xref: /sqlite-3.40.0/src/whereexpr.c (revision 2e27d28f)
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.
10 **
11 *************************************************************************
12 ** This module contains C code that generates VDBE code used to process
13 ** the WHERE clause of SQL statements.
14 **
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.
18 */
19 #include "sqliteInt.h"
20 #include "whereInt.h"
21 
22 /* Forward declarations */
23 static void exprAnalyze(SrcList*, WhereClause*, int);
24 
25 /*
26 ** Deallocate all memory associated with a WhereOrInfo object.
27 */
28 static void whereOrInfoDelete(sqlite3 *db, WhereOrInfo *p){
29   sqlite3WhereClauseClear(&p->wc);
30   sqlite3DbFree(db, p);
31 }
32 
33 /*
34 ** Deallocate all memory associated with a WhereAndInfo object.
35 */
36 static void whereAndInfoDelete(sqlite3 *db, WhereAndInfo *p){
37   sqlite3WhereClauseClear(&p->wc);
38   sqlite3DbFree(db, p);
39 }
40 
41 /*
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.
48 **
49 ** This routine will increase the size of the pWC->a[] array as necessary.
50 **
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.
54 **
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.
59 */
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);
71       }
72       pWC->a = pOld;
73       return 0;
74     }
75     memcpy(pWC->a, pOld, sizeof(pWC->a[0])*pWC->nTerm);
76     if( pOld!=pWC->aStatic ){
77       sqlite3DbFree(db, pOld);
78     }
79     pWC->nSlot = sqlite3DbMallocSize(db, pWC->a)/sizeof(pWC->a[0]);
80   }
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;
86   }
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;
94 }
95 
96 /*
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"
100 */
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;
107 }
108 
109 /*
110 ** Commute a comparison operator.  Expressions of the form "X op Y"
111 ** are converted into "Y op X".
112 **
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.
120 */
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;
136     }
137   }
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;
146   }
147 }
148 
149 /*
150 ** Translate from TK_xx operator to WO_xx bitmask.
151 */
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));
164   }
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;
174 }
175 
176 
177 #ifndef SQLITE_OMIT_LIKE_OPTIMIZATION
178 /*
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.
182 **
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.
189 */
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 */
196 ){
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 */
207 
208   if( !sqlite3IsLikeFunction(db, pExpr, pnoCase, wc) ){
209     return 0;
210   }
211 #ifdef SQLITE_EBCDIC
212   if( *pnoCase ) return 0;
213 #endif
214   pList = pExpr->x.pList;
215   pLeft = pList->a[1].pExpr;
216 
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);
225     }
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;
230   }
231   if( z ){
232 
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.
239     */
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 */
244       ){
245         sqlite3ValueFree(pVal);
246         return 0;
247       }
248     }
249 
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 ){
255         if( z[cnt++]>0xc0 ) while( (z[cnt]&0xc0)==0x80 ){ cnt++; }
256       }
257     }
258 
259     /* The optimization is possible only if (1) the pattern does not begin
260     ** with a wildcard and if (2) the non-wildcard prefix does not end with
261     ** an (illegal 0xff) character.  The second condition is necessary so
262     ** that we can increment the prefix key to find an upper bound for the
263     ** range search.
264     */
265     if( cnt!=0 && 255!=(u8)z[cnt-1] ){
266       Expr *pPrefix;
267 
268       /* A "complete" match if the pattern ends with "*" or "%" */
269       *pisComplete = c==wc[0] && z[cnt+1]==0;
270 
271       /* Get the pattern prefix.  Remove all escapes from the prefix. */
272       pPrefix = sqlite3Expr(db, TK_STRING, (char*)z);
273       if( pPrefix ){
274         int iFrom, iTo;
275         char *zNew = pPrefix->u.zToken;
276         zNew[cnt] = 0;
277         for(iFrom=iTo=0; iFrom<cnt; iFrom++){
278           if( zNew[iFrom]==wc[3] ) iFrom++;
279           zNew[iTo++] = zNew[iFrom];
280         }
281         zNew[iTo] = 0;
282       }
283       *ppPrefix = pPrefix;
284 
285       /* If the RHS pattern is a bound parameter, make arrangements to
286       ** reprepare the statement when that parameter is rebound */
287       if( op==TK_VARIABLE ){
288         Vdbe *v = pParse->pVdbe;
289         sqlite3VdbeSetVarmask(v, pRight->iColumn);
290         if( *pisComplete && pRight->u.zToken[1] ){
291           /* If the rhs of the LIKE expression is a variable, and the current
292           ** value of the variable means there is no need to invoke the LIKE
293           ** function, then no OP_Variable will be added to the program.
294           ** This causes problems for the sqlite3_bind_parameter_name()
295           ** API. To work around them, add a dummy OP_Variable here.
296           */
297           int r1 = sqlite3GetTempReg(pParse);
298           sqlite3ExprCodeTarget(pParse, pRight, r1);
299           sqlite3VdbeChangeP3(v, sqlite3VdbeCurrentAddr(v)-1, 0);
300           sqlite3ReleaseTempReg(pParse, r1);
301         }
302       }
303     }else{
304       z = 0;
305     }
306   }
307 
308   rc = (z!=0);
309   sqlite3ValueFree(pVal);
310   return rc;
311 }
312 #endif /* SQLITE_OMIT_LIKE_OPTIMIZATION */
313 
314 
315 #ifndef SQLITE_OMIT_VIRTUALTABLE
316 /*
317 ** Check to see if the given expression is of the form
318 **
319 **         column OP expr
320 **
321 ** where OP is one of MATCH, GLOB, LIKE or REGEXP and "column" is a
322 ** column of a virtual table.
323 **
324 ** If it is then return TRUE.  If not, return FALSE.
325 */
326 static int isMatchOfColumn(
327   Expr *pExpr,                    /* Test this expression */
328   unsigned char *peOp2            /* OUT: 0 for MATCH, or else an op2 value */
329 ){
330   static const struct Op2 {
331     const char *zOp;
332     unsigned char eOp2;
333   } aOp[] = {
334     { "match",  SQLITE_INDEX_CONSTRAINT_MATCH },
335     { "glob",   SQLITE_INDEX_CONSTRAINT_GLOB },
336     { "like",   SQLITE_INDEX_CONSTRAINT_LIKE },
337     { "regexp", SQLITE_INDEX_CONSTRAINT_REGEXP }
338   };
339   ExprList *pList;
340   Expr *pCol;                     /* Column reference */
341   int i;
342 
343   if( pExpr->op!=TK_FUNCTION ){
344     return 0;
345   }
346   pList = pExpr->x.pList;
347   if( pList==0 || pList->nExpr!=2 ){
348     return 0;
349   }
350   pCol = pList->a[1].pExpr;
351   if( pCol->op!=TK_COLUMN || !IsVirtual(pCol->pTab) ){
352     return 0;
353   }
354   for(i=0; i<ArraySize(aOp); i++){
355     if( sqlite3StrICmp(pExpr->u.zToken, aOp[i].zOp)==0 ){
356       *peOp2 = aOp[i].eOp2;
357       return 1;
358     }
359   }
360   return 0;
361 }
362 #endif /* SQLITE_OMIT_VIRTUALTABLE */
363 
364 /*
365 ** If the pBase expression originated in the ON or USING clause of
366 ** a join, then transfer the appropriate markings over to derived.
367 */
368 static void transferJoinMarkings(Expr *pDerived, Expr *pBase){
369   if( pDerived ){
370     pDerived->flags |= pBase->flags & EP_FromJoin;
371     pDerived->iRightJoinTable = pBase->iRightJoinTable;
372   }
373 }
374 
375 /*
376 ** Mark term iChild as being a child of term iParent
377 */
378 static void markTermAsChild(WhereClause *pWC, int iChild, int iParent){
379   pWC->a[iChild].iParent = iParent;
380   pWC->a[iChild].truthProb = pWC->a[iParent].truthProb;
381   pWC->a[iParent].nChild++;
382 }
383 
384 /*
385 ** Return the N-th AND-connected subterm of pTerm.  Or if pTerm is not
386 ** a conjunction, then return just pTerm when N==0.  If N is exceeds
387 ** the number of available subterms, return NULL.
388 */
389 static WhereTerm *whereNthSubterm(WhereTerm *pTerm, int N){
390   if( pTerm->eOperator!=WO_AND ){
391     return N==0 ? pTerm : 0;
392   }
393   if( N<pTerm->u.pAndInfo->wc.nTerm ){
394     return &pTerm->u.pAndInfo->wc.a[N];
395   }
396   return 0;
397 }
398 
399 /*
400 ** Subterms pOne and pTwo are contained within WHERE clause pWC.  The
401 ** two subterms are in disjunction - they are OR-ed together.
402 **
403 ** If these two terms are both of the form:  "A op B" with the same
404 ** A and B values but different operators and if the operators are
405 ** compatible (if one is = and the other is <, for example) then
406 ** add a new virtual AND term to pWC that is the combination of the
407 ** two.
408 **
409 ** Some examples:
410 **
411 **    x<y OR x=y    -->     x<=y
412 **    x=y OR x=y    -->     x=y
413 **    x<=y OR x<y   -->     x<=y
414 **
415 ** The following is NOT generated:
416 **
417 **    x<y OR x>y    -->     x!=y
418 */
419 static void whereCombineDisjuncts(
420   SrcList *pSrc,         /* the FROM clause */
421   WhereClause *pWC,      /* The complete WHERE clause */
422   WhereTerm *pOne,       /* First disjunct */
423   WhereTerm *pTwo        /* Second disjunct */
424 ){
425   u16 eOp = pOne->eOperator | pTwo->eOperator;
426   sqlite3 *db;           /* Database connection (for malloc) */
427   Expr *pNew;            /* New virtual expression */
428   int op;                /* Operator for the combined expression */
429   int idxNew;            /* Index in pWC of the next virtual term */
430 
431   if( (pOne->eOperator & (WO_EQ|WO_LT|WO_LE|WO_GT|WO_GE))==0 ) return;
432   if( (pTwo->eOperator & (WO_EQ|WO_LT|WO_LE|WO_GT|WO_GE))==0 ) return;
433   if( (eOp & (WO_EQ|WO_LT|WO_LE))!=eOp
434    && (eOp & (WO_EQ|WO_GT|WO_GE))!=eOp ) return;
435   assert( pOne->pExpr->pLeft!=0 && pOne->pExpr->pRight!=0 );
436   assert( pTwo->pExpr->pLeft!=0 && pTwo->pExpr->pRight!=0 );
437   if( sqlite3ExprCompare(0,pOne->pExpr->pLeft, pTwo->pExpr->pLeft, -1) ) return;
438   if( sqlite3ExprCompare(0,pOne->pExpr->pRight, pTwo->pExpr->pRight,-1) )return;
439   /* If we reach this point, it means the two subterms can be combined */
440   if( (eOp & (eOp-1))!=0 ){
441     if( eOp & (WO_LT|WO_LE) ){
442       eOp = WO_LE;
443     }else{
444       assert( eOp & (WO_GT|WO_GE) );
445       eOp = WO_GE;
446     }
447   }
448   db = pWC->pWInfo->pParse->db;
449   pNew = sqlite3ExprDup(db, pOne->pExpr, 0);
450   if( pNew==0 ) return;
451   for(op=TK_EQ; eOp!=(WO_EQ<<(op-TK_EQ)); op++){ assert( op<TK_GE ); }
452   pNew->op = op;
453   idxNew = whereClauseInsert(pWC, pNew, TERM_VIRTUAL|TERM_DYNAMIC);
454   exprAnalyze(pSrc, pWC, idxNew);
455 }
456 
457 #if !defined(SQLITE_OMIT_OR_OPTIMIZATION) && !defined(SQLITE_OMIT_SUBQUERY)
458 /*
459 ** Analyze a term that consists of two or more OR-connected
460 ** subterms.  So in:
461 **
462 **     ... WHERE  (a=5) AND (b=7 OR c=9 OR d=13) AND (d=13)
463 **                          ^^^^^^^^^^^^^^^^^^^^
464 **
465 ** This routine analyzes terms such as the middle term in the above example.
466 ** A WhereOrTerm object is computed and attached to the term under
467 ** analysis, regardless of the outcome of the analysis.  Hence:
468 **
469 **     WhereTerm.wtFlags   |=  TERM_ORINFO
470 **     WhereTerm.u.pOrInfo  =  a dynamically allocated WhereOrTerm object
471 **
472 ** The term being analyzed must have two or more of OR-connected subterms.
473 ** A single subterm might be a set of AND-connected sub-subterms.
474 ** Examples of terms under analysis:
475 **
476 **     (A)     t1.x=t2.y OR t1.x=t2.z OR t1.y=15 OR t1.z=t3.a+5
477 **     (B)     x=expr1 OR expr2=x OR x=expr3
478 **     (C)     t1.x=t2.y OR (t1.x=t2.z AND t1.y=15)
479 **     (D)     x=expr1 OR (y>11 AND y<22 AND z LIKE '*hello*')
480 **     (E)     (p.a=1 AND q.b=2 AND r.c=3) OR (p.x=4 AND q.y=5 AND r.z=6)
481 **     (F)     x>A OR (x=A AND y>=B)
482 **
483 ** CASE 1:
484 **
485 ** If all subterms are of the form T.C=expr for some single column of C and
486 ** a single table T (as shown in example B above) then create a new virtual
487 ** term that is an equivalent IN expression.  In other words, if the term
488 ** being analyzed is:
489 **
490 **      x = expr1  OR  expr2 = x  OR  x = expr3
491 **
492 ** then create a new virtual term like this:
493 **
494 **      x IN (expr1,expr2,expr3)
495 **
496 ** CASE 2:
497 **
498 ** If there are exactly two disjuncts and one side has x>A and the other side
499 ** has x=A (for the same x and A) then add a new virtual conjunct term to the
500 ** WHERE clause of the form "x>=A".  Example:
501 **
502 **      x>A OR (x=A AND y>B)    adds:    x>=A
503 **
504 ** The added conjunct can sometimes be helpful in query planning.
505 **
506 ** CASE 3:
507 **
508 ** If all subterms are indexable by a single table T, then set
509 **
510 **     WhereTerm.eOperator              =  WO_OR
511 **     WhereTerm.u.pOrInfo->indexable  |=  the cursor number for table T
512 **
513 ** A subterm is "indexable" if it is of the form
514 ** "T.C <op> <expr>" where C is any column of table T and
515 ** <op> is one of "=", "<", "<=", ">", ">=", "IS NULL", or "IN".
516 ** A subterm is also indexable if it is an AND of two or more
517 ** subsubterms at least one of which is indexable.  Indexable AND
518 ** subterms have their eOperator set to WO_AND and they have
519 ** u.pAndInfo set to a dynamically allocated WhereAndTerm object.
520 **
521 ** From another point of view, "indexable" means that the subterm could
522 ** potentially be used with an index if an appropriate index exists.
523 ** This analysis does not consider whether or not the index exists; that
524 ** is decided elsewhere.  This analysis only looks at whether subterms
525 ** appropriate for indexing exist.
526 **
527 ** All examples A through E above satisfy case 3.  But if a term
528 ** also satisfies case 1 (such as B) we know that the optimizer will
529 ** always prefer case 1, so in that case we pretend that case 3 is not
530 ** satisfied.
531 **
532 ** It might be the case that multiple tables are indexable.  For example,
533 ** (E) above is indexable on tables P, Q, and R.
534 **
535 ** Terms that satisfy case 3 are candidates for lookup by using
536 ** separate indices to find rowids for each subterm and composing
537 ** the union of all rowids using a RowSet object.  This is similar
538 ** to "bitmap indices" in other database engines.
539 **
540 ** OTHERWISE:
541 **
542 ** If none of cases 1, 2, or 3 apply, then leave the eOperator set to
543 ** zero.  This term is not useful for search.
544 */
545 static void exprAnalyzeOrTerm(
546   SrcList *pSrc,            /* the FROM clause */
547   WhereClause *pWC,         /* the complete WHERE clause */
548   int idxTerm               /* Index of the OR-term to be analyzed */
549 ){
550   WhereInfo *pWInfo = pWC->pWInfo;        /* WHERE clause processing context */
551   Parse *pParse = pWInfo->pParse;         /* Parser context */
552   sqlite3 *db = pParse->db;               /* Database connection */
553   WhereTerm *pTerm = &pWC->a[idxTerm];    /* The term to be analyzed */
554   Expr *pExpr = pTerm->pExpr;             /* The expression of the term */
555   int i;                                  /* Loop counters */
556   WhereClause *pOrWc;       /* Breakup of pTerm into subterms */
557   WhereTerm *pOrTerm;       /* A Sub-term within the pOrWc */
558   WhereOrInfo *pOrInfo;     /* Additional information associated with pTerm */
559   Bitmask chngToIN;         /* Tables that might satisfy case 1 */
560   Bitmask indexable;        /* Tables that are indexable, satisfying case 2 */
561 
562   /*
563   ** Break the OR clause into its separate subterms.  The subterms are
564   ** stored in a WhereClause structure containing within the WhereOrInfo
565   ** object that is attached to the original OR clause term.
566   */
567   assert( (pTerm->wtFlags & (TERM_DYNAMIC|TERM_ORINFO|TERM_ANDINFO))==0 );
568   assert( pExpr->op==TK_OR );
569   pTerm->u.pOrInfo = pOrInfo = sqlite3DbMallocZero(db, sizeof(*pOrInfo));
570   if( pOrInfo==0 ) return;
571   pTerm->wtFlags |= TERM_ORINFO;
572   pOrWc = &pOrInfo->wc;
573   memset(pOrWc->aStatic, 0, sizeof(pOrWc->aStatic));
574   sqlite3WhereClauseInit(pOrWc, pWInfo);
575   sqlite3WhereSplit(pOrWc, pExpr, TK_OR);
576   sqlite3WhereExprAnalyze(pSrc, pOrWc);
577   if( db->mallocFailed ) return;
578   assert( pOrWc->nTerm>=2 );
579 
580   /*
581   ** Compute the set of tables that might satisfy cases 1 or 3.
582   */
583   indexable = ~(Bitmask)0;
584   chngToIN = ~(Bitmask)0;
585   for(i=pOrWc->nTerm-1, pOrTerm=pOrWc->a; i>=0 && indexable; i--, pOrTerm++){
586     if( (pOrTerm->eOperator & WO_SINGLE)==0 ){
587       WhereAndInfo *pAndInfo;
588       assert( (pOrTerm->wtFlags & (TERM_ANDINFO|TERM_ORINFO))==0 );
589       chngToIN = 0;
590       pAndInfo = sqlite3DbMallocRawNN(db, sizeof(*pAndInfo));
591       if( pAndInfo ){
592         WhereClause *pAndWC;
593         WhereTerm *pAndTerm;
594         int j;
595         Bitmask b = 0;
596         pOrTerm->u.pAndInfo = pAndInfo;
597         pOrTerm->wtFlags |= TERM_ANDINFO;
598         pOrTerm->eOperator = WO_AND;
599         pAndWC = &pAndInfo->wc;
600         memset(pAndWC->aStatic, 0, sizeof(pAndWC->aStatic));
601         sqlite3WhereClauseInit(pAndWC, pWC->pWInfo);
602         sqlite3WhereSplit(pAndWC, pOrTerm->pExpr, TK_AND);
603         sqlite3WhereExprAnalyze(pSrc, pAndWC);
604         pAndWC->pOuter = pWC;
605         if( !db->mallocFailed ){
606           for(j=0, pAndTerm=pAndWC->a; j<pAndWC->nTerm; j++, pAndTerm++){
607             assert( pAndTerm->pExpr );
608             if( allowedOp(pAndTerm->pExpr->op)
609              || pAndTerm->eOperator==WO_MATCH
610             ){
611               b |= sqlite3WhereGetMask(&pWInfo->sMaskSet, pAndTerm->leftCursor);
612             }
613           }
614         }
615         indexable &= b;
616       }
617     }else if( pOrTerm->wtFlags & TERM_COPIED ){
618       /* Skip this term for now.  We revisit it when we process the
619       ** corresponding TERM_VIRTUAL term */
620     }else{
621       Bitmask b;
622       b = sqlite3WhereGetMask(&pWInfo->sMaskSet, pOrTerm->leftCursor);
623       if( pOrTerm->wtFlags & TERM_VIRTUAL ){
624         WhereTerm *pOther = &pOrWc->a[pOrTerm->iParent];
625         b |= sqlite3WhereGetMask(&pWInfo->sMaskSet, pOther->leftCursor);
626       }
627       indexable &= b;
628       if( (pOrTerm->eOperator & WO_EQ)==0 ){
629         chngToIN = 0;
630       }else{
631         chngToIN &= b;
632       }
633     }
634   }
635 
636   /*
637   ** Record the set of tables that satisfy case 3.  The set might be
638   ** empty.
639   */
640   pOrInfo->indexable = indexable;
641   pTerm->eOperator = indexable==0 ? 0 : WO_OR;
642 
643   /* For a two-way OR, attempt to implementation case 2.
644   */
645   if( indexable && pOrWc->nTerm==2 ){
646     int iOne = 0;
647     WhereTerm *pOne;
648     while( (pOne = whereNthSubterm(&pOrWc->a[0],iOne++))!=0 ){
649       int iTwo = 0;
650       WhereTerm *pTwo;
651       while( (pTwo = whereNthSubterm(&pOrWc->a[1],iTwo++))!=0 ){
652         whereCombineDisjuncts(pSrc, pWC, pOne, pTwo);
653       }
654     }
655   }
656 
657   /*
658   ** chngToIN holds a set of tables that *might* satisfy case 1.  But
659   ** we have to do some additional checking to see if case 1 really
660   ** is satisfied.
661   **
662   ** chngToIN will hold either 0, 1, or 2 bits.  The 0-bit case means
663   ** that there is no possibility of transforming the OR clause into an
664   ** IN operator because one or more terms in the OR clause contain
665   ** something other than == on a column in the single table.  The 1-bit
666   ** case means that every term of the OR clause is of the form
667   ** "table.column=expr" for some single table.  The one bit that is set
668   ** will correspond to the common table.  We still need to check to make
669   ** sure the same column is used on all terms.  The 2-bit case is when
670   ** the all terms are of the form "table1.column=table2.column".  It
671   ** might be possible to form an IN operator with either table1.column
672   ** or table2.column as the LHS if either is common to every term of
673   ** the OR clause.
674   **
675   ** Note that terms of the form "table.column1=table.column2" (the
676   ** same table on both sizes of the ==) cannot be optimized.
677   */
678   if( chngToIN ){
679     int okToChngToIN = 0;     /* True if the conversion to IN is valid */
680     int iColumn = -1;         /* Column index on lhs of IN operator */
681     int iCursor = -1;         /* Table cursor common to all terms */
682     int j = 0;                /* Loop counter */
683 
684     /* Search for a table and column that appears on one side or the
685     ** other of the == operator in every subterm.  That table and column
686     ** will be recorded in iCursor and iColumn.  There might not be any
687     ** such table and column.  Set okToChngToIN if an appropriate table
688     ** and column is found but leave okToChngToIN false if not found.
689     */
690     for(j=0; j<2 && !okToChngToIN; j++){
691       pOrTerm = pOrWc->a;
692       for(i=pOrWc->nTerm-1; i>=0; i--, pOrTerm++){
693         assert( pOrTerm->eOperator & WO_EQ );
694         pOrTerm->wtFlags &= ~TERM_OR_OK;
695         if( pOrTerm->leftCursor==iCursor ){
696           /* This is the 2-bit case and we are on the second iteration and
697           ** current term is from the first iteration.  So skip this term. */
698           assert( j==1 );
699           continue;
700         }
701         if( (chngToIN & sqlite3WhereGetMask(&pWInfo->sMaskSet,
702                                             pOrTerm->leftCursor))==0 ){
703           /* This term must be of the form t1.a==t2.b where t2 is in the
704           ** chngToIN set but t1 is not.  This term will be either preceded
705           ** or follwed by an inverted copy (t2.b==t1.a).  Skip this term
706           ** and use its inversion. */
707           testcase( pOrTerm->wtFlags & TERM_COPIED );
708           testcase( pOrTerm->wtFlags & TERM_VIRTUAL );
709           assert( pOrTerm->wtFlags & (TERM_COPIED|TERM_VIRTUAL) );
710           continue;
711         }
712         iColumn = pOrTerm->u.leftColumn;
713         iCursor = pOrTerm->leftCursor;
714         break;
715       }
716       if( i<0 ){
717         /* No candidate table+column was found.  This can only occur
718         ** on the second iteration */
719         assert( j==1 );
720         assert( IsPowerOfTwo(chngToIN) );
721         assert( chngToIN==sqlite3WhereGetMask(&pWInfo->sMaskSet, iCursor) );
722         break;
723       }
724       testcase( j==1 );
725 
726       /* We have found a candidate table and column.  Check to see if that
727       ** table and column is common to every term in the OR clause */
728       okToChngToIN = 1;
729       for(; i>=0 && okToChngToIN; i--, pOrTerm++){
730         assert( pOrTerm->eOperator & WO_EQ );
731         if( pOrTerm->leftCursor!=iCursor ){
732           pOrTerm->wtFlags &= ~TERM_OR_OK;
733         }else if( pOrTerm->u.leftColumn!=iColumn ){
734           okToChngToIN = 0;
735         }else{
736           int affLeft, affRight;
737           /* If the right-hand side is also a column, then the affinities
738           ** of both right and left sides must be such that no type
739           ** conversions are required on the right.  (Ticket #2249)
740           */
741           affRight = sqlite3ExprAffinity(pOrTerm->pExpr->pRight);
742           affLeft = sqlite3ExprAffinity(pOrTerm->pExpr->pLeft);
743           if( affRight!=0 && affRight!=affLeft ){
744             okToChngToIN = 0;
745           }else{
746             pOrTerm->wtFlags |= TERM_OR_OK;
747           }
748         }
749       }
750     }
751 
752     /* At this point, okToChngToIN is true if original pTerm satisfies
753     ** case 1.  In that case, construct a new virtual term that is
754     ** pTerm converted into an IN operator.
755     */
756     if( okToChngToIN ){
757       Expr *pDup;            /* A transient duplicate expression */
758       ExprList *pList = 0;   /* The RHS of the IN operator */
759       Expr *pLeft = 0;       /* The LHS of the IN operator */
760       Expr *pNew;            /* The complete IN operator */
761 
762       for(i=pOrWc->nTerm-1, pOrTerm=pOrWc->a; i>=0; i--, pOrTerm++){
763         if( (pOrTerm->wtFlags & TERM_OR_OK)==0 ) continue;
764         assert( pOrTerm->eOperator & WO_EQ );
765         assert( pOrTerm->leftCursor==iCursor );
766         assert( pOrTerm->u.leftColumn==iColumn );
767         pDup = sqlite3ExprDup(db, pOrTerm->pExpr->pRight, 0);
768         pList = sqlite3ExprListAppend(pWInfo->pParse, pList, pDup);
769         pLeft = pOrTerm->pExpr->pLeft;
770       }
771       assert( pLeft!=0 );
772       pDup = sqlite3ExprDup(db, pLeft, 0);
773       pNew = sqlite3PExpr(pParse, TK_IN, pDup, 0);
774       if( pNew ){
775         int idxNew;
776         transferJoinMarkings(pNew, pExpr);
777         assert( !ExprHasProperty(pNew, EP_xIsSelect) );
778         pNew->x.pList = pList;
779         idxNew = whereClauseInsert(pWC, pNew, TERM_VIRTUAL|TERM_DYNAMIC);
780         testcase( idxNew==0 );
781         exprAnalyze(pSrc, pWC, idxNew);
782         pTerm = &pWC->a[idxTerm];
783         markTermAsChild(pWC, idxNew, idxTerm);
784       }else{
785         sqlite3ExprListDelete(db, pList);
786       }
787       pTerm->eOperator = WO_NOOP;  /* case 1 trumps case 3 */
788     }
789   }
790 }
791 #endif /* !SQLITE_OMIT_OR_OPTIMIZATION && !SQLITE_OMIT_SUBQUERY */
792 
793 /*
794 ** We already know that pExpr is a binary operator where both operands are
795 ** column references.  This routine checks to see if pExpr is an equivalence
796 ** relation:
797 **   1.  The SQLITE_Transitive optimization must be enabled
798 **   2.  Must be either an == or an IS operator
799 **   3.  Not originating in the ON clause of an OUTER JOIN
800 **   4.  The affinities of A and B must be compatible
801 **   5a. Both operands use the same collating sequence OR
802 **   5b. The overall collating sequence is BINARY
803 ** If this routine returns TRUE, that means that the RHS can be substituted
804 ** for the LHS anyplace else in the WHERE clause where the LHS column occurs.
805 ** This is an optimization.  No harm comes from returning 0.  But if 1 is
806 ** returned when it should not be, then incorrect answers might result.
807 */
808 static int termIsEquivalence(Parse *pParse, Expr *pExpr){
809   char aff1, aff2;
810   CollSeq *pColl;
811   const char *zColl1, *zColl2;
812   if( !OptimizationEnabled(pParse->db, SQLITE_Transitive) ) return 0;
813   if( pExpr->op!=TK_EQ && pExpr->op!=TK_IS ) return 0;
814   if( ExprHasProperty(pExpr, EP_FromJoin) ) return 0;
815   aff1 = sqlite3ExprAffinity(pExpr->pLeft);
816   aff2 = sqlite3ExprAffinity(pExpr->pRight);
817   if( aff1!=aff2
818    && (!sqlite3IsNumericAffinity(aff1) || !sqlite3IsNumericAffinity(aff2))
819   ){
820     return 0;
821   }
822   pColl = sqlite3BinaryCompareCollSeq(pParse, pExpr->pLeft, pExpr->pRight);
823   if( pColl==0 || sqlite3StrICmp(pColl->zName, "BINARY")==0 ) return 1;
824   pColl = sqlite3ExprCollSeq(pParse, pExpr->pLeft);
825   zColl1 = pColl ? pColl->zName : 0;
826   pColl = sqlite3ExprCollSeq(pParse, pExpr->pRight);
827   zColl2 = pColl ? pColl->zName : 0;
828   return sqlite3_stricmp(zColl1, zColl2)==0;
829 }
830 
831 /*
832 ** Recursively walk the expressions of a SELECT statement and generate
833 ** a bitmask indicating which tables are used in that expression
834 ** tree.
835 */
836 static Bitmask exprSelectUsage(WhereMaskSet *pMaskSet, Select *pS){
837   Bitmask mask = 0;
838   while( pS ){
839     SrcList *pSrc = pS->pSrc;
840     mask |= sqlite3WhereExprListUsage(pMaskSet, pS->pEList);
841     mask |= sqlite3WhereExprListUsage(pMaskSet, pS->pGroupBy);
842     mask |= sqlite3WhereExprListUsage(pMaskSet, pS->pOrderBy);
843     mask |= sqlite3WhereExprUsage(pMaskSet, pS->pWhere);
844     mask |= sqlite3WhereExprUsage(pMaskSet, pS->pHaving);
845     if( ALWAYS(pSrc!=0) ){
846       int i;
847       for(i=0; i<pSrc->nSrc; i++){
848         mask |= exprSelectUsage(pMaskSet, pSrc->a[i].pSelect);
849         mask |= sqlite3WhereExprUsage(pMaskSet, pSrc->a[i].pOn);
850       }
851     }
852     pS = pS->pPrior;
853   }
854   return mask;
855 }
856 
857 /*
858 ** Expression pExpr is one operand of a comparison operator that might
859 ** be useful for indexing.  This routine checks to see if pExpr appears
860 ** in any index.  Return TRUE (1) if pExpr is an indexed term and return
861 ** FALSE (0) if not.  If TRUE is returned, also set aiCurCol[0] to the cursor
862 ** number of the table that is indexed and aiCurCol[1] to the column number
863 ** of the column that is indexed, or XN_EXPR (-2) if an expression is being
864 ** indexed.
865 **
866 ** If pExpr is a TK_COLUMN column reference, then this routine always returns
867 ** true even if that particular column is not indexed, because the column
868 ** might be added to an automatic index later.
869 */
870 static SQLITE_NOINLINE int exprMightBeIndexed2(
871   SrcList *pFrom,        /* The FROM clause */
872   Bitmask mPrereq,       /* Bitmask of FROM clause terms referenced by pExpr */
873   int *aiCurCol,         /* Write the referenced table cursor and column here */
874   Expr *pExpr            /* An operand of a comparison operator */
875 ){
876   Index *pIdx;
877   int i;
878   int iCur;
879   for(i=0; mPrereq>1; i++, mPrereq>>=1){}
880   iCur = pFrom->a[i].iCursor;
881   for(pIdx=pFrom->a[i].pTab->pIndex; pIdx; pIdx=pIdx->pNext){
882     if( pIdx->aColExpr==0 ) continue;
883     for(i=0; i<pIdx->nKeyCol; i++){
884       if( pIdx->aiColumn[i]!=XN_EXPR ) continue;
885       if( sqlite3ExprCompareSkip(pExpr, pIdx->aColExpr->a[i].pExpr, iCur)==0 ){
886         aiCurCol[0] = iCur;
887         aiCurCol[1] = XN_EXPR;
888         return 1;
889       }
890     }
891   }
892   return 0;
893 }
894 static int exprMightBeIndexed(
895   SrcList *pFrom,        /* The FROM clause */
896   Bitmask mPrereq,       /* Bitmask of FROM clause terms referenced by pExpr */
897   int *aiCurCol,         /* Write the referenced table cursor & column here */
898   Expr *pExpr,           /* An operand of a comparison operator */
899   int op                 /* The specific comparison operator */
900 ){
901   /* If this expression is a vector to the left or right of a
902   ** inequality constraint (>, <, >= or <=), perform the processing
903   ** on the first element of the vector.  */
904   assert( TK_GT+1==TK_LE && TK_GT+2==TK_LT && TK_GT+3==TK_GE );
905   assert( TK_IS<TK_GE && TK_ISNULL<TK_GE && TK_IN<TK_GE );
906   assert( op<=TK_GE );
907   if( pExpr->op==TK_VECTOR && (op>=TK_GT && ALWAYS(op<=TK_GE)) ){
908     pExpr = pExpr->x.pList->a[0].pExpr;
909   }
910 
911   if( pExpr->op==TK_COLUMN ){
912     aiCurCol[0] = pExpr->iTable;
913     aiCurCol[1] = pExpr->iColumn;
914     return 1;
915   }
916   if( mPrereq==0 ) return 0;                 /* No table references */
917   if( (mPrereq&(mPrereq-1))!=0 ) return 0;   /* Refs more than one table */
918   return exprMightBeIndexed2(pFrom,mPrereq,aiCurCol,pExpr);
919 }
920 
921 /*
922 ** The input to this routine is an WhereTerm structure with only the
923 ** "pExpr" field filled in.  The job of this routine is to analyze the
924 ** subexpression and populate all the other fields of the WhereTerm
925 ** structure.
926 **
927 ** If the expression is of the form "<expr> <op> X" it gets commuted
928 ** to the standard form of "X <op> <expr>".
929 **
930 ** If the expression is of the form "X <op> Y" where both X and Y are
931 ** columns, then the original expression is unchanged and a new virtual
932 ** term of the form "Y <op> X" is added to the WHERE clause and
933 ** analyzed separately.  The original term is marked with TERM_COPIED
934 ** and the new term is marked with TERM_DYNAMIC (because it's pExpr
935 ** needs to be freed with the WhereClause) and TERM_VIRTUAL (because it
936 ** is a commuted copy of a prior term.)  The original term has nChild=1
937 ** and the copy has idxParent set to the index of the original term.
938 */
939 static void exprAnalyze(
940   SrcList *pSrc,            /* the FROM clause */
941   WhereClause *pWC,         /* the WHERE clause */
942   int idxTerm               /* Index of the term to be analyzed */
943 ){
944   WhereInfo *pWInfo = pWC->pWInfo; /* WHERE clause processing context */
945   WhereTerm *pTerm;                /* The term to be analyzed */
946   WhereMaskSet *pMaskSet;          /* Set of table index masks */
947   Expr *pExpr;                     /* The expression to be analyzed */
948   Bitmask prereqLeft;              /* Prerequesites of the pExpr->pLeft */
949   Bitmask prereqAll;               /* Prerequesites of pExpr */
950   Bitmask extraRight = 0;          /* Extra dependencies on LEFT JOIN */
951   Expr *pStr1 = 0;                 /* RHS of LIKE/GLOB operator */
952   int isComplete = 0;              /* RHS of LIKE/GLOB ends with wildcard */
953   int noCase = 0;                  /* uppercase equivalent to lowercase */
954   int op;                          /* Top-level operator.  pExpr->op */
955   Parse *pParse = pWInfo->pParse;  /* Parsing context */
956   sqlite3 *db = pParse->db;        /* Database connection */
957   unsigned char eOp2;              /* op2 value for LIKE/REGEXP/GLOB */
958   int nLeft;                       /* Number of elements on left side vector */
959 
960   if( db->mallocFailed ){
961     return;
962   }
963   pTerm = &pWC->a[idxTerm];
964   pMaskSet = &pWInfo->sMaskSet;
965   pExpr = pTerm->pExpr;
966   assert( pExpr->op!=TK_AS && pExpr->op!=TK_COLLATE );
967   prereqLeft = sqlite3WhereExprUsage(pMaskSet, pExpr->pLeft);
968   op = pExpr->op;
969   if( op==TK_IN ){
970     assert( pExpr->pRight==0 );
971     if( sqlite3ExprCheckIN(pParse, pExpr) ) return;
972     if( ExprHasProperty(pExpr, EP_xIsSelect) ){
973       pTerm->prereqRight = exprSelectUsage(pMaskSet, pExpr->x.pSelect);
974     }else{
975       pTerm->prereqRight = sqlite3WhereExprListUsage(pMaskSet, pExpr->x.pList);
976     }
977   }else if( op==TK_ISNULL ){
978     pTerm->prereqRight = 0;
979   }else{
980     pTerm->prereqRight = sqlite3WhereExprUsage(pMaskSet, pExpr->pRight);
981   }
982   pMaskSet->bVarSelect = 0;
983   prereqAll = sqlite3WhereExprUsage(pMaskSet, pExpr);
984   if( pMaskSet->bVarSelect ) pTerm->wtFlags |= TERM_VARSELECT;
985   if( ExprHasProperty(pExpr, EP_FromJoin) ){
986     Bitmask x = sqlite3WhereGetMask(pMaskSet, pExpr->iRightJoinTable);
987     prereqAll |= x;
988     extraRight = x-1;  /* ON clause terms may not be used with an index
989                        ** on left table of a LEFT JOIN.  Ticket #3015 */
990     if( (prereqAll>>1)>=x ){
991       sqlite3ErrorMsg(pParse, "ON clause references tables to its right");
992       return;
993     }
994   }
995   pTerm->prereqAll = prereqAll;
996   pTerm->leftCursor = -1;
997   pTerm->iParent = -1;
998   pTerm->eOperator = 0;
999   if( allowedOp(op) ){
1000     int aiCurCol[2];
1001     Expr *pLeft = sqlite3ExprSkipCollate(pExpr->pLeft);
1002     Expr *pRight = sqlite3ExprSkipCollate(pExpr->pRight);
1003     u16 opMask = (pTerm->prereqRight & prereqLeft)==0 ? WO_ALL : WO_EQUIV;
1004 
1005     if( pTerm->iField>0 ){
1006       assert( op==TK_IN );
1007       assert( pLeft->op==TK_VECTOR );
1008       pLeft = pLeft->x.pList->a[pTerm->iField-1].pExpr;
1009     }
1010 
1011     if( exprMightBeIndexed(pSrc, prereqLeft, aiCurCol, pLeft, op) ){
1012       pTerm->leftCursor = aiCurCol[0];
1013       pTerm->u.leftColumn = aiCurCol[1];
1014       pTerm->eOperator = operatorMask(op) & opMask;
1015     }
1016     if( op==TK_IS ) pTerm->wtFlags |= TERM_IS;
1017     if( pRight
1018      && exprMightBeIndexed(pSrc, pTerm->prereqRight, aiCurCol, pRight, op)
1019     ){
1020       WhereTerm *pNew;
1021       Expr *pDup;
1022       u16 eExtraOp = 0;        /* Extra bits for pNew->eOperator */
1023       assert( pTerm->iField==0 );
1024       if( pTerm->leftCursor>=0 ){
1025         int idxNew;
1026         pDup = sqlite3ExprDup(db, pExpr, 0);
1027         if( db->mallocFailed ){
1028           sqlite3ExprDelete(db, pDup);
1029           return;
1030         }
1031         idxNew = whereClauseInsert(pWC, pDup, TERM_VIRTUAL|TERM_DYNAMIC);
1032         if( idxNew==0 ) return;
1033         pNew = &pWC->a[idxNew];
1034         markTermAsChild(pWC, idxNew, idxTerm);
1035         if( op==TK_IS ) pNew->wtFlags |= TERM_IS;
1036         pTerm = &pWC->a[idxTerm];
1037         pTerm->wtFlags |= TERM_COPIED;
1038 
1039         if( termIsEquivalence(pParse, pDup) ){
1040           pTerm->eOperator |= WO_EQUIV;
1041           eExtraOp = WO_EQUIV;
1042         }
1043       }else{
1044         pDup = pExpr;
1045         pNew = pTerm;
1046       }
1047       exprCommute(pParse, pDup);
1048       pNew->leftCursor = aiCurCol[0];
1049       pNew->u.leftColumn = aiCurCol[1];
1050       testcase( (prereqLeft | extraRight) != prereqLeft );
1051       pNew->prereqRight = prereqLeft | extraRight;
1052       pNew->prereqAll = prereqAll;
1053       pNew->eOperator = (operatorMask(pDup->op) + eExtraOp) & opMask;
1054     }
1055   }
1056 
1057 #ifndef SQLITE_OMIT_BETWEEN_OPTIMIZATION
1058   /* If a term is the BETWEEN operator, create two new virtual terms
1059   ** that define the range that the BETWEEN implements.  For example:
1060   **
1061   **      a BETWEEN b AND c
1062   **
1063   ** is converted into:
1064   **
1065   **      (a BETWEEN b AND c) AND (a>=b) AND (a<=c)
1066   **
1067   ** The two new terms are added onto the end of the WhereClause object.
1068   ** The new terms are "dynamic" and are children of the original BETWEEN
1069   ** term.  That means that if the BETWEEN term is coded, the children are
1070   ** skipped.  Or, if the children are satisfied by an index, the original
1071   ** BETWEEN term is skipped.
1072   */
1073   else if( pExpr->op==TK_BETWEEN && pWC->op==TK_AND ){
1074     ExprList *pList = pExpr->x.pList;
1075     int i;
1076     static const u8 ops[] = {TK_GE, TK_LE};
1077     assert( pList!=0 );
1078     assert( pList->nExpr==2 );
1079     for(i=0; i<2; i++){
1080       Expr *pNewExpr;
1081       int idxNew;
1082       pNewExpr = sqlite3PExpr(pParse, ops[i],
1083                              sqlite3ExprDup(db, pExpr->pLeft, 0),
1084                              sqlite3ExprDup(db, pList->a[i].pExpr, 0));
1085       transferJoinMarkings(pNewExpr, pExpr);
1086       idxNew = whereClauseInsert(pWC, pNewExpr, TERM_VIRTUAL|TERM_DYNAMIC);
1087       testcase( idxNew==0 );
1088       exprAnalyze(pSrc, pWC, idxNew);
1089       pTerm = &pWC->a[idxTerm];
1090       markTermAsChild(pWC, idxNew, idxTerm);
1091     }
1092   }
1093 #endif /* SQLITE_OMIT_BETWEEN_OPTIMIZATION */
1094 
1095 #if !defined(SQLITE_OMIT_OR_OPTIMIZATION) && !defined(SQLITE_OMIT_SUBQUERY)
1096   /* Analyze a term that is composed of two or more subterms connected by
1097   ** an OR operator.
1098   */
1099   else if( pExpr->op==TK_OR ){
1100     assert( pWC->op==TK_AND );
1101     exprAnalyzeOrTerm(pSrc, pWC, idxTerm);
1102     pTerm = &pWC->a[idxTerm];
1103   }
1104 #endif /* SQLITE_OMIT_OR_OPTIMIZATION */
1105 
1106 #ifndef SQLITE_OMIT_LIKE_OPTIMIZATION
1107   /* Add constraints to reduce the search space on a LIKE or GLOB
1108   ** operator.
1109   **
1110   ** A like pattern of the form "x LIKE 'aBc%'" is changed into constraints
1111   **
1112   **          x>='ABC' AND x<'abd' AND x LIKE 'aBc%'
1113   **
1114   ** The last character of the prefix "abc" is incremented to form the
1115   ** termination condition "abd".  If case is not significant (the default
1116   ** for LIKE) then the lower-bound is made all uppercase and the upper-
1117   ** bound is made all lowercase so that the bounds also work when comparing
1118   ** BLOBs.
1119   */
1120   if( pWC->op==TK_AND
1121    && isLikeOrGlob(pParse, pExpr, &pStr1, &isComplete, &noCase)
1122   ){
1123     Expr *pLeft;       /* LHS of LIKE/GLOB operator */
1124     Expr *pStr2;       /* Copy of pStr1 - RHS of LIKE/GLOB operator */
1125     Expr *pNewExpr1;
1126     Expr *pNewExpr2;
1127     int idxNew1;
1128     int idxNew2;
1129     const char *zCollSeqName;     /* Name of collating sequence */
1130     const u16 wtFlags = TERM_LIKEOPT | TERM_VIRTUAL | TERM_DYNAMIC;
1131 
1132     pLeft = pExpr->x.pList->a[1].pExpr;
1133     pStr2 = sqlite3ExprDup(db, pStr1, 0);
1134 
1135     /* Convert the lower bound to upper-case and the upper bound to
1136     ** lower-case (upper-case is less than lower-case in ASCII) so that
1137     ** the range constraints also work for BLOBs
1138     */
1139     if( noCase && !pParse->db->mallocFailed ){
1140       int i;
1141       char c;
1142       pTerm->wtFlags |= TERM_LIKE;
1143       for(i=0; (c = pStr1->u.zToken[i])!=0; i++){
1144         pStr1->u.zToken[i] = sqlite3Toupper(c);
1145         pStr2->u.zToken[i] = sqlite3Tolower(c);
1146       }
1147     }
1148 
1149     if( !db->mallocFailed ){
1150       u8 c, *pC;       /* Last character before the first wildcard */
1151       pC = (u8*)&pStr2->u.zToken[sqlite3Strlen30(pStr2->u.zToken)-1];
1152       c = *pC;
1153       if( noCase ){
1154         /* The point is to increment the last character before the first
1155         ** wildcard.  But if we increment '@', that will push it into the
1156         ** alphabetic range where case conversions will mess up the
1157         ** inequality.  To avoid this, make sure to also run the full
1158         ** LIKE on all candidate expressions by clearing the isComplete flag
1159         */
1160         if( c=='A'-1 ) isComplete = 0;
1161         c = sqlite3UpperToLower[c];
1162       }
1163       *pC = c + 1;
1164     }
1165     zCollSeqName = noCase ? "NOCASE" : "BINARY";
1166     pNewExpr1 = sqlite3ExprDup(db, pLeft, 0);
1167     pNewExpr1 = sqlite3PExpr(pParse, TK_GE,
1168            sqlite3ExprAddCollateString(pParse,pNewExpr1,zCollSeqName),
1169            pStr1);
1170     transferJoinMarkings(pNewExpr1, pExpr);
1171     idxNew1 = whereClauseInsert(pWC, pNewExpr1, wtFlags);
1172     testcase( idxNew1==0 );
1173     exprAnalyze(pSrc, pWC, idxNew1);
1174     pNewExpr2 = sqlite3ExprDup(db, pLeft, 0);
1175     pNewExpr2 = sqlite3PExpr(pParse, TK_LT,
1176            sqlite3ExprAddCollateString(pParse,pNewExpr2,zCollSeqName),
1177            pStr2);
1178     transferJoinMarkings(pNewExpr2, pExpr);
1179     idxNew2 = whereClauseInsert(pWC, pNewExpr2, wtFlags);
1180     testcase( idxNew2==0 );
1181     exprAnalyze(pSrc, pWC, idxNew2);
1182     pTerm = &pWC->a[idxTerm];
1183     if( isComplete ){
1184       markTermAsChild(pWC, idxNew1, idxTerm);
1185       markTermAsChild(pWC, idxNew2, idxTerm);
1186     }
1187   }
1188 #endif /* SQLITE_OMIT_LIKE_OPTIMIZATION */
1189 
1190 #ifndef SQLITE_OMIT_VIRTUALTABLE
1191   /* Add a WO_MATCH auxiliary term to the constraint set if the
1192   ** current expression is of the form:  column MATCH expr.
1193   ** This information is used by the xBestIndex methods of
1194   ** virtual tables.  The native query optimizer does not attempt
1195   ** to do anything with MATCH functions.
1196   */
1197   if( pWC->op==TK_AND && isMatchOfColumn(pExpr, &eOp2) ){
1198     int idxNew;
1199     Expr *pRight, *pLeft;
1200     WhereTerm *pNewTerm;
1201     Bitmask prereqColumn, prereqExpr;
1202 
1203     pRight = pExpr->x.pList->a[0].pExpr;
1204     pLeft = pExpr->x.pList->a[1].pExpr;
1205     prereqExpr = sqlite3WhereExprUsage(pMaskSet, pRight);
1206     prereqColumn = sqlite3WhereExprUsage(pMaskSet, pLeft);
1207     if( (prereqExpr & prereqColumn)==0 ){
1208       Expr *pNewExpr;
1209       pNewExpr = sqlite3PExpr(pParse, TK_MATCH,
1210                               0, sqlite3ExprDup(db, pRight, 0));
1211       if( ExprHasProperty(pExpr, EP_FromJoin) && pNewExpr ){
1212         ExprSetProperty(pNewExpr, EP_FromJoin);
1213       }
1214       idxNew = whereClauseInsert(pWC, pNewExpr, TERM_VIRTUAL|TERM_DYNAMIC);
1215       testcase( idxNew==0 );
1216       pNewTerm = &pWC->a[idxNew];
1217       pNewTerm->prereqRight = prereqExpr;
1218       pNewTerm->leftCursor = pLeft->iTable;
1219       pNewTerm->u.leftColumn = pLeft->iColumn;
1220       pNewTerm->eOperator = WO_MATCH;
1221       pNewTerm->eMatchOp = eOp2;
1222       markTermAsChild(pWC, idxNew, idxTerm);
1223       pTerm = &pWC->a[idxTerm];
1224       pTerm->wtFlags |= TERM_COPIED;
1225       pNewTerm->prereqAll = pTerm->prereqAll;
1226     }
1227   }
1228 #endif /* SQLITE_OMIT_VIRTUALTABLE */
1229 
1230   /* If there is a vector == or IS term - e.g. "(a, b) == (?, ?)" - create
1231   ** new terms for each component comparison - "a = ?" and "b = ?".  The
1232   ** new terms completely replace the original vector comparison, which is
1233   ** no longer used.
1234   **
1235   ** This is only required if at least one side of the comparison operation
1236   ** is not a sub-select.  */
1237   if( pWC->op==TK_AND
1238   && (pExpr->op==TK_EQ || pExpr->op==TK_IS)
1239   && (nLeft = sqlite3ExprVectorSize(pExpr->pLeft))>1
1240   && sqlite3ExprVectorSize(pExpr->pRight)==nLeft
1241   && ( (pExpr->pLeft->flags & EP_xIsSelect)==0
1242     || (pExpr->pRight->flags & EP_xIsSelect)==0)
1243   ){
1244     int i;
1245     for(i=0; i<nLeft; i++){
1246       int idxNew;
1247       Expr *pNew;
1248       Expr *pLeft = sqlite3ExprForVectorField(pParse, pExpr->pLeft, i);
1249       Expr *pRight = sqlite3ExprForVectorField(pParse, pExpr->pRight, i);
1250 
1251       pNew = sqlite3PExpr(pParse, pExpr->op, pLeft, pRight);
1252       transferJoinMarkings(pNew, pExpr);
1253       idxNew = whereClauseInsert(pWC, pNew, TERM_DYNAMIC);
1254       exprAnalyze(pSrc, pWC, idxNew);
1255     }
1256     pTerm = &pWC->a[idxTerm];
1257     pTerm->wtFlags = TERM_CODED|TERM_VIRTUAL;  /* Disable the original */
1258     pTerm->eOperator = 0;
1259   }
1260 
1261   /* If there is a vector IN term - e.g. "(a, b) IN (SELECT ...)" - create
1262   ** a virtual term for each vector component. The expression object
1263   ** used by each such virtual term is pExpr (the full vector IN(...)
1264   ** expression). The WhereTerm.iField variable identifies the index within
1265   ** the vector on the LHS that the virtual term represents.
1266   **
1267   ** This only works if the RHS is a simple SELECT, not a compound
1268   */
1269   if( pWC->op==TK_AND && pExpr->op==TK_IN && pTerm->iField==0
1270    && pExpr->pLeft->op==TK_VECTOR
1271    && pExpr->x.pSelect->pPrior==0
1272   ){
1273     int i;
1274     for(i=0; i<sqlite3ExprVectorSize(pExpr->pLeft); i++){
1275       int idxNew;
1276       idxNew = whereClauseInsert(pWC, pExpr, TERM_VIRTUAL);
1277       pWC->a[idxNew].iField = i+1;
1278       exprAnalyze(pSrc, pWC, idxNew);
1279       markTermAsChild(pWC, idxNew, idxTerm);
1280     }
1281   }
1282 
1283 #ifdef SQLITE_ENABLE_STAT3_OR_STAT4
1284   /* When sqlite_stat3 histogram data is available an operator of the
1285   ** form "x IS NOT NULL" can sometimes be evaluated more efficiently
1286   ** as "x>NULL" if x is not an INTEGER PRIMARY KEY.  So construct a
1287   ** virtual term of that form.
1288   **
1289   ** Note that the virtual term must be tagged with TERM_VNULL.
1290   */
1291   if( pExpr->op==TK_NOTNULL
1292    && pExpr->pLeft->op==TK_COLUMN
1293    && pExpr->pLeft->iColumn>=0
1294    && OptimizationEnabled(db, SQLITE_Stat34)
1295   ){
1296     Expr *pNewExpr;
1297     Expr *pLeft = pExpr->pLeft;
1298     int idxNew;
1299     WhereTerm *pNewTerm;
1300 
1301     pNewExpr = sqlite3PExpr(pParse, TK_GT,
1302                             sqlite3ExprDup(db, pLeft, 0),
1303                             sqlite3ExprAlloc(db, TK_NULL, 0, 0));
1304 
1305     idxNew = whereClauseInsert(pWC, pNewExpr,
1306                               TERM_VIRTUAL|TERM_DYNAMIC|TERM_VNULL);
1307     if( idxNew ){
1308       pNewTerm = &pWC->a[idxNew];
1309       pNewTerm->prereqRight = 0;
1310       pNewTerm->leftCursor = pLeft->iTable;
1311       pNewTerm->u.leftColumn = pLeft->iColumn;
1312       pNewTerm->eOperator = WO_GT;
1313       markTermAsChild(pWC, idxNew, idxTerm);
1314       pTerm = &pWC->a[idxTerm];
1315       pTerm->wtFlags |= TERM_COPIED;
1316       pNewTerm->prereqAll = pTerm->prereqAll;
1317     }
1318   }
1319 #endif /* SQLITE_ENABLE_STAT3_OR_STAT4 */
1320 
1321   /* Prevent ON clause terms of a LEFT JOIN from being used to drive
1322   ** an index for tables to the left of the join.
1323   */
1324   testcase( pTerm!=&pWC->a[idxTerm] );
1325   pTerm = &pWC->a[idxTerm];
1326   pTerm->prereqRight |= extraRight;
1327 }
1328 
1329 /***************************************************************************
1330 ** Routines with file scope above.  Interface to the rest of the where.c
1331 ** subsystem follows.
1332 ***************************************************************************/
1333 
1334 /*
1335 ** This routine identifies subexpressions in the WHERE clause where
1336 ** each subexpression is separated by the AND operator or some other
1337 ** operator specified in the op parameter.  The WhereClause structure
1338 ** is filled with pointers to subexpressions.  For example:
1339 **
1340 **    WHERE  a=='hello' AND coalesce(b,11)<10 AND (c+12!=d OR c==22)
1341 **           \________/     \_______________/     \________________/
1342 **            slot[0]            slot[1]               slot[2]
1343 **
1344 ** The original WHERE clause in pExpr is unaltered.  All this routine
1345 ** does is make slot[] entries point to substructure within pExpr.
1346 **
1347 ** In the previous sentence and in the diagram, "slot[]" refers to
1348 ** the WhereClause.a[] array.  The slot[] array grows as needed to contain
1349 ** all terms of the WHERE clause.
1350 */
1351 void sqlite3WhereSplit(WhereClause *pWC, Expr *pExpr, u8 op){
1352   Expr *pE2 = sqlite3ExprSkipCollate(pExpr);
1353   pWC->op = op;
1354   if( pE2==0 ) return;
1355   if( pE2->op!=op ){
1356     whereClauseInsert(pWC, pExpr, 0);
1357   }else{
1358     sqlite3WhereSplit(pWC, pE2->pLeft, op);
1359     sqlite3WhereSplit(pWC, pE2->pRight, op);
1360   }
1361 }
1362 
1363 /*
1364 ** Initialize a preallocated WhereClause structure.
1365 */
1366 void sqlite3WhereClauseInit(
1367   WhereClause *pWC,        /* The WhereClause to be initialized */
1368   WhereInfo *pWInfo        /* The WHERE processing context */
1369 ){
1370   pWC->pWInfo = pWInfo;
1371   pWC->pOuter = 0;
1372   pWC->nTerm = 0;
1373   pWC->nSlot = ArraySize(pWC->aStatic);
1374   pWC->a = pWC->aStatic;
1375 }
1376 
1377 /*
1378 ** Deallocate a WhereClause structure.  The WhereClause structure
1379 ** itself is not freed.  This routine is the inverse of
1380 ** sqlite3WhereClauseInit().
1381 */
1382 void sqlite3WhereClauseClear(WhereClause *pWC){
1383   int i;
1384   WhereTerm *a;
1385   sqlite3 *db = pWC->pWInfo->pParse->db;
1386   for(i=pWC->nTerm-1, a=pWC->a; i>=0; i--, a++){
1387     if( a->wtFlags & TERM_DYNAMIC ){
1388       sqlite3ExprDelete(db, a->pExpr);
1389     }
1390     if( a->wtFlags & TERM_ORINFO ){
1391       whereOrInfoDelete(db, a->u.pOrInfo);
1392     }else if( a->wtFlags & TERM_ANDINFO ){
1393       whereAndInfoDelete(db, a->u.pAndInfo);
1394     }
1395   }
1396   if( pWC->a!=pWC->aStatic ){
1397     sqlite3DbFree(db, pWC->a);
1398   }
1399 }
1400 
1401 
1402 /*
1403 ** These routines walk (recursively) an expression tree and generate
1404 ** a bitmask indicating which tables are used in that expression
1405 ** tree.
1406 */
1407 Bitmask sqlite3WhereExprUsage(WhereMaskSet *pMaskSet, Expr *p){
1408   Bitmask mask;
1409   if( p==0 ) return 0;
1410   if( p->op==TK_COLUMN ){
1411     return sqlite3WhereGetMask(pMaskSet, p->iTable);
1412   }
1413   mask = (p->op==TK_IF_NULL_ROW) ? sqlite3WhereGetMask(pMaskSet, p->iTable) : 0;
1414   assert( !ExprHasProperty(p, EP_TokenOnly) );
1415   if( p->pLeft ) mask |= sqlite3WhereExprUsage(pMaskSet, p->pLeft);
1416   if( p->pRight ){
1417     mask |= sqlite3WhereExprUsage(pMaskSet, p->pRight);
1418     assert( p->x.pList==0 );
1419   }else if( ExprHasProperty(p, EP_xIsSelect) ){
1420     if( ExprHasProperty(p, EP_VarSelect) ) pMaskSet->bVarSelect = 1;
1421     mask |= exprSelectUsage(pMaskSet, p->x.pSelect);
1422   }else if( p->x.pList ){
1423     mask |= sqlite3WhereExprListUsage(pMaskSet, p->x.pList);
1424   }
1425   return mask;
1426 }
1427 Bitmask sqlite3WhereExprListUsage(WhereMaskSet *pMaskSet, ExprList *pList){
1428   int i;
1429   Bitmask mask = 0;
1430   if( pList ){
1431     for(i=0; i<pList->nExpr; i++){
1432       mask |= sqlite3WhereExprUsage(pMaskSet, pList->a[i].pExpr);
1433     }
1434   }
1435   return mask;
1436 }
1437 
1438 
1439 /*
1440 ** Call exprAnalyze on all terms in a WHERE clause.
1441 **
1442 ** Note that exprAnalyze() might add new virtual terms onto the
1443 ** end of the WHERE clause.  We do not want to analyze these new
1444 ** virtual terms, so start analyzing at the end and work forward
1445 ** so that the added virtual terms are never processed.
1446 */
1447 void sqlite3WhereExprAnalyze(
1448   SrcList *pTabList,       /* the FROM clause */
1449   WhereClause *pWC         /* the WHERE clause to be analyzed */
1450 ){
1451   int i;
1452   for(i=pWC->nTerm-1; i>=0; i--){
1453     exprAnalyze(pTabList, pWC, i);
1454   }
1455 }
1456 
1457 /*
1458 ** For table-valued-functions, transform the function arguments into
1459 ** new WHERE clause terms.
1460 **
1461 ** Each function argument translates into an equality constraint against
1462 ** a HIDDEN column in the table.
1463 */
1464 void sqlite3WhereTabFuncArgs(
1465   Parse *pParse,                    /* Parsing context */
1466   struct SrcList_item *pItem,       /* The FROM clause term to process */
1467   WhereClause *pWC                  /* Xfer function arguments to here */
1468 ){
1469   Table *pTab;
1470   int j, k;
1471   ExprList *pArgs;
1472   Expr *pColRef;
1473   Expr *pTerm;
1474   if( pItem->fg.isTabFunc==0 ) return;
1475   pTab = pItem->pTab;
1476   assert( pTab!=0 );
1477   pArgs = pItem->u1.pFuncArg;
1478   if( pArgs==0 ) return;
1479   for(j=k=0; j<pArgs->nExpr; j++){
1480     while( k<pTab->nCol && (pTab->aCol[k].colFlags & COLFLAG_HIDDEN)==0 ){k++;}
1481     if( k>=pTab->nCol ){
1482       sqlite3ErrorMsg(pParse, "too many arguments on %s() - max %d",
1483                       pTab->zName, j);
1484       return;
1485     }
1486     pColRef = sqlite3ExprAlloc(pParse->db, TK_COLUMN, 0, 0);
1487     if( pColRef==0 ) return;
1488     pColRef->iTable = pItem->iCursor;
1489     pColRef->iColumn = k++;
1490     pColRef->pTab = pTab;
1491     pTerm = sqlite3PExpr(pParse, TK_EQ, pColRef,
1492                          sqlite3ExprDup(pParse->db, pArgs->a[j].pExpr, 0));
1493     whereClauseInsert(pWC, pTerm, TERM_DYNAMIC);
1494   }
1495 }
1496