/* ** Commute a comparison operator. Expressions of the form "X op Y" ** are converted into "Y op X". ** ** If left/right precedence rules come into play when determining the ** collating sequence, then COLLATE operators are adjusted to ensure ** that the collating sequence does not change. For example: ** "Y collate NOCASE op X" becomes "X op Y" because any collation sequence on ** the left hand side of a comparison overrides any collation sequence ** attached to the right. For the same reason the EP_Collate flag ** is not commuted. */ static void exprCommute(Parse *pParse, Expr *pExpr){ u16 expRight = (pExpr->pRight->flags & EP_Collate); u16 expLeft = (pExpr->pLeft->flags & EP_Collate); assert( allowedOp(pExpr->op) && pExpr->op!=TK_IN ); if( expRight==expLeft ){ /* Either X and Y both have COLLATE operator or neither do */ if( expRight ){ /* Both X and Y have COLLATE operators. Make sure X is always ** used by clearing the EP_Collate flag from Y. */ pExpr->pRight->flags &= ~EP_Collate; }else if( sqlite3ExprCollSeq(pParse, pExpr->pLeft)!=0 ){ /* Neither X nor Y have COLLATE operators, but X has a non-default ** collating sequence. So add the EP_Collate marker on X to cause ** it to be searched first. */ pExpr->pLeft->flags |= EP_Collate; } } SWAP(Expr*,pExpr->pRight,pExpr->pLeft); if( pExpr->op>=TK_GT ){ assert( TK_LT==TK_GT+2 ); assert( TK_GE==TK_LE+2 ); assert( TK_GT>TK_EQ ); assert( TK_GT<TK_LE ); assert( pExpr->op>=TK_GT && pExpr->op<=TK_GE ); pExpr->op = ((pExpr->op-TK_GT)^2)+TK_GT; } }
/* ** The input to this routine is an ExprInfo structure with only the ** "p" field filled in. The job of this routine is to analyze the ** subexpression and populate all the other fields of the ExprInfo ** structure. */ static void exprAnalyze(SrcList *pSrc, ExprMaskSet *pMaskSet, ExprInfo *pInfo){ Expr *pExpr = pInfo->p; pInfo->prereqLeft = exprTableUsage(pMaskSet, pExpr->pLeft); pInfo->prereqRight = exprTableUsage(pMaskSet, pExpr->pRight); pInfo->prereqAll = exprTableUsage(pMaskSet, pExpr); pInfo->indexable = 0; pInfo->idxLeft = -1; pInfo->idxRight = -1; if( allowedOp(pExpr->op) && (pInfo->prereqRight & pInfo->prereqLeft)==0 ){ if( pExpr->pRight && pExpr->pRight->op==TK_COLUMN ){ pInfo->idxRight = pExpr->pRight->iTable; pInfo->indexable = 1; } if( pExpr->pLeft->op==TK_COLUMN ){ pInfo->idxLeft = pExpr->pLeft->iTable; pInfo->indexable = 1; } } if( pInfo->indexable ){ assert( pInfo->idxLeft!=pInfo->idxRight ); /* We want the expression to be of the form "X = expr", not "expr = X". ** So flip it over if necessary. If the expression is "X = Y", then ** we want Y to come from an earlier table than X. ** ** The collating sequence rule is to always choose the left expression. ** So if we do a flip, we also have to move the collating sequence. */ if( tableOrder(pSrc,pInfo->idxLeft)<tableOrder(pSrc,pInfo->idxRight) ){ assert( pExpr->op!=TK_IN ); SWAP(CollSeq*,pExpr->pRight->pColl,pExpr->pLeft->pColl); SWAP(Expr*,pExpr->pRight,pExpr->pLeft); if( pExpr->op>=TK_GT ){ assert( TK_LT==TK_GT+2 ); assert( TK_GE==TK_LE+2 ); assert( TK_GT>TK_EQ ); assert( TK_GT<TK_LE ); assert( pExpr->op>=TK_GT && pExpr->op<=TK_GE ); pExpr->op = ((pExpr->op-TK_GT)^2)+TK_GT; } SWAP(unsigned, pInfo->prereqLeft, pInfo->prereqRight); SWAP(short int, pInfo->idxLeft, pInfo->idxRight); }
/* ** The input to this routine is an ExprInfo structure with only the ** "p" field filled in. The job of this routine is to analyze the ** subexpression and populate all the other fields of the ExprInfo ** structure. */ static void exprAnalyze(ExprMaskSet *pMaskSet, ExprInfo *pInfo){ Expr *pExpr = pInfo->p; pInfo->prereqLeft = exprTableUsage(pMaskSet, pExpr->pLeft); pInfo->prereqRight = exprTableUsage(pMaskSet, pExpr->pRight); pInfo->prereqAll = exprTableUsage(pMaskSet, pExpr); pInfo->indexable = 0; pInfo->idxLeft = -1; pInfo->idxRight = -1; if( allowedOp(pExpr->op) && (pInfo->prereqRight & pInfo->prereqLeft)==0 ){ if( pExpr->pRight && pExpr->pRight->op==TK_COLUMN ){ pInfo->idxRight = pExpr->pRight->iTable; pInfo->indexable = 1; } if( pExpr->pLeft->op==TK_COLUMN ){ pInfo->idxLeft = pExpr->pLeft->iTable; pInfo->indexable = 1; } } }
/* ** Translate from TK_xx operator to WO_xx bitmask. */ static u16 operatorMask(int op){ u16 c; assert( allowedOp(op) ); if( op==TK_IN ){ c = WO_IN; }else if( op==TK_ISNULL ){ c = WO_ISNULL; }else if( op==TK_IS ){ c = WO_IS; }else{ assert( (WO_EQ<<(op-TK_EQ)) < 0x7fff ); c = (u16)(WO_EQ<<(op-TK_EQ)); } assert( op!=TK_ISNULL || c==WO_ISNULL ); assert( op!=TK_IN || c==WO_IN ); assert( op!=TK_EQ || c==WO_EQ ); assert( op!=TK_LT || c==WO_LT ); assert( op!=TK_LE || c==WO_LE ); assert( op!=TK_GT || c==WO_GT ); assert( op!=TK_GE || c==WO_GE ); assert( op!=TK_IS || c==WO_IS ); return c; }
/* ** The input to this routine is an WhereTerm structure with only the ** "pExpr" field filled in. The job of this routine is to analyze the ** subexpression and populate all the other fields of the WhereTerm ** structure. ** ** If the expression is of the form "<expr> <op> X" it gets commuted ** to the standard form of "X <op> <expr>". ** ** If the expression is of the form "X <op> Y" where both X and Y are ** columns, then the original expression is unchanged and a new virtual ** term of the form "Y <op> X" is added to the WHERE clause and ** analyzed separately. The original term is marked with TERM_COPIED ** and the new term is marked with TERM_DYNAMIC (because it's pExpr ** needs to be freed with the WhereClause) and TERM_VIRTUAL (because it ** is a commuted copy of a prior term.) The original term has nChild=1 ** and the copy has idxParent set to the index of the original term. */ static void exprAnalyze( SrcList *pSrc, /* the FROM clause */ WhereClause *pWC, /* the WHERE clause */ int idxTerm /* Index of the term to be analyzed */ ){ WhereInfo *pWInfo = pWC->pWInfo; /* WHERE clause processing context */ WhereTerm *pTerm; /* The term to be analyzed */ WhereMaskSet *pMaskSet; /* Set of table index masks */ Expr *pExpr; /* The expression to be analyzed */ Bitmask prereqLeft; /* Prerequesites of the pExpr->pLeft */ Bitmask prereqAll; /* Prerequesites of pExpr */ Bitmask extraRight = 0; /* Extra dependencies on LEFT JOIN */ Expr *pStr1 = 0; /* RHS of LIKE/GLOB operator */ int isComplete = 0; /* RHS of LIKE/GLOB ends with wildcard */ int noCase = 0; /* uppercase equivalent to lowercase */ int op; /* Top-level operator. pExpr->op */ Parse *pParse = pWInfo->pParse; /* Parsing context */ sqlite3 *db = pParse->db; /* Database connection */ if( db->mallocFailed ){ return; } pTerm = &pWC->a[idxTerm]; pMaskSet = &pWInfo->sMaskSet; pExpr = pTerm->pExpr; assert( pExpr->op!=TK_AS && pExpr->op!=TK_COLLATE ); prereqLeft = sqlite3WhereExprUsage(pMaskSet, pExpr->pLeft); op = pExpr->op; if( op==TK_IN ){ assert( pExpr->pRight==0 ); if( ExprHasProperty(pExpr, EP_xIsSelect) ){ pTerm->prereqRight = exprSelectUsage(pMaskSet, pExpr->x.pSelect); }else{ pTerm->prereqRight = sqlite3WhereExprListUsage(pMaskSet, pExpr->x.pList); } }else if( op==TK_ISNULL ){ pTerm->prereqRight = 0; }else{ pTerm->prereqRight = sqlite3WhereExprUsage(pMaskSet, pExpr->pRight); } prereqAll = sqlite3WhereExprUsage(pMaskSet, pExpr); if( ExprHasProperty(pExpr, EP_FromJoin) ){ Bitmask x = sqlite3WhereGetMask(pMaskSet, pExpr->iRightJoinTable); prereqAll |= x; extraRight = x-1; /* ON clause terms may not be used with an index ** on left table of a LEFT JOIN. Ticket #3015 */ } pTerm->prereqAll = prereqAll; pTerm->leftCursor = -1; pTerm->iParent = -1; pTerm->eOperator = 0; if( allowedOp(op) ){ Expr *pLeft = sqlite3ExprSkipCollate(pExpr->pLeft); Expr *pRight = sqlite3ExprSkipCollate(pExpr->pRight); u16 opMask = (pTerm->prereqRight & prereqLeft)==0 ? WO_ALL : WO_EQUIV; if( pLeft->op==TK_COLUMN ){ pTerm->leftCursor = pLeft->iTable; pTerm->u.leftColumn = pLeft->iColumn; pTerm->eOperator = operatorMask(op) & opMask; } if( op==TK_IS ) pTerm->wtFlags |= TERM_IS; if( pRight && pRight->op==TK_COLUMN ){ WhereTerm *pNew; Expr *pDup; u16 eExtraOp = 0; /* Extra bits for pNew->eOperator */ if( pTerm->leftCursor>=0 ){ int idxNew; pDup = sqlite3ExprDup(db, pExpr, 0); if( db->mallocFailed ){ sqlite3ExprDelete(db, pDup); return; } idxNew = whereClauseInsert(pWC, pDup, TERM_VIRTUAL|TERM_DYNAMIC); if( idxNew==0 ) return; pNew = &pWC->a[idxNew]; markTermAsChild(pWC, idxNew, idxTerm); if( op==TK_IS ) pNew->wtFlags |= TERM_IS; pTerm = &pWC->a[idxTerm]; pTerm->wtFlags |= TERM_COPIED; if( termIsEquivalence(pParse, pDup) ){ pTerm->eOperator |= WO_EQUIV; eExtraOp = WO_EQUIV; } }else{ pDup = pExpr; pNew = pTerm; } exprCommute(pParse, pDup); pLeft = sqlite3ExprSkipCollate(pDup->pLeft); pNew->leftCursor = pLeft->iTable; pNew->u.leftColumn = pLeft->iColumn; testcase( (prereqLeft | extraRight) != prereqLeft ); pNew->prereqRight = prereqLeft | extraRight; pNew->prereqAll = prereqAll; pNew->eOperator = (operatorMask(pDup->op) + eExtraOp) & opMask; } } #ifndef SQLITE_OMIT_BETWEEN_OPTIMIZATION /* If a term is the BETWEEN operator, create two new virtual terms ** that define the range that the BETWEEN implements. For example: ** ** a BETWEEN b AND c ** ** is converted into: ** ** (a BETWEEN b AND c) AND (a>=b) AND (a<=c) ** ** The two new terms are added onto the end of the WhereClause object. ** The new terms are "dynamic" and are children of the original BETWEEN ** term. That means that if the BETWEEN term is coded, the children are ** skipped. Or, if the children are satisfied by an index, the original ** BETWEEN term is skipped. */ else if( pExpr->op==TK_BETWEEN && pWC->op==TK_AND ){ ExprList *pList = pExpr->x.pList; int i; static const u8 ops[] = {TK_GE, TK_LE}; assert( pList!=0 ); assert( pList->nExpr==2 ); for(i=0; i<2; i++){ Expr *pNewExpr; int idxNew; pNewExpr = sqlite3PExpr(pParse, ops[i], sqlite3ExprDup(db, pExpr->pLeft, 0), sqlite3ExprDup(db, pList->a[i].pExpr, 0), 0); transferJoinMarkings(pNewExpr, pExpr); idxNew = whereClauseInsert(pWC, pNewExpr, TERM_VIRTUAL|TERM_DYNAMIC); testcase( idxNew==0 ); exprAnalyze(pSrc, pWC, idxNew); pTerm = &pWC->a[idxTerm]; markTermAsChild(pWC, idxNew, idxTerm); } } #endif /* SQLITE_OMIT_BETWEEN_OPTIMIZATION */ #if !defined(SQLITE_OMIT_OR_OPTIMIZATION) && !defined(SQLITE_OMIT_SUBQUERY) /* Analyze a term that is composed of two or more subterms connected by ** an OR operator. */ else if( pExpr->op==TK_OR ){ assert( pWC->op==TK_AND ); exprAnalyzeOrTerm(pSrc, pWC, idxTerm); pTerm = &pWC->a[idxTerm]; } #endif /* SQLITE_OMIT_OR_OPTIMIZATION */ #ifndef SQLITE_OMIT_LIKE_OPTIMIZATION /* Add constraints to reduce the search space on a LIKE or GLOB ** operator. ** ** A like pattern of the form "x LIKE 'aBc%'" is changed into constraints ** ** x>='ABC' AND x<'abd' AND x LIKE 'aBc%' ** ** The last character of the prefix "abc" is incremented to form the ** termination condition "abd". If case is not significant (the default ** for LIKE) then the lower-bound is made all uppercase and the upper- ** bound is made all lowercase so that the bounds also work when comparing ** BLOBs. */ if( pWC->op==TK_AND && isLikeOrGlob(pParse, pExpr, &pStr1, &isComplete, &noCase) ){ Expr *pLeft; /* LHS of LIKE/GLOB operator */ Expr *pStr2; /* Copy of pStr1 - RHS of LIKE/GLOB operator */ Expr *pNewExpr1; Expr *pNewExpr2; int idxNew1; int idxNew2; const char *zCollSeqName; /* Name of collating sequence */ const u16 wtFlags = TERM_LIKEOPT | TERM_VIRTUAL | TERM_DYNAMIC; pLeft = pExpr->x.pList->a[1].pExpr; pStr2 = sqlite3ExprDup(db, pStr1, 0); /* Convert the lower bound to upper-case and the upper bound to ** lower-case (upper-case is less than lower-case in ASCII) so that ** the range constraints also work for BLOBs */ if( noCase && !pParse->db->mallocFailed ){ int i; char c; pTerm->wtFlags |= TERM_LIKE; for(i=0; (c = pStr1->u.zToken[i])!=0; i++){ pStr1->u.zToken[i] = sqlite3Toupper(c); pStr2->u.zToken[i] = sqlite3Tolower(c); } } if( !db->mallocFailed ){ u8 c, *pC; /* Last character before the first wildcard */ pC = (u8*)&pStr2->u.zToken[sqlite3Strlen30(pStr2->u.zToken)-1]; c = *pC; if( noCase ){ /* The point is to increment the last character before the first ** wildcard. But if we increment '@', that will push it into the ** alphabetic range where case conversions will mess up the ** inequality. To avoid this, make sure to also run the full ** LIKE on all candidate expressions by clearing the isComplete flag */ if( c=='A'-1 ) isComplete = 0; c = sqlite3UpperToLower[c]; } *pC = c + 1; } zCollSeqName = noCase ? "NOCASE" : "BINARY"; pNewExpr1 = sqlite3ExprDup(db, pLeft, 0); pNewExpr1 = sqlite3PExpr(pParse, TK_GE, sqlite3ExprAddCollateString(pParse,pNewExpr1,zCollSeqName), pStr1, 0); transferJoinMarkings(pNewExpr1, pExpr); idxNew1 = whereClauseInsert(pWC, pNewExpr1, wtFlags); testcase( idxNew1==0 ); exprAnalyze(pSrc, pWC, idxNew1); pNewExpr2 = sqlite3ExprDup(db, pLeft, 0); pNewExpr2 = sqlite3PExpr(pParse, TK_LT, sqlite3ExprAddCollateString(pParse,pNewExpr2,zCollSeqName), pStr2, 0); transferJoinMarkings(pNewExpr2, pExpr); idxNew2 = whereClauseInsert(pWC, pNewExpr2, wtFlags); testcase( idxNew2==0 ); exprAnalyze(pSrc, pWC, idxNew2); pTerm = &pWC->a[idxTerm]; if( isComplete ){ markTermAsChild(pWC, idxNew1, idxTerm); markTermAsChild(pWC, idxNew2, idxTerm); } } #endif /* SQLITE_OMIT_LIKE_OPTIMIZATION */ #ifndef SQLITE_OMIT_VIRTUALTABLE /* Add a WO_MATCH auxiliary term to the constraint set if the ** current expression is of the form: column MATCH expr. ** This information is used by the xBestIndex methods of ** virtual tables. The native query optimizer does not attempt ** to do anything with MATCH functions. */ if( isMatchOfColumn(pExpr) ){ int idxNew; Expr *pRight, *pLeft; WhereTerm *pNewTerm; Bitmask prereqColumn, prereqExpr; pRight = pExpr->x.pList->a[0].pExpr; pLeft = pExpr->x.pList->a[1].pExpr; prereqExpr = sqlite3WhereExprUsage(pMaskSet, pRight); prereqColumn = sqlite3WhereExprUsage(pMaskSet, pLeft); if( (prereqExpr & prereqColumn)==0 ){ Expr *pNewExpr; pNewExpr = sqlite3PExpr(pParse, TK_MATCH, 0, sqlite3ExprDup(db, pRight, 0), 0); idxNew = whereClauseInsert(pWC, pNewExpr, TERM_VIRTUAL|TERM_DYNAMIC); testcase( idxNew==0 ); pNewTerm = &pWC->a[idxNew]; pNewTerm->prereqRight = prereqExpr; pNewTerm->leftCursor = pLeft->iTable; pNewTerm->u.leftColumn = pLeft->iColumn; pNewTerm->eOperator = WO_MATCH; markTermAsChild(pWC, idxNew, idxTerm); pTerm = &pWC->a[idxTerm]; pTerm->wtFlags |= TERM_COPIED; pNewTerm->prereqAll = pTerm->prereqAll; } } #endif /* SQLITE_OMIT_VIRTUALTABLE */ #ifdef SQLITE_ENABLE_STAT3_OR_STAT4 /* When sqlite_stat3 histogram data is available an operator of the ** form "x IS NOT NULL" can sometimes be evaluated more efficiently ** as "x>NULL" if x is not an INTEGER PRIMARY KEY. So construct a ** virtual term of that form. ** ** Note that the virtual term must be tagged with TERM_VNULL. */ if( pExpr->op==TK_NOTNULL && pExpr->pLeft->op==TK_COLUMN && pExpr->pLeft->iColumn>=0 && OptimizationEnabled(db, SQLITE_Stat34) ){ Expr *pNewExpr; Expr *pLeft = pExpr->pLeft; int idxNew; WhereTerm *pNewTerm; pNewExpr = sqlite3PExpr(pParse, TK_GT, sqlite3ExprDup(db, pLeft, 0), sqlite3PExpr(pParse, TK_NULL, 0, 0, 0), 0); idxNew = whereClauseInsert(pWC, pNewExpr, TERM_VIRTUAL|TERM_DYNAMIC|TERM_VNULL); if( idxNew ){ pNewTerm = &pWC->a[idxNew]; pNewTerm->prereqRight = 0; pNewTerm->leftCursor = pLeft->iTable; pNewTerm->u.leftColumn = pLeft->iColumn; pNewTerm->eOperator = WO_GT; markTermAsChild(pWC, idxNew, idxTerm); pTerm = &pWC->a[idxTerm]; pTerm->wtFlags |= TERM_COPIED; pNewTerm->prereqAll = pTerm->prereqAll; } } #endif /* SQLITE_ENABLE_STAT3_OR_STAT4 */ /* Prevent ON clause terms of a LEFT JOIN from being used to drive ** an index for tables to the left of the join. */ pTerm->prereqRight |= extraRight; }
/* ** Analyze a term that consists of two or more OR-connected ** subterms. So in: ** ** ... WHERE (a=5) AND (b=7 OR c=9 OR d=13) AND (d=13) ** ^^^^^^^^^^^^^^^^^^^^ ** ** This routine analyzes terms such as the middle term in the above example. ** A WhereOrTerm object is computed and attached to the term under ** analysis, regardless of the outcome of the analysis. Hence: ** ** WhereTerm.wtFlags |= TERM_ORINFO ** WhereTerm.u.pOrInfo = a dynamically allocated WhereOrTerm object ** ** The term being analyzed must have two or more of OR-connected subterms. ** A single subterm might be a set of AND-connected sub-subterms. ** Examples of terms under analysis: ** ** (A) t1.x=t2.y OR t1.x=t2.z OR t1.y=15 OR t1.z=t3.a+5 ** (B) x=expr1 OR expr2=x OR x=expr3 ** (C) t1.x=t2.y OR (t1.x=t2.z AND t1.y=15) ** (D) x=expr1 OR (y>11 AND y<22 AND z LIKE '*hello*') ** (E) (p.a=1 AND q.b=2 AND r.c=3) OR (p.x=4 AND q.y=5 AND r.z=6) ** (F) x>A OR (x=A AND y>=B) ** ** CASE 1: ** ** If all subterms are of the form T.C=expr for some single column of C and ** a single table T (as shown in example B above) then create a new virtual ** term that is an equivalent IN expression. In other words, if the term ** being analyzed is: ** ** x = expr1 OR expr2 = x OR x = expr3 ** ** then create a new virtual term like this: ** ** x IN (expr1,expr2,expr3) ** ** CASE 2: ** ** If there are exactly two disjuncts and one side has x>A and the other side ** has x=A (for the same x and A) then add a new virtual conjunct term to the ** WHERE clause of the form "x>=A". Example: ** ** x>A OR (x=A AND y>B) adds: x>=A ** ** The added conjunct can sometimes be helpful in query planning. ** ** CASE 3: ** ** If all subterms are indexable by a single table T, then set ** ** WhereTerm.eOperator = WO_OR ** WhereTerm.u.pOrInfo->indexable |= the cursor number for table T ** ** A subterm is "indexable" if it is of the form ** "T.C <op> <expr>" where C is any column of table T and ** <op> is one of "=", "<", "<=", ">", ">=", "IS NULL", or "IN". ** A subterm is also indexable if it is an AND of two or more ** subsubterms at least one of which is indexable. Indexable AND ** subterms have their eOperator set to WO_AND and they have ** u.pAndInfo set to a dynamically allocated WhereAndTerm object. ** ** From another point of view, "indexable" means that the subterm could ** potentially be used with an index if an appropriate index exists. ** This analysis does not consider whether or not the index exists; that ** is decided elsewhere. This analysis only looks at whether subterms ** appropriate for indexing exist. ** ** All examples A through E above satisfy case 3. But if a term ** also satisfies case 1 (such as B) we know that the optimizer will ** always prefer case 1, so in that case we pretend that case 3 is not ** satisfied. ** ** It might be the case that multiple tables are indexable. For example, ** (E) above is indexable on tables P, Q, and R. ** ** Terms that satisfy case 3 are candidates for lookup by using ** separate indices to find rowids for each subterm and composing ** the union of all rowids using a RowSet object. This is similar ** to "bitmap indices" in other database engines. ** ** OTHERWISE: ** ** If none of cases 1, 2, or 3 apply, then leave the eOperator set to ** zero. This term is not useful for search. */ static void exprAnalyzeOrTerm( SrcList *pSrc, /* the FROM clause */ WhereClause *pWC, /* the complete WHERE clause */ int idxTerm /* Index of the OR-term to be analyzed */ ){ WhereInfo *pWInfo = pWC->pWInfo; /* WHERE clause processing context */ Parse *pParse = pWInfo->pParse; /* Parser context */ sqlite3 *db = pParse->db; /* Database connection */ WhereTerm *pTerm = &pWC->a[idxTerm]; /* The term to be analyzed */ Expr *pExpr = pTerm->pExpr; /* The expression of the term */ int i; /* Loop counters */ WhereClause *pOrWc; /* Breakup of pTerm into subterms */ WhereTerm *pOrTerm; /* A Sub-term within the pOrWc */ WhereOrInfo *pOrInfo; /* Additional information associated with pTerm */ Bitmask chngToIN; /* Tables that might satisfy case 1 */ Bitmask indexable; /* Tables that are indexable, satisfying case 2 */ /* ** Break the OR clause into its separate subterms. The subterms are ** stored in a WhereClause structure containing within the WhereOrInfo ** object that is attached to the original OR clause term. */ assert( (pTerm->wtFlags & (TERM_DYNAMIC|TERM_ORINFO|TERM_ANDINFO))==0 ); assert( pExpr->op==TK_OR ); pTerm->u.pOrInfo = pOrInfo = sqlite3DbMallocZero(db, sizeof(*pOrInfo)); if( pOrInfo==0 ) return; pTerm->wtFlags |= TERM_ORINFO; pOrWc = &pOrInfo->wc; sqlite3WhereClauseInit(pOrWc, pWInfo); sqlite3WhereSplit(pOrWc, pExpr, TK_OR); sqlite3WhereExprAnalyze(pSrc, pOrWc); if( db->mallocFailed ) return; assert( pOrWc->nTerm>=2 ); /* ** Compute the set of tables that might satisfy cases 1 or 3. */ indexable = ~(Bitmask)0; chngToIN = ~(Bitmask)0; for(i=pOrWc->nTerm-1, pOrTerm=pOrWc->a; i>=0 && indexable; i--, pOrTerm++){ if( (pOrTerm->eOperator & WO_SINGLE)==0 ){ WhereAndInfo *pAndInfo; assert( (pOrTerm->wtFlags & (TERM_ANDINFO|TERM_ORINFO))==0 ); chngToIN = 0; pAndInfo = sqlite3DbMallocRaw(db, sizeof(*pAndInfo)); if( pAndInfo ){ WhereClause *pAndWC; WhereTerm *pAndTerm; int j; Bitmask b = 0; pOrTerm->u.pAndInfo = pAndInfo; pOrTerm->wtFlags |= TERM_ANDINFO; pOrTerm->eOperator = WO_AND; pAndWC = &pAndInfo->wc; sqlite3WhereClauseInit(pAndWC, pWC->pWInfo); sqlite3WhereSplit(pAndWC, pOrTerm->pExpr, TK_AND); sqlite3WhereExprAnalyze(pSrc, pAndWC); pAndWC->pOuter = pWC; testcase( db->mallocFailed ); if( !db->mallocFailed ){ for(j=0, pAndTerm=pAndWC->a; j<pAndWC->nTerm; j++, pAndTerm++){ assert( pAndTerm->pExpr ); if( allowedOp(pAndTerm->pExpr->op) ){ b |= sqlite3WhereGetMask(&pWInfo->sMaskSet, pAndTerm->leftCursor); } } } indexable &= b; } }else if( pOrTerm->wtFlags & TERM_COPIED ){ /* Skip this term for now. We revisit it when we process the ** corresponding TERM_VIRTUAL term */ }else{ Bitmask b; b = sqlite3WhereGetMask(&pWInfo->sMaskSet, pOrTerm->leftCursor); if( pOrTerm->wtFlags & TERM_VIRTUAL ){ WhereTerm *pOther = &pOrWc->a[pOrTerm->iParent]; b |= sqlite3WhereGetMask(&pWInfo->sMaskSet, pOther->leftCursor); } indexable &= b; if( (pOrTerm->eOperator & WO_EQ)==0 ){ chngToIN = 0; }else{ chngToIN &= b; } } } /* ** Record the set of tables that satisfy case 3. The set might be ** empty. */ pOrInfo->indexable = indexable; pTerm->eOperator = indexable==0 ? 0 : WO_OR; /* For a two-way OR, attempt to implementation case 2. */ if( indexable && pOrWc->nTerm==2 ){ int iOne = 0; WhereTerm *pOne; while( (pOne = whereNthSubterm(&pOrWc->a[0],iOne++))!=0 ){ int iTwo = 0; WhereTerm *pTwo; while( (pTwo = whereNthSubterm(&pOrWc->a[1],iTwo++))!=0 ){ whereCombineDisjuncts(pSrc, pWC, pOne, pTwo); } } } /* ** chngToIN holds a set of tables that *might* satisfy case 1. But ** we have to do some additional checking to see if case 1 really ** is satisfied. ** ** chngToIN will hold either 0, 1, or 2 bits. The 0-bit case means ** that there is no possibility of transforming the OR clause into an ** IN operator because one or more terms in the OR clause contain ** something other than == on a column in the single table. The 1-bit ** case means that every term of the OR clause is of the form ** "table.column=expr" for some single table. The one bit that is set ** will correspond to the common table. We still need to check to make ** sure the same column is used on all terms. The 2-bit case is when ** the all terms are of the form "table1.column=table2.column". It ** might be possible to form an IN operator with either table1.column ** or table2.column as the LHS if either is common to every term of ** the OR clause. ** ** Note that terms of the form "table.column1=table.column2" (the ** same table on both sizes of the ==) cannot be optimized. */ if( chngToIN ){ int okToChngToIN = 0; /* True if the conversion to IN is valid */ int iColumn = -1; /* Column index on lhs of IN operator */ int iCursor = -1; /* Table cursor common to all terms */ int j = 0; /* Loop counter */ /* Search for a table and column that appears on one side or the ** other of the == operator in every subterm. That table and column ** will be recorded in iCursor and iColumn. There might not be any ** such table and column. Set okToChngToIN if an appropriate table ** and column is found but leave okToChngToIN false if not found. */ for(j=0; j<2 && !okToChngToIN; j++){ pOrTerm = pOrWc->a; for(i=pOrWc->nTerm-1; i>=0; i--, pOrTerm++){ assert( pOrTerm->eOperator & WO_EQ ); pOrTerm->wtFlags &= ~TERM_OR_OK; if( pOrTerm->leftCursor==iCursor ){ /* This is the 2-bit case and we are on the second iteration and ** current term is from the first iteration. So skip this term. */ assert( j==1 ); continue; } if( (chngToIN & sqlite3WhereGetMask(&pWInfo->sMaskSet, pOrTerm->leftCursor))==0 ){ /* This term must be of the form t1.a==t2.b where t2 is in the ** chngToIN set but t1 is not. This term will be either preceded ** or follwed by an inverted copy (t2.b==t1.a). Skip this term ** and use its inversion. */ testcase( pOrTerm->wtFlags & TERM_COPIED ); testcase( pOrTerm->wtFlags & TERM_VIRTUAL ); assert( pOrTerm->wtFlags & (TERM_COPIED|TERM_VIRTUAL) ); continue; } iColumn = pOrTerm->u.leftColumn; iCursor = pOrTerm->leftCursor; break; } if( i<0 ){ /* No candidate table+column was found. This can only occur ** on the second iteration */ assert( j==1 ); assert( IsPowerOfTwo(chngToIN) ); assert( chngToIN==sqlite3WhereGetMask(&pWInfo->sMaskSet, iCursor) ); break; } testcase( j==1 ); /* We have found a candidate table and column. Check to see if that ** table and column is common to every term in the OR clause */ okToChngToIN = 1; for(; i>=0 && okToChngToIN; i--, pOrTerm++){ assert( pOrTerm->eOperator & WO_EQ ); if( pOrTerm->leftCursor!=iCursor ){ pOrTerm->wtFlags &= ~TERM_OR_OK; }else if( pOrTerm->u.leftColumn!=iColumn ){ okToChngToIN = 0; }else{ int affLeft, affRight; /* If the right-hand side is also a column, then the affinities ** of both right and left sides must be such that no type ** conversions are required on the right. (Ticket #2249) */ affRight = sqlite3ExprAffinity(pOrTerm->pExpr->pRight); affLeft = sqlite3ExprAffinity(pOrTerm->pExpr->pLeft); if( affRight!=0 && affRight!=affLeft ){ okToChngToIN = 0; }else{ pOrTerm->wtFlags |= TERM_OR_OK; } } } } /* At this point, okToChngToIN is true if original pTerm satisfies ** case 1. In that case, construct a new virtual term that is ** pTerm converted into an IN operator. */ if( okToChngToIN ){ Expr *pDup; /* A transient duplicate expression */ ExprList *pList = 0; /* The RHS of the IN operator */ Expr *pLeft = 0; /* The LHS of the IN operator */ Expr *pNew; /* The complete IN operator */ for(i=pOrWc->nTerm-1, pOrTerm=pOrWc->a; i>=0; i--, pOrTerm++){ if( (pOrTerm->wtFlags & TERM_OR_OK)==0 ) continue; assert( pOrTerm->eOperator & WO_EQ ); assert( pOrTerm->leftCursor==iCursor ); assert( pOrTerm->u.leftColumn==iColumn ); pDup = sqlite3ExprDup(db, pOrTerm->pExpr->pRight, 0); pList = sqlite3ExprListAppend(pWInfo->pParse, pList, pDup); pLeft = pOrTerm->pExpr->pLeft; } assert( pLeft!=0 ); pDup = sqlite3ExprDup(db, pLeft, 0); pNew = sqlite3PExpr(pParse, TK_IN, pDup, 0, 0); if( pNew ){ int idxNew; transferJoinMarkings(pNew, pExpr); assert( !ExprHasProperty(pNew, EP_xIsSelect) ); pNew->x.pList = pList; idxNew = whereClauseInsert(pWC, pNew, TERM_VIRTUAL|TERM_DYNAMIC); testcase( idxNew==0 ); exprAnalyze(pSrc, pWC, idxNew); pTerm = &pWC->a[idxTerm]; markTermAsChild(pWC, idxNew, idxTerm); }else{ sqlite3ExprListDelete(db, pList); } pTerm->eOperator = WO_NOOP; /* case 1 trumps case 3 */ } } }