pLevel, iReleaseReg);
3905 addrNxt = pLevel->addrNxt;
3906 sqlite3VdbeAddOp2(v, OP_MustBeInt, iRowidReg, addrNxt);
3907 sqlite3VdbeAddOp3(v, OP_NotExists, iCur, addrNxt, iRowidReg);
3908 sqlite3ExprCacheStore(pParse, iCur, -1, iRowidReg);
3909 VdbeComment((v, "pk"));
3910 pLevel->op = OP_Noop;
3911 }else if( pLevel->plan.wsFlags & WHERE_ROWID_RANGE ){
3912 /* Case 2: We have an inequality comparison against the ROWID field.
3914 int testOp = OP_Noop;
3915 int start;
3916 int memEndValue = 0;
3917 WhereTerm *pStart, *pEnd;
3919 assert( omitTable==0 );
3920 pStart = findTerm(pWC, iCur, -1, notReady, WO_GT|WO_GE, 0);
3921 pEnd = findTerm(pWC, iCur, -1, notReady, WO_LT|WO_LE, 0);
3922 if( bRev ){
3923 pTerm = pStart;
3924 pStart = pEnd;
3925 pEnd = pTerm;
3927 if( pStart ){
3928 Expr *pX; /* The expression that defines the start bound */
3929 int r1, rTemp; /* Registers for holding the start boundary */
3931 /* The following constant maps TK_xx codes into corresponding
3932 ** seek opcodes. It depends on a particular ordering of TK_xx
3934 const u8 aMoveOp[] = {
3935 /* TK_GT */ OP_SeekGt,
3936 /* TK_LE */ OP_SeekLe,
3937 /* TK_LT */ OP_SeekLt,
3938 /* TK_GE */ OP_SeekGe
3940 assert( TK_LE==TK_GT+1 ); /* Make sure the ordering.. */
3941 assert( TK_LT==TK_GT+2 ); /* ... of the TK_xx values... */
3942 assert( TK_GE==TK_GT+3 ); /* ... is correcct. */
3944 testcase( pStart->wtFlags & TERM_VIRTUAL ); /* EV: R-30575-11662 */
3945 pX = pStart->pExpr;
3946 assert( pX!=0 );
3947 assert( pStart->leftCursor==iCur );
3948 r1 = sqlite3ExprCodeTemp(pParse, pX->pRight, &rTemp);
3949 sqlite3VdbeAddOp3(v, aMoveOp[pX->op-TK_GT], iCur, addrBrk, r1);
3950 VdbeComment((v, "pk"));
3951 sqlite3ExprCacheAffinityChange(pParse, r1, 1);
3952 sqlite3ReleaseTempReg(pParse, rTemp);
3953 disableTerm(pLevel, pStart);
3954 }else{
3955 sqlite3VdbeAddOp2(v, bRev ? OP_Last : OP_Rewind, iCur, addrBrk);
3957 if( pEnd ){
3958 Expr *pX;
3959 pX = pEnd->pExpr;
3960 assert( pX!=0 );
3961 assert( pEnd->leftCursor==iCur );
3962 testcase( pEnd->wtFlags & TERM_VIRTUAL ); /* EV: R-30575-11662 */
3963 memEndValue = ++pParse->nMem;
3964 sqlite3ExprCode(pParse, pX->pRight, memEndValue);
3965 if( pX->op==TK_LT || pX->op==TK_GT ){
3966 testOp = bRev ? OP_Le : OP_Ge;
3967 }else{
3968 testOp = bRev ? OP_Lt : OP_Gt;
3970 disableTerm(pLevel, pEnd);
3972 start = sqlite3VdbeCurrentAddr(v);
3973 pLevel->op = bRev ? OP_Prev : OP_Next;
3974 pLevel->p1 = iCur;
3975 pLevel->p2 = start;
3976 if( pStart==0 && pEnd==0 ){
3977 pLevel->p5 = SQLITE_STMTSTATUS_FULLSCAN_STEP;
3978 }else{
3979 assert( pLevel->p5==0 );
3981 if( testOp!=OP_Noop ){
3982 iRowidReg = iReleaseReg = sqlite3GetTempReg(pParse);
3983 sqlite3VdbeAddOp2(v, OP_Rowid, iCur, iRowidReg);
3984 sqlite3ExprCacheStore(pParse, iCur, -1, iRowidReg);
3985 sqlite3VdbeAddOp3(v, testOp, memEndValue, addrBrk, iRowidReg);
3986 sqlite3VdbeChangeP5(v, SQLITE_AFF_NUMERIC | SQLITE_JUMPIFNULL);
3988 }else if( pLevel->plan.wsFlags & (WHERE_COLUMN_RANGE|WHERE_COLUMN_EQ) ){
3989 /* Case 3: A scan using an index.
3991 ** The WHERE clause may contain zero or more equality
3992 ** terms ("==" or "IN" operators) that refer to the N
3993 ** left-most columns of the index. It may also contain
3994 ** inequality constraints (>, <, >= or <=) on the indexed
3995 ** column that immediately follows the N equalities. Only
3996 ** the right-most column can be an inequality - the rest must
3997 ** use the "==" and "IN" operators. For example, if the
3998 ** index is on (x,y,z), then the following clauses are all
3999 ** optimized:
4001 ** x=5
4002 ** x=5 AND y=10
4003 ** x=5 AND y<10
4004 ** x=5 AND y>5 AND y<10
4005 ** x=5 AND y=5 AND z<=10
4007 ** The z<10 term of the following cannot be used, only
4008 ** the x=5 term:
4010 ** x=5 AND z<10
4012 ** N may be zero if there are inequality constraints.
4013 ** If there are no inequality constraints, then N is at
4014 ** least one.
4016 ** This case is also used when there are no WHERE clause
4017 ** constraints but an index is selected anyway, in order
4018 ** to force the output order to conform to an ORDER BY.
4020 static const u8 aStartOp[] = {
4023 OP_Rewind, /* 2: (!start_constraints && startEq && !bRev) */
4024 OP_Last, /* 3: (!start_constraints && startEq && bRev) */
4025 OP_SeekGt, /* 4: (start_constraints && !startEq && !bRev) */
4026 OP_SeekLt, /* 5: (start_constraints && !startEq && bRev) */
4027 OP_SeekGe, /* 6: (start_constraints && startEq && !bRev) */
4028 OP_SeekLe /* 7: (start_constraints && startEq && bRev) */
4030 static const u8 aEndOp[] = {
4031 OP_Noop, /* 0: (!end_constraints) */
4032 OP_IdxGE, /* 1: (end_constraints && !bRev) */
4033 OP_IdxLT /* 2: (end_constraints && bRev) */
4035 int nEq = pLevel->plan.nEq; /* Number of == or IN terms */
4036 int isMinQuery = 0; /* If this is an optimized SELECT min(x).. */
4037 int regBase; /* Base register holding constraint values */
4038 int r1; /* Temp register */
4039 WhereTerm *pRangeStart = 0; /* Inequality constraint at range start */
4040 WhereTerm *pRangeEnd = 0; /* Inequality constraint at range end */
4041 int startEq; /* True if range start uses ==, >= or <= */
4042 int endEq; /* True if range end uses ==, >= or <= */
4043 int start_constraints; /* Start of range is constrained */
4044 int nConstraint; /* Number of constraint terms */
4045 Index *pIdx; /* The index we will be using */
4046 int iIdxCur; /* The VDBE cursor for the index */
4047 int nExtraReg = 0; /* Number of extra registers needed */
4048 int op; /* Instruction opcode */
4049 char *zStartAff; /* Affinity for start of range constraint */
4050 char *zEndAff; /* Affinity for end of range constraint */
4052 pIdx = pLevel->plan.u.pIdx;
4053 iIdxCur = pLevel->iIdxCur;
4054 k = pIdx->aiColumn[nEq]; /* Column for inequality constraints */
4056 /* If this loop satisfies a sort order (pOrderBy) request that
4057 ** was passed to this function to implement a "SELECT min(x) ..."
4058 ** query, then the caller will only allow the loop to run for
4059 ** a single iteration. This means that the first row returned
4060 ** should not have a NULL value stored in 'x'. If column 'x' is
4061 ** the first one after the nEq equality constraints in the index,
4062 ** this requires some special handling.
4064 if( (wctrlFlags&WHERE_ORDERBY_MIN)!=0
4065 && (pLevel->plan.wsFlags&WHERE_ORDERBY)
4066 && (pIdx->nColumn>nEq)
4068 /* assert( pOrderBy->nExpr==1 ); */
4069 /* assert( pOrderBy->a[0].pExpr->iColumn==pIdx->aiColumn[nEq] ); */
4070 isMinQuery = 1;
4071 nExtraReg = 1;
4074 /* Find any inequality constraint terms for the start and end
4075 ** of the range.
4077 if( pLevel->plan.wsFlags & WHERE_TOP_LIMIT ){
4078 pRangeEnd = findTerm(pWC, iCur, k, notReady, (WO_LT|WO_LE), pIdx);
4079 nExtraReg = 1;
4081 if( pLevel->plan.wsFlags & WHERE_BTM_LIMIT ){
4082 pRangeStart = findTerm(pWC, iCur, k, notReady, (WO_GT|WO_GE), pIdx);
4083 nExtraReg = 1;
4086 /* Generate code to evaluate all constraint terms using == or IN
4087 ** and store the values of those terms in an array of registers
4088 ** starting at regBase.
4090 regBase = codeAllEqualityTerms(
4091 pParse, pLevel, pWC, notReady, nExtraReg, &zStartAff
4093 zEndAff = sqlite3DbStrDup(pParse->db, zStartAff);
4094 addrNxt = pLevel->addrNxt;
4096 /* If we are doing a reverse order scan on an ascending index, or
4097 ** a forward order scan on a descending index, interchange the
4098 ** start and end terms (pRangeStart and pRangeEnd).
4100 if( nEq<pIdx->nColumn && bRev==(pIdx->aSortOrder[nEq]==SQLITE_SO_ASC) ){
4101 SWAP(WhereTerm *, pRangeEnd, pRangeStart);
4104 testcase( pRangeStart && pRangeStart->eOperator & WO_LE );
4105 testcase( pRangeStart && pRangeStart->eOperator & WO_GE );
4106 testcase( pRangeEnd && pRangeEnd->eOperator & WO_LE );
4107 testcase( pRangeEnd && pRangeEnd->eOperator & WO_GE );
4108 startEq = !pRangeStart || pRangeStart->eOperator & (WO_LE|WO_GE);
4109 endEq = !pRangeEnd || pRangeEnd->eOperator & (WO_LE|WO_GE);
4110 start_constraints = pRangeStart || nEq>0;
4112 /* Seek the index cursor to the start of the range. */
4113 nConstraint = nEq;
4114 if( pRangeStart ){
4115 Expr *pRight = pRangeStart->pExpr->pRight;
4116 sqlite3ExprCode(pParse, pRight, regBase+nEq);
4117 if( (pRangeStart->wtFlags & TERM_VNULL)==0 ){
4118 sqlite3ExprCodeIsNullJump(v, pRight, regBase+nEq, addrNxt);
4120 if( zStartAff ){
4121 if( sqlite3CompareAffinity(pRight, zStartAff[nEq])==SQLITE_AFF_NONE){
4122 /* Since the comparison is to be performed with no conversions
4123 ** applied to the operands, set the affinity to apply to pRight to
4124 ** SQLITE_AFF_NONE. */
4125 zStartAff[nEq] = SQLITE_AFF_NONE;
4127 if( sqlite3ExprNeedsNoAffinityChange(pRight, zStartAff[nEq]) ){
4128 zStartAff[nEq] = SQLITE_AFF_NONE;
4131 nConstraint++;
4132 testcase( pRangeStart->wtFlags & TERM_VIRTUAL ); /* EV: R-30575-11662 */
4133 }else if( isMinQuery ){
4134 sqlite3VdbeAddOp2(v, OP_Null, 0, regBase+nEq);
4135 nConstraint++;
4136 startEq = 0;
4137 start_constraints = 1;
4139 codeApplyAffinity(pParse, regBase, nConstraint, zStartAff);
4140 op = aStartOp[(start_constraints<<2) + (startEq<<1) + bRev];
4141 assert( op!=0 );
4142 testcase( op==OP_Rewind );
4143 testcase( op==OP_Last );
4144 testcase( op==OP_SeekGt );
4145 testcase( op==OP_SeekGe );
4146 testcase( op==OP_SeekLe );
4147 testcase( op==OP_SeekLt );
4148 sqlite3VdbeAddOp4Int(v, op, iIdxCur, addrNxt, regBase, nConstraint);
4150 /* Load the value for the inequality constraint at the end of the
4151 ** range (if any).
4153 nConstraint = nEq;
4154 if( pRangeEnd ){
4155 Expr *pRight = pRangeEnd->pExpr->pRight;
4156 sqlite3ExprCacheRemove(pParse, regBase+nEq, 1);
4157 sqlite3ExprCode(pParse, pRight, regBase+nEq);
4158 if( (pRangeEnd->wtFlags & TERM_VNULL)==0 ){
4159 sqlite3ExprCodeIsNullJump(v, pRight, regBase+nEq, addrNxt);
4161 if( zEndAff ){
4162 if( sqlite3CompareAffinity(pRight, zEndAff[nEq])==SQLITE_AFF_NONE){
4163 /* Since the comparison is to be performed with no conversions
4164 ** applied to the operands, set the affinity to apply to pRight to
4165 ** SQLITE_AFF_NONE. */
4166 zEndAff[nEq] = SQLITE_AFF_NONE;
4168 if( sqlite3ExprNeedsNoAffinityChange(pRight, zEndAff[nEq]) ){
4169 zEndAff[nEq] = SQLITE_AFF_NONE;
4172 codeApplyAffinity(pParse, regBase, nEq+1, zEndAff);
4173 nConstraint++;
4174 testcase( pRangeEnd->wtFlags & TERM_VIRTUAL ); /* EV: R-30575-11662 */
4176 sqlite3DbFree(pParse->db, zStartAff);
4177 sqlite3DbFree(pParse->db, zEndAff);
4179 /* Top of the loop body */
4180 pLevel->p2 = sqlite3VdbeCurrentAddr(v);
4182 /* Check if the index cursor is past the end of the range. */
4183 op = aEndOp[(pRangeEnd || nEq) * (1 + bRev)];
4184 testcase( op==OP_Noop );
4185 testcase( op==OP_IdxGE );
4186 testcase( op==OP_IdxLT );
4187 if( op!=OP_Noop ){
4188 sqlite3VdbeAddOp4Int(v, op, iIdxCur, addrNxt, regBase, nConstraint);
4189 sqlite3VdbeChangeP5(v, endEq!=bRev ?1:0);
4192 /* If there are inequality constraints, check that the value
4193 ** of the table column that the inequality contrains is not NULL.
4194 ** If it is, jump to the next iteration of the loop.
4196 r1 = sqlite3GetTempReg(pParse);
4197 testcase( pLevel->plan.wsFlags & WHERE_BTM_LIMIT );
4198 testcase( pLevel->plan.wsFlags & WHERE_TOP_LIMIT );
4199 if( (pLevel->plan.wsFlags & (WHERE_BTM_LIMIT|WHERE_TOP_LIMIT))!=0 ){
4200 sqlite3VdbeAddOp3(v, OP_Column, iIdxCur, nEq, r1);
4201 sqlite3VdbeAddOp2(v, OP_IsNull, r1, addrCont);
4203 sqlite3ReleaseTempReg(pParse, r1);
4205 /* Seek the table cursor, if required */
4206 disableTerm(pLevel, pRangeStart);
4207 disableTerm(pLevel, pRangeEnd);
4208 if( !omitTable ){
4209 iRowidReg = iReleaseReg = sqlite3GetTempReg(pParse);
4210 sqlite3VdbeAddOp2(v, OP_IdxRowid, iIdxCur, iRowidReg);
4211 sqlite3ExprCacheStore(pParse, iCur, -1, iRowidReg);
4212 sqlite3VdbeAddOp2(v, OP_Seek, iCur, iRowidReg); /* Deferred seek */
4215 /* Record the instruction used to terminate the loop. Disable
4216 ** WHERE clause terms made redundant by the index range scan.
4218 if( pLevel->plan.wsFlags & WHERE_UNIQUE ){
4219 pLevel->op = OP_Noop;
4220 }else if( bRev ){
4221 pLevel->op = OP_Prev;
4222 }else{
4223 pLevel->op = OP_Next;
4225 pLevel->p1 = iIdxCur;
4226 }else
4228 #ifndef SQLITE_OMIT_OR_OPTIMIZATION
4229 if( pLevel->plan.wsFlags & WHERE_MULTI_OR ){
4230 /* Case 4: Two or more separately indexed terms connected by OR
4232 ** Example:
4234 ** CREATE TABLE t1(a,b,c,d);
4235 ** CREATE INDEX i1 ON t1(a);
4236 ** CREATE INDEX i2 ON t1(b);
4237 ** CREATE INDEX i3 ON t1(c);
4239 ** SELECT * FROM t1 WHERE a=5 OR b=7 OR (c=11 AND d=13)
4241 ** In the example, there are three indexed terms connected by OR.
4242 ** The top of the loop looks like this:
4244 ** Null 1 # Zero the rowset in reg 1
4246 ** Then, for each indexed term, the following. The arguments to
4247 ** RowSetTest are such that the rowid of the current row is inserted
4248 ** into the RowSet. If it is already present, control skips the
4249 ** Gosub opcode and jumps straight to the code generated by WhereEnd().
4251 ** sqlite3WhereBegin(<term>)
4252 ** RowSetTest # Insert rowid into rowset
4253 ** Gosub 2 A
4254 ** sqlite3WhereEnd()
4256 ** Following the above, code to terminate the loop. Label A, the target
4257 ** of the Gosub above, jumps to the instruction right after the Goto.
4259 ** Null 1 # Zero the rowset in reg 1
4260 ** Goto B # The loop is finished.
4262 ** A: <loop body> # Return data, whatever.
4264 ** Return 2 # Jump back to the Gosub
4266 ** B: <after the loop>
4269 WhereClause *pOrWc; /* The OR-clause broken out into subterms */
4270 SrcList *pOrTab; /* Shortened table list or OR-clause generation */
4272 int regReturn = ++pParse->nMem; /* Register used with OP_Gosub */
4273 int regRowset = 0; /* Register for RowSet object */
4274 int regRowid = 0; /* Register holding rowid */
4275 int iLoopBody = sqlite3VdbeMakeLabel(v); /* Start of loop body */
4276 int iRetInit; /* Address of regReturn init */
4277 int untestedTerms = 0; /* Some terms not completely tested */
4278 int ii; /* Loop counter */
4279 Expr *pAndExpr = 0; /* An ".. AND (...)" expression */
4281 pTerm = pLevel->plan.u.pTerm;
4282 assert( pTerm!=0 );
4283 assert( pTerm->eOperator==WO_OR );
4284 assert( (pTerm->wtFlags & TERM_ORINFO)!=0 );
4285 pOrWc = &pTerm->u.pOrInfo->wc;
4286 pLevel->op = OP_Return;
4287 pLevel->p1 = regReturn;
4289 /* Set up a new SrcList ni pOrTab containing the table being scanned
4290 ** by this loop in the a[0] slot and all notReady tables in a[1..] slots.
4291 ** This becomes the SrcList in the recursive call to sqlite3WhereBegin().
4293 if( pWInfo->nLevel>1 ){
4294 int nNotReady; /* The number of notReady tables */
4295 struct SrcList_item *origSrc; /* Original list of tables */
4296 nNotReady = pWInfo->nLevel - iLevel - 1;
4297 pOrTab = sqlite3StackAllocRaw(pParse->db,
4298 sizeof(*pOrTab)+ nNotReady*sizeof(pOrTab->a[0]));
4299 if( pOrTab==0 ) return notReady;
4300 pOrTab->nAlloc = (i16)(nNotReady + 1);
4301 pOrTab->nSrc = pOrTab->nAlloc;
4302 memcpy(pOrTab->a, pTabItem, sizeof(*pTabItem));
4303 origSrc = pWInfo->pTabList->a;
4304 for(k=1; k<=nNotReady; k++){
4305 memcpy(&pOrTab->a[k], &origSrc[pLevel[k].iFrom], sizeof(pOrTab->a[k]));
4307 }else{
4308 pOrTab = pWInfo->pTabList;
4311 /* Initialize the rowset register to contain NULL. An SQL NULL is
4312 ** equivalent to an empty rowset.
4314 ** Also initialize regReturn to contain the address of the instruction
4315 ** immediately following the OP_Return at the bottom of the loop. This
4316 ** is required in a few obscure LEFT JOIN cases where control jumps
4317 ** over the top of the loop into the body of it. In this case the
4318 ** correct response for the end-of-loop code (the OP_Return) is to
4319 ** fall through to the next instruction, just as an OP_Next does if
4320 ** called on an uninitialized cursor.
4322 if( (wctrlFlags & WHERE_DUPLICATES_OK)==0 ){
4323 regRowset = ++pParse->nMem;
4324 regRowid = ++pParse->nMem;
4325 sqlite3VdbeAddOp2(v, OP_Null, 0, regRowset);
4327 iRetInit = sqlite3VdbeAddOp2(v, OP_Integer, 0, regReturn);
4329 /* If the original WHERE clause is z of the form: (x1 OR x2 OR ...) AND y
4330 ** Then for every term xN, evaluate as the subexpression: xN AND z
4331 ** That way, terms in y that are factored into the disjunction will
4332 ** be picked up by the recursive calls to sqlite3WhereBegin() below.
4334 if( pWC->nTerm>1 ){
4335 pAndExpr = sqlite3ExprAlloc(pParse->db, TK_AND, 0, 0);
4336 pAndExpr->pRight = pWhere;
4339 for(ii=0; ii<pOrWc->nTerm; ii++){
4340 WhereTerm *pOrTerm = &pOrWc->a[ii];
4341 if( pOrTerm->leftCursor==iCur || pOrTerm->eOperator==WO_AND ){
4342 WhereInfo *pSubWInfo; /* Info for single OR-term scan */
4343 Expr *pOrExpr = pOrTerm->pExpr;
4344 if( pAndExpr ){
4345 pAndExpr->pLeft = pOrExpr;
4346 pOrExpr = pAndExpr;
4348 /* Loop through table entries that match term pOrTerm. */
4349 pSubWInfo = sqlite3WhereBegin(pParse, pOrTab, pOrExpr, 0, 0,
4350 WHERE_OMIT_OPEN_CLOSE | WHERE_AND_ONLY |
4351 WHERE_FORCE_TABLE | WHERE_ONETABLE_ONLY);
4352 if( pSubWInfo ){
4353 explainOneScan(
4354 pParse, pOrTab, &pSubWInfo->a[0], iLevel, pLevel->iFrom, 0
4356 if( (wctrlFlags & WHERE_DUPLICATES_OK)==0 ){
4357 int iSet = ((ii==pOrWc->nTerm-1)?-1:ii);
4358 int r;
4359 r = sqlite3ExprCodeGetColumn(pParse, pTabItem->pTab, -1, iCur,
4360 regRowid);
4361 sqlite3VdbeAddOp4Int(v, OP_RowSetTest, regRowset,
4362 sqlite3VdbeCurrentAddr(v)+2, r, iSet);
4364 sqlite3VdbeAddOp2(v, OP_Gosub, regReturn, iLoopBody);
4366 /* The pSubWInfo->untestedTerms flag means that this OR term
4367 ** contained one or more AND term from a notReady table. The
4368 ** terms from the notReady table could not be tested and will
4369 ** need to be tested later.
4371 if( pSubWInfo->untestedTerms ) untestedTerms = 1;
4373 /* Finish the loop through table entries that match term pOrTerm. */
4374 sqlite3WhereEnd(pSubWInfo);
4378 sqlite3DbFree(pParse->db, pAndExpr);
4379 sqlite3VdbeChangeP1(v, iRetInit, sqlite3VdbeCurrentAddr(v));
4380 sqlite3VdbeAddOp2(v, OP_Goto, 0, pLevel->addrBrk);
4381 sqlite3VdbeResolveLabel(v, iLoopBody);
4383 if( pWInfo->nLevel>1 ) sqlite3StackFree(pParse->db, pOrTab);
4384 if( !untestedTerms ) disableTerm(pLevel, pTerm);
4385 }else
4386 #endif /* SQLITE_OMIT_OR_OPTIMIZATION */
4389 /* Case 5: There is no usable index. We must do a complete
4390 ** scan of the entire table.
4392 static const u8 aStep[] = { OP_Next, OP_Prev };
4393 static const u8 aStart[] = { OP_Rewind, OP_Last };
4394 assert( bRev==0 || bRev==1 );
4395 assert( omitTable==0 );
4396 pLevel->op = aStep[bRev];
4397 pLevel->p1 = iCur;
4398 pLevel->p2 = 1 + sqlite3VdbeAddOp2(v, aStart[bRev], iCur, addrBrk);
4399 pLevel->p5 = SQLITE_STMTSTATUS_FULLSCAN_STEP;
4401 notReady &= ~getMask(pWC->pMaskSet, iCur);
4403 /* Insert code to test every subexpression that can be completely
4404 ** computed using the current set of tables.
4406 ** IMPLEMENTATION-OF: R-49525-50935 Terms that cannot be satisfied through
4407 ** the use of indices become tests that are evaluated against each row of
4408 ** the relevant input tables.
4410 for(pTerm=pWC->a, j=pWC->nTerm; j>0; j--, pTerm++){
4411 Expr *pE;
4412 testcase( pTerm->wtFlags & TERM_VIRTUAL ); /* IMP: R-30575-11662 */
4413 testcase( pTerm->wtFlags & TERM_CODED );
4414 if( pTerm->wtFlags & (TERM_VIRTUAL|TERM_CODED) ) continue;
4415 if( (pTerm->prereqAll & notReady)!=0 ){
4416 testcase( pWInfo->untestedTerms==0
4417 && (pWInfo->wctrlFlags & WHERE_ONETABLE_ONLY)!=0 );
4418 pWInfo->untestedTerms = 1;
4419 continue;
4421 pE = pTerm->pExpr;
4422 assert( pE!=0 );
4423 if( pLevel->iLeftJoin && !ExprHasProperty(pE, EP_FromJoin) ){
4424 continue;
4426 sqlite3ExprIfFalse(pParse, pE, addrCont, SQLITE_JUMPIFNULL);
4427 pTerm->wtFlags |= TERM_CODED;
4430 /* For a LEFT OUTER JOIN, generate code that will record the fact that
4431 ** at least one row of the right table has matched the left table.
4433 if( pLevel->iLeftJoin ){
4434 pLevel->addrFirst = sqlite3VdbeCurrentAddr(v);
4435 sqlite3VdbeAddOp2(v, OP_Integer, 1, pLevel->iLeftJoin);
4436 VdbeComment((v, "record LEFT JOIN hit"));
4437 sqlite3ExprCacheClear(pParse);
4438 for(pTerm=pWC->a, j=0; j<pWC->nTerm; j++, pTerm++){
4439 testcase( pTerm->wtFlags & TERM_VIRTUAL ); /* IMP: R-30575-11662 */
4440 testcase( pTerm->wtFlags & TERM_CODED );
4441 if( pTerm->wtFlags & (TERM_VIRTUAL|TERM_CODED) ) continue;
4442 if( (pTerm->prereqAll & notReady)!=0 ){
4443 assert( pWInfo->untestedTerms );
4444 continue;
4446 assert( pTerm->pExpr );
4447 sqlite3ExprIfFalse(pParse, pTerm->pExpr, addrCont, SQLITE_JUMPIFNULL);
4448 pTerm->wtFlags |= TERM_CODED;
4451 sqlite3ReleaseTempReg(pParse, iReleaseReg);
4453 return notReady;
4456 #if defined(SQLITE_TEST)
4458 ** The following variable holds a text description of query plan generated
4459 ** by the most recent call to sqlite3WhereBegin(). Each call to WhereBegin
4460 ** overwrites the previous. This information is used for testing and
4461 ** analysis only.
4463 char sqlite3_query_plan[BMS*2*40]; /* Text of the join */
4464 static int nQPlan = 0; /* Next free slow in _query_plan[] */
4466 #endif /* SQLITE_TEST */
4470 ** Free a WhereInfo structure
4472 static void whereInfoFree(sqlite3 *db, WhereInfo *pWInfo){
4473 if( ALWAYS(pWInfo) ){
4474 int i;
4475 for(i=0; i<pWInfo->nLevel; i++){
4476 sqlite3_index_info *pInfo = pWInfo->a[i].pIdxInfo;
4477 if( pInfo ){
4478 /* assert( pInfo->needToFreeIdxStr==0 || db->mallocFailed ); */
4479 if( pInfo->needToFreeIdxStr ){
4480 sqlite3_free(pInfo->idxStr);
4482 sqlite3DbFree(db, pInfo);
4484 if( pWInfo->a[i].plan.wsFlags & WHERE_TEMP_INDEX ){
4485 Index *pIdx = pWInfo->a[i].plan.u.pIdx;
4486 if( pIdx ){
4487 sqlite3DbFree(db, pIdx->zColAff);
4488 sqlite3DbFree(db, pIdx);
4492 whereClauseClear(pWInfo->pWC);
4493 sqlite3DbFree(db, pWInfo);
4499 ** Generate the beginning of the loop used for WHERE clause processing.
4500 ** The return value is a pointer to an opaque structure that contains
4501 ** information needed to terminate the loop. Later, the calling routine
4502 ** should invoke sqlite3WhereEnd() with the return value of this function
4503 ** in order to complete the WHERE clause processing.
4505 ** If an error occurs, this routine returns NULL.
4507 ** The basic idea is to do a nested loop, one loop for each table in
4508 ** the FROM clause of a select. (INSERT and UPDATE statements are the
4509 ** same as a SELECT with only a single table in the FROM clause.) For
4510 ** example, if the SQL is this:
4512 ** SELECT * FROM t1, t2, t3 WHERE ...;
4514 ** Then the code generated is conceptually like the following:
4516 ** foreach row1 in t1 do \ Code generated
4517 ** foreach row2 in t2 do |-- by sqlite3WhereBegin()
4518 ** foreach row3 in t3 do /
4519 ** ...
4520 ** end \ Code generated
4521 ** end |-- by sqlite3WhereEnd()
4522 ** end /
4524 ** Note that the loops might not be nested in the order in which they
4525 ** appear in the FROM clause if a different order is better able to make
4526 ** use of indices. Note also that when the IN operator appears in
4527 ** the WHERE clause, it might result in additional nested loops for
4528 ** scanning through all values on the right-hand side of the IN.
4530 ** There are Btree cursors associated with each table. t1 uses cursor
4531 ** number pTabList->a[0].iCursor. t2 uses the cursor pTabList->a[1].iCursor.
4532 ** And so forth. This routine generates code to open those VDBE cursors
4533 ** and sqlite3WhereEnd() generates the code to close them.
4535 ** The code that sqlite3WhereBegin() generates leaves the cursors named
4536 ** in pTabList pointing at their appropriate entries. The [...] code
4537 ** can use OP_Column and OP_Rowid opcodes on these cursors to extract
4538 ** data from the various tables of the loop.
4540 ** If the WHERE clause is empty, the foreach loops must each scan their
4541 ** entire tables. Thus a three-way join is an O(N^3) operation. But if
4542 ** the tables have indices and there are terms in the WHERE clause that
4543 ** refer to those indices, a complete table scan can be avoided and the
4544 ** code will run much faster. Most of the work of this routine is checking
4545 ** to see if there are indices that can be used to speed up the loop.
4547 ** Terms of the WHERE clause are also used to limit which rows actually
4548 ** make it to the "..." in the middle of the loop. After each "foreach",
4549 ** terms of the WHERE clause that use only terms in that loop and outer
4550 ** loops are evaluated and if false a jump is made around all subsequent
4551 ** inner loops (or around the "..." if the test occurs within the inner-
4552 ** most loop)
4554 ** OUTER JOINS
4556 ** An outer join of tables t1 and t2 is conceptally coded as follows:
4558 ** foreach row1 in t1 do
4559 ** flag = 0
4560 ** foreach row2 in t2 do
4561 ** start:
4562 ** ...
4563 ** flag = 1
4564 ** end
4565 ** if flag==0 then
4566 ** move the row2 cursor to a null row
4567 ** goto start
4568 ** fi
4569 ** end
4571 ** ORDER BY CLAUSE PROCESSING
4573 ** *ppOrderBy is a pointer to the ORDER BY clause of a SELECT statement,
4574 ** if there is one. If there is no ORDER BY clause or if this routine
4575 ** is called from an UPDATE or DELETE statement, then ppOrderBy is NULL.
4577 ** If an index can be used so that the natural output order of the table
4578 ** scan is correct for the ORDER BY clause, then that index is used and
4579 ** *ppOrderBy is set to NULL. This is an optimization that prevents an
4580 ** unnecessary sort of the result set if an index appropriate for the
4581 ** ORDER BY clause already exists.
4583 ** If the where clause loops cannot be arranged to provide the correct
4584 ** output order, then the *ppOrderBy is unchanged.
4586 WhereInfo *sqlite3WhereBegin(
4587 Parse *pParse, /* The parser context */
4588 SrcList *pTabList, /* A list of all tables to be scanned */
4589 Expr *pWhere, /* The WHERE clause */
4590 ExprList **ppOrderBy, /* An ORDER BY clause, or NULL */
4591 ExprList *pDistinct, /* The select-list for DISTINCT queries - or NULL */
4592 u16 wctrlFlags /* One of the WHERE_* flags defined in sqliteInt.h */
4594 int i; /* Loop counter */
4595 int nByteWInfo; /* Num. bytes allocated for WhereInfo struct */
4596 int nTabList; /* Number of elements in pTabList */
4597 WhereInfo *pWInfo; /* Will become the return value of this function */
4598 Vdbe *v = pParse->pVdbe; /* The virtual database engine */
4599 Bitmask notReady; /* Cursors that are not yet positioned */
4600 WhereMaskSet *pMaskSet; /* The expression mask set */
4601 WhereClause *pWC; /* Decomposition of the WHERE clause */
4602 struct SrcList_item *pTabItem; /* A single entry from pTabList */
4603 WhereLevel *pLevel; /* A single level in the pWInfo list */
4604 int iFrom; /* First unused FROM clause element */
4605 int andFlags; /* AND-ed combination of all pWC->a[].wtFlags */
4606 sqlite3 *db; /* Database connection */
4608 /* The number of tables in the FROM clause is limited by the number of
4609 ** bits in a Bitmask
4611 testcase( pTabList->nSrc==BMS );
4612 if( pTabList->nSrc>BMS ){
4613 sqlite3ErrorMsg(pParse, "at most %d tables in a join", BMS);
4614 return 0;
4617 /* This function normally generates a nested loop for all tables in
4618 ** pTabList. But if the WHERE_ONETABLE_ONLY flag is set, then we should
4619 ** only generate code for the first table in pTabList and assume that
4620 ** any cursors associated with subsequent tables are uninitialized.
4622 nTabList = (wctrlFlags & WHERE_ONETABLE_ONLY) ? 1 : pTabList->nSrc;
4624 /* Allocate and initialize the WhereInfo structure that will become the
4625 ** return value. A single allocation is used to store the WhereInfo
4626 ** struct, the contents of WhereInfo.a[], the WhereClause structure
4627 ** and the WhereMaskSet structure. Since WhereClause contains an 8-byte
4628 ** field (type Bitmask) it must be aligned on an 8-byte boundary on
4629 ** some architectures. Hence the ROUND8() below.
4631 db = pParse->db;
4632 nByteWInfo = ROUND8(sizeof(WhereInfo)+(nTabList-1)*sizeof(WhereLevel));
4633 pWInfo = sqlite3DbMallocZero(db,
4634 nByteWInfo +
4635 sizeof(WhereClause) +
4636 sizeof(WhereMaskSet)
4638 if( db->mallocFailed ){
4639 sqlite3DbFree(db, pWInfo);
4640 pWInfo = 0;
4641 goto whereBeginError;
4643 pWInfo->nLevel = nTabList;
4644 pWInfo->pParse = pParse;
4645 pWInfo->pTabList = pTabList;
4646 pWInfo->iBreak = sqlite3VdbeMakeLabel(v);
4647 pWInfo->pWC = pWC = (WhereClause *)&((u8 *)pWInfo)[nByteWInfo];
4648 pWInfo->wctrlFlags = wctrlFlags;
4649 pWInfo->savedNQueryLoop = pParse->nQueryLoop;
4650 pMaskSet = (WhereMaskSet*)&pWC[1];
4652 /* Disable the DISTINCT optimization if SQLITE_DistinctOpt is set via
4653 ** sqlite3_test_ctrl(SQLITE_TESTCTRL_OPTIMIZATIONS,...) */
4654 if( db->flags & SQLITE_DistinctOpt ) pDistinct = 0;
4656 /* Split the WHERE clause into separate subexpressions where each
4657 ** subexpression is separated by an AND operator.
4659 initMaskSet(pMaskSet);
4660 whereClauseInit(pWC, pParse, pMaskSet, wctrlFlags);
4661 sqlite3ExprCodeConstants(pParse, pWhere);
4662 whereSplit(pWC, pWhere, TK_AND); /* IMP: R-15842-53296 */
4664 /* Special case: a WHERE clause that is constant. Evaluate the
4665 ** expression and either jump over all of the code or fall thru.
4667 if( pWhere && (nTabList==0 || sqlite3ExprIsConstantNotJoin(pWhere)) ){
4668 sqlite3ExprIfFalse(pParse, pWhere, pWInfo->iBreak, SQLITE_JUMPIFNULL);
4669 pWhere = 0;
4672 /* Assign a bit from the bitmask to every term in the FROM clause.
4674 ** When assigning bitmask values to FROM clause cursors, it must be
4675 ** the case that if X is the bitmask for the N-th FROM clause term then
4676 ** the bitmask for all FROM clause terms to the left of the N-th term
4677 ** is (X-1). An expression from the ON clause of a LEFT JOIN can use
4678 ** its Expr.iRightJoinTable value to find the bitmask of the right table
4679 ** of the join. Subtracting one from the right table bitmask gives a
4680 ** bitmask for all tables to the left of the join. Knowing the bitmask
4681 ** for all tables to the left of a left join is important. Ticket #3015.
4683 ** Configure the WhereClause.vmask variable so that bits that correspond
4684 ** to virtual table cursors are set. This is used to selectively disable
4685 ** the OR-to-IN transformation in exprAnalyzeOrTerm(). It is not helpful
4686 ** with virtual tables.
4688 ** Note that bitmasks are created for all pTabList->nSrc tables in
4689 ** pTabList, not just the first nTabList tables. nTabList is normally
4690 ** equal to pTabList->nSrc but might be shortened to 1 if the
4691 ** WHERE_ONETABLE_ONLY flag is set.
4693 assert( pWC->vmask==0 && pMaskSet->n==0 );
4694 for(i=0; i<pTabList->nSrc; i++){
4695 createMask(pMaskSet, pTabList->a[i].iCursor);
4696 #ifndef SQLITE_OMIT_VIRTUALTABLE
4697 if( ALWAYS(pTabList->a[i].pTab) && IsVirtual(pTabList->a[i].pTab) ){
4698 pWC->vmask |= ((Bitmask)1 << i);
4700 #endif
4702 #ifndef NDEBUG
4704 Bitmask toTheLeft = 0;
4705 for(i=0; i<pTabList->nSrc; i++){
4706 Bitmask m = getMask(pMaskSet, pTabList->a[i].iCursor);
4707 assert( (m-1)==toTheLeft );
4708 toTheLeft |= m;
4711 #endif
4713 /* Analyze all of the subexpressions. Note that exprAnalyze() might
4714 ** add new virtual terms onto the end of the WHERE clause. We do not
4715 ** want to analyze these virtual terms, so start analyzing at the end
4716 ** and work forward so that the added virtual terms are never processed.
4718 exprAnalyzeAll(pTabList, pWC);
4719 if( db->mallocFailed ){
4720 goto whereBeginError;
4723 /* Check if the DISTINCT qualifier, if there is one, is redundant.
4724 ** If it is, then set pDistinct to NULL and WhereInfo.eDistinct to
4725 ** WHERE_DISTINCT_UNIQUE to tell the caller to ignore the DISTINCT.
4727 if( pDistinct && isDistinctRedundant(pParse, pTabList, pWC, pDistinct) ){
4728 pDistinct = 0;
4729 pWInfo->eDistinct = WHERE_DISTINCT_UNIQUE;
4732 /* Chose the best index to use for each table in the FROM clause.
4734 ** This loop fills in the following fields:
4736 ** pWInfo->a[].pIdx The index to use for this level of the loop.
4737 ** pWInfo->a[].wsFlags WHERE_xxx flags associated with pIdx
4738 ** pWInfo->a[].nEq The number of == and IN constraints
4739 ** pWInfo->a[].iFrom Which term of the FROM clause is being coded
4740 ** pWInfo->a[].iTabCur The VDBE cursor for the database table
4741 ** pWInfo->a[].iIdxCur The VDBE cursor for the index
4742 ** pWInfo->a[].pTerm When wsFlags==WO_OR, the OR-clause term
4744 ** This loop also figures out the nesting order of tables in the FROM
4745 ** clause.
4747 notReady = ~(Bitmask)0;
4748 andFlags = ~0;
4749 WHERETRACE(("*** Optimizer Start ***\n"));
4750 for(i=iFrom=0, pLevel=pWInfo->a; i<nTabList; i++, pLevel++){
4751 WhereCost bestPlan; /* Most efficient plan seen so far */
4752 Index *pIdx; /* Index for FROM table at pTabItem */
4753 int j; /* For looping over FROM tables */
4754 int bestJ = -1; /* The value of j */
4755 Bitmask m; /* Bitmask value for j or bestJ */
4756 int isOptimal; /* Iterator for optimal/non-optimal search */
4757 int nUnconstrained; /* Number tables without INDEXED BY */
4758 Bitmask notIndexed; /* Mask of tables that cannot use an index */
4760 memset(&bestPlan, 0, sizeof(bestPlan));
4761 bestPlan.rCost = SQLITE_BIG_DBL;
4762 WHERETRACE(("*** Begin search for loop %d ***\n", i));
4764 /* Loop through the remaining entries in the FROM clause to find the
4765 ** next nested loop. The loop tests all FROM clause entries
4766 ** either once or twice.
4768 ** The first test is always performed if there are two or more entries
4769 ** remaining and never performed if there is only one FROM clause entry
4770 ** to choose from. The first test looks for an "optimal" scan. In
4771 ** this context an optimal scan is one that uses the same strategy
4772 ** for the given FROM clause entry as would be selected if the entry
4773 ** were used as the innermost nested loop. In other words, a table
4774 ** is chosen such that the cost of running that table cannot be reduced
4775 ** by waiting for other tables to run first. This "optimal" test works
4776 ** by first assuming that the FROM clause is on the inner loop and finding
4777 ** its query plan, then checking to see if that query plan uses any
4778 ** other FROM clause terms that are notReady. If no notReady terms are
4779 ** used then the "optimal" query plan works.
4781 ** Note that the WhereCost.nRow parameter for an optimal scan might
4782 ** not be as small as it would be if the table really were the innermost
4783 ** join. The nRow value can be reduced by WHERE clause constraints
4784 ** that do not use indices. But this nRow reduction only happens if the
4785 ** table really is the innermost join.
4787 ** The second loop iteration is only performed if no optimal scan
4788 ** strategies were found by the first iteration. This second iteration
4789 ** is used to search for the lowest cost scan overall.
4791 ** Previous versions of SQLite performed only the second iteration -
4792 ** the next outermost loop was always that with the lowest overall
4793 ** cost. However, this meant that SQLite could select the wrong plan
4794 ** for scripts such as the following:
4796 ** CREATE TABLE t1(a, b);
4797 ** CREATE TABLE t2(c, d);
4798 ** SELECT * FROM t2, t1 WHERE t2.rowid = t1.a;
4800 ** The best strategy is to iterate through table t1 first. However it
4801 ** is not possible to determine this with a simple greedy algorithm.
4802 ** Since the cost of a linear scan through table t2 is the same
4803 ** as the cost of a linear scan through table t1, a simple greedy
4804 ** algorithm may choose to use t2 for the outer loop, which is a much
4805 ** costlier approach.
4807 nUnconstrained = 0;
4808 notIndexed = 0;
4809 for(isOptimal=(iFrom<nTabList-1); isOptimal>=0 && bestJ<0; isOptimal--){
4810 Bitmask mask; /* Mask of tables not yet ready */
4811 for(j=iFrom, pTabItem=&pTabList->a[j]; j<nTabList; j++, pTabItem++){
4812 int doNotReorder; /* True if this table should not be reordered */
4813 WhereCost sCost; /* Cost information from best[Virtual]Index() */
4814 ExprList *pOrderBy; /* ORDER BY clause for index to optimize */
4815 ExprList *pDist; /* DISTINCT clause for index to optimize */
4817 doNotReorder = (pTabItem->jointype & (JT_LEFT|JT_CROSS))!=0;
4818 if( j!=iFrom && doNotReorder ) break;
4819 m = getMask(pMaskSet, pTabItem->iCursor);
4820 if( (m & notReady)==0 ){
4821 if( j==iFrom ) iFrom++;
4822 continue;
4824 mask = (isOptimal ? m : notReady);
4825 pOrderBy = ((i==0 && ppOrderBy )?*ppOrderBy:0);
4826 pDist = (i==0 ? pDistinct : 0);
4827 if( pTabItem->pIndex==0 ) nUnconstrained++;
4829 WHERETRACE(("=== trying table %d with isOptimal=%d ===\n",
4830 j, isOptimal));
4831 assert( pTabItem->pTab );
4832 #ifndef SQLITE_OMIT_VIRTUALTABLE
4833 if( IsVirtual(pTabItem->pTab) ){
4834 sqlite3_index_info **pp = &pWInfo->a[j].pIdxInfo;
4835 bestVirtualIndex(pParse, pWC, pTabItem, mask, notReady, pOrderBy,
4836 &sCost, pp);
4837 }else
4838 #endif
4840 bestBtreeIndex(pParse, pWC, pTabItem, mask, notReady, pOrderBy,
4841 pDist, &sCost);
4843 assert( isOptimal || (sCost.used&notReady)==0 );
4845 /* If an INDEXED BY clause is present, then the plan must use that
4846 ** index if it uses any index at all */
4847 assert( pTabItem->pIndex==0
4848 || (sCost.plan.wsFlags & WHERE_NOT_FULLSCAN)==0
4849 || sCost.plan.u.pIdx==pTabItem->pIndex );
4851 if( isOptimal && (sCost.plan.wsFlags & WHERE_NOT_FULLSCAN)==0 ){
4852 notIndexed |= m;
4855 /* Conditions under which this table becomes the best so far:
4857 ** (1) The table must not depend on other tables that have not
4858 ** yet run.
4860 ** (2) A full-table-scan plan cannot supercede indexed plan unless
4861 ** the full-table-scan is an "optimal" plan as defined above.
4863 ** (3) All tables have an INDEXED BY clause or this table lacks an
4864 ** INDEXED BY clause or this table uses the specific
4865 ** index specified by its INDEXED BY clause. This rule ensures
4866 ** that a best-so-far is always selected even if an impossible
4867 ** combination of INDEXED BY clauses are given. The error
4868 ** will be detected and relayed back to the application later.
4869 ** The NEVER() comes about because rule (2) above prevents
4870 ** An indexable full-table-scan from reaching rule (3).
4872 ** (4) The plan cost must be lower than prior plans or else the
4873 ** cost must be the same and the number of rows must be lower.
4875 if( (sCost.used&notReady)==0 /* (1) */
4876 && (bestJ<0 || (notIndexed&m)!=0 /* (2) */
4877 || (bestPlan.plan.wsFlags & WHERE_NOT_FULLSCAN)==0
4878 || (sCost.plan.wsFlags & WHERE_NOT_FULLSCAN)!=0)
4879 && (nUnconstrained==0 || pTabItem->pIndex==0 /* (3) */
4880 || NEVER((sCost.plan.wsFlags & WHERE_NOT_FULLSCAN)!=0))
4881 && (bestJ<0 || sCost.rCost<bestPlan.rCost /* (4) */
4882 || (sCost.rCost<=bestPlan.rCost
4883 && sCost.plan.nRow<bestPlan.plan.nRow))
4885 WHERETRACE(("=== table %d is best so far"
4886 " with cost=%g and nRow=%g\n",
4887 j, sCost.rCost, sCost.plan.nRow));
4888 bestPlan = sCost;
4889 bestJ = j;
4891 if( doNotReorder ) break;
4894 assert( bestJ>=0 );
4895 assert( notReady & getMask(pMaskSet, pTabList->a[bestJ].iCursor) );
4896 WHERETRACE(("*** Optimizer selects table %d for loop %d"
4897 " with cost=%g and nRow=%g\n",
4898 bestJ, pLevel-pWInfo->a, bestPlan.rCost, bestPlan.plan.nRow));
4899 /* The ALWAYS() that follows was added to hush up clang scan-build */
4900 if( (bestPlan.plan.wsFlags & WHERE_ORDERBY)!=0 && ALWAYS(ppOrderBy) ){
4901 *ppOrderBy = 0;
4903 if( (bestPlan.plan.wsFlags & WHERE_DISTINCT)!=0 ){
4904 assert( pWInfo->eDistinct==0 );
4905 pWInfo->eDistinct = WHERE_DISTINCT_ORDERED;
4907 andFlags &= bestPlan.plan.wsFlags;
4908 pLevel->plan = bestPlan.plan;
4909 testcase( bestPlan.plan.wsFlags & WHERE_INDEXED );
4910 testcase( bestPlan.plan.wsFlags & WHERE_TEMP_INDEX );
4911 if( bestPlan.plan.wsFlags & (WHERE_INDEXED|WHERE_TEMP_INDEX) ){
4912 pLevel->iIdxCur = pParse->nTab++;
4913 }else{
4914 pLevel->iIdxCur = -1;
4916 notReady &= ~getMask(pMaskSet, pTabList->a[bestJ].iCursor);
4917 pLevel->iFrom = (u8)bestJ;
4918 if( bestPlan.plan.nRow>=(double)1 ){
4919 pParse->nQueryLoop *= bestPlan.plan.nRow;
4922 /* Check that if the table scanned by this loop iteration had an
4923 ** INDEXED BY clause attached to it, that the named index is being
4924 ** used for the scan. If not, then query compilation has failed.
4925 ** Return an error.
4927 pIdx = pTabList->a[bestJ].pIndex;
4928 if( pIdx ){
4929 if( (bestPlan.plan.wsFlags & WHERE_INDEXED)==0 ){
4930 sqlite3ErrorMsg(pParse, "cannot use index: %s", pIdx->zName);
4931 goto whereBeginError;
4932 }else{
4933 /* If an INDEXED BY clause is used, the bestIndex() function is
4934 ** guaranteed to find the index specified in the INDEXED BY clause
4935 ** if it find an index at all. */
4936 assert( bestPlan.plan.u.pIdx==pIdx );
4940 WHERETRACE(("*** Optimizer Finished ***\n"));
4941 if( pParse->nErr || db->mallocFailed ){
4942 goto whereBeginError;
4945 /* If the total query only selects a single row, then the ORDER BY
4946 ** clause is irrelevant.
4948 if( (andFlags & WHERE_UNIQUE)!=0 && ppOrderBy ){
4949 *ppOrderBy = 0;
4952 /* If the caller is an UPDATE or DELETE statement that is requesting
4953 ** to use a one-pass algorithm, determine if this is appropriate.
4954 ** The one-pass algorithm only works if the WHERE clause constraints
4955 ** the statement to update a single row.
4957 assert( (wctrlFlags & WHERE_ONEPASS_DESIRED)==0 || pWInfo->nLevel==1 );
4958 if( (wctrlFlags & WHERE_ONEPASS_DESIRED)!=0 && (andFlags & WHERE_UNIQUE)!=0 ){
4959 pWInfo->okOnePass = 1;
4960 pWInfo->a[0].plan.wsFlags &= ~WHERE_IDX_ONLY;
4963 /* Open all tables in the pTabList and any indices selected for
4964 ** searching those tables.
4966 sqlite3CodeVerifySchema(pParse, -1); /* Insert the cookie verifier Goto */
4967 notReady = ~(Bitmask)0;
4968 pWInfo->nRowOut = (double)1;
4969 for(i=0, pLevel=pWInfo->a; i<nTabList; i++, pLevel++){
4970 Table *pTab; /* Table to open */
4971 int iDb; /* Index of database containing table/index */
4973 pTabItem = &pTabList->a[pLevel->iFrom];
4974 pTab = pTabItem->pTab;
4975 pLevel->iTabCur = pTabItem->iCursor;
4976 pWInfo->nRowOut *= pLevel->plan.nRow;
4977 iDb = sqlite3SchemaToIndex(db, pTab->pSchema);
4978 if( (pTab->tabFlags & TF_Ephemeral)!=0 || pTab->pSelect ){
4979 /* Do nothing */
4980 }else
4981 #ifndef SQLITE_OMIT_VIRTUALTABLE
4982 if( (pLevel->plan.wsFlags & WHERE_VIRTUALTABLE)!=0 ){
4983 const char *pVTab = (const char *)sqlite3GetVTable(db, pTab);
4984 int iCur = pTabItem->iCursor;
4985 sqlite3VdbeAddOp4(v, OP_VOpen, iCur, 0, 0, pVTab, P4_VTAB);
4986 }else
4987 #endif
4988 if( (pLevel->plan.wsFlags & WHERE_IDX_ONLY)==0
4989 && (wctrlFlags & WHERE_OMIT_OPEN_CLOSE)==0 ){
4990 int op = pWInfo->okOnePass ? OP_OpenWrite : OP_OpenRead;
4991 sqlite3OpenTable(pParse, pTabItem->iCursor, iDb, pTab, op);
4992 testcase( pTab->nCol==BMS-1 );
4993 testcase( pTab->nCol==BMS );
4994 if( !pWInfo->okOnePass && pTab->nCol<BMS ){
4995 Bitmask b = pTabItem->colUsed;
4996 int n = 0;
4997 for(; b; b=b>>1, n++){}
4998 sqlite3VdbeChangeP4(v, sqlite3VdbeCurrentAddr(v)-1,
4999 SQLITE_INT_TO_PTR(n), P4_INT32);
5000 assert( n<=pTab->nCol );
5002 }else{
5003 sqlite3TableLock(pParse, iDb, pTab->tnum, 0, pTab->zName);
5005 #ifndef SQLITE_OMIT_AUTOMATIC_INDEX
5006 if( (pLevel->plan.wsFlags & WHERE_TEMP_INDEX)!=0 ){
5007 constructAutomaticIndex(pParse, pWC, pTabItem, notReady, pLevel);
5008 }else
5009 #endif
5010 if( (pLevel->plan.wsFlags & WHERE_INDEXED)!=0 ){
5011 Index *pIx = pLevel->plan.u.pIdx;
5012 KeyInfo *pKey = sqlite3IndexKeyinfo(pParse, pIx);
5013 int iIdxCur = pLevel->iIdxCur;
5014 assert( pIx->pSchema==pTab->pSchema );
5015 assert( iIdxCur>=0 );
5016 sqlite3VdbeAddOp4(v, OP_OpenRead, iIdxCur, pIx->tnum, iDb,
5017 (char*)pKey, P4_KEYINFO_HANDOFF);
5018 VdbeComment((v, "%s", pIx->zName));
5020 sqlite3CodeVerifySchema(pParse, iDb);
5021 notReady &= ~getMask(pWC->pMaskSet, pTabItem->iCursor);
5023 pWInfo->iTop = sqlite3VdbeCurrentAddr(v);
5024 if( db->mallocFailed ) goto whereBeginError;
5026 /* Generate the code to do the search. Each iteration of the for
5027 ** loop below generates code for a single nested loop of the VM
5028 ** program.
5030 notReady = ~(Bitmask)0;
5031 for(i=0; i<nTabList; i++){
5032 pLevel = &pWInfo->a[i];
5033 explainOneScan(pParse, pTabList, pLevel, i, pLevel->iFrom, wctrlFlags);
5034 notReady = codeOneLoopStart(pWInfo, i, wctrlFlags, notReady, pWhere);
5035 pWInfo->iContinue = pLevel->addrCont;
5038 #ifdef SQLITE_TEST /* For testing and debugging use only */
5039 /* Record in the query plan information about the current table
5040 ** and the index used to access it (if any). If the table itself
5041 ** is not used, its name is just '{}'. If no index is used
5042 ** the index is listed as "{}". If the primary key is used the
5043 ** index name is '*'.
5045 for(i=0; i<nTabList; i++){
5046 char *z;
5047 int n;
5048 pLevel = &pWInfo->a[i];
5049 pTabItem = &pTabList->a[pLevel->iFrom];
5050 z = pTabItem->zAlias;
5051 if( z==0 ) z = pTabItem->pTab->zName;
5052 n = sqlite3Strlen30(z);
5053 if( n+nQPlan < sizeof(sqlite3_query_plan)-10 ){
5054 if( pLevel->plan.wsFlags & WHERE_IDX_ONLY ){
5055 memcpy(&sqlite3_query_plan[nQPlan], "{}", 2);
5056 nQPlan += 2;
5057 }else{
5058 memcpy(&sqlite3_query_plan[nQPlan], z, n);
5059 nQPlan += n;
5061 sqlite3_query_plan[nQPlan++] = ' ';
5063 testcase( pLevel->plan.wsFlags & WHERE_ROWID_EQ );
5064 testcase( pLevel->plan.wsFlags & WHERE_ROWID_RANGE );
5065 if( pLevel->plan.wsFlags & (WHERE_ROWID_EQ|WHERE_ROWID_RANGE) ){
5066 memcpy(&sqlite3_query_plan[nQPlan], "* ", 2);
5067 nQPlan += 2;
5068 }else if( (pLevel->plan.wsFlags & WHERE_INDEXED)!=0 ){
5069 n = sqlite3Strlen30(pLevel->plan.u.pIdx->zName);
5070 if( n+nQPlan < sizeof(sqlite3_query_plan)-2 ){
5071 memcpy(&sqlite3_query_plan[nQPlan], pLevel->plan.u.pIdx->zName, n);
5072 nQPlan += n;
5073 sqlite3_query_plan[nQPlan++] = ' ';
5075 }else{
5076 memcpy(&sqlite3_query_plan[nQPlan], "{} ", 3);
5077 nQPlan += 3;
5080 while( nQPlan>0 && sqlite3_query_plan[nQPlan-1]==' ' ){
5081 sqlite3_query_plan[--nQPlan] = 0;
5083 sqlite3_query_plan[nQPlan] = 0;
5084 nQPlan = 0;
5085 #endif /* SQLITE_TEST // Testing and debugging use only */
5087 /* Record the continuation address in the WhereInfo structure. Then
5088 ** clean up and return.
5090 return pWInfo;
5092 /* Jump here if malloc fails */
5093 whereBeginError:
5094 if( pWInfo ){
5095 pParse->nQueryLoop = pWInfo->savedNQueryLoop;
5096 whereInfoFree(db, pWInfo);
5098 return 0;
5102 ** Generate the end of the WHERE loop. See comments on
5103 ** sqlite3WhereBegin() for additional information.
5105 void sqlite3WhereEnd(WhereInfo *pWInfo){
5106 Parse *pParse = pWInfo->pParse;
5107 Vdbe *v = pParse->pVdbe;
5108 int i;
5109 WhereLevel *pLevel;
5110 SrcList *pTabList = pWInfo->pTabList;
5111 sqlite3 *db = pParse->db;
5113 /* Generate loop termination code.
5115 sqlite3ExprCacheClear(pParse);
5116 for(i=pWInfo->nLevel-1; i>=0; i--){
5117 pLevel = &pWInfo->a[i];
5118 sqlite3VdbeResolveLabel(v, pLevel->addrCont);
5119 if( pLevel->op!=OP_Noop ){
5120 sqlite3VdbeAddOp2(v, pLevel->op, pLevel->p1, pLevel->p2);
5121 sqlite3VdbeChangeP5(v, pLevel->p5);
5123 if( pLevel->plan.wsFlags & WHERE_IN_ABLE && pLevel->u.in.nIn>0 ){
5124 struct InLoop *pIn;
5125 int j;
5126 sqlite3VdbeResolveLabel(v, pLevel->addrNxt);
5127 for(j=pLevel->u.in.nIn, pIn=&pLevel->u.in.aInLoop[j-1]; j>0; j--, pIn--){
5128 sqlite3VdbeJumpHere(v, pIn->addrInTop+1);
5129 sqlite3VdbeAddOp2(v, OP_Next, pIn->iCur, pIn->addrInTop);
5130 sqlite3VdbeJumpHere(v, pIn->addrInTop-1);
5132 sqlite3DbFree(db, pLevel->u.in.aInLoop);
5134 sqlite3VdbeResolveLabel(v, pLevel->addrBrk);
5135 if( pLevel->iLeftJoin ){
5136 int addr;
5137 addr = sqlite3VdbeAddOp1(v, OP_IfPos, pLevel->iLeftJoin);
5138 assert( (pLevel->plan.wsFlags & WHERE_IDX_ONLY)==0
5139 || (pLevel->plan.wsFlags & WHERE_INDEXED)!=0 );
5140 if( (pLevel->plan.wsFlags & WHERE_IDX_ONLY)==0 ){
5141 sqlite3VdbeAddOp1(v, OP_NullRow, pTabList->a[i].iCursor);
5143 if( pLevel->iIdxCur>=0 ){
5144 sqlite3VdbeAddOp1(v, OP_NullRow, pLevel->iIdxCur);
5146 if( pLevel->op==OP_Return ){
5147 sqlite3VdbeAddOp2(v, OP_Gosub, pLevel->p1, pLevel->addrFirst);
5148 }else{
5149 sqlite3VdbeAddOp2(v, OP_Goto, 0, pLevel->addrFirst);
5151 sqlite3VdbeJumpHere(v, addr);
5155 /* The "break" point is here, just past the end of the outer loop.
5156 ** Set it.
5158 sqlite3VdbeResolveLabel(v, pWInfo->iBreak);
5160 /* Close all of the cursors that were opened by sqlite3WhereBegin.
5162 assert( pWInfo->nLevel==1 || pWInfo->nLevel==pTabList->nSrc );
5163 for(i=0, pLevel=pWInfo->a; i<pWInfo->nLevel; i++, pLevel++){
5164 struct SrcList_item *pTabItem = &pTabList->a[pLevel->iFrom];
5165 Table *pTab = pTabItem->pTab;
5166 assert( pTab!=0 );
5167 if( (pTab->tabFlags & TF_Ephemeral)==0
5168 && pTab->pSelect==0
5169 && (pWInfo->wctrlFlags & WHERE_OMIT_OPEN_CLOSE)==0
5171 int ws = pLevel->plan.wsFlags;
5172 if( !pWInfo->okOnePass && (ws & WHERE_IDX_ONLY)==0 ){
5173 sqlite3VdbeAddOp1(v, OP_Close, pTabItem->iCursor);
5175 if( (ws & WHERE_INDEXED)!=0 && (ws & WHERE_TEMP_INDEX)==0 ){
5176 sqlite3VdbeAddOp1(v, OP_Close, pLevel->iIdxCur);
5180 /* If this scan uses an index, make code substitutions to read data
5181 ** from the index in preference to the table. Sometimes, this means
5182 ** the table need never be read from. This is a performance boost,
5183 ** as the vdbe level waits until the table is read before actually
5184 ** seeking the table cursor to the record corresponding to the current
5185 ** position in the index.
5187 ** Calls to the code generator in between sqlite3WhereBegin and
5188 ** sqlite3WhereEnd will have created code that references the table
5189 ** directly. This loop scans all that code looking for opcodes
5190 ** that reference the table and converts them into opcodes that
5191 ** reference the index.
5193 if( (pLevel->plan.wsFlags & WHERE_INDEXED)!=0 && !db->mallocFailed){
5194 int k, j, last;
5195 VdbeOp *pOp;
5196 Index *pIdx = pLevel->plan.u.pIdx;
5198 assert( pIdx!=0 );
5199 pOp = sqlite3VdbeGetOp(v, pWInfo->iTop);
5200 last = sqlite3VdbeCurrentAddr(v);
5201 for(k=pWInfo->iTop; k<last; k++, pOp++){
5202 if( pOp->p1!=pLevel->iTabCur ) continue;
5203 if( pOp->opcode==OP_Column ){
5204 for(j=0; j<pIdx->nColumn; j++){
5205 if( pOp->p2==pIdx->aiColumn[j] ){
5206 pOp->p2 = j;
5207 pOp->p1 = pLevel->iIdxCur;
5208 break;
5211 assert( (pLevel->plan.wsFlags & WHERE_IDX_ONLY)==0
5212 || j<pIdx->nColumn );
5213 }else if( pOp->opcode==OP_Rowid ){
5214 pOp->p1 = pLevel->iIdxCur;
5215 pOp->opcode = OP_IdxRowid;
5221 /* Final cleanup
5223 pParse->nQueryLoop = pWInfo->savedNQueryLoop;
5224 whereInfoFree(db, pWInfo);
5225 return;