4 ** The author disclaims copyright to this source code. In place of
5 ** a legal notice, here is a blessing:
7 ** May you do good and not evil.
8 ** May you find forgiveness for yourself and forgive others.
9 ** May you share freely, never taking more than you give.
11 *************************************************************************
12 ** This module contains C code that generates VDBE code used to process
13 ** the WHERE clause of SQL statements. This module is responsible for
14 ** generating the code that loops through a table looking for applicable
15 ** rows. Indices are selected and used to speed the search when doing
16 ** so is applicable. Because this module is responsible for selecting
17 ** indices, you might also think of this module as the "query optimizer".
19 #include "sqliteInt.h"
23 ** Extra information appended to the end of sqlite3_index_info but not
24 ** visible to the xBestIndex function, at least not directly. The
25 ** sqlite3_vtab_collation() interface knows how to reach it, however.
27 ** This object is not an API and can be changed from one release to the
28 ** next. As long as allocateIndexInfo() and sqlite3_vtab_collation()
29 ** agree on the structure, all will be well.
31 typedef struct HiddenIndexInfo HiddenIndexInfo
;
32 struct HiddenIndexInfo
{
33 WhereClause
*pWC
; /* The Where clause being analyzed */
34 Parse
*pParse
; /* The parsing context */
37 /* Forward declaration of methods */
38 static int whereLoopResize(sqlite3
*, WhereLoop
*, int);
40 /* Test variable that can be set to enable WHERE tracing */
41 #if defined(SQLITE_TEST) || defined(SQLITE_DEBUG)
42 /***/ int sqlite3WhereTrace
= 0;
47 ** Return the estimated number of output rows from a WHERE clause
49 LogEst
sqlite3WhereOutputRowCount(WhereInfo
*pWInfo
){
50 return pWInfo
->nRowOut
;
54 ** Return one of the WHERE_DISTINCT_xxxxx values to indicate how this
55 ** WHERE clause returns outputs for DISTINCT processing.
57 int sqlite3WhereIsDistinct(WhereInfo
*pWInfo
){
58 return pWInfo
->eDistinct
;
62 ** Return TRUE if the WHERE clause returns rows in ORDER BY order.
63 ** Return FALSE if the output needs to be sorted.
65 int sqlite3WhereIsOrdered(WhereInfo
*pWInfo
){
66 return pWInfo
->nOBSat
;
70 ** Return TRUE if the innermost loop of the WHERE clause implementation
71 ** returns rows in ORDER BY order for complete run of the inner loop.
73 ** Across multiple iterations of outer loops, the output rows need not be
74 ** sorted. As long as rows are sorted for just the innermost loop, this
75 ** routine can return TRUE.
77 int sqlite3WhereOrderedInnerLoop(WhereInfo
*pWInfo
){
78 return pWInfo
->bOrderedInnerLoop
;
82 ** Return the VDBE address or label to jump to in order to continue
83 ** immediately with the next row of a WHERE clause.
85 int sqlite3WhereContinueLabel(WhereInfo
*pWInfo
){
86 assert( pWInfo
->iContinue
!=0 );
87 return pWInfo
->iContinue
;
91 ** Return the VDBE address or label to jump to in order to break
92 ** out of a WHERE loop.
94 int sqlite3WhereBreakLabel(WhereInfo
*pWInfo
){
95 return pWInfo
->iBreak
;
99 ** Return ONEPASS_OFF (0) if an UPDATE or DELETE statement is unable to
100 ** operate directly on the rowis returned by a WHERE clause. Return
101 ** ONEPASS_SINGLE (1) if the statement can operation directly because only
102 ** a single row is to be changed. Return ONEPASS_MULTI (2) if the one-pass
103 ** optimization can be used on multiple
105 ** If the ONEPASS optimization is used (if this routine returns true)
106 ** then also write the indices of open cursors used by ONEPASS
107 ** into aiCur[0] and aiCur[1]. iaCur[0] gets the cursor of the data
108 ** table and iaCur[1] gets the cursor used by an auxiliary index.
109 ** Either value may be -1, indicating that cursor is not used.
110 ** Any cursors returned will have been opened for writing.
112 ** aiCur[0] and aiCur[1] both get -1 if the where-clause logic is
113 ** unable to use the ONEPASS optimization.
115 int sqlite3WhereOkOnePass(WhereInfo
*pWInfo
, int *aiCur
){
116 memcpy(aiCur
, pWInfo
->aiCurOnePass
, sizeof(int)*2);
117 #ifdef WHERETRACE_ENABLED
118 if( sqlite3WhereTrace
&& pWInfo
->eOnePass
!=ONEPASS_OFF
){
119 sqlite3DebugPrintf("%s cursors: %d %d\n",
120 pWInfo
->eOnePass
==ONEPASS_SINGLE
? "ONEPASS_SINGLE" : "ONEPASS_MULTI",
124 return pWInfo
->eOnePass
;
128 ** Move the content of pSrc into pDest
130 static void whereOrMove(WhereOrSet
*pDest
, WhereOrSet
*pSrc
){
132 memcpy(pDest
->a
, pSrc
->a
, pDest
->n
*sizeof(pDest
->a
[0]));
136 ** Try to insert a new prerequisite/cost entry into the WhereOrSet pSet.
138 ** The new entry might overwrite an existing entry, or it might be
139 ** appended, or it might be discarded. Do whatever is the right thing
140 ** so that pSet keeps the N_OR_COST best entries seen so far.
142 static int whereOrInsert(
143 WhereOrSet
*pSet
, /* The WhereOrSet to be updated */
144 Bitmask prereq
, /* Prerequisites of the new entry */
145 LogEst rRun
, /* Run-cost of the new entry */
146 LogEst nOut
/* Number of outputs for the new entry */
150 for(i
=pSet
->n
, p
=pSet
->a
; i
>0; i
--, p
++){
151 if( rRun
<=p
->rRun
&& (prereq
& p
->prereq
)==prereq
){
152 goto whereOrInsert_done
;
154 if( p
->rRun
<=rRun
&& (p
->prereq
& prereq
)==p
->prereq
){
158 if( pSet
->n
<N_OR_COST
){
159 p
= &pSet
->a
[pSet
->n
++];
163 for(i
=1; i
<pSet
->n
; i
++){
164 if( p
->rRun
>pSet
->a
[i
].rRun
) p
= pSet
->a
+ i
;
166 if( p
->rRun
<=rRun
) return 0;
171 if( p
->nOut
>nOut
) p
->nOut
= nOut
;
176 ** Return the bitmask for the given cursor number. Return 0 if
177 ** iCursor is not in the set.
179 Bitmask
sqlite3WhereGetMask(WhereMaskSet
*pMaskSet
, int iCursor
){
181 assert( pMaskSet
->n
<=(int)sizeof(Bitmask
)*8 );
182 for(i
=0; i
<pMaskSet
->n
; i
++){
183 if( pMaskSet
->ix
[i
]==iCursor
){
191 ** Create a new mask for cursor iCursor.
193 ** There is one cursor per table in the FROM clause. The number of
194 ** tables in the FROM clause is limited by a test early in the
195 ** sqlite3WhereBegin() routine. So we know that the pMaskSet->ix[]
196 ** array will never overflow.
198 static void createMask(WhereMaskSet
*pMaskSet
, int iCursor
){
199 assert( pMaskSet
->n
< ArraySize(pMaskSet
->ix
) );
200 pMaskSet
->ix
[pMaskSet
->n
++] = iCursor
;
204 ** Advance to the next WhereTerm that matches according to the criteria
205 ** established when the pScan object was initialized by whereScanInit().
206 ** Return NULL if there are no more matching WhereTerms.
208 static WhereTerm
*whereScanNext(WhereScan
*pScan
){
209 int iCur
; /* The cursor on the LHS of the term */
210 i16 iColumn
; /* The column on the LHS of the term. -1 for IPK */
211 Expr
*pX
; /* An expression being tested */
212 WhereClause
*pWC
; /* Shorthand for pScan->pWC */
213 WhereTerm
*pTerm
; /* The term being tested */
214 int k
= pScan
->k
; /* Where to start scanning */
216 assert( pScan
->iEquiv
<=pScan
->nEquiv
);
219 iColumn
= pScan
->aiColumn
[pScan
->iEquiv
-1];
220 iCur
= pScan
->aiCur
[pScan
->iEquiv
-1];
223 for(pTerm
=pWC
->a
+k
; k
<pWC
->nTerm
; k
++, pTerm
++){
224 if( pTerm
->leftCursor
==iCur
225 && pTerm
->u
.leftColumn
==iColumn
227 || sqlite3ExprCompareSkip(pTerm
->pExpr
->pLeft
,
228 pScan
->pIdxExpr
,iCur
)==0)
229 && (pScan
->iEquiv
<=1 || !ExprHasProperty(pTerm
->pExpr
, EP_FromJoin
))
231 if( (pTerm
->eOperator
& WO_EQUIV
)!=0
232 && pScan
->nEquiv
<ArraySize(pScan
->aiCur
)
233 && (pX
= sqlite3ExprSkipCollate(pTerm
->pExpr
->pRight
))->op
==TK_COLUMN
236 for(j
=0; j
<pScan
->nEquiv
; j
++){
237 if( pScan
->aiCur
[j
]==pX
->iTable
238 && pScan
->aiColumn
[j
]==pX
->iColumn
){
242 if( j
==pScan
->nEquiv
){
243 pScan
->aiCur
[j
] = pX
->iTable
;
244 pScan
->aiColumn
[j
] = pX
->iColumn
;
248 if( (pTerm
->eOperator
& pScan
->opMask
)!=0 ){
249 /* Verify the affinity and collating sequence match */
250 if( pScan
->zCollName
&& (pTerm
->eOperator
& WO_ISNULL
)==0 ){
252 Parse
*pParse
= pWC
->pWInfo
->pParse
;
254 if( !sqlite3IndexAffinityOk(pX
, pScan
->idxaff
) ){
258 pColl
= sqlite3BinaryCompareCollSeq(pParse
,
259 pX
->pLeft
, pX
->pRight
);
260 if( pColl
==0 ) pColl
= pParse
->db
->pDfltColl
;
261 if( sqlite3StrICmp(pColl
->zName
, pScan
->zCollName
) ){
265 if( (pTerm
->eOperator
& (WO_EQ
|WO_IS
))!=0
266 && (pX
= pTerm
->pExpr
->pRight
)->op
==TK_COLUMN
267 && pX
->iTable
==pScan
->aiCur
[0]
268 && pX
->iColumn
==pScan
->aiColumn
[0]
270 testcase( pTerm
->eOperator
& WO_IS
);
282 if( pScan
->iEquiv
>=pScan
->nEquiv
) break;
283 pWC
= pScan
->pOrigWC
;
291 ** Initialize a WHERE clause scanner object. Return a pointer to the
292 ** first match. Return NULL if there are no matches.
294 ** The scanner will be searching the WHERE clause pWC. It will look
295 ** for terms of the form "X <op> <expr>" where X is column iColumn of table
296 ** iCur. Or if pIdx!=0 then X is column iColumn of index pIdx. pIdx
297 ** must be one of the indexes of table iCur.
299 ** The <op> must be one of the operators described by opMask.
301 ** If the search is for X and the WHERE clause contains terms of the
302 ** form X=Y then this routine might also return terms of the form
303 ** "Y <op> <expr>". The number of levels of transitivity is limited,
304 ** but is enough to handle most commonly occurring SQL statements.
306 ** If X is not the INTEGER PRIMARY KEY then X must be compatible with
309 static WhereTerm
*whereScanInit(
310 WhereScan
*pScan
, /* The WhereScan object being initialized */
311 WhereClause
*pWC
, /* The WHERE clause to be scanned */
312 int iCur
, /* Cursor to scan for */
313 int iColumn
, /* Column to scan for */
314 u32 opMask
, /* Operator(s) to scan for */
315 Index
*pIdx
/* Must be compatible with this index */
317 pScan
->pOrigWC
= pWC
;
321 pScan
->zCollName
= 0;
324 iColumn
= pIdx
->aiColumn
[j
];
325 if( iColumn
==XN_EXPR
){
326 pScan
->pIdxExpr
= pIdx
->aColExpr
->a
[j
].pExpr
;
327 pScan
->zCollName
= pIdx
->azColl
[j
];
328 }else if( iColumn
==pIdx
->pTable
->iPKey
){
330 }else if( iColumn
>=0 ){
331 pScan
->idxaff
= pIdx
->pTable
->aCol
[iColumn
].affinity
;
332 pScan
->zCollName
= pIdx
->azColl
[j
];
334 }else if( iColumn
==XN_EXPR
){
337 pScan
->opMask
= opMask
;
339 pScan
->aiCur
[0] = iCur
;
340 pScan
->aiColumn
[0] = iColumn
;
343 return whereScanNext(pScan
);
347 ** Search for a term in the WHERE clause that is of the form "X <op> <expr>"
348 ** where X is a reference to the iColumn of table iCur or of index pIdx
349 ** if pIdx!=0 and <op> is one of the WO_xx operator codes specified by
350 ** the op parameter. Return a pointer to the term. Return 0 if not found.
352 ** If pIdx!=0 then it must be one of the indexes of table iCur.
353 ** Search for terms matching the iColumn-th column of pIdx
354 ** rather than the iColumn-th column of table iCur.
356 ** The term returned might by Y=<expr> if there is another constraint in
357 ** the WHERE clause that specifies that X=Y. Any such constraints will be
358 ** identified by the WO_EQUIV bit in the pTerm->eOperator field. The
359 ** aiCur[]/iaColumn[] arrays hold X and all its equivalents. There are 11
360 ** slots in aiCur[]/aiColumn[] so that means we can look for X plus up to 10
361 ** other equivalent values. Hence a search for X will return <expr> if X=A1
362 ** and A1=A2 and A2=A3 and ... and A9=A10 and A10=<expr>.
364 ** If there are multiple terms in the WHERE clause of the form "X <op> <expr>"
365 ** then try for the one with no dependencies on <expr> - in other words where
366 ** <expr> is a constant expression of some kind. Only return entries of
367 ** the form "X <op> Y" where Y is a column in another table if no terms of
368 ** the form "X <op> <const-expr>" exist. If no terms with a constant RHS
369 ** exist, try to return a term that does not use WO_EQUIV.
371 WhereTerm
*sqlite3WhereFindTerm(
372 WhereClause
*pWC
, /* The WHERE clause to be searched */
373 int iCur
, /* Cursor number of LHS */
374 int iColumn
, /* Column number of LHS */
375 Bitmask notReady
, /* RHS must not overlap with this mask */
376 u32 op
, /* Mask of WO_xx values describing operator */
377 Index
*pIdx
/* Must be compatible with this index, if not NULL */
379 WhereTerm
*pResult
= 0;
383 p
= whereScanInit(&scan
, pWC
, iCur
, iColumn
, op
, pIdx
);
386 if( (p
->prereqRight
& notReady
)==0 ){
387 if( p
->prereqRight
==0 && (p
->eOperator
&op
)!=0 ){
388 testcase( p
->eOperator
& WO_IS
);
391 if( pResult
==0 ) pResult
= p
;
393 p
= whereScanNext(&scan
);
399 ** This function searches pList for an entry that matches the iCol-th column
402 ** If such an expression is found, its index in pList->a[] is returned. If
403 ** no expression is found, -1 is returned.
405 static int findIndexCol(
406 Parse
*pParse
, /* Parse context */
407 ExprList
*pList
, /* Expression list to search */
408 int iBase
, /* Cursor for table associated with pIdx */
409 Index
*pIdx
, /* Index to match column of */
410 int iCol
/* Column of index to match */
413 const char *zColl
= pIdx
->azColl
[iCol
];
415 for(i
=0; i
<pList
->nExpr
; i
++){
416 Expr
*p
= sqlite3ExprSkipCollate(pList
->a
[i
].pExpr
);
418 && p
->iColumn
==pIdx
->aiColumn
[iCol
]
421 CollSeq
*pColl
= sqlite3ExprNNCollSeq(pParse
, pList
->a
[i
].pExpr
);
422 if( 0==sqlite3StrICmp(pColl
->zName
, zColl
) ){
432 ** Return TRUE if the iCol-th column of index pIdx is NOT NULL
434 static int indexColumnNotNull(Index
*pIdx
, int iCol
){
437 assert( iCol
>=0 && iCol
<pIdx
->nColumn
);
438 j
= pIdx
->aiColumn
[iCol
];
440 return pIdx
->pTable
->aCol
[j
].notNull
;
445 return 0; /* Assume an indexed expression can always yield a NULL */
451 ** Return true if the DISTINCT expression-list passed as the third argument
454 ** A DISTINCT list is redundant if any subset of the columns in the
455 ** DISTINCT list are collectively unique and individually non-null.
457 static int isDistinctRedundant(
458 Parse
*pParse
, /* Parsing context */
459 SrcList
*pTabList
, /* The FROM clause */
460 WhereClause
*pWC
, /* The WHERE clause */
461 ExprList
*pDistinct
/* The result set that needs to be DISTINCT */
468 /* If there is more than one table or sub-select in the FROM clause of
469 ** this query, then it will not be possible to show that the DISTINCT
470 ** clause is redundant. */
471 if( pTabList
->nSrc
!=1 ) return 0;
472 iBase
= pTabList
->a
[0].iCursor
;
473 pTab
= pTabList
->a
[0].pTab
;
475 /* If any of the expressions is an IPK column on table iBase, then return
476 ** true. Note: The (p->iTable==iBase) part of this test may be false if the
477 ** current SELECT is a correlated sub-query.
479 for(i
=0; i
<pDistinct
->nExpr
; i
++){
480 Expr
*p
= sqlite3ExprSkipCollate(pDistinct
->a
[i
].pExpr
);
481 if( p
->op
==TK_COLUMN
&& p
->iTable
==iBase
&& p
->iColumn
<0 ) return 1;
484 /* Loop through all indices on the table, checking each to see if it makes
485 ** the DISTINCT qualifier redundant. It does so if:
487 ** 1. The index is itself UNIQUE, and
489 ** 2. All of the columns in the index are either part of the pDistinct
490 ** list, or else the WHERE clause contains a term of the form "col=X",
491 ** where X is a constant value. The collation sequences of the
492 ** comparison and select-list expressions must match those of the index.
494 ** 3. All of those index columns for which the WHERE clause does not
495 ** contain a "col=X" term are subject to a NOT NULL constraint.
497 for(pIdx
=pTab
->pIndex
; pIdx
; pIdx
=pIdx
->pNext
){
498 if( !IsUniqueIndex(pIdx
) ) continue;
499 for(i
=0; i
<pIdx
->nKeyCol
; i
++){
500 if( 0==sqlite3WhereFindTerm(pWC
, iBase
, i
, ~(Bitmask
)0, WO_EQ
, pIdx
) ){
501 if( findIndexCol(pParse
, pDistinct
, iBase
, pIdx
, i
)<0 ) break;
502 if( indexColumnNotNull(pIdx
, i
)==0 ) break;
505 if( i
==pIdx
->nKeyCol
){
506 /* This index implies that the DISTINCT qualifier is redundant. */
516 ** Estimate the logarithm of the input value to base 2.
518 static LogEst
estLog(LogEst N
){
519 return N
<=10 ? 0 : sqlite3LogEst(N
) - 33;
523 ** Convert OP_Column opcodes to OP_Copy in previously generated code.
525 ** This routine runs over generated VDBE code and translates OP_Column
526 ** opcodes into OP_Copy when the table is being accessed via co-routine
527 ** instead of via table lookup.
529 ** If the bIncrRowid parameter is 0, then any OP_Rowid instructions on
530 ** cursor iTabCur are transformed into OP_Null. Or, if bIncrRowid is non-zero,
531 ** then each OP_Rowid is transformed into an instruction to increment the
532 ** value stored in its output register.
534 static void translateColumnToCopy(
535 Parse
*pParse
, /* Parsing context */
536 int iStart
, /* Translate from this opcode to the end */
537 int iTabCur
, /* OP_Column/OP_Rowid references to this table */
538 int iRegister
, /* The first column is in this register */
539 int bIncrRowid
/* If non-zero, transform OP_rowid to OP_AddImm(1) */
541 Vdbe
*v
= pParse
->pVdbe
;
542 VdbeOp
*pOp
= sqlite3VdbeGetOp(v
, iStart
);
543 int iEnd
= sqlite3VdbeCurrentAddr(v
);
544 if( pParse
->db
->mallocFailed
) return;
545 for(; iStart
<iEnd
; iStart
++, pOp
++){
546 if( pOp
->p1
!=iTabCur
) continue;
547 if( pOp
->opcode
==OP_Column
){
548 pOp
->opcode
= OP_Copy
;
549 pOp
->p1
= pOp
->p2
+ iRegister
;
552 }else if( pOp
->opcode
==OP_Rowid
){
554 /* Increment the value stored in the P2 operand of the OP_Rowid. */
555 pOp
->opcode
= OP_AddImm
;
559 pOp
->opcode
= OP_Null
;
568 ** Two routines for printing the content of an sqlite3_index_info
569 ** structure. Used for testing and debugging only. If neither
570 ** SQLITE_TEST or SQLITE_DEBUG are defined, then these routines
573 #if !defined(SQLITE_OMIT_VIRTUALTABLE) && defined(WHERETRACE_ENABLED)
574 static void TRACE_IDX_INPUTS(sqlite3_index_info
*p
){
576 if( !sqlite3WhereTrace
) return;
577 for(i
=0; i
<p
->nConstraint
; i
++){
578 sqlite3DebugPrintf(" constraint[%d]: col=%d termid=%d op=%d usabled=%d\n",
580 p
->aConstraint
[i
].iColumn
,
581 p
->aConstraint
[i
].iTermOffset
,
582 p
->aConstraint
[i
].op
,
583 p
->aConstraint
[i
].usable
);
585 for(i
=0; i
<p
->nOrderBy
; i
++){
586 sqlite3DebugPrintf(" orderby[%d]: col=%d desc=%d\n",
588 p
->aOrderBy
[i
].iColumn
,
589 p
->aOrderBy
[i
].desc
);
592 static void TRACE_IDX_OUTPUTS(sqlite3_index_info
*p
){
594 if( !sqlite3WhereTrace
) return;
595 for(i
=0; i
<p
->nConstraint
; i
++){
596 sqlite3DebugPrintf(" usage[%d]: argvIdx=%d omit=%d\n",
598 p
->aConstraintUsage
[i
].argvIndex
,
599 p
->aConstraintUsage
[i
].omit
);
601 sqlite3DebugPrintf(" idxNum=%d\n", p
->idxNum
);
602 sqlite3DebugPrintf(" idxStr=%s\n", p
->idxStr
);
603 sqlite3DebugPrintf(" orderByConsumed=%d\n", p
->orderByConsumed
);
604 sqlite3DebugPrintf(" estimatedCost=%g\n", p
->estimatedCost
);
605 sqlite3DebugPrintf(" estimatedRows=%lld\n", p
->estimatedRows
);
608 #define TRACE_IDX_INPUTS(A)
609 #define TRACE_IDX_OUTPUTS(A)
612 #ifndef SQLITE_OMIT_AUTOMATIC_INDEX
614 ** Return TRUE if the WHERE clause term pTerm is of a form where it
615 ** could be used with an index to access pSrc, assuming an appropriate
618 static int termCanDriveIndex(
619 WhereTerm
*pTerm
, /* WHERE clause term to check */
620 struct SrcList_item
*pSrc
, /* Table we are trying to access */
621 Bitmask notReady
/* Tables in outer loops of the join */
624 if( pTerm
->leftCursor
!=pSrc
->iCursor
) return 0;
625 if( (pTerm
->eOperator
& (WO_EQ
|WO_IS
))==0 ) return 0;
626 if( (pSrc
->fg
.jointype
& JT_LEFT
)
627 && !ExprHasProperty(pTerm
->pExpr
, EP_FromJoin
)
628 && (pTerm
->eOperator
& WO_IS
)
630 /* Cannot use an IS term from the WHERE clause as an index driver for
631 ** the RHS of a LEFT JOIN. Such a term can only be used if it is from
635 if( (pTerm
->prereqRight
& notReady
)!=0 ) return 0;
636 if( pTerm
->u
.leftColumn
<0 ) return 0;
637 aff
= pSrc
->pTab
->aCol
[pTerm
->u
.leftColumn
].affinity
;
638 if( !sqlite3IndexAffinityOk(pTerm
->pExpr
, aff
) ) return 0;
639 testcase( pTerm
->pExpr
->op
==TK_IS
);
645 #ifndef SQLITE_OMIT_AUTOMATIC_INDEX
647 ** Generate code to construct the Index object for an automatic index
648 ** and to set up the WhereLevel object pLevel so that the code generator
649 ** makes use of the automatic index.
651 static void constructAutomaticIndex(
652 Parse
*pParse
, /* The parsing context */
653 WhereClause
*pWC
, /* The WHERE clause */
654 struct SrcList_item
*pSrc
, /* The FROM clause term to get the next index */
655 Bitmask notReady
, /* Mask of cursors that are not available */
656 WhereLevel
*pLevel
/* Write new index here */
658 int nKeyCol
; /* Number of columns in the constructed index */
659 WhereTerm
*pTerm
; /* A single term of the WHERE clause */
660 WhereTerm
*pWCEnd
; /* End of pWC->a[] */
661 Index
*pIdx
; /* Object describing the transient index */
662 Vdbe
*v
; /* Prepared statement under construction */
663 int addrInit
; /* Address of the initialization bypass jump */
664 Table
*pTable
; /* The table being indexed */
665 int addrTop
; /* Top of the index fill loop */
666 int regRecord
; /* Register holding an index record */
667 int n
; /* Column counter */
668 int i
; /* Loop counter */
669 int mxBitCol
; /* Maximum column in pSrc->colUsed */
670 CollSeq
*pColl
; /* Collating sequence to on a column */
671 WhereLoop
*pLoop
; /* The Loop object */
672 char *zNotUsed
; /* Extra space on the end of pIdx */
673 Bitmask idxCols
; /* Bitmap of columns used for indexing */
674 Bitmask extraCols
; /* Bitmap of additional columns */
675 u8 sentWarning
= 0; /* True if a warnning has been issued */
676 Expr
*pPartial
= 0; /* Partial Index Expression */
677 int iContinue
= 0; /* Jump here to skip excluded rows */
678 struct SrcList_item
*pTabItem
; /* FROM clause term being indexed */
679 int addrCounter
= 0; /* Address where integer counter is initialized */
680 int regBase
; /* Array of registers where record is assembled */
682 /* Generate code to skip over the creation and initialization of the
683 ** transient index on 2nd and subsequent iterations of the loop. */
686 addrInit
= sqlite3VdbeAddOp0(v
, OP_Once
); VdbeCoverage(v
);
688 /* Count the number of columns that will be added to the index
689 ** and used to match WHERE clause constraints */
692 pWCEnd
= &pWC
->a
[pWC
->nTerm
];
693 pLoop
= pLevel
->pWLoop
;
695 for(pTerm
=pWC
->a
; pTerm
<pWCEnd
; pTerm
++){
696 Expr
*pExpr
= pTerm
->pExpr
;
697 assert( !ExprHasProperty(pExpr
, EP_FromJoin
) /* prereq always non-zero */
698 || pExpr
->iRightJoinTable
!=pSrc
->iCursor
/* for the right-hand */
699 || pLoop
->prereq
!=0 ); /* table of a LEFT JOIN */
701 && (pTerm
->wtFlags
& TERM_VIRTUAL
)==0
702 && !ExprHasProperty(pExpr
, EP_FromJoin
)
703 && sqlite3ExprIsTableConstant(pExpr
, pSrc
->iCursor
) ){
704 pPartial
= sqlite3ExprAnd(pParse
->db
, pPartial
,
705 sqlite3ExprDup(pParse
->db
, pExpr
, 0));
707 if( termCanDriveIndex(pTerm
, pSrc
, notReady
) ){
708 int iCol
= pTerm
->u
.leftColumn
;
709 Bitmask cMask
= iCol
>=BMS
? MASKBIT(BMS
-1) : MASKBIT(iCol
);
710 testcase( iCol
==BMS
);
711 testcase( iCol
==BMS
-1 );
713 sqlite3_log(SQLITE_WARNING_AUTOINDEX
,
714 "automatic index on %s(%s)", pTable
->zName
,
715 pTable
->aCol
[iCol
].zName
);
718 if( (idxCols
& cMask
)==0 ){
719 if( whereLoopResize(pParse
->db
, pLoop
, nKeyCol
+1) ){
720 goto end_auto_index_create
;
722 pLoop
->aLTerm
[nKeyCol
++] = pTerm
;
728 pLoop
->u
.btree
.nEq
= pLoop
->nLTerm
= nKeyCol
;
729 pLoop
->wsFlags
= WHERE_COLUMN_EQ
| WHERE_IDX_ONLY
| WHERE_INDEXED
732 /* Count the number of additional columns needed to create a
733 ** covering index. A "covering index" is an index that contains all
734 ** columns that are needed by the query. With a covering index, the
735 ** original table never needs to be accessed. Automatic indices must
736 ** be a covering index because the index will not be updated if the
737 ** original table changes and the index and table cannot both be used
738 ** if they go out of sync.
740 extraCols
= pSrc
->colUsed
& (~idxCols
| MASKBIT(BMS
-1));
741 mxBitCol
= MIN(BMS
-1,pTable
->nCol
);
742 testcase( pTable
->nCol
==BMS
-1 );
743 testcase( pTable
->nCol
==BMS
-2 );
744 for(i
=0; i
<mxBitCol
; i
++){
745 if( extraCols
& MASKBIT(i
) ) nKeyCol
++;
747 if( pSrc
->colUsed
& MASKBIT(BMS
-1) ){
748 nKeyCol
+= pTable
->nCol
- BMS
+ 1;
751 /* Construct the Index object to describe this index */
752 pIdx
= sqlite3AllocateIndexObject(pParse
->db
, nKeyCol
+1, 0, &zNotUsed
);
753 if( pIdx
==0 ) goto end_auto_index_create
;
754 pLoop
->u
.btree
.pIndex
= pIdx
;
755 pIdx
->zName
= "auto-index";
756 pIdx
->pTable
= pTable
;
759 for(pTerm
=pWC
->a
; pTerm
<pWCEnd
; pTerm
++){
760 if( termCanDriveIndex(pTerm
, pSrc
, notReady
) ){
761 int iCol
= pTerm
->u
.leftColumn
;
762 Bitmask cMask
= iCol
>=BMS
? MASKBIT(BMS
-1) : MASKBIT(iCol
);
763 testcase( iCol
==BMS
-1 );
764 testcase( iCol
==BMS
);
765 if( (idxCols
& cMask
)==0 ){
766 Expr
*pX
= pTerm
->pExpr
;
768 pIdx
->aiColumn
[n
] = pTerm
->u
.leftColumn
;
769 pColl
= sqlite3BinaryCompareCollSeq(pParse
, pX
->pLeft
, pX
->pRight
);
770 pIdx
->azColl
[n
] = pColl
? pColl
->zName
: sqlite3StrBINARY
;
775 assert( (u32
)n
==pLoop
->u
.btree
.nEq
);
777 /* Add additional columns needed to make the automatic index into
778 ** a covering index */
779 for(i
=0; i
<mxBitCol
; i
++){
780 if( extraCols
& MASKBIT(i
) ){
781 pIdx
->aiColumn
[n
] = i
;
782 pIdx
->azColl
[n
] = sqlite3StrBINARY
;
786 if( pSrc
->colUsed
& MASKBIT(BMS
-1) ){
787 for(i
=BMS
-1; i
<pTable
->nCol
; i
++){
788 pIdx
->aiColumn
[n
] = i
;
789 pIdx
->azColl
[n
] = sqlite3StrBINARY
;
793 assert( n
==nKeyCol
);
794 pIdx
->aiColumn
[n
] = XN_ROWID
;
795 pIdx
->azColl
[n
] = sqlite3StrBINARY
;
797 /* Create the automatic index */
798 assert( pLevel
->iIdxCur
>=0 );
799 pLevel
->iIdxCur
= pParse
->nTab
++;
800 sqlite3VdbeAddOp2(v
, OP_OpenAutoindex
, pLevel
->iIdxCur
, nKeyCol
+1);
801 sqlite3VdbeSetP4KeyInfo(pParse
, pIdx
);
802 VdbeComment((v
, "for %s", pTable
->zName
));
804 /* Fill the automatic index with content */
805 sqlite3ExprCachePush(pParse
);
806 pTabItem
= &pWC
->pWInfo
->pTabList
->a
[pLevel
->iFrom
];
807 if( pTabItem
->fg
.viaCoroutine
){
808 int regYield
= pTabItem
->regReturn
;
809 addrCounter
= sqlite3VdbeAddOp2(v
, OP_Integer
, 0, 0);
810 sqlite3VdbeAddOp3(v
, OP_InitCoroutine
, regYield
, 0, pTabItem
->addrFillSub
);
811 addrTop
= sqlite3VdbeAddOp1(v
, OP_Yield
, regYield
);
813 VdbeComment((v
, "next row of \"%s\"", pTabItem
->pTab
->zName
));
815 addrTop
= sqlite3VdbeAddOp1(v
, OP_Rewind
, pLevel
->iTabCur
); VdbeCoverage(v
);
818 iContinue
= sqlite3VdbeMakeLabel(v
);
819 sqlite3ExprIfFalse(pParse
, pPartial
, iContinue
, SQLITE_JUMPIFNULL
);
820 pLoop
->wsFlags
|= WHERE_PARTIALIDX
;
822 regRecord
= sqlite3GetTempReg(pParse
);
823 regBase
= sqlite3GenerateIndexKey(
824 pParse
, pIdx
, pLevel
->iTabCur
, regRecord
, 0, 0, 0, 0
826 sqlite3VdbeAddOp2(v
, OP_IdxInsert
, pLevel
->iIdxCur
, regRecord
);
827 sqlite3VdbeChangeP5(v
, OPFLAG_USESEEKRESULT
);
828 if( pPartial
) sqlite3VdbeResolveLabel(v
, iContinue
);
829 if( pTabItem
->fg
.viaCoroutine
){
830 sqlite3VdbeChangeP2(v
, addrCounter
, regBase
+n
);
831 testcase( pParse
->db
->mallocFailed
);
832 translateColumnToCopy(pParse
, addrTop
, pLevel
->iTabCur
,
833 pTabItem
->regResult
, 1);
834 sqlite3VdbeGoto(v
, addrTop
);
835 pTabItem
->fg
.viaCoroutine
= 0;
837 sqlite3VdbeAddOp2(v
, OP_Next
, pLevel
->iTabCur
, addrTop
+1); VdbeCoverage(v
);
839 sqlite3VdbeChangeP5(v
, SQLITE_STMTSTATUS_AUTOINDEX
);
840 sqlite3VdbeJumpHere(v
, addrTop
);
841 sqlite3ReleaseTempReg(pParse
, regRecord
);
842 sqlite3ExprCachePop(pParse
);
844 /* Jump here when skipping the initialization */
845 sqlite3VdbeJumpHere(v
, addrInit
);
847 end_auto_index_create
:
848 sqlite3ExprDelete(pParse
->db
, pPartial
);
850 #endif /* SQLITE_OMIT_AUTOMATIC_INDEX */
852 #ifndef SQLITE_OMIT_VIRTUALTABLE
854 ** Allocate and populate an sqlite3_index_info structure. It is the
855 ** responsibility of the caller to eventually release the structure
856 ** by passing the pointer returned by this function to sqlite3_free().
858 static sqlite3_index_info
*allocateIndexInfo(
859 Parse
*pParse
, /* The parsing context */
860 WhereClause
*pWC
, /* The WHERE clause being analyzed */
861 Bitmask mUnusable
, /* Ignore terms with these prereqs */
862 struct SrcList_item
*pSrc
, /* The FROM clause term that is the vtab */
863 ExprList
*pOrderBy
, /* The ORDER BY clause */
864 u16
*pmNoOmit
/* Mask of terms not to omit */
868 struct sqlite3_index_constraint
*pIdxCons
;
869 struct sqlite3_index_orderby
*pIdxOrderBy
;
870 struct sqlite3_index_constraint_usage
*pUsage
;
871 struct HiddenIndexInfo
*pHidden
;
874 sqlite3_index_info
*pIdxInfo
;
877 /* Count the number of possible WHERE clause constraints referring
878 ** to this virtual table */
879 for(i
=nTerm
=0, pTerm
=pWC
->a
; i
<pWC
->nTerm
; i
++, pTerm
++){
880 if( pTerm
->leftCursor
!= pSrc
->iCursor
) continue;
881 if( pTerm
->prereqRight
& mUnusable
) continue;
882 assert( IsPowerOfTwo(pTerm
->eOperator
& ~WO_EQUIV
) );
883 testcase( pTerm
->eOperator
& WO_IN
);
884 testcase( pTerm
->eOperator
& WO_ISNULL
);
885 testcase( pTerm
->eOperator
& WO_IS
);
886 testcase( pTerm
->eOperator
& WO_ALL
);
887 if( (pTerm
->eOperator
& ~(WO_EQUIV
))==0 ) continue;
888 if( pTerm
->wtFlags
& TERM_VNULL
) continue;
889 assert( pTerm
->u
.leftColumn
>=(-1) );
893 /* If the ORDER BY clause contains only columns in the current
894 ** virtual table then allocate space for the aOrderBy part of
895 ** the sqlite3_index_info structure.
899 int n
= pOrderBy
->nExpr
;
901 Expr
*pExpr
= pOrderBy
->a
[i
].pExpr
;
902 if( pExpr
->op
!=TK_COLUMN
|| pExpr
->iTable
!=pSrc
->iCursor
) break;
909 /* Allocate the sqlite3_index_info structure
911 pIdxInfo
= sqlite3DbMallocZero(pParse
->db
, sizeof(*pIdxInfo
)
912 + (sizeof(*pIdxCons
) + sizeof(*pUsage
))*nTerm
913 + sizeof(*pIdxOrderBy
)*nOrderBy
+ sizeof(*pHidden
) );
915 sqlite3ErrorMsg(pParse
, "out of memory");
919 /* Initialize the structure. The sqlite3_index_info structure contains
920 ** many fields that are declared "const" to prevent xBestIndex from
921 ** changing them. We have to do some funky casting in order to
922 ** initialize those fields.
924 pHidden
= (struct HiddenIndexInfo
*)&pIdxInfo
[1];
925 pIdxCons
= (struct sqlite3_index_constraint
*)&pHidden
[1];
926 pIdxOrderBy
= (struct sqlite3_index_orderby
*)&pIdxCons
[nTerm
];
927 pUsage
= (struct sqlite3_index_constraint_usage
*)&pIdxOrderBy
[nOrderBy
];
928 *(int*)&pIdxInfo
->nConstraint
= nTerm
;
929 *(int*)&pIdxInfo
->nOrderBy
= nOrderBy
;
930 *(struct sqlite3_index_constraint
**)&pIdxInfo
->aConstraint
= pIdxCons
;
931 *(struct sqlite3_index_orderby
**)&pIdxInfo
->aOrderBy
= pIdxOrderBy
;
932 *(struct sqlite3_index_constraint_usage
**)&pIdxInfo
->aConstraintUsage
=
936 pHidden
->pParse
= pParse
;
937 for(i
=j
=0, pTerm
=pWC
->a
; i
<pWC
->nTerm
; i
++, pTerm
++){
939 if( pTerm
->leftCursor
!= pSrc
->iCursor
) continue;
940 if( pTerm
->prereqRight
& mUnusable
) continue;
941 assert( IsPowerOfTwo(pTerm
->eOperator
& ~WO_EQUIV
) );
942 testcase( pTerm
->eOperator
& WO_IN
);
943 testcase( pTerm
->eOperator
& WO_IS
);
944 testcase( pTerm
->eOperator
& WO_ISNULL
);
945 testcase( pTerm
->eOperator
& WO_ALL
);
946 if( (pTerm
->eOperator
& ~(WO_EQUIV
))==0 ) continue;
947 if( pTerm
->wtFlags
& TERM_VNULL
) continue;
948 assert( pTerm
->u
.leftColumn
>=(-1) );
949 pIdxCons
[j
].iColumn
= pTerm
->u
.leftColumn
;
950 pIdxCons
[j
].iTermOffset
= i
;
951 op
= pTerm
->eOperator
& WO_ALL
;
952 if( op
==WO_IN
) op
= WO_EQ
;
954 pIdxCons
[j
].op
= pTerm
->eMatchOp
;
955 }else if( op
& (WO_ISNULL
|WO_IS
) ){
957 pIdxCons
[j
].op
= SQLITE_INDEX_CONSTRAINT_ISNULL
;
959 pIdxCons
[j
].op
= SQLITE_INDEX_CONSTRAINT_IS
;
962 pIdxCons
[j
].op
= (u8
)op
;
963 /* The direct assignment in the previous line is possible only because
964 ** the WO_ and SQLITE_INDEX_CONSTRAINT_ codes are identical. The
965 ** following asserts verify this fact. */
966 assert( WO_EQ
==SQLITE_INDEX_CONSTRAINT_EQ
);
967 assert( WO_LT
==SQLITE_INDEX_CONSTRAINT_LT
);
968 assert( WO_LE
==SQLITE_INDEX_CONSTRAINT_LE
);
969 assert( WO_GT
==SQLITE_INDEX_CONSTRAINT_GT
);
970 assert( WO_GE
==SQLITE_INDEX_CONSTRAINT_GE
);
971 assert( pTerm
->eOperator
&(WO_IN
|WO_EQ
|WO_LT
|WO_LE
|WO_GT
|WO_GE
|WO_AUX
) );
973 if( op
& (WO_LT
|WO_LE
|WO_GT
|WO_GE
)
974 && sqlite3ExprIsVector(pTerm
->pExpr
->pRight
)
976 if( i
<16 ) mNoOmit
|= (1 << i
);
977 if( op
==WO_LT
) pIdxCons
[j
].op
= WO_LE
;
978 if( op
==WO_GT
) pIdxCons
[j
].op
= WO_GE
;
984 for(i
=0; i
<nOrderBy
; i
++){
985 Expr
*pExpr
= pOrderBy
->a
[i
].pExpr
;
986 pIdxOrderBy
[i
].iColumn
= pExpr
->iColumn
;
987 pIdxOrderBy
[i
].desc
= pOrderBy
->a
[i
].sortOrder
;
995 ** The table object reference passed as the second argument to this function
996 ** must represent a virtual table. This function invokes the xBestIndex()
997 ** method of the virtual table with the sqlite3_index_info object that
998 ** comes in as the 3rd argument to this function.
1000 ** If an error occurs, pParse is populated with an error message and a
1001 ** non-zero value is returned. Otherwise, 0 is returned and the output
1002 ** part of the sqlite3_index_info structure is left populated.
1004 ** Whether or not an error is returned, it is the responsibility of the
1005 ** caller to eventually free p->idxStr if p->needToFreeIdxStr indicates
1006 ** that this is required.
1008 static int vtabBestIndex(Parse
*pParse
, Table
*pTab
, sqlite3_index_info
*p
){
1009 sqlite3_vtab
*pVtab
= sqlite3GetVTable(pParse
->db
, pTab
)->pVtab
;
1012 TRACE_IDX_INPUTS(p
);
1013 rc
= pVtab
->pModule
->xBestIndex(pVtab
, p
);
1014 TRACE_IDX_OUTPUTS(p
);
1016 if( rc
!=SQLITE_OK
){
1017 if( rc
==SQLITE_NOMEM
){
1018 sqlite3OomFault(pParse
->db
);
1019 }else if( !pVtab
->zErrMsg
){
1020 sqlite3ErrorMsg(pParse
, "%s", sqlite3ErrStr(rc
));
1022 sqlite3ErrorMsg(pParse
, "%s", pVtab
->zErrMsg
);
1025 sqlite3_free(pVtab
->zErrMsg
);
1029 /* This error is now caught by the caller.
1030 ** Search for "xBestIndex malfunction" below */
1031 for(i
=0; i
<p
->nConstraint
; i
++){
1032 if( !p
->aConstraint
[i
].usable
&& p
->aConstraintUsage
[i
].argvIndex
>0 ){
1033 sqlite3ErrorMsg(pParse
,
1034 "table %s: xBestIndex returned an invalid plan", pTab
->zName
);
1039 return pParse
->nErr
;
1041 #endif /* !defined(SQLITE_OMIT_VIRTUALTABLE) */
1043 #ifdef SQLITE_ENABLE_STAT3_OR_STAT4
1045 ** Estimate the location of a particular key among all keys in an
1046 ** index. Store the results in aStat as follows:
1048 ** aStat[0] Est. number of rows less than pRec
1049 ** aStat[1] Est. number of rows equal to pRec
1051 ** Return the index of the sample that is the smallest sample that
1052 ** is greater than or equal to pRec. Note that this index is not an index
1053 ** into the aSample[] array - it is an index into a virtual set of samples
1054 ** based on the contents of aSample[] and the number of fields in record
1057 static int whereKeyStats(
1058 Parse
*pParse
, /* Database connection */
1059 Index
*pIdx
, /* Index to consider domain of */
1060 UnpackedRecord
*pRec
, /* Vector of values to consider */
1061 int roundUp
, /* Round up if true. Round down if false */
1062 tRowcnt
*aStat
/* OUT: stats written here */
1064 IndexSample
*aSample
= pIdx
->aSample
;
1065 int iCol
; /* Index of required stats in anEq[] etc. */
1066 int i
; /* Index of first sample >= pRec */
1067 int iSample
; /* Smallest sample larger than or equal to pRec */
1068 int iMin
= 0; /* Smallest sample not yet tested */
1069 int iTest
; /* Next sample to test */
1070 int res
; /* Result of comparison operation */
1071 int nField
; /* Number of fields in pRec */
1072 tRowcnt iLower
= 0; /* anLt[] + anEq[] of largest sample pRec is > */
1074 #ifndef SQLITE_DEBUG
1075 UNUSED_PARAMETER( pParse
);
1078 assert( pIdx
->nSample
>0 );
1079 assert( pRec
->nField
>0 && pRec
->nField
<=pIdx
->nSampleCol
);
1081 /* Do a binary search to find the first sample greater than or equal
1082 ** to pRec. If pRec contains a single field, the set of samples to search
1083 ** is simply the aSample[] array. If the samples in aSample[] contain more
1084 ** than one fields, all fields following the first are ignored.
1086 ** If pRec contains N fields, where N is more than one, then as well as the
1087 ** samples in aSample[] (truncated to N fields), the search also has to
1088 ** consider prefixes of those samples. For example, if the set of samples
1091 ** aSample[0] = (a, 5)
1092 ** aSample[1] = (a, 10)
1093 ** aSample[2] = (b, 5)
1094 ** aSample[3] = (c, 100)
1095 ** aSample[4] = (c, 105)
1097 ** Then the search space should ideally be the samples above and the
1098 ** unique prefixes [a], [b] and [c]. But since that is hard to organize,
1099 ** the code actually searches this set:
1112 ** For each sample in the aSample[] array, N samples are present in the
1113 ** effective sample array. In the above, samples 0 and 1 are based on
1114 ** sample aSample[0]. Samples 2 and 3 on aSample[1] etc.
1116 ** Often, sample i of each block of N effective samples has (i+1) fields.
1117 ** Except, each sample may be extended to ensure that it is greater than or
1118 ** equal to the previous sample in the array. For example, in the above,
1119 ** sample 2 is the first sample of a block of N samples, so at first it
1120 ** appears that it should be 1 field in size. However, that would make it
1121 ** smaller than sample 1, so the binary search would not work. As a result,
1122 ** it is extended to two fields. The duplicates that this creates do not
1123 ** cause any problems.
1125 nField
= pRec
->nField
;
1127 iSample
= pIdx
->nSample
* nField
;
1129 int iSamp
; /* Index in aSample[] of test sample */
1130 int n
; /* Number of fields in test sample */
1132 iTest
= (iMin
+iSample
)/2;
1133 iSamp
= iTest
/ nField
;
1135 /* The proposed effective sample is a prefix of sample aSample[iSamp].
1136 ** Specifically, the shortest prefix of at least (1 + iTest%nField)
1137 ** fields that is greater than the previous effective sample. */
1138 for(n
=(iTest
% nField
) + 1; n
<nField
; n
++){
1139 if( aSample
[iSamp
-1].anLt
[n
-1]!=aSample
[iSamp
].anLt
[n
-1] ) break;
1146 res
= sqlite3VdbeRecordCompare(aSample
[iSamp
].n
, aSample
[iSamp
].p
, pRec
);
1148 iLower
= aSample
[iSamp
].anLt
[n
-1] + aSample
[iSamp
].anEq
[n
-1];
1150 }else if( res
==0 && n
<nField
){
1151 iLower
= aSample
[iSamp
].anLt
[n
-1];
1158 }while( res
&& iMin
<iSample
);
1159 i
= iSample
/ nField
;
1162 /* The following assert statements check that the binary search code
1163 ** above found the right answer. This block serves no purpose other
1164 ** than to invoke the asserts. */
1165 if( pParse
->db
->mallocFailed
==0 ){
1167 /* If (res==0) is true, then pRec must be equal to sample i. */
1168 assert( i
<pIdx
->nSample
);
1169 assert( iCol
==nField
-1 );
1170 pRec
->nField
= nField
;
1171 assert( 0==sqlite3VdbeRecordCompare(aSample
[i
].n
, aSample
[i
].p
, pRec
)
1172 || pParse
->db
->mallocFailed
1175 /* Unless i==pIdx->nSample, indicating that pRec is larger than
1176 ** all samples in the aSample[] array, pRec must be smaller than the
1177 ** (iCol+1) field prefix of sample i. */
1178 assert( i
<=pIdx
->nSample
&& i
>=0 );
1179 pRec
->nField
= iCol
+1;
1180 assert( i
==pIdx
->nSample
1181 || sqlite3VdbeRecordCompare(aSample
[i
].n
, aSample
[i
].p
, pRec
)>0
1182 || pParse
->db
->mallocFailed
);
1184 /* if i==0 and iCol==0, then record pRec is smaller than all samples
1185 ** in the aSample[] array. Otherwise, if (iCol>0) then pRec must
1186 ** be greater than or equal to the (iCol) field prefix of sample i.
1187 ** If (i>0), then pRec must also be greater than sample (i-1). */
1189 pRec
->nField
= iCol
;
1190 assert( sqlite3VdbeRecordCompare(aSample
[i
].n
, aSample
[i
].p
, pRec
)<=0
1191 || pParse
->db
->mallocFailed
);
1194 pRec
->nField
= nField
;
1195 assert( sqlite3VdbeRecordCompare(aSample
[i
-1].n
, aSample
[i
-1].p
, pRec
)<0
1196 || pParse
->db
->mallocFailed
);
1200 #endif /* ifdef SQLITE_DEBUG */
1203 /* Record pRec is equal to sample i */
1204 assert( iCol
==nField
-1 );
1205 aStat
[0] = aSample
[i
].anLt
[iCol
];
1206 aStat
[1] = aSample
[i
].anEq
[iCol
];
1208 /* At this point, the (iCol+1) field prefix of aSample[i] is the first
1209 ** sample that is greater than pRec. Or, if i==pIdx->nSample then pRec
1210 ** is larger than all samples in the array. */
1211 tRowcnt iUpper
, iGap
;
1212 if( i
>=pIdx
->nSample
){
1213 iUpper
= sqlite3LogEstToInt(pIdx
->aiRowLogEst
[0]);
1215 iUpper
= aSample
[i
].anLt
[iCol
];
1218 if( iLower
>=iUpper
){
1221 iGap
= iUpper
- iLower
;
1228 aStat
[0] = iLower
+ iGap
;
1229 aStat
[1] = pIdx
->aAvgEq
[nField
-1];
1232 /* Restore the pRec->nField value before returning. */
1233 pRec
->nField
= nField
;
1236 #endif /* SQLITE_ENABLE_STAT3_OR_STAT4 */
1239 ** If it is not NULL, pTerm is a term that provides an upper or lower
1240 ** bound on a range scan. Without considering pTerm, it is estimated
1241 ** that the scan will visit nNew rows. This function returns the number
1242 ** estimated to be visited after taking pTerm into account.
1244 ** If the user explicitly specified a likelihood() value for this term,
1245 ** then the return value is the likelihood multiplied by the number of
1246 ** input rows. Otherwise, this function assumes that an "IS NOT NULL" term
1247 ** has a likelihood of 0.50, and any other term a likelihood of 0.25.
1249 static LogEst
whereRangeAdjust(WhereTerm
*pTerm
, LogEst nNew
){
1252 if( pTerm
->truthProb
<=0 ){
1253 nRet
+= pTerm
->truthProb
;
1254 }else if( (pTerm
->wtFlags
& TERM_VNULL
)==0 ){
1255 nRet
-= 20; assert( 20==sqlite3LogEst(4) );
1262 #ifdef SQLITE_ENABLE_STAT3_OR_STAT4
1264 ** Return the affinity for a single column of an index.
1266 char sqlite3IndexColumnAffinity(sqlite3
*db
, Index
*pIdx
, int iCol
){
1267 assert( iCol
>=0 && iCol
<pIdx
->nColumn
);
1268 if( !pIdx
->zColAff
){
1269 if( sqlite3IndexAffinityStr(db
, pIdx
)==0 ) return SQLITE_AFF_BLOB
;
1271 return pIdx
->zColAff
[iCol
];
1276 #ifdef SQLITE_ENABLE_STAT3_OR_STAT4
1278 ** This function is called to estimate the number of rows visited by a
1279 ** range-scan on a skip-scan index. For example:
1281 ** CREATE INDEX i1 ON t1(a, b, c);
1282 ** SELECT * FROM t1 WHERE a=? AND c BETWEEN ? AND ?;
1284 ** Value pLoop->nOut is currently set to the estimated number of rows
1285 ** visited for scanning (a=? AND b=?). This function reduces that estimate
1286 ** by some factor to account for the (c BETWEEN ? AND ?) expression based
1287 ** on the stat4 data for the index. this scan will be peformed multiple
1288 ** times (once for each (a,b) combination that matches a=?) is dealt with
1291 ** It does this by scanning through all stat4 samples, comparing values
1292 ** extracted from pLower and pUpper with the corresponding column in each
1293 ** sample. If L and U are the number of samples found to be less than or
1294 ** equal to the values extracted from pLower and pUpper respectively, and
1295 ** N is the total number of samples, the pLoop->nOut value is adjusted
1298 ** nOut = nOut * ( min(U - L, 1) / N )
1300 ** If pLower is NULL, or a value cannot be extracted from the term, L is
1301 ** set to zero. If pUpper is NULL, or a value cannot be extracted from it,
1304 ** Normally, this function sets *pbDone to 1 before returning. However,
1305 ** if no value can be extracted from either pLower or pUpper (and so the
1306 ** estimate of the number of rows delivered remains unchanged), *pbDone
1309 ** If an error occurs, an SQLite error code is returned. Otherwise,
1312 static int whereRangeSkipScanEst(
1313 Parse
*pParse
, /* Parsing & code generating context */
1314 WhereTerm
*pLower
, /* Lower bound on the range. ex: "x>123" Might be NULL */
1315 WhereTerm
*pUpper
, /* Upper bound on the range. ex: "x<455" Might be NULL */
1316 WhereLoop
*pLoop
, /* Update the .nOut value of this loop */
1317 int *pbDone
/* Set to true if at least one expr. value extracted */
1319 Index
*p
= pLoop
->u
.btree
.pIndex
;
1320 int nEq
= pLoop
->u
.btree
.nEq
;
1321 sqlite3
*db
= pParse
->db
;
1323 int nUpper
= p
->nSample
+1;
1325 u8 aff
= sqlite3IndexColumnAffinity(db
, p
, nEq
);
1328 sqlite3_value
*p1
= 0; /* Value extracted from pLower */
1329 sqlite3_value
*p2
= 0; /* Value extracted from pUpper */
1330 sqlite3_value
*pVal
= 0; /* Value extracted from record */
1332 pColl
= sqlite3LocateCollSeq(pParse
, p
->azColl
[nEq
]);
1334 rc
= sqlite3Stat4ValueFromExpr(pParse
, pLower
->pExpr
->pRight
, aff
, &p1
);
1337 if( pUpper
&& rc
==SQLITE_OK
){
1338 rc
= sqlite3Stat4ValueFromExpr(pParse
, pUpper
->pExpr
->pRight
, aff
, &p2
);
1339 nUpper
= p2
? 0 : p
->nSample
;
1345 for(i
=0; rc
==SQLITE_OK
&& i
<p
->nSample
; i
++){
1346 rc
= sqlite3Stat4Column(db
, p
->aSample
[i
].p
, p
->aSample
[i
].n
, nEq
, &pVal
);
1347 if( rc
==SQLITE_OK
&& p1
){
1348 int res
= sqlite3MemCompare(p1
, pVal
, pColl
);
1349 if( res
>=0 ) nLower
++;
1351 if( rc
==SQLITE_OK
&& p2
){
1352 int res
= sqlite3MemCompare(p2
, pVal
, pColl
);
1353 if( res
>=0 ) nUpper
++;
1356 nDiff
= (nUpper
- nLower
);
1357 if( nDiff
<=0 ) nDiff
= 1;
1359 /* If there is both an upper and lower bound specified, and the
1360 ** comparisons indicate that they are close together, use the fallback
1361 ** method (assume that the scan visits 1/64 of the rows) for estimating
1362 ** the number of rows visited. Otherwise, estimate the number of rows
1363 ** using the method described in the header comment for this function. */
1364 if( nDiff
!=1 || pUpper
==0 || pLower
==0 ){
1365 int nAdjust
= (sqlite3LogEst(p
->nSample
) - sqlite3LogEst(nDiff
));
1366 pLoop
->nOut
-= nAdjust
;
1368 WHERETRACE(0x10, ("range skip-scan regions: %u..%u adjust=%d est=%d\n",
1369 nLower
, nUpper
, nAdjust
*-1, pLoop
->nOut
));
1373 assert( *pbDone
==0 );
1376 sqlite3ValueFree(p1
);
1377 sqlite3ValueFree(p2
);
1378 sqlite3ValueFree(pVal
);
1382 #endif /* SQLITE_ENABLE_STAT3_OR_STAT4 */
1385 ** This function is used to estimate the number of rows that will be visited
1386 ** by scanning an index for a range of values. The range may have an upper
1387 ** bound, a lower bound, or both. The WHERE clause terms that set the upper
1388 ** and lower bounds are represented by pLower and pUpper respectively. For
1389 ** example, assuming that index p is on t1(a):
1391 ** ... FROM t1 WHERE a > ? AND a < ? ...
1396 ** If either of the upper or lower bound is not present, then NULL is passed in
1397 ** place of the corresponding WhereTerm.
1399 ** The value in (pBuilder->pNew->u.btree.nEq) is the number of the index
1400 ** column subject to the range constraint. Or, equivalently, the number of
1401 ** equality constraints optimized by the proposed index scan. For example,
1402 ** assuming index p is on t1(a, b), and the SQL query is:
1404 ** ... FROM t1 WHERE a = ? AND b > ? AND b < ? ...
1406 ** then nEq is set to 1 (as the range restricted column, b, is the second
1407 ** left-most column of the index). Or, if the query is:
1409 ** ... FROM t1 WHERE a > ? AND a < ? ...
1411 ** then nEq is set to 0.
1413 ** When this function is called, *pnOut is set to the sqlite3LogEst() of the
1414 ** number of rows that the index scan is expected to visit without
1415 ** considering the range constraints. If nEq is 0, then *pnOut is the number of
1416 ** rows in the index. Assuming no error occurs, *pnOut is adjusted (reduced)
1417 ** to account for the range constraints pLower and pUpper.
1419 ** In the absence of sqlite_stat4 ANALYZE data, or if such data cannot be
1420 ** used, a single range inequality reduces the search space by a factor of 4.
1421 ** and a pair of constraints (x>? AND x<?) reduces the expected number of
1422 ** rows visited by a factor of 64.
1424 static int whereRangeScanEst(
1425 Parse
*pParse
, /* Parsing & code generating context */
1426 WhereLoopBuilder
*pBuilder
,
1427 WhereTerm
*pLower
, /* Lower bound on the range. ex: "x>123" Might be NULL */
1428 WhereTerm
*pUpper
, /* Upper bound on the range. ex: "x<455" Might be NULL */
1429 WhereLoop
*pLoop
/* Modify the .nOut and maybe .rRun fields */
1432 int nOut
= pLoop
->nOut
;
1435 #ifdef SQLITE_ENABLE_STAT3_OR_STAT4
1436 Index
*p
= pLoop
->u
.btree
.pIndex
;
1437 int nEq
= pLoop
->u
.btree
.nEq
;
1439 if( p
->nSample
>0 && nEq
<p
->nSampleCol
){
1440 if( nEq
==pBuilder
->nRecValid
){
1441 UnpackedRecord
*pRec
= pBuilder
->pRec
;
1443 int nBtm
= pLoop
->u
.btree
.nBtm
;
1444 int nTop
= pLoop
->u
.btree
.nTop
;
1446 /* Variable iLower will be set to the estimate of the number of rows in
1447 ** the index that are less than the lower bound of the range query. The
1448 ** lower bound being the concatenation of $P and $L, where $P is the
1449 ** key-prefix formed by the nEq values matched against the nEq left-most
1450 ** columns of the index, and $L is the value in pLower.
1452 ** Or, if pLower is NULL or $L cannot be extracted from it (because it
1453 ** is not a simple variable or literal value), the lower bound of the
1454 ** range is $P. Due to a quirk in the way whereKeyStats() works, even
1455 ** if $L is available, whereKeyStats() is called for both ($P) and
1456 ** ($P:$L) and the larger of the two returned values is used.
1458 ** Similarly, iUpper is to be set to the estimate of the number of rows
1459 ** less than the upper bound of the range query. Where the upper bound
1460 ** is either ($P) or ($P:$U). Again, even if $U is available, both values
1461 ** of iUpper are requested of whereKeyStats() and the smaller used.
1463 ** The number of rows between the two bounds is then just iUpper-iLower.
1465 tRowcnt iLower
; /* Rows less than the lower bound */
1466 tRowcnt iUpper
; /* Rows less than the upper bound */
1467 int iLwrIdx
= -2; /* aSample[] for the lower bound */
1468 int iUprIdx
= -1; /* aSample[] for the upper bound */
1471 testcase( pRec
->nField
!=pBuilder
->nRecValid
);
1472 pRec
->nField
= pBuilder
->nRecValid
;
1474 /* Determine iLower and iUpper using ($P) only. */
1477 iUpper
= p
->nRowEst0
;
1479 /* Note: this call could be optimized away - since the same values must
1480 ** have been requested when testing key $P in whereEqualScanEst(). */
1481 whereKeyStats(pParse
, p
, pRec
, 0, a
);
1483 iUpper
= a
[0] + a
[1];
1486 assert( pLower
==0 || (pLower
->eOperator
& (WO_GT
|WO_GE
))!=0 );
1487 assert( pUpper
==0 || (pUpper
->eOperator
& (WO_LT
|WO_LE
))!=0 );
1488 assert( p
->aSortOrder
!=0 );
1489 if( p
->aSortOrder
[nEq
] ){
1490 /* The roles of pLower and pUpper are swapped for a DESC index */
1491 SWAP(WhereTerm
*, pLower
, pUpper
);
1492 SWAP(int, nBtm
, nTop
);
1495 /* If possible, improve on the iLower estimate using ($P:$L). */
1497 int n
; /* Values extracted from pExpr */
1498 Expr
*pExpr
= pLower
->pExpr
->pRight
;
1499 rc
= sqlite3Stat4ProbeSetValue(pParse
, p
, &pRec
, pExpr
, nBtm
, nEq
, &n
);
1500 if( rc
==SQLITE_OK
&& n
){
1502 u16 mask
= WO_GT
|WO_LE
;
1503 if( sqlite3ExprVectorSize(pExpr
)>n
) mask
= (WO_LE
|WO_LT
);
1504 iLwrIdx
= whereKeyStats(pParse
, p
, pRec
, 0, a
);
1505 iNew
= a
[0] + ((pLower
->eOperator
& mask
) ? a
[1] : 0);
1506 if( iNew
>iLower
) iLower
= iNew
;
1512 /* If possible, improve on the iUpper estimate using ($P:$U). */
1514 int n
; /* Values extracted from pExpr */
1515 Expr
*pExpr
= pUpper
->pExpr
->pRight
;
1516 rc
= sqlite3Stat4ProbeSetValue(pParse
, p
, &pRec
, pExpr
, nTop
, nEq
, &n
);
1517 if( rc
==SQLITE_OK
&& n
){
1519 u16 mask
= WO_GT
|WO_LE
;
1520 if( sqlite3ExprVectorSize(pExpr
)>n
) mask
= (WO_LE
|WO_LT
);
1521 iUprIdx
= whereKeyStats(pParse
, p
, pRec
, 1, a
);
1522 iNew
= a
[0] + ((pUpper
->eOperator
& mask
) ? a
[1] : 0);
1523 if( iNew
<iUpper
) iUpper
= iNew
;
1529 pBuilder
->pRec
= pRec
;
1530 if( rc
==SQLITE_OK
){
1531 if( iUpper
>iLower
){
1532 nNew
= sqlite3LogEst(iUpper
- iLower
);
1533 /* TUNING: If both iUpper and iLower are derived from the same
1534 ** sample, then assume they are 4x more selective. This brings
1535 ** the estimated selectivity more in line with what it would be
1536 ** if estimated without the use of STAT3/4 tables. */
1537 if( iLwrIdx
==iUprIdx
) nNew
-= 20; assert( 20==sqlite3LogEst(4) );
1539 nNew
= 10; assert( 10==sqlite3LogEst(2) );
1544 WHERETRACE(0x10, ("STAT4 range scan: %u..%u est=%d\n",
1545 (u32
)iLower
, (u32
)iUpper
, nOut
));
1549 rc
= whereRangeSkipScanEst(pParse
, pLower
, pUpper
, pLoop
, &bDone
);
1550 if( bDone
) return rc
;
1554 UNUSED_PARAMETER(pParse
);
1555 UNUSED_PARAMETER(pBuilder
);
1556 assert( pLower
|| pUpper
);
1558 assert( pUpper
==0 || (pUpper
->wtFlags
& TERM_VNULL
)==0 );
1559 nNew
= whereRangeAdjust(pLower
, nOut
);
1560 nNew
= whereRangeAdjust(pUpper
, nNew
);
1562 /* TUNING: If there is both an upper and lower limit and neither limit
1563 ** has an application-defined likelihood(), assume the range is
1564 ** reduced by an additional 75%. This means that, by default, an open-ended
1565 ** range query (e.g. col > ?) is assumed to match 1/4 of the rows in the
1566 ** index. While a closed range (e.g. col BETWEEN ? AND ?) is estimated to
1567 ** match 1/64 of the index. */
1568 if( pLower
&& pLower
->truthProb
>0 && pUpper
&& pUpper
->truthProb
>0 ){
1572 nOut
-= (pLower
!=0) + (pUpper
!=0);
1573 if( nNew
<10 ) nNew
= 10;
1574 if( nNew
<nOut
) nOut
= nNew
;
1575 #if defined(WHERETRACE_ENABLED)
1576 if( pLoop
->nOut
>nOut
){
1577 WHERETRACE(0x10,("Range scan lowers nOut from %d to %d\n",
1578 pLoop
->nOut
, nOut
));
1581 pLoop
->nOut
= (LogEst
)nOut
;
1585 #ifdef SQLITE_ENABLE_STAT3_OR_STAT4
1587 ** Estimate the number of rows that will be returned based on
1588 ** an equality constraint x=VALUE and where that VALUE occurs in
1589 ** the histogram data. This only works when x is the left-most
1590 ** column of an index and sqlite_stat3 histogram data is available
1591 ** for that index. When pExpr==NULL that means the constraint is
1592 ** "x IS NULL" instead of "x=VALUE".
1594 ** Write the estimated row count into *pnRow and return SQLITE_OK.
1595 ** If unable to make an estimate, leave *pnRow unchanged and return
1598 ** This routine can fail if it is unable to load a collating sequence
1599 ** required for string comparison, or if unable to allocate memory
1600 ** for a UTF conversion required for comparison. The error is stored
1601 ** in the pParse structure.
1603 static int whereEqualScanEst(
1604 Parse
*pParse
, /* Parsing & code generating context */
1605 WhereLoopBuilder
*pBuilder
,
1606 Expr
*pExpr
, /* Expression for VALUE in the x=VALUE constraint */
1607 tRowcnt
*pnRow
/* Write the revised row estimate here */
1609 Index
*p
= pBuilder
->pNew
->u
.btree
.pIndex
;
1610 int nEq
= pBuilder
->pNew
->u
.btree
.nEq
;
1611 UnpackedRecord
*pRec
= pBuilder
->pRec
;
1612 int rc
; /* Subfunction return code */
1613 tRowcnt a
[2]; /* Statistics */
1617 assert( nEq
<=p
->nColumn
);
1618 assert( p
->aSample
!=0 );
1619 assert( p
->nSample
>0 );
1620 assert( pBuilder
->nRecValid
<nEq
);
1622 /* If values are not available for all fields of the index to the left
1623 ** of this one, no estimate can be made. Return SQLITE_NOTFOUND. */
1624 if( pBuilder
->nRecValid
<(nEq
-1) ){
1625 return SQLITE_NOTFOUND
;
1628 /* This is an optimization only. The call to sqlite3Stat4ProbeSetValue()
1629 ** below would return the same value. */
1630 if( nEq
>=p
->nColumn
){
1635 rc
= sqlite3Stat4ProbeSetValue(pParse
, p
, &pRec
, pExpr
, 1, nEq
-1, &bOk
);
1636 pBuilder
->pRec
= pRec
;
1637 if( rc
!=SQLITE_OK
) return rc
;
1638 if( bOk
==0 ) return SQLITE_NOTFOUND
;
1639 pBuilder
->nRecValid
= nEq
;
1641 whereKeyStats(pParse
, p
, pRec
, 0, a
);
1642 WHERETRACE(0x10,("equality scan regions %s(%d): %d\n",
1643 p
->zName
, nEq
-1, (int)a
[1]));
1648 #endif /* SQLITE_ENABLE_STAT3_OR_STAT4 */
1650 #ifdef SQLITE_ENABLE_STAT3_OR_STAT4
1652 ** Estimate the number of rows that will be returned based on
1653 ** an IN constraint where the right-hand side of the IN operator
1654 ** is a list of values. Example:
1656 ** WHERE x IN (1,2,3,4)
1658 ** Write the estimated row count into *pnRow and return SQLITE_OK.
1659 ** If unable to make an estimate, leave *pnRow unchanged and return
1662 ** This routine can fail if it is unable to load a collating sequence
1663 ** required for string comparison, or if unable to allocate memory
1664 ** for a UTF conversion required for comparison. The error is stored
1665 ** in the pParse structure.
1667 static int whereInScanEst(
1668 Parse
*pParse
, /* Parsing & code generating context */
1669 WhereLoopBuilder
*pBuilder
,
1670 ExprList
*pList
, /* The value list on the RHS of "x IN (v1,v2,v3,...)" */
1671 tRowcnt
*pnRow
/* Write the revised row estimate here */
1673 Index
*p
= pBuilder
->pNew
->u
.btree
.pIndex
;
1674 i64 nRow0
= sqlite3LogEstToInt(p
->aiRowLogEst
[0]);
1675 int nRecValid
= pBuilder
->nRecValid
;
1676 int rc
= SQLITE_OK
; /* Subfunction return code */
1677 tRowcnt nEst
; /* Number of rows for a single term */
1678 tRowcnt nRowEst
= 0; /* New estimate of the number of rows */
1679 int i
; /* Loop counter */
1681 assert( p
->aSample
!=0 );
1682 for(i
=0; rc
==SQLITE_OK
&& i
<pList
->nExpr
; i
++){
1684 rc
= whereEqualScanEst(pParse
, pBuilder
, pList
->a
[i
].pExpr
, &nEst
);
1686 pBuilder
->nRecValid
= nRecValid
;
1689 if( rc
==SQLITE_OK
){
1690 if( nRowEst
> nRow0
) nRowEst
= nRow0
;
1692 WHERETRACE(0x10,("IN row estimate: est=%d\n", nRowEst
));
1694 assert( pBuilder
->nRecValid
==nRecValid
);
1697 #endif /* SQLITE_ENABLE_STAT3_OR_STAT4 */
1700 #ifdef WHERETRACE_ENABLED
1702 ** Print the content of a WhereTerm object
1704 static void whereTermPrint(WhereTerm
*pTerm
, int iTerm
){
1706 sqlite3DebugPrintf("TERM-%-3d NULL\n", iTerm
);
1710 memcpy(zType
, "...", 4);
1711 if( pTerm
->wtFlags
& TERM_VIRTUAL
) zType
[0] = 'V';
1712 if( pTerm
->eOperator
& WO_EQUIV
) zType
[1] = 'E';
1713 if( ExprHasProperty(pTerm
->pExpr
, EP_FromJoin
) ) zType
[2] = 'L';
1714 if( pTerm
->eOperator
& WO_SINGLE
){
1715 sqlite3_snprintf(sizeof(zLeft
),zLeft
,"left={%d:%d}",
1716 pTerm
->leftCursor
, pTerm
->u
.leftColumn
);
1717 }else if( (pTerm
->eOperator
& WO_OR
)!=0 && pTerm
->u
.pOrInfo
!=0 ){
1718 sqlite3_snprintf(sizeof(zLeft
),zLeft
,"indexable=0x%lld",
1719 pTerm
->u
.pOrInfo
->indexable
);
1721 sqlite3_snprintf(sizeof(zLeft
),zLeft
,"left=%d", pTerm
->leftCursor
);
1724 "TERM-%-3d %p %s %-12s prob=%-3d op=0x%03x wtFlags=0x%04x",
1725 iTerm
, pTerm
, zType
, zLeft
, pTerm
->truthProb
,
1726 pTerm
->eOperator
, pTerm
->wtFlags
);
1727 if( pTerm
->iField
){
1728 sqlite3DebugPrintf(" iField=%d\n", pTerm
->iField
);
1730 sqlite3DebugPrintf("\n");
1732 sqlite3TreeViewExpr(0, pTerm
->pExpr
, 0);
1737 #ifdef WHERETRACE_ENABLED
1739 ** Show the complete content of a WhereClause
1741 void sqlite3WhereClausePrint(WhereClause
*pWC
){
1743 for(i
=0; i
<pWC
->nTerm
; i
++){
1744 whereTermPrint(&pWC
->a
[i
], i
);
1749 #ifdef WHERETRACE_ENABLED
1751 ** Print a WhereLoop object for debugging purposes
1753 static void whereLoopPrint(WhereLoop
*p
, WhereClause
*pWC
){
1754 WhereInfo
*pWInfo
= pWC
->pWInfo
;
1755 int nb
= 1+(pWInfo
->pTabList
->nSrc
+3)/4;
1756 struct SrcList_item
*pItem
= pWInfo
->pTabList
->a
+ p
->iTab
;
1757 Table
*pTab
= pItem
->pTab
;
1758 Bitmask mAll
= (((Bitmask
)1)<<(nb
*4)) - 1;
1759 sqlite3DebugPrintf("%c%2d.%0*llx.%0*llx", p
->cId
,
1760 p
->iTab
, nb
, p
->maskSelf
, nb
, p
->prereq
& mAll
);
1761 sqlite3DebugPrintf(" %12s",
1762 pItem
->zAlias
? pItem
->zAlias
: pTab
->zName
);
1763 if( (p
->wsFlags
& WHERE_VIRTUALTABLE
)==0 ){
1765 if( p
->u
.btree
.pIndex
&& (zName
= p
->u
.btree
.pIndex
->zName
)!=0 ){
1766 if( strncmp(zName
, "sqlite_autoindex_", 17)==0 ){
1767 int i
= sqlite3Strlen30(zName
) - 1;
1768 while( zName
[i
]!='_' ) i
--;
1771 sqlite3DebugPrintf(".%-16s %2d", zName
, p
->u
.btree
.nEq
);
1773 sqlite3DebugPrintf("%20s","");
1777 if( p
->u
.vtab
.idxStr
){
1778 z
= sqlite3_mprintf("(%d,\"%s\",%x)",
1779 p
->u
.vtab
.idxNum
, p
->u
.vtab
.idxStr
, p
->u
.vtab
.omitMask
);
1781 z
= sqlite3_mprintf("(%d,%x)", p
->u
.vtab
.idxNum
, p
->u
.vtab
.omitMask
);
1783 sqlite3DebugPrintf(" %-19s", z
);
1786 if( p
->wsFlags
& WHERE_SKIPSCAN
){
1787 sqlite3DebugPrintf(" f %05x %d-%d", p
->wsFlags
, p
->nLTerm
,p
->nSkip
);
1789 sqlite3DebugPrintf(" f %05x N %d", p
->wsFlags
, p
->nLTerm
);
1791 sqlite3DebugPrintf(" cost %d,%d,%d\n", p
->rSetup
, p
->rRun
, p
->nOut
);
1792 if( p
->nLTerm
&& (sqlite3WhereTrace
& 0x100)!=0 ){
1794 for(i
=0; i
<p
->nLTerm
; i
++){
1795 whereTermPrint(p
->aLTerm
[i
], i
);
1802 ** Convert bulk memory into a valid WhereLoop that can be passed
1803 ** to whereLoopClear harmlessly.
1805 static void whereLoopInit(WhereLoop
*p
){
1806 p
->aLTerm
= p
->aLTermSpace
;
1808 p
->nLSlot
= ArraySize(p
->aLTermSpace
);
1813 ** Clear the WhereLoop.u union. Leave WhereLoop.pLTerm intact.
1815 static void whereLoopClearUnion(sqlite3
*db
, WhereLoop
*p
){
1816 if( p
->wsFlags
& (WHERE_VIRTUALTABLE
|WHERE_AUTO_INDEX
) ){
1817 if( (p
->wsFlags
& WHERE_VIRTUALTABLE
)!=0 && p
->u
.vtab
.needFree
){
1818 sqlite3_free(p
->u
.vtab
.idxStr
);
1819 p
->u
.vtab
.needFree
= 0;
1820 p
->u
.vtab
.idxStr
= 0;
1821 }else if( (p
->wsFlags
& WHERE_AUTO_INDEX
)!=0 && p
->u
.btree
.pIndex
!=0 ){
1822 sqlite3DbFree(db
, p
->u
.btree
.pIndex
->zColAff
);
1823 sqlite3DbFreeNN(db
, p
->u
.btree
.pIndex
);
1824 p
->u
.btree
.pIndex
= 0;
1830 ** Deallocate internal memory used by a WhereLoop object
1832 static void whereLoopClear(sqlite3
*db
, WhereLoop
*p
){
1833 if( p
->aLTerm
!=p
->aLTermSpace
) sqlite3DbFreeNN(db
, p
->aLTerm
);
1834 whereLoopClearUnion(db
, p
);
1839 ** Increase the memory allocation for pLoop->aLTerm[] to be at least n.
1841 static int whereLoopResize(sqlite3
*db
, WhereLoop
*p
, int n
){
1843 if( p
->nLSlot
>=n
) return SQLITE_OK
;
1845 paNew
= sqlite3DbMallocRawNN(db
, sizeof(p
->aLTerm
[0])*n
);
1846 if( paNew
==0 ) return SQLITE_NOMEM_BKPT
;
1847 memcpy(paNew
, p
->aLTerm
, sizeof(p
->aLTerm
[0])*p
->nLSlot
);
1848 if( p
->aLTerm
!=p
->aLTermSpace
) sqlite3DbFreeNN(db
, p
->aLTerm
);
1855 ** Transfer content from the second pLoop into the first.
1857 static int whereLoopXfer(sqlite3
*db
, WhereLoop
*pTo
, WhereLoop
*pFrom
){
1858 whereLoopClearUnion(db
, pTo
);
1859 if( whereLoopResize(db
, pTo
, pFrom
->nLTerm
) ){
1860 memset(&pTo
->u
, 0, sizeof(pTo
->u
));
1861 return SQLITE_NOMEM_BKPT
;
1863 memcpy(pTo
, pFrom
, WHERE_LOOP_XFER_SZ
);
1864 memcpy(pTo
->aLTerm
, pFrom
->aLTerm
, pTo
->nLTerm
*sizeof(pTo
->aLTerm
[0]));
1865 if( pFrom
->wsFlags
& WHERE_VIRTUALTABLE
){
1866 pFrom
->u
.vtab
.needFree
= 0;
1867 }else if( (pFrom
->wsFlags
& WHERE_AUTO_INDEX
)!=0 ){
1868 pFrom
->u
.btree
.pIndex
= 0;
1874 ** Delete a WhereLoop object
1876 static void whereLoopDelete(sqlite3
*db
, WhereLoop
*p
){
1877 whereLoopClear(db
, p
);
1878 sqlite3DbFreeNN(db
, p
);
1882 ** Free a WhereInfo structure
1884 static void whereInfoFree(sqlite3
*db
, WhereInfo
*pWInfo
){
1886 assert( pWInfo
!=0 );
1887 for(i
=0; i
<pWInfo
->nLevel
; i
++){
1888 WhereLevel
*pLevel
= &pWInfo
->a
[i
];
1889 if( pLevel
->pWLoop
&& (pLevel
->pWLoop
->wsFlags
& WHERE_IN_ABLE
) ){
1890 sqlite3DbFree(db
, pLevel
->u
.in
.aInLoop
);
1893 sqlite3WhereClauseClear(&pWInfo
->sWC
);
1894 while( pWInfo
->pLoops
){
1895 WhereLoop
*p
= pWInfo
->pLoops
;
1896 pWInfo
->pLoops
= p
->pNextLoop
;
1897 whereLoopDelete(db
, p
);
1899 sqlite3DbFreeNN(db
, pWInfo
);
1903 ** Return TRUE if all of the following are true:
1905 ** (1) X has the same or lower cost that Y
1906 ** (2) X uses fewer WHERE clause terms than Y
1907 ** (3) Every WHERE clause term used by X is also used by Y
1908 ** (4) X skips at least as many columns as Y
1909 ** (5) If X is a covering index, than Y is too
1911 ** Conditions (2) and (3) mean that X is a "proper subset" of Y.
1912 ** If X is a proper subset of Y then Y is a better choice and ought
1913 ** to have a lower cost. This routine returns TRUE when that cost
1914 ** relationship is inverted and needs to be adjusted. Constraint (4)
1915 ** was added because if X uses skip-scan less than Y it still might
1916 ** deserve a lower cost even if it is a proper subset of Y. Constraint (5)
1917 ** was added because a covering index probably deserves to have a lower cost
1918 ** than a non-covering index even if it is a proper subset.
1920 static int whereLoopCheaperProperSubset(
1921 const WhereLoop
*pX
, /* First WhereLoop to compare */
1922 const WhereLoop
*pY
/* Compare against this WhereLoop */
1925 if( pX
->nLTerm
-pX
->nSkip
>= pY
->nLTerm
-pY
->nSkip
){
1926 return 0; /* X is not a subset of Y */
1928 if( pY
->nSkip
> pX
->nSkip
) return 0;
1929 if( pX
->rRun
>= pY
->rRun
){
1930 if( pX
->rRun
> pY
->rRun
) return 0; /* X costs more than Y */
1931 if( pX
->nOut
> pY
->nOut
) return 0; /* X costs more than Y */
1933 for(i
=pX
->nLTerm
-1; i
>=0; i
--){
1934 if( pX
->aLTerm
[i
]==0 ) continue;
1935 for(j
=pY
->nLTerm
-1; j
>=0; j
--){
1936 if( pY
->aLTerm
[j
]==pX
->aLTerm
[i
] ) break;
1938 if( j
<0 ) return 0; /* X not a subset of Y since term X[i] not used by Y */
1940 if( (pX
->wsFlags
&WHERE_IDX_ONLY
)!=0
1941 && (pY
->wsFlags
&WHERE_IDX_ONLY
)==0 ){
1942 return 0; /* Constraint (5) */
1944 return 1; /* All conditions meet */
1948 ** Try to adjust the cost of WhereLoop pTemplate upwards or downwards so
1951 ** (1) pTemplate costs less than any other WhereLoops that are a proper
1952 ** subset of pTemplate
1954 ** (2) pTemplate costs more than any other WhereLoops for which pTemplate
1955 ** is a proper subset.
1957 ** To say "WhereLoop X is a proper subset of Y" means that X uses fewer
1958 ** WHERE clause terms than Y and that every WHERE clause term used by X is
1961 static void whereLoopAdjustCost(const WhereLoop
*p
, WhereLoop
*pTemplate
){
1962 if( (pTemplate
->wsFlags
& WHERE_INDEXED
)==0 ) return;
1963 for(; p
; p
=p
->pNextLoop
){
1964 if( p
->iTab
!=pTemplate
->iTab
) continue;
1965 if( (p
->wsFlags
& WHERE_INDEXED
)==0 ) continue;
1966 if( whereLoopCheaperProperSubset(p
, pTemplate
) ){
1967 /* Adjust pTemplate cost downward so that it is cheaper than its
1969 WHERETRACE(0x80,("subset cost adjustment %d,%d to %d,%d\n",
1970 pTemplate
->rRun
, pTemplate
->nOut
, p
->rRun
, p
->nOut
-1));
1971 pTemplate
->rRun
= p
->rRun
;
1972 pTemplate
->nOut
= p
->nOut
- 1;
1973 }else if( whereLoopCheaperProperSubset(pTemplate
, p
) ){
1974 /* Adjust pTemplate cost upward so that it is costlier than p since
1975 ** pTemplate is a proper subset of p */
1976 WHERETRACE(0x80,("subset cost adjustment %d,%d to %d,%d\n",
1977 pTemplate
->rRun
, pTemplate
->nOut
, p
->rRun
, p
->nOut
+1));
1978 pTemplate
->rRun
= p
->rRun
;
1979 pTemplate
->nOut
= p
->nOut
+ 1;
1985 ** Search the list of WhereLoops in *ppPrev looking for one that can be
1986 ** replaced by pTemplate.
1988 ** Return NULL if pTemplate does not belong on the WhereLoop list.
1989 ** In other words if pTemplate ought to be dropped from further consideration.
1991 ** If pX is a WhereLoop that pTemplate can replace, then return the
1992 ** link that points to pX.
1994 ** If pTemplate cannot replace any existing element of the list but needs
1995 ** to be added to the list as a new entry, then return a pointer to the
1996 ** tail of the list.
1998 static WhereLoop
**whereLoopFindLesser(
2000 const WhereLoop
*pTemplate
2003 for(p
=(*ppPrev
); p
; ppPrev
=&p
->pNextLoop
, p
=*ppPrev
){
2004 if( p
->iTab
!=pTemplate
->iTab
|| p
->iSortIdx
!=pTemplate
->iSortIdx
){
2005 /* If either the iTab or iSortIdx values for two WhereLoop are different
2006 ** then those WhereLoops need to be considered separately. Neither is
2007 ** a candidate to replace the other. */
2010 /* In the current implementation, the rSetup value is either zero
2011 ** or the cost of building an automatic index (NlogN) and the NlogN
2012 ** is the same for compatible WhereLoops. */
2013 assert( p
->rSetup
==0 || pTemplate
->rSetup
==0
2014 || p
->rSetup
==pTemplate
->rSetup
);
2016 /* whereLoopAddBtree() always generates and inserts the automatic index
2017 ** case first. Hence compatible candidate WhereLoops never have a larger
2018 ** rSetup. Call this SETUP-INVARIANT */
2019 assert( p
->rSetup
>=pTemplate
->rSetup
);
2021 /* Any loop using an appliation-defined index (or PRIMARY KEY or
2022 ** UNIQUE constraint) with one or more == constraints is better
2023 ** than an automatic index. Unless it is a skip-scan. */
2024 if( (p
->wsFlags
& WHERE_AUTO_INDEX
)!=0
2025 && (pTemplate
->nSkip
)==0
2026 && (pTemplate
->wsFlags
& WHERE_INDEXED
)!=0
2027 && (pTemplate
->wsFlags
& WHERE_COLUMN_EQ
)!=0
2028 && (p
->prereq
& pTemplate
->prereq
)==pTemplate
->prereq
2033 /* If existing WhereLoop p is better than pTemplate, pTemplate can be
2034 ** discarded. WhereLoop p is better if:
2035 ** (1) p has no more dependencies than pTemplate, and
2036 ** (2) p has an equal or lower cost than pTemplate
2038 if( (p
->prereq
& pTemplate
->prereq
)==p
->prereq
/* (1) */
2039 && p
->rSetup
<=pTemplate
->rSetup
/* (2a) */
2040 && p
->rRun
<=pTemplate
->rRun
/* (2b) */
2041 && p
->nOut
<=pTemplate
->nOut
/* (2c) */
2043 return 0; /* Discard pTemplate */
2046 /* If pTemplate is always better than p, then cause p to be overwritten
2047 ** with pTemplate. pTemplate is better than p if:
2048 ** (1) pTemplate has no more dependences than p, and
2049 ** (2) pTemplate has an equal or lower cost than p.
2051 if( (p
->prereq
& pTemplate
->prereq
)==pTemplate
->prereq
/* (1) */
2052 && p
->rRun
>=pTemplate
->rRun
/* (2a) */
2053 && p
->nOut
>=pTemplate
->nOut
/* (2b) */
2055 assert( p
->rSetup
>=pTemplate
->rSetup
); /* SETUP-INVARIANT above */
2056 break; /* Cause p to be overwritten by pTemplate */
2063 ** Insert or replace a WhereLoop entry using the template supplied.
2065 ** An existing WhereLoop entry might be overwritten if the new template
2066 ** is better and has fewer dependencies. Or the template will be ignored
2067 ** and no insert will occur if an existing WhereLoop is faster and has
2068 ** fewer dependencies than the template. Otherwise a new WhereLoop is
2069 ** added based on the template.
2071 ** If pBuilder->pOrSet is not NULL then we care about only the
2072 ** prerequisites and rRun and nOut costs of the N best loops. That
2073 ** information is gathered in the pBuilder->pOrSet object. This special
2074 ** processing mode is used only for OR clause processing.
2076 ** When accumulating multiple loops (when pBuilder->pOrSet is NULL) we
2077 ** still might overwrite similar loops with the new template if the
2078 ** new template is better. Loops may be overwritten if the following
2079 ** conditions are met:
2081 ** (1) They have the same iTab.
2082 ** (2) They have the same iSortIdx.
2083 ** (3) The template has same or fewer dependencies than the current loop
2084 ** (4) The template has the same or lower cost than the current loop
2086 static int whereLoopInsert(WhereLoopBuilder
*pBuilder
, WhereLoop
*pTemplate
){
2087 WhereLoop
**ppPrev
, *p
;
2088 WhereInfo
*pWInfo
= pBuilder
->pWInfo
;
2089 sqlite3
*db
= pWInfo
->pParse
->db
;
2092 /* If pBuilder->pOrSet is defined, then only keep track of the costs
2095 if( pBuilder
->pOrSet
!=0 ){
2096 if( pTemplate
->nLTerm
){
2097 #if WHERETRACE_ENABLED
2098 u16 n
= pBuilder
->pOrSet
->n
;
2101 whereOrInsert(pBuilder
->pOrSet
, pTemplate
->prereq
, pTemplate
->rRun
,
2103 #if WHERETRACE_ENABLED /* 0x8 */
2104 if( sqlite3WhereTrace
& 0x8 ){
2105 sqlite3DebugPrintf(x
?" or-%d: ":" or-X: ", n
);
2106 whereLoopPrint(pTemplate
, pBuilder
->pWC
);
2113 /* Look for an existing WhereLoop to replace with pTemplate
2115 whereLoopAdjustCost(pWInfo
->pLoops
, pTemplate
);
2116 ppPrev
= whereLoopFindLesser(&pWInfo
->pLoops
, pTemplate
);
2119 /* There already exists a WhereLoop on the list that is better
2120 ** than pTemplate, so just ignore pTemplate */
2121 #if WHERETRACE_ENABLED /* 0x8 */
2122 if( sqlite3WhereTrace
& 0x8 ){
2123 sqlite3DebugPrintf(" skip: ");
2124 whereLoopPrint(pTemplate
, pBuilder
->pWC
);
2132 /* If we reach this point it means that either p[] should be overwritten
2133 ** with pTemplate[] if p[] exists, or if p==NULL then allocate a new
2134 ** WhereLoop and insert it.
2136 #if WHERETRACE_ENABLED /* 0x8 */
2137 if( sqlite3WhereTrace
& 0x8 ){
2139 sqlite3DebugPrintf("replace: ");
2140 whereLoopPrint(p
, pBuilder
->pWC
);
2141 sqlite3DebugPrintf(" with: ");
2143 sqlite3DebugPrintf(" add: ");
2145 whereLoopPrint(pTemplate
, pBuilder
->pWC
);
2149 /* Allocate a new WhereLoop to add to the end of the list */
2150 *ppPrev
= p
= sqlite3DbMallocRawNN(db
, sizeof(WhereLoop
));
2151 if( p
==0 ) return SQLITE_NOMEM_BKPT
;
2155 /* We will be overwriting WhereLoop p[]. But before we do, first
2156 ** go through the rest of the list and delete any other entries besides
2157 ** p[] that are also supplated by pTemplate */
2158 WhereLoop
**ppTail
= &p
->pNextLoop
;
2161 ppTail
= whereLoopFindLesser(ppTail
, pTemplate
);
2162 if( ppTail
==0 ) break;
2164 if( pToDel
==0 ) break;
2165 *ppTail
= pToDel
->pNextLoop
;
2166 #if WHERETRACE_ENABLED /* 0x8 */
2167 if( sqlite3WhereTrace
& 0x8 ){
2168 sqlite3DebugPrintf(" delete: ");
2169 whereLoopPrint(pToDel
, pBuilder
->pWC
);
2172 whereLoopDelete(db
, pToDel
);
2175 rc
= whereLoopXfer(db
, p
, pTemplate
);
2176 if( (p
->wsFlags
& WHERE_VIRTUALTABLE
)==0 ){
2177 Index
*pIndex
= p
->u
.btree
.pIndex
;
2178 if( pIndex
&& pIndex
->tnum
==0 ){
2179 p
->u
.btree
.pIndex
= 0;
2186 ** Adjust the WhereLoop.nOut value downward to account for terms of the
2187 ** WHERE clause that reference the loop but which are not used by an
2190 ** For every WHERE clause term that is not used by the index
2191 ** and which has a truth probability assigned by one of the likelihood(),
2192 ** likely(), or unlikely() SQL functions, reduce the estimated number
2193 ** of output rows by the probability specified.
2195 ** TUNING: For every WHERE clause term that is not used by the index
2196 ** and which does not have an assigned truth probability, heuristics
2197 ** described below are used to try to estimate the truth probability.
2198 ** TODO --> Perhaps this is something that could be improved by better
2199 ** table statistics.
2201 ** Heuristic 1: Estimate the truth probability as 93.75%. The 93.75%
2202 ** value corresponds to -1 in LogEst notation, so this means decrement
2203 ** the WhereLoop.nOut field for every such WHERE clause term.
2205 ** Heuristic 2: If there exists one or more WHERE clause terms of the
2206 ** form "x==EXPR" and EXPR is not a constant 0 or 1, then make sure the
2207 ** final output row estimate is no greater than 1/4 of the total number
2208 ** of rows in the table. In other words, assume that x==EXPR will filter
2209 ** out at least 3 out of 4 rows. If EXPR is -1 or 0 or 1, then maybe the
2210 ** "x" column is boolean or else -1 or 0 or 1 is a common default value
2211 ** on the "x" column and so in that case only cap the output row estimate
2212 ** at 1/2 instead of 1/4.
2214 static void whereLoopOutputAdjust(
2215 WhereClause
*pWC
, /* The WHERE clause */
2216 WhereLoop
*pLoop
, /* The loop to adjust downward */
2217 LogEst nRow
/* Number of rows in the entire table */
2219 WhereTerm
*pTerm
, *pX
;
2220 Bitmask notAllowed
= ~(pLoop
->prereq
|pLoop
->maskSelf
);
2222 LogEst iReduce
= 0; /* pLoop->nOut should not exceed nRow-iReduce */
2224 assert( (pLoop
->wsFlags
& WHERE_AUTO_INDEX
)==0 );
2225 for(i
=pWC
->nTerm
, pTerm
=pWC
->a
; i
>0; i
--, pTerm
++){
2226 if( (pTerm
->wtFlags
& TERM_VIRTUAL
)!=0 ) break;
2227 if( (pTerm
->prereqAll
& pLoop
->maskSelf
)==0 ) continue;
2228 if( (pTerm
->prereqAll
& notAllowed
)!=0 ) continue;
2229 for(j
=pLoop
->nLTerm
-1; j
>=0; j
--){
2230 pX
= pLoop
->aLTerm
[j
];
2231 if( pX
==0 ) continue;
2232 if( pX
==pTerm
) break;
2233 if( pX
->iParent
>=0 && (&pWC
->a
[pX
->iParent
])==pTerm
) break;
2236 if( pTerm
->truthProb
<=0 ){
2237 /* If a truth probability is specified using the likelihood() hints,
2238 ** then use the probability provided by the application. */
2239 pLoop
->nOut
+= pTerm
->truthProb
;
2241 /* In the absence of explicit truth probabilities, use heuristics to
2242 ** guess a reasonable truth probability. */
2244 if( pTerm
->eOperator
&(WO_EQ
|WO_IS
) ){
2245 Expr
*pRight
= pTerm
->pExpr
->pRight
;
2246 testcase( pTerm
->pExpr
->op
==TK_IS
);
2247 if( sqlite3ExprIsInteger(pRight
, &k
) && k
>=(-1) && k
<=1 ){
2252 if( iReduce
<k
) iReduce
= k
;
2257 if( pLoop
->nOut
> nRow
-iReduce
) pLoop
->nOut
= nRow
- iReduce
;
2261 ** Term pTerm is a vector range comparison operation. The first comparison
2262 ** in the vector can be optimized using column nEq of the index. This
2263 ** function returns the total number of vector elements that can be used
2264 ** as part of the range comparison.
2266 ** For example, if the query is:
2268 ** WHERE a = ? AND (b, c, d) > (?, ?, ?)
2272 ** CREATE INDEX ... ON (a, b, c, d, e)
2274 ** then this function would be invoked with nEq=1. The value returned in
2277 static int whereRangeVectorLen(
2278 Parse
*pParse
, /* Parsing context */
2279 int iCur
, /* Cursor open on pIdx */
2280 Index
*pIdx
, /* The index to be used for a inequality constraint */
2281 int nEq
, /* Number of prior equality constraints on same index */
2282 WhereTerm
*pTerm
/* The vector inequality constraint */
2284 int nCmp
= sqlite3ExprVectorSize(pTerm
->pExpr
->pLeft
);
2287 nCmp
= MIN(nCmp
, (pIdx
->nColumn
- nEq
));
2288 for(i
=1; i
<nCmp
; i
++){
2289 /* Test if comparison i of pTerm is compatible with column (i+nEq)
2290 ** of the index. If not, exit the loop. */
2291 char aff
; /* Comparison affinity */
2292 char idxaff
= 0; /* Indexed columns affinity */
2293 CollSeq
*pColl
; /* Comparison collation sequence */
2294 Expr
*pLhs
= pTerm
->pExpr
->pLeft
->x
.pList
->a
[i
].pExpr
;
2295 Expr
*pRhs
= pTerm
->pExpr
->pRight
;
2296 if( pRhs
->flags
& EP_xIsSelect
){
2297 pRhs
= pRhs
->x
.pSelect
->pEList
->a
[i
].pExpr
;
2299 pRhs
= pRhs
->x
.pList
->a
[i
].pExpr
;
2302 /* Check that the LHS of the comparison is a column reference to
2303 ** the right column of the right source table. And that the sort
2304 ** order of the index column is the same as the sort order of the
2305 ** leftmost index column. */
2306 if( pLhs
->op
!=TK_COLUMN
2307 || pLhs
->iTable
!=iCur
2308 || pLhs
->iColumn
!=pIdx
->aiColumn
[i
+nEq
]
2309 || pIdx
->aSortOrder
[i
+nEq
]!=pIdx
->aSortOrder
[nEq
]
2314 testcase( pLhs
->iColumn
==XN_ROWID
);
2315 aff
= sqlite3CompareAffinity(pRhs
, sqlite3ExprAffinity(pLhs
));
2316 idxaff
= sqlite3TableColumnAffinity(pIdx
->pTable
, pLhs
->iColumn
);
2317 if( aff
!=idxaff
) break;
2319 pColl
= sqlite3BinaryCompareCollSeq(pParse
, pLhs
, pRhs
);
2320 if( pColl
==0 ) break;
2321 if( sqlite3StrICmp(pColl
->zName
, pIdx
->azColl
[i
+nEq
]) ) break;
2327 ** Adjust the cost C by the costMult facter T. This only occurs if
2328 ** compiled with -DSQLITE_ENABLE_COSTMULT
2330 #ifdef SQLITE_ENABLE_COSTMULT
2331 # define ApplyCostMultiplier(C,T) C += T
2333 # define ApplyCostMultiplier(C,T)
2337 ** We have so far matched pBuilder->pNew->u.btree.nEq terms of the
2338 ** index pIndex. Try to match one more.
2340 ** When this function is called, pBuilder->pNew->nOut contains the
2341 ** number of rows expected to be visited by filtering using the nEq
2342 ** terms only. If it is modified, this value is restored before this
2343 ** function returns.
2345 ** If pProbe->tnum==0, that means pIndex is a fake index used for the
2346 ** INTEGER PRIMARY KEY.
2348 static int whereLoopAddBtreeIndex(
2349 WhereLoopBuilder
*pBuilder
, /* The WhereLoop factory */
2350 struct SrcList_item
*pSrc
, /* FROM clause term being analyzed */
2351 Index
*pProbe
, /* An index on pSrc */
2352 LogEst nInMul
/* log(Number of iterations due to IN) */
2354 WhereInfo
*pWInfo
= pBuilder
->pWInfo
; /* WHERE analyse context */
2355 Parse
*pParse
= pWInfo
->pParse
; /* Parsing context */
2356 sqlite3
*db
= pParse
->db
; /* Database connection malloc context */
2357 WhereLoop
*pNew
; /* Template WhereLoop under construction */
2358 WhereTerm
*pTerm
; /* A WhereTerm under consideration */
2359 int opMask
; /* Valid operators for constraints */
2360 WhereScan scan
; /* Iterator for WHERE terms */
2361 Bitmask saved_prereq
; /* Original value of pNew->prereq */
2362 u16 saved_nLTerm
; /* Original value of pNew->nLTerm */
2363 u16 saved_nEq
; /* Original value of pNew->u.btree.nEq */
2364 u16 saved_nBtm
; /* Original value of pNew->u.btree.nBtm */
2365 u16 saved_nTop
; /* Original value of pNew->u.btree.nTop */
2366 u16 saved_nSkip
; /* Original value of pNew->nSkip */
2367 u32 saved_wsFlags
; /* Original value of pNew->wsFlags */
2368 LogEst saved_nOut
; /* Original value of pNew->nOut */
2369 int rc
= SQLITE_OK
; /* Return code */
2370 LogEst rSize
; /* Number of rows in the table */
2371 LogEst rLogSize
; /* Logarithm of table size */
2372 WhereTerm
*pTop
= 0, *pBtm
= 0; /* Top and bottom range constraints */
2374 pNew
= pBuilder
->pNew
;
2375 if( db
->mallocFailed
) return SQLITE_NOMEM_BKPT
;
2376 WHERETRACE(0x800, ("BEGIN %s.addBtreeIdx(%s), nEq=%d\n",
2377 pProbe
->pTable
->zName
,pProbe
->zName
, pNew
->u
.btree
.nEq
));
2379 assert( (pNew
->wsFlags
& WHERE_VIRTUALTABLE
)==0 );
2380 assert( (pNew
->wsFlags
& WHERE_TOP_LIMIT
)==0 );
2381 if( pNew
->wsFlags
& WHERE_BTM_LIMIT
){
2382 opMask
= WO_LT
|WO_LE
;
2384 assert( pNew
->u
.btree
.nBtm
==0 );
2385 opMask
= WO_EQ
|WO_IN
|WO_GT
|WO_GE
|WO_LT
|WO_LE
|WO_ISNULL
|WO_IS
;
2387 if( pProbe
->bUnordered
) opMask
&= ~(WO_GT
|WO_GE
|WO_LT
|WO_LE
);
2389 assert( pNew
->u
.btree
.nEq
<pProbe
->nColumn
);
2391 saved_nEq
= pNew
->u
.btree
.nEq
;
2392 saved_nBtm
= pNew
->u
.btree
.nBtm
;
2393 saved_nTop
= pNew
->u
.btree
.nTop
;
2394 saved_nSkip
= pNew
->nSkip
;
2395 saved_nLTerm
= pNew
->nLTerm
;
2396 saved_wsFlags
= pNew
->wsFlags
;
2397 saved_prereq
= pNew
->prereq
;
2398 saved_nOut
= pNew
->nOut
;
2399 pTerm
= whereScanInit(&scan
, pBuilder
->pWC
, pSrc
->iCursor
, saved_nEq
,
2402 rSize
= pProbe
->aiRowLogEst
[0];
2403 rLogSize
= estLog(rSize
);
2404 for(; rc
==SQLITE_OK
&& pTerm
!=0; pTerm
= whereScanNext(&scan
)){
2405 u16 eOp
= pTerm
->eOperator
; /* Shorthand for pTerm->eOperator */
2407 LogEst nOutUnadjusted
; /* nOut before IN() and WHERE adjustments */
2409 #ifdef SQLITE_ENABLE_STAT3_OR_STAT4
2410 int nRecValid
= pBuilder
->nRecValid
;
2412 if( (eOp
==WO_ISNULL
|| (pTerm
->wtFlags
&TERM_VNULL
)!=0)
2413 && indexColumnNotNull(pProbe
, saved_nEq
)
2415 continue; /* ignore IS [NOT] NULL constraints on NOT NULL columns */
2417 if( pTerm
->prereqRight
& pNew
->maskSelf
) continue;
2419 /* Do not allow the upper bound of a LIKE optimization range constraint
2420 ** to mix with a lower range bound from some other source */
2421 if( pTerm
->wtFlags
& TERM_LIKEOPT
&& pTerm
->eOperator
==WO_LT
) continue;
2423 /* Do not allow constraints from the WHERE clause to be used by the
2424 ** right table of a LEFT JOIN. Only constraints in the ON clause are
2426 if( (pSrc
->fg
.jointype
& JT_LEFT
)!=0
2427 && !ExprHasProperty(pTerm
->pExpr
, EP_FromJoin
)
2432 if( IsUniqueIndex(pProbe
) && saved_nEq
==pProbe
->nKeyCol
-1 ){
2433 pBuilder
->bldFlags
|= SQLITE_BLDF_UNIQUE
;
2435 pBuilder
->bldFlags
|= SQLITE_BLDF_INDEXED
;
2437 pNew
->wsFlags
= saved_wsFlags
;
2438 pNew
->u
.btree
.nEq
= saved_nEq
;
2439 pNew
->u
.btree
.nBtm
= saved_nBtm
;
2440 pNew
->u
.btree
.nTop
= saved_nTop
;
2441 pNew
->nLTerm
= saved_nLTerm
;
2442 if( whereLoopResize(db
, pNew
, pNew
->nLTerm
+1) ) break; /* OOM */
2443 pNew
->aLTerm
[pNew
->nLTerm
++] = pTerm
;
2444 pNew
->prereq
= (saved_prereq
| pTerm
->prereqRight
) & ~pNew
->maskSelf
;
2447 || (pNew
->wsFlags
& WHERE_COLUMN_NULL
)!=0
2448 || (pNew
->wsFlags
& WHERE_COLUMN_IN
)!=0
2449 || (pNew
->wsFlags
& WHERE_SKIPSCAN
)!=0
2453 Expr
*pExpr
= pTerm
->pExpr
;
2454 if( ExprHasProperty(pExpr
, EP_xIsSelect
) ){
2455 /* "x IN (SELECT ...)": TUNING: the SELECT returns 25 rows */
2457 nIn
= 46; assert( 46==sqlite3LogEst(25) );
2459 /* The expression may actually be of the form (x, y) IN (SELECT...).
2460 ** In this case there is a separate term for each of (x) and (y).
2461 ** However, the nIn multiplier should only be applied once, not once
2462 ** for each such term. The following loop checks that pTerm is the
2463 ** first such term in use, and sets nIn back to 0 if it is not. */
2464 for(i
=0; i
<pNew
->nLTerm
-1; i
++){
2465 if( pNew
->aLTerm
[i
] && pNew
->aLTerm
[i
]->pExpr
==pExpr
) nIn
= 0;
2467 }else if( ALWAYS(pExpr
->x
.pList
&& pExpr
->x
.pList
->nExpr
) ){
2468 /* "x IN (value, value, ...)" */
2469 nIn
= sqlite3LogEst(pExpr
->x
.pList
->nExpr
);
2470 assert( nIn
>0 ); /* RHS always has 2 or more terms... The parser
2471 ** changes "x IN (?)" into "x=?". */
2473 if( pProbe
->hasStat1
){
2474 LogEst M
, logK
, safetyMargin
;
2476 ** N = the total number of rows in the table
2477 ** K = the number of entries on the RHS of the IN operator
2478 ** M = the number of rows in the table that match terms to the
2479 ** to the left in the same index. If the IN operator is on
2480 ** the left-most index column, M==N.
2482 ** Given the definitions above, it is better to omit the IN operator
2483 ** from the index lookup and instead do a scan of the M elements,
2484 ** testing each scanned row against the IN operator separately, if:
2486 ** M*log(K) < K*log(N)
2488 ** Our estimates for M, K, and N might be inaccurate, so we build in
2489 ** a safety margin of 2 (LogEst: 10) that favors using the IN operator
2490 ** with the index, as using an index has better worst-case behavior.
2491 ** If we do not have real sqlite_stat1 data, always prefer to use
2494 M
= pProbe
->aiRowLogEst
[saved_nEq
];
2496 safetyMargin
= 10; /* TUNING: extra weight for indexed IN */
2497 if( M
+ logK
+ safetyMargin
< nIn
+ rLogSize
){
2499 ("Scan preferred over IN operator on column %d of \"%s\" (%d<%d)\n",
2500 saved_nEq
, pProbe
->zName
, M
+logK
+10, nIn
+rLogSize
));
2504 ("IN operator preferred on column %d of \"%s\" (%d>=%d)\n",
2505 saved_nEq
, pProbe
->zName
, M
+logK
+10, nIn
+rLogSize
));
2508 pNew
->wsFlags
|= WHERE_COLUMN_IN
;
2509 }else if( eOp
& (WO_EQ
|WO_IS
) ){
2510 int iCol
= pProbe
->aiColumn
[saved_nEq
];
2511 pNew
->wsFlags
|= WHERE_COLUMN_EQ
;
2512 assert( saved_nEq
==pNew
->u
.btree
.nEq
);
2514 || (iCol
>=0 && nInMul
==0 && saved_nEq
==pProbe
->nKeyCol
-1)
2516 if( iCol
==XN_ROWID
|| pProbe
->uniqNotNull
2517 || (pProbe
->nKeyCol
==1 && pProbe
->onError
&& eOp
==WO_EQ
)
2519 pNew
->wsFlags
|= WHERE_ONEROW
;
2521 pNew
->wsFlags
|= WHERE_UNQ_WANTED
;
2524 }else if( eOp
& WO_ISNULL
){
2525 pNew
->wsFlags
|= WHERE_COLUMN_NULL
;
2526 }else if( eOp
& (WO_GT
|WO_GE
) ){
2527 testcase( eOp
& WO_GT
);
2528 testcase( eOp
& WO_GE
);
2529 pNew
->wsFlags
|= WHERE_COLUMN_RANGE
|WHERE_BTM_LIMIT
;
2530 pNew
->u
.btree
.nBtm
= whereRangeVectorLen(
2531 pParse
, pSrc
->iCursor
, pProbe
, saved_nEq
, pTerm
2535 if( pTerm
->wtFlags
& TERM_LIKEOPT
){
2536 /* Range contraints that come from the LIKE optimization are
2537 ** always used in pairs. */
2539 assert( (pTop
-(pTerm
->pWC
->a
))<pTerm
->pWC
->nTerm
);
2540 assert( pTop
->wtFlags
& TERM_LIKEOPT
);
2541 assert( pTop
->eOperator
==WO_LT
);
2542 if( whereLoopResize(db
, pNew
, pNew
->nLTerm
+1) ) break; /* OOM */
2543 pNew
->aLTerm
[pNew
->nLTerm
++] = pTop
;
2544 pNew
->wsFlags
|= WHERE_TOP_LIMIT
;
2545 pNew
->u
.btree
.nTop
= 1;
2548 assert( eOp
& (WO_LT
|WO_LE
) );
2549 testcase( eOp
& WO_LT
);
2550 testcase( eOp
& WO_LE
);
2551 pNew
->wsFlags
|= WHERE_COLUMN_RANGE
|WHERE_TOP_LIMIT
;
2552 pNew
->u
.btree
.nTop
= whereRangeVectorLen(
2553 pParse
, pSrc
->iCursor
, pProbe
, saved_nEq
, pTerm
2556 pBtm
= (pNew
->wsFlags
& WHERE_BTM_LIMIT
)!=0 ?
2557 pNew
->aLTerm
[pNew
->nLTerm
-2] : 0;
2560 /* At this point pNew->nOut is set to the number of rows expected to
2561 ** be visited by the index scan before considering term pTerm, or the
2562 ** values of nIn and nInMul. In other words, assuming that all
2563 ** "x IN(...)" terms are replaced with "x = ?". This block updates
2564 ** the value of pNew->nOut to account for pTerm (but not nIn/nInMul). */
2565 assert( pNew
->nOut
==saved_nOut
);
2566 if( pNew
->wsFlags
& WHERE_COLUMN_RANGE
){
2567 /* Adjust nOut using stat3/stat4 data. Or, if there is no stat3/stat4
2568 ** data, using some other estimate. */
2569 whereRangeScanEst(pParse
, pBuilder
, pBtm
, pTop
, pNew
);
2571 int nEq
= ++pNew
->u
.btree
.nEq
;
2572 assert( eOp
& (WO_ISNULL
|WO_EQ
|WO_IN
|WO_IS
) );
2574 assert( pNew
->nOut
==saved_nOut
);
2575 if( pTerm
->truthProb
<=0 && pProbe
->aiColumn
[saved_nEq
]>=0 ){
2576 assert( (eOp
& WO_IN
) || nIn
==0 );
2577 testcase( eOp
& WO_IN
);
2578 pNew
->nOut
+= pTerm
->truthProb
;
2581 #ifdef SQLITE_ENABLE_STAT3_OR_STAT4
2585 && pNew
->u
.btree
.nEq
<=pProbe
->nSampleCol
2586 && ((eOp
& WO_IN
)==0 || !ExprHasProperty(pTerm
->pExpr
, EP_xIsSelect
))
2588 Expr
*pExpr
= pTerm
->pExpr
;
2589 if( (eOp
& (WO_EQ
|WO_ISNULL
|WO_IS
))!=0 ){
2590 testcase( eOp
& WO_EQ
);
2591 testcase( eOp
& WO_IS
);
2592 testcase( eOp
& WO_ISNULL
);
2593 rc
= whereEqualScanEst(pParse
, pBuilder
, pExpr
->pRight
, &nOut
);
2595 rc
= whereInScanEst(pParse
, pBuilder
, pExpr
->x
.pList
, &nOut
);
2597 if( rc
==SQLITE_NOTFOUND
) rc
= SQLITE_OK
;
2598 if( rc
!=SQLITE_OK
) break; /* Jump out of the pTerm loop */
2600 pNew
->nOut
= sqlite3LogEst(nOut
);
2601 if( pNew
->nOut
>saved_nOut
) pNew
->nOut
= saved_nOut
;
2608 pNew
->nOut
+= (pProbe
->aiRowLogEst
[nEq
] - pProbe
->aiRowLogEst
[nEq
-1]);
2609 if( eOp
& WO_ISNULL
){
2610 /* TUNING: If there is no likelihood() value, assume that a
2611 ** "col IS NULL" expression matches twice as many rows
2619 /* Set rCostIdx to the cost of visiting selected rows in index. Add
2620 ** it to pNew->rRun, which is currently set to the cost of the index
2621 ** seek only. Then, if this is a non-covering index, add the cost of
2622 ** visiting the rows in the main table. */
2623 rCostIdx
= pNew
->nOut
+ 1 + (15*pProbe
->szIdxRow
)/pSrc
->pTab
->szTabRow
;
2624 pNew
->rRun
= sqlite3LogEstAdd(rLogSize
, rCostIdx
);
2625 if( (pNew
->wsFlags
& (WHERE_IDX_ONLY
|WHERE_IPK
))==0 ){
2626 pNew
->rRun
= sqlite3LogEstAdd(pNew
->rRun
, pNew
->nOut
+ 16);
2628 ApplyCostMultiplier(pNew
->rRun
, pProbe
->pTable
->costMult
);
2630 nOutUnadjusted
= pNew
->nOut
;
2631 pNew
->rRun
+= nInMul
+ nIn
;
2632 pNew
->nOut
+= nInMul
+ nIn
;
2633 whereLoopOutputAdjust(pBuilder
->pWC
, pNew
, rSize
);
2634 rc
= whereLoopInsert(pBuilder
, pNew
);
2636 if( pNew
->wsFlags
& WHERE_COLUMN_RANGE
){
2637 pNew
->nOut
= saved_nOut
;
2639 pNew
->nOut
= nOutUnadjusted
;
2642 if( (pNew
->wsFlags
& WHERE_TOP_LIMIT
)==0
2643 && pNew
->u
.btree
.nEq
<pProbe
->nColumn
2645 whereLoopAddBtreeIndex(pBuilder
, pSrc
, pProbe
, nInMul
+nIn
);
2647 pNew
->nOut
= saved_nOut
;
2648 #ifdef SQLITE_ENABLE_STAT3_OR_STAT4
2649 pBuilder
->nRecValid
= nRecValid
;
2652 pNew
->prereq
= saved_prereq
;
2653 pNew
->u
.btree
.nEq
= saved_nEq
;
2654 pNew
->u
.btree
.nBtm
= saved_nBtm
;
2655 pNew
->u
.btree
.nTop
= saved_nTop
;
2656 pNew
->nSkip
= saved_nSkip
;
2657 pNew
->wsFlags
= saved_wsFlags
;
2658 pNew
->nOut
= saved_nOut
;
2659 pNew
->nLTerm
= saved_nLTerm
;
2661 /* Consider using a skip-scan if there are no WHERE clause constraints
2662 ** available for the left-most terms of the index, and if the average
2663 ** number of repeats in the left-most terms is at least 18.
2665 ** The magic number 18 is selected on the basis that scanning 17 rows
2666 ** is almost always quicker than an index seek (even though if the index
2667 ** contains fewer than 2^17 rows we assume otherwise in other parts of
2668 ** the code). And, even if it is not, it should not be too much slower.
2669 ** On the other hand, the extra seeks could end up being significantly
2670 ** more expensive. */
2671 assert( 42==sqlite3LogEst(18) );
2672 if( saved_nEq
==saved_nSkip
2673 && saved_nEq
+1<pProbe
->nKeyCol
2674 && pProbe
->noSkipScan
==0
2675 && pProbe
->aiRowLogEst
[saved_nEq
+1]>=42 /* TUNING: Minimum for skip-scan */
2676 && (rc
= whereLoopResize(db
, pNew
, pNew
->nLTerm
+1))==SQLITE_OK
2679 pNew
->u
.btree
.nEq
++;
2681 pNew
->aLTerm
[pNew
->nLTerm
++] = 0;
2682 pNew
->wsFlags
|= WHERE_SKIPSCAN
;
2683 nIter
= pProbe
->aiRowLogEst
[saved_nEq
] - pProbe
->aiRowLogEst
[saved_nEq
+1];
2684 pNew
->nOut
-= nIter
;
2685 /* TUNING: Because uncertainties in the estimates for skip-scan queries,
2686 ** add a 1.375 fudge factor to make skip-scan slightly less likely. */
2688 whereLoopAddBtreeIndex(pBuilder
, pSrc
, pProbe
, nIter
+ nInMul
);
2689 pNew
->nOut
= saved_nOut
;
2690 pNew
->u
.btree
.nEq
= saved_nEq
;
2691 pNew
->nSkip
= saved_nSkip
;
2692 pNew
->wsFlags
= saved_wsFlags
;
2695 WHERETRACE(0x800, ("END %s.addBtreeIdx(%s), nEq=%d, rc=%d\n",
2696 pProbe
->pTable
->zName
, pProbe
->zName
, saved_nEq
, rc
));
2701 ** Return True if it is possible that pIndex might be useful in
2702 ** implementing the ORDER BY clause in pBuilder.
2704 ** Return False if pBuilder does not contain an ORDER BY clause or
2705 ** if there is no way for pIndex to be useful in implementing that
2708 static int indexMightHelpWithOrderBy(
2709 WhereLoopBuilder
*pBuilder
,
2717 if( pIndex
->bUnordered
) return 0;
2718 if( (pOB
= pBuilder
->pWInfo
->pOrderBy
)==0 ) return 0;
2719 for(ii
=0; ii
<pOB
->nExpr
; ii
++){
2720 Expr
*pExpr
= sqlite3ExprSkipCollate(pOB
->a
[ii
].pExpr
);
2721 if( pExpr
->op
==TK_COLUMN
&& pExpr
->iTable
==iCursor
){
2722 if( pExpr
->iColumn
<0 ) return 1;
2723 for(jj
=0; jj
<pIndex
->nKeyCol
; jj
++){
2724 if( pExpr
->iColumn
==pIndex
->aiColumn
[jj
] ) return 1;
2726 }else if( (aColExpr
= pIndex
->aColExpr
)!=0 ){
2727 for(jj
=0; jj
<pIndex
->nKeyCol
; jj
++){
2728 if( pIndex
->aiColumn
[jj
]!=XN_EXPR
) continue;
2729 if( sqlite3ExprCompareSkip(pExpr
,aColExpr
->a
[jj
].pExpr
,iCursor
)==0 ){
2738 /* Check to see if a partial index with pPartIndexWhere can be used
2739 ** in the current query. Return true if it can be and false if not.
2741 static int whereUsablePartialIndex(int iTab
, WhereClause
*pWC
, Expr
*pWhere
){
2744 Parse
*pParse
= pWC
->pWInfo
->pParse
;
2745 while( pWhere
->op
==TK_AND
){
2746 if( !whereUsablePartialIndex(iTab
,pWC
,pWhere
->pLeft
) ) return 0;
2747 pWhere
= pWhere
->pRight
;
2749 if( pParse
->db
->flags
& SQLITE_EnableQPSG
) pParse
= 0;
2750 for(i
=0, pTerm
=pWC
->a
; i
<pWC
->nTerm
; i
++, pTerm
++){
2751 Expr
*pExpr
= pTerm
->pExpr
;
2752 if( (!ExprHasProperty(pExpr
, EP_FromJoin
) || pExpr
->iRightJoinTable
==iTab
)
2753 && sqlite3ExprImpliesExpr(pParse
, pExpr
, pWhere
, iTab
)
2762 ** Add all WhereLoop objects for a single table of the join where the table
2763 ** is identified by pBuilder->pNew->iTab. That table is guaranteed to be
2764 ** a b-tree table, not a virtual table.
2766 ** The costs (WhereLoop.rRun) of the b-tree loops added by this function
2767 ** are calculated as follows:
2769 ** For a full scan, assuming the table (or index) contains nRow rows:
2771 ** cost = nRow * 3.0 // full-table scan
2772 ** cost = nRow * K // scan of covering index
2773 ** cost = nRow * (K+3.0) // scan of non-covering index
2775 ** where K is a value between 1.1 and 3.0 set based on the relative
2776 ** estimated average size of the index and table records.
2778 ** For an index scan, where nVisit is the number of index rows visited
2779 ** by the scan, and nSeek is the number of seek operations required on
2780 ** the index b-tree:
2782 ** cost = nSeek * (log(nRow) + K * nVisit) // covering index
2783 ** cost = nSeek * (log(nRow) + (K+3.0) * nVisit) // non-covering index
2785 ** Normally, nSeek is 1. nSeek values greater than 1 come about if the
2786 ** WHERE clause includes "x IN (....)" terms used in place of "x=?". Or when
2787 ** implicit "x IN (SELECT x FROM tbl)" terms are added for skip-scans.
2789 ** The estimated values (nRow, nVisit, nSeek) often contain a large amount
2790 ** of uncertainty. For this reason, scoring is designed to pick plans that
2791 ** "do the least harm" if the estimates are inaccurate. For example, a
2792 ** log(nRow) factor is omitted from a non-covering index scan in order to
2793 ** bias the scoring in favor of using an index, since the worst-case
2794 ** performance of using an index is far better than the worst-case performance
2795 ** of a full table scan.
2797 static int whereLoopAddBtree(
2798 WhereLoopBuilder
*pBuilder
, /* WHERE clause information */
2799 Bitmask mPrereq
/* Extra prerequesites for using this table */
2801 WhereInfo
*pWInfo
; /* WHERE analysis context */
2802 Index
*pProbe
; /* An index we are evaluating */
2803 Index sPk
; /* A fake index object for the primary key */
2804 LogEst aiRowEstPk
[2]; /* The aiRowLogEst[] value for the sPk index */
2805 i16 aiColumnPk
= -1; /* The aColumn[] value for the sPk index */
2806 SrcList
*pTabList
; /* The FROM clause */
2807 struct SrcList_item
*pSrc
; /* The FROM clause btree term to add */
2808 WhereLoop
*pNew
; /* Template WhereLoop object */
2809 int rc
= SQLITE_OK
; /* Return code */
2810 int iSortIdx
= 1; /* Index number */
2811 int b
; /* A boolean value */
2812 LogEst rSize
; /* number of rows in the table */
2813 LogEst rLogSize
; /* Logarithm of the number of rows in the table */
2814 WhereClause
*pWC
; /* The parsed WHERE clause */
2815 Table
*pTab
; /* Table being queried */
2817 pNew
= pBuilder
->pNew
;
2818 pWInfo
= pBuilder
->pWInfo
;
2819 pTabList
= pWInfo
->pTabList
;
2820 pSrc
= pTabList
->a
+ pNew
->iTab
;
2822 pWC
= pBuilder
->pWC
;
2823 assert( !IsVirtual(pSrc
->pTab
) );
2825 if( pSrc
->pIBIndex
){
2826 /* An INDEXED BY clause specifies a particular index to use */
2827 pProbe
= pSrc
->pIBIndex
;
2828 }else if( !HasRowid(pTab
) ){
2829 pProbe
= pTab
->pIndex
;
2831 /* There is no INDEXED BY clause. Create a fake Index object in local
2832 ** variable sPk to represent the rowid primary key index. Make this
2833 ** fake index the first in a chain of Index objects with all of the real
2834 ** indices to follow */
2835 Index
*pFirst
; /* First of real indices on the table */
2836 memset(&sPk
, 0, sizeof(Index
));
2839 sPk
.aiColumn
= &aiColumnPk
;
2840 sPk
.aiRowLogEst
= aiRowEstPk
;
2841 sPk
.onError
= OE_Replace
;
2843 sPk
.szIdxRow
= pTab
->szTabRow
;
2844 aiRowEstPk
[0] = pTab
->nRowLogEst
;
2846 pFirst
= pSrc
->pTab
->pIndex
;
2847 if( pSrc
->fg
.notIndexed
==0 ){
2848 /* The real indices of the table are only considered if the
2849 ** NOT INDEXED qualifier is omitted from the FROM clause */
2854 rSize
= pTab
->nRowLogEst
;
2855 rLogSize
= estLog(rSize
);
2857 #ifndef SQLITE_OMIT_AUTOMATIC_INDEX
2858 /* Automatic indexes */
2859 if( !pBuilder
->pOrSet
/* Not part of an OR optimization */
2860 && (pWInfo
->wctrlFlags
& WHERE_OR_SUBCLAUSE
)==0
2861 && (pWInfo
->pParse
->db
->flags
& SQLITE_AutoIndex
)!=0
2862 && pSrc
->pIBIndex
==0 /* Has no INDEXED BY clause */
2863 && !pSrc
->fg
.notIndexed
/* Has no NOT INDEXED clause */
2864 && HasRowid(pTab
) /* Not WITHOUT ROWID table. (FIXME: Why not?) */
2865 && !pSrc
->fg
.isCorrelated
/* Not a correlated subquery */
2866 && !pSrc
->fg
.isRecursive
/* Not a recursive common table expression. */
2868 /* Generate auto-index WhereLoops */
2870 WhereTerm
*pWCEnd
= pWC
->a
+ pWC
->nTerm
;
2871 for(pTerm
=pWC
->a
; rc
==SQLITE_OK
&& pTerm
<pWCEnd
; pTerm
++){
2872 if( pTerm
->prereqRight
& pNew
->maskSelf
) continue;
2873 if( termCanDriveIndex(pTerm
, pSrc
, 0) ){
2874 pNew
->u
.btree
.nEq
= 1;
2876 pNew
->u
.btree
.pIndex
= 0;
2878 pNew
->aLTerm
[0] = pTerm
;
2879 /* TUNING: One-time cost for computing the automatic index is
2880 ** estimated to be X*N*log2(N) where N is the number of rows in
2881 ** the table being indexed and where X is 7 (LogEst=28) for normal
2882 ** tables or 0.5 (LogEst=-10) for views and subqueries. The value
2883 ** of X is smaller for views and subqueries so that the query planner
2884 ** will be more aggressive about generating automatic indexes for
2885 ** those objects, since there is no opportunity to add schema
2886 ** indexes on subqueries and views. */
2887 pNew
->rSetup
= rLogSize
+ rSize
;
2888 if( pTab
->pSelect
==0 && (pTab
->tabFlags
& TF_Ephemeral
)==0 ){
2893 ApplyCostMultiplier(pNew
->rSetup
, pTab
->costMult
);
2894 if( pNew
->rSetup
<0 ) pNew
->rSetup
= 0;
2895 /* TUNING: Each index lookup yields 20 rows in the table. This
2896 ** is more than the usual guess of 10 rows, since we have no way
2897 ** of knowing how selective the index will ultimately be. It would
2898 ** not be unreasonable to make this value much larger. */
2899 pNew
->nOut
= 43; assert( 43==sqlite3LogEst(20) );
2900 pNew
->rRun
= sqlite3LogEstAdd(rLogSize
,pNew
->nOut
);
2901 pNew
->wsFlags
= WHERE_AUTO_INDEX
;
2902 pNew
->prereq
= mPrereq
| pTerm
->prereqRight
;
2903 rc
= whereLoopInsert(pBuilder
, pNew
);
2907 #endif /* SQLITE_OMIT_AUTOMATIC_INDEX */
2909 /* Loop over all indices. If there was an INDEXED BY clause, then only
2910 ** consider index pProbe. */
2911 for(; rc
==SQLITE_OK
&& pProbe
;
2912 pProbe
=(pSrc
->pIBIndex
? 0 : pProbe
->pNext
), iSortIdx
++
2914 if( pProbe
->pPartIdxWhere
!=0
2915 && !whereUsablePartialIndex(pSrc
->iCursor
, pWC
, pProbe
->pPartIdxWhere
) ){
2916 testcase( pNew
->iTab
!=pSrc
->iCursor
); /* See ticket [98d973b8f5] */
2917 continue; /* Partial index inappropriate for this query */
2919 if( pProbe
->bNoQuery
) continue;
2920 rSize
= pProbe
->aiRowLogEst
[0];
2921 pNew
->u
.btree
.nEq
= 0;
2922 pNew
->u
.btree
.nBtm
= 0;
2923 pNew
->u
.btree
.nTop
= 0;
2928 pNew
->prereq
= mPrereq
;
2930 pNew
->u
.btree
.pIndex
= pProbe
;
2931 b
= indexMightHelpWithOrderBy(pBuilder
, pProbe
, pSrc
->iCursor
);
2932 /* The ONEPASS_DESIRED flags never occurs together with ORDER BY */
2933 assert( (pWInfo
->wctrlFlags
& WHERE_ONEPASS_DESIRED
)==0 || b
==0 );
2934 if( pProbe
->tnum
<=0 ){
2935 /* Integer primary key index */
2936 pNew
->wsFlags
= WHERE_IPK
;
2938 /* Full table scan */
2939 pNew
->iSortIdx
= b
? iSortIdx
: 0;
2940 /* TUNING: Cost of full table scan is (N*3.0). */
2941 pNew
->rRun
= rSize
+ 16;
2942 ApplyCostMultiplier(pNew
->rRun
, pTab
->costMult
);
2943 whereLoopOutputAdjust(pWC
, pNew
, rSize
);
2944 rc
= whereLoopInsert(pBuilder
, pNew
);
2949 if( pProbe
->isCovering
){
2950 pNew
->wsFlags
= WHERE_IDX_ONLY
| WHERE_INDEXED
;
2953 m
= pSrc
->colUsed
& pProbe
->colNotIdxed
;
2954 pNew
->wsFlags
= (m
==0) ? (WHERE_IDX_ONLY
|WHERE_INDEXED
) : WHERE_INDEXED
;
2957 /* Full scan via index */
2960 || pProbe
->pPartIdxWhere
!=0
2962 && pProbe
->bUnordered
==0
2963 && (pProbe
->szIdxRow
<pTab
->szTabRow
)
2964 && (pWInfo
->wctrlFlags
& WHERE_ONEPASS_DESIRED
)==0
2965 && sqlite3GlobalConfig
.bUseCis
2966 && OptimizationEnabled(pWInfo
->pParse
->db
, SQLITE_CoverIdxScan
)
2969 pNew
->iSortIdx
= b
? iSortIdx
: 0;
2971 /* The cost of visiting the index rows is N*K, where K is
2972 ** between 1.1 and 3.0, depending on the relative sizes of the
2973 ** index and table rows. */
2974 pNew
->rRun
= rSize
+ 1 + (15*pProbe
->szIdxRow
)/pTab
->szTabRow
;
2976 /* If this is a non-covering index scan, add in the cost of
2977 ** doing table lookups. The cost will be 3x the number of
2978 ** lookups. Take into account WHERE clause terms that can be
2979 ** satisfied using just the index, and that do not require a
2981 LogEst nLookup
= rSize
+ 16; /* Base cost: N*3 */
2983 int iCur
= pSrc
->iCursor
;
2984 WhereClause
*pWC2
= &pWInfo
->sWC
;
2985 for(ii
=0; ii
<pWC2
->nTerm
; ii
++){
2986 WhereTerm
*pTerm
= &pWC2
->a
[ii
];
2987 if( !sqlite3ExprCoveredByIndex(pTerm
->pExpr
, iCur
, pProbe
) ){
2990 /* pTerm can be evaluated using just the index. So reduce
2991 ** the expected number of table lookups accordingly */
2992 if( pTerm
->truthProb
<=0 ){
2993 nLookup
+= pTerm
->truthProb
;
2996 if( pTerm
->eOperator
& (WO_EQ
|WO_IS
) ) nLookup
-= 19;
3000 pNew
->rRun
= sqlite3LogEstAdd(pNew
->rRun
, nLookup
);
3002 ApplyCostMultiplier(pNew
->rRun
, pTab
->costMult
);
3003 whereLoopOutputAdjust(pWC
, pNew
, rSize
);
3004 rc
= whereLoopInsert(pBuilder
, pNew
);
3010 pBuilder
->bldFlags
= 0;
3011 rc
= whereLoopAddBtreeIndex(pBuilder
, pSrc
, pProbe
, 0);
3012 if( pBuilder
->bldFlags
==SQLITE_BLDF_INDEXED
){
3013 /* If a non-unique index is used, or if a prefix of the key for
3014 ** unique index is used (making the index functionally non-unique)
3015 ** then the sqlite_stat1 data becomes important for scoring the
3017 pTab
->tabFlags
|= TF_StatsUsed
;
3019 #ifdef SQLITE_ENABLE_STAT3_OR_STAT4
3020 sqlite3Stat4ProbeFree(pBuilder
->pRec
);
3021 pBuilder
->nRecValid
= 0;
3028 #ifndef SQLITE_OMIT_VIRTUALTABLE
3031 ** Argument pIdxInfo is already populated with all constraints that may
3032 ** be used by the virtual table identified by pBuilder->pNew->iTab. This
3033 ** function marks a subset of those constraints usable, invokes the
3034 ** xBestIndex method and adds the returned plan to pBuilder.
3036 ** A constraint is marked usable if:
3038 ** * Argument mUsable indicates that its prerequisites are available, and
3040 ** * It is not one of the operators specified in the mExclude mask passed
3041 ** as the fourth argument (which in practice is either WO_IN or 0).
3043 ** Argument mPrereq is a mask of tables that must be scanned before the
3044 ** virtual table in question. These are added to the plans prerequisites
3045 ** before it is added to pBuilder.
3047 ** Output parameter *pbIn is set to true if the plan added to pBuilder
3048 ** uses one or more WO_IN terms, or false otherwise.
3050 static int whereLoopAddVirtualOne(
3051 WhereLoopBuilder
*pBuilder
,
3052 Bitmask mPrereq
, /* Mask of tables that must be used. */
3053 Bitmask mUsable
, /* Mask of usable tables */
3054 u16 mExclude
, /* Exclude terms using these operators */
3055 sqlite3_index_info
*pIdxInfo
, /* Populated object for xBestIndex */
3056 u16 mNoOmit
, /* Do not omit these constraints */
3057 int *pbIn
/* OUT: True if plan uses an IN(...) op */
3059 WhereClause
*pWC
= pBuilder
->pWC
;
3060 struct sqlite3_index_constraint
*pIdxCons
;
3061 struct sqlite3_index_constraint_usage
*pUsage
= pIdxInfo
->aConstraintUsage
;
3065 WhereLoop
*pNew
= pBuilder
->pNew
;
3066 Parse
*pParse
= pBuilder
->pWInfo
->pParse
;
3067 struct SrcList_item
*pSrc
= &pBuilder
->pWInfo
->pTabList
->a
[pNew
->iTab
];
3068 int nConstraint
= pIdxInfo
->nConstraint
;
3070 assert( (mUsable
& mPrereq
)==mPrereq
);
3072 pNew
->prereq
= mPrereq
;
3074 /* Set the usable flag on the subset of constraints identified by
3075 ** arguments mUsable and mExclude. */
3076 pIdxCons
= *(struct sqlite3_index_constraint
**)&pIdxInfo
->aConstraint
;
3077 for(i
=0; i
<nConstraint
; i
++, pIdxCons
++){
3078 WhereTerm
*pTerm
= &pWC
->a
[pIdxCons
->iTermOffset
];
3079 pIdxCons
->usable
= 0;
3080 if( (pTerm
->prereqRight
& mUsable
)==pTerm
->prereqRight
3081 && (pTerm
->eOperator
& mExclude
)==0
3083 pIdxCons
->usable
= 1;
3087 /* Initialize the output fields of the sqlite3_index_info structure */
3088 memset(pUsage
, 0, sizeof(pUsage
[0])*nConstraint
);
3089 assert( pIdxInfo
->needToFreeIdxStr
==0 );
3090 pIdxInfo
->idxStr
= 0;
3091 pIdxInfo
->idxNum
= 0;
3092 pIdxInfo
->orderByConsumed
= 0;
3093 pIdxInfo
->estimatedCost
= SQLITE_BIG_DBL
/ (double)2;
3094 pIdxInfo
->estimatedRows
= 25;
3095 pIdxInfo
->idxFlags
= 0;
3096 pIdxInfo
->colUsed
= (sqlite3_int64
)pSrc
->colUsed
;
3098 /* Invoke the virtual table xBestIndex() method */
3099 rc
= vtabBestIndex(pParse
, pSrc
->pTab
, pIdxInfo
);
3103 assert( pNew
->nLSlot
>=nConstraint
);
3104 for(i
=0; i
<nConstraint
; i
++) pNew
->aLTerm
[i
] = 0;
3105 pNew
->u
.vtab
.omitMask
= 0;
3106 pIdxCons
= *(struct sqlite3_index_constraint
**)&pIdxInfo
->aConstraint
;
3107 for(i
=0; i
<nConstraint
; i
++, pIdxCons
++){
3109 if( (iTerm
= pUsage
[i
].argvIndex
- 1)>=0 ){
3111 int j
= pIdxCons
->iTermOffset
;
3112 if( iTerm
>=nConstraint
3115 || pNew
->aLTerm
[iTerm
]!=0
3116 || pIdxCons
->usable
==0
3118 sqlite3ErrorMsg(pParse
,"%s.xBestIndex malfunction",pSrc
->pTab
->zName
);
3119 testcase( pIdxInfo
->needToFreeIdxStr
);
3120 return SQLITE_ERROR
;
3122 testcase( iTerm
==nConstraint
-1 );
3124 testcase( j
==pWC
->nTerm
-1 );
3126 pNew
->prereq
|= pTerm
->prereqRight
;
3127 assert( iTerm
<pNew
->nLSlot
);
3128 pNew
->aLTerm
[iTerm
] = pTerm
;
3129 if( iTerm
>mxTerm
) mxTerm
= iTerm
;
3130 testcase( iTerm
==15 );
3131 testcase( iTerm
==16 );
3132 if( iTerm
<16 && pUsage
[i
].omit
) pNew
->u
.vtab
.omitMask
|= 1<<iTerm
;
3133 if( (pTerm
->eOperator
& WO_IN
)!=0 ){
3134 /* A virtual table that is constrained by an IN clause may not
3135 ** consume the ORDER BY clause because (1) the order of IN terms
3136 ** is not necessarily related to the order of output terms and
3137 ** (2) Multiple outputs from a single IN value will not merge
3139 pIdxInfo
->orderByConsumed
= 0;
3140 pIdxInfo
->idxFlags
&= ~SQLITE_INDEX_SCAN_UNIQUE
;
3141 *pbIn
= 1; assert( (mExclude
& WO_IN
)==0 );
3145 pNew
->u
.vtab
.omitMask
&= ~mNoOmit
;
3147 pNew
->nLTerm
= mxTerm
+1;
3148 for(i
=0; i
<=mxTerm
; i
++){
3149 if( pNew
->aLTerm
[i
]==0 ){
3150 /* The non-zero argvIdx values must be contiguous. Raise an
3151 ** error if they are not */
3152 sqlite3ErrorMsg(pParse
,"%s.xBestIndex malfunction",pSrc
->pTab
->zName
);
3153 testcase( pIdxInfo
->needToFreeIdxStr
);
3154 return SQLITE_ERROR
;
3157 assert( pNew
->nLTerm
<=pNew
->nLSlot
);
3158 pNew
->u
.vtab
.idxNum
= pIdxInfo
->idxNum
;
3159 pNew
->u
.vtab
.needFree
= pIdxInfo
->needToFreeIdxStr
;
3160 pIdxInfo
->needToFreeIdxStr
= 0;
3161 pNew
->u
.vtab
.idxStr
= pIdxInfo
->idxStr
;
3162 pNew
->u
.vtab
.isOrdered
= (i8
)(pIdxInfo
->orderByConsumed
?
3163 pIdxInfo
->nOrderBy
: 0);
3165 pNew
->rRun
= sqlite3LogEstFromDouble(pIdxInfo
->estimatedCost
);
3166 pNew
->nOut
= sqlite3LogEst(pIdxInfo
->estimatedRows
);
3168 /* Set the WHERE_ONEROW flag if the xBestIndex() method indicated
3169 ** that the scan will visit at most one row. Clear it otherwise. */
3170 if( pIdxInfo
->idxFlags
& SQLITE_INDEX_SCAN_UNIQUE
){
3171 pNew
->wsFlags
|= WHERE_ONEROW
;
3173 pNew
->wsFlags
&= ~WHERE_ONEROW
;
3175 rc
= whereLoopInsert(pBuilder
, pNew
);
3176 if( pNew
->u
.vtab
.needFree
){
3177 sqlite3_free(pNew
->u
.vtab
.idxStr
);
3178 pNew
->u
.vtab
.needFree
= 0;
3180 WHERETRACE(0xffff, (" bIn=%d prereqIn=%04llx prereqOut=%04llx\n",
3181 *pbIn
, (sqlite3_uint64
)mPrereq
,
3182 (sqlite3_uint64
)(pNew
->prereq
& ~mPrereq
)));
3188 ** If this function is invoked from within an xBestIndex() callback, it
3189 ** returns a pointer to a buffer containing the name of the collation
3190 ** sequence associated with element iCons of the sqlite3_index_info.aConstraint
3191 ** array. Or, if iCons is out of range or there is no active xBestIndex
3192 ** call, return NULL.
3194 const char *sqlite3_vtab_collation(sqlite3_index_info
*pIdxInfo
, int iCons
){
3195 HiddenIndexInfo
*pHidden
= (HiddenIndexInfo
*)&pIdxInfo
[1];
3196 const char *zRet
= 0;
3197 if( iCons
>=0 && iCons
<pIdxInfo
->nConstraint
){
3199 int iTerm
= pIdxInfo
->aConstraint
[iCons
].iTermOffset
;
3200 Expr
*pX
= pHidden
->pWC
->a
[iTerm
].pExpr
;
3202 pC
= sqlite3BinaryCompareCollSeq(pHidden
->pParse
, pX
->pLeft
, pX
->pRight
);
3204 zRet
= (pC
? pC
->zName
: "BINARY");
3210 ** Add all WhereLoop objects for a table of the join identified by
3211 ** pBuilder->pNew->iTab. That table is guaranteed to be a virtual table.
3213 ** If there are no LEFT or CROSS JOIN joins in the query, both mPrereq and
3214 ** mUnusable are set to 0. Otherwise, mPrereq is a mask of all FROM clause
3215 ** entries that occur before the virtual table in the FROM clause and are
3216 ** separated from it by at least one LEFT or CROSS JOIN. Similarly, the
3217 ** mUnusable mask contains all FROM clause entries that occur after the
3218 ** virtual table and are separated from it by at least one LEFT or
3221 ** For example, if the query were:
3223 ** ... FROM t1, t2 LEFT JOIN t3, t4, vt CROSS JOIN t5, t6;
3225 ** then mPrereq corresponds to (t1, t2) and mUnusable to (t5, t6).
3227 ** All the tables in mPrereq must be scanned before the current virtual
3228 ** table. So any terms for which all prerequisites are satisfied by
3229 ** mPrereq may be specified as "usable" in all calls to xBestIndex.
3230 ** Conversely, all tables in mUnusable must be scanned after the current
3231 ** virtual table, so any terms for which the prerequisites overlap with
3232 ** mUnusable should always be configured as "not-usable" for xBestIndex.
3234 static int whereLoopAddVirtual(
3235 WhereLoopBuilder
*pBuilder
, /* WHERE clause information */
3236 Bitmask mPrereq
, /* Tables that must be scanned before this one */
3237 Bitmask mUnusable
/* Tables that must be scanned after this one */
3239 int rc
= SQLITE_OK
; /* Return code */
3240 WhereInfo
*pWInfo
; /* WHERE analysis context */
3241 Parse
*pParse
; /* The parsing context */
3242 WhereClause
*pWC
; /* The WHERE clause */
3243 struct SrcList_item
*pSrc
; /* The FROM clause term to search */
3244 sqlite3_index_info
*p
; /* Object to pass to xBestIndex() */
3245 int nConstraint
; /* Number of constraints in p */
3246 int bIn
; /* True if plan uses IN(...) operator */
3248 Bitmask mBest
; /* Tables used by best possible plan */
3251 assert( (mPrereq
& mUnusable
)==0 );
3252 pWInfo
= pBuilder
->pWInfo
;
3253 pParse
= pWInfo
->pParse
;
3254 pWC
= pBuilder
->pWC
;
3255 pNew
= pBuilder
->pNew
;
3256 pSrc
= &pWInfo
->pTabList
->a
[pNew
->iTab
];
3257 assert( IsVirtual(pSrc
->pTab
) );
3258 p
= allocateIndexInfo(pParse
, pWC
, mUnusable
, pSrc
, pBuilder
->pOrderBy
,
3260 if( p
==0 ) return SQLITE_NOMEM_BKPT
;
3262 pNew
->wsFlags
= WHERE_VIRTUALTABLE
;
3264 pNew
->u
.vtab
.needFree
= 0;
3265 nConstraint
= p
->nConstraint
;
3266 if( whereLoopResize(pParse
->db
, pNew
, nConstraint
) ){
3267 sqlite3DbFree(pParse
->db
, p
);
3268 return SQLITE_NOMEM_BKPT
;
3271 /* First call xBestIndex() with all constraints usable. */
3272 WHERETRACE(0x800, ("BEGIN %s.addVirtual()\n", pSrc
->pTab
->zName
));
3273 WHERETRACE(0x40, (" VirtualOne: all usable\n"));
3274 rc
= whereLoopAddVirtualOne(pBuilder
, mPrereq
, ALLBITS
, 0, p
, mNoOmit
, &bIn
);
3276 /* If the call to xBestIndex() with all terms enabled produced a plan
3277 ** that does not require any source tables (IOW: a plan with mBest==0),
3278 ** then there is no point in making any further calls to xBestIndex()
3279 ** since they will all return the same result (if the xBestIndex()
3280 ** implementation is sane). */
3281 if( rc
==SQLITE_OK
&& (mBest
= (pNew
->prereq
& ~mPrereq
))!=0 ){
3282 int seenZero
= 0; /* True if a plan with no prereqs seen */
3283 int seenZeroNoIN
= 0; /* Plan with no prereqs and no IN(...) seen */
3285 Bitmask mBestNoIn
= 0;
3287 /* If the plan produced by the earlier call uses an IN(...) term, call
3288 ** xBestIndex again, this time with IN(...) terms disabled. */
3290 WHERETRACE(0x40, (" VirtualOne: all usable w/o IN\n"));
3291 rc
= whereLoopAddVirtualOne(
3292 pBuilder
, mPrereq
, ALLBITS
, WO_IN
, p
, mNoOmit
, &bIn
);
3294 mBestNoIn
= pNew
->prereq
& ~mPrereq
;
3301 /* Call xBestIndex once for each distinct value of (prereqRight & ~mPrereq)
3302 ** in the set of terms that apply to the current virtual table. */
3303 while( rc
==SQLITE_OK
){
3305 Bitmask mNext
= ALLBITS
;
3307 for(i
=0; i
<nConstraint
; i
++){
3309 pWC
->a
[p
->aConstraint
[i
].iTermOffset
].prereqRight
& ~mPrereq
3311 if( mThis
>mPrev
&& mThis
<mNext
) mNext
= mThis
;
3314 if( mNext
==ALLBITS
) break;
3315 if( mNext
==mBest
|| mNext
==mBestNoIn
) continue;
3316 WHERETRACE(0x40, (" VirtualOne: mPrev=%04llx mNext=%04llx\n",
3317 (sqlite3_uint64
)mPrev
, (sqlite3_uint64
)mNext
));
3318 rc
= whereLoopAddVirtualOne(
3319 pBuilder
, mPrereq
, mNext
|mPrereq
, 0, p
, mNoOmit
, &bIn
);
3320 if( pNew
->prereq
==mPrereq
){
3322 if( bIn
==0 ) seenZeroNoIN
= 1;
3326 /* If the calls to xBestIndex() in the above loop did not find a plan
3327 ** that requires no source tables at all (i.e. one guaranteed to be
3328 ** usable), make a call here with all source tables disabled */
3329 if( rc
==SQLITE_OK
&& seenZero
==0 ){
3330 WHERETRACE(0x40, (" VirtualOne: all disabled\n"));
3331 rc
= whereLoopAddVirtualOne(
3332 pBuilder
, mPrereq
, mPrereq
, 0, p
, mNoOmit
, &bIn
);
3333 if( bIn
==0 ) seenZeroNoIN
= 1;
3336 /* If the calls to xBestIndex() have so far failed to find a plan
3337 ** that requires no source tables at all and does not use an IN(...)
3338 ** operator, make a final call to obtain one here. */
3339 if( rc
==SQLITE_OK
&& seenZeroNoIN
==0 ){
3340 WHERETRACE(0x40, (" VirtualOne: all disabled and w/o IN\n"));
3341 rc
= whereLoopAddVirtualOne(
3342 pBuilder
, mPrereq
, mPrereq
, WO_IN
, p
, mNoOmit
, &bIn
);
3346 if( p
->needToFreeIdxStr
) sqlite3_free(p
->idxStr
);
3347 sqlite3DbFreeNN(pParse
->db
, p
);
3348 WHERETRACE(0x800, ("END %s.addVirtual(), rc=%d\n", pSrc
->pTab
->zName
, rc
));
3351 #endif /* SQLITE_OMIT_VIRTUALTABLE */
3354 ** Add WhereLoop entries to handle OR terms. This works for either
3355 ** btrees or virtual tables.
3357 static int whereLoopAddOr(
3358 WhereLoopBuilder
*pBuilder
,
3362 WhereInfo
*pWInfo
= pBuilder
->pWInfo
;
3365 WhereTerm
*pTerm
, *pWCEnd
;
3369 WhereLoopBuilder sSubBuild
;
3370 WhereOrSet sSum
, sCur
;
3371 struct SrcList_item
*pItem
;
3373 pWC
= pBuilder
->pWC
;
3374 pWCEnd
= pWC
->a
+ pWC
->nTerm
;
3375 pNew
= pBuilder
->pNew
;
3376 memset(&sSum
, 0, sizeof(sSum
));
3377 pItem
= pWInfo
->pTabList
->a
+ pNew
->iTab
;
3378 iCur
= pItem
->iCursor
;
3380 for(pTerm
=pWC
->a
; pTerm
<pWCEnd
&& rc
==SQLITE_OK
; pTerm
++){
3381 if( (pTerm
->eOperator
& WO_OR
)!=0
3382 && (pTerm
->u
.pOrInfo
->indexable
& pNew
->maskSelf
)!=0
3384 WhereClause
* const pOrWC
= &pTerm
->u
.pOrInfo
->wc
;
3385 WhereTerm
* const pOrWCEnd
= &pOrWC
->a
[pOrWC
->nTerm
];
3390 sSubBuild
= *pBuilder
;
3391 sSubBuild
.pOrderBy
= 0;
3392 sSubBuild
.pOrSet
= &sCur
;
3394 WHERETRACE(0x200, ("Begin processing OR-clause %p\n", pTerm
));
3395 for(pOrTerm
=pOrWC
->a
; pOrTerm
<pOrWCEnd
; pOrTerm
++){
3396 if( (pOrTerm
->eOperator
& WO_AND
)!=0 ){
3397 sSubBuild
.pWC
= &pOrTerm
->u
.pAndInfo
->wc
;
3398 }else if( pOrTerm
->leftCursor
==iCur
){
3399 tempWC
.pWInfo
= pWC
->pWInfo
;
3400 tempWC
.pOuter
= pWC
;
3404 sSubBuild
.pWC
= &tempWC
;
3409 #ifdef WHERETRACE_ENABLED
3410 WHERETRACE(0x200, ("OR-term %d of %p has %d subterms:\n",
3411 (int)(pOrTerm
-pOrWC
->a
), pTerm
, sSubBuild
.pWC
->nTerm
));
3412 if( sqlite3WhereTrace
& 0x400 ){
3413 sqlite3WhereClausePrint(sSubBuild
.pWC
);
3416 #ifndef SQLITE_OMIT_VIRTUALTABLE
3417 if( IsVirtual(pItem
->pTab
) ){
3418 rc
= whereLoopAddVirtual(&sSubBuild
, mPrereq
, mUnusable
);
3422 rc
= whereLoopAddBtree(&sSubBuild
, mPrereq
);
3424 if( rc
==SQLITE_OK
){
3425 rc
= whereLoopAddOr(&sSubBuild
, mPrereq
, mUnusable
);
3427 assert( rc
==SQLITE_OK
|| sCur
.n
==0 );
3432 whereOrMove(&sSum
, &sCur
);
3436 whereOrMove(&sPrev
, &sSum
);
3438 for(i
=0; i
<sPrev
.n
; i
++){
3439 for(j
=0; j
<sCur
.n
; j
++){
3440 whereOrInsert(&sSum
, sPrev
.a
[i
].prereq
| sCur
.a
[j
].prereq
,
3441 sqlite3LogEstAdd(sPrev
.a
[i
].rRun
, sCur
.a
[j
].rRun
),
3442 sqlite3LogEstAdd(sPrev
.a
[i
].nOut
, sCur
.a
[j
].nOut
));
3448 pNew
->aLTerm
[0] = pTerm
;
3449 pNew
->wsFlags
= WHERE_MULTI_OR
;
3452 memset(&pNew
->u
, 0, sizeof(pNew
->u
));
3453 for(i
=0; rc
==SQLITE_OK
&& i
<sSum
.n
; i
++){
3454 /* TUNING: Currently sSum.a[i].rRun is set to the sum of the costs
3455 ** of all sub-scans required by the OR-scan. However, due to rounding
3456 ** errors, it may be that the cost of the OR-scan is equal to its
3457 ** most expensive sub-scan. Add the smallest possible penalty
3458 ** (equivalent to multiplying the cost by 1.07) to ensure that
3459 ** this does not happen. Otherwise, for WHERE clauses such as the
3460 ** following where there is an index on "y":
3462 ** WHERE likelihood(x=?, 0.99) OR y=?
3464 ** the planner may elect to "OR" together a full-table scan and an
3465 ** index lookup. And other similarly odd results. */
3466 pNew
->rRun
= sSum
.a
[i
].rRun
+ 1;
3467 pNew
->nOut
= sSum
.a
[i
].nOut
;
3468 pNew
->prereq
= sSum
.a
[i
].prereq
;
3469 rc
= whereLoopInsert(pBuilder
, pNew
);
3471 WHERETRACE(0x200, ("End processing OR-clause %p\n", pTerm
));
3478 ** Add all WhereLoop objects for all tables
3480 static int whereLoopAddAll(WhereLoopBuilder
*pBuilder
){
3481 WhereInfo
*pWInfo
= pBuilder
->pWInfo
;
3482 Bitmask mPrereq
= 0;
3485 SrcList
*pTabList
= pWInfo
->pTabList
;
3486 struct SrcList_item
*pItem
;
3487 struct SrcList_item
*pEnd
= &pTabList
->a
[pWInfo
->nLevel
];
3488 sqlite3
*db
= pWInfo
->pParse
->db
;
3491 u8 priorJointype
= 0;
3493 /* Loop over the tables in the join, from left to right */
3494 pNew
= pBuilder
->pNew
;
3495 whereLoopInit(pNew
);
3496 for(iTab
=0, pItem
=pTabList
->a
; pItem
<pEnd
; iTab
++, pItem
++){
3497 Bitmask mUnusable
= 0;
3499 pNew
->maskSelf
= sqlite3WhereGetMask(&pWInfo
->sMaskSet
, pItem
->iCursor
);
3500 if( ((pItem
->fg
.jointype
|priorJointype
) & (JT_LEFT
|JT_CROSS
))!=0 ){
3501 /* This condition is true when pItem is the FROM clause term on the
3502 ** right-hand-side of a LEFT or CROSS JOIN. */
3505 priorJointype
= pItem
->fg
.jointype
;
3506 #ifndef SQLITE_OMIT_VIRTUALTABLE
3507 if( IsVirtual(pItem
->pTab
) ){
3508 struct SrcList_item
*p
;
3509 for(p
=&pItem
[1]; p
<pEnd
; p
++){
3510 if( mUnusable
|| (p
->fg
.jointype
& (JT_LEFT
|JT_CROSS
)) ){
3511 mUnusable
|= sqlite3WhereGetMask(&pWInfo
->sMaskSet
, p
->iCursor
);
3514 rc
= whereLoopAddVirtual(pBuilder
, mPrereq
, mUnusable
);
3516 #endif /* SQLITE_OMIT_VIRTUALTABLE */
3518 rc
= whereLoopAddBtree(pBuilder
, mPrereq
);
3520 if( rc
==SQLITE_OK
&& pBuilder
->pWC
->hasOr
){
3521 rc
= whereLoopAddOr(pBuilder
, mPrereq
, mUnusable
);
3523 mPrior
|= pNew
->maskSelf
;
3524 if( rc
|| db
->mallocFailed
) break;
3527 whereLoopClear(db
, pNew
);
3532 ** Examine a WherePath (with the addition of the extra WhereLoop of the 6th
3533 ** parameters) to see if it outputs rows in the requested ORDER BY
3534 ** (or GROUP BY) without requiring a separate sort operation. Return N:
3536 ** N>0: N terms of the ORDER BY clause are satisfied
3537 ** N==0: No terms of the ORDER BY clause are satisfied
3538 ** N<0: Unknown yet how many terms of ORDER BY might be satisfied.
3540 ** Note that processing for WHERE_GROUPBY and WHERE_DISTINCTBY is not as
3541 ** strict. With GROUP BY and DISTINCT the only requirement is that
3542 ** equivalent rows appear immediately adjacent to one another. GROUP BY
3543 ** and DISTINCT do not require rows to appear in any particular order as long
3544 ** as equivalent rows are grouped together. Thus for GROUP BY and DISTINCT
3545 ** the pOrderBy terms can be matched in any order. With ORDER BY, the
3546 ** pOrderBy terms must be matched in strict left-to-right order.
3548 static i8
wherePathSatisfiesOrderBy(
3549 WhereInfo
*pWInfo
, /* The WHERE clause */
3550 ExprList
*pOrderBy
, /* ORDER BY or GROUP BY or DISTINCT clause to check */
3551 WherePath
*pPath
, /* The WherePath to check */
3552 u16 wctrlFlags
, /* WHERE_GROUPBY or _DISTINCTBY or _ORDERBY_LIMIT */
3553 u16 nLoop
, /* Number of entries in pPath->aLoop[] */
3554 WhereLoop
*pLast
, /* Add this WhereLoop to the end of pPath->aLoop[] */
3555 Bitmask
*pRevMask
/* OUT: Mask of WhereLoops to run in reverse order */
3557 u8 revSet
; /* True if rev is known */
3558 u8 rev
; /* Composite sort order */
3559 u8 revIdx
; /* Index sort order */
3560 u8 isOrderDistinct
; /* All prior WhereLoops are order-distinct */
3561 u8 distinctColumns
; /* True if the loop has UNIQUE NOT NULL columns */
3562 u8 isMatch
; /* iColumn matches a term of the ORDER BY clause */
3563 u16 eqOpMask
; /* Allowed equality operators */
3564 u16 nKeyCol
; /* Number of key columns in pIndex */
3565 u16 nColumn
; /* Total number of ordered columns in the index */
3566 u16 nOrderBy
; /* Number terms in the ORDER BY clause */
3567 int iLoop
; /* Index of WhereLoop in pPath being processed */
3568 int i
, j
; /* Loop counters */
3569 int iCur
; /* Cursor number for current WhereLoop */
3570 int iColumn
; /* A column number within table iCur */
3571 WhereLoop
*pLoop
= 0; /* Current WhereLoop being processed. */
3572 WhereTerm
*pTerm
; /* A single term of the WHERE clause */
3573 Expr
*pOBExpr
; /* An expression from the ORDER BY clause */
3574 CollSeq
*pColl
; /* COLLATE function from an ORDER BY clause term */
3575 Index
*pIndex
; /* The index associated with pLoop */
3576 sqlite3
*db
= pWInfo
->pParse
->db
; /* Database connection */
3577 Bitmask obSat
= 0; /* Mask of ORDER BY terms satisfied so far */
3578 Bitmask obDone
; /* Mask of all ORDER BY terms */
3579 Bitmask orderDistinctMask
; /* Mask of all well-ordered loops */
3580 Bitmask ready
; /* Mask of inner loops */
3583 ** We say the WhereLoop is "one-row" if it generates no more than one
3584 ** row of output. A WhereLoop is one-row if all of the following are true:
3585 ** (a) All index columns match with WHERE_COLUMN_EQ.
3586 ** (b) The index is unique
3587 ** Any WhereLoop with an WHERE_COLUMN_EQ constraint on the rowid is one-row.
3588 ** Every one-row WhereLoop will have the WHERE_ONEROW bit set in wsFlags.
3590 ** We say the WhereLoop is "order-distinct" if the set of columns from
3591 ** that WhereLoop that are in the ORDER BY clause are different for every
3592 ** row of the WhereLoop. Every one-row WhereLoop is automatically
3593 ** order-distinct. A WhereLoop that has no columns in the ORDER BY clause
3594 ** is not order-distinct. To be order-distinct is not quite the same as being
3595 ** UNIQUE since a UNIQUE column or index can have multiple rows that
3596 ** are NULL and NULL values are equivalent for the purpose of order-distinct.
3597 ** To be order-distinct, the columns must be UNIQUE and NOT NULL.
3599 ** The rowid for a table is always UNIQUE and NOT NULL so whenever the
3600 ** rowid appears in the ORDER BY clause, the corresponding WhereLoop is
3601 ** automatically order-distinct.
3604 assert( pOrderBy
!=0 );
3605 if( nLoop
&& OptimizationDisabled(db
, SQLITE_OrderByIdxJoin
) ) return 0;
3607 nOrderBy
= pOrderBy
->nExpr
;
3608 testcase( nOrderBy
==BMS
-1 );
3609 if( nOrderBy
>BMS
-1 ) return 0; /* Cannot optimize overly large ORDER BYs */
3610 isOrderDistinct
= 1;
3611 obDone
= MASKBIT(nOrderBy
)-1;
3612 orderDistinctMask
= 0;
3614 eqOpMask
= WO_EQ
| WO_IS
| WO_ISNULL
;
3615 if( wctrlFlags
& WHERE_ORDERBY_LIMIT
) eqOpMask
|= WO_IN
;
3616 for(iLoop
=0; isOrderDistinct
&& obSat
<obDone
&& iLoop
<=nLoop
; iLoop
++){
3617 if( iLoop
>0 ) ready
|= pLoop
->maskSelf
;
3619 pLoop
= pPath
->aLoop
[iLoop
];
3620 if( wctrlFlags
& WHERE_ORDERBY_LIMIT
) continue;
3624 if( pLoop
->wsFlags
& WHERE_VIRTUALTABLE
){
3625 if( pLoop
->u
.vtab
.isOrdered
) obSat
= obDone
;
3628 pLoop
->u
.btree
.nIdxCol
= 0;
3630 iCur
= pWInfo
->pTabList
->a
[pLoop
->iTab
].iCursor
;
3632 /* Mark off any ORDER BY term X that is a column in the table of
3633 ** the current loop for which there is term in the WHERE
3634 ** clause of the form X IS NULL or X=? that reference only outer
3637 for(i
=0; i
<nOrderBy
; i
++){
3638 if( MASKBIT(i
) & obSat
) continue;
3639 pOBExpr
= sqlite3ExprSkipCollate(pOrderBy
->a
[i
].pExpr
);
3640 if( pOBExpr
->op
!=TK_COLUMN
) continue;
3641 if( pOBExpr
->iTable
!=iCur
) continue;
3642 pTerm
= sqlite3WhereFindTerm(&pWInfo
->sWC
, iCur
, pOBExpr
->iColumn
,
3643 ~ready
, eqOpMask
, 0);
3644 if( pTerm
==0 ) continue;
3645 if( pTerm
->eOperator
==WO_IN
){
3646 /* IN terms are only valid for sorting in the ORDER BY LIMIT
3647 ** optimization, and then only if they are actually used
3648 ** by the query plan */
3649 assert( wctrlFlags
& WHERE_ORDERBY_LIMIT
);
3650 for(j
=0; j
<pLoop
->nLTerm
&& pTerm
!=pLoop
->aLTerm
[j
]; j
++){}
3651 if( j
>=pLoop
->nLTerm
) continue;
3653 if( (pTerm
->eOperator
&(WO_EQ
|WO_IS
))!=0 && pOBExpr
->iColumn
>=0 ){
3654 if( sqlite3ExprCollSeqMatch(pWInfo
->pParse
,
3655 pOrderBy
->a
[i
].pExpr
, pTerm
->pExpr
)==0 ){
3658 testcase( pTerm
->pExpr
->op
==TK_IS
);
3660 obSat
|= MASKBIT(i
);
3663 if( (pLoop
->wsFlags
& WHERE_ONEROW
)==0 ){
3664 if( pLoop
->wsFlags
& WHERE_IPK
){
3668 }else if( (pIndex
= pLoop
->u
.btree
.pIndex
)==0 || pIndex
->bUnordered
){
3671 nKeyCol
= pIndex
->nKeyCol
;
3672 nColumn
= pIndex
->nColumn
;
3673 assert( nColumn
==nKeyCol
+1 || !HasRowid(pIndex
->pTable
) );
3674 assert( pIndex
->aiColumn
[nColumn
-1]==XN_ROWID
3675 || !HasRowid(pIndex
->pTable
));
3676 isOrderDistinct
= IsUniqueIndex(pIndex
);
3679 /* Loop through all columns of the index and deal with the ones
3680 ** that are not constrained by == or IN.
3683 distinctColumns
= 0;
3684 for(j
=0; j
<nColumn
; j
++){
3685 u8 bOnce
= 1; /* True to run the ORDER BY search loop */
3687 assert( j
>=pLoop
->u
.btree
.nEq
3688 || (pLoop
->aLTerm
[j
]==0)==(j
<pLoop
->nSkip
)
3690 if( j
<pLoop
->u
.btree
.nEq
&& j
>=pLoop
->nSkip
){
3691 u16 eOp
= pLoop
->aLTerm
[j
]->eOperator
;
3693 /* Skip over == and IS and ISNULL terms. (Also skip IN terms when
3694 ** doing WHERE_ORDERBY_LIMIT processing).
3696 ** If the current term is a column of an ((?,?) IN (SELECT...))
3697 ** expression for which the SELECT returns more than one column,
3698 ** check that it is the only column used by this loop. Otherwise,
3699 ** if it is one of two or more, none of the columns can be
3700 ** considered to match an ORDER BY term. */
3701 if( (eOp
& eqOpMask
)!=0 ){
3702 if( eOp
& WO_ISNULL
){
3703 testcase( isOrderDistinct
);
3704 isOrderDistinct
= 0;
3707 }else if( ALWAYS(eOp
& WO_IN
) ){
3708 /* ALWAYS() justification: eOp is an equality operator due to the
3709 ** j<pLoop->u.btree.nEq constraint above. Any equality other
3710 ** than WO_IN is captured by the previous "if". So this one
3711 ** always has to be WO_IN. */
3712 Expr
*pX
= pLoop
->aLTerm
[j
]->pExpr
;
3713 for(i
=j
+1; i
<pLoop
->u
.btree
.nEq
; i
++){
3714 if( pLoop
->aLTerm
[i
]->pExpr
==pX
){
3715 assert( (pLoop
->aLTerm
[i
]->eOperator
& WO_IN
) );
3723 /* Get the column number in the table (iColumn) and sort order
3724 ** (revIdx) for the j-th column of the index.
3727 iColumn
= pIndex
->aiColumn
[j
];
3728 revIdx
= pIndex
->aSortOrder
[j
];
3729 if( iColumn
==pIndex
->pTable
->iPKey
) iColumn
= XN_ROWID
;
3735 /* An unconstrained column that might be NULL means that this
3736 ** WhereLoop is not well-ordered
3740 && j
>=pLoop
->u
.btree
.nEq
3741 && pIndex
->pTable
->aCol
[iColumn
].notNull
==0
3743 isOrderDistinct
= 0;
3746 /* Find the ORDER BY term that corresponds to the j-th column
3747 ** of the index and mark that ORDER BY term off
3750 for(i
=0; bOnce
&& i
<nOrderBy
; i
++){
3751 if( MASKBIT(i
) & obSat
) continue;
3752 pOBExpr
= sqlite3ExprSkipCollate(pOrderBy
->a
[i
].pExpr
);
3753 testcase( wctrlFlags
& WHERE_GROUPBY
);
3754 testcase( wctrlFlags
& WHERE_DISTINCTBY
);
3755 if( (wctrlFlags
& (WHERE_GROUPBY
|WHERE_DISTINCTBY
))==0 ) bOnce
= 0;
3756 if( iColumn
>=XN_ROWID
){
3757 if( pOBExpr
->op
!=TK_COLUMN
) continue;
3758 if( pOBExpr
->iTable
!=iCur
) continue;
3759 if( pOBExpr
->iColumn
!=iColumn
) continue;
3761 Expr
*pIdxExpr
= pIndex
->aColExpr
->a
[j
].pExpr
;
3762 if( sqlite3ExprCompareSkip(pOBExpr
, pIdxExpr
, iCur
) ){
3766 if( iColumn
!=XN_ROWID
){
3767 pColl
= sqlite3ExprNNCollSeq(pWInfo
->pParse
, pOrderBy
->a
[i
].pExpr
);
3768 if( sqlite3StrICmp(pColl
->zName
, pIndex
->azColl
[j
])!=0 ) continue;
3770 pLoop
->u
.btree
.nIdxCol
= j
+1;
3774 if( isMatch
&& (wctrlFlags
& WHERE_GROUPBY
)==0 ){
3775 /* Make sure the sort order is compatible in an ORDER BY clause.
3776 ** Sort order is irrelevant for a GROUP BY clause. */
3778 if( (rev
^ revIdx
)!=pOrderBy
->a
[i
].sortOrder
) isMatch
= 0;
3780 rev
= revIdx
^ pOrderBy
->a
[i
].sortOrder
;
3781 if( rev
) *pRevMask
|= MASKBIT(iLoop
);
3786 if( iColumn
==XN_ROWID
){
3787 testcase( distinctColumns
==0 );
3788 distinctColumns
= 1;
3790 obSat
|= MASKBIT(i
);
3792 /* No match found */
3793 if( j
==0 || j
<nKeyCol
){
3794 testcase( isOrderDistinct
!=0 );
3795 isOrderDistinct
= 0;
3799 } /* end Loop over all index columns */
3800 if( distinctColumns
){
3801 testcase( isOrderDistinct
==0 );
3802 isOrderDistinct
= 1;
3804 } /* end-if not one-row */
3806 /* Mark off any other ORDER BY terms that reference pLoop */
3807 if( isOrderDistinct
){
3808 orderDistinctMask
|= pLoop
->maskSelf
;
3809 for(i
=0; i
<nOrderBy
; i
++){
3812 if( MASKBIT(i
) & obSat
) continue;
3813 p
= pOrderBy
->a
[i
].pExpr
;
3814 mTerm
= sqlite3WhereExprUsage(&pWInfo
->sMaskSet
,p
);
3815 if( mTerm
==0 && !sqlite3ExprIsConstant(p
) ) continue;
3816 if( (mTerm
&~orderDistinctMask
)==0 ){
3817 obSat
|= MASKBIT(i
);
3821 } /* End the loop over all WhereLoops from outer-most down to inner-most */
3822 if( obSat
==obDone
) return (i8
)nOrderBy
;
3823 if( !isOrderDistinct
){
3824 for(i
=nOrderBy
-1; i
>0; i
--){
3825 Bitmask m
= MASKBIT(i
) - 1;
3826 if( (obSat
&m
)==m
) return i
;
3835 ** If the WHERE_GROUPBY flag is set in the mask passed to sqlite3WhereBegin(),
3836 ** the planner assumes that the specified pOrderBy list is actually a GROUP
3837 ** BY clause - and so any order that groups rows as required satisfies the
3840 ** Normally, in this case it is not possible for the caller to determine
3841 ** whether or not the rows are really being delivered in sorted order, or
3842 ** just in some other order that provides the required grouping. However,
3843 ** if the WHERE_SORTBYGROUP flag is also passed to sqlite3WhereBegin(), then
3844 ** this function may be called on the returned WhereInfo object. It returns
3845 ** true if the rows really will be sorted in the specified order, or false
3848 ** For example, assuming:
3850 ** CREATE INDEX i1 ON t1(x, Y);
3854 ** SELECT * FROM t1 GROUP BY x,y ORDER BY x,y; -- IsSorted()==1
3855 ** SELECT * FROM t1 GROUP BY y,x ORDER BY y,x; -- IsSorted()==0
3857 int sqlite3WhereIsSorted(WhereInfo
*pWInfo
){
3858 assert( pWInfo
->wctrlFlags
& WHERE_GROUPBY
);
3859 assert( pWInfo
->wctrlFlags
& WHERE_SORTBYGROUP
);
3860 return pWInfo
->sorted
;
3863 #ifdef WHERETRACE_ENABLED
3864 /* For debugging use only: */
3865 static const char *wherePathName(WherePath
*pPath
, int nLoop
, WhereLoop
*pLast
){
3866 static char zName
[65];
3868 for(i
=0; i
<nLoop
; i
++){ zName
[i
] = pPath
->aLoop
[i
]->cId
; }
3869 if( pLast
) zName
[i
++] = pLast
->cId
;
3876 ** Return the cost of sorting nRow rows, assuming that the keys have
3877 ** nOrderby columns and that the first nSorted columns are already in
3880 static LogEst
whereSortingCost(
3886 /* TUNING: Estimated cost of a full external sort, where N is
3887 ** the number of rows to sort is:
3889 ** cost = (3.0 * N * log(N)).
3891 ** Or, if the order-by clause has X terms but only the last Y
3892 ** terms are out of order, then block-sorting will reduce the
3895 ** cost = (3.0 * N * log(N)) * (Y/X)
3897 ** The (Y/X) term is implemented using stack variable rScale
3899 LogEst rScale
, rSortCost
;
3900 assert( nOrderBy
>0 && 66==sqlite3LogEst(100) );
3901 rScale
= sqlite3LogEst((nOrderBy
-nSorted
)*100/nOrderBy
) - 66;
3902 rSortCost
= nRow
+ rScale
+ 16;
3904 /* Multiple by log(M) where M is the number of output rows.
3905 ** Use the LIMIT for M if it is smaller */
3906 if( (pWInfo
->wctrlFlags
& WHERE_USE_LIMIT
)!=0 && pWInfo
->iLimit
<nRow
){
3907 nRow
= pWInfo
->iLimit
;
3909 rSortCost
+= estLog(nRow
);
3914 ** Given the list of WhereLoop objects at pWInfo->pLoops, this routine
3915 ** attempts to find the lowest cost path that visits each WhereLoop
3916 ** once. This path is then loaded into the pWInfo->a[].pWLoop fields.
3918 ** Assume that the total number of output rows that will need to be sorted
3919 ** will be nRowEst (in the 10*log2 representation). Or, ignore sorting
3920 ** costs if nRowEst==0.
3922 ** Return SQLITE_OK on success or SQLITE_NOMEM of a memory allocation
3925 static int wherePathSolver(WhereInfo
*pWInfo
, LogEst nRowEst
){
3926 int mxChoice
; /* Maximum number of simultaneous paths tracked */
3927 int nLoop
; /* Number of terms in the join */
3928 Parse
*pParse
; /* Parsing context */
3929 sqlite3
*db
; /* The database connection */
3930 int iLoop
; /* Loop counter over the terms of the join */
3931 int ii
, jj
; /* Loop counters */
3932 int mxI
= 0; /* Index of next entry to replace */
3933 int nOrderBy
; /* Number of ORDER BY clause terms */
3934 LogEst mxCost
= 0; /* Maximum cost of a set of paths */
3935 LogEst mxUnsorted
= 0; /* Maximum unsorted cost of a set of path */
3936 int nTo
, nFrom
; /* Number of valid entries in aTo[] and aFrom[] */
3937 WherePath
*aFrom
; /* All nFrom paths at the previous level */
3938 WherePath
*aTo
; /* The nTo best paths at the current level */
3939 WherePath
*pFrom
; /* An element of aFrom[] that we are working on */
3940 WherePath
*pTo
; /* An element of aTo[] that we are working on */
3941 WhereLoop
*pWLoop
; /* One of the WhereLoop objects */
3942 WhereLoop
**pX
; /* Used to divy up the pSpace memory */
3943 LogEst
*aSortCost
= 0; /* Sorting and partial sorting costs */
3944 char *pSpace
; /* Temporary memory used by this routine */
3945 int nSpace
; /* Bytes of space allocated at pSpace */
3947 pParse
= pWInfo
->pParse
;
3949 nLoop
= pWInfo
->nLevel
;
3950 /* TUNING: For simple queries, only the best path is tracked.
3951 ** For 2-way joins, the 5 best paths are followed.
3952 ** For joins of 3 or more tables, track the 10 best paths */
3953 mxChoice
= (nLoop
<=1) ? 1 : (nLoop
==2 ? 5 : 10);
3954 assert( nLoop
<=pWInfo
->pTabList
->nSrc
);
3955 WHERETRACE(0x002, ("---- begin solver. (nRowEst=%d)\n", nRowEst
));
3957 /* If nRowEst is zero and there is an ORDER BY clause, ignore it. In this
3958 ** case the purpose of this call is to estimate the number of rows returned
3959 ** by the overall query. Once this estimate has been obtained, the caller
3960 ** will invoke this function a second time, passing the estimate as the
3961 ** nRowEst parameter. */
3962 if( pWInfo
->pOrderBy
==0 || nRowEst
==0 ){
3965 nOrderBy
= pWInfo
->pOrderBy
->nExpr
;
3968 /* Allocate and initialize space for aTo, aFrom and aSortCost[] */
3969 nSpace
= (sizeof(WherePath
)+sizeof(WhereLoop
*)*nLoop
)*mxChoice
*2;
3970 nSpace
+= sizeof(LogEst
) * nOrderBy
;
3971 pSpace
= sqlite3DbMallocRawNN(db
, nSpace
);
3972 if( pSpace
==0 ) return SQLITE_NOMEM_BKPT
;
3973 aTo
= (WherePath
*)pSpace
;
3974 aFrom
= aTo
+mxChoice
;
3975 memset(aFrom
, 0, sizeof(aFrom
[0]));
3976 pX
= (WhereLoop
**)(aFrom
+mxChoice
);
3977 for(ii
=mxChoice
*2, pFrom
=aTo
; ii
>0; ii
--, pFrom
++, pX
+= nLoop
){
3981 /* If there is an ORDER BY clause and it is not being ignored, set up
3982 ** space for the aSortCost[] array. Each element of the aSortCost array
3983 ** is either zero - meaning it has not yet been initialized - or the
3984 ** cost of sorting nRowEst rows of data where the first X terms of
3985 ** the ORDER BY clause are already in order, where X is the array
3987 aSortCost
= (LogEst
*)pX
;
3988 memset(aSortCost
, 0, sizeof(LogEst
) * nOrderBy
);
3990 assert( aSortCost
==0 || &pSpace
[nSpace
]==(char*)&aSortCost
[nOrderBy
] );
3991 assert( aSortCost
!=0 || &pSpace
[nSpace
]==(char*)pX
);
3993 /* Seed the search with a single WherePath containing zero WhereLoops.
3995 ** TUNING: Do not let the number of iterations go above 28. If the cost
3996 ** of computing an automatic index is not paid back within the first 28
3997 ** rows, then do not use the automatic index. */
3998 aFrom
[0].nRow
= MIN(pParse
->nQueryLoop
, 48); assert( 48==sqlite3LogEst(28) );
4000 assert( aFrom
[0].isOrdered
==0 );
4002 /* If nLoop is zero, then there are no FROM terms in the query. Since
4003 ** in this case the query may return a maximum of one row, the results
4004 ** are already in the requested order. Set isOrdered to nOrderBy to
4005 ** indicate this. Or, if nLoop is greater than zero, set isOrdered to
4006 ** -1, indicating that the result set may or may not be ordered,
4007 ** depending on the loops added to the current plan. */
4008 aFrom
[0].isOrdered
= nLoop
>0 ? -1 : nOrderBy
;
4011 /* Compute successively longer WherePaths using the previous generation
4012 ** of WherePaths as the basis for the next. Keep track of the mxChoice
4013 ** best paths at each generation */
4014 for(iLoop
=0; iLoop
<nLoop
; iLoop
++){
4016 for(ii
=0, pFrom
=aFrom
; ii
<nFrom
; ii
++, pFrom
++){
4017 for(pWLoop
=pWInfo
->pLoops
; pWLoop
; pWLoop
=pWLoop
->pNextLoop
){
4018 LogEst nOut
; /* Rows visited by (pFrom+pWLoop) */
4019 LogEst rCost
; /* Cost of path (pFrom+pWLoop) */
4020 LogEst rUnsorted
; /* Unsorted cost of (pFrom+pWLoop) */
4021 i8 isOrdered
= pFrom
->isOrdered
; /* isOrdered for (pFrom+pWLoop) */
4022 Bitmask maskNew
; /* Mask of src visited by (..) */
4023 Bitmask revMask
= 0; /* Mask of rev-order loops for (..) */
4025 if( (pWLoop
->prereq
& ~pFrom
->maskLoop
)!=0 ) continue;
4026 if( (pWLoop
->maskSelf
& pFrom
->maskLoop
)!=0 ) continue;
4027 if( (pWLoop
->wsFlags
& WHERE_AUTO_INDEX
)!=0 && pFrom
->nRow
<3 ){
4028 /* Do not use an automatic index if the this loop is expected
4029 ** to run less than 1.25 times. It is tempting to also exclude
4030 ** automatic index usage on an outer loop, but sometimes an automatic
4031 ** index is useful in the outer loop of a correlated subquery. */
4032 assert( 10==sqlite3LogEst(2) );
4036 /* At this point, pWLoop is a candidate to be the next loop.
4037 ** Compute its cost */
4038 rUnsorted
= sqlite3LogEstAdd(pWLoop
->rSetup
,pWLoop
->rRun
+ pFrom
->nRow
);
4039 rUnsorted
= sqlite3LogEstAdd(rUnsorted
, pFrom
->rUnsorted
);
4040 nOut
= pFrom
->nRow
+ pWLoop
->nOut
;
4041 maskNew
= pFrom
->maskLoop
| pWLoop
->maskSelf
;
4043 isOrdered
= wherePathSatisfiesOrderBy(pWInfo
,
4044 pWInfo
->pOrderBy
, pFrom
, pWInfo
->wctrlFlags
,
4045 iLoop
, pWLoop
, &revMask
);
4047 revMask
= pFrom
->revLoop
;
4049 if( isOrdered
>=0 && isOrdered
<nOrderBy
){
4050 if( aSortCost
[isOrdered
]==0 ){
4051 aSortCost
[isOrdered
] = whereSortingCost(
4052 pWInfo
, nRowEst
, nOrderBy
, isOrdered
4055 rCost
= sqlite3LogEstAdd(rUnsorted
, aSortCost
[isOrdered
]);
4058 ("---- sort cost=%-3d (%d/%d) increases cost %3d to %-3d\n",
4059 aSortCost
[isOrdered
], (nOrderBy
-isOrdered
), nOrderBy
,
4063 rUnsorted
-= 2; /* TUNING: Slight bias in favor of no-sort plans */
4066 /* Check to see if pWLoop should be added to the set of
4067 ** mxChoice best-so-far paths.
4069 ** First look for an existing path among best-so-far paths
4070 ** that covers the same set of loops and has the same isOrdered
4071 ** setting as the current path candidate.
4073 ** The term "((pTo->isOrdered^isOrdered)&0x80)==0" is equivalent
4074 ** to (pTo->isOrdered==(-1))==(isOrdered==(-1))" for the range
4075 ** of legal values for isOrdered, -1..64.
4077 for(jj
=0, pTo
=aTo
; jj
<nTo
; jj
++, pTo
++){
4078 if( pTo
->maskLoop
==maskNew
4079 && ((pTo
->isOrdered
^isOrdered
)&0x80)==0
4081 testcase( jj
==nTo
-1 );
4086 /* None of the existing best-so-far paths match the candidate. */
4088 && (rCost
>mxCost
|| (rCost
==mxCost
&& rUnsorted
>=mxUnsorted
))
4090 /* The current candidate is no better than any of the mxChoice
4091 ** paths currently in the best-so-far buffer. So discard
4092 ** this candidate as not viable. */
4093 #ifdef WHERETRACE_ENABLED /* 0x4 */
4094 if( sqlite3WhereTrace
&0x4 ){
4095 sqlite3DebugPrintf("Skip %s cost=%-3d,%3d,%3d order=%c\n",
4096 wherePathName(pFrom
, iLoop
, pWLoop
), rCost
, nOut
, rUnsorted
,
4097 isOrdered
>=0 ? isOrdered
+'0' : '?');
4102 /* If we reach this points it means that the new candidate path
4103 ** needs to be added to the set of best-so-far paths. */
4105 /* Increase the size of the aTo set by one */
4108 /* New path replaces the prior worst to keep count below mxChoice */
4112 #ifdef WHERETRACE_ENABLED /* 0x4 */
4113 if( sqlite3WhereTrace
&0x4 ){
4114 sqlite3DebugPrintf("New %s cost=%-3d,%3d,%3d order=%c\n",
4115 wherePathName(pFrom
, iLoop
, pWLoop
), rCost
, nOut
, rUnsorted
,
4116 isOrdered
>=0 ? isOrdered
+'0' : '?');
4120 /* Control reaches here if best-so-far path pTo=aTo[jj] covers the
4121 ** same set of loops and has the same isOrdered setting as the
4122 ** candidate path. Check to see if the candidate should replace
4123 ** pTo or if the candidate should be skipped.
4125 ** The conditional is an expanded vector comparison equivalent to:
4126 ** (pTo->rCost,pTo->nRow,pTo->rUnsorted) <= (rCost,nOut,rUnsorted)
4128 if( pTo
->rCost
<rCost
4129 || (pTo
->rCost
==rCost
4131 || (pTo
->nRow
==nOut
&& pTo
->rUnsorted
<=rUnsorted
)
4135 #ifdef WHERETRACE_ENABLED /* 0x4 */
4136 if( sqlite3WhereTrace
&0x4 ){
4138 "Skip %s cost=%-3d,%3d,%3d order=%c",
4139 wherePathName(pFrom
, iLoop
, pWLoop
), rCost
, nOut
, rUnsorted
,
4140 isOrdered
>=0 ? isOrdered
+'0' : '?');
4141 sqlite3DebugPrintf(" vs %s cost=%-3d,%3d,%3d order=%c\n",
4142 wherePathName(pTo
, iLoop
+1, 0), pTo
->rCost
, pTo
->nRow
,
4143 pTo
->rUnsorted
, pTo
->isOrdered
>=0 ? pTo
->isOrdered
+'0' : '?');
4146 /* Discard the candidate path from further consideration */
4147 testcase( pTo
->rCost
==rCost
);
4150 testcase( pTo
->rCost
==rCost
+1 );
4151 /* Control reaches here if the candidate path is better than the
4152 ** pTo path. Replace pTo with the candidate. */
4153 #ifdef WHERETRACE_ENABLED /* 0x4 */
4154 if( sqlite3WhereTrace
&0x4 ){
4156 "Update %s cost=%-3d,%3d,%3d order=%c",
4157 wherePathName(pFrom
, iLoop
, pWLoop
), rCost
, nOut
, rUnsorted
,
4158 isOrdered
>=0 ? isOrdered
+'0' : '?');
4159 sqlite3DebugPrintf(" was %s cost=%-3d,%3d,%3d order=%c\n",
4160 wherePathName(pTo
, iLoop
+1, 0), pTo
->rCost
, pTo
->nRow
,
4161 pTo
->rUnsorted
, pTo
->isOrdered
>=0 ? pTo
->isOrdered
+'0' : '?');
4165 /* pWLoop is a winner. Add it to the set of best so far */
4166 pTo
->maskLoop
= pFrom
->maskLoop
| pWLoop
->maskSelf
;
4167 pTo
->revLoop
= revMask
;
4170 pTo
->rUnsorted
= rUnsorted
;
4171 pTo
->isOrdered
= isOrdered
;
4172 memcpy(pTo
->aLoop
, pFrom
->aLoop
, sizeof(WhereLoop
*)*iLoop
);
4173 pTo
->aLoop
[iLoop
] = pWLoop
;
4174 if( nTo
>=mxChoice
){
4176 mxCost
= aTo
[0].rCost
;
4177 mxUnsorted
= aTo
[0].nRow
;
4178 for(jj
=1, pTo
=&aTo
[1]; jj
<mxChoice
; jj
++, pTo
++){
4179 if( pTo
->rCost
>mxCost
4180 || (pTo
->rCost
==mxCost
&& pTo
->rUnsorted
>mxUnsorted
)
4182 mxCost
= pTo
->rCost
;
4183 mxUnsorted
= pTo
->rUnsorted
;
4191 #ifdef WHERETRACE_ENABLED /* >=2 */
4192 if( sqlite3WhereTrace
& 0x02 ){
4193 sqlite3DebugPrintf("---- after round %d ----\n", iLoop
);
4194 for(ii
=0, pTo
=aTo
; ii
<nTo
; ii
++, pTo
++){
4195 sqlite3DebugPrintf(" %s cost=%-3d nrow=%-3d order=%c",
4196 wherePathName(pTo
, iLoop
+1, 0), pTo
->rCost
, pTo
->nRow
,
4197 pTo
->isOrdered
>=0 ? (pTo
->isOrdered
+'0') : '?');
4198 if( pTo
->isOrdered
>0 ){
4199 sqlite3DebugPrintf(" rev=0x%llx\n", pTo
->revLoop
);
4201 sqlite3DebugPrintf("\n");
4207 /* Swap the roles of aFrom and aTo for the next generation */
4215 sqlite3ErrorMsg(pParse
, "no query solution");
4216 sqlite3DbFreeNN(db
, pSpace
);
4217 return SQLITE_ERROR
;
4220 /* Find the lowest cost path. pFrom will be left pointing to that path */
4222 for(ii
=1; ii
<nFrom
; ii
++){
4223 if( pFrom
->rCost
>aFrom
[ii
].rCost
) pFrom
= &aFrom
[ii
];
4225 assert( pWInfo
->nLevel
==nLoop
);
4226 /* Load the lowest cost path into pWInfo */
4227 for(iLoop
=0; iLoop
<nLoop
; iLoop
++){
4228 WhereLevel
*pLevel
= pWInfo
->a
+ iLoop
;
4229 pLevel
->pWLoop
= pWLoop
= pFrom
->aLoop
[iLoop
];
4230 pLevel
->iFrom
= pWLoop
->iTab
;
4231 pLevel
->iTabCur
= pWInfo
->pTabList
->a
[pLevel
->iFrom
].iCursor
;
4233 if( (pWInfo
->wctrlFlags
& WHERE_WANT_DISTINCT
)!=0
4234 && (pWInfo
->wctrlFlags
& WHERE_DISTINCTBY
)==0
4235 && pWInfo
->eDistinct
==WHERE_DISTINCT_NOOP
4239 int rc
= wherePathSatisfiesOrderBy(pWInfo
, pWInfo
->pResultSet
, pFrom
,
4240 WHERE_DISTINCTBY
, nLoop
-1, pFrom
->aLoop
[nLoop
-1], ¬Used
);
4241 if( rc
==pWInfo
->pResultSet
->nExpr
){
4242 pWInfo
->eDistinct
= WHERE_DISTINCT_ORDERED
;
4245 if( pWInfo
->pOrderBy
){
4246 if( pWInfo
->wctrlFlags
& WHERE_DISTINCTBY
){
4247 if( pFrom
->isOrdered
==pWInfo
->pOrderBy
->nExpr
){
4248 pWInfo
->eDistinct
= WHERE_DISTINCT_ORDERED
;
4251 pWInfo
->nOBSat
= pFrom
->isOrdered
;
4252 pWInfo
->revMask
= pFrom
->revLoop
;
4253 if( pWInfo
->nOBSat
<=0 ){
4256 u32 wsFlags
= pFrom
->aLoop
[nLoop
-1]->wsFlags
;
4257 if( (wsFlags
& WHERE_ONEROW
)==0
4258 && (wsFlags
&(WHERE_IPK
|WHERE_COLUMN_IN
))!=(WHERE_IPK
|WHERE_COLUMN_IN
)
4261 int rc
= wherePathSatisfiesOrderBy(pWInfo
, pWInfo
->pOrderBy
, pFrom
,
4262 WHERE_ORDERBY_LIMIT
, nLoop
-1, pFrom
->aLoop
[nLoop
-1], &m
);
4263 testcase( wsFlags
& WHERE_IPK
);
4264 testcase( wsFlags
& WHERE_COLUMN_IN
);
4265 if( rc
==pWInfo
->pOrderBy
->nExpr
){
4266 pWInfo
->bOrderedInnerLoop
= 1;
4267 pWInfo
->revMask
= m
;
4273 if( (pWInfo
->wctrlFlags
& WHERE_SORTBYGROUP
)
4274 && pWInfo
->nOBSat
==pWInfo
->pOrderBy
->nExpr
&& nLoop
>0
4276 Bitmask revMask
= 0;
4277 int nOrder
= wherePathSatisfiesOrderBy(pWInfo
, pWInfo
->pOrderBy
,
4278 pFrom
, 0, nLoop
-1, pFrom
->aLoop
[nLoop
-1], &revMask
4280 assert( pWInfo
->sorted
==0 );
4281 if( nOrder
==pWInfo
->pOrderBy
->nExpr
){
4283 pWInfo
->revMask
= revMask
;
4289 pWInfo
->nRowOut
= pFrom
->nRow
;
4291 /* Free temporary memory and return success */
4292 sqlite3DbFreeNN(db
, pSpace
);
4297 ** Most queries use only a single table (they are not joins) and have
4298 ** simple == constraints against indexed fields. This routine attempts
4299 ** to plan those simple cases using much less ceremony than the
4300 ** general-purpose query planner, and thereby yield faster sqlite3_prepare()
4301 ** times for the common case.
4303 ** Return non-zero on success, if this query can be handled by this
4304 ** no-frills query planner. Return zero if this query needs the
4305 ** general-purpose query planner.
4307 static int whereShortCut(WhereLoopBuilder
*pBuilder
){
4309 struct SrcList_item
*pItem
;
4318 pWInfo
= pBuilder
->pWInfo
;
4319 if( pWInfo
->wctrlFlags
& WHERE_OR_SUBCLAUSE
) return 0;
4320 assert( pWInfo
->pTabList
->nSrc
>=1 );
4321 pItem
= pWInfo
->pTabList
->a
;
4323 if( IsVirtual(pTab
) ) return 0;
4324 if( pItem
->fg
.isIndexedBy
) return 0;
4325 iCur
= pItem
->iCursor
;
4327 pLoop
= pBuilder
->pNew
;
4330 pTerm
= sqlite3WhereFindTerm(pWC
, iCur
, -1, 0, WO_EQ
|WO_IS
, 0);
4332 testcase( pTerm
->eOperator
& WO_IS
);
4333 pLoop
->wsFlags
= WHERE_COLUMN_EQ
|WHERE_IPK
|WHERE_ONEROW
;
4334 pLoop
->aLTerm
[0] = pTerm
;
4336 pLoop
->u
.btree
.nEq
= 1;
4337 /* TUNING: Cost of a rowid lookup is 10 */
4338 pLoop
->rRun
= 33; /* 33==sqlite3LogEst(10) */
4340 for(pIdx
=pTab
->pIndex
; pIdx
; pIdx
=pIdx
->pNext
){
4342 assert( pLoop
->aLTermSpace
==pLoop
->aLTerm
);
4343 if( !IsUniqueIndex(pIdx
)
4344 || pIdx
->pPartIdxWhere
!=0
4345 || pIdx
->nKeyCol
>ArraySize(pLoop
->aLTermSpace
)
4347 opMask
= pIdx
->uniqNotNull
? (WO_EQ
|WO_IS
) : WO_EQ
;
4348 for(j
=0; j
<pIdx
->nKeyCol
; j
++){
4349 pTerm
= sqlite3WhereFindTerm(pWC
, iCur
, j
, 0, opMask
, pIdx
);
4350 if( pTerm
==0 ) break;
4351 testcase( pTerm
->eOperator
& WO_IS
);
4352 pLoop
->aLTerm
[j
] = pTerm
;
4354 if( j
!=pIdx
->nKeyCol
) continue;
4355 pLoop
->wsFlags
= WHERE_COLUMN_EQ
|WHERE_ONEROW
|WHERE_INDEXED
;
4356 if( pIdx
->isCovering
|| (pItem
->colUsed
& pIdx
->colNotIdxed
)==0 ){
4357 pLoop
->wsFlags
|= WHERE_IDX_ONLY
;
4360 pLoop
->u
.btree
.nEq
= j
;
4361 pLoop
->u
.btree
.pIndex
= pIdx
;
4362 /* TUNING: Cost of a unique index lookup is 15 */
4363 pLoop
->rRun
= 39; /* 39==sqlite3LogEst(15) */
4367 if( pLoop
->wsFlags
){
4368 pLoop
->nOut
= (LogEst
)1;
4369 pWInfo
->a
[0].pWLoop
= pLoop
;
4370 assert( pWInfo
->sMaskSet
.n
==1 && iCur
==pWInfo
->sMaskSet
.ix
[0] );
4371 pLoop
->maskSelf
= 1; /* sqlite3WhereGetMask(&pWInfo->sMaskSet, iCur); */
4372 pWInfo
->a
[0].iTabCur
= iCur
;
4373 pWInfo
->nRowOut
= 1;
4374 if( pWInfo
->pOrderBy
) pWInfo
->nOBSat
= pWInfo
->pOrderBy
->nExpr
;
4375 if( pWInfo
->wctrlFlags
& WHERE_WANT_DISTINCT
){
4376 pWInfo
->eDistinct
= WHERE_DISTINCT_UNIQUE
;
4387 ** Helper function for exprIsDeterministic().
4389 static int exprNodeIsDeterministic(Walker
*pWalker
, Expr
*pExpr
){
4390 if( pExpr
->op
==TK_FUNCTION
&& ExprHasProperty(pExpr
, EP_ConstFunc
)==0 ){
4394 return WRC_Continue
;
4398 ** Return true if the expression contains no non-deterministic SQL
4399 ** functions. Do not consider non-deterministic SQL functions that are
4400 ** part of sub-select statements.
4402 static int exprIsDeterministic(Expr
*p
){
4404 memset(&w
, 0, sizeof(w
));
4406 w
.xExprCallback
= exprNodeIsDeterministic
;
4407 w
.xSelectCallback
= sqlite3SelectWalkFail
;
4408 sqlite3WalkExpr(&w
, p
);
4413 ** Generate the beginning of the loop used for WHERE clause processing.
4414 ** The return value is a pointer to an opaque structure that contains
4415 ** information needed to terminate the loop. Later, the calling routine
4416 ** should invoke sqlite3WhereEnd() with the return value of this function
4417 ** in order to complete the WHERE clause processing.
4419 ** If an error occurs, this routine returns NULL.
4421 ** The basic idea is to do a nested loop, one loop for each table in
4422 ** the FROM clause of a select. (INSERT and UPDATE statements are the
4423 ** same as a SELECT with only a single table in the FROM clause.) For
4424 ** example, if the SQL is this:
4426 ** SELECT * FROM t1, t2, t3 WHERE ...;
4428 ** Then the code generated is conceptually like the following:
4430 ** foreach row1 in t1 do \ Code generated
4431 ** foreach row2 in t2 do |-- by sqlite3WhereBegin()
4432 ** foreach row3 in t3 do /
4434 ** end \ Code generated
4435 ** end |-- by sqlite3WhereEnd()
4438 ** Note that the loops might not be nested in the order in which they
4439 ** appear in the FROM clause if a different order is better able to make
4440 ** use of indices. Note also that when the IN operator appears in
4441 ** the WHERE clause, it might result in additional nested loops for
4442 ** scanning through all values on the right-hand side of the IN.
4444 ** There are Btree cursors associated with each table. t1 uses cursor
4445 ** number pTabList->a[0].iCursor. t2 uses the cursor pTabList->a[1].iCursor.
4446 ** And so forth. This routine generates code to open those VDBE cursors
4447 ** and sqlite3WhereEnd() generates the code to close them.
4449 ** The code that sqlite3WhereBegin() generates leaves the cursors named
4450 ** in pTabList pointing at their appropriate entries. The [...] code
4451 ** can use OP_Column and OP_Rowid opcodes on these cursors to extract
4452 ** data from the various tables of the loop.
4454 ** If the WHERE clause is empty, the foreach loops must each scan their
4455 ** entire tables. Thus a three-way join is an O(N^3) operation. But if
4456 ** the tables have indices and there are terms in the WHERE clause that
4457 ** refer to those indices, a complete table scan can be avoided and the
4458 ** code will run much faster. Most of the work of this routine is checking
4459 ** to see if there are indices that can be used to speed up the loop.
4461 ** Terms of the WHERE clause are also used to limit which rows actually
4462 ** make it to the "..." in the middle of the loop. After each "foreach",
4463 ** terms of the WHERE clause that use only terms in that loop and outer
4464 ** loops are evaluated and if false a jump is made around all subsequent
4465 ** inner loops (or around the "..." if the test occurs within the inner-
4470 ** An outer join of tables t1 and t2 is conceptally coded as follows:
4472 ** foreach row1 in t1 do
4474 ** foreach row2 in t2 do
4480 ** move the row2 cursor to a null row
4485 ** ORDER BY CLAUSE PROCESSING
4487 ** pOrderBy is a pointer to the ORDER BY clause (or the GROUP BY clause
4488 ** if the WHERE_GROUPBY flag is set in wctrlFlags) of a SELECT statement
4489 ** if there is one. If there is no ORDER BY clause or if this routine
4490 ** is called from an UPDATE or DELETE statement, then pOrderBy is NULL.
4492 ** The iIdxCur parameter is the cursor number of an index. If
4493 ** WHERE_OR_SUBCLAUSE is set, iIdxCur is the cursor number of an index
4494 ** to use for OR clause processing. The WHERE clause should use this
4495 ** specific cursor. If WHERE_ONEPASS_DESIRED is set, then iIdxCur is
4496 ** the first cursor in an array of cursors for all indices. iIdxCur should
4497 ** be used to compute the appropriate cursor depending on which index is
4500 WhereInfo
*sqlite3WhereBegin(
4501 Parse
*pParse
, /* The parser context */
4502 SrcList
*pTabList
, /* FROM clause: A list of all tables to be scanned */
4503 Expr
*pWhere
, /* The WHERE clause */
4504 ExprList
*pOrderBy
, /* An ORDER BY (or GROUP BY) clause, or NULL */
4505 ExprList
*pResultSet
, /* Query result set. Req'd for DISTINCT */
4506 u16 wctrlFlags
, /* The WHERE_* flags defined in sqliteInt.h */
4507 int iAuxArg
/* If WHERE_OR_SUBCLAUSE is set, index cursor number
4508 ** If WHERE_USE_LIMIT, then the limit amount */
4510 int nByteWInfo
; /* Num. bytes allocated for WhereInfo struct */
4511 int nTabList
; /* Number of elements in pTabList */
4512 WhereInfo
*pWInfo
; /* Will become the return value of this function */
4513 Vdbe
*v
= pParse
->pVdbe
; /* The virtual database engine */
4514 Bitmask notReady
; /* Cursors that are not yet positioned */
4515 WhereLoopBuilder sWLB
; /* The WhereLoop builder */
4516 WhereMaskSet
*pMaskSet
; /* The expression mask set */
4517 WhereLevel
*pLevel
; /* A single level in pWInfo->a[] */
4518 WhereLoop
*pLoop
; /* Pointer to a single WhereLoop object */
4519 int ii
; /* Loop counter */
4520 sqlite3
*db
; /* Database connection */
4521 int rc
; /* Return code */
4522 u8 bFordelete
= 0; /* OPFLAG_FORDELETE or zero, as appropriate */
4524 assert( (wctrlFlags
& WHERE_ONEPASS_MULTIROW
)==0 || (
4525 (wctrlFlags
& WHERE_ONEPASS_DESIRED
)!=0
4526 && (wctrlFlags
& WHERE_OR_SUBCLAUSE
)==0
4529 /* Only one of WHERE_OR_SUBCLAUSE or WHERE_USE_LIMIT */
4530 assert( (wctrlFlags
& WHERE_OR_SUBCLAUSE
)==0
4531 || (wctrlFlags
& WHERE_USE_LIMIT
)==0 );
4533 /* Variable initialization */
4535 memset(&sWLB
, 0, sizeof(sWLB
));
4537 /* An ORDER/GROUP BY clause of more than 63 terms cannot be optimized */
4538 testcase( pOrderBy
&& pOrderBy
->nExpr
==BMS
-1 );
4539 if( pOrderBy
&& pOrderBy
->nExpr
>=BMS
) pOrderBy
= 0;
4540 sWLB
.pOrderBy
= pOrderBy
;
4542 /* Disable the DISTINCT optimization if SQLITE_DistinctOpt is set via
4543 ** sqlite3_test_ctrl(SQLITE_TESTCTRL_OPTIMIZATIONS,...) */
4544 if( OptimizationDisabled(db
, SQLITE_DistinctOpt
) ){
4545 wctrlFlags
&= ~WHERE_WANT_DISTINCT
;
4548 /* The number of tables in the FROM clause is limited by the number of
4549 ** bits in a Bitmask
4551 testcase( pTabList
->nSrc
==BMS
);
4552 if( pTabList
->nSrc
>BMS
){
4553 sqlite3ErrorMsg(pParse
, "at most %d tables in a join", BMS
);
4557 /* This function normally generates a nested loop for all tables in
4558 ** pTabList. But if the WHERE_OR_SUBCLAUSE flag is set, then we should
4559 ** only generate code for the first table in pTabList and assume that
4560 ** any cursors associated with subsequent tables are uninitialized.
4562 nTabList
= (wctrlFlags
& WHERE_OR_SUBCLAUSE
) ? 1 : pTabList
->nSrc
;
4564 /* Allocate and initialize the WhereInfo structure that will become the
4565 ** return value. A single allocation is used to store the WhereInfo
4566 ** struct, the contents of WhereInfo.a[], the WhereClause structure
4567 ** and the WhereMaskSet structure. Since WhereClause contains an 8-byte
4568 ** field (type Bitmask) it must be aligned on an 8-byte boundary on
4569 ** some architectures. Hence the ROUND8() below.
4571 nByteWInfo
= ROUND8(sizeof(WhereInfo
)+(nTabList
-1)*sizeof(WhereLevel
));
4572 pWInfo
= sqlite3DbMallocRawNN(db
, nByteWInfo
+ sizeof(WhereLoop
));
4573 if( db
->mallocFailed
){
4574 sqlite3DbFree(db
, pWInfo
);
4576 goto whereBeginError
;
4578 pWInfo
->pParse
= pParse
;
4579 pWInfo
->pTabList
= pTabList
;
4580 pWInfo
->pOrderBy
= pOrderBy
;
4581 pWInfo
->pWhere
= pWhere
;
4582 pWInfo
->pResultSet
= pResultSet
;
4583 pWInfo
->aiCurOnePass
[0] = pWInfo
->aiCurOnePass
[1] = -1;
4584 pWInfo
->nLevel
= nTabList
;
4585 pWInfo
->iBreak
= pWInfo
->iContinue
= sqlite3VdbeMakeLabel(v
);
4586 pWInfo
->wctrlFlags
= wctrlFlags
;
4587 pWInfo
->iLimit
= iAuxArg
;
4588 pWInfo
->savedNQueryLoop
= pParse
->nQueryLoop
;
4589 memset(&pWInfo
->nOBSat
, 0,
4590 offsetof(WhereInfo
,sWC
) - offsetof(WhereInfo
,nOBSat
));
4591 memset(&pWInfo
->a
[0], 0, sizeof(WhereLoop
)+nTabList
*sizeof(WhereLevel
));
4592 assert( pWInfo
->eOnePass
==ONEPASS_OFF
); /* ONEPASS defaults to OFF */
4593 pMaskSet
= &pWInfo
->sMaskSet
;
4594 sWLB
.pWInfo
= pWInfo
;
4595 sWLB
.pWC
= &pWInfo
->sWC
;
4596 sWLB
.pNew
= (WhereLoop
*)(((char*)pWInfo
)+nByteWInfo
);
4597 assert( EIGHT_BYTE_ALIGNMENT(sWLB
.pNew
) );
4598 whereLoopInit(sWLB
.pNew
);
4600 sWLB
.pNew
->cId
= '*';
4603 /* Split the WHERE clause into separate subexpressions where each
4604 ** subexpression is separated by an AND operator.
4606 initMaskSet(pMaskSet
);
4607 sqlite3WhereClauseInit(&pWInfo
->sWC
, pWInfo
);
4608 sqlite3WhereSplit(&pWInfo
->sWC
, pWhere
, TK_AND
);
4610 /* Special case: No FROM clause
4613 if( pOrderBy
) pWInfo
->nOBSat
= pOrderBy
->nExpr
;
4614 if( wctrlFlags
& WHERE_WANT_DISTINCT
){
4615 pWInfo
->eDistinct
= WHERE_DISTINCT_UNIQUE
;
4617 ExplainQueryPlan((pParse
, 0, "SCAN CONSTANT ROW"));
4619 /* Assign a bit from the bitmask to every term in the FROM clause.
4621 ** The N-th term of the FROM clause is assigned a bitmask of 1<<N.
4623 ** The rule of the previous sentence ensures thta if X is the bitmask for
4624 ** a table T, then X-1 is the bitmask for all other tables to the left of T.
4625 ** Knowing the bitmask for all tables to the left of a left join is
4626 ** important. Ticket #3015.
4628 ** Note that bitmasks are created for all pTabList->nSrc tables in
4629 ** pTabList, not just the first nTabList tables. nTabList is normally
4630 ** equal to pTabList->nSrc but might be shortened to 1 if the
4631 ** WHERE_OR_SUBCLAUSE flag is set.
4635 createMask(pMaskSet
, pTabList
->a
[ii
].iCursor
);
4636 sqlite3WhereTabFuncArgs(pParse
, &pTabList
->a
[ii
], &pWInfo
->sWC
);
4637 }while( (++ii
)<pTabList
->nSrc
);
4641 for(ii
=0; ii
<pTabList
->nSrc
; ii
++){
4642 Bitmask m
= sqlite3WhereGetMask(pMaskSet
, pTabList
->a
[ii
].iCursor
);
4650 /* Analyze all of the subexpressions. */
4651 sqlite3WhereExprAnalyze(pTabList
, &pWInfo
->sWC
);
4652 if( db
->mallocFailed
) goto whereBeginError
;
4654 /* Special case: WHERE terms that do not refer to any tables in the join
4655 ** (constant expressions). Evaluate each such term, and jump over all the
4656 ** generated code if the result is not true.
4658 ** Do not do this if the expression contains non-deterministic functions
4659 ** that are not within a sub-select. This is not strictly required, but
4660 ** preserves SQLite's legacy behaviour in the following two cases:
4662 ** FROM ... WHERE random()>0; -- eval random() once per row
4663 ** FROM ... WHERE (SELECT random())>0; -- eval random() once overall
4665 for(ii
=0; ii
<sWLB
.pWC
->nTerm
; ii
++){
4666 WhereTerm
*pT
= &sWLB
.pWC
->a
[ii
];
4667 if( pT
->wtFlags
& TERM_VIRTUAL
) continue;
4668 if( pT
->prereqAll
==0 && (nTabList
==0 || exprIsDeterministic(pT
->pExpr
)) ){
4669 sqlite3ExprIfFalse(pParse
, pT
->pExpr
, pWInfo
->iBreak
, SQLITE_JUMPIFNULL
);
4670 pT
->wtFlags
|= TERM_CODED
;
4674 if( wctrlFlags
& WHERE_WANT_DISTINCT
){
4675 if( isDistinctRedundant(pParse
, pTabList
, &pWInfo
->sWC
, pResultSet
) ){
4676 /* The DISTINCT marking is pointless. Ignore it. */
4677 pWInfo
->eDistinct
= WHERE_DISTINCT_UNIQUE
;
4678 }else if( pOrderBy
==0 ){
4679 /* Try to ORDER BY the result set to make distinct processing easier */
4680 pWInfo
->wctrlFlags
|= WHERE_DISTINCTBY
;
4681 pWInfo
->pOrderBy
= pResultSet
;
4685 /* Construct the WhereLoop objects */
4686 #if defined(WHERETRACE_ENABLED)
4687 if( sqlite3WhereTrace
& 0xffff ){
4688 sqlite3DebugPrintf("*** Optimizer Start *** (wctrlFlags: 0x%x",wctrlFlags
);
4689 if( wctrlFlags
& WHERE_USE_LIMIT
){
4690 sqlite3DebugPrintf(", limit: %d", iAuxArg
);
4692 sqlite3DebugPrintf(")\n");
4694 if( sqlite3WhereTrace
& 0x100 ){ /* Display all terms of the WHERE clause */
4695 sqlite3WhereClausePrint(sWLB
.pWC
);
4699 if( nTabList
!=1 || whereShortCut(&sWLB
)==0 ){
4700 rc
= whereLoopAddAll(&sWLB
);
4701 if( rc
) goto whereBeginError
;
4703 #ifdef WHERETRACE_ENABLED
4704 if( sqlite3WhereTrace
){ /* Display all of the WhereLoop objects */
4707 static const char zLabel
[] = "0123456789abcdefghijklmnopqrstuvwyxz"
4708 "ABCDEFGHIJKLMNOPQRSTUVWYXZ";
4709 for(p
=pWInfo
->pLoops
, i
=0; p
; p
=p
->pNextLoop
, i
++){
4710 p
->cId
= zLabel
[i
%(sizeof(zLabel
)-1)];
4711 whereLoopPrint(p
, sWLB
.pWC
);
4716 wherePathSolver(pWInfo
, 0);
4717 if( db
->mallocFailed
) goto whereBeginError
;
4718 if( pWInfo
->pOrderBy
){
4719 wherePathSolver(pWInfo
, pWInfo
->nRowOut
+1);
4720 if( db
->mallocFailed
) goto whereBeginError
;
4723 if( pWInfo
->pOrderBy
==0 && (db
->flags
& SQLITE_ReverseOrder
)!=0 ){
4724 pWInfo
->revMask
= ALLBITS
;
4726 if( pParse
->nErr
|| NEVER(db
->mallocFailed
) ){
4727 goto whereBeginError
;
4729 #ifdef WHERETRACE_ENABLED
4730 if( sqlite3WhereTrace
){
4731 sqlite3DebugPrintf("---- Solution nRow=%d", pWInfo
->nRowOut
);
4732 if( pWInfo
->nOBSat
>0 ){
4733 sqlite3DebugPrintf(" ORDERBY=%d,0x%llx", pWInfo
->nOBSat
, pWInfo
->revMask
);
4735 switch( pWInfo
->eDistinct
){
4736 case WHERE_DISTINCT_UNIQUE
: {
4737 sqlite3DebugPrintf(" DISTINCT=unique");
4740 case WHERE_DISTINCT_ORDERED
: {
4741 sqlite3DebugPrintf(" DISTINCT=ordered");
4744 case WHERE_DISTINCT_UNORDERED
: {
4745 sqlite3DebugPrintf(" DISTINCT=unordered");
4749 sqlite3DebugPrintf("\n");
4750 for(ii
=0; ii
<pWInfo
->nLevel
; ii
++){
4751 whereLoopPrint(pWInfo
->a
[ii
].pWLoop
, sWLB
.pWC
);
4756 /* Attempt to omit tables from the join that do not affect the result.
4757 ** For a table to not affect the result, the following must be true:
4759 ** 1) The query must not be an aggregate.
4760 ** 2) The table must be the RHS of a LEFT JOIN.
4761 ** 3) Either the query must be DISTINCT, or else the ON or USING clause
4762 ** must contain a constraint that limits the scan of the table to
4763 ** at most a single row.
4764 ** 4) The table must not be referenced by any part of the query apart
4765 ** from its own USING or ON clause.
4767 ** For example, given:
4769 ** CREATE TABLE t1(ipk INTEGER PRIMARY KEY, v1);
4770 ** CREATE TABLE t2(ipk INTEGER PRIMARY KEY, v2);
4771 ** CREATE TABLE t3(ipk INTEGER PRIMARY KEY, v3);
4773 ** then table t2 can be omitted from the following:
4775 ** SELECT v1, v3 FROM t1
4776 ** LEFT JOIN t2 USING (t1.ipk=t2.ipk)
4777 ** LEFT JOIN t3 USING (t1.ipk=t3.ipk)
4781 ** SELECT DISTINCT v1, v3 FROM t1
4783 ** LEFT JOIN t3 USING (t1.ipk=t3.ipk)
4785 notReady
= ~(Bitmask
)0;
4786 if( pWInfo
->nLevel
>=2
4787 && pResultSet
!=0 /* guarantees condition (1) above */
4788 && OptimizationEnabled(db
, SQLITE_OmitNoopJoin
)
4791 Bitmask tabUsed
= sqlite3WhereExprListUsage(pMaskSet
, pResultSet
);
4792 if( sWLB
.pOrderBy
){
4793 tabUsed
|= sqlite3WhereExprListUsage(pMaskSet
, sWLB
.pOrderBy
);
4795 for(i
=pWInfo
->nLevel
-1; i
>=1; i
--){
4796 WhereTerm
*pTerm
, *pEnd
;
4797 struct SrcList_item
*pItem
;
4798 pLoop
= pWInfo
->a
[i
].pWLoop
;
4799 pItem
= &pWInfo
->pTabList
->a
[pLoop
->iTab
];
4800 if( (pItem
->fg
.jointype
& JT_LEFT
)==0 ) continue;
4801 if( (wctrlFlags
& WHERE_WANT_DISTINCT
)==0
4802 && (pLoop
->wsFlags
& WHERE_ONEROW
)==0
4806 if( (tabUsed
& pLoop
->maskSelf
)!=0 ) continue;
4807 pEnd
= sWLB
.pWC
->a
+ sWLB
.pWC
->nTerm
;
4808 for(pTerm
=sWLB
.pWC
->a
; pTerm
<pEnd
; pTerm
++){
4809 if( (pTerm
->prereqAll
& pLoop
->maskSelf
)!=0 ){
4810 if( !ExprHasProperty(pTerm
->pExpr
, EP_FromJoin
)
4811 || pTerm
->pExpr
->iRightJoinTable
!=pItem
->iCursor
4817 if( pTerm
<pEnd
) continue;
4818 WHERETRACE(0xffff, ("-> drop loop %c not used\n", pLoop
->cId
));
4819 notReady
&= ~pLoop
->maskSelf
;
4820 for(pTerm
=sWLB
.pWC
->a
; pTerm
<pEnd
; pTerm
++){
4821 if( (pTerm
->prereqAll
& pLoop
->maskSelf
)!=0 ){
4822 pTerm
->wtFlags
|= TERM_CODED
;
4825 if( i
!=pWInfo
->nLevel
-1 ){
4826 int nByte
= (pWInfo
->nLevel
-1-i
) * sizeof(WhereLevel
);
4827 memmove(&pWInfo
->a
[i
], &pWInfo
->a
[i
+1], nByte
);
4833 WHERETRACE(0xffff,("*** Optimizer Finished ***\n"));
4834 pWInfo
->pParse
->nQueryLoop
+= pWInfo
->nRowOut
;
4836 /* If the caller is an UPDATE or DELETE statement that is requesting
4837 ** to use a one-pass algorithm, determine if this is appropriate.
4839 ** A one-pass approach can be used if the caller has requested one
4840 ** and either (a) the scan visits at most one row or (b) each
4841 ** of the following are true:
4843 ** * the caller has indicated that a one-pass approach can be used
4844 ** with multiple rows (by setting WHERE_ONEPASS_MULTIROW), and
4845 ** * the table is not a virtual table, and
4846 ** * either the scan does not use the OR optimization or the caller
4847 ** is a DELETE operation (WHERE_DUPLICATES_OK is only specified
4850 ** The last qualification is because an UPDATE statement uses
4851 ** WhereInfo.aiCurOnePass[1] to determine whether or not it really can
4852 ** use a one-pass approach, and this is not set accurately for scans
4853 ** that use the OR optimization.
4855 assert( (wctrlFlags
& WHERE_ONEPASS_DESIRED
)==0 || pWInfo
->nLevel
==1 );
4856 if( (wctrlFlags
& WHERE_ONEPASS_DESIRED
)!=0 ){
4857 int wsFlags
= pWInfo
->a
[0].pWLoop
->wsFlags
;
4858 int bOnerow
= (wsFlags
& WHERE_ONEROW
)!=0;
4860 0!=(wctrlFlags
& WHERE_ONEPASS_MULTIROW
)
4861 && 0==(wsFlags
& WHERE_VIRTUALTABLE
)
4862 && (0==(wsFlags
& WHERE_MULTI_OR
) || (wctrlFlags
& WHERE_DUPLICATES_OK
))
4864 pWInfo
->eOnePass
= bOnerow
? ONEPASS_SINGLE
: ONEPASS_MULTI
;
4865 if( HasRowid(pTabList
->a
[0].pTab
) && (wsFlags
& WHERE_IDX_ONLY
) ){
4866 if( wctrlFlags
& WHERE_ONEPASS_MULTIROW
){
4867 bFordelete
= OPFLAG_FORDELETE
;
4869 pWInfo
->a
[0].pWLoop
->wsFlags
= (wsFlags
& ~WHERE_IDX_ONLY
);
4874 /* Open all tables in the pTabList and any indices selected for
4875 ** searching those tables.
4877 for(ii
=0, pLevel
=pWInfo
->a
; ii
<nTabList
; ii
++, pLevel
++){
4878 Table
*pTab
; /* Table to open */
4879 int iDb
; /* Index of database containing table/index */
4880 struct SrcList_item
*pTabItem
;
4882 pTabItem
= &pTabList
->a
[pLevel
->iFrom
];
4883 pTab
= pTabItem
->pTab
;
4884 iDb
= sqlite3SchemaToIndex(db
, pTab
->pSchema
);
4885 pLoop
= pLevel
->pWLoop
;
4886 if( (pTab
->tabFlags
& TF_Ephemeral
)!=0 || pTab
->pSelect
){
4889 #ifndef SQLITE_OMIT_VIRTUALTABLE
4890 if( (pLoop
->wsFlags
& WHERE_VIRTUALTABLE
)!=0 ){
4891 const char *pVTab
= (const char *)sqlite3GetVTable(db
, pTab
);
4892 int iCur
= pTabItem
->iCursor
;
4893 sqlite3VdbeAddOp4(v
, OP_VOpen
, iCur
, 0, 0, pVTab
, P4_VTAB
);
4894 }else if( IsVirtual(pTab
) ){
4898 if( (pLoop
->wsFlags
& WHERE_IDX_ONLY
)==0
4899 && (wctrlFlags
& WHERE_OR_SUBCLAUSE
)==0 ){
4900 int op
= OP_OpenRead
;
4901 if( pWInfo
->eOnePass
!=ONEPASS_OFF
){
4903 pWInfo
->aiCurOnePass
[0] = pTabItem
->iCursor
;
4905 sqlite3OpenTable(pParse
, pTabItem
->iCursor
, iDb
, pTab
, op
);
4906 assert( pTabItem
->iCursor
==pLevel
->iTabCur
);
4907 testcase( pWInfo
->eOnePass
==ONEPASS_OFF
&& pTab
->nCol
==BMS
-1 );
4908 testcase( pWInfo
->eOnePass
==ONEPASS_OFF
&& pTab
->nCol
==BMS
);
4909 if( pWInfo
->eOnePass
==ONEPASS_OFF
&& pTab
->nCol
<BMS
&& HasRowid(pTab
) ){
4910 Bitmask b
= pTabItem
->colUsed
;
4912 for(; b
; b
=b
>>1, n
++){}
4913 sqlite3VdbeChangeP4(v
, -1, SQLITE_INT_TO_PTR(n
), P4_INT32
);
4914 assert( n
<=pTab
->nCol
);
4916 #ifdef SQLITE_ENABLE_CURSOR_HINTS
4917 if( pLoop
->u
.btree
.pIndex
!=0 ){
4918 sqlite3VdbeChangeP5(v
, OPFLAG_SEEKEQ
|bFordelete
);
4922 sqlite3VdbeChangeP5(v
, bFordelete
);
4924 #ifdef SQLITE_ENABLE_COLUMN_USED_MASK
4925 sqlite3VdbeAddOp4Dup8(v
, OP_ColumnsUsed
, pTabItem
->iCursor
, 0, 0,
4926 (const u8
*)&pTabItem
->colUsed
, P4_INT64
);
4929 sqlite3TableLock(pParse
, iDb
, pTab
->tnum
, 0, pTab
->zName
);
4931 if( pLoop
->wsFlags
& WHERE_INDEXED
){
4932 Index
*pIx
= pLoop
->u
.btree
.pIndex
;
4934 int op
= OP_OpenRead
;
4935 /* iAuxArg is always set to a positive value if ONEPASS is possible */
4936 assert( iAuxArg
!=0 || (pWInfo
->wctrlFlags
& WHERE_ONEPASS_DESIRED
)==0 );
4937 if( !HasRowid(pTab
) && IsPrimaryKeyIndex(pIx
)
4938 && (wctrlFlags
& WHERE_OR_SUBCLAUSE
)!=0
4940 /* This is one term of an OR-optimization using the PRIMARY KEY of a
4941 ** WITHOUT ROWID table. No need for a separate index */
4942 iIndexCur
= pLevel
->iTabCur
;
4944 }else if( pWInfo
->eOnePass
!=ONEPASS_OFF
){
4945 Index
*pJ
= pTabItem
->pTab
->pIndex
;
4946 iIndexCur
= iAuxArg
;
4947 assert( wctrlFlags
& WHERE_ONEPASS_DESIRED
);
4948 while( ALWAYS(pJ
) && pJ
!=pIx
){
4953 pWInfo
->aiCurOnePass
[1] = iIndexCur
;
4954 }else if( iAuxArg
&& (wctrlFlags
& WHERE_OR_SUBCLAUSE
)!=0 ){
4955 iIndexCur
= iAuxArg
;
4958 iIndexCur
= pParse
->nTab
++;
4960 pLevel
->iIdxCur
= iIndexCur
;
4961 assert( pIx
->pSchema
==pTab
->pSchema
);
4962 assert( iIndexCur
>=0 );
4964 sqlite3VdbeAddOp3(v
, op
, iIndexCur
, pIx
->tnum
, iDb
);
4965 sqlite3VdbeSetP4KeyInfo(pParse
, pIx
);
4966 if( (pLoop
->wsFlags
& WHERE_CONSTRAINT
)!=0
4967 && (pLoop
->wsFlags
& (WHERE_COLUMN_RANGE
|WHERE_SKIPSCAN
))==0
4968 && (pWInfo
->wctrlFlags
&WHERE_ORDERBY_MIN
)==0
4969 && pWInfo
->eDistinct
!=WHERE_DISTINCT_ORDERED
4971 sqlite3VdbeChangeP5(v
, OPFLAG_SEEKEQ
); /* Hint to COMDB2 */
4973 VdbeComment((v
, "%s", pIx
->zName
));
4974 #ifdef SQLITE_ENABLE_COLUMN_USED_MASK
4978 for(ii
=0; ii
<pIx
->nColumn
; ii
++){
4979 jj
= pIx
->aiColumn
[ii
];
4980 if( jj
<0 ) continue;
4981 if( jj
>63 ) jj
= 63;
4982 if( (pTabItem
->colUsed
& MASKBIT(jj
))==0 ) continue;
4983 colUsed
|= ((u64
)1)<<(ii
<63 ? ii
: 63);
4985 sqlite3VdbeAddOp4Dup8(v
, OP_ColumnsUsed
, iIndexCur
, 0, 0,
4986 (u8
*)&colUsed
, P4_INT64
);
4988 #endif /* SQLITE_ENABLE_COLUMN_USED_MASK */
4991 if( iDb
>=0 ) sqlite3CodeVerifySchema(pParse
, iDb
);
4993 pWInfo
->iTop
= sqlite3VdbeCurrentAddr(v
);
4994 if( db
->mallocFailed
) goto whereBeginError
;
4996 /* Generate the code to do the search. Each iteration of the for
4997 ** loop below generates code for a single nested loop of the VM
5000 for(ii
=0; ii
<nTabList
; ii
++){
5003 pLevel
= &pWInfo
->a
[ii
];
5004 wsFlags
= pLevel
->pWLoop
->wsFlags
;
5005 #ifndef SQLITE_OMIT_AUTOMATIC_INDEX
5006 if( (pLevel
->pWLoop
->wsFlags
& WHERE_AUTO_INDEX
)!=0 ){
5007 constructAutomaticIndex(pParse
, &pWInfo
->sWC
,
5008 &pTabList
->a
[pLevel
->iFrom
], notReady
, pLevel
);
5009 if( db
->mallocFailed
) goto whereBeginError
;
5012 addrExplain
= sqlite3WhereExplainOneScan(
5013 pParse
, pTabList
, pLevel
, wctrlFlags
5015 pLevel
->addrBody
= sqlite3VdbeCurrentAddr(v
);
5016 notReady
= sqlite3WhereCodeOneLoopStart(pWInfo
, ii
, notReady
);
5017 pWInfo
->iContinue
= pLevel
->addrCont
;
5018 if( (wsFlags
&WHERE_MULTI_OR
)==0 && (wctrlFlags
&WHERE_OR_SUBCLAUSE
)==0 ){
5019 sqlite3WhereAddScanStatus(v
, pTabList
, pLevel
, addrExplain
);
5024 VdbeModuleComment((v
, "Begin WHERE-core"));
5027 /* Jump here if malloc fails */
5030 pParse
->nQueryLoop
= pWInfo
->savedNQueryLoop
;
5031 whereInfoFree(db
, pWInfo
);
5037 ** Part of sqlite3WhereEnd() will rewrite opcodes to reference the
5038 ** index rather than the main table. In SQLITE_DEBUG mode, we want
5039 ** to trace those changes if PRAGMA vdbe_addoptrace=on. This routine
5042 #ifndef SQLITE_DEBUG
5043 # define OpcodeRewriteTrace(D,K,P) /* no-op */
5045 # define OpcodeRewriteTrace(D,K,P) sqlite3WhereOpcodeRewriteTrace(D,K,P)
5046 static void sqlite3WhereOpcodeRewriteTrace(
5051 if( (db
->flags
& SQLITE_VdbeAddopTrace
)==0 ) return;
5052 sqlite3VdbePrintOp(0, pc
, pOp
);
5057 ** Generate the end of the WHERE loop. See comments on
5058 ** sqlite3WhereBegin() for additional information.
5060 void sqlite3WhereEnd(WhereInfo
*pWInfo
){
5061 Parse
*pParse
= pWInfo
->pParse
;
5062 Vdbe
*v
= pParse
->pVdbe
;
5066 SrcList
*pTabList
= pWInfo
->pTabList
;
5067 sqlite3
*db
= pParse
->db
;
5069 /* Generate loop termination code.
5071 VdbeModuleComment((v
, "End WHERE-core"));
5072 sqlite3ExprCacheClear(pParse
);
5073 for(i
=pWInfo
->nLevel
-1; i
>=0; i
--){
5075 pLevel
= &pWInfo
->a
[i
];
5076 pLoop
= pLevel
->pWLoop
;
5077 if( pLevel
->op
!=OP_Noop
){
5078 #ifndef SQLITE_DISABLE_SKIPAHEAD_DISTINCT
5082 if( pWInfo
->eDistinct
==WHERE_DISTINCT_ORDERED
5083 && i
==pWInfo
->nLevel
-1 /* Ticket [ef9318757b152e3] 2017-10-21 */
5084 && (pLoop
->wsFlags
& WHERE_INDEXED
)!=0
5085 && (pIdx
= pLoop
->u
.btree
.pIndex
)->hasStat1
5086 && (n
= pLoop
->u
.btree
.nIdxCol
)>0
5087 && pIdx
->aiRowLogEst
[n
]>=36
5089 int r1
= pParse
->nMem
+1;
5092 sqlite3VdbeAddOp3(v
, OP_Column
, pLevel
->iIdxCur
, j
, r1
+j
);
5094 pParse
->nMem
+= n
+1;
5095 op
= pLevel
->op
==OP_Prev
? OP_SeekLT
: OP_SeekGT
;
5096 addrSeek
= sqlite3VdbeAddOp4Int(v
, op
, pLevel
->iIdxCur
, 0, r1
, n
);
5097 VdbeCoverageIf(v
, op
==OP_SeekLT
);
5098 VdbeCoverageIf(v
, op
==OP_SeekGT
);
5099 sqlite3VdbeAddOp2(v
, OP_Goto
, 1, pLevel
->p2
);
5101 #endif /* SQLITE_DISABLE_SKIPAHEAD_DISTINCT */
5102 /* The common case: Advance to the next row */
5103 sqlite3VdbeResolveLabel(v
, pLevel
->addrCont
);
5104 sqlite3VdbeAddOp3(v
, pLevel
->op
, pLevel
->p1
, pLevel
->p2
, pLevel
->p3
);
5105 sqlite3VdbeChangeP5(v
, pLevel
->p5
);
5107 VdbeCoverageIf(v
, pLevel
->op
==OP_Next
);
5108 VdbeCoverageIf(v
, pLevel
->op
==OP_Prev
);
5109 VdbeCoverageIf(v
, pLevel
->op
==OP_VNext
);
5110 #ifndef SQLITE_DISABLE_SKIPAHEAD_DISTINCT
5111 if( addrSeek
) sqlite3VdbeJumpHere(v
, addrSeek
);
5114 sqlite3VdbeResolveLabel(v
, pLevel
->addrCont
);
5116 if( pLoop
->wsFlags
& WHERE_IN_ABLE
&& pLevel
->u
.in
.nIn
>0 ){
5119 sqlite3VdbeResolveLabel(v
, pLevel
->addrNxt
);
5120 for(j
=pLevel
->u
.in
.nIn
, pIn
=&pLevel
->u
.in
.aInLoop
[j
-1]; j
>0; j
--, pIn
--){
5121 sqlite3VdbeJumpHere(v
, pIn
->addrInTop
+1);
5122 if( pIn
->eEndLoopOp
!=OP_Noop
){
5124 assert( pLoop
->wsFlags
& WHERE_IN_EARLYOUT
);
5125 sqlite3VdbeAddOp4Int(v
, OP_IfNoHope
, pLevel
->iIdxCur
,
5126 sqlite3VdbeCurrentAddr(v
)+2,
5127 pIn
->iBase
, pIn
->nPrefix
);
5130 sqlite3VdbeAddOp2(v
, pIn
->eEndLoopOp
, pIn
->iCur
, pIn
->addrInTop
);
5132 VdbeCoverageIf(v
, pIn
->eEndLoopOp
==OP_Prev
);
5133 VdbeCoverageIf(v
, pIn
->eEndLoopOp
==OP_Next
);
5135 sqlite3VdbeJumpHere(v
, pIn
->addrInTop
-1);
5138 sqlite3VdbeResolveLabel(v
, pLevel
->addrBrk
);
5139 if( pLevel
->addrSkip
){
5140 sqlite3VdbeGoto(v
, pLevel
->addrSkip
);
5141 VdbeComment((v
, "next skip-scan on %s", pLoop
->u
.btree
.pIndex
->zName
));
5142 sqlite3VdbeJumpHere(v
, pLevel
->addrSkip
);
5143 sqlite3VdbeJumpHere(v
, pLevel
->addrSkip
-2);
5145 #ifndef SQLITE_LIKE_DOESNT_MATCH_BLOBS
5146 if( pLevel
->addrLikeRep
){
5147 sqlite3VdbeAddOp2(v
, OP_DecrJumpZero
, (int)(pLevel
->iLikeRepCntr
>>1),
5148 pLevel
->addrLikeRep
);
5152 if( pLevel
->iLeftJoin
){
5153 int ws
= pLoop
->wsFlags
;
5154 addr
= sqlite3VdbeAddOp1(v
, OP_IfPos
, pLevel
->iLeftJoin
); VdbeCoverage(v
);
5155 assert( (ws
& WHERE_IDX_ONLY
)==0 || (ws
& WHERE_INDEXED
)!=0 );
5156 if( (ws
& WHERE_IDX_ONLY
)==0 ){
5157 assert( pLevel
->iTabCur
==pTabList
->a
[pLevel
->iFrom
].iCursor
);
5158 sqlite3VdbeAddOp1(v
, OP_NullRow
, pLevel
->iTabCur
);
5160 if( (ws
& WHERE_INDEXED
)
5161 || ((ws
& WHERE_MULTI_OR
) && pLevel
->u
.pCovidx
)
5163 sqlite3VdbeAddOp1(v
, OP_NullRow
, pLevel
->iIdxCur
);
5165 if( pLevel
->op
==OP_Return
){
5166 sqlite3VdbeAddOp2(v
, OP_Gosub
, pLevel
->p1
, pLevel
->addrFirst
);
5168 sqlite3VdbeGoto(v
, pLevel
->addrFirst
);
5170 sqlite3VdbeJumpHere(v
, addr
);
5172 VdbeModuleComment((v
, "End WHERE-loop%d: %s", i
,
5173 pWInfo
->pTabList
->a
[pLevel
->iFrom
].pTab
->zName
));
5176 /* The "break" point is here, just past the end of the outer loop.
5179 sqlite3VdbeResolveLabel(v
, pWInfo
->iBreak
);
5181 assert( pWInfo
->nLevel
<=pTabList
->nSrc
);
5182 for(i
=0, pLevel
=pWInfo
->a
; i
<pWInfo
->nLevel
; i
++, pLevel
++){
5186 struct SrcList_item
*pTabItem
= &pTabList
->a
[pLevel
->iFrom
];
5187 Table
*pTab
= pTabItem
->pTab
;
5189 pLoop
= pLevel
->pWLoop
;
5191 /* For a co-routine, change all OP_Column references to the table of
5192 ** the co-routine into OP_Copy of result contained in a register.
5193 ** OP_Rowid becomes OP_Null.
5195 if( pTabItem
->fg
.viaCoroutine
){
5196 testcase( pParse
->db
->mallocFailed
);
5197 translateColumnToCopy(pParse
, pLevel
->addrBody
, pLevel
->iTabCur
,
5198 pTabItem
->regResult
, 0);
5202 /* If this scan uses an index, make VDBE code substitutions to read data
5203 ** from the index instead of from the table where possible. In some cases
5204 ** this optimization prevents the table from ever being read, which can
5205 ** yield a significant performance boost.
5207 ** Calls to the code generator in between sqlite3WhereBegin and
5208 ** sqlite3WhereEnd will have created code that references the table
5209 ** directly. This loop scans all that code looking for opcodes
5210 ** that reference the table and converts them into opcodes that
5211 ** reference the index.
5213 if( pLoop
->wsFlags
& (WHERE_INDEXED
|WHERE_IDX_ONLY
) ){
5214 pIdx
= pLoop
->u
.btree
.pIndex
;
5215 }else if( pLoop
->wsFlags
& WHERE_MULTI_OR
){
5216 pIdx
= pLevel
->u
.pCovidx
;
5219 && (pWInfo
->eOnePass
==ONEPASS_OFF
|| !HasRowid(pIdx
->pTable
))
5220 && !db
->mallocFailed
5222 last
= sqlite3VdbeCurrentAddr(v
);
5223 k
= pLevel
->addrBody
;
5225 if( db
->flags
& SQLITE_VdbeAddopTrace
){
5226 printf("TRANSLATE opcodes in range %d..%d\n", k
, last
-1);
5229 pOp
= sqlite3VdbeGetOp(v
, k
);
5230 for(; k
<last
; k
++, pOp
++){
5231 if( pOp
->p1
!=pLevel
->iTabCur
) continue;
5232 if( pOp
->opcode
==OP_Column
5233 #ifdef SQLITE_ENABLE_OFFSET_SQL_FUNC
5234 || pOp
->opcode
==OP_Offset
5238 assert( pIdx
->pTable
==pTab
);
5239 if( !HasRowid(pTab
) ){
5240 Index
*pPk
= sqlite3PrimaryKeyIndex(pTab
);
5241 x
= pPk
->aiColumn
[x
];
5244 x
= sqlite3ColumnOfIndex(pIdx
, x
);
5247 pOp
->p1
= pLevel
->iIdxCur
;
5248 OpcodeRewriteTrace(db
, k
, pOp
);
5250 assert( (pLoop
->wsFlags
& WHERE_IDX_ONLY
)==0 || x
>=0
5251 || pWInfo
->eOnePass
);
5252 }else if( pOp
->opcode
==OP_Rowid
){
5253 pOp
->p1
= pLevel
->iIdxCur
;
5254 pOp
->opcode
= OP_IdxRowid
;
5255 OpcodeRewriteTrace(db
, k
, pOp
);
5256 }else if( pOp
->opcode
==OP_IfNullRow
){
5257 pOp
->p1
= pLevel
->iIdxCur
;
5258 OpcodeRewriteTrace(db
, k
, pOp
);
5262 if( db
->flags
& SQLITE_VdbeAddopTrace
) printf("TRANSLATE complete\n");
5269 pParse
->nQueryLoop
= pWInfo
->savedNQueryLoop
;
5270 whereInfoFree(db
, pWInfo
);