Merge the latest changes from trunk.
[sqlite.git] / src / where.c
blob31b72d965f52bcac235861e0c014d8e84cbba46c
1 /*
2 ** 2001 September 15
3 **
4 ** The author disclaims copyright to this source code. In place of
5 ** a legal notice, here is a blessing:
6 **
7 ** May you do good and not evil.
8 ** May you find forgiveness for yourself and forgive others.
9 ** May you share freely, never taking more than you give.
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"
20 #include "whereInt.h"
22 /* Forward declaration of methods */
23 static int whereLoopResize(sqlite3*, WhereLoop*, int);
25 /* Test variable that can be set to enable WHERE tracing */
26 #if defined(SQLITE_TEST) || defined(SQLITE_DEBUG)
27 /***/ int sqlite3WhereTrace = 0;
28 #endif
32 ** Return the estimated number of output rows from a WHERE clause
34 LogEst sqlite3WhereOutputRowCount(WhereInfo *pWInfo){
35 return pWInfo->nRowOut;
39 ** Return one of the WHERE_DISTINCT_xxxxx values to indicate how this
40 ** WHERE clause returns outputs for DISTINCT processing.
42 int sqlite3WhereIsDistinct(WhereInfo *pWInfo){
43 return pWInfo->eDistinct;
47 ** Return TRUE if the WHERE clause returns rows in ORDER BY order.
48 ** Return FALSE if the output needs to be sorted.
50 int sqlite3WhereIsOrdered(WhereInfo *pWInfo){
51 return pWInfo->nOBSat;
55 ** Return TRUE if the innermost loop of the WHERE clause implementation
56 ** returns rows in ORDER BY order for complete run of the inner loop.
58 ** Across multiple iterations of outer loops, the output rows need not be
59 ** sorted. As long as rows are sorted for just the innermost loop, this
60 ** routine can return TRUE.
62 int sqlite3WhereOrderedInnerLoop(WhereInfo *pWInfo){
63 return pWInfo->bOrderedInnerLoop;
67 ** Return the VDBE address or label to jump to in order to continue
68 ** immediately with the next row of a WHERE clause.
70 int sqlite3WhereContinueLabel(WhereInfo *pWInfo){
71 assert( pWInfo->iContinue!=0 );
72 return pWInfo->iContinue;
76 ** Return the VDBE address or label to jump to in order to break
77 ** out of a WHERE loop.
79 int sqlite3WhereBreakLabel(WhereInfo *pWInfo){
80 return pWInfo->iBreak;
84 ** Return ONEPASS_OFF (0) if an UPDATE or DELETE statement is unable to
85 ** operate directly on the rowis returned by a WHERE clause. Return
86 ** ONEPASS_SINGLE (1) if the statement can operation directly because only
87 ** a single row is to be changed. Return ONEPASS_MULTI (2) if the one-pass
88 ** optimization can be used on multiple
90 ** If the ONEPASS optimization is used (if this routine returns true)
91 ** then also write the indices of open cursors used by ONEPASS
92 ** into aiCur[0] and aiCur[1]. iaCur[0] gets the cursor of the data
93 ** table and iaCur[1] gets the cursor used by an auxiliary index.
94 ** Either value may be -1, indicating that cursor is not used.
95 ** Any cursors returned will have been opened for writing.
97 ** aiCur[0] and aiCur[1] both get -1 if the where-clause logic is
98 ** unable to use the ONEPASS optimization.
100 int sqlite3WhereOkOnePass(WhereInfo *pWInfo, int *aiCur){
101 memcpy(aiCur, pWInfo->aiCurOnePass, sizeof(int)*2);
102 #ifdef WHERETRACE_ENABLED
103 if( sqlite3WhereTrace && pWInfo->eOnePass!=ONEPASS_OFF ){
104 sqlite3DebugPrintf("%s cursors: %d %d\n",
105 pWInfo->eOnePass==ONEPASS_SINGLE ? "ONEPASS_SINGLE" : "ONEPASS_MULTI",
106 aiCur[0], aiCur[1]);
108 #endif
109 return pWInfo->eOnePass;
113 ** Move the content of pSrc into pDest
115 static void whereOrMove(WhereOrSet *pDest, WhereOrSet *pSrc){
116 pDest->n = pSrc->n;
117 memcpy(pDest->a, pSrc->a, pDest->n*sizeof(pDest->a[0]));
121 ** Try to insert a new prerequisite/cost entry into the WhereOrSet pSet.
123 ** The new entry might overwrite an existing entry, or it might be
124 ** appended, or it might be discarded. Do whatever is the right thing
125 ** so that pSet keeps the N_OR_COST best entries seen so far.
127 static int whereOrInsert(
128 WhereOrSet *pSet, /* The WhereOrSet to be updated */
129 Bitmask prereq, /* Prerequisites of the new entry */
130 LogEst rRun, /* Run-cost of the new entry */
131 LogEst nOut /* Number of outputs for the new entry */
133 u16 i;
134 WhereOrCost *p;
135 for(i=pSet->n, p=pSet->a; i>0; i--, p++){
136 if( rRun<=p->rRun && (prereq & p->prereq)==prereq ){
137 goto whereOrInsert_done;
139 if( p->rRun<=rRun && (p->prereq & prereq)==p->prereq ){
140 return 0;
143 if( pSet->n<N_OR_COST ){
144 p = &pSet->a[pSet->n++];
145 p->nOut = nOut;
146 }else{
147 p = pSet->a;
148 for(i=1; i<pSet->n; i++){
149 if( p->rRun>pSet->a[i].rRun ) p = pSet->a + i;
151 if( p->rRun<=rRun ) return 0;
153 whereOrInsert_done:
154 p->prereq = prereq;
155 p->rRun = rRun;
156 if( p->nOut>nOut ) p->nOut = nOut;
157 return 1;
161 ** Return the bitmask for the given cursor number. Return 0 if
162 ** iCursor is not in the set.
164 Bitmask sqlite3WhereGetMask(WhereMaskSet *pMaskSet, int iCursor){
165 int i;
166 assert( pMaskSet->n<=(int)sizeof(Bitmask)*8 );
167 for(i=0; i<pMaskSet->n; i++){
168 if( pMaskSet->ix[i]==iCursor ){
169 return MASKBIT(i);
172 return 0;
176 ** Create a new mask for cursor iCursor.
178 ** There is one cursor per table in the FROM clause. The number of
179 ** tables in the FROM clause is limited by a test early in the
180 ** sqlite3WhereBegin() routine. So we know that the pMaskSet->ix[]
181 ** array will never overflow.
183 static void createMask(WhereMaskSet *pMaskSet, int iCursor){
184 assert( pMaskSet->n < ArraySize(pMaskSet->ix) );
185 pMaskSet->ix[pMaskSet->n++] = iCursor;
189 ** Advance to the next WhereTerm that matches according to the criteria
190 ** established when the pScan object was initialized by whereScanInit().
191 ** Return NULL if there are no more matching WhereTerms.
193 static WhereTerm *whereScanNext(WhereScan *pScan){
194 int iCur; /* The cursor on the LHS of the term */
195 i16 iColumn; /* The column on the LHS of the term. -1 for IPK */
196 Expr *pX; /* An expression being tested */
197 WhereClause *pWC; /* Shorthand for pScan->pWC */
198 WhereTerm *pTerm; /* The term being tested */
199 int k = pScan->k; /* Where to start scanning */
201 assert( pScan->iEquiv<=pScan->nEquiv );
202 pWC = pScan->pWC;
203 while(1){
204 iColumn = pScan->aiColumn[pScan->iEquiv-1];
205 iCur = pScan->aiCur[pScan->iEquiv-1];
206 assert( pWC!=0 );
208 for(pTerm=pWC->a+k; k<pWC->nTerm; k++, pTerm++){
209 if( pTerm->leftCursor==iCur
210 && pTerm->u.leftColumn==iColumn
211 && (iColumn!=XN_EXPR
212 || sqlite3ExprCompareSkip(pTerm->pExpr->pLeft,
213 pScan->pIdxExpr,iCur)==0)
214 && (pScan->iEquiv<=1 || !ExprHasProperty(pTerm->pExpr, EP_FromJoin))
216 if( (pTerm->eOperator & WO_EQUIV)!=0
217 && pScan->nEquiv<ArraySize(pScan->aiCur)
218 && (pX = sqlite3ExprSkipCollate(pTerm->pExpr->pRight))->op==TK_COLUMN
220 int j;
221 for(j=0; j<pScan->nEquiv; j++){
222 if( pScan->aiCur[j]==pX->iTable
223 && pScan->aiColumn[j]==pX->iColumn ){
224 break;
227 if( j==pScan->nEquiv ){
228 pScan->aiCur[j] = pX->iTable;
229 pScan->aiColumn[j] = pX->iColumn;
230 pScan->nEquiv++;
233 if( (pTerm->eOperator & pScan->opMask)!=0 ){
234 /* Verify the affinity and collating sequence match */
235 if( pScan->zCollName && (pTerm->eOperator & WO_ISNULL)==0 ){
236 CollSeq *pColl;
237 Parse *pParse = pWC->pWInfo->pParse;
238 pX = pTerm->pExpr;
239 if( !sqlite3IndexAffinityOk(pX, pScan->idxaff) ){
240 continue;
242 assert(pX->pLeft);
243 pColl = sqlite3BinaryCompareCollSeq(pParse,
244 pX->pLeft, pX->pRight);
245 if( pColl==0 ) pColl = pParse->db->pDfltColl;
246 if( sqlite3StrICmp(pColl->zName, pScan->zCollName) ){
247 continue;
250 if( (pTerm->eOperator & (WO_EQ|WO_IS))!=0
251 && (pX = pTerm->pExpr->pRight)->op==TK_COLUMN
252 && pX->iTable==pScan->aiCur[0]
253 && pX->iColumn==pScan->aiColumn[0]
255 testcase( pTerm->eOperator & WO_IS );
256 continue;
258 pScan->pWC = pWC;
259 pScan->k = k+1;
260 return pTerm;
264 pWC = pWC->pOuter;
265 k = 0;
266 }while( pWC!=0 );
267 if( pScan->iEquiv>=pScan->nEquiv ) break;
268 pWC = pScan->pOrigWC;
269 k = 0;
270 pScan->iEquiv++;
272 return 0;
276 ** Initialize a WHERE clause scanner object. Return a pointer to the
277 ** first match. Return NULL if there are no matches.
279 ** The scanner will be searching the WHERE clause pWC. It will look
280 ** for terms of the form "X <op> <expr>" where X is column iColumn of table
281 ** iCur. Or if pIdx!=0 then X is column iColumn of index pIdx. pIdx
282 ** must be one of the indexes of table iCur.
284 ** The <op> must be one of the operators described by opMask.
286 ** If the search is for X and the WHERE clause contains terms of the
287 ** form X=Y then this routine might also return terms of the form
288 ** "Y <op> <expr>". The number of levels of transitivity is limited,
289 ** but is enough to handle most commonly occurring SQL statements.
291 ** If X is not the INTEGER PRIMARY KEY then X must be compatible with
292 ** index pIdx.
294 static WhereTerm *whereScanInit(
295 WhereScan *pScan, /* The WhereScan object being initialized */
296 WhereClause *pWC, /* The WHERE clause to be scanned */
297 int iCur, /* Cursor to scan for */
298 int iColumn, /* Column to scan for */
299 u32 opMask, /* Operator(s) to scan for */
300 Index *pIdx /* Must be compatible with this index */
302 pScan->pOrigWC = pWC;
303 pScan->pWC = pWC;
304 pScan->pIdxExpr = 0;
305 pScan->idxaff = 0;
306 pScan->zCollName = 0;
307 if( pIdx ){
308 int j = iColumn;
309 iColumn = pIdx->aiColumn[j];
310 if( iColumn==XN_EXPR ){
311 pScan->pIdxExpr = pIdx->aColExpr->a[j].pExpr;
312 pScan->zCollName = pIdx->azColl[j];
313 }else if( iColumn==pIdx->pTable->iPKey ){
314 iColumn = XN_ROWID;
315 }else if( iColumn>=0 ){
316 pScan->idxaff = pIdx->pTable->aCol[iColumn].affinity;
317 pScan->zCollName = pIdx->azColl[j];
319 }else if( iColumn==XN_EXPR ){
320 return 0;
322 pScan->opMask = opMask;
323 pScan->k = 0;
324 pScan->aiCur[0] = iCur;
325 pScan->aiColumn[0] = iColumn;
326 pScan->nEquiv = 1;
327 pScan->iEquiv = 1;
328 return whereScanNext(pScan);
332 ** Search for a term in the WHERE clause that is of the form "X <op> <expr>"
333 ** where X is a reference to the iColumn of table iCur or of index pIdx
334 ** if pIdx!=0 and <op> is one of the WO_xx operator codes specified by
335 ** the op parameter. Return a pointer to the term. Return 0 if not found.
337 ** If pIdx!=0 then it must be one of the indexes of table iCur.
338 ** Search for terms matching the iColumn-th column of pIdx
339 ** rather than the iColumn-th column of table iCur.
341 ** The term returned might by Y=<expr> if there is another constraint in
342 ** the WHERE clause that specifies that X=Y. Any such constraints will be
343 ** identified by the WO_EQUIV bit in the pTerm->eOperator field. The
344 ** aiCur[]/iaColumn[] arrays hold X and all its equivalents. There are 11
345 ** slots in aiCur[]/aiColumn[] so that means we can look for X plus up to 10
346 ** other equivalent values. Hence a search for X will return <expr> if X=A1
347 ** and A1=A2 and A2=A3 and ... and A9=A10 and A10=<expr>.
349 ** If there are multiple terms in the WHERE clause of the form "X <op> <expr>"
350 ** then try for the one with no dependencies on <expr> - in other words where
351 ** <expr> is a constant expression of some kind. Only return entries of
352 ** the form "X <op> Y" where Y is a column in another table if no terms of
353 ** the form "X <op> <const-expr>" exist. If no terms with a constant RHS
354 ** exist, try to return a term that does not use WO_EQUIV.
356 WhereTerm *sqlite3WhereFindTerm(
357 WhereClause *pWC, /* The WHERE clause to be searched */
358 int iCur, /* Cursor number of LHS */
359 int iColumn, /* Column number of LHS */
360 Bitmask notReady, /* RHS must not overlap with this mask */
361 u32 op, /* Mask of WO_xx values describing operator */
362 Index *pIdx /* Must be compatible with this index, if not NULL */
364 WhereTerm *pResult = 0;
365 WhereTerm *p;
366 WhereScan scan;
368 p = whereScanInit(&scan, pWC, iCur, iColumn, op, pIdx);
369 op &= WO_EQ|WO_IS;
370 while( p ){
371 if( (p->prereqRight & notReady)==0 ){
372 if( p->prereqRight==0 && (p->eOperator&op)!=0 ){
373 testcase( p->eOperator & WO_IS );
374 return p;
376 if( pResult==0 ) pResult = p;
378 p = whereScanNext(&scan);
380 return pResult;
384 ** This function searches pList for an entry that matches the iCol-th column
385 ** of index pIdx.
387 ** If such an expression is found, its index in pList->a[] is returned. If
388 ** no expression is found, -1 is returned.
390 static int findIndexCol(
391 Parse *pParse, /* Parse context */
392 ExprList *pList, /* Expression list to search */
393 int iBase, /* Cursor for table associated with pIdx */
394 Index *pIdx, /* Index to match column of */
395 int iCol /* Column of index to match */
397 int i;
398 const char *zColl = pIdx->azColl[iCol];
400 for(i=0; i<pList->nExpr; i++){
401 Expr *p = sqlite3ExprSkipCollate(pList->a[i].pExpr);
402 if( p->op==TK_COLUMN
403 && p->iColumn==pIdx->aiColumn[iCol]
404 && p->iTable==iBase
406 CollSeq *pColl = sqlite3ExprCollSeq(pParse, pList->a[i].pExpr);
407 if( pColl && 0==sqlite3StrICmp(pColl->zName, zColl) ){
408 return i;
413 return -1;
417 ** Return TRUE if the iCol-th column of index pIdx is NOT NULL
419 static int indexColumnNotNull(Index *pIdx, int iCol){
420 int j;
421 assert( pIdx!=0 );
422 assert( iCol>=0 && iCol<pIdx->nColumn );
423 j = pIdx->aiColumn[iCol];
424 if( j>=0 ){
425 return pIdx->pTable->aCol[j].notNull;
426 }else if( j==(-1) ){
427 return 1;
428 }else{
429 assert( j==(-2) );
430 return 0; /* Assume an indexed expression can always yield a NULL */
436 ** Return true if the DISTINCT expression-list passed as the third argument
437 ** is redundant.
439 ** A DISTINCT list is redundant if any subset of the columns in the
440 ** DISTINCT list are collectively unique and individually non-null.
442 static int isDistinctRedundant(
443 Parse *pParse, /* Parsing context */
444 SrcList *pTabList, /* The FROM clause */
445 WhereClause *pWC, /* The WHERE clause */
446 ExprList *pDistinct /* The result set that needs to be DISTINCT */
448 Table *pTab;
449 Index *pIdx;
450 int i;
451 int iBase;
453 /* If there is more than one table or sub-select in the FROM clause of
454 ** this query, then it will not be possible to show that the DISTINCT
455 ** clause is redundant. */
456 if( pTabList->nSrc!=1 ) return 0;
457 iBase = pTabList->a[0].iCursor;
458 pTab = pTabList->a[0].pTab;
460 /* If any of the expressions is an IPK column on table iBase, then return
461 ** true. Note: The (p->iTable==iBase) part of this test may be false if the
462 ** current SELECT is a correlated sub-query.
464 for(i=0; i<pDistinct->nExpr; i++){
465 Expr *p = sqlite3ExprSkipCollate(pDistinct->a[i].pExpr);
466 if( p->op==TK_COLUMN && p->iTable==iBase && p->iColumn<0 ) return 1;
469 /* Loop through all indices on the table, checking each to see if it makes
470 ** the DISTINCT qualifier redundant. It does so if:
472 ** 1. The index is itself UNIQUE, and
474 ** 2. All of the columns in the index are either part of the pDistinct
475 ** list, or else the WHERE clause contains a term of the form "col=X",
476 ** where X is a constant value. The collation sequences of the
477 ** comparison and select-list expressions must match those of the index.
479 ** 3. All of those index columns for which the WHERE clause does not
480 ** contain a "col=X" term are subject to a NOT NULL constraint.
482 for(pIdx=pTab->pIndex; pIdx; pIdx=pIdx->pNext){
483 if( !IsUniqueIndex(pIdx) ) continue;
484 for(i=0; i<pIdx->nKeyCol; i++){
485 if( 0==sqlite3WhereFindTerm(pWC, iBase, i, ~(Bitmask)0, WO_EQ, pIdx) ){
486 if( findIndexCol(pParse, pDistinct, iBase, pIdx, i)<0 ) break;
487 if( indexColumnNotNull(pIdx, i)==0 ) break;
490 if( i==pIdx->nKeyCol ){
491 /* This index implies that the DISTINCT qualifier is redundant. */
492 return 1;
496 return 0;
501 ** Estimate the logarithm of the input value to base 2.
503 static LogEst estLog(LogEst N){
504 return N<=10 ? 0 : sqlite3LogEst(N) - 33;
508 ** Convert OP_Column opcodes to OP_Copy in previously generated code.
510 ** This routine runs over generated VDBE code and translates OP_Column
511 ** opcodes into OP_Copy when the table is being accessed via co-routine
512 ** instead of via table lookup.
514 ** If the bIncrRowid parameter is 0, then any OP_Rowid instructions on
515 ** cursor iTabCur are transformed into OP_Null. Or, if bIncrRowid is non-zero,
516 ** then each OP_Rowid is transformed into an instruction to increment the
517 ** value stored in its output register.
519 static void translateColumnToCopy(
520 Parse *pParse, /* Parsing context */
521 int iStart, /* Translate from this opcode to the end */
522 int iTabCur, /* OP_Column/OP_Rowid references to this table */
523 int iRegister, /* The first column is in this register */
524 int bIncrRowid /* If non-zero, transform OP_rowid to OP_AddImm(1) */
526 Vdbe *v = pParse->pVdbe;
527 VdbeOp *pOp = sqlite3VdbeGetOp(v, iStart);
528 int iEnd = sqlite3VdbeCurrentAddr(v);
529 if( pParse->db->mallocFailed ) return;
530 for(; iStart<iEnd; iStart++, pOp++){
531 if( pOp->p1!=iTabCur ) continue;
532 if( pOp->opcode==OP_Column ){
533 pOp->opcode = OP_Copy;
534 pOp->p1 = pOp->p2 + iRegister;
535 pOp->p2 = pOp->p3;
536 pOp->p3 = 0;
537 }else if( pOp->opcode==OP_Rowid ){
538 if( bIncrRowid ){
539 /* Increment the value stored in the P2 operand of the OP_Rowid. */
540 pOp->opcode = OP_AddImm;
541 pOp->p1 = pOp->p2;
542 pOp->p2 = 1;
543 }else{
544 pOp->opcode = OP_Null;
545 pOp->p1 = 0;
546 pOp->p3 = 0;
553 ** Two routines for printing the content of an sqlite3_index_info
554 ** structure. Used for testing and debugging only. If neither
555 ** SQLITE_TEST or SQLITE_DEBUG are defined, then these routines
556 ** are no-ops.
558 #if !defined(SQLITE_OMIT_VIRTUALTABLE) && defined(WHERETRACE_ENABLED)
559 static void TRACE_IDX_INPUTS(sqlite3_index_info *p){
560 int i;
561 if( !sqlite3WhereTrace ) return;
562 for(i=0; i<p->nConstraint; i++){
563 sqlite3DebugPrintf(" constraint[%d]: col=%d termid=%d op=%d usabled=%d\n",
565 p->aConstraint[i].iColumn,
566 p->aConstraint[i].iTermOffset,
567 p->aConstraint[i].op,
568 p->aConstraint[i].usable);
570 for(i=0; i<p->nOrderBy; i++){
571 sqlite3DebugPrintf(" orderby[%d]: col=%d desc=%d\n",
573 p->aOrderBy[i].iColumn,
574 p->aOrderBy[i].desc);
577 static void TRACE_IDX_OUTPUTS(sqlite3_index_info *p){
578 int i;
579 if( !sqlite3WhereTrace ) return;
580 for(i=0; i<p->nConstraint; i++){
581 sqlite3DebugPrintf(" usage[%d]: argvIdx=%d omit=%d\n",
583 p->aConstraintUsage[i].argvIndex,
584 p->aConstraintUsage[i].omit);
586 sqlite3DebugPrintf(" idxNum=%d\n", p->idxNum);
587 sqlite3DebugPrintf(" idxStr=%s\n", p->idxStr);
588 sqlite3DebugPrintf(" orderByConsumed=%d\n", p->orderByConsumed);
589 sqlite3DebugPrintf(" estimatedCost=%g\n", p->estimatedCost);
590 sqlite3DebugPrintf(" estimatedRows=%lld\n", p->estimatedRows);
592 #else
593 #define TRACE_IDX_INPUTS(A)
594 #define TRACE_IDX_OUTPUTS(A)
595 #endif
597 #ifndef SQLITE_OMIT_AUTOMATIC_INDEX
599 ** Return TRUE if the WHERE clause term pTerm is of a form where it
600 ** could be used with an index to access pSrc, assuming an appropriate
601 ** index existed.
603 static int termCanDriveIndex(
604 WhereTerm *pTerm, /* WHERE clause term to check */
605 struct SrcList_item *pSrc, /* Table we are trying to access */
606 Bitmask notReady /* Tables in outer loops of the join */
608 char aff;
609 if( pTerm->leftCursor!=pSrc->iCursor ) return 0;
610 if( (pTerm->eOperator & (WO_EQ|WO_IS))==0 ) return 0;
611 if( (pTerm->prereqRight & notReady)!=0 ) return 0;
612 if( pTerm->u.leftColumn<0 ) return 0;
613 aff = pSrc->pTab->aCol[pTerm->u.leftColumn].affinity;
614 if( !sqlite3IndexAffinityOk(pTerm->pExpr, aff) ) return 0;
615 testcase( pTerm->pExpr->op==TK_IS );
616 return 1;
618 #endif
621 #ifndef SQLITE_OMIT_AUTOMATIC_INDEX
623 ** Generate code to construct the Index object for an automatic index
624 ** and to set up the WhereLevel object pLevel so that the code generator
625 ** makes use of the automatic index.
627 static void constructAutomaticIndex(
628 Parse *pParse, /* The parsing context */
629 WhereClause *pWC, /* The WHERE clause */
630 struct SrcList_item *pSrc, /* The FROM clause term to get the next index */
631 Bitmask notReady, /* Mask of cursors that are not available */
632 WhereLevel *pLevel /* Write new index here */
634 int nKeyCol; /* Number of columns in the constructed index */
635 WhereTerm *pTerm; /* A single term of the WHERE clause */
636 WhereTerm *pWCEnd; /* End of pWC->a[] */
637 Index *pIdx; /* Object describing the transient index */
638 Vdbe *v; /* Prepared statement under construction */
639 int addrInit; /* Address of the initialization bypass jump */
640 Table *pTable; /* The table being indexed */
641 int addrTop; /* Top of the index fill loop */
642 int regRecord; /* Register holding an index record */
643 int n; /* Column counter */
644 int i; /* Loop counter */
645 int mxBitCol; /* Maximum column in pSrc->colUsed */
646 CollSeq *pColl; /* Collating sequence to on a column */
647 WhereLoop *pLoop; /* The Loop object */
648 char *zNotUsed; /* Extra space on the end of pIdx */
649 Bitmask idxCols; /* Bitmap of columns used for indexing */
650 Bitmask extraCols; /* Bitmap of additional columns */
651 u8 sentWarning = 0; /* True if a warnning has been issued */
652 Expr *pPartial = 0; /* Partial Index Expression */
653 int iContinue = 0; /* Jump here to skip excluded rows */
654 struct SrcList_item *pTabItem; /* FROM clause term being indexed */
655 int addrCounter = 0; /* Address where integer counter is initialized */
656 int regBase; /* Array of registers where record is assembled */
658 /* Generate code to skip over the creation and initialization of the
659 ** transient index on 2nd and subsequent iterations of the loop. */
660 v = pParse->pVdbe;
661 assert( v!=0 );
662 addrInit = sqlite3VdbeAddOp0(v, OP_Once); VdbeCoverage(v);
664 /* Count the number of columns that will be added to the index
665 ** and used to match WHERE clause constraints */
666 nKeyCol = 0;
667 pTable = pSrc->pTab;
668 pWCEnd = &pWC->a[pWC->nTerm];
669 pLoop = pLevel->pWLoop;
670 idxCols = 0;
671 for(pTerm=pWC->a; pTerm<pWCEnd; pTerm++){
672 Expr *pExpr = pTerm->pExpr;
673 assert( !ExprHasProperty(pExpr, EP_FromJoin) /* prereq always non-zero */
674 || pExpr->iRightJoinTable!=pSrc->iCursor /* for the right-hand */
675 || pLoop->prereq!=0 ); /* table of a LEFT JOIN */
676 if( pLoop->prereq==0
677 && (pTerm->wtFlags & TERM_VIRTUAL)==0
678 && !ExprHasProperty(pExpr, EP_FromJoin)
679 && sqlite3ExprIsTableConstant(pExpr, pSrc->iCursor) ){
680 pPartial = sqlite3ExprAnd(pParse->db, pPartial,
681 sqlite3ExprDup(pParse->db, pExpr, 0));
683 if( termCanDriveIndex(pTerm, pSrc, notReady) ){
684 int iCol = pTerm->u.leftColumn;
685 Bitmask cMask = iCol>=BMS ? MASKBIT(BMS-1) : MASKBIT(iCol);
686 testcase( iCol==BMS );
687 testcase( iCol==BMS-1 );
688 if( !sentWarning ){
689 sqlite3_log(SQLITE_WARNING_AUTOINDEX,
690 "automatic index on %s(%s)", pTable->zName,
691 pTable->aCol[iCol].zName);
692 sentWarning = 1;
694 if( (idxCols & cMask)==0 ){
695 if( whereLoopResize(pParse->db, pLoop, nKeyCol+1) ){
696 goto end_auto_index_create;
698 pLoop->aLTerm[nKeyCol++] = pTerm;
699 idxCols |= cMask;
703 assert( nKeyCol>0 );
704 pLoop->u.btree.nEq = pLoop->nLTerm = nKeyCol;
705 pLoop->wsFlags = WHERE_COLUMN_EQ | WHERE_IDX_ONLY | WHERE_INDEXED
706 | WHERE_AUTO_INDEX;
708 /* Count the number of additional columns needed to create a
709 ** covering index. A "covering index" is an index that contains all
710 ** columns that are needed by the query. With a covering index, the
711 ** original table never needs to be accessed. Automatic indices must
712 ** be a covering index because the index will not be updated if the
713 ** original table changes and the index and table cannot both be used
714 ** if they go out of sync.
716 extraCols = pSrc->colUsed & (~idxCols | MASKBIT(BMS-1));
717 mxBitCol = MIN(BMS-1,pTable->nCol);
718 testcase( pTable->nCol==BMS-1 );
719 testcase( pTable->nCol==BMS-2 );
720 for(i=0; i<mxBitCol; i++){
721 if( extraCols & MASKBIT(i) ) nKeyCol++;
723 if( pSrc->colUsed & MASKBIT(BMS-1) ){
724 nKeyCol += pTable->nCol - BMS + 1;
727 /* Construct the Index object to describe this index */
728 pIdx = sqlite3AllocateIndexObject(pParse->db, nKeyCol+1, 0, &zNotUsed);
729 if( pIdx==0 ) goto end_auto_index_create;
730 pLoop->u.btree.pIndex = pIdx;
731 pIdx->zName = "auto-index";
732 pIdx->pTable = pTable;
733 n = 0;
734 idxCols = 0;
735 for(pTerm=pWC->a; pTerm<pWCEnd; pTerm++){
736 if( termCanDriveIndex(pTerm, pSrc, notReady) ){
737 int iCol = pTerm->u.leftColumn;
738 Bitmask cMask = iCol>=BMS ? MASKBIT(BMS-1) : MASKBIT(iCol);
739 testcase( iCol==BMS-1 );
740 testcase( iCol==BMS );
741 if( (idxCols & cMask)==0 ){
742 Expr *pX = pTerm->pExpr;
743 idxCols |= cMask;
744 pIdx->aiColumn[n] = pTerm->u.leftColumn;
745 pColl = sqlite3BinaryCompareCollSeq(pParse, pX->pLeft, pX->pRight);
746 pIdx->azColl[n] = pColl ? pColl->zName : sqlite3StrBINARY;
747 n++;
751 assert( (u32)n==pLoop->u.btree.nEq );
753 /* Add additional columns needed to make the automatic index into
754 ** a covering index */
755 for(i=0; i<mxBitCol; i++){
756 if( extraCols & MASKBIT(i) ){
757 pIdx->aiColumn[n] = i;
758 pIdx->azColl[n] = sqlite3StrBINARY;
759 n++;
762 if( pSrc->colUsed & MASKBIT(BMS-1) ){
763 for(i=BMS-1; i<pTable->nCol; i++){
764 pIdx->aiColumn[n] = i;
765 pIdx->azColl[n] = sqlite3StrBINARY;
766 n++;
769 assert( n==nKeyCol );
770 pIdx->aiColumn[n] = XN_ROWID;
771 pIdx->azColl[n] = sqlite3StrBINARY;
773 /* Create the automatic index */
774 assert( pLevel->iIdxCur>=0 );
775 pLevel->iIdxCur = pParse->nTab++;
776 sqlite3VdbeAddOp2(v, OP_OpenAutoindex, pLevel->iIdxCur, nKeyCol+1);
777 sqlite3VdbeSetP4KeyInfo(pParse, pIdx);
778 VdbeComment((v, "for %s", pTable->zName));
780 /* Fill the automatic index with content */
781 sqlite3ExprCachePush(pParse);
782 pTabItem = &pWC->pWInfo->pTabList->a[pLevel->iFrom];
783 if( pTabItem->fg.viaCoroutine ){
784 int regYield = pTabItem->regReturn;
785 addrCounter = sqlite3VdbeAddOp2(v, OP_Integer, 0, 0);
786 sqlite3VdbeAddOp3(v, OP_InitCoroutine, regYield, 0, pTabItem->addrFillSub);
787 addrTop = sqlite3VdbeAddOp1(v, OP_Yield, regYield);
788 VdbeCoverage(v);
789 VdbeComment((v, "next row of \"%s\"", pTabItem->pTab->zName));
790 }else{
791 addrTop = sqlite3VdbeAddOp1(v, OP_Rewind, pLevel->iTabCur); VdbeCoverage(v);
793 if( pPartial ){
794 iContinue = sqlite3VdbeMakeLabel(v);
795 sqlite3ExprIfFalse(pParse, pPartial, iContinue, SQLITE_JUMPIFNULL);
796 pLoop->wsFlags |= WHERE_PARTIALIDX;
798 regRecord = sqlite3GetTempReg(pParse);
799 regBase = sqlite3GenerateIndexKey(
800 pParse, pIdx, pLevel->iTabCur, regRecord, 0, 0, 0, 0
802 sqlite3VdbeAddOp2(v, OP_IdxInsert, pLevel->iIdxCur, regRecord);
803 sqlite3VdbeChangeP5(v, OPFLAG_USESEEKRESULT);
804 if( pPartial ) sqlite3VdbeResolveLabel(v, iContinue);
805 if( pTabItem->fg.viaCoroutine ){
806 sqlite3VdbeChangeP2(v, addrCounter, regBase+n);
807 testcase( pParse->db->mallocFailed );
808 translateColumnToCopy(pParse, addrTop, pLevel->iTabCur,
809 pTabItem->regResult, 1);
810 sqlite3VdbeGoto(v, addrTop);
811 pTabItem->fg.viaCoroutine = 0;
812 }else{
813 sqlite3VdbeAddOp2(v, OP_Next, pLevel->iTabCur, addrTop+1); VdbeCoverage(v);
815 sqlite3VdbeChangeP5(v, SQLITE_STMTSTATUS_AUTOINDEX);
816 sqlite3VdbeJumpHere(v, addrTop);
817 sqlite3ReleaseTempReg(pParse, regRecord);
818 sqlite3ExprCachePop(pParse);
820 /* Jump here when skipping the initialization */
821 sqlite3VdbeJumpHere(v, addrInit);
823 end_auto_index_create:
824 sqlite3ExprDelete(pParse->db, pPartial);
826 #endif /* SQLITE_OMIT_AUTOMATIC_INDEX */
828 #ifndef SQLITE_OMIT_VIRTUALTABLE
830 ** Allocate and populate an sqlite3_index_info structure. It is the
831 ** responsibility of the caller to eventually release the structure
832 ** by passing the pointer returned by this function to sqlite3_free().
834 static sqlite3_index_info *allocateIndexInfo(
835 Parse *pParse,
836 WhereClause *pWC,
837 Bitmask mUnusable, /* Ignore terms with these prereqs */
838 struct SrcList_item *pSrc,
839 ExprList *pOrderBy,
840 u16 *pmNoOmit /* Mask of terms not to omit */
842 int i, j;
843 int nTerm;
844 struct sqlite3_index_constraint *pIdxCons;
845 struct sqlite3_index_orderby *pIdxOrderBy;
846 struct sqlite3_index_constraint_usage *pUsage;
847 WhereTerm *pTerm;
848 int nOrderBy;
849 sqlite3_index_info *pIdxInfo;
850 u16 mNoOmit = 0;
852 /* Count the number of possible WHERE clause constraints referring
853 ** to this virtual table */
854 for(i=nTerm=0, pTerm=pWC->a; i<pWC->nTerm; i++, pTerm++){
855 if( pTerm->leftCursor != pSrc->iCursor ) continue;
856 if( pTerm->prereqRight & mUnusable ) continue;
857 assert( IsPowerOfTwo(pTerm->eOperator & ~WO_EQUIV) );
858 testcase( pTerm->eOperator & WO_IN );
859 testcase( pTerm->eOperator & WO_ISNULL );
860 testcase( pTerm->eOperator & WO_IS );
861 testcase( pTerm->eOperator & WO_ALL );
862 if( (pTerm->eOperator & ~(WO_ISNULL|WO_EQUIV|WO_IS))==0 ) continue;
863 if( pTerm->wtFlags & TERM_VNULL ) continue;
864 assert( pTerm->u.leftColumn>=(-1) );
865 nTerm++;
868 /* If the ORDER BY clause contains only columns in the current
869 ** virtual table then allocate space for the aOrderBy part of
870 ** the sqlite3_index_info structure.
872 nOrderBy = 0;
873 if( pOrderBy ){
874 int n = pOrderBy->nExpr;
875 for(i=0; i<n; i++){
876 Expr *pExpr = pOrderBy->a[i].pExpr;
877 if( pExpr->op!=TK_COLUMN || pExpr->iTable!=pSrc->iCursor ) break;
879 if( i==n){
880 nOrderBy = n;
884 /* Allocate the sqlite3_index_info structure
886 pIdxInfo = sqlite3DbMallocZero(pParse->db, sizeof(*pIdxInfo)
887 + (sizeof(*pIdxCons) + sizeof(*pUsage))*nTerm
888 + sizeof(*pIdxOrderBy)*nOrderBy );
889 if( pIdxInfo==0 ){
890 sqlite3ErrorMsg(pParse, "out of memory");
891 return 0;
894 /* Initialize the structure. The sqlite3_index_info structure contains
895 ** many fields that are declared "const" to prevent xBestIndex from
896 ** changing them. We have to do some funky casting in order to
897 ** initialize those fields.
899 pIdxCons = (struct sqlite3_index_constraint*)&pIdxInfo[1];
900 pIdxOrderBy = (struct sqlite3_index_orderby*)&pIdxCons[nTerm];
901 pUsage = (struct sqlite3_index_constraint_usage*)&pIdxOrderBy[nOrderBy];
902 *(int*)&pIdxInfo->nConstraint = nTerm;
903 *(int*)&pIdxInfo->nOrderBy = nOrderBy;
904 *(struct sqlite3_index_constraint**)&pIdxInfo->aConstraint = pIdxCons;
905 *(struct sqlite3_index_orderby**)&pIdxInfo->aOrderBy = pIdxOrderBy;
906 *(struct sqlite3_index_constraint_usage**)&pIdxInfo->aConstraintUsage =
907 pUsage;
909 for(i=j=0, pTerm=pWC->a; i<pWC->nTerm; i++, pTerm++){
910 u8 op;
911 if( pTerm->leftCursor != pSrc->iCursor ) continue;
912 if( pTerm->prereqRight & mUnusable ) continue;
913 assert( IsPowerOfTwo(pTerm->eOperator & ~WO_EQUIV) );
914 testcase( pTerm->eOperator & WO_IN );
915 testcase( pTerm->eOperator & WO_IS );
916 testcase( pTerm->eOperator & WO_ISNULL );
917 testcase( pTerm->eOperator & WO_ALL );
918 if( (pTerm->eOperator & ~(WO_ISNULL|WO_EQUIV|WO_IS))==0 ) continue;
919 if( pTerm->wtFlags & TERM_VNULL ) continue;
920 assert( pTerm->u.leftColumn>=(-1) );
921 pIdxCons[j].iColumn = pTerm->u.leftColumn;
922 pIdxCons[j].iTermOffset = i;
923 op = (u8)pTerm->eOperator & WO_ALL;
924 if( op==WO_IN ) op = WO_EQ;
925 if( op==WO_MATCH ){
926 op = pTerm->eMatchOp;
928 pIdxCons[j].op = op;
929 /* The direct assignment in the previous line is possible only because
930 ** the WO_ and SQLITE_INDEX_CONSTRAINT_ codes are identical. The
931 ** following asserts verify this fact. */
932 assert( WO_EQ==SQLITE_INDEX_CONSTRAINT_EQ );
933 assert( WO_LT==SQLITE_INDEX_CONSTRAINT_LT );
934 assert( WO_LE==SQLITE_INDEX_CONSTRAINT_LE );
935 assert( WO_GT==SQLITE_INDEX_CONSTRAINT_GT );
936 assert( WO_GE==SQLITE_INDEX_CONSTRAINT_GE );
937 assert( WO_MATCH==SQLITE_INDEX_CONSTRAINT_MATCH );
938 assert( pTerm->eOperator & (WO_IN|WO_EQ|WO_LT|WO_LE|WO_GT|WO_GE|WO_MATCH) );
940 if( op & (WO_LT|WO_LE|WO_GT|WO_GE)
941 && sqlite3ExprIsVector(pTerm->pExpr->pRight)
943 if( i<16 ) mNoOmit |= (1 << i);
944 if( op==WO_LT ) pIdxCons[j].op = WO_LE;
945 if( op==WO_GT ) pIdxCons[j].op = WO_GE;
948 j++;
950 for(i=0; i<nOrderBy; i++){
951 Expr *pExpr = pOrderBy->a[i].pExpr;
952 pIdxOrderBy[i].iColumn = pExpr->iColumn;
953 pIdxOrderBy[i].desc = pOrderBy->a[i].sortOrder;
956 *pmNoOmit = mNoOmit;
957 return pIdxInfo;
961 ** The table object reference passed as the second argument to this function
962 ** must represent a virtual table. This function invokes the xBestIndex()
963 ** method of the virtual table with the sqlite3_index_info object that
964 ** comes in as the 3rd argument to this function.
966 ** If an error occurs, pParse is populated with an error message and a
967 ** non-zero value is returned. Otherwise, 0 is returned and the output
968 ** part of the sqlite3_index_info structure is left populated.
970 ** Whether or not an error is returned, it is the responsibility of the
971 ** caller to eventually free p->idxStr if p->needToFreeIdxStr indicates
972 ** that this is required.
974 static int vtabBestIndex(Parse *pParse, Table *pTab, sqlite3_index_info *p){
975 sqlite3_vtab *pVtab = sqlite3GetVTable(pParse->db, pTab)->pVtab;
976 int rc;
978 TRACE_IDX_INPUTS(p);
979 rc = pVtab->pModule->xBestIndex(pVtab, p);
980 TRACE_IDX_OUTPUTS(p);
982 if( rc!=SQLITE_OK ){
983 if( rc==SQLITE_NOMEM ){
984 sqlite3OomFault(pParse->db);
985 }else if( !pVtab->zErrMsg ){
986 sqlite3ErrorMsg(pParse, "%s", sqlite3ErrStr(rc));
987 }else{
988 sqlite3ErrorMsg(pParse, "%s", pVtab->zErrMsg);
991 sqlite3_free(pVtab->zErrMsg);
992 pVtab->zErrMsg = 0;
994 #if 0
995 /* This error is now caught by the caller.
996 ** Search for "xBestIndex malfunction" below */
997 for(i=0; i<p->nConstraint; i++){
998 if( !p->aConstraint[i].usable && p->aConstraintUsage[i].argvIndex>0 ){
999 sqlite3ErrorMsg(pParse,
1000 "table %s: xBestIndex returned an invalid plan", pTab->zName);
1003 #endif
1005 return pParse->nErr;
1007 #endif /* !defined(SQLITE_OMIT_VIRTUALTABLE) */
1009 #ifdef SQLITE_ENABLE_STAT3_OR_STAT4
1011 ** Estimate the location of a particular key among all keys in an
1012 ** index. Store the results in aStat as follows:
1014 ** aStat[0] Est. number of rows less than pRec
1015 ** aStat[1] Est. number of rows equal to pRec
1017 ** Return the index of the sample that is the smallest sample that
1018 ** is greater than or equal to pRec. Note that this index is not an index
1019 ** into the aSample[] array - it is an index into a virtual set of samples
1020 ** based on the contents of aSample[] and the number of fields in record
1021 ** pRec.
1023 static int whereKeyStats(
1024 Parse *pParse, /* Database connection */
1025 Index *pIdx, /* Index to consider domain of */
1026 UnpackedRecord *pRec, /* Vector of values to consider */
1027 int roundUp, /* Round up if true. Round down if false */
1028 tRowcnt *aStat /* OUT: stats written here */
1030 IndexSample *aSample = pIdx->aSample;
1031 int iCol; /* Index of required stats in anEq[] etc. */
1032 int i; /* Index of first sample >= pRec */
1033 int iSample; /* Smallest sample larger than or equal to pRec */
1034 int iMin = 0; /* Smallest sample not yet tested */
1035 int iTest; /* Next sample to test */
1036 int res; /* Result of comparison operation */
1037 int nField; /* Number of fields in pRec */
1038 tRowcnt iLower = 0; /* anLt[] + anEq[] of largest sample pRec is > */
1040 #ifndef SQLITE_DEBUG
1041 UNUSED_PARAMETER( pParse );
1042 #endif
1043 assert( pRec!=0 );
1044 assert( pIdx->nSample>0 );
1045 assert( pRec->nField>0 && pRec->nField<=pIdx->nSampleCol );
1047 /* Do a binary search to find the first sample greater than or equal
1048 ** to pRec. If pRec contains a single field, the set of samples to search
1049 ** is simply the aSample[] array. If the samples in aSample[] contain more
1050 ** than one fields, all fields following the first are ignored.
1052 ** If pRec contains N fields, where N is more than one, then as well as the
1053 ** samples in aSample[] (truncated to N fields), the search also has to
1054 ** consider prefixes of those samples. For example, if the set of samples
1055 ** in aSample is:
1057 ** aSample[0] = (a, 5)
1058 ** aSample[1] = (a, 10)
1059 ** aSample[2] = (b, 5)
1060 ** aSample[3] = (c, 100)
1061 ** aSample[4] = (c, 105)
1063 ** Then the search space should ideally be the samples above and the
1064 ** unique prefixes [a], [b] and [c]. But since that is hard to organize,
1065 ** the code actually searches this set:
1067 ** 0: (a)
1068 ** 1: (a, 5)
1069 ** 2: (a, 10)
1070 ** 3: (a, 10)
1071 ** 4: (b)
1072 ** 5: (b, 5)
1073 ** 6: (c)
1074 ** 7: (c, 100)
1075 ** 8: (c, 105)
1076 ** 9: (c, 105)
1078 ** For each sample in the aSample[] array, N samples are present in the
1079 ** effective sample array. In the above, samples 0 and 1 are based on
1080 ** sample aSample[0]. Samples 2 and 3 on aSample[1] etc.
1082 ** Often, sample i of each block of N effective samples has (i+1) fields.
1083 ** Except, each sample may be extended to ensure that it is greater than or
1084 ** equal to the previous sample in the array. For example, in the above,
1085 ** sample 2 is the first sample of a block of N samples, so at first it
1086 ** appears that it should be 1 field in size. However, that would make it
1087 ** smaller than sample 1, so the binary search would not work. As a result,
1088 ** it is extended to two fields. The duplicates that this creates do not
1089 ** cause any problems.
1091 nField = pRec->nField;
1092 iCol = 0;
1093 iSample = pIdx->nSample * nField;
1095 int iSamp; /* Index in aSample[] of test sample */
1096 int n; /* Number of fields in test sample */
1098 iTest = (iMin+iSample)/2;
1099 iSamp = iTest / nField;
1100 if( iSamp>0 ){
1101 /* The proposed effective sample is a prefix of sample aSample[iSamp].
1102 ** Specifically, the shortest prefix of at least (1 + iTest%nField)
1103 ** fields that is greater than the previous effective sample. */
1104 for(n=(iTest % nField) + 1; n<nField; n++){
1105 if( aSample[iSamp-1].anLt[n-1]!=aSample[iSamp].anLt[n-1] ) break;
1107 }else{
1108 n = iTest + 1;
1111 pRec->nField = n;
1112 res = sqlite3VdbeRecordCompare(aSample[iSamp].n, aSample[iSamp].p, pRec);
1113 if( res<0 ){
1114 iLower = aSample[iSamp].anLt[n-1] + aSample[iSamp].anEq[n-1];
1115 iMin = iTest+1;
1116 }else if( res==0 && n<nField ){
1117 iLower = aSample[iSamp].anLt[n-1];
1118 iMin = iTest+1;
1119 res = -1;
1120 }else{
1121 iSample = iTest;
1122 iCol = n-1;
1124 }while( res && iMin<iSample );
1125 i = iSample / nField;
1127 #ifdef SQLITE_DEBUG
1128 /* The following assert statements check that the binary search code
1129 ** above found the right answer. This block serves no purpose other
1130 ** than to invoke the asserts. */
1131 if( pParse->db->mallocFailed==0 ){
1132 if( res==0 ){
1133 /* If (res==0) is true, then pRec must be equal to sample i. */
1134 assert( i<pIdx->nSample );
1135 assert( iCol==nField-1 );
1136 pRec->nField = nField;
1137 assert( 0==sqlite3VdbeRecordCompare(aSample[i].n, aSample[i].p, pRec)
1138 || pParse->db->mallocFailed
1140 }else{
1141 /* Unless i==pIdx->nSample, indicating that pRec is larger than
1142 ** all samples in the aSample[] array, pRec must be smaller than the
1143 ** (iCol+1) field prefix of sample i. */
1144 assert( i<=pIdx->nSample && i>=0 );
1145 pRec->nField = iCol+1;
1146 assert( i==pIdx->nSample
1147 || sqlite3VdbeRecordCompare(aSample[i].n, aSample[i].p, pRec)>0
1148 || pParse->db->mallocFailed );
1150 /* if i==0 and iCol==0, then record pRec is smaller than all samples
1151 ** in the aSample[] array. Otherwise, if (iCol>0) then pRec must
1152 ** be greater than or equal to the (iCol) field prefix of sample i.
1153 ** If (i>0), then pRec must also be greater than sample (i-1). */
1154 if( iCol>0 ){
1155 pRec->nField = iCol;
1156 assert( sqlite3VdbeRecordCompare(aSample[i].n, aSample[i].p, pRec)<=0
1157 || pParse->db->mallocFailed );
1159 if( i>0 ){
1160 pRec->nField = nField;
1161 assert( sqlite3VdbeRecordCompare(aSample[i-1].n, aSample[i-1].p, pRec)<0
1162 || pParse->db->mallocFailed );
1166 #endif /* ifdef SQLITE_DEBUG */
1168 if( res==0 ){
1169 /* Record pRec is equal to sample i */
1170 assert( iCol==nField-1 );
1171 aStat[0] = aSample[i].anLt[iCol];
1172 aStat[1] = aSample[i].anEq[iCol];
1173 }else{
1174 /* At this point, the (iCol+1) field prefix of aSample[i] is the first
1175 ** sample that is greater than pRec. Or, if i==pIdx->nSample then pRec
1176 ** is larger than all samples in the array. */
1177 tRowcnt iUpper, iGap;
1178 if( i>=pIdx->nSample ){
1179 iUpper = sqlite3LogEstToInt(pIdx->aiRowLogEst[0]);
1180 }else{
1181 iUpper = aSample[i].anLt[iCol];
1184 if( iLower>=iUpper ){
1185 iGap = 0;
1186 }else{
1187 iGap = iUpper - iLower;
1189 if( roundUp ){
1190 iGap = (iGap*2)/3;
1191 }else{
1192 iGap = iGap/3;
1194 aStat[0] = iLower + iGap;
1195 aStat[1] = pIdx->aAvgEq[nField-1];
1198 /* Restore the pRec->nField value before returning. */
1199 pRec->nField = nField;
1200 return i;
1202 #endif /* SQLITE_ENABLE_STAT3_OR_STAT4 */
1205 ** If it is not NULL, pTerm is a term that provides an upper or lower
1206 ** bound on a range scan. Without considering pTerm, it is estimated
1207 ** that the scan will visit nNew rows. This function returns the number
1208 ** estimated to be visited after taking pTerm into account.
1210 ** If the user explicitly specified a likelihood() value for this term,
1211 ** then the return value is the likelihood multiplied by the number of
1212 ** input rows. Otherwise, this function assumes that an "IS NOT NULL" term
1213 ** has a likelihood of 0.50, and any other term a likelihood of 0.25.
1215 static LogEst whereRangeAdjust(WhereTerm *pTerm, LogEst nNew){
1216 LogEst nRet = nNew;
1217 if( pTerm ){
1218 if( pTerm->truthProb<=0 ){
1219 nRet += pTerm->truthProb;
1220 }else if( (pTerm->wtFlags & TERM_VNULL)==0 ){
1221 nRet -= 20; assert( 20==sqlite3LogEst(4) );
1224 return nRet;
1228 #ifdef SQLITE_ENABLE_STAT3_OR_STAT4
1230 ** Return the affinity for a single column of an index.
1232 char sqlite3IndexColumnAffinity(sqlite3 *db, Index *pIdx, int iCol){
1233 assert( iCol>=0 && iCol<pIdx->nColumn );
1234 if( !pIdx->zColAff ){
1235 if( sqlite3IndexAffinityStr(db, pIdx)==0 ) return SQLITE_AFF_BLOB;
1237 return pIdx->zColAff[iCol];
1239 #endif
1242 #ifdef SQLITE_ENABLE_STAT3_OR_STAT4
1244 ** This function is called to estimate the number of rows visited by a
1245 ** range-scan on a skip-scan index. For example:
1247 ** CREATE INDEX i1 ON t1(a, b, c);
1248 ** SELECT * FROM t1 WHERE a=? AND c BETWEEN ? AND ?;
1250 ** Value pLoop->nOut is currently set to the estimated number of rows
1251 ** visited for scanning (a=? AND b=?). This function reduces that estimate
1252 ** by some factor to account for the (c BETWEEN ? AND ?) expression based
1253 ** on the stat4 data for the index. this scan will be peformed multiple
1254 ** times (once for each (a,b) combination that matches a=?) is dealt with
1255 ** by the caller.
1257 ** It does this by scanning through all stat4 samples, comparing values
1258 ** extracted from pLower and pUpper with the corresponding column in each
1259 ** sample. If L and U are the number of samples found to be less than or
1260 ** equal to the values extracted from pLower and pUpper respectively, and
1261 ** N is the total number of samples, the pLoop->nOut value is adjusted
1262 ** as follows:
1264 ** nOut = nOut * ( min(U - L, 1) / N )
1266 ** If pLower is NULL, or a value cannot be extracted from the term, L is
1267 ** set to zero. If pUpper is NULL, or a value cannot be extracted from it,
1268 ** U is set to N.
1270 ** Normally, this function sets *pbDone to 1 before returning. However,
1271 ** if no value can be extracted from either pLower or pUpper (and so the
1272 ** estimate of the number of rows delivered remains unchanged), *pbDone
1273 ** is left as is.
1275 ** If an error occurs, an SQLite error code is returned. Otherwise,
1276 ** SQLITE_OK.
1278 static int whereRangeSkipScanEst(
1279 Parse *pParse, /* Parsing & code generating context */
1280 WhereTerm *pLower, /* Lower bound on the range. ex: "x>123" Might be NULL */
1281 WhereTerm *pUpper, /* Upper bound on the range. ex: "x<455" Might be NULL */
1282 WhereLoop *pLoop, /* Update the .nOut value of this loop */
1283 int *pbDone /* Set to true if at least one expr. value extracted */
1285 Index *p = pLoop->u.btree.pIndex;
1286 int nEq = pLoop->u.btree.nEq;
1287 sqlite3 *db = pParse->db;
1288 int nLower = -1;
1289 int nUpper = p->nSample+1;
1290 int rc = SQLITE_OK;
1291 u8 aff = sqlite3IndexColumnAffinity(db, p, nEq);
1292 CollSeq *pColl;
1294 sqlite3_value *p1 = 0; /* Value extracted from pLower */
1295 sqlite3_value *p2 = 0; /* Value extracted from pUpper */
1296 sqlite3_value *pVal = 0; /* Value extracted from record */
1298 pColl = sqlite3LocateCollSeq(pParse, p->azColl[nEq]);
1299 if( pLower ){
1300 rc = sqlite3Stat4ValueFromExpr(pParse, pLower->pExpr->pRight, aff, &p1);
1301 nLower = 0;
1303 if( pUpper && rc==SQLITE_OK ){
1304 rc = sqlite3Stat4ValueFromExpr(pParse, pUpper->pExpr->pRight, aff, &p2);
1305 nUpper = p2 ? 0 : p->nSample;
1308 if( p1 || p2 ){
1309 int i;
1310 int nDiff;
1311 for(i=0; rc==SQLITE_OK && i<p->nSample; i++){
1312 rc = sqlite3Stat4Column(db, p->aSample[i].p, p->aSample[i].n, nEq, &pVal);
1313 if( rc==SQLITE_OK && p1 ){
1314 int res = sqlite3MemCompare(p1, pVal, pColl);
1315 if( res>=0 ) nLower++;
1317 if( rc==SQLITE_OK && p2 ){
1318 int res = sqlite3MemCompare(p2, pVal, pColl);
1319 if( res>=0 ) nUpper++;
1322 nDiff = (nUpper - nLower);
1323 if( nDiff<=0 ) nDiff = 1;
1325 /* If there is both an upper and lower bound specified, and the
1326 ** comparisons indicate that they are close together, use the fallback
1327 ** method (assume that the scan visits 1/64 of the rows) for estimating
1328 ** the number of rows visited. Otherwise, estimate the number of rows
1329 ** using the method described in the header comment for this function. */
1330 if( nDiff!=1 || pUpper==0 || pLower==0 ){
1331 int nAdjust = (sqlite3LogEst(p->nSample) - sqlite3LogEst(nDiff));
1332 pLoop->nOut -= nAdjust;
1333 *pbDone = 1;
1334 WHERETRACE(0x10, ("range skip-scan regions: %u..%u adjust=%d est=%d\n",
1335 nLower, nUpper, nAdjust*-1, pLoop->nOut));
1338 }else{
1339 assert( *pbDone==0 );
1342 sqlite3ValueFree(p1);
1343 sqlite3ValueFree(p2);
1344 sqlite3ValueFree(pVal);
1346 return rc;
1348 #endif /* SQLITE_ENABLE_STAT3_OR_STAT4 */
1351 ** This function is used to estimate the number of rows that will be visited
1352 ** by scanning an index for a range of values. The range may have an upper
1353 ** bound, a lower bound, or both. The WHERE clause terms that set the upper
1354 ** and lower bounds are represented by pLower and pUpper respectively. For
1355 ** example, assuming that index p is on t1(a):
1357 ** ... FROM t1 WHERE a > ? AND a < ? ...
1358 ** |_____| |_____|
1359 ** | |
1360 ** pLower pUpper
1362 ** If either of the upper or lower bound is not present, then NULL is passed in
1363 ** place of the corresponding WhereTerm.
1365 ** The value in (pBuilder->pNew->u.btree.nEq) is the number of the index
1366 ** column subject to the range constraint. Or, equivalently, the number of
1367 ** equality constraints optimized by the proposed index scan. For example,
1368 ** assuming index p is on t1(a, b), and the SQL query is:
1370 ** ... FROM t1 WHERE a = ? AND b > ? AND b < ? ...
1372 ** then nEq is set to 1 (as the range restricted column, b, is the second
1373 ** left-most column of the index). Or, if the query is:
1375 ** ... FROM t1 WHERE a > ? AND a < ? ...
1377 ** then nEq is set to 0.
1379 ** When this function is called, *pnOut is set to the sqlite3LogEst() of the
1380 ** number of rows that the index scan is expected to visit without
1381 ** considering the range constraints. If nEq is 0, then *pnOut is the number of
1382 ** rows in the index. Assuming no error occurs, *pnOut is adjusted (reduced)
1383 ** to account for the range constraints pLower and pUpper.
1385 ** In the absence of sqlite_stat4 ANALYZE data, or if such data cannot be
1386 ** used, a single range inequality reduces the search space by a factor of 4.
1387 ** and a pair of constraints (x>? AND x<?) reduces the expected number of
1388 ** rows visited by a factor of 64.
1390 static int whereRangeScanEst(
1391 Parse *pParse, /* Parsing & code generating context */
1392 WhereLoopBuilder *pBuilder,
1393 WhereTerm *pLower, /* Lower bound on the range. ex: "x>123" Might be NULL */
1394 WhereTerm *pUpper, /* Upper bound on the range. ex: "x<455" Might be NULL */
1395 WhereLoop *pLoop /* Modify the .nOut and maybe .rRun fields */
1397 int rc = SQLITE_OK;
1398 int nOut = pLoop->nOut;
1399 LogEst nNew;
1401 #ifdef SQLITE_ENABLE_STAT3_OR_STAT4
1402 Index *p = pLoop->u.btree.pIndex;
1403 int nEq = pLoop->u.btree.nEq;
1405 if( p->nSample>0 && nEq<p->nSampleCol ){
1406 if( nEq==pBuilder->nRecValid ){
1407 UnpackedRecord *pRec = pBuilder->pRec;
1408 tRowcnt a[2];
1409 int nBtm = pLoop->u.btree.nBtm;
1410 int nTop = pLoop->u.btree.nTop;
1412 /* Variable iLower will be set to the estimate of the number of rows in
1413 ** the index that are less than the lower bound of the range query. The
1414 ** lower bound being the concatenation of $P and $L, where $P is the
1415 ** key-prefix formed by the nEq values matched against the nEq left-most
1416 ** columns of the index, and $L is the value in pLower.
1418 ** Or, if pLower is NULL or $L cannot be extracted from it (because it
1419 ** is not a simple variable or literal value), the lower bound of the
1420 ** range is $P. Due to a quirk in the way whereKeyStats() works, even
1421 ** if $L is available, whereKeyStats() is called for both ($P) and
1422 ** ($P:$L) and the larger of the two returned values is used.
1424 ** Similarly, iUpper is to be set to the estimate of the number of rows
1425 ** less than the upper bound of the range query. Where the upper bound
1426 ** is either ($P) or ($P:$U). Again, even if $U is available, both values
1427 ** of iUpper are requested of whereKeyStats() and the smaller used.
1429 ** The number of rows between the two bounds is then just iUpper-iLower.
1431 tRowcnt iLower; /* Rows less than the lower bound */
1432 tRowcnt iUpper; /* Rows less than the upper bound */
1433 int iLwrIdx = -2; /* aSample[] for the lower bound */
1434 int iUprIdx = -1; /* aSample[] for the upper bound */
1436 if( pRec ){
1437 testcase( pRec->nField!=pBuilder->nRecValid );
1438 pRec->nField = pBuilder->nRecValid;
1440 /* Determine iLower and iUpper using ($P) only. */
1441 if( nEq==0 ){
1442 iLower = 0;
1443 iUpper = p->nRowEst0;
1444 }else{
1445 /* Note: this call could be optimized away - since the same values must
1446 ** have been requested when testing key $P in whereEqualScanEst(). */
1447 whereKeyStats(pParse, p, pRec, 0, a);
1448 iLower = a[0];
1449 iUpper = a[0] + a[1];
1452 assert( pLower==0 || (pLower->eOperator & (WO_GT|WO_GE))!=0 );
1453 assert( pUpper==0 || (pUpper->eOperator & (WO_LT|WO_LE))!=0 );
1454 assert( p->aSortOrder!=0 );
1455 if( p->aSortOrder[nEq] ){
1456 /* The roles of pLower and pUpper are swapped for a DESC index */
1457 SWAP(WhereTerm*, pLower, pUpper);
1458 SWAP(int, nBtm, nTop);
1461 /* If possible, improve on the iLower estimate using ($P:$L). */
1462 if( pLower ){
1463 int n; /* Values extracted from pExpr */
1464 Expr *pExpr = pLower->pExpr->pRight;
1465 rc = sqlite3Stat4ProbeSetValue(pParse, p, &pRec, pExpr, nBtm, nEq, &n);
1466 if( rc==SQLITE_OK && n ){
1467 tRowcnt iNew;
1468 u16 mask = WO_GT|WO_LE;
1469 if( sqlite3ExprVectorSize(pExpr)>n ) mask = (WO_LE|WO_LT);
1470 iLwrIdx = whereKeyStats(pParse, p, pRec, 0, a);
1471 iNew = a[0] + ((pLower->eOperator & mask) ? a[1] : 0);
1472 if( iNew>iLower ) iLower = iNew;
1473 nOut--;
1474 pLower = 0;
1478 /* If possible, improve on the iUpper estimate using ($P:$U). */
1479 if( pUpper ){
1480 int n; /* Values extracted from pExpr */
1481 Expr *pExpr = pUpper->pExpr->pRight;
1482 rc = sqlite3Stat4ProbeSetValue(pParse, p, &pRec, pExpr, nTop, nEq, &n);
1483 if( rc==SQLITE_OK && n ){
1484 tRowcnt iNew;
1485 u16 mask = WO_GT|WO_LE;
1486 if( sqlite3ExprVectorSize(pExpr)>n ) mask = (WO_LE|WO_LT);
1487 iUprIdx = whereKeyStats(pParse, p, pRec, 1, a);
1488 iNew = a[0] + ((pUpper->eOperator & mask) ? a[1] : 0);
1489 if( iNew<iUpper ) iUpper = iNew;
1490 nOut--;
1491 pUpper = 0;
1495 pBuilder->pRec = pRec;
1496 if( rc==SQLITE_OK ){
1497 if( iUpper>iLower ){
1498 nNew = sqlite3LogEst(iUpper - iLower);
1499 /* TUNING: If both iUpper and iLower are derived from the same
1500 ** sample, then assume they are 4x more selective. This brings
1501 ** the estimated selectivity more in line with what it would be
1502 ** if estimated without the use of STAT3/4 tables. */
1503 if( iLwrIdx==iUprIdx ) nNew -= 20; assert( 20==sqlite3LogEst(4) );
1504 }else{
1505 nNew = 10; assert( 10==sqlite3LogEst(2) );
1507 if( nNew<nOut ){
1508 nOut = nNew;
1510 WHERETRACE(0x10, ("STAT4 range scan: %u..%u est=%d\n",
1511 (u32)iLower, (u32)iUpper, nOut));
1513 }else{
1514 int bDone = 0;
1515 rc = whereRangeSkipScanEst(pParse, pLower, pUpper, pLoop, &bDone);
1516 if( bDone ) return rc;
1519 #else
1520 UNUSED_PARAMETER(pParse);
1521 UNUSED_PARAMETER(pBuilder);
1522 assert( pLower || pUpper );
1523 #endif
1524 assert( pUpper==0 || (pUpper->wtFlags & TERM_VNULL)==0 );
1525 nNew = whereRangeAdjust(pLower, nOut);
1526 nNew = whereRangeAdjust(pUpper, nNew);
1528 /* TUNING: If there is both an upper and lower limit and neither limit
1529 ** has an application-defined likelihood(), assume the range is
1530 ** reduced by an additional 75%. This means that, by default, an open-ended
1531 ** range query (e.g. col > ?) is assumed to match 1/4 of the rows in the
1532 ** index. While a closed range (e.g. col BETWEEN ? AND ?) is estimated to
1533 ** match 1/64 of the index. */
1534 if( pLower && pLower->truthProb>0 && pUpper && pUpper->truthProb>0 ){
1535 nNew -= 20;
1538 nOut -= (pLower!=0) + (pUpper!=0);
1539 if( nNew<10 ) nNew = 10;
1540 if( nNew<nOut ) nOut = nNew;
1541 #if defined(WHERETRACE_ENABLED)
1542 if( pLoop->nOut>nOut ){
1543 WHERETRACE(0x10,("Range scan lowers nOut from %d to %d\n",
1544 pLoop->nOut, nOut));
1546 #endif
1547 pLoop->nOut = (LogEst)nOut;
1548 return rc;
1551 #ifdef SQLITE_ENABLE_STAT3_OR_STAT4
1553 ** Estimate the number of rows that will be returned based on
1554 ** an equality constraint x=VALUE and where that VALUE occurs in
1555 ** the histogram data. This only works when x is the left-most
1556 ** column of an index and sqlite_stat3 histogram data is available
1557 ** for that index. When pExpr==NULL that means the constraint is
1558 ** "x IS NULL" instead of "x=VALUE".
1560 ** Write the estimated row count into *pnRow and return SQLITE_OK.
1561 ** If unable to make an estimate, leave *pnRow unchanged and return
1562 ** non-zero.
1564 ** This routine can fail if it is unable to load a collating sequence
1565 ** required for string comparison, or if unable to allocate memory
1566 ** for a UTF conversion required for comparison. The error is stored
1567 ** in the pParse structure.
1569 static int whereEqualScanEst(
1570 Parse *pParse, /* Parsing & code generating context */
1571 WhereLoopBuilder *pBuilder,
1572 Expr *pExpr, /* Expression for VALUE in the x=VALUE constraint */
1573 tRowcnt *pnRow /* Write the revised row estimate here */
1575 Index *p = pBuilder->pNew->u.btree.pIndex;
1576 int nEq = pBuilder->pNew->u.btree.nEq;
1577 UnpackedRecord *pRec = pBuilder->pRec;
1578 int rc; /* Subfunction return code */
1579 tRowcnt a[2]; /* Statistics */
1580 int bOk;
1582 assert( nEq>=1 );
1583 assert( nEq<=p->nColumn );
1584 assert( p->aSample!=0 );
1585 assert( p->nSample>0 );
1586 assert( pBuilder->nRecValid<nEq );
1588 /* If values are not available for all fields of the index to the left
1589 ** of this one, no estimate can be made. Return SQLITE_NOTFOUND. */
1590 if( pBuilder->nRecValid<(nEq-1) ){
1591 return SQLITE_NOTFOUND;
1594 /* This is an optimization only. The call to sqlite3Stat4ProbeSetValue()
1595 ** below would return the same value. */
1596 if( nEq>=p->nColumn ){
1597 *pnRow = 1;
1598 return SQLITE_OK;
1601 rc = sqlite3Stat4ProbeSetValue(pParse, p, &pRec, pExpr, 1, nEq-1, &bOk);
1602 pBuilder->pRec = pRec;
1603 if( rc!=SQLITE_OK ) return rc;
1604 if( bOk==0 ) return SQLITE_NOTFOUND;
1605 pBuilder->nRecValid = nEq;
1607 whereKeyStats(pParse, p, pRec, 0, a);
1608 WHERETRACE(0x10,("equality scan regions %s(%d): %d\n",
1609 p->zName, nEq-1, (int)a[1]));
1610 *pnRow = a[1];
1612 return rc;
1614 #endif /* SQLITE_ENABLE_STAT3_OR_STAT4 */
1616 #ifdef SQLITE_ENABLE_STAT3_OR_STAT4
1618 ** Estimate the number of rows that will be returned based on
1619 ** an IN constraint where the right-hand side of the IN operator
1620 ** is a list of values. Example:
1622 ** WHERE x IN (1,2,3,4)
1624 ** Write the estimated row count into *pnRow and return SQLITE_OK.
1625 ** If unable to make an estimate, leave *pnRow unchanged and return
1626 ** non-zero.
1628 ** This routine can fail if it is unable to load a collating sequence
1629 ** required for string comparison, or if unable to allocate memory
1630 ** for a UTF conversion required for comparison. The error is stored
1631 ** in the pParse structure.
1633 static int whereInScanEst(
1634 Parse *pParse, /* Parsing & code generating context */
1635 WhereLoopBuilder *pBuilder,
1636 ExprList *pList, /* The value list on the RHS of "x IN (v1,v2,v3,...)" */
1637 tRowcnt *pnRow /* Write the revised row estimate here */
1639 Index *p = pBuilder->pNew->u.btree.pIndex;
1640 i64 nRow0 = sqlite3LogEstToInt(p->aiRowLogEst[0]);
1641 int nRecValid = pBuilder->nRecValid;
1642 int rc = SQLITE_OK; /* Subfunction return code */
1643 tRowcnt nEst; /* Number of rows for a single term */
1644 tRowcnt nRowEst = 0; /* New estimate of the number of rows */
1645 int i; /* Loop counter */
1647 assert( p->aSample!=0 );
1648 for(i=0; rc==SQLITE_OK && i<pList->nExpr; i++){
1649 nEst = nRow0;
1650 rc = whereEqualScanEst(pParse, pBuilder, pList->a[i].pExpr, &nEst);
1651 nRowEst += nEst;
1652 pBuilder->nRecValid = nRecValid;
1655 if( rc==SQLITE_OK ){
1656 if( nRowEst > nRow0 ) nRowEst = nRow0;
1657 *pnRow = nRowEst;
1658 WHERETRACE(0x10,("IN row estimate: est=%d\n", nRowEst));
1660 assert( pBuilder->nRecValid==nRecValid );
1661 return rc;
1663 #endif /* SQLITE_ENABLE_STAT3_OR_STAT4 */
1666 #ifdef WHERETRACE_ENABLED
1668 ** Print the content of a WhereTerm object
1670 static void whereTermPrint(WhereTerm *pTerm, int iTerm){
1671 if( pTerm==0 ){
1672 sqlite3DebugPrintf("TERM-%-3d NULL\n", iTerm);
1673 }else{
1674 char zType[4];
1675 char zLeft[50];
1676 memcpy(zType, "...", 4);
1677 if( pTerm->wtFlags & TERM_VIRTUAL ) zType[0] = 'V';
1678 if( pTerm->eOperator & WO_EQUIV ) zType[1] = 'E';
1679 if( ExprHasProperty(pTerm->pExpr, EP_FromJoin) ) zType[2] = 'L';
1680 if( pTerm->eOperator & WO_SINGLE ){
1681 sqlite3_snprintf(sizeof(zLeft),zLeft,"left={%d:%d}",
1682 pTerm->leftCursor, pTerm->u.leftColumn);
1683 }else if( (pTerm->eOperator & WO_OR)!=0 && pTerm->u.pOrInfo!=0 ){
1684 sqlite3_snprintf(sizeof(zLeft),zLeft,"indexable=0x%lld",
1685 pTerm->u.pOrInfo->indexable);
1686 }else{
1687 sqlite3_snprintf(sizeof(zLeft),zLeft,"left=%d", pTerm->leftCursor);
1689 sqlite3DebugPrintf(
1690 "TERM-%-3d %p %s %-12s prob=%-3d op=0x%03x wtFlags=0x%04x",
1691 iTerm, pTerm, zType, zLeft, pTerm->truthProb,
1692 pTerm->eOperator, pTerm->wtFlags);
1693 if( pTerm->iField ){
1694 sqlite3DebugPrintf(" iField=%d\n", pTerm->iField);
1695 }else{
1696 sqlite3DebugPrintf("\n");
1698 sqlite3TreeViewExpr(0, pTerm->pExpr, 0);
1701 #endif
1703 #ifdef WHERETRACE_ENABLED
1705 ** Show the complete content of a WhereClause
1707 void sqlite3WhereClausePrint(WhereClause *pWC){
1708 int i;
1709 for(i=0; i<pWC->nTerm; i++){
1710 whereTermPrint(&pWC->a[i], i);
1713 #endif
1715 #ifdef WHERETRACE_ENABLED
1717 ** Print a WhereLoop object for debugging purposes
1719 static void whereLoopPrint(WhereLoop *p, WhereClause *pWC){
1720 WhereInfo *pWInfo = pWC->pWInfo;
1721 int nb = 1+(pWInfo->pTabList->nSrc+3)/4;
1722 struct SrcList_item *pItem = pWInfo->pTabList->a + p->iTab;
1723 Table *pTab = pItem->pTab;
1724 Bitmask mAll = (((Bitmask)1)<<(nb*4)) - 1;
1725 sqlite3DebugPrintf("%c%2d.%0*llx.%0*llx", p->cId,
1726 p->iTab, nb, p->maskSelf, nb, p->prereq & mAll);
1727 sqlite3DebugPrintf(" %12s",
1728 pItem->zAlias ? pItem->zAlias : pTab->zName);
1729 if( (p->wsFlags & WHERE_VIRTUALTABLE)==0 ){
1730 const char *zName;
1731 if( p->u.btree.pIndex && (zName = p->u.btree.pIndex->zName)!=0 ){
1732 if( strncmp(zName, "sqlite_autoindex_", 17)==0 ){
1733 int i = sqlite3Strlen30(zName) - 1;
1734 while( zName[i]!='_' ) i--;
1735 zName += i;
1737 sqlite3DebugPrintf(".%-16s %2d", zName, p->u.btree.nEq);
1738 }else{
1739 sqlite3DebugPrintf("%20s","");
1741 }else{
1742 char *z;
1743 if( p->u.vtab.idxStr ){
1744 z = sqlite3_mprintf("(%d,\"%s\",%x)",
1745 p->u.vtab.idxNum, p->u.vtab.idxStr, p->u.vtab.omitMask);
1746 }else{
1747 z = sqlite3_mprintf("(%d,%x)", p->u.vtab.idxNum, p->u.vtab.omitMask);
1749 sqlite3DebugPrintf(" %-19s", z);
1750 sqlite3_free(z);
1752 if( p->wsFlags & WHERE_SKIPSCAN ){
1753 sqlite3DebugPrintf(" f %05x %d-%d", p->wsFlags, p->nLTerm,p->nSkip);
1754 }else{
1755 sqlite3DebugPrintf(" f %05x N %d", p->wsFlags, p->nLTerm);
1757 sqlite3DebugPrintf(" cost %d,%d,%d\n", p->rSetup, p->rRun, p->nOut);
1758 if( p->nLTerm && (sqlite3WhereTrace & 0x100)!=0 ){
1759 int i;
1760 for(i=0; i<p->nLTerm; i++){
1761 whereTermPrint(p->aLTerm[i], i);
1765 #endif
1768 ** Convert bulk memory into a valid WhereLoop that can be passed
1769 ** to whereLoopClear harmlessly.
1771 static void whereLoopInit(WhereLoop *p){
1772 p->aLTerm = p->aLTermSpace;
1773 p->nLTerm = 0;
1774 p->nLSlot = ArraySize(p->aLTermSpace);
1775 p->wsFlags = 0;
1779 ** Clear the WhereLoop.u union. Leave WhereLoop.pLTerm intact.
1781 static void whereLoopClearUnion(sqlite3 *db, WhereLoop *p){
1782 if( p->wsFlags & (WHERE_VIRTUALTABLE|WHERE_AUTO_INDEX) ){
1783 if( (p->wsFlags & WHERE_VIRTUALTABLE)!=0 && p->u.vtab.needFree ){
1784 sqlite3_free(p->u.vtab.idxStr);
1785 p->u.vtab.needFree = 0;
1786 p->u.vtab.idxStr = 0;
1787 }else if( (p->wsFlags & WHERE_AUTO_INDEX)!=0 && p->u.btree.pIndex!=0 ){
1788 sqlite3DbFree(db, p->u.btree.pIndex->zColAff);
1789 sqlite3DbFreeNN(db, p->u.btree.pIndex);
1790 p->u.btree.pIndex = 0;
1796 ** Deallocate internal memory used by a WhereLoop object
1798 static void whereLoopClear(sqlite3 *db, WhereLoop *p){
1799 if( p->aLTerm!=p->aLTermSpace ) sqlite3DbFreeNN(db, p->aLTerm);
1800 whereLoopClearUnion(db, p);
1801 whereLoopInit(p);
1805 ** Increase the memory allocation for pLoop->aLTerm[] to be at least n.
1807 static int whereLoopResize(sqlite3 *db, WhereLoop *p, int n){
1808 WhereTerm **paNew;
1809 if( p->nLSlot>=n ) return SQLITE_OK;
1810 n = (n+7)&~7;
1811 paNew = sqlite3DbMallocRawNN(db, sizeof(p->aLTerm[0])*n);
1812 if( paNew==0 ) return SQLITE_NOMEM_BKPT;
1813 memcpy(paNew, p->aLTerm, sizeof(p->aLTerm[0])*p->nLSlot);
1814 if( p->aLTerm!=p->aLTermSpace ) sqlite3DbFreeNN(db, p->aLTerm);
1815 p->aLTerm = paNew;
1816 p->nLSlot = n;
1817 return SQLITE_OK;
1821 ** Transfer content from the second pLoop into the first.
1823 static int whereLoopXfer(sqlite3 *db, WhereLoop *pTo, WhereLoop *pFrom){
1824 whereLoopClearUnion(db, pTo);
1825 if( whereLoopResize(db, pTo, pFrom->nLTerm) ){
1826 memset(&pTo->u, 0, sizeof(pTo->u));
1827 return SQLITE_NOMEM_BKPT;
1829 memcpy(pTo, pFrom, WHERE_LOOP_XFER_SZ);
1830 memcpy(pTo->aLTerm, pFrom->aLTerm, pTo->nLTerm*sizeof(pTo->aLTerm[0]));
1831 if( pFrom->wsFlags & WHERE_VIRTUALTABLE ){
1832 pFrom->u.vtab.needFree = 0;
1833 }else if( (pFrom->wsFlags & WHERE_AUTO_INDEX)!=0 ){
1834 pFrom->u.btree.pIndex = 0;
1836 return SQLITE_OK;
1840 ** Delete a WhereLoop object
1842 static void whereLoopDelete(sqlite3 *db, WhereLoop *p){
1843 whereLoopClear(db, p);
1844 sqlite3DbFreeNN(db, p);
1848 ** Free a WhereInfo structure
1850 static void whereInfoFree(sqlite3 *db, WhereInfo *pWInfo){
1851 if( ALWAYS(pWInfo) ){
1852 int i;
1853 for(i=0; i<pWInfo->nLevel; i++){
1854 WhereLevel *pLevel = &pWInfo->a[i];
1855 if( pLevel->pWLoop && (pLevel->pWLoop->wsFlags & WHERE_IN_ABLE) ){
1856 sqlite3DbFree(db, pLevel->u.in.aInLoop);
1859 sqlite3WhereClauseClear(&pWInfo->sWC);
1860 while( pWInfo->pLoops ){
1861 WhereLoop *p = pWInfo->pLoops;
1862 pWInfo->pLoops = p->pNextLoop;
1863 whereLoopDelete(db, p);
1865 sqlite3DbFreeNN(db, pWInfo);
1870 ** Return TRUE if all of the following are true:
1872 ** (1) X has the same or lower cost that Y
1873 ** (2) X is a proper subset of Y
1874 ** (3) X skips at least as many columns as Y
1876 ** By "proper subset" we mean that X uses fewer WHERE clause terms
1877 ** than Y and that every WHERE clause term used by X is also used
1878 ** by Y.
1880 ** If X is a proper subset of Y then Y is a better choice and ought
1881 ** to have a lower cost. This routine returns TRUE when that cost
1882 ** relationship is inverted and needs to be adjusted. The third rule
1883 ** was added because if X uses skip-scan less than Y it still might
1884 ** deserve a lower cost even if it is a proper subset of Y.
1886 static int whereLoopCheaperProperSubset(
1887 const WhereLoop *pX, /* First WhereLoop to compare */
1888 const WhereLoop *pY /* Compare against this WhereLoop */
1890 int i, j;
1891 if( pX->nLTerm-pX->nSkip >= pY->nLTerm-pY->nSkip ){
1892 return 0; /* X is not a subset of Y */
1894 if( pY->nSkip > pX->nSkip ) return 0;
1895 if( pX->rRun >= pY->rRun ){
1896 if( pX->rRun > pY->rRun ) return 0; /* X costs more than Y */
1897 if( pX->nOut > pY->nOut ) return 0; /* X costs more than Y */
1899 for(i=pX->nLTerm-1; i>=0; i--){
1900 if( pX->aLTerm[i]==0 ) continue;
1901 for(j=pY->nLTerm-1; j>=0; j--){
1902 if( pY->aLTerm[j]==pX->aLTerm[i] ) break;
1904 if( j<0 ) return 0; /* X not a subset of Y since term X[i] not used by Y */
1906 return 1; /* All conditions meet */
1910 ** Try to adjust the cost of WhereLoop pTemplate upwards or downwards so
1911 ** that:
1913 ** (1) pTemplate costs less than any other WhereLoops that are a proper
1914 ** subset of pTemplate
1916 ** (2) pTemplate costs more than any other WhereLoops for which pTemplate
1917 ** is a proper subset.
1919 ** To say "WhereLoop X is a proper subset of Y" means that X uses fewer
1920 ** WHERE clause terms than Y and that every WHERE clause term used by X is
1921 ** also used by Y.
1923 static void whereLoopAdjustCost(const WhereLoop *p, WhereLoop *pTemplate){
1924 if( (pTemplate->wsFlags & WHERE_INDEXED)==0 ) return;
1925 for(; p; p=p->pNextLoop){
1926 if( p->iTab!=pTemplate->iTab ) continue;
1927 if( (p->wsFlags & WHERE_INDEXED)==0 ) continue;
1928 if( whereLoopCheaperProperSubset(p, pTemplate) ){
1929 /* Adjust pTemplate cost downward so that it is cheaper than its
1930 ** subset p. */
1931 WHERETRACE(0x80,("subset cost adjustment %d,%d to %d,%d\n",
1932 pTemplate->rRun, pTemplate->nOut, p->rRun, p->nOut-1));
1933 pTemplate->rRun = p->rRun;
1934 pTemplate->nOut = p->nOut - 1;
1935 }else if( whereLoopCheaperProperSubset(pTemplate, p) ){
1936 /* Adjust pTemplate cost upward so that it is costlier than p since
1937 ** pTemplate is a proper subset of p */
1938 WHERETRACE(0x80,("subset cost adjustment %d,%d to %d,%d\n",
1939 pTemplate->rRun, pTemplate->nOut, p->rRun, p->nOut+1));
1940 pTemplate->rRun = p->rRun;
1941 pTemplate->nOut = p->nOut + 1;
1947 ** Search the list of WhereLoops in *ppPrev looking for one that can be
1948 ** replaced by pTemplate.
1950 ** Return NULL if pTemplate does not belong on the WhereLoop list.
1951 ** In other words if pTemplate ought to be dropped from further consideration.
1953 ** If pX is a WhereLoop that pTemplate can replace, then return the
1954 ** link that points to pX.
1956 ** If pTemplate cannot replace any existing element of the list but needs
1957 ** to be added to the list as a new entry, then return a pointer to the
1958 ** tail of the list.
1960 static WhereLoop **whereLoopFindLesser(
1961 WhereLoop **ppPrev,
1962 const WhereLoop *pTemplate
1964 WhereLoop *p;
1965 for(p=(*ppPrev); p; ppPrev=&p->pNextLoop, p=*ppPrev){
1966 if( p->iTab!=pTemplate->iTab || p->iSortIdx!=pTemplate->iSortIdx ){
1967 /* If either the iTab or iSortIdx values for two WhereLoop are different
1968 ** then those WhereLoops need to be considered separately. Neither is
1969 ** a candidate to replace the other. */
1970 continue;
1972 /* In the current implementation, the rSetup value is either zero
1973 ** or the cost of building an automatic index (NlogN) and the NlogN
1974 ** is the same for compatible WhereLoops. */
1975 assert( p->rSetup==0 || pTemplate->rSetup==0
1976 || p->rSetup==pTemplate->rSetup );
1978 /* whereLoopAddBtree() always generates and inserts the automatic index
1979 ** case first. Hence compatible candidate WhereLoops never have a larger
1980 ** rSetup. Call this SETUP-INVARIANT */
1981 assert( p->rSetup>=pTemplate->rSetup );
1983 /* Any loop using an appliation-defined index (or PRIMARY KEY or
1984 ** UNIQUE constraint) with one or more == constraints is better
1985 ** than an automatic index. Unless it is a skip-scan. */
1986 if( (p->wsFlags & WHERE_AUTO_INDEX)!=0
1987 && (pTemplate->nSkip)==0
1988 && (pTemplate->wsFlags & WHERE_INDEXED)!=0
1989 && (pTemplate->wsFlags & WHERE_COLUMN_EQ)!=0
1990 && (p->prereq & pTemplate->prereq)==pTemplate->prereq
1992 break;
1995 /* If existing WhereLoop p is better than pTemplate, pTemplate can be
1996 ** discarded. WhereLoop p is better if:
1997 ** (1) p has no more dependencies than pTemplate, and
1998 ** (2) p has an equal or lower cost than pTemplate
2000 if( (p->prereq & pTemplate->prereq)==p->prereq /* (1) */
2001 && p->rSetup<=pTemplate->rSetup /* (2a) */
2002 && p->rRun<=pTemplate->rRun /* (2b) */
2003 && p->nOut<=pTemplate->nOut /* (2c) */
2005 return 0; /* Discard pTemplate */
2008 /* If pTemplate is always better than p, then cause p to be overwritten
2009 ** with pTemplate. pTemplate is better than p if:
2010 ** (1) pTemplate has no more dependences than p, and
2011 ** (2) pTemplate has an equal or lower cost than p.
2013 if( (p->prereq & pTemplate->prereq)==pTemplate->prereq /* (1) */
2014 && p->rRun>=pTemplate->rRun /* (2a) */
2015 && p->nOut>=pTemplate->nOut /* (2b) */
2017 assert( p->rSetup>=pTemplate->rSetup ); /* SETUP-INVARIANT above */
2018 break; /* Cause p to be overwritten by pTemplate */
2021 return ppPrev;
2025 ** Insert or replace a WhereLoop entry using the template supplied.
2027 ** An existing WhereLoop entry might be overwritten if the new template
2028 ** is better and has fewer dependencies. Or the template will be ignored
2029 ** and no insert will occur if an existing WhereLoop is faster and has
2030 ** fewer dependencies than the template. Otherwise a new WhereLoop is
2031 ** added based on the template.
2033 ** If pBuilder->pOrSet is not NULL then we care about only the
2034 ** prerequisites and rRun and nOut costs of the N best loops. That
2035 ** information is gathered in the pBuilder->pOrSet object. This special
2036 ** processing mode is used only for OR clause processing.
2038 ** When accumulating multiple loops (when pBuilder->pOrSet is NULL) we
2039 ** still might overwrite similar loops with the new template if the
2040 ** new template is better. Loops may be overwritten if the following
2041 ** conditions are met:
2043 ** (1) They have the same iTab.
2044 ** (2) They have the same iSortIdx.
2045 ** (3) The template has same or fewer dependencies than the current loop
2046 ** (4) The template has the same or lower cost than the current loop
2048 static int whereLoopInsert(WhereLoopBuilder *pBuilder, WhereLoop *pTemplate){
2049 WhereLoop **ppPrev, *p;
2050 WhereInfo *pWInfo = pBuilder->pWInfo;
2051 sqlite3 *db = pWInfo->pParse->db;
2052 int rc;
2054 /* If pBuilder->pOrSet is defined, then only keep track of the costs
2055 ** and prereqs.
2057 if( pBuilder->pOrSet!=0 ){
2058 if( pTemplate->nLTerm ){
2059 #if WHERETRACE_ENABLED
2060 u16 n = pBuilder->pOrSet->n;
2061 int x =
2062 #endif
2063 whereOrInsert(pBuilder->pOrSet, pTemplate->prereq, pTemplate->rRun,
2064 pTemplate->nOut);
2065 #if WHERETRACE_ENABLED /* 0x8 */
2066 if( sqlite3WhereTrace & 0x8 ){
2067 sqlite3DebugPrintf(x?" or-%d: ":" or-X: ", n);
2068 whereLoopPrint(pTemplate, pBuilder->pWC);
2070 #endif
2072 return SQLITE_OK;
2075 /* Look for an existing WhereLoop to replace with pTemplate
2077 whereLoopAdjustCost(pWInfo->pLoops, pTemplate);
2078 ppPrev = whereLoopFindLesser(&pWInfo->pLoops, pTemplate);
2080 if( ppPrev==0 ){
2081 /* There already exists a WhereLoop on the list that is better
2082 ** than pTemplate, so just ignore pTemplate */
2083 #if WHERETRACE_ENABLED /* 0x8 */
2084 if( sqlite3WhereTrace & 0x8 ){
2085 sqlite3DebugPrintf(" skip: ");
2086 whereLoopPrint(pTemplate, pBuilder->pWC);
2088 #endif
2089 return SQLITE_OK;
2090 }else{
2091 p = *ppPrev;
2094 /* If we reach this point it means that either p[] should be overwritten
2095 ** with pTemplate[] if p[] exists, or if p==NULL then allocate a new
2096 ** WhereLoop and insert it.
2098 #if WHERETRACE_ENABLED /* 0x8 */
2099 if( sqlite3WhereTrace & 0x8 ){
2100 if( p!=0 ){
2101 sqlite3DebugPrintf("replace: ");
2102 whereLoopPrint(p, pBuilder->pWC);
2103 sqlite3DebugPrintf(" with: ");
2104 }else{
2105 sqlite3DebugPrintf(" add: ");
2107 whereLoopPrint(pTemplate, pBuilder->pWC);
2109 #endif
2110 if( p==0 ){
2111 /* Allocate a new WhereLoop to add to the end of the list */
2112 *ppPrev = p = sqlite3DbMallocRawNN(db, sizeof(WhereLoop));
2113 if( p==0 ) return SQLITE_NOMEM_BKPT;
2114 whereLoopInit(p);
2115 p->pNextLoop = 0;
2116 }else{
2117 /* We will be overwriting WhereLoop p[]. But before we do, first
2118 ** go through the rest of the list and delete any other entries besides
2119 ** p[] that are also supplated by pTemplate */
2120 WhereLoop **ppTail = &p->pNextLoop;
2121 WhereLoop *pToDel;
2122 while( *ppTail ){
2123 ppTail = whereLoopFindLesser(ppTail, pTemplate);
2124 if( ppTail==0 ) break;
2125 pToDel = *ppTail;
2126 if( pToDel==0 ) break;
2127 *ppTail = pToDel->pNextLoop;
2128 #if WHERETRACE_ENABLED /* 0x8 */
2129 if( sqlite3WhereTrace & 0x8 ){
2130 sqlite3DebugPrintf(" delete: ");
2131 whereLoopPrint(pToDel, pBuilder->pWC);
2133 #endif
2134 whereLoopDelete(db, pToDel);
2137 rc = whereLoopXfer(db, p, pTemplate);
2138 if( (p->wsFlags & WHERE_VIRTUALTABLE)==0 ){
2139 Index *pIndex = p->u.btree.pIndex;
2140 if( pIndex && pIndex->tnum==0 ){
2141 p->u.btree.pIndex = 0;
2144 return rc;
2148 ** Adjust the WhereLoop.nOut value downward to account for terms of the
2149 ** WHERE clause that reference the loop but which are not used by an
2150 ** index.
2152 ** For every WHERE clause term that is not used by the index
2153 ** and which has a truth probability assigned by one of the likelihood(),
2154 ** likely(), or unlikely() SQL functions, reduce the estimated number
2155 ** of output rows by the probability specified.
2157 ** TUNING: For every WHERE clause term that is not used by the index
2158 ** and which does not have an assigned truth probability, heuristics
2159 ** described below are used to try to estimate the truth probability.
2160 ** TODO --> Perhaps this is something that could be improved by better
2161 ** table statistics.
2163 ** Heuristic 1: Estimate the truth probability as 93.75%. The 93.75%
2164 ** value corresponds to -1 in LogEst notation, so this means decrement
2165 ** the WhereLoop.nOut field for every such WHERE clause term.
2167 ** Heuristic 2: If there exists one or more WHERE clause terms of the
2168 ** form "x==EXPR" and EXPR is not a constant 0 or 1, then make sure the
2169 ** final output row estimate is no greater than 1/4 of the total number
2170 ** of rows in the table. In other words, assume that x==EXPR will filter
2171 ** out at least 3 out of 4 rows. If EXPR is -1 or 0 or 1, then maybe the
2172 ** "x" column is boolean or else -1 or 0 or 1 is a common default value
2173 ** on the "x" column and so in that case only cap the output row estimate
2174 ** at 1/2 instead of 1/4.
2176 static void whereLoopOutputAdjust(
2177 WhereClause *pWC, /* The WHERE clause */
2178 WhereLoop *pLoop, /* The loop to adjust downward */
2179 LogEst nRow /* Number of rows in the entire table */
2181 WhereTerm *pTerm, *pX;
2182 Bitmask notAllowed = ~(pLoop->prereq|pLoop->maskSelf);
2183 int i, j, k;
2184 LogEst iReduce = 0; /* pLoop->nOut should not exceed nRow-iReduce */
2186 assert( (pLoop->wsFlags & WHERE_AUTO_INDEX)==0 );
2187 for(i=pWC->nTerm, pTerm=pWC->a; i>0; i--, pTerm++){
2188 if( (pTerm->wtFlags & TERM_VIRTUAL)!=0 ) break;
2189 if( (pTerm->prereqAll & pLoop->maskSelf)==0 ) continue;
2190 if( (pTerm->prereqAll & notAllowed)!=0 ) continue;
2191 for(j=pLoop->nLTerm-1; j>=0; j--){
2192 pX = pLoop->aLTerm[j];
2193 if( pX==0 ) continue;
2194 if( pX==pTerm ) break;
2195 if( pX->iParent>=0 && (&pWC->a[pX->iParent])==pTerm ) break;
2197 if( j<0 ){
2198 if( pTerm->truthProb<=0 ){
2199 /* If a truth probability is specified using the likelihood() hints,
2200 ** then use the probability provided by the application. */
2201 pLoop->nOut += pTerm->truthProb;
2202 }else{
2203 /* In the absence of explicit truth probabilities, use heuristics to
2204 ** guess a reasonable truth probability. */
2205 pLoop->nOut--;
2206 if( pTerm->eOperator&(WO_EQ|WO_IS) ){
2207 Expr *pRight = pTerm->pExpr->pRight;
2208 testcase( pTerm->pExpr->op==TK_IS );
2209 if( sqlite3ExprIsInteger(pRight, &k) && k>=(-1) && k<=1 ){
2210 k = 10;
2211 }else{
2212 k = 20;
2214 if( iReduce<k ) iReduce = k;
2219 if( pLoop->nOut > nRow-iReduce ) pLoop->nOut = nRow - iReduce;
2223 ** Term pTerm is a vector range comparison operation. The first comparison
2224 ** in the vector can be optimized using column nEq of the index. This
2225 ** function returns the total number of vector elements that can be used
2226 ** as part of the range comparison.
2228 ** For example, if the query is:
2230 ** WHERE a = ? AND (b, c, d) > (?, ?, ?)
2232 ** and the index:
2234 ** CREATE INDEX ... ON (a, b, c, d, e)
2236 ** then this function would be invoked with nEq=1. The value returned in
2237 ** this case is 3.
2239 static int whereRangeVectorLen(
2240 Parse *pParse, /* Parsing context */
2241 int iCur, /* Cursor open on pIdx */
2242 Index *pIdx, /* The index to be used for a inequality constraint */
2243 int nEq, /* Number of prior equality constraints on same index */
2244 WhereTerm *pTerm /* The vector inequality constraint */
2246 int nCmp = sqlite3ExprVectorSize(pTerm->pExpr->pLeft);
2247 int i;
2249 nCmp = MIN(nCmp, (pIdx->nColumn - nEq));
2250 for(i=1; i<nCmp; i++){
2251 /* Test if comparison i of pTerm is compatible with column (i+nEq)
2252 ** of the index. If not, exit the loop. */
2253 char aff; /* Comparison affinity */
2254 char idxaff = 0; /* Indexed columns affinity */
2255 CollSeq *pColl; /* Comparison collation sequence */
2256 Expr *pLhs = pTerm->pExpr->pLeft->x.pList->a[i].pExpr;
2257 Expr *pRhs = pTerm->pExpr->pRight;
2258 if( pRhs->flags & EP_xIsSelect ){
2259 pRhs = pRhs->x.pSelect->pEList->a[i].pExpr;
2260 }else{
2261 pRhs = pRhs->x.pList->a[i].pExpr;
2264 /* Check that the LHS of the comparison is a column reference to
2265 ** the right column of the right source table. And that the sort
2266 ** order of the index column is the same as the sort order of the
2267 ** leftmost index column. */
2268 if( pLhs->op!=TK_COLUMN
2269 || pLhs->iTable!=iCur
2270 || pLhs->iColumn!=pIdx->aiColumn[i+nEq]
2271 || pIdx->aSortOrder[i+nEq]!=pIdx->aSortOrder[nEq]
2273 break;
2276 testcase( pLhs->iColumn==XN_ROWID );
2277 aff = sqlite3CompareAffinity(pRhs, sqlite3ExprAffinity(pLhs));
2278 idxaff = sqlite3TableColumnAffinity(pIdx->pTable, pLhs->iColumn);
2279 if( aff!=idxaff ) break;
2281 pColl = sqlite3BinaryCompareCollSeq(pParse, pLhs, pRhs);
2282 if( pColl==0 ) break;
2283 if( sqlite3StrICmp(pColl->zName, pIdx->azColl[i+nEq]) ) break;
2285 return i;
2289 ** Adjust the cost C by the costMult facter T. This only occurs if
2290 ** compiled with -DSQLITE_ENABLE_COSTMULT
2292 #ifdef SQLITE_ENABLE_COSTMULT
2293 # define ApplyCostMultiplier(C,T) C += T
2294 #else
2295 # define ApplyCostMultiplier(C,T)
2296 #endif
2299 ** We have so far matched pBuilder->pNew->u.btree.nEq terms of the
2300 ** index pIndex. Try to match one more.
2302 ** When this function is called, pBuilder->pNew->nOut contains the
2303 ** number of rows expected to be visited by filtering using the nEq
2304 ** terms only. If it is modified, this value is restored before this
2305 ** function returns.
2307 ** If pProbe->tnum==0, that means pIndex is a fake index used for the
2308 ** INTEGER PRIMARY KEY.
2310 static int whereLoopAddBtreeIndex(
2311 WhereLoopBuilder *pBuilder, /* The WhereLoop factory */
2312 struct SrcList_item *pSrc, /* FROM clause term being analyzed */
2313 Index *pProbe, /* An index on pSrc */
2314 LogEst nInMul /* log(Number of iterations due to IN) */
2316 WhereInfo *pWInfo = pBuilder->pWInfo; /* WHERE analyse context */
2317 Parse *pParse = pWInfo->pParse; /* Parsing context */
2318 sqlite3 *db = pParse->db; /* Database connection malloc context */
2319 WhereLoop *pNew; /* Template WhereLoop under construction */
2320 WhereTerm *pTerm; /* A WhereTerm under consideration */
2321 int opMask; /* Valid operators for constraints */
2322 WhereScan scan; /* Iterator for WHERE terms */
2323 Bitmask saved_prereq; /* Original value of pNew->prereq */
2324 u16 saved_nLTerm; /* Original value of pNew->nLTerm */
2325 u16 saved_nEq; /* Original value of pNew->u.btree.nEq */
2326 u16 saved_nBtm; /* Original value of pNew->u.btree.nBtm */
2327 u16 saved_nTop; /* Original value of pNew->u.btree.nTop */
2328 u16 saved_nSkip; /* Original value of pNew->nSkip */
2329 u32 saved_wsFlags; /* Original value of pNew->wsFlags */
2330 LogEst saved_nOut; /* Original value of pNew->nOut */
2331 int rc = SQLITE_OK; /* Return code */
2332 LogEst rSize; /* Number of rows in the table */
2333 LogEst rLogSize; /* Logarithm of table size */
2334 WhereTerm *pTop = 0, *pBtm = 0; /* Top and bottom range constraints */
2336 pNew = pBuilder->pNew;
2337 if( db->mallocFailed ) return SQLITE_NOMEM_BKPT;
2338 WHERETRACE(0x800, ("BEGIN addBtreeIdx(%s), nEq=%d\n",
2339 pProbe->zName, pNew->u.btree.nEq));
2341 assert( (pNew->wsFlags & WHERE_VIRTUALTABLE)==0 );
2342 assert( (pNew->wsFlags & WHERE_TOP_LIMIT)==0 );
2343 if( pNew->wsFlags & WHERE_BTM_LIMIT ){
2344 opMask = WO_LT|WO_LE;
2345 }else{
2346 assert( pNew->u.btree.nBtm==0 );
2347 opMask = WO_EQ|WO_IN|WO_GT|WO_GE|WO_LT|WO_LE|WO_ISNULL|WO_IS;
2349 if( pProbe->bUnordered ) opMask &= ~(WO_GT|WO_GE|WO_LT|WO_LE);
2351 assert( pNew->u.btree.nEq<pProbe->nColumn );
2353 saved_nEq = pNew->u.btree.nEq;
2354 saved_nBtm = pNew->u.btree.nBtm;
2355 saved_nTop = pNew->u.btree.nTop;
2356 saved_nSkip = pNew->nSkip;
2357 saved_nLTerm = pNew->nLTerm;
2358 saved_wsFlags = pNew->wsFlags;
2359 saved_prereq = pNew->prereq;
2360 saved_nOut = pNew->nOut;
2361 pTerm = whereScanInit(&scan, pBuilder->pWC, pSrc->iCursor, saved_nEq,
2362 opMask, pProbe);
2363 pNew->rSetup = 0;
2364 rSize = pProbe->aiRowLogEst[0];
2365 rLogSize = estLog(rSize);
2366 for(; rc==SQLITE_OK && pTerm!=0; pTerm = whereScanNext(&scan)){
2367 u16 eOp = pTerm->eOperator; /* Shorthand for pTerm->eOperator */
2368 LogEst rCostIdx;
2369 LogEst nOutUnadjusted; /* nOut before IN() and WHERE adjustments */
2370 int nIn = 0;
2371 #ifdef SQLITE_ENABLE_STAT3_OR_STAT4
2372 int nRecValid = pBuilder->nRecValid;
2373 #endif
2374 if( (eOp==WO_ISNULL || (pTerm->wtFlags&TERM_VNULL)!=0)
2375 && indexColumnNotNull(pProbe, saved_nEq)
2377 continue; /* ignore IS [NOT] NULL constraints on NOT NULL columns */
2379 if( pTerm->prereqRight & pNew->maskSelf ) continue;
2381 /* Do not allow the upper bound of a LIKE optimization range constraint
2382 ** to mix with a lower range bound from some other source */
2383 if( pTerm->wtFlags & TERM_LIKEOPT && pTerm->eOperator==WO_LT ) continue;
2385 /* Do not allow IS constraints from the WHERE clause to be used by the
2386 ** right table of a LEFT JOIN. Only constraints in the ON clause are
2387 ** allowed */
2388 if( (pSrc->fg.jointype & JT_LEFT)!=0
2389 && !ExprHasProperty(pTerm->pExpr, EP_FromJoin)
2390 && (eOp & (WO_IS|WO_ISNULL))!=0
2392 testcase( eOp & WO_IS );
2393 testcase( eOp & WO_ISNULL );
2394 continue;
2397 if( IsUniqueIndex(pProbe) && saved_nEq==pProbe->nKeyCol-1 ){
2398 pBuilder->bldFlags |= SQLITE_BLDF_UNIQUE;
2399 }else{
2400 pBuilder->bldFlags |= SQLITE_BLDF_INDEXED;
2402 pNew->wsFlags = saved_wsFlags;
2403 pNew->u.btree.nEq = saved_nEq;
2404 pNew->u.btree.nBtm = saved_nBtm;
2405 pNew->u.btree.nTop = saved_nTop;
2406 pNew->nLTerm = saved_nLTerm;
2407 if( whereLoopResize(db, pNew, pNew->nLTerm+1) ) break; /* OOM */
2408 pNew->aLTerm[pNew->nLTerm++] = pTerm;
2409 pNew->prereq = (saved_prereq | pTerm->prereqRight) & ~pNew->maskSelf;
2411 assert( nInMul==0
2412 || (pNew->wsFlags & WHERE_COLUMN_NULL)!=0
2413 || (pNew->wsFlags & WHERE_COLUMN_IN)!=0
2414 || (pNew->wsFlags & WHERE_SKIPSCAN)!=0
2417 if( eOp & WO_IN ){
2418 Expr *pExpr = pTerm->pExpr;
2419 pNew->wsFlags |= WHERE_COLUMN_IN;
2420 if( ExprHasProperty(pExpr, EP_xIsSelect) ){
2421 /* "x IN (SELECT ...)": TUNING: the SELECT returns 25 rows */
2422 int i;
2423 nIn = 46; assert( 46==sqlite3LogEst(25) );
2425 /* The expression may actually be of the form (x, y) IN (SELECT...).
2426 ** In this case there is a separate term for each of (x) and (y).
2427 ** However, the nIn multiplier should only be applied once, not once
2428 ** for each such term. The following loop checks that pTerm is the
2429 ** first such term in use, and sets nIn back to 0 if it is not. */
2430 for(i=0; i<pNew->nLTerm-1; i++){
2431 if( pNew->aLTerm[i] && pNew->aLTerm[i]->pExpr==pExpr ) nIn = 0;
2433 }else if( ALWAYS(pExpr->x.pList && pExpr->x.pList->nExpr) ){
2434 /* "x IN (value, value, ...)" */
2435 nIn = sqlite3LogEst(pExpr->x.pList->nExpr);
2436 assert( nIn>0 ); /* RHS always has 2 or more terms... The parser
2437 ** changes "x IN (?)" into "x=?". */
2439 }else if( eOp & (WO_EQ|WO_IS) ){
2440 int iCol = pProbe->aiColumn[saved_nEq];
2441 pNew->wsFlags |= WHERE_COLUMN_EQ;
2442 assert( saved_nEq==pNew->u.btree.nEq );
2443 if( iCol==XN_ROWID
2444 || (iCol>0 && nInMul==0 && saved_nEq==pProbe->nKeyCol-1)
2446 if( iCol>=0 && pProbe->uniqNotNull==0 ){
2447 pNew->wsFlags |= WHERE_UNQ_WANTED;
2448 }else{
2449 pNew->wsFlags |= WHERE_ONEROW;
2452 }else if( eOp & WO_ISNULL ){
2453 pNew->wsFlags |= WHERE_COLUMN_NULL;
2454 }else if( eOp & (WO_GT|WO_GE) ){
2455 testcase( eOp & WO_GT );
2456 testcase( eOp & WO_GE );
2457 pNew->wsFlags |= WHERE_COLUMN_RANGE|WHERE_BTM_LIMIT;
2458 pNew->u.btree.nBtm = whereRangeVectorLen(
2459 pParse, pSrc->iCursor, pProbe, saved_nEq, pTerm
2461 pBtm = pTerm;
2462 pTop = 0;
2463 if( pTerm->wtFlags & TERM_LIKEOPT ){
2464 /* Range contraints that come from the LIKE optimization are
2465 ** always used in pairs. */
2466 pTop = &pTerm[1];
2467 assert( (pTop-(pTerm->pWC->a))<pTerm->pWC->nTerm );
2468 assert( pTop->wtFlags & TERM_LIKEOPT );
2469 assert( pTop->eOperator==WO_LT );
2470 if( whereLoopResize(db, pNew, pNew->nLTerm+1) ) break; /* OOM */
2471 pNew->aLTerm[pNew->nLTerm++] = pTop;
2472 pNew->wsFlags |= WHERE_TOP_LIMIT;
2473 pNew->u.btree.nTop = 1;
2475 }else{
2476 assert( eOp & (WO_LT|WO_LE) );
2477 testcase( eOp & WO_LT );
2478 testcase( eOp & WO_LE );
2479 pNew->wsFlags |= WHERE_COLUMN_RANGE|WHERE_TOP_LIMIT;
2480 pNew->u.btree.nTop = whereRangeVectorLen(
2481 pParse, pSrc->iCursor, pProbe, saved_nEq, pTerm
2483 pTop = pTerm;
2484 pBtm = (pNew->wsFlags & WHERE_BTM_LIMIT)!=0 ?
2485 pNew->aLTerm[pNew->nLTerm-2] : 0;
2488 /* At this point pNew->nOut is set to the number of rows expected to
2489 ** be visited by the index scan before considering term pTerm, or the
2490 ** values of nIn and nInMul. In other words, assuming that all
2491 ** "x IN(...)" terms are replaced with "x = ?". This block updates
2492 ** the value of pNew->nOut to account for pTerm (but not nIn/nInMul). */
2493 assert( pNew->nOut==saved_nOut );
2494 if( pNew->wsFlags & WHERE_COLUMN_RANGE ){
2495 /* Adjust nOut using stat3/stat4 data. Or, if there is no stat3/stat4
2496 ** data, using some other estimate. */
2497 whereRangeScanEst(pParse, pBuilder, pBtm, pTop, pNew);
2498 }else{
2499 int nEq = ++pNew->u.btree.nEq;
2500 assert( eOp & (WO_ISNULL|WO_EQ|WO_IN|WO_IS) );
2502 assert( pNew->nOut==saved_nOut );
2503 if( pTerm->truthProb<=0 && pProbe->aiColumn[saved_nEq]>=0 ){
2504 assert( (eOp & WO_IN) || nIn==0 );
2505 testcase( eOp & WO_IN );
2506 pNew->nOut += pTerm->truthProb;
2507 pNew->nOut -= nIn;
2508 }else{
2509 #ifdef SQLITE_ENABLE_STAT3_OR_STAT4
2510 tRowcnt nOut = 0;
2511 if( nInMul==0
2512 && pProbe->nSample
2513 && pNew->u.btree.nEq<=pProbe->nSampleCol
2514 && ((eOp & WO_IN)==0 || !ExprHasProperty(pTerm->pExpr, EP_xIsSelect))
2516 Expr *pExpr = pTerm->pExpr;
2517 if( (eOp & (WO_EQ|WO_ISNULL|WO_IS))!=0 ){
2518 testcase( eOp & WO_EQ );
2519 testcase( eOp & WO_IS );
2520 testcase( eOp & WO_ISNULL );
2521 rc = whereEqualScanEst(pParse, pBuilder, pExpr->pRight, &nOut);
2522 }else{
2523 rc = whereInScanEst(pParse, pBuilder, pExpr->x.pList, &nOut);
2525 if( rc==SQLITE_NOTFOUND ) rc = SQLITE_OK;
2526 if( rc!=SQLITE_OK ) break; /* Jump out of the pTerm loop */
2527 if( nOut ){
2528 pNew->nOut = sqlite3LogEst(nOut);
2529 if( pNew->nOut>saved_nOut ) pNew->nOut = saved_nOut;
2530 pNew->nOut -= nIn;
2533 if( nOut==0 )
2534 #endif
2536 pNew->nOut += (pProbe->aiRowLogEst[nEq] - pProbe->aiRowLogEst[nEq-1]);
2537 if( eOp & WO_ISNULL ){
2538 /* TUNING: If there is no likelihood() value, assume that a
2539 ** "col IS NULL" expression matches twice as many rows
2540 ** as (col=?). */
2541 pNew->nOut += 10;
2547 /* Set rCostIdx to the cost of visiting selected rows in index. Add
2548 ** it to pNew->rRun, which is currently set to the cost of the index
2549 ** seek only. Then, if this is a non-covering index, add the cost of
2550 ** visiting the rows in the main table. */
2551 rCostIdx = pNew->nOut + 1 + (15*pProbe->szIdxRow)/pSrc->pTab->szTabRow;
2552 pNew->rRun = sqlite3LogEstAdd(rLogSize, rCostIdx);
2553 if( (pNew->wsFlags & (WHERE_IDX_ONLY|WHERE_IPK))==0 ){
2554 pNew->rRun = sqlite3LogEstAdd(pNew->rRun, pNew->nOut + 16);
2556 ApplyCostMultiplier(pNew->rRun, pProbe->pTable->costMult);
2558 nOutUnadjusted = pNew->nOut;
2559 pNew->rRun += nInMul + nIn;
2560 pNew->nOut += nInMul + nIn;
2561 whereLoopOutputAdjust(pBuilder->pWC, pNew, rSize);
2562 rc = whereLoopInsert(pBuilder, pNew);
2564 if( pNew->wsFlags & WHERE_COLUMN_RANGE ){
2565 pNew->nOut = saved_nOut;
2566 }else{
2567 pNew->nOut = nOutUnadjusted;
2570 if( (pNew->wsFlags & WHERE_TOP_LIMIT)==0
2571 && pNew->u.btree.nEq<pProbe->nColumn
2573 whereLoopAddBtreeIndex(pBuilder, pSrc, pProbe, nInMul+nIn);
2575 pNew->nOut = saved_nOut;
2576 #ifdef SQLITE_ENABLE_STAT3_OR_STAT4
2577 pBuilder->nRecValid = nRecValid;
2578 #endif
2580 pNew->prereq = saved_prereq;
2581 pNew->u.btree.nEq = saved_nEq;
2582 pNew->u.btree.nBtm = saved_nBtm;
2583 pNew->u.btree.nTop = saved_nTop;
2584 pNew->nSkip = saved_nSkip;
2585 pNew->wsFlags = saved_wsFlags;
2586 pNew->nOut = saved_nOut;
2587 pNew->nLTerm = saved_nLTerm;
2589 /* Consider using a skip-scan if there are no WHERE clause constraints
2590 ** available for the left-most terms of the index, and if the average
2591 ** number of repeats in the left-most terms is at least 18.
2593 ** The magic number 18 is selected on the basis that scanning 17 rows
2594 ** is almost always quicker than an index seek (even though if the index
2595 ** contains fewer than 2^17 rows we assume otherwise in other parts of
2596 ** the code). And, even if it is not, it should not be too much slower.
2597 ** On the other hand, the extra seeks could end up being significantly
2598 ** more expensive. */
2599 assert( 42==sqlite3LogEst(18) );
2600 if( saved_nEq==saved_nSkip
2601 && saved_nEq+1<pProbe->nKeyCol
2602 && pProbe->noSkipScan==0
2603 && pProbe->aiRowLogEst[saved_nEq+1]>=42 /* TUNING: Minimum for skip-scan */
2604 && (rc = whereLoopResize(db, pNew, pNew->nLTerm+1))==SQLITE_OK
2606 LogEst nIter;
2607 pNew->u.btree.nEq++;
2608 pNew->nSkip++;
2609 pNew->aLTerm[pNew->nLTerm++] = 0;
2610 pNew->wsFlags |= WHERE_SKIPSCAN;
2611 nIter = pProbe->aiRowLogEst[saved_nEq] - pProbe->aiRowLogEst[saved_nEq+1];
2612 pNew->nOut -= nIter;
2613 /* TUNING: Because uncertainties in the estimates for skip-scan queries,
2614 ** add a 1.375 fudge factor to make skip-scan slightly less likely. */
2615 nIter += 5;
2616 whereLoopAddBtreeIndex(pBuilder, pSrc, pProbe, nIter + nInMul);
2617 pNew->nOut = saved_nOut;
2618 pNew->u.btree.nEq = saved_nEq;
2619 pNew->nSkip = saved_nSkip;
2620 pNew->wsFlags = saved_wsFlags;
2623 WHERETRACE(0x800, ("END addBtreeIdx(%s), nEq=%d, rc=%d\n",
2624 pProbe->zName, saved_nEq, rc));
2625 return rc;
2629 ** Return True if it is possible that pIndex might be useful in
2630 ** implementing the ORDER BY clause in pBuilder.
2632 ** Return False if pBuilder does not contain an ORDER BY clause or
2633 ** if there is no way for pIndex to be useful in implementing that
2634 ** ORDER BY clause.
2636 static int indexMightHelpWithOrderBy(
2637 WhereLoopBuilder *pBuilder,
2638 Index *pIndex,
2639 int iCursor
2641 ExprList *pOB;
2642 ExprList *aColExpr;
2643 int ii, jj;
2645 if( pIndex->bUnordered ) return 0;
2646 if( (pOB = pBuilder->pWInfo->pOrderBy)==0 ) return 0;
2647 for(ii=0; ii<pOB->nExpr; ii++){
2648 Expr *pExpr = sqlite3ExprSkipCollate(pOB->a[ii].pExpr);
2649 if( pExpr->op==TK_COLUMN && pExpr->iTable==iCursor ){
2650 if( pExpr->iColumn<0 ) return 1;
2651 for(jj=0; jj<pIndex->nKeyCol; jj++){
2652 if( pExpr->iColumn==pIndex->aiColumn[jj] ) return 1;
2654 }else if( (aColExpr = pIndex->aColExpr)!=0 ){
2655 for(jj=0; jj<pIndex->nKeyCol; jj++){
2656 if( pIndex->aiColumn[jj]!=XN_EXPR ) continue;
2657 if( sqlite3ExprCompare(0, pExpr,aColExpr->a[jj].pExpr,iCursor)==0 ){
2658 return 1;
2663 return 0;
2667 ** Return a bitmask where 1s indicate that the corresponding column of
2668 ** the table is used by an index. Only the first 63 columns are considered.
2670 static Bitmask columnsInIndex(Index *pIdx){
2671 Bitmask m = 0;
2672 int j;
2673 for(j=pIdx->nColumn-1; j>=0; j--){
2674 int x = pIdx->aiColumn[j];
2675 if( x>=0 ){
2676 testcase( x==BMS-1 );
2677 testcase( x==BMS-2 );
2678 if( x<BMS-1 ) m |= MASKBIT(x);
2681 return m;
2684 /* Check to see if a partial index with pPartIndexWhere can be used
2685 ** in the current query. Return true if it can be and false if not.
2687 static int whereUsablePartialIndex(int iTab, WhereClause *pWC, Expr *pWhere){
2688 int i;
2689 WhereTerm *pTerm;
2690 Parse *pParse = pWC->pWInfo->pParse;
2691 while( pWhere->op==TK_AND ){
2692 if( !whereUsablePartialIndex(iTab,pWC,pWhere->pLeft) ) return 0;
2693 pWhere = pWhere->pRight;
2695 if( pParse->db->flags & SQLITE_EnableQPSG ) pParse = 0;
2696 for(i=0, pTerm=pWC->a; i<pWC->nTerm; i++, pTerm++){
2697 Expr *pExpr = pTerm->pExpr;
2698 if( (!ExprHasProperty(pExpr, EP_FromJoin) || pExpr->iRightJoinTable==iTab)
2699 && sqlite3ExprImpliesExpr(pParse, pExpr, pWhere, iTab)
2701 return 1;
2704 return 0;
2708 ** Add all WhereLoop objects for a single table of the join where the table
2709 ** is identified by pBuilder->pNew->iTab. That table is guaranteed to be
2710 ** a b-tree table, not a virtual table.
2712 ** The costs (WhereLoop.rRun) of the b-tree loops added by this function
2713 ** are calculated as follows:
2715 ** For a full scan, assuming the table (or index) contains nRow rows:
2717 ** cost = nRow * 3.0 // full-table scan
2718 ** cost = nRow * K // scan of covering index
2719 ** cost = nRow * (K+3.0) // scan of non-covering index
2721 ** where K is a value between 1.1 and 3.0 set based on the relative
2722 ** estimated average size of the index and table records.
2724 ** For an index scan, where nVisit is the number of index rows visited
2725 ** by the scan, and nSeek is the number of seek operations required on
2726 ** the index b-tree:
2728 ** cost = nSeek * (log(nRow) + K * nVisit) // covering index
2729 ** cost = nSeek * (log(nRow) + (K+3.0) * nVisit) // non-covering index
2731 ** Normally, nSeek is 1. nSeek values greater than 1 come about if the
2732 ** WHERE clause includes "x IN (....)" terms used in place of "x=?". Or when
2733 ** implicit "x IN (SELECT x FROM tbl)" terms are added for skip-scans.
2735 ** The estimated values (nRow, nVisit, nSeek) often contain a large amount
2736 ** of uncertainty. For this reason, scoring is designed to pick plans that
2737 ** "do the least harm" if the estimates are inaccurate. For example, a
2738 ** log(nRow) factor is omitted from a non-covering index scan in order to
2739 ** bias the scoring in favor of using an index, since the worst-case
2740 ** performance of using an index is far better than the worst-case performance
2741 ** of a full table scan.
2743 static int whereLoopAddBtree(
2744 WhereLoopBuilder *pBuilder, /* WHERE clause information */
2745 Bitmask mPrereq /* Extra prerequesites for using this table */
2747 WhereInfo *pWInfo; /* WHERE analysis context */
2748 Index *pProbe; /* An index we are evaluating */
2749 Index sPk; /* A fake index object for the primary key */
2750 LogEst aiRowEstPk[2]; /* The aiRowLogEst[] value for the sPk index */
2751 i16 aiColumnPk = -1; /* The aColumn[] value for the sPk index */
2752 SrcList *pTabList; /* The FROM clause */
2753 struct SrcList_item *pSrc; /* The FROM clause btree term to add */
2754 WhereLoop *pNew; /* Template WhereLoop object */
2755 int rc = SQLITE_OK; /* Return code */
2756 int iSortIdx = 1; /* Index number */
2757 int b; /* A boolean value */
2758 LogEst rSize; /* number of rows in the table */
2759 LogEst rLogSize; /* Logarithm of the number of rows in the table */
2760 WhereClause *pWC; /* The parsed WHERE clause */
2761 Table *pTab; /* Table being queried */
2763 pNew = pBuilder->pNew;
2764 pWInfo = pBuilder->pWInfo;
2765 pTabList = pWInfo->pTabList;
2766 pSrc = pTabList->a + pNew->iTab;
2767 pTab = pSrc->pTab;
2768 pWC = pBuilder->pWC;
2769 assert( !IsVirtual(pSrc->pTab) );
2771 if( pSrc->pIBIndex ){
2772 /* An INDEXED BY clause specifies a particular index to use */
2773 pProbe = pSrc->pIBIndex;
2774 }else if( !HasRowid(pTab) ){
2775 pProbe = pTab->pIndex;
2776 }else{
2777 /* There is no INDEXED BY clause. Create a fake Index object in local
2778 ** variable sPk to represent the rowid primary key index. Make this
2779 ** fake index the first in a chain of Index objects with all of the real
2780 ** indices to follow */
2781 Index *pFirst; /* First of real indices on the table */
2782 memset(&sPk, 0, sizeof(Index));
2783 sPk.nKeyCol = 1;
2784 sPk.nColumn = 1;
2785 sPk.aiColumn = &aiColumnPk;
2786 sPk.aiRowLogEst = aiRowEstPk;
2787 sPk.onError = OE_Replace;
2788 sPk.pTable = pTab;
2789 sPk.szIdxRow = pTab->szTabRow;
2790 aiRowEstPk[0] = pTab->nRowLogEst;
2791 aiRowEstPk[1] = 0;
2792 pFirst = pSrc->pTab->pIndex;
2793 if( pSrc->fg.notIndexed==0 ){
2794 /* The real indices of the table are only considered if the
2795 ** NOT INDEXED qualifier is omitted from the FROM clause */
2796 sPk.pNext = pFirst;
2798 pProbe = &sPk;
2800 rSize = pTab->nRowLogEst;
2801 rLogSize = estLog(rSize);
2803 #ifndef SQLITE_OMIT_AUTOMATIC_INDEX
2804 /* Automatic indexes */
2805 if( !pBuilder->pOrSet /* Not part of an OR optimization */
2806 && (pWInfo->wctrlFlags & WHERE_OR_SUBCLAUSE)==0
2807 && (pWInfo->pParse->db->flags & SQLITE_AutoIndex)!=0
2808 && pSrc->pIBIndex==0 /* Has no INDEXED BY clause */
2809 && !pSrc->fg.notIndexed /* Has no NOT INDEXED clause */
2810 && HasRowid(pTab) /* Not WITHOUT ROWID table. (FIXME: Why not?) */
2811 && !pSrc->fg.isCorrelated /* Not a correlated subquery */
2812 && !pSrc->fg.isRecursive /* Not a recursive common table expression. */
2814 /* Generate auto-index WhereLoops */
2815 WhereTerm *pTerm;
2816 WhereTerm *pWCEnd = pWC->a + pWC->nTerm;
2817 for(pTerm=pWC->a; rc==SQLITE_OK && pTerm<pWCEnd; pTerm++){
2818 if( pTerm->prereqRight & pNew->maskSelf ) continue;
2819 if( termCanDriveIndex(pTerm, pSrc, 0) ){
2820 pNew->u.btree.nEq = 1;
2821 pNew->nSkip = 0;
2822 pNew->u.btree.pIndex = 0;
2823 pNew->nLTerm = 1;
2824 pNew->aLTerm[0] = pTerm;
2825 /* TUNING: One-time cost for computing the automatic index is
2826 ** estimated to be X*N*log2(N) where N is the number of rows in
2827 ** the table being indexed and where X is 7 (LogEst=28) for normal
2828 ** tables or 1.375 (LogEst=4) for views and subqueries. The value
2829 ** of X is smaller for views and subqueries so that the query planner
2830 ** will be more aggressive about generating automatic indexes for
2831 ** those objects, since there is no opportunity to add schema
2832 ** indexes on subqueries and views. */
2833 pNew->rSetup = rLogSize + rSize + 4;
2834 if( pTab->pSelect==0 && (pTab->tabFlags & TF_Ephemeral)==0 ){
2835 pNew->rSetup += 24;
2837 ApplyCostMultiplier(pNew->rSetup, pTab->costMult);
2838 if( pNew->rSetup<0 ) pNew->rSetup = 0;
2839 /* TUNING: Each index lookup yields 20 rows in the table. This
2840 ** is more than the usual guess of 10 rows, since we have no way
2841 ** of knowing how selective the index will ultimately be. It would
2842 ** not be unreasonable to make this value much larger. */
2843 pNew->nOut = 43; assert( 43==sqlite3LogEst(20) );
2844 pNew->rRun = sqlite3LogEstAdd(rLogSize,pNew->nOut);
2845 pNew->wsFlags = WHERE_AUTO_INDEX;
2846 pNew->prereq = mPrereq | pTerm->prereqRight;
2847 rc = whereLoopInsert(pBuilder, pNew);
2851 #endif /* SQLITE_OMIT_AUTOMATIC_INDEX */
2853 /* Loop over all indices
2855 for(; rc==SQLITE_OK && pProbe; pProbe=pProbe->pNext, iSortIdx++){
2856 if( pProbe->pPartIdxWhere!=0
2857 && !whereUsablePartialIndex(pSrc->iCursor, pWC, pProbe->pPartIdxWhere) ){
2858 testcase( pNew->iTab!=pSrc->iCursor ); /* See ticket [98d973b8f5] */
2859 continue; /* Partial index inappropriate for this query */
2861 rSize = pProbe->aiRowLogEst[0];
2862 pNew->u.btree.nEq = 0;
2863 pNew->u.btree.nBtm = 0;
2864 pNew->u.btree.nTop = 0;
2865 pNew->nSkip = 0;
2866 pNew->nLTerm = 0;
2867 pNew->iSortIdx = 0;
2868 pNew->rSetup = 0;
2869 pNew->prereq = mPrereq;
2870 pNew->nOut = rSize;
2871 pNew->u.btree.pIndex = pProbe;
2872 b = indexMightHelpWithOrderBy(pBuilder, pProbe, pSrc->iCursor);
2873 /* The ONEPASS_DESIRED flags never occurs together with ORDER BY */
2874 assert( (pWInfo->wctrlFlags & WHERE_ONEPASS_DESIRED)==0 || b==0 );
2875 if( pProbe->tnum<=0 ){
2876 /* Integer primary key index */
2877 pNew->wsFlags = WHERE_IPK;
2879 /* Full table scan */
2880 pNew->iSortIdx = b ? iSortIdx : 0;
2881 /* TUNING: Cost of full table scan is (N*3.0). */
2882 pNew->rRun = rSize + 16;
2883 ApplyCostMultiplier(pNew->rRun, pTab->costMult);
2884 whereLoopOutputAdjust(pWC, pNew, rSize);
2885 rc = whereLoopInsert(pBuilder, pNew);
2886 pNew->nOut = rSize;
2887 if( rc ) break;
2888 }else{
2889 Bitmask m;
2890 if( pProbe->isCovering ){
2891 pNew->wsFlags = WHERE_IDX_ONLY | WHERE_INDEXED;
2892 m = 0;
2893 }else{
2894 m = pSrc->colUsed & ~columnsInIndex(pProbe);
2895 pNew->wsFlags = (m==0) ? (WHERE_IDX_ONLY|WHERE_INDEXED) : WHERE_INDEXED;
2898 /* Full scan via index */
2899 if( b
2900 || !HasRowid(pTab)
2901 || pProbe->pPartIdxWhere!=0
2902 || ( m==0
2903 && pProbe->bUnordered==0
2904 && (pProbe->szIdxRow<pTab->szTabRow)
2905 && (pWInfo->wctrlFlags & WHERE_ONEPASS_DESIRED)==0
2906 && sqlite3GlobalConfig.bUseCis
2907 && OptimizationEnabled(pWInfo->pParse->db, SQLITE_CoverIdxScan)
2910 pNew->iSortIdx = b ? iSortIdx : 0;
2912 /* The cost of visiting the index rows is N*K, where K is
2913 ** between 1.1 and 3.0, depending on the relative sizes of the
2914 ** index and table rows. */
2915 pNew->rRun = rSize + 1 + (15*pProbe->szIdxRow)/pTab->szTabRow;
2916 if( m!=0 ){
2917 /* If this is a non-covering index scan, add in the cost of
2918 ** doing table lookups. The cost will be 3x the number of
2919 ** lookups. Take into account WHERE clause terms that can be
2920 ** satisfied using just the index, and that do not require a
2921 ** table lookup. */
2922 LogEst nLookup = rSize + 16; /* Base cost: N*3 */
2923 int ii;
2924 int iCur = pSrc->iCursor;
2925 WhereClause *pWC2 = &pWInfo->sWC;
2926 for(ii=0; ii<pWC2->nTerm; ii++){
2927 WhereTerm *pTerm = &pWC2->a[ii];
2928 if( !sqlite3ExprCoveredByIndex(pTerm->pExpr, iCur, pProbe) ){
2929 break;
2931 /* pTerm can be evaluated using just the index. So reduce
2932 ** the expected number of table lookups accordingly */
2933 if( pTerm->truthProb<=0 ){
2934 nLookup += pTerm->truthProb;
2935 }else{
2936 nLookup--;
2937 if( pTerm->eOperator & (WO_EQ|WO_IS) ) nLookup -= 19;
2941 pNew->rRun = sqlite3LogEstAdd(pNew->rRun, nLookup);
2943 ApplyCostMultiplier(pNew->rRun, pTab->costMult);
2944 whereLoopOutputAdjust(pWC, pNew, rSize);
2945 rc = whereLoopInsert(pBuilder, pNew);
2946 pNew->nOut = rSize;
2947 if( rc ) break;
2951 pBuilder->bldFlags = 0;
2952 rc = whereLoopAddBtreeIndex(pBuilder, pSrc, pProbe, 0);
2953 if( pBuilder->bldFlags==SQLITE_BLDF_INDEXED ){
2954 /* If a non-unique index is used, or if a prefix of the key for
2955 ** unique index is used (making the index functionally non-unique)
2956 ** then the sqlite_stat1 data becomes important for scoring the
2957 ** plan */
2958 pTab->tabFlags |= TF_StatsUsed;
2960 #ifdef SQLITE_ENABLE_STAT3_OR_STAT4
2961 sqlite3Stat4ProbeFree(pBuilder->pRec);
2962 pBuilder->nRecValid = 0;
2963 pBuilder->pRec = 0;
2964 #endif
2966 /* If there was an INDEXED BY clause, then only that one index is
2967 ** considered. */
2968 if( pSrc->pIBIndex ) break;
2970 return rc;
2973 #ifndef SQLITE_OMIT_VIRTUALTABLE
2976 ** Argument pIdxInfo is already populated with all constraints that may
2977 ** be used by the virtual table identified by pBuilder->pNew->iTab. This
2978 ** function marks a subset of those constraints usable, invokes the
2979 ** xBestIndex method and adds the returned plan to pBuilder.
2981 ** A constraint is marked usable if:
2983 ** * Argument mUsable indicates that its prerequisites are available, and
2985 ** * It is not one of the operators specified in the mExclude mask passed
2986 ** as the fourth argument (which in practice is either WO_IN or 0).
2988 ** Argument mPrereq is a mask of tables that must be scanned before the
2989 ** virtual table in question. These are added to the plans prerequisites
2990 ** before it is added to pBuilder.
2992 ** Output parameter *pbIn is set to true if the plan added to pBuilder
2993 ** uses one or more WO_IN terms, or false otherwise.
2995 static int whereLoopAddVirtualOne(
2996 WhereLoopBuilder *pBuilder,
2997 Bitmask mPrereq, /* Mask of tables that must be used. */
2998 Bitmask mUsable, /* Mask of usable tables */
2999 u16 mExclude, /* Exclude terms using these operators */
3000 sqlite3_index_info *pIdxInfo, /* Populated object for xBestIndex */
3001 u16 mNoOmit, /* Do not omit these constraints */
3002 int *pbIn /* OUT: True if plan uses an IN(...) op */
3004 WhereClause *pWC = pBuilder->pWC;
3005 struct sqlite3_index_constraint *pIdxCons;
3006 struct sqlite3_index_constraint_usage *pUsage = pIdxInfo->aConstraintUsage;
3007 int i;
3008 int mxTerm;
3009 int rc = SQLITE_OK;
3010 WhereLoop *pNew = pBuilder->pNew;
3011 Parse *pParse = pBuilder->pWInfo->pParse;
3012 struct SrcList_item *pSrc = &pBuilder->pWInfo->pTabList->a[pNew->iTab];
3013 int nConstraint = pIdxInfo->nConstraint;
3015 assert( (mUsable & mPrereq)==mPrereq );
3016 *pbIn = 0;
3017 pNew->prereq = mPrereq;
3019 /* Set the usable flag on the subset of constraints identified by
3020 ** arguments mUsable and mExclude. */
3021 pIdxCons = *(struct sqlite3_index_constraint**)&pIdxInfo->aConstraint;
3022 for(i=0; i<nConstraint; i++, pIdxCons++){
3023 WhereTerm *pTerm = &pWC->a[pIdxCons->iTermOffset];
3024 pIdxCons->usable = 0;
3025 if( (pTerm->prereqRight & mUsable)==pTerm->prereqRight
3026 && (pTerm->eOperator & mExclude)==0
3028 pIdxCons->usable = 1;
3032 /* Initialize the output fields of the sqlite3_index_info structure */
3033 memset(pUsage, 0, sizeof(pUsage[0])*nConstraint);
3034 assert( pIdxInfo->needToFreeIdxStr==0 );
3035 pIdxInfo->idxStr = 0;
3036 pIdxInfo->idxNum = 0;
3037 pIdxInfo->orderByConsumed = 0;
3038 pIdxInfo->estimatedCost = SQLITE_BIG_DBL / (double)2;
3039 pIdxInfo->estimatedRows = 25;
3040 pIdxInfo->idxFlags = 0;
3041 pIdxInfo->colUsed = (sqlite3_int64)pSrc->colUsed;
3043 /* Invoke the virtual table xBestIndex() method */
3044 rc = vtabBestIndex(pParse, pSrc->pTab, pIdxInfo);
3045 if( rc ) return rc;
3047 mxTerm = -1;
3048 assert( pNew->nLSlot>=nConstraint );
3049 for(i=0; i<nConstraint; i++) pNew->aLTerm[i] = 0;
3050 pNew->u.vtab.omitMask = 0;
3051 pIdxCons = *(struct sqlite3_index_constraint**)&pIdxInfo->aConstraint;
3052 for(i=0; i<nConstraint; i++, pIdxCons++){
3053 int iTerm;
3054 if( (iTerm = pUsage[i].argvIndex - 1)>=0 ){
3055 WhereTerm *pTerm;
3056 int j = pIdxCons->iTermOffset;
3057 if( iTerm>=nConstraint
3058 || j<0
3059 || j>=pWC->nTerm
3060 || pNew->aLTerm[iTerm]!=0
3061 || pIdxCons->usable==0
3063 rc = SQLITE_ERROR;
3064 sqlite3ErrorMsg(pParse,"%s.xBestIndex malfunction",pSrc->pTab->zName);
3065 return rc;
3067 testcase( iTerm==nConstraint-1 );
3068 testcase( j==0 );
3069 testcase( j==pWC->nTerm-1 );
3070 pTerm = &pWC->a[j];
3071 pNew->prereq |= pTerm->prereqRight;
3072 assert( iTerm<pNew->nLSlot );
3073 pNew->aLTerm[iTerm] = pTerm;
3074 if( iTerm>mxTerm ) mxTerm = iTerm;
3075 testcase( iTerm==15 );
3076 testcase( iTerm==16 );
3077 if( iTerm<16 && pUsage[i].omit ) pNew->u.vtab.omitMask |= 1<<iTerm;
3078 if( (pTerm->eOperator & WO_IN)!=0 ){
3079 /* A virtual table that is constrained by an IN clause may not
3080 ** consume the ORDER BY clause because (1) the order of IN terms
3081 ** is not necessarily related to the order of output terms and
3082 ** (2) Multiple outputs from a single IN value will not merge
3083 ** together. */
3084 pIdxInfo->orderByConsumed = 0;
3085 pIdxInfo->idxFlags &= ~SQLITE_INDEX_SCAN_UNIQUE;
3086 *pbIn = 1; assert( (mExclude & WO_IN)==0 );
3090 pNew->u.vtab.omitMask &= ~mNoOmit;
3092 pNew->nLTerm = mxTerm+1;
3093 assert( pNew->nLTerm<=pNew->nLSlot );
3094 pNew->u.vtab.idxNum = pIdxInfo->idxNum;
3095 pNew->u.vtab.needFree = pIdxInfo->needToFreeIdxStr;
3096 pIdxInfo->needToFreeIdxStr = 0;
3097 pNew->u.vtab.idxStr = pIdxInfo->idxStr;
3098 pNew->u.vtab.isOrdered = (i8)(pIdxInfo->orderByConsumed ?
3099 pIdxInfo->nOrderBy : 0);
3100 pNew->rSetup = 0;
3101 pNew->rRun = sqlite3LogEstFromDouble(pIdxInfo->estimatedCost);
3102 pNew->nOut = sqlite3LogEst(pIdxInfo->estimatedRows);
3104 /* Set the WHERE_ONEROW flag if the xBestIndex() method indicated
3105 ** that the scan will visit at most one row. Clear it otherwise. */
3106 if( pIdxInfo->idxFlags & SQLITE_INDEX_SCAN_UNIQUE ){
3107 pNew->wsFlags |= WHERE_ONEROW;
3108 }else{
3109 pNew->wsFlags &= ~WHERE_ONEROW;
3111 rc = whereLoopInsert(pBuilder, pNew);
3112 if( pNew->u.vtab.needFree ){
3113 sqlite3_free(pNew->u.vtab.idxStr);
3114 pNew->u.vtab.needFree = 0;
3116 WHERETRACE(0xffff, (" bIn=%d prereqIn=%04llx prereqOut=%04llx\n",
3117 *pbIn, (sqlite3_uint64)mPrereq,
3118 (sqlite3_uint64)(pNew->prereq & ~mPrereq)));
3120 return rc;
3125 ** Add all WhereLoop objects for a table of the join identified by
3126 ** pBuilder->pNew->iTab. That table is guaranteed to be a virtual table.
3128 ** If there are no LEFT or CROSS JOIN joins in the query, both mPrereq and
3129 ** mUnusable are set to 0. Otherwise, mPrereq is a mask of all FROM clause
3130 ** entries that occur before the virtual table in the FROM clause and are
3131 ** separated from it by at least one LEFT or CROSS JOIN. Similarly, the
3132 ** mUnusable mask contains all FROM clause entries that occur after the
3133 ** virtual table and are separated from it by at least one LEFT or
3134 ** CROSS JOIN.
3136 ** For example, if the query were:
3138 ** ... FROM t1, t2 LEFT JOIN t3, t4, vt CROSS JOIN t5, t6;
3140 ** then mPrereq corresponds to (t1, t2) and mUnusable to (t5, t6).
3142 ** All the tables in mPrereq must be scanned before the current virtual
3143 ** table. So any terms for which all prerequisites are satisfied by
3144 ** mPrereq may be specified as "usable" in all calls to xBestIndex.
3145 ** Conversely, all tables in mUnusable must be scanned after the current
3146 ** virtual table, so any terms for which the prerequisites overlap with
3147 ** mUnusable should always be configured as "not-usable" for xBestIndex.
3149 static int whereLoopAddVirtual(
3150 WhereLoopBuilder *pBuilder, /* WHERE clause information */
3151 Bitmask mPrereq, /* Tables that must be scanned before this one */
3152 Bitmask mUnusable /* Tables that must be scanned after this one */
3154 int rc = SQLITE_OK; /* Return code */
3155 WhereInfo *pWInfo; /* WHERE analysis context */
3156 Parse *pParse; /* The parsing context */
3157 WhereClause *pWC; /* The WHERE clause */
3158 struct SrcList_item *pSrc; /* The FROM clause term to search */
3159 sqlite3_index_info *p; /* Object to pass to xBestIndex() */
3160 int nConstraint; /* Number of constraints in p */
3161 int bIn; /* True if plan uses IN(...) operator */
3162 WhereLoop *pNew;
3163 Bitmask mBest; /* Tables used by best possible plan */
3164 u16 mNoOmit;
3166 assert( (mPrereq & mUnusable)==0 );
3167 pWInfo = pBuilder->pWInfo;
3168 pParse = pWInfo->pParse;
3169 pWC = pBuilder->pWC;
3170 pNew = pBuilder->pNew;
3171 pSrc = &pWInfo->pTabList->a[pNew->iTab];
3172 assert( IsVirtual(pSrc->pTab) );
3173 p = allocateIndexInfo(pParse, pWC, mUnusable, pSrc, pBuilder->pOrderBy,
3174 &mNoOmit);
3175 if( p==0 ) return SQLITE_NOMEM_BKPT;
3176 pNew->rSetup = 0;
3177 pNew->wsFlags = WHERE_VIRTUALTABLE;
3178 pNew->nLTerm = 0;
3179 pNew->u.vtab.needFree = 0;
3180 nConstraint = p->nConstraint;
3181 if( whereLoopResize(pParse->db, pNew, nConstraint) ){
3182 sqlite3DbFree(pParse->db, p);
3183 return SQLITE_NOMEM_BKPT;
3186 /* First call xBestIndex() with all constraints usable. */
3187 WHERETRACE(0x40, (" VirtualOne: all usable\n"));
3188 rc = whereLoopAddVirtualOne(pBuilder, mPrereq, ALLBITS, 0, p, mNoOmit, &bIn);
3190 /* If the call to xBestIndex() with all terms enabled produced a plan
3191 ** that does not require any source tables (IOW: a plan with mBest==0),
3192 ** then there is no point in making any further calls to xBestIndex()
3193 ** since they will all return the same result (if the xBestIndex()
3194 ** implementation is sane). */
3195 if( rc==SQLITE_OK && (mBest = (pNew->prereq & ~mPrereq))!=0 ){
3196 int seenZero = 0; /* True if a plan with no prereqs seen */
3197 int seenZeroNoIN = 0; /* Plan with no prereqs and no IN(...) seen */
3198 Bitmask mPrev = 0;
3199 Bitmask mBestNoIn = 0;
3201 /* If the plan produced by the earlier call uses an IN(...) term, call
3202 ** xBestIndex again, this time with IN(...) terms disabled. */
3203 if( bIn ){
3204 WHERETRACE(0x40, (" VirtualOne: all usable w/o IN\n"));
3205 rc = whereLoopAddVirtualOne(
3206 pBuilder, mPrereq, ALLBITS, WO_IN, p, mNoOmit, &bIn);
3207 assert( bIn==0 );
3208 mBestNoIn = pNew->prereq & ~mPrereq;
3209 if( mBestNoIn==0 ){
3210 seenZero = 1;
3211 seenZeroNoIN = 1;
3215 /* Call xBestIndex once for each distinct value of (prereqRight & ~mPrereq)
3216 ** in the set of terms that apply to the current virtual table. */
3217 while( rc==SQLITE_OK ){
3218 int i;
3219 Bitmask mNext = ALLBITS;
3220 assert( mNext>0 );
3221 for(i=0; i<nConstraint; i++){
3222 Bitmask mThis = (
3223 pWC->a[p->aConstraint[i].iTermOffset].prereqRight & ~mPrereq
3225 if( mThis>mPrev && mThis<mNext ) mNext = mThis;
3227 mPrev = mNext;
3228 if( mNext==ALLBITS ) break;
3229 if( mNext==mBest || mNext==mBestNoIn ) continue;
3230 WHERETRACE(0x40, (" VirtualOne: mPrev=%04llx mNext=%04llx\n",
3231 (sqlite3_uint64)mPrev, (sqlite3_uint64)mNext));
3232 rc = whereLoopAddVirtualOne(
3233 pBuilder, mPrereq, mNext|mPrereq, 0, p, mNoOmit, &bIn);
3234 if( pNew->prereq==mPrereq ){
3235 seenZero = 1;
3236 if( bIn==0 ) seenZeroNoIN = 1;
3240 /* If the calls to xBestIndex() in the above loop did not find a plan
3241 ** that requires no source tables at all (i.e. one guaranteed to be
3242 ** usable), make a call here with all source tables disabled */
3243 if( rc==SQLITE_OK && seenZero==0 ){
3244 WHERETRACE(0x40, (" VirtualOne: all disabled\n"));
3245 rc = whereLoopAddVirtualOne(
3246 pBuilder, mPrereq, mPrereq, 0, p, mNoOmit, &bIn);
3247 if( bIn==0 ) seenZeroNoIN = 1;
3250 /* If the calls to xBestIndex() have so far failed to find a plan
3251 ** that requires no source tables at all and does not use an IN(...)
3252 ** operator, make a final call to obtain one here. */
3253 if( rc==SQLITE_OK && seenZeroNoIN==0 ){
3254 WHERETRACE(0x40, (" VirtualOne: all disabled and w/o IN\n"));
3255 rc = whereLoopAddVirtualOne(
3256 pBuilder, mPrereq, mPrereq, WO_IN, p, mNoOmit, &bIn);
3260 if( p->needToFreeIdxStr ) sqlite3_free(p->idxStr);
3261 sqlite3DbFreeNN(pParse->db, p);
3262 return rc;
3264 #endif /* SQLITE_OMIT_VIRTUALTABLE */
3267 ** Add WhereLoop entries to handle OR terms. This works for either
3268 ** btrees or virtual tables.
3270 static int whereLoopAddOr(
3271 WhereLoopBuilder *pBuilder,
3272 Bitmask mPrereq,
3273 Bitmask mUnusable
3275 WhereInfo *pWInfo = pBuilder->pWInfo;
3276 WhereClause *pWC;
3277 WhereLoop *pNew;
3278 WhereTerm *pTerm, *pWCEnd;
3279 int rc = SQLITE_OK;
3280 int iCur;
3281 WhereClause tempWC;
3282 WhereLoopBuilder sSubBuild;
3283 WhereOrSet sSum, sCur;
3284 struct SrcList_item *pItem;
3286 pWC = pBuilder->pWC;
3287 pWCEnd = pWC->a + pWC->nTerm;
3288 pNew = pBuilder->pNew;
3289 memset(&sSum, 0, sizeof(sSum));
3290 pItem = pWInfo->pTabList->a + pNew->iTab;
3291 iCur = pItem->iCursor;
3293 for(pTerm=pWC->a; pTerm<pWCEnd && rc==SQLITE_OK; pTerm++){
3294 if( (pTerm->eOperator & WO_OR)!=0
3295 && (pTerm->u.pOrInfo->indexable & pNew->maskSelf)!=0
3297 WhereClause * const pOrWC = &pTerm->u.pOrInfo->wc;
3298 WhereTerm * const pOrWCEnd = &pOrWC->a[pOrWC->nTerm];
3299 WhereTerm *pOrTerm;
3300 int once = 1;
3301 int i, j;
3303 sSubBuild = *pBuilder;
3304 sSubBuild.pOrderBy = 0;
3305 sSubBuild.pOrSet = &sCur;
3307 WHERETRACE(0x200, ("Begin processing OR-clause %p\n", pTerm));
3308 for(pOrTerm=pOrWC->a; pOrTerm<pOrWCEnd; pOrTerm++){
3309 if( (pOrTerm->eOperator & WO_AND)!=0 ){
3310 sSubBuild.pWC = &pOrTerm->u.pAndInfo->wc;
3311 }else if( pOrTerm->leftCursor==iCur ){
3312 tempWC.pWInfo = pWC->pWInfo;
3313 tempWC.pOuter = pWC;
3314 tempWC.op = TK_AND;
3315 tempWC.nTerm = 1;
3316 tempWC.a = pOrTerm;
3317 sSubBuild.pWC = &tempWC;
3318 }else{
3319 continue;
3321 sCur.n = 0;
3322 #ifdef WHERETRACE_ENABLED
3323 WHERETRACE(0x200, ("OR-term %d of %p has %d subterms:\n",
3324 (int)(pOrTerm-pOrWC->a), pTerm, sSubBuild.pWC->nTerm));
3325 if( sqlite3WhereTrace & 0x400 ){
3326 sqlite3WhereClausePrint(sSubBuild.pWC);
3328 #endif
3329 #ifndef SQLITE_OMIT_VIRTUALTABLE
3330 if( IsVirtual(pItem->pTab) ){
3331 rc = whereLoopAddVirtual(&sSubBuild, mPrereq, mUnusable);
3332 }else
3333 #endif
3335 rc = whereLoopAddBtree(&sSubBuild, mPrereq);
3337 if( rc==SQLITE_OK ){
3338 rc = whereLoopAddOr(&sSubBuild, mPrereq, mUnusable);
3340 assert( rc==SQLITE_OK || sCur.n==0 );
3341 if( sCur.n==0 ){
3342 sSum.n = 0;
3343 break;
3344 }else if( once ){
3345 whereOrMove(&sSum, &sCur);
3346 once = 0;
3347 }else{
3348 WhereOrSet sPrev;
3349 whereOrMove(&sPrev, &sSum);
3350 sSum.n = 0;
3351 for(i=0; i<sPrev.n; i++){
3352 for(j=0; j<sCur.n; j++){
3353 whereOrInsert(&sSum, sPrev.a[i].prereq | sCur.a[j].prereq,
3354 sqlite3LogEstAdd(sPrev.a[i].rRun, sCur.a[j].rRun),
3355 sqlite3LogEstAdd(sPrev.a[i].nOut, sCur.a[j].nOut));
3360 pNew->nLTerm = 1;
3361 pNew->aLTerm[0] = pTerm;
3362 pNew->wsFlags = WHERE_MULTI_OR;
3363 pNew->rSetup = 0;
3364 pNew->iSortIdx = 0;
3365 memset(&pNew->u, 0, sizeof(pNew->u));
3366 for(i=0; rc==SQLITE_OK && i<sSum.n; i++){
3367 /* TUNING: Currently sSum.a[i].rRun is set to the sum of the costs
3368 ** of all sub-scans required by the OR-scan. However, due to rounding
3369 ** errors, it may be that the cost of the OR-scan is equal to its
3370 ** most expensive sub-scan. Add the smallest possible penalty
3371 ** (equivalent to multiplying the cost by 1.07) to ensure that
3372 ** this does not happen. Otherwise, for WHERE clauses such as the
3373 ** following where there is an index on "y":
3375 ** WHERE likelihood(x=?, 0.99) OR y=?
3377 ** the planner may elect to "OR" together a full-table scan and an
3378 ** index lookup. And other similarly odd results. */
3379 pNew->rRun = sSum.a[i].rRun + 1;
3380 pNew->nOut = sSum.a[i].nOut;
3381 pNew->prereq = sSum.a[i].prereq;
3382 rc = whereLoopInsert(pBuilder, pNew);
3384 WHERETRACE(0x200, ("End processing OR-clause %p\n", pTerm));
3387 return rc;
3391 ** Add all WhereLoop objects for all tables
3393 static int whereLoopAddAll(WhereLoopBuilder *pBuilder){
3394 WhereInfo *pWInfo = pBuilder->pWInfo;
3395 Bitmask mPrereq = 0;
3396 Bitmask mPrior = 0;
3397 int iTab;
3398 SrcList *pTabList = pWInfo->pTabList;
3399 struct SrcList_item *pItem;
3400 struct SrcList_item *pEnd = &pTabList->a[pWInfo->nLevel];
3401 sqlite3 *db = pWInfo->pParse->db;
3402 int rc = SQLITE_OK;
3403 WhereLoop *pNew;
3404 u8 priorJointype = 0;
3406 /* Loop over the tables in the join, from left to right */
3407 pNew = pBuilder->pNew;
3408 whereLoopInit(pNew);
3409 for(iTab=0, pItem=pTabList->a; pItem<pEnd; iTab++, pItem++){
3410 Bitmask mUnusable = 0;
3411 pNew->iTab = iTab;
3412 pNew->maskSelf = sqlite3WhereGetMask(&pWInfo->sMaskSet, pItem->iCursor);
3413 if( ((pItem->fg.jointype|priorJointype) & (JT_LEFT|JT_CROSS))!=0 ){
3414 /* This condition is true when pItem is the FROM clause term on the
3415 ** right-hand-side of a LEFT or CROSS JOIN. */
3416 mPrereq = mPrior;
3418 priorJointype = pItem->fg.jointype;
3419 #ifndef SQLITE_OMIT_VIRTUALTABLE
3420 if( IsVirtual(pItem->pTab) ){
3421 struct SrcList_item *p;
3422 for(p=&pItem[1]; p<pEnd; p++){
3423 if( mUnusable || (p->fg.jointype & (JT_LEFT|JT_CROSS)) ){
3424 mUnusable |= sqlite3WhereGetMask(&pWInfo->sMaskSet, p->iCursor);
3427 rc = whereLoopAddVirtual(pBuilder, mPrereq, mUnusable);
3428 }else
3429 #endif /* SQLITE_OMIT_VIRTUALTABLE */
3431 rc = whereLoopAddBtree(pBuilder, mPrereq);
3433 if( rc==SQLITE_OK ){
3434 rc = whereLoopAddOr(pBuilder, mPrereq, mUnusable);
3436 mPrior |= pNew->maskSelf;
3437 if( rc || db->mallocFailed ) break;
3440 whereLoopClear(db, pNew);
3441 return rc;
3445 ** Examine a WherePath (with the addition of the extra WhereLoop of the 6th
3446 ** parameters) to see if it outputs rows in the requested ORDER BY
3447 ** (or GROUP BY) without requiring a separate sort operation. Return N:
3449 ** N>0: N terms of the ORDER BY clause are satisfied
3450 ** N==0: No terms of the ORDER BY clause are satisfied
3451 ** N<0: Unknown yet how many terms of ORDER BY might be satisfied.
3453 ** Note that processing for WHERE_GROUPBY and WHERE_DISTINCTBY is not as
3454 ** strict. With GROUP BY and DISTINCT the only requirement is that
3455 ** equivalent rows appear immediately adjacent to one another. GROUP BY
3456 ** and DISTINCT do not require rows to appear in any particular order as long
3457 ** as equivalent rows are grouped together. Thus for GROUP BY and DISTINCT
3458 ** the pOrderBy terms can be matched in any order. With ORDER BY, the
3459 ** pOrderBy terms must be matched in strict left-to-right order.
3461 static i8 wherePathSatisfiesOrderBy(
3462 WhereInfo *pWInfo, /* The WHERE clause */
3463 ExprList *pOrderBy, /* ORDER BY or GROUP BY or DISTINCT clause to check */
3464 WherePath *pPath, /* The WherePath to check */
3465 u16 wctrlFlags, /* WHERE_GROUPBY or _DISTINCTBY or _ORDERBY_LIMIT */
3466 u16 nLoop, /* Number of entries in pPath->aLoop[] */
3467 WhereLoop *pLast, /* Add this WhereLoop to the end of pPath->aLoop[] */
3468 Bitmask *pRevMask /* OUT: Mask of WhereLoops to run in reverse order */
3470 u8 revSet; /* True if rev is known */
3471 u8 rev; /* Composite sort order */
3472 u8 revIdx; /* Index sort order */
3473 u8 isOrderDistinct; /* All prior WhereLoops are order-distinct */
3474 u8 distinctColumns; /* True if the loop has UNIQUE NOT NULL columns */
3475 u8 isMatch; /* iColumn matches a term of the ORDER BY clause */
3476 u16 eqOpMask; /* Allowed equality operators */
3477 u16 nKeyCol; /* Number of key columns in pIndex */
3478 u16 nColumn; /* Total number of ordered columns in the index */
3479 u16 nOrderBy; /* Number terms in the ORDER BY clause */
3480 int iLoop; /* Index of WhereLoop in pPath being processed */
3481 int i, j; /* Loop counters */
3482 int iCur; /* Cursor number for current WhereLoop */
3483 int iColumn; /* A column number within table iCur */
3484 WhereLoop *pLoop = 0; /* Current WhereLoop being processed. */
3485 WhereTerm *pTerm; /* A single term of the WHERE clause */
3486 Expr *pOBExpr; /* An expression from the ORDER BY clause */
3487 CollSeq *pColl; /* COLLATE function from an ORDER BY clause term */
3488 Index *pIndex; /* The index associated with pLoop */
3489 sqlite3 *db = pWInfo->pParse->db; /* Database connection */
3490 Bitmask obSat = 0; /* Mask of ORDER BY terms satisfied so far */
3491 Bitmask obDone; /* Mask of all ORDER BY terms */
3492 Bitmask orderDistinctMask; /* Mask of all well-ordered loops */
3493 Bitmask ready; /* Mask of inner loops */
3496 ** We say the WhereLoop is "one-row" if it generates no more than one
3497 ** row of output. A WhereLoop is one-row if all of the following are true:
3498 ** (a) All index columns match with WHERE_COLUMN_EQ.
3499 ** (b) The index is unique
3500 ** Any WhereLoop with an WHERE_COLUMN_EQ constraint on the rowid is one-row.
3501 ** Every one-row WhereLoop will have the WHERE_ONEROW bit set in wsFlags.
3503 ** We say the WhereLoop is "order-distinct" if the set of columns from
3504 ** that WhereLoop that are in the ORDER BY clause are different for every
3505 ** row of the WhereLoop. Every one-row WhereLoop is automatically
3506 ** order-distinct. A WhereLoop that has no columns in the ORDER BY clause
3507 ** is not order-distinct. To be order-distinct is not quite the same as being
3508 ** UNIQUE since a UNIQUE column or index can have multiple rows that
3509 ** are NULL and NULL values are equivalent for the purpose of order-distinct.
3510 ** To be order-distinct, the columns must be UNIQUE and NOT NULL.
3512 ** The rowid for a table is always UNIQUE and NOT NULL so whenever the
3513 ** rowid appears in the ORDER BY clause, the corresponding WhereLoop is
3514 ** automatically order-distinct.
3517 assert( pOrderBy!=0 );
3518 if( nLoop && OptimizationDisabled(db, SQLITE_OrderByIdxJoin) ) return 0;
3520 nOrderBy = pOrderBy->nExpr;
3521 testcase( nOrderBy==BMS-1 );
3522 if( nOrderBy>BMS-1 ) return 0; /* Cannot optimize overly large ORDER BYs */
3523 isOrderDistinct = 1;
3524 obDone = MASKBIT(nOrderBy)-1;
3525 orderDistinctMask = 0;
3526 ready = 0;
3527 eqOpMask = WO_EQ | WO_IS | WO_ISNULL;
3528 if( wctrlFlags & WHERE_ORDERBY_LIMIT ) eqOpMask |= WO_IN;
3529 for(iLoop=0; isOrderDistinct && obSat<obDone && iLoop<=nLoop; iLoop++){
3530 if( iLoop>0 ) ready |= pLoop->maskSelf;
3531 if( iLoop<nLoop ){
3532 pLoop = pPath->aLoop[iLoop];
3533 if( wctrlFlags & WHERE_ORDERBY_LIMIT ) continue;
3534 }else{
3535 pLoop = pLast;
3537 if( pLoop->wsFlags & WHERE_VIRTUALTABLE ){
3538 if( pLoop->u.vtab.isOrdered ) obSat = obDone;
3539 break;
3540 }else{
3541 pLoop->u.btree.nIdxCol = 0;
3543 iCur = pWInfo->pTabList->a[pLoop->iTab].iCursor;
3545 /* Mark off any ORDER BY term X that is a column in the table of
3546 ** the current loop for which there is term in the WHERE
3547 ** clause of the form X IS NULL or X=? that reference only outer
3548 ** loops.
3550 for(i=0; i<nOrderBy; i++){
3551 if( MASKBIT(i) & obSat ) continue;
3552 pOBExpr = sqlite3ExprSkipCollate(pOrderBy->a[i].pExpr);
3553 if( pOBExpr->op!=TK_COLUMN ) continue;
3554 if( pOBExpr->iTable!=iCur ) continue;
3555 pTerm = sqlite3WhereFindTerm(&pWInfo->sWC, iCur, pOBExpr->iColumn,
3556 ~ready, eqOpMask, 0);
3557 if( pTerm==0 ) continue;
3558 if( pTerm->eOperator==WO_IN ){
3559 /* IN terms are only valid for sorting in the ORDER BY LIMIT
3560 ** optimization, and then only if they are actually used
3561 ** by the query plan */
3562 assert( wctrlFlags & WHERE_ORDERBY_LIMIT );
3563 for(j=0; j<pLoop->nLTerm && pTerm!=pLoop->aLTerm[j]; j++){}
3564 if( j>=pLoop->nLTerm ) continue;
3566 if( (pTerm->eOperator&(WO_EQ|WO_IS))!=0 && pOBExpr->iColumn>=0 ){
3567 const char *z1, *z2;
3568 pColl = sqlite3ExprCollSeq(pWInfo->pParse, pOrderBy->a[i].pExpr);
3569 if( !pColl ) pColl = db->pDfltColl;
3570 z1 = pColl->zName;
3571 pColl = sqlite3ExprCollSeq(pWInfo->pParse, pTerm->pExpr);
3572 if( !pColl ) pColl = db->pDfltColl;
3573 z2 = pColl->zName;
3574 if( sqlite3StrICmp(z1, z2)!=0 ) continue;
3575 testcase( pTerm->pExpr->op==TK_IS );
3577 obSat |= MASKBIT(i);
3580 if( (pLoop->wsFlags & WHERE_ONEROW)==0 ){
3581 if( pLoop->wsFlags & WHERE_IPK ){
3582 pIndex = 0;
3583 nKeyCol = 0;
3584 nColumn = 1;
3585 }else if( (pIndex = pLoop->u.btree.pIndex)==0 || pIndex->bUnordered ){
3586 return 0;
3587 }else{
3588 nKeyCol = pIndex->nKeyCol;
3589 nColumn = pIndex->nColumn;
3590 assert( nColumn==nKeyCol+1 || !HasRowid(pIndex->pTable) );
3591 assert( pIndex->aiColumn[nColumn-1]==XN_ROWID
3592 || !HasRowid(pIndex->pTable));
3593 isOrderDistinct = IsUniqueIndex(pIndex);
3596 /* Loop through all columns of the index and deal with the ones
3597 ** that are not constrained by == or IN.
3599 rev = revSet = 0;
3600 distinctColumns = 0;
3601 for(j=0; j<nColumn; j++){
3602 u8 bOnce = 1; /* True to run the ORDER BY search loop */
3604 assert( j>=pLoop->u.btree.nEq
3605 || (pLoop->aLTerm[j]==0)==(j<pLoop->nSkip)
3607 if( j<pLoop->u.btree.nEq && j>=pLoop->nSkip ){
3608 u16 eOp = pLoop->aLTerm[j]->eOperator;
3610 /* Skip over == and IS and ISNULL terms. (Also skip IN terms when
3611 ** doing WHERE_ORDERBY_LIMIT processing).
3613 ** If the current term is a column of an ((?,?) IN (SELECT...))
3614 ** expression for which the SELECT returns more than one column,
3615 ** check that it is the only column used by this loop. Otherwise,
3616 ** if it is one of two or more, none of the columns can be
3617 ** considered to match an ORDER BY term. */
3618 if( (eOp & eqOpMask)!=0 ){
3619 if( eOp & WO_ISNULL ){
3620 testcase( isOrderDistinct );
3621 isOrderDistinct = 0;
3623 continue;
3624 }else if( ALWAYS(eOp & WO_IN) ){
3625 /* ALWAYS() justification: eOp is an equality operator due to the
3626 ** j<pLoop->u.btree.nEq constraint above. Any equality other
3627 ** than WO_IN is captured by the previous "if". So this one
3628 ** always has to be WO_IN. */
3629 Expr *pX = pLoop->aLTerm[j]->pExpr;
3630 for(i=j+1; i<pLoop->u.btree.nEq; i++){
3631 if( pLoop->aLTerm[i]->pExpr==pX ){
3632 assert( (pLoop->aLTerm[i]->eOperator & WO_IN) );
3633 bOnce = 0;
3634 break;
3640 /* Get the column number in the table (iColumn) and sort order
3641 ** (revIdx) for the j-th column of the index.
3643 if( pIndex ){
3644 iColumn = pIndex->aiColumn[j];
3645 revIdx = pIndex->aSortOrder[j];
3646 if( iColumn==pIndex->pTable->iPKey ) iColumn = -1;
3647 }else{
3648 iColumn = XN_ROWID;
3649 revIdx = 0;
3652 /* An unconstrained column that might be NULL means that this
3653 ** WhereLoop is not well-ordered
3655 if( isOrderDistinct
3656 && iColumn>=0
3657 && j>=pLoop->u.btree.nEq
3658 && pIndex->pTable->aCol[iColumn].notNull==0
3660 isOrderDistinct = 0;
3663 /* Find the ORDER BY term that corresponds to the j-th column
3664 ** of the index and mark that ORDER BY term off
3666 isMatch = 0;
3667 for(i=0; bOnce && i<nOrderBy; i++){
3668 if( MASKBIT(i) & obSat ) continue;
3669 pOBExpr = sqlite3ExprSkipCollate(pOrderBy->a[i].pExpr);
3670 testcase( wctrlFlags & WHERE_GROUPBY );
3671 testcase( wctrlFlags & WHERE_DISTINCTBY );
3672 if( (wctrlFlags & (WHERE_GROUPBY|WHERE_DISTINCTBY))==0 ) bOnce = 0;
3673 if( iColumn>=(-1) ){
3674 if( pOBExpr->op!=TK_COLUMN ) continue;
3675 if( pOBExpr->iTable!=iCur ) continue;
3676 if( pOBExpr->iColumn!=iColumn ) continue;
3677 }else{
3678 if( sqlite3ExprCompare(0,
3679 pOBExpr,pIndex->aColExpr->a[j].pExpr,iCur) ){
3680 continue;
3683 if( iColumn>=0 ){
3684 pColl = sqlite3ExprCollSeq(pWInfo->pParse, pOrderBy->a[i].pExpr);
3685 if( !pColl ) pColl = db->pDfltColl;
3686 if( sqlite3StrICmp(pColl->zName, pIndex->azColl[j])!=0 ) continue;
3688 pLoop->u.btree.nIdxCol = j+1;
3689 isMatch = 1;
3690 break;
3692 if( isMatch && (wctrlFlags & WHERE_GROUPBY)==0 ){
3693 /* Make sure the sort order is compatible in an ORDER BY clause.
3694 ** Sort order is irrelevant for a GROUP BY clause. */
3695 if( revSet ){
3696 if( (rev ^ revIdx)!=pOrderBy->a[i].sortOrder ) isMatch = 0;
3697 }else{
3698 rev = revIdx ^ pOrderBy->a[i].sortOrder;
3699 if( rev ) *pRevMask |= MASKBIT(iLoop);
3700 revSet = 1;
3703 if( isMatch ){
3704 if( iColumn==XN_ROWID ){
3705 testcase( distinctColumns==0 );
3706 distinctColumns = 1;
3708 obSat |= MASKBIT(i);
3709 }else{
3710 /* No match found */
3711 if( j==0 || j<nKeyCol ){
3712 testcase( isOrderDistinct!=0 );
3713 isOrderDistinct = 0;
3715 break;
3717 } /* end Loop over all index columns */
3718 if( distinctColumns ){
3719 testcase( isOrderDistinct==0 );
3720 isOrderDistinct = 1;
3722 } /* end-if not one-row */
3724 /* Mark off any other ORDER BY terms that reference pLoop */
3725 if( isOrderDistinct ){
3726 orderDistinctMask |= pLoop->maskSelf;
3727 for(i=0; i<nOrderBy; i++){
3728 Expr *p;
3729 Bitmask mTerm;
3730 if( MASKBIT(i) & obSat ) continue;
3731 p = pOrderBy->a[i].pExpr;
3732 mTerm = sqlite3WhereExprUsage(&pWInfo->sMaskSet,p);
3733 if( mTerm==0 && !sqlite3ExprIsConstant(p) ) continue;
3734 if( (mTerm&~orderDistinctMask)==0 ){
3735 obSat |= MASKBIT(i);
3739 } /* End the loop over all WhereLoops from outer-most down to inner-most */
3740 if( obSat==obDone ) return (i8)nOrderBy;
3741 if( !isOrderDistinct ){
3742 for(i=nOrderBy-1; i>0; i--){
3743 Bitmask m = MASKBIT(i) - 1;
3744 if( (obSat&m)==m ) return i;
3746 return 0;
3748 return -1;
3753 ** If the WHERE_GROUPBY flag is set in the mask passed to sqlite3WhereBegin(),
3754 ** the planner assumes that the specified pOrderBy list is actually a GROUP
3755 ** BY clause - and so any order that groups rows as required satisfies the
3756 ** request.
3758 ** Normally, in this case it is not possible for the caller to determine
3759 ** whether or not the rows are really being delivered in sorted order, or
3760 ** just in some other order that provides the required grouping. However,
3761 ** if the WHERE_SORTBYGROUP flag is also passed to sqlite3WhereBegin(), then
3762 ** this function may be called on the returned WhereInfo object. It returns
3763 ** true if the rows really will be sorted in the specified order, or false
3764 ** otherwise.
3766 ** For example, assuming:
3768 ** CREATE INDEX i1 ON t1(x, Y);
3770 ** then
3772 ** SELECT * FROM t1 GROUP BY x,y ORDER BY x,y; -- IsSorted()==1
3773 ** SELECT * FROM t1 GROUP BY y,x ORDER BY y,x; -- IsSorted()==0
3775 int sqlite3WhereIsSorted(WhereInfo *pWInfo){
3776 assert( pWInfo->wctrlFlags & WHERE_GROUPBY );
3777 assert( pWInfo->wctrlFlags & WHERE_SORTBYGROUP );
3778 return pWInfo->sorted;
3781 #ifdef WHERETRACE_ENABLED
3782 /* For debugging use only: */
3783 static const char *wherePathName(WherePath *pPath, int nLoop, WhereLoop *pLast){
3784 static char zName[65];
3785 int i;
3786 for(i=0; i<nLoop; i++){ zName[i] = pPath->aLoop[i]->cId; }
3787 if( pLast ) zName[i++] = pLast->cId;
3788 zName[i] = 0;
3789 return zName;
3791 #endif
3794 ** Return the cost of sorting nRow rows, assuming that the keys have
3795 ** nOrderby columns and that the first nSorted columns are already in
3796 ** order.
3798 static LogEst whereSortingCost(
3799 WhereInfo *pWInfo,
3800 LogEst nRow,
3801 int nOrderBy,
3802 int nSorted
3804 /* TUNING: Estimated cost of a full external sort, where N is
3805 ** the number of rows to sort is:
3807 ** cost = (3.0 * N * log(N)).
3809 ** Or, if the order-by clause has X terms but only the last Y
3810 ** terms are out of order, then block-sorting will reduce the
3811 ** sorting cost to:
3813 ** cost = (3.0 * N * log(N)) * (Y/X)
3815 ** The (Y/X) term is implemented using stack variable rScale
3816 ** below. */
3817 LogEst rScale, rSortCost;
3818 assert( nOrderBy>0 && 66==sqlite3LogEst(100) );
3819 rScale = sqlite3LogEst((nOrderBy-nSorted)*100/nOrderBy) - 66;
3820 rSortCost = nRow + rScale + 16;
3822 /* Multiple by log(M) where M is the number of output rows.
3823 ** Use the LIMIT for M if it is smaller */
3824 if( (pWInfo->wctrlFlags & WHERE_USE_LIMIT)!=0 && pWInfo->iLimit<nRow ){
3825 nRow = pWInfo->iLimit;
3827 rSortCost += estLog(nRow);
3828 return rSortCost;
3832 ** Given the list of WhereLoop objects at pWInfo->pLoops, this routine
3833 ** attempts to find the lowest cost path that visits each WhereLoop
3834 ** once. This path is then loaded into the pWInfo->a[].pWLoop fields.
3836 ** Assume that the total number of output rows that will need to be sorted
3837 ** will be nRowEst (in the 10*log2 representation). Or, ignore sorting
3838 ** costs if nRowEst==0.
3840 ** Return SQLITE_OK on success or SQLITE_NOMEM of a memory allocation
3841 ** error occurs.
3843 static int wherePathSolver(WhereInfo *pWInfo, LogEst nRowEst){
3844 int mxChoice; /* Maximum number of simultaneous paths tracked */
3845 int nLoop; /* Number of terms in the join */
3846 Parse *pParse; /* Parsing context */
3847 sqlite3 *db; /* The database connection */
3848 int iLoop; /* Loop counter over the terms of the join */
3849 int ii, jj; /* Loop counters */
3850 int mxI = 0; /* Index of next entry to replace */
3851 int nOrderBy; /* Number of ORDER BY clause terms */
3852 LogEst mxCost = 0; /* Maximum cost of a set of paths */
3853 LogEst mxUnsorted = 0; /* Maximum unsorted cost of a set of path */
3854 int nTo, nFrom; /* Number of valid entries in aTo[] and aFrom[] */
3855 WherePath *aFrom; /* All nFrom paths at the previous level */
3856 WherePath *aTo; /* The nTo best paths at the current level */
3857 WherePath *pFrom; /* An element of aFrom[] that we are working on */
3858 WherePath *pTo; /* An element of aTo[] that we are working on */
3859 WhereLoop *pWLoop; /* One of the WhereLoop objects */
3860 WhereLoop **pX; /* Used to divy up the pSpace memory */
3861 LogEst *aSortCost = 0; /* Sorting and partial sorting costs */
3862 char *pSpace; /* Temporary memory used by this routine */
3863 int nSpace; /* Bytes of space allocated at pSpace */
3865 pParse = pWInfo->pParse;
3866 db = pParse->db;
3867 nLoop = pWInfo->nLevel;
3868 /* TUNING: For simple queries, only the best path is tracked.
3869 ** For 2-way joins, the 5 best paths are followed.
3870 ** For joins of 3 or more tables, track the 10 best paths */
3871 mxChoice = (nLoop<=1) ? 1 : (nLoop==2 ? 5 : 10);
3872 assert( nLoop<=pWInfo->pTabList->nSrc );
3873 WHERETRACE(0x002, ("---- begin solver. (nRowEst=%d)\n", nRowEst));
3875 /* If nRowEst is zero and there is an ORDER BY clause, ignore it. In this
3876 ** case the purpose of this call is to estimate the number of rows returned
3877 ** by the overall query. Once this estimate has been obtained, the caller
3878 ** will invoke this function a second time, passing the estimate as the
3879 ** nRowEst parameter. */
3880 if( pWInfo->pOrderBy==0 || nRowEst==0 ){
3881 nOrderBy = 0;
3882 }else{
3883 nOrderBy = pWInfo->pOrderBy->nExpr;
3886 /* Allocate and initialize space for aTo, aFrom and aSortCost[] */
3887 nSpace = (sizeof(WherePath)+sizeof(WhereLoop*)*nLoop)*mxChoice*2;
3888 nSpace += sizeof(LogEst) * nOrderBy;
3889 pSpace = sqlite3DbMallocRawNN(db, nSpace);
3890 if( pSpace==0 ) return SQLITE_NOMEM_BKPT;
3891 aTo = (WherePath*)pSpace;
3892 aFrom = aTo+mxChoice;
3893 memset(aFrom, 0, sizeof(aFrom[0]));
3894 pX = (WhereLoop**)(aFrom+mxChoice);
3895 for(ii=mxChoice*2, pFrom=aTo; ii>0; ii--, pFrom++, pX += nLoop){
3896 pFrom->aLoop = pX;
3898 if( nOrderBy ){
3899 /* If there is an ORDER BY clause and it is not being ignored, set up
3900 ** space for the aSortCost[] array. Each element of the aSortCost array
3901 ** is either zero - meaning it has not yet been initialized - or the
3902 ** cost of sorting nRowEst rows of data where the first X terms of
3903 ** the ORDER BY clause are already in order, where X is the array
3904 ** index. */
3905 aSortCost = (LogEst*)pX;
3906 memset(aSortCost, 0, sizeof(LogEst) * nOrderBy);
3908 assert( aSortCost==0 || &pSpace[nSpace]==(char*)&aSortCost[nOrderBy] );
3909 assert( aSortCost!=0 || &pSpace[nSpace]==(char*)pX );
3911 /* Seed the search with a single WherePath containing zero WhereLoops.
3913 ** TUNING: Do not let the number of iterations go above 28. If the cost
3914 ** of computing an automatic index is not paid back within the first 28
3915 ** rows, then do not use the automatic index. */
3916 aFrom[0].nRow = MIN(pParse->nQueryLoop, 48); assert( 48==sqlite3LogEst(28) );
3917 nFrom = 1;
3918 assert( aFrom[0].isOrdered==0 );
3919 if( nOrderBy ){
3920 /* If nLoop is zero, then there are no FROM terms in the query. Since
3921 ** in this case the query may return a maximum of one row, the results
3922 ** are already in the requested order. Set isOrdered to nOrderBy to
3923 ** indicate this. Or, if nLoop is greater than zero, set isOrdered to
3924 ** -1, indicating that the result set may or may not be ordered,
3925 ** depending on the loops added to the current plan. */
3926 aFrom[0].isOrdered = nLoop>0 ? -1 : nOrderBy;
3929 /* Compute successively longer WherePaths using the previous generation
3930 ** of WherePaths as the basis for the next. Keep track of the mxChoice
3931 ** best paths at each generation */
3932 for(iLoop=0; iLoop<nLoop; iLoop++){
3933 nTo = 0;
3934 for(ii=0, pFrom=aFrom; ii<nFrom; ii++, pFrom++){
3935 for(pWLoop=pWInfo->pLoops; pWLoop; pWLoop=pWLoop->pNextLoop){
3936 LogEst nOut; /* Rows visited by (pFrom+pWLoop) */
3937 LogEst rCost; /* Cost of path (pFrom+pWLoop) */
3938 LogEst rUnsorted; /* Unsorted cost of (pFrom+pWLoop) */
3939 i8 isOrdered = pFrom->isOrdered; /* isOrdered for (pFrom+pWLoop) */
3940 Bitmask maskNew; /* Mask of src visited by (..) */
3941 Bitmask revMask = 0; /* Mask of rev-order loops for (..) */
3943 if( (pWLoop->prereq & ~pFrom->maskLoop)!=0 ) continue;
3944 if( (pWLoop->maskSelf & pFrom->maskLoop)!=0 ) continue;
3945 if( (pWLoop->wsFlags & WHERE_AUTO_INDEX)!=0 && pFrom->nRow<10 ){
3946 /* Do not use an automatic index if the this loop is expected
3947 ** to run less than 2 times. */
3948 assert( 10==sqlite3LogEst(2) );
3949 continue;
3951 /* At this point, pWLoop is a candidate to be the next loop.
3952 ** Compute its cost */
3953 rUnsorted = sqlite3LogEstAdd(pWLoop->rSetup,pWLoop->rRun + pFrom->nRow);
3954 rUnsorted = sqlite3LogEstAdd(rUnsorted, pFrom->rUnsorted);
3955 nOut = pFrom->nRow + pWLoop->nOut;
3956 maskNew = pFrom->maskLoop | pWLoop->maskSelf;
3957 if( isOrdered<0 ){
3958 isOrdered = wherePathSatisfiesOrderBy(pWInfo,
3959 pWInfo->pOrderBy, pFrom, pWInfo->wctrlFlags,
3960 iLoop, pWLoop, &revMask);
3961 }else{
3962 revMask = pFrom->revLoop;
3964 if( isOrdered>=0 && isOrdered<nOrderBy ){
3965 if( aSortCost[isOrdered]==0 ){
3966 aSortCost[isOrdered] = whereSortingCost(
3967 pWInfo, nRowEst, nOrderBy, isOrdered
3970 rCost = sqlite3LogEstAdd(rUnsorted, aSortCost[isOrdered]);
3972 WHERETRACE(0x002,
3973 ("---- sort cost=%-3d (%d/%d) increases cost %3d to %-3d\n",
3974 aSortCost[isOrdered], (nOrderBy-isOrdered), nOrderBy,
3975 rUnsorted, rCost));
3976 }else{
3977 rCost = rUnsorted;
3978 rUnsorted -= 2; /* TUNING: Slight bias in favor of no-sort plans */
3981 /* Check to see if pWLoop should be added to the set of
3982 ** mxChoice best-so-far paths.
3984 ** First look for an existing path among best-so-far paths
3985 ** that covers the same set of loops and has the same isOrdered
3986 ** setting as the current path candidate.
3988 ** The term "((pTo->isOrdered^isOrdered)&0x80)==0" is equivalent
3989 ** to (pTo->isOrdered==(-1))==(isOrdered==(-1))" for the range
3990 ** of legal values for isOrdered, -1..64.
3992 for(jj=0, pTo=aTo; jj<nTo; jj++, pTo++){
3993 if( pTo->maskLoop==maskNew
3994 && ((pTo->isOrdered^isOrdered)&0x80)==0
3996 testcase( jj==nTo-1 );
3997 break;
4000 if( jj>=nTo ){
4001 /* None of the existing best-so-far paths match the candidate. */
4002 if( nTo>=mxChoice
4003 && (rCost>mxCost || (rCost==mxCost && rUnsorted>=mxUnsorted))
4005 /* The current candidate is no better than any of the mxChoice
4006 ** paths currently in the best-so-far buffer. So discard
4007 ** this candidate as not viable. */
4008 #ifdef WHERETRACE_ENABLED /* 0x4 */
4009 if( sqlite3WhereTrace&0x4 ){
4010 sqlite3DebugPrintf("Skip %s cost=%-3d,%3d,%3d order=%c\n",
4011 wherePathName(pFrom, iLoop, pWLoop), rCost, nOut, rUnsorted,
4012 isOrdered>=0 ? isOrdered+'0' : '?');
4014 #endif
4015 continue;
4017 /* If we reach this points it means that the new candidate path
4018 ** needs to be added to the set of best-so-far paths. */
4019 if( nTo<mxChoice ){
4020 /* Increase the size of the aTo set by one */
4021 jj = nTo++;
4022 }else{
4023 /* New path replaces the prior worst to keep count below mxChoice */
4024 jj = mxI;
4026 pTo = &aTo[jj];
4027 #ifdef WHERETRACE_ENABLED /* 0x4 */
4028 if( sqlite3WhereTrace&0x4 ){
4029 sqlite3DebugPrintf("New %s cost=%-3d,%3d,%3d order=%c\n",
4030 wherePathName(pFrom, iLoop, pWLoop), rCost, nOut, rUnsorted,
4031 isOrdered>=0 ? isOrdered+'0' : '?');
4033 #endif
4034 }else{
4035 /* Control reaches here if best-so-far path pTo=aTo[jj] covers the
4036 ** same set of loops and has the same isOrdered setting as the
4037 ** candidate path. Check to see if the candidate should replace
4038 ** pTo or if the candidate should be skipped.
4040 ** The conditional is an expanded vector comparison equivalent to:
4041 ** (pTo->rCost,pTo->nRow,pTo->rUnsorted) <= (rCost,nOut,rUnsorted)
4043 if( pTo->rCost<rCost
4044 || (pTo->rCost==rCost
4045 && (pTo->nRow<nOut
4046 || (pTo->nRow==nOut && pTo->rUnsorted<=rUnsorted)
4050 #ifdef WHERETRACE_ENABLED /* 0x4 */
4051 if( sqlite3WhereTrace&0x4 ){
4052 sqlite3DebugPrintf(
4053 "Skip %s cost=%-3d,%3d,%3d order=%c",
4054 wherePathName(pFrom, iLoop, pWLoop), rCost, nOut, rUnsorted,
4055 isOrdered>=0 ? isOrdered+'0' : '?');
4056 sqlite3DebugPrintf(" vs %s cost=%-3d,%3d,%3d order=%c\n",
4057 wherePathName(pTo, iLoop+1, 0), pTo->rCost, pTo->nRow,
4058 pTo->rUnsorted, pTo->isOrdered>=0 ? pTo->isOrdered+'0' : '?');
4060 #endif
4061 /* Discard the candidate path from further consideration */
4062 testcase( pTo->rCost==rCost );
4063 continue;
4065 testcase( pTo->rCost==rCost+1 );
4066 /* Control reaches here if the candidate path is better than the
4067 ** pTo path. Replace pTo with the candidate. */
4068 #ifdef WHERETRACE_ENABLED /* 0x4 */
4069 if( sqlite3WhereTrace&0x4 ){
4070 sqlite3DebugPrintf(
4071 "Update %s cost=%-3d,%3d,%3d order=%c",
4072 wherePathName(pFrom, iLoop, pWLoop), rCost, nOut, rUnsorted,
4073 isOrdered>=0 ? isOrdered+'0' : '?');
4074 sqlite3DebugPrintf(" was %s cost=%-3d,%3d,%3d order=%c\n",
4075 wherePathName(pTo, iLoop+1, 0), pTo->rCost, pTo->nRow,
4076 pTo->rUnsorted, pTo->isOrdered>=0 ? pTo->isOrdered+'0' : '?');
4078 #endif
4080 /* pWLoop is a winner. Add it to the set of best so far */
4081 pTo->maskLoop = pFrom->maskLoop | pWLoop->maskSelf;
4082 pTo->revLoop = revMask;
4083 pTo->nRow = nOut;
4084 pTo->rCost = rCost;
4085 pTo->rUnsorted = rUnsorted;
4086 pTo->isOrdered = isOrdered;
4087 memcpy(pTo->aLoop, pFrom->aLoop, sizeof(WhereLoop*)*iLoop);
4088 pTo->aLoop[iLoop] = pWLoop;
4089 if( nTo>=mxChoice ){
4090 mxI = 0;
4091 mxCost = aTo[0].rCost;
4092 mxUnsorted = aTo[0].nRow;
4093 for(jj=1, pTo=&aTo[1]; jj<mxChoice; jj++, pTo++){
4094 if( pTo->rCost>mxCost
4095 || (pTo->rCost==mxCost && pTo->rUnsorted>mxUnsorted)
4097 mxCost = pTo->rCost;
4098 mxUnsorted = pTo->rUnsorted;
4099 mxI = jj;
4106 #ifdef WHERETRACE_ENABLED /* >=2 */
4107 if( sqlite3WhereTrace & 0x02 ){
4108 sqlite3DebugPrintf("---- after round %d ----\n", iLoop);
4109 for(ii=0, pTo=aTo; ii<nTo; ii++, pTo++){
4110 sqlite3DebugPrintf(" %s cost=%-3d nrow=%-3d order=%c",
4111 wherePathName(pTo, iLoop+1, 0), pTo->rCost, pTo->nRow,
4112 pTo->isOrdered>=0 ? (pTo->isOrdered+'0') : '?');
4113 if( pTo->isOrdered>0 ){
4114 sqlite3DebugPrintf(" rev=0x%llx\n", pTo->revLoop);
4115 }else{
4116 sqlite3DebugPrintf("\n");
4120 #endif
4122 /* Swap the roles of aFrom and aTo for the next generation */
4123 pFrom = aTo;
4124 aTo = aFrom;
4125 aFrom = pFrom;
4126 nFrom = nTo;
4129 if( nFrom==0 ){
4130 sqlite3ErrorMsg(pParse, "no query solution");
4131 sqlite3DbFreeNN(db, pSpace);
4132 return SQLITE_ERROR;
4135 /* Find the lowest cost path. pFrom will be left pointing to that path */
4136 pFrom = aFrom;
4137 for(ii=1; ii<nFrom; ii++){
4138 if( pFrom->rCost>aFrom[ii].rCost ) pFrom = &aFrom[ii];
4140 assert( pWInfo->nLevel==nLoop );
4141 /* Load the lowest cost path into pWInfo */
4142 for(iLoop=0; iLoop<nLoop; iLoop++){
4143 WhereLevel *pLevel = pWInfo->a + iLoop;
4144 pLevel->pWLoop = pWLoop = pFrom->aLoop[iLoop];
4145 pLevel->iFrom = pWLoop->iTab;
4146 pLevel->iTabCur = pWInfo->pTabList->a[pLevel->iFrom].iCursor;
4148 if( (pWInfo->wctrlFlags & WHERE_WANT_DISTINCT)!=0
4149 && (pWInfo->wctrlFlags & WHERE_DISTINCTBY)==0
4150 && pWInfo->eDistinct==WHERE_DISTINCT_NOOP
4151 && nRowEst
4153 Bitmask notUsed;
4154 int rc = wherePathSatisfiesOrderBy(pWInfo, pWInfo->pResultSet, pFrom,
4155 WHERE_DISTINCTBY, nLoop-1, pFrom->aLoop[nLoop-1], &notUsed);
4156 if( rc==pWInfo->pResultSet->nExpr ){
4157 pWInfo->eDistinct = WHERE_DISTINCT_ORDERED;
4160 if( pWInfo->pOrderBy ){
4161 if( pWInfo->wctrlFlags & WHERE_DISTINCTBY ){
4162 if( pFrom->isOrdered==pWInfo->pOrderBy->nExpr ){
4163 pWInfo->eDistinct = WHERE_DISTINCT_ORDERED;
4165 }else{
4166 pWInfo->nOBSat = pFrom->isOrdered;
4167 pWInfo->revMask = pFrom->revLoop;
4168 if( pWInfo->nOBSat<=0 ){
4169 pWInfo->nOBSat = 0;
4170 if( nLoop>0 ){
4171 u32 wsFlags = pFrom->aLoop[nLoop-1]->wsFlags;
4172 if( (wsFlags & WHERE_ONEROW)==0
4173 && (wsFlags&(WHERE_IPK|WHERE_COLUMN_IN))!=(WHERE_IPK|WHERE_COLUMN_IN)
4175 Bitmask m = 0;
4176 int rc = wherePathSatisfiesOrderBy(pWInfo, pWInfo->pOrderBy, pFrom,
4177 WHERE_ORDERBY_LIMIT, nLoop-1, pFrom->aLoop[nLoop-1], &m);
4178 testcase( wsFlags & WHERE_IPK );
4179 testcase( wsFlags & WHERE_COLUMN_IN );
4180 if( rc==pWInfo->pOrderBy->nExpr ){
4181 pWInfo->bOrderedInnerLoop = 1;
4182 pWInfo->revMask = m;
4188 if( (pWInfo->wctrlFlags & WHERE_SORTBYGROUP)
4189 && pWInfo->nOBSat==pWInfo->pOrderBy->nExpr && nLoop>0
4191 Bitmask revMask = 0;
4192 int nOrder = wherePathSatisfiesOrderBy(pWInfo, pWInfo->pOrderBy,
4193 pFrom, 0, nLoop-1, pFrom->aLoop[nLoop-1], &revMask
4195 assert( pWInfo->sorted==0 );
4196 if( nOrder==pWInfo->pOrderBy->nExpr ){
4197 pWInfo->sorted = 1;
4198 pWInfo->revMask = revMask;
4204 pWInfo->nRowOut = pFrom->nRow;
4206 /* Free temporary memory and return success */
4207 sqlite3DbFreeNN(db, pSpace);
4208 return SQLITE_OK;
4212 ** Most queries use only a single table (they are not joins) and have
4213 ** simple == constraints against indexed fields. This routine attempts
4214 ** to plan those simple cases using much less ceremony than the
4215 ** general-purpose query planner, and thereby yield faster sqlite3_prepare()
4216 ** times for the common case.
4218 ** Return non-zero on success, if this query can be handled by this
4219 ** no-frills query planner. Return zero if this query needs the
4220 ** general-purpose query planner.
4222 static int whereShortCut(WhereLoopBuilder *pBuilder){
4223 WhereInfo *pWInfo;
4224 struct SrcList_item *pItem;
4225 WhereClause *pWC;
4226 WhereTerm *pTerm;
4227 WhereLoop *pLoop;
4228 int iCur;
4229 int j;
4230 Table *pTab;
4231 Index *pIdx;
4233 pWInfo = pBuilder->pWInfo;
4234 if( pWInfo->wctrlFlags & WHERE_OR_SUBCLAUSE ) return 0;
4235 assert( pWInfo->pTabList->nSrc>=1 );
4236 pItem = pWInfo->pTabList->a;
4237 pTab = pItem->pTab;
4238 if( IsVirtual(pTab) ) return 0;
4239 if( pItem->fg.isIndexedBy ) return 0;
4240 iCur = pItem->iCursor;
4241 pWC = &pWInfo->sWC;
4242 pLoop = pBuilder->pNew;
4243 pLoop->wsFlags = 0;
4244 pLoop->nSkip = 0;
4245 pTerm = sqlite3WhereFindTerm(pWC, iCur, -1, 0, WO_EQ|WO_IS, 0);
4246 if( pTerm ){
4247 testcase( pTerm->eOperator & WO_IS );
4248 pLoop->wsFlags = WHERE_COLUMN_EQ|WHERE_IPK|WHERE_ONEROW;
4249 pLoop->aLTerm[0] = pTerm;
4250 pLoop->nLTerm = 1;
4251 pLoop->u.btree.nEq = 1;
4252 /* TUNING: Cost of a rowid lookup is 10 */
4253 pLoop->rRun = 33; /* 33==sqlite3LogEst(10) */
4254 }else{
4255 for(pIdx=pTab->pIndex; pIdx; pIdx=pIdx->pNext){
4256 int opMask;
4257 assert( pLoop->aLTermSpace==pLoop->aLTerm );
4258 if( !IsUniqueIndex(pIdx)
4259 || pIdx->pPartIdxWhere!=0
4260 || pIdx->nKeyCol>ArraySize(pLoop->aLTermSpace)
4261 ) continue;
4262 opMask = pIdx->uniqNotNull ? (WO_EQ|WO_IS) : WO_EQ;
4263 for(j=0; j<pIdx->nKeyCol; j++){
4264 pTerm = sqlite3WhereFindTerm(pWC, iCur, j, 0, opMask, pIdx);
4265 if( pTerm==0 ) break;
4266 testcase( pTerm->eOperator & WO_IS );
4267 pLoop->aLTerm[j] = pTerm;
4269 if( j!=pIdx->nKeyCol ) continue;
4270 pLoop->wsFlags = WHERE_COLUMN_EQ|WHERE_ONEROW|WHERE_INDEXED;
4271 if( pIdx->isCovering || (pItem->colUsed & ~columnsInIndex(pIdx))==0 ){
4272 pLoop->wsFlags |= WHERE_IDX_ONLY;
4274 pLoop->nLTerm = j;
4275 pLoop->u.btree.nEq = j;
4276 pLoop->u.btree.pIndex = pIdx;
4277 /* TUNING: Cost of a unique index lookup is 15 */
4278 pLoop->rRun = 39; /* 39==sqlite3LogEst(15) */
4279 break;
4282 if( pLoop->wsFlags ){
4283 pLoop->nOut = (LogEst)1;
4284 pWInfo->a[0].pWLoop = pLoop;
4285 assert( pWInfo->sMaskSet.n==1 && iCur==pWInfo->sMaskSet.ix[0] );
4286 pLoop->maskSelf = 1; /* sqlite3WhereGetMask(&pWInfo->sMaskSet, iCur); */
4287 pWInfo->a[0].iTabCur = iCur;
4288 pWInfo->nRowOut = 1;
4289 if( pWInfo->pOrderBy ) pWInfo->nOBSat = pWInfo->pOrderBy->nExpr;
4290 if( pWInfo->wctrlFlags & WHERE_WANT_DISTINCT ){
4291 pWInfo->eDistinct = WHERE_DISTINCT_UNIQUE;
4293 #ifdef SQLITE_DEBUG
4294 pLoop->cId = '0';
4295 #endif
4296 return 1;
4298 return 0;
4302 ** Helper function for exprIsDeterministic().
4304 static int exprNodeIsDeterministic(Walker *pWalker, Expr *pExpr){
4305 if( pExpr->op==TK_FUNCTION && ExprHasProperty(pExpr, EP_ConstFunc)==0 ){
4306 pWalker->eCode = 0;
4307 return WRC_Abort;
4309 return WRC_Continue;
4313 ** Return true if the expression contains no non-deterministic SQL
4314 ** functions. Do not consider non-deterministic SQL functions that are
4315 ** part of sub-select statements.
4317 static int exprIsDeterministic(Expr *p){
4318 Walker w;
4319 memset(&w, 0, sizeof(w));
4320 w.eCode = 1;
4321 w.xExprCallback = exprNodeIsDeterministic;
4322 sqlite3WalkExpr(&w, p);
4323 return w.eCode;
4327 ** Generate the beginning of the loop used for WHERE clause processing.
4328 ** The return value is a pointer to an opaque structure that contains
4329 ** information needed to terminate the loop. Later, the calling routine
4330 ** should invoke sqlite3WhereEnd() with the return value of this function
4331 ** in order to complete the WHERE clause processing.
4333 ** If an error occurs, this routine returns NULL.
4335 ** The basic idea is to do a nested loop, one loop for each table in
4336 ** the FROM clause of a select. (INSERT and UPDATE statements are the
4337 ** same as a SELECT with only a single table in the FROM clause.) For
4338 ** example, if the SQL is this:
4340 ** SELECT * FROM t1, t2, t3 WHERE ...;
4342 ** Then the code generated is conceptually like the following:
4344 ** foreach row1 in t1 do \ Code generated
4345 ** foreach row2 in t2 do |-- by sqlite3WhereBegin()
4346 ** foreach row3 in t3 do /
4347 ** ...
4348 ** end \ Code generated
4349 ** end |-- by sqlite3WhereEnd()
4350 ** end /
4352 ** Note that the loops might not be nested in the order in which they
4353 ** appear in the FROM clause if a different order is better able to make
4354 ** use of indices. Note also that when the IN operator appears in
4355 ** the WHERE clause, it might result in additional nested loops for
4356 ** scanning through all values on the right-hand side of the IN.
4358 ** There are Btree cursors associated with each table. t1 uses cursor
4359 ** number pTabList->a[0].iCursor. t2 uses the cursor pTabList->a[1].iCursor.
4360 ** And so forth. This routine generates code to open those VDBE cursors
4361 ** and sqlite3WhereEnd() generates the code to close them.
4363 ** The code that sqlite3WhereBegin() generates leaves the cursors named
4364 ** in pTabList pointing at their appropriate entries. The [...] code
4365 ** can use OP_Column and OP_Rowid opcodes on these cursors to extract
4366 ** data from the various tables of the loop.
4368 ** If the WHERE clause is empty, the foreach loops must each scan their
4369 ** entire tables. Thus a three-way join is an O(N^3) operation. But if
4370 ** the tables have indices and there are terms in the WHERE clause that
4371 ** refer to those indices, a complete table scan can be avoided and the
4372 ** code will run much faster. Most of the work of this routine is checking
4373 ** to see if there are indices that can be used to speed up the loop.
4375 ** Terms of the WHERE clause are also used to limit which rows actually
4376 ** make it to the "..." in the middle of the loop. After each "foreach",
4377 ** terms of the WHERE clause that use only terms in that loop and outer
4378 ** loops are evaluated and if false a jump is made around all subsequent
4379 ** inner loops (or around the "..." if the test occurs within the inner-
4380 ** most loop)
4382 ** OUTER JOINS
4384 ** An outer join of tables t1 and t2 is conceptally coded as follows:
4386 ** foreach row1 in t1 do
4387 ** flag = 0
4388 ** foreach row2 in t2 do
4389 ** start:
4390 ** ...
4391 ** flag = 1
4392 ** end
4393 ** if flag==0 then
4394 ** move the row2 cursor to a null row
4395 ** goto start
4396 ** fi
4397 ** end
4399 ** ORDER BY CLAUSE PROCESSING
4401 ** pOrderBy is a pointer to the ORDER BY clause (or the GROUP BY clause
4402 ** if the WHERE_GROUPBY flag is set in wctrlFlags) of a SELECT statement
4403 ** if there is one. If there is no ORDER BY clause or if this routine
4404 ** is called from an UPDATE or DELETE statement, then pOrderBy is NULL.
4406 ** The iIdxCur parameter is the cursor number of an index. If
4407 ** WHERE_OR_SUBCLAUSE is set, iIdxCur is the cursor number of an index
4408 ** to use for OR clause processing. The WHERE clause should use this
4409 ** specific cursor. If WHERE_ONEPASS_DESIRED is set, then iIdxCur is
4410 ** the first cursor in an array of cursors for all indices. iIdxCur should
4411 ** be used to compute the appropriate cursor depending on which index is
4412 ** used.
4414 WhereInfo *sqlite3WhereBegin(
4415 Parse *pParse, /* The parser context */
4416 SrcList *pTabList, /* FROM clause: A list of all tables to be scanned */
4417 Expr *pWhere, /* The WHERE clause */
4418 ExprList *pOrderBy, /* An ORDER BY (or GROUP BY) clause, or NULL */
4419 ExprList *pResultSet, /* Query result set. Req'd for DISTINCT */
4420 u16 wctrlFlags, /* The WHERE_* flags defined in sqliteInt.h */
4421 int iAuxArg /* If WHERE_OR_SUBCLAUSE is set, index cursor number
4422 ** If WHERE_USE_LIMIT, then the limit amount */
4424 int nByteWInfo; /* Num. bytes allocated for WhereInfo struct */
4425 int nTabList; /* Number of elements in pTabList */
4426 WhereInfo *pWInfo; /* Will become the return value of this function */
4427 Vdbe *v = pParse->pVdbe; /* The virtual database engine */
4428 Bitmask notReady; /* Cursors that are not yet positioned */
4429 WhereLoopBuilder sWLB; /* The WhereLoop builder */
4430 WhereMaskSet *pMaskSet; /* The expression mask set */
4431 WhereLevel *pLevel; /* A single level in pWInfo->a[] */
4432 WhereLoop *pLoop; /* Pointer to a single WhereLoop object */
4433 int ii; /* Loop counter */
4434 sqlite3 *db; /* Database connection */
4435 int rc; /* Return code */
4436 u8 bFordelete = 0; /* OPFLAG_FORDELETE or zero, as appropriate */
4438 assert( (wctrlFlags & WHERE_ONEPASS_MULTIROW)==0 || (
4439 (wctrlFlags & WHERE_ONEPASS_DESIRED)!=0
4440 && (wctrlFlags & WHERE_OR_SUBCLAUSE)==0
4443 /* Only one of WHERE_OR_SUBCLAUSE or WHERE_USE_LIMIT */
4444 assert( (wctrlFlags & WHERE_OR_SUBCLAUSE)==0
4445 || (wctrlFlags & WHERE_USE_LIMIT)==0 );
4447 /* Variable initialization */
4448 db = pParse->db;
4449 memset(&sWLB, 0, sizeof(sWLB));
4451 /* An ORDER/GROUP BY clause of more than 63 terms cannot be optimized */
4452 testcase( pOrderBy && pOrderBy->nExpr==BMS-1 );
4453 if( pOrderBy && pOrderBy->nExpr>=BMS ) pOrderBy = 0;
4454 sWLB.pOrderBy = pOrderBy;
4456 /* Disable the DISTINCT optimization if SQLITE_DistinctOpt is set via
4457 ** sqlite3_test_ctrl(SQLITE_TESTCTRL_OPTIMIZATIONS,...) */
4458 if( OptimizationDisabled(db, SQLITE_DistinctOpt) ){
4459 wctrlFlags &= ~WHERE_WANT_DISTINCT;
4462 /* The number of tables in the FROM clause is limited by the number of
4463 ** bits in a Bitmask
4465 testcase( pTabList->nSrc==BMS );
4466 if( pTabList->nSrc>BMS ){
4467 sqlite3ErrorMsg(pParse, "at most %d tables in a join", BMS);
4468 return 0;
4471 /* This function normally generates a nested loop for all tables in
4472 ** pTabList. But if the WHERE_OR_SUBCLAUSE flag is set, then we should
4473 ** only generate code for the first table in pTabList and assume that
4474 ** any cursors associated with subsequent tables are uninitialized.
4476 nTabList = (wctrlFlags & WHERE_OR_SUBCLAUSE) ? 1 : pTabList->nSrc;
4478 /* Allocate and initialize the WhereInfo structure that will become the
4479 ** return value. A single allocation is used to store the WhereInfo
4480 ** struct, the contents of WhereInfo.a[], the WhereClause structure
4481 ** and the WhereMaskSet structure. Since WhereClause contains an 8-byte
4482 ** field (type Bitmask) it must be aligned on an 8-byte boundary on
4483 ** some architectures. Hence the ROUND8() below.
4485 nByteWInfo = ROUND8(sizeof(WhereInfo)+(nTabList-1)*sizeof(WhereLevel));
4486 pWInfo = sqlite3DbMallocRawNN(db, nByteWInfo + sizeof(WhereLoop));
4487 if( db->mallocFailed ){
4488 sqlite3DbFree(db, pWInfo);
4489 pWInfo = 0;
4490 goto whereBeginError;
4492 pWInfo->pParse = pParse;
4493 pWInfo->pTabList = pTabList;
4494 pWInfo->pOrderBy = pOrderBy;
4495 pWInfo->pWhere = pWhere;
4496 pWInfo->pResultSet = pResultSet;
4497 pWInfo->aiCurOnePass[0] = pWInfo->aiCurOnePass[1] = -1;
4498 pWInfo->nLevel = nTabList;
4499 pWInfo->iBreak = pWInfo->iContinue = sqlite3VdbeMakeLabel(v);
4500 pWInfo->wctrlFlags = wctrlFlags;
4501 pWInfo->iLimit = iAuxArg;
4502 pWInfo->savedNQueryLoop = pParse->nQueryLoop;
4503 memset(&pWInfo->nOBSat, 0,
4504 offsetof(WhereInfo,sWC) - offsetof(WhereInfo,nOBSat));
4505 memset(&pWInfo->a[0], 0, sizeof(WhereLoop)+nTabList*sizeof(WhereLevel));
4506 assert( pWInfo->eOnePass==ONEPASS_OFF ); /* ONEPASS defaults to OFF */
4507 pMaskSet = &pWInfo->sMaskSet;
4508 sWLB.pWInfo = pWInfo;
4509 sWLB.pWC = &pWInfo->sWC;
4510 sWLB.pNew = (WhereLoop*)(((char*)pWInfo)+nByteWInfo);
4511 assert( EIGHT_BYTE_ALIGNMENT(sWLB.pNew) );
4512 whereLoopInit(sWLB.pNew);
4513 #ifdef SQLITE_DEBUG
4514 sWLB.pNew->cId = '*';
4515 #endif
4517 /* Split the WHERE clause into separate subexpressions where each
4518 ** subexpression is separated by an AND operator.
4520 initMaskSet(pMaskSet);
4521 sqlite3WhereClauseInit(&pWInfo->sWC, pWInfo);
4522 sqlite3WhereSplit(&pWInfo->sWC, pWhere, TK_AND);
4524 /* Special case: No FROM clause
4526 if( nTabList==0 ){
4527 if( pOrderBy ) pWInfo->nOBSat = pOrderBy->nExpr;
4528 if( wctrlFlags & WHERE_WANT_DISTINCT ){
4529 pWInfo->eDistinct = WHERE_DISTINCT_UNIQUE;
4533 /* Assign a bit from the bitmask to every term in the FROM clause.
4535 ** The N-th term of the FROM clause is assigned a bitmask of 1<<N.
4537 ** The rule of the previous sentence ensures thta if X is the bitmask for
4538 ** a table T, then X-1 is the bitmask for all other tables to the left of T.
4539 ** Knowing the bitmask for all tables to the left of a left join is
4540 ** important. Ticket #3015.
4542 ** Note that bitmasks are created for all pTabList->nSrc tables in
4543 ** pTabList, not just the first nTabList tables. nTabList is normally
4544 ** equal to pTabList->nSrc but might be shortened to 1 if the
4545 ** WHERE_OR_SUBCLAUSE flag is set.
4547 for(ii=0; ii<pTabList->nSrc; ii++){
4548 createMask(pMaskSet, pTabList->a[ii].iCursor);
4549 sqlite3WhereTabFuncArgs(pParse, &pTabList->a[ii], &pWInfo->sWC);
4551 #ifdef SQLITE_DEBUG
4553 Bitmask mx = 0;
4554 for(ii=0; ii<pTabList->nSrc; ii++){
4555 Bitmask m = sqlite3WhereGetMask(pMaskSet, pTabList->a[ii].iCursor);
4556 assert( m>=mx );
4557 mx = m;
4560 #endif
4562 /* Analyze all of the subexpressions. */
4563 sqlite3WhereExprAnalyze(pTabList, &pWInfo->sWC);
4564 if( db->mallocFailed ) goto whereBeginError;
4566 /* Special case: WHERE terms that do not refer to any tables in the join
4567 ** (constant expressions). Evaluate each such term, and jump over all the
4568 ** generated code if the result is not true.
4570 ** Do not do this if the expression contains non-deterministic functions
4571 ** that are not within a sub-select. This is not strictly required, but
4572 ** preserves SQLite's legacy behaviour in the following two cases:
4574 ** FROM ... WHERE random()>0; -- eval random() once per row
4575 ** FROM ... WHERE (SELECT random())>0; -- eval random() once overall
4577 for(ii=0; ii<sWLB.pWC->nTerm; ii++){
4578 WhereTerm *pT = &sWLB.pWC->a[ii];
4579 if( pT->prereqAll==0 && (nTabList==0 || exprIsDeterministic(pT->pExpr)) ){
4580 sqlite3ExprIfFalse(pParse, pT->pExpr, pWInfo->iBreak, SQLITE_JUMPIFNULL);
4581 pT->wtFlags |= TERM_CODED;
4585 if( wctrlFlags & WHERE_WANT_DISTINCT ){
4586 if( isDistinctRedundant(pParse, pTabList, &pWInfo->sWC, pResultSet) ){
4587 /* The DISTINCT marking is pointless. Ignore it. */
4588 pWInfo->eDistinct = WHERE_DISTINCT_UNIQUE;
4589 }else if( pOrderBy==0 ){
4590 /* Try to ORDER BY the result set to make distinct processing easier */
4591 pWInfo->wctrlFlags |= WHERE_DISTINCTBY;
4592 pWInfo->pOrderBy = pResultSet;
4596 /* Construct the WhereLoop objects */
4597 #if defined(WHERETRACE_ENABLED)
4598 if( sqlite3WhereTrace & 0xffff ){
4599 sqlite3DebugPrintf("*** Optimizer Start *** (wctrlFlags: 0x%x",wctrlFlags);
4600 if( wctrlFlags & WHERE_USE_LIMIT ){
4601 sqlite3DebugPrintf(", limit: %d", iAuxArg);
4603 sqlite3DebugPrintf(")\n");
4605 if( sqlite3WhereTrace & 0x100 ){ /* Display all terms of the WHERE clause */
4606 sqlite3WhereClausePrint(sWLB.pWC);
4608 #endif
4610 if( nTabList!=1 || whereShortCut(&sWLB)==0 ){
4611 rc = whereLoopAddAll(&sWLB);
4612 if( rc ) goto whereBeginError;
4614 #ifdef WHERETRACE_ENABLED
4615 if( sqlite3WhereTrace ){ /* Display all of the WhereLoop objects */
4616 WhereLoop *p;
4617 int i;
4618 static const char zLabel[] = "0123456789abcdefghijklmnopqrstuvwyxz"
4619 "ABCDEFGHIJKLMNOPQRSTUVWYXZ";
4620 for(p=pWInfo->pLoops, i=0; p; p=p->pNextLoop, i++){
4621 p->cId = zLabel[i%(sizeof(zLabel)-1)];
4622 whereLoopPrint(p, sWLB.pWC);
4625 #endif
4627 wherePathSolver(pWInfo, 0);
4628 if( db->mallocFailed ) goto whereBeginError;
4629 if( pWInfo->pOrderBy ){
4630 wherePathSolver(pWInfo, pWInfo->nRowOut+1);
4631 if( db->mallocFailed ) goto whereBeginError;
4634 if( pWInfo->pOrderBy==0 && (db->flags & SQLITE_ReverseOrder)!=0 ){
4635 pWInfo->revMask = ALLBITS;
4637 if( pParse->nErr || NEVER(db->mallocFailed) ){
4638 goto whereBeginError;
4640 #ifdef WHERETRACE_ENABLED
4641 if( sqlite3WhereTrace ){
4642 sqlite3DebugPrintf("---- Solution nRow=%d", pWInfo->nRowOut);
4643 if( pWInfo->nOBSat>0 ){
4644 sqlite3DebugPrintf(" ORDERBY=%d,0x%llx", pWInfo->nOBSat, pWInfo->revMask);
4646 switch( pWInfo->eDistinct ){
4647 case WHERE_DISTINCT_UNIQUE: {
4648 sqlite3DebugPrintf(" DISTINCT=unique");
4649 break;
4651 case WHERE_DISTINCT_ORDERED: {
4652 sqlite3DebugPrintf(" DISTINCT=ordered");
4653 break;
4655 case WHERE_DISTINCT_UNORDERED: {
4656 sqlite3DebugPrintf(" DISTINCT=unordered");
4657 break;
4660 sqlite3DebugPrintf("\n");
4661 for(ii=0; ii<pWInfo->nLevel; ii++){
4662 whereLoopPrint(pWInfo->a[ii].pWLoop, sWLB.pWC);
4665 #endif
4666 /* Attempt to omit tables from the join that do not effect the result */
4667 if( pWInfo->nLevel>=2
4668 && pResultSet!=0
4669 && OptimizationEnabled(db, SQLITE_OmitNoopJoin)
4671 Bitmask tabUsed = sqlite3WhereExprListUsage(pMaskSet, pResultSet);
4672 if( sWLB.pOrderBy ){
4673 tabUsed |= sqlite3WhereExprListUsage(pMaskSet, sWLB.pOrderBy);
4675 while( pWInfo->nLevel>=2 ){
4676 WhereTerm *pTerm, *pEnd;
4677 pLoop = pWInfo->a[pWInfo->nLevel-1].pWLoop;
4678 if( (pWInfo->pTabList->a[pLoop->iTab].fg.jointype & JT_LEFT)==0 ) break;
4679 if( (wctrlFlags & WHERE_WANT_DISTINCT)==0
4680 && (pLoop->wsFlags & WHERE_ONEROW)==0
4682 break;
4684 if( (tabUsed & pLoop->maskSelf)!=0 ) break;
4685 pEnd = sWLB.pWC->a + sWLB.pWC->nTerm;
4686 for(pTerm=sWLB.pWC->a; pTerm<pEnd; pTerm++){
4687 if( (pTerm->prereqAll & pLoop->maskSelf)!=0
4688 && !ExprHasProperty(pTerm->pExpr, EP_FromJoin)
4690 break;
4693 if( pTerm<pEnd ) break;
4694 WHERETRACE(0xffff, ("-> drop loop %c not used\n", pLoop->cId));
4695 pWInfo->nLevel--;
4696 nTabList--;
4699 WHERETRACE(0xffff,("*** Optimizer Finished ***\n"));
4700 pWInfo->pParse->nQueryLoop += pWInfo->nRowOut;
4702 /* If the caller is an UPDATE or DELETE statement that is requesting
4703 ** to use a one-pass algorithm, determine if this is appropriate.
4705 assert( (wctrlFlags & WHERE_ONEPASS_DESIRED)==0 || pWInfo->nLevel==1 );
4706 if( (wctrlFlags & WHERE_ONEPASS_DESIRED)!=0 ){
4707 int wsFlags = pWInfo->a[0].pWLoop->wsFlags;
4708 int bOnerow = (wsFlags & WHERE_ONEROW)!=0;
4709 if( bOnerow
4710 || ((wctrlFlags & WHERE_ONEPASS_MULTIROW)!=0
4711 && 0==(wsFlags & WHERE_VIRTUALTABLE))
4713 pWInfo->eOnePass = bOnerow ? ONEPASS_SINGLE : ONEPASS_MULTI;
4714 if( HasRowid(pTabList->a[0].pTab) && (wsFlags & WHERE_IDX_ONLY) ){
4715 if( wctrlFlags & WHERE_ONEPASS_MULTIROW ){
4716 bFordelete = OPFLAG_FORDELETE;
4718 pWInfo->a[0].pWLoop->wsFlags = (wsFlags & ~WHERE_IDX_ONLY);
4723 /* Open all tables in the pTabList and any indices selected for
4724 ** searching those tables.
4726 for(ii=0, pLevel=pWInfo->a; ii<nTabList; ii++, pLevel++){
4727 Table *pTab; /* Table to open */
4728 int iDb; /* Index of database containing table/index */
4729 struct SrcList_item *pTabItem;
4731 pTabItem = &pTabList->a[pLevel->iFrom];
4732 pTab = pTabItem->pTab;
4733 iDb = sqlite3SchemaToIndex(db, pTab->pSchema);
4734 pLoop = pLevel->pWLoop;
4735 if( (pTab->tabFlags & TF_Ephemeral)!=0 || pTab->pSelect ){
4736 /* Do nothing */
4737 }else
4738 #ifndef SQLITE_OMIT_VIRTUALTABLE
4739 if( (pLoop->wsFlags & WHERE_VIRTUALTABLE)!=0 ){
4740 const char *pVTab = (const char *)sqlite3GetVTable(db, pTab);
4741 int iCur = pTabItem->iCursor;
4742 sqlite3VdbeAddOp4(v, OP_VOpen, iCur, 0, 0, pVTab, P4_VTAB);
4743 }else if( IsVirtual(pTab) ){
4744 /* noop */
4745 }else
4746 #endif
4747 if( (pLoop->wsFlags & WHERE_IDX_ONLY)==0
4748 && (wctrlFlags & WHERE_OR_SUBCLAUSE)==0 ){
4749 int op = OP_OpenRead;
4750 if( pWInfo->eOnePass!=ONEPASS_OFF ){
4751 op = OP_OpenWrite;
4752 pWInfo->aiCurOnePass[0] = pTabItem->iCursor;
4754 sqlite3OpenTable(pParse, pTabItem->iCursor, iDb, pTab, op);
4755 assert( pTabItem->iCursor==pLevel->iTabCur );
4756 testcase( pWInfo->eOnePass==ONEPASS_OFF && pTab->nCol==BMS-1 );
4757 testcase( pWInfo->eOnePass==ONEPASS_OFF && pTab->nCol==BMS );
4758 if( pWInfo->eOnePass==ONEPASS_OFF && pTab->nCol<BMS && HasRowid(pTab) ){
4759 Bitmask b = pTabItem->colUsed;
4760 int n = 0;
4761 for(; b; b=b>>1, n++){}
4762 sqlite3VdbeChangeP4(v, -1, SQLITE_INT_TO_PTR(n), P4_INT32);
4763 assert( n<=pTab->nCol );
4765 #ifdef SQLITE_ENABLE_CURSOR_HINTS
4766 if( pLoop->u.btree.pIndex!=0 ){
4767 sqlite3VdbeChangeP5(v, OPFLAG_SEEKEQ|bFordelete);
4768 }else
4769 #endif
4771 sqlite3VdbeChangeP5(v, bFordelete);
4773 #ifdef SQLITE_ENABLE_COLUMN_USED_MASK
4774 sqlite3VdbeAddOp4Dup8(v, OP_ColumnsUsed, pTabItem->iCursor, 0, 0,
4775 (const u8*)&pTabItem->colUsed, P4_INT64);
4776 #endif
4777 }else{
4778 sqlite3TableLock(pParse, iDb, pTab->tnum, 0, pTab->zName);
4780 if( pLoop->wsFlags & WHERE_INDEXED ){
4781 Index *pIx = pLoop->u.btree.pIndex;
4782 int iIndexCur;
4783 int op = OP_OpenRead;
4784 /* iAuxArg is always set if to a positive value if ONEPASS is possible */
4785 assert( iAuxArg!=0 || (pWInfo->wctrlFlags & WHERE_ONEPASS_DESIRED)==0 );
4786 if( !HasRowid(pTab) && IsPrimaryKeyIndex(pIx)
4787 && (wctrlFlags & WHERE_OR_SUBCLAUSE)!=0
4789 /* This is one term of an OR-optimization using the PRIMARY KEY of a
4790 ** WITHOUT ROWID table. No need for a separate index */
4791 iIndexCur = pLevel->iTabCur;
4792 op = 0;
4793 }else if( pWInfo->eOnePass!=ONEPASS_OFF ){
4794 Index *pJ = pTabItem->pTab->pIndex;
4795 iIndexCur = iAuxArg;
4796 assert( wctrlFlags & WHERE_ONEPASS_DESIRED );
4797 while( ALWAYS(pJ) && pJ!=pIx ){
4798 iIndexCur++;
4799 pJ = pJ->pNext;
4801 op = OP_OpenWrite;
4802 pWInfo->aiCurOnePass[1] = iIndexCur;
4803 }else if( iAuxArg && (wctrlFlags & WHERE_OR_SUBCLAUSE)!=0 ){
4804 iIndexCur = iAuxArg;
4805 op = OP_ReopenIdx;
4806 }else{
4807 iIndexCur = pParse->nTab++;
4809 pLevel->iIdxCur = iIndexCur;
4810 assert( pIx->pSchema==pTab->pSchema );
4811 assert( iIndexCur>=0 );
4812 if( op ){
4813 sqlite3VdbeAddOp3(v, op, iIndexCur, pIx->tnum, iDb);
4814 sqlite3VdbeSetP4KeyInfo(pParse, pIx);
4815 if( (pLoop->wsFlags & WHERE_CONSTRAINT)!=0
4816 && (pLoop->wsFlags & (WHERE_COLUMN_RANGE|WHERE_SKIPSCAN))==0
4817 && (pWInfo->wctrlFlags&WHERE_ORDERBY_MIN)==0
4818 && pWInfo->eDistinct!=WHERE_DISTINCT_ORDERED
4820 sqlite3VdbeChangeP5(v, OPFLAG_SEEKEQ); /* Hint to COMDB2 */
4822 VdbeComment((v, "%s", pIx->zName));
4823 #ifdef SQLITE_ENABLE_COLUMN_USED_MASK
4825 u64 colUsed = 0;
4826 int ii, jj;
4827 for(ii=0; ii<pIx->nColumn; ii++){
4828 jj = pIx->aiColumn[ii];
4829 if( jj<0 ) continue;
4830 if( jj>63 ) jj = 63;
4831 if( (pTabItem->colUsed & MASKBIT(jj))==0 ) continue;
4832 colUsed |= ((u64)1)<<(ii<63 ? ii : 63);
4834 sqlite3VdbeAddOp4Dup8(v, OP_ColumnsUsed, iIndexCur, 0, 0,
4835 (u8*)&colUsed, P4_INT64);
4837 #endif /* SQLITE_ENABLE_COLUMN_USED_MASK */
4840 if( iDb>=0 ) sqlite3CodeVerifySchema(pParse, iDb);
4842 pWInfo->iTop = sqlite3VdbeCurrentAddr(v);
4843 if( db->mallocFailed ) goto whereBeginError;
4845 /* Generate the code to do the search. Each iteration of the for
4846 ** loop below generates code for a single nested loop of the VM
4847 ** program.
4849 notReady = ~(Bitmask)0;
4850 for(ii=0; ii<nTabList; ii++){
4851 int addrExplain;
4852 int wsFlags;
4853 pLevel = &pWInfo->a[ii];
4854 wsFlags = pLevel->pWLoop->wsFlags;
4855 #ifndef SQLITE_OMIT_AUTOMATIC_INDEX
4856 if( (pLevel->pWLoop->wsFlags & WHERE_AUTO_INDEX)!=0 ){
4857 constructAutomaticIndex(pParse, &pWInfo->sWC,
4858 &pTabList->a[pLevel->iFrom], notReady, pLevel);
4859 if( db->mallocFailed ) goto whereBeginError;
4861 #endif
4862 addrExplain = sqlite3WhereExplainOneScan(
4863 pParse, pTabList, pLevel, ii, pLevel->iFrom, wctrlFlags
4865 pLevel->addrBody = sqlite3VdbeCurrentAddr(v);
4866 notReady = sqlite3WhereCodeOneLoopStart(pWInfo, ii, notReady);
4867 pWInfo->iContinue = pLevel->addrCont;
4868 if( (wsFlags&WHERE_MULTI_OR)==0 && (wctrlFlags&WHERE_OR_SUBCLAUSE)==0 ){
4869 sqlite3WhereAddScanStatus(v, pTabList, pLevel, addrExplain);
4873 /* Done. */
4874 VdbeModuleComment((v, "Begin WHERE-core"));
4875 return pWInfo;
4877 /* Jump here if malloc fails */
4878 whereBeginError:
4879 if( pWInfo ){
4880 pParse->nQueryLoop = pWInfo->savedNQueryLoop;
4881 whereInfoFree(db, pWInfo);
4883 return 0;
4887 ** Generate the end of the WHERE loop. See comments on
4888 ** sqlite3WhereBegin() for additional information.
4890 void sqlite3WhereEnd(WhereInfo *pWInfo){
4891 Parse *pParse = pWInfo->pParse;
4892 Vdbe *v = pParse->pVdbe;
4893 int i;
4894 WhereLevel *pLevel;
4895 WhereLoop *pLoop;
4896 SrcList *pTabList = pWInfo->pTabList;
4897 sqlite3 *db = pParse->db;
4899 /* Generate loop termination code.
4901 VdbeModuleComment((v, "End WHERE-core"));
4902 sqlite3ExprCacheClear(pParse);
4903 for(i=pWInfo->nLevel-1; i>=0; i--){
4904 int addr;
4905 pLevel = &pWInfo->a[i];
4906 pLoop = pLevel->pWLoop;
4907 if( pLevel->op!=OP_Noop ){
4908 #ifndef SQLITE_DISABLE_SKIPAHEAD_DISTINCT
4909 int addrSeek = 0;
4910 Index *pIdx;
4911 int n;
4912 if( pWInfo->eDistinct==WHERE_DISTINCT_ORDERED
4913 && (pLoop->wsFlags & WHERE_INDEXED)!=0
4914 && (pIdx = pLoop->u.btree.pIndex)->hasStat1
4915 && (n = pLoop->u.btree.nIdxCol)>0
4916 && pIdx->aiRowLogEst[n]>=36
4918 int r1 = pParse->nMem+1;
4919 int j, op;
4920 for(j=0; j<n; j++){
4921 sqlite3VdbeAddOp3(v, OP_Column, pLevel->iIdxCur, j, r1+j);
4923 pParse->nMem += n+1;
4924 op = pLevel->op==OP_Prev ? OP_SeekLT : OP_SeekGT;
4925 addrSeek = sqlite3VdbeAddOp4Int(v, op, pLevel->iIdxCur, 0, r1, n);
4926 VdbeCoverageIf(v, op==OP_SeekLT);
4927 VdbeCoverageIf(v, op==OP_SeekGT);
4928 sqlite3VdbeAddOp2(v, OP_Goto, 1, pLevel->p2);
4930 #endif /* SQLITE_DISABLE_SKIPAHEAD_DISTINCT */
4931 /* The common case: Advance to the next row */
4932 sqlite3VdbeResolveLabel(v, pLevel->addrCont);
4933 sqlite3VdbeAddOp3(v, pLevel->op, pLevel->p1, pLevel->p2, pLevel->p3);
4934 sqlite3VdbeChangeP5(v, pLevel->p5);
4935 VdbeCoverage(v);
4936 VdbeCoverageIf(v, pLevel->op==OP_Next);
4937 VdbeCoverageIf(v, pLevel->op==OP_Prev);
4938 VdbeCoverageIf(v, pLevel->op==OP_VNext);
4939 #ifndef SQLITE_DISABLE_SKIPAHEAD_DISTINCT
4940 if( addrSeek ) sqlite3VdbeJumpHere(v, addrSeek);
4941 #endif
4942 }else{
4943 sqlite3VdbeResolveLabel(v, pLevel->addrCont);
4945 if( pLoop->wsFlags & WHERE_IN_ABLE && pLevel->u.in.nIn>0 ){
4946 struct InLoop *pIn;
4947 int j;
4948 sqlite3VdbeResolveLabel(v, pLevel->addrNxt);
4949 for(j=pLevel->u.in.nIn, pIn=&pLevel->u.in.aInLoop[j-1]; j>0; j--, pIn--){
4950 sqlite3VdbeJumpHere(v, pIn->addrInTop+1);
4951 if( pIn->eEndLoopOp!=OP_Noop ){
4952 sqlite3VdbeAddOp2(v, pIn->eEndLoopOp, pIn->iCur, pIn->addrInTop);
4953 VdbeCoverage(v);
4954 VdbeCoverageIf(v, pIn->eEndLoopOp==OP_PrevIfOpen);
4955 VdbeCoverageIf(v, pIn->eEndLoopOp==OP_NextIfOpen);
4957 sqlite3VdbeJumpHere(v, pIn->addrInTop-1);
4960 sqlite3VdbeResolveLabel(v, pLevel->addrBrk);
4961 if( pLevel->addrSkip ){
4962 sqlite3VdbeGoto(v, pLevel->addrSkip);
4963 VdbeComment((v, "next skip-scan on %s", pLoop->u.btree.pIndex->zName));
4964 sqlite3VdbeJumpHere(v, pLevel->addrSkip);
4965 sqlite3VdbeJumpHere(v, pLevel->addrSkip-2);
4967 #ifndef SQLITE_LIKE_DOESNT_MATCH_BLOBS
4968 if( pLevel->addrLikeRep ){
4969 sqlite3VdbeAddOp2(v, OP_DecrJumpZero, (int)(pLevel->iLikeRepCntr>>1),
4970 pLevel->addrLikeRep);
4971 VdbeCoverage(v);
4973 #endif
4974 if( pLevel->iLeftJoin ){
4975 int ws = pLoop->wsFlags;
4976 addr = sqlite3VdbeAddOp1(v, OP_IfPos, pLevel->iLeftJoin); VdbeCoverage(v);
4977 assert( (ws & WHERE_IDX_ONLY)==0 || (ws & WHERE_INDEXED)!=0 );
4978 if( (ws & WHERE_IDX_ONLY)==0 ){
4979 sqlite3VdbeAddOp1(v, OP_NullRow, pTabList->a[i].iCursor);
4981 if( (ws & WHERE_INDEXED)
4982 || ((ws & WHERE_MULTI_OR) && pLevel->u.pCovidx)
4984 sqlite3VdbeAddOp1(v, OP_NullRow, pLevel->iIdxCur);
4986 if( pLevel->op==OP_Return ){
4987 sqlite3VdbeAddOp2(v, OP_Gosub, pLevel->p1, pLevel->addrFirst);
4988 }else{
4989 sqlite3VdbeGoto(v, pLevel->addrFirst);
4991 sqlite3VdbeJumpHere(v, addr);
4993 VdbeModuleComment((v, "End WHERE-loop%d: %s", i,
4994 pWInfo->pTabList->a[pLevel->iFrom].pTab->zName));
4997 /* The "break" point is here, just past the end of the outer loop.
4998 ** Set it.
5000 sqlite3VdbeResolveLabel(v, pWInfo->iBreak);
5002 assert( pWInfo->nLevel<=pTabList->nSrc );
5003 for(i=0, pLevel=pWInfo->a; i<pWInfo->nLevel; i++, pLevel++){
5004 int k, last;
5005 VdbeOp *pOp;
5006 Index *pIdx = 0;
5007 struct SrcList_item *pTabItem = &pTabList->a[pLevel->iFrom];
5008 Table *pTab = pTabItem->pTab;
5009 assert( pTab!=0 );
5010 pLoop = pLevel->pWLoop;
5012 /* For a co-routine, change all OP_Column references to the table of
5013 ** the co-routine into OP_Copy of result contained in a register.
5014 ** OP_Rowid becomes OP_Null.
5016 if( pTabItem->fg.viaCoroutine ){
5017 testcase( pParse->db->mallocFailed );
5018 translateColumnToCopy(pParse, pLevel->addrBody, pLevel->iTabCur,
5019 pTabItem->regResult, 0);
5020 continue;
5023 /* If this scan uses an index, make VDBE code substitutions to read data
5024 ** from the index instead of from the table where possible. In some cases
5025 ** this optimization prevents the table from ever being read, which can
5026 ** yield a significant performance boost.
5028 ** Calls to the code generator in between sqlite3WhereBegin and
5029 ** sqlite3WhereEnd will have created code that references the table
5030 ** directly. This loop scans all that code looking for opcodes
5031 ** that reference the table and converts them into opcodes that
5032 ** reference the index.
5034 if( pLoop->wsFlags & (WHERE_INDEXED|WHERE_IDX_ONLY) ){
5035 pIdx = pLoop->u.btree.pIndex;
5036 }else if( pLoop->wsFlags & WHERE_MULTI_OR ){
5037 pIdx = pLevel->u.pCovidx;
5039 if( pIdx
5040 && (pWInfo->eOnePass==ONEPASS_OFF || !HasRowid(pIdx->pTable))
5041 && !db->mallocFailed
5043 last = sqlite3VdbeCurrentAddr(v);
5044 k = pLevel->addrBody;
5045 pOp = sqlite3VdbeGetOp(v, k);
5046 for(; k<last; k++, pOp++){
5047 if( pOp->p1!=pLevel->iTabCur ) continue;
5048 if( pOp->opcode==OP_Column ){
5049 int x = pOp->p2;
5050 assert( pIdx->pTable==pTab );
5051 if( !HasRowid(pTab) ){
5052 Index *pPk = sqlite3PrimaryKeyIndex(pTab);
5053 x = pPk->aiColumn[x];
5054 assert( x>=0 );
5056 x = sqlite3ColumnOfIndex(pIdx, x);
5057 if( x>=0 ){
5058 pOp->p2 = x;
5059 pOp->p1 = pLevel->iIdxCur;
5061 assert( (pLoop->wsFlags & WHERE_IDX_ONLY)==0 || x>=0
5062 || pWInfo->eOnePass );
5063 }else if( pOp->opcode==OP_Rowid ){
5064 pOp->p1 = pLevel->iIdxCur;
5065 pOp->opcode = OP_IdxRowid;
5066 }else if( pOp->opcode==OP_IfNullRow ){
5067 pOp->p1 = pLevel->iIdxCur;
5073 /* Final cleanup
5075 pParse->nQueryLoop = pWInfo->savedNQueryLoop;
5076 whereInfoFree(db, pWInfo);
5077 return;