Enhance the command-line completion extension to return the names of
[sqlite.git] / src / whereexpr.c
blob313c5ee9bc79222490c1b36604b3394869bb5abf
1 /*
2 ** 2015-06-08
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.
15 ** This file was originally part of where.c but was split out to improve
16 ** readability and editabiliity. This file contains utility routines for
17 ** analyzing Expr objects in the WHERE clause.
19 #include "sqliteInt.h"
20 #include "whereInt.h"
22 /* Forward declarations */
23 static void exprAnalyze(SrcList*, WhereClause*, int);
26 ** Deallocate all memory associated with a WhereOrInfo object.
28 static void whereOrInfoDelete(sqlite3 *db, WhereOrInfo *p){
29 sqlite3WhereClauseClear(&p->wc);
30 sqlite3DbFree(db, p);
34 ** Deallocate all memory associated with a WhereAndInfo object.
36 static void whereAndInfoDelete(sqlite3 *db, WhereAndInfo *p){
37 sqlite3WhereClauseClear(&p->wc);
38 sqlite3DbFree(db, p);
42 ** Add a single new WhereTerm entry to the WhereClause object pWC.
43 ** The new WhereTerm object is constructed from Expr p and with wtFlags.
44 ** The index in pWC->a[] of the new WhereTerm is returned on success.
45 ** 0 is returned if the new WhereTerm could not be added due to a memory
46 ** allocation error. The memory allocation failure will be recorded in
47 ** the db->mallocFailed flag so that higher-level functions can detect it.
49 ** This routine will increase the size of the pWC->a[] array as necessary.
51 ** If the wtFlags argument includes TERM_DYNAMIC, then responsibility
52 ** for freeing the expression p is assumed by the WhereClause object pWC.
53 ** This is true even if this routine fails to allocate a new WhereTerm.
55 ** WARNING: This routine might reallocate the space used to store
56 ** WhereTerms. All pointers to WhereTerms should be invalidated after
57 ** calling this routine. Such pointers may be reinitialized by referencing
58 ** the pWC->a[] array.
60 static int whereClauseInsert(WhereClause *pWC, Expr *p, u16 wtFlags){
61 WhereTerm *pTerm;
62 int idx;
63 testcase( wtFlags & TERM_VIRTUAL );
64 if( pWC->nTerm>=pWC->nSlot ){
65 WhereTerm *pOld = pWC->a;
66 sqlite3 *db = pWC->pWInfo->pParse->db;
67 pWC->a = sqlite3DbMallocRawNN(db, sizeof(pWC->a[0])*pWC->nSlot*2 );
68 if( pWC->a==0 ){
69 if( wtFlags & TERM_DYNAMIC ){
70 sqlite3ExprDelete(db, p);
72 pWC->a = pOld;
73 return 0;
75 memcpy(pWC->a, pOld, sizeof(pWC->a[0])*pWC->nTerm);
76 if( pOld!=pWC->aStatic ){
77 sqlite3DbFree(db, pOld);
79 pWC->nSlot = sqlite3DbMallocSize(db, pWC->a)/sizeof(pWC->a[0]);
81 pTerm = &pWC->a[idx = pWC->nTerm++];
82 if( p && ExprHasProperty(p, EP_Unlikely) ){
83 pTerm->truthProb = sqlite3LogEst(p->iTable) - 270;
84 }else{
85 pTerm->truthProb = 1;
87 pTerm->pExpr = sqlite3ExprSkipCollate(p);
88 pTerm->wtFlags = wtFlags;
89 pTerm->pWC = pWC;
90 pTerm->iParent = -1;
91 memset(&pTerm->eOperator, 0,
92 sizeof(WhereTerm) - offsetof(WhereTerm,eOperator));
93 return idx;
97 ** Return TRUE if the given operator is one of the operators that is
98 ** allowed for an indexable WHERE clause term. The allowed operators are
99 ** "=", "<", ">", "<=", ">=", "IN", "IS", and "IS NULL"
101 static int allowedOp(int op){
102 assert( TK_GT>TK_EQ && TK_GT<TK_GE );
103 assert( TK_LT>TK_EQ && TK_LT<TK_GE );
104 assert( TK_LE>TK_EQ && TK_LE<TK_GE );
105 assert( TK_GE==TK_EQ+4 );
106 return op==TK_IN || (op>=TK_EQ && op<=TK_GE) || op==TK_ISNULL || op==TK_IS;
110 ** Commute a comparison operator. Expressions of the form "X op Y"
111 ** are converted into "Y op X".
113 ** If left/right precedence rules come into play when determining the
114 ** collating sequence, then COLLATE operators are adjusted to ensure
115 ** that the collating sequence does not change. For example:
116 ** "Y collate NOCASE op X" becomes "X op Y" because any collation sequence on
117 ** the left hand side of a comparison overrides any collation sequence
118 ** attached to the right. For the same reason the EP_Collate flag
119 ** is not commuted.
121 static void exprCommute(Parse *pParse, Expr *pExpr){
122 u16 expRight = (pExpr->pRight->flags & EP_Collate);
123 u16 expLeft = (pExpr->pLeft->flags & EP_Collate);
124 assert( allowedOp(pExpr->op) && pExpr->op!=TK_IN );
125 if( expRight==expLeft ){
126 /* Either X and Y both have COLLATE operator or neither do */
127 if( expRight ){
128 /* Both X and Y have COLLATE operators. Make sure X is always
129 ** used by clearing the EP_Collate flag from Y. */
130 pExpr->pRight->flags &= ~EP_Collate;
131 }else if( sqlite3ExprCollSeq(pParse, pExpr->pLeft)!=0 ){
132 /* Neither X nor Y have COLLATE operators, but X has a non-default
133 ** collating sequence. So add the EP_Collate marker on X to cause
134 ** it to be searched first. */
135 pExpr->pLeft->flags |= EP_Collate;
138 SWAP(Expr*,pExpr->pRight,pExpr->pLeft);
139 if( pExpr->op>=TK_GT ){
140 assert( TK_LT==TK_GT+2 );
141 assert( TK_GE==TK_LE+2 );
142 assert( TK_GT>TK_EQ );
143 assert( TK_GT<TK_LE );
144 assert( pExpr->op>=TK_GT && pExpr->op<=TK_GE );
145 pExpr->op = ((pExpr->op-TK_GT)^2)+TK_GT;
150 ** Translate from TK_xx operator to WO_xx bitmask.
152 static u16 operatorMask(int op){
153 u16 c;
154 assert( allowedOp(op) );
155 if( op==TK_IN ){
156 c = WO_IN;
157 }else if( op==TK_ISNULL ){
158 c = WO_ISNULL;
159 }else if( op==TK_IS ){
160 c = WO_IS;
161 }else{
162 assert( (WO_EQ<<(op-TK_EQ)) < 0x7fff );
163 c = (u16)(WO_EQ<<(op-TK_EQ));
165 assert( op!=TK_ISNULL || c==WO_ISNULL );
166 assert( op!=TK_IN || c==WO_IN );
167 assert( op!=TK_EQ || c==WO_EQ );
168 assert( op!=TK_LT || c==WO_LT );
169 assert( op!=TK_LE || c==WO_LE );
170 assert( op!=TK_GT || c==WO_GT );
171 assert( op!=TK_GE || c==WO_GE );
172 assert( op!=TK_IS || c==WO_IS );
173 return c;
177 #ifndef SQLITE_OMIT_LIKE_OPTIMIZATION
179 ** Check to see if the given expression is a LIKE or GLOB operator that
180 ** can be optimized using inequality constraints. Return TRUE if it is
181 ** so and false if not.
183 ** In order for the operator to be optimizible, the RHS must be a string
184 ** literal that does not begin with a wildcard. The LHS must be a column
185 ** that may only be NULL, a string, or a BLOB, never a number. (This means
186 ** that virtual tables cannot participate in the LIKE optimization.) The
187 ** collating sequence for the column on the LHS must be appropriate for
188 ** the operator.
190 static int isLikeOrGlob(
191 Parse *pParse, /* Parsing and code generating context */
192 Expr *pExpr, /* Test this expression */
193 Expr **ppPrefix, /* Pointer to TK_STRING expression with pattern prefix */
194 int *pisComplete, /* True if the only wildcard is % in the last character */
195 int *pnoCase /* True if uppercase is equivalent to lowercase */
197 const u8 *z = 0; /* String on RHS of LIKE operator */
198 Expr *pRight, *pLeft; /* Right and left size of LIKE operator */
199 ExprList *pList; /* List of operands to the LIKE operator */
200 int c; /* One character in z[] */
201 int cnt; /* Number of non-wildcard prefix characters */
202 char wc[4]; /* Wildcard characters */
203 sqlite3 *db = pParse->db; /* Database connection */
204 sqlite3_value *pVal = 0;
205 int op; /* Opcode of pRight */
206 int rc; /* Result code to return */
208 if( !sqlite3IsLikeFunction(db, pExpr, pnoCase, wc) ){
209 return 0;
211 #ifdef SQLITE_EBCDIC
212 if( *pnoCase ) return 0;
213 #endif
214 pList = pExpr->x.pList;
215 pLeft = pList->a[1].pExpr;
217 pRight = sqlite3ExprSkipCollate(pList->a[0].pExpr);
218 op = pRight->op;
219 if( op==TK_VARIABLE && (db->flags & SQLITE_EnableQPSG)==0 ){
220 Vdbe *pReprepare = pParse->pReprepare;
221 int iCol = pRight->iColumn;
222 pVal = sqlite3VdbeGetBoundValue(pReprepare, iCol, SQLITE_AFF_BLOB);
223 if( pVal && sqlite3_value_type(pVal)==SQLITE_TEXT ){
224 z = sqlite3_value_text(pVal);
226 sqlite3VdbeSetVarmask(pParse->pVdbe, iCol);
227 assert( pRight->op==TK_VARIABLE || pRight->op==TK_REGISTER );
228 }else if( op==TK_STRING ){
229 z = (u8*)pRight->u.zToken;
231 if( z ){
233 /* If the RHS begins with a digit or a minus sign, then the LHS must
234 ** be an ordinary column (not a virtual table column) with TEXT affinity.
235 ** Otherwise the LHS might be numeric and "lhs >= rhs" would be false
236 ** even though "lhs LIKE rhs" is true. But if the RHS does not start
237 ** with a digit or '-', then "lhs LIKE rhs" will always be false if
238 ** the LHS is numeric and so the optimization still works.
240 if( sqlite3Isdigit(z[0]) || z[0]=='-' ){
241 if( pLeft->op!=TK_COLUMN
242 || sqlite3ExprAffinity(pLeft)!=SQLITE_AFF_TEXT
243 || IsVirtual(pLeft->pTab) /* Value might be numeric */
245 sqlite3ValueFree(pVal);
246 return 0;
250 /* Count the number of prefix characters prior to the first wildcard */
251 cnt = 0;
252 while( (c=z[cnt])!=0 && c!=wc[0] && c!=wc[1] && c!=wc[2] ){
253 cnt++;
254 if( c==wc[3] && z[cnt]!=0 ) cnt++;
257 /* The optimization is possible only if (1) the pattern does not begin
258 ** with a wildcard and if (2) the non-wildcard prefix does not end with
259 ** an (illegal 0xff) character. The second condition is necessary so
260 ** that we can increment the prefix key to find an upper bound for the
261 ** range search.
263 if( cnt!=0 && 255!=(u8)z[cnt-1] ){
264 Expr *pPrefix;
266 /* A "complete" match if the pattern ends with "*" or "%" */
267 *pisComplete = c==wc[0] && z[cnt+1]==0;
269 /* Get the pattern prefix. Remove all escapes from the prefix. */
270 pPrefix = sqlite3Expr(db, TK_STRING, (char*)z);
271 if( pPrefix ){
272 int iFrom, iTo;
273 char *zNew = pPrefix->u.zToken;
274 zNew[cnt] = 0;
275 for(iFrom=iTo=0; iFrom<cnt; iFrom++){
276 if( zNew[iFrom]==wc[3] ) iFrom++;
277 zNew[iTo++] = zNew[iFrom];
279 zNew[iTo] = 0;
281 *ppPrefix = pPrefix;
283 /* If the RHS pattern is a bound parameter, make arrangements to
284 ** reprepare the statement when that parameter is rebound */
285 if( op==TK_VARIABLE ){
286 Vdbe *v = pParse->pVdbe;
287 sqlite3VdbeSetVarmask(v, pRight->iColumn);
288 if( *pisComplete && pRight->u.zToken[1] ){
289 /* If the rhs of the LIKE expression is a variable, and the current
290 ** value of the variable means there is no need to invoke the LIKE
291 ** function, then no OP_Variable will be added to the program.
292 ** This causes problems for the sqlite3_bind_parameter_name()
293 ** API. To work around them, add a dummy OP_Variable here.
295 int r1 = sqlite3GetTempReg(pParse);
296 sqlite3ExprCodeTarget(pParse, pRight, r1);
297 sqlite3VdbeChangeP3(v, sqlite3VdbeCurrentAddr(v)-1, 0);
298 sqlite3ReleaseTempReg(pParse, r1);
301 }else{
302 z = 0;
306 rc = (z!=0);
307 sqlite3ValueFree(pVal);
308 return rc;
310 #endif /* SQLITE_OMIT_LIKE_OPTIMIZATION */
313 #ifndef SQLITE_OMIT_VIRTUALTABLE
315 ** Check to see if the pExpr expression is a form that needs to be passed
316 ** to the xBestIndex method of virtual tables. Forms of interest include:
318 ** Expression Virtual Table Operator
319 ** ----------------------- ---------------------------------
320 ** 1. column MATCH expr SQLITE_INDEX_CONSTRAINT_MATCH
321 ** 2. column GLOB expr SQLITE_INDEX_CONSTRAINT_GLOB
322 ** 3. column LIKE expr SQLITE_INDEX_CONSTRAINT_LIKE
323 ** 4. column REGEXP expr SQLITE_INDEX_CONSTRAINT_REGEXP
324 ** 5. column != expr SQLITE_INDEX_CONSTRAINT_NE
325 ** 6. expr != column SQLITE_INDEX_CONSTRAINT_NE
326 ** 7. column IS NOT expr SQLITE_INDEX_CONSTRAINT_ISNOT
327 ** 8. expr IS NOT column SQLITE_INDEX_CONSTRAINT_ISNOT
328 ** 9. column IS NOT NULL SQLITE_INDEX_CONSTRAINT_ISNOTNULL
330 ** In every case, "column" must be a column of a virtual table. If there
331 ** is a match, set *ppLeft to the "column" expression, set *ppRight to the
332 ** "expr" expression (even though in forms (6) and (8) the column is on the
333 ** right and the expression is on the left). Also set *peOp2 to the
334 ** appropriate virtual table operator. The return value is 1 or 2 if there
335 ** is a match. The usual return is 1, but if the RHS is also a column
336 ** of virtual table in forms (5) or (7) then return 2.
338 ** If the expression matches none of the patterns above, return 0.
340 static int isAuxiliaryVtabOperator(
341 Expr *pExpr, /* Test this expression */
342 unsigned char *peOp2, /* OUT: 0 for MATCH, or else an op2 value */
343 Expr **ppLeft, /* Column expression to left of MATCH/op2 */
344 Expr **ppRight /* Expression to left of MATCH/op2 */
346 if( pExpr->op==TK_FUNCTION ){
347 static const struct Op2 {
348 const char *zOp;
349 unsigned char eOp2;
350 } aOp[] = {
351 { "match", SQLITE_INDEX_CONSTRAINT_MATCH },
352 { "glob", SQLITE_INDEX_CONSTRAINT_GLOB },
353 { "like", SQLITE_INDEX_CONSTRAINT_LIKE },
354 { "regexp", SQLITE_INDEX_CONSTRAINT_REGEXP }
356 ExprList *pList;
357 Expr *pCol; /* Column reference */
358 int i;
360 pList = pExpr->x.pList;
361 if( pList==0 || pList->nExpr!=2 ){
362 return 0;
364 pCol = pList->a[1].pExpr;
365 if( pCol->op!=TK_COLUMN || !IsVirtual(pCol->pTab) ){
366 return 0;
368 for(i=0; i<ArraySize(aOp); i++){
369 if( sqlite3StrICmp(pExpr->u.zToken, aOp[i].zOp)==0 ){
370 *peOp2 = aOp[i].eOp2;
371 *ppRight = pList->a[0].pExpr;
372 *ppLeft = pCol;
373 return 1;
376 }else if( pExpr->op==TK_NE || pExpr->op==TK_ISNOT || pExpr->op==TK_NOTNULL ){
377 int res = 0;
378 Expr *pLeft = pExpr->pLeft;
379 Expr *pRight = pExpr->pRight;
380 if( pLeft->op==TK_COLUMN && IsVirtual(pLeft->pTab) ){
381 res++;
383 if( pRight && pRight->op==TK_COLUMN && IsVirtual(pRight->pTab) ){
384 res++;
385 SWAP(Expr*, pLeft, pRight);
387 *ppLeft = pLeft;
388 *ppRight = pRight;
389 if( pExpr->op==TK_NE ) *peOp2 = SQLITE_INDEX_CONSTRAINT_NE;
390 if( pExpr->op==TK_ISNOT ) *peOp2 = SQLITE_INDEX_CONSTRAINT_ISNOT;
391 if( pExpr->op==TK_NOTNULL ) *peOp2 = SQLITE_INDEX_CONSTRAINT_ISNOTNULL;
392 return res;
394 return 0;
396 #endif /* SQLITE_OMIT_VIRTUALTABLE */
399 ** If the pBase expression originated in the ON or USING clause of
400 ** a join, then transfer the appropriate markings over to derived.
402 static void transferJoinMarkings(Expr *pDerived, Expr *pBase){
403 if( pDerived ){
404 pDerived->flags |= pBase->flags & EP_FromJoin;
405 pDerived->iRightJoinTable = pBase->iRightJoinTable;
410 ** Mark term iChild as being a child of term iParent
412 static void markTermAsChild(WhereClause *pWC, int iChild, int iParent){
413 pWC->a[iChild].iParent = iParent;
414 pWC->a[iChild].truthProb = pWC->a[iParent].truthProb;
415 pWC->a[iParent].nChild++;
419 ** Return the N-th AND-connected subterm of pTerm. Or if pTerm is not
420 ** a conjunction, then return just pTerm when N==0. If N is exceeds
421 ** the number of available subterms, return NULL.
423 static WhereTerm *whereNthSubterm(WhereTerm *pTerm, int N){
424 if( pTerm->eOperator!=WO_AND ){
425 return N==0 ? pTerm : 0;
427 if( N<pTerm->u.pAndInfo->wc.nTerm ){
428 return &pTerm->u.pAndInfo->wc.a[N];
430 return 0;
434 ** Subterms pOne and pTwo are contained within WHERE clause pWC. The
435 ** two subterms are in disjunction - they are OR-ed together.
437 ** If these two terms are both of the form: "A op B" with the same
438 ** A and B values but different operators and if the operators are
439 ** compatible (if one is = and the other is <, for example) then
440 ** add a new virtual AND term to pWC that is the combination of the
441 ** two.
443 ** Some examples:
445 ** x<y OR x=y --> x<=y
446 ** x=y OR x=y --> x=y
447 ** x<=y OR x<y --> x<=y
449 ** The following is NOT generated:
451 ** x<y OR x>y --> x!=y
453 static void whereCombineDisjuncts(
454 SrcList *pSrc, /* the FROM clause */
455 WhereClause *pWC, /* The complete WHERE clause */
456 WhereTerm *pOne, /* First disjunct */
457 WhereTerm *pTwo /* Second disjunct */
459 u16 eOp = pOne->eOperator | pTwo->eOperator;
460 sqlite3 *db; /* Database connection (for malloc) */
461 Expr *pNew; /* New virtual expression */
462 int op; /* Operator for the combined expression */
463 int idxNew; /* Index in pWC of the next virtual term */
465 if( (pOne->eOperator & (WO_EQ|WO_LT|WO_LE|WO_GT|WO_GE))==0 ) return;
466 if( (pTwo->eOperator & (WO_EQ|WO_LT|WO_LE|WO_GT|WO_GE))==0 ) return;
467 if( (eOp & (WO_EQ|WO_LT|WO_LE))!=eOp
468 && (eOp & (WO_EQ|WO_GT|WO_GE))!=eOp ) return;
469 assert( pOne->pExpr->pLeft!=0 && pOne->pExpr->pRight!=0 );
470 assert( pTwo->pExpr->pLeft!=0 && pTwo->pExpr->pRight!=0 );
471 if( sqlite3ExprCompare(0,pOne->pExpr->pLeft, pTwo->pExpr->pLeft, -1) ) return;
472 if( sqlite3ExprCompare(0,pOne->pExpr->pRight, pTwo->pExpr->pRight,-1) )return;
473 /* If we reach this point, it means the two subterms can be combined */
474 if( (eOp & (eOp-1))!=0 ){
475 if( eOp & (WO_LT|WO_LE) ){
476 eOp = WO_LE;
477 }else{
478 assert( eOp & (WO_GT|WO_GE) );
479 eOp = WO_GE;
482 db = pWC->pWInfo->pParse->db;
483 pNew = sqlite3ExprDup(db, pOne->pExpr, 0);
484 if( pNew==0 ) return;
485 for(op=TK_EQ; eOp!=(WO_EQ<<(op-TK_EQ)); op++){ assert( op<TK_GE ); }
486 pNew->op = op;
487 idxNew = whereClauseInsert(pWC, pNew, TERM_VIRTUAL|TERM_DYNAMIC);
488 exprAnalyze(pSrc, pWC, idxNew);
491 #if !defined(SQLITE_OMIT_OR_OPTIMIZATION) && !defined(SQLITE_OMIT_SUBQUERY)
493 ** Analyze a term that consists of two or more OR-connected
494 ** subterms. So in:
496 ** ... WHERE (a=5) AND (b=7 OR c=9 OR d=13) AND (d=13)
497 ** ^^^^^^^^^^^^^^^^^^^^
499 ** This routine analyzes terms such as the middle term in the above example.
500 ** A WhereOrTerm object is computed and attached to the term under
501 ** analysis, regardless of the outcome of the analysis. Hence:
503 ** WhereTerm.wtFlags |= TERM_ORINFO
504 ** WhereTerm.u.pOrInfo = a dynamically allocated WhereOrTerm object
506 ** The term being analyzed must have two or more of OR-connected subterms.
507 ** A single subterm might be a set of AND-connected sub-subterms.
508 ** Examples of terms under analysis:
510 ** (A) t1.x=t2.y OR t1.x=t2.z OR t1.y=15 OR t1.z=t3.a+5
511 ** (B) x=expr1 OR expr2=x OR x=expr3
512 ** (C) t1.x=t2.y OR (t1.x=t2.z AND t1.y=15)
513 ** (D) x=expr1 OR (y>11 AND y<22 AND z LIKE '*hello*')
514 ** (E) (p.a=1 AND q.b=2 AND r.c=3) OR (p.x=4 AND q.y=5 AND r.z=6)
515 ** (F) x>A OR (x=A AND y>=B)
517 ** CASE 1:
519 ** If all subterms are of the form T.C=expr for some single column of C and
520 ** a single table T (as shown in example B above) then create a new virtual
521 ** term that is an equivalent IN expression. In other words, if the term
522 ** being analyzed is:
524 ** x = expr1 OR expr2 = x OR x = expr3
526 ** then create a new virtual term like this:
528 ** x IN (expr1,expr2,expr3)
530 ** CASE 2:
532 ** If there are exactly two disjuncts and one side has x>A and the other side
533 ** has x=A (for the same x and A) then add a new virtual conjunct term to the
534 ** WHERE clause of the form "x>=A". Example:
536 ** x>A OR (x=A AND y>B) adds: x>=A
538 ** The added conjunct can sometimes be helpful in query planning.
540 ** CASE 3:
542 ** If all subterms are indexable by a single table T, then set
544 ** WhereTerm.eOperator = WO_OR
545 ** WhereTerm.u.pOrInfo->indexable |= the cursor number for table T
547 ** A subterm is "indexable" if it is of the form
548 ** "T.C <op> <expr>" where C is any column of table T and
549 ** <op> is one of "=", "<", "<=", ">", ">=", "IS NULL", or "IN".
550 ** A subterm is also indexable if it is an AND of two or more
551 ** subsubterms at least one of which is indexable. Indexable AND
552 ** subterms have their eOperator set to WO_AND and they have
553 ** u.pAndInfo set to a dynamically allocated WhereAndTerm object.
555 ** From another point of view, "indexable" means that the subterm could
556 ** potentially be used with an index if an appropriate index exists.
557 ** This analysis does not consider whether or not the index exists; that
558 ** is decided elsewhere. This analysis only looks at whether subterms
559 ** appropriate for indexing exist.
561 ** All examples A through E above satisfy case 3. But if a term
562 ** also satisfies case 1 (such as B) we know that the optimizer will
563 ** always prefer case 1, so in that case we pretend that case 3 is not
564 ** satisfied.
566 ** It might be the case that multiple tables are indexable. For example,
567 ** (E) above is indexable on tables P, Q, and R.
569 ** Terms that satisfy case 3 are candidates for lookup by using
570 ** separate indices to find rowids for each subterm and composing
571 ** the union of all rowids using a RowSet object. This is similar
572 ** to "bitmap indices" in other database engines.
574 ** OTHERWISE:
576 ** If none of cases 1, 2, or 3 apply, then leave the eOperator set to
577 ** zero. This term is not useful for search.
579 static void exprAnalyzeOrTerm(
580 SrcList *pSrc, /* the FROM clause */
581 WhereClause *pWC, /* the complete WHERE clause */
582 int idxTerm /* Index of the OR-term to be analyzed */
584 WhereInfo *pWInfo = pWC->pWInfo; /* WHERE clause processing context */
585 Parse *pParse = pWInfo->pParse; /* Parser context */
586 sqlite3 *db = pParse->db; /* Database connection */
587 WhereTerm *pTerm = &pWC->a[idxTerm]; /* The term to be analyzed */
588 Expr *pExpr = pTerm->pExpr; /* The expression of the term */
589 int i; /* Loop counters */
590 WhereClause *pOrWc; /* Breakup of pTerm into subterms */
591 WhereTerm *pOrTerm; /* A Sub-term within the pOrWc */
592 WhereOrInfo *pOrInfo; /* Additional information associated with pTerm */
593 Bitmask chngToIN; /* Tables that might satisfy case 1 */
594 Bitmask indexable; /* Tables that are indexable, satisfying case 2 */
597 ** Break the OR clause into its separate subterms. The subterms are
598 ** stored in a WhereClause structure containing within the WhereOrInfo
599 ** object that is attached to the original OR clause term.
601 assert( (pTerm->wtFlags & (TERM_DYNAMIC|TERM_ORINFO|TERM_ANDINFO))==0 );
602 assert( pExpr->op==TK_OR );
603 pTerm->u.pOrInfo = pOrInfo = sqlite3DbMallocZero(db, sizeof(*pOrInfo));
604 if( pOrInfo==0 ) return;
605 pTerm->wtFlags |= TERM_ORINFO;
606 pOrWc = &pOrInfo->wc;
607 memset(pOrWc->aStatic, 0, sizeof(pOrWc->aStatic));
608 sqlite3WhereClauseInit(pOrWc, pWInfo);
609 sqlite3WhereSplit(pOrWc, pExpr, TK_OR);
610 sqlite3WhereExprAnalyze(pSrc, pOrWc);
611 if( db->mallocFailed ) return;
612 assert( pOrWc->nTerm>=2 );
615 ** Compute the set of tables that might satisfy cases 1 or 3.
617 indexable = ~(Bitmask)0;
618 chngToIN = ~(Bitmask)0;
619 for(i=pOrWc->nTerm-1, pOrTerm=pOrWc->a; i>=0 && indexable; i--, pOrTerm++){
620 if( (pOrTerm->eOperator & WO_SINGLE)==0 ){
621 WhereAndInfo *pAndInfo;
622 assert( (pOrTerm->wtFlags & (TERM_ANDINFO|TERM_ORINFO))==0 );
623 chngToIN = 0;
624 pAndInfo = sqlite3DbMallocRawNN(db, sizeof(*pAndInfo));
625 if( pAndInfo ){
626 WhereClause *pAndWC;
627 WhereTerm *pAndTerm;
628 int j;
629 Bitmask b = 0;
630 pOrTerm->u.pAndInfo = pAndInfo;
631 pOrTerm->wtFlags |= TERM_ANDINFO;
632 pOrTerm->eOperator = WO_AND;
633 pAndWC = &pAndInfo->wc;
634 memset(pAndWC->aStatic, 0, sizeof(pAndWC->aStatic));
635 sqlite3WhereClauseInit(pAndWC, pWC->pWInfo);
636 sqlite3WhereSplit(pAndWC, pOrTerm->pExpr, TK_AND);
637 sqlite3WhereExprAnalyze(pSrc, pAndWC);
638 pAndWC->pOuter = pWC;
639 if( !db->mallocFailed ){
640 for(j=0, pAndTerm=pAndWC->a; j<pAndWC->nTerm; j++, pAndTerm++){
641 assert( pAndTerm->pExpr );
642 if( allowedOp(pAndTerm->pExpr->op)
643 || pAndTerm->eOperator==WO_AUX
645 b |= sqlite3WhereGetMask(&pWInfo->sMaskSet, pAndTerm->leftCursor);
649 indexable &= b;
651 }else if( pOrTerm->wtFlags & TERM_COPIED ){
652 /* Skip this term for now. We revisit it when we process the
653 ** corresponding TERM_VIRTUAL term */
654 }else{
655 Bitmask b;
656 b = sqlite3WhereGetMask(&pWInfo->sMaskSet, pOrTerm->leftCursor);
657 if( pOrTerm->wtFlags & TERM_VIRTUAL ){
658 WhereTerm *pOther = &pOrWc->a[pOrTerm->iParent];
659 b |= sqlite3WhereGetMask(&pWInfo->sMaskSet, pOther->leftCursor);
661 indexable &= b;
662 if( (pOrTerm->eOperator & WO_EQ)==0 ){
663 chngToIN = 0;
664 }else{
665 chngToIN &= b;
671 ** Record the set of tables that satisfy case 3. The set might be
672 ** empty.
674 pOrInfo->indexable = indexable;
675 pTerm->eOperator = indexable==0 ? 0 : WO_OR;
677 /* For a two-way OR, attempt to implementation case 2.
679 if( indexable && pOrWc->nTerm==2 ){
680 int iOne = 0;
681 WhereTerm *pOne;
682 while( (pOne = whereNthSubterm(&pOrWc->a[0],iOne++))!=0 ){
683 int iTwo = 0;
684 WhereTerm *pTwo;
685 while( (pTwo = whereNthSubterm(&pOrWc->a[1],iTwo++))!=0 ){
686 whereCombineDisjuncts(pSrc, pWC, pOne, pTwo);
692 ** chngToIN holds a set of tables that *might* satisfy case 1. But
693 ** we have to do some additional checking to see if case 1 really
694 ** is satisfied.
696 ** chngToIN will hold either 0, 1, or 2 bits. The 0-bit case means
697 ** that there is no possibility of transforming the OR clause into an
698 ** IN operator because one or more terms in the OR clause contain
699 ** something other than == on a column in the single table. The 1-bit
700 ** case means that every term of the OR clause is of the form
701 ** "table.column=expr" for some single table. The one bit that is set
702 ** will correspond to the common table. We still need to check to make
703 ** sure the same column is used on all terms. The 2-bit case is when
704 ** the all terms are of the form "table1.column=table2.column". It
705 ** might be possible to form an IN operator with either table1.column
706 ** or table2.column as the LHS if either is common to every term of
707 ** the OR clause.
709 ** Note that terms of the form "table.column1=table.column2" (the
710 ** same table on both sizes of the ==) cannot be optimized.
712 if( chngToIN ){
713 int okToChngToIN = 0; /* True if the conversion to IN is valid */
714 int iColumn = -1; /* Column index on lhs of IN operator */
715 int iCursor = -1; /* Table cursor common to all terms */
716 int j = 0; /* Loop counter */
718 /* Search for a table and column that appears on one side or the
719 ** other of the == operator in every subterm. That table and column
720 ** will be recorded in iCursor and iColumn. There might not be any
721 ** such table and column. Set okToChngToIN if an appropriate table
722 ** and column is found but leave okToChngToIN false if not found.
724 for(j=0; j<2 && !okToChngToIN; j++){
725 pOrTerm = pOrWc->a;
726 for(i=pOrWc->nTerm-1; i>=0; i--, pOrTerm++){
727 assert( pOrTerm->eOperator & WO_EQ );
728 pOrTerm->wtFlags &= ~TERM_OR_OK;
729 if( pOrTerm->leftCursor==iCursor ){
730 /* This is the 2-bit case and we are on the second iteration and
731 ** current term is from the first iteration. So skip this term. */
732 assert( j==1 );
733 continue;
735 if( (chngToIN & sqlite3WhereGetMask(&pWInfo->sMaskSet,
736 pOrTerm->leftCursor))==0 ){
737 /* This term must be of the form t1.a==t2.b where t2 is in the
738 ** chngToIN set but t1 is not. This term will be either preceded
739 ** or follwed by an inverted copy (t2.b==t1.a). Skip this term
740 ** and use its inversion. */
741 testcase( pOrTerm->wtFlags & TERM_COPIED );
742 testcase( pOrTerm->wtFlags & TERM_VIRTUAL );
743 assert( pOrTerm->wtFlags & (TERM_COPIED|TERM_VIRTUAL) );
744 continue;
746 iColumn = pOrTerm->u.leftColumn;
747 iCursor = pOrTerm->leftCursor;
748 break;
750 if( i<0 ){
751 /* No candidate table+column was found. This can only occur
752 ** on the second iteration */
753 assert( j==1 );
754 assert( IsPowerOfTwo(chngToIN) );
755 assert( chngToIN==sqlite3WhereGetMask(&pWInfo->sMaskSet, iCursor) );
756 break;
758 testcase( j==1 );
760 /* We have found a candidate table and column. Check to see if that
761 ** table and column is common to every term in the OR clause */
762 okToChngToIN = 1;
763 for(; i>=0 && okToChngToIN; i--, pOrTerm++){
764 assert( pOrTerm->eOperator & WO_EQ );
765 if( pOrTerm->leftCursor!=iCursor ){
766 pOrTerm->wtFlags &= ~TERM_OR_OK;
767 }else if( pOrTerm->u.leftColumn!=iColumn ){
768 okToChngToIN = 0;
769 }else{
770 int affLeft, affRight;
771 /* If the right-hand side is also a column, then the affinities
772 ** of both right and left sides must be such that no type
773 ** conversions are required on the right. (Ticket #2249)
775 affRight = sqlite3ExprAffinity(pOrTerm->pExpr->pRight);
776 affLeft = sqlite3ExprAffinity(pOrTerm->pExpr->pLeft);
777 if( affRight!=0 && affRight!=affLeft ){
778 okToChngToIN = 0;
779 }else{
780 pOrTerm->wtFlags |= TERM_OR_OK;
786 /* At this point, okToChngToIN is true if original pTerm satisfies
787 ** case 1. In that case, construct a new virtual term that is
788 ** pTerm converted into an IN operator.
790 if( okToChngToIN ){
791 Expr *pDup; /* A transient duplicate expression */
792 ExprList *pList = 0; /* The RHS of the IN operator */
793 Expr *pLeft = 0; /* The LHS of the IN operator */
794 Expr *pNew; /* The complete IN operator */
796 for(i=pOrWc->nTerm-1, pOrTerm=pOrWc->a; i>=0; i--, pOrTerm++){
797 if( (pOrTerm->wtFlags & TERM_OR_OK)==0 ) continue;
798 assert( pOrTerm->eOperator & WO_EQ );
799 assert( pOrTerm->leftCursor==iCursor );
800 assert( pOrTerm->u.leftColumn==iColumn );
801 pDup = sqlite3ExprDup(db, pOrTerm->pExpr->pRight, 0);
802 pList = sqlite3ExprListAppend(pWInfo->pParse, pList, pDup);
803 pLeft = pOrTerm->pExpr->pLeft;
805 assert( pLeft!=0 );
806 pDup = sqlite3ExprDup(db, pLeft, 0);
807 pNew = sqlite3PExpr(pParse, TK_IN, pDup, 0);
808 if( pNew ){
809 int idxNew;
810 transferJoinMarkings(pNew, pExpr);
811 assert( !ExprHasProperty(pNew, EP_xIsSelect) );
812 pNew->x.pList = pList;
813 idxNew = whereClauseInsert(pWC, pNew, TERM_VIRTUAL|TERM_DYNAMIC);
814 testcase( idxNew==0 );
815 exprAnalyze(pSrc, pWC, idxNew);
816 pTerm = &pWC->a[idxTerm];
817 markTermAsChild(pWC, idxNew, idxTerm);
818 }else{
819 sqlite3ExprListDelete(db, pList);
821 pTerm->eOperator = WO_NOOP; /* case 1 trumps case 3 */
825 #endif /* !SQLITE_OMIT_OR_OPTIMIZATION && !SQLITE_OMIT_SUBQUERY */
828 ** We already know that pExpr is a binary operator where both operands are
829 ** column references. This routine checks to see if pExpr is an equivalence
830 ** relation:
831 ** 1. The SQLITE_Transitive optimization must be enabled
832 ** 2. Must be either an == or an IS operator
833 ** 3. Not originating in the ON clause of an OUTER JOIN
834 ** 4. The affinities of A and B must be compatible
835 ** 5a. Both operands use the same collating sequence OR
836 ** 5b. The overall collating sequence is BINARY
837 ** If this routine returns TRUE, that means that the RHS can be substituted
838 ** for the LHS anyplace else in the WHERE clause where the LHS column occurs.
839 ** This is an optimization. No harm comes from returning 0. But if 1 is
840 ** returned when it should not be, then incorrect answers might result.
842 static int termIsEquivalence(Parse *pParse, Expr *pExpr){
843 char aff1, aff2;
844 CollSeq *pColl;
845 if( !OptimizationEnabled(pParse->db, SQLITE_Transitive) ) return 0;
846 if( pExpr->op!=TK_EQ && pExpr->op!=TK_IS ) return 0;
847 if( ExprHasProperty(pExpr, EP_FromJoin) ) return 0;
848 aff1 = sqlite3ExprAffinity(pExpr->pLeft);
849 aff2 = sqlite3ExprAffinity(pExpr->pRight);
850 if( aff1!=aff2
851 && (!sqlite3IsNumericAffinity(aff1) || !sqlite3IsNumericAffinity(aff2))
853 return 0;
855 pColl = sqlite3BinaryCompareCollSeq(pParse, pExpr->pLeft, pExpr->pRight);
856 if( pColl==0 || sqlite3StrICmp(pColl->zName, "BINARY")==0 ) return 1;
857 return sqlite3ExprCollSeqMatch(pParse, pExpr->pLeft, pExpr->pRight);
861 ** Recursively walk the expressions of a SELECT statement and generate
862 ** a bitmask indicating which tables are used in that expression
863 ** tree.
865 static Bitmask exprSelectUsage(WhereMaskSet *pMaskSet, Select *pS){
866 Bitmask mask = 0;
867 while( pS ){
868 SrcList *pSrc = pS->pSrc;
869 mask |= sqlite3WhereExprListUsage(pMaskSet, pS->pEList);
870 mask |= sqlite3WhereExprListUsage(pMaskSet, pS->pGroupBy);
871 mask |= sqlite3WhereExprListUsage(pMaskSet, pS->pOrderBy);
872 mask |= sqlite3WhereExprUsage(pMaskSet, pS->pWhere);
873 mask |= sqlite3WhereExprUsage(pMaskSet, pS->pHaving);
874 if( ALWAYS(pSrc!=0) ){
875 int i;
876 for(i=0; i<pSrc->nSrc; i++){
877 mask |= exprSelectUsage(pMaskSet, pSrc->a[i].pSelect);
878 mask |= sqlite3WhereExprUsage(pMaskSet, pSrc->a[i].pOn);
879 if( pSrc->a[i].fg.isTabFunc ){
880 mask |= sqlite3WhereExprListUsage(pMaskSet, pSrc->a[i].u1.pFuncArg);
884 pS = pS->pPrior;
886 return mask;
890 ** Expression pExpr is one operand of a comparison operator that might
891 ** be useful for indexing. This routine checks to see if pExpr appears
892 ** in any index. Return TRUE (1) if pExpr is an indexed term and return
893 ** FALSE (0) if not. If TRUE is returned, also set aiCurCol[0] to the cursor
894 ** number of the table that is indexed and aiCurCol[1] to the column number
895 ** of the column that is indexed, or XN_EXPR (-2) if an expression is being
896 ** indexed.
898 ** If pExpr is a TK_COLUMN column reference, then this routine always returns
899 ** true even if that particular column is not indexed, because the column
900 ** might be added to an automatic index later.
902 static SQLITE_NOINLINE int exprMightBeIndexed2(
903 SrcList *pFrom, /* The FROM clause */
904 Bitmask mPrereq, /* Bitmask of FROM clause terms referenced by pExpr */
905 int *aiCurCol, /* Write the referenced table cursor and column here */
906 Expr *pExpr /* An operand of a comparison operator */
908 Index *pIdx;
909 int i;
910 int iCur;
911 for(i=0; mPrereq>1; i++, mPrereq>>=1){}
912 iCur = pFrom->a[i].iCursor;
913 for(pIdx=pFrom->a[i].pTab->pIndex; pIdx; pIdx=pIdx->pNext){
914 if( pIdx->aColExpr==0 ) continue;
915 for(i=0; i<pIdx->nKeyCol; i++){
916 if( pIdx->aiColumn[i]!=XN_EXPR ) continue;
917 if( sqlite3ExprCompareSkip(pExpr, pIdx->aColExpr->a[i].pExpr, iCur)==0 ){
918 aiCurCol[0] = iCur;
919 aiCurCol[1] = XN_EXPR;
920 return 1;
924 return 0;
926 static int exprMightBeIndexed(
927 SrcList *pFrom, /* The FROM clause */
928 Bitmask mPrereq, /* Bitmask of FROM clause terms referenced by pExpr */
929 int *aiCurCol, /* Write the referenced table cursor & column here */
930 Expr *pExpr, /* An operand of a comparison operator */
931 int op /* The specific comparison operator */
933 /* If this expression is a vector to the left or right of a
934 ** inequality constraint (>, <, >= or <=), perform the processing
935 ** on the first element of the vector. */
936 assert( TK_GT+1==TK_LE && TK_GT+2==TK_LT && TK_GT+3==TK_GE );
937 assert( TK_IS<TK_GE && TK_ISNULL<TK_GE && TK_IN<TK_GE );
938 assert( op<=TK_GE );
939 if( pExpr->op==TK_VECTOR && (op>=TK_GT && ALWAYS(op<=TK_GE)) ){
940 pExpr = pExpr->x.pList->a[0].pExpr;
943 if( pExpr->op==TK_COLUMN ){
944 aiCurCol[0] = pExpr->iTable;
945 aiCurCol[1] = pExpr->iColumn;
946 return 1;
948 if( mPrereq==0 ) return 0; /* No table references */
949 if( (mPrereq&(mPrereq-1))!=0 ) return 0; /* Refs more than one table */
950 return exprMightBeIndexed2(pFrom,mPrereq,aiCurCol,pExpr);
954 ** The input to this routine is an WhereTerm structure with only the
955 ** "pExpr" field filled in. The job of this routine is to analyze the
956 ** subexpression and populate all the other fields of the WhereTerm
957 ** structure.
959 ** If the expression is of the form "<expr> <op> X" it gets commuted
960 ** to the standard form of "X <op> <expr>".
962 ** If the expression is of the form "X <op> Y" where both X and Y are
963 ** columns, then the original expression is unchanged and a new virtual
964 ** term of the form "Y <op> X" is added to the WHERE clause and
965 ** analyzed separately. The original term is marked with TERM_COPIED
966 ** and the new term is marked with TERM_DYNAMIC (because it's pExpr
967 ** needs to be freed with the WhereClause) and TERM_VIRTUAL (because it
968 ** is a commuted copy of a prior term.) The original term has nChild=1
969 ** and the copy has idxParent set to the index of the original term.
971 static void exprAnalyze(
972 SrcList *pSrc, /* the FROM clause */
973 WhereClause *pWC, /* the WHERE clause */
974 int idxTerm /* Index of the term to be analyzed */
976 WhereInfo *pWInfo = pWC->pWInfo; /* WHERE clause processing context */
977 WhereTerm *pTerm; /* The term to be analyzed */
978 WhereMaskSet *pMaskSet; /* Set of table index masks */
979 Expr *pExpr; /* The expression to be analyzed */
980 Bitmask prereqLeft; /* Prerequesites of the pExpr->pLeft */
981 Bitmask prereqAll; /* Prerequesites of pExpr */
982 Bitmask extraRight = 0; /* Extra dependencies on LEFT JOIN */
983 Expr *pStr1 = 0; /* RHS of LIKE/GLOB operator */
984 int isComplete = 0; /* RHS of LIKE/GLOB ends with wildcard */
985 int noCase = 0; /* uppercase equivalent to lowercase */
986 int op; /* Top-level operator. pExpr->op */
987 Parse *pParse = pWInfo->pParse; /* Parsing context */
988 sqlite3 *db = pParse->db; /* Database connection */
989 unsigned char eOp2 = 0; /* op2 value for LIKE/REGEXP/GLOB */
990 int nLeft; /* Number of elements on left side vector */
992 if( db->mallocFailed ){
993 return;
995 pTerm = &pWC->a[idxTerm];
996 pMaskSet = &pWInfo->sMaskSet;
997 pExpr = pTerm->pExpr;
998 assert( pExpr->op!=TK_AS && pExpr->op!=TK_COLLATE );
999 prereqLeft = sqlite3WhereExprUsage(pMaskSet, pExpr->pLeft);
1000 op = pExpr->op;
1001 if( op==TK_IN ){
1002 assert( pExpr->pRight==0 );
1003 if( sqlite3ExprCheckIN(pParse, pExpr) ) return;
1004 if( ExprHasProperty(pExpr, EP_xIsSelect) ){
1005 pTerm->prereqRight = exprSelectUsage(pMaskSet, pExpr->x.pSelect);
1006 }else{
1007 pTerm->prereqRight = sqlite3WhereExprListUsage(pMaskSet, pExpr->x.pList);
1009 }else if( op==TK_ISNULL ){
1010 pTerm->prereqRight = 0;
1011 }else{
1012 pTerm->prereqRight = sqlite3WhereExprUsage(pMaskSet, pExpr->pRight);
1014 pMaskSet->bVarSelect = 0;
1015 prereqAll = sqlite3WhereExprUsage(pMaskSet, pExpr);
1016 if( pMaskSet->bVarSelect ) pTerm->wtFlags |= TERM_VARSELECT;
1017 if( ExprHasProperty(pExpr, EP_FromJoin) ){
1018 Bitmask x = sqlite3WhereGetMask(pMaskSet, pExpr->iRightJoinTable);
1019 prereqAll |= x;
1020 extraRight = x-1; /* ON clause terms may not be used with an index
1021 ** on left table of a LEFT JOIN. Ticket #3015 */
1022 if( (prereqAll>>1)>=x ){
1023 sqlite3ErrorMsg(pParse, "ON clause references tables to its right");
1024 return;
1027 pTerm->prereqAll = prereqAll;
1028 pTerm->leftCursor = -1;
1029 pTerm->iParent = -1;
1030 pTerm->eOperator = 0;
1031 if( allowedOp(op) ){
1032 int aiCurCol[2];
1033 Expr *pLeft = sqlite3ExprSkipCollate(pExpr->pLeft);
1034 Expr *pRight = sqlite3ExprSkipCollate(pExpr->pRight);
1035 u16 opMask = (pTerm->prereqRight & prereqLeft)==0 ? WO_ALL : WO_EQUIV;
1037 if( pTerm->iField>0 ){
1038 assert( op==TK_IN );
1039 assert( pLeft->op==TK_VECTOR );
1040 pLeft = pLeft->x.pList->a[pTerm->iField-1].pExpr;
1043 if( exprMightBeIndexed(pSrc, prereqLeft, aiCurCol, pLeft, op) ){
1044 pTerm->leftCursor = aiCurCol[0];
1045 pTerm->u.leftColumn = aiCurCol[1];
1046 pTerm->eOperator = operatorMask(op) & opMask;
1048 if( op==TK_IS ) pTerm->wtFlags |= TERM_IS;
1049 if( pRight
1050 && exprMightBeIndexed(pSrc, pTerm->prereqRight, aiCurCol, pRight, op)
1052 WhereTerm *pNew;
1053 Expr *pDup;
1054 u16 eExtraOp = 0; /* Extra bits for pNew->eOperator */
1055 assert( pTerm->iField==0 );
1056 if( pTerm->leftCursor>=0 ){
1057 int idxNew;
1058 pDup = sqlite3ExprDup(db, pExpr, 0);
1059 if( db->mallocFailed ){
1060 sqlite3ExprDelete(db, pDup);
1061 return;
1063 idxNew = whereClauseInsert(pWC, pDup, TERM_VIRTUAL|TERM_DYNAMIC);
1064 if( idxNew==0 ) return;
1065 pNew = &pWC->a[idxNew];
1066 markTermAsChild(pWC, idxNew, idxTerm);
1067 if( op==TK_IS ) pNew->wtFlags |= TERM_IS;
1068 pTerm = &pWC->a[idxTerm];
1069 pTerm->wtFlags |= TERM_COPIED;
1071 if( termIsEquivalence(pParse, pDup) ){
1072 pTerm->eOperator |= WO_EQUIV;
1073 eExtraOp = WO_EQUIV;
1075 }else{
1076 pDup = pExpr;
1077 pNew = pTerm;
1079 exprCommute(pParse, pDup);
1080 pNew->leftCursor = aiCurCol[0];
1081 pNew->u.leftColumn = aiCurCol[1];
1082 testcase( (prereqLeft | extraRight) != prereqLeft );
1083 pNew->prereqRight = prereqLeft | extraRight;
1084 pNew->prereqAll = prereqAll;
1085 pNew->eOperator = (operatorMask(pDup->op) + eExtraOp) & opMask;
1089 #ifndef SQLITE_OMIT_BETWEEN_OPTIMIZATION
1090 /* If a term is the BETWEEN operator, create two new virtual terms
1091 ** that define the range that the BETWEEN implements. For example:
1093 ** a BETWEEN b AND c
1095 ** is converted into:
1097 ** (a BETWEEN b AND c) AND (a>=b) AND (a<=c)
1099 ** The two new terms are added onto the end of the WhereClause object.
1100 ** The new terms are "dynamic" and are children of the original BETWEEN
1101 ** term. That means that if the BETWEEN term is coded, the children are
1102 ** skipped. Or, if the children are satisfied by an index, the original
1103 ** BETWEEN term is skipped.
1105 else if( pExpr->op==TK_BETWEEN && pWC->op==TK_AND ){
1106 ExprList *pList = pExpr->x.pList;
1107 int i;
1108 static const u8 ops[] = {TK_GE, TK_LE};
1109 assert( pList!=0 );
1110 assert( pList->nExpr==2 );
1111 for(i=0; i<2; i++){
1112 Expr *pNewExpr;
1113 int idxNew;
1114 pNewExpr = sqlite3PExpr(pParse, ops[i],
1115 sqlite3ExprDup(db, pExpr->pLeft, 0),
1116 sqlite3ExprDup(db, pList->a[i].pExpr, 0));
1117 transferJoinMarkings(pNewExpr, pExpr);
1118 idxNew = whereClauseInsert(pWC, pNewExpr, TERM_VIRTUAL|TERM_DYNAMIC);
1119 testcase( idxNew==0 );
1120 exprAnalyze(pSrc, pWC, idxNew);
1121 pTerm = &pWC->a[idxTerm];
1122 markTermAsChild(pWC, idxNew, idxTerm);
1125 #endif /* SQLITE_OMIT_BETWEEN_OPTIMIZATION */
1127 #if !defined(SQLITE_OMIT_OR_OPTIMIZATION) && !defined(SQLITE_OMIT_SUBQUERY)
1128 /* Analyze a term that is composed of two or more subterms connected by
1129 ** an OR operator.
1131 else if( pExpr->op==TK_OR ){
1132 assert( pWC->op==TK_AND );
1133 exprAnalyzeOrTerm(pSrc, pWC, idxTerm);
1134 pTerm = &pWC->a[idxTerm];
1136 #endif /* SQLITE_OMIT_OR_OPTIMIZATION */
1138 #ifndef SQLITE_OMIT_LIKE_OPTIMIZATION
1139 /* Add constraints to reduce the search space on a LIKE or GLOB
1140 ** operator.
1142 ** A like pattern of the form "x LIKE 'aBc%'" is changed into constraints
1144 ** x>='ABC' AND x<'abd' AND x LIKE 'aBc%'
1146 ** The last character of the prefix "abc" is incremented to form the
1147 ** termination condition "abd". If case is not significant (the default
1148 ** for LIKE) then the lower-bound is made all uppercase and the upper-
1149 ** bound is made all lowercase so that the bounds also work when comparing
1150 ** BLOBs.
1152 if( pWC->op==TK_AND
1153 && isLikeOrGlob(pParse, pExpr, &pStr1, &isComplete, &noCase)
1155 Expr *pLeft; /* LHS of LIKE/GLOB operator */
1156 Expr *pStr2; /* Copy of pStr1 - RHS of LIKE/GLOB operator */
1157 Expr *pNewExpr1;
1158 Expr *pNewExpr2;
1159 int idxNew1;
1160 int idxNew2;
1161 const char *zCollSeqName; /* Name of collating sequence */
1162 const u16 wtFlags = TERM_LIKEOPT | TERM_VIRTUAL | TERM_DYNAMIC;
1164 pLeft = pExpr->x.pList->a[1].pExpr;
1165 pStr2 = sqlite3ExprDup(db, pStr1, 0);
1167 /* Convert the lower bound to upper-case and the upper bound to
1168 ** lower-case (upper-case is less than lower-case in ASCII) so that
1169 ** the range constraints also work for BLOBs
1171 if( noCase && !pParse->db->mallocFailed ){
1172 int i;
1173 char c;
1174 pTerm->wtFlags |= TERM_LIKE;
1175 for(i=0; (c = pStr1->u.zToken[i])!=0; i++){
1176 pStr1->u.zToken[i] = sqlite3Toupper(c);
1177 pStr2->u.zToken[i] = sqlite3Tolower(c);
1181 if( !db->mallocFailed ){
1182 u8 c, *pC; /* Last character before the first wildcard */
1183 pC = (u8*)&pStr2->u.zToken[sqlite3Strlen30(pStr2->u.zToken)-1];
1184 c = *pC;
1185 if( noCase ){
1186 /* The point is to increment the last character before the first
1187 ** wildcard. But if we increment '@', that will push it into the
1188 ** alphabetic range where case conversions will mess up the
1189 ** inequality. To avoid this, make sure to also run the full
1190 ** LIKE on all candidate expressions by clearing the isComplete flag
1192 if( c=='A'-1 ) isComplete = 0;
1193 c = sqlite3UpperToLower[c];
1195 *pC = c + 1;
1197 zCollSeqName = noCase ? "NOCASE" : "BINARY";
1198 pNewExpr1 = sqlite3ExprDup(db, pLeft, 0);
1199 pNewExpr1 = sqlite3PExpr(pParse, TK_GE,
1200 sqlite3ExprAddCollateString(pParse,pNewExpr1,zCollSeqName),
1201 pStr1);
1202 transferJoinMarkings(pNewExpr1, pExpr);
1203 idxNew1 = whereClauseInsert(pWC, pNewExpr1, wtFlags);
1204 testcase( idxNew1==0 );
1205 exprAnalyze(pSrc, pWC, idxNew1);
1206 pNewExpr2 = sqlite3ExprDup(db, pLeft, 0);
1207 pNewExpr2 = sqlite3PExpr(pParse, TK_LT,
1208 sqlite3ExprAddCollateString(pParse,pNewExpr2,zCollSeqName),
1209 pStr2);
1210 transferJoinMarkings(pNewExpr2, pExpr);
1211 idxNew2 = whereClauseInsert(pWC, pNewExpr2, wtFlags);
1212 testcase( idxNew2==0 );
1213 exprAnalyze(pSrc, pWC, idxNew2);
1214 pTerm = &pWC->a[idxTerm];
1215 if( isComplete ){
1216 markTermAsChild(pWC, idxNew1, idxTerm);
1217 markTermAsChild(pWC, idxNew2, idxTerm);
1220 #endif /* SQLITE_OMIT_LIKE_OPTIMIZATION */
1222 #ifndef SQLITE_OMIT_VIRTUALTABLE
1223 /* Add a WO_AUX auxiliary term to the constraint set if the
1224 ** current expression is of the form "column OP expr" where OP
1225 ** is an operator that gets passed into virtual tables but which is
1226 ** not normally optimized for ordinary tables. In other words, OP
1227 ** is one of MATCH, LIKE, GLOB, REGEXP, !=, IS, IS NOT, or NOT NULL.
1228 ** This information is used by the xBestIndex methods of
1229 ** virtual tables. The native query optimizer does not attempt
1230 ** to do anything with MATCH functions.
1232 if( pWC->op==TK_AND ){
1233 Expr *pRight = 0, *pLeft = 0;
1234 int res = isAuxiliaryVtabOperator(pExpr, &eOp2, &pLeft, &pRight);
1235 while( res-- > 0 ){
1236 int idxNew;
1237 WhereTerm *pNewTerm;
1238 Bitmask prereqColumn, prereqExpr;
1240 prereqExpr = sqlite3WhereExprUsage(pMaskSet, pRight);
1241 prereqColumn = sqlite3WhereExprUsage(pMaskSet, pLeft);
1242 if( (prereqExpr & prereqColumn)==0 ){
1243 Expr *pNewExpr;
1244 pNewExpr = sqlite3PExpr(pParse, TK_MATCH,
1245 0, sqlite3ExprDup(db, pRight, 0));
1246 if( ExprHasProperty(pExpr, EP_FromJoin) && pNewExpr ){
1247 ExprSetProperty(pNewExpr, EP_FromJoin);
1249 idxNew = whereClauseInsert(pWC, pNewExpr, TERM_VIRTUAL|TERM_DYNAMIC);
1250 testcase( idxNew==0 );
1251 pNewTerm = &pWC->a[idxNew];
1252 pNewTerm->prereqRight = prereqExpr;
1253 pNewTerm->leftCursor = pLeft->iTable;
1254 pNewTerm->u.leftColumn = pLeft->iColumn;
1255 pNewTerm->eOperator = WO_AUX;
1256 pNewTerm->eMatchOp = eOp2;
1257 markTermAsChild(pWC, idxNew, idxTerm);
1258 pTerm = &pWC->a[idxTerm];
1259 pTerm->wtFlags |= TERM_COPIED;
1260 pNewTerm->prereqAll = pTerm->prereqAll;
1262 SWAP(Expr*, pLeft, pRight);
1265 #endif /* SQLITE_OMIT_VIRTUALTABLE */
1267 /* If there is a vector == or IS term - e.g. "(a, b) == (?, ?)" - create
1268 ** new terms for each component comparison - "a = ?" and "b = ?". The
1269 ** new terms completely replace the original vector comparison, which is
1270 ** no longer used.
1272 ** This is only required if at least one side of the comparison operation
1273 ** is not a sub-select. */
1274 if( pWC->op==TK_AND
1275 && (pExpr->op==TK_EQ || pExpr->op==TK_IS)
1276 && (nLeft = sqlite3ExprVectorSize(pExpr->pLeft))>1
1277 && sqlite3ExprVectorSize(pExpr->pRight)==nLeft
1278 && ( (pExpr->pLeft->flags & EP_xIsSelect)==0
1279 || (pExpr->pRight->flags & EP_xIsSelect)==0)
1281 int i;
1282 for(i=0; i<nLeft; i++){
1283 int idxNew;
1284 Expr *pNew;
1285 Expr *pLeft = sqlite3ExprForVectorField(pParse, pExpr->pLeft, i);
1286 Expr *pRight = sqlite3ExprForVectorField(pParse, pExpr->pRight, i);
1288 pNew = sqlite3PExpr(pParse, pExpr->op, pLeft, pRight);
1289 transferJoinMarkings(pNew, pExpr);
1290 idxNew = whereClauseInsert(pWC, pNew, TERM_DYNAMIC);
1291 exprAnalyze(pSrc, pWC, idxNew);
1293 pTerm = &pWC->a[idxTerm];
1294 pTerm->wtFlags |= TERM_CODED|TERM_VIRTUAL; /* Disable the original */
1295 pTerm->eOperator = 0;
1298 /* If there is a vector IN term - e.g. "(a, b) IN (SELECT ...)" - create
1299 ** a virtual term for each vector component. The expression object
1300 ** used by each such virtual term is pExpr (the full vector IN(...)
1301 ** expression). The WhereTerm.iField variable identifies the index within
1302 ** the vector on the LHS that the virtual term represents.
1304 ** This only works if the RHS is a simple SELECT, not a compound
1306 if( pWC->op==TK_AND && pExpr->op==TK_IN && pTerm->iField==0
1307 && pExpr->pLeft->op==TK_VECTOR
1308 && pExpr->x.pSelect->pPrior==0
1310 int i;
1311 for(i=0; i<sqlite3ExprVectorSize(pExpr->pLeft); i++){
1312 int idxNew;
1313 idxNew = whereClauseInsert(pWC, pExpr, TERM_VIRTUAL);
1314 pWC->a[idxNew].iField = i+1;
1315 exprAnalyze(pSrc, pWC, idxNew);
1316 markTermAsChild(pWC, idxNew, idxTerm);
1320 #ifdef SQLITE_ENABLE_STAT3_OR_STAT4
1321 /* When sqlite_stat3 histogram data is available an operator of the
1322 ** form "x IS NOT NULL" can sometimes be evaluated more efficiently
1323 ** as "x>NULL" if x is not an INTEGER PRIMARY KEY. So construct a
1324 ** virtual term of that form.
1326 ** Note that the virtual term must be tagged with TERM_VNULL.
1328 if( pExpr->op==TK_NOTNULL
1329 && pExpr->pLeft->op==TK_COLUMN
1330 && pExpr->pLeft->iColumn>=0
1331 && OptimizationEnabled(db, SQLITE_Stat34)
1333 Expr *pNewExpr;
1334 Expr *pLeft = pExpr->pLeft;
1335 int idxNew;
1336 WhereTerm *pNewTerm;
1338 pNewExpr = sqlite3PExpr(pParse, TK_GT,
1339 sqlite3ExprDup(db, pLeft, 0),
1340 sqlite3ExprAlloc(db, TK_NULL, 0, 0));
1342 idxNew = whereClauseInsert(pWC, pNewExpr,
1343 TERM_VIRTUAL|TERM_DYNAMIC|TERM_VNULL);
1344 if( idxNew ){
1345 pNewTerm = &pWC->a[idxNew];
1346 pNewTerm->prereqRight = 0;
1347 pNewTerm->leftCursor = pLeft->iTable;
1348 pNewTerm->u.leftColumn = pLeft->iColumn;
1349 pNewTerm->eOperator = WO_GT;
1350 markTermAsChild(pWC, idxNew, idxTerm);
1351 pTerm = &pWC->a[idxTerm];
1352 pTerm->wtFlags |= TERM_COPIED;
1353 pNewTerm->prereqAll = pTerm->prereqAll;
1356 #endif /* SQLITE_ENABLE_STAT3_OR_STAT4 */
1358 /* Prevent ON clause terms of a LEFT JOIN from being used to drive
1359 ** an index for tables to the left of the join.
1361 testcase( pTerm!=&pWC->a[idxTerm] );
1362 pTerm = &pWC->a[idxTerm];
1363 pTerm->prereqRight |= extraRight;
1366 /***************************************************************************
1367 ** Routines with file scope above. Interface to the rest of the where.c
1368 ** subsystem follows.
1369 ***************************************************************************/
1372 ** This routine identifies subexpressions in the WHERE clause where
1373 ** each subexpression is separated by the AND operator or some other
1374 ** operator specified in the op parameter. The WhereClause structure
1375 ** is filled with pointers to subexpressions. For example:
1377 ** WHERE a=='hello' AND coalesce(b,11)<10 AND (c+12!=d OR c==22)
1378 ** \________/ \_______________/ \________________/
1379 ** slot[0] slot[1] slot[2]
1381 ** The original WHERE clause in pExpr is unaltered. All this routine
1382 ** does is make slot[] entries point to substructure within pExpr.
1384 ** In the previous sentence and in the diagram, "slot[]" refers to
1385 ** the WhereClause.a[] array. The slot[] array grows as needed to contain
1386 ** all terms of the WHERE clause.
1388 void sqlite3WhereSplit(WhereClause *pWC, Expr *pExpr, u8 op){
1389 Expr *pE2 = sqlite3ExprSkipCollate(pExpr);
1390 pWC->op = op;
1391 if( pE2==0 ) return;
1392 if( pE2->op!=op ){
1393 whereClauseInsert(pWC, pExpr, 0);
1394 }else{
1395 sqlite3WhereSplit(pWC, pE2->pLeft, op);
1396 sqlite3WhereSplit(pWC, pE2->pRight, op);
1401 ** Initialize a preallocated WhereClause structure.
1403 void sqlite3WhereClauseInit(
1404 WhereClause *pWC, /* The WhereClause to be initialized */
1405 WhereInfo *pWInfo /* The WHERE processing context */
1407 pWC->pWInfo = pWInfo;
1408 pWC->pOuter = 0;
1409 pWC->nTerm = 0;
1410 pWC->nSlot = ArraySize(pWC->aStatic);
1411 pWC->a = pWC->aStatic;
1415 ** Deallocate a WhereClause structure. The WhereClause structure
1416 ** itself is not freed. This routine is the inverse of
1417 ** sqlite3WhereClauseInit().
1419 void sqlite3WhereClauseClear(WhereClause *pWC){
1420 int i;
1421 WhereTerm *a;
1422 sqlite3 *db = pWC->pWInfo->pParse->db;
1423 for(i=pWC->nTerm-1, a=pWC->a; i>=0; i--, a++){
1424 if( a->wtFlags & TERM_DYNAMIC ){
1425 sqlite3ExprDelete(db, a->pExpr);
1427 if( a->wtFlags & TERM_ORINFO ){
1428 whereOrInfoDelete(db, a->u.pOrInfo);
1429 }else if( a->wtFlags & TERM_ANDINFO ){
1430 whereAndInfoDelete(db, a->u.pAndInfo);
1433 if( pWC->a!=pWC->aStatic ){
1434 sqlite3DbFree(db, pWC->a);
1440 ** These routines walk (recursively) an expression tree and generate
1441 ** a bitmask indicating which tables are used in that expression
1442 ** tree.
1444 Bitmask sqlite3WhereExprUsage(WhereMaskSet *pMaskSet, Expr *p){
1445 Bitmask mask;
1446 if( p==0 ) return 0;
1447 if( p->op==TK_COLUMN ){
1448 return sqlite3WhereGetMask(pMaskSet, p->iTable);
1450 mask = (p->op==TK_IF_NULL_ROW) ? sqlite3WhereGetMask(pMaskSet, p->iTable) : 0;
1451 assert( !ExprHasProperty(p, EP_TokenOnly) );
1452 if( p->pLeft ) mask |= sqlite3WhereExprUsage(pMaskSet, p->pLeft);
1453 if( p->pRight ){
1454 mask |= sqlite3WhereExprUsage(pMaskSet, p->pRight);
1455 assert( p->x.pList==0 );
1456 }else if( ExprHasProperty(p, EP_xIsSelect) ){
1457 if( ExprHasProperty(p, EP_VarSelect) ) pMaskSet->bVarSelect = 1;
1458 mask |= exprSelectUsage(pMaskSet, p->x.pSelect);
1459 }else if( p->x.pList ){
1460 mask |= sqlite3WhereExprListUsage(pMaskSet, p->x.pList);
1462 return mask;
1464 Bitmask sqlite3WhereExprListUsage(WhereMaskSet *pMaskSet, ExprList *pList){
1465 int i;
1466 Bitmask mask = 0;
1467 if( pList ){
1468 for(i=0; i<pList->nExpr; i++){
1469 mask |= sqlite3WhereExprUsage(pMaskSet, pList->a[i].pExpr);
1472 return mask;
1477 ** Call exprAnalyze on all terms in a WHERE clause.
1479 ** Note that exprAnalyze() might add new virtual terms onto the
1480 ** end of the WHERE clause. We do not want to analyze these new
1481 ** virtual terms, so start analyzing at the end and work forward
1482 ** so that the added virtual terms are never processed.
1484 void sqlite3WhereExprAnalyze(
1485 SrcList *pTabList, /* the FROM clause */
1486 WhereClause *pWC /* the WHERE clause to be analyzed */
1488 int i;
1489 for(i=pWC->nTerm-1; i>=0; i--){
1490 exprAnalyze(pTabList, pWC, i);
1495 ** For table-valued-functions, transform the function arguments into
1496 ** new WHERE clause terms.
1498 ** Each function argument translates into an equality constraint against
1499 ** a HIDDEN column in the table.
1501 void sqlite3WhereTabFuncArgs(
1502 Parse *pParse, /* Parsing context */
1503 struct SrcList_item *pItem, /* The FROM clause term to process */
1504 WhereClause *pWC /* Xfer function arguments to here */
1506 Table *pTab;
1507 int j, k;
1508 ExprList *pArgs;
1509 Expr *pColRef;
1510 Expr *pTerm;
1511 if( pItem->fg.isTabFunc==0 ) return;
1512 pTab = pItem->pTab;
1513 assert( pTab!=0 );
1514 pArgs = pItem->u1.pFuncArg;
1515 if( pArgs==0 ) return;
1516 for(j=k=0; j<pArgs->nExpr; j++){
1517 while( k<pTab->nCol && (pTab->aCol[k].colFlags & COLFLAG_HIDDEN)==0 ){k++;}
1518 if( k>=pTab->nCol ){
1519 sqlite3ErrorMsg(pParse, "too many arguments on %s() - max %d",
1520 pTab->zName, j);
1521 return;
1523 pColRef = sqlite3ExprAlloc(pParse->db, TK_COLUMN, 0, 0);
1524 if( pColRef==0 ) return;
1525 pColRef->iTable = pItem->iCursor;
1526 pColRef->iColumn = k++;
1527 pColRef->pTab = pTab;
1528 pTerm = sqlite3PExpr(pParse, TK_EQ, pColRef,
1529 sqlite3ExprDup(pParse->db, pArgs->a[j].pExpr, 0));
1530 whereClauseInsert(pWC, pTerm, TERM_DYNAMIC);