/* ** 2001 September 15 ** ** The author disclaims copyright to this source code. In place of ** a legal notice, here is a blessing: ** ** May you do good and not evil. ** May you find forgiveness for yourself and forgive others. ** May you share freely, never taking more than you give. ** ************************************************************************* ** This module contains C code that generates VDBE code used to process ** the WHERE clause of SQL statements. This module is responsible for ** generating the code that loops through a table looking for applicable ** rows. Indices are selected and used to speed the search when doing ** so is applicable. Because this module is responsible for selecting ** indices, you might also think of this module as the "query optimizer". */ #include "sqliteInt.h" #include "whereInt.h" /* ** Return the estimated number of output rows from a WHERE clause */ u64 sqlite3WhereOutputRowCount(WhereInfo *pWInfo){ return sqlite3LogEstToInt(pWInfo->nRowOut); } /* ** Return one of the WHERE_DISTINCT_xxxxx values to indicate how this ** WHERE clause returns outputs for DISTINCT processing. */ int sqlite3WhereIsDistinct(WhereInfo *pWInfo){ return pWInfo->eDistinct; } /* ** Return TRUE if the WHERE clause returns rows in ORDER BY order. ** Return FALSE if the output needs to be sorted. */ int sqlite3WhereIsOrdered(WhereInfo *pWInfo){ return pWInfo->nOBSat; } /* ** Return the VDBE address or label to jump to in order to continue ** immediately with the next row of a WHERE clause. */ int sqlite3WhereContinueLabel(WhereInfo *pWInfo){ assert( pWInfo->iContinue!=0 ); return pWInfo->iContinue; } /* ** Return the VDBE address or label to jump to in order to break ** out of a WHERE loop. */ int sqlite3WhereBreakLabel(WhereInfo *pWInfo){ return pWInfo->iBreak; } /* ** Return TRUE if an UPDATE or DELETE statement can operate directly on ** the rowids returned by a WHERE clause. Return FALSE if doing an ** UPDATE or DELETE might change subsequent WHERE clause results. ** ** If the ONEPASS optimization is used (if this routine returns true) ** then also write the indices of open cursors used by ONEPASS ** into aiCur[0] and aiCur[1]. iaCur[0] gets the cursor of the data ** table and iaCur[1] gets the cursor used by an auxiliary index. ** Either value may be -1, indicating that cursor is not used. ** Any cursors returned will have been opened for writing. ** ** aiCur[0] and aiCur[1] both get -1 if the where-clause logic is ** unable to use the ONEPASS optimization. */ int sqlite3WhereOkOnePass(WhereInfo *pWInfo, int *aiCur){ memcpy(aiCur, pWInfo->aiCurOnePass, sizeof(int)*2); return pWInfo->okOnePass; } /* ** Move the content of pSrc into pDest */ static void whereOrMove(WhereOrSet *pDest, WhereOrSet *pSrc){ pDest->n = pSrc->n; memcpy(pDest->a, pSrc->a, pDest->n*sizeof(pDest->a[0])); } /* ** Try to insert a new prerequisite/cost entry into the WhereOrSet pSet. ** ** The new entry might overwrite an existing entry, or it might be ** appended, or it might be discarded. Do whatever is the right thing ** so that pSet keeps the N_OR_COST best entries seen so far. */ static int whereOrInsert( WhereOrSet *pSet, /* The WhereOrSet to be updated */ Bitmask prereq, /* Prerequisites of the new entry */ LogEst rRun, /* Run-cost of the new entry */ LogEst nOut /* Number of outputs for the new entry */ ){ u16 i; WhereOrCost *p; for(i=pSet->n, p=pSet->a; i>0; i--, p++){ if( rRun<=p->rRun && (prereq & p->prereq)==prereq ){ goto whereOrInsert_done; } if( p->rRun<=rRun && (p->prereq & prereq)==p->prereq ){ return 0; } } if( pSet->na[pSet->n++]; p->nOut = nOut; }else{ p = pSet->a; for(i=1; in; i++){ if( p->rRun>pSet->a[i].rRun ) p = pSet->a + i; } if( p->rRun<=rRun ) return 0; } whereOrInsert_done: p->prereq = prereq; p->rRun = rRun; if( p->nOut>nOut ) p->nOut = nOut; return 1; } /* ** Initialize a preallocated WhereClause structure. */ static void whereClauseInit( WhereClause *pWC, /* The WhereClause to be initialized */ WhereInfo *pWInfo /* The WHERE processing context */ ){ pWC->pWInfo = pWInfo; pWC->pOuter = 0; pWC->nTerm = 0; pWC->nSlot = ArraySize(pWC->aStatic); pWC->a = pWC->aStatic; } /* Forward reference */ static void whereClauseClear(WhereClause*); /* ** Deallocate all memory associated with a WhereOrInfo object. */ static void whereOrInfoDelete(sqlite3 *db, WhereOrInfo *p){ whereClauseClear(&p->wc); sqlite3DbFree(db, p); } /* ** Deallocate all memory associated with a WhereAndInfo object. */ static void whereAndInfoDelete(sqlite3 *db, WhereAndInfo *p){ whereClauseClear(&p->wc); sqlite3DbFree(db, p); } /* ** Deallocate a WhereClause structure. The WhereClause structure ** itself is not freed. This routine is the inverse of whereClauseInit(). */ static void whereClauseClear(WhereClause *pWC){ int i; WhereTerm *a; sqlite3 *db = pWC->pWInfo->pParse->db; for(i=pWC->nTerm-1, a=pWC->a; i>=0; i--, a++){ if( a->wtFlags & TERM_DYNAMIC ){ sqlite3ExprDelete(db, a->pExpr); } if( a->wtFlags & TERM_ORINFO ){ whereOrInfoDelete(db, a->u.pOrInfo); }else if( a->wtFlags & TERM_ANDINFO ){ whereAndInfoDelete(db, a->u.pAndInfo); } } if( pWC->a!=pWC->aStatic ){ sqlite3DbFree(db, pWC->a); } } /* ** Add a single new WhereTerm entry to the WhereClause object pWC. ** The new WhereTerm object is constructed from Expr p and with wtFlags. ** The index in pWC->a[] of the new WhereTerm is returned on success. ** 0 is returned if the new WhereTerm could not be added due to a memory ** allocation error. The memory allocation failure will be recorded in ** the db->mallocFailed flag so that higher-level functions can detect it. ** ** This routine will increase the size of the pWC->a[] array as necessary. ** ** If the wtFlags argument includes TERM_DYNAMIC, then responsibility ** for freeing the expression p is assumed by the WhereClause object pWC. ** This is true even if this routine fails to allocate a new WhereTerm. ** ** WARNING: This routine might reallocate the space used to store ** WhereTerms. All pointers to WhereTerms should be invalidated after ** calling this routine. Such pointers may be reinitialized by referencing ** the pWC->a[] array. */ static int whereClauseInsert(WhereClause *pWC, Expr *p, u8 wtFlags){ WhereTerm *pTerm; int idx; testcase( wtFlags & TERM_VIRTUAL ); if( pWC->nTerm>=pWC->nSlot ){ WhereTerm *pOld = pWC->a; sqlite3 *db = pWC->pWInfo->pParse->db; pWC->a = sqlite3DbMallocRaw(db, sizeof(pWC->a[0])*pWC->nSlot*2 ); if( pWC->a==0 ){ if( wtFlags & TERM_DYNAMIC ){ sqlite3ExprDelete(db, p); } pWC->a = pOld; return 0; } memcpy(pWC->a, pOld, sizeof(pWC->a[0])*pWC->nTerm); if( pOld!=pWC->aStatic ){ sqlite3DbFree(db, pOld); } pWC->nSlot = sqlite3DbMallocSize(db, pWC->a)/sizeof(pWC->a[0]); } pTerm = &pWC->a[idx = pWC->nTerm++]; if( p && ExprHasProperty(p, EP_Unlikely) ){ pTerm->truthProb = sqlite3LogEst(p->iTable) - 270; }else{ pTerm->truthProb = 1; } pTerm->pExpr = sqlite3ExprSkipCollate(p); pTerm->wtFlags = wtFlags; pTerm->pWC = pWC; pTerm->iParent = -1; return idx; } /* ** This routine identifies subexpressions in the WHERE clause where ** each subexpression is separated by the AND operator or some other ** operator specified in the op parameter. The WhereClause structure ** is filled with pointers to subexpressions. For example: ** ** WHERE a=='hello' AND coalesce(b,11)<10 AND (c+12!=d OR c==22) ** \________/ \_______________/ \________________/ ** slot[0] slot[1] slot[2] ** ** The original WHERE clause in pExpr is unaltered. All this routine ** does is make slot[] entries point to substructure within pExpr. ** ** In the previous sentence and in the diagram, "slot[]" refers to ** the WhereClause.a[] array. The slot[] array grows as needed to contain ** all terms of the WHERE clause. */ static void whereSplit(WhereClause *pWC, Expr *pExpr, u8 op){ pWC->op = op; if( pExpr==0 ) return; if( pExpr->op!=op ){ whereClauseInsert(pWC, pExpr, 0); }else{ whereSplit(pWC, pExpr->pLeft, op); whereSplit(pWC, pExpr->pRight, op); } } /* ** Initialize a WhereMaskSet object */ #define initMaskSet(P) (P)->n=0 /* ** Return the bitmask for the given cursor number. Return 0 if ** iCursor is not in the set. */ static Bitmask getMask(WhereMaskSet *pMaskSet, int iCursor){ int i; assert( pMaskSet->n<=(int)sizeof(Bitmask)*8 ); for(i=0; in; i++){ if( pMaskSet->ix[i]==iCursor ){ return MASKBIT(i); } } return 0; } /* ** Create a new mask for cursor iCursor. ** ** There is one cursor per table in the FROM clause. The number of ** tables in the FROM clause is limited by a test early in the ** sqlite3WhereBegin() routine. So we know that the pMaskSet->ix[] ** array will never overflow. */ static void createMask(WhereMaskSet *pMaskSet, int iCursor){ assert( pMaskSet->n < ArraySize(pMaskSet->ix) ); pMaskSet->ix[pMaskSet->n++] = iCursor; } /* ** These routines walk (recursively) an expression tree and generate ** a bitmask indicating which tables are used in that expression ** tree. */ static Bitmask exprListTableUsage(WhereMaskSet*, ExprList*); static Bitmask exprSelectTableUsage(WhereMaskSet*, Select*); static Bitmask exprTableUsage(WhereMaskSet *pMaskSet, Expr *p){ Bitmask mask = 0; if( p==0 ) return 0; if( p->op==TK_COLUMN ){ mask = getMask(pMaskSet, p->iTable); return mask; } mask = exprTableUsage(pMaskSet, p->pRight); mask |= exprTableUsage(pMaskSet, p->pLeft); if( ExprHasProperty(p, EP_xIsSelect) ){ mask |= exprSelectTableUsage(pMaskSet, p->x.pSelect); }else{ mask |= exprListTableUsage(pMaskSet, p->x.pList); } return mask; } static Bitmask exprListTableUsage(WhereMaskSet *pMaskSet, ExprList *pList){ int i; Bitmask mask = 0; if( pList ){ for(i=0; inExpr; i++){ mask |= exprTableUsage(pMaskSet, pList->a[i].pExpr); } } return mask; } static Bitmask exprSelectTableUsage(WhereMaskSet *pMaskSet, Select *pS){ Bitmask mask = 0; while( pS ){ SrcList *pSrc = pS->pSrc; mask |= exprListTableUsage(pMaskSet, pS->pEList); mask |= exprListTableUsage(pMaskSet, pS->pGroupBy); mask |= exprListTableUsage(pMaskSet, pS->pOrderBy); mask |= exprTableUsage(pMaskSet, pS->pWhere); mask |= exprTableUsage(pMaskSet, pS->pHaving); if( ALWAYS(pSrc!=0) ){ int i; for(i=0; inSrc; i++){ mask |= exprSelectTableUsage(pMaskSet, pSrc->a[i].pSelect); mask |= exprTableUsage(pMaskSet, pSrc->a[i].pOn); } } pS = pS->pPrior; } return mask; } /* ** Return TRUE if the given operator is one of the operators that is ** allowed for an indexable WHERE clause term. The allowed operators are ** "=", "<", ">", "<=", ">=", "IN", and "IS NULL" */ static int allowedOp(int op){ assert( TK_GT>TK_EQ && TK_GTTK_EQ && TK_LTTK_EQ && TK_LE=TK_EQ && op<=TK_GE) || op==TK_ISNULL; } /* ** 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_GTop>=TK_GT && pExpr->op<=TK_GE ); pExpr->op = ((pExpr->op-TK_GT)^2)+TK_GT; } } /* ** 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{ 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 ); return c; } /* ** Advance to the next WhereTerm that matches according to the criteria ** established when the pScan object was initialized by whereScanInit(). ** Return NULL if there are no more matching WhereTerms. */ static WhereTerm *whereScanNext(WhereScan *pScan){ int iCur; /* The cursor on the LHS of the term */ int iColumn; /* The column on the LHS of the term. -1 for IPK */ Expr *pX; /* An expression being tested */ WhereClause *pWC; /* Shorthand for pScan->pWC */ WhereTerm *pTerm; /* The term being tested */ int k = pScan->k; /* Where to start scanning */ while( pScan->iEquiv<=pScan->nEquiv ){ iCur = pScan->aEquiv[pScan->iEquiv-2]; iColumn = pScan->aEquiv[pScan->iEquiv-1]; while( (pWC = pScan->pWC)!=0 ){ for(pTerm=pWC->a+k; knTerm; k++, pTerm++){ if( pTerm->leftCursor==iCur && pTerm->u.leftColumn==iColumn && (pScan->iEquiv<=2 || !ExprHasProperty(pTerm->pExpr, EP_FromJoin)) ){ if( (pTerm->eOperator & WO_EQUIV)!=0 && pScan->nEquivaEquiv) ){ int j; pX = sqlite3ExprSkipCollate(pTerm->pExpr->pRight); assert( pX->op==TK_COLUMN ); for(j=0; jnEquiv; j+=2){ if( pScan->aEquiv[j]==pX->iTable && pScan->aEquiv[j+1]==pX->iColumn ){ break; } } if( j==pScan->nEquiv ){ pScan->aEquiv[j] = pX->iTable; pScan->aEquiv[j+1] = pX->iColumn; pScan->nEquiv += 2; } } if( (pTerm->eOperator & pScan->opMask)!=0 ){ /* Verify the affinity and collating sequence match */ if( pScan->zCollName && (pTerm->eOperator & WO_ISNULL)==0 ){ CollSeq *pColl; Parse *pParse = pWC->pWInfo->pParse; pX = pTerm->pExpr; if( !sqlite3IndexAffinityOk(pX, pScan->idxaff) ){ continue; } assert(pX->pLeft); pColl = sqlite3BinaryCompareCollSeq(pParse, pX->pLeft, pX->pRight); if( pColl==0 ) pColl = pParse->db->pDfltColl; if( sqlite3StrICmp(pColl->zName, pScan->zCollName) ){ continue; } } if( (pTerm->eOperator & WO_EQ)!=0 && (pX = pTerm->pExpr->pRight)->op==TK_COLUMN && pX->iTable==pScan->aEquiv[0] && pX->iColumn==pScan->aEquiv[1] ){ continue; } pScan->k = k+1; return pTerm; } } } pScan->pWC = pScan->pWC->pOuter; k = 0; } pScan->pWC = pScan->pOrigWC; k = 0; pScan->iEquiv += 2; } return 0; } /* ** Initialize a WHERE clause scanner object. Return a pointer to the ** first match. Return NULL if there are no matches. ** ** The scanner will be searching the WHERE clause pWC. It will look ** for terms of the form "X " where X is column iColumn of table ** iCur. The must be one of the operators described by opMask. ** ** If the search is for X and the WHERE clause contains terms of the ** form X=Y then this routine might also return terms of the form ** "Y ". The number of levels of transitivity is limited, ** but is enough to handle most commonly occurring SQL statements. ** ** If X is not the INTEGER PRIMARY KEY then X must be compatible with ** index pIdx. */ static WhereTerm *whereScanInit( WhereScan *pScan, /* The WhereScan object being initialized */ WhereClause *pWC, /* The WHERE clause to be scanned */ int iCur, /* Cursor to scan for */ int iColumn, /* Column to scan for */ u32 opMask, /* Operator(s) to scan for */ Index *pIdx /* Must be compatible with this index */ ){ int j; /* memset(pScan, 0, sizeof(*pScan)); */ pScan->pOrigWC = pWC; pScan->pWC = pWC; if( pIdx && iColumn>=0 ){ pScan->idxaff = pIdx->pTable->aCol[iColumn].affinity; for(j=0; pIdx->aiColumn[j]!=iColumn; j++){ if( NEVER(j>pIdx->nColumn) ) return 0; } pScan->zCollName = pIdx->azColl[j]; }else{ pScan->idxaff = 0; pScan->zCollName = 0; } pScan->opMask = opMask; pScan->k = 0; pScan->aEquiv[0] = iCur; pScan->aEquiv[1] = iColumn; pScan->nEquiv = 2; pScan->iEquiv = 2; return whereScanNext(pScan); } /* ** Search for a term in the WHERE clause that is of the form "X " ** where X is a reference to the iColumn of table iCur and is one of ** the WO_xx operator codes specified by the op parameter. ** Return a pointer to the term. Return 0 if not found. ** ** The term returned might by Y= if there is another constraint in ** the WHERE clause that specifies that X=Y. Any such constraints will be ** identified by the WO_EQUIV bit in the pTerm->eOperator field. The ** aEquiv[] array holds X and all its equivalents, with each SQL variable ** taking up two slots in aEquiv[]. The first slot is for the cursor number ** and the second is for the column number. There are 22 slots in aEquiv[] ** so that means we can look for X plus up to 10 other equivalent values. ** Hence a search for X will return if X=A1 and A1=A2 and A2=A3 ** and ... and A9=A10 and A10=. ** ** If there are multiple terms in the WHERE clause of the form "X " ** then try for the one with no dependencies on - in other words where ** is a constant expression of some kind. Only return entries of ** the form "X Y" where Y is a column in another table if no terms of ** the form "X " exist. If no terms with a constant RHS ** exist, try to return a term that does not use WO_EQUIV. */ static WhereTerm *findTerm( WhereClause *pWC, /* The WHERE clause to be searched */ int iCur, /* Cursor number of LHS */ int iColumn, /* Column number of LHS */ Bitmask notReady, /* RHS must not overlap with this mask */ u32 op, /* Mask of WO_xx values describing operator */ Index *pIdx /* Must be compatible with this index, if not NULL */ ){ WhereTerm *pResult = 0; WhereTerm *p; WhereScan scan; p = whereScanInit(&scan, pWC, iCur, iColumn, op, pIdx); while( p ){ if( (p->prereqRight & notReady)==0 ){ if( p->prereqRight==0 && (p->eOperator&WO_EQ)!=0 ){ return p; } if( pResult==0 ) pResult = p; } p = whereScanNext(&scan); } return pResult; } /* Forward reference */ static void exprAnalyze(SrcList*, WhereClause*, int); /* ** Call exprAnalyze on all terms in a WHERE clause. */ static void exprAnalyzeAll( SrcList *pTabList, /* the FROM clause */ WhereClause *pWC /* the WHERE clause to be analyzed */ ){ int i; for(i=pWC->nTerm-1; i>=0; i--){ exprAnalyze(pTabList, pWC, i); } } #ifndef SQLITE_OMIT_LIKE_OPTIMIZATION /* ** Check to see if the given expression is a LIKE or GLOB operator that ** can be optimized using inequality constraints. Return TRUE if it is ** so and false if not. ** ** In order for the operator to be optimizible, the RHS must be a string ** literal that does not begin with a wildcard. */ static int isLikeOrGlob( Parse *pParse, /* Parsing and code generating context */ Expr *pExpr, /* Test this expression */ Expr **ppPrefix, /* Pointer to TK_STRING expression with pattern prefix */ int *pisComplete, /* True if the only wildcard is % in the last character */ int *pnoCase /* True if uppercase is equivalent to lowercase */ ){ const char *z = 0; /* String on RHS of LIKE operator */ Expr *pRight, *pLeft; /* Right and left size of LIKE operator */ ExprList *pList; /* List of operands to the LIKE operator */ int c; /* One character in z[] */ int cnt; /* Number of non-wildcard prefix characters */ char wc[3]; /* Wildcard characters */ sqlite3 *db = pParse->db; /* Database connection */ sqlite3_value *pVal = 0; int op; /* Opcode of pRight */ if( !sqlite3IsLikeFunction(db, pExpr, pnoCase, wc) ){ return 0; } #ifdef SQLITE_EBCDIC if( *pnoCase ) return 0; #endif pList = pExpr->x.pList; pLeft = pList->a[1].pExpr; if( pLeft->op!=TK_COLUMN || sqlite3ExprAffinity(pLeft)!=SQLITE_AFF_TEXT || IsVirtual(pLeft->pTab) ){ /* IMP: R-02065-49465 The left-hand side of the LIKE or GLOB operator must ** be the name of an indexed column with TEXT affinity. */ return 0; } assert( pLeft->iColumn!=(-1) ); /* Because IPK never has AFF_TEXT */ pRight = sqlite3ExprSkipCollate(pList->a[0].pExpr); op = pRight->op; if( op==TK_VARIABLE ){ Vdbe *pReprepare = pParse->pReprepare; int iCol = pRight->iColumn; pVal = sqlite3VdbeGetBoundValue(pReprepare, iCol, SQLITE_AFF_NONE); if( pVal && sqlite3_value_type(pVal)==SQLITE_TEXT ){ z = (char *)sqlite3_value_text(pVal); } sqlite3VdbeSetVarmask(pParse->pVdbe, iCol); assert( pRight->op==TK_VARIABLE || pRight->op==TK_REGISTER ); }else if( op==TK_STRING ){ z = pRight->u.zToken; } if( z ){ cnt = 0; while( (c=z[cnt])!=0 && c!=wc[0] && c!=wc[1] && c!=wc[2] ){ cnt++; } if( cnt!=0 && 255!=(u8)z[cnt-1] ){ Expr *pPrefix; *pisComplete = c==wc[0] && z[cnt+1]==0; pPrefix = sqlite3Expr(db, TK_STRING, z); if( pPrefix ) pPrefix->u.zToken[cnt] = 0; *ppPrefix = pPrefix; if( op==TK_VARIABLE ){ Vdbe *v = pParse->pVdbe; sqlite3VdbeSetVarmask(v, pRight->iColumn); if( *pisComplete && pRight->u.zToken[1] ){ /* If the rhs of the LIKE expression is a variable, and the current ** value of the variable means there is no need to invoke the LIKE ** function, then no OP_Variable will be added to the program. ** This causes problems for the sqlite3_bind_parameter_name() ** API. To work around them, add a dummy OP_Variable here. */ int r1 = sqlite3GetTempReg(pParse); sqlite3ExprCodeTarget(pParse, pRight, r1); sqlite3VdbeChangeP3(v, sqlite3VdbeCurrentAddr(v)-1, 0); sqlite3ReleaseTempReg(pParse, r1); } } }else{ z = 0; } } sqlite3ValueFree(pVal); return (z!=0); } #endif /* SQLITE_OMIT_LIKE_OPTIMIZATION */ #ifndef SQLITE_OMIT_VIRTUALTABLE /* ** Check to see if the given expression is of the form ** ** column MATCH expr ** ** If it is then return TRUE. If not, return FALSE. */ static int isMatchOfColumn( Expr *pExpr /* Test this expression */ ){ ExprList *pList; if( pExpr->op!=TK_FUNCTION ){ return 0; } if( sqlite3StrICmp(pExpr->u.zToken,"match")!=0 ){ return 0; } pList = pExpr->x.pList; if( pList->nExpr!=2 ){ return 0; } if( pList->a[1].pExpr->op != TK_COLUMN ){ return 0; } return 1; } #endif /* SQLITE_OMIT_VIRTUALTABLE */ /* ** If the pBase expression originated in the ON or USING clause of ** a join, then transfer the appropriate markings over to derived. */ static void transferJoinMarkings(Expr *pDerived, Expr *pBase){ if( pDerived ){ pDerived->flags |= pBase->flags & EP_FromJoin; pDerived->iRightJoinTable = pBase->iRightJoinTable; } } /* ** Mark term iChild as being a child of term iParent */ static void markTermAsChild(WhereClause *pWC, int iChild, int iParent){ pWC->a[iChild].iParent = iParent; pWC->a[iChild].truthProb = pWC->a[iParent].truthProb; pWC->a[iParent].nChild++; } #if !defined(SQLITE_OMIT_OR_OPTIMIZATION) && !defined(SQLITE_OMIT_SUBQUERY) /* ** 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) ** ** 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 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 " where C is any column of table T and ** 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 2. 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 2 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 2 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 neither case 1 nor case 2 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; whereClauseInit(pOrWc, pWInfo); whereSplit(pOrWc, pExpr, TK_OR); exprAnalyzeAll(pSrc, pOrWc); if( db->mallocFailed ) return; assert( pOrWc->nTerm>=2 ); /* ** Compute the set of tables that might satisfy cases 1 or 2. */ 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; whereClauseInit(pAndWC, pWC->pWInfo); whereSplit(pAndWC, pOrTerm->pExpr, TK_AND); exprAnalyzeAll(pSrc, pAndWC); pAndWC->pOuter = pWC; testcase( db->mallocFailed ); if( !db->mallocFailed ){ for(j=0, pAndTerm=pAndWC->a; jnTerm; j++, pAndTerm++){ assert( pAndTerm->pExpr ); if( allowedOp(pAndTerm->pExpr->op) ){ b |= getMask(&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 = getMask(&pWInfo->sMaskSet, pOrTerm->leftCursor); if( pOrTerm->wtFlags & TERM_VIRTUAL ){ WhereTerm *pOther = &pOrWc->a[pOrTerm->iParent]; b |= getMask(&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 2. The set might be ** empty. */ pOrInfo->indexable = indexable; pTerm->eOperator = indexable==0 ? 0 : WO_OR; /* ** 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 & getMask(&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==getMask(&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 2 */ } } } #endif /* !SQLITE_OMIT_OR_OPTIMIZATION && !SQLITE_OMIT_SUBQUERY */ /* ** 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 " X" it gets commuted ** to the standard form of "X ". ** ** If the expression is of the form "X Y" where both X and Y are ** columns, then the original expression is unchanged and a new virtual ** term of the form "Y 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; /* LIKE/GLOB distinguishes case */ 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 = exprTableUsage(pMaskSet, pExpr->pLeft); op = pExpr->op; if( op==TK_IN ){ assert( pExpr->pRight==0 ); if( ExprHasProperty(pExpr, EP_xIsSelect) ){ pTerm->prereqRight = exprSelectTableUsage(pMaskSet, pExpr->x.pSelect); }else{ pTerm->prereqRight = exprListTableUsage(pMaskSet, pExpr->x.pList); } }else if( op==TK_ISNULL ){ pTerm->prereqRight = 0; }else{ pTerm->prereqRight = exprTableUsage(pMaskSet, pExpr->pRight); } prereqAll = exprTableUsage(pMaskSet, pExpr); if( ExprHasProperty(pExpr, EP_FromJoin) ){ Bitmask x = getMask(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( 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); pTerm = &pWC->a[idxTerm]; pTerm->wtFlags |= TERM_COPIED; if( pExpr->op==TK_EQ && !ExprHasProperty(pExpr, EP_FromJoin) && OptimizationEnabled(db, SQLITE_Transitive) ){ 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( 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; Token sCollSeqName; /* Name of collating sequence */ pLeft = pExpr->x.pList->a[1].pExpr; pStr2 = sqlite3ExprDup(db, pStr1, 0); 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; } sCollSeqName.z = noCase ? "NOCASE" : "BINARY"; sCollSeqName.n = 6; pNewExpr1 = sqlite3ExprDup(db, pLeft, 0); pNewExpr1 = sqlite3PExpr(pParse, TK_GE, sqlite3ExprAddCollateToken(pParse,pNewExpr1,&sCollSeqName), pStr1, 0); transferJoinMarkings(pNewExpr1, pExpr); idxNew1 = whereClauseInsert(pWC, pNewExpr1, TERM_VIRTUAL|TERM_DYNAMIC); testcase( idxNew1==0 ); exprAnalyze(pSrc, pWC, idxNew1); pNewExpr2 = sqlite3ExprDup(db, pLeft, 0); pNewExpr2 = sqlite3PExpr(pParse, TK_LT, sqlite3ExprAddCollateToken(pParse,pNewExpr2,&sCollSeqName), pStr2, 0); transferJoinMarkings(pNewExpr2, pExpr); idxNew2 = whereClauseInsert(pWC, pNewExpr2, TERM_VIRTUAL|TERM_DYNAMIC); 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 = exprTableUsage(pMaskSet, pRight); prereqColumn = exprTableUsage(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. This ** TERM_VNULL tag will suppress the not-null check at the beginning ** of the loop. Without the TERM_VNULL flag, the not-null check at ** the start of the loop will prevent any results from being returned. */ 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; } /* ** This function searches pList for an entry that matches the iCol-th column ** of index pIdx. ** ** If such an expression is found, its index in pList->a[] is returned. If ** no expression is found, -1 is returned. */ static int findIndexCol( Parse *pParse, /* Parse context */ ExprList *pList, /* Expression list to search */ int iBase, /* Cursor for table associated with pIdx */ Index *pIdx, /* Index to match column of */ int iCol /* Column of index to match */ ){ int i; const char *zColl = pIdx->azColl[iCol]; for(i=0; inExpr; i++){ Expr *p = sqlite3ExprSkipCollate(pList->a[i].pExpr); if( p->op==TK_COLUMN && p->iColumn==pIdx->aiColumn[iCol] && p->iTable==iBase ){ CollSeq *pColl = sqlite3ExprCollSeq(pParse, pList->a[i].pExpr); if( ALWAYS(pColl) && 0==sqlite3StrICmp(pColl->zName, zColl) ){ return i; } } } return -1; } /* ** Return true if the DISTINCT expression-list passed as the third argument ** is redundant. ** ** A DISTINCT list is redundant if the database contains some subset of ** columns that are unique and non-null. */ static int isDistinctRedundant( Parse *pParse, /* Parsing context */ SrcList *pTabList, /* The FROM clause */ WhereClause *pWC, /* The WHERE clause */ ExprList *pDistinct /* The result set that needs to be DISTINCT */ ){ Table *pTab; Index *pIdx; int i; int iBase; /* If there is more than one table or sub-select in the FROM clause of ** this query, then it will not be possible to show that the DISTINCT ** clause is redundant. */ if( pTabList->nSrc!=1 ) return 0; iBase = pTabList->a[0].iCursor; pTab = pTabList->a[0].pTab; /* If any of the expressions is an IPK column on table iBase, then return ** true. Note: The (p->iTable==iBase) part of this test may be false if the ** current SELECT is a correlated sub-query. */ for(i=0; inExpr; i++){ Expr *p = sqlite3ExprSkipCollate(pDistinct->a[i].pExpr); if( p->op==TK_COLUMN && p->iTable==iBase && p->iColumn<0 ) return 1; } /* Loop through all indices on the table, checking each to see if it makes ** the DISTINCT qualifier redundant. It does so if: ** ** 1. The index is itself UNIQUE, and ** ** 2. All of the columns in the index are either part of the pDistinct ** list, or else the WHERE clause contains a term of the form "col=X", ** where X is a constant value. The collation sequences of the ** comparison and select-list expressions must match those of the index. ** ** 3. All of those index columns for which the WHERE clause does not ** contain a "col=X" term are subject to a NOT NULL constraint. */ for(pIdx=pTab->pIndex; pIdx; pIdx=pIdx->pNext){ if( !IsUniqueIndex(pIdx) ) continue; for(i=0; inKeyCol; i++){ i16 iCol = pIdx->aiColumn[i]; if( 0==findTerm(pWC, iBase, iCol, ~(Bitmask)0, WO_EQ, pIdx) ){ int iIdxCol = findIndexCol(pParse, pDistinct, iBase, pIdx, i); if( iIdxCol<0 || pTab->aCol[iCol].notNull==0 ){ break; } } } if( i==pIdx->nKeyCol ){ /* This index implies that the DISTINCT qualifier is redundant. */ return 1; } } return 0; } /* ** Estimate the logarithm of the input value to base 2. */ static LogEst estLog(LogEst N){ return N<=10 ? 0 : sqlite3LogEst(N) - 33; } /* ** Two routines for printing the content of an sqlite3_index_info ** structure. Used for testing and debugging only. If neither ** SQLITE_TEST or SQLITE_DEBUG are defined, then these routines ** are no-ops. */ #if !defined(SQLITE_OMIT_VIRTUALTABLE) && defined(WHERETRACE_ENABLED) static void TRACE_IDX_INPUTS(sqlite3_index_info *p){ int i; if( !sqlite3WhereTrace ) return; for(i=0; inConstraint; i++){ sqlite3DebugPrintf(" constraint[%d]: col=%d termid=%d op=%d usabled=%d\n", i, p->aConstraint[i].iColumn, p->aConstraint[i].iTermOffset, p->aConstraint[i].op, p->aConstraint[i].usable); } for(i=0; inOrderBy; i++){ sqlite3DebugPrintf(" orderby[%d]: col=%d desc=%d\n", i, p->aOrderBy[i].iColumn, p->aOrderBy[i].desc); } } static void TRACE_IDX_OUTPUTS(sqlite3_index_info *p){ int i; if( !sqlite3WhereTrace ) return; for(i=0; inConstraint; i++){ sqlite3DebugPrintf(" usage[%d]: argvIdx=%d omit=%d\n", i, p->aConstraintUsage[i].argvIndex, p->aConstraintUsage[i].omit); } sqlite3DebugPrintf(" idxNum=%d\n", p->idxNum); sqlite3DebugPrintf(" idxStr=%s\n", p->idxStr); sqlite3DebugPrintf(" orderByConsumed=%d\n", p->orderByConsumed); sqlite3DebugPrintf(" estimatedCost=%g\n", p->estimatedCost); sqlite3DebugPrintf(" estimatedRows=%lld\n", p->estimatedRows); } #else #define TRACE_IDX_INPUTS(A) #define TRACE_IDX_OUTPUTS(A) #endif #ifndef SQLITE_OMIT_AUTOMATIC_INDEX /* ** Return TRUE if the WHERE clause term pTerm is of a form where it ** could be used with an index to access pSrc, assuming an appropriate ** index existed. */ static int termCanDriveIndex( WhereTerm *pTerm, /* WHERE clause term to check */ struct SrcList_item *pSrc, /* Table we are trying to access */ Bitmask notReady /* Tables in outer loops of the join */ ){ char aff; if( pTerm->leftCursor!=pSrc->iCursor ) return 0; if( (pTerm->eOperator & WO_EQ)==0 ) return 0; if( (pTerm->prereqRight & notReady)!=0 ) return 0; if( pTerm->u.leftColumn<0 ) return 0; aff = pSrc->pTab->aCol[pTerm->u.leftColumn].affinity; if( !sqlite3IndexAffinityOk(pTerm->pExpr, aff) ) return 0; return 1; } #endif #ifndef SQLITE_OMIT_AUTOMATIC_INDEX /* ** Generate code to construct the Index object for an automatic index ** and to set up the WhereLevel object pLevel so that the code generator ** makes use of the automatic index. */ static void constructAutomaticIndex( Parse *pParse, /* The parsing context */ WhereClause *pWC, /* The WHERE clause */ struct SrcList_item *pSrc, /* The FROM clause term to get the next index */ Bitmask notReady, /* Mask of cursors that are not available */ WhereLevel *pLevel /* Write new index here */ ){ int nKeyCol; /* Number of columns in the constructed index */ WhereTerm *pTerm; /* A single term of the WHERE clause */ WhereTerm *pWCEnd; /* End of pWC->a[] */ Index *pIdx; /* Object describing the transient index */ Vdbe *v; /* Prepared statement under construction */ int addrInit; /* Address of the initialization bypass jump */ Table *pTable; /* The table being indexed */ int addrTop; /* Top of the index fill loop */ int regRecord; /* Register holding an index record */ int n; /* Column counter */ int i; /* Loop counter */ int mxBitCol; /* Maximum column in pSrc->colUsed */ CollSeq *pColl; /* Collating sequence to on a column */ WhereLoop *pLoop; /* The Loop object */ char *zNotUsed; /* Extra space on the end of pIdx */ Bitmask idxCols; /* Bitmap of columns used for indexing */ Bitmask extraCols; /* Bitmap of additional columns */ u8 sentWarning = 0; /* True if a warnning has been issued */ Expr *pPartial = 0; /* Partial Index Expression */ int iContinue = 0; /* Jump here to skip excluded rows */ /* Generate code to skip over the creation and initialization of the ** transient index on 2nd and subsequent iterations of the loop. */ v = pParse->pVdbe; assert( v!=0 ); addrInit = sqlite3CodeOnce(pParse); VdbeCoverage(v); /* Count the number of columns that will be added to the index ** and used to match WHERE clause constraints */ nKeyCol = 0; pTable = pSrc->pTab; pWCEnd = &pWC->a[pWC->nTerm]; pLoop = pLevel->pWLoop; idxCols = 0; for(pTerm=pWC->a; pTermprereq==0 && (pTerm->wtFlags & TERM_VIRTUAL)==0 && sqlite3ExprIsTableConstant(pTerm->pExpr, pSrc->iCursor) ){ pPartial = sqlite3ExprAnd(pParse->db, pPartial, sqlite3ExprDup(pParse->db, pTerm->pExpr, 0)); } if( termCanDriveIndex(pTerm, pSrc, notReady) ){ int iCol = pTerm->u.leftColumn; Bitmask cMask = iCol>=BMS ? MASKBIT(BMS-1) : MASKBIT(iCol); testcase( iCol==BMS ); testcase( iCol==BMS-1 ); if( !sentWarning ){ sqlite3_log(SQLITE_WARNING_AUTOINDEX, "automatic index on %s(%s)", pTable->zName, pTable->aCol[iCol].zName); sentWarning = 1; } if( (idxCols & cMask)==0 ){ if( whereLoopResize(pParse->db, pLoop, nKeyCol+1) ){ goto end_auto_index_create; } pLoop->aLTerm[nKeyCol++] = pTerm; idxCols |= cMask; } } } assert( nKeyCol>0 ); pLoop->u.btree.nEq = pLoop->nLTerm = nKeyCol; pLoop->wsFlags = WHERE_COLUMN_EQ | WHERE_IDX_ONLY | WHERE_INDEXED | WHERE_AUTO_INDEX; /* Count the number of additional columns needed to create a ** covering index. A "covering index" is an index that contains all ** columns that are needed by the query. With a covering index, the ** original table never needs to be accessed. Automatic indices must ** be a covering index because the index will not be updated if the ** original table changes and the index and table cannot both be used ** if they go out of sync. */ extraCols = pSrc->colUsed & (~idxCols | MASKBIT(BMS-1)); mxBitCol = MIN(BMS-1,pTable->nCol); testcase( pTable->nCol==BMS-1 ); testcase( pTable->nCol==BMS-2 ); for(i=0; icolUsed & MASKBIT(BMS-1) ){ nKeyCol += pTable->nCol - BMS + 1; } /* Construct the Index object to describe this index */ pIdx = sqlite3AllocateIndexObject(pParse->db, nKeyCol+1, 0, &zNotUsed); if( pIdx==0 ) goto end_auto_index_create; pLoop->u.btree.pIndex = pIdx; pIdx->zName = "auto-index"; pIdx->pTable = pTable; n = 0; idxCols = 0; for(pTerm=pWC->a; pTermu.leftColumn; Bitmask cMask = iCol>=BMS ? MASKBIT(BMS-1) : MASKBIT(iCol); testcase( iCol==BMS-1 ); testcase( iCol==BMS ); if( (idxCols & cMask)==0 ){ Expr *pX = pTerm->pExpr; idxCols |= cMask; pIdx->aiColumn[n] = pTerm->u.leftColumn; pColl = sqlite3BinaryCompareCollSeq(pParse, pX->pLeft, pX->pRight); pIdx->azColl[n] = ALWAYS(pColl) ? pColl->zName : "BINARY"; n++; } } } assert( (u32)n==pLoop->u.btree.nEq ); /* Add additional columns needed to make the automatic index into ** a covering index */ for(i=0; iaiColumn[n] = i; pIdx->azColl[n] = "BINARY"; n++; } } if( pSrc->colUsed & MASKBIT(BMS-1) ){ for(i=BMS-1; inCol; i++){ pIdx->aiColumn[n] = i; pIdx->azColl[n] = "BINARY"; n++; } } assert( n==nKeyCol ); pIdx->aiColumn[n] = -1; pIdx->azColl[n] = "BINARY"; /* Create the automatic index */ assert( pLevel->iIdxCur>=0 ); pLevel->iIdxCur = pParse->nTab++; sqlite3VdbeAddOp2(v, OP_OpenAutoindex, pLevel->iIdxCur, nKeyCol+1); sqlite3VdbeSetP4KeyInfo(pParse, pIdx); VdbeComment((v, "for %s", pTable->zName)); /* Fill the automatic index with content */ sqlite3ExprCachePush(pParse); addrTop = sqlite3VdbeAddOp1(v, OP_Rewind, pLevel->iTabCur); VdbeCoverage(v); if( pPartial ){ iContinue = sqlite3VdbeMakeLabel(v); sqlite3ExprIfFalse(pParse, pPartial, iContinue, SQLITE_JUMPIFNULL); pLoop->wsFlags |= WHERE_PARTIALIDX; } regRecord = sqlite3GetTempReg(pParse); sqlite3GenerateIndexKey(pParse, pIdx, pLevel->iTabCur, regRecord, 0, 0, 0, 0); sqlite3VdbeAddOp2(v, OP_IdxInsert, pLevel->iIdxCur, regRecord); sqlite3VdbeChangeP5(v, OPFLAG_USESEEKRESULT); if( pPartial ) sqlite3VdbeResolveLabel(v, iContinue); sqlite3VdbeAddOp2(v, OP_Next, pLevel->iTabCur, addrTop+1); VdbeCoverage(v); sqlite3VdbeChangeP5(v, SQLITE_STMTSTATUS_AUTOINDEX); sqlite3VdbeJumpHere(v, addrTop); sqlite3ReleaseTempReg(pParse, regRecord); sqlite3ExprCachePop(pParse); /* Jump here when skipping the initialization */ sqlite3VdbeJumpHere(v, addrInit); end_auto_index_create: sqlite3ExprDelete(pParse->db, pPartial); } #endif /* SQLITE_OMIT_AUTOMATIC_INDEX */ #ifndef SQLITE_OMIT_VIRTUALTABLE /* ** Allocate and populate an sqlite3_index_info structure. It is the ** responsibility of the caller to eventually release the structure ** by passing the pointer returned by this function to sqlite3_free(). */ static sqlite3_index_info *allocateIndexInfo( Parse *pParse, WhereClause *pWC, struct SrcList_item *pSrc, ExprList *pOrderBy ){ int i, j; int nTerm; struct sqlite3_index_constraint *pIdxCons; struct sqlite3_index_orderby *pIdxOrderBy; struct sqlite3_index_constraint_usage *pUsage; WhereTerm *pTerm; int nOrderBy; sqlite3_index_info *pIdxInfo; /* Count the number of possible WHERE clause constraints referring ** to this virtual table */ for(i=nTerm=0, pTerm=pWC->a; inTerm; i++, pTerm++){ if( pTerm->leftCursor != pSrc->iCursor ) continue; assert( IsPowerOfTwo(pTerm->eOperator & ~WO_EQUIV) ); testcase( pTerm->eOperator & WO_IN ); testcase( pTerm->eOperator & WO_ISNULL ); testcase( pTerm->eOperator & WO_ALL ); if( (pTerm->eOperator & ~(WO_ISNULL|WO_EQUIV))==0 ) continue; if( pTerm->wtFlags & TERM_VNULL ) continue; nTerm++; } /* If the ORDER BY clause contains only columns in the current ** virtual table then allocate space for the aOrderBy part of ** the sqlite3_index_info structure. */ nOrderBy = 0; if( pOrderBy ){ int n = pOrderBy->nExpr; for(i=0; ia[i].pExpr; if( pExpr->op!=TK_COLUMN || pExpr->iTable!=pSrc->iCursor ) break; } if( i==n){ nOrderBy = n; } } /* Allocate the sqlite3_index_info structure */ pIdxInfo = sqlite3DbMallocZero(pParse->db, sizeof(*pIdxInfo) + (sizeof(*pIdxCons) + sizeof(*pUsage))*nTerm + sizeof(*pIdxOrderBy)*nOrderBy ); if( pIdxInfo==0 ){ sqlite3ErrorMsg(pParse, "out of memory"); return 0; } /* Initialize the structure. The sqlite3_index_info structure contains ** many fields that are declared "const" to prevent xBestIndex from ** changing them. We have to do some funky casting in order to ** initialize those fields. */ pIdxCons = (struct sqlite3_index_constraint*)&pIdxInfo[1]; pIdxOrderBy = (struct sqlite3_index_orderby*)&pIdxCons[nTerm]; pUsage = (struct sqlite3_index_constraint_usage*)&pIdxOrderBy[nOrderBy]; *(int*)&pIdxInfo->nConstraint = nTerm; *(int*)&pIdxInfo->nOrderBy = nOrderBy; *(struct sqlite3_index_constraint**)&pIdxInfo->aConstraint = pIdxCons; *(struct sqlite3_index_orderby**)&pIdxInfo->aOrderBy = pIdxOrderBy; *(struct sqlite3_index_constraint_usage**)&pIdxInfo->aConstraintUsage = pUsage; for(i=j=0, pTerm=pWC->a; inTerm; i++, pTerm++){ u8 op; if( pTerm->leftCursor != pSrc->iCursor ) continue; assert( IsPowerOfTwo(pTerm->eOperator & ~WO_EQUIV) ); testcase( pTerm->eOperator & WO_IN ); testcase( pTerm->eOperator & WO_ISNULL ); testcase( pTerm->eOperator & WO_ALL ); if( (pTerm->eOperator & ~(WO_ISNULL|WO_EQUIV))==0 ) continue; if( pTerm->wtFlags & TERM_VNULL ) continue; pIdxCons[j].iColumn = pTerm->u.leftColumn; pIdxCons[j].iTermOffset = i; op = (u8)pTerm->eOperator & WO_ALL; if( op==WO_IN ) op = WO_EQ; pIdxCons[j].op = op; /* The direct assignment in the previous line is possible only because ** the WO_ and SQLITE_INDEX_CONSTRAINT_ codes are identical. The ** following asserts verify this fact. */ assert( WO_EQ==SQLITE_INDEX_CONSTRAINT_EQ ); assert( WO_LT==SQLITE_INDEX_CONSTRAINT_LT ); assert( WO_LE==SQLITE_INDEX_CONSTRAINT_LE ); assert( WO_GT==SQLITE_INDEX_CONSTRAINT_GT ); assert( WO_GE==SQLITE_INDEX_CONSTRAINT_GE ); assert( WO_MATCH==SQLITE_INDEX_CONSTRAINT_MATCH ); assert( pTerm->eOperator & (WO_IN|WO_EQ|WO_LT|WO_LE|WO_GT|WO_GE|WO_MATCH) ); j++; } for(i=0; ia[i].pExpr; pIdxOrderBy[i].iColumn = pExpr->iColumn; pIdxOrderBy[i].desc = pOrderBy->a[i].sortOrder; } return pIdxInfo; } /* ** The table object reference passed as the second argument to this function ** must represent a virtual table. This function invokes the xBestIndex() ** method of the virtual table with the sqlite3_index_info object that ** comes in as the 3rd argument to this function. ** ** If an error occurs, pParse is populated with an error message and a ** non-zero value is returned. Otherwise, 0 is returned and the output ** part of the sqlite3_index_info structure is left populated. ** ** Whether or not an error is returned, it is the responsibility of the ** caller to eventually free p->idxStr if p->needToFreeIdxStr indicates ** that this is required. */ static int vtabBestIndex(Parse *pParse, Table *pTab, sqlite3_index_info *p){ sqlite3_vtab *pVtab = sqlite3GetVTable(pParse->db, pTab)->pVtab; int i; int rc; TRACE_IDX_INPUTS(p); rc = pVtab->pModule->xBestIndex(pVtab, p); TRACE_IDX_OUTPUTS(p); if( rc!=SQLITE_OK ){ if( rc==SQLITE_NOMEM ){ pParse->db->mallocFailed = 1; }else if( !pVtab->zErrMsg ){ sqlite3ErrorMsg(pParse, "%s", sqlite3ErrStr(rc)); }else{ sqlite3ErrorMsg(pParse, "%s", pVtab->zErrMsg); } } sqlite3_free(pVtab->zErrMsg); pVtab->zErrMsg = 0; for(i=0; inConstraint; i++){ if( !p->aConstraint[i].usable && p->aConstraintUsage[i].argvIndex>0 ){ sqlite3ErrorMsg(pParse, "table %s: xBestIndex returned an invalid plan", pTab->zName); } } return pParse->nErr; } #endif /* !defined(SQLITE_OMIT_VIRTUALTABLE) */ #ifdef SQLITE_ENABLE_STAT3_OR_STAT4 /* ** Estimate the location of a particular key among all keys in an ** index. Store the results in aStat as follows: ** ** aStat[0] Est. number of rows less than pVal ** aStat[1] Est. number of rows equal to pVal ** ** Return the index of the sample that is the smallest sample that ** is greater than or equal to pRec. */ static int whereKeyStats( Parse *pParse, /* Database connection */ Index *pIdx, /* Index to consider domain of */ UnpackedRecord *pRec, /* Vector of values to consider */ int roundUp, /* Round up if true. Round down if false */ tRowcnt *aStat /* OUT: stats written here */ ){ IndexSample *aSample = pIdx->aSample; int iCol; /* Index of required stats in anEq[] etc. */ int iMin = 0; /* Smallest sample not yet tested */ int i = pIdx->nSample; /* Smallest sample larger than or equal to pRec */ int iTest; /* Next sample to test */ int res; /* Result of comparison operation */ #ifndef SQLITE_DEBUG UNUSED_PARAMETER( pParse ); #endif assert( pRec!=0 ); iCol = pRec->nField - 1; assert( pIdx->nSample>0 ); assert( pRec->nField>0 && iColnSampleCol ); do{ iTest = (iMin+i)/2; res = sqlite3VdbeRecordCompare(aSample[iTest].n, aSample[iTest].p, pRec); if( res<0 ){ iMin = iTest+1; }else{ i = iTest; } }while( res && iMinnSample ); assert( 0==sqlite3VdbeRecordCompare(aSample[i].n, aSample[i].p, pRec) || pParse->db->mallocFailed ); }else{ /* Otherwise, pRec must be smaller than sample $i and larger than ** sample ($i-1). */ assert( i==pIdx->nSample || sqlite3VdbeRecordCompare(aSample[i].n, aSample[i].p, pRec)>0 || pParse->db->mallocFailed ); assert( i==0 || sqlite3VdbeRecordCompare(aSample[i-1].n, aSample[i-1].p, pRec)<0 || pParse->db->mallocFailed ); } #endif /* ifdef SQLITE_DEBUG */ /* At this point, aSample[i] is the first sample that is greater than ** or equal to pVal. Or if i==pIdx->nSample, then all samples are less ** than pVal. If aSample[i]==pVal, then res==0. */ if( res==0 ){ aStat[0] = aSample[i].anLt[iCol]; aStat[1] = aSample[i].anEq[iCol]; }else{ tRowcnt iLower, iUpper, iGap; if( i==0 ){ iLower = 0; iUpper = aSample[0].anLt[iCol]; }else{ i64 nRow0 = sqlite3LogEstToInt(pIdx->aiRowLogEst[0]); iUpper = i>=pIdx->nSample ? nRow0 : aSample[i].anLt[iCol]; iLower = aSample[i-1].anEq[iCol] + aSample[i-1].anLt[iCol]; } aStat[1] = pIdx->aAvgEq[iCol]; if( iLower>=iUpper ){ iGap = 0; }else{ iGap = iUpper - iLower; } if( roundUp ){ iGap = (iGap*2)/3; }else{ iGap = iGap/3; } aStat[0] = iLower + iGap; } return i; } #endif /* SQLITE_ENABLE_STAT3_OR_STAT4 */ /* ** If it is not NULL, pTerm is a term that provides an upper or lower ** bound on a range scan. Without considering pTerm, it is estimated ** that the scan will visit nNew rows. This function returns the number ** estimated to be visited after taking pTerm into account. ** ** If the user explicitly specified a likelihood() value for this term, ** then the return value is the likelihood multiplied by the number of ** input rows. Otherwise, this function assumes that an "IS NOT NULL" term ** has a likelihood of 0.50, and any other term a likelihood of 0.25. */ static LogEst whereRangeAdjust(WhereTerm *pTerm, LogEst nNew){ LogEst nRet = nNew; if( pTerm ){ if( pTerm->truthProb<=0 ){ nRet += pTerm->truthProb; }else if( (pTerm->wtFlags & TERM_VNULL)==0 ){ nRet -= 20; assert( 20==sqlite3LogEst(4) ); } } return nRet; } #ifdef SQLITE_ENABLE_STAT3_OR_STAT4 /* ** This function is called to estimate the number of rows visited by a ** range-scan on a skip-scan index. For example: ** ** CREATE INDEX i1 ON t1(a, b, c); ** SELECT * FROM t1 WHERE a=? AND c BETWEEN ? AND ?; ** ** Value pLoop->nOut is currently set to the estimated number of rows ** visited for scanning (a=? AND b=?). This function reduces that estimate ** by some factor to account for the (c BETWEEN ? AND ?) expression based ** on the stat4 data for the index. this scan will be peformed multiple ** times (once for each (a,b) combination that matches a=?) is dealt with ** by the caller. ** ** It does this by scanning through all stat4 samples, comparing values ** extracted from pLower and pUpper with the corresponding column in each ** sample. If L and U are the number of samples found to be less than or ** equal to the values extracted from pLower and pUpper respectively, and ** N is the total number of samples, the pLoop->nOut value is adjusted ** as follows: ** ** nOut = nOut * ( min(U - L, 1) / N ) ** ** If pLower is NULL, or a value cannot be extracted from the term, L is ** set to zero. If pUpper is NULL, or a value cannot be extracted from it, ** U is set to N. ** ** Normally, this function sets *pbDone to 1 before returning. However, ** if no value can be extracted from either pLower or pUpper (and so the ** estimate of the number of rows delivered remains unchanged), *pbDone ** is left as is. ** ** If an error occurs, an SQLite error code is returned. Otherwise, ** SQLITE_OK. */ static int whereRangeSkipScanEst( Parse *pParse, /* Parsing & code generating context */ WhereTerm *pLower, /* Lower bound on the range. ex: "x>123" Might be NULL */ WhereTerm *pUpper, /* Upper bound on the range. ex: "x<455" Might be NULL */ WhereLoop *pLoop, /* Update the .nOut value of this loop */ int *pbDone /* Set to true if at least one expr. value extracted */ ){ Index *p = pLoop->u.btree.pIndex; int nEq = pLoop->u.btree.nEq; sqlite3 *db = pParse->db; int nLower = -1; int nUpper = p->nSample+1; int rc = SQLITE_OK; int iCol = p->aiColumn[nEq]; u8 aff = iCol>=0 ? p->pTable->aCol[iCol].affinity : SQLITE_AFF_INTEGER; CollSeq *pColl; sqlite3_value *p1 = 0; /* Value extracted from pLower */ sqlite3_value *p2 = 0; /* Value extracted from pUpper */ sqlite3_value *pVal = 0; /* Value extracted from record */ pColl = sqlite3LocateCollSeq(pParse, p->azColl[nEq]); if( pLower ){ rc = sqlite3Stat4ValueFromExpr(pParse, pLower->pExpr->pRight, aff, &p1); nLower = 0; } if( pUpper && rc==SQLITE_OK ){ rc = sqlite3Stat4ValueFromExpr(pParse, pUpper->pExpr->pRight, aff, &p2); nUpper = p2 ? 0 : p->nSample; } if( p1 || p2 ){ int i; int nDiff; for(i=0; rc==SQLITE_OK && inSample; i++){ rc = sqlite3Stat4Column(db, p->aSample[i].p, p->aSample[i].n, nEq, &pVal); if( rc==SQLITE_OK && p1 ){ int res = sqlite3MemCompare(p1, pVal, pColl); if( res>=0 ) nLower++; } if( rc==SQLITE_OK && p2 ){ int res = sqlite3MemCompare(p2, pVal, pColl); if( res>=0 ) nUpper++; } } nDiff = (nUpper - nLower); if( nDiff<=0 ) nDiff = 1; /* If there is both an upper and lower bound specified, and the ** comparisons indicate that they are close together, use the fallback ** method (assume that the scan visits 1/64 of the rows) for estimating ** the number of rows visited. Otherwise, estimate the number of rows ** using the method described in the header comment for this function. */ if( nDiff!=1 || pUpper==0 || pLower==0 ){ int nAdjust = (sqlite3LogEst(p->nSample) - sqlite3LogEst(nDiff)); pLoop->nOut -= nAdjust; *pbDone = 1; WHERETRACE(0x10, ("range skip-scan regions: %u..%u adjust=%d est=%d\n", nLower, nUpper, nAdjust*-1, pLoop->nOut)); } }else{ assert( *pbDone==0 ); } sqlite3ValueFree(p1); sqlite3ValueFree(p2); sqlite3ValueFree(pVal); return rc; } #endif /* SQLITE_ENABLE_STAT3_OR_STAT4 */ /* ** This function is used to estimate the number of rows that will be visited ** by scanning an index for a range of values. The range may have an upper ** bound, a lower bound, or both. The WHERE clause terms that set the upper ** and lower bounds are represented by pLower and pUpper respectively. For ** example, assuming that index p is on t1(a): ** ** ... FROM t1 WHERE a > ? AND a < ? ... ** |_____| |_____| ** | | ** pLower pUpper ** ** If either of the upper or lower bound is not present, then NULL is passed in ** place of the corresponding WhereTerm. ** ** The value in (pBuilder->pNew->u.btree.nEq) is the number of the index ** column subject to the range constraint. Or, equivalently, the number of ** equality constraints optimized by the proposed index scan. For example, ** assuming index p is on t1(a, b), and the SQL query is: ** ** ... FROM t1 WHERE a = ? AND b > ? AND b < ? ... ** ** then nEq is set to 1 (as the range restricted column, b, is the second ** left-most column of the index). Or, if the query is: ** ** ... FROM t1 WHERE a > ? AND a < ? ... ** ** then nEq is set to 0. ** ** When this function is called, *pnOut is set to the sqlite3LogEst() of the ** number of rows that the index scan is expected to visit without ** considering the range constraints. If nEq is 0, then *pnOut is the number of ** rows in the index. Assuming no error occurs, *pnOut is adjusted (reduced) ** to account for the range constraints pLower and pUpper. ** ** In the absence of sqlite_stat4 ANALYZE data, or if such data cannot be ** used, a single range inequality reduces the search space by a factor of 4. ** and a pair of constraints (x>? AND x123" Might be NULL */ WhereTerm *pUpper, /* Upper bound on the range. ex: "x<455" Might be NULL */ WhereLoop *pLoop /* Modify the .nOut and maybe .rRun fields */ ){ int rc = SQLITE_OK; int nOut = pLoop->nOut; LogEst nNew; #ifdef SQLITE_ENABLE_STAT3_OR_STAT4 Index *p = pLoop->u.btree.pIndex; int nEq = pLoop->u.btree.nEq; if( p->nSample>0 && nEqnSampleCol ){ if( nEq==pBuilder->nRecValid ){ UnpackedRecord *pRec = pBuilder->pRec; tRowcnt a[2]; u8 aff; /* Variable iLower will be set to the estimate of the number of rows in ** the index that are less than the lower bound of the range query. The ** lower bound being the concatenation of $P and $L, where $P is the ** key-prefix formed by the nEq values matched against the nEq left-most ** columns of the index, and $L is the value in pLower. ** ** Or, if pLower is NULL or $L cannot be extracted from it (because it ** is not a simple variable or literal value), the lower bound of the ** range is $P. Due to a quirk in the way whereKeyStats() works, even ** if $L is available, whereKeyStats() is called for both ($P) and ** ($P:$L) and the larger of the two returned values is used. ** ** Similarly, iUpper is to be set to the estimate of the number of rows ** less than the upper bound of the range query. Where the upper bound ** is either ($P) or ($P:$U). Again, even if $U is available, both values ** of iUpper are requested of whereKeyStats() and the smaller used. ** ** The number of rows between the two bounds is then just iUpper-iLower. */ tRowcnt iLower; /* Rows less than the lower bound */ tRowcnt iUpper; /* Rows less than the upper bound */ int iLwrIdx = -2; /* aSample[] for the lower bound */ int iUprIdx = -1; /* aSample[] for the upper bound */ if( pRec ){ testcase( pRec->nField!=pBuilder->nRecValid ); pRec->nField = pBuilder->nRecValid; } if( nEq==p->nKeyCol ){ aff = SQLITE_AFF_INTEGER; }else{ aff = p->pTable->aCol[p->aiColumn[nEq]].affinity; } /* Determine iLower and iUpper using ($P) only. */ if( nEq==0 ){ iLower = 0; iUpper = p->nRowEst0; }else{ /* Note: this call could be optimized away - since the same values must ** have been requested when testing key $P in whereEqualScanEst(). */ whereKeyStats(pParse, p, pRec, 0, a); iLower = a[0]; iUpper = a[0] + a[1]; } assert( pLower==0 || (pLower->eOperator & (WO_GT|WO_GE))!=0 ); assert( pUpper==0 || (pUpper->eOperator & (WO_LT|WO_LE))!=0 ); assert( p->aSortOrder!=0 ); if( p->aSortOrder[nEq] ){ /* The roles of pLower and pUpper are swapped for a DESC index */ SWAP(WhereTerm*, pLower, pUpper); } /* If possible, improve on the iLower estimate using ($P:$L). */ if( pLower ){ int bOk; /* True if value is extracted from pExpr */ Expr *pExpr = pLower->pExpr->pRight; rc = sqlite3Stat4ProbeSetValue(pParse, p, &pRec, pExpr, aff, nEq, &bOk); if( rc==SQLITE_OK && bOk ){ tRowcnt iNew; iLwrIdx = whereKeyStats(pParse, p, pRec, 0, a); iNew = a[0] + ((pLower->eOperator & (WO_GT|WO_LE)) ? a[1] : 0); if( iNew>iLower ) iLower = iNew; nOut--; pLower = 0; } } /* If possible, improve on the iUpper estimate using ($P:$U). */ if( pUpper ){ int bOk; /* True if value is extracted from pExpr */ Expr *pExpr = pUpper->pExpr->pRight; rc = sqlite3Stat4ProbeSetValue(pParse, p, &pRec, pExpr, aff, nEq, &bOk); if( rc==SQLITE_OK && bOk ){ tRowcnt iNew; iUprIdx = whereKeyStats(pParse, p, pRec, 1, a); iNew = a[0] + ((pUpper->eOperator & (WO_GT|WO_LE)) ? a[1] : 0); if( iNewpRec = pRec; if( rc==SQLITE_OK ){ if( iUpper>iLower ){ nNew = sqlite3LogEst(iUpper - iLower); /* TUNING: If both iUpper and iLower are derived from the same ** sample, then assume they are 4x more selective. This brings ** the estimated selectivity more in line with what it would be ** if estimated without the use of STAT3/4 tables. */ if( iLwrIdx==iUprIdx ) nNew -= 20; assert( 20==sqlite3LogEst(4) ); }else{ nNew = 10; assert( 10==sqlite3LogEst(2) ); } if( nNewwtFlags & TERM_VNULL)==0 ); nNew = whereRangeAdjust(pLower, nOut); nNew = whereRangeAdjust(pUpper, nNew); /* TUNING: If there is both an upper and lower limit and neither limit ** has an application-defined likelihood(), assume the range is ** reduced by an additional 75%. This means that, by default, an open-ended ** range query (e.g. col > ?) is assumed to match 1/4 of the rows in the ** index. While a closed range (e.g. col BETWEEN ? AND ?) is estimated to ** match 1/64 of the index. */ if( pLower && pLower->truthProb>0 && pUpper && pUpper->truthProb>0 ){ nNew -= 20; } nOut -= (pLower!=0) + (pUpper!=0); if( nNew<10 ) nNew = 10; if( nNewnOut>nOut ){ WHERETRACE(0x10,("Range scan lowers nOut from %d to %d\n", pLoop->nOut, nOut)); } #endif pLoop->nOut = (LogEst)nOut; return rc; } #ifdef SQLITE_ENABLE_STAT3_OR_STAT4 /* ** Estimate the number of rows that will be returned based on ** an equality constraint x=VALUE and where that VALUE occurs in ** the histogram data. This only works when x is the left-most ** column of an index and sqlite_stat3 histogram data is available ** for that index. When pExpr==NULL that means the constraint is ** "x IS NULL" instead of "x=VALUE". ** ** Write the estimated row count into *pnRow and return SQLITE_OK. ** If unable to make an estimate, leave *pnRow unchanged and return ** non-zero. ** ** This routine can fail if it is unable to load a collating sequence ** required for string comparison, or if unable to allocate memory ** for a UTF conversion required for comparison. The error is stored ** in the pParse structure. */ static int whereEqualScanEst( Parse *pParse, /* Parsing & code generating context */ WhereLoopBuilder *pBuilder, Expr *pExpr, /* Expression for VALUE in the x=VALUE constraint */ tRowcnt *pnRow /* Write the revised row estimate here */ ){ Index *p = pBuilder->pNew->u.btree.pIndex; int nEq = pBuilder->pNew->u.btree.nEq; UnpackedRecord *pRec = pBuilder->pRec; u8 aff; /* Column affinity */ int rc; /* Subfunction return code */ tRowcnt a[2]; /* Statistics */ int bOk; assert( nEq>=1 ); assert( nEq<=p->nColumn ); assert( p->aSample!=0 ); assert( p->nSample>0 ); assert( pBuilder->nRecValidnRecValid<(nEq-1) ){ return SQLITE_NOTFOUND; } /* This is an optimization only. The call to sqlite3Stat4ProbeSetValue() ** below would return the same value. */ if( nEq>=p->nColumn ){ *pnRow = 1; return SQLITE_OK; } aff = p->pTable->aCol[p->aiColumn[nEq-1]].affinity; rc = sqlite3Stat4ProbeSetValue(pParse, p, &pRec, pExpr, aff, nEq-1, &bOk); pBuilder->pRec = pRec; if( rc!=SQLITE_OK ) return rc; if( bOk==0 ) return SQLITE_NOTFOUND; pBuilder->nRecValid = nEq; whereKeyStats(pParse, p, pRec, 0, a); WHERETRACE(0x10,("equality scan regions: %d\n", (int)a[1])); *pnRow = a[1]; return rc; } #endif /* SQLITE_ENABLE_STAT3_OR_STAT4 */ #ifdef SQLITE_ENABLE_STAT3_OR_STAT4 /* ** Estimate the number of rows that will be returned based on ** an IN constraint where the right-hand side of the IN operator ** is a list of values. Example: ** ** WHERE x IN (1,2,3,4) ** ** Write the estimated row count into *pnRow and return SQLITE_OK. ** If unable to make an estimate, leave *pnRow unchanged and return ** non-zero. ** ** This routine can fail if it is unable to load a collating sequence ** required for string comparison, or if unable to allocate memory ** for a UTF conversion required for comparison. The error is stored ** in the pParse structure. */ static int whereInScanEst( Parse *pParse, /* Parsing & code generating context */ WhereLoopBuilder *pBuilder, ExprList *pList, /* The value list on the RHS of "x IN (v1,v2,v3,...)" */ tRowcnt *pnRow /* Write the revised row estimate here */ ){ Index *p = pBuilder->pNew->u.btree.pIndex; i64 nRow0 = sqlite3LogEstToInt(p->aiRowLogEst[0]); int nRecValid = pBuilder->nRecValid; int rc = SQLITE_OK; /* Subfunction return code */ tRowcnt nEst; /* Number of rows for a single term */ tRowcnt nRowEst = 0; /* New estimate of the number of rows */ int i; /* Loop counter */ assert( p->aSample!=0 ); for(i=0; rc==SQLITE_OK && inExpr; i++){ nEst = nRow0; rc = whereEqualScanEst(pParse, pBuilder, pList->a[i].pExpr, &nEst); nRowEst += nEst; pBuilder->nRecValid = nRecValid; } if( rc==SQLITE_OK ){ if( nRowEst > nRow0 ) nRowEst = nRow0; *pnRow = nRowEst; WHERETRACE(0x10,("IN row estimate: est=%d\n", nRowEst)); } assert( pBuilder->nRecValid==nRecValid ); return rc; } #endif /* SQLITE_ENABLE_STAT3_OR_STAT4 */ /* ** Disable a term in the WHERE clause. Except, do not disable the term ** if it controls a LEFT OUTER JOIN and it did not originate in the ON ** or USING clause of that join. ** ** Consider the term t2.z='ok' in the following queries: ** ** (1) SELECT * FROM t1 LEFT JOIN t2 ON t1.a=t2.x WHERE t2.z='ok' ** (2) SELECT * FROM t1 LEFT JOIN t2 ON t1.a=t2.x AND t2.z='ok' ** (3) SELECT * FROM t1, t2 WHERE t1.a=t2.x AND t2.z='ok' ** ** The t2.z='ok' is disabled in the in (2) because it originates ** in the ON clause. The term is disabled in (3) because it is not part ** of a LEFT OUTER JOIN. In (1), the term is not disabled. ** ** Disabling a term causes that term to not be tested in the inner loop ** of the join. Disabling is an optimization. When terms are satisfied ** by indices, we disable them to prevent redundant tests in the inner ** loop. We would get the correct results if nothing were ever disabled, ** but joins might run a little slower. The trick is to disable as much ** as we can without disabling too much. If we disabled in (1), we'd get ** the wrong answer. See ticket #813. */ static void disableTerm(WhereLevel *pLevel, WhereTerm *pTerm){ if( pTerm && (pTerm->wtFlags & TERM_CODED)==0 && (pLevel->iLeftJoin==0 || ExprHasProperty(pTerm->pExpr, EP_FromJoin)) && (pLevel->notReady & pTerm->prereqAll)==0 ){ pTerm->wtFlags |= TERM_CODED; if( pTerm->iParent>=0 ){ WhereTerm *pOther = &pTerm->pWC->a[pTerm->iParent]; if( (--pOther->nChild)==0 ){ disableTerm(pLevel, pOther); } } } } /* ** Code an OP_Affinity opcode to apply the column affinity string zAff ** to the n registers starting at base. ** ** As an optimization, SQLITE_AFF_NONE entries (which are no-ops) at the ** beginning and end of zAff are ignored. If all entries in zAff are ** SQLITE_AFF_NONE, then no code gets generated. ** ** This routine makes its own copy of zAff so that the caller is free ** to modify zAff after this routine returns. */ static void codeApplyAffinity(Parse *pParse, int base, int n, char *zAff){ Vdbe *v = pParse->pVdbe; if( zAff==0 ){ assert( pParse->db->mallocFailed ); return; } assert( v!=0 ); /* Adjust base and n to skip over SQLITE_AFF_NONE entries at the beginning ** and end of the affinity string. */ while( n>0 && zAff[0]==SQLITE_AFF_NONE ){ n--; base++; zAff++; } while( n>1 && zAff[n-1]==SQLITE_AFF_NONE ){ n--; } /* Code the OP_Affinity opcode if there is anything left to do. */ if( n>0 ){ sqlite3VdbeAddOp2(v, OP_Affinity, base, n); sqlite3VdbeChangeP4(v, -1, zAff, n); sqlite3ExprCacheAffinityChange(pParse, base, n); } } /* ** Generate code for a single equality term of the WHERE clause. An equality ** term can be either X=expr or X IN (...). pTerm is the term to be ** coded. ** ** The current value for the constraint is left in register iReg. ** ** For a constraint of the form X=expr, the expression is evaluated and its ** result is left on the stack. For constraints of the form X IN (...) ** this routine sets up a loop that will iterate over all values of X. */ static int codeEqualityTerm( Parse *pParse, /* The parsing context */ WhereTerm *pTerm, /* The term of the WHERE clause to be coded */ WhereLevel *pLevel, /* The level of the FROM clause we are working on */ int iEq, /* Index of the equality term within this level */ int bRev, /* True for reverse-order IN operations */ int iTarget /* Attempt to leave results in this register */ ){ Expr *pX = pTerm->pExpr; Vdbe *v = pParse->pVdbe; int iReg; /* Register holding results */ assert( iTarget>0 ); if( pX->op==TK_EQ ){ iReg = sqlite3ExprCodeTarget(pParse, pX->pRight, iTarget); }else if( pX->op==TK_ISNULL ){ iReg = iTarget; sqlite3VdbeAddOp2(v, OP_Null, 0, iReg); #ifndef SQLITE_OMIT_SUBQUERY }else{ int eType; int iTab; struct InLoop *pIn; WhereLoop *pLoop = pLevel->pWLoop; if( (pLoop->wsFlags & WHERE_VIRTUALTABLE)==0 && pLoop->u.btree.pIndex!=0 && pLoop->u.btree.pIndex->aSortOrder[iEq] ){ testcase( iEq==0 ); testcase( bRev ); bRev = !bRev; } assert( pX->op==TK_IN ); iReg = iTarget; eType = sqlite3FindInIndex(pParse, pX, IN_INDEX_LOOP, 0); if( eType==IN_INDEX_INDEX_DESC ){ testcase( bRev ); bRev = !bRev; } iTab = pX->iTable; sqlite3VdbeAddOp2(v, bRev ? OP_Last : OP_Rewind, iTab, 0); VdbeCoverageIf(v, bRev); VdbeCoverageIf(v, !bRev); assert( (pLoop->wsFlags & WHERE_MULTI_OR)==0 ); pLoop->wsFlags |= WHERE_IN_ABLE; if( pLevel->u.in.nIn==0 ){ pLevel->addrNxt = sqlite3VdbeMakeLabel(v); } pLevel->u.in.nIn++; pLevel->u.in.aInLoop = sqlite3DbReallocOrFree(pParse->db, pLevel->u.in.aInLoop, sizeof(pLevel->u.in.aInLoop[0])*pLevel->u.in.nIn); pIn = pLevel->u.in.aInLoop; if( pIn ){ pIn += pLevel->u.in.nIn - 1; pIn->iCur = iTab; if( eType==IN_INDEX_ROWID ){ pIn->addrInTop = sqlite3VdbeAddOp2(v, OP_Rowid, iTab, iReg); }else{ pIn->addrInTop = sqlite3VdbeAddOp3(v, OP_Column, iTab, 0, iReg); } pIn->eEndLoopOp = bRev ? OP_PrevIfOpen : OP_NextIfOpen; sqlite3VdbeAddOp1(v, OP_IsNull, iReg); VdbeCoverage(v); }else{ pLevel->u.in.nIn = 0; } #endif } disableTerm(pLevel, pTerm); return iReg; } /* ** Generate code that will evaluate all == and IN constraints for an ** index scan. ** ** For example, consider table t1(a,b,c,d,e,f) with index i1(a,b,c). ** Suppose the WHERE clause is this: a==5 AND b IN (1,2,3) AND c>5 AND c<10 ** The index has as many as three equality constraints, but in this ** example, the third "c" value is an inequality. So only two ** constraints are coded. This routine will generate code to evaluate ** a==5 and b IN (1,2,3). The current values for a and b will be stored ** in consecutive registers and the index of the first register is returned. ** ** In the example above nEq==2. But this subroutine works for any value ** of nEq including 0. If nEq==0, this routine is nearly a no-op. ** The only thing it does is allocate the pLevel->iMem memory cell and ** compute the affinity string. ** ** The nExtraReg parameter is 0 or 1. It is 0 if all WHERE clause constraints ** are == or IN and are covered by the nEq. nExtraReg is 1 if there is ** an inequality constraint (such as the "c>=5 AND c<10" in the example) that ** occurs after the nEq quality constraints. ** ** This routine allocates a range of nEq+nExtraReg memory cells and returns ** the index of the first memory cell in that range. The code that ** calls this routine will use that memory range to store keys for ** start and termination conditions of the loop. ** key value of the loop. If one or more IN operators appear, then ** this routine allocates an additional nEq memory cells for internal ** use. ** ** Before returning, *pzAff is set to point to a buffer containing a ** copy of the column affinity string of the index allocated using ** sqlite3DbMalloc(). Except, entries in the copy of the string associated ** with equality constraints that use NONE affinity are set to ** SQLITE_AFF_NONE. This is to deal with SQL such as the following: ** ** CREATE TABLE t1(a TEXT PRIMARY KEY, b); ** SELECT ... FROM t1 AS t2, t1 WHERE t1.a = t2.b; ** ** In the example above, the index on t1(a) has TEXT affinity. But since ** the right hand side of the equality constraint (t2.b) has NONE affinity, ** no conversion should be attempted before using a t2.b value as part of ** a key to search the index. Hence the first byte in the returned affinity ** string in this example would be set to SQLITE_AFF_NONE. */ static int codeAllEqualityTerms( Parse *pParse, /* Parsing context */ WhereLevel *pLevel, /* Which nested loop of the FROM we are coding */ int bRev, /* Reverse the order of IN operators */ int nExtraReg, /* Number of extra registers to allocate */ char **pzAff /* OUT: Set to point to affinity string */ ){ u16 nEq; /* The number of == or IN constraints to code */ u16 nSkip; /* Number of left-most columns to skip */ Vdbe *v = pParse->pVdbe; /* The vm under construction */ Index *pIdx; /* The index being used for this loop */ WhereTerm *pTerm; /* A single constraint term */ WhereLoop *pLoop; /* The WhereLoop object */ int j; /* Loop counter */ int regBase; /* Base register */ int nReg; /* Number of registers to allocate */ char *zAff; /* Affinity string to return */ /* This module is only called on query plans that use an index. */ pLoop = pLevel->pWLoop; assert( (pLoop->wsFlags & WHERE_VIRTUALTABLE)==0 ); nEq = pLoop->u.btree.nEq; nSkip = pLoop->nSkip; pIdx = pLoop->u.btree.pIndex; assert( pIdx!=0 ); /* Figure out how many memory cells we will need then allocate them. */ regBase = pParse->nMem + 1; nReg = pLoop->u.btree.nEq + nExtraReg; pParse->nMem += nReg; zAff = sqlite3DbStrDup(pParse->db, sqlite3IndexAffinityStr(v, pIdx)); if( !zAff ){ pParse->db->mallocFailed = 1; } if( nSkip ){ int iIdxCur = pLevel->iIdxCur; sqlite3VdbeAddOp1(v, (bRev?OP_Last:OP_Rewind), iIdxCur); VdbeCoverageIf(v, bRev==0); VdbeCoverageIf(v, bRev!=0); VdbeComment((v, "begin skip-scan on %s", pIdx->zName)); j = sqlite3VdbeAddOp0(v, OP_Goto); pLevel->addrSkip = sqlite3VdbeAddOp4Int(v, (bRev?OP_SeekLT:OP_SeekGT), iIdxCur, 0, regBase, nSkip); VdbeCoverageIf(v, bRev==0); VdbeCoverageIf(v, bRev!=0); sqlite3VdbeJumpHere(v, j); for(j=0; jaiColumn[j]>=0 ); VdbeComment((v, "%s", pIdx->pTable->aCol[pIdx->aiColumn[j]].zName)); } } /* Evaluate the equality constraints */ assert( zAff==0 || (int)strlen(zAff)>=nEq ); for(j=nSkip; jaLTerm[j]; assert( pTerm!=0 ); /* The following testcase is true for indices with redundant columns. ** Ex: CREATE INDEX i1 ON t1(a,b,a); SELECT * FROM t1 WHERE a=0 AND b=0; */ testcase( (pTerm->wtFlags & TERM_CODED)!=0 ); testcase( pTerm->wtFlags & TERM_VIRTUAL ); r1 = codeEqualityTerm(pParse, pTerm, pLevel, j, bRev, regBase+j); if( r1!=regBase+j ){ if( nReg==1 ){ sqlite3ReleaseTempReg(pParse, regBase); regBase = r1; }else{ sqlite3VdbeAddOp2(v, OP_SCopy, r1, regBase+j); } } testcase( pTerm->eOperator & WO_ISNULL ); testcase( pTerm->eOperator & WO_IN ); if( (pTerm->eOperator & (WO_ISNULL|WO_IN))==0 ){ Expr *pRight = pTerm->pExpr->pRight; if( sqlite3ExprCanBeNull(pRight) ){ sqlite3VdbeAddOp2(v, OP_IsNull, regBase+j, pLevel->addrBrk); VdbeCoverage(v); } if( zAff ){ if( sqlite3CompareAffinity(pRight, zAff[j])==SQLITE_AFF_NONE ){ zAff[j] = SQLITE_AFF_NONE; } if( sqlite3ExprNeedsNoAffinityChange(pRight, zAff[j]) ){ zAff[j] = SQLITE_AFF_NONE; } } } } *pzAff = zAff; return regBase; } #ifndef SQLITE_OMIT_EXPLAIN /* ** This routine is a helper for explainIndexRange() below ** ** pStr holds the text of an expression that we are building up one term ** at a time. This routine adds a new term to the end of the expression. ** Terms are separated by AND so add the "AND" text for second and subsequent ** terms only. */ static void explainAppendTerm( StrAccum *pStr, /* The text expression being built */ int iTerm, /* Index of this term. First is zero */ const char *zColumn, /* Name of the column */ const char *zOp /* Name of the operator */ ){ if( iTerm ) sqlite3StrAccumAppend(pStr, " AND ", 5); sqlite3StrAccumAppendAll(pStr, zColumn); sqlite3StrAccumAppend(pStr, zOp, 1); sqlite3StrAccumAppend(pStr, "?", 1); } /* ** Argument pLevel describes a strategy for scanning table pTab. This ** function appends text to pStr that describes the subset of table ** rows scanned by the strategy in the form of an SQL expression. ** ** For example, if the query: ** ** SELECT * FROM t1 WHERE a=1 AND b>2; ** ** is run and there is an index on (a, b), then this function returns a ** string similar to: ** ** "a=? AND b>?" */ static void explainIndexRange(StrAccum *pStr, WhereLoop *pLoop, Table *pTab){ Index *pIndex = pLoop->u.btree.pIndex; u16 nEq = pLoop->u.btree.nEq; u16 nSkip = pLoop->nSkip; int i, j; Column *aCol = pTab->aCol; i16 *aiColumn = pIndex->aiColumn; if( nEq==0 && (pLoop->wsFlags&(WHERE_BTM_LIMIT|WHERE_TOP_LIMIT))==0 ) return; sqlite3StrAccumAppend(pStr, " (", 2); for(i=0; i=nSkip ){ explainAppendTerm(pStr, i, z, "="); }else{ if( i ) sqlite3StrAccumAppend(pStr, " AND ", 5); sqlite3XPrintf(pStr, 0, "ANY(%s)", z); } } j = i; if( pLoop->wsFlags&WHERE_BTM_LIMIT ){ char *z = aiColumn[j] < 0 ? "rowid" : aCol[aiColumn[j]].zName; explainAppendTerm(pStr, i++, z, ">"); } if( pLoop->wsFlags&WHERE_TOP_LIMIT ){ char *z = aiColumn[j] < 0 ? "rowid" : aCol[aiColumn[j]].zName; explainAppendTerm(pStr, i, z, "<"); } sqlite3StrAccumAppend(pStr, ")", 1); } /* ** This function is a no-op unless currently processing an EXPLAIN QUERY PLAN ** command, or if either SQLITE_DEBUG or SQLITE_ENABLE_STMT_SCANSTATUS was ** defined at compile-time. If it is not a no-op, a single OP_Explain opcode ** is added to the output to describe the table scan strategy in pLevel. ** ** If an OP_Explain opcode is added to the VM, its address is returned. ** Otherwise, if no OP_Explain is coded, zero is returned. */ static int explainOneScan( Parse *pParse, /* Parse context */ SrcList *pTabList, /* Table list this loop refers to */ WhereLevel *pLevel, /* Scan to write OP_Explain opcode for */ int iLevel, /* Value for "level" column of output */ int iFrom, /* Value for "from" column of output */ u16 wctrlFlags /* Flags passed to sqlite3WhereBegin() */ ){ int ret = 0; #if !defined(SQLITE_DEBUG) && !defined(SQLITE_ENABLE_STMT_SCANSTATUS) if( pParse->explain==2 ) #endif { struct SrcList_item *pItem = &pTabList->a[pLevel->iFrom]; Vdbe *v = pParse->pVdbe; /* VM being constructed */ sqlite3 *db = pParse->db; /* Database handle */ int iId = pParse->iSelectId; /* Select id (left-most output column) */ int isSearch; /* True for a SEARCH. False for SCAN. */ WhereLoop *pLoop; /* The controlling WhereLoop object */ u32 flags; /* Flags that describe this loop */ char *zMsg; /* Text to add to EQP output */ StrAccum str; /* EQP output string */ char zBuf[100]; /* Initial space for EQP output string */ pLoop = pLevel->pWLoop; flags = pLoop->wsFlags; if( (flags&WHERE_MULTI_OR) || (wctrlFlags&WHERE_ONETABLE_ONLY) ) return 0; isSearch = (flags&(WHERE_BTM_LIMIT|WHERE_TOP_LIMIT))!=0 || ((flags&WHERE_VIRTUALTABLE)==0 && (pLoop->u.btree.nEq>0)) || (wctrlFlags&(WHERE_ORDERBY_MIN|WHERE_ORDERBY_MAX)); sqlite3StrAccumInit(&str, zBuf, sizeof(zBuf), SQLITE_MAX_LENGTH); str.db = db; sqlite3StrAccumAppendAll(&str, isSearch ? "SEARCH" : "SCAN"); if( pItem->pSelect ){ sqlite3XPrintf(&str, 0, " SUBQUERY %d", pItem->iSelectId); }else{ sqlite3XPrintf(&str, 0, " TABLE %s", pItem->zName); } if( pItem->zAlias ){ sqlite3XPrintf(&str, 0, " AS %s", pItem->zAlias); } if( (flags & (WHERE_IPK|WHERE_VIRTUALTABLE))==0 ){ const char *zFmt = 0; Index *pIdx; assert( pLoop->u.btree.pIndex!=0 ); pIdx = pLoop->u.btree.pIndex; assert( !(flags&WHERE_AUTO_INDEX) || (flags&WHERE_IDX_ONLY) ); if( !HasRowid(pItem->pTab) && IsPrimaryKeyIndex(pIdx) ){ if( isSearch ){ zFmt = "PRIMARY KEY"; } }else if( flags & WHERE_PARTIALIDX ){ zFmt = "AUTOMATIC PARTIAL COVERING INDEX"; }else if( flags & WHERE_AUTO_INDEX ){ zFmt = "AUTOMATIC COVERING INDEX"; }else if( flags & WHERE_IDX_ONLY ){ zFmt = "COVERING INDEX %s"; }else{ zFmt = "INDEX %s"; } if( zFmt ){ sqlite3StrAccumAppend(&str, " USING ", 7); sqlite3XPrintf(&str, 0, zFmt, pIdx->zName); explainIndexRange(&str, pLoop, pItem->pTab); } }else if( (flags & WHERE_IPK)!=0 && (flags & WHERE_CONSTRAINT)!=0 ){ const char *zRange; if( flags&(WHERE_COLUMN_EQ|WHERE_COLUMN_IN) ){ zRange = "(rowid=?)"; }else if( (flags&WHERE_BOTH_LIMIT)==WHERE_BOTH_LIMIT ){ zRange = "(rowid>? AND rowid?)"; }else{ assert( flags&WHERE_TOP_LIMIT); zRange = "(rowidu.vtab.idxNum, pLoop->u.vtab.idxStr); } #endif #ifdef SQLITE_EXPLAIN_ESTIMATED_ROWS if( pLoop->nOut>=10 ){ sqlite3XPrintf(&str, 0, " (~%llu rows)", sqlite3LogEstToInt(pLoop->nOut)); }else{ sqlite3StrAccumAppend(&str, " (~1 row)", 9); } #endif zMsg = sqlite3StrAccumFinish(&str); ret = sqlite3VdbeAddOp4(v, OP_Explain, iId, iLevel, iFrom, zMsg,P4_DYNAMIC); } return ret; } #else # define explainOneScan(u,v,w,x,y,z) 0 #endif /* SQLITE_OMIT_EXPLAIN */ #ifdef SQLITE_ENABLE_STMT_SCANSTATUS /* ** Configure the VM passed as the first argument with an ** sqlite3_stmt_scanstatus() entry corresponding to the scan used to ** implement level pLvl. Argument pSrclist is a pointer to the FROM ** clause that the scan reads data from. ** ** If argument addrExplain is not 0, it must be the address of an ** OP_Explain instruction that describes the same loop. */ static void addScanStatus( Vdbe *v, /* Vdbe to add scanstatus entry to */ SrcList *pSrclist, /* FROM clause pLvl reads data from */ WhereLevel *pLvl, /* Level to add scanstatus() entry for */ int addrExplain /* Address of OP_Explain (or 0) */ ){ const char *zObj = 0; WhereLoop *pLoop = pLvl->pWLoop; if( (pLoop->wsFlags & (WHERE_IPK|WHERE_VIRTUALTABLE))==0 ){ zObj = pLoop->u.btree.pIndex->zName; }else{ zObj = pSrclist->a[pLvl->iFrom].zName; } sqlite3VdbeScanStatus( v, addrExplain, pLvl->addrBody, pLvl->addrVisit, pLoop->nOut, zObj ); } #else # define addScanStatus(a, b, c, d) ((void)d) #endif /* ** Generate code for the start of the iLevel-th loop in the WHERE clause ** implementation described by pWInfo. */ static Bitmask codeOneLoopStart( WhereInfo *pWInfo, /* Complete information about the WHERE clause */ int iLevel, /* Which level of pWInfo->a[] should be coded */ Bitmask notReady /* Which tables are currently available */ ){ int j, k; /* Loop counters */ int iCur; /* The VDBE cursor for the table */ int addrNxt; /* Where to jump to continue with the next IN case */ int omitTable; /* True if we use the index only */ int bRev; /* True if we need to scan in reverse order */ WhereLevel *pLevel; /* The where level to be coded */ WhereLoop *pLoop; /* The WhereLoop object being coded */ WhereClause *pWC; /* Decomposition of the entire WHERE clause */ WhereTerm *pTerm; /* A WHERE clause term */ Parse *pParse; /* Parsing context */ sqlite3 *db; /* Database connection */ Vdbe *v; /* The prepared stmt under constructions */ struct SrcList_item *pTabItem; /* FROM clause term being coded */ int addrBrk; /* Jump here to break out of the loop */ int addrCont; /* Jump here to continue with next cycle */ int iRowidReg = 0; /* Rowid is stored in this register, if not zero */ int iReleaseReg = 0; /* Temp register to free before returning */ pParse = pWInfo->pParse; v = pParse->pVdbe; pWC = &pWInfo->sWC; db = pParse->db; pLevel = &pWInfo->a[iLevel]; pLoop = pLevel->pWLoop; pTabItem = &pWInfo->pTabList->a[pLevel->iFrom]; iCur = pTabItem->iCursor; pLevel->notReady = notReady & ~getMask(&pWInfo->sMaskSet, iCur); bRev = (pWInfo->revMask>>iLevel)&1; omitTable = (pLoop->wsFlags & WHERE_IDX_ONLY)!=0 && (pWInfo->wctrlFlags & WHERE_FORCE_TABLE)==0; VdbeModuleComment((v, "Begin WHERE-loop%d: %s",iLevel,pTabItem->pTab->zName)); /* Create labels for the "break" and "continue" instructions ** for the current loop. Jump to addrBrk to break out of a loop. ** Jump to cont to go immediately to the next iteration of the ** loop. ** ** When there is an IN operator, we also have a "addrNxt" label that ** means to continue with the next IN value combination. When ** there are no IN operators in the constraints, the "addrNxt" label ** is the same as "addrBrk". */ addrBrk = pLevel->addrBrk = pLevel->addrNxt = sqlite3VdbeMakeLabel(v); addrCont = pLevel->addrCont = sqlite3VdbeMakeLabel(v); /* If this is the right table of a LEFT OUTER JOIN, allocate and ** initialize a memory cell that records if this table matches any ** row of the left table of the join. */ if( pLevel->iFrom>0 && (pTabItem[0].jointype & JT_LEFT)!=0 ){ pLevel->iLeftJoin = ++pParse->nMem; sqlite3VdbeAddOp2(v, OP_Integer, 0, pLevel->iLeftJoin); VdbeComment((v, "init LEFT JOIN no-match flag")); } /* Special case of a FROM clause subquery implemented as a co-routine */ if( pTabItem->viaCoroutine ){ int regYield = pTabItem->regReturn; sqlite3VdbeAddOp3(v, OP_InitCoroutine, regYield, 0, pTabItem->addrFillSub); pLevel->p2 = sqlite3VdbeAddOp2(v, OP_Yield, regYield, addrBrk); VdbeCoverage(v); VdbeComment((v, "next row of \"%s\"", pTabItem->pTab->zName)); pLevel->op = OP_Goto; }else #ifndef SQLITE_OMIT_VIRTUALTABLE if( (pLoop->wsFlags & WHERE_VIRTUALTABLE)!=0 ){ /* Case 1: The table is a virtual-table. Use the VFilter and VNext ** to access the data. */ int iReg; /* P3 Value for OP_VFilter */ int addrNotFound; int nConstraint = pLoop->nLTerm; sqlite3ExprCachePush(pParse); iReg = sqlite3GetTempRange(pParse, nConstraint+2); addrNotFound = pLevel->addrBrk; for(j=0; jaLTerm[j]; if( pTerm==0 ) continue; if( pTerm->eOperator & WO_IN ){ codeEqualityTerm(pParse, pTerm, pLevel, j, bRev, iTarget); addrNotFound = pLevel->addrNxt; }else{ sqlite3ExprCode(pParse, pTerm->pExpr->pRight, iTarget); } } sqlite3VdbeAddOp2(v, OP_Integer, pLoop->u.vtab.idxNum, iReg); sqlite3VdbeAddOp2(v, OP_Integer, nConstraint, iReg+1); sqlite3VdbeAddOp4(v, OP_VFilter, iCur, addrNotFound, iReg, pLoop->u.vtab.idxStr, pLoop->u.vtab.needFree ? P4_MPRINTF : P4_STATIC); VdbeCoverage(v); pLoop->u.vtab.needFree = 0; for(j=0; ju.vtab.omitMask>>j)&1 ){ disableTerm(pLevel, pLoop->aLTerm[j]); } } pLevel->op = OP_VNext; pLevel->p1 = iCur; pLevel->p2 = sqlite3VdbeCurrentAddr(v); sqlite3ReleaseTempRange(pParse, iReg, nConstraint+2); sqlite3ExprCachePop(pParse); }else #endif /* SQLITE_OMIT_VIRTUALTABLE */ if( (pLoop->wsFlags & WHERE_IPK)!=0 && (pLoop->wsFlags & (WHERE_COLUMN_IN|WHERE_COLUMN_EQ))!=0 ){ /* Case 2: We can directly reference a single row using an ** equality comparison against the ROWID field. Or ** we reference multiple rows using a "rowid IN (...)" ** construct. */ assert( pLoop->u.btree.nEq==1 ); pTerm = pLoop->aLTerm[0]; assert( pTerm!=0 ); assert( pTerm->pExpr!=0 ); assert( omitTable==0 ); testcase( pTerm->wtFlags & TERM_VIRTUAL ); iReleaseReg = ++pParse->nMem; iRowidReg = codeEqualityTerm(pParse, pTerm, pLevel, 0, bRev, iReleaseReg); if( iRowidReg!=iReleaseReg ) sqlite3ReleaseTempReg(pParse, iReleaseReg); addrNxt = pLevel->addrNxt; sqlite3VdbeAddOp2(v, OP_MustBeInt, iRowidReg, addrNxt); VdbeCoverage(v); sqlite3VdbeAddOp3(v, OP_NotExists, iCur, addrNxt, iRowidReg); VdbeCoverage(v); sqlite3ExprCacheAffinityChange(pParse, iRowidReg, 1); sqlite3ExprCacheStore(pParse, iCur, -1, iRowidReg); VdbeComment((v, "pk")); pLevel->op = OP_Noop; }else if( (pLoop->wsFlags & WHERE_IPK)!=0 && (pLoop->wsFlags & WHERE_COLUMN_RANGE)!=0 ){ /* Case 3: We have an inequality comparison against the ROWID field. */ int testOp = OP_Noop; int start; int memEndValue = 0; WhereTerm *pStart, *pEnd; assert( omitTable==0 ); j = 0; pStart = pEnd = 0; if( pLoop->wsFlags & WHERE_BTM_LIMIT ) pStart = pLoop->aLTerm[j++]; if( pLoop->wsFlags & WHERE_TOP_LIMIT ) pEnd = pLoop->aLTerm[j++]; assert( pStart!=0 || pEnd!=0 ); if( bRev ){ pTerm = pStart; pStart = pEnd; pEnd = pTerm; } if( pStart ){ Expr *pX; /* The expression that defines the start bound */ int r1, rTemp; /* Registers for holding the start boundary */ /* The following constant maps TK_xx codes into corresponding ** seek opcodes. It depends on a particular ordering of TK_xx */ const u8 aMoveOp[] = { /* TK_GT */ OP_SeekGT, /* TK_LE */ OP_SeekLE, /* TK_LT */ OP_SeekLT, /* TK_GE */ OP_SeekGE }; assert( TK_LE==TK_GT+1 ); /* Make sure the ordering.. */ assert( TK_LT==TK_GT+2 ); /* ... of the TK_xx values... */ assert( TK_GE==TK_GT+3 ); /* ... is correcct. */ assert( (pStart->wtFlags & TERM_VNULL)==0 ); testcase( pStart->wtFlags & TERM_VIRTUAL ); pX = pStart->pExpr; assert( pX!=0 ); testcase( pStart->leftCursor!=iCur ); /* transitive constraints */ r1 = sqlite3ExprCodeTemp(pParse, pX->pRight, &rTemp); sqlite3VdbeAddOp3(v, aMoveOp[pX->op-TK_GT], iCur, addrBrk, r1); VdbeComment((v, "pk")); VdbeCoverageIf(v, pX->op==TK_GT); VdbeCoverageIf(v, pX->op==TK_LE); VdbeCoverageIf(v, pX->op==TK_LT); VdbeCoverageIf(v, pX->op==TK_GE); sqlite3ExprCacheAffinityChange(pParse, r1, 1); sqlite3ReleaseTempReg(pParse, rTemp); disableTerm(pLevel, pStart); }else{ sqlite3VdbeAddOp2(v, bRev ? OP_Last : OP_Rewind, iCur, addrBrk); VdbeCoverageIf(v, bRev==0); VdbeCoverageIf(v, bRev!=0); } if( pEnd ){ Expr *pX; pX = pEnd->pExpr; assert( pX!=0 ); assert( (pEnd->wtFlags & TERM_VNULL)==0 ); testcase( pEnd->leftCursor!=iCur ); /* Transitive constraints */ testcase( pEnd->wtFlags & TERM_VIRTUAL ); memEndValue = ++pParse->nMem; sqlite3ExprCode(pParse, pX->pRight, memEndValue); if( pX->op==TK_LT || pX->op==TK_GT ){ testOp = bRev ? OP_Le : OP_Ge; }else{ testOp = bRev ? OP_Lt : OP_Gt; } disableTerm(pLevel, pEnd); } start = sqlite3VdbeCurrentAddr(v); pLevel->op = bRev ? OP_Prev : OP_Next; pLevel->p1 = iCur; pLevel->p2 = start; assert( pLevel->p5==0 ); if( testOp!=OP_Noop ){ iRowidReg = ++pParse->nMem; sqlite3VdbeAddOp2(v, OP_Rowid, iCur, iRowidReg); sqlite3ExprCacheStore(pParse, iCur, -1, iRowidReg); sqlite3VdbeAddOp3(v, testOp, memEndValue, addrBrk, iRowidReg); VdbeCoverageIf(v, testOp==OP_Le); VdbeCoverageIf(v, testOp==OP_Lt); VdbeCoverageIf(v, testOp==OP_Ge); VdbeCoverageIf(v, testOp==OP_Gt); sqlite3VdbeChangeP5(v, SQLITE_AFF_NUMERIC | SQLITE_JUMPIFNULL); } }else if( pLoop->wsFlags & WHERE_INDEXED ){ /* Case 4: A scan using an index. ** ** The WHERE clause may contain zero or more equality ** terms ("==" or "IN" operators) that refer to the N ** left-most columns of the index. It may also contain ** inequality constraints (>, <, >= or <=) on the indexed ** column that immediately follows the N equalities. Only ** the right-most column can be an inequality - the rest must ** use the "==" and "IN" operators. For example, if the ** index is on (x,y,z), then the following clauses are all ** optimized: ** ** x=5 ** x=5 AND y=10 ** x=5 AND y<10 ** x=5 AND y>5 AND y<10 ** x=5 AND y=5 AND z<=10 ** ** The z<10 term of the following cannot be used, only ** the x=5 term: ** ** x=5 AND z<10 ** ** N may be zero if there are inequality constraints. ** If there are no inequality constraints, then N is at ** least one. ** ** This case is also used when there are no WHERE clause ** constraints but an index is selected anyway, in order ** to force the output order to conform to an ORDER BY. */ static const u8 aStartOp[] = { 0, 0, OP_Rewind, /* 2: (!start_constraints && startEq && !bRev) */ OP_Last, /* 3: (!start_constraints && startEq && bRev) */ OP_SeekGT, /* 4: (start_constraints && !startEq && !bRev) */ OP_SeekLT, /* 5: (start_constraints && !startEq && bRev) */ OP_SeekGE, /* 6: (start_constraints && startEq && !bRev) */ OP_SeekLE /* 7: (start_constraints && startEq && bRev) */ }; static const u8 aEndOp[] = { OP_IdxGE, /* 0: (end_constraints && !bRev && !endEq) */ OP_IdxGT, /* 1: (end_constraints && !bRev && endEq) */ OP_IdxLE, /* 2: (end_constraints && bRev && !endEq) */ OP_IdxLT, /* 3: (end_constraints && bRev && endEq) */ }; u16 nEq = pLoop->u.btree.nEq; /* Number of == or IN terms */ int regBase; /* Base register holding constraint values */ WhereTerm *pRangeStart = 0; /* Inequality constraint at range start */ WhereTerm *pRangeEnd = 0; /* Inequality constraint at range end */ int startEq; /* True if range start uses ==, >= or <= */ int endEq; /* True if range end uses ==, >= or <= */ int start_constraints; /* Start of range is constrained */ int nConstraint; /* Number of constraint terms */ Index *pIdx; /* The index we will be using */ int iIdxCur; /* The VDBE cursor for the index */ int nExtraReg = 0; /* Number of extra registers needed */ int op; /* Instruction opcode */ char *zStartAff; /* Affinity for start of range constraint */ char cEndAff = 0; /* Affinity for end of range constraint */ u8 bSeekPastNull = 0; /* True to seek past initial nulls */ u8 bStopAtNull = 0; /* Add condition to terminate at NULLs */ pIdx = pLoop->u.btree.pIndex; iIdxCur = pLevel->iIdxCur; assert( nEq>=pLoop->nSkip ); /* If this loop satisfies a sort order (pOrderBy) request that ** was passed to this function to implement a "SELECT min(x) ..." ** query, then the caller will only allow the loop to run for ** a single iteration. This means that the first row returned ** should not have a NULL value stored in 'x'. If column 'x' is ** the first one after the nEq equality constraints in the index, ** this requires some special handling. */ assert( pWInfo->pOrderBy==0 || pWInfo->pOrderBy->nExpr==1 || (pWInfo->wctrlFlags&WHERE_ORDERBY_MIN)==0 ); if( (pWInfo->wctrlFlags&WHERE_ORDERBY_MIN)!=0 && pWInfo->nOBSat>0 && (pIdx->nKeyCol>nEq) ){ assert( pLoop->nSkip==0 ); bSeekPastNull = 1; nExtraReg = 1; } /* Find any inequality constraint terms for the start and end ** of the range. */ j = nEq; if( pLoop->wsFlags & WHERE_BTM_LIMIT ){ pRangeStart = pLoop->aLTerm[j++]; nExtraReg = 1; } if( pLoop->wsFlags & WHERE_TOP_LIMIT ){ pRangeEnd = pLoop->aLTerm[j++]; nExtraReg = 1; if( pRangeStart==0 && (j = pIdx->aiColumn[nEq])>=0 && pIdx->pTable->aCol[j].notNull==0 ){ bSeekPastNull = 1; } } assert( pRangeEnd==0 || (pRangeEnd->wtFlags & TERM_VNULL)==0 ); /* Generate code to evaluate all constraint terms using == or IN ** and store the values of those terms in an array of registers ** starting at regBase. */ regBase = codeAllEqualityTerms(pParse,pLevel,bRev,nExtraReg,&zStartAff); assert( zStartAff==0 || sqlite3Strlen30(zStartAff)>=nEq ); if( zStartAff ) cEndAff = zStartAff[nEq]; addrNxt = pLevel->addrNxt; /* If we are doing a reverse order scan on an ascending index, or ** a forward order scan on a descending index, interchange the ** start and end terms (pRangeStart and pRangeEnd). */ if( (nEqnKeyCol && bRev==(pIdx->aSortOrder[nEq]==SQLITE_SO_ASC)) || (bRev && pIdx->nKeyCol==nEq) ){ SWAP(WhereTerm *, pRangeEnd, pRangeStart); SWAP(u8, bSeekPastNull, bStopAtNull); } testcase( pRangeStart && (pRangeStart->eOperator & WO_LE)!=0 ); testcase( pRangeStart && (pRangeStart->eOperator & WO_GE)!=0 ); testcase( pRangeEnd && (pRangeEnd->eOperator & WO_LE)!=0 ); testcase( pRangeEnd && (pRangeEnd->eOperator & WO_GE)!=0 ); startEq = !pRangeStart || pRangeStart->eOperator & (WO_LE|WO_GE); endEq = !pRangeEnd || pRangeEnd->eOperator & (WO_LE|WO_GE); start_constraints = pRangeStart || nEq>0; /* Seek the index cursor to the start of the range. */ nConstraint = nEq; if( pRangeStart ){ Expr *pRight = pRangeStart->pExpr->pRight; sqlite3ExprCode(pParse, pRight, regBase+nEq); if( (pRangeStart->wtFlags & TERM_VNULL)==0 && sqlite3ExprCanBeNull(pRight) ){ sqlite3VdbeAddOp2(v, OP_IsNull, regBase+nEq, addrNxt); VdbeCoverage(v); } if( zStartAff ){ if( sqlite3CompareAffinity(pRight, zStartAff[nEq])==SQLITE_AFF_NONE){ /* Since the comparison is to be performed with no conversions ** applied to the operands, set the affinity to apply to pRight to ** SQLITE_AFF_NONE. */ zStartAff[nEq] = SQLITE_AFF_NONE; } if( sqlite3ExprNeedsNoAffinityChange(pRight, zStartAff[nEq]) ){ zStartAff[nEq] = SQLITE_AFF_NONE; } } nConstraint++; testcase( pRangeStart->wtFlags & TERM_VIRTUAL ); }else if( bSeekPastNull ){ sqlite3VdbeAddOp2(v, OP_Null, 0, regBase+nEq); nConstraint++; startEq = 0; start_constraints = 1; } codeApplyAffinity(pParse, regBase, nConstraint - bSeekPastNull, zStartAff); op = aStartOp[(start_constraints<<2) + (startEq<<1) + bRev]; assert( op!=0 ); sqlite3VdbeAddOp4Int(v, op, iIdxCur, addrNxt, regBase, nConstraint); VdbeCoverage(v); VdbeCoverageIf(v, op==OP_Rewind); testcase( op==OP_Rewind ); VdbeCoverageIf(v, op==OP_Last); testcase( op==OP_Last ); VdbeCoverageIf(v, op==OP_SeekGT); testcase( op==OP_SeekGT ); VdbeCoverageIf(v, op==OP_SeekGE); testcase( op==OP_SeekGE ); VdbeCoverageIf(v, op==OP_SeekLE); testcase( op==OP_SeekLE ); VdbeCoverageIf(v, op==OP_SeekLT); testcase( op==OP_SeekLT ); /* Load the value for the inequality constraint at the end of the ** range (if any). */ nConstraint = nEq; if( pRangeEnd ){ Expr *pRight = pRangeEnd->pExpr->pRight; sqlite3ExprCacheRemove(pParse, regBase+nEq, 1); sqlite3ExprCode(pParse, pRight, regBase+nEq); if( (pRangeEnd->wtFlags & TERM_VNULL)==0 && sqlite3ExprCanBeNull(pRight) ){ sqlite3VdbeAddOp2(v, OP_IsNull, regBase+nEq, addrNxt); VdbeCoverage(v); } if( sqlite3CompareAffinity(pRight, cEndAff)!=SQLITE_AFF_NONE && !sqlite3ExprNeedsNoAffinityChange(pRight, cEndAff) ){ codeApplyAffinity(pParse, regBase+nEq, 1, &cEndAff); } nConstraint++; testcase( pRangeEnd->wtFlags & TERM_VIRTUAL ); }else if( bStopAtNull ){ sqlite3VdbeAddOp2(v, OP_Null, 0, regBase+nEq); endEq = 0; nConstraint++; } sqlite3DbFree(db, zStartAff); /* Top of the loop body */ pLevel->p2 = sqlite3VdbeCurrentAddr(v); /* Check if the index cursor is past the end of the range. */ if( nConstraint ){ op = aEndOp[bRev*2 + endEq]; sqlite3VdbeAddOp4Int(v, op, iIdxCur, addrNxt, regBase, nConstraint); testcase( op==OP_IdxGT ); VdbeCoverageIf(v, op==OP_IdxGT ); testcase( op==OP_IdxGE ); VdbeCoverageIf(v, op==OP_IdxGE ); testcase( op==OP_IdxLT ); VdbeCoverageIf(v, op==OP_IdxLT ); testcase( op==OP_IdxLE ); VdbeCoverageIf(v, op==OP_IdxLE ); } /* Seek the table cursor, if required */ disableTerm(pLevel, pRangeStart); disableTerm(pLevel, pRangeEnd); if( omitTable ){ /* pIdx is a covering index. No need to access the main table. */ }else if( HasRowid(pIdx->pTable) ){ iRowidReg = ++pParse->nMem; sqlite3VdbeAddOp2(v, OP_IdxRowid, iIdxCur, iRowidReg); sqlite3ExprCacheStore(pParse, iCur, -1, iRowidReg); sqlite3VdbeAddOp2(v, OP_Seek, iCur, iRowidReg); /* Deferred seek */ }else if( iCur!=iIdxCur ){ Index *pPk = sqlite3PrimaryKeyIndex(pIdx->pTable); iRowidReg = sqlite3GetTempRange(pParse, pPk->nKeyCol); for(j=0; jnKeyCol; j++){ k = sqlite3ColumnOfIndex(pIdx, pPk->aiColumn[j]); sqlite3VdbeAddOp3(v, OP_Column, iIdxCur, k, iRowidReg+j); } sqlite3VdbeAddOp4Int(v, OP_NotFound, iCur, addrCont, iRowidReg, pPk->nKeyCol); VdbeCoverage(v); } /* Record the instruction used to terminate the loop. Disable ** WHERE clause terms made redundant by the index range scan. */ if( pLoop->wsFlags & WHERE_ONEROW ){ pLevel->op = OP_Noop; }else if( bRev ){ pLevel->op = OP_Prev; }else{ pLevel->op = OP_Next; } pLevel->p1 = iIdxCur; pLevel->p3 = (pLoop->wsFlags&WHERE_UNQ_WANTED)!=0 ? 1:0; if( (pLoop->wsFlags & WHERE_CONSTRAINT)==0 ){ pLevel->p5 = SQLITE_STMTSTATUS_FULLSCAN_STEP; }else{ assert( pLevel->p5==0 ); } }else #ifndef SQLITE_OMIT_OR_OPTIMIZATION if( pLoop->wsFlags & WHERE_MULTI_OR ){ /* Case 5: Two or more separately indexed terms connected by OR ** ** Example: ** ** CREATE TABLE t1(a,b,c,d); ** CREATE INDEX i1 ON t1(a); ** CREATE INDEX i2 ON t1(b); ** CREATE INDEX i3 ON t1(c); ** ** SELECT * FROM t1 WHERE a=5 OR b=7 OR (c=11 AND d=13) ** ** In the example, there are three indexed terms connected by OR. ** The top of the loop looks like this: ** ** Null 1 # Zero the rowset in reg 1 ** ** Then, for each indexed term, the following. The arguments to ** RowSetTest are such that the rowid of the current row is inserted ** into the RowSet. If it is already present, control skips the ** Gosub opcode and jumps straight to the code generated by WhereEnd(). ** ** sqlite3WhereBegin() ** RowSetTest # Insert rowid into rowset ** Gosub 2 A ** sqlite3WhereEnd() ** ** Following the above, code to terminate the loop. Label A, the target ** of the Gosub above, jumps to the instruction right after the Goto. ** ** Null 1 # Zero the rowset in reg 1 ** Goto B # The loop is finished. ** ** A: # Return data, whatever. ** ** Return 2 # Jump back to the Gosub ** ** B: ** ** Added 2014-05-26: If the table is a WITHOUT ROWID table, then ** use an ephemeral index instead of a RowSet to record the primary ** keys of the rows we have already seen. ** */ WhereClause *pOrWc; /* The OR-clause broken out into subterms */ SrcList *pOrTab; /* Shortened table list or OR-clause generation */ Index *pCov = 0; /* Potential covering index (or NULL) */ int iCovCur = pParse->nTab++; /* Cursor used for index scans (if any) */ int regReturn = ++pParse->nMem; /* Register used with OP_Gosub */ int regRowset = 0; /* Register for RowSet object */ int regRowid = 0; /* Register holding rowid */ int iLoopBody = sqlite3VdbeMakeLabel(v); /* Start of loop body */ int iRetInit; /* Address of regReturn init */ int untestedTerms = 0; /* Some terms not completely tested */ int ii; /* Loop counter */ u16 wctrlFlags; /* Flags for sub-WHERE clause */ Expr *pAndExpr = 0; /* An ".. AND (...)" expression */ Table *pTab = pTabItem->pTab; pTerm = pLoop->aLTerm[0]; assert( pTerm!=0 ); assert( pTerm->eOperator & WO_OR ); assert( (pTerm->wtFlags & TERM_ORINFO)!=0 ); pOrWc = &pTerm->u.pOrInfo->wc; pLevel->op = OP_Return; pLevel->p1 = regReturn; /* Set up a new SrcList in pOrTab containing the table being scanned ** by this loop in the a[0] slot and all notReady tables in a[1..] slots. ** This becomes the SrcList in the recursive call to sqlite3WhereBegin(). */ if( pWInfo->nLevel>1 ){ int nNotReady; /* The number of notReady tables */ struct SrcList_item *origSrc; /* Original list of tables */ nNotReady = pWInfo->nLevel - iLevel - 1; pOrTab = sqlite3StackAllocRaw(db, sizeof(*pOrTab)+ nNotReady*sizeof(pOrTab->a[0])); if( pOrTab==0 ) return notReady; pOrTab->nAlloc = (u8)(nNotReady + 1); pOrTab->nSrc = pOrTab->nAlloc; memcpy(pOrTab->a, pTabItem, sizeof(*pTabItem)); origSrc = pWInfo->pTabList->a; for(k=1; k<=nNotReady; k++){ memcpy(&pOrTab->a[k], &origSrc[pLevel[k].iFrom], sizeof(pOrTab->a[k])); } }else{ pOrTab = pWInfo->pTabList; } /* Initialize the rowset register to contain NULL. An SQL NULL is ** equivalent to an empty rowset. Or, create an ephemeral index ** capable of holding primary keys in the case of a WITHOUT ROWID. ** ** Also initialize regReturn to contain the address of the instruction ** immediately following the OP_Return at the bottom of the loop. This ** is required in a few obscure LEFT JOIN cases where control jumps ** over the top of the loop into the body of it. In this case the ** correct response for the end-of-loop code (the OP_Return) is to ** fall through to the next instruction, just as an OP_Next does if ** called on an uninitialized cursor. */ if( (pWInfo->wctrlFlags & WHERE_DUPLICATES_OK)==0 ){ if( HasRowid(pTab) ){ regRowset = ++pParse->nMem; sqlite3VdbeAddOp2(v, OP_Null, 0, regRowset); }else{ Index *pPk = sqlite3PrimaryKeyIndex(pTab); regRowset = pParse->nTab++; sqlite3VdbeAddOp2(v, OP_OpenEphemeral, regRowset, pPk->nKeyCol); sqlite3VdbeSetP4KeyInfo(pParse, pPk); } regRowid = ++pParse->nMem; } iRetInit = sqlite3VdbeAddOp2(v, OP_Integer, 0, regReturn); /* If the original WHERE clause is z of the form: (x1 OR x2 OR ...) AND y ** Then for every term xN, evaluate as the subexpression: xN AND z ** That way, terms in y that are factored into the disjunction will ** be picked up by the recursive calls to sqlite3WhereBegin() below. ** ** Actually, each subexpression is converted to "xN AND w" where w is ** the "interesting" terms of z - terms that did not originate in the ** ON or USING clause of a LEFT JOIN, and terms that are usable as ** indices. ** ** This optimization also only applies if the (x1 OR x2 OR ...) term ** is not contained in the ON clause of a LEFT JOIN. ** See ticket http://www.sqlite.org/src/info/f2369304e4 */ if( pWC->nTerm>1 ){ int iTerm; for(iTerm=0; iTermnTerm; iTerm++){ Expr *pExpr = pWC->a[iTerm].pExpr; if( &pWC->a[iTerm] == pTerm ) continue; if( ExprHasProperty(pExpr, EP_FromJoin) ) continue; testcase( pWC->a[iTerm].wtFlags & TERM_ORINFO ); testcase( pWC->a[iTerm].wtFlags & TERM_VIRTUAL ); if( pWC->a[iTerm].wtFlags & (TERM_ORINFO|TERM_VIRTUAL) ) continue; if( (pWC->a[iTerm].eOperator & WO_ALL)==0 ) continue; pExpr = sqlite3ExprDup(db, pExpr, 0); pAndExpr = sqlite3ExprAnd(db, pAndExpr, pExpr); } if( pAndExpr ){ pAndExpr = sqlite3PExpr(pParse, TK_AND, 0, pAndExpr, 0); } } /* Run a separate WHERE clause for each term of the OR clause. After ** eliminating duplicates from other WHERE clauses, the action for each ** sub-WHERE clause is to to invoke the main loop body as a subroutine. */ wctrlFlags = WHERE_OMIT_OPEN_CLOSE | WHERE_FORCE_TABLE | WHERE_ONETABLE_ONLY; for(ii=0; iinTerm; ii++){ WhereTerm *pOrTerm = &pOrWc->a[ii]; if( pOrTerm->leftCursor==iCur || (pOrTerm->eOperator & WO_AND)!=0 ){ WhereInfo *pSubWInfo; /* Info for single OR-term scan */ Expr *pOrExpr = pOrTerm->pExpr; /* Current OR clause term */ int j1 = 0; /* Address of jump operation */ if( pAndExpr && !ExprHasProperty(pOrExpr, EP_FromJoin) ){ pAndExpr->pLeft = pOrExpr; pOrExpr = pAndExpr; } /* Loop through table entries that match term pOrTerm. */ WHERETRACE(0xffff, ("Subplan for OR-clause:\n")); pSubWInfo = sqlite3WhereBegin(pParse, pOrTab, pOrExpr, 0, 0, wctrlFlags, iCovCur); assert( pSubWInfo || pParse->nErr || db->mallocFailed ); if( pSubWInfo ){ WhereLoop *pSubLoop; int addrExplain = explainOneScan( pParse, pOrTab, &pSubWInfo->a[0], iLevel, pLevel->iFrom, 0 ); addScanStatus(v, pOrTab, &pSubWInfo->a[0], addrExplain); /* This is the sub-WHERE clause body. First skip over ** duplicate rows from prior sub-WHERE clauses, and record the ** rowid (or PRIMARY KEY) for the current row so that the same ** row will be skipped in subsequent sub-WHERE clauses. */ if( (pWInfo->wctrlFlags & WHERE_DUPLICATES_OK)==0 ){ int r; int iSet = ((ii==pOrWc->nTerm-1)?-1:ii); if( HasRowid(pTab) ){ r = sqlite3ExprCodeGetColumn(pParse, pTab, -1, iCur, regRowid, 0); j1 = sqlite3VdbeAddOp4Int(v, OP_RowSetTest, regRowset, 0, r,iSet); VdbeCoverage(v); }else{ Index *pPk = sqlite3PrimaryKeyIndex(pTab); int nPk = pPk->nKeyCol; int iPk; /* Read the PK into an array of temp registers. */ r = sqlite3GetTempRange(pParse, nPk); for(iPk=0; iPkaiColumn[iPk]; sqlite3ExprCodeGetColumn(pParse, pTab, iCol, iCur, r+iPk, 0); } /* Check if the temp table already contains this key. If so, ** the row has already been included in the result set and ** can be ignored (by jumping past the Gosub below). Otherwise, ** insert the key into the temp table and proceed with processing ** the row. ** ** Use some of the same optimizations as OP_RowSetTest: If iSet ** is zero, assume that the key cannot already be present in ** the temp table. And if iSet is -1, assume that there is no ** need to insert the key into the temp table, as it will never ** be tested for. */ if( iSet ){ j1 = sqlite3VdbeAddOp4Int(v, OP_Found, regRowset, 0, r, nPk); VdbeCoverage(v); } if( iSet>=0 ){ sqlite3VdbeAddOp3(v, OP_MakeRecord, r, nPk, regRowid); sqlite3VdbeAddOp3(v, OP_IdxInsert, regRowset, regRowid, 0); if( iSet ) sqlite3VdbeChangeP5(v, OPFLAG_USESEEKRESULT); } /* Release the array of temp registers */ sqlite3ReleaseTempRange(pParse, r, nPk); } } /* Invoke the main loop body as a subroutine */ sqlite3VdbeAddOp2(v, OP_Gosub, regReturn, iLoopBody); /* Jump here (skipping the main loop body subroutine) if the ** current sub-WHERE row is a duplicate from prior sub-WHEREs. */ if( j1 ) sqlite3VdbeJumpHere(v, j1); /* The pSubWInfo->untestedTerms flag means that this OR term ** contained one or more AND term from a notReady table. The ** terms from the notReady table could not be tested and will ** need to be tested later. */ if( pSubWInfo->untestedTerms ) untestedTerms = 1; /* If all of the OR-connected terms are optimized using the same ** index, and the index is opened using the same cursor number ** by each call to sqlite3WhereBegin() made by this loop, it may ** be possible to use that index as a covering index. ** ** If the call to sqlite3WhereBegin() above resulted in a scan that ** uses an index, and this is either the first OR-connected term ** processed or the index is the same as that used by all previous ** terms, set pCov to the candidate covering index. Otherwise, set ** pCov to NULL to indicate that no candidate covering index will ** be available. */ pSubLoop = pSubWInfo->a[0].pWLoop; assert( (pSubLoop->wsFlags & WHERE_AUTO_INDEX)==0 ); if( (pSubLoop->wsFlags & WHERE_INDEXED)!=0 && (ii==0 || pSubLoop->u.btree.pIndex==pCov) && (HasRowid(pTab) || !IsPrimaryKeyIndex(pSubLoop->u.btree.pIndex)) ){ assert( pSubWInfo->a[0].iIdxCur==iCovCur ); pCov = pSubLoop->u.btree.pIndex; wctrlFlags |= WHERE_REOPEN_IDX; }else{ pCov = 0; } /* Finish the loop through table entries that match term pOrTerm. */ sqlite3WhereEnd(pSubWInfo); } } } pLevel->u.pCovidx = pCov; if( pCov ) pLevel->iIdxCur = iCovCur; if( pAndExpr ){ pAndExpr->pLeft = 0; sqlite3ExprDelete(db, pAndExpr); } sqlite3VdbeChangeP1(v, iRetInit, sqlite3VdbeCurrentAddr(v)); sqlite3VdbeAddOp2(v, OP_Goto, 0, pLevel->addrBrk); sqlite3VdbeResolveLabel(v, iLoopBody); if( pWInfo->nLevel>1 ) sqlite3StackFree(db, pOrTab); if( !untestedTerms ) disableTerm(pLevel, pTerm); }else #endif /* SQLITE_OMIT_OR_OPTIMIZATION */ { /* Case 6: There is no usable index. We must do a complete ** scan of the entire table. */ static const u8 aStep[] = { OP_Next, OP_Prev }; static const u8 aStart[] = { OP_Rewind, OP_Last }; assert( bRev==0 || bRev==1 ); if( pTabItem->isRecursive ){ /* Tables marked isRecursive have only a single row that is stored in ** a pseudo-cursor. No need to Rewind or Next such cursors. */ pLevel->op = OP_Noop; }else{ pLevel->op = aStep[bRev]; pLevel->p1 = iCur; pLevel->p2 = 1 + sqlite3VdbeAddOp2(v, aStart[bRev], iCur, addrBrk); VdbeCoverageIf(v, bRev==0); VdbeCoverageIf(v, bRev!=0); pLevel->p5 = SQLITE_STMTSTATUS_FULLSCAN_STEP; } } #ifdef SQLITE_ENABLE_STMT_SCANSTATUS pLevel->addrVisit = sqlite3VdbeCurrentAddr(v); #endif /* Insert code to test every subexpression that can be completely ** computed using the current set of tables. */ for(pTerm=pWC->a, j=pWC->nTerm; j>0; j--, pTerm++){ Expr *pE; testcase( pTerm->wtFlags & TERM_VIRTUAL ); testcase( pTerm->wtFlags & TERM_CODED ); if( pTerm->wtFlags & (TERM_VIRTUAL|TERM_CODED) ) continue; if( (pTerm->prereqAll & pLevel->notReady)!=0 ){ testcase( pWInfo->untestedTerms==0 && (pWInfo->wctrlFlags & WHERE_ONETABLE_ONLY)!=0 ); pWInfo->untestedTerms = 1; continue; } pE = pTerm->pExpr; assert( pE!=0 ); if( pLevel->iLeftJoin && !ExprHasProperty(pE, EP_FromJoin) ){ continue; } sqlite3ExprIfFalse(pParse, pE, addrCont, SQLITE_JUMPIFNULL); pTerm->wtFlags |= TERM_CODED; } /* Insert code to test for implied constraints based on transitivity ** of the "==" operator. ** ** Example: If the WHERE clause contains "t1.a=t2.b" and "t2.b=123" ** and we are coding the t1 loop and the t2 loop has not yet coded, ** then we cannot use the "t1.a=t2.b" constraint, but we can code ** the implied "t1.a=123" constraint. */ for(pTerm=pWC->a, j=pWC->nTerm; j>0; j--, pTerm++){ Expr *pE, *pEAlt; WhereTerm *pAlt; if( pTerm->wtFlags & (TERM_VIRTUAL|TERM_CODED) ) continue; if( pTerm->eOperator!=(WO_EQUIV|WO_EQ) ) continue; if( pTerm->leftCursor!=iCur ) continue; if( pLevel->iLeftJoin ) continue; pE = pTerm->pExpr; assert( !ExprHasProperty(pE, EP_FromJoin) ); assert( (pTerm->prereqRight & pLevel->notReady)!=0 ); pAlt = findTerm(pWC, iCur, pTerm->u.leftColumn, notReady, WO_EQ|WO_IN, 0); if( pAlt==0 ) continue; if( pAlt->wtFlags & (TERM_CODED) ) continue; testcase( pAlt->eOperator & WO_EQ ); testcase( pAlt->eOperator & WO_IN ); VdbeModuleComment((v, "begin transitive constraint")); pEAlt = sqlite3StackAllocRaw(db, sizeof(*pEAlt)); if( pEAlt ){ *pEAlt = *pAlt->pExpr; pEAlt->pLeft = pE->pLeft; sqlite3ExprIfFalse(pParse, pEAlt, addrCont, SQLITE_JUMPIFNULL); sqlite3StackFree(db, pEAlt); } } /* For a LEFT OUTER JOIN, generate code that will record the fact that ** at least one row of the right table has matched the left table. */ if( pLevel->iLeftJoin ){ pLevel->addrFirst = sqlite3VdbeCurrentAddr(v); sqlite3VdbeAddOp2(v, OP_Integer, 1, pLevel->iLeftJoin); VdbeComment((v, "record LEFT JOIN hit")); sqlite3ExprCacheClear(pParse); for(pTerm=pWC->a, j=0; jnTerm; j++, pTerm++){ testcase( pTerm->wtFlags & TERM_VIRTUAL ); testcase( pTerm->wtFlags & TERM_CODED ); if( pTerm->wtFlags & (TERM_VIRTUAL|TERM_CODED) ) continue; if( (pTerm->prereqAll & pLevel->notReady)!=0 ){ assert( pWInfo->untestedTerms ); continue; } assert( pTerm->pExpr ); sqlite3ExprIfFalse(pParse, pTerm->pExpr, addrCont, SQLITE_JUMPIFNULL); pTerm->wtFlags |= TERM_CODED; } } return pLevel->notReady; } #ifdef WHERETRACE_ENABLED /* ** Print the content of a WhereTerm object */ static void whereTermPrint(WhereTerm *pTerm, int iTerm){ if( pTerm==0 ){ sqlite3DebugPrintf("TERM-%-3d NULL\n", iTerm); }else{ char zType[4]; memcpy(zType, "...", 4); if( pTerm->wtFlags & TERM_VIRTUAL ) zType[0] = 'V'; if( pTerm->eOperator & WO_EQUIV ) zType[1] = 'E'; if( ExprHasProperty(pTerm->pExpr, EP_FromJoin) ) zType[2] = 'L'; sqlite3DebugPrintf("TERM-%-3d %p %s cursor=%-3d prob=%-3d op=0x%03x\n", iTerm, pTerm, zType, pTerm->leftCursor, pTerm->truthProb, pTerm->eOperator); sqlite3TreeViewExpr(0, pTerm->pExpr, 0); } } #endif #ifdef WHERETRACE_ENABLED /* ** Print a WhereLoop object for debugging purposes */ static void whereLoopPrint(WhereLoop *p, WhereClause *pWC){ WhereInfo *pWInfo = pWC->pWInfo; int nb = 1+(pWInfo->pTabList->nSrc+7)/8; struct SrcList_item *pItem = pWInfo->pTabList->a + p->iTab; Table *pTab = pItem->pTab; sqlite3DebugPrintf("%c%2d.%0*llx.%0*llx", p->cId, p->iTab, nb, p->maskSelf, nb, p->prereq); sqlite3DebugPrintf(" %12s", pItem->zAlias ? pItem->zAlias : pTab->zName); if( (p->wsFlags & WHERE_VIRTUALTABLE)==0 ){ const char *zName; if( p->u.btree.pIndex && (zName = p->u.btree.pIndex->zName)!=0 ){ if( strncmp(zName, "sqlite_autoindex_", 17)==0 ){ int i = sqlite3Strlen30(zName) - 1; while( zName[i]!='_' ) i--; zName += i; } sqlite3DebugPrintf(".%-16s %2d", zName, p->u.btree.nEq); }else{ sqlite3DebugPrintf("%20s",""); } }else{ char *z; if( p->u.vtab.idxStr ){ z = sqlite3_mprintf("(%d,\"%s\",%x)", p->u.vtab.idxNum, p->u.vtab.idxStr, p->u.vtab.omitMask); }else{ z = sqlite3_mprintf("(%d,%x)", p->u.vtab.idxNum, p->u.vtab.omitMask); } sqlite3DebugPrintf(" %-19s", z); sqlite3_free(z); } if( p->wsFlags & WHERE_SKIPSCAN ){ sqlite3DebugPrintf(" f %05x %d-%d", p->wsFlags, p->nLTerm,p->nSkip); }else{ sqlite3DebugPrintf(" f %05x N %d", p->wsFlags, p->nLTerm); } sqlite3DebugPrintf(" cost %d,%d,%d\n", p->rSetup, p->rRun, p->nOut); if( p->nLTerm && (sqlite3WhereTrace & 0x100)!=0 ){ int i; for(i=0; inLTerm; i++){ whereTermPrint(p->aLTerm[i], i); } } } #endif /* ** Convert bulk memory into a valid WhereLoop that can be passed ** to whereLoopClear harmlessly. */ static void whereLoopInit(WhereLoop *p){ p->aLTerm = p->aLTermSpace; p->nLTerm = 0; p->nLSlot = ArraySize(p->aLTermSpace); p->wsFlags = 0; } /* ** Clear the WhereLoop.u union. Leave WhereLoop.pLTerm intact. */ static void whereLoopClearUnion(sqlite3 *db, WhereLoop *p){ if( p->wsFlags & (WHERE_VIRTUALTABLE|WHERE_AUTO_INDEX) ){ if( (p->wsFlags & WHERE_VIRTUALTABLE)!=0 && p->u.vtab.needFree ){ sqlite3_free(p->u.vtab.idxStr); p->u.vtab.needFree = 0; p->u.vtab.idxStr = 0; }else if( (p->wsFlags & WHERE_AUTO_INDEX)!=0 && p->u.btree.pIndex!=0 ){ sqlite3DbFree(db, p->u.btree.pIndex->zColAff); sqlite3KeyInfoUnref(p->u.btree.pIndex->pKeyInfo); sqlite3DbFree(db, p->u.btree.pIndex); p->u.btree.pIndex = 0; } } } /* ** Deallocate internal memory used by a WhereLoop object */ static void whereLoopClear(sqlite3 *db, WhereLoop *p){ if( p->aLTerm!=p->aLTermSpace ) sqlite3DbFree(db, p->aLTerm); whereLoopClearUnion(db, p); whereLoopInit(p); } /* ** Increase the memory allocation for pLoop->aLTerm[] to be at least n. */ static int whereLoopResize(sqlite3 *db, WhereLoop *p, int n){ WhereTerm **paNew; if( p->nLSlot>=n ) return SQLITE_OK; n = (n+7)&~7; paNew = sqlite3DbMallocRaw(db, sizeof(p->aLTerm[0])*n); if( paNew==0 ) return SQLITE_NOMEM; memcpy(paNew, p->aLTerm, sizeof(p->aLTerm[0])*p->nLSlot); if( p->aLTerm!=p->aLTermSpace ) sqlite3DbFree(db, p->aLTerm); p->aLTerm = paNew; p->nLSlot = n; return SQLITE_OK; } /* ** Transfer content from the second pLoop into the first. */ static int whereLoopXfer(sqlite3 *db, WhereLoop *pTo, WhereLoop *pFrom){ whereLoopClearUnion(db, pTo); if( whereLoopResize(db, pTo, pFrom->nLTerm) ){ memset(&pTo->u, 0, sizeof(pTo->u)); return SQLITE_NOMEM; } memcpy(pTo, pFrom, WHERE_LOOP_XFER_SZ); memcpy(pTo->aLTerm, pFrom->aLTerm, pTo->nLTerm*sizeof(pTo->aLTerm[0])); if( pFrom->wsFlags & WHERE_VIRTUALTABLE ){ pFrom->u.vtab.needFree = 0; }else if( (pFrom->wsFlags & WHERE_AUTO_INDEX)!=0 ){ pFrom->u.btree.pIndex = 0; } return SQLITE_OK; } /* ** Delete a WhereLoop object */ static void whereLoopDelete(sqlite3 *db, WhereLoop *p){ whereLoopClear(db, p); sqlite3DbFree(db, p); } /* ** Free a WhereInfo structure */ static void whereInfoFree(sqlite3 *db, WhereInfo *pWInfo){ if( ALWAYS(pWInfo) ){ whereClauseClear(&pWInfo->sWC); while( pWInfo->pLoops ){ WhereLoop *p = pWInfo->pLoops; pWInfo->pLoops = p->pNextLoop; whereLoopDelete(db, p); } sqlite3DbFree(db, pWInfo); } } /* ** Return TRUE if all of the following are true: ** ** (1) X has the same or lower cost that Y ** (2) X is a proper subset of Y ** (3) X skips at least as many columns as Y ** ** By "proper subset" we mean that X uses fewer WHERE clause terms ** than Y and that every WHERE clause term used by X is also used ** by Y. ** ** If X is a proper subset of Y then Y is a better choice and ought ** to have a lower cost. This routine returns TRUE when that cost ** relationship is inverted and needs to be adjusted. The third rule ** was added because if X uses skip-scan less than Y it still might ** deserve a lower cost even if it is a proper subset of Y. */ static int whereLoopCheaperProperSubset( const WhereLoop *pX, /* First WhereLoop to compare */ const WhereLoop *pY /* Compare against this WhereLoop */ ){ int i, j; if( pX->nLTerm-pX->nSkip >= pY->nLTerm-pY->nSkip ){ return 0; /* X is not a subset of Y */ } if( pY->nSkip > pX->nSkip ) return 0; if( pX->rRun >= pY->rRun ){ if( pX->rRun > pY->rRun ) return 0; /* X costs more than Y */ if( pX->nOut > pY->nOut ) return 0; /* X costs more than Y */ } for(i=pX->nLTerm-1; i>=0; i--){ if( pX->aLTerm[i]==0 ) continue; for(j=pY->nLTerm-1; j>=0; j--){ if( pY->aLTerm[j]==pX->aLTerm[i] ) break; } if( j<0 ) return 0; /* X not a subset of Y since term X[i] not used by Y */ } return 1; /* All conditions meet */ } /* ** Try to adjust the cost of WhereLoop pTemplate upwards or downwards so ** that: ** ** (1) pTemplate costs less than any other WhereLoops that are a proper ** subset of pTemplate ** ** (2) pTemplate costs more than any other WhereLoops for which pTemplate ** is a proper subset. ** ** To say "WhereLoop X is a proper subset of Y" means that X uses fewer ** WHERE clause terms than Y and that every WHERE clause term used by X is ** also used by Y. */ static void whereLoopAdjustCost(const WhereLoop *p, WhereLoop *pTemplate){ if( (pTemplate->wsFlags & WHERE_INDEXED)==0 ) return; for(; p; p=p->pNextLoop){ if( p->iTab!=pTemplate->iTab ) continue; if( (p->wsFlags & WHERE_INDEXED)==0 ) continue; if( whereLoopCheaperProperSubset(p, pTemplate) ){ /* Adjust pTemplate cost downward so that it is cheaper than its ** subset p. */ WHERETRACE(0x80,("subset cost adjustment %d,%d to %d,%d\n", pTemplate->rRun, pTemplate->nOut, p->rRun, p->nOut-1)); pTemplate->rRun = p->rRun; pTemplate->nOut = p->nOut - 1; }else if( whereLoopCheaperProperSubset(pTemplate, p) ){ /* Adjust pTemplate cost upward so that it is costlier than p since ** pTemplate is a proper subset of p */ WHERETRACE(0x80,("subset cost adjustment %d,%d to %d,%d\n", pTemplate->rRun, pTemplate->nOut, p->rRun, p->nOut+1)); pTemplate->rRun = p->rRun; pTemplate->nOut = p->nOut + 1; } } } /* ** Search the list of WhereLoops in *ppPrev looking for one that can be ** supplanted by pTemplate. ** ** Return NULL if the WhereLoop list contains an entry that can supplant ** pTemplate, in other words if pTemplate does not belong on the list. ** ** If pX is a WhereLoop that pTemplate can supplant, then return the ** link that points to pX. ** ** If pTemplate cannot supplant any existing element of the list but needs ** to be added to the list, then return a pointer to the tail of the list. */ static WhereLoop **whereLoopFindLesser( WhereLoop **ppPrev, const WhereLoop *pTemplate ){ WhereLoop *p; for(p=(*ppPrev); p; ppPrev=&p->pNextLoop, p=*ppPrev){ if( p->iTab!=pTemplate->iTab || p->iSortIdx!=pTemplate->iSortIdx ){ /* If either the iTab or iSortIdx values for two WhereLoop are different ** then those WhereLoops need to be considered separately. Neither is ** a candidate to replace the other. */ continue; } /* In the current implementation, the rSetup value is either zero ** or the cost of building an automatic index (NlogN) and the NlogN ** is the same for compatible WhereLoops. */ assert( p->rSetup==0 || pTemplate->rSetup==0 || p->rSetup==pTemplate->rSetup ); /* whereLoopAddBtree() always generates and inserts the automatic index ** case first. Hence compatible candidate WhereLoops never have a larger ** rSetup. Call this SETUP-INVARIANT */ assert( p->rSetup>=pTemplate->rSetup ); /* Any loop using an appliation-defined index (or PRIMARY KEY or ** UNIQUE constraint) with one or more == constraints is better ** than an automatic index. */ if( (p->wsFlags & WHERE_AUTO_INDEX)!=0 && (pTemplate->wsFlags & WHERE_INDEXED)!=0 && (pTemplate->wsFlags & WHERE_COLUMN_EQ)!=0 && (p->prereq & pTemplate->prereq)==pTemplate->prereq ){ break; } /* If existing WhereLoop p is better than pTemplate, pTemplate can be ** discarded. WhereLoop p is better if: ** (1) p has no more dependencies than pTemplate, and ** (2) p has an equal or lower cost than pTemplate */ if( (p->prereq & pTemplate->prereq)==p->prereq /* (1) */ && p->rSetup<=pTemplate->rSetup /* (2a) */ && p->rRun<=pTemplate->rRun /* (2b) */ && p->nOut<=pTemplate->nOut /* (2c) */ ){ return 0; /* Discard pTemplate */ } /* If pTemplate is always better than p, then cause p to be overwritten ** with pTemplate. pTemplate is better than p if: ** (1) pTemplate has no more dependences than p, and ** (2) pTemplate has an equal or lower cost than p. */ if( (p->prereq & pTemplate->prereq)==pTemplate->prereq /* (1) */ && p->rRun>=pTemplate->rRun /* (2a) */ && p->nOut>=pTemplate->nOut /* (2b) */ ){ assert( p->rSetup>=pTemplate->rSetup ); /* SETUP-INVARIANT above */ break; /* Cause p to be overwritten by pTemplate */ } } return ppPrev; } /* ** Insert or replace a WhereLoop entry using the template supplied. ** ** An existing WhereLoop entry might be overwritten if the new template ** is better and has fewer dependencies. Or the template will be ignored ** and no insert will occur if an existing WhereLoop is faster and has ** fewer dependencies than the template. Otherwise a new WhereLoop is ** added based on the template. ** ** If pBuilder->pOrSet is not NULL then we care about only the ** prerequisites and rRun and nOut costs of the N best loops. That ** information is gathered in the pBuilder->pOrSet object. This special ** processing mode is used only for OR clause processing. ** ** When accumulating multiple loops (when pBuilder->pOrSet is NULL) we ** still might overwrite similar loops with the new template if the ** new template is better. Loops may be overwritten if the following ** conditions are met: ** ** (1) They have the same iTab. ** (2) They have the same iSortIdx. ** (3) The template has same or fewer dependencies than the current loop ** (4) The template has the same or lower cost than the current loop */ static int whereLoopInsert(WhereLoopBuilder *pBuilder, WhereLoop *pTemplate){ WhereLoop **ppPrev, *p; WhereInfo *pWInfo = pBuilder->pWInfo; sqlite3 *db = pWInfo->pParse->db; /* If pBuilder->pOrSet is defined, then only keep track of the costs ** and prereqs. */ if( pBuilder->pOrSet!=0 ){ #if WHERETRACE_ENABLED u16 n = pBuilder->pOrSet->n; int x = #endif whereOrInsert(pBuilder->pOrSet, pTemplate->prereq, pTemplate->rRun, pTemplate->nOut); #if WHERETRACE_ENABLED /* 0x8 */ if( sqlite3WhereTrace & 0x8 ){ sqlite3DebugPrintf(x?" or-%d: ":" or-X: ", n); whereLoopPrint(pTemplate, pBuilder->pWC); } #endif return SQLITE_OK; } /* Look for an existing WhereLoop to replace with pTemplate */ whereLoopAdjustCost(pWInfo->pLoops, pTemplate); ppPrev = whereLoopFindLesser(&pWInfo->pLoops, pTemplate); if( ppPrev==0 ){ /* There already exists a WhereLoop on the list that is better ** than pTemplate, so just ignore pTemplate */ #if WHERETRACE_ENABLED /* 0x8 */ if( sqlite3WhereTrace & 0x8 ){ sqlite3DebugPrintf(" skip: "); whereLoopPrint(pTemplate, pBuilder->pWC); } #endif return SQLITE_OK; }else{ p = *ppPrev; } /* If we reach this point it means that either p[] should be overwritten ** with pTemplate[] if p[] exists, or if p==NULL then allocate a new ** WhereLoop and insert it. */ #if WHERETRACE_ENABLED /* 0x8 */ if( sqlite3WhereTrace & 0x8 ){ if( p!=0 ){ sqlite3DebugPrintf("replace: "); whereLoopPrint(p, pBuilder->pWC); } sqlite3DebugPrintf(" add: "); whereLoopPrint(pTemplate, pBuilder->pWC); } #endif if( p==0 ){ /* Allocate a new WhereLoop to add to the end of the list */ *ppPrev = p = sqlite3DbMallocRaw(db, sizeof(WhereLoop)); if( p==0 ) return SQLITE_NOMEM; whereLoopInit(p); p->pNextLoop = 0; }else{ /* We will be overwriting WhereLoop p[]. But before we do, first ** go through the rest of the list and delete any other entries besides ** p[] that are also supplated by pTemplate */ WhereLoop **ppTail = &p->pNextLoop; WhereLoop *pToDel; while( *ppTail ){ ppTail = whereLoopFindLesser(ppTail, pTemplate); if( ppTail==0 ) break; pToDel = *ppTail; if( pToDel==0 ) break; *ppTail = pToDel->pNextLoop; #if WHERETRACE_ENABLED /* 0x8 */ if( sqlite3WhereTrace & 0x8 ){ sqlite3DebugPrintf(" delete: "); whereLoopPrint(pToDel, pBuilder->pWC); } #endif whereLoopDelete(db, pToDel); } } whereLoopXfer(db, p, pTemplate); if( (p->wsFlags & WHERE_VIRTUALTABLE)==0 ){ Index *pIndex = p->u.btree.pIndex; if( pIndex && pIndex->tnum==0 ){ p->u.btree.pIndex = 0; } } return SQLITE_OK; } /* ** Adjust the WhereLoop.nOut value downward to account for terms of the ** WHERE clause that reference the loop but which are not used by an ** index. ** ** In the current implementation, the first extra WHERE clause term reduces ** the number of output rows by a factor of 10 and each additional term ** reduces the number of output rows by sqrt(2). */ static void whereLoopOutputAdjust( WhereClause *pWC, /* The WHERE clause */ WhereLoop *pLoop, /* The loop to adjust downward */ LogEst nRow /* Number of rows in the entire table */ ){ WhereTerm *pTerm, *pX; Bitmask notAllowed = ~(pLoop->prereq|pLoop->maskSelf); int i, j; int nEq = 0; /* Number of = constraints not within likely()/unlikely() */ for(i=pWC->nTerm, pTerm=pWC->a; i>0; i--, pTerm++){ if( (pTerm->wtFlags & TERM_VIRTUAL)!=0 ) break; if( (pTerm->prereqAll & pLoop->maskSelf)==0 ) continue; if( (pTerm->prereqAll & notAllowed)!=0 ) continue; for(j=pLoop->nLTerm-1; j>=0; j--){ pX = pLoop->aLTerm[j]; if( pX==0 ) continue; if( pX==pTerm ) break; if( pX->iParent>=0 && (&pWC->a[pX->iParent])==pTerm ) break; } if( j<0 ){ if( pTerm->truthProb<=0 ){ pLoop->nOut += pTerm->truthProb; }else{ pLoop->nOut--; if( pTerm->eOperator&WO_EQ ) nEq++; } } } /* TUNING: If there is at least one equality constraint in the WHERE ** clause that does not have a likelihood() explicitly assigned to it ** then do not let the estimated number of output rows exceed half ** the number of rows in the table. */ if( nEq && pLoop->nOut>nRow-10 ){ pLoop->nOut = nRow - 10; } } /* ** Adjust the cost C by the costMult facter T. This only occurs if ** compiled with -DSQLITE_ENABLE_COSTMULT */ #ifdef SQLITE_ENABLE_COSTMULT # define ApplyCostMultiplier(C,T) C += T #else # define ApplyCostMultiplier(C,T) #endif /* ** We have so far matched pBuilder->pNew->u.btree.nEq terms of the ** index pIndex. Try to match one more. ** ** When this function is called, pBuilder->pNew->nOut contains the ** number of rows expected to be visited by filtering using the nEq ** terms only. If it is modified, this value is restored before this ** function returns. ** ** If pProbe->tnum==0, that means pIndex is a fake index used for the ** INTEGER PRIMARY KEY. */ static int whereLoopAddBtreeIndex( WhereLoopBuilder *pBuilder, /* The WhereLoop factory */ struct SrcList_item *pSrc, /* FROM clause term being analyzed */ Index *pProbe, /* An index on pSrc */ LogEst nInMul /* log(Number of iterations due to IN) */ ){ WhereInfo *pWInfo = pBuilder->pWInfo; /* WHERE analyse context */ Parse *pParse = pWInfo->pParse; /* Parsing context */ sqlite3 *db = pParse->db; /* Database connection malloc context */ WhereLoop *pNew; /* Template WhereLoop under construction */ WhereTerm *pTerm; /* A WhereTerm under consideration */ int opMask; /* Valid operators for constraints */ WhereScan scan; /* Iterator for WHERE terms */ Bitmask saved_prereq; /* Original value of pNew->prereq */ u16 saved_nLTerm; /* Original value of pNew->nLTerm */ u16 saved_nEq; /* Original value of pNew->u.btree.nEq */ u16 saved_nSkip; /* Original value of pNew->nSkip */ u32 saved_wsFlags; /* Original value of pNew->wsFlags */ LogEst saved_nOut; /* Original value of pNew->nOut */ int iCol; /* Index of the column in the table */ int rc = SQLITE_OK; /* Return code */ LogEst rSize; /* Number of rows in the table */ LogEst rLogSize; /* Logarithm of table size */ WhereTerm *pTop = 0, *pBtm = 0; /* Top and bottom range constraints */ pNew = pBuilder->pNew; if( db->mallocFailed ) return SQLITE_NOMEM; assert( (pNew->wsFlags & WHERE_VIRTUALTABLE)==0 ); assert( (pNew->wsFlags & WHERE_TOP_LIMIT)==0 ); if( pNew->wsFlags & WHERE_BTM_LIMIT ){ opMask = WO_LT|WO_LE; }else if( pProbe->tnum<=0 || (pSrc->jointype & JT_LEFT)!=0 ){ opMask = WO_EQ|WO_IN|WO_GT|WO_GE|WO_LT|WO_LE; }else{ opMask = WO_EQ|WO_IN|WO_ISNULL|WO_GT|WO_GE|WO_LT|WO_LE; } if( pProbe->bUnordered ) opMask &= ~(WO_GT|WO_GE|WO_LT|WO_LE); assert( pNew->u.btree.nEqnColumn ); iCol = pProbe->aiColumn[pNew->u.btree.nEq]; pTerm = whereScanInit(&scan, pBuilder->pWC, pSrc->iCursor, iCol, opMask, pProbe); saved_nEq = pNew->u.btree.nEq; saved_nSkip = pNew->nSkip; saved_nLTerm = pNew->nLTerm; saved_wsFlags = pNew->wsFlags; saved_prereq = pNew->prereq; saved_nOut = pNew->nOut; pNew->rSetup = 0; rSize = pProbe->aiRowLogEst[0]; rLogSize = estLog(rSize); for(; rc==SQLITE_OK && pTerm!=0; pTerm = whereScanNext(&scan)){ u16 eOp = pTerm->eOperator; /* Shorthand for pTerm->eOperator */ LogEst rCostIdx; LogEst nOutUnadjusted; /* nOut before IN() and WHERE adjustments */ int nIn = 0; #ifdef SQLITE_ENABLE_STAT3_OR_STAT4 int nRecValid = pBuilder->nRecValid; #endif if( (eOp==WO_ISNULL || (pTerm->wtFlags&TERM_VNULL)!=0) && (iCol<0 || pSrc->pTab->aCol[iCol].notNull) ){ continue; /* ignore IS [NOT] NULL constraints on NOT NULL columns */ } if( pTerm->prereqRight & pNew->maskSelf ) continue; pNew->wsFlags = saved_wsFlags; pNew->u.btree.nEq = saved_nEq; pNew->nLTerm = saved_nLTerm; if( whereLoopResize(db, pNew, pNew->nLTerm+1) ) break; /* OOM */ pNew->aLTerm[pNew->nLTerm++] = pTerm; pNew->prereq = (saved_prereq | pTerm->prereqRight) & ~pNew->maskSelf; assert( nInMul==0 || (pNew->wsFlags & WHERE_COLUMN_NULL)!=0 || (pNew->wsFlags & WHERE_COLUMN_IN)!=0 || (pNew->wsFlags & WHERE_SKIPSCAN)!=0 ); if( eOp & WO_IN ){ Expr *pExpr = pTerm->pExpr; pNew->wsFlags |= WHERE_COLUMN_IN; if( ExprHasProperty(pExpr, EP_xIsSelect) ){ /* "x IN (SELECT ...)": TUNING: the SELECT returns 25 rows */ nIn = 46; assert( 46==sqlite3LogEst(25) ); }else if( ALWAYS(pExpr->x.pList && pExpr->x.pList->nExpr) ){ /* "x IN (value, value, ...)" */ nIn = sqlite3LogEst(pExpr->x.pList->nExpr); } assert( nIn>0 ); /* RHS always has 2 or more terms... The parser ** changes "x IN (?)" into "x=?". */ }else if( eOp & (WO_EQ) ){ pNew->wsFlags |= WHERE_COLUMN_EQ; if( iCol<0 || (nInMul==0 && pNew->u.btree.nEq==pProbe->nKeyCol-1) ){ if( iCol>=0 && !IsUniqueIndex(pProbe) ){ pNew->wsFlags |= WHERE_UNQ_WANTED; }else{ pNew->wsFlags |= WHERE_ONEROW; } } }else if( eOp & WO_ISNULL ){ pNew->wsFlags |= WHERE_COLUMN_NULL; }else if( eOp & (WO_GT|WO_GE) ){ testcase( eOp & WO_GT ); testcase( eOp & WO_GE ); pNew->wsFlags |= WHERE_COLUMN_RANGE|WHERE_BTM_LIMIT; pBtm = pTerm; pTop = 0; }else{ assert( eOp & (WO_LT|WO_LE) ); testcase( eOp & WO_LT ); testcase( eOp & WO_LE ); pNew->wsFlags |= WHERE_COLUMN_RANGE|WHERE_TOP_LIMIT; pTop = pTerm; pBtm = (pNew->wsFlags & WHERE_BTM_LIMIT)!=0 ? pNew->aLTerm[pNew->nLTerm-2] : 0; } /* At this point pNew->nOut is set to the number of rows expected to ** be visited by the index scan before considering term pTerm, or the ** values of nIn and nInMul. In other words, assuming that all ** "x IN(...)" terms are replaced with "x = ?". This block updates ** the value of pNew->nOut to account for pTerm (but not nIn/nInMul). */ assert( pNew->nOut==saved_nOut ); if( pNew->wsFlags & WHERE_COLUMN_RANGE ){ /* Adjust nOut using stat3/stat4 data. Or, if there is no stat3/stat4 ** data, using some other estimate. */ whereRangeScanEst(pParse, pBuilder, pBtm, pTop, pNew); }else{ int nEq = ++pNew->u.btree.nEq; assert( eOp & (WO_ISNULL|WO_EQ|WO_IN) ); assert( pNew->nOut==saved_nOut ); if( pTerm->truthProb<=0 && iCol>=0 ){ assert( (eOp & WO_IN) || nIn==0 ); testcase( eOp & WO_IN ); pNew->nOut += pTerm->truthProb; pNew->nOut -= nIn; }else{ #ifdef SQLITE_ENABLE_STAT3_OR_STAT4 tRowcnt nOut = 0; if( nInMul==0 && pProbe->nSample && pNew->u.btree.nEq<=pProbe->nSampleCol && ((eOp & WO_IN)==0 || !ExprHasProperty(pTerm->pExpr, EP_xIsSelect)) ){ Expr *pExpr = pTerm->pExpr; if( (eOp & (WO_EQ|WO_ISNULL))!=0 ){ testcase( eOp & WO_EQ ); testcase( eOp & WO_ISNULL ); rc = whereEqualScanEst(pParse, pBuilder, pExpr->pRight, &nOut); }else{ rc = whereInScanEst(pParse, pBuilder, pExpr->x.pList, &nOut); } if( rc==SQLITE_NOTFOUND ) rc = SQLITE_OK; if( rc!=SQLITE_OK ) break; /* Jump out of the pTerm loop */ if( nOut ){ pNew->nOut = sqlite3LogEst(nOut); if( pNew->nOut>saved_nOut ) pNew->nOut = saved_nOut; pNew->nOut -= nIn; } } if( nOut==0 ) #endif { pNew->nOut += (pProbe->aiRowLogEst[nEq] - pProbe->aiRowLogEst[nEq-1]); if( eOp & WO_ISNULL ){ /* TUNING: If there is no likelihood() value, assume that a ** "col IS NULL" expression matches twice as many rows ** as (col=?). */ pNew->nOut += 10; } } } } /* Set rCostIdx to the cost of visiting selected rows in index. Add ** it to pNew->rRun, which is currently set to the cost of the index ** seek only. Then, if this is a non-covering index, add the cost of ** visiting the rows in the main table. */ rCostIdx = pNew->nOut + 1 + (15*pProbe->szIdxRow)/pSrc->pTab->szTabRow; pNew->rRun = sqlite3LogEstAdd(rLogSize, rCostIdx); if( (pNew->wsFlags & (WHERE_IDX_ONLY|WHERE_IPK))==0 ){ pNew->rRun = sqlite3LogEstAdd(pNew->rRun, pNew->nOut + 16); } ApplyCostMultiplier(pNew->rRun, pProbe->pTable->costMult); nOutUnadjusted = pNew->nOut; pNew->rRun += nInMul + nIn; pNew->nOut += nInMul + nIn; whereLoopOutputAdjust(pBuilder->pWC, pNew, rSize); rc = whereLoopInsert(pBuilder, pNew); if( pNew->wsFlags & WHERE_COLUMN_RANGE ){ pNew->nOut = saved_nOut; }else{ pNew->nOut = nOutUnadjusted; } if( (pNew->wsFlags & WHERE_TOP_LIMIT)==0 && pNew->u.btree.nEqnColumn ){ whereLoopAddBtreeIndex(pBuilder, pSrc, pProbe, nInMul+nIn); } pNew->nOut = saved_nOut; #ifdef SQLITE_ENABLE_STAT3_OR_STAT4 pBuilder->nRecValid = nRecValid; #endif } pNew->prereq = saved_prereq; pNew->u.btree.nEq = saved_nEq; pNew->nSkip = saved_nSkip; pNew->wsFlags = saved_wsFlags; pNew->nOut = saved_nOut; pNew->nLTerm = saved_nLTerm; /* Consider using a skip-scan if there are no WHERE clause constraints ** available for the left-most terms of the index, and if the average ** number of repeats in the left-most terms is at least 18. ** ** The magic number 18 is selected on the basis that scanning 17 rows ** is almost always quicker than an index seek (even though if the index ** contains fewer than 2^17 rows we assume otherwise in other parts of ** the code). And, even if it is not, it should not be too much slower. ** On the other hand, the extra seeks could end up being significantly ** more expensive. */ assert( 42==sqlite3LogEst(18) ); if( saved_nEq==saved_nSkip && saved_nEq+1nKeyCol && pProbe->noSkipScan==0 && pProbe->aiRowLogEst[saved_nEq+1]>=42 /* TUNING: Minimum for skip-scan */ && (rc = whereLoopResize(db, pNew, pNew->nLTerm+1))==SQLITE_OK ){ LogEst nIter; pNew->u.btree.nEq++; pNew->nSkip++; pNew->aLTerm[pNew->nLTerm++] = 0; pNew->wsFlags |= WHERE_SKIPSCAN; nIter = pProbe->aiRowLogEst[saved_nEq] - pProbe->aiRowLogEst[saved_nEq+1]; pNew->nOut -= nIter; /* TUNING: Because uncertainties in the estimates for skip-scan queries, ** add a 1.375 fudge factor to make skip-scan slightly less likely. */ nIter += 5; whereLoopAddBtreeIndex(pBuilder, pSrc, pProbe, nIter + nInMul); pNew->nOut = saved_nOut; pNew->u.btree.nEq = saved_nEq; pNew->nSkip = saved_nSkip; pNew->wsFlags = saved_wsFlags; } return rc; } /* ** Return True if it is possible that pIndex might be useful in ** implementing the ORDER BY clause in pBuilder. ** ** Return False if pBuilder does not contain an ORDER BY clause or ** if there is no way for pIndex to be useful in implementing that ** ORDER BY clause. */ static int indexMightHelpWithOrderBy( WhereLoopBuilder *pBuilder, Index *pIndex, int iCursor ){ ExprList *pOB; int ii, jj; if( pIndex->bUnordered ) return 0; if( (pOB = pBuilder->pWInfo->pOrderBy)==0 ) return 0; for(ii=0; iinExpr; ii++){ Expr *pExpr = sqlite3ExprSkipCollate(pOB->a[ii].pExpr); if( pExpr->op!=TK_COLUMN ) return 0; if( pExpr->iTable==iCursor ){ if( pExpr->iColumn<0 ) return 1; for(jj=0; jjnKeyCol; jj++){ if( pExpr->iColumn==pIndex->aiColumn[jj] ) return 1; } } } return 0; } /* ** Return a bitmask where 1s indicate that the corresponding column of ** the table is used by an index. Only the first 63 columns are considered. */ static Bitmask columnsInIndex(Index *pIdx){ Bitmask m = 0; int j; for(j=pIdx->nColumn-1; j>=0; j--){ int x = pIdx->aiColumn[j]; if( x>=0 ){ testcase( x==BMS-1 ); testcase( x==BMS-2 ); if( xa; inTerm; i++, pTerm++){ if( sqlite3ExprImpliesExpr(pTerm->pExpr, pWhere, iTab) ) return 1; } return 0; } /* ** Add all WhereLoop objects for a single table of the join where the table ** is idenfied by pBuilder->pNew->iTab. That table is guaranteed to be ** a b-tree table, not a virtual table. ** ** The costs (WhereLoop.rRun) of the b-tree loops added by this function ** are calculated as follows: ** ** For a full scan, assuming the table (or index) contains nRow rows: ** ** cost = nRow * 3.0 // full-table scan ** cost = nRow * K // scan of covering index ** cost = nRow * (K+3.0) // scan of non-covering index ** ** where K is a value between 1.1 and 3.0 set based on the relative ** estimated average size of the index and table records. ** ** For an index scan, where nVisit is the number of index rows visited ** by the scan, and nSeek is the number of seek operations required on ** the index b-tree: ** ** cost = nSeek * (log(nRow) + K * nVisit) // covering index ** cost = nSeek * (log(nRow) + (K+3.0) * nVisit) // non-covering index ** ** Normally, nSeek is 1. nSeek values greater than 1 come about if the ** WHERE clause includes "x IN (....)" terms used in place of "x=?". Or when ** implicit "x IN (SELECT x FROM tbl)" terms are added for skip-scans. ** ** The estimated values (nRow, nVisit, nSeek) often contain a large amount ** of uncertainty. For this reason, scoring is designed to pick plans that ** "do the least harm" if the estimates are inaccurate. For example, a ** log(nRow) factor is omitted from a non-covering index scan in order to ** bias the scoring in favor of using an index, since the worst-case ** performance of using an index is far better than the worst-case performance ** of a full table scan. */ static int whereLoopAddBtree( WhereLoopBuilder *pBuilder, /* WHERE clause information */ Bitmask mExtra /* Extra prerequesites for using this table */ ){ WhereInfo *pWInfo; /* WHERE analysis context */ Index *pProbe; /* An index we are evaluating */ Index sPk; /* A fake index object for the primary key */ LogEst aiRowEstPk[2]; /* The aiRowLogEst[] value for the sPk index */ i16 aiColumnPk = -1; /* The aColumn[] value for the sPk index */ SrcList *pTabList; /* The FROM clause */ struct SrcList_item *pSrc; /* The FROM clause btree term to add */ WhereLoop *pNew; /* Template WhereLoop object */ int rc = SQLITE_OK; /* Return code */ int iSortIdx = 1; /* Index number */ int b; /* A boolean value */ LogEst rSize; /* number of rows in the table */ LogEst rLogSize; /* Logarithm of the number of rows in the table */ WhereClause *pWC; /* The parsed WHERE clause */ Table *pTab; /* Table being queried */ pNew = pBuilder->pNew; pWInfo = pBuilder->pWInfo; pTabList = pWInfo->pTabList; pSrc = pTabList->a + pNew->iTab; pTab = pSrc->pTab; pWC = pBuilder->pWC; assert( !IsVirtual(pSrc->pTab) ); if( pSrc->pIndex ){ /* An INDEXED BY clause specifies a particular index to use */ pProbe = pSrc->pIndex; }else if( !HasRowid(pTab) ){ pProbe = pTab->pIndex; }else{ /* There is no INDEXED BY clause. Create a fake Index object in local ** variable sPk to represent the rowid primary key index. Make this ** fake index the first in a chain of Index objects with all of the real ** indices to follow */ Index *pFirst; /* First of real indices on the table */ memset(&sPk, 0, sizeof(Index)); sPk.nKeyCol = 1; sPk.nColumn = 1; sPk.aiColumn = &aiColumnPk; sPk.aiRowLogEst = aiRowEstPk; sPk.onError = OE_Replace; sPk.pTable = pTab; sPk.szIdxRow = pTab->szTabRow; aiRowEstPk[0] = pTab->nRowLogEst; aiRowEstPk[1] = 0; pFirst = pSrc->pTab->pIndex; if( pSrc->notIndexed==0 ){ /* The real indices of the table are only considered if the ** NOT INDEXED qualifier is omitted from the FROM clause */ sPk.pNext = pFirst; } pProbe = &sPk; } rSize = pTab->nRowLogEst; rLogSize = estLog(rSize); #ifndef SQLITE_OMIT_AUTOMATIC_INDEX /* Automatic indexes */ if( !pBuilder->pOrSet && (pWInfo->pParse->db->flags & SQLITE_AutoIndex)!=0 && pSrc->pIndex==0 && !pSrc->viaCoroutine && !pSrc->notIndexed && HasRowid(pTab) && !pSrc->isCorrelated && !pSrc->isRecursive ){ /* Generate auto-index WhereLoops */ WhereTerm *pTerm; WhereTerm *pWCEnd = pWC->a + pWC->nTerm; for(pTerm=pWC->a; rc==SQLITE_OK && pTermprereqRight & pNew->maskSelf ) continue; if( termCanDriveIndex(pTerm, pSrc, 0) ){ pNew->u.btree.nEq = 1; pNew->nSkip = 0; pNew->u.btree.pIndex = 0; pNew->nLTerm = 1; pNew->aLTerm[0] = pTerm; /* TUNING: One-time cost for computing the automatic index is ** estimated to be X*N*log2(N) where N is the number of rows in ** the table being indexed and where X is 7 (LogEst=28) for normal ** tables or 1.375 (LogEst=4) for views and subqueries. The value ** of X is smaller for views and subqueries so that the query planner ** will be more aggressive about generating automatic indexes for ** those objects, since there is no opportunity to add schema ** indexes on subqueries and views. */ pNew->rSetup = rLogSize + rSize + 4; if( pTab->pSelect==0 && (pTab->tabFlags & TF_Ephemeral)==0 ){ pNew->rSetup += 24; } ApplyCostMultiplier(pNew->rSetup, pTab->costMult); /* TUNING: Each index lookup yields 20 rows in the table. This ** is more than the usual guess of 10 rows, since we have no way ** of knowing how selective the index will ultimately be. It would ** not be unreasonable to make this value much larger. */ pNew->nOut = 43; assert( 43==sqlite3LogEst(20) ); pNew->rRun = sqlite3LogEstAdd(rLogSize,pNew->nOut); pNew->wsFlags = WHERE_AUTO_INDEX; pNew->prereq = mExtra | pTerm->prereqRight; rc = whereLoopInsert(pBuilder, pNew); } } } #endif /* SQLITE_OMIT_AUTOMATIC_INDEX */ /* Loop over all indices */ for(; rc==SQLITE_OK && pProbe; pProbe=pProbe->pNext, iSortIdx++){ if( pProbe->pPartIdxWhere!=0 && !whereUsablePartialIndex(pSrc->iCursor, pWC, pProbe->pPartIdxWhere) ){ testcase( pNew->iTab!=pSrc->iCursor ); /* See ticket [98d973b8f5] */ continue; /* Partial index inappropriate for this query */ } rSize = pProbe->aiRowLogEst[0]; pNew->u.btree.nEq = 0; pNew->nSkip = 0; pNew->nLTerm = 0; pNew->iSortIdx = 0; pNew->rSetup = 0; pNew->prereq = mExtra; pNew->nOut = rSize; pNew->u.btree.pIndex = pProbe; b = indexMightHelpWithOrderBy(pBuilder, pProbe, pSrc->iCursor); /* The ONEPASS_DESIRED flags never occurs together with ORDER BY */ assert( (pWInfo->wctrlFlags & WHERE_ONEPASS_DESIRED)==0 || b==0 ); if( pProbe->tnum<=0 ){ /* Integer primary key index */ pNew->wsFlags = WHERE_IPK; /* Full table scan */ pNew->iSortIdx = b ? iSortIdx : 0; /* TUNING: Cost of full table scan is (N*3.0). */ pNew->rRun = rSize + 16; ApplyCostMultiplier(pNew->rRun, pTab->costMult); whereLoopOutputAdjust(pWC, pNew, rSize); rc = whereLoopInsert(pBuilder, pNew); pNew->nOut = rSize; if( rc ) break; }else{ Bitmask m; if( pProbe->isCovering ){ pNew->wsFlags = WHERE_IDX_ONLY | WHERE_INDEXED; m = 0; }else{ m = pSrc->colUsed & ~columnsInIndex(pProbe); pNew->wsFlags = (m==0) ? (WHERE_IDX_ONLY|WHERE_INDEXED) : WHERE_INDEXED; } /* Full scan via index */ if( b || !HasRowid(pTab) || ( m==0 && pProbe->bUnordered==0 && (pProbe->szIdxRowszTabRow) && (pWInfo->wctrlFlags & WHERE_ONEPASS_DESIRED)==0 && sqlite3GlobalConfig.bUseCis && OptimizationEnabled(pWInfo->pParse->db, SQLITE_CoverIdxScan) ) ){ pNew->iSortIdx = b ? iSortIdx : 0; /* The cost of visiting the index rows is N*K, where K is ** between 1.1 and 3.0, depending on the relative sizes of the ** index and table rows. If this is a non-covering index scan, ** also add the cost of visiting table rows (N*3.0). */ pNew->rRun = rSize + 1 + (15*pProbe->szIdxRow)/pTab->szTabRow; if( m!=0 ){ pNew->rRun = sqlite3LogEstAdd(pNew->rRun, rSize+16); } ApplyCostMultiplier(pNew->rRun, pTab->costMult); whereLoopOutputAdjust(pWC, pNew, rSize); rc = whereLoopInsert(pBuilder, pNew); pNew->nOut = rSize; if( rc ) break; } } rc = whereLoopAddBtreeIndex(pBuilder, pSrc, pProbe, 0); #ifdef SQLITE_ENABLE_STAT3_OR_STAT4 sqlite3Stat4ProbeFree(pBuilder->pRec); pBuilder->nRecValid = 0; pBuilder->pRec = 0; #endif /* If there was an INDEXED BY clause, then only that one index is ** considered. */ if( pSrc->pIndex ) break; } return rc; } #ifndef SQLITE_OMIT_VIRTUALTABLE /* ** Add all WhereLoop objects for a table of the join identified by ** pBuilder->pNew->iTab. That table is guaranteed to be a virtual table. */ static int whereLoopAddVirtual( WhereLoopBuilder *pBuilder, /* WHERE clause information */ Bitmask mExtra ){ WhereInfo *pWInfo; /* WHERE analysis context */ Parse *pParse; /* The parsing context */ WhereClause *pWC; /* The WHERE clause */ struct SrcList_item *pSrc; /* The FROM clause term to search */ Table *pTab; sqlite3 *db; sqlite3_index_info *pIdxInfo; struct sqlite3_index_constraint *pIdxCons; struct sqlite3_index_constraint_usage *pUsage; WhereTerm *pTerm; int i, j; int iTerm, mxTerm; int nConstraint; int seenIn = 0; /* True if an IN operator is seen */ int seenVar = 0; /* True if a non-constant constraint is seen */ int iPhase; /* 0: const w/o IN, 1: const, 2: no IN, 2: IN */ WhereLoop *pNew; int rc = SQLITE_OK; pWInfo = pBuilder->pWInfo; pParse = pWInfo->pParse; db = pParse->db; pWC = pBuilder->pWC; pNew = pBuilder->pNew; pSrc = &pWInfo->pTabList->a[pNew->iTab]; pTab = pSrc->pTab; assert( IsVirtual(pTab) ); pIdxInfo = allocateIndexInfo(pParse, pWC, pSrc, pBuilder->pOrderBy); if( pIdxInfo==0 ) return SQLITE_NOMEM; pNew->prereq = 0; pNew->rSetup = 0; pNew->wsFlags = WHERE_VIRTUALTABLE; pNew->nLTerm = 0; pNew->u.vtab.needFree = 0; pUsage = pIdxInfo->aConstraintUsage; nConstraint = pIdxInfo->nConstraint; if( whereLoopResize(db, pNew, nConstraint) ){ sqlite3DbFree(db, pIdxInfo); return SQLITE_NOMEM; } for(iPhase=0; iPhase<=3; iPhase++){ if( !seenIn && (iPhase&1)!=0 ){ iPhase++; if( iPhase>3 ) break; } if( !seenVar && iPhase>1 ) break; pIdxCons = *(struct sqlite3_index_constraint**)&pIdxInfo->aConstraint; for(i=0; inConstraint; i++, pIdxCons++){ j = pIdxCons->iTermOffset; pTerm = &pWC->a[j]; switch( iPhase ){ case 0: /* Constants without IN operator */ pIdxCons->usable = 0; if( (pTerm->eOperator & WO_IN)!=0 ){ seenIn = 1; } if( pTerm->prereqRight!=0 ){ seenVar = 1; }else if( (pTerm->eOperator & WO_IN)==0 ){ pIdxCons->usable = 1; } break; case 1: /* Constants with IN operators */ assert( seenIn ); pIdxCons->usable = (pTerm->prereqRight==0); break; case 2: /* Variables without IN */ assert( seenVar ); pIdxCons->usable = (pTerm->eOperator & WO_IN)==0; break; default: /* Variables with IN */ assert( seenVar && seenIn ); pIdxCons->usable = 1; break; } } memset(pUsage, 0, sizeof(pUsage[0])*pIdxInfo->nConstraint); if( pIdxInfo->needToFreeIdxStr ) sqlite3_free(pIdxInfo->idxStr); pIdxInfo->idxStr = 0; pIdxInfo->idxNum = 0; pIdxInfo->needToFreeIdxStr = 0; pIdxInfo->orderByConsumed = 0; pIdxInfo->estimatedCost = SQLITE_BIG_DBL / (double)2; pIdxInfo->estimatedRows = 25; rc = vtabBestIndex(pParse, pTab, pIdxInfo); if( rc ) goto whereLoopAddVtab_exit; pIdxCons = *(struct sqlite3_index_constraint**)&pIdxInfo->aConstraint; pNew->prereq = mExtra; mxTerm = -1; assert( pNew->nLSlot>=nConstraint ); for(i=0; iaLTerm[i] = 0; pNew->u.vtab.omitMask = 0; for(i=0; i=0 ){ j = pIdxCons->iTermOffset; if( iTerm>=nConstraint || j<0 || j>=pWC->nTerm || pNew->aLTerm[iTerm]!=0 ){ rc = SQLITE_ERROR; sqlite3ErrorMsg(pParse, "%s.xBestIndex() malfunction", pTab->zName); goto whereLoopAddVtab_exit; } testcase( iTerm==nConstraint-1 ); testcase( j==0 ); testcase( j==pWC->nTerm-1 ); pTerm = &pWC->a[j]; pNew->prereq |= pTerm->prereqRight; assert( iTermnLSlot ); pNew->aLTerm[iTerm] = pTerm; if( iTerm>mxTerm ) mxTerm = iTerm; testcase( iTerm==15 ); testcase( iTerm==16 ); if( iTerm<16 && pUsage[i].omit ) pNew->u.vtab.omitMask |= 1<eOperator & WO_IN)!=0 ){ if( pUsage[i].omit==0 ){ /* Do not attempt to use an IN constraint if the virtual table ** says that the equivalent EQ constraint cannot be safely omitted. ** If we do attempt to use such a constraint, some rows might be ** repeated in the output. */ break; } /* A virtual table that is constrained by an IN clause may not ** consume the ORDER BY clause because (1) the order of IN terms ** is not necessarily related to the order of output terms and ** (2) Multiple outputs from a single IN value will not merge ** together. */ pIdxInfo->orderByConsumed = 0; } } } if( i>=nConstraint ){ pNew->nLTerm = mxTerm+1; assert( pNew->nLTerm<=pNew->nLSlot ); pNew->u.vtab.idxNum = pIdxInfo->idxNum; pNew->u.vtab.needFree = pIdxInfo->needToFreeIdxStr; pIdxInfo->needToFreeIdxStr = 0; pNew->u.vtab.idxStr = pIdxInfo->idxStr; pNew->u.vtab.isOrdered = (i8)(pIdxInfo->orderByConsumed ? pIdxInfo->nOrderBy : 0); pNew->rSetup = 0; pNew->rRun = sqlite3LogEstFromDouble(pIdxInfo->estimatedCost); pNew->nOut = sqlite3LogEst(pIdxInfo->estimatedRows); whereLoopInsert(pBuilder, pNew); if( pNew->u.vtab.needFree ){ sqlite3_free(pNew->u.vtab.idxStr); pNew->u.vtab.needFree = 0; } } } whereLoopAddVtab_exit: if( pIdxInfo->needToFreeIdxStr ) sqlite3_free(pIdxInfo->idxStr); sqlite3DbFree(db, pIdxInfo); return rc; } #endif /* SQLITE_OMIT_VIRTUALTABLE */ /* ** Add WhereLoop entries to handle OR terms. This works for either ** btrees or virtual tables. */ static int whereLoopAddOr(WhereLoopBuilder *pBuilder, Bitmask mExtra){ WhereInfo *pWInfo = pBuilder->pWInfo; WhereClause *pWC; WhereLoop *pNew; WhereTerm *pTerm, *pWCEnd; int rc = SQLITE_OK; int iCur; WhereClause tempWC; WhereLoopBuilder sSubBuild; WhereOrSet sSum, sCur; struct SrcList_item *pItem; pWC = pBuilder->pWC; pWCEnd = pWC->a + pWC->nTerm; pNew = pBuilder->pNew; memset(&sSum, 0, sizeof(sSum)); pItem = pWInfo->pTabList->a + pNew->iTab; iCur = pItem->iCursor; for(pTerm=pWC->a; pTermeOperator & WO_OR)!=0 && (pTerm->u.pOrInfo->indexable & pNew->maskSelf)!=0 ){ WhereClause * const pOrWC = &pTerm->u.pOrInfo->wc; WhereTerm * const pOrWCEnd = &pOrWC->a[pOrWC->nTerm]; WhereTerm *pOrTerm; int once = 1; int i, j; sSubBuild = *pBuilder; sSubBuild.pOrderBy = 0; sSubBuild.pOrSet = &sCur; WHERETRACE(0x200, ("Begin processing OR-clause %p\n", pTerm)); for(pOrTerm=pOrWC->a; pOrTermeOperator & WO_AND)!=0 ){ sSubBuild.pWC = &pOrTerm->u.pAndInfo->wc; }else if( pOrTerm->leftCursor==iCur ){ tempWC.pWInfo = pWC->pWInfo; tempWC.pOuter = pWC; tempWC.op = TK_AND; tempWC.nTerm = 1; tempWC.a = pOrTerm; sSubBuild.pWC = &tempWC; }else{ continue; } sCur.n = 0; #ifdef WHERETRACE_ENABLED WHERETRACE(0x200, ("OR-term %d of %p has %d subterms:\n", (int)(pOrTerm-pOrWC->a), pTerm, sSubBuild.pWC->nTerm)); if( sqlite3WhereTrace & 0x400 ){ for(i=0; inTerm; i++){ whereTermPrint(&sSubBuild.pWC->a[i], i); } } #endif #ifndef SQLITE_OMIT_VIRTUALTABLE if( IsVirtual(pItem->pTab) ){ rc = whereLoopAddVirtual(&sSubBuild, mExtra); }else #endif { rc = whereLoopAddBtree(&sSubBuild, mExtra); } if( rc==SQLITE_OK ){ rc = whereLoopAddOr(&sSubBuild, mExtra); } assert( rc==SQLITE_OK || sCur.n==0 ); if( sCur.n==0 ){ sSum.n = 0; break; }else if( once ){ whereOrMove(&sSum, &sCur); once = 0; }else{ WhereOrSet sPrev; whereOrMove(&sPrev, &sSum); sSum.n = 0; for(i=0; inLTerm = 1; pNew->aLTerm[0] = pTerm; pNew->wsFlags = WHERE_MULTI_OR; pNew->rSetup = 0; pNew->iSortIdx = 0; memset(&pNew->u, 0, sizeof(pNew->u)); for(i=0; rc==SQLITE_OK && irRun = sSum.a[i].rRun + 1; pNew->nOut = sSum.a[i].nOut; pNew->prereq = sSum.a[i].prereq; rc = whereLoopInsert(pBuilder, pNew); } WHERETRACE(0x200, ("End processing OR-clause %p\n", pTerm)); } } return rc; } /* ** Add all WhereLoop objects for all tables */ static int whereLoopAddAll(WhereLoopBuilder *pBuilder){ WhereInfo *pWInfo = pBuilder->pWInfo; Bitmask mExtra = 0; Bitmask mPrior = 0; int iTab; SrcList *pTabList = pWInfo->pTabList; struct SrcList_item *pItem; sqlite3 *db = pWInfo->pParse->db; int nTabList = pWInfo->nLevel; int rc = SQLITE_OK; u8 priorJoinType = 0; WhereLoop *pNew; /* Loop over the tables in the join, from left to right */ pNew = pBuilder->pNew; whereLoopInit(pNew); for(iTab=0, pItem=pTabList->a; iTabiTab = iTab; pNew->maskSelf = getMask(&pWInfo->sMaskSet, pItem->iCursor); if( ((pItem->jointype|priorJoinType) & (JT_LEFT|JT_CROSS))!=0 ){ mExtra = mPrior; } priorJoinType = pItem->jointype; if( IsVirtual(pItem->pTab) ){ rc = whereLoopAddVirtual(pBuilder, mExtra); }else{ rc = whereLoopAddBtree(pBuilder, mExtra); } if( rc==SQLITE_OK ){ rc = whereLoopAddOr(pBuilder, mExtra); } mPrior |= pNew->maskSelf; if( rc || db->mallocFailed ) break; } whereLoopClear(db, pNew); return rc; } /* ** Examine a WherePath (with the addition of the extra WhereLoop of the 5th ** parameters) to see if it outputs rows in the requested ORDER BY ** (or GROUP BY) without requiring a separate sort operation. Return N: ** ** N>0: N terms of the ORDER BY clause are satisfied ** N==0: No terms of the ORDER BY clause are satisfied ** N<0: Unknown yet how many terms of ORDER BY might be satisfied. ** ** Note that processing for WHERE_GROUPBY and WHERE_DISTINCTBY is not as ** strict. With GROUP BY and DISTINCT the only requirement is that ** equivalent rows appear immediately adjacent to one another. GROUP BY ** and DISTINCT do not require rows to appear in any particular order as long ** as equivalent rows are grouped together. Thus for GROUP BY and DISTINCT ** the pOrderBy terms can be matched in any order. With ORDER BY, the ** pOrderBy terms must be matched in strict left-to-right order. */ static i8 wherePathSatisfiesOrderBy( WhereInfo *pWInfo, /* The WHERE clause */ ExprList *pOrderBy, /* ORDER BY or GROUP BY or DISTINCT clause to check */ WherePath *pPath, /* The WherePath to check */ u16 wctrlFlags, /* Might contain WHERE_GROUPBY or WHERE_DISTINCTBY */ u16 nLoop, /* Number of entries in pPath->aLoop[] */ WhereLoop *pLast, /* Add this WhereLoop to the end of pPath->aLoop[] */ Bitmask *pRevMask /* OUT: Mask of WhereLoops to run in reverse order */ ){ u8 revSet; /* True if rev is known */ u8 rev; /* Composite sort order */ u8 revIdx; /* Index sort order */ u8 isOrderDistinct; /* All prior WhereLoops are order-distinct */ u8 distinctColumns; /* True if the loop has UNIQUE NOT NULL columns */ u8 isMatch; /* iColumn matches a term of the ORDER BY clause */ u16 nKeyCol; /* Number of key columns in pIndex */ u16 nColumn; /* Total number of ordered columns in the index */ u16 nOrderBy; /* Number terms in the ORDER BY clause */ int iLoop; /* Index of WhereLoop in pPath being processed */ int i, j; /* Loop counters */ int iCur; /* Cursor number for current WhereLoop */ int iColumn; /* A column number within table iCur */ WhereLoop *pLoop = 0; /* Current WhereLoop being processed. */ WhereTerm *pTerm; /* A single term of the WHERE clause */ Expr *pOBExpr; /* An expression from the ORDER BY clause */ CollSeq *pColl; /* COLLATE function from an ORDER BY clause term */ Index *pIndex; /* The index associated with pLoop */ sqlite3 *db = pWInfo->pParse->db; /* Database connection */ Bitmask obSat = 0; /* Mask of ORDER BY terms satisfied so far */ Bitmask obDone; /* Mask of all ORDER BY terms */ Bitmask orderDistinctMask; /* Mask of all well-ordered loops */ Bitmask ready; /* Mask of inner loops */ /* ** We say the WhereLoop is "one-row" if it generates no more than one ** row of output. A WhereLoop is one-row if all of the following are true: ** (a) All index columns match with WHERE_COLUMN_EQ. ** (b) The index is unique ** Any WhereLoop with an WHERE_COLUMN_EQ constraint on the rowid is one-row. ** Every one-row WhereLoop will have the WHERE_ONEROW bit set in wsFlags. ** ** We say the WhereLoop is "order-distinct" if the set of columns from ** that WhereLoop that are in the ORDER BY clause are different for every ** row of the WhereLoop. Every one-row WhereLoop is automatically ** order-distinct. A WhereLoop that has no columns in the ORDER BY clause ** is not order-distinct. To be order-distinct is not quite the same as being ** UNIQUE since a UNIQUE column or index can have multiple rows that ** are NULL and NULL values are equivalent for the purpose of order-distinct. ** To be order-distinct, the columns must be UNIQUE and NOT NULL. ** ** The rowid for a table is always UNIQUE and NOT NULL so whenever the ** rowid appears in the ORDER BY clause, the corresponding WhereLoop is ** automatically order-distinct. */ assert( pOrderBy!=0 ); if( nLoop && OptimizationDisabled(db, SQLITE_OrderByIdxJoin) ) return 0; nOrderBy = pOrderBy->nExpr; testcase( nOrderBy==BMS-1 ); if( nOrderBy>BMS-1 ) return 0; /* Cannot optimize overly large ORDER BYs */ isOrderDistinct = 1; obDone = MASKBIT(nOrderBy)-1; orderDistinctMask = 0; ready = 0; for(iLoop=0; isOrderDistinct && obSat0 ) ready |= pLoop->maskSelf; pLoop = iLoopaLoop[iLoop] : pLast; if( pLoop->wsFlags & WHERE_VIRTUALTABLE ){ if( pLoop->u.vtab.isOrdered ) obSat = obDone; break; } iCur = pWInfo->pTabList->a[pLoop->iTab].iCursor; /* Mark off any ORDER BY term X that is a column in the table of ** the current loop for which there is term in the WHERE ** clause of the form X IS NULL or X=? that reference only outer ** loops. */ for(i=0; ia[i].pExpr); if( pOBExpr->op!=TK_COLUMN ) continue; if( pOBExpr->iTable!=iCur ) continue; pTerm = findTerm(&pWInfo->sWC, iCur, pOBExpr->iColumn, ~ready, WO_EQ|WO_ISNULL, 0); if( pTerm==0 ) continue; if( (pTerm->eOperator&WO_EQ)!=0 && pOBExpr->iColumn>=0 ){ const char *z1, *z2; pColl = sqlite3ExprCollSeq(pWInfo->pParse, pOrderBy->a[i].pExpr); if( !pColl ) pColl = db->pDfltColl; z1 = pColl->zName; pColl = sqlite3ExprCollSeq(pWInfo->pParse, pTerm->pExpr); if( !pColl ) pColl = db->pDfltColl; z2 = pColl->zName; if( sqlite3StrICmp(z1, z2)!=0 ) continue; } obSat |= MASKBIT(i); } if( (pLoop->wsFlags & WHERE_ONEROW)==0 ){ if( pLoop->wsFlags & WHERE_IPK ){ pIndex = 0; nKeyCol = 0; nColumn = 1; }else if( (pIndex = pLoop->u.btree.pIndex)==0 || pIndex->bUnordered ){ return 0; }else{ nKeyCol = pIndex->nKeyCol; nColumn = pIndex->nColumn; assert( nColumn==nKeyCol+1 || !HasRowid(pIndex->pTable) ); assert( pIndex->aiColumn[nColumn-1]==(-1) || !HasRowid(pIndex->pTable)); isOrderDistinct = IsUniqueIndex(pIndex); } /* Loop through all columns of the index and deal with the ones ** that are not constrained by == or IN. */ rev = revSet = 0; distinctColumns = 0; for(j=0; ju.btree.nEq && pLoop->nSkip==0 && ((i = pLoop->aLTerm[j]->eOperator) & (WO_EQ|WO_ISNULL))!=0 ){ if( i & WO_ISNULL ){ testcase( isOrderDistinct ); isOrderDistinct = 0; } continue; } /* Get the column number in the table (iColumn) and sort order ** (revIdx) for the j-th column of the index. */ if( pIndex ){ iColumn = pIndex->aiColumn[j]; revIdx = pIndex->aSortOrder[j]; if( iColumn==pIndex->pTable->iPKey ) iColumn = -1; }else{ iColumn = -1; revIdx = 0; } /* An unconstrained column that might be NULL means that this ** WhereLoop is not well-ordered */ if( isOrderDistinct && iColumn>=0 && j>=pLoop->u.btree.nEq && pIndex->pTable->aCol[iColumn].notNull==0 ){ isOrderDistinct = 0; } /* Find the ORDER BY term that corresponds to the j-th column ** of the index and mark that ORDER BY term off */ bOnce = 1; isMatch = 0; for(i=0; bOnce && ia[i].pExpr); testcase( wctrlFlags & WHERE_GROUPBY ); testcase( wctrlFlags & WHERE_DISTINCTBY ); if( (wctrlFlags & (WHERE_GROUPBY|WHERE_DISTINCTBY))==0 ) bOnce = 0; if( pOBExpr->op!=TK_COLUMN ) continue; if( pOBExpr->iTable!=iCur ) continue; if( pOBExpr->iColumn!=iColumn ) continue; if( iColumn>=0 ){ pColl = sqlite3ExprCollSeq(pWInfo->pParse, pOrderBy->a[i].pExpr); if( !pColl ) pColl = db->pDfltColl; if( sqlite3StrICmp(pColl->zName, pIndex->azColl[j])!=0 ) continue; } isMatch = 1; break; } if( isMatch && (wctrlFlags & WHERE_GROUPBY)==0 ){ /* Make sure the sort order is compatible in an ORDER BY clause. ** Sort order is irrelevant for a GROUP BY clause. */ if( revSet ){ if( (rev ^ revIdx)!=pOrderBy->a[i].sortOrder ) isMatch = 0; }else{ rev = revIdx ^ pOrderBy->a[i].sortOrder; if( rev ) *pRevMask |= MASKBIT(iLoop); revSet = 1; } } if( isMatch ){ if( iColumn<0 ){ testcase( distinctColumns==0 ); distinctColumns = 1; } obSat |= MASKBIT(i); }else{ /* No match found */ if( j==0 || jmaskSelf; for(i=0; ia[i].pExpr; mTerm = exprTableUsage(&pWInfo->sMaskSet,p); if( mTerm==0 && !sqlite3ExprIsConstant(p) ) continue; if( (mTerm&~orderDistinctMask)==0 ){ obSat |= MASKBIT(i); } } } } /* End the loop over all WhereLoops from outer-most down to inner-most */ if( obSat==obDone ) return (i8)nOrderBy; if( !isOrderDistinct ){ for(i=nOrderBy-1; i>0; i--){ Bitmask m = MASKBIT(i) - 1; if( (obSat&m)==m ) return i; } return 0; } return -1; } /* ** If the WHERE_GROUPBY flag is set in the mask passed to sqlite3WhereBegin(), ** the planner assumes that the specified pOrderBy list is actually a GROUP ** BY clause - and so any order that groups rows as required satisfies the ** request. ** ** Normally, in this case it is not possible for the caller to determine ** whether or not the rows are really being delivered in sorted order, or ** just in some other order that provides the required grouping. However, ** if the WHERE_SORTBYGROUP flag is also passed to sqlite3WhereBegin(), then ** this function may be called on the returned WhereInfo object. It returns ** true if the rows really will be sorted in the specified order, or false ** otherwise. ** ** For example, assuming: ** ** CREATE INDEX i1 ON t1(x, Y); ** ** then ** ** SELECT * FROM t1 GROUP BY x,y ORDER BY x,y; -- IsSorted()==1 ** SELECT * FROM t1 GROUP BY y,x ORDER BY y,x; -- IsSorted()==0 */ int sqlite3WhereIsSorted(WhereInfo *pWInfo){ assert( pWInfo->wctrlFlags & WHERE_GROUPBY ); assert( pWInfo->wctrlFlags & WHERE_SORTBYGROUP ); return pWInfo->sorted; } #ifdef WHERETRACE_ENABLED /* For debugging use only: */ static const char *wherePathName(WherePath *pPath, int nLoop, WhereLoop *pLast){ static char zName[65]; int i; for(i=0; iaLoop[i]->cId; } if( pLast ) zName[i++] = pLast->cId; zName[i] = 0; return zName; } #endif /* ** Return the cost of sorting nRow rows, assuming that the keys have ** nOrderby columns and that the first nSorted columns are already in ** order. */ static LogEst whereSortingCost( WhereInfo *pWInfo, LogEst nRow, int nOrderBy, int nSorted ){ /* TUNING: Estimated cost of a full external sort, where N is ** the number of rows to sort is: ** ** cost = (3.0 * N * log(N)). ** ** Or, if the order-by clause has X terms but only the last Y ** terms are out of order, then block-sorting will reduce the ** sorting cost to: ** ** cost = (3.0 * N * log(N)) * (Y/X) ** ** The (Y/X) term is implemented using stack variable rScale ** below. */ LogEst rScale, rSortCost; assert( nOrderBy>0 && 66==sqlite3LogEst(100) ); rScale = sqlite3LogEst((nOrderBy-nSorted)*100/nOrderBy) - 66; rSortCost = nRow + estLog(nRow) + rScale + 16; /* TUNING: The cost of implementing DISTINCT using a B-TREE is ** similar but with a larger constant of proportionality. ** Multiply by an additional factor of 3.0. */ if( pWInfo->wctrlFlags & WHERE_WANT_DISTINCT ){ rSortCost += 16; } return rSortCost; } /* ** Given the list of WhereLoop objects at pWInfo->pLoops, this routine ** attempts to find the lowest cost path that visits each WhereLoop ** once. This path is then loaded into the pWInfo->a[].pWLoop fields. ** ** Assume that the total number of output rows that will need to be sorted ** will be nRowEst (in the 10*log2 representation). Or, ignore sorting ** costs if nRowEst==0. ** ** Return SQLITE_OK on success or SQLITE_NOMEM of a memory allocation ** error occurs. */ static int wherePathSolver(WhereInfo *pWInfo, LogEst nRowEst){ int mxChoice; /* Maximum number of simultaneous paths tracked */ int nLoop; /* Number of terms in the join */ Parse *pParse; /* Parsing context */ sqlite3 *db; /* The database connection */ int iLoop; /* Loop counter over the terms of the join */ int ii, jj; /* Loop counters */ int mxI = 0; /* Index of next entry to replace */ int nOrderBy; /* Number of ORDER BY clause terms */ LogEst mxCost = 0; /* Maximum cost of a set of paths */ LogEst mxUnsorted = 0; /* Maximum unsorted cost of a set of path */ int nTo, nFrom; /* Number of valid entries in aTo[] and aFrom[] */ WherePath *aFrom; /* All nFrom paths at the previous level */ WherePath *aTo; /* The nTo best paths at the current level */ WherePath *pFrom; /* An element of aFrom[] that we are working on */ WherePath *pTo; /* An element of aTo[] that we are working on */ WhereLoop *pWLoop; /* One of the WhereLoop objects */ WhereLoop **pX; /* Used to divy up the pSpace memory */ LogEst *aSortCost = 0; /* Sorting and partial sorting costs */ char *pSpace; /* Temporary memory used by this routine */ int nSpace; /* Bytes of space allocated at pSpace */ pParse = pWInfo->pParse; db = pParse->db; nLoop = pWInfo->nLevel; /* TUNING: For simple queries, only the best path is tracked. ** For 2-way joins, the 5 best paths are followed. ** For joins of 3 or more tables, track the 10 best paths */ mxChoice = (nLoop<=1) ? 1 : (nLoop==2 ? 5 : 10); assert( nLoop<=pWInfo->pTabList->nSrc ); WHERETRACE(0x002, ("---- begin solver. (nRowEst=%d)\n", nRowEst)); /* If nRowEst is zero and there is an ORDER BY clause, ignore it. In this ** case the purpose of this call is to estimate the number of rows returned ** by the overall query. Once this estimate has been obtained, the caller ** will invoke this function a second time, passing the estimate as the ** nRowEst parameter. */ if( pWInfo->pOrderBy==0 || nRowEst==0 ){ nOrderBy = 0; }else{ nOrderBy = pWInfo->pOrderBy->nExpr; } /* Allocate and initialize space for aTo, aFrom and aSortCost[] */ nSpace = (sizeof(WherePath)+sizeof(WhereLoop*)*nLoop)*mxChoice*2; nSpace += sizeof(LogEst) * nOrderBy; pSpace = sqlite3DbMallocRaw(db, nSpace); if( pSpace==0 ) return SQLITE_NOMEM; aTo = (WherePath*)pSpace; aFrom = aTo+mxChoice; memset(aFrom, 0, sizeof(aFrom[0])); pX = (WhereLoop**)(aFrom+mxChoice); for(ii=mxChoice*2, pFrom=aTo; ii>0; ii--, pFrom++, pX += nLoop){ pFrom->aLoop = pX; } if( nOrderBy ){ /* If there is an ORDER BY clause and it is not being ignored, set up ** space for the aSortCost[] array. Each element of the aSortCost array ** is either zero - meaning it has not yet been initialized - or the ** cost of sorting nRowEst rows of data where the first X terms of ** the ORDER BY clause are already in order, where X is the array ** index. */ aSortCost = (LogEst*)pX; memset(aSortCost, 0, sizeof(LogEst) * nOrderBy); } assert( aSortCost==0 || &pSpace[nSpace]==(char*)&aSortCost[nOrderBy] ); assert( aSortCost!=0 || &pSpace[nSpace]==(char*)pX ); /* Seed the search with a single WherePath containing zero WhereLoops. ** ** TUNING: Do not let the number of iterations go above 25. If the cost ** of computing an automatic index is not paid back within the first 25 ** rows, then do not use the automatic index. */ aFrom[0].nRow = MIN(pParse->nQueryLoop, 46); assert( 46==sqlite3LogEst(25) ); nFrom = 1; assert( aFrom[0].isOrdered==0 ); if( nOrderBy ){ /* If nLoop is zero, then there are no FROM terms in the query. Since ** in this case the query may return a maximum of one row, the results ** are already in the requested order. Set isOrdered to nOrderBy to ** indicate this. Or, if nLoop is greater than zero, set isOrdered to ** -1, indicating that the result set may or may not be ordered, ** depending on the loops added to the current plan. */ aFrom[0].isOrdered = nLoop>0 ? -1 : nOrderBy; } /* Compute successively longer WherePaths using the previous generation ** of WherePaths as the basis for the next. Keep track of the mxChoice ** best paths at each generation */ for(iLoop=0; iLooppLoops; pWLoop; pWLoop=pWLoop->pNextLoop){ LogEst nOut; /* Rows visited by (pFrom+pWLoop) */ LogEst rCost; /* Cost of path (pFrom+pWLoop) */ LogEst rUnsorted; /* Unsorted cost of (pFrom+pWLoop) */ i8 isOrdered = pFrom->isOrdered; /* isOrdered for (pFrom+pWLoop) */ Bitmask maskNew; /* Mask of src visited by (..) */ Bitmask revMask = 0; /* Mask of rev-order loops for (..) */ if( (pWLoop->prereq & ~pFrom->maskLoop)!=0 ) continue; if( (pWLoop->maskSelf & pFrom->maskLoop)!=0 ) continue; /* At this point, pWLoop is a candidate to be the next loop. ** Compute its cost */ rUnsorted = sqlite3LogEstAdd(pWLoop->rSetup,pWLoop->rRun + pFrom->nRow); rUnsorted = sqlite3LogEstAdd(rUnsorted, pFrom->rUnsorted); nOut = pFrom->nRow + pWLoop->nOut; maskNew = pFrom->maskLoop | pWLoop->maskSelf; if( isOrdered<0 ){ isOrdered = wherePathSatisfiesOrderBy(pWInfo, pWInfo->pOrderBy, pFrom, pWInfo->wctrlFlags, iLoop, pWLoop, &revMask); }else{ revMask = pFrom->revLoop; } if( isOrdered>=0 && isOrderedisOrdered^isOrdered)&0x80)==0" is equivalent ** to (pTo->isOrdered==(-1))==(isOrdered==(-1))" for the range ** of legal values for isOrdered, -1..64. */ for(jj=0, pTo=aTo; jjmaskLoop==maskNew && ((pTo->isOrdered^isOrdered)&0x80)==0 ){ testcase( jj==nTo-1 ); break; } } if( jj>=nTo ){ /* None of the existing best-so-far paths match the candidate. */ if( nTo>=mxChoice && (rCost>mxCost || (rCost==mxCost && rUnsorted>=mxUnsorted)) ){ /* The current candidate is no better than any of the mxChoice ** paths currently in the best-so-far buffer. So discard ** this candidate as not viable. */ #ifdef WHERETRACE_ENABLED /* 0x4 */ if( sqlite3WhereTrace&0x4 ){ sqlite3DebugPrintf("Skip %s cost=%-3d,%3d order=%c\n", wherePathName(pFrom, iLoop, pWLoop), rCost, nOut, isOrdered>=0 ? isOrdered+'0' : '?'); } #endif continue; } /* If we reach this points it means that the new candidate path ** needs to be added to the set of best-so-far paths. */ if( nTo=0 ? isOrdered+'0' : '?'); } #endif }else{ /* Control reaches here if best-so-far path pTo=aTo[jj] covers the ** same set of loops and has the sam isOrdered setting as the ** candidate path. Check to see if the candidate should replace ** pTo or if the candidate should be skipped */ if( pTo->rCostrCost==rCost && pTo->nRow<=nOut) ){ #ifdef WHERETRACE_ENABLED /* 0x4 */ if( sqlite3WhereTrace&0x4 ){ sqlite3DebugPrintf( "Skip %s cost=%-3d,%3d order=%c", wherePathName(pFrom, iLoop, pWLoop), rCost, nOut, isOrdered>=0 ? isOrdered+'0' : '?'); sqlite3DebugPrintf(" vs %s cost=%-3d,%d order=%c\n", wherePathName(pTo, iLoop+1, 0), pTo->rCost, pTo->nRow, pTo->isOrdered>=0 ? pTo->isOrdered+'0' : '?'); } #endif /* Discard the candidate path from further consideration */ testcase( pTo->rCost==rCost ); continue; } testcase( pTo->rCost==rCost+1 ); /* Control reaches here if the candidate path is better than the ** pTo path. Replace pTo with the candidate. */ #ifdef WHERETRACE_ENABLED /* 0x4 */ if( sqlite3WhereTrace&0x4 ){ sqlite3DebugPrintf( "Update %s cost=%-3d,%3d order=%c", wherePathName(pFrom, iLoop, pWLoop), rCost, nOut, isOrdered>=0 ? isOrdered+'0' : '?'); sqlite3DebugPrintf(" was %s cost=%-3d,%3d order=%c\n", wherePathName(pTo, iLoop+1, 0), pTo->rCost, pTo->nRow, pTo->isOrdered>=0 ? pTo->isOrdered+'0' : '?'); } #endif } /* pWLoop is a winner. Add it to the set of best so far */ pTo->maskLoop = pFrom->maskLoop | pWLoop->maskSelf; pTo->revLoop = revMask; pTo->nRow = nOut; pTo->rCost = rCost; pTo->rUnsorted = rUnsorted; pTo->isOrdered = isOrdered; memcpy(pTo->aLoop, pFrom->aLoop, sizeof(WhereLoop*)*iLoop); pTo->aLoop[iLoop] = pWLoop; if( nTo>=mxChoice ){ mxI = 0; mxCost = aTo[0].rCost; mxUnsorted = aTo[0].nRow; for(jj=1, pTo=&aTo[1]; jjrCost>mxCost || (pTo->rCost==mxCost && pTo->rUnsorted>mxUnsorted) ){ mxCost = pTo->rCost; mxUnsorted = pTo->rUnsorted; mxI = jj; } } } } } #ifdef WHERETRACE_ENABLED /* >=2 */ if( sqlite3WhereTrace & 0x02 ){ sqlite3DebugPrintf("---- after round %d ----\n", iLoop); for(ii=0, pTo=aTo; iirCost, pTo->nRow, pTo->isOrdered>=0 ? (pTo->isOrdered+'0') : '?'); if( pTo->isOrdered>0 ){ sqlite3DebugPrintf(" rev=0x%llx\n", pTo->revLoop); }else{ sqlite3DebugPrintf("\n"); } } } #endif /* Swap the roles of aFrom and aTo for the next generation */ pFrom = aTo; aTo = aFrom; aFrom = pFrom; nFrom = nTo; } if( nFrom==0 ){ sqlite3ErrorMsg(pParse, "no query solution"); sqlite3DbFree(db, pSpace); return SQLITE_ERROR; } /* Find the lowest cost path. pFrom will be left pointing to that path */ pFrom = aFrom; for(ii=1; iirCost>aFrom[ii].rCost ) pFrom = &aFrom[ii]; } assert( pWInfo->nLevel==nLoop ); /* Load the lowest cost path into pWInfo */ for(iLoop=0; iLoopa + iLoop; pLevel->pWLoop = pWLoop = pFrom->aLoop[iLoop]; pLevel->iFrom = pWLoop->iTab; pLevel->iTabCur = pWInfo->pTabList->a[pLevel->iFrom].iCursor; } if( (pWInfo->wctrlFlags & WHERE_WANT_DISTINCT)!=0 && (pWInfo->wctrlFlags & WHERE_DISTINCTBY)==0 && pWInfo->eDistinct==WHERE_DISTINCT_NOOP && nRowEst ){ Bitmask notUsed; int rc = wherePathSatisfiesOrderBy(pWInfo, pWInfo->pResultSet, pFrom, WHERE_DISTINCTBY, nLoop-1, pFrom->aLoop[nLoop-1], ¬Used); if( rc==pWInfo->pResultSet->nExpr ){ pWInfo->eDistinct = WHERE_DISTINCT_ORDERED; } } if( pWInfo->pOrderBy ){ if( pWInfo->wctrlFlags & WHERE_DISTINCTBY ){ if( pFrom->isOrdered==pWInfo->pOrderBy->nExpr ){ pWInfo->eDistinct = WHERE_DISTINCT_ORDERED; } }else{ pWInfo->nOBSat = pFrom->isOrdered; if( pWInfo->nOBSat<0 ) pWInfo->nOBSat = 0; pWInfo->revMask = pFrom->revLoop; } if( (pWInfo->wctrlFlags & WHERE_SORTBYGROUP) && pWInfo->nOBSat==pWInfo->pOrderBy->nExpr ){ Bitmask revMask = 0; int nOrder = wherePathSatisfiesOrderBy(pWInfo, pWInfo->pOrderBy, pFrom, 0, nLoop-1, pFrom->aLoop[nLoop-1], &revMask ); assert( pWInfo->sorted==0 ); if( nOrder==pWInfo->pOrderBy->nExpr ){ pWInfo->sorted = 1; pWInfo->revMask = revMask; } } } pWInfo->nRowOut = pFrom->nRow; /* Free temporary memory and return success */ sqlite3DbFree(db, pSpace); return SQLITE_OK; } /* ** Most queries use only a single table (they are not joins) and have ** simple == constraints against indexed fields. This routine attempts ** to plan those simple cases using much less ceremony than the ** general-purpose query planner, and thereby yield faster sqlite3_prepare() ** times for the common case. ** ** Return non-zero on success, if this query can be handled by this ** no-frills query planner. Return zero if this query needs the ** general-purpose query planner. */ static int whereShortCut(WhereLoopBuilder *pBuilder){ WhereInfo *pWInfo; struct SrcList_item *pItem; WhereClause *pWC; WhereTerm *pTerm; WhereLoop *pLoop; int iCur; int j; Table *pTab; Index *pIdx; pWInfo = pBuilder->pWInfo; if( pWInfo->wctrlFlags & WHERE_FORCE_TABLE ) return 0; assert( pWInfo->pTabList->nSrc>=1 ); pItem = pWInfo->pTabList->a; pTab = pItem->pTab; if( IsVirtual(pTab) ) return 0; if( pItem->zIndex ) return 0; iCur = pItem->iCursor; pWC = &pWInfo->sWC; pLoop = pBuilder->pNew; pLoop->wsFlags = 0; pLoop->nSkip = 0; pTerm = findTerm(pWC, iCur, -1, 0, WO_EQ, 0); if( pTerm ){ pLoop->wsFlags = WHERE_COLUMN_EQ|WHERE_IPK|WHERE_ONEROW; pLoop->aLTerm[0] = pTerm; pLoop->nLTerm = 1; pLoop->u.btree.nEq = 1; /* TUNING: Cost of a rowid lookup is 10 */ pLoop->rRun = 33; /* 33==sqlite3LogEst(10) */ }else{ for(pIdx=pTab->pIndex; pIdx; pIdx=pIdx->pNext){ assert( pLoop->aLTermSpace==pLoop->aLTerm ); if( !IsUniqueIndex(pIdx) || pIdx->pPartIdxWhere!=0 || pIdx->nKeyCol>ArraySize(pLoop->aLTermSpace) ) continue; for(j=0; jnKeyCol; j++){ pTerm = findTerm(pWC, iCur, pIdx->aiColumn[j], 0, WO_EQ, pIdx); if( pTerm==0 ) break; pLoop->aLTerm[j] = pTerm; } if( j!=pIdx->nKeyCol ) continue; pLoop->wsFlags = WHERE_COLUMN_EQ|WHERE_ONEROW|WHERE_INDEXED; if( pIdx->isCovering || (pItem->colUsed & ~columnsInIndex(pIdx))==0 ){ pLoop->wsFlags |= WHERE_IDX_ONLY; } pLoop->nLTerm = j; pLoop->u.btree.nEq = j; pLoop->u.btree.pIndex = pIdx; /* TUNING: Cost of a unique index lookup is 15 */ pLoop->rRun = 39; /* 39==sqlite3LogEst(15) */ break; } } if( pLoop->wsFlags ){ pLoop->nOut = (LogEst)1; pWInfo->a[0].pWLoop = pLoop; pLoop->maskSelf = getMask(&pWInfo->sMaskSet, iCur); pWInfo->a[0].iTabCur = iCur; pWInfo->nRowOut = 1; if( pWInfo->pOrderBy ) pWInfo->nOBSat = pWInfo->pOrderBy->nExpr; if( pWInfo->wctrlFlags & WHERE_WANT_DISTINCT ){ pWInfo->eDistinct = WHERE_DISTINCT_UNIQUE; } #ifdef SQLITE_DEBUG pLoop->cId = '0'; #endif return 1; } return 0; } /* ** Generate the beginning of the loop used for WHERE clause processing. ** The return value is a pointer to an opaque structure that contains ** information needed to terminate the loop. Later, the calling routine ** should invoke sqlite3WhereEnd() with the return value of this function ** in order to complete the WHERE clause processing. ** ** If an error occurs, this routine returns NULL. ** ** The basic idea is to do a nested loop, one loop for each table in ** the FROM clause of a select. (INSERT and UPDATE statements are the ** same as a SELECT with only a single table in the FROM clause.) For ** example, if the SQL is this: ** ** SELECT * FROM t1, t2, t3 WHERE ...; ** ** Then the code generated is conceptually like the following: ** ** foreach row1 in t1 do \ Code generated ** foreach row2 in t2 do |-- by sqlite3WhereBegin() ** foreach row3 in t3 do / ** ... ** end \ Code generated ** end |-- by sqlite3WhereEnd() ** end / ** ** Note that the loops might not be nested in the order in which they ** appear in the FROM clause if a different order is better able to make ** use of indices. Note also that when the IN operator appears in ** the WHERE clause, it might result in additional nested loops for ** scanning through all values on the right-hand side of the IN. ** ** There are Btree cursors associated with each table. t1 uses cursor ** number pTabList->a[0].iCursor. t2 uses the cursor pTabList->a[1].iCursor. ** And so forth. This routine generates code to open those VDBE cursors ** and sqlite3WhereEnd() generates the code to close them. ** ** The code that sqlite3WhereBegin() generates leaves the cursors named ** in pTabList pointing at their appropriate entries. The [...] code ** can use OP_Column and OP_Rowid opcodes on these cursors to extract ** data from the various tables of the loop. ** ** If the WHERE clause is empty, the foreach loops must each scan their ** entire tables. Thus a three-way join is an O(N^3) operation. But if ** the tables have indices and there are terms in the WHERE clause that ** refer to those indices, a complete table scan can be avoided and the ** code will run much faster. Most of the work of this routine is checking ** to see if there are indices that can be used to speed up the loop. ** ** Terms of the WHERE clause are also used to limit which rows actually ** make it to the "..." in the middle of the loop. After each "foreach", ** terms of the WHERE clause that use only terms in that loop and outer ** loops are evaluated and if false a jump is made around all subsequent ** inner loops (or around the "..." if the test occurs within the inner- ** most loop) ** ** OUTER JOINS ** ** An outer join of tables t1 and t2 is conceptally coded as follows: ** ** foreach row1 in t1 do ** flag = 0 ** foreach row2 in t2 do ** start: ** ... ** flag = 1 ** end ** if flag==0 then ** move the row2 cursor to a null row ** goto start ** fi ** end ** ** ORDER BY CLAUSE PROCESSING ** ** pOrderBy is a pointer to the ORDER BY clause (or the GROUP BY clause ** if the WHERE_GROUPBY flag is set in wctrlFlags) of a SELECT statement ** if there is one. If there is no ORDER BY clause or if this routine ** is called from an UPDATE or DELETE statement, then pOrderBy is NULL. ** ** The iIdxCur parameter is the cursor number of an index. If ** WHERE_ONETABLE_ONLY is set, iIdxCur is the cursor number of an index ** to use for OR clause processing. The WHERE clause should use this ** specific cursor. If WHERE_ONEPASS_DESIRED is set, then iIdxCur is ** the first cursor in an array of cursors for all indices. iIdxCur should ** be used to compute the appropriate cursor depending on which index is ** used. */ WhereInfo *sqlite3WhereBegin( Parse *pParse, /* The parser context */ SrcList *pTabList, /* FROM clause: A list of all tables to be scanned */ Expr *pWhere, /* The WHERE clause */ ExprList *pOrderBy, /* An ORDER BY (or GROUP BY) clause, or NULL */ ExprList *pResultSet, /* Result set of the query */ u16 wctrlFlags, /* One of the WHERE_* flags defined in sqliteInt.h */ int iIdxCur /* If WHERE_ONETABLE_ONLY is set, index cursor number */ ){ int nByteWInfo; /* Num. bytes allocated for WhereInfo struct */ int nTabList; /* Number of elements in pTabList */ WhereInfo *pWInfo; /* Will become the return value of this function */ Vdbe *v = pParse->pVdbe; /* The virtual database engine */ Bitmask notReady; /* Cursors that are not yet positioned */ WhereLoopBuilder sWLB; /* The WhereLoop builder */ WhereMaskSet *pMaskSet; /* The expression mask set */ WhereLevel *pLevel; /* A single level in pWInfo->a[] */ WhereLoop *pLoop; /* Pointer to a single WhereLoop object */ int ii; /* Loop counter */ sqlite3 *db; /* Database connection */ int rc; /* Return code */ /* Variable initialization */ db = pParse->db; memset(&sWLB, 0, sizeof(sWLB)); /* An ORDER/GROUP BY clause of more than 63 terms cannot be optimized */ testcase( pOrderBy && pOrderBy->nExpr==BMS-1 ); if( pOrderBy && pOrderBy->nExpr>=BMS ) pOrderBy = 0; sWLB.pOrderBy = pOrderBy; /* Disable the DISTINCT optimization if SQLITE_DistinctOpt is set via ** sqlite3_test_ctrl(SQLITE_TESTCTRL_OPTIMIZATIONS,...) */ if( OptimizationDisabled(db, SQLITE_DistinctOpt) ){ wctrlFlags &= ~WHERE_WANT_DISTINCT; } /* The number of tables in the FROM clause is limited by the number of ** bits in a Bitmask */ testcase( pTabList->nSrc==BMS ); if( pTabList->nSrc>BMS ){ sqlite3ErrorMsg(pParse, "at most %d tables in a join", BMS); return 0; } /* This function normally generates a nested loop for all tables in ** pTabList. But if the WHERE_ONETABLE_ONLY flag is set, then we should ** only generate code for the first table in pTabList and assume that ** any cursors associated with subsequent tables are uninitialized. */ nTabList = (wctrlFlags & WHERE_ONETABLE_ONLY) ? 1 : pTabList->nSrc; /* Allocate and initialize the WhereInfo structure that will become the ** return value. A single allocation is used to store the WhereInfo ** struct, the contents of WhereInfo.a[], the WhereClause structure ** and the WhereMaskSet structure. Since WhereClause contains an 8-byte ** field (type Bitmask) it must be aligned on an 8-byte boundary on ** some architectures. Hence the ROUND8() below. */ nByteWInfo = ROUND8(sizeof(WhereInfo)+(nTabList-1)*sizeof(WhereLevel)); pWInfo = sqlite3DbMallocZero(db, nByteWInfo + sizeof(WhereLoop)); if( db->mallocFailed ){ sqlite3DbFree(db, pWInfo); pWInfo = 0; goto whereBeginError; } pWInfo->aiCurOnePass[0] = pWInfo->aiCurOnePass[1] = -1; pWInfo->nLevel = nTabList; pWInfo->pParse = pParse; pWInfo->pTabList = pTabList; pWInfo->pOrderBy = pOrderBy; pWInfo->pResultSet = pResultSet; pWInfo->iBreak = pWInfo->iContinue = sqlite3VdbeMakeLabel(v); pWInfo->wctrlFlags = wctrlFlags; pWInfo->savedNQueryLoop = pParse->nQueryLoop; pMaskSet = &pWInfo->sMaskSet; sWLB.pWInfo = pWInfo; sWLB.pWC = &pWInfo->sWC; sWLB.pNew = (WhereLoop*)(((char*)pWInfo)+nByteWInfo); assert( EIGHT_BYTE_ALIGNMENT(sWLB.pNew) ); whereLoopInit(sWLB.pNew); #ifdef SQLITE_DEBUG sWLB.pNew->cId = '*'; #endif /* Split the WHERE clause into separate subexpressions where each ** subexpression is separated by an AND operator. */ initMaskSet(pMaskSet); whereClauseInit(&pWInfo->sWC, pWInfo); whereSplit(&pWInfo->sWC, pWhere, TK_AND); /* Special case: a WHERE clause that is constant. Evaluate the ** expression and either jump over all of the code or fall thru. */ for(ii=0; iinTerm; ii++){ if( nTabList==0 || sqlite3ExprIsConstantNotJoin(sWLB.pWC->a[ii].pExpr) ){ sqlite3ExprIfFalse(pParse, sWLB.pWC->a[ii].pExpr, pWInfo->iBreak, SQLITE_JUMPIFNULL); sWLB.pWC->a[ii].wtFlags |= TERM_CODED; } } /* Special case: No FROM clause */ if( nTabList==0 ){ if( pOrderBy ) pWInfo->nOBSat = pOrderBy->nExpr; if( wctrlFlags & WHERE_WANT_DISTINCT ){ pWInfo->eDistinct = WHERE_DISTINCT_UNIQUE; } } /* Assign a bit from the bitmask to every term in the FROM clause. ** ** When assigning bitmask values to FROM clause cursors, it must be ** the case that if X is the bitmask for the N-th FROM clause term then ** the bitmask for all FROM clause terms to the left of the N-th term ** is (X-1). An expression from the ON clause of a LEFT JOIN can use ** its Expr.iRightJoinTable value to find the bitmask of the right table ** of the join. Subtracting one from the right table bitmask gives a ** bitmask for all tables to the left of the join. Knowing the bitmask ** for all tables to the left of a left join is important. Ticket #3015. ** ** Note that bitmasks are created for all pTabList->nSrc tables in ** pTabList, not just the first nTabList tables. nTabList is normally ** equal to pTabList->nSrc but might be shortened to 1 if the ** WHERE_ONETABLE_ONLY flag is set. */ for(ii=0; iinSrc; ii++){ createMask(pMaskSet, pTabList->a[ii].iCursor); } #ifndef NDEBUG { Bitmask toTheLeft = 0; for(ii=0; iinSrc; ii++){ Bitmask m = getMask(pMaskSet, pTabList->a[ii].iCursor); assert( (m-1)==toTheLeft ); toTheLeft |= m; } } #endif /* Analyze all of the subexpressions. Note that exprAnalyze() might ** add new virtual terms onto the end of the WHERE clause. We do not ** want to analyze these virtual terms, so start analyzing at the end ** and work forward so that the added virtual terms are never processed. */ exprAnalyzeAll(pTabList, &pWInfo->sWC); if( db->mallocFailed ){ goto whereBeginError; } if( wctrlFlags & WHERE_WANT_DISTINCT ){ if( isDistinctRedundant(pParse, pTabList, &pWInfo->sWC, pResultSet) ){ /* The DISTINCT marking is pointless. Ignore it. */ pWInfo->eDistinct = WHERE_DISTINCT_UNIQUE; }else if( pOrderBy==0 ){ /* Try to ORDER BY the result set to make distinct processing easier */ pWInfo->wctrlFlags |= WHERE_DISTINCTBY; pWInfo->pOrderBy = pResultSet; } } /* Construct the WhereLoop objects */ WHERETRACE(0xffff,("*** Optimizer Start ***\n")); #if defined(WHERETRACE_ENABLED) /* Display all terms of the WHERE clause */ if( sqlite3WhereTrace & 0x100 ){ int i; for(i=0; inTerm; i++){ whereTermPrint(&sWLB.pWC->a[i], i); } } #endif if( nTabList!=1 || whereShortCut(&sWLB)==0 ){ rc = whereLoopAddAll(&sWLB); if( rc ) goto whereBeginError; /* Display all of the WhereLoop objects if wheretrace is enabled */ #ifdef WHERETRACE_ENABLED /* !=0 */ if( sqlite3WhereTrace ){ WhereLoop *p; int i; static char zLabel[] = "0123456789abcdefghijklmnopqrstuvwyxz" "ABCDEFGHIJKLMNOPQRSTUVWYXZ"; for(p=pWInfo->pLoops, i=0; p; p=p->pNextLoop, i++){ p->cId = zLabel[i%sizeof(zLabel)]; whereLoopPrint(p, sWLB.pWC); } } #endif wherePathSolver(pWInfo, 0); if( db->mallocFailed ) goto whereBeginError; if( pWInfo->pOrderBy ){ wherePathSolver(pWInfo, pWInfo->nRowOut+1); if( db->mallocFailed ) goto whereBeginError; } } if( pWInfo->pOrderBy==0 && (db->flags & SQLITE_ReverseOrder)!=0 ){ pWInfo->revMask = (Bitmask)(-1); } if( pParse->nErr || NEVER(db->mallocFailed) ){ goto whereBeginError; } #ifdef WHERETRACE_ENABLED /* !=0 */ if( sqlite3WhereTrace ){ int ii; sqlite3DebugPrintf("---- Solution nRow=%d", pWInfo->nRowOut); if( pWInfo->nOBSat>0 ){ sqlite3DebugPrintf(" ORDERBY=%d,0x%llx", pWInfo->nOBSat, pWInfo->revMask); } switch( pWInfo->eDistinct ){ case WHERE_DISTINCT_UNIQUE: { sqlite3DebugPrintf(" DISTINCT=unique"); break; } case WHERE_DISTINCT_ORDERED: { sqlite3DebugPrintf(" DISTINCT=ordered"); break; } case WHERE_DISTINCT_UNORDERED: { sqlite3DebugPrintf(" DISTINCT=unordered"); break; } } sqlite3DebugPrintf("\n"); for(ii=0; iinLevel; ii++){ whereLoopPrint(pWInfo->a[ii].pWLoop, sWLB.pWC); } } #endif /* Attempt to omit tables from the join that do not effect the result */ if( pWInfo->nLevel>=2 && pResultSet!=0 && OptimizationEnabled(db, SQLITE_OmitNoopJoin) ){ Bitmask tabUsed = exprListTableUsage(pMaskSet, pResultSet); if( sWLB.pOrderBy ) tabUsed |= exprListTableUsage(pMaskSet, sWLB.pOrderBy); while( pWInfo->nLevel>=2 ){ WhereTerm *pTerm, *pEnd; pLoop = pWInfo->a[pWInfo->nLevel-1].pWLoop; if( (pWInfo->pTabList->a[pLoop->iTab].jointype & JT_LEFT)==0 ) break; if( (wctrlFlags & WHERE_WANT_DISTINCT)==0 && (pLoop->wsFlags & WHERE_ONEROW)==0 ){ break; } if( (tabUsed & pLoop->maskSelf)!=0 ) break; pEnd = sWLB.pWC->a + sWLB.pWC->nTerm; for(pTerm=sWLB.pWC->a; pTermprereqAll & pLoop->maskSelf)!=0 && !ExprHasProperty(pTerm->pExpr, EP_FromJoin) ){ break; } } if( pTerm drop loop %c not used\n", pLoop->cId)); pWInfo->nLevel--; nTabList--; } } WHERETRACE(0xffff,("*** Optimizer Finished ***\n")); pWInfo->pParse->nQueryLoop += pWInfo->nRowOut; /* If the caller is an UPDATE or DELETE statement that is requesting ** to use a one-pass algorithm, determine if this is appropriate. ** The one-pass algorithm only works if the WHERE clause constrains ** the statement to update a single row. */ assert( (wctrlFlags & WHERE_ONEPASS_DESIRED)==0 || pWInfo->nLevel==1 ); if( (wctrlFlags & WHERE_ONEPASS_DESIRED)!=0 && (pWInfo->a[0].pWLoop->wsFlags & WHERE_ONEROW)!=0 ){ pWInfo->okOnePass = 1; if( HasRowid(pTabList->a[0].pTab) ){ pWInfo->a[0].pWLoop->wsFlags &= ~WHERE_IDX_ONLY; } } /* Open all tables in the pTabList and any indices selected for ** searching those tables. */ notReady = ~(Bitmask)0; for(ii=0, pLevel=pWInfo->a; iia[pLevel->iFrom]; pTab = pTabItem->pTab; iDb = sqlite3SchemaToIndex(db, pTab->pSchema); pLoop = pLevel->pWLoop; if( (pTab->tabFlags & TF_Ephemeral)!=0 || pTab->pSelect ){ /* Do nothing */ }else #ifndef SQLITE_OMIT_VIRTUALTABLE if( (pLoop->wsFlags & WHERE_VIRTUALTABLE)!=0 ){ const char *pVTab = (const char *)sqlite3GetVTable(db, pTab); int iCur = pTabItem->iCursor; sqlite3VdbeAddOp4(v, OP_VOpen, iCur, 0, 0, pVTab, P4_VTAB); }else if( IsVirtual(pTab) ){ /* noop */ }else #endif if( (pLoop->wsFlags & WHERE_IDX_ONLY)==0 && (wctrlFlags & WHERE_OMIT_OPEN_CLOSE)==0 ){ int op = OP_OpenRead; if( pWInfo->okOnePass ){ op = OP_OpenWrite; pWInfo->aiCurOnePass[0] = pTabItem->iCursor; }; sqlite3OpenTable(pParse, pTabItem->iCursor, iDb, pTab, op); assert( pTabItem->iCursor==pLevel->iTabCur ); testcase( !pWInfo->okOnePass && pTab->nCol==BMS-1 ); testcase( !pWInfo->okOnePass && pTab->nCol==BMS ); if( !pWInfo->okOnePass && pTab->nColcolUsed; int n = 0; for(; b; b=b>>1, n++){} sqlite3VdbeChangeP4(v, sqlite3VdbeCurrentAddr(v)-1, SQLITE_INT_TO_PTR(n), P4_INT32); assert( n<=pTab->nCol ); } }else{ sqlite3TableLock(pParse, iDb, pTab->tnum, 0, pTab->zName); } if( pLoop->wsFlags & WHERE_INDEXED ){ Index *pIx = pLoop->u.btree.pIndex; int iIndexCur; int op = OP_OpenRead; /* iIdxCur is always set if to a positive value if ONEPASS is possible */ assert( iIdxCur!=0 || (pWInfo->wctrlFlags & WHERE_ONEPASS_DESIRED)==0 ); if( !HasRowid(pTab) && IsPrimaryKeyIndex(pIx) && (wctrlFlags & WHERE_ONETABLE_ONLY)!=0 ){ /* This is one term of an OR-optimization using the PRIMARY KEY of a ** WITHOUT ROWID table. No need for a separate index */ iIndexCur = pLevel->iTabCur; op = 0; }else if( pWInfo->okOnePass ){ Index *pJ = pTabItem->pTab->pIndex; iIndexCur = iIdxCur; assert( wctrlFlags & WHERE_ONEPASS_DESIRED ); while( ALWAYS(pJ) && pJ!=pIx ){ iIndexCur++; pJ = pJ->pNext; } op = OP_OpenWrite; pWInfo->aiCurOnePass[1] = iIndexCur; }else if( iIdxCur && (wctrlFlags & WHERE_ONETABLE_ONLY)!=0 ){ iIndexCur = iIdxCur; if( wctrlFlags & WHERE_REOPEN_IDX ) op = OP_ReopenIdx; }else{ iIndexCur = pParse->nTab++; } pLevel->iIdxCur = iIndexCur; assert( pIx->pSchema==pTab->pSchema ); assert( iIndexCur>=0 ); if( op ){ sqlite3VdbeAddOp3(v, op, iIndexCur, pIx->tnum, iDb); sqlite3VdbeSetP4KeyInfo(pParse, pIx); VdbeComment((v, "%s", pIx->zName)); } } if( iDb>=0 ) sqlite3CodeVerifySchema(pParse, iDb); notReady &= ~getMask(&pWInfo->sMaskSet, pTabItem->iCursor); } pWInfo->iTop = sqlite3VdbeCurrentAddr(v); if( db->mallocFailed ) goto whereBeginError; /* Generate the code to do the search. Each iteration of the for ** loop below generates code for a single nested loop of the VM ** program. */ notReady = ~(Bitmask)0; for(ii=0; iia[ii]; wsFlags = pLevel->pWLoop->wsFlags; #ifndef SQLITE_OMIT_AUTOMATIC_INDEX if( (pLevel->pWLoop->wsFlags & WHERE_AUTO_INDEX)!=0 ){ constructAutomaticIndex(pParse, &pWInfo->sWC, &pTabList->a[pLevel->iFrom], notReady, pLevel); if( db->mallocFailed ) goto whereBeginError; } #endif addrExplain = explainOneScan( pParse, pTabList, pLevel, ii, pLevel->iFrom, wctrlFlags ); pLevel->addrBody = sqlite3VdbeCurrentAddr(v); notReady = codeOneLoopStart(pWInfo, ii, notReady); pWInfo->iContinue = pLevel->addrCont; if( (wsFlags&WHERE_MULTI_OR)==0 && (wctrlFlags&WHERE_ONETABLE_ONLY)==0 ){ addScanStatus(v, pTabList, pLevel, addrExplain); } } /* Done. */ VdbeModuleComment((v, "Begin WHERE-core")); return pWInfo; /* Jump here if malloc fails */ whereBeginError: if( pWInfo ){ pParse->nQueryLoop = pWInfo->savedNQueryLoop; whereInfoFree(db, pWInfo); } return 0; } /* ** Generate the end of the WHERE loop. See comments on ** sqlite3WhereBegin() for additional information. */ void sqlite3WhereEnd(WhereInfo *pWInfo){ Parse *pParse = pWInfo->pParse; Vdbe *v = pParse->pVdbe; int i; WhereLevel *pLevel; WhereLoop *pLoop; SrcList *pTabList = pWInfo->pTabList; sqlite3 *db = pParse->db; /* Generate loop termination code. */ VdbeModuleComment((v, "End WHERE-core")); sqlite3ExprCacheClear(pParse); for(i=pWInfo->nLevel-1; i>=0; i--){ int addr; pLevel = &pWInfo->a[i]; pLoop = pLevel->pWLoop; sqlite3VdbeResolveLabel(v, pLevel->addrCont); if( pLevel->op!=OP_Noop ){ sqlite3VdbeAddOp3(v, pLevel->op, pLevel->p1, pLevel->p2, pLevel->p3); sqlite3VdbeChangeP5(v, pLevel->p5); VdbeCoverage(v); VdbeCoverageIf(v, pLevel->op==OP_Next); VdbeCoverageIf(v, pLevel->op==OP_Prev); VdbeCoverageIf(v, pLevel->op==OP_VNext); } if( pLoop->wsFlags & WHERE_IN_ABLE && pLevel->u.in.nIn>0 ){ struct InLoop *pIn; int j; sqlite3VdbeResolveLabel(v, pLevel->addrNxt); for(j=pLevel->u.in.nIn, pIn=&pLevel->u.in.aInLoop[j-1]; j>0; j--, pIn--){ sqlite3VdbeJumpHere(v, pIn->addrInTop+1); sqlite3VdbeAddOp2(v, pIn->eEndLoopOp, pIn->iCur, pIn->addrInTop); VdbeCoverage(v); VdbeCoverageIf(v, pIn->eEndLoopOp==OP_PrevIfOpen); VdbeCoverageIf(v, pIn->eEndLoopOp==OP_NextIfOpen); sqlite3VdbeJumpHere(v, pIn->addrInTop-1); } sqlite3DbFree(db, pLevel->u.in.aInLoop); } sqlite3VdbeResolveLabel(v, pLevel->addrBrk); if( pLevel->addrSkip ){ sqlite3VdbeAddOp2(v, OP_Goto, 0, pLevel->addrSkip); VdbeComment((v, "next skip-scan on %s", pLoop->u.btree.pIndex->zName)); sqlite3VdbeJumpHere(v, pLevel->addrSkip); sqlite3VdbeJumpHere(v, pLevel->addrSkip-2); } if( pLevel->iLeftJoin ){ addr = sqlite3VdbeAddOp1(v, OP_IfPos, pLevel->iLeftJoin); VdbeCoverage(v); assert( (pLoop->wsFlags & WHERE_IDX_ONLY)==0 || (pLoop->wsFlags & WHERE_INDEXED)!=0 ); if( (pLoop->wsFlags & WHERE_IDX_ONLY)==0 ){ sqlite3VdbeAddOp1(v, OP_NullRow, pTabList->a[i].iCursor); } if( pLoop->wsFlags & WHERE_INDEXED ){ sqlite3VdbeAddOp1(v, OP_NullRow, pLevel->iIdxCur); } if( pLevel->op==OP_Return ){ sqlite3VdbeAddOp2(v, OP_Gosub, pLevel->p1, pLevel->addrFirst); }else{ sqlite3VdbeAddOp2(v, OP_Goto, 0, pLevel->addrFirst); } sqlite3VdbeJumpHere(v, addr); } VdbeModuleComment((v, "End WHERE-loop%d: %s", i, pWInfo->pTabList->a[pLevel->iFrom].pTab->zName)); } /* The "break" point is here, just past the end of the outer loop. ** Set it. */ sqlite3VdbeResolveLabel(v, pWInfo->iBreak); assert( pWInfo->nLevel<=pTabList->nSrc ); for(i=0, pLevel=pWInfo->a; inLevel; i++, pLevel++){ int k, last; VdbeOp *pOp; Index *pIdx = 0; struct SrcList_item *pTabItem = &pTabList->a[pLevel->iFrom]; Table *pTab = pTabItem->pTab; assert( pTab!=0 ); pLoop = pLevel->pWLoop; /* For a co-routine, change all OP_Column references to the table of ** the co-routine into OP_SCopy of result contained in a register. ** OP_Rowid becomes OP_Null. */ if( pTabItem->viaCoroutine && !db->mallocFailed ){ last = sqlite3VdbeCurrentAddr(v); k = pLevel->addrBody; pOp = sqlite3VdbeGetOp(v, k); for(; kp1!=pLevel->iTabCur ) continue; if( pOp->opcode==OP_Column ){ pOp->opcode = OP_Copy; pOp->p1 = pOp->p2 + pTabItem->regResult; pOp->p2 = pOp->p3; pOp->p3 = 0; }else if( pOp->opcode==OP_Rowid ){ pOp->opcode = OP_Null; pOp->p1 = 0; pOp->p3 = 0; } } continue; } /* Close all of the cursors that were opened by sqlite3WhereBegin. ** Except, do not close cursors that will be reused by the OR optimization ** (WHERE_OMIT_OPEN_CLOSE). And do not close the OP_OpenWrite cursors ** created for the ONEPASS optimization. */ if( (pTab->tabFlags & TF_Ephemeral)==0 && pTab->pSelect==0 && (pWInfo->wctrlFlags & WHERE_OMIT_OPEN_CLOSE)==0 ){ int ws = pLoop->wsFlags; if( !pWInfo->okOnePass && (ws & WHERE_IDX_ONLY)==0 ){ sqlite3VdbeAddOp1(v, OP_Close, pTabItem->iCursor); } if( (ws & WHERE_INDEXED)!=0 && (ws & (WHERE_IPK|WHERE_AUTO_INDEX))==0 && pLevel->iIdxCur!=pWInfo->aiCurOnePass[1] ){ sqlite3VdbeAddOp1(v, OP_Close, pLevel->iIdxCur); } } /* If this scan uses an index, make VDBE code substitutions to read data ** from the index instead of from the table where possible. In some cases ** this optimization prevents the table from ever being read, which can ** yield a significant performance boost. ** ** Calls to the code generator in between sqlite3WhereBegin and ** sqlite3WhereEnd will have created code that references the table ** directly. This loop scans all that code looking for opcodes ** that reference the table and converts them into opcodes that ** reference the index. */ if( pLoop->wsFlags & (WHERE_INDEXED|WHERE_IDX_ONLY) ){ pIdx = pLoop->u.btree.pIndex; }else if( pLoop->wsFlags & WHERE_MULTI_OR ){ pIdx = pLevel->u.pCovidx; } if( pIdx && !db->mallocFailed ){ last = sqlite3VdbeCurrentAddr(v); k = pLevel->addrBody; pOp = sqlite3VdbeGetOp(v, k); for(; kp1!=pLevel->iTabCur ) continue; if( pOp->opcode==OP_Column ){ int x = pOp->p2; assert( pIdx->pTable==pTab ); if( !HasRowid(pTab) ){ Index *pPk = sqlite3PrimaryKeyIndex(pTab); x = pPk->aiColumn[x]; } x = sqlite3ColumnOfIndex(pIdx, x); if( x>=0 ){ pOp->p2 = x; pOp->p1 = pLevel->iIdxCur; } assert( (pLoop->wsFlags & WHERE_IDX_ONLY)==0 || x>=0 ); }else if( pOp->opcode==OP_Rowid ){ pOp->p1 = pLevel->iIdxCur; pOp->opcode = OP_IdxRowid; } } } } /* Final cleanup */ pParse->nQueryLoop = pWInfo->savedNQueryLoop; whereInfoFree(db, pWInfo); return; }