4 ** The author disclaims copyright to this source code. In place of
5 ** a legal notice, here is a blessing:
7 ** May you do good and not evil.
8 ** May you find forgiveness for yourself and forgive others.
9 ** May you share freely, never taking more than you give.
11 *************************************************************************
12 ** This module contains C code that generates VDBE code used to process
13 ** the WHERE clause of SQL statements.
15 ** This file was originally part of where.c but was split out to improve
16 ** readability and editabiliity. This file contains utility routines for
17 ** analyzing Expr objects in the WHERE clause.
19 #include "sqliteInt.h"
22 /* Forward declarations */
23 static void exprAnalyze(SrcList
*, WhereClause
*, int);
26 ** Deallocate all memory associated with a WhereOrInfo object.
28 static void whereOrInfoDelete(sqlite3
*db
, WhereOrInfo
*p
){
29 sqlite3WhereClauseClear(&p
->wc
);
34 ** Deallocate all memory associated with a WhereAndInfo object.
36 static void whereAndInfoDelete(sqlite3
*db
, WhereAndInfo
*p
){
37 sqlite3WhereClauseClear(&p
->wc
);
42 ** Add a single new WhereTerm entry to the WhereClause object pWC.
43 ** The new WhereTerm object is constructed from Expr p and with wtFlags.
44 ** The index in pWC->a[] of the new WhereTerm is returned on success.
45 ** 0 is returned if the new WhereTerm could not be added due to a memory
46 ** allocation error. The memory allocation failure will be recorded in
47 ** the db->mallocFailed flag so that higher-level functions can detect it.
49 ** This routine will increase the size of the pWC->a[] array as necessary.
51 ** If the wtFlags argument includes TERM_DYNAMIC, then responsibility
52 ** for freeing the expression p is assumed by the WhereClause object pWC.
53 ** This is true even if this routine fails to allocate a new WhereTerm.
55 ** WARNING: This routine might reallocate the space used to store
56 ** WhereTerms. All pointers to WhereTerms should be invalidated after
57 ** calling this routine. Such pointers may be reinitialized by referencing
58 ** the pWC->a[] array.
60 static int whereClauseInsert(WhereClause
*pWC
, Expr
*p
, u16 wtFlags
){
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 );
69 if( wtFlags
& TERM_DYNAMIC
){
70 sqlite3ExprDelete(db
, p
);
75 memcpy(pWC
->a
, pOld
, sizeof(pWC
->a
[0])*pWC
->nTerm
);
76 if( pOld
!=pWC
->aStatic
){
77 sqlite3DbFree(db
, pOld
);
79 pWC
->nSlot
= sqlite3DbMallocSize(db
, pWC
->a
)/sizeof(pWC
->a
[0]);
81 pTerm
= &pWC
->a
[idx
= pWC
->nTerm
++];
82 if( p
&& ExprHasProperty(p
, EP_Unlikely
) ){
83 pTerm
->truthProb
= sqlite3LogEst(p
->iTable
) - 270;
87 pTerm
->pExpr
= sqlite3ExprSkipCollateAndLikely(p
);
88 pTerm
->wtFlags
= wtFlags
;
91 memset(&pTerm
->eOperator
, 0,
92 sizeof(WhereTerm
) - offsetof(WhereTerm
,eOperator
));
97 ** Return TRUE if the given operator is one of the operators that is
98 ** allowed for an indexable WHERE clause term. The allowed operators are
99 ** "=", "<", ">", "<=", ">=", "IN", "IS", and "IS NULL"
101 static int allowedOp(int op
){
102 assert( TK_GT
>TK_EQ
&& TK_GT
<TK_GE
);
103 assert( TK_LT
>TK_EQ
&& TK_LT
<TK_GE
);
104 assert( TK_LE
>TK_EQ
&& TK_LE
<TK_GE
);
105 assert( TK_GE
==TK_EQ
+4 );
106 return op
==TK_IN
|| (op
>=TK_EQ
&& op
<=TK_GE
) || op
==TK_ISNULL
|| op
==TK_IS
;
110 ** Commute a comparison operator. Expressions of the form "X op Y"
111 ** are converted into "Y op X".
113 static u16
exprCommute(Parse
*pParse
, Expr
*pExpr
){
114 if( pExpr
->pLeft
->op
==TK_VECTOR
115 || pExpr
->pRight
->op
==TK_VECTOR
116 || sqlite3BinaryCompareCollSeq(pParse
, pExpr
->pLeft
, pExpr
->pRight
) !=
117 sqlite3BinaryCompareCollSeq(pParse
, pExpr
->pRight
, pExpr
->pLeft
)
119 pExpr
->flags
^= EP_Commuted
;
121 SWAP(Expr
*,pExpr
->pRight
,pExpr
->pLeft
);
122 if( pExpr
->op
>=TK_GT
){
123 assert( TK_LT
==TK_GT
+2 );
124 assert( TK_GE
==TK_LE
+2 );
125 assert( TK_GT
>TK_EQ
);
126 assert( TK_GT
<TK_LE
);
127 assert( pExpr
->op
>=TK_GT
&& pExpr
->op
<=TK_GE
);
128 pExpr
->op
= ((pExpr
->op
-TK_GT
)^2)+TK_GT
;
134 ** Translate from TK_xx operator to WO_xx bitmask.
136 static u16
operatorMask(int op
){
138 assert( allowedOp(op
) );
141 }else if( op
==TK_ISNULL
){
143 }else if( op
==TK_IS
){
146 assert( (WO_EQ
<<(op
-TK_EQ
)) < 0x7fff );
147 c
= (u16
)(WO_EQ
<<(op
-TK_EQ
));
149 assert( op
!=TK_ISNULL
|| c
==WO_ISNULL
);
150 assert( op
!=TK_IN
|| c
==WO_IN
);
151 assert( op
!=TK_EQ
|| c
==WO_EQ
);
152 assert( op
!=TK_LT
|| c
==WO_LT
);
153 assert( op
!=TK_LE
|| c
==WO_LE
);
154 assert( op
!=TK_GT
|| c
==WO_GT
);
155 assert( op
!=TK_GE
|| c
==WO_GE
);
156 assert( op
!=TK_IS
|| c
==WO_IS
);
161 #ifndef SQLITE_OMIT_LIKE_OPTIMIZATION
163 ** Check to see if the given expression is a LIKE or GLOB operator that
164 ** can be optimized using inequality constraints. Return TRUE if it is
165 ** so and false if not.
167 ** In order for the operator to be optimizible, the RHS must be a string
168 ** literal that does not begin with a wildcard. The LHS must be a column
169 ** that may only be NULL, a string, or a BLOB, never a number. (This means
170 ** that virtual tables cannot participate in the LIKE optimization.) The
171 ** collating sequence for the column on the LHS must be appropriate for
174 static int isLikeOrGlob(
175 Parse
*pParse
, /* Parsing and code generating context */
176 Expr
*pExpr
, /* Test this expression */
177 Expr
**ppPrefix
, /* Pointer to TK_STRING expression with pattern prefix */
178 int *pisComplete
, /* True if the only wildcard is % in the last character */
179 int *pnoCase
/* True if uppercase is equivalent to lowercase */
181 const u8
*z
= 0; /* String on RHS of LIKE operator */
182 Expr
*pRight
, *pLeft
; /* Right and left size of LIKE operator */
183 ExprList
*pList
; /* List of operands to the LIKE operator */
184 u8 c
; /* One character in z[] */
185 int cnt
; /* Number of non-wildcard prefix characters */
186 u8 wc
[4]; /* Wildcard characters */
187 sqlite3
*db
= pParse
->db
; /* Database connection */
188 sqlite3_value
*pVal
= 0;
189 int op
; /* Opcode of pRight */
190 int rc
; /* Result code to return */
192 if( !sqlite3IsLikeFunction(db
, pExpr
, pnoCase
, (char*)wc
) ){
196 if( *pnoCase
) return 0;
198 pList
= pExpr
->x
.pList
;
199 pLeft
= pList
->a
[1].pExpr
;
201 pRight
= sqlite3ExprSkipCollate(pList
->a
[0].pExpr
);
203 if( op
==TK_VARIABLE
&& (db
->flags
& SQLITE_EnableQPSG
)==0 ){
204 Vdbe
*pReprepare
= pParse
->pReprepare
;
205 int iCol
= pRight
->iColumn
;
206 pVal
= sqlite3VdbeGetBoundValue(pReprepare
, iCol
, SQLITE_AFF_BLOB
);
207 if( pVal
&& sqlite3_value_type(pVal
)==SQLITE_TEXT
){
208 z
= sqlite3_value_text(pVal
);
210 sqlite3VdbeSetVarmask(pParse
->pVdbe
, iCol
);
211 assert( pRight
->op
==TK_VARIABLE
|| pRight
->op
==TK_REGISTER
);
212 }else if( op
==TK_STRING
){
213 z
= (u8
*)pRight
->u
.zToken
;
217 /* Count the number of prefix characters prior to the first wildcard */
219 while( (c
=z
[cnt
])!=0 && c
!=wc
[0] && c
!=wc
[1] && c
!=wc
[2] ){
221 if( c
==wc
[3] && z
[cnt
]!=0 ) cnt
++;
224 /* The optimization is possible only if (1) the pattern does not begin
225 ** with a wildcard and if (2) the non-wildcard prefix does not end with
226 ** an (illegal 0xff) character, or (3) the pattern does not consist of
227 ** a single escape character. The second condition is necessary so
228 ** that we can increment the prefix key to find an upper bound for the
229 ** range search. The third is because the caller assumes that the pattern
230 ** consists of at least one character after all escapes have been
232 if( cnt
!=0 && 255!=(u8
)z
[cnt
-1] && (cnt
>1 || z
[0]!=wc
[3]) ){
235 /* A "complete" match if the pattern ends with "*" or "%" */
236 *pisComplete
= c
==wc
[0] && z
[cnt
+1]==0;
238 /* Get the pattern prefix. Remove all escapes from the prefix. */
239 pPrefix
= sqlite3Expr(db
, TK_STRING
, (char*)z
);
242 char *zNew
= pPrefix
->u
.zToken
;
244 for(iFrom
=iTo
=0; iFrom
<cnt
; iFrom
++){
245 if( zNew
[iFrom
]==wc
[3] ) iFrom
++;
246 zNew
[iTo
++] = zNew
[iFrom
];
251 /* If the LHS is not an ordinary column with TEXT affinity, then the
252 ** pattern prefix boundaries (both the start and end boundaries) must
253 ** not look like a number. Otherwise the pattern might be treated as
254 ** a number, which will invalidate the LIKE optimization.
256 ** Getting this right has been a persistent source of bugs in the
257 ** LIKE optimization. See, for example:
258 ** 2018-09-10 https://sqlite.org/src/info/c94369cae9b561b1
259 ** 2019-05-02 https://sqlite.org/src/info/b043a54c3de54b28
260 ** 2019-06-10 https://sqlite.org/src/info/fd76310a5e843e07
261 ** 2019-06-14 https://sqlite.org/src/info/ce8717f0885af975
262 ** 2019-09-03 https://sqlite.org/src/info/0f0428096f17252a
264 if( pLeft
->op
!=TK_COLUMN
265 || sqlite3ExprAffinity(pLeft
)!=SQLITE_AFF_TEXT
266 || IsVirtual(pLeft
->y
.pTab
) /* Value might be numeric */
270 isNum
= sqlite3AtoF(zNew
, &rDummy
, iTo
, SQLITE_UTF8
);
272 if( iTo
==1 && zNew
[0]=='-' ){
276 isNum
= sqlite3AtoF(zNew
, &rDummy
, iTo
, SQLITE_UTF8
);
281 sqlite3ExprDelete(db
, pPrefix
);
282 sqlite3ValueFree(pVal
);
289 /* If the RHS pattern is a bound parameter, make arrangements to
290 ** reprepare the statement when that parameter is rebound */
291 if( op
==TK_VARIABLE
){
292 Vdbe
*v
= pParse
->pVdbe
;
293 sqlite3VdbeSetVarmask(v
, pRight
->iColumn
);
294 if( *pisComplete
&& pRight
->u
.zToken
[1] ){
295 /* If the rhs of the LIKE expression is a variable, and the current
296 ** value of the variable means there is no need to invoke the LIKE
297 ** function, then no OP_Variable will be added to the program.
298 ** This causes problems for the sqlite3_bind_parameter_name()
299 ** API. To work around them, add a dummy OP_Variable here.
301 int r1
= sqlite3GetTempReg(pParse
);
302 sqlite3ExprCodeTarget(pParse
, pRight
, r1
);
303 sqlite3VdbeChangeP3(v
, sqlite3VdbeCurrentAddr(v
)-1, 0);
304 sqlite3ReleaseTempReg(pParse
, r1
);
313 sqlite3ValueFree(pVal
);
316 #endif /* SQLITE_OMIT_LIKE_OPTIMIZATION */
319 #ifndef SQLITE_OMIT_VIRTUALTABLE
321 ** Check to see if the pExpr expression is a form that needs to be passed
322 ** to the xBestIndex method of virtual tables. Forms of interest include:
324 ** Expression Virtual Table Operator
325 ** ----------------------- ---------------------------------
326 ** 1. column MATCH expr SQLITE_INDEX_CONSTRAINT_MATCH
327 ** 2. column GLOB expr SQLITE_INDEX_CONSTRAINT_GLOB
328 ** 3. column LIKE expr SQLITE_INDEX_CONSTRAINT_LIKE
329 ** 4. column REGEXP expr SQLITE_INDEX_CONSTRAINT_REGEXP
330 ** 5. column != expr SQLITE_INDEX_CONSTRAINT_NE
331 ** 6. expr != column SQLITE_INDEX_CONSTRAINT_NE
332 ** 7. column IS NOT expr SQLITE_INDEX_CONSTRAINT_ISNOT
333 ** 8. expr IS NOT column SQLITE_INDEX_CONSTRAINT_ISNOT
334 ** 9. column IS NOT NULL SQLITE_INDEX_CONSTRAINT_ISNOTNULL
336 ** In every case, "column" must be a column of a virtual table. If there
337 ** is a match, set *ppLeft to the "column" expression, set *ppRight to the
338 ** "expr" expression (even though in forms (6) and (8) the column is on the
339 ** right and the expression is on the left). Also set *peOp2 to the
340 ** appropriate virtual table operator. The return value is 1 or 2 if there
341 ** is a match. The usual return is 1, but if the RHS is also a column
342 ** of virtual table in forms (5) or (7) then return 2.
344 ** If the expression matches none of the patterns above, return 0.
346 static int isAuxiliaryVtabOperator(
347 sqlite3
*db
, /* Parsing context */
348 Expr
*pExpr
, /* Test this expression */
349 unsigned char *peOp2
, /* OUT: 0 for MATCH, or else an op2 value */
350 Expr
**ppLeft
, /* Column expression to left of MATCH/op2 */
351 Expr
**ppRight
/* Expression to left of MATCH/op2 */
353 if( pExpr
->op
==TK_FUNCTION
){
354 static const struct Op2
{
358 { "match", SQLITE_INDEX_CONSTRAINT_MATCH
},
359 { "glob", SQLITE_INDEX_CONSTRAINT_GLOB
},
360 { "like", SQLITE_INDEX_CONSTRAINT_LIKE
},
361 { "regexp", SQLITE_INDEX_CONSTRAINT_REGEXP
}
364 Expr
*pCol
; /* Column reference */
367 pList
= pExpr
->x
.pList
;
368 if( pList
==0 || pList
->nExpr
!=2 ){
372 /* Built-in operators MATCH, GLOB, LIKE, and REGEXP attach to a
373 ** virtual table on their second argument, which is the same as
374 ** the left-hand side operand in their in-fix form.
376 ** vtab_column MATCH expression
377 ** MATCH(expression,vtab_column)
379 pCol
= pList
->a
[1].pExpr
;
380 testcase( pCol
->op
==TK_COLUMN
&& pCol
->y
.pTab
==0 );
381 if( ExprIsVtab(pCol
) ){
382 for(i
=0; i
<ArraySize(aOp
); i
++){
383 if( sqlite3StrICmp(pExpr
->u
.zToken
, aOp
[i
].zOp
)==0 ){
384 *peOp2
= aOp
[i
].eOp2
;
385 *ppRight
= pList
->a
[0].pExpr
;
392 /* We can also match against the first column of overloaded
393 ** functions where xFindFunction returns a value of at least
394 ** SQLITE_INDEX_CONSTRAINT_FUNCTION.
396 ** OVERLOADED(vtab_column,expression)
398 ** Historically, xFindFunction expected to see lower-case function
399 ** names. But for this use case, xFindFunction is expected to deal
400 ** with function names in an arbitrary case.
402 pCol
= pList
->a
[0].pExpr
;
403 testcase( pCol
->op
==TK_COLUMN
&& pCol
->y
.pTab
==0 );
404 if( ExprIsVtab(pCol
) ){
406 sqlite3_module
*pMod
;
407 void (*xNotUsed
)(sqlite3_context
*,int,sqlite3_value
**);
409 pVtab
= sqlite3GetVTable(db
, pCol
->y
.pTab
)->pVtab
;
411 assert( pVtab
->pModule
!=0 );
412 pMod
= (sqlite3_module
*)pVtab
->pModule
;
413 if( pMod
->xFindFunction
!=0 ){
414 i
= pMod
->xFindFunction(pVtab
,2, pExpr
->u
.zToken
, &xNotUsed
, &pNotUsed
);
415 if( i
>=SQLITE_INDEX_CONSTRAINT_FUNCTION
){
417 *ppRight
= pList
->a
[1].pExpr
;
423 }else if( pExpr
->op
==TK_NE
|| pExpr
->op
==TK_ISNOT
|| pExpr
->op
==TK_NOTNULL
){
425 Expr
*pLeft
= pExpr
->pLeft
;
426 Expr
*pRight
= pExpr
->pRight
;
427 testcase( pLeft
->op
==TK_COLUMN
&& pLeft
->y
.pTab
==0 );
428 if( ExprIsVtab(pLeft
) ){
431 testcase( pRight
&& pRight
->op
==TK_COLUMN
&& pRight
->y
.pTab
==0 );
432 if( pRight
&& ExprIsVtab(pRight
) ){
434 SWAP(Expr
*, pLeft
, pRight
);
438 if( pExpr
->op
==TK_NE
) *peOp2
= SQLITE_INDEX_CONSTRAINT_NE
;
439 if( pExpr
->op
==TK_ISNOT
) *peOp2
= SQLITE_INDEX_CONSTRAINT_ISNOT
;
440 if( pExpr
->op
==TK_NOTNULL
) *peOp2
= SQLITE_INDEX_CONSTRAINT_ISNOTNULL
;
445 #endif /* SQLITE_OMIT_VIRTUALTABLE */
448 ** If the pBase expression originated in the ON or USING clause of
449 ** a join, then transfer the appropriate markings over to derived.
451 static void transferJoinMarkings(Expr
*pDerived
, Expr
*pBase
){
453 pDerived
->flags
|= pBase
->flags
& EP_FromJoin
;
454 pDerived
->iRightJoinTable
= pBase
->iRightJoinTable
;
459 ** Mark term iChild as being a child of term iParent
461 static void markTermAsChild(WhereClause
*pWC
, int iChild
, int iParent
){
462 pWC
->a
[iChild
].iParent
= iParent
;
463 pWC
->a
[iChild
].truthProb
= pWC
->a
[iParent
].truthProb
;
464 pWC
->a
[iParent
].nChild
++;
468 ** Return the N-th AND-connected subterm of pTerm. Or if pTerm is not
469 ** a conjunction, then return just pTerm when N==0. If N is exceeds
470 ** the number of available subterms, return NULL.
472 static WhereTerm
*whereNthSubterm(WhereTerm
*pTerm
, int N
){
473 if( pTerm
->eOperator
!=WO_AND
){
474 return N
==0 ? pTerm
: 0;
476 if( N
<pTerm
->u
.pAndInfo
->wc
.nTerm
){
477 return &pTerm
->u
.pAndInfo
->wc
.a
[N
];
483 ** Subterms pOne and pTwo are contained within WHERE clause pWC. The
484 ** two subterms are in disjunction - they are OR-ed together.
486 ** If these two terms are both of the form: "A op B" with the same
487 ** A and B values but different operators and if the operators are
488 ** compatible (if one is = and the other is <, for example) then
489 ** add a new virtual AND term to pWC that is the combination of the
494 ** x<y OR x=y --> x<=y
495 ** x=y OR x=y --> x=y
496 ** x<=y OR x<y --> x<=y
498 ** The following is NOT generated:
500 ** x<y OR x>y --> x!=y
502 static void whereCombineDisjuncts(
503 SrcList
*pSrc
, /* the FROM clause */
504 WhereClause
*pWC
, /* The complete WHERE clause */
505 WhereTerm
*pOne
, /* First disjunct */
506 WhereTerm
*pTwo
/* Second disjunct */
508 u16 eOp
= pOne
->eOperator
| pTwo
->eOperator
;
509 sqlite3
*db
; /* Database connection (for malloc) */
510 Expr
*pNew
; /* New virtual expression */
511 int op
; /* Operator for the combined expression */
512 int idxNew
; /* Index in pWC of the next virtual term */
514 if( (pOne
->eOperator
& (WO_EQ
|WO_LT
|WO_LE
|WO_GT
|WO_GE
))==0 ) return;
515 if( (pTwo
->eOperator
& (WO_EQ
|WO_LT
|WO_LE
|WO_GT
|WO_GE
))==0 ) return;
516 if( (eOp
& (WO_EQ
|WO_LT
|WO_LE
))!=eOp
517 && (eOp
& (WO_EQ
|WO_GT
|WO_GE
))!=eOp
) return;
518 assert( pOne
->pExpr
->pLeft
!=0 && pOne
->pExpr
->pRight
!=0 );
519 assert( pTwo
->pExpr
->pLeft
!=0 && pTwo
->pExpr
->pRight
!=0 );
520 if( sqlite3ExprCompare(0,pOne
->pExpr
->pLeft
, pTwo
->pExpr
->pLeft
, -1) ) return;
521 if( sqlite3ExprCompare(0,pOne
->pExpr
->pRight
, pTwo
->pExpr
->pRight
,-1) )return;
522 /* If we reach this point, it means the two subterms can be combined */
523 if( (eOp
& (eOp
-1))!=0 ){
524 if( eOp
& (WO_LT
|WO_LE
) ){
527 assert( eOp
& (WO_GT
|WO_GE
) );
531 db
= pWC
->pWInfo
->pParse
->db
;
532 pNew
= sqlite3ExprDup(db
, pOne
->pExpr
, 0);
533 if( pNew
==0 ) return;
534 for(op
=TK_EQ
; eOp
!=(WO_EQ
<<(op
-TK_EQ
)); op
++){ assert( op
<TK_GE
); }
536 idxNew
= whereClauseInsert(pWC
, pNew
, TERM_VIRTUAL
|TERM_DYNAMIC
);
537 exprAnalyze(pSrc
, pWC
, idxNew
);
540 #if !defined(SQLITE_OMIT_OR_OPTIMIZATION) && !defined(SQLITE_OMIT_SUBQUERY)
542 ** Analyze a term that consists of two or more OR-connected
545 ** ... WHERE (a=5) AND (b=7 OR c=9 OR d=13) AND (d=13)
546 ** ^^^^^^^^^^^^^^^^^^^^
548 ** This routine analyzes terms such as the middle term in the above example.
549 ** A WhereOrTerm object is computed and attached to the term under
550 ** analysis, regardless of the outcome of the analysis. Hence:
552 ** WhereTerm.wtFlags |= TERM_ORINFO
553 ** WhereTerm.u.pOrInfo = a dynamically allocated WhereOrTerm object
555 ** The term being analyzed must have two or more of OR-connected subterms.
556 ** A single subterm might be a set of AND-connected sub-subterms.
557 ** Examples of terms under analysis:
559 ** (A) t1.x=t2.y OR t1.x=t2.z OR t1.y=15 OR t1.z=t3.a+5
560 ** (B) x=expr1 OR expr2=x OR x=expr3
561 ** (C) t1.x=t2.y OR (t1.x=t2.z AND t1.y=15)
562 ** (D) x=expr1 OR (y>11 AND y<22 AND z LIKE '*hello*')
563 ** (E) (p.a=1 AND q.b=2 AND r.c=3) OR (p.x=4 AND q.y=5 AND r.z=6)
564 ** (F) x>A OR (x=A AND y>=B)
568 ** If all subterms are of the form T.C=expr for some single column of C and
569 ** a single table T (as shown in example B above) then create a new virtual
570 ** term that is an equivalent IN expression. In other words, if the term
571 ** being analyzed is:
573 ** x = expr1 OR expr2 = x OR x = expr3
575 ** then create a new virtual term like this:
577 ** x IN (expr1,expr2,expr3)
581 ** If there are exactly two disjuncts and one side has x>A and the other side
582 ** has x=A (for the same x and A) then add a new virtual conjunct term to the
583 ** WHERE clause of the form "x>=A". Example:
585 ** x>A OR (x=A AND y>B) adds: x>=A
587 ** The added conjunct can sometimes be helpful in query planning.
591 ** If all subterms are indexable by a single table T, then set
593 ** WhereTerm.eOperator = WO_OR
594 ** WhereTerm.u.pOrInfo->indexable |= the cursor number for table T
596 ** A subterm is "indexable" if it is of the form
597 ** "T.C <op> <expr>" where C is any column of table T and
598 ** <op> is one of "=", "<", "<=", ">", ">=", "IS NULL", or "IN".
599 ** A subterm is also indexable if it is an AND of two or more
600 ** subsubterms at least one of which is indexable. Indexable AND
601 ** subterms have their eOperator set to WO_AND and they have
602 ** u.pAndInfo set to a dynamically allocated WhereAndTerm object.
604 ** From another point of view, "indexable" means that the subterm could
605 ** potentially be used with an index if an appropriate index exists.
606 ** This analysis does not consider whether or not the index exists; that
607 ** is decided elsewhere. This analysis only looks at whether subterms
608 ** appropriate for indexing exist.
610 ** All examples A through E above satisfy case 3. But if a term
611 ** also satisfies case 1 (such as B) we know that the optimizer will
612 ** always prefer case 1, so in that case we pretend that case 3 is not
615 ** It might be the case that multiple tables are indexable. For example,
616 ** (E) above is indexable on tables P, Q, and R.
618 ** Terms that satisfy case 3 are candidates for lookup by using
619 ** separate indices to find rowids for each subterm and composing
620 ** the union of all rowids using a RowSet object. This is similar
621 ** to "bitmap indices" in other database engines.
625 ** If none of cases 1, 2, or 3 apply, then leave the eOperator set to
626 ** zero. This term is not useful for search.
628 static void exprAnalyzeOrTerm(
629 SrcList
*pSrc
, /* the FROM clause */
630 WhereClause
*pWC
, /* the complete WHERE clause */
631 int idxTerm
/* Index of the OR-term to be analyzed */
633 WhereInfo
*pWInfo
= pWC
->pWInfo
; /* WHERE clause processing context */
634 Parse
*pParse
= pWInfo
->pParse
; /* Parser context */
635 sqlite3
*db
= pParse
->db
; /* Database connection */
636 WhereTerm
*pTerm
= &pWC
->a
[idxTerm
]; /* The term to be analyzed */
637 Expr
*pExpr
= pTerm
->pExpr
; /* The expression of the term */
638 int i
; /* Loop counters */
639 WhereClause
*pOrWc
; /* Breakup of pTerm into subterms */
640 WhereTerm
*pOrTerm
; /* A Sub-term within the pOrWc */
641 WhereOrInfo
*pOrInfo
; /* Additional information associated with pTerm */
642 Bitmask chngToIN
; /* Tables that might satisfy case 1 */
643 Bitmask indexable
; /* Tables that are indexable, satisfying case 2 */
646 ** Break the OR clause into its separate subterms. The subterms are
647 ** stored in a WhereClause structure containing within the WhereOrInfo
648 ** object that is attached to the original OR clause term.
650 assert( (pTerm
->wtFlags
& (TERM_DYNAMIC
|TERM_ORINFO
|TERM_ANDINFO
))==0 );
651 assert( pExpr
->op
==TK_OR
);
652 pTerm
->u
.pOrInfo
= pOrInfo
= sqlite3DbMallocZero(db
, sizeof(*pOrInfo
));
653 if( pOrInfo
==0 ) return;
654 pTerm
->wtFlags
|= TERM_ORINFO
;
655 pOrWc
= &pOrInfo
->wc
;
656 memset(pOrWc
->aStatic
, 0, sizeof(pOrWc
->aStatic
));
657 sqlite3WhereClauseInit(pOrWc
, pWInfo
);
658 sqlite3WhereSplit(pOrWc
, pExpr
, TK_OR
);
659 sqlite3WhereExprAnalyze(pSrc
, pOrWc
);
660 if( db
->mallocFailed
) return;
661 assert( pOrWc
->nTerm
>=2 );
664 ** Compute the set of tables that might satisfy cases 1 or 3.
666 indexable
= ~(Bitmask
)0;
667 chngToIN
= ~(Bitmask
)0;
668 for(i
=pOrWc
->nTerm
-1, pOrTerm
=pOrWc
->a
; i
>=0 && indexable
; i
--, pOrTerm
++){
669 if( (pOrTerm
->eOperator
& WO_SINGLE
)==0 ){
670 WhereAndInfo
*pAndInfo
;
671 assert( (pOrTerm
->wtFlags
& (TERM_ANDINFO
|TERM_ORINFO
))==0 );
673 pAndInfo
= sqlite3DbMallocRawNN(db
, sizeof(*pAndInfo
));
679 pOrTerm
->u
.pAndInfo
= pAndInfo
;
680 pOrTerm
->wtFlags
|= TERM_ANDINFO
;
681 pOrTerm
->eOperator
= WO_AND
;
682 pAndWC
= &pAndInfo
->wc
;
683 memset(pAndWC
->aStatic
, 0, sizeof(pAndWC
->aStatic
));
684 sqlite3WhereClauseInit(pAndWC
, pWC
->pWInfo
);
685 sqlite3WhereSplit(pAndWC
, pOrTerm
->pExpr
, TK_AND
);
686 sqlite3WhereExprAnalyze(pSrc
, pAndWC
);
687 pAndWC
->pOuter
= pWC
;
688 if( !db
->mallocFailed
){
689 for(j
=0, pAndTerm
=pAndWC
->a
; j
<pAndWC
->nTerm
; j
++, pAndTerm
++){
690 assert( pAndTerm
->pExpr
);
691 if( allowedOp(pAndTerm
->pExpr
->op
)
692 || pAndTerm
->eOperator
==WO_AUX
694 b
|= sqlite3WhereGetMask(&pWInfo
->sMaskSet
, pAndTerm
->leftCursor
);
700 }else if( pOrTerm
->wtFlags
& TERM_COPIED
){
701 /* Skip this term for now. We revisit it when we process the
702 ** corresponding TERM_VIRTUAL term */
705 b
= sqlite3WhereGetMask(&pWInfo
->sMaskSet
, pOrTerm
->leftCursor
);
706 if( pOrTerm
->wtFlags
& TERM_VIRTUAL
){
707 WhereTerm
*pOther
= &pOrWc
->a
[pOrTerm
->iParent
];
708 b
|= sqlite3WhereGetMask(&pWInfo
->sMaskSet
, pOther
->leftCursor
);
711 if( (pOrTerm
->eOperator
& WO_EQ
)==0 ){
720 ** Record the set of tables that satisfy case 3. The set might be
723 pOrInfo
->indexable
= indexable
;
725 pTerm
->eOperator
= WO_OR
;
728 pTerm
->eOperator
= WO_OR
;
731 /* For a two-way OR, attempt to implementation case 2.
733 if( indexable
&& pOrWc
->nTerm
==2 ){
736 while( (pOne
= whereNthSubterm(&pOrWc
->a
[0],iOne
++))!=0 ){
739 while( (pTwo
= whereNthSubterm(&pOrWc
->a
[1],iTwo
++))!=0 ){
740 whereCombineDisjuncts(pSrc
, pWC
, pOne
, pTwo
);
746 ** chngToIN holds a set of tables that *might* satisfy case 1. But
747 ** we have to do some additional checking to see if case 1 really
750 ** chngToIN will hold either 0, 1, or 2 bits. The 0-bit case means
751 ** that there is no possibility of transforming the OR clause into an
752 ** IN operator because one or more terms in the OR clause contain
753 ** something other than == on a column in the single table. The 1-bit
754 ** case means that every term of the OR clause is of the form
755 ** "table.column=expr" for some single table. The one bit that is set
756 ** will correspond to the common table. We still need to check to make
757 ** sure the same column is used on all terms. The 2-bit case is when
758 ** the all terms are of the form "table1.column=table2.column". It
759 ** might be possible to form an IN operator with either table1.column
760 ** or table2.column as the LHS if either is common to every term of
763 ** Note that terms of the form "table.column1=table.column2" (the
764 ** same table on both sizes of the ==) cannot be optimized.
767 int okToChngToIN
= 0; /* True if the conversion to IN is valid */
768 int iColumn
= -1; /* Column index on lhs of IN operator */
769 int iCursor
= -1; /* Table cursor common to all terms */
770 int j
= 0; /* Loop counter */
772 /* Search for a table and column that appears on one side or the
773 ** other of the == operator in every subterm. That table and column
774 ** will be recorded in iCursor and iColumn. There might not be any
775 ** such table and column. Set okToChngToIN if an appropriate table
776 ** and column is found but leave okToChngToIN false if not found.
778 for(j
=0; j
<2 && !okToChngToIN
; j
++){
781 for(i
=pOrWc
->nTerm
-1; i
>=0; i
--, pOrTerm
++){
782 assert( pOrTerm
->eOperator
& WO_EQ
);
783 pOrTerm
->wtFlags
&= ~TERM_OR_OK
;
784 if( pOrTerm
->leftCursor
==iCursor
){
785 /* This is the 2-bit case and we are on the second iteration and
786 ** current term is from the first iteration. So skip this term. */
790 if( (chngToIN
& sqlite3WhereGetMask(&pWInfo
->sMaskSet
,
791 pOrTerm
->leftCursor
))==0 ){
792 /* This term must be of the form t1.a==t2.b where t2 is in the
793 ** chngToIN set but t1 is not. This term will be either preceded
794 ** or follwed by an inverted copy (t2.b==t1.a). Skip this term
795 ** and use its inversion. */
796 testcase( pOrTerm
->wtFlags
& TERM_COPIED
);
797 testcase( pOrTerm
->wtFlags
& TERM_VIRTUAL
);
798 assert( pOrTerm
->wtFlags
& (TERM_COPIED
|TERM_VIRTUAL
) );
801 iColumn
= pOrTerm
->u
.leftColumn
;
802 iCursor
= pOrTerm
->leftCursor
;
803 pLeft
= pOrTerm
->pExpr
->pLeft
;
807 /* No candidate table+column was found. This can only occur
808 ** on the second iteration */
810 assert( IsPowerOfTwo(chngToIN
) );
811 assert( chngToIN
==sqlite3WhereGetMask(&pWInfo
->sMaskSet
, iCursor
) );
816 /* We have found a candidate table and column. Check to see if that
817 ** table and column is common to every term in the OR clause */
819 for(; i
>=0 && okToChngToIN
; i
--, pOrTerm
++){
820 assert( pOrTerm
->eOperator
& WO_EQ
);
821 if( pOrTerm
->leftCursor
!=iCursor
){
822 pOrTerm
->wtFlags
&= ~TERM_OR_OK
;
823 }else if( pOrTerm
->u
.leftColumn
!=iColumn
|| (iColumn
==XN_EXPR
824 && sqlite3ExprCompare(pParse
, pOrTerm
->pExpr
->pLeft
, pLeft
, -1)
828 int affLeft
, affRight
;
829 /* If the right-hand side is also a column, then the affinities
830 ** of both right and left sides must be such that no type
831 ** conversions are required on the right. (Ticket #2249)
833 affRight
= sqlite3ExprAffinity(pOrTerm
->pExpr
->pRight
);
834 affLeft
= sqlite3ExprAffinity(pOrTerm
->pExpr
->pLeft
);
835 if( affRight
!=0 && affRight
!=affLeft
){
838 pOrTerm
->wtFlags
|= TERM_OR_OK
;
844 /* At this point, okToChngToIN is true if original pTerm satisfies
845 ** case 1. In that case, construct a new virtual term that is
846 ** pTerm converted into an IN operator.
849 Expr
*pDup
; /* A transient duplicate expression */
850 ExprList
*pList
= 0; /* The RHS of the IN operator */
851 Expr
*pLeft
= 0; /* The LHS of the IN operator */
852 Expr
*pNew
; /* The complete IN operator */
854 for(i
=pOrWc
->nTerm
-1, pOrTerm
=pOrWc
->a
; i
>=0; i
--, pOrTerm
++){
855 if( (pOrTerm
->wtFlags
& TERM_OR_OK
)==0 ) continue;
856 assert( pOrTerm
->eOperator
& WO_EQ
);
857 assert( pOrTerm
->leftCursor
==iCursor
);
858 assert( pOrTerm
->u
.leftColumn
==iColumn
);
859 pDup
= sqlite3ExprDup(db
, pOrTerm
->pExpr
->pRight
, 0);
860 pList
= sqlite3ExprListAppend(pWInfo
->pParse
, pList
, pDup
);
861 pLeft
= pOrTerm
->pExpr
->pLeft
;
864 pDup
= sqlite3ExprDup(db
, pLeft
, 0);
865 pNew
= sqlite3PExpr(pParse
, TK_IN
, pDup
, 0);
868 transferJoinMarkings(pNew
, pExpr
);
869 assert( !ExprHasProperty(pNew
, EP_xIsSelect
) );
870 pNew
->x
.pList
= pList
;
871 idxNew
= whereClauseInsert(pWC
, pNew
, TERM_VIRTUAL
|TERM_DYNAMIC
);
872 testcase( idxNew
==0 );
873 exprAnalyze(pSrc
, pWC
, idxNew
);
874 /* pTerm = &pWC->a[idxTerm]; // would be needed if pTerm where used again */
875 markTermAsChild(pWC
, idxNew
, idxTerm
);
877 sqlite3ExprListDelete(db
, pList
);
882 #endif /* !SQLITE_OMIT_OR_OPTIMIZATION && !SQLITE_OMIT_SUBQUERY */
885 ** We already know that pExpr is a binary operator where both operands are
886 ** column references. This routine checks to see if pExpr is an equivalence
888 ** 1. The SQLITE_Transitive optimization must be enabled
889 ** 2. Must be either an == or an IS operator
890 ** 3. Not originating in the ON clause of an OUTER JOIN
891 ** 4. The affinities of A and B must be compatible
892 ** 5a. Both operands use the same collating sequence OR
893 ** 5b. The overall collating sequence is BINARY
894 ** If this routine returns TRUE, that means that the RHS can be substituted
895 ** for the LHS anyplace else in the WHERE clause where the LHS column occurs.
896 ** This is an optimization. No harm comes from returning 0. But if 1 is
897 ** returned when it should not be, then incorrect answers might result.
899 static int termIsEquivalence(Parse
*pParse
, Expr
*pExpr
){
902 if( !OptimizationEnabled(pParse
->db
, SQLITE_Transitive
) ) return 0;
903 if( pExpr
->op
!=TK_EQ
&& pExpr
->op
!=TK_IS
) return 0;
904 if( ExprHasProperty(pExpr
, EP_FromJoin
) ) return 0;
905 aff1
= sqlite3ExprAffinity(pExpr
->pLeft
);
906 aff2
= sqlite3ExprAffinity(pExpr
->pRight
);
908 && (!sqlite3IsNumericAffinity(aff1
) || !sqlite3IsNumericAffinity(aff2
))
912 pColl
= sqlite3ExprCompareCollSeq(pParse
, pExpr
);
913 if( sqlite3IsBinary(pColl
) ) return 1;
914 return sqlite3ExprCollSeqMatch(pParse
, pExpr
->pLeft
, pExpr
->pRight
);
918 ** Recursively walk the expressions of a SELECT statement and generate
919 ** a bitmask indicating which tables are used in that expression
922 static Bitmask
exprSelectUsage(WhereMaskSet
*pMaskSet
, Select
*pS
){
925 SrcList
*pSrc
= pS
->pSrc
;
926 mask
|= sqlite3WhereExprListUsage(pMaskSet
, pS
->pEList
);
927 mask
|= sqlite3WhereExprListUsage(pMaskSet
, pS
->pGroupBy
);
928 mask
|= sqlite3WhereExprListUsage(pMaskSet
, pS
->pOrderBy
);
929 mask
|= sqlite3WhereExprUsage(pMaskSet
, pS
->pWhere
);
930 mask
|= sqlite3WhereExprUsage(pMaskSet
, pS
->pHaving
);
931 if( ALWAYS(pSrc
!=0) ){
933 for(i
=0; i
<pSrc
->nSrc
; i
++){
934 mask
|= exprSelectUsage(pMaskSet
, pSrc
->a
[i
].pSelect
);
935 mask
|= sqlite3WhereExprUsage(pMaskSet
, pSrc
->a
[i
].pOn
);
936 if( pSrc
->a
[i
].fg
.isTabFunc
){
937 mask
|= sqlite3WhereExprListUsage(pMaskSet
, pSrc
->a
[i
].u1
.pFuncArg
);
947 ** Expression pExpr is one operand of a comparison operator that might
948 ** be useful for indexing. This routine checks to see if pExpr appears
949 ** in any index. Return TRUE (1) if pExpr is an indexed term and return
950 ** FALSE (0) if not. If TRUE is returned, also set aiCurCol[0] to the cursor
951 ** number of the table that is indexed and aiCurCol[1] to the column number
952 ** of the column that is indexed, or XN_EXPR (-2) if an expression is being
955 ** If pExpr is a TK_COLUMN column reference, then this routine always returns
956 ** true even if that particular column is not indexed, because the column
957 ** might be added to an automatic index later.
959 static SQLITE_NOINLINE
int exprMightBeIndexed2(
960 SrcList
*pFrom
, /* The FROM clause */
961 Bitmask mPrereq
, /* Bitmask of FROM clause terms referenced by pExpr */
962 int *aiCurCol
, /* Write the referenced table cursor and column here */
963 Expr
*pExpr
/* An operand of a comparison operator */
968 for(i
=0; mPrereq
>1; i
++, mPrereq
>>=1){}
969 iCur
= pFrom
->a
[i
].iCursor
;
970 for(pIdx
=pFrom
->a
[i
].pTab
->pIndex
; pIdx
; pIdx
=pIdx
->pNext
){
971 if( pIdx
->aColExpr
==0 ) continue;
972 for(i
=0; i
<pIdx
->nKeyCol
; i
++){
973 if( pIdx
->aiColumn
[i
]!=XN_EXPR
) continue;
974 if( sqlite3ExprCompareSkip(pExpr
, pIdx
->aColExpr
->a
[i
].pExpr
, iCur
)==0 ){
976 aiCurCol
[1] = XN_EXPR
;
983 static int exprMightBeIndexed(
984 SrcList
*pFrom
, /* The FROM clause */
985 Bitmask mPrereq
, /* Bitmask of FROM clause terms referenced by pExpr */
986 int *aiCurCol
, /* Write the referenced table cursor & column here */
987 Expr
*pExpr
, /* An operand of a comparison operator */
988 int op
/* The specific comparison operator */
990 /* If this expression is a vector to the left or right of a
991 ** inequality constraint (>, <, >= or <=), perform the processing
992 ** on the first element of the vector. */
993 assert( TK_GT
+1==TK_LE
&& TK_GT
+2==TK_LT
&& TK_GT
+3==TK_GE
);
994 assert( TK_IS
<TK_GE
&& TK_ISNULL
<TK_GE
&& TK_IN
<TK_GE
);
996 if( pExpr
->op
==TK_VECTOR
&& (op
>=TK_GT
&& ALWAYS(op
<=TK_GE
)) ){
997 pExpr
= pExpr
->x
.pList
->a
[0].pExpr
;
1000 if( pExpr
->op
==TK_COLUMN
){
1001 aiCurCol
[0] = pExpr
->iTable
;
1002 aiCurCol
[1] = pExpr
->iColumn
;
1005 if( mPrereq
==0 ) return 0; /* No table references */
1006 if( (mPrereq
&(mPrereq
-1))!=0 ) return 0; /* Refs more than one table */
1007 return exprMightBeIndexed2(pFrom
,mPrereq
,aiCurCol
,pExpr
);
1011 ** The input to this routine is an WhereTerm structure with only the
1012 ** "pExpr" field filled in. The job of this routine is to analyze the
1013 ** subexpression and populate all the other fields of the WhereTerm
1016 ** If the expression is of the form "<expr> <op> X" it gets commuted
1017 ** to the standard form of "X <op> <expr>".
1019 ** If the expression is of the form "X <op> Y" where both X and Y are
1020 ** columns, then the original expression is unchanged and a new virtual
1021 ** term of the form "Y <op> X" is added to the WHERE clause and
1022 ** analyzed separately. The original term is marked with TERM_COPIED
1023 ** and the new term is marked with TERM_DYNAMIC (because it's pExpr
1024 ** needs to be freed with the WhereClause) and TERM_VIRTUAL (because it
1025 ** is a commuted copy of a prior term.) The original term has nChild=1
1026 ** and the copy has idxParent set to the index of the original term.
1028 static void exprAnalyze(
1029 SrcList
*pSrc
, /* the FROM clause */
1030 WhereClause
*pWC
, /* the WHERE clause */
1031 int idxTerm
/* Index of the term to be analyzed */
1033 WhereInfo
*pWInfo
= pWC
->pWInfo
; /* WHERE clause processing context */
1034 WhereTerm
*pTerm
; /* The term to be analyzed */
1035 WhereMaskSet
*pMaskSet
; /* Set of table index masks */
1036 Expr
*pExpr
; /* The expression to be analyzed */
1037 Bitmask prereqLeft
; /* Prerequesites of the pExpr->pLeft */
1038 Bitmask prereqAll
; /* Prerequesites of pExpr */
1039 Bitmask extraRight
= 0; /* Extra dependencies on LEFT JOIN */
1040 Expr
*pStr1
= 0; /* RHS of LIKE/GLOB operator */
1041 int isComplete
= 0; /* RHS of LIKE/GLOB ends with wildcard */
1042 int noCase
= 0; /* uppercase equivalent to lowercase */
1043 int op
; /* Top-level operator. pExpr->op */
1044 Parse
*pParse
= pWInfo
->pParse
; /* Parsing context */
1045 sqlite3
*db
= pParse
->db
; /* Database connection */
1046 unsigned char eOp2
= 0; /* op2 value for LIKE/REGEXP/GLOB */
1047 int nLeft
; /* Number of elements on left side vector */
1049 if( db
->mallocFailed
){
1052 pTerm
= &pWC
->a
[idxTerm
];
1053 pMaskSet
= &pWInfo
->sMaskSet
;
1054 pExpr
= pTerm
->pExpr
;
1055 assert( pExpr
->op
!=TK_AS
&& pExpr
->op
!=TK_COLLATE
);
1056 prereqLeft
= sqlite3WhereExprUsage(pMaskSet
, pExpr
->pLeft
);
1059 assert( pExpr
->pRight
==0 );
1060 if( sqlite3ExprCheckIN(pParse
, pExpr
) ) return;
1061 if( ExprHasProperty(pExpr
, EP_xIsSelect
) ){
1062 pTerm
->prereqRight
= exprSelectUsage(pMaskSet
, pExpr
->x
.pSelect
);
1064 pTerm
->prereqRight
= sqlite3WhereExprListUsage(pMaskSet
, pExpr
->x
.pList
);
1066 }else if( op
==TK_ISNULL
){
1067 pTerm
->prereqRight
= 0;
1069 pTerm
->prereqRight
= sqlite3WhereExprUsage(pMaskSet
, pExpr
->pRight
);
1071 pMaskSet
->bVarSelect
= 0;
1072 prereqAll
= sqlite3WhereExprUsageNN(pMaskSet
, pExpr
);
1073 if( pMaskSet
->bVarSelect
) pTerm
->wtFlags
|= TERM_VARSELECT
;
1074 if( ExprHasProperty(pExpr
, EP_FromJoin
) ){
1075 Bitmask x
= sqlite3WhereGetMask(pMaskSet
, pExpr
->iRightJoinTable
);
1077 extraRight
= x
-1; /* ON clause terms may not be used with an index
1078 ** on left table of a LEFT JOIN. Ticket #3015 */
1079 if( (prereqAll
>>1)>=x
){
1080 sqlite3ErrorMsg(pParse
, "ON clause references tables to its right");
1084 pTerm
->prereqAll
= prereqAll
;
1085 pTerm
->leftCursor
= -1;
1086 pTerm
->iParent
= -1;
1087 pTerm
->eOperator
= 0;
1088 if( allowedOp(op
) ){
1090 Expr
*pLeft
= sqlite3ExprSkipCollate(pExpr
->pLeft
);
1091 Expr
*pRight
= sqlite3ExprSkipCollate(pExpr
->pRight
);
1092 u16 opMask
= (pTerm
->prereqRight
& prereqLeft
)==0 ? WO_ALL
: WO_EQUIV
;
1094 if( pTerm
->iField
>0 ){
1095 assert( op
==TK_IN
);
1096 assert( pLeft
->op
==TK_VECTOR
);
1097 pLeft
= pLeft
->x
.pList
->a
[pTerm
->iField
-1].pExpr
;
1100 if( exprMightBeIndexed(pSrc
, prereqLeft
, aiCurCol
, pLeft
, op
) ){
1101 pTerm
->leftCursor
= aiCurCol
[0];
1102 pTerm
->u
.leftColumn
= aiCurCol
[1];
1103 pTerm
->eOperator
= operatorMask(op
) & opMask
;
1105 if( op
==TK_IS
) pTerm
->wtFlags
|= TERM_IS
;
1107 && exprMightBeIndexed(pSrc
, pTerm
->prereqRight
, aiCurCol
, pRight
, op
)
1111 u16 eExtraOp
= 0; /* Extra bits for pNew->eOperator */
1112 assert( pTerm
->iField
==0 );
1113 if( pTerm
->leftCursor
>=0 ){
1115 pDup
= sqlite3ExprDup(db
, pExpr
, 0);
1116 if( db
->mallocFailed
){
1117 sqlite3ExprDelete(db
, pDup
);
1120 idxNew
= whereClauseInsert(pWC
, pDup
, TERM_VIRTUAL
|TERM_DYNAMIC
);
1121 if( idxNew
==0 ) return;
1122 pNew
= &pWC
->a
[idxNew
];
1123 markTermAsChild(pWC
, idxNew
, idxTerm
);
1124 if( op
==TK_IS
) pNew
->wtFlags
|= TERM_IS
;
1125 pTerm
= &pWC
->a
[idxTerm
];
1126 pTerm
->wtFlags
|= TERM_COPIED
;
1128 if( termIsEquivalence(pParse
, pDup
) ){
1129 pTerm
->eOperator
|= WO_EQUIV
;
1130 eExtraOp
= WO_EQUIV
;
1136 pNew
->wtFlags
|= exprCommute(pParse
, pDup
);
1137 pNew
->leftCursor
= aiCurCol
[0];
1138 pNew
->u
.leftColumn
= aiCurCol
[1];
1139 testcase( (prereqLeft
| extraRight
) != prereqLeft
);
1140 pNew
->prereqRight
= prereqLeft
| extraRight
;
1141 pNew
->prereqAll
= prereqAll
;
1142 pNew
->eOperator
= (operatorMask(pDup
->op
) + eExtraOp
) & opMask
;
1146 #ifndef SQLITE_OMIT_BETWEEN_OPTIMIZATION
1147 /* If a term is the BETWEEN operator, create two new virtual terms
1148 ** that define the range that the BETWEEN implements. For example:
1150 ** a BETWEEN b AND c
1152 ** is converted into:
1154 ** (a BETWEEN b AND c) AND (a>=b) AND (a<=c)
1156 ** The two new terms are added onto the end of the WhereClause object.
1157 ** The new terms are "dynamic" and are children of the original BETWEEN
1158 ** term. That means that if the BETWEEN term is coded, the children are
1159 ** skipped. Or, if the children are satisfied by an index, the original
1160 ** BETWEEN term is skipped.
1162 else if( pExpr
->op
==TK_BETWEEN
&& pWC
->op
==TK_AND
){
1163 ExprList
*pList
= pExpr
->x
.pList
;
1165 static const u8 ops
[] = {TK_GE
, TK_LE
};
1167 assert( pList
->nExpr
==2 );
1171 pNewExpr
= sqlite3PExpr(pParse
, ops
[i
],
1172 sqlite3ExprDup(db
, pExpr
->pLeft
, 0),
1173 sqlite3ExprDup(db
, pList
->a
[i
].pExpr
, 0));
1174 transferJoinMarkings(pNewExpr
, pExpr
);
1175 idxNew
= whereClauseInsert(pWC
, pNewExpr
, TERM_VIRTUAL
|TERM_DYNAMIC
);
1176 testcase( idxNew
==0 );
1177 exprAnalyze(pSrc
, pWC
, idxNew
);
1178 pTerm
= &pWC
->a
[idxTerm
];
1179 markTermAsChild(pWC
, idxNew
, idxTerm
);
1182 #endif /* SQLITE_OMIT_BETWEEN_OPTIMIZATION */
1184 #if !defined(SQLITE_OMIT_OR_OPTIMIZATION) && !defined(SQLITE_OMIT_SUBQUERY)
1185 /* Analyze a term that is composed of two or more subterms connected by
1188 else if( pExpr
->op
==TK_OR
){
1189 assert( pWC
->op
==TK_AND
);
1190 exprAnalyzeOrTerm(pSrc
, pWC
, idxTerm
);
1191 pTerm
= &pWC
->a
[idxTerm
];
1193 #endif /* SQLITE_OMIT_OR_OPTIMIZATION */
1195 #ifndef SQLITE_OMIT_LIKE_OPTIMIZATION
1196 /* Add constraints to reduce the search space on a LIKE or GLOB
1199 ** A like pattern of the form "x LIKE 'aBc%'" is changed into constraints
1201 ** x>='ABC' AND x<'abd' AND x LIKE 'aBc%'
1203 ** The last character of the prefix "abc" is incremented to form the
1204 ** termination condition "abd". If case is not significant (the default
1205 ** for LIKE) then the lower-bound is made all uppercase and the upper-
1206 ** bound is made all lowercase so that the bounds also work when comparing
1210 && isLikeOrGlob(pParse
, pExpr
, &pStr1
, &isComplete
, &noCase
)
1212 Expr
*pLeft
; /* LHS of LIKE/GLOB operator */
1213 Expr
*pStr2
; /* Copy of pStr1 - RHS of LIKE/GLOB operator */
1218 const char *zCollSeqName
; /* Name of collating sequence */
1219 const u16 wtFlags
= TERM_LIKEOPT
| TERM_VIRTUAL
| TERM_DYNAMIC
;
1221 pLeft
= pExpr
->x
.pList
->a
[1].pExpr
;
1222 pStr2
= sqlite3ExprDup(db
, pStr1
, 0);
1224 /* Convert the lower bound to upper-case and the upper bound to
1225 ** lower-case (upper-case is less than lower-case in ASCII) so that
1226 ** the range constraints also work for BLOBs
1228 if( noCase
&& !pParse
->db
->mallocFailed
){
1231 pTerm
->wtFlags
|= TERM_LIKE
;
1232 for(i
=0; (c
= pStr1
->u
.zToken
[i
])!=0; i
++){
1233 pStr1
->u
.zToken
[i
] = sqlite3Toupper(c
);
1234 pStr2
->u
.zToken
[i
] = sqlite3Tolower(c
);
1238 if( !db
->mallocFailed
){
1239 u8 c
, *pC
; /* Last character before the first wildcard */
1240 pC
= (u8
*)&pStr2
->u
.zToken
[sqlite3Strlen30(pStr2
->u
.zToken
)-1];
1243 /* The point is to increment the last character before the first
1244 ** wildcard. But if we increment '@', that will push it into the
1245 ** alphabetic range where case conversions will mess up the
1246 ** inequality. To avoid this, make sure to also run the full
1247 ** LIKE on all candidate expressions by clearing the isComplete flag
1249 if( c
=='A'-1 ) isComplete
= 0;
1250 c
= sqlite3UpperToLower
[c
];
1254 zCollSeqName
= noCase
? "NOCASE" : sqlite3StrBINARY
;
1255 pNewExpr1
= sqlite3ExprDup(db
, pLeft
, 0);
1256 pNewExpr1
= sqlite3PExpr(pParse
, TK_GE
,
1257 sqlite3ExprAddCollateString(pParse
,pNewExpr1
,zCollSeqName
),
1259 transferJoinMarkings(pNewExpr1
, pExpr
);
1260 idxNew1
= whereClauseInsert(pWC
, pNewExpr1
, wtFlags
);
1261 testcase( idxNew1
==0 );
1262 exprAnalyze(pSrc
, pWC
, idxNew1
);
1263 pNewExpr2
= sqlite3ExprDup(db
, pLeft
, 0);
1264 pNewExpr2
= sqlite3PExpr(pParse
, TK_LT
,
1265 sqlite3ExprAddCollateString(pParse
,pNewExpr2
,zCollSeqName
),
1267 transferJoinMarkings(pNewExpr2
, pExpr
);
1268 idxNew2
= whereClauseInsert(pWC
, pNewExpr2
, wtFlags
);
1269 testcase( idxNew2
==0 );
1270 exprAnalyze(pSrc
, pWC
, idxNew2
);
1271 pTerm
= &pWC
->a
[idxTerm
];
1273 markTermAsChild(pWC
, idxNew1
, idxTerm
);
1274 markTermAsChild(pWC
, idxNew2
, idxTerm
);
1277 #endif /* SQLITE_OMIT_LIKE_OPTIMIZATION */
1279 #ifndef SQLITE_OMIT_VIRTUALTABLE
1280 /* Add a WO_AUX auxiliary term to the constraint set if the
1281 ** current expression is of the form "column OP expr" where OP
1282 ** is an operator that gets passed into virtual tables but which is
1283 ** not normally optimized for ordinary tables. In other words, OP
1284 ** is one of MATCH, LIKE, GLOB, REGEXP, !=, IS, IS NOT, or NOT NULL.
1285 ** This information is used by the xBestIndex methods of
1286 ** virtual tables. The native query optimizer does not attempt
1287 ** to do anything with MATCH functions.
1289 if( pWC
->op
==TK_AND
){
1290 Expr
*pRight
= 0, *pLeft
= 0;
1291 int res
= isAuxiliaryVtabOperator(db
, pExpr
, &eOp2
, &pLeft
, &pRight
);
1294 WhereTerm
*pNewTerm
;
1295 Bitmask prereqColumn
, prereqExpr
;
1297 prereqExpr
= sqlite3WhereExprUsage(pMaskSet
, pRight
);
1298 prereqColumn
= sqlite3WhereExprUsage(pMaskSet
, pLeft
);
1299 if( (prereqExpr
& prereqColumn
)==0 ){
1301 pNewExpr
= sqlite3PExpr(pParse
, TK_MATCH
,
1302 0, sqlite3ExprDup(db
, pRight
, 0));
1303 if( ExprHasProperty(pExpr
, EP_FromJoin
) && pNewExpr
){
1304 ExprSetProperty(pNewExpr
, EP_FromJoin
);
1305 pNewExpr
->iRightJoinTable
= pExpr
->iRightJoinTable
;
1307 idxNew
= whereClauseInsert(pWC
, pNewExpr
, TERM_VIRTUAL
|TERM_DYNAMIC
);
1308 testcase( idxNew
==0 );
1309 pNewTerm
= &pWC
->a
[idxNew
];
1310 pNewTerm
->prereqRight
= prereqExpr
;
1311 pNewTerm
->leftCursor
= pLeft
->iTable
;
1312 pNewTerm
->u
.leftColumn
= pLeft
->iColumn
;
1313 pNewTerm
->eOperator
= WO_AUX
;
1314 pNewTerm
->eMatchOp
= eOp2
;
1315 markTermAsChild(pWC
, idxNew
, idxTerm
);
1316 pTerm
= &pWC
->a
[idxTerm
];
1317 pTerm
->wtFlags
|= TERM_COPIED
;
1318 pNewTerm
->prereqAll
= pTerm
->prereqAll
;
1320 SWAP(Expr
*, pLeft
, pRight
);
1323 #endif /* SQLITE_OMIT_VIRTUALTABLE */
1325 /* If there is a vector == or IS term - e.g. "(a, b) == (?, ?)" - create
1326 ** new terms for each component comparison - "a = ?" and "b = ?". The
1327 ** new terms completely replace the original vector comparison, which is
1330 ** This is only required if at least one side of the comparison operation
1331 ** is not a sub-select. */
1333 && (pExpr
->op
==TK_EQ
|| pExpr
->op
==TK_IS
)
1334 && (nLeft
= sqlite3ExprVectorSize(pExpr
->pLeft
))>1
1335 && sqlite3ExprVectorSize(pExpr
->pRight
)==nLeft
1336 && ( (pExpr
->pLeft
->flags
& EP_xIsSelect
)==0
1337 || (pExpr
->pRight
->flags
& EP_xIsSelect
)==0)
1340 for(i
=0; i
<nLeft
; i
++){
1343 Expr
*pLeft
= sqlite3ExprForVectorField(pParse
, pExpr
->pLeft
, i
);
1344 Expr
*pRight
= sqlite3ExprForVectorField(pParse
, pExpr
->pRight
, i
);
1346 pNew
= sqlite3PExpr(pParse
, pExpr
->op
, pLeft
, pRight
);
1347 transferJoinMarkings(pNew
, pExpr
);
1348 idxNew
= whereClauseInsert(pWC
, pNew
, TERM_DYNAMIC
);
1349 exprAnalyze(pSrc
, pWC
, idxNew
);
1351 pTerm
= &pWC
->a
[idxTerm
];
1352 pTerm
->wtFlags
|= TERM_CODED
|TERM_VIRTUAL
; /* Disable the original */
1353 pTerm
->eOperator
= 0;
1356 /* If there is a vector IN term - e.g. "(a, b) IN (SELECT ...)" - create
1357 ** a virtual term for each vector component. The expression object
1358 ** used by each such virtual term is pExpr (the full vector IN(...)
1359 ** expression). The WhereTerm.iField variable identifies the index within
1360 ** the vector on the LHS that the virtual term represents.
1362 ** This only works if the RHS is a simple SELECT (not a compound) that does
1363 ** not use window functions.
1365 if( pWC
->op
==TK_AND
&& pExpr
->op
==TK_IN
&& pTerm
->iField
==0
1366 && pExpr
->pLeft
->op
==TK_VECTOR
1367 && pExpr
->x
.pSelect
->pPrior
==0
1368 #ifndef SQLITE_OMIT_WINDOWFUNC
1369 && pExpr
->x
.pSelect
->pWin
==0
1373 for(i
=0; i
<sqlite3ExprVectorSize(pExpr
->pLeft
); i
++){
1375 idxNew
= whereClauseInsert(pWC
, pExpr
, TERM_VIRTUAL
);
1376 pWC
->a
[idxNew
].iField
= i
+1;
1377 exprAnalyze(pSrc
, pWC
, idxNew
);
1378 markTermAsChild(pWC
, idxNew
, idxTerm
);
1382 #ifdef SQLITE_ENABLE_STAT4
1383 /* When sqlite_stat4 histogram data is available an operator of the
1384 ** form "x IS NOT NULL" can sometimes be evaluated more efficiently
1385 ** as "x>NULL" if x is not an INTEGER PRIMARY KEY. So construct a
1386 ** virtual term of that form.
1388 ** Note that the virtual term must be tagged with TERM_VNULL.
1390 if( pExpr
->op
==TK_NOTNULL
1391 && pExpr
->pLeft
->op
==TK_COLUMN
1392 && pExpr
->pLeft
->iColumn
>=0
1393 && !ExprHasProperty(pExpr
, EP_FromJoin
)
1394 && OptimizationEnabled(db
, SQLITE_Stat4
)
1397 Expr
*pLeft
= pExpr
->pLeft
;
1399 WhereTerm
*pNewTerm
;
1401 pNewExpr
= sqlite3PExpr(pParse
, TK_GT
,
1402 sqlite3ExprDup(db
, pLeft
, 0),
1403 sqlite3ExprAlloc(db
, TK_NULL
, 0, 0));
1405 idxNew
= whereClauseInsert(pWC
, pNewExpr
,
1406 TERM_VIRTUAL
|TERM_DYNAMIC
|TERM_VNULL
);
1408 pNewTerm
= &pWC
->a
[idxNew
];
1409 pNewTerm
->prereqRight
= 0;
1410 pNewTerm
->leftCursor
= pLeft
->iTable
;
1411 pNewTerm
->u
.leftColumn
= pLeft
->iColumn
;
1412 pNewTerm
->eOperator
= WO_GT
;
1413 markTermAsChild(pWC
, idxNew
, idxTerm
);
1414 pTerm
= &pWC
->a
[idxTerm
];
1415 pTerm
->wtFlags
|= TERM_COPIED
;
1416 pNewTerm
->prereqAll
= pTerm
->prereqAll
;
1419 #endif /* SQLITE_ENABLE_STAT4 */
1421 /* Prevent ON clause terms of a LEFT JOIN from being used to drive
1422 ** an index for tables to the left of the join.
1424 testcase( pTerm
!=&pWC
->a
[idxTerm
] );
1425 pTerm
= &pWC
->a
[idxTerm
];
1426 pTerm
->prereqRight
|= extraRight
;
1429 /***************************************************************************
1430 ** Routines with file scope above. Interface to the rest of the where.c
1431 ** subsystem follows.
1432 ***************************************************************************/
1435 ** This routine identifies subexpressions in the WHERE clause where
1436 ** each subexpression is separated by the AND operator or some other
1437 ** operator specified in the op parameter. The WhereClause structure
1438 ** is filled with pointers to subexpressions. For example:
1440 ** WHERE a=='hello' AND coalesce(b,11)<10 AND (c+12!=d OR c==22)
1441 ** \________/ \_______________/ \________________/
1442 ** slot[0] slot[1] slot[2]
1444 ** The original WHERE clause in pExpr is unaltered. All this routine
1445 ** does is make slot[] entries point to substructure within pExpr.
1447 ** In the previous sentence and in the diagram, "slot[]" refers to
1448 ** the WhereClause.a[] array. The slot[] array grows as needed to contain
1449 ** all terms of the WHERE clause.
1451 void sqlite3WhereSplit(WhereClause
*pWC
, Expr
*pExpr
, u8 op
){
1452 Expr
*pE2
= sqlite3ExprSkipCollateAndLikely(pExpr
);
1454 if( pE2
==0 ) return;
1456 whereClauseInsert(pWC
, pExpr
, 0);
1458 sqlite3WhereSplit(pWC
, pE2
->pLeft
, op
);
1459 sqlite3WhereSplit(pWC
, pE2
->pRight
, op
);
1464 ** Initialize a preallocated WhereClause structure.
1466 void sqlite3WhereClauseInit(
1467 WhereClause
*pWC
, /* The WhereClause to be initialized */
1468 WhereInfo
*pWInfo
/* The WHERE processing context */
1470 pWC
->pWInfo
= pWInfo
;
1474 pWC
->nSlot
= ArraySize(pWC
->aStatic
);
1475 pWC
->a
= pWC
->aStatic
;
1479 ** Deallocate a WhereClause structure. The WhereClause structure
1480 ** itself is not freed. This routine is the inverse of
1481 ** sqlite3WhereClauseInit().
1483 void sqlite3WhereClauseClear(WhereClause
*pWC
){
1486 sqlite3
*db
= pWC
->pWInfo
->pParse
->db
;
1487 for(i
=pWC
->nTerm
-1, a
=pWC
->a
; i
>=0; i
--, a
++){
1488 if( a
->wtFlags
& TERM_DYNAMIC
){
1489 sqlite3ExprDelete(db
, a
->pExpr
);
1491 if( a
->wtFlags
& TERM_ORINFO
){
1492 whereOrInfoDelete(db
, a
->u
.pOrInfo
);
1493 }else if( a
->wtFlags
& TERM_ANDINFO
){
1494 whereAndInfoDelete(db
, a
->u
.pAndInfo
);
1497 if( pWC
->a
!=pWC
->aStatic
){
1498 sqlite3DbFree(db
, pWC
->a
);
1504 ** These routines walk (recursively) an expression tree and generate
1505 ** a bitmask indicating which tables are used in that expression
1508 Bitmask
sqlite3WhereExprUsageNN(WhereMaskSet
*pMaskSet
, Expr
*p
){
1510 if( p
->op
==TK_COLUMN
&& !ExprHasProperty(p
, EP_FixedCol
) ){
1511 return sqlite3WhereGetMask(pMaskSet
, p
->iTable
);
1512 }else if( ExprHasProperty(p
, EP_TokenOnly
|EP_Leaf
) ){
1513 assert( p
->op
!=TK_IF_NULL_ROW
);
1516 mask
= (p
->op
==TK_IF_NULL_ROW
) ? sqlite3WhereGetMask(pMaskSet
, p
->iTable
) : 0;
1517 if( p
->pLeft
) mask
|= sqlite3WhereExprUsageNN(pMaskSet
, p
->pLeft
);
1519 mask
|= sqlite3WhereExprUsageNN(pMaskSet
, p
->pRight
);
1520 assert( p
->x
.pList
==0 );
1521 }else if( ExprHasProperty(p
, EP_xIsSelect
) ){
1522 if( ExprHasProperty(p
, EP_VarSelect
) ) pMaskSet
->bVarSelect
= 1;
1523 mask
|= exprSelectUsage(pMaskSet
, p
->x
.pSelect
);
1524 }else if( p
->x
.pList
){
1525 mask
|= sqlite3WhereExprListUsage(pMaskSet
, p
->x
.pList
);
1527 #ifndef SQLITE_OMIT_WINDOWFUNC
1528 if( (p
->op
==TK_FUNCTION
|| p
->op
==TK_AGG_FUNCTION
) && p
->y
.pWin
){
1529 mask
|= sqlite3WhereExprListUsage(pMaskSet
, p
->y
.pWin
->pPartition
);
1530 mask
|= sqlite3WhereExprListUsage(pMaskSet
, p
->y
.pWin
->pOrderBy
);
1531 mask
|= sqlite3WhereExprUsage(pMaskSet
, p
->y
.pWin
->pFilter
);
1536 Bitmask
sqlite3WhereExprUsage(WhereMaskSet
*pMaskSet
, Expr
*p
){
1537 return p
? sqlite3WhereExprUsageNN(pMaskSet
,p
) : 0;
1539 Bitmask
sqlite3WhereExprListUsage(WhereMaskSet
*pMaskSet
, ExprList
*pList
){
1543 for(i
=0; i
<pList
->nExpr
; i
++){
1544 mask
|= sqlite3WhereExprUsage(pMaskSet
, pList
->a
[i
].pExpr
);
1552 ** Call exprAnalyze on all terms in a WHERE clause.
1554 ** Note that exprAnalyze() might add new virtual terms onto the
1555 ** end of the WHERE clause. We do not want to analyze these new
1556 ** virtual terms, so start analyzing at the end and work forward
1557 ** so that the added virtual terms are never processed.
1559 void sqlite3WhereExprAnalyze(
1560 SrcList
*pTabList
, /* the FROM clause */
1561 WhereClause
*pWC
/* the WHERE clause to be analyzed */
1564 for(i
=pWC
->nTerm
-1; i
>=0; i
--){
1565 exprAnalyze(pTabList
, pWC
, i
);
1570 ** For table-valued-functions, transform the function arguments into
1571 ** new WHERE clause terms.
1573 ** Each function argument translates into an equality constraint against
1574 ** a HIDDEN column in the table.
1576 void sqlite3WhereTabFuncArgs(
1577 Parse
*pParse
, /* Parsing context */
1578 struct SrcList_item
*pItem
, /* The FROM clause term to process */
1579 WhereClause
*pWC
/* Xfer function arguments to here */
1586 if( pItem
->fg
.isTabFunc
==0 ) return;
1589 pArgs
= pItem
->u1
.pFuncArg
;
1590 if( pArgs
==0 ) return;
1591 for(j
=k
=0; j
<pArgs
->nExpr
; j
++){
1593 while( k
<pTab
->nCol
&& (pTab
->aCol
[k
].colFlags
& COLFLAG_HIDDEN
)==0 ){k
++;}
1594 if( k
>=pTab
->nCol
){
1595 sqlite3ErrorMsg(pParse
, "too many arguments on %s() - max %d",
1599 pColRef
= sqlite3ExprAlloc(pParse
->db
, TK_COLUMN
, 0, 0);
1600 if( pColRef
==0 ) return;
1601 pColRef
->iTable
= pItem
->iCursor
;
1602 pColRef
->iColumn
= k
++;
1603 pColRef
->y
.pTab
= pTab
;
1604 pRhs
= sqlite3PExpr(pParse
, TK_UPLUS
,
1605 sqlite3ExprDup(pParse
->db
, pArgs
->a
[j
].pExpr
, 0), 0);
1606 pTerm
= sqlite3PExpr(pParse
, TK_EQ
, pColRef
, pRhs
);
1607 if( pItem
->fg
.jointype
& JT_LEFT
){
1608 sqlite3SetJoinExpr(pTerm
, pItem
->iCursor
);
1610 whereClauseInsert(pWC
, pTerm
, TERM_DYNAMIC
);