xref: /sqlite-3.40.0/src/whereexpr.c (revision 87f500ce)
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 char *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[3];                /* 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   if( pLeft->op!=TK_COLUMN
217    || sqlite3ExprAffinity(pLeft)!=SQLITE_AFF_TEXT
218    || IsVirtual(pLeft->pTab)  /* Value might be numeric */
219   ){
220     /* IMP: R-02065-49465 The left-hand side of the LIKE or GLOB operator must
221     ** be the name of an indexed column with TEXT affinity. */
222     return 0;
223   }
224   assert( pLeft->iColumn!=(-1) ); /* Because IPK never has AFF_TEXT */
225 
226   pRight = sqlite3ExprSkipCollate(pList->a[0].pExpr);
227   op = pRight->op;
228   if( op==TK_VARIABLE ){
229     Vdbe *pReprepare = pParse->pReprepare;
230     int iCol = pRight->iColumn;
231     pVal = sqlite3VdbeGetBoundValue(pReprepare, iCol, SQLITE_AFF_BLOB);
232     if( pVal && sqlite3_value_type(pVal)==SQLITE_TEXT ){
233       z = (char *)sqlite3_value_text(pVal);
234     }
235     sqlite3VdbeSetVarmask(pParse->pVdbe, iCol);
236     assert( pRight->op==TK_VARIABLE || pRight->op==TK_REGISTER );
237   }else if( op==TK_STRING ){
238     z = pRight->u.zToken;
239   }
240   if( z ){
241     cnt = 0;
242     while( (c=z[cnt])!=0 && c!=wc[0] && c!=wc[1] && c!=wc[2] ){
243       cnt++;
244     }
245     if( cnt!=0 && 255!=(u8)z[cnt-1] ){
246       Expr *pPrefix;
247       *pisComplete = c==wc[0] && z[cnt+1]==0;
248       pPrefix = sqlite3Expr(db, TK_STRING, z);
249       if( pPrefix ) pPrefix->u.zToken[cnt] = 0;
250       *ppPrefix = pPrefix;
251       if( op==TK_VARIABLE ){
252         Vdbe *v = pParse->pVdbe;
253         sqlite3VdbeSetVarmask(v, pRight->iColumn);
254         if( *pisComplete && pRight->u.zToken[1] ){
255           /* If the rhs of the LIKE expression is a variable, and the current
256           ** value of the variable means there is no need to invoke the LIKE
257           ** function, then no OP_Variable will be added to the program.
258           ** This causes problems for the sqlite3_bind_parameter_name()
259           ** API. To work around them, add a dummy OP_Variable here.
260           */
261           int r1 = sqlite3GetTempReg(pParse);
262           sqlite3ExprCodeTarget(pParse, pRight, r1);
263           sqlite3VdbeChangeP3(v, sqlite3VdbeCurrentAddr(v)-1, 0);
264           sqlite3ReleaseTempReg(pParse, r1);
265         }
266       }
267     }else{
268       z = 0;
269     }
270   }
271 
272   rc = (z!=0);
273   sqlite3ValueFree(pVal);
274   return rc;
275 }
276 #endif /* SQLITE_OMIT_LIKE_OPTIMIZATION */
277 
278 
279 #ifndef SQLITE_OMIT_VIRTUALTABLE
280 /*
281 ** Check to see if the given expression is of the form
282 **
283 **         column OP expr
284 **
285 ** where OP is one of MATCH, GLOB, LIKE or REGEXP and "column" is a
286 ** column of a virtual table.
287 **
288 ** If it is then return TRUE.  If not, return FALSE.
289 */
290 static int isMatchOfColumn(
291   Expr *pExpr,                    /* Test this expression */
292   unsigned char *peOp2            /* OUT: 0 for MATCH, or else an op2 value */
293 ){
294   static const struct Op2 {
295     const char *zOp;
296     unsigned char eOp2;
297   } aOp[] = {
298     { "match",  SQLITE_INDEX_CONSTRAINT_MATCH },
299     { "glob",   SQLITE_INDEX_CONSTRAINT_GLOB },
300     { "like",   SQLITE_INDEX_CONSTRAINT_LIKE },
301     { "regexp", SQLITE_INDEX_CONSTRAINT_REGEXP }
302   };
303   ExprList *pList;
304   Expr *pCol;                     /* Column reference */
305   int i;
306 
307   if( pExpr->op!=TK_FUNCTION ){
308     return 0;
309   }
310   pList = pExpr->x.pList;
311   if( pList==0 || pList->nExpr!=2 ){
312     return 0;
313   }
314   pCol = pList->a[1].pExpr;
315   if( pCol->op!=TK_COLUMN || !IsVirtual(pCol->pTab) ){
316     return 0;
317   }
318   for(i=0; i<ArraySize(aOp); i++){
319     if( sqlite3StrICmp(pExpr->u.zToken, aOp[i].zOp)==0 ){
320       *peOp2 = aOp[i].eOp2;
321       return 1;
322     }
323   }
324   return 0;
325 }
326 #endif /* SQLITE_OMIT_VIRTUALTABLE */
327 
328 /*
329 ** If the pBase expression originated in the ON or USING clause of
330 ** a join, then transfer the appropriate markings over to derived.
331 */
332 static void transferJoinMarkings(Expr *pDerived, Expr *pBase){
333   if( pDerived ){
334     pDerived->flags |= pBase->flags & EP_FromJoin;
335     pDerived->iRightJoinTable = pBase->iRightJoinTable;
336   }
337 }
338 
339 /*
340 ** Mark term iChild as being a child of term iParent
341 */
342 static void markTermAsChild(WhereClause *pWC, int iChild, int iParent){
343   pWC->a[iChild].iParent = iParent;
344   pWC->a[iChild].truthProb = pWC->a[iParent].truthProb;
345   pWC->a[iParent].nChild++;
346 }
347 
348 /*
349 ** Return the N-th AND-connected subterm of pTerm.  Or if pTerm is not
350 ** a conjunction, then return just pTerm when N==0.  If N is exceeds
351 ** the number of available subterms, return NULL.
352 */
353 static WhereTerm *whereNthSubterm(WhereTerm *pTerm, int N){
354   if( pTerm->eOperator!=WO_AND ){
355     return N==0 ? pTerm : 0;
356   }
357   if( N<pTerm->u.pAndInfo->wc.nTerm ){
358     return &pTerm->u.pAndInfo->wc.a[N];
359   }
360   return 0;
361 }
362 
363 /*
364 ** Subterms pOne and pTwo are contained within WHERE clause pWC.  The
365 ** two subterms are in disjunction - they are OR-ed together.
366 **
367 ** If these two terms are both of the form:  "A op B" with the same
368 ** A and B values but different operators and if the operators are
369 ** compatible (if one is = and the other is <, for example) then
370 ** add a new virtual AND term to pWC that is the combination of the
371 ** two.
372 **
373 ** Some examples:
374 **
375 **    x<y OR x=y    -->     x<=y
376 **    x=y OR x=y    -->     x=y
377 **    x<=y OR x<y   -->     x<=y
378 **
379 ** The following is NOT generated:
380 **
381 **    x<y OR x>y    -->     x!=y
382 */
383 static void whereCombineDisjuncts(
384   SrcList *pSrc,         /* the FROM clause */
385   WhereClause *pWC,      /* The complete WHERE clause */
386   WhereTerm *pOne,       /* First disjunct */
387   WhereTerm *pTwo        /* Second disjunct */
388 ){
389   u16 eOp = pOne->eOperator | pTwo->eOperator;
390   sqlite3 *db;           /* Database connection (for malloc) */
391   Expr *pNew;            /* New virtual expression */
392   int op;                /* Operator for the combined expression */
393   int idxNew;            /* Index in pWC of the next virtual term */
394 
395   if( (pOne->eOperator & (WO_EQ|WO_LT|WO_LE|WO_GT|WO_GE))==0 ) return;
396   if( (pTwo->eOperator & (WO_EQ|WO_LT|WO_LE|WO_GT|WO_GE))==0 ) return;
397   if( (eOp & (WO_EQ|WO_LT|WO_LE))!=eOp
398    && (eOp & (WO_EQ|WO_GT|WO_GE))!=eOp ) return;
399   assert( pOne->pExpr->pLeft!=0 && pOne->pExpr->pRight!=0 );
400   assert( pTwo->pExpr->pLeft!=0 && pTwo->pExpr->pRight!=0 );
401   if( sqlite3ExprCompare(pOne->pExpr->pLeft, pTwo->pExpr->pLeft, -1) ) return;
402   if( sqlite3ExprCompare(pOne->pExpr->pRight, pTwo->pExpr->pRight, -1) )return;
403   /* If we reach this point, it means the two subterms can be combined */
404   if( (eOp & (eOp-1))!=0 ){
405     if( eOp & (WO_LT|WO_LE) ){
406       eOp = WO_LE;
407     }else{
408       assert( eOp & (WO_GT|WO_GE) );
409       eOp = WO_GE;
410     }
411   }
412   db = pWC->pWInfo->pParse->db;
413   pNew = sqlite3ExprDup(db, pOne->pExpr, 0);
414   if( pNew==0 ) return;
415   for(op=TK_EQ; eOp!=(WO_EQ<<(op-TK_EQ)); op++){ assert( op<TK_GE ); }
416   pNew->op = op;
417   idxNew = whereClauseInsert(pWC, pNew, TERM_VIRTUAL|TERM_DYNAMIC);
418   exprAnalyze(pSrc, pWC, idxNew);
419 }
420 
421 #if !defined(SQLITE_OMIT_OR_OPTIMIZATION) && !defined(SQLITE_OMIT_SUBQUERY)
422 /*
423 ** Analyze a term that consists of two or more OR-connected
424 ** subterms.  So in:
425 **
426 **     ... WHERE  (a=5) AND (b=7 OR c=9 OR d=13) AND (d=13)
427 **                          ^^^^^^^^^^^^^^^^^^^^
428 **
429 ** This routine analyzes terms such as the middle term in the above example.
430 ** A WhereOrTerm object is computed and attached to the term under
431 ** analysis, regardless of the outcome of the analysis.  Hence:
432 **
433 **     WhereTerm.wtFlags   |=  TERM_ORINFO
434 **     WhereTerm.u.pOrInfo  =  a dynamically allocated WhereOrTerm object
435 **
436 ** The term being analyzed must have two or more of OR-connected subterms.
437 ** A single subterm might be a set of AND-connected sub-subterms.
438 ** Examples of terms under analysis:
439 **
440 **     (A)     t1.x=t2.y OR t1.x=t2.z OR t1.y=15 OR t1.z=t3.a+5
441 **     (B)     x=expr1 OR expr2=x OR x=expr3
442 **     (C)     t1.x=t2.y OR (t1.x=t2.z AND t1.y=15)
443 **     (D)     x=expr1 OR (y>11 AND y<22 AND z LIKE '*hello*')
444 **     (E)     (p.a=1 AND q.b=2 AND r.c=3) OR (p.x=4 AND q.y=5 AND r.z=6)
445 **     (F)     x>A OR (x=A AND y>=B)
446 **
447 ** CASE 1:
448 **
449 ** If all subterms are of the form T.C=expr for some single column of C and
450 ** a single table T (as shown in example B above) then create a new virtual
451 ** term that is an equivalent IN expression.  In other words, if the term
452 ** being analyzed is:
453 **
454 **      x = expr1  OR  expr2 = x  OR  x = expr3
455 **
456 ** then create a new virtual term like this:
457 **
458 **      x IN (expr1,expr2,expr3)
459 **
460 ** CASE 2:
461 **
462 ** If there are exactly two disjuncts and one side has x>A and the other side
463 ** has x=A (for the same x and A) then add a new virtual conjunct term to the
464 ** WHERE clause of the form "x>=A".  Example:
465 **
466 **      x>A OR (x=A AND y>B)    adds:    x>=A
467 **
468 ** The added conjunct can sometimes be helpful in query planning.
469 **
470 ** CASE 3:
471 **
472 ** If all subterms are indexable by a single table T, then set
473 **
474 **     WhereTerm.eOperator              =  WO_OR
475 **     WhereTerm.u.pOrInfo->indexable  |=  the cursor number for table T
476 **
477 ** A subterm is "indexable" if it is of the form
478 ** "T.C <op> <expr>" where C is any column of table T and
479 ** <op> is one of "=", "<", "<=", ">", ">=", "IS NULL", or "IN".
480 ** A subterm is also indexable if it is an AND of two or more
481 ** subsubterms at least one of which is indexable.  Indexable AND
482 ** subterms have their eOperator set to WO_AND and they have
483 ** u.pAndInfo set to a dynamically allocated WhereAndTerm object.
484 **
485 ** From another point of view, "indexable" means that the subterm could
486 ** potentially be used with an index if an appropriate index exists.
487 ** This analysis does not consider whether or not the index exists; that
488 ** is decided elsewhere.  This analysis only looks at whether subterms
489 ** appropriate for indexing exist.
490 **
491 ** All examples A through E above satisfy case 3.  But if a term
492 ** also satisfies case 1 (such as B) we know that the optimizer will
493 ** always prefer case 1, so in that case we pretend that case 3 is not
494 ** satisfied.
495 **
496 ** It might be the case that multiple tables are indexable.  For example,
497 ** (E) above is indexable on tables P, Q, and R.
498 **
499 ** Terms that satisfy case 3 are candidates for lookup by using
500 ** separate indices to find rowids for each subterm and composing
501 ** the union of all rowids using a RowSet object.  This is similar
502 ** to "bitmap indices" in other database engines.
503 **
504 ** OTHERWISE:
505 **
506 ** If none of cases 1, 2, or 3 apply, then leave the eOperator set to
507 ** zero.  This term is not useful for search.
508 */
509 static void exprAnalyzeOrTerm(
510   SrcList *pSrc,            /* the FROM clause */
511   WhereClause *pWC,         /* the complete WHERE clause */
512   int idxTerm               /* Index of the OR-term to be analyzed */
513 ){
514   WhereInfo *pWInfo = pWC->pWInfo;        /* WHERE clause processing context */
515   Parse *pParse = pWInfo->pParse;         /* Parser context */
516   sqlite3 *db = pParse->db;               /* Database connection */
517   WhereTerm *pTerm = &pWC->a[idxTerm];    /* The term to be analyzed */
518   Expr *pExpr = pTerm->pExpr;             /* The expression of the term */
519   int i;                                  /* Loop counters */
520   WhereClause *pOrWc;       /* Breakup of pTerm into subterms */
521   WhereTerm *pOrTerm;       /* A Sub-term within the pOrWc */
522   WhereOrInfo *pOrInfo;     /* Additional information associated with pTerm */
523   Bitmask chngToIN;         /* Tables that might satisfy case 1 */
524   Bitmask indexable;        /* Tables that are indexable, satisfying case 2 */
525 
526   /*
527   ** Break the OR clause into its separate subterms.  The subterms are
528   ** stored in a WhereClause structure containing within the WhereOrInfo
529   ** object that is attached to the original OR clause term.
530   */
531   assert( (pTerm->wtFlags & (TERM_DYNAMIC|TERM_ORINFO|TERM_ANDINFO))==0 );
532   assert( pExpr->op==TK_OR );
533   pTerm->u.pOrInfo = pOrInfo = sqlite3DbMallocZero(db, sizeof(*pOrInfo));
534   if( pOrInfo==0 ) return;
535   pTerm->wtFlags |= TERM_ORINFO;
536   pOrWc = &pOrInfo->wc;
537   memset(pOrWc->aStatic, 0, sizeof(pOrWc->aStatic));
538   sqlite3WhereClauseInit(pOrWc, pWInfo);
539   sqlite3WhereSplit(pOrWc, pExpr, TK_OR);
540   sqlite3WhereExprAnalyze(pSrc, pOrWc);
541   if( db->mallocFailed ) return;
542   assert( pOrWc->nTerm>=2 );
543 
544   /*
545   ** Compute the set of tables that might satisfy cases 1 or 3.
546   */
547   indexable = ~(Bitmask)0;
548   chngToIN = ~(Bitmask)0;
549   for(i=pOrWc->nTerm-1, pOrTerm=pOrWc->a; i>=0 && indexable; i--, pOrTerm++){
550     if( (pOrTerm->eOperator & WO_SINGLE)==0 ){
551       WhereAndInfo *pAndInfo;
552       assert( (pOrTerm->wtFlags & (TERM_ANDINFO|TERM_ORINFO))==0 );
553       chngToIN = 0;
554       pAndInfo = sqlite3DbMallocRawNN(db, sizeof(*pAndInfo));
555       if( pAndInfo ){
556         WhereClause *pAndWC;
557         WhereTerm *pAndTerm;
558         int j;
559         Bitmask b = 0;
560         pOrTerm->u.pAndInfo = pAndInfo;
561         pOrTerm->wtFlags |= TERM_ANDINFO;
562         pOrTerm->eOperator = WO_AND;
563         pAndWC = &pAndInfo->wc;
564         memset(pAndWC->aStatic, 0, sizeof(pAndWC->aStatic));
565         sqlite3WhereClauseInit(pAndWC, pWC->pWInfo);
566         sqlite3WhereSplit(pAndWC, pOrTerm->pExpr, TK_AND);
567         sqlite3WhereExprAnalyze(pSrc, pAndWC);
568         pAndWC->pOuter = pWC;
569         if( !db->mallocFailed ){
570           for(j=0, pAndTerm=pAndWC->a; j<pAndWC->nTerm; j++, pAndTerm++){
571             assert( pAndTerm->pExpr );
572             if( allowedOp(pAndTerm->pExpr->op)
573              || pAndTerm->eOperator==WO_MATCH
574             ){
575               b |= sqlite3WhereGetMask(&pWInfo->sMaskSet, pAndTerm->leftCursor);
576             }
577           }
578         }
579         indexable &= b;
580       }
581     }else if( pOrTerm->wtFlags & TERM_COPIED ){
582       /* Skip this term for now.  We revisit it when we process the
583       ** corresponding TERM_VIRTUAL term */
584     }else{
585       Bitmask b;
586       b = sqlite3WhereGetMask(&pWInfo->sMaskSet, pOrTerm->leftCursor);
587       if( pOrTerm->wtFlags & TERM_VIRTUAL ){
588         WhereTerm *pOther = &pOrWc->a[pOrTerm->iParent];
589         b |= sqlite3WhereGetMask(&pWInfo->sMaskSet, pOther->leftCursor);
590       }
591       indexable &= b;
592       if( (pOrTerm->eOperator & WO_EQ)==0 ){
593         chngToIN = 0;
594       }else{
595         chngToIN &= b;
596       }
597     }
598   }
599 
600   /*
601   ** Record the set of tables that satisfy case 3.  The set might be
602   ** empty.
603   */
604   pOrInfo->indexable = indexable;
605   pTerm->eOperator = indexable==0 ? 0 : WO_OR;
606 
607   /* For a two-way OR, attempt to implementation case 2.
608   */
609   if( indexable && pOrWc->nTerm==2 ){
610     int iOne = 0;
611     WhereTerm *pOne;
612     while( (pOne = whereNthSubterm(&pOrWc->a[0],iOne++))!=0 ){
613       int iTwo = 0;
614       WhereTerm *pTwo;
615       while( (pTwo = whereNthSubterm(&pOrWc->a[1],iTwo++))!=0 ){
616         whereCombineDisjuncts(pSrc, pWC, pOne, pTwo);
617       }
618     }
619   }
620 
621   /*
622   ** chngToIN holds a set of tables that *might* satisfy case 1.  But
623   ** we have to do some additional checking to see if case 1 really
624   ** is satisfied.
625   **
626   ** chngToIN will hold either 0, 1, or 2 bits.  The 0-bit case means
627   ** that there is no possibility of transforming the OR clause into an
628   ** IN operator because one or more terms in the OR clause contain
629   ** something other than == on a column in the single table.  The 1-bit
630   ** case means that every term of the OR clause is of the form
631   ** "table.column=expr" for some single table.  The one bit that is set
632   ** will correspond to the common table.  We still need to check to make
633   ** sure the same column is used on all terms.  The 2-bit case is when
634   ** the all terms are of the form "table1.column=table2.column".  It
635   ** might be possible to form an IN operator with either table1.column
636   ** or table2.column as the LHS if either is common to every term of
637   ** the OR clause.
638   **
639   ** Note that terms of the form "table.column1=table.column2" (the
640   ** same table on both sizes of the ==) cannot be optimized.
641   */
642   if( chngToIN ){
643     int okToChngToIN = 0;     /* True if the conversion to IN is valid */
644     int iColumn = -1;         /* Column index on lhs of IN operator */
645     int iCursor = -1;         /* Table cursor common to all terms */
646     int j = 0;                /* Loop counter */
647 
648     /* Search for a table and column that appears on one side or the
649     ** other of the == operator in every subterm.  That table and column
650     ** will be recorded in iCursor and iColumn.  There might not be any
651     ** such table and column.  Set okToChngToIN if an appropriate table
652     ** and column is found but leave okToChngToIN false if not found.
653     */
654     for(j=0; j<2 && !okToChngToIN; j++){
655       pOrTerm = pOrWc->a;
656       for(i=pOrWc->nTerm-1; i>=0; i--, pOrTerm++){
657         assert( pOrTerm->eOperator & WO_EQ );
658         pOrTerm->wtFlags &= ~TERM_OR_OK;
659         if( pOrTerm->leftCursor==iCursor ){
660           /* This is the 2-bit case and we are on the second iteration and
661           ** current term is from the first iteration.  So skip this term. */
662           assert( j==1 );
663           continue;
664         }
665         if( (chngToIN & sqlite3WhereGetMask(&pWInfo->sMaskSet,
666                                             pOrTerm->leftCursor))==0 ){
667           /* This term must be of the form t1.a==t2.b where t2 is in the
668           ** chngToIN set but t1 is not.  This term will be either preceded
669           ** or follwed by an inverted copy (t2.b==t1.a).  Skip this term
670           ** and use its inversion. */
671           testcase( pOrTerm->wtFlags & TERM_COPIED );
672           testcase( pOrTerm->wtFlags & TERM_VIRTUAL );
673           assert( pOrTerm->wtFlags & (TERM_COPIED|TERM_VIRTUAL) );
674           continue;
675         }
676         iColumn = pOrTerm->u.leftColumn;
677         iCursor = pOrTerm->leftCursor;
678         break;
679       }
680       if( i<0 ){
681         /* No candidate table+column was found.  This can only occur
682         ** on the second iteration */
683         assert( j==1 );
684         assert( IsPowerOfTwo(chngToIN) );
685         assert( chngToIN==sqlite3WhereGetMask(&pWInfo->sMaskSet, iCursor) );
686         break;
687       }
688       testcase( j==1 );
689 
690       /* We have found a candidate table and column.  Check to see if that
691       ** table and column is common to every term in the OR clause */
692       okToChngToIN = 1;
693       for(; i>=0 && okToChngToIN; i--, pOrTerm++){
694         assert( pOrTerm->eOperator & WO_EQ );
695         if( pOrTerm->leftCursor!=iCursor ){
696           pOrTerm->wtFlags &= ~TERM_OR_OK;
697         }else if( pOrTerm->u.leftColumn!=iColumn ){
698           okToChngToIN = 0;
699         }else{
700           int affLeft, affRight;
701           /* If the right-hand side is also a column, then the affinities
702           ** of both right and left sides must be such that no type
703           ** conversions are required on the right.  (Ticket #2249)
704           */
705           affRight = sqlite3ExprAffinity(pOrTerm->pExpr->pRight);
706           affLeft = sqlite3ExprAffinity(pOrTerm->pExpr->pLeft);
707           if( affRight!=0 && affRight!=affLeft ){
708             okToChngToIN = 0;
709           }else{
710             pOrTerm->wtFlags |= TERM_OR_OK;
711           }
712         }
713       }
714     }
715 
716     /* At this point, okToChngToIN is true if original pTerm satisfies
717     ** case 1.  In that case, construct a new virtual term that is
718     ** pTerm converted into an IN operator.
719     */
720     if( okToChngToIN ){
721       Expr *pDup;            /* A transient duplicate expression */
722       ExprList *pList = 0;   /* The RHS of the IN operator */
723       Expr *pLeft = 0;       /* The LHS of the IN operator */
724       Expr *pNew;            /* The complete IN operator */
725 
726       for(i=pOrWc->nTerm-1, pOrTerm=pOrWc->a; i>=0; i--, pOrTerm++){
727         if( (pOrTerm->wtFlags & TERM_OR_OK)==0 ) continue;
728         assert( pOrTerm->eOperator & WO_EQ );
729         assert( pOrTerm->leftCursor==iCursor );
730         assert( pOrTerm->u.leftColumn==iColumn );
731         pDup = sqlite3ExprDup(db, pOrTerm->pExpr->pRight, 0);
732         pList = sqlite3ExprListAppend(pWInfo->pParse, pList, pDup);
733         pLeft = pOrTerm->pExpr->pLeft;
734       }
735       assert( pLeft!=0 );
736       pDup = sqlite3ExprDup(db, pLeft, 0);
737       pNew = sqlite3PExpr(pParse, TK_IN, pDup, 0);
738       if( pNew ){
739         int idxNew;
740         transferJoinMarkings(pNew, pExpr);
741         assert( !ExprHasProperty(pNew, EP_xIsSelect) );
742         pNew->x.pList = pList;
743         idxNew = whereClauseInsert(pWC, pNew, TERM_VIRTUAL|TERM_DYNAMIC);
744         testcase( idxNew==0 );
745         exprAnalyze(pSrc, pWC, idxNew);
746         pTerm = &pWC->a[idxTerm];
747         markTermAsChild(pWC, idxNew, idxTerm);
748       }else{
749         sqlite3ExprListDelete(db, pList);
750       }
751       pTerm->eOperator = WO_NOOP;  /* case 1 trumps case 3 */
752     }
753   }
754 }
755 #endif /* !SQLITE_OMIT_OR_OPTIMIZATION && !SQLITE_OMIT_SUBQUERY */
756 
757 /*
758 ** We already know that pExpr is a binary operator where both operands are
759 ** column references.  This routine checks to see if pExpr is an equivalence
760 ** relation:
761 **   1.  The SQLITE_Transitive optimization must be enabled
762 **   2.  Must be either an == or an IS operator
763 **   3.  Not originating in the ON clause of an OUTER JOIN
764 **   4.  The affinities of A and B must be compatible
765 **   5a. Both operands use the same collating sequence OR
766 **   5b. The overall collating sequence is BINARY
767 ** If this routine returns TRUE, that means that the RHS can be substituted
768 ** for the LHS anyplace else in the WHERE clause where the LHS column occurs.
769 ** This is an optimization.  No harm comes from returning 0.  But if 1 is
770 ** returned when it should not be, then incorrect answers might result.
771 */
772 static int termIsEquivalence(Parse *pParse, Expr *pExpr){
773   char aff1, aff2;
774   CollSeq *pColl;
775   const char *zColl1, *zColl2;
776   if( !OptimizationEnabled(pParse->db, SQLITE_Transitive) ) return 0;
777   if( pExpr->op!=TK_EQ && pExpr->op!=TK_IS ) return 0;
778   if( ExprHasProperty(pExpr, EP_FromJoin) ) return 0;
779   aff1 = sqlite3ExprAffinity(pExpr->pLeft);
780   aff2 = sqlite3ExprAffinity(pExpr->pRight);
781   if( aff1!=aff2
782    && (!sqlite3IsNumericAffinity(aff1) || !sqlite3IsNumericAffinity(aff2))
783   ){
784     return 0;
785   }
786   pColl = sqlite3BinaryCompareCollSeq(pParse, pExpr->pLeft, pExpr->pRight);
787   if( pColl==0 || sqlite3StrICmp(pColl->zName, "BINARY")==0 ) return 1;
788   pColl = sqlite3ExprCollSeq(pParse, pExpr->pLeft);
789   zColl1 = pColl ? pColl->zName : 0;
790   pColl = sqlite3ExprCollSeq(pParse, pExpr->pRight);
791   zColl2 = pColl ? pColl->zName : 0;
792   return sqlite3_stricmp(zColl1, zColl2)==0;
793 }
794 
795 /*
796 ** Recursively walk the expressions of a SELECT statement and generate
797 ** a bitmask indicating which tables are used in that expression
798 ** tree.
799 */
800 static Bitmask exprSelectUsage(WhereMaskSet *pMaskSet, Select *pS){
801   Bitmask mask = 0;
802   while( pS ){
803     SrcList *pSrc = pS->pSrc;
804     mask |= sqlite3WhereExprListUsage(pMaskSet, pS->pEList);
805     mask |= sqlite3WhereExprListUsage(pMaskSet, pS->pGroupBy);
806     mask |= sqlite3WhereExprListUsage(pMaskSet, pS->pOrderBy);
807     mask |= sqlite3WhereExprUsage(pMaskSet, pS->pWhere);
808     mask |= sqlite3WhereExprUsage(pMaskSet, pS->pHaving);
809     if( ALWAYS(pSrc!=0) ){
810       int i;
811       for(i=0; i<pSrc->nSrc; i++){
812         mask |= exprSelectUsage(pMaskSet, pSrc->a[i].pSelect);
813         mask |= sqlite3WhereExprUsage(pMaskSet, pSrc->a[i].pOn);
814       }
815     }
816     pS = pS->pPrior;
817   }
818   return mask;
819 }
820 
821 /*
822 ** Expression pExpr is one operand of a comparison operator that might
823 ** be useful for indexing.  This routine checks to see if pExpr appears
824 ** in any index.  Return TRUE (1) if pExpr is an indexed term and return
825 ** FALSE (0) if not.  If TRUE is returned, also set *piCur to the cursor
826 ** number of the table that is indexed and *piColumn to the column number
827 ** of the column that is indexed, or XN_EXPR (-2) if an expression is being
828 ** indexed.
829 **
830 ** If pExpr is a TK_COLUMN column reference, then this routine always returns
831 ** true even if that particular column is not indexed, because the column
832 ** might be added to an automatic index later.
833 */
834 static int exprMightBeIndexed(
835   SrcList *pFrom,        /* The FROM clause */
836   int op,                /* The specific comparison operator */
837   Bitmask mPrereq,       /* Bitmask of FROM clause terms referenced by pExpr */
838   Expr *pExpr,           /* An operand of a comparison operator */
839   int *piCur,            /* Write the referenced table cursor number here */
840   int *piColumn          /* Write the referenced table column number here */
841 ){
842   Index *pIdx;
843   int i;
844   int iCur;
845 
846   /* If this expression is a vector to the left or right of a
847   ** inequality constraint (>, <, >= or <=), perform the processing
848   ** on the first element of the vector.  */
849   assert( TK_GT+1==TK_LE && TK_GT+2==TK_LT && TK_GT+3==TK_GE );
850   assert( TK_IS<TK_GE && TK_ISNULL<TK_GE && TK_IN<TK_GE );
851   assert( op<=TK_GE );
852   if( pExpr->op==TK_VECTOR && (op>=TK_GT && ALWAYS(op<=TK_GE)) ){
853     pExpr = pExpr->x.pList->a[0].pExpr;
854   }
855 
856   if( pExpr->op==TK_COLUMN ){
857     *piCur = pExpr->iTable;
858     *piColumn = pExpr->iColumn;
859     return 1;
860   }
861   if( mPrereq==0 ) return 0;                 /* No table references */
862   if( (mPrereq&(mPrereq-1))!=0 ) return 0;   /* Refs more than one table */
863   for(i=0; mPrereq>1; i++, mPrereq>>=1){}
864   iCur = pFrom->a[i].iCursor;
865   for(pIdx=pFrom->a[i].pTab->pIndex; pIdx; pIdx=pIdx->pNext){
866     if( pIdx->aColExpr==0 ) continue;
867     for(i=0; i<pIdx->nKeyCol; i++){
868       if( pIdx->aiColumn[i]!=XN_EXPR ) continue;
869       if( sqlite3ExprCompare(pExpr, pIdx->aColExpr->a[i].pExpr, iCur)==0 ){
870         *piCur = iCur;
871         *piColumn = XN_EXPR;
872         return 1;
873       }
874     }
875   }
876   return 0;
877 }
878 
879 /*
880 ** The input to this routine is an WhereTerm structure with only the
881 ** "pExpr" field filled in.  The job of this routine is to analyze the
882 ** subexpression and populate all the other fields of the WhereTerm
883 ** structure.
884 **
885 ** If the expression is of the form "<expr> <op> X" it gets commuted
886 ** to the standard form of "X <op> <expr>".
887 **
888 ** If the expression is of the form "X <op> Y" where both X and Y are
889 ** columns, then the original expression is unchanged and a new virtual
890 ** term of the form "Y <op> X" is added to the WHERE clause and
891 ** analyzed separately.  The original term is marked with TERM_COPIED
892 ** and the new term is marked with TERM_DYNAMIC (because it's pExpr
893 ** needs to be freed with the WhereClause) and TERM_VIRTUAL (because it
894 ** is a commuted copy of a prior term.)  The original term has nChild=1
895 ** and the copy has idxParent set to the index of the original term.
896 */
897 static void exprAnalyze(
898   SrcList *pSrc,            /* the FROM clause */
899   WhereClause *pWC,         /* the WHERE clause */
900   int idxTerm               /* Index of the term to be analyzed */
901 ){
902   WhereInfo *pWInfo = pWC->pWInfo; /* WHERE clause processing context */
903   WhereTerm *pTerm;                /* The term to be analyzed */
904   WhereMaskSet *pMaskSet;          /* Set of table index masks */
905   Expr *pExpr;                     /* The expression to be analyzed */
906   Bitmask prereqLeft;              /* Prerequesites of the pExpr->pLeft */
907   Bitmask prereqAll;               /* Prerequesites of pExpr */
908   Bitmask extraRight = 0;          /* Extra dependencies on LEFT JOIN */
909   Expr *pStr1 = 0;                 /* RHS of LIKE/GLOB operator */
910   int isComplete = 0;              /* RHS of LIKE/GLOB ends with wildcard */
911   int noCase = 0;                  /* uppercase equivalent to lowercase */
912   int op;                          /* Top-level operator.  pExpr->op */
913   Parse *pParse = pWInfo->pParse;  /* Parsing context */
914   sqlite3 *db = pParse->db;        /* Database connection */
915   unsigned char eOp2;              /* op2 value for LIKE/REGEXP/GLOB */
916   int nLeft;                       /* Number of elements on left side vector */
917 
918   if( db->mallocFailed ){
919     return;
920   }
921   pTerm = &pWC->a[idxTerm];
922   pMaskSet = &pWInfo->sMaskSet;
923   pExpr = pTerm->pExpr;
924   assert( pExpr->op!=TK_AS && pExpr->op!=TK_COLLATE );
925   prereqLeft = sqlite3WhereExprUsage(pMaskSet, pExpr->pLeft);
926   op = pExpr->op;
927   if( op==TK_IN ){
928     assert( pExpr->pRight==0 );
929     if( sqlite3ExprCheckIN(pParse, pExpr) ) return;
930     if( ExprHasProperty(pExpr, EP_xIsSelect) ){
931       pTerm->prereqRight = exprSelectUsage(pMaskSet, pExpr->x.pSelect);
932     }else{
933       pTerm->prereqRight = sqlite3WhereExprListUsage(pMaskSet, pExpr->x.pList);
934     }
935   }else if( op==TK_ISNULL ){
936     pTerm->prereqRight = 0;
937   }else{
938     pTerm->prereqRight = sqlite3WhereExprUsage(pMaskSet, pExpr->pRight);
939   }
940   prereqAll = sqlite3WhereExprUsage(pMaskSet, pExpr);
941   if( ExprHasProperty(pExpr, EP_FromJoin) ){
942     Bitmask x = sqlite3WhereGetMask(pMaskSet, pExpr->iRightJoinTable);
943     prereqAll |= x;
944     extraRight = x-1;  /* ON clause terms may not be used with an index
945                        ** on left table of a LEFT JOIN.  Ticket #3015 */
946     if( (prereqAll>>1)>=x ){
947       sqlite3ErrorMsg(pParse, "ON clause references tables to its right");
948       return;
949     }
950   }
951   pTerm->prereqAll = prereqAll;
952   pTerm->leftCursor = -1;
953   pTerm->iParent = -1;
954   pTerm->eOperator = 0;
955   if( allowedOp(op) ){
956     int iCur, iColumn;
957     Expr *pLeft = sqlite3ExprSkipCollate(pExpr->pLeft);
958     Expr *pRight = sqlite3ExprSkipCollate(pExpr->pRight);
959     u16 opMask = (pTerm->prereqRight & prereqLeft)==0 ? WO_ALL : WO_EQUIV;
960 
961     if( pTerm->iField>0 ){
962       assert( op==TK_IN );
963       assert( pLeft->op==TK_VECTOR );
964       pLeft = pLeft->x.pList->a[pTerm->iField-1].pExpr;
965     }
966 
967     if( exprMightBeIndexed(pSrc, op, prereqLeft, pLeft, &iCur, &iColumn) ){
968       pTerm->leftCursor = iCur;
969       pTerm->u.leftColumn = iColumn;
970       pTerm->eOperator = operatorMask(op) & opMask;
971     }
972     if( op==TK_IS ) pTerm->wtFlags |= TERM_IS;
973     if( pRight
974      && exprMightBeIndexed(pSrc, op, pTerm->prereqRight, pRight, &iCur,&iColumn)
975     ){
976       WhereTerm *pNew;
977       Expr *pDup;
978       u16 eExtraOp = 0;        /* Extra bits for pNew->eOperator */
979       assert( pTerm->iField==0 );
980       if( pTerm->leftCursor>=0 ){
981         int idxNew;
982         pDup = sqlite3ExprDup(db, pExpr, 0);
983         if( db->mallocFailed ){
984           sqlite3ExprDelete(db, pDup);
985           return;
986         }
987         idxNew = whereClauseInsert(pWC, pDup, TERM_VIRTUAL|TERM_DYNAMIC);
988         if( idxNew==0 ) return;
989         pNew = &pWC->a[idxNew];
990         markTermAsChild(pWC, idxNew, idxTerm);
991         if( op==TK_IS ) pNew->wtFlags |= TERM_IS;
992         pTerm = &pWC->a[idxTerm];
993         pTerm->wtFlags |= TERM_COPIED;
994 
995         if( termIsEquivalence(pParse, pDup) ){
996           pTerm->eOperator |= WO_EQUIV;
997           eExtraOp = WO_EQUIV;
998         }
999       }else{
1000         pDup = pExpr;
1001         pNew = pTerm;
1002       }
1003       exprCommute(pParse, pDup);
1004       pNew->leftCursor = iCur;
1005       pNew->u.leftColumn = iColumn;
1006       testcase( (prereqLeft | extraRight) != prereqLeft );
1007       pNew->prereqRight = prereqLeft | extraRight;
1008       pNew->prereqAll = prereqAll;
1009       pNew->eOperator = (operatorMask(pDup->op) + eExtraOp) & opMask;
1010     }
1011   }
1012 
1013 #ifndef SQLITE_OMIT_BETWEEN_OPTIMIZATION
1014   /* If a term is the BETWEEN operator, create two new virtual terms
1015   ** that define the range that the BETWEEN implements.  For example:
1016   **
1017   **      a BETWEEN b AND c
1018   **
1019   ** is converted into:
1020   **
1021   **      (a BETWEEN b AND c) AND (a>=b) AND (a<=c)
1022   **
1023   ** The two new terms are added onto the end of the WhereClause object.
1024   ** The new terms are "dynamic" and are children of the original BETWEEN
1025   ** term.  That means that if the BETWEEN term is coded, the children are
1026   ** skipped.  Or, if the children are satisfied by an index, the original
1027   ** BETWEEN term is skipped.
1028   */
1029   else if( pExpr->op==TK_BETWEEN && pWC->op==TK_AND ){
1030     ExprList *pList = pExpr->x.pList;
1031     int i;
1032     static const u8 ops[] = {TK_GE, TK_LE};
1033     assert( pList!=0 );
1034     assert( pList->nExpr==2 );
1035     for(i=0; i<2; i++){
1036       Expr *pNewExpr;
1037       int idxNew;
1038       pNewExpr = sqlite3PExpr(pParse, ops[i],
1039                              sqlite3ExprDup(db, pExpr->pLeft, 0),
1040                              sqlite3ExprDup(db, pList->a[i].pExpr, 0));
1041       transferJoinMarkings(pNewExpr, pExpr);
1042       idxNew = whereClauseInsert(pWC, pNewExpr, TERM_VIRTUAL|TERM_DYNAMIC);
1043       testcase( idxNew==0 );
1044       exprAnalyze(pSrc, pWC, idxNew);
1045       pTerm = &pWC->a[idxTerm];
1046       markTermAsChild(pWC, idxNew, idxTerm);
1047     }
1048   }
1049 #endif /* SQLITE_OMIT_BETWEEN_OPTIMIZATION */
1050 
1051 #if !defined(SQLITE_OMIT_OR_OPTIMIZATION) && !defined(SQLITE_OMIT_SUBQUERY)
1052   /* Analyze a term that is composed of two or more subterms connected by
1053   ** an OR operator.
1054   */
1055   else if( pExpr->op==TK_OR ){
1056     assert( pWC->op==TK_AND );
1057     exprAnalyzeOrTerm(pSrc, pWC, idxTerm);
1058     pTerm = &pWC->a[idxTerm];
1059   }
1060 #endif /* SQLITE_OMIT_OR_OPTIMIZATION */
1061 
1062 #ifndef SQLITE_OMIT_LIKE_OPTIMIZATION
1063   /* Add constraints to reduce the search space on a LIKE or GLOB
1064   ** operator.
1065   **
1066   ** A like pattern of the form "x LIKE 'aBc%'" is changed into constraints
1067   **
1068   **          x>='ABC' AND x<'abd' AND x LIKE 'aBc%'
1069   **
1070   ** The last character of the prefix "abc" is incremented to form the
1071   ** termination condition "abd".  If case is not significant (the default
1072   ** for LIKE) then the lower-bound is made all uppercase and the upper-
1073   ** bound is made all lowercase so that the bounds also work when comparing
1074   ** BLOBs.
1075   */
1076   if( pWC->op==TK_AND
1077    && isLikeOrGlob(pParse, pExpr, &pStr1, &isComplete, &noCase)
1078   ){
1079     Expr *pLeft;       /* LHS of LIKE/GLOB operator */
1080     Expr *pStr2;       /* Copy of pStr1 - RHS of LIKE/GLOB operator */
1081     Expr *pNewExpr1;
1082     Expr *pNewExpr2;
1083     int idxNew1;
1084     int idxNew2;
1085     const char *zCollSeqName;     /* Name of collating sequence */
1086     const u16 wtFlags = TERM_LIKEOPT | TERM_VIRTUAL | TERM_DYNAMIC;
1087 
1088     pLeft = pExpr->x.pList->a[1].pExpr;
1089     pStr2 = sqlite3ExprDup(db, pStr1, 0);
1090 
1091     /* Convert the lower bound to upper-case and the upper bound to
1092     ** lower-case (upper-case is less than lower-case in ASCII) so that
1093     ** the range constraints also work for BLOBs
1094     */
1095     if( noCase && !pParse->db->mallocFailed ){
1096       int i;
1097       char c;
1098       pTerm->wtFlags |= TERM_LIKE;
1099       for(i=0; (c = pStr1->u.zToken[i])!=0; i++){
1100         pStr1->u.zToken[i] = sqlite3Toupper(c);
1101         pStr2->u.zToken[i] = sqlite3Tolower(c);
1102       }
1103     }
1104 
1105     if( !db->mallocFailed ){
1106       u8 c, *pC;       /* Last character before the first wildcard */
1107       pC = (u8*)&pStr2->u.zToken[sqlite3Strlen30(pStr2->u.zToken)-1];
1108       c = *pC;
1109       if( noCase ){
1110         /* The point is to increment the last character before the first
1111         ** wildcard.  But if we increment '@', that will push it into the
1112         ** alphabetic range where case conversions will mess up the
1113         ** inequality.  To avoid this, make sure to also run the full
1114         ** LIKE on all candidate expressions by clearing the isComplete flag
1115         */
1116         if( c=='A'-1 ) isComplete = 0;
1117         c = sqlite3UpperToLower[c];
1118       }
1119       *pC = c + 1;
1120     }
1121     zCollSeqName = noCase ? "NOCASE" : "BINARY";
1122     pNewExpr1 = sqlite3ExprDup(db, pLeft, 0);
1123     pNewExpr1 = sqlite3PExpr(pParse, TK_GE,
1124            sqlite3ExprAddCollateString(pParse,pNewExpr1,zCollSeqName),
1125            pStr1);
1126     transferJoinMarkings(pNewExpr1, pExpr);
1127     idxNew1 = whereClauseInsert(pWC, pNewExpr1, wtFlags);
1128     testcase( idxNew1==0 );
1129     exprAnalyze(pSrc, pWC, idxNew1);
1130     pNewExpr2 = sqlite3ExprDup(db, pLeft, 0);
1131     pNewExpr2 = sqlite3PExpr(pParse, TK_LT,
1132            sqlite3ExprAddCollateString(pParse,pNewExpr2,zCollSeqName),
1133            pStr2);
1134     transferJoinMarkings(pNewExpr2, pExpr);
1135     idxNew2 = whereClauseInsert(pWC, pNewExpr2, wtFlags);
1136     testcase( idxNew2==0 );
1137     exprAnalyze(pSrc, pWC, idxNew2);
1138     pTerm = &pWC->a[idxTerm];
1139     if( isComplete ){
1140       markTermAsChild(pWC, idxNew1, idxTerm);
1141       markTermAsChild(pWC, idxNew2, idxTerm);
1142     }
1143   }
1144 #endif /* SQLITE_OMIT_LIKE_OPTIMIZATION */
1145 
1146 #ifndef SQLITE_OMIT_VIRTUALTABLE
1147   /* Add a WO_MATCH auxiliary term to the constraint set if the
1148   ** current expression is of the form:  column MATCH expr.
1149   ** This information is used by the xBestIndex methods of
1150   ** virtual tables.  The native query optimizer does not attempt
1151   ** to do anything with MATCH functions.
1152   */
1153   if( pWC->op==TK_AND && isMatchOfColumn(pExpr, &eOp2) ){
1154     int idxNew;
1155     Expr *pRight, *pLeft;
1156     WhereTerm *pNewTerm;
1157     Bitmask prereqColumn, prereqExpr;
1158 
1159     pRight = pExpr->x.pList->a[0].pExpr;
1160     pLeft = pExpr->x.pList->a[1].pExpr;
1161     prereqExpr = sqlite3WhereExprUsage(pMaskSet, pRight);
1162     prereqColumn = sqlite3WhereExprUsage(pMaskSet, pLeft);
1163     if( (prereqExpr & prereqColumn)==0 ){
1164       Expr *pNewExpr;
1165       pNewExpr = sqlite3PExpr(pParse, TK_MATCH,
1166                               0, sqlite3ExprDup(db, pRight, 0));
1167       idxNew = whereClauseInsert(pWC, pNewExpr, TERM_VIRTUAL|TERM_DYNAMIC);
1168       testcase( idxNew==0 );
1169       pNewTerm = &pWC->a[idxNew];
1170       pNewTerm->prereqRight = prereqExpr;
1171       pNewTerm->leftCursor = pLeft->iTable;
1172       pNewTerm->u.leftColumn = pLeft->iColumn;
1173       pNewTerm->eOperator = WO_MATCH;
1174       pNewTerm->eMatchOp = eOp2;
1175       markTermAsChild(pWC, idxNew, idxTerm);
1176       pTerm = &pWC->a[idxTerm];
1177       pTerm->wtFlags |= TERM_COPIED;
1178       pNewTerm->prereqAll = pTerm->prereqAll;
1179     }
1180   }
1181 #endif /* SQLITE_OMIT_VIRTUALTABLE */
1182 
1183   /* If there is a vector == or IS term - e.g. "(a, b) == (?, ?)" - create
1184   ** new terms for each component comparison - "a = ?" and "b = ?".  The
1185   ** new terms completely replace the original vector comparison, which is
1186   ** no longer used.
1187   **
1188   ** This is only required if at least one side of the comparison operation
1189   ** is not a sub-select.  */
1190   if( pWC->op==TK_AND
1191   && (pExpr->op==TK_EQ || pExpr->op==TK_IS)
1192   && (nLeft = sqlite3ExprVectorSize(pExpr->pLeft))>1
1193   && sqlite3ExprVectorSize(pExpr->pRight)==nLeft
1194   && ( (pExpr->pLeft->flags & EP_xIsSelect)==0
1195     || (pExpr->pRight->flags & EP_xIsSelect)==0)
1196   ){
1197     int i;
1198     for(i=0; i<nLeft; i++){
1199       int idxNew;
1200       Expr *pNew;
1201       Expr *pLeft = sqlite3ExprForVectorField(pParse, pExpr->pLeft, i);
1202       Expr *pRight = sqlite3ExprForVectorField(pParse, pExpr->pRight, i);
1203 
1204       pNew = sqlite3PExpr(pParse, pExpr->op, pLeft, pRight);
1205       transferJoinMarkings(pNew, pExpr);
1206       idxNew = whereClauseInsert(pWC, pNew, TERM_DYNAMIC);
1207       exprAnalyze(pSrc, pWC, idxNew);
1208     }
1209     pTerm = &pWC->a[idxTerm];
1210     pTerm->wtFlags = TERM_CODED|TERM_VIRTUAL;  /* Disable the original */
1211     pTerm->eOperator = 0;
1212   }
1213 
1214   /* If there is a vector IN term - e.g. "(a, b) IN (SELECT ...)" - create
1215   ** a virtual term for each vector component. The expression object
1216   ** used by each such virtual term is pExpr (the full vector IN(...)
1217   ** expression). The WhereTerm.iField variable identifies the index within
1218   ** the vector on the LHS that the virtual term represents.
1219   **
1220   ** This only works if the RHS is a simple SELECT, not a compound
1221   */
1222   if( pWC->op==TK_AND && pExpr->op==TK_IN && pTerm->iField==0
1223    && pExpr->pLeft->op==TK_VECTOR
1224    && pExpr->x.pSelect->pPrior==0
1225   ){
1226     int i;
1227     for(i=0; i<sqlite3ExprVectorSize(pExpr->pLeft); i++){
1228       int idxNew;
1229       idxNew = whereClauseInsert(pWC, pExpr, TERM_VIRTUAL);
1230       pWC->a[idxNew].iField = i+1;
1231       exprAnalyze(pSrc, pWC, idxNew);
1232       markTermAsChild(pWC, idxNew, idxTerm);
1233     }
1234   }
1235 
1236 #ifdef SQLITE_ENABLE_STAT3_OR_STAT4
1237   /* When sqlite_stat3 histogram data is available an operator of the
1238   ** form "x IS NOT NULL" can sometimes be evaluated more efficiently
1239   ** as "x>NULL" if x is not an INTEGER PRIMARY KEY.  So construct a
1240   ** virtual term of that form.
1241   **
1242   ** Note that the virtual term must be tagged with TERM_VNULL.
1243   */
1244   if( pExpr->op==TK_NOTNULL
1245    && pExpr->pLeft->op==TK_COLUMN
1246    && pExpr->pLeft->iColumn>=0
1247    && OptimizationEnabled(db, SQLITE_Stat34)
1248   ){
1249     Expr *pNewExpr;
1250     Expr *pLeft = pExpr->pLeft;
1251     int idxNew;
1252     WhereTerm *pNewTerm;
1253 
1254     pNewExpr = sqlite3PExpr(pParse, TK_GT,
1255                             sqlite3ExprDup(db, pLeft, 0),
1256                             sqlite3ExprAlloc(db, TK_NULL, 0, 0));
1257 
1258     idxNew = whereClauseInsert(pWC, pNewExpr,
1259                               TERM_VIRTUAL|TERM_DYNAMIC|TERM_VNULL);
1260     if( idxNew ){
1261       pNewTerm = &pWC->a[idxNew];
1262       pNewTerm->prereqRight = 0;
1263       pNewTerm->leftCursor = pLeft->iTable;
1264       pNewTerm->u.leftColumn = pLeft->iColumn;
1265       pNewTerm->eOperator = WO_GT;
1266       markTermAsChild(pWC, idxNew, idxTerm);
1267       pTerm = &pWC->a[idxTerm];
1268       pTerm->wtFlags |= TERM_COPIED;
1269       pNewTerm->prereqAll = pTerm->prereqAll;
1270     }
1271   }
1272 #endif /* SQLITE_ENABLE_STAT3_OR_STAT4 */
1273 
1274   /* Prevent ON clause terms of a LEFT JOIN from being used to drive
1275   ** an index for tables to the left of the join.
1276   */
1277   testcase( pTerm!=&pWC->a[idxTerm] );
1278   pTerm = &pWC->a[idxTerm];
1279   pTerm->prereqRight |= extraRight;
1280 }
1281 
1282 /***************************************************************************
1283 ** Routines with file scope above.  Interface to the rest of the where.c
1284 ** subsystem follows.
1285 ***************************************************************************/
1286 
1287 /*
1288 ** This routine identifies subexpressions in the WHERE clause where
1289 ** each subexpression is separated by the AND operator or some other
1290 ** operator specified in the op parameter.  The WhereClause structure
1291 ** is filled with pointers to subexpressions.  For example:
1292 **
1293 **    WHERE  a=='hello' AND coalesce(b,11)<10 AND (c+12!=d OR c==22)
1294 **           \________/     \_______________/     \________________/
1295 **            slot[0]            slot[1]               slot[2]
1296 **
1297 ** The original WHERE clause in pExpr is unaltered.  All this routine
1298 ** does is make slot[] entries point to substructure within pExpr.
1299 **
1300 ** In the previous sentence and in the diagram, "slot[]" refers to
1301 ** the WhereClause.a[] array.  The slot[] array grows as needed to contain
1302 ** all terms of the WHERE clause.
1303 */
1304 void sqlite3WhereSplit(WhereClause *pWC, Expr *pExpr, u8 op){
1305   Expr *pE2 = sqlite3ExprSkipCollate(pExpr);
1306   pWC->op = op;
1307   if( pE2==0 ) return;
1308   if( pE2->op!=op ){
1309     whereClauseInsert(pWC, pExpr, 0);
1310   }else{
1311     sqlite3WhereSplit(pWC, pE2->pLeft, op);
1312     sqlite3WhereSplit(pWC, pE2->pRight, op);
1313   }
1314 }
1315 
1316 /*
1317 ** Initialize a preallocated WhereClause structure.
1318 */
1319 void sqlite3WhereClauseInit(
1320   WhereClause *pWC,        /* The WhereClause to be initialized */
1321   WhereInfo *pWInfo        /* The WHERE processing context */
1322 ){
1323   pWC->pWInfo = pWInfo;
1324   pWC->pOuter = 0;
1325   pWC->nTerm = 0;
1326   pWC->nSlot = ArraySize(pWC->aStatic);
1327   pWC->a = pWC->aStatic;
1328 }
1329 
1330 /*
1331 ** Deallocate a WhereClause structure.  The WhereClause structure
1332 ** itself is not freed.  This routine is the inverse of
1333 ** sqlite3WhereClauseInit().
1334 */
1335 void sqlite3WhereClauseClear(WhereClause *pWC){
1336   int i;
1337   WhereTerm *a;
1338   sqlite3 *db = pWC->pWInfo->pParse->db;
1339   for(i=pWC->nTerm-1, a=pWC->a; i>=0; i--, a++){
1340     if( a->wtFlags & TERM_DYNAMIC ){
1341       sqlite3ExprDelete(db, a->pExpr);
1342     }
1343     if( a->wtFlags & TERM_ORINFO ){
1344       whereOrInfoDelete(db, a->u.pOrInfo);
1345     }else if( a->wtFlags & TERM_ANDINFO ){
1346       whereAndInfoDelete(db, a->u.pAndInfo);
1347     }
1348   }
1349   if( pWC->a!=pWC->aStatic ){
1350     sqlite3DbFree(db, pWC->a);
1351   }
1352 }
1353 
1354 
1355 /*
1356 ** These routines walk (recursively) an expression tree and generate
1357 ** a bitmask indicating which tables are used in that expression
1358 ** tree.
1359 */
1360 Bitmask sqlite3WhereExprUsage(WhereMaskSet *pMaskSet, Expr *p){
1361   Bitmask mask;
1362   if( p==0 ) return 0;
1363   if( p->op==TK_COLUMN ){
1364     mask = sqlite3WhereGetMask(pMaskSet, p->iTable);
1365     return mask;
1366   }
1367   assert( !ExprHasProperty(p, EP_TokenOnly) );
1368   mask = p->pRight ? sqlite3WhereExprUsage(pMaskSet, p->pRight) : 0;
1369   if( p->pLeft ) mask |= sqlite3WhereExprUsage(pMaskSet, p->pLeft);
1370   if( ExprHasProperty(p, EP_xIsSelect) ){
1371     mask |= exprSelectUsage(pMaskSet, p->x.pSelect);
1372   }else if( p->x.pList ){
1373     mask |= sqlite3WhereExprListUsage(pMaskSet, p->x.pList);
1374   }
1375   return mask;
1376 }
1377 Bitmask sqlite3WhereExprListUsage(WhereMaskSet *pMaskSet, ExprList *pList){
1378   int i;
1379   Bitmask mask = 0;
1380   if( pList ){
1381     for(i=0; i<pList->nExpr; i++){
1382       mask |= sqlite3WhereExprUsage(pMaskSet, pList->a[i].pExpr);
1383     }
1384   }
1385   return mask;
1386 }
1387 
1388 
1389 /*
1390 ** Call exprAnalyze on all terms in a WHERE clause.
1391 **
1392 ** Note that exprAnalyze() might add new virtual terms onto the
1393 ** end of the WHERE clause.  We do not want to analyze these new
1394 ** virtual terms, so start analyzing at the end and work forward
1395 ** so that the added virtual terms are never processed.
1396 */
1397 void sqlite3WhereExprAnalyze(
1398   SrcList *pTabList,       /* the FROM clause */
1399   WhereClause *pWC         /* the WHERE clause to be analyzed */
1400 ){
1401   int i;
1402   for(i=pWC->nTerm-1; i>=0; i--){
1403     exprAnalyze(pTabList, pWC, i);
1404   }
1405 }
1406 
1407 /*
1408 ** For table-valued-functions, transform the function arguments into
1409 ** new WHERE clause terms.
1410 **
1411 ** Each function argument translates into an equality constraint against
1412 ** a HIDDEN column in the table.
1413 */
1414 void sqlite3WhereTabFuncArgs(
1415   Parse *pParse,                    /* Parsing context */
1416   struct SrcList_item *pItem,       /* The FROM clause term to process */
1417   WhereClause *pWC                  /* Xfer function arguments to here */
1418 ){
1419   Table *pTab;
1420   int j, k;
1421   ExprList *pArgs;
1422   Expr *pColRef;
1423   Expr *pTerm;
1424   if( pItem->fg.isTabFunc==0 ) return;
1425   pTab = pItem->pTab;
1426   assert( pTab!=0 );
1427   pArgs = pItem->u1.pFuncArg;
1428   if( pArgs==0 ) return;
1429   for(j=k=0; j<pArgs->nExpr; j++){
1430     while( k<pTab->nCol && (pTab->aCol[k].colFlags & COLFLAG_HIDDEN)==0 ){k++;}
1431     if( k>=pTab->nCol ){
1432       sqlite3ErrorMsg(pParse, "too many arguments on %s() - max %d",
1433                       pTab->zName, j);
1434       return;
1435     }
1436     pColRef = sqlite3ExprAlloc(pParse->db, TK_COLUMN, 0, 0);
1437     if( pColRef==0 ) return;
1438     pColRef->iTable = pItem->iCursor;
1439     pColRef->iColumn = k++;
1440     pColRef->pTab = pTab;
1441     pTerm = sqlite3PExpr(pParse, TK_EQ, pColRef,
1442                          sqlite3ExprDup(pParse->db, pArgs->a[j].pExpr, 0));
1443     whereClauseInsert(pWC, pTerm, TERM_DYNAMIC);
1444   }
1445 }
1446